paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1612.02878 | 1 | 1612 | 2016-12-09T00:55:52 | Deformations and generalized derivations of Lie conformal superalgebras | [
"math.RA"
] | The purpose of this paper is to extend the cohomology and conformal derivation theories of the classical Lie conformal algebras to Lie conformal superalgebras. Firstly, we construct the semidirect product of a Lie conformal superalgebra and its conformal module, and study derivations of this semidirect product. Secondly, we develop cohomology theory of Lie conformal superalgebras and discuss some applications to the study of deformations of Lie conformal superalgebras. Finally, we introduce generalized derivations of Lie conformal superalgebras and study their properties. | math.RA | math | Deformations and generalized derivations of
Lie conformal superalgebras
6
1
0
2
c
e
D
9
Jun Zhao1, Liangyun Chen1∗, Lamei Yuan2
1School of Mathematics and Statistics, Northeast Normal University,
Changchun 130024, China
2Academy of Fundamental and Interdisciplinary Sciences,
Harbin Institute of Technology, Harbin 150080, China
]
.
A
R
h
t
a
m
[
1
v
8
7
8
2
0
.
2
1
6
1
:
v
i
X
r
a
Abstract
The purpose of this paper is to extend the cohomology and conformal deriva-
tion theories of the classical Lie conformal algebras to Lie conformal superalgebras.
Firstly, we construct the semidirect product of a Lie conformal superalgebra and
its conformal module, and study derivations of this semidirect product. Secondly,
we develop cohomology theory of Lie conformal superalgebras and discuss some
applications to the study of deformations of Lie conformal superalgebras. Finally,
we introduce generalized derivations of Lie conformal superalgebras and study their
properties.
Key words: Lie conformal superalgebras, semidirect product, cohomology, deforma-
tions, generalized derivations
Mathematics Subject Classification(2010): 16S70, 17A42, 17B10, 17B56, 17B70
1
Introduction
Lie conformal superalgebras encode the singular part of the operator product expansion
of chiral fields in two-dimensional quantum field theory. On the other hand, they are
closely connected to the notion of a formal distribution Lie superalgebra (g, F ), that is
a Lie superalgebra g spanned by the coefficients of a family F of mutually local formal
distributions. Conversely, to a Lie conformal superalgebra R one can associate a formal
distribution Lie superalgebra (LieR, R) which establishes an equivalence between the
category of Lie conformal superalgebras and the category of equivalence classes of formal
Corresponding author(L. Chen): [email protected].
Supported by NNSF of China (Nos. 11171055, 11471090 and 11301109).
1
distribution Lie superalgebras obtained as quotients of Lie R by irregular ideals. See
[2, 3, 5] in details.
The notion of a Lie conformal superalgebra was introduced in [5], which gave deriva-
tions and representations of Lie conformal superalgebras. The cohomology theory of Lie
conformal algebras was developed in [1]. Lately, the generalized derivation theory of Lie
conformal algebras were studied in [6], extending the generalized derivation theory of Lie
algebras given in [8]. [7] study conformal derivations of the semidirect product. [9] study
Hom-Nijienhuis operators and T *-extensions of hom-Lie superalgebras.
In the present
paper, we aim to do same in [7] for Lie conformal superalgebras, extend the cohomology
theory and study the generalized derivations of Lie conformal algebras to the super case.
Furthermore, by a 2-cocycle Nijienhuis operator generate a deformation.
The paper is organized as follows. In Section 2, we recall notions of derivations and
modules of a Lie conformal superalgebra. Moreover, we construct the semidirect product
of a Lie conformal superalgebra and its conformal module, and study derivations of this
semidirect product.
In Section 3, we define cohomology and Nijienhuis operators of Lie conformal super-
algebras, and show that the deformation generated by a 2-cocycle Nijienhuis operator is
trivial.
In Section 4, we introduce different kinds of generalized derivations of Lie confor-
mal superalgebras, and study their properties and connections, extending some results
obtained in [8].
Throughout this paper, all vector spaces, linear maps, and tensor products are over
the complex field C. In addition to the standard notations Z and R, we use Z+ to denote
the set of nonnegative integers. R is a Z2-graded C[∂]-module, x ∈ R means x is a
homogeneous element, a ∈ Z2 is the degree of a.
2 Conformal derivations of semidirect products of
Lie conformal superalgebras and their conformal
modules
First we present the definition of Lie conformal superalgebras given in [2, Definition 2.1].
Definition 2.1. A Lie conformal superalgebra R is a left Z2-graded C[∂]-module, and
for any n ∈ Z≥0 there is a family of C-bilinear n-products from R ⊗ R to R satisfying
the following conditions:
(C0) For any a, b ∈ R, there is an N such that a(n)b = 0 for n ≫ N,
(C1) For any a, b ∈ R and n ∈ Z≥0, (∂a)(n)b = −na(n)b,
2
(C2) For any a, b ∈ R and n ∈ Z≥0,
a(n)b = −(−1)ab
∞
Xj=0
(−1)j+n 1
j!
∂j(b(n+j)a),
(C3) For any a, b, c ∈ R and m, n ∈ Z≥0,
a(m)(b(n)c) =
m
Xj=0
j )(a(j)b)(m+n−j)c + (−1)abb(n)(a(m)c).
(m
Note that if we define λ-bracket [−λ−]:
[aλb] =
∞
Xn=0
λn
n!
a(n)b, a, b ∈ R.
(2.1)
That is, R is a Lie conformal superalgebra if and only if [−λ−] satisfies the following
axioms:
(C1)λ Conformal sesquilinearity :
(C2)λ
(C3)λ
Skew − symmetry :
Jacobi identity :
[(∂a)λb] = −λ[aλb];
[aλb] = −(−1)ab[b−∂−λa];
[aλ[bµc]] = [[aλb]λ+µc] + (−1)ab[bµ[aλc]].
Example 2.2. Let R = C[∂]L ⊕ C[∂]E be a free Z2-graded C[∂]-module. Define
[LλL] = (∂ + 2λ)L, [LλE] = (∂ +
3
2
λ)E, [EλL] = (
1
2
∂ +
3
2
λ)E, [EλE] = 0,
where R¯0 = C[∂]L and R¯1 = C[∂]E. Then R is a Lie conformal superalgebra.
Example 2.3. (Neveu-Schwarz Lie conformal superalgebra) Let NS = C[∂]L ⊕ C[∂]G be
a free Z2-graded C[∂]-module. Define
[LλL] = (∂ + 2λ)L, [LλG] = (∂ +
3
2
λ)G, [GλL] = (
1
2
∂ +
3
2
λ)G, [GλG] = L,
where NS¯0 = C[∂]L and NS¯1 = C[∂]G. Then NS is a Lie conformal superalgebra. We call
it Neveu-Schwarz Lie conformal superalgebra.
Proposition 2.4. Let R = C[∂]x ⊕ C[∂]y be a Lie conformal superalgebra that is a free
Z2-graded C[∂]-module, where R¯0 = C[∂]x and R¯1 = C[∂]y. Then R is isomorphic to one
of the following Lie conformal superalgebras:
(1) R1 : [xλx] = 0, [xλy] = 0, [yλy] = p(∂)x, ∀ p(∂) ∈ C[∂];
(2) R2 : [xλx] = 0, [yλy] = 0, [xλy] = q(λ)y, ∀ q(λ) ∈ C[λ];
(3) R3 : [xλx] = (∂ + 2λ)x, [xλy] = 0, [yλy] = 0;
3
(4) R4 : [xλx] = (∂ + 2λ)x, [xλy] = (∂ + βλ)y, [yλy] = 0, ∀ β ∈ C;
(5) R5 : [xλx] = (∂ + 2λ)x, [xλy] = (∂ + 3
2 λ)y, [yλy] = αx, ∀ α ∈ C.
Proof. Let us set [xλx] = a(∂, λ)x, [xλy] = b(∂, λ)y, and [yλy] = d(∂)x (because of
i=0 bi(λ)∂i, and d(∂, λ) =
i=0 di∂i, with an(λ) 6= 0, bm(λ) 6= 0, and dq 6= 0. Consider the Jacobi identity [xλ[xµx]] =
i=0 ai(λ)∂i, b(∂, λ) = Pm
[yλy] = [y−∂−λy]), where a(∂, λ) = Pn
Pq
[[xλx]λ+µx] + [xµ[xλx]], this is equivalent to saying that
a(∂ + λ, µ)a(∂, λ)x = a(−λ − µ, λ)a(∂, λ + µ)x + a(∂ + µ, λ)a(∂, µ)x.
(2.2)
Equating terms of degree 2n−1 in ∂ in both sides, we get, if n > 1 : n(λ−µ)an(λ)an(µ) =
0, which shows that an(λ) = 0. So a(∂, λ) = a0(λ) + a1(λ)∂. If we put λ = µ in Eq.(2.2),
we get a(∂, 2λ)a(−2λ, λ) = 0, hence a0(λ) = 2λa1(λ), therefore a(∂, λ) = a1(λ)(∂ + 2λ).
Plugging this into [xλx] = −[x−∂−λx], we get that a1(λ) is a constant C. So [xλx] =
C(∂ + 2λ)x.
Case 1
If C = 0, Consider the Jacobi identity [xλ[xµy]] = [[xλx]λ+µy] + [xµ[xλy]], this is
equivalent to saying that
b(∂ + λ, µ)b(∂, λ)y = b(∂ + µ, λ)b(∂, µ)y.
Equating terms of degree 2m−1 in ∂ in both sides, we get, if m ≥ 1 : m(λ−µ)bm(λ)bm(µ) =
0, which shows that bm(λ) = 0. So b(∂, λ) = b0(λ). So [xλy] = b0(λ)y.
(1) If b0(λ) = 0, then [xλy] = [yλx] = 0. In this situation, R is isomorphic to R1.
(2) If b0(λ) 6= 0, consider the Jacobi identity [xλ[yµy]] = [[xλy]λ+µy] + [yµ[xλy]], this
is equivalent to saying that
0 = b0(λ)d(∂)x + b0(λ)d(∂)x.
So d(∂) = 0, thus [yλy] = 0. Then R is isomorphic to R2.
Case 2
If C 6= 0, we can let C = 1(otherwise, change x by a complex multiple). Then
[xλx] = (∂ + 2λ)x. Consider the Jacobi identity [xλ[xµy]] = [[xλx]λ+µy] + [xµ[xλy]], this is
equivalent to saying that
b(∂ + λ, µ)b(∂, λ)y = (λ − µ)b(∂, λ + µ)y + b(∂ + µ, λ)b(∂, µ)y.
(2.3)
Equating terms of degree 2m−1 in ∂ in both sides, we get, if m > 1 : m(λ−µ)bm(λ)bm(µ) =
0, which shows that bm(λ) = 0. So b(∂, λ) = b0(λ) + b1(λ)∂. Substituting into Eq.(2.3),
we get
b1(λ)b1(µ) = b1(λ + µ)
(2.4)
4
and
λb0(λ)b1(µ) = (λ − µ)b0(λ + µ) + µb1(λ)b0(µ).
(2.5)
By Eq.(2.4), we have b1(λ) = 0 or 1.
(1) If b1(λ) = 0, then (λ − µ)b0(λ + µ) = 0, so b0(λ) = 0, that is [xλy] = 0. Consider
the Jacobi identity [xλ[yµy]] = [[xλy]λ+µy] + [yµ[xλy]], this is equivalent to saying that
d(∂ + λ)(∂ + 2λ) = 0, i.e. [yλy] = 0. Then R is isomorphic to R3.
(2) If b1(λ) = 1, Substituting into Eq.(2.5), we get λb0(λ) − µb0(µ) = (λ − µ)b0(λ +
µ). So b0(λ) = βλ, and [xλy] = (∂ + βλ)y. Consider the Jacobi identity [xλ[yµy]] =
[[xλy]λ+µy] + [yµ[xλy]], this is equivalent to saying that
(∂ + 2λ)d(∂ + λ)x = (∂ − λ + 2βλ)d(∂)x.
(2.6)
Equating terms of the most degree in λ in both sides, we get d(∂) = C1.
(i) If C1 = 0, then [yλy] = 0. Then R is isomorphic to R4.
(ii) If C1 6= 0, Substituting into Eq.(2.6), we get β = 3
2. Then R is isomorphic to R5.
Definition 2.5. A linear map ρ : R → R′ is a homomorphism of Lie conformal superal-
gebras if ρ satisfies ρ∂ = ∂ρ and ρ([aλb]) = [ρ(a)λρ(b)], ∀a, b ∈ R.
Example 2.6. [5] Let g = g¯0 ⊕ g¯1 be a complex Lie superalgebra with Lie bracket [−, −].
Let (Curg)θ := C[∂] ⊗ gθ be the free C[∂]-module. Then Curg = (Curg)¯0 ⊕ (Curg)¯1 is
a Lie conformal superalgebra, called current Lie conformal superalgebra, with λ-bracket
given by:
[(f (∂) ⊗ a)λ(g(∂) ⊗ b)] := f (−λ)g(∂ + λ) ⊗ [a, b], a, b ∈ R.
(2.7)
Definition 2.7. [5] Let M and N be Z2-graded C[∂]-modules. A conformal linear map
of degree θ from M to N is a sequence f = {f(n)}n∈Z≥0 of f(n) ∈ HomC(M, N) satisfying
that
∂N f(n) − f(n)∂M = −nf(n−1), n ∈ Z≥0
and f(n)(Mµ) ⊆ Nµ+θ, µ, θ ∈ Z2.
Set fλ = P∞
only if
n=0
λn
n! f(n). Then f = {f(n)}n∈Z≥0 is a conformal linear map of degree θ if and
fλ∂M = (∂N + λ)fλ
and fλ(Mµ) ⊆ Nµ+θ[λ], µ, θ ∈ Z2.
Definition 2.8. An associative conformal superalgebra R is a left Z2-graded C[∂]-module
endowed with a λ-product from R ⊗ R to C[λ] ⊗ R, for any a, b, c ∈ R, satisfying the
following conditions:
(1) (∂a)λb = −λaλb, aλ(∂b) = (∂ + λ)(aλb),
(2) aλ(bµc) = (aλb)λ+µc.
5
Let Chom(M, N)θ denote the set of conformal linear maps of degree θ from M to N.
Then Chom(M, N) = Chom(M, N)¯0 ⊕ Chom(M, N)¯1 is a Z2-graded C[∂]-module via:
∂f(n) = −nf(n−1), equivalently, ∂fλ = −λfλ.
The composition fλg : L → N ⊗C[λ] of conformal linear maps f : M → N and g : L → M
is given by
(fλg)λ+µ = fλgµ,
∀ f, g ∈ Chom(M, N).
If M is a finitely generated Z2-graded C[∂]-module, then Cend(M) := Chom(M, M)
is an associative conformal superalgebra with respect to the above composition. Thus,
Cend(M) becomes a Lie conformal superalgebra, denoted as gc(M), with respect to the
following λ-bracket(see [5, Example1.1]):
[fλg]µ = fλgµ−λ − (−1)f ggµ−λfλ.
(2.8)
Hereafter all Z2-graded C[∂]-modules are supposed to be finitely generated.
Definition 2.9. [2, Definition 2.2] A Z2-graded C[∂]-module M is a conformal module of
a Lie conformal superalgebra R if there is a homomorphism of Lie conformal superalgebras
ρ : R → Cend(M).
Furthermore, we call (ρ, M) is a representation of the Lie conformal superalgebra R.
If (ρ, M) is a representation of a Lie conformal superalgebra R, it is obvious that we have
the following relations;
m
Xj=0
j )ρ(a(j)b)(m+n−j) + (−1)abρ(b)(n)ρ(a)(m), ρ(∂(a))(n) = −nρ(a)(n);
(m
ρ(a)(m)ρ(b)(n) =
equivalently,
ρ(a)λρ(b)µ − (−1)abρ(b)µρ(a)λ = [ρ(a)λρ(b)]λ+µ = ρ([aλb])λ+µ, ρ(∂(a))λ = −λρ(a)λ.
Example 2.10. Let g = g¯0 ⊕ g¯1 be a finite dimensional complex Lie superalgebra, π :
g → End(M) a finite dimensional representation of g. Then the free Z2-graded C[∂]-
module C[∂] ⊗ M is a conformal module of Curg, with module structure ρ : Curg →
Cend(C[∂] ⊗ M) given by:
ρ(f (∂) ⊗ a)λ(g(∂) ⊗ m) = f (−λ)g(∂ + λ) ⊗ π(a)(m), ∀ f (∂), g(∂) ∈ C[∂], a ∈ g, m ∈ M.
Definition 2.11. [5] Let R be a Lie conformal superalgebra. d ∈ Cend(R) is a conformal
derivation if for any a, b ∈ R it holds that
d(m)(a(n)b) =
m
Xj=0
j )(d(j)a)(m+n−j)b + (−1)ada(n)(d(m)(b));
(m
equivalently,
dλ([aµb]) = [(dλ(a))λ+µb] + (−1)ad[aµ(dλ(b))].
(2.9)
6
For any r ∈ R, dr
λ is called an inner conformal derivation of R if dr
λ(r′) = [rλr′],
∀r′ ∈ R.
Example 2.12. Let R = C[∂]L ⊕ Curg be a free Z2-graded C[∂]-module, where Curg
is current Lie conformal superalgebra. Define dL : Curg → Curg by dL
λ g = (∂ + λ)g for
every g ∈ g ⊂ Curg. Then dL is a conformal derivation degree 0 of Curg. Furthermore, if
we define
[LλL] = (∂ + 2λ)L, [Lλg] = dL
λ g, [gλh] = [g, h],
where R¯0 = C[∂]L ⊕ Curg¯0 and R¯1 = Curg¯1. Then R is a Lie conformal superalgebra.
Example 2.13. Every conformal derivation of Neveu-Schwarz Lie conformal superalgebra
is an inner conformal derivation.
Proof. Case 1
Suppose that d = 0, and dλL = p(∂, λ)L, dλG = g(∂, λ)G, where p(∂, λ) =
i=0 di(λ)∂i. If d is a conformal derivation, it needs to satisfy
dλ[LµL] = [(dλL)λ+µL] + [Lµ(dλL)], this is equivalent to saying that
i=0 ai(λ)∂i, g(∂, λ) = Pq
Pn
(∂ + λ + 2µ)p(∂, λ)L = p(−λ − µ, λ)(∂ + 2λ + 2µ)L + p(∂ + µ, λ)(∂ + 2µ)L.
(2.10)
Equating terms of degree n in ∂ in both sides, we get, if n > 1 : (λ − nµ)an(λ) = 0, which
shows that an(λ) = 0. So p(∂, λ) = a0(λ) + a1(λ)∂, substituting into Eq.(2.10), we get
a0(λ) = 2λa1(λ), hence dλL = a1(λ)(∂ + 2λ)L.
Similarly, dλ[GµG] = [(dλG)λ+µG] + [Gµ(dλG)], it is equivalent to
p(∂, λ)L = g(−λ − µ, λ)L + g(∂ + µ, λ)L.
(2.11)
The degree of ∂ in p(∂, λ) is at most one, so g(∂, λ) = d0(λ) + d1(λ)∂, substituting into
Eq.(2.11), we get d0(λ) = 3
2 λa1(λ), hence dλG = a1(λ)(∂ + 3
2 λd1(λ) = 3
2 λ)G.
So dλ = a1(λ)adLλ.
Case 2
Suppose that d = 1, and dλL = q(∂, λ)G, dλG = f (∂, λ)L, where q(∂, λ) =
i=0 ci(λ)∂i. If d is a conformal derivation, it needs to satisfy
dλ[LµL] = [(dλL)λ+µL] + [Lµ(dλL)], this is equivalent to saying that
i=0 bi(λ)∂i, f (∂, λ) = Pp
Pm
(∂ + λ + 2µ)q(∂, λ)G = q(−λ − µ, λ)(
1
2
∂ +
3
2
λ +
3
2
µ)G + q(∂ + µ, λ)(∂ +
3
2
µ)G. (2.12)
Equating terms of degree m in ∂ in both sides, we get, if m > 1 : (λ + 1
2 µ − mµ)bm(λ) = 0,
which shows that bm(λ) = 0. So q(∂, λ) = b0(λ) + b1(λ)∂, substituting into Eq.(2.12), we
get b0(λ) = 3λb1(λ), hence dλL = b1(λ)(∂ + 3λ)G.
Similarly, dλ[GµG] = [(dλG)λ+µG] − [Gµ(dλG)], it is equivalent to
q(∂, λ)G = f (−λ − µ, λ)(∂ +
µ)G − f (∂ + µ, λ)(
1
2
∂ +
3
2
µ)G.
(2.13)
3
2
λ +
3
2
7
The degree of ∂ in q(∂, λ) is at most one, so f (∂, λ) = c0(λ), substituting into Eq.(2.13),
we get c0(λ) = 2b1(λ), hence dλG = 2b1(λ)L.
So dλ = 2b1(λ)adGλ. Then every conformal derivation of Neveu-Schwarz Lie confor-
mal superalgebra is an inner conformal derivation.
Definition 2.14. Let R be a Lie conformal superalgebra with λ-bracket [−λ−], M a
conformal module of R and ρ : R → C[λ] ⊗ Cend(M) : r 7→ ρ(r)λ the corresponding
representation. A conformal linear map d ∈ Chom(R, M) is a conformal derivation from
R to M if
dλ([rµr′]) = (−1)rdρ(r)µ(dλ(r′)) − (−1)(r+d)r′ρ(r′)−∂−λ−µ(dλ(r)).
(2.14)
Note that if ρ is the adjoint representation (that is ρ(a)λb = [aλb]) of R, then Eq.(2.14)
is Eq.(2.9). We have the following:
Lemma 2.15. Let R be a Lie conformal superalgebra with λ-bracket [−λ−], M a confor-
mal module of R and ρ : R → C[λ] ⊗ Cend(M) : r 7→ ρ(r)λ the corresponding represen-
tation. For any m ∈ M, define a C-linear map
dm : R → C[λ] ⊗ M : dm
λ (r) = −(−1)rmρ(r)−∂−λ(m), r ∈ R.
(2.15)
Then dm is a conformal derivation from R to M.
Proof. For any r ∈ R, m ∈ Mθ, according to Definition 2.9 and Eq.(2.15), we have
λ (∂r) = −(−1)rmρ(∂r)−∂−λ(m) = −(−1)rm(∂ + λ)ρ(r)−∂−λ(m) = (∂ + λ)dm
dm
λ (r),
and dm(Rη) ⊆ Mη+θ, which means that dm ∈ Cend(R, M)θ. For any r′ ∈ R,
− dm
λ ([rµr′]) = −Xs≥0
µs
s!
λ (r(s)r′)
dm
= (−1)(r+r′)mXs≥0
= (−1)(r+r′)m Xs,t≥0
µs
s!
ρ(r(s)r′)−∂−λ(m)
µs
s!
(−∂ − λ)t
t!
ρ(r(s)r′)(t)(m).
(2.16)
Since ρ(r) ∈ Cend(M)r is a conformal linear map, it follows from
ρ(r)µ((−∂ − λ)tm′) = (−∂ − λ − µ)tρ(r)µ(m′), m′ ∈ M,
and hence,
−(−1)rmρ(r)µ(dm
λ (r′))
= (−1)rm(−1)r′mρ(r)µ(ρ(r′)−∂−λ(m))
8
= (−1)(r+r′)mρ(r)µ(Xt≥0
= (−1)(r+r′)m Xs,t≥0
µs
s!
(−∂ − λ)t
t!
ρ(r′)(t)(m))
(−∂ − λ − µ)t
t!
ρ(r)(s)(ρ(r′)(t)(m)).
(2.17)
Similarly, ρ(r′) ∈ Cend(M)r′ is a conformal linear map, it follows from
ρ(r′)−∂−λ−µ((−∂ − λ)tm′) = µtρ(r′)−∂−λ−µ(m′),
and hence,
−(−1)r′(r+m)ρ(r′)−∂−λ−µ(dm
λ (r))
= (−1)r′(r+m)+mrρ(r′)−∂−λ−µ(ρ(r)−∂−λ(m))
= (−1)r′(r+m)+mr Xs,t≥0
µs
s!
(−∂ − λ − µ)t
t!
ρ(r′)(t)(ρ(r)(s)(m)).
(2.18)
Since ρ is a representation, by Definition 2.9, Eqs.(2.16), (2.17) and (2.18), dm is a
conformal derivation from R to M.
Definition 2.16. Let R be a Lie conformal superalgebra with λ-bracket [−λ−], M a
conformal module of R and ρ : R → C[λ] ⊗ Cend(M) : r 7→ ρ(r)λ the corresponding
representation. dm given by Eq.(2.15) is called an inner conformal derivation from R to
M.
Remark 2.17. If ρ is the adjoint representation of R, then an inner conformal derivation
from R to R is nothing rather than an inner conformal derivation of R. In fact, for any
r, r′ ∈ R, by Definition 2.1 (C2)λ it follows from
λ(r′) = −(−1)rr′ρ(r′)−∂−λ(r) = −(−1)rr′[r′
dr
−∂−λr] = [rλr′].
(2.19)
Lemma 2.18. If f, g ∈ Cend(M), then for any m ∈ M,
(1) fλ(g−∂−µm) = (fλg)−∂−µm,
(2) f−∂−λ(gµm) = (f−∂−µg)−∂−λ+µm,
(3) f−∂−λ(g−∂−µm) = (f−∂+µ−λg)−∂−µm.
Lemma 2.19. Let R be a Lie conformal superalgebra with λ-bracket [−λ−], M a confor-
mal module of R and ρ : R → C[λ] ⊗ Cend(M) : r 7→ ρ(r)λ the corresponding represen-
tation. Define a λ-bracket [−λ−]M on R ⊕ M = (R ⊕ M)¯0 ⊕ (R ⊕ M)¯1 by
[(r + m)λ(r′ + m′)]M = [rλr′] + ρ(r)λm′ − (−1)r′mρ(r′)−∂−λm,
(2.20)
∀ r + m, r′ + m′ ∈ R ⊕ M, where (R ⊕ M)θ = Rθ ⊕ Mθ, θ ∈ Z2. Then R ⊕ M is a Lie
conformal superalgebra, called the semidirect product of R and M, and denote by R ⋉ M.
9
Proof. ∀ r + m, r′ + m′, r′′ + m′′ ∈ R ⊕ M, note that R ⊕ M is equipped with a
C[∂]-module structure via
∂(r + m) = ∂(r) + ∂(m).
A direct computation shows that
[∂(r + m)λ(r′ + m′)]M = [(∂r + ∂m)λ(r′ + m′)]M
= [(∂r)λr′] + ρ(∂r)λ(m′) − (−1)r′mρ(r′)−∂−λ(∂m)
= −λ[rλr′] − λρ(r)λ(m′) − (−1)r′m(∂ − λ − ∂)ρ(r′)−∂−λ(m)
= −λ([rλr′] + ρ(r)λ(m′) − (−1)r′mρ(r′)−∂−λ(m))
= −λ[(r + m)λ(r′ + m′)]M .
Thus (C1)λ holds. (C2)λ follows from
[(r′ + m′)−∂−λ(r + m)]M = [r′
−∂−λr] + ρ(r′)−∂−λ(m) − (−1)r′mρ(r)λ(m′)
= −(−1)r′m([rλr′] + ρ(r)λ(m′) − (−1)r′mρ(r′)−∂−λ(m))
= −(−1)r′m[(r + m)λ(r′ + m′)]M .
To check the Jacobi identity, we compute
[(r + m)λ[(r′ + m′)µ(r′′ + m′′)]M ]M
= [(r + m)λ([r′
= [rλ[r′
µr′′] + ρ(r′)µ(m′′) − (−1)r′′m′ρ(r′′)−∂−µ(m′)]M
µr′′]] + ρ(r)λ(ρ(r′)µ(m′′)) − (−1)r′′m′ρ(r)λ(ρ(r′′)−∂−µ(m′))
−(−1)(r′′+r′)mρ([r′
µr′′])−∂−λ(m),
(−1)rr′[(r′ + m′)µ[(r + m)λ(r′′ + m′′)]M ]M
= (−1)rr′[r′
µ[rλr′′]] + (−1)rr′ρ(r′)µ(ρ(r)λ(m′′))
(2.21)
−(−1)rr′+r′′mρ(r′)µ(ρ(r′′)−∂−λ(m)) − (−1)r′′m′ρ([rλr′′])−∂−µ(m′) (2.22)
and
[[(r + m)λ(r′ + m′)]M (λ+µ)(r′′ + m′′)]M
= [([rλr′] + ρ(r)λ(m′) − (−1)r′mρ(r′)−∂−λ(m))λ+µ(r′′ + m′′)]M
= [[rλr′]λ+µr′′] + ρ([rλr′])λ+µ(m′′) − (−1)r′′(r+m′)ρ(r′′)−∂−λ−µ(ρ(r)λm′)
+(−1)r′′(r′+m)+r′mρ(r′′)−∂−λ−µ(ρ(r′)−∂−λm).
(2.23)
By Eqs.(2.21)–(2.23), we only need to show that
ρ(r)λ(ρ(r′)−∂−µ′ m′′) − (−1)rr′ρ(r′)−∂−µ′ −λ(ρ(r)−∂−µ′ m′′) = ρ([rλr′])−∂−µ′ (m′′). (2.24)
Since (ρ, M) is a representation of R,
ρ(r)λ(ρ(r′)µm′′) − (−1)rr′ρ(r′)µ(ρ(r)λm′′) = ρ([rλr′])λ+µ(m′′).
(2.25)
10
Replacing µ by −λ − µ
′ − ∂ in Eq.(2.25) and using (C1)λ, we obtain
ρ(r)λ(ρ(r′)−∂−µ′m′′) − (−1)rr′ρ(r′)−∂−µ′−λ(ρ(r)λm′′) = ρ([rλr′])−∂−µ′(m′′).
(2.26)
By (C1)λ again, Eq.(2.26) is equivalent to Eq.(2.24). This implies
[(r + m)λ[(r′ + m′)µ(r′′ + m′′)]M ]M
= (−1)rr′[(r′ + m′)µ[(r + m)λ(r′′ + m′′)]M ]M + [[(r + m)λ(r′ + m′)]M (λ+µ)(r′′ + m′′)]M .
Then R ⊕ M is a Lie conformal superalgebra.
Let g be a complex Lie superalgebra and M a finite dimensional g-module. Then we
have the Current Lie conformal superalgebra Cur(g ⋉ M) (see Example 2.6). By Example
2.10, C[∂] ⊗ M is a conformal module of Curg. Thus, by the above lemma we have the
Lie conformal superalgebra Curg ⋉ (C[∂] ⊗ M).
Corollary 2.20. Cur(g ⋉ M) is a proper Lie conformal subalgebra of Curg ⋉ (C[∂] ⊗ M).
Proof. Let ρ : g → C[λ] ⊗ EndC(M) be the corresponding representation. For any
f (∂) ⊗ (r + m), g(∂) ⊗ (r′ + m′) ∈ Cur(g ⋉ M), where r, r′ ∈ g, m, m′ ∈ M. By the
λ-bracket on Cur(g ⋉ M)(Example 2.6) it follows from
[(f (∂) ⊗ (r + m))λ(g(∂) ⊗ (r′ + m′))]
= f (−λ)g(∂ + λ) ⊗ [r + m, r′ + m′]
= f (−λ)g(∂ + λ) ⊗ ([r, r′] + ρ(r)(m′) − (−1)r′mρ(r′)(m)).
(2.27)
By the module structure on C[∂] ⊗ M(see Example 2.10) it follows from
ρ(g(∂) ⊗ r′)−∂−λ(f (∂) ⊗ m) = g(∂ + λ)f (−λ) ⊗ ρ(r′)(m) = f (−λ)g(∂ + λ) ⊗ ρ(r′)(m).
So, by the λ-bracket on Curg ⋉ (C[∂] ⊗ M), we have
[(f (∂) ⊗ (r + m))λ(g(∂) ⊗ (r′ + m′))]
= [(f (∂) ⊗ r + f (∂) ⊗ m))λ(g(∂) ⊗ r′ + g(∂) ⊗ m′)]
= [(f (∂) ⊗ r)λ(g(∂) ⊗ r′)] + ρ(f (∂) ⊗ r)λ(g(∂) ⊗ m′))
−(−1)r′mρ(g(∂) ⊗ r′)−∂−λ(f (∂) ⊗ m))
= f (−λ)g(∂ + λ) ⊗ [r, r′] + f (−λ)g(∂ + λ) ⊗ ρ(r)(m′)
−(−1)r′mf (−λ)g(∂ + λ) ⊗ ρ(r′)(m)
= f (−λ)g(∂ + λ) ⊗ ([r, r′] + ρ(r)(m′) − (−1)r′mρ(r′)(m)).
(2.28)
By Eqs.(2.27) and (2.28), the λ-bracket on Cur(g ⋉ M) is induced by that on Curg ⋉
(C[∂] ⊗ M).
Lemma 2.21. Let R be a Lie conformal superalgebra with λ-bracket [−λ−], M a C[∂]-
module, ρ ∈ Chom(R, Cend(M)), and f ∈ Chom(M, R). Then for any r ∈ R and m ∈ M,
the following two equations are equivalent:
11
(1) fλ(ρ(r)µ(m)) = (−1)rf [rµ(fλ(m))].
(2) fλ(ρ(r)−∂−µ(m)) = −(−1)rm[(fλ(m))λ+µr].
Proof.
If fλ(ρ(r)µ(m)) = (−1)rf [rµ(fλ(m))], replacing µ by −λ − µ
′ − ∂ and using
(C1)λ, we obtain
fλ(ρ(r)−∂−µ′(m)) = −(−1)rm[(fλ(m))λ+µ′r],
that is (2) holds. The reverse conclusion follows similarly.
Lemma 2.22. Let R be a Lie conformal superalgebra with λ-bracket [−λ−], M a C[∂]-
module, ρ ∈ Chom(R, Cend(M)), f ∈ Cend(M)θ and g ∈ Cend(R)θ. Then for any r ∈ R
and m ∈ M, the following two equations are equivalent:
(1) fλ(ρ(r)µ(m)) = ρ(gλ(r))λ+µ(m) + (−1)θrρ(r)µ(fλ(m)).
(2) fλ(ρ(r)−∂−µ(m)) = ρ(gλ(r))−∂−µ(m) + (−1)θrρ(r)−∂−λ−µ(fλ(m)).
Proof.
Its proof is similar to Lemma 2.21.
Let R be a Lie conformal superalgebra, and M a conformal module of R. Hereafter
we denote C-linear maps from the C[∂]-module R ⊕ M to C[λ] ⊗ (R ⊕ M) in the following
form of matrices:
(∗)
dλ = (cid:18) d11 d12
d22 (cid:19)λ
d21
: R ⊕ M → C[λ] ⊗ (R ⊕ M), where
d11λ : R → C[λ] ⊗ R, d12λ : M → C[λ] ⊗ R,
d21λ : R → C[λ] ⊗ M, d22λ : M → C[λ] ⊗ M.
Lemma 2.23. Let R be a Lie conformal superalgebra, and M a conformal module of R.
Then d is given by (∗) is a conformal linear map of degree θ if and only if all d11, d12, d21
and d22 are conformal linear maps of degree θ.
Proof.
It is straightforward by Definition 2.7.
Theorem 2.24. Let R be a Lie conformal superalgebra with λ-bracket[−λ−], M a con-
formal module of R, and ρ : R → C[λ] ⊗ Cend(M) : r → ρ(r)λ the corresponding
representation. Then a conformal linear map dλ given by (∗) is a conformal derivation of
degree θ of R ⋉ M if and only if dλ satisfies the following conditions:
(1) d11 is a conformal derivation of degree θ of R.
(2) For any r ∈ R and m, m′ ∈ M, d12λ(ρ(r)µm) = (−1)θr[rµ(d12λm)],
ρ(d12λm)λ+µ(m′) = (−1)mm′ρ(d12λm′)−∂−µ(m).
12
(3) d21 is a conformal derivation of degree θ from R to M.
(4) For any r ∈ R and m ∈ M,
d22λ(ρ(r)µm) = ρ(d11λr)λ+µ(m) + (−1)θrρ(r)µ(d22λm).
Proof. For any r + m, r′ + m′ ∈ R ⋉ M, by Eq.(2.20), we have
dλ([(r + m)µ(r′ + m′)])
= d11λ([rµr′]) + d12λ(ρ(r)µm′) − (−1)r′md12λ(ρ(r′)−∂−µm)
+d21λ([rµr′]) + d22λ(ρ(r)µm′) − (−1)r′md22λ(ρ(r′)−∂−µm),
[(dλ(r + m))λ+µ(r′ + m′)]
= [(d11λ(r) + d12λ(m) + d21λ(r) + d22λ(m))λ+µ(r′ + m′)]
= [(d11λ(r))λ+µr′] + [(d12λ(m))λ+µr′] + ρ(d11λ(r))λ+µ(m′) + ρ(d12λ(m))λ+µ(m′)
−(−1)r′(r+θ)ρ(r′)−∂−λ−µ(d21λ(r)) − (−1)r′(m+θ)ρ(r′)−∂−λ−µ(d22λ(m))
and
(−1)θr[(r + m)µ(dλ(r′ + m′))]
= (−1)θr[(r + m)µ(d11λ(r′) + d12λ(m′) + d21λ(r′) + d22λ(m′))]
= (−1)θr[rµ(d11λ(r′))] + (−1)θr[rµ(d12λ(m′))] + (−1)θrρ(r)µ(d21λ(r′))
+(−1)θrρ(r)µ(d22λ(m′)) − (−1)rr′ρ(d11λ(r′))−∂−µ(m)
−(−1)mm′ρ(d12λ(m′))−∂−µ(m).
Suppose that dλ is a conformal derivation of degree θ of R ⋉ M. By Eq.(2.9) and
taking m = m′ = 0 in the above identities, we get
d11λ[rµr′] = [(d11λ(r))λ+µr′] + (−1)θr[rµ(d11λ(r′))],
(2.29)
d21λ[rµr′] = (−1)θrρ(r)µ(d21λ(r′)) − (−1)r′(r+θ)ρ(r′)−∂−λ−µ(d21λ(r)); (2.30)
Taking r = r′ = 0,
ρ(d12λ(m))λ+µ(m′) − (−1)mm′ρ(d12λ(m′))−∂−µ(m) = 0;
(2.31)
Taking m = 0, r′ = 0,
d12λ(ρ(r)µ(m′)) = (−1)θr[rµ(d12λ(m′))],
d22λ(ρ(r)µ(m′)) = ρ(d11λ(r))λ+µ(m′) + (−1)θrρ(r)µ(d22λ(m′));
(2.32)
(2.33)
Taking m′ = 0, r = 0,
−(−1)r′md12λ(ρ(r′)−∂−µm) = [(d12λ(m))λ+µr′],
(2.34)
d22λ(ρ(r′)−∂−µm) = ρ(d11λ(r′))−∂−µ(m) + (−1)θr′ρ(r′)−∂−λ−µ(d22λ(m)). (2.35)
By Lemmas 2.21 and 2.22, Eq.(2.32) is equivalent to Eq.(2.34), and Eq.(2.33) is equivalent
to Eq.(2.35). By Definition 2.14, Lemmas 2.21, 2.22 and 2.23, dλ is a conformal derivation
of degree θ of R ⋉ M if and only if Eqs.(2.29)-(2.35) are satisfied. Eqs.(2.29)-(2.35) are
equivalent to conditions (1) − (4) of Theorem 2.24.
13
3 Deformations of Lie conformal superalgebras
In the following we aim to develop cohomology theory of Lie conformal superalgebras. To
do this, we need the following concept.
Definition 3.1. An n-cochain (n ∈ Z+) of a Lie conformal superalgebra R with coeffi-
cients in a module M is a C-linear map of degree θ
γ : Rn → M[λ1, · · · , λn],
(a1, · · · , an)
7→ γλ1,··· ,λn(a1, · · · , an),
where M[λ1, · · · , λn] denotes the space of polynomials with coefficients in M, satisfying
the following conditions:
Conformal antilinearity:
γλ1,··· ,λn(a1, · · · , ∂ai, · · · , an) = −λiγλ1,··· ,λn(a1, · · · , ai, · · · , an);
Skew-symmetry:
γλ1,··· ,λi,··· ,λj ,··· ,λn(a1, · · · , ai, · · · , aj, · · · , an)
= −(−1)aiajγλ1,··· ,λj ,··· ,λi,··· ,λn(a1, · · · , aj, · · · , ai, · · · , an).
Let R⊗0 = C as usual, so that a 0-cochain γ is an element of M. Define a differential
d of a cochain γ by
(dγ)λ1,··· ,λn+1(a1, · · · , an+1)
n+1
=
+
Pi=1
(−1)i+1(−1)(γ+a1+···+ai−1)aiρ(ai)λiγλ1,··· , λi,··· ,λn+1(a1, · · · , ai, · · · , an+1)
P1≤i<j
(−1)i+j(−1)(a1+···+ai−1)ai+(a1+···+aj−1)aj+aiaj
n+1
γλi+λj ,λ1,··· , λi,··· ,λj ,··· ,λn+1
([aiλiaj], a1, · · · , ai, · · · , aj, · · · , an+1),
where ρ is the corresponding representation of M, and γ is extended linearly over the
polynomials in λi. In particular, if γ is a 0-cochain, then (dγ)λa = aλγ.
Remark 3.2. Conformal antilinearity implies the following relation for an n-cochain γ:
γλ+µ,λ1,···([aλb], a1, · · · ) = γλ+µ,λ1,···([a−∂−µb], a1, · · · ).
Proposition 3.3. dγ is a cochain and d2 = 0.
Proof. Let γ be an n-cochain. As discussed in the proof of [1, Lemma 2.1], dγ
satisfies conformal antilinearity and skew-symmetry. Thus dγ is an (n + 1)-cochain.
A straightforward computation shows that
(d2γ)λ1,··· ,λn+2(a1, · · · , an+2)
14
=
=
n+2
+
n+2
n+2
(−1)i+1+γai+Aiρ(ai)λi(dγ)λ1,··· ,λi,··· ,λn+2(a1, · · · , ai, · · · , an+2)
Xi=1
X1≤i<j
X1≤j<i
(−1)i+j+γ(ai+aj)+Ai+Aj ρ(ai)λi(ρ(aj)λj γλ1,··· ,λj,i,··· ,λn+2(a1, · · · , aj,i, · · · , an+2))
(−1)i+j+Ai+Aj+aiaj(dγ)λi+λj ,λ1,··· ,λi,j ,··· ,λn+2([aiλiaj], a1, · · · , ai,j, · · · , an+2)
+
n+2
X1≤i<j
(−1)i+j+1+γ(ai+aj)+Ai+(Aj−ai)
ρ(ai)λi(ρ(aj)λj γλ1,··· ,λi,j ,··· ,λn+2(a1, · · · , ai,j, · · · , an+2))
+
n+2
X1≤j<k<i
(−1)i+j+k+1+γai+Ai+Aj +Ak+ajak
ρ(ai)λiγλj+λk,λ1,··· ,λj,k,i,··· ,λn+2([aj λj
ak], a1, · · · , aj,k,i, · · · , an+2)
+
n+2
X1≤j<i<k
(−1)i+j+k+γai+Ai+Aj +(Ak−ai)+aj ak
ρ(ai)λiγλj+λk,λ1,··· ,λj,i,k,··· ,λn+2([aj λj
ak], a1, · · · , aj,i,k, · · · , an+2)
+
n+2
X1≤i<j<k
(−1)i+j+k+1+γai+Ai+(Aj −ai)+(Ak−ai)+aj ak
ρ(ai)λiγλj+λk,λ1,··· ,λi,j,k,··· ,λn+2([aj λj
ak], a1, · · · , ai,j,k, · · · , an+2)
+
n+2
X1≤k<i<j
(−1)i+j+k+Ai+Aj+Ak+aiaj +(γ+ai+aj)ak
ρ(ak)λkγλi+λj ,λ1,··· ,λk,i,j ,··· ,λn+2([aiλiaj], a1, · · · , ak,i,j, · · · , an+2)
+
n+2
X1≤i<k<j
(−1)i+j+k+1+Ai+Aj+Ak+aiaj +(γ+aj)ak
ρ(ak)λkγλi+λj ,λ1,··· ,λi,k,j ,··· ,λn+2
([aiλiaj], a1, · · · , ai,k,j, · · · , an+2)
+
n+2
X1≤i<j<k
(−1)i+j+k+Ai+Aj+Ak+aiaj +γak
ρ(ak)λkγλi+λj ,λ1,··· ,λi,j,k,··· ,λn+2
([aiλiaj], a1, · · · , ai,j,k, · · · , an+2)
+
n+2
X1≤i<j
(−1)i+j+Ai+Aj+aiaj+γ(ai+aj)
ρ([aiλiaj])λi+λj γλ1,··· ,λj ,··· ,λi,··· ,λn+2(a1, · · · , aj, · · · , ai, · · · , an+2)
15
(3.1)
(3.2)
(3.3)
(3.4)
(3.5)
(3.6)
(3.7)
(3.8)
(3.9)
n+2
+
Xdistinct i,j,k,l,i<j,k<l
(−1)i+j+k+lsign{i, j, k, l}(−1)Ai+Aj +aiaj+(ai+aj)(ak+al)
(−1)(Ak−ai−aj)+(Al−ai−aj)+akal
γλk+λl,λi+λj ,λ1,··· ,λi,j,k,l,··· ,λn+2([ak λkal], [aiλiaj], a1, · · · , ai,j,k,l, · · · , an+2)
(3.10)
+
n+2
Xi,j,k=1,i<j,k6=i,j
(−1)i+j+k+1sign{i, j, k}(−1)Ai+Aj+aiaj+(Ak−ai−aj)
γλi+λj +λk,λ1,··· ,λi,j,k,··· ,λn+2([[aiλiaj]λi+λj ak], a1, · · · , ai,j,k, · · · , an+2),
(3.11)
where Ai = (a1 + · · · + ai−1)ai, Aj − ai = (a1 + · · · + ai + · · · + aj−1)aj and
sign{i1, · · · , ip} is the sign of the permutation putting the indices in increasing order and
ai,j,··· means that ai, aj, · · · are omitted.
It is obvious that Eqs.(3.3) and (3.8) summations cancel each other. The same is
true for Eqs.(3.4) and (3.7), (3.5) and (3.6). The Jacobi identity implies Eq.(3.11) = 0,
whereas skew-symmetry of γ gives Eq.(3.10) = 0. As M is an R-module,
−ρ(ai)λi(ρ(aj)λj m) + (−1)aiajρ(aj)λj (ρ(ai)λim) + ρ([aiλiaj])λi+λj (m) = 0.
Eqs.(3.1), (3.2) and (3.9) summations cancel. This proves d2γ = 0.
Thus the cochains of a Lie conformal superalgebra R with coefficients in a module
M form a comlex, which is denoted by
C• = C•(R, M) = Mn∈Z+
Cn(R, M).
Where Cn(R, M) = Cn(R, M)¯0 ⊕Cn(R, M)¯1, C n(R, M)θ is the set of n-cochain of degree θ
(θ ∈ Z2). The cohomology H•(R, M) of a Lie conformal superalgebra R with coefficients
in a module M is the cohomology of complex C•.
Let R be a Lie conformal superalgebra. Define
ρ(a)λb = [aλb], ∀ a, b ∈ R.
(3.12)
Proposition 3.4. (ρ, R) is a representation with the λ-action given in Eq.(3.12).
Proof.
It only consists of checking the axioms from Definition 2.9.
Let γ ∈ C n(R, R). Define an operator d : C n(R, R) → C n+1(R, R) by
(dγ)λ1,··· ,λn+1(a1, · · · , an+1)
n+1
=
Xi=1
(−1)i+1(−1)(γ+a1+···+ai−1)ai[aiλiγλ1,··· , λi,··· ,λn+1
(a1, · · · , ai, · · · , an+1)]
16
+
n+1
X1≤i<j
(−1)i+j(−1)(a1+···+ai−1)ai+(a1+···+aj−1)aj +aiaj
γλi+λj ,λ1,··· , λi,··· ,λj ,··· ,λn+1([aiλiaj], a1, · · · , ai, · · · , aj, · · · , an+1).
Obviously, the operator d is induced from the differential d. Thus d preserves the space
of cochains and satisfies d2 = 0. In the following the complex C •(R, R) is assumed to be
associated with the differential d.
For ψ ∈ C 2(R, R)¯0, we consider a t-parameterized family of bilinear operations on R
[aλb]t = [aλb] + tψλ,−∂−λ(a, b), ∀ a, b ∈ R.
(3.13)
If [·λ·]t endows (R, [·λ·]t, α) with a Lie conformal superalgebra structure, we say that ψ
generates a deformation of the Lie conformal superalgebra R. It is easy to see that [·λ·]t
satisfies (C1)λ and (C2)λ. If it is true for (C3)λ, expanding the Jacobi identity for [·λ·]t
gives
[aλ[bµc]] + t([aλ(ψµ,−∂−µ(b, c))] + ψλ,−∂−λ(a, [bµc]))
+t2ψλ,−∂−λ(a, ψµ,−∂−µ(b, c))
= (−1)ab[bµ[aλc]] + (−1)abt([bµ(ψλ,−∂−λ(a, c))] + ψµ,−∂−µ(b, [aλc]))
+(−1)abt2ψµ,−∂−µ(b, ψλ,−∂−λ(a, c))
+[[aλb]λ+µc] + t([(ψλ,−∂−λ(a, b))λ+µc] + ψλ+µ,−∂−λ−µ([aλb], c))
+t2ψλ+µ,−∂−λ−µ(ψλ,−∂−λ(a, b), c).
This is equivalent to the following relations
[aλ(ψµ,−∂−µ(b, c))] + ψλ,−∂−λ(a, [bµc])
= (−1)ab[bµ(ψλ,−∂−λ(a, c))] + (−1)abψµ,−∂−µ(b, [aλc])
+[(ψλ,−∂−λ(a, b))λ+µc] + ψλ+µ,−∂−λ−µ([aλb], c)
(3.14)
and
ψλ,−∂−λ(a, ψµ,−∂−µ(b, c))
= (−1)abψµ,−∂−µ(b, ψλ,−∂−λ(a, c)) + ψλ+µ,−∂−λ−µ(ψλ,−∂−λ(a, b), c).
(3.15)
By conformal antilinearity of ψ, we have
[(ψλ,−∂−λ(a, b))λ+µc] = [ψλ,µ(a, b)λ+µc].
(3.16)
On the other hand, let ψ be a cocycle, i.e., dψ = 0. Explicitly,
0 = (dψ)λ,µ,γ(a, b, c)
= [aλ(ψµ,γ(b, c))] − (−1)ab[bµ(ψλ,γ(a, c))] + (−1)(a+b)c[cγ(ψλ,µ(a, b))]
−ψλ+µ,γ([aλb], c) + (−1)bcψλ+γ,µ([aλc], b) − (−1)ab+acψµ+γ,λ([bµc], a)
17
= [aλ(ψµ,γ(b, c))] − (−1)ab[bµ(ψλ,γ(a, c))] − [(ψλ,µ(a, b))−∂−γc]
+ψλ,µ+γ(a, [bµc]) − (−1)abψµ,λ+γ(b, [aλc]) − ψλ+µ,γ([aλb], c).
(3.17)
By (C1)λ, Eq.(3.16) and replacing γ by −λ − µ − ∂ in Eq.(3.17), we obtain
0 = [aλ(ψµ,−∂−µ(b, c))] − (−1)ab[bµ(ψλ,−∂−λ(a, c))] − [(ψλ,µ(a, b))λ+µc]
+ψλ,−∂−λ(a, [bµc]) − (−1)abψµ,−∂−µ(b, [aλc]) − ψλ+µ,−∂−λ−µ([aλb], c),
which is exactly Eq.(3.14). Thus, when ψ is a 2-cocycle satisfying Eq.(3.15), (R, [·λ·]t, α)
forms a Lie conformal superalgebra. In this case, ψ generates a deformation of the Lie
conformal superalgebra R.
A deformation is said to be trivial if there is a linear operator f ∈ C 1(R, R)¯0 such
that for Ttλ = id + tfλ, there holds
Tt−∂([aλb]t) = [(Ttλ(a))λTt−∂(b)], ∀ a, b ∈ R.
Definition 3.5. A linear operator f ∈ C 1(R, R)¯0 is called a Nijienhuis operator if
[(fλ(a))λ(fµ(b))] = fλ+µ([aλb]N ), ∀ a, b ∈ R,
where the bracket [·λ·]N is defined by
(3.18)
(3.19)
[aλb]N = [(fλ(a))λb] + [aλ(f−∂(b))] − f−∂([aλb]) ∀ a, b ∈ R.
(3.20)
Remark 3.6. In particular, by (C1)λ and setting µ = −∂ − λ in Eq.(3.19), we obtain
[(fλ(a))λf−∂(b)] = f−∂([aλb]N ), ∀ a, b ∈ R.
(3.21)
Theorem 3.7. Let R be a Lie conformal superalgebra, and f ∈ C 1(R, R)¯0 a Nijienhuis
operator. Then a deformation of R can be obtained by putting
ψλ,−∂−λ(a, b) := (df )λ,−∂−λ(a, b) = [aλb]N , ∀ a, b ∈ R.
(3.22)
Furthermore, this deformation is trivial.
Proof. Since ψ = df , dψ = 0 is valid. To see that ψ generates a deformation of R,
we need to check Eq.(3.15) for ψ. By Eqs.(3.20) and (3.22), we compute and get
ψλ,−∂−λ(a, ψµ,−∂−µ(b, c)) = [aλ[bµc]N ]N
= [(fλ(a))λ([bµc]N )] + [aλ(f−∂([bµc]N ))] − f−∂([aλ[bµc]N ])
= [(fλ(a))λ[(fµ(b))µc]] + [(fλ(a))λ[bµ(f−∂(c))]] − [(fλ(a))λ(f−∂([bµc]))] + [aλ(f−∂([bµc]N ))]
−f−∂([aλ[(fµ(b))µc]]) − f−∂([aλ[bµ(f−∂(c))]]) + f−∂([aλ(f−∂([bµc]))])
= [(fλ(a))λ[(fµ(b))µc]]
+ [(fλ(a))λ[bµ(f−∂(c))]]
−[(fλ(a))λ(f−∂([bµc]))] + [aλ[(fµ(b))µ(f−∂(c))]]
−f−∂([aλ[(fµ(b))µc]])
−f−∂([aλ[bµ(f−∂(c))]])
+f−∂([aλ(f−∂([bµc]))]).
(3)
{z
}
(1)
{z
{z
(4)
}
}
(2)
{z
{z
(5)
}
}
18
In the same way, we have
(−1)abψµ,−∂−µ(b, ψλ,−∂−λ(a, c))
= (−1)ab[(fµ(b))µ[(fλ(a))λc]]
+ (−1)ab[(fµ(b))µ[aλ(f−∂(c))]]
−(−1)ab[(fµ(b))µ(f−∂([aλc]))] + (−1)ab[bµ[(fλ(a))λ(f−∂(c))]]
−(−1)abf−∂([bµ[(fλ(a))λc]])
−(−1)abf−∂([bµ[aλ(f−∂(c))]])
}
}
(3)′
{z
(2)′
{z
}
}
(5)′
{z
}
+(−1)abf−∂([bµ(f−∂([aλc]))])
and
ψλ+µ,−∂−λ−µ(ψλ,−∂−λ(a, b), c)
= [(fλ+µ([aλb]N ))λ+µc] + [([aλb]N )λ+µ(f−∂(c))] − f−∂([([aλb]N )λ+µ)c])
= [[(fλ(a))λ(fµ(b))]λ+µc]
(1)′′
{z
(2)′′
{z
+ [[(fλ(a))λb]λ+µ(f−∂(c))]
+ [[aλ(f−∂(b))]λ+µ(f−∂(c))]
−[(f−∂([aλb]))λ+µ(f−∂(c))]
−f−∂([(fλ(a))λb]λ+µc])
−f−∂([[aλ(f−∂(b))]λ+µ
c]) + f−∂([(f−∂([aλb]))λ+µc]).
}
(3)′′
{z
}
(1)′
{z
(6)′
{z
}
(6)′′
{z
}
(4)′′
{z
}
Since f is a Nijienhuis operator and by Eq.(3.21), we get
−[(fλ(a))λ(f−∂([bµc]))] + f−∂([aλ(f−∂([bµc]))])
= −f−∂([(fλ(a))λ[bµc]])
+ f 2
−∂([aλ[bµc]])
,
(6)
{z
}
(7)
{z
}
−(−1)ab[(fµ(b))µ(f−∂([aλc]))] + (−1)abf−∂([bµ(f−∂([aλc]))])
= −(−1)abf−∂([(fµ(b))µ[aλc]])
+ (−1)abf 2
−∂([bµ[aλc]])
.
(4)′
{z
}
(7)′
{z
}
By (C1)λ and Eq.(3.21),
−[(f−∂([aλb]))λ+µ(f−∂(c))] + f−∂([(f−∂([aλb]))λ+µc])
= −[(fλ+µ([aλb]))λ+µ(f−∂(c))] + f−∂([(fλ+µ([aλb]))λ+µc])
= −f−∂([[aλb]λ+µ(f−∂(c))])
−∂([[aλb]λ+µc])
+ f 2
.
(5)′′
{z
}
(7)′′
{z
}
Note that according to the Jacobi identity and (C1)λ for a, b, c ∈ R,
[aλ[(fµ(b))µ(f−∂(c))]] = [[aλfµ(b)]λ+µ(f−∂(c))] + (−1)ab[(fµ(b))µ[aλ(f−∂(c))]]
19
is equivalent to
[aλ[(fµ(b))µ(f−∂(c))]] = [[aλf−∂(b)]λ+µ(f−∂(c))] + (−1)ab[(fµ(b))µ[aλ(f−∂(c))]].
Thus (i) + (i)′ + (i)′′ = 0, for i = 1, · · · , 7. This proves that ψ generates a deformation of
the Lie conformal superalgebra R.
Let Ttλ = id + tfλ. By Eqs.(3.13) and (3.22),
Tt−∂([aλb]t) = (id + tf−∂)([aλb] + tψλ,−∂−λ(a, b))
= (id + tf−∂)([aλb] + t[aλb]N )
= [aλb] + t([aλb]N + f−∂([aλb])) + t2f−∂([aλb]N ).
(3.23)
Furthermore,
[(Ttλ(a))λTt−∂(b)] = [(a + tfλ(a))λ(b + tf−∂(b))]
= [aλb] + t([(fλ(a))λb] + [aλ(f−∂(b))]) + t2[(fλ(a))λf−∂(b)].(3.24)
Combining Eqs.(3.23) with (3.24) gives Tt−∂([aλb]t) = [(Ttλ(a))λTt−∂(b)]. Therefore the
deformation is trivial.
4 Generalized derivations of Lie conformal superal-
gebras
Let R be a Lie conformal superalgebra. Define Der(R)θ is the set of all derivations of
degree θ, then it is obvious that Der(R) = Der(R)¯0 ⊕Der(R)¯1 is a subalgebra of Cend(R).
Definition 4.1. An element f in Cend(R)θ is called
• a generalized derivation of degree θ of R, if there exist f
′
, f
′′ ∈ Cend(R)θ such that
[(fλ(a))λ+µb] + (−1)θa[aµ(f
′
λ(b))] = f
λ ([aµb]), ∀ a, b ∈ R.
• an quasiderivation of degree θ of R, if there is f
′ ∈ Cend(R)θ such that
[(fλ(a))λ+µb] + (−1)θa[aµ(fλ(b))] = f
λ([aµb]), ∀ a, b ∈ R.
• an centroid of degree θ of R, if it satisfies
[(fλ(a))λ+µb] = (−1)θa[aµ(fλ(b))] = fλ([aµb]), ∀ a, b ∈ R.
(4.3)
• an quasicentroid of degree θ of R, if it satisfies
[(fλ(a))λ+µb] = (−1)θa[aµ(fλ(b))], ∀ a, b ∈ R.
(4.4)
20
′′
′
(4.1)
(4.2)
• an central derivation of degree θ of R, if it satisfies
[(fλ(a))λ+µb] = fλ([aµb]) = 0, ∀ a, b ∈ R.
(4.5)
Denote by GDer(R)θ, QDer(R)θ, C(R)θ, QC(R)θ and ZDer(R)θ the sets of all gen-
eralized derivations, quasiderivations, centroids, quasicentroids and central derivations of
degree θ of R. It is easy to see that
ZDer(R) ⊆ Der(R) ⊆ QDer(R) ⊆ GDer(R) ⊆ Cend(R), C(R) ⊆ QC(R) ⊆ GDer(R).
(4.6)
Proposition 4.2. Let R be a Lie conformal superalgebra. Then
(1) GDer(R), QDer(R) and C(R) are subalgebras of Cend(R).
(2) ZDer(R) is an ideal of Der(R).
Proof.
(1) We only prove that GDer(R) is a subalgebra of Cend(R). The proof for
the other two cases is exactly analogous.
For f ∈ GDer(R)σ, g ∈ GDer(R)ϑ, a, b ∈ R, there exist f
′′ ∈ GDer(R)ϑ ) such that Eq.(4.1) holds for f (resp. g). We only need to show
, f
′′ ∈ GDer(R)σ (resp.
′
′
g
, g
′′
[f
′′
λ g
]θ([aµb]) = [([fλg]θ(a))µ+θb] + (−1)(σ+ϑ)a[aµ([f
′
λg
′
]θ(b))].
(4.7)
By Eq.(2.8), we have
[([fλg]θ(a))µ+θb] = [(fλ(gθ−λ(a)))µ+θb] − (−1)σϑ[(gθ−λ(fλ(a)))µ+θb].
(4.8)
By Eq.(4.1), we obtain
= f
= f
= g
= g
′
′′
′′
′′
′′
′′
λ ([aµ(g
θ−λ(b))])
θ−λ([aµb])) − (−1)ϑaf
[(fλ(gθ−λ(a)))µ+θb]
λ ([(gθ−λ(a))µ+θ−λb]) − (−1)σ(ϑ+a)[(gθ−λ(a))µ+θ−λ(f
λ (g
−(−1)σ(ϑ+a)g
[(gθ−λ(fλ(a)))µ+θb]
θ−λ([(fλ(a))λ+µb]) − (−1)ϑ(σ+a)[fλ(a)λ+µ(g
θ−λ(f
−(−1)ϑ(σ+a)f
λ(b))]) + (−1)σ(ϑ+a)+ϑa[aµ(g
θ−λ(b))]) + (−1)ϑ(σ+a)+σa[aµ(f
λ ([aµb])) − (−1)σag
θ−λ([aµ(f
λ(b))])
θ−λ([aµ(f
θ−λ(b))]
λ ([aµ(g
′′
′′
′′
′′
′′
′
′
′
′
′
λ(b))]
′
′
θ−λ(f
λ(b)))],
(4.9)
′
λ(g
′
θ−λ(b))).
(4.10)
Substituting Eqs.(4.9) and (4.10) into Eq.(4.8) gives Eq.(4.7). Hence [fλg] ∈ GDer(R)[λ],
and GDer(R) is a subalgebra of Cend(R).
(2) For f ∈ ZDer(R), g ∈ Der(R), and a, b ∈ R, by Eq.(4.5), we have
[fλg]θ([aµb]) = fλ(gθ−λ([aµb])) − (−1)f ggθ−λ(fλ([aµb])) = fλ(gθ−λ([aµb]))
21
= fλ([(gθ−λ(a))µ+θ−λb] + (−1)ag[aµ(gθ−λ(b))]) = 0,
[[fλg]θ(a)µ+θb] = [(fλ(gθ−λ(a)) − (−1)f ggθ−λ(fλ(a)))µ+θb]
= [−(−1)f g(gθ−λ(fλa)))µ+θb]
= −(−1)f ggθ−λ([fλ(a)λ+µb]) + (−1)ga[fλ(a)λ+µgθ−λ(b)]
= 0.
This shows that [fλg] ∈ ZDer(R)[λ]. Thus ZDer(R) is an ideal of Der(R).
Lemma 4.3. Let R be a Lie conformal superalgebra. Then
(1) [Der(R)λC(R)] ⊆ C(R)[λ],
(2) [QDer(R)λQC(R)] ⊆ QC(R)[λ],
(3) [QC(R)λQC(R)] ⊆ QDer(R)[λ].
Proof.
It is straightforward.
Theorem 4.4. Let R be a Lie conformal superalgebra. Then
GDer(R) = QDer(R) + QC(R).
Proof. For f ∈ GDer(R)θ, there exist f
′
, f
′′ ∈ Cend(R)θ such that
[(fλ(a))λ+µb] + (−1)θa[aµ(f
′
λ(b))] = f
λ ([aµb]), ∀ a, b ∈ R.
′′
By (C2)λ and Eq.(4.11), we get
(−1)θb[b−∂−λ−µ(fλ(a))] + [(f
′
λ(b))−∂−µa] = f
λ ([b−∂−µa]).
′′
By (C1)λ and setting µ = −∂ − λ − µ′ in Eq.(4.12), we obtain
(−1)θb[bµ′(fλ(a))] + [(f
λ(b))λ+µ′a] = f
λ ([bµ′ a]).
Then, changing the place of a, b and replacing µ′ by µ in Eq.(4.13) give
′
′
′′
′′
(−1)θa[aµ(fλ(b))] + [(f
λ(a)))λ+µb] = f
λ ([aµb]).
(4.11)
(4.12)
(4.13)
(4.14)
Combining Eqs.(4.11) and (4.14) gives
[(
[(
′
fλ + f
λ
2
′
fλ − f
λ
2
(a))λ+µb] + (−1)θa[aµ(
(a))λ+µb] − (−1)θa[aµ(
′
fλ + f
λ
2
′
fλ − f
λ
2
(b))] = f
λ ([aµb]),
′′
(b))] = 0.
It follows that fλ+f
2 ∈ QDer(R) and fλ−f
2 ∈ QC(R). Hence
′
λ
′
λ
fλ =
′
fλ + f
λ
2
+
′
fλ − f
λ
2
∈ QDer(R) + QC(R),
proving that GDer(R) ⊆ QDer(R) + QC(R). The reverse inclusion relation follows from
Eq.(4.6) and Lemma 4.3.
22
Theorem 4.5. Let R be a Lie conformal superalgebra and Z(R) the center of R. Then
[C(R)λQC(R)] ⊆ Chom(R, Z(R))[λ]. Moreover, if Z(R) = 0, then [C(R)λQC(R)] = 0.
Proof. For f ∈ C(R), g ∈ QC(R), and a, b ∈ R, by Eqs.(4.3)and (4.4), we have
[([fλg]θ(a))µ+θb] = [(fλ(gθ−λ(a)))µ+θb] − (−1)f g[(gθ−λ(fλ(a)))µ+θb]
= fλ([(gθ−λ(a))µ+θ−λb]) − (−1)f g+g(f +a)[(fλ(a))λ+µ(gθ−λ(b))]
= fλ([(gθ−λ(a))µ+θ−λb]) − (−1)gafλ([aµ(gθ−λ(b))])
= fλ([(gθ−λ(a))µ+θ−λb] − (−1)ga[aµ(gθ−λ(b))])
= 0.
Hence [fλg](a) ∈ Z(R)[λ], and then [fλg] ∈ Chom(R, Z(R))[λ].
[fλg](a) = 0, ∀ a ∈ hg(R). Thus [C(R)λQC(R)] = 0.
If Z(R) = 0, then
Lemma 4.6. Let R be a Lie conformal superalgebra. If f ∈ QC(R)θ ∩ QDer(R)θ and
f, f ′ satisfy Eq.(4.2), then we have
[aλ(fγ([bµc]))] = [aλ[(fγ(b))µ+γc]] = (−1)θb[aλ[bµ(fγ(c))]].
Proof. Let f ′ = 2ϕ. Since f ∈ QC(R)θ ∩ QDer(R)θ and f, f ′ satisfy Eq.(4.2), we
get
[(fγ(a))λ+γb] = (−1)θa[aλ(fγ(b))],
ϕγ([aλb]) =
1
2
f ′
γ([aλb]) =
1
2
([(fγ(a))λ+γb] + (−1)θa[aλ(fγ(b))])
= [(fγ(a))λ+γb] = (−1)θa[aλ(fγ(b))].
(4.15)
By Eq.(4.15), we have
[(ϕγ([aλb]))λ+µ+γc] = [[(fγ(a))λ+γb]λ+µ+γc]
= [(fγ(a))λ+γ[bµc]] − (−1)(θ+a)b[bµ[(fγ(a))λ+γc]]
= [(fγ(a))λ+γ[bµc]] + (−1)bc[[(fγ(a))λ+γc]−∂−µb]
= ϕγ([aλ[bµc]]) + (−1)bc[(ϕγ([aλc])−∂−µb]
= ϕγ([aλ[bµc]]) − (−1)(a+b)c[(ϕγ([c−∂−λa])−∂−µb],
(4.16)
it is equivalent to
(−1)ac[(ϕγ([aλb]))λ+µ+γc] + (−1)bc[(ϕγ([c−∂−λa])−∂−µb] = (−1)acϕγ([aλ[bµc]]).(4.17)
On the one hand, by (C1)λ and setting µ = −∂ − λ − γ − µ′ in Eq.(4.17), we obtain
(−1)ac[(ϕγ([aλb]))−∂−µ′c] + (−1)bc[ϕγ([c−∂−λa])λ+γ+µ′ b] = (−1)acϕγ([aλ[b−∂−µ′ c]]),
it is equivalent to
(−1)ac[ϕγ([aλb])−∂−µ′ c] + (−1)bc[ϕγ([cµ′a])λ+γ+µ′b] = (−1)acϕγ([aλ[b−∂−µ′ c]]).(4.18)
23
By (C1)λ and setting λ = −∂ − λ′ − γ − µ′ in Eq.(4.18), we get
(−1)ac[ϕγ([a−∂−λ′b])−∂−µ′ c] + (−1)bc[ϕγ([cµ′a])−∂−λ′b]
= (−1)acϕγ([a−∂−λ′−µ′[b−∂−µ′c]]).
(4.19)
On the other hand, replacing µ, λ by λ, µ in Eq.(4.17) give
(−1)ac[(ϕγ([aµb]))λ+µ+γc] + (−1)bc[(ϕγ([c−∂−µa])−∂−λb] = (−1)acϕγ([aµ[bλc]]).(4.20)
By (C1)λ and setting λ = −∂ − λ′ − γ − µ in Eq.(4.20), we get
(−1)ac[(ϕγ([aµb]))−∂−λ′c] + (−1)bc[(ϕγ([c−∂−µa])λ′+γ+µb] = (−1)acϕγ([aµ[b−∂−λ′c]]),
it is equivalent to
(−1)ac[(ϕγ([aµb]))−∂−λ′c] + (−1)bc[(ϕγ([cλ′a])λ′+γ+µb] = (−1)acϕγ([aµ[b−∂−λ′c]]).(4.21)
By Eqs.(4.17), (4.19) and (4.21), we have
2((−1)ac[(ϕγ([aλb]))λ+µ+γc] + (−1)bc[(ϕγ([c−∂−λa])−∂−µb]
+(−1)ab[(ϕγ([bµc]))−∂−λa])
= (−1)ac[(ϕγ([aλb]))λ+µ+γc] + (−1)bc[(ϕγ([c−∂−λa])−∂−µb]
+(−1)bc[(ϕγ([c−∂−λa])−∂−µb] + (−1)ab[(ϕγ([bµc]))−∂−λa]
+(−1)ab[(ϕγ([bµc]))−∂−λa] + (−1)ac[(ϕγ([aλb]))λ+µ+γc]
= (−1)acϕγ([aλ[bµc]]) + (−1)bcϕγ([c−∂−λ−µ[a−∂−µb]]) + (−1)abϕγ([bµ[c−∂−λa]])
= (−1)acϕγ([aλ[bµc]]) + (−1)bcϕγ([c−∂−λ−µ[aλb]]) + (−1)abϕγ([bµ[c−∂−λa]])
= 0,
so
(−1)ac[(ϕγ([aλb]))λ+µ+γc] + (−1)bc[(ϕγ([c−∂−λa])−∂−µb]
+(−1)ab[(ϕγ([bµc]))−∂−λa] = 0.
(4.22)
Combining Eqs.(4.17) with (4.22), we get
(−1)acϕγ([aλ[bµc]]) = −(−1)ab[(ϕγ([bµc]))−∂−λa] = (−1)ac+θa[aλ(ϕγ([bµc]))],
that is
ϕγ([aλ[bµc]]) = (−1)θa[aλ(ϕγ([bµc]))].
(4.23)
By Eqs.(4.23) and (4.15), we obtain
[aλ(fγ([bµc]))] = (−1)θaϕγ([aλ[bµc]] = [aλ(ϕγ([bµc]))]
= [aλ[(fγ(b))µ+γc]] = (−1)θb[aλ[bµ(fγ(c))]].
This shows the lemma.
24
Lemma 4.7. Let R be a Lie conformal superalgebra. If f ∈ QC(R)θ ∩ QDer(R)θ and
Z(R) = 0, then f ∈ C(R)θ.
Proof.
It is straightforward by Lemma 4.6.
Theorem 4.8. Let R be a Lie conformal superalgebra.
QDer(R) = C(R).
If Z(R) = 0, then QC(R) ∩
Proof. C(R) ⊆ QC(R) ∩ QDer(R) is obvious. By Lemma 4.7, if Z(R) = 0, we have
QC(R)θ ∩ QDer(R)θ ⊆ C(R)θ. So QC(R) ∩ QDer(R) = C(R).
Proposition 4.9. Let R be a Lie conformal superalgebra. If Z(R) = 0, then QC(R) is a
Lie conformal superalgebra if and only if [QC(R)λQC(R)] = 0.
Proof.
(⇒) Suppose that QC(R) is a Lie conformal superalgebra. For f ∈ QC(R)θ, g ∈
QC(R)ϑ, [fλg] ∈ QC(R)θ+ϑ[λ]. For a, b ∈ R, by Eq.(4.4), we have
[([fλg]θ(a))µ+θb] = (−1)(θ+ϑ)a[aµ([fλg]θ(b))].
(4.24)
By Eqs.(2.8) and (4.4), we obtain
[([fλg]θ(a))µ+θb]
= [(fλ(gθ−λ(a)))µ+θb] − (−1)θϑ[(gθ−λ(fλ(a)))µ+θb]
= (−1)θ(ϑ+a)[(gθ−λ(a))µ+θ−λ(fλ(b))] − (−1)θϑ+ϑ(θ+a)[(fλ(a))λ+µ(gθ−λ(b))]
= (−1)θ(ϑ+a)+ϑa[aµ(gθ−λ(fλ(b)))] − (−1)θϑ+ϑ(θ+a)+θa[aλ(fλ(gθ−λ(b)))]
= −(−1)(ϑ+θ)a[aµ([fλg]θ(b))].
(4.25)
Combining Eqs.(4.24) and (4.25) gives
[([fλg]θ(a))µ+θb] = 0,
and thus [fλg]θ(a) ∈ Z(R)[λ] = 0(Z(R) = 0). Therefore, [fλg] = 0.
(⇐) It is clear.
References
[1] B. Bakalov, V. Kac, A. Voronov, Cohomology of conformal algebras, Comm. Math.
Phys. 200 (1999), no. 3, 561-598.
[2] C. Boyallian, V. Kac, J. Liberati, A. Rudakov, Representations of simple finite Lie
conformal superalgebras of type W and S. J. Math. Phys. 47 (2006), no. 4, 043513,
25 pp.
[3] Z. Chang, Automorphisms and twisted forms of differential lie conformal superalge-
bras. Thesis (Ph.D.)-University of Alberta (Canada). 2013. 117 pp.
25
[4] A. D'Andrea, V. Kac, Structure theory of finite conformal algebras, Selecta Math.
(N.S.) 4 (1998), no. 3, 377-418.
[5] D. Fattori, V. Kac, A. Retakh, Structure theory of finite Lie conformal superalgebras.
Lie theory and its applications in physics V, 27-63, World Sci. Publ., River Edge, NJ,
2004.
[6] G. Fan, Y. Hong, Y. Su, Generalized conformal derivations of Lie conformal algebras,
arXiv:1602.01159 (2016).
[7] J. Guo, Y. Tan, Conformal derivations of semidirect products of Lie conformal al-
gebras and their conformal modules. Proc. Amer. Math. Soc. 142 (2014), no. 5,
1471C1483.
[8] G. Leger, E. Luks, Generalized derivations of Lie algebras, J. Algebra, 228 (2000)
165-203.
[9] Y. Liu, L. Chen, Y. Ma, Hom-Nijienhuis operators and T *-extensions of hom-Lie
superalgebras, Linear Algebra Appl. 439 (2013), no. 7, 2131-2144.
26
|
1004.1125 | 1 | 1004 | 2010-04-07T16:29:21 | Special Freudenthal-Kantor triple systems and Lie algebras with dicyclic symmetry | [
"math.RA"
] | Lie algebras endowed with an action by automorphisms of the dicyclic group of degree 3 are considered. The close connections of these algebras with Lie algebras graded over the nonreduced root system BC1, with J-ternary algebras and with Freudenthal-Kantor triple systems are explored. | math.RA | math |
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
AND LIE ALGEBRAS WITH DICYCLIC SYMMETRY
ALBERTO ELDUQUE⋆ AND SUSUMU OKUBO⋆⋆
Abstract. Lie algebras endowed with an action by automorphisms of
the dicyclic group of degree 3 are considered. The close connections of
these algebras with Lie algebras graded over the nonreduced root sys-
tem BC1, with J-ternary algebras and with Freudenthal-Kantor triple
systems are explored.
1. Introduction
Gradings of Lie algebras by abelian groups constitute a classical subject
(see [Koc09] for a survey of results). In characteristic 0, any grading by a
finitely generated abelian group is given by an action by automorphisms of
an abelian group (the group of characters) on the Lie algebra.
Recently the authors have considered actions of some small nonabelian
groups by automorphisms on a Lie algebra. More precisely, and motivated
by the Principle of Triality that appears related to the classical Lie algebra
of type D4, Lie algebras with an action of the symmetric groups S3 and
S4 by automorphisms were considered in [EO07], [Eld07], [EO08], [EO09].
It turns out that Lie algebras with an action of S4 by automorphisms are
coordinatized by some nonassociative algebras [EO07]. The unital algebras
in this class are precisely the structurable algebras of Allison [All78]. The
Lie algebras with an action of S3 by automorphisms are coordinatized by
a class of algebras which generalize the Malcev algebras [EO09]. These are
some binary-ternary algebras termed generalized Malcev algebras. Jordan
triple systems are instances of these generalized Malcev algebras.
As proved in [EO08, Section 7], the Lie algebras graded by the nonreduced
root system BC1 of type B1 are naturally among the Lie algebras with S4-
symmetry (that is, with an action of S4 by automorphisms).
The Lie algebras graded by the nonreduced root system BC1 but of type
C1 present another type of symmetry given by the dicyclic group
Dic3 = hθ, φ : θ4 = 1 = φ3, φθφ = θi.
Our aim in this paper is to study those Lie algebras with such a symmetry.
It will be shown that Lie algebras with Dic3-symmetry are strongly related
to the BC1-graded Lie algebras of type C1, to the class of J-ternary algebras
introduced in [All76], and to the Freudenthal-Kantor triple systems defined
Date: April 7, 2010.
⋆ Supported by the Spanish Ministerio de Educaci´on y Ciencia and FEDER (MTM
2007-67884-C04-02) and by the Diputaci´on General de Arag´on (Grupo de Investigaci´on
de ´Algebra).
⋆⋆ Supported in part by U.S. Department of Energy Grant No. DE-FG02-91 ER40685.
1
2
ALBERTO ELDUQUE AND SUSUMU OKUBO
in [YO84]. Moreover, these Lie algebras are coordinatized by elements in
a class of nonassociative algebraic systems consisting of binary-ternary al-
gebras endowed with an involutive automorphism which will be introduced
here and named dicyclic ternary algebras.
In Section 2 the Lie algebras graded over the nonreduced root system
BC1 of type C1 will be reviewed, as well as their connections with J-ternary
algebras. These J-ternary algebras are closely connected with a subclass of
the Freudenthal-Kantor triple systems introduced by Yamaguti and Ono in
[YO84]. This subclass will be defined in Section 3 and has its own indepen-
dent interest.
Most of the arguments can be superized and a class of strictly BC1-graded
Lie superalgebras of type C1 will be considered in Section 4, as well as the
associated J-ternary (−1)-algebras and (−1, −1) Freudenthal-Kantor triple
systems.
Section 5 will introduce the dicyclic ternary algebras, which coordina-
tize the Lie algebras with dicyclic symmetry. The dicyclic ternary algebras
attached to J-ternary algebras will be the object of study of Section 6. Fi-
nally, Section 7 will be devoted to show that the 5-graded Lie algebra (or
superalgebra) attached to any (ǫ, δ) Freudenthal-Kantor triple system (not
necessarily special) is always endowed with an action of Dic3 by automor-
phisms. For δ = −ǫ this action reduces to an action of the symmetric group
S3. For ǫ = δ = 1 we get Lie algebras with Dic3-symmetry whose attached
dicyclic ternary algebra is computed.
Throughout the paper all the algebraic systems considered will be defined
over a ground field F whose characteristic will always be assumed to be
6= 2, 3.
2. BC1-graded Lie algebras of type C1
and J -ternary algebras
The purpose of this section is to review the relationship between Lie
algebras graded over the non reduced root system BC1 of type C1, and a
class of nonassociative algebraic systems named J-ternary algebras. This
goes back to [All76] (see [Hei75] and [Kan73] for closely related work). It
must be mentioned that the Lie algebras graded by the root system BC1 (of
any type) are characterized in [BS03], which completes the work in [ABG],
The proof given here is different from the one in [All76].
Let g be a Lie algebra over F containing a subalgebra isomorphic to sl(V ),
for a two-dimensional vector space V over F (that is, a subalgebra isomorphic
to sl2(F)), and such that, as a module for sl(V ), the Lie algebra g is a direct
sum of modules which are either trivial, or isomorphic to V , or to the adjoint
module sl(V ). These are precisely the required conditions for g to be a BC1-
graded Lie algebra of type C1.
Then, as a module for sl(V ), we may write:
where d s the sum of the trivial modules (this is the centralizer in g of the
subalgebra sl(V ) = sl(V ) ⊗ 1, and hence it is a subalgebra), and J and T
g =(cid:0)sl(V ) ⊗ J(cid:1) ⊕(cid:0)V ⊗ T(cid:1) ⊕ d,
(2.1)
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
3
are vector spaces, with J containing a distinguished element 1, such that
sl(V ) ⊗ 1 is the subalgebra sl(V ) we have started with.
In the following lemma we collect some well-known results. Unadorned
tensor products are always considered over the ground field F.
Lemma 2.1. Let V be a two-dimensional vector space.
(i) The skew symmetric map sl(V ) ⊗ sl(V ) → sl(V ), f ⊗ g 7→ [f, g] =
f g − gf , spans the space of sl(V )-invariant maps Homsl(V )(sl(V ) ⊗
sl(V ), sl(V )).
(ii) The symmetric map sl(V ) ⊗ sl(V ) → F, f ⊗ g 7→ trace(f g), spans
the space of sl(V )-invariant maps Homsl(V )(sl(V ) ⊗ sl(V ), F).
(iii) For any f, g, h ∈ sl(V ), one has
[[f, g], h] = 2(cid:0)trace(gh)f − trace(f h)g(cid:1).
(2.2)
(iv) The map sl(V ) ⊗ V → V , f ⊗ u 7→ f (u), spans the space of sl(V )-
invariant maps Homsl(V )(sl(V ) ⊗ V, V ).
(v) The space Homsl(V )(V ⊗ V, F) of sl(V )-invariant maps from V ⊗
V into the trivial module F is one-dimensional and spanned by a
nonzero skew symmetric bilinear map (cid:0)uv(cid:1).
(vi) The symmetric map V ⊗ V → sl(V ), u ⊗ v 7→ γu,v = (cid:0)u.(cid:1)v +
(cid:0)v.(cid:1)u ((..) as in (v)), spans the space of sl(V )-invariant maps
Homsl(V )(V ⊗ V, sl(V )).
(vii) There are no nontrivial sl(V )-invariant maps sl(V )⊗sl(V ) → V and
V ⊗V → V . That is, Homsl(V )(sl(V )⊗sl(V ), V ) = 0 = Homsl(V )(V ⊗
V, V ).
(cid:3)
Note that for any f ∈ sl(V ), f 2 = 1
2 trace(f 2)1, and hence f g + gf =
trace(f g)1 for any f, g ∈ sl(V ). This gives f gf = trace(f g)f − 1
2 trace(f 2)g,
so f gh + hgf = trace(f g)h + trace(gh)f − trace(f h)g, and finally [[f, g], h] =
(f gh+hgf )−(gf h+hf g) = 2(cid:0)trace(gh)f −trace(f h)g(cid:1) for any f, g ∈ sl(V ),
thus getting (2.2).
From now on, we will fix a nonzero skew symmetric bilinear form (..) on
our two-dimensional vector space V .
The sl(V )-invariance of the Lie bracket in our Lie algebra g in (2.1) gives,
for any f, g ∈ sl(V ), u, v ∈ V and ϕ ∈ d, the following conditions:
[f ⊗ a, g ⊗ b] = [f, g] ⊗ a · b + 2 trace(f g)Da,b,
[f ⊗ a, u ⊗ x] = f (u) ⊗ a • x,
[u ⊗ x, v ⊗ y] = γu,v ⊗ hxyi +(cid:0)uv(cid:1)dx,y,
[ϕ, f ⊗ a] = f ⊗ ϕ(a),
[ϕ, u ⊗ x] = u ⊗ ϕ(x),
(2.3)
4
ALBERTO ELDUQUE AND SUSUMU OKUBO
for suitable bilinear maps
J × J → J,
(a, b) 7→ a · b,
(symmetric),
J × J → d,
J × T → T,
(a, b) 7→ Da,b,
(a, x) 7→ a • x,
(skew symmetric),
T × T → J,
(x, y) 7→ hxyi
(skew symmetric),
(2.4)
T × T → d,
d ×J → J,
(x, y) 7→ dx,y
(ϕ, a) 7→ ϕ(a),
d ×T → T,
(ϕ, x) 7→ ϕ(x),
(symmetric),
such that 1 · a = a, D1,a = 0 and 1 • x = x for any a ∈ J and x ∈ T .
The Jacobi identity on g shows that all these maps are invariant under
the action of the Lie subalgebra d. It also gives some other conditions to be
satisfied by these maps:
• For any f, g, h ∈ sl(V ) and a, b, c ∈ J we have,
[[f ⊗ a, g ⊗ b], h ⊗ c] = [[f, g], h] ⊗ (a · b) · c
+ 2 trace(f g)h ⊗ Da,b(c) + 2 trace([f, g]h)Da·b,c
For linearly independent f, g, h (generic case), equation (2.2) implies
that Pcyclic[[f ⊗ a, g ⊗ b], h ⊗ c] equals 0 if and only if the following
two conditions hold:
Da·b,c + Db.c,a + Dc·a,b = 0,
Da,b(c) = a · (b · c) − b · (a · c),
(2.5)
for any a, b, c ∈ J. Since the characteristic is 6= 3, the first equation
in (2.5) is equivalent to Da·a,a = 0, which in view of the second
equation is equivalent to:
(a · a) · (b · a) = ((a · a) · b) · a
for any a, b ∈ J. That is, J is a Jordan algebra.
This goes back to [Tits62], where the case in which T = 0 is
considered and the well-known Tits-Kantor-Koecher construction of
Lie algebras out of Jordan algebras appeared for the first time (see
also [Jac67]).
Therefore the Jacobi identity for elements in sl(V )⊗J is equivalent
to (J, ·) being a Jordan algebra and to the condition Da,b(c) = a ·
(b · c) − b · (a · c) for any a, b, c ∈ J (that is, Da,b acts as the natural
inner derivation of J attached to the elements a, b ∈ J [Jac67]).
• For any f, g ∈ sl(V ), a, b ∈ J, w ∈ V and x ∈ T we get:
[[f ⊗ a, g ⊗ b], w ⊗ x] = [[f, g] ⊗ a · b + 2 trace(f g)Da,b, w ⊗ x]
= [f, g](w) ⊗ (a · b) • x + 2 trace(f g)w ⊗ Da,b(x),
[[f ⊗ a, w ⊗ x], g ⊗ b] = [f (w) ⊗ a • x, g ⊗ b] = −g(f (w)) ⊗ b • (a • x),
[f ⊗ a, [g ⊗ b, w ⊗ x]] = [f ⊗ a, g(w) ⊗ b • x] = f (g(w)) ⊗ a • (b • x).
2 trace(f 2)1, Jacobi identity gives:
With f = g, and since f 2 = 1
4Da,b(x) = a • (b • x) − b • (a • x).
(2.6)
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
5
Now, using this fact and since f g+gf = trace(f g)1, with f, g ∈ sl(V )
and w ∈ W such that f (g(w)) and g(f (w)) are linearly independent
(generic case), we get
(a · b) • x =
1
2(cid:16)a • (b • x) + b • (a • x)(cid:17).
(2.7)
That is, the map J → EndF(T ), a 7→ λa, with λa(x) = a • x,
is a homomorphism of Jordan algebra, where EndF(T ) is a Jordan
algebra under the product α◦β = 1
algebra is denoted by EndF(T )+. In other words, T is a special unital
Jordan module for J.
2(cid:0)αβ +βα). This (special) Jordan
Conversely, equations (2.6) and (2.7) imply the Jacobi identity for
the elements above.
• For any f ∈ sl(V ), a ∈ J, u, v ∈ V , and x, y ∈ T , we get:
[f ⊗ a, [u ⊗ x, v ⊗ y]] = [f ⊗ a, γu,v ⊗ hxyi + (uv)dx,y]
= [f, γu,v] ⊗ a · hxyi − (uv)f ⊗ dx,y(a)
+ 2 trace(f γu,v)Da,hxyi,
[[f ⊗ a, u ⊗ x], v ⊗ y] = [f (u) ⊗ a • x, v ⊗ y]
[u ⊗ x, [f ⊗ a, v ⊗ y]] = [u ⊗ x, f (v) ⊗ a • y]
= γf (u),v ⊗ ha • xyi + (f (u)v)da•x,y,
= γu,f (v) ⊗ hxa • yi + (uf (v))dx,a•y.
But trace(f γu,v) = trace(cid:0)f(cid:0)(u.)v + (v.)u(cid:1)(cid:1) = (uf (v)) + (vf (u)) =
−2(cid:0)f (u)v) = 2(uf (v)), so the component in d of the above equa-
tions give:
4Da,hxyi = −da•x,y + dx,a•y,
(2.8)
for any a ∈ J and x, y ∈ T .
Also [f, γu,v] = γf (u),v + γu,f (v). Hence with u = v, [f, γu,u] =
2γf (u),u = 2γu,f (u), and we get
2a · hxyi = ha • xyi + hxa • yi,
(2.9)
for any a ∈ J and x, y ∈ T .
Since the dimension of V is 2, the skew symmetry forces
(w1w2)w3 = 0.
Xcyclic
Hence it follows that for any u, v, w ∈ V , (uv)f (w) + (f (w)u)v +
(vf (w))u = 0, or
(uv)f = −(f (u).)v + (f (v).)u,
(2.10)
for any u, v ∈ V and f ∈ sl(V ). Now, take u and v linearly indepen-
dent in V , and take f = γu,v. Then [f, γu,v] = 0, γf (u),v = −(uv)f =
−γu,f (v), so we obtain:
dx,y(a) = ha • xyi − hxa • yi.
(2.11)
Conversely, equations (2.8), (2.9) and (2.11) imply the Jacobi
identity for the elements above.
6
ALBERTO ELDUQUE AND SUSUMU OKUBO
• Finally, for u, v, w ∈ V and x, y, z ∈ T ,
[[u ⊗ x, v ⊗ y], w ⊗ z] = [γu,v ⊗ hxyi + (uv)dx,y, w ⊗ z]
= γu,v(w) ⊗ hxyi • z + (uv)w ⊗ dx,y(z).
With u and v such that (uv) = 1 and w = u, γu,v(u) = −u = γv,u(u),
γu,u(v) = 2v and the Jacobi identity gives:
dx,y(z) − dz,y(x) = hxyi • z − hzyi • x + 2hxzi • y
(2.12)
for any x, y, z ∈ T . And conversely, this equation is sufficient to
obtain the validity of the Jacobi identity for elements in V ⊗ T .
Let us collect the information obtained so far in the following result:
Theorem 2.2. Let g be a Lie algebra. Then g is BC1-graded of type C1 if
and only if there is a two-dimensional vector space V such that that g is, up
to isomorphism, the Lie algebra in (2.1), with Lie bracket given in (2.3), for
a unital Jordan algebra (J, ·), a special unital Jordan module T for J (with
action denoted by a • x for a ∈ J and x ∈ T ), and a Lie subalgebra d of g,
endowed with d-invariant bilinear maps:
(a, b) 7→ Da,b
(x, y) 7→ hxyi
(skew symmetric),
(skew symmetric),
J × J → d,
T × T → J,
T × T → d,
d ×J → J,
(x, y) 7→ dx,y
(d, a) 7→ d(a),
d ×T → T,
(d, x) 7→ d(x),
(symmetric),
satisfying equations (2.5), (2.6), (2.8), (2.9), (2.11), and (2.12), and such
that for any d ∈ d, the map J ⊕ T → J ⊕ T , a + x 7→ d(a) + d(x) is a
derivation of the algebra J ⊕ T , with the product given by the formula:
(a + x) ⋄ (b + y) =(cid:0)a · b + hxyi(cid:1) +(cid:0)a • y + b • x(cid:1). (cid:3)
Note that (2.13) above is a consequence of the bilinear maps involved
being invariant under the action of d, this being forced by the Jacobi identity
on g.
(2.13)
All this is strongly related to the J-ternary algebras considered by Allison
[All76]:
Definition 2.3. (see [ABG, (3.12)]) Let J be a unital Jordan algebra with
multiplication a · b. Let T be a unital special Jordan module for J with
action a • x for a ∈ J and x ∈ T . Assume h..i : T × T → J is a skew
symmetric bilinear map and (., ., .) : T × T × T → T is a trilinear product on
T . Then the pair (J, T ) is called a J-ternary algebra if the following axioms
hold for any a ∈ J and x, y, z, w, v ∈ T :
(JT1): a · hxyi =
1
2(cid:16)ha • xyi + hxa • yi(cid:17),
(JT2): a • (x, y, z) = (a • x, y, z) − (x, a • y, z) + (x, y, a • z),
(JT3): (x, y, z) = (z, y, x) − hxzi • y,
(JT4): (x, y, z) = (y, x, z) + hxyi • z,
(JT5): h(x, y, z)wi + hz(x, y, w)i = hxhzwi • yi,
(JT6): (x, y, (z, w, v)) = ((x, y, z), w, v)+(z, (y, x, w), v)+(z, w, (x, y, w)).
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
7
Theorem 2.4. Let (J, ·) be a unital Jordan algebra, let (T, •) be a special
unital Jordan module for (J, ·), and let d be a Lie algebra acting on both
vector spaces J and T (that is, both J and T are modules for d). Assume
D.,. : J × J → d, (a, b) 7→ Da,b, and h..i : T × T → J, (x, y) 7→ hxyi, are two
skew symmetric d-invariant bilinear maps, d.,. : T × T → d, (x, y) 7→ dx,y is
a symmetric d-invariant map, and the action of d on the direct sum J ⊕ T
is an action by derivations of the algebra (J ⊕ T, ⋄) in (2.13). Then if (2.5),
(2.6), (2.8), (2.9), (2.11) and (2.12) are satisfied, the pair (J, T ) becomes a
J-ternary algebra with the triple product (., ., .) : T × T × T → T given by
(x, y, z) =
1
2(cid:16)−dx,y(z) + hxyi • z(cid:17).
(2.14)
Conversely, if (J, T ) is a J-ternary algebra, then the image of the bilinear
maps
where
D.,. : J × J → EndF(J ⊕ T ), (a, b) 7→ Da,b,
Da,b(c) = a · (b · c) − b · (a · c),
Da,b(x) =
for a, b, c ∈ J and x ∈ T , and
1
4(cid:16)a • (b • x) − b • (a • x)(cid:17),
d.,. : T × T → EndF(J ⊕ T ), (x, y) 7→ dx,y,
where
dx,y(a) = ha • xyi − hxa • yi,
dx,y(z) = hxyi • z − 2(x, y, z),
(2.15)
(2.16)
for a ∈ J and x, y, z ∈ T , is contained in the Lie algebra of derivations of
the algebra (J ⊕ T, ⋄) defined in (2.13), and the equations (2.5), (2.8), (2.9)
and (2.12) are satisfied.
Proof. Assume first that (J, ·), (T, •), h., .i, D.,. and d.,. satisfy the conditions
in the theorem, then use (2.14) to define the trilinear product (., ., .) : T ×
T × T → T . Then condition (JT1) is equivalent to (2.9), while condition
(JT3) is equivalent to (2.12). Also (JT4) is a consequence of the symmetry
of d.,. and the skew symmetry of h..i.
For a ∈ J and x, y, z ∈ T we get:
a•(x, y, z) − (x, y, a • z)
=
=
=
1
1
1
2(cid:16)dx,y(a • z) − a • dx,y(z) + a •(cid:0)hxyi • z(cid:1) − hxyi • (a • z)(cid:17)
2(cid:16)dx,y(a) • z + 4Da,hxyi(z)(cid:17)
2(cid:16)ha • xyi • z − hxa • yi • z − da•x,y(z) + dx,a•y(z)(cid:17)
(using (2.6) and the fact that dx,y is a derivation)
(because of (2.8) and (2.11))
= (a • x, y, z) − (x, a • y, z),
8
ALBERTO ELDUQUE AND SUSUMU OKUBO
thus obtaining the validity of (JT2). Now, for x, y, z, w ∈ T we have:
h(x, y, z)wi + hz(x, y, w)i
1
1
=
=
=
2(cid:16)−hdx,y(z)wi − hzdx,y(w)i + hhxyi • zwi + hzhxyi • wi(cid:17)
2(cid:16)−dx,y(cid:0)hzwi(cid:1) + 2hxyi · hzwi(cid:17)
2(cid:16)−hhzwi • xyi + hxhzwi • yi
(using (2.9) and the fact that dx,y is a derivation)
+ hhzwi • xyi + hxhzwi • yi(cid:17) (using (2.11) and (2.9))
1
= hxhzwi • yi,
thus getting (JT5). Now the symmetry of d.,. and the fact that (., ., .) is a
d-invariant map show that (JT6) is a consequence of (JT2) and the skew
symmetry of h..i.
Conversely, let (J, T ) be a J-ternary algebra, and consider d = der(J ⊕
T, ⋄) and D.,. and d.,. defined by equations (2.15) and (2.16). Then both D.,.
and d.,. are d-invariant maps. Let us check that Da,b and dx,y belong to d
for any a, b ∈ J and x, y ∈ T .
Actually, Da,b acts as a derivation of (J, ·) since this is a Jordan algebra.
Besides, the equation Da,b(c • x) = Da,b(c) • x + c • Da,b(x) holds for any
a, b, c ∈ J and x ∈ T since T is a special module for J, and the condition
Da,b(hxyi) = hDa,b(x)yi + hxDa,b(y)i, for a, b ∈ J and x, y ∈ T , is a
consequence of (JT1). Therefore Da,b ∈ der(J ⊕ T, ⋄) for any a, b ∈ J.
Also, for x, y ∈ T , dx,y acts as a derivation of (J, ·) because of (JT1) and
(2.16). For x, y, z, w ∈ T , (JT5) is equivalent to the equation
hdx,y(z)wi + hzdx,y(w)i −(cid:16)hhxyi • zwi + hzhxyi • wi(cid:17)
= −2hxhzwi • yi,
and the expression in parenthesis above is, due to (JT1), 2hxyi · hzwi =
2hzwihxyi, which equals (again by (JT1)) hhzwi • xyi + hxhzwi • yi.
Hence (JT5) becomes, due to (JT1), the equation
hdx,y(z)wi + hzdx,y(w)i = hhzwi • xyi − hxhzwi • yi = dx,y(cid:0)hzwi(cid:1).
Finally, (JT4) shows that dx,y(z) = −(cid:16)(x, y, z) + (y, x, z)(cid:17), which is sym-
metric on x, y. For a ∈ J and x, y, z ∈ T we get:
dx,y(a • z) − dx,y(a) • z − a • dx,y(z)
= −(x, y, a • z) − (y, x, a • z) − ha • xyi • z + hxa • yi • z
+ a • (x, y, z) + a • (y, x, z)
=(cid:16)(a • x, y, z) − (y, a • x, z) − ha • xyi • z(cid:17)
+(cid:16)(a • y, x, z) − (x, a • y, z) − ha • yxi • z(cid:17) (thanks to (JT2))
= 0 (because of (JT4) and using the skew symmetry of h..i).
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
9
Therefore, the images of the maps D.,. and d.,. are contained in the Lie
algebra of derivations of (J ⊕ T, ⋄).
Now, (2.5) is a consequence of J being a Jordan algebra, (2.9) is equivalent
to (JT1), (2.12) is equivalent to (JT3). And as for the validity of (2.8):
4Da,hxyi = −da•x,y + dx,a•y for any a ∈ J and x, y ∈ T , this is valid on J as
a consequence of (JT1), while for z ∈ T :
4Da,hxyi(z) + da•x,y(z) − dx,a•y(z)
= a •(cid:0)hxyi • z(cid:1) − hxyi • (a • z)
−(cid:16)(a • x, y, z) + (y, a • x, z)(cid:17) +(cid:16)(x, a • y, z) + (a • y, x, z)(cid:17)
= a •(cid:16)(x, y, z) − (y, x, z)(cid:17) −(cid:16)(x, y, a • z) − (y, x, a • z)(cid:17)
−(cid:16)(a • x, y, z) + (y, a • x, z)(cid:17) +(cid:16)(x, a • y, z) + (a • y, x, z)(cid:17)
(where we have used (JT4))
= 0
(because of (JT2)). (cid:3)
Corollary 2.5. Let (J, T ) be a J-ternary algebra and let V be a two-
dimensional vector space endowed with a nonzero skew symmetric bilinear
map (..). Then the vector space
g(J, T ) =(cid:0)sl(V ) ⊗ J) ⊕(cid:0)V ⊗ T(cid:1) ⊕ d,
where d is the linear span of the operators Da,b and dx,y in EndF(J ⊕ T )
defined in (2.15) and (2.16), for a, b ∈ J and x, y ∈ T , and with bracket
given in (2.3), is a BC1-graded Lie algebra of type C1.
(cid:3)
(2.17)
3. J -ternary algebras and special Freudenthal-Kantor triple
systems
In [YO84], Yamaguti and Ono considered a wide class of triple sys-
tems: the (ǫ, δ) Freudenthal-Kantor triple systems, which extend the clas-
sical Freudenthal triple systems and are useful tools in the construction of
Lie algebras and superalgebras.
An (ǫ, δ) Freudental-Kantor triple system (ǫ, δ are either 1 or −1) over
a ground field F is a triple system (U, xyz) such that, if L(x, y), K(x, y) ∈
EndF(U ) are defined by
( L(x, y)z = xyz
K(x, y)z = xzy − δyzx,
then
[L(u, v), L(x, y)] = L(cid:0)L(u, v)x, y(cid:1) + ǫL(cid:0)x, L(v, u)y(cid:1),
K(cid:0)K(u, v)x, y(cid:1) = L(y, x)K(u, v) − ǫK(u, v)L(x, y),
hold for any x, y, u, v ∈ U .
(3.1)
(3.2a)
(3.2b)
For ǫ = −1 and δ = 1, these are the so called Kantor triple systems (or
generalized Jordan triple systems of second order [Kan73]).
10
ALBERTO ELDUQUE AND SUSUMU OKUBO
The aim of this section is to show the close relationship between the
J-ternary algebras and some particular (1, 1) Freudenthal-Kantor triple sys-
tems.
Lemma 3.1. Let (U, xyz) be a triple system satisfying equation (3.2a), with
ǫ ∈ {±1}, and define the endomorphisms S(x, y) and T (x, y) ∈ EndF(U ) by
S(x, y) = L(x, y) + ǫL(y, x)
T (x, y) = L(y, x) − ǫL(x, y).
(3.3)
Then for any u, v ∈ U , S(u, v) is a derivation of the triple system (U, xyz),
while T (u, v) satisfies
for any x, y, z ∈ U . As a consequence, the following equations hold:
T (u, v)(cid:0)xyz(cid:1) =(cid:0)T (u, v)x(cid:1)yz − x(cid:0)T (u, v)y(cid:1)z + xy(cid:0)T (u, v)z(cid:1)
[S(u, v), L(x, y)] = L(cid:0)S(u, v)x, y(cid:1) + L(cid:0)x, S(u, v)y(cid:1)
[T (u, v), L(x, y)] = L(cid:0)T (u, v)x, y(cid:1) − L(cid:0)x, T (u, v)y(cid:1)
[S(u, v), T (x, y)] = T(cid:0)S(u, v)x, y(cid:1) + T(cid:0)x, S(u, v)y(cid:1)
[T (u, v), S(x, y)] = −ǫT(cid:0)T (u, v)x, y(cid:1) + ǫT(cid:0)x, T (u, v)y(cid:1)
[T (u, v), T (x, y)] = −ǫS(cid:0)T (u, v)x, y(cid:1) + ǫS(cid:0)x, T (u, v)y(cid:1)
(3.4a)
(3.4b)
(3.4c)
(3.4d)
(3.4e)
Proof. The fact that S(u, v) is a derivation and that T (u, v) satisfies the
equation above when applied to a product xyz follow at once from (3.2a).
These conditions are equivalent to equations (3.4a) and (3.4b). The other
equations are obtained from these.
(cid:3)
Remark 3.2. Let (U, xyz) be a triple system satisfying equation (3.2a),
with ǫ ∈ {±1}, and denote by L (respectively S, T ) the linear span of the
operators L(x, y) (respectively S(x, y), T (x, y)) for x, y ∈ U . Lemma 3.1
shows that S is a subalgebra of L, and that the conditions L = S + T ,
[S, T ] ⊆ T and [T , T ] ⊆ S hold. In particular, this shows that S ∩ T is an
ideal of L, and modulo this ideal we get a Z2-graded Lie algebra.
Definition 3.3. Let U be an (ǫ, δ) Freudenthal-Kantor triple system. Then
U is said to be special in case
K(x, y) = ǫδL(y, x) − ǫL(x, y)
(3.5)
holds for any x, y ∈ U .
Moreover, U is said to be unitary in case the identity map belongs to
K(U, U ) (the linear span of the endomorphisms K(x, y)).
Especially, any balanced system (that is, K(x, y) = b(x, y)1 for some
nonzero bilinear form b(x, y)) is unitary.
Therefore, an (ǫ, ǫ) Freudenthal-Kantor triple system is special if the op-
erators K(x, y) coincide with our previous T (x, y) considered in Lemma 3.1.
Proposition 3.4. Let U 6= 0 be a unitary (ǫ, δ) Freudenthal-Kantor triple
system. Then ǫ = δ and U is special.
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
11
Proof. If U is unitary, then 1 =Pm
i=1 K(ui, vi) for some ui, vi ∈ U . Replace
u and v by ui and vi in (3.2b) and sum over i to get K(x, y) = L(y, x) −
ǫL(x, y). For ǫ = δ we get (3.5), while for ǫ = −δ this shows K(x, y) =
−ǫK(y, x), and (3.1) shows that K(x, y) = −δK(y, x), so that K(x, y) = 0
for any x, y ∈ U , a contradiction to U being unitary.
(cid:3)
Proposition 3.5.
(i) Let U be a special (ǫ, ǫ) Freudenthal-Kantor triple system. Then we
have
K(u, v)K(x, y) + K(x, y)K(u, v)
for any u, v, x, y ∈ U .
= K(cid:0)K(u, v)x, y(cid:1) + K(cid:0)x, K(u, v)y(cid:1),
(3.6)
Conversely, if the triple product on U satisfies equations (3.2a),
(3.5) and (3.6), then it defines a special (ǫ, ǫ) Freudenthal-Kantor
triple system.
(ii) Let U be a special (ǫ, −ǫ) Freudenthal-Kantor triple system. Then
K(x, y) is a derivation of U for any x, y ∈ U , and we have
xyz + ǫxzy = 0,
(3.7a)
K(u, v)T (x, y) + T (x, y)K(u, v) = K(cid:0)K(u, v)x, y(cid:1) − K(cid:0)x, K(u, v)y(cid:1),
(3.7b)
for any u, v, x, y ∈ U .
Conversely, if ǫ = −δ and the triple product on U satisfies (3.2a),
(3.7a) and (3.7b), then it defines a special (ǫ, −ǫ) Freudenthal-Kantor
triple system.
Proof. For (i), using (3.2b) we compute:
K(cid:0)K(u, v)x, y(cid:1) + K(cid:0)x, K(u, v)y(cid:1) = K(cid:0)K(u, v)x, y(cid:1) − ǫK(cid:0)K(u, v)y, x(cid:1)
=(cid:16)L(y, x) − ǫL(x, y)(cid:17)K(u, v) + K(u, v)(cid:16)−ǫL(x, y) + L(y, x)(cid:17)
= K(x, y)K(u, v) + K(u, v)K(x, y),
which proves (3.6).
For the converse statement, we must check the validity of (3.2b). Note
that L(x, y) = 1
(3.5). Hence we compute:
2(cid:0)S(x, y) − ǫK(x, y)(cid:1), and K(x, y) is our previous T (x, y) by
L(y,x)K(u, v) − ǫK(u, v)L(x, y)
2(cid:16)S(y, x) − ǫK(y, x)(cid:17)K(u, v) −
1
ǫ[S(x, y), K(u, v)] +
ǫ[K(u, v), S(x, y)] +
=
=
1
1
2
= −
1
2
1
2
ǫK(u, v)(cid:16)S(x, y) − ǫK(x, y)(cid:17)
2(cid:16)K(x, y)K(u, v) + K(u, v)K(x, y)(cid:17)
2(cid:16)K(x, y)K(u, v) + K(u, v)K(x, y)(cid:17)
1
= K(cid:0)K(u, v)x, y(cid:1) because of (3.4d) and (3.6).
12
ALBERTO ELDUQUE AND SUSUMU OKUBO
Now for (ii), assume U is a special (ǫ, −ǫ) Freudenthal-Kantor triple sys-
tem. Then K(x, y) = −ǫS(x, y) for any x, y and hence K(x, y) is a derivation
of U (Lemma 3.1). Then using (3.2b) with δ = −ǫ we get:
K(cid:0)K(u, v)x, y(cid:1) − K(cid:0)x, K(u, v)y(cid:1)
= K(cid:0)K(u, v)x, y(cid:1) − ǫK(cid:0)K(u, v)y, x(cid:1)
=(cid:16)L(y, x)K(u, v) − ǫK(u, v)L(x, y)(cid:17)
− ǫ(cid:16)L(x, y)K(u, v) − ǫK(u, v)L(y, x)(cid:17)
= T (x, y)K(u, v) + K(u, v)T (x, y)
for any u, v, x, y ∈ U . Consider too the trilinear map Λ(x, y, z) = xyz+ǫxzy.
Equations (3.1) and (3.5) give:
xzy − δyzx = ǫδyxz − ǫxyz,
which shows Λ(x, y, z) = δΛ(y, x, z) for any x, y, z ∈ U . But by its own
definition Λ(x, y, z) = ǫΛ(x, z, y). Hence with ǫ = −δ this gives Λ(x, z, y) =
−Λ(y, x, z), and hence
Λ(x, z, y) = −Λ(y, x, z) = Λ(z, y, x) = −Λ(x, z, y),
and we get Λ(x, y, z) = 0 for any x, y, z ∈ U . (Note that the argument above
shows that for special (ǫ, ǫ) Freudenthal-Kantor triple systems, Λ(x, y, z) is
symmetric on its arguments for ǫ = 1 and alternating for ǫ = −1.)
Conversely, with ǫ = −δ, by (3.7a) K(x, y) = −L(y, x) − ǫL(x, y) =
−ǫS(x, y) is a derivation of U , and T (x, y) = L(y, x) − ǫL(x, y), so L(x, y) =
−ǫ
2 (cid:0)T (x, y) + K(x, y)(cid:1) and L(y, x) = 1
that K(u, v) is a derivation of U and (3.7b) we obtain
2(cid:0)T (x, y) − K(x, y)(cid:1). Therefore, using
L(y, x)K(u, v) − ǫK(u, v)L(x, y)
1
2
K(u, v)(cid:0)T (x, y) + K(x, y)(cid:1)
[K(u, v), K(x, y)]
1
1
1
=
=
=
1
2
2(cid:0)T (x, y) − K(x, y)(cid:1)K(u, v) +
2(cid:0)K(u, v)T (x, y) + T (x, y)K(u, v)(cid:1) +
2(cid:16)K(cid:0)K(u, v)x, y(cid:1) − K(cid:0)x, K(u, v)y(cid:1)(cid:17)
= K(cid:0)K(u, v)x, y(cid:1)
2(cid:16)K(cid:0)K(u, v)x, y(cid:1) + K(cid:0)x, K(u, v)y(cid:1)(cid:17)
+
1
for any u, v, x, y ∈ U , thus getting (3.2b).
(cid:3)
Remark 3.6. Given an (ǫ, ǫ) Freudenthal-Kantor triple system, equation
(3.6) shows that the linear subspace F1 + span {K(x, y) : x, y ∈ U } is a
Jordan subalgebra of the special Jordan algebra EndF(U )+ (with product
f · g = 1
2 (f g + gf )).
Theorem 3.7. Let (J, T ) be a J-ternary algebra. Then T , endowed with its
triple product (x, y, z), is a special (1, 1) Freudenthal-Kantor triple system.
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
13
Conversely, let (U, xyz) be a special (1, 1) Freudenthal-Kantor triple sys-
tem. Let J be the F subspace of EndF(U ) spanned by the identity map
and by the operators K(x, y) for x, y ∈ U : J = F1 + K(U, U ). Then J
is a subalgebra of the special Jordan algebra EndF(U )+ (with multiplication
f · g = 1
2 (f g + gf ), and the pair (J, U ), endowed with the natural action of
J on U : a • x = a(x), and the multilinear maps:
(., ., .) : U × U × U → U,
(x, y, z) = xyz,
h..i : U × U → J,
hxyi = −K(x, y),
is a J-ternary algebra.
Proof. Assume first that (J, T ) is a J-ternary algebra and consider the space
T endowed with the triple product xyz = (x, y, z). Take ǫ = δ = 1. Then
(3.2a) is equivalent to (JT6), while K(x, y)z = (x, z, y)−(y, z, x) = −hxyi•
z by (JT3). Then (JT4) shows that K(x, y)z = −hxyi • z = (y, x, z) −
(x, y, z) = (cid:0)L(y, x) − L(x, y)(cid:1)z, which gives the speciality condition (3.5).
Finally, (JT5) gives
−K(cid:0)L(x, y)z, w(cid:1) − K(cid:0)z, L(x, y)w(cid:1) = K(cid:0)x, K(z, w)y(cid:1),
K(cid:0)K(z, w)y, x(cid:1) = K(cid:0)L(x, y)z, w(cid:1) + K(cid:0)z, L(x, y)w(cid:1).
But (3.2a) gives, for any x, y, z, w, u:
By skew symmetry we get
(3.8)
L(x, y)K(z, w)u = L(x, y)(cid:0)zuw − wuz)
so that
= K(cid:0)L(x, y)z, w(cid:1)u + K(cid:0)z, L(x, y)w(cid:1)u + K(z, w)L(y, x)u,
K(cid:0)L(x, y)z, w(cid:1) + K(cid:0)z, L(x, y)w(cid:1) = L(x, y)K(z, w) − K(z, w)L(y, x),
and (3.8) becomes (3.2b).
Conversely, let (U, xyz) be a special (1, 1) Freudenthal-Kantor triple sys-
tem, and let J = F1 + K(U, U ). Then (3.6) shows that J is a Jordan sub-
algebra of EndF(U )+, and hence U becomes a unital special Jordan module
for J. Now consider the trilinear product (x, y, z) = xyz and the skew
symmetric bilinear map h..i : U × U → J given by hxyi = −K(x, y).
Then equation (3.6) is equivalent to the condition (JT1), equation (3.2a)
and (3.5) give (JT2), equation (JT3) is a consequence of the definition
of the operator K(x, y) in (3.1), while (JT4) is equivalent to (3.5). Now
(JT5) follows from (3.2b) and the skew symmetry of K(., .), and (JT6) is
equivalent to (3.2a). Therefore, (J, U ) is a J-ternary algebra.
(cid:3)
Corollary 3.8. Let g be a BC1-graded Lie algebra of type C1 as in (2.1).
Then the vector space T , endowed with the triple product
(x, y, z) =
1
2(cid:0)−dx,y(z) + hxyi • z(cid:1)
in (2.14), is a special (1, 1) Freudenthal-Kantor triple system.
Conversely, given a special (1, 1) Freudenthal-Kantor triple system (U, xyz),
let d be the linear span of the derivations S(x, y) = L(x, y) + L(y, x) in (3.3)
(which is a subalgebra of the Lie algebra of derivations of der U by Lemma
14
ALBERTO ELDUQUE AND SUSUMU OKUBO
3.1) and let J be the Jordan subalgebra F1 + K(U, U ) of EndF(U )+. Then
the vector space
g =(cid:0)sl(V ) ⊗ J(cid:1) ⊕(cid:0)V ⊗ U(cid:1) ⊕ d,
with Lie bracket given by the bracket in d and, for any f, g ∈ sl(V ), u, v ∈ V ,
a, b ∈ J, x, y ∈ U and ϕ ∈ d, by the formulae:
2 (ab + ba) + 1
2 trace(f g)[a, b],
• [f ⊗ a, g ⊗ b] = [f, g] ⊗ 1
• [f ⊗ a, u ⊗ x] = f (u) ⊗ a(x),
• [u ⊗ x, v ⊗ y] = −γu,v ⊗ K(x, y) − (uv)(cid:0)L(x, y) + L(y, x)(cid:1),
• [ϕ, f ⊗ a] = f ⊗ ϕ(a),
• [ϕ, u ⊗ x] = u ⊗ ϕ(x),
is a BC1-graded Lie algebra of type C1.
Proof. It is enough to note that if (U, xyz) is a special (1, 1) Freudenthal-
Kantor triple system and (J, U ) is the associated J-ternary algebra, with
J = F1 + K(U, U ), then for x, y, z ∈ U , and a, b, c ∈ J:
dx,y(z) = hxyi • z − 2(x, y, z)
dx,y(a) = ha • xyi − hxa • yi
=(cid:0)−K(x, y) − 2L(x, y)(cid:1)(z)
= −(cid:0)L(x, y) + L(y, x)(cid:1)(z)
= −K(cid:0)a(x), y(cid:1) − K(cid:0)a(y), x(cid:1)
4(cid:0)a(b(x)) − b(a(x))(cid:1) =
= −[L(x, y) + L(y, x), a]
1
4
1
1
4
(because of (3.4d))
Da,b(x) =
[a, b](x),
Da,b(c) = a · (b · c) − b · (a · c) =
[[a, b], c].
(Note that we have 4(cid:16)a · (b · c) − b · (a · c)(cid:17) = a(bc + cb) + (bc + cb)a − b(ac +
ca) − (ac + ca)b = (ab − ba)c − c(ab − ba) = [[a, b], c].)
(cid:3)
Remark 3.9. One may change the skew symmetric bilinear form on V to
its negative and hence change the Lie bracket accordingly. In this way one
has [u ⊗ x, v ⊗ y] = γu,v ⊗ K(x, y) + (uv)(cid:0)L(x, y) + L(y, x)(cid:1), which is nicer
than the bracket above.
Remark 3.10. Let (U, xyz) be a special (1, 1) Freudenthal-Kantor triple
system and let g be the attached Lie algebra as in Corollary 3.8. Then the
subspace(cid:0)sl(V )⊗K(U, U )(cid:1)⊕(cid:0)V ⊗U(cid:1)⊕d is an ideal of g. Hence if g is simple
we get J = K(U, U ), so that (U, xyz) is unitary. Moreover, if dim J = 1,
then (U, xyz) is balanced.
4. Superalgebras
Any BC1-graded Lie algebra of type C1 (as in (2.1)) is Z2-graded, with
(4.1)
g¯1 = V ⊗ T.
In a similar vein we may consider Lie superalgebras g = g¯0 ⊕ g¯1 such that
g¯0 contains a subalgebra isomorphic to sl(V ) so that g¯0 is a direct sum of
g¯0 =(cid:0)sl(V ) ⊗ J(cid:1) ⊕ d,
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
15
copies of the adjoint module for sl(V ) and copies of the trivial module, and
g¯1 is a direct sum of copies of the two-dimensional module V . These will be
called strictly BC1-graded Lie superalgebras of type C1.
The arguments in the two previous sections apply here. The bracket on
the superalgebra g is given by (2.3) for suitable bilinear maps as in (2.4) with
the difference that h..i : T × T → J is symmetric here, while d.,. : T × T → d
is skew symmetric.
Theorem 2.2 has the following counterpart (whose proof is obtained by
superizing the arguments there):
Theorem 4.1. Let g be a Lie superalgebra. Then g is strictly BC1-graded
of type C1 if and only if there is a two-dimensional vector space V such
that that g is, up to isomorphism, the Lie superalgebra in (2.1), with g¯0 =
(cid:0)sl(V ) ⊗ J(cid:1) ⊕ d and g¯1 = V ⊗ T , and Lie bracket as in (2.3), for a unital
Jordan algebra (J, ·), a special unital Jordan module T for J (with action
denoted by a • x for a ∈ J and x ∈ T ), and a Lie subalgebra d of g, endowed
with d-invariant bilinear maps:
J × J → d,
T × T → J,
T × T → d,
d ×J → J,
(a, b) 7→ Da,b
(x, y) 7→ hxyi
(x, y) 7→ dx,y
(d, a) 7→ d(a),
d ×T → T,
(d, x) 7→ d(x),
(skew symmetric),
(symmetric),
(skew symmetric),
satisfying the following equations for any a, b, c ∈ J and x, y, z ∈ T :
Da·b,c + Db·c,a + Dc·a,b = 0, Da,b(c) = a · (b · c) − b · (a · c),
4Da,b(x) = a • (b • x) − b • (a • x),
4Da,hxyi = −da•x,y + dx,a•y,
2a · hxyi = ha • xyi + hxa • yi,
dx,y(a) = ha • xyi − hxa • yi,
dx,y(z) − dy,z(x) = hxyi • z + hyzi • x − 2hzxi • y,
(4.2)
(4.3)
(4.4)
(4.5)
(4.6)
(4.7)
and such that for any d ∈ d, the map J ⊕ T → J ⊕ T , a + x 7→ d(a) + d(x)
is a derivation of the algebra J ⊕ T , with the product given by the formula
(2.13).
The definition of a J-ternary algebra is changed to:
Definition 4.2. Let J be a unital Jordan algebra with multiplication a · b.
Let T be a unital special Jordan module for J with action a • x for a ∈ J
and x ∈ T . Assume h..i : T × T → J is a symmetric bilinear map and
(., ., .) : T × T × T → T is a trilinear product on T . Then the pair (J, T ) is
called a J-ternary (−1)-algebra if the following axioms hold for any a ∈ J
and x, y, z, w, v ∈ T :
1
2(cid:16)ha • xyi + hxa • yi(cid:17),
(-JT1): a · hxyi =
(-JT2): a • (x, y, z) = (a • x, y, z) − (x, a • y, z) + (x, y, a • z),
(-JT3): (x, y, z) + (z, y, x) = hxzi • y,
(-JT4): (x, y, z) + (y, x, z) = hxyi • z,
16
ALBERTO ELDUQUE AND SUSUMU OKUBO
(-JT5): h(x, y, z)wi + hz(x, y, w)i = hxhzwi • yi,
(-JT6): (x, y, (z, w, v)) = ((x, y, z), w, v)−(z, (y, x, w), v)+(z, w, (x, y, w)).
Now Theorem 2.4 has the following counterpart:
Theorem 4.3. Let (J, ·) be a unital Jordan algebra, let (T, •) be a special
unital Jordan module for (J, ·), and let d be a Lie algebra acting on both
vector spaces J and T (that is, both J and T are modules for d). Assume
D.,. : J × J → d, (a, b) 7→ Da,b, and d.,. : T × T → d, (x, y) 7→ dx,y, are two
skew symmetric d-invariant bilinear maps, h..i : T × T → J, (x, y) 7→ hxyi,
is a symmetric d-invariant bilinear map, and the action of d on the direct
sum J ⊕ T is an action by derivations of the algebra (J ⊕ T, ⋄) in (2.13).
Then if (4.2), (4.3), (4.4), (4.5), (4.6) and (4.7) are satisfied, the pair (J, T )
becomes a J-ternary (−1)-algebra with the triple product (., ., .) : T ×T ×T →
T given by (2.14):
(x, y, z) =
1
2(cid:16)−dx,y(z) + hxyi • z(cid:17).
Conversely, if (J, T ) is a J-ternary (−1)-algebra, then the image of the
bilinear maps
D.,. : J × J → EndF(J ⊕ T ), (a, b) 7→ Da,b,
where, as in (2.15):
Da,b(c) = a · (b · c) − b · (a · c),
Da,b(x) =
for a, b, c ∈ J and x ∈ T , and
1
4(cid:16)a • (b • x) − b • (a • x)(cid:17),
d.,. : T × T → EndF(J ⊕ T ), (x, y) 7→ dx,y,
where, as in (2.16)
dx,y(a) = ha • xyi − hxa • yi,
dx,y(z) = hxyi • z − 2(x, y, z),
for a ∈ J and x, y, z ∈ T , is contained in the Lie algebra of derivations of
the algebra (J ⊕ T, ⋄) defined in (2.13), and the equations (4.2), (4.4), (4.5)
and (4.7) are satisfied.
Finally, Theorem 3.7 becomes:
Theorem 4.4. Let (J, T ) be a J-ternary (−1)-algebra. Then T , endowed
with its triple product (x, y, z), is a special (−1, −1) Freudenthal-Kantor
triple system.
Conversely, let (U, xyz) be a special (−1, −1) Freudenthal-Kantor triple
system. Let J be the F subspace of EndF(U ) spanned by the identity map
and by the operators K(x, y) for x, y ∈ U . Then J is a subalgebra of the
special Jordan algebra EndF(U )+ (with multiplication f · g = 1
2 (f g + gf )),
and the pair (J, U ), endowed with the natural action of J on U : a•x = a(x),
and the multilinear maps:
(., ., .) : U × U × U → U,
(x, y, z) = xyz,
h..i : U × U → J,
hxyi = K(x, y),
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
17
is a J-ternary (−1)-algebra.
Remark 4.5. Note that K(x, y) = −hxyi in the (1, 1)-setting, but K(x, y) =
hxyi in the (−1, −1)-setting.
triple system.
Corollary 4.6. Let g be a strictly BC1-graded Lie superalgebra of type C1 as
in (2.1). Then the vector space T , endowed with the triple product (x, y, z) =
1
2(cid:0)−dx,y(z) + hxyi • z(cid:1) in (2.14), is a special (−1, −1) Freudenthal-Kantor
Conversely, given a special (−1, −1) Freudenthal-Kantor triple system T ,
let d be the linear span of the derivations S(x, y) = L(x, y) − L(y, x) and
let J be the Jordan subalgebra F1 + K(T, T ) of EndF(T )+. Then the vector
space
g =(cid:0)sl(V ) ⊗ J(cid:1) ⊕(cid:0)V ⊗ T(cid:1) ⊕ d,
with Lie bracket given by the bracket in d and, for any f, g ∈ sl(V ), u, v ∈ V ,
a, b ∈ J, x, y ∈ T and ϕ ∈ d, by the formulae:
2 (ab + ba) + 1
2 trace(f g)[a, b],
• [f ⊗ a, g ⊗ b] = [f, g] ⊗ 1
• [f ⊗ a, u ⊗ x] = f (u) ⊗ a • x,
• [u ⊗ x, v ⊗ y] = γu,v ⊗ K(x, y) − (uv)(cid:0)L(x, y) − L(y, x)(cid:1),
• [ϕ, f ⊗ a] = f ⊗ ϕ(a),
• [ϕ, u ⊗ x] = u ⊗ ϕ(x),
is a strictly BC1-graded Lie superalgebra of type C1, with g¯0 =(cid:0)sl(V )⊗J(cid:1)⊕d
and g¯1 = V ⊗ T .
5. Dicyclic symmetry
Let V be a two-dimensional vector space endowed with a nonzero (and
hence nondegenerate) skew symmetric bilinear form (..). Let {u, v} be a
symplectic basis of V (that is, (uv) = 1). Assume that the ground field
contains the primitive cubic roots ω and ω2 of 1. Then the correspond-
ing symplectic group Sp(V ) = SL(V ) contains the elements θ, φ, whose
coordinate matrices on the basis above are:
1
θ ↔(cid:18) 0
φ ↔(cid:18)ω
−1 0(cid:19) ,
0 ω2(cid:19) .
(so θ(u) = −v, θ(v) = u),
0
These elements satisfy
θ4 = 1 = φ3, φθφ = θ,
(5.1)
(5.2)
and hence the group they generate is the dicyclic group Dic3 (semidirect
product of C3 by C4).
It turns out that conjugation by elements of Dic3 gives an action of this
group by automorphisms on sl(V ) (σ(f ) = σf σ−1 for any σ ∈ Dic3 and
f ∈ sl(V )), and hence that any BC1 graded Lie algebra of type C1 (as in
(2.1)) is endowed with an action of Dic3 by automorphisms.
The aim of this section is the study of those Lie algebras endowed with
an action of Dic3 by automorphisms.
18
ALBERTO ELDUQUE AND SUSUMU OKUBO
Let g be a Lie algebra endowed with such an action. Then φ induces a
grading over Z3:
(5.3)
where g1 = {x ∈ g : φ(x) = x}, gω = {x ∈ g : φ(x) = ωx} and gω2 = {x ∈
g : φ(x) = ω2x}.
g = g1 ⊕ gω ⊕ gω2,
Since φθφ = θ, the following conditions hold:
θ(g1) = g1,
θ(gω) = gω2,
θ(gω2) = gω.
This allows us to define a unary, binary and ternary operations on gω as
follows:
¯x = θ2(x),
x ∗ y = θ−1(cid:0)[x, y](cid:1),
{x, y, z} = [[x, θ(y)], z],
(5.4a)
(5.4b)
(5.4c)
for x, y, z ∈ gω. It is easy to check that these are indeed well defined. For
instance, for x, y, z ∈ gω, θ(y) belongs to gω2, and hence we have [x, θ(y)] ∈
[gω, gω2 ] ⊆ g1 and {x, y, z} = [[x, θ(y)], z] ∈ [g1, gω] ⊆ gω.
Theorem 5.1. Let g be a Lie algebra over a field F containing the primitive
cubic roots of 1, endowed with an action by automorphisms of the dicyclic
group Dic3. Let ¯x, x ∗ y and {x, y, z} be defined as in (5.4) on gω = {x ∈
g : φ(x) = ωx}. Then the binary product x ∗ y is anticommutative and
the map x 7→ ¯x is an order 2 automorphism of both the binary and the
ternary products: x ∗ y = ¯x ∗ ¯y, {x, y, z} = {¯x, ¯y, ¯z}. Moreover, the following
identities hold for any u, v, , w, x, y, z ∈ gω:
(D1): {x, z, y} − {y, z, x} = (¯x ∗ ¯y) ∗ ¯z,
(D2): {u, ¯v, x ∗ y} + {v, u, x} ∗ y + x ∗ {v, u, y} = 0,
(D3): {x, y ∗ z, w} + {y, z ∗ x, w} + {z, x ∗ y, w} = 0,
(D4): {¯x ∗ ¯y, z, w} + {¯y ∗ ¯z, x, w} + {¯z ∗ ¯x, y, w} = 0,
(D5): {u, v, {x, y, z}} = {{u, v, x}, y, z} − {x, {v, ¯u, y}, z} + {x, y, {u, v, z}}.
Proof. The anticommutativity of x ∗ y is clear. Also, since θ2 generates the
center of the dicyclic group Dic3, it follows easily that x 7→ ¯x = θ2(x) is an
order 2 automorphism of both the binary and ternary products.
Now for x, y, z ∈ gω:
{x, z, y} − {y, z, x} = [[x, θ(z)], y] − [[y, θ(z)], x]
= [[x, y], θ(z)] = θ(cid:0)[θ−1([x, y]), z](cid:1)
= θ(cid:0)[x ∗ y, z]) = θ2(cid:0)θ−1([x ∗ y, z])(cid:1)
= (x ∗ y) ∗ z = (¯x ∗ ¯y) ∗ ¯z,
which gives (D1).
For any d ∈ g1 we have
[d, x ∗ y] = [d, θ−1([x, y])] = θ−1(cid:0)[θ(d), [x, y]](cid:1)
= θ−1(cid:0)[[θ(d), x], y] + [x, [θ(d), y]](cid:1)
= [θ(d), x] ∗ y + x ∗ [θ(d), y],
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
19
and thus
{u, ¯v, x ∗ y} = [[u, θ(¯v)], x ∗ y]
= [θ([u, θ(¯v)]), x] ∗ y + x ∗ [θ([u, θ(¯v)]), y]
= −[[v, θ(u)], x] ∗ y − x ∗ [[v, θ(u)], y]
= −{v, u, x} ∗ y − x ∗ {v, u, y},
which gives (D2).
Equation (D3) is a direct consequence of the Jacobi identity, since {x, y ∗
z, w} = [[x, θ(y ∗ z)], w] = [[x, [y, z]], w]. The same happens for (D4), as {¯x ∗
¯y, z, w} = [[¯x ∗ ¯y, θ(z)], w] = [[x ∗ y, θ(z)], w] = [[θ2(cid:0)θ−1([x, y])(cid:1), θ(z)], w] =
[[[θ(x), θ(y)], θ(z)], w].
Finally,
{u, v, {x, y, z}}
= [[u, θ(v)], [[x, θ(y)], z]]
= [[[[u, θ(v)], x], θ(y)], z] + [[x, [[u, θ(v)], θ(y)]], z] + [[x, θ(y)], [[u, θ(v)], z]]
= {{u, v, x}, y, z} − [[x, θ(cid:0)[[v, θ−1(u)], y](cid:1)], z] + {x, y, {u, v, z}}
= {{u, v, x}, y, z} − {x, {v, ¯u, y}, z} + {x, y, {u, v, z}},
thus obtaining (D5).
(cid:3)
Definition 5.2. Let (A,¯, ∗, {...}) be a vector space endowed with a linear
endomorphism x 7→ ¯x, a binary product x ∗ y and a triple product {x, y, z}.
Then (A,¯, ∗, {...}) is said to be a dicyclic ternary algebra (or simply a D-
ternary algebra) if the linear map ¯ is an order 2 automorphism of both the
binary and the triple products, the binary product is anticommutative, and
conditions (D1) -- (D5) are satisfied.
Theorem 5.1 has a natural converse:
Theorem 5.3. Let (A,¯, ∗, {...}) be a dicyclic algebra. Let ι1(A) and ι2(A)
denote two copies of A, and for any u, v ∈ A denote by τ (u, v) ∈ EndF(A) the
linear map x 7→ {u, v, x}, and by τ (A, A) the linear span of these operators.
Then the vector space direct sum
g(A) = τ (A, A) ⊕ ι1(A) ⊕ ι2(A)
is a Lie algebra under the bracket:
[τ (u, v), τ (x, y)] = τ(cid:0){u, v, x}, y(cid:1) − τ(cid:0)x, {v, ¯u, y}(cid:1),
[τ (u, v), ι1(x)] = ι1(cid:0){u, v, x}(cid:1),
[τ (u, v), ι2(x)] = −ι2(cid:0){v, ¯u, x}(cid:1),
[ι1(x), ι1(y)] = ι2(x ∗ y),
[ι2(x), ι2(y)] = ι1(x ∗ y),
[ι1(x), ι2(y)] = τ (x, y),
for any u, v, x, y ∈ A.
20
ALBERTO ELDUQUE AND SUSUMU OKUBO
Moreover, if the ground field F contains the primitive cubic roots ω and
ω2 of 1, then g is endowed with an action by automorphisms of the dicyclic
group Dic3 as follows:
φ(cid:0)τ (x, y)(cid:1) = τ (x, y)
θ(cid:0)τ (x, y)(cid:1) = −τ (¯y, x)
for any x, y ∈ A.
φ(cid:0)ι1(x)(cid:1) = ωι1(x)
θ(cid:0)ι1(x)(cid:1) = ι2(x)
φ(cid:0)ι2(x)(cid:1) = ω2ι2(x)
θ(cid:0)ι2(x)(cid:1) = ι1(¯x)
(cid:3)
The proof consists of straightforward computations using (D1) -- (D5).
Note that the bracket [τ (u, v), τ (x, y)] is just, due to (D5), the bracket of
these operators in EndF(A).
In ending this section, we simply note that the generalized Malcev algebra
based upon the S3 symmetry considered in [EO09] is a special case of the
D-ternary algebra in which we have ¯x = x for any x ∈ U .
6. Dicyclic and J -ternary algebras
Let (J, T ) be a J-ternary algebra and consider the attached Lie algebra
in (2.17):
g(J, T ) =(cid:0)sl(V ) ⊗ J) ⊕(cid:0)V ⊗ T(cid:1) ⊕ d .
Then, if the ground field F contains the primitive cubic roots ω, ω2 of 1, the
dicyclic group Dic3 (which is a subgroup of the symplectic group Sp(V ) =
SL(V )) acts on g(J, T ) by automorphisms. Recall that the action of Dic3
on sl(V ) is by conjugation. The action on d is trivial.
Let {u, v} be a symplectic basis of V : (uv) = 1, and consider the standard
basis {H, E, F } of sl(V ), with coordinate matrices in this basis:
0
H ↔(cid:18)1
which satisfy
0 −1(cid:19) , E ↔(cid:18)0 1
0 0(cid:19) , F ↔(cid:18)0 0
1 0(cid:19) ,
(6.1)
[H, E] = 2E,
[H, F ] = −2F,
[E, F ] = H.
Let γw1,w2 = (w1.)w2 + (w2.)w1 be as in Lemma 2.1. Then, relative to our
symplectic basis we have:
γu,v = −H,
γu,u = 2E,
γv,v = −2F.
(6.2)
The action of the elements θ, φ ∈ Dic3 in (5.1) induces the action by conju-
gation on sl(V ), which satisfies:
φHφ−1 = H,
θHθ−1 = −H,
φEφ−1 = ω2E,
θEθ−1 = −F,
φF φ−1 = ωF,
θF θ−1 = −E.
(Note that σγw1,w2σ−1 = γσ(w1),σ(w2) for any σ ∈ Sp(V ) and w1, w2 ∈ V .)
Then the Lie algebra g = g(J, T ) decomposes under the action of φ as in
(5.3) with
g1 =(cid:0)H ⊗ J(cid:1) ⊕ d,
gω =(cid:0)F ⊗ J(cid:1) ⊕(cid:0)u ⊗ T(cid:1),
gω2 =(cid:0)E ⊗ J(cid:1) ⊕(cid:0)v ⊗ T(cid:1).
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
21
Therefore gω =(cid:0)F ⊗ J(cid:1) ⊕(cid:0)u ⊗ T(cid:1) becomes a dicyclic ternary algebra with
order 2 automorphism given by
F ⊗ a = F ⊗ a,
u ⊗ x = −u ⊗ x,
for any a ∈ J and x ∈ T (recall from (5.4) that ¯ is the restriction to gω of
θ2), with binary and triple products given by the following formulas (where
we use (2.3)):
(F ⊗ a) ∗ (F ⊗ b) = θ−1(cid:0)[F ⊗ a, F ⊗ b](cid:1) = 0,
(F ⊗ a) ∗ (u ⊗ x) = θ−1(cid:0)[F ⊗ a, u ⊗ x](cid:1)
= θ−1(cid:0)v ⊗ a • x(cid:1)
= −u ⊗ a • x
(u ⊗ x) ∗ (u ⊗ y) = θ−1(cid:0)[u ⊗ x, u ⊗ y](cid:1)
= θ−1(cid:0)2E ⊗ hxyi(cid:1)
= −2F ⊗ hxyi,
and with triple product given by:
{F ⊗ a, u ⊗ x, X} = [[F ⊗ a, θ(u ⊗ x)], X]
= −[[F ⊗ a, v ⊗ x], X] = 0,
{u ⊗ x, F ⊗ a, X} = [[u ⊗ x, θ(F ⊗ a)], X]
= −[[u ⊗ x, E ⊗ a], X] = 0,
for any a ∈ J, x ∈ T , and by:
{F ⊗ a, F ⊗ b, F ⊗ c} = [[F ⊗ a, θ(F ⊗ b)], F ⊗ c]
= −[[F ⊗ a, E ⊗ b], F ⊗ c]
= [H ⊗ a · b − 2Da,b, F ⊗ c]
(as trace(F E) = 1)
{F ⊗ a, F ⊗ b, u ⊗ x} = [H ⊗ a · b − 2Da,b, u ⊗ x]
= −2F ⊗(cid:0)(a · b) · c + Da,b(c)(cid:1)
= −2F ⊗(cid:0)(a · b) · c + a · (b · c) − (a · c) · b(cid:1),
= u ⊗(cid:0)(a · b) • x − 2Da,b(x)(cid:1)
= u ⊗ (b • (a • x)),
(using (2.6) and (2.7))
{u ⊗ x, u ⊗ y, F ⊗ a} = [[u ⊗ x, θ(u ⊗ y)], F ⊗ a]
= −[[u ⊗ x, v ⊗ y], F ⊗ a]
= [H ⊗ hxyi − dx,y, F ⊗ a]
= F ⊗(cid:0)−2hxyi · a − dx,y(a)(cid:1)
= −2F ⊗ ha • xyi,
(using (2.9) and (2.11))
{u ⊗ x, u ⊗ y, u ⊗ z} = [H ⊗ hxyi − dx,y, u ⊗ z]
= u ⊗(cid:0)hxyi • z − dx,y(z)(cid:1)
= 2u ⊗ (x, y, z),
(using (2.14))
for any a, b ∈ J, x, y, z ∈ T and X ∈ gω.
22
ALBERTO ELDUQUE AND SUSUMU OKUBO
We may identify gω = (cid:0)F ⊗ J(cid:1) ⊕(cid:0)u ⊗ T(cid:1) with A = J ⊕ T , by means of
(F ⊗ a) + (u ⊗ x) ↔ a + x ∈ J ⊕ T .
The above computations are summarized in the next result.
Proposition 6.1. Let (J, T ) be a J-ternary algebra over a field F containing
the primitive cubic roots ω, ω2 of 1, and let g = g(J, T ) be the attached BC1-
graded Lie algebra of type C1 with the above action of the dicyclic group Dic3
by automorphisms. Then the dicyclic ternary algebra defined on gω by means
of (5.4) is isomorphic to the dicyclic ternary algebra (A = J ⊕ T,¯, ∗, {...})
with ¯a = a, ¯x = −x, for any a ∈ J and x ∈ T , and with
a ∗ b = 0,
a ∗ x = −x ∗ a = −a • x,
x ∗ y = −2hxyi,
{a, x, w} = 0 = {x, a, w},
{a, b, c} = −2(cid:0)(a · b) · c + a · (b · c) − (a · c) · b(cid:1),
{a, b, x} = b • (a • x),
{x, y, a} = −2ha • xyi,
{x, y, z} = 2(x, y, z),
(6.3)
for any a, b, c ∈ J, x, y, z ∈ T and w ∈ J ⊕ T .
(cid:3)
Let (J, T ) be a J-ternary algebra, and let A = J ⊕ T be the associated
dicyclic ternary algebra as in Proposition 6.1. Due to (6.3), the unity 1 of
the Jordan algebra J verifies the following conditions:
{1, 1, a} = −2a = {a, 1, 1},
{1, 1, x} = x,
{x, 1, 1} = 0,
{a, 1, b} = −2a · b,
1 ∗ a = 0,
1 ∗ x = −x,
(6.4)
for any a ∈ J and x ∈ T .
Our next purpose is to show that the existence of an element with these
properties characterizes the dicyclic ternary algebras coming from a J-
ternary algebra. Actually a stronger result (Corollary 6.7) will be proved.
Given a dicyclic ternary algebra (A,¯, ∗, {...}), the automorphism x 7→ ¯x
induces a Z2-grading: A = A0 ⊕ A1, where A0 = {x ∈ A : ¯x = x} and
A1 = {x ∈ A : ¯x = −x}. Usually the elements of A0 will be denoted by
a, b, ..., while the elements of A1 by x, y, ... in what follows.
Lemma 6.2. Let (A,¯, ∗, {...}) be a dicyclic ternary algebra which contains
an element e ∈ A0 such that {e, e, e} = −2e and τ (e, e) (= {e, e, .}) acts as a
scalar on both A0 and A1. Then {e, e, a} = −2a for any a ∈ A0, A0 ∗ A0 = 0
and either A ∗ A = 0 or {e, e, x} = x for any x ∈ A1.
Proof. As {e, e, e} = −2e, our hypotheses imply {e, e, a} = −2a for any
a ∈ A0. Let α ∈ F such that {e, e, x} = αx for any x ∈ A1. Now equation
(D2) implies {e, e, s ∗ t} + {e, e, s} ∗ t + s ∗ {e, e, t} = 0 for any s, t ∈ A.
For s = a ∈ A0 and t = b ∈ A0 we get −6a ∗ b = 0, so that A0 ∗ A0 = 0.
For s = a ∈ A0 and t = x ∈ A1 we get (2α − 2)a ∗ x = 0, so either α = 1
or A0 ∗ A1 = 0. Finally, for s = x ∈ A1 and t = y ∈ A1, we obtain
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
23
(2 − 2α)x ∗ y = 0, so either α = 1 or A1 ∗ A1 = 0. Therefore either α = 1 or
A ∗ A = 0 as required.
(cid:3)
Lemma 6.3. Let (A,¯, ∗, {...}) be a dicyclic ternary algebra which contains
an element e ∈ A0 such that {e, e, a} = −2a for any a ∈ A0, and {e, e, x} =
x and {x, e, e} = 0 for any x ∈ A1. Then {A0, A1, A} = 0 = {A1, A0, A}.
Proof. Note first that for x ∈ A1,
{e, x, e} = {e, e, {e, x, e}}
= {{e, e, e}, x, e} − {e, {e, e, x}, e} + {e, x, {e, e, e}} using (D5)
= −2{e, x, e} − {e, x, e} − 2{e, x, e} = −5{e, x, e},
so we get {e, A1, e} = 0.
For a ∈ A0 and x ∈ A1, (D3) gives {e, a∗x, e}+{a, x∗e, e}+{x, e∗a, e} =
0, but e ∗ a = 0 (Lemma 6.2), and {e, a ∗ x, e} ∈ {e, A1, e} = 0. We conclude
that {A0, A1 ∗ e, e} = 0 holds. But (D1) gives x = {e, e, x} − {x, e, e} =
−(e ∗ x) ∗ e = (x ∗ e) ∗ e. Hence A1 ∗ e = A1 and {A0, A1, e} = 0.
Now (D4) gives {a∗x, e, e}−{x∗e, a, e}+{e∗a, x, e} = 0, but {a∗x, e, e} ∈
{A1, e, e} = 0 and e ∗ A0 ∈ A0 ∗ A0 = 0, so we conclude {A1, A0, e} = 0.
Therefore, {A0, A1, e} = {A1, A0, e} = 0. Finally, (D5) gives
[τ (e, e), τ (a, x)] = τ(cid:0){e, e, a}, x(cid:1) − τ(cid:0)a, {e, e, x}(cid:1) = −3τ (a, x),
but also
[τ (e, e), τ (a, x)] = −[τ (a, x), τ (e, e)] = −τ(cid:0){a, x, e}, e(cid:1) − τ(cid:0)e, {x, a, e}(cid:1) = 0.
Thus τ (a, x) = 0, and in the same vein we get τ (x, a) = 0, as required. (cid:3)
Under the conditions of the previous Lemma, the Lie algebra g(A) =
τ (A, A)⊕ι1(A)⊕ι2(A) in Theorem 5.3 satisfies that its subalgebra τ (A, A) =
τ (A0, A0) + τ (A1, A1) contains the elements
H = τ (e, e), E = −ι2(e), F = ι1(e),
which satisfy [E, F ] = H, [H, E] = 2E, and [H, F ] = −2F , and hence span
a subalgebra isomorphic to sl2(F). Moreover, the following conditions hold:
[E, ι1(A1)] = −τ (A1, e) = 0,
[F, ι1(x)] = ι2(e ∗ x),
[H, ι1(x)] = ι1(x),
[F, ι2(A1)] = τ (e, A1) = 0,
[E, ι2(x)] = ι1(e ∗ x),
[H, ι2(x)] = −ι2(x),
for any x ∈ A1.
Hence for any 0 6= x ∈ A1, the vector space Fι1(x) + Fι2(e ∗ x) is a two-
dimensional irreducible module for our copy of sl2(F), since e ∗ (e ∗ x) = x.
Also, we have [E, ι2(A0)] = 0 = [F, ι1(A0)] as e ∗ A0 = 0, and there-
fore, for any 0 6= a ∈ A0, the vector space Fι1(a) + Fτ (a, e) + Fι2({e, a, e})
is isomorphic to the adjoint module for sl2(F) under the map F 7→ ι1(a),
H 7→ τ (a, e) and E 7→ 1
2 ι2({e, a, e}), because [E, ι1(a)] = −[ι2(e), ι1(a)] =
τ (a, e), and [E, τ (a, e)] = [τ (a, e), ι2(e)] = −ι2({e, a, e} (while [E, H] =
−2E). Note that for any a ∈ A0, (D5) gives 4a = {a, e, {e, e, e}} =
{{a, e, e}, e, e} − {e, {e, a, e, }, e} + {e, e, {a, e, e}} = 8a − {e, {e, a, e}, e}, so
that {e, {e, a, e}, e} = 4a, and in particular {e, a, e} 6= 0 if a 6= 0.
We summarize these arguments in the following result:
24
ALBERTO ELDUQUE AND SUSUMU OKUBO
Lemma 6.4. Let (A,¯, ∗, {...}) be a dicyclic ternary algebra which contains
an element e ∈ A0 such that {e, e, a} = −2a for any a ∈ A0, and {e, e, x} =
x and {x, e, e} = 0 for any x ∈ A1. Then
• For any 0 6= a ∈ A0, the vector space Fι1(a)+ Fτ (a, e)+ Fι2({e, a, e})
is isomorphic to the adjoint module for sl2(F) under the map F 7→
ι1(a), H 7→ τ (a, e) and E 7→ 1
2 ι2({e, a, e})
• For any 0 6= x ∈ A1, the vector space Fι1(x) + Fι2(e ∗ x) is a two-
(cid:3)
dimensional irreducible module for sl2(F).
An extra previous result is still needed:
Lemma 6.5. Let (A,¯, ∗, {...}) be a dicyclic ternary algebra which con-
tains an element e ∈ A0 such that {e, e, a} = −2a for any a ∈ A0, and
{e, e, x} = x and {x, e, e} = 0 for any x ∈ A1. Then τ (A, A) = τ (A0, e) ⊕
Centg(A)(cid:0)sl2(F)(cid:1) (where Centg(A)(cid:0)sl2(F)(cid:1) denotes the centralizer in g(A) of
our subalgebra sl2(F) = FH ⊕ FE ⊕ FF ).
Proof. The element H = τ (e, e) acts diagonally with eigenvalues −2 on
ι1(A0), 2 on ι2(A0), 1 on ι1(A1), −1 on ι2(A1) and 0 on τ (A, A) = τ (A0, A0)+
τ (A1, A1) (see Theorem 5.3 for the bracket on the Lie algebra g(A)). In par-
ticular we have
Centg(A)(cid:0)sl2(F)(cid:1) ⊆ {X ∈ g(A) : [H, X] = 0} = τ (A, A).
Also note that for 0 6= a ∈ A0 and 0 6= x ∈ A1, [F, τ (a, e)] = −[τ (a, e), ι1(e)] =
−ι1({a, e, e}) = 2ι1(a) 6= 0, so we get
Now, for x, y ∈ A1, (D4) gives
τ (A0, e) ∩ Centg(A)(cid:0)sl2(F)(cid:1) = 0.
−τ (x ∗ e, y) − τ (e ∗ y, x) + τ (y ∗ x, e) = 0,
which we rewrite as
τ (x ∗ y, e) = τ (e ∗ x, y) − τ (e ∗ y, x).
Hence we get
τ (x ∗ y, e) +
1
1
2
and
τ (e ∗ x, y) =
2(cid:0)τ (e ∗ x, y) + τ (e ∗ y, x)(cid:1),
[τ (e ∗ x, y) + τ (e ∗ y, x), ι1(e)] = ι1(cid:0){e ∗ x, y, e} + {e ∗ y, x, e}(cid:1)
[τ (e ∗ x, y) + τ (e ∗ y, x), ι2(e)] = ι2(cid:0){y, e ∗ x, e} + {x, e ∗ y, e}(cid:1).
But τ (A0, A1) = 0 (Lemma 6.3) and (D1) gives:
(6.5)
(6.6)
(6.7)
{e ∗ x, y, e} = ((e ∗ x) ∗ e) ∗ y = −{x, e, e} ∗ y = −x ∗ y,
so
{e ∗ x, y, e} + {e ∗ y, x, e} = −(x ∗ y + y ∗ x) = 0.
In a similar way we get {x, e ∗ y, e} = (x ∗ e) ∗ (e ∗ y), so
{x, e ∗ y, e} + {y, e ∗ x, e} = −(cid:0)(e ∗ x) ∗ (e ∗ y) + (e ∗ y) ∗ (e ∗ x)(cid:1) = 0.
It follows now from (6.7) that
[τ (e ∗ x, y) + τ (e ∗ y, x), E] = 0 = [τ (e ∗ x, y) + τ (e ∗ y, x), F ],
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
25
so that the element τ (e ∗ x, y) + τ (e ∗ y, x) is in Centrg(A)(cid:0)sl2(F)(cid:1), and hence
(6.6) shows that
τ (A1, A1) = τ (e ∗ A1, A1) ⊆ τ (A0, e) + Centrg(A)(cid:0)sl2(F)(cid:1).
Finally, for any a, b ∈ A0 we have:
(6.8)
[τ (a, e), τ (b, e)] = τ ({a, e, b}, e) − τ (b, {e, a, e}),
[τ (a, e), τ (b, e)] = −[τ (b, e), τ (a, e)] = −τ ({b, e, a}, e) + τ (a, {e, b, e}).
Hence, as {a, e, b} = {b, e, a} by (D1) (as A0 ∗ A0 = 0 by Lemma 6.2), we
get
2τ ({a, e, b}, e) = τ (a, {e, b, e}) + τ (b, {e, a, e}),
and
τ (a, {e, b, e}) = τ ({a, e, b}, e) +
Also,
−2{a, e, b} = {a, e, {b, e, e}}
1
2(cid:0)τ (a, {e, b, e} − τ (b, {e, a, e}(cid:1).
(6.9)
= {{a, e, b}, e, e} − {b, {e, a, e}, e} + {b, e, {a, e, e}}
= −2{a, e, b} − {b, {e, a, e}, e} − 2{b, e, a}.
Hence {b, {e, a, e}, e} = 2{b, e, a}, which is symmetric on a, b by (D1).
Therefore,
[τ (a, {e, b, e}) − τ (b, {e, a, e}), F ] = [τ (a, {e, b, e}) − τ (b, {e, a, e}), ι1(e)]
And because of (D1) and (D5) we have:
= ι1(cid:0){a, {e, b, e}, e} − {b, {e, a, e}, e}(cid:1) = 0.
{{e, a, e}, b, e} = {e, b, {e, a, e}}
= {{e, b, e}, a, e} − {e, {b, e, a}, e} + {e, a, {e, b, e}}
= 2{{e, b, e}, a, e} − {e, {b, e, a}, e}.
Thus 2{{e, b, e}, a, e} − {{e, a, e}, b, e} = {e, {b, e, a}, e} is symmetric on a, b.
Adding the symmetric element {{e, b, e}, a, e} + {{e, a, e}, b, e} we obtain
that {{e, b, e}, a, e} is symmetric on a, b. Then we get:
[τ (a, {e, b, e}) − τ (b, {e, a, e}), E] = −[τ (a, {e, b, e}) − τ (b, {e, a, e}), ι2(e)]
= ι2(cid:0){{e, b, e}, a, e} − {{e, a, e}, b, e}(cid:1) = 0.
Therefore τ (a, {e, b, e}) − τ (b, {e, a, e}) belongs to Centrg(A)(cid:0)sl2(F)(cid:1), and
equation (6.9) gives:
τ (A0, A0) = τ (A0, {e, A0, e}) ⊆ τ (A0, e) + Centrg(A)(cid:0)sl2(F)(cid:1).
Equations (6.5), (6.8) and (6.10) give the desired result.
(6.10)
(cid:3)
Theorem 6.6. Let (A,¯, ∗, {...}) be a dicyclic ternary algebra which contains
an element e ∈ A0 such that {e, e, a} = −2a for any a ∈ A0, and {e, e, x} =
x and {x, e, e} = 0 for any x ∈ A1. Then the attached Lie algebra g(A) in
Theorem 5.3 is a BC1-graded Lie algebra of type C1.
26
ALBERTO ELDUQUE AND SUSUMU OKUBO
Proof. The two previous lemmas show that we have a decomposition:
g(A) =(cid:16)τ (A0, e)⊕ ι1(A0)⊕ ι2(A0)(cid:17)⊕(cid:16)ι1(A1)⊕ ι2(A1)(cid:17) ⊕ Centrg(A)(cid:0)sl2(F)(cid:1),
where
• τ (A0, e) ⊕ ι1(A0) ⊕ ι2(A0) is a direct sum of copies of the adjoint
module for sl2(F) = FH + FE + FF ,
• ι1(A1) ⊕ ι2(A1) is a direct sum of copies of the two-dimensional
irreducible module for sl2(F),
• Centrg(A)(cid:0)sl2(F)(cid:1) is a trivial module for sl2(F),
and this proves the Theorem.
(cid:3)
Therefore, under the hypotheses of this Theorem, the Lie algebra g(A)
can be identified to
where we fix a symplectic basis {u, v} of V , take H, E, F the endomorphisms
of V with coordinate matrices given in (6.1), and identify elements as follows:
(cid:0)sl(V ) ⊗ A0(cid:1) ⊕(cid:0)V ⊗ A1(cid:1) ⊕ d,
H ⊗ a ↔ τ (a, e)
E ⊗ a ↔
1
2
ι2({e, a, e})
F ⊗ a ↔ ι1(a)
u ⊗ x ↔ ι1(x)
v ⊗ x ↔ ι2(e ∗ x)
for any a ∈ A0 and x ∈ A1. Also we take the subalgebra d = Centrg(A)(cid:0)sl2(F)(cid:1) =
Centrτ (A,A)(cid:0)sl2(F)(cid:1).
The Lie bracket in g(A) then induces some operations as in (2.3) and (2.4)
on A0 and A1, which we determine now. We take a, b ∈ A0 and x, y, z ∈ A1.
• [H ⊗ a, F ⊗ b] = −2F ⊗ a · b, but [τ (a, e), ι1(b)] = ι1({a, e, b}). Hence,
a · b = −
1
2
{a, e, b}.
• [H ⊗ a, u ⊗ x] = u ⊗ a • x, and [τ (a, e), ι1(x)] = ι1({a, e, x}), so
a • x = {a, e, x}.
• [u ⊗ x, v ⊗ y] = −H ⊗ hxyi + dx,y, and [ι1(x), ι2(e ∗ y)] = τ (x, e ∗ y).
Since (x ∗ e) ∗ e = −{x, e, e} + {e, e, x} = x (see (D1)), equation
(6.6) gives
τ (x, e ∗ y) = τ (e ∗ (e ∗ x), e ∗ y)
=
1
2
so we get:
τ ((e ∗ x) ∗ (e ∗ y), e) +
1
2(cid:0)τ (x, e ∗ y) + τ (y, e ∗ x)(cid:1),
hxyi = −
1
2
(e ∗ x) ∗ (e ∗ y),
dx,y =
1
2(cid:0)τ (x, e ∗ y) + τ (y, e ∗ x)(cid:1),
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
27
and, because of (2.14),
(x, y, z) =
1
= −
= −
1
1
4(cid:0)−{x, e ∗ y, z} − {y, e ∗ x, z} − {(e ∗ x) ∗ (e ∗ y), e, z}(cid:1)
4(cid:0){x, e ∗ y, z} + {y, e ∗ x, z}
+ {(e ∗ y) ∗ e, e ∗ x, z} + {e ∗ (e ∗ x), e ∗ y, z}(cid:1)
4(cid:0)2{x, e ∗ y, z}(cid:1) = −
{x, e ∗ y, z},
1
2
where (D1) and (D4) have been used.
This proves the following result:
Corollary 6.7. Let (A,¯, ∗, {...}) be a dicyclic ternary algebra which con-
tains an element e ∈ A0 such that {e, e, a} = −2a for any a ∈ A0, and
{e, e, x} = x and {x, e, e} = 0 for any x ∈ A1. Then the pair (A0, A1) is a
J-ternary algebra with operations:
a · b = −
1
2
{a, e, b},
a • x = {a, e, x},
hxyi = −
(x, y, z) = −
1
2
1
2
(e ∗ x) ∗ (e ∗ y),
{x, e ∗ y, z},
for any a, b ∈ A0 and x, y, z ∈ A1.
(cid:3)
Remark 6.8. Had we started with the dicyclic ternary algebra A = J ⊕ T
attached to a J-ternary algebra, with e = 1 ∈ J as in Proposition 6.1, then
the element e would satisfy, besides the conditions {e, e, a} = −2a for any
a ∈ A0 = J and {e, e, x} = x, {x, e, e} = 0 for any x ∈ A1 = T , the extra
condition e ∗ x = −1 • x = −x for any x ∈ T in (6.4), and hence we would
have obtained
a · b = −
1
2
{a, e, b},
a • x = {a, e, x} = −(a ∗ x) ∗ e = −a ∗ x (using (D1)),
hxyi = −
1
2
x ∗ y,
(x, y, z) =
1
2
{x, y, z},
for a, b ∈ J and x, y, z ∈ T , which recover the original operations in J and
T (see (6.3)).
7. (ǫ, δ) Freudenthal-Kantor triple systems and dicyclic
algebras
The results in Section 6 show that for a J-ternary algebra (J, T ), the
Lie algebra g(J, T ) is a Lie algebra with dicyclic symmetry, because the
dicyclic group is contained in the symplectic group Sp(V ) which acts by
In particular (see Section 3) given a special
automorphisms on g(J, T ).
28
ALBERTO ELDUQUE AND SUSUMU OKUBO
(1, 1) Freudenthal-Kantor triple system (U, xyz), the pair (J, U ) is endowed
with a structure of J-ternary algebra (Theorem 3.7), where J is the special
Jordan algebra F1 + K(U, U ), and hence we get that the corresponding Lie
algebra g(J, U ) is a Lie algebra with dicyclic symmetry (that is, there is an
action of the dicyclic group Dic3 by automorphisms on g(J, U )).
In this section, given an arbitrary (ǫ, δ) Freudenthal-Kantor triple system
(U, xyz), a 5-graded Lie algebra, for δ = 1, or Lie superalgebra, for δ = −1,
with dicyclic symmetry will be defined. For a special (1, 1) Freudenthal-
Kantor triple system this algebra is quite close to our previous g(J, U ).
The construction of g(U ) is based on the pioneering work of Yamaguti
and Ono [YO84].
Thus, let (U, xyz) be a (ǫ, δ) Freudenthal-Kantor triple system, then the
space of 2 × 1 matrices over U :
T =(cid:26)(cid:18)x
y(cid:19) : x, y ∈ U(cid:27)
becomes a Lie triple system for δ = 1 and an anti-Lie triple system for
δ = −1 (see [YO84, Section 3]) by means of the triple product:
b2(cid:19)(cid:18)a3
b3(cid:19)(cid:21)
(cid:20)(cid:18)a1
b1(cid:19)(cid:18)a2
=(cid:18)L(a1, b2) − δL(a2, b1)
−ǫK(b1, b2)
and, therefore, the vector space
δK(a1, a2)
ǫL(b2, a1) − ǫδL(b1, a2)(cid:19)(cid:18)a3
b3(cid:19)
L = span(cid:26)(cid:18) L(a, b) K(c, d)
ǫL(b, a)(cid:19) : a, b, c, d, e, f ∈ U(cid:27)
K(e, f )
(7.1)
(7.2)
is a Lie subalgebra of Mat2(cid:0)EndF(U )(cid:1)−. (Given an associative algebra A,
A− denotes the Lie algebra defined on A with product given by the usual
Lie bracket [x, y] = xy − yx.)
Hence we get either a Z2-graded Lie algebra (for δ = 1) or a superalgebra
(for δ = −1)
g(U ) = L ⊕ T
(7.3)
where L is the even part and T the odd part. The bracket in g(U ) is given
by:
• the given bracket in L as a subalgebra of Mat2(cid:0)EndF(U )(cid:1)−,
• [M, X] = M (X) for any M ∈ L and X ∈ T (note that Mat2(cid:0)EndF(U )(cid:1) ≃
EndF(T )),
• for any a1, a2, b1, b2 ∈ U :
(cid:20)(cid:18)a1
b1(cid:19) ,(cid:18)a2
b2(cid:19)(cid:21) =(cid:18)L(a1, b2) − δL(a2, b1)
−ǫK(b1, b2)
δK(a1, a2)
ǫL(b2, a1) − ǫδL(b1, a2)(cid:19) .
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
29
This (super)algebra g(U ) is consistently Z-graded as follows:
0
ǫL(b, a)(cid:19) : a, b ∈ U(cid:27) ,
0
g(U )(0) = span(cid:26)(cid:18)L(a, b)
g(U )(1) =(cid:18)U
0(cid:19) ,
g(U )(−1) =(cid:18) 0
U(cid:19) ,
g(U )(2) = span(cid:26)(cid:18)0 K(a, b)
g(U )(−2) = span(cid:26)(cid:18) 0
0
0 (cid:19) : a, b ∈ U(cid:27) ,
K(a, b) 0(cid:19) : a, b ∈ U(cid:27) ,
0
so that g(U ) is 5-graded and
L = g(U )(−2) ⊕ g(U )(0) ⊕ g(U )(2),
T = g(U )(−1) ⊕ g(U )(1).
On the other hand, if U is a special (ǫ, ǫ) Freudenthal-Kantor triple sys-
tem, Corollaries 3.8 and 4.6 give another construction of a Lie (super)algebra
g(J, U ) attached to U (Lie algebra for ǫ = 1 and superalgebra for ǫ = −1),
where J = F1 + K(U, U ):
with d = S(U, U ) (the linear span of the operators S(x, y) defined in (3.3)).
g(J, U ) =(cid:0)sl(V ) ⊗ J(cid:1) ⊕(cid:0)V ⊗ U(cid:1) ⊕ d,
Proposition 7.1. Let (U, xyz) be a special (ǫ, ǫ) Freudenthal-Kantor triple
system, and let {u, v} be a fixed symplectic basis of the two-dimensional
vector space V (so that (uv) = 1 for a fixed nonzero skew symmetric bilinear
map (..)). Then the linear map g(U ) → g(J, U ) given by:
(cid:18)L(a, b)
0
0
(cid:18)a
b(cid:19) ∈ T 7→ u ⊗ a + v ⊗ b ∈ V ⊗ U,
ǫL(b, a)(cid:19) 7→ −ǫγu,v ⊗ K(a, b) − S(a, b) ∈(cid:0)sl(V ) ⊗ J(cid:1) ⊕ d,
(cid:18)0 K(a, b)
0 (cid:19) 7→ −γu,u ⊗ K(a, b) ∈ sl(V ) ⊗ J,
(cid:18) 0
K(a, b) 0(cid:19) 7→ γv,v ⊗ K(a, b) ∈ sl(V ) ⊗ J,
0
0
for any a, b ∈ U , is a one-to-one Lie (super)algebra homomorphism.
Proof. This is done by straightforward computations.
(cid:3)
Under this homomorphism g(U )(1) (respectively g(U )(−1)) is identified to
u ⊗ U (respectively v ⊗ U ).
In particular, for a special (ǫ, ǫ) Freudenthal-Kantor triple system, the Lie
(super)algebra g(U ) inherits the dicyclic symmetry from g(J, U ) considered
in Section 3 (assuming the primitive cubic roots ω and ω2 of 1 are contained
in the ground field F).
Our next purpose is to check that for any arbitrary (ǫ, δ) Freudenthal-
Kantor triple system (U, xyz) the Lie (super)algebra g(U ) presents dicyclic
30
ALBERTO ELDUQUE AND SUSUMU OKUBO
symmetry (assuming ω ∈ F), which restricts to the previous one for the
special (ǫ, ǫ) Freudenthal-Kantor triple systems.
First, the fact that g(U ) is Z-graded gives a group homomorphism:
F× −→ Aut(cid:0)g(U )(cid:1)
µ 7→ φµ : X 7→ µiX, for X ∈ g(U )(i), −2 ≤ i ≤ 2.
Also, the map
θ : T −→ T
(cid:18)a
b(cid:19) 7→ (cid:18)−ǫb
δa (cid:19) for a, b ∈ U ,
is an automorphism of the (anti)Lie triple system T , and hence extends
naturally to an automorphism, also denoted by θ, of g(U ) = L ⊕ T , with
θ(M ) = θM θ−1 for any M ∈ L ⊆ EndF(T ).
Note that θ2 = 1 for ǫ = −δ, while θ4 = 1 6= θ2 for ǫ = δ. For ǫ = δ, θ2 is
−1 on T and 1 on L, so it coincides with the Z2-grading automorphism of
g(U ). A few computations give precise formulas for the action of θ:
θ(cid:18)L(a, b)
0
0
0
θ(cid:18)0 K(a, b)
ǫL(b, a)(cid:19) =(cid:20)θ(cid:18)a
0(cid:19) , θ(cid:18)0
b(cid:19)(cid:21)
0(cid:19)(cid:21) =(cid:18)ǫL(b, a)
a(cid:19) ,(cid:18)b
= −ǫδ(cid:20)(cid:18)0
0 (cid:19) = δ(cid:20)θ(cid:18)a
0(cid:19) , θ(cid:18)b
0(cid:19)(cid:21)
= δ(cid:20)(cid:18)0
a(cid:19) ,(cid:18)0
b(cid:19)(cid:21) = −ǫδ(cid:18) 0
b(cid:19)(cid:21)
a(cid:19) , θ(cid:18)0
K(a, b) 0(cid:19) = −ǫ(cid:20)θ(cid:18)0
θ(cid:18) 0
= −ǫ(cid:20)(cid:18)a
0(cid:19) ,(cid:18)b
0(cid:19)(cid:21) = −ǫδ(cid:18)0 K(a, b)
0 (cid:19) .
K(a, b) 0(cid:19) ,
0
0
0
0
L(a, b)(cid:19) ,
0
Assuming ω ∈ F, the automorphisms φ = φω and θ satisfy (5.2). The
action of φ induces a Z3-grading: g(U ) = g(U )1 ⊕ g(U )ω ⊕ g(U )ω2, with
g(U )1 = g(U )(0), g(U )ω = g(U )(−2) ⊕ g(U )(1) ≃ K(U, U ) ⊕ U , and g(U )ω2 =
g(U )(2) ⊕ g(U )(−1) ≃ K(U, U ) ⊕ U
Remark 7.2. Actually, for δ = −ǫ we get θ2 = 1, and hence φ and θ
generate a subgroup of automorphisms of g(U ) isomorphic to the symmetric
group S3.
In particular, for δ = −ǫ = 1, g(U ) is a Lie algebra with S3-
symmetry, and this defines a structure of generalized Malcev algebra (see
[EO09]) on gω ≃ K(U, U ) ⊕ U . This extends [EO09, Proposition 4.1].
For ǫ = δ = 1, g(U ) becomes a Lie algebra with dicyclic symmetry, and
this shows that A = g(U )ω ≃ K(U, U ) ⊕ U becomes a dicyclic ternary
algebra (see Section 5), with A0 = g(U )(−2) ≃ K(U, U ) (fixed by θ2) and
A1 = g(U )(1) ≃ U . Identify U with A1 = g(U )(1) and K(U, U ) with A0 =
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
31
g(U )(−2) in the natural way:
u ↔(cid:18)u
0(cid:19) ,
K(a, b) ↔(cid:18) 0
K(a, b) 0(cid:19) .
0
After these identifications, we can compute the binary and ternary prod-
ucts in (5.4).
First note that A0∗A0 = 0 as [g(U )(−2), g(U )(−2)] = 0. Also, for a, b, x, y ∈
U we have:
0
θ−1(cid:18)(cid:20)(cid:18) 0
θ−1(cid:18)(cid:20)(cid:18)x
K(a, b) 0(cid:19) ,(cid:18)0
0(cid:19) ,(cid:18)y
x(cid:19)(cid:21)(cid:19) = θ−1(cid:18)
0(cid:19)(cid:21)(cid:19) = θ−1(cid:18)0 K(x, y)
K(a, b)x(cid:19) =(cid:18)K(a, b)x
(cid:19) ,
0 (cid:19) = −(cid:18) 0
K(x, y) 0(cid:19) .
0
0
0
0
Therefore, the binary product in A = K(U, U ) ⊕ U is given by:
M1 ∗ M2 = 0,
M ∗ x = M x,
x1 ∗ x2 = −K(x1, x2),
(7.4)
for any M, M1, M2 ∈ K(U, U ) and x, x1, x2 ∈ U .
On the other hand we have:
[g(U )(1), θ(cid:0)g(U )(−2)(cid:1)] = [g(U )(1), g(U )(2)] = 0 = [g(U )(−2), θ(cid:0)g(U )(1)(cid:1)],
so we get the following instances of the triple product:
{A0, A1, A} = 0 = {A1, A0, A}.
Now, for any a1, a2, b1, b2 ∈ U , we get
(cid:20)(cid:18)
0
0
0
0
K(a2, b2) 0(cid:19)(cid:21)
K(a1, b1) 0(cid:19) ,(cid:18)0 K(a2, b2)
K(a1, b1) 0(cid:19) , θ(cid:18)
= −(cid:20)(cid:18)
=(cid:18)K(a2, b2)K(a1, b1)
0
0
0
0
0
(cid:19)(cid:21)
0
−K(a1, b1)K(a2, b2)(cid:19) .
The fact that this last matrix belongs to g(U )(0) is a consequence of the next
result (see [YO84, (2.9) and (2.10)]).
Lemma 7.3. Let (U, xyz) be an (ǫ, δ) Freudenthal-Kantor triple system,
then for any a, b, c, d ∈ U the following identities hold:
ǫK(a, b)K(c, d) + L(cid:0)K(a, b)c, d(cid:1) − δL(cid:0)K(a, b)d, c(cid:1) = 0,
K(c, d)K(a, b) + δL(cid:0)c, K(a, b)d(cid:1) − L(cid:0)d, K(a, b)c(cid:1) = 0.
Proof. Equation (3.2b) gives
(7.5a)
(7.5b)
K(cid:0)K(u, v)x, y(cid:1)z − δK(cid:0)K(u, v)z, y(cid:1)x
But (3.1) gives K(cid:0)K(u, v)x, y(cid:1)z + δL(y, z)K(u, v)x = L(cid:0)K(u, v)x, z)y and
a similar relation holds interchanging x and z, whence (7.5a).
= L(y, x)K(u, v)z − δL(y, z)K(u, v)x − ǫK(u, v)K(x, z)y,
32
ALBERTO ELDUQUE AND SUSUMU OKUBO
On the other hand, equation (3.2a) gives:
[L(u, v), L(x, y)] − δ[L(u, y), L(x, v)]
= L(cid:0)L(u, v)x, y(cid:1) − δL(cid:0)L(u, y)x, v(cid:1) + ǫL(cid:0)x, K(v, y)u(cid:1).
Interchange u and x to get
[L(x, v), L(u, y)] − δ[L(x, y), L(u, v)]
= L(cid:0)L(x, v)u, y(cid:1) − δL(cid:0)L(x, y)u, v(cid:1) + ǫL(cid:0)u, K(v, y)x(cid:1),
and this gives
δL(cid:0)L(u, v)x, y(cid:1) − L(cid:0)L(u, y)x, v(cid:1) + ǫδL(cid:0)x, K(v, y)u(cid:1)
= L(cid:0)L(x, v)u, y(cid:1) − δL(cid:0)L(x, y)u, v(cid:1) + ǫL(cid:0)u, K(v, y)x(cid:1),
or
δL(cid:0)x, K(v, y)u(cid:1) − L(cid:0)u, K(v, y)x(cid:1) = ǫL(cid:0)K(x, u)v, y(cid:1) + ǫL(cid:0)K(u, x)y, v(cid:1),
whose right hand side equals −K(x, u)K(v, y), due to (7.5a), thus getting
(7.5b).
(cid:3)
Coming back to the case ǫ = δ = 1, for a1, a2, a3, b1, b2, b3, x ∈ U we have
0
0
K(a2, b2) 0(cid:19)(cid:21) ,(cid:18)
0
0
K(a3, b3) 0(cid:19)(cid:21)
0
−K(a1, b1)K(a2, b2)K(a3, b3) − K(a3, b3)K(a2, b2)K(a1, b1)
0
0
K(a1, b1) 0(cid:19) , θ(cid:18)
(cid:20)(cid:20)(cid:18)
=(cid:18)
(cid:20)(cid:20)(cid:18)
K(a1, b1) 0(cid:19) , θ(cid:18)
0
0
0
0 (cid:19) ,
(cid:19) .
0
0
K(a2, b2) 0(cid:19)(cid:21) ,(cid:18)x
0(cid:19)(cid:21) =(cid:18)K(a2, b2)K(a1, b1)x
0
Therefore, we get the following instances of the triple product on A =
K(U, U ) ⊕ U :
{M1, M2, M3} = −(cid:0)M1M2M3 + M3M2M1(cid:1),
{M1, M2, x} = M2M1x,
for any M1, M2, M3 ∈ K(U, U ) and x ∈ U . (Note that the triple product
{M1, M2, M3} is, up to the sign, the usual Jordan triple product in the
associative algebra EndF(U ).)
Finally, for any x1, x2 ∈ U :
(cid:20)(cid:18)x1
0(cid:19) , θ(cid:18)x2
0(cid:19)(cid:21) =(cid:20)(cid:18)x1
0(cid:19) ,(cid:18) 0
x2(cid:19)(cid:21) =(cid:18)L(x1, x2)
0
0
ǫL(x2, x1)(cid:19) ,
SPECIAL FREUDENTHAL-KANTOR TRIPLE SYSTEMS
33
so that we obtain, for any x1, x2, x3, a, b ∈ U :
0
0
0(cid:19)(cid:21) ,(cid:18) 0
K(a, b) 0(cid:19)(cid:21)
=(cid:18)
L(x2, x1)K(a, b) − K(a, b)L(x1, x2) 0(cid:19)
=(cid:18)
K(cid:0)K(a, b)x1, x2(cid:1) 0(cid:19) (see (3.2b))
0(cid:19)(cid:21) ,(cid:18)x3
0(cid:19)(cid:21) =(cid:18)L(x1, x2)x3
(cid:19) =(cid:18)x1x2x3
0 (cid:19) .
0
0
0
0
(cid:20)(cid:20)(cid:18)x1
0(cid:19) , θ(cid:18)x2
(cid:20)(cid:20)(cid:18)x1
0(cid:19) , θ(cid:18)x2
Hence we get the following instances of the triple product on A = K(U, U )⊕
U :
{x1, x2, M } = K(cid:0)M x1, x2(cid:1),
{x1, x2, x3} = x1x2x3,
for any x1, x2, x3 ∈ U and M ∈ K(U, U ).
Let us summarize the previous arguments in our last result, which at-
taches a dicyclic ternary algebra to an arbitrary (1, 1) Freudenthal-Kantor
triple system:
Theorem 7.4. Let (U, xyz) be a (1, 1) Freudenthal-Kantor triple system.
Then (cid:0)K(U, U ) ⊕ U,¯, ∗, {., ., .}(cid:1) is a dicyclic ternary algebra, where
¯M = M,
¯x = −x,
M1 ∗ M2 = 0, M ∗ x = M x,
x1 ∗ x2 = −K(x1, x2),
{M1, x1, M2} = {M1, x1, x2} = {x1, M1, M2} = {x1, M1, x2} = 0,
{M1, M2, M3} = −(cid:0)M1M2M3 + M3M2M1(cid:1),
{x1, x2, M } = K(cid:0)M x1, x2(cid:1),
for any x, x1, x2, x3 ∈ U and M, M1, M2, M3 ∈ K(U, U ).
{x1, x2, x3} = x1x2x3,
{M1, M2, x} = M2M1x,
(cid:3)
References
[All76]
[All78]
[ABG]
[BS03]
[Eld07]
[EO07]
[EO08]
[EO09]
B.N. Allison, A construction of Lie algebras from J -ternary algebras, Amer.
J. Math. 98 (1976), no. 2, 285 -- 294.
A class of nonassociative algebras with involution containing the class
of Jordan algebras, Math. Ann. 237 (1978), no. 2, 133-156.
B.N. Allison, G. Benkart, and Y. Gao, Lie Algebras Graded by the Root System
BCr, r ≥ 2, Mem. Amer. Math. Soc. 158 (2002), no. 751, x+158 pp.
G. Benkart and O. Smirnov, Lie Algebras Graded by the Root System BC1, J.
Lie Theory 13 (2003), no. 1, 91 -- 132.
A. Elduque, The S4-action on the tetrahedron algebra, Proc. Royal Soc. Edin-
burgh 137A (2007), no. 6, 1227 -- 1248.
A. Elduque and S. Okubo, Lie algebras with S4-action and structurable alge-
bras, J. Algebra 307 (2007), no. 2, 864 -- 890.
S4-symmetry on the Tits construction of exceptional Lie algebras and
superalgebras, Publ. Mat. 52 (2008), no. 2, 315 -- 346.
Lie algebras with S3 or S4-action and generalized Malcev algebras, Proc.
Royal Soc. Edinburgh 139A (2009), no. 2, 321 -- 354.
34
[Hei75]
[Jac67]
[Kan73]
[Koc09]
[Tits62]
[YO84]
ALBERTO ELDUQUE AND SUSUMU OKUBO
W. Hein, A construction of Lie algebras by triple systems, Trans. Amer. Math.
Soc. 205 (1975), 79 -- 95.
N. Jacobson Structure and representations of Jordan algebras, American Math-
ematical Society Colloquium Publications, Vol. XXXIX, American Mathemat-
ical Society, Providence, R.I. 1968 x+453 pp.
I.L. Kantor, Models of exceptional simple Lie algebras, Dokl. Akad. Nauk SSSR
208 (1973), 1276 -- 1279.
M. Kochetov, Gradings on finite-dimensional simple Lie algebras, Acta Appl.
Math. 108 (2009), no. 1, 101 -- 127.
J. Tits, Une classe d'alg`ebres de Lie en relation avec les alg`ebres de Jordan,
Nederl. Akad. Wetensch. Proc. Ser. A 65 = Indag. Math. 24 (1962), 530 -- 535.
K. Yamaguti and A. Ono, On representations of Freudenthal-Kantor triple
systems U (ǫ, δ), Bull. Fac. School Ed. Hiroshima Univ., Part II, 7 (1984),
43 -- 51.
Departamento de Matem´aticas e Instituto Universitario de Matem´aticas y
Aplicaciones, Universidad de Zaragoza, 50009 Zaragoza, Spain
E-mail address: [email protected]
Department of Physics and Astronomy, University of Rochester, Rochester,
NY 14627, USA
E-mail address: [email protected]
|
1703.00718 | 2 | 1703 | 2017-05-02T08:36:09 | The automorphisms of Petit's algebras | [
"math.RA"
] | Let $\sigma$ be an automorphism of a field $K$ with fixed field $F$. We study the automorphisms of nonassociative unital algebras which are canonical generalizations of the associative quotient algebras $K[t;\sigma]/fK[t;\sigma]$ obtained when the twisted polynomial $f\in K[t;\sigma]$ is invariant, and were first defined by Petit. We compute all their automorphisms if $\sigma$ commutes with all automorphisms in ${\rm Aut}_F(K)$ and $n\geq m-1$, where $n$ is the order of $\sigma$ and $m$ the degree of $f$,and obtain partial results for $n<m-1$. In the case where $K/F$ is a finite Galois field extension, we obtain more detailed information on the structure of the automorphism groups of these nonassociative unital algebras over $F$. We also briefly investigate when two such algebras are isomorphic. | math.RA | math |
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
C. BROWN AND S. PUMPL UN
Abstract. Let σ be an automorphism of a field K with fixed field F . We study the
automorphisms of nonassociative unital algebras which are canonical generalizations of
the associative quotient algebras K[t; σ]/f K[t; σ] obtained when the twisted polynomial
f ∈ K[t; σ] is invariant, and were first defined by Petit. We compute all their automor-
phisms if σ commutes with all automorphisms in AutF (K) and n ≥ m − 1, where n is
the order of σ and m the degree of f , and obtain partial results for n < m − 1. In the
case where K/F is a finite Galois field extension, we obtain more detailed information
on the structure of the automorphism groups of these nonassociative unital algebras over
F . We also briefly investigate when two such algebras are isomorphic.
Introduction
Let D be a division algebra, σ an injective endomorphism of D, δ a left σ-derivation and
R = D[t; σ, δ] a skew polynomial ring (for instance, c.f. [16, § 3.4]). For an invariant skew
polynomial f ∈ R, i.e. when the ideal Rf is a two-sided principal ideal, the quotient alge-
bra R/Rf appears in classical constructions of associative central simple algebras, usually
employing an irreducible f ∈ R to get examples of division algebras, e.g. see [15].
In 1967, Petit [22, 23] introduced a class of unital nonassociative algebras Sf , which
canonically generalize the quotient algebras R/Rf obtained when factoring out an invariant
f ∈ R of degree m. The algebra Sf = D[t; σ, δ]/D[t; σ, δ]f is defined on the additive subgroup
{h ∈ R deg(h) < m} of R by using right division by f to define the algebra multiplication
g ◦ h = gh modrf . The properties of the algebras Sf were studied in detail in [22, 23], and
for D a finite base field (hence w.l.o.g. δ = 0) in [20].
Even earlier, the algebra Sf with f (t) = t2 − i ∈ C[t;
appeared in [8] as the first example of a nonassociative division algebra.
],
the complex conjugation,
Although the algebras themselves have received little attention so far, the right nucleus
of Sf (the eigenspace of f ∈ R) already appeared implicitly in classical constructions by
Amitsur [2, 3, 4], but also in results on computational aspects of operator algebras; they
are for instance used in algorithms factoring skew polynomials over Fq(t) or finite fields, cf.
[11, 12, 13, 14]. The role of classical algebraic constructions in coding theory is well known
(cf. [16, Chapter 9], [17, 1, 7]).
Moreover, recently space-time block codes, coset codes and wire-tap codes were obtained
[9, 10, 21, 24, 27, 28, 29], and they also
employing the algebras Sf over number fields, cf.
appear useful for linear cyclic codes [25, 26].
2010 Mathematics Subject Classification. Primary: 17A35; Secondary: 17A60, 17A36, 16S36.
Key words and phrases. Skew polynomial ring, skew polynomials, Ore polynomials, automorphisms,
nonassociative algebras.
1
2
C. BROWN AND S. PUMPL UN
If K is a finite field, F the fixed field of σ, K/F a finite Galois field extension and
f ∈ K[t; σ] = K[t; σ, 0] irreducible and invariant, the Sf are Jha-Johnson semifields (also
called cyclic semifields) [20, Theorem 15], and were studied for instance by Wene [35] and
more recently by Lavrauw and Sheekey [20]. The main motivation for our paper comes from
the question how the automorphism groups of Jha-Johnson semifields look like. The results
presented here are applied to some Jha-Johnson semifields in [6].
The structure of this paper is as follows: In Section 1, we introduce the terminology and
define the algebras Sf . We limit our observations to the algebras which are not associative.
Given a field extension K, σ ∈ Aut(K) of order n with fixed field F , such that σ com-
mutes with all τ ∈ AutF (K), and f ∈ K[t; σ] of degree m not invariant, we compute the
automorphisms of Sf in Section 2. We obtain all automorphisms for n ≥ m − 1 and some
partial results for n < m − 1 (Theorems 4 and 5). For n ≥ m − 1, the automorphisms in
AutF (Sf ) are canonically induced by the F -automorphism G of R = K[t; σ] which satisfy
G(f (t)) = af (t) for some a ∈ K ×, and on K restrict to an automorphism that commutes
with τ .
The automorphisms groups of Sf where f (t) = tm − a ∈ K[t; σ], a ∈ K \ F , play a special
i=0 biti ∈ K[t; σ] and b0 = a, AutF (Sg)
role, as for all nonassociative Sg with g(t) = tm −Pm−1
is a subgroup of AutF (Sf ) when n ≥ m − 1.
We then focus on the situation that K/F is a finite Galois field extension such that σ
commutes with all τ ∈ Gal(K/F ). In many cases, either AutF (Sf ) ∼= Gal(K/F ) or is trivial
(Theorem 10). Necessary conditions for extending Galois automorphisms τ ∈ Gal(K/F )
to Sf are studied in Sections 3 and 4. The existence of cyclic subgroups of AutF (Sf ) is
investigated in Section 5.
For f (t) = tm − a ∈ K[t; σ] and K/F a cyclic field extension of degree m, the algebra
Sf is also called a nonassociative cyclic algebra and denoted by (K/F, σ, a). These algebras
are canonical generalizations of associative cyclic algebras, but also generalizations of the
algebras in [3, 15]. The automorphisms of nonassociative cyclic algebras are investigated
in Section 6. All the automorphisms of A = (K/F, σ, a) extending idK are inner and
form a cyclic subgroup of AutF (A) isomorphic to ker(NK/F ).
In some cases, this is the
whole automorphism group, e.g.
In these cases, every
automorphism of A leaves K fixed and is inner. We explain when the automorphism group
of a nonassociative quaternion algebra A (where m = 2) contains a dicyclic group and when
it contains a subgroup isomorphic to the semidirect product of two cyclic groups.
if F has no mth root of unity.
In Section 7 we briefly investigate isomorphisms between two algebras Sf and Sg.
This work is part of the first author's PhD thesis [5] written under the supervision of
the second author. For results on the automorphisms of the more general algebras defined
using f ∈ D[t; σ], or a more detailed study and the (less relevant) cases left out in this
paper the reader is referred to [5]. For examples of applications of the associated classical
constructions the readers are referred to [1, 7, 16, 17].
1. Preliminaries
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
3
1.1. Nonassociative algebras. Let F be a field and let A be an F -vector space. A is an
algebra over F if there exists an F -bilinear map A × A → A, (x, y) 7→ x · y, denoted simply
by juxtaposition xy, the multiplication of A. An algebra A is called unital if there is an
element in A, denoted by 1, such that 1x = x1 = x for all x ∈ A. We will only consider
unital algebras from now on without explicitly saying so.
Associativity in A is measured by the associator [x, y, z] = (xy)z − x(yz). The left
nucleus of A is defined as Nucl(A) = {x ∈ A [x, A, A] = 0}, the middle nucleus of A is
Nucm(A) = {x ∈ A [A, x, A] = 0} and the right nucleus of A is defined as Nucr(A) = {x ∈
A [A, A, x] = 0}. Nucl(A), Nucm(A), and Nucr(A) are associative subalgebras of A. Their
intersection Nuc(A) = {x ∈ A [x, A, A] = [A, x, A] = [A, A, x] = 0} is the nucleus of A.
Nuc(A) is an associative subalgebra of A containing F 1 and x(yz) = (xy)z whenever one of
the elements x, y, z lies in Nuc(A). The center of A is C(A) = {x ∈ A x ∈ Nuc(A) and xy =
yx for all y ∈ A}.
An F -algebra A 6= 0 is called a division algebra if for any a ∈ A, a 6= 0, the left
multiplication with a, La(x) = ax, and the right multiplication with a, Ra(x) = xa, are
bijective. If A has finite dimension over F , then A is a division algebra if and only if A
has no zero divisors [31, pp. 15, 16]. An element 0 6= a ∈ A has a left inverse al ∈ A, if
Ra(al) = ala = 1, and a right inverse ar ∈ A, if La(ar) = aar = 1.
An automorphism G ∈ AutF (A) is an inner automorphism if there is an element m ∈ A
with left inverse ml such that G(x) = (mlx)m for all x ∈ A. Given an inner automorphism
Gm ∈ AutF (A) and some H ∈ AutF (A), then clearly H −1 ◦ Gm ◦ H ∈ AutF (A) is an inner
automorphism. [34, Lemma 2, Theorem 3, 4] generalize to any nonassociative algebra:
Proposition 1. Let A be an algebra over F .
(i) For all invertible n ∈ Nuc(A), Gn(x) = (n−1x)n is an inner automorphism of A.
(ii) If Gm is an inner automorphism of A, then so is Gnm(x) = ((mln−1)x)(nm) for all
invertible n ∈ Nuc(A).
(iii) If Gm is an inner automorphism of A, and a, b ∈ Nuc(A) are invertible, then Gam =
Gbm if and only if ab−1 ∈ C(A).
(iv) For invertible n, m ∈ Nuc(A), Gm = Gn if and only if n−1m ∈ C(A).
The set {Gm m ∈ Nuc(A) invertible} is a subgroup of AutF (A). For each invertible
m ∈ Nuc(A) \ C(A), Gm generates a cyclic subgroup which has finite order s if ms ∈ C(A),
so in particular if m has order s.
Note that if the nucleus is commutative, then for all invertible n ∈ Nuc(A), Gn(x) =
(n−1x)n is an inner automorphism of A such that GnNuc(A) = idNuc(A).
1.2. Twisted polynomial rings. Let K be a field and σ an automorphism of K. The
twisted polynomial ring K[t; σ] is the set of polynomials a0 + a1t + ··· + antn with ai ∈ K,
where addition is defined term-wise and multiplication by ta = σ(a)t for all a ∈ K. For
f = a0 + a1t + ··· + antn with an 6= 0 define deg(f ) = n and put deg(0) = −∞. Then
deg(f g) = deg(f ) + deg(g). An element f ∈ R is irreducible in R if it is not a unit and it
has no proper factors, i.e. if there do not exist g, h ∈ R with deg(g), deg(h) < deg(f ) such
that f = gh.
4
C. BROWN AND S. PUMPL UN
R = K[t; σ] is a left and right principal ideal domain and there is a right division algorithm
in R: for all g, f ∈ R, g 6= 0, there exist unique r, q ∈ R with deg(r) < deg(f ), such that
g = qf + r. There is also a left division algorithm in R [15, p. 3 and Prop. 1.1.14]. (Our
terminology is the one used by Petit [22] and Lavrauw and Sheekey [20]; it is different from
Jacobson's, who calls what we call right a left division algorithm and vice versa.) Define
F = Fix(σ).
1.3. Nonassociative algebras obtained from skew polynomial rings. Let K be a
field, σ an automorphism of K with F = Fix(σ), and f ∈ R = K[t; σ] of degree m. Let
modrf denote the remainder of right division by f . Then then additive abelian group
Rm = {g ∈ K[t; σ] deg(g) < m} together with the multiplication g ◦ h = gh modrf is a
unital nonassociative algebra Sf = (Rm,◦) over F0 = {a ∈ K ah = ha for all h ∈ Sf}. F0
is a subfield of K [22, (7)] and it is straightforward to see that if f (t) = tm −Pm−1
i=0 aiti and
a0 6= 0 then F0 = F [25, Remark 9]. The algebra Sf is also denoted by R/Rf [22, 23] if we
want to make clear which ring R is involved in the construction. In the following, we call
the algebras Sf Petit algebras and denote their multiplication simply by juxtaposition.
Using left division by f and the remainder modlf of left division by f instead, we can
analogously define the multiplication for another unital nonassociative algebra on Rm over
F0, called f S. We will only consider the Petit algebras Sf , since every algebra f S is the
opposite algebra of some Petit algebra [22, (1)].
Theorem 2. (cf. [22, (2), (5), (9)]) Let f (t) ∈ R = K[t; σ].
(i) If Sf is not associative then Nucl(Sf ) = Nucm(Sf ) = K and Nucr(Sf ) = {g ∈ R f g ∈
Rf}.
(ii) The powers of t are associative if and only if tmt = ttm if and only if t ∈ Nucr(Sf ) if
and only if f t ∈ Rf.
(iii) Let f ∈ R be irreducible and Sf a finite-dimensional F -vector space or free of finite
rank as a right Nucr(Sf )-module. Then Sf is a division algebra.
Conversely, if Sf is a division algebra then f is irreducible.
(iv) Sf is associative if and only if f is invariant. In that case, Sf is the usual quotient
algebra.
(v) Let f (t) = tm − Pm−1
aiσi(z) for all z ∈ K, i ∈ {0, . . . , m − 1} and ai ∈ F for all i ∈ {0, . . . , m − 1}.
i=0 aiti ∈ R = K[t; σ]. Then f is invariant if and only if σm(z)ai =
Note that if f is not invariant, then the nucleus of any Sf = K[t; σ]/K[t; σ]f is a subfield
of K = Nucl(Sf ). If Nuc(Sf ) is larger than F , then {Gm 0 6= m ∈ Nuc(A)} is a non-trivial
subgroup of AutF (Sf ) and each inner automorphism Gm in this subgroup extends idNuc(A)
by Proposition 1.
Proposition 3. Let f (t) ∈ F [t] = F [t; σ] ⊂ K[t; σ].
(i) F [t]/(f (t)) is a commutative subring of Sf and F [t]/(f (t)) ∼= F ⊕ F t ⊕ ··· ⊕ F tm−1 ⊂
Nucr(Sf ). In particular, then f t ∈ Rf which is equivalent to the powers of t being associative,
which again is equivalent to tmt = ttm.
(ii) If f (t) is irreducible in F [t], F [t]/(f (t)) is an algebraic subfield of degree m contained
in the right nucleus.
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
5
Proof. Sf contains the commutative subring F [t]/(f (t)). If f (t) is irreducible in F [t], this
is an algebraic field extension of F . This subring is isomorphic to the ring consisting of the
i=0 aiti with ai ∈ F .
elements Pm−1
Clearly F ⊂ Nucr(Sf ). For all a, b, c ∈ K, i, j ∈ {0, . . . , m − 1} we have [ati, btj, t] =
(aσi(b)ti+j )t − (ati)(btj+1) = aσi(b)ti+j+1 − aσi(bc)ti+j = 0. Thus t ∈ Nucr(Sf ) which
implies that F ⊕ F t⊕···⊕ F tm−1 ⊂ Nucr(Sf ), hence the assertion. The rest is obvious. (cid:3)
We will assume throughout the paper that deg(f ) = m ≥ 2 (since if f is constant then
Sf ∼= K) and that σ 6= id. Without loss of generality, we only consider monic polynomials
f , since Sf = Saf for all non-zero a ∈ K.
2. Automorphisms of Sf
2.1. Let K be a field, σ an automorphism of K of order n (which may be infinite), F =
Fix(σ), and f (t) = tm − Pm−1
i=0 aiti ∈ K[t; σ] a twisted polynomial which is not invariant.
Theorem 4. Suppose σ commutes with all τ ∈ AutF (K). Let n ≥ m − 1. Then H is an
automorphism of Sf if and only if H = Hτ,k with
m−1
Hτ,k(
X
i=0
xiti) = τ (x0) + τ (x1)kt + τ (x2)kσ(k)t2 + ··· + τ (xm−1)kσ(k)··· σm−2(k)tm−1,
where τ ∈ AutF (K) and k ∈ K × is such that
τ (ai) = (cid:16)
(1)
m−1
Y
l=i
σl(k)(cid:17)ai
for all i ∈ {0, . . . , m − 1}.
Proof. Let H : Sf → Sf be an automorphism. Since Sf is not associative, Nucl(Sf ) = K
by Theorem 2 (i). Since any automorphism preserves the left nucleus, H(K) = K and so
HK = τ for some τ ∈ AutF (K). Suppose H(t) = Pm−1
i=0 kiti for some ki ∈ K. Then we
have
(2)
and
(3)
H(tz) = H(t)H(z) = (
kiti)τ (z) =
X
i=0
m−1
m−1
X
i=0
kiσi(τ (z))ti
H(tz) = H(σ(z)t) =
m−1
X
i=0
τ (σ(z))kiti
for all z ∈ K. Comparing the coefficients of ti in (2) and (3) we obtain
kiσi(τ (z)) = kiτ (σi(z)) = τ (σ(z))ki = kiτ (σ(z))
(4)
for all i ∈ {0, . . . , m − 1} and all z ∈ K. This implies ki(τ(cid:0)σi(z) − σ(z)(cid:1)) = 0 for all
i ∈ {0, . . . , m − 1} and all z ∈ K since σ and τ commute, i.e.
(5)
ki = 0 or σi(z) = σ(z)
for all i ∈ {0, . . . , m − 1} and all z ∈ K.
6
C. BROWN AND S. PUMPL UN
Since σ has order n ≥ m − 1, which means σi 6= σ for all i ∈ {0, . . . , m − 1}, i 6= 1, (5)
implies ki = 0 for all i ∈ {0, . . . , m − 1}, i 6= 1. Therefore H(t) = kt for some k ∈ K ×.
Furthermore, we have H(zti) = H(z)H(t)i = τ (z)(kt)i = τ (z)(cid:16)Qi−1
l=0 σl(k)(cid:17)ti for all i ∈
{1, . . . , m − 1} and z ∈ K. Thus H has the form
xiti) = τ (x0) +
σl(k)ti,
Hτ,k(
τ (xi)
(6)
m−1
m−1
X
i=0
X
i=1
i−1
Y
l=0
for some k ∈ K ×. Moreover, with tm = ttm−1, also
(7)
H(tm) = H(cid:16)
aiti(cid:17) =
m−1
X
i=0
m−1
H(ai)H(t)i = τ (a0) +
X
i=0
m−1
X
i=1
τ (ai)(cid:0)
i−1
Y
l=0
σl(k)(cid:1)ti
and H(ttm−1) = H(t)H(tm−1) = H(t)H(t)m−1, i.e.
(8)
H(t)m = H(t)H(t)m−1 = kσ(k)··· σm−1(k)tm = kσ(k)··· σm−1(k)
m−1
X
i=0
aiti.
q=i σq(k)(cid:17)ai for all i ∈ {0, . . . , m− 1}. Thus H is
Comparing (7) and (8) gives τ (ai) = (cid:16)Qm−1
as in (6) where k ∈ K × is such that (1) holds for all i ∈ {0, . . . , m − 1}.
The Hτ,k are indeed automorphisms of Sf : Let G be an automorphism of R = K[t; σ].
i=0 biti ∈ K[t; σ] we have G(h(t)) = τ (b0) + Pm−1
Then for h(t) = Pr
l=0 σl(k)ti for
some τ ∈ Aut(K) such that σ ◦ τ = τ ◦ σ and some k ∈ K × (the proof of [20, Lemma 1]
works for any R = K[t; σ], or cf. [18, p. 75]). It is straightforward to see that Sf ∼= SG(f )
[20, Theorem 7], the proof works for any R = K[t; σ]). In particular, this means that
(cf.
if k ∈ K × satisfies (1) then G(f (t)) = (cid:0)Qm−1
l=0 σl(k)(cid:1)f (t) = af (t) with a ∈ K × being
the product of the σl(k), and so G induces an isomorphism of Sf with Saf = Sf , i.e. an
automorphism of Sf . The automorphisms of AutF (Sf ) are therefore all canonically induced
by the F -automorphisms G of R = K[t; σ] which satisfy (1).
(cid:3)
i=i τ (bi)Qi−1
The assumption that n ≥ m − 1 is needed in (4) to conclude that ki = 0 for i =
0, 2, 3, . . . , m − 1 and so H(t) = kt. If n < m − 1 we still obtain:
Theorem 5. Suppose σ commutes with all τ ∈ AutF (K). Let n < m − 1.
(i) For all k ∈ K × satisfying (1) for all i ∈ {0, . . . , m − 1}, the maps Hτ,k from Theorem 4
are automorphisms of Sf and form a subgroup of AutF (Sf ).
(ii) Let H ∈ AutF (Sf ) and N = Nucr(Sf ). Then HK = τ for some τ ∈ AutF (K),
HN ∈ AutF (N ) and H(t) = g(t) with g(t) = k1t + k1+nt1+n + k1+2nt1+2n + . . .+ k1+snt1+sn
for some k1+ln ∈ K, 0 ≤ l ≤ s. Moreover, g(t)i is well defined for all i ≤ m − 1, i.e., all
powers of g(t) are associative for all i ≤ m − 1, and g(t)g(t)m−1 = Pm−1
i=0 τ (ai)g(t)i. Thus
m−1
m−1
H(
X
i=0
xiti) =
X
i=0
τ (xi)g(t)i.
Proof. (i) is straightforward, using the relevant parts of the proof of Theorem 4. Note that
the inverse of Hτ,k is Hτ −1,τ −1(k−1) and Hτ,k ◦ Hρ,b = Hτ ρ,τ (b)k.
(ii) Let H : Sf → Sf be an automorphism. As in Theorem 4, HK = τ for some τ ∈
AutF (K), and HN ∈ AutF (N ). Suppose H(t) = Pm−1
i=0 kiti for some ki ∈ K. Comparing
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
7
the coefficients of t in H(tz) = H(t)H(z) = H(σ(z)t) we obtain (5) for all i ∈ {0, . . . , m− 1}
and all z ∈ K. Since σ has order n < m − 1, here, (5) only implies ki = 0 for i ∈ {0, . . . , n},
i 6= 1. Therefore H(t) = k1t+Pm−1
i=n+1 kiti for some ki ∈ K. However, σi(z) = σ(z) for all z ∈
K if and only if i = nl+1 for some l ∈ Z since σ has order n. Therefore (5) implies ki = 0 for
every i 6= 1+nl, l ∈ N0, i ∈ {0, . . . , m − 1}. Thus H(t) = k1t+k1+ntn+1+. . .+k1+snt1+sn for
some s, sn < m− 1. Furthermore, H(tm) = H(Pm−1
i=0 τ (ai)(k1t + k1+nt1+n +
. . . + k1+snt1+sn)i and H(tm) = (k1t + k1+nt1+n + . . . + k1+snt1+sn)m. Similarly, H(t)i =
(k1t + k1+nt1+n + . . . + k1+snt1+sn)i. Together these imply the assertion.
(cid:3)
i=0 aiti) = Pm−1
A closer look at the proof of Theorems 4 and 5 reveals that in fact the following holds
without requiring σ to commute with all τ ∈ AutF (K):
Proposition 6. (i) For every k ∈ K × satisfying (1) for all i ∈ {0, . . . , m − 1} for τ = id,
Hid,k is an automorphism of Sf and generates a subgroup of AutF (Sf ).
(ii) If any H ∈ AutF (Sf ) restricts to some τ ∈ AutF (K) such that τ ◦ σ = σ ◦ τ then
H = Hτ,k with k ∈ K × as in Theorem 4. Moreover, {Hτ,k τ ∈ AutF (Sf ), τ ◦ σ = σ ◦ τ, k ∈
K × with τ (ai) = (Qm−1
(iii) If m = 2, H ∈ Aut(Sf ) if and only if H = Hτ,k with τ ◦ σ = σ ◦ τ , τ (a0) = kσ(k)a0,
and τ (a1) = σ(k)a1.
l=i σl(k))ai for all i ∈ {0, . . . , m − 1}} is a subgroup of AutF (Sf ).
2.2. The automorphisms groups of Sf for f (t) = tm − a ∈ K[t; σ], a ∈ K \ F , are crucial in
the understanding of the automorphism groups of all the algebras Sg, as for all nonassociative
Sg with g(t) = tm − Pm−1
i=0 biti ∈ K[t; σ] such that b0 = a, AutF (Sg) is a subgroup of
AutF (Sf ):
i=0 biti ∈ K[t; σ] not be invariant.
i=0 aiti ∈ K[t; σ] not be invariant and assume bi ∈ {0, ai} for all
Theorem 7. Suppose σ commutes with all τ ∈ Gal(K/F ). Let n ≥ m − 1 and g(t) =
tm − Pm−1
(i) If f (t) = tm − b0 ∈ K[t; σ], b0 ∈ K \ F , then AutF (Sg) ⊂ AutF (Sf ) is a subgroup.
(ii) Let f (t) = tm − Pm−1
i ∈ {0, . . . , m − 1}. Then AutF (Sg) ⊂ AutF (Sf ) is a subgroup.
Proof. (i) Let H ∈ AutF (Sg). By Theorem 4, H has the form H(Pm−1
Pm−1
for all i = 0, . . . , m − 1.
automorphism of Sf , again by Theorem 4.
(ii) The proof is analogous to (i).
i=0 xiti) = τ (x0) +
j=i σj(k)(cid:17)bi
In particular, τ (b0) = kσ(k)··· σm−1(k)b0 and so H is also an
i=1 τ (xi)Qi−1
l=0 σl(k)ti, where τ ∈ AutF (K) and k ∈ K × satisfy τ (bi) = (cid:16)Qm−1
(cid:3)
Similarily, for n < m − 1 employing Theorem 5 we obtain:
i=0 biti ∈ K[t; σ] not be invariant.
Theorem 8. Suppose σ commutes with all τ ∈ AutF (K). Let n < m − 1 and g(t) =
tm − Pm−1
(i) If f (t) = tm − b0 ∈ K[t; σ], b0 ∈ K \ F , then {H ∈ AutF (Sg) H = Hτ,k} is a subgroup
of {H ∈ AutF (Sf ) H = Hτ,k}.
(ii) Let f (t) = tm − Pm−1
i=0 aiti ∈ K[t; σ] not be invariant such that bi ∈ {0, ai} for all
i ∈ {0, . . . , m− 1} Then {H ∈ AutF (Sg) H = Hτ,k} is a subgroup of {H ∈ AutF (Sf ) H =
Hτ,k}.
8
C. BROWN AND S. PUMPL UN
The automorphism groups of Sf with f (t) = tm − a ∈ K[t; σ] are therefore particularly
relevant.
3. Necessary conditions for extending Galois automorphisms to Sf
From now on we restrict ourselves to the situation that R = K[t; σ] and F = Fix(σ),
where K/F is a finite Galois field extension and σ of order n.
We take a closer look at Equality (1), which gives necessary conditions on how to choose
the elements k ∈ K × used to extend τ ∈ Gal(K/F ) to AutF (Sf ). These become more
restrictive for the choice of the elements k, the more coefficients in f (t) are non-zero. Let
NK/F : K → F be the norm of K/F . All monic polynomials f considered in the following
are assumed to not be invariant and of degree m.
Proposition 9. Suppose that σ and τ commute. Let f (t) = tm − Pm−1
k ∈ K × such that
m−1
i=0 aiti ∈ K[t; σ] and
(1)
τ (ai) = (cid:16)
σl(k)(cid:17)ai
Y
l=i
for all i ∈ {0, . . . , m − 1}. Then:
(i) For all i ∈ {0, . . . , m − 1} with ai 6= 0, NK/F (k) is an (m − i)th root of unity.
In particular, if a0 6= 0 (e.g., if f (t) is irreducible) then NK/F (k) is an mth root of unity,
and if am−1 6= 0 then NK/F (k) = 1.
If am−1 ∈ Fix(τ )× then k = 1.
(ii) If τ 6= id and there is some i such that ai is not contained in Fix(τ ), then k 6= 1.
(iii) Suppose that there is some ai 6= 0 and F does not contain any non-trivial (m − i)th
roots of unity. Then NK/F (k) = 1.
(iv) If there is an i ∈ {0, . . . , m − 1} such that ai ∈ Fix(τ )×, then 1 = Qm−1
particular, if n = m, σ generates Gal(K/F ), and a0 ∈ Fix(τ )× then k ∈ ker(NK/F ).
(v) Suppose τ = idK. Then for all i ∈ {0, . . . , m − 1} with ai 6= 0, 1 = Qm−1
l=i σl(k). In
particular, if n = m, σ generates Gal(K/F ) and a0 6= 0 then k ∈ ker(NK/F ). In this case,
the automorphisms extending idK are in one-one correspondence with those k ∈ ker(NK/F )
satisfying (1).
l=i σl(k). In
Proof. (i) Equality (1) states that τ (ai) = (cid:16)Qm−1
l=i σl(k)(cid:17)ai for all i ∈ {0, . . . , m− 1}. Thus
NK/F (ai) = Qm−1
l=i NK/F (σl(k))NK/F (ai) (apply NK/F to both sides of (1)), and there-
fore NK/F (ai) = NK/F (k)m−iNK/F (ai) for all i ∈ {0, . . . , m − 1} is a necessary condition
on k. For all ai 6= 0, this yields 1 = NK/F (k)m−i therefore NK/F (k) ∈ F × must be an
(m − i)th root of unity, for all i ∈ {0, . . . , m − 1}, with ai 6= 0. Hence if am−1 6= 0 then
τ (am−1) = σm−1(k)am−1, thus NK/F (am−1) = NK/F (k)NK/F (am−1), i.e. NK/F (k) = 1. If
even am−1 ∈ Fix(τ )× then am−1 = σm−1(k)am−1 means σm−1(k) = 1, i.e. k = 1.
(ii) k = 1 implies τ (ai) = ai, i.e. ai ∈ Fix(τ ) for all i ∈ {0, . . . , m − 1}.
(iii) By (i), NK/F (k) ∈ F × is an (m − i)th root of unity, for all i ∈ {0, . . . , m − 1} with
ai 6= 0. If F does not contain any non-trivial (m − i)th roots of unity, then NK/F (k) = 1.
(iv) If there is an i ∈ {0, . . . , m − 1} such that ai ∈ Fix(τ )×, then (1) becomes 1 =
Qm−1
l=i σl(k).
In particular, if a0 ∈ Fix(τ )×, m = n and σ generates Gal(K/F ), then
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
9
NK/F (k) = 1 is a necessary condition on k.
(v) Here, (1) becomes 1 = Qm−1
l=i σl(k) for all i ∈ {0, . . . , m − 1} with ai 6= 0. In particular,
if n = m, σ generates Gal(K/F ) and a0 6= 0 (which happens if f (t) is irreducible) then
NK/F (k) = 1 is a necessary condition on k.
(cid:3)
For instance, Proposition 9 (i) yields that k = 1 if am−1 ∈ Fix(τ )× and so Theorems 4
and 5 become:
i=0 aiti ∈
Theorem 10. Suppose σ commutes with all τ ∈ Gal(K/F ) and f (t) = tm − Pm−1
K[t; σ] is not invariant with am−1 ∈ F ×.
(i) Let n ≥ m − 1. If ai 6∈ Fix(τ ) for all τ 6= id and all non-zero ai, i 6= m − 1, then
AutF (Sf ) = {id}.
If f (t) ∈ F [t; σ], any automorphism H of Sf has the form Hτ,1 where τ ∈ Gal(K/F ), and
AutF (Sf ) ∼= Gal(K/F ).
(ii) Let n < m − 1. If f (t) ∈ F [t; σ] is not invariant, the maps Hτ,1 are automorphisms of
Sf for all τ ∈ Gal(K/F ) and Gal(K/F ) is isomorphic to a subgroup of AutF (Sf ).
Proof. (i) H is an automorphism of Sf if and only if H has the form Hτ,k, where τ ∈
Gal(K/F ) and k ∈ K × is such that τ (ai) = (cid:16)Qm−1
l=i σl(k)(cid:17)ai for all i ∈ {0, . . . , m − 1}.
Since am−1 ∈ F × we have am−1 ∈ Fix(τ )× for all τ which forces k = 1 as the only possibility
for any τ ∈ Gal(K/F ) by Proposition 9 (i). This in turn means that any extension Hτ,k has
the form Hτ,1. In particular, the existence of an extension Hτ,k, τ 6= id, implies τ (ai) = ai
for all non-zero ai, i 6= m − 1, that is ai ∈ Fix(τ ) for all non-zero ai.
that extends to an automorphism of Sf and AutF (Sf ) = {Hid,1} = {id}.
Thus if ai 6∈ Fix(τ ) for all τ 6= id and all i ∈ {0, . . . , m − 2} then there is no non-trivial τ
If f (t) ∈ F [t; σ] then AutF (Sf ) = {Hτ,1} ∼= Gal(K/F ).
(ii) follows from (i) and Theorem 8.
(cid:3)
Note that indeed Condition (1) heavily restricts the choice of available k to k = 1 in most
cases.
Corollary 11. Suppose σ commutes with all τ ∈ Gal(K/F ). Let n ≥ m − 1 and f (t) =
tm − a0 ∈ K[t; σ], a0 ∈ K \ F .
(i) H ∈ AutF (Sf ) if and only if H = Hτ,k where k ∈ K × is such that τ (a0) = (cid:16)Qm−1
In particular, here NK/F (k) is an mth root of unity.
(ii) For all g(t) = tm − Pm−1
AutF (Sf ).
i=0 aiti ∈ K[t; σ] with a0 ∈ K \ F , AutF (Sg) is a subgroup of
l=0 σl(k)(cid:17)a0.
Proof. (i) follows from Theorem 4 and Proposition 9.
(ii) follows from Theorem 7.
(cid:3)
For f (t) = tm − a0 ∈ K[t; σ], a0 ∈ K \ F , the automorphisms Hid,k extending idK thus
are in one-to-one correspondence with those k satisfying Qm−1
l=0 σl(k) = 1 (in particular, we
have NK/F (k)m = 1). Analogously, we still obtain for n < m − 1 employing Theorem 5 and
Theorem 8:
10
C. BROWN AND S. PUMPL UN
Corollary 12. Suppose σ commutes with all τ ∈ Gal(K/F ). Let n < m − 1 and f (t) =
tm − a0 ∈ K[t; σ], a0 ∈ K \ F .
(i) For all k ∈ K × with NK/F (k) an mth root of unity and τ (a0) = (cid:16)Qm−1
l=0 σl(k)(cid:17)a0, the
maps Hτ,k are automorphisms of Sf .
(ii) For all g(t) = tm − Pm−1
is a subgroup of {H ∈ AutF (Sf ) H = Hτ,k}.
i=0 aiti ∈ K[t; σ] with a0 ∈ K \ F , {H ∈ AutF (Sg) H = Hτ,k}
For m = n and K/F a cyclic field extension, the algebras considered in Corollary 11 are
called nonassociative cyclic algebras of degree m, as they can be seen as canonical general-
izations of associative cyclic algebras. These algebras are treated in Section 6.
4. Automorphisms extending idK when K/F is a cyclic field extension
Let f (t) = tm −Pm−1
i=0 aiti ∈ K[t; σ] not be invariant. In general, we know that if Sf has
nucleus K then every inner automorphism Gc with c ∈ K ×, extends idK. Conversely, an
extension Hid,k of idK is inner for the right choice of k:
Lemma 13. Let k = c−1σ(c) with c ∈ K ×, then Hid,k ∈ AutF (Sf ) is an inner automor-
phism.
i=0 xiti(cid:17) = (cid:16)c−1 Pm−1
Proof. A simple calculation shows that Gc(cid:16)Pm−1
i=1 xic−1σi(c)ti = Hid,k(cid:16)Pm−1
Pm−1
Let now K/F be a cyclic Galois field extension of degree n with Gal(K/F ) = hσi and
norm NK/F : K → F, NK/F (k) = kσ(k)σ2(k)··· σn−1(k). By Hilbert's Theorem 90,
ker(NK/F ) = ∆σ(1), where ∆σ(l) = {σ(c)lc−1 c ∈ K ×} is the σ-conjugacy class of l ∈ K ×
[19].
i=0 xiti(cid:17).
i=0 xiti(cid:17)c = x0 +
(cid:3)
Theorem 14. (i) Every automorphism Hid,k ∈ AutF (Sf ) such that NK/F (k) = 1 is an
inner automorphism.
(ii) If n ≥ m − 1 and am−1 6= 0, or if n = m, ai = 0 for all i 6= 0 and a0 ∈ K \ F , then
these are all the automorphisms extending idK.
Proof. (i) Suppose there is Hid,k ∈ AutF (Sf ) with NK/F (k) = 1, then by Hilbert 90, there
exists c ∈ K × such that k = c−1σ(c). Thus Hid,k = Hid,c−1σ(c) for c ∈ K × and so Gc = Hid,k
by Lemma 13.
(ii) By Theorem 4 and Proposition 9 (i), these are all the automorphisms extending idK
when n ≥ m − 1 if am−1 6= 0. The remaining assertion is proved analogoulsy.
(cid:3)
5. Cyclic subgroups of AutF (Sf )
For any Galois field extension K/F and σ ∈ Gal(K/F ) of order n, we now give some
conditions for AutF (Sf ) to have cyclic subgroups.
Theorem 15. Suppose F contains an sth root of unity ω. Suppose that either f (t) = ts−a ∈
K[t; σ] where a ∈ K \ F , or f (t) = tsl −Pl−1
i=0 aistis ∈ K[t; σ] such that Sf is not associative.
Then hHid,ωi is a cyclic subgroup of AutF (Sf ) of order at most s and of order s, if ω is a
primitive root of unity.
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
11
i=0 aistis. Then we have Qls−1
q=is σq(ωj) = ωj(ls−is) = 1 for all i =
q=is σq(ωj)(cid:17)ais for all i = 0, . . . , l− 1 and so Hid,ωj ∈ AutF (Sf )
Proof. (i) Let f (t) = ts−a. Then ωjσ(ωj)··· σs−1(ωj) = ωjs = 1 and so Hid,ωj ∈ AutF (Sf )
for all 0 ≤ j ≤ s − 1 by Proposition 6.
(ii) Let f (t) = tsl − Pl−1
0, . . . , l− 1. Hence ais = (cid:16)Qls−1
for all 0 ≤ j ≤ s − 1 by Proposition 6.
In both (i) and (ii), hHid,ωi is a cyclic subgroup of Aut(Sf ) of order less or equal to s, since
Hid,ωj ◦ Hid,ωr = Hid,ωj+r for all 0 ≤ j, r ≤ sl − 1.
Lemma 16. Let F have characteristic not two, m be even and f (t) = tm−P(m−2)/2
K[t; σ] not invariant. Then {Hid,1, Hid,−1} is a subgroup of Sf of order 2.
Proof. The maps Hid,1 and Hid,−1 are automorphisms of Sf by Proposition 6, and Hid,−1 ◦
Hid,−1 = Hid,1.
a2it2i ∈
(cid:3)
(cid:3)
i=0
If f ∈ F [t] ⊂ K[t; σ], we obtain:
Theorem 17. Suppose σ commutes with all τ ∈ Gal(K/F ), and f (t) = tm − Pm−1
F [t; σ] ⊂ K[t; σ] is not invariant.
(i) hHσ,1i ∼= Z/nZ is a cyclic subgroup of AutF (Sf ).
(ii) Suppose Gal(K/F ) = hσi, n = m is prime, a0 6= 0 and not all of a1, . . . , am−1 are zero.
Then AutF (Sf ) = hHσ,1i ∼= Z/mZ.
Proof. Let j ∈ {0, . . . , n − 1}. Since τ (ai) = ai for all i, here (1) becomes
i=0 aiti ∈
(9)
ai = (cid:16)
m−1
Y
q=i
σq(k)(cid:17)ai
for all i ∈ {0, . . . , m − 1}.
(i) Clearly, (9) is satisfied for k = 1 and all i ∈ {0, . . . , m − 1}, therefore the maps
Hτ,1 are automorphisms of Sf for all τ ∈ Gal(K/F ) by Theorems 4 and 5. We have
Hσj ,1 ◦ Hσl,1 = Hσj+l,1 and Hσn,1 = Hid,1. Hence hHσ,1i = (cid:8)Hid,1, Hσ,1, . . . , Hσm−1,1(cid:9) is a
cyclic subgroup of order n.
(ii) By Theorem 4, the automorphisms of Sf are exactly the maps Hσj ,k where j ∈ {0, . . . , n − 1}
and k ∈ K × satisfies (9) for all i ∈ {0, . . . , m − 1}. The maps Hσj ,1 are therefore automor-
phisms of Sf for all j ∈ {0, . . . , n − 1}. We prove that these are the only automorphisms of
Sf : a0 6= 0 and so NK/F (k) = 1 by (9). Therefore, by Hilbert 90, there exists α ∈ K such
that k = σ(α)/α. Let l ∈ {1, . . . , m − 1} be such that al 6= 0. Then by (9),
1 =
m−1
Y
q=l
σq(k) =
m−1
Y
q=l
σq(cid:0)
σ(α)
α (cid:1) = Qm
q=l+1 σq(α)
Qm−1
q=l σq(α)
=
α
σl(α)
.
Thus α ∈ Fix(σj) = F since m is prime. Therefore k = σ(α)/α = α/α = 1 as required. (cid:3)
This complements our results from Theorem 10, which in case Gal(K/F ) is cyclic of
degree n mean the following:
Corollary 18. Suppose Gal(K/F ) is cyclic of degree n and f (t) = tm−Pm−1
not invariant with am−1 ∈ F ×.
i=0 aiti ∈ F [t; σ]
12
C. BROWN AND S. PUMPL UN
(i) Let n ≥ m − 1 then for all τ ∈ Gal(K/F ) the maps Hτ,1 are exactly the automorphisms
of Sf and AutF (Sf ) ∼= Gal(K/F ) ∼= Z/nZ.
(ii) Let n < m − 1 then for all τ ∈ Gal(K/F ) the maps Hτ,1 are automorphisms of Sf and
Gal(K/F ) ∼= Z/nZ is isomorphic to a subgroup of AutF (Sf ).
6. Nonassociative cyclic algebras
6.1. Let K/F be a cyclic Galois extension of degree m with Gal(K/F ) = hσi and f (t) =
tm − a ∈ K[t; σ]. Then (K/F, σ, a) = K[t; σ]/K[t; σ](tm − a) is called a nonassociative cyclic
algebra of degree m over F . It is not associative for all a ∈ K \ F and a cyclic associative
central simple algebra over F for a ∈ F ×. We will only consider the case that a ∈ K \ F . If
1, a, a2, . . . , am−1 are linearly independent over F then (K/F, σ, a) is a division algebra (cf.
[32], [30] for finite F ). In particular, if K/F is of prime degree then (K/F, σ, a) is a division
algebra for every a ∈ K \ F .
Theorem 19. Let A = (K/F, σ, a) be a nonassociative cyclic algebra of degree m.
(i) All the automorphisms of A which extend idK are inner automorphisms and of the form
Hid,l for all l ∈ K × such that NK/F (l) = 1.
The subgroup they generate in AutF (A) is isomorphic to ker(NK/F ).
(ii) An automorphism σj 6= id can be extended to H ∈ AutF (A), if and only if there is
some l ∈ K such that σj (a) = NK/F (l)a. In that case, H = Hσj ,l and if m is prime then
NK/F (l) = ω for an mth root of unity 1 6= ω ∈ F .
(iii) Let c ∈ K \ F and suppose there exists r ∈ N such that cr ∈ F ×. Let r be minimal.
Then hGci is a cyclic subgroup of AutF (Sf ) of order r.
Proof. Theorem 4, Theorem 14 (ii), and Proposition 9 imply (i) and (ii).
(iii) Let c ∈ K \ F . Then Gc is an automorphism, because K is the nucleus of A. Since
Gc ◦ Gc = Gc2 , Gc ◦ Gc ◦ Gc = Gc3 and so on, we have Gcr = id if and only if cr ∈ F . If
r ∈ N is smallest possible then hGci is a cyclic subgroup of AutF (Sf ) of order r.
(cid:3)
Note that different roots of unity yield different l in Theorem 19 (ii). This yields:
Theorem 20. Let A = (K/F, σ, a) be a nonassociative cyclic algebra of degree m. Suppose
F contains a non-trivial mth root of unity ω.
(i) hHid,ωi is a cyclic subgroup of AutF (A) of order at most m. If ω is a primitive mth root
of unity, then hHid,ωi has order m.
(ii) If there is an element l ∈ K, such that NK/F (l) = ω for ω a primitive mth root of unity
and σ(d) = ωd, then the subgroup generated by Hσ,l has order m2.
Proof. (i) follows from Theorem 15.
(ii) Suppose σ can be extended to an F -automorphism H of A. Then by Theorem 19, there
is an element l ∈ K, such that NK/F (l) = ω, ω 6= 1 and σ(d) = ωd, and H = Hσ,l. (If
1 = NK/F (l), then σ(d) = d, contradiction.)
The subgroup generated by H = Hσ,l has order greater than m, since Hσ,l ◦ ··· ◦ Hσ,l
(m-times) becomes Hσm,b = Hid,ω with ω = NK/F (l). Hid,ω has order m, so the subgroup
generated by H = Hσ,l has order m2.
(cid:3)
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
13
6.2. The case that m is prime. Let us now assume that the cyclic field extension K/F
has prime degree m = deg(f ). Suppose that F contains a primitive mth root of unity, where
m is prime to the characteristic of F . Then K = F (d), where d is a root of an irreducible
polynomial tm − c ∈ F [t].
Lemma 21. (cf. [33, Lemma 6.2.7]) The eigenvalues of σj ∈ Gal(K/F ) are precisely the
mth roots of unity. Moreover, the only possible eigenvectors are of the form edi for some i,
0 ≤ i ≤ m − 1 and some e ∈ F .
Let f (t) = tm − a ∈ K[t; σ],
a 6∈ F. Then we get the following strong restriction for
automorphisms of Sf :
Theorem 22. H is an automorphism of Sf extending σj 6= id if and only if H = Hσj ,k for
some k ∈ K ×, where NK/F (k) is an mth root of unity and a = eds for some e ∈ F × and
some ds.
Proof. H is an automorphism of Sf if and only if H = Hσj ,k where j ∈ {0, . . . , m − 1} and
k ∈ K × is such that σj(a) = (cid:16)Qm−1
l=0 σl(k)(cid:17)a = NK/F (k)a. For all σj 6= id, by Lemma 21
this condition is equivalent to NK/F (k) being an mth root of unity and a = eds for some ds
and e ∈ F ×, for all k ∈ K ×.
(cid:3)
Applying Theorem 7, our results for the automorphisms of a nonassociative cyclic algebra
A = (K/F, σ, a) of degree m yield the following observations for more general algebras Sg:
Corollary 23. Suppose Gal(K/F ) = hσi is cyclic of degree m and g(t) = tm−Pm−1
K[t; σ] is not invariant with a0 ∈ K \ F . Suppose one of the following holds:
i=0 aiti ∈
• F has no mth root of unity.
• m is prime and F contains a primitive mth root of unity, where m is prime to the
characteristic of F . Let K = F (d) as in Section 6.2 and a0 6= edi, e ∈ F ×.
Then every F -automorphism of Sg leaves K fixed, is inner and AutF (Sg) is a subgroup of
ker(NK/F ), thus cyclic. In particular, if ker(NK/F ) has prime order, then either AutF (Sg)
is trivial or AutF (Sg) ∼= ker(NK/F ).
6.3. The automorphism groups of nonassociative quaternion algebras. Recall the
dicyclic group
(10)
Dicl = hx, y y2l = 1, x2 = yl, x−1yx = y−1i
of order 4l. The semidirect product Z/sZ ⋊l Z/nZ between the cyclic groups Z/sZ and
Z/nZ corresponds to a choice of an integer l such that ln ≡ 1 mod s. It can be described
by the presentation Z/sZ ⋊l Z/nZ = hx, y xs = 1, yn = 1, yxy−1 = xli.
We obtain the following result for the automorphism groups of nonassociative quaternion
algebras (where m = 2):
Theorem 24. Suppose K = F (√b) is a quadratic field extension of F , char(F ) 6= 2,
and consider the nonassociative quaternion algebra A = (K/F, σ, λ√b) for some λ ∈ F ×.
Suppose there exists k ∈ K × such that kσ(k) = −1.
14
C. BROWN AND S. PUMPL UN
For every c ∈ K \ F for which there is a positive integer j such that cj ∈ F ×, pick the
smallest such j.
(i) If j is even then AutF (Sf ) contains the dicyclic group of order 2j.
(ii) If j is odd then AutF (Sf ) contains a subgroup isomorphic to the semidirect product
Z/jZ ⋊j−1 Z/4Z. In particular, AutF (A) always contains a subgroup isomorphic to Z/4Z.
Proof. Since σ(√b) = −√b and kσ(k) = −1, Hσ,k ∈ AutF (Sf ) by Theorem 19. A simple
calculation shows that hHσ,ki = {Hσ,k, Hid,−1, Hσ,−k, Hid,1}. hGci is a cyclic subgroup of
AutF (Sf ) of order j by Theorem 19 (iii).
(i) Suppose j is even and write j = 2l. We prove first that Gcl = Hid,−1. Write cl =
µ0 + µ1√b for some µ0, µ1 ∈ F . Then cj = c2l = µ2
1b + 2µ0µ1√b ∈ F which implies
2µ0µ1 = 0. Hence µ0 = 0 or µ1 = 0. Since j is minimal, cl /∈ F so µ0 = 0 and cl = µ1√b.
0 + µ2
We obtain
Gcl(x0 + x1t) = x0 + x1(µ1√b)−1σ(µ1√b)t
1 b−1√b(−µ1√b)t
= x0 + x1µ−1
= x0 − x1t = Hid,−1(x0 + x1t)
which implies Gcl = Hid,−1. Next we prove (Hσ,k)−1GcHσ,k = G−1
show (Hσ,k)−1 = Hσ,−k and G−1
c = Gσ(c). We have
c
. Simple calculations
Hσ,−k(cid:0)Gc(cid:0)Hσ,k(cid:0)x0 + x1t(cid:1)(cid:1)(cid:1) = Hσ,−k(cid:0)Gc(cid:0)σ(x0) + σ(x1)kt(cid:1)(cid:1)
= Hσ,−k(cid:0)σ(x0) + σ(x1)kc−1σ(c)t(cid:1)
= x0 − x1σ(k)σ(c−1)ckt
= x0 + x1σ(c−1)ct = Gσ(c)(cid:0)x0 + x1t(cid:1)
and so (Hσ,k)−1GcHσ,k = G−1
c
.
Thus H 2
σ,k = Hid,−1 = Gcl = Gl
c, G2l
c = id and (Hσ,k)−1GcHσ,k = G−1
c
. Hence hHσ,k, Gci
has the presentation (10) as required.
(ii) Suppose j is odd. Then hGci does not contain Hid,−1 as Hid,−1 has order 2 which implies
hHσ,ki ∩ hGci = {id}. Furthermore (Hσ,k)−1GcHσ,k = G−1
c = Gcj−1 can be shown
similarly as in (i). Note that (j − 1)4 = j4 − 4j3 + 6j2 − 4j + 1 ≡ 1 mod (j). Thus AutF (Sf )
contains the subgroup hGci ⋊j−1 hHσ,ki ∼= Z/jZ ⋊j−1 Z/4Z as required.
In particular, choose c = √b in (i), so that j = 2. This implies AutF (A) contains the dicyclic
group of order 4, which is the cyclic group of order 4.
c = Gj−1
(cid:3)
Example 25. (i) Let F = Q(i), K = F (√−3), σ(√−3) = −√−3 and A = (K/F, σ, λ√−3)
be a nonassociative quaternion algebra with some λ ∈ F ×. Note that for k = i we have
iσ(i) = −1. Let c = 1 +√−3. Then c2 = −2 + 2√−3 and c3 = −8 which implies j = 3 here.
Therefore AutF (Sf ) contains a subgroup isomorphic to the semidirect product Z/3Z⋊2Z/4Z
by Theorem 24.
(ii) Let F = Q(i), K = F (p−1/12), σ(p−1/12) = −p−1/12 and A = (K/F, σ, λp−1/12)
be a nonassociative quaternion algebra for some λ ∈ F ×. Again for k = i we have iσ(i) = −1.
Let c = 1 + 2p−1/12. Then c2 = 2/3 + 2i/√3, c3 = 8i/3√3, c4 = −8/9 + 8i/3√3,
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
15
c5 = −16/9 + 16i/9√3 and c6 = −64/27. Hence c, c2, c3, c4, c5 ∈ K \ F and c6 ∈ F .
Therefore AutF (A) contains the dicyclic group of order 12 by Theorem 24.
7. Isomorphisms between Sf and Sg
The proofs of the previous sections can be adapted to check when two Petit algebras are
isomorphic and when not. This is not the main focus of this paper so we just point out how
some of the results can be transferred.
If K and L are fields, and Sf = K[t; σ]/K[t; σ]f (t) ∼= L[t; σ′]/L[t; σ′]g(t) = Sg, then
K ∼= L, N ucr(Sf ) ∼= N ucr(Sg), deg(f ) = deg(g), and Fix(σ) ∼= Fix(σ′), since isomorphic
algebras have the same dimensions, and isomorphic nuclei and center.
If G is an automorphism of R = K[t; σ] which restricts to an automorphism τ on K which
commutes with σ, f ∈ R is irreducible and g(t) = G(f (t)), then G induces an isomorphism
Sf ∼= SG(g) [20, Theorem 7] (the proof works for any base field).
Pm−1
From now on let F be the fixed field of σ, σ have order n, and both f (t) = tm −
i=0 aiti and g(t) = tm − Pm−1
i=0 biti ∈ K[t; σ] be not invariant. Then the following is
proved analogously to Theorem 4, Theorem 5 and Proposition 6:
Theorem 26. Suppose σ commutes with all τ ∈ AutF (K) and n ≥ m − 1. Then Sf ∼= Sg
if and only if there exists τ ∈ AutF (K) and k ∈ K × such that
(11)
τ (ai) = (cid:16)
σl(k)(cid:17)bi
m−1
Y
l=i
for all i ∈ {0, . . . , m − 1}. Every such τ and k yield a unique isomorphism Gτ,k : Sf → Sg,
m−1
Gτ,k(
X
i=0
xiti) = τ (x0) +
m−1
X
i=1
τ (xi)
i−1
Y
l=0
σl(k)ti.
If n < m − 1 we still get a partial result:
Theorem 27. Suppose there exists τ ∈ AutF (K) and k ∈ K × such that τ ◦ σ = σ ◦ τ and
such that (11) holds for all i ∈ {0, . . . , m− 1}. Then Sf ∼= Sg with an isomorphism given by
m−1
Gτ,k(
X
i=0
xiti) = τ (x0) +
m−1
X
i=1
τ (xi)
i−1
Y
l=0
σl(k)ti
as in Theorem 26.
Corollary 28. For every k ∈ K × such that ai = (cid:16)Qm−1
Gid,k : Sf → Sg is an isomorphism.
l=i σl(k)(cid:17)bi for all i ∈ {0, . . . , m− 1},
As a direct consequence of Theorem 26 we obtain:
Theorem 29. Suppose σ commutes with all τ ∈ AutF (K) and n ≥ m− 1. If Sf ∼= Sg, then
ai = 0 if and only if bi = 0, for all i ∈ {0, . . . , m − 1}.
Proof. If Sf ∼= Sg then by Theorem 26, there exists j ∈ {0, . . . , n − 1} and k ∈ K × such
that τ (ai) = (cid:16)Qm−1
l=i σl(k)(cid:17)bi for all i ∈ {0, . . . , m − 1}. This implies ai = 0 if and only if
bi = 0, for all i ∈ {0, . . . , m − 1}.
(cid:3)
16
C. BROWN AND S. PUMPL UN
From now on we restrict ourselves to the situation that R = K[t; σ] and F = Fix(σ),
where K/F is a finite Galois field extension and σ of order n. We take a closer look at the
consequences of Equality (11):
l=i σl(k)(cid:17)bi for all i ∈ {0, . . . , m− 1}.
Proposition 30. Let k ∈ K × such that τ (ai) = (cid:16)Qm−1
Then ai = 0 if and only if bi = 0 and:
(i) For all i ∈ {0, . . . , m − 1} with ai 6= 0, NK/F (ai) = NK/F (k)m−iNK/F (bi).
(ii) If there is some i ∈ {0, . . . , m − 1} such that ai ∈ Fix(τ )×, then ai/bi = Qm−1
l=i σl(k).
In particular, if am−1 ∈ F × and bm−1 ∈ F ×, then k ∈ F × and ai = km−ibi for all
i ∈ {0, . . . , m − 1}.
(iii) If a0 ∈ Fix(τ )×, m = n and Gal(K/F ) = hσi then a0 = NK/F (k)b0.
Proof. (i) Equality (11) implies that NK/F (ai) = Qm−1
l=i NK/F (σl(k))NK/F (bi) (simply
apply NK/F to both sides of (11)), therefore NK/F (ai) = NK/F (k)m−iNK/F (bi) for all
i ∈ {0, . . . , m − 1} is a necessary condition on k.
(ii) If there is an i ∈ {0, . . . , m − 1} such that ai ∈ Fix(τ )×, then (11) implies that
ai = (cid:16)Qm−1
l=i σl(k)(cid:17)bi, so that we obtain ai/bi = Qm−1
l=i σl(k).
Alternatively, if am−1 ∈ F × and bm−1 ∈ F ×, then am−1 = σm−1(k)bm−1 imply k ∈ F ×,
hence ai = km−1−ibi for all i ∈ {0, . . . , m − 1}.
(iii) In particular, if a0 ∈ Fix(τ )×, m = n and σ generates Gal(K/F ), then a0/b0 =
Qm−1
l=0 σl(k) = NK/F (k) is a necessary condition on k.
Corollary 31. Suppose σ commutes with all τ ∈ Gal(K/F ) and n ≥ m − 1. Assume that
one of the following holds:
(i) There exists i ∈ {0, . . . , m − 1} such that bi 6= 0 and NK/F (aib−1
(ii) m = n, a0 ∈ F × and b0 ∈ K \ F .
Then Sf ≇ Sg.
Corollary 32. Suppose Gal(K/F ) = hσi and n = m. Let f (t) = tm − a, g(t) = tm − b ∈
K[t; σ] where a, b ∈ K \ F .
(i) Sf ∼= Sg if and only if there exists τ ∈ Gal(K/F ) and k ∈ K × such that τ (a) = NK/F (k)b.
(ii) If σj(a) 6= NK/F (k)b for all k ∈ K ×, j = 0, . . . , m − 1, then Sf 6∼= Sg.
) 6∈ NK/F (K ×)m−i;
(cid:3)
i
These follow from Proposition 30. Note that Corollary 32 canonically generalizes well-
known criteria for associative cyclic algebras.
References
[1] R. Alfaro, A. V. Kelarev, Recent results on ring constructions for error-correcting codes. Algebraic
Structures and their Representations, XV Coloquio Latinoamericano de Algebra (Cocoyoc, Morelos,
Mexico, July 20-26, 2003), Contemporary Math. 376 (2005), 1-12.
[2] A. S. Amitsur, Differential polynomials and division algebras. Annals of Mathematics, Vol. 59 (2) (1954)
245-278.
[3] A. S. Amitsur, Non-commutative cyclic fields. Duke Math. J. 21 (1954), 87-105.
[4] A. S. Amitsur, Generic splitting fields of central simple algebras. Ann. of Math. 62 (2) (1955), 8-43.
[5] C. Brown, PhD Thesis University of Nottingham, in preparation.
[6] C. Brown, S. Pumplun, A. Steele, Automorphisms and isomorphisms of Jha-Johnson semifields obtained
from skew polynomial rings. Online at arXiv:1703.02356 [math.RA]
THE AUTOMORPHISMS OF PETIT'S ALGEBRAS
17
[7] J. Cazaran, A. V. Kelarev, S. J. Quinn, D. Vertigan, An algorithm for computing the minimum distances
of extensions of BCH codes embedded in semigroup rings. Semigroup Forum 73 (2006), 317-329.
[8] L. E. Dickson, Linear algebras in which division is always uniquely possible. Trans. Amer. Math. Soc. 7
(3) (1906), 370-390.
[9] J. Ducoat, F. Oggier, Lattice encoding of cyclic codes from skew polynomial rings. Proc. of the 4th
International Castle Meeting on Coding Theory and Applications, Palmela, 2014.
[10] J. Ducoat, F. Oggier, On skew polynomial codes and lattices from quotients of cyclic division algebras.
Adv. Math. Commun. 10 (1) (2016), 79-94.
[11] M. Giesbrecht, Factoring in skew-polynomial rings over finite fields. J. Symbolic Comput. 26 (4) (1998),
463-486.
[12] M. Giesbrecht, Y. Zhang, Factoring and decomposing Ore polynomials over Fq(t). Proceedings of the
2003 International Symposium on Symbolic and Algebraic Computation, 127134, ACM, New York, 2003.
[13] J. G`omez-Torrecillas, F. J. Lobillo, G. Navarro, Factoring Ore polynomials over Fq(t) is difficult. Online
at arXiv:1505.07252[math.RA]
[14] J. G`omez-Torrecillas, Basic module theory over non-commutative rings with computational aspects of
operator algebras. With an appendix by V. Levandovskyy. Lecture Notes in Comput. Sci. 8372, Algebraic
and algorithmic aspects of differential and integral operators, Springer, Heidelberg (2014) 23-82.
[15] N. Jacobson, "Finite-dimensional division algebras over fields." Springer Verlag, Berlin-Heidelberg-New
York, 1996.
[16] A. V. Kelarev, "Ring Constructions and Applications", World Scientific, River Edge, New Jork, 2002.
[17] A. V. Kelarev, P. Sole, Error-correcting codes as ideals in group rings. Contemporary Mathematics 273
(2001), 11-18.
[18] K. Kishimoto, On cyclic extensions of simple rings. J. Fac. Sci. Hokkaido Univ. Ser. I 19 (1966), 74-85.
[19] T. Y. Lam, A. Leroy, Hilbert 90 theorems over division rings. Trans. Amer. Math. Soc. 345 (2) (1994),
595-622.
[20] M. Lavrauw, J. Sheekey, Semifields from skew-polynomial rings. Adv. Geom. 13 (4) (2013), 583-604.
[21] F. Oggier, B. A. Sethuraman, Quotients of orders in cyclic algebras and space-time codes. Adv. Math.
Commun. 7 (4) (2013), 441-461.
[22] J.-C. Petit, Sur certains quasi-corps g´en´eralisant un type d'anneau-quotient. S´eminaire Dubriel. Alg`ebre
et th´eorie des nombres 20 (1966 - 67), 1-18.
[23] J.-C. Petit, Sur les quasi-corps distributifes `a base momog`ene. C. R. Acad. Sc. Paris 266 (1968), S´erie
A, 402-404.
[24] S. Pumplun, Quotients of orders in algebras obtained from skew polynomials and possible applications.
Online at arXiv:1609.04201 [math.RA]
[25] S. Pumplun, Finite nonassociative algebras obtained from skew polynomials and possible appli-
cations to (f, σ, δ)-codes. To appear in Advances in Mathematics of Communications. Online at
arXiv:1507.01491[cs.IT]
[26] S. Pumplun, How to obtain lattices from (f, σ, δ)-codes via a generalization of Construction A. Online
at arXiv:1607.03787 [cs.IT]
[27] S. Pumplun, A. Steele, The nonassociative algebras used to build fast-decodable space-time block codes.
Advances in Mathematics of Communications 9 (4) 2015, 449-469.
[28] S. Pumplun, A. Steele, Fast-decodable MIDO codes from nonassociative algebras. Int. J. of Information
and Coding Theory (IJICOT) 3 (1) 2015, 15-38.
[29] A. Steele, S. Pumplun, F. Oggier, MIDO space-time codes from associative and non-associative cyclic
algebras. Information Theory Workshop (ITW) 2012 IEEE (2012), 192-196.
[30] R. Sandler, Autotopism groups of some finite non-associative algebras. Amer. J. Math. 84 (1962), 239-
264.
[31] R. D. Schafer, "An Introduction to Nonassociative Algebras." Dover Publ., Inc., New York, 1995.
[32] A. Steele, Nonassociative cyclic algebras. Israel Journal of Mathematics 200 (1) (2014), 361-387.
[33] A. Steele, Some new classes of division algebras and potential applications to space-time block coding.
PhD Thesis, University of Nottingham 2013, online at eprints.nottingham.ac.uk/13934/
18
C. BROWN AND S. PUMPL UN
[34] G. P. Wene, Inner automorphisms of semifields. Note Mat. 29 (2009), suppl. 1, 231-242.
[35] G. P. Wene, Finite semifields three-dimensional over the left nuclei. Nonassociative algebra and its
applications (Sao Paulo, 1998), Lecture Notes in Pure and Appl. Math., 211, Dekker, New York, 2000,
447-456.
E-mail address: [email protected]; [email protected]
School of Mathematical Sciences, University of Nottingham, University Park, Nottingham
NG7 2RD, United Kingdom
|
1208.5298 | 1 | 1208 | 2012-08-27T06:39:19 | Submaximal Integral Domains | [
"math.RA"
] | It is proved that if $D$ is a $UFD$ and $R$ is a $D$-algebra, such that $U(R)\cap D\neq U(D)$, then $R$ has a maximal subring. In particular, if $R$ is a ring which either contains a unit $x$ which is not algebraic over the prime subring of $R$, or $R$ has zero characteristic and there exists a natural number $n>1$ such that $\frac{1}{n}\in R$, then $R$ has a maximal subring. It is shown that if $R$ is a reduced ring with $|R|>2^{2^{\aleph_0}}$ or $J(R)\neq 0$, then any $R$-algebra has a maximal subring. Residually finite rings without maximal subrings are fully characterized. It is observed that every uncountable $UFD$ has a maximal subring. The existence of maximal subrings in a noetherian integral domain $R$, in relation to either the cardinality of the set of divisors of some of its elements or the height of its maximal ideals, is also investigated. | math.RA | math |
SUBMAXIMAL INTEGRAL DOMAINS
A. AZARANG
Department of Mathematics, Chamran University, Ahvaz, Iran
a−[email protected]
Abstract. It is proved that if D is a U F D and R is a D-algebra, such that
U (R) ∩ D 6= U (D), then R has a maximal subring. In particular, if R is a ring
which either contains a unit x which is not algebraic over the prime subring of
R, or R has zero characteristic and there exists a natural number n > 1 such
that 1
n ∈ R, then R has a maximal subring. It is shown that if R is a reduced
ℵ0 or J(R) 6= 0, then any R-algebra has a maximal subring.
ring with R > 22
Residually finite rings without maximal subrings are fully characterized. It is
observed that every uncountable U F D has a maximal subring. The existence
of maximal subrings in a noetherian integral domain R, in relation to either
the cardinality of the set of divisors of some of its elements or the height of its
maximal ideals, is also investigated.
Introduction
All rings in this article are commutative with 1 6= 0; all modules are unital. If S
is a subring of a ring R, then 1R ∈ S. In this paper the characteristic of a ring R
is denoted by Char(R), and the set of all maximal ideals of a ring R is denoted by
M ax(R). For any ring R, let Z = Z · 1R = {n · 1R n ∈ Z}, be the prime subring of
R. Rings with maximal subrings are called submaximal rings in [4] and [7]. Some
important rings such as uncountable artinian rings, zero-dimensional rings which
are either not integral over Z or with zero characteristic, noetherian rings R with
R > 2ℵ0 and infinite direct product of rings are submaximal, see [4-7]. We should
remind the reader that all finite rings except Zn (up to isomorphism), where n is
a natural number, are submaximal. It is also interesting to note that whenever S
is a finite maximal subring of a ring R, then R must be finite, see [8, Theorem 8],
[19], [17] and [20]. The latter interesting fact is also an easy consequence of [5, the
proof of Theorem 2.9 ] or [6, Theorem 3.8]. Recently S.S. Korobkov determined
which finite rings have exactly two maximal subrings, see [18].
We remind the reader that whenever S is a maximal subring of a ring R, then
R is called minimal ring extension of S. Recently, D.E. Dobbs and J. Shapiro have
extended the results in [15], to integral domains and certain commutative rings,
see [13] and [14], respectively. Also see [23], [11] and [21]. T.G. Lucas, in [21],
characterized minimal ring extensions of certain commutative rings especially in
the case of minimal integral extension. It is interesting to know that every com-
mutative ring R has a minimal ring extension, for if M is a simple R-module then
the idealization R(+)M is a minimal ring extension of R (note, for any R-module
2000 Mathematics Subject Classification. 13B02, 13G05, 13E05.
Key words and phrases. Maximal Subring, Integral Domains, Noetherian, UFD.
1
2
A. AZARANG
M , every R-subalgebra of R(+)M has the form R(+)N , where N is a submodule
of M , see [12]). For a generalization of minimal ring extensions, see also [10].
Unlike maximal ideals (resp. minimal ring extension) whose existence is guar-
anteed either by Zorn Lemma or noetherianity of rings (resp. by idealization or
other techniques, see [12]), maximal subrings need not always exist, see [7] for such
examples and in particular, for example of rings of any infinite cardinality, which
are not submaximal. In fact by the above comment about the idealization, one can
easily see that if K is any field with zero characteristic, then the ring Z(+)K is not
submaximal, see [7, Example 3.19]. Also, in the latter reference and in [4] a good
motivations for the study of maximal subrings related to algebraic geometry and
elliptic curves are given.
In this paper, we are merely interested in finding submaximal integral domains,
especially atomic and noetherian integral domains. A brief outline of this paper is
as follow. Section 1, contains some preliminaries and also some generalizations of
results which are to be appeared in [7]. It is observed that whenever D is a U F D
and R is a D-algebra in which a non-unit of D is invertible, then R is submaximal.
In particular, if R is a Z-algebra such that Q ∩ R 6= Z, then R is submaximal.
Moreover, if D is a P ID and D ⊆ R is an integral domain such that D is integrally
closed in R and U (R) 6= U (D), then R is submaximal. Consequently it is proved
that, if R is a Z-algebra, then either R is submaximal or for any prime number p,
there exists a maximal ideal M of R such that Char( R
M ) = p. It is observed that
every ring either is submaximal or is Hilbert. In particular, if R is a reduced ring
with R > 22ℵ0 or J(R) 6= 0, then any R-algebra is submaximal. Consequently, it
is shown that if R is a reduced non-submaximal ring with zero characteristic, then
Tp∈P Rp = 0. It is proved that if R is a residue finite non-submaximal ring, then
R is a countable principal ideal ring which is either an integral domain with zero
characteristic or it is an artinian ring with nonzero characteristic. Finally in Section
1, the existence of maximal subring in semi-local rings and localization of rings are
investigated. In particular, it is proved that if R is a ring and S is a multiplicatively
closed set in R such that RS is semi-local, then either RS is submaximal or every
prime ideal of RS has the form PS, for some P ∈ M ax(R) ∩ M in(R). Moreover,
in the latter case, if RS is submaximal, then R is submaximal too. Section 2, is
devoted to the existence of maximal subrings in unique factorization domains, noe-
therian integral domains and certain atomic domains. It is observed that, every
uncountable U F D is submaximal. We also generalized the latter result to certain
uncountable atomic domains. In particular, it is proved that if R is an uncountable
noetherian Z-algebra, in which every natural number has at most countably many
(irreducible) divisors, then R is submaximal. It is shown that, if R is a noetherian
integral domain with zero characteristic and tr.degZR = n ≥ 1 (resp. with nonzero
characteristic and tr.degZ R = n ≥ 2) such that the height of every maximal ideal
of R is greater or equal to n + 1 (resp. greater or equal to n) and Z[X] ⊆ R is
a residually algebraic extension, where X is a transcendental basis for R over Z,
then R is submaximal. Finally, we show that every uncountable Dedekind domain
D with M ax(D) ≤ ℵ0, is submaximal.
SUBMAXIMAL INTEGRAL DOMAINS
3
local) if M ax(R) is finite (resp.
Next, let us recall some standard definitions and notations in commutative rings,
see [16], which are used in the sequel. An integral domain D is called G-domain if
the quotient field of D is finitely generated as a ring over D. A prime ideal P of
a ring R is called G-ideal if R
P is a G-domain. A ring R is called Hilbert if every
G-ideal of R is maximal. We also call a ring R, not necessarily noetherian, semi-
M ax(R) = 1). An integral domain
local (resp.
D is called atomic, if every nonzero non-unit of D is a finite product of irreducible
elements, not necessarily unique. An integral domain D is called idf -domain if
every nonzero non-unit element of D has at most finitely many irreducible divisors,
see [1]. In this paper the set of minimal prime ideals and prime ideals of a ring R
are denoted by M in(R) and Spec(R), respectively. As usual, let U (R) denote the
set of all units of a ring R. The Jacobson and the nil radical ideals of a ring R are
also denoted by J(R) and N (R), respectively. If P is a prime ideal of a ring R, then
the height of P is denoted by ht(P ). If D is an integral domain, then we denote the
set of all non-associate irreducible elements of D by Ir(D). We recall that if D ⊆ R
is an extension of integral domains, then as for the existence a transcendental basis
for field extensions, one can easily see that there exists a subset X of R which is
algebraically independent over D and R is algebraic over D[X] (hence every integral
domain is algebraic over a U F D). Moreover, in the latter case X = tr.degF (E),
where E and F are the quotient fields of R and D, respectively. Hence, similar to
the field extensions, we can define the transcendental degree of R over D which is
denoted by tr.degD(R). Finally, we denote the set of all natural prime numbers by
P.
1. Preliminaries and Generalizations
We begin this section with the following useful fact about the existence of max-
imal subrings in subrings of a submaximal ring, which is the converse of [6, Propo-
sition 2.1]. We remind the reader that a ring R is submaximal if and only if there
exist a proper subring S of R and an element x ∈ R \ S such that S[x] = R, see [3,
Theorem 2.5]. Now the following is in order, and although its proof is in [7], but
we present it for the sake of the reader.
Proposition 1.1. [7, Theorem 2.19]. Let R ⊆ T be rings. If there exists a maximal
subring V for T such that V is integrally closed in T and U (R) * V , then R is
submaximal.
Proof. First, we claim that whenever x ∈ U (R) \ V , then x−1 ∈ V . To see this,
we observe that x−1 ∈ R ⊆ T = V [x] (note, V is a maximal subring of T ).
Consequently, x−1 = a0 + a1x + · · · + anxn, where a0, a1, . . . , an ∈ V . Now by
multiplying the latter equality by x−n, we infer that x−1 is integral over V , hence
x−1 ∈ V . But U (R) * V implies that V ∩ R is a proper subring of R and there
exists x ∈ U (R) \ V with T = V [x]. Finally, we claim that R = (V ∩ R)[x], which
by the preceding comment, it implies that R is submaximal. To this end, let y ∈ R,
hence y ∈ V [x] and therefore y = b0 + b1x + · · · + bmxm, where b0, b1, . . . , bm ∈ V ,
implies that yx−m ∈ V ∩ R (note, x−1 ∈ V ), i.e., y ∈ (R ∩ V )[x] and we are
done.
(cid:3)
Next, we have the following fact which is needed in the sequel.
4
A. AZARANG
Theorem 1.2. Let R be a ring and D be a subring of R which is a U F D. If there
exists an irreducible element p ∈ D such that 1
p ∈ R, then R is submaximal. In
particular, if U (R) ∩ D 6= U (D), then R is submaximal.
Proof. We first prove the theorem by the assumption that R is algebraic over D.
We also may assume that R is an integral domain (note, if not, then there exists
a prime ideal Q of R such that D ∩ Q = 0 and therefore R
Q contains a copy of
D). Now suppose that K and E are the quotient fields of D and R, respectively.
Thus K/E is an algebraic extension, since R is algebraic over D. Now, note that
K has a maximal subring V such that 1
p /∈ V (for example V = D(p)). Hence E
has a maximal subring W such that W ∩ K = V , by [6, Proposition 2.1]. Therefore
1
p /∈ W . Thus we have U (R) * W which implies that R is submaximal by the
above proposition. Finally, assume that R is not algebraic over D, but by the
preceding comment we may suppose that R is an integral domain too. Let X be a
transcendental basis for R over D. Thus R is algebraic over D[X]. Now note that
D[X] is a U F D and p is an irreducible element in it. Hence we are done by the
first part of the proof. The final part is evident.
(cid:3)
The following fact also justifies the two cases proofs of [4, Proposition 2.10].
Remark 1.3. Let R be a ring satisfying the conditions of the above theorem, then
there exists a maximal subring of R which dose not contain 1
p . In particular, if
K is a field with zero characteristic, then for any prime number p, there exists a
maximal subring Vp of K such that 1
p /∈ Vp. Hence if M is the unique nonzero
prime ideal of Vp, we infer that Char( Vp
M ) = p.
The next three interesting facts are now immediate.
Corollary 1.4. Let R be a U F D and S be a multiplicatively closed subset of R
which contains a non-unit of R, then RS is submaximal.
Corollary 1.5. Let R be a ring with zero characteristic. If there exists a natural
number n > 1 such that 1
n ∈ R, then R is submaximal.
Corollary 1.6. Let D be an integral domain with zero characteristic and X be a
set of independent indeterminates over it. Then for any x ∈ X and every natural
number n > 1, the ring
(nx−1)D[X] is submaximal.
D[X]
Corollary 1.7. If R is a ring with 0 = Char(R) 6= Char( R
T is submaximal.
J(R) ), then any R-algebra
Proof. Assume that Char( R
1 − kn ∈ U (R) ⊆ U (T ) and therefore we are done by Corollary 1.5.
J(R) ) = n, thus n ∈ J(R). Hence for any k ∈ Z, we have
(cid:3)
Corollary 1.8. Let R be a ring with zero characteristic which is not submaximal.
Then {Char( R
M ) M ∈ M ax(R) } = P and therefore M ax(R) is infinite.
Proof. Since R is not submaximal, we infer that Char( R
M ) 6= 0 for each maximal
ideal M of R, by Corollary 1.5. Hence {Char( R
M ) M ∈ M ax(R) } ⊆ P. Now
for each prime number q we claim that there exists a maximal ideal M of R with
Char( R
M ) = q which proves the lemma. To see this, we note that qR 6= R, by
Corollary 1.5. Consequently, there exists a maximal ideal M of R with qR ⊆ M ,
i.e., Char( R
(cid:3)
M ) = q.
SUBMAXIMAL INTEGRAL DOMAINS
5
For more observations we need the following lemma.
Lemma 1.9. Let R be a ring and x ∈ R is not algebraic over the prime subring of
R. Then at least one of the following conditions holds.
(1) If Char(R) = 0, then there exists a prime ideal Q of R such that R/Q
contains a copy of Z[x].
(2) If Char(R) = n > 0, then for any prime divisor p of n, there exists a prime
ideal Q of R such that R/Q contains a copy of Zp[x].
Proof. If R has zero (or prime) characteristic, we are done since Z[x] \ {0} is a
multiplicatively closed set in R. Now, suppose that R has nonzero characteristic,
say n, which is also not a prime number. Assume that p is a prime divisor of n.
Since dim Zn[x] = 1 and P = pZ
nZ [x] is a non-maximal prime ideal of Zn[x], hence
we infer that P is a minimal prime ideal of Zn[x]. Thus, there exists a minimal
prime ideal Q of R such that Q ∩ Zn[x] = P . Now we have Zp[x] ∼=
Q∩Zn[x] ⊆ R
and therefore we are done.
Zn[x]
Q
(cid:3)
Remark 1.10. In fact in Corollary 1.8, we see that if R is not submaximal and
Z ⊆ R, then M ax(R) ≥ P. We can generalize the previous fact to any non-
submaximal ring which contains a U F D as follow. First, we recall the reader that
if R is a ring with M ax(R) > 2ℵ0, then R is submaximal, see [4, Proposition 2.6
]. Now assume that D is a U F D and let Ir′(D) be a subset of Ir(D) such that for
any p 6= q in Ir′(D), we have pD + qD = D. Now, if R is a ring which contains
D, then either R is submaximal or Ir′(D) ≤ M ax(R) ≤ 2ℵ0. To see this assume
that R is not submaximal, then for any q ∈ Ir′(D) we have qR 6= R, by Theorem
1.2 and hence there exists a maximal ideal Mq of R, such that qR ⊆ Mq.
It is
clear that whenever p 6= q in Ir′(D), then we have Mq 6= Mp, and therefore we are
done. In particular, if R is a non-submaximal ring with nonzero characteristic, say
n, which is not algebraic over Zn, then M ax(R) is infinite. To see this note that
by part (2) of the above lemma, for any prime divisor p of n there exists a prime
ideal Q of R such that R/Q contains a copy of Zp[x]. Hence we are done by the
first part of the proof.
The following proof greatly simplifies the proof of [7, Theorem 2.1 and Theorem
2.4].
Corollary 1.11. [7, Theorem 2.4]. Let R be a ring with a unit element x ∈ R
which is not algebraic over Z. Then R is submaximal (in fact every R-algebra is
submaximal).
Proof. In view of Theorem 1.2 and Lemma 1.9 we are done.
(cid:3)
Corollary 1.12. Let R be a ring. Then either R is submaximal or J(R) is algebraic
over Z.
We need the following immediate corollary in the next section.
Corollary 1.13. Let R be a ring which is not algebraic over Z. Then either R
is submaximal or for any non-algebraic element x ∈ R over Z and every natural
number n > 1, we have Z ∩ (nx − 1)R 6= 0.
Proof. If (nx − 1)R = R, then we are done by Corollary 1.11, and if not, then by
using Corollary 1.5, we are done.
(cid:3)
6
A. AZARANG
Corollary 1.14. Let D be a P ID and R ⊇ D be an integral domain.
integrally closed in R and U (R) 6= U (D), then R is submaximal.
If D is
Proof. Let x ∈ U (R) \ U (D). If x is not algebraic over D, then we are done, by
Corollary 1.11. Hence assume that x is algebraic over D, thus there exists b ∈ D
such that bx is integral over D and since D is integrally closed in R, we must have
bx = a ∈ D. Therefore x = a
b . Now, since x /∈ U (D), we infer that either x /∈ D or
x−1 /∈ D. Therefore, in any case, there must exist r, s ∈ D such that (r, s) = 1 and
z = r
s ∈ R. Thus we are done,
by Theorem 1.2.
(cid:3)
s ∈ U (R) \ D. Now since D is a P ID, we infer that 1
Lemma 1.15. Let R be a ring with nonzero characteristic n which is square free
(in particular, if R is reduced ring with nonzero characteristic). Then either R is
submaximal or U (R) is a torsion group.
Proof. Without lose of generality we may assume that Char(R) = p, where p is
a prime number. Now suppose that R is not submaximal, then U (R) must be
algebraic over Zp, by Corollary 1.11. Assume that x ∈ U (R), thus we infer that
Zp[x] ∼=
Zp[t]
I where I is a nonzero ideal of the polynomial ring Zp[t]. Hence we
infer that Zp[x] is a finite ring, and therefore x is a torsion element. Thus U (R) is
a torsion group.
(cid:3)
We recall the reader that zero dimensional rings (in particular von Neumann
regular rings) with zero characteristic are submaximal, see [6, Corollary 3.11]. We
also have the following.
Proposition 1.16. Let R be a von Neumann regular ring. Then either R is sub-
maximal or R is a periodic ring.
Proof. If R is not submaximal then by the above comment R has nonzero charac-
teristic. Hence by the above lemma U (R) is torsion. But it is well-known that von
Neumann regular rings are unit regular, that is to say, for any x ∈ R, there exists
u ∈ U (R) such that x = x2u. Hence by the above lemma, if un = 1, then we have
xn = x2n and thus we are done.
(cid:3)
In fact the above result holds for any zero-dimensional ring R. For proof note
that if R is not submaximal then R has nonzero characteristic, say n, and R is
integral over Zn, by [6, Corollary 3.14]. Now note that for any x ∈ R, the ring
Zn[x] is finite and hence we are done. The next remark shows in some rings R, the
group U (R) may not be torsion.
Remark 1.17. Let R be a ring. If R is von Neumann regular with zero characteristic
then clearly U (R) is not torsion, by the proof of the above proposition, since R is
not periodic. Also, if R is a ring with J(R) 6= 0, then U (R) is not torsion. To see
this note that if U (R) is torsion, then for any 0 6= x ∈ J(R), there exists a natural
number n such that (1 + x)n = 1. Hence we infer that x = 0, which is absurd.
By Corollary 1.12, if R is a ring then either R is submaximal or every element
of J(R) is algebraic over Z. Now we also have the following result.
Proposition 1.18. Let R be a ring with zero characteristic and J(R) 6= 0. Then
either R is submaximal or for any x ∈ J(R) and f (t) ∈ Z[t], if f (x) = 0, then
f (0) = 0. In particular J(R) consists of zero divisors.
SUBMAXIMAL INTEGRAL DOMAINS
7
Proof. Assume that R is not submaximal and x ∈ J(R), f (t) ∈ Z[t], and f (x) = 0.
Now since x ∈ J(R), we infer that if u is one of the elements 1 + f (0) or 1 − f (0),
then u ∈ U (R) ∩ Z. Thus by Corollary 1.5, we have u = 1 or u = −1. This implies
that either f (0) = 0, and therefore we are done, or f (0) ∈ {2, −2}. But in the
latter case, we have 2 ∈ J(R) and therefore 1 − 2n ∈ U (R), for each n ∈ Z, which
is impossible by Corollary 1.5.
(cid:3)
The following is a generalization of [7, Corollary 2.24].
Corollary 1.19. Let R be an integral domain with J(R) 6= 0. Then any R-algebra
T is submaximal. In particular, any algebra over a non-field G-domain is submax-
imal.
Proof. If R has nonzero characteristic or Char(R) = Char( R
J(R) ) = 0, then one
can easily see that J(R) is not algebraic over Z (note, if 0 6= x ∈ J(R) and
anxn + · · · + a1x + a0 = 0, where n ∈ N, ai ∈ Z and a0 6= 0, then we infer that
a0 ∈ J(R) which is absurd). Therefore U (R) is not algebraic over Z. Thus U (T )
is not algebraic over Z and therefore T is submaximal, by Corollary 1.11. Hence
we may assume that 0 = Char(R) 6= Char( R
J(R) ) and hence T is submaximal by
Corollary 1.7. The last part is now evident.
(cid:3)
Remark 1.20. One can prove the above corollary by using the proof of Proposition
1.18, Lemma 1.15 and Remark 1.17.
Proposition 1.21. Let R ⊆ T be an extension of commutative rings with lying-
over property and R is not Hilbert. Then T is submaximal.
Proof. Let P be a prime ideal in R such that P is not an intersection of a family
of maximal ideals in R. Now assume Q is a prime ideal in T lying over P . Thus
R/P ⊆ T /Q and since J(R/P ) 6= 0, we infer that T /Q is submaximal by Corollary
1.19.
(cid:3)
We recall the reader that if R is a ring with M ax(R) > 2ℵ0 , then R is submax-
imal, see [4, Proposition 2.6 ].
Corollary 1.22. Let R be a ring. Then either R is submaximal or it is a Hilbert
ring with Spec(R) ≤ 22ℵ0 .
Proof. If R is not submaximal, then for any prime ideal P of R, the integral domain
R/P is not submaximal too. Hence we infer that J(R/P ) = 0, by Corollary 1.19,
i.e., R is Hilbert and therefore P is an intersection of a set of maximal ideals of R.
Thus by the above comment we infer that Spec(R) ≤ 22ℵ0 .
(cid:3)
Remark 1.23. In fact if R is not submaximal, then for any prime ideal P and
subring S of R, the prime ideal P ∩ S is an intersection of a family of maximal
ideals of S. To see this note that R/P contains a copy of S/(P ∩ S), and since R
is not submaximal, we infer that J(S/(P ∩ S)) = 0, by Corollary 1.19. Hence we
are done.
Lemma 1.24. Let R be a ring. Then at least one of the following conditions holds,
(1) There exists a maximal ideal M of R, such that R/M is not an algebraic
extension of a finite field (i.e., R/M is not absolutely algebraic field). In
particular, R/M and therefore R are submaximal.
(2) For any subring S of R, we have J(S) ⊆ J(R).
8
A. AZARANG
Proof. If (1) does not hold, then for any maximal ideal M of R, the field R/M is
algebraic over a finite field. Hence we infer that every subring of R/M is a field.
Now note that if S is a subring of R, then (S + M )/M is a subring of R/M and
therefore (S + M )/M is a field. Thus S ∩ M is a maximal ideal of S, for any
maximal ideal M of R. This shows that J(S) ⊆ J(R). For the final part in (1),
note that by [4, Theorem 1.8 ], if R/M is not algebraic over a finite field, then R/M
and therefore R are submaximal.
(cid:3)
In [4, Proposition 2.9] it is proved that if R is a ring with R/J(R) > 22ℵ0 , then
R is submaximal.
Corollary 1.25. Let R be a reduced ring. If either J(R) 6= 0 or R > 22ℵ0 , then
R is submaximal. Moreover, every R-algebra T , is submaximal too.
Proof. If R is not submaximal, then by Corollary 1.22, R is Hilbert ring and there-
fore J(R) = N (R). Hence we infer that J(R) = 0 and by the above comment also
we have R ≤ 22ℵ0 which are absurd. For the final part note that T /N (T ) contains
a copy of R, hence by our assumptions, either by the above lemma J(T /N (T )) 6= 0,
or T /N (T ) > 22ℵ0 . Thus by the first part, T /N (T ) and therefore T are submaxi-
mal.
(cid:3)
Hence by the above corollary if T is a non-submaximal ring, then for any reduced
subring R of T we have J(R) = 0. More generally, for any subring R of T we have
N (R) = J(R). To see this, note that R + N (T ) is a subring of T . Now since
T /N (T ) contains a copy of R/N (R), we infer that J(R/N (R)) = 0, hence we are
done. The following is also interesting.
Corollary 1.26. Let R be a reduced ring with zero characteristic, then either R is
submaximal or Tp∈P Rp = 0.
Proof. If R is not submaximal then by Corollary 1.8, we infer that Tp∈P Rp ⊆ J(R).
But by the above corollary we also have J(R) = 0. Hence we are done.
(cid:3)
For a ring R, let CoHt1(R) be the set of all prime ideal P of R, such that if
P ( Q, where Q ∈ Spec(R), then Q ∈ M ax(R).
Proposition 1.27. Let R be a noetherian Hilbert ring, then any P ∈ CoHt1(R)
is an intersection of infinite countably many maximal ideals. Consequently, if R
is one dimensional noetherian (Hilbert) domain, then for any infinite family A of
maximal ideals of R, we have T A = 0. In particular,
(1) For any noetherian ring R, either R is submaximal or CoHt1(R) ≤ 2ℵ0.
(2) If R is noetherian domain and dim(R) = 2, then either R is submaximal
or Spec(R) ≤ 2ℵ0.
Proof. Let {Mi}i∈I be a family of maximal ideals of R such that P = Ti∈I Mi. We
first show that for any infinite countable subfamily {Mn}n∈N of {Mi}i∈I we have
P = Tn∈N Mn. Assume A = Tn∈N Mn, hence we infer that P ⊆ A ⊆ Mn, for all
n ∈ N. Hence, if P 6= A, then M in(R/A) is infinite which is a contradiction, thus
P = A and we are done. Now, if furthermore R is one dimensional domain, then
any infinite intersection of maximal ideals is zero, by the previous proof. To see
(1), assume that R is not submaximal, hence M ax(R) ≤ 2ℵ0, by the comment
preceding Corollary 1.22, and since any P ∈ CoHt1(R) is a countable intersection
of maximal ideals, we infer that CoHt1(R) ≤ 2ℵ0. Part (2) is now evident.
(cid:3)
SUBMAXIMAL INTEGRAL DOMAINS
9
We recall that each zero dimensional ring with nonzero characteristic which is
not integral over its prime subring, is submaximal, see [6, Corollary 3.14]. The
following is a generalization of the existence of maximal subrings in artinian rings,
see [5].
Corollary 1.28. Let R be a semi-local ring. Then either R is submaximal or
R has nonzero characteristic, say n, which is integral over Zn (thus R is zero-
dimensional). In particular, every semi-local ring with zero characteristic is sub-
maximal. Consequently,
(1) Non-submaximal semi-local integral domains are exactly non-submaximal
fields.
(2) Every non-submaximal noetherian semi-local ring, is countable artinian.
Proof. If Char(R) = 0, then we are done by Corollary 1.8. Hence assume that R
has nonzero characteristic. If R is not submaximal, then R is a Hilbert ring by
Corollary 1.22. Therefore every non-maximal prime ideal of R is an intersection of
infinitely many maximal ideals. Hence we infer that R is zero dimensional, since
M ax(R) < ℵ0. Now by the above comment we infer that R must be integral
over its prime subring. For part (1), we note that the prime subring of an integral
domain with nonzero characteristic is a field; and for (2) note that R is a zero-
dimensional ring. Hence R is artinian. Thus by [5, Proposition 2.4], R must be
countable too.
(cid:3)
Proposition 1.29. Let R1 ⊆ R2 be extension of rings. Assume that R1 is semi-
local. Then either R2 is submaximal or R1 is zero-dimensional. In other words,
every algebra over a semi-local ring which is not zero dimensional, is submaximal.
Proof. First note that, if P is a prime ideal of R2, then the ring R2/P contains a
copy of S = R1/(R1 ∩ P ). Hence if J(S) 6= 0, then R2/P and therefore R2 are
submaximal by Corollary 1.19. If not, then we infer that P ∩ R1 is a maximal ideal
of R1, since R1 is semi-local. Hence, we may assume that for any prime ideal P of
R2, R1 ∩ P is a maximal ideal in R1. Now, if Q is a prime ideal in R1, then there
exists a prime ideal P of R2 such that P ∩ R1 ⊆ Q (note, there exists a prime ideal
P of R2 with P ∩ (R1 \ Q) = ∅). Hence we infer that Q = P ∩ R1 and therefore Q
is maximal in R1. Hence R1 is a zero dimensional ring.
(cid:3)
We recall the reader that a ring R is called residue finite if R/I is a finite ring
for every nonzero ideal I of R. It is clear that if R is a residue finite ring, then
dim(R) ≤ 1 and in fact dim(R) = 1 if and only if R is a non-field integral domain.
In the next theorem we give the structure of non-submaximal residue finite rings.
Theorem 1.30. Let R be a residue finite ring which is not submaximal. Then R
is a countable principal ideal ring. Moreover, exactly one of the following holds:
(1) If dim(R) = 1, then R = U (R)Z and R is algebraic over Z.
(2) If dim(R) = 0, then R is an artinian ring with nonzero characteristic, say
n, which is also integral over Zn. Moreover, R has only finitely many ideals.
In particular, if R is reduced then R is finite.
Proof. First note that since R is not submaximal then for any nonzero ideal I of R
we infer that R/I ∼= Zm for some natural number m (note, it is clear that all finite
rings except Zn, up to isomorphism, where n is a natural number, are submaximal).
This shows that I = Rm and therefore R is a principal ideal ring. Now by the above
10
A. AZARANG
comment we have two cases, either dim(R) = 1 or dim(R) = 0. First assume that
dim(R) = 1 and therefore R is a non-field integral domain. Hence we have two
cases.
(1) If R has nonzero characteristic, say p (where p ∈ P), then we infer that
for any nonzero ideal I of R we have R/I ∼= Zp, which is absurd, by the
preceding comment (note, in this case R ∼= Zp which is impossible).
(2) If R has zero characteristic. Then R is a P ID with Ir(R) = P, by the
first part of the proof. Hence we infer that R = U (R)Z. Also note that by
Corollary 1.11, U (R) is algebraic over Z. Therefore U (R) is countable and
hence R is countable too. Thus we are done.
Now assume that R is zero-dimensional ring. Thus R is artinian, since R is
noetherian (note, every ideal of R is principal). Thus by [5, Proposition 2.4], R is
countable and has a nonzero characteristic, say n, which is also integral over Zn, by
[5, Corollary 2.5]. Moreover, by the first part of the proof since every nonzero ideal
of R has the form I = Rm where mn, we infer that R has only finitely many ideals.
Also note that if R is reduced, then by Corollary 1.25, we infer that J(R) = 0 and
therefore R ∼= Zn (where n is square free) and hence we are done.
(cid:3)
Proposition 1.31. Let D be an integral domain and S be a multiplicatively closed
set in it such that S * U (D). If DS is not submaximal then the following conditions
hold.
(1) D has zero characteristic. S is algebraic over Z and therefore S ≤ ℵ0. In
particular Z is not integrally closed in DS.
(2) There exists an infinite subset M of M ax(D) such that M ax(DS) = {QS Q ∈
M}. In particular, T M = 0.
(3) For any non-maximal prime ideal PS of DS, either Char( D
P ) = 0 or DS
PS
∼=
D
P .
∼= ( D
P ) ¯S = F rac( D
Proof. Since DS is not submaximal then by Corollary 1.11, we infer that S is
algebraic over Z. Hence if Char(D) 6= 0, then S ⊆ U (D) which is absurd. Thus
D has zero characteristic. Hence by Corollary 1.14, Z is not integrally closed in
DS. Now assume PS is a maximal ideal in DS and P ∈ Spec(D) \ M ax(D). Thus
we have DS
P ) is
PS
the quotient field of D
P ) is
submaximal by Corollary 1.11, and therefore DS is submaximal which is absurd.
Hence P ∈ M ax(D). Also, note that by Corollary 1.28, M ax(DS) is infinite since
DS is not submaximal; and by Corollary 1.19 we have J(DS) = 0 and therefore
T M = 0. Finally, for part (3), assume that Char( D
P ) = q > 0, then by part (1),
either ¯S ⊆ U ( D
P and we are done; or ¯S * U ( D
P ) and therefore DS
P )
and therefore ( D
P ) ¯S is submaximal. Thus DS is submaximal which is absurd. (cid:3)
P ), where ¯S = {s + P s ∈ S} and F rac( D
P , and since P is not maximal we infer that F rac( D
P ) ¯S = D
∼= ( D
PS
∼= D
Note that in the above proposition clearly for any maximal ideal QS of DS we
also have DS
Q . More generally, if R is a ring and S be a multiplicatively
QS
closed set in R, then the non-submaximality of RS implies that every maximal
ideal of RS has the form PS for some maximal ideal P of R, by the preceding proof.
In particular, if R has nonzero characteristic then one can easily see that, by a
similar proof, for every prime ideal PS of RS we have RS
P . The following is a
PS
generalization of [7, Theorem 3.2 ].
∼= R
SUBMAXIMAL INTEGRAL DOMAINS
11
Theorem 1.32. Let R be a ring and S be a multiplicatively closed set in R, such
that RS is semi-local. Then at least one of the following holds.
(1) RS is submaximal.
(2) Spec(RS) = M ax(RS) = {PS P ∈ M}, where M is a finite subset of
M in(R) ∩ M ax(R).
In particular if (2) holds and RS is submaximal, then R is submaximal too.
Proof. Assume that RS is not submaximal, then by the above comment M ax(RS) =
{PS P ∈ M}, where M is a finite subset of M ax(R). But since RS is a semi-local
non-submaximal ring, then we infer that RS is zero-dimensional, by Corollary 1.28.
Hence M ⊆ M in(R). Now assume that (2) holds and RS is submaximal. Thus by
[7, Theorem 2.26], at least one of the following holds (note RS is zero-dimensional).
(1) There exists a maximal ideal PS of RS, such that RS/PS is submaximal.
But, since RS/PS ∼= R/P , we infer that R/P and therefore R are submax-
imal.
(2) There exist distinct maximal ideals PS and QS of RS such that RS/PS ∼=
RS/QS. Hence similar to (1), we infer that R/P ∼= R/Q and therefore R
is submaximal, by [3, Theorem 2.2].
(3) There exist an ideal IS and a maximal ideal PS of RS, such that (PS)2 ⊆
IS ⊆ PS and RS/IS ∼= K[x]/(x2), for some field K. Hence we infer that
I is a P -primary ideal in R. Therefore R/I is a local ring with unique
prime ideal P/I. Thus RS/IS ∼= (R/I) ¯S
∼= (R/I)P/I = R/I, where ¯S =
{s + I s ∈ S}, see [16, P. 24, Ex. 7]. Hence R/I and therefore R are
submaximal.
(cid:3)
The following remark which is a generalization of [4, Corollary 1.15] is interesting.
Remark 1.33. Let F be the set of all fields, up to isomorphism, which are not
submaximal and let D be the class of all integral domains (or reduced rings), up to
isomorphism, which are not submaximal. Now for any D ∈ D we have the following
facts:
(1) For any M ∈ M ax(D), we have D/M ∈ F .
(2) For any M, N ∈ M ax(D), with M 6= N we have D/M ≇ D/N (For
otherwise, by [3, Theorem 2.2], D is submaximal). In other words there
exists an injection ΦD from M ax(D) into F , sending M into D/M .
(3) M ax(D) ≤ F, by [4, Proposition 2.6] or (2).
(4) J(D) = 0, by Corollary 1.19 or 1.25.
(5) Hence we have the natural rings embedding D ֒→ QM∈Max(D) D/M ֒→
QE∈F E (i.e., every non-submaximal integral domain (or reduced ring)
can be embedded in QE∈F E).
Now, for any D ∈ D, let RdM ax(D) = Im(ΦD). Two non submaximal integral
domains D1 and D2 are called RdM ax-equivalent, if RdM ax(D1) = RdM ax(D2).
Now, let D′ be the set of equivalent classes of this relation. We claim that D′ ≤
2F = 22ℵ0 and F ⊆ D′. To show this, it is clear that F ⊆ D′. Also note that
for any [D] ∈ D′, the function that send [D] into RdM ax(D) is well-defined and
one-one from D′ into P (F ), the set of all subsets of F . Hence we are done, since
F = 2ℵ0 , by [4, Corollary 1.15].
12
A. AZARANG
2. Submaximal Integral Domains
In [5, Corollary 1.3], it is proved that uncountable fields, are submaximal. The
following interesting result is a generalization of this fact.
Theorem 2.1. Let R be an uncountable U F D, then R is submaximal.
Proof. If U (R) is uncountable, then we are done by Corollary 1.11. Hence we may
assume that U (R) is countable. Thus Ir(R) = R. Now, note that there exists
a p ∈ Ir(R) such that 1 − p /∈ U (R). Hence let p be an element in Ir(R), such
that 1 − p /∈ U (R) and q0 ∈ Ir(R) such that q01 − p. Thus q0 ∈ A = {q ∈
Ir(R) pR + qR = R }. Now we show that A must be an uncountable set. Let us
assume that A is countable and put B = {pq + 1 q ∈ Ir(R) \ A}. It is clear that B
is an uncountable set and therefore there exists a non-unit element x ∈ B such that
x has an irreducible divisor q′ ∈ Ir(R) \ A (note, U (R) and A are countable, thus
the set of all elements which are of the form uq1 · · · qn, where u ∈ U (R) and qi ∈ A,
n ∈ N ∪ {0} must be a countable set). Hence q′R + pR = R and q′ /∈ A, which
is a contradiction. Thus A must be uncountable. Now for any q ∈ A, p + (q) is a
unit in the ring R/(q), hence if there exists q ∈ A such that p + (q) is not algebraic
over the prime subring of R/(q), then by Corollary 1.11, R/(q) and therefore R are
submaximal. Consequently, we may assume that for any q ∈ A, p + (q) is algebraic
over the prime subring of R/(q). Thus for any q ∈ A, Z[p] ∩ (q) 6= 0, where Z is the
prime subring of R. But Z[p] is a countable set and the set {(q)}q∈A is uncountable,
thus there exists a nonzero element f ∈ Z[p] which belongs to an infinite (in fact
uncountable) number of (q), where q ∈ A, which is a contradiction. This proves
the theorem.
(cid:3)
Corollary 2.2. Let R be a non-submaximal non-field P ID, then R is countable
and Ir(R) = R.
Proof. By the above theorem, R and Ir(R) are countable. Now note that if Ir(R)
is finite then R is a G-domain and therefore R is submaximal by Corollary 1.19,
hence we are done.
(cid:3)
Corollary 2.3. Every localization of an uncountable U F D is submaximal.
Proposition 2.4. Let D be an uncountable atomic (or noetherian) domain. As-
sume that there exists an irreducible element p of D such that 1 − p /∈ U (D) and
every element of Z[p] has at most countably many (irreducible) divisors. Then D is
submaximal. In particular, if D is an uncountable atomic (or noetherian) domain
such that every element of it has at most countably many (irreducible) divisors,
then D is submaximal. Consequently, every uncountable noetherian idf -domain is
submaximal.
Proof. Note that any noetherian integral domain is an atomic domain, and by
using the proof of the previous theorem word-for-word, one can easily complete the
proof.
(cid:3)
Theorem 2.5. Let R be an uncountable atomic (or noetherian) integral domain
with zero characteristic. If every n ∈ N has at most countably many (irreducible)
divisors, then R is submaximal.
Proof. We may assume that U (R) is algebraic over Z and therefore it is count-
able, by Corollary 1.11. Hence we infer that Ir(R) = R and therefore Ir(R) is
SUBMAXIMAL INTEGRAL DOMAINS
13
uncountable. Let X be a transcendental basis for R over Z. Now, if there exist a
natural number n > 1 and x ∈ X such that Z∩(nx−1)R = 0, then R is submaximal
by Corollary 1.13. Hence we may assume that Z ∩ (nx − 1)R 6= 0 for any natural
number n > 1 and x ∈ X. Since X is uncountable and the number of ideals of Z is
countable, we infer that there exists an uncountable subset Y of X such that for any
y ∈ Y we have Z ∩ (ny − 1)R = mZ for some fixed natural numbers n > 1 and m.
Hence for any y ∈ Y we have ny −1m. Now we show that m has uncountably many
irreducible divisors. Assume that P = {q ∈ Ir(R) : qny − 1, for some y ∈ Y }. If
P is countable, then we infer that {ny − 1 : y ∈ Y } is countable too (note U (R)
is countable) which is a contradiction. Hence P is uncountable. Now note that
any q ∈ P is an irreducible divisor of m, i.e., m has uncountable many irreducible
divisors, which is a contradiction. Hence, for any natural number n > 1, the set
{x ∈ X Z ∩ (nx − 1)R 6= 0} is countable, and therefore R is submaximal, by
Corollary 1.13.
(cid:3)
For more observations we need the following definition, see [2].
Definition 2.6. An extension R ⊆ T of rings is called residually algebraic exten-
sion, if for any prime ideal Q of T , the ring T /Q is algebraic over R/(Q ∩ R).
One can easily see that if R ⊆ T is a residually algebraic extension then T
must be algebraic over R. Also see [2] for more interesting results about residually
algebraic extensions. In particular, see [2, Section 4, b-Maximal subrings] which
contains interesting results related to the subject of this paper. The following
lemma is needed for the next theorem.
Lemma 2.7. Let R ⊆ T be a residually algebraic extension of rings and R has
finite dimension. Then T has finite dimension too and we have dim(T ) ≤ dim(R).
Proof. Assume that n = dim(R). First suppose that T is an integral domain, and
we prove the lemma by induction on n. If n = 0, then R is a field and therefore
T is a field too, hence we are done. Thus assume that n ≥ 1 and the lemma
holds for any residually algebraic extension (of integral domains) R ⊆ T with
dim(R) < n. Now assume that R ⊆ T is a residually algebraic extension of integral
domains with dim(R) = n. Hence for any nonzero prime ideal Q of T , the extension
R/(Q ∩ R) ⊆ T /Q is also a residually algebraic extension of integral domains and
dim(R/(Q ∩ R)) < n (note that T is algebraic over R and Q 6= 0, hence Q ∩ R 6= 0).
Hence we infer that dim(T /Q) < n. This immediately implies that dim(T ) ≤ n
and therefore we are done. Now assume that R ⊆ T be any residually algebraic
extension, dim(R) = n and Q be a prime ideal of T . Hence R/(Q ∩ R) ⊆ T /Q
is a residually algebraic extension of integral domains and dim(R/(Q ∩ R)) ≤ n.
Thus by the first part of the proof we infer that dim(T /Q) ≤ n and since the latter
inequality holds for any prime ideal Q of T , we must have dim(T ) ≤ n. Therefore
we are done.
(cid:3)
The following is now in order.
Theorem 2.8. Let R be a noetherian integral domain with tr.degZ (R) = n < ℵ0
and assume that X is a transcendental basis for R over Z. Moreover let Z[X] ⊆ R
be a residually algebraic extension and at least one of the following holds.
(1) If Char(R) = 0 and n ≥ 1, then for any maximal ideal M of R we have
ht(M ) ≥ n + 1.
14
A. AZARANG
(2) If Char(R) = p > 0 and n ≥ 2, then for any maximal ideal M of R we
have ht(M ) ≥ n.
Then R is submaximal.
Proof. (1) Let x ∈ X, if R(2x − 1) = R, then we are done, by Corollary 1.11.
Hence assume that R(2x − 1) 6= R and P be a prime ideal of R which is min-
imal over R(2x − 1). Thus by the Krull's principal ideal theorem we infer that
ht(P ) = 1 and therefore by our assumption P is not a maximal ideal in R.
Hence (0) ( (2x − 1)Z[X] ⊆ P ∩ Z[X]. Thus we have two cases. First,
if
(2x − 1)Z[X] = P ∩ Z[X], then
P ), hence we
are done, by Corollary 1.5. Thus we may assume that Q = P ∩Z[X] 6= (2x−1)Z[X].
Therefore ht(Q) ≥ 2 and since dim(Z[X]) = n + 1, we infer that dim( Z[X]
Q ) ≤ n − 1.
But Z[X]
P is a residually algebraic extension, hence by the above lemma con-
clude that dim( R
P ) ≤ n − 1. Now since ht(P ) = 1, the latter inequality immediately
implies that ht(M ) ≤ n, for any maximal ideal M ⊇ P , which is absurd. Thus we
are done.
P and therefore 1
(2x−1)Z[X] ⊆ R
Q ⊆ R
Z[X]
2 ∈ U ( R
Zp[X]
(xy−1)Zp[X] ⊆ R
(2) Let x, y ∈ X and x 6= y. If R(1 − xy) = R, then we are done by Corollary
1.11. Hence assume that R(1 − xy) 6= R and P be a prime ideal of R which is
minimal over R(xy − 1). Thus by the Krull's principal ideal theorem we infer that
ht(P ) = 1 and therefore by our assumption P is not a maximal ideal in R. Hence
(0) ( (xy −1)Zp[X] ⊆ P ∩Zp[X]. Thus we have two cases. First, if (xy −1)Zp[X] =
P ∩ Zp[X], then
P ), hence
we are done by Corollary 1.11 (note, x + (xy − 1)Zp[X] is not algebraic over Zp,
since Zp[X] is a U F D). Thus assume that Q = P ∩ Zp[X] 6= (xy − 1)Zp[X].
Therefore ht(Q) ≥ 2 and since dim(Zp[X]) = n we infer that dim( Zp[X]
Q ) ≤ n − 2.
But Zp[X]
Q ⊆ R
P is a residually algebraic extension, hence by the above lemma we
infer that dim( R
P ) ≤ n − 2. Now since ht(P ) = 1, the latter inequality immediately
implies that ht(M ) ≤ n − 1, for any maximal ideal M ⊇ P , which is absurd. Thus
we are done.
(cid:3)
P and therefore x + (xy − 1)Zp[X] ∈ U ( R
Proposition 2.9. Let a non-singleton X 6= ∅ be a set of independence indeter-
minates in a noetherian ring R over Z, where Z is the prime subring of R. If
Char(R) ∈ P ∪ {0} and R is integral over Z[X], then R is submaximal.
Proof. Let x, y ∈ X and x 6= y. If R(1 − xy) = R, then we are done. Hence assume
that P is a minimal prime ideal of R(1 − xy). Hence ht(P ) ≤ 1, by the Krull's
principal ideal theorem. Thus ht(P ∩ Z[X]) ≤ 1. But (1 − xy)Z[X] is a prime ideal
in Z[X], which is contained in P ∩ Z[X]. So we infer that (1 − xy)Z[X] = P ∩ Z[X]
and therefore T = Z[X]/(1 − xy)Z[X] ⊆ R/P . Now ¯x and ¯y are units in T , which
are not algebraic over the prime subring of T (note, Z[X] is a U F D). Hence R/P
has unit elements which are not algebraic over its prime subring and therefore we
are done, by Corollary 1.11.
(cid:3)
We conclude this article with the following fact about Dedekind domains.
Proposition 2.10. Let D be an uncountable Dedekind domain with countable set
of maximal ideals. Then U (D) is uncountable. In particular, D is submaximal.
SUBMAXIMAL INTEGRAL DOMAINS
15
Proof. Let U (D) be countable and seek a contradiction. It is now clear that Ir(D)
is uncountable. Hence we infer that the set of principal ideals of D is uncountable.
But since D is a Dedekind domain, every nonzero ideal of D is a finite product of
prime ideals. Since the set of prime ideals of D is countable, we infer that the set
of ideals D is countable too, which is a contradiction. Thus U (D) is uncountable
and therefore we are done, by Corollary 1.11.
(cid:3)
References
[1] D.D. Anderson and Bernadette Mullins, Finite factorization domains, Proc. A.M.S. 124 (2)
(2008) 389-396.
[2] A. Ayache and A. Jaballah, Residually algebraic pairs of rings, Math Z. 225 (1997) 49-65.
[3] A. Azarang, On maximal subrings, Far East J. Math. Sci (FJMS) 32 (1) (2009) 107-118.
[4] A. Azarang, O.A.S. Karamzadeh, Which fields have no maximal subrings?, Rend. Sem.
Mat. Univ. Padova, 126 (2011) 213-228.
[5] A. Azarang, O.A.S. Karamzadeh, On the existence of maximal subrings in commutative
artinian rings, J. Algebra Appl. 9 (5) (2010) 771-778.
[6] A. Azarang, O.A.S. Karamzadeh, On Maximal Subrings of Commutative Rings, to appear
in Algebra Colloquium.
[7] A. Azarang, O.A.S. Karamzadeh, Most Commutative Rings Have Maximal Subrings, to
appear in Algebra Colloquium.
[8] H.E. Bell and F. Guerriero, Some condition for finiteness and commutativity of rings, J.
Math. Math. Sci. 13 (3) (1990) 535-544.
[9] H.E. Bell and A.A. Klein, On finiteness of rings with finite maximal subrings, J. Math.
Math. Sci. 16 (2) (1993) 351-354.
[10] P.J. Cahen, D.E. Dobbs and T.J. Lucas, Valuative domains, J. Algebra Appl. 9 (1) (2010)
43-72.
[11] P.J. Cahen, D.E. Dobbs and T.J. Lucas, Characterizing minimal ring extensions, Rocky
Mountain Journal of Math 41 (4) (2011) 1081-1125.
[12] David E. Dobbs, Every commutative ring has a minimal ring extension, Comm. Algebra 34
(2006) 3875-3881.
[13] D.E. Dobbs and J. Shapiro, A classification of the minimal ring extensions of an integral
domain, J. Algebra, 305 (2006), 185-193.
[14] D.E. Dobbs and J. Shapiro, A classification of the minimal ring extensions of certain com-
mutative rings, J. Algebra, 308 (2007), 800-821.
[15] D. Ferrand, J.-P. Olivier, Homomorphismes minimaux danneaux, J. Algebra 16 (1970)
461-471.
[16] I. Kaplansky, Commutative Rings, Revised edn. (University of Chicago Press, Chicago
1974).
[17] Abraham A. Klein, The finiteness of a ring with a finite maximal subrings, Comm. Algebra
21 (4) (1993) 1389-1392.
[18] S.S. Korobkov, Finite Rings with Exactly Two Maximal Subrings, Russian Mathematics
(Iz. VUZ), 55 (6) (2011) 4652.
[19] Thomas J. Laffey, A finiteness theorem for rings, Proc. R. Ir. Acad. 92 (2) (1992) 285-288.
[20] T. Kwen Lee, K. Shan Liu. Algebra with a finite-dimensional maximal subalgebra, Comm.
Algebra 33 (1) (2005) 339-342.
[21] T.G. Lucas, Minimal integral ring extensions, Journal of Comm Alg. 3 (1) (2011) 47-81.
[22] M. L. Modica, Maximal Subrings, Ph.D. Dissertation, (University of Chicago, 1975).
[23] Gabriel Picavet and Martine Picavet-L'Hermitte, About minimal morphisms, in: Multiplica-
tive ideal theory in commutative algebra, Springer-Verlag, New York, (2006) pp. 369-386.
|
1705.01562 | 2 | 1705 | 2018-04-07T16:15:52 | Hermitian and skew hermitian forms over local rings | [
"math.RA"
] | We study the classification problem of possibly degenerate hermitian and skew hermitian bilinear forms over local rings where 2 is a unit. | math.RA | math |
HERMITIAN AND SKEW HERMITIAN FORMS OVER LOCAL RINGS
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
Abstract. We investigate the structure of possibly degenerate ε-hermitian forms over
local rings. We prove classification theorems in the cases where the ring is complete and
either the form is nondegenerate or the ring is a discrete valuation ring. In the latter case
we describe a complete set of invariants for such forms based on a generalisation of the
classical notion of the radical of the form.
1. Introduction
We are concerned with the classification problem of possibly degenerate hermitian or
skew hermitian bilinear forms over a local ring in which 2 is a unit. Symmetric and skew
symmetric forms are included as a special case, as we allow the underlying involution to
be trivial.
In foundational papers of Durfee ([7]) and O'Meara ([14]) the integral theory of quadratic
forms over local fields is developed. This is equivalent to the theory of symmetric bilinear
forms over a complete local principal ideal domain with finite residue field. Jacobowitz
([8]) extended this integral theory to the case of hermitian forms over local fields equipped
with an involution.
More recently, Levchuk and Starikova ([12]) have proved the existence and uniqueness
of normal forms for symmetric matrices over local principal ideal domains under certain
assumptions on the unit group of the ring.
In a wider setting Bayer-Fluckiger and Fainsilber ([1]) have considered the general prob-
lem of equivalence of hermitian or skew hermitian forms over arbitrary rings and are able to
prove some quite general reduction theorems for this problem. A broad study of sesquilin-
ear forms and their connection to hermitian forms was carried out by Bayer-Fluckiger,
First and Modovan ([2]) as well as by Bayer-Fluckiger and Moldovan ([3]).
For general information on quadratic and hermitian forms over rings we refer the reader
to the textbook of Knus ([9]).
1.1. Outline of the paper. Given the extensive existing literature on this topic and the
fact that we aim to provide a unified and self contained exposition, some overlap with
previously known results is inevitable.
In this section we summarise the results of the
paper, emphasising the contributions that we have made.
Throughout, A is a commutative local ring with maximal ideal r and ∗ is an involution of
A, that is to say an automorphism of order at most two. Clearly, ∗ induces an involution
2010 Mathematics Subject Classification. 11E08, 11E39, 15A63.
Key words and phrases. local ring, discrete valuation ring, hermitian, skew hermitian.
1
2
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
on the residue field A/r. If this induced involution is trivial, we say that ∗ is ramified,
otherwise it is unramified - this distinction will play a decisive role in the sequel. Thus
broadly speaking there are four types of forms to consider, ramified hermitian, ramified
skew hermitian, unramified hermitian and unramified skew hermitian. However, the last
two can be merged into one case since they differ only up to multiplication by a unit.
In Section 2 we present some elementary results on nondegenerate forms of rank one or
two in various cases. These essentially form the building blocks for our later classification
theorems. In light of this we include all the details so as to make the paper self contained.
In Section 3 we introduce completeness and use it to derive the existence of basis vectors
with certain specified properties. A key new result here is Lemma 5 which guarantees the
existence of a symplectic pair under appropriate conditions. We also give an example to
show that this lemma can fail in the absence of completeness.
In Section 4 we use the results of the previous sections to analyse nondegenerate forms
of any rank over arbitrary local rings. In this case there is an associated vector space and
form over the residue field and in the case that the ring is also complete we show that
equivalence of forms over the local ring reduces to equivalence of these associated forms
over the residue field. In particular we demonstrate the existence of a symplectic basis for
a nondegenerate skew hermitian form over a complete local ring with ramified involution
- see Proposition 1.
The remaining sections of the paper deal with possibly degenerate forms. In this case it
is not true without further assumptions that such forms have a nice - in a sense precisely
defined in Section 5 - decomposition. However in the case of a discrete valuation ring, that
is to say a local principal ideal domain, we are able to prove that any form has such a nice
decomposition - this is the content of Theorem 3.
Following this we introduce a generalisation of the classical notion of the radical of a
form. In our case we have one generalised radical module for each nonnegative integer.
Using these generalised radicals we are able to show that the equivalence problem for a
pair of forms over a complete discrete valuation ring can be reduced to the equivalence of
a series of pairs of nondegenerate forms over the associated residue field - see Theorem 4.
This is a broad generalisation of Theorem D of [4] which dealt only with the special case
of symmetric matrices.
Under some additional hypotheses, these spaces essentially turn out to give rise to a
complete set of invariants for hermitian and skew hermitian forms over a complete discrete
valuation ring. This can be found in Theorem 6.
Also under these hypotheses, we are able to prove, in Theorems 7 and 9, the existence and
uniqueness of normal forms in all possible cases. As a consequence of these normal forms
we also prove that the congruence class of a hermitian or skew hermitian matrix, over a
complete local principal ideal domain satisfying the aforementioned additional hypotheses,
depends only on the invariant factors of the matrix - this is stated in Theorem 8. Results
analogous to this last theorem have been obtained elsewhere, for polynomial rings over
algebraically closed fields see Theorem 4.5 of [6], and for skew symmetric matrices over
principal ideal domains see Exercise 4 of Chapter XIV of [11].
HERMITIAN AND SKEW HERMITIAN FORMS OVER LOCAL RINGS
3
1.2. Terminology and notation. We shall (sometimes) use the shorthand a for the
residue class a + r. Moreover, we extend this shorthand to matrices in the obvious way.
We write U(A) for the multiplicative group of units of A and recall the fundamental fact
that, since A is local, r = A − U(A).
Let R be the subring of A consisting of elements fixed by ∗. Observe that R is also a
local ring, with maximal ideal m = R ∩ r. Let S = {a ∈ A : a∗ = −a}. Since 2 is assumed
to be a unit of A we have A = R ⊕ S. If S ⊂ r we say that ∗ is ramified. Otherwise we
say that it is unramified. Since r is ∗-invariant, ∗ induces an involution on the residue field
A/r, which we also denote by ∗. As indicated in the introduction, ∗ is ramified if and only
if the induced involution on the residue field is trivial.
We write Mm(A) for the set of m × m matrices over A. The group of invertible m × m
Let V be a free A-module of rank m > 0 and let h : V × V → A be an ε-hermitian
form, where ε = ±1. That is to say h is bi-additive, A-linear in the second variable and
h(v, u) = εh(u, v)∗ for all u, v ∈ V .
The Gram matrix of a list of vectors v1,· · · , vk with respect to a form h is the k × k
matrix whose (i, j)-entry is h(vi, vj). The form h is said to be nondegenerate if the Gram
matrix of any basis of V belongs to GLm(A).
matrices is denoted by GLm(A).
A matrix C ∈ Mm(A) is said to be ε-hermitian if C ′∗ = εC where C ′ denotes the
transpose of A. Of course h is ε-hermitian if and only if G is ε-hermitian, where G is the
Gram matrix of any basis. On the other hand, given an ε-hermitian C ∈ Mm(A), we can
construct an ε-hermitian form hC on V = Am, by hC(u, v) = u′∗Cv.
Matrices C, D ∈ Mm(A) are said to be ∗-congruent if there exists X ∈ GLm(A) such
that D = X ′∗CX. Given an ε-hermitian form h, then the Gram matrices of any two bases
are ∗-congruent. On the other hand we say that forms h1 on V1, respectively h2 on V2 are
equivalent if there is an A-isomorphisn ϕ : V1 → V2 such that h1(u, v) = h2(ϕ(u), ϕ(v)).
Now, hC and hD as above are equivalent if and only C and D are ∗-congruent.
2. Small nondegenerate submodules
In contrast to the field case, a linearly independent set is not necessarily a subset of a
basis - this is apparent even in the rank one case. In the presence of the form h, however,
we have some sufficient conditions.
Given an A-submodule W of V , let W ⊥ = {v ∈ V : h(w, v) = 0}. Then W ⊥ is an
A-submodule of V .
Lemma 1. Let u ∈ V and suppose that h(u, u) ∈ U(A). Then V = Au ⊕ Au⊥. Moreover,
both Au and Au⊥ are free A-modules.
Proof. Let v1,· · · , vm be a basis of V . So u = P civi. Now h(u, u) = P cih(u, vi) ∈ U(A).
Since A is local this implies that cj ∈ U(A) for some j. Without loss of generality assume
that c1 ∈ U(A). Therefore u, v2, . . . , vm is a basis of V . Now, for i = 2,· · · m, let wi =
vi − h(u, vi)h(u, u)−1u. Clearly u, w2,· · · , wm is a basis of V and w2,· · · , wm is a basis of
Au⊥.
(cid:3)
4
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
In the case that ∗ is ramified and h is skew hermitian, it is immediate that h(u, u) ∈ r
for all u ∈ V . So we cannot have any nondegenerate rank one submodules in that case.
However we have the following lemma concerning rank two submodules.
Lemma 2. Suppose that ∗ is ramified and that h is skew hermitian. If u, v ∈ V satisfy
h(u, v) ∈ U(A), then V = (Au + Av) ⊕ (Au + Av)⊥. Moreover both summands in this
decomposition are free A-modules.
Proof. Let v1,· · · , vm be a basis of V . So u = Pi aivi. Suppose that ai ∈ r for all i.
Then clearly h(u, v) ∈ r, contradicting our assumption. So without loss of generality
we may assume that a1 ∈ U(A).
It follows that u, v2,· · · , vn is a basis of V . Write
v = b1u+Pm
j=2 bjvj. If b2,· · · , bn ∈ r, then since h(u, u) ∈ r it would follow that h(u, v) ∈ r,
contradicting our hypothesis. Without loss of generality we may assume that b2 ∈ U(A)
and it follows that u, v, v3,· · · , vm is a basis of V . Since the Gram matrix of u, v is
invertible in this case it follows that, for i = 3,· · · , m, there exist (unique) ai, bi ∈ A such
that h(u, aiu + biv) = h(u, vi) and h(v, aiu + biv) = h(v, vi). Let wi = vi− (aiu + biv). Then
u, v, w3,· · · , wm is a basis of V and w2,· · · , w3 is a basis of (Au + Av)⊥ as required.
(cid:3)
Now we show, excepting the ramified skew hermitian case, that if h has any unit value,
then there is some element w ∈ V such that h(w, w) is a unit.
Lemma 3. Let u, v ∈ V . Suppose that h(u, v) ∈ U(A) and that either h is hermitian or
that ∗ is unramified. Then there is some w ∈ V such that h(w, w) ∈ U(A).
Proof. Replacing v by h(u, v)−1v if necessary, we can assume without loss of generality
that h(u, v) = 1. Now suppose that h is hermitian. Then
2 = h(u, v) + h(v, u) = h(u + v, u + v) − h(u, u) − h(v, v).
Now since 2 ∈ U(A) and A is local, we conclude that at least one of h(u, u), h(v, v) or
h(u + v, u + v) belongs to U(A). Finally in the case where h is skew hermitian and ∗ is
unramified, choose b ∈ U(A) ∩ S. Now observe that bh is hermitian and that bh(u, v) ∈
U(A) if and only if h(u, v) ∈ U(A).
(cid:3)
In summary the results of this section show that if h has any unit value then it is possible
to break off a nondegenerate submodule of rank one except in the ramified skew hermitian
case. In the latter case, no nondegenerate rank one submodules can possibly exist, but it
is possible to break off a nondegenerate submodule of rank two.
3. Complete local rings
Having given conditions sufficient to ensure the existence of nondegenerate rank one and
two submodules, in this section we introduce some natural conditions on the ring that
will guarantee the existence of unit length basis vectors, or in the ramified skew hermitian
case, the existence of a symplectic pair. Recall that a symplectic pair is an ordered pair
of vectors (u, v) ∈ V 2 satisfying h(u, u) = h(v, v) = 0 and h(u, v) = −h(v, u) = 1. We also
recall that a symplectic basis is a basis v1, w1, v2, w2,· · · , vl, wl such that each pair (vi, wi)
HERMITIAN AND SKEW HERMITIAN FORMS OVER LOCAL RINGS
5
i=1
is a symplectic pair and such that h(vi, vj) = h(wi, wj) = h(vi, wj) = h(wj, vi) = 0 for
i 6= j.
Observe that h(au, au) = a∗ah(u, u), so it is natural to investigate the image of the
so-called norm map, N : a 7→ a∗a. To this end, we follow Cohen ([5]) and say that A is
complete if T∞
ri = 0 and A is metrically complete with respect to its r-adic metric. We
observe that in the present context, where we seek solutions to certain quadratic equations,
it is natural to restrict our attention to complete rings.
Lemma 4. Suppose that A is complete. Suppose that b ∈ U(R) and that a∗a ≡ b mod r
for some a ∈ A. Then there is some c ∈ A such that c∗c = b.
Proof. First we observe that since A is complete, R is also complete. Since 2 ∈ R, it follows
from Hensel's Lemma (see [5], Theorem 4) that the squaring map from 1 + m to 1 + m is
a surjection. Now a ∈ U(A) since b ∈ U(A), therefore b(a∗a)−1 ∈ 1 + m. Hence there is
some δ ∈ m such that b(a∗a)−1 = (1 + δ)2. So b = (a(1 + δ))∗a(1 + δ).
Corollary 1. Suppose that A is complete and that u ∈ V , b ∈ U(A) and b∗ = εb. If
h(u, u) ≡ b mod r then there is some w ∈ Au such that h(w, w) = b.
Proof. Note that h(u, u)−1b ∈ 1 + m. By Lemma 4 there is some c ∈ A such that c∗c =
h(u, u)−1b. Let w = cu.
(cid:3)
(cid:3)
Now we turn to the nondegenerate two dimensional submodules in the case where ∗ is
ramified and h is skew hermitian.
Lemma 5. Suppose that A is complete and that ∗ is ramified. Suppose also that h is skew
hermitian and that h(u, v) ∈ U(A). Then there is some symplectic pair (u′, v′) ∈ V 2 such
that Au′ + Av′ = Au + Av.
Proof. By replacing v by h(u, v)−1v we can assume that h(u, v) = 1. Now observe that for
b ∈ A, h(u + bv, u + bv) = 0 if and only if
(1)
b∗bh(v, v) + (b − b∗) + h(u, u) = 0
Since both h(v, v) and h(u, u) belong to S we can apply Lemma 6 (below) to conclude that
Equation (1) has a solution in r. So assume that b ∈ r satisfies (1) and let u′ = u + bv.
Clearly, Au + Av = Au′ + Av. Moreover, h(u′, v) = 1 + b∗h(v, v) ∈ U(A). Now, let v′ =
h(u′, v)−1(v + 1
2h(v, v)(h(u′, v)∗)−1u′). A straightforward calculation, using h(u′, u′) = 0,
demonstrates that (u′, v′) is a symplectic pair. Moreover, since h(u′, v) is a unit, it is clear
that Au′ + Av′ = Au′ + Av = Au + Av.
(cid:3)
To complete the proof of Lemma 5 we need the following.
Lemma 6. Suppose that A is complete and that ∗ is ramified. Given α, β, γ ∈ A satisfying
α, γ ∈ S and β ∈ U(A), there is some t ∈ Aγ such that
αt∗t + βt − t∗β ∗ + γ = 0.
6
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
Proof. We define sequences (βi) and (γi) as follows. Let γ1 = γ and β1 = β. Given βi and
γi, let
γi+1 = −
βi+1 = βi +
,
αγ2
i
4β ∗
i βi
αγi
2β ∗
i
.
Observe that βi+1 is a unit since βi+1 ≡ βi mod r. Also γi+1 is skew hermitian, since
α and γi (inductively) are both skew hermitian. Now we check that − 1
i=1 γi/βi is the
required solution. Define fi(t) = αt∗t + βit − t∗β ∗
i + γi. Using the fact that γi is skew
hermitian, an easy calculation shows that fi(t − γi/2βi) = fi+1(t). Therefore fk+1(0) =
fk(−γk/2βk) = fk−1(−γk/2βk−γk−1/2βk−1) = · · · = f1(− 1
i=1 γi/βi). Now fk+1(0) = γk.
Since γ1 ∈ r, it is clear that γk ∈ rk for all k. Therefore f1(− 1
i=1 γi/βi) ∈ rk as
required.
2 P∞
2 Pk
(cid:3)
2 Pk
The following example shows that if A is not complete it may be possible to find a
nondegenerate rank two skew hermitian submodule that does not have a symplectic basis.
Example 1. Fix an odd prime p such that p + 1 is not the sum of two squares (e.g.
p = 5) and let A be the extension of R = Z(p) obtained by adjoining a square root of p.
So A = R ⊕ R√p is a local principal ideal domain with maximal ideal A√p. Let ∗
be the involution of A that fixes elements of R and maps √p to −√p and consider the
nondegenerate skew hermitian form hM where M = (cid:18)√p
−1 √p(cid:19). We claim that there is
no isotropic basis vector of A2 . Suppose that (cid:18)
a + b√p(cid:19) was such an vector (a similar
argument applies to the case (cid:18)a + b√p
(cid:19)). A straightforward calculation shows that we
must have pb2 − 2b− 1 = a2. Writing b = c
d for integers c and d, we see that pc2 − 2cd − d2
must be the square of an integer. But pc2 − 2cd − d2 = (p + 1)c2 − (d + c)2, so (p + 1)c2 is
the sum of two squares. By assumption, p + 1 is not the sum of two squares, so neither is
(p + 1)c2 for any integer c.
1
1
1
4. Nondegenerate forms
We set V (−1) = 0 and
V (i) = {v ∈ V : h(V, v) ⊂ r
i},
i ≥ 0.
Observe that V (i) is an A-submodule of V , V (i + 1) ⊂ V (i) and rV (i − 1) ⊂ V (i) for all
ri = 0 then Ti≥0 V (i) = {v ∈ V h(v, V ) = 0} is
i ≥ 0. We note in passing that, if Ti≥0
the radical of h. Now, since rV (i) ⊂ V (i + 1), we see that
inherits a A/r-vector space structure. The form h induces a map
W (i) = V (i)/(rV (i − 1) + V (i + 1))
hi : W (i) × W (i) → r
i/r
i+1.
HERMITIAN AND SKEW HERMITIAN FORMS OVER LOCAL RINGS
7
In the special case i = 0, we see that h0 is an ε-hermitian form V /V (1) × V /V (1) → A/r.
Observe that h is nondegenerate if and only if h0 is nondegenerate.
Lemma 7. Suppose that h is nondegenerate. Then V (i) = riV for all i ≥ 0.
Proof. The inclusion riV ⊂ V (i) is clear. For the reverse inclusion, let v ∈ V (i) and let
{v1, . . . , vm} be a basis of V , so that v = a1v1 + · · · + amvm for some ai ∈ A. Since h is
nondegenerate, given any 1 ≤ j ≤ m, there is u ∈ V such that h(u, vj) = 1 and h(u, vk) = 0
for all k 6= j. It follows that h(u, v) = aj, whence aj ∈ ri and a fortiori v ∈ riV .
Theorem 1. Suppose that h is nondegenerate.
(cid:3)
diagonal.
(1) If ∗ is unramified or if h is hermitian then V has a basis whose Gram matrix is
(2) If ∗ is ramified and h is skew hermitian then V has a basis whose Gram matrix is
the direct sum of m/2 matrices, each of which is an invertible 2× 2 skew hermitian
matrix.
Proof. Since h is nondegenerate the Gram matrix of a basis is invertible. Since A is local
some entry of this Gram matrix must be a unit. Therefore there exist vectors u, v such
that h(u, v) ∈ U(A). The results of Section 2 show that in the ramified skew hermitian
case there is a nondegenerate rank two submodule of V and that in all other cases there is
a nondegenerate rank one submodule. Moreover, if U is this rank one or two submodule
then by Lemma 1 or Lemma 2, V = U ⊕ U ⊥ and U ⊥ is a free submodule of V . Now, it
is clear that hU ⊥ is also nondegenerate. The required conclusions follow by induction on
the rank of V .
(cid:3)
Theorem 2. Suppose that A is complete and let h and h′ be nondegenerate ε-hermitian
forms on V . Then h and h′ are equivalent if and only if h0 and h′
0 are equivalent. In
particular, if ∗ is ramified and h and h′ are skew hermitian, then h and h′ are equivalent.
Theorem 2 (Matrix version). Suppose that A is complete and let C, D ∈ GLm(A). Then
C and D are ∗-congruent if and only C and D are ∗-congruent over A/r. In particular, if
∗ is ramified and C and D are skew hermitian, then C and D are ∗-congruent.
Proof. The only if direction is obvious. For the other direction suppose that C and D
are ∗-congruent. So there is some matrix X ∈ Mm(A) such that X ∈ GLm(A/r) and
X ′∗CX ≡ D mod r. Since A is local, X ∈ GLm(A) and C is ∗-congruent to X ′∗CX. So
replacing C by X ′∗CX we can, without loss of generality, reduce to the case that C ≡ D
mod r. Equivalently, we may assume without loss of generality that h(u, v) ≡ h′(u, v)
mod r for all u, v ∈ V .
Now we deal with the case where ∗ is ramified and h is skew hermitian. In this case,
Lemma 5 implies that V has symplectic bases relative to h and h′, whence they are equiv-
alent.
In all other cases, Lemmas 1 and 3 ensure the existence of a basis v1,· · · , vm of V
whose corresponding Gram matrix relative to h, say M, is diagonal with only units on
the diagonal. By Corollary 1 there is some w1 ∈ Av1 such that h′(w1, w1) = h(v1, v1).
8
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
Moreover, it is clear that, for j ≥ 2, h′(w1, vj) ≡ h′(v1, vj) ≡ h(v1, vj) ≡ 0 mod r. For
j ≥ 2, let wj = vj − h′(w1, vj)h′(w1, w1)−1w1. Now w1,· · · , wm is a basis of V whose Gram
matrix with respect to h′ is of the form
(cid:18)M11
Moreover N ≡ diag(M22,· · · , Mmm) mod r.
Inductively, we may assume that there is
some X ∈ GLm−1(A) such that X ′∗NX = diag(M22,· · · , Mmm). It follows that there is
a basis of V whose Gram matrix with respect to h′ is equal to M. Note that Corollary 1
provides the base case of the induction.
(cid:3)
0 N(cid:19) .
0
We list some noteworthy corollaries and special cases of Theorem 2.
Proposition 1. If A is complete, ∗ is ramified and h is skew hermitian and nondegenerate,
then V has a symplectic basis.
(cid:3)
We have a canonical imbedding R/m ֒→ A/r and we will view R/m as a subfield of A/r by
means of this imbedding. If ∗ is ramified then R/m = A/r. Suppose ∗ is unramified. Then
R/m is the fixed field of an automorphism of A/r of order 2. If A/r is quadratically closed
(in the sense that it has no extensions of degree 2), then by the Diller-Dress theorem (see
[10], p. 235), R/m is a Euclidean field (this is an ordered field wherein every nonnegative
element is a square).
Proposition 2. Suppose that A is complete and that A/r is quadratically closed. If ∗ is
unramified and h is a nondegenerate hermitian form then V has a basis whose Gram matrix
is diagonal and such that all the diagonal entries are ±1. Moreover, given any two such
bases, the signatures of the corresponding Gram matrices are the same.
Proof. By Theorem 2 it suffices to prove this result in the case that A is a field (i.e. r = 0).
By our above remarks, A is the quadratic extension of a Euclidean field. The result in this
case goes exactly as in a classical case when A = C.
(cid:3)
Similarly we have
Proposition 3. Suppose that A is complete and that A/r is quadratically closed. If ∗ is
unramified and h is a nondegenerate skew hermitian form then, given b ∈ U(A) ∩ S, V
has a basis whose Gram matrix is diagonal and such that all the diagonal entries are ±b.
Moreover, given any two such bases, the number of occurrences of b are the same in each
of the corresponding Gram matrices.
Proof. Apply the previous proposition to the hermitian form b−1h.
Proposition 4. Suppose that A is complete and that A/r is quadratically closed. If ∗ is
ramified and h is a nondegenerate hermitian form then V has an orthonormal basis.
(cid:3)
Proof. This follows by combining Theorem 2 with the classical result that every nonde-
generate symmetric form over a quadratically closed field of characteristic not 2 has an
orthonormal basis.
(cid:3)
HERMITIAN AND SKEW HERMITIAN FORMS OVER LOCAL RINGS
9
For the remainder of the paper A will be a discrete valuation ring, that is, a local
5. Discrete Valuation Rings
principal ideal domain. A uniformiser of A is a generator of the maximal ideal.
Lemma 8. If ∗ is a nontrivial involution of A then A has a uniformiser y such that
y∗ = −y.
Proof. Suppose that r = Az. Now z = 1
2 (z − z∗) and since A is local, at least
one of 1
2(z − z∗) must lie in r − r2. So we can certainly choose some w such
that w∗ = ±w and r = Aw. Suppose that w∗ = w. Since ∗ is nontrivial there is some
nonzero u ∈ A such that u∗ = −u. Now u = bwk for some b ∈ U(A) and since w∗ = w, we
conclude that b∗ = −b. Now y = bw is the required generator of r.
2(z + z∗) or 1
2(z + z∗) + 1
(cid:3)
Following the previous lemma we fix y so that Ay = r and so that y∗ = −y in the case
Following O'Meara [14] we say that an ε-hermitian matrix M ∈ Mm(A) has an O'Meara
that ∗ is a nontrivial involution. We agree that y∞ = 0.
decomposition if
M =
s
Mi=0
aiMi,
where each ai ∈ A and every Mi is an invertible hermitian or skew hermitian matrix.
Our next result is an extension of Theorem 1 to arbitrary forms, possibly degenerate,
and establishes the existence of an O'Meara decomposition for the Gram matrices of such
forms.
Theorem 3.
(1) If ∗ is unramified then V has a basis whose Gram matrix is diagonal.
(2) If ∗ is ramified then V has a basis whose Gram matrix is the direct sum of a diagonal
matrix (possibly of size zero) and a number (possibly zero) of 2 × 2 blocks each of
the form ydB where B is an invertible skew hermitian 2 × 2 matrix.
Proof. We prove this by induction on the rank of V . If h is identically zero then the theorem
is true. Now suppose that h is not identically zero. Let d = min{j : h(V, V ) ⊂ rj}. Since A
is a domain there is a unique form h′ on V such that h = ydh′. Clearly h′ is ε-hermitian if
y∗ = y and is (−1)dε-hermitian if y∗ = −y. Moreover, by construction, there exist u, v ∈ V
such that h′(u, v) ∈ U(A).
If h′ is skew hermitian and ∗ is ramified then, by Lemma 2 there is some nondegenerate
rank two submodule U such that V = U ⊥ U ⊥. If h′ is hermitian or if ∗ is unramified
then, by Lemma 1 and Lemma 3 there is some nondegenerate rank one submodule U such
that V = U ⊥ U ⊥.
In either case U has an O'Meara decomposition of the required form with respect to h′
and by induction U ⊥ has an O'Meara decomposition of the required form with respect to
h′. Together these yield the required decomposition for h.
(cid:3)
Next we consider the structure of the space W (i) in more detail. By Theorem 3 we have
(2)
V = U0 ⊥ U1 ⊥ U2 ⊥ · · ·· · · ⊥ U∞
10
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
where Ui = 0 for all but finitely many i and, for Ui 6= 0, the Gram matrix of a basis, say Bi,
of Ui is equal to yiMi with Mi invertible. Thus, if B is the union of all Bi then B is a basis
of V with Gram matrix
(3)
M = M0 ⊕ y1M1 ⊕ y2M2 ⊕ · · ·· · · ⊕ y∞M∞,
where almost all summands have size 0.
Lemma 9. For every nonnegative integer i, we have
V (i) = r
iU0 ⊥ r
i−1U1 ⊥ r
i−2U2 ⊥ · · · ⊥ Ui ⊥ Ui+1 ⊥ Ui+2 ⊥ · · · ⊥ U∞.
Proof. The operation of passing from V to V (i) is compatible with orthogonal decomposi-
tions, so the result follows immediately from Lemma 7 and the decomposition (2).
(cid:3)
Corollary 2. For every nonnegative integer i, we have dimA/r W (i) = rankA Ui.
Making use of the uniformiser y for A, we may alter the above map W (i)×W (i) → ri/ri+1
and define an A/r-valued form hi : W (i) × W (i) → A/r by
hi(u + rV (i − 1) + V (i + 1), v + rV (i − 1) + V (i + 1)) = y−ih(u, v) + r,
u, v ∈ V (i).
Note that hi is ε-hermitian if ∗ is trivial and is ε(−1)i-hermitian if ∗ is nontrivial.
Corollary 3. For every nonnegative integer i, the form hi is nondegenerate. In fact, the
Gram matrix of the basis Bi of W (i) relative to hi is Mi, which is invertible or has size 0.
It is clear that if h and h′ are equivalent forms on V , then hi and h′
i are equivalent over
A/r for i ≥ 0. The converse is, of course, not necessarily true (see Example 1).
Theorem 4. Suppose that A is complete and that h and h′ are ε-hermitian forms on V .
If, for each nonnegative integer i, the forms hi and h′
i are equivalent over A/r then h and
h′ are equivalent.
Proof. This is an immediate consequence of Theorem 2 (in its matrix version), Corollary 3
and the decomposition (3).
(cid:3)
We next obtain the following classification theorem for ∗-congruence classes of ε-hermitian
matrices
Theorem 5. Assume that A is complete.
(1) Suppose that M ∈ Mm(A) is ε-hermitian. Then M is ∗-congruent to a matrix of
i=0 yiMi, where for each Mi is either of size zero or is an invertible
the form L∞
matrix. Moreover, for 0 ≤ i < ∞,
(a) if ∗ is unramified then Mi is a diagonal matrix.
(b) if ∗ is ramified and (yi)∗ = εyi then Mi is diagonal. Whereas if ∗ is ramified
and (yi)∗ = −εyi then Mi is a direct sum of copies of (cid:18) 0
i=0 yiMi and N = L∞
(2) Given ε-hermitian matrices M = L∞
that each Mi (resp. Ni) is either of size zero or is an invertible matrix, then
1
−1 0(cid:19).
i=0 yiNi in Mm(A) such
HERMITIAN AND SKEW HERMITIAN FORMS OVER LOCAL RINGS
11
(a) if ∗ is unramified, M and N are ∗-congruent over A if and only if for each
0 ≤ i < ∞, Mi is ∗-congruent to Ni over A/r.
(b) if ∗ is ramified, M and N are ∗-congruent over A if and only if for each
0 ≤ i < ∞ such that (yi)∗ = εyi, Mi is ∗-congruent to Ni over A/r, and for
each 0 ≤ i < ∞ such that (yi)∗ = −εyi, the size of Mi is equal to the size
of Ni.
Proof. This follows immediately from Theorem 2, Theorem 3 and Lemma 5.
(cid:3)
Thus, for matrices over a complete discrete valuation ring with residue field of charac-
teristic not 2, the ∗-congruence problem essentially reduces to the ∗-congruence problem
over its residue field.
Example 2. (cf. Example 1.1 of [12]) Given any field F , consider the complete, local,
nonprincipal domain A = F [[X, Y ]] and the ∗-hermitian (with trivial ∗) matrix
M = (cid:18) 0 X
X Y (cid:19) .
We claim that M has no O'Meara decomposition. Suppose, to the contrary, that it does.
Since M is not invertible, then M must be ∗-congruent to a diagonal matrix D = diag(a, b).
Thus D = T ′MT for some T ∈ GL2(A). Taking determinants yields ab = −t2X 2, where
t ∈ U(A). Since (X) is a prime ideal, it follows that a ∈ (X) or b ∈ (X). Suppose, without
loss of generality, that a = cX for some c ∈ A. Cancelling, we obtain cb = −t2X.
If
b /∈ (X) then a repetition of the preceding argument yields that b ∈ U(A), against the fact
that the hermitian form h = hM takes no unit values. We are thus forced to conclude that
b = dX for some d ∈ A, which implies that c, d ∈ U(A). Now, since M is congruent to D,
there is u ∈ V = A2 such that h(u, u) = cX. Let e1, e2 be the canonical basis of V . Then
u = f e1 + ge2 for some f, g ∈ A, whence 2f gX + g2Y = cX, that is, g2Y = (c − 2f g)X.
We infer that g ∈ (X) and a fortiori c − 2f g ∈ U(A). Thus (c − 2f g)X ∈ (Y ) but neither
(c − 2f g) nor X are in (Y ), which contradicts the fact that (Y ) is a prime ideal.
The following proposition is not needed for the sequel. However, we include it to shed
light on the structure of the rings under consideration.
Proposition 5. Suppose ∗ is a nontrivial involution of A. Then R is a discrete valuation
ring and A = R[z] = R ⊕ Rz, where z∗ = −z and z2 = x ∈ R. Moreover, x ∈ U(R) if ∗ is
unramified and x ∈ m otherwise. In either case, x is not a square in R.
Proof. By Lemma 8 we can choose y ∈ S ∩ (r − r2).
Suppose first that S ⊂ r. Since m = r ∩ R, we know that if b ∈ m then b = cy for some
c ∈ A. Clearly, since b ∈ R and y ∈ S, we must have c ∈ S ⊂ r. Therefore c ∈ Ay and
we conclude that b ∈ Ry2. So in this case m = Ry2. Since 2 ∈ U(A), we have A = R ⊕ S.
If s ∈ S then sy ∈ r ∩ R = m, whence sy = ry2 for some r ∈ R, so s = ry and a fortiori
S = Ry. Moreover, in this case, let x = y2 and observe that x cannot be a square of any
element of R.
12
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
On the other hand, if S 6⊂ r then we may choose w ∈ U(A) such that w∗ = −w. Then
z = wy ∈ R ∩ (r − r2) and it is clear that m = Rz in this case. If t ∈ S then tw−1 ∈ R,
so t ∈ Rw, which gives S = Rw. Now let x = w2 ∈ R and once again, one readily checks
that x is not a square in R.
In both cases R is a local domain with principal maximal ideal, so every ideal is principal.
(cid:3)
6. Invariants and normal forms
We are finally in a position to show that, after imposing additional hypotheses, the
sequence di = dimA/r W (i), i ≥ 0, is a complete set of invariants for equivalence classes of
ε-hermitian forms.
Let B stand for the fixed field of the involution that ∗ induces on A/r. Then B = A/r if
∗ is ramified and B = R/m (viewed as a subfield of A/r) if ∗ is unramified. In either case,
we have a norm map N : A/r → B given by a + r → aa∗ + r. One of the aforementioned
hypotheses is that N be surjective.
Theorem 6. Suppose that A is complete and N is surjective. Let V and V ′ be free A-
modules, both of rank m, and let h : V × V → A and h : V ′ × V ′ → A be ε-hermitian
forms. Then h and h′ are equivalent if and only if di = d′
Proof. This follows from Corollary 3 and Theorem 4 using the fact that N is surjective. (cid:3)
i for all i ≥ 0.
We may use Theorem 6 to obtain normal forms for ε-hermitian matrices.
Theorem 7. Suppose that A is complete and N is surjective.
∗-congruent to one and only one matrix of the form L∞
size > 0 is equal to the direct sum of copies of (cid:18) 0
1
matrix).
1
and equal to the identity matrix if i is odd (resp. even).
is ∗-congruent to one and only one matrix of the form L∞
of size > 0 is equal to the direct sum of copies of (cid:18) 0
(1) If ∗ is nontrivial, let M ∈ Mm(A) be skew hermitian (resp. hermitian). Then M
i=0 yiMi, where every Mi
−1 0(cid:19) if i is even (resp. odd)
(2) If ∗ is trivial, let M ∈ Mm(A) be skew symmetric (resp. symmetric). Then M is
i=0 yiMi, where every Mi of
−1 0(cid:19) (resp. equal to the identity
Recall that two matrices M, N ∈ Mm(A) are said to be equivalent if P MQ = N for
some P, Q ∈ GLm(A). Assuming that A is complete and N is surjective, we may use
Corollary 3 and Theorem 4 to see that the problem of ∗-congruence of matrices reduces to
the problem of equivalence of matrices, whose answer is well-known. Since every invertible
matrix is equivalent to the identity matrix, we have
Theorem 8. Suppose A is complete and N is surjective. Then two ε-hermitian matrices
M, N ∈ Mm(A) are ∗-congruent if and only if they have the same invariant factors.
HERMITIAN AND SKEW HERMITIAN FORMS OVER LOCAL RINGS
13
What are the possible invariant factors of an ε-hermitian matrix? By above this amounts
to asking what are the sequences that arise as (di)i≥0 for some ε-hermitian form. Let us
call such sequences ε-realisable. The answer is an immediate consequence of Theorem 7.
Proposition 6. Suppose that A is complete and N is surjective. Let (di) be a sequence of
nonnegative integers.
(1) If ∗ is trivial and ε = 1 then (di) is ε-realisable.
(2) If ∗ is trivial and ε = −1 then (di) is ε-realisable if and only if di is even for all i.
(3) If ∗ nontrivial, then (di) is ε-realisable if and only if di is even for i such that
(−1)i = −ε.
The analogue of Theorem 7 in the case when ∗ is unramified and A/r is assumed to be
quadratically closed is an immediate consequence of Proposition 2 and Theorem 5.
Theorem 9. Suppose that A is complete with quadratically closed residue field. Assume
that ∗ is unramified and fix b ∈ S ∩ U(A). Let M ∈ Mm(A) be hermitian (resp. skew
hermitian). Then M is ∗-congruent to one and only one matrix of the form L∞
i=0 yiMi,
where every Mi of size > 0 is equal to a diagonal matrix diag(1, . . . , 1,−1, . . . ,−1) (resp.
diag(b, . . . , b,−b, . . . ,−b)).
References
[1] Eva Bayer-Fluckiger and Laura Fainsilber, Non-unimodular Hermitian forms, Invent. Math. 123
(1996), no. 2, 233–240, DOI 10.1007/s002220050024. MR1374198
[2] Eva Bayer-Fluckiger, Uriya A. First, and Daniel A. Moldovan, Hermitian categories, extension
of scalars and systems of sesquilinear forms, Pacific J. Math. 270 (2014), no. 1, 1–26, DOI
10.2140/pjm.2014.270.1. MR3245846
[3] Eva Bayer-Fluckiger and Daniel Arnold Moldovan, Sesquilinear forms over rings with involution, J.
Pure Appl. Algebra 218 (2014), no. 3, 417–423, DOI 10.1016/j.jpaa.2013.06.012. MR3124208
[4] Yonglin Cao and Fernando Szechtman, Congruence of symmetric matrices over local rings, Linear
Algebra Appl. 431 (2009), no. 9, 1687–1690, DOI 10.1016/j.laa.2009.06.003. MR2555069
[5] I.S. Cohen, On the structure and ideal theory of complete local rings, Transactions of the American
Mathematical Society 59 (1946), no. 1, 54-106.
[6] Dragomir Z. ¯Dokovi´c and Fernando Szechtman, Solution of the congruence problem for arbitrary
Hermitian and skew-Hermitian matrices over polynomial rings, Math. Res. Lett. 10 (2003), no. 1,
1–10, DOI 10.4310/MRL.2003.v10.n1.a1. MR1960118
[7] William H. Durfee, Congruence of quadratic forms over valuation rings, Duke Math. J. 11 (1944),
687–697. MR0011073
[8] Ronald Jacobowitz, Hermitian forms over local fields, Amer. J. Math. 84 (1962), 441–465, DOI
10.2307/2372982. MR0150128
[9] Max-Albert Knus, Quadratic and Hermitian forms over rings, Grundlehren der Mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences], vol. 294, Springer-Verlag, Berlin,
1991. With a foreword by I. Bertuccioni. MR1096299
[10] T.Y. Lam, Introduction to Quadratic Forms over Fields, Graduate Studies in Mathematics, vol. 67,
American Mathematical Society, Providence, RI, 2005.
[11] Serge Lang, Algebra, 2nd ed., Addison-Wesley Publishing Company, Advanced Book Program, Read-
ing, MA, 1984. MR783636
[12] V. Levchuk and O. Starikova, Quadratic forms of projective spaces over rings, Sb. Math. 197 (2006),
887-899.
14
JAMES CRUICKSHANK, RACHEL QUINLAN, AND FERNANDO SZECHTMAN
[13] V. Levchuk and Starikova O., A normal form and schemes of quadratic forms, J. Math. Sci. 152
(2008), 558-570.
[14] O. T. O'Meara, Quadratic forms over local fields, Amer. J. Math. 77 (1955), 87–116, DOI
10.2307/2372423. MR0067163
School of Mathematics, Statistics and Applied Mathematics, National University of
Ireland Galway, Ireland
E-mail address: [email protected]
School of Mathematics, Statistics and Applied Mathematics, National University of
Ireland Galway, Ireland
E-mail address: [email protected]
Department of Mathematics and Statistics, University of Regina, Canada
E-mail address: [email protected]
|
1109.5062 | 1 | 1109 | 2011-09-23T13:01:30 | Invertible unital bimodules over rings with local units, and related exact sequences of groups II | [
"math.RA",
"math.CT"
] | Let $R$ be a ring with a set of local units, and a homomorphism of groups $\underline{\Theta} : \G \to \Picar{R}$ to the Picard group of $R$. We study under which conditions $\underline{\Theta}$ is determined by a factor map, and, henceforth, it defines a generalized crossed product with a same set of local units. Given a ring extension $R \subseteq S$ with the same set of local units and assuming that $\underline{\Theta}$ is induced by a homomorphism of groups $\G \to \Inv{R}{S}$ to the group of all invertible $R$-sub-bimodules of $S$, then we construct an analogue of the Chase-Harrison-Rosenberg seven terms exact sequence of groups attached to the triple $(R \subseteq S, \underline{\Theta})$, which involves the first, the second and the third cohomology groups of $\G$ with coefficients in the group of all $R$-bilinear automorphisms of $R$. Our approach generalizes the works by Kanzaki and Miyashita in the unital case. | math.RA | math |
Invertible unital bimodules over rings with local units, and
related exact sequences of groups, II ∗
L. El Kaoutit
Departamento de Álgebra
Facultad de Educación y Humanidades
Universidad de Granada
El Greco N0 10, E-51002 Ceuta, España
e-mail:[email protected]
J. Gómez-Torrecillas
Departamento de Álgebra
Facultad de Ciencias
Universidad de Granada
E18071 Granada, España
e-mail: [email protected]
Abstract
Let R be a ring with a set of local units, and a homomorphism of groups Θ : G →
Pic(R) to the Picard group of R. We study under which conditions Θ is determined
by a factor map, and, henceforth, it defines a generalized crossed product with a same
set of local units. Given a ring extension R ⊆ S with the same set of local units and
assuming that Θ is induced by a homomorphism of groups G → InvR(S) to the group of
all invertible R-sub-bimodules of S, then we construct an analogue of the Chase-Harrison-
Rosenberg seven terms exact sequence of groups attached to the triple (R ⊆ S, Θ), which
involves the first, the second and the third cohomology groups of G with coefficients in
the group of all R-bilinear automorphisms of R. Our approach generalizes the works by
Kanzaki and Miyashita in the unital case.
Introduction
For a Galois extension R/k of commutative rings with identity and finite Galois group G,
Chase, Harrison and Rosenberg, gave in [5, Corollary 5.5] (see also [6]) the following seven
terms exact sequence of groups
1 → H 1(G, U (R)) → Pick(k) → PicR(R)G → H 2(G, U (R)) → B(R/k)
→ H 1(G, PicR(R)) → H 3(G, U (R)),
where for a ring A with identity, PicA(A) = {[P ] ∈ Pic(A) ap = pa, ∀ a ∈ A, p ∈ P } is a
subgroup of the Picard group Pic(A), and U (A) the group of units. The group B(R/k) is the
Brauer group of Azumaya k-algebras split by R, and the other terms are the usual cohomology
groups of G with coefficients in abelian groups. As one can realize, the case of finite Galois
extension of fields reduces to fundamental theorems in Galois cohomology of fields extensions,
namely, the Hilbert's 90 Theorem, and the isomorphism of the Brauer group of a field with
the second cohomology group.
∗Research supported by the grant MTM2010-20940-C02-01 from the Ministerio de Ciencia e Innovación
and from FEDER
1
The construction of this sequence of groups, as was given in [5, 6], is ultimately based
on the consideration of a certain spectral sequence already introduced by Grothendieck, and
the generalized Amitsur cohomology. The same sequence was also constructed after that by
Kanzaki [13, Theorem, p.187], using only elementary methods that employ the novel notion of
generalized crossed products. This notion has been the key success which allowed Miyashita
to extend in [15, Theorem 2.12] the above framework to the context of a noncommutative
unital ring extension R ⊆ S with a homomorphism of groups to the group of all invertible
sub-R-bimodules of S. In this setting the seven terms exact sequence takes of course a slightly
different form, although the first, the fourth and the last terms remain the cohomology groups
of the acting group with coefficients in the group of units of the base ring. The commutative
Galois case is then recovered by first considering a tower k ⊆ R ⊆ S, where k ⊆ R is a finite
Galois extension and S is a fixed crossed product constructed from the Galois group; secondly
by considering the homomorphism of groups which sends any element in the Galois group to
it corresponding homogeneous component of S (components which belong to the group of all
invertible sub-R-bimodules of S).
In this paper we construct an analogue of the above seven terms exact sequence in the
context of rings with local units (see the definition below). The most common class of examples
are the rings with enough orthogonal idempotents or Gabriel's rings which are defined from
small additive categories. It is noteworthy that up to our knowledge there is no satisfactory
notion in the literature of morphism of rings with local units that could be, for instance,
extracted from a given additive functor between additive small categories. Here we restrict
ourselves to the naive case of extensions of the form R ⊆ S which have the same set of local
units. This at least includes the situation where two small additive categories have the same
set of objects and differ only in the size of the sets of morphisms, as it frequently happens in
the theory of localization in abelian categories. Our methods are inspired from Miyashita's
work [15] and use the results of our earlier work [9]. The present paper is in fact a continuation
of [9]. It is worth noticing that, although the steps for constructing the seven terms sequence
in our context are slightly parallel to [15], this construction is not an easy task since it has its
own challenges and difficulties.
The paper is organized as follows.
In Section 1 we give some preliminary results on
similar and invertible unital bimodules. Most of results on similar unital bimodules are in
fact the non unital version of [11, 12], except perhaps the construction of the twisting natural
transformations and its weak associativity (Proposition 1.3 and Lemma 1.4). The generalized
crossed product is introduced in Section 2, where we also give useful constructions on the
normalized 2 and 3-cocycles with coefficient in the unit group of the base ring. Section 3 is
devoted to the construction of an abelian group whose elements consist of isomorphic classes
of generalized crossed products. We show that it contains a subgroup which is isomorphic
to the 2-cohomology group (Proposition 3.2). Our main results is contained in Section 4,
which contains a number of exact sequences of groups that culminate in the seven terms exact
sequence that generalizes Chase-Harrison-Rosenberg's one (Theorem 4.12 ).
Notations and basic notions:
In this paper ring means an associative and not necessarily unital ring. We denote by Z(R)
(resp. Z(G)) the center of a ring R (resp. of a group G), and if R is unital we will denote by
U (R) the unit group of R, that is, the set of all invertible elements of R.
Let R be a ring and E ⊂ R a set of idempotent elements; R is said to be a ring with set of
local units E, provided that for every finite subset {r1, · · · , rn} ⊂ R, there exists e ∈ E such
2
that
e ri = ri e = ri,
for every i = 1, · · · , n,
In this way, given a finite set {r1, · · · , rn} of elements in R, we denote
U nit{r1, · · · , rn} = ne ∈ E eri = rie = ri,
for every i = 1, 2, · · · , no .
Observe that our definition differs from that in [1, Definition 1.1], since we do not assume
that the idempotents of E commute. In fact, our rings generalize those of [1] since, when the
idempotents are commuting, it is enough to require that for each element r ∈ R there exists
e ∈ E such that er = re = r (see [1, Lemma 1.2]). The Rees matrix rings considered in [2],
are for instance, rings with a set of local units with noncommuting idempotents.
Let R be a ring with a set of local units E. A right R-module X is said to be unital
provided one of the following equivalent conditions holds
(i) X ⊗R R ∼= X via the right R-action of X,
(ii) XR := {Pf inite xiri ri ∈ R, xi ∈ X} = X,
(iii) for every element x ∈ X, there exists an element e ∈ E such that xe = x.
Left unital R-modules are analogously defined. Obviously any right R-module X contains
XR as the largest right unital R-submodule.
Let R and S be rings with a fixed set of local units, respectively, E and E′. In the sequel,
an extension of rings φ : (R, E) → (S, E′) with the same set of local units, stand for an additive
and multiplicative map φ which satisfies φ(E) = E′. Note that since R and S have the same
set of local units, any right (resp. left) unital S-module can be considered as right (resp. left)
unital R-module by restricting scalars. Furthermore, for any right S-module X, we have the
equality XR = XS.
1 Similar and invertible unital bimodules.
Let R be a ring with a set of local units E. In the first part of this section we give some prelim-
inary results on unital R-bimodules, which we will use frequently in the sequel. Most of them
were already formulated by K. Hirata in [11] and [12] when R is unital. For the convenience
of the reader we will give in our case complete proofs. The second part provides a relation
between the automorphism group of an invertible unital R -- bimodule and the automorphisms
of the base ring R. This result will be also used in the forthcoming sections.
A unital R-bimodule is an R-bimodule which is left and right unital. In the sequel, the
expression R-bilinear map means homomorphism of R-bimodules. Note that if M is a unital
R-bimodule, then for every element m ∈ M, there exists a two sided unit e for m. That is,
there exists e ∈ E such that m = me = em, see [7, p. 733].
If several R-bimodules are
handled, say M, N, P with a finite number of elements {mi, ni, pi}i, mi ∈ M, ni ∈ N, pi ∈ P ,
then we denote by U nit{mi, ni, pi} the set of common two sided units of the mi's, ni's and
the pi's. This means, that e ∈ U nit{mi, ni, pi} ⊆ E if and only if
emi = mi = mie, eni = ni = nie, epi = pi = pie,
for every i = 1, · · · , n.
3
Given a unital R-bimodule M, we denote its invariant sub-bimodule, that is, the set of all
R-bilinear maps from R to M, by
M R := HomR−R(R, M ).
It is canonically a module over the commutative ring of all R-bimodule endomorphisms Z :=
EndR−R(R). The Z-action is given as follows: for every z ∈ Z and f ∈ M R, we have
z.f : RRR → RMR,(cid:16)r 7→ f (z(r)) = z(e)f (r) = f (r)z(e), where e ∈ U nit{r}(cid:17) .
Lemma 1.1. Let R be a ring with local units and M a unital R-bimodule such that M
that is, M is isomorphic as an R-bimodule to a direct summand of direct sum of n copies of
R. Then
/ R(n) ,
⊕
(i) M R is a finitely generated and projective Z-module;
(ii) there is an isomorphism of R-bimodules
M ∼= R ⊗Z M R,
where the R-bimodule structure of the right hand term is induced by that of R. That is,
r(s ⊗Z f )t = (rst) ⊗Z f,
(s ⊗Z f ∈ R ⊗Z M R, r, t ∈ R).
Proof. Let us consider the following canonical R-bimodule homomorphisms:
ϕi : M
ι
/ R(n)
πi
/ R,
ψi : R
ιi
/ R(n)
π /
/ M,
i = 1, · · · , n,
where πi and ιi are the canonical projections and injections.
(i) It is clear that {(ψi, ψ∗
i is defined by the right composition with ϕi, that is, ψ∗
i )}i ∈ M R × HomZ (M R, Z) is a finite dual Z-basis for M R,
i (f ) = ϕi ◦ f , for every
where each ψ∗
f ∈ M R.
(ii) We know that there is a well defined R-bilinear map
R ⊗Z M R
r ⊗Z f
η
/ M
/ f (r).
We can easily show that
η−1
M
m
is actually the inverse map of η.
/ R ⊗Z M R
i ϕi(m) ⊗Z ψi,
/ Pn
We will follow Miyashita's notation concerning this class of R-bimodules. That is, if we
have an R-bimodule M which is a direct summand as a bimodule of finitely many copies of
another R-bimodule N, we shall denote this situation by M N. It is clear that this relation
is reflexive and transitive. Furthermore, it is compatible with the tensor product over R, in
the sense that we have M ⊗R QN ⊗R Q and Q ⊗R M Q ⊗R N, for every unital R-bimodule
Q, whenever M N. In this way, we set
M ∼ N,
if and only if, M N and N M,
(1)
and call M is similar to N. This in fact defines an equivalence relation which is, in the above
sense, compatible with the tensor product over R.
4
/
/
/
/
/
/
/
/
/
Lemma 1.2. Let R be a ring with local units and M , N are unital R-bimodules such that
M R and N R as bimodules. Then
(i) for every finitely generated and projective Z-module P , we have an isomorphism of Z-
modules
(R ⊗Z P )R ∼= P ;
(ii) there is a natural (on both components) Z-linear isomorphism
(M ⊗R N )R ∼= M R ⊗Z N R.
Proof. (i) Let {(pi, p∗
define a finite family of elements in Z:
i )}1≤i≤n be a dual Z-basis for P . Given an element f ∈ (R ⊗Z P )R, we
zf, i
:= (R ⊗Z p∗
i ) ◦ f : R → R ⊗Z P → R ⊗Z Z ∼= R.
Now, an easy computation shows that the following maps are mutually inverse
P
p
/ (R ⊗Z P )R
/ hr 7−→ r ⊗Z pi ,
(R ⊗Z P )R
f
/ P
i zf, i pi,
/ Pn
which gives the stated isomorphism.
(ii) By Lemma 1.1(ii) we know that
Therefore, we have
M ∼= R ⊗Z M R and N ∼= R ⊗Z N R.
M ⊗R N ∼= R ⊗Z (cid:16)M R ⊗Z N R(cid:17) .
Since by Lemma 1.1(i), M R ⊗Z N R is a finitely generated and projective Z-module, we obtain
the desired isomorphism by applying item (i) just proved above.
Keep the notations of the first paragraph in the proof of Lemma 1.1.
Proposition 1.3. Let R be a ring with local units and M , N are unital R-bimodules such
that M R and N R as bimodules. Then there is a natural R-bilinear isomorphism M ⊗R N ∼=
N ⊗R M given explicitly by
TM,N : M ⊗R N
x ⊗R y
∼=
where e ∈ U nit{x, y}.
/ N ⊗R M
i ϕi(x) y ⊗R ψi(e),
/ Pn
Proof. By Lemmata 1.1 and 1.2, we have the following chain of R-bilinear isomorphisms
M ⊗R N
∼= /
/ R ⊗Z M R ⊗Z N R
∼= /
/ R ⊗Z N R ⊗Z M R
∼= /
/ N ⊗R (R ⊗Z M R)
∼= /
/ N ⊗R M.
Writing down explicitly the composition of these isomorphisms, we obtain the stated one.
5
/
/
/
/
/
/
The natural transformation T−,− is associative in the following sense.
Lemma 1.4. Let R be ring with a local units. Consider X, Y, P, Q, U, V unital R-bimodules
which satisfy QR, (Y ⊗R V )R, (X ⊗R U )R and (P ⊗R Y )R, all as bimodules. Then the
following diagram is commutative:
X ⊗R U ⊗R P ⊗R Y ⊗R V ⊗R Q
X⊗U ⊗P ⊗TY ⊗V, Q
X ⊗R U ⊗R P ⊗R Q ⊗R Y ⊗R V
TX⊗U, P ⊗Y ⊗V ⊗Q
TX⊗U, P ⊗Q⊗Y ⊗V
P ⊗R Y ⊗R X ⊗R U ⊗R V ⊗R Q
P ⊗TY ⊗X⊗U ⊗V, Q
/ P ⊗R Q ⊗R Y ⊗R X ⊗R U ⊗R V.
Proof. Straightforward.
Recall form [9, p. 227] that a unital R-bimodule X is said to be invertible if there exists
another unital R-bimodule Y with two R-bilinear isomorphisms
X ⊗R Y
l
∼=
/ R
r
∼=
Y ⊗R X.
As was shown in [9], one can choose the isomorphisms l, r such that
X ⊗R r = l ⊗R X,
Y ⊗R l = r ⊗R Y.
In others words, such that (l, r−1) is a Morita context from R to R.
In this way, Y is
isomorphic as an R-bimodule to the right unital part X ∗R of the right dual R-module
X ∗ = Hom−R(X, R) of X. This isomorphism is explicitly given by
∼=−→ X ∗R, (cid:16)y 7−→ [x 7→ r(y ⊗ x)](cid:17) .
Y
It is worth mentioning that X is not necessarily finitely generated as right unital R-module;
but a direct limit of finitely generated and projective right unital R-modules.
Given e ∈ E consider its decomposition with respect to the invertible unital R-bimodule
X, that is,
e = X
(e)
xeye,
(i.e. e = X
(e)
l(xe ⊗R ye)).
It is clear that the xe's and the ye's can be chosen such that
exe = xe, and ye = yee,
which means that e is a right unit for the ye's and left unit for the xe's.
Remark 1.5. It is noteworthy that a two sided unit of the set of elements {xe, ye} do not
need in general to coincide with e. However, any two sided unit e1 ∈ U nit{xe, ye} is also a
unit for e. In other words, the elements e and P(e) xe e ye belong to the same unital ring eRe,
but they are not necessarily equal. This makes a difference in technical difficulties compared
with the case of unital rings.
The following lemma gives explicitly the relation between the automorphisms group of uni-
tal invertible R-bimodule and the unit group of the commutative unital ring Z = EndR−R(R).
6
/
/
/
/
o
o
Lemma 1.6. Let R be a ring with local units and X an invertible unital R-bimodule. Then
there is an isomorphism of groups, from the group of R-bilinear automorphisms of X, AutR−R(X)
to the unit group of Z, defined by
AutR−R(X)
σ
g(−)
/ AutR−R(R) = U (Z)
(2)
/ eσ
where for each r ∈ R with unit e ∈ U nit{r}, we have
eσ(r) = X
(e)
σ(xe)yer = X
(e)
rσ(xe)ye.
In particular, for every element t ∈ X and σ ∈ AutR−R(X), we have
Proof. Follows directly from [4, Lemma 2.1].
σ(t) = eσ(e)t, whenever e ∈ U nit{t}.
The Picard group of R is denoted by Pic(R). It consists of all isomorphisms classes of
invertible unital R-bimodules, with multiplication induced by the tensor product. As in the
unital case [10, Theorem 2.(i)], there is a canonical homomorphism of groups α : Pic(R) →
Aut(U (Z)). This homomorphism is explicitly given as follows. Given [X] ∈ Pic(R) and
u ∈ U (Z), we consider the R-bilinear automorphism of X defined by
σu : X −→ X,
(cid:16)t 7−→ t u(e), t ∈ X, and e ∈ U nit{t}(cid:17) .
We clearly have σuv = σu ◦ σv, and so gσuv = fσu ◦fσv, for u, v ∈ U (Z) It follows from Lemma
1.6 that
This equation implies that the element fσu ∈ U (Z) is independent from the choice of the
representative X of the class [X] ∈ Pic(R). We have thus a well defined map
σu(rt) = fσu(r)t, for every t ∈ X, r ∈ R.
(3)
α : Pic(R) → Aut(U (Z)),
[X] 7→ α[X](u) = fσu
which is a homomorphism of groups by (3). This is exactly the homomorphism induced by
the map α of [4, p. 135]. We have the formula:
α[X](u)(r) = fσu(r) = X
(e)
rxeu(e1)ye,
(i.e.X
(e)
rl(xe ⊗R u(e1)ye)),
(4)
where e ∈ U nit{r}, e1 ∈ U nit{xe, ye}, and as before e = P(e) xeye.
2 Generalized crossed products with local units.
Let G be any group with neutral element 1. Consider a ring R with a set of local units E,
and a group homomorphism Θ : G → Pic(R). This is equivalent to give sets of invertible
R-bimodules and isomorphisms of bimodules
{Θx : x ∈ G},
{F Θ
x,y : Θx ⊗R Θy
∼=−→ Θxy, x, y ∈ G},
7
/
/
where Θ(x) = [Θx] for every x ∈ G. Consider the product on the direct sum Lx∈G Θx
determined by the maps F Θ
x,y. There is no reason to expect that this product, after the choice
of the representatives Θx, will be associative. In fact, it becomes an associative multiplication
if and only if the diagram
Θx ⊗R Θy ⊗R Θz
F Θ
xy ⊗Θz
Θxy ⊗R Θz
(5)
Θx⊗F Θ
yz
Θx ⊗R Θyz
F Θ
x,yz
F Θ
xy,z
/ Θxyz
commutes for every x, y, z ∈ G. Even in this case, it is not clear wether this multiplication has
a set of local units. On the other hand, we know that there is an isomorphism of R-bimodules
ι : R → Θ1, so a natural candidate for such a set should be E′ = ι(E). In fact, this is a set of
local units if and only if, for every x ∈ G, the following diagrams are commutative
Θx ⊗R R
≃
Θx
,
R ⊗R Θx
≃
Θx
(6)
'NNNNNNNNNNN
ι⊗Θx
9tttttttttt
F Θ
1,x
Θ1 ⊗R Θx
'NNNNNNNNNNN
Θx⊗ι
9tttttttttt
F Θ
x,1
Θx ⊗R Θ1
The previous discussion suggest then the following definition.
Definition 2.1. A factor map for the morphism Θ : G → Pic(R) consists of sets of invertible
R-bimodules and isomorphisms of bimodules
{Θx : x ∈ G},
{F Θ
x,y : Θx ⊗R Θy
∼=−→ Θxy, x, y ∈ G},
where Θ(x) = [Θx] for every x ∈ G, and one isomorphism of R -- bimodules ι : R → Θ1 such
that the diagrams (5) and (6) are commutative.
Now we come back to our initial homomorphism of groups Θ : G → Pic(R) and its
associated family of isomorphisms {F Θ
x,y}x,y,∈ G. Our next aim is to find conditions under
which we can extract from these data a factor map for R and G. First of all we should assume
that the diagrams of (6) are commutative. We can define using the homomorphism of groups
α : Pic(R) → Aut(U (Z)) given by equation (4), a left G-action on the group of units U (Z),
where Z = EndR−R(R). Written down explicitly the composition α ◦ Θ : G → Pic(R) →
Aut(U (Z)) the outcome action is described as follows. Given a unit e ∈ E, its decomposition
with respect to the invertible R-bimodule Θx is written as e = P(e) xexe ∈ ι−1(Θ1) = R.
Using formula (4), the left G-action on U (Z) is then defined by
G × U (Z)
(x, u)
/ U (Z)
/ xu = hr 7→ P(e) rxeu(e1)xei ,
(7)
where e ∈ U nit{r} and e1 ∈ U nit{xe, xe}. Recall that in this case e1 ∈ U nit{e}, and so
e1 ∈ U nit{r}.
The abelian group of all n-cocycles with respect to this action is denoted by Z n
Θ(G, U (Z));
Θ(G, U (Z)) we denote the n-coboundary subgroup. The n-cohomology group, is
while by Bn
then denoted by H n
Θ(G, U (Z)).
8
/
/
/
/
/
'
9
/
/
'
9
/
/
Remark 2.2. To each homomorphism of groups Θ : G → Pic(R), it corresponds as before
a G-group structure on U (Z), and so it leads to the cohomology groups H ∗
Θ(G, U (Z)). The
nature of this correspondence stills, up to our knowledge, unexplored, even in the case of unital
rings. However, as we will see, it is in fact the starting key point of any theory of generalized
crossed products. Following the commutative and unital case, see Remark 2.13 below, here we
will also extract our results from one fixed G-action, this is to say fixed cohomology groups. For
instance, one could start with the cohomology which corresponds to the trivial homomorphism
of groups, that is, I : G → Pic(R) sending x 7→ [R], of course here the cohomology is not at
all trivial.
Here is a non trivial example.
Example 2.3. ([4]) Let R be a ring with E as a set of local units. Assume that there is
a homomorphism of groups Φ : G → Aut(R) to the group of ring automorphisms of R (we
assume that Φ(x)(E) = E, for every x ∈ G). Composing with the canonical homomorphism
Aut(R) → Pic(R) which sends x 7→ [Rx] (i.e.
the R-bimodule whose underlying left R-
module is RR and its right module structure is r.s = rx(s)), we obtain a homomorphism
Θ : G → Pic(R). The corresponding family of maps {F Θ
x,y}x, y∈G is given by
F Θ
x,y : Rx ⊗R Ry −→ Rxy, (r ⊗R s 7−→ rx(s)) .
Here of course the associativity and unitary properties of the multiplication of R, both imply
that {F Θ
x,y}x, y∈G is actually a factor map for R and G. A routine computation shows now that
the cohomology H ∗
Θ(G, U (Z)) which corresponds to this Θ is exactly the one considerd in [4,
p. 135], since the G-action (7) coincides for this case with that given in [4].
Lemma 2.4. Keep the above G-action on U (Z). Fix an element x ∈ G.
(1) For every t ∈ Θx and u ∈ U (Z), we have
t u(e) = xu(e) t,
where e ∈ U nit{t}.
(2) Given similar bimodules M ∼ Θx; for every m ∈ M and u ∈ U (Z), we have
m u(e) = xu(e) m,
with e ∈ U nit{m}.
(8)
(9)
Proof. (1) This is equation (3).
(2) The identity (8) clearly lifts to finite direct sums Θ(n)
Θ(n)
x .
x , and, hence, for subbimodules M of
Proposition 2.5. Let R be a ring with a set of local units E and G any group. Assume that
there is a homomorphism of groups Θ : G → Pic(R) whose family of maps {F Θ
x,y}x,y ∈G satisfy
the commutativity of diagrams (6). Denote by αx,y,z : Θxyz → Θxyz the isomorphisms that
satisfy
y,z(cid:17) = F Θ
Then ^αx,y,z defines a normalized element of Z 3
Θ(G, U (Z)), where ^αx,y,z is the image of αx,y,z
under the isomorphism of groups stated in Lemma 1.6. Furthermore, if ^α−,−,− is cohomolog-
ically trivial, that is, there is a normalized map σ−,− : G × G → U (Z) such that
x,yz ◦(cid:16)Θx ⊗R F Θ
xy,z ◦(cid:16)F Θ
x,y ⊗R Θz(cid:17) .
αx,y,z ◦ F Θ
(10)
^αx,y,z = σx,yz
xσy,z σ−1
x,y σ−1
xy,z,
9
for every x, y, z ∈ G, then there is a factor map {F
Θ
x,y}x.y ∈G for Θ, defined by
F
Θ
x,y : Θx ⊗R Θy
ux ⊗R uy
/ Θxy
/ σx,y(e)F Θ
x,y(ux ⊗R uy),
where e ∈ U nit{ux, uy}.
Proof. Consider x, y, z, t ∈ G, and ux ∈ Θx, uy ∈ Θy, uz ∈ Θz and ut ∈ Θt, denote by uxuy the
image of ux ⊗R uy under F Θ
x,y. There are several isomorphisms from Θx ⊗R Θy ⊗R Θz ⊗R Θt to
Θxyzt which are defined by the family F Θ
−,−. So it will be convenient to adopt some notations.
In order to do this, we rewrite the stated equation (10) satisfied by the α−,−,−'s, as follows
αx,y,z(cid:16)ux(uyuz)(cid:17) = (cid:16)(uxuy)uz(cid:17) .
Using Lemma 1.6, this is equivalent to
βx,y,z(e)(cid:16)ux(uyuz)(cid:17) = (cid:16)(uxuy)uz(cid:17) ,
where e ∈ U nit{ux, uy, uz} and β−,−,− := ^α−,−,− is the image of αx,y,z under the homomor-
phism of groups defined in Lemma 1.6. In this way, if we fix a unit e ∈ U nit{ux, uy, uz, ut},
then we have
(cid:16)(uxuy)uz(cid:17) ut = βx,y,z(e)(cid:16)ux(uyuz)(cid:17) ut
= βx,y,z(e)βx,yz,t(e)(cid:16)ux(cid:16)(uyuz)ut(cid:17)(cid:17)
= βx,y,z(e)βx,yz,t(e)(cid:16)ux(cid:16)βy,z,t(e)(cid:16)uy(uzut)(cid:17)(cid:17)(cid:17)
= βx,y,z(e)βx,yz,t(e)(cid:16)uxβy,z,t(e)(cid:16)uy(uz ut)(cid:17)(cid:17)
= βx,y,z(e)βx,yz,t(e)(cid:16)xβy,z,t(e)ux(cid:16)uy(uzut)(cid:17)(cid:17)
= βx,y,z(e)βx,yz,t(e) xβy,z,t(e)(cid:16)ux(cid:16)uy(uz ut)(cid:17)(cid:17)
(8)
= βx,y,z(e)βx,yz,t(e) xβy,z,t(e)β−1
= βx,y,z(e)βx,yz,t(e) xβy,z,t(e)β−1
x,y,zt(e)(cid:16)(uxuy)(uz ut)(cid:17)
x,y,zt(e)β−1
xy,z,t(e)(cid:16)(uxuy)uz(cid:17) ut.
Given h ∈ E, using the above notations, we can consider the following units
(h)
h = X
h1 = X
h2 = X
h3 = X
(h1)
(h2)
yh1yh1,
zh2zh2,
th3th3,
(h3)
xhxh,
h1 ∈ U nit{xh, xh};
h2 ∈ U nit{yh1, yh1};
h3 ∈ U nit{zh2, zh2};
h4 ∈ U nit{th3, th3}.
10
/
/
It is clear that h4 is a unit for all the handled elements, in particular h4 ∈ U nit{xh, yh1, zh2 , th3}.
According to the equality showed previously, for every set {xh, yh1, zh2, th3}, we have
(cid:16)(xhyh1)zh2(cid:17) th3 = βx,y,z(h4)βx,yz,t(h4) xβy,z,t(h4)β−1
x,y,zt(h4)β−1
xy,z,t(h4)(cid:16)(xhyh1)zh2(cid:17) th3.
Using the map F Θ
t,t−1, and summing up, we get
(cid:16)(xhyh1)zh2(cid:17) th3th3 = βx,y,z(h4)βx,yz,t(h4) xβy,z,t(h4)β−1
x,y,zt(h4)β−1
xy,z,t(h4)(cid:16)(xhyh1)zh2(cid:17) th3th3,
X
(h3)
which means that
(xhyh1)zh2 = βx,y,z(h4)βx,yz,t(h4) xβy,z,t(h4)β−1
x,y,zt(h4)β−1
xy,z,t(h4)(xhyh1)zh2, equality in Θxyz.
Repeating thrice the same process, we end up with
h = βx,y,z(h4)βx,yz,t(h4) xβy,z,t(h4)β−1
x,y,zt(h4)β−1
xy,z,t(h4)h
= βx,y,z(h)βx,yz,t(h) xβy,z,t(h)β−1
x,y,zt(h)β−1
xy,z,t(h),
since h4h = hh4 = h and the involved maps are R-bilinear. Therefore, for any unit h ∈ E, we
have
βxy,z,t ◦ βx,y,zt(h) = xβy,z,t ◦ βx,yz,t ◦ βx,y,z(h).
Since the β−,−,−'s are R-bilinear, this equality implies the 3-cocycle condition, that is,
βxy,z,t ◦ βx,y,zt = xβy,z,t ◦ βx,yz,t ◦ βx,y,z,
for every x, y, z, t ∈ G.
A routine computation, using equation (8), shows the last statement.
Corollary 2.6. Let R be a ring with local units and G any group, and fix a factor map
{Θx}x ∈G for R and G with its associated cohomology groups. Assume that there is a family of
invertible unital R-bimodules {Ωx}x ∈ G such that Ωx ∼ Θx (they are similar), for every x ∈ G,
with a family of R-bilinear isomorphisms
Fx,y : Ωx ⊗R Ωy −→ Ωxy.
For x, y, z ∈ G, we denote by αx,y,z : Ωxyz → Ωxyz the resulting isomorphisms which satisfy
αx,y,z ◦ Fx,yz ◦(cid:16)Ωx ⊗R Fy,z(cid:17) = Fxy,z ◦(cid:16)Fx,y ⊗R Ωz(cid:17) .
(11)
Then ^αx,y,z defines an element of Z 3
isomorphism of groups stated in Lemma 1.6.
Θ(G, U (Z)), where ^αx,y,z is the image of αx,y,z under the
Proof. This is a direct consequence of Proposition 2.5 and Lemma 2.4(2).
We need to clarify the situation when two different classes of representatives are chosen.
11
Proposition 2.7. Let R be a ring with a set of local units E and G any group. Consider
two factor maps {F Θ
x,y}x,y ∈G such that, for every x ∈ G, there exists an R-
∼=→ Θx. Denote by τx,y, where x, y ∈ G, the following R-bilinear
x,y}x,y ∈G and {F Γ
bilinear isomorphism ax : Γx
isomorphism
τx,y : Θxy
F Θ
x,y
−1
/ Θx ⊗R Θy
∼=
/ Γx ⊗R Γy
F Γ
x,y
/ Γxy
∼=
/ Θxy,
That is,
τx,y ◦ F Θ
x,y ◦ (ax ⊗R ay) = axy ◦ F Γ
x,y.
Then gτx,y defines a normalized element of Z 2
the isomorphism of groups stated in Lemma 1.6.
Θ(G, U (Z)), where gτx,y is the image of τx,y under
Proof. Fix a set of elements x, y, z ∈ G. Assume that we have
x gτx,yz(h) gτx,yz(h) = gτx,y(h)gτxy,z(h),
So taking any element r ∈ R with unit f ∈ U nit{r}, we obtain
for every unit h ∈ E.
(12)
gτx,yz ◦ x gτx,yz(r) = gτx,yz ◦ x gτx,yz(f r)
= x gτx,yz(f )gτx,yz(r)
= x gτx,yz(f )gτx,yz(f )r
= gτx,y(f )gτxy,z(f )r
= gτxy,z(cid:16)gτx,y(f )r(cid:17)
= gτxy,z ◦ gτx,y(r),
(12)
which shows that gτ−,− is a 2-cocycle. In fact gτ−,− is by definition a normalized 2-cocycle.
Equation (12) is fulfilled by following analogue steps as in the proof of Proposition 2.5.
Proposition 2.8. The hypothesis are that of Corollary 2.6. Assume further that there are
two families of invertible unital R-bimodules {Ωx}x ∈ G and {Γx}x ∈ G as in Corollary 2.6 (i.e.
Γx ∼ Θx ∼ Ωx), together with families of R-bilinear isomorphisms
F Ω
x,y : Ωx ⊗R Ωy −→ Ωxy, F Γ
x,y : Γx ⊗R Γy −→ Γxy.
Assume also that there are R-bilinear isomorphisms ax : Ωx → Γx, x ∈ G. Consider the
associated 3-cocycles βΩ
Θ(G, U (Z)) given by Proposition 2.5. Then,
−,−,−, βΓ
−,−,− ∈ Z 3
−,−,− ◦(cid:16)βΓ
βΩ
−,−,−(cid:17)−1
−,−,−] = [βΓ
∈ B2
Θ(G, U (Z)).
That is, they are cohomologous, or [βΩ
−,−,−] in H 3
Θ(G, U (Z)).
Proof. Let us denote by αΩ
−,−,− the R-bilinear isomorphisms satisfying equation
(11), respectively, for {Ωx}x ∈ G and {Γx}x ∈G. For any pair x, y ∈ G, we consider the R-bilinear
isomorphism bxy : Ωxy → Ωxy defined by the following equation
−,−,− and αΓ
axy ◦ bxy ◦ F Ω
x,y = F Γ
x,y ◦(cid:16)ax ⊗R ay(cid:17) .
(13)
12
/
/
/
/
We denote by γx,y, x, y ∈ G, the image of bxy under the homomorphism of groups defined in
Lemma 1.6.
Fix x, y, z ∈ G, and consider (tx, ty, tz) ∈ Ωx × Ωy × Ωz and (ux, uy, uz) ∈ Γx × Γy × Γz with
a common unit e ∈ U nit{tx, ty, tz, ux, uy, uz}. Using the notation of the proof of Proposition
2.5, we can write
βΩ
x,y,z(e) tx(tytz) = (txty)tz.
βΓ
x,y,z(e) ux(uyuz) = (uxuy)uz.
γx,y(e) axy(txty) = ax(tx)ay(ty).
(14)
(15)
(16)
A routine computation as in the proof of Proposition 2.5 using equation (14), (15) and (16),
show that, for every unit h ∈ E, we have
(cid:16)βΩ
x,y,z(h)(cid:17)−1
x,y,z(h) = (cid:16)γx,yz(h)(cid:17)−1(cid:16)xγy,z(h)(cid:17)−1
βΓ
γx,y(h)γxy,z(h).
This implies that
x,y,z ◦(cid:16)βΩ
x,y,z(cid:17)−1
βΓ
= γxy,z ◦ γx,y ◦(cid:16)xγy,z(cid:17)−1
◦(cid:16)γx,yz(cid:17)−1
, in U (Z),
which finishes the proof.
Now we are able to give the right definition of generalized crossed product. We hope it
will result cleaner than the one considered in [13, 15] for rings with identity. In the sequel, R
denotes a ring with a set of local units E.
Definition 2.9. Given a factor map as in Definition 2.1
{F Θ
x,y : Θx ⊗R Θy
∼=−→ Θxy, x, y ∈ G},
with an isomorphism of R-bimodules ι : R → Θ1, we define its associated generalized crossed
product ∆(Θ) = Lx∈G Θx with multiplication
θxθy = F Θ
x,y(θx ⊗ θy), for θx ∈ Θx, θy ∈ Θy.
It follows that Θ1 is a subring of ∆(Θ) and the map ι : R → Θ1 is a ring isomorphism.
Therefore, ι(E) is a set of local units for Θ1 that serves as well as a set of local units for ∆(Θ).
Normally, we will identify R with Θ1, and, thus, E will be a set of local units for the generalised
crossed product ∆(Θ). Obviously, ∆(Θ) is a G -- graded ring such that ΘxΘy = Θxy for every
x, y ∈ G (thus, it could be referred to as a strongly graded ring after [16]). For instance, if we
take Θ as in Example 2.3, then ∆(Θ) is the skew group ring R ∗ G, see [4].
Lx∈G Γx with ring isomorphism ι : R → Γ1 and multiplication given by γxγy = F Γ
Remark 2.10. A factor map as in Definition 2.1 determines a generalized crossed product Γ =
x,y(γx ⊗R γy)
for γx ∈ Γx, γy ∈ Γy. In other words, generalized crossed products and factor maps are just
two ways to define the same mathematical object. A generalized crossed product Γ gives
clearly a group homomorphism
Γ : G
/ Pic(R) ,
(x 7→ [Γx]).
Whether a general group homomorphism G → Pic(R) gives some generalized crossed product
is not so clear. However, as we have seen in Proposition 2.5, this is possible, if the underlying
maps satisfy the commutativity of diagrams (6) and the associated 3-cocycle is trivial or at
least cohomologically trivial.
13
/
Definition 2.11. A morphism of generalized crossed products (∆(Θ), ν), (∆(Γ), ι) is a graded
ring homomorphism f : ∆(Θ) → ∆(Γ) such that f ◦ ν = ι. Equivalently, it consists of a set of
homomorphisms of R -- bimodules
such that all the diagrams
{fx : Θx → Γx : x ∈ G}
Θx ⊗R Θy
fx⊗fy
Γx ⊗R Γy
F Θ
x,y
F Γ
x,y
/ Θxy
fxy
/ Γxy
commute, and f1 ◦ ν = ι.
Let R ⊆ S be an extension of rings with the same set of local units E. We denote as in
[9] by InvR(S) the group of all invertible unital R-sub-bimodules of S. An R-sub-bimodule
X ⊆ S belongs to the group InvR(S) if and only if there exists an R-sub-bimodule Y ⊆ S
such that
XY = R = Y X,
where the first and last terms are obviously defined by the multiplication of S. Clearly, there
is a homomorphism of groups µ : InvR(S) → Pic(R).
Remark 2.12. A generalized crossed product Γ of R with G gives in particular the ring
extension ι : (R, E) → (Γ, E′) with local untis. In this way, Γx is a unital R-subbimodule of
Γ for every x ∈ G, and the isomorphism ι : R → Γ1 becomes R -- bilinear. We will identify
Γ1 with R and E with E′. Let us denote by F Γ : Γ ⊗R Γ → Γ the multiplication map of the
ring Γ. It follows from [9, Lemma 1.1] that its restriction to Γx ⊗R Γy gives an R-bilinear
isomorphism
F Γ
xy : Γx ⊗R Γy
≃ /
/ ΓxΓy = Γxy
for every x, y ∈ G, since, obviously, Γx ∈ InvR(Γ) for every x ∈ G. Clearly, our generalized
crossed product determines a factor map in the sense of Definition 2.1. On the other hand,
the associated homomorphisms of groups Γ : G → Pic(R) actually factors throughout the
morphism µ. That is, we have a commutative diagram of groups
Γ
Pic(R)
(17)
G
D
D
D
Γ
D
D
:uuuuuuuuuuu
µ
!D
InvR(Γ).
As we have seen in Remark 2.12, one can start with a fixed ring extension R ⊆ S with the
same set of local units E, and then work with the Γ's defined this time by Γ : G → InvR(S)
instead of the Γ's. In this case the situation becomes much more manageable, in the sense
that each morphism Γ induces automatically a factor map and vice versa.
14
/
/
/
/
!
:
Remark 2.13. As was shown in [13, Remarks 1., 2. and 3.], the above notion generalizes
the classical notion of crossed product in the commutative Galois setting. Specifically, given a
commutative Galois extension k ⊆ R with a finite Galois group G, and consider the canonical
homomorphism of groups Φ0 : G → Pick(R), where Pick(R) = {[P ] ∈ Pic(R) kp = pk, ∀ k ∈
k, p ∈ P }, which sends x → [Rx] to the class of the R-bimodules Rx whose underlying left
R-module is RR and its right R-module structure was twisted by the automorphism x, see
Example 2.3. In this case, the over ring is chosen to be S = ⊕x∈GΦ0(x) defined as the usual
crossed product attached to Φ0, which is not necessary trivial, that is, defined using any 2-
cocycle. Thus, in the Galois theory of commutative rings, the homomorphism of groups which
leads to the study of intermediate crossed products rings, is the obvious homomorphism of
groups Θ which satisfies Φ0 = µ ◦ Θ, where µ is as in diagram (17).
3 An abelian group of generalized crossed products.
From now on, we fix a ring extension R ⊆ S with the same set of local units E. Let G be any
group and assume given a morphism of groups
Θ : G −→ InvR(S), (cid:16)x 7−→ Θx(cid:17) ,
where we have used the notation of Remark 2.12. The multiplication of S induces then a
factor map
{F Θ
x,y : Θx ⊗R Θy −→ Θxy, : x, y ∈ G}.
Clearly R = Θ1, whence E is a set of local units for ∆(Θ). We introduce in this section
an abelian group C(Θ/R) whose elements are the isomorphism classes of generalized crossed
products whose homogeneous components are similar as R-bimodules to that of ∆(Θ), see
Section 2 for definition. Our definition is inspired from that of Miyashita [15, §2.], however
the proofs presented here are slightly different. This group will be crucial for the construction
of the seven terms exact sequence of groups in the next section.
Let us consider thus the set C(Θ/R) of isomorphism classes of generalized crossed products
[∆(Γ)] such that, for each x ∈ G, we have Γx ∼ Θx, that is they are similar as R-bimodules,
see Section 1.
Proposition 3.1. Consider the set of isomorphisms classes of generalized crossed products
C(Θ/R) defined above. Then C(Θ/R) has the structure of an abelian group. Moreover the
subset C0(Θ/R) consisting of classes [∆(Λ)] such that for every x ∈ G, Λx ∼= Θx as R-
bimodules, is a sub-group of C(Θ/R).
Proof. Given two classes [∆(Ω)], [∆(Γ)] ∈ C(Θ/R), their multiplication is defined by
[∆(Ω)] [∆(Γ)] = [ ⊕
x∈G
Ωx ⊗R Θx−1 ⊗R Γx].
The factor maps of its representative generalized crossed product are given by the following
15
composition
Ωx ⊗R Θx−1 ⊗R Γx ⊗R Ωy ⊗R Θy−1 ⊗R Γy
UUUUUUUUUUUUUUU
Ωx⊗TΘ
!
"
#
%
&
(
*
,
/
2
5
9
=
F ΩΓ
x,y
B
G
K
x−1 ⊗ Γx, Ωy ⊗RΘ
y−1 ⊗Γy
*UUUUUUUUUUUUUUU
Ωx ⊗R Ωy ⊗R Θy−1 ⊗R Θx−1 ⊗R Γx ⊗R Γy
F Ω
x,y ⊗F Θ
y−1,x−1 ⊗F Γ
x,y
N Q T W Y [ ] _ a b d e
Ωxy ⊗R Θ(xy)−1 ⊗R Γxy,
where TΘx−1 ⊗Γx, Ωy⊗RΘy−1 is the twist R-bilinear map defined in Proposition 1.3. The asso-
ciativity of the F ΩΓ
−,−'s is easily deduced using Lemma 1.4. This multiplication is commutative
since for any two classes [∆(Ω)], [∆(Γ)] ∈ C(Θ/R), we have a family of R-bilinear isomorphisms
Γx ⊗R Θx−1 ⊗R Ωx
∼=
Γx ⊗R Θx−1 ⊗R Ωx ⊗R Θx−1 ⊗R Θx
,
/
2
6
;
T⊗Θx
Ωx ⊗R Θx−1 ⊗R Γx ⊗R Θx−1 ⊗R Θx
A
∼=
G
L
Q
∼=
T X [
^
Ωx ⊗R Θx−1 ⊗R Γx,
which is easily shown to be compatible with both factors maps of Γ and Ω. Thus, it induces
an isomorphism at the level of generalized crossed products.
It is clear that the class of [∆(Θ)] is the unit of this multiplication. The inverse of any
class [∆(Ω)] is given by
[∆(Ω)]−1 = [ ⊕
x∈G
Θx ⊗R Ωx−1 ⊗R Θx].
Lastly, the fact that C0(Θ/R) is a sub-group of C(Θ/R), can be immediately deduced from
definitions.
Proposition 3.2. Consider the abelian group C0(Θ/R) of Proposition 3.1. Then there is an
isomorphism of groups
C0(Θ/R) ∼= H 2
Θ(G, U (Z)).
Proof. The stated isomorphism is given by the following map:
ζ : C0(Θ/R) −→ H 2
Θ(G, U (Z)), (cid:16)[∆(Λ)] 7−→ [eτ−,−](cid:17) ,
16
*
2
2
/
/
/
/
where eτ−,− is the normalized 2-cocycle defined in Proposition 2.7 and [eτ−,−] denotes its
equivalence class in H 2
Θ(G, U (Z)). First we need to show that this map is in fact well de-
∼=→ ∆(Σ)
fined. To do so, we take two isomorphic generalized crossed products χ : ∆(Λ)
which represent the same element in the subgroup C0(Θ/R), with R-bilinear isomorphisms
ax : Λx → Θx ← Σx : bx, for all x ∈ G. Then we check that the associated 2-cocycles are
cohomologous. So let us denote these cocycles by eτ−,− and eγ−,− associated, respectively, to
∆(Λ) and ∆(Σ). We need to show that [eτ−,−] = [eγ−,−]. Recall from Proposition 2.7 that we
have
τx,y ◦ F Θ
x,y ◦(cid:16)ax ⊗R ay(cid:17) = axy ◦ F Λ
x,y,
γx,y ◦ F Θ
x,y ◦(cid:16)bx ⊗R by(cid:17) = bxy ◦ F Σ
x,y,
for every x, y ∈ G, where F Λ
Σ. For each x ∈ G, we set the following R-bilinear isomorphism
−,− and F Σ
−,− are, respectively, the factor maps relative to Λ and
βx : Θx
a−1
x
/ Λx
χx /
/ Σx
bx /
/ Θx ,
where χx is the x-homogeneous component of the isomorphism χ. Then it is easily seen, using
the fact that χ is morphism of generalized crossed products, that, for each pair of elements
x, y ∈ G, we have
βxy ◦ τx,y ◦ F Θ
x,y = γx,y ◦ F Θ
x,y ◦(cid:16)βx ⊗R βy(cid:17) .
(18)
(20)
We claim that
So let e ∈ E, and consider its decomposition e = P(e) xexe with respect to Θx, for a given
x ∈ G. Let e1 ∈ U nit{xe, xe}. Using equation (18), we get the following equality in S
eτx,y(e)eβxy(e) = eγx,y(e)eβx(e)xeβy(e),
for all e ∈ E.
(19)
X
(e), (e1)
βxy ◦ τx,y(xeye1)ye1xe = X
(e), (e1)
Routine computations show that,
γx,y(cid:16)βx(xe)βy(ye1)(cid:17) ye1xe.
(e), (e1)
X
γxy(cid:16)βx(xe)βy(ye1)(cid:17) ye1xe
X
(e), (e1)
βxyτx,y(xeye1)ye1xe = eτx,y(e1)eβxy(e1)e = eτx,y(e)eβxy(e), and
(8)
= eγx,y(e1)eβx(e1)xeβy(e1)e = eγx,y(e)eβx(e)xeβy(e),
which by (20) imply that
which is the claimed equality. On the other hand we define
eτx,y(e)eβxy(e) = eγx,y(e)eβx(e)xeβy(e),
h : G
x
/ U (Z) = AutR−R(R)
/ (cid:16)hx : r 7−→ P(e) rβx(xe)xe(cid:17)
where e ∈ U nit{r}, and e = P(e) xexe is the decomposition of e in R = ΘxΘx−1. It is clear
form definitions that eβx = hx, for ever x ∈ G. Taking an arbitrary element r ∈ R, we can
show using equation (19) that
eτx,y ◦ hxy(r) = eγx,y ◦ xhy ◦ hx(r).
17
/
/
/
Therefore, eτx,y ◦ hxy = eγx,y ◦ xhy ◦ hx, which means that eτ−,− and eγ−,− are cohomologous.
By the same way, we show that, if eτ−,− and eγ−,− are cohomologous, then ∆(Λ/R) and
∆(Σ/R) are isomorphic as generalized crossed products, which means that the stated map ζ
is injective. To show that this map is surjective, we start with a given normalized 2-cocycle
σ−,− : G2 → U (Z), and consider the following R-bilinear automorphisms: for every x ∈ G
ϑx : Θx −→ Θx, (cid:16)u 7−→ σx,x(e)u(cid:17) ,
where e ∈ U nit{u}. Now set Σx = Θx as an R-bimodule with the R-bilinear isomorphism
ϑx : Σx → Θx previously defined. We define new factor maps relative to those Σx's, by
x,y : Σx ⊗R Σy → Σxy, (cid:16)ux ⊗R uy 7−→ σx,y(e)F Θ
x,y(ux ⊗R uy)(cid:17) ,
F Σ
where e ∈ U nit{ux, uy}. The 2-cocycle condition on σ−,− gives the associativity of F Σ
−,−, while
the normalized condition gives the unitary property. Thus, ∆(Σ/R) is a generalized crossed
product whose isomorphic class belongs by definition to the sub-group C0(Θ/R).
Lastly, by a routine computation, using the definition of the twist natural map given in
Proposition 1.3, we show that the map ζ is a homomorphism of groups.
4 The Chase-Harrison-Rosenberg seven terms exact sequence.
In this section we show the analogue of Chase-Harrison-Rosenberg's [5] seven terms exact
sequence, generalizing by this the case of commutative Galois extensions with finite Galois
group due of T. Kanzaki [13, Theorem p.187] and the noncommutative unital case treated by
Y. Miyashita in [15, Theorem 2.12].
From now on we fix a ring extension R ⊆ S with a same set of local units E, together
with a morphism of groups Θ : G → InvR(S) and the associated generalized crossed product
∆(Θ/R) having F Θ
−,− as a factor map relative to Θ.
We denote by Pic(R) and Pic(S), respectively, the Picard group of R and S. There is a
canonical left G-action on Pic(R) induced by the obvious homomorphism of groups InvR(S) →
Pic(R) given explicitly by
x[P ] = [Θx ⊗R P ⊗R Θx−1],
for every x ∈ G, and [P ] ∈ Pic(R).
The corresponding G-invariant sub-group is denoted by Pic(R)G. We will consider the sub-
group PicZ (R) whose elements are represented by Z-invariant R-bimodules, in the sense that
[P ] ∈ PicZ (R), if and only if
[P ] ∈ Pic(R), and z(e)pe′ = epz(e′), ∀p ∈ P,
for every pair of units e, e′ ∈ E, and every element z ∈ Z. In this way, we set
PicZ (R)G := PicZ (R) ∩ Pic(R)G.
Recall from [9, Sect. 3] the group P(S/R):
P(S/R) = n [P ]
[φ]
3 [X] [P ] ∈ Pic(R), [X] ∈ Pic(S), and a map φ : RPR → RXRo
18
+
where at least one of the maps φl : P ⊗R S → X, or φr : S ⊗R P → X, canonically attached
to φ, is an isomorphisms of bimodules. For more details on the structure group of this set we
refer to [9, Section 3.]. Obviously there is a homomorphism of groups
Ol : P(S/R) −→ Pic(R),(cid:16)( [P ]
[φ]
3 [X] ) 7−→ [P ](cid:17) .
(21)
This group also inherits the above G-action. That is, there is a canonical left G-action on this
group which is induced by Θ and given as follows:
x(cid:16) [P ]
[φ]
3 [X] (cid:17) = (cid:16) [Θx ⊗R P ⊗R Θx−1]
[ψ]
3 [X] (cid:17) ,
where ψ := Θx⊗Rφ⊗RΘx−1 (composed with the multiplication of S). The subgroup P(S/R)G
of G-invariant elements of P(S/R) has a sightly simpler description:
Lemma 4.1. Keeping the above notations, we have
P(S/R)G = n [P ]
[φ]
3 [X] ∈ P(S/R) φ(P )Θx = Θxφ(P ), for every x ∈ Go ,
where φ(P )Θx and Θxφ(P ) stand for the obvious subsets of the S-bimodule X, e.g.
φ(P ) Θx =
X
f inite
φ(pi)ui
x pi ∈ P, ui
x ∈ Θx
.
Proof. Is based up on two main facts. The first one is that the R-bilinear map φ : P → X
3 [X] ∈ P(S/R), is always injective, see [9, Lemma 3.1].
defining the given element [P ]
The second fact consists on the following equivalence: For a fixed x ∈ G, saying that φ(P )Θx =
Θxφ(P ) in SXS is equivalent to say that there is an isomorphism fx : P ∼= Θx ⊗R P ⊗R Θx−1
which completes the commutativity of the following diagram
[φ]
P
fx
φ
Θx ⊗R P ⊗R Θx−1
/ Θx ⊗R X ⊗R Θx−1
Θx⊗φ⊗Θx−1
Y Y Z Z [ [ \ \ ] ] ^ ^ _ _ _ ` ` a a b b c c d d e e f f
/ S ⊗R X ⊗R S
ψ
X
/ X.
The later means, by the definition of P(S/R), that
(cid:16) [P ]
[φ]
3 [X] (cid:17) = (cid:16) [Θx ⊗ P ⊗ Θx−1]
[ψ]
3 [X] (cid:17) .
4.1 The first exact sequence.
Now we set
PZ (S/R) = n [P ]
[φ]
3 [X] ∈ P(S/R) [P ] ∈ PicZ (R)o ,
19
+
+
+
+
+
/
/
3
3
/
/
+
+
+
and
PZ (S/R)G = PZ (S/R) ∩ P(S/R)G.
On the other hand, we consider the group AutR−ring(S) of all R-ring automorphisms of
S (with same set of local units). It has the following subgroup
AutR−ring(S)(G) = nf ∈ AutR−ring(S) f (Θx) = Θx, for all x ∈ Go .
There are two interesting maps which will be used in the sequel: The first one is given by
E : AutR−ring(S) −→ P(S/R),(cid:16)f 7−→ ( [R]
[ιf ]
3 [Sf ] )(cid:17) ,
where ιf is the inclusion R ⊂ Sf , and Sf is the S-bimodule induced from the automorphism
f , that is, the left S-action is that of SS, while the right one has been altered by f , that is,
we have a new left S-action
s.s′ = s f (s′), for all s, s′ ∈ S.
The second connects the group U (Z) with AutR−ring(S), and it is defined by
F : U (Z) = AutR−R(R) −→ AutR−ring(S),(cid:16)σ 7−→ hF(σ) : s 7→ σ−1(e) s σ(e)i(cid:17) ,
(22)
where e ∈ U nit{s}, s ∈ S.
Proposition 4.2. Keeping the previous notations, there is a commutative diagram whose rows
are exact sequences of groups
U (Z) = AutR−R(R)
F
/ AutR−ring(S)
E
/ P(S/R)
Ol
/ Pic(R)
U (Z) = AutR−R(R)
/ AutR−ring(S)(G)
PZ (S/R)G
PicZ (R)G .
Proof. The exactness of the first row was proved in [9, Proposition 5.1]. The commutativity as
well as the exactness of the second row, using routine computations, are immediately deduced
from the definition of the involved maps.
The conditions under assumption, that is, the existence of Θ : G → InvR(S) a homo-
morphism of groups with a given extension of rings R ⊆ S with the same set of local units
E, are satisfied for the extension R ⊆ ∆(Θ). Precisely, the map Θ factors thought out a
′
homomorphism of groups Θ
rings with the same set of local units E. Thus, applying Proposition 4.2 to this extension, we
obtain the first statement of the following corollary.
: G → InvR(∆(Θ)),(cid:16)x 7→ Θx(cid:17), as R ⊆ ∆(Θ) is an extension of
Corollary 4.3. Consider the generalized crossed product ∆ := ∆(Θ) of S with G. Then there
is an exact sequence of groups
1
/ U (Z)
/ AutR−ring(∆)(G)
/ PZ (∆/R)G
/ PicZ (R)G.
In particular, we have the following exact sequence of groups
1
/ H 1
Θ(G, U (Z))
S1
/ PZ (∆/R)G
S2
/ PicZ (R)G.
20
+
/
/
/
/
/
/
?
O
/
/
?
O
?
O
/
/
/
/
/
/
/
Proof. We only need to check the particular statement. So let f ∈ AutR−ring(∆)(G), this
gives a family of R-bilinear isomorphisms fx : Θx → Θx, x ∈ G, which by Lemma 1.6 they
lead to a map
σ : G −→ U (Z), (cid:16)x 7−→ efx(cid:17) ,
as Θx ∈ InvR(S). An easy verification, using (8), shows that, for every unit e ∈ E, we have
Therefore,
efxy(e) = efx(e)xefy(e).
σxy = σx ◦ xσy,
for every x, y ∈ G.
(23)
That is, σ is a normalized 1-cocycle, i.e. σ ∈ Z 1
Θ(G, U (Z)).
Conversely, given a normalized 1-cocycle γ ∈ Z 1
Θ(G, U (Z)), we consider the R-bilinear
isomorphisms
gx : Θx −→ Θx, (cid:16)u 7−→ γx(e)u(cid:17) , where e ∈ U nit{u}, u ∈ Θx.
Clearly, the direct sum g := ⊕x∈Ggx defines an automorphism of generalized crossed product.
Hence, g ∈ AutR−ring(∆)(G).
In conclusion we have constructed a mutually inverse maps which in fact establishes an
isomorphism of groups
AutR−ring(∆)(G) ∼= Z 1
Θ(G, U (Z)).
(24)
Now let us consider en element u ∈ U (Z) = AutR−R(R), and its image F(u) under the map
F defined in equation (22) using the extension R ⊆ ∆. Then, for every element a ∈ Θx, we
have
F(u)(a) = u−1(e) a u(e),
e ∈ U nit{a}
(8)
= u−1(e) xu(e) a.
Therefore, u defines, under the isomorphism of (24), a 1-coboundary,
G −→ U (Z), (cid:16)x 7−→ xu ◦ u−1(cid:17) .
We have then constructed an isomorphism of groups
AutR−ring(∆)(G)/AutR−R(R) ∼= H 1
Θ(G, U (Z)).
The exactness follows now from the first statement which is a particular case of Proposition
4.2.
4.2 The second exact sequence.
For any element [P ] ∈ Pic(R), we will denote by [P −1] its inverse element. Let us consider
the following groups
PicZ (R)(G)
:= n[P ] ∈ PicZ (R) P ⊗R Θx ⊗R P −1 ∼ Θx,
for all x ∈ Go .
21
A routine computation shows that this is a subgroup of PicZ (R) which contains the sub-
group of all G-invariant elements. That is, there is a homomorphism of groups PicZ (R)G →
PicZ (R)(G).
Given an element [P ] ∈ PicZ (R)(G) and set Ωx := P ⊗R Θx ⊗R P −1, for every x ∈ G.
Consider now the family of R-bilinear isomorphisms
F Ω
x,y : Ωx ⊗R Ωy → Ωxy
x,y and the isomorphism P ⊗RP −1 ∼= R. This is clearly
which is defined using the factor maps F Θ
a factor map relative to the homomorphism of groups Ω : G → InvR(S) sending x 7→ Ωx.
This gives us a generalized crossed product ∆(Ω), and in fact establishes a homomorphism of
groups
L : PicZ (R)(G) → C(Θ/R),
where the right hand group was defined in Section 3. The proof of the following lemma is left
to the reader.
Lemma 4.4. Keep the above notations. There is a commutative diagram of groups
PicZ (R)G
PicZ (R)(G)
L
L
/ C0(Θ/R)
/ C(Θ/R).
Now we can show our second exact sequence.
Proposition 4.5. Let ∆(Θ) be a generalized crossed product of S with G. Then there is an
exact sequence of groups
PZ (∆(Θ)/R)G
S2
/ PicZ (R)G
S3
/ C0(Θ/R),
where S2 and S3 are, respectively, the restrictions of the map Ol given in equation (21) attached
to the extension R ⊆ ∆(Θ), and of the map defined in Lemma 4.4.
Proof. Let [P ] ∈ PicZ (R)G such that S3([P ]) = [∆(Ω)] = [∆(Θ)], where for every x ∈ G,
Ωx = P ⊗RΘx⊗RP −1. This means that we have an isomorphism ∆(Ω) ∼= ∆(Θ) of generalized
crossed products, and so a family of R-bilinear isomorphisms ιx : Θx ⊗R P ∼= P ⊗R Θx, x ∈ G.
We then obtain an R-bilinear isomorphism
ι = ⊕
x∈ G
ιx : ∆(Θ) ⊗R P −→ P ⊗R ∆(Θ),
which in fact satisfies the conditions stated in [8, Eq. (5.1) and (5.2), p. 161]. This implies
that ∆(Θ) ⊗R P admits a structure of ∆(Θ)-bimodule whose underlying left structure is given
by that of ∆(Θ)∆(Θ). In this way, the previous map ι becomes an isomorphism of (∆(Θ), R)-
bimodules, and so we have
(∆(Θ) ⊗R P ) ⊗∆(Θ) (P −1 ⊗R ∆(Θ)) ∼= (∆(Θ) ⊗R P ) ⊗∆(Θ) (∆(Θ) ⊗R P −1)
∼= ∆(Θ) ⊗R P ⊗R P −1
22
_
/
_
/
/
/
∼= ∆(Θ),
an isomorphism of ∆(Θ)-bimodules. Whence the isomorphic class of ∆(Θ) ⊗R P belongs
to Pic(∆(Θ)). We have then construct an element [P ]
3 [∆(Θ) ⊗R P ] of the group
P(∆(Θ)/R), where φ is the obvious map
[φ]
φ : P −→ ∆(Θ) ⊗R P, (cid:16)p 7−→ e ⊗R p(cid:17) ,
where e ∈ U nit{p}. It is clear that, for every x ∈ G, the equality φ(P )Θx = Θxφ(P ) holds
true in ∆(Θ) ⊗R P . Thus,
3 [∆(Θ) ⊗R P ] ∈ PZ (∆(Θ)/R)(G) and
[P ]
[φ]
S2(cid:16) [P ]
[φ]
3 [∆(Θ) ⊗R P ] (cid:17) = [P ].
This shows the inclusion Ker(S3) ⊆ Im(S2).
Conversely, let [Q] ∈ Im(S2), that is, there exists an element [Q]
[ψ]
3 [Y ] ∈ PZ (∆(Θ)/R)G.
We need to compute the image S3([Q]). From the choice of [Q] we know that [Q] ∈ PicZ (R)
and that, for every x ∈ G, ψ(Q)Θx = Θxψ(Q) in the ∆(Θ)-bimodule Y . As in the proof of
Lemma 4.1, the later means that, there are R-bilinear isomorphisms:
∼=−→ Θx ⊗R Q, for every x ∈ G,
fx : Q ⊗R Θx
which convert commutative the following diagrams
Q ⊗R Θx ⊗R Θy
Q⊗F Θ
x,y
Q ⊗R Θxy
'NNNNNNNNNNNN
7pppppppppppp
Y
fx⊗Θy
fxy
Θx ⊗R Q ⊗R Θy
Θxy ⊗R Q
*TTTTTTTTTTTTTTTT
Θx⊗fy
5kkkkkkkkkkkkkk
x,y⊗Q
F Θ
Θx ⊗R Θy ⊗R Q
Coming back to the image S3([Q]) := [∆(Γ)], we know that for every x ∈ G, we have Γx =
Q ⊗R Θx ⊗R Q−1 ∼= Θx, via the fx's. The above commutative diagram shows that these
x,y. Therefore, ∆(Θ) ∼= ∆(Γ)
isomorphisms are compatible with the factor maps F Γ
as generalized crossed products, and so S3([Q]) = [∆(Θ)] the neutral element of the group
C0(Θ/R).
x,y and F Θ
4.3 Comparison with a previous alternative exact sequence.
σ
→ Aut(R) → Pic(R), as in Example 2.3. We consider
Let us assume here that Θ : G
the extension R ⊂ RG, where S := RG = ∆(Θ) is the skew group ring of R by G. Thus
Θ : G → InvR(S) factors though out µ as in diagram (17).
In what follows, we want to
explain the relation between the five terms exact sequence given in [4, Theorem 2.8], and the
resulting one by combining Proposition 4.5 and Corollary 4.3, that is the sequence
1
/ H 1
Θ(G, U (Z))
S1
/ PZ (∆(Θ)/R)G
/ PicZ (R)G
S3
/ H 2
Θ(G, U (Z)) ∼= C0(Θ/R).
S2
23
+
+
+
+
/
/
t
*
*
5
/
/
/
/
As we argued in Example 2.3, the cohomology H ∗
[4]. Thus, following the notation of [4], the result [4, Theorem 2.8] says that
Θ(G, U (Z)) coincides with that considered in
1
/ H 1
Θ(G, U (Z))
ϕσ /
/ G-Pic(R)
Fσ /
/ Pic(R)G Φσ /
/ H 2
Θ(G, U (Z))
It is not difficult to see that the restriction of Φσ
is an exact sequence (of pointed sets).
to PicZ (R)G gives exactly S3. On the other hand, for any element
3 [X] ∈
PZ (∆(Θ)/R)G, we will define an object (P, τ ) where τ : G → AutZ(P ) (to the group of
additive automorphisms of P ) is a G-homomorphism [4, Definition 2.3]. As in the proof of
Lemma 4.1, we know that there is a family of R-bilinear isomorphisms jx : Rx⊗RP → P ⊗RRx,
x ∈ G, so we define τ by the following composition
[P ]
[φ]
τx : x−1P
∼= /
/ Rx ⊗R P
jx
/ P ⊗R Rx
∼= /
/ P x,
here P y denotes the R-bimodule constructed from P by twisting its right module structure
using the automorphism y, and the same construction is used on the left. To check that τ is
a homomorphism of groups, one uses the equality
jxy ◦ (F Θ
x,y ⊗R P ) = (P ⊗R F Θ
x,y) ◦ (jx ⊗R Ry) ◦ (Rx ⊗R jy),
for all x, y ∈ G.
Now, if we take another representative, that is, if we assume that
[P ]
[φ]
3 [X] = [P ′]
[φ′]
3 [X ′] ∈ PZ (∆(Θ)/R)G ,
then, by the definition of the group P(∆(Θ)/R), see [9, Section 3], there are bilinear isomor-
phisms f : P → P ′ and g : X → X ′ such that
P
f
P ′
φ
X
g
/ X ′
φ′
commutes. In this way we have also the following commutative diagram
Rx ⊗R P
Rx⊗f
Rx ⊗R P ′
jx
j′
x
P ⊗R Rx
f ⊗Rx
/ P ′ ⊗R Rx,
x ◦ f = f ◦ τx. This implies that [P, τ ] = [P ′, τ ′] in the group G-Pic(R).
which says that τ ′
We then have constructed a monomorphism of groups PZ (∆(Θ)/R)G → G-Pic(R) such that
the restriction of the map Fσ to PZ (∆(Θ)/R)G is exactly S2. We leave to the reader to check
that we furthermore obtain a commutative diagram
1
1
/ H 1
Θ(G, U (Z))
S1
/ PZ (∆(Θ)/R)G
S2
/ PicZ (R)G
S3
/ H 2
Θ(G, U (Z))
/ H 1
Θ(G, U (Z))
ϕσ
/ G-Pic(R)
Fσ
/ Pic(R)G
Φσ
/ H 2
Θ(G, U (Z)).
24
/
+
/
+
+
/
/
/
/
/
/
/
/
_
/
_
/
/
/
/
/
4.4 The third exact sequence.
We define the group B(Θ/R) as the quotient group
PicZ (R)(G)
L
/ C(Θ/R)
/ B(Θ/R)
/ 1,
(25)
where L is the map defined in Lemma 4.4.
The third exact sequence of groups is stated in the following
Proposition 4.6. Let ∆(Θ) be a generalized crossed product of S with G. Then there is an
exact sequence of groups
PicZ (R)G
S3
/ C0(Θ/R)
S4
/ B(Θ/R),
where S3 and S4 are constructed from the sequence (25) and the commutativity of the diagram
stated in Lemma 4.4.
Proof. The inclusion Im(S3) ⊆ Ker(S4) is by construction trivial. Now, let [∆(Γ)] ∈ C0(Θ/R)
such that S4([∆(Γ)]) = 1. By the definition of the group B(Θ/R), there exists [P ] ∈
PicZ (R)(G) such that L([P ]) = [∆(Γ)]. Therefore, we obtain a chain of R-bilinear iso-
morphisms:
Θx ∼= Γx ∼= P ⊗R Θx ⊗R P −1, for all x ∈ G,
which means that [P ] ∈ PicZ (R)G, and so [∆(Γ)] = S3([P ]) which completes the proof.
Remark 4.7. In the case of a unital commutative Galois extension k ⊆ R with a finite Galois
group, see for instance [3, p. 396], and taking the extension R ⊆ S and Θ as in Remark
2.13. The homomorphism S4 coincides, up to the isomorphism of Proposition 3.2, with the
homomorphism of groups defined in [3, Theorem A 12, p. 406] and where the group B(Θ/R)
coincides with the Brauer group of the k-Azumaya algebras split by R.
4.5 The fourth exact sequence.
Consider the following subgroup of the Picard group Pic(R):
Pic0(R) = n[P ] ∈ Pic(R) P ∼ R, as bimoduleso .
As we have seen in Proposition 1.3, this is an abelian group. Clearly it inherits the G-module
structure of Pic(R): The action is given by
x[P ] = [Θx ⊗R P ⊗R Θx−1], for every x ∈ G.
Lemma 4.8. Keep the above notations. Then the map
C(Θ/R)
[∆(Γ)]
ζ
/ Z 1(G, Pic0(R))
/ hx 7−→ [Γx][Θx−1]i
defines a homomorphism of groups. Furthermore, there is an exact sequence of groups
1
/ C0(Θ/R)
/ C(Θ/R)
ζ
/ Z 1(G, Pic0(R)).
25
/
/
/
/
/
/
/
/
/
/
Proof. The first claim is immediate from the definitions, as well as the inclusion C0(Θ/R) ⊆
Ker(ζ). Conversely, given an element [∆(Γ)] ∈ C(Θ/R) such that ζ([∆(Γ)]) = 1. This
implies that for every x ∈ G, there is an isomorphism of R-bimodule Γx ⊗R Θx−1 ∼= R. Hence,
Γx ∼= Θx, for every x ∈ G, which means that [∆(Γ)] ∈ C0(Θ/R).
We define another abelian group, which in the commutative Galois case [13], coincides with
the first cohomology group of G with coefficients in Pic0(R). It is defined by the following
commutative diagram
PicZ (R)(G)
/__________
Z 1(G, Pic0(R))
1
/ H
(G, Pic0(R))
/ 1
&MMMMMMMMMMM
L
7oooooooooooo
ζ
C(Θ/R)
whose row is an exact sequence. Our fourth exact sequence is given by the following.
Proposition 4.9. Let ∆(Θ) be a generalized crossed product of S with G. Then there is an
exact sequence of groups
C0(Θ/R)
S4
/ B(Θ/R)
S5
1
/ H
(G, Pic0(R)).
Proof. First we need to define S5. We construct it as the map which completes the commuta-
tivity of the diagram
C0(Θ/R)
S4
)RRRRRRRRRRRRRR
Lc
PicZ (R)(G)
L
/ C(Θ/R)
ζ
B(Θ/R)
S5
Z 1(G, Pic0(R))
1
/ H
(G, Pic0(R))
/ 1
/ 1,
whose first row is exact.
The fact that Im(S4) ⊆ Ker(S5) is easily deduced from the previous diagram and Lemma
4.8. Conversely, let Ξ = Lc([∆(Γ)]), for some [∆(Γ)] ∈ C(Θ/R), be an element in B(Θ/R)
(here f c denotes the cokernel map of f ), such that S5(Ξ) = S5 ◦ Lc([∆(Γ)]) = 1. Then
(ζ L)c ◦ ζ([∆(Γ)]) = 1, which implies that ζ([∆(Γ)]) ∈ ζ(cid:16)L(PicZ (R)(G))(cid:17). Therefore, there
exists an element [P ] ∈ PicZ (R)(G) such that [∆(Γ)]L([P ])−1 ∈ Ker(ζ) = C0(Θ/R) by
Lemma 4.8. Now, we have
S4(cid:16)[∆(Γ)]L([P ])−1(cid:17) = S4(cid:16)[∆(Γ)]L([P −1])(cid:17)
= Lc([∆(Γ)])LcL([P −1])
= Lc([∆(Γ)]) = Ξ,
and this shows that Ξ ∈ Im(S4).
26
/
&
/
/
7
/
/
_
)
/
/
/
/
/
/
4.6 The fifth exact sequence and the main Theorem.
Keep from Subsection 4.5 the definition of the group Pic0(R) with its structure of G-module.
Before stating the fifth sequence, we will need first to give a homomorphism of group from
the group H
(G, Pic0(R)) to the third cohomology group H 3
Θ(G, U (Z)).
Given a normalized 1-cocycle g ∈ Z 1(G, Pic0(R)) and put gx = [∇x]. Then, for every
1
pair of elements x, y ∈ G, one can easily shows that
xgy [Θx] = [Θx] gy.
(26)
Now, for every x ∈ G, we set
By the cocycle condition, we obtain
[Ux]
:= gx[Θx], in Pic(R).
[Uxy] = gxy[Θxy]
xgy[Θx][Θy]
= gx
(26)
= gx[Θx] gy[Θy]
= [Ux][Uy].
This means that there are R-bilinear isomorphisms
F g
x,y : Ux ⊗R Uy −→ Uxy,
(27)
1,x = id = F g
with F g
x,1 for every x ∈ G. By Proposition 2.5, we have a 3-cocycle in
Θ(G, U (Z)) attached to these maps F g
Z 3
If there is an-
other class [Vx] ∈ Pic(R) such that [Vx] = gx[Θx], for every x ∈ G. Then the families
of invertible R-bimodules {Vx}x ∈G and {Ux}x ∈G satisfy the conditions of Proposition 2.8.
Therefore, the associated 3-cocycles are cohomologous. This means that the correspondence
g 7→ [αg
Θ(G, U (Z)) is a well defined map. Henceforth, we have a well defined
homomorphism of groups
−,−, which we denote by αg
−,−,−] ∈ H 3
−,−,−.
Z 1(G, Pic0(R))
S13
g
/ H 3
Θ(G, U (Z))
/ [αg
−,−,−].
(28)
The proof of the fact that S13 is a multiplicative map uses Lemma 1.1(ii) and the twisted
natural transformation of Proposition 1.3. Completes and detailed steps of this proof are
omitted.
Proposition 4.10. Let R ⊆ S be an extension of rings with the same set of local units, and
∆(Θ) a generalized crossed product of S with G. Then the homomorphism of equation (28)
satisfies S13 ◦ ζ = [1], where ζ is the homomorphism of Lemma 4.8. Furthermore, there is a
commutative diagram of groups
PicZ (R)(G)
(ζ◦L)
/ Z 1(G, Pic0(R))
1
/ H
(G, Pic0(R))
/ 1
S13
vl l l l l l l l l l
S6
H 3
Θ(G, U (Z))
with exact row.
27
/
/
/
/
v
/
Proof. We need to check that S13 ◦ ζ([∆(Γ)]) = [1], for any element [∆(Γ)] ∈ C(Θ/R). Set
f = ζ([∆(Γ)]), so we have fx[Θx] = [Γx], for every x ∈ G. Thus, the corresponding maps
F f
−,− of ∆(Γ) relative to Γ. They
define an associative multiplication on ∆(Γ), and so induce a trivial 3-cocycle, which means
that S13(f ) = [1]. The last statement is now clear.
−.− of equation (27), are exactly given by the factors maps F Γ
At this level we are ending up with the following commutative diagram
PicZ (R)G
1
/ PicZ (R)(G)
S3
L
/ C0(Θ/R)
/ C(Θ/R)
ζ
S4
*UUUUUUUUUUUUUUUU
Lc
/ B(Θ/R)
S5
PicZ (R)(G)
(ζ◦L)
/ Z 1(G, Pic0(R))
(ζ◦L)c
1
/ H
(G, Pic0(R))
1
/ 1
S13
tjjjjjjjjjjjjj
S6
H 3
Θ(G, U (Z))
Proposition 4.11. Let ∆(Θ) be a generalized crossed product of S with G. Then there is an
exact sequence of groups
B(Θ/R)
S5
1
/ H
(G, Pic0(R))
S6
/ H 3
Θ(G, U (Z)).
Proof. It is clear from the above diagram that Im(S5) ⊆ Ker(S6) since by Proposition 4.10
(G, Pic0(R)) such that S6([h]) = 1,
S13 ◦ ζ = [1]. Conversely, let us consider a class [h] ∈ H
for some element h ∈ Z 1(G, Pic0(R)). By the same diagram, we also have that the class
S13(h) = [1] ∈ H 3
Θ(G, U (Z)), and so
there exists σ−,− : G × G → U (Z) such that
Θ(G, U (Z)). This says that S13(h) = [β], for some β ∈ B2
1
βx,y,z = σx,yz
xσy,z σ−1
x,y σ−1
xy,z,
for every x, y, , z ∈ G. Now, for each x ∈ G, we put hx[Θx] = [Ωx], and consider F h
associated family of maps as in equation (27), F h
definition the β−,−,−'s satisfy
−,− the
x,y : Ωx ⊗R Ωy → Ωxy, where x, y ∈ G. By
βx,y,z ◦ F h
x,yz ◦(cid:16)F h
x,y ⊗R Ωz(cid:17) = F h
xy,z ◦(cid:16)Ωx ⊗R F h
y,z(cid:17) ,
for every x, y, , z ∈ G. In this way, there are factor maps F Ω
−,− relative to Ω, defined by
F Ω
x,y : Ωx ⊗R Ωy
ωx ⊗R ωy
/ Ωxy
/ σx,y(e)F h
x,y(ωx ⊗R ωy),
where e ∈ U nit{ωx, ωy}. A routine computation shows that ∆(Ω) is actually a generalized
crossed product of S with G with factor maps F Ω
−,− relative to Ω.
On the other hand, since for each x ∈ G, hx ∈ Pic0(R), we have Ωx ∼ Θx. Thus
[∆(Ω)] ∈ C(Θ/R) with ζ([∆(Ω)]) = h. Whence
(ζ L)c(h) = [h] = (ζ L)c ◦ ζ([∆(Ω)]) = S5(Lc([∆(Ω)])),
which implies that [h] ∈ Im(S5), and this completes the proof.
28
_
/
_
*
/
/
/
/
/
/
/
t
/
/
/
/
/
The following diagram summarizes the information we have show so far.
H 1
Θ(G, U (Z))
S1
/ PZ (∆/R)
S2
/ PicZ (R)G
S3
/ H 2
Θ(G, U (Z))
PicZ (R)(G)
L
/ C(Θ/R)
/ B(Θ/R)
S4
)TTTTTTTTTTTTT
Lc
PicZ (R)(G)
(ζ◦L)
/ Z 1(G, Pic0(R))
(ζ◦L)c
1
/ H
(G, Pic0(R))
ζ
S5
S13
ujjjjjjjjjjjj
S6
H 3
Θ(G, U (Z))
We are now in position to announce our main result.
Theorem 4.12. Let R ⊆ S be an extension of rings with the same set of local units, and
∆(Θ) a generalized crossed product of S with a group G. Then there is an exact sequence of
groups
1
/ H 1
Θ(G, U(Z))
S1
/ PZ (∆/R)G
S2
/ PicZ (R)G
S3
/ H 2
Θ(G, U(Z))
S4
/ B(Θ/R)
S5
1
/ H
(G, Pic0(R))
S6
/ H 3
Θ(G, U(Z)).
Proof. This is a direct consequence of Corollary 4.3 and Propositions 4.5, 4.6, 4.9, and 4.11.
References
[1] G. D. Abrams, Morita equivalence for rings with local units, Commun. Algebra 11 (1983),
no. 8, 801 -- 837.
[2] P.N. Ánh and L. Márki, Rees matrix rings, J. Algebra 81 (1983), 340 -- 369.
[3] M. Auslander and O. Goldman, The Brauer group of a commutative ring, Trans. Amer.
Math. Soc. 97 (1960), 367 -- 409.
[4] M. Beattie and A. Del Río, Graded equivalences and Picard groups, J. Pure Appl. Algebra,
141 (1999), 131 -- 152.
[5] S.U. Chase, D.K. Harrison and A. Rosenberg, Galois Theory and Galois homology of
commutative rings, Mem. Amer. Math. Soc. 58 (1965), 15 -- 33.
[6] S.U. Chase and D.K. A. Rosenberg, Amitsur cohomology and the Brauer groups, Mem.
Amer. Math. Soc. 58 (1965), 34 -- 79.
[7] L. El Kaoutit, Corings over rings with local units, Math. Nachr. 282, N. 5, 726 -- 747,
(2009).
29
/
/
_
/
_
)
/
/
/
/
u
/
/
/
/
/
/
/
[8] L. El Kaoutit, Extended distributive law. Cowreath over corings., J. Algebra Appl. 9
(2010), 135 -- 171.
[9] L. El Kaoutit and J. Gómez-Torrecillas, Invertible unital bimodules over rings with local
units, and related exact sequences of groups. J. Algebra 323 (2010), 224 -- 240.
[10] A. Fröhlich, The Picard group of noncommutative rings, in particular of orders, Trans.
Amer. Math. Soc. 180 (1973), 1 -- 45.
[11] K. Hirata, Some types of separable extensions of rings, Nagoya Math. J. 33 (1968), 108 --
115.
[12] K. Hirata, Separable extensions and centralizers of rings, Nagoya Math. J. 35 (1969),
31 -- 45.
[13] T. Kanzaki, On generalized crossed product and Brauer group, Osaka J. Math. 5 (1968),
175 -- 188.
[14] Y. Miyashita, On Galois extensions and crossed product, J. Fac. Sci. Hokkaido Univ., Ser.
I, 21 (1970), 97 -- 121.
[15] Y. Miyashita, An exact sequence associated with generalized crossed product, Nagoya
Math. J. 49 (1973), 21 -- 51.
[16] C. Năstăsescu and F. Van Oystaeyen, Graded Ring Theory. North-Holland Mathematical
Library, Vol. 28, 1982.
30
|
1111.6625 | 4 | 1111 | 2012-05-08T09:44:24 | An elementary proof of the vanishing of the second cohomology of the Witt and Virasoro algebra with values in the adjoint module | [
"math.RA",
"math-ph",
"math-ph",
"math.QA",
"math.RT"
] | By elementary and direct calculations the vanishing of the (algebraic) second Lie algebra cohomology of the Witt and the Virasoro algebra with values in the adjoint module is shown. This yields infinitesimal and formal rigidity or these algebras. The first (and up to now only) proof of this important result was given 1989 by Fialowski in an unpublished note. It is based on cumbersome calculations. Compared to the original proof the presented one is quite elegant and considerably simpler. | math.RA | math |
AN ELEMENTARY PROOF OF THE VANISHING OF THE SECOND
COHOMOLOGY OF THE WITT AND VIRASORO ALGEBRA WITH
VALUES IN THE ADJOINT MODULE
MARTIN SCHLICHENMAIER
Abstract. By elementary and direct calculations the vanishing of the (algebraic) second
Lie algebra cohomology of the Witt and the Virasoro algebra with values in the adjoint
module is shown. This yields infinitesimal and formal rigidity or these algebras. The
first (and up to now only) proof of this important result was given 1989 by Fialowski in
an unpublished note. It is based on cumbersome calculations. Compared to the original
proof the presented one is quite elegant and considerably simpler.
1. Introduction
The simplest nontrivial infinite dimensional Lie algebras are the Witt algebra and its
central extension the Virasoro algebra. The Witt algebra is related to the Lie algebra of
the group of diffeomorphisms of the unit circle. The central extension comes into play as
one is typically forced to consider projective actions if one wants to quantize a classical
system or wants to regularize a field theory. There is a huge amount of literature about
the application of these algebras. Here it is not the place to give even a modest overview
about these applications.
The Witt algebra W is the graded Lie algebra generated as vector space by the elements
{en n ∈ Z} with Lie structure
[en, em] = (m − n)en+m, n, m ∈ Z.
(1.1)
The Virasoro algebra V is the universal one-dimensional central extension. Detailed def-
initions and descriptions are given in Section 2. The goal of this article is to give an
elementary proof that the second Lie algebra cohomology of both algebras with values in
the adjoint modules will vanish. The result will be stated in Theorem 3.1 below. These
cohomology spaces are related to deformations of the algebras. In particular, if they vanish
the algebras will be infinitesimal and formally rigid, see Corollary 3.2. Note that this does
Date: 28.11.2011, rev. 27.1.2012, reference update after publication 8.5.2012.
1991 Mathematics Subject Classification. Primary: 17B56; Secondary: 17B68, 17B65, 17B66, 14D15,
81R10, 81T40.
Key words and phrases. Witt algebra; Virasoro algebra; Lie algebra cohomology; Deformations of alge-
bras; rigidity; conformal field theory.
This work was finalized during a visit to the Institut Mittag-Leffler (Djursholm, Sweden) in the frame
of the program "Complex Analysis and Integrable Systems", Fall 2011. Partial support by the ESF
networking programme HCAA, and the Internal Research Project GEOMQ08, University of Luxembourg,
is acknowledged.
1
2
MARTIN SCHLICHENMAIER
not mean that the algebras are analytically or geometrically rigid. Indeed, together with
Alice Fialowski we showed in [5] that there exist natural, geometrically defined, nontrivial
families of Lie algebras given by Krichever-Novikov type algebras associated to elliptic
curves. These families appeared in [13].
In these families the special fiber is the Witt
(resp. Virasoro) algebra but all other fibers are non-isomorphic to it. Similar results for
the current, resp. affine Lie algebras can be found in [6], [7].
The result on the vanishing of the second Lie algebra cohomology of the Witt algebra
should clearly be attributed to Fialowski. There exists an unpublished manuscript [1] by
her, dating from 1989, where she does explicit calculations 1. These calculations were
quite cumbersome not really appealing for journal publication.
In 1990 Fialowski gave
the statement of the rigidity of the Witt and Virasoro algebra in [3] without proof. Later,
in the above-mentioned joint paper of the author with her [5] a sketch of its proof was
presented. It is based on the calculations of the cohomology of the Lie algebra V ect(S1)
of vector fields on S1 with values in the adjoint module. Hence, this proof is not purely
algebraic. Based on results of Tsujishita [14], Reshetnikov [12], and Goncharova [10] we
showed that
H∗(V ect(S1), V ect(S1)) = {0}.
(1.2)
See also the book of Guieu and Roger [11]. We argued that by density arguments the
vanishing of the cohomology of the Witt algebra will follow. This is indeed true if one
considers continuous cohomology. But here we are dealing with arbitrary cohomology. In
a recent attempt to fill the details how to extend the arguments to this setting, we run
into troubles. We would have been forced to make considerable cohomological changes
to obtain finally certain boundedness properties to guarantee convergence with respect to
the topology coming from the Lie algebra of vector fields. Hence, there was an incentive
to return to a direct algebraic and elementary proof. Indeed, I found a very elementary,
computational, but nevertheless reasonable short and elegant proof which is simpler than
Fialowski's original calculations [1].
As the Witt and Virasoro algebra are of fundamental importance inside of mathematics
and in the applications, and up to now there is no complete published proof, the pre-
sented proof is for sure worth-while to publish. The proof avoids the heavy machinery of
Tsujishita, Reshetnikov, and Goncharova.2
Furthermore the presented proof has the advantage that by its algebraic nature it will
be valid for every field K of characteristic zero. Moreover, this is one of the rare occasions
where algebraic cohomology is calculated. From the very beginning it was desirable to
have a purely algebraic proof of the vanishing of the algebraic cohomology.
In Section 2 we give the definition of both the Witt and Viraosoro algebra and make
some remarks on their graded structure. The graded structure will play an important role
in the article.
In Section 3 after recalling the definition of general Lie algebra cohomology the 2-
cohomology of a Lie algebra with values in the adjoint module is considered in more
1Note added after publication of the current article: The content of the mentioned note of Fialowski is
in the meantime accessible on the archive and will appear in Jour. Math. Phys. See [1] for details.
2As Fuks writes in his book [8, p.119] "Theorem 2.3.1 [Goncharova's theorem] still remains one of the
most difficult theorems in the cohomology theory of infinite-dimensional Lie algebras."
VANISHING OF THE SECOND COHOMOLOGY OF THE VIRASORO ALGEBRA
3
detail. Some facts about its relation to deformations of the algebra are quoted to allow
to judge the importance of this cohomology [7]. The main result about the vanishing of
the second cohomology and the rigidity for the Witt and Virasoro algebra is formulated
in Theorem 3.1 and Corollary 3.2. The cohomology considered in this article is algebraic
cohomology without any restriction on the cocycles.
In Section 4 we use the graded structure of the algebras to decompose their cohomology
into graded pieces. It is quite easy then to show that for degree d 6= 0, the degree d parts
will vanish. Hence, everything is reduced to degree zero. This is due to the fact that the
grading is induced by the action of a special element. Here it is the element e0.
The vanishing of the degree zero part is more involved. It will be presented for the Witt
algebra in Section 5. It is the computational core of the article, but the computations are
elementary.
In Section 6 we extend this to the Virasoro algebra V. We show that the vanishing of
the cohomology for W implies the same for its central extension V.
As the core of the proof is elementary, I want to keep this spirit throughout the article.
Hence, it will be rather self-contained and elementary. We will not use things like long
exact cohomology sequences, etc. The proofs are purely algebraic and the vanishing and
rigidity theorems are true over every field K of characteristic zero.
2. The algebras
The Witt algebra W is the Lie algebra generated as vector space over a field K by the
elements {en n ∈ Z} with Lie structure
[en, em] = (m − n)en+m, n, m ∈ Z.
(2.1)
As in the cohomology computations we have to divide by arbitrary integers we assume for
the whole article that char(K) = 0.
Remark 2.1. In conformal field theory K will be C and the Witt algebra can be realized
as complexification of the Lie algebra of polynomial vector fields V ectpol(S1) on the circle
S1, which is a subalgebra of V ect(S1), die Lie algebra of all C ∞ vector fields on the circle.
In this realization en = exp(i n ϕ) d
dϕ . The Lie product is the usual bracket of vector fields.
An alternative realization is given as the algebra of meromorphic vector fields on the
Riemann sphere P1(C) which are holomorphic outside {0} and {∞}. In this realization
en = zn+1 d
dz .
A very important fact is that the Witt algebra is a Z-graded Lie algebra. We define
the degree by setting deg(en) := n, then the Lie product between elements of degree n
and of degree m is of degree n + m (if nonzero). The homogeneous spaces Wn of degree n
are one-dimensional with basis en. Crucial for the following is the additional fact that the
eigenspace decomposition of the element e0, acting via the adjoint action on W coincides
with the decomposition into homogeneous subspaces. This follows from
[e0, en] = n en = deg(en) en.
(2.2)
4
MARTIN SCHLICHENMAIER
Another property, which will play a role, is that [W, W] = W. That means that W is a
perfect Lie algebra. In fact,
en =
1
n
[eo, en], n 6= 0,
e0 =
1
2
[e−1, e1].
(2.3)
The Virasoro algebra V is the universal one-dimensional central extension of W. As
If we set for x ∈ W, x := (0, x), and
vector space it is the direct sum V = K ⊕ W.
t := (1, 0) then its basis elements are en, n ∈ Z and t with the Lie product
[en, em] = (m − n)en+m −
1
12
(n3 − n)δ−m
n
t,
[en, t] = [t, t] = 0,
(2.4)
for all n, m ∈ Z. Here δl
k is the Kronecker delta which is equal to 1 if k = l, otherwise
zero. If we set deg(en) := deg(en) = n and deg(t) := 0 then V becomes a graded algebra.
Let ν be the Lie homomorphism mapping the central element t to 0 and x to x inducing
the following short exact sequence of Lie algebras
0 −−−−→ K −−−−→ V
ν−−−−→ W −−−−→ 0 .
(2.5)
This sequence does not split, i.e. it is a non-trivial central extension.
The expression 1
n
In some abuse of notation we identify the element x ∈ V with x ∈ W and after identifi-
cation we have Vn = Wn for n 6= 0 and V0 = he0, tiK. Note that the relation (2.2) inducing
the eigenspace decomposition for the grading element e0 = e0 remains true.
12 (n3−n)δ−m
is the defining cocycle for the central extension. This form
is given in a standard normalisation -- others are possible. It is a Lie algebra two-cocycle
of W with values in the trivial module. The equivalence classes of central extensions are
in 1:1 correspondence to the cohomology classes H2(W, K), see the next section for more
details. It is well-known that dim H2(W, K) = 1, and that the class of the above cocycle
is a generator.
3. Cohomology and Deformations
Let us recall for completeness and further reference the definition of the Lie algebra
cohomology of a Lie algebra W with values in a Lie module M over W . We denote the
Lie module structure by W × M → M, (x, m) 7→ x . m. A k-cochain is an alternating
k-multilinear map W × W × · · · × W → M (k copies of W ). The vector space of k-cochains
is denoted by C k(W ; M ). Here we will deal exclusively with algebraic cohomology.
Next we have the family of coboundary operators
δk : C k(W ; M ) → C k+1(W ; M ),
k ∈ N, with δk+1 ◦ δk = 0.
(3.1)
Here we will only consider the second cohomology
δ2ψ(x, y, z) := ψ([x, y], z) + ψ([y, z], x) + ψ([z, x], y)
− x . ψ(y, z) + y . ψ(x, z) − z . ψ(x, y),
(3.2)
(δ1φ)(x, y) := φ([x, y]) − x . φ(y) + y . φ(x).
A k-cochain ψ is called a k-cocycle if it lies in the kernel of the k-coboundary operator δk.
It is called a k-coboundary if it lies in the image of the (k − 1)-coboundary operator.
By δk ◦ δk−1 = 0 the vector space quotient of cocycles modulo coboundaries is well-
defined. It is called the vector space of k-Lie algebra cohomology of W with values in the
VANISHING OF THE SECOND COHOMOLOGY OF THE VIRASORO ALGEBRA
5
module M . It is denoted by Hk(W ; M ). Two cocycles which are in the same cohomology
class are called cohomologous.
The trivial module is K with the Lie action x . m = 0, for all x ∈ W and m ∈ K. The
second cohomology with values in the trivial module classifies equivalence classes of central
extensions of W . It is well-known that for the Witt algebra W we have dim H2(W; K) = 1
and that the class of the cocycle defining V gives a basis.
The second cohomology of W with values in the adjoint module, H2(W ; W ), i.e. with
module structure x . y := [x, y], is the cohomology to be studied here. As we will need the
formula for explicit calculations later let me specialize the 2-cocycle condition (3.2)
δ2ψ(x, y, z) := 0 = ψ([x, y], z) + ψ([y, z], x) + ψ([z, x], y)
− [x, ψ(y, z)] + [y, ψ(x, z)] − [z, ψ(x, y)].
(3.3)
A 2-cocycle will be a coboundary if it lies in the image of the (1-)coboundary operator,
i.e. there exists a linear map φ : W → W such that
ψ(x, y) = (δ1φ)(x, y) := φ([x, y]) − [φ(x), y] − [x, φ(y)].
(3.4)
The second cohomology H2(W, W ) is related to the deformations of the Lie algebra W .
(1) H2(W, W ) classifies infinitesimal deformations of W up to equivalence (Gersten-
haber [9]).
(2) If dim H2(W, W ) < ∞ then there exists a versal formal family for the formal
deformations of W whose base is formally embedded into H2(W, W ). This is due
to Fialowski [2], and Fialowski and Fuks [4].
A Lie algebra W is called rigid if every deformation is locally equivalent to the trivial
family. Hence, if H2(W, W ) = 0 then W is infinitesimally and formally rigid. See [7] for
more information on the connection between cohomology and deformations.
After having recalled the general definitions, I formulate the main result of this article.
Theorem 3.1. Both the second cohomology of the Witt algebra W and of the Virasoro
algebra V (over a field K with char(K) = 0) with values in the adjoint module vanishes,
i.e.
H2(W; W) = {0},
H2(V; V) = {0}.
(3.5)
Corollary 3.2. Both W and V are formally and infinitesimally rigid.
I refer to Section 1, the Introduction, for the history of this theorem. Here I only like to
repeat that the first proof (at least in the Witt case) was given by Alice Fialowski by very
cumbersome unpublished calculations [1] 3. The proof which I present in the following is
elementary, but much more accessible and elegant as the original proof.
3Note added after publication of the current article: In the meantime they are published, see [1] for
details
6
MARTIN SCHLICHENMAIER
4. The degree decomposition of the cohomology
Let W be an arbitrary Z-graded Lie algebra, i.e.
W = Mn∈Z
Wn.
(4.1)
Recall that we consider algebraic cohomology, i.e. our 2-cochains ψ ∈ C 2(W ; W ) are
arbitrary alternating bilinear maps in the usual sense, i.e.
for all v, w ∈ W the cochain
ψ(v, w) will be a finite linear combination (depending on v and w) of basis elements in
W . We call a k-cochain ψ homogeneous of degree d if there exists a d ∈ Z such that for
all i1, i2, . . . , ik ∈ Z and homogeneous elements xil ∈ W , of deg(xil) = il, for l = 1, . . . , k
we have that
ψ(xi1, xi2, . . . , xik ) ∈ Wn, with n =
k
Xl=1
il + d.
(4.2)
The corresponding subspace of degree d homogeneous k-cochains is denoted by C k
Every k-cochain can be written as a formal infinite sum
(d)(W ; W ).
ψ = Xd∈Z
ψ(d).
(4.3)
Note that evaluated for a fixed k-tuple of elements only a finite number of the summands
will produce values different from zero.
An inspection of the coboundary operators (3.3) and (3.4) for homogeneous elements
x, y, z shows
Proposition 4.1. The coboundary operators δk are operators of degree zero, i.e. applied
to a k-cocycle of degree d they will produce a (k + 1)-cocycle also of degree d.
In the following we will concentrate on k = 2 or k = 1. If ψ = Pd ψ(d) is a 2-cocycle
then δ2ψ = Pd δ2ψ(d) = 0. By Proposition 4.1 δ2ψ(d) is either zero or of degree d. As we
sum over different degrees and the terms cannot cancel if different from zero we obtain
that ψ is cocycle if and only if all degree d components ψ(d) will be individually 2-cocycles.
Moreover, if ψ(d) is 2-coboundary, i.e. ψ(d) = δ1φ with a 1-cochain φ, then we can find
another 1-cochain φ′ of degree d such that ψ(d) = δ1φ′.
We summarize as follows. Every cohomology class α ∈ H2(W ; W ) can be decomposed
as formal sum
α = Xd∈Z
α(d),
α(d) ∈ H2
(d)(W ; W ),
(4.4)
where the latter space consists of classes of cocycles of degree d modulo coboundaries of
degree d.
For the rest of this section let W be either W or V and d 6= 0. We will show that for
these algebras the cohomology spaces of degree d 6= 0 will vanish. The degree zero case
needs a more involved treatment and will be done in the next sections. We start with a
cocycle of degree d 6= 0 and make first a cohomological change ψ′ = ψ − δ1φ with
φ : W → W,
x 7→ φ(x) =
ψ(x, e0)
d
.
(4.5)
VANISHING OF THE SECOND COHOMOLOGY OF THE VIRASORO ALGEBRA
7
Recall e0 is the element of either W or V which gives the degree decomposition. This
implies (note that by definition φ(e0) = 0)
ψ′(x, e0) = ψ(x, e0) − (δ1φ)(x, e0) = ψ(x, e0) − φ([x, e0]) + [φ(x), e0]
= d φ(x) + deg(x)φ(x) − (deg(x) + d)φ(x) = 0.
(4.6)
We evaluate (3.3) for the cocycle ψ′ on the triple (x, y, e0) and leave out the cocycle values
which vanish due to (4.6):
0 = ψ′([y, e0], x) + ψ′([e0, x], y) − [e0, ψ′(x, y)]
= (deg(y) + deg(x) − (deg(x) + deg(y) + d))ψ′(x, y) = −dψ′(x, y).
(4.7)
As d 6= 0 we obtain ψ′(x, y) = 0 for all x, y ∈ W . We conclude
Proposition 4.2. The following hold:
(a)
(b)
H2
(d)(W; W) = H2
H2(W; W) = H2
(d)(V; V) = {0},
(0)(W; W),
for d 6= 0.
H2(V; V) = H2
(0)(V; V).
Remark 4.3. In fact the arguments also work for every Z-graded Lie algebra W for which
there exists an element e0 such that the homogeneous spaces Wn are just the eigenspaces
of e0 under the adjoint action to the eigenvalue n. Such Lie algebras are called internally
graded. See also Theorem 1.5.2 in [8].
5. Degree zero for the Witt algebra
It remains the degree zero part.
In this section we consider only the Witt algebra.
Recall that the homogeneous subspaces of degree n are one-dimensional and generated
by en. Hence, a degree zero cocycle can be written as ψ(ei, ej) = ψi,jei+j and if it
is a coboundary then it can be given as a coboundary of a linear form of degree zero:
φ(ei) = φiei. The systems of ψi,j and φi for i, j ∈ Z fix ψ and φ completely. If we evaluate
(3.3) for the triple (ei, ej, ek) we get for the coefficients
0 =(j − i)ψi+j,k − (k − i)ψi+k,j + (k − j)ψj+k,i
− (j + k − i)ψj,k + (i + k − j)ψi,k − (i + j − k)ψi,j.
For the coboundary we obtain
(δφ)i,j = (j − i)(φi+j − φj − φi).
(5.1)
(5.2)
Hence, ψ is a coboundary if and only if there exists a system of φk ∈ K, k ∈ Z such that
ψi,j = (j − i)(φi+j − φj − φi),
∀i, j ∈ Z.
(5.3)
A degree zero 1-cochain φ will be a 1-cocycle (i.e. δ1φ = 0) if and only if
(5.4)
This has the solution φi = i φ1, ∀i ∈ Z. Hence, given a φ we can always find a φ′ with
(φ′)1 = 0 and δ1φ = δ1φ′.
In the following we will always choose such a φ′ for our
2-coboundaries.
φi+j − φj − φi = 0.
Step 1: We make a cohomological change
8
MARTIN SCHLICHENMAIER
We start with a 2-cocycle ψ given by the system of ψi,j and will modify it by adding a
coboundary δ1φ with suitable φ to obtain ψ′ = ψ − δ1φ. We will determine φ inspired by
the intended relation (5.3)
ψi,1 = (1 − i)(φi+1 − φ1 − φi) = (1 − i)(φi+1 − φi).
(5.5)
Note that we could put φ1 = 0 by our normalization.
(a) Starting from φ0 := −ψ0,1 we set in descending order for i ≤ −1
1 − i
(b) φ2 cannot be fixed by (5.5), instead we use (5.3)
φi := φi+1 −
ψi,1.
1
ψ−1,2 = 3(−φ2 − φ−1),
yielding φ2 := −φ−1 −
1
3
ψ−1,2.
Then we have ψ′
(c) We use again (5.5) to calculate recursively in ascending order φi, i ≥ 3 by
−1,2 = 0.
φi+1 := φi +
1
1 − i
ψi,1.
For the cohomologous cocycle ψ′ we obtain by construction
−1,2 = ψ′
and ψ′
∀i ∈ Z,
ψ′
i,1 = 0,
2,−1 = 0.
(5.6)
(5.7)
(5.8)
(5.9)
Step2: Show that ψ′ is identical zero.
To avoid cumbersome notation we will denote the cohomologous cocycle by ψ.
Lemma 5.1. Let ψ be a 2-cocycle of degree zero such that ψi,1 = 0, ∀i ∈ Z and ψ−1,2 = 0,
then ψ will be identical zero.
Before we proof the lemma we use it to show
Proof. (Witt part of Theorem 3.1.) By Proposition 4.2 it is enough to consider degree
zero cohomology. By the cohomological change done in Step 1 every degree zero cocycle
ψ is cohomologous to a cocycle fulfilling the conditions of Lemma 5.1. But such a cocycle
vanishes by the lemma. Hence, the original cocycle ψ is cohomological trivial.
(cid:3)
Proof. (Lemma 5.1.) First, we note two special cases of (5.1) which will be useful in the
following. For the index triple (i, −1, k) we obtain
0 = − (i + 1)ψi−1,k − (k − i)ψi+k,−1 + (k + 1)ψk−1,i
− (−1 + k − i)ψ−1,k + (i + k + 1)ψi,k − (i − 1 − k)ψi,−1.
(5.10)
For the triple (i, 1, k), and ignoring terms of the type ψi,1 which are zero by assumption,
we obtain
0 = (1 − i)ψi+1,k + (k − 1)ψk+1,i + (i + k − 1)ψi,k.
(5.11)
We will consider ψi,m for certain values of m ≤ 2 and finally make ascending and de-
scending induction on m. We will call the coefficient ψi,m coefficients of level m (and of
level i by antisymmetry). By assumption the cocycle values of level 1 are all zero.
m = 0
VANISHING OF THE SECOND COHOMOLOGY OF THE VIRASORO ALGEBRA
9
By the antisymmetry we have ψ1,i = −ψi,1 = 0. In (5.11) we consider k = 0 this gives
0 = (1 − i) (ψi+1,0 − ψi,0) .
(5.12)
Starting from ψ1,0 = 0 this implies for i ≤ 0 that ψi,0 = 0 and for i ≥ 3 that ψi,0 = ψ2,0.
Next we consider (5.10) for k = 2, i = 0 and obtain
0 = −ψ−1,2 − 2ψ2,−1 + 3ψ1,0 − ψ−1,2 + 3ψ0,2 + 3ψ0,−1.
(5.13)
The ψ−1,2 terms cancel and we know already ψ1,0 = ψ−1,0 = 0, hence ψ2,0 = 0. This
implies
m = −1
In (5.11) we set k = −1 and obtain (with ψ0,i = 0)
ψi,0 = 0 ∀i ∈ Z.
− (i − 1) ψi+1,−1 + (i − 2) ψi,−1 = 0.
(5.14)
(5.15)
Hence,
ψi,−1 =
i − 1
i − 2
ψi+1,−1,
for i 6= 2,
ψi+1,−1 =
i − 2
i − 1
ψi,−1,
for i 6= 1.
(5.16)
The first formula implies starting from ψ1,−1 = −ψ−1,1 = 0 that ψi,−1 = 0, for all i ≤ 1.
The second formula for i = 2 implies ψ3,−1 = 0 and hence ψi,−1 = 0 for i ≥ 3. But by
assumption ψ2,−1 = ψ−1,2 = 0. Hence,
ψi,−1 = 0 ∀i ∈ Z.
(5.17)
m = −2
We plug the value k = −2 into (5.11) and get for the terms not yet identified as zero
(1 − i)ψi+1,−2 + (i − 3)ψi,−2 = 0.
(5.18)
This yields
ψi+1,−2 =
i − 3
i − 1
ψi,−2,
for i 6= 1,
ψi,−2 =
i − 1
i − 3
ψi+1,−2,
for i 6= 3.
(5.19)
From the first formula we get ψ3,−2 = −ψ2,−2, ψ4,−2 = 0 · ψ3,−2, and hence ψi,−2 = 0, for
all i ≥ 4.
From the second formula we get starting from ψ1,−2 = 0 that ψi,−2 = 0 for i ≤ 1.
Altogether, ψi,−2 = 0 for i 6= 2, 3. The value of ψ2,−2 = −ψ3,−2 stays undetermined for
the moment.
m = 2
We start from (5.10) for k = 2 and recall that terms of levels 0, 1, -1 are zero. This gives
− (i + 1) ψi−1,2 + (i + 3) ψi,2 = 0.
(5.20)
Hence,
ψi,2 =
i + 1
i + 3
ψi−1,2,
for i 6= −3,
ψi−1,2 =
i + 3
i + 1
ψi,2,
for i 6= −1.
(5.21)
From the first formula we start from ψ−1,2 = 0 and get ψi,2 = 0, ∀i ≥ −1. From the second
we get ψ−3,2 = −ψ−2,2, then ψ−4,2 = 0 and then altogether ψi,2 = 0 for all i 6= −2, −3.
The value ψ−3,2 = −ψ−2,2 stays undetermined for the moment.
10
MARTIN SCHLICHENMAIER
To find it we consider the index triple (2, −2, 4) in (5.1) and obtain after leaving out
terms which are obviously zero
0 = −2 ψ6,−2 − 8 ψ4,2 + 4 ψ2,−2.
(5.22)
From the level m = 2 discussion we get ψ4,2 = 0, from m = −2 we get ψ6,−2 = 0. This
shows that ψ2,−2 = ψ3,−2 = ψ−3,2 = 0 and we can conclude
ψi,−2 = ψi,2 = 0,
∀i ∈ Z.
(5.23)
m < −2
We make induction assuming it is true for m = 2, 1, 0, −1, −2. We start from (5.10) for k
and put the k − 1 level element on the l.h.s. By this
(k + 1)ψi,k−1 = terms of level k and −1.
(5.24)
By induction the terms on the r.h.s. are zero. Note that in this region k < −1, hence
k + 1 6= 0, and ψi,k−1 = 0, too.
m > 2
Again we make induction. Starting from (5.11) we get
(k − 1) ψi,k+1 = (1 − i) ψi+1,k + (i + k − 1) ψi,k.
(5.25)
As k ≥ 2 the value of 1 − k 6= 0 and we get by induction trivially the statement for k + 1.
(cid:3)
Altogether we obtain ψi,k = 0, ∀i, k ∈ Z.
The presented calculation is completely different and much simpler as in [1]
6. Extension to the Virasoro algebra
Here we show in an elementary way that also for the Virasoro algebra
H2(V, V) = 0.
(6.1)
Hence V is also infinitesimally and formally rigid. This shows the Virasoro part of Theo-
rem 3.1. It is our intention not to use any higher techniques, but see Remark 6.4 at the
end of this section.
First recall that by Proposition 4.2 it is enough to consider degree zero cocycles. We
start with a degree zero 2-cocycle ψ : V × V → V of the Virasoro algebra. If we apply the
Lie homomorphism ν we get the bilinear map ν ◦ ψ which we restrict to W × W
ψ′ = ν ◦ ψ : W × W → W.
(6.2)
Unfortunately, in general ψ′ will not be a 2-cocycle for the Witt algebra. We have to be
careful as the Lie product for V differs from that of W by multiples of the central element.
But we are allowed to make cohomologous changes.
Proposition 6.1. Given a cocycle ψ ∈ C 2(V, V) there exists a cohomologous one ψ ∈
C 2(V, V) such that the bilinear map ψ′ = ν ◦ ψ ∈ C 2(W, W).
VANISHING OF THE SECOND COHOMOLOGY OF THE VIRASORO ALGEBRA
11
Proof. Let x, y, z ∈ W. We have [x, y]V = [x, y]W + α(x, y) · t. We consider (3.3) with the
bracket [., .]V and rewrite it in terms of [., .]W (but drop the index) and α. For the second
group of terms we use that ν is a Lie homomorphism and they will not see the central
elements. Only in the first group they will play a role. We get
0 = ψ′([x, y], z) + ψ′([y, z], x) + ψ′([z, x], y)
+ α([x, y], z) ν ◦ ψ(t, z) + α([y, z], x) ν ◦ ψ(t, x) + α([z, x], y) ν ◦ ψ(t, y)
− [x, ψ′(y, z)] + [y, ψ′(x, z)] − [z, ψ′(x, y)].
(6.3)
We will make a cohomological change for the cocycle ψ in V such that ψ(t, z) will have
only central terms, i.e. ν ◦ ψ(t, z) will vanish. Restricting it to W × W and projecting it
by ν will define a 2-cocycle for W.
As ψ is a degree zero cocycle, we have ψ(e1, t) = a e1. We set φ(t) := a e0, and φ(en) := 0
for all n ∈ Z. Let ψ := ψ − δ1φ. We calculate
ψ(e1, t) = ψ(e1, t) − (δ1φ)(e1, t) = a e1 − φ([e1, t]) + [e1, φ(t)] + [φ(e1), t]
= a e1 − 0 − a e1 + 0 = 0.
The coefficients an (and b, but it will not play any role) are given by
ψ(e0, t) = a0en + b · t.
ψ(en, t) = anen, n 6= 0,
(6.4)
(6.5)
Note that a1 = 0. If we evaluate (3.3) for the triple (en, em, t) and leave out terms which
are zero due to the fact, that t is central we get
0 = ψ([en, em]V , t) − [en, ψ(em, t)]V + [em, ψ(t, en)]V .
This implies (by evaluating the coefficients at the element em+n)
(m − n)(an+m − am − an) = 0.
(6.6)
(6.7)
If we plug in m = 1 and use a1 = 0 we get an = 0 for all n ≤ 1 and an = a2 for all n ≥ 2.
Now plugging in m = 2, n = −2 we obtain a2 = −a−2 = 0. Hence altogether an = 0 for
all n, and ν ◦ ψ(w, t) = 0, ∀w ∈ V.
(cid:3)
Hence, after a cohomologous change we may assume that our restricted and projected
ψ′ will be a 2-cocycle for the algebra W with values in the adjoint module W. In the last
section we showed that it is cohomologically trivial, i.e. there exists a φ′ : W → W such
that δW
1 φ′ = ψ′ = ν ◦ ψ. We denote for a moment the corresponding coboundary operators
of the two Lie algebras by δW and δV . By setting φ(t) := 0 we extend φ′ to a linear map
φ : V → V. In particular, ν ◦ φ = φ′ if restricted to W.
The 2-cocycle ψ = ψ − δV
1 φ will be a cohomologous cocycle for V. If we apply ν then
ν ψ = νψ − νδV
1 φ = ψ′ − δW
1 φ′ = 0.
(6.8)
Hence, ψ takes values in the kernel of ν, i.e.
ψ : V × V → K · t,
(6.9)
with t the central element.
12
MARTIN SCHLICHENMAIER
This implies that it is enough to show that every class ψ : V → V with values in the
central ideal K · t will be a coboundary to show the vanishing of H2(V, V). In the following
let ψ be of this kind.
As ψ(x, y) = ψx,yt will be central the 2-cocycle condition (3.3) will reduce to
(δ2ψ)(x, y, z) = ψ([x, y], z) + ψ([y, z], x) + ψ([z, x], y) = 0.
(6.10)
The coboundary condition for φ : V → K · t reduces to
We can reformulate this as that the component function ψx,y : V × V → K is a Lie algebra
2-cocycle for V with values in the trivial module.
(δ1φ)(x, y) = φ([x, y]).
(6.11)
Lemma 6.2. We have
ψ(x, t) = 0,
∀x ∈ V.
(6.12)
Proof. We evaluate (6.10) for (ei, ej, t). As t is central [ej, t] = [t, ei] = 0, hence only the
first term in (6.10) will survive and we get (j − i)ψ(ei+j , t) = 0. Choosing j = 0 we obtain
ψ(ei, t) = 0 for i 6= 0, choosing i = −1, j = +1 we obtain also ψ(e0, t) = 0. As ψ(t, t) = 0
is automatic and {en, n ∈ Z, t} is a basis of V this shows the result.
(cid:3)
Next let φ : V → K·t be a linear map. We write for the component functions φ(ei) = φit
and φ(t) = c · t. If we evaluate the 1-coboundary operator (6.11) for pairs of basis elements
we get
δ1φ(ei, ej ) = φ([ei, ej ]) = (cid:18)(j − i)φi+j −
1
12
(i3 − i)δ−j
i c(cid:19) t,
i, j ∈ Z,
(6.13)
and by Lemma 6.2 we get δ1φ(ei, t) = 0, for all i ∈ Z.
With respect to (6.13) we choose φ such that the cohomologous cocycle ψ′ = ψ − δ1φ
fulfills
ψ′(ei, e0) = 0, ∀i ∈ Z,
and ψ′(e1, e−1) = ψ′(e2, e−2) = 0.
(6.14)
This is obtained by putting
φi := −
1
i
ψei,e0, i ∈ Z, i 6= 0, φ0 := −(1/2)ψe1,e−1,
c := (−2 ψe2,e−2 − 8φ0).
(6.15)
Lemma 6.3. The cocycle ψ′ is identically zero.
Proof. First we consider the 2-cocycle condition (6.10) for the triple (ei, ej, e0) and obtain
using ψ′(ei, t) = 0 that
0 = 0 + (−j)ψ′(ej, ei) + i ψ′(ei, ej) = (i + j)ψ′(ei, ej).
(6.16)
This yields that
(6.17)
Next we show by induction that ψ′(ei, e−i) = 0 for all i ≥ 0 (and by antisymmetry also
for i ≤ 0). Note that this is true for i = 0, 1 and 2, see (6.14). Consider (6.10) for the
triple (en, e−(n−1), e−1), n > 2. After evaluation of the Lie products we get
for j 6= −i.
ψ′(ei, ej ) = 0
0 = (−2n + 1)ψ′(e1, e−1) + (n − 2)ψ′(e−n, en) + (−n + 1)ψ′(en−1, e−(n−1)).
(6.18)
VANISHING OF THE SECOND COHOMOLOGY OF THE VIRASORO ALGEBRA
13
The first and last term vanishes by induction. As n > 2 this implies
As we have always ψ′(x, t) = 0 this shows the lemma.
ψ′(en, e−n) = 0.
(6.19)
(cid:3)
Finally we obtain that the ψ we started with was a coboundary. This implies indeed
H2(V; V) = {0}.
Remark 6.4. What has been done in this section can also be interpreted in the framework
of long exact cohomology sequences in Lie algebra cohomology. The exact sequence (2.5)
of Lie algebras is also a exact sequence of Lie modules over V. For such sequences we have
a long exact sequence in cohomology. If we consider only level two we get
· · · −−−−→ H2(V; K) −−−−→ H2(V; V)
ν∗−−−−→ H2(V; W) −−−−→ · · · .
(6.20)
Proposition 6.1 also shows that naturally H2(V; W) ∼= H2(W; W) (which is not a conse-
quence from the existence of the long exact sequence). The calculations of Lemma 6.2 and
Lemma 6.3 showed H2(V; K) = {0}. From Section 5 we know H2(W; W) = {0}. Hence,
also H2(V; V) = {0}.
References
[1] Fialowski, A.: Unpublished Note, 1989, in the meantime published as arXiv:1202.3132, to appear in
Jour. Math. Phys. 2012
[2] Fialowski, A.: An example of formal deformations of Lie algebras. In: Proceedings of NATO Con-
ference on Deformation Theory of Algebras and Applications, Il Ciocco, Italy, 1986, pp. 375 -- 401,
Kluwer, Dordrecht, 1988.
[3] Fialowski, A.: Deformations of some infinite-dimensional Lie algebras. J. Math. Phys. 31(6) 1340 -- 1343
(1990)
[4] Fialowski, A., and Fuks, D.: Construction of miniversal deformations of Lie algebras. J. Funct. Anal.
161, 76-110 (1999).
[5] Fialowski, A., and Schlichenmaier, M.: Global deformations of the Witt algebra of Krichever Novikov
type. Comm. Contemp. Math. 5, 921 -- 945 (2003).
[6] Fialowski, A., and Schlichenmaier, M.: Global Geometric Deformations of Current Algebras as
Krichever -- Novikov Type Algebras. Comm. Math. Phys. (2005), DOI 10.1007/s00220-005-1423-5.
[7] Fialowski, A., and Schlichenmaier, M.: Global Geometric Deformations of the Virasoro algebra, cur-
rent and affine algebras by Krichever-Novikov type algebras, International Journal of Theoretical
Physics. Vol. 46, No. 11 (2007) pp.2708 - 2724
[8] Fuks, D.: Cohomology of Infinite-dimensional Lie Algebras, Consultants Bureau, N.Y., London, 1986.
[9] Gerstenhaber, M.: On the deformation of rings and algebras I,II,III Ann. Math. 79, 59-10 (1964),
84, 1-19 (1966), 88, 1-34 (1968).
[10] Goncharova, I.V.: Cohomology of Lie algebras of formal vector fields on the line. Funct. Anal. Appl.
7, No.2, 6-14 (1973)
[11] Guieu, L., Roger, C.: L'alg`ebre et le groupe de Virasoro. Les publications CRM, Montreal 2007.
[12] Reshetnikov, V.N.: On the cohomology of the Lie algebra of vector fields on a circle. Usp. Mat. Nauk
26, 231-232 (1971)
[13] Schlichenmaier, M.: Degenerations of generalized Krichever-Novikov algebras on tori, Jour. Math.
Phys. 34, 3809-3824 (1993).
[14] Tsujishita, T.: On the continuous cohomology of the Lie algebra of vector fields. Proc. Japan Acad.
53, Sec. A, 134-138 (1977)
14
MARTIN SCHLICHENMAIER
University of Luxembourg, Mathematics Research Unit, FSTC, Campus Kirchberg, 6, rue
Coudenhove-Kalergi, L-1359 Luxembourg-Kirchberg, Luxembourg
E-mail address: [email protected]
|
1711.11322 | 2 | 1711 | 2019-02-12T08:39:21 | On the Prime Graph Question for Integral Group Rings of Conway simple groups | [
"math.RA",
"math.GR",
"math.RT"
] | The Prime Graph Question for integral group rings asks if it is true that if the normalized unit group of the integral group ring of a finite group $G$ contains an element of order $pq$, for some primes $p$ and $q$, also $G$ contains an element of that order. We answer this question for the three Conway sporadic simple groups after reducing it to a combinatorial question about Young tableaus and Littlewood-Richardson coefficients. This finishes work of V. Bovdi, A. Konovalov and S. Linton. | math.RA | math |
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF
CONWAY SIMPLE GROUPS
LEO MARGOLIS
Abstract. The Prime Graph Question for integral group rings asks if it is true that if the
normalized unit group of the integral group ring of a finite group G contains an element of order
pq, for some primes p and q, also G contains an element of that order. We answer this question
for the three Conway sporadic simple groups after reducing it to a combinatorial question about
Young tableaux and Littlewood-Richardson coefficients. This finishes work of V. Bovdi, A.
Konovalov and S. Linton.
1. Introduction
The unit group of the integral group ring ZG of a finite group G has given rise to many inter-
esting research questions and results, as recorded e.g. in the monographs [Seh93, JdR16]. Many
of these questions concern the connection between finite subgroups of units of ZG and the group
base G, e.g. the Isomorphism Problem which was open for over 50 years until answered by Her-
tweck [Her01] or the recently answered Zassenhaus Conjecture [EM18] which was open for over
40 years. For both questions there turned out to be counterexamples. In this light the weaker
versions of the former Zassenhaus Conjecture, which are still open, become more important. One
of these versions, the so-called Prime Graph Question, gained attention after being introduced by
Kimmerle in [Kim06]. To formulate it denote by V(ZG) the group of the so-called normalized
units in ZG, i.e. the units whose coefficients sum up to 1.
Prime Graph Question: Let G be a finite group and p and q some primes. If V(ZG) contains
an element of order pq, does G contain an element of order pq?
The prime graph Γ(G) of a group G is a graph whose vertices are the primes appearing as orders
of elements in G and two vertices p and q are connected if and only if G contains an element of
order pq. Hence the Prime Graph Question can also be formulated as: For G a finite group, does
Γ(G) = Γ(V(ZG)) hold?
The Prime Graph Question seems particularly approachable, since, in contrast to other ques-
tions on the finite subgroups in V(ZG), a reduction result is available here: The Prime Graph
Question holds for a group G, if it holds for all almost simple images of G [KK17, Theorem
2.1]. Recall that a group A is called almost simple if it is isomorphic to a subgroup of the auto-
morphism group of a non-abelian simple group S containing the inner automorphisms of S, i.e.
S ∼= Inn(S) ≤ A ≤ Aut(S). In this case S is called the socle of A.
It is known that the Prime Graph Question has a positive answer for almost simple groups
with socle isomorphic to a projective special linear group PSL(2, p) or PSL(2, p2) for p a prime
[BM17b, Theorem A] or an alternating group of degree up to 17 [BC17]. The question also has
a positive answer for groups whose order is divisible by exactly three pairwise different primes
[KK17, BM17c] and many almost simple groups whose order is divisible by four pairwise different
primes [BM18b].
This research was partially supported by the European Commission individual Marie Curie-Skłodowska Fellow-
ship in the H2020 program under Grant 705112-ZC and the FWO (Research Foundation Flanders).
2010 Mathematics Subject Classification. Primary 16U60, 16S34. Secondary 20C05, 20C20, 05E10.
Key words and phrases. Unit Group, Group Ring, Prime Graph Question, Sporadic simple groups.
1
2
LEO MARGOLIS
Employing a computer implementation of a method introduced by Luthar and Passi [LP89] and
Hertweck [Her07], nowadays known as the HeLP method, in a series of papers between 2007 and
2012 Bovdi, Konovalov and different collaborators of them studied the Prime Graph Question for
sporadic simple groups [BKS07, BK07a, BK07b, BKL08, BKM08, BK08, BGK09, BK09, BK10,
BJK11, BKL11, BK12]. Overall they studied 17 sporadic simple groups and were able to prove
the Prime Graph Question for 13 of these groups, an overview of their results can be found in
[KK15, Section 5]. Also in [KK15, Corollary 5.4] it was recorded that the Prime Graph Question
holds for almost simple groups with a socle isomorphic to one of the 13 sporadic simple groups
for which Bovdi, Konovalov et al. proved the Prime Graph Question. The four groups studied by
Bovdi, Konovalov et al. for which they were not able to prove the Prime Graph Question were the
O'Nan simple group [BGK09] and the three Conway simple groups [BKL11]. Developing further
the combinatorial side of a method introduced in [BM17c] in this note we obtain a positive answer
for the Prime Graph Question for the latter groups. This is the first contribution to the Prime
Graph Question for sporadic simple groups since the papers of Bovdi, Konovalov et al.
Theorem 1.1. The Prime Graph Question has a positive answer for the sporadic Conway simple
groups Co3, Co2 and Co1.
Note that the outer automorphism group of each Conway simple group is trivial, so Theorem 1.1
proves the Prime Graph Question also for every almost simple group whose socle is isomorphic to
a Conway simple group.
To describe our main tool let us first introduce some notation. Let G be a finite group, u ∈
V(ZG) a unit of order n and χ an ordinary character of G with corresponding representation D.
Then we can extend D linearly to ZG and afterwards restrict it to V(ZG), obtaining an extension
of D and χ. In particular D(u) is then a matrix of finite order dividing n. Let ξ be an n-th root
of unity. Then we denote by µ(ξ, u, χ) the multiplicity of ξ as an eigenvalue of D(u). Moreover
for an integer k we will denote by ζk a primitive k-th root of unity. Then the results given above
will follow from an application of the following theorem.
Theorem 1.2. Let G be a finite group, p an odd prime and m a positive integer not divisible by
p. Assume that some p-block of G is a Brauer Tree Algebra with ordinary characters χ1, ...,χp
and with the Brauer tree being a line of the form
χ1
χ2
χ3
χp−2
χp−1
χp
.
Moreover assume that in the rings of values of χ1, ...,χp the prime p is unramified.
Let u ∈ V(ZG) be a unit of order pm and let ξ be some m-th root of unity. Then the inequality
µ(ξ · ζp, u, χp−1) − µ(ξ · ζp, u, χp) ≤ µ(ξ, u, χ1) −
p−2
X
(−1)iµ(ξ · ζp, u, χi)
holds.
i=1
This theorem might be regarded as a generalization of [BM18b, Proposition 3.2]. The condition
on the unit can also be formulated in terms of character values which makes the inequality easier
to check by hand, although it becomes much longer then, cf. Lemma 2.3.
Theorem 1.2 will be achieved using the so-called lattice method introduced in [BM17c] and
developed further in [BM18b]. It is typical for the study of group rings that different fields of
mathematics, such as group theory, ring theory, representation theory and number theory are
combined to achieve results. In the present paper the main tool is combinatorics, more precisely
calculations with Young tableaus and Littlewood-Richardson coefficients, cf. the proof of Theo-
rem 1.2. It is quite plausible that Theorem 1.2, which is the main tool to prove Theorem 1.1, can
be applied also to study the Prime Graph Question or related questions for many other groups.
Also the methods presented here could be used to prove variations of Theorem 1.2, in particular
for other forms of Brauer trees.
We will also need to use the HeLP method to obtain enough restrictions on torsion units which
allow the application of Theorem 1.2. We will start by recalling the needed methods in Section 2,
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
3
develop the necessary combinatorics in Section 3, continue to prove Theorem 1.2 and finally show
how this can be applied to prove Theorem 1.1 in Section 4.
2. Preliminaries and methods
A fundamental notion when studying torsion units of integral group rings are so-called partial
augmentations. Let u ∈ ZG be an element of the form Pg∈G zgg and denote by xG the conjugacy
class of an element x in G. Then εxG(u) = Pg∈xG zg is called the partial augmentation of u at
x. A fundamental theorem of Hertweck [Her07, Theorem 2.3] states that if u ∈ V(ZG) is a unit
of order n then εxG(u) 6= 0 implies that the order of x divides n. This implies in particular that
the exponents of G and V(ZG) coincide, a result due originally to Cohn and Livingstone [CL65].
Moreover ε1(u) = 0 unless u = 1 by the Berman-Higman Theorem [JdR16, Proposition 1.5.1].
2.1. The HeLP method. Partial augmentations are class functions and as such can be investi-
gated using representation theory. Note that if χ is an ordinary character of G and u ∈ V(ZG)
a torsion unit of order n then χ(u) = PxG εxG(u)χ(g), where the sum runs over the conjugacy
classes of G. It was shown by Hertweck [Her07, Theorem 3.2] that this also holds for p-Brauer
characters, if p is not a divisor of n and the sum is understood to run on p-regular conjugacy
classes of G. On the other hand if we are given the partial augmentations of u and its powers
we can compute the eigenvalues, including multiplicities, of u under any ordinary or p-modular
representation of G, again assuming p is not dividing n. These multiplicities can be expressed in
explicit formulas, depending on the partial augmentations of u and its powers, and that the results
of these formulas are non-negative integers is the basic idea of the HeLP method, cf.
[BM17b,
Section 2] for a detailed explanation and [Her07] for the proofs.
Hence there is an algorithmic method which allows to obtain restrictions on the possible partial
augmentations of torsion units in V(ZG), provided we have some knowledge on the characters of
G. If we apply the HeLP method to the whole character table of G then, for any fixed n, we will
obtain a finite number of possibilities for the partial augmentations of units of order n in V(ZG).
If it happens that we obtain no possibility for units of some given order n, then we know that
there exist no units at all in V(ZG) of order n. This is the basic idea in the application of the
HeLP method for the study of the Prime Graph Question and in this way Bovdi, Konovalov et
al. gave their proofs for 13 sporadic groups. We will also rely on the HeLP method to produce
a finite number of possibilities for the partial augmentations of units of certain orders and then
Theorem 1.2 will do the rest. We will apply here a computer implementation of the HeLP method
as a package [BM17a] in the computer algebra system GAP [GAP17]. This package can also be
used to reproduce the results of Bovdi, Konovalov et al.
2.2. The lattice method. The lattice method is an other method which makes use of the mul-
tiplicities of eigenvalues of a torsion unit u ∈ V(ZG) of order n under ordinary representations of
G. This method was introduced in [BM17c] and the basic idea consists in obtaining restrictions
on the simple kG-modules when viewed as khui-modules, where k is a field of characteristic p
for a divisor p of n. These simple kG-modules are the p-modular composition factors of ordinary
representations of G and studying different ordinary representations with common composition
factors can finally produce a contradiction to the existence of u.
We will only recall those parts of the lattice method necessary for our means in this article.
We will cite the articles [BM17c, BM18b] where this method is explained, but many facts can be
found in text books on representation theory. First of all let C = hci be a cyclic group of order
pm where p is a prime and m an integer not divisible by p. Let k be a field of characteristic
p containing a primitive m-th root of unity ξ. Then there are, up to isomorphism, exactly pm
indecomposable kC-modules determined by a pair (i, j) where 1 ≤ i ≤ p, 1 ≤ j ≤ m such that a
module M corresponding to (i, j) is i-dimensional as a k-vector space and cp acts as ξj on M . We
denote this module by I j
i and if the action of cp is clear from the context or if m = 1 we simply
write Ii. The module I j
i is simple if and only if i = 1 and each I j
i is uniserial [BM17c, Proposition
2.2].
4
LEO MARGOLIS
Now assume that M is a kC-module, d-dimensional as k-vector space, such that cp acts as ξj
on M for some fixed j. Then by the above we know that M ∼= apIp ⊕ ap−1Ip−1 ⊕ · · · ⊕ a1I1 for
certain non-negative integers a1, ..., ap. Moreover app + ap−1(p − 1) + ... + a1 = d and hence
λ = (p, ..., p
{z }
ap
, p − 1, ..., p − 1
}
{z
ap−1
, . . . , 1, ..., 1
)
{z }a1
is a partition of d. We call λ the partition corresponding to M .
A common framework to work with partitions using combinatorics is provided by considering
Young tableaux and several notions connected with these tableaux. We recall here quickly those
of them we will need in this paper. They all can be found in several textbooks on algebraic
combinatoric as e.g. [Ful97]. A Young diagram corresponding to a partition λ = (λ1, ..., λr) is
a diagram consisting of boxes ordered in rows and columns such that the first row contains λ1
boxes, the second row λ2 boxes etc. A skew diagram is obtained by removing a smaller Young
diagram from a larger one that contains it. I.e. if µ = (µ1, ..., µr) ≤ λ = (λ1, ..., λr), meaning
that µi ≤ λi for all i, then the skew diagram corresponding to λ/µ is obtained by erasing from
the Young diagram corresponding to λ, always from left to right, µ1 boxes in the first row, µ2
boxes in the second row etc. If we fill a skew diagram with entries from an alphabet (in our case
the alphabet always will be the positive integers) a skew diagram becomes a skew tableau T . By
"filling" we mean writing a letter in every box. T is called semistandard if reading a row from left
to right the entries do not decrease and reading a column from top to bottom the entries strictly
increase.
For a box b in a skew tableau T we denote by w(b) the word which we obtain reading T from
top to bottom and from right to left until b, including the entry in b. We denote by w(T ) the
word which we obtain reading all boxes of T in this way, i.e.
if b is the lowest box in the first
(i.e. most left) column of T then w(b) = w(T ). We say that T satisfies the lattice property if for
every box b in T the word w(b) contains the letter 1 at least as many times as the letter 2, the
letter 2 at least as many times as 3 etc. If the maximal letter of w(T ) is s and w(T ) contains
ν1 times the letter 1, ν2 times the letter 2 etc. then ν = (ν1, ..., νs) is called the content of T .
Note that if T is a skew tableau satisfying the lattice property then the content ν = (ν1, ..., νs)
of T is a partition of ν1 + ... + νs as then ν1 ≥ ν2 ≥ ... ≥ νs. If now ν is some partition then the
Littlewood-Richardson coefficient cλ
µ,ν is the number of semistandard skew tableaux corresponding
to λ/µ which satisfy the lattice property and have content ν. A fact of fundamental importance
to us is that the Littlewood-Richardson coefficient cλ
µ,ν is symmetric in µ and ν.
These combinatorial objects play a role in the study of torsion units of group rings via the
following observation [BM18b, Theorem 2.8].
Theorem 2.1. Let C = hci be a cyclic group of order p, for p a prime, and k a field of charac-
teristic p. Let M, U and Q be kC-modules with corresponding partitions λ, µ and ν respectively.
Then M contains a submodule U isomorphic to U such that M/ U ∼= Q if and only if cλ
µ,ν 6= 0.
To pass from multiplicities of eigenvalues under ordinary representations to modules over mod-
ular group algebras we will employ the following [BM17c, Propositions 2.3, 2.4].
Proposition 2.2. Let C = hci be a cyclic group of order pm such that p does not divide m. Let
R be a complete local ring of characteristic 0 containing a primitive m-th root of unity ξ such that
p is contained in the maximal ideal of R and not ramified in R. Denote by k the residue class field
of R and adopt the bar-notation for reduction modulo the maximal ideal of R.
Let D be a representation of C such that the eigenvalues of D(u) in the algebraic closure of the
quotient field of R, with multiplicities, are ξA1, ξ2A2, . . . , ξmAm for certain multisets Ai consisting
of p-th roots of unity. Here also Ai = ∅ is possible. Let ζ be a non-trivial p-th root of unity. Note
that since the sum of the eigenvalues of D(u) is an element in R we know for every i that if Ai
contains ζ exactly r times then Ai contains also ζ2, ...,ζ p−1 exactly r times.
Let M be an RC-lattice affording the representation D. Then
M ∼= M1 ⊕ M2 ⊕ ... ⊕ Mm
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
5
such that for every i we have: If Ai contains ζ exactly r times and 1 exactly s times then
Mi ∼= aI i
p ⊕ (r − a)I i
p−1 ⊕ (s − a)I i
1
for some non-negative integer a ≤ min{r, s}.
For u ∈ V(ZG) a unit of order pm for which we know the partial augmentations, including
those of its powers, these facts allow to derive information on the isomorphism type of simple
kG-modules, for k a big enough field of characteristic p, from the eigenvalues of u under ordinary
representations of G, see the proof of Theorem 1.2, or [BM17c] for a more detailed sketch of the
method.
2.3. A reformulation using character values. The multiplicities in Theorem 1.2 can be com-
puted from the partial augmentations of u and its proper powers. This can be done by hand or using
the GAP-package HeLP [BM18a], more precisely the command HeLP_MultiplicitiesOfEigenvalues.
The theorem can however also be translated in a condition on character values, so that the condi-
tion of Theorem 1.2 can be more easily checked without using GAP (or just to look up the character
values). This is particularly handy if the characters in question only have integral values and the
following lemma provides a way to do this.
Lemma 2.3. Let G be a finite group and u ∈ V(ZG) a unit of order pq where p and q are different
primes. Let χ be a character of G which takes only integral values. Set χ(1) = d, χ(up) = x,
χ(uq) = y and χ(u) = z. Then the following formulas hold.
pq · µ(1, u, χ) = d + (q − 1)x + (p − 1)y + (p − 1)(q − 1)z,
pq · µ(ζp, u, χ) = d + (q − 1)x − y − (q − 1)z,
pq · µ(ζq, u, χ) = d − x + (p − 1)y − (p − 1)z,
pq · µ(ζpq, u, χ) = d − x − y + z.
p, u, χ), for i not divisible by q we have µ(ζq, u, χ) = µ(ζ i
Proof. Recall that the sum over all primitive p-th or q-th roots of unity is −1 while the sum over
all primitive pq-th roots of unity is 1. Note that for i not divisible by p we have µ(ζp, u, χ) =
µ(ζ i
q, u, χ) and for i not divisible by
pq also µ(ζpq, u, χ) = µ(ζ i
pq, u, χ). Denote by D a representation corresponding to χ. So if e.g.
µ(ζp, u, χ) = k, i.e. ζp appears k times as an eigenvalue of D(u), then also each other primitive
p-th root of unity appears k times and together they contribute −k to χ(u) = tr(D(u)). Moreover
thinking of D(u) as a diagonalized matrix and taking the p-th power in D(up) each of these
eigenvalues becomes 1 and so they contribute (p − 1) · k to χ(up).
Arguing in the same manner for primitive pq-th roots of unity we obtain that they contribute
µ(ζpq, u, χ) to χ(u), while contributing e.g. (q − 1)µ(ζpq, u, χ) to χ(uq) as they become primitive
primitive p-roots of unity as eigenvalues of D(uq).
Employing the same arguments for all roots of unity involved in D(u) we obtain:
χ(1) = µ(1, u, χ) + (p − 1)µ(ζp, u, χ) + (q − 1)µ(ζq, u, χ) + (p − 1)(q − 1)µ(ζpq, u, χ),
χ(uq) = µ(1, u, χ) − µ(ζp, u, χ) + (q − 1)µ(ζq, u, χ) − (q − 1)µ(ζpq, u, χ),
χ(up) = µ(1, u, χ) + (p − 1)µ(ζp, u, χ) − µ(ζq, u, χ) − (p − 1)µ(ζpq, u, χ),
χ(u) = µ(1, u, χ) − µ(ζp, u, χ) − µ(ζq, u, χ) + µ(ζpq, u, χ).
The lemma now follows by linear transformations.
(cid:3)
3. Combinatorics
The main result of this section, which is the main ingredient in the proof of Theorem 1.2,
is Proposition 3.5.
It provides bounds on the possible entries in a semistandard skew tableau
satisfying the lattice property for a special form of skew tableaux which are typical for our intended
application.
6
LEO MARGOLIS
Definition 3.1. We call a skew diagram of form A if it consists of p columns such that the second
to (p − 1)-th column all end at the same row, say the m-th row. This is equivalent to saying that
there is a number m such that for a box b in the first to (p − 2)-th column and in the k-th row,
for k ≤ m, there is always a box to the right of b and moreover the second column contains no
box in the (m + 1)-th row. Furthermore we also allow the p-th column to be empty.
We call the boxes in the first to (p − 1)-th column which lie in the m-th row and above the body
of the skew diagram. The boxes of the first column lying below the m-th row are called the tail
and the p-th column the head of the tableau.
For a semistandard skew tableau T satisfying the lattice property we denote by γi(T ) the
number of letters which appear at least i times in w(T ).
Example 3.2. In Figure 1 the upper skew diagram is of form A while the lower one is not. In the
first skew diagram the boxes in the head are marked with h, the boxes of the body with b and the
boxes of the tail with t.
h
b h
b
b
b
b
b
b
b
b
b
b
t
t
Figure 1. A skew diagram of form A and not of form A
Remark: We clarify our way of speaking about the location of a box in a skew tableau relative
to another box, so that the following proofs become more readable. Let b be a box in a skew
tableau T . If we speak about the box lying right from b we mean the unique box neighbouring b
on the right. This box will be denoted by br then. If we speak about a box lying right from b this
can be any box lying in a column of T which lies to the right of the column containing b. We will
also use a different notation for these kind of boxes.
We speak the same way of boxes lying to the left, above or below b.
Lemma 3.3. Let T be a semistandard skew tableau of form A with p columns satisfying the lattice
property. Let b be a box in the k-th column lying in the body of T with entry e where 1 ≤ k ≤ p − 1.
Then w(b) contains e at least (p − k) times.
Proof. We will argue by decreasing induction. For k = p − 1 the claim is clear, since being the
entry in b the letter e appears in w(b) at least once. So let k < p − 1 and let er be the entry in
the box br right from b. Note that br exists by the assumption that T is of form A and b a box
in the body of T . Then by induction the letter er appears at least p − k − 1 times in w(br). As
e ≤ er, since T is semistandard, and w(br) satisfies the lattice property also e must appear at least
p − k − 1 times in w(br). Hence e appears at least p − k times in w(b).
(cid:3)
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
7
Lemma 3.4. Let T be a semistandard skew tableau of form A with p columns satisfying the lattice
property. Let 1 ≤ k ≤ p − 2 and let b be a box in the k-th column of T inside the body of T with
entry e. Assume that the box right from b is the h-th box in the (k + 1)-th column of T . Then
e ≤ h.
Proof. Assume that reading from top to bottom and right to left b is the first box contradicting
the claim. In particular, e > h. Let br be the box right from b with entry er. We will show that
the (k + 1)-th column contains every letter 1, 2, ...,e − 1. As T is semistandrad the entries in the
(k + 1)-th column of T are strictly increasing when read from top to bottom. So assuming that
the column contains all entries between 1 and e − 1 the entry in the first box of the column must
in fact be 1, the entry in the second box 2, ..., the entry in the (e − 1)-th box must be e − 1. In
particular, as e > h, the h-th box will contain h, meaning h = er < e, contradicting that T is a
semistandard tableau. So once we show that the (k + 1)-th column contains every letter 1, 2, ...,
e − 1 the lemma will follow by contradiction.
We will show also that when bk+1 is a box in the (k + 1)-th column containing an entry es
smaller than e then this entry appears exactly p − k times in w(bk+1) and if the box left from bk+1
exists then the entry of this box is strictly smaller than es.
We will argue by decreasing induction. So we will first show that the (k + 1)-th column contains
the letter e − 1, that if bk+1 denotes the box in the (k + 1)-th column containing e − 1, then e − 1
appears in w(bk+1) exactly p − k times and if the box left from bk+1 exists its entry is strictly
smaller than e − 1. By Lemma 3.3 the word w(b) contains the letter e at least p − k times and
so w(b) also contains every letter smaller than e at least p − k times, as w(b) satisfies the lattice
property. By assumption b is the first box violating the lemma when going through the tableau
top to bottom and right to left. So if the box ba above b exists and has entry ea then the box right
from ba is the (h − 1)-th box in the (k + 1)-th column. So ea ≤ h − 1 and e − ea > h − (h − 1) = 1.
In any case, if ba exists or not, we have that the k-th column does not contain the entry e − 1, as
the entries in the k-th column must strictly increase and there are no boxes between the entries
ea and e. Note that then also e − 1 is not the entry in a box lying above b and to the left of b,
since this would be a box contradicting the fact that T is semistandard. Since w(b) contains e − 1
at least p − k times and w(b) does not take into account boxes lying below b as well as boxes lying
left from b and not above b, we conclude that every column right from b must contain e − 1. In
particular, the (k + 1)-th column. Moreover if bk+1 is the box in the (k + 1)-th column containing
e − 1 then e − 1 appears exactly p − k times in w(bk+1), as the boxes between bk+1 and b (counted
right to left and top to bottom) do not contain the letter e − 1. Furthermore bk+1 is a box above
br, since otherwise er ≤ e − 1 < e, which would contradict the fact that T is semistandard. Since
the k-th column does not contain e − 1 the box lying left from bk+1, if it exists, must contain an
entry smaller than e − 1, since T is semistandard. This proofs the base case of our induction.
The arguments for the induction step are very similar to the arguments for the base case. So
assume that es ≤ e − 1 is an entry contained in the box bk+1 which lies in the (k + 1)-th column,
that es appears exactly p − k times in w(bk+1) and that the box left from bk+1, if existent, contains
an entry smaller than es. Denote the box left from bk+1, if existent, by (bk+1)l. Then (bk+1)l
contains an entry strictly smaller than es by induction. So the the box above (bk+1)l, if existent,
has maximal entry es − 2, as T is semistandard. In particular es − 1 is not an entry in a box lying
above and to the left of bk+1. But, by Lemma 3.3, the word w(bk+1) contains es at least p − k
times, so it contains es − 1 at least p − k times and we get that es − 1 must be an entry in the
(k + 1)-th column. So it must lie in the box (bk+1)a above bk+1. Then es − 1 appears exactly p − k
times in w((bk+1)a), as it does not appear in a box above and to the left of (bk+1)a. Also, the box
left from (bk+1)a, if existent, contains an entry smaller than es − 1. This finishes the induction
step and thus the lemma is proved.
(cid:3)
Proposition 3.5. Let T be a semistandard skew tableau of form A with p columns satisfying the
lattice property. Let 1 ≤ k ≤ p − 1 and let h be the height of the k-th column inside the body of
T , i.e. the number of boxes in the k-th column of T which lie in the body (so if k = 1 we do not
count the boxes in the tail).
Then γp−k(T ) ≥ h and if k ≥ 2 then γp−k+2(T ) ≤ h.
8
LEO MARGOLIS
Proof. Let b be the lowest box in the k-th column and assume the entry of b is e.
Then just for the reason that the entries in a column are strictly increasing we have h ≤ e. By
Lemma 3.3 the entry e appears in w(b) at least p − k times, so each letter 1, 2 ..., h appears at
least p − k times in w(T ) and hence γp−k(T ) ≥ h.
So assume k ≥ 2 and assume s = γp−k+2(T ) > h. By Lemma 3.4 the maximal possible entry
in the box left from b, if this box exists, is h. So also an entry in a box to the left from b which
lies in the body is at most h (note that there is no box below b anywhere in the body, as b is the
lowest box in its column and T of form A). So inside the body an entry bigger than h does not
appear in the 1st to (k − 1)-th column. Hence each letter h + 1, ..., s must appear in the tail and
in the k-th column and every column right from the k-th column, as otherwise it can not appear
p − k + 2 or more times in total. We will show that the k-th column then contains every letter
1,..., s, which will yield the final contradiction, since clearly the k-th column can not contain more
than h entries. The proof is similar to the proof of Lemma 3.4.
Let b be the box in the k-th column containing s. Note that also s can not appear in a box
lying below and to the right of b, as T is semistandard. Then s appears in w(b) exactly p − k + 1
times, since it appears at least overall p − k + 2 times in w(T ) and it does not appear in the 1st
to (k − 1)-th column inside the body. The box left from b, if existent, contains an entry smaller
than h, so also smaller than s, by Lemma 3.4. We will show by decreasing induction that the
k-th column contains every entry e between 1 and s such that if be is the box containing e then e
appears exactly p − k + 1 times in w(be) and the box left from be, if existent, contains an entry
smaller than e. By the above this holds for e = s, providing the base case of the induction.
So assume that this holds for a certain e + 1 and let be+1 be the box in the k-th column
containing e + 1. Denote by (be+1)a, if it exists, the box above be+1 and by (be+1)l, if it exists,
the box to the left of be+1. If (be+1)l exists its entry is at most e, by induction. So if the box
above (be+1)l exists its entry is at most e − 1. This implies that any entry lying to the left and
not below (be+1)a is smaller than e. Since by induction e + 1 appears exactly p − k + 1 times in
w(be+1), also e appears at least p + k − 1 times in w(be+1). Combining this with the conclusion
before, we see that (be+1)a must exist and its entry must be e. Moreover the letter e appears
exactly p − k + 1 times in w((be+1)a) and the box to the left of (be+1)a contains an entry smaller
than e. This finishes the induction step and hence the proof.
(cid:3)
4. Proofs of main results
We proceed to prove Theorem 1.2. For the basic facts about Brauer trees we refer to [HL89].
Here we will only need that if (K, R, k) is a p-modular system such that the characters labelling
the vertices of the Brauer tree are afforded by RG-modules then each edge of the Brauer tree
corresponds to a simple kG-module, each vertex of the Brauer tree corresponds to a simple RG-
module with the given character and after reducing such a simple RG-module modulo the maximal
ideal of R the composition factors of this module are those corresponding to the adjacent edges,
each with multiplicity 1.
Proof of Theorem 1.2: Let R be a complete discrete valuation ring containing a primitive m-th
root of unity such that R is an algebraic extension of Zp, representations affording the characters
..., χp can be realized over R and p is unramified in R. Such a ring R exists by a result
χ1,
of Fong [BM18b, Remark 2.5]. Let k be the residue class field of R. We use the bar-notation
to denote reduction modulo the maximal ideal of R also with respect to modules. Let M be a
khui-module. Adapting the notation from skew tableaus we denote by γi(M ) the number of direct
indecomposable summands of M of dimension at least i.
Let M ′
1, ..., M ′
1, ..., S ′
p be RG-modules corresponding to the characters χ1, ..., χp and let, as explained
p respectively on which up acts
in Proposition 2.2, M1, ..., Mp be the direct summands of M ′
as ξp. Let S ′
p−1 be the simple kG-modules corresponding to the edges of the Brauer tree
in the natural order, i.e. S ′
i corresponds to the edge with vertices χi and χi+1. Let S1, ..., Sp−1
p−1 respectively on which up acts as ξp. We will view Mi
be the direct summand of S ′
1,
and Si as khui-modules, unless explicitly stated otherwise. Let λi be the partition corresponding
to Mi and µi be the partition corresponding to Si, in the sense of Section 2.2. Note that by
1, ..., M ′
..., S ′
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS
9
Proposition 2.2 if Sj is a submodule of Mi then the skew diagram corresponding to λi/µj is of
form A with p columns, since the indecomposable summands of Mi are all of dimension 1, p − 1
or p, i.e. the possible length of rows in a Young diagram corresponding to λi are 1, p − 1 and p.
We will use the following fact several times.
Claim: If we want to compute the possible isomorphism types of Si for some 1 < i < p − 1
using Theorem 2.1 we can consider a semistandard skew tableau satisfying the lattice property of
form λi/µi−1 or λi+1/µi+1.
Proof of the claim: The composition factors of M ′
properties of Brauer trees. If S ′
i−1 is a submodule of M ′
Si corresponds to a skew tableau of form λi/µi−1 and cλi
submodule of M ′
µi,µi−1 6= 0. As the Littlewood-Richardson coefficient is symmetric this also implies cλi
cλi
So it suffices to consider λi/µi−1, or rather the question if cλi
i as kG-module are S ′
i by the
i as kG-module then Si ∼= Mi/Si−1. So
µi−1,µi 6= 0. If on the other hand S ′
i is a
i as kG-module then Si−1 corresponds to a skew tableau of the form λi/µi and
µi−1,µi 6= 0.
i−1 and S ′
µi−1,µi 6= 0.
If S ′
i is a submodule of M ′
On the other hand we can consider M ′
i+1 as kG-module which has composition factors S ′
S ′
i+1.
a skew tableau of form λi+1/µi and cλi+1
corresponds to a skew tableau of the form λi+1/µi+1 and cλi+1
Littlewood-Richardson coefficient implies that it suffices to consider if cλi+1
the proof of the claim.
i and
i as kG-module then Si+1 ∼= Mi+1/Si. So Si corresponds to
µi,µi+1 6= 0. On the other hand, as in the case before, Si
6= 0. The symmetry of the
µi+1,µi 6= 0. This finishes
µi+1,µi
We will next show that
(1)
and also
(2)
γp−2(Sp−2) ≥ µ(ξ · ζp, u, χp−1) − µ(ξ · ζp, u, χp)
γp−2(Sp−2) ≤ µ(ξ, u, χ1) −
p−2
X
(−1)iµ(ξ · ζp, u, χi)
i=1
which will imply the theorem. First by the claim above the isomorphism type of Sp−2 can be
described by a semistandard skew tableau T satisfying the lattice property of shape λp−1/µp−1.
Note that µp−1 = λp since Mp ∼= Sp−1, as Mp corresponds to a vertex of the Brauer tree which
has only one neighbour. By Proposition 3.5 we know that γp−2(T ) = γp−2(Sp−2) is at least as big
as the height of the second column of T . Now by Proposition 2.2 the number of indecomposable
direct summands of dimension at least 2 in Mp−1, i.e. the number of rows of length at least 2 in a
Young diagram corresponding to λp−1, is µ(ξ · ζp, u, χp−1) and the analogues statement holds for
λp. So the height of the second column of T is µ(ξ · ζp, u, χp−1) − µ(ξ · ζp, u, χp), proving (1).
To prove (2) we will prove by induction on r that if r is odd then
and if r is even then
γr(Sr) ≤ µ(ξ, u, χ1) −
r
X
i=1
(−1)iµ(ξ · ζp, u, χi)
γp+1−r(Sr) ≥ −µ(ξ, u, χ1) +
r
X
i=1
(−1)iµ(ξ · ζp, u, χi).
Once this induction reaches r = p − 2, which is an odd number, the first inequality will be
exactly (2) and so the proof of the theorem will be finished.
For the base case consider r = 1. Then γ1(S1) is the number of indecomposable direct summands
of S1 ∼= M1 which by Proposition 2.2 is at most µ(ξ, u, χ1) + µ(ξ · ζp, u, χ1).
For the induction step assume that the statement holds for all numbers smaller than r. First
assume that r is even. Now Sr corresponds to a semistandard skew tableau T satisfying the lattice
property of form λr/µr−1. So by Proposition 3.5 the number γp+1−r(Sr) = γp+1−r(T ) is at least
as big as the height of the (r − 1)-th column in the body of T . Since r ≤ p this numbers equals
the difference of the number of direct indecomposable summands of Mr which is at least (r − 1)-
dimensional, but not 1-dimensional, and the number of indecomposable direct summands of Sr−1
10
LEO MARGOLIS
of dimension at least r−1, i.e. γr−1(Sr−1). The number of indecomposable direct summands of Mr
which are at least (r − 1)-dimensional, but not 1-dimensional, is µ(ξ · ζp, u, χr) by Proposition 2.2.
Since by induction γr−1(Sr−1) ≤ µ(ξ, u, χ1) − Pr−1
i=1 (−1)iµ(ξ · ζp, u, χi) we obtain that the height
of the (r − 1)-th column in the body of T is at least
µ(ξ · ζp, u, χr) − (µ(ξ, u, χ1) −
r−1
X
i=1
(−1)iµ(ξ · ζp, u, χi)) = −µ(ξ, u, χ1) +
r
X
i=1
(−1)iµ(ξ · ζp, u, χi).
Now assume that r is odd, bigger than 2 and smaller than p−1. Also in this case Sr corresponds
to a semistandard skew tableau T satisfying the lattice property of form λr/µr−1. So the number
γr(Sr) = γr(T ) is at most as big as the height of the (p + 2 − r)-th column in the body of T
by Proposition 3.5. This is the difference of the number of indecomposable direct summands of
dimension at least p + 2 − r in Mr, i.e. γp+2−r(Mr), and the number of indecomposable direct
summands of Sr−1 of dimension at least p + 2 − r, i.e. γp+2−r(Sr−1) = γp+1−(r−1)(Sr−1). By
Proposition 2.2 we know γp+2−r(Mr) = µ(ξ · ζp, u, χr) and by induction γp+1−(r−1)(Sr−1) is at
least −µ(ξ, u, χ1) + Pr−1
i=1 (−1)iµ(ξ · ζp, u, χi). So the height of the (p + 2 − r)-th column of T is
at most
µ(ξ · ζp, u, χr) − (−µ(ξ, u, χ1) +
r−1
X
i=1
(−1)iµ(ξ · ζp, u, χi)) = µ(ξ, u, χ1) −
r
X
i=1
(−1)iµ(ξ · ζp, u, χi).
This finishes the proof of the induction claim and hence the proof of the theorem.
(cid:3)
The proof of the Prime Graph Question for the Conway simple groups now boils down to the
application of Theorem 1.2 for various cases described in [BKL11]. The information contained in
[BKL11] is not completely sufficient for our purposes, since to compute multiplicities of eigenvalues
we do not only need to know the partial augmentations of a unit u ∈ V(ZG), but also the partial
augmentations of its proper powers. These can be computed using the GAP-package HeLP [BM17a]
and we will indicate in all cases we need to consider which characters are sufficient to obtain these
possible partial augmentations using the command HeLP_WithGivenOrder. Also in the case of the
first Conway group we will use stronger results obtainable by the HeLP method than those given
in [BKL11].
We will denote by χi the i-th irreducible complex character of a group G as given in the GAP
character table library [Bre12]. We will also use names for conjugacy classes as in [Bre12]. The
statements we will need about the p-blocks of various groups, their defect and their corresponding
Brauer trees can all be derived from GAP, but they are also given in [HL89].
Proof of Theorem 1.1: We will study the three cases of interest separately.
Case G = Co3: The information on the orders of elements in G is contained in [CCN+85].
Using this information and [BKL11, Theorem 1(i)] we only have to consider units of order 35 in
V(ZG). Using χ2 to compute partial augmentations for units of order 5 and χ2, χ3 to compute
the partial augmentations of units of order 35 we are left with the following possibilities:
(ε5a(u7), ε5b(u7), ε5a(u), ε5b(u), ε7a(u)) ∈ {(−4, 5, 3, 12, −14), (−3, 4, 4, 11, −14)}.
Note that there is only one class of elements of order 7 in G, so we do not need to consider
the partial augmentations of u5, since the only class in which a partial augmentation at u5 is
non-vanishing is 7a.
G possesses a 5-block of defect 1 whose Brauer tree is a line of form
χ5
χ29
χ39
χ35
χ12
.
All irreducible ordinary characters in this block have only integral values and 5 is of course not
ramified in Z. So we can apply Theorem 1.2. We provide the multiplicities of the needed eigen-
values of u, for the two critical distributions of partial augmentations, in Table 1. Here the first
entry in each column contains the possible values of (ε5a(u7), ε5b(u7), ε5a(u), ε5b(u), ε7a(u)).
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS 11
(−4, 5, 3, 12, −14)
(−3, 4, 4, 11, −14)
µ(1, u, χ) µ(ζ5, u, χ) µ(1, u, χ) µ(ζ5, u, χ)
33
29
2
2119
7029
5071
104
3
2118
7030
5070
105
χ
χ5
χ29
χ39
χ35
χ12
χ
χ5
χ29
χ39
χ35
χ12
Table 1. Multiplicities of eigenvalues for G = Co3 for units of order 35.
So using Theorem 1.2 with ξ = 1 we would have, e.g.
if the partial augmentations are
(ε5a(u7), ε5b(u7), ε5a(u), ε5b(u), ε7a(u)) = (−4, 5, 3, 12, −14) that
µ(ζ5, u, χ35) − µ(ζ5, u, χ12) = 4967 ≤ µ(1, u, χ5) + µ(ζ5, u, χ5) − µ(ζ5, u, χ29) + µ(ζ5, u, χ39) = 4945,
a contradiction. The same argument applies to the other possibility of distributions of partial
augmentations of u, in which case we get 4965 ≤ 4944.
Case G = Co2: The information on the orders of elements in G is contained in [CCN+85].
Using this information and [BKL11, Theorem 2(i)] we only have to consider units of order 35 in
ZG. Using χ2 to compute partial augmentations of units of order 5 and χ2, χ3 for order 35 we get
(ε5a(u7), ε5b(u7), ε5a(u), ε5b(u), ε7a(u)) ∈ {(−4, 5, 3, 12, −14), (−3, 4, 4, 11, −14)}.
Again there is only one class of elements of order 7, so we do not need to consider u5.
Also in this case G possesses a 5-block of defect 1 with Brauer tree of the form
χ4
χ24
χ43
χ38
χ20
.
All ordinary characters in this block only have integral values. The necessary multiplicities are
provided in Table 2
(−4, 5, 3, 12, −14)
(−3, 4, 4, 11, −14)
µ(1, u, χ) µ(ζ5, u, χ) µ(1, u, χ) µ(ζ5, u, χ)
33
29
2
3269
13354
11396
1254
3
3268
13355
11395
1255
Table 2. Multiplicities of eigenvalues for G = Co2 for units of order 35.
In the first case using Theorem 1.2 with ξ = 1 we get 10142 ≤ 10120 and in the second case
10140 ≤ 10119.
Case G = Co1: The information on the orders of elements in G is contained in [CCN+85].
Using this information and [BKL11, Theorem 3(i)] we only have to consider units of order 55 and
65 in ZG. We will first apply the HeLP method here to obtain stronger results than in [BKL11].
The 2-, 3- and 5-modular Brauer tables are not available in GAP, but the online version of the
Atlas, also implemented as a GAP-package [WPN+11], contains several representations in these
characteristics and we will use a character ψ coming from a 24-dimensional representation of G
over F2. The values of ψ on classes of interest can be found in Table 3.
1a
5a
ψ 24 −6
5b
5c
4 −1
11a 13a
−2
2
Table 3. A 2-modular Brauer character for Co1.
12
LEO MARGOLIS
The results we obtain using the HeLP method are stronger than those in [BKL11], since we
include ψ. Using χ2, χ3 and ψ we obtain 98 possibilities for the partial augmentations of units of
order 5. Then using χ2, χ4, χ5 and ψ we can exclude the existence of units of order 65 in ZG.
For units of order 55 we apply the characters χ2, χ3, χ4 and ψ. Note that there is only one class
of elements of order 11 in G.
We get that (ε5a(u11), ε5b(u11), ε5c(u11), ε5a(u), ε5b(u), ε5c(u), ε11a(u)) is one of the four possi-
bilities
(1, 5, −5, 1, −6, −5, 11), (1, 6, −6, 1, −5, −6, 11), (2, 6, −7, 2, −5, −7, 11), (2, 7, −8, 2, −4, −8, 11).
The principal 11-block of G has a Brauer tree of the form
χ1
χ12
χ34
χ41
χ64
χ79
χ85
χ73
χ60
χ38
χ14
.
Also in this block all ordinary characters only have integral values. The necessary multiplicities
are provided in Table 4.
(1, 5, −5, 1, −6, −5, 11)
(1, 5, −5, 1, −6, −5, 11)
(1, 5, −5, 1, −6, −5, 11)
(1, 5, −5, 1, −6, −5, 11)
µ(1, u, χ) µ(ζ11, u, χ) µ(1, u, χ) µ(ζ11, u, χ) µ(1, u, χ) µ(ζ11, u, χ) µ(1, u, χ) µ(ζ11, u, χ)
1
1
1
1
0
5670
138600
391385
1929870
4495195
5326495
3734505
1522180
297285
6875
0
5638
138520
391350
1929856
4495195
5326570
3734540
1522180
297303
6880
χ
χ1
χ12
χ34
χ41
χ64
χ79
χ85
χ73
χ60
χ38
χ14
0
5668
138600
391385
1929876
4495195
5326485
3734505
1522180
297293
6885
0
5640
138520
391350
1929850
4495195
5326580
3734540
1522180
297295
6870
Table 4. Multiplicities of eigenvalues for G = Co1 for units of order 55.
From the four possibilities we obtain by Theorem 1.2 the inequalities 290408 ≤ 290389, 290410 ≤
(cid:3)
290391, 290423 ≤ 290304 and 290425 ≤ 290306, all of which do not hold.
Remark 4.1. In the proof of Theorem 1.1 to handle the first Conway group Co1 we used Theo-
rem 1.2 with p = 11 instead of p = 5 for two reasons. One is to show that Theorem 1.2 can also be
applied with other primes than 5. The other, more importantly, is that the 11-Brauer trees can be
computed using GAP which can not be done as easily with 5-Brauer trees, since the 5-Brauer table
is not available in GAP. The 5-Brauer trees are given in [HL89], but the calculations with p = 11
are easier to verify for a reader who does not have [HL89] available. Using the first Brauer tree
given in [HL89, Section 6.22] and Theorem 1.2 with p = 5 and ξ = 1 one can also get the same
result.
Remark 4.2. Of the 17 sporadic groups studied by Bovdi, Konovalov et al. the only group for
which they could not prove the Prime Graph Question, apart from the Conway groups, was the
O'Nan simple group [BGK09]. Here it remains to exclude the existence of units of order 33 and
57. Theorem 1.2 can be applied here with p = 11 to handle units of order 33, but not for order
57. This is not possible since for p ∈ {3, 19} every p-block of G which is a Brauer Tree Algebra
contains characters such that p is ramified in the ring of values of these characters.
Acknowledgements
The setting handled in this paper was discussed at various stages with Alexander Konovalov,
Andreas Bächle, Alexander Thumm and Eugenio Giannelli. I am thankful to them all for their
input.
I would also like to thank the referee whose comments helped to improve the readability, par-
ticularly of the combinatorial proofs, significantly.
ON THE PRIME GRAPH QUESTION FOR INTEGRAL GROUP RINGS OF CONWAY SIMPLE GROUPS 13
References
[BC17]
A. Bächle and M. Caicedo, On the prime graph question for almost simple groups with an alternating
socle, Internat. J. Algebra Comput. 27 (2017), no. 3, 333 -- 347.
[BGK09] V. Bovdi, A. Grishkov, and A. Konovalov, Kimmerle conjecture for the Held and O'Nan sporadic simple
[BJK11]
[BK07a]
groups, Sci. Math. Jpn. 69 (2009), no. 3, 353 -- 361.
V. Bovdi, E. Jespers, and A. Konovalov, Torsion units in integral group rings of Janko simple groups,
Math. Comp. 80 (2011), no. 273, 593 -- 615.
V. Bovdi and A. Konovalov, Integral group ring of the first Mathieu simple group, Groups St. Andrews
2005. Vol. 1, London Math. Soc. Lecture Note Ser., vol. 339, Cambridge Univ. Press, Cambridge, 2007,
pp. 237 -- 245.
[BK07b]
, Integral group ring of the McLaughlin simple group, Algebra Discrete Math. (2007), no. 2,
[BK08]
[BK09]
[BK10]
[BK12]
[BKL08]
[BKL11]
43 -- 53.
2680.
, Integral group ring of the Mathieu simple group M23, Comm. Algebra 36 (2008), no. 7, 2670 --
, Integral group ring of Rudvalis simple group, Ukraïn. Mat. Zh. 61 (2009), no. 1, 3 -- 13.
, Torsion units in integral group ring of Higman-Sims simple group, Studia Sci. Math. Hungar.
47 (2010), no. 1, 1 -- 11.
, Integral group ring of the Mathieu simple group M24, J. Algebra Appl. 11 (2012), no. 1,
1250016, 10.
V. Bovdi, A. Konovalov, and S. Linton, Torsion units in integral group ring of the Mathieu simple
group M22, LMS J. Comput. Math. 11 (2008), 28 -- 39.
, Torsion units in integral group rings of Conway simple groups, Internat. J. Algebra Comput.
21 (2011), no. 4, 615 -- 634.
[BKM08] V. Bovdi, A. Konovalov, and E. Marcos, Integral group ring of the Suzuki sporadic simple group, Publ.
[BKS07]
[BM17a]
Math. Debrecen 72 (2008), no. 3-4, 487 -- 503.
V. Bovdi, A. Konovalov, and S. Siciliano, Integral group ring of the Mathieu simple group M12, Rend.
Circ. Mat. Palermo (2) 56 (2007), no. 1, 125 -- 136.
A. Bächle and L. Margolis, HeLP -- Hertweck-Luthar-Passi method, GAP package, Version 3.1,
http://homepages.vub.ac.be/abachle/help/, 2017.
[BM17b]
, On the prime graph question for integral group rings of 4-primary groups I, Internat. J. Algebra
Comput. 27 (2017), no. 6, 731 -- 767.
[BM17c]
[BM18a]
, Rational Conjugacy of Torsion Units in Integral Group Rings of Non-Solvable Groups, Proc.
Edinb. Math. Soc. (2) 60 (2017), no. 4, 813 -- 830.
, HeLP: a GAP package for torsion units in integral group rings, J. Softw. Algebra Geom. 8
(2018), 1 -- 9.
[BM18b]
, On the Prime Graph Question for Integral Group Rings of 4-primary groups II, Algebr.
[Bre12]
Represent. Theory (2018), 1 -- 21, in press, doi: 10.1007/s10468-018-9776-6.
T.
http://www.math.rwth-aachen.de/~Thomas.Breuer/ctbllib, May 2012, GAP package.
Character
Breuer,
Library,
Table
The
GAP
Version
1.2.1,
[CL65]
[EM18]
[CCN+85] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, and R. A. Wilson, Atlas of finite groups, Oxford
University Press, Eynsham, 1985, Maximal subgroups and ordinary characters for simple groups, With
computational assistance from J. G. Thackray.
J. A. Cohn and D. Livingstone, On the structure of group algebras. I, Canad. J. Math. 17 (1965),
583 -- 593.
F. Eisele and L. Margolis, A counterexample to the first Zassenhaus conjecture, Adv. Math. 339 (2018),
599 -- 641.
W. Fulton, Young tableaux, London Mathematical Society Student Texts, vol. 35, Cambridge University
Press, Cambridge, 1997, With applications to representation theory and geometry.
The GAP Group, GAP -- Groups, Algorithms,
http://www.gap-system.org.
M. Hertweck, A counterexample to the isomorphism problem for integral group rings, Ann. of Math.
(2) 154 (2001), no. 1, 115 -- 138.
and Programming, Version 4.8.8,
[GAP17]
[Her01]
[Ful97]
2017,
[Her07]
[HL89]
[JdR16]
, Partial augmentations and Brauer character values of
torsion units in group rings,
arXiv:0612429v2 [math.RA], 2004 - 2007.
G. Hiss and K. Lux, Brauer trees of sporadic groups, Oxford Science Publications, The Clarendon
Press, Oxford University Press, New York, 1989.
E. Jespers and Á. del Río, Group ring groups. Vol. 1. Orders and generic constructions of units, De
Gruyter Graduate, De Gruyter, Berlin, 2016.
[Kim06] W. Kimmerle, On the prime graph of the unit group of integral group rings of finite groups, Groups,
[KK15]
rings and algebras, Contemp. Math., vol. 420, Amer. Math. Soc., Providence, RI, 2006, pp. 215 -- 228.
W. Kimmerle and A. Konovalov, Recent advances on torsion subgroups of integral group rings, Groups
St Andrews 2013, London Math. Soc. Lecture Note Ser., vol. 422, Cambridge Univ. Press, Cambridge,
2015, pp. 331 -- 347.
14
[KK17]
[LP89]
[Seh93]
LEO MARGOLIS
, On the Gruenberg -- Kegel graph of integral group rings of finite groups, Internat. J. Algebra
Comput. 27 (2017), no. 6, 619 -- 631.
I. S. Luthar and I. B. S. Passi, Zassenhaus conjecture for A5, Proc. Indian Acad. Sci. Math. Sci. 99
(1989), no. 1, 1 -- 5.
S. K. Sehgal, Units in integral group rings, Pitman Monographs and Surveys in Pure and Applied
Mathematics, vol. 69, Longman Scientific & Technical, Harlow, 1993.
J.N. Bray,
and T. Breuer,
R.A.
A GAP
Rep,
http://www.math.rwth-aachen.de/~Thomas.Breuer/atlasrep, July 2011, Refereed GAP package.
of Group Representations,
Interface
Version
[WPN+11] R.A. Wilson,
Parker,
to
S. Nickerson,
the Atlas
Atlas-
1.5,
Vrije Universiteit Brussel, Faculty of Sciences, Department of Mathematics, Pleinlaan 2, B-1050
Brussels, Belgium
E-mail address: [email protected]
|
1304.5767 | 1 | 1304 | 2013-04-21T17:01:06 | Cohomology of n-ary-Nambu-Lie superalgebras and super_1 3-algebra | [
"math.RA"
] | The purpose of this paper is define the representation and the cohomology of n-ary-Nambu-Lie superalgebras. Morever we study extensions and provide the computation of the derivations and second cohomology group of super w_1 3-algebra | math.RA | math |
Cohomology of n-ary-Nambu-Lie superalgebras and
super w∞ 3-algebra
Faouzi AMMAR
Universit´e de Sfax
Facult´e des Sciences,
Nejib Saadaoui
Universit´e de Gab`es
Facult´e des Sciences,
B.P. 1171, Sfax 3000, Tunisie,
Campus Universitaire, Gab`es 6072, Tunisie,
[email protected]
[email protected]
October 16, 2018
Abstract
The purpose of this paper is define the representation and the cohomology of n-ary-Nambu-
Lie superalgebras. Morever we study central extensions and provide as application the compu-
tations of the derivations and second cohomology group of super w∞ 3-algebra.
Introduction
The first instances of n-ary algebras in Physics appeared with a generalization of the Hamiltonian
mechanics proposed in 1973 by Nambu [5]. More recent motivation comes from string theory and
M-branes involving naturally an algebra with ternary operation called Bagger-Lambert algebra
which gives impulse to a significant development. It was used in [6] as one of the main ingredients
in the construction of a new type of supersymmetric gauge theory that is consistent with all the
symmetries expected of a multiple M2-brane theory: 16 supersymmetries, conformal invariance,
and an SO(8) R-symmetry that acts on the eight transverse scalars. On the other hand in the
study of supergravity solutions describing M2-branes ending on M5-branes, the Lie algebra ap-
pearing in the original Nahm equations has to be replaced with a generalization involving ternary
bracket in the lifted Nahm equations, see [7]. For other applications in Physics see [8, 9, 10].
The algebraic formulation of Nambu mechanics is due to Takhtajan [11, 19] while the abstract
definition of n-ary Nambu algebras or n-ary Nambu-Lie algebras (when the bracket is skew sym-
metric) was given by Filippov in 1985 see [13]. The Leibniz n-ary algebras were introduced and
studied in [14]. For deformation theory and cohomologies of n-ary algebras of Lie type, we refer to
[15, 16, 17, 18, 19].
The paper is organized as follows.
In the first section we give the definitions and some key
constructions of n-ary-Nambu-Lie superalgebras.
In Section 2 we define a derivation of n-ary-
Nambu-Lie superalgebra. Section 3 is dedicated to the representation theory n-ary-Nambu-Lie
superalgebras including adjoint representation. In Section 3 we construct family of cohomologies
1
of n-ary-Nambu-Lie superalgebras. In Section 4, we discuss extensions of n-ary-Nambu-Lie super-
algebras and their connection to cohomology. In the last section we compute the derivations and
cohomology group of the super w∞ 3-algebra.
1 The n-ary-Nambu-Lie superalgebra
Let N be a linear superspace over a field K that is a Z2-graded linear space with a direct sum
N = N0 ⊕ N1.
The elements of Nj, j ∈ Z2, are said to be homogenous of parity j. The parity of a homogeneous
element x is denoted by x.
Let x1, · · · , xn be n homogenous elements of N , we denote by (x1, · · · , xn) = x1 + · · · + xn the
parity of an element (x1, · · · , xn) in N n.
The space End(N ) is Z2 graded with a direct sum End(N ) = (End(N ))0 ⊕ (End(N ))1 where
(End(N ))j = {f ∈ End(N )/f (Ni) ⊂ Ni+j}. The elements of (End(N ))j are said to be homogenous
of parity j.
Definition 1.1. [2] An n-ary-Nambu superalgebra is a pair (N , [., · · · , .]) constiting of a vector
superspace N and even n-linear map [., . . . , .] : N n → N , satisfying
hy2, . . . , yn, [x1, . . . , xn]i =
nXi=1
(−1)(y2+···+yn)(x1+···+xi−1)hx1, . . . , [y2, . . . , yn, xi], . . . , xni
(1.1)
for all (x1, · · · , xn) ∈ N n, (y2, · · · , yn) ∈ N n−1.
The identity (1.1) is called Nambu identity.
Definition 1.2. An n-ary-Nambu superalgebra (N , [., · · · , .]) is called n-ary-Nambu-Lie superal-
gebra if the bracket is skew-symmetric that is
[x1, . . . , xi−1, xi, . . . , xn] = −(−1)xi−1xi[x1, . . . , xi, xi−1, . . . , xn].
(1.2)
Definition 1.3. Let (N , [., . . . , .]) and (N ′, [., . . . , .]′) be two n-ary-Nambu-Lie superalgebra. An
homomorphism f : N → N ′ is said to be morphism of n-ary-Nambu-Lie superalgebra if
[f (x1), . . . , f (xn)]′ = f ([x1, . . . , xn]) ∀x1, . . . , xn ∈ N .
(1.3)
Proposition 1.4. Let f be an even endomorphism of n-ary-Nambu-Lie superalgebra (N , [., · · · , .]).
We can define the new n-ary-Nambu-Lie superalgebra (N , [., · · · , .]′) , where [x1, . . . , xn]′ = f ([x1, · · · , xn]).
2 Derivation of n-ary-Nambu-Lie superalgebra
Definition 2.1. [2] We call D ∈ Endi(N ), where i is in Z2, a derivation of the n-ary-Nambu-Lie
superalgebra (N , [., · · · , .]) if
D([x1, · · · , xn]) =
nXk=1
(−1)D(x1+···+xk−1)[x1, · · · , D(xk), · · · , xn], f or all homogeneous x1, · · · , xn ∈ N .
2
We denote by Der(N ) = Der0(N ) ⊕ Der1(N ) the set of derivation of the n-ary-Nambu-Lie
superalgebra N .
The subspace Der(N ) ⊂ End(N ) is easily seen to be closed under the bracket
[D1, D2] = D1 ◦ D2 − (−1)D1D2D2 ◦ D1
(2.1)
(known as the supercommutator) and it is called the superalgebra of derivations of N .
With above notation, Der(N ) is a Lie superalgebra, in wich the Lie bracket is given by (2.1).
Fix n−1 homogeneous elements x1, · · · , xn−1 ∈ N , and define the transformations ad(x1, · · · , xn−1) ∈
End(N ) by the rule
ad(x1, · · · , xn−1)(x) = [x1, · · · , xn−1, x].
(2.2)
Then ad(x1, · · · , xn−1) is a derivation of N , wich we call inner derivation of N .
Indeed we have
ad(y2, · · · , yn)(cid:16)[x1, · · · , xn](cid:17) = hy2, · · · , yn, [x1, · · · , xn]i
=
=
(−1)(y2+···+yn)(x1+···+xi−1)hx1, · · · , ad(y)(xi), · · · , xni
nXi=1
(−1)ad(y)(x1+···+xi−1)hx1, · · · , ad(y)(xi), · · · , xni.
nXi=1
3 Representations of n-ary-Nambu-Lie superalgebra
We provide in the following a graded version of the study of representations of n-ary-Nambu-Lie
algebra stated in [3].
Let (N , [., · · · , .]) be a n-ary-Nambu-Lie superalgebra and V = V0 ⊕ V1 an arbitrary vector
, Vj]V ⊂ Vi+j where
superspace. Let [., .]V : N n−1 × V −→ V be a bilinear map satisfying [N n−1
i, j ∈ Z2.
i
Definition 3.1. The pair (V, [., .]V ) is called a module on the n-ary-Nambu-Lie superalgebra N =
N0 ⊕ N1 or N -module V if the even multilinear mapping [., . . . , .]V satisfies
=
had(x)(xn), y2, . . . , yn−1, viV
(−1)n−i+xi(xi+1+···+xn)hx1, . . . ,bxi . . . , xn,(cid:2)xi, y2, . . . , yn−1, v(cid:3)ViV
nXi=1
(−1)y(x1+···+xi−1)hx1, . . . , ad(y)(xi), . . . , xn−1, viV
n−1Xi=1
,
(3.1)
=hy, [x, v]V(cid:3)V − (−1)yx(cid:2)x, [y, v]ViV
(3.2)
for all homogeneous x, y in N n−1 and v in V . It will also say that (V, [., . . . , .]V ) is a representation
of N .
3
Example 3.2. Let (N , [., . . . , .]) be a n-ary-Nambu-Lie superalgebra and the map ad defined in
(2.2). Then (N , ad) is a representation of N .
Remark 3.3. When [., .]V is the zero map, we say that the module V is trivial.
4 Cohomology of n-ary-Nambu-Lie superalgebra induced by co-
homology of Leibniz algebras
In this section, we aim to extend to n-ary Nambu superalgebra type of process introduced by
Takhtajan to construct a complex of n-ary Nambu superalgebra starting from a complex of binary
algebras (see [20]).
Definition 4.1. A Leibniz superalgebra is a pair (L, [., .]) consisting of a vector superspace L and
bilinear map [., .] : L × L → L satisfying
hx, [y, z]i =h[x, y], z(cid:3) + (−1)xy(cid:2)y, [x, z]i.
(4.1)
Let (L, [., .]) be a Leibniz superalgebra and W be an arbitrary vector superspace. Let [., .]W :
L × W −→ W be a even bilinear map satisfiying
h[x, y], w(cid:3)W =hx, [y, w]WiW
The pair (W, [., .]W is called an L-module.
− (−1)xy(cid:2)y, [x, v]wiW
.
Let (N , [., . . . , .]) be a n-ary-Nambu-Lie superalgebra and (V, [., . . . , .]V ) be a N -module.
We denote by L(N ) the space ∧n−1N and we call it the fundamental set.
We define a bilinear map [., .]L : L(N ) × L(N ) → L(N ) and ad : N × V −→ V respectivly by
[x, y]L =
and
(−1)x(y1+···+yi−1)y1 ∧ · · · ∧ ad(x)(yi) ∧ · · · ∧ yn−1
n−1Xi=1
for all x = x1 ∧ · · · ∧ xn−1, y = y1 ∧ · · · ∧ yn−1 ∈ L(N ), v ∈ V.
ad(x, v) = [x, v]V
(4.2)
(4.3)
Lemma 4.2. Let (N , [., . . . , .]) be a n-ary-Nambu-Lie superalgebra and be V be a N -module. The
map ad satisfies
ad([x, y]L)(v) = ad(x)(ad(y)(v)) − (−1)xyad(y)(ad(x)(v))
(4.4)
for all x, y ∈ L(N ), v ∈ V .
4
Proof. By (4.3) and (3.2) we have
ad([x, y]L)(v) = (cid:2)[x, y]L, v(cid:3)V
=
(−1)x(y1+···+yi−1)[y1 ∧ · · · ∧ ad(x)(yi) ∧ · · · ∧ yn−1, v]V .
n−1Xi=1
= (cid:2)x1, . . . , xn−1, [y1, . . . , yn−1, v](cid:3)V − (−1)x(y1+···+yn−1)(cid:2)y1, . . . , yn−1, [x1, . . . , xn−1, v](cid:3)V
= ad(x)(ad(y)(v)) − (−1)xyad(y)(ad(x)(v)).
Proposition 4.3. The pair (L(N ), [., .]L) is a Leibniz superalgebra.
Let (V, [., . . . , .]V ) be a N -module. We denote
W = L(N , V ) = {x1 ∧ · · · ∧ xn−2 ∧ v, xi ∈ N , v ∈ V }.
Note that
u1 ∧ . . . uj−1 ∧ uj ∧ · · · ∧ un−1 = −(−1)uj−1uju1 ∧ . . . ui ∧ ui−1 ∧ · · · ∧ un−1,
(4.5)
for all homogenous element u = u1 ∧ · · · ∧ un−1 of W .
Define a bilinear map [., .]W : L(N ) −→ W by
[x, y1 ∧ · · · ∧ y1 ∧ v]W =
(−1)x(y1+···+yi−1)y1 ∧ · · · ∧ ad(x)(yi) ∧ · · · ∧ yn−1 ∧ v
n−1Xi=1
+(−1)x(y1+···+yn−1)y1 ∧ · · · ∧ yn−1 ∧ ad(x)(v).
Proposition 4.4. The pair (W, [., .]W ) is a L-module.
In the following, the expression [x, y] means:
• [x, y]L if x, y ∈ L(N ).
• [x1, . . . , xn−1, y] if x = x1 ∧ · · · ∧ xn−1 ∈ L(N ), y ∈ N .
Definition 4.5. We call k-cochain of a n-ary super-algebra N with values in V a multilinear map
Denote C k(N , V ) the set of k-cochains on N with values in V .
ϕ : L(N )k × N −→ V.
Theorem 4.6. Let (N , [., · · · , .]) be a n-ary-Nambu-Lie superalgebra and C k(L(N ), V ) the set of
k-cochain.
5
We define a coboundary operator dk : C k(L(N ), W ) → C k+1(L(N ), W ) by df (x) = −[x, f ] when
f ∈ C 0(L(N ), V ) = V and for k ≥ 1,
dk(f )(x0, . . . , xk) = − X0≤s<t≤k
(−1)s+xs(xs+1+···+xt−1)f (x0, . . . , bxs, . . . , xt−1, [xs, xt], xt+1, . . . , xk, z)
(−1)s+xs(f +x0+···+xs−1)hxs, f (x0, . . . , bxs, . . . , xk)iW
k−1Xs=0
+(−1)khf (x0, . . . , xk−1), xki′
+
W
,
where
hx1 ∧ · · · ∧ xn−2 ∧ v, y1 ∧ · · · ∧ yn−1i′
n−1Xi=1
W
= −
(−1)(x1+···+xn−2+v)(y1+···+yi−1)+vyiy1 ∧ · · · ∧ ad(x1 ∧ · · · ∧ xn−2 ∧ yi)(v) ∧ . . . yn−1
Let ∆k : C k−1(N , V ) → C k(L(N ), W ) be the linear map defined for k = 0 by
∆(f )(x0) =
and for k > 0 by
∆k(f )(x0, . . . , xk)
(−1)f (x1
0+···+xi−1
0
n−1Xi=1
)x1
0 ∧ · · · ∧ f (xi
0) ∧ · · · ∧ xn−1
0
=
(−1)(f +x0+···+xk−1)(x1
k+···+xi−1
k
)x1
k ∧ · · · ∧ f (x0, . . . , xk−1, xi
k) ∧ xi+1
k ∧ . . . xn−1
k
n−1Xi=1
where we set xj = x1
n-ary-Nambu-Lie superalgebra such that
j ∧ · · · ∧ xn−1
j
. Then there exists a cohomology complex (C k(N , N ), δ) for
The cobondary map δk+1 : Ck(N , V ) → Ck+1(N , V ) is defined by
dk ◦ ∆k−1 = ∆k−1 ◦ δk−1
δk+1(f )(x0, . . . , xk, z)
−
= − X0≤s<t≤k
(−1)s+xs(xs+1+···+xt−1)f (x0, . . . ,bxs, . . . , xt−1, [xs, xt], xt+1, . . . , xk, z)
(−1)s+xs(xs+1+···+xk)f(cid:16)x0, . . . , xs−1, xs+1, . . . , xk, ad(xs)(z)(cid:17)
kXs=0
(−1)s+xs(f +x0+···+xs−1)hxs, f (x0, . . . , bxs, . . . , xk, z)iV
kXs=0
n−1Xi=1
"z ∧ x1
k(cid:17)#V
, f(cid:16)x0, . . . , xk−1, xi
(−1)k−i+(f +x0+···+xk−1+xi+1
k ∧ · · · ∧ xn−1
+···+xn−1
)(xi
+
+
k
.
.
k
k
k ∧ . . .cxi
k+xk)+z(f +x0+···+xk)+xk(xi+1
k
+···+xn−1
k
)
6
Proof. For anyf ∈ C k−2(N , V ), by calculation we obtain dk+1 ◦ dk(f )(x0, . . . , xk) = 0 and
dk ◦ ∆k−1(f )(x0, . . . , xk) = ∆k−1 ◦ δk−1(f )(x0, . . . , xk).
One has ∆k+1◦δk = dk◦∆k, then ∆k+1◦δk◦δk−1 = dk◦dk−1◦∆k−1 = 0, because dk◦dk−1 = 0
Definition 4.7.
• The k-cocycles space is defined as Z k(N , V ) = ker δk. The even (resp. odd)
k-cocycles space is defined as Z k(N , V )0 = Z k(N , V ) ∩ (C k(N , V ))0 (resp. Z k(N , V )1 =
Z k(N , V ) ∩ (C k(N , V ))1).
• The k-coboundaries space is defined as Bk(N , V ) = Im δk−1. The even (resp. odd) k-
1 (N , V ) = Bk(N , V ) ∩
0 (N , V ) = Bk(N , V ) ∩ (C k(N , V ))0 (resp. Bk
coboundaries space is Bk
(C k(N , V ))0).
• The kth cohomology space is the quotient H k(N , V ) = Z k(N , V )/Bk(N , V ). It decomposes
as well as even and odd kth cohomology spaces.
Finally, we denote by H k(N , V ) = H k
0 (N , V ) ⊕ H k
1 (N , V ) the kth cohomology space and by
⊕k≥0H k(N , V ) the r-cohomology group of the Hom-Lie superalgebra N with values in V .
Remark 4.8. The subspace Z 1(N , N ) is the set of derivation of N .
5 Extensions of n-ary-Nambu-Lie superalgebra
An extension theory of Hom-Lie superalgebras was stated in [4].
An extension of a n-ary-Nambu-Lie superalgebra (N , [., . . . , .]) by N -module (V, [., . . . , .]V ) is
an exact sequence
0 −→ (V, [., . . . , .]V )
i−→ (eN , ^[., . . . , .])
We say that the extension is central if [L(eN ), i(V )] eN = 0.
Two extensions
π−→ (N , [., . . . , .]) −→ 0.
0 −→ (V, [., . . . , .]V )
πk−→ (N , [., . . . , .]) −→ 0 (k = 1, 2)
ik−→ (fNk, ^[., . . . , .])
are equivalent if there is an isomorpism ϕ : (N1, [., . . . , .]1) −→ (N2, [., . . . , .]2) such that ϕo i1 = i2
and π2 o ϕ = π1.
Proposition 5.1. Let (N , [., · · · , .]) be a n-ary-Nambu-Lie superalgebra and V be a N -module. The
second cohomology space H 2(N , V ) = Z 2(N , V )/B2(G, V ) is in one-to-one correspondence with the
set of the equivalence classes of central extensions of (N , [., . . . , .]) by (V, [., . . . , .]V ).
be a central extension of n-ary-Nambu-Lie superalgebra (N , [., . . . , .]) by (V, [., . . . , .]V ), so there is
i−→ (eN , ^[., . . . , .])
π−→ (N , [., . . . , .]) −→ 0.
Proof. Let
0 −→ (V, [., . . . , .]V )
a space H such that eN = H ⊕ i(V ).
7
The map π/H : H → N (resp k : V → i(V )) defined by π/H(x) = π(x) (resp. k(v) = i(v)) is
so [s(x), s(z)] eN − s([x, z]) ∈ i(V ).
We set [s(x), s(z)] − s([x, z]) = G(x, z) ∈ i(V ) then F (x, z) = l ◦ G(x, z) ∈ V, it easy to see that
s(x) + i(v), it is easy to verify that ϕ is a bijective.
For all x = x1 ∧ . . . xn−1 ∈ L(N ) (resp. v = v1 ∧ . . . vn−1 ∈ L(V )) we denote s(x1) ∧ . . . s(xn−1)
(resp. i(v1) ∧ · · · ∧ i(vn−1)) by s(x). (resp. i(v)).
bijective, its inverse s (resp. l) note. Considering the map ϕ : N × V → eN defined by ϕ(x, v) =
Since π is homomorphism of n-ary-Nambu-Lie superalgebra then π(cid:16)[s(x), s(z)] eN − s([x, z])(cid:17) = 0
F ∈ C 2(N , V ) is a 2-cochain that defines a bracket on eN . In fact, we can identify as a superspace
N × V and eN by ϕ : (x, v) → s(x) + i(v) where the bracket is
Viewed as elements of N × V we have h(x, v), (z, w)i = (cid:16)[x, z], F (x, z)(cid:17) and the homogeneous
elements (x, v) of N × V are such that x = v and we have in this case (x, v) = x.
We deduce that for every central extension
[s(x) + i(v), s(z) + i(w)] eN = [s(x), s(z)] eN = s([x, z]) + F (x, z).
0 −→ (V, [., . . . , .]V )
π−→ (N , [., . . . , .]) −→ 0.
One may associate a two cocycle F ∈ Z 2(N , V ). Indeed, for x ∈ L(N ), z ∈ N , if we set
i−→ (eN , ^[., . . . , .])
F (x, z) = l(cid:16)[s(x), s(z)] − s([x, z])(cid:17) ∈ V,
then, we have F (x, z) ∈ V and F satisfies the 2-cocycle conditions.
Conversely, for each f ∈ Z 2(N , V ), one can define a central extension
0 −→ (V, [., . . . , .]V ) −→ (Nf , [., . . . , .]f ) −→ (N , [., . . . , .]) −→ 0,
by
h(x, v), (y, w)if
=(cid:16)[x, y], f (x, y)(cid:17),
where x ∈ L(N ), z ∈ N and v ∈ L(V ), w ∈ V.
Let f and g be two elements of Z 2(N , V ) such that f − g ∈ B2(G, V ) i.e. (f − g)(x, z) = h([x, z]),
where h : N → V is a linear map . Now we prove that the extensions defined by f and g are
equivalent. Let us define Φ : Nf × V → Ng × V by
Φ(x, v) = (x, v − h(x)).
It is clear that Φ is bijective. Let us check that Φ is a homomorphism of n-ary-Nambu-Lie super-
algebra . We have
hΦ((x, v)), Φ((z, w))ig
= h(x, v − h(x)), (z, w − h(z))ig
= (cid:16)[x, z], g(x, z)(cid:17)
= (cid:16)[x, z], f (x, z) − h([x, z])(cid:17)
= Φ(cid:16)([x, z], f (x, z))(cid:17)
= Φ(cid:16)[(x, v), (z, w)]f(cid:17).
8
Next, we show that for f, g ∈ Z 2(N , V ) such that the central extensions
0 → (V, [., . . . , .]V ) → (Nf , ^[., . . . , .]f ) → (N , [., . . . , .]) → 0, and
0 → (V, [., . . . , .]V ) → (Ng, ^[., . . . , .]g) → (N , [., . . . , .]) → 0, are equivalent, we have f − g ∈
B2(N , V ). Let Φ be a homomorphism of n-ary-Nambu-Lie superalgebra . such that
0
0
/ (V, [., . . . , .]V )
i1
/ (Nf , ^[., . . . , .])
π1
/ (N , [., . . . , .])
0
idV
Φ
idN
/ (V, [., . . . , .]V )
i2
/ (Ng, ^[., . . . , .])
π2
/ (N , [., . . . , .])
/ 0
commutes. We can express Φ(x, v) = (x, v − h(x)) for some linear map h : N → V. Then we have
Φ([(x, v), (z, w)]f ) = Φ(([x, z], f (x, z)))
= ([x, z], f (x, z) − h([x, z])),
[Φ((x, v)), Φ((z, w))]g = [(x, v − h(x)), (z, w − h(y))]g
= ([x, z], g(x, z)),
and thus (f − g)(x, y) = h([x, z]) (i.e. f − g ∈ B2(N , V )), so we have completed the proof.
5.1 Deformation of n-ary-Nambu-Lie superalgebra.
Definition 5.2. Let (N , [., .]) be a n-ary-Nambu-Lie superalgebra. A one -parameter formal Lie
super deformation of N is given by the K[[t]]-multilinear map [., . . . , .]t : N n−1[[t]]×N [[t]] −→ N [[t]]
of the form
[., . . . , .]t =Xi≥0
ti[., . . . , .]i,
where each [., . . . , .]i is a even multiilinear map [., . . . , .]i
K[[t]]-multilinear), [., . . . , .] = [., . . . , .]0 and satisfying the following conditions
: N n−1 × N −→ N (extended to be
[x1, . . . , xi, . . . , xj, . . . , xn]t = (−1)j−i+xi(xi+1+···+xj−1)[x1, . . . , xj, . . . , xi, . . . , xn]t, (5.1)
[x, [y, z]t]t − [[x, y]L, z]t − (−1)yx[y, [x, z]t]t = 0
(5.2)
for all homogeneous elements x, y ∈ N n−1 and z ∈ N . The super deformation is said to be of
order k if [., . . . , .]t =
ti[., . . . , .]i.
kXi=0
t =Xi≥0
Given two deformation Nt = (N , [., . . . , .]t) and N ′
t = (N ′, [., . . . , .]′
ti[., . . . , .]i and [., . . . , .]′
ti[., . . . , .]′
i with [., . . . , .]0 = [., .]′
Xi≥0
N ′
t are equivalent if there exists a formal automorphism Φt =
Φ0 = idN , such that
Φt([x, z]t) = [Φt(x), Φt(z)]′
t.
9
t) of N where [., . . . , .]t =
0 = [., . . . , .]. We say that Nt and
Φiti where Φi ∈ End(N ) and
kXi=0
/
/
/
/
/
/
/
/
/
A deformation Nt is said to be trivial if and only if Nt is equivalent to N ( viewd as superalgebra
on N [[t]].)
The identity (5.2) is called a deformation equation and it is equivalent to
i.e
or
Xi≥0,j≥0(cid:18)[x, [y, z]i]j − [[x, y]L, z]t − (−1)yx[y, [x, z]i]j(cid:19)ti+j = 0;
Xi≥0,s≥0(cid:18)[x, [y, z]i]s−i − [[x, y]L, z]s−i − (−1)yx[y, [x, z]i]s−i(cid:19)ts = 0,
tsXi≥0(cid:18)[x, [y, z]i]s−i − [[x, y]L, z]s−i − (−1)yx[y, [x, z]i]s−i(cid:19) = 0.
Xs≥0
sXi=0(cid:18)[x, [y, z]i]s−i − [[x, y]L, z]s−i − (−1)yx[y, [x, z]i]s−i(cid:19) = 0.
The deformation equation is equivalent to the follwing infinite system
(5.3)
In particular, for s = 0 we have [x, [y, z]0]0 − [[x, y]L, z]0 − (−1)yx[y, [x, z]0]0 wich is the super
Jacobie identity of N .
The equation for s = 1, is equivalent to δ2
For s ≥ 2, the identities (5.3) are equivalent to:
0([., . . . , .]1) = 0. Then [., . . . , .]1 is a 2-cocycle.
δ2([., . . . , .]s)(x, y, z) = −
One may also prove
s−1Xi=1(cid:18)[x, [y, z]i]s−i − [[x, y]L, z]s−i − (−1)yx, [y, [x, z]i]s−i(cid:19).
Theorem 5.3. Let (N , [., . . . , .]) be a n-ary-Nambu-Lie superalgebra and Nt = (N , [., . . . , .]t) be
ti[., . . . , .]i. Then there ex-
a one-parameter formal deformation of N , where [., . . . , .]t = Xi≥0
ists an equivalent deformation N ′
t = (N , [., . . . , .]′
t) where [., . . . , .]′
ti[., . . . , .]′
i such that
t = Xi≥0
[., . . . , .]′
i ∈ Z 2(N , N ) and doesn't belong to B2(N , N ).
Hence, if H 2(N , N ) = 0 then every formal deformation is equivalent to a trivial deformation.
The n-ary-Nambu-Lie superalgebra is called rigid.
6 Cohomology of the super w∞ 3-algebra
The generators of CHOVW algebra are given by
m = (−1)iλi− 1
Li
L
i
m = (−1)iλi+ 3
α+ 1
h
2
r
α+ 1
h
2
r
= (−1)α+1λα+ 1
= (−1)α+1λα+ 1
10
2 zn+iθ
2 zn+i ∂i
∂zi ,
∂i
∂
∂zi ,
∂θ
∂α
2 zr+α ∂
∂zα ,
∂θ
∂α
∂zα .
2 zr+αθ
The commutation relation is defined by
[a, b] = ab − (−1)abba.
Let us define a super 3-bracket as follows:
[a, b, c] = [a, b]c + (−1)a(b+c)[b, c]a + (−1)c(b+c).
(6.1)
(6.2)
Using (6.1), (6.2) and taking the scaling limit λ → 0, then we obtain the following super w∞-algebra.
n, Lh
h
n, L
[Li
[Li
[Li
m, Lj
m, Lj
m, Lj
m, Lj
[Li
[Li
n, h
α+ 1
m, h
2
p
α+ 1
n, h
2
p
m+n+k ,
i+j+h−1
m+n+k ,
k] =(cid:0)h(n − m) + j(m − k) + i(k − n)(cid:1)Li+j+h−1
k] =(cid:0)h(n − m) + j(m − k) + i(k − n)(cid:1)L
] =(cid:0)α(n − m) + j(m − p) + i(p − n)(cid:1)h
] =(cid:0)α(n − m) + j(m − r) + i(r − n)(cid:1)h
] =(cid:0)i(p − r) + α(r − m) + β(m − p)(cid:1)L
β+ 1
2
r
α+ 1
2
r
, h
i+j+α−1+ 1
2
m+n+p
i+j+α−1+ 1
2
m+n+r
,
,
i+α+β−1
m+r+p
.
The other brakets are obtained by supersymmetry or equals 0.
This algebra is Z2 graded with
w∞ = (w∞)0 ⊕ (w∞)1 where (w∞)0 = Mn∈Z,i∈N
< Li
n, L
i
n > and (w∞)1 = Mn∈Z,i∈N
i+ 1
< h
2
n
, h
i+ 1
n > .
2
In the following, we describe a super w∞ 3-algebra obtained in [1] and we compute its derivations
and second cohomology group.
6.1 Derivations of the super w∞ 3-algebra.
An even derivation D (resp. odd) is said of degree (s, t) if there exists (s, t) ∈ Z × N such that,
i+ 1
2
i+ 1
m > ⊕ < h
2
for all (m, i) ∈ Z × N, we have D(cid:0) < Li
m >(cid:1) ⊂(cid:0) < h
D(cid:0) < h
(cid:0) < h
m+s >(cid:1) and D(cid:0) < h
It easy to check that Der(w∞) = M(s,t)∈Z×N(cid:0)Der(w∞)(s,t)
m > ⊕ < L
i+t+ 1
m+s > ⊕ < h
2
m > (cid:1) ⊂ (cid:0) < Li+t
m+s >(cid:1) (resp. D(cid:0) < Li
m >(cid:1) ⊂(cid:0) < Li+t
(cid:1).
0 ⊕ Der(w∞)(s,t)
i+t+ 1
m+s > ⊕ < h
2
i+ 1
m > ⊕ < h
2
i+t+ 1
2
i+t+ 1
2
i+ 1
2
Let f be a homogeneous derivation
1
i
m+s > ⊕ < L
m > ⊕ < L
m+s > ⊕ < L
i
i+t
m+s > (cid:1) and
m >(cid:1) ⊂
m+s >(cid:1)).
i+t
f ([x1, x2, x3]) = [f (x1), x2, x3] + (−1)f x1[x1, f (x2), x3] + (−1)f (x1+x2)[x1, f (x2), f (x3)].
We deduce that
(cid:16)h(n − m) + j(m − k) + i(k − n)(cid:17)f (Li+j+h−1
m+n+k ) = hf (Li
ki +hLi
m), Lj
n, Lh
m, f (Lj
n), Lh
m, Lj
n, f (Lh
ki +hLi
k)i,
(6.3)
11
m), Lj
n, L
h
ki +hLi
m, f (Lj
n), L
h
ki +hLi
m, Lj
n, f (L
h
k)i
(6.4)
(cid:16)h(n − m) + j(m − k) + i(k − n)(cid:17)f (L
(cid:16)α(n − m) + j(m − r) + i(r − n)(cid:17)f (h
(cid:16)α(n − m) + j(m − r) + i(r − n)(cid:17)f (h
(cid:16)α(m − n) + j(r − m) + i(n − r)(cid:17)f (L
i+j+h−1
i+j+α−1+ 1
2
m+n+r
m+n+k ) = hf (Li
) = hf (Li
+hLi
) = hf (Li
+hLi
m+n+r ) = hf (Li
i+j+α−1+ 1
2
m+n+r
i+j+α−1
m), Lj
α+ 1
n, h
2
r
m, Lj
α+ 1
n, f (h
2
r
m), Lj
α+ 1
n, h
2
r
m, Lj
α+ 1
n, f (h
2
r
i +hLi
)i
i +hLi
)i
i +hLi
α+ 1
2
r
m, f (Lj
α+ 1
n), h
2
r
m, f (Lj
α+ 1
n), h
2
r
i
i
(6.5)
(6.6)
i
(6.7)
j+ 1
m), h
2
n
, h
α+ 1
2
r
j+ 1
m, f (h
2
n
), h
α+ 1
2
r
+(−1)f hLi
j+ 1
m, h
2
n
, f (h
)i
6.1.1
Even derivations of the super w∞ 3-algebra
Let f be an even derivation of degree (s, t),
f (Li,t
m,s) = ai,t
m,sLi+t
m+s + bi,t
m,sL
i+t
m+s,
i
m) = ci,t
f (L
m,sLi+t
m+s + di,t
m,sL
i+t
m+s,
α+ 1
f (h
2
r
) = eα,t
r,s h
α+t+ 1
r+s + f α,t
r,s h
2
α+ 1
2
r
and
By (6.3) we have
β+ 1
f (h
2
r
) = xβ,t
r,s h
β+t+ 1
r+s + yβ,t
r,s h
2
β+t+ 1
2
r+s
.
and
= ai,t
+aj,t
+ah,t
ai+j+h−1,t
m+n+k,s (cid:16)h(n − m) + j(m − k) + i(k − n)(cid:17)
m,s(cid:16)h(n − m) + j(m − k) + i(k − n) − hs + js + t(k − n)(cid:17)
n,s(cid:16)h(n − m) + (j)(m − k) + i(k − n) + hs + t(m − k) − is(cid:17)
k,s(cid:16)h(n − m) + j(m − k) + i(k − n) + t(n − m) − js + is(cid:17),
m+n+k,s(cid:16)h(n − m) + j(m − k) + i(k − n)(cid:17)
m,s(cid:16)h(n − m) + j(m − k) + i(k − n) + t(k − n) − hs + js(cid:17)
n,s(cid:16)h(n − m) + j(m − k) + i(k − n) + t(m − k) + hs − is(cid:17)
k,s(cid:16)h(n − m) + j(m − k) + i(k − n) + t(n − m) − js + is(cid:17).
bi+j+h,t
+bh,t
+bj,t
= bi,t
12
(6.8)
(6.9)
By (6.4), we have
(cid:16)h(n − m) + j(m − k) + i(k − n)(cid:17)ci+j+h−1,t
m+n+k,s = (cid:16)h(n − m) + j(m − k) + i(k − n) + t(n − m) − js + is(cid:17)ch,t
k,s
(6.10)
and
m+n+k,s
(cid:16)h(n − m) + j(m − k) + i(k − n)(cid:17)di+j+h−1,t
= (cid:16)h(n − m) + j(m − k) + i(k − n) − hs + js + t(k − n)(cid:17)ai,t
+(cid:16)h(n − m) + j(m − k) + i(k − n) + hs + t(m − k) − is(cid:17)aj,t
+(cid:16)h(n − m) + j(m − k) + i(k − n) + t(n − m) − js + is(cid:17)dh,t
m,s
n,s
k,s.
By (6.5) and (6.6), we obtain, exactly, the same equation as (6.11).
By (6.7), we obtain
m+n+r,s
(cid:16)α(m − n) + j(r − m) + i(n − r)(cid:17)di+j+h−1,t
= (cid:16)α(m − n) + j(r − m) + i(n − r) − js + αs + t(n − r)(cid:17)ai,t
+(cid:16)α(m − n) + j(r − m) + i(n − r) − αs + t(r − m) + is(cid:17)ej,t
+(cid:16)α(m − n) + j(r − m) + i(n − r) + t(m − n) + js − is(cid:17)yα,t
(cid:16)α(m − n) + j(r − m) + i(n − r)(cid:17)ci+j+α−1,t
m+n+r,s = 0.
m,s
n,s
r,s ,
and
(6.11)
(6.12)
(6.13)
Taking m = n = i = 0, j = 1 (resp. m = 1, n = −1, r = 0, i = 1, j = 0), we obtain
cα,t
r,s = 0, ∀(r, s) ∈ Z × N.
Proposition 6.1.
• If s + 2t 6= 0 we have
Der(w∞)(s,t)
0
= < ad(Lt
1+s, L1
−1) > ⊕ < ad(L0
1, L1+t
−1+s) > ⊕ < ad(L
• If s + 2t = 0 and t 6= 0 we have
Der(w∞)(s,t)
0
= < ad(L1+t
1+s, L0
−1) > ⊕ < ad(L1
1, Lt
−1+s) > ⊕ < ad(L
• If s + 2t = 0 and t = 0 we have
t
1+s, L1
−1) > ⊕ < ad(L0
1, L
1+t
1+s, L0
−1) > ⊕ < ad(L1
1, L
1+t
−1+s) >,
t
−1+s) > .
= < ad(L1
−1, L0
1) > ⊕ < ad(L1
0, L0
1+ 1
0) > ⊕ < ad(h
2
0
, h
1
2
0+ 1
0 ) > ⊕ < ad(h
2
1
, h
1+ 1
−1 ) >
2
⊕ < ϕ1 > ⊕ < ϕ2 >,
Der(w∞)(0,0)
0
where
ϕ1(Lh
ϕ2(Lh
h+ 1
k) = ϕ1(h
2
k
α+ 1
2
k
k) = ϕ2(h
) = 0, ϕ1(L
) = 0, ϕ2(L
h
k) = L
h
k) = L
α+ 1
h
k, ϕ1(h
2
k
α+ 1
h
k, ϕ2(h
2
k
α+ 1
) = h
2
k
α+ 1
) = h
2
k
.
.
13
Proof. Taking m = 1, n = −1, i = 0, j = 1 in (6.3) we obtain
(−2h + 1 − k)f (Lh
k) = a0,t
1+s, L1
k) + b0,t
t
−1)(Lh
1+s, L1
1,sad(Lt
k )
1+t
k ) + b1,t
+a1,t
−1+s)(Lh
−1,sad(L0
k)
1, L
+(−2h − 2t + 1 − k − s)f (Lh
k),
−1)(Lh
−1+s)(Lh
−1,sad(L0
1, L1+t
1,sad(L
which gives ( if s + 2t 6= 0)
f (X h
k ) =
ad(Lt
1+s, L1
−1)(X h
k ) +
ad(X
t
1+s, L1
−1)(X h
k )
ad(L0
1, L1+t
−1+s)(X h
k ) +
ad(L0
1, L
1+t
−1+s)(X h
k ) for any X h
k ∈ w∞
a0,t
1,s
s + 2t
a1,t
−1,s
s + 2t
+
b0,t
1,s
s + 2t
b1,t
−1,s
s + 2t
Taking m = 1, n = −1, i = 1, j = 0 in (6.3) we obtain
(−2h + 1 + k)f (X h
k ) = a1,t
k ) + b1,t
1+s, L0
1, Lt
1+t
1,sad(L1+t
−1)(X h
1+s, L0
k )
k ) + b0,t
+a0,t
−1,sad(L1
−1,sad(L1
1, L
+(−2h − 2t + 1 + k + s)f (X h
k ),
−1)(X h
−1+s)(X h
1,sad(L
t
−1+s)(X h
k )
which gives
f (X h
k ) =
a1,t
1,s
4t
a0,t
−1,s
4t
+
ad(L1+t
1+s, L0
−1)(X h
k ) +
b0,t
1,s
4t
ad(L
1+t
1+s, L0
−1)(X h
k )
ad(L1
1, Lt
−1+s)(X h
k ) +
b1,t
−1,s
4t
ad(L1
1, L
t
−1+s)(X h
k ) if s + 2t = 0 and t 6= 0.
Taking s = 0, t = 0, m = 0, n = 0, i = 2, j = 0 in (6.8), we obtain
ah,0
k,0 = ha2,0
0,0 + ha0,0
0,0 + a0,0
k,0, ∀k ∈ Z∗.
(6.14)
Taking s = 0, t = 0, h = 0, m = 1, i = 0, n = 0, j = 1 in (6.8), we obtain
(1 − k)a0,0
k+1,0 = (1 − k)a0,0
1,0 + (1 − k)a1,0
0,0 + (1 − k)a0,0
k,0.
We deduce that
a0,0
k,0 = (k − 1)a0,0
1,0 + (k − 1)a1,0
0,0 + a0,0
1,0 ∀k ∈ Z∗.
Then ah,0
k,0 = ha2,0
0,0 + ha0,0
0,0 + (k − 1)a0,0
1,0 + (k − 1)a1,0
0,0 + a0,0
1,0. So,
a2,0
1,0 = 2a2,0
0,0 + 2a0,0
0,0 + a0,0
1,0 = 2a2,0
0,0 − 2a1,0
0,0 + a0,0
1,0
Taking m = 1, n = 0, k = 0, i = 0, j = 1, h = 2 in (6.8), we obtain
a2,0
1,0 = a1,0
0,0 + a0,0
1,0 + a2,0
0,0
14
So
We deduce that
Then
0,0 = 3a1,0
a2,0
0,0.
ah,0
k,0 = (2h + k − 1)a1,0
0,0 + ka0,0
1,0.
(6.15)
bh,0
k,0 = (2h + k − 1)b1,0
0,0 + kb0,0
1,0.
Taking s = 0, t = 0, m = 1, n = −1, i = j = 1 in (6.12), we obtain
dh,0
k,0 = (h − 2)(a1,0
1,0 + a1,0
−1,0) + d2,0
k,0
1,0 + a1,0
Since a1,0
Taking s = 0, t = 0, m = 1, n = −0, i = 0, j = 1, h = 2 in (6.12) we obtain
1,0 we deduce that dh,0
k,0 = 2(h − 2)a1,0
−1,0 = a0,0
0,0 + d2,0
k,0.
−(k + 1)d2,0
k+1,0 = −(k + 1)a0,0
1,0 − (k + 1)a1,0
0,0 − (k + 1)d3,0
k,0.
One can deduce that
Then,
and
dh,0
k,0 = (2h + k − 1)a1,0
0,0 + ka0,0
1,0 + d2,0
0,0 − 3a1,0
0,0, ∀k ∈ Z \ {−1}.
eh,0
k,0 = (2h + k − 1)a1,0
0,0 + ka0,0
1,0 + e2,0
0,0 − 3a1,0
0,0, ∀k ∈ Z \ {−1},
r,0 = (2α + r − 1)a1,0
yα,0
0,0 + ra0,0
1,0 + y2,0
0,0 − 3a1,0
0,0, ∀r ∈ Z \ {−1}.
(6.16)
(6.17)
(6.18)
Taking m = 1, n = −1, i = j = 1 in (6.5), we obtain xα,0
So, f α,0
Using (6.15), (6.16), (6.17) and (6.18) in (6.12), we obtain d2,0
r = 0, ∀α ∈ N∗ \ {1}.
r,0 = f 2
r,0 = x2
k = 0, ∀h ∈ N∗ \ {1}.
0,0 = −3a1,0
0,0 + e2,0
0,0 + y2,0
0,0.
Therefore,
f = a1,0
0,0ad(L1
−1, L0
1) + a0,0
1,0ad(L1
0, L0
0) + b1,0
0,0ad(h0
1, h
1
−1) − b0,0
1,0ad(h1
0, h
0
0) + (d2,0
0,0 − 3a1,0
0,0)ϕ1 + (e2,0
0,0 − 3a1,0
0,0)ϕ2,
where,
ϕ1(Lh
ϕ2(Lh
h+ 1
k ) = ϕ1(h
2
k
α+ 1
k ) = ϕ2(h
2
k
) = 0, ϕ1(L
) = 0, ϕ2(L
α+ 1
h
h
k) = L
k, ϕ1(h
2
k
α+ 1
h
h
k) = L
k, ϕ2(h
2
k
α+ 1
) = h
2
k
α+ 1
) = h
2
k
,
.
15
6.2 Odd derivation
Let f be an odd derivation of degree (s, t),
f (Li,t
m,s) = ai,t
i
m) = ci,t
f (L
) = eα,t
α+ 1
f (h
2
r
β+ 1
f (h
2
r
) = xβ,t
i+t 1
m+s + bi,t
2
m+s + di,t
r+s + f α,t
r+s + yβ,t
m,sh
m,shi+t
r,s lα+t
r,s lβ+t
r,s L
r,s L
m,sh
m,sh
α+t
r+s
β+t
r+s
i+t+ 1
m+s ,
2
i+t
m+s,
(6.19)
Proposition 6.2.
• If s + 2t 6= 0, then we have
.
Der(w∞)(s,t)
1
t+ 1
1+s , L1
2
= < ad(h
−1) > ⊕ < ad(h
• If s + 2t = 0 and t 6= 0 we have
t+ 1
1+s , L1
2
−1) > ⊕ < ad(L0
1, h
1+t+ 1
−1+s ) > ⊕ < ad(L0
2
1, h
1+t+ 1
−1+s ) > .
2
Der(w∞)(s,t)
1+t+ 1
2
1+s
1
= < ad(h
, L0
−1) > ⊕ < ad(h
1+t+ 1
2
1+s
, L0
−1) > ⊕ < ad(L1
1, h
t+ 1
−1+s) > ⊕ < ad(L1
2
1, h
t+ 1
−1+s) > .
2
• If s + 2t = 0 and t = 0
Der(w∞)(0,0)
1
= < ad(L0
1, h
1+ 1
−1 ) > ⊕ < ad(L0
2
1, h
1+ 1
−1 ) > ⊕ < ad(L0
2
1+ 1
0, h
2
0
) > ⊕ < ad(L0
0, h
1+ 1
2
0
) >
Proof. Setting m = 1, n = −1, i = 0, j = 1 in (6.3), one has
(−2h + 1 − k)f (Lh
k) = a0,t
k ) + b0,t
t+ 1
1+s , L1
2
1,sad(h
+a1,t
+(−2h − 2t + 1 − k − s)f (Lh
−1)(Lh
1+t+ 1
−1+s )(Lh
2
t+ 1
1+s , L1
1,sad(h
2
k ) + b1,t
−1,sad(L0
k),
−1,sad(L0
1, h
−1)(Lh
k)
1+t+ 1
−1+s )(Lh
k)
2
1, h
which gives
f (Lh
k) =
a1,t
−1,s
s + 2t
+
a0,t
1,s
s + 2t
ad(h
t+ 1
1+s , L1
2
ad(L0
1, h
1+t+ 1
−1+s )(Lh
2
−1)(Lh
k) +
b1,t
−1,s
s + 2t
k ) +
b0,t
1,s
s + 2t
ad(h
t+ 1
1+s , L1
2
−1)(Lh
k )
ad(L0
1, h
1+t+ 1
−1+s )(Lh
2
k) if s + 2t 6= 0.
Furthermore , taking m = 1, n = −1, i = 1, j = 0 in (6.3), one has
f (Lh
k) =
ad(h
1+t+ 1
2
1+s
, L0
−1)(Lh
k) +
ad(h
1+t+ 1
2
1+s
, L0
−1)(Lh
k )
ad(L1
1, h
t+ 1
−1+s)(Lh
2
k ) if t 6= 0.
a1,t
1,s
4t
a0,t
−1,s
4t
+
b0,t
1,s
4t
b1,t
−1,s
4t
ad(L1
1, h
t+ 1
−1+s)(Lh
2
k ) +
16
By (6.3) and (6.19) we obtain, exactly, the same equation as (6.8) and (6.9).
By (6.4) and (6.19) we obtain, exactly, the same equation as ( 6.10) and (6.11).
By (6.5) and (6.19) we obtain (6.11) and (6.10).
By (6.6) and (6.19) we obtain (6.11) and (6.10).
Therefore,
ah,0
k,0 = (2h + k − 1)a1,0
0,0 + ka0,0
1,0
bh,0
k,0 = (2h + k − 1)b1,0
0,0 + kb0,0
1,0.
dh,0
k,0 = (2h + k − 1)a1,0
0,0 + ka0,0
1,0 + d2,0
0,0 − 3a1,0
0,0, ∀k ∈ Z \ {−1}.
We deduce that
f = a1,0
0,0ad(L0
1, h
1+ 1
−1 ) + a0,0
2
1,0ad(L0
1, h
1+ 1
−1 ) + b1,0
2
0,0ad(L0
1+ 1
0, h
2
0
) + b0,0
1,0ad(L0
0, h
1+ 1
2
0
)
6.3 Cohomology space H 2
0 (w∞, C) of w∞
In the following we describe the cohomology space H 2
class of an element f .
0 (w∞, C) . We denote by [f ] the cohomology
Theorem 6.3. The second cohomology of the super w∞ 3-algebra with values in the trivial module
vanishes, i.e
H 2
0 (w∞, C) = {0}.
Proof. For all f ∈ Z 1
0 (w∞, C) (resp. f ∈ Z 2
0 (w∞, C)) we have, respectively,
δ1(f )(x0, z) = f ([x1, x2, x3]) = 0,
(6.20)
δ2(f )(x0, x1, z) = −f(cid:16)[x0, x1], z(cid:17) − (−1)x0x1f(cid:16)x1, ad(x0)(z)(cid:17) + f(cid:16)x0, ad(x1)(z)(cid:17) = 0.
(6.21)
n, Lh
k), (Li
m, Lj
n, L
h
k), (Li
m, Lj
α+ 1
n, h
2
r
), (Li
m, Lj
n, h
α+ 1
2
r
α+ 1
), (h
2
r
, Li
m, h
β+ 1
2
s
)(cid:27).
k )(cid:17) = 0.
m, Lj
k ) ∈(cid:26)(Li
k(cid:17) − f(cid:16)X v
q], X h
Case 1 (X v
p , X l
q, X h
Using (6.21), we have
−f(cid:16)[Li
Since [x, y]l =
then
n−1Xi=1
m ∧ Lj
n, X v
p ∧ X l
p ∧ X l
q, ad(Li
m ∧ Lj
n)(X h
m ∧ Lj
n, ad(X v
p ∧ X l
q)(X h
k )(cid:17) + f(cid:16)Li
(−1)x(y1+···+yi−1)y1 ∧ · · · ∧ ad(x)(yi) ∧ · · · ∧ yn−1, and ad(a ∧ b)(z) = [a, b, z]
−f(cid:16)[Li
−f(cid:16)X v
m ∧ Lj
n, X v
p ] ∧ X l
q, X h
p , [Li
m ∧ Lj
n, X l
q], X h
p ∧ X l
q, [Li
m ∧ Lj
n, X h
m ∧ Lj
n, [X v
p ∧ X l
q, X h
k(cid:17) − f(cid:16)X v
k ](cid:17) + f(cid:16)Li
17
k(cid:17)
k ](cid:17) = 0.
So,
m+n+p ∧ X l
−(cid:18)v(n − m) + j(m − p) + i(p − n)(cid:19)f(cid:16)X i+j+v−1
−(cid:18)l(n − m) + j(m − q) + i(q − n)(cid:19)f(cid:16)X v
−(cid:18)h(n − m) + j(m − k) + i(k − n)(cid:19)f(cid:16)X v
+(cid:18)h(q − p) + l(p − k) + i(k − q)(cid:19)f(cid:16)Li
m ∧ Lj
p ∧ X l
q, X h
k(cid:17)
k(cid:17)
m+n+k (cid:17)
p+q+k (cid:17) = 0,
q, X i+j+h−1
p ∧ X i+j+l−1
m+n+q , X h
n, X v+l+h−1
where p is the integer part of p.
Setting m = 0, n = 0, i = 1, j = 0, in (6.22), we obtain
(6.22)
f (X v
p ∧ X l
q, X h
k(cid:17)) =
h(q − p) + l(p − k) + (k − q)
p + q + k
f (L1
0 ∧ L0
0, X v+t+h−1
p+q+k
) (p + q + k 6= 0).
Setting m = 1, n = −1, i = 1, j = 0, k = −p − q (6.22), we obtain
f (X v
p ∧ X l
q, X h
−p−q) = −
h(q − p) + l(2p − q) + (−p − 2q)
2l + 2v + 2h − 3
f (L1
1 ∧ L0
−1, X v+l+h−1
0
).
Setting m = 1, j = 0, i = 1, p = 1, q = −1, v = 1, l = 0, k = 0 (6.22), we obtain
(−2h − 1)f(cid:16)L1
1 ∧ L0
−1, X h
0(cid:17) = 0.
Then f (L1
1 ∧ L0
We deduce that
−1, X h
0 ) = 0.
We define an endomorphism g of w∞ by g(X h
By (6.20) we obtain
k f (L1
1 ∧ L0
−1, X h
k ) and g(X h
0 ) = 0.
f (X v
p ∧ X l
−p−q) = 0.
q, X h
k ) = − 1
δ1(g)(X v
p ∧ X l
q, X h
k ) = −g(ad(X v
p ∧ X l
q)(X h
k )) = f (Li
m ∧ Lj
n, X h
k ).
Case 2 (X v
p , X l
q, X h
Using (6.21), we have
), (Li
m, Lj
n, h
n, L
n, Lh
m, Lj
m, Lj
m, Lj
k), (Li
h
k), (Li
α+ 1
n, h
2
r
k ) /∈(cid:26)(Li
−(cid:16)v(n − m) + j(m − p) + i(p − n)(cid:17)f(cid:16)X i+j+v−1
−(cid:16)l(n − m) + j(m − q) + i(q − n)(cid:17)f(cid:16)X v
−(cid:16)h(n − m) + j(m − k) + i(k − n)(cid:17)f(cid:16)X v
+f(cid:16)Li
n, 0(cid:17) = 0.
m ∧ Lj
p ∧ X l
m+n+p ∧ X l
p ∧ X i+j+l−1
q, X h
k(cid:17)
k(cid:17)
m+n+k (cid:17)
m+n+q , X h
q, X i+j+h−1
α+ 1
2
r
α+ 1
), (h
2
r
, Li
β+ 1
m, h
2
s
)(cid:27).
(6.23)
Setting m = 0, n = 0, i = 1, j = 0, in (6.23), we obtain f (X v
Setting m = 1, n = −1, i = 1, j = 0, in (6.23), we obtain f (X v
p ∧ X l
q, X h
k ) = 0 (p + q + k 6= 0).
p ∧ X l
q, X h
k ) = 0 .
18
6.4 Second cohomology H 2(w∞, w∞).
In this section, we aim compute the second cohomology group of w∞ with values in itself.
Theorem 6.4. The second cohomology of the 3-ary-Nambu-Lie superalgebra w∞) with values in
the adjoint module vaniches, i.e.
H 2(w∞, w∞) = {0}.
Then, every formal deformation is equivalent to trivial deformation.
Proof. For all g ∈ Z 1
0 (w∞) (resp. f ∈ Z 2
0,s(w∞)) we have respectively
p ∧ Ll
p ∧ Ll
q, X h
k )
q, X h
δ1(g)(Lv
= −g(cid:16)(cid:2)Lv
k(cid:3)(cid:17) +hLv
p ∧ Ll
q, g(X h
k ∧ Ll
q, g(Lv
k ∧ Lv
p, g(Ll
k )i −hX h
p)i +hX h
q)i = 0.
m ∧ Lj
n, [X v
p ∧ X l
q, X h
m ∧ Lj
n, f (X v
p ∧ X l
q, X h
p ∧ X l
q, f (Li
m ∧ Lj
n, X h
p , [Li
m ∧ Lj
n, X l
q], X h
p ∧ X l
q, [Li
m ∧ Lj
n, X h
m ∧ Lj
m ∧ Lj
n, X v
n, X v
p ∧ X l
p ] ∧ X l
q, X h
k )
q, X h
δ2(f )(Li
= −f(cid:16)[Li
+f(cid:16)Li
+hX h
Case 1: s 6= 0.
k(cid:17) − f(cid:16)X v
k ](cid:17) +hLi
p )i −hX h
k ∧ X l
q, f (Li
m ∧ Lj
n, X v
k ∧ X v
p , f (Li
m ∧ Lj
n, X l
k(cid:17) − f(cid:16)X v
k )i −hX v
q)i = 0.
k ](cid:17)
k )i
Setting m = 0, n = 0, i = 0, j = 1, in (6.22), we obtain
q, X h
p ∧ X l
(cid:0)p + q + k(cid:1)f(cid:0)X v
k(cid:1)
+(cid:0)h(q − p) + l(p − k) + v(k − q)(cid:1)f(cid:0)L0
k )(cid:3) +(cid:2)X h
−(cid:2)X v
q, f (L0
p ∧ X l
0 ∧ L1
0, X h
0 ∧ L1
0, X v+l+h−1
0, X v
p+q+k
0 ∧ L1
k ∧ X l
q, f (L0
0 ∧ L1
0, f (X v
p ∧ X l
q, X h
k ∧ X v
p , f (L0
0 ∧ L1
0, X l
(cid:1) +(cid:2)L0
p )(cid:3) −(cid:2)X h
s f (L0
0 ∧ L1
0, X h
k ).
k )(cid:3)
q)(cid:3) = 0.
We denote by g the linear map defined on w∞ by g(X h
It easy to verify that δ1(g)(X v
p ∧ X l
Case 2: s = 0.
k ) = f (X v
p ∧ X l
q, X h
k ) = − 1
q, X h
k ).
Setting m = 1, n = −1, i = 1, j = 0 in (6.22), we obtain
p ∧ X l
q, X h
q, f (L1
k(cid:1) +(cid:0)h(q − p) + l(p − k) + v(k − q)(cid:1)f(cid:0)L1
q, f (L1
−1, X h
−1, X v
k ∧ X l
1 ∧ L0
1 ∧ L0
p ∧ X l
−2f(cid:0)X v
−(cid:2)X v
k )(cid:3) +(cid:2)X h
1 ∧ L0
−1, X v+l+h−1
p+q+k
k ∧ X v
p )(cid:3) −(cid:2)X h
(cid:1)
p , f (L1
1 ∧ L0
−1, X l
q)(cid:3) = 0
We denote by g the linear map defined on w∞ by g(X h
q, X h
It easy to verify that δ1(g)(X v
k ) = f (X v
p ∧ X l
p ∧ X l
q, X h
2 f (L1
k ) = − 1
−1, X h
k ), that ends the proof.
1 ∧ L0
k ).
19
References
[1] Min-Rue Chen,Ke Wua and Wei-Zhong Zhao:
Super w1 3-algebra, Sissa-Springer 090, no.
10, 1007 (2011).
[2] P. D. Beites and A. P. Pozhidaev: ON SIMPLE FILIPPOV SUPERALGEBRAS OF TYPE
A(m, n), arxiv.: 1008.0715v1, , 1007 (2010).
[3] Ruipu Bai and Ying Li: T ∗
θ . -Extensions of n-Lie Algebras, International Scholarly Research
Network ISRN Algebra: 10.5402,381875 , (2011).
[4] Faouzi Ammar, Abdenacer Makhlouf, Nejib Saadoui: Cohomology of Hom-Lie superalgebras
and q-deformed Witt superalgebra, arXiv: 1204.6244, (2012).
[5] Nambu Y: Generalized Hamiltonian mechanics, Phys. Rev. D7: 17405-2412, (1973).
[6] Bagger J. and Lambert N.: Gauge Symmetry and Supersymmetry of multiple M2Branes,
arXiv:0711.0955v2[hep-th] (2007) Phys. Rev. D77: 065008, (2008).
[7] Basu A. and Harvey J.A.: The M2-M5 brane system and a generalised nahm equation Nucl,
Phys. B713: (2005).
[8] Kerner R. : Ternary algebraic structures and their application in physics, in the "Proc.
BTLP 23rd international Colloqium on Group Theoretical Methods in Physics" arXiv math-
ph:0011023,(2000)
[9] R. Kerner, Z3-graded algebras and non-commutative gauge theories, dans le livre, in: Z.
Oziewicz, B. Jancewicz, A. Borowiec (Eds.), Spinors, Twistors, Clifford Algebras and Quantum
Deformations, Kluwer Academic Publishers, 1993, pp. 349357.
[10] -- -- -- -- :The cubic chessboard : Geometry and physics, Classical Quantum Gravity 14, A203-
A225 (1997).
[11] Daletskii Y.L. and Takhtajan L.A.: Leibniz and Lie Structures for Nambu algebra, Letters in
Mathematical Physics 39, (1997) 127141.
[12] Takhtajan L. :TOn foundation of the generalized Nambu mechanics, Comm. Math. Phys.
160(1994), 295-315.
[13] Filippov V., n-Lie algebras, Sibirsk. Mat. Zh. 26, 126-140 (1985) (English transl.: Siberian
Math. J. 26, 879-891 (1985)).
[14] Cassas J.M., Loday J.-L. and Pirashvili Leibniz n-algebras, Forum Math. 14 (2002), 189207.
[15] Ataguema H., Makhlouf A. Deformations of ternary algebras, Journal of Generalized Lie The-
ory and Applications, vol. 1, (2007), 4145.
[16] -- -- -- Notes on cohomologies of ternary algebras of associative type, Journal of Generalized
Lie Theory and Applica- tions 3 no. 3, (2009), 157174
20
[17] Gautheron P.: Some Remarks Concerning Nambu Mechanics, Letters in Mathematical Physics
37 (1996) 103116.
[18] J.A. De Azcarraga, J.M. Izquierdo, Cohomology of Filippov algebras and an analogue of
Whiteheads lemma, J. Phys. Conf. Ser. 175 (2009) 012001
[19] Takhtajan L., :On foundation of the generalized Nambu mechanics, Comm. Math. Phys. 160
(1994), 295-315.
[20] L. Takhtajan. A higher order analog of Chevally-Eilenberg complex and deformation theory of
n-algebras, St. Petersburg Math. J., 6 (1995), 429-438.
21
|
1809.05711 | 1 | 1809 | 2018-09-15T12:56:48 | Zinbiel algebras and bialgebras: main properties and related algebraic structures | [
"math.RA"
] | This work provides a characterization of left and right Zinbiel algebras.Basic identities are established and discussed, showing that Zinbiel algebras are center-symmetric, and therefore Lie-admissible algebras. Their bimodules are given, and used to build a Zinbiel algebra structure on the direct sum of the underlying vector space and a finite-dimensional vector space. In addition, their matched pair is built, and related to the matched pair of their sub-adjacent Lie algebras. Besides, Zinbiel coalgebras are introduced, and linked to their underlying Lie coalgebras and coassociative coalgebras. Moreover, the related Manin triple is defined, and used to characterize Zinbiel bialgebras, and their equivalence to the associated matched pair. | math.RA | math |
Zinbiel algebras and bialgebras: main properties
and related algebraic structures
Mahouton Norbert Hounkonnou
University of Abomey-Calavi, International Chair in Mathematical Physics and
Applications, ICMPA-UNESCO Chair, 072 BP 50, Cotonou, Rep. of Benin
E-mail: [email protected], with copy to [email protected]
Mafoya Landry Dassoundo
University of Abomey-Calavi, International Chair in Mathematical Physics and
Applications, ICMPA-UNESCO Chair, 072 BP 50, Cotonou, Rep. of Benin
E-mail: [email protected]
Abstract. This work provides a characterization of left and right Zinbiel algebras.
Basic identities are established and discussed, showing that Zinbiel algebras are center-
symmetric, and therefore Lie-admissible algebras. Their bimodules are given, and used
to build a Zinbiel algebra structure on the direct sum of the underlying vector space and
a finite-dimensional vector space. In addition, their matched pair is built, and related
to the matched pair of their sub-adjacent Lie algebras. Besides, Zinbiel coalgebras are
introduced, and linked to their underlying Lie coalgebras and coassociative coalgebras.
Moreover, the related Manin triple is defined, and used to characterize Zinbiel bialgebras,
and their equivalence to the associated matched pair.
1. Introduction
One of the important subclasses of nonassociative algebras is that of Lie algebras. Their
properties lead to some generalizations of associative algebras such as Lie-admissible
algebras [1], left-symmetric algebras [2], flexible algebras, alternative algebras, Malcev
algebras, Leibniz algebras, etc.
Leibniz algebra is an algebra (A, [, ]) such that the following identity is satisfied:
[x, [y, z]] = [[x, y], z] − [[x, z], y], ∀x, y, z ∈ A.
(1.0.1)
The category of Zinbiel algebras is Koszul dual to the category of Leibniz algebras in the
sense of J-L. Loday [3], where Zinbiel algebras were called Leibniz algebras. They appear
in various domains in mathematics and physics, e.g. , in the theory of nonlinear geometric
control [4] characterized by the following system:
x1 = u1,
x2 = u2,
x3 =
x1u2 − u2x1.
u1 6 1
u2 6 1
(1.0.2)
This system is completely controllable,
every point can be reached from every
other point, since the distribution spanned by the vector fields f1 = ∂x1 − x2∂x3 and
f2 = ∂x1 − x3∂x3 is nonintegrable due to the nonvanishing commutator [f1, f2] = 2∂x3.
Performing the global coordinate change y1 = x1, y2 = x2, y3 = 1
2 (x3 + x1x2) transforms
the initial system into a less symmetric but simpler form:
i.e.
Now setting
y1 = u1,
y2 = u2,
y3 = y1u2.
u1 6 1
u2 6 1
u3(t) = (u1 ⋆ u2)(t) :=(cid:18)Z t
0
u1(s)ds(cid:19) u2(t)
(1.0.3)
(1.0.4)
yields the Zinbiel identity
(u1 ⋆ (u2 ⋆ u3))(t) = ((u1 ⋆ u2 + u2 ⋆ u1) ⋆ u3)(t).
(1.0.5)
Besides, Zinbiel algebras under q-commutator given by x ◦q y = x ◦ y + qy ◦ x, where
q ∈ C, were investigated in [5].
In this work, we provide a characterization of left and right Zinbiel algebras. We
establish and discussed their basic identities, showing that Zinbiel algebras are center-
symmetric, and, therefore, Lie-admissible algebras. Furthermore, the bimodules are
defined and used to build a Zinbiel algebra structure on the direct sum of the underlying
vector space and a finite-dimensional vector space.
In addition, their matched pair is
constructed and related to the matched pair of their sub-adjacent Lie algebras. Besides,
Zinbiel coalgebras are introduced, and linked to their underlying Lie coalgebras and
coassociative algebras. Moreover, the related Manin triple is defined and used to
characterize Zinbiel bialgebras, and their equivalence to the associated matched pair.
2. Zinbiel algebras: basic properties and consequences
2.1. Zinbiel algebras
Definition 2.1 A left (resp. right) Zinbiel algebra is a vector space A endowed with a
bilinear product ∗ satisfying, for all x, y, z ∈ A,
(x ∗ y) ∗ z = x ∗ (y ∗ z) + x ∗ (z ∗ y),
(2.1.1)
(
resp.
x ∗ (y ∗ z) = (x ∗ y) ∗ z + (y ∗ x) ∗ z
),
(2.1.2)
or, equivalently,
(x, y, z) = x ∗ (z ∗ y)
(
resp.
(x, y, z) = −(y ∗ x) ∗ z)
),
(2.1.3)
(2.1.4)
where, ∀x, y, z ∈ A, (x, y, x) = (x ∗ y) ∗ z − x ∗ (y ∗ z) is the associator associated to ∗.
Proposition 2.2 Let A be a vector space, µ : A × A → A, (x, y) 7→ µ(x, y) = x ∗ y, a
bilinear product, τ an exchange map defined on A ⊗ A, and id the identity map defined
on A. Then, if (A, ∗) is a left Zinbiel algebra, the relation
x ∗ (y ∗ z) = y ∗ (x ∗ z),
∀x, y, z ∈ A,
equivalent to the two equations:
µ ◦ (id ⊗µ) = µ ◦ (id ⊗µ) ◦ (τ ⊗ id),
and
holds.
µ ◦ (id ⊗µ) = (µ ◦ τ ) ◦ (µ ⊗ id) ◦ (id ⊗τ ),
Similarly, if (A, ∗) is a right Zinbiel algebra, the relation
(x ∗ y) ∗ z = (x ∗ z) ∗ y,
∀x, y, z ∈ A,
equivalent to the two identities:
µ ◦ (µ ⊗ id) = µ ◦ (µ ⊗ id) ◦ (id ⊗τ ),
(2.1.5)
(2.1.6)
(2.1.7)
(2.1.8)
(2.1.9)
and
is satisfied.
µ ◦ (µ ⊗ id) = (µ ◦ τ ) ◦ (id ⊗µ) ◦ (τ ⊗ id).
(2.1.10)
2.2. Basic properties
Definition 2.3 The opposite algebra of the algebra (A, ∗) is the algebra denoted by
Aopp = (A, ∗opp) whose product is given by x ∗opp y = y ∗ x, for all x, y ∈ A.
Remark 2.4 (i) If (A, ∗) is a commutative algebra, then Aopp = (A, ∗). Conversely, if
Aopp = (A, ∗), then (A, ∗) is a commutative algebra.
(ii) The opposite algebra of the left (resp. right) Zinbiel algebra is a right (resp.
left)
Zinbiel algebra under the same underlying vector space.
In the sequel, both the left and right Zinbiel algebras are simply called Zinbiel algebras.
Proposition 2.5 [6] Let (A, ∗) be a Zinbiel algebra. Then (A, {, }∗), such that, for all
x, y ∈ A, {x, y}∗ = x ∗ y + y ∗ x, is a commutative associative algebra.
Proposition 2.6 Consider a Zinbiel algebra (A, µ), µ : A × A → A, (x, y) 7→ µ(x, y) =
x ∗ y, and the bilinear map τ : A ⊗ A → A ⊗ A given by τ (x ⊗ y) = y ⊗ x for all x, y ∈ A.
Then, the following identities hold:
µ ◦ (id ⊗(µ ◦ τ )) = µ ◦ (µ ⊗ id) ◦ (id ⊗τ ) + (µ ◦ τ ) ◦ (id ⊗(µ ◦ τ )) ◦ (τ ⊗ id)
(2.2.1)
(µ ◦ τ ) ◦ (µ ⊗ id) = µ ◦ (µ ⊗ id) ◦ (id ⊗τ ) + (µ ◦ τ ) ◦ (id ⊗(µ ◦ τ )) ◦ (τ ⊗ id)
(2.2.2)
(µ ◦ τ ) ◦ ((µ ◦ τ ) ⊗ id) = (µ ◦ τ ) ◦ (id ⊗µ) + (µ ◦ τ ) ◦ (id ⊗(µ ◦ τ ))
(2.2.3)
µ ◦ (id ⊗µ) = µ ◦ (id ⊗µ) ◦ (τ ⊗ id)
(2.2.4)
Theorem 2.7 Let (A, ∗) be a Zinbiel algebra. Then (A, [, ]∗) is a Lie algebra, where, for
all x, y ∈ A, [x, y]∗ = x ∗ y − y ∗ x.
Proof.
Consider a Zinbiel algebra (A, µ), where for all x, y ∈ A, µ(x, y) = x ∗ y. By its
definition, the commutator of a bilinear product is bilinear and skew symmetric. Thus,
proving that (A, [, ] := µ−µ◦τ ) is a Lie algebra, it only remains to prove that [, ] := µ−µ◦τ
satisfies the Jacobi identity. Indeed, we have, for all x, y, z ∈ A,
[x, [y, z]] + [y, [z, x]] + [z, [x, y]] = [x, y ∗ z − z ∗ y] + [y, z ∗ x − x ∗ z] + [z, x ∗ y − x ∗ y]
= {x ∗ (y ∗ z) − y ∗ (x ∗ z)} + {y ∗ (z ∗ x) − z ∗ (y ∗ x)}
+ {z ∗ (x ∗ y) − x ∗ (z ∗ y)} + {(y ∗ x) ∗ z − (y ∗ z) ∗ x}
+ {(z ∗ y) ∗ x − (z ∗ x) ∗ y} + {(x ∗ z) ∗ y − (x ∗ y) ∗ z}
= {µ ◦ (id ⊗µ) − µ ◦ (id ⊗µ) ◦ (τ ⊗ id)}(x, y, z)
+ {µ ◦ (id ⊗µ) − µ ◦ (id ⊗µ) ◦ (τ ⊗ id)}(y, z, x)
+ {µ ◦ (id ⊗µ) − µ ◦ (id ⊗µ) ◦ (τ ⊗ id)}(z, x, y)
+ {(µ ◦ τ ) ◦ (id ⊗(µ ◦ τ ))
− µ ◦ ((µ ◦ τ ) ⊗ id) ◦ (id ⊗τ )}(z, x, y)
+ {(µ ◦ τ ) ◦ (id ⊗(µ ◦ τ ))
− µ ◦ ((µ ◦ τ ) ⊗ id) ◦ (id ⊗τ )}(x, y, z)
+ {(µ ◦ τ ) ◦ (id ⊗(µ ◦ τ ))
− µ ◦ ((µ ◦ τ ) ⊗ id) ◦ (id ⊗τ )}(y, z, x).
Using the relations (2.1.6), (2.1.7), (2.1.10) and (2.1.9), the right hand side of the last
equality of this equation vanishes. Therefore, for all x, y, z ∈ A,
[x, [y, z]∗]∗ + [y, [z, x]∗]∗ + [z, [x, y]∗]∗ = 0,
i.e. the Jacobi identity associated to the bilinear product [, ]∗ holds. Hence, (A, [, ]∗) is a
Lie algebra.
(cid:3)
From the definition of a center-symmetric algebra given in [7] by, for all x, y, z ∈ A,
(x, y, z)◦ := (z, y, x)◦, where (x, y, z)◦ := (x ◦ y) ◦ z − x ◦ (y ◦ z) is the associator of the
bilinear product ◦, we have:
Proposition 2.8 Any Zinbiel algebra is a center-symmetric algebra.
Example 2.9 Consider a vector space A given by A = C[X] :=(a =Xk∈N
On A, we define the bilinear products given by, for all a, b ∈ A,
akX k, ak ∈ C).
a ∗ b := bZ X
0
a(t)dt
and
a ◦ b :=Z X
0
b(t)∂t(a(t))dt,
(2.2.5)
where ∂t := d
dt
. Then, the associators of these products are given, for any a, b, c ∈ A, by
(a, b, c)∗ = −(b ∗ a) ∗ c,
(a, b, c)◦ = a ◦ (c ◦ b).
(2.2.6)
(2.2.7)
Therefore, (C[X], ∗) and (C[X], ◦) are right and left Zinbiel algebras, respectively.
2.3. Bimodule
Definition 2.10 A bimodule of a Zinbiel algebra is a triple (l, r, A), where A is a vector
space endowed with a Zinbiel algebra structure, V is a vector space, and l, r : A → gl(V )
are two linear maps satisfying the following relations, for all x, y ∈ A,
lxly = lx·y + ly·x,
lxry = rx·y = ryrx + rylx.
(2.3.1)
(2.3.2)
Proposition 2.11 Let (A, ·) be a Zinbiel algebra. Consider a vector space V over a field
K and two linear maps l, r : A → gl(V ). The triple (l, r, V ) is a bimodule of A if and only
if there is a Zinbiel algebra structure on the vector space A ⊕ V given by, ∀x, y ∈ A and
all u, v ∈ V,
(x + u) ∗ (y + v) = x · y + (lxv + ryu).
(2.3.3)
Proposition 2.12 Let (l, r, V ) be a bimodule of a Zinbiel algebra (A, ·), where V is a
vector space and l, r : A → gl(V ) are two linear maps. Then,
(i) the following conditions are satisfied, for all x, y ∈ A,
lxy = rylx,
rxry = ryrx.
(2.3.4)
(2.3.5)
(ii) the linear map l − r : A → gl(V ), x 7→ lx − rx is a representation of the sub-adjacent
Lie algebra G(A) := (A, [, ]·) associated to (A, ·).
2.4. Matched pair
Definition 2.13 [8] Let (G, ·) and (H, ◦) be two commutative associative algebras, and
ρ : H → gl(G) and µ : G → gl(H) be two K-linear maps which are representations of H
and G, respectively, satisfying the following relations: for all x, y ∈ G and all a, b ∈ H,
µ(x)(a ◦ b) = (µ(x)a) ◦ b + µ(ρ(a)x)b,
ρ(a)(x · y) = (ρ(a)x) · b + ρ(µ(x)a)y.
(2.4.1)
(2.4.2)
Then, (G, H, ρ, µ) is called a matched pair of the commutative associative algebras G and
H, denoted by G ⊲⊳µ
ρ H.
In this case, (G ⊕ H, ∗) defines a commutative associative algebra with respect to the
product ∗, given, for all x, y ∈ G and all a, b ∈ H, by
(x + a) ∗ (y + b) = x · y + µ(a)y + µ(b)x + a ◦ b + ρ(x)b + ρ(y)a.
Definition 2.14 [9] Let (G, [, ]G ) and (H, [, ]H ) be two Lie algebras and let µ : H → gl(G)
and ρ : G → gl(H) be two Lie algebra representations satisfying the following relations, for
all x, y ∈ G and all a, b ∈ H,
ρ(x) [a, b]
H
− [ρ(x)a, b]
H
− [a, ρ(x)b]
H
+ ρ(µ(a)x)b − ρ(µ(b)x)a = 0,
(2.4.3)
µ(a) [x, y]
G
− [µ(a)x, y]
− [x, µ(a)y]
G
G
+ µ(ρ(x)a)y − µ(ρ(y)a)x = 0.
(2.4.4)
Then, (G, H, ρ, µ) is called a matched pair of the Lie algebras G and H, denoted by H ⊲⊳ρ
µ G.
In this case, (G ⊕ H, [, ]G⊕H ) is a Lie algebra with respect to the product [, ]G⊕H defined on
the direct sum vector space G ⊕ H by, ∀x, y ∈ G and all a, b ∈ H,
[(x + a), (y + b)]G⊕H = [x, y]G + µ(a)y − µ(b)x + [a, b]H + ρ(x)b − ρ(y)a.
(2.4.5)
Theorem 2.15 Let (A, ·) and (B, ◦) be two Zinbiel algebras. Consider the four linear
maps defined as lA, rA : A → gl(B) and lB, rB : B → gl(A). Suppose that (lA, rA, B) and
(lB, rB, A) are bimodules of A and B, respectively, obeying the relations, for all x, y ∈ A
and all a, b ∈ B,
rB(a)(x · y + y · x) − x · (rB(a)y) − rB(lA(y)a)x = 0,
rA(x)(a ◦ b + b ◦ a) − a ◦ (rA(x)b) − rA(lB(b)x)a = 0,
lB(a)(x · y) = ((lB + rB)(a)x) · y + lB((lA + rA)(x)a)y
= x · (lB(a)y) + rB(rA(y)a)x,
(2.4.6)
(2.4.7)
(2.4.8)
lA(x)(a ◦ b) = lA((lB + rB)(a)x)b + ((lA + rA)(x)a) ◦ b
(2.4.9)
= a ◦ (lA(x)b) + rA(rB(b)x)a.
Then, there is a Zinbiel algebra structure on A ⊕ B given by:
(x + a) ∗ (y + b) = (x · y + lB(a)y + rB(b)x) + (a ◦ b + lA(x)b + rA(y)a).
(2.4.10)
Proof.
Suppose x, y, z ∈ A and a, b, c ∈ B. Then, using the bilinear product ∗ defined on A⊕B
by (x + a) ∗ (y + b) = {x · y + lB(x)y + rB(b)x} + {a ◦ b + lA(x)b + rA(y)a}, we have:
{(x + a) ∗ (y + b)} ∗ (z + c) = (x · y) · z + ((lB(a)y) · z + lB(rA(y)az)) + lB(a ◦ b)z
+rB(c)(x · y) + ((rB(b)x) · z + lB(lA(x)b)z) + rB(c)(lB(a)y) + rB(c)(rB(b)x)
+(a ◦ b) ◦ c + ((lA(x)b) ◦ c + lA(rB(b)x)c) + lA(x · y)c + rA(z)(a ◦ b)
+rA(z)(lA(x)b) + ((rA(y)a) ◦ c + lA(lB(a)y)c) + rA(z)(rA(y)a),
(2.4.11)
{(y + b) ∗ (x + a)} ∗ (z + c) = (y · x) · z + ((lB(b)x) · z + lB(rA(x)bz)) + lB(b ◦ a)z
+rB(c)(y · x) + ((rB(a)y) · z + lB(lA(y)a)z) + rB(c)(lB(b)x) + rB(c)(rB(a)y)
+(b ◦ a) ◦ c + ((lA(y)a) ◦ c + lA(rB(a)y)c) + lA(y · x)c + rA(z)(b ◦ a)
+((rA(x)b) ◦ c + lA(lB(b)x)c) + rA(z)(lA(y)a) + rA(z)(rA(x)b),
(2.4.12)
(x + a) ∗ {(y + b) ∗ (z + c)} = x · (y · z) + (x · (lB(b)z) + rB(rA(z)b)x) + lB(a)(y · z)
+lB(a)(lB(b)z) + lB(a)(rB(c)y) + rB(b ◦ c)x + (x · (rB(c)y) + rB(lA(y)c)x)
+a ◦ (b ◦ c) + (a ◦ (lA(y)c) + rA(rB(c)y)a) + (a ◦ (rA(z)b) + rA(lB(b)z)a)
+lA(x)(b ◦ c) + lA(x)(lA(y)c) + lA(x)(rA(z)b) + rA(y · z)a.
(2.4.13)
Summing (2.4.11) and (2.4.12) gives (2.4.13), what is equivalent to say that (lA, rA, B)
and (lB, rB, A) are bimodules of the Zinbiel algebras (A, ·) and (B, ◦), respectively, and
the four linear maps lA, rA, lB, and rB satisfy the equations (2.4.6), (2.4.7), (2.4.8), and
(2.4.9).
(cid:3)
Corollary 2.16 Let (A, B, lA, rA, lB, rB) be a matched pair of the Zinbiel algebras (A, ·)
and (B, ◦). Then,
(i) (Gass(A), Gass(B), lA + rA, lB + rB) is a matched pair of the commutative associative
algebras Gass(A) := (A, {, }· ) and Gass(B) := (B, {, }∗ ).
(ii) (G(A), G(B), lA − rA, lB − rB) is a matched pair of the Lie algebras G(A) := (A, [, ]· )
and G(B) := (B, [, ]◦ ).
3. Zinbiel coalgebras, Manin triple and bialgebra
The structures of nonassociative coalgebras and their basic properties are investigated
and discussed in [10] and references therein. This section is devoted to the construction
of Zinbiel bialgebra using their associated coalgebras and Manin triple.
3.1. Zinbiel coalgebras
Definition 3.1 Let A be a vector space equipped with the linear map ∆ : A → A ⊗ A.
The couple (A, ∆) is a right Zinbiel coalgebra if ∆ satisfies the following identity
(id ⊗∆) ◦ ∆ = (∆ ⊗ id) ◦ ∆ + ((τ ◦ ∆) ⊗ id) ◦ ∆,
(3.1.1)
illustrated by the following commutative diagram:
∆
A
∆
A ⊗ A
id ⊗∆
A ⊗ A
(∆+(τ ◦∆))⊗id
/ A ⊗ A ⊗ A
(3.1.2)
where id is the identity map on A, and τ is the exchange map defined on A ⊗ A.
Definition 3.2 Let A be a vector space endowed with the linear map ∆ : A → A ⊗ A.
The couple (A, ∆) is a left Zinbiel coalgebra if ∆ satisfies the following identity:
(∆ ⊗ id) ◦ ∆ = (id ⊗∆) ◦ ∆ + (id ⊗(τ ◦ ∆)) ◦ ∆,
(3.1.3)
illustrated by the following commutative diagram:
∆
A
∆
A ⊗ A
∆⊗id
A ⊗ A
id ⊗(∆+(τ ◦∆))
/ A ⊗ A ⊗ A
(3.1.4)
where id is the identity map on A, and τ is the exchange map defined on A ⊗ A.
Proposition 3.3 Let A be a vector space equipped with a linear map ∆ : A → A ⊗ A, τ
be the exchange map defined on A ⊗ A, and id be the identity map defined on A.
(i) If (A, ∆) is a right Zinbiel coalgebra, then the relation
(id ⊗∆) ◦ ∆ = (τ ⊗ id) ◦ (id ⊗∆) ◦ ∆ = (τ ⊗ id) ◦ (∆ ⊗ id) ◦ (τ ◦ ∆)
(3.1.5)
is satisfied.
(ii) If (A, ∆) is a left Zinbiel coalgebra, then the relation
(∆ ⊗ id) ◦ ∆ = (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ = (id ⊗τ ) ◦ (id ⊗∆) ◦ (τ ◦ ∆)
(3.1.6)
holds.
Definition 3.4 A commutative infinitesimal bialgebra (A, ·, ∆) is an infinitesimal
bialgebra (A, ·, ∆) such that (A, ·) is a commutative algebra and (A, ∆) is a cocommutative
coalgebra.
/
/
/
/
/
/
Definition 3.5 The opposite coalgebra of the coalgebra (A, ∆) is a coalgebra (A, τ ◦ ∆),
where the coproduct τ ◦ ∆ is defined by: τ ◦ ∆ : A −→ A ⊗ A.
Proposition 3.6 The opposite coalgebra of a right Zinbiel coalgebra is a left Zinbiel
coalgebra. Conversely, the opposite coalgebra of a left Zinbiel coalgebra is a right Zinbiel
coalgebra.
Proposition 3.7 Consider a right Zinbiel coalgebra (A, ∆), and τ the exchange map
given by τ : A ⊗ A → A ⊗ A, for all x, y ∈ A, τ (x ⊗ y) = y ⊗ x. Then, the following
equations hold:
(id ⊗(τ ◦ ∆)) ◦ ∆ = (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ + (τ ⊗ id) ◦ (id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆), (3.1.7)
(∆ ⊗ id) ◦ (τ ◦ ∆) = (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ + (τ ⊗ id) ◦ (id ⊗(τ ◦ ∆) ◦ (τ ◦ ∆), (3.1.8)
((τ ◦ ∆) ⊗ id) ◦ (τ ◦ ∆) = (id ⊗∆) ◦ (τ ◦ ∆) + (id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆).
(3.1.9)
Proof.
Consider a right Zinbiel coalgebra (A, ∆) and the exchange map τ on A ⊗ A. We have:
(id ⊗(τ ◦ ∆)) ◦ ∆ = (id ⊗τ ) ◦ (id ⊗∆) ◦ ∆
= (id ⊗τ ) ◦ ((∆ ⊗ id) ◦ ∆) + (id ⊗∆) ◦ ((τ ⊗ ∆) ⊗ id) ◦ ∆
= (id ⊗∆) ◦ (∆ ⊗ id) ◦ ∆ + (id ⊗τ ) ◦ (id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆)
= (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ + (id ⊗∆) ◦ (τ ◦ ∆)
= (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ + (∆ ⊗ id) ◦ ∆
= (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ + (τ ⊗ id) ◦ ((τ ◦ ∆) ⊗ id) ◦ ∆
(id ⊗(τ ◦ ∆)) ◦ ∆ = (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ + (τ ⊗ id) ◦ (id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆).
Hence, the relation (3.1.7) holds.
Besides,
(id ⊗(τ ◦ ∆)) ◦ ∆ = ((τ ◦ ∆) ⊗ id) ◦ (τ ◦ ∆) = (τ ⊗) ◦ (∆ ⊗ id) ◦ (τ ◦ ∆)
= (id ⊗∆) ◦ ∆ = (∆ ⊗ id) ◦ (τ ◦ ∆),
which is the relation (3.1.8).
Using the following identities,
((τ ◦ ∆) ⊗ id) ◦ (τ ◦ ∆) = (id ⊗(τ ◦ ∆)) ◦ ∆,
(id ⊗∆) ◦ (τ ◦ ∆) = (∆ ⊗ id) ◦ ∆ = (τ ⊗ id) ◦ ((τ ◦ ∆) ⊗ id) ◦ ∆
(id ⊗∆) ◦ (τ ◦ ∆) = (τ ⊗ id) ◦ (id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆),
and
(id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆ = (id ⊗τ ) ◦ (id ⊗∆) ◦ (τ ◦ ∆) = (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆,
we obtain that the relations (3.1.7) and (3.1.9) are equivalent.
(cid:3)
Definition 3.8 A cocommutative coassociative coalgebra is the couple (A, ∆), where A
is a vector space and ∆ : A → A ⊗ A a linear map such that the following relations hold:
∆ = τ ◦ ∆,
(∆ ⊗ id) ◦ ∆ = (id ⊗∆) ◦ ∆.
(3.1.10)
(3.1.11)
Proposition 3.9 Consider a vector
space A and ∆ a linear map given by
∆ : A → A ⊗ A. The couple (A, ∆) is a Zinbiel coalgebra if (A, ∆{,}) is a cocommutative
coassociative coalgebra, where ∆{,} : A → A ⊗ A is given by ∆{,} := ∆ + τ ◦ ∆, and τ is
the exchange map defined on A ⊗ A.
Definition 3.10 Let A be a vector space equipped with the linear map ∆ : A → A ⊗ A.
Then, the couple (A, ∆) is a Lie coalgebra if the following relations are satisfied:
∆ = −τ ◦ ∆,
(id ⊗∆) ◦ ∆ + (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆ − (∆ ⊗ id) ◦ ∆ = 0.
(3.1.12)
(3.1.13)
Proposition 3.11 Consider a vector space A equipped with a linear map ∆ : A → A ⊗ A.
Then, the couple (A, ∆) is a Zinbiel coalgebra if the linear map ∆[,] : A → A ⊗ A given
by ∆[,] := ∆ − τ ◦ ∆ satisfies the relations (3.1.12) and (3.1.13).
Proof.
Let (A, µ) be a Zinbiel algebra and (A, ∆) be a Zinbiel coalgebra. From the definition
of the exchange map, we have:
∆[,] = id ◦∆[,] = τ ◦ (τ ◦ ∆[,]) = τ ◦ (τ ◦ (∆ − τ ◦ ∆))
= τ ◦ (τ ◦ ∆ − ∆) = −τ ◦ (∆ − τ ◦ ∆),
and, hence, ∆[,] = −τ ◦ ∆[,] which is the relation (3.1.12).
Besides,
(id ⊗∆[,]) ◦ ∆[,] = (id ⊗(∆ − τ ◦ ∆)) ◦ (∆ − τ ◦ ∆)
= (id ⊗∆) ◦ ∆ − (id ⊗∆) ◦ (τ ◦ ∆) − (id ⊗(τ ◦ ∆)) ◦ ∆
+ (id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆).
Then, the relation
(id ⊗∆[,]) ◦ ∆[,] = (id ⊗∆) ◦ ∆ − (id ⊗∆) ◦ (τ ◦ ∆) − (id ⊗(τ ◦ ∆)) ◦ ∆
+ (id ⊗(τ ◦ ∆)) ◦ (τ ◦ ∆)
(3.1.14)
holds. In addition,
(∆[,] ⊗ id) ◦ ∆[,] = ((∆ − τ ◦ ∆) ⊗ id) ◦ (∆ − τ ◦ ∆)
= (∆ ⊗ id) ◦ (∆ − τ ◦ ∆)) − ((τ ◦ ∆) ⊗ id) ◦ (∆ − (τ ◦ ∆))
= (∆ ⊗ id) ◦ ∆ − (∆ ⊗ id) ◦ (τ ◦ ∆) − ((τ ◦ ∆) ⊗ id) ◦ ∆
+ ((τ ◦ ∆) ⊗ id) ◦ (τ ◦ ∆,
and the relation
(∆[,] ⊗ id) ◦ ∆[,] = (∆ ⊗ id) ◦ ∆ − (∆ ⊗ id) ◦ (τ ◦ ∆) − ((τ ◦ ∆) ⊗ id) ◦ ∆
+ ((τ ◦ ∆) ⊗ id) ◦ (τ ◦ ∆)
(3.1.15)
is satisfied.
Furthermore, we have
(id ⊗τ ) ◦ (∆[,] ⊗ id) ◦ ∆[,] = (id ⊗τ ) ◦ ((τ ◦ ∆) ⊗ id) ◦ (τ ◦ ∆)
− (id ⊗τ ) ◦ (∆ ⊗ id) ◦ (τ ◦ ∆)
− (id ⊗τ ) ◦ ((τ ◦ ∆) ⊗ id) ◦ ∆
+ (id ⊗τ ) ◦ (∆ ⊗ id) ◦ ∆.
(3.1.16)
In addition, by considering the right hand side of the relations (3.1.14), (3.1.15),
and (3.1.16), and using the relations (3.1.7), (3.1.8), and (3.1.9), we obtain the identity
(3.1.13).
(cid:3)
3.2. Manin triple and Zinbiel bialgebras
Definition 3.12 Consider a nonassociative algebra (A, ·) and B a bilinear form on A.
• B is symmetric if B(x, y) = B(y, x), ∀x, y ∈ A.
• B is invariant if B(x · y, z) = B(x, y · z), ∀x, y, z ∈ A.
• B is nondegenerate if the set {x ∈ A, B(x, y) = 0, ∀y ∈ A} is reduce to {0}.
Definition 3.13 A Manin triple of Zinbiel algebras is a triple (A, A+, A−) endowed with
a nondegenerate symmetric bilinear form B ( ; ), invariant on A, and satisfying:
(i) A = A+ ⊕ A− as K-vector space;
(ii) A+ and A− are Zinbiel subalgebras of A;
(iii) A+ and A− are isotropic with respect to B(; ), i.e., B(A+; A+) = 0 = B(A−; A−).
Theorem 3.14 Let (A, ·) and (A∗, ◦) be two Zinbiel algebras.
the six-
tuple (A, A∗, R∗
if
and only if
is a Manin triple endowed with the bilinear form
B(x + a∗, y + b∗) =< x, b∗ > + < y, a∗ >, ∀x, y ∈ A and ∀a∗, b∗ ∈ A∗, where <, > is the
natural pairing between A and A∗.
◦) is a matched pair of Zinbiel algebras A and A∗
· , L∗
(A ⊕ A∗, A, A∗)
· ; R∗
◦, L∗
Then,
Proposition 3.15 Let (A, ·) be a Zinbiel algebra and (A∗, ◦) be a Zinbiel algebra on its
dual space A∗. Then the following conditions are equivalent:
(i) (A ⊕ A∗, A, A∗) is the standard Manin triple of considered Zinbiel algebra;
(ii) (G(A), G(A∗), −ad∗
◦) is a matched pair of sub-adjacent Lie algebras;
(iii) (A, A∗, R∗
(iv) (A, A∗) is a Zinbiel bialgebra.
◦) is a matched pair of Zinbiel algebras;
· , −ad∗
· , R∗
· , L∗
◦, L∗
Acknowledgments
This work is supported by TWAS Research Grant RGA No.17-542 RG/ MATHS/AF/AC g-fr3240300147.
The ICMPA-UNESCO Chair is in partnership with the Association pour la Promotion Sci-
entifique de l'Afrique (APSA), France, and Daniel Iagolnitzer Foundation (DIF), France,
supporting the development of mathematical physics in Africa.
References
[1] Santilli R M 1968 Nuovo Cim. Suppl. 6 1225-1249
[2] Bai C 2008 Commun. Contemp. Math., 10 No.2 221-260
[3] Loday J-L 1995 Math. Scand. 77 189-196
[4] Kawski M 2001 J. Math. Sci. 103 No.6 725-744
[5] Dzhumadil'daev A S 2007 J. Math. Sci., 144 No.2 3909-3925
[6] Aguiar M 2000 Lett. Math. Phys. 54 263-277
[7] Hounkonnou M.N. and Dassoundo M.L. 2015 Trends Math. 261-273
[8] Ni X and Bai C 2013 J. Math. Phys. 54 023515
[9] Majid S 1990 Pacific J. Math. 141 No.2 311-332
[10] Anquela J A, Cort´es T and Montaner F 1994 22 No12 4693-4716
|
1011.5411 | 2 | 1011 | 2010-12-13T15:40:51 | (Quasi-)Poisson enveloping algebras | [
"math.RA",
"math.CT"
] | We introduce the quasi-Poisson enveloping algebra and Poisson enveloping algebra for a non-commutative Poisson algebra. We prove that for a non-commutative Poisson algebra, the category of quasi-Poisson modules is equivalent to the category of left modules over its quasi-Poisson enveloping algebra, and the category of Poisson modules is equivalent to the category of left modules over its Poisson enveloping algebra. | math.RA | math |
(QUASI-)POISSON ENVELOPING ALGEBRAS
YAN-HONG YANG, YUAN YAO, AND YU YE∗
Abstract. We introduce the quasi-Poisson enveloping algebra and Poisson en-
veloping algebra for a non-commutative Poisson algebra. We prove that for a
non-commutative Poisson algebra, the category of quasi-Poisson modules is equiv-
alent to the category of left modules over its quasi-Poisson enveloping algebra, and
the category of Poisson modules is equivalent to the category of left modules over
its Poisson enveloping algebra.
0. Introduction
Poisson algebras appear naturally in Hamiltonian mechanics, and play a central
role in the study of Poisson geometry and quantum groups. Poisson manifolds are
known as smooth manifolds equipped with a Poisson algebra structure. One can
consult [14] for topics in Poisson geometry.
With the development of non-commutative geometry in the past decades, versions
of non-commutative Poisson-like structures were introduced and studied by many
authors from different perspectives [8, 10, 16, 5, 12, 2, 15, 6, 3, 17]. Non-commutative
Poisson algebra (NCPA) is one of them. The notion of NCPA is first suggested by P.
Xu [16], which is especially suitable for geometric situation, and developed by Flato,
Gerstenhaber and Voronov to a more general case [5].
An important and practical idea to understand an algebraic object is to study
the representations, or equivalently, the modules over it. So is for Poisson algebra.
Poisson modules over an NCPA are defined in a natural way [5], see also in [9, Section
0]. Note that when restricted to a usual Poisson algebra, this notion of Poisson module
is different from the usual one studied in [11, Definition 1] and [4]; see Remark 1.8.
Key words and phrases. Non-commutative Poisson algebra, Poisson module, Poisson enveloping
algebra.
2000 Mathematics Subject Classification. 16S40, 17B63.
∗ The corresponding author.
1
2
YAN-HONG YANG, YUAN YAO, AND YU YE∗
In this paper, we introduce quasi-Poisson enveloping algebra and Poisson envelop-
ing algebra for a given NCPA. It turns out that the category of quasi-Poisson modules
is equivalent to the category of left modules over its quasi-Poisson enveloping algebra,
and the category of Poisson modules is equivalent to the category of left modules over
its Poisson enveloping algebra.
As a consequence, we give an affirmative answer to a question raised in [5, Sec-
tion 2]: does there exist enough injectives or projectives in the category of Poisson
modules? This will enable us to construct certain cohomology theory for NCPA in
a standard way by using projective and injective resolutions [1]. A possible applica-
tion for algebraist is that cohomology theory of the standard Poisson algebras (see
Example 1.3) may give some brand new invariants of associative algebras.
Note that another version of Poisson enveloping algebra for a commutative Poisson
algebra was introduced in [11, Definition 3]. We would like to emphasize that Poisson
enveloping algebra in our sense, when restricted to a commutative Poisson algebra,
is different from that one; see Remark 2.6.
The paper is organized as follows.
In section 1, we recall some definitions and
notations. A characterization of Poisson-simple algebras is also given. Section 2 deals
with the construction of quasi-Poisson enveloping algebra and Poisson enveloping
algebra for a given NCPA. In Section 3, we give our main result: the category of
quasi-Poisson modules is equivalent to the category of left modules over the quasi-
Poisson enveloping algebra; the category of Poisson modules is equivalent to the
category of left modules over the Poisson enveloping algebra.
1. Poisson algebras and modules
Throughout K will be a fixed field of characteristic 0 and algebras considered are
over K. All unadorned ⊗ will mean ⊗K. Without loss of generality, we assume that
associative algebras considered here have an identity.
Definition 1.1. A non-commutative Poisson algebra (NCPA) over K is a triple
A = (A, •, {−, −}), where A is a K-space, (A, •) is an associative algebra and (A, {−, −})
is a Lie algebra, such that the Leibniz rule {ab, c} = a{b, c} + {a, c}b holds for all
a, b, c ∈ A, here we write a•b as ab for brevity. The operator • is called the multipli-
cation and {−, −} the Poisson bracket of A.
(QUASI-)POISSON ENVELOPING ALGEBRAS
3
Remark 1.2. Note that a usual Poisson algebra is by definition an NCPA with
commutative multiplication. In this paper, when we use the term Poisson algebra we
always mean a commutative one, otherwise we use the notion NCPA.
Example 1.3. Any associative algebra A is also a Lie algebra with the Lie bracket
given by the commutator [−, −], say [a, b] = ab − ba for any a, b ∈ A. One shows
easily that (A, •, [−, −]) is an NCPA, which we called the standard NCPA of A.
Definition 1.4. Let A be an NCPA. A quasi-Poisson A-module M is a an A-A-
bimodule M together with a K-bilinear map {−, −}∗ : A × M −→ M, which satisfies
(1.1)
(1.2)
(1.3)
{a, b · m}∗ = {a, b} · m + b · {a, m}∗,
{a, m · b}∗ = m · {a, b} + {a, m}∗ · b,
{{a, b}, m}∗ = {a, {b, m}∗}∗ − {b, {a, m}∗}∗
for all a, b ∈ A and m ∈ M , and a · m and m · a denote the left and right A-action
on M respectively. Note that the last condition is just equivalent to say that M is a
Lie module over A. If moreover,
(1.4)
{ab, m}∗ = a · {b, m}∗ + {a, m}∗ · b
holds for all a, b ∈ A and m ∈ M , then M is called a Poisson A-module. A
homomorphism of quasi-Poisson A-modules (resp. Poisson modules) is a K-linear
map f : M → N such that
(1.5)
(1.6)
f (a · m) = a · f (m), f (m · a) = f (m) · a,
f ({a, m}∗) = {a, f (m)}∗
hold for all a ∈ A and m ∈ M . We denote the category of quasi-Poisson A-modules
(resp. Poisson modules) by Q(A) (resp. P(A)).
Remark 1.5. The definition of Poisson A-module is suggested in [5], see also [9,
Section 0]. One of its immediate merits is that A itself can be naturally viewed as a
Poisson A-module with the regular A-A-bimodule structure and the action {−, −}∗
given by the Poisson bracket, i.e., {a, b}∗ = {a, b} for all a, b ∈ A.
Example 1.6. Let A be an NCPA. Then A ⊗ A admits a quasi-Poisson module
structure with the actions given by a · (b ⊗ c) = ab ⊗ c, (a ⊗ b) · c = a ⊗ bc and
{a, b ⊗ c}∗ = {a, b} ⊗ c + b ⊗ {a, c} for all a, b, c ∈ A. Note that this does not give to
a Poisson module in general.
4
YAN-HONG YANG, YUAN YAO, AND YU YE∗
Example 1.7. Let A be an NCPA with trivial Poisson bracket, say {−, −} = 0. Then
any A-bimodule M is a Poisson module under trivial Lie action, i.e. {−, −}∗ = 0.
Note that this is not true if the Poisson bracket is non-trivial.
Remark 1.8. The usual Poisson modules over commutative Poisson algebras ([11,
Definition 1]) are just the Poisson modules defined here with coincide left and right
associative actions.
Concerning the relations between Hochschild cohomology and deformation theory,
Poisson modules in our sense by using bimodules seems to be more suitable for
deformation theory, even for commutative Poisson algebras.
Recall that a Poisson derivation is a linear endomorphism of A which is both
an associative algebra derivation and a Lie algebra derivation. We have the following
observation, generalizing [4, Example 7].
Proposition 1.9. Let A be an NCPA and C(A) the center of A in the associative
sense. If a K-bilinear form {−, −}∗ : A × A → A makes the regular A-A-bimodule A
a Poisson A-module, then the induced map
ψ : A → A, ψ(a) = {a, 1A}∗, ∀a ∈ A
is a Poisson derivation satisfying ψ(A) ⊆ C(A) and
(1.7)
ψ(a)[b, c] = 0, ∀ a, b, c ∈ A,
here [b, c] = bc − cb denotes the commutator. Conversely, a Poisson derivation
ψ : A → A satisfying ψ(A) ⊆ C(A) and (1.7) makes the regular bimodule A a Poisson
A-module with {−, −}∗ given by
{a, b}∗ = {a, b} + ψ(a)b, ∀a, b ∈ A.
Moreover, if dimK C(A) = 1, then there is a unique Poisson A-module structure
(A, {−, −}∗) on the regular bimodule A, which is given by {a, b}∗ = {a, b}.
Proof. Suppose A is a Poisson A-module. Set ψ(a) = {a, 1A}∗. We will show that ψ
is a Poisson derivation with ψ(A) ⊆ C(A). For any a, b ∈ A,
{a, b}∗ = {a, b} + b{a, 1A}∗
= {a, b} + {a, 1A}∗b,
where the equalities follow from b = b1A = 1Ab and (1.1) and (1.2). Hence ψ(a) ∈
C(A) for all a ∈ A.
(QUASI-)POISSON ENVELOPING ALGEBRAS
5
Next, we have ψ(ab) = {ab, 1A}∗ = {a, 1A}∗b + a{b, 1A}∗ = ψ(a)b + aψ(b) by (1.4),
which implies that ψ is an associative derivation.
Moreover, the following equalities implies that ψ is also a Lie algebra derivation.
ψ({a, b}) = {{a, b}, 1A}∗ = {a, {b, 1A}∗}∗ − {b, {a, 1A}∗}∗
= {a, {b, 1A}∗} + {b, 1A}∗{a, 1A}∗ − {b, {a, 1A}∗} − {a, 1A}∗{b, 1A}∗
= {a, ψ(b)} + {ψ(a), b}
At last, ∀ a, b, c ∈ A we have
{ab, c}∗ = {ab, c} + ψ(ab)c = a{b, c} + {a, c}b + aψ(b)c + ψ(a)bc,
a{b, c}∗ + {a, c}∗b = a{b, c} + aψ(b)c + {a, c}b + ψ(a)cb.
Combined with (1.4), we get the equality (1.7).
The proof of the converse part is easy and left to the reader.
Now suppose that dimK C(A) = 1. It suffices to consider the case dimK A ≥ 3.
Take a basis {e1, e2, e3 · · · } of A with e1 = 1A. Thus C(A) = Ke1. Suppose that
(A, {−, −}∗) is a Poisson A-module and ψ the induced derivation. Then
ψ(eiej) = ψ(ei)ej + eiψ(ej), ∀i 6= j ≥ 2.
Note that ψ(eiej) ∈ Ke1, ψ(ei)ej ∈ Kej and eiψ(ej ) ∈ Kei. By assumption e1, ei, ej
are linearly independent and it forces ψ(ei) = ψ(ej ) = 0, ∀i, j ≥ 2. Since ψ(1A) = 0
for any Poisson derivation ψ, we obtain that ψ(a) = 0, ∀a ∈ A, and hence {a, b}∗ =
(cid:3)
{a, b} + ψ(a)b = {a, b}, ∀a, b ∈ A.
Remark 1.10. For an NCPA, it is still not known whether the above correspondence
between the set of Poisson module structures on the regular A-bimodule and the set
of Poisson derivations is bijective. This is just equivalent to ask whether any Poisson
derivation for an NCPA has the property (1.7).
In the rest of this section, we will characterize the simplicity of an NCPA A by
studying its Poisson modules. Recall that a left (resp. right, two-sided) Poisson
ideal of A is defined to be an ideal in the Lie algebra sense and a left (resp. right,
two-sided) ideal in the associative algebra sense. When I is a two-sided Poisson ideal
of A, it is easy to check that A/I is a Poisson A-module with the weak Poisson action
given by {a, b + I}∗ , {a, b} + I for any a, b ∈ A. Moreover, A/I is an NCPA with
the obvious bracket {a + I, b + I} , {a, b} + I.
6
YAN-HONG YANG, YUAN YAO, AND YU YE∗
An NCPA A is left (resp. right) Poisson-simple provided that any left (resp.
right) Poisson ideal of A is either 0 or A. A is said to be Poisson-simple if any two-
sided Poisson ideal of A is either 0 or A itself. Clearly if A is simple as an associative
algebra, then A is Poisson-simple.
Let M be a quasi-Poisson A-module. The annihilator AnnA(M ) of M in A is
defined as AnnA(M ) , {a ∈ A a · m = m · a = 0, ∀m ∈ M }. Clearly AnnA(M ) is a
two-sided ideal of A in the associative sense. Moreover, AnnA(M ) is also a Lie ideal
and hence a Poisson ideal, just as shown in the following lemma.
Lemma 1.11. Let A be an NCPA and M a quasi-Poisson A-module. Then AnnA(M )
is a two-sided Poisson ideal of A.
Proof. Let a ∈ AnnA(M ) and b ∈ A. Then we have (ba)m = b(am) = 0 and
m(ba) = (mb)a = 0, hence ba ∈ AnnA(M ). Similarly we have ab ∈ AnnA(M ).
Therefore AnnA(M ) is a two-sided ideal in the associative sense.
Since M is a quasi-Poisson module by assumption, (1.1) and (1.2) implies that
{b, a}m = {b, am}∗ − a{b, m}∗ = 0 and m{b, a} = {b, ma}∗ − {b, m}∗a = 0, thus
(cid:3)
{b, a} ∈ AnnA(M ). It concludes that Ann(M ) is a two-sided Poisson ideal.
Proposition 1.12. Let A be an NCPA. The following statements are equivalent:
(i) A is Poisson-simple;
(ii) For every nonzero quasi-Poisson A-module M , AnnA(M ) = 0;
(iii) For every nonzero Poisson A-module M , AnnA(M ) = 0.
Proof. (i)=⇒(ii). Assume that A is Poisson simple and M is a quasi-Poisson A-
module. AnnA(M ) is a two-sided Poisson ideal of A by Lemma 1.11. M 6= 0 implies
that AnnA(M ) 6= A, for 1A /∈ AnnA(M ) in this case. Now the simplicity of A forces
AnnA(M ) = 0.
(ii)=⇒(iii) is obvious, for a Poisson module is always a quasi-Poisson module.
(iii)=⇒(i). Suppose that for any Poisson A-module M 6= 0, AnnA(M ) = 0. As-
sume that A is not Poisson-simple and let I 6= 0 be a proper two-sided Poisson
ideal, then A/I is a Poisson A-module such that AnnA(A/I) = I, which leads to a
(cid:3)
contradiction. Thus A must be Poisson-simple.
(QUASI-)POISSON ENVELOPING ALGEBRAS
7
2. Poisson enveloping algebras
The main purpose of this section is to introduce the notion of Quasi-Poisson en-
veloping algebra and Poisson enveloping algebra for an NCPA. For this we need some
preparation.
Recall that the opposite algebra Aop of an associative algebra A has the same
underlying vector space as A and the multiplication is given in a reversed order, i.e.
a ·Aop b = b ·A a. The associative algebra A ⊗ Aop is called the enveloping algebra
of A and denoted by Ae. We denote a ·A b by a•b and a ·Aop b by a◦b to simplify
the notation. Note that if A is an NCPA , then Aop is also an NCPA with the same
Poisson bracket as defined on A.
Let H be a Hopf algebra and A an H-module algebra, i.e. A is an associative
algebra and a left H-module, such that h(1A) = ǫ(h)1A, h(ab) = P(h1a)(h2b) for all
h ∈ H and a ∈ A, here we use the Sweedler's notation ∆(h) = P h1 ⊗ h2. Then we
may define a multiplication ⋆ on the vector space A ⊗ H by setting
(a ⊗ h) ⋆ (a′ ⊗ h′) = X ah1(a′) ⊗ h2h′
for all a, a′ ∈ A and h, h′ ∈ H. It is not hard to show that (A ⊗ H, ⋆) is an associative
algebra with identity 1A ⊗ 1H, which we call the smash product of A and H and
denote by A#H. Elements in A#H are usually written as a#h rather than a ⊗ h.
We refer to [13] for more details about Hopf algebras, module algebras over a Hopf
algebra and their smash product.
Remark 2.1. Let H be a cocommutative Hopf algebra. Recall that H is cocommu-
tative if ∆(h) = P h1 ⊗ h2 = P h2 ⊗ h1 for any h ∈ H. Then for any H-module
algebra A, the opposite algebra Aop is also a H-module algebra with the same H-
action as the one on A. Moreover, if A and B are both H-module algebra, then
so is A ⊗ B, here the H-action on A ⊗ B is given by the usual tensor product, say
h · (a ⊗ b) = P h1 ·A a ⊗ h2 ·B b for any h ∈ H, a ∈ A and b ∈ B. In particular, the
enveloping algebra Ae of an H-module algebra A is also an H-module algebra.
Let A be a K-vector space and T (A) the tensor algebra. By definition there is a
decomposition of vector spaces T (A) = L∞
i=0 T i(A), making T (A) a positively graded
algebra, here T i(A) = A⊗i. Any a ∈ T i(A) is said to be a homogeneous element of
degree i, denoted by deg(a) = i, particularly deg(1T (A)) = 0. We may also identify
A with a subspace of T (A), say T 1(A), in an obvious way.
The following notations will be handy later on:
8
YAN-HONG YANG, YUAN YAO, AND YU YE∗
(i) Let {vi}i∈S be a fixed K−basis of A, here S is an index set. Let α =
(i(1), i(2), · · · , i(r)) ∈ Sr be a sequence in S. r is called the degree of α and denoted
by r = deg(α). We denote the element vi(1)vi(2) · · · vi(r) = vi(1) ⊗ vi(2) ⊗ · · · ⊗ vi(r) by
−→
−→α and also denote 1T (A) by
∅ for consistency, where ∅ is the empty sequence. Thus
{−→α α ∈ Sr, r > 0} ∪ {1T (A)} form a basis of T (A). For given α = (i(1), · · · , i(r))
and β = (j(1), · · · , j(s)), we define α ∨ β , (i(1), · · · , i(r), j(1), · · · , j(s)), hence
−→α
−−−→
α ∨ β.
−→
β =
(ii) Let α = (i(1), · · · , i(r)) and X ⊔Y be an ordered bipartition of r = {1, · · · , r},
here "ordered" means that X ⊔ Y and Y ⊔ X give different bipartitions, which differs
from the usual one. We would like to emphasize that X and Y are allowed to
be empty here. Suppose X = {X1, X2, · · · , Xr1 } and Y = {Y1, Y2, · · · , Yr2} with
X1 < X2 < · · · < Xr1 and Y1 < Y2 < · · · < Yr2 , where r1 = X, r2 = Y . Set
αX = (i(X1), i(X2), · · · , i(Xr1 )) and αY = (i(Y1), i(Y2), · · · , i(Yr2)). By definition
α = αX ⊔ αY is called an ordered bipartition of α with respect to the ordered
bipartition r = X ⊔ Y . Similarly, one defines ordered n-partitions α = α1 ⊔
α2 · · · ⊔ αn for any n ≥ 2.
Remark 2.2. Note that by definition, different partitions r = X ⊔ Y and r = X ′ ⊔ Y ′
give rise to different partitions α = αX ⊔ αY and α = αX ′ ⊔ αY ′, even if αX = αX ′
and αY = αY ′ as sequences in S.
For example, let i ∈ S and α = (i, i). Then there are exactly 4 ordered bipartitions
of α:
{1, 2} = ∅ ⊔ {1, 2},
{1, 2} = {1} ⊔ {2},
{1, 2} = {2} ⊔ {1},
{1, 2} = {1, 2} ⊔ ∅,
α = ∅ ⊔ (i, i);
α = (i) ⊔ (i);
α = (i) ⊔ (i);
α = (i, i) ⊔ ∅.
By the expression Pα=α1 ⊔α2, we mean to take the sum over each possible ordered
bipartition of {1, · · · , deg(−→α )}. Similarly, Pα=α1⊔α2⊔α3 means the sum is taken over
all possible ordered tripartitions of {1, · · · , deg(−→α )}, and so on.
One shows that T (A) admits a Hopf algebra structure. The counit ǫ is given by
ǫ(1T (A)) = 1 and ǫ(−→α ) = 0 for all α ∈ Sr with r > 0. The coproduct ∆ is given by
∆(−→α ) = X
−→α1 ⊗ −→α2
α=α1⊔α2
(QUASI-)POISSON ENVELOPING ALGEBRAS
9
for all α ∈ Sr with r ≥ 0.
In particular, ∆(1T (A)) = 1T (A) ⊗ 1T (A) and ∆(a) =
1T (A) ⊗ a + a ⊗ 1T (A) for any a ∈ A. The coproduct ∆ is known as the "shuffle
coproduct". By definition ∆ is cocommutative, and hence T (A) is a cocommutative
Hopf algebra.
Assume further that A is a Lie algebra and consider the universal enveloping
algebra U (A) of A. By definition U (A) = T (A)/I, where I is the two-sided ideal
of T (A) generated by {x ⊗ y − y ⊗ x − {x, y} x, y ∈ A}. We denote the canonical
projection by π : T (A) ։ U (A). Moreover, I is easily shown to be a Hopf ideal of
T (A) and hence U (A) inherits a cocommutative Hopf algebra structure.
Let {vi}i∈S be a fixed K−basis of A with S an index set. We may assume that S
is totally ordered. According to the Poincare-Birkhoff-Witt theorem, the elements
{−→α α = (i(1), i(2), · · · , i(r)) ∈ Sr, i(1) ≤ i(2) ≤ · · · ≤ i(r), r ≥ 1},
along with 1U (A), form a basis of U (A), here we also denote the element π(−→α ) in U (A)
by −→α by abuse of notation. Clearly we may identify A with π(T 1(A)), a subspace of
U (A).
Recall that for a Lie algebra A, one of the advantages of U (A) is that the cat-
egory of Lie modules over A is equivalent to the category of left modules over
U (A). Particularly, the Lie bracket makes A a Lie module and hence a U (A)-
module with the action given by −→α · a = {vi(1), {vi(2), {, · · · , {vi(r), a} · · · }}} for all
α = (i(1), i(2), · · · , i(r)) ∈ Sr and a ∈ A. We denote the action −→α · a by −→α (a). Note
that Aop has the same underlying space as A and hence the same U (A)-action. Since
U (A) is a Hopf algebra, Ae = A ⊗ Aop is a tensor product of U (A)-modules and hence
−→α1(a) ⊗ −→α2(b) for all α ∈ Sr with
a U (A)-module. Explicitly, −→α (a ⊗ b) = Pα=α1 ⊔α2
r ≥ 0, a, b ∈ A.
Proposition 2.3. Let A = (A, •, {−, −}) be an NCPA. Then A, Aop and Ae are all
U (A)-module algebras with the above U (A)-module structure.
Proof. We begin with the algebra A. Clearly 1A is a center element of the Lie algebra
A, hence x(1A) = ǫ(x)1A for any x ∈ U (A). Next we prove that
(2.1)
x(a•b) = X(x1(a))•(x2(b))
10
YAN-HONG YANG, YUAN YAO, AND YU YE∗
for all x ∈ U (A) and a, b ∈ A. We need only to check it for a set of generators. In
fact, if x, y ∈ U (A) are both satisfying (2.1), then
(xy)(a•b) = x(y(a•b)) = x(X y1(a)•y2(b))
= X(x1(y1(a)))•(x2(y2(b)))
= X((x1y1)(a))•((x2y2)(b))
= X((xy)1(a))•((xy)2(b)),
which implies that xy also satisfies (2.1).
Note that for any v ∈ A, ∆(v) = 1U (A) ⊗ v + v ⊗ 1U (A). By (1.4) we have
v(a•b) = {v, a•b} = a•{v, b} + {v, a}•b = X(v1(a))•(v2(b))
for all a, b ∈ A. Since A generates U (A), it follows that that A is a U (A)-module
algebra. By Remark 2.1, Aop and Ae are both U (A)-module algebras.
(cid:3)
With the above preparation we give our main definition as follows.
Definition 2.4. Let A = (A, •, {−, −}) be an NCPA. The smash product Ae#U (A)
is called the quasi-Poisson enveloping algebra of A and denoted by Q(A). The
Poisson enveloping algebra of A, denoted by P(A), is defined to be the quotient
algebra Q(A)/J, here J is the ideal of Q(A) generated by
{1A ⊗ 1Aop#(a•b) − a ⊗ 1Aop#b − 1A ⊗ b#a a, b ∈ A}.
Remark 2.5. By definition, Q(A) = A ⊗ Aop ⊗ U (A) as a vector space. Thus Q(A)
has a PBW-type basis given by
{vi ⊗ vj#−→α i, j ∈ S, α = (i(1), i(2), · · · , i(r)) ∈ Sr, i(1) ≤ · · · ≤ i(r), r ≥ 0}.
The identity in Q(A) is given by 1A ⊗ 1Aop#1U (A), and the multiplication in Q(A)
can be written down explicitly as
( vi1 ⊗ vj1#−→α )(vi2 ⊗ vj2#
−→
β ) = X
(vi1
α=α1 ⊔α2⊔α3
−→α1(vi2)) ⊗ (vj1
•
−→α2(vj2))#(−→α3
◦
−→
β ),
here • and ◦ denote the multiplications in A and Aop respectively.
Remark 2.6. Recall that for a usual Poisson algebra A, another version of Poisson
enveloping algebra, denoted by P ′(A) as distinguished from P(A), has already been
developed; see [11, Definition 3], for example. This notion differs from ours, the
reason is that the Poisson modules concerned are different, just as mentioned in
Remark 1.8. In fact, P ′(A) = P(A)/h(a ⊗ 1Aop − 1A ⊗ a)#1U (A), a ∈ Ai.
(QUASI-)POISSON ENVELOPING ALGEBRAS
11
Remark 2.7. We have natural morphisms of associative algebras:
i : A ֒→ Q(A), a 7→ a ⊗ 1Aop#1U (A);
k : Aop ֒→ Q(A), a 7→ 1A ⊗ a#1U (A);
i′ : A ֒→ Q(A) ։ P(A);
k′ : Aop ֒→ Q(A) ։ P(A);
and morphisms of Lie algebras:
j : A ֒→ Q(A), a 7→ 1A ⊗ 1Aop#a;
j′ : A ֒→ Q(A) ։ P(A).
By construction of smash product we know that i, j, k are all injective maps. Let
f : P(A) → A be the linear map given by f (a ⊗ b#−→α ) = ǫ(−→α )ab. Clearly f ◦ i′ =
f ◦ k′ = IdA, which implies that i′ and k′ are both injective maps. Note that j′ is not
an injective map, for 0 6= 1A ∈ Ker(j′).
Clearly i(A), j(A) and k(A) generate Q(A) and hence P(A). Moreover, the gen-
erating relations can be written down explicitly.
Proposition 2.8. Let A be an NCPA. Then Q(A) is generated by the generators
i(a), j(a), k(a), a ∈ A subject to relations:
i(a)i(b) = i(a•b), k(a)k(b) = k(a◦b), i(a)k(b) = k(b)i(a);
j(a)j(b) − j(b)j(a) = j({a, b});
j(a)i(b) = i(b)j(a) + i({a, b}), j(a)k(b) = k(b)j(a) + k({a, b}), ∀a, b ∈ A.
The proof is given by routine check and we omit it here.
Example 2.9. In case A is an NCPA with trivial Poisson bracket, it is plain to see
that Q(A) = A ⊗ Aop ⊗ S(A) as an associative algebra, here S(A) is the polynomial
ring of A.
Example 2.10. Let A = K × K as an associative algebra. Then there exists a unique
Lie bracket, say the commutative one, making A an NCPA. We claim that in this
case, P(A) ∼= K × K × K[x] × K[x].
In fact, 1A = e1 + e2 is an orthogonal decomposition of identity, here e1 = (1, 0)
and e2 = (0, 1). Set est = i′(es)k′(et) for all 1 ≤ s, t ≤ 2, α12 = i′(e1)j′(e1)k′(e2) and
α21 = i′(e2)j′(e2)k′(e1). For any subset S ⊆ P(A), we denote by K⌈S⌉ the subalgebra
of P(A) generated by S, note that we do not require that 1P(A) ∈ K⌈S⌉ here.
Consider subalgebras A11 = K⌈e11⌉, A22 = K⌈e22⌉, A12 = K⌈e12, α12⌉ and A21 =
K⌈e21, α21⌉. Clearly A11 ∼= K ∼= A22 and A12 ∼= K[x] ∼= A21 as algebras. Moreover,
P(A) = A11 × A22 × A12 × A21 and the assertion follows.
12
YAN-HONG YANG, YUAN YAO, AND YU YE∗
Example 2.11. Let A = Khx1, x2, · · · , xni/J 2 be a 2-truncated algebra, where J =
hx1, x2, · · · , xni is the graded Jacobson radical. Let {−, −} be a Lie bracket on A.
Then (A, •, {−, −}) is an NCPA if and only if 1A is a center element in the Lie
algebra A.
Set αi = i′(xi), βi = k′(xi) and γi = j′(xi) for any 1 ≤ i ≤ n, here i′, j′, k′ be given
as in Remark 2.7. Then P(A) has a basis consisting of
αi, βj , αiβj, 1 ≤ i, j ≤ n,
γa1
1 γa2
2 · · · γan
n , ai ≥ 0, i = 1, 2, · · · , n,
αiγa1
1 γa2
2 · · · γan
n , ai ≥ 0, i = 1, 2, · · · , n,
n
X
i=1
ai > 0.
Remark 2.12. To calculate the Poisson enveloping algebras of an NCPA could be
very difficult in general. An open question is to ask whether there exist PBW-like
bases for Poisson enveloping algebras.
3. Poisson modules are modules
We are in a position to exhibit the equivalence between the category of (quasi)-
Poisson modules over an NCPA and the category of modules over its (quasi)-Poisson
enveloping algebras.
Let A be an NCPA and M ∈ Q(A) a quasi-Poisson module. To avoid confusion we
fix some notations. The left, right and Lie action of A on M are denoted by ·L, ·R and
{−, −}∗ respectively. The induced action of U (A) on M is denoted by x(m) for any
x ∈ U (A) and m ∈ M , to be precise, −→α (m) = {vi(1), {vi(2), {, · · · , {vi(r), a}∗ · · · }∗}∗}∗
for any α = (i1, · · · , ir) ∈ Sr and m ∈ M . By the similar argument used in the
proof of Proposition 2.3, we show that x(a ·L m) = P x1(a) ·L x2(m) and x(m · b) =
P x1(m) ·R x2(b) for any x ∈ U (A), a, b ∈ A and m ∈ M , again we use the Sweedler's
notation here. The action of Q(A) is always denoted by ·. The multiplications in A
and Aop are denoted by • and ◦, respectively.
Now we define F : Q(A) −→ Mod(Q(A)) as follows. Let M ∈ Q(A)). We define
an action of Q(A) on the underlying space M by setting
(vi ⊗ vj#−→α ) · m , vi ·L (−→α (m)) ·R vj
(QUASI-)POISSON ENVELOPING ALGEBRAS
13
for any i, j ∈ S, α ∈ Sr and m ∈ M . We claim that this action makes M a Q(A)-
module. For this it suffices to show that the associativity
(3.1)
(vi1 ⊗ vj1#−→α )((vi2 ⊗ vj2#
−→
β )(m)) = ((vi1 ⊗ vj1#−→α )(vi2 ⊗ vj2#
−→
β ))(m)
holds for any m ∈ M , i1, i2, j1, j2 ∈ S, α and β. One uses Proposition 2.8 to give a
proof; or one can give a direct argument as follows. In fact, the left hand side
LHS = (vi1 ⊗ vj1#−→α )(vi2 ·L (
−→
β (m)) ·R vj2)
−→
β (m) ·R vj2) ·R vj1
−→
vi1 ·L (−→α1(vi2 ) ·L (−→α3
β )(m) ·R
−→α2(vj2)) ·R vj1
(vi1
−→α1(vi2)) ·L (−→α3
•
−→
β )(m) ·R (vj1
−→α2(vj2))
◦
(vi1
−→α1(vi2) ⊗ vj1
•
−→α2(vj2)#(−→α3
◦
−→
β ))(m)
= vi1 ·L
−→α (vi2 ·L
= X
α=α1 ⊔α2⊔α3
= X
α=α1 ⊔α2⊔α3
= X
α=α1 ⊔α2⊔α3
= RHS,
the right hand side. We denote by F M the U (A)-module with underlying space M
and the above action.
Next we define G : Mod(Q(A)) −→ Q(A). Given M ∈ Mod(Q(A)), we may endow
the underlying space M a quasi-Poisson A-module structure by setting
(3.2)
(3.3)
a ·L m , i(a) · m, m ·R a , k(a) · m, ∀a ∈ A, m ∈ M ;
{a, m}∗ , j(a) · m, ∀a ∈ A, m ∈ M.
Clearly M is an A-bimodule as well as a Lie module with the above actions, since i, k
are homomorphisms of associative algebras and j is a homomorphism of Lie algebras.
We claim that M is a quasi-Poisson module with the above actions. For this it suffices
to show that (1.1) and (1.2), which can be translated to
j(a) · (i(b) · m) = i({a, b}) · m + i(b) · (j(a) · m),
j(a) · (k(b) · m) = k({a, b}) · m + k(b) · (j(a) · m),
hold for any a, b ∈ A and m ∈ M . This is obvious since in Q(A),
j(a)i(b) = i({a, b}) + i(b)j(a),
j(a)k(b) = k({a, b}) + k(b)j(a).
We denote by GM the quasi-Poisson A-module with the underlying space M and the
above actions.
14
YAN-HONG YANG, YUAN YAO, AND YU YE∗
Recall that there is a forgetful functor, denoted by F
Q(A) from Q(A) to Mod(K),
the category of K-vector spaces. Similarly we have forgetful functors FMod(Q(A)),
F
P(A) and F
Mod(P(A)).
Lemma 3.1. Let F : Q(A) → Mod(Q(A)) and G : Mod(Q(A)) → Q(A) be given
as above. Then F and G give rise to functors respecting the forgetful functors to
Mod(K), i.e. FQ(A) = FMod(Q(A)) ◦ F and FMod(Q(A)) = FQ(A) ◦ G.
Proof. We consider F first. Note that for each M ∈ Q(A), we have defined an object
F M ∈ Mod(Q(A)) with F M sharing the same underlying space with M . To extend
F to a functor, we need to define a map
F : HomQ(A)(M, N ) −→ HomQ(A)(F M, F N )
for each pair of M, N ∈ Q(A), such that certain compatible conditions are satisfied.
Now the requirement to respect the forgetful functor means that for each M ∈ Q(A),
M and F M have the same underlying vector space, and for each f ∈ HomQ(A)(M, N ),
F f = f as morphisms of vector spaces.
To complete the proof we need only to show that if f : M −→ N is a morphism
of quasi-Poisson modules, then f is a morphism of Q(A)-modules. This is easy to
show. In fact, we have f (vi ·L
−→α (f (m)) ·R vj and hence
−→α (m) ·R vj) = vi ·L
f ((vi ⊗ vj#−→α ) · m) = (vi ⊗ vj#−→α ) · f (m)
for all m ∈ M, i, j ∈ S and α.
Similar argument works for G and the lemma follows.
(cid:3)
Recall that the Poisson enveloping algebra P(A) of A is by definition the quotient
algebra Q(A)/J, where J is the ideal of Q(A) generated by
{1A ⊗ 1Aop#(a•b) − a ⊗ 1Aop#b − 1A ⊗ b#a a, b ∈ A}.
Thus Mod(P(A)) is a full subcategory of Mod(Q(A)), and a Q(A)-module M is a
P(A)-module if and only if J annihilates M .
Now consider the restriction of the functor F to P(A), the full subcategory con-
sisting of Poisson modules. Let M ∈ P(A). One checks easily that in F M ,
(1A ⊗ 1Aop#(a•b) − a ⊗ 1Aop#b − 1A ⊗ b#a) · m
={a•b, m}∗ − a ·L {b, m}∗ − {a, m}∗ ·R b = 0.
(QUASI-)POISSON ENVELOPING ALGEBRAS
15
The first equality follows from the construction of F M and the latter one from the
definition of Poisson modules. Thus J annihilates F M and hence F M is a P(A)-
module. Therefore F P(A) gives a functor F ′ : P(A) −→ Mod(P(A)).
Conversely, the above equalities also implies that for each P(A)-module M , GM
satisfies (1.4), and hence is a Poisson A-module. Thus GMod(P(A)) gives a functor
G′ : Mod(P(A)) −→ P(A).
Now our main result is stated as follows.
Theorem 3.2. Let A be an NCPA. Then there are isomorphisms of categories
Q(A) ∼= Mod(Q(A)) and P(A) ∼= Mod(P(A)). In fact, we have
(3.4)
(3.5)
F ◦ G = IdMod(Q(A)), G ◦ F = IdQ(A),
F ′ ◦ G′ = IdMod(P(A)), G′ ◦ F ′ = IdP(A),
here F, G, F ′, G′ are functors given as above.
Proof. First we show that G ◦ F (M ) = M . Since G ◦ F (M ) is defined to have the
same underlying vector space as M , it is only left to prove that they have the same
actions of A. By definitions of F and G, we have
a ·G◦F (M ) (m) = i(a) ·F (M ) (m) = a ·M (m),
m ·G◦F (M ) a = k(a) ·F (M ) (m) = m ·M a,
{a, m}∗G◦F (M ) = j(a) ·F (M ) (m) = {a, m}∗M
for all a ∈ A and m ∈ M , where ·G◦F (M ), {−, −}∗G◦F (M ), ·M and {−, −}∗M denote
the actions of A on the modules G ◦ F (M ) and M respectively. Thus we have shown
that G ◦ F (M ) = M as quasi-Poisson A-modules.
Next we show that F ◦ G(M ) = M for all M ∈ Mod(Q(A)), again it suffices to
show that
(vi ⊗ vj#−→α ) ·F ◦G(M ) m = (vi ⊗ vj#−→α ) ·M m
holds for any vi ⊗ vj#−→α ∈ Q(A) and m ∈ M . This is true since by definition
(vi ⊗ vj#−→α ) ·F ◦G(M ) m
=vi ·G(M ) (−→α G(M )(m)) ·G1(M ) vj
=((1A ⊗ vj#1U (A))(vi ⊗ 1Aop#1U (A))(1A ⊗ 1Aop#−→α )) ·M m
=(vi ⊗ vj#−→α ) ·M m.
16
YAN-HONG YANG, YUAN YAO, AND YU YE∗
Recall that F and G are both functors respecting the forgetful functor, which
It follows that
means G ◦ F (f ) = f for each morphism f in the category Q(A).
G ◦ F = IdQ(A) as functors.
Similarly, we obtain that F ◦ G = IdMod(Q(A)). Since F ′ is a restriction of F and
G′ is a restriction of G, we know that F ′ ◦ G′ is a restriction of F ◦ G, and hence we
deduce easily that F ′ ◦ G′ = IdMod(P(A)). Similarly we have G′ ◦ F ′ = IdP(A). The
(cid:3)
proof is completed.
Remark 3.3. An easy consequence of the theorem is that the categories Q(A) and
P(A) are both abelian categories with enough projectives and injectives, which an-
swers a question raised in [5, Section 2]. The quasi-Poisson and Poisson enveloping
algebras also enable us to apply the ring theoretic methods to the study of NCPA,
especially to the study of cohomology theory of NCPA.
Remark 3.4. An NCPA A itself can be viewed as a P(A)-module via the functor
F ′ and a Poisson ideal of A is just a P(A)-submodule of A. Now A is Poisson simple
if and only if A is simple as a P(A)-module.
Example 3.5. Let A be an associative algebra and consider the standard NCPA of
A. It is direct to check that any A-bimodule M is a Poisson A-module under the
Lie action given by {a, m}∗ = a · m − m · a. This gives a functor from the category
Mod(A ⊗ Aop) to the category P(A), and hence to Mod(P(A)). In fact, this functor
is induced from an isomorphism of associative algebras ψ : A ⊗ Aop ∼= P(A)/I, where
I = hj(a) − i(a) + k(a), a ∈ Ai is a two-sided ideal of P(A).
−→ P(A)
The isomorphism ψ is given by the composition map A ⊗ Aop f
π−→ P(A)/I,
here f is defined by f (a ⊗ b) = i′(a)k′(b) for any a, b ∈ A and π is the canonical map.
Obviously ψ is an epimorphism, hence to show that ψ is an isomorphism it suffices to
show that it is also injective. For this, one uses the fact that the A-bimodule A ⊗ Aop
is a Poisson module under the Lie action given by {c, a ⊗ b}∗ = ca ⊗ b − a ⊗ bc for
all a, b, c ∈ A. Thus A ⊗ Aop is a P(A)-module. Moreover, the ideal I annihilates
A ⊗ Aop and hence A ⊗ Aop is a P(A)/I-module. By definition of the functor F ,
ψ(x) · (1A ⊗ 1Aop) = x for any x ∈ A ⊗ Aop.
It follows that ψ(x) 6= 0 for any
0 6= x ∈ A ⊗ Aop, which implies that ψ is injective.
Acknowledgements: This work is supported by National Natural Science Founda-
tion of China (No. 10971206). The authors would like to thank Dr. Yan-Hong Bao
and Dr. Xiao-Wu Chen for helpful suggestions.
(QUASI-)POISSON ENVELOPING ALGEBRAS
17
References
[1] Y.H. Bao and Y. Ye, On quasi-Poisson cohomology, in preparation.
[2] W. Crawley-Boevey, P. Etingof and V. Ginzburg, Noncommutative geometry and quiver alge-
bras. Adv. Math. 209 (2007) 274-336.
[3] J.M. Casas and T. Pirashvili, Algebras with bracket. Manuscripta Math. 119 (2006) 1-15.
[4] D. Farkas, Modules for Poisson Algebras. Comm. in Algebra 28 (7) (2000) 3293-3306.
[5] M. Flato, M. Gerstenhaber and A.A. Voronov, Cohomology and deformation of Leibniz pairs.
Lett. Math. Phys. 34 (1995), 77-90.
[6] D. Farkas and G. Letzter, Ring Theory from Symplectic Geometry. Journal of Pure and Applied
Algebra 225(1998) 255-290.
[7] V. Ginzburg and M. Kapranov, Koszul duality for operads. Duke Math. J. 76 (1) (1994) 203-272.
[8] M. Kontsevich, Formal (non)commutative symplectic geometry. The Gelfand Math. Seminars,
1990-1992, Birkhauser, Boston, MA, 173-187.
[9] F. Kubo, Finite-dimensional Simple Leibniz Pairs and Simple Poisson Modules. Lett. Math.Phys.
43 (1) (1998) 21-29.
[10] F. Loose, Symplectic algebras and Poisson algebras. Commun. Algebras 21 (7) (1993) 2395-2416.
[11] Sei-Qwon Oh, Poisson Enveloping Algebras. Comm. Algebra 27 (5) (1999) 2181-2186.
[12] N. Reshetikhin, A. A. Voronov and A. Weinstein, Semiquantum geometry. J. Math. Sci. 82 (1)
(1996) 3255 -- 3267.
[13] M.E. Sweedler, Hopf algebras, W. A. Benjamin, Inc., New York, 1969.
[14] I. Vaisman, Lectures on the geometry of Poisson manifolds, Progress in Math., Vol. 118,
Birkhauser, Basel, 1994.
[15] M. Van den Bergh, Double Poisson Algebras. Trans. Amer. Math. Soc. 360 (11) (2008) 5711 --
5769.
[16] P. Xu, Noncommutative Poisson algebras. Amer. J. Math. 116 (1) (1994) 101-125.
[17] Y. Yao, Y.Ye and P.Zhang, Quiver Poisson Algebras. J. Algebra 312 (2) (2007), 570 -- 589.
Department of Mathematics, Columbia University, New York, New York 10027, USA;
E-mail address: [email protected]
Department of Mathematics, The University of Texas, 1 University Station C1200,
Austin, TX 78712, USA;
E-mail address: [email protected]
Yu Ye, Department of Mathematics, University of Science and Technology of China,
Hefei 230026, Anhui, PR China.
E-mail address: [email protected]
|
1806.04211 | 1 | 1806 | 2018-06-08T10:33:55 | A parallel algorithm for Gaussian elimination over finite fields | [
"math.RA",
"cs.DC",
"math.NA",
"math.GR"
] | In this paper we describe a parallel Gaussian elimination algorithm for matrices with entries in a finite field. Unlike previous approaches, our algorithm subdivides a very large input matrix into smaller submatrices by subdividing both rows and columns into roughly square blocks sized so that computing with individual blocks on individual processors provides adequate concurrency. The algorithm also returns the transformation matrix, which encodes the row operations used. We go to some lengths to avoid storing any unnecessary data as we keep track of the row operations, such as block columns of the transformation matrix known to be zero. The algorithm is accompanied by a concurrency analysis which shows that the improvement in concurrency is of the same order of magnitude as the number of blocks. An implementation of the algorithm has been tested on matrices as large as $1 000 000\times 1 000 000$ over small finite fields. | math.RA | math |
A parallel algorithm for Gaussian elimination over
finite fields
Stephen Linton, Gabriele Nebe, Alice Niemeyer, Richard Parker, Jon Thackray
June 13, 2018
School of Computer Science, University of St. Andrews, St. Andrews,
Fife KY169SX, Scotland E-mail address: [email protected]
Lehrstuhl D für Mathematik, RWTH Aachen University, 52056 Aachen,
Germany E-mail address: [email protected]
Lehrstuhl B für Mathematik, RWTH Aachen University, 52056 Aachen,
Germany E-mail address: [email protected]
70 York St. Cambridge CB1 2PY, UK
E-mail address: [email protected]
E-mail address: [email protected]
Abstract
In this paper we describe a parallel Gaussian elimination algorithm for matrices
with entries in a finite field. Unlike previous approaches, our algorithm subdivides
a very large input matrix into smaller submatrices by subdividing both rows and
columns into roughly square blocks sized so that computing with individual blocks
on individual processors provides adequate concurrency. The algorithm also returns
the transformation matrix, which encodes the row operations used. We go to
some lengths to avoid storing any unnecessary data as we keep track of the row
operations, such as block columns of the transformation matrix known to be zero.
The algorithm is accompanied by a concurrency analysis which shows that the
improvement in concurrency is of the same order of magnitude as the number of
blocks. An implementation of the algorithm has been tested on matrices as large
as 1000000 × 1000000 over small finite fields.
Keywords: Gaussian elimination, parallel algorithm, finite fields
MSC 2010:15A06, 68W10, 68W05
1
1 Introduction
Already employed for solving equations by hand in China over 2000 years ago [9], the
Gaussian elimination method has become an invaluable tool in many areas of science.
When computing machines became available, this algorithm was one of the first to be
implemented on a computer. In general the algorithm takes as input a matrix with entries
in a field (or a division ring) and transforms this matrix to a matrix in row echelon form.
It can be employed for several different purposes, and computer implementations can be
tailored to suit the intended application. Different variants of the Gaussian elimination
algorithm can be envisaged, for example computing the rank of a matrix, computing a
row echelon form or a reduced row echelon form of a matrix, or computing one of these
echelon forms together with the transformation matrix. Often one of these versions
of the Gaussian elimination algorithm lies at the heart of other algorithms for solving
problems in a broad range of areas and their overall performance is often dictated by the
performance of the underlying Gaussian elimination algorithm. Thus an implementation
of a Gaussian elimination algorithm is required to display exceptional performance.
Since their invention, computers have become faster and more powerful every year.
Yet, for over a decade this increase in computing power is no longer primarily due to
faster CPUs but rather to the number of different processors an individual computer has,
paired with the increasingly sophisticated memory hierarchy. It is therefore paramount
that modern algorithms are tailored to modern computers.
In particular this means
that they need to be able to perform computations in parallel and store the data for the
current computations readily in cache.
With the advance of parallel computers comes the need to design a parallel algo-
rithm to perform Gaussian elimination on a matrix. Such a parallel algorithm would
immediately result in immense speedups of higher level algorithms calling the Gaus-
sian elimination without having to introduce parallelism to these algorithms themselves.
When designing a parallel Gaussian elimination algorithm it is important to keep the ap-
plications of the algorithms in mind. Several versions of a parallel Gaussian elimination
algorithm have been described when working over the field of real or complex numbers,
see [6] for a survey and PLASMA [11] for implementations. In this paper we describe a
parallel version of the Gaussian elimination algorithm which, given a matrix with entries
in a finite field, computes a reduced row echelon form together with the transformation
matrix. We note that when working with dense matrices over finite fields we are not
concerned with sparsity, the selection of suitable pivot elements nor numerical accuracy.
In particular, we can always choose the first non-zero element of a given row as our pivot
element, ensuring that all entries to the left of a pivot are known to be zero. Moreover,
we need not be concerned with producing elements in a field which require more and
more memory to store them. Avoiding producing very large field elements would again
complicate pivot selection. Thus our main concern is to design a parallel algorithm which
makes optimal use of modern parallel computers.
We assume an underlying shared memory computational model in which we have
access to k different processors, each of which can run a job independently from any
2
other. The processors communicate with each other through the shared memory system.
Our design must take account of the limited total memory bandwidth of the system. The
aim of our parallel algorithm is to achieve adequate concurrency by dividing the necessary
computational work into smaller jobs and scheduling these to run simultaneously on the
k processors. This in turn calls for a very careful organization of the jobs so that different
jobs do not interfere with each other. We will address these issues in Section 2.1.
It is well known, see for example [5, Theorems 28.7, 28.8], that the asymptotic com-
plexities of matrix inversion and matrix multiplication are equal. An algorithm that
shows this reduces inversion of a 2n × 2n matrix to 6 multiplications of n × n matrices
together with two inversions, also of n× n matrices. Applying this approach recursively,
almost all of the run-time of inversion is spent in multiplications of various sizes. It is
not difficult to see that this extends to our somewhat more general computation of a
reduced row echelon form, with transformation matrix.
We envisage that we are given a very large matrix over a finite field for which we
need to compute a reduced row echelon form together with a transformation matrix.
Several approaches to achieving this already exist for finite fields. One approach takes
advantage of the reduction to multiplication mentioned above, and delegates the problem
of parallelising the computation primarily to the much easier problem of parallelising the
larger multiplications. Another approach represents finite field elements as floating point
real numbers of various sizes in such a way that (with care) the exact result over the
finite field can still be recovered. The problem can then be delegated to any of a number
of highly efficient parallel floating point linear algebra systems [7]. Another approach by
Albrecht et al. works over small finite fields of characteristic 2 (see [2] and [3]). A fourth
approach in unpublished work by Lübeck repeatedly divides the matrix horizontally,
echelonising each block of rows in parallel and then sorting the rows of the matrix, so
that rows with similar length initial sequences of zeros come together.
Large modern computers typically have a large number of cores but may well have
an order of magnitude less real memory bandwidth per core than a typical laptop or
desktop computer. On such a large modern computer, at the lowest level, a core can
only be fully occupied if essentially all its data is in the smallest, fastest level of cache
memory (L1). At the next level out, this work can only be started if essentially all its
data is in the next level of cache (L2). A similar statement is true for L3 cache. To use a
modern computer effectively for matrix operations, it is therefore necessary to repeatedly
subdivide matrices in both directions producing matrices that are roughly square at a
scale commensurate with the size of the cache at each level.
It is not too hard to design an algorithm for matrix multiplication with these prop-
erties. An approach along these lines to Gaussian elimination for these roughly square
submatrices is described in Section 2.3.2.
For the whole matrix, we also need to subdivide to achieve concurrency.
Unlike previous approaches, we subdivide our very large input matrix into smaller
submatrices, called blocks, by subdividing both rows and columns into roughly square
blocks sized so that computing with individual blocks on individual processors provides
adequate concurrency. We will show that we gain a concurrency improvement in the
3
same order of magnitude as the number of blocks, see Proposition 5.3 and Theorem 5.4.
As well as computing the reduced row echelon form of the input matrix, we compute
the transformation matrix, which encodes the row operations used. We go to some
lengths to avoid storing any unnecessary data as we keep track of the row operations, such
as block columns of the transformation matrix known to be zero. Our experiments show
that the memory usage during the execution of the algorithm remains fairly stable, and is
similar to storing the input matrix. The runtime is broadly comparable to multiplication
of matrices to the same size as the input matrix. This gives evidence that we have
succeeded in keeping the cost of computing the transformation matrix as small as possible
and is in accordance with the theoretical analysis of Bürgisser et al.
in particular [4,
Theorem 16.12].
The parallel Gauss algorithm is designed with three distinct environments in view,
although we have only implemented the first so far.
The first (and original) target is to use a single machine with multiple cores with the
matrix in shared memory. Here the objective is to subdivide the overall task into many
subtasks that can run concurrently.
The second target is to use a single machine where the matrix does not fit in memory,
but does fit on disk. Here the objective is to subdivide the matrix so that each piece fits
into memory.
The third target is where several computers (probably each with multiple cores) are
to work on a single problem simultaneously. Here the objective is again to subdivide the
overall task into many subtasks that can run concurrently on different computers with
access to the same central disk.
2 Preliminaries
2.1 Computational Model
Our general approach to parallel computing is to decompose the work to be done into
relatively small units with explicit dependencies. Units all of whose dependencies are
met are called "runnable" and a fixed pool of worker threads carry out the runnable units,
thereby discharging the dependencies of other units and making them in turn runnable.
A module, called the scheduler, is charged with keeping track of the dependencies and
scheduling the execution of the individual units. Data describing the units of work and
their input and output data reside in a shared data store, but we take considerable care
to ensure that the speed of this store is not critical to overall performance. It can thus
be the large amount of shared DRAM on a multicore server, or disk storage local to a
single server (for the case where we only have one server, but data too large for its RAM)
or shared disk provided sufficiently strong consistency can be guaranteed.
In our implementations, we have used two variations of this model. In one, the task-
model, the units of work are tasks and are relatively coarse-grained. A task represents,
roughly speaking, all the work that needs to be done in a particular part of the matrix at
a particular stage of the algorithm. Dependencies are between tasks, so one task cannot
4
execute until certain others have completed and it will find the data it needs where
those previous tasks have stored it. To simplify understanding, we collect different data
into data packages, the input and output of the tasks. For example a typical output of
the task ClearDown is the data package A = (A, M, K, ρ(cid:48), E, λ) with six components
which we refer to as A.A, A.M, etc.
In the other model, the job-model, the units of work are the jobs and are significantly
more fine-grained and represent a single elementary computation such as a submatrix
multiplication. More importantly, the dependencies are between the jobs and the data
they produce and consume. Each job requires zero or more input data objects and
produces one or more output objects. A job is runnable when all of its input data has
been produced. This finer grain approach allows more concurrency. Further gain in
efficiency can be achieved by giving the scheduler guidance as to which jobs or tasks
are urgent and which are less so. This aspect is mainly ignored in this paper. The
implementation of Meataxe64 [10] uses the job-model and identifies the urgent jobs. An
implementation in HPC-GAP by Jendrik Brachter and Sergio Siccha is based on the
task-model.
The parallel Gaussian elimination algorithm is described in Section 3 as a program
called the Chief charged with defining the tasks and the data packages they work on.
The Chief is described as a sequential program but the reader should be warned that
the tasks are executed in an unpredictable order which may bear little resemblance to
the order in which the tasks are submitted by the Chief. The result of running the
Chief is a plan consisting of a list of tasks, respectively jobs, together with their inputs
and outputs whose collective execution performs the Gaussian elimination.
Below we specify The Chief in terms of tasks, specified in turn by jobs, for which
the reader will find a more or less specific description in Section 4.
2.2 Gaussian elimination
This subsection describes the Gaussian elimination process in a way we hope is familiar
to the reader, but in our notation. Given a matrix H with entries in a field F the output
of the Gauss algorithm consists of matrices M, K and R and permutation matrices Pρ
and Pγ such that
PρHPγ =
.
(1)
(cid:18)M 0
(cid:19)
K 1
(cid:18)−1 R
(cid:19)
0
0
The matrices Pρ and Pγ perform row, respectively column, permutations on the input
matrix H such that the top left-hand part of the resulting matrix PρHPγ is invertible
(with inverse −M) with the same rank as H. It should be noticed that the permutation
matrices Pρ and Pγ in Equation (1) are not uniquely defined. All that matters is that
they put the pivotal rows and columns into the top left-hand corner of the matrix. We
therefore choose to only specify the sets of row and column numbers containing pivotal
elements. As we are chopping our matrix into blocks, we use the word selected to specify
a row or column in which a pivot has been already been found. Hence we have to apply
a permutation to the rows during the course of the algorithm to ensure that our pivots
5
remain located in columns with increasing indices. We formalize how we store these
permutation matrices in the Definition 2.1.
Definition 2.1. When we enumerate elements of a set, we always implicitly assume that
these are in order. To a subset ρ = {ρ1, . . . , ρρ} ⊆ {1, . . . , α} associate a 0/1 matrix
(cid:40)
ρ ∈ Fρ×α with ρi,j =
1
0
j = ρi
else.
We call ρ the row-select matrix and ρ the row-nonselect matrix associated to the set
ρ and its complement ρ = {1, . . . , α} \ ρ.
Remark 2.2. Note that the matrix
(cid:18) ρ
ρ
(cid:19)
∈ Fα×α
Pρ =
is a permutation matrix associated to the permutation pρ with pρ(i) = ρi for i ≤ ρ and
pρ(i) = ρi−ρ for i > ρ. Note that pρ = 1 if and only if ρ = {1, . . . , i} for some 0 ≤ i ≤ α
but for the other subsets ρ we may recover ρ from pρ as the image {pρ(1), . . . , pρ(i)}
if pρ(i + 1) < pρ(i).
In this sense we keep switching between permutations, subsets,
and bitstrings in {0, 1}α which represent the characteristic function of the subset. In
particular, this means that a subset of cardinality 2α − α of special permutations of all
the α! permutations of {1, . . . , α} is sufficient for all our purposes.
We also note that for a matrix H ∈ Fα×β and ρ ⊆ {1, . . . , α} and γ ⊆ {1, . . . , β} the
matrix ρ × H ∈ Fρ×β consists of those rows of H whose indices lie in ρ (retaining the
ordering) and the matrix H × γtr ∈ Fα×γ consists of those columns of H whose indices
lie in γ. Therefore we also call γtr the column-select matrix and γtr the column-nonselect
matrix associated to γ.
Remark 2.3. Let H ∈ Fα×β be of rank r. Then the echelonisation algorithm will
produce sets ρ ⊆ {1, . . . , α} and γ ⊆ {1, . . . , β} of cardinality γ = ρ = r and matrices
M ∈ Fr×r, K ∈ F(α−r)×r, R ∈ Fr×(β−r) such that
(cid:18) M 0
(cid:19)(cid:18) ρ
(cid:19)
H(cid:0) γtr γtr (cid:1) =
(cid:18) −1 R
(cid:19)
0
0
K 1
ρ
We will refer to these matrices as
(M, K, R, ρ, γ) := ECH(H).
Strictly speaking, the job ECH computes the negative row reduced echelon form of
the input matrix H. However we will simply call this the echelon form of H.
We assume we have an implementation of ECH suitable for use on a single core. The
goal of this paper is to show how to scale it up.
6
Example
We now present an example to highlight some of the structure of the algorithm. Consider
the matrix C ∈ F6×6
and divided into four blocks as shown. Our first step is to echelonise the top left block.
This yields
3
given by
P{1,3}
0 2 0
1 0 0
P{1,3} =
0 0 1
0 1 0
0 2 2 0 1 0
0 2 2 1 2 2
1 0 1 0 2 1
2 0 1 0 2 2
0 1 1 2 1 1
1 2 2 2 0 0
0 2 2
P{1,2} =
2 0 2
and P{1,2} =
1 0 0
,
(cid:19)
(cid:18)0 2
, K =(cid:0)2 0(cid:1), and R =
0 2 2
0 0 0
0 1 0
0 0 1
0 2 2
1 0 1
(cid:18)2
(cid:19)
1 0 0
2 0 1
matching Equation 1. Here
where the bars separate selected from non-selected rows or columns. So the outputs
from this step are the multiplier M =
as well as the
row-select set ρ, selecting the first and third row and the column-select set γ, selecting
the first two columns.
1 0
2
0 2 0
Mathematically, we need to left multiply the top right block by
We take advantage of the known structure of this matrix to speed up this computation
as follows. We divide the top right block into the selected rows, as described by ρ:
After this, two further steps are available, our parallel implementation will do both
concurrently. We mimic the row transformations, applied to the top left block, on the
top right block. We also use the echelonised top left block to clean out some columns in
the block beneath it. We explain them in that order.
P{1,3}.
(cid:18)0 1 0
(cid:19)
and the rest(cid:0)1 2 2(cid:1). We add the product of K and the selected rows to
the non-selected rows, giving(cid:0)1 1 2(cid:1), and then multiply the selected rows by M giving
(cid:18)0 1 2
(cid:19)
(cid:18) 2 0 2
The other step requires us to add multiples of the pivot rows from the echelonised top
to the bottom left block, so as to clear the pivotal columns. We
1 0 0
2 0 1
left block
0 1 0
0 2 1
(cid:19)
.
0 2 2
7
again take advantage of the known structure of this matrix to speed up this computation
as follows. We divide the bottom left block into the selected columns, as described by γ:
1
. Now we add the product of the selected
and the non-selected columns
2 0
0 1
1 2
columns and R to these non-selected columns and obtain
0
2
1
2
2
.
2 0
and
0 1 0
.
.
2 2 1
2 0 2
0 1
1 2
2 0 2 0 1 2
0 2 2 0 1 0
0 0 0 1 1 2
0 0 2 0 1 0
0 0 0 2 2 1
0 0 2 2 0 2
We must now mimic the row transformations used to clear the pivotal columns of
the bottom left block by adding multiples of rows from the top right hand block to the
bottom right hand block, i.e. we add the product of
to the
(cid:18)0 1 2
(cid:19)
0 1 0
bottom right hand block, which becomes
After all these steps, the overall matrix is
At this stage we have dealt with all the consequences of the pivots found in the top
left block. What remains to be done is to echelonise the top-right and bottom-left
submatrices in the picture above and deal with the consequences of any pivots found.
Finally part of the bottom right hand block will need to be echelonised. This example,
chopped 2 × 2 does not demonstrate pivotal row merging as described in Section 2.3.1
2.3 Some guiding points
2.3.1 Subdividing a huge matrix
Our parallel version of the Gaussian elimination algorithm takes as input a huge matrix
and subdivides it, both horizontally and vertically, into blocks. We do this partly to
obtain concurrency and partly to reduce the row length for cache reasons. Once the top-
left block has been echelonised, the same row operations can be applied simultaneously
to all the blocks along the top row. Putting the rest of the left-most block column into
echelon form requires addressing the blocks sequentially from the top down. However,
in the common case where the co-rank of the top-left block is small, it is not a great
deal of computation. Once the top block row and left-most block column are in echelon
form, we can update each of the blocks in the other rows and columns concurrently.
8
It should be remarked that nothing else can happen until this first block is echelonised,
suggesting that we should make this block smaller than the rest. A similar comment
applies to the last (bottom-right) block.
Proceeding down the left-most block column sequentially enables us to merge into a
single row of blocks all those rows whose pivots lie there. This merging is done to reduce
the amount of data access. Usually, the work of doing this merging is not great, so that
soon after the first block echelonisation is complete, a large block multiply-and-add can
be done to every block of the matrix. Without the merging, this multiply-and-add would
be done piecemeal, requiring multiple passes through the block.
2.3.2 Echelonisation of a single block
The performance of the echelonisation of a single block (as defined in Remark 2.3) can
have a considerable impact on the concurrency, as many later jobs may depend on each
echelonisation.
The sequential algorithm used to echelonise individual blocks is recursive, combin-
ing elements of the recursive inversion algorithm already mentioned, with the greater
generality of the technique described in this paper. A block is divided into two, either
horizontally or vertically, and the top (resp. left) part is echelonised. Using a simplified
version of ClearDown (resp. UpdateRow), see Sections 4.2.1 and 4.2.2, the remain-
der of the matrix and the transformation matrix are updated, producing a second block
which must also be echelonised. The results of the two echelonisations can be combined
to compute the data package consisting of M, K, R, Pρ and Pγ. Using this technique
recursively for all matrices bigger than a threshold size (about 300 dimensions) and a
simple direct Gaussian elimination algorithm below this size, echelonisation of a block
takes essentially the same time as a block multiply.
2.3.3 Blocks change size and shape
In our description of the algorithm, especially in Equations (3) and (4), we imagine that
rows are permuted so that those containing pivots (in blocks to the left) are moved to
the top, and the rest are moved to the bottom.
In the program, however, these two
sets of rows are held in different matrices, but it seemed better to try to include all the
information in one matrix, attempting to clarify the relationships between the parts.
As a consequence of moving rows about, the sizes and shapes of the blocks change
during the course of the algorithm. We use a superscript to indicate the "stage" of a
ik is the matrix block at the (i, k)-position in
particular block, so that, for example, Cj
In some ways the original input matrix C (see Equation (2) below) is
its j'th stage.
gradually converted into the matrix R. An intermediate matrix B collects the pivotal
rows from C which are subsequently deleted from C.
9
2.3.4 Riffles
Our subset-stored permutations are used both to pull matrices apart and to merge them
together. We call the pulling apart an 'extract' where one matrix is separated into two
matrices with the selected rows (or columns) going to one, and the remaining rows to the
other. We call the merging a 'riffle' where one matrix is assembled from two inputs with
the subset-stored permutation directing from which input each row (or column) comes.
2.3.5 Transformation matrix abbreviation
To compute the transformation matrix, we could apply the same row operations we
performed on the input matrix to the identity matrix. In practice this would be wasteful,
both of memory and computational effort, since initially the identity matrix contains no
information at all, and the early row operations would mainly be adding zeros to other
zeros. Considerable effort has been expended in this paper to avoid storing any parts
of matrices whose values are known a priori, thereby saving both memory, and work
manipulating them. The graphs shown in Section 6 suggest that this has been successful.
The details of this optimisation are the subject of Section 4.2.3, which may be skipped
on first reading.
Note that although the output matrix M in Equation 2 below is square, the number
of blocks in each row may differ from the number of blocks in each column, since the
block rows are indexed by the block column in which the pivot was found and vice versa.
3 A parallel Gauss algorithm
3.1 The structure of the algorithm
We now describe a parallel version of the Gaussian elimination algorithm which takes as
input a huge matrix and subdivides it, both horizontally and vertically, into blocks.
To distinguish between the huge matrix and its blocks we use different fonts.
Let F be a finite field and C ∈ Fm×n a huge matrix of rank r. We describe a
parallel version of the well-known Gauss algorithm, which computes matrices R, a
transformation matrix T =
and subsets ⊆ {1, . . . m} and Υ ⊆
{1, . . . , n}, both of cardinality r, such that
K 1(m−r)×(m−r)
(cid:19)
(cid:18)M 0r×(m−r)
(cid:19)
C(cid:16)
(cid:19)(cid:18)
(cid:18)M 0r×(m−r)
K 1(m−r)×(m−r)
tr(cid:17)
(cid:18)−1r×r R
(cid:19)
0
0
Υtr Υ
=
(2)
is in (negative row reduced) echelon form.
Comparing to Remark 2.3, we see that our task is to chop our huge input matrix
into smaller blocks and then, using ECH and other jobs on the blocks, to effect the
same operation on the huge matrix as ECH does on a single block. We therefore choose
10
positive integers ai, bj such that
a(cid:88)
ai = m,
b(cid:88)
bj = n
i=1
j=1
and our algorithm copies the block-submatrices of the input matrix C into data packages
Cij ∈ Fai×bj, called blocks, and performs tasks (as described in Section 4) on these smaller
blocks.
We call the a matrices C1j, . . . , Caj the j-th block column and the b matrices Ci1, . . . , Cib
the i-th block row of C. Echelonising the overall matrix C is achieved by performing an
echelonisation algorithm on individual blocks and using the resulting data to modify
others.
The result of the Gaussian elimination as well as the intermediate matrices are par-
titioned into blocks: When the Gauss algorithm has completed, the matrix R, which
occurs in Equation (2), consists of blocks and has the form
R(cid:48)
R1 R(cid:48)
. . .
R(cid:48)
0 R2 R(cid:48)
...
...
...
0 Rb−1 R(cid:48)
0
b−1,b
0
Rb
. . .
j=1 rj is the rank of C and for
j=1 rj is the rank of the submatrix of C consisting of the first k
block columns.
The time-consuming parts of the Gaussian elimination algorithm consist of Step 1
and Step 3, whereas the intermediate Step 2 is not. After the first step the matrix C has
been transformed into an upper triangular matrix and prior to permuting columns the
matrix (−1r×rR) ∈ Fr×n has the shape
jk ∈ Frj×(bk−rk). Here r =(cid:80)b
with Rj ∈ Frj×(bj−rj ) and R(cid:48)
k = 1, . . . , b the sum(cid:80)k
...
. . .
. . .
R =
. . .
. . .
...
0
(3)
1b
2b
12
23
R =
−1 R2 X23 R23
−1 R1 X12 R12
0 0
...
0 0
0 0
...
. . .
. . .
. . .
...
0 0
. . .
. . .
. . .
...
0 0
,
X1b R1b
X2b R2b
...
−1 Rb
−1 Rb−1 Xb−1,b Rb−1,b
(4)
where Rjk ∈ Frj×(bk−rk) and Xjk ∈ Frj×rk. To simplify notation in the algorithms
below we define the data packages Bjk = (XjkRjk) ∈ Frj×bk for 1 ≤ j < k ≤ b and
Dj = (Dj.R, Dj.γ), where Dj.R = Rj and Dj.γ ⊆ {1, . . . , bj} is the set of indices of
pivotal columns in the block column j.
3.1.1 Computing the transformation matrix
During the course of the algorithm we also compute the transformation matrix T ∈
GLm(F) as given in Equation (2). Of course this could be achieved by simply performing
11
the same row operations on an identity matrix that were performed on C. This involves
considerable work on blocks known to be either zero or identity matrices. For example,
if C is invertible, the work required to compute its inverse is needlessly increased by
50%. To avoid such extra work, we only store the relevant parts of the blocks of the
transformation matrix.
During the computation, the data packages Mji (j = 1, . . . , b; i = 1, . . . , a) and Kih
(i, h = 1, . . . , a) record the matching status of the transformation matrix. If Cik denotes
the i, k-block of the original input matrix C, then initially Cik = Cik and Bjk has no
rows. Likewise, initially Kih is the identity matrix if i = h and the zero matrix otherwise
and the matrix Mji has no rows and no columns. When storing the data packages K
and M, we omit the columns and rows in the blocks that are zero or still unchanged
from the identity matrix.
To obtain the row select matrix we maintain further data packages Eij = (Eij.ρ, Eij.δ)
(1 ≤ i ≤ a, 1 ≤ j ≤ b) with Eij.ρ ⊆ {1, . . . , ai} is the set of indices of pivotal rows in
block row i with pivot in some block column 1, . . . , j and Eij.δ ∈ {0, 1}Eij .ρ records
which indices already occurred up to block column j − 1.
During the algorithm, having handled block row i in Step 1, we have
a(cid:88)
a(cid:88)
h=1
Mjh × Ehi.ρ × Chk = Bjk
K(cid:96)h × Ehi.ρ × Chk = C(cid:96)k.
and
The final column select matrix is the block diagonal matrix
h=1
Υ = diag(Da
1.γ, . . . , Da
b .γ).
and the row select matrix is
= diag(E1b.ρ, . . . , Eab.ρ).
3.2 Step 1
Step 1 loops over the block rows. For the j-th column of the i-th block row, the algorithm
calls Task ClearDown with the two data packages Cij and Di−1
as input. Task
ClearDown amalgamates the pivots in Di−1
.γ with the pivots in the matrix Cij to
j.R of the (negative) echelon
produce the enlarged set Di
form (−1 Di
j.R) followed by 0 rows (up to column permutations which are remembered
j.γ). Moreover, the task ClearDown records in its output data package Aij the
in Di
row operations performed. With the help of these data packages, the first step then
propagates the same elementary row operations to the remaining blocks in block row i
as well as to block row i of the transition matrix using Task UpdateRow.
j.γ as well as a new matrix Di
j
j
12
Hence Step 1 assembles in the block row(cid:0)0 . . . 0 −1Dj.R Bj,j+1
and the number of rows of C is m −(cid:80)j
rows of the original input matrix whose pivotal entries lie in block column j for j ≤ b.
These rows are then deleted from the data package C. Thus having treated the j-th
block column the matrix C contains no rows whose pivots lie in block columns 1, . . . , j
(cid:1) the
. . . Bjb
k=1 rk.
In particular, during the course of the entire algorithm the block rows Ci,− con-
tain fewer and fewer rows, whereas the number of rows of the block row Bj,− increases
accordingly. After completing the block column j the matrices Bjk remain stable.
Similarly for the transformation matrix, the matrix M gains rows whereas K loses
rows. However, things here are slightly more complicated due to the fact that we do
not store the full transformation matrix. As we only store columns of Kih that are not
known a priori to be zero or columns of an identity matrix, we have to ensure that all
the needed columns are present when calling UpdateRowTrafo. The columns that
are not yet present are precisely the columns that correspond to the positions of the
pivot rows and are stored in Ehj.δ. If i = h this means we need to insert into the correct
positions columns of an identity matrix and if i (cid:54)= h then columns of a zero matrix. To
achieve this, and also to efficiently deal with the cases where the matrices Kih or Mjh
are to be initialized we adapted the task UpdateRow to obtain UpdateRowTrafo.
3.3 Step 2
This intermediate step becomes necessary as we do only store the relevant parts of the
transformation matrices Mji. Before the upwards cleaning in Step 3 we need to riffle in
zero columns in Mji so that the number of columns in Mji is equal to the number of
columns in Mbi for all j.
3.4 Step 3
Then back cleaning only performs upwards row operations on the matrix from Equation
(4) to eliminate the Xjk. The matrices Rjk from Equation (4) are stored in the data
packages Rjk in the Chief. Having cleaned block columns b, . . . , k − 1 the algorithm
adds the Xjk multiple of block row k to block row j for all j ≤ k−1 to clear block column
k. The same row operations are performed on the relevant part M of the transformation
matrix.
13
Algorithm 1: The Chief
Input : C =: (C1
Output: R = (Rk−j
ik)i=1,...,a,k=1,...,b, where C1
jk )j≤k=1,...,b, M = (M2a
ik ∈ Fai×bk
jh)j=1,...,b,h=1,...,a, K = (Ka
select matrix ⊆ {1, . . . , m}, the concatenation of the
Eib.ρ ⊆ {1, . . . , ai} (i = 1, . . . , a), and a column select matrix
Υ ⊆ {1, . . . , n}, the concatenation of the Da
j .γ ⊆ {1, . . . , bj}
(j = 1, . . . , b), such that
(cid:19)(cid:18)
(cid:18)M 0K 1
(cid:19)
C(cid:0)Υ Υ(cid:1) =
(cid:18)−1 R
(cid:19)
0
0
ih)i,h=1,...,a, a row
Step 1:
for i from 1 to a do
for j from 1 to b do
j; Aij) := ClearDown(Cj
(Di
Eij :=Extend(Aij, Ei,j−1, j);
for k from j + 1 to b do
ij, Di−1
j
, i);
(Cj+1
ik , Bi
jk) := UpdateRow(Aij, Cj
ik, Bi−1
jk , i);
for h from 1 to i do
(Kj+1
ih , Mi
jh) :=UpdateRowTrafo(Aij, Kj
ih, Mi−1
jh , Ehj, i, h, j);
Step 2:
for j from 1 to b do
for h from 1 to a do
Ma+1
jh
:=RowLengthen(Ma
jh, Ehj, Ehb);
Step 3:
for k from 1 to b do
kk :=Copy(Da
k);
R0
for k from b downto 1 do
for j from 1 to k − 1 do
jk) := PreClearUp(Ba
(Xjk, R0
for (cid:96) from k to b do
R(cid:96)−k+1
j(cid:96)
:= ClearUp(R(cid:96)−k
j(cid:96)
k);
jk, Da
, Xjk, R(cid:96)−k
k(cid:96) );
for h from 1 to a do
Ma+h
jh
:= ClearUp(Ma+h−1
jh
, Xjk, Ma+h−1
kh
);
14
4 Jobs and Tasks
4.1 The jobs
In this section we describe the jobs. These are fundamental steps that are later used to
define the tasks. Many of the jobs take as input one or more matrices. While the input
and output matrices of the jobs within the global context of the parallel Gauss algorithm
are blocks computed from a huge input matrix, the jobs described in this section work
locally only on these matrices. In current implementations, each job can be performed
by a single threaded computation, entirely in RAM and in a reasonable amount of time.
cpy This task simply copies the input matrix to the output.
mul This job performs a matrix multiplication. It takes as input two matrices A ∈ Fα×β
and B ∈ Fβ×δ and returns as output the matrix A × B ∈ Fα×δ.
mad This job performs a matrix multiplication followed by a matrix addition. It takes
as input matrices A ∈ Fα×δ and B ∈ Fα×β and C ∈ Fβ×δ and returns the matrix
A + B × C ∈ Fα×δ.
ech This job performs an echelonisation as described in Remark 2.3. We will refer to
the job as
(M, K, R, ρ, γ) := ECH(H).
cex This job performs two column extracts. It takes as input a matrix H ∈ Fα×β and
a subset γ ⊆ {1, . . . , β} and returns the matrices H × γtr and H × γtr, consisting
of all those columns of H whose indices lie in γ, respectively do not lie in γ, as
described in Definition 2.1.
rex This job performs two row extracts. It takes as input a matrix H ∈ Fα×β and a
subset ρ ⊆ {1, . . . , α} and returns the matrices ρ × H and ρ × H, consisting of all
those rows of H whose indices lie in ρ, respectively do not lie in ρ, as described in
Definition 2.1.
unh This job performs a union plus history. It takes as input a subset ρ1 ⊆ {1, . . . , α}
and, for α0 = α−ρ1, a subset ρ2 ⊆ {1, . . . , α0} and returns a subset ρ ⊆ {1, . . . , α}
defined as follows. Write {1, . . . , α} \ ρ1 = {x1, . . . , xα0}. Define
ρ = ρ1 ∪ {xi i ∈ ρ2} =: {y1, . . . , yr}
as an ordered set. Then u ∈ {0, 1}r with u(cid:96) = 0 if y(cid:96) ∈ ρ1 and u(cid:96) = 1 otherwise.
We refer to this job as
(ρ, u) := UNH(ρ1, ρ2).
un0 This job does the same as unh except that the first input set of unh is omitted
and assumed empty.
15
mkr It takes two sets ρ1 ⊆ ρ2 and produces a bitstring λ ⊆ {0, 1}ρ2 with 1s corre-
sponding to the elements in ρ1 and 0s corresponding to the elements in ρ2\ρ1.
rrf This job performs a row riffle. The input consists of a bit string u ∈ {0, 1}r and
two matrices B ∈ Fα×β and C ∈ Fγ×β with α + γ = r, where the number of 0s in
u is α and the number of 1s in u is γ. The job returns the new matrix A ∈ Fr×b
whose rows are the rows of B and C combined according to u. In some sense this
is the inverse of row extract.
crz Similarly to row riffles we also need column riffles, but we only need to riffle in zero
columns.
adi It takes as input a matrix K ∈ Fα×β and a bitstring δ ∈ {0, 1}β and puts Ki,ji := 1
if ji is the position of the ith 0 in δ. Note that combining crz with adi allows us
to riffle in columns of the identity matrix.
4.2 The tasks
We now describe the tasks on which our Gaussian elimination algorithm depends. As
mentioned above, a task receives data packages as input, which in turn may consist of
several components, and returns data packages as output.
Definition 4.1. A data package is a record of one or several components. A data package
is called ready (for a given scheduler) if the task that produces it as output has finished,
regardless whether this task has computed all its components. If there is no task having
this data package as an output, then we also consider it ready.
Example: Task UpdateRow.
If called with the parameter i = 1 the task UpdateRow can start, even though the com-
ponent A.A of the data package A has not been computed, after the task ClearDown
for i = 1 has completed. Note that for i = 1, no job in the task UpdateRow takes the
jk is an input to UpdateRow
component A.A as an input. Also the data package B0
but not computed by any task in the Chief. So therefore it is also considered ready.
Task 1: Extend
Input : A = (A, M, K, ρ(cid:48), E, λ), E = (ρ, δ) with ρ ⊂ {1, . . . , α}, δ a riffle, j
Output: E.
j = 1
j (cid:54)= 1
(E.ρ, E.δ) := UN0(A.ρ(cid:48));
(E.ρ, E.δ) := UNH(E.ρ, A.ρ(cid:48));
(UN0):
(UNH):
16
Task 2: RowLengthen
Input : M ∈ Fα×g1, E1.ρ ⊆ E2.ρ ⊆ {1, . . . , α} of sizes g1, g2 with g1 ≤ g2.
Output: M ∈ Fα×g2.
(MKR): λ := MKR(E1.ρ, E2.ρ);
(CRZ): M := CRZ(M, λ);
Task 3: ClearUp
Input : R ∈ Fα×β, X ∈ Fα×γ, M ∈ Fγ×β.
Output: R ∈ Fα×β.
(MAD): R := R + X × M;
Task 4: PreClearUp
Input : B ∈ Fα×β, D, with D.γ ⊆ {1, . . . , β} of cardinality g.
Output: X ∈ Fα×g, R ∈ Fα×(β−g).
(CEX): X := B × D.γtr; R := B × D.γ
tr;
Task 5: Copy
Input : D, with D.R ∈ Fα×β.
Output: R ∈ Fα×β.
(CPY): R := D.R;
4.2.1 Task ClearDown
Task ClearDown works on block columns. Suppose that j ∈ {1, . . . , b} and ClearDown
works on block column j which contains bj columns. Task ClearDown assumes that
block column j truncated after row i − 1 is in row echelon form and the aim of task
ClearDown is to replace the block column j truncated after row i by its row echelon
form.
Task ClearDown takes two data packages C and D as input. The first data package
C is the block Cij which is the block in the i-th block row of block column j. The second
data set D contains two data elements. The data element D.R is a matrix such that
block column j truncated after block row i−1 is in row echelon form (−1 D.R) followed
by 0 rows. The data element D.γ ⊆ {1, . . . , bj} contains indices of the pivots assembled
in block column j truncated after block row i − 1.
The task produces two data packages A and D as outputs. The data elements stored
in the data package A are required to propagate row operations performed during the
17
call to Task UpdateRow to other blocks in block row i. The data elements stored in
the data package D are required for a subsequent call to ClearDown for the block
Ci+1,j in block column j.
We begin by partitioning the input block C according to D.γ into pivotal and non
pivotal columns C = (A.A A(cid:48)). Using the rows of the matrix (−1 D.R) we can reduce
C to (0 H(cid:48)) where H(cid:48) = A(cid:48) +A.A×D.R. The next step is to call job ECH to echelonise
H(cid:48) and obtain
(A.M, A.K, R, A.ρ(cid:48), γ(cid:48)) := ECH(H(cid:48)),
where A.ρ(cid:48) is the set of pivotal rows of H(cid:48) and γ(cid:48) the set of pivotal columns.
As block column j truncated after block row i−1 is in row echelon form (−1r×r D.R)
followed by 0 rows, we now wish to determine a new remnant matrix R (which will become
the new D.R) such that block column j truncated after block row i is in row echelon
form (−1(r+r(cid:48))×(r+r(cid:48)) R) followed by 0 rows. To achieve this, the we have to add the r(cid:48)
pivots of H(cid:48) to −1r×r and reduce D.R according to (−1r(cid:48)×r(cid:48) R). This amounts to first
separating the columns of D.R into those containing pivot entries of H(cid:48) and those that
do not, i.e. writing D.R = (A.E R(cid:48)) with the help of the row select and row non-select
matrices γ(cid:48) and γ(cid:48). We then use the rows of the matrix (−1r(cid:48)×r(cid:48) R) to reduce D.R to
(0 R(cid:48) +A.E×D.R). The new set D.γ of all pivotal columns of block column j truncated
after block row i is now obtained by combining the old set D.γ and γ(cid:48). We record in
A.λ the information which of these indices came from the r(cid:48) pivots of H(cid:48). Finally, the
new remnant R is obtained by interleaving the rows of R(cid:48) + A.E× R with the rows of R
according to A.λ and storing the resulting matrix as the new D.R.
The following pseudo code details Task ClearDown:
Task 6: ClearDown
Input : C ∈ Fα×β, D.γ ⊆ {1, . . . , β} of cardinality r, D.R ∈ Fr×(β−r), i;
Output: D.R ∈ F(r+r(cid:48))×(β−r−r(cid:48)), D.γ ⊆ {1, . . . , β} of cardinality r + r(cid:48) and
A = (A, M, K, ρ(cid:48), E, λ) where A ∈ Fα×r, M ∈ Fr(cid:48)×r(cid:48), E ∈ Fr×r(cid:48),
K ∈ F(α−r(cid:48))×r(cid:48), ρ(cid:48) ⊆ {1, . . . , α − r} of cardinality r(cid:48), λ ∈ {0, 1}r+r(cid:48).
if i = 1 then
else
(ECH):
(A.M, A.K, D.R, A.ρ(cid:48), D.γ) := ECH(C);
tr;
(CEX): A.A := C × D.γtr; A(cid:48) := C × D.γ
(MAD): H := A(cid:48) + A.A × D.R;
(A.M, A.K, R, A.ρ(cid:48), γ(cid:48)) := ECH(H);
(ECH):
(CEX): A.E := D.R × (γ(cid:48))tr, R(cid:48) := D.R × (γ(cid:48))tr;
(MAD): R(cid:48) := R(cid:48) + A.E × R;
(UNH):
(RRF): D.R := RRF(A.λ, R(cid:48), R);
(D.γ, A.λ) := UNH(D.γ, γ(cid:48));
end
18
4.2.2 Task UpdateRow
Given i ∈ {1, . . . , a}, the Task UpdateRow works on block C = Cik in block row i and
block column k. It takes as input data packages A, C and B, where the data package
A encodes the necessary information computed by ClearDown when transforming an
earlier block in the same block row i into echelon form.
The same row operations that were performed on this earlier block now need to be
performed on C. This subroutine also assembles in the matrix B the rows in block
column k whose pivotal entry lies in block column j for j + 1 ≤ k ≤ b. The new
data package C returned by Tasks UpdateRow then is equal to the transformed input
matrix C with these rows deleted.
The following pseudo code details Task UpdateRow:
Task 7: UpdateRow
Input : A = (A, M, K, ρ(cid:48), E, λ), C ∈ Fα×β, B ∈ Fr×β, i.
Output: C ∈ F(α−r(cid:48))×β, B ∈ F(r+r(cid:48))×β.
(1)
i (cid:54)= 1
i = 1
always
always
i (cid:54)= 1
i (cid:54)= 1
i = 1
always
(MAD): Z := C + A.A × B;
(CPY): Z := C;
(REX): V := A.ρ(cid:48) × Z; and W := A.ρ(cid:48) × Z;
(MUL): X := A.M × V ;
(MAD): S := B + A.E × X;
(RRF): B := RRF(A.λ, S, X);
(CPY): B := X;
(MAD): C := W + A.K × V ;
(2)
(3)
(4)
(5)
(6)
Remark 4.2. In the case i = 1 in Task UpdateRow we work with the first block row.
Therefore we do not need to perform the upwards cleaning on the data package C and
the data package B is initialized accordingly. Note that for i = 1 the task ClearDown
did not compute the components A.A, A.E and A.λ and also the input data package B
of UpdateRow is not present.
4.2.3 Task UpdateRowTrafo
If one is not too concerned about performance, then it would be possible to generate
an identity matrix K and apply the UpdateRow task replacing C by K and B by M
to mimic the relevant row operations performed to obtain the transformation matrix M
and the cleaner matrix K. This would result in a lot of needless work performed on zero
or identity matrices. The main difference is that we never store any columns known to
belong to an identity or a zero matrix. Instead we insert these columns just before they
are needed. Moreover, we never add a matrix known to be zero or multiply by a matrix
known to be the identity.
As a result, we require a separate procedure, UpdateRowTrafo, to mimic the
row operations on the transformation matrix. UpdateRowTrafo still performs the
19
same steps as UpdateRow, identified by the same numbers, however it requires some
additional steps, indicated by the symbol + in the first column and which insert some
unstored columns into the matrix K. The various instances of a given step are due to
the fact that we can often avoid unnecessary work. In particular, UpdateRowTrafo
takes as an additional input the integers i, j, h, with h ≤ i. The integer i indicates the
current block row and h the current block column, on which to mimic the row operations
performed during UpdateRow on block row i. If j > 1 then we already computed some
input K into which we need to riffle in zero (if i (cid:54)= h) or the relevant columns of the
identity matrix (if i = h). It turns out that it is easier to always riffle in zero (the first
line marked with +) and mimic the special case i = h by adding the correct 0/1 matrix
to V later in the other line marked with +. If j = 1 then we should initialise K with
zero (if i (cid:54)= h) or the relevant columns of the identity matrix (if i = h). As we only need
the input K to define V and W in line (2), we mimic this by remembering that W = 0
in this case and V is either 0 (if i (cid:54)= h) or the identity matrix if i = h. So for j = 1 and
h = i we omit the multiplication by the identity in lines (3) and (6).
that h ≤ i, this only happens if h = i = 1.
If i = 1 again Remark 4.2 applies accordingly to line (4). Note that due to the fact
In the following pseudo code describing Task UpdateRowTrafo we indicate in the
last column which unstored matrices are implicitly known to be 0 or the identity 1.
20
Task 8: UpdateRowTrafo
Input : A = (A, M, K, ρ(cid:48), E, λ), K ∈ Fα×β, M ∈ Fr×β(cid:48), E = (ρ, δ), i, h, j.
Output: K ∈ F(α−r(cid:48))×(β+δ), M ∈ F(r+r(cid:48))×β(cid:48).
command
−
remark
K is 0
job
(CRZ): K := CRZ(K, E.δ);
(MAD): Z := K + A.A × M;
(MUL): Z := A.A × M;
(CPY): Z := K;
−
(REX): V := A.ρ(cid:48) × Z;
W := A.ρ(cid:48) × Z;
−
V := ADI(V, E.δ);
−
(ADI):
(MUL): X := A.M × V ;
(CPY): X := A.M
(MAD): S := M + A.E × X;
(MUL): S := A.E × X;
−
case
+ j (cid:54)= 1
j = 1
(1) h (cid:54)= i, j (cid:54)= 1
h (cid:54)= i, j = 1
h = i, j (cid:54)= 1
h = i, j = 1
(2) ¬(j = 1 ∧ h = i)
j = 1 ∧ h = i
+ j (cid:54)= 1 ∧ h = i
j = 1 ∧ h = i
(3) ¬(j = 1 ∧ h = i)
j = 1 ∧ h = i
(4) h (cid:54)= i
h = i (cid:54)= 1
h = i = 1
(5) ¬(h = i = 1)
(6) ¬(j = 1 ∧ h = i)
h = i = 1
j = 1 ∧ h = i
K is 0
M is 0
Z is 0
V , W are 0
V is 1
V is 1
M is 0
S no rows
S no rows
V is 1, W is 0
(RRF): M := RRF(A.λ, S, X);
(CPY): M := X;
(MAD): K := W + A.K × V ;
(CPY): K := A.K;
5 Concurrency analysis
To measure the degree of concurrency of our algorithm we assign costs to each of the
tasks. We perform a relative analysis, comparing the cost of a parallel Gauss algorithm
with the cost of a sequential algorithm. Therefore, to simplify our analysis, we assume
that the cost of a matrix multiplication of (α× β)· (β × γ) possibly followed by addition
is αβγ and the cost of echelonising an α × β matrix of rank r is αβr. It seems plausible
that when assuming that these costs are homogeneous functions of some degree, bounded
below by ω (see [4, Chapter 16]), then in the results of Proposition 5.3 and Theorem 5.4
the degree of concurrency can be replaced by some constant times aω−1. We also assume
that all blocks are square matrices of size α × α, where α is not too small and
α =
=
n
a
m
b
.
Then the tasks Extend, RowLengthen, PreClearUp, and Copy do not perform
any time consuming operations (compared to ClearDown and UpdateRow), so we
assign cost 0 to these tasks.
21
↓(i, j-1)
→(i, j-1, j)
↓(i-1, j)
→(i-1, j, k)
→(i, j-1, k)
↓(i, j)
→(i, j, k)
Figure 1: Extract of task dependency graph
Lemma 5.1. The cost of ClearDown is bounded above by α3 and the cost of Up-
dateRow is bounded above by 1.25α3.
Proof. We start analysing the task ClearDown: For i = 1 only one job ECH is
performed contributing cost α3. Otherwise the first call of MAD multiplies a matrix
A.A of size α× r with a matrix D.R of size r× (α− r), contributing cost αr(α− r). The
echelonisation is done on an α×(α−r) matrix of rank r(cid:48) and the second MAD multiplies
an r × r(cid:48) matrix by an r(cid:48) × (α − r − r(cid:48)) matrix. So in total the cost of ClearDown is
αr(α − r) + α(α − r)r(cid:48) + rr(cid:48)(α − r − r(cid:48)) = α2(r + r(cid:48)) − rr(cid:48)(α + r + r(cid:48)) ≤ α3
as r + r(cid:48) ≤ α.
For the task UpdateRow we similarly obtain the cost αrα for the MAD in row (1),
r(cid:48)r(cid:48)α for the MUL in row (3), rr(cid:48)α for the MAD in row (4), and (α− r(cid:48))r(cid:48)α for the MAD
in row (6). Summing up we obtain
α2(r + r(cid:48)) + αrr(cid:48) ≤ 1.25α3
again since r + r(cid:48) ≤ α.
Ignoring all tasks of cost 0 Step 1 only involves the tasks ClearDown and Up-
dateRow. The graph of task dependencies decomposes naturally into layers accord-
ij, Di−1
ing to the value of i + j. We abbreviate the call ClearDown(Cj
, i) with data
packages depending on i, j by ↓(i, j) and similarly UpdateRow(Aij, Cj
ik, Bi−1
jk , i) by
→(i, j, k).Then Figure 1 displays the local task dependencies in Step 1 for layers i + j − 1
and i + j, where k = j + 1, . . . , b.
j
Recall that a critical path in a task dependency graph is the longest directed path
between any pair of start node and finish node. Its length is weighted by the cost of the
nodes along the path.
22
Proposition 5.2.
1. The weighted length of a critical path in the task dependency
graph of Step 1 is 2.25α3(a + b − 1).
2. The weighted length of a critical path in Step 3 is max((b − 1)α3, aα3).
Proof. 1.) The task dependency graph splits naturally into a + b − 1 layers according
to the value of i + j ∈ {2, . . . , a + b}. Within each layer, the critical paths involve
ClearDown and UpdateRow exactly once. So in total the length of a critical path
is (a + b − 1)(α3 + 1.25α3).
2.) Step 3 only involves the task ClearUp of non-zero cost α3. The data package
Rjh is only changed by the data packages below in the same column so h − j times.
The maximum of h − j is achieved at R1b contributing the cost α3(b − 1). For the
transformation matrix, which can be done independently, each of the Mjh is touched a
times contributing aα3.
To determine the average degree of concurrency we divide the cost of the sequential
Gaussalgorithm (with transformation matrix) applied to the m×n-matrix ClearDown
by the weighted length of a critical path. For simplicity we assume that m = n, a = b
and that all blocks are of the same size α × β with α = β.
Proposition 5.3. Under the assumptions above and using Lemma 5.1 the average degree
of concurrency of the Chief is 1
Proof. By our assumptions n = m and the cost of the sequential Gaussalgorithm (with
transformation matrix) applied to the huge matrix ClearDown is n3 = a3α3. By
Proposition 5.2 the weighted length of a critical path in the complete algorithm the
Chief is (2.25(2a − 1) + a)α3 = (5.5a − 2.25)α3 ∼ 5.5aα3.
5.5a2.
In practical examples the gain of performance is much better. This is partly due
to the fact that we split our matrix into blocks that fit into the computer's memory;
an echelonisation of the huge matrix, however, would require to permanently read and
write data to the hard disk. The other reason is that in random examples the length
of a critical path is much shorter. To make this precise we assume, in addition to the
assumptions above of starting with a square matrix partitioned into square blocks of
equal size α, that our input matrix is well-conditioned, by which we mean that the a
top-left square submatrices of the input matrix of size jα (j = 1, . . . , a) have full rank.
Then, properly implemented, the cost of ClearDown is α3, if it is called for i = j
and 0 otherwise. Also in Update Row r(cid:48) = 0 and so the cost of Update Row is α3
(this can be shown without using the assumptions in Lemma 5.1). In particular in the
dependency graph above, the weighted length of a critical path in any odd layer is α3
and in an even layer, this length is 2α3 (resp. α3 for the last layer). In total this shows
the following
Theorem 5.4. For a well-conditioned square matrix, the length of a critical path in
the dependency graph is α3(3a − 2) and hence the average degree of concurrency of the
Chief is 1
3a2.
23
Remark 5.5. The concurrency analysis above is not sufficient to ensure the effective
use of all processors throughout the entire run that we see in the experimental results
below.
Although the steps are estimated at their worst case in practice many tasks early in the
task dependency graph execute a lot faster. Provided there are sufficiently many blocks,
enough work is available for execution considerably earlier than the task dependency graph
might suggest. Assigning a priority i + j to a task pertaining to the i-th row and j-th
column directs an appropriate scheduling.
6 Experimental results
We give timings for the two implementations mentioned in Section 2.1.
The Meataxe
In order to demonstrate the power of this algorithm the following tests were done on a
machine with 64-piledriver cores with 512 GB of memory running at 2.7 gigaherz.
We chose as our first example a random 1, 000, 000 × 1, 000, 000-matrix with entries
in the field of order 2. To put this matrix into reduced echelon form with transformation
matrix we chose to chop the matrix into blocks of size 50, 000 × 50, 000. This run took
520min.
The following graph shows the progress of the calculation. The red shows that over
60 cores were used for the vast majority of the time. The blue shows that during Steps 1
and 2 the memory footprint was fairly constant but at the transition to Step 3 about
30% more memory was needed, due mainly to the expansion of the matrix M.
For a second example, on the same machine, we echelonised with transformation
matrix a random 600, 000 × 600, 000-matrix with entries in the field of order 3 in 460
min, using a block size of 30, 000. We do not give a graph for this run, as it is almost
indistinguishable from the one given above.
24
0 10 20 30 40 50 60 70 0 100 200 300 400 500 600 0 50 100 150 200 250CPUsGBElapsed Time(min)CPU usagememory usage (GB) The above examples were done with a carefully chosen block size. The following two
examples highlight the effect of too large a block size. We echelonised the same random
300, 000 × 300, 000-matrix with entries in the field of order 3 using block sizes of 30, 000
and 15, 000, respectively. In the first graph we see that 15, 000 is again a good choice of
block size. Almost all cores are used for most of the time. In the second graph we see
that 30, 000 is too large a block size, so that the 64 available cores are seldomly utilised
at once, leading to a run time which is about 50% longer.
Note that in general terms the necessary block size agrees with the concurrency
analysis.
Our final three examples are intended to demonstrate that our methods are not
restricted to tiny fields, nor to prime fields. A random 200, 000 × 200, 000-matrix with
entries in the field with 193 took 445 mins, a random 200, 000 × 200, 000-matrix with
entries in the field with 1331 = 113 took 615 mins, and a random 100, 000 × 100, 000-
matrix with entries in the field 50653 = 373 elements took 200 mins.
7 Acknowledgements
We thank Martin Albrecht for discussions in the early stages of the algorithm design.
The first full implementation of this algorithm was developed by Jendrik Brachter in
GAP [8] which was essential to getting the details of the algorithm right. A parallel
version of this implementation, using HPC-GAP, was produced by Jendrik Brachter and
Sergio Siccha.
The second and third author acknowledge the support of the SFB-TR 195.
References
[1] Martin R. Albrecht The M4RIE library for dense linear algebra over small fields with
even characteristic. ISSAC 2012-Proceedings of the 37th International Symposium
on Symbolic and Algebraic Computation, 28–34, ACM, New York, 2012.
25
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 5 10 15 20 25 30 35 40CPUsGBElapsed Time(min)CPU usagememory usage (GB) 0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 80 90 100 5 10 15 20 25 30 35 40 45 50 55 60CPUsGBElapsed Time(min)CPU usagememory usage (GB) [2] Martin Albrecht and Gregory Bard and William Hart. Algorithm 898: Efficient
multiplication of dense matrices over GF (2) ACM Transactions on Mathematical
Software (TOMS) 37 (1), Article 9, 2010.
[3] Martin Albrecht et al.,Linear Algebra over F2 (and F2e)
https://malb.bitbucket.io/m4ri-e-website-2008-2015/further.html
[4] Peter Bürgisser and Michael Clausen and M. Amin Shokrollahi, Algebraic complexity
theory. With the collaboration of Thomas Lickteig. Grundlehren der Mathematis-
chen Wissenschaften 315. Springer-Verlag, Berlin, 1997.
[5] T.H. Cormen, C.E. Leiserson, R.L. Rivest, C. Stein, Introduction to Algorithms.
(Second ed.), The MIT Press, Cambridge, Massachusetts, London McGraw-Hill
Book Company, Boston Burr Ridge, IL Dubuque, IA Madison, WI New York San
Francisco St. Louis Montreal Toronto, (2001).
[6] Simplice Donfack and Jack Dongarra and Mathieu Faverge and Mark Gates and
Jakub Kurzak and Piotr Luszczek and Ichitaro Yamazaki, A survey of recent de-
velopments in parallel implementations of Gaussian elimination. Concurrency and
Computat.: Pract. Exper. (2014). DOI: 10.1002/cpe.3306
[7] Jean-Guillaume Dumas, Pascal Giorgi and Clément Pernet, Dense Linear Algebra
over Word-Size Prime Fields: the FFLAS and FFPACK Packages. ACM Trans. on
Mathematical Software (TOMS) 35(3), ACM Press, NY, USA, (2008), 1–42. DOI:
10.1145/1391989.1391992
[8] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.9.1;
2018. (https://www.gap-system.org)
[9] Joseph F. Grcar, Mathematicians of Gaussian elimination. Notices Amer. Math.
Soc. 58 (6), (2011), 782–792.
[10] Richard Parker, Meataxe64 Blog, https://meataxe64.wordpress.com
[11] PLASMA software package for
http://icl.cs.utk.edu/plasma/
solving problems
in dense linear algebra.
26
|
1604.01391 | 2 | 1604 | 2017-02-24T07:44:55 | Poisson Centralizer of the Trace | [
"math.RA",
"math.QA"
] | The Poisson centralizer of the trace element is determined in the coordinate ring of SL_n endowed with the Poisson structure obtained as the semiclassical limit of its quantized coordinate ring. It turns out that this maximal Poisson-commutative subalgebra coincides with the subalgebra of invariants with respect to the adjoint action. | math.RA | math |
POISSON CENTRALIZER OF THE TRACE
SZABOLCS MÉSZÁROS
Abstract. The Poisson centralizer of the trace element Pi
xi,i is determined
in the coordinate ring of SLn endowed with the Poisson structure obtained
as the semiclassical limit of its quantized coordinate ring. It turns out that
this maximal Poisson-commutative subalgebra coincides with the subalgebra
of invariants with respect to the adjoint action.
1. Introduction
The semiclassical limit Poisson structure on O(SLn) received considerable atten-
tion recently because of the connection between the primitive ideals of the quantized
coordinate ring Oq(SLn) and the symplectic leaves of the Poisson manifold SLn
(see for example [HL2],[G],[Y]). In this paper, we present another relation between
O(SLn) endowed with the semiclassical limit Poisson structure and Oq(SLn).
the trace element σ1 = Pi xi,i in Oq(SLn) (resp.
In [M] it was shown that if q ∈ C× is not a root of unity then the centralizer of
in Oq(Mn) and Oq(GLn)) is a
maximal commutative subalgebra, generated by certain sums of principal quantum
minors. By Theorem 2.4 and 5.1 in [DL2], this subalgebra coincides with the
subalgebra of cocommutative elements in Oq(SLn) and also with the subalgebra of
invariants of the adjoint coaction. (This result is generalized in [AZ] for arbitrary
characteristic and q being a root of unity.)
On the Poisson algebra side, the corresponding Poisson-subalgebra of O(SLn)
is generated by the coefficients of the characteristic polynomial c1, . . . , cn−1. We
prove the following:
Theorem 1.1. For n ≥ 1 the subalgebra C[c1, . . . , cn−1] (resp. C[c1, . . . , cn] and
C[c1, . . . , cn, c−1
n ]) is maximal Poisson-commutative in O(SLn) (resp. O(Mn) and
O(GLn)) with respect to the semiclassical limit Poisson structure.
It is easy to deduce from [DL1] or [DL2] that {ci, cj} = 0 (1 ≤ i, j ≤ n) in O(Mn)
(see Proposition 5.1 below). Therefore, Theorem 1.1 is a direct consequence of the
following statement:
Theorem 1.2. For n ≥ 1 the Poisson-centralizer of c1 in O(SLn) (resp. c1 ∈
O(Mn) and O(GLn)) equipped with the semiclassical limit Poisson bracket is gen-
erated as a subalgebra by
• c1, . . . cn−1 in the case of O(SLn),
• c1, . . . , cn in the case of O(Mn), and
• c1, . . . , cn, c−1
n in the case of O(GLn).
2010 Mathematics Subject Classification. 16T20, 17B63 (primary), 16W70, 20G42 (secondary).
Keywords. Quantized coordinate ring, semiclassical limit, Poisson algebra, complete involutive
system, maximal Poisson-commutative subalgebra.
1
Poisson Centralizer of the Trace
Szabolcs Mészáros
The proof is based on modifying the Poisson bracket of the algebras that makes
an induction possible. A similar idea is used in the proof of the analogous result in
the quantum setup (see [M]).
It is well known that the coefficient functions c1, . . . , cn ∈ O(Mn) of the char-
acteristic polynomial generate the subalgebra O(Mn)GLn of GLn-invariants with
respect to the adjoint action. This implies that the subalgebra coincides with the
Poisson center of the coordinate ring O(Mn) endowed with the Kirillov-Kostant-
Souriau (KKS) Poisson bracket. Hence, Theorem 1.1 for O(Mn) can be interpreted
as an interesting interplay between the KKS and the semiclassical limit Poisson
structure. Namely, while the subalgebra O(Mn)GLn is contained in every maxi-
mal Poisson-commutative subalgebra with respect to the former Poisson bracket,
it is contained in only one maximal Poisson-commutative subalgebra (itself) with
respect to the latter Poisson bracket.
A Poisson-commutative subalgebra is also called an involutive (or Hamiltonian)
system, while a maximal one is called a complete involutive system (see Section 2
or [V]). Such a system is integrable if the (Krull) dimension of the generated sub-
algebra is sufficiently large. In our case, the subalgebra generated by the elements
c1, . . . , cn−1 is not integrable, as its dimension is n − 1 (resp. n for GLn) instead
The article is organized as follows: First, we introduce the required notions, and
in Section 3 we prove that the three statements in Theorem 1.2 are equivalent.
In Section 3.1, we prove Theorem 1.2 for n = 2 as a starting case of an induction
presented in Section 5 that completes the proof of the theorem. In the article, every
algebra is understood over the field C.
of the required (cid:0)n+1
2 (cid:1) − 1 (resp. (cid:0)n+1
2 (cid:1) for GLn), see Remark 5.3.
2. Preliminaries
2.1. Poisson algebras. First, we collect the basic notions about Poisson algebras
we use in the article. For further details about Poisson algebras, see [V].
A commutative Poisson algebra (cid:0)A, {., .}(cid:1) is a unital commutative associative
algebra A together with a bilinear operation {., .} : A × A → A called the Poisson
bracket such that it is antisymmetric, satisfies the Jacobi identity, and for any
a ∈ A, {a, .} : A → A is a derivation. For commutative Poisson algebras A and
B, the map ϕ : A → B is a morphism of Poisson algebras if it is both an algebra
homomorphism and a Lie-homomorphism.
There is a natural notion of Poisson subalgebra (i.e. a subalgebra that is also
a Lie-subalgebra), Poisson ideal (i.e an ideal that is also a Lie-ideal) and quotient
Poisson algebra (as the quotient Lie-algebra inherits the bracket). The Poisson
centralizer C(a) of an element a ∈ A is defined as {b ∈ A {a, b} = 0}. Clearly, it is
a Poisson subalgebra. Analogously, a ∈ A is called Poisson-central if C(a) = A. One
says that a subalgebra C ≤ A is Poisson-commutative (or involutive) if {c, d} = 0
for all c, d ∈ C and it is maximal Poisson-commutative (or maximal involutive) if
there is no Poisson-commutative subalgebra in A that strictly contains C.
The Poisson center (or Casimir subalgebra) of A is Z(A) := {a ∈ A C(a) =
A}. Let A be a reduced, finitely generated commutative Poisson algebra. The
rank Rk{., .} of the Poisson structure {., .} is defined by the rank of the matrix
(cid:0){gi, gj})i,j ∈ AN ×N for a generating system g1, . . . , gN ∈ A. (One can prove that it
is independent of the chosen generating system.) A maximal Poisson-commutative
2
Poisson Centralizer of the Trace
Szabolcs Mészáros
subalgebra C is called integrable if
dim C = dim A −
1
2
Rk{., .}
The inequality ≤ holds for any Poisson-commutative subalgebra (Proposition II.3.4
in [V]), hence integrability is a maximality condition on the size of C that does not
necessarily hold for every maximal involutive system.
2.2. Filtered Poisson algebras.
Definition 2.1. A filtered Poisson algebra is a Poisson algebra together with an
ascending chain of subspaces {F d}d∈N in A such that
• A = ∪d∈NF d,
• F d · F e ⊆ F d+e for all d, e ∈ N, and
• {F d, F e} ⊆ F d+e for all d, e ∈ N.
Together with the filtration preserving morphisms of Poisson algebras, they form a
category.
For a filtered Poisson algebra A, we may define its associated graded Poisson
algebra grA as
gr(A) :=Md∈N
F d/F d−1
where we used the simplifying notation F −1 = {0}. The multiplication of gr(A) is
defined the usual way:
F d/F d−1 × F e/F e−1 → F d+e/F d+e−1
(cid:0)x + F d−1, y + F e−1(cid:1) 7→ xy + F d+e−1
Analogously, the Poisson structure of gr(A) is defined by (cid:0)x + F d−1, y + F e−1(cid:1) 7→
{x, y} + F d+e−1. One can check that this way gr(A) is a Poisson algebra.
Let (S, +) be an abelian monoid. (We will only use this definition for S = N and
S = Z/nZ for some n ∈ N.) An S-graded Poisson algebra R is a Poisson algebra
together with a fixed grading
R = ⊕d∈SRd
such that R is both a graded algebra (i.e. Rd · Re ⊆ Rd+e for all d, e ∈ S) and a
graded Lie algebra (i.e. {Rd, Re} ⊆ Rd+e for all d, e ∈ S) with respect to the given
grading.
The above construction A 7→ gr(A) yields an N-graded Poisson algebra. In fact,
for a morphism of filtered Poisson algebras
gr(.) can be turned into a functor:
f :(cid:0)A, {F d}d∈N(cid:1) →(cid:0)B, {Gd}d∈N(cid:1) we define
gr(f ) : gr(A) → gr(B)
One can check that it is indeed well defined and preserves composition.
(cid:0)xd + F d−1(cid:1)d∈N 7→(cid:0)f (xd) + Gd−1(cid:1)d∈N
Remark 2.2. Given an N-graded Poisson algebra R = ⊕d∈NRd, one has a natural
way to associate a filtered Poisson algebra to it. Namely, let F d := ⊕k≤dRk. In
this case, the associated graded Poisson algebra grR of(cid:0)R, {F d}d∈N(cid:1) is isomorphic
to R.
3
Poisson Centralizer of the Trace
Szabolcs Mészáros
2.3. The Kirillov-Kostant-Souriau bracket. A classical example of a Poisson
algebra is given by the Kirillov-Kostant-Souriau (KKS) bracket on O(g∗), the coor-
dinate ring of the dual of a finite-dimensional (real or complex) Lie algebra(cid:0)g, [., .](cid:1)
(see [ChP] Example 1.1.3, or [W] Section 3).
It is defined as follows: a function f ∈ O(g∗) at a point v ∈ g∗ has a differential
0 g∗ ∼= g∗∗ ∼= g.
v g∗ where we can canonically identify the spaces T ∗
dfv ∈ T ∗
Hence, we may define the Poisson bracket on O(g∗) as
v g∗ ∼= T ∗
{f, g}(v) := [dfv, dgv](v)
for all f, g ∈ O(g∗) and v ∈ g∗. It is clear that it is a Lie-bracket but it can be
checked that the Leibniz-identity is also satisfied. For g = gln, it gives a Poisson
bracket on O(Mn).
Alternatively, one can define this Poisson structure via semiclassical limits.
2.4. Semiclassical limits. Let A = ∪d∈ZAd be a Z-filtered algebra such that its
associated graded algebra gr(A) := ⊕d∈ZAd/Ad−1 is commutative. The Rees ring
of A is defined as
Rees(A) :=Md∈Z
Adhd ⊆ A[h, h−1]
Using the obvious multiplication, it is a Z-graded algebra. The semiclassical limit
of A is the Poisson algebra Rees(A)/hRees(A) together with the bracket
{a + hAm, b + hAn} :=
1
h
[a, b] + An+m−2 ∈ An+m−1/An+m−2
for all homogeneous elements a + hAm ∈ Am+1/hAm, b + hAn ∈ An+1/hAn. The
definition is valid as the underlying algebra of Rees(A)/hRees(A) is gr(A) that is
assumed to be commutative, hence [a, b] ∈ hAm+n−1.
The Poisson algebra O(g∗) with the KKS bracket can be obtained as the semi-
classical limit of U g, see [G], Example 2.6.
2.5. Quantized coordinate rings. Assume that n ∈ N+ and define Ot(Mn) as
the unital C-algebra generated by the n2 generators xi,j for 1 ≤ i, j ≤ n over
C[t, t−1] that are subject to the following relations:
xi,j xk,l =
xk,lxi,j + (t − t−1)xi,lxk,j
txk,lxi,j
xk,lxi,j
if i < k and j < l
if (i = k and j < l) or (j = l and i < k)
if (i > k and j < l) or (j > l and i < k)
It turns out to be a finitely generated C[t, t−1]-algebra
for all 1 ≤ i, j, k, l ≤ n.
that is a Noetherian domain. (For a detailed exposition, see [BG].) Furthermore, it
can be endowed with a coalgebra structure by setting ε(xi,j) = δi,j and ∆(xi,j ) =
k=1 xi,k ⊗ xk,j . It turns Ot(Mn) into a bialgebra.
For q ∈ C×, the quantized coordinate ring of n × n matrices with parameter q is
Pn
defined as the C-algebra
Oq(Mn) := Ot(Mn)/(t − q)
In this article, we only deal with the case when q is not a root of unity, then the
algebra is called the generic quantized coordinate ring of Mn.
Similarly, one can define the non-commutative deformations of the coordinate
rings of GLn and SLn using the quantum determinant
4
Poisson Centralizer of the Trace
Szabolcs Mészáros
detq := Xs∈Sn
(−q)ℓ(s)x1,s(1)x2,s(2) . . . xn,s(n)
where ℓ(σ) stands for the length of σ in the Coxeter group Sn. Then -- analogously
to the classical case -- one defines
Oq(SLn) := Oq(Mn)/(detq − 1)
by localizing at the central element detq.
Oq(GLn) := Oq(Mn)(cid:2)det−1
q (cid:3)
2.6. Semiclassical limits of quantized coordinate rings. The semiclassical
limits of Oq(SLn) can be obtained via the slight modification of process of Section
2.4 (see [G], Example 2.2). The algebra R := Ot(Mn) can be endowed with a Z-
filtration by defining F n to be the span of monomials that are the product of at most
n variables. However, instead of defining a Poisson structure on Rees(R)/hRees(R)
with respect to this filtration, consider the algebra R/(t − 1)R that is isomorphic
to O(Mn) as an algebra. The semiclassical limit Poisson bracket is defined as
{¯a, ¯b} :=
1
t − 1
(ab − ba) + (t − 1)R ∈ R/(t − 1)R
for any two representing elements a, b ∈ R for ¯a, ¯b ∈ R/(t − 1)R. One can check
that it is a well-defined Poisson bracket.
This Poisson structure of O(Mn) can be given explicitly by the following rela-
tions:
{xi,j, xk,l} =
2xi,lxk,j
xi,j xk,l
0
if i < k and j < l
if (i = k and j < l) or (j = l and i < k)
otherwise
extended according to the Leibniz-rule (see [G]). It is a quadratic Poisson structure
in the sense of [V], Definition II.2.6. The semiclassical limit for GLn and SLn
is defined analogously using Oq(GLn) and Oq(SLn) or by localization (resp. by
taking quotient) at the Poisson central element det (resp. det − 1) in O(Mn).
2.7. Coefficients of the characteristic polynomial. Consider the characteristic
polynomial function Mn → C[x], A 7→ det(A − xI). Let us define the elements
c0, c1, . . . , cn ∈ O(Mn) as
n
det(A − xI) =
(−1)icixn−i
Xi=0
In particular, c0 = 1, c1 = tr and cn = det. Their images in O(SLn) ∼= O(Mn)/(det−
1) are denoted by c1, . . . , cn−1. If ambiguity may arise, we will write ci(A) for the
element corresponding to ci for an algebra A with a fixed isomorphism A ∼= O(Mk)
for some k.
The coefficient functions c1, . . . , cn can also be expressed via matrix minors as
follows: For I, J ⊆ {1, . . . , n}, I = (i1, . . . , ik) and J = (j1, . . . , jk) define
[I J] := Xs∈Sk
sgn(s)xi1 ,js(1) . . . xik ,js(k)
i.e. it is the determinant of the subalgebra generated by {xi,j}i∈I,j∈J that can be
identified with O(Mk). Then
5
Poisson Centralizer of the Trace
Szabolcs Mészáros
[I I] ∈ O(Mn)
ci = XI=i
for all 1 ≤ i ≤ n. It is well.known that c1, . . . , cn generate the same subalgebra of
O(Mn) as the trace functions A 7→ Tr(Ak), namely, the subalgebra O(Mn)GLn of
GLn-invariants with respect to the adjoint action.
3. Equivalence of the statements
Consider O(Mn) endowed with the semiclassical limits Poisson bracket. As it is
discussed in the Introduction, Theorem 1.1 follows directly from Theorem 1.2 and
Proposition 5.1.
The following proposition shows that it is enough to prove Theorem 1.2 for the
case of O(Mn).
Proposition 3.1. For any n ∈ N+ the following are equivalent:
(1) The Poisson-centralizer of c1 ∈ O(Mn) is generated by c1, . . . , cn.
(2) The Poisson-centralizer of c1 ∈ O(GLn) is generated by c1, . . . , cn, c−1
n .
(3) The Poisson-centralizer of c1 ∈ O(SLn) is generated by c1, . . . , cn−1.
Proof. The first and second statements are equivalent as det is a Poisson-central
element, so we have {c1, h · detk} = {c1, h} · detk for any h ∈ O(GLn) and k ∈ Z.
Hence,
proving 1) ⇐⇒ 2).
O(GLn) ⊇ C(c1) =(cid:0)O(Mn) ∩ C(c1)(cid:1)[det−1]
1) ⇐⇒ 3): First, assume 1) and let h ∈ O(SLn) such that {c1, h} = 0.
Since O(SLn) is Z/nZ-graded (inherited from the N-grading of O(Mn)) and c1 is
homogeneous with respect to this grading, its Poisson-centralizer is generated by
Z/nZ-homogeneous elements, so we may assume that h is Z/nZ-homogeneous.
Let k = deg(h) ∈ Z/nZ. Let h ∈ O(Mn) be a lift of h ∈ O(SLn) and consider the
j=0 hjn+k of h, where hjn+k is homogeneous
N-homogeneous decomposition h =Pd
of degree jn + k for all j ∈ N. Define
d
hjn+kdetd−j ∈ O(Mn)dn+k
h′ :=
Xj=0
that is a homogeneous element of degree dn + k representing h ∈ O(SLn) in
O(Mn). Then {c1, h′} ∈ (det −1) ∩ O(Mn)dn+k+1 since {c1, h′} = {c1, h} =
0, c1 is homogeneous of degree 1 and the Poisson-structure is graded. Clearly,
(det −1)∩O(Mn)dn+k+1 = 0 hence {c1, h′} = 0. Applying 1) gives h′ ∈ C[c1, . . . , cn]
so h ∈ C[c1, . . . , cn−1] as we claimed.
Conversely, assume 3) and let h ∈ O(Mn) such that {c1, h} = 0. Since c1 is
N-homogeneous, we may assume that h is also N-homogeneous and so the image
h ∈ O(SLn) of h is Z/nZ-homogeneous. By the assumption, h = p(c1, . . . , cn−1) for
some p ∈ C[t1, . . . , tn−1]. Endow C[t1, . . . , tn] with the N-grading deg(ti) = i. As h
is Z/nZ-homogeneous, we may choose p ∈ C[t1, . . . , tn−1] so that its homogeneous
components are all of degree dn + deg(h) ∈ N with respect to the above grading for
some d ∈ N.
By h − p(c1, . . . , cn−1) ∈ (det −1) and the assumptions on degrees, we may
choose a polynomial q ∈ C[t1, . . . , tn] that is homogeneous with respect to the above
6
Poisson Centralizer of the Trace
Szabolcs Mészáros
grading and q(t1, . . . , tn−1, 1) = p. Let h′ := h·detr where r := 1
so deg(h′) = deg(q) ∈ N. Then
n (deg q −deg h) ∈ Z
h′ − q(c1, . . . , cn) ∈ (det −1) ∩ O(Mn)degq = 0
hence h′ ∈ C[c1, . . . , cn] and h ∈ C[c1, . . . , cn, c−1
O(Mn) = C[c1, . . . , cn] by the definitions.
n ]. This is enough as C[c1, . . . , cn, c−1
n ]∩
(cid:3)
4. Case of O(SL2)
In this section, we prove Theorem 1.2 for O(SL2) that is the first step of the
induction in the proof of the general case.
We denote by a, b, c, d the generators x1,1, x1,2, x2,1, x2,2 ∈ O(SL2) and tr :=
c1 = a + d.
Proposition 4.1. The centralizer of tr ∈ O(SL2) is C[tr].
By ad − bc = 1 we have a monomial basis of O(SL2) consisting of
aibkcl, bkcldj, bkcl
(i, j ∈ N+, k, l ∈ N)
The Poisson bracket on the generators is the following:
{a, b} = ab
{a, c} = ac
{a, d} = 2bc
{b, c} = 0
{b, d} = bd
{c, d} = cd
The action of {tr, .} on the basis elements can be written as
= (k + l)ai+1bkcl − 2iai−1bk+1cl+1 − (k + l)aibkcld
(cid:8)(a + d), aibkcl(cid:9) =
= (k + l)ai+1bkcl − (2i + k + l)ai−1bk+1cl+1 − (k + l)ai−1bkcl
By the same computation on bkcl and bkcldj one obtains
(cid:8)(a + d), bkcl(cid:9) = (k + l)abkcl − (k + l)bkcld
(cid:8)(a + d), bkcldj(cid:9) = (k + l + 2j)bk+1cl+1dj−1 + (k + l)bkcldj−1 − (k + l)bkcldj+1
Hence, for a polynomial p ∈ C[t1, t2] and i ≥ 1:
(4.1)
(cid:8)(a + d), aip(b, c)(cid:9) = ai+1Xm
m · pm(b, c)
where pm is the m-th homogeneous component of p. The analogous computations
for p(b, c)dj (j ≥ 1) and p(b, c) give
−ai−1Xm (cid:0)(2i + m)bc + m(cid:1)pm(b, c)
(4.2)
(4.3)
(cid:8)(a + d), p(b, c)dj(cid:9) = −dj+1Xm
m · pm(b, c)
+dj−1Xm (cid:0)(m + 2j)bc + m(cid:1)pm(b, c)
(cid:8)(a + d), p(b, c)(cid:9) = (a − d)Xm
7
m · pm(b, c)
Poisson Centralizer of the Trace
Szabolcs Mészáros
Proof of Proposition 4.1. Assume that 0 6= g ∈ C(tr) and write it as
g =
α
Xi=1
airi +
β
Xj=1
sjdj + u
where ri, sj and u are elements of C[b, c], and α and β are the highest powers of a
and d appearing in the decomposition.
We prove that rα ∈ C · 1. If α = 0 then rα = u so the aibkcl terms (i > 0)
in {a + d, g} are the same as the aibkcl terms in {a + d, u} by Eq. 4.1, 4.2 and
4.3. However, by 4.3, these terms are nonzero if u /∈ C and that is a contradiction.
Assume that α ≥ 1 and for a fixed k ∈ N define the subspace
alC[b, c, d] ⊆ O(SL2)
Ak :=Xl≤k
By (cid:8)tr, Aα−1(cid:9) ⊆ Aα we have
Aα = {tr, g} + Aα =(cid:8)tr, aαrα + Aα−1(cid:9) + Aα =
= aα{tr, rα} + αaα−1bcrα + Aα = aα{tr, rα} + Aα
By Eq. 4.3 it is possible only if {tr, rα} = 0 so rα ∈ C[b, c] ∩ C(tr) = C · 1.
If α > 0 we may simplify g by subtracting polynomials of tr from it. Indeed,
by rα ∈ C× we have g − rαtrα ∈ Aα−1 ∩ C(tr) so we can replace g by g − rαtrα.
Hence, we may assume that α = 0. Then, again, rα = u ∈ C · 1 ⊆ C(tr) so we may
also assume that u = 0.
If g is nonzero after the simplification, we get a contradiction.
Indeed, let
p(b, c)dγ be the summand of g with the smallest γ ∈ N. By the above simplifi-
cations, γ ≥ 1. Then the coefficient of dγ−1 in {tr, g} is the same as the coefficient
of dγ−1 in
{tr, p(b, c)dγ} = {tr, p(b, c)}dγ + 2γbcp(b, c)dγ−1
so it is 2γbcp(b, c)dγ−1 that is nonzero if p(b, c) 6= 0 and γ ≥ 1. That is a contra-
diction.
(cid:3)
5. Proof of the main result
Let n ≥ 2 and let us denote An := O(Mn).
Proposition 5.1. C[σ1, . . . , σn] ≤ An is a Poisson-commutative subalgebra.
Proof. Consider the principal quantum minor sums
σi = XI=iXs∈Si
t−ℓ(s)xi1,is(1) . . . xit ,is(t) ∈ Ot(Mn)
When An is viewed as the semiclassical limit R/(t − 1)R where R = Ot(Mn) (see
Subsection 2.5), one can see that σi represents ci ∈ R/(t − 1)R ∼= O(Mn). In [DL1],
it is proved that σiσj = σj σi in Oq(Mn) if q is not a root of unity, in particular, if
q is transcendental.
Since the algebra Oq(Mn) is defined over Z[q, q−1], the elements σ1, . . . , σn (that
are defined over Z[q, q−1]) commute in Oq(cid:0)Mn(Z)(cid:1) ≤ Oq(cid:0)Mn(C)(cid:1) as well. Hence,
σ1, . . . , σn also commute after extension of scalars, i.e. in the ring Oq(cid:0)Mn(Z)(cid:1) ⊗Z
C ∼= Ot(cid:0)Mn(C)(cid:1). Consequently, in An ∼= R/(t − 1)R the subalgebra C[c1, . . . , cn]
is a Poisson-commutative subalgebra, by the definition of semiclassical limit.
(cid:3)
8
Poisson Centralizer of the Trace
Szabolcs Mészáros
By Proposition 5.1, C[c1, . . . , cn] is in the Poisson-centralizer C(c1). To prove
the converse, Theorem 1.2, we need further notations. Consider the Poisson ideal
I := (x1,j, xi,1 2 ≤ i, j ≤ n) ⊳ An
We will denote its quotient Poisson algebra by B2,n := An/I and the natural
surjection by ϕ : An → B2,n. Note that B2,n ∼= An−1[t] as Poisson algebras by
xi,j + I 7→ xi−1,j−1 (2 ≤ i, j ≤ n) and x1,1 7→ t where the bracket of An−1[t] is the
trivial extension of the bracket of An−1 by {t, a} = 0 for all a ∈ An−1[t].
Furthermore, Dn will stand for C[t1, . . . , tn] endowed with the zero Poisson
bracket. Define the map δ : B2,n → Dn as xi,j + I 7→ δi,jti that is morphism
of Poisson algebras by {xi,i, xj,j} ∈ I. Note that (δ ◦ ϕ)(ci) = si, the elementary
symmetric polynomial in t1, . . . , tn. In particular, δ ◦ ϕ restricted to C[c1, . . . , cn] is
an isomorphism onto the symmetric polynomials in t1, . . . , tn by the fundamental
theorem of symmetric polynomials. In the proof of Theorem 1.2 we verify the same
property for C(σ1).
Although the algebras An, B2,n and Dn are N-graded Poisson algebras (see
Section 2) using the total degree of An and the induced gradings on the quotients,
we will instead consider them as filtered Poisson algebras where the filtration is not
the one that corresponds to this grading. For each d ∈ N, let us define
Ad = {a ∈ An degx1,1 (a) ≤ d}
This is indeed a filtration on An. Note, that the grading degx1,1 is incompatible
with the bracket by {x1,1, x2,2} = x1,2x2,1. The algebras B2,n, Dn and C(c1) inherit
a filtered Poisson algebra structure as they are Poisson sub- and quotient algebras
of An so we may take Bd := ϕ(Ad), Dd := (δ ◦ ϕ)(Ad) and Cd = Ad ∩ C(c1). This
way, the natural surjections ϕ and δ and the embedding C(c1) ֒→ An are maps of
filtered Poisson algebras.
In the proof of Theorem 1.2 we use the associated graded Poisson algebras of
B2,n, Dn and C(c1) (see Section 2). First, we describe the structure of these. The
filtrations on B2,n and Dn are induced by the x1,1- and t1-degrees, hence we have
grB2,n ∼= B2,n and grDn ∼= Dn as graded Poisson algebras (and grδ = δ), so we
identify them in the following.
The underlying graded algebra of grAn is isomorphic to An using the x1,1-degree
but the Poisson bracket is different: it is the same on the generators xi,j and xk,l
for (i, j) 6= (1, 1) 6= (k, l) but
{x1,1, xi,j }gr = 0
{x1,1, x1,j}gr = x1,1x1,j
{x1,1, xi,1}gr = x1,1xi,1
(2 ≤ i, j ≤ n)
(2 ≤ j ≤ n)
(2 ≤ i ≤ n)
where {., .}gr stands for the Poisson bracket of grAn. Consequently, as maps we
have grϕ = ϕ, we still have {ci, cj}gr = 0 for all i, j, and the underlying algebra of
grC(c1) can be identified with C(c1).
Note, that C(c1) is defined by the original Poisson structure {., .} of An and not
by {., .}gr, even if it will be considered as a Poisson subalgebra of grAn. The reason
of this slightly ambiguous notation is that we will also introduce Cgr(x1,1) ⊆ grAn
as the centralizer of x1,1 with respect to {., .}gr.
Our associated graded setup can be summarized as follows:
C(c1) ⊆
grAn
/ B2,n
δ
/ Dn
ϕ
9
/
/
/
/
/
/
Poisson Centralizer of the Trace
Szabolcs Mészáros
Proof of Theorem 1.2. We prove the statement by induction on n. The statement
is verified for O(SL2) in Section 4 so, by Proposition 3.1 the case n = 2 is proved.
Assume that n ≥ 3. We shall prove that
• (δ ◦ ϕ)C(c1) : C(c1) → Dn is injective, and
• the image (δ ◦ ϕ)(cid:0)C(c1)(cid:1) is in DSn
n .
These imply that the restriction of δ ◦ ϕ to C(c1) is an isomorphism onto DSn
n since
C(c1) ∋ ci for i = 1, . . . , n (see Section 2) and δ ◦ ϕ restricted to C[c1, . . . , cn] is
surjective onto DSn
n . The statement of the theorem follows.
To prove that δ ◦ ϕ is injective on C(c1) it is enough to prove that δ is injective
on C(cid:0)ϕ(c1)(cid:1) and that ϕ is injective on C(c1). Indeed, as ϕ is a Poisson map we
have ϕ(cid:0)C(c1)(cid:1) ⊆ C(cid:0)ϕ(c1)(cid:1).
First, we prove δ is injective on C(cid:0)ϕ(c1)(cid:1). By B2,n ∼= An−1[t] where t is Poisson-
B2,n ⊇ C(cid:0)ϕ(c1)(cid:1) ∼= CAn−1[t](cid:0)t + c1(An−1)(cid:1) = CAn−1(cid:0)c1(An−1)(cid:1)[t] ⊆ An−1[t]
By the induction hypothesis
central, we have
CAn−1(cid:0)c1(An−1)(cid:1) = C(cid:2)c1(An−1), . . . , cn−1(An−1)(cid:3)
where si is the symmetric polynomial in the variables t2, . . . , tn. In particular, δ is
Therefore, δ restricted to C(cid:0)ϕ(c1)(cid:1) is an isomorphism onto C[s1, . . . , sn−1][t1] ⊆ Dn
injective on C(cid:0)ϕ(c1)(cid:1).
To verify the injectivity of ϕ on C(c1), define
Cgr(x1,1) := {a ∈ grAn {x1,1, a}gr = 0}
The subalgebra C(c1) is contained in Cgr(x1,1) since for a homogeneous element a
of degree d, we have
Ad+1/Ad ∋ {x1,1, a}gr + Ad = {x1,1 + A0, a + Ad−1} + Ad = {c1, a} + Ad
hence {c1, a} = 0 implies {x1,1, a}gr = 0 ∈ grAn. Our setup can be visualized on
the following diagram:
gr(An)
ϕ
/ B2,n
δ
/ Dn
S
S
;✈
✈
✈
✈
✈
✈
✈
✈
✈
Cgr(x1,1)
C(cid:0)ϕ(c1)(cid:1)
8r
r
r
r
r
r
S
C(c1)
r
r
r
r
Now, it is enough to prove that ϕ restricted to Cgr(x1,1) is injective.
We can give an explicit description of Cgr(x1,1) in the following form:
Cgr(x1,1) = C[x1,1, xi,j 2 ≤ i, j ≤ n] ≤ grAn
Indeed,
{x1,1, xi,j}gr =(x1,1xi,j
0
if j 6= i = 1 or i 6= j = 1
otherwise
Therefore, the map adgrx1,1 : a 7→ {x1,1, a}gr acts on a monomial m ∈ grAn as
{x1,1, m}gr = c(m) · x1,1m where c(m) is the sum of the exponents of the x1,j 's
10
/
/
/
/
/
/
,
;
8
Poisson Centralizer of the Trace
Szabolcs Mészáros
and xi,1's (2 ≤ i, j ≤ n) in m. Hence, adgrx1,1 maps the monomial basis of grAn
injectively into itself. In particular,
Cgr(x1,1) = Ker(cid:0)adgrx1,1(cid:1) = {a ∈ grAn c(m) = 0} ∼= An−1[t]
using the isomorphism x1,1 7→ t and xi,j 7→ xi−1,j−1.
The injectivity part of the theorem follows: ϕ is injective on Cgr(x1,1) (in fact it
injective.
is an isomorphism onto B2,n), and ϕ maps C(c1) into C(cid:0)ϕ(c1)(cid:1) on which δ is also
To prove (δ ◦ ϕ)(cid:0)C(c1)(cid:1) ⊆ DSn
n , first note that in the above we have proved that
(δ ◦ ϕ)(cid:0)C(c1)(cid:1) ⊆ δ(cid:0)C(cid:0)ϕ(c1)(cid:1)(cid:1) ⊆ DSn−1
n
where Sn−1 acts on Dn by permuting t2, . . . , tn. Consider the automorphism γ
of An given by the reflection to the off-diagonal: γ(xi,j ) = xn+1−i,n+1−j .
It is
not a Poisson map but a Poisson antimap (using the terminology of [ChP]), i.e.
γ({a, b}) = −{γ(a), γ(b)}. It maps c1 into itself and consequently C(c1) into itself.
For the analogous involution γ : Dn → Dn, ti 7→ tn+1−i (i = 1, . . . , n) we have
(δ ◦ ϕ) ◦ γ = γ ◦ (δ ◦ ϕ). Hence,
(δ ◦ ϕ)(cid:0)C(c1)(cid:1) = (δ ◦ ϕ ◦ γ)(cid:0)C(c1)(cid:1) = (γ ◦ δ ◦ ϕ)(cid:0)C(c1)(cid:1) ⊆ γ(cid:0)DSn−1
proving the symmetry of (δ ◦ ϕ)(cid:0)C(c1)(cid:1) in t1, . . . , tn−1, so it is symmetric in all the
variables by n ≥ 3.
(cid:1)
(cid:3)
n
Remark 5.2. In contrast with Theorem 1.1, in the case of the KKS Poisson struc-
ture, every Poisson-commutative subalgebra contains the Poisson center C[c1, . . . , cn],
see [W]. For a maximal commutative subalgebra with respect to the KKS bracket,
see [KW].
Remark 5.3. We prove that C[c1, . . . cn−1] is not an integrable complete involutive
system (see Section 2). First, observe that the rank of the semiclassical Poisson
bracket of O(SLn) is n(n − 1).
Indeed, by Section 2, the rank is the maximal dimension of the symplectic leaves
in SLn. The symplectic leaves in SLn are classified in [HL1], Theorem A.2.1, based
on the work of Lu, Weinstein and Semenov-Tian-Shansky [LW], [S]. The dimension
of a symplectic leaf is determined by an associated element of W ×W where W = Sn
is the Weyl group of SLn. According to Proposition A.2.2, if (w+, w−) ∈ W × W
then the dimension of the corresponding leaves is
(5.1)
ℓ(w+) + ℓ(w−) + min{m ∈ N w+w−1
− = r1 · · · · · rm ri is a transposition for all i}
where ℓ(.) is the length function of the Weyl group that -- in the case of SLn --
is the number of inversions in a permutation. By the definition of inversion using
elementary transpositions, the above quantity is bounded by
ℓ(w+) + ℓ(w−) + ℓ(w+w−1
− )
The maximum of the latter is n(n−1) since ℓ(w+) =(cid:0)n
stands for the longest element of Sn. Therefore,
ℓ(w+)+ℓ(w−)+ℓ(w+w−1
because ℓ(gh) ≤ ℓ(g) + ℓ(h) = ℓ(g) + ℓ(h−1) for all g, h ∈ Sn. This maximum is
attained on w+ = w− = t, even for the original quantity in Equation 5.1. Hence,
2(cid:1)−ℓ(w+t) where t = (n . . . 1)
− ) = n(n−1)−ℓ(w+t)−ℓ(w−t)+ℓ(cid:0)(w+t)(w−t)−1(cid:1) ≤ n(n−1)
11
Poisson Centralizer of the Trace
Szabolcs Mészáros
Rk{., .} = n(n − 1) for SLn and Rk{., .} = n(n − 1) + 1 for Mn and GLn. However,
a complete integrable system should have dimension
dim SLn −
1
2
Rk{., .} = n2 − 1 −(cid:18)n
2(cid:19) =(cid:18)n + 1
2 (cid:19) − 1
So it does not equal to dim C[c1, . . . cn−1] = n − 1 if n > 1. Similarly, the system is
non-integrable for Mn and GLn.
References
[AZ] V. V. Antonov, A. N. Zubkov, Coinvariants for a coadjoint action of quantum matrices,
Algebra and Logic, 48 (2009), no 4, 239.
[BG] K. A. Brown, K. R. Goodearl, Lectures on Algebraic Quantum Groups, Birkhauser, 2002.
[ChP] V. Chari, A. N. Pressley, A Guide to Quantum Groups, Cambridge Univ. Press, 1995.
[DL1] M. Domokos, T. H. Lenagan, Conjugation coinvariants of quantum matrices, Bull. London
Math. Soc. 35 (2003) 117-127.
[DL2] M. Domokos, T. H. Lenagan, Representation rings of quantum groups, J. Algebra 282
[G]
(2004), 103-128.
K. R. Goodearl, Semiclassical Limits of Quantized Coordinate Rings, Advances in Ring
Theory (2010), Part of the series Trends in Mathematics, 165-204.
[HL1] T. J. Hodges, T. Levasseur, Primitive Ideals of Cq[SL(3)], Commun. Math. Phys. 156,
(1993) 581-605.
[HL2] T. J. Hodges, T. Levasseur, Primitive ideals of Cq[SL(n)], J. Algebra 168 (1994), 455-468.
[KW] B. Kostant, N. Wallach, Gelfand-Zeitlin theory from the perspective of classical mechanics.
I, Progr. Math. 243, Birkhauser, Boston, 2006, Studies in Lie theory, 319 -- 364.
[LW] J.-H. Lu, A. Weinstein, Poisson Lie groups, dressing transformations and Bruhat decom-
[M]
[S]
positions, J. Differential Geometry 31, (1990) 501-526 .
Sz. Mészáros, Cocommutative Elements Form a Maximal Commutative Subalgebra in
Quantum Matrices (submitted), arxiv:1512.04353
M.A. Semenov-Tian-Shansky, Dressing transformations and Poisson group actions, Publ.
RIMS, Kyoto Univ., 21, (1985) 1237-1260.
P. Vanhaecke, Integrable Systems in the Realm of Algebraic Geometry, Springer (1996).
[V]
[W] A. Weinstein, The local structure of Poisson manifolds, J. Differential Geometry 18 (1983),
523-557.
[Y] M. Yakimov, On the spectra of quantum groups, Memoirs Amer. Math. Soc. 229 (2014),
no. 1078.
Department of Mathematics, Central European University, Budapest, 1051
E-mail address: [email protected]
12
|
1609.05594 | 1 | 1609 | 2016-09-19T04:05:08 | Geometric classification of nilpotent Jordan algebras of dimension five | [
"math.RA"
] | The variety $JorN_{5}$ of five-dimensional nilpotent Jordan algebras structures over an algebraically closed field is investigated. We show that $JorN_{5}$ is the union of five irreducible components, four of them correspond to the Zariski closure of the $GL_5$-orbits of four rigid algebras and the other one is the Zariski closure of an union of orbits of infinite family of algebras, none of them being rigid. | math.RA | math | Geometric classification of nilpotent Jordan algebras of dimension
five
Iryna Kashubaa,1,∗, María Eugenia Martina,2
aInstituto de Matemática e Estatística, Universidade de São Paulo, R. do Matão 1010, 05508-090, São Paulo,
Brazil.
Abstract
The variety JorN5 of five-dimensional nilpotent Jordan algebras structures over an alge-
braically closed field is investigated. We show that JorN5 is the union of five irreducible
components, four of them correspond to the Zariski closure of the GL5-orbits of four rigid
algebras and the other one is the Zariski closure of an union of orbits of infinite family of
algebras, none of them being rigid.
Keywords: Jordan algebra, deformation of algebras, rigidity
2010 MSC: 14J10, 17C55, 17C10.
6
1
0
2
p
e
S
9
1
]
.
A
R
h
t
a
m
[
1
v
4
9
5
5
0
.
9
0
6
1
:
v
i
X
r
a
1 Introduction
In this paper we study the geometry of nilpotent Jordan algebras over algebraically
closed fields k of characteristic (cid:54)= 2, 3. Namely, let V be an n-dimensional k-vector space,
the bilinear maps V × V → V form an n3-dimensional algebraic variety Homk(V ⊗ V, V ).
Moreover, Jordan algebra structures satisfy polynomial equations (originating from both
commutativity and Jordan identity) and therefore determine a Zariski-closed subset, which
we will call the variety of Jordan algebras of dimension n. As any affine variety, Jorn
admits decomposition in a finite number of irreducible components. Furthermore, nilpo-
tency conditions are also of polynomial type and thus the set of nilpotent Jordan algebras
in dimension n, which will be denoted by JorNn, possess a structure of algebraic variety
embedded in Jorn. The general linear group G = GL(V) acts on both Jorn and JorNn
via "change of basis", decomposing them into G-orbits which correspond to the classes of
isomorphic Jordan algebras. We study the geometry of JorN(n): dimension, decomposition
into irreducible components and their description. The crucial tool for it is the notion of
deformation of algebras: we say that J1 is a deformation of J2 if the G-orbit of J2 belongs
to Zariski closure of the G-orbit of J1. Algebra J whose orbit is an open subset in Jorn is
of particular interest, since it does not admit non-trivial deformations and its closure rises
to irreducible component. We will call such algebras rigid.
Analogously to JorNn one defines the varieties of nilpotent Lie and associative algebras.
We will denote them LieNn and AssocNn, respectively. The geometry of these two varieties
is rather different.
∗Corresponding author
Email addresses: [email protected] (Iryna Kashuba), [email protected] (María Eugenia Martin)
1The author was supported by CNPq (309742/2013-7).
2The author was supported by the CAPES scholarship for postdoc program of Mathematics, IME-USP.
1
In 1968, F. Flanigan posed a question whether every irreducible component is the closure
of a G-orbit of some rigid algebra J . In [1] he proved that every component of Assocn has
to carry an open subset of non-singular points which is either the orbit of a single rigid
algebra or an infinite union of orbits of algebras which differ only in the multiplication of
the radical. Flanigan has discovered that the second alternative does in fact occur: there is
a component of Assoc3 which consists entirely of the orbits of an infinite family of three-
dimensional nilpotent algebras.
In [2], G. Mazzola has proved that the subvariety of commutative algebras inside AssocNn
is irreducible for n ≤ 6 and it admits at least two irreducible components in dimension 7.
For the non-commutative case the description is known only for n ≤ 5. A. MakhLouf showed
that AssocN2 is irreducible, while the number of irreducible components of AssocNn when
n is equal to 3, 4 and 5 is two, four and thirteen, respectively. Further, the author estimated
a lower bound of the number of irreducible components in any dimension, see [3].
For the variety of nilpotent Lie algebras LieNn the geometric classification is known for
n ≤ 8. In 1988, F. Grunewald and J. O'Halloran in [4] established that LieNn is irreducible
when n ≤ 5. Two years later, S. Craig proved that this is the case for dimension 6 as well,
see [5]. The first reducible variety of nilpotent Lie algebras appears in dimension 7, see [6].
It was shown by M. Goze and J. M. Ancochea-Bermudez. that LieN7 decomposes into two
algebraic components each corresponding to infinite family of algebras. Finally, in [7] the
authors described all eight components of the variety of nilpotent Lie algebras of dimension
8.
As for the variety of Jordan algebras the references are rather recent. In [8, 2005], [9,
2014] it was shown that in Jorn for n ≤ 4 any irreducible component is determined by a
rigid algebra. In [10, 2011], the authors establish that the subvariety of nilpotent algebras
JorNn is irreducible for n ≤ 3 and admits two irreducible components if n = 4.
In this paper, we will show that JorN5 is the union of five irreducible components, there
are four rigid algebras and the closure of the union of G-orbits of a certain infinite family
gives rise to the fifth component.
The paper is organized as follows. Section 2 begins with some preliminaries on Jor-
dan algebras, further we establish the list of non-isomorphic nilpotent Jordan algebras of
dimension 5. Among the non-isomorphic nilpotent Jordan algebras there are four infinite
families depending on one or two parameters from k. In Section 3, we define the variety
of Jordan algebras JorNn, determine useful invariants of deformation and show how to
construct deformation between algebras or how to ensure that such deformation does not
exist. In Section 4 we give several examples for the deformation both of and into the infinite
families, while in Section 5, we prove the main result of this paper describing deformations
between the algebras in JorN5 and characterizing its the irreducible components.
2 Jordan algebras: definition and basic concepts, list of nilpotent
algebras up to dimension five
In this section we present the basic concepts, notations for Jordan algebras and the
algebraic classification of nilpotent Jordan algebras of small dimensions.
Definition 1. A Jordan k-algebra is a commutative k-algebra J satisfying the Jordan
identity, namely
((x · x) · y) · x = (x · x) · (y · x),
x, y ∈ J .
(1)
2
In a Jordan algebra J we define inductively a series of subsets by setting
J 1 = J ,
J n = J n−1 · J + J n−2 · J 2 + ··· + J · J n−1.
The subset J n is called the n-th power of the algebra J .
Definition 2. A Jordan algebra J is said to be nilpotent if there exists an integer n ∈ N
such that J n = 0. The least such number is called the index of nilpotency of the algebra
J , nilindex(J ).
Further in this section we establish a list of all non-isomorphic nilpotent algebras of
dimension five. For convenience we first describe all indecomposable nilpotent algebras
of dimension up to four, using the notations from [11]. Also we drop · and denote a
multiplication in J simply as xy. Products that we do not specify are understood to be 0 or
determined by commutativity.
Example 3 (Indecomposable nilpotent Jordan algebras of dimension one, and two three). All
nilpotent Jordan algebras of dimension less than four are associative. There is only one
nilpotent indecomposable algebra in one and two: kn, with n2 = 0 and B3 generated by n1,
with n3
1 = 0, respectively. There are two indecomposable nilpotent algebras in dimension
three, T3 generated by n1, such that n4
1 = 0 and T4 = (cid:104)n1, n2 n2
2(cid:105).
1 = n2
In [11] all Jordan algebras of dimension four over an algebraically closed field of char-
acteristic not 2 were described. In the Table 1 we list eight non-isomorphic indecomposable
nilpotent Jordan algebras.
Table 1: Indecomposable four-dimensional nilpotent Jordan
algebras.
F
F61
F62
F63
F64
F65
F66
F69
F70
1 = n2 n2
n2
Multiplication Table
2 = n4 n1n2 = n3 n1n3 = n4
4 = n2 n1n2 = n3
n2
1 = n2 n2
1 = n2 n2
n2
4 = −n2 − n3 n1n2 = n3 n2n4 = n3
n2
1 = n2 n2
n2
1 = n2 n1n2 = n3 n2n4 = n3
n2
1 = n2 n2
4 = n3 n1n2 = n3
4 = −n2 n1n2 = n3 n2n4 = n3
n2
1 = n2 n1n3 = n4
n2
1 = n2 n3n4 = n2
2.1. Nilpotent Jordan algebras of dimension five
In [2], G. Mazzola has classified nilpotent commutative associative (and thus Jordan)
five-dimensional algebras. Keeping the notations from the paper we list the 25 non-isomorphic
such algebras in Table 2. For each algebra we calculate the dimension of its automorphism
group Aut(J ), of its annihilator Ann(J ) = {a ∈ J aJ = 0}, of its second power J 2 and
the index of nilpotency nilindex(J ).
3
Table 2: Five-dimensional nilpotent associative Jordan alge-
bras.
Multiplication Table
dim
Aut
dim
Ann
dim
2
nil-
index
n2
1 = n2 n2
2 = n1n3 = n4
n1n2 = n3 n1n4 = n2n3 = n5
n2
1 = n2 n2
2 = n1n3 = n4
n1n2 = n3
n2
1 = n2 n2
4 = n5
n1n2 = n4n5 = n3
n2
1 = n3 n1n2 = n4
n1n4 = n2n3 = n5
F61 ⊕ kn5
T3 ⊕ B3
n1n3 = n4 n2n3 = n5
n1n2 = n4 − n5
n2
1 = n3 n2
2 = n1n3 = n4
n1n2 = n5
n1n3 = n5
n1n2 = n4 n2n3 = −n5
n2
1 = n3 n1n3 = n4 n1n2 = n5
n2
1 = n2n3 = n4 n1n3 = n5
n1n2 = n4 n1n3 = n5
n2
3 = n4 n1n2 = n3n4 = n5
F66 ⊕ kn5
T4 ⊕ B3
2 = n5 n1n2 = n4
3 = n1n2 = n5
1 = n3 n2
n2
n2
1 = n4 n2
T3 ⊕ kn4 ⊕ kn5
B3 ⊕ B3 ⊕ kn5
n2n3 = n1n4 = n5
F69 ⊕ kn5
F70 ⊕ kn5
T4 ⊕ kn4 ⊕ kn5
B3 ⊕ kn3 ⊕ kn4 ⊕ kn5
kn1 ⊕ kn2 ⊕ kn3 ⊕ kn4 ⊕ kn5
5
6
6
7
7
7
7
8
8
9
9
11
8
9
9
10
10
11
11
12
11
12
14
17
25
1
1
1
1
2
2
2
2
2
2
2
2
1
2
2
3
2
3
3
3
1
2
3
4
5
4
3
3
3
3
3
2
3
2
3
2
2
2
2
2
3
2
2
2
2
1
1
1
1
0
6
5
4
4
5
4
3
4
3
4
3
3
4
4
3
3
3
4
3
3
3
3
3
3
2
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
Recently, in [12] a method to construct nilpotent Jordan algebras using central exten-
sion of lower dimension algebras was presented. The authors showed that any nilpotent
algebra can be obtained via this methods and provided a list of 35 single non-associative
Jordan algebras together with 6 families of algebras depending on parameters in k. The list
of these algebras is given in Table 3 (using the notations [12]) together with the dimen-
sions of Aut(J ), Ann(J ), J 2 and of the associative center Z(J ) = {a ∈ J (a,J ,J ) =
(J , a,J ) = (J ,J , a) = 0}.
4
Table 3: Five-dimensional nilpotent non-associative Jordan
algebras.
15
J
J1
J2
J3
J4
J5
J6
J7
J8
J9
J10
J11
J12
J13
J14
J α
J16
J17
J18
J19
J20
J21
J22
J β
J γ
J 0
J − 1
J25
J δ
23
24
24
24
26
2
Multiplication Table
dim
Aut
dim
Ann
8
9
8
7
6
7
6
5
4
5
7
6
6
7
6
6
5
4
5
4
3
4
6
7
7
8
7
6
2
2
2
2
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
2
1
1
1
F65 ⊕ kn5
F64 ⊕ kn5
F63 ⊕ kn5
F62 ⊕ kn5
n2
1 = n2 n2
4 = n2n3 = n5
n2
1 = n2 n1n4 = n2n3 = n5
n1n2 = n3 n3n4 = n5
n1n2 = n3 n3n4 = n2
1 = n5
n3n4 = n2
n1n2 = n3
1 = n2
n1n2 = n3
2 = n5
n1n3 = n2
4 = n2n3 = n5
n1n2 = n3 n1n3 = n2
4 = n5
n1n3 = n2
n1n2 = n3
4 = n2
n1n2 = n3
2 = n5
n1n4 = n1n3 = n2n3 = n5
n1n2 = n3 n1n3 = n2n4 = n5
1 = n2 n1n2 = n3 n2
n2
4 = αn5
n2
2 = n1n3 = n2n4 = n5, α∈k
n1n2 = n2
3 = n4 n1n4 = n5
n1n2 = n2
3 = n4
n1n4 = n2n3 = n5
n1n2 = n2
3 = n4
n1n4 = n2
2 = n5
n1n2 = n2
3 = n4 n3n4 = n5
n1n2 = n2
3 = n4 n3n4 = n2
n1n2 = n2
3 = n4
1 = n5
1 = n2
n3n4 = n2
1 = n2 n1n2 = n2
n2
n2
2 = n1n4 = n5
2 = n5
3 = n4
n2
1 = n3 n1n3 = n2n4 = n5
n1n2 = n4 n2n3 = βn5, β∈k
n2
1 = n3 n1n2 = n4 n2n3 = n5
n1n4 = γn5, γ∈k∗−{1,− 1
2}
n2
1 = n3 n1n2 = n4 n2n3 = n5
n2
1 = n3 n1n2 = n4 n2n3 = n5
n1n4 = − 1
n2
1 = n3 n1n2 = n4
2 n5
n1n3 = n1n4 = n5 n2n3 = −2n5
n2
1 = n3 n2
2 = n4 n2n4 = δn5
n1n3 = n1n4 = n5, δ∈k∗
5
dim
J 2
2
2
2
2
2
2
2
2
dim
Z
3
3
3
3
3
3
2
2
2
2
2
2
2
2
3
2
2
2
2
2
2
3
3
3
3
3
3
3
2
3
3
3
3
3
3
2
2
2
2
2
2
3
3
3
3
3
3
3
Table 3: Five-dimensional nilpotent non-associative Jordan
algebras.
J
J 0
26
J ε,φ
27
27
27
J ε,ε−1
J −1,−1
J28
J29
J30
J31
J32
J33
J34
J35
J36
J37
J38
J39
J40
J λ
41
n2
1 = n2 n2n3 = n4 n1n2 = n5
n2
1 = n2 n2n3 = n4
n1n3 = n2
3 = n5
n2
1 = n2 n2n3 = n4
n1n2 = n1n3 = n5
n2
1 = n2 n2n3 = n4
n1n2 = n2
3 = n5
Multiplication Table
n2
1 = n3 n2
2 = n4
n1n3 = n1n4 = n5
n2
1 = n3 n2n3 = n1n4 = n5
n2
2 = n4 n2n4 = εn5
n1n3 = φn5 ε,φ∈k, εφ(cid:54)=1
n2
1 = n3 n2n3 = n1n4 = n5
n2
2 = n4 n2n4 = εn5
n1n3 = ε−1n5
ε∈k∗−{−1}
n2
1 = n3 n2n3 = n1n4 = n5
n2
2 = n4 n2n4 = n1n3 = −n5
n2
1 = n2 n2n3 = n4 n2
3 = n5
n2
1 = n2 n2n3 = n4 n1n3 = n5
n1n2 = n3 n1n3 = n4 n2
n1n2 = n3 n1n3 = n4 n2
n1n2 = n3 n1n3 = n4
1 = n5
2 = n5
1 = n2
n2
n2n3 = n5
2 = n5 n1n2 = n3
n1n3 = n4
n1n2 = n3 n1n3 = n4
n2
1 = n2n3 = n5
n1n2 = n3 n2
1 = n5
n2
2 = n1n3 = n4
n1n2 = n3 n2
2 = n1n3 = n4
n2
1 = n2n3 = n5
n2
1 = n5 n1n2 = n3 n2
2 = λn5
n1n3 = n2n3 = n4, λ∈k
dim
Aut
dim
Ann
6
6
6
6
7
7
6
7
6
5
8
7
6
6
5
7
4
6
2
1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
2
dim
J 2
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
dim
Z
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
3
Theorem 4. In Tables 2 and 3 all algebras are one-to-one non-isomorphic except for
1. J α1
2. J β1
3. J α1,β1
4. J λ1
5. J 1
27
15 (cid:39) J α2
23 (cid:39) J β2
24 (cid:39) J α2
26 (cid:39) J β2
24 if and only if α1 = α2;
26 if and only if β1 = ±β2;
15 , J α1
23 , J β1
(cid:39) J α2,β2
41 if and only if λ1 = λ2 or λ1 = λ−1
2 ;
27
41 (cid:39) J λ2
24 is in fact associative and corresponds to 4 from Table 2;
if and only if (α1, β1) = (α2, β2) or (α1, β1) = (β2, α2);
6
6. J29 (cid:39) J 0
24;
7. J37 (cid:39) J −1,−1
8. The family J λ
27
isomorphic to J 0
.
41 is isomorphic to J ε,ε−1
26. Remaining only J 0
27
for λ (cid:54)= 0, 1 and ε = − (−1+√λ)2
41, which will be denoted simply as J41.
(1+√λ)2 , while J 1
41 is
Proof. Firstly, in [12] the authors showed the isomorphism from items 1-4. The isomor-
phism 5 follows immediately from the multiplication tables of the corresponding algebras.
To see 6-8 consider the change of basis
27
J29 (cid:39) J 0
e1 = n1
e2 = n3
e3 = n2
e4 = n5
e5 = n4
24 J37 (cid:39) J −1,−1
e1 = n1 − n2
e2 = −n1 − n2
e3 = −2(n3 − n5)
e4 = 2(n3 + n5)
e5 = 2n4
26
J 1
41 (cid:39) J 0
e1 = n1 + n2
e2 = in1 − in2
e3 = 2(n3 + n5)
e4 = 2(n3 − n5)
e5 = 4n4
then the transformation
(cid:33)
n2
−1 + √λ
1 + √λ
2(−1 + √λ)2λ
(1 + √λ)2
n5
n3 +
n4
and for any λ (cid:54)= 0, 1 consider ε = − (−1+√λ)2
(1+√λ)2
(cid:32)
(cid:33)
e1 = √λn1 + n2
1 + √λ
n1 +
e2 =
e3 = 2√λn3 + 2λn5
(cid:32)√λ − λ
2(−1 + √λ)2√λ
2(−1 + √λ)2√λ
(1 + √λ)2
e4 = −
1 + √λ
e5 = −
41 (cid:39) J ε,ε−1
27
.
gives the isomorphism J λ
Further the dimension of the automorphism group, the annihilator, any power of alge-
bra, the associative center, index of nilpotency together with associative and non-associative
conditions, are invariants of class of isomorphic algebras. Therefore any two algebras whose
last four columns in Table 2 or Table 3 are not equal are non-isomorphic, and one has to ver-
ify that the algebras are non-isomorphic in the following cases:
1. J9 − J18 − J20: the four-dimensional non-associative nilpotent Jordan algebra F64 is
a subalgebra of J20 but it is not a subalgebra nor of J9 neither of J18. Analogously,
F70 is a subalgebra of J9 and is not of J18.
2. J8−J17−J19: the dimension of the second cohomology group of J19 with coefficients
in J19, H 2(J19,J19), is 6 while dim H 2(J8,J8) = dim H 2(J17,J17) = 5. For the
definition of the cohomology group for Jordan algebras we refer to [9]. On the other
hand, the four-dimensional algebra F63 is a subalgebra of J8 but is not of J17.
3. J33 (cid:54)(cid:39) J38: the three-dimensional nilpotent Jordan algebra T4 is a subalgebra of J38
but is not of J33.
7
4. J5−J12−J13: dim H 2(J12,J12) = dim H 2(J13,J13) = 13 while dim H 2(J5,J5) = 12
But the dimension of Jacobi space of J12, i.e.
Jac(J ) = {a ∈ J a(xy) = (ax)y + x(ay), ∀x, y ∈ J } ,
is 4 whereas dim Jac(J13) = 3.
5. J7 (cid:54)(cid:39) J16: the four-dimensional algebra F71 = T4 ⊕ kn4 is a subalgebra of J7 but is
not of J16.
6. J6 − J11 − J14: F63 ⊆ J14 while F63 (cid:54)⊆ J6, adding dim H 2(J11,J11) = 15 while
dim H 2(J6,J6) = dim H 2(J14,J14) = 14 we obtain the result.
24 (cid:54)(cid:39) J25: for any γ (cid:54)= 0, algebra F69 is a subalgebra of J γ
of J25. Since dim Ann(J 0
24) = 2, J25 (cid:54)(cid:39) J 0
24.
7. J γ
24 but is not a subalgebra
24) = 12 , while dim Jac(J28) =
9. J α
8. J28 − J 0
24,J 0
24 − J31 − J35 − J39: one compares the dimensions of both second co-
homology group and Jacobi space, dim H 2(J35,J35) = 10, dim H 2(J31,J31) = 8,
dim H 2(J28,J28) = dim H 2(J39,J39) = 11, and dim H 2(J 0
3 and dim Jac(J39) = 4.
26 − J ε,φ
15 − J β
algebras. Analogously, dim H 2(J β
26) = dim H 2(J ε,φ
dim H 2(J δ
Next, suppose that J ε,φ
is given by a matrix A = (aij)5
following conditions: aij = ai1ai2 = ak5 = 0 for i = 1, 2, j = 3, 4, 5, and a3k = a2
a4k = a2
for k = 3, 4. Moreover,
27 : the nilindex(J α
23,J β
27 ,J ε,φ
26 for some δ, ε, φ ∈ k and suppose that this isomorphism
i,j=1. Then the coefficients aij have to satisfy the
1(k−2),
2(k−2), a5k = +2a2(k−2)(a3(k−2) + a4(k−2)ε) + 2a1(k−2)(a4(k−2) + a3(k−2)φ) = 0,
15) = 5 for any α ∈ k while it is 4 for the other
23) = 9, while
23) = 8 if β (cid:54)= 0 and dim H 2(J 0
27 ) = 7 if δ (cid:54)= 0 and εφ (cid:54)= 1.
23 − J δ
26,J δ
27 (cid:39) J δ
23,J 0
10. J 0
a21a32 + a11a42 + a21a42ε + a41(a12 + a22ε) + a11a32φ + a31(a22 + a12φ) = 0
11φ = 0
12φ) = 0
11(a22 + a12φ) = 0
12φ = 0.
a11a21(a11 + a21) − a55 + a3
22ε) + a11(a2
a2
21(a12 + a22ε) + a2
−a55 + a21(a2
21ε + a3
22 + a2
12 + a2
22ε + a3
27 − J30 − J32 − J36 − J −1,−1
a12a22(a12 + a22) − a55δ + a3
But the system allows only trivial solutions ai5 = 0 for 1 ≤ i ≤ 5.
26 − J ε,ε−1
− J41: the nilpotency type of J32 and J36
is (2, 1, 2) while for all other algebras it is (2, 2, 1) (for definition of nilpotency type
of Jordan algebra we refer to [11]). Suppose that J36 (cid:39) J32 and the isomorphism is
given by the matrix A = (aij)5
i,j=1. Then aij's have to satisfy the following conditions:
aij = a3k = 2ai3a33 = 0 for i = 1, 2, j = 2, 4, 5, k = 4, 5, an2 = 2a11a(n−1)1, for
n = 3, 4, and a52 = 2a31a21, 2a13a23 = 0, 2a11a21a23 = a54, 2a11a21a13 = a44, a31a23 +
11a21, a21a13 + a11a23 = 0. This
a21a33 = a55 = 2a11a2
implies that ai4 = 0 for 1 ≤ i ≤ 5.
Comparing the dimensions of the cohomologies we rule out some cases:
21, a31a13 + a11a33 = a45 = 2a2
27
dim H 2(J −1,−1
dim H 2(J ε,ε−1
27
27
27
,J −1,−1
,J ε,ε−1
dim H 2(J 0
) = dim H 2(J41,J41) = 8,
) = dim H 2(J30,J30) = 7
26,J 0
26) = 10.
27
8
27
27
27
On the other hand, dim Jac(J −1,−1
check that J ε,ε−1
from J30 to J ε,ε−1
11 = a23, 2a21a31 = a43, 2a11a31 + a2
a2
a11a32 + a31(a12 + a32) = 0, a32a24ε−1 = a31a23ε = a45 = a31a24 = a32a23, a2
2a22a32 = a44 and 2a12a32 + a2
a contradiction.
) = 4 whereas dim Jac(J41) = 3. It remains to
i,j=1 which gives an isomorphism
(cid:54)(cid:39) J30. Suppose that A = (aij)5
, then aij = ak5 = 0 for i = 1, 3, j = 3, 4, 5 and k = 2, 5. Moreover,
31 = a53, a11a12 = 0, a31a22 + a21a32 = 0,
12 = a24,
32 = a54. All the solutions lead to ai5 = 0 for 1 ≤ i ≤ 5,
3 Deformation, variety of Jordan algebras, constructing defor-
mations
k=1 ck
namely, ei · ej =(cid:80)n
Let V be a n-dimensional k-vector space with a fixed basis {e1, e2,··· , en}. To endow V
with a Jordan k-algebra structure, (J ,·), it suffices to specify n3 structure constants ck
ij ∈ k,
ij) ∈ kn3 defines Jordan algebra
its coordinates satisfy certain polynomial equations which follow from commutativity and
Jordan identity. These equations cut out an algebraic variety Jorn in kn3. The points
ij) ∈ Jorn are in one-to-one correspondence with n-dimensional k-algebras along with
(ck
a particular choice of basis. A change of basis in J is given by the general linear group
G = GL(V) action on Jorn, namely:
ijek, i, j ∈ {1, 2, . . . , n}. If a point (ck
g(J ,·) (cid:55)→ (J ,·g),
x ·g y = g(g−1x · g−1y),
(2)
for any J ∈ Jorn, g ∈ G and x, y ∈ V. The G-orbit of a Jordan algebra J , that is, the set of
all images of J under the action of G, is denoted by J G. The set of different G-orbits of this
action is in one-to-one correspondence with the set of isomorphism classes of n-dimensional
Jordan algebras. The Zariski topology of the affine space provides the following relation of
orbits, namely, we say that J1 is a deformation of J2 or that J1 dominates J2, if the orbit
1 . We denote it by J1 → J2. A Jordan
2 is contained in the Zariski closure of the orbit J G
J G
algebra J is called rigid if its G-orbit is a Zariski-open set in Jorn. Rigid algebra J does
1 ∩ J G (cid:54)= ∅ and therefore
not admit non-trivial deformations, indeed if J1 → J then J G
J1 (cid:39) J .
As we have already mentioned in the introduction the rigid algebras are of particular
interest. Recall that any affine variety could be decomposed into its irreducible components.
When J ∈ Jorn is rigid, there exists an irreducible component C such that J G ∩C is a non-
empty open subspace in C, and therefore the closure of J G coincides with C. The converse
is not true: in the sections 5 we will show that there exists an irreducible component which
does not contain a single open G-orbit but rather an open subset which consists of infinite
union of orbits of algebras.
Any (ck
ij) ∈ Jorn which corresponds to a nilpotent Jordan algebra has to satisfy polyno-
mial equation, thus the set of nilpotent Jordan algebras in dimension n possess a structure
of algebraic variety embedded in JorNn ⊆ Jorn. Further in this section we collect the prop-
erties of deformation which will be used to describe JorN5. The first fact is straightforward
but very useful.
Lemma 5. [9] Let Ji and J (cid:48)i be in Jorni, for i = 1, 2. If Ji → J (cid:48)i , then J1 ⊕ J2 → J (cid:48)1 ⊕ J (cid:48)2.
Example 6. From the description of JorN4 in [10] it follows that F62 → F63 → F64 and
F62 → F65, consequently J4 → J3 → J2 and J4 → J1.
9
The other way to deform one algebra into another is to build a curve as specified in
Lemma 7.
Lemma 7. [2]Suppose that there exists a curve γ in Jorn which generically lies in a subvariety
Example 8. Let γ(t) =(cid:8)et
U and which cuts J G in special point then J ∈ U.
1 = √2t2n1 −
1,et
√2t2n2 + 2t2n3 − t2n4, et
2 = 2√2t4n2 + 4t4n4, et
4 = 4t6(n4 + n5) and
5 = 8t8n5. Thus γ defines a curve in JorN5 which for any t (cid:54)= 0 belongs to J G
21 (observe
et
4 and for t → 0 we get the multiplication table
that det γ(t) = −28t23). However, et
2et
of J18. Thus, from Lemma 7, it follows that J18 ∈ J G
(cid:9) be a change of basis of J21 given by: et
3 = √2t3(n1 + n2), et
21, i.e., J21 → J18.
3 = tet
2,et
3,et
4,et
5
The non-existence of deformation for a given pair of algebras J1,J2 ∈ Jorn will follow
from the violation of one of the conditions below.
Theorem 9. If J1 → J2 then:
1. dim Aut(J1) < dim Aut(J2);
2. dim Ann(J1) ≤ dim Ann(J2);
3. dim(J m
4. dim Z(J1) ≤ dim Z(J2).
1 ) ≥ dim(J m
2 ), for any positive integer m;
basis for J ∈ Jorn and a =(cid:80)n
Proof. For the proof of (1) to (3) we refer to [9]. To see (4), it suffices to prove that the sets
{J ∈ Jorn dim Z(J ) ≥ s} are Zariski-closed in Jorn for any s ∈ N. Let e1, e2, . . . , en be a
i=1 αiei ∈ Z(J ). Substituting a in each of the 3n2 equations
(a, ei, ej) = 0, (ei, a, ej) = 0 and (ei, ej, a) = 0 for i, j = 1, . . . , n, we obtain a system with
3n3 equations having αi as unknowns. Then dim Z(J ) = n − rank(P3n3,n), where P3n3,n
is the matrix of the system, thus the condition dim Z(J ) ≥ s is equivalent to the fact that
all (n − s + 1)-minors of P3n3,n vanish, which is defined by a finite number of polynomial
identities in the structure constants.
As direct consequences of (3) and (4), respectively, we have:
Corollary 10.
1. If J1 and J2 are nilpotent Jordan algebras such as J1 → J2 then nilindex(J1) ≥
nilindex(J2).
2. Any deformation of a non-associative algebra is a non-associative algebra.
4 Deformations of infinite families of algebras
Whenever the class of algebras consists of only finitely many non-isomorphic algebras,
the irreducible components are completely determined by rigid algebras, see Jorn, n ≤ 4.
When we have infinite number of non-isomorphic algebras it is not necessarily that any
irreducible component contains an open orbit. In [1], Flanigan showed that there exists an
infinite family of nilpotent algebras of dimension three such that the closure of a union of
their orbits is an irreducible component in AssocN3, while it carries no open orbit. In the
next section, we will show that such irreducible component exists in JorN5. In general,
the behaviour of infinite families under deformation is of particular interest and in what
10
follows we will look into examples of deformations in JorN5 which involve families. First,
to establish the notation denote by
(cid:91)
α∈T
N #
i =
i )G ,
(J α
the union of the orbits of algebras in the family J α
4.1. Infinite family is dominated by a single algebra
i , α ∈ T where T is some subset of k.
3 )t2n4 + α2
(cid:1)2 = t4et
Since(cid:0)et
The infinite family J α
15, α ∈ k, is dominated by algebra J22. To see that, for any
α ∈ k consider a change of basis of J22 parametrized by γα(t): et
3 n2 + tn3,
3 n4 and
2 = t4n2 + (1 − 2α
et
5 = t8n5. Observe that det(γα) = t25 thus, for any t (cid:54)= 0, the curve γα(t) belongs to J G
22.
et
15. Therefore
4
N #
22.
15 ⊂ J G
4.2. Infinite family is a deformation of a single algebra
5 it follows that γα(0) gives the multiplication table of J α
3 = t6n4 − (−1 + α)t4n5, et
4 = t4n2 − t5n3 + αt2
1 = t2n1 − α
9 n5, et
3 +αet
23, meaning that, for any t (cid:54)= 0, γ(t) ⊂ N #
23 and cuts J G
23, β ∈ k to J41 is quite remarkable, since no single
Indeed dim Aut(J41) = dim Aut(J β
23) for all β ∈ k.
2 n2,
5 = t2n3 − tn4. So the curve γ(t) lies transversely
41 in t = 0. Thus
1 = tn1 − 1
, where β(t) = 1
t , namely: et
23
The deformation from the family J β
4 = t
2 n3, et
4 n5 and et
member of J β
23 can dominate J41.
Consider the "variable" change of basis of J β(t)
2 n2, et
3 = − t
2 = − 1
et
to the orbits of J β
J41 ∈ N #
as a deformation of N #
defines the multiplication in J α0
(cid:0)
this case J 1
2 = n2 + (−1 + t)n4, et
et
J 1
(cid:1)G and cuts J G
k . We use the basis of J α
15
23. We will abuse the notation and will denote this fact by J β
Observe that in the same way we obtain that any algebra J α
23 → J41.
k of a family can be viewed
k from Table 3, then for any fixed α0 γ(α)
k while for all other values of α algebra belongs to N #
k .
In a rather different way, we obtain the dominance of J39 by the family J α
15 (a single member of J α
15, α ∈ k) dominates J39. Consider the curve: et
4 = t2n5 and et
(cid:0)
39 in t = 0 and G
5 in t = 1, respectively. Then we have
15, α ∈ k. In
1 = tn1,
5 = t2n2. For any t (cid:54)= 0, 1, the curve lies in
(cid:0)
3 = tn3, et
(cid:91)
(cid:1)G ⊆
α∈k
(cid:1)G ⊆ N #
15
J α
15
J39 ∈
J 1
15
and, therefore, J α
4.3. Deformations between families
15 → J39.
The family J ε,φ
27 since dim Aut(J δ
26 deforms into any single member of J ε,φ
27 , ε, φ ∈ k and εφ (cid:54)= 1, dominates the family J δ
26, δ ∈ k∗. Again no single
26) = dim Aut(J ε,φ
27 ),
member of J δ
for all δ, ε, φ ∈ k. For any fixed δ ∈ k∗, we construct a curve γδ(t) which lie trans-
versely to the orbits of J ε,φ
, where
27
3 = t2δ2n3, et
4 = δ2n4 and
1 = tδn1, et
(ε(t), φ(t)) = (tδ, 1
5 = tδ3n5. For any t (cid:54)= 0, the curve lies in N #
et
t = 0. Thus J δ
Differently, we obtain the dominance of the family J γ
J β
23, β ∈ k.
where β0 = (1−2γ0)i
27 . Consider the "variable" change of basis of J ε(t),φ(t)
J δ
26 ∈ N #
In this case, fixing γ0 ∈ k∗ − {1,− 1
1 = − i√γ0
√γ0
2} by the family
24 deforms into J β0
23 ,
n4,
3 = − 1
2}, the algebra J γ0
n1 + n2, et
2 = tn2, et
2 = δn2, et
27 and, since et
5, cuts (cid:0)
24, γ ∈ k∗ − {1,− 1
27, for any δ ∈ k∗.
t2 ), defined as: et
. Define the curve: et
(cid:1)G in
n3 − 2i√γ0
3 = tet
2et
γ0
26
11
n5. For any t (cid:54)= 0, the curve lies in the G-orbit of J β0
23 and,
n4 and et
4 = − it√γ0
et
since et
2et
4 = tγ0et
5, cuts(cid:0)
5 = − it√γ3
J γ0
24
0
(cid:1)G in t = 0. Therefore
(cid:17)G
(cid:16)
J γ0
24 ∈
J β0
24, for all γ ∈ k∗ − {1,− 1
2}.
23
⊆ N #
23
23 → J γ
consequently J β
5 The algebraic variety JorN5
In this section, we will describe the algebraic variety of five-dimensional Jordan algebras
JorN5. First we show that any structure can be deformed either into one of the given four
algebras or into the algebras in family J ε,φ
27 . The main tool of the proof is Lemma 7, to
define a curve γ(t) we use notations and basis of Table 3.
40 or N #
Lemma 11. Every structure J ∈ JorN5 belongs to G
27.
Proof. All associative structures, namely i for i = 1, . . . , 25, belong to G
1 . In fact, G. Maz-
zola in [2] showed that the subvariety of associative algebras is an irreducible subvariety in
JorN5 and that the orbit G
1 is open, thus its closure contains all associative structure. From
Corollary 10 it follows that there is no non-associative structure which deforms into 1.
22, J G
21, J G
1 , J G
All algebras Ji for i = 1, . . . , 20 except for the family J α
15, are dominated by J21. Exam-
ple 8 gives J21 → J18. For the deformation J18 → J17, consider a family of algebras with
the following basis
8t2 n4,
1 = −t2n1 + 1
2 n2 − tn3 + 1
et
2 = −t8n2 − 3t6
et
2 n4,
3 = t4n2 − t5n3 + t2
et
et
4 = t10n4 + t8n5,
et
5 = −t12n5.
2 n4,
(3)
algebras isomorphic to J18, while for t = 0 one obtains J17 (indeed, (cid:0)et
Here {n1, . . . , n5} is the basis of J18 given in Table 3, t ∈ k and therefore this family can
be viewed as a curve γ(t) ∈ JorN5. For any fixed t (cid:54)= 0 the change of basis (3) determines
5).
Analogously, the change of basis et
5 = t2n5 of
J17 does the trick for J17 → J16. Applying Lemma 7 to given transformations we obtain
the following deformations:
(cid:1)2 = −t4et
4 = t2n4 and et
2 = t2n2, et
3 = tn3, et
1 = n1, et
2
2 n2 + n3
t3 n1
t n2
t2 n3
t4 n4
t6 n5
J19 → J12
1 = n1 − 1
et
et
2 = tn3 + n4
et
3 = tn4 + n5
et
4 = it(n2 − n3)
J21 → J20 J20 → J19
1 = 1
1 = t2n1
et
et
2 = 1
et
et
2 = n2
3 = 1
et
et
3 = tn3
4 = 1
et
4 = t2n4
et
5 = 1
et
et
5 = t3n5
J12 → J11
et
1 = n1
et
2 = t2n2
et
3 = t2n3
et
4 = tn4
5 = t2n5
et
To obtain J21 → J9 consider the following change of basis et
5 = 1
(cid:1) n4 and et
n2 − n3 +(cid:0) 1
n1 + 1√2
2t2 n4, et
et
5 = tn5
4t
n2,
1 = − 1√2t
2t2 n5 of J21. Deformations
4 = − 1√2
3 = − 1
et
J4 → J3 → J2 and J4 → J1 follow from Example 6. Suppose {n1, . . . , n5} is the basis of
J9, then et
1 = tn1 − 1
2 n3 + t3n4
5 = t4n5 for any t (cid:54)= 0 defines algebras which are isomorphic to J9. Meanwhile
and et
5 thus, when t tends to zero, we get the structure constants of algebra J10. The
4 = tet
et
3et
2 − 1
2 n3 + t2n4, et
3 = t2(n3 − n5), et
2 = tn2 − 1
2 n3 + t2n4, et
2 = 1√2t
4 = t
n1, et
J20 → J8
et
1 = tn1
et
2 = n2
et
3 = tn4
4 = tn3 + t2n4
et
et
5 = t2n5
12
1 = n1− t2n2 +√2tn4, et
n4
5 = −t6n5 of J10 does the trick for J10 → J13. To complete the description of orbits
21 use Lemma 7 for the following cases:
2 = −t2n1 + n3−√2tn4, et
3 = t4n3− 4t2n5, et
2 n3− t5
2√2
4 = t4
basis et
and et
inside J G
1 = it√2
et
(n1 − n2)
J13 → J14
et
1 = n2
et
3 = n3
et
4 = n5
et
5 = tn1
et
2 = t(−n1 + n2)
J7 → J4
et
1 = n1 + n4
et
2 = n2 + n4
J8 → J7
et
1 = n2
et
2 = tn1
et
3 = tn3
et
4 = n4
et
5 = tn5
J10 → J5
et
2 = t2n3
3 = t2
et
2 (n1 + n2)
et
4 = t2n4
5 = t4n5
et
15, α ∈ k deforms into algebra
15 → J39 follow from Sections 4.1 and 4.2. To show that
3 = tn3,
5 = n5 of J39. The deformation J22 → J33 is given by the change of basis
2 = t4n2 + n4, et
5 = t6n4 + t2n5 of J22. The
J22. Deformations J22 → J α
J39 → J34 we use again Lemma 7 applying it to the curve γ(t) : et
4 = tn4 and et
et
4 = t5n5 and et
1 = t2n1 + n3, et
et
curve γ(t) :
Any algebra J28, J30 -J34, J39 as well as the family J α
J5 → J6
et
1 = n1 + n4
et
2 = n2 + n5
et
3 = tn3
et
4 = tn4
et
5 = tn5
et
3 = −tn3
et
4 = n4
et
5 = −tn5
2 = tn2, et
1 = n1, et
3 = tn2 − t3n3, et
et
1 = t2n1 + n2 + tn3,
et
2 = t4n2 + 2tn4 + 3t2n5,
3 = t4
et
4 = 2t9
et
et
5 = t5n4 + t6n5
3 n2 + 2t5
3 n4,
3 n3,
27 n4 + t2
2 n3, et
4 = t5
1 = n1, et
2 = n2, et
2 n4 and et
3 = 1
3 = 1
12 n2 + t
2 n2 + n3, et
4 = −tn4 and et
belongs to J G
basis et
3 = t(n2 − n3), et
(cid:1)2 = 2tet
defines algebra isomorphic to J32. Observe that (cid:0)et
33 for any t (cid:54)= 0 and intersects J32 when t = 0, thus J33 → J32. The new
5 = n5 of J32 for any t (cid:54)= 0
4, thus we get the structure
3 n3,
4 n5 of J33 and
5 = n4 + n5 of J30 give J33 → J30
t2 n1,
1 = 1
3 = tn3,
J31 when t tends to zero, i.e. J32 → J31. The changes of basis et
3 n5, et
2 = t4n2 + 2t
5 = t
et
2 = t2n2, et
1 = t(n1 + n2), et
et
and J30 → J28, respectively.
4 = 1
2 = 1
et
4 = tn5 and et
et
Next, J40 → J38 → J36. Indeed, we apply Lemma 7 to the change of basis et
2 = tn1, et
t4 n5 of J40 and et
5 = 1
5 = t2n4, of J38, respectively.
Finally, the orbits of remaining structures, i.e. families J β
J25, J35, J41, belong to the closure of N #
23 → J γ
J β
24 follow from Sections 4.2 and 4.3.
Further, we show that J ε,φ
J ε(t),φ(t)
:
23, J γ
27. The deformations J ε,φ
26 as well as algebras
23 → J41 and
23 for any fix β0 ∈ k. Consider the change of basis of
1 = t2n1 + 1
12 n4 + t2
24, J δ
27 → J δ
4 = t2n4 and et
27 → J β0
t5 n4 and et
1 = n2, et
t n2, et
3 = 1
t3 n3, et
26, J β
9 n2 + t
27
3
(cid:18)
et
1 = (−1 − tβ0)n1 − n2,
et
2 = tn2,
et
3 = (1 + tβ0)2n3 + n4,
et
4 = −tn4,
et
5 = t2(1 + tβ0)n5,
(ε(t), φ(t)) =
(cid:19)
.
−1 − tβ0, −1 − t2 − tβ0
(1 + tβ0)2
Note that ε(t)φ(t) = 1 + t2
for t = 0 it gives J β0
23 .
and et
1+tβ0
Applying Lemma 7 to the basis transformation et
5 = −2t2n3 + t2n5 of J41 we obtain J41 → J35.
then, while for any t (cid:54)= 0 this basis defines structure in N #
27,
1 = n2, et
2 = tn1−tn2, et
3 = tn3, et
4 = tn4
13
24
27
(cid:17)2
27
2 = −2tn2, et
, γ(t) = t−1
24. For any t (cid:54)= 0, φ(t) ⊂ N #
27 is not contained in the Zariski closure of the orbits of any
To show that J γ
1 = n1 + n2, et
5 = tn5, where {n1, . . . , n5} is a basis of J γ(t)
24 → J25, define a curve φ(t) ⊂ JorN5, as been the set of structure
3 = n3 + 2n4, et
4 = −2tn4
2 . Analogously to Section
24, while φ(0)
constants of algebras with the basis et
and et
4.2, the curve lies transversely to the orbits of J γ
corresponds to the structure constants of J25.
Lemma 12. The infinite union N #
algebra in JorN5.
(cid:16)
Proof. By Corollary 10 the non-associative algebra J ε,φ
can not be deformed into the
associative algebra 1 for any ε, φ ∈ k. Observe that for any ε, φ ∈ k with εφ (cid:54)= 1,
27 ) = 1. Applying Theorem 9 item 3 and item 2 we
dim
obtain that J21 (cid:57) J ε,φ
27 , first observe that J22 does not deform into the other three
algebras from Lemma 11. Indeed, comparing the dimensions of the automorphism group
we deduce that 1 (cid:57) J22 and J40 (cid:57) J22, while the dimensions of the second power provide
that J G
22 is a irreducible
component of JorN5. Since N #
27 is contained in the irreducible component C and both has
the same dimension dimN #
27 is a dense subset of C and
22 is open in C then there exists J ∈ N #
22, i.e., J (cid:39) J22, a contradiction
J G
by Theorem 4.
J ε,φ
To see that J22 (cid:57) J ε,φ
= 3 and dim Ann(J ε,φ
27 and J40 (cid:57) J ε,φ
22 then J22 is rigid and C = J G
27 , when εφ (cid:54)= 1, respectively.
27 = 21 = dimC, N #
21. Suppose that N #
27 such as J ∈ J G
27 = C. So, N #
(cid:42) J G
22
27 ⊆ J G
27
(cid:16)
(cid:17)G
27 belong
and
In particular, we proved
27 is a product of two irreducible varieties
27 determines a component of the variety JorN5.
Corollary 13. N #
that a finite-dimensional Jordan algebras not necessarily deforms into rigid one.
Proof. From Lemma 12 it follows that we only have to show that all algebras in N #
to the same component. In fact N #
J ε,φ
k2 = {(ε, φ) ε, φ ∈ k} and is therefore irreducible.
Lemma 14. Algebras 1, J21, J22 and J40 are rigid.
Proof. By a similar argument as in proof of Lemma 12 we can show that J G
27 and
therefore J22 is rigid. To show that the other three algebras are rigid it is sufficient to prove
that there are no deformations between them and that none of these algebras is dominated
by the family J ε,φ
27 .
Using the dimension argument for the second power of an algebra we have J21 (cid:57) 1,
J22 (cid:57) 1 and J21 (cid:57) J40, while comparing the dimension of the annihilator we obtain
J40 (cid:57) 1. Thus we have shown that 1 is a rigid algebra in JorN5 too. The associative al-
gebra 1 can not dominate non-associative algebras, thus 1 (cid:57) J21 and 1 (cid:57) J40. Also, due
to the dimension argument for the automorphism group we get J22 (cid:57) J40 and comparing
the dimension of the associative center we have J22 (cid:57) J21 and J40 (cid:57) J21. On the other
(cid:42) N #
hand dimJ G
27 and therefore J21 is a rigid
algebra. Finally suppose that J40 belongs to the component N #
27. Note that the dimensions
40 and N #
of J G
27 are both equal to 21, therefore
21 = 22 while dimN #
27 = 21 leading to J G
22 (cid:54)⊆ N #
21
Which is impossible since dim Ann(J40) = 2, while dim Ann(J ε,φ
εφ (cid:54)= 1. Consequently J40 is a rigid algebra.
27 ) = 1 for ε, φ ∈ k with
40 = N #
J G
27.
14
Finally, note that the change of basis given by identity matrix I5 multiplied by t guar-
antees J → 25 for any J ∈ JorN5. Consequently, JorN5 is a connected affine variety.
Therefore we have proved the following theorem.
Theorem 15. The algebraic variety JorN5 of five-dimensional nilpotent Jordan algebras is
a connected affine variety with an infinite number of orbits under the action of the group G.
Furthermore, it has four irreducible components given by Zariski closures of the orbits of the
rigid algebras 1, J21, J22 and J40 and one more component that is the Zariski closure of the
infinite union N #
27.
The diagram of JorN5 was represented in Figure 1.
15
Figure 1: The irreducible components of JorN5
16
222120191817160all associative algebrasdimCJ1J2J3J4J5J6J7J8J9J10J11J12J13J14Jα15J16J17J18J19J20J21J22Jβ23Jγ24J024J−1224J25Jδ26J026Jε,φ27Jε,ε−127J−1,−127J28J30J31J32J33J34J35J36J38J39J40J41ǫ1ǫ25 6 References
[1] F. Flanigan, Algebraic Geography: Varieties of Structure Constants, Pacific Journal of Mathematics 27 (1)
(1968) 71–79.
[2] G. Mazzola, Generic Finite Schemes and Hochschild Cocycles, Commentarii Mathematici Helvetici 55
(1980) 267–293.
[3] A. Makhlouf, The irreducible components of the nilpotent associative algebras, Revista Matemática Com-
plutense 6 (1) (1993) 27.
[4] F. Grunewald, J. O'Halloran, Varieties of nilpotent Lie algebras of dimension less than six, Journal of
Algebra 112 (2) (1988) 315–325.
[5] S. Craig, Degenerations of 6-dimensional nilpotent lie algebras over C, Communications in Algebra
18 (10) (1990) 3493–3505.
[6] M. Goze, J. A. Bermudez, On the varieties of nilpotent lie algebras of dimension 7 and 8, Journal of pure
and applied algebra 77 (2) (1992) 131–140.
[7] J. M. Ancochea-Bermudez, J. R. Gómez-Martin, G. Valeiras, M. Goze, Sur les composantes irréductibles
de la varieté des lois d'algèbres de lie nilpotentes, Journal of Pure and Applied Algebra 106 (1) (1996)
11 – 22.
[8] I. Kashuba, I. Shestakov, Jordan Algebras of Dimension Three: Geometric Classification and Represen-
tation Type, Actas del XVI Coloquio Latinoamericano de Álgebra (Colonia del Sacramento, Uruguay).
Biblioteca de la Revista Matemática Iberoamericana 1 (2005) 295–315.
[9] I. Kashuba, M. E. Martin, Deformations of Jordan Algebras of Dimension Four, Journal of Algebra 399
(2014) 277–289.
[10] J. M. Ancochea Bermúdez, J. Fresán, J. M. Bentabol, Contractions of Low-Dimensional Nilpotent Jordan
Algebras, Communications in Algebra 39 (3) (2011) 1139–1151.
[11] M. E. Martin, Four Dimensional Jordan Algebras, International Journal of Mathematics, Game Theory
and Algebra 20 (4) (2013) 41–59.
[12] A. Hegazi, H. Abdelwahab, Classification of five dimensional nilpotent jordan algebras, arXiv preprint
arXiv:1401.3992.
17
|
0901.4888 | 3 | 0901 | 2012-02-17T14:18:48 | Polynomial functions over bounded distributive lattices | [
"math.RA",
"math.CO"
] | Let L be a bounded distributive lattice. We give several characterizations of those L^n --> L mappings that are polynomial functions, i.e., functions which can be obtained from projections and constant functions using binary joins and meets. Moreover, we discuss the disjunctive normal form representations of these polynomial functions. | math.RA | math |
POLYNOMIAL FUNCTIONS OVER BOUNDED
DISTRIBUTIVE LATTICES
MIGUEL COUCEIRO AND JEAN-LUC MARICHAL
Abstract. Let L be a bounded distributive lattice. We give several
characterizations of those Ln
→ L mappings that are polynomial func-
tions, i.e., functions which can be obtained from projections and con-
stant functions using binary joins and meets. Moreover, we discuss the
disjunctive normal form representations of these polynomial functions.
Keywords: Distributive lattice; polynomial function; normal form; func-
tional equation.
1. Introduction
Let (L; ∧, ∨) be a lattice. With no danger of ambiguity, we denote lattices
by their universes. By a (lattice) polynomial function we simply mean a map
f : Ln → L which can be obtained by composition of the binary operations ∧
and ∨, the projections, and the constant functions; see, e.g., page 93 in [2]. If
constant functions are not used, then these polynomial functions are usually
referred to as term functions. For general background, see [2, 5, 6, 10]
For a finite lattice L, the set of all polynomial functions on L is well
understood.
Indeed, Kindermann [11] reduces the problem of describing
polynomial functions to tolerances, and reasonable descriptions for the latter
have been provided in Czedli and Klukovits [4] and Chajda [3].
The goal of the current paper is to present a more direct approach to
polynomial functions and provide alternative descriptions, different in nature
and flavor, in the case when L is distributive, with 0 and 1 as bottom and
top elements. Notably enough, instead of finiteness it suffices to assume that
L is a bounded distributive lattice. So, throughout the paper, we assume
that L is a bounded distributive lattice. Also, functions that are not order-
preserving cannot be polynomial functions. Thus, our main result focuses
on order-preserving functions.
We shall make use of the following notation. The ternary median term
(x∨y)∧(x∨z)∧(y∨z) will be denoted by med(x, y, z). For c ∈ L, the constant
tuple (c, . . . , c) in Ln will be denoted by c. For k ∈ [n] = {1, . . . , n}, a ∈ L,
k be the tuple in Ln whose ith component is
and x = (x1, . . . , xn) ∈ Ln, let xa
a, if i = k, and xi, otherwise. Let [x]c (resp. [x]c) denote the n-tuple whose
ith component is 0 (resp. 1), if xi 6 c (resp. xi > c), and xi, otherwise.
Date: December 2, 2009.
1
2
MIGUEL COUCEIRO AND JEAN-LUC MARICHAL
k).
−k(x) = f (ax
The range of a function f : Ln → L is defined by R(f ) = {f (x) : x ∈ Ln}.
For k ∈ [n] and a = (a1, . . . , an) ∈ Ln, we also define the unary function
−k : L → L as f a
f a
Note that if f : Ln → L is a polynomial function, then every unary func-
tion g obtained from f by substituting constants for n−m of its variables and
identifying the remaining m variables is also a polynomial function. In fact,
it is not difficult to see that such a function g is of the form med(a, x, b),
for some a, b ∈ L, and thus every such function g preserves ∧ and ∨. In
particular, we have the following fact.
Fact 1. Every polynomial function f : Ln → L satisfies condition (1) and
its dual, where
(1)
f (xa∧b
k
) = f (xa
k) ∧ f (xb
k)
for all x ∈ Ln, a, b ∈ L, and k ∈ [n].
Moreover, every function f : Ln → L satisfying (1) or its dual is order-
preserving.
The following result reassembles the various characterizations of polyno-
mial functions provided in this paper, and its proof is given in Section 2.
Main Theorem. Let L be a bounded distributive lattice and f : Ln → L,
n > 1, be an order-preserving function. The following conditions are equiv-
alent:
(i) f is a polynomial function;
(ii) for every x ∈ Ln and k ∈ [n],
(2)
f (x) = med(cid:0)f (x0
k), xk, f (x1
k)(cid:1);
(iii) f satisfies (1) and its dual, the sets R(f ) and R(f a
−k), for every
a ∈ Ln and k ∈ [n], are convex, and for every x ∈ Ln and k ∈ [n],
(3)
f (x)
f (x
k
) = f (x);
(iv) f satisfies condition (4) and its dual, where
(4)
(5)
f (x ∧ c) = f (x) ∧ c
for all c ∈ [f (0), f (1)];
(v) f satisfies the dual of (1), and (4), and
f (x) = f (x ∧ c) ∨ f ([x]c)
for all c ∈ [f (0), f (1)];
(vi) f satisfies (1) and its dual, (5) and its dual, and
(6)
f (c) = c
for all c ∈ [f (0), f (1)].
Even though not evident at the first sight, note that from the Main Theo-
rem it follows that (v) is equivalent to its dual. Note also that every function
satisfying (2) is order-preserving. The equivalence between (i) and (ii) was
first established in [14, Theorem 17]. Moreover, by Fact 1, it follows that
the order-preservation condition is redundant for all except assertion (iv).
POLYNOMIAL FUNCTIONS OVER BOUNDED DISTRIBUTIVE LATTICES
3
Let 2[n] denote the set of all subsets of [n]. If α : 2[n] → L is a mapping,
then
(7)
_
I⊆[n]
(cid:0)α(I) ∧ ^
i∈I
xi(cid:1)
is called a disjunctive normal form over L. For a function f : Ln → L,
let DNF(f ) denote the set of those maps α : 2[n] → L for which (7), as an
Ln → L mapping, coincides with f . Observe that DNF(f ) = ∅ if f is not
a polynomial function. For I ⊆ [n], let eI be the characteristic vector of
I, i.e., the tuple in Ln whose ith component is 1 if i ∈ I, and 0 otherwise.
Define αf : 2[n] → L, I 7→ f (eI ).
Lemma 2 (Goodstein [8]). If L is a bounded distributive lattice and f : Ln →
L a polynomial function, then αf ∈ DNF(f ). In particular, each polynomial
function has a disjunctive normal form representation.
By Lemma 2, for each polynomial function f : Ln → L, we have that
(8)
f is uniquely determined by its restriction to {0, 1}n.
It is noteworthy that, by (8), term functions are exactly those polynomial
functions f : Ln → L for which {0}, {1}, and {0, 1} constitute subalgebras
of (L, f ). In addition to the Main Theorem, we prove the following result
strengthening Lemma 2.
Proposition 3. Let L be a bounded distributive lattice, f : Ln → L a poly-
nomial function, and α : 2[n] → L a mapping. Then α ∈ DNF(f ) if and only
if WJ⊆I α(J) = αf (I) for all I ⊆ [n].
Using Proposition 3, it is straightforward to construct examples of lattices
L and polynomial functions f : Ln → L for which DNF(f ) > 1, and to
provide some technical conditions characterizing those polynomial functions
f for which DNF(f ) = 1. The trivial details are left to the reader.
2. Technicalities and proofs
In this section we provide the proofs of the Main Theorem and Proposition
3. First, we prove the latter.
Proof of Proposition 3. Let L be a bounded distributive lattice, f : Ln →
L a polynomial function, and α : 2[n] → L a mapping.
Suppose first that α ∈ DNF(f ). Then, for every I ⊆ [n], αf (I) = f (eI) =
WJ⊆I α(J). Now suppose that WJ⊆I α(J) = αf (I), for all I ⊆ [n], and let
g : Ln → L be the polynomial function such that α ∈ DNF(g). Clearly, for
every I ⊆ [n], we have g(eI ) = WJ⊆I α(J) = αf (I) = f (eI ). From (8) it
follows that g = f and hence, α ∈ DNF(f ).
(cid:3)
To prove the Main Theorem we will need some auxiliary results. We
proceed by focusing first on the conditions given in (iv), (v), and (vi).
Lemma 4. Every polynomial function f : Ln → L satisfies (4) and its dual.
4
MIGUEL COUCEIRO AND JEAN-LUC MARICHAL
Proof. Let f : Ln → L be a polynomial function. For any c ∈ [f (0), f (1)],
we have
f (x ∧ c) = _
(xi ∧ c)(cid:17) = f (0) ∨ _
(cid:16)αf (I) ∧ ^
(cid:16)αf (I) ∧ ^
(xi ∧ c)(cid:17)
I⊆[n]
i∈I
I⊆[n]
I6=∅
i∈I
= (cid:16)f (0) ∨ _
I⊆[n]
I6=∅
(cid:0)αf (I) ∧ ^
i∈I
xi(cid:1)(cid:17) ∧ c = f (x) ∧ c.
Similarly, it follows that f satisfies the dual of (4).
Lemma 5. Let f : Ln → L be an order-preserving function. If f satisfies
(4) or its dual, then it satisfies (6).
In particular, R(f ) coincides with
[f (0), f (1)] and thus is convex.
(cid:3)
Proof. If f satisfies (4), then for any c ∈ [f (0), f (1)], we have f (c) = f (1 ∧
c) = f (1) ∧ c = c and thus f satisfies (6). The dual statement follows
similarly. The last claim follows immediately from (6).
(cid:3)
Lemma 6. Let f : L → L be an order-preserving function. If f satisfies (4)
and its dual, then it preserves ∧ and ∨.
Proof. If f : L → L satisfies (4) and its dual, by Lemma 5, it satisfies (6).
Then, for every x ∈ L, we have
f (x) = med(f (0), f (x), f (1)) = f (med(f (0), x, f (1)))
= med(f (0), x, f (1)).
It is then immediate to see that f preserves ∧ and ∨.
Lemma 7. Let f : Ln → L be an order-preserving function. If f satisfies
(4) and its dual, then it satisfies (5) and its dual. Moreover, f satisfies (1)
and its dual.
Proof. Suppose that f satisfies (4) and its dual. For any x ∈ Ln and any
c ∈ [f (0), f (1)], we have
(cid:3)
f (x ∧ c) ∨ f ([x]c) = (cid:0)f (x) ∧ c(cid:1) ∨ f ([x]c)
= (cid:0)f (x) ∨ f ([x]c)(cid:1) ∧ (cid:0)c ∨ f ([x]c)(cid:1)
= f (x) ∧ f (c ∨ [x]c) = f (x)
and hence, f satisfies (5). The dual statement follows similarly.
Claim 1. For every k ∈ [n] and a ∈ Ln, the unary function f a
and its dual.
Proof of Claim 1. Let k ∈ [n], a ∈ Ln, and consider c ∈ [f a
(5), (4), and the dual of (4), for x ∈ L, we have
−k ([x ∨ c]c) = (cid:0)f a
−k (x ∨ c)
−k(x ∨ c) = f a∧c
f a
= c ∨ f
−k ([x ∨ c]c) = f a∨c
−k(1) ∧ c(cid:1) ∨ f
−k (c) ∨ f
[a]c
[a]c
−k satisfies (4)
−k(0), f a
−k(1)]. By
[a]c
−k ([x ∨ c]c)
= f a
−k(x) ∨ c.
POLYNOMIAL FUNCTIONS OVER BOUNDED DISTRIBUTIVE LATTICES
5
Therefore f a
satisfies (4).
−k satisfies the dual of (4). Similarly, we can prove that it also
(cid:3)
By Claim 1, each function f a
preserves ∧ and ∨.
−k satisfies (4) and its dual. By Lemma 6, it
(cid:3)
As mentioned, the equivalence between (i) and (ii) in the Main Theorem
was shown in [14, Theorem 17]. For the sake of self-containment, we provide
a simpler proof here.
Proposition 8 ([14]). A function f : Ln → L is a polynomial function if
and only if it satisfies (2) for every x ∈ Ln and k ∈ [n].
Proof. On the one hand, if f : Ln → L is a polynomial function, then, for
every x ∈ Ln and k ∈ [n], we have
f (x0
k) = _
I⊆[n]
I6∋k
(cid:16)αf (I) ∧ ^
i∈I
xi(cid:17)
and
f (x1
k) = _
I⊆[n]
I6∋k
(cid:16)αf (I) ∧ ^
i∈I
xi(cid:17) ∨ _
I⊆[n]
I∋k
(cid:16)αf (I) ∧ ^
i∈I\{k}
xi(cid:17)
= f (x0
k) ∨ _
I⊆[n]
I∋k
(cid:16)αf (I) ∧ ^
i∈I\{k}
xi(cid:17)
and hence
med(cid:0)f (x0
k), xk, f (x1
k)(cid:1) = f (x0
k) ∨ (cid:0)xk ∧ f (x1
k)(cid:1) = f (x),
which shows that f satisfies (2) for every x ∈ Ln and k ∈ [n].
On the other hand, any function obtained from a function in this class by
substituting constants for variables, is also in the class. Thus, if a function
f : Ln → L satisfies (2) for every x ∈ Ln and k ∈ [n], then by repeated
applications of (2), we can easily verify that f can be obtained by compo-
sition of the binary operations ∨ and ∧, the projections, and the constant
functions. That is, f is a polynomial function.
(cid:3)
Now we focus on the conditions given in (iii), (v), and (vi) of the Main
Theorem. We shall make use of the following general result.
Lemma 9. Let C be a class of functions f : Ln → L (n > 1) such that
(i) the unary members of C are polynomial functions;
(ii) for n > 1, any unary function obtained from an n-ary function f in
C by substituting constants for n − 1 variables of f is also in C.
Then C is a class of polynomial functions.
6
MIGUEL COUCEIRO AND JEAN-LUC MARICHAL
Proof. Let C be a class of functions satisfying the conditions of the lemma.
We show that each f : Ln → L in C is a polynomial function. By condition
(i), the claim holds for n = 1. So suppose that n > 1. By Proposition 8, it
is enough to show that f satisfies (2) for every x ∈ Ln and k ∈ [n]. So let
a ∈ Ln and k ∈ [n]. By condition (ii), we have that f a
−k ∈ C, and hence f a
−k
−k satisfies (ii) of the Main
is a polynomial function. By Proposition 8, f a
Theorem, and hence,
f (a) = f a
−k(ak) = med(cid:0)f a
−k(0), ak, f a
−k(1)(cid:1) = med(cid:0)f (a0
k), ak, f (a1
k)(cid:1).
Since the above holds for every a ∈ Ln and k ∈ [n], it follows that f satisfies
(ii) of the Main Theorem, and thus it is a polynomial function.
(cid:3)
Note that, for n = 1, (3) reduces to the well-known idempotency equation
f ◦ f = f ; see for instance Kuczma et al. [12, §11.9E].
Proposition 10. A unary function f : L → L is a polynomial function if
and only if R(f ) is convex and f is a solution of the idempotency equation
that preserves ∧ and ∨.
Proof. By Proposition 8, every unary polynomial function f : L → L is of
the form f (x) = med(f (0), x, f (1)) and thus satisfies the conditions stated
in the proposition.
Conversely, let f : L → L be a solution of the idempotency equation
that preserves ∧ and ∨ and such that R(f ) is convex, and let x ∈ L. If
x ∈ [f (0), f (1)] = R(f ), then there is z ∈ L such that x = f (z) and hence
f (x) = f (f (z)) = f (z) = x = med(f (0), x, f (1)).
Otherwise, let z = med(f (0), x, f (1)) ∈ [f (0), f (1)]. Then, since f preserves
∧ and ∨, it is order-preserving, and we have
f (x) = med(cid:0)f (0), f (x), f (1)(cid:1) = med(cid:0)f (f (0)), f (x), f (f (1))(cid:1)
= f(cid:0) med(f (0), x, f (1))(cid:1) = f (z) = z = med(f (0), x, f (1)),
which shows that f is a polynomial function.
(cid:3)
Proposition 11. An order-preserving function f : Ln → L is a polynomial
function if and only if f satisfies (1), its dual, (3), and the sets R(f ) and
R(f a
−k), for every a ∈ Ln and k ∈ [n], are convex.
Proof. By Fact 1 and Lemmas 4 and 5, if f is a polynomial function, then f
satisfies (1), its dual, and R(f ) is convex. By Proposition 8, every polyno-
−k), for a ∈ Ln
mial function f satisfies (2) (and hence (3)) and each set R(f a
and k ∈ [n], is convex.
To prove the converse claim, consider the class C of order-preserving
functions f : Ln → L (n > 1) satisfying the conditions of the proposition.
Clearly, C satisfies condition (ii) of Lemma 9. By Proposition 10, C sat-
isfies also condition (i) of Lemma 9, and hence, C is a class of polynomial
functions.
(cid:3)
POLYNOMIAL FUNCTIONS OVER BOUNDED DISTRIBUTIVE LATTICES
7
Proposition 12. Let f : L → L be a function. The following conditions are
equivalent:
(i) f is a polynomial function;
(ii) f satisfies (4) and preserves ∨;
(iii) f satisfies (6) and preserves ∧ and ∨.
Proof. The implication (i) ⇒ (iii) follows from Lemmas 4, 5, and 6. The
implication (iii) ⇒ (ii) follows from the fact that if f satisfies (6) and
preserves ∧ then it satisfies (4). Finally, to see that the implication (ii) ⇒ (i)
holds, observe first that (4) implies (6) by Lemma 5. Since f preserves ∨, we
have that f satisfies the dual of (4). Moreover, f is clearly order-preserving,
and we have
f (x) = med(cid:0)f (0), f (x), f (1)(cid:1) = f (0) ∨ (cid:0)f (x) ∧ f (1)(cid:1)
= f(cid:0)f (0) ∨ (x ∧ f (1))(cid:1) = f(cid:0) med(f (0), x, f (1))(cid:1)
= med(f (0), x, f (1)),
which shows that f is a polynomial function.
(cid:3)
Proposition 13. An order-preserving function f : Ln → L is a polynomial
function if and only if f satisfies the dual of (1), and (4), and (5).
Proof. By Lemmas 4 and 7 it follows that the conditions are necessary.
To prove the converse claim, we make use of Lemma 9. Let C be the class
of order-preserving functions f : Ln → L (n > 1) satisfying the conditions
of the proposition. By Proposition 12, C satisfies condition (i) of Lemma 9.
To complete the proof, it is enough to show that (4) and (5) are preserved
under substituting constants for n − 1 variables, since then condition (ii) of
Lemma 9 will be also fulfilled. Thus, take f : Ln → L satisfying (4) and (5).
−k satisfies (4), just note that if
Let k ∈ [n] and a ∈ Ln. To see that f a
x ∈ L and c ∈ [f a
−k(0), f a
−k(1)] then,
−k(x ∧ c) = f a∧c
f a
= (f a
−k (x ∧ c) ∨ f
−k(x) ∧ c) ∨ f
−k(x) ∧ c.
= f a
[a]c
−k (0)
[a]c
−k (0)
(by (5))
(by (4))
To see that f a
have f (ax
−k satisfies (5), let x ∈ L and c ∈ [f a
k ∧ c) 6 f a
−k(x) = f (ax
f a
−k(x ∧ c) and f ([ax
k ∧ c) ∨ f ([ax
k]c) 6 f a
k]c) 6 f a
−k(x ∧ c) ∨ f a
−k(0), f a
−k(1)]. We clearly
−k([x]c). Hence, by (5), we get
−k([x]c) 6 f a
−k(x).
−k satisfies (5).
(cid:3)
Since the above holds for every x ∈ L, we have that f a
Proposition 14. An order-preserving function f : Ln → L is a polynomial
function if and only if f satisfies (1) and its dual, (5) and its dual, and (6).
Proof. By Lemmas 4, 5, and 7, it follows that the conditions are necessary.
To prove the converse claim, we make use of Lemma 9. Let C be the class
of order-preserving functions f : Ln → L (n > 1) satisfying the conditions
8
MIGUEL COUCEIRO AND JEAN-LUC MARICHAL
of the proposition. By Proposition 12, C satisfies condition (i) of Lemma 9.
To complete the proof, it is enough to show that (5), its dual, and (6)
are preserved under substituting constants for n − 1 variables, since then
condition (ii) of Lemma 9 will be also fulfilled. Thus, take f : Ln → L
satisfying (5), its dual, and (6).
Let k ∈ [n] and a ∈ Ln. To see that f a
−k satisfies (6), just note that if
c ∈ [f a
−k(0), f a
−k(1)] then, by (5) and its dual,
−k (0) 6 c 6 f a∨c
−k (c) ∨ f
[a]c
−k(c) = f a∧c
f a
−k (c) ∧ f
[a]c
−k (1) = f a
−k(c).
The proof that f a
of Proposition 13. The dual claim follows similarly.
−k satisfies (5) follows exactly the same steps as in the proof
(cid:3)
We can now provide a proof of the Main Theorem.
Proof of the Main Theorem. The equivalences (i) ⇔ (ii) ⇔ (iii) are
given by Propositions 8 and 11. The equivalence (i) ⇔ (iv) follows from
Lemmas 4 and 7, and Proposition 13. The equivalences (i) ⇔ (v) and
(i) ⇔ (vi) follow from Propositions 13 and 14, respectively.
(cid:3)
3. Concluding remarks
By the equivalence (i) ⇔ (iii), polynomial functions f : Ln → L with
f (0) = 0 and f (1) = 1 coincide exactly with those for which R(f ) = L.
These are referred to as discrete Sugeno integrals and were studied in [13]
where equivalence (i) ⇔ (iv) of the Main Theorem was established for this
particular case when L is an interval of the real line. Also, the implica-
tion (i) ⇒ (v) of the Main Theorem reduces to that established by Ben-
venuti, Mesiar, and Vivona [1] when L is an interval of the real line, since
in this case the dual of (1) becomes redundant. Condition (4) and its dual,
when strengthened to all c ∈ L, are referred to as ∧-homogeneity and ∨-
homogeneity, respectively; see [9]. These were used by Fodor and Roubens [7]
to axiomatize certain classes of aggregation functions over the reals.
Recall that the property of being order-preserving is a consequence of
all except assertion (iv) of the Main Theorem. Also, given the nature of
statements (iii) -- (vi), it is natural to ask whether the equivalences between
these and (i) continue to hold over non-distributive lattices. The reader can
easily verify that (already for unary polynomial functions) this is not the
case.
References
[1] P. Benvenuti, R. Mesiar, and D. Vivona. Monotone set functions-based integrals. In
Handbook of measure theory, Vol. I, II, pages 1329 -- 1379. North-Holland, Amsterdam,
2002.
[2] S. Burris and H. P. Sankappanavar. A Course in Universal Algebra, volume 78 of
Graduate Texts in Mathematics. Springer-Verlag, New York, 1981.
[3] I. Chajda. Albebraic Theory of Tolerance Relations. ISBN 80-7067-042-8, Olomouc,
1991.
POLYNOMIAL FUNCTIONS OVER BOUNDED DISTRIBUTIVE LATTICES
9
[4] G. Cz´edli and L. Klukovits. A note on tolerances of idempotent algebras. Glasnik
Matematicki, 18(38):35 -- 38, 1983.
[5] K. Denecke and S.L. Wismath. Hyperidentities and clones, volume 14 of Algebra,
Logic and Applications. Gordon and Breach Science Publishers, Amsterdam, 2000.
[6] K. Denecke and S.L. Wismath. Universal algebra and coalgebra. World Scientific Pub-
lishing, Hackensack, NJ, 2009.
[7] J. C. Fodor and M. Roubens. Characterization of weighted maximum and some re-
lated operations. Inform. Sci., 84(3-4):173 -- 180, 1995.
[8] R. L. Goodstein. The solution of equations in a lattice. Proc. Roy. Soc. Edinburgh
Sect. A, 67:231 -- 242, 1965/1967.
[9] M. Grabisch, J.-L. Marichal, R. Mesiar, and E. Pap. Aggregation functions. Encyclo-
pedia of Mathematics and its Applications. Cambridge University Press, Cambridge,
UK, 2009.
[10] G. Gratzer. General Lattice Theory. Birkhauser Verlag, Berlin, 2003. Second edition.
[11] M. Kindermann. Uber die Aquivalenz von Ordnungspolynomvollstandigkeit und Tol-
eranzeinfachheit endlicher Verbande. Contributions to General Algebra (Proceedings
of the Klagenfurt conference 1978), 145 -- 149, 1979.
[12] M. Kuczma, B. Choczewski, and R. Ger. Iterative Functional Equations. Cambridge
University Press, Cambridge, UK, 1990.
[13] J.-L. Marichal. On Sugeno integral as an aggregation function. Fuzzy Sets and Sys-
tems, 114(3):347 -- 365, 2000.
[14] J.-L. Marichal. Weighted lattice polynomials. Discrete Mathematics, 309(4):814 -- 820,
2009.
Mathematics Research Unit, FSTC, University of Luxembourg, 6, rue Coudenhove-
Kalergi, L-1359 Luxembourg-Kirchberg, Luxembourg.
E-mail address: miguel.couceiro[at]uni.lu
Mathematics Research Unit, FSTC, University of Luxembourg, 6, rue Coudenhove-
Kalergi, L-1359 Luxembourg-Kirchberg, Luxembourg.
E-mail address: jean-luc.marichal[at]uni.lu
|
0901.1395 | 3 | 0901 | 2018-05-01T10:58:54 | Invariants of Lie algebras extended over commutative algebras without unit | [
"math.RA",
"math.KT"
] | We establish results about the second cohomology with coefficients in the trivial module, symmetric invariant bilinear forms and derivations of a Lie algebra extended over a commutative associative algebra without unit. These results provide a simple unified approach to a number of questions treated earlier in completely separated ways: periodization of semisimple Lie algebras (Anna Larsson), derivation algebras, with prescribed semisimple part, of nilpotent Lie algebras (Benoist), and presentations of affine Kac-Moody algebras. | math.RA | math |
INVARIANTS OF LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS
WITHOUT UNIT
PASHA ZUSMANOVICH
ABSTRACT. We establish results about the second cohomology with coefficients in the trivial module,
symmetric invariant bilinear forms, and derivations of a Lie algebra extended over a commutative associa-
tive algebra without unit. These results provide a simple unified approach to a number of questions treated
earlier in completely separated ways: periodization of semisimple Lie algebras (Anna Larsson), derivation
algebras, with prescribed semisimple part, of nilpotent Lie algebras (Benoist), and presentations of affine
Kac-Moody algebras.
INTRODUCTION
In this paper we consider current Lie algebras, i.e., Lie algebras of the form L ⊗ A, where L is a Lie
algebra, A is a commutative associative algebra, and the multiplication in L ⊗ A being defined by the
formula
[x ⊗ a, y ⊗ b] = [x, y] ⊗ ab
for any x, y ∈ L, a, b ∈ A. Note that A can be considered as the algebra of functions on the spectrum of A,
and L ⊗ A can therefore be interpreted as the algebra of "currents", as physicists say, on this spectrum.
Berezin and Karpelevich [BK] were among the first to study certain class of representations of current
algebras over local finite dimensional commutative algebras A. In that, somewhat obscure, paper they
showed that cohomology of such algebras can be reduced to cohomology of L. In what follows we will
consider other types of commutative algebras A and the description of cohomology in our case is more
involved and interesting. We are primarily interested in the second cohomology of L ⊗ A with trivial
coefficients, the space of symmetric invariant bilinear forms on L ⊗ A, and the algebras of derivations of
L ⊗ A.
These invariants were determined for numerous particular cases of current Lie algebras (see, for ex-
ample, [S]), the general formulae for the second homology with trivial coefficients in terms of invariants
of L and A were obtained in [Ha], [Z1] and [NW], and similar formulae for the space of symmetric
invariant bilinear forms and derivation algebras were obtained in [Z1] and [Z2], respectively.
So why return to these settled questions? In all considerations until now, the algebra A was supposed
to have a unit. However, there are many interesting examples of current algebras where A is not unital.
For example, in [La1], the so-called periodization of semisimple Lie algebras g was considered, which
is nothing but g ⊗ t C[t]. It is known that the second homology of any nilpotent Lie algebra with trivial
coefficients has interpretation in terms of presentation of the algebra, so allowing A to be nilpotent
allows us to obtain presentation of L ⊗ A irrespective of the properties of L.
It turns out that elementary arguments similar to those in [Z2] allow us to extend the above mentioned
results to the case of non-unital A. In particular, concerning the second cohomology and symmetric
invariant bilinear forms, we provide another proof, considerably shorter than all the previous ones even
in the case of unital A.
The contents of this paper are as follows. In §§1 -- 3 we establish the general formulae for 2-cocycles,
symmetric invariant bilinear forms, and get partial results about derivations of the current Lie algebras,
Date: last revised December 26, 2017.
2000 Mathematics Subject Classification. 17B20, 17B40, 17B55, 17B56, 17B67.
Key words and phrases. Current Lie algebra; Kac-Moody algebras; second cohomology;
invariant bilinear form;
derivation.
arXiv:0901.1395; J. Nonlin. Math. Phys. 17 (2010), Suppl. 1 (special issue in memory of F.A. Berezin), 87 -- 102.
1
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
2
respectively. This is followed by applications: in §4 we reprove the result from [La1] about presentations
of periodizations of the semisimple Lie algebras. In passing, we also mention how to derive from our
results the theorem from [Be] about semisimple components of the derivation algebras of certain current
Lie algebras, and the Serre defining relations between Chevalley generators of the non-twisted affine
Kac-Moody algebra.
In all these cases, the absence of unit in A is essential. All these proofs are significantly shorter than
the original ones, and reveal various almost trivial, but so far unnoticed or unpublished, links between
different concepts and results. These links are, perhaps, the main virtue of this paper.
It seems that most if not everything considered here can be extended in a straightforward way to
twisted, Leibniz and super settings, but we will not venture into this, at least for now.
NOTATION AND CONVENTIONS
All algebras and vector spaces are defined over an arbitrary field K of characteristic different from 2
and 3, unless stated otherwise (some of the results are valid in characteristic 3, but we will not go into
this).
In what follows, L denotes a Lie algebra, A is an associative commutative algebra.
Given an L-module M, let Bn(L, M), Zn(L, M), Cn(L, M) and H n(L, M) denote the space of nth degree
coboundaries, cocycles, cochains, and cohomology of L with coefficients in M, respectively (we will
be mainly interested in the particular cases of degree 2 and the trivial module K, or degree 1 and the
adjoint module or its dual). Note that C2(L, K) is the space of all skew-symmetric bilinear forms on L.
The space of all symmetric bilinear forms on L will be denoted as S2(L, K).
Let Z(L), [L, L] and Der(L) denote the center, the commutant (the derived algebra), and the Lie
algebra of derivations of L, respectively. Similarly, let Ann(A) = {a ∈ A Aa = 0} and AA denote the
annulator and the square of A, respectively, and let HC∗(A) denotes its cyclic cohomology.
A bilinear form ϕ : L × L → K is said to be cyclic if
ϕ([x, y], z) = ϕ([z, x], y)
for any x, y, z ∈ L. Note that if ϕ is symmetric, this condition is equivalent to the invariance of the form
ϕ:
ϕ([x, y], z) + ϕ(y, [x, z]) = 0,
while if ϕ is skew-symmetric, the notions of cyclic and invariant forms differ. Let B(L) denote the space
of all symmetric bilinear invariant (=cyclic) forms on L.
Similarly, a bilinear form α : A × A → K is said to be cyclic if
and invariant if
α(ab, c) = α(ca, b),
α(ab, c) = α(a, bc)
for any a, b, c ∈ A. If the form α is symmetric, it is cyclic if and only if it is invariant.
1. THE SECOND COHOMOLOGY
Theorem 1. Let L be a Lie algebra, A an associative commutative algebra, and at least one of L and A
be finite-dimensional. Then each cocycle in Z2(L⊗A, K) can be represented as the sum of decomposable
cocycles ϕ⊗ α, where ϕ : L × L → K and α : A × A → K are of one of the following 8 types:
(i) ϕ([x, y], z) + ϕ([z, x], y) + ϕ([y, z], x) = 0 and α is cyclic,
(ii) ϕ is cyclic and α(ab, c) + α(ca, b) + α(bc, a) = 0,
(iii) ϕ([L, L], L) = 0,
(iv) α(AA, A) = 0,
where each of these 4 types splits into two subtypes: with ϕ skew-symmetric and α symmetric, and with
ϕ symmetric and α skew-symmetric.
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
3
Proof. Each cocycle Φ ∈ Z2(L ⊗ A, K), being an element of
End(L ⊗ A ⊗ L ⊗ A, K) ≃ End(L ⊗ L, K) ⊗ End(A ⊗ A, K),
can be written in the form Φ = ∑
i∈I
ϕi ⊗ αi, where ϕi : L × L → K and αi : A × A → K are bilinear maps,
and I is a finite set of indices (this is the place where the assumption of finite-dimensionality is needed).
Using this representation, write the cocycle equation for an arbitrary triple x ⊗ a, y ⊗ b, z ⊗ c, where
x, y, z ∈ L, a, b, c ∈ A:
(1)
∑
i∈I
ϕi([x, y], z) ⊗ αi(ab, c) + ϕi([z, x], y) ⊗ αi(ca, b) + ϕi([y, z], x) ⊗ αi(bc, a) = 0.
Symmetrizing this equality with respect to x, y, we get:
∑
i∈I(cid:16)ϕi([x, z], y) + ϕi([y, z], x)(cid:17) ⊗(cid:16)αi(bc, a) − αi(ca, b)(cid:17) = 0.
On the other hand, cyclically permuting x, y, z in (1) and summing up the 3 equalities obtained, we get:
∑
i∈I(cid:16)ϕi([x, y], z) + ϕi([z, x], y) + ϕi([y, z], x)(cid:17) ⊗(cid:16)αi(ab, c) + αi(bc, a) + αi(ca, b)(cid:17) = 0.
Applying Lemma 1.1 from [Z2] to the last two equalities, we get a partition of the index set I = I1 ∪ I2 ∪
I3 ∪ I4 such that
ϕi([x, z], y) + ϕi([y, z], x) = 0, ϕi([x, y], z) + ϕi([z, x], y) + ϕi([y, z], x) = 0
ϕi([x, z], y) + ϕi([y, z], x) = 0, αi(ab, c) + αi(bc, a) + αi(ca, b) = 0
ϕi([x, y], z) + ϕi([z, x], y) + ϕi([y, z], x) = 0, αi(bc, a) − αi(ca, b) = 0
αi(bc, a) − αi(ca, b) = 0, αi(ab, c) + αi(bc, a) + αi(ca, b) = 0
for i ∈ I1
for i ∈ I2
for i ∈ I3
for i ∈ I4.
It is obvious that if the characteristic of K is different from 3, then ϕi([L, L], L) = 0 for i ∈ I1, and
αi(AA, A) = 0 for i ∈ I4.
It is obvious also that ϕi ⊗ αi satisfies the cocycle equation (1) for each
i ∈ I1, I2, I3, I4.
Now write the condition of skew-symmetry of Φ:
(2)
ϕi(x, y) ⊗ αi(a, b) + ϕi(y, x) ⊗ αi(b, a) = 0
∑
i∈I
and symmetrize it with respect to x, y:
(ϕi(x, y) − ϕi(y, x)) ⊗ (αi(a, b) − αi(b, a)) = 0
(ϕi(x, y) + ϕi(y, x)) ⊗ (αi(a, b) + αi(b, a)) = 0.
∑
i∈I
∑
i∈I
From the last two equalities, using again Lemma 1.1 from [Z2], we see that each set I1, I2, I3, I4 can be
split further into two subsets, one having skew-symmetric ϕi and symmetric αi, and the other one having
symmetric ϕi and skew-symmetric αi.
(cid:3)
Remark 1. As all our bilinear maps are K-valued, the cocycles of the form ϕ ⊗ α are, of course, just
products of bilinear maps ϕα. However, we have retained the symbol ⊗, to make it easier to track
dependence on the more general situation of [Z2].
(3)
Remark 2. The second subdivision in the statement of Theorem 1, which follows from equality (2), is
merely a manifestation of the vector space isomorphism
C2(L ⊗ A, K) ≃(cid:0)S2(L, K) ⊗C2(A, K)(cid:1) ⊕(cid:0)C2(L, K) ⊗ S2(A, K)(cid:1) .
Let dΩ be the 2-coboundary defined by a given linear map Ω : L ⊗ A → K. The latter can be written in
the form Ω = ∑
i∈I
ωi ⊗ βi for some linear maps ωi : L → K and βi : A → K. Then
dΩ(x ⊗ a, y ⊗ b) = ∑
i∈I
ωi([x, y]) ⊗ βi(ab),
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
4
i.e., coboundaries always lie in the direct summand C2(L, K) ⊗ S2(A, K). Consequently, nonzero cocy-
cles from different direct summands in (3) can never be cohomologically dependent, and the cocycles
from the direct summand S2(L, K) ⊗ C2(A, K) are cohomologically independent if and only if they are
linearly independent.
One may try to formulate Theorem 1 as a statement about H 2(L ⊗ A, K), but in full generality this will
lead only to cumbersome complications. In each case of interest, one can easily obtain an information
about the cohomology. For example, assuming A contains a unit, one immediately sees that cocycles of
type (i) necessarily have ϕ skew-symmetric and α symmetric, cocycles of type (ii) necessarily have ϕ
symmetric and α skew-symmetric, and cocycles of type (iv) vanish. This leads to a known formula for
H 2(L ⊗ A, K), where the cocycles of type (i) contribute to the term H 2(L, K) ⊗ A∗, the cocycles of type
(ii) contribute to the term B(L)⊗HC1(A), and the cocycles of type (iii) are non-essential (in terminology
of [Z1]).
Another, more concrete, application is given in §4.
2. SYMMETRIC INVARIANT BILINEAR FORMS
Theorem 2. Let L be a Lie algebra, A an associative commutative algebra, and at least one of L and
A be finite-dimensional. Then each symmetric invariant bilinear form on L ⊗ A can be represented as a
sum of decomposable forms ϕ⊗ α, ϕ : L × L → K, α : A × A → K of one of the following 6 types:
(i) both ϕ and α are cyclic,
(ii) ϕ([L, L], L) = 0,
(iii) α(AA, A) = 0,
where each of these 3 types splits into two subtypes: with both ϕ and α symmetric, and with both ϕ and
α skew-symmetric.
Proof. The proof is absolutely similar to that of Theorem 1. As in the proof of Theorem 1, we may write
a symmetric invariant bilinear form Φ on L ⊗ A as ∑
ϕi ⊗ αi for suitable bilinear maps ϕi : L × L → K
i∈I
and αi : A × A → K. The invariance condition, written for a given triple x ⊗ a, y ⊗ b, z ⊗ c, reads:
(4)
∑
i∈I
ϕi([x, y], z) ⊗ αi(ab, c) + ϕi([x, z], y) ⊗ αi(ca, b) = 0.
Symmetrizing this with respect to x, y, we get:
∑
i∈I(cid:16)ϕi([x, z], y) + ϕi([y, z], x)(cid:17) ⊗ αi(ca, b) = 0.
Hence the index set can be partitioned I = I1 ∪ I2 in such a way that
ϕi([x, z], y) + ϕi([y, z], x) = 0
for any i ∈ I1, and αi(AA, A) = 0 for any i ∈ I2. Then (4) can be rewritten as
∑
i∈I1
ϕi([x, y], z) ⊗(cid:16)αi(ab, c) − αi(ca, b)(cid:17) = 0.
Hence there is a partition I1 = I11 ∪ I12 such that ϕi([L, L], L) = 0 for any i ∈ I11, and αi(ab, c) = αi(ac, b)
for any i ∈ I12.
The condition of symmetry of Φ:
ϕi(x, y) ⊗ αi(a, b) − ϕi(y, x) ⊗ αi(b, a) = 0,
∑
i∈I
being symmetrized with respect to x, y, allows us to partition further each of the sets I11, I12, I2 into
two subsets, one having both ϕi and αi symmetric, and the other one having both ϕi and αi skew-
symmetric.
(cid:3)
This generalizes [Z1, Theorem 4.1], where a similar statement is proved for A unital.
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
5
3. DERIVATIONS
Naturally, one may try to apply the same approach to description of the derivations of a given current
algebra L ⊗ A (for unital A, see [Z2, Corollary 2.2]). Indeed, each derivation D of L ⊗ A, being an
element of
End(L ⊗ A, L ⊗ A) ≃ End(L, L) ⊗ End(A, A),
can be expressed in the form D = ∑
i∈I
ϕi ⊗ αi, where ϕi : L → L and αi : A → A are linear maps. The
condition that D is a derivation, written for an arbitrary pair x ⊗ a and y ⊗ b, where x, y ∈ L and a, b ∈ A,
reads:
ϕi([x, y]) ⊗ αi(ab) − [ϕi(x), y] ⊗ αi(a)b − [x,ϕi(y)] ⊗ aαi(b) = 0.
∑
i∈I
Symmetrizing this equality with respect to a, b (this is equivalent to symmetrization with respect to x,
y):
∑
i∈I(cid:16)[ϕi(x), y] − [x,ϕi(y)](cid:17) ⊗(cid:16)aαi(b) − bαi(a)(cid:17) = 0,
we get a partition of the index set I into two subsets with conditions [ϕi(x), y] = [x,ϕi(y)] and aαi(b) =
bαi(a), respectively. But as there are only two variables in each of L and A, no other symmetrization is
possible, so the last equality is all what we can get in this way.
The failure of this method can be also explained by looking at the simple example of the Lie algebra
sl(2) ⊗ tK[t]. In sl(2), fix a basis {e−, h, e+} with multiplication
(5)
[h, e−] = −e−,
[h, e+] = e+,
[e−, e+] = h.
It is easy to check that the map defined, for any f (t) ∈ tK[t], by the formulae
e− ⊗ f (t) 7→ e− ⊗ t(cid:16) d f (t)
e+ ⊗ f (t) 7→ e+ ⊗ t(cid:16) d f (t)
dt
d f (t)
dt
h ⊗ f (t) 7→ h ⊗ t
dt
− f (t)(cid:17)
+ f (t)(cid:17)
is a derivation of sl(2) ⊗ tK[t]. It is obvious that this map is not a decomposable one, i.e., of the form
ϕ ⊗ α for some ϕ : sl(2) → sl(2) and α : tK[t] → tK[t]. But for this approach to succeed, all the maps
in question should be representable in such a way in the end.
However, under additional assumption on L ⊗ A, we can derive information about Der(L ⊗ A) from
the results of the preceding sections, using the relationship between H 1(L, L∗), H 2(L, K), and B(L). In
the literature, this relationship was noted many times in a slightly different form, and goes back to the
classical works of Koszul and Hochschild -- Serre [HS]. Namely, there is an exact sequence
(6)
0 → H 2(L, K) u→ H 1(L, L∗) v→ B(L) w→ H 3(L, K)
where for the representative ϕ ∈ Z2(L, K) of a given cohomology class, we have to take the class of
u(ϕ), the latter being given by
(u(ϕ)(x))(y) = ϕ(x, y)
for any x, y ∈ L, v is sending the class of a given cocycle D ∈ Z1(L, L∗) to the bilinear form v(D) :
L × L → K defined by the formula
v(D)(x, y) = D(x)(y) + D(y)(x),
and w is sending a given symmetric bilinear invariant form ψ : L × L → K to the class of the cocycle
ω ∈ Z3(L, K) defined by
ω(x, y, z) = ψ([x, y], z)
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
6
(see, for example, [D1], where a certain long exact sequence is obtained, of which this one is the
beginning, and references therein for many earlier particular versions; this exact sequence was also
established in [NW, Proposition 7.2] with two additional terms on the right).
In the case where L ≃ L∗ as L-modules, the sequence (6) provides a way to evaluate H 1(L, L) given
H 2(L, K) and B(L). The L-module isomorphism L ≃ L∗ implies the existence of a symmetric invariant
non-degenerate form h·, ·i on L. In terms of this form, u is sending the class of a given cocycle ϕ ∈
Z2(L, K) to the class of the cocycle u(ϕ) ∈ Z1(L, L) defined by
h(u(ϕ))(x), yi = ϕ(x, y),
and v is sending the class of a given cocycle D ∈ Z1(L, L) to the bilinear form v(D) : L × L → K defined
by the formula
v(D)(x, y) = hD(x), yi + hx, D(y)i.
Turning to current Lie algebras, we will make even stronger assumption: that L ≃ L∗ and A ≃ A∗.
Then, utilizing the results of preceding sections about H 2(L ⊗ A, K) and B(L ⊗ A), we will derive results
about Der(L ⊗ A).
In the literature, given H 1(L, L∗), the space H 2(L, K) was computed for various Lie algebras L (see,
for example, [S], [D1] and references therein). Here we utilize this connection in the other direction.
Theorem 3. Let L be a nonabelian Lie algebra, A an associative commutative algebra, both L and A
finite-dimensional and with symmetric invariant non-degenerate bilinear form. Then each derivation of
L ⊗ A can be represented as the sum of decomposable linear maps d ⊗β, where d : L → L and β : A → A
are of one of the following types:
(i) d([x, y]) = λ([d(x), y] + [x, d(y)]), β(ab) = µβ(a)b for certain λ, µ ∈ K such that λµ = 1,
(ii) d([x, y]) = λ[d(x), y], β(ab) = µ(β(a)b + aβ(b)) for certain λ, µ ∈ K such that λµ = 1,
(iii) [d(x), y] + [x, d(y)] = 0, β(AA) = 0, β(a)b = aβ(b),
(iv) d([L, L]) = 0, [d(x), y] + [x, d(y)] = 0, β(a)b = aβ(b),
(v) d([L, L]) = 0, [d(x), x] = 0, β(a)b + aβ(b) = 0,
(vi) [d(x), x] = 0, β(AA) = 0, β(a)b + aβ(b) = 0,
(vii) d([L, L]) = 0, d(L) ⊆ Z(L),
(viii) d([L, L]) = 0, β(A) ⊆ Ann(A),
(ix) d(L) ⊆ Z(L), β(AA) = 0,
(x) β(AA) = 0, β(A) ⊆ Ann(A).
Proof. By abuse of notation, let h· , ·i denote a symmetric invariant non-degenerate bilinear form both
on L and A. Obviously, the tensor product of these forms defines a symmetric invariant non-degenerate
bilinear form on L ⊗ A, for which by even bigger abuse of notation we will use the same symbol:
hx ⊗ a, y ⊗ bi = hx, yiha, bi.
We have L∗ ≃ L as L-modules, A∗ ≃ A as A-modules, and (L ⊗ A)∗ ≃ L ⊗ A as L ⊗ A-modules.
As a vector space, H 1(L ⊗ A, L ⊗ A) can be represented as the direct sum of Ker v and Im v, and the
exact sequence (6) tells that Ker v = Im u and Im v = Ker w.
By Theorem 1, H 2(L ⊗ A, K) is spanned by cohomology classes which can be represented by decom-
posable cocycles ϕ⊗ α for appropriate ϕ : L × L → L and α : A × A → K. For each such pair ϕ and α,
there are unique linear maps d : L → L and β : A → A such that
(7)
for any x, y ∈ L, and
(8)
hd(x), yi = ϕ(x, y)
hβ(a), bi = α(a, b)
for any a, b ∈ A. Hence the decomposable linear map d ⊗ β : L ⊗ A → L ⊗ A satisfies
h(d ⊗ β)(x ⊗ a), y ⊗ bi = (ϕ⊗ α)(x ⊗ a, y ⊗ b),
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
7
i.e., coincides with u(ϕ ⊗ α). Thus, Im u is spanned by cohomology classes whose representatives are
decomposable derivations.
Similarly, by the proof of Theorem 2, B(L ⊗ A) is spanned by decomposable elements ϕ ⊗ α, and
Ker w is spanned by such elements of types (ii) and (iii), i.e., either ϕ([L, L], L) = 0 or α(AA, A) = 0.
Again, for each such element we can find d : L → L and β : A → A satisfying (7) and (8) respectively.
Furthermore, we may assume that for each such ϕ⊗α, the maps ϕ and α are either both symmetric, or
both skew-symmetric, and hence both d and β are either self-adjoint or skew-self-adjoint, respectively,
with respect to h·, ·i. In both cases we have:
(9)
h(d ⊗ β)(x ⊗ a), y ⊗ bi + hx ⊗ a, (d ⊗ β)(y ⊗ b)i
= hd(x), yihβ(a), bi + hx, d(y)iha,β(b)i = 2hd(x), yihβ(a), bi
= 2ϕ(x, y) ⊗ α(a, b)
for any x, y ∈ L and a, b ∈ A.
The condition ϕ([L, L], L) = 0 ensures that d([L, L]) = 0, and an equivalent condition ϕ(L, [L, L]) = 0
ensures that
h[d(x), z], yi = −hd(x), [y, z]i = 0,
implying d(L) ⊆ Z(L), and hence d ⊗ β is a derivation of L ⊗ A. Quite analogously, the condition
α(AA, A) = α(A, AA) = 0 implies also that d ⊗β is a derivation of L ⊗ A. The equality (9) ensures that v
maps the cohomology class of this derivation to 2ϕ⊗α. Thus Im v is spanned by images of cohomology
classes whose representatives are decomposable derivations.
Putting all this together, we see that H 1(L ⊗ A, L ⊗ A) is spanned by the cohomology classes whose
representatives are decomposable derivations. As inner derivations of L ⊗ A are, obviously, also spanned
by decomposable inner derivations ad (x ⊗ a) = ad x ⊗ Ra, where Ra is the multiplication on a ∈ A, any
derivation of L ⊗ A is representable as the sum of decomposable derivations.
The rest is easy. The condition that d ⊗ β is a derivation, reads:
(10)
d([x, y]) ⊗ β(ab) − [d(x), y] ⊗ β(a)b − [x, d(y)] ⊗ aβ(b) = 0
for any x, y ∈ L, a, b ∈ A. Symmetrizing (10) as in the beginning of this section, we see that either
[d(x), y] = [x, d(y)] for any x, y ∈ L, or β(a)b = aβ(b) for any a, b ∈ A. The equation (10) is equivalent
to
d([x, y]) ⊗ β(ab) − [d(x), y] ⊗ (β(a)b + aβ(b)) = 0
in the first case, and to
d([x, y]) ⊗ β(ab) − ([d(x), y] + [x, d(y)]) ⊗ β(a)b = 0.
in the second case. Now trivial case-by-case considerations involving vanishing and linear dependence
of the linear operators occurring as tensor product factors in these two equalities, produce the final list
of derivations.
(cid:3)
Theorem 3 can be applied, for example, to the Lie algebra g ⊗ tK[t]/(t n), n > 2, where g is a se-
misimple finite-dimensional Lie algebra over any field of characteristic 0, to obtain a very short proof
of the result of Benoist [Be] about realization of any semisimple Lie algebra as semisimple part of
the Lie algebra of derivations of a nilpotent Lie algebra (another short proof with direct calculation of
Der(g ⊗ tK[t]/(t 3)) follows from [LL, Proposition 3.5]). Indeed, as noted, for example, in [BB, Lemma
2.2], tK[t]/(t n) possesses a symmetric nondegenerate invariant bilinear form B, hence g ⊗ tK[t]/(t n)
possesses such a form (being the product of the Killing form on g and B), so Theorem 3 is applicable.
As g is perfect and centerless, the derivations of types (iv), (v), (vii), (viii), (ix) vanish. The remaining
types can be handled, for example, by appealing to results of [Ho], [F] or [LL], which imply that in
the case g 6≃ sl(2), the corresponding mappings d vanish also for types (ii) and (iii), and for the rest of
the types are either inner derivations of g, or multiplications by scalar. Then, performing elementary
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
8
calculations with conditions imposed on β's in the remaining types, and rearranging the obtained spaces
of derivations, we get the following isomorphism of vector spaces:
Der(g ⊗ tK[t]/(t n)) ≃ (g ⊗ K[t]/(t n)) ⊕ (End(g)/g) ⊕ K.
(11)
Elements of the first summand are assembled from types (i) and (x), and act on g ⊗ tK[t]/(t n) as the
Lie multiplication by an element of g and the associative commutative multiplication by an element of
K[t]/(t n). Elements of the second summand are assembled from types (i), (vi) and (x), and act by the
rule
7→ F(x) ⊗ t n−1
x ⊗ t
x ⊗ t k 7→ 0
if k ≥ 2,
where x ∈ g, F ∈ End(g), and elements of g are assumed to be embedded into End(g) as inner deriva-
tions. Elements of the third, one-dimensional, summand are assembled from types (i) and (vi), and are
proportional to the following derivation:
7→ x ⊗ t n−2
x ⊗ t
x ⊗ t 2 7→ 2x ⊗ t n−1
x ⊗ t k 7→ 0
if k ≥ 3,
where x ∈ g.
All this implies that, as a Lie algebra, Der(g ⊗ tK[t]/(t n)) splits into the semidirect sum of the semi-
simple part isomorphic to g (identified with the part g⊗ 1 of the first summand in (11)), and the nilpotent
radical consisting of g ⊗ tK[t]/(t n) from the first summand, and the whole second and third summands.
The case g = sl(2) can be treated separately and easily.
4. PERIODIZATION OF SEMISIMPLE LIE ALGEBRAS
For a given Lie algebra L, its periodization is an N-graded Lie algebra, with component in each degree
isomorphic to L. In other words, the periodization of L is L ⊗ tK[t].
In [La1], Anna Larsson studied periodization of semisimple finite-dimensional Lie algebras g over
any field K of characteristic 0. She proved that, unless g contains direct summands isomorphic to sl(2),
its periodization possesses a presentation with only quadratic relations. Since generators and relations
of (generalized graded) nilpotent Lie algebras can be interpreted as the first and second homology,
Larsson's statement can be formulated in homological terms.
Note that in [La2] this result was generalized to some classes of Lie superalgebras. We do not touch
upon superalgebras here, and it seems to be an interesting task to tackle the results and questions from
[La2] from this paper's viewpoint. Interest in periodizations of Lie superalgebras, and whether they
admit generators subject to quadratic relations, arose from an earlier work of Lofwall and Roos [LR]
about some amazing Hopf algebras (for further details, see [La1] and [La2]).
As noted in [La1], the whole space H∗(g⊗t C[t], C) was studied by much more sophisticated methods
in the celebrated paper by Garland and Lepowsky [GL] (actually, a particular case interesting for us
here was already sketched in [G]; the case of g = sl(2) was also treated in [FF, p. 233]). Garland
and Lepowsky determined the eigenvectors of the Laplacian on the corresponding Chevalley-Eilenberg
(co)chain complex. However, to extract from [GL] exact results about (co)homology of interest requires
nontrivial combinatorics with the Weyl group, as demonstrated in [HW], and case-by-case analysis for
each series of the simple Lie algebras. Here we derive results for the second cohomology in a uniform
way from the results of §1 using elementary methods providing an alternative short proof of Larsson's
result. Also, our approach clearly shows why the case of sl(2) is exceptional.
Theorem 4 (Larsson). Let g be a finite-dimensional semisimple Lie algebra over an algebraically closed
field K of characteristic 0. Then
H 2(g ⊗ tK[t], K) ≃ C2(g, K)/B2(g, K)M(cid:16)M Sα(cid:17) ,
(12)
(13)
Φ(x ⊗ t i, y ⊗ t j) =(ϕ(x, y)
Ψ(x ⊗ t i, y ⊗ t j) =
ψ(x, y)
−ψ(x, y)
0
if i = j = 1,
otherwise,
0
if i = 1, j = 2,
if i = 2, j = 1,
otherwise,
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
9
where the second direct sum is taken over all simple direct summands of g isomorphic to sl(2), and each
Sα is a certain 5-dimensional space of symmetric bilinear forms on the corresponding direct summand.
The basic cocycles can be chosen among cocycles of the form
where x, y ∈ g and ϕ is a skew-symmetric bilinear form on g, and the cocycles Ψ whose only non-
vanishing values on all pairs of the simple direct summands of g are determined by the formula
where x, y belong to the corresponding sl(2)-direct summand, and ψ is a symmetric bilinear form on
this direct summand satisfying
(14)
in the standard sl(2)-basis (5).
ψ(e−, e+) =
1
2
ψ(h, h)
Proof. Consider the cocycles appearing in Theorem 1 for L = g and A = tK[t] case-by-case.
Cocycles of type (i). Writing the cyclicity condition of α for the triple t i−1, t j, t, we get
(15)
for any i ≥ 2, j ≥ 1.
α(t i+ j−1, t) = α(t i, t j)
(ia). ϕ is skew-symmetric, thus ϕ ∈ Z2(g, K), and α is symmetric. Since H 2(g, K) = 0, it follows that
ϕ = dω for some linear map ω : g → K. Define a linear map Ω : g ⊗ tK[t] → K by setting, for any x ∈ g,
Ω(x ⊗ t i) =(0
ω(x)α(t i−1, t)
if i = 1,
if i ≥ 2.
Then, taking (15) into account, we get ϕ⊗ α = dΩ, i.e., cocycles of this type are trivial.
(ib). ϕ is symmetric, α is skew-symmetric.
If g = sl(2), direct calculation shows that the space of corresponding ϕ's coincides with the space
of all symmetric bilinear forms on sl(2) satisfying the condition (14), and hence is 5-dimensional (an
equivalent calculation is contained in [D2, Theorem 6.5]). This case is exceptional, as shows the fol-
lowing
Lemma 1 (Dzhumadil'daev -- Bakirova). Let g 6≃ sl(2) be simple. Then any symmetric bilinear form ϕ
on g satisfying
(16)
ϕ([x, y], z) + ϕ([z, x], y) + ϕ([y, z], x) = 0
for any x, y, z ∈ g vanishes.
This Lemma is proved in [DB] by considering the Chevalley basis of g and performing computations
with the corresponding root system. In [D2] and [DB], symmetric bilinear forms satisfying the condition
(16) are called commutative 2-cocycles and arise naturally in connection with classification of algebras
satisfying skew-symmetric identities. We will give a different proof which stresses the connection with
yet other notions and results.
Proof. Consider a map from the space of bilinear forms on a Lie algebra L to the space of linear maps
from L to L∗, by sending a bilinear form ϕ to the linear map D : L → L∗ such that D(x)(y) = ϕ(x, y). It
is easy to see that a symmetric bilinear form ϕ satisfying (16) maps to a linear map D satisfying
D([x, y]) = −y • D(x) + x • D(y),
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
10
where • denotes the standard L-action on the dual module L∗ (this is completely analogous to the em-
bedding of H 2(L, K) into H 1(L, L∗) mentioned in §3).
For any finite dimensional simple Lie algebra, there is an isomorphism of g-modules g ≃ g∗, and we
have an embedding of the space of bilinear forms in question into the space of linear maps D : g → g
satisfying the condition
D([x, y]) = −[D(x), y] − [x, D(y)].
Such maps are called antiderivations and were studied in several papers. In particular, in [Ho, Theorem
5.1] it is proved that central simple classical Lie algebras of dimension > 3 do not have nonzero an-
tiderivations (generalizations of this result in different directions obtained further in the series of papers
by Filippov, of which [F] is the latest, and in [LL]). Hence, these Lie algebras do not have nonzero
symmetric bilinear forms satisfying (16) either.
(cid:3)
Coupled with the well-known fact that H 2(g, K) = 0, Lemma 1 implies that any (symmetric, skew-
symmetric or mixed) bilinear form ϕ on g 6≃ sl(2) satisfying the cocycle equation (16), is a 2-coboundary†.
This can be also compared with the fact that Leibniz cohomology (and, in particular, the second Leibniz
cohomology with trivial coefficients) of g vanishes (for homological version, see [P, Proposition 2.1] or
[N]). The condition for a bilinear form ϕ to be a Leibniz 2-cocycle can be expressed as
ϕ([x, y], z) + ϕ([z, x], y) − ϕ(x, [y, z]) = 0.
In the general case, where g is a direct sum of simple ideals g1 ⊕ · · · ⊕ gn, it is easy to see that
ϕ(gi, g j) = 0 for i 6= j, and hence the space of symmetric bilinear forms on g satisfying (16) decomposes
into the direct sum of appropriate spaces on each of gi. The latter are determined by (14) if gi ≃ sl(2)
and vanishes otherwise.
Now, turning to α's, and permuting i and j in (15), we get α(t i, t j) = 0 for all i, j ≥ 2 and α(t k, t) = 0
for all k ≥ 3. Conversely, it is easy to see that a skew-symmetric map α satisfying these conditions is
cyclic. Hence, the space of skew-symmetric cyclic maps on tK[t] is 1-dimensional and each cocycle of
this type can be written in the form (13).
Cocycles of type (ii).
(iia). ϕ is skew-symmetric, α is symmetric.
Lemma 2. Any skew-symmetric cyclic bilinear form on g vanishes.
Proof. The proof is almost identical to the proof of the well-known fact that any skew-symmetric in-
variant bilinear form on g vanishes (see, for example, [Bo, Chapter 1, §6, Exercises 7(b) and 18(a,b)]).
Namely, let ϕ be a skew-symmetric cyclic form on g. First, let g be simple. There is a linear map
σ : g → g such that ϕ(x, y) = hσ(x), yi, where h·, ·i is the Killing form on g. Then, for any x, y, z ∈ g:
hσ([x, y]), zi = ϕ([x, y], z) = −ϕ(y, [z, x]) = −hσ(y), [z, x]i = −h[x,σ(y)], zi.
Hence σ anticommutes with each ad x and, in particular, [σ(x), x] = 0 for any x ∈ g. But then by [Be,
Lemme 2] (or by more general results from [LL]), σ belongs to the centroid of g. Hence σ is a scalar
and necessarily vanishes.
In the general case where g is the direct sum of simple ideals g1 ⊕ · · · ⊕ gn, it is easy to see that
ϕ(gi, g j) = 0 for any i 6= j. But by just proved ϕ vanishes also on each gi, and hence vanishes on the
whole of g.
(cid:3)
(iib). ϕ is symmetric and α is skew-symmetric, so α ∈ HC1(tK[t]). There is an isomorphism of
graded algebras
HC∗(K[t]) ≃ HC∗(K) ⊕ HC∗(tK[t])
† Added December 26, 2017: The statement is true, but it does not follow, at least in a straightforward way, from Lemma
1 and the fact that H 2(g, K) = 0. For a correct proof, see a forthcoming text (with A. Makhlouf) "Hom-Lie algebra structures
on Kac-Moody algebras".
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
11
(for a general relationship between cyclic homology of augmented algebra and its augmentation ideal,
see [LQ, §4]; the cohomological version can be obtained in exactly the same way). On the other hand,
HC∗(K[t]) ≃ HC∗(K) ⊕ (terms concentrated in degree 0)
(see, for example, [Lo, §3.1.7]). Hence HC1(tK[t]) = 0. Of course, the vanishing of HC1(tK[t]) can be
established also by direct easy calculations.
Cocycles of type (iii) vanish.
Cocycles of type (iv). Obviously, α(t i, t j) = 0 for (i, j) 6= (1, 1). Hence α is symmetric, and each
cocycle of this type has the form (12).
To summarize: cocycles of type (ia), (ii) and (iii) either are trivial or vanish, and cocycles of type (ib)
are given by formula (13) and vanish if g does not contain direct summands isomorphic to sl(2). So in
the latter case, all nontrivial cocycles are of type (iv) and given by formula (12). Considering the natural
grading in g ⊗ tK[t] by degrees of t, observing that the second cohomology is finite-dimensional and
hence is dual to the second homology, and turning to interpretation of the 2-cycles as relations between
generators, we get the assertion proved in [La1]: that g ⊗ tK[t] admits a presentation with quadratic
relations, provided g does not contain direct summands isomorphic to sl(2).
Let us decide now when cocycles given by (12) and (13) are cohomologically independent. According
to Remark 2 in §1, any cohomological dependency beyond linear dependency can occur for cocycles of
type (12) only. Writing, as in Remark 2, Φ = dΩ for Ω = ∑
ωi ⊗ βi, where ωi : g → g and βi : tK[t] →
i∈I
tK[t] are some linear maps, we get:
and
ωi([x, y])βi(t 2) = ϕ(x, y)
∑
i∈I
ωi([x, y])βi(t k) = 0
∑
i∈I
for k ≥ 3. Hence it is clear that cocycles of type (12) are cohomologically independent if and only if the
corresponding skew-symmetric bilinear forms are independent modulo 2-coboundaries.
(cid:3)
In [La1], the author speculates about the possibility to derive Theorem 4 from the standard presenta-
tion of the affine Kac-Moody algebra. Let us indicate briefly how one can do the opposite: namely, how
Theorem 4 allows us to recover the Serre relations between Chevalley generators of non-twisted affine
Kac-Moody algebras.
Let g be a finite-dimensional simple Lie algebra over an algebraically closed field K of characteristic
0, andbL = g ⊗ K[t, t−1] be a loop algebra, aka "centerless, derivation-free" part of an affine non-twisted
Kac-Moody algebra.
rearrangements of terms, can be written in the form
It is well-known that it admits a triangular decomposition which, with slight
bL =(cid:16)(g ⊗ t−1K[t−1]) ⊕ (n− ⊗ 1)(cid:17) ⊕(cid:16)h ⊗ 1(cid:17) ⊕(cid:16)(g ⊗ tK[t]) ⊕ (n+ ⊗ 1)(cid:17),
where g = n− ⊕ h ⊕ n+ is the triangular decomposition of the simple Lie algebra g ([K, §7.6]; all direct
sums are direct sums of vector spaces).
Consider the Hochschild -- Serre spectral sequence abutting to H∗((g ⊗ tK[t]) ⊕ (n+ ⊗ 1), K) with re-
spect to the ideal g ⊗ tK[t]. The general result implicitly contained in [HS] and explicitly, for example,
in [R, Lemma 1], tells that this spectral sequence degenerates at the E2 term. For the sake of simplicity,
let us exclude the case where g = sl(2). Then the E2 terms affecting the second cohomology in question
are:
E 02
∞ = E 02
E 11
∞ = E 11
E 20
∞ = E 20
2 = H 2(n+, H 0(g ⊗ tK[t], K)) ≃ H 2(n+, K),
2 = H 1(n+, H 1(g ⊗ tK[t], K)) ≃ H 1(n+, g),
2 = H 0(n+, H 2(g ⊗ tK[t], K)) ≃ (C2(g, K)/B2(g, K))n+.
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
12
The first and second isomorphisms here are obvious, the third one follows from Theorem 4. Here the first
term corresponds to relations between elements of n+ ⊗ 1, which are (classical) Serre relations for the
finite-dimensional Lie algebra g, the second term corresponds to relations between elements of g ⊗tK[t]
and n+ ⊗ 1, and the third one corresponds to relations between elements of g ⊗ tK[t]. Expressing these
elements by means of Chevalley generators in terms of the corresponding Cartan matrix, we get the
corresponding part of the Serre relations.
Repeating similar reasonings for the "minus" part, and completing relations in an obvious manner
between the "plus" and "minus" parts and the "Cartan subalgebra" h ⊗ 1, we get the complete set of
well-known construction which adds central extension and derivation, and its presentation readily fol-
defining relations forbL. The whole affine non-twisted Kac-Moody algebra is obtained frombL by the
lows from the presentation ofbL.
This approach is by no means new. For example, we find in [LP, Remark in §2]: "Similar calculations
by induction on the rank for simple finite-dimensional and loop algebras give the shortest known to us
proof of completeness of the Serre defining relations". "Induction on the rank" means induction with
repetitive application of the Hochschild -- Serre spectral sequence relative to a Kac-Moody algebra build
upon the simple finite-dimensional subalgebra of g. Here we use the Hochschild-Serre spectral sequence
in a different way, and only once, thus getting even shorter proof.
It is mentioned in [K, §9.16] that "a simple cohomological proof" of the completeness of the Serre
defining relations "was found by O. Mathieu (unpublished)". We presume that the approach outlined
here is similar to that unpublished proof.
ACKNOWLEDGEMENTS
Thanks are due to Semyon Konstein, Dimitry Leites, and Irina Shchepochkina for a careful reading
of preliminary versions of the manuscript and many suggestions which improved the presentation, to
Askar Dzhumadil'daev for drawing attention to papers [D2] and [DB] which provide an interesting
connection between the topic of this paper and classification of skew-symmetric identities of algebras,
to Clas Lofwall for drawing attention to the thesis [La2], to Anna Tihinen (Larsson) for supplying it,
and to Rutwig Campoamor-Stursberg for supplying some other needed literature.
REFERENCES
I. Bajo and S. Benayadi, Lie algebras admitting a unique quadratic structure, Comm. Algebra 25 (1997), 2795 -- 2805.
[BB]
[Be] Y. Benoist, La partie semi-simple de l'alg`ebre des d´erivations d'une alg`ebre de Lie nilpotente, Comptes Rendus
Acad. Sci. Paris 307 (1988), 901 -- 904.
[BK] F. Berezin and F.I. Karpelevich, Lie algebras with supplementary structure, Math. USSR Sbornik 6 (1968), 185 -- 203.
[Bo] N. Bourbaki, Lie Groups and Lie Algebras, Chapters 1 -- 3, Springer, 1989.
[D1] A.S. Dzhumadil'daev, Symmetric (co)homologies of Lie algebras, Comptes Rendus Acad. Sci. Paris 324 (1997),
497 -- 502.
[D2]
[DB]
[FF] B.L. Feigin and D.B. Fuchs, Cohomology of some nilpotent subalgebras of the Virasoro and Kac-Moody algebras, J.
, Algebras with skew-symmetric identity of degree 3, J. Math. Sci. 161 (2009), 11 -- 30.
and A.B. Bakirova, Simple two-sided anti-Lie-admissible algebras, J. Math. Sci. 161 (2009), 31 -- 36.
Geom. Phys. 5 (1988), 209 -- 235.
V.T. Filippov, δ-derivations of prime Lie algebras, Siber. Math. J. 40 (1999), 174 -- 184.
[F]
[G] H. Garland, Dedekind's η-function and the cohomology of infinite dimensional Lie algebras, Proc. Nat. Acad. Sci.
USA 72 (1975), 2493 -- 2495.
and J. Lepowsky, Lie algebra homology and the Macdonald-Kac formulas, Invent. Math. 34 (1976), 37 -- 76.
[GL]
[Ha] A. Haddi, Homologie des alg`ebres de Lie ´etendues `a une alg`ebre commutative, Comm. Algebra 20 (1992), 1145 --
1166.
[HW] P. Hanlon and D.B. Wales, On the homology of sln(tC[t]) and a theorem of Stembridge, J. Algebra 269 (2003), 1 -- 17.
[HS] G. Hochschild and J.-P. Serre, Cohomology of Lie algebras, Ann. Math. 57 (1953), 591 -- 603.
[Ho] N.C. Hopkins, Generalized derivations of nonassociative algebras, Nova J. Math. Game Theory Algebra 5 (1996),
215 -- 224.
[K] V. Kac, Infinite Dimensional Lie Algebras, 3rd ed., Cambridge Univ. Press, 1995.
[La1] A. Larsson, A periodisation of semisimple Lie algebras, Homology, Homotopy and Appl. 4 (2002), 337 -- 355.
LIE ALGEBRAS EXTENDED OVER COMMUTATIVE ALGEBRAS WITHOUT UNIT
13
[La2]
, Periodisations of Contragredient Lie Superalgebras and Their Presentations, PhD thesis, Stockholm Uni-
versity, 2003.
[LL] G.F. Leger and E.M. Luks, Generalized derivations of Lie algebras, J. Algebra 228 (2000), 165 -- 203.
[LP] D. Leites and E. Poletaeva, Defining relations for classical Lie algebras of polynomial vector fields, Math. Scand. 81
(1997), 5 -- 19; arXiv:math/0510019.
J.-L. Loday, Cyclic Homology, 2nd ed., Springer, 1998.
and D. Quillen, Cyclic homology and the Lie algebra homology of matrices, Comment. Math. Helv. 59
[Lo]
[LQ]
(1984), 565 -- 591.
[LR] C. Lofwall and J.-E. Roos, A nonnilpotent 1-2-presented graded Hopf algebra whose Hilbert series converges in the
unit circle, Adv. Math. 130 (1997), 161 -- 200.
[NW] K.-H. Neeb and F. Wagemann, The second cohomology of current algebras of general Lie algebras, Canad. J. Math.
[N]
[P]
[R]
[S]
60 (2008), 892 -- 922; arXiv:math/0511260.
P. Ntolo, Homologie de Leibniz d'alg`ebres de Lie semi-simples, Comptes Rendus Acad. Sci. Paris 318 (1994), 707 --
710.
T. Pirashvili, On Leibniz homology, Ann. Inst. Fourier 44 (1994), 401 -- 411.
B.I. Rozenfeld, Cohomology of the algebra of formal universal differential operators, Funct. Anal. Appl. 9 (1975),
126 -- 130.
L.J. Santharoubane, The second cohomology group for Kac-Moody Lie algebras and Kahler differentials, J. Algebra
125 (1989), 13 -- 26.
[Z1] P. Zusmanovich, The second homology group of current Lie algebras, Ast´erisque 226 (1994), 435 -- 452;
arXiv:0808.0217.
[Z2]
, Low-dimensional cohomology of current Lie algebras and analogs of the Riemann tensor for loop manifolds,
Lin. Algebra Appl. 407 (2005), 71 -- 104; arXiv:math/0302334.
E-mail address: [email protected]
|
1710.04190 | 4 | 1710 | 2018-07-05T19:31:32 | Hom-associative Ore extensions and weak unitalizations | [
"math.RA"
] | We introduce hom-associative Ore extensions as non-unital, non-associative Ore extensions with a hom-associative multiplication, and give some necessary and sufficient conditions when such exist. Within this framework, we construct families of hom-associative quantum planes, universal enveloping algebras of a Lie algebra, and Weyl algebras, all being hom-associative generalizations of their classical counterparts, as well as prove that the latter are simple. We also provide a way of embedding any multiplicative hom-associative algebra into a multiplicative, weakly unital hom-associative algebra, which we call a weak unitalization. | math.RA | math |
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK
UNITALIZATIONS
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
Abstract. We introduce hom-associative Ore extensions as non-unital, non-
associative Ore extensions with a hom-associative multiplication, and give
some necessary and sufficient conditions when such exist. Within this frame-
work, we construct families of hom-associative quantum planes, universal en-
veloping algebras of a Lie algebra, and Weyl algebras, all being hom-associative
generalizations of their classical counterparts, as well as prove that the latter
are simple. We also provide a way of embedding any multiplicative hom-
associative algebra into a multiplicative, weakly unital hom-associative alge-
bra, which we call a weak unitalization.
1. Introduction
Hom-Lie algebras and related hom-algebra structures have recently become a
subject of growing interest and extensive investigations, in part due to the prospect
of providing a general framework in which one can produce many types of natural
deformations of (Lie) algebras, in particular q-deformations which are of interest
both in mathematics and in physics. One of the main initial motivations for this
development came from mathematical physics works on q-deformations of infinite-
dimensional algebras, primarily the q-deformed Heisenberg algebras (q-deformed
Weyl algebras), oscillator algebras, and the Virasoro algebra [1 -- 8, 11, 12, 16 -- 18].
Quasi-Lie algebras, subclasses of quasi-hom-Lie algebras, and hom-Lie algebras
as well as their general colored (graded) counterparts were introduced between 2003
and 2005 in [10,13 -- 15,25]. Further on, between 2006 and 2008, Makhlouf and Silve-
strov introduced the notions of hom-associative algebras, hom-(co, bi)algebras and
hom-Hopf algebras, and also studied their properties [19 -- 21]. A hom-associative al-
gebra, being a generalization of an associative algebra with the associativity axiom
extended by a linear twisting map, is always hom-Lie admissible, meaning that the
commutator multiplication in any hom-associative algebra yields a hom-Lie alge-
bra [19]. Whereas associativity is replaced by hom-associativity in hom-associative
algebras, hom-coassociativity for hom-coalgebras can be considered in a similar
way.
One of the main tools in these important developments and in many construc-
tions of examples and classes of hom-algebra structures in physics and in math-
ematics are based on twisted derivations, or σ-derivations, which are generalized
derivations twisting the Leibniz rule by means of a linear map. These types of
twisted derivation maps are central for the associative Ore extension algebras, or
2010 Mathematics Subject Classification. 17A30, 17A01.
Key words and phrases. hom-associative Ore extensions, hom-associative Weyl algebras, hom-
associative algebras.
1
2
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
rings, introduced in algebra in the 1930s, generalizing crossed product (semidirect
product) algebras, or rings, incorporating both actions and twisted derivations.
Non-associative Ore extensions on the other hand were first introduced in 2015
and in the unital case, by Nystedt, Oinert, and Richter [23] (see also [24] for
an extension to monoid Ore extensions).
In the present article, we generalize
this construction to the non-unital case, as well as investigate when these non-
unital, non-associative Ore extensions are hom-associative. Finding necessary and
sufficient conditions for such to exist, we are also able to construct families of
hom-associative quantum planes (Example 5.9), universal enveloping algebras of
a Lie algebra (Example 5.10), and Weyl algebras (Example 5.11), all being hom-
associative generalizations of their classical counterparts. We do not make use of
any previous results about non-associative Ore extensions, but our construction of
hom-associative Weyl algebras has some similarities to the non-associative Weyl
algebras in [23]; for instance they both are simple. At last, in Section 6, we prove
constructively that any multiplicative hom-associative algebra can be embedded in
a multiplicative, weakly unital hom-associative algebra.
2. Preliminaries
In this section, we present some definitions and review some results from the
theory of hom-associative algebras and that of non-associative Ore extensions.
2.1. Hom-associative algebras. Here we define what we mean for an algebraic
structure to be hom-associative, and review a couple of results concerning the con-
struction of them. First, throughout this paper, by non-associative algebras we
mean algebras which are not necessarily associative, which includes in particular
associative algebras by definition. We also follow the convention of calling a non-
associative algebra A unital
if there exist an element 1 ∈ A such that for any
element a ∈ A, a · 1 = 1 · a = a. By non-unital algebras, we mean algebras which
are not necessarily unital, which includes also unital algebras as a subclass.
Definition 2.1 (Hom-associative algebra). A hom-associative algebra over an as-
sociative, commutative, and unital ring R, is a triple (M, ·, α) consisting of an
R-module M , a binary operation · : M × M → M linear over R in both arguments,
and an R-linear map α : M → M satisfying, for all a, b, c ∈ M ,
(1)
α(a) · (b · c) = (a · b) · α(c).
Since α twists the associativity, we will refer to it as the twisting map, and unless
otherwise stated, it is understood that α without any further reference will always
denote the twisting map of a hom-associative algebra.
Remark 2.2. A hom-associative algebra over R is in particular a non-unital, non-
associative R-algebra, and in case α is the identity map, a non-unital, associative
R-algebra.
Furthermore, if the twisting map α is also multiplicative, i.e.
if α(a · b) =
α(a) · α(b) for all elements a and b in the algebra, then we say that the hom-
associative algebra is multiplicative.
Definition 2.3 (Morphism of hom-associative algebras). A morphism between two
hom-associative algebras A and A′ with twisting maps α and α′ respectively, is an
algebra homomorphism f : A → A′ such that f ◦ α = α′ ◦ f . If f is also bijective,
the two are isomorphic, written A ∼= A′.
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
3
Definition 2.4 (Hom-associative subalgebra). Let A be a hom-associative algebra
with twisting map α. A hom-associative subalgebra B of A is a subalgebra of A
that is also a hom-associative algebra with twisting map given by the restriction of
α to B.
Definition 2.5 (Hom-ideal). A hom-ideal of a hom-associative algebra is an alge-
bra ideal I such that α(I) ⊆ I.
In the classical setting, an ideal is in particular a subalgebra. With the above
definition, the analogue is also true for a hom-associative algebra, in that a hom-
ideal is a hom-associative subalgebra.
Definition 2.6 (Hom-simplicity). We say that a hom-associative algebra A is hom-
simple provided its only hom-ideals are 0 and A.
In particular, we see that any simple hom-associative algebra is also hom-simple,
while the converse need not be true; there may exist ideals that are not invariant
under α.
Definition 2.7 (Hom-associative ring). A hom-associative ring can be seen as a
hom-associative algebra over the ring of integers.
Definition 2.8 (Weakly unital hom-associative algebra). Let A be a hom-associ-
ative algebra. If for all a ∈ A, e · a = a · e = α(a) for some e ∈ A, we say that A is
weakly unital with weak unit e.
Remark 2.9. Any unital, hom-associative algebra with twisting map α is weakly
unital with weak unit α(1), since by hom-associativity
α(1) · a = α(1) · (1 · a) = (1 · 1) · α(a) = α(a) = α(a) · (1 · 1) = (a · 1) · α(1) = a · α(1).
Any non-unital, associative algebra can be extended to a non-trivial hom-associ-
ative algebra, which the following proposition demonstrates:
Proposition 2.10 ([26]). Let A be a non-unital, associative algebra, α an algebra
endomorphism on A and define ∗ : A × A → A by a ∗ b := α(a · b) for all a, b ∈ A.
Then (A, ∗, α) is a hom-associative algebra.
Proof. Linearity follows immediately, while for all a, b, c ∈ A, we have
α(a) ∗ (b ∗ c) = α(a) ∗ (α(b · c)) = α(α(a) · α(b · c)) = α(α(a · b · c)),
(a ∗ b) ∗ α(c) = α(a · b) ∗ α(c) = α(α(a · b) · α(c)) = α(α(a · b · c)),
which proves that (A, ∗, α) is hom-associative.
(cid:3)
Note that we are abusing the notation in Definition 2.1 a bit here; A in (A, ∗, α)
does really denote the algebra and not only its module structure. From now on, we
will always refer to this construction when writing ∗.
Corollary 2.11 ([9]). If A is a unital, associative algebra, then (A, ∗, α) is weakly
unital with weak unit 1.
Proof. 1 ∗ x = α(1 · x) = α(x) = α(x · 1) = x ∗ 1.
(cid:3)
4
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
2.2. Non-unital, non-associative Ore extensions. In this section, we define
non-unital, non-associative Ore extensions, together with some new terminology.
Definition 2.12 (Left R-additivity). If R is a non-unital, non-associative ring, we
say that a map β : R → R is left R-additive if for all r, s, t ∈ R, r · β(s + t) =
r · (β(s) + β(t)).
In what follows, N will always denote the set of non-negative integers, and N>0
the set of positive integers. Now, given a non-unital, non-associative ring R with
left R-additive maps δ : R → R and σ : R → R, by a non-unital, non-associative
Ore extension of R, R[X; σ, δ], we mean the set of formal sumsPi∈N aiX i, ai ∈ R,
called polynomials, with finitely many ai nonzero, endowed with the addition
Xi∈N
aiX i +Xi∈N
biX i =Xi∈N
(ai + bi)X i,
ai, bi ∈ R,
where two polynomials are equal if and only if their corresponding coefficients are
equal, and for all a, b ∈ R and m, n ∈ N, a multiplication
(2)
aX m · bX n =Xi∈N
(a · πm
i (b)) X i+n.
i denotes the sum of all (cid:0)m
i(cid:1) possible compositions of i copies of σ and
Here πm
m − i copies of δ in arbitrary order. Then, for example π0
1 =
σ ◦ δ ◦ δ + δ ◦ σ ◦ δ + δ ◦ δ ◦ σ. We also extend the definition of πm
i by setting
πm
i ≡ 0 whenever i < 0, or i > m. Imposing distributivity of the multiplication
over addition makes R[X; σ, δ] a ring. In the special case when σ = idR, we say
that R[X; idR, δ] is a non-unital, non-associative differential polynomial ring, and
when δ ≡ 0, R[X; σ, 0] is said to be a non-unital, non-associative skew polynomial
ring.
0 = idR and π3
Note that when m = n = 0, aX 0 · bX 0 = Pi∈N(cid:0)a · π0
R ∼= RX 0 by the isomorphism r 7→ rX 0 for any r ∈ R. Since RX 0 is a subring of
R[X; σ, δ], we can view R as a subring of R[X; σ, δ], making sense of expressions
like a · bX 0.
i (b)(cid:1) X i = (a · b)X 0, so
Remark 2.13. If R contains a unit, we write X for the formal sum Pi∈N aiX i
with a1 = 1 and ai = 0 when i 6= 1. It does not necessarily make sense to think of
X as an element of the non-associative Ore extension if R is not unital.
The left-distributivity of the multiplication over addition forces δ and σ to be
left R-additive: for any r, s, t ∈ R, rX · (s + t) = rX · s + rX · t, and by expanding
the left- and right-hand side,
rX · (s + t) = r · σ(s + t)X + r · δ(s + t),
rX · s + rX · t = r · σ(s)X + r · δ(s) + r · σ(t)X + r · δ(t),
so by comparing coefficients, we arrive at the desired conclusion.
Definition 2.14 (σ-derivation). Let R be a non-unital, non-associative ring where
σ is an endomorphism and δ an additive map on R. Then δ is called a σ-derivation
if δ(a · b) = σ(a) · δ(b) + δ(a) · b holds for all a, b ∈ R. If σ = idR, δ is a derivation.
Remark 2.15. If R and σ are unital and δ a σ-derivation, then δ(1) = δ(1 · 1) =
2 · δ(1), so that δ(1) = 0. Furthermore, if R is also associative, then it is both a
necessary and sufficient condition that σ be an endomorphism and δ a σ-derivation
on R for the unital, associative Ore extension R[X; σ, δ] to exist.
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
5
Definition 2.16 (Homogeneous map). Let R[X; σ, δ] be a non-unital non-associ-
ative Ore extension of a non-unital, non-associative ring R. Then we say that a map
β : R[X; σ, δ] → R[X; σ, δ] is homogeneous if for all a ∈ R and m ∈ N, β(aX m) =
β(a)X m. If γ : R → R is any (additive) map, we may extend it homogeneously to
R[X; σ, δ] by defining γ(aX m) := γ(a)X m (imposing additivity).
3. Non-associative Ore extensions of non-associative rings
We use this small section to present a couple of results that hold true for any
non-unital, non-associative Ore extension of a non-unital, non-associative ring.
Lemma 3.1 (Homogeneously extended ring endomorphism). Let R[X; σ, δ] be a
non-unital, non-associative Ore extension of a non-unital, non-associative ring R.
If γ is an endomorphism on R, then the homogeneously extended map is an endo-
morphism on R[X; σ, δ] if and only if
(3)
γ(a) · πm
i (γ(b)) = γ(a) · γ(πm
i (b)),
for all i, m ∈ N and a, b ∈ R.
Proof. Additivity follows from the definition, while for any monomials aX m and
bX n,
γ(aX m · bX n) = γ Xi∈N
γ(aX m) · γ(bX n) = γ(a)X m · γ(b)X n =Xi∈N
i (b)X i+n! =Xi∈N
a · πm
γ(a) · γ(cid:0)πm
i (b)(cid:1)X i+n,
γ(a) · πm
i (γ(b))X i+n.
Comparing coefficients between the two completes the proof.
(cid:3)
Corollary 3.2 (Homogeneously extended unital ring endomorphism). Let R[X; σ, δ]
be a unital, non-associative Ore extension of a unital, non-associative ring R. If
α is an endomorphism on R and there exists an a ∈ R such that α(a) = 1, then
the homogeneously extended map on R[X; σ, δ] is an endomorphism if and only if
α commutes with δ and σ.
Proof. This follows from Lemma 3.1 by choosing a so that α(a) = 1: if α commutes
with δ and σ, then πm
i (α(b)) =
α(πm
0(b)) =
α(δ(b)). If choosing m = 1 and i = 1, π1
1 (b)) = α(σ(b)). (cid:3)
i (b)), then by choosing m = 1 and i = 0, π1
i (b)). On the other hand, if πm
0(α(b)) = δ(α(b)), and α(π1
1(α(b)) = σ(α(b)), α(π1
i (α(b)) = α(πm
4. Hom-associative Ore extensions of non-associative rings
The following section is devoted to the question what non-unital, non-associative
Ore extensions of non-unital, non-associative rings R are hom-associative?
Proposition 4.1 (Hom-associative Ore extension). Let R[X; σ, δ] be a non-unital,
non-associative Ore extension of a non-unital, non-associative ring R. Furthermore,
let αi,j (a) ∈ R be dependent on a ∈ R and i, j ∈ N>0, and put for an additive map
α : R[X; σ, δ] → R[X; σ, δ],
(4)
α (aX m) =Xi∈N
αi+1,m+1(a)X i,
∀a ∈ R, ∀m ∈ N.
6
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
Then R[X; σ, δ] is hom-associative with the twisting map α if and only if for all
a, b, c ∈ R and k, m, n, p ∈ N,
(5)
Xj∈NXi∈N
αi+1,m+1(a) · πi
k−j(cid:0)b · πn
j−p(c)(cid:1) =Xj∈NXi∈N
Proof. For any a, b, c ∈ R and m, n, p ∈ N,
(a · πm
i (b)) · πi+n
k−j (αj+1,p+1(c)) .
α (aX m) · (bX n · cX p) = α (aX m) ·
Xq∈N(cid:0)b · πn
α (aX m) ·(cid:0)(cid:0)b · πn
q (c)(cid:1) X q+p
q (c)(cid:1) X q+p(cid:1)
αi+1,m+1(a)X i ·(cid:0)(cid:0)b · πn
l(cid:0)b · πn
l(cid:0)b · πn
k−j(cid:0)b · πn
i (b)) X i+n! · α (cX p)
=Xq∈N
=Xq∈NXi∈N
=Xq∈NXi∈NXl∈N
=Xl∈NXj∈NXi∈N
=Xk∈NXj∈NXi∈N
αi+1,m+1(a) · πi
αi+1,m+1(a) · πi
αi+1,m+1(a) · πi
(a · πm
q (c)(cid:1) X q+p(cid:1)
q (c)(cid:1) X l+q+p
j−p(c)(cid:1) X l+j
j−p(c)(cid:1) X k,
i (b)) X i+n · α (cX p)
i (b)) X i+n ·Xj∈N
αj+1,p+1(c)X j
(a · πm
i (b)) X i+n · αj+1,p+1(c)X j
(a · πm
i (b)) · πi+n
l
(αj+1,p+1(c)) X l+j
(a · πm
i (b)) · πi+n
k−j (αj+1,p+1(c)) X k.
(a · πm
(aX m · bX n) · α (cX p) = Xi∈N
=Xi∈N
=Xi∈N
=Xi∈NXj∈N
=Xi∈NXj∈NXl∈N
=Xk∈NXj∈NXi∈N
(a · πm
Comparing coefficients completes the proof.
(cid:3)
Corollary 4.2. Let R[X; σ, δ] be a non-unital, hom-associative Ore extension of
a non-unital, non-associative ring R, with twisting map defined by (4). Then the
following assertions hold for all a, b, c ∈ R and k, p ∈ N:
I0,a
αi+1,1(a) · πi
k−p(b · c) =(a · b) · αk+1,p+1(c),
(6)
I0,a
Xi=k−p−1
Xi=k−p
αi+1,1(a) ·(cid:0)πi
k−p−1(b · σ(c))(cid:1)
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
7
+
I0,a
Xi=k−p
ai+1,1(cid:0)πi
k−p(b · δ(c))(cid:1) =(a · b) · (δ(αk+1,p+1(c)) + σ(αk,p+1(c)))
=(a · b) · (αk+1,p+1(δ(c)) + αk,p+1(σ(c))),
I1,a
Xi=k−p
αi+1,2(a) · πi
k−p(b · c) = (a · σ(b)) · (δ(αk+1,p+1(c)) + σ(αk,p+1(c)))
+ (a · δ(b)) · αk+1,p+1(c),
(7)
(8)
where α0,p+1(·) := 0, and Ip,a is the smallest natural number, depending on p and
a, such that αi+1,p(a) = 0 for all i > Ip,a.
Proof. We get (6), the first equality in (7), and (8) immediatly from the cases
m = n = 0, m = 0, n = 1, and m = 1, n = 0 in (5), respectively. The second
equality in (7) follows from comparison with (6).
(cid:3)
Remark 4.3. In case k < p, or k > I0,a in (6), (a · b) · αk+1,p+1(c) = 0. The
statement is analogous for (7) and (8).
Corollary 4.4. Let R[X; σ, δ] be a non-unital, hom-associative Ore extension of
a non-unital, non-associative ring R, with twisting map defined by (4). Then the
following assertions hold for all a, b, c ∈ R and j, p ∈ N:
(9)
(a · b) · σ(αI+1,p+1(c)) = (a · b) · αI+1,p+1(σ(c)),
(10)
I = max(Ip,c, Ip,δ(c)),
(a · b) · δ(α1,p+1(c)) = (a · b) · α1,p+1(δ(c)) =((a · b) · αj+1,j+1(δ(c))
0
if p = 0,
if p 6= 0.
Proof. Put k = max(Ip,c, Ip,δ(c)) and k = 0 in (7), respectively.
(cid:3)
5. Hom-associative Ore extensions of hom-associative rings
In this section, we will continue our previous investigation, but narrowed down
to hom-associative Ore extensions of hom-associative rings.
Corollary 5.1. Let R[X; σ, δ] be a non-unital, non-associative Ore extension of
a non-unital, hom-associative ring R, and extend the twisting map α : R → R
homogeneously to R[X; σ, δ]. Then R[X; σ, δ] is hom-associative if and only if for
all a, b, c ∈ R and l, m, n ∈ N,
(11)
Xi∈N
α(a) · πm
i (cid:0)b · πn
l−i(c)(cid:1) =Xi∈N
(a · πm
i (b)) · πi+n
l
(α(c)) .
Proof. A homogeneous α corresponds to αi+1,m+1(a) = α(a)·δi,m and αj+1,p+1(c) =
α(c) · δj,p in Proposition 4.1, where δi,m is the Kronecker delta. Then the left-hand
8
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
side reads
Xj∈NXi∈N
αi+1,m+1(a) · πi
k−j(cid:0)b · πn
j−p(c)(cid:1) =Xj∈NXi∈N
j−p(c)(cid:1)
α(a) · δi,m · πi
k−j(cid:0)b · πn
j−p(c)(cid:1)
k−p−i(c)(cid:1)
l−i(c)(cid:1) ,
k−j(cid:0)b · πn
i (cid:0)b · πn
i (cid:0)b · πn
α(a) · πm
α(a) · πm
α(a) · πm
=Xj∈N
=Xi∈N
=Xi∈N
and the right-hand side
Xj∈NXi∈N
(a · πm
i (b)) · πi+n
k−j (αj+1,p+1(c)) =Xj∈NXi∈N
(a · πm
i (b)) · πi+n
k−j (α(c) · δj,p)
(a · πm
i (b)) · πi+n
k−p (α(c))
(a · πm
i (b)) · πi+n
l
(α(c)) ,
=Xi∈N
=Xi∈N
which completes the proof.
(cid:3)
Corollary 5.2. Let R[X; σ, δ] be a non-unital, hom-associative Ore extension of
a non-unital, hom-associative ring R, with the twisting map α : R → R extended
homogeneously to R[X; σ, δ]. Then, for all a, b, c ∈ R,
(12)
(13)
(14)
(15)
(a · b) · δ(α(c)) = (a · b) · α(δ(c)),
(a · b) · σ(α(c)) = (a · b) · α(σ(c)),
α(a) · δ(b · c) = α(a) · (δ(b) · c + σ(b) · δ(c)),
α(a) · σ(b · c) = α(a) · (σ(b) · σ(c)) .
Proof. Using the same technique as in the proof of Corollary 5.1, this follows from
Corollary 4.2 with a homogeneous α.
(cid:3)
For the two last equations, it is worth noting the resemblance to the unital and
associative case (see the latter part of Remark 2.15).
Corollary 5.3. Assume α : R → R is the twisting map of a non-unital, hom-
associative ring R, and extend the map homogeneously to R[X; σ, δ]. Assume fur-
ther that α commutes with δ and σ. Then R[X; σ, δ] is hom-associative if and only
if for all a, b, c ∈ R and l, m, n ∈ N,
(16)
α(a) ·Xi∈N
πm
i (cid:0)b · πn
l−i(c)(cid:1) = α(a) ·Xi∈N(cid:0)πm
i (b) · πi+n
l
(c)(cid:1) .
Proof. Using Corollary 5.1, we know that R[X; σ, δ] is hom-associative if and only
if for all a, b, c ∈ R and l, m, n ∈ N,
Xi∈N
α(a) · πm
i (cid:0)b · πn
l−i(c)(cid:1) =Xi∈N
(a · πm
i (b)) · πi+n
l
(α(c)) .
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
9
However, since α commutes with both δ and σ, and R is hom-associative, the
right-hand side can be rewritten as
(a · πm
i (b)) · πi+n
l
Xi∈N
(α(c)) =Xi∈N
=Xi∈N
(a · πm
i (b)) · α(cid:0)πi+n
l
i (b) · πi+n
l
α(a) ·(cid:0)πm
(c)(cid:1)
(c)(cid:1) .
As a last step, we use left-distributivity to pull out α(a) from the sums.
(cid:3)
Proposition 5.4. Assume α : R → R is the twisting map of a non-unital, hom-
associative ring R, and extend the map homogeneously to R[X; σ, δ]. Assume fur-
ther that α commutes with δ and σ, and that σ is an endomorphism and δ a
σ-derivation. Then R[X; σ, δ] is hom-associative.
Proof. We refer the reader to the proof in [22], where it is seen that neither asso-
ciativity, nor unitality is used to prove that for all b, c ∈ R and l, m, n ∈ N,
(17)
Xi∈N
πm
i (cid:0)b · πn
l−i(c)(cid:1) =Xi∈N
i (b) · πi+n
πm
l
(c) ,
and therefore also (16) holds.
(cid:3)
One may further ask oneself whether it is possible to construct non-trivial hom-
associative Ore extensions, starting from associative rings? The answer is affirma-
tive, and the remaining part of this section will be devoted to show that.
Proposition 5.5. Let R[X; σ, δ] be a non-unital, associative Ore extension of a
non-unital, associative ring R, and α : R → R a ring endomorphism that commutes
with δ and σ. Then (R[X; σ, δ], ∗, α) is a multiplicative, non-unital, hom-associative
Ore extension with α extended homogeneously to R[X; σ, δ].
Proof. Since α is an endomorphism on R that commutes with δ and σ, (3) holds,
so by Lemma 3.1, the homogeneously extended map α on R[X; σ, δ] is an endomor-
phism. Referring to Proposition 2.10, (R[X; σ, δ], ∗, α) is thus a hom-associative
ring. Furthermore, we see that ∗ is the multiplication (2) of a non-unital, non-
associative Ore extension, since for all a, b ∈ R and m, n ∈ N,
aX m ∗ bX n = α Xi∈N
=Xi∈N
(a · πm
i (b)) X i+n! =Xi∈N
(a ∗ πm
i (b)) X i+n.
α (a · πm
i (b)) X i+n
(cid:3)
Remark 5.6. Note in particular that if R[X; σ, δ] is unital, then (R[X; σ, δ], ∗, α)
is weakly unital with weak unit 1 due to Corollary 2.11.
Proposition 5.7 (Hom-associative σ-derivation). Let A be an associative algebra,
α and σ algebra endomorphisms, and δ a σ-derivation on A. Assume α commutes
with δ and σ. Then σ is an algebra endomorphism and δ a σ-derivation on (A, ∗, α).
10
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
Proof. Linearity follows immediately, while for any a, b ∈ A,
σ(a ∗ b) = σ(α(a · b)) = α(σ(a · b)) = α(σ(a) · σ(b)) = σ(a) ∗ σ(b),
δ(a ∗ b) = δ(α(a · b)) = α(δ(a · b)) = α(σ(a) · δ(b) + δ(a) · b)
= α(σ(a) · δ(b)) + α(δ(a) · b) = σ(a) ∗ δ(b) + δ(a) ∗ b,
which completes the proof.
(cid:3)
Remark 5.8. For a non-unital, associative skew polynomial ring R[X; σ, 0], one can
always achieve a deformation into a non-unital, hom-associative skew polynomial
ring using Proposition 5.5 by defining the twisting map α as σ, due to the fact that
σ always commutes with itself and the zero map.
Example 5.9 (Hom-associative quantum planes). The quantum plane can be de-
fined as the unital, associative skew polynomial ring K[Y ][X; σ, 0] =: A where K
is a field of characteristic zero and σ the unital K-algebra automorphism of K[Y ]
such that σ(Y ) = qY and q ∈ K ×, K × being the multiplicative group of nonzero
elements in K. From Remark 5.8, we know that at least one nontrivial deformation
of K[Y ][X; σ, 0] into a hom-associative skew polynomial ring exist, so let us try to
see if there are others as well. Putting α(Y ) = amY m + . . . + a1Y + a0 for some
constants am, . . . , a0 ∈ K and m ∈ N and then comparing σ(α(Y )) = amqmY m +
. . . + a1qY + a0 and α(σ(Y )) = α(qY ) = qα(Y ) = amqY m + . . . + a1qY + a0q
gives α(Y ) = a1Y since q ∈ K × is arbitrary. By the same kind of argument,
α(σ(1)) = σ(α(1)) if and only if α(1) = 1. For such α and any monomial bnY n
where bn ∈ K and n ∈ N>0,
α(σ(bnY n)) =bnα(σ(Y n)) = bnα(σn(Y )) = bnαn(σ(Y )) = bnσn(α(Y ))
=bnσ(αn(Y )) = bnσ(α(Y n)) = σ(α(bnY n)).
By linearity, α commutes with σ on any polynomial in K[Y ][X; σ, 0], and by ex-
cluding the possibility α ≡ 0, we put the twisting map to be αk(Y ) = kY for
some element k ∈ K ×, the index k making evident that the map depends on the
parameter k. This K-algebra endomorphism gives us a family of hom-associative
quantum planes (A, ∗, αk), each value of k giving a weakly unital hom-associative
skew polynomial ring, the member for which k = 1 corresponding to the unital,
associative quantum plane.
If k 6= 1, we get nontrivial deformations, since for
instance X ∗ (Y ∗ Y ) = k4q2Y 2X, while (X ∗ Y ) ∗ Y = k3q2Y 2X. Now that
Proposition 5.7 guarantees that σ is a K-algebra endomorphism on any member of
(A, ∗, αk) as well, we call these members hom-associative quantum planes, satisfying
the commutation relation X ∗ Y = kqY ∗ X.
Example 5.10 (Hom-associative universal enveloping algebras). The two-dimen-
sional Lie algebra L with basis {X, Y } over the field K of characteristic zero is
defined by the Lie bracket [X, Y ]L = Y . Its universal enveloping algebra, U (L), can
be written as the unital, associative differential polynomial ring K[Y ][X; idK[Y ], δ]
where δ = Y d
dY . Put for the K-algebra endomorphism α(Y ) = anY n+. . .+a1Y +a0
where an, . . . , a0 ∈ K and n ∈ N. Then α(δ(Y )) = α(Y ) = anY n + . . . + a1Y + a0
and δ(α(Y )) = nanY n + . . . + a1Y , so by comparing coefficients, a1 is the only
nonzero such. Using the same kind of argument, α(δ(1)) = δ(α(1)) if and only if
α(1) = 1. Let bn ∈ K be arbitrary and m ∈ N>0. Then α(δ(bnY n)) = nbnα(Y n) =
nbnαn(Y ) = nbnan
1 Y n.
Since it is sufficient to check commutativity of α and δ on an arbitrary monomial, we
1 Y n, and δ(α(bnY n)) = δ(bnαn(Y )) = δ(bnan
1 Y n) = nbnan
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
11
define the twisting map as αk(Y ) = kY, k ∈ K ×, giving a family of hom-associative
universal enveloping algebras of L, (U (L), ∗, αk), where the commutation relation
X · Y − Y · X = Y is deformed to X ∗ Y − Y ∗ X = kY .
Example 5.11 (Hom-associative Weyl algebras). Consider the first Weyl algebra
exhibited as a unital, associative differential polynomial ring, K[Y ][X; idK[Y ], δ] =:
A, where K is a field of characteristic zero and δ = d
dY . Clearly any alge-
bra endomorphism α on K[Y ] commutes with idK[Y ], but what about δ? Since
α(δ(Y )) = α(1) = 1, we need to have δ(α(Y )) = 1 which implies α(Y ) = Y + k,
for some k ∈ K. On the other hand, if α is an algebra endomorphism such that
α(Y ) = Y + k for any k ∈ K, then for any monomial aY m where m ∈ N>0,
α(δ(aY m)) = amα(Y m−1) = amαm−1(Y ) = am(Y + k)m−1,
δ(α(aY m)) = aδ(αm(Y )) = aδ((Y + k)m) = am(Y + k)m−1.
Hence any algebra endomorphism α on K[Y ] that satisfies α(Y ) = Y + k for any
k ∈ K will commute with δ (and any algebra endomorphism that commutes with δ
will be on this form). Since α commutes with δ and σ, we know from Corollary 3.2
that α extends to a ring endomorphism on A as well by α(aX m) = α(a)X m.
Linearity over K follows from the definition, so in fact α extends to an algebra
endomorphism on A. Appealing to Proposition 5.5 and Remark 5.6, we thus have
a family of hom-associative, weakly-unital differential polynomial rings (A, ∗, αk)
with weak unit 1, where k ∈ K and αk is the K-algebra endomorphism defined by
αk (p(Y )X m) = p(Y + k)X m for all polynomials p(Y ) ∈ K[Y ] and m ∈ N. Since
Proposition 5.7 assures δ to be a K-linear σ-derivation on any member (A, ∗, αk)
as well, we call these hom-associative Weyl algebras, including the associative Weyl
algebra in the member corresponding to k = 0. One can note that the hom-
associative Weyl algebras all satisfy the commutation relation X ∗ Y − Y ∗ X = 1,
where 1 is a weak unit.
Lemma 5.12. Let R be a non-unital, non-associative ring. Then in R[X; idR, δ],
(18)
aX n · b =
n
Xi=0(cid:18)(cid:18)n
i(cid:19) · a · δn−i(b)(cid:19) X i,
for any a, b ∈ R and n ∈ N.
Proof. This follows from (2) with σ = idR.
(cid:3)
Lemma 5.13. Let R be a weakly unital, hom-associative ring with weak unit e and
twisting map α commuting with the derivation δ on R, and extend α homogeneously
to R[X; idR, δ]. Then the following hold:
(i) a · δn(e) = δn(e) · a = 0 for any a ∈ R and n ∈ N>0,
(ii) e is a weak unit in R[X; idR, δ],
(iii) eX · q − q · eX =Pn
Proof. First, note that
i=0 α(δ(qi))X i for any q =Pn
i=0 qiX i ∈ R[X; idR, δ].
δ(a · e) =a · δ(e) + δ(a) · e = a · δ(e) + e · δ(a),
δ(a · e) =δ(e · a) = e · δ(a) + δ(e) · a,
and hence δ(e) · a = a · δ(e). Moreover, δ(a · e) = δ(α(a)) = α(δ(a)) = e · δ(a), so
δ(e) · a = 0. Assume δn(e) · a = a · δn(e) = 0 for all n ∈ N>0. Then, since a is
12
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
arbitrary, δn(e) · δ(a) = δ(a) · δn(e) = 0 as well, and hence
0 =δ(0) = δ (a · δn(e)) = a · δn+1(e) + δ(a) · δn(e) = a · δn+1(e),
0 =δ(0) = δ (δn(e) · a) = δn(e) · δ(a) + δn+1(e) · a = δn+1(e) · a,
so the first assertion holds by induction. The second assertion follows from the first
and Lemma 5.12 with b = e, since for any m ∈ N,
aX m · e = (a · e)X m = α(a)X m = α (aX m) = (e · a)X m = e · (aX m) ,
and by distributivity of the multiplication, e·q = q·e = α(q) for any q ∈ R[X; idR; δ].
The last assertion follows from a direct computation using the first assertion and
Lemma 5.12.
(cid:3)
A well-known fact about the associative Weyl algebras are that they are simple.
This fact is also true in the case of the non-associative Weyl algebras introduced in
[23], and it turns out that the hom-associative Weyl algebras have this property as
well.
Proposition 5.14. The hom-associative Weyl algebras are simple.
Proof. The main part of the proof follows the same line of reasoning that can be
applied to the unital and associative case; let (A, ∗, αk) be any hom-associative
Weyl algebra, and I any nonzero ideal of it. Let p = Pi∈N pi(Y )X i ∈ I be an
arbitrary nonzero polynomial with pi(Y ) ∈ K[Y ], and put m := maxi(deg(pi(Y ))).
Then, since 1 ∈ A is a weak unit in (A, ∗, αk), we may use Lemma 5.13 and the
commutator [·, ·] to compute
[X, p] =Xi∈N
αk(cid:0)p′
i(Y )X i(cid:1) =Xi∈N
i(Y + k)X i.
p′
Since maxi(deg(p′
i(Y + k)) = m − 1, by applying the commutator to the resulting
polynomial with X m times, we get a polynomial Pj∈N ajX j of degree n, where
an ∈ K is nonzero. Then
Xj∈N
Y ∗Xj∈N
ajX j ∗ Y =Xj∈NXi∈N
ajX j = αk
Y Xj∈N
=Xj∈N
aj ∗ Y X j.
aj ∗ πj
i (Y )X i =Xj∈N(cid:0)aj ∗ X j−1 + aj ∗ Y X j(cid:1) ,
= αk
ajX j
Xj∈N
=Xj∈N
ajY X j
αk(cid:0)ajY X j(cid:1)
Therefore deg(cid:16)hPj∈N ajX j, Yi(cid:17) = n − 1, where deg(·) now denotes the degree of
a polynomial in X. By applying the commutator to the resulting polynomial with
Y n times, we get an ∗ 1 ∈ I;
an ∗ 1 = αk(an) = an ∈ I =⇒ a−1
n ∗ (an ∗ 1) = a−1
n ∗ an = αk(1) = 1 ∈ I.
Take any polynomial q =Pi∈N qi(Y )X i in (A, ∗, αk). Then
qi(Y )X i = q ∈ I, and therefore I = (A, ∗, αk).
1 ∗Xi∈N
qi(Y − k)X i =Xi∈N
(cid:3)
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
13
6. Weak unitalizations of hom-associative algebras
For a non-unital, associative R-algebra A consisting of an R-module M endowed
with a multiplication, one can always find an embedding of the algebra into a unital,
associative algebra by taking the direct sum M ⊕ R and defining multiplication by
(m1, r1)·(m2, r2) := (m1 ·m2 +r1 ·m2+r2 ·m1, r1 ·r2), m1, m2 ∈ M and r1, r2 ∈ R.
A can then be embedded by the injection map M → M ⊕ 0, being an isomorphism
into the unital, associative algebra M ⊕ R with the unit given by (0, 1).
In [9], Fr´egier and Gohr showed that not all hom-associative algebras can be em-
bedded into even a weakly unital hom-associative algebra. In this section, we prove
that any multiplicative hom-associative algebra can be embedded into a multiplica-
tive, weakly unital hom-associative algebra by twisting the above unitalization of
a non-unital, associative algebra with α. We call this a weak unitalization.
Proposition 6.1. Let M be a non-unital, non-associative R-algebra and α a linear
map on M . Endow M ⊕ R with the following multiplication:
(19)
(m1, r1) • (m2, r2) :=(m1 · m2 + r1 · α(m2) + r2 · α(m1), r1 · r2),
for any m1, m2 ∈ M and r1, r2 ∈ R. Then M ⊕ R is a non-unital, non-associative
R-algebra.
Proof. R can be seen as a module over itself, and since any direct sum of modules
over R is again a module over R, M ⊕ R is a module over R. For any m1, m2 ∈ M
and λ, r1, r2 ∈ R,
λ · ((m1, r1) • (m2, r2)) = λ · (m1 · m2 + r1 · α(m2) + r2 · α(m1), r1 · r2)
= (λ · m1 · m2 + λ · r1 · α(m2) + λ · r2 · α(m1), λ · r1 · r2)
= (λ · m1 · m2 + λ · r1 · α(m2) + r2 · α(λ · m1), λ · r1 · r2)
= (λ · m1, λ · r1) • (m2, r2) = (λ · (m1, r1)) • (m2, r2),
((m1, r1) + (m2, r2)) • (m3, r3) =(m1 + m2, r1 + r2) • (m3, r3)
= ((m1 + m2) · m3 + (r1 + r2) · α(m3)
+r3 · α(m1 + m2), (r1 + r2) · r3))
= (m1 · m3 + r1 · α(m3) + r3 · α(m1), r1 · r3)
+ (m2 · m3 + m2 · α(m3) + r3 · α(m2), r2 · r3)
=(m1, r1) • (m3, r3) + (m2, r2) • (m3, r3),
so the binary operation • is linear in the first argument, and by symmetry, also
linear in the second argument.
(cid:3)
Proposition 6.2 (Weak unitalization). If (M, ·, α) is a multiplicative hom-associ-
ative algebra over an associative, commutative, and unital ring R, then (M ⊕
R, •, βα) is a multiplicative, weakly unital hom-associative algebra over R with
weak unit (0, 1). Here, • is given by (19) and βα : M ⊕ R → M ⊕ R by
(20)
βα((m1, r1)) :=(α(m1), r1),
for any m1 ∈ M and r1 ∈ R.
We call (M ⊕ R, •, βα) a weak unitalization of (M, ·, α).
14
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
Proof. We proved in Proposition 6.1 that the multiplication • made M ⊕ R a non-
unital, non-associative algebra, and due to the fact that α is linear, it follows that
βα is also linear. Multiplicativity of βα also follows from that of α, since for any
m1, m2, m3 ∈ M and r1, r2, r3 ∈ R,
βα ((m1, r1)) • βα ((m2, r2)) =(α(m1), r1) • (α(m2), r2)
= (α(m1) · α(m2) + r1 · α(α(r2)) + r2 · α(α(r1)), r1 · r2)
= (α(m1 · m2) + r1 · α(α(r2)) + r2 · α(α(r1)), r1 · r2)
= (α (m1 · m2 + r1 · α(r2) + r2 · α(r1)) , r1 · r2)
=βα ((m1, r1) • (m2, r2)) ,
while hom-associativity can be proved by the following calculation:
βα ((m1, r1)) • ((m2, r2) • (m3, r3)) =(α(m1), r1) • (m2 · m3 + r2 · α(m3)
+ r3 · α(m2), r2 · r3)
= (α(m1) · (m2 · m3) + r2 · α(m1) · α(m3)
+ r3 · α(m1) · α(m2)
+ r1 · α(m2 · m3 + r2 · α(m3) + r3 · α(m2))
+r2 · r3 · α(α(m1)), r1 · r2 · r3)
= ((m1 · m2) · α(m3) + r2 · α(m1) · α(m3)
+ r3 · α(m1) · α(m2)
+ r1 · α(m2 · m3 + r2 · α(m3) + r3 · α(m2))
+r2 · r3 · α(α(m1)), r1 · r2 · r3)
= ((m1 · m2 + r1 · α(m2) + r2 · α(m1)) · α(m3)
+ r1 · r2 · α(α(m3)) + r3 · α(m1 · m2)
+ r3 · α(r1 · α(m2) + r2 · α(m1)), r1 · r2 · r3)
= ((m1, r1) • (m2, r2)) • βα((m3, r3)).
At last, (m1, r1) • (0, 1) = (0, 1) • (m1, r1) = (1 · α(m1), 1 · r1) = βα((m1, r1)). (cid:3)
Remark 6.3. In case α is the identity map, so that the algebra is associative, the
weak unitalization is the unitalization described in the beginning of this section,
thus giving a unital algebra.
Corollary 6.4. (M, ·, α) ∼= (M ⊕ 0, •, βα).
Proof. The projection map π : M ⊕ 0 → M is a bijective algebra homomorphism.
For any m ∈ M , π(βα(m, 0)) = π(α(m), 0) = α(m) and α(π(m, 0)) = α(m),
therefore α ◦ π = π ◦ βα, so by Definition 2.3, (M ⊕ 0, •, βα) ∼= (M, ·, α).
(cid:3)
Using Corollary 6.4, we identify (M, ·, α) with its image in (M ⊕ R, •, βα), seeing
the former as embedded in the latter.
Lemma 6.5. All ideals in a weakly unital hom-associative algebra are hom-ideals.
Proof. Let I be an ideal, a ∈ I and e a weak unit in a hom-associative algebra.
Then α(a) = e · a ∈ I, so α(I) ⊆ I.
(cid:3)
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
15
A simple hom-associative algebra is always hom-simple, the hom-associative
Weyl algebras in Example 5.11 being examples thereof. The converse is also true if
the algebra has a weak unit, due to Lemma 6.5.
Corollary 6.6. (M, ·, α) is a hom-ideal in (M ⊕ R, •, βα).
Proof. For any m1, m2 and r1 ∈ R, (m1, r1)•(m2, 0) = (m1 ·m2+r1 ·α(m2), 0) ∈ M ,
and (m2, 0) • (m1, r1) = (m2 · m1 + r1 · α(m2), 0) ∈ M , so (M, ·, α) is an ideal in
a weakly unital hom-associative algebra, and by Lemma 6.5 therefore also a hom-
ideal.
(cid:3)
Recall that for a ring R, if there is a positive integer n such that n · a = 0 for all
a ∈ R, then the smallest such n is the characteristic of the ring R, char(R). If no
such positive integer exists, then one defines char(R) = 0.
Proposition 6.7. Let R be a weakly unital hom-associative ring with weak unit
e and injective or surjective twisting map α.
If n · e 6= 0 for all n ∈ Z>0, then
char(R) = 0.
If n · e = 0 for some n ∈ Z>0, then the smallest such n is the
characteristic of R.
Proof. If n·e 6= 0 for all n ∈ Z>0, then clearly we cannot have n·a = 0 for all a ∈ R,
and hence char(R) = 0. Now assume n is a positive integer such that n · e = 0. If
α is injective, then for all a ∈ R,
α(n · a) = n · α(a) = n · (e · a) = (n · e) · a = 0 · a = 0 ⇐⇒ n · a = 0.
On the other hand, if α is surjective, then for all a ∈ R, a = α(b) for some b ∈ R,
and hence n · a = n · α(b) = n · (e · b) = (n · e) · b = 0 · b = 0.
(cid:3)
Proposition 6.8. Let R := (M, ·, α) be a hom-associative ring, and define
S :=((M ⊕ Z, •, βα),
(M ⊕ Zn, •, βα),
if char(R) = 0,
if char(R) = n.
Then the weak unitalization S of R has the same characteristic as R.
Proof. This follows immediately by using the definition of the characteristic.
(cid:3)
The main conclusion to draw from this section is that any multiplicative hom-
associative algebra can be seen as a multiplicative, weakly unital hom-associative
algebra by its weak unitalization. The converse, that any weakly unital hom-
associative algebra is necessarily multiplicative if also α(e) = e, where e is a weak
unit, should be known. However, since we have not been able to find this statement
elsewhere, we provide a short proof of it here for the convenience of the reader.
Proposition 6.9. If e is a weak unit in a weakly unital hom-associative algebra
A, and α(e) = e, then A is multiplicative.
Proof. For any a, b ∈ A, α(e)·(a · b) = e ·(a · b) = α (a · b). Using hom-associativity,
α(e) · (a · b) = (e · a) · α(b) = α(a) · α(b).
(cid:3)
Acknowledgment. We would like to thank Lars Hellstrom for discussions leading
to some of the results presented in the article.
16
PER B ACK, JOHAN RICHTER, AND SERGEI SILVESTROV
References
[1] N. Aizawa and H. Sato, q-deformation of the Virasoro algebra with central extension, Phys.
Lett. B, 256(2) (1991), 185 -- 90 (Preprint Hiroshima University HUPD-9012 (1990)).
[2] M. Chaichian, D. Ellinas, and Z. Popowicz, Quantum conformal algebra with central exten-
sion, Phys. Lett. B, 248 (1990), 95 -- 99.
[3] M. Chaichian, A. P. Isaev, J. Lukierski, Z. Popowic, and P. Presnajder, q-deformations of
Virasoro algebra and conformal dimensions, Phys. Lett. B, 262(1) (1991), 32 -- 38.
[4] M. Chaichian, P. Kulish, and J. Lukierski, q-deformed Jacobi identity, q-oscillators and q-
deformed infinite-dimensional algebras, Phys. Lett. B, 237(3 -- 4) (1990), 401 -- 406.
[5] M. Chaichian, Z. Popowicz, and P. Presnajder, q-Virasoro algebra and its relation to the
q-deformed KdV system Phys. Lett. B, 249(1) (1990), 63 -- 65.
[6] T. L. Curtright and C. K. Zachos, Deforming maps for quantum algebras, Phys. Lett. B,
243(3) (1990), 237 -- 244.
[7] T. L. Curtright, D. B. Fairlie, and C. K. Zachos, Ternary Virasoro-Witt algebra, Phys. Lett.
B, 666(4) (2008), 386 -- 390.
[8] C. Daskaloyannis, Generalized deformed Virasoro algebras, Modern Phys. Lett. A, 7(9)
(1992), 809 -- 816.
[9] Y. Fregier and A. Gohr, On unitality conditions
for Hom-associative algebras,
arXiv:0904.4874 (2009).
[10] J. T. Hartwig, D. Larsson, and S. D. Silvestrov, Deformations of Lie algebras using σ-
derivations, J. of Algebra, 295(2) (2006), 314 -- 361 (Preprint Lund University LUTFMA-
5036-2003 (2003)).
[11] N. Hu, q-Witt algebras, q-Lie algebras, q-holomorph structure and representations, Algebra
Colloq., 6(1) (1999), 51 -- 70.
[12] C. Kassel, Cyclic homology of differential operators, the Virasoro algebra and a q-analogue,
Comm. Math. Phys., 146(2) (1992), 343 -- 356.
[13] D. Larsson and S. D. Silvestrov, Quasi-Hom-Lie algebras, central extensions and 2-cocycle-
like identities, J. Algebra, 288(2) (2005), 321 -- 344 (Preprint Lund University LUTFMA-5038-
2004 (2004)).
[14] D. Larsson and S. D. Silvestrov, Quasi-Lie Algebras, in "Noncommutative geometry and
representation theory in mathematical physics", eds. J. Fuchs, J. Mickelsson, G. Rozenblioum,
A. Stolin, A. Westerberg, Amer. Math. Soc., Providence, RI, Contemp. Math., vol. 391,
(2005), 241-248. (Preprint Lund University LUTFMA-5049-2004 (2004)).
[15] D. Larsson and S. D. Silvestrov, Graded quasi-Lie algebras, Czechoslovak J. Phys., 55(11)
(2005), 1473 -- 1478.
[16] K. Q. Liu, Quantum central extensions, C. R. Math. Rep. Acad. Sci. Canada. 13 (1991),
135 -- 140.
[17] K. Q. Liu, Characterizations of the quantum Witt algebra, Lett. Math. Phys., 24(4) (1992),
257 -- 265.
[18] K. Q. Liu, The Quantum Witt Algebra and Quantization of Some Modules over Witt Algebra,
PhD thesis, University of Alberta, 1992.
[19] A. Makhlouf and S. D. Silvestrov, Hom-algebra structures, J. Gen. Lie Theory Appl., 2(2)
(2008), 51 -- 64 (Preprint Lund University LUTFMA-5074-2006 (2006)).
[20] A. Makhlouf and S. D. Silvestrov, Hom-algebras and hom-coalgebras, J. Algebra Appl., 9(4)
(2010), 553 -- 589 (Preprint Lund University LUTFMA-5103-2008 (2008) and arXiv:0811.0400
(2008)).
[21] A. Makhlouf A and S. D. Silvestrov, Hom-Lie admissible hom-coalgebras and hom-Hopf
algebras, in Generalized Lie Theory in Mathematics, Physics and Beyond, S. Silvestrov, E.
Paal et al eds., Springer, Berlin, 2009, chapter 17 pp 189 -- 206 (Preprint Lund University
LUTFMA-5091-2007 (2007) and arXiv:0709.2413 (2007)).
[22] P. Nystedt, A combinatorial proof of associativity of Ore extensions, Discrete Math., 313(23)
(2013), 2748 -- 2750.
[23] P. Nystedt, J. Oinert, and J. Richter, Non-associative Ore extensions, Isr. J. Math., 224(1)
(2018), 263 -- 292, arXiv:1509.01436 (2015).
[24] P. Nystedt, J. Oinert, and J. Richter, Monoid Ore extensions, arXiv:1705.02778 (2017).
[25] G. Sigurdsson and S. Silvestrov, Graded quasi-Lie algebras of Witt type, Czech J Phys.,
56(10 -- 11) (2006), 1287 -- 1291.
HOM-ASSOCIATIVE ORE EXTENSIONS AND WEAK UNITALIZATIONS
17
[26] D. Yau, Hom-algebras and homology, J. Lie Theory, 19 (2009), 409 -- 421.
Division of Applied Mathematics, The School of Education, Culture and Communi-
cation, Malardalen University, Box 883, SE-721 23 Vasteras, Sweden
E-mail addresses: [email protected], [email protected], and [email protected]
|
1410.8306 | 2 | 1410 | 2017-10-23T16:08:45 | Algebraic entropy of amenable group actions | [
"math.RA",
"math.GR",
"math.RT"
] | Let $R$ be a ring, let $G$ be an amenable group and let $R\ast G$ be a crossed product. The goal of this paper is to construct, starting with a suitable additive function $L$ on the category of left modules over $R$, an additive function on a subcategory of the category of left modules over $R\ast G$, which coincides with the whole category if $L({}_RR) <\infty$. This construction can be performed using a dynamical invariant associated with the original function $L$, called algebraic $L$-entropy. We apply our results to two classical problems on group rings: the Stable Finiteness and the Zero-Divisors Conjectures. | math.RA | math |
Algebraic entropy of amenable group actions
Simone Virili∗
August 24, 2018
Abstract
Let R be a ring, let G be an amenable group and let RG be a crossed product. The goal of this
paper is to construct, starting with a suitable additive function L on the category of left modules over
R, an additive function on a subcategory of the category of left modules over RG, which coincides
with the whole category if LpRRq ă 8. This construction can be performed using a dynamical invariant
associated with the original function L, called algebraic L-entropy. We apply our results to two classical
problems on group rings: the Stable Finiteness and the Zero-Divisors Conjectures.
-- -- -- -- -- -- -- -- -- -- --
2010 Mathematics Subject Classification. 16S35, 43A07, 16D10, 18E35.
Key words and phrases. Length functions, Gabriel dimension, algebraic entropy, amenable groups,
zero divisors, stable finiteness.
Contents
1 Introduction
2 Length functions and torsion theories
2.1 Length functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Gabriel localization and Gabriel dimension . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Length functions compatible with self-equivalences . . . . . . . . . . . . . . . . . . . . . .
2.4 Examples of length functions
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
5
5
7
9
9
3 Amenable groups and crossed products
10
3.1 Amenable groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Non-negative real functions on finite subsets of an amenable group . . . . . . . . . . . . . 11
3.3 Crossed products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4 The algebraic L-entropy
13
4.1 Definition and basic properties
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 The algebraic entropy is upper continuous . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3 The algebraic entropy is additive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4 Values on (sub)shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5 Applications
18
5.1 Stable finiteness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.2 Zero-Divisors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
∗The author was partially supported by DGI MINECO MTM2011-28992-C02-01, MINECO MTM2014-53644-P, by the
Comissionat per Universitats i Recerca de la Generalitat de Catalunya and by the Fondazione Cassa di Risparmio di Padova
e Rovigo (Progetto di Eccellenza "Algebraic structures and their applications")
1
1
Introduction
Let V be a finite dimensional vector space over a given field K, and consider a surjective endomorphism
φ : V Ñ V . Then φ has to be injective, and one way to show this is to compute dimensions:
dimKpV q " dimKpφpV qq ` dimKpKerpφqq " dimKpV q ` dimKpKerpφqq
so that
dimKpKerpφqq " 0 ,
hence Kerpφq " 0. A left module M over a ring R is said to be Hopfian if any of its surjective
endomorphisms is an automorphism. Thus, the above argument shows that finite dimensional vector
spaces are Hopfian modules. Repeating these computations, but with composition length ℓp´q of modules
used instead of dimension, one can show that modules of finite length over any ring are Hopfian.
Let us try to abstract the above idea: fix a ring R and consider a function L : R-Mod Ñ Rě0 Y t8u,
that associates to any left R-module M the value LpM q P Rě0 Y t8u. We say that L is
-- additive if, given a short exact sequence 0 Ñ A Ñ B Ñ C Ñ 0 in R-Mod, LpBq " LpAq ` LpCq;
-- faithful provided LpM q " 0 if and only if M " 0.
If we have a faithful additive function on R-Mod, then we can conclude in the exact same way that all
the modules in the class FinpLq :" tM P R-Mod : LpM q ă 8u are Hopfian.
A rich source of Hopfian modules is the class of Noetherian modules. The usual proof that a Noethe-
rian module M is Hopfian is as follows: any surjective endomorphism φ : M Ñ M induces isomorphisms
Kerpφq -- Ð Kerpφ2q{Kerpφq -- Ð Kerpφ3q{Kerpφ2q -- Ð . . .
By the ascending chain condition, the chain Kerpφq Ď Kerpφ2q Ď Kerpφ3q Ď . . . stabilizes, so there exists
n P N such that 0 -- Kerpφn`1q{Kerpφnq -- Kerpφq. There is another way to show that Noetherian
modules are Hopfian: using additive functions. In fact, in this generality, it is not possible to use just
one faithful additive function, but we need to introduce an entire family of (not always faithful) additive
functions that, when used together, allow us to conclude:
The first step is to notice that, given a hereditary torsion class T Ď R-Mod, there is an additive
function
ℓT : R-Mod Ñ Rě0 Y t8u ,
defined as
ℓT pM q " sup
ℓT pσq ,
σ
where σ ranges in all the finite sequences of the form σ : 0 " M0 Ď M1 Ď . . . Ď Mn " M and
ℓT pσq :" #ti P t1, . . . , nu : Mi{Mi´1 R T u. In other words, ℓT pM q is the composition length of M in
the Gabriel quotient R-Mod{T (see Section 2). Notice that ℓT pM q " 0 if and only if M P T .
Now that we have a procedure that takes as an input a hereditary torsion class and produces an
additive function, we need a source for torsion classes. Thus, our second step is to recall that the Gabriel
filtration of R-Mod is a transfinite sequence of hereditary torsion classes Cα Ď R-Mod such that:
-- C0 is the class containing just the 0-modules;
-- if Cα is given, Cα`1 is the smallest hereditary torsion class containing Cα and all the modules M
such that, given N ď M , exactly one between N and M {N is in Cα (these are those M that become
simple in the Gabriel quotient R-Mod{Cα);
-- if λ is a limit ordinal, Cλ is the smallest hereditary torsion class containing all the Cα, for α ă λ.
For any ordinal α, we denote by ℓα : R-Mod Ñ Rě0 Y t8u the additive function induced by Cα with the
above procedure.
At this point, we can conclude with few observations: given a Noetherian left R-module M and
a surjective endomorphism φ : M Ñ M suppose, looking for a contradiction, that Kerpφq ‰ 0. Using
that Kerpφq is Noetherian, we can find an ordinal α such that Kerpφq P Cα`1zCα. Considering the
α ` 1-torsion part of M , Tα`1M :" tx P M : Rx P Cα`1u, we obtain the following short exact sequence:
0 Ñ Kerpφq Ñ Tα`1M Ñ Tα`1M Ñ 0 .
Furthermore, ℓαpKerpφqq ď ℓαpTα`1M q ă 8 (where the finiteness of these quantities is a consequence
of Noetherianity, see Lemma 2.15. A simple way to see this is that any object in the quotient category
Cα`1{Cα is semi-Artinian and objects that are both Noetherian and semi-Artinian do have finite length).
Thus, by the additivity of ℓα, ℓαpKerpφqq " 0, that is, Kerpαq P Cα, which is a contradiction. Thus
Kerpφq " 0 and φ is an automorphism as desired.
2
To summarize, the above strategy consists in: finding a suitable filtration of R-Mod by hereditary
torsion classes, associate to any of these torsion classes an additive function, and finally select a suitable
length function in our family to prove the injectivity of the given map. One of the main observation of
this paper is that this same strategy can be adapted to find a partial solution of the following famous
conjecture:
Stable Finiteness Conjecture (Kaplansky, 1972). Let K be a field and let G be a group. Then, the
group ring KrGs is stably finite.
The above conjecture was verified by Kaplansky [Kap95] in case K " C, and similar arguments allow
to solve the conjecture in the positive for any field of zero characteristic. Furthermore, it was proved by
Ara, O'Meara and Perera [AOP02] that KrGs is stably finite whenever G is residually amenable and K
is an arbitrary division ring. Soon after, this last result was extended by Elek and Szab´o [ES04], who
proved that the group ring KrGs is stably finite for K any division ring and G a sofic group. Let us also
remark that Ara et al. proved that any crossed product KG (not just the group ring KrGs, see Section
3.3) is stably finite, provided K is a division ring and G is an amenable group (see Section 3.1).
The relation between the above conjecture and Hopficity is clear: the stable finiteness of the group
ring KrGs means exactly that all the free KrGs-modules of finite rank KrGsn are Hopfian. In this paper
we extend, and provide a new proof for, the known positive solution of the above conjecture in case G
is an amenable group. Our main result in this direction will be the following:
Theorem A. Let R be a left Noetherian ring, let G be a finitely generated amenable group and let RG
be a fixed crossed product. Then, for any finitely generated left R-module RK P R-Mod, any RG-
submodule of RG bR K is Hopfian.
In particular, EndRGpM q is stably finite for any submodule RGM ď RG bR K.
Notice that the above theorem is an extremely strong version of the Stable Finiteness Conjecture
that cannot be extended to sofic groups. In fact, the above statement fails for free non-Abelian groups
(see Example 5.6 and Problem 5.7). On the other hand, recently Li and Liang [LL15] were able to adapt
our strategy to show that RrGs is stably finite whenever R is left Noetherian and G is sofic. It is still
an open question if the same is true for more general crossed products.
Let us briefly describe our strategy to prove Theorem A. The idea is very close to the one we sketched
above in proving the Hopficity of Noetherian modules using additive functions. Fix a crossed product
RG, with G finitely generated amenable and R Noetherian. With a little abuse of notation, for any
ordinal α we denote by Cα the hereditary torsion class of those RG-modules RGM whose underlying
R-module RM is in Cα (the α-th layer of the Gabriel filtration of R-Mod, as described above). Now
we need suitable additive functions of RG-Mod, one for each α. To do this, we need to "extend" the
additive functions ℓα : R-Mod Ñ Rě0 Y t8u from R-Mod to RG-Mod. This problem of extension of a
given additive function is an interesting matter of dynamical nature, that deserves a brief discussion in its
own (see Theorem B below). For the moment, it is enough for us to know that this is the point where we
use the amenability of G to produce, for any ordinal α, an additive function Lα : RG-Mod Ñ Rě0 Yt8u
such that:
-- "Lα extends ℓα", that is, LαpRG bR Kq " ℓαpKq for any left R-module K such that ℓαpKq ă 8;
-- "Lα is as faithful as it can be", that is, LαpRGN q " 0 if and only if ℓαpRN q " 0, for any
RG-submodule N ď RG bR K, with K as above.
let K be a finitely generated left R-module, consider an RG-submodule
Now it is easy to conclude:
M ď RGbRK, and let φ : M Ñ M be a surjective endomorphism. Suppose, looking for a contradiction,
that Kerpφq ‰ 0. Using the Noetherianity of K, it is possible to find an ordinal α such that RKerpφq P
Cα`1zCα. Then, there is a short exact sequence of RG-modules
Furthermore, since Tα`1pRG bR Kq -- RG bR Tα`1K,
0 Ñ Kerpφq Ñ Tα`1M Ñ Tα`1M Ñ 0 .
LαpKerpφqq ď LαpTα`1M q ď LαpRG bR Tα`1Kq " ℓαpTα`1Kq ă 8
where the equality uses the first property of Lα listed above. By the additivity of Lα, we obtain that
LαpKerpφqq " 0 and so, by the second property listed above, ℓαpRKerpφqq " 0, so Kerpφq P Cα, which is
a contradiction, so Kerpφq " 0 and φ is an automorphism as desired.
As we explained above, in order to apply the machinery of torsion classes and additive functions to
Kaplansky's Stable Finiteness Conjecture we need to answer (at least partially), the following question:
3
Let R be a ring, let G be an amenable group and fix a crossed product RG. Is it possible to
find a map
tadditive functions on R-Modu ÝÑ tadditive functions on RG-Modu
L ÞÝÑ LRG
such that LRGpN q " 0 implies LpN q " 0 for any N ď RG bR K, and LRGpRG bR Kq "
LpKq, for any K P FinpLq?
In fact we will restrict our attention to a particular class of additive functions called length functions,
where an additive function L : R-Mod Ñ Rě0 Y t8u is a length function provided
LpM q " suptLpF q : F ď M finitely generatedu .
The composition length ℓ, and more generally all the additive functions of the form ℓα arising from the
Gabriel filtration, are in fact length functions. Let us also remark that, even when R is a field and L
is the dimension of vector spaces, Elek [Ele99] showed that it is not always possible to extend L to a
length function of RrGs-Mod if G is not an amenable group, so it is essential to include amenability
in our hypotheses. In fact, Elek [Ele03] showed that (again for R a field and L the dimension) this is
always possible when G is finitely generated amenable.
Given a ring R (no hypothesis is needed on the ring) and a finitely generated amenable group G, fix
a crossed product RG and a length function L : R-Mod Ñ Rě0 Y t8u which is compatible with RG
(see Definition 3.6). In particular, all the functions ℓα arising from some layer in the Gabriel filtration
satisfy these hypotheses.
Given a left RG-module RGM , let FinLpM q :" tRK ď M : LpKq ă 8u. If FinLpM q contains all the
finitely generated R-submodules of M we say that M is locally L-finite and we denote by lFinLpRGq
the class of such modules. For example, lFinLpRGq contains all the left RG-modules RGM such that
LpM q ă 8. Furthermore, given a left R-module RK such that LpKq ă 8, then the left RG-module
RG bR K is in lFinLpRGq. In Section 4 we define the algebraic L-entropy as a function
entL : lFinLpRGq Ñ Rě0 Y t8u .
The main result of Section 4 is the following theorem, showing that the algebraic entropy provides a
partial answer to above extension problem:
Theorem B. In the above hypotheses and notation, the invariant entL : lFinLpRGq Ñ Rě0 Y t8u
satisfies the following properties:
(1) entL is a length function;
(2) entLpRG bR Kq " LpKq for any L-finite left R-module K;
(3) entLpRGN q " 0 if and only if LpRN q " 0, for any RG-submodule N ď RG bR K.
The proof of the above theorem will take almost all Section 4, in particular, part (1) of the theorem
will be proved in Sections 4.2 and 4.3, while parts (2) and (3) in Section 4.4.
The final Section 5 is devoted to two applications of the algebraic entropy: the already mentioned
Stable Finiteness Conjecture, and the Zero Divisors Conjecture, which states that, for a torsion-free
group G and a field K, the group ring KrGs is a domain. Using the algebraic entropy we will give a new
proof of the fact, known for group rings, that, provided G is amenable, a given crossed product K G is
a domain if and only if it is an Ore domain. Furthermore, using an approach similar to that of Chung
and Thom [CT13], we will show that the Zero Divisors Conjecture is equivalent to a statement about
the possible range of the algebraic entropy, extending their results from finite fields to arbitrary skew
fields, and from group rings to crossed products (see Theorem 5.9).
It is a pleasure for me to thank Peter V´amos for giving me a copy of his
Acknowledgements.
Ph.D. thesis and for useful discussions started in Padova in 2010. I am also sincerely grateful to my
Ph.D. advisor Dolors Herbera for her encouragement and trust: she gave me time, freedom and several
suggestions that helped me to work independently on this project. Finally, I would like to thank Pere
Ara, Ferran Ced´o, Hanfeng Li, Bingbing Liang and Nhan-Phu Chung for some useful discussions on
preliminary versions of this paper.
4
2 Length functions and torsion theories
All along this section, unless otherwise stated, C denotes a Grothendieck category, that is, a cocomplete
Abelian category, with exact directed colimits (i.e., C satisfies Grothendieck's axiom (Ab.5)) and with a
generator. For unexplained terms and notations we refer to [Ste71].
Remember that any Grothendieck category is well-powered (see [Ste71, Proposition 6, p.94]) and
that the set of all subobjects of a given object M is a bounded and bicomplete lattice, which we denote
by LpM q; the minimum of LpM q is the 0-object and the maximum is M itself. For every family
tNi : i P Iu of subobjects of M we denote by řI Ni (resp., ŞI Ni) the join (resp., meet) of the Ni in
LpM q. If tNi : i P Iu is a directed system of subobjects, řI Ni is called the direct union of the Ni
and it is sometimes denoted also by ŤI Ni. With this notation we can state the equivalent form of
Grothendieck's axiom Ab.5 stating that, given an object M in C, a subobject K of M and a directed
system tNi : i P Iu of subobjects of M ,
ÿI
Ni¸ X K "ÿI
pNi X Kq .
Let us recall the following fundamental properties of Grothendieck categories:
Fact 2.1.
[Ste71, Corollaries 4.3 and 4.4, pp.222 -- 223] Let C be a Grothendieck category, then
(1) C is complete;
(2) C has enough injective objects.
Given a family tMi : i P Iu of objects of C, we denote by śI Mi and ÀI Mi the product and the
coproduct respectively. Given an object M of C, we denote by EpM q the injective envelope of M . Given
an endomorphism φ : M Ñ M , a subobject N of M is φ-invariant if φpN q ď N . Any subobject of a
quotient (or, equivalently, a quotient of a subobject) of an object M is generically called a segment of
M . A series of M is a finite chain of subobjects
σ : 0 " N0 ď N1 ď ď Nn " M .
(2.1)
The factors of σ are the segments of M of the form Ni`1{Ni for some i ă n. Two series of M are
equivalent if they have the isomorphic factors up to reordering. If all the factors of σ are simple objects,
then we say that σ is a composition series for M . If σ1 0 " N 1
m " M is another
series of M , we say that σ1 is a refinement of σ if tN0, . . . , Nnu Ď tN 1
mu. The following fact is
well-known.
1 ď ď N 1
0 ď N 1
0, . . . , N 1
Fact 2.2 (Artin-Schreier's Refinement Theorem). Let C be a Grothendieck category and let M be an
object of C. If σ1 and σ2 are two series of M , then they admit equivalent refinements.
2.1 Length functions
In any category C it is possible to define (real-valued) invariants to measure various finiteness properties
of the objects. In general, we call invariant of C, any map i : C Ñ Rě0 Y t8u such that ipXq " ipX 1q
whenever X and X 1 are isomorphic objects in C.
If we make some stronger assumption on the structure of the category C, we can refine our definition
of invariant in order to obtain a more treatable notion. Indeed, suppose that C is an Abelian category
(or, more generally, an exact category). The information that one usually wants to encode is about
homological properties. In this setting it seems natural to ask that, given a short exact sequence
0 Ñ X1 Ñ X2 Ñ X3 Ñ 0
(2.2)
in C, we have ipX2q " ipX1q ` ipX3q. In this case, we say that i is additive on the sequence (2.2). If i is
additive on all the short exact sequences of C and ip0q " 0, then we say that i is an additive invariant
(or additive function).
In the following lemma we collect some useful properties of additive functions.
Lemma 2.3. Let C be an Abelian category and let i : C Ñ Rě0 Y t8u be an additive function. Then
(1) ipXq ě ipY q for every segment Y of X P C;
(2) ipX1 ` X2q ` ipX1 X X2q " ipX1q ` ipX2q for every pair of subobjects X1, X2 of X P C;
5
A natural assumption in the context of Grothendieck categories is the upper continuity of invariants,
(3) řjodd ipXj q "řjeven ipXjq for every exact sequence 0 Ñ X1 Ñ X2 Ñ Ñ Xn Ñ 0 in C.
given an object X P C and a directed set S " tXα : α P Λu of subobjects of X such that řΛ Xα " X,
we say that i is continuous on S if
ipXq " suptipXαq : α P Λu .
(2.3)
If i is continuous on all the direct unions as above, we say that i is upper continuous. Upper continuity
can be defined in arbitrary Abelian categories even if it seems more meaningful when direct limits exist
and are exact.
Definition 2.4 ( [NR65]). Let C be an Abelian category. An additive and upper continuous invariant
i : C Ñ Rě0 Y t8u is said to be a length function.
In what follows we generally denote length functions by the symbol L. We remark that the usual
definition of length function is given in module categories, which are in particular locally finitely gener-
ated Grothendieck categories. In this special setting, the usual definition of upper continuity is different
(see part (3) of the following proposition). We now show that we are not defining a new notion of upper
continuity but just generalizing this concept to arbitrary Grothendieck categories (similar observations,
with analogous proofs, were already present in [V´am68] for module categories).
Recall that, given an object M of a Grothendieck category C and an ordinal κ, a set tMα : α ă κu
of subobjectsis of M is a continuous chain if
(1) Mα ď Mβ, provided α, β ă κ and α ď β;
(2) Ťαăλ Mα " Mλ for any limit ordinal λ ă κ.
Proposition 2.5. Let C be a Grothendieck category and L : C Ñ Rě0 Y t8u be an additive function.
Consider the following statements:
(1) L is a length function;
(2) given an object M P C, an ordinal κ and a continuous chain tMα : α ă κu of subobjects of M such
that M "Ťαăκ Mα, we have that LpM q " suptLpMαq : α ă κu;
(3) for every object M P C we have that LpM q " suptLpF q : F finitely generated subobject of M u.
Then (1)ô(2) and (2)ð(3). If C is locally finitely generated, then the above statements are all equivalent.
In the last part of this subsection we show that, under suitable conditions, to prove the additivity of
an invariant it is enough to show that it is additive on short exact sequences of finitely generated objects
(see Proposition 2.7). This will be extremely useful in proving that the algebraic entropy is additive.
Lemma 2.6. Let I be a set, let tai : i P Iu and tbi : i P Iu be bounded subsets of Rě0, and let a :" supi ai
and b :" inf i bi. If ai ď a, b ď bi and ai ` bi " c for all i P I, then a ` b " c.
Proof. For any ε ą 0 there exists i P I such that a ´ ai ă ε and b ´ bi ă ε. Thus,
c ´ ε " ai ` pbi ´ εq ă a ` b ă pai ` εq ` bi " c ` ε .
Proposition 2.7. Let R be a ring and let C Ď R-Mod be a subclass closed under subobjects and quotients.
An upper continuous invariant L : C Ñ Rě0 Y t8u such that LpF q ă 8 for any finitely generated F P C
is additive if and only if the following properties hold true:
(1) LpK2q " LpK1q ` LpK3q for any short exact sequence 0 Ñ K1 Ñ K2 Ñ K3 Ñ 0 of finitely
generated modules in C;
(2) given a finitely generated K P C and a directed system of submodules tHiui, LpK{Hq " inf i LpK{Hiq,
where H :"Ťi Hi.
Proof. Consider a short exact sequence 0 Ñ N Ñ M Ñ M {N Ñ 0 in C. Then, LpM q " supF LpF q,
with F ranging in the finitely generated submodules of M . Furthermore, for any such F we have that
LpN X F q " supG LpGq, LpF {pN X F qq " inf G LpF {Gq and LpGq ` LpF {Gq " LpF q, for any G ď
N X F finitely generated (where for these three equalities we used upper continuity, hypothesis (2) and
6
hypothesis (1), respectively). By the above lemma, this implies that LpF q " LpN X F q ` LpF {pN X F qq.
Thus,
LpM q " sup
LpF q
F ďM f.g.
" sup
LpN X F q ` LpF {pN X F qq
F ďM f.g.
" sup
LpN X F q ` sup
LppF ` N q{N q
F ďM f.g.
F ďM f.g.
" LpN q ` LpM {N q
On the other hand, if L is additive, then (1) is clear. Furthermore, given a finitely generated K P C and
a directed system of submodules tHiui, then
LpK{Hq " LpKq ´ LpHq " LpKq ´ sup
i
LpHiq " inf
i
pLpKq ´ LpHiqq " inf
i
LpK{Hiq .
2.2 Gabriel localization and Gabriel dimension
Let C be a Grothendieck category and let A Ď C be a subclass. Recall that A is a semi-closed if, given
a short exact sequence
0 Ñ A Ñ B Ñ C Ñ 0 ,
B P A implies that both A and C belong to A. A Serre class is a semi-closed class that is closed under
extensions. Furthermore, A is a hereditary torsion class (or localizing class) if it is Serre and it is closed
under taking arbitrary direct sums. On the other hand, A is a hereditary torsion free class provided it
is closed under taking sub-objects, extensions, products and injective envelopes.
Definition 2.8. A hereditary torsion theory τ in C is a pair of classes pT , Fq such that
-- the class T of τ -torsion objects is a hereditary torsion class;
-- the class F of τ -torsion free objects is a hereditary torsion free class;
-- pT qK " F and KpFq " T .
In this paper the symbols τ , T and F are always used to denote a torsion theory, a torsion class and
a torsion free class respectively. Since all the torsion theories in the sequel are hereditary, we just
say "torsion theory", "torsion class" and "torsion free class" to mean respectively "hereditary torsion
theory", "hereditary torsion class" and "hereditary torsion free class".
Given a torsion theory τ " pT , Fq, the inclusion T Ñ C has a right adjoint Tτ : C Ñ T , which is
a subfunctor of the identity, called the torsion functor. For any object X P C there is a short exact
sequence of the form
0 Ñ Tτ X Ñ X Ñ X{Tτ X Ñ 0
such that Tτ X P T and X{Tτ X P F.
Let us also recall the close relation between torsion theories and localization of Abelian categories.
We start with the following definition:
Definition 2.9. A localization of C is a pair of adjoint functors Q : C Õ D : S, where D is an Abelian
category, S is fully faithful, and Q is exact and essentially surjective. In this situation, D is called a
quotient category, Q is a quotient functor and S is a section functor. The composition L " SQ : C Ñ C
is called the localization functor.
One can encounter slightly different definitions of localization in other contexts, see for example
[Ran06]. We can now explain the connection between localizations and torsion theories. Indeed, starting
with a localization Q : C Õ D : S and letting L " S Q,
KerpLq " tX P C : LpXq " 0u " tX P C : QpXq " 0u " KerpQq
is a torsion class (use the exactness of Q and the fact that it is a left adjoint). Hence, the localization
pQ, Sq induces a torsion theory pKerpQq, KerpQqKq. On the other hand, one can construct a localization
out of a torsion theory.
Definition 2.10. Let C be a Grothendieck category and τ " pT , Fq a torsion theory. An object X P C
is τ -local if X P F and EpXq{X P F. The full subcategory of C of all the τ -local objects is denoted by
C{T .
7
The definition of the localization induced by a torsion theory depends on the following lemma.
[Gab62] Let C be a Grothendieck category and let τ " pT , Fq be a torsion theory. Then,
Lemma 2.11.
the canonical inclusion C{T Ñ C has a left adjoint functor which is exact.
Definition 2.12. Let C be a Grothendieck category and let τ " pT , Fq be a torsion theory. The inclusion
Sτ : C{T Ñ C is called τ -section functor, while its left adjoint functor Qτ : C Ñ C{T is called τ -quotient
functor. The composition Lτ " Sτ Qτ is called τ -localization functor.
Let us now recall that the Gabriel filtration of C is a transfinite chain t0u " C´1 Ď C0 Ď Ď Cα Ď . . .
of torsion classes defined as follows:
-- C´1 " t0u;
-- suppose that α is an ordinal for which Cα has already been defined. An object C P C is said to
be α-cocritical if C is τα-torsion free and every proper quotient of C is τα-torsion, where τα is the
unique torsion theory whose torsion class is Cα. In the sequel, we will just say α-torsion (resp.,
torsion free) instead of τα-torsion (resp., torsion free). We let Cα`1 be the smallest hereditary
torsion class containing Cα and all the α-cocritical objects;
-- if λ is a limit ordinal, we let Cλ be the smallest hereditary torsion class containing Ťαăλ
For any ordinal α, we let Tα : C Ñ Cα and Qα : C Ñ C{Cα be respectively the torsion and the quotient
functors. Abusing notation, we use the same symbols for the functors Tα : Cα`1 Ñ Cα and Qα : Cα`1 Ñ
Cα`1{Cα, induced by restriction.
Remark 2.13. By definition, a Grothendieck category C has a generator G. Furthermore, G has just
a set (as opposed to a proper class) of subobject, equivalently, it has just a set of quotients. One
can show that Cα`1 is the smallest torsion class containing Cα and the α-cocritical quotients of G.
As a consequences we obtain that there is an ordinal κ such that Cα " Cκ for all α ě κ (just take
κ " suptα : there are α-cocritical quotients of Gu).
Cα.
Consider the union ¯C "Ťα
Cα of the Gabriel filtration (this makes sense by the above remark). An
object belonging to ¯C is said to be an object with Gabriel dimension. The Gabriel dimension of M is
the minimal ordinal δ such that M P Cδ, in symbols G.dimpM q " δ. In general, it may happen that
C ‰ ¯C; if C " ¯C, we say that C is a Gabriel category with Gabriel dimension G.dimpCq " κ, where κ is
the smallest ordinal such that Cκ " C.
In the following example we specialize some of the above notions to categories of modules. Even
if the example is stated for modules, part (1) holds for any Grothendieck category, taking a generator
in place of RR. Part (2) can be generalized to any Grothendieck category with a set of Noetherian
generators.
Example 2.14. Let R be a ring and let C " R-Mod be the category of left R-modules.
(1) If C has Gabriel dimension G.dimpCq " κ, then G.dimpRRq " κ. In fact, the inequality G.dimpCq ě
G.dimpRRq is trivial. For the converse, just notice that if α is an ordinal for which RR P Cα, then
Cα " C. This is because Cα is closed under arbitrary direct sums and quotients, so it contains any
quotient of a free module, that is, any module.
(2) If R is left Noetherian then C is a Gabriel category.
In fact, given an ordinal α, Cα " Cα`1
implies that RR P Cα. To show this notice that, if α is an ordinal for which RR R Cα, the set
I " tRI ď R : RpR{Iq R Cαu is non-empty and, by Noetherianity, I has a maximal element ¯I.
Then, RpR{ ¯Iq R Cα but any of its proper quotients is in Cα. So, RpR{ ¯Iq is α-cocritic and it belongs
to Cα`1, which therefore properly contains Cα.
In what follows we collect some well-known properties of Gabriel dimension (for a proof see [GR74]).
Recall that a Grothendieck category C is said to be semi-Artinian if G.dimpCq " 0.
Lemma 2.15. Given a Gabriel category C, the following statements hold true:
(1) G.dimpCq " suptG. dimpM q : M P Cu;
(2) if N ď M P C, then G.dimpM q " maxtG.dimpN q, G.dimpM {N qu;
(3) let α be an ordinal ă G.dimpCq and M P C, then M P Cα`1 if and only if there exists an ordinal σ
and a continuous chain 0 " N0 ď N1 ď ď Nσ " M , such that Ni`1{Ni is either α-cocritical or
α-torsion for every i ă σ;
(4) let λ be a limit ordinal ď G.dimpCq and M P C, then M P Cλ if and only if M "Ťαăλ TαpM q;
(5) Cα`1{Cα is semi-Artinian for all α ă G.dimpCq;
(6) let N P C be a Noetherian object, then G.dimpN q is a successor ordinal. Furthermore, there exists
a finite series 0 " N0 ď N1 ď ď Nk " N such that Ni{Ni´1 is cocritical for all i " 1, . . . , k.
8
2.3 Length functions compatible with self-equivalences
All along this section, let C be a Grothendieck category, let L : C Ñ Rě0 Y t8u be a length function and
let F : C Ñ C be a self-equivalence, that is,
(Eq. 1) F is essentially surjective, i.e., for all X P C, there exists Y P C such that F pXq -- Y ;
(Eq. 2) F is fully faithful, i.e., for all X, Y P C, the natural morphism HomCpX, Y q Ñ HomCpF pXq, F pY qq
is an isomorphism.
A consequence of the definition is that F preserves any structure defined by universal properties, in par-
ticular, F commutes with direct and inverse limits and it preserves exactness of sequences. Furthermore,
F commutes with injective envelopes and it preserves lattices of subobjects. It is easily seen that
LF : C Ñ Rě0 Y t8u
such that LF pM q " LpF pM qq ,
(2.4)
for all M P C is a length function. In what follows we are going to study to what extent LF can differ
from L. The following example shows that L and LF may be very different.
Example 2.16. Consider a field K and consider the category K K-Mod -- K-Mod K-Mod. This
category is semi-Artinian and it has a self-equivalence F : K K-Mod Ñ K K-Mod such that pM, N q ÞÑ
pN, M q and pφ, ψq ÞÑ pψ, φq. If we take L to be the length function such that LppM, N qq " dimK pM q,
then clearly LF ppM, 0qq " 0 ‰ dimKpM q " LppM, 0qq, provided M ‰ 0.
Definition 2.17. Given a Grothendieck category C and a self-equivalence F : C Ñ C, we say that a
length function L : C Ñ Rě0 Y t8u is compatible with F provided LF pM q " LpF pM qq " LpM q for all
M P C.
Our motivation for studying compatibility of length functions with self-equivalences is the following:
given a ring R and a ring automorphism φ : R Ñ R, we define a self-equivalence
Fφ : R-Mod Ñ R-Mod ,
(2.5)
which is the scalar restriction along φ, that is: given a left R-module RM such that R acts on M via a
ring homomorphism λ : R Ñ EndZpM q, FφpM q is isomorphic to M as an Abelian group, while R acts
on FφpM q via the ring homomorphism λ φ : R Ñ EndZpFφpM qq. We are interested in finding length
functions L such that LpFφpM qq " LpM q.
2.4 Examples of length functions
Example 2.18. Let D be a left Ore domain, let Σ " Dzt0u and let Q " Σ´1D be the ring of left
fractions of D. For any M P D-Mod, the torsion free rank rkpM q is defined to be the dimension of the
left module Σ´1M " Q bD M over the skew field Q. The torsion free rank rk : D-Mod Ñ Rě0 Y t8u,
where rkpM q " 8 whenever the rank of M is not finite, is a length function. In particular, the dimension
of vector spaces over a field is a length function.
Example 2.19. The logarithm of the cardinality log ´ : Z-Mod Ñ Rě0 Y t8u, where log G " 8
whenever G is not finite, is a length function.
In what follows we define the length of a lattice. This can be applied to the lattice of subobjects
of a given object in a Grothendieck category obtaining the so-called composition length, which is the
invariant inspiring the abstract notion of length function.
Definition 2.20. Let pL, ď, 0, 1q be a bounded lattice. A sequence of elements σ : 0 " x0 ň x1 ň ň
xn " 1 is said to be a series of L. Furthermore, the number of strict inequalities in a series σ, is called
the length of σ, in particular the above series has length n; we write it ℓpσq " n. The length of L is
ℓpLq " suptℓpσq : σ a series of Lu .
Example 2.21. Let C be a Grothendieck category. Notice that a series of LpM q is just a series of M .
We can define a function ℓ : C Ñ Rě0 Y t8u, ℓpM q " ℓpLpM qq. Clearly, ℓpM q " 0 if and only if M " 0,
ℓpM q " 1 if and only if M is a simple object and ℓpM q ă 8 if and only if M is both Noetherian and
Artinian. It is a standard result that ℓ is a length function (see for example [Fac98] for a proof in module
categories).
In the following proposition we introduce a family of length functions that will be extremely useful
for our applications of algebraic entropy:
9
Proposition 2.22. Let C be a Grothendieck category, let α be an ordinal, and let Cα be the α-th layer
in the Gabriel filtration of C. Define an invariant
ℓα : C Ñ Rě0 Y t8u
such that
ℓαpM q :" ℓpLpQαM qq
is the length of the lattice of subobjects of QαM in C{Cα. Then, ℓα is a length function, Kerpℓαq " Cα,
and ℓα is compatible with any self-equivalence of C.
Proof. The fact that ℓα is a length function follows since Qα is an exact left adjoint (so it commutes with
direct limits and it sends short exact sequences to short exact sequences). The fact that Kerpℓαq " Cα
follows since QαpXq " 0 if and only if X P Cα, and the unique object in C{Cα whose composition length
is zero is the 0-object. The fact that ℓα is compatible with any self-equivalence of C follows directly
from the observation that, given X P C, the value ℓαpXq just depends on the lattice of subobjects LpXq,
and any self-equivalence preserves lattices of subobjects. To show this one needs just to use the lattice
theoretic description of Gabriel dimension (for this see [Nv87]), and the lattice theoretic description of
composition length given in Definition 2.20.
3 Amenable groups and crossed products
3.1 Amenable groups
Amenable groups were defined by John von Neumann in 1929 as groups admitting a left-invariant mean.
We adopt here an equivalent definition of amenability given by Følner [Fø55].
Indeed, consider two
subsets A, C Ď G, then
-- the C-interior of A is InCpAq " tx P G : xC Ď Au;
-- the C-exterior of A is OutC pAq " tx P G : xC X A ‰ Hu;
-- the C-boundary of A is BCpAq " OutC pAqzInCpAq.
If e P C, one can imagine the above notions as in the following picture
G
G
G
A
x.
A
x.
A x.
e. C
e. C
e. C
BCpAq
InC pAq
OutC pAq
Definition 3.1. A group G is amenable if and only if there exists a directed set pI, ďq and a net
tFi : i P Iu of finite subsets of G such that, for any finite subset C of G,
lim
I
BCpFiq
Fi
" 0 .
(3.1)
Any such net is called a Følner net.
If G is countable, then one can take I " N and just speak about Følner sequences. Furthermore, given
a countable group G and a Følner sequence s " tFn : n P Nu, we say that s is a Følner exhaustion if
(1) e P F0 and Fn Ď Fn`1 for all n P N;
(2) ŤnPN Fn " G.
It can be proved that a finitely generated group is amenable if and only if it admits a Følner exhaustion
(see for example [Pet13, Lemma 5.3]).
Example 3.2.
(1) Any finite group and any finitely generated Abelian group is amenable;
10
(2) it is known that the class of amenable groups is closed under the operations of taking subgroups,
taking factors over normal subgroups, taking extensions and taking increasing unions. We obtain
that a group G is amenable if and only if all its finitely generated subgroups are amenable, in
particular, arbitrary Abelian groups and locally finite groups are amenable;
(3) consider a finitely generated group G and let S " S´1 be a finite set of generators containing the
unit of G. For all n P N`, let BnpSq " ts1 . . . sn : si P Su and define a function:
fS : N` Ñ N` such that fSpnq " BnpSq .
The group G is said to be of sub-exponential growth if the growth of fS is sub-exponential (it can
be shown that this notion does not depend on the choice of the generating set S). If G is of sub-
exponential growth, then tBnpSq : n P N`u is a Følner exhaustion for G (see for example [CSC10,
Section 6.11]);
(4) non-commutative free groups are not amenable. So, any group which contains a free subgroup of
rank 2 is not amenable. There exist amenable groups with exponential growth.
3.2 Non-negative real functions on finite subsets of an amenable group
In their seminal paper [OW87], Ornstein and Weiss introduced a notion of entropy for actions of amenable
groups on metric spaces. Using the theory of quasi-tiles, they were able to prove that the lim sup defining
their entropy is a true limit. In what follows, we recall some of these deep results as they can be applied
with just minor changes to our algebraic setting. The following terminology and results are due to
Ornstein and Weiss [OW87] (see also [HYZ11] and [WZ92]).
Denote by FpGq the family of all finite subsets of G. Let A1, . . . , Ak P FpGq and ε P p0, 1q. The
family tA1, . . . , Aku is ε-disjoint if there are A1
(1) A1
(2) A1
i Ď Ai and A1
i X A1
j " H if 1 ď i ‰ j ď k.
i{Ai ą 1 ´ ε for i " 1, . . . , k;
1, . . . , A1
k P FpGq such that
Given α P p0, 1s and A P FpGq, tA1, . . . , Aku is an α-cover of A if
Finally, tA1, . . . , Aku ε-quasi-tiles A if there exist C1, . . . , Ck P FpGq such that
A X pŤk
i"1 Aiq
A
ě α .
(1) CiAi Ď A, for all i " 1, . . . , k, and tcAi : c P Ciu forms an ε-disjoint family;
(2) CiAi X Cj Aj " H, if 1 ď i ‰ j ď k;
(3) tCiAi : i " 1, . . . , ku forms a (1 ´ ε)-cover of A.
The subsets C1, . . . , Ck are called tiling centers. It is a deep result, due to Ornstein and Weiss, that
whenever G is an amenable group and tFnunPN is a Følner exhaustion, for any (small enough) ε ą 0, one
can find a nice family of subsets of G that ε-quasi-tiles Fn for all (big enough) n P N. More precisely:
[OW87] Let G be a finitely generated amenable group and let tFnunPN be a Følner
Theorem 3.3.
exhaustion of G. Then, for all ε P p0, 1{4q and ¯n P N, there exist positive integers n1, . . . , nk such that
¯n ď n1 ď ď nk and tFn1 , . . . , Fnk u ε-quasi-tiles Fm, for any big enough m.
In the rest of the subsection we recall some results and terminology about non-negative invariants
for the finite subsets of G, that is, functions f : FpGq Ñ Rě0. In particular, we say that f is
(1) monotone if f pAq ď f pA1q, for all A Ď A1 P FpGq;
(2) sub-additive if f pA Y A1q ď f pAq ` f pA1q, for all A, A1 P FpGq;
(3) (left) G-equivariant if f pgAq " f pAq, for all A P FpGq and g P G.
A consequence of (3) above is that f ptguq " f pteuq, for all g P G. Thus by (2), f pAq ď řgPA f ptguq "
Af peq, for all A P FpGq.
The following result, generally known as "Ornstein-Weiss Lemma", is proved in [OW87] for a suitable
class of locally compact amenable groups (a direct proof, along the same lines, in the discrete case can
be found in [WZ92], while a nice alternative argument, based on ideas of Gromov, is given in [Kri07]).
11
Lemma 3.4. Let G be a finitely generated amenable group and consider a monotone, sub-additive
and G-equivariant function f : FpGq Ñ Rě0. Then, for any Følner sequence tFnunPN, the sequence
pf pFnq{FnqnPN converges and the value of the limit limnPN f pFnq{Fn is the same for any choice of the
Følner sequence.
We conclude this paragraph with the following consequence of Theorem 3.3. The proof is essentially
given in [HYZ11], we give it here for completeness sake, as our statement is slightly different.
Corollary 3.5. Let G be a finitely generated amenable group and tFnunPN be a Følner exhaustion of G.
Then, for any ε P p0, 1{4q and ¯n P N there exist integers n1, . . . , nk such that ¯n ď n1 ď ď nk and, for
any sub-additive and G-equivariant f : FpGq Ñ R we have
lim sup
nÑ8
f pFnq
Fn
ď M ε `
1
1 ´ ε
max
1ďiďk
f pFni q
Fni
,
where M " f pteuq.
It is important to underline that the choice of the n1, . . . , nk does not depend on the function f , but
we can really find a family tn1, . . . , nku, which works for all f at the same time.
Proof. Let ε P p0, 1{4q and ¯n P N. By Theorem 3.3, there exist positive integers n1, . . . , nk such that
¯n ď n1 ď ď nk and tFn1 , . . . , Fnk u ε-quasi-tiles Fn, for all n ě ¯n. We let C 1
k be the tiling
centers for Fn. Thus, when n ě ¯n, we have
1 , . . . , C n
Fn Ě
k
ďi"1
C n
i Fni
and
ě max#p1 ´ εqFn , p1 ´ εq
C n
i Fni + .
k
ÿi"1
Now, let f : FpGq Ñ Rě0 be a sub-additive and G-equivariant function, we obtain that
f pFnq
Fn
ď
`
C n
k
ďi"1
i"1 C n
Fn
i Fni
i Fni¯
f´FnzŤk
i Fni
ď M
FnzŤk
i Fni¯
f´Ťk
i Fni
Ťk
řk
p1 ´ εqřk
i"1 C n
i"1 C n
i"1 C n
Fn
i"1 C n
ď M ε `
ď M ε `
f´Ťk
f´Ťk
Ťk
i"1 C n
Fn
i"1 C n
i"1 C n
i Fni¯
i Fni¯
i Fni
`
i f pFni q
i"1 C n
i Fni
ď M ε `
1
1 ´ ε
max
1ďiďk
f pFni q
Fni
.
3.3 Crossed products
Given a group G and a ring R, a crossed product RG of R with G is a ring constructed as follows: as
a set, RG is the collection of all the formal sums of the form
rgg ,
ÿgPG
with rg P R and rg " 0 for all but finite g P G, and where the g are symbols uniquely assigned to each
g P G. Sum in RG is as expected, that is, it is defined component-wise exploiting the addition in R:
ÿgPG
rgg¸ `ÿgPG
sgg¸ " ÿgPG
prg ` sgqg .
In order to define a product in RG, we need to take two maps
σ : G Ñ AutringpRq and ρ : G G Ñ U pRq ,
where AutringpRq denotes the group of automorphisms of R in the category of (unitary) rings, while
U pRq is the group of units of R. Given g P G and r P R we denote the image of r via the automorphism
σpgq by rσpgq. We suppose also that σ and ρ satisfy the following conditions for all r P R and g1, g2 and
g3 P G:
12
(Cross.1) ρpg1, g2qρpg1g2, g3q " ρpg2, g3qσpg1qρpg1, g2g3q;
(Cross.2) rσpg2qσpg1q " ρpg1, g2qrσpg1g2qρpg1, g2q´1;
(Cross.3) ρpg, eq " ρpe, gq " 1 (for all g P G) and σpeq " 1.
The product in RG is defined by the rule prgqpshq " rsσpgqρpg, hqgh, together with bilinearity, that is
ÿgPG
rgg¸ÿgPG
sgg¸ " ÿgPG ÿh1h2"g
rh1 sσph1q
h2
ρph1, h2q¸ g .
By (Cross.1) and (Cross.2) above, RG is an associative and unitary ring, while by (Cross.3) 1RG " e.
For more details on this kind of construction we refer to [Pas89].
Notice that there is a canonical injective ring homomorphism R Ñ RG such that r ÞÑ re; for this
we identify R with a subring of RG. This allows one to consider a forgetful functor from RG-Mod Ñ
R-Mod. On the other hand, in general there is no natural map G Ñ RG which is compatible with
the operations, anyway the obvious assignment g ÞÑ g respects the operations modulo some units of R.
Thus, given a left RG-module M , there is a canonical map
λ : G Ñ AutZpM q ,
g ÞÑ λg ,
λgpmq " gm ,
(3.2)
which is not in general a homomorphism of groups. Given an R-submodule RK ď M and an element
g P G, λgpKq is again an R-submodule of M but it is not in general isomorphic to RK. As described in
(2.5), there is a self-equivalence of the category R-Mod, induced by the ring automorphism σpgq
Fσpgq : R-Mod Ñ R-Mod .
It follows by the definitions that λgpKq " gK -- FσpgqK.
In particular, if L is a length function
compatible with the equivalence Fσpgq, then LpλgpKqq " LpKq. This useful fact motivates the following
Definition 3.6. Let RG be a crossed product and let L : R-Mod Ñ Rě0 Y t8u be a length function.
Then, L is said to be compatible with RG provided L is compatible with Fσpgq, for all g P G.
4 The algebraic L-entropy
We fix all along this section a ring R, an infinite finitely generated amenable group G, a crossed product
RG and a length function L : R-Mod Ñ Rě0 Y t8u compatible with RG.
Let us remark that, if R is a (skew) field, L " dimR is compatible with any crossed product RG;
more generally, this happens for all the functions ℓα described in Subsection 2.3. On the other hand, if
RG " RrGs, then any length function is trivially compatible with RG.
4.1 Definition and basic properties
In this subsection we define the algebraic entropy as an invariant for left RG-modules. It turns out
that this notion is not well-behaved on all the RG-modules but just on a subclass of RG-Mod of all
the RG-modules M for which the family of submodules FinLpM q :" tRK ď M : LpKq ă 8u is big
enough. More precisely:
Definition 4.1. A left R-module M is said to be locally L-finite if FinLpM q contains all the finitely
generated submodules of M . We denote by lFinpLq the class of all the locally L-finite left R-modules
while we let lFinLpRGq be the class of all the left RG-modules RGM such that RM P lFinpLq, that is
lFinLpRGq :" tM P RG-Mod : LpKq ă 8 for any f.g. RK ď M u .
Notice that lFinpLq is closed under taking direct limits, quotients and submodules but not in general
under taking extensions (see [SVV13]).
Let M P lFinLpRGq. Then, given K P FinLpM q and F P FpGq we let
TF pKq :" ÿgPF
gK .
By the additivity of L and the compatibility of L with RG, LpTF pKqq ďřgPF LpgKq " F LpKq ă 8.
In particular, for any L-finite submodule K of M we obtain a function
fK : FpGq Ñ Rě0
fK pF q :" LpTF pKqq .
We now verify that the above function satisfies the hypotheses of the Ornstein-Weiss Lemma:
13
Lemma 4.2. In the above notation, fK is monotone, sub-additive and G-equivariant.
Proof. Let F, F 1 P FpGq. The proof follows by the properties of L, in fact, TF pKq` TF 1pKq " TF YF 1pKq
and so fK pF q ` fK pF 1q ě fK pF Y F 1q, proving sub-additivity. Furthermore, if F Ď F 1 then TF pKq ď
TF 1 pKq and so fK pF q ď fK pF 1q, proving monotonicity. Finally, for any g P G,
TgF pKq " ÿf PF
gf K " ÿf PF
g f pρpg, f qσpghqq´1K " g ÿf PF
f K " gTF pKq ,
since pρpg, f qσpghqq´1 is a unit of R and so pρpg, f qσpghqq´1K " K. Thus, by the compatibility of L with
RG, LpTgF pKqq " LpgTF pKqq " LpTF pKqq, proving that fK is G-equivariant.
Thus, by the Ornstein-Weiss Lemma, the limit in the following definition exists and it does not
depend on the choice of the Følner sequence.
Definition 4.3. Let M P lFinLpRGq, let tFnunPN be a Følner sequence for G and let RK P FinLpM q.
The L-entropy of RGM with respect to K is
entLpRGM, Kq :" lim
nÑ8
LpTFnpKqq
Fn
.
The L-entropy of the RG-module RGM is entLpRGM q :" suptentLpM, Kq : K P FinLpM qu.
Remark 4.4. We defined the L-entropy for left RG-modules in case G is finitely generated. Anyway,
the exact same procedure allows one to define this invariant when G is just countable (but not necessarily
finitely generated). Furthermore, standard variations of the above arguments using Følner nets allow
one to define a similar invariant in case G is not countable.
Example 4.5. Given RGM P lFinLpRGq such that LpRM q ă 8, entLpRGM q " 0. Indeed, given
K P FinLpM q, entLpM, Kq ď limnÑ8 LpM q{Fn ď limnÑ8 LpM q{n " 0 (use the fact that, as G is
infinite, we can take a Følner sequence such that Fn ň Fn`1 for all n P N, thus Fn ě n).
Given a locally L-finite RG-module RGM and an RG-submodule N ď M , there is an inclusion
FinLpN q Ď FinLpM q, this easily implies the following lemma:
Lemma 4.6. Let RGM P lFinLpRGq and RGN ď M . Then, entLpRGM q ě entLpRGN q.
In fact, entropy is also monotone under taking quotients, but we need some more work before showing
that (see Corollary 4.11). Let us conclude this subsection with the following consequence of Corollary
3.5:
Corollary 4.7. Let tFnunPN be a Følner exhaustion of G. Then, for any ε P p0, 1{4q and n P N there
exist n1, . . . , nk P N with n ď n1 ď ď nk such that, given an L-finite submodule K ď M ,
entLpM, Kq ď ε LpKq `
1
1 ´ ε
max
1ďiďk
LpTFni pKqq
Fni
.
Proof. This is a straightforward application of Corollary 3.5. In fact, the function fK : FpGq Ñ Rě0
such that fK pF q " LpTF pKqq satisfies the hypotheses of such corollary for any L-finite R-submodule K
of M , by Lemma 4.2. Furthermore, limnÑ8 fK pFnq{Fn " entLpM, Kq by the definition of entropy.
4.2 The algebraic entropy is upper continuous
The following result allows us to redefine the algebraic entropy in terms of finitely generated submodules.
Proposition 4.8. Let RGM P lFinLpRGq and let H ď K P FinLpM q.
(1) if LpK{Hq ă ε, then entLpM, Kq ´ entLpM, Hq ă ε;
(2) entLpRGM q " suptentLpM, Kq : RK finitely generatedu.
Proof. For any g P G, gK{gH -- FσpgqpK{Hq so, since L is compatible with Fσpgq, LpgK{gHq ă ε.
Thus, by the additivity of L,
LpTFn pKq{TFnpHqq ď ÿgPFn
LppgK ` TFn pHqq{TFnpHqq ă Fnε ,
for all n P N, where tFnunPN is a Følner sequence. Therefore, entLpM, Kq ´ entLpM, Hq ă ε. We can
now verify part (2). Indeed, the inequality "ď" comes directly from the definition of entropy. On the
other hand, given K P FinLpM q and ε ą 0, by the upper continuity of L there exists H ď K finitely
generated such that LpKq ´ LpHq ă ε. By part (1), entLpM, Kq ă entLpM, Hq ` ε, which easily yields
the claim.
14
The following corollary deals with the case when M is generated (as RG-module) by an L-finite
In such situation one does not need to take a supremum to
R-submodule K, that is, M " TGpKq.
compute entropy.
Corollary 4.9. Let M be a left RG-module such that M " TGpKq for some K P FinLpM q, then
Proof. Given a finitely generated R-submodule H of M , we can find a finite subset e P F Ď G such that
H Ď TF pKq. This shows that entLpM, Hq ď entLpM, TF pKqq. By the Følner condition,
entLpM q " entLpM, Kq .
lim
nÑ8
FnF
Fn
ď lim
nÑ8
Fn YŤf PF BF pFnqf
Fn
ď 1 ` lim
nÑ8 ÿf PF
BF pFnqf
Fn
" 1 .
On the other hand, FnF {Fn ě 1 so limnÑ8 FnF {Fn " 1. We obtain that
entLpM, TF pKqq " lim
nÑ8
TFn pTF pKqq
Fn
" lim
nÑ8
TFnF pKq
Fn
Fn
FnF
" entLpM, Kq ,
where the last equality comes from the fact that tFnF unPN is a Følner sequence (use the fact that, for any
C Ď FpGq and n P N, one has the inclusion BCpFnF q Ď BCF ´1 pFnq and apply the Følner condition for
tFnunPN) and the definition of entL does not depend on the choice of a particular Følner sequence. Thus,
entLpM, Hq ď entLpM, Kq for any finitely generated H ď M ; one concludes by Proposition 4.8.
The upper continuity of entL can now be verified easily using the above lemma and Proposition 4.8:
Corollary 4.10. The invariant entL : lFinLpRGq Ñ Rě0 Y t8u is upper continuous.
Proof. The fact that entL is an invariant can be derived by the definition and the fact that L is an
invariant. Now, let M P lFinLpRGq, then by Proposition 4.8 and Corollary 4.9 we get
entLpRGM q " suptentLpM, Kq : K finitely generated R-submodule of M u
" suptentLpRGpTGpKqq : K finitely generated R-submodule of M u
" suptentLpRGN q : N finitely generated RG-submodule of M u .
Another consequence of Proposition 4.8 is the monotonicity of entropy under taking quotients:
Corollary 4.11. Let RGM P lFinLpRGq and RGN ď M . Then, entLpRGM q ě entLpRGpM {N qq.
Proof. Given a finitely generated submodule ¯K ď M {N , there exists a finitely generated (thus L-finite)
submodule K ď M such that pK ` N q{N -- ¯K. Given a Følner sequence tFnunPN,
TFnp ¯Kq " pTFn pKq ` N q{N and so LpTFn p ¯Kqq ď LpTFnpKqq for all n P N .
Dividing by Fn and passing to the limit we get entLpM {N, ¯Kq ď entLpM, Kq ď entLpRGM q.
We conclude this subsection with the following property that will be extremely useful in proving the
additivity of entropy.
Proposition 4.12. Let M P lFinLpRGq be finitely generated and let tNi : i P Iu be a directed system
of RG-submodules. Letting N :"Ťi Ni, entLpM {N q " inf i entLpM {Niq.
Proof. The inequality "ď" follows by Corollary 4.11. On the other hand, let K P FinLpM q be a finitely
generated R-submodule such that TGpKq " M ; let also ¯K :" pK ` N q{N and ¯Ki :" pK ` Niq{Ni for
all i P I. By Lemma 4.9, entLpM {N q " entLpM {N, ¯Kq and entLpM {Niq " entLpM {Ni, ¯Kiq, for all i.
Now fix (arbitrarily) a constant ε P p0, 1{4q and notice that:
(a) there exists m P N such that
LpTFnp ¯Kqq
Fn
ď entLpM {N q ` ε for all n ě m;
(b) there exist m ď m1 ď ď mk P N such that entLpM {Niq ď εLp ¯Kiq`
for all i P I, by Corollary 4.7;
1
1 ´ ε
max
1ďtďk
LpTFmt p ¯Kiqq
Fmt
,
(c) for any t P t1, . . . , ku there exists jt P I such that LpTFmt p ¯Kjqq ď inf
LpTFmt p ¯Kiqq ` ε "
LpTFmt p ¯Kqq ` ε, whenever Nj ě Njt , where for the equality we use Proposition 2.7.
i
15
Since tNi : i P Iu is directed, there exists ¯j P I such that řk
inf
i
entLpM {Niq ď entLpM {N¯jq
t"1 Njt ď N¯j . Thus,
pbq
ď ε Lp ¯K¯jq `
1
1 ´ ε
pcq
ď ε Lp ¯K¯jq `
1
1 ´ ε
LpTFmt p ¯K¯jqq
max
1ďtďk
Fmt
LpTFmt p ¯Kqq ` ε
Fmt
paq
ď ε Lp ¯K¯jq `
1
1 ´ ε
pentLpM {N q ` 2εq .
As this holds for any small enough ε ą 0, the conclusion follows.
4.3 The algebraic entropy is additive
In this subsection we complete the proof of the fact that
entL : lFinLpRGq Ñ Rě0 Y t8u
is a length function. In fact we have already seen that entL is an upper continuous invariant. To verify
the additivity of entL, by Propositions 2.7 and 4.12, it is enough to check that entL is additive on short
exact sequences of finitely generated modules.
The computations in the following proposition are freely inspired to the proof of the Abramov-Rokhlin
Formula given in [WZ92]. The context (and even the statements) in that paper is quite different but
the ideas contained there can be perfectly adapted to our needs.
Proposition 4.13. Let 0 Ñ N Ñ M Ñ M {N Ñ 0 be a short exact sequence of finitely generated left
RG-modules, then entL pRGM q " entL pRGN q ` entL pRGpM {N qq.
Proof. Fix a Følner exhaustion tFnunPN of G and finitely generated R-submodules K ď M and K1 ď
K XN such that TGpK1q " N and TGpKq " M . If we let ¯K :" pK `N q{N ď M {N , then TGp ¯Kq " M {N .
By Lemma 4.9, entLpM q " entLpM, Kq, entLpN q " entLpN, K1q and entLpM {N q " entLpM {N, ¯Kq.
Let ε P p0, 1{4q, by the existence of the limits defining L-entropy, we can find ¯n P N such that
LpTFn pKqq
Fn
´ entLpM q
ă ε ,
LpTFn pK1qq
Fn
ă ε ,
LpTFnp ¯Kqq
Fn
´ entLpN q
for all n ě ¯n. Moreover, LpTFm pKqq " LpTFm pK1qq`LpTFmpKq{TFm pK1qq ě LpTFm pK1qq`LpTFmp ¯Kqq,
for all m P N, and so, for all m ě ¯n,
ă ε ,
´ entLpM {N q
entLpM q ě
LpTFm pKqq
Fm
´ ε ě
LpTFmpK1qq
Fm
`
LpTFm p ¯Kqq
Fm
´ ε ě entLpN q ` entLpM {N q ´ 3ε ,
yielding the inequality entLpM q ě entLpN q ` entLpM {N q. For the converse inequality, notice that
entLpM q ď
LpTFm pKqq
Fm
` ε "
LpTFmpK1qq
Fm
`
LpTFm pKq{TmpK1qq
Fm
` ε
ď entLpN q `
LpTFm pKq{TmpK1qq
Fm
` 2ε .
To conclude we should verify that pLpTFm pKq{TmpK1qqq{Fm is close enough to entLpM {N q.
Since tFnu is a Følner exhaustion, N "ŤnPN TFnpK1q and so, for any L-finite submodule H ď M , LpH X
N q " limnÑ8 LpH X TFnpK1qq, by upper continuity. Using additivity, LppH ` N q{N q " limnÑ8 LppH `
TFnpK1qq{TFn pK1qq, so that LppH ` N q{N q ´ LppH ` TFnpK1qq{TFnpK1qq ă ε, for any big enough
n P N. By Theorem 3.3, there exist ¯n ă n1 ă ă nk P N and ¯n1 such that tFn1 , . . . , Fnk u ε-quasi-tiles
Fm for all m ě ¯n1. Applying the above argument with H " TFni pKq (for all i " 1, . . . , k), we can find
n ě maxt¯n1, nku such that
for all n ě n and all i " 1, . . . , k. From now on we fix m ě ¯¯n such that BF ´1
pFmq. Thus,
that, given i P t1, . . . , ku, Fni Ď F¯¯n, and so BF ¯¯npFmqFni Ď BF ´1
¯¯n F ¯¯n
¯¯n F ¯¯n
pFmq{Fm ď ε. Notice
¸ ´ LpTFni p ¯Kqq
ă ε
(4.1)
L TFni pKq ` TFnpK1q
TFnpK1q
BF ¯¯npFmqFni {Fm ă ε ,
for all i " 1, . . . , k.
16
We choose tiling centers C1, . . . , Ck for Fm, so that
Consider the following estimate:
k
ďi"1
CiFni
Fm ě
L TFmzŤi CiFni pKq ` TFm pK1q
TFm pK1q
1
Fm
ě max#p1 ´ εqFm , p1 ´ εq
CiFni + ,
k
ÿi"1
¸ ď
ď
Fm
LpTFmzŤi CiFni pKqq
FmzŤi CiFni
Fm
LpKq ď εLpKq ,
Given i P t1, . . . , ku, using that CiFni Ď pCizBFnpFmqqFni Y BFnpFmqFni , we get:
1
Fm
L TCiFni
pKq ` TFm pK 1q
TFm pK 1q
¸ ď
L TcFni
L TFni
pFmq
pFmq
pKq ` TFn
pK 1q
TFn
p ¯Kqq ` Ciε
n
ÿcPCizBF
ÿcPCizBF
L TFni
n
ď
1
Fm
ď
1
Fm
Ci
Fm
ď
ď
CiLpTFni
Fm
pKq ` TFm pK 1q
TFm pK 1q
TBF
n
pFmqFni
pKq ` TFm pK 1q
TFm pK 1q
1
Fm
L
¸ `
¸ `
pKq ` T
c´1Fm pK 1q
BFn
pFmqFni LpKq
c´1Fm pK 1q
T
Fm
pK 1q
¸ ` εLpKq
` εLpKq ď
CiFni
Fm
entLpM {N q ` ε CiFni
Fm
`
Ci
Fm
` LpKq
(4.4)
where the second inequality follows by the compatibility of L with RG, the third one is true since
Fn Ď c´1Fm for all c P CizBFnpFmq, and the fourth one follows by (4.1).
Let us assemble together the above computations:
LpTFm pKq{TFm pK 1qq
Fm
ď
1
Fm
pKq ` TFm pK 1q
¸ `
1
Fm
L TFmzŤi Ci Fni
TFm pK 1q
pKq ` TFm pK 1q
¸
(4.2)
(4.3)
‚
TFm pK 1q
L TŤi Ci Fni
ÿi"1
L TCi Fni
k
1
Fm
pKq ` TFm pK 1q
TFm pK 1q
¸ ` εLpKq
k
ÿi"1 CiFni
Fm
entLpM {N q
entLpM {N q ` ε CiFni
` pk ` 1qLpKq .
1 ´ ε
Fm
` ε 2
1 ´ ε
(4.3)
ď
(4.4)
ď
ď
`
Ci
Fm
` LpKq ` εLpKq
Hence, we have obtained that
entLpM q ď entLpN q `
entLpM {N q
1 ´ ε
` ε 2
1 ´ ε
` pk ` 1qLpKq ,
holds for any ε P p0, 1{4q, and so entLpM q ď entLpN q ` entLpM {N q, concluding the proof.
4.4 Values on (sub)shifts
This subsection is devoted to compute the values of the entropy on the left RG-modules of the form
M " RG bR K, for some left R-module K, and their RG-submodules. This will conclude the proof
of Theorem B.
Definition 4.14. A left RG-module of the form M " RG bR K is said to be a Bernoulli shift while
any of its RG-submodules is a subshift.
Notice that, given a Bernoulli shift M " RG bR K, there is a direct sum decomposition, as a left
R-module, RM -- ÀgPG gK. Thus, for any F P FpGq,
TF pgKq " àhPF
In the following example we compute the algebraic entropy of Bernoulli shifts.
hgK and
TGpgKq " M .
(4.5)
17
Example 4.15. In the above notation, suppose LpKq ă 8. By Lemma 4.9 and (4.5), we obtain that
entLpRGM q " entLpM, Kq. Furthermore, again by (4.5), LpTF pKqq{F " LpKq, for all F P FpGq.
Therefore, entLpM q " LpKq.
The computation in the above example shows that the entropy of M " RG bR K is 0 if and only
if LpKq " 0, if and only if LpM q " 0. Our next goal is to show that, if RGN is a subshift of M , then
entLpRGN q " 0 if and only if LpRN q " 0. This will be proved in Proposition 4.18 but first we need to
recall some useful terminology and results from [CSC06].
Definition 4.16. Let E and F be subsets of G. A subset N Ď G is an pE, F q-net if it satisfies the
following conditions:
(1) the subsets pgEqgPN are pairwise disjoint, that is, gE X g1E " H for all g ‰ g1 P N ;
(2) G "ŤgPN gF .
The following lemma is a variation of [CSC06, Lemmas 2.2 and 4.3].
Lemma 4.17. Let tFnunPN be a Følner sequence, E Ď F Ď G be finite subsets with e P F and N an
pE, F q-net. Then,
(1) there exists an pE, EE´1q-net;
(2) there exist α P p0, 1s and n0 P N such that Fn X N ě α Fn, for all n ą n0.
Proof. Part (1) is proved in [CSC06, Lemma 2.2], so let us concentrate on part (2). For each n P N, let
F `F
n zpFn X N q Ď BF pFnq. Furthermore, since Fn is covered by
the sets gF , g P F `F
n " OutF pFnq X N and notice that F `F
n , we have Fn ď F F `F
n . Let now α1 " 1{F , thus
α1Fn ´ Fn X N ď F `F
n ´ Fn X N ď F `F
n zpFn X N q ď BF pFnq .
Let α2 P p0, α1q. By the Følner condition, there exists n0 P N such that BF pFnq{Fn ď α2 for all
n ą n0. Thus, letting α " α1 ´ α2 P p0, 1s, we get Fn X N ě α1Fn ´ BF pFnq ě αFn, for all
n ą n0.
Proposition 4.18. Let K be an L-finite left R-module and RGN a subshift of M " RG bR K. Then,
entLpRGN q " 0
if and only if LpRN q " 0 .
Proof. Suppose LpRN q ‰ 0, then there exists x P N such that LpRxq ‰ 0. Let E be the set of all
elements h P G such that, writing x " řgPG gxg, the component xh is not 0. We fix an pE, EE´1q-net
N . Notice that, given f1 ‰ f2 P N , then βf1 pRxq X βf2 pRxq " 0. Thus, by Lemma 4.17, we can find
n0 P N and α P p0, 1q such that
LpTFn pRxqq ě LpTFnXN pRxqq " Fn X N LpRxq ě αFnLpRxq
for all n ą n0. In particular, entLpRGN q ě entLpN, Rxq ě αLpRxq ‰ 0.
5 Applications
5.1 Stable finiteness
Let R be a ring. Recall from the introduction that a left R-module M is said to be hopfian if any of
its surjective endomorphisms is bijective. Recall also that a ring R is directly finite if xy " 1 implies
yx " 1 for all x, y P R. Furthermore, R is stably finite if the ring MatkpRq of k k square matrices with
coefficients in R, is directly finite for all k P N`.
The following lemma gives a natural connection between stable finiteness and Hopficity:
Lemma 5.1. Given a ring R and a positive integer k, the following are equivalent:
(1) Rk
R is hopfian (as a right R-module);
(2) RRk is hopfian (as a left R-module);
(3) MatkpRq is directly finite.
18
As we said, our interest in Hopficity is motivated by the Stable Finiteness Conjecture, stating that
for any field K and any group G, the group ring KrGs is stably finite. In fact, a fairly general case of
this conjecture was recently verified by Elek and Szab´o [ES04] that proved that KrGs is stably finite
for any division ring K and any sofic group G. A straightforward consequence is that, under these
hypotheses, MatnpKrGsq is stably finite. Now, by the Artin-Wedderburn Theorem, given a semisimple
Artinian ring R, there exist positive integers k, n1, . . . , nk P N` and division rings K1, . . . , Kk such that
R -- Matn1 pK1q Matnk pKkq. This implies that, RrGs -- Matn1 pK1rGsq Matnk pKkrGsq,
thus a consequence of the above theorem is that RrGs is stably finite whenever R is semisimple Artinian
and G sofic. This result can be further generalized as follows:
Remark 5.2. [Ferran Ced´o, private communication (2012)] If R is a ring with left Krull dimension
(for example see [MR01]) and G is sofic, then RrGs is stably finite. First of all, notice that, if I is a
nilpotent ideal of R, then IrGs " RrGsI is a nilpotent ideal of RrGs and so one can reduce the problem
modulo nilpotent ideals. Now, by [MR01, Corollary 6.3.8], the prime radical N of R is nilpotent and
N " P1 X XPm, where P1, . . . , Pm are minimal prime ideals. Thus, by [MR01, Proposition 6.3.5], R{N
is a semiprime Goldie ring and so, by [MR01, Theorem 2.3.6] R{N has a classical semisimple Artinian
ring of quotients S. In particular, pR{N qrGs embeds in SrGs and it is therefore stably finite.
Both the proof of the residually amenable case due to Ara, O'Meara and Perera, and the proof of
the sofic case due to Elek and Szab´o, consist in finding a suitable embedding of KrGs in a ring which is
known to be stably finite. Such methods are really effective but, as far as we know, cannot be used to
obtain information on the modules over KrGs. It seems natural to ask the following
Question 5.3. Let G be a group, R a ring and let RG be a fixed crossed product. Given a finitely
generated left RG-module RGM , when is RGM an hopfian module?
Using the theory of algebraic entropy we can now prove that a large class of left RG-modules is
hereditarily Hopfian (i.e., any submodule is Hopfian), in case R is left Noetherian and G amenable
(see the statement of Theorem A). We remark that this is a very strong version of Kaplasky's Stable
Finiteness Conjecture in the amenable case, which can be re-obtained as a corollary. The proof of
Theorem A makes use of the full force of the localization techniques introduced in Section 2. Such heavy
machinery hides in some sense the idea behind the proof; this is the reason for which we prefer to give
first the proof of the following more elementary statement, whose proof is far more transparent:
Lemma 5.4. Let K be a division ring, let G be a finitely generated amenable group and fix a crossed
product KG. For all n P N`, pKGqn " KG b Kn is a hereditarily hopfian left KG-module.
Proof. Let n P N` and choose KG-submodules N ď M ď pKGqn such that there exists a short exact
sequence
0 Ñ N Ñ M Ñ M Ñ 0 ,
we have to show that N " 0. The length function dim : K-Mod Ñ Rě0 Y t8u is compatible with any
crossed product, so we can consider the dim-entropy of left KG-modules. In particular,
entdimpM q " entdimpM q ` entdimpN q
and
0 ď entdimpN q ď entdimpM q ď entdimpKG b Knq " n .
Thus, entdimpN q " 0. By Proposition 4.18, this implies that dimpN q " 0, that is, N " 0.
The same argument of the above proof can be used to prove Theorem A, modulo the fundamental
tool of Gabriel dimension:
Proof of Theorem A. Consider a left RG-submodule M ď RG bR K and a short exact sequence of
left RG-modules
φ
0 Ñ Kerpφq Ñ M
ÝÑ M Ñ 0 .
In order to go further with the proof we need to show that, as a left R-module, the Gabriel dimension
of Kerpφq is a successor ordinal whenever it is not ´1 (i.e., whenever Kerpφq ‰ 0). This follows by the
following
Lemma 5.5. In the hypotheses of Theorem A, G.dimpRN q is a successor ordinal for any non-trivial
R-submodule N ď RG bR K.
19
Proof. A consequence of Lemma 2.15 (6) is that Tα`1pKq{TαpKq ‰ 0 for just finitely many
ordinals α. Notice that TαpRG bR Kq -- RG bR TαpKq, as left R-modules, for any ordinal
α. Thus, Tα`1pRG bR Kq{TαpRG bR Kq ‰ 0 for finitely many ordinals. Notice also that
TαpN q " TαpRG bR Kq X N for all α, thus,
Tα`1pN q
TαpN q
"
Tα`1pRG bR Kq X N
TαpRG bR Kq X N
--
--
pTα`1pRG bR Kq X N q ` TαpRG bR Kq
TαpRG bR Kq
ď
Tα`1pRG bR Kq
TαpRG bR Kq
is different from zero for finitely many ordinals α. Thus,
G.dimpN q " suptα ` 1 : Tα`1pN q{TαpN q ‰ 0u " maxtα ` 1 : Tα`1pN q{TαpN q ‰ 0u
is clearly a successor ordinal.
Now, suppose that Kerpφq ‰ 0 and let G.dimpKerpφqq " α ` 1. We want to show that
φaeTα`1pM q : Tα`1pM q Ñ Tα`1pM q
is surjective. Indeed, if there is x P Tα`1pM qzφpTα`1pM qq, it means that there exists y P M zTα`1pM q
such that φpyq " x (by the surjectivity of φ). This is to say that there is a short exact sequence
0 Ñ Kerpφq X RGy Ñ RGy Ñ RGx Ñ 0 ,
with G.dimpRpRGyqq ŋ α ` 1 ě maxtG.dimpRpKerpφq X RGyqq, G.dimpRpRGxqqu, which contradicts
Lemma 2.15(4). Thus, we have a short exact sequence of left RG-modules
0 Ñ Kerpφq Ñ Tα`1pM q Ñ Tα`1pM q Ñ 0 .
Consider the length function ℓα : R-Mod Ñ Rě0 Y t8u described in Subsection 2.3 and recall that
Kerpℓαq is exactly the class of all left R-modules with Gabriel dimension ď α. Furthermore, Tα`1pKq
is a Noetherian module, thus, QαpTα`1pKqq is a Noetherian object in a semi-Artinian category, that is,
an object with finite composition length, for this reason ℓαpTα`1pKqq " ℓpQαpTα`1pKqqq ă 8. Using
the computations of Example 4.15 and the Addition Theorem, we get
entℓα pTα`1pRGbKqqq " ℓαpTα`1pKqq ă 8 and entℓα pTα`1pM qq " entℓα pTα`1pM qq`entℓα pKerpφqq .
Hence, entℓα pKerpφqq " 0 which, by Proposition 4.18, is equivalent to say that ℓαpKerpφqq " 0, contra-
dicting the fact that G.dimpKerpφqq " α ` 1.
In the above proof we made use of the Addition Theorem for the algebraic entropy, which is quite
a deep result. We want to underline that if one is only interested in the second part of the statement,
that is, stable finiteness of endomorphism rings, then it is sufficient to use the weaker additivity of the
algebraic entropy on direct sums, which can be independently verified as an exercise.
Example 5.6. Let G be a free group of rank ě 2 and let K be a field. It is well-known that KrGs is
not left (nor right) Noetherian so we can find a left ideal KrGsI ď KrGs which is not finitely generated.
Furthermore, by [Coh06, Corollary 7.11.8], KrGs is a free ideal ring, so I is free. This means that I is
isomorphic to a coproduct of the form KrGspNq which is obviously not hopfian.
Let us conclude this subsection with the following problem:
Problem 5.7. Study the class of finitely generated groups G such that the group algebra KrGs is heredi-
tarily Hopfian for any skew field K. Is it true that this property characterizes the class of finitely generated
amenable groups?
One could also state an analogous problem including all the possible crossed products KG, instead
of just the group algebras KrGs.
20
5.2 Zero-Divisors
In this last section of the paper we discuss another classical conjecture due to Kaplansky about group
rings connecting it to the theory of algebraic entropy:
Conjecture 5.8 (Kaplansky). Let K be a field and G be a torsion-free group. Then KrGs is a domain.
Some cases of the above conjecture are known to be true but the conjecture is fairly open in general
(for a classical reference on this conjecture see for example [Pas85]). In most of the known cases, the
strategy for the proof is to find an immersion of KrGs in some division ring. This is clearly sufficient
but, in principle, it is a stronger property. To the best of the author's knowledge, the following question
remains open: Is it true that KrGs is a domain if and only if KrGs is a subring of a division ring?
The above question is known to have positive answer if G is amenable (see [Luc02, Example 8.16]). In
this section we provide an alternative argument to answer the above question for amenable groups (in
the more general setting of crossed group rings) and we translate the amenable case of Conjecture 5.8
into an equivalent statement about algebraic entropy. This approach is inspired to the work of Nhan-Phu
Chung and Andreas Thom [CT13]. Indeed, we can prove the following
Theorem 5.9. Let K be a division ring and let G be a finitely generated amenable group. For any fixed
crossed product KG, the following are equivalent:
(1) KG is a left (and right) Ore domain;
(2) KG is a domain;
(3) entdimpKGM q " 0, for every proper quotient M of KG;
(4) Impentdimq " N Y t8u.
Before proving the above theorem we recall some useful properties about Ore domains. We start
recalling that a domain D is left Ore if Dx X Dy ‰ t0u for all x, y P Dzt0u. It can be shown that this is
equivalent to say that D is a left flat subring of a division ring.
Proposition 5.10. A domain D is left Ore if and only if there is a length function L : D-Mod Ñ
Rě0 Y t8u such that LpDq " 1.
Proof. If D is left Ore, then D is a flat subring of a division ring K. Then there is an exact functor
K bD ´ : D-Mod Ñ K-Mod which commutes with direct limits. Thus, we can define the desired length
function L simply letting LpDM q :" dimKpK bD M q.
On the other hand, suppose that there is a length function L : D-Mod Ñ Rě0 Y t8u such that LpDq " 1
and choose x, y P Dzt0u. Since D is a domain, both Dx and Dy contain (and are contained in) a copy
of D, thus LpDxq " LpDyq " 1. If, looking for a contradiction Dx X Dy " t0u, then
1 " LpDq ě LpDx ` Dyq " LpDx ' Dyq " LpDxq ` LpDyq " 2,
which is a contradiction.
It is a classical result that any left Noetherian domain is left Ore (see for example [MR01, Theorem
1.15 in Chapter 2.1]). By the above proposition we can generalize this result as follows:
Corollary 5.11. A domain with left Gabriel dimension is necessarily left Ore.
Proof. Let D be a domain with left Gabriel dimension. First of all we verify that G.dimpDDq is not a
limit ordinal. Indeed, if G.dimpDDq " λ is a limit ordinal, then D " Ťαăλ TαpDq. This means that,
for any non-zero x P D, there exists α ă λ such that Dx P TαpDq. Choose a non-zero x P D, as D
is a domain, there is a copy of D inside Dx. Thus, G.dimpDq ď G.dimpDxq ď α for some α ă λ, a
contradiction.
If G.dimpDDq " α ` 1 for some ordinal α, then we can consider the length function
ℓα : D-Mod Ñ Rě0 Y t8u ,
ℓαpM q " ℓpQαpM qq .
To conclude one has to show that ℓαpDq " 1, that is, QαpDq is a simple object. Since Cα`1{Cα is semi-
Artinian, there is a simple subobject S of QαpDq. Then SαpSq is a sub-module of SαQαpDq. Identify
SαpSq, SαQαpDq and D with submodules of EpDq, since D is essential in EpDq, there is 0 ‰ x such
that x P SαpSq X D, but then SαpSq contains an isomorphic copy of D. Thus QαSαpSq " S contains
an isomorphic copy of QαpDq, which is therefore simple.
We can finally prove our result:
21
Proof of Theorem 5.9. (1)ñ(2) is trivial while (2)ñ(1) follows by Proposition 5.10 and the fact that
the algebraic dim-entropy is a length function on KG-Mod such that entdimpKGKGq " 1.
(2)ñ(3). Consider a short exact sequence 0 Ñ KGI Ñ KGKG Ñ KGM Ñ 0, with I ‰ 0. Choose
0 ‰ x P I, then KGx -- KG, and so entdimpKGM q " entdimpKGKGq ´ entdimpKGIq ď 1 ´ 1 " 0.
(3)ñ(4). Let us show first that for any finitely generate left KG-module KGF , entdimpKGF q P N. In
fact, choose a finite set of generators x1, . . . , xn for F and, letting F0 " 0 and Fi " KGx1 ` ` KGxi
for all i " 1, . . . , n, consider the filtration 0 Ď F1 Ď F2 Ď Ď Fn " F . By additivity,
entdimpF q "
n
ÿi"1
entdimpFi{Fi´1q .
All the modules F1{F0, . . . , Fn{Fn´1 are cyclics (i.e. quotients of KG), thus entdimpFi{Fi´1q P t0, 1u
by hypothesis. Hence, entdimpF q P N. To conclude one argues by upper continuity that the algebraic
dim-entropy of an arbitrary left KG-module is the supremum of a subset of N, thus it belongs to
N Y t8u.
(4)ñ(2). Let x P KG and consider the short exact sequence
0 Ñ I Ñ KG Ñ KGx Ñ 0
where I " ty P KG : yx " 0u. Suppose that x is a zero-divisor, that is, I ‰ 0 or, equivalently,
dimpKIq ‰ 0. By Proposition 4.18, entdimpIq ą 0 and, by our assumption, entLpIq ě 1. Hence, using
additivity, entdimpKGxq " 0. Again by Proposition 4.18, this implies dimpKGxq " 0 and consequently
KGx " 0, that is, x " 0. Thus, the unique zero-divisor in KG is 0.
References
[AOP02] Pere Ara, Kevin C. O'Meara, and Francesc Perera. Stable finiteness of group rings in arbitrary
characteristic. Adv. Math., 170(2):224 -- 238, 2002.
[Coh06] P. M. Cohn. Free ideal rings and localization in general rings, volume 3 of New Mathematical
Monographs. Cambridge University Press, Cambridge, 2006.
[CSC06] Tullio Ceccherini-Silberstein and Michel Coornaert. The Garden of Eden theorem for linear
cellular automata. Ergodic Theory Dynam. Systems, 26(1):53 -- 68, 2006.
[CSC10] Tullio Ceccherini-Silberstein and Michel Coornaert. Cellular automata and groups. Springer
Monographs in Mathematics. Springer-Verlag, Berlin, 2010.
[CT13] Nhan-Phu Chung and Andreas Thom. Some remarks on the entropy for algebraic actions of
amenable groups. http://arxiv.org/abs/1302.5813, 2013.
[Ele99] G´abor Elek. The Euler characteristic of discrete groups and Yuzvinskii's entropy addition
formula. Bull. London Math. Soc., 31(6):661 -- 664, 1999.
[Ele03] G´abor Elek. The rank of finitely generated modules over group algebras. Proc. Amer. Math.
Soc., 131(11):3477 -- 3485 (electronic), 2003.
[ES04] G´abor Elek and Endre Szab´o. Sofic groups and direct finiteness. J. Algebra, 280(2):426 -- 434,
2004.
[Fac98] Alberto Facchini. Module theory. Modern Birkhauser Classics. Birkhauser/Springer Basel AG,
Basel, 1998. Endomorphism rings and direct sum decompositions in some classes of modules,
[2012 reprint of the 1998 original] [MR1634015].
[Fø55] Erling Følner. On groups with full Banach mean value. Math. Scand., 3:243 -- 254, 1955.
[Gab62] Pierre Gabriel. Des cat´egories ab´eliennes. Bull. Soc. Math. France, 90:323 -- 448, 1962.
[GR74] Robert Gordon and J. Chris Robson. The Gabriel dimension of a module. J. Algebra, 29:459 --
473, 1974.
[HYZ11] Wen Huang, Xiangdong Ye, and Guohua Zhang. Local entropy theory for a countable discrete
amenable group action. J. Funct. Anal., 261(4):1028 -- 1082, 2011.
[Kap95] Irving Kaplansky. Fields and rings. Chicago Lectures in Mathematics. University of Chicago
Press, Chicago, IL, 1995. Reprint of the second (1972) edition.
22
[Kri07] Fabrice Krieger. Le lemme d'Ornstein-Weiss d'apr`es Gromov.
In Dynamics, ergodic theory,
and geometry, volume 54 of Math. Sci. Res. Inst. Publ., pages 99 -- 111. Cambridge Univ. Press,
Cambridge, 2007.
[LL15] Hanfeng
Li
and
Bingbing
Liang.
Sofic
mean
length.
arXiv:
https: // arxiv. org/ abs/ 1510. 07655 , 2015.
[Luc02] Wolfgang Luck. L2-invariants: theory and applications to geometry and K-theory, volume 44
of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in
Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys
in Mathematics]. Springer-Verlag, Berlin, 2002.
[MR01] J. C. McConnell and J. C. Robson. Noncommutative Noetherian rings, volume 30 of Graduate
Studies in Mathematics. American Mathematical Society, Providence, RI, revised edition, 2001.
With the cooperation of L. W. Small.
[NR65] D. G. Northcott and M. Reufel. A generalization of the concept of length. Quart. J. Math.
Oxford Ser. (2), 16:297 -- 321, 1965.
[Nv87] Constantin Nastasescu and Freddy van Oystaeyen. Dimensions of ring theory. Mathematics and
its Applications, 36. Dordrecht etc.: D. Reidel Publishing Company, a member of the Kluwer
Academic Publishers Group. XI, 360 p.;(1987)., 1987.
[OW87] Donald S. Ornstein and Benjamin Weiss. Entropy and isomorphism theorems for actions of
amenable groups. J. Analyse Math., 48:1 -- 141, 1987.
[Pas85] Donald S. Passman. The algebraic structure of group rings. Robert E. Krieger Publishing Co.
Inc., Melbourne, FL, 1985. Reprint of the 1977 original.
[Pas89] Donald S. Passman. Infinite crossed products, volume 135 of Pure and Applied Mathematics.
Academic Press Inc., Boston, MA, 1989.
[Pet13] G´abor Pete. Probability and geometry on groups. http://www.math.bme.hu/gabor/PGG.pdf,
May 2013.
[Ran06] Andrew Ranicki, editor. Non-commutative localization in algebra and topology, volume 330
of London Mathematical Society Lecture Note Series, Cambridge, 2006. Cambridge University
Press.
[Ste71] Bo Stenstrom. Rings and modules of quotients. Lecture Notes in Mathematics, Vol. 237. Springer-
Verlag, Berlin, 1971.
[SVV13] Luigi Salce, Peter V´amos, and Simone Virili. Length functions, multiplicities and algebraic
entropy. Forum Math., 25:255 -- 282, 2013.
[V´am68] Peter V´amos. Length Functions on Modules. PhD thesis, University of Sheffield, 1968.
[WZ92] Thomas Ward and Qing Zhang. The Abramov-Rokhlin entropy addition formula for amenable
group actions. Monatsh. Math., 114(3-4):317 -- 329, 1992.
Address of the author:
Simone Virili - [email protected]
Facultad de matem´aticas, Universidad de Murcia, Campus Espinardo, 30100 Murcia, Spain.
23
|
1505.07123 | 1 | 1505 | 2015-05-26T20:14:19 | Inverse semigroups associated with labelled spaces and their tight spectra | [
"math.RA",
"math.OA"
] | The notion of a labelled space was introduced by Bates and Pask in generalizing certain classes of C*-algebras. Motivated by Exel's work on inverse semigroups and combinatorial C*-algebras, we associate each weakly left resolving labelled space with an inverse semigroup, and characterize the tight spectrum of the latter in a way that is reminiscent of the description of the boundary path space of a directed graph. | math.RA | math |
Inverse semigroups associated with
labelled spaces and their tight spectra
Giuliano Boava, Gilles G. de Castro, and Fernando de L. Mortari
Abstract. The notion of a labelled space was introduced by
Bates and Pask in generalizing certain classes of C*-algebras. Mo-
tivated by Exel's work on inverse semigroups and combinatorial
C*-algebras, we associate each weakly left resolving labelled space
with an inverse semigroup, and characterize the tight spectrum of
the latter in a way that is reminiscent of the description of the
boundary path space of a directed graph.
1. Introduction
When studying C*-algebras, it is often useful do describe them as
C*-algebras associated with groupoids [11]. For certain classes of C*-
algebras, such descriptions have been achieved with the assistance of
inverse semigroups, most notably with the work of Paterson on graph
C*-algebras [10] and of others inspired by his approach [7, 8, 9]. The
desired groupoid appears there as the groupoid of germs for an action
of an inverse semigroup (defined from the original object considered,
e.g. a directed graph in [10]) on a certain topological space. A natural
candidate for a topological space for the inverse semigroup to act on
is the set of characters defined on the idempotents of the semigroup;
however, the C*-algebra constructed from the groupoid of germs that
comes from this action is not usually isomorphic to the original C*-
algebra.
What was done in the examples studied to somehow fix this is
to restrict the space of characters using information from the original
object; one of the main goals of Exel in [7] is to investigate whether this
restriction can be done using only information contained in the inverse
2010 Mathematics Subject Classification. Primary: 20M18, Secondary: 05C20,
05C78.
Key words and phrases. Inverse semigroup, tight spectrum, labelled space.
1
2
G. BOAVA, G. DE CASTRO, AND F. MORTARI
semigroup, without returning to the original object. He showed that it
is possible to do so, by considering only those characters that are tight,
that is, using the tight spectrum of the idempotent semilattice.
Many of the C*-algebras studied in the above-mentioned work can
be realized as C*-algebras associated with labelled graphs (or more
precisely, labelled spaces), as defined by Bates and Pask [2] and further
studied with collaborators [1, 3, 4]. Examples include C*-algebras
associated with ultragraphs and the Matsumoto algebras of (two-sided)
subshifts [2], as well as the Carlsen-Matsumoto algebra of a one-sided
shift [1].
The main goal of this paper is taking the first step in bringing
inverse semigroup theory to the study of labelled spaces in general.
Motivated by the inverse semigroup treatment given to graph, ultra-
graph and higher rank graph C*-algebras in [7, 8, 9, 10], we construct
an inverse semigroup from a labelled space that is weakly left resolving,
and proceed to give a description of the tight spectrum for this inverse
semigroup. This is achieved through a study and complete characteri-
zation of filters in the idempotent semilattice of this inverse semigroup.
The resulting description resembles the characterization of the bound-
ary path space of a directed graph as the set consisting of all infinite
paths on the graph, and of all finite paths that end in a singular ver-
tex [13]; indeed we show that, for certain classes of labelled spaces,
we arrive exactly at the boundary path space of the graph, but that
this may not always be the case. Although motivated by the theory of
C*-algebras, we work strictly in the realm of inverse semigroup theory.
This paper is organized as follows: Section 2 is a review of some of
the necessary terminology and results on labelled spaces and filters and
characters in semilattices. Section 3 deals with the construction of an
inverse semigroup associated with a labelled space. Sections 4, 5 and 6
give a description of the filters, ultrafilters and tight filters, respectively,
on the semilattice of idempotents of the inverse semigroup defined in
Section 3, and examples are studied.
2. Preliminaries
2.1. Labelled graphs.
Definition 2.1. A (directed) graph E = (E0, E1, r, s) consists of
nonempty countable sets E0, E1 and functions r, s : E1 → E0; an el-
ement of E0 is called a vertex of the graph, and an element of E1 is
called an edge. For an edge e, we say that r(e) is the range of e and
s(e) is the source of e.
INVERSE SEMIGROUPS AND LABELLED SPACES
3
Definition 2.2. A vertex v in a graph E is called a source if
r−1(v) = ∅, a sink if s−1(v) = ∅, and an infinite emitter if s−1(v)
is an infinite set. The vertex v is singular if it is either a sink or an
infinite emitter, and regular otherwise.
Definition 2.3. A path of length n in a graph E is a sequence
λ = λ1λ2 . . . λn of edges such that r(λi) = s(λi+1) for all i = 1, . . . , n−1.
We write λ = n for the length of λ and regard vertices as paths
of length 0. We denote by En the set of all paths of length n and
E∗ = ∪n≥0En. We extend the range and source maps to E∗ by defining
s(λ) = s(λ1) and r(λ) = r(λn) if n ≥ 2 and s(v) = v = r(v) for n = 0.
Similarly, we define a path of infinite length (or an infinite path) as an
infinite sequence λ = λ1λ2 . . . of edges such that r(λi) = s(λi+1) for all
i ≥ 1. The set of all infinite paths will be denoted by E∞.
Definition 2.4. Given a nonempty set A, called an alphabet, and
whose elements are referred to as letters, a labelled graph (E, L) over the
alphabet A consists of a graph E = (E0, E1, r, s) and a map L : E1 → A,
called the labelling map.
Without loss of generality, we may assume that L is onto.
Let A∗ be the set of all finite words over A, including the empty
word ω, and extend the map L to a map L : En → A∗ defined by
L(λ) = L(λ1) . . . L(λn). Similarly, one can define the map L : E∞ →
A∞, where A∞ is the set of infinite countable words over A.
Given A ⊆ P(E0) (where P(E0) is the power set of E0) define
(2.1)
L(AE1) = {L(e) e ∈ E1 and s(e) ∈ A}.
Definition 2.5. The elements of Ln := L(En) are called labelled
paths of length n and the elements of L∞ := L(E∞) are called labelled
paths of infinite length. If α is a labelled path, a path on the graph
λ such that L(λ) = α is called a representative of α. Since any two
representatives of α have the same length, one can define the length of
α, denoted by α, as the length of any one of its representatives. We
also consider ω as a labelled path, with ω = 0. If 1 ≤ i ≤ j ≤ α,
let αi,j = αiαi+1 . . . αj if j < ∞ and αi,j = αiαi+1 . . . if j = ∞.
If
j < i set αi,j = ω. The set L≥1 = ∪n≥1Ln is the set of all labelled
paths of positive finite length. We also define L∗ = {ω} ∪ L≥1, and
L≤∞ = L∗ ∪ L∞.
Since labelled paths are words over an alphabet, we can concatenate
two labelled paths α and β to obtain a new word αβ. However, it is
not always true that αβ is also a labelled path, for there may not be a
path on the graph whose label is αβ.
4
G. BOAVA, G. DE CASTRO, AND F. MORTARI
Definition 2.6. Given two labelled paths α, β, we say that α is a
beginning of β if β = αβ′ for some labelled path β′. We say that α and
β are comparable if α is a a beginning of β or β is a beginning of α.
Definition 2.7. For α ∈ L∗ and A ∈ P(E0), the relative range of
α with respect to A, denoted by r(A, α), is the set
r(A, α) = {r(λ) λ ∈ E∗, L(λ) = α, s(λ) ∈ A}
if α ∈ L≥1 and r(A, ω) = A if α = ω. The range of α, denoted by r(α),
is the set
r(α) = r(E0, α).
For α ∈ L≥1 we also define the source of α as the set
s(α) = {s(λ) ∈ E0 L(λ) = α}.
Observe that, in particular, r(ω) = E0, and if α ∈ L≥1 then r(α) =
{r(λ) ∈ E0 L(λ) = α}. The definitions above give maps s, r : L≥1 →
P(E0). Also, if α, β ∈ L∗ are such that αβ ∈ L∗ then r(r(A, α), β) =
r(A, αβ). Finally, for A, B ∈ P(E0) and α ∈ L∗, it holds that r(A ∪
B, α) = r(A, α) ∪ r(B, α).
The following notions were introduced by Bates and Pask in [2].
Definition 2.8. Let (E, L) be a labelled graph and B ⊆ P(E0).
We say that B is closed under relative ranges if r(A, α) ∈ B for all
A ∈ B and all α ∈ L∗. If additionally B is closed under finite inter-
sections and unions and contains all r(α) for α ∈ L≥1, we say that B
is accommodating for (E, L) and in this case we say that (E, L, B) is a
labelled space.
Definition 2.9. We say that a labelled space (E, L, B) is weakly
left resolving if for all A, B ∈ B and all α ∈ L≥1 we have r(A ∩ B, α) =
r(A, α) ∩ r(B, α).
Observe that in a weakly left resolving labelled space, if A, B ∈ B
are disjoint then for all α ∈ L∗, r(A, α) and r(B, α) are disjoint.
Definition 2.10. For a given α ∈ L∗, let Bα := B ∩ P(r(α)).
Note that Bω = B. These sets are fundamental to the analysis that
follows.
Remark 2.11. Throughout most of the paper, we assume that the
accommodating family for the labelled space (E, L, B) is also closed
under relative complements; in this case the set Bα is a Boolean algebra
for each α ∈ L≥1, and Bω = B is a generalized Boolean algebra as in
[12]. By the duality given by Stone's representation theorem there is
a topological space associated with each Bα with α ∈ L∗, which we
INVERSE SEMIGROUPS AND LABELLED SPACES
5
denote by Xα, consisting of the set of ultrafilters in Bα (for details, see
[6] for instance).
Lemma 2.12. Let (E, L, B) be a weakly left resolving labelled space
such that B is closed under relative complements. If A, B ∈ B and
α ∈ L∗ then r(A \ B, α) = r(A, α) \ r(B, α).
Proof. Let A, B ∈ B. Clearly,
r(A \ B, ω) = A \ B = r(A, ω) \ r(B, ω).
For α ∈ L≥1, since A = (A \ B) ⊔ (A ∩ B) (where ⊔ stands for the
union of disjoint sets) and the labelled space is weakly left resolving,
r(A, α) = r((A \ B) ⊔ (A ∩ B), α) = r(A \ B, α) ⊔ r(A ∩ B, α)
so that
r(A \ B, α) = r(A, α) \ (r(A ∩ B, α))
= r(A, α) \ (r(A, α) ∩ r(B, α))
= r(A, α) \ r(B, α).
(cid:3)
2.2. Filters and characters. Following [7], we recall some basic
definitions and properties about filters and characters.
Definition 2.13. Let P be a partially ordered set with least ele-
ment 0. A filter in P is a nonempty subset ξ ⊆ P such that
(i) 0 /∈ ξ;
(ii) if x ∈ ξ and x ≤ y, then y ∈ ξ;
(iii) if x, y ∈ ξ, there exists z ∈ ξ such that z ≤ x and z ≤ y.
An ultrafilter is a filter which is not properly contained in any filter.
Remark 2.14. When P is a (meet) semilattice, condition (iii) may
be replaced by x ∧ y ∈ ξ if x, y ∈ ξ.
Let P be a partially ordered set with least element 0 and x ∈ P .
We denote the principal filter generated by x by ↑x, that is,
↑x = {y ∈ P x ≤ y}.
Similarly, the principal ideal generated by x is
↓x = {y ∈ P y ≤ x}.
When X and Y are subsets of P , we use
↑X = [x∈X
↑x = {y ∈ P x ≤ y for some x ∈ X},
6
G. BOAVA, G. DE CASTRO, AND F. MORTARI
and ↑Y X = Y ∩ ↑X; the sets ↑Y x, ↓Y x, ↓X and ↓Y X are defined
analogously. We remark that ↑X and ↑Y X are not filters, in general.
The next proposition characterizes the subsets X of P for which ↑X is
a filter.
Proposition 2.15. Let P be a partially ordered set with least ele-
ment 0 and X be a nonempty subset of P . Then ↑X is a filter if and
only if 0 /∈ X and for every x, y ∈ X, there exists z ∈ X such that
z ≤ x and z ≤ y. In particular, if P is a semilattice then ↑X is a filter
if and only if 0 /∈ X and for every x, y ∈ X one has x ∧ y ∈ X.
Proof. Trivial.
(cid:3)
Proposition 2.16 ([7], Lemma 12.3). Let E be a semilattice with
0. A filter ξ in E is an ultrafilter if and only if
{y ∈ E y ∧ x 6= 0 ∀ x ∈ ξ} ⊆ ξ.
When E is a Boolean algebra, we have another characterization of
ultrafilters.
Proposition 2.17 ([5], Theorem IV.3.12). Let E be a Boolean
algebra. A filter ξ in E is an ultrafilter if and only if for every x ∈ E,
exactly one of x and ¬x is in ξ.
Definition 2.18. Let E be a semilattice with 0. A character of E
is a nonzero function φ from E to the Boolean algebra {0, 1} such that
(a) φ(0) = 0;
(b) φ(x ∧ y) = φ(x) ∧ φ(y), for all x, y ∈ E.
The set of all characters of E is denoted by E0 and we endow E0 with
the topology of pointwise convergence.
The next definitions are adaptations of [7].
Definition 2.19. Let E be a semilattice with 0 and x ∈ E. A set
Z ⊆ ↓x is said to be a cover for x if for all nonzero y ∈ ↓x, there exists
z ∈ Z such that z ∧ x 6= 0.
Definition 2.20. Let E be a semilattice with 0. A character φ of
E is tight if for every x ∈ E and every finite cover Z for x, we have
φ(z) = φ(x).
_z∈Z
The set of all tight characters of E is denoted by Etight, and called the
tight spectrum of E.
INVERSE SEMIGROUPS AND LABELLED SPACES
7
We can associate each filter ξ in a semilattice E with a character
φξ of E given by
φξ(x) = (cid:26) 1,
0,
if x ∈ E,
otherwise.
Conversely, when φ is a character, ξφ = {x ∈ E φ(x) = 1} is a filter
in E. Clearly, these maps give a bijection between E0 and the set of
filters in E. We denote by E∞ the set of all characters φ of E such
that ξφ is an ultrafilter, and a filter ξ in E is said to be tight if φξ is a
tight character.
Proposition 2.21 ([7], Proposition 12.7). Every ultrafilter is tight,
that is, E∞ ⊆ Etight.
Definition 2.20 can be rephrased to characterize tight filters, as in
the following result.
Proposition 2.22. A filter ξ in E is tight if and only if for every
x ∈ ξ and every finite cover Z for x, one has Z ∩ ξ 6= ∅.
When E is a Boolean algebra, tight filters are easily described:
Proposition 2.23 ([7], Proposition 11.9). Let E be a Boolean al-
gebra and ξ a filter in E. The following are equivalent:
(i) ξ is tight;
(ii) ξ is an ultrafilter;
(iii) φξ is a morphism of Boolean algebras.
The next theorem is the main result relating ultrafilters and tight
characters.
Theorem 2.24 ([7], Theorem 12.9). Let E be a semilattice with 0.
Then Etight is the closure of E∞ in E0.
The convergence of characters in E0 can be translated to the context
of filters; thus, a net {ξλ}λ∈Λ of filters in E converges to a filter ξ if for
each x ∈ E, there exists λ0 ∈ Λ such that for all λ ≥ λ0, x ∈ ξλ if and
only if x ∈ ξ.
3. Definition of the inverse semigroup
For this section let us fix a labelled space (E, L, B), and consider
the set T of all triples (α, A, β) ∈ L∗ × B × L∗ for which A ∈ Bα ∩ Bβ,
together with an extra element, say z. A binary operation on T is
8
G. BOAVA, G. DE CASTRO, AND F. MORTARI
defined as follows: for all t ∈ T define zt = tz = z, and for elements
(α, A, β), (γ, B, δ) in L∗ × B × L∗ put
(α, A, β)(γ, B, δ) =
(αγ′, r(A, γ′) ∩ B, δ),
(α, A ∩ r(B, β′), δβ′),
z,
if γ = βγ′,
if β = γβ′,
otherwise.
Thus, nontrivial products occur only if β and γ are comparable. This
product is indeed well-defined, for one can see for instance in the first
case that A ∈ Bα implies r(A, γ′) ∈ Bαγ′, which gives r(A, γ′) ∩ B ∈
Bαγ′. Similarly, B ∈ Bδ implies r(A, γ′) ∩ B ∈ Bδ, establishing the
condition r(A, γ′) ∩ B ∈ Bαγ′ ∩ Bδ.
Remark 3.1. Note that if β = γ then γ′ = β′ = ω and, therefore,
r(A, γ′) = A, r(B, β′) = B, whence
(α, A, β)(β, B, δ) = (α, A ∩ B, δ).
Proposition 3.2. Suppose that the labelled space (E, L, B) is weakly
left resolving; then, the set T together with the operation above is a
semigroup with zero element given by z.
Proof. It must be shown that the operation is associative. Let
s, t, u ∈ T be given. Clearly associativity holds if any of these equals
z, so suppose they are all in L∗ × B × L∗, say
s = (α, A, β), t = (γ, B, δ), u = (µ, C, ν).
Begin by considering the case st = z and tu 6= z, observing that st = z
implies β and γ are not comparable, while tu 6= z says the first entry in
tu is either γ (if δ = µδ′) or γµ′ (if µ = δµ′); but β is not comparable
with any of these, so s(tu) = z = zu = (st)u.
Evidently the case st 6= z and tu = z follows similarly, and the
case st = tu = z is immediate, so it remains to consider st 6= z and
tu 6= z. There are four cases to investigate, according to how the labels
β, γ, δ, µ compare.
If γ = βγ′ and µ = δµ′, note that γµ′ = βγ′µ′ and
(st)u = (αγ′, r(A, γ′) ∩ B, δ)(µ, C, ν)
= (αγ′µ′, r(r(A, γ′) ∩ B, µ′) ∩ C, ν)
= (αγ′µ′, [r(r(A, γ′), µ′) ∩ r(B, µ′)] ∩ C, ν)
= (αγ′µ′, r(A, γ′µ′) ∩ [r(B, µ′) ∩ C], ν)
= (α, A, β)(γµ′, r(B, µ′) ∩ C, ν)
= s(tu),
INVERSE SEMIGROUPS AND LABELLED SPACES
9
where in the third equality the weakly left resolving hypothesis was
used.
The case β = γβ′ and δ = µδ′ is analogous, and the case γ = βγ′
and δ = µδ′ follows directly from the definition, by comparing the two
sides.
Finally, if β = γβ′ and µ = δµ′, one has st = (α, A ∩ r(B, β′), δβ′)
and tu = (γµ′, r(B, µ′) ∩ C, ν). To evaluate (st)u, compare δβ′ with µ.
Suppose µ begins with δβ′, say µ = δβ′µ′′; since µ = δµ′, this is the
same as µ′ = β′µ′′ and, in particular, γµ′ = γ(β′µ′′) = (γβ′)µ′′ = βµ′′,
so that
(st)u = (α, A ∩ r(B, β′), δβ′)(µ, C, ν)
= (αµ′′, r(A ∩ r(B, β′), µ′′) ∩ C, ν)
= (αµ′′, [r(A, µ′′) ∩ r(r(B, β′), µ′′)] ∩ C, ν)
= (αµ′′, [r(A, µ′′) ∩ r(B, β′µ′′)] ∩ C, ν)
= (αµ′′, r(A, µ′′) ∩ [r(B, µ′) ∩ C], ν)
= (α, A, β)(γµ′, r(B, µ′) ∩ C, ν)
= s(tu).
The case where δβ′ begins with µ = δµ′ is analogous, and the case
in which they are not comparable is trivial, since then both (st)u and
s(tu) equal z.
(cid:3)
The next step is to obtain a set S by identifying with z all elements
in T of the form (α, ∅, β), collapsing them all to a single element, rep-
resented by 0, and leaving the other elements in T as they are. The
operation on the semigroup T passes to this quotient S, turning it into
a semigroup with zero element 0.
In fact, to show the operation is well defined it is enough to see it is
well behaved when operating with the zero class, which is easily done:
if s = (α, ∅, β) and t = (γ, B, δ), note that
• if β and γ are not comparable, then st = z;
• if γ = βγ′, then
st = (αγ′, r(∅, γ′) ∩ B, δ) = (αγ′, ∅ ∩ B, δ) = (αγ′, ∅, δ);
• if β = γβ′ then st = (α, ∅, δβ′).
Thus, in all three cases the class of st in the quotient equals 0, and the
same can be said for the class of ts.
Remark 3.3. Given s = (α, A, β) and t = (γ, B, δ) in S, their
product st may be equal to 0 even when β and γ are comparable, for it
10
G. BOAVA, G. DE CASTRO, AND F. MORTARI
may be the case that r(A, γ′)∩B = ∅ (when γ = βγ′) or A∩r(B, β′) = ∅
(when β = γβ′).
Proposition 3.4. Suppose the labelled space (E, L, B) is weakly
left resolving. Then S is an inverse semigroup with zero element 0.
Proof. Let a nonzero s = (α, A, β) ∈ S be given. Then the ele-
ment s∗ defined by (β, A, α) is such that s∗s = (β, A, β) (see Remark
3.1), giving
ss∗s = s(s∗s) = (α, A, β)(β, A, β) = (α, A, α) = s,
and also s∗ss∗ = s∗.
To prove uniqueness of the inverse, suppose t ∈ S is such that
sts = s and tst = t, noting that the first of these identities implies
t 6= 0. Writing t = (γ, B, δ), from st 6= 0 it follows that β and γ are
comparable, and from ts 6= 0 one sees δ and α are comparable.
Let us show that β = γ.
Indeed, if γ = βγ′ with γ′
6= ω, by
definition one has
s = (st)s = (cid:26) (αγ′α′, r(r(A, γ′) ∩ B, α′) ∩ A, β),
(αγ′, [r(A, γ′) ∩ B] ∩ r(A, δ′), βδ′),
if α = δα′,
if δ = αδ′,
a contradiction since both first entries, αγ′α′ and αγ′, are not equal
to α. The case β = γβ′ with β′
6= ω reaches a similar contradiction,
using t = t(st) and looking at the resulting third entries instead. An
analogous argument establishes that α = δ, and so it can be concluded
that t = (β, B, α).
Finally, s = (st)s and t = (ts)t now give A∩B = A and B ∩A = B,
(cid:3)
whence A = B, that is, t = s∗.
In what follows we describe the filters, ultrafilters and tight filters
in E(S).
4. Filters in E(S)
Recall that if T is an inverse semigroup then its set of idempotents
E(T ) is a semilattice; the meet between two idempotents p and q is
p ∧ q = pq = qp (idempotents in an inverse semigroup commute), and
the induced order is given by p ≤ q when pq = p.
Let (E, L, B) be a labelled space which is weakly left resolving and
consider the inverse semigroup S as in the previous section. In this case,
the set E(S) of idempotents is a semilattice with 0. In this section, we
describe the elements, order and filters in E(S).
Let p = (α, A, β) be a nonzero element in S. If p is an idempotent,
then either β is a beginning of α or α is a beginning of β. In first case,
INVERSE SEMIGROUPS AND LABELLED SPACES
11
writing α = βα′,
p2 = (α, A, β)(α, A, β) = (αα′, r(A, α′) ∩ A, β).
Since p2 = p, we must have α′ = ω, hence α = β. Analogously, when α
is a beginning of β, we also conclude that α = β. On the other hand,
every element in S of the form (α, A, α) is an idempotent. Thus,
E(S) = {(α, A, α) α ∈ L∗ and A ∈ Bα} ∪ {0}.
Proposition 4.1. Let p = (α, A, α) and q = (β, B, β) be nonzero
elements in E(S). Then p ≤ q if and only if α = βα′ and A ⊆ r(B, α′).
Proof. Suppose p ≤ q; by definition this occurs if and only if
(α, A, α)(β, B, β) = (α, A, α), which gives α = βα′. The multiplication
results in (α, A ∩ r(B, α′), βα′) from where it can be deduced that
A ⊆ r(B, α′).
For the converse, it is clear that if α = βα′ and A ⊆ r(B, α′), then
(cid:3)
pq = p.
Remark 4.2. The preceding proposition says that "bigger" ele-
ments in E(S) have "shorter" words and "bigger" sets (when compared
in the same Boolean Algebra). Furthermore, for a fixed α ∈ L∗ the
restricted order in the subset of E(S) given by {(α, A, α) A ∈ Bα}
coincides with the natural inclusion order in Bα.
Let ξ be a filter in E(S) and p = (α, A, α) and q = (β, B, β) in ξ.
Since ξ is a filter, then pq 6= 0; hence, α and β are comparable. This
says the words of any two elements in a filter are comparable.
Proposition 4.3. Let α ∈ L∗ and F be a filter in Bα. Then
ξ = [A∈F
↑(α, A, α)
= {(α1,i, A, α1,i) ∈ E(S) 0 ≤ i ≤ α and r(A, αi+1,α) ∈ F }
is a filter in E(S).
Proof. First we show that the sets in the statement are equal: let
A ∈ F and p = (β, B, β) ∈ ↑(α, A, α). From the order in E(S) and
Proposition 4.1, we obtain α = βα′ and A ⊆ r(B, α′) so that β = α1,i
for some i and r(B, αi+1,α) = r(B, α′) ∈ F , since A ∈ F and F is a
filter.
Conversely, let p = (α1,i, A, α1,i) be such that r(A, αi+1,α) ∈ F.
Together with (α, r(A, αi+1,α), α) ≤ (α1,i, A, α1,i), this shows that p ∈
SA∈F ↑(α, A, α).
Now we show that ξ is a filter: clearly ξ is nonempty, 0 /∈ ξ and if
p ∈ ξ and p ≥ q then q ∈ ξ. Let p, q ∈ ξ and choose A and B in F
12
G. BOAVA, G. DE CASTRO, AND F. MORTARI
such that p ≥ (α, A, α) and q ≥ (α, B, α). We have that A ∩ B ∈ F,
given that F is a filter, whence r = (α, A ∩ B, α) ∈ ξ. Since r ≤ p and
r ≤ q, the proof is complete.
(cid:3)
Proposition 4.4. Let ξ be a filter in E(S) and suppose there exists
a word α ∈ L∗ such that (α, A, α) ∈ ξ for some A ∈ Bα and α has the
largest length among all β such that (β, B, β) ∈ ξ for some B ∈ Bβ.
Define
F = {A ∈ B (α, A, α) ∈ ξ}.
Then F is a filter in Bα and
ξ = [A∈F
↑(α, A, α).
Proof. When we restrict the order of E(S) to elements with the
same word α, the order coincides with the order in Bα, hence F is a
filter. It is easy to see that SA∈F ↑(α, A, α) ⊆ ξ. On the other hand,
let p = (β, B, β) ∈ ξ and choose A ∈ Bα such that q = (α, A, α) ∈ ξ.
Since ξ is a filter, then p ∧ q = pq ∈ ξ and pq 6= 0. Therefore, β must be
a beginning of α (because α has the largest length), say α = βα′, and
pq = (α, A ∩ r(B, α′), α). From this we conclude that A ∩ r(B, α′) ∈ F,
(cid:3)
whence p ∈ ↑(α, A ∩ r(B, α′), α) ⊆ SA∈F ↑(α, A, α).
In Proposition 4.3, we construct a filter with a largest word in E(S)
from a (finite) word α and a filter in Bα, whereas with Proposition 4.4,
we go in the opposite direction. It is easy to see that these constructions
are each other's inverses, and therefore there is a bijective correspon-
dence between filters in E(S) with largest word and pairs (α, F), where
α is a finite word and F is a filter in Bα.
Our next goal is to obtain a characterization of the filters in E(S)
which do not admit a word with largest length.
Definition 4.5. Let α ∈ L∞ and {F n}n≥0 be a family such that
F n is a filter in Bα1,n for every n > 0 and F 0 is a filter in B or F 0 = ∅.
The family {F n}n≥0 is said to be admissible for α if
F n ⊆ {A ∈ Bα1,n r(A, αn+1) ∈ F n+1}
for all n ≥ 0, and is said to be complete for α if
F n = {A ∈ Bα1,n r(A, αn+1) ∈ F n+1}
for all n ≥ 0.
Lemma 4.6. Let α ∈ L∞ and {F n}n≥0 be a family such that F n
is a filter in Bα1,n for every n > 0 and F 0 is a filter in B or F 0 = ∅.
INVERSE SEMIGROUPS AND LABELLED SPACES
13
Then {F n}n≥0 is complete if and only if
F n = {A ∈ Bα1,n r(A, αn+1,m) ∈ F m}
for all n ≥ 0 and all m > n.
Proof. The "if" part is clear and the "only if" part follows easily
by induction using that r(r(A, β), γ) = r(A, βγ) if β, γ ∈ L∗ are such
that βγ ∈ L∗.
(cid:3)
Proposition 4.7. Let α ∈ L∞, {F n}n≥0 be an admissible family
for α and define
ξ =
∞
[n=0 [A∈F n
↑(α1,n, A, α1,n).
Then ξ is a filter in E(S).
Proof. We already know that, for each n ≥ 0,
ξn = [A∈F n
↑(α1,n, A, α1,n)
is a filter in E(S) (except if F 0 = ∅, which would then give ξ0 = ∅).
We claim that ξn is contained in ξn+1 for every n ≥ 0.
Indeed, let
p ∈ ξn and choose q = (α1,n, A, α1,n) ∈ ξn such that q ≤ p. Since the
family is admissible, we conclude that r(A, αn+1) ∈ F n+1 and, hence,
r = (α1,n+1, r(A, αn+1), α1,n+1) ∈ ξn+1. Given that ξn+1 is a filter and
r ≤ q ≤ p, then p ∈ ξn+1. This claim says the filters ξn are nested,
from where we deduce that their union is a filter.
(cid:3)
Proposition 4.8. Let ξ be a filter in E(S) and suppose there is
no word in γ ∈ L∗ such that (γ, C, γ) ∈ ξ for some C ∈ Bγ and γ
possesses largest length among all β such that (β, B, β) ∈ ξ for some
B ∈ Bβ. Then there exists α ∈ L∞ such that every p ∈ ξ can be written
as p = (α1,n, A, α1,n) for some n ≥ 0 and some A ∈ Bα1,n. Moreover,
if we define for each n ≥ 0,
F n = {A ∈ B (α1,n, A, α1,n) ∈ ξ},
then {F n}n≥0 is a complete family for α.
Proof. For each n > 0, choose p = (β, B, β) ∈ ξ such that β ≥ n
and define αn = βn. Since any two elements in ξ have comparable
words, then αn is well defined. Hence, we have a word α = α1α2 . . . ∈
L∞. Let F n as in the statement. We claim that F n is nonempty
for all n > 0. Indeed, fix n > 0 e choose m ≥ n such that there is
p = (α1,m, A, α1,m) ∈ ξ. Observe that q = (α1,n, r(α1,n), α1,n) ≥ p and,
because ξ is a filter, q ∈ ξ. This shows that r(α1,n) ∈ F n.
14
G. BOAVA, G. DE CASTRO, AND F. MORTARI
Again, since the order in E(S) is the same as the order in Bα1,n
when we consider only elements with the word α1,n, we conclude that
F n is a filter in Bα1,n for all n ≥ 0 (except for n = 0 where F 0 may be
the empty set).
It remains to show that the family is complete. Fix n ≥ 0 and
denote by G the set {A ∈ Bα1,n r(A, αn+1) ∈ F n+1}. Let A ∈ F n.
Since F n+1 is nonempty, we can choose B ∈ F n+1. Because ξ is a filter
and both p = (α1,n, A, α1,n) and q = (α1,n+1, B, α1,n+1) belong to ξ,
it follows that p ∧ q = (α1,n+1, r(A, αn+1) ∩ B, α1,n+1) ∈ ξ. This says
r(A, αn+1) ∩ B is nonempty and belongs to F n+1. Since F n+1 is a filter
in Bα1,n+1 and r(A, αn+1) ∈ Bα1,n+1, we have that r(A, αn+1) ∈ F n+1
and, therefore, A ∈ G. On the other hand, let A ∈ G and observe that
p = (α1,n+1, r(A, αn+1), α1,n+1) ∈ ξ. Clearly, p ≤ q = (α1,n, A, α1,n).
Again using that ξ is a filter, we obtain q ∈ ξ, which says A ∈ F n. (cid:3)
Proposition 4.9. There is a bijective correspondence between fil-
ters in E(S) without largest word and pairs (α, {F n}n≥0), where α ∈
L∞ and {F n}n≥0 is a complete family for α. These correspondences
are given by Propositions 4.7 and 4.8.
Proof. Let ξ be a filter in E(S) without a largest word, consider
(α, {F n}n≥0) as in Proposition 4.8 and let η be the filter given by
Proposition 4.7 from (α, {F n}n≥0); we will show that ξ = η: clearly
ξ ⊆ η from the definition of η. Let p ∈ η and choose n ≥ 0 and
A ∈ F n such that (α1,n, A, α1,n) ≤ p. By the definition of F n we have
(α1,n, A, α1,n) ∈ ξ and, since ξ is a filter, it follows that p ∈ ξ and the
equality is established.
Conversely, let α ∈ L∞ and {F n}n≥0 be a complete family for α,
consider the filter ξ constructed from them as in Proposition 4.7 and
let (β, {Gn}n≥0) be given from ξ by Proposition 4.8. We need to show
that (α, {F n}n≥0) = (β, {Gn}n≥0). By construction, we see that α = β.
Fix m ≥ 0 and let A ∈ F m. Then (α1,m, A, α1,m) ∈ ξ, hence A ∈ Gm.
On the other hand, given B ∈ G m the definition of Gm ensures that
(α1,m, B, α1,m) ∈ ξ; therefore, there exists k ≥ m and C ∈ F k such
that (α1,k, C, α1,k) ≤ (α1,m, B, α1,m). This says that C ⊆ r(B, αm+1,k)
and, since F k is a filter in Bα1,k , it follows that r(B, αm+1,k) ∈ F k.
Completeness of {F n}n≥0 and Lemma 4.6 now give B ∈ F m. Thus,
(α, {F n}n≥0) = (β, {Gn}n≥0) and the proof is complete.
(cid:3)
The previous results classify the filters in E(S) in two types: the
ones as in Propositions 4.3 and 4.4, which we call filters of finite type;
and the filters as in Propositions 4.7, 4.8 and 4.9, called filters of infinite
type.
INVERSE SEMIGROUPS AND LABELLED SPACES
15
Next, we present some properties about filters in E(S) that will be
useful later; furthermore, we introduce a new notation for the filters.
Proposition 4.10. Let α, β ∈ L∗ be such that αβ ∈ L∗, F be a
filter in Bα and G be a filter in Bαβ. The following are equivalent:
(i) {r(A, β) A ∈ F} ⊆ G;
(ii) F ⊆ {A ∈ Bα r(A, β) ∈ G};
(iii) [A∈F
↑(α, A, α) ⊆ [B∈G
↑(αβ, B, αβ).
Proof. The equivalence between (i) and (ii) is obvious. Suppose
(iii) is valid and let A ∈ F; by (iii), there exists B ∈ G such that
(αβ, B, αβ) ≤ (α, A, α), hence B ⊆ r(A, β). Since G is a filter in Bαβ
and r(A, β) ∈ Bαβ, therefore r(A, β) ∈ G, proving (ii).
that (α, A, α) ≤ p. We have (αβ, r(A, β), αβ) ≤ (α, A, α) ≤ p and, by
Now assume (ii), let p ∈ SA∈F ↑(α, A, α) and choose A ∈ F such
(ii), r(A, β) ∈ G, so that p ∈ SB∈G ↑(αβ, B, αβ) which gives (iii). (cid:3)
Proposition 4.11. Let α ∈ L∗ and {F n}n≥0 be an admissible fam-
ily for α. For each n ≥ 0, define
F n = {A ∈ Bα1,n r(A, αn+1,m) ∈ F m for some m ≥ n}.
Then:
(i) F n ⊆ F n, for all n ≥ 0;
(ii) {F n}n≥0 is a complete family for α;
(iii) (α, {F n}n≥0) and (α, {F n}n≥0) generate the same filter in E(S)
(see Proposition 4.7).
(iv) If ξ is the filter associated with (α, {F n}n≥0), then the pair asso-
ciated with ξ (by Proposition 4.8) is (α, {F n}n≥0).
Proof. Item (i) is immediate from the definition of F n. We claim
F n is a filter in Bα1,n (or possibly the empty set when n = 0): clearly,
∅ /∈ F n. Let A1 and A2 be in F n and choose m1, m2 ≥ n such
that r(A1, αn+1,m1) ∈ F m1 and r(A2, αn+1,m2) ∈ F m2. Define m =
max{m1, m2} and note that r(A1, αn+1,m) and r(A2, αn+1,m) both be-
long to F m, given that {F n}n≥0 is admissible. Because F m is a filter
in Bα1,m and the labelled space is weakly left resolving, we have that
r(A1 ∩ A2, αn+1,m) = r(A1, αn+1,m) ∩ r(A2, αn+1,m) ∈ F m
and therefore A1 ∩ A2 ∈ F n. Now, let A ∈ F n and B ∈ Bα1,n be such
that B ⊇ A. Choose m ≥ n such that r(A, αn+1,m) ∈ F m and observe
that r(B, αn+1,m) ⊇ r(A, αn+1,m). Since F m is a filter, we have that
r(B, αn+1,m) ∈ F m, which says B ∈ F n. This completes the proof of
our claim.
16
G. BOAVA, G. DE CASTRO, AND F. MORTARI
Next we show that {F n}n≥0 is complete: Let
Gn = {A ∈ Bα1,n r(A, αn+1) ∈ F n+1}.
We need to show that F n = Gn. To see this, let A ∈ F n and choose
m ≥ n such that r(A, αn+1,m) ∈ F m. If m = n then A ∈ F n and, since
{F n}n≥0 is admissible, we have r(A, αn+1) ∈ F n+1 ⊆ F n+1, which says
A ∈ Gn. If m > n, then r(r(A, αn+1), αn+2,m) = r(A, αn+1,m) ∈ F m
and therefore r(A, αn+1) ∈ F n+1, whence A ∈ Gn. On the other hand,
let B ∈ Gn. Then r(B, αn+1) ∈ F n+1 and we can take m ≥ n + 1
such that r(B, αn+1,m) = r(r(B, αn+1), αn+2,m) ∈ F m. It follows that
B ∈ F n and the proof of (ii) is complete.
In order to show (iii), denote by ξ and ¯ξ the filters associated
with (α, {F n}n≥0) and (α, {F n}n≥0), respectively. Clearly, ξ ⊆ ¯ξ; let
p ∈ ¯ξ and choose n ≥ 0 and A ∈ F n such that (α1,n, A, α1,n) ≤ p.
Take m ≥ n such that r(A, αn+1,m) ∈ F m and observe that q =
(α1,m, r(A, αn+1,m), α1,m) ∈ ξ and q ≤ p. ξ is a filter, therefore p ∈ ξ,
which shows (iii).
Item (iv) is consequence from (iii) and Proposition 4.9.
(cid:3)
Proposition 4.12. Let ξ be a filter in E(S) of finite type and
(α, F) be its associated pair. For each n ∈ {0, 1, . . . , α}, define
F n = {A ∈ B (α1,n, A, α1,n) ∈ ξ}.
Then:
(i) F n is a filter in Bα1,n for every n ∈ {1, . . . , α} and F 0 is either
empty or a filter in B;
(ii) F n = {A ∈ Bα1,n r(A, αn+1,m) ∈ F m}, for all 0 ≤ n < m ≤ α.
Proof. These results are proved exactly in the same way as Propo-
(cid:3)
sition 4.8 and Lemma 4.6.
We summarize the results obtained until now. There are two kinds
of filters in E(S): of infinite and finite type. The former is characterized
by an infinite word α and a complete family {F n}n≥0 for α. Clearly,
for n ≤ m, F n is completely determined from F m. When we need to
create a filter of infinite type, we need only an infinite word α and an
admissible family for α. Indeed, an admissible family can be completed
to a complete family by using Proposition 4.11. Sometimes, we use
Proposition 4.10 to obtain equivalent formulations for the admissibility
condition in Definition 4.5.
The filters of finite type are determined by a finite word α and a
filter F in Bα. Aiming to obtain a standard way to work with all filters,
INVERSE SEMIGROUPS AND LABELLED SPACES
17
we now describe filters of finite type in a slightly different form: analo-
gous to Definition 4.5, to each finite word α we can define admissible or
complete families {F n}0≤n≤α for α (when α is the empty word ω, we
do not allow F 0 = ∅). Due to Proposition 4.11, admissible families can
be completed and, since the filter F α determines the family (Proposi-
tion 4.12), then it is clear that there is a correspondence between filters
of finite type and pairs (α, {F n}0≤n≤α), where α is a finite word and
{F n}0≤n≤α is a complete family for α.
Combining the results obtained in this section, we obtain the fol-
lowing description for the filters in E(S).
Theorem 4.13. Let (E, L, B) be a labelled space which is weakly
left resolving, and let S be its associated inverse semigroup. Then
there is a bijective correspondence between filters in E(S) and pairs
(α, {F n}0≤n≤α), where α ∈ L≤∞ and {F n}0≤n≤α is a complete fam-
ily for α (understanding that 0 ≤ n ≤ α means 0 ≤ n < ∞ when
α ∈ L∞).
In view of this, we will use the following notation: when a filter ξ
in E(S) with associated word α ∈ L≤∞ is given, we write ξ = ξα to
stress the word α. Furthermore, we denote by ξα
n (or ξn when there is
no ambiguity) the filter F n in Bα1,n of the complete family associated
with ξα (obviously, n ≤ α). To be specific:
and the family {ξα
ξα
n = {A ∈ B (α1,n, A, α1,n) ∈ ξα}
n }0≤n≤α satisfies
ξα
n = {A ∈ Bα1,n r(A, αn+1,m) ∈ ξα
m}
for all 0 ≤ n < m ≤ α.
Finally we remark that, in order to create a filter of finite type, we
need only a word α ∈ L∗ and a filter in Bα, whereas to create a filter
of infinite type, it is enough to have a word α ∈ L∞ and an admissible
family for α.
For future reference, we denote by F the set of all filters in E(S)
and by Fα the subset of F of those filters whose associated word is
α ∈ L≤∞.
5. Ultrafilters in E(S)
Next we describe the ultrafilters in E(S). Since the notion of ultra-
filter is associated with inclusion among filters, we enunciate an easy,
but useful, proposition which characterizes the order (given by inclu-
sion) in F.
18
G. BOAVA, G. DE CASTRO, AND F. MORTARI
Proposition 5.1. Let ξα, ηβ ∈ F. Then ξα ⊆ ηβ if and only if α
is a beginning of (or equal to) β and ξn ⊆ ηn for all 0 ≤ n ≤ α.
Proof. Trivial.
(cid:3)
Remark 5.2. The proposition above can be simplified in the case
α is a finite word: the statement would then be ξα ⊆ ηβ if and only if
α is a beginning of (or equal to) β and ξα ⊆ ηα.
Proposition 5.3. Let ξα ∈ F be a filter of infinite type, that is, α ∈
L∞. Then ξα is an ultrafilter if and only if {ξn}n≥0 is maximal among
all complete families for α, where maximal means that if {F n}n≥0 is
complete family for α such that ξn ⊆ F n for all n ≥ 0, then ξn = F n
for all n ≥ 0.
Proof. First suppose ξα is an ultrafilter and let {F n}n≥0 be a
complete family for α such that ξn ⊆ F n for all n ≥ 0.
If η is the
filter associated with (α, {F n}n≥0) then ξα ⊆ η, by the previous result.
Since ξα is an ultrafilter, therefore ξα = η, hence ξn = F n for all n ≥ 0,
that is, {ξn}n≥0 is maximal.
On the other hand, suppose {ξn}n≥0 is maximal and let ηβ ∈ F be
such that ξα ⊆ ηβ. By the previous result we must have β = α and
ξn ⊆ ηn for all n ≥ 0; since {ξn}n≥0 is maximal, we have that ξn = ηn
for all n and therefore ξα = ηβ, showing that ξα is an ultrafilter.
(cid:3)
Proposition 5.4. Let ξα ∈ F be a filter of finite type, that is,
α ∈ L∗. Then ξα is an ultrafilter if and only if ξα is an ultrafilter in
Bα and for each letter b ∈ A, there exists A ∈ ξα such that r(A, b) = ∅.
Proof. (⇒) Suppose ξα is an ultrafilter, and let F be a filter in
Bα such that ξα ⊆ F . The filter η associated with (α, F) contains ξα,
therefore η = ξα and so F = ξα, whence ξα is an ultrafilter.
Now, suppose by contradiction there exists b ∈ A such that r(A, b) 6=
∅ for all A ∈ ξα. Consider the subset of Bαb given by
Y = {r(A, b) A ∈ ξα}.
Given that ξα is a filter and the labelled space is weakly left resolv-
ing, we have that Y is closed under intersections and, by hypothesis,
∅ /∈ Y ; it follows from Proposition 2.15 that ↑B
αb Y is a filter in Bαb.
Consider then the filter η associated with the pair (αb, ↑B
αb Y ). Since
ηα = {A ∈ Bα r(A, b) ∈ ↑Bαb Y } ⊇ ξα, we would then have ξα $ η
by Remark 5.2, contradicting the fact that ξα is an ultrafilter.
(⇐) Suppose ξα is not an ultrafilter, so that there is a ηβ ∈ F such
that ξα $ ηβ, noting that α is a beginning of β, by Proposition 5.1. If
β = α then ξα $ ηα (by Remark 5.2), hence ξα is not an ultrafilter in
INVERSE SEMIGROUPS AND LABELLED SPACES
19
Bα. If β 6= α, then β = αγ with γ 6= ω. Let b be the first letter of γ and
observe that ξα ⊆ ηα = {A ∈ Bα r(A, b) ∈ ηα+1}; it follows from
Proposition 4.10 that {r(A, b) A ∈ ξα} ⊆ ηα+1 and, because ηα+1
is a filter, we have r(A, b) 6= ∅ for all A ∈ ξα, a contradiction.
(cid:3)
Remark 5.5. The condition "for each letter b ∈ A, there exists A ∈
ξα such that r(A, b) = ∅" in the previous result has other variations;
indeed, if ξα ∈ F is a filter of finite type, it is then easy to see that the
following conditions are equivalent:
(i) For each letter b ∈ A, there exists A ∈ ξα such that r(A, b) = ∅;
(ii) For each letter b ∈ A such that αb ∈ L∗, there exists A ∈ ξα such
that r(A, b) = ∅;
(iii) For each word β ∈ L∗, there exists A ∈ ξα such that r(A, β) = ∅;
(iv) For each word β ∈ L∗ such that αβ ∈ L∗, there exists A ∈ ξα
such that r(A, β) = ∅.
The previous two results characterize the ultrafilters in E(S), but
natural questions related to Propositions 5.3 and 5.4 can be raised; for
instance, if α ∈ L∞ and ξα is a filter, it is clear that if ξn is ultrafilter
for every n ≥ 0, then ξα is an ultrafilter. The converse, however, is not
true in general, as the next example illustrates.
Example 5.6. Consider the labelled graph (E, L) given by
a1
v1
a1
v2
v4
a2
a2
a2
v3
v5
a3
v6
a4
v7
· · ·
a5
and define B as the family of all subsets A ⊆ E0 such that v2 ∈ A
whenever v4 ∈ A and v3 ∈ A whenever v5 ∈ A.
It is not difficult
to check that B is a weakly left resolving accommodating family for
(E, L).
Consider the labelled path α = a1a2a3 · · · . We claim there ex-
ists a unique complete family {F n}n≥0 for α (which is therefore max-
imal), given by F 0 = {{v1}}, F 1 = {{v2, v4}}, F 2 = {{v3, v5}} and
F n = {{vn+3}} for n ≥ 3: to prove this, first observe that Bα1,1 =
{∅, {v2}, {v2, v4}}, Bα1,2 = {∅, {v3}, {v3, v5}} and Bα1,n = {∅, {vn+3}}
for n ≥ 3. It is easily checked that the given family is complete for α,
and we see that for n ≥ 3 the only possible filter in Bα1,n is {{vn+3}}.
Since in a complete family each filter determines uniquely the ones that
come before it, this establishes the desired uniqueness (for instance,
20
G. BOAVA, G. DE CASTRO, AND F. MORTARI
F 2 = {A ∈ Bα1,2 : r(A, a3) ∈ F 3} and, since r({v3}, a3) = ∅, then
F 2 = {{v3, v5}}. Similarly, the only possibility for F 1 is {{v2, v4}}).
By Proposition 5.3, we obtain an ultrafilter ξα on E(S) such that
ξn = F n for all n; however, ξ2 = {{v3, v5}} is not an ultrafilter on Bα1,2
because it is contained in the filter {{v3}, {v3, v5}}.
On the other hand, if we assume that the family B is closed under
relative complements, that is, for every A, B ∈ B one has A \ B ∈ B,
then the situation described in the previous example does not occur
and,
in view of this, we assume from now on that the accommo-
dating family B of the labelled space (E, L, B) is closed under rela-
tive complements; this ensures in particular that B = Bω is a gen-
eralized Boolean algebra, Bα is a Boolean algebra for α ∈ L≥1 and
r(A\ B, β) = r(A, β)\ r(B, β) for β ∈ L∗, see Remark 2.11 and Lemma
2.12. We recall this hypothesis was already used by Bates, Pask and
Willis in [4], and Bates, Carlsen and Pask in [1].
Proposition 5.7. Suppose the accommodating family B is closed
under relative complements and let ξα be a filter in E(S). If ξn is an
ultrafilter for n with 0 < n ≤ α, then ξm is an ultrafilter for every
0 < m < n. If furthermore ξ0 is nonempty, then ξ0 is also an ultrafilter.
Proof. Fix 0 < m < n and consider the relative range func-
tion, r( · , αm+1,n) : Bα1,m −→ Bα1,n. The fact that Bα1,m and Bα1,n
are Boolean algebras and r( · , αm+1,n) is a morphism of Boolean al-
gebras ensures that the inverse image of an ultrafilter in Bα1,n under
r( · , αm+1,n) is an ultrafilter in Bα1,m. By the definition of ξm and ξn,
we see that ξm is the inverse image of ξn; thus, if ξn is an ultrafilter,
then ξm is an ultrafilter.
Now, suppose that ξn is an ultrafilter and ξ0 6= ∅. We will show
that ξ0 satisfies Proposition 2.16 and will use Proposition 2.17 to char-
acterize ξn. Let C ∈ B be such that C ∩ B 6= ∅ for all B ∈ ξ0 and
fix any A ∈ ξ0. By the condition of C, A \ C does not belong to ξ0.
This says that r(A, α1,n) ∈ ξn and r(A \ C, α1,n) /∈ ξn. Thus, using the
notation Dc = r(α1,n) \ D,
r(A \ C, α1,n) /∈ ξn ⇒ r(A, α1,n) \ r(C, α1,n) /∈ ξn
⇒ r(A, α1,n) ∩ r(C, α1,n)c /∈ ξn
⇒ r(A, α1,n)c ∪ r(C, α1,n) ∈ ξn
⇒ r(A, α1,n) ∩ (r(A, α1,n)c ∪ r(C, α1,n)) ∈ ξn
⇒ r(A, α1,n) ∩ r(C, α1,n) ∈ ξn
⇒ r(C, α1,n) ∈ ξn
⇒ C ∈ ξ0.
(by 2.17)
This shows that ξ0 is an ultrafilter.
(cid:3)
INVERSE SEMIGROUPS AND LABELLED SPACES
21
Remark 5.8. Let ξα be an ultrafilter as in Proposition 5.4. If the
accommodating family B is closed under relative complements, then
the previous proposition applies and therefore ξn is an ultrafilter for all
0 ≤ n ≤ α (except, possibly, for n = 0).
Proposition 5.9. Suppose the accommodating family B is closed
under relative complements and let ξα be a filter of infinite type. Then
ξα is an ultrafilter if and on if ξn is an ultrafilter for every n > 0 and
ξ0 is either an ultrafilter or the empty set.
Proof. The "if" part is obvious, so assume ξα is an ultrafilter and
fix n > 0. Let A ∈ Bα1,n be such that Ac = r(α1,n) \ A does not
belong to ξn. By using the characterization given by Proposition 2.17
for ultrafilters in Boolean algebras, it suffices to verify that A ∈ ξn to
show that ξn is an ultrafilter.
Since p = (α1,n, Ac, α1,n) /∈ ξα and ξα is an ultrafilter, it follows from
Proposition 2.16 that there exists q = (α1,m, B, α1,m) ∈ ξα such that
pq = 0. There are three cases to consider: if m = n, then B ∩ Ac = ∅,
which would give B ⊆ A and thus A ∈ ξn, since ξn is a filter. The
case m < n is analogous using q = (α1,n, r(B, αm+1,n), α1,n) instead of
q, and the case m > n follows by using p = (α1,m, r(Ac, αn+1,m), α1,m)
instead of p to obtain that r(A, αn+1,m) ∈ ξm, and therefore A ∈ ξn.
This shows that ξn is an ultrafilter for n > 0.
The case n = 0 follows from Proposition 5.7.
(cid:3)
The results of this section therefore give the following description
of the ultrafilters in E(S).
Theorem 5.10. Let (E, L, B) be a labelled space which is weakly
left resolving, and let S be its associated inverse semigroup. Then the
ultrafilters in E(S) are:
(i) The filters of finite type ξα such that ξα is an ultrafilter in Bα
and for each b ∈ A there exists A ∈ ξα such that r(A, b) = ∅.
(ii) The filters of infinite type ξα such that the family {ξn}n≥0 is max-
imal among all complete families for α.
Suppose in addition that the accommodating family B is closed under
relative complements. Then (ii) can be replaced with
(ii)' The filters of infinite type ξα such that ξn is an ultrafilter for
every n > 0 and ξ0 is either an ultrafilter or the empty set.
Remark 5.11. When the accommodating family B is closed under
relative complements, we can use Proposition 5.7 to conclude that a
filter ξα of infinite type is an ultrafilter if and only for every n ∈ N
there exists m ≥ n such that ξm is an ultrafilter.
22
G. BOAVA, G. DE CASTRO, AND F. MORTARI
6. Tight Filters in E(S)
Let Ftight be the set of all tight filters in E(S).
If we give to it
the topology induced from the topology of pointwise convergence of
characters, via the bijection between tight characters and tight filters
given after Definition 2.20, it follows that Ftight is (homeomorphic to)
the tight spectrum of E(S).
Now we characterize the tight filters in E(S), a goal that is accom-
plished with Theorem 6.7. From now on the accommodating family
B is required to be closed under relative complements, as above. We
begin with some auxiliary results.
Proposition 6.1. Suppose the accommodating family B is closed
under relative complements and let ξα be a tight filter in E(S). Then,
for every 0 ≤ n ≤ α, ξn is an ultrafilter in Bα1,n (ξ0 may also be the
empty set).
Proof. Fix n > 0 and suppose ξn is not an ultrafilter. There exists
then A ∈ Bα1,n such that A and Ac = r(α1,n) \ A does not belong to
ξn.
Let p = (α1,n, r(α1,n), α1,n), Z = {(α1,n, A, α1,n), (α1,n, Ac, α1,n)}
and note that p ∈ ξα and Z is a cover for p; indeed, for every q ≤ p, we
must have q = (α1,nβ, B, α1,nβ) for some β ∈ L∗ and some B ∈ Bα1,nβ.
Since B ⊆ r(α1,nβ) = r(A, β) ∪ r(Ac, β), then B ∩ r(A, β) 6= ∅ or B ∩
r(Ac, β) 6= ∅ and, therefore, q · (α1,n, A, α1,n) 6= 0 or q · (α1,n, Ac, α1,n) 6=
0. This shows Z is a cover for p. Since Z ∩ ξα = ∅, we see that ξα is
not tight (see Proposition 2.22), a contradiction.
Now suppose ξ0 is nonempty and not an ultrafilter. By Proposition
2.16, there is A ∈ B such that A /∈ ξ0 and A ∩ B 6= ∅ for all B ∈ ξ0.
Fix any B ∈ ξ0 and define p = (ω, B, ω) and
Z = {(ω, B ∩ A, ω), (ω, B \ A, ω)}.
An argument similar to the one used above shows that Z is a cover
for p. Since p ∈ ξα and Z ∩ ξα = ∅ (because B ∩ A ⊆ A /∈ ξ0 and
A ∩ (B \ A) = ∅), it follows that ξα is not tight, again a contradiction
and the result is proved.
(cid:3)
Corollary 6.2. Suppose the accommodating family B is closed
under relative complements and let ξ be a filter of infinite type. Then
ξ is tight if and only if ξ is an ultrafilter.
Proof. This follows from Propositions 6.1, 2.21 and Theorem 5.10.
(cid:3)
INVERSE SEMIGROUPS AND LABELLED SPACES
23
Corollary 6.3. Suppose the accommodating family B is closed
under relative complements and let ξα be a filter of finite type. If ξα
is not an ultrafilter, then ξα is not tight.
Proof. Obvious from Proposition 6.1.
Before stating our main theorem, for each α ∈ L∗ recall from Re-
(cid:3)
mark 2.11 that
Xα = {F ⊆ Bα F is an ultrafilter in Bα}
and also define
X sink
α = {F ∈ Xα ∀ b ∈ A, ∃ A ∈ F such that r(A, b) = ∅}.
Observe that X sink
Remark 5.5.
α
can be defined in many equivalent ways, using
Suppose the accommodating family B to be closed under relative
complements. In this case, for every α, β ∈ L≥1 such that αβ ∈ L≥1,
the map r( · , β) : Bα −→ Bαβ is a morphism of Boolean algebras and,
: Xαβ −→ Xα given by
therefore, we have its dual morphism fα[β]
fα[β](F ) = {A ∈ Bα r(A, β) ∈ F} (this is just the inverse image of
F under r( · , β)), as in the proof of Proposition 5.7. When α = ω, if
F ∈ Bβ then {A ∈ B r(A, β) ∈ F} is either an ultrafilter in B = Bω or
the empty set, and we can therefore consider fω[β] : Xβ −→ Xω ∪ {∅}.
Employing this new notation, if ξα is a filter in E(S) and 0 ≤ m <
n ≤ α, then ξm = fα1,m[αm+1,n](ξn).
If we endow the sets Xα with the topology given by the convergence
of filters stated at the end of Section 2.2 (this is the pointwise conver-
gence of characters), it is clear that the functions fα[β] are continuous.
Furthermore, it is easy to see that fα[βγ] = fα[β] ◦ fαβ[γ].
Corollary 6.2 says there are no tight filters other than the ultrafilters
among the filters of infinite type. The next result classifies the tight
filters of finite type.
Proposition 6.4. Suppose the accommodating family B to be closed
under relative complements and let ξα be a filter of finite type. Then
ξα is a tight filter if and only if ξα is an ultrafilter and at least one of
the following conditions hold:
(a) There is a net {F λ}λ∈Λ ⊆ X sink
(b) There is a net {(tλ, F λ)}λ∈Λ, where tλ is a letter in A and F λ ∈
Xαtλ for each λ ∈ Λ, such that {fα[tλ](F λ)}λ∈Λ converges to ξα
and for every b ∈ A there is λb ∈ Λ such that tλ 6= b for all λ ≥ λb.
converging to ξα.
α
Proof. We show the "if" part first. Suppose that ξα is an ultrafil-
ter and that (a) holds. For each λ ∈ Λ, consider the filter ηλ associated
24
G. BOAVA, G. DE CASTRO, AND F. MORTARI
with the pair (α, F λ) (see Proposition 4.3). By Theorem 5.10, {ηλ}λ∈Λ
is a net of ultrafilters in F. Furthermore, since {ηλ
α}λ∈Λ converges to
ξα in Xα and the functions fα1,m[αm+1,α] are continuous, then {ηλ
m}λ∈Λ
converges to ξm for every 0 ≤ m ≤ α. From this, it is easy to see that
{ηλ}λ∈Λ converges to ξα, whence ξα is tight, by Theorem 2.24.
Now, suppose that ξα is an ultrafilter and that (b) holds. For
each λ ∈ Λ, consider an ultrafilter ηλ containing the filter associated
with the pair (αtλ, F λ) and observe that ηλ
α+1 = F λ. We claim that
{ηλ}λ∈Λ converges to ξα. To see this, take (β, B, β) ∈ E(S). We divide
the proof in three cases.
Case 1: Suppose β is a beginning of (or equal to) α. In this case, β =
α1,m for some m ≤ α. Since {ηλ
α+1)}λ∈Λ converges
to ξα and fα1,m[αm+1,α] is continuous, then {ηλ
m}λ∈Λ converges to ξm.
From this, it is easy to see there is λ0 ∈ Λ such that for every λ ≥ λ0,
(β, B, β) ∈ ηλ if and only if (β, B, β) ∈ ξα.
Case 2: Suppose β and α are not comparable. In this case, (β, B, β) /∈
ξα and (β, B, β) /∈ ηλ for all λ ∈ Λ.
Case 3: Suppose α is a beginning of (and not equal to) β. Write β = αγ
and denote by b the first letter of γ. By (b), we can choose λb ∈ Λ
such that tλ 6= b for all λ ≥ λb. This says that (β, B, β) /∈ ηλ if λ ≥ λb.
Since (β, B, β) /∈ ξα, we are done.
α}λ∈Λ = {fα[tλ](ηλ
Now, for "only if" side: suppose ξα is a tight filter. By Corollary
6.3, ξα is an ultrafilter in Bα. By Theorem 2.24, we can choose a net
{ηλ}λ∈Λ of ultrafilters in E(S) which converges to ξα. Consider the
following disjoint subsets of Λ:
Λ1 = {λ ∈ Λ the word of ηλ is a beginning of (and not equal to) α},
Λ2 = {λ ∈ Λ the word of ηλ is equal to α},
Λ3 = {λ ∈ Λ α is a beginning of (and not equal to) the word of ηλ},
Λ4 = {λ ∈ Λ the word of ηλ is not comparable with α}.
α}λ∈Λ2 is a net in X sink
α
Because {ηλ}λ∈Λ converges to ξα, Λ1 and Λ4 cannot be cofinal subsets of
Λ. Since Λ = Λ1 ∪Λ2 ∪Λ3 ∪Λ4, then Λ2 or Λ3 are cofinal. If Λ2 is cofinal
then, by Theorem 5.10, {ηλ
. Furthermore, it is
clear that {ηλ
α}λ∈Λ2 converges to ξα, establishing (a). If Λ3 is cofinal,
denote by tλ the first letter after α in the word of ηλ, for each λ ∈ Λ3.
Thus, {(tλ, ηλ
α+1 ∈ Xαtλ. Since {ηλ}λ∈Λ3
converges to ξα, then {fα[tλ](ηλ
α}λ∈Λ3 converges to ξα.
This establishes the first part of (b).
α+1)}λ∈Λ3 is a net such that ηλ
α+1)}λ∈Λ3 = {ηλ
Finally, let b ∈ A be given. If αb is not a labelled path, then tλ 6= b
for every λ ∈ Λ3. If αb is a labelled path, then (αb, r(αb), αb) ∈ E(S) is
INVERSE SEMIGROUPS AND LABELLED SPACES
25
such that (αb, r(αb), αb) /∈ ξα. By the definition of convergence, there
exists λb ∈ Λ3 such that (αb, r(αb), αb) /∈ ηλ for all λ ≥ λb, which is
equivalent to tλ 6= b for all λ ≥ λb. This completes (b).
(cid:3)
Remark 6.5. The following conditions are easily seen to be equiv-
alent (and therefore condition (b) in the above result can be replaced
with its equivalent formulation below):
(i) There is a net {(tλ, F λ)}λ∈Λ, where tλ is a letter in A and F λ ∈
Xαtλ for each λ ∈ Λ, such that {fα[tλ](F λ)}λ∈Λ converges to ξα
and for every b ∈ A there is λb ∈ Λ such that tλ 6= b for all λ ≥ λb.
(ii) There is a net {(βλ, F λ)}λ∈Λ, where βλ is a word in L≥1 and
F λ ∈ Xαβλ for each λ ∈ Λ, such that {fα[βλ](F λ)}λ∈Λ converges
to ξα and for every γ ∈ L≥1 there is λγ ∈ Λ such that βλ and γ
are not comparable for all λ ≥ λb.
Furthermore, if we give A the discrete topology, the condition "for
every b ∈ A there is λb ∈ Λ such that tλ 6= b for all λ ≥ λb" is the same
as saying the net {tλ}λ∈Λ ⊆ A converges to infinity.
The next result gives a more algebraic description for the tight
filters of finite type in E(S).
Proposition 6.6. Suppose the accommodating family B is closed
under relative complements and let ξα be a filter of finite type. Then ξα
is a tight filter if and only if ξα is an ultrafilter and for each A ∈ ξα
at least one of the following conditions hold:
(a) L(AE1) given by (2.1) is infinite.
(b) There exists B ∈ Bα such that ∅ 6= B ⊆ A ∩ E0
sink.
α
α
sink.
Proof. Let ξα be a tight filter as in part (a) of Proposition 6.4
and A ∈ ξα be given. If L(AE1) is infinite there is nothing to prove.
Suppose then that L(AE1) is finite. Since A ∈ ξα there exists λ such
, for each b ∈ L(AE1)
that A ∈ F λ ∈ X sink
there exists Cb ∈ F λ such that r(Cb, b) = ∅. Since L(AE1) is finite,
if C is the intersection of all such Cb then C ∈ F λ.
It follows that
r(A ∩ C, b) = ∅ for all b ∈ A and hence A ∩ C ⊆ E0
sink. Defining
B = A ∩ C, we have that B ∈ F λ so that B ∈ Bα and clearly ∅ 6= B ⊆
A ∩ E0
. By the definition of X sink
Now let ξα be a tight filter as in part (b) of Proposition 6.4 and
A ∈ ξα be given. There exists λ0 such that A ∈ fα[tλ](F λ) for all
λ ≥ λ0. That means that r(A, tλ) 6= ∅ for all λ ≥ λ0. Observe that
there are infinitely many different letters in the set of all tλ for λ ≥ λ0.
It follows that L(AE1) is infinite.
26
G. BOAVA, G. DE CASTRO, AND F. MORTARI
For the converse, we use Proposition 2.22, and observe it is enough
to consider x ∈ ξα of the form x = (α, A, α): indeed, if x = (α′, A, α′) ∈
ξα and Z is a finite cover for x observe that α = α′α′′; if we define
x′ = (α, r(A, α′′), α) and Z ′ = {zx′ z ∈ Z} \ {0} (which is a finite
cover for x′), it is easy to see that if Z ′ ∩ ξα 6= ∅ then Z ∩ ξα 6= ∅.
Consider therefore x = (α, A, α) ∈ ξα and le Z be a finite cover for
x. We claim that there exists ∅ 6= D ∈ Bα such that (α, D, α) ∈ Z.
Indeed, if L(AE1) is infinite then there are infinitely many elements of
the form yi = (αbi, r(A, bi), αbi) with distinct bi's in ↓x. For each yi
there exists z = (αβ, C, αβ) ∈ Z such that yiz 6= 0. Since Z is finite
and αβ and αbi are comparable, there must exist (α, D, α) ∈ Z for
some ∅ 6= D ∈ Bα. If there exists B ∈ Bα such that ∅ 6= B ⊆ A ∩ E0
sink
then y = (α, B, α) ∈ ↓x. Since B ⊆ E0
sink, for an element of Z to
intersect y it has to be of the form (α, D, α) ∈ Z for some ∅ 6= D ∈ Bα.
Now let D1, . . . , Dn be all the sets in Bα such that (α, Di, α) ∈ Z.
As we have seen above, n ≥ 1. Observe that D1 ∪ · · · ∪ Dn ∈ ξα.
If not, since ξα is an ultrafilter there would exist C ∈ ξα such that
C ∩ (D1 ∪ · · · ∪ Dn ∈ ξα) = ∅. Now A ∩ C ∈ ξα so that it satisfies (a)
or (b). Arguing as above it can be shown that there exists D such that
(α, D, α) is in Z and intersects (α, A ∩ C, α) which is a contradiction.
Since ξα is an ultrafilter, it is a prime filter so that Di ∈ ξα for some
i = 1, . . . , n. It follows that ξα ∩ Z 6= ∅.
(cid:3)
Note that, in using the proposition above, it is enough to verify
condition (b) only in the case L(AE1) is finite.
Joining Corollary 6.2 and the previous proposition, we get the main
result of this section.
Theorem 6.7. Let (E, L, B) be a labelled space which is weakly left
resolving and whose accommodating family B is closed under relative
complements, and let S be its associated inverse semigroup. Then the
tight filters in E(S) are:
(i) The ultrafilters of infinite type.
(ii) The filters of finite type ξα such that ξα is an ultrafilter in Bα
and for each A ∈ ξα at least one of the following conditions hold:
(a) L(AE1) is infinite.
(b) There exists B ∈ Bα such that ∅ 6= B ⊆ A ∩ E0
sink.
Example 6.8. Let E = (E0, E1, r, s) be any graph and define a la-
belled graph by setting A = E1 and L : A → E1 the identity map.
Let B be the sets of all finite subsets of E0, then B is an accommo-
dating family for (E, L) that is closed under relative complements. We
INVERSE SEMIGROUPS AND LABELLED SPACES
27
prove that for the labelled space (E, L, B), Ftight is homeomorphic to
the boundary path space of the graph ∂E as in [13].
Recall that ∂E = E∞ ∪ {λ ∈ E∗ r(λ) is singular}. On E∗ ∪ E∞ we
give the initial topology induced by the map θ : E∗ ∪ E∞ → {0, 1}E∗
defined by θ(λ)(µ) = 1 if µ is a beginning of λ and 0 otherwise. We
then give the subspace topology on ∂E.
Now, observe that for α ∈ L≥1, Bα = {∅, r(α)} so that it admits
only one ultrafilter, which is {r(α)}. And in Bω = B the ultrafilters
are exactly the principal filters ↑B{v} where v ∈ E0. It follows that for
each α ∈ L≥1 ∪ L∞ there is only one filter ξ in E with word α. For the
empty word the filters ξω such that ξ0 is an ultrafilter are in a bijective
correspondence with vertices v ∈ E0.
By Theorem 6.7 we see that the tight filters are those associated
with an infinite word α ∈ L∞, those associated with α ∈ L≥1 such
that r(α) is singular or ξω where ξ0 is the principal filter generated
by a singular vertex. It follows that there is a bijection between Ftight
and ∂E which is a homeomorphism since both topologies are essentially
given by pointwise convergence.
Proposition 6.9. Let (E, L) be a left resolving labelled graph such
that E0 is a finite set and let B = P(E0). Then the tight spectrum Ftight
of the labelled space (E, L, B) is homeomorphic to the boundary path
space ∂E of the graph E.
n = ↑Bα1,n
Proof. Let λ = λ1λ2 · · · λm be a finite path in ∂E, set α = L(λ)
and, for each 0 ≤ n ≤ m, let ξα
{r(λn)} (which for the given
accomodating family means simply the set of all subsets of r(α1,n)
that contain the element r(λn)). When n = 0, we set λ0 = s(λ).
Similarly, if λ = λ1λ2 · · · is an infinite path in ∂E, set α = L(λ) and
let ξα
n = ↑Bα1,n {r(λn)} for n ≥ 0. In both cases, we can consider the
pair (α, {ξα
n }n≥0) associated with λ.
We claim that {ξα
(with n < λ if λ is finite) and let A ∈ ξα
r(λn+1) ∈ r(A, αn+1), hence
n }n≥0 is a complete family for α. Indeed, fix n ≥ 0
n . Since r(λn) ∈ A, then
ξα
n ⊆ {A ∈ Bα1,n r(A, αn+1) ∈ ξα
n+1}.
On the other hand, let A ∈ Bα1,n be such that r(A, αn+1) ∈ ξα
n+1.
Therefore there exists v ∈ A and e ∈ E1 such that s(e) = v, r(v) =
r(λn+1) and L(e) = αn+1. The labelled graph is left resolving, so we
must have e = λn+1. This says r(λn) = s(λn+1) = v ∈ A, showing that
{A ∈ Bα1,n r(A, αn+1) ∈ ξα
n+1} ⊆ ξα
n
28
G. BOAVA, G. DE CASTRO, AND F. MORTARI
and that the family {ξα
that ξα
n is an ultrafilter in Bα1,n for all n.
n }n≥0 is complete for α, as claimed. Also, observe
n }n≥0).
Let ξα be the filter in E(S) associated with the pair (α, {ξα
We show next that ξα is tight. There are three cases to consider.
Case 1: λ is an infinite path. In this case, ξα is a filter of infinite type
and it is clear that it is an ultrafilter. It follows from Theorem 6.7 that
ξα is tight.
Case 2: λ is a finite path and s−1(r(λ)) is an infinite set. In this case,
ξα is a filter of finite type. Let A ∈ ξα
α. Using that r(λ) ∈ A, that
the labelled graph is left resolving and that E0 is finite, we see that
L(AE1) = ∞. It follows that ξα is tight, by Theorem 6.7.
Case 3: λ is a finite path and r(λ) ∈ E0
finite type. For every A ∈ ξα
and ∅ 6= B ⊆ A ∩ E0
sink. In this case, ξα is a filter of
α, letting B = {r(λ)} we have that B ∈ B
sink. Again Theorem 6.7 ensures ξα is tight.
We can thus define
Φ : ∂E −→ Ftight
λ 7−→ ξα
.
By construction, Φ is clearly injective. Let us show surjectivity:
let
ξα ∈ Ftight. For each n ≥ 0 (with n ≤ α if ξα is of finite type)
the set r(α1,n) is finite, and thus all ultrafilters in Bα1,n = P(r(α1,n))
are principal; in particular, there exists vn ∈ r(α1,n) such that ξα
n =
↑Bα1,n
{vn}.
Observe that for each n ≥ 1, there exists a unique λn ∈ E1 such that
s(λ) = vn−1, r(λn) = vn and L(λ) = αn. Indeed, using that {ξα
n }n≥0
is a complete family, we must have r({vn−1}, αn) ∈ ξα
n . Hence, there
is λn ∈ E1 such that s(λ) = vn−1, r(λn) = vn and L(λ) = αn. The
uniqueness of λn follows from the fact that the labelled graph is left
resolving.
Thus, we have constructed a path λ = λ1λ2 · · · such that L(λ) = α.
Let us see that λ ∈ ∂E: if α = ∞, there is nothing to do. Suppose
If L(AE1) = ∞, then
α < ∞ and consider A = {r(λ)} = {vα}.
If L(AE1) < ∞, then
s−1(r(λ)) is clearly infinite, whence λ ∈ ∂E.
there is B ∈ B such that ∅ 6= B ⊆ A ∩ E0
sink. Since A is unitary, that
means B = A = {r(λ)} and r(λ) ∈ E0
sink, which shows that λ ∈ ∂E.
By construction, we have Φ(λ) = ξα, establishing that Φ is a bijection.
To see that Φ is a homeomorphism, observe that under the condi-
tions stated above, a net {ξ(j)}j∈J converges to ξ in Ftight if and only if
for all (β, {v}, β) ∈ E(S), there exists j0 ∈ J such that for all j ≥ j0,
(β, {v}, β) ∈ ξ(j) ⇐⇒ (β, {v}, β) ∈ ξ.
INVERSE SEMIGROUPS AND LABELLED SPACES
29
Also observe that a net {λ(j)}j∈J in ∂E converges to λ ∈ ∂E if and
only if for all µ ∈ E∗ there exists j0 ∈ J such that for all j ≥ j0, λi
begins with µ if and only if λ begins with µ (see example 6.8 above).
In the same way as was done above, the map between E∗ and
the set {(β, A, β) ∈ E(S) A = {v} is singleton} given by λ 7→
(L(λ), {r(λ)}, L(λ)) is a bijection. From these observations, it is not
hard to see that Φ is a homeomorphism.
(cid:3)
Example 6.10. This example shows that, under the hypotheses
of the previous result, if we consider an accommodating family other
than P(E0) then the boundary path space of the graph may no longer
be identified with the tight spectrum as was done in the proof above.
For instance, consider the labelled graph below:
2
a
a
1
a
a
3
4
It is readily verified that
B = {∅, {1}, {3}, {1, 3}, {2, 4}, {1, 2, 4}, {2, 3, 4}, {1, 2, 3, 4}}
is an accommodating family for the given labelled graph, that is closed
under relative complements, and that the resulting labelled space is
weakly left resolving (indeed, the labelled graph is left resolving, even).
We can see that for each n ≥ 0, the boundary path space of the
underlying graph has two elements of length n. Let us show that, for
the only labelled path of length n, there is a single tight filter associated
with it.
Note that the only ultrafilters in B are G1 = ↑B{1}, G2 = ↑B{2, 4}
and G3 = ↑B{3}. Also, r(α) = E0 and therefore Bα = B for all α ∈ L∗.
Finally, for n ≥ 0 observe that
fan[a](G1) = G2, fan[a](G2) = G1, fan[a](G3) = G1.
From this and using Theorem 6.7, we see that the tight filters for
this labelled space are classified as follows:
There are two ultrafilters of infinite type. The first is associated
with the complete family {F n}n where F n equals G1 if n is odd and
G2 if n is even. The second ultrafilter is the opposite: F n equals G1 if
n is even and G2 if n is odd.
There is a single tight filter for each α ∈ L∗. This happens for
two reasons: first, condition (ii)(a) of Theorem 6.7 does not occur here
30
G. BOAVA, G. DE CASTRO, AND F. MORTARI
since the alphabet A consists of a single letter; second, G3 is the only
ultrafilter in B that satisfies condition (ii)(b).
Thus, a filter ξα in E(S) is tight if and only if ξα = G 3. Addition-
ally, if α = n + 1 (that is, α = an+1) with n ≥ 0 then for each k ≤ n
we have that
ξk = (cid:26) G1,
G2,
if n and k have different parities,
if n and k have the same parity.
The forced relations between the ultrafilters that produce the com-
plete families can be seen with the following directed graph, (where the
arrows indicate maps of type far[a], r ≥ 0):
G2
G1
G3
Example 6.11. In this example, we show that the tight spectrum
of a labelled space does not depend only on its labelled paths. Consider
the two labelled graphs (E1, L1) and (E2, L2) given respectively by
1
v1
0
0
v2
and
1
0
0
v2
v1
1
v3
0
It is easy to see that both labelled graphs produce the same la-
belled paths. As accommodating families for each, choose B1 = P(E0
1)
and B2 = P(E0
2), respectively. Proposition 6.9 says the tight spec-
tra of (E1, L1, B1) and (E2, L2, B2) are homeomorphic to ∂E1 and ∂E2,
respectively.
However, ∂E1 has no isolated points, as can be readily verified,
whereas the infinite path determined by the loop at vertex v3 is an
isolated point for ∂E2; therefore, they are not homeomorphic.
References
[1] T. Bates, T. Meier Carlsen, and D. Pask. C ∗-algebras of labelled graphs III -
K-theory computations. ArXiv e-prints, Mar. 2012.
[2] T. Bates and D. Pask. C ∗-algebras of labelled graphs. J. Operator Theory,
57(1):207 -- 226, 2007.
INVERSE SEMIGROUPS AND LABELLED SPACES
31
[3] T. Bates and D. Pask. C ∗-algebras of labelled graphs. II. Simplicity results.
Math. Scand., 104(2):249 -- 274, 2009.
[4] T. Bates, D. Pask, and P. Willis. Group actions on labeled graphs and their
C*-algebras. ArXiv e-prints, June 2011.
[5] S. Burris and H. P. Sankappanavar. A course in universal algebra, volume 78
of Graduate Texts in Mathematics. Springer-Verlag, New York-Berlin, 1981.
[6] H. P. Doctor. The categories of Boolean lattices, Boolean rings and Boolean
spaces. Canad. Math. Bull., 7:245 -- 252, 1964.
[7] R. Exel. Inverse semigroups and combinatorial C ∗-algebras. Bull. Braz. Math.
Soc. (N.S.), 39(2):191 -- 313, 2008.
[8] C. Farthing, P. S. Muhly, and T. Yeend. Higher-rank graph C ∗-algebras: an
inverse semigroup and groupoid approach. Semigroup Forum, 71(2):159 -- 187,
2005.
[9] A. E. Marrero and P. S. Muhly. Groupoid and inverse semigroup presentations
of ultragraph C ∗-algebras. Semigroup Forum, 77(3):399 -- 422, 2008.
[10] A. L. T. Paterson. Graph inverse semigroups, groupoids and their C ∗-algebras.
J. Operator Theory, 48(3, suppl.):645 -- 662, 2002.
[11] J. Renault. A groupoid approach to C ∗-algebras, volume 793 of Lecture Notes
in Mathematics. Springer, Berlin, 1980.
[12] M. H. Stone. Postulates for Boolean Algebras and Generalized Boolean Alge-
bras. Amer. J. Math., 57(4):703 -- 732, 1935.
[13] S. B. G. Webster. The path space of a directed graph. Proc. Amer. Math. Soc.,
142(1):213 -- 225, 2014.
Departamento de Matem´atica, Universidade Federal de Santa Cata-
rina, 88040-970 Florian´opolis SC, Brazil.
E-mail address: [email protected], [email protected],
[email protected]
|
1002.2752 | 2 | 1002 | 2010-06-16T19:55:31 | An octonion algebra originating in combinatorics | [
"math.RA",
"math.CO"
] | C.H. Yang discovered a polynomial version of the classical Lagrange identity expressing the product of two sums of four squares as another sum of four squares. He used it to give short proofs of some important theorems on composition of delta-codes (now known as T-sequences). We investigate the possible new versions of his polynomial Lagrange identity. Our main result shows that all such identities are equivalent to each other. | math.RA | math |
AN OCTONION ALGEBRA ORIGINATING IN
COMBINATORICS
DRAGOMIR Z. D -- OKOVI ´C AND KAIMING ZHAO
Abstract. C.H. Yang discovered a polynomial version of the clas-
sical Lagrange identity expressing the product of two sums of four
squares as another sum of four squares. He used it to give short
proofs of some important theorems on composition of δ-codes (now
known as T -sequences). We investigate the possible new versions
of his polynomial Lagrange identity. Our main result shows that
all such identities are equivalent to each other.
2000 Mathematics Subject Classification 17A75, 05B30
1. Introduction
C.H. Yang [8, 9, 10] discovered a polynomial version (see below)
of the classical Lagrange identity on the product of two sums of four
squares. He used his identity to give short and elegant proofs of some
important theorems [10] in combinatorics of binary and ternary se-
quences. These results provide new methods for the construction of
several classes of combinatorial objects such as T -sequences, orthogo-
nal designs, and Hadamard matrices [6, 10]. Our motivation and the
main goal was to investigate the possible new versions of the polyno-
mial Lagrange identity. However, our main result shows that all such
identities are equivalent to each other (for the precise statement see
Theorem 4.1).
Let A = Z[z, z−1] be the Laurent polynomial ring over the integers
Z. For any f = f (z) = Pk akzk ∈ A, ak ∈ Z, we define its conjugate
as f ∗ = f (z−1). We also say that a0 is the constant term of f and write
CT(f ) = a0. Note that CT(f f ∗) = Pk a2
k. Let A0 be the fixed subring
of A under the conjugation, i.e., A0 = {f ∈ A f = f ∗}. It is easy to
see that A0 = Z[z + z−1]. We embed the polynomial ring Z[t] into A by
sending t → z + z−1 and view A as a Z[t]-algebra. It is easy to check
that A = A0 ⊕ A0z. Thus A is a free Z[t]-module of rank 2.
Key words and phrases. Laurent polynomial ring, polynomial Lagrange identity,
octonion algebra, composition algebra.
The first author was supported by an NSERC Discovery Grant, and the second
by the NSERC and the NSF of China (Grant 10871192).
1
2
D.Z. D -- OKOVI ´C AND K. ZHAO
We are mainly interested in the free A-module E = A4. When
viewed as a Z[t]-module it is again free but now its rank is 8. For any
x = (x0, x1, x2, x3) ∈ E we define its norm as N(x) = Pk xkx∗
k. Thus
N : E → A0 is a quadratic form on the A0-module E.
Now we state the Lagrange identity for Laurent polynomials (mod-
ified Theorem 1 in [10]). Let x = (xk), y = (yk) ∈ E and define
(p, q, r, s) ∈ E by Yang's formulae:
p = x0y0 − x1y∗
q = x0y1 + x1y∗
1y∗
r = x0y2 − x∗
s = x0y3 + x∗
1y∗
1 − x2y∗
2y∗
0 + x∗
3 + x2y∗
2 − x∗
2y∗
2 − x3y∗
3,
3y∗
3 − x∗
2,
3y∗
0 + x∗
1,
1 + x3y∗
0.
We can use these formulae to define an A0-bilinear multiplication "◦"
on E by x ◦ y = (p, q, r, s).
It is straightforward to verify that the
Lagrange identity N(x ◦ y) = N(x)N(y) is indeed valid.
We shall see in the next section that (E, ◦) is in fact an octonion
algebra over Z[t]. In Section 3 we give an explicit description of the
orthogonal group O(N) of the pair (E, N), see Theorem 3.4. In Section
4, we show that all A0-bilinear multiplications on E satisfying the La-
grange identity are equivalent, in the sense defined there, to the above
multiplication "◦".
Our main result can probably be generalized by replacing Z with a
more general commutative ring. We decided not to pursue this here in
order to preserve the essentially combinatorial flavour of the original
problem.
We are grateful to the referee for correcting a couple of errors in the
original proof of Lemma 4.3, for giving the stronger result in Theorem
4.3, and for his other detailed comments and suggestions.
2. Yang formulae define an octonion algebra
We make A = Z[z, z−1] into a Z[t]-algebra via the homomorphism
Z[t] → A sending t → z + z−1. We shall often identify Z[t] with its
image A0 under this homomorphism. According to the definition in
[1, Chapter III, §2], A is a quadratic Z[t]-algebra. By using the basis
{1, z}, we see that its type is (−1, t). Indeed we have z2 = −1 + tz.
Moreover, (A, ∗) is an example of a Cayley algebra (see loc. cit.).
Let H = A × A, a free A-module of rank 2. We shall also view it
as a free Z[t]-module of rank 4 with basis (1, 0), (z, 0), (0, 1), (0, z).
We make H into an associative noncommutative algebra by using the
Cayley -- Dickson process, i.e., we define the multiplication in H by
(a, b)(c, d) = (ac − d∗b, bc∗ + da),
a, b, c, d ∈ A.
AN OCTONION ALGEBRA
3
The involution "∗" on A extends to an involutory anti-automorphism
of H by setting
(a, b)∗ = (a∗, −b),
a, b ∈ A.
Thus H is an example of a quaternion algebra over Z[t], see loc. cit.
No. 5, Example 2. As such, it has type (−1, t, −1).
Let E = H × H, a free A-module of rank 4. We shall also view it as
a free Z[t]-module of rank 8 with basis
e0 = (1, 0, 0, 0),
e1 = (0, 1, 0, 0),
e2 = (0, 0, 1, 0),
e3 = (0, 0, 0, 1),
e′
0 = ze0 = (z, 0, 0, 0),
e′
1 = ze1 = (0, z, 0, 0),
e′
2 = ze2 = (0, 0, z, 0),
e′
3 = ze3 = (0, 0, 0, z).
We make E into a nonassociative algebra by using the Cayley -- Dickson
process once more. Thus, we define the multiplication in E by
(u, v)(x, y) = (ux − y∗v, vx∗ + yu),
u, v, x, y ∈ H.
The involution "∗" on H extends to one on E by setting
(u, v)∗ = (u∗, −v),
u, v ∈ H.
Thus E is an example of an octonion algebra over Z[t], see [1, Chapter
III, Appendix]. By using the above basis, we find that the type of this
octonion algebra is (−1, t, −1, −1).
Let us write the above octonion multiplication in terms of the A-
basis {e0, e1, e2, e3}. For x = (x0, x1, x2, x3) and y = (y0, y1, y2, y3), we
find that
1 − x2y∗
0 + x∗
xy = (x0y0 − x1y∗
3y3,
2y3 − x3y∗
x0y1 + x1y∗
2,
0 + x3y∗
x0y2 − x∗
1,
x∗
0y3 + x1y2 − x2y1 + x3y0).
2 − x∗
1y3 + x2y∗
Hence, the map sending (x0, x1, x2, x3) → (x0, x1, x2, x∗
3) is an isomor-
phism of this octonion algebra with the algebra (E, ◦) defined by the
Yang formulae.
3. Orthogonal group
It is obvious that U1 = {x ∈ A xx∗ = 1} is the group of invertible
elements of A. It consists of the elements ±zk, k ∈ Z. In our proofs
below we shall often use the following obvious fact: The subset U1
generates A as an additive group.
4
D.Z. D -- OKOVI ´C AND K. ZHAO
The ring homomorphisms ϕ : A → C are parametrized by nonzero
complex numbers w: By definition, the homomorphism ϕw correspond-
ing to w sends the indeterminate z to w. The homomorphism ϕw is
compatible with the involutions (conjugation on A and complex conju-
gation on C), i.e, ϕw(x∗) = ϕw(x) for all x ∈ A, if and only if w = 1.
The following lemma will be used in the next section.
Lemma 3.1. If m ∈ Z is not a square, then the equation xx∗ = m has
no solution in A.
Proof. Assume that f f ∗ = m for some f ∈ A. By applying the homo-
morphism ϕ1 we obtain f (1)f ∗(1) = m. As f ∗(1) = f (1) ∈ Z, we have
a contradiction.
(cid:3)
We also remark that, for x ∈ A, CT(xx∗) = 0 implies x = 0, and
CT(xx∗) = 1 implies xx∗ = 1.
Denote the general linear group of the A0-module E by GL(A0, E).
We introduce the orthogonal groups
O(N) = {ϕ ∈ GL(A0, E) N(ϕ(u)) = N(u), ∀u ∈ E},
O4(Z) = {ϕ ∈ GL4(Z) N(ϕ(u)) = N(u), ∀u ∈ Z4},
where Z4 is considered as an additive subgroup of E = A4.
Let Q : E × E → A0 be the polar form of N, i.e.,
Q(x, y) = N(x + y) − N(x) − N(y)
=
3
X
k=0
(x∗
kyk + xky∗
k) .
Note that Q is a symmetric A0-bilinear O(N)-invariant form, and that
Q is nondegenerate, i.e., its kernel is 0.
Let us begin with a useful remark.
Remark 3.2. An A0-linear map ϕ : E → E preserving N is automati-
cally bijective, and so ϕ ∈ O(N). Indeed, as Q is nondegenerate, the
injectivity is obvious.
If S is the matrix of Q with respect to some
A0-basis of E, then det(ϕ)2 det(S) = det(S) 6= 0, forcing det(ϕ) = ±1,
and ϕ is bijective.
The following result is crucial to this paper.
Theorem 3.3. (a) The group O4(Z) consists of entry permutations on
Z4 with arbitrary sign changes.
(b) The set U ′
(c) The unit sphere U4 = {x ∈ E N(x) = 1} is equal to ∪3
(d) The sphere U ′
1 = {x ∈ A xx∗ = −(z−z−1)2} is equal to (z−z−1)U1.
k=0U1ek.
4 = {x ∈ E N(x) = −(z − z−1)2} is equal to
(z − z−1)U4.
AN OCTONION ALGEBRA
5
Proof. (a) Fix α ∈ O4(Z). Let α(ei) = Pj aijej where aij ∈ Z. From
1 = N(α(ei)) = Pj a2
ij we see that, for each i, exactly one of aij is ±1
and the other ones vanish. The rest follows easily.
(b) By applying the homomorphisms ϕ±1 to xx∗ = −(z − z−1)2,
we conclude that x is divisible by z − 1 and z + 1, and so we have
x = (z − z−1)y for some y ∈ A. As xx∗ = −(z − z−1)2yy∗, we deduce
that yy∗ = 1, i.e., y ∈ U1.
(c) Suppose x = (x0, x1, x2, x3) ∈ U4. Then P3
k = 1. By
comparing the constant terms, we conclude that exactly one of the xk
belongs to U1 while the other vanish.
k=0 xkx∗
(d) Suppose x = (xk) ∈ U ′
k = 2 − z2 − z−2. Hence
Pk CT(xkx∗
k) = 2, which by the above remark implies that exactly one
of the xks is nonzero. This nonzero component belongs to (z − z−1)U1
by (b).
(cid:3)
4. Then Pk xkx∗
Note that O4(Z) = E16Σ4 is the semidirect product of the elementary
abelian group E16 of order 16 which acts on Z4 by multiplying the
coordinates with ±1 and the symmetric group Σ4 of degree 4 which
permutes the coordinates.
For u = (u0, u1, u2, u3) ∈ U 4
1 we define σu ∈ O(N) by σu(x) = (ukxk),
where x = (xk) ∈ E. Then
(3.1)
T = {σu u ∈ U 4
1 }
is an abelian subgroup of O(N).
Any α ∈ O4(Z) extends uniquely to an A-linear element of O(N),
still denoted by α. Thus we consider O4(Z) as a subgroup of O(N).
Define τ0 ∈ O(N) by τ0(x0, x1, x2, x3) = (x∗
0, x1, x2, x3), and define
similarly τk for k = 1, 2, 3. Let Γ denote the group of order 16 generated
by the τks. The group O(N) admits the following factorization.
Theorem 3.4. O(N) = T Σ4Γ.
Proof. Let ϕ ∈ O(N). Since ϕ(U4) = U4, for any index i there exist
ui ∈ U1 and an index i′ such that ϕ(ei) = uiei′. Since the form Q is
ϕ-invariant, the map i → i′ must be a permutation of {0, 1, 2, 3}.
For any a ∈ A we have a + a∗ ∈ A0 and so
ϕ((a + a∗)ei) = (a + a∗)uiei′.
Now let a ∈ U1, a 6= ±1. Then each of the terms ϕ(aei), ϕ(a∗ei), auiei′
and a∗uiei′ belongs to U4, and the sum of the first two terms is not 0.
It follows that ϕ(aei) is equal to auiei′ or a∗uiei′.
Assume that there exist a, b ∈ U1 different from ±1 such that ϕ(aei)
= auiei′ while ϕ(bei) = b∗uiei′. Then ϕ((a + b)ei) = (a + b∗)uiei′, and
by taking the norms we obtain the contradiction (a − a∗)(b − b∗) = 0.
6
D.Z. D -- OKOVI ´C AND K. ZHAO
Consequently, one of the following two identities: ϕ(aei) = auiei′ or
ϕ(aei) = a∗uiei′ must hold for all a ∈ A. Equivalently, there exists
εi ∈ {0, 1} such that ϕτ εi
i (aei) = auiei′ for all a ∈ A. Let
u = (ui) ∈ U 4
1
and τ =
3
Y
i=0
τ εi
i ∈ Γ.
The composite β = σ−1
u ϕτ is A-linear and we have σ−1
all i. Hence, ϕ = σuβτ with σu ∈ T , β ∈ Σ4 and τ ∈ Γ.
u ϕτ (ei) = ei′ for
(cid:3)
4. Composition algebra structures on E
In this section we determine all A0-bilinear multiplications "·" on E
which satisfy the (polynomial) Lagrange identity
(4.1)
N(x · y) = N(x)N(y).
Since E is a free A0-module of rank 8 and the form Q is nondegenerate,
such algebra (E, ·) will be a composition algebra in the sense of the
definition in [4, p. 305] provided that it has an identity element.
We say that two A0-bilinear multiplications ⋆ and ⋄ are equivalent if
there exist σ1, σ2, τ ∈ O(N) such that x ⋄ y = τ (σ1(x) ⋆ σ2(y)) for all
x, y ∈ E.
The following theorem is our main result.
Theorem 4.1. Any A0-bilinear multiplication "·" on E satisfying the
Lagrange identity (4.1) is equivalent to the multiplication "◦" defined
by Yang's formulae.
Proof. We shall reduce the proof to the special case where e0 is the
identity element of the A0-algebra (E, ·). Then the assertion of the
theorem follows from the next theorem.
Define A0-linear maps L, R : E → E by L(u) = e0 · u and R(u) =
u · e0. The Lagrange identity implies that L and R preserve N. By the
remark 3.2, L, R ∈ O(N).
Next we use a well known argument due to, at least, Kaplansky
[3]. We have N(R−1(x) · L−1(y)) = N(x)N(y) for x, y ∈ E. With
x ⋆ y = R−1(x) · L−1(y), we have
(e0 · e0) ⋆ x = R−1(e0 · e0) · L−1(x) = e0 · L−1(x) = x,
x ⋆ (e0 · e0) = R−1(x) · L−1(e0 · e0) = R−1(x) · e0 = x.
Thus e0 · e0 is the identity element of the algebra (E, ⋆). As e0 · e0 ∈ U4,
we have e0 · e0 = aei for some a ∈ U1 and some index i. Hence, there
exists σ ∈ O(N) such that σ(e0) = e0 · e0. If x ⋄ y = σ−1(σ(x) ⋆ σ(y)),
then e0 is the identity element of the algebra (A, ⋄).
(cid:3)
AN OCTONION ALGEBRA
7
We assume now that e0 is the identity element of (E, ·). Conse-
quently, this is a composition algebra and the alternative laws
(4.2)
x · (x · y) = (x · x) · y,
(x · y) · y = x · (y · y)
are valid [4, p. 306]. The reader should consult this book for additional
properties of composition algebras. We state only a few properties that
we need.
Define the A0-linear map trace T : E → A0 by T (x) = Q(e0, x).
Then for any x ∈ E we have
(4.3)
x · x − T (x)x + N(x)e0 = 0.
Linearizing gives
(4.4)
x · y + y · x = T (x)y + T (y)x − Q(x, y)e0, ∀ x, y ∈ E.
Clearly Ae0 ⊥ A3 with respect to Q, where A3 = Ae1 + Ae2 + Ae3.
Recall that we have extended the conjugation ∗ from A to E in
Section 2. Moreover, note that
(ae0 + v)∗ = a∗e0 − v,
a ∈ A, v ∈ A3.
It is easy to verify that
(4.5)
Q(x, z · y∗) = Q(x · y, z) = Q(y, x∗ · z), ∀x, y, z ∈ E.
The fact that in the next theorem one can assert that the two alge-
bras are actually isomorphic is due to the referee.
Theorem 4.2. Every A0-algebra (E, ·) satisfying the Lagrange identity
(4.1) and having e0 as the identity element is isomorphic to the Yang
algebra (E, ◦).
We shall break the proof into several lemmas. 1
Lemma 4.3. By replacing "·" with an isomorphic multiplication, we
may also assume that
(4.6)
(ae0) · y = ay
for all a ∈ A and y ∈ E.
Proof. Let us fix an index i and b ∈ U1. Then the right multiplication
by bei belongs to O(N), and Theorem 3.4 yields an index j as well as
uj ∈ U1 such that (ae0) · (bei) = ujaej for all a ∈ A or (ae0) · (bei) =
uja∗ej for all a ∈ A. Specializing a to 1 implies that j = i and uj = b,
hence (ae0) · (bei) = abei for all a ∈ A or (ae0) · (bei) = a∗bei for all
a ∈ A.
1For a simpler proof see Section 5.
8
D.Z. D -- OKOVI ´C AND K. ZHAO
Assume that (ae0) · (b1ei) = ab1ei holds for all a ∈ U1 and (ae0) ·
(b2ei) = a∗b2ei holds for all a ∈ U1, for some units b1, b2 different from
±1. Then (ae0) · ((b1 + b2)ei) = (ab1 + a∗b2)ei holds for all a ∈ U1. By
taking norms on both sides, we obtain that
(1 − a2)b1b∗
2 + (1 − (a∗)2)b∗
1b2 = 0
for all a ∈ U1. This is clearly a contradiction.
Since U1 generates A as an abelian group, we conclude that either
(ae0) · (bei) = abei for all a, b ∈ A or (ae0) · (bei) = a∗bei for all a, b ∈ A.
In particular, (ae0) · (bei) is A-linear in b.
If (ae0) · (be0) = a∗be0 for all a, b ∈ A, then setting b = 1 gives the
contradiction: ae0 = a∗e0 for all a ∈ A. Thus we have (ae0) · (be0) =
abe0 for all a, b ∈ A. For i 6= 0, there is an εi ∈ {0, 1} such that
i (τ εi
τ εi
If τ is the product of the τ εi
i with i 6= 0, then τ 2 = 1 and the
multiplication "⋆" defined by x ⋆ y = τ (τ (x) · τ (y)) satisfies all the
requirements.
i (bei)) = abei for all a, b ∈ A.
i (ae0) · τ εi
Moreover, the map τ : (E, ⋆) → (E, ·) is an isomorphism of unital
(cid:3)
A0-algebras.
We assume from now on that the identity (4.6) holds. This property
will be shared by the modified multiplication which will be introduced
in Lemma 4.6.
Lemma 4.4. For a ∈ A and x ∈ Ae1 + Ae2 + Ae3 we have x · (ae0) =
a∗x.
Proof. Recall that Ae1 + Ae2 + Ae3 is the kernel of the trace T and is
perpendicular to Ae0 relative to Q. By using (4.4), we obtain that
x · (ae0) = x · (ae0) + (ae0) · x − (ae0) · x
= T (ae0)x + T (x)ae0 − Q(x, ae0)e0 − ax
= (a + a∗ − a)x = a∗x.
(cid:3)
Lemma 4.5. For a, b ∈ A and i 6= 0 we have (aei) · (bei) = −ab∗e0.
Proof. Since N is anisotropic, it suffices to show that N((aei) · (bei) +
ab∗e0) = 0. By expanding the LHS and applying (4.5) gives
N((aei) · (bei) + ab∗e0) = 2N(a)N(b) + Q((aei) · (bei), ab∗e0)
= 2N(a)N(b) + Q(aei, (ab∗e0) · (bei)∗)
= 2N(a)N(b) − Q(aei, (ab∗e0) · (bei))
= 2N(a)N(b) − N(b)Q(aei, aei)
= 2N(a)N(b) − 2N(a)N(b) = 0.
AN OCTONION ALGEBRA
9
(cid:3)
Lemma 4.6. By replacing "·" with an isomorphic multiplication, we
may assume that
(4.7)
ei · ej = ek = −ej · ei
for any cyclic permutation (i, j, k) of (1, 2, 3).
Proof. Let (i, j, k) be a cyclic permutation of (1, 2, 3). Since (e0 + ei) ·
(e0 + ej) = e0 + ei + ej + ei · ej and N(e0 + ej) = N(e0 + ej) = 2, we
have N(e0 + ei + ej + ei · ej) = 4. As ei · ej ∈ U4, Lemma 3.1 implies
that ei · ej ∈ U1ek. A similar argument shows that ej · ei ∈ U1ek.
In particular we have e1 · e2 = ue3 for some u ∈ U1. The multiplica-
tion "⋆" defined by x ⋆ y = σ−1
a (σa(x) · σa(y)), where a = (1, 1, 1, u), has
the previously stated property and sends (e1, e2) → e3. By replacing
"·" with this new multiplication, we may assume that e1 · e2 = e3.
Since (ei + ej) · (ei + ej) = ei · ej + ej · ei − 2e0, we have N(ei · ej +
ej · ei − 2e0) = 4. Since ei · ej + ej · ei ∈ Aek, it follows that this sum is
0, i.e., ej · ei = −ei · ej. In particular, we have e2 · e1 = −e3. Thus the
assertion of the lemma is valid if (i, j, k) = (1, 2, 3).
The remaining equalities (4.7) follow easily by using the alternative
laws (4.2). For instance, we have e1 · e3 = e1 · (e1 · e2) = (e1 · e1) · e2 =
(−e0) · e2 = −e2.
To finish the proof, we point out that the map σa : (E, ⋆) → (E, ·)
(cid:3)
is an isomorphism of unital A0-algebras.
In view of the last lemma, we may assume now (and we do) that the
identities (4.7) are valid for any cyclic permutation (i, j, k) of (1, 2, 3).
Lemma 4.7. For any cyclic permutation (i, j, k) of (1, 2, 3) we have
(4.8)
(aei) · (bej) = a∗b∗ek = −(bej) · (aei)
for all a, b ∈ A. Consequently, the multiplications "·" and "◦" coincide.
Proof. In view of (4.4), it suffices to prove the first equality. We first
treat the case a = 1. By applying the linearized left alternative law,
we find that
ei · (bej) = ei · ((be0) · ej)
= (ei · (be0) + (be0) · ei) · ej − (be0) · (ei · ej)
= (b + b∗)ek − bek = b∗ek.
10
D.Z. D -- OKOVI ´C AND K. ZHAO
For arbitrary a, by using this special case, the linearized right alterna-
tive law and Lemma 4.4, we obtain that
(aei) · (bej) = (ei · (a∗e0)) · (bej)
= ei · ((a∗e0) · (bej) + (bej) · (a∗e0)) − (ei · (bej)) · (a∗e0)
= ei · ((a + a∗)bej) − (b∗ek) · (a∗e0)
= ((a + a∗)b∗ − ab∗)ek = a∗b∗ek.
(cid:3)
This concludes the proof of Theorem 4.2, and also of Theorem 4.1.
The referee supplied an alternative proof of Theorem 4.2. His proof
proceeds first by changing scalars from the base ring Z[t] to its quo-
tient field Q(t), and then applying a theorem of Thakur [7] (see also
[5, Chapter VIII, Exercise 6]) which establishes a connection between
octonion algebras and ternary hermitian forms. In the case of Yang
algebra, let us write E = Ae0 ⊕ A3, where A3 is the column A-space.
Extending the conjugation of A componentwise to A3, we have a her-
mitian form h : A3 × A3 → A given by h(v, w) = vtw∗. With this
notation, the Yang multiplication is given by
(ae0 ⊕ v) ◦ (be0 ⊕ w) = (ab − h(v, w))e0 ⊕ (aw + b∗v + v∗ × w∗)
where a, b ∈ A and v, w ∈ A3. The sign × stands for the ordinary cross
product in 3-space.
5. Addendum
Alberto Elduque [2] has found a simpler proof of Theorem 4.2. Here
we present his simplification. We point out that the involution "∗"
coincides with the the involution "−" in [4, p. 305] and so it is an
involution of the algebra (E, ·).
The first paragraph of the proof of Lemma 4.3 implies that Ae0 · ei =
Aei for each i. By applying the involution "∗" we obtain that also
ei · Ae0 = Aei for each i. By (4.3) we have (ze0) · (ze0) = −e0 + tze0 =
z2e0. It follows that (xe0) · (ye0) = xye0 for all x, y ∈ A. By (4.3), we
also have e1 · e1 = −e0. By applying [4, Lemma (7.1.7)], with B = Ae0
and v = e1, we conclude that P = Ae0 + Ae1 is a subalgebra of (E, ·).
From the proof of that lemma it also follows that the multiplications
"·" and "◦" coincide on P .
From (4.4) we deduce that e2 · P ⊆ P ⊥ = Ae2 + Ae3. Since x → e2 · x
is an orthogonal transformation on E, we have Ae2 = e2 ·Ae0 ⊥ e2 ·Ae1.
It follows that e2 · Ae1 ⊆ Ae3. Similarly, e2 · Ae3 ⊆ Ae1. Thus we must
have e2 · Ae1 = Ae3 and e2 · Ae3 = Ae1, and so E = P ⊕ e2 · P . We
now apply [4, Lemma (7.1.7)] for the second time by taking B = P
AN OCTONION ALGEBRA
11
and v = e2. We conclude that the multiplications "·" and "◦" coincide
on E.
Note that this proof does not use any of the lemmas 4.4-7, and uses
only the first paragraph of the proof of Lemma 4.3.
References
[1] N. Bourbaki, Alg`ebre, Chapitres 1 `a 3, Hermann, Paris, 1970.
[2] A. Elduque, Private communication, March 2, 2010.
[3] I. Kaplansky, Infinite-dimensional quadratic forms admitting composition,
Proc. Amer. Math. Soc. 4 (1953), 956 -- 960.
[4] M.-A. Knus, Quadratic and Hermitian Forms over Rings, Grundlehren der
mathematischen Wissenschaften 294, Springer-Verlag, 1991.
[5] M.-A. Knus, A. Merkurjev, M. Rost and J.-P. Tignol, The Book of Involutions,
Amer. Math. Soc. Colloq. Publ. 44, Amer. Math. Soc., Providence, RI, 1998.
[6] J. Seberry and M. Yamada, Hadamard matrices, sequences and block designs,
in Contemporary Design Theory: A Collection of Surveys, Eds. J.H. Dinitz
and D.R. Stinson, J. Wiley, New York, 1992, pp. 431 -- 560.
[7] M. L. Thakur, Cayley algebra bundles on A2
K revisited, Comm. Algebra 23
(1995), no. 13, 5119 -- 5130.
[8] C. H. Yang, Lagrange identity for polynomials and δ-codes of lengths 7t and
13t, Proc. Amer. Math. Soc. 88 (1983), 746 -- 750.
[9] C. H. Yang, A composition theorem for δ-codes, Proc. Amer. Math. Soc. 89
(1983), 375 -- 378.
[10] C. H. Yang, On composition of four-symbol δ-codes and Hadamard matrices,
Proc. Amer. Math. Soc. 107 (1989), 763 -- 776.
Department of Pure Mathematics, University of Waterloo, Water-
loo, Ontario, N2L 3G1, Canada
Department of Mathematics, Wilfrid Laurier University, Water-
loo, Ontario, N2L 3C5, Canada and Institute of Mathematics, Acad-
emy of Mathematics and System Sciences, Chinese Academy of Sci-
ences, Beijing 100190, P.R. China
E-mail address: [email protected]
E-mail address: [email protected]
|
1412.5776 | 2 | 1412 | 2018-07-10T15:43:04 | A relational description of higher commutators in Mal'cev varieties | [
"math.RA"
] | We give a relational description of higher commutator operators, which were introduced by Bulatov, in varieties with a Mal'cev term. Furthermore, we use this result to prove that for every algebra with a Mal'cev term there exists a largest clone on the same underlying set containing the Mal'cev operation and having the same congruence lattice and the same higher commutator operators as the original algebra. A local variant of this theorem is given. | math.RA | math | A RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
IN MAL'CEV VARIETIES
JAKUB OPRSAL
Abstract. We give a relational description of higher commutator operators,
which were introduced by Bulatov, in varieties with a Mal'cev term. Further-
more, we use this result to prove that for every algebra with a Mal'cev term
there exists a largest clone containing the Mal'cev operation and having the
same congruence lattice and the same higher commutator operators as the
original algebra. We also give a local variant of this theorem.
8
1
0
2
l
u
J
0
1
]
.
A
R
h
t
a
m
[
2
v
6
7
7
5
.
2
1
4
1
:
v
i
X
r
a
1. Introduction
Two algebras are called polynomially equivalent if they have the same under-
lying set and the same clone of all polynomial operations. One of the invariants
to distinguish polynomially inequivalent algebras is the congruence lattice of the
corresponding algebra, and the binary commutator operation [·, ·] on this lattice
(the theory describing this commutator have been developed in the 80's, and is de-
scribed in the book by Freese and McKenzie [8]). In fact, from the results of Idziak
[10] and Bulatov [6], one can see that on the three-element set, every Mal'cev al-
gebra is up to polynomial equivalence described by its congruence lattice, and the
binary commutator operation. This is no longer true for sets with at least four
elements. But one can generalize the binary commutator operator to higher arities.
These higher arity commutators have been introduced by Bulatov [5]. From the
description of polynomial clones with a Mal'cev operation on the four-element set
[7], one can obtain that every four-element Mal'cev algebra is determined up to
polynomial equivalence by its unary polynomials, congruence lattice, and higher
commutator operators on this lattice. The higher commutators are defined by the
following 'term-condition'.
Definition 1.1 (Bulatov's higher commutator operators). Let α0, . . . , αn−1, and
γ be congruences of some algebra A. We say that α0, . . . , αn−2 centralize αn−1
modulo γ if for all tuples ai, bi, i = 0, . . . , n − 1, and all terms t of A such that
(1) ai 6= bi, but the corresponding entries are congruent modulo αi for all
i ∈ {0, . . . , n − 1}, and
(2) t(x0, . . . , xn−2, an−1) ≡γ t(x0, . . . , xn−2, bn−1) for all (x0, . . . , xn−2) ∈ ({a0, b0}×
· · · × {an−2, bn−2}) \ {(b0, . . . , bn−2)},
Date: July 11, 2018.
2010 Mathematics Subject Classification. 08A40.
Key words and phrases. commutator theory, Mal'cev algebras, clone.
Supported by Austrian Science Fund (FWF), P24077: Algebraic approaches to the description
of Mal'cev clones; Czech Science Foundation, GA CR 13-01832S; and Charles University in Prague,
project SVV-2014-260107.
1
2
we have
J. OPRSAL
t(b0, . . . , bn−2, an−1) ≡γ t(b0, . . . , bn−2, bn−1).
The n-ary commutator [α0, . . . , αn−1] is then defined as the smallest congruence γ
such that α0, . . . , αn−2 centralize αn−1 modulo γ. We define the nulary commutator
to be trivially the full congruence on A, and for the unary commutator of α we put
[α] = α.
One of important notions that came from higher commutators is a notion of su-
pernilpotence: an algebra is k-supernilpotent if it satisfies the commutator identity
[1, 1, . . . , 1
] = 0.
}
k+1
{z
If an algebra is k-supernilpotent for some k we say that is it supernilpotent. For
general algebras supernilpotence is a strictly stronger notion then nilpotence; i.e.,
there is a nilpotent algebra which is not supernilpotent. However, this is not the
case in the variety of groups where both notions coincide. Therefore supernilpotent
algebras can be viewed as natural generalization of nilpotent groups. They also
share several properties with nilpotent groups, in particular a Mal'cev algebra of
finite type is supernilpotent if and only if it is a product of prime power order
supernilpotent algebras [2]. It has been shown in [3] that there are two expansions
of the same group that are both 2-supernilpotent, but the clone given as the join of
their clones is not. In this paper we establish additional properties to ensure that the
join of two k-supernilpotent clones sharing a Mal'cev operation is k-supernilpotent.
To achieve that goal we give a description of higher commutators using a certain
2n-ary relation denoted ∆(α0, . . . , αn−1) (see Definition 3.1). A similar relation
have been also defined in [13]. The relation ∆(α0, . . . , αn−1) encodes the value
of [α0, . . . , αn−1] as its forks at the last coordinate -- by a fork of a relation R at
a coordinate i we mean a pair (a, b) such that there exists c, d ∈ R with ci = a,
di = b, and cj = dj for all j 6= i; and we denote ψi(R), the set of all forks of R at
i. A similar notion has been used to investigate some properties of algebras with
a cube term [4, 1]. The description of higher commutators is then given by the
following theorem.
Theorem 1.2. If A is an algebra with a Mal'cev term, and α0, . . . , αn−1 are con-
gruences of A then
[α0, . . . , αn−1] = ψ2n−1(∆(α0, . . . , αn−1)).
Further we show that ∆(α0, . . . , αn−1) encodes not only the commutator [α0, . . . , αn−1]
but also all smaller-arity commutators that can be obtained by omitting one or more
of the congruences αi. We show that if we take the clone of all polymorphisms of
the relation ∆(α0, . . . , αn−1) we get exactly the clone C(α0, . . . , αn−1) with the
properties described in the following theorem, and consequently one can construct
a largest clone with the same commutator operators as the original Mal'cev algebra.
Theorem 1.3. Let A be an algebra with Mal'cev term q, and let α0, . . . , αn−1 be
congruences of A. Then there exists a largest clone C(α0, . . . , αn−1) on A contain-
ing q such that it preserves congruences α0, . . . , αn−1, and all commutators of the
form [αi0 , . . . , αik−1 ] (where k ≤ n and 0 ≤ i0 < · · · < ik−1 < n) agree in A and
(A, C(α0, . . . , αn−1)).
RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
3
Corollary 1.4. Let A be an algebra with a Mal'cev term q, then there exists
a largest clone on A containing q such that the algebra corresponding to this clone
has the same congruence lattice as A and the same higher commutator operators
as A.
Proof of Corollary 1.4 given Theorem 1.3. The largest such clone is the intersec-
tion of all clones C(α1, . . . , αn) from Theorem 1.3 for all n and all tuples α1, . . . , αn
of congruences of A.
(cid:3)
Although our main motivation of developing this theory lies in the application
to Mal'cev algebras on a finite domain, the same results are valid even for algebras
with infinite domains. Moreover, since the largest clone in the previous theorem is
described as a polymorphism clone, we know that such clone is closed in the natural
topology given by pointwise convergence by a result of Romov [12]. More on clones
on infinite sets can be found in [9].
The theory developed to prove Theorem 1.3 is strong enough to give relatively
short proofs of several basic properties of higher commutators (usually referred as
(HC1) -- (HC8)) that have been established in [5], their proofs have been published
in [2]. Our alternative proofs of some of these properties are given in the last section
of this paper.
2. Preliminaries and notation
Algebras are denoted by bold letters, the underlying set of an algebra is denoted
by the same letter in italic, Con A denotes the set of all congruences of an algebra
A, Clo A the set (clone) of all term operations of A, Cg X denotes the congruence
generated by X, Sg Y denotes a subalgebra generated by Y , and if α is a congruence
then we use the symbol a ≡α b to denote (a, b) ∈ α. Furthermore, if R is a relation,
we use symbol Pol R to denote the clone of all polymorphisms of R. The symbol 2
will denote both the natural number 2 and the two-element set {0, 1}.
We denote tuples by bold letters. The i-th coordinate of tuple a is denoted by
either ai, or a(i). So, a = (a0, . . . , an−1) and (a0, . . . , an−1)(i) = ai. Tuples will
be usually numbered by an increasing sequence of consecutive integers starting at
0. So every n-ary relation is a subset of A{0,...,n−1} = An. The only exception will
be elements of the relation ∆(α0, . . . , αn−1). In the theory of binary commutator
described in [8], it is usual to denote the elements of 4-ary relation ∆α,β (we will
denote the same relation ∆(α, β)) as 2×2 matrices. Similarly, when generalizing this
concept to ∆(α0, . . . , αn−1) one should write elements of this relation as 2 × · · · × 2
n-dimensional matrices. We will denote those elements by tuples whose coordinates
will be labeled by the set 2n = {0, 1}n, we will write these coordinates as binary
sequences omitting brackets and commas, and if needed we will view them as reverse
binary expansions of natural numbers 0, . . . , 2n − 1; i.e., the tuple k = k0 . . . kn−1
represents the number P ki2i. This gives us a natural linear ordering of the set
2n that we will use to write the elements of A2n
as linear 2n-tuples. So, the tuple
a ∈ A2n
will be written as (a00...0, a10...0, a010...0, a110...0, . . . , a11...1). For d ∈ {0, 1}
we will use symbol d for the negation of d, i.e., 0 = 1 and 1 = 0. We will also
refer to forks of these relations at some coordinate k the same way as if all the
coordinates would be integer.
The last piece of notation has a close connection to a simple lemma about forks
of a relation. For any map e : J → I and a ∈ AI the symbol ae denotes the J-tuple
4
J. OPRSAL
defined by ae(j) = ae(j). Similarly, for a relation R ≤ AI , Re denotes the relation
{ae a ∈ R}.
Lemma 2.1. Let A be an algebra, R ≤ AI , S ≤ AJ , e : J → I, and Re ⊆ S.
If i ∈ I and there is a unique j ∈ J such that e(j) = i then ψi(R) ⊆ ψj(S). In
particular,
(i) if R ⊆ S then ψi(R) ⊆ ψi(S) for every i ∈ I;
(ii) if e : I → I is bijective then ψe(i)(R) = ψi(Re) for every i ∈ I.
Proof. Suppose that (a, b) ∈ ψi(R); i.e., there are tuples a, b ∈ R such that ai = a,
bi = b, and ak = bk for all k 6= i. Then from Re ⊆ S we know that ae, be ∈ S.
These tuples witness that (a, b) ∈ ψj(S), because ae(j) = ai = a, be(j) = bi = b,
and ae(k) = be(k) for k 6= j.
The statement (i) is a special case of the former for I = J, and e being the
identity mapping. To prove (ii), suppose that e is a bijection on the set I. Then
from the statement for S = Re we get that ψe(i)(R) ⊆ ψi(Re). For the other
inclusion substitute e with e−1, R with Re, and i with e(i).
(cid:3)
We recall two simple well-known lemmata for Mal'cev algebras.
Lemma 2.2. Let A be a Mal'cev algebra. Then any binary reflexive compatible
relation on A is a congruence.
(cid:3)
Lemma 2.3. Let A be a Mal'cev algebra, and let R be n-ary compatible relation
on A. If (a, b) ∈ ψi(R), and (c0, . . . , ci−1, a, ci+1, . . . , cn−1) ∈ R then
(c0, . . . , ci−1, b, ci+1, . . . , cn−1) ∈ R.
Proof. Without loss of generality suppose that i = 0. Let q be a Mal'cev term of
A, and let (a, u1, . . . , un−1) and (b, u1, . . . , un−1) be witnesses for (a, b) ∈ ψ0(R).
Then
b
c1
...
cn−1
=
a
c1
...
cn−1
since we know that R is compatible with q.
un−1 un−1
q
b
u1
...
a
u1
∈ R,
(cid:3)
3. Description of higher commutators
Definition 3.1. Let A be an algebra, and α0, . . . , αn−1 ∈ Con A. First, for each
congruence αi choose one dimension in the n-dimensional space. We define the
relation ∆A(α0, . . . , αn−1) as the 2n-ary relation indexed by the set 2n generated
by tuples that are constant on two opposing (n − 1)-dimensional hyperfaces of the
hypercube orthogonal to the dimension corresponding to αi and these constants
are αi congruent.
We will use faced
i (a) to denote the (d + 1)-th hyperface orthogonal to dimen-
i (a) = afi,d where fi,d(k) = k0 . . . ki−1dki . . . kn−2. The generating
i (a, b). By defini-
i (a, b) = (b, . . . , b); or equiva-
i (a, b) = (a, b)di where di : 2n → 2 is defined by k 7→ k(i). Finally,
sion i, i.e., faced
tuples of the relation ∆A(α0, . . . , αn−1) will be denoted cuben
tion, face0
lently, cuben
∆A(α0, . . . , αn−1) is defined as
i (a, b) = (a, . . . , a), and face1
i cuben
i cuben
∆A(α0, . . . , αn−1) := Sg(cid:8) cuben
i (a, b) i < n, a ≡αi b(cid:9) = _i<n
cuben
i (αi).
RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
5
For the trivial case when n = 0, we put ∆A() := A. If the algebra is clear from the
context, we will write just ∆(α0, . . . , αn−1) instead of ∆A(α0, . . . , αn−1), and if C is
a clone on the set A, we will write ∆C(α0, . . . , αn−1) instead of ∆(A,C)(α0, . . . , αn−1).
Example. We will describe generators of ∆(α0, α1, α2) for three congruences α0,
α1, α2 of an algebra A. The elements of ∆(α0, α1, α2) are indexed by vertices
of a three-dimensional hypercube. The generators are tuples of one of the fol-
lowing forms (a, b, a, b, a, b, a, b), where a ≡α0 b, (a, a, b, b, a, a, b, b) where a ≡α1 b,
and (a, a, a, a, b, b, b, b) where a ≡α2 b. Their graphical representation is given in
Figure 1.
a
a
b
b
a
a
α0
b
b
b
a
b
a
b
a
α1
b
a
a
a
α2
a
a
b
b
b
b
Figure 1. Generators of ∆(α0, α1, α2)
Before we get to the proof of Theorem 1.2, we will describe some basic proper-
ties of the relation ∆(α0, . . . , αn−1). The first lemma gives a term description of
∆(α0, . . . , αn−1). This description gives a clear connection of ∆(α0, . . . , αn−1) and
the term condition.
Lemma 3.2. For every algebra A, and congruences αi ∈ Con A, i < n,
∆(α0, . . . , αn−1)
=(cid:8)(cid:0)t(a0, . . . , an−1), t(b0, a1 . . . , an−1), . . . , t(b0, b1 . . . , bn−1)(cid:1)
∀i < n : mi ∈ N0, ai, bi ∈ Ami , ai ≡αi bi; t ∈ CloP mi A(cid:9)
where the elements of the 2n-tuple include the term t applied to all the combinations
of corresponding ai's and bi's.
Proof. The relation ∆(α0, . . . , αn−1) is generated by tuples cuben
So, ∆(α0, . . . , αn−1) is the set of all tuples of the form
i (a, b) for a ≡αi b.
t(cuben
i0 (a0, b0), . . . , cuben
ik−1 (ak−1, bk−1))
where t ∈ Clok A, and for all j < k we have ij < n, aj ≡ bj modulo αij . The de-
scription in the statement of the lemma can be obtained from this by grouping
together cuben
ij (aj, bj)'s with the same index ij, and applying the term t coordi-
natewise.
(cid:3)
Example. In the ternary commutator case, the lemma tells that
∆(α0, . . . , αn−1) =(cid:8)(cid:0)t(a0, a1, a2), t(b0, a1, a2), t(a0, b1, a2), t(b0, b1, a2),
t(a0, a1, b2), t(b0, a1, b2), t(a0, b1, b2), t(b0, b1, b2)(cid:1)
m0, m1, m2 ∈ N0, t ∈ Clom0+m1+m2 A, ∀i < 3 : ai, bi ∈ Ami , ai ≡αi bi(cid:9).
The graphical representation of a typical element of this relation is given in Figure 2.
6
J. OPRSAL
t(a0, b1, a2)
t(a0, b1, b2)
t(b0, b1, a2)
t(b0, b1, b2)
t(a0, a1, a2)
t(a0, a1, b2)
t(b0, a1, a2)
t(b0, a1, b2)
Figure 2.
Lemma 3.3. Let A be an algebra, and α0, . . . , αn−1, γ ∈ Con A. Then α0, . . . , αn−2
centralize αn−1 modulo γ if for every tuple
(a0...0, . . . , a1...1, b0...0, . . . , b1...1) ∈ ∆(α0, . . . , αn−1)
such that ak ≡γ bk for all k ∈ 2n−1 \ {1 . . . 1} we have also a1...1 ≡γ b1...1.
(cid:3)
Lemma 3.4. Let A be an algebra with congruences α0, . . . , αn, i < n, and si : k 7→
k(0) . . . k(i−1)k(i)k(i+1) . . . k(n−1). Then
(i) if d ∈ {0, 1} and a ∈ ∆(α0, . . . , αn−1) then
faced
i a ∈ ∆(α0, . . . , αi−1, αi+1, . . . , αn−1);
(ii) if a ∈ A2n−1
face0
such that
i a = face1
i a ∈ ∆(α0, . . . , αi−1, αi+1, . . . , αn−1),
then a ∈ ∆(α0, . . . , αn−1);
(iii) ∆(α0, . . . , αn−1)si = ∆(α0, . . . , αn−1).
Furthermore, if A is a Mal'cev algebra then the binary relation
is a congruence on ∆(α0, . . . , αi−1, αi+1, . . . , αn).
δ =(cid:8)(cid:0)face0
i a, face1
i a(cid:1) a ∈ ∆(α0, . . . , αn−1)(cid:9)
Proof. Since all the considered relations are compatible, we can check the validity
of particular inclusions on the generators of the relations, i.e., tuples cubei,n(a, b).
In detail, to prove (i) one has to observe that
faced
i cuben
j (a, b) =
and consequently
(a, b)
j
cuben−1
(c, . . . , c)
cuben−1
j−1 (a, b)
if j < i,
where c ∈ {a, b} if j = i,
if j > i,
faced
i {cuben
j (a, b) j < n, a ≡αj b}
= {cuben−1
j
(a, b) j < i, a ≡αj b} ∪ {cuben−1
j−1 (a, b) i < j < n, a ≡αj b}.
Hence the relation generated by the left hand side is the same as the relation
generated by the right hand side which gives the desired equality.
For (ii), first observe that
{a face0
i a = face1
i a ∈ ∆(α0, . . . , αi−1, αi+1, . . . , αn−1)}
RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
7
is the 2n-ary relation generated by tuples a such that face0
a ≡ b modulo αj, or αj+1 when j < i, or j ≥ i respectively. Second, if a is such
a tuple then a = cuben
In either
case, a ∈ ∆(α0, . . . , αn−1) for a, b that are congruent modulo the corresponding
congruence.
j (a, b) if j < i, or a = cuben
j+1(a, b) if j ≥ i.
i a = face1
i a = cubej,n−1(a, b),
The item (iii) is a consequence of the fact that
(cuben
j (a, b))si =(cuben
cuben
j (b, a)
if j = i, or
j (a, b) otherwise.
Hence (cuben
nition of ∆(α0, . . . , αn−1), we get that
j αj)si = cuben
j αj from the symmetry of congruences. From the defi-
∆(α0, . . . , αn−1)si = ∆(α0, . . . , αn−1).
From items (i) -- (iii) we already know that the binary relation δ is a reflexive
symmetric binary relation on ∆(α0, . . . , αi−1, αi+1, . . . , αn−1). The rest follows
from Lemma 2.2.
(cid:3)
Lemma 3.5. Let A be a Mal'cev algebra with congruences α0, . . . , αn−1. Then
ψj(∆(α0, . . . , αn−1)) = ψk(∆(α0, . . . , αn−1)) for all j, k ∈ 2n.
Proof. For simplicity, let ψk = ψk(∆(α0, . . . , αn−1)). If si is the permutation on 2n
defined by si(k) = k0 . . . ki−1kiki+1 . . . kn−1 then from Lemma 3.4(iii), we know that
∆(α0, . . . , αn−1) = ∆(α0, . . . , αn−1)si . This gives, by Lemma 2.1, that ψk = ψsi(k)
for all i < n. But, if i0, . . . , im−1 are exactly those indices i such that ki 6= ji then
j = si0 ◦ · · · ◦ sim−1 (k), and consequently ψk = ψj for all j, k.
(cid:3)
Instead of proving Theorem 1.2 directly, we will prove the following refinement.
The theorem is then given by equivalence of (1) and (4).
Proposition 3.6. If A is a Mal'cev algebra, α0, . . . , αn−1 ∈ Con A, and a, b ∈ A;
then the following is equivalent
(1) (a, b) ∈ ψ1...1(∆(α0, . . . , αn−1));
(2) (a, . . . , a, b) ∈ ∆(α0, . . . , αn−1);
(3) there exists c0, . . . , c2n−1−2 such that
(c0, . . . , c2n−1−2, a, c0, . . . , c2n−1−2, b) ∈ ∆(α0, . . . , αn−1).
(4) a ≡[α0,...,αn−1] b;
Proof. The implication (1) → (2) is direct corollary of Lemma 2.3, given that
∆(α0, . . . , αn−1) contains all constant tuples, in particular (a, . . . , a). (2) → (3)
is trivial. For (3) → (4) suppose that (3) is satisfied for a pair (a, b) then since
ci ≡ ci mod [α0, . . . , αn−1] for all i < 2n−1 − 1 and α0, . . . , αn−2 centralize αn−1
modulo [α0, . . . , αn−1], we have a ≡ b mod [α0, . . . , αn−1] from Lemma 3.3.
The last to prove is (4) → (1), in other words that
(3.1)
[α0, . . . , αn−1] ≤ ψ1...1(∆(α0, . . . , αn−1)).
Let ψ = ψ1...1(∆(α0, . . . , αn−1)); from Lemma 3.5 we know that
ψ = ψ1...1(∆(α0, . . . , αn−1)) = ψk(∆(α0, . . . , αn−1))
for all k ∈ 2n, so we do not have to distinguish between forks at different coordi-
nates. To prove (3.1) by the definition of the commutator, it is enough to prove that
8
J. OPRSAL
α0, . . . , αn−2 centralize αn−1 modulo ψ. For that we will use Lemma 3.3. Suppose
that
(a0...0, . . . , a1...1, b0...0, . . . , b1...1) ∈ ∆(α0, . . . , αn−1),
and ai ≡ψ bi for all i ∈ 2n−1 \ {1 . . . 1}. By repeated use of Lemma 2.3 we can
replace b00...0, . . . , b01...1 by the respective ai's. Hence,
(a00...0, . . . , a01...1, a11...1, a00...0, . . . , a01...1, b11...1) ∈ ∆(α0, . . . , αn−1).
In addition, we know from Lemma 3.4(i) and 3.4(ii) that
(a00...0, . . . , a01...1, a11...1, a00...0, . . . , a01...1, a11...1) ∈ ∆(α0, . . . , αn−1).
So, a11...1 ≡ψ b11...1 which concludes the proof that α0, . . . , αn−2 centralize αn−1
modulo ψ.
(cid:3)
In the last proposition some parts have been already known. The proposition (in
the case of Mal'cev algebras) generalize Theorem 4.9 of [8] which gives equivalence
of (2), (3), and (4) for binary commutators in congruence modular varieties. The
omitted equivalence of (1) and (3) in the binary case can be easily derived from the
known fact that ∆(α0, α1) is a congruence on rows. Furthermore, for the higher
commutators, the implication (3) → (4) for the variety of groups has appeared in
[13]; and if all αi's are principal congruences, the equivalence of (2) and (4) is given
by [2, Lemma 6.13] (via an easy translation similar to Lemma 3.2).
4. Strong cube terms, and clones of operations preserving
commutators
We will use terms that generalize Mal'cev terms. These terms will play similar
role as a difference term in the case of binary commutator. A (2n − 1)-ary term qn
is a strong n-cube term if it satisfies
qn(x00...0, x10...0, . . . , x01...1) ≈ x11...1
. This
whenever there is i < n such that for all k we have xk = xk0...ki−1kiki+1...kn−1
gives a set of n identities, each with 2n−1 variables. The two identities for strong
2-cube term are
q2(x, y, x) ≈ y
and q2(x, x, y) ≈ y.
So, the term q2(y, x, z) is a Mal'cev term, and if q is a Mal'cev term then q(y, x, z)
is a strong 2-cube term. The three identities for strong 3-cube term are
q3(x, y, z, w, x, y, z) ≈ w
q3(x, y, x, y, z, w, z) ≈ w
q3(x, x, y, y, z, z, w) ≈ w.
Strong cube terms are stricter version of cube terms introduced in [4]; a (2n − 1)-
ary term is an n-cube term if it satisfies
qn(x00...0, x10...0, . . . , x01...1) ≈ x11...1
whenever there is i < n such that xk = xl for all k, l such that ki = li. This gives
n two-variable identities compared to the 2n−1 variables used in the identities for
strong cube terms. One can see that for n ≥ 2 every strong n-cube term satisfies
the identities of an n-cube term; the i-th identity of a cube term is implied by
almost any identity for a strong cube term except for the i-th one.
Lemma 4.1. The following is equivalent for any algebra A.
RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
9
(1) A has a strong n-cube term for all n ≥ 2,
(2) A has a strong n-cube term for some n ≥ 2,
(3) A has a Mal'cev term.
Proof. From the observation in the above paragraphs, we know that the condition
(3) is equivalent to
(3') A has a strong 2-cube term.
We will prove equivalence of (1), (2), and (3'). The implication (1) → (2) is
trivial. For (2) → (3′) observe that if qn is a strong n-cube term then q2(x, y, z) =
qn(x, . . . , x, x, y, z) is a strong 2-cube term. For the last implication (3′) → (1) we
can construct strong cube terms by the recursion:
qn+1(x0, . . . , x2n+1−1) = q2(qn(x0, . . . , x2n−2), x2n−1, qn(x2n , . . . , x2n+1−2))
It is easy to check that if qn is a strong n-cube term and q2 is a strong 2-cube term
then qn+1 is a strong (n + 1)-cube term.
(cid:3)
The following lemma is the key for proving Theorem 1.3 and a lot of properties
of higher commutators in Mal'cev varieties.
Lemma 4.2. Let A be an algebra with a strong n-cube term qn, α1, . . . , αn ∈
Con A. Then a ∈ ∆(α1, . . . , αn) if and only if
face0
i a ∈ ∆(α0, . . . , αi−1, αi+1, . . . , αn−1)
for each i and qn(a00...0, a10...0, . . . , a01...1) ≡[α0,...,αn−1] a11...1.
Proof. We will prove the lemma in two steps.
Claim 1. If face0
a10...0, . . . , a01...1) = a11...1 then a ∈ ∆(α1, . . . , αn−1);
i a ∈ ∆(α0, . . . , αi−1, αi+1, . . . , αn−1) for each i < n, and qn(a00...0,
For k ∈ 2n and b ∈ A2n
let ak denotes the 2n-tuple
ak = (a00...0, ak00...0, a0k1...0, . . . , ak0k1...kn−1).
i ak = face1
Note that if ki = 0 then face0
i ak0...ki−11ki+1...kn−1 . Sup-
pose that the tuple a satisfies the premise of the claim. The fact that ak ∈
∆(α0, . . . , αn−1) for all k 6= 11 . . . 1 follows from induction on the number of 0's in
k -- the base step is given by Lemma 3.4(ii) and the assumption; the induction step
is given by Lemma 3.4(i) and (ii) and the above observation. Next we claim that
i ak = face0
(4.1)
qA2n
n
(a00...0, . . . , a01...1) = (a00...0, . . . , a01...1, qA
n (a00...0, . . . , a01...1));
i.e., we have to show that
qn(a00...0, aj00...0, . . . , a0j1...jn−1 ) = aj
for every coordinate j 6= 11 . . . 1. The above is trivial for the coordinate j = 11 . . . 1;
for j with ji = 0 the identity follows from the i-th identity for a strong cube term.
Finally, since the relation ∆(α0, . . . , αn−1) is compatible with qn we know that the
right hand side of (4.1) is in ∆(α0, . . . , αn−1).
Claim 2. If a ∈ ∆(α0, . . . , αn−1) then qn(a00...0, . . . , a01...1) ≡[α0,...,αn−1] a11...1.
10
J. OPRSAL
If (a00...0, . . . , a11...1) ∈ ∆(α0, . . . , αn−1) then from Lemma 3.4(i) we know that
the tuple (a00...0, . . . , a01...1, qn(a00...0, . . . , a01...1)) satisfies the prerequisites of the
first claim. Hence, we know that
(a00...0, . . . , a01...1, qn(a00...0, . . . , a01...1)) ∈ ∆(α0, . . . , αn−1)
which shows that
(a11...1, qn(a00...0, . . . , a01...1)) ∈ ψ11...1(∆(α0, . . . , αn−1)) = [α0, . . . , αn−1].
Finally, we get to the statement of this lemma. The 'only if' part is Claim 2
together with Lemma 3.4(i). For the 'if' part, if
face0
i a ∈ ∆(α0, . . . , αi−1, αi+1, . . . , αn−1)
for all i < n, we know from Claim 1 that
(a00...0, . . . , a01...1, qn(a00...0, . . . , a01...1)) ∈ ∆(α0, . . . , αn−1).
From the last condition and Theorem 1.2, we know that qn(a00...0, . . . , a01...1) and
a11...1 are congruent modulo ψ11...1(∆(α0, . . . , αn−1)). Therefore, a ∈ ∆(α0, . . . , αn−1)
from Lemma 2.3.
(cid:3)
In the rest of this chapter we use symbol [α0, . . . , αn−1]X to denote the commu-
tator [α0, . . . , αn−1] computed in the algebra X.
Corollary 4.3. Let A, B are algebras that share a universe, a Mal'cev operation,
and congruences α0, . . . , αn−1. Then
[αi0 , . . . , αik−1 ]A = [αi0 , . . . , αik−1 ]B
for all k ≤ n and 0 ≤ i0 < · · · < ik−1 < n if and only if ∆A(α0, . . . , αn−1) =
∆B(α0, . . . , αn−1).
Proof. We will prove the corollary by induction on n. The case n = 1 is trivial; for
the induction step suppose that for all congruences β0, . . . , βn−1 ∈ Con A ∩ Con B
such that the commutators [βi0 , . . . , βik−1 ] agree in A and B for all k < n and
i0, . . . , ik−1 pairwise distinct elements from {0, . . . , n−1}, we have ∆A(βi0 , . . . , βik−1 ) =
∆B(βi0 , . . . , βik−1 ).
In particular, we have ∆A(α0, . . . , αi−1, αi+1, . . . , αn−1) =
∆B(α0, . . . , αi−1, αi+1, . . . , αn−1) for all i < n. Let qn be a common strong n-
cube operation of A and B (it can be derived from the common Mal'cev oper-
ation). From Lemma 4.2 we know that whenever X is a Mal'cev algebra then
a ∈ ∆X(α0, . . . , αn−1) if and only if
(4.2)
face0
j a ∈ ∆X(α0, . . . , αi−1, αi+1, . . . , αn−1) for all j < n, and
a2n−1 ≡[α0,...,αn−1]X qn(a0, . . . , a2n−2).
Now, suppose that a ∈ ∆A(α0, . . . , αn−1) hence (4.2) is valid for X = A. But since
the operation qn, the commutators [α0, . . . , αn−1], and the relations ∆(α0, . . . , αi−1,
αi+1, . . . , αn−1) agree in A and B, we get that (4.2) is also true for X = B, hence
a ∈ ∆B(α0, . . . , αn−1). This shows that ∆A(α0, . . . , αn−1) ⊆ ∆B(α0, . . . , αn−1).
The other inclusion is analogous.
For the 'only if' part, suppose that ∆A(α0, . . . , αn−1) = ∆B(α0, . . . , αn−1).
From Lemma 3.4(i) it follows that ∆A(αi0 , . . . , αik−1 ) = ∆B(αi0 , . . . , αik−1 ) for
all {i0, . . . , ik−1} ⊆ {0, . . . , n − 1} with i's pairwise distinct; and consequently from
Theorem 1.2, [αi0 , . . . , αik−1 ]A = [αi0 , . . . , αik−1 ]B.
(cid:3)
RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
11
Finally, we get to the proof of the Theorem 1.3. We restate the theorem once
again.
Theorem 1.3. Let A be an algebra with Mal'cev term q, and α0, . . . , αn−1 be
congruences of A. Then there exists a largest clone C on A containing q such that it
preserves congruences α0, . . . , αn−1, and all commutators of the form [αi0 , . . . , αik ]
where 0 ≤ i0 < · · · < ik−1 < n agree in A and (A, C).
Proof. We will show that the largest clone satisfying the required properties is
the clone C of all polymorphisms of the relation ∆A(α0, . . . , αn−1). Obviously
C ⊇ Clo A which implies that q ∈ C and ∆C(α0, . . . , αn−1) ⊇ ∆A(α0, . . . , αn−1).
But since the relation ∆A(α0, . . . , αn−1) is compatible with C, and the relation
∆C(α0, . . . , αn−1) is the smallest compatible relation containing cuben
i (a, b) for all
a ≡αi b and i < n, we have ∆C(α0, . . . , αn−1) = ∆A(α0, . . . , αn−1).
From Corollary 4.3, we know that C satisfies the specified property. The rest is to
prove that C is the largest such clone. Let B be another clone satisfying the property.
Then from the same corollary we get ∆A(α0, . . . , αn−1) = ∆B(α0, . . . , αn−1), and
consequently B ⊆ Pol(∆A(α0, . . . , αn−1)) = C.
(cid:3)
5. Proofs of basic properties of higher commutators
In this chapter we will present alternative proofs of basic properties of higher
commutators formulated in [5, 2]. For an arbitrary algebra A and its congruences
α0, . . . , αn−1, β0, . . . , βn−1, γ, and η the following is satisfied
(HC1) [α0, . . . , αn−1] ≤ α0 ∧ · · · ∧ αn−1;
(HC2) if αi ≤ βi for all i then [α0, . . . , αn−1] ≤ [β0, . . . , βn−1];
(HC3) [α0, . . . , αn−1] ≤ [α1, . . . , αn−1].
Furthermore, if A is a Mal'cev algebra then
(HC4) if σ is a permutation on the set {0, . . . , n − 1} then
[α0, . . . , αn−1] = [ασ(0), . . . , ασ(n−1)];
(HC5) congruences α0, . . . , αn−2 centralize αn−1 modulo γ if and only if
(HC6) if η ≤ α0, . . . , αn−1 then
[α0, . . . , αn−1] ≤ γ;
[α0/η, . . . , αn−1/η]A/η = ([α0, . . . , αn−1]A ∨ η)/η;
(HC7) if I is a non-empty set, and ρi are congruences of A for all i ∈ I then
_i∈I
[α0, . . . , αn−2, ρi] = [α0, . . . , αn−2,_i∈I
ρi];
(HC8) if i = 1, . . . , n − 1 then [[α0, . . . , αi−1], αi, . . . , αn−1] ≤ [α0, . . . , αn−1].
Although properties (HC1 -- 3) can be derived directly from Theorem 1.2 and the
definition of ∆(α0, . . . , αn−1), we will omit these proofs because methods would
work only for Mal'cev algebras; the general case have been proved in [11, Propo-
sition 1.3]. We will prove properties (HC4), (HC5), (HC7), and (HC8) -- the last
property (HC6) is a corollary of (HC5).
Proposition 5.1 (HC4, [2, Proposition 6.1]). Let A be a Mal'cev algebra and
let α0, . . . , αn−1 ∈ Con A. Then [α0, . . . , αn−1] = [ασ(0), . . . , ασ(n−1)] for each
permutation σ of {0, . . . , n − 1}.
12
J. OPRSAL
Proof. We claim that the relations ∆(ασ(0), . . . , ασ(n−1)) and ∆(α0, . . . , αn−1) are
identical up to permuting coordinates; precisely
∆(ασ(0), . . . , ασ(n−1)) = ∆(α0, . . . , αn−1)σ′
where σ′ is defined by σ′(k) = kσ−1(0) . . . kσ−1(n−1). One can check this fact by
observing that σ′(k)(σ(i)) = ki, and consequently
∆(ασ(0), . . . , ασ(n−1)) = _i<n
cuben
σ(i)(ασ(i))(cid:17)σ′
i (ασ(i)) = _i<n(cid:16)cuben
=(cid:0)_i<n
cubei,n(αi)(cid:1)σ′
= ∆(α0, . . . , αn−1)σ′
.
Finally, σ′(11 . . . 1) = 11 . . . 1, so the statement is true from Theorem 1.2 and
Lemma 2.1(ii).
(cid:3)
Proposition 5.2 (HC5, [2, Lemma 6.2]). Let A be a Mal'cev algebra and α0, . . . , αn−1,
γ be congruences of A. Then α0, . . . , αn−2 centralizes αn−1 modulo γ if and only
if γ ≥ [α0, . . . , αn−1].
Proof. The 'only if' part is given by the definition of the commutator, to prove the
'if' part suppose that a ∈ ∆(α0, . . . , αn−1) such that ak0...kn−20 ≡γ ak0...kn−21 for
all k ∈ 2n−1 \ {1 . . . 1}. We want to prove that a11...10 ≡γ a11...1. By Lemma 4.2
we know that
a11...1 ≡[α0,...,αn−1] qn(a00...0, . . . , a01...1)
but the right hand side is modulo γ congruent to
qn(a00...0, . . . , a01...10, a1...10, a00...0, . . . , a01...10) = a1...10.
So, we know that a1...11 ≡ a1...10 modulo γ since γ ≥ [α0, . . . , αn−1]. And finally,
α0, . . . , αn−2 centralizes αn−1 modulo γ from Lemma 3.3.
(cid:3)
The condition (HC6) is a direct corollary of condition (HC5); for a proof see [2,
Corollary 6.3]. The following two lemmata prepare for the proof of (HC7).
Lemma 5.3. Let A be an algebra, I a non-empty set, ρi ∈ Con A for all i ∈ I,
and α0, . . . , αn−2 ∈ Con A. Then
∆(α0, . . . , αn−2,_i∈I
ρi) = _i∈I
∆(α0, . . . , αn−2, ρi).
Proof. The statement can be derived directly from the definition of the relation
∆(α0, . . . , αn−1) by a simple calculation:
∆(α0, . . . , αn−2,_i∈I
ρi) =(cid:16)cuben
= _i∈I(cid:0)cuben
= _i∈I
0 α0 ∨ · · · ∨ cuben
n−2 αn−2 ∨ cuben
0 α0 ∨ · · · ∨ cuben
n−2 αn−2 ∨ cuben
ρi(cid:17)
n−1_i∈I
n−1 ρi(cid:1)
∆(α0, . . . , αn−2, ρi).
(cid:3)
The following lemma was in fact proved during the proof of [2, Lemma 6.4].
RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
13
Lemma 5.4. Let A be an algebra, I a non-empty set, ρi ∈ Con A for all i ∈ I,
and α0, . . . , αn−2 ∈ Con A. Then
[α0, . . . , αn−2,_i∈I
ρi] =
_{i0,...,ik−1}⊆I,
k<∞
[α0, . . . , αn−2, _i∈{i0,...,ik−1}
ρi].
k<∞
ρi].
η =
Proof. Throughout the proof we will extensively use compactness of subuniverses
of some fixed algebra that is if Ai for i ∈ J are subalgebras of some algebra A,
and a ∈ Wi∈J Ai then there exists a finite set K ⊆ J such that a ∈ Wi∈K Ai. To
shorten the notation let η denotes the right hand side of the statement. Hence
_{i0,...,ik−1}⊆I,
[α0, . . . , αn−2, _i∈{i0,...,ik−1}
First, we prove that α0, . . . , αn−2 centralizeWi∈I ρi modulo η.
Claim 1. If a ∈ ∆(α0, . . . , αn−2,Wi∈I ρi) then there is a finite set S ⊆ I such that
a ∈ ∆(α0, . . . , αn−2,Wi∈S ρi).
modulo [α0, . . . , αn−2,Wi∈T ρi].
j=0 [α0, . . . , αn−2,Wi∈Tj
Wk−1
Claim 2. If a ≡η b then there is a finite set T ⊆ I such that a and b are congruent
Again, there are finite sets T0, . . . , Tk−1 such that a and b are congruent modulo
The claim follows from Lemma 5.3 and the note at the beginning of this proof.
ρi]. And,
k−1
_j=0
[α0, . . . , αn−2, _i∈Tj
ρi] ≤ [α0, . . . , αn−2, _ρ∈T
ρ]
j=0 Tj. Which completes the proof of the second claim.
k ∈ 2n−1 \ {1 . . . 1}. Let S be a finite set from Claim 1, and let Tk be finite sets
where T =Sk−1
Suppose that a ∈ ∆(α0, . . . , αn−2,Wi∈I ρi) and ak0...kn−20 ≡η ak0...kn−21 for all
such that ak0...kn−20 and ak0...kn−21 are congruent modulo [α0, . . . , αn−2,Wρ∈Tk ρ];
set) and η′ =Wi∈U ρi. Then
(1) a ∈ ∆(α0, . . . , αn−2, η′), and
(2) ak0...kn−20 ≡[α0,...,αn−1,η′] ak0...kn−21 for all k ∈ 2n−1 \ {1 . . . 1}.
such sets exist by Claim 2. Let U = S ∪ T0...0 ∪ · · · ∪ T1...10 (note that U is a finite
So, from the Lemma 3.3 we know that a1...10 ≡[α0,...,αn−1,η′] a1...11. Finally, [α0, . . . , αn−1, η′] ≤
[α0, . . . , αn−1, η] because η′ ≤ η. Hence α0, . . . , αn−2 centralize Wi∈I ρi modulo η,
and consequently [α0, . . . , αn−2,Wi∈I ρi] ≤ η.
The other inclusion is obvious from the fact that
[α0, . . . , αn−2,_i∈I
ρi] ≥ [α0, . . . , αn−2,_i∈J
ρi]
for every finite set J ⊆ I.
(cid:3)
Lemma 5.5 ([2, Corollary 6.6]). Let A be a Mal'cev algebra and α1, . . . , αn−1,
ρ1, . . . , ρk congruences of A. Then
k
k
[α0, . . . , αn−2,
ρi] =
[α0, . . . , αn−2, ρi].
_i=1
_i=1
14
J. OPRSAL
Proof. It suffices to prove the statement just for k = 2. We will write
(a0, . . . , a2n−1−1) ≡∆(α0,...,αn−2,ρi) (b0, . . . , b2n−1−1)
if (a0, . . . , a2n−1−1, b0, . . . , b2n−1−1) ∈ ∆(α0, . . . , αn−2, ρi). Note that from Lemma
3.4, we know that the binary relation {(a, b) : a ≡∆(α0,...,αn−2,ρi) b} is a congruence
on ∆(α0, . . . , αn−2). From Lemma 5.3, we know that
∆(α0, . . . , αn−2, ρ1 ∨ ρ2) = ∆(α0, . . . , αn−2, ρ1) ∨ ∆(α0, . . . , αn−2, ρ2).
Since in Mal'cev algebras α ◦ β = Sg(α ∪ β) for any pair of congruences α, β, we
have that for all a, b ∈ ∆(α0, . . . , αn−2), a ≡∆(α0,...,αn−2,ρ1∨ρ2) b if and only if
there exists c such that a ≡∆(α0,...,αn−2,ρ1) c and b ≡∆(α0,...,αn−2,ρ2) c.
Now, we prove that
(5.1)
[α0, . . . , αn−2, ρ1 ∨ ρ2] ≤ [α0, . . . , αn−2, ρ1] ∨ [α0, . . . , αn−2, ρ2].
Suppose that a and b are congruent modulo [α0, . . . , αn−2, ρ1 ∨ ρ2], hence from
Proposition 3.6 there are e0, . . . , e2n−1−2 such that
(e0, . . . , e2n−1−2, a) ≡∆(α0,...,αn−2,ρ1∨ρ2) (e0, . . . , e2n−1−2, b).
From the above observation, we know that there is a tuple c ∈ ∆(α0, . . . , αn−2)
such that
(5.2)
and
(5.3)
c ≡∆(α0,...,αn−2,ρ1) (e0, . . . , e2n−1−2, a)
c ≡∆(α0,...,αn−2,ρ2) (e0, . . . , e2n−1−2, b)
If we use Lemma 4.2 for (5.2), we get that
a ≡[α0,...,αn−2,ρ1] qn(c0, . . . , c2n−1−1, e0, . . . , e2n−1−2);
similarly for (5.3), we get that
b ≡[α0,...,αn−2,ρ2] qn(c0, . . . , c2n−1−1, e0, . . . , e2n−1−2).
Altogether, a and b are congruent modulo [α0, . . . , αn−2, ρ1] ∨ [α0, . . . , αn−2, ρ2].
Which completes the proof of (5.1). The other inclusion is given by (HC2).
(cid:3)
Proposition 5.6 (HC7, [2, Lemma 6.7]). Let A be a Mal'cev algebra with congru-
ences α0, . . . , αn−2, and ρi, i ∈ I for I non-empty set. Then
[α0, . . . , αn−2,_i∈I
ρi] = _i∈I
[α0, . . . , αn−2, ρi].
Proof. If I is a finite set then the proposition is given by Lemma 5.5. If I is infinite,
we can first use Lemma 5.4 to get
ρi] =
_{i0,...,ik−1}⊆I
[α0, . . . , αn−2, _i∈{i0,...,ik−1}
ρi].
Then by using the finite case, the right hand side is equal to
[α0, . . . , αn−2,_i∈I
_{i0,...,ik−1}⊆I,
i∈{i0,...,ik−1}
[α0, . . . , αn−2, ρi] = _i∈I
[α0, . . . , αn−2, ρi].
(cid:3)
Proposition 5.7 (HC8, [2, Proposition 6.14]). Let A be a Mal'cev algebra with
congruences α0, . . . , αn−1, and i ∈ {1, . . . , n − 1}. Then
[[α0, . . . , αi−1], αi, . . . , αn−1] ≤ [α0, . . . , αn−1].
RELATIONAL DESCRIPTION OF HIGHER COMMUTATORS
15
Proof. For m ≥ i, define the map em : 2m → 2m−i+1 by k 7→ k′ki . . . km−1 where
k′ = k0 · k1 · . . . · ki−1. We claim that
∆([α0, . . . , αi−1], αi, . . . , αm−1)em ≤ ∆(α0, . . . , αm−1).
(5.4)
Because ∆([α0, . . . , αi−1], αi, . . . , αm−1)em is clearly a subuniverse of A2m
by the set
generated
(cubem−i+1
0
[α0, . . . , αi−1])em ∪ [j<m−i
(cubem−i+1
j+1
αi+j )em ,
it suffices to prove that this is a subset of ∆(α0, . . . , αm−1). The inclusions
(cubem−i+1
j+1
αi+j )em ⊆ ∆(α0, . . . , αn−1)
are consequences of the fact that em(k)(j + 1) = kj+i for all j < m − i. The
other inclusion, (cubem−i+1
[α0, . . . , αi−1])em ⊆ ∆(α0, . . . , αm−1), can be proved by
induction on m. If m = i then
0
(cube1
0[α0, . . . , αi−1])em = {(a, . . . , a, b) a ≡[α0,...,αi−1] b},
and all the elements of this set are in ∆(α0, . . . , αi−1) from Proposition 3.6. For
the induction step, observe that
face0
m(cid:0)(cube(m+1)−i+1
0
(a, b))em+1(cid:1) = face1
m(cid:0)(cube(m+1)−i+1
0
(a, b))em+1(cid:1)
=(cid:0)cubem−i+1
0
(a, b)(cid:1)em .
So the inclusion follows from Lemma 3.4(ii). Finally, from (5.4) for m = n, the fact
that en(k) = 1 . . . 1 if and only if k = 1 . . . 1, and Lemma 2.1 we get
ψ1...1(∆([α0, . . . , αi−1], αi, . . . , αn−1)) ≤ ψ1...1(∆(α0, . . . , αn−1)).
The rest is Theorem 1.2.
(cid:3)
Acknowledgements. The author would like to thank Erhard Aichinger for many
hours of fruitful discussions and many helpful comments on the manuscript.
References
[1] Erhard Aichinger, Peter Mayr, and Ralph McKenzie. On the number of finite algebraic struc-
tures. J. Eur. Math. Soc. (JEMS), 16(8):1673 -- 1686, 2014.
[2] Erhard Aichinger and Nebojsa Mudrinski. Some applications of higher commutators in
Mal'cev algebras. Algebra Universalis, 63(4):367 -- 403, 2010.
[3] Erhard Aichinger and Nebojsa Mudrinski. On various concepts of nilpotence for expansions
of groups. Publ. Math. Debrecen, 83(4):583 -- 604, 2013.
[4] Joel Berman, Pawe l Idziak, Petar Markovi´c, Ralph McKenzie, Matthew Valeriote, and Ross
Willard. Varieties with few subalgebras of powers. Trans. Amer. Math. Soc., 362(3):1445 --
1473, 2010.
[5] Andrei Bulatov. On the number of finite Mal'tsev algebras. In Contributions to general alge-
bra, 13 (Velk´e Karlovice, 1999/Dresden, 2000), pages 41 -- 54. Heyn, Klagenfurt, 2001.
[6] Andrei A. Bulatov. Three-element Mal'tsev algebras. Acta Sci. Math. (Szeged), 71(3-4):469 --
500, 2005.
[7] Andrei A. Bulatov and Pawe l M. Idziak. Counting Mal'tsev clones on small sets. Discrete
Math., 268(1 -- 3):59 -- 80, 2003.
[8] Ralph Freese and Ralph McKenzie. Commutator theory for congruence modular varieties,
volume 125 of London Mathematical Society Lecture Note Series. Cambridge University Press,
Cambridge, 1987.
[9] Martin Goldstern and Michael Pinsker. A survey of clones on infinite sets. Algebra Universalis,
59(3-4):365 -- 403, 2008.
16
J. OPRSAL
[10] Pawe l M. Idziak. Clones containing Mal'tsev operations. Internat. J. Algebra Comput.,
9(2):213 -- 226, 1999.
[11] Nebojsa Mudrinski. On Polynomials in Mal'cev Algebras. PhD thesis, University of Novi Sad,
2009.
[12] B.A. Romov. Galois correspondence between iterative post algebras and relations on an infi-
nite set. Cybernetics, 13(3):377 -- 379, 1977.
[13] Jason Shaw. Commutator relations and the clones of finite groups. Algebra Universalis,
72(1):29 -- 52, 2014.
Department of Algebra, Faculty of Mathematics and Physics, Charles University
in Prague, Sokolovsk´a 83, 186 75 Praha, Czech Republic
Institute for Algebra, Johannes Kepler University, 4040 Linz, Austria
E-mail address: [email protected]
|
1702.02257 | 3 | 1702 | 2017-11-17T13:22:42 | Closure operators, frames, and neatest representations | [
"math.RA"
] | Given a poset $P$ and a standard closure operator $\Gamma:\wp(P)\to\wp(P)$ we give a necessary and sufficient condition for the lattice of $\Gamma$-closed sets of $\wp(P)$ to be a frame in terms of the recursive construction of the $\Gamma$-closure of sets. We use this condition to show that given a set $\mathcal{U}$ of distinguished joins from $P$, the lattice of $\mathcal{U}$-ideals of $P$ fails to be a frame if and only if it fails to be $\sigma$-distributive, with $\sigma$ depending on the cardinalities of sets in $\mathcal{U}$. From this we deduce that if a poset has the property that whenever $a\wedge(b\vee c)$ is defined for $a,b,c\in P$ it is necessarily equal to $(a\wedge b)\vee (a\wedge c)$, then it has an $(\omega,3)$-representation. This answers a question from the literature. | math.RA | math |
Closure operators, frames, and neatest
representations
Rob Egrot
Abstract
Given a poset P and a standard closure operator Γ : ℘(P ) → ℘(P ) we
give a necessary and sufficient condition for the lattice of Γ-closed sets of
℘(P ) to be a frame in terms of the recursive construction of the Γ-closure
of sets. We use this condition to show that given a set U of distinguished
joins from P , the lattice of U-ideals of P fails to be a frame if and only if
it fails to be σ-distributive, with σ depending on the cardinalities of sets
in U. From this we deduce that if a poset has the property that whenever
a ∧ (b ∨ c) is defined for a, b, c ∈ P it is necessarily equal to (a ∧ b) ∨ (a ∧ c),
then it has an (ω, 3)-representation. This answers a question from the
literature.
1 Introduction
Schein [15] defines a meet-semilattice S to be distributive if it satisfies the first
order definable condition that whenever a ∧ (b ∨ c) is defined, (a ∧ b) ∨ (a ∧ c)
is also defined and the two are equal. A dual definition can, of course, be
made for join-semilattices. Here for simplicity all semilattices are considered
to be meet-semilattices, and the order duals to our results are left unstated.
Note that Schein's version of distributivity is strictly weaker than the notion
of distributivity introduced by Gratzer and Schmidt [10] (see for example [20]),
though they coincide for lattices. Schein's distributivity (hereafter referred to
as 3-distributivity, for reasons that will become clear later) is equivalent to
the property of being embeddable into a powerset algebra via a semilattice
homomorphism preserving all existing binary joins [2].
Schein's 3-distributivity can be generalized to the concept of α-distributivity
for cardinals α (see definition 3.1). This notion has been studied when α = n <
ω [11], when α = ω [20, 4], and when α is any regular cardinal [12]. Note that if
m, n ≤ ω with m < n then n-distributivity trivially implies m-distributivity, but
the converse is not true [13]. A recurring theme in these investigations is that
a semilattice S is α-distributive if and only if the complete lattice of downsets
of S that are closed under existing α-small joins is a frame (i.e. it satisfies the
complete distributivity condition that x ∧ W Y = WY (x ∧ y) for all elements x
and subsets Y ).
1
Notions relating to distributivity have also been studied in the more general
setting of partially ordered sets (posets) [18, 12]. Note that the concepts in [18]
and [12] are not equivalent, even for semilattices (that of [18] is a straightforward
generalization of the distributivity of Gratzer and Schmidt [10], while that of
[12] is more closely related to that of Schein [15]).
As the transition from lattices to semilattices causes previously equivalent
formulations of distributivity to diverge, so too does the transition from semi-
lattices to posets. The following example is of particular interest to us. For
a semilattice, being an ω-distributive poset as defined in [12] is equivalent to
being an ω-distributive semilattice in the sense used here [12, proposition 2.3],
which is in turn equivalent to being embeddable into a powerset algebra via a
map preserving all existing finite meets and joins [2]. However, this is not true
for arbitrary posets, as the ω-distributivity of [12] is strictly stronger than the
powerset algebra embedding property [12, example 4.2 and theorem 4.8].
The property of being embeddable into a powerset algebra via an embedding
preserving meets and joins of certain cardinalities has been studied using the
terminology representable [3, 14, 7] (see definition 4.1). A notable departure
from the semilattice case is that the first order theory of representable posets is
considerably more complex. For example, while the class of (m, n)-representable
posets is elementary for all m, n with 2 < m, n ≤ ω [7], explicit first order axioms
are not known, and the class cannot be finitely axiomatized [6]. This contrasts
with the semilattice case where intuitive first order axioms are known, and only
a finite number are required to ensure (ω, n)-representability for finite n [2].
In [3] a condition for posets, which we will refer to as MD, generalizing the
distributive property for semilattices is defined (see definition 4.2). The authors
conjecture ([3, section 3]) that this condition is sufficient to ensure a poset has a
representation preserving existing binary joins and all existing finite meets (they
call this a neatest representation). It is easily seen that this is not a necessary
condition for such a representation, e.g. [7, example 1.5]).
In the original phrasing of [3, definition 1] it is unclear whether a neatest
representation must preserve arbitrary existing meets or only finite ones. From
context we assume the latter, as the with the former definition the conjecture is
false. To see this note that every Boolean algebra satisfies MD, and indeed the
stronger condition that if x ∧ W Y is defined then WY (x ∧ y) is also defined and
the two are equal (a Boolean algebra also satisfies the dual to this condition,
so a complete Boolean algebra is structurally both a frame and a co-frame).
However, it is well known that a Boolean algebra has a representation preserving
arbitrary meets and/or joins if and only if it atomic [1, corollary 1].
However, if we only demand that finite meets are preserved then the con-
jecture is indeed correct, which we prove in this note. The main step in our
proof is a result classifying the standard closure operators on ℘(P ) whose lat-
tices of closed sets are frames as being precisely those that can be constructed
using a certain recursive procedure (theorem 3.5). This can be viewed as a
generalization of [12, theorem 2.7]. The proof of the conjecture about neatest
representations is then an easy corollary.
We note that the argument in the solution of this conjecture boils down to a
2
straightforward extension of the forward direction of the argument for semilat-
tices in [11, theorem 2] to the poset setting. The result however is nevertheless
perhaps surprising given that the class of posets with this kind of representation
cannot be finitely axiomatized (see section 5 for a discussion of this).
We proceed as follows. In section 2 we define introduce join-specifications
and the concept of the radius of a standard closure operator, and assorted sup-
porting terminology. In section 3 we use this to prove the central theorem (the-
orem 3.5), and in the fourth section we prove a general version of the conjecture
on neatest representations (corollary 4.6). The final section is devoted to a short
discussion of the implications of this for the theory of poset representations, and
some indirectly related questions in complexity theory.
2 Join-specifications and standard closure oper-
ators
We begin with a brief introduction to closure operators and their relationship
with poset completions. This topic has been studied in detail, and a recent sur-
vey can be found in [8]. Nevertheless it will be useful to present some well known
results with a consistent terminology, and to make explicit some facts that are
only implicit elsewhere. We will give specific references when appropriate.
Definition 2.1 (closure operator). Given a set X a closure operator on X is a
map from ℘(X) to itself that is extensive, monotone and idempotent. I.e. such
that for all S, T ⊆ X we have:
1. S ⊆ Γ(S),
2. S ⊆ T =⇒ Γ(S) ⊆ Γ(T ), and
3. Γ(Γ(S)) = Γ(S).
Given a set S ⊆ P we denote {p ∈ P : p ≤ s for some s ∈ S} by S↓, and
p↓ is used a shorthand for {p}↓. Given a poset P we are interested in closure
operators on P . In particular we are interested in standard closure operators,
that is, closure operators such that Γ({p}) = p↓ for all p ∈ P .
Definition 2.2 (join-completion). Given a poset P a join-completion of P is
a complete lattice L and an order embedding e : P → L such that e[P ] (the
image of P under e) is join-dense in L.
Given a standard closure operator Γ : ℘(P ) → ℘(P ), the Γ-closed sets form a
complete lattice when ordered by inclusion (which we denote with Γ[℘(P )]), and
the map φΓ : P → Γ[℘(P )] defined by φΓ(p) = p↓ is a join-completion of P (we
usually omit the subscript and just write φ). If I is an indexing set and Ci is a
Γ-closed set for all i ∈ I then VI Ci = TI Ci and WI Ci = Γ(SI Ci) (see e.g. [8,
section 2.1]). Conversely, given a join-completion e : P → L the sets e−1[x↓] for
x ∈ L define the closed sets of a standard closure operator Γe : ℘(P ) → ℘(P ).
This well known connection can be expressed as the following proposition.
3
Proposition 2.3. There is a dual isomorphism between the complete lattice
of standard closure operators on P (ordered by pointwise inclusion) and the
complete lattice of (e[P ] preserving isomorphism classes of ) join-completions of
P (ordered by defining (e1 : P → L1) ≤ (e2 : P → L2) if and only if there is an
order embedding ψ : L1 → L2 such that ψ ◦ e1 = e2. If ψ exists it will necessarily
preserve all existing meets).
Proof. A direct argument is straightforward. See for example [8, propositions
2.1 and 2.19], or the introduction to [18], for equivalent results.
We are interested in join-completions e : P → L where the embedding e
preserves certain existing joins from P . Often this is done by making some
uniform selection, such the joins of all sets smaller than some fixed cardinal
(when they exist, see e.g. [17]). We intend to be more general, and to that end
we make the following definition.
Definition 2.4 (join-specification). Let P be a poset. Let U be a subset of
℘(P ). Then U is a join-specification (of P ) if it satisfies the following conditions:
1. W S exists in P for all S ∈ U,
2. {p} ∈ U for all p ∈ P , and
3. ∅ /∈ U.
Definition 2.4 is similar to that of a subset selection (see e.g. [8, section 2]).
The difference is that we demand that the selection contains the singletons and
that every selected set has a defined join. This serves to tidy up some of the
later definitions.
Definition 2.5 (radius of U). Given a join-specification U we define the radius
of U to be the smallest cardinal σ such that σ > S for all S ∈ U.
Definition 2.6 (ΓU ). Given a join-specification U with radius σ and S ⊆ P we
define the following subsets of P using transfinite recursion.
• Γ0(S) = S↓.
• If α + 1 is a successor ordinal then Γα+1(S) = {W T : T ∈ U and T ⊆
Γα(S)}↓.
• If λ is a limit ordinal Γλ(S) = Sβ<λ Γβ(S).
We define ΓU : ℘(P ) → ℘(P ) by ΓU (S) = Γχ(S) for all S ∈ ℘(P ), where χ is
the smallest regular cardinal with σ ≤ χ.
Definition 2.7 (U-ideal). Given a join-specification U we define a U-ideal of
P to be a down-set closed under joins from U. We define the empty set to be a
U-ideal.
The following is a generalization of [12, proposition 1.2].
4
Proposition 2.8. If U is a join-specification then ΓU is the standard closure
operator taking S ⊆ P to the smallest U-ideal containing S.
Proof. Let S ∈ ℘(P ) \ {∅}, and let I be the smallest U-ideal containing S. Then
we must have ΓU (S) ⊆ I by the closure requirements of I.
It is easy to see
that ΓU (S) is a down-set, so it remains only to show that ΓU (S) is closed under
joins from U. So let X ∈ U and suppose X ⊆ ΓU (S). Then by definition of
ΓU we have X ⊆ Sβ<χ Γβ(S), so for each x ∈ X there is some βx < χ with
x ∈ Γβx(S). Since χ is regular there must be β′ with βx ≤ β′ < χ for all x ∈ X,
and so W X ∈ Γβ ′+1(S) ⊆ Γχ(S) = ΓU (S) as required. It is straightforward to
show that the function taking sets to the smallest U-ideal containing them is a
standard closure operator.
Proposition 2.9. Given a join-specification U, the canonical map φ : P →
ΓU [℘(P )] preserves arbitrary existing meets and the joins of all sets from U.
Proof. This is well known, but we give a short proof for the sake of completeness.
First recall that arbitrary intersections of Γ-closed sets are also Γ-closed. So if
V S = t in P , then V φ[S] = TS s↓ = t↓ = φ(t) as required. Preservation of
joins from U follows from lemma 2.11 and corollary 2.13(2) below.
Note that φ may also preserve joins of sets that are not in U. Given a join-
specification U there will generally be more than one join-completion preserving
the specified joins. We are interested in ΓU [℘(P )], which is in fact the largest
such join-completion ([16] attributes this result to [5]). This is easily seen by
noting that if e : P → L is a join-completion such that e(W S) = W e[S] for some
S ⊆ P , then whenever x ∈ L and S ⊆ e−1[x↓] we must have W S ∈ e−1[x↓] as
otherwise W e[S] 6= e(W S). So in particular if e : P → L preserves joins from U
then e−1[x↓] is a U-ideal for all x ∈ L. Since L is isomorphic to {e−1[x↓] : x ∈ L}
considered as a lattice ordered by inclusion the result follows as we can think of
L as being a subset of the set of all U-ideals.
Definition 2.10 (UΓ). Any standard closure operator Γ : ℘(P ) → ℘(P ) defines
a join-specification UΓ by
S ∈ UΓ ⇐⇒ _ S exists and for all Γ-closed sets C we have S ⊆ C =⇒ _ S ∈ C
Lemma 2.11. For UΓ as in definition 2.10 and the canonical φ : P → Γ[℘(P )]
we have
S ∈ UΓ ⇐⇒ _ S exists and φ(_ S) = _ φ[S]
Proof. Let S ∈ UΓ. Then W φ[S] is the smallest Γ-closed set containing S, so
must therefore contain W S, and is indeed equal to (W S)↓ = φ(W S). Conversely,
if W S exists and W φ[S] = φ(W S) then every Γ-closed set containing S must
contain W S, and so S ∈ UΓ by definition.
The join-specifications of a poset P are a subset of ℘(℘(P )), and as this
subset is closed under taking arbitrary unions and intersections they form a
complete lattice when ordered by inclusion. This leads us to the following
result.
5
Proposition 2.12. Let J be the lattice of join-specifications of P , and let C be
the lattice of standard closure operators on P . Define f : J → C and g : C → J
by f (U) = ΓU and g(Γ) = UΓ. Then f and g form a Galois connection between
C and J .
for all Γ ∈ C and U ∈ J we have ΓU (T ) ≤ Γ(T ) for all
T ∈ ℘(P ) ⇐⇒ U ⊆ UΓ.
I.e.
Proof. Let Γ ∈ C and let U ∈ J . Suppose ΓU ≤ Γ. Since J is ordered by
inclusion we let S ∈ U, and we aim to show that S ∈ UΓ. Now, ΓU (S) is the
smallest U-ideal containing S, so W S ∈ ΓU (S), and since ΓU (S) ≤ Γ(S) this
means any Γ-closed set containing S must contain W S. But then by definition
S ∈ UΓ so we are done.
For the converse suppose U ⊆ UΓ and let T ∈ ℘(P ). We aim to show that
ΓU (T ) ⊆ Γ(T ). Since ΓU (T ) is the smallest U-ideal containing T it is sufficient
to show that Γ(T ) is also a U-ideal. But since U ⊆ UΓ this follows directly from
the definition of UΓ.
Corollary 2.13.
(1) For all standard closure operators Γ : ℘(P ) → ℘(P ) we have ΓUΓ ≤ Γ, but
we do not necessarily have Γ = ΓUΓ.
(2) For all join-specifications U of P we have U ⊆ UΓU , but we do not neces-
sarily have U = UΓU .
(3) If U ′ = UΓU then ΓU = ΓU ′.
(4) If Γ′ = ΓUΓ then UΓ = UΓ′.
Proof. This all follows from the fact that f and g from proposition 2.12 form a
Galois connection, with examples 2.14 and 2.15 witnessing lack of equality for
parts (1) and (2) respectively.
Example 2.14. Let P be the three element antichain {a, b, c}, and let the
Γ-closed sets be ∅, {a}, {b}, {c}, and {a, b, c}. Then UΓ = {{a}, {b}, {c}} (so
Γ[℘(P )] is the MacNeille completion of P ), and the set of UΓ-ideals of P is just
℘(P ) (so ΓUΓ [℘(P )] is the Alexandroff completion composed of all down-sets of P
in this case). Then, for example, Γ({a, b}) = {a, b, c}, but ΓUΓ({a, b}) = {a, b}.
Example 2.14 also demonstrates that not every standard closure operator
arises from a join-specification. This is because the only join-specification on
P is {{a}, {b}, {c}}, and as we saw in the example the induced closure opera-
tor produces the Alexandroff completion, and not, for example, the MacNeille
completion. This issue is also discussed in [16, section 2].
Example 2.15. Let P be a poset containing elements x, x′, y, y′, and z. Let the
non-trivial orderings be x < x′ < z, and y < y′ < z, so x ∨ y = x′ ∨ y′ = z. Let
U = {{x}, {x′}, {y}, {y′}, {z}, {x, y}} be a join-specification. Then {x′, y′} ∈
UΓU \ U.
6
It follows from corollary 2.13 that different join-specifications can define
the same closure operator. Given a standard closure operator Γ arising from
a join-specification, while there is not necessarily a minimal generating join-
specification, definition 2.5 allows us to define a class of join-specifications gen-
erating Γ whose sets have the smallest possible maximum size. This puts an
upper bound on the number of iterations required in a recursive construction of Γ
(though we can't expect to do better than ω, even if all the sets in the generating
join-specification have bounded finite size). These minimal join-specifications
(definition 2.17 below) are also relevant when discussing the distributivity of
the lattice of Γ-closed sets (see corollary 3.6).
Definition 2.16 (radius of Γ). Given a standard closure operator Γ such that
Γ = ΓU for some join-specification U, we define the radius of Γ to be the
minimum of {χ : χ is the radius of a join-specification U ′ of P with ΓU ′ = Γ}.
Definition 2.17 (minimal join-specification). A join-specification U of P is
minimal if the radius of U is equal to the radius of ΓU .
The following technical lemma will be used in the next section.
Lemma 2.18. Let U be a join-specification. Then the following hold for all
S ∈ ℘(P ):
(1) If ΓU (S) = p↓ then p = W S.
(2) If p = W S and p ∈ ΓU (S) then S ∈ UΓU .
Proof. For the first part note that p must be an upper bound for S, so if p 6= W S
then S has another upper bound q with p 6≤ q. But ΓU (S) ⊆ p↓ ∩ q↓ by
proposition 2.8, and thus p /∈ ΓU (S), which would be a contradiction. For the
second part note that since W S = p, by definition 2.10 we have S ∈ UΓU ⇐⇒
for all ΓU -closed sets C we have S ⊆ C =⇒ p ∈ C. Since ΓU (S) is the smallest
ΓU -ideal containing S, if p ∈ ΓU (S) then every ΓU -ideal containing S must also
contain p and we are done.
3 When is a lattice of U -ideals a frame?
As mentioned in the introduction, definitions of distributivity in semilattices
modeled on that of Schein [15] give rise to results that can be stated in our
terminology as a semilattice S is α-distributive if and only if the lattice of Uα-
ideals of S is α-distributive, where Uα contains all sets smaller than α whose joins
are defined (where α is some finite or regular cardinal [11, 4, 12]). Moreover,
it turns out that if this lattice of Uα-ideals is α-distributive then it will be a
frame (see definition 3.1 below). We can extend this to posets and arbitrary
join-specifications, in the sense that if the lattice of U-ideals of P fails to be
a frame then it must be because distributivity fails for the embedded images
of some element of P and some set in U (see corollary 3.7). The key result is
theorem 3.5, which can be seen as a partial generalization of [12, theorem 2.7]
to arbitrary join-specifications.
7
Definition 3.1 (α-distributive). Given a cardinal α we say a lattice (or a
semilattice) L is α-distributive if given {x} ∪ Y ⊆ L such that Y < α, if
x ∧W Y exists then WY (x ∧ y) also exists and the two are equal. Note that when
3 ≤ n ≤ ω, in the lattice case n-distributivity is just distributivity. When L is
a complete lattice and L is α-distributive for all α we say L is a frame.
From now on we fix a minimal join-specification V and we define Φ = ΓV .
Let σ be the radius of V and let χ be the smallest regular cardinal with σ ≤ χ.
Similarly let σ′ be the radius of UΦ (recall definition 2.10), and let χ′ be the
smallest regular cardinal with σ′ ≤ χ′.
Definition 3.2 (Υ). Given S ⊆ P we define the following subsets of P using
transfinite recursion.
• Υ0(S) = S↓.
• If α + 1 is a successor ordinal then Υα+1(S) = {W T : T ∈ UΦ and
T ⊆ Υα(S)}.
• If λ is a limit ordinal Υλ(S) = Sβ<λ Υβ(S).
We define Υ : ℘(P ) → ℘(P ) by Υ(S) = Υχ′ (S) for all S ∈ ℘(P ).
This definition differs from definition 2.6 in that we do not close downwards
during the successor steps. Also note the use of χ′ in place of χ. This is
important in the proof of corollary 3.6. Note that Υ will not necessarily be a
closure operator as it may not be idempotent.
Lemma 3.3. Υ(S) ⊆ Φ(S) for all S ∈ ℘(P ).
Proof. We have Φ = ΓV = ΓUΦ by proposition 2.13(3), and by definition of Υ
we have Υ(S) ⊆ ΓUΦ(S) for all S ∈ ℘(P ).
Lemma 3.4. Let J be an indexing set, and let I, Kj ∈ Φ[℘(P )] for all j ∈ J.
Then I ∩ Υ(SJ Kj) ⊆ Φ(SJ (I ∩ Kj)).
Proof. We note that Υ(SJ Kj) = Sα<χ′ Υα(SJ Kj). We proceed by showing
that I ∩ Υα(SJ Kj) ⊆ Φ(SJ (I ∩ Kj)) for all α using transfinite induction on
α.
If α = 0 the result is trivial so consider the successor ordinal α + 1 and
assume the appropriate inductive hypothesis. Let p ∈ I ∩ Υα+1(SJ Kj). Then
p = W T for some T ∈ UΦ with T ⊆ Υα(SJ Kj), and so since p ∈ I we have
T ⊆ I ∩Υα(SJ Kj). By the inductive hypothesis this means T ⊆ Φ(SJ (I ∩Kj)),
and thus by definition of UΦ we have p ∈ Φ(SJ (I ∩ Kj)) as required. The limit
case is trivial.
Theorem 3.5. The following are equivalent:
1. Φ(S) = Υ(S) for all S ∈ ℘(S).
2. Φ[℘(P )] is a frame when considered to be a lattice ordered by inclusion.
8
Proof. (1 =⇒ 2). Let J be an indexing set and let I, Kj ∈ Φ[℘(P )] for all
j ∈ J. We must show that I ∩ Φ(SJ Kj) = Φ(SJ (I ∩ Kj)). Since we are
assuming Φ = Υ and the right side is always included in the left it remains only
to show that I ∩ Υ(SJ Kj) ⊆ Φ(SJ (I ∩ Kj)), and this is lemma 3.4.
(2 =⇒ 1). We note that Φ0(S) = Υ0(S) for all S ∈ ℘(P ) by definition,
and we proceed by using transfinite induction to show that Φα(S) ⊆ Υα(S) for
all cardinals α and for all S ∈ ℘(P ). As Υ(S) ⊆ Φ(S) by lemma 3.3 the result
then follows.
Let S ∈ ℘(P ) and suppose Φα(S) ⊆ Υα(S) for some cardinal α. Let T ⊆
Φα(S) and suppose T ∈ U. Let p ≤ W T . Then in Φ[℘(P )] we have p↓ ∩WT t↓ =
p↓ ∩ (W T )↓ = p↓ by proposition 2.9. Since Φ[℘(P )] is a frame we also have
WT (p↓ ∩ t↓) = p↓, and thus WT (p↓ ∩ t↓) = Φ(ST (p↓ ∩ t↓)) = Φ(T ↓ ∩ p↓) = p↓.
So by lemma 2.18(1) we have p = W(T ↓ ∩ p↓), and thus by lemma 2.18(2) we
have T ↓ ∩ p↓ ∈ UΦ. But T ↓ ⊆ Φα(S) ⊆ Υα(S), and thus T ↓ ∩ p↓ ⊆ Υα(S), and
so p ∈ Υα+1(S) as required. The limit case is trivial and so we are done.
Corollary 3.6. If Φ[℘(P )] is not a frame then there is T ∈ V and p ∈ P such
that
1. p ≤ W T , and
2. φ(p) ∧ W φ[T ] 6= WT (φ(p) ∧ φ(t)) in Φ[℘(P )].
Here φ is the canonical map from P into Φ[℘(P )] taking p to p↓.
Proof. By theorem 3.5 if Φ[℘(P )] is not a frame then there is S ∈ ℘(P ) with
Φ(S) 6= Υ(S). Let α be the smallest cardinal such that Φα(S) 6⊆ Υ(S). This
exists by lemma 3.3 and the assumption that Φ(S) 6= Υ(S). Moreover, α cannot
be 0 or a limit cardinal so α = β + 1 for some β. Choose any p ∈ Φα(S) \ Υ(S).
Then by minimality of α we must have p /∈ Φβ(S), and p ≤ W T for some
In Φ[℘(P )] we have p↓ ∩ WT t↓ = p↓. However, if
T ∈ V with T ⊆ Φβ(S).
p↓ = WT (p↓ ∩ t↓) = Φ(T ↓ ∩ p↓), then p = W(T ↓ ∩ p↓) and T ↓ ∩ p↓ ∈ UΦ by
lemma 2.18. Since T ↓ ∩ p↓ ⊆ Φβ(S), by choice of α we have T ↓ ∩ p↓ ⊆ Υ(S).
Since Υ(S) is closed under joins from UΦ (by an argument similar to that in the
proof of proposition 2.8), and p /∈ Υ(S) by choice of p, we must therefore have
p↓ 6= WT (p↓ ∩ t↓) to avoid contradiction.
Here the minimality of V is relevant as it gives a smaller upper bound on the
possible size of a cardinal α for which α-distributivity can fail. Two immediate
consequences of corollary 3.6 are the following.
Corollary 3.7. If Φ[℘(P )] is not a frame then Φ[℘(P )] must fail to be σ-
distributive (recall that σ is the radius of V).
Proof. If Φ[℘(P )] is not a frame then distributivity fails for φ(p) ∧ W φ[T ] for
some T ∈ V, and T < σ by definition.
Corollary 3.8. Given a poset P the lattice of all down-sets of P closed under
existing finite joins is a frame if and only if it is distributive.
9
Proof. This lattice is produced by the closure operator arising from the join-
specification containing all non-empty finite sets with defined joins. The radius
of this join-specification is ω, and the result then follows from corollary 3.7.
In the case where κ is a regular cardinal and V = Uκ = {S ∈ ℘(P ) \ {∅} :
S < κ and W S exists} theorem 3.5 and its corollaries are superseded by [12,
theorem 2.7].
Indeed, in this case the mentioned theorem shows, using our
notation, that Φ[℘(P )] is a frame if and only if Φ(S) = {W T : T ∈ V and
T ⊆ S↓} for all S ∈ ℘(P ).
Various distributivity properties for join-completions of posets (and qua-
siorders) are investigated in [9]. See in particular [9, theorem 2.1] for a summary
of other properties equivalent to the join-completion corresponding to a given
standard closure operator being a frame.
4 Sentences guaranteeing representability
Definition 4.1 ((α, β)-representable). For cardinals α and β a poset P is (α, β)-
representable if there is an embedding h : P → F , where F is a powerset algebra,
such that h preserves meets of sets with cardinalities strictly less than α, and
joins of sets with cardinalities strictly less than β. When α = β we just write
α-representable.
The following definition appears under a slightly different name as [3, defi-
nition 2].
Definition 4.2 (MD). A poset P is MD (meet distributive) if for all a, b, c ∈ P ,
if a ∧ (b ∨ c) is defined then (a ∧ b) ∨ (a ∧ c) is also defined and the two are equal.
Given a cardinal α ≥ 3 we can generalize this to the following.
Definition 4.3 (MDα). P is MDα if whenever {a} ∪ X ⊆ P , X < α and
a ∧ W X is defined in P , we have WX (a ∧ x) is also defined in P and the two are
equal.
Note that for semilattices MDα is equivalent to α-distributivity. We use our
terminology to avoid confusion with the distributivity for posets defined in [12].
Given a cardinal γ suppose V = Uγ is the set of all sets S ⊆ P such that S < γ
and W S exists in P , and define Φ and Υ for this V as in section 3. Note that in
section 3 we say V is minimal, but minimality is not essential in the definitions
of Φ and Υ, or the theory developed therein. So the fact that Uγ may not be
minimal is not a problem here. We have the following theorem.
Theorem 4.4. Let γ be any cardinal strictly greater than 2. If P is MDγ then
Φ[℘(P )] is a frame.
Proof. By corollary 3.6 if Φ[℘(P )] fails to be a frame there must be T ∈ V and
p ∈ P such that p ≤ W T and φ(p) ∧ W φ[T ] 6= WT (φ(p) ∧ φ(t)) in Φ[℘(P )].
But by definition of V we must have T < γ, and so by proposition 2.9 and the
assumption that P is MDγ we have φ(p)∧W φ[T ] = φ(p∧W T ) = φ(WT (p∧t)) =
WT (φ(p) ∧ φ(t)), so Φ[℘(P )] must be a frame after all.
10
Corollary 4.5. Every MD poset has an (ω, 3)-representation.
Proof. Let V = U3 and let Φ be defined as in section 3. Then as Φ[℘(P )] is dis-
tributive it embeds into a powerset algebra F via a map preserving finite meets
and joins. Since φ : P → Φ[℘(P )] preserves binary joins and arbitrary meets
we obtain an (ω, 3)-representation for P by composing φ with the embedding of
Φ[℘(P )] into F .
By setting the value of γ appropriately we also obtain the following result.
Corollary 4.6. If 3 ≤ n ≤ ω then every MDn poset has an (ω, n)-representation.
When α > ω we do not obtain a result corresponding to corollary 4.6 because
a frame is not necessarily (ω, α)-representable in this case. We know this because
for every α > ω, every non-atomic countable Boolean satisfies MDα but is not
(ω, α)-representable, as discussed in the introduction.
Note that it follows from theorem 4.4 and [12, theorem 2.7] that if κ is a
regular cardinal we have MDκ =⇒ κ-distributive in the sense of [12] (which
for convenience we shall dub HMκ). Also by [12, theorem 2.7], being HMκ is
equivalent to having Φ(S) = {W T : T ∈ Uκ and T ⊆ S↓} for all S ∈ ℘(P ) (as
mentioned in the passage following corollary 3.8). Using this we can show that
6=⇒ MDκ for all regular κ, as example 4.7 below provides a poset that
HMκ
is HMκ, but fails to be MDκ, for all regular κ.
Example 4.7. Let P be the poset in the diagram below and let κ be any
regular cardinal. Given S ∈ ℘(P ) define Φ′(S) = {W T : T ∈ Uκ and T ⊆ S↓}.
Then a ∧ (b ∨ c) = c, but a ∧ b does not exist. So P fails to be MD3, and thus
fails to be MDκ. However, the only non-trivial non-principal downsets of P are
{b, c} and {a, b, c}, and Φ({b, c}) = (b ∨ c)↓ = Φ′({b, c}), and Φ({a, b, c}) = P =
Φ′({a, b, c}), and so P is HMκ for all regular κ.
•
•a
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
⑤
•b
•c
Combining the preceding discussion with [12, example 4.2 and theorem 4.8]
we obtain the following chain of strictly one way implications.
MDω =⇒ HMω =⇒ ω-representable
5 Obstructions to representability
Another way to view these results is to think about obstructions to a poset
being, for example, 3-representable. For the sake of this discussion we say P
has a triple (a, b, c) if there are a, b, c ∈ P such that a ∧ (b ∨ c) is defined. If
(a, b, c) is a triple of P and (a ∧ b) ∨ (a ∧ c) is defined but not equal to a ∧ (b ∨ c)
11
we call (a, b, c) a split triple, and we say a triple (a, b, c) is an indeterminate
triple if (a ∧ b) ∨ (a ∧ c) is not defined.
It's easy to see that having a split triple is a sufficient condition for a poset
P to fail to be 3-representable. Moreover, existence of a split triple can be
defined in first order logic, so if having a split triple were also a necessary con-
dition for a poset to fail to be 3-representable it would follow that the class
of 3-representable posets is finitely axiomatizable. This is not the case [6], so
there must be other, less obvious obstructions to 3-representability. This is not
surprising, as even in the simpler semilattice case the existence of a split triple
is not necessary for failure of 3-representability. However, the semilattices that
fail to be (ω, 3)-representable can be characterized as those that contain either
a split triple or an indeterminate triple (this is [2, theorem 2.2] phrased in the
terminology of triples), which is also a first order property. Since the class of
(m, n)-representable posets is elementary for all m and n with 2 < m, n ≤ ω, it
follows that the class of posets that fail to be 3-representable cannot be axiom-
atized in first order logic at all (otherwise it would be finitely axiomatizable, in
contradiction with [6]). This contrasts starkly with the intuitive finite axioma-
tization of the semilattice case.
Putting this another way, let S be the class of posets containing a split
triple, let I be the class of posets containing an indeterminate triple but no split
triple, let L be the class of MD-posets, and let R be the class of 3-representable
posets. Then S, I, and L are all basic elementary and partition the class of all
posets. Moreover, L ⊂ R, and S ⊂ ¯R. However, I ∩ R is elementary but not
finitely axiomatizable, and I ∩ ¯R is not even elementary. This contrasts with
the semilattice case where L and R coincide.
On the other hand, one reason we might expect these obstructions to repre-
sentability to defy simple characterization comes from computational complexity
theory. It was shown in [19] that the problem of deciding whether a finite poset
has an (m, n)-representation is NP-complete for countable m, n with at least
one greater than 3. So if there were a simple characterization of the finite posets
that fail to be (m, n)-representable, such as exists in the semilattice case, we
could potentially use this to prove that this decision problem is also in coNP,
which would imply the unexpected coincidence NP = coNP. More precisely,
by Fagin's theorem we have NP = coNP if and only if the finite posets that
fail to be (m, n)-representable (for suitable m and n) have an existential second
order characterization.
References
[1] A. Abian.
'Boolean rings with isomorphisms preserving suprema and in-
fima'. J. Lond. Math. Soc. (2) 3 (1971), 618–620.
[2] R. Balbes. 'A representation theory for prime and implicative semilattices'.
Trans. Amer. Math. Soc. 136 (1969), 261–267.
12
[3] Y. Cheng and P. Kemp. 'Representation of posets'. Zeitschr. f. math. Logik
und Grundlagen d. Math 38 (1992), 269–276.
[4] W. Cornish and R. Hickman.
'Weakly distributive semilattices'. Acta.
Math. Acad. Sci. Hungar. 32 (1978), 5–16.
[5] H. P. Doctor. Extensions of a partially ordered set (1967). Thesis (Ph.D.)–
McMaster University (Canada).
[6] R. Egrot. 'No finite axiomatizations for posets embeddable into distributive
lattices'. preprint ArXiv:1610.00858.
[7] R. Egrot. 'Representable posets'. J. Appl. Log. 16 (2016), 60–71.
[8] M. Ern´e.
'Closure'.
In Beyond topology, Contemp. Math., Volume 486
(Amer. Math. Soc., Providence, RI, 2009), 163–238.
[9] M. Ern´e and G. Wilke. 'Standard completions for quasiordered sets'. Semi-
group Forum 27 (1-4) (1983), 351–376.
[10] G. Gratzer and E.T. Schmidt.
'On congruence lattices of lattices'. Acta
Math. Acad. Sci. Hungar. 13 (1962), 179–185.
[11] R. Hickman. 'Distributivity in semilattices'. Acta Math. Acad. Sci. Hungar.
32 (1978), 35–45.
[12] R.C. Hickman and G.P. Monro. 'Distributive partially ordered sets'. Fund.
Math. 120 (1984), 151–166.
[13] K. Kearnes. 'The class of prime semilattices is not finitely axiomatizable'.
Semigroup Forum 55 (1997), 133–134.
[14] P. Kemp.
'Representation of partially ordered sets'. Algebra Universalis
30 (1993), 348–351.
[15] B. Schein. 'On the definition of distributive semilattices'. Algebra Univer-
salis 2 (1972), 1–2.
[16] J. Schmidt. 'Each join-completion of a partially ordered set is the solution
of a universal problem'. J. Austral. Math. Soc. 17 (1974), 406–413.
[17] J. Schmidt.
'Universal and internal properties of some completions of k-
join-semilattices and k-join-distributive partially ordered sets'. J. Reine
Angew. Math. 255 (1972), 8–22.
[18] J. Schmidt.
'Universal and internal properties of some extensions of par-
tially ordered sets'. J. Reine Angew. Math. 253 (1972), 28–42.
[19] C. Van Alten. 'Embedding Ordered Sets into Distributive Lattices'. Order
33 (2016), 419–427.
[20] J.C. Varlet.
'On separation properties in semilattices'. Semigroup Forum
10 (1975), 220–228.
13
|
1106.6191 | 3 | 1106 | 2011-12-21T09:00:09 | Splitting full matrix algebras over algebraic number fields | [
"math.RA",
"cs.SC",
"math.NT"
] | Let K be an algebraic number field of degree d and discriminant D over Q. Let A be an associative algebra over K given by structure constants such that A is isomorphic to the algebra M_n(K) of n by n matrices over K for some positive integer n. Suppose that d, n and D are bounded. Then an isomorphism of A with M_n(K) can be constructed by a polynomial time ff-algorithm. (An ff-algorithm is a deterministic procedure which is allowed to call oracles for factoring integers and factoring univariate polynomials over finite fields.)
As a consequence, we obtain a polynomial time ff-algorithm to compute isomorphisms of central simple algebras of bounded degree over K. | math.RA | math | Splitting full matrix algebras over algebraic number fields
G´abor Ivanyos
Computer
Automation
Research Institute, Hungarian
Acad. Sci;
and
[email protected]
1
1
0
2
c
e
D
1
2
Josef Schicho
Johann Radon Institute of Com-
putational and Applied Mathe-
matics;
Austrian Academy of Sciences
[email protected]
Lajos R´onyai
Computer
Automation
and
Research Institute, Hungarian
Acad. Sci.
Dept.
Univ.
nomics
of Algebra, Budapest
of Technology and Eco-
[email protected]
September 26, 2018
Abstract
]
.
A
R
h
t
a
m
[
3
v
1
9
1
6
.
6
0
1
1
:
v
i
X
r
a
Let K be an algebraic number field of degree d and discriminant ∆ over Q. Let A be
an associative algebra over K given by structure constants such that A ∼= Mn(K) holds for
some positive integer n. Suppose that d, n and ∆ are bounded. Then an isomorphism
A → Mn(K) can be constructed by a polynomial time ff-algorithm. An ff-algorithm is a
deterministic procedure which is allowed to call oracles for factoring integers and factoring
univariate polynomials over finite fields.
As a consequence, we obtain a polynomial time ff-algorithm to compute ismorphisms
of central simple algebras of bounded degree over K.
1
Introduction
In this paper we consider the following algorithmic problem, which we call explicit isomorphism
problem: let K be an algebraic number field, A an associative algebra over K. Suppose that A is
isomorphic to the full matrix algebra Mn(K). Construct explicitly an isomorphism A → Mn(K).
Or, equivalently, give an irreducible A module.
Recall that for an algebra A over a field K and a K-basis a1, . . . , am of A over K the products
aiaj can be expressed as linear combinations of the ai
aiaj = γij1a1 + γij2a2 + · · · + γijmam.
The elements γijk ∈ K are called structure constants. In this paper an algebra is considered to
be given as a collection of structure constants. The usual representation of a number field K
0 Key words and phrases: Central simple algebra, splitting, splitting element, Minkowski's theorem on
convex bodies, maximal order, real and complex embedding, lattice basis reduction, parametrization, Severi-
Brauer surfaces, n-descent on elliptic curves.
The work of Ivanyos and R´onyai was supported in part by OTKA grants NK 72845, K77476, and K77778.
1
over Q with the minimal polynomial f ∈ Z[x] of an algebraic integer α ∈ K with K = Q(α)
can also be considered this way.
For basic definitions and facts from the theory of finite dimensional associative algebras the
reader is referred to [37] and [39]. Let A be a finite dimensional associative algebra over K, which
is either a finite field or an algebraic number field. In [19] and [41] polynomial time algorithms
were proposed for the computation of the radical Rad(A), and for the computation of the
Wedderburn decomposition (the minimal two-sided ideals) of the semisimple part A/Rad(A).
The algorithm for the Wedderburn decomposition is probabilistic (Las Vegas) in the finite case,
the others are deterministic. Alternative methods, improvements and related results have been
obtained in [12], [13], [7], [15], [16], [17], [20], [23], [24],[1]. A recent survey is [3].
To obtain a decomposition of A into minimal left ideals, one has to be able to solve the
explicit isomorphism problem for simple algebras over K. In [41] this was shown to be possible
in randomized polynomial time when K is finite. This method was derandomized recently in [25]
in the case when the dimension of A over K is bounded. In [40] and [47] evidence (randomized
reduction) is presented, that over algebraic number fields the explicit isomorphism problem
problem is at least as difficult as the task of factoring integers, a problem not known to be
amenable to polynomial time algorithms. For simple algebras over a number field K polynomial
time Las Vegas algorithms were given in [12] and [2] to find a number field L ⊇ K such that
A ⊗K L ∼= Mn(L) for a suitable n, together an explicit representation of the isomorphism. In
[14] a real version was established: if K ⊂ R, and A splits over R, then it can be achieved that
L ⊂ R. These results were derandomized in part in [43], and completely in [20].
Following [42] we recall the notion of an ff-algorithm. It is an algorithm which is allowed to
call an oracle for two types of subproblems. These are the problem of factoring integers, and
the problem of factoring polynomials over finite fields. We have no deterministic polynomial
time algorithms for these problems (but the latter one admits polynomial time randomized
algorithms). In both cases the cost of the oracle call is the length of the input to the call.
In [42] the problem of deciding if A ∼= Mn(K) holds for an algebra A over a number field
K was shown to be in NP ∩ coNP . The proof relies on properties of maximal orders Λ ≤ A
for central simple algebras A over K. Maximal orders are in many ways analogous to the full
ring of algebraic integers in K. The principal result of [26] is a polynomial time ff-algorithm to
construct maximal orders in simple algebras over Q. A very similar algorithm is presented in
[35]. In [47] a more direct method is given for quaternion algebras.
Several of the algorithms mentioned here have implementations in the computer algebra
system Magma, see for example [33].
We mention also a somewhat surprising application of the algorithms for orders: they have
been applied in the construction and analysis of high performance space time block codes for
wireless communication, see [22], [45]. In fact, in addition to an application of the algorithm of
[26], in [22] an improvement is suggested for the orders relevant there.
The main result of this paper is a polynomial time ff-algorithm for the case when A is a
central simple algebra of bounded dimension over a small extension field K of Q. This was
known before only in the smallest nontrivial case dimQ A = 4, see [27] and the more recent
papers [11], [46], [47]. More precisely we have the following.
Theorem 1. Let K be an algebraic number field of degree d and discriminant ∆ over Q. Let
A be an associative algebra over K given by structure constants such that A ∼= Mn(K) holds
for some positive integer n. Suppose that d, n and ∆ are bounded. Then an isomorphism
A → Mn(K) can be constructed by a polynomial time ff-algorithm.
2
We remark, that the algorithm of Theorem 1 gives an explicit isomorphism even if we do not
assume that log ∆, n, and d are bounded. However, the running time then may be exponential
in these parameters. This holds also for the algorithmic applications given in the last section
of the paper.
In addition to computational representation theory where the problem naturally originates
from, the explicit isomorphism problem arises also in connection with computational problems
of arithmetic geometry:
in a series of seminal papers [8], [9], and [10] the n-Selmer group of
an elliptic curve E over a number field K is studied. A method is developed to represent the
elements of the Selmer group as genus one normal curves of degree n. One of the key ingredients
of their method is to solve the explicit isomorphism problem for Mn(K). In [10] an algorithm
is outlined for the explicit isomorphism problem over K = Q, and is detailed for the cases
n = 3, 5. Our approach is based on similar ideas.
An algorithm for explicit isomorphisms is useful also for computing parametrizations in
algebraic geometry:
[11] considers parametrizations of conics, and [21] gives algorithms for
rational parametrization of Severi-Brauer surfaces. In fact, in [21] an algorithm is given which
solves the explicit isomorphism problem when A ∼= M3(Q). This, however, uses a procedure for
solving norm equations whose complexity was not clear so far. For example it was not known
if they can be solved in ff-polynomial time. The case A ∼= M4(Q) is treated similarly in [38].
The organization of the paper is as follows. First, in Section 2 we prove Theorem 1 in
the simpler case K = Q. This combines the approach of Fisher [18] (that is used in [10] as
well), which considers a real embedding of A, with an application of Minkowski's theorem on
convex bodies, and with approximate lattice basis reduction. In the next section the argument
is extended to number fields. An important role is played here by the traditional map in
algebraic number theory which maps K into Rd, see Section 13, Chapter I. in [29].
In the last section two applications are presented. One of these is a polynomial time ff-
algorithm to compute ismorphisms of central simple algebras of bounded degree over K.
Acknowledgement. We are grateful to G´eza K´os and S´andor Z. Kiss for discussions on the
subject. We thank Jacques-Arthur Weil for calling our attention to [18].
2 Full matrix algebras over Q
Here we consider the case K = Q of Theorem 1. We prove first a statement on the existence of
small and highly singular elements in maximal orders.
Theorem 2. Let A be a Q-subalgebra of Mn(R) isomorphic to Mn(Q) and let Λ be a maximal
Z-order in A. Then there exists an element C ∈ Λ which has rank 1 as a matrix, and whose
Frobenius norm kCk is less than n.
Remark 3. When we apply the above theorem, the Frobenius norm k · k will be inherited from
Mn(R), with respect to an arbitrary embedding of A into Mn(R). Recall that for a matrix
X ∈ Mn(R) we have kXk =pT r(X T X).
Proof. The isomorphism A ∼= Mn(Q) extends to an automorphism of Mn(R). Therefore, by the
Noether-Skolem Theorem, there exists a matrix P ∈ Mn(R) such that A = P Mn(Q)P −1. Let Λ′
denote the standard maximal order Mn(Z) in Mn(Q). The theory of maximal orders in central
3
simple algebras over Q implies that there exists an invertible rational matrix P ′ ∈ Mn(Q)
such that it gives us P −1ΛP from Λ′: P −1ΛP = P ′Λ′P ′−1, whence Λ = P P ′Λ′P ′−1P −1. Set
Q = P P ′/( det P det P ′)1/n. Clearly Q ∈ Mn(R), det Q is ±1 and
Λ = QΛ′Q−1.
Let ρ denote the left ideal of Λ′ consisting of all integer matrices which have 0 everywhere
except in the first column. Clearly ρ is a lattice of determinant 1 in the linear space S of
all real matrices having nonzeros only in the first column. Now the lattice L = Qρ will be a
sublattice of S, with determinant 1 (see Subsection 2.2 from [34] for basic facts on lattices in
real Euclidean spaces).
We can apply Minkowski's theorem on lattice points in convex bodies to L in S, and to the
ball of radius √n in S centered at the zero matrix (we refer here to the Euclidean distance,
that is, the Frobenius norm on Mn(R)). The volume (calculated in S) of the ball is more than
2n, as it contains 2n internally disjoint copies of the n-dimensional unit cube, and more. We
infer that there exists an element B ∈ ρ such that QB is a nonzero matrix whose length is less
than √n. Clearly B and hence QB is a rank 1 matrix.
Next consider the "transpose" of this argument with Q−1 in the place of Q: there exists a
nonzero integer matrix B′, which is zero everywhere except in the first row, such that B′Q−1
is nonzero, and has Euclidean length less than √n.
Now
C = QBB′Q−1
meets the requirements of the statement.
length less than n because the Frobenius norm is submultiplicative:
Indeed, it is in Λ because BB′ ∈ Mn(Z).
It has
kCk = k(QB)(B′Q−1)k ≤ kQBk · kB′Q−1k < (√n)2 = n.
Obviously, C has rank at most 1, as B and B′ are of rank 1. Finally, from the shape of B and
B′ we see, that BB′ 6= 0, hence rank BB′ = rank C = 1. This finishes the proof.
Remark. Essentially the above reasoning shows the existence of a rank one C ∈ Λ such that
kCk ≤ γn, where γn is Hermite's constant (see Chapter IX, [5]). This bound is achieved if we
select B and B′ whose norm is at most most √γn. This gives a better bound for large values
of n.
small are necessarily zero divisors.
The following two lemmas point out that elements X form an order Λ ⊂ Mn(Q) with kXk
Lemma 4. Let X ∈ Mn(C) be a matrix such that det X is an integer, and kXk < √n. Then
X is a singular matrix.
Proof. The argument is essentially from [18]. Let X = QR be the QR decomposition of X,
with Q unitary and R an upper triangular matrix whose diagonal entries are r1, r2, . . . , rn. We
have
1
nkXk2 < 1.
det X2/n = (r12r22 · · ·rn2)1/n ≤
Here we used the fact that kXk =pT r(X ∗X) = pT r(R∗R) because Q∗Q = I. We conclude
(r12 + r22 + · · ·rn2) ≤
1
nkRk2 =
that det X = 0.
1
n
4
The next statement has a similar flavour. It was pointed out to us by our colleague G´eza
K´os.
Lemma 5. Let X ∈ Mn(Q) be a matrix whose characteristic polynomial has integral coeffi-
cients, and kXk < 1. Then X is a nilpotent matrix.
Proof. The eigenvalues of X are algebraic integers, hence the eigenvalues of X t are algebraic
integers as well, for any positive integer t. We infer that the characteristic polynomial of X t
has integral coefficients. Also, the norm condition implies that X t tends to the zero matrix O
as t → ∞, hence X t = O for a sufficiently large t.
The following argument is from H. W. Lenstra, see p. 546 in [32]. Informally, it states that
the coefficients with respect to a reduced basis of a vector v with small length v from a lattice
Γ are relatively small.
Lemma 6. Let Γ be a full lattice in Rm. Suppose that we have a basis b1, . . . , bm of Γ over Z
such that
holds for a real number cm > 0. Suppose that
b1 · b2 · · ·bm ≤ cm · det(Γ)
(1)
v =
m
Xi=1
γibi ∈ Γ,
γi ∈ Z.
Then we have γi ≤ cm
Proof. From Cramer's rule we obtain
v
bi for i = 1, . . . , m.
.
γi = det(b1, b2, . . . , bi−1, v, bi+1, . . . , bm)
= v
bi · b1 · · ·bi−1 · bi · bi+1 · · ·bm
det(Γ)
det(Γ)
det(Γ)
≤ b1 · · ·bi−1 · v · bi+1 . . .bm
≤ v
bi · cm ·
= cm · v
bi
det(Γ)
det(Γ)
.
=
We remark that the LLL algorithm gives a basis with cm = 2m(m−1)/4 in formula (1), see
[31]. We shall have a lattice of vectors with nonrational coordinates, and thus invoke the
approximate version of the LLL algorithm developed by Buchmann, see Corollary 4 of [4]. This
will provide a reduced basis with
cm := (γm)
m(m−1)
2
.
2
(2)
m
2(cid:19)m
2 (cid:18)3
Here γm is Hermite's constant. It is known that γm ≤ m for all integers m ≥ 1, and γm
1
πe + o(1) for m large.
m ≤
We can describe now the algorithm of Theorem 1 for the case K = Q. Suppose that, as
input, we have an algebra A over Q, given to us by structure constants. Suppose also that
A is isomorphic to the full matrix algebra Mn(Q). Our objective is to give this isomorphism
explicitly. More specifically the algorithm outputs an element C ∈ A which has rank 1 in
Mn(Q). Then the left action of A on AC provides an A → Mn(Q) isomorphism. The major
steps of the algorithm are the following.
5
1. Use the Ivanyos-R´onyai algorithm [26] to construct a maximal order Λ in A. This is a
polynomial time ff-algorithm1.
2. Compute an embedding of A into Mn(R). One uses here the the deterministic polynomial
time algorithm obtained via the derandomization by de Graaf and Ivanyos [20] of the Las
Vegas algorithm of Eberly [14]. This way we have a Frobenius norm on A. For X ∈ A we
can set kXk = pT r(X T X). Also, via this embedding Λ can be viewed as a full lattice
in Rm, where m = n2. The length v of a lattice vector v is just the Frobenius norm of
v as a matrix.
3. Compute a rational approximation A of our basis B of Λ with precision q0(B, 1
2 )
(see Section 2 in [4] for the definition of the precision parameter q0). One can use here
the Algorithm of Schonhage2 [44].
2, 2
m−1
4. Compute a reduced basis b1, . . . , bm of the lattice Λ ⊂ Rm by applying the LLL algorithm
to A. For cm we have the value from (2).
5. If some of the basis elements bi is a zero divisor in A, then there are two cases.
If
rank bi = 1, then we are done and stop with the output C := bi. Otherwise, if 1 <
rank bi < n, then we compute the the right identity element e of the left ideal Abi by
solving the straightforward system of linear equations, set A := eAe and go back to Step
1.
6. At this point we know that bi ≥ √n holds for every i. Generate all integral linear
bi ≤ cm√n until a C
combinations C ′ = Pm
i=1 γibi, where the γi are integers, γi ≤ cm
is found with rank C = 1. Output this C.
n
Proof of theorem 1 for K = Q. As for the correctness of the algorithm, let b1, . . . , bm the basis
of Λ obtained at Step 4 with kb1k ≤ · · · ≤ kbmk. Then by Corollary 4 from [4] we have
kbik ≤
3
2 · 2
m−1
2 λi
for i = 1, . . . , m,
where λi is the i-th successive minimum of Λ. From this we infer
2(cid:19)m
kb1kkb2k · · ·kbmk ≤(cid:18)3
m(m−1)
2
2
λ1 · · · λm ≤ (γm)
m
2(cid:19)m
2 (cid:18)3
m(m−1)
2
2
det(Λ),
as claimed. At the last inequality we used Minkowski's inequality on successive minima (see
Chapter VIII in [5]).
We remark also that, if at Step 5 we have rank e = k, then it is easy to see that eAe ∼= Mk(Q).
Moreover, a rank one element of eAe will have rank one in A as well. At Step 6 the bi are
nonsingular matrices, hence kbik ≥ √n holds by Lemma 4. Finally, Theorem 2 and Lemma 6
(this is applied for v := C and v ≤ n) show that an element C with rank one exists among
the linear combinations enumerated.
1It performs well if the integers to be factored are not very big. The method has been implemented in
Magma by de Graaf.
2For a more recent method see [36].
6
Considering the timing of the algorithm, Step 1 runs in polynomial time as an ff-algorithm.
Steps 2, 4 and 5 can be done in deterministic polynomial time. At Step 3 the precision paramater
q0 is polynomial in the input size, hence Schonhage's approximation algorithm (see also Section
3 of [30]) runs in polynomial time.
The number of jumps back to Step 1 is also bounded, hence each Step is carried out in a
bounded number of times. Finally, the number of elements C ′ enumerated at Step 6 is at most
(2cm√n + 1)m, this is also bounded by our assumption.
Remarks. 1.
In Step 4 of the preceding algorithm one may also consider the idempotent
f = I − e, where I is the identity element of A. If rankf = 1, then we can stop with C := f .
Otherwise, if rankf < ranke, then we may work with fAf instead of eAe.
2. We could avoid jumps back to Step 1 if we had a good lower bound on the quantities kbik.
Unfortunately, we do not have such a bound in general. The difficulty here may come from
the fact, that the closure of the similarity-orbit of nilpotent matrices contains the zero matrix.
This is illustrated by the matrices
X =(cid:18) t 0
0 1
t (cid:19) , E =(cid:18) 0 1
0 0 (cid:19) .
We have XEX −1 = t2E, hence kXEX −1k gets arbitrarily small as t → 0.
3. We could have used Lemma 5 instead of Lemma 4. In this case we test in Step 4 if there is
a nilpotent element among the bi. Also, then in Step 5 we have to enumerate integral linear
combinations Pm
i=1 γibi with γi ≤ cm · n.
3 The general case
Let K be a number field of degree d over Q, the maximal order of K is denoted by R and the
positive discriminant of R is ∆. Let A be a central simple algebra over K such that A ∼= Mn(K),
and let Λ be a maximal order in A.
It is known (see Reiner [39], Corollary 27.6) that there is an isomorphism ψ : A → Mn(K)
such that the image of Λ is
Λ′ := ψ(Λ) =
. . .
R · · · R J −1
...
...
R · · · R J −1
J · · · J R
,
where J is a fractional ideal of R in K. (The notation with a matrix having sets as entries refers
to all matrices (xij)n
i,j=1 whose elements belong to the designated sets, for example, x11 ∈ R,
xn1 ∈ J, etc.) Let σ1, . . . , σr be the embeddings of K into R and σr+1, σr+1, . . . , σr+s, σr+s be
the non-real embeddings of K into C; here we have d = r + 2s.
For each 1 ≤ i ≤ r + s let us consider an embedding φi of A into Mn(C), which extends
σi (for i ≤ r we require φi(A) ≤ Mn(R)). We remark that such embeddings can actually be
computed efficiently by the methods of [14] and [20]. For x ∈ A the matrices φi(x) are in
Mn(C), hence we may speak about the absolute value of their entries. Set
b = (cid:18) 2
π(cid:19)2sn
∆n!
1
nd
2s
d
1
d .
∆
=(cid:18) 2
π(cid:19)
7
Theorem 7. There exists a rank one element x ∈ Λ such that the entries of the matrices φi(x)
for i = 1, . . . , s + r all have absolute value at most b.
Proof. Let ψi : A → Mn(C) be the composition of ψ with the natural extension of σi to Mn(C).
These maps are shown at the diagram below. The vertical map is the extension of σi from K
to Mn(K). The triangle is commutative.
Mn(C)
φi←− A
ψ
ψi
−→ Mn(K)
ց ↓σi
Mn(C)
Then the C-linear extensions of the composite maps φiψ−1
from ψi(A) to Mn(C) are C-
algebra automorphisms of Mn(C) (whose restrictions, for i = 1, . . . , r, to the real matrices
are automorphisms of Mn(R)). As these automorphisms must be inner, there exist matri-
ces A1, . . . , Ar ∈ GLn(R) with determinant ±1 and Ar+1, . . . , Ar+s ∈ SLn(C) such that for
i = 1, . . . , r + s we have
i
φi(Λ) = A−1
i Λ′Ai = A−1
i
σi(R)
...
· · · σi(R) σi(J −1)
. . .
...
σi(R)
σi(J)
· · · σi(R) σi(J −1)
σi(R)
· · · σi(J)
Ai.
i := (A−1
i )T . We show that there exist nonzero vectors u ∈ (R, . . . , R, J) ⊂ Kn and
Put A′
v ∈ (R, . . . , R, J −1) ⊂ Kn such that for every index i = 1, . . . , r, all the coordinates of σi(u)A′
i
and σi(v)Ai are of "small" absolute values. Then all the entries of the matrix φiψ−1(uT v) will
be small, demonstrating that there exist a rank one element of Λ, namely ψ−1(uT v), which is
small in all the embeddings φi.
To this end, we consider the set M of row vectors of length nd of the form
(σ1(u), . . . , σr(u), σr+1(u), σr+1(u), . . . , σr+s(u), σr+s(u)),
(3)
where u ∈ (R, . . . , R, J). M is a lattice in the linear space Cdn whose rank is nd because of
the linear independence of field automorphisms, see Theorem I.3 in [28]. The determinant of
lattice M is
∆n/2N(J),
where N(J) is the norm of the fractional ideal J (see Proposition 13.4, Chapter I. in [29]). Next
we consider the set M′ of vectors of the form
(σ1(u)A′
1, . . . , σr(u)A′
r, σr+1(u)A′
r+1, σr+1(u)A′
r+1, . . . , σr+s(u)A′
r+1, σr+s(u)A′
r+s).
This set is obtained by multiplying vectors from M by the block diagonal matrix
1, . . . , A′
r, A′
r+1, A′
r+1, . . . , A′
r+1, A′
diag(cid:0)A′
r+s(cid:1) .
Here each block has determinant ±1, therefore the determinant of M′ remains ∆n/2N(J).
Finally we apply the block diagonal matrix
diag(cid:18)I, . . . , I,(cid:18) 1
2I −ι
2 I
ι
1
2I
2I (cid:19) , . . . ,(cid:18) 1
2 I −ι
2 I
ι
1
2 I
2I (cid:19)(cid:19) ,
8
where I stands for the n by n identity matrix, and we have r blocks of I. The determinant of
this matrix is (ι/2)ns. From M′ we obtain the lattice L of rank nd in Rnd ⊂ Cnd consisting of
the vectors
(σ1(u)A′
1, . . . , σr(u)A′
r,ℜ(σr+1(u)A′
r+1),ℑ(σr+1(u)A′
r+1),
(4)
where u runs over (R, . . . , R, J) ⊂ Kn. The determinant of L is 2−sn∆n/2N(J). We apply now
Minkowski's theorem on convex bodies to the lattice L and to the product of rn one-dimensional
balls and sn two-dimensional balls of radius
r+1), . . . ,ℜ(σr+s(u)A′
r+1),ℑ(σr+s(u)A′
r(J) =(cid:18)(cid:18) 2
π(cid:19)sn
N(J)∆n/2(cid:19)
1
nd
.
This is a closed convex centrally symmetric (with respect to the origin) body of volume
(2r(J))rn(cid:0)πr(J)2(cid:1)sn
.
This volume is 2nd det L. The theorem tells us that there exists a nonzero u ∈ (R, . . . , R, J)
such that for every 1 ≤ i ≤ r + s, all the coordinates of σi(u)A′
i have absolute value at most
r(J).
Similarly, there exists a nonzero vector v ∈ (R, . . . , R, J −1) such that for every 1 ≤ i ≤ r +s,
all the coordinates of σi(v)Ai have absolute value at most r(J −1) where
r(J −1) =(cid:18)(cid:18) 2
π(cid:19)sn
N(J)−1∆n/2(cid:19)
1
nd
.
Then x = ψ−1(uT v) is a rank one element of Λ such that for every i, all the entries of the
matrix φi(x) have absolute value at most
r(J)r(J −1) = (cid:18) 2
π(cid:19)2sn
∆n!
1
nd
2s
d
=(cid:18) 2
π(cid:19)
1
d = b.
∆
We point out two interesting consequences:
1. If K = Q, R = Z, then ∆ = 1, s = 0, hence b = 1. We have an element x of our maximal
order Λ which has rank 1 as a matrix from Mn(Q), and with respect to our selected embedding
of A into Mn(R) has elements of absolute value at most 1. This is essentially Theorem 2.
If D is a positive squarefree integer, K = Q(√D), then ∆ = D, if D is congruent to 1
2.
modulo 4, and ∆ = 4D, if D is congruent to 3 modulo 4. Then s = 0, d = 2, hence b ≤ 2√D.
Lemma 8. Let y ∈ Λ be an element such that kφi(y)k < √n holds for i = 1, . . . , r + s. Then
y is a zero divisor in A.
Proof. As in Lemma 4 we obtain that
To our algorithm we shall need a more general variant of Lemma 4.
det φi(y) < 1 for i = 1, . . . , r + s.
(5)
9
Note that det φi(y) = σi(n(y)), where n(y) is the reduced norm of y (see Section 9 in [39]).
Inequality (5) implies that
σ1(n(y)) · · · σr(n(y))σr+1(n(y))σr+1(n(y))· · · σr+s(n(y))σr+s(n(y)) < 1.
Moreover, by Theorem 10.1 from [39] n(y) ∈ R, therefore the number on the left is a rational
integer, giving that det φi(y) = 0 for at least one (and hence for all) i. This implies that y is a
zero divisor in A.
To be able to use lattice basis reduction techniques, we use a transformation which turns a
maximal order in A into a full lattice in a suitable real linear space. To this end for y ∈ Λ we
form the vectors
Φ(y) := (φ1(y), . . . , φr(y),ℜ(φr+1(y)),ℑ(φr+1(y)), . . . ,ℜ(φr+s(y)),ℑ(φr+s(y))).
As with (3) and (4), we infer that Γ := Φ(Λ) is a full lattice in the real linear space in Rm, with
m = n2d.
We give now the algorithm of Theorem 1 for the general case: as input, we have an algebra
A over K, given to us by structure constants. Suppose further, that A ∼= Mn(K). Our algorithm
outputs an element x ∈ A which has rank 1 in Mn(K).
1. Use the Ivanyos-R´onyai algorithm [26] to construct a maximal order Λ in A.
2. Compute the embeddings φi of A into Mn(C) for i = 1, . . . , r + s (they are embeddings
into Mn(R) for i ≤ r) by the deterministic variant [20] of Eberly's algorithm [14].
3. Form a basis of the full rank lattice Γ ⊂ Rm with m = n2d. Note that for the Euclidean
length in Γ we have
Φ(y)2 =
r+s
Xi=1
kφi(y)k2.
4. Compute a reduced basis b1, . . . , bm of the lattice Γ ⊂ Rm by using Buchmann's approx-
imate version the LLL algorithm to achieve the value in (2) for the reducedness factor
cm.
5. If an element y = Φ−1(bi) is a zero divisor in A, then there are two cases. If rank y = 1,
then we are done and stop with the output x := y. Otherwise, if 1 < rank y < n, then we
compute the the right identity element e of the left ideal Ay, set A := eAe and go back
to Step 1.
6. At this point we know that bi ≥ √n holds for every i. Generate all linear combinations
w =Pm
i=1 γibi, where the γi are rational integers with
2s
d
γi ≤ cm
bn√r + s
bi
π(cid:19)
≤ cmbpn(r + s) = cm(cid:18) 2
∆
1
dpn(r + s)
until a w is found such that rank x = 1 holds for the x ∈ Λ with Φ(x) = w. Output this
x.
10
Proof of Theorem 1. The proof is essentially the same as in the simpler case K = Q. At Step 6
Φ−1(bi) is necessarily a nonsingular element of Λ for i = 1, . . . , r + s. By Lemma 8 there must
be a j such that kφj(Φ−1(bi))k ≥ √n, giving that bi ≥ √n. Theorem 7 and Lemma 6, the
latter is applied with v ≤ bn√r + s, show that an element w with rank Φ−1(w) = 1 exists
among the linear combinations enumerated.
Here also each Step is carried out in a bounded number of times. The number of elements
w enumerated at Step 6 is at most (2cmbpn(r + s) + 1)m. This is also bounded by our
assumptions.
4 Two consequences
From the elementary theory of the Brauer group (see for example Section 12.5 from [37]) we
know that for two central simple algebras A and B of the same dimension n2 over a field K we
have A ∼= B if and only if
(6)
A ⊗K Bop ∼= Mn2(K).
We outline next that, over an infinite K, how one can efficiently recover from an isomorphism
(6) an isomorphism σ : A → B.
Having isomorphism (6) explicitly implies that we have in our hands a left A⊗KBop- module
V of dimension n2 over K. Then V , as a left A-module, is isomorphic to the regular left A-
module because they have the same dimension over K. There exists an element v ∈ V such
that the map φv : a 7→ av is a left A-module isomorphism from A to V . The elements v of V
which do not generate V as a left A-module are zeros of a certain polynomial on V of degree n2
(the determinant of the linear map a 7→ av). Similarly, the elements v of V for which the map
ψv : b 7→ vb is not a right Bop-module isomorphism between Bop and V are zeros of a polynomial
on V of degree n2. Therefore, by the Schwartz-Zippel Lemma there exists an element v ∈ V
for which the maps φv and ψv are simultaneously left and right isomorphisms, respectively.
The methods of [2] or [6] for finding large cyclic submodules can be used to obtain first a
left A-module generator V and then essentially the same method can be applied to gradually
transform v to a generator of V as a right Bop-module while preserving the property that v is a
left A-module generator for V . For example, the the method of Lemma 8 from [6] can be used
here. We recall the statement of the lemma for the reader's convenience.
Lemma 9. Let V be an r-dimensional module over the semisimple K-algebra A and v1, . . . , vr
be a K-basis of V . Assume that v ∈ V is an element of non-maximal rank. Let Ω be a subset
of K∗ consisting of at least rk v + 1 elements. Then there exists a scalar ω ∈ Ω and a basis
element u ∈ {v1, . . . , vr} such that rk(v + ωu) > rk v. (Here the rank rk v of v is defined as the
dimension of the A-submodule of V generated by v.)
We claim that if v ∈ V is an element such that φv and ψv are simultaneously isomorphisms
of the respective modules, then σ = ψ−1
v φv is an algebra isomorphism between A and B. It is
obvious that σ is a K-linear isomorphism between A and B. Note that for a ∈ A, σa is the
unique element b ∈ B with the property av = vb. Therefore σ(a1a2) is the unique element of B
with a1a2v = vb. But a1a2v = a1v(σa2) = v(σa1)(σa2), whence σ(a1a2) = (σa1)(σa2).
Combining this argument with the algorithm of Theorem 1 for constructing a suitable
module V , we obtain the following:
11
Corollary 10. Let K be an algebraic number field of degree d and discriminant ∆ over Q. Let
A,B be central simple algebras over K of the same dimension n2 given by structure constants.
Suppose that d, n and ∆ are bounded.
If A and B are isomorphic, then an isomorphism
A → B can be constructed by a polynomial time ff-algorithm.
The next statement is quite modest. It formulates a very plausible claim, but, to the best
of our knowledge, it was not proven before.
Corollary 11. Let K be an algebraic number field and A be an associative algebra over K given
by structure constants such that A ∼= Mn(K) holds for some integer n > 1. Then there exists
a zero divisor x ∈ A which admits polynomially bounded coordinates with respect to the input
basis of A. Moreover, such a zero divisor x can be obtained by a polynomial space bounded
computation.
Proof. A slight modification of the algorithm of Therorem 1 will provide a reasonably small
zero divisor: at Step 5 we stop if y is a zero divisor. Note that y has polynomial size as Steps
1-5 constitute a polynomial time ff-alghorithm.
If no zero divisor is found at Step 5, then
we proceed directly to Step 6. The integral linear combinations considered there have size
polynomial in the input length, and their enumeration can be carried out using polynomial
space only.
Remark. A more direct, but perhaps algorithmically less efficient proof of Corollary 11 is
possible. Let c1, . . . , cn2 be the basis of Λ given by the Ivanyos R´onyai algorithm. Express the
element x of Theorem 7 in this basis:
x = α1c1 + α2c2 + · · · + αn2cn2,
with αi ∈ Z. Using that kxk ≤ bn, and that the vectors ci have polynomial size, Cramer's rule
implies a polynomial bound on the size of the coefficients αi.
By the well known connection between split cyclic algebras and relative norm equations (see
Theorem 30.4 in Reiner [39]), our results imply that for a number field K and a cyclic extension
L of K if a norm equation NL/K(x) = a is solvable, then there is a solution whose standard
representation has polynomial size (in terms of the size of the standard representation of a and
a basis of L). Furthermore, for fixed K and fixed degree L : K, a solution can be found by a
polynomial time ff-algorithm.
We have given here a polynomial time ff-algorithm for the explicit isomorphism problem for
central simple algebras A of fixed dimension over a fixed number field K. Potential directions
to extend this result may be allowing the dimension of the algebra over K to grow or allowing K
to vary (even if its degree over Q remains fixed), or both. Existence of ff-algorithms for finding
an explict isomorphism of a non-split central simple algebra with the algebra of matrices over a
skewfield is also left open (even in the case of fixed base field, or fixed dimension). It would be
interesting also to develop practical variants and programs for the algorithms presented here.
References
[1] V. Acciaro, Solvability of norm equations over cyclic number fields of prime degree, Math.
Comp. 65(1996), 1663-1674.
12
[2] L. Babai, L. R´onyai, Computing irreducible representations of finite groups; Mathematics
of Computation, 192(1990), 705-722.
[3] M. R. Bremner, How to compute the Wedderburn decomposition of a finite-dimensional
associative algebra, manuscript, 2010, http://arxiv.org/pdf/1008.2006
[4] J. Buchmann, Reducing lattice bases by means of approximations, in: Algorithmic num-
ber theory, LNCS 877, Springer-Verlag, 1994, 160 -- 168.
[5] J. W. S. Cassels, An introduction to the geometry of numbers, Springer-Verlag, 1971.
[6] A. Chistov, G. Ivanyos, and M. Karpinski, Polynomial time algorithms for modules
over finite dimensional algebras, in Proceedings of the 1997 International Symposium on
Symbolic and Algebraic Comput, ACM, New York, 1997, pp. 68 -- 74.
[7] A. M. Cohen, G. Ivanyos and D. B. Wales, Finding the radical of an algebra of linear
transformations, J. Pure Appl. Algebra, 117/118(1997) 177-193.
[8] J. Cremona, T. Fisher, C. O'Neil, D. Simon, M. Stoll, Explicit n-descent on elliptic curves
I. Algebra, Journal fur die reine und angewandte Mathematik, 615(2008), 121 -- 155.
[9] J. Cremona, T. Fisher, C. O'Neil, D. Simon, M. Stoll, Explicit n-descent on elliptic curves
II. Geometry, Journal fur die reine und angewandte Mathematik, 632(2009), 63 -- 84.
[10] J. Cremona, T. Fisher, C. O'Neil, D. Simon, M. Stoll, Explicit n-descent on elliptic
curves III. Algorithms, ArXiv 1107.3516.
[11] J. E. Cremona, D. Rusin, Efficient solution of rational conics, Math. Comp. 72(2003),
1417-1441.
[12] W. M. Eberly, Computations for algebras and group representations, Ph. D. Thesis,
Dept. of Computer Science, University of Toronto, 1989.
[13] W. M. Eberly, Decomposition of algebras over finite fields and number fields; Compu-
tational Complexity, 1(1991), 179-206.
[14] W. M. Eberly, Decompositions of algebras over R and C; Computational Complexity,
1(1991), 207-230.
[15] W. Eberly, M. Giesbrecht, Efficient decomposition of associative algebras, Proceedings,
1996 International Symposium on Symbolic and Algebraic Computation, Zurich, Switzer-
land, pp. 170-178.
[16] W. Eberly, M. Giesbrecht, Efficient decomposition of associative algebras over finite
fields, J. Symbolic. Comput. 29(2000), 441-488.
[17] W. Eberly, M. Giesbrecht, Efficient decomposition of separable algebras. J. Symbolic
Comput. 37(2004), 35-81.
[18] T. Fisher, How to trivialise an algebra, Abstracts, Rational points on curves and higher
theory and explicit methods, Workshop at Jacobs University,
dimensional varieties:
Bremen, July 2007.
http://www.faculty.jacobs-university.de/mstoll/workshop2007/fisher2.pdf
13
[19] K. Friedl, L. R´onyai, Polynomial time solutions of some problems in computational
algebra; Proc. 17th ACM STOC, Providence, Rhode Island, 1985, 153-162.
[20] W. A. de Graaf, G. Ivanyos, Finding maximal tori and splitting elements in matrix
algebras, In: F. van Oystaeyen, M. Saorin (eds.), Interactions Between Ring Theory and
Representations of Algebras, (Proc. Euroconference in Murcia, 1998), Lecture Notes in
Pure and Applied Mathematics 210, Marcel Dekker 2000, 95-105.
[21] W. A. de Graaf, M. Harrison, J. P´ılnikov´a, J. Schicho, A Lie algebra method for rational
parametrization of Severi-Brauer surfaces, Journal of Algebra, 303(2006), 514 -- 529.
[22] C. Hollanti, J. Lahtnonen, K. Ranto, R. Vehkalahti, On densest MIMO lattices from
cyclic division algebras, IEEE Trans. Inform. Theory, 55(2009), 3751-3780.
[23] G. Ivanyos, Finding the radical of matrix algebras using Fitting decompositions, Journal
of Pure and Applied Algebra, 139(1999), 159-182.
[24] G. Ivanyos, Fast randomized algorithms for the structure of matrix algebras over finite
fields (extended abstract), Proc. 2000 Int. Symp. Symbolic Algebraic Computation (St.
Andrews), (ACM, New York, 2000), pp. 175-183
[25] G. Ivanyos, M. Karpinski, L. R´onyai, N. Saxena, Trading GRH for algebra: algorithms for
factoring polynomials and related structures, Mathematics of Computation, to appear.
[26] G. Ivanyos, L. R´onyai, Finding maximal orders in semisimple algebras over Q, Comput.
complexity, 3(1993), 245 -- 261.
[27] G. Ivanyos, ´A. Sz´ant´o, Lattice basis reduction for indefinite forms and an application,
Discrete Math. 153(1996), 177 -- 188.
[28] N. Jacobson, Lectures in abstract algebra III, Theory of fields and Galois theory, Springer-
Verlag, 1964.
[29] G.J. Janusz, Algebraic number fields, American Math. Soc., 1996.
[30] R. Kannan, A. K. Lenstra and L. Lov´asz, Polynomial factorization and nonrandomness
of bits of algebraic and some transcendental numbers, Mathematics of Computation,
50(1988), 235 -- 250.
[31] A. K. Lenstra, H. W. Lenstra, L. Lov´asz, Factoring polynomials with rational coefficients,
Math. Ann. 261(1982), 515 -- 534.
[32] H. W. Lenstra, Jr., Integer programming with a fixed number of variables, Mathematics
of Operations Research, 8(1983), 538 -- 548.
[33] http://magma.maths.usyd.edu.au/magma/handbook/text/840
[34] J. Matousek, Lectures on discrete geometry, GTM 212, Springer-Verlag, 2002.
[35] G. Nebe, A. Steel, Recognition of division algebras, Journal of Algebra, 322(2009), 903 --
909.
14
[36] V. Pan, Univariate polynomials: nearly optimal algorithms for numerical factorization
and root finding, Journal of Symbolic Computation, 33(2002), 701 -- 733.
[37] R. S. Pierce, Associative algebras, Springer-Verlag, 1982.
[38] J. P´ılnikov´a, Trivializing a central simple algebra of degree 4 over the rational numbers,
J. Symbolic Comput. 42(2007), 579 -- 586.
[39] I. Reiner, Maximal orders, Academic Press, 1975.
[40] L. R´onyai, Zero divisors in quaternion algebras, Journal of Algorithms, 9(1988) 494 -- 506.
[41] L. R´onyai, Computing the structure of finite algebras, Journal of Symbolic Computation,
9(1990) 355 -- 373.
[42] L. R´onyai, Algorithmic properties of maximal orders in simple algebras over Q, Comput.
Complexity, 2(1992), 225 -- 243.
[43] L. R´onyai, A deterministic method for computing splitting elements in semisimple alge-
bras over Q, Journal of Algorithms, 6(1994), 24 -- 32.
[44] A. Schonhage, The fundamental theorem of algebra in terms of computational complex-
ity, Preliminary report, Universitat Tubingen, 1982.
[45] B. A. Sethuraman, Division algebras and wireless communication, Notices Amer. Math.
Soc. 57(2010), 1432 -- 1439.
[46] D. Simon, Solving quadratic equations using reduced unimodular quadratic equations,
Math. Comp. 74(2005), 1531 -- 1543.
[47] J. Voight, Identifying the matrix ring: algorithms for quaternion algebras and quadratic
forms, manuscript, 2010.
15
|
1711.08381 | 1 | 1711 | 2017-11-19T14:59:15 | Symplectic, product and complex structures on 3-Lie algebras | [
"math.RA",
"math-ph",
"math.DG",
"math-ph"
] | In this paper, first we introduce the notion of a phase space of a 3-Lie algebra and show that a 3-Lie algebra has a phase space if and only if it is sub-adjacent to a 3-pre-Lie algebra. Then we introduce the notion of a product structure on a 3-Lie algebra using the Nijenhuis condition as the integrability condition. A 3-Lie algebra enjoys a product structure if and only if it is the direct sum (as vector spaces) of two subalgebras. We find that there are four types special integrability conditions, and each of them gives rise to a special decomposition of the original 3-Lie algebra. They are also related to $\huaO$-operators, Rota-Baxter operators and matched pairs of 3-Lie algebras. Parallelly, we introduce the notion of a complex structure on a 3-Lie algebra and there are also four types special integrability conditions. Finally, we add compatibility conditions between a complex structure and a product structure, between a symplectic structure and a paracomplex structure, between a symplectic structure and a complex structure, to introduce the notions of a complex product structure, a para-K\"{a}hler structure and a pseudo-K\"{a}hler structure on a 3-Lie algebra. We use 3-pre-Lie algebras to construct these structures. Furthermore, a Levi-Civita product is introduced associated to a pseudo-Riemannian 3-Lie algebra and deeply studied. | math.RA | math | Symplectic, product and complex structures on 3-Lie algebras
7
1
0
2
Yunhe Sheng and Rong Tang
Abstract
v
o
N
9
1
]
.
A
R
h
t
a
m
[
1
v
1
8
3
8
0
.
1
1
7
1
:
v
i
X
r
a
In this paper, first we introduce the notion of a phase space of a 3-Lie algebra and show
that a 3-Lie algebra has a phase space if and only if it is sub-adjacent to a 3-pre-Lie algebra.
Then we introduce the notion of a product structure on a 3-Lie algebra using the Nijenhuis
condition as the integrability condition. A 3-Lie algebra enjoys a product structure if and
only if it is the direct sum (as vector spaces) of two subalgebras. We find that there are four
types special integrability conditions, and each of them gives rise to a special decomposition
of the original 3-Lie algebra. They are also related to O-operators, Rota-Baxter operators and
matched pairs of 3-Lie algebras. Parallelly, we introduce the notion of a complex structure
on a 3-Lie algebra and there are also four types special integrability conditions. Finally, we
add compatibility conditions between a complex structure and a product structure, between
a symplectic structure and a paracomplex structure, between a symplectic structure and a
complex structure, to introduce the notions of a complex product structure, a para-Kähler
structure and a pseudo-Kähler structure on a 3-Lie algebra. We use 3-pre-Lie algebras to
construct these structures. Furthermore, a Levi-Civita product is introduced associated to a
pseudo-Riemannian 3-Lie algebra and deeply studied.
Contents
1 Introduction
2 Preliminaries
3 Representations of 3-pre-Lie algebras
4 Symplectic structures and phase spaces of 3-Lie algebras
5 Product structures on 3-Lie algebras
6 Complex structures on 3-Lie algebras
7 Complex product structures on 3-Lie algebras
8 Para-Kähler structures on 3-Lie algebras
9 Pseudo-Kähler structures on 3-Lie algebras
2
4
6
8
13
18
25
29
33
0Keyword: 3-Lie algebra, symplectic structure, complex structure, product structure, complex product structure,
para-Kähler structure, Kähler structure
1
1
Introduction
A symplectic structure on a Lie algebra g is a nondegenerate 2-cocycle ω ∈ ∧2g∗. The underlying
structure of a symplectic Lie algebra is a quadratic pre-Lie algebra [17]. An almost product
structure on a Lie algebra g is a linear map E satisfying E2 = Id. If in addition, E also satisfies
the following integrability condition
[Ex, Ey] = E([Ex, y] + [x, Ey] − E[x, y]),
∀x, y ∈ g,
then E is called a product structure. The above integrability condition is called the Nijenhuis
condition. An equivalent characterization of a product structure is that g is the direct sum (as
vector spaces) of two subalgebras. An almost complex structure on a Lie algebra g is a linear
map J satisfying J 2 = −Id. A complex structure on a Lie algebra is an almost complex struc-
ture that satisfies the Nijenhuis condition. Adding compatibility conditions between a complex
structure and a product structure, between a symplectic structure and a paracomplex structure,
between a symplectic structure and a complex structure, one obtains a complex product struc-
ture, a paraKähler structure and a pseudo-Kähler structure respectively. These structures play
important roles in algebra, geometry and mathematical physics, and are widely studied. See
[1, 2, 3, 4, 5, 8, 9, 12, 14, 15, 16, 18, 19, 29, 35] for more details.
Generalizations of Lie algebras to higher arities, including 3-Lie algebras and more generally, n-
Lie algebras [24, 28, 37], have attracted attention from several fields of mathematics and physics. It
is the algebraic structure corresponding to Nambu mechanics [25, 32, 36]. In particular, the study
of 3-Lie algebras plays an important role in string theory. In [13], Basu and Harvey suggested to
replace the Lie algebra appearing in the Nahm equation by a 3-Lie algebra for the lifted Nahm
equations. Furthermore, in the context of Bagger-Lambert-Gustavsson model of multiple M2-
branes, Bagger-Lambert managed to construct, using a ternary bracket, an N = 2 supersymmetric
version of the worldvolume theory of the M-theory membrane, see [6]. An extensive literatures are
related to this pioneering work, see [7, 26, 27, 33]. See the review article [20] for more details. In
particular, metric 3-algebras were deeply studied in the seminal works [21, 22, 23]. In [31], the
authors introduced the notion a Nijenhuis operator on an n-Lie algebra, which generates a trivial
deformation.
The purpose of this paper is to study symplectic structures, product structure and complex
structures on 3-Lie algebras and these combined structures.
In the case of Lie algebras, pre-
Lie algebras play important roles in these studies. It is believable that 3-pre-Lie algebras will play
important roles in the corresponding studies. Thus, first we introduce the notion of a representation
of a 3-pre-Lie algebra and construct the associated semidirect product 3-pre-Lie algebra. Several
important properties of representations of 3-pre-Lie algebras are studied.
Note that the notion of a symplectic structure on a 3-Lie algebra was introduced in [11] and
it is shown that the underlying structure of a symplectic 3-Lie algebra is a quadratic 3-pre-Lie
algebra. We introduce the notion of a phase space of a 3-Lie algebra g, which is a symplectic 3-Lie
algebra g ⊕ g∗ satisfying some conditions, and show that a 3-Lie algebra has a phase space if and
only if it is sub-adjacent to a 3-pre-Lie algebra. We also introduce the notion of a Manin triple
of 3-pre-Lie algebras and show that there is a one-to-one correspondence between Manin triples of
3-pre-Lie algebras and phase spaces of 3-Lie algebras.
An almost product structure on a 3-Lie algebra g is defined to be a linear map E : g −→ g satis-
fying E2 = Id. It is challengeable to add an integrability condition on an almost product structure
to obtain a product structure on a 3-Lie algebra. We note that the Nijenhuis condition (see (5))
given in [31] is the correct integrability condition. Let us explain this issue. Denote by g± the
2
eigenspaces corresponding to eigenvalues ±1 of an almost product structure E. Then it is obvious
that g = g+ ⊕ g− as vector spaces. The Nijenhuis condition ensures that both g+ and g− are sub-
algebras. This is what "integrability" means. Moreover, we find that there are four types special
integrability conditions, which are called strict product structure, abelian product structure, strong
abelian product structure and perfect product structure respectively, each of them gives rise to a
special decomposition of the original 3-Lie algebra. See the following table for a precise description:
product
structure
E[x, y, z]g = [Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−E([Ex, Ey, z]g + [Ex, y, Ez]g + [x, Ey, Ez]g)
strict product
structure
+[Ex, Ey, Ez]g
E[x, y, z]g = [Ex, y, z]g
abelian product
[x, y, z]g = −[x, Ey, Ez]g − [Ex, y, Ez]g − [Ex, Ey, z]g
structure
strong abelian
product structure
perfect product
structure
[x, y, z]g = E[Ex, y, z]g + E[x, Ey, z]g + E[x, y, Ez]g
O-operators
Rota-Baxter operators
E[x, y, z]g = [Ex, Ey, Ez]g
involutive automorphisms
g = g+ ⊕ g−
[g+, g+, g+]g ⊂ g+
[g−, g−, g−]g ⊂ g−
[g+, g+, g−]g = 0
[g−, g−, g+]g = 0
[g+, g+, g+]g = 0
[g−, g−, g−]g = 0
[g+, g+, g+]g = 0
[g−, g−, g−]g = 0
[g+, g+, g−]g ⊂ g+
[g−, g−, g+]g ⊂ g−
[g+, g+, g−]g ⊂ g−
[g−, g−, g+]g ⊂ g+
It is surprised that a strong abelian product structure is also an O-operator on a 3-Lie algebra
associated to the adjoint representation. Since an O-operator on a 3-Lie algebra associated to
the adjoint representation is also a Rota-Baxter operator [10, 34], it turns out that involutive
Rota-Baxter operator can also serve as an integrability condition. This is totally different from the
case of Lie algebras. Furthermore, by the definition of a perfect product structure, an involutive
automorphism of a 3-Lie algebra can also serve as an integrability condition. This is also a new
phenomenon. Note that the decomposition that a perfect product structure gives is exactly the
condition required in the definition of a matched pair of 3-Lie algebras [11]. Thus, this kind of
product structure will be frequently used in our studies.
An almost complex structure on a 3-Lie algebra g is defined to be a linear map J : g −→ g
satisfying J 2 = −Id. With the above motivation, we define a complex structure on a 3-Lie algebra
g to be an almost complex structure satisfying the Nijenhuis condition. Then gi and g−i, which are
eigenspaces of eigenvalues ±i of a complex linear map JC (the complexification of J) are subalgebras
of the 3-Lie algebra gC, the complexification of g. Parallel to the case of product structures, there
are also four types special integrability conditions, and each of them gives rise to a special decompo-
complex
structure
J[x, y, z]g = [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
+J([J x, J y, z]g + [J x, y, J z]g + [x, J y, J z]g)
strict complex
structure
−[J x, J y, J z]g
J[x, y, z]g = [J x, y, z]g
abelian complex
[x, y, z]g = [x, J y, J z]g + [J x, y, J z]g + [J x, J y, z]g
sition of gC:
structure
gC = gi ⊕ g−i
[gi, gi, gi]gC ⊂ gi
[g−i, g−i, g−i]gC ⊂ g−i
[gi, gi, g−i]gC = 0
[g−i, g−i, gi]gC = 0
[gi, gi, gi]gC = 0
[g−i, g−i, g−i]gC = 0
[x, y, z]g = −J([J x, y, z]g + [x, J y, z]g + [x, y, J z]g)
[gi, gi, gi]gC = 0
strong abelian
complex structure
perfect complex
structure
O-operators
Rota-Baxter operators
J[x, y, z]g = −[J x, J y, J z]g
anti-involutive automorphisms
3
[g−i, g−i, g−i]gC = 0
[gi, gi, g−i]gC ⊂ gi
[g−i, g−i, gi]gC ⊂ g−i
[gi, gi, g−i]gC ⊂ g−i
[g−i, g−i, gi]gC ⊂ gi
Then we add a compatibility condition between a complex structure and a product structure
on a 3-Lie algebra to define a complex product structure on a 3-Lie algebra. We give an equivalent
characterization of a complex product structure on a 3-Lie algebra g using the decomposition of g.
We add a compatibility condition between a symplectic structure and a paracomplex structure on a
3-Lie algebra to define a paraKähler structure on a 3-Lie algebra. An equivalent characterization of
a paraKähler structure on a 3-Lie algebra g is also given using the decomposition of g. Associated
to a paraKähler structure on a 3-Lie algebra, there is also a pseudo-Riemannian structure. We
introduce the notion of a Livi-Civita product associated to a pseudo-Riemannian 3-Lie algebra, and
give its precise formulas. Finally, we add a compatibility condition between a symplectic structure
and a complex structure on a 3-Lie algebra to define a pseudo-Kähler structure on a 3-Lie algebra.
The relation between a paraKähler structure and a pseudo-Kähler structure on a 3-Lie algebra
is investigated. We construct complex product structures, paraKähler structures and pseudo-
Kähler structures in terms of 3-pre-Lie algebras. We also give examples of symplectic structures,
product structures, complex structures, complex product structure, paraKähler structures and
pseudo-Kähler structures on the 4-dimensional Euclidean 3-Lie algebra A4 given in [6].
The paper is organized as follows. In Section 2, we recall Nijenhuis operators on 3-Lie algebras
and 3-pre-Lie algebras. In Section 3, we study representations of 3-pre-Lie algebras. In Section 4,
we introduce the notion of a phase space of a 3-Lie algebra and show that a 3-Lie algebra has a
phase space if and only if it is sub-adjacent to a 3-pre-Lie algebra. We also introduce the notion of
a Manin triple of 3-pre-Lie algebras and study its relation with phase spaces of 3-Lie algebras. In
Section 5, we introduce the notion of a product structure on a 3-Lie algebra and give four special
integrability conditions. In Section 6, we introduce the notion of a complex structure on a 3-Lie
algebra and give four special integrability conditions. In Section 7, we introduce the notion of a
complex product structure on a 3-Lie algebra and give its equivalent characterization. In Section
8, we introduce the notion of a paraKähler structure on a 3-Lie algebra and give its equivalent
characterization. Moreover, we give a detailed study on the associated Levi-Civita product. In
Section 9, we introduce the notion of a pseudo-Kähler structure on a 3-Lie algebra and study the
relation with a paraKähler structure.
In this paper, we work over the real field R and the complex field C and all the vector spaces
are finite-dimensional.
Acknowledgement: We give our warmest thanks to Chengming Bai for very useful comments
and discussions. This research is supported by NSFC (11471139) and NSF of Jilin Province
(20170101050JC).
2 Preliminaries
In this section, first we recall the notion of a Nijenhuis operator on a 3-Lie algebra, which will
be frequently used as the integrability condition in our later studies. Then we recall the notion
of a 3-pre-Lie algebra, which is the main tool to construct examples of symplectic, product and
complex structures on 3-Lie algebras.
Definition 2.1. A 3-Lie algebra is a vector space g together with a trilinear skew-symmetric
bracket [·, ·, ·]g : ∧3g −→ g such that the following fundamental identity holds:
[x, y, [z, w, v]g]g = [[x, y, z]g, w, v]g + [z, [x, y, w]g, v]g + [z, w, [x, y, v]g]g,
∀x, y, z, w, v ∈ g.
(1)
For x, y ∈ g, define ad : ∧2g −→ gl(g) by
adx,yz = [x, y, z]g,
∀z ∈ g.
4
Then (1) is equivalent to that adx,y is a derivation, i.e.
adx,y[z, w, v]g = [adx,yz, w, v]g + [z, adx,yw, v]g + [z, w, adx,yv]g,
∀x, y ∈ g.
Let (g, [·, ·, ·]g) be a 3-Lie algebra, and N : g −→ g a linear map. Define a 3-ary bracket
[·, ·, ·]1
N : ∧3g −→ g by
[x, y, z]1
N = [N x, y, z]g + [x, N y, z]g + [x, y, N z]g − N [x, y, z]g.
Then we define 3-ary bracket [·, ·, ·]2
N : ∧3g −→ g by
[x, y, z]2
N = [N x, N y, z]g + [x, N y, N z]g + [N x, y, N z]g − N [x, y, z]1
N .
(2)
(3)
Definition 2.2. ([31]) Let (g, [·, ·, ·]g) be a 3-Lie algebra. A linear map N : g −→ g is called a
Nijenhuis operator if the following Nijenhuis condition is satisfied:
[N x, N y, N z]g = N [x, y, z]2
N ,
∀x, y, z ∈ g.
(4)
More precisely, a linear map N : g −→ g of a 3-Lie algebra (g, [·, ·, ·]g) is a Nijenhuis operator
if and only if
[N x, N y, N z]g = N [N x, N y, z]g + N [x, N y, N z]g + N [N x, y, N z]g
−N 2[N x, y, z]g − N 2[x, N y, z]g − N 2[x, y, N z]g
+N 3[x, y, z]g.
(5)
Definition 2.3. ([28]) A representation of a 3-Lie algebra (g, [·, ·, ·]g) on a vector space V is a
linear map ρ : ∧2g −→ gl(V ), such that for all x1, x2, x3, x4 ∈ g, there holds:
ρ([x1, x2, x3]g, x4) + ρ(x3, [x1, x2, x4]g) = [ρ(x1, x2), ρ(x3, x4)];
ρ([x1, x2, x3]g, x4) = ρ(x1, x2) ◦ ρ(x3, x4) + ρ(x2, x3) ◦ ρ(x1, x4) + ρ(x3, x1) ◦ ρ(x2, x4).
Example 2.4. Let (g, [·, ·, ·]g) be a 3-Lie algebra. The linear map ad : ∧2g −→ gl(g) defines a
representation of the 3-Lie algebra g on itself, which we call the adjoint representation of g.
Let A be a vector space. For a linear map φ : A ⊗ A → gl(V ), we define a linear map
φ∗ : A ⊗ A → gl(V ∗) by
hφ∗(x, y)α, vi = −hα, φ(x, y)vi, ∀α ∈ V ∗, x, y ∈ g, v ∈ V.
Lemma 2.5. ([11]) Let (V, ρ) be a representation of a 3-Lie algebra (g, [·, ·, ·]g). Then (V ∗, ρ∗) is
a representation of the 3-Lie algebra (g, [·, ·, ·]g), which is called the dual representation.
Lemma 2.6. Let g be a 3-Lie algebra, V a vector space and ρ : ∧2g → gl(V ) a skew-symmetric
linear map. Then (V ; ρ) is a representation of g if and only if there is a 3-Lie algebra structure
(called the semidirect product) on the direct sum of vector spaces g ⊕ V , defined by
[x1 + v1, x2 + v2, x3 + v3]ρ = [x1, x2, x3]g + ρ(x1, x2)v3 + ρ(x2, x3)v1 + ρ(x3, x1)v2,
(6)
for all xi ∈ g, vi ∈ V, 1 ≤ i ≤ 3. We denote this semidirect product 3-Lie algebra by g ⋉ρ V.
5
Definition 2.7. Let A be a vector space with a linear map {·, ·, ·} : ⊗3A → A. The pair (A, {·, ·, ·})
is called a 3-pre-Lie algebra if the following identities hold:
{x, y, z} = −{y, x, z}
{x1, x2, {x3, x4, x5}} = {[x1, x2, x3]C , x4, x5} + {x3, [x1, x2, x4]C , x5}
{[x1, x2, x3]C , x4, x5} = {x1, x2, {x3, x4, x5}} + {x2, x3, {x1, x4, x5}}
+{x3, x4, {x1, x2, x5}}
+{x3, x1, {x2, x4, x5}},
where x, y, z, xi ∈ A, 1 ≤ i ≤ 5 and [·, ·, ·]C is defined by
[x, y, z]C , {x, y, z} + {y, z, x} + {z, x, y}, ∀x, y, z ∈ A.
(7)
(8)
(9)
(10)
Proposition 2.8. ([11, Proposition 3.21]) Let (A, {·, ·, ·}) be a 3-pre-Lie algebra. Then (A, [·, ·, ·]C )
is a 3-Lie algebra, which is called the sub-adjacent 3-Lie algebra of A, and denoted by Ac. (A, {·, ·, ·})
is called the compatible 3-pre-Lie algebra structure on the 3-Lie algebra Ac.
Define the left multiplication L : ∧2A −→ gl(A) by L(x, y)z = {x, y, z} for all x, y, z ∈ A. Then
(A, L) is a representation of the 3-Lie algebra Ac. Moreover, we define the right multiplication
R : ⊗2A → gl(A) by R(x, y)z = {z, x, y}. If there is a 3-pre-Lie algebra structure on its dual
space A∗, we denote the left multiplication and right multiplication by L and R respectively.
Definition 2.9. ([11, Definition 3.16]) Let (g, [·, ·, ·]g) be a 3-Lie algebra and (V, ρ) a representation.
A linear operator T : V → g is called an O-operator associated to (V, ρ) if T satisfies:
[T u, T v, T w]g = T (ρ(T u, T v)w + ρ(T v, T w)u + ρ(T w, T u)v), ∀u, v, w ∈ V.
(11)
Proposition 2.10. ([11, Proposition 3.27]) Let (g, [·, ·, ·]g) be a 3-Lie algebra. Then there is
a compatible 3-pre-Lie algebra if and only if there exists an invertible O-operator T : V → g
associated to a representation (V, ρ). Furthermore, the compatible 3-pre-Lie structure on g is given
by
{x, y, z} = T ρ(x, y)T −1(z), ∀x, y, z ∈ g.
(12)
3 Representations of 3-pre-Lie algebras
In this section, we introduce the notion of a representation of a 3-pre-Lie algebra, construct the
corresponding semidirect product 3-pre-Lie algebra and give the dual representation.
Definition 3.1. A representation of a 3-pre-Lie algebra (A, {·, ·, ·}) on a vector space V consists
of a pair (ρ, µ), where ρ : ∧2A → gl(V ) is a representation of the 3-Lie algebra Ac on V and
6
µ : ⊗2A → gl(V ) is a linear map such that for all x1, x2, x3, x4 ∈ A, the following equalities hold:
ρ(x1, x2)µ(x3, x4) = µ(x3, x4)ρ(x1, x2) − µ(x3, x4)µ(x2, x1)
µ([x1, x2, x3]C , x4) = ρ(x1, x2)µ(x3, x4) + ρ(x2, x3)µ(x1, x4) + ρ(x3, x1)µ(x2, x4),
µ(x1, {x2, x3, x4}) = µ(x3, x4)µ(x1, x2) + µ(x3, x4)ρ(x1, x2)
+µ(x3, x4)µ(x1, x2) + µ([x1, x2, x3]C , x4) + µ(x3, {x1, x2, x4}), (13)
(14)
−µ(x3, x4)µ(x2, x1) − µ(x2, x4)µ(x1, x3)
−µ(x2, x4)ρ(x1, x3) + µ(x2, x4)µ(x3, x1) + ρ(x2, x3)µ(x1, x4),
(15)
µ(x3, x4)ρ(x1, x2) = µ(x3, x4)µ(x2, x1) − µ(x3, x4)µ(x1, x2)
+ρ(x1, x2)µ(x3, x4) − µ(x2, {x1, x3, x4}) + µ(x1, {x2, x3, x4}).
(16)
Let (A, {·, ·, ·}) be a 3-pre-Lie algebra and ρ a representation of the sub-adjacent 3-Lie alge-
bra (Ac, [·, ·, ·]C ) on the vector space V . Then (ρ, 0) is a representation of the 3-pre-Lie algebra
(A, {·, ·, ·}) on the vector space V . It is obvious that (L, R) is a representation of a 3-pre-Lie algebra
on itself, which is called the regular representation.
Let (V, ρ, µ) be a representation of a 3-pre-Lie algebra (A, {·, ·, ·}). Define a trilinear bracket
operation {·, ·, ·}ρ,µ : ⊗3(A ⊕ V ) → A ⊕ V by
{x1 + v1, x2 + v2, x3 + v3}ρ,µ , {x1, x2, x3} + ρ(x1, x2)v3 + µ(x2, x3)v1 − µ(x1, x3)v2.
(17)
By straightforward computations, we have
Theorem 3.2. With the above notation, (A ⊕ V, {·, ·, ·}ρ,µ) is a 3-pre-Lie algebra.
This 3-pre-Lie algebra is called the semidirect product of the 3-pre-Lie algebra (A, {·, ·, ·})
and (V, ρ, µ), and denoted by A ⋉ρ,µ V .
Let V be a vector space. Define the switching operator τ : ⊗2V −→ ⊗2V by
τ (T ) = x2 ⊗ x1,
∀T = x1 ⊗ x2 ∈ ⊗2V.
Proposition 3.3. Let (ρ, µ) be a representation of a 3-pre-Lie algebra (A, {·, ·, ·}) on a vector
space V . Then ρ − µτ + µ is a representation of the sub-adjacent 3-Lie algebra (Ac, [·, ·, ·]C ) on the
vector space V .
Proof. By Theorem 3.2, we have the semidirect product 3-pre-Lie algebra A ⋉ρ,µ V . Consider its
sub-adjacent 3-Lie algebra structure [·, ·, ·]C , we have
[x1 + v1, x2 + v2, x3 + v3]C = {x1 + v1, x2 + v2, x3 + v3}ρ,µ + {x2 + v2, x3 + v3, x1 + v1}ρ,µ
{x3 + v3, x1 + v1, x2 + v2}ρ,µ
= {x1, x2, x3} + ρ(x1, x2)v3 + µ(x2, x3)v1 − µ(x1, x3)v2
+{x2, x3, x1} + ρ(x2, x3)v1 + µ(x3, x1)v2 − µ(x2, x1)v3
+{x3, x1, x2} + ρ(x3, x1)v2 + µ(x1, x2)v3 − µ(x3, x2)v1
= [x1, x2, x3]C + ((ρ − µτ + µ)(x1, x2))v3
+((ρ − µτ + µ)(x2, x3))v1 + ((ρ − µτ + µ)(x3, x1))v2.
(18)
By Lemma 2.6, ρ − µτ + µ is a representation of the sub-adjacent 3-Lie algebra (Ac, [·, ·, ·]C ) on
the vector space V . The proof is finished.
If (ρ, µ) = (L, R) is the regular representation of a 3-pre-Lie algebra (A, {·, ·, ·}), then ρ − µτ +
µ = ad is the adjoint representation of the sub-adjacent 3-Lie algebra (Ac, [·, ·, ·]C ).
7
Corollary 3.4. Let (ρ, µ) be a representation of a 3-pre-Lie algebra (A, {·, ·, ·}) on a vector space
V . Then the semidirect product 3-pre-Lie algebras A ⋉ρ,µ V and A ⋉ρ−µτ +µ,0 V given by the
representations (ρ, µ) and (ρ − µτ + µ, 0) respectively have the same sub-adjacent 3-Lie algebra
Ac ⋉ρ−µτ +µ V given by (18), which is the semidirect product of the 3-Lie algebra (Ac, [·, ·, ·]C ) and
its representation (V, ρ − µτ + µ).
Proposition 3.5. Let (ρ, µ) be a representation of a 3-pre-Lie algebra (A, {·, ·, ·}) on a vector
space V . Then (ρ∗ − µ∗τ + µ∗, −µ∗) is a representation of the 3-pre-Lie algebra (A, {·, ·, ·}) on the
vector space V ∗, which is called the dual representation of the representation (V, ρ, µ).
Proof. By Proposition 3.3, ρ − µτ + µ is a representation of the sub-adjacent 3-Lie algebra
(Ac, [·, ·, ·]C ) on the vector space V . By Lemma 2.5, ρ∗ − µ∗τ + µ∗ is a representation of the
sub-adjacent 3-Lie algebra (Ac, [·, ·, ·]C ) on the dual vector space V ∗.
It is straightforward to
deduce that other conditions in Definition 3.1 also hold. We leave details to readers.
Corollary 3.6. Let (V, ρ, µ) be a representation of a 3-pre-Lie algebra (A, {·, ·, ·}). Then the
semidirect product 3-pre-Lie algebras A ⋉ρ∗,0 V ∗ and A ⋉ρ∗−µ∗τ +µ∗,−µ∗ V ∗ given by the repre-
sentations (ρ∗, 0) and (ρ∗ − µ∗τ + µ∗, −µ∗) respectively have the same sub-adjacent 3-Lie algebra
Ac ⋉ρ∗ V ∗, which is the semidirect product of the 3-Lie algebra (Ac, [·, ·, ·]C ) and its representation
(V ∗, ρ∗).
If (ρ, µ) = (L, R) is the regular representation of a 3-pre-Lie algebra (A, {·, ·, ·}), then (ρ∗ −
µ∗τ + µ∗, −µ∗) = (ad∗, −R∗) and the corresponding semidirect product 3-Lie algebra is Ac ⋉L∗ A∗,
which is the key object when we construct phase spaces of 3-Lie algebras in the next section.
4 Symplectic structures and phase spaces of 3-Lie algebras
In this section, we introduce the notion of a phase space of a 3-Lie algebra and show that a 3-
Lie algebra has a phase space if and only if it is sub-adjacent to a 3-pre-Lie algebra. Moreover,
we introduce the notion of a Manin triple of 3-pre-Lie algebras and show that there is a one-to-
one correspondence between Manin triples of 3-pre-Lie algebras and perfect phase spaces of 3-Lie
algebras.
Definition 4.1. ([11]) A symplectic structure on a 3-Lie algebra (g, [·, ·, ·]g) is a nondegenerate
skew-symmetric bilinear form ω ∈ ∧2g∗ satisfying the following equality:
ω([x, y, z]g, w) − ω([y, z, w]g, x) + ω([z, w, x]g, y) − ω([w, x, y]g, z) = 0,
∀x, y, z, w ∈ g.
(19)
Example 4.2. Consider the 4-dimensional Euclidean 3-Lie algebra A4 given in [6]. The underlying
vector space is R4. Relative to an orthogonal basis {e1, e2, e3, e4}, the 3-Lie bracket is given by
[e1, e2, e3] = e4,
[e2, e3, e4] = e1,
[e1, e3, e4] = e2,
[e1, e2, e4] = e3.
Then it is straightforward to see that any nondegenerate skew-symmetric bilinear form is a sym-
plectic structure on A4. In particular,
ω1 = e∗
ω4 = e∗
3 ∧ e∗
1 ∧ e∗
1 + e∗
2 + e∗
4 ∧ e∗
4 ∧ e∗
2, ω2 = e∗
3, ω5 = e∗
2 ∧ e∗
1 ∧ e∗
1 + e∗
2 + e∗
4 ∧ e∗
3 ∧ e∗
3, ω3 = e∗
4, ω6 = e∗
2 ∧ e∗
1 ∧ e∗
1 + e∗
3 + e∗
3 ∧ e∗
4,
2 ∧ e∗
4
are symplectic structures on A4, where {e∗
4} are the dual basis.
1, e∗
2, e∗
3, e∗
8
Proposition 4.3. ([11]) Let (g, [·, ·, ·]g, ω) be a symplectic 3-Lie algebra. Then there exists a
compatible 3-pre-Lie algebra structure {·, ·, ·} on g given by
ω({x, y, z}, w) = −ω(z, [x, y, w]g),
∀x, y, z, w ∈ g.
(20)
A quadratic 3-pre-Lie algebra is a 3-pre-Lie algebra (A, {·, ·, ·}) equipped with a nondegen-
erate skew-symmetric bilinear form ω ∈ ∧2A∗ such that the following invariant condition holds:
ω({x, y, z}, w) = −ω(z, [x, y, w]C),
∀x, y, z, w ∈ A.
(21)
Proposition 4.3 tells us that quadratic 3-pre-Lie algebras are the underlying structures of symplectic
3-Lie algebras.
Let V be a vector space and V ∗ = Hom(V, R) its dual space. Then there is a natural nonde-
generate skew-symmetric bilinear form ω on T ∗V = V ⊕ V ∗ given by:
ω(x + α, y + β) = hα, yi − hβ, xi, ∀x, y ∈ V, α, β ∈ V ∗.
(22)
Definition 4.4. Let (h, [·, ·, ·]h) be a 3-Lie algebra and h∗ its dual space.
• If there is a 3-Lie algebra structure [·, ·, ·] on the direct sum vector space T ∗h = h ⊕ h∗ such
that (h ⊕ h∗, [·, ·, ·], ω) is a symplectic 3-Lie algebra, where ω given by (22), and (h, [·, ·, ·]h)
and (h∗, [·, ·, ·]h∗ ) are 3-Lie subalgebras of (h ⊕ h∗, [·, ·, ·]), then the symplectic 3-Lie algebra
(h ⊕ h∗, [·, ·, ·], ω) is called a phase space of the 3-Lie algebra (h, [·, ·, ·]h).
• A phase space (h ⊕ h∗, [·, ·, ·], ω) is called perfect if the following conditions are satisfied:
[x, y, α] ∈ h∗,
[α, β, x] ∈ h,
∀x, y ∈ h, α, β ∈ h∗.
(23)
3-pre-Lie algebras play important role in the study of phase spaces of 3-Lie algebras.
Theorem 4.5. A 3-Lie algebra has a phase space if and only if it is sub-adjacent to a 3-pre-Lie
algebra.
Proof. Let (A, {·, ·, ·}) be a 3-pre-Lie algebra. By Proposition 2.8, the left multiplication L is a
representation of the sub-adjacent 3-Lie algebra Ac on A. By Lemma 2.5, L∗ is a representation
of the sub-adjacent 3-Lie algebra Ac on A∗. Thus, we have the semidirect product 3-Lie algebra
Ac ⋉L∗ A∗ = (Ac ⊕ A∗, [·, ·, ·]L∗ ). Then (Ac ⋉L∗ A∗, ω) is a symplectic 3-Lie algebra, which is a
phase space of the sub-adjacent 3-Lie algebra (Ac, [·, ·, ·]C ). In fact, for all x1, x2, x3, x4 ∈ A and
α1, α2, α3, α4 ∈ A∗, we have
ω([x1 + α1, x2 + α2, x3 + α3]L∗ , x4 + α4)
= ω([x1, x2, x3]C + L∗(x1, x2)α3 + L∗(x2, x3)α1 + L∗(x3, x1)α2, x4 + α4)
= hL∗(x1, x2)α3 + L∗(x2, x3)α1 + L∗(x3, x1)α2, x4i − hα4, [x1, x2, x3]Ci
= −hα3, {x1, x2, x4}i − hα1, {x2, x3, x4}i − hα2, {x3, x1, x4}i
−hα4, {x1, x2, x3}i − hα4, {x2, x3, x1}i − hα4, {x3, x1, x2}i.
9
Similarly, we have
ω([x2 + α2, x3 + α3, x4 + α4]L∗ , x1 + α1)
= −hα4, {x2, x3, x1}i − hα2, {x3, x4, x1}i − hα3, {x4, x2, x1}i
−hα1, {x2, x3, x4}i − hα1, {x3, x4, x2}i − hα1, {x4, x2, x3}i,
ω([x3 + α3, x4 + α4, x1 + α1]L∗ , x2 + α2)
= −hα1, {x3, x4, x2}i − hα3, {x4, x1, x2}i − hα4, {x1, x3, x2}i
−hα2, {x3, x4, x1}i − hα2, {x4, x1, x3}i − hα2, {x1, x3, x4}i,
ω([x4 + α4, x1 + α1, x2 + α2]L∗ , x3 + α3)
= −hα2, {x4, x1, x3}i − hα4, {x1, x2, x3}i − hα1, {x2, x4, x3}i
−hα3, {x4, x1, x2}i − hα3, {x1, x2, x4}i − hα3, {x2, x4, x1}i.
Since {x1, x2, x3} = −{x2, x1, x3}, we deduce that ω is a symplectic structure on the semidirect
product 3-Lie algebra Ac ⋉L∗ A∗. Moreover, (Ac, [·, ·, ·]C ) is a subalgebra of Ac ⋉L∗ A∗ and A∗ is
an abelian subalgebra of Ac ⋉L∗ A∗. Thus, the symplectic 3-Lie algebra (Ac ⋉L∗ A∗, ω) is a phase
space of the sub-adjacent 3-Lie algebra (Ac, [·, ·, ·]C ).
Conversely, let (T ∗h = h ⊕ h∗, [·, ·, ·], ω) be a phase space of a 3-Lie algebra (h, [·, ·, ·]h). By
Proposition 4.3, there exists a compatible 3-pre-Lie algebra structure {·, ·, ·} on T ∗h given by (20).
Since (h, [·, ·, ·]h) is a subalgebra of (h ⊕ h∗, [·, ·, ·]), we have
ω({x, y, z}, w) = −ω(z, [x, y, w]) = −ω(z, [x, y, w]h) = 0,
∀x, y, z, w ∈ h.
Thus, {x, y, z} ∈ h, which implies that (h, {·, ·, ·}h) is a subalgebra of the 3-pre-Lie algebra
Its sub-adjacent 3-Lie algebra (hc, [·, ·, ·]C ) is exactly the original 3-Lie algebra
(T ∗h, {·, ·, ·}).
(h, [·, ·, ·]h).
Corollary 4.6. Let (T ∗h = h ⊕ h∗, [·, ·, ·], ω) be a phase space of a 3-Lie algebra (h, [·, ·, ·]h) and
(h ⊕ h∗, {·, ·, ·}) the associated 3-pre-Lie algebra. Then both (h, {·, ·, ·}h) and (h∗, {·, ·, ·}h∗) are
subalgebras of the 3-pre-Lie algebra (h ⊕ h∗, {·, ·, ·}).
Corollary 4.7. If (h ⊕ h∗, [·, ·, ·], ω) is a phase space of a 3-Lie algebra (h, [·, ·, ·]h) such that the 3-
Lie algebra (h ⊕ h∗, [·, ·, ·]) is a semidirect product h ⋉ρ∗ h∗, where ρ is a representation of (h, [·, ·, ·]h)
on h and ρ∗ is its dual representation, then
{x, y, z} , ρ(x, y)z,
∀x, y, z ∈ h,
defines a 3-pre-Lie algebra structure on h.
Proof. For all x, y, z ∈ h and α ∈ h∗, we have
hα, {x, y, z}i = −ω({x, y, z}, α) = ω(z, [x, y, α]g⊕g∗ ) = ω(z, ρ∗(x, y)α) = −hρ∗(x, y)α, zi
= hα, ρ(x, y)zi.
Therefore, {x, y, z} = ρ(x, y)z.
Example 4.8. Let (A, {·, ·, ·}A) be a 3-pre-Lie algebra. Since there is a semidirect product 3-pre-
Lie algebra structure (A ⋉L∗,0 A∗, {·, ·, ·}L∗,0) on the phase space T ∗Ac = Ac ⋉L∗ A∗, one can
10
construct a new phase space T ∗Ac ⋉L∗ (T ∗Ac)∗. This process can be continued indefinitely. Hence,
there exist a series of phase spaces {A(n)}n≥2 :
A(1) = Ac, A(2) = T ∗A(1) = Ac ⋉L∗ A∗, · · · , A(n) = T ∗A(n−1), · · · .
A(n) (n ≥ 2) is called the symplectic double of A(n−1).
At the end of this section, we introduce the notion of a Manin triple of 3-pre-Lie algebras.
Definition 4.9. A Manin triple of 3-pre-Lie algebras is a triple (A; A, A′), where
• (A, {·, ·, ·}, ω) is a quadratic 3-pre-Lie algebra;
• both A and A′ are isotropic subalgebras of (A, {·, ·, ·});
• A = A ⊕ A′ as vector spaces;
• for all x, y ∈ A and α, β ∈ A′, there holds:
{x, y, α} ∈ A′,
{α, x, y} ∈ A′,
{α, β, x} ∈ A,
{x, α, β} ∈ A.
(24)
In a Manin triple of 3-pre-Lie algebras, since the skewsymmetric bilinear form ω is nondegen-
erate, A′ can be identified with A∗ via
hα, xi , ω(α, x),
∀x ∈ A, α ∈ A′.
Thus, A is isomorphic to A ⊕ A∗ naturally and the bilinear form ω is exactly given by (22). By
the invariant condition (21), we can obtain the precise form of the 3-pre-Lie structure {·, ·, ·} on
A ⊕ A∗.
Proposition 4.10. Let (A ⊕ A∗; A, A∗) be a Manin triple of 3-pre-Lie algebras, where the non-
degenerate skewsymmetric bilinear form ω on the 3-pre-Lie algebra is given by (22). Then we
have
{x, y, α} = (L∗ − R∗τ + R∗)(x, y)α,
{α, x, y} = −R∗(x, y)α,
{α, β, x} = (L∗ − R∗τ + R∗)(α, β)x,
{x, α, β} = −R∗(α, β)x.
(25)
(26)
(27)
(28)
Proof. For all x, y, z ∈ A, α ∈ A∗, we have
h{x, y, α}, zi = ω({x, y, α}, z) = −ω(α, [x, y, z]C )
= −ω(α, {x, y, z} + {y, z, x} + {z, x, y})
= −ω(α, L(x, y)z − R(y, x)z + R(x, y)z)
= −hα, L(x, y)z − R(y, x)z + R(x, y)zi
= h(L∗ − R∗τ + R∗)(x, y)α, zi,
which implies that (25) holds. We have
h{α, x, y}, zi = ω({α, x, y}, z) = −ω(y, [α, x, z]C ) = ω(y, [z, x, α]C ) = −ω({z, x, y}, α)
= hα, R(x, y)zi = −hR∗(x, y)α, zi,
which implies that (26) holds. Similarly, we can deduce that (27) and (28) hold.
11
Theorem 4.11. There is a one-to-one correspondence between Manin triples of 3-pre-Lie algebras
and perfect phase spaces of 3-Lie algebras. More precisely, if (A ⊕ A∗; A, A∗) is a Manin triple
of 3-pre-Lie algebras, then (A ⊕ A∗, [·, ·, ·]C , ω) is a symplectic 3-Lie algebra, where ω is given by
(22). Conversely, if (h ⊕ h∗, [·, ·, ·], ω) is a perfect phase space of a 3-Lie algebra (h, [·, ·, ·]h), then
(h⊕h∗; h, h∗) is a Manin triple of 3-pre-Lie algebras, where the 3-pre-Lie algebra structure on h⊕h∗
is given by (20).
Proof. Let (A⊕A∗; A, A∗) be a Manin triple of 3-pre-Lie algebras. Denote by {·, ·, ·}A and {·, ·, ·}A∗
the 3-pre-Lie algebra structure on A and A∗ respectively, and denote by [·, ·, ·]A and [·, ·, ·]A∗ the
corresponding sub-adjacent 3-Lie algebra structure on A and A∗ respectively. By Proposition 4.10,
it is straightforward to deduce that the corresponding 3-Lie algebra structure [·, ·, ·]C on A ⊕ A∗ is
given by
[x + α, y + β, z + γ]C = [x, y, z]A + L∗(α, β)z + L∗(β, γ)x + L∗(γ, α)y
+[α, β, γ]A∗ + L∗(x, y)γ + L∗(y, z)α + L∗(z, x)β.
(29)
For all x1, x2, x3, x4 ∈ A and α1, α2, α3, α4 ∈ A∗, we have
ω([x1 + α1, x2 + α2, x3 + α3]C , x4 + α4)
= ω([x1, x2, x3]A + L∗(α1, α2)x3 + L∗(α2, α3)x1 + L∗(α3, α1)x2
+[α1, α2, α3]A∗ + L∗(x1, x2)α3 + L∗(x2, x3)α1 + L∗(x3, x1)α2, x4 + α4)
= h[α1, α2, α3]A∗ + L∗(x1, x2)α3 + L∗(x2, x3)α1 + L∗(x3, x1)α2, x4i
−hα4, [x1, x2, x3]A + L∗(α1, α2)x3 + L∗(α2, α3)x1 + L∗(α3, α1)x2i
= h[α1, α2, α3]A∗ , x4i − hα3, {x1, x2, x4}Ai − hα1, {x2, x3, x4}Ai − hα2, {x3, x1, x4}Ai
−hα4, [x1, x2, x3]Ai + h{α1, α2, α4}A∗, x3i + h{α2, α3, α4}A∗, x1i + h{α3, α1, α4}A∗, x2i.
Similarly, we have
ω([x2 + α2, x3 + α3, x4 + α4], x1 + α1)
= h[α2, α3, α4]C , x1i − hα4, {x2, x3, x1}Ai − hα2, {x3, x4, x1}Ai − hα3, {x4, x2, x1}Ai
−hα1, [x2, x3, x4]C i + h{α2, α3, α1}A∗, x4i + h{α3, α4, α1}A∗, x2i + h{α4, α2, α1}A∗, x3i,
ω([x3 + α3, x4 + α4, x1 + α1], x2 + α2)
= h[α3, α4, α1]C , x2i − hα1, {x3, x4, x2}Ai − hα3, {x4, x1, x2}Ai − hα4, {x1, x3, x2}Ai
−hα2, [x3, x4, x1]C i + h{α3, α4, α2}A∗, x1i + h{α4, α1, α2}A∗, x3i + h{α1, α3, α2}A∗, x4i,
ω([x4 + α4, x1 + α1, x2 + α2], x3 + α3)
= h[α4, α1, α2]C , x3i − hα2, {x4, x1, x3}Ai − hα4, {x1, x2, x3}Ai − hα1, {x2, x4, x3}Ai
−hα3, [x4, x1, x2]C i + h{α4, α1, α3}A∗, x2i + h{α1, α2, α3}A∗, x4i + h{α2, α4, α3}A∗, x1i.
By {x1, x2, x3}A = −{x2, x1, x3}A and {α1, α2, α3}A∗ = −{α2, α1, α3}A∗ , we deduce that ω is a
symplectic structure on the 3-Lie algebra (A ⊕ A∗, [·, ·, ·]C ). Therefore, it is a phase space.
Conversely, let (h⊕h∗, [·, ·, ·], ω) be a phase space of the 3-Lie algebra (h, [·, ·, ·]h). By Proposition
4.3, there exists a 3-pre-Lie algebra structure {·, ·, ·} on h ⊕ h∗ given by (20) such that (h ⊕
h∗, {·, ·, ·}, ω) is a quadratic 3-pre-Lie algebra. By Corollary 4.6, (h, {·, ·, ·}h) and (h∗, {·, ·, ·}h∗)
are 3-pre-Lie subalgebras of (h ⊕ h∗, {·, ·, ·}). It is obvious that both h and h∗ are isotropic. Thus,
we only need to show that (24) holds. By (23), for all x1, x2 ∈ h and α1, α2 ∈ h∗, we have
ω({x1, x2, α1}, α2) = −ω(α1, [x1, x2, α2]C ) = 0,
12
which implies that {x1, x2, α1} ∈ h∗. Similarly, we can show that the other conditions in (24) also
hold. The proof is finished.
Remark 4.12. The notions of a matched pair of 3-Lie algebras and a Manin triple of 3-Lie
algebras were introduced in [11]. By (29), we obtain that (Ac, A∗c; L∗, L∗) is a matched pair of
3-Lie algebras and the phase space is exactly the double of this matched pair. However, one should
note that a Manin triple of 3-pre-Lie algebras does not give rise to a Manin triple of 3-Lie algebras.
Remark 4.13. For pre-Lie algebras, there are equivalent description between Manin triples of pre-
Lie algebras, matched pairs of pre-Lie algebras associated to the dual representations of the regular
representations and pre-Lie bialgebras [8]. Here we only study Manin triples of 3-pre-Lie algebras,
which are closely related to phase spaces of 3-Lie algebras and para-Kähler 3-Lie algebras that
studied in Section 8. We postpone the study of matched pairs of 3-pre-Lie algebras and 3-pre-Lie
bialgebras in the future.
5 Product structures on 3-Lie algebras
In this section, we introduce the notion of a product structure on a 3-Lie algebra using the Nijenhuis
condition as the integrability condition. We find four special integrability conditions, each of them
gives a special decomposition of the original 3-Lie algebra. At the end of this section, we introduce
the notion of a (perfect) paracomplex structure on a 3-Lie algebra and give examples.
Definition 5.1. Let (g, [·, ·, ·]g) be a 3-Lie algebra. An almost product structure on the 3-Lie
algebra (g, [·, ·, ·]g) is a linear endomorphism E : g → g satisfying E2 = Id (E 6= ±Id). An almost
product structure is called a product structure if the following integrability condition is satisfied:
E[x, y, z]g = [Ex, Ey, Ez]g + [Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−E[Ex, Ey, z]g − E[x, Ey, Ez]g − E[Ex, y, Ez]g.
(30)
Remark 5.2. One can understand a product structure on a 3-Lie algebra as a Nijenhuis operator
E on a 3-Lie algebra satisfying E2 = Id.
Theorem 5.3. Let (g, [·, ·, ·]g) be a 3-Lie algebra. Then (g, [·, ·, ·]g) has a product structure if and
only if g admits a decomposition:
g = g+ ⊕ g−,
(31)
where g+ and g− are subalgebras of g.
Proof. Let E be a product structure on g. By E2 = Id , we have g = g+ ⊕ g−, where g+ and g−
are the eigenspaces of g associated to the eigenvalues ±1. For all x1, x2, x3 ∈ g+, we have
E[x1, x2, x3]g = [Ex1, Ex2, Ex3]g + [Ex1, x2, x3]g + [x1, Ex2, x3]g + [x1, x2, Ex3]g
−E[Ex1, Ex2, x3]g − E[x1, Ex2, Ex3]g − E[Ex1, x2, Ex3]g
= 4[x1, x2, x3]g − 3E[x1, x2, x3]g.
Thus, we have [x1, x2, x3]g ∈ g+, which implies that g+ is a subalgebra. Similarly, we can show
that g− is a subalgebra.
13
Conversely, we define a linear endomorphism E : g → g by
E(x + α) = x − α, ∀x ∈ g+, α ∈ g−.
(32)
Obviously we have E2 = Id. Since g+ is a subalgebra of g, for all x1, x2, x3 ∈ g+, we have
[Ex1, Ex2, Ex3]g + [Ex1, x2, x3]g + [x1, Ex2, x3]g + [x1, x2, Ex3]g
−E[Ex1, Ex2, x3]g − E[x1, Ex2, Ex3]g − E[Ex1, x2, Ex3]g
= 4[x1, x2, x3]g − 3E[x1, x2, x3]g = [x1, x2, x3]g
= E[x1, x2, x3]g,
which implies that (30) holds for all x1, x2, x3 ∈ g+. Similarly, we can show that (30) holds for all
x, y, z ∈ g. Therefore, E is a product structure on g.
Lemma 5.4. Let E be an almost product structure on a 3-Lie algebra (g, [·, ·, ·]g). If E satisfies
the following equation
E[x, y, z]g = [Ex, y, z]g,
(33)
then E is a product structure on g such that [g+, g+, g−]g = 0 and [g−, g−, g+]g = 0, i.e. g is the
3-Lie algebra direct sum of g+ and g−.
Proof. By (33) and E2 = Id, we have
[Ex, Ey, Ez]g + [Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−E[Ex, Ey, z]g − E[x, Ey, Ez]g − E[Ex, y, Ez]g
= [Ex, Ey, Ez]g + E[x, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−[E2x, Ey, z]g − [Ex, Ey, Ez]g − [E2x, y, Ez]g
= E[x, y, z]g.
Thus, E is a product structure on g. For all x1, x2 ∈ g+, α1 ∈ g−, on one hand we have
E[α1, x1, x2]g = [Eα1, x1, x2]g = −[α1, x1, x2]g.
On the other hand, we have
E[α1, x1, x2]g = E[x1, x2, α1]g = [Ex1, x2, α1]g = [x1, x2, α1]g.
Thus, we obtain [g+, g+, g−]g = 0. Similarly, we have [g−, g−, g+]g = 0. The proof is finished.
Definition 5.5. (Integrability condition I) An almost product structure E on a 3-Lie algebra
(g, [·, ·, ·]g) is called a strict product structure if (33) holds.
Corollary 5.6. Let (g, [·, ·, ·]g) be a 3-Lie algebra. Then (g, [·, ·, ·]g) has a strict product structure
if and only if g admits a decomposition:
where g+ and g− are subalgebras of g such that [g+, g+, g−]g = 0 and [g−, g−, g+]g = 0.
Proof. We leave the details to readers.
g = g+ ⊕ g−,
14
Lemma 5.7. Let E be an almost product structure on a 3-Lie algebra (g, [·, ·, ·]g). If E satisfies
the following equation
[x, y, z]g = −[x, Ey, Ez]g − [Ex, y, Ez]g − [Ex, Ey, z]g,
(34)
then E is a product structure on g.
Proof. By (34) and E2 = Id, we have
[Ex, Ey, Ez]g + [Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−E[Ex, Ey, z]g − E[x, Ey, Ez]g − E[Ex, y, Ez]g
= −[Ex, E2y, E2z]g − [E2x, Ey, E2z]g − [E2x, E2y, Ez]g
+[Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g + E[x, y, z]g
= E[x, y, z]g.
Thus, E is a product structure on g.
Definition 5.8. (Integrability condition II) An almost product structure E on a 3-Lie algebra
(g, [·, ·, ·]g) is called an abelian product structure if (34) holds.
Corollary 5.9. Let (g, [·, ·, ·]g) be a 3-Lie algebra. Then (g, [·, ·, ·]g) has an abelian product structure
if and only if g admits a decomposition:
g = g+ ⊕ g−,
where g+ and g− are abelian subalgebras of g.
Proof. Let E be an abelian product structure on g. For all x1, x2, x3 ∈ g+, we have
[x1, x2, x3]g = −[Ex1, Ex2, x3]g − [x1, Ex2, Ex3]g − [Ex1, x2, Ex3]g
= −3[x1, x2, x3]g,
which implies that [x1, x2, x3]g = 0. Similarly, for all α1, α2, α3 ∈ g−, we also have [α1, α2, α3]g = 0.
Thus, both g+ and g− are abelian subalgebras.
Conversely, define a linear endomorphism E : g → g by (32). Then it is straightforward to
deduce that E is an abelian product structure on g.
Lemma 5.10. Let E be an almost product structure on a 3-Lie algebra (g, [·, ·, ·]g). If E satisfies
the following equation
[x, y, z]g = E[Ex, y, z]g + E[x, Ey, z]g + E[x, y, Ez]g,
(35)
then E is an abelian product structure on g such that [g+, g+, g−]g ⊂ g+ and [g−, g−, g+]g ⊂ g−.
Proof. By (35) and E2 = Id, we have
[Ex, Ey, Ez]g + [Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−E[Ex, Ey, z]g − E[x, Ey, Ez]g − E[Ex, y, Ez]g
= E[x, Ey, Ez]g + E[Ex, y, Ez]g + E[Ex, Ey, z]g + E[x, y, z]g
−E[Ex, Ey, z]g − E[x, Ey, Ez]g − E[Ex, y, Ez]g
= E[x, y, z]g.
15
Thus, we obtain that E is a product structure on g. For all x1, x2, x3 ∈ g+, by (35), we have
[x1, x2, x3]g = E[Ex1, x2, x3]g + E[x1, Ex2, x3]g + E[x1, x2, Ex3]g
= 3E[x1, x2, x3]g = 3[x1, x2, x3]g.
Thus, we obtain [g+, g+, g+]g = 0. Similarly, we have [g−, g−, g−]g = 0. By Corollary 5.9, E is an
abelian product structure on g. Moreover, for all x1, x2 ∈ g+, α1 ∈ g−, we have
[x1, x2, α1]g = E[Ex1, x2, α1]g + E[x1, Ex2, α1]g + E[x1, x2, Eα1]g
= E[x1, x2, α1]g,
which implies that [g+, g+, g−]g ⊂ g+. Similarly, we have [g−, g−, g+]g ⊂ g−.
Definition 5.11. (Integrability condition III) An almost product structure E on a 3-Lie
algebra (g, [·, ·, ·]g) is called a strong abelian product structure if (35) holds.
Corollary 5.12. Let (g, [·, ·, ·]g) be a 3-Lie algebra. Then (g, [·, ·, ·]g) has a strong abelian product
structure if and only if g admits a decomposition:
g = g+ ⊕ g−,
where g+ and g− are abelian subalgebras of g such that [g+, g+, g−]g ⊂ g+ and [g−, g−, g+]g ⊂ g−.
Remark 5.13. Let E be a strong abelian product structure on a 3-Lie algebra (g, [·, ·, ·]g). Then
we can define ν+ : g+ −→ Hom(∧2g−, g−) and ν− : g− −→ Hom(∧2g+, g+) by
ν+(x)(α, β) = [α, β, x]g,
ν−(α)(x, y) = [x, y, α]g,
∀x, y ∈ g+, α, β ∈ g−.
It turns out ν+ and ν− are generalized representations of abelian 3-Lie algebras g+ and g− on g−
and g+ respectively. See [30] for more details about generalized representations of 3-Lie algebras.
More surprisingly, a strong abelian product structure is an O-operator as well as a Rota-Baxter
operator [10, 34]. Thus, some O-operators and Rota-Baxter operators on 3-Lie algebras can serve
as integrability conditions.
Proposition 5.14. Let E be an almost product structure on a 3-Lie algebra (g, [·, ·, ·]g). Then
E is a strong abelian structure on g if and only if E is an O-operator associated to the adjoint
representation (g, ad). Furthermore, there exists a compatible 3-pre-Lie algebra (g, {·, ·, ·}) on the
3-Lie algebra (g, [·, ·, ·]g), here the 3-pre-Lie algebra structure on g is given by
{x, y, z} = E[x, y, Ez]g, ∀x, y, z ∈ g.
(36)
Proof. By (35), for all x, y, z ∈ g we have
[Ex, Ey, Ez]g = E[E2x, Ey, Ez]g + E[Ex, E2y, Ez]g + E[Ex, Ey, E2z]g
= E(adEx,Eyz + adEy,Ezx + adEz,Exy).
Thus, E is an O-operator associated to the adjoint representation (g, ad).
Conversely, if for all x, y, z ∈ g, we have
[Ex, Ey, Ez]g = E(adEx,Eyz + adEy,Ezx + adEz,Exy)
= E([Ex, Ey, z]g + [x, Ey, Ez]g + [Ex, y, Ez]g),
16
then [x, y, z]g = E[x, y, Ez]g + E[Ex, y, z]g + E[x, Ey, z]g by E−1 = E. Thus, E is a strong abelian
structure on g.
Furthermore, by E−1 = E and Proposition 2.10, there exists a compatible 3-pre-Lie algebra on
g given by {x, y, z} = Eadx,yE−1(z) = E[x, y, Ez]g. The proof is finished.
There is a new phenomenon that an involutive automorphism of a 3-Lie algebra also serves as
an integrability condition.
Lemma 5.15. Let E be an almost product structure on a 3-Lie algebra (g, [·, ·, ·]g). If E satisfies
the following equation
then E is a product structure on g such that
E[x, y, z]g = [Ex, Ey, Ez]g,
Proof. By (37) and E2 = Id, we have
[g+, g+, g−]g ⊂ g−,
[g−, g−, g+]g ⊂ g+.
(37)
(38)
[Ex, Ey, Ez]g + [Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−E[Ex, Ey, z]g − E[x, Ey, Ez]g − E[Ex, y, Ez]g
= E[x, y, z]g + [Ex, y, z]g + [x, Ey, z]g + [x, y, Ez]g
−[E2x, E2y, Ez]g − [Ex, E2y, E2z]g − [E2x, Ey, E2z]g
= E[x, y, z]g.
Thus, E is a product structure on g. Moreover, for all x1, x2 ∈ g+, α1 ∈ g−, we have
E[x1, x2, α1]g = [Ex1, Ex2, Eα1]g = −[x1, x2, α1]g,
which implies that [g+, g+, g−]g ⊂ g−. Similarly, we have [g−, g−, g+]g ⊂ g+.
Definition 5.16. (Integrability condition IV) An almost product structure E on a 3-Lie alge-
bra (g, [·, ·, ·]g) is called a perfect product structure if (37) holds.
Corollary 5.17. Let (g, [·, ·, ·]g) be a 3-Lie algebra. Then (g, [·, ·, ·]g) has a perfect product structure
if and only if g admits a decomposition:
g = g+ ⊕ g−,
where g+ and g− are subalgebras of g such that [g+, g+, g−]g ⊂ g− and [g−, g−, g+]g ⊂ g+.
Proof. We leave the details to readers.
Corollary 5.18. A strict product structure on a 3-Lie algebra is a perfect product structure.
Remark 5.19. Let E be a product structure on a 3-Lie algebra (g, [·, ·, ·]g). By Theorem 5.3, g+
and g− are subalgebras. However, the brackets of mixed terms are very complicated. But a perfect
product structure E on (g, [·, ·, ·]g) ensures [g+, g+, g−]g ⊂ g− and [g−, g−, g+]g ⊂ g+. Note that
this is exactly the condition required in the definition of a matched pair of 3-Lie algebras [11]. Thus,
E is a perfect product structure if and only if (g+, g−) is a matched pair of 3-Lie algebras. This
type of product structures are very important in our later studies.
17
Definition 5.20.
(i) A paracomplex structure on a 3-Lie algebra (g, [·, ·, ·]g) is a product
structure E on g such that the eigenspaces of g associated to the eigenvalues ±1 have the
same dimension, i.e. dim(g+) = dim(g−).
(i) A perfect paracomplex structure on a 3-Lie algebra (g, [·, ·, ·]g) is a perfect product struc-
ture E on g such that the eigenspaces of g associated to the eigenvalues ±1 have the same
dimension, i.e. dim(g+) = dim(g−).
Proposition 5.21. Let (A, {·, ·, ·}) be a 3-pre-Lie algebra. Then, on the semidirect product 3-Lie
algebra Ac ⋉L∗ A∗, there is a perfect paracomplex structure E : Ac ⋉L∗ A∗ → Ac ⋉L∗ A∗ given by
E(x + α) = x − α, ∀x ∈ Ac, α ∈ A∗.
(39)
Proof. It is obvious that E2 = Id. Moreover, we have (Ac ⋉L∗ A∗)+ = A, (Ac ⋉L∗ A∗)− = A∗ and
they are two subalgebras of the semidirect product 3-Lie algebra Ac ⋉L∗ A∗. By Theorem 5.3, E is
a product structure on Ac ⋉L∗ A∗. Since A and A∗ have the same dimension, E is a paracomplex
structure on Ac ⋉L∗ A∗. It is obvious that E is perfect.
At the end of this section, we give some examples of product structures.
Example 5.22. There is a unique non-trivial 3-dimensional 3-Lie algebra. It has a basis {e1, e2, e3}
with respect to which the non-zero product is given by
[e1, e2, e3] = e1.
1
0
0
Then E =
−1 0
1
0
0
0
E =
0
0
1
0
1
0
0 −1 0
0
1
and E =
0
0
1
0
0 −1
is a perfect product structure.
0
are strong abelian product structures and
Example 5.23. Consider the 4-dimensional Euclidean 3-Lie algebra A4 given in Example 4.2.
Then
E1 =
E4 =
0
1 0
0 1
0
0 0 −1
0 0
−1 0
1
0
0
0
0
0
0
0
0
0 −1
, E2 =
0
0
0
0
0
1
0 −1
, E5 =
, E3 =
0
0
0
0
1
0
0
0 −1 0
0
0
1
0
0 −1
0
−1 0
1
0
0
0 −1 0
0
1
0
0
0
0
0
, E6 =
1
0
0 −1
0
0
0
0
0
0
0 −1 0
1
0
0
0
−1
0 0
0 −1 0 0
0
1 0
0 1
0
0
0
,
are perfect and abelian product structures.
6 Complex structures on 3-Lie algebras
In this section, we introduce the notion of a complex structure on a real 3-Lie algebra using the
Nijenhuis condition as the integrability condition. Parallel to the case of product structures, we
also find four special integrability conditions.
18
Definition 6.1. Let (g, [·, ·, ·]g) be a real 3-Lie algebra. An almost complex structure on g is
a linear endomorphism J : g → g satisfying J 2 = −Id. An almost complex structure is called a
complex structure if the following integrability condition is satisfied:
J[x, y, z]g = −[J x, J y, J z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
+J[J x, J y, z]g + J[x, J y, J z]g + J[J x, y, J z]g.
(40)
Remark 6.2. One can understand a complex structure on a 3-Lie algebra as a Nijenhuis operator
J on a 3-Lie algebra satisfying J 2 = −Id.
Remark 6.3. One can also use definition 6.1 to define the notion of a complex structure on a
complex 3-Lie algebra, considering J to be C-linear. However, this is not very interesting since
for a complex 3-Lie algebra, there is a one-to-one correspondence between such C-linear complex
structures and product structures (see Proposition 7.1).
Consider gC = g ⊗R C ∼= {x + iyx, y ∈ g}, the complexification of the real Lie algebra g, which
turns out to be a complex 3-Lie algebra by extending the 3-Lie bracket on g complex trilinearly,
and we denote it by (gC, [·, ·, ·]gC ). We have an equivalent description of the integrability condition
given in Definition 6.1. We denote by σ the conjugation in gC with respect to the real form g, that
is, σ(x + iy) = x − iy, x, y ∈ g. Then, σ is a complex antilinear, involutive automorphism of the
complex vector space gC.
Theorem 6.4. Let (g, [·, ·, ·]g) be a real 3-Lie algebra. Then g has a complex structure if and only
if gC admits a decomposition:
gC = q ⊕ p,
(41)
where q and p = σ(q) are complex subalgebras of gC.
Proof. We extend the complex structure J complex linearly, which is denoted by JC, i.e. JC :
gC −→ gC is defined by
JC(x + iy) = J x + iJ y,
∀x, y ∈ g.
(42)
Then JC is a complex linear endomorphism on gC satisfying J 2
C = −Id and the integrability condi-
tion (40) on gC. Denote by g±i the corresponding eigenspaces of gC associated to the eigenvalues
±i and there holds:
gC = gi ⊕ g−i.
It is straightforward to see that gi = {x − iJ xx ∈ g} and g−i = {x + iJ xx ∈ g}. Therefore, we
have g−i = σ(gi).
For all X, Y, Z ∈ gi, we have
JC[X, Y, Z]gC = −[JCX, JCY, JCZ]gC + [JCX, Y, Z]gC + [X, JCY, Z]gC + [X, Y, JCZ]gC
+JC[JCX, JCY, Z]gC + JC[X, JCY, JCZ]gC + JC[JCX, Y, JCZ]gC
= 4i[X, Y, Z]gC − 3JC[X, Y, Z]gC .
Thus, we have [X, Y, Z]gC ∈ gi, which implies that gi is a subalgebra. Similarly, we can show that
g−i is also a subalgebra.
Conversely, we define a complex linear endomorphism JC : gC → gC by
JC(X + σ(Y )) = iX − iσ(Y ), ∀X, Y ∈ q.
(43)
19
Since σ is a complex antilinear, involutive automorphism of gC, we have
J 2
C(X + σ(Y )) = JC(iX − iσ(Y )) = JC(iX + σ(iY )) = i(iX) − iσ(iY ) = −X − σ(Y ),
i.e. J 2
C = −Id. Since q is a subalgebra of gC, for all X, Y, Z ∈ q, we have
−[JCX, JCY, JCZ]gC + [JCX, Y, Z]gC + [X, JCY, Z]gC + [X, Y, JCZ]gC
+JC[JCX, JCY, Z]gC + JC[X, JCY, JCZ]gC + JC[JCX, Y, JCZ]gC
= 4i[X, Y, Z]gC − 3JC[X, Y, Z]gC = i[X, Y, Z]gC
= JC[X, Y, Z]gC,
which implies that JC satisfies (40) for all X, Y, Z ∈ q. Similarly, we can show that JC satisfies (40)
for all X , Y, Z ∈ gC. Since gC = q ⊕ p, we can write X ∈ gC as X = X + σ(Y ), for some X, Y ∈ q.
Since σ is a complex antilinear, involutive automorphism of gC, we have
(JC ◦ σ)(X + σ(Y )) = JC(Y + σ(X)) = iY − iσ(X) = σ(iX − iσ(Y )) = (σ ◦ JC)(X + σ(Y )),
which implies that JC ◦ σ = σ ◦ JC. Moreover, since σ(X ) = X is equivalent to X ∈ g, we deduce
that the set of fixed points of σ is the real vector space g. By JC ◦ σ = σ ◦ JC, there is a well-defined
J ∈ gl(g) given by
J , JCg.
Follows from that JC satisfies (40) and J 2
C = −Id on gC, J is a complex structure on g.
Lemma 6.5. Let J be an almost complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). If J satisfies
J[x, y, z]g = [J x, y, z]g, ∀x, y, z ∈ g,
(44)
then J is a complex structure on (g, [·, ·, ·]g).
Proof. By (44) and J 2 = −Id, we have
−[J x, J y, J z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
+J[J x, J y, z]g + J[x, J y, J z]g + J[J x, y, J z]g
= −[J x, J y, J z]g + J[x, y, z]g + [x, J y, z]g + [x, y, J z]g
+[J 2x, J y, z]g + [J x, J y, J z]g + [J 2x, y, J z]g
= J[x, y, z]g.
Thus, we obtain that J is a complex structure on g.
Definition 6.6. (Integrability condition I) An almost complex structure J on a real 3-Lie
algebra (g, [·, ·, ·]g) is called a strict complex structure if (44) holds.
Corollary 6.7. Let (g, [·, ·, ·]g) be a real 3-Lie algebra. Then there is a strict complex structure on
(g, [·, ·, ·]g) if and only if gC admits a decomposition:
gC = q ⊕ p,
(45)
where q and p = σ(q) are complex subalgebras of gC such that [q, q, p]gC = 0 and [p, p, q]gC = 0, i.e.
gC is a 3-Lie algebra direct sum of q and p.
20
Proof. Let J be a strict complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then, JC is a strict
complex structure on the complex 3-Lie algebra (gC, [·, ·, ·]gC). For all X, Y ∈ gi and σ(Z) ∈ g−i,
on one hand we have
JC[X, Y, σ(Z)]gC = [JCX, Y, σ(Z)]gC = i[X, Y, σ(Z)]gC .
On the other hand, we have
JC[X, Y, σ(Z)]gC = JC[σ(Z), X, Y ]gC = [JCσ(Z), X, Y ]gC = −i[σ(Z), X, Y ]gC .
Thus, we obtain [gi, gi, g−i]gC = 0. Similarly, we can show [g−i, g−i, gi]gC = 0.
Conversely, define a complex linear endomorphism JC : gC → gC by (43). Then it is straight-
forward to deduce that J 2
C = −Id. Since q is a subalgebra of gC, for all X, Y, Z ∈ q, we have
JC[X, Y, Z]gC = i[X, Y, Z]gC = [JCX, Y, Z]gC,
which implies that JC satisfies (44) for all X, Y, Z ∈ q. Similarly, we can show that JC satisfies (44)
for all X , Y, Z ∈ gC. By the proof of Theorem 6.4, we obtain that J , JCg is a strict complex
structure on the real 3-Lie algebra (g, [·, ·, ·]g). The proof is finished.
Let J be an almost complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). We can define a
complex vector space structure on the real vector space g by
(a + bi)x , ax + bJ x, ∀a, b ∈ R, x ∈ g.
(46)
Define two maps ϕ : g → gi and ψ : g → g−i as following:
ϕ(x) =
ψ(x) =
1
2
1
2
(x − iJ x),
(x + iJ x).
It is straightforward to deduce that ϕ is complex linear isomorphism and ψ = σ ◦ ϕ is a complex
antilinear isomorphism between complex vector spaces.
Let J be a strict complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then with the complex
vector space structure defined above, (g, [·, ·, ·]g) is a complex 3-Lie algebra. In fact, the fact that
the 3-Lie bracket is complex trilinear follows from
[(a + bi)x, y, z]g = [ax + bJ x, y, z]g = a[x, y, z]g + b[J x, y, z]g
= a[x, y, z]g + bJ[x, y, z]g = (a + bi)[x, y, z]g
using (44) and (46).
Let J be a complex structure on g. Define a new bracket [·, ·, ·]J : ∧3g → g by
[x, y, z]J , 1
4
([x, y, z]g − [x, J y, J z]g − [J x, y, J z]g − [J x, J y, z]g), ∀x, y, z ∈ g.
(47)
Proposition 6.8. Let J be a complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then (g, [·, ·, ·]J )
is a real 3-Lie algebra. Moreover, J is a strict complex structure on (g, [·, ·, ·]J ) and the correspond-
ing complex 3-Lie algebra (g, [·, ·, ·]J ) is isomorphic to the complex 3-Lie algebra gi.
21
Proof. One can show that (g, [·, ·, ·]J ) is a real 3-Lie algebra directly. Here we use a different
approach to prove this result. By (40), for all x, y, z ∈ g, we have
[ϕ(x), ϕ(y), ϕ(z)]gC =
=
=
1
8
1
8
[x − iJ x, y − iJ y, z − iJ z]gC
([x, y, z]g − [x, J y, J z]g − [J x, y, J z]g − [J x, J y, z]g)
−
1
8
i([x, y, J z]g + [x, J y, z]g + [J x, y, z]g − [J x, J y, J z]g)
1
8
([x, y, z]g − [x, J y, J z]g − [J x, y, J z]g − [J x, J y, z]g)
−
1
8
iJ([x, y, z]g − [x, J y, J z]g − [J x, y, J z]g − [J x, J y, z]g)
= ϕ[x, y, z]J .
(48)
Thus, we have [x, y, z]J = ϕ−1[ϕ(x), ϕ(y), ϕ(z)]gC . Since J is a complex structure, gi is a 3-Lie
subalgebra. Therefore, (g, [·, ·, ·]J ) is a real 3-Lie algebra.
By (40), for all x, y, z ∈ g, we have
J[x, y, z]J =
=
1
4
1
4
J([x, y, z]g − [x, J y, J z]g − [J x, y, J z]g − [J x, J y, z]g)
(−[J x, J y, J z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g)
= [J x, y, z]J ,
which implies that J is a strict complex structure on (g, [·, ·, ·]J ). By (48), ϕ is a complex 3-Lie
algebra isomorphism. The proof is finished.
Proposition 6.9. Let J be a complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then J is a
strict complex structure on (g, [·, ·, ·]g) if and only if [·, ·, ·]J = [·, ·, ·]g.
Proof. If J is a strict complex structure on (g, [·, ·, ·]g), by J[x, y, z]g = [J x, y, z]g, we have
[x, y, z]J =
1
4
([x, y, z]g − [x, J y, J z]g − [J x, y, J z]g − [J x, J y, z]g) = [x, y, z]g.
Conversely, if [·, ·, ·]J = [·, ·, ·]g, we have
−3[x, y, z]g = [x, J y, J z]g + [J x, y, J z]g + [J x, J y, z]g.
Then by the integrability condition of J, we obtain
4J[x, y, z]J = −[J x, J y, J z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
= 3[J x, y, z]g + [J x, y, z]g
= 4[J x, y, z]g,
which implies that J[x, y, z]g = [J x, y, z]g. The proof is finished.
Lemma 6.10. Let J be an almost complex structure on a real 3-Lie algebra (g, [·, ·, ·]g).
satisfies the following equation
If J
[x, y, z]g = [x, J y, J z]g + [J x, y, J z]g + [J x, J y, z]g,
(49)
Then, J is a complex structure on g.
22
Proof. By (49) and J 2 = −Id, we have
−[J x, J y, J z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
+J[J x, J y, z]g + J[x, J y, J z]g + J[J x, y, J z]g
= −[J x, J 2y, J 2z]g − [J 2x, J y, J 2z]g − [J 2x, J 2y, J z]g
+[J x, y, z]g + [x, J y, z]g + [x, y, J z]g + J[x, y, z]g
= J[x, y, z]g.
Thus, we obtain that J is a complex structure on g.
Definition 6.11. (Integrability condition II) An almost complex structure J on a real 3-Lie
algebra (g, [·, ·, ·]g) is called an abelian complex structure if (49) holds.
Remark 6.12. Let J be an abelian complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then
(g, [·, ·, ·]J ) is an abelian 3-Lie algebra.
Corollary 6.13. Let (g, [·, ·, ·]g) be a real 3-Lie algebra. Then g has an abelian complex structure
if and only if gC admits a decomposition:
where q and p = σ(q) are complex abelian subalgebras of gC.
gC = q ⊕ p,
Proof. Let J be an abelian complex structure on g. By Proposition 6.8, we obtain that ϕ is a
complex 3-Lie algebra isomorphism from (g, [·, ·, ·]J ) to (gi, [·, ·, ·]gC ). Since J is abelian, (g, [·, ·, ·]J )
is an abelian 3-Lie algebra. Therefore, q = gi is an abelian subalgebra of gC. Since p = g−i = σ(gi),
for all x1 + iy1, x2 + iy2, x3 + iy3 ∈ gi, we have
[σ(x1 + iy1), σ(x2 + iy2), σ(x3 + iy3)]gC
= [x1 − iy1, x2 − iy2, x3 − iy3]gC
= ([x1, x2, x3]g − [x1, y2, y3]g − [y1, x2, y3]g − [y1, y2, x3]g)
−i([x1, x2, y3]g + [x1, y2, x3]g + [y1, x2, x3]g − [y1, y2, y3]g)
= σ[x1 + iy1, x2 + iy2, x3 + iy3]gC
= 0.
Thus, p is an abelian subalgebra of gC.
Conversely, by Theorem 6.4, there is a complex structure J on g. Moreover, by Proposition 6.8,
we have a complex 3-Lie algebra isomorphism ϕ from (g, [·, ·, ·]J ) to (q, [·, ·, ·]gC). Thus, (g, [·, ·, ·]J )
is an abelian 3-Lie algebra. By the definition of [·, ·, ·]J , we obtain that J is an abelian complex
structure on g. The proof is finished.
Lemma 6.14. Let J be an almost complex structure on a real 3-Lie algebra (g, [·, ·, ·]g).
satisfies the following equation
If J
[x, y, z]g = −J[J x, y, z]g − J[x, J y, z]g − J[x, y, J z]g,
(50)
then J is a complex structure on g.
23
Proof. By (50) and J 2 = −Id, we have
−[J x, J y, J z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
+J[J x, J y, z]g + J[x, J y, J z]g + J[J x, y, J z]g
= J[J 2x, J y, J z]g + J[J x, J 2y, J z]g + J[J x, J y, J 2z]g + J[x, y, z]g
+J[J x, J y, z]g + J[x, J y, J z]g + J[J x, y, J z]g
= J[x, y, z]g.
Thus, J is a complex structure on g.
Definition 6.15. (Integrability condition III) An almost complex structure J on a real 3-Lie
algebra (g, [·, ·, ·]g) is called a strong abelian complex structure if (50) holds.
Corollary 6.16. Let (g, [·, ·, ·]g) be a real 3-Lie algebra. Then g has a strong abelian complex
structure if and only if gC admits a decomposition:
gC = q ⊕ p,
where q and p = σ(q) are abelian complex subalgebras of gC such that [q, q, p]gC ⊂ q and [p, p, q]gC ⊂
p.
Parallel to the case of strong abelian product structures on a 3-Lie algebra, strong abelian
complex structures on a 3-Lie algebra are also O-operators associated to the adjoint representation.
Proposition 6.17. Let J be an almost complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then
J is a strong abelian complex structure on a 3-Lie algebra (g, [·, ·, ·]g) if and only if −J is an O-
operator on (g, [·, ·, ·]g) associated to the adjoint representation (g, ad). Furthermore, there exists a
compatible 3-pre-Lie algebra (g, {·, ·, ·}) on the 3-Lie algebra (g, [·, ·, ·]g), here the 3-pre-Lie algebra
structure on g is given by
{x, y, z} = −J[x, y, J z]g, ∀x, y, z ∈ g.
(51)
Proof. By (50), for all x, y, z ∈ g we have
[−J x, −J y, −J z]g = J[J 2x, J y, J z]g + J[J x, J 2y, J z]g + J[J x, J y, J 2z]g
= −J(ad−Jx,−Jyz + ad−Jy,−Jzx + ad−Jz,−Jxy).
Thus, −J is an O-operator associated to the adjoint representation (g, ad).
Conversely, if for all x, y, z ∈ g, we have
[−J x, −J y, −J z]g = −J(ad−Jx,−Jyz + ad−Jy,−Jzx + ad−Jz,−Jxy)
= −J([−J x, −J y, z]g + [x, −J y, −J z]g + [−J x, y, −J z]g),
then we obtain [x, y, z]g = −J[x, y, J z]g − J[J x, y, z]g − J[x, J y, z]g by (−J)−1 = J.
Furthermore, by (−J)−1 = J and Proposition 2.10, there exists a compatible 3-pre-Lie algebra
on g given by {x, y, z} = −Jadx,y(−J −1(z)) = −J[x, y, J z]g. The proof is finished.
Lemma 6.18. Let J be an almost complex structure on a real 3-Lie algebra (g, [·, ·, ·]g).
satisfies the following equation
If J
J[x, y, z]g = −[J x, J y, J z]g,
(52)
then J is a complex structure on g.
24
Proof. By (52) and J 2 = Id, we have
−[J x, J y, J z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
+J[J x, J y, z]g + J[x, J y, J z]g + J[J x, y, J z]g
= J[x, y, z]g + [J x, y, z]g + [x, J y, z]g + [x, y, J z]g
−[J 2x, J 2y, J z]g − [J x, J 2y, J 2z]g − [J 2x, J y, J 2z]g
= J[x, y, z]g.
Thus, J is a complex structure on g.
Definition 6.19. (Integrability condition IV) An almost complex structure J on a real 3-Lie
algebra (g, [·, ·, ·]g) is called a perfect complex structure if (52) holds.
Corollary 6.20. Let (g, [·, ·, ·]g) be a real 3-Lie algebra. Then g has a perfect complex structure if
and only if gC admits a decomposition:
gC = q ⊕ p,
where q and p = σ(q) are complex subalgebras of gC such that [q, q, p]gC ⊂ p and [p, p, q]gC ⊂ q.
Corollary 6.21. Let J be a strict complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then J
is a perfect complex structure on g.
Example 6.22. Consider the 4-dimensional Euclidean 3-Lie algebra A4 given in Example 4.2.
Then
J1 =
J4 =
0
0
1
0
0 −1
0
0
1
0
0 −1
0
0
0
0
, J2 =
1 0
0
−1 0 0
0
0
0
0
0 0 −1
0 1
0
, J5 =
0
0
0
0
0 −1 0
0
1
0
0 −1
0
0
1
0
0
1
0
0
−1 0
0
0
1
0 −1 0
0
0
0
0
, J3 =
, J6 =
,
0
0
0
0
0 −1
0
0
1
0
0
1
0
0 −1 0
1 0
0
0 1
0
−1
0 0
0 −1 0 0
0
0
0
are abelian complex structures. Moreover, J1, J6 are strong abelian complex structures and
J2, J3, J4, J5 are perfect complex structures.
7 Complex product structures on 3-Lie algebras
In this section, we add a compatibility condition between a complex structure and a product
structure on a 3-Lie algebra to introduce the notion of a complex product structure. We construct
complex product structures using 3-pre-Lie algebras. First we illustrate the relation between a
complex structure and a product structure on a complex 3-Lie algebra.
Proposition 7.1. Let (g, [·, ·, ·]g) be a complex 3-Lie algebra. Then E is a product structure on g
if and only if J = iE is a complex structure on g.
25
Proof. Let E be a product structure on g. We have J 2 = i2E2 = −Id. Thus, J is an almost
complex structure on g. Since E satisfies the integrability condition (30), we have
J[x, y, z]g = iE[x, y, z]g
= −[iEx, iEy, iEz]g + [iEx, y, z]g + [x, iEy, z]g + [x, y, iEz]g
+iE[iEx, iEy, z]g + iE[x, iEy, iEz]g + iE[iEx, y, iEz]g.
Thus, J is a complex structure on the complex 3-Lie algebra g.
The converse part can be proved similarly and we omit details.
Corollary 7.2. Let J be a complex structure on a real 3-Lie algebra (g, [·, ·, ·]g). Then, −iJC is a
paracomplex structure on the complex 3-Lie algebra (gC, [·, ·, ·]gC ), where JC is defined by (42).
Proof. By Theorem 6.4, gC = gi ⊕ g−i and g−i = σ(gi), where gi and g−i are subalgebras of gC.
It is obvious that dim(gi) = dim(g−i). By Proposition 5.3, there is a paracomplex structure on gC.
On the other hand, it is obvious that JC is a complex structure on gC. By Proposition 7.1, −iJC
a product structure on the complex 3-Lie algebra (gC, [·, ·, ·]gC ). It is straightforward to see that
gi and g−i are eigenspaces of −iJC corresponding to +1 and −1. Thus, −iJC is a paracomplex
structure.
Definition 7.3. Let (g, [·, ·, ·]g) be a real 3-Lie algebra. A complex product structure on the
3-Lie algebra g is a pair {J, E} of a complex structure J and a product structure E satisfying
J ◦ E = −E ◦ J.
(53)
If E is perfect, we call {J, E} a perfect complex product structure on g.
Remark 7.4. Let {J, E} be a complex product structure on a real 3-Lie algebra (g, [·, ·, ·]g). For
all x ∈ g+, by (53), we have E(J x) = −J x, which implies that J(g+) ⊂ g−. Analogously, we
obtain J(g−) ⊂ g+. Thus, we get J(g−) = g+ and J(g+) = g−. Therefore, dim(g+) = dim(g−)
and E is a paracomplex structure on g.
Theorem 7.5. Let (g, [·, ·, ·]g) be a real 3-Lie algebra (g, [·, ·, ·]g). Then the following statements
are equivalent:
(i) g has a complex product structure;
(ii) g has a complex structure J and can be decomposed as g = g+ ⊕ g−, where g+, g− are 3-Lie
subalgebras of g and g− = J g+.
Proof. Let {J, E} be a complex product structure and let g± denote the eigenspaces corresponding
to the eigenvalues ±1 of E. By Theorem 5.3, both g+ and g− are 3-Lie subalgebras of g and
J ◦ E = −E ◦ J implies g− = J g+.
Conversely, we can define a linear map E : g → g by
E(x + α) = x − α, ∀x ∈ g+, α ∈ g−.
By Theorem 5.3, E is a product structure on g. By g− = J g+ and J 2 = −Id, we have
E(J(x + α)) = E(J(x) + J(α)) = −J(x) + J(α) = −J(E(x + α)).
Thus, {J, E} is a complex product structure on g. The proof is finished.
26
Example 7.6. Consider the product structures and the complex structures on the 4-dimensional
Euclidean 3-Lie algebra A4 given in Example 5.23 and Example 6.22 respectively. Then {Ji, Ei}
for i = 1, 2, 3, 4, 5, 6 are complex product structures on A4.
We give a characterization of a perfect complex product structure on a 3-Lie algebra.
Proposition 7.7. Let E be a perfect paracomplex structure on a real 3-Lie algebra (g, [·, ·, ·]g).
Then there is a perfect complex product structure {J, E} on g if and only if there exists a linear
isomorphism φ : g+ → g− satisfying the following equation
φ[x, y, z]g = −[φ(x), φ(y), φ(z)]g + [φ(x), y, z]g + [x, φ(y), z]g + [x, y, φ(z)]g
+φ[φ(x), φ(y), z]g + φ[x, φ(y), φ(z)]g + φ[φ(x), y, φ(z)]g,
∀x, y, z ∈ g+. (54)
Proof. Let {J, E} be a perfect complex product structure on g. Define a linear isomorphism
φ : g+ → g− by φ , Jg+ : g+ → g−. By the compatibility condition (40) that the complex
structure J satisfies and the coherence condition (38) that a perfect product structure E satisfies,
we deduce that (54) holds.
Conversely, we define an endomorphism J of g by
J(x + α) = −φ−1(α) + φ(x), ∀x ∈ g+, α ∈ g−.
(55)
It is obvious that J is an almost complex structure on g and J ◦ E = −E ◦ J. For all α, β, γ ∈ g−,
let x, y, z ∈ g+ such that φ(x) = α, φ(y) = β and φ(z) = γ. By (54) and (38), we have
−[J α, J β, J γ]g + [J α, β, γ]g + [α, J β, γ]g + [α, β, J γ]g
+J[J α, J β, γ]g + J[α, J β, J γ]g + J[J α, β, J γ]g
= [x, y, z]g − [x, φ(y), φ(z)]g − [φ(x), y, φ(z)]g − [φ(x), φ(y), z]g
−φ−1[x, y, φ(z)]g − φ−1[φ(x), y, z]g − φ−1[x, φ(y), z]g
= −φ−1[φ(x), φ(y), φ(z)]g
= J[α, β, γ]g,
which implies that (40) holds for all α, β, γ ∈ g−. Similarly, we can deduce that (40) holds for all
the other cases. Thus, J is a complex structure and {J, E} is a perfect complex product structure
on the 3-Lie algebra g.
At the end of this section, we construct perfect complex product structure using 3-pre-Lie
algebras.
A nondegenerate symmetric bilinear form B ∈ A∗ ⊗ A∗ on a real 3-pre-Lie algebra (A, {·, ·, ·})
is called invariant if
B({x, y, z}, w) = −B(z, {x, y, w}), ∀x, y, z, w ∈ A.
Then B induces a linear isomorphism B♯ : A → A∗ by
hB♯(x), yi = B(x, y), ∀x, y ∈ A.
(56)
(57)
Proposition 7.8. Let (A, {·, ·, ·}) be a real 3-pre-Lie algebra with a nondegenerate symmetric
bilinear from B. Then there is a perfect complex product structure {J, E} on the semidirect product
3-Lie algebra Ac ⋉L∗ A∗, where E is given by (39) and the complex structure J is given as follows:
J(x + α) = −B♯−1
(α) + B♯(x), ∀x ∈ A, α ∈ A∗.
(58)
27
Proof. By Proposition 5.21, E is a perfect product structure on Ac ⋉L∗ A∗. For all x, y, z ∈ Ac,
we have
−[B♯(x), B♯(y), B♯(z)]L∗ + [B♯(x), y, z]L∗ + [x, B♯(y), z]L∗ + [x, y, B♯(z)]L∗
+B♯[B♯(x), B♯(y), z]L∗ + B♯[x, B♯(y), B♯(z)]L∗ + B♯[B♯(x), y, B♯(z)]L∗
= [B♯(x), y, z]L∗ + [x, B♯(y), z]L∗ + [x, y, B♯(z)]L∗
= L∗(x, y)B♯(z) + L∗(y, z)B♯(x) + L∗(z, x)B♯(y).
By (56), we have
hB♯[x, y, z]C , wi = hB♯{x, y, z}, wi + hB♯{y, z, x}, wi + hB♯{z, x, y}, wi
= B({x, y, z}, w) + B({y, z, x}, w) + B({z, x, y}, w)
= −B(z, {x, y, w}) − B(x, {y, z, w}) − B(y, {z, x, w})
= −hB♯(z), {x, y, w}i − hB♯(x), {y, z, w}i − hB♯(y), {z, x, w}i
= hL∗(x, y)B♯(z), wi + hL∗(y, z)B♯(x), wi + hL∗(z, x)B♯(y), wi,
which implies that
Thus, we have
B♯[x, y, z]C = L∗(x, y)B♯(z) + L∗(y, z)B♯(x) + L∗(z, x)B♯(y).
B♯[x, y, z]C = −[B♯(x), B♯(y), B♯(z)]L∗ + [B♯(x), y, z]L∗ + [x, B♯(y), z]L∗ + [x, y, B♯(z)]L∗
+B♯[B♯(x), B♯(y), z]L∗ + B♯[x, B♯(y), B♯(z)]L∗ + B♯[B♯(x), y, B♯(z)]L∗.
By Proposition 7.7, we obtain that {J, E} is a perfect complex product structure on Ac ⋉L∗ A∗.
Let (A, {·, ·, ·}) be a real 3-pre-Lie algebra. On the real 3-Lie algebra aff(A) = Ac ⋉L A, we
consider two endomorphisms J and E given by
J(x, y) = (−y, x), E(x, y) = (x, −y), ∀x, y ∈ A.
(59)
Proposition 7.9. With the above notations, {J, E} is a perfect complex product structure on the
3-Lie algebra aff(A).
Proof. It is obvious that E is a perfect product structure on aff(A). Moreover, we have J 2 = −Id
and J ◦ E = −E ◦ J. Obviously aff(A)+ = {(x, 0)x ∈ A}, aff(A)− = {(0, y)y ∈ A}. Define
φ : aff(A)+ → aff(A)− by φ , Jaff(A)+ : aff(A)+ → aff(A)−. More precisely, φ(x, 0) = (0, x).
Then for all (x, 0), (y, 0), (z, 0) ∈ aff(A)+, we have
−[φ(x, 0), φ(y, 0), φ(z, 0)]L + [φ(x, 0), (y, 0), (z, 0)]L + [(x, 0), φ(y, 0), (z, 0)]L
+[(x, 0), (y, 0), φ(z, 0)]L + φ[φ(x, 0), φ(y, 0), (z, 0)]L + φ[(x, 0), φ(y, 0), φ(z, 0)]L
+φ[φ(x, 0), (y, 0), φ(z, 0)]L
= [φ(x, 0), (y, 0), (z, 0)]L + [(x, 0), φ(y, 0), (z, 0)]L + [(x, 0), (y, 0), φ(z, 0)]L
= (0, {y, z, x}) + (0, {z, x, y}) + (0, {x, y, z})
= φ[(x, 0), (y, 0), (z, 0)]L.
By Proposition 7.7, {J, E} is a perfect complex product structure on the 3-Lie algebra aff(A).
28
8 Para-Kähler structures on 3-Lie algebras
In this section, we add a compatibility condition between a symplectic structure and a paracomplex
structure on a 3-Lie algebra to introduce the notion of a para-Kähler structure on a 3-Lie algebra.
A para-Kähler structure gives rise to a pseudo-Riemannian structure. We introduce the notion of a
Livi-Civita product associated to a pseudo-Riemannian 3-Lie algebra and give its precise formulas
using the decomposition of the original 3-Lie algebra.
Definition 8.1. Let ω be a symplectic structure and E a paracomplex structure on a 3-Lie algebra
(g, [·, ·, ·]g). The triple (g, ω, E) is called a para-Kähler 3-Lie algebra if the following equality
holds:
ω(Ex, Ey) = −ω(x, y),
∀x, y ∈ g.
(60)
If E is perfect, we call (g, ω, E) a perfect para-Kähler 3-Lie algebra.
Proposition 8.2. Let (A, {·, ·, ·}) be a 3-pre-Lie algebra. Then (Ac ⋉L∗ A∗, ω, E) is a perfect
para-Kähler 3-Lie algebra, where ω is given by (22) and E is defined by (39).
Proof. By Theorem 4.5, (Ac ⋉L∗ A∗, ω) is a symplectic 3-Lie algebra. By Proposition 5.21, E is a
perfect paracomplex structure on the phase space T ∗Ac. For all x1, x2 ∈ A, α1, α2 ∈ A∗, we have
ω(E(x1 + α1), E(x2 + α2)) = ω(x1 − α1, x2 − α2) = h−α1, x2i − h−α2, x1i
= −ω(x1 + α1, x2 + α2).
Therefore, (T ∗Ac = Ac ⋉L∗ A∗, ω, E) is a perfect paraKähler 3-Lie algebra.
Similar as the case of para-Kähler Lie algebras, we have the following equivalent description of
a para-Kähler 3-Lie algebra.
Theorem 8.3. Let (g, ω) be a symplectic 3-Lie algebra. Then there exists a paracomplex structure
E on the 3-Lie algebra (g, [·, ·, ·]g) such that (g, ω, E) is a para-Kähler 3-Lie algebra if and only if
there exist two isotropic 3-Lie subalgebras g+ and g− such that g = g+ ⊕ g− as the direct sum of
vector spaces.
Proof. Let (g, ω, E) be a para-Kähler 3-Lie algebra. Since E is a paracomplex structure on g, we
have g = g+ ⊕ g−, where g+ and g− are 3-Lie subalgebras of g. For all x1, x2 ∈ g+, by (60), we
have
ω(Ex1, Ex2) = ω(x1, x2) = −ω(x1, x2),
which implies that ω(g+, g+) = 0. Thus, g+ is isotropic. Similarly, g− is also isotropic.
Conversely, since g+ and g− are subalgebras, g = g+ ⊕ g− as vector spaces, there is a product
structure E on g defined by (32). Moreover, since g = g+ ⊕ g− as vector spaces and both g+ and
g− are isotropic, we obtain that dim g+=dim g−. Thus, E is a paracomplex structure on g. For
all x1, x2 ∈ g+, α1, α2 ∈ g−, since g+ and g− are isotropic, we have
ω(E(x1 + α1), E(x2 + α2)) = ω(x1 − α1, x2 − α2) = −ω(x1, α2) − ω(α1, x2)
= −ω(x1 + α1, x2 + α2).
Thus, (g, ω, E) is a para-Kähler 3-Lie algebra. The proof is finished.
29
Example 8.4. Consider the symplectic structures and the perfect paracomplex structures on the
4-dimensional Euclidean 3-Lie algebra A4 given in Example 4.2 and Example 5.23 respectively.
Then {ωi, Ei} for i = 1, 2, 3, 4, 5, 6 are perfect para-Kähler structures on A4.
Example 8.5. Let (h, [·, ·, ·]h be a 3-Lie algebra and (h ⊕ h∗, ω) its (perfect) phase space, where ω
is given by (22). Then E : h ⊕ h∗ −→ h ⊕ h∗ defined by
E(x + α) = x − α,
∀x ∈ h, α ∈ h∗,
(61)
is a (perfect) paracomplex structure and (h ⊕ h∗, ω, E) is a (perfect) para-Kähler 3-Lie algebra.
Let (g, ω, E) be a para-Kähler 3-Lie algebra. Then it is obvious that g− is isomorphic to g∗
+
via the symplectic structure ω. Moreover, it is straightforward to deduce that
Proposition 8.6. Any para-Kähler 3-Lie algebra is isomorphic to the para-Kähler 3-Lie algebra
associated to a phase space of a 3-Lie algebra.
In the sequel, we study the Levi-Civita product associated to a perfect para-Kähler 3-Lie
algebra.
Definition 8.7. A pseudo-Riemannian 3-Lie algebra is a 3-Lie algebra (g, [·, ·, ·]g) endowed
with a nondegenerate symmetric bilinear form S. The associated Levi-Civita product is the product
on g, ∇ : ⊗3g −→ g with (x, y, z) 7−→ ∇x,yz, given by the following formula:
3S(∇x,yz, w) = S([x, y, z]g, w) − 2S([x, y, w]g, z) + S([y, z, w]g, x) + S([z, x, w]g, y).
(62)
Proposition 8.8. Let (g, S) be a pseudo-Riemannian 3-Lie algebra. Then the Levi-Civita product
{·, ·, ·} satisfies the following equations:
∇x,yz = −∇y,xz,
∇x,yz + ∇y,zx + ∇z,xy = [x, y, z]g.
(63)
(64)
Proof. For all w ∈ g, it is obvious that
3S(∇y,xz, w) = S([y, x, z]g, w) − 2S([y, x, w]g, z) + S([x, z, w]g, y) + S([z, y, w]g, x)
= −3S(∇x,yz, w).
By the nondegeneracy of S, we obtain ∇x,yz = −∇y,xz.
For all x, y, z, w ∈ g, we have
3S(∇x,yz, w) = S([x, y, z]g, w) − 2S([x, y, w]g, z) + S([y, z, w]g, x) + S([z, x, w]g, y),
3S(∇y,zx, w) = S([y, z, x]g, w) − 2S([y, z, w]g, x) + S([z, x, w]g, y) + S([x, y, w]g, z),
3S(∇z,xy, w) = S([z, x, y]g, w) − 2S([z, x, w]g, y) + S([x, y, w]g, z) + S([y, z, w]g, x).
Add up the three equations, we have
3S(∇x,yz + ∇y,zx + ∇z,xy, w) = 3S([x, y, z]g, w),
which implies that ∇x,yz + ∇y,zx + ∇z,xy = [x, y, z]g. The proof is finished.
Let (g, ω, E) be a perfect para-Kähler 3-Lie algebra. Define a bilinear form S on g by
S(x, y) , ω(x, Ey), ∀x, y ∈ g.
(65)
30
Proposition 8.9. With the above notations, (g, S) is a pseudo-Riemannian 3-Lie algebra. More-
over, the associated Levi-Civita product ∇ and the perfect paracomplex structure E satisfy the
following compatibility condition:
E∇x,yz = ∇Ex,EyEz.
(66)
Proof. Since ω is skewsymmetric and ω(Ex, Ey) = −ω(x, y), we have
S(y, x) = ω(y, Ex) = −ω(Ey, E2x) = −ω(Ey, x) = ω(x, Ey) = S(x, y),
which implies that S is symmetric. Moreover, since ω is nondegenerate and E2 = Id, it is obvious
that S is nondegenerate. Thus, S is a pseudo-Riemannian metric on the 3-Lie algebra g. Moreover,
we have
3S(∇Ex,EyEz, w)
= S([Ex, Ey, Ez]g, w) − 2S([Ex, Ey, w]g, Ez) + S([Ey, Ez, w]g, Ex) + S([Ez, Ex, w]g, Ey)
= S(E[x, y, z]g, w) − 2S(E[x, y, Ew]g, Ez) + S(E[y, z, Ew]g, Ex) + S(E[z, x, Ew]g, Ey)
= −(S([x, y, z]g, Ew) − 2S([x, y, Ew]g, z) + S([y, z, Ew]g, x) + S([z, x, Ew]g, y))
= −3S(∇x,yz, Ew)
= 3S(E∇x,yz, w).
Thus, we have E∇x,yz = ∇Ex,EyEz.
The following two propositions clarifies the relationship between the Levi-Civita product and
the 3-pre-Lie multiplication on a para-Kähler 3-Lie algebra.
Proposition 8.10. Let (g, ω, E) be a para-Kähler 3-Lie algebra and ∇ the associated Levi-Civita
product. Then for all x1, x2, x3 ∈ g+ and α1, α2, α3 ∈ g−, we have
∇x1,x2x3 = {x1, x2, x3}, ∇α1,α2α3 = {α1, α2, α3}.
Proof. Since (g, ω, E) is a para-Kähler 3-Lie algebra, 3-Lie subalgebras g+ and g− are isotropic
and g = g+ ⊕ g− as vector spaces. For all x1, x2, x3, x4 ∈ g+, we have
3ω(∇x1,x2x3, x4)
= 3S(∇x1,x2x3, Ex4) = 3S(∇x1,x2x3, x4)
= S([x1, x2, x3]g, x4) − 2S([x1, x2, x4]g, x3) + S([x2, x3, x4]g, x1) + S([x3, x1, x4]g, x2)
= ω([x1, x2, x3]g, x4) − 2ω([x1, x2, x4]g, x3) + ω([x2, x3, x4]g, x1) + ω([x3, x1, x4]g, x2)
= 0.
By (g+)⊥ = g+, we obtain ∇x1,x2x3 ∈ g+. Similarly, for all α1, α2, α3 ∈ g−, ∇α1,α2 α3 ∈ g−.
Furthermore, for all x1, x2, x3 ∈ g+, and α ∈ g−, we have
3ω(∇x1,x2x3, α)
= 3S(∇x1,x2x3, Eα) = −3S(∇x1,x2x3, α)
= −S([x1, x2, x3]g, α) + 2S([x1, x2, α]g, x3) − S([x2, x3, α]g, x1) − S([x3, x1, α]g, x2)
= ω([x1, x2, x3]g, α) + 2ω([x1, x2, α]g, x3) − ω([x2, x3, α]g, x1) − ω([x3, x1, α]g, x2)
= ω([α, x1, x2]g, x3) + 2ω([x1, x2, α]g, x3)
= −3ω(x3, [x1, x2, α]g)
= 3ω({x1, x2, x3}, α).
31
Thus, ∇x1,x2x3 = {x1, x2, x3}. Similarly, we have ∇α1,α2α3 = {α1, α2, α3}. The proof is finished.
Proposition 8.11. Let (g, ω, E) be a perfect para-Kähler 3-Lie algebra and ∇ the associated Levi-
Civita product. Then for all x1, x2 ∈ g+ and α1, α2 ∈ g−, we have
∇x1,x2α1 = {x1, x2, α1} +
2
3
({x2, α1, x1} + {α1, x1, x2}),
∇α1,x1x2 = −
1
3
{α1, x1, x2} +
2
3
{x2, α1, x1},
∇α1,α2 x1 = {α1, α2, x1} +
2
3
({α2, x1, α1} + {x1, α1, α2}),
∇x1,α1α2 = −
1
3
{x1, α1, α2} +
2
3
{α2, x1, α1}.
(67)
(68)
(69)
(70)
Proof. Since (g, ω, E) is a perfect para-Kähler 3-Lie algebra, 3-Lie subalgebras g+ and g− are
isotropic and g = g+ ⊕ g− as vector spaces. Thus, we have S(g+, g+) = S(g−, g−) = 0. For all
x1, x2 ∈ g+ and α1, α2 ∈ g−, we have
3S(∇x1,x2α1, α2)
= S([x1, x2, α1]g, α2) − 2S([x1, x2, α2]g, α1) + S([x2, α1, α2]g, x1) + S([α1, x1, α2]g, x2) = 0.
Since S is nondegenerate, we have ∇x1,x2α1 ∈ g−. Moreover, For all x1, x2, x3 ∈ g+ and α1 ∈ g−,
we have
3ω(∇x1,x2α1, x3)
= 3S(∇x1,x2α1, Ex3) = 3S(∇x1,x2α1, x3)
= S([x1, x2, α1]g, x3) − 2S([x1, x2, x3]g, α1) + S([x2, α1, x3]g, x1) + S([α1, x1, x3]g, x2)
= ω([x1, x2, α1]g, Ex3) − 2ω([x1, x2, x3]g, Eα1) + ω([x2, α1, x3]g, Ex1) + ω([α1, x1, x3]g, Ex2)
= ω([x1, x2, α1]g, x3) + 2ω([x1, x2, x3]g, α1) + ω([x2, α1, x3]g, x1) + ω([α1, x1, x3]g, x2)
= ω([x1, x2, α1]g, x3) + 2ω({x1, x2, α1}, x3) + ω({x2, α1, x1}, x3) + ω({α1, x1, x2}, x3).
Thus, we obtain
∇x1,x2α1 = {x1, x2, α1} +
2
3
({x2, α1, x1} + {α1, x1, x2}),
which implies that (67) holds.
For all x1, x2 ∈ g+ and α1, α2 ∈ g−, we have
3S(∇α1,x1x2, α2)
= S([α1, x1, x2]g, α2) − 2S([α1, x1, α2]g, x2) + S([x1, x2, α2]g, α1) + S([x2, α1, α2]g, x1) = 0.
Since S is nondegenerate, we have ∇α1,x1x2 ∈ g−. Moreover, For all x1, x2, x3 ∈ g+ and α1 ∈ g−,
we have
3ω(∇α1,x1x2, x3)
= 3S(∇α1,x1x2, Ex3) = 3S(∇α1,x1x2, x3)
= S([α1, x1, x2]g, x3) − 2S([α1, x1, x3]g, x2) + S([x1, x2, x3]g, α1) + S([x2, α1, x3]g, x1)
= ω([α1, x1, x2]g, x3) − 2ω([α1, x1, x3]g, x2) − ω([x1, x2, x3]g, α1) + ω([x2, α1, x3]g, x1)
= ω([α1, x1, x2]g, x3) − 2ω({α1, x1, x2}, x3) − ω({x1, x2, α1}, x3) + ω({x2, α1, x1}, x3).
32
Thus, we obtain
∇α1,x1x2 = −
1
3
{α1, x1, x2} +
2
3
{x2, α1, x1},
which implies that (68) holds.
(69) and (70) can be proved similarly. We omit details. The proof is finished.
Under the isomorphism given in Proposition 8.6 and the correspondence given in Theorem 4.11,
using the formulas provided in Proposition 4.10, we get
Corollary 8.12. For the perfect para-Kähler 3-Lie algebra (h ⊕ h∗, ω, E) given in Example 8.5,
for all x1, x2 ∈ h and α1, α2 ∈ h∗, we have
∇x1,x2α1 = (L∗(x1, x2) −
∇α1,x1x2 = (
1
3
∇α1,α2 x1 = (L∗(α1, α2) −
1
3
R∗(x1, x2) +
1
3
R∗(α1, α2) +
R∗(x2, x1) +
1
3
R∗(x1, x2))α1,
R∗(x2, x1))α1,
2
3
R∗(α2, α1) +
1
3
R∗(α1, α2))x1,
2
3
R∗(α2, α1))x1.
∇x1,α1α2 = (
1
3
(71)
(72)
(73)
(74)
9 Pseudo-Kähler structures on 3-Lie algebras
In this section, we add a compatibility condition between a symplectic structure and a complex
structure on a 3-Lie algebra to introduce the notion of a pseudo-Kähler structure on a 3-Lie algebra.
The relation between para-Kähler structures and pseudo-Kähler structures on a 3-Lie algebra is
investigated.
Definition 9.1. Let ω be a symplectic structure and J a complex structure on a real 3-Lie algebra
(g, [·, ·, ·]g). The triple (g, ω, J) is called a real pseudo-Kähler 3-Lie algebra if
ω(J x, J y) = ω(x, y),
∀x, y ∈ g.
(75)
Example 9.2. Consider the symplectic structures and the complex structures on the 4-dimensional
Euclidean 3-Lie algebra A4 given in Example 4.2 and Example 6.22 respectively. Then {ωi, Ji} for
i = 1, 2, 3, 4, 5, 6 are pseudo-Kähler structures on A4.
Proposition 9.3. Let (g, ω, J) be a real pseudo-Kähler 3-Lie algebra. Define a bilinear form S on
g by
S(x, y) , ω(x, J y), ∀x, y ∈ g.
(76)
Then (g, S) is a pseudo-Riemannian 3-Lie algebra.
Proof. By (75), we have
S(y, x) = ω(y, J x) = ω(J y, J 2x) = −ω(J y, x) = ω(x, J y) = S(x, y),
which implies that S is symmetric. Moreover, since ω is nondegenerate and J 2 = −Id, it is obvious
that S is nondegenerate. Thus, S is a pseudo-Riemannian metric on the 3-Lie algebra g.
33
Definition 9.4. Let (g, ω, J) be a real pseudo-Kähler 3-Lie algebra.
Riemannian metric is positive definite, we call (g, ω, J) a real Kähler 3-Lie algebra.
If the associated pseudo-
Theorem 9.5. Let (g, ω, E) be a complex para-Kähler 3-Lie algebra. Then (gR, ωR, J) is a real
pseudo-Kähler 3-Lie algebra, where gR is the underlying real 3-Lie algebra, J = iE and ωR = Re(ω)
is the real part of ω.
Proof. By Proposition 7.1, J = iE is a complex structure on the complex 3-Lie algebra g. Thus,
J is also a complex structure on the real 3-Lie algebra gR. It is obvious that ωR is skew-symmetric.
If for all x ∈ g, ωR(x, y) = 0. Then we have
ω(x, y) = ωR(x, y) + iωR(−ix, y) = 0.
By the nondegeneracy of ω, we obtain y = 0. Thus, ωR is nondegenerate. Therefore, ωR is a
symplectic structure on the real 3-Lie algebra gR. By ω(Ex, Ey) = −ω(x, y), we have
ωR(J x, J y) = Re(ω(iEx, iEy)) = Re(−ω(Ex, Ey)) = Re(ω(x, y)) = ωR(x, y).
Thus, (gR, iE, ωR) is a real pseudo-Kähler 3-Lie algebra.
Conversely, we have
Theorem 9.6. Let (g, ω, J) be a real pseudo-Kähler 3-Lie algebra. Then (gC, ωC, E) is a complex
para-Kähler 3-Lie algebra, where gC = g ⊗R C is the complexification of g, E = −iJC and ωC is the
complexification of ω, more precisely,
ωC(x1 + iy1, x2 + iy2) = ω(x1, x2) − ω(y1, y2) + iω(x1, y2) + iω(y1, x2),
∀x1, x2, y1, y2 ∈ g. (77)
Proof. By Corollary 7.2, E = −iJC is a paracomplex structure on the complex 3-Lie algebra gC.
It is obvious that ωC is skew-symmetric and nondegenerate. Moreover, since ω is a symplectic
structure on g, we deduce that ωC is a symplectic structure on gC. Finally, by ω(J x, J y) = ω(x, y),
we have
ωC(E(x1 + iy1), E(x2 + iy2)) = ωC(J y1 − iJ x1, J y2 − iJ x2)
= ω(J y1, J y2) − ω(J x1, J x2) − iω(J x1, J y2) − iω(J y1, J x2)
= ω(y1, y2) − ω(x1, x2) − iω(x1, y2) − iω(y1, x2)
= −ωC(x1 + iy1, x2 + iy2).
Therefore, (gC, ωC, −iJC) is a complex para-Kähler 3-Lie algebra.
At the end of this section, we construct a Kähler 3-Lie algebra using a 3-pre-Lie algebra with
a symmetric and positive definite invariant bilinear form.
Proposition 9.7. Let (A, {·, ·, ·}) be a real 3-pre-Lie algebra with a symmetric and positive definite
invariant bilinear form B. Then (Ac ⋉L∗ A∗, ω, −J) is a real Kähler 3-Lie algebra, where J is given
by (58) and ω is given by (22).
Proof. By Theorem 4.5 and Proposition 7.8, ω is a symplectic structure and J is a perfect complex
structure on the semidirect product 3-Lie algebra (Ac ⋉L∗ A∗, [·, ·, ·]L∗ ). Obviously, −J is also a
perfect complex structure on Ac ⋉L∗ A∗. Let {e1, · · · , en} be a basis of A such that B(ei, ej) = δij
and e∗
n be the dual basis of A∗. Then for all i, j, k, l, we have
1, · · · , e∗
ω(ei + e∗
j , ek + e∗
j ), −J(ek + e∗
l ) = δjk − δli,
l )) = ω(ej − e∗
ω(−J(ei + e∗
i , el − e∗
k) = −δil + δkj ,
34
which implies that ω(−J(x + α), −J(y + β)) = ω(x + α, y + β) for all x, y ∈ A and α, β ∈ A∗.
Therefore, (Ac ⋉L∗ A∗, ω, −J) is a pseudo-Kähler 3-Lie algebra. Finally, Let x = Pn
i=1 λiei ∈
A, α = Pn
i ∈ A∗ such that x + α 6= 0. We have
i=1 µie∗
S(x + α, x + α) = ω(x + α, −J(x + α))
= ω(cid:0)
n
X
i=1
λiei +
n
X
i=1
µie∗
i ,
n
X
i=1
µiei −
n
X
i=1
λie∗
i )(cid:1)
=
n
X
i=1
µ2
i +
n
X
i=1
λ2
i > 0.
Thus, S is positive definite. Therefore, {Ac ⋉L∗ A∗, ω, −J} is a real Kähler 3-Lie algebra.
References
[1] D. V. Alekseevsky and A. M. Perelomov, Poisson and symplectic structures on Lie algebras.
I. J. Geom. Phys. 22 (1997), no. 3, 191-211.
[2] A. Andrada, Complex product structures on 6-dimensional nilpotent Lie algebras. Forum
Math. 20 (2008), no. 2, 285–315.
[3] A. Andrada, M. L. Barberis and I. Dotti, Classification of abelian complex structures on
6-dimensional Lie algebras. J. Lond. Math. Soc. (2) 83 (2011), no. 1, 232-255.
[4] A. Andrada, M. L. Barberis and I. Dotti and G. P. Ovando, Product structures on four
dimensional solvable Lie algebras. Homology Homotopy Appl. 7 (2005), no. 1, 9–37.
[5] A. Andrada and S. Salamon, Complex product structures on Lie algebras. Forum Math. 17
(2005), no. 2, 261–295.
[6] J. Bagger and N. Lambert, Gauge symmetry and supersymmetry of multiple M2-branes
gauge theories. Phys. Rev. D 77 (2008), 065008.
[7] J. Bagger and N. Lambert, Three-algebras and N=6 Chern-Simons gauge theories. Phys.
Rev. D 79 (2009), no. 2, 025002, 8 pp.
[8] C. Bai, Left-symmetric bialgebras and an analogue of the classical Yang-Baxter equation.
Commun. Contemp. Math. 10 (2008), no. 2, 221-260.
[9] C. Bai, A further study on non-abelian phase spaces: left-symmetric algebraic approach and
related geometry. Rev. Math. Phys. 18 (2006), no. 5, 545–564.
[10] R. Bai, L. Guo, J. Li and Y. Wu, Rota-Baxter 3-Lie algebras. J. Math. Phys. 54 (2013),
063504.
[11] C. Bai, L. Guo and Y. Sheng, Bialgebras, the classical Yang-Baxter equation and Manin
triples for 3-Lie algebras. arXiv:1604.05996.
[12] I. Bajo and S. Benayadi, Abelian para-Kähler structures on Lie algebras. Differential Geom.
Appl. 29 (2011), no. 2, 160–173.
35
[13] A. Basu and J. A. Harvey, The M2-M5 brane system and a generalized Nahm's equation.
Nucl. Phys. B 713, 136 (2005), 136–150.
[14] S. Benayadi and M. Boucetta, On para-Kähler and hyper-para-Kähler Lie algebras. J. Algebra
436 (2015), 61–101.
[15] G. Calvaruso, A complete classification of four-dimensional paraKähler Lie algebras. Complex
Manifolds 2 (2015), 1–10.
[16] G. Calvaruso, Four-dimensional paraKähler Lie algebras: classification and geometry. Hous-
ton J. Math. 41 (2015), no. 3, 733–748.
[17] B. Y. Chu, Symplectic homogeneous spaces. Trans. Amer. Math. Soc. 197 (1974), 145-159.
[18] R. Cleyton, J. Lauret and Y. Poon, Weak mirror symmetry of Lie algebras. J. Symplectic
Geom. 8 (2010), no. 1, 37–55.
[19] R. Cleyton, Y. Poon and G. P. Ovando, Weak mirror symmetry of complex symplectic Lie
algebras. J. Geom. Phys. 61 (2011), no. 8, 1553-1563.
[20] J. A. de Azc´arraga and J. M. Izquierdo, n-ary algebras: a review with applications, J. Phys.
A: Math. Theor. 43 (2010), 293001.
[21] P. de Medeiros, J. Figueroa-O'Farrill and E. Méndez-Escobar, Metric Lie 3-algebras in
Bagger-Lambert theory. J. High Energy Phys. 2008, no. 8, 045, 41 pp.
[22] P. de Medeiros, J. Figueroa-O'Farrill, E. Méndez-Escobar and P. Ritter, On the Lie-algebraic
origin of metric 3-algebras. Comm. Math. Phys. 290 (2009), no. 3, 871-902.
[23] P. de Medeiros, J. Figueroa-O'Farrill, E. Méndez-Escobar and P. Ritter, Metric 3-Lie algebras
for unitary Bagger-Lambert theories. J. High Energy Phys. 2009, no. 4, 037, 44 pp.
[24] V. T. Filippov, n-Lie algebras. Sib. Mat. Zh. 26 (1985) 126–140.
[25] P. Gautheron, Some remarks concerning Nambu mechanics. Lett. Math. Phys. 37 (1996)
103–116.
[26] J. Gomis, D. Rodríguez-Gómez, M. Van Raamsdonk and H. Verlinde, Supersymmetric Yang-
Mills theory from Lorentzian three-algebras. J. High Energy Phys. no. 8, 094, 18 pp (2008).
[27] P. Ho, R. Hou and Y. Matsuo, Lie 3-algebra and multiple M2-branes. J. High Energy Phys.
no. 6, 020, 30 pp (2008).
[28] Sh. M. Kasymov, On a theory of n-Lie algebras. (Russian) Algebra i Logika 26, no. 3 (1987)
277-297.
[29] T. Li and A. Tomassini, Almost Kähler structures on four dimensional unimodular Lie alge-
bras. J. Geom. Phys. 62 (2012), no. 7, 1714–1731.
[30] J. Liu, A. Makhlouf and Y. Sheng, A new approach to representations of 3-Lie algebras and
abelian extensions. Algebr Represent Theor. (2017), doi:10.1007/s10468-017-9693-0.
[31] J. Liu, Y. Sheng, Y. Zhou and C. Bai, Nijenhuis operators on n-Lie algebras. Commun.
Theor. Phys. (Beijing) 65 (2016), no. 6, 659–670.
36
[32] Y. Nambu, Generalized Hamiltonian dynamics. Phys. Rev. D 7 (1973) 2405-2412.
[33] G. Papadopoulos, M2-branes, 3-Lie algebras and Plucker relations. J. High Energy Phys.
(2008), no. 5, 054, 9 pp.
[34] J. Pei, C. Bai and L. Guo, Splitting of operads and Rota-Baxter operators on operads. Appl.
Categ. Structures 25 (2017), no. 4, 505–538.
[35] S. M. Salamon, Complex structures on nilpotent Lie algebras. J. Pure Appl. Algebra 157
(2001), no. 2-3, 311-333.
[36] L. Takhtajan, On foundation of the generalized Nambu mechanics. Comm. Math. Phys. 160
(1994) 295–315.
[37] L. Takhtajan, A higher order analog of Chevalley-Eilenberg complex and deformation theory
of n-algebras. St. Petersburg Math. J. 6 (1995) 429–438.
Department of Mathematics, Jilin University, Changchun 130012, China
Email: [email protected], [email protected]
37
|
1301.4238 | 1 | 1301 | 2013-01-17T21:14:02 | Equalities and inequalities for Hermitian solutions and Hermitian definite solutions of the two matrix equations $AX = B$ and $AXA^* = B$ | [
"math.RA",
"math.OA"
] | This paper studies algebraic properties of Hermitian solutions and Hermitian definite solutions of the two types of matrix equation $AX = B$ and $AXA^* = B$. We first establish a variety of rank and inertia formulas for calculating the maximal and minimal ranks and inertias of Hermitian solutions and Hermitian definite solutions of the matrix equations $AX = B$ and $AXA^* = B$, and then use them to characterize many qualities and inequalities for Hermitian solutions and Hermitian definite solutions of the two matrix equations and their variations. | math.RA | math |
Equalities and inequalities for Hermitian solutions and Her-
mitian definite solutions of the two matrix equations AX = B
and AX A∗
= B
YONGGE TIAN
Abstract. This paper studies algebraic properties of Hermitian solutions and Hermitian definite
solutions of the two types of matrix equation AX = B and AXA∗ = B. We first establish a variety
of rank and inertia formulas for calculating the maximal and minimal ranks and inertias of Hermi-
tian solutions and Hermitian definite solutions of the matrix equations AX = B and AXA∗ = B,
and then use them to characterize many qualities and inequalities for Hermitian solutions and
Hermitian definite solutions of the two matrix equations and their variations.
Mathematics Subject Classifications. 15A03, 15A09, 15A24, 15B57.
Keywords. Matrix equation, Hermitian solution, Hermitian definite solution, generalized inverse,
rank, inertia, matrix equality, matrix inequality, Lowner partial ordering.
1
Introduction
Consider the following two well-known linear matrix equations
and
AX = B
AXA∗ = B,
(1.1)
(1.2)
both of which are simplest cases of various types of linear matrix equation (with symmetric pat-
terns), and are the starting point of many advanced study on complicated matrix equations. A
huge amount of results about the two equations and applications were given in the literature. In
particular, many problems on algebraic properties of solutions of the two matrix equations were ex-
plicitly characterized by using formulas for ranks and inertias of matrices. In this paper, the author
focuses on Hermitian solutions or Hermitian definite solutions of (1.1) and (1.2), and studies the
following optimization problems on the ranks and inertias of Hermitian solutions and Hermitian
definite solutions of (1.1) and (1.2):
Problem 1.1 Let A, B ∈ Cm×n be given, and assume that (1.1) has a Hermitian solution or
Hermitian definite solution X ∈ Cn×n. In this case, establish formulas for calculating the extremal
ranks and inertias of
X − P,
(1.3)
where P ∈ Cn×n is a given Hermitian matrix, and use the formulas to characterize behaviors
of these Hermitian solution and definite solution, in particular, to give necessary and sufficient
conditions for the four inequalities
X ≻ P, X < P, X ≺ P, X 4 P
(1.4)
in the Lowner partial ordering to hold, respectively.
Problem 1.2 Let A, B, C, D ∈ Cm×n be given, X, Y ∈ Cn×n be two unknown matrices, and
assume that the two linear matrix equations
AX = B, CY = D
(1.5)
have Hermitian solutions, respectively. In this case, establish formulas for calculating the extremal
ranks and inertias of the difference
X − Y
(1.6)
This work was supported partially by National Natural Science Foundation of China (Grant No. 11271384).
1
of the Hermitian solutions, and use the formulas to derive necessary and sufficient conditions for
X ≻ Y, X < Y, X ≺ Y, X 4 Y
(1.7)
to hold in the Lowner partial ordering, respectively.
Problem 1.3 Let A ∈ Cm×n and B ∈ Cm
H be given, and assume that (1.2) has a Hermitian
solution. Pre- and post-multiplying a matrix T ∈ Cp×n and its conjugate transpose T ∗ on both
sides of (1.2) yields a transformed equation as follows
Further define
T AXA∗T ∗ = T BT ∗.
S = { X ∈ Cn
T = { Y ∈ Cn
H AXA∗ = B },
H T AY A∗T ∗ = T BT ∗ }.
(1.8)
(1.9)
(1.10)
In this case, give necessary and sufficient conditions for S = T to hold, as well as necessary and
sufficient conditions for X ≻ Y, X < Y, X ≺ Y and X 4 Y to hold for X ∈ S and Y ∈ T ,
respectively.
Problem 1.4 Assume that (1.2) has a Hermitian solution, and define
S = { X ∈ Cn
T = { (X1 + X2)/2 T1AX1A∗T ∗
H AXA∗ = B} ,
(1.11)
(1.12)
H } .
1 = T1BT ∗
1 , T2AX2A∗T ∗
2 = T2BT ∗
2 , X1, X2 ∈ Cn
In this case, give necessary and sufficient conditions for S = T to hold.
Problem 1.5 Denote the sets of all least-squares solutions and least-rank Hermitian solutions of
(1.2) as
S = { X ∈ Cn
T = { Y ∈ Cn
H k B − AXA∗ kF = min } ,
H r( B − AY A∗ ) = min } .
(1.13)
(1.14)
In this case, establish necessary and sufficient conditions for X ≻ Y, X < Y, X ≺ Y and X 4 Y
to hold for X ∈ S and Y ∈ T , respectively.
Matrix equations have been a prominent concerns in matrix theory and applications. As is
known to all, two key tasks in solving a matrix equation is to give identifying condition for the
existence of a solution of the equation, and to give general solution of the equation. Once general
solution is given, the subsequent work is to describe behaviors of solutions of the matrix equation,
such as, the uniqueness of solutions; the norms of solutions; the ranks and ranges of solutions, the
definiteness of solutions, equalities and inequalities of solutions, etc. Problems 1.1 and 1.2 describe
the inequalities for solutions of (1.1), as well as relations between solutions of two linear matrix
equations.
Throughout this paper, Cm×n and Cm
H stand for the sets of all m × n complex matrices and all
m × m complex Hermitian matrices, respectively; the symbols A∗, r(A) and R(A) stand for the
transpose, conjugate transpose, rank and range (column space) of a matrix A ∈ Cm×n, respectively;
Im denotes the identity matrix of order m; [ A, B ] denotes a row block matrix consisting of A and
B. The Moore -- Penrose inverse of a matrix A ∈ Cm×n, denoted by A†, is defined to be the unique
matrix X ∈ Cn×m satisfying the following four matrix equations
(i) AXA = A,
(ii) XAX = X,
(iii) (AX)∗ = AX,
(iv) (XA)∗ = XA.
Further, let EA = Im − AA† and FA = In − A†A, which ranks are given by r(EA) = m − r(A)
and r(FA) = n − r(A). A well-known property of the Moore -- Penrose inverse is (A†)∗ = (A∗)†. In
particular, both (A†)∗ = A† and AA† = A†A hold if A is Hermitian, i.e., A = A∗. A < 0 (A ≻ 0)
means that A is Hermitian positive semi-definite (Hermitian positive definite). Two A, B ∈ Cm
H are
said to satisfy the inequality A < B (A ≻ B) in the Lowner partial ordering if A − B is Hermitian
2
positive semi-definite (Hermitian positive definite). i±(A) denotes the numbers of the positive and
negative eigenvalues of a Hermitian matrix A counted with multiplicities, respectively.
The results on ranks and inertias of matrices in Lemmas 1.6 and 1.7 below are obvious or
well-known (see also [10, 11] for their references), while the closed-form formulas for matrix ranks
and inertias in Lemmas 1.9 and 1.12 -- 1.15 were established by the present author, which we shall
use in the latter part of this paper to derive analytical solutions to Problems 1.1 -- 1.5.
Lemma 1.6 Let S, S1 and S2 be three sets consisting of (square) matrices over Cm×n, and let H
be a set consisting of Hermitian matrices over Cm
H . Then, the following hold.
(a) Under m = n, S has a nonsingular matrix if and only if maxX∈S r(X) = m.
(b) Under m = n, all X ∈ S are nonsingular if and only if minX∈S r(X) = m.
(c) 0 ∈ S if and only if minX∈S r(X) = 0.
(d) S = {0} if and only if maxX∈S r(X) = 0.
(e) H has a matrix X ≻ 0 (X ≺ 0) if and only if maxX∈H i+(X) = m (maxX∈H i−(X) = m).
(f) All X ∈ H satisfy X ≻ 0 (X ≺ 0) if and only if minX∈H i+(X) = m (minX∈H i−(X) = m ).
(g) H has a matrix X < 0 (X 4 0) if and only if minX∈H i−(X) = 0 (minX∈H i+(X) = 0 ).
(h) All X ∈ H satisfy X < 0 (X 4 0) if and only if maxX∈H i−(X) = 0 (maxX∈H i+( X) = 0 ).
(i) The following hold
S1 ∩ S2 6= ∅ ⇔
min
X1∈S1, X2∈S2
r( X1 − X2 ) = 0,
S1 ⊆ S2 ⇔ max
X1∈S1
min
X2∈S2
r( X1 − X2 ) = 0,
S1 ⊇ S2 ⇔ max
X2∈S2
min
X1∈S1
r( X1 − X2 ) = 0,
(1.15)
(1.16)
(1.17)
there exist X1 ∈ S1 and X2 ∈ S2 such that X1 ≻ X2 ⇔ max
i+( X1 − X2 ) = m,
X1∈S1, X2∈S2
(1.18)
there exist X1 ∈ S1 and X2 ∈ S2 such that X1 < X2 ⇔ min
i−( X1 − X2 ) = 0.
X1∈S1, X2∈S2
Lemma 1.7 Let A ∈ Cm
H , B ∈ Cn
H, Q ∈ Cm×n, and assume that P ∈ Cm×m is nonsingular. Then,
(1.19)
i±(P AP ∗) = i±(A),
i∓(A)
if λ > 0
if λ < 0
,
i±(λA) =(cid:26) i±(A)
i±(cid:20) A 0
i+(cid:20) 0 Q
0 B (cid:21) = i±(A) + i±(B),
0 (cid:21) = i−(cid:20) 0 Q
Q∗
Q∗
0 (cid:21) = r(Q).
Lemma 1.8 ([7]) Let A ∈ Cm×n, B ∈ Cm×k, C ∈ Cl×n and D ∈ Cl×k. Then,
r[ A, B ] = r(A) + r(EAB) = r(B) + r(EB A),
r(cid:20) A
C(cid:21) = r(A) + r(CFA) = r(C) + r(AFC ),
r(cid:20) A B
C 0 (cid:21) = r(B) + r(C) + r(EBAFC ).
3
(1.20)
(1.21)
(1.22)
(1.23)
(1.24)
(1.25)
(1.26)
Lemma 1.9 ([10]) Let A ∈ Cm
H , B ∈ Cm×n, D ∈ Cn
H, and let
M1 =(cid:20) A B
0 (cid:21), M2 =(cid:20) A B
B∗ D(cid:21).
B∗
Then, the following expansion formulas hold
i±(M1) = r(B) + i±(EBAEB),
EAB
i±(M2) = i±(A) + i±(cid:20)
0
B∗EA D − B∗A†B(cid:21), r(M2) = r(A) + r(cid:20)
0
EAB
B∗EA D − B∗A†B(cid:21).
r(M1) = 2r(B) + r(EBAEB),
(1.27)
(1.28)
(1.29)
Under the condition A < 0,
i+(M1) = r[ A, B ],
i−(M1) = r(B),
r(M1) = r[ A, B ] + r(B).
(1.30)
Under the condition R(B) ⊆ R(A),
i±(M2) = i±(A) + i±( D − B∗A†B ), r(M2) = r(A) + r( D − B∗A†B ).
(1.31)
Some general rank and inertia expansion formulas derived from (1.24) -- (1.29) are given below
C
B
BFQ
EP C
− r(Q),
A B
C 0
A B 0
C 0 P
r(cid:20) A
EP C 0 (cid:21) = r(cid:20) A B 0
C 0 P (cid:21) − r(P ),
0 Q
r(cid:20) A BFQ
0 Q 0
r(cid:20) A
i±(cid:20) A
i±(cid:20) EQAEQ EQB
0 (cid:21) = r
0 (cid:21) = r
0 (cid:21) = i±
D (cid:21) = i±
A B 0
0 P ∗
B∗
0
P
A B Q
B∗ D 0
Q∗
0
0
B∗EQ
BFP
FP B∗
(1.32)
(1.33)
(1.34)
(1.35)
(1.36)
− r(P ) − r(Q),
− r(P ),
− r(Q).
0
We shall use them to simplify ranks and inertias of block matrices involving Moore -- Penrose inverses
of matrices.
Lemma 1.10 ([2]) Let A, B ∈ Cm×n be given. Then, the following hold.
(a) Eq. (1.1) has a solution X ∈ Cn
H if and only if R(B) ⊆ R(A) and AB∗ = BA∗. In this case,
the general Hermitian solution of (1.1) can be written as
X = A†B + (A†B)∗ − A†BA†A + FAU FA,
(1.37)
where U ∈ Cn
H is arbitrary.
(b) The matrix equation in (1.1) has a solution 0 4 X ∈ Cn
H if and only if R(B) ⊆ R(A),
H of
AB∗ < 0 and R(AB∗) = R(BA∗) = R(B). In this case, the general solution 0 4 X ∈ Cn
(1.1) can be written as
X = B∗(AB∗)†B + FAU FA,
(1.38)
where 0 4 U ∈ Cn
H is arbitrary.
Lemma 1.11 Eq. (1.2) has a solution X ∈ Cn
AA†B = B. In this case, the general Hermitian solution of AXA∗ = B can be written as
H if and only if R(B) ⊆ R(A), or equivalently,
X = A†B(A†)∗ + FAU + U ∗FA,
(1.39)
where U ∈ Cn×n is arbitrary.
4
Lemma 1.12 ([11]) Let Aj ∈ Cmj ×n and Bj ∈ C
mj
H be given, j = 1, 2, and assume that
A1X1A∗
1 = B1 and A2X2A∗
2 = B2
are solvable for X1, X2 ∈ Cn
H. Also define
Sj =(cid:8) Xj ∈ Cn
H AjXjA∗
j = Bj(cid:9) ,
j = 1, 2, M =
0
A1
B1
0 −B2 A2
A∗
1 A∗
2
.
0
Then,
max
X1∈S1,X2∈S2
min
X1∈S1,X2∈S2
r( X1 − X2 ) = min { n, r(M ) + 2n − 2r(A1) − 2r(A2) } ,
r( X1 − X2 ) = r(M ) − 2r[ A∗
1, A∗
2 ],
max
X1∈S1,X2∈S2
min
X1∈S1,X2∈S2
i±( X1 − X2 ) = i±(M ) + n − r(A1) − r(A2),
i±( X1 − X2 ) = i±(M ) − r[ A∗
1, A∗
2 ].
(1.40)
(1.41)
(1.42)
(1.43)
(1.44)
(1.45)
Consequently, the following hold.
(a) There exist X1 ∈ S1 and X2 ∈ S2 such that X1 − X2 is nonsingular if and only if r(M ) >
2r(A1) + 2r(A2) − n.
(b) X1 − X2 is nonsingular for all X1 ∈ S1 and X2 ∈ S2 if and only if r(M ) = 2r[ A∗
1, A∗
2 ] + n.
(c) There exist X1 ∈ S1 and X2 ∈ S2 such that X1 = X2 if and only if R(Bj) ⊆ R(Aj) and
r(M ) = 2r[ A∗
1, A∗
2 ], j = 1, 2.
(d) The rank of X1 − X2 is invariant for all X1 ∈ S1 and X2 ∈ S2 if and only if r(M ) =
2r[ A∗
1, A∗
2 ] − n or r(A1) = r(A2) = n.
(e) There exist X1 ∈ S1 and X2 ∈ S2 such that X1 ≻ X2 (X1 ≺ X2) if and only if i+(M ) =
r(A1) + r(A2) ( i−(M ) = r(A1) + r(A2) ) .
(f) X1 ≻ X2 (X1 ≺ X2) for all X1 ∈ S1 and X2 ∈ S2 if and only if i+(M ) = r[ A∗
1, A∗
2 ] + n
( i−(M ) = r[ A∗
1, A∗
2 ] + n ) .
(g) There exist X1 ∈ S1 and X2 ∈ S2 such that X1 < X2 (X1 4 X2) if and only if i−(M ) =
r[ A∗
1, A∗
2 ] ( i+(M ) = r[ A∗
1, A∗
2 ] ) .
(h) X1 < X2 (X1 4 X2) for all X1 ∈ S1 and X2 ∈ S2 if and only if i−(M ) = r(A1) + r(A2) − n
( i+(M ) = r(A1) + r(A2) − n ) .
(i) i+( X1 − X2 ) is invariant for all X1 ∈ S1 and X2 ∈ S2 ⇔ i−( X1 − X2 ) is invariant for all
X1 ∈ S1 and X2 ∈ S2 ⇔ r(A1) = r(A2) = n.
Lemma 1.13 Let A ∈ Cm
following hold.
H and B ∈ Cm×n be given, and denote M = (cid:20) A B
0 (cid:21). Then, the
B∗
(a) [10, 21] The extremal ranks and inertias of A − BXB∗ subject to X ∈ Cn
H are given by
max
X∈Cn
H
min
X∈Cn
H
r( A − BXB∗ ) = r[ A, B ],
r( A − BXB∗ ) = 2r[ A, B ] − r(M ),
max
X∈Cn
H
min
X∈Cn
H
i±( A − BXB∗ ) = i±(M ),
i±( A − BXB∗ ) = r[ A, B ] − i∓(M ).
5
(1.46)
(1.47)
(1.48)
(1.49)
(b) [15] The extremal ranks and inertias of A ± BXB∗ subject to 0 4 X ∈ Cn
H are given by
max
r( A + BXB∗ ) = r[ A, B ], min
r( A + BXB∗ ) = i+(A) + r[ A, B ] − i+(M ), (1.50)
max
i+( A + BXB∗ ) = i+(M ), min
i+( A + BXB∗ ) = i+(A),
04X∈Cn
H
04X∈Cn
H
H
i−( A + BXB∗ ) = i−(A), min
04X∈Cn
r( A − BXB∗ ) = r[ A, B ], min
H
04X∈Cn
H
H
i+( A − BXB∗ ) = i+(A), min
04X∈Cn
i−( A − BXB∗ ) = i−(M ), min
H
04X∈Cn
H
i−( A + BXB∗ ) = r[ A, B ] − i+(M ),
r( A − BXB∗ ) = i−(A) + r[ A, B ] − i−(M ), (1.53)
i+( A − BXB∗ ) = r[ A, B ] − i−(M ),
i−( A − BXB∗ ) = i−(A).
04X∈Cn
H
04X∈Cn
H
max
04X∈Cn
max
04X∈Cn
H
max
04X∈Cn
max
04X∈Cn
H
Lemma 1.14 ([4, 10]) Let A ∈ Cm
H , B ∈ Cm×n and C ∈ Cm×k be given. Then,
max
H, Y ∈Ck
H
X∈Cn
r( A − BXB∗ − CY C ∗ ) = r[ A, B, C ],
min
H, Y ∈Ck
H
X∈Cn
r( A − BXB∗ − CY C ∗ ) = 2r[ A, B, C ] + r(cid:20) A B
0 (cid:21) − r(cid:20) A B C
0 (cid:21)
B∗
C ∗
0
0
C ∗
− r(cid:20) A B C
0 (cid:21),
i±( A − BXB∗ − CY C ∗ ) = i±
0
i±( A − BXB∗ − CY C ∗ ) = r[ A, B, C ] − i∓
A B C
B∗
0
C ∗
0
0
,
max
H, Y ∈Ck
H
X∈Cn
min
H, Y ∈Ck
H
X∈Cn
A B C
B∗
0
C ∗
0
0
.
0
Lemma 1.15 ([5, 11]) Let A ∈ Cm
R(B) ⊆ R(C ∗). Then,
H , B ∈ Cm×n and C ∈ Cp×m be given and assume that
(1.51)
(1.52)
(1.54)
(1.55)
(1.56)
(1.57)
(1.58)
(1.59)
(1.60)
(1.61)
(1.62)
(1.63)
(1.64)
(1.65)
(1.66)
(1.67)
max
X∈Cn×p
min
X∈Cn×p
0 (cid:21)(cid:27),
B∗
r[ A − BXC − (BXC)∗ ] = min(cid:26)r[ A, C ∗ ], r(cid:20) A B
r[ A − BXC − (BXC)∗ ] = 2r[ A, C ∗ ] + r(cid:20) A B
i±[ A − BXC − (BXC)∗ ] = i±(cid:20) A B
0 (cid:21),
i±[ A − BXC − (BXC)∗ ] = r[ A, C ∗ ] + i±(cid:20) A B
B∗
B∗
B∗
0 (cid:21) − 2r(cid:20) A B
C 0 (cid:21),
max
X∈Cn×p
min
X∈Cn×p
and
0 (cid:21) − r(cid:20) A B
C 0 (cid:21),
0 (cid:21)(cid:27),
max
X∈Cn×m
min
X∈Cn×m
max
X∈Cn×m
min
X∈Cn×m
B∗
r[ A − BX − (BX)∗ ] = min(cid:26)m, r(cid:20) A B
r[ A − BX − (BX)∗ ] = r(cid:20) A B
0 (cid:21) − 2r(B),
i±[ A − BX − (BX)∗ ] = i±(cid:20) A B
0 (cid:21),
i±[ A − BX − (BX)∗ ] = i±(cid:20) A B
0 (cid:21) − r(B).
B∗
B∗
B∗
2 Properties of Hermitian solutions and Hermitian definite
solutions of AX = B
Some formulas for calculating the ranks and inertias of Hermitian solutions and Hermitian definite
solutions of the matrix equation in (1.1) were established in [11]. In this section, we reconsider the
6
ranks and inertias of these solutions and give a group of complete results.
Theorem 2.1 Assume that (1.1) has a Hermitian solution, and let P ∈ Cn
H. Also, define
S = { X ∈ Cn
H AX = B }.
Then,
r( X − P ) = r( B − AP ) − r(A) + n,
r( X − P ) = 2r( B − AP ) − r( BA∗ − AP A∗ ),
max
X∈S
min
X∈S
max
X∈S
min
X∈S
i±( X − P ) = i±( BA∗ − AP A∗ ) − r(A) + n,
i±( X − P ) = r( B − AP ) − i∓( BA∗ − AP A∗ ).
(2.1)
(2.2)
(2.3)
(2.4)
(2.5)
In consequence, the following hold.
(a) There exists an X ∈ S such that X − P is nonsingular if and only if R( AP − B ) = R(A).
(b) X − P is nonsingular for all X ∈ S if and only if 2r( B − AP ) = r( BA∗ − AP A∗ ) + n.
(c) There exists an X ∈ S such that X ≻ P (X ≺ P ) if and only if
R( BA∗ − AP A∗ ) = R(A) and BA∗ < AP A∗
( R( BA∗ − AP A∗ ) = R(A) and BA∗ 4 AP A∗ ) .
(d) X ≻ P (X ≺ P ) holds for all X ∈ S if and only if
r( B − AP ) = n and BA∗ < AP A∗
( r( B − AP ) = n and AB∗ 4 AP A∗ ) .
(e) There exists an X ∈ S such that X < P (X 4 P ) if and only if
R( B − AP ) = R( BA∗ − AP A∗ ) and BA∗ < AP A∗
( R( B − AP ) = R( BA∗ − AP A∗ ) and BA∗ 4 AP A∗ ).
(f) X < P (X 4 P ) holds for all X ∈ S if and only if
BA∗ < AP A∗ and r(A) = n ( BA∗ 4 AP A∗ and r(A) = n ) .
In particular, the following hold.
(g) There exists an X ∈ S such that X is nonsingular if and only if R(B) = R(A).
(h) X is nonsingular for all X ∈ S if and only if r( B ) = n.
(i) There exists an X ∈ S such that X ≻ 0 (X ≺ 0) if and only if
R(B) = R(A) and BA∗ < 0 ( R(B) = R(A) and BA∗ 4 0 ) .
(j) X ≻ 0 (X ≺ 0) holds for all X ∈ S if and only if
r(B) = n and BA∗ < 0 ( r(B) = n and AB∗ 4 0 ) .
(k) There exists an X ∈ S such that X < 0 (X 4 0) if and only if
R(B) = R(BA∗) and BA∗ < 0 ( R(B) = R(BA∗) and BA∗ 4 0 ).
(l) X < 0 (X 4 0) holds for all X ∈ S if and only if
BA∗ < 0 and r(A) = n ( BA∗ 4 0 and r(A) = n ) .
7
Proof By Lemma 1.10(a), X − P can be written as
X − P = X0 − P + FAU FA,
(2.6)
where X0 = A†B + (A†B)∗ − A†BA†A and U ∈ Cn
gives
H is arbitrary. Applying Lemma 1.13(a) to (2.6)
max
X∈S
r( X − P ) = max
U ∈Cn
H
r( X0 − P + FAU FA) = r[ X0 − P, FA ],
min
X∈S
r( X − P ) = min
U ∈Cn
H
max
X∈S
i±( X − P ) = min
U ∈Cn
H
min
X∈S
i±( X − P ) = min
U ∈Cn
H
r( X0 − P + FAU FA) = 2r[ X0 − P, FA ] − r(cid:20) X0 − P FA
0 (cid:21),
0 (cid:21),
i±( X0 − P + FAU FA) = i±(cid:20) X0 − P FA
0 (cid:21).
i±( X0 − P + FAU FA) = r[ X0 − P, FA ] − i∓(cid:20) X0 − P FA
FA
FA
FA
(2.7)
(2.8)
(2.9)
(2.10)
Applying (1.24) and (1.28) to the block matrices in (2.7) -- (2.10) and simplifying by (1.20) and
elementary block matrix operations, we obtain
r[ X0 − P, FA ] = r( A†AX0 − A†AP ) + r(FA) = r( B − AP ) + n − r(A),
FA
i±(cid:20) X0 − P FA
r(cid:20) X0 − P FA
0 (cid:21) = r(FA) + i±[ A†A( X0 − P )A†A ] = n − r(A) + i±( BA∗ − AP A∗ ),
0 (cid:21) = 2r(FA) + r[ A†A( X0 − P )A†A ] = 2n − 2r(A) + r( BA∗ − AP A∗ ).
FA
(2.11)
(2.12)
(2.13)
Substituting (2.11) -- (2.13) into (2.7) -- (2.10) and simplifying leads to (2.2) -- (2.5). Results (a) -- (l)
follow from applying Lemma 1.6 to (2.2) -- (2.5).
(cid:3)
Theorem 2.2 Assume that (1.1) has a Hermitian solution X < 0, and let 0 4 P ∈ Cn
define
H. Also,
S = { 0 4 X ∈ Cn
H AX = B }, M =(cid:20) AB∗ B
P (cid:21).
B∗
Then,
(2.14)
(2.15)
(2.16)
(2.17)
(2.18)
(2.19)
(2.20)
max
X∈S
min
X∈S
max
X∈S
min
X∈S
max
X∈S
min
X∈S
r( X − P ) = r( B − AP ) − r(A) + n,
r( X − P ) = i−(M ) + r( B − AP ) − i+( BA∗ − AP A∗ ),
i+( X − P ) = i+( BA∗ − AP A∗ ) − r(A) + n,
i+( X − P ) = i−(M ),
i−( X − P ) = i+(M ) − r(B),
i−( X − P ) = r( B − AP ) − i+( BA∗ − AP A∗ ).
Consequently, the following hold.
(a) There exists an X ∈ S such that X − P is nonsingular if and only if R( B − AP ) = R(A).
(b) X −P is nonsingular for all X ∈ S if and only if i−(M )+r( B−AP ) = i+( BA∗−AP A∗ )+n.
(c) There exists an X ∈ S such that X ≻ P if and only if R( BA∗ − AP A∗ ) = R(A) and
BA∗ < AP A∗.
(d) X ≻ P holds for all X ∈ S if and only if i−(M ) = n.
(e) There exists an X ∈ S such that X ≺ P if and only if i+(M ) = r(B) + n.
8
(f) X ≺ P holds for all X ∈ S if and only if r( B − AP ) = n and BA∗ 4 AP A∗.
(g) There exists an X ∈ S such that X < P if and only if R( B − AP ) = R( BA∗ − AP A∗ ) and
BA∗ < AP A∗.
(h) X < P holds for all X ∈ S if and only if i+(M ) = r(B).
(i) There exists an X ∈ S such that X 4 P if and only if M < 0.
(j) X 4 P holds for all X ∈ S if and only if i+( BA∗ − AP A∗ ) = n − r(A).
In particular, the following hold.
(k) There exists an X ∈ S such that X − In is nonsingular if and only if R( B − A ) = R(A).
(l) X − In is nonsingular for all X ∈ S if and only if i−( BA∗ − BB∗ ) + r( B − A ) = i+( BA∗ −
AA∗ ) + n.
(m) There exists an X ∈ S such that X ≻ In if and only if R( BA∗ − AA∗ ) = R(A) and
BA∗ < AA∗.
(n) X ≻ In holds for all X ∈ S if and only if BA∗ ≺ BB∗.
(o) There exists an X ∈ S such that X ≺ In if and only if R( BA∗ − BB∗ ) = R(B) and
BA∗ < BB∗.
(p) X ≺ In holds for all X ∈ S if and only if r( B − A ) = n and BA∗ 4 AA∗.
(q) There exists an X ∈ S such that X < In if and only if R( B − A ) = R( BA∗ − AA∗ ) and
BA∗ < AA∗.
(r) X < In holds for all X ∈ S if and only if r(B) = n and BA∗ 4 BB∗.
(s) There exists an X ∈ S such that X 4 In if and only if BA∗ < BB∗.
(t) X 4 In holds for all X ∈ S if and only if i+( BA∗ − AA∗ ) = n − r(A).
Proof By Lemma 1.10(b), X − P can be written as
where X0 = B∗(AB∗)†B and 0 4 U ∈ Cn
H is arbitrary. Applying Lemma 1.13(b) to (2.21) gives
X − P = X0 − P + FAU FA,
(2.21)
max
X∈S
min
X∈S
r( X − P ) = max
04U ∈Cn
H
r( X − P ) = min
04U ∈Cn
H
r( X0 − P + FAU FA) = r[ X0 − P, FA ],
r( X0 − P + FAU FA) = i+( X0 − P ) + r[ X0 − P, FA ]
(2.22)
(2.23)
(2.24)
(2.25)
(2.26)
0 (cid:21),
− i+(cid:20) X0 − P FA
FA
i+( X0 − P + FAU FA) = i+(cid:20) X0 − P FA
0 (cid:21),
FA
i+( X0 − P + FAU FA) = i+( X0 − P ),
i−( X0 − P + FAU FA) = i−( X0 − P ),
i+( X − P ) = min
04U ∈Cn
H
i+( X − P ) = min
04U ∈Cn
H
i−( X − P ) = min
04U ∈Cn
H
max
X∈S
min
X∈S
max
X∈S
min
X∈S
i−( X − P ) = min
04U ∈Cn
H
i±( X0 − P + FAU FA) = r[ X0 − P, FA ] − i+(cid:20) X0 − P FA
0 (cid:21). (2.27)
FA
Applying (1.21) and (1.31) to X0 − P gives
i±( X0 − P ) = i∓[ P − B∗(AB∗)†B∗ ] = i∓(cid:20) AB∗ B∗
P (cid:21) − i∓(AB∗) = i∓(M ) − i∓(AB∗),
B
9
so that
i+( X0 − P ) = i−(M ),
i+( X0 − P ) = i+(M ) − r(B).
(2.28)
Substituting (2.11) -- (2.13) and (2.28) into (2.22) -- (2.27) and simplifying leads to (2.15) -- (2.20).
Results (a) -- (t) follow from applying Lemma 1.6 to (2.15) -- (2.20).
(cid:3)
A general problem related to Hermitian solutions and Hermitian definite solutions of AX = B
is to establish formulas for calculating the extremal ranks and inertias of P − QXQ∗ subject to
the Hermitian solutions and Hermitian definite solutions of AX = B. The results obtained can be
used to solve optimization problems of P − QXQ∗ subject to AX = B.
3 Relations between Hermitian solutions of AX = B and
C Y = D
In order to compare Hermitian solutions of matrix equations, we first establish some fundamental
formulas for calculating the extremal ranks and inertias of difference of Hermitian solutions of
the two matrix equations AX = B and CY = D, and then use them to characterize relationship
between the Hermitian solutions.
Theorem 3.1 Assume that each of the matrix equations in (1.5) has a Hermitian solution, and
let
S = { X ∈ Cn
H AX = B }, T = { Y ∈ Cn
H CY = D }.
AB∗
0
A∗
M =
0
A
−CD∗ C
C ∗
0
, N =(cid:20) A B
C D(cid:21).
r( X − Y ) = n + r(N ) − r(A) − r(C),
Also denote
Then,
max
X∈S, Y ∈T
min
X∈S, Y ∈T
max
X∈S, Y ∈T
min
X∈S, Y ∈T
(3.1)
(3.2)
(3.3)
(3.4)
(3.5)
r( X − Y ) = 2r(N ) + r( AD∗ − BC ∗ ) − r(cid:20) A BA∗
C DA∗(cid:21) − r(cid:20) A BC ∗
C DC ∗(cid:21),
i±( X − Y ) = n + i±(M ) − r(A) − r(C),
i±( X − Y ) = r(N ) − i∓(M ).
In consequence, the following hold.
(a) There exist X ∈ S and Y ∈ T such that X − Y is nonsingular if and only if r(N ) =
r(A) + r(C).
(b) X − Y is nonsingular for all X ∈ S and Y ∈ T if and only if
2r(N ) + r( AD∗ − BC ∗ ) = r(cid:20) A BA∗
C DA∗(cid:21) + r(cid:20) A BC ∗
C(cid:21)[ B∗, D∗ ] =(cid:20) B
C DC ∗(cid:21) + n.
D(cid:21) [ A∗, C ∗ ].
(c) S ∩ T 6= ∅ if and only if R(cid:20) B
D(cid:21) ⊆ R(cid:20) A
C(cid:21), (cid:20) A
(d) There exist X ∈ S and Y ∈ T such that X ≻ Y (X ≺ Y ) if and only if i+(M ) = r(A) + r(C)
( i−(M ) = r(A) + r(C) ) .
(e) X ≻ Y (X ≺ Y ) holds for all X ∈ S and Y ∈ T if and only if i−(M ) = r(N ) − n
( i+(M ) = r(N ) − n ).
(f) There exist X ∈ S and Y ∈ T such that X < Y (X 4 Y ) holds if and only if i+(M ) = r(N )
( i−(M ) = r(N ) ).
10
(g) X < Y (X 4 Y ) holds for all X ∈ S and Y ∈ T if and only if i−(M ) = r(A) + r(C) − n
( i+(M ) = r(A) + r(C) − n ) .
Proof By Lemma 1.10(a), X − Y can be written as
X − Y = X0 − Y0 + FAU FA − FC V FC ,
(3.6)
where X0 = A†B + (A†B)∗ − A†BA†A and Y0 = C †D + (C †D)∗ − C †DC †C, and U, V ∈ Cn
arbitrary. Applying Lemma 1.14 to (3.6) gives
H are
max
X∈S, Y ∈T
r( X − Y ) = max
U, V
r( X0 − Y0 + FAU FA − FC V FC ) = r[ X0 − Y0, FA, FC ],
(3.7)
min
X∈S, Y ∈T
r( X − Y ) = min
U, V
r( X0 − Y0 + FAU FA − FC V FC )
= 2r[ X0 − Y0, FA, FC ] + r(cid:20) X0 − Y0 FA
0 (cid:21),
− r(cid:20) X0 − Y0 FA FC
FC
FC
0
i±( X0 − Y0 + FAU FA − FC V FC )
i±( X − Y ) = max
U, V
0 (cid:21) − r(cid:20) X0 − Y0 FA FC
0 (cid:21)
FA
0
(3.8)
(3.9)
(3.10)
max
X∈S, Y ∈T
min
X∈S, Y ∈T
i±( X − Y ) = min
U, V
i±( X0 − Y0 + FAU FA − FC V FC )
0
0
FA
FC
= i±
X0 − Y0 FA FC
0
0
,
= r[ X0 − Y0, FA, FC ] − i∓
X0 − Y0 FA FC
0
FA
FC
0
0
.
0
Applying (1.32) -- (1.34) to the block matrices in (3.7) -- (3.10) and simplifying by Lemma 1.7, (1.20)
and elementary block matrix operations, we obtain
X0 − Y0
0
0
In
In
A 0
0 C
0
0
0
In
In
In
−B A 0
−B 0 −A
D 0
D 0 C
C
A 0
0
In D∗
0 C ∗
In
0
0 C ∗
0
In
−B A
X0 − Y0
In
0
r[ X0 − Y0, FA, FC ] = r
= r
= r
0 (cid:21) = r
= r
= r
r(cid:20) X0 − Y0 FA
FC
= n + r(N ) − r(A) − r(C),
− r(A) − r(C)
− r(A) − r(C)
− r(A) − r(C)
− r(A) − r(C)
− r(A) − r(C)
0
In
0
In
0
0
0
0 BC ∗ − AD∗
= 2n + r( BC ∗ − AD∗ ) − r(A) − r(C),
− r(A) − r(C)
11
(3.11)
(3.12)
In
0
A 0
0 C
In
0
0 A∗
0
0
− 2r(A) − r(C)
− 2r(A) − r(C)
− 2r(A) − r(C)
In
0
0 A∗
0
0
0
0
0
0
In
0
0 −A BA∗
C −DA∗
0
0
FA
In
0
In
0
0
X0 − Y0
0
In
−B A 0
D 0 C
r(cid:20) X0 − Y0 FA FC
0 (cid:21) = r
= r
= r
= 2n + r(cid:20) A BA∗
0 (cid:21) = 2n + r(cid:20) A BC ∗
r(cid:20) X0 − Y0 FA FC
i±
0
= i±
X0 − Y0 FA FC
0
In
In
0
0
0
In
0
0
X0 − Y0
FA
FC
FC
0
0
0
(3.13)
(3.14)
C DA∗(cid:21) − 2r(A) − r(C),
C DC ∗(cid:21) − r(A) − 2r(C),
In
0
In
0 A∗
0
0
0
A 0
0 C
0
0
0 C ∗
0
0
0
0
In
0
0
0
In
In
In −B∗/2 D∗/2
0
0
−B/2 A 0
D/2
0 C
0
C ∗
0
0
A∗
0
0
0
− r(A) − r(C)
− r(A) − r(C)
= i±
= n + i±
= n + i±
0
In
0
0
0
= i±
0
0
C ∗
−AD∗/2
0
− r(A) − r(C)
− r(A) − r(C)
C ∗
−AD∗/2
0
0
0
−A∗
AB∗
0
0
0
In
0
0
0 −A
0
0
C −DA∗/2
−A∗
AB∗
−A
C −DA∗/2
C ∗
0
A∗
0
A AB∗
C
0
−CD∗
= n + i±(M ) − r(A) − r(C).
− r(A) − r(C)
(3.15)
Substituting (3.11) -- (3.15) into (3.7) -- (3.10) and simplifying leads to (3.2) -- (3.5). Results (a) -- (g)
follow from applying Lemma 1.6 to (3.2) -- (3.5).
(cid:3)
A direct consequence for AX = B and its perturbation equation ( A + δA )Y = ( B + δB ) is
given below.
Corollary 3.2 Assume that both AX = B and its perturbation equation ( A + δA )Y = ( B + δB )
have Hermitian solutions, and let
S = { X ∈ Cn
H AX = B }, T = { Y ∈ Cn
H ( A + δA )Y = ( B + δB ) }.
Also denote
AB∗
0
A∗
0
A
−( A + δA )( B + δB )∗
( A + δA )
( A + δA )∗
0
M =
Then, the following hold.
, N =(cid:20) A B
δA δB(cid:21).
12
(a) There exist X ∈ S and Y ∈ T such that X − Y is nonsingular if and only if
r(cid:20) A B
δA δB(cid:21) = r(A) + r( A + δA ).
(b) X − Y is nonsingular for all X ∈ S and Y ∈ T if and only if
2r(N ) + r[ A(δB)∗ − B(δA)∗ ] = r(cid:20) A
δB(cid:21) ⊆ R(cid:20) A
(c) S ∩ T 6= ∅ if and only if R(cid:20) B
δA (δB)A∗(cid:21) + r(cid:20) A
δA(cid:21) and A(δB)∗ = B(δA)∗.
BA∗
BA∗ + B(δA)∗
δA (δB)A∗ + (δB)(δA)∗(cid:21) + n.
(d) There exist X ∈ S and Y ∈ T such that X ≻ Y (X ≺ Y ) if and only if i+(M ) = r(A) +
r( A + δA ) ( i−(M ) = r(A) + r( A + δA ) ) .
(e) X ≻ Y (X ≺ Y ) holds for all X ∈ S and Y ∈ T if and only if i−(M ) = r(N ) − n
( i+(M ) = r(N ) − n ).
(f) There exist X ∈ S and Y ∈ T such that X < Y (X 4 Y ) if and only if i+(M ) = r(N )
( i−(M ) = r(N ) ).
(g) X < Y (X 4 Y ) holds for all X ∈ S and Y ∈ T if and only if
i−(M ) = r(A) + r( A + δA ) − n ( i+(M ) = r(A) + r( A + δA ) − n ) .
4 Equalities and inequalities for Hermitian solutions of AX A∗
=
B and its transformed equations
The transformation matrix equation in (1.8) may reasonably occur in the investigation of (1.2)
and its variations. For instance,
(a) setting T = A∗ in (1.8) yields A∗AXA∗A = A∗BA, which is the well-known normal equation
of (1.2) corresponding to trace[( B − AXA∗ )( B − AXA∗ )∗ ] = min;
(b) partitioning A and B in (1.2) as A = (cid:20) A1
A2(cid:21) and B =(cid:20) B1 B2
2 B3(cid:21), and setting T = [ Im1 , 0 ]
and T = [ 0, Im2 ] in (1.8), respectively, we obtain two small equations A1XA∗
A2XA∗
1 = B1 and
B∗
2 = B3;
(c) partitioning A and X in (1.2) as A = [ A1, A2 ] and X = (cid:20) X1 X2
and T = EA2, respectively, we obtain two small equations EA1 A2X2A∗
EA2A1X1A∗
1EA2 = EA2 BEA2, respectively;
2 X3(cid:21), and setting T = EA1
2EA1 = EA1 BEA1 and
X ∗
(d) decomposing A in (1.2) as a sum A = A1+A2 and setting T = EA1 and T = EA2 , respectively,
1EA2 =
2EA1 = EA1BEA1 and EA2 A1XA∗
we obtain two transformed equations EA1 A2XA∗
EA2BEA2 , respectively.
Since solutions of (1.2) and its transformed equations are not necessarily the same, it is necessary
to consider relations between the solutions of (1.2) and its transformed equations.
Theorem 4.1 Assume that (1.2) has a Hermitian solution, and let S and T be defined as in (1.9)
and (1.10). Then, the following hold.
(a) S ⊆ T always holds, namely, all Hermitian solutions of (1.2) are solutions of (1.8).
(b) S = T if and only if r(T A) = r(A), or equivalently, R(A∗T ∗) = R(A∗).
13
If X is a Hermitian solution of (1.2), then T AXA∗T ∗ = T BT ∗ holds as well. Hence, we
Proof
have (a). Note from (1.17) that the set inclusion S ⊇ T is equivalent to the following max-min
rank problem
Then, applying (1.65) and simplifying by (1.29), we obtain
max
Y ∈T
min
X∈S
r( X − Y ) = 0.
min
X∈S
r( X − Y ) = min
V
r(cid:2) A†B(A∗)† + FAV + V ∗FA − Y (cid:3)
= r(cid:20) A†B(A†)∗ − Y FA
0 (cid:21) − 2r(FA)
FA
= r[ A†B(A†)∗ − A†AY A†A ] = r( B − AY A∗ ).
From Lemma 1.11, the general Hermitian solution of (1.8) can be written as
Y = (T A)†T BT ∗(A∗T ∗)† − FT AU − U ∗FT A,
where U is arbitrary. Applying (1.60), we obtain
r( B − AY A∗ )
max
Y ∈T
= max
U
r[ B − A(T A)†T BT ∗(A∗T ∗)†A∗ + AFT AU A∗ + AU ∗FT AA∗ ]
= min(cid:26)r[ B − A(T A)†T BT ∗(A∗T ∗)†A∗, A ], r(cid:20) B − A(T A)†T BT ∗(A∗T ∗)†A∗ AFT A
= min(cid:26)r(A), r(cid:20) B − A(T A)†T BT ∗(A∗T ∗)†A∗ AFT A
(cid:21)(cid:27).
FT AA∗
FT AA∗
0
0
Applying (1.35) and simplifying by (1.20) and (1.23), we obtain
(cid:21)
0
A∗
0
FT AA∗
B − A(T A)†T BT ∗(A∗T ∗)†A∗ A
i±(cid:20) B − A(T A)†T BT ∗(A∗T ∗)†A∗ AFT A
= i±
= i±
= i±
0 A 0
A∗
0
0
−T B + T BT ∗(A∗T ∗)†A∗
0
A∗
0
0
0
0
− r(T A) = r(A) − r(T A),
0 A∗T ∗
T A
0
0
− r(T A)
0
0
B − A(T A)†T BT ∗(A∗T ∗)†A∗ A −BT ∗ + A(T A)†T BT ∗
− r(T A)
so that
(4.1)
(4.2)
(4.3)
(cid:21)(cid:27)
(4.4)
(4.5)
(4.6)
r(cid:20) B − A(T A)†T BT ∗(A∗T ∗)†A∗ AFT A
FT AA∗
0
(cid:21) = 2r(A) − 2r(T A).
Substituting (4.6) into (4.4), and then (4.4) into (4.2) yields
max
Y ∈T
min
X∈S1
r( X − Y ) = min{r(A), 2r(A) − 2r(T A)}.
(4.7)
Finally, substituting (4.7) into (4.1) leads to the result in (b).
(cid:3)
Theorem 4.2 Assume that (1.2) has a Hermitian solution, and let S and T be as given in (1.9)
and (1.10). Then, the following hold.
(a) There exist X ∈ S and Y ∈ T such that X ≻ Y (X ≺ Y ) if and only if T A = 0. So that if
(1.8) is not null equation, there don't exist X ∈ S and Y ∈ T such that X ≻ Y (X ≺ Y ).
14
(b) There always exist X ∈ S and Y ∈ T such that X < Y (X 4 Y ).
Proof From (1.18) and (1.19), there exist X ∈ S and Y ∈ T such that X ≻ Y (X ≺ Y ) if and
only if
max
X∈S, Y ∈T
i+( X − Y ) = n (cid:18) max
X∈S, Y ∈T
i−( X − Y ) = n(cid:19) ;
there exist X ∈ S and Y ∈ T such that X < Y (X 4 Y ) if and only if
min
X∈S, Y ∈T
i−( X − Y ) = 0 (cid:18) min
X∈S1, Y ∈T
i+( X − Y ) = 0(cid:19) .
From (1.39) and (4.3), the difference of X − Y can be written as
(4.8)
(4.9)
X − Y = A†B(A∗)† − (T A)†T BT ∗(A∗T ∗)† + FAV + V ∗FA + FT AU + U ∗FT A
= A†B(A∗)† − (T A)†T BT ∗(A∗T ∗)† + [ FA, FT A ](cid:20) V
FT A(cid:21).
U(cid:21) + [ V ∗, U ∗ ](cid:20) FA
(4.10)
Applying (1.66) and simplifying by (1.28) and (4.5), we obtain
max
X∈S, Y ∈T
i±( X − Y ) = i±
A†B(A∗)† − (T A)†T BT ∗(A∗T ∗)† FA FT A
FA
FT A
0
0
= i±(cid:20) B − A(T A)†T BT ∗(A∗T ∗)†A∗ AFT A
FAA∗
= r(A) − r(T A) + n − r(A) = n − r(T A).
0
0
0
(cid:21) + r(FA)
Substituting (4.11) into (4.8) leads to the result in (a).
Applying (1.67) and simplifying by (1.25), (1.28) and (4.5), we obtain
min
X∈S, Y ∈T
i±( X − Y ) = i±
A†B(A∗)† − (T A)†T BT ∗(A∗T ∗)† FA FT A
FA
FT A
0
0
− r[ FA, FT A ]
(4.11)
(4.12)
= i±(cid:20) B − A(T A)†T BT ∗(A∗T ∗)†A∗ AFT A
FAA∗
0
= r(A) − r(T A) − r(A) + r(T A) = 0.
0
0
(cid:21) − r(AFT A)
Substituting (4.12) into (4.9) leads to the result in (b).
(cid:3)
5 Average equalities for Hermitian solutions of AX A∗
= B
and its two transformed equations
In this section, we study the relations between the two sets in (1.11) and (1.12).
Theorem 5.1 Assume that (1.2) has a Hermitian solution, and let S and T be defined as in (1.11)
and (1.12), in which T1 ∈ Cp1×n and T2 ∈ Cp2×n. Then,
(a) S ⊆ T always holds.
(b) S = T if and only if
r(cid:20) T1A
T2A(cid:21) = r(T1A) + r(T2A) − r(A).
(5.1)
(c) In particular, if r(T1A) = r(T2A) = r(A), then S = T .
15
Proof If X is a Hermitian solution of (1.2), then T1AXA∗T ∗
hold as well, and
1 = T1BT ∗
1 and T2AXA∗T ∗
2 = T2BT ∗
2
T1BT ∗
1
0
A∗T ∗
1
0
−T2BT ∗
2
A∗T ∗
2
r
T1A
T2A
0
= 2r(cid:20) T1A
T2A(cid:21).
(5.2)
holds by Lemma 1.12(c). This fact means that any Hermitian solution X of (1.2) can be written
as X = ( X + X )/2, the average of the two Hermitian solutions of the two equations in (1.12).
Hence, we have (a).
Note from (1.17) that the set inclusion S ⊇ T is equivalent to
max
Y ∈T
min
X∈S
r( X − Y ) = 0.
(5.3)
Applying (1.65), we first obtain
min
X∈S
r( X − Y ) = min
V
r(cid:2) A†B(A∗)† + FAV + V ∗FA − Y (cid:3) = r( B − AY A∗ ).
(5.4)
From Lemma 1.11, the general expression of the matrices of the two equations in (1.12) can be
written as
Y = (T1A)†T1BT ∗
− FT1AU1 − U ∗
1 (A∗T ∗
1 FT1A − FT2AU2 − U ∗
1 )†/2 + (T2A)†T2BT ∗
2 FT2A,
2 (A∗T ∗
2 )†/2
where U1 and U2 are arbitrary. Then
B − AY A∗ = B − A(T1A)†T1BT ∗
− AFT1AU1A∗ − AU ∗
1 (A∗T ∗
1 )†A∗/2 + A(T2A)†T2BT ∗
1 FT1AA∗ − AFT2 AU2A∗ − AU ∗
2 (A∗T ∗
2 FT2AA∗
2 )†A∗/2
= G − [ AFT1A, AFT2A ](cid:20) U1
U2 (cid:21) A∗ − A[ U ∗
1 , U ∗
2 ](cid:20) FT1AA∗
FT2AA∗ (cid:21),
(5.5)
(5.6)
where G = B − A(T1A)†T1BT ∗
1 (A∗T ∗
1 )†A∗/2 − A(T2A)†T2BT ∗
2 (A∗T ∗
2 )†A∗/2. Applying (1.60) gives
max
Y ∈T
r( B − AY A∗ ) = max
U1, U2
1 , U ∗
2 ](cid:20) FT1AA∗
FT2AA∗(cid:21)(cid:19)
(5.7)
r(cid:18)G − [ AFT1A, AFT2A ](cid:20) U1
r(A), r
FT1AA∗
FT2AA∗
0
0
G
U2(cid:21) A∗ − A[ U ∗
0
0
,
AFT1 A AFT2A
= min
16
where applying (1.35) and simplifying by elementary matrix operations, we obtain
G
AFT1A AFT2 A
FT1AA∗
FT2AA∗
0
0
0
0
r
2B − A(T1A)†T1BT ∗
1 (A∗T ∗
1 )†A∗ − A(T2A)†T2BT ∗
2 (A∗T ∗
− 2r(T1A) − 2r(T2A)
A∗
A∗
0
0
0
A∗
A∗
= r
= r
= r
= r
= r
= r
= 2r(cid:20) T1A
0
0
0
0
A∗T ∗
2
0
0
0
0
A∗T ∗
2
0
0
2 )†A∗ A
0
0
0
A
0 A∗T ∗
1
0
T1A 0
0
T2A
0
0
0
0
0
0
A
0
0
0 −T1A
T2A
0
0
A∗T ∗
1
0
0
0
0
A∗T ∗
2
0
0
−2T1B + T1BT ∗
1 (A∗T ∗
1 )†A∗ + T1A(T2A)†T2BT ∗
2 (A∗T ∗
2 )†A∗
− 2r(T1A) − 2r(T2A)
0
0
A∗
0
0 −T1A 2T1BT ∗
0
T2A
+ r(A) − 2r(T1A) − 2r(T2A)
0
0
−A∗T ∗
1
0
1 − T1BT ∗
1 − T1A(T2A)†T2BT ∗
2 (A∗T ∗
2 )†A∗T ∗
1
A∗T ∗
2
0
0
+ 2r(A) − 2r(T1A) − 2r(T2A)
+ 2r(A) − 2r(T1A) − 2r(T2A)
−A∗T ∗
1
0
A∗T ∗
2
−T1A T1BT ∗
T2A
1 − T1A(T2A)†T2BT ∗
2 (A∗T ∗
2 )†A∗T ∗
1
−A∗T ∗
1
−T1A T1BT ∗
1 −T1A(T2A)†T2BT ∗
2
T2A
0
0
0
−A∗T ∗
1
−T1A T1BT ∗
1
T2A
0
A∗T ∗
2
0
−T2BT ∗
2
+ 2r(A) − 2r(T1A) − 2r(T2A)
T2A(cid:21) + 2r(A) − 2r(T1A) − 2r(T2A) (by (5.2)).
Hence,
max
Y ∈T
r( B − AY A∗ ) = min(cid:26)r(A), 2r(cid:20) T1A
T2A(cid:21)+ 2r(A) − 2r(T1A) − 2r(T2A)(cid:27).
Setting the both sides to zero leads to the equality in (5.1). Under r(T1A) = r(T2A) = r(A), both
sides of (5.1) reduces to 2r(A). Hence, (c) follows.
(cid:3)
We next give some consequences of Theorem 5.1. Partitioning A and B in (1.2) as
A2(cid:21) X[ A∗
(cid:20) A1
1, A∗
3 B2(cid:21),
2 ] =(cid:20) B1 B3
B∗
where A1 ∈ Cm1×n A2 ∈ Cm2×n, B1 ∈ Cm1
T1 = [ Im1 , 0 ] and T = [ 0, Im2 ] in (1.8), respectively, we obtain two small equations
H , B3 ∈ Cm1×m2, m1 + m2 = m, and setting
H , B2 ∈ Cm2
A1XA∗
1 = B1 and A2XA∗
2 = B2.
Let
T = { (X1 + X2)/2 A1X1A∗
1 = B1, A2X2A∗
2 = B2, X1, X2 ∈ Cn
H } .
(5.10)
Applying Theorem 5.1 to (1.11) and (5.10), we obtain the following result.
17
(5.8)
(5.9)
Corollary 5.2 Assume that (1.2) has a solution, and let S be as given in (1.11) and T as in
(5.10). Then, S = T if and only if R(A∗
1) = R(A∗
2).
Decomposing A in (1.2) as A = A1 + A2 and setting T = EA1 and T = EA2 respectively yields
the following two transformed equations
EA1 A2XA∗
2EA1 = EA1 BEA1 , EA2 A1XA∗
1EA2 = EA2BEA2 .
Also let
T = {( X1 + X2 )/2 EA2A1X1A∗
1EA2 = EA2BEA2 , EA1A2X2A∗
2EA1 = EA1BEA1 , X1, X2 ∈ Cn
H} .
(5.11)
Corollary 5.3 Assume that (1.2) has a solution, and let S and T be as given in (1.11) and (5.11).
Then, the following hold.
(a) S = T if and only if
r(cid:20) A1
0 A2
0 A2 A1(cid:21) = 2r[ A1, A2 ] − r(A).
(b) Under R(A1) ∩ R(A2) = {0}, S = T if and only if R(A∗
1) = R(A∗
2).
Proof From Theorem 5.1(b), S = T if and only if
EA1A2(cid:21) = r(EA2 A1) + r(EA1 A2) − r(A),
r(cid:20) EA2A1
(5.12)
(5.13)
which is equivalent to (5.12) by (1.24).
(cid:3)
In addition to the average equalities of Hermitian solutions of (1.2) and its two transformed
equations, it would be of interest to consider weighted average equalities for the Hermitian solutions
of (1.2) and its k transformed equations
TiAXiA∗T ∗
i = TiBT ∗
i ,
i = 1, . . . , k,
(5.14)
where Ti ∈ Cpi×n. As usual, define
TiAXiA∗T ∗
i = TiBT ∗
i , Xi = X ∗
i ,
λi = 1, λi > 0, i = 1, . . . , k).
(5.15)
k
Xi=1
An open problem is to establish necessary and sufficient conditions for S = T to hold.
T =( k
Xi=1
λiXi (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
6 Equalities and inequalities between least-squares and least-
rank Hermitian solutions of AX A∗
= B
It is well known that the normal equation corresponding to the norm minimization problem in
(1.13) is given by
A∗AXA∗A = A∗BA,
(6.1)
see [1], while the normal equation corresponding to the rank minimization problem in (1.14) is
given by
ET1 Y ET1 = −ET1 T M †T ∗ET1 ,
(6.2)
A∗
where M =(cid:20) B A
0(cid:21), T = [ 0, In ] and T1 = T FM ; see [10]. Both (6.1) and (6.2) are transformed
equations of AXA∗ = B. Relations between solutions of the two equations were considered in [19].
In this section, we establish a group of formulas for calculating the extremal ranks and inertias of
X − Y for their Hermitian solutions, and use the formulas to solve Problem 1.5.
18
Lemma 6.1 Let A ∈ Cm×n and B ∈ Cm
H be given. Then the following hold.
(a) [1] The solution of (1.13) is
min
X∈Cn
H
arg min
X∈Cn
H
k B − AXA∗ kF = k B − AA†BAA† kF ,
k B − AXA∗ kF = A†B(A†)∗ + FAU + U ∗FA,
where U ∈ Cn×n is arbitrary.
(b) [10] The solution of (1.14) is
r( B − AY A∗ ) = 2r[ A, B ] − r(cid:20) B A
0(cid:21),
A∗
r( B − AY A∗ ) = −T M †T ∗ + T1V + V ∗T ∗
1 ,
min
Y ∈Cn
H
arg min
Y ∈Cn
H
where V ∈ C(m+n)×n is arbitrary.
Theorem 6.2 Let S and T be as given in (1.13) and (1.14), and define
M =(cid:20) B A
A∗
0(cid:21) , N =
B
A∗B
A∗
Then,
BA A
0
0
0
0
.
max
X∈S,Y ∈T
min
X∈S,Y ∈T
r( X − Y ) = min{ n, 2n + r(N ) − 2r(A) − r(M )},
r( X − Y ) = r(N ) + r(M ) − 2r[ A, B ] − 2r(A),
max
X∈S,Y ∈T
min
X∈S,Y ∈T
i±( X − Y ) = i∓(N ) + n − r(A) − i∓(M ),
i±( X − Y ) = i∓(N ) + i±(M ) − r[ A, B ] − r(A)
hold. Under the condition B < 0,
max
X∈S,Y ∈T
min
X∈S,Y ∈T
r( X − Y ) = min{ n, 2n + r[ A, BA ] − 3r(A)},
r( X − Y ) = r[ A, BA ] − r(A),
max
X∈S,Y ∈T
max
X∈S,Y ∈T
min
X∈S,Y ∈T
min
X∈S,Y ∈T
i+( X − Y ) = r[ A, BA ] + n − 2r(A),
i−( X − Y ) = n − r(A),
i+( X − Y ) = r[ A, BA ] − r(A),
i−( X − Y ) = 0.
(6.3)
(6.4)
(6.5)
(6.6)
(6.7)
(6.8)
(6.9)
(6.10)
(6.11)
(6.12)
(6.13)
(6.14)
(6.15)
(6.16)
In consequence, the following hold.
(a) There exist X ∈ S and Y ∈ T such that X − Y is nonsingular if and only if r(N ) >
2r(A) + r(M ) − n.
(b) X − Y is nonsingular for all X ∈ S and Y ∈ T if and only if r(N ) + r(M ) = 2r[ A, B ] +
2r(A) + n.
(c) There exist X ∈ S and Y ∈ T such that X = Y if and only if r(N ) + r(M ) = 2r[ A, B ] +
2r(A).
(d) There exist X ∈ S and Y ∈ T such that X ≻ Y if and only if i−(N ) = i−(M ) + r(A).
19
(e) X ≻ Y holds for all X ∈ S and Y ∈ T if and only if i−(N ) = r(A) + r[ A, B ] − i+(M ) + n.
(f) There exist X ∈ S and Y ∈ T such that X < Y if and only if i+(N ) = r(A) + r[ A, B ] −
i−(M ).
(g) X < Y holds for all X ∈ S and Y ∈ T if and only if i+(N ) = i+(M ) + r(A) − n.
(h) There exist X ∈ S and Y ∈ T such that X ≺ Y if and only if i+(N ) = i+(M ) + r(A).
(i) X ≺ Y holds for all X ∈ S and Y ∈ T if and only if i+(N ) = r(A) + r[ A, B ] − i−(M ) + n.
(j) There exist X ∈ S and Y ∈ T such that X 4 Y if and only if i−(N ) = r(A) + r[ A, B ] −
i+(M ).
(k) X 4 Y holds for all X ∈ S and Y ∈ T if and only if i−(N ) = i−(M ) + r(A) − n.
Under the condition B < 0, the following hold.
(l) There exist X ∈ S and Y ∈ T such that X − Y is nonsingular if and only if r[ A, BA ] >
3r(A) − n.
(m) X − Y is nonsingular for all X ∈ S and Y ∈ T if and only if r[ A, BA ] = r(A) + n.
(n) There exist X ∈ S and Y ∈ T such that X = Y if and only if R(BA) ⊆ R(A).
(o) There exist X ∈ S and Y ∈ T such that X ≻ Y if and only if r[ A, BA ] = 2r(A).
(p) X ≻ Y holds for all X ∈ S and Y ∈ T if and only if r[ A, BA ] = r(A) + n.
(q) There always exist X ∈ S and Y ∈ T such that X < Y.
(r) X < Y holds for all X ∈ S and Y ∈ T if and only if r(A) = n.
(s) There exist X ∈ S and Y ∈ T such that X ≺ Y if and only if A = 0.
(t) There exist X ∈ S and Y ∈ T such that X 4 Y if and only if R(BA) ⊆ R(A).
(u) X 4 Y holds for all X ∈ S and Y ∈ T if and only if r[ A, BA ] = 2r(A) − n.
Proof Applying (1.42) -- (1.45) to (6.1) and (6.2) gives
A∗BA
0
A∗A
n, r
0
A∗A
ET1 T M †T ∗ET1 ET1
ET1
0
+ 2n − 2r(A) − 2r(ET1 )
,
(6.17)
− 2r[ A∗A, ET1 ],
(6.18)
max
X∈S,Y ∈T
min
X∈S,Y ∈T
max
X∈S,Y ∈T
min
X∈S,Y ∈T
r( X − Y ) =
r( X − Y ) = r
i±( X − Y ) = i±
i±( X − Y ) = i±
0
ET1
0
0
0
0
A∗BA
A∗A
A∗BA
0
A∗A
ET1 T M †T ∗ET1 ET1
A∗A
A∗BA
A∗A
ET1 T M †T ∗ET1 ET1
0
0
0
r(ET1 ) = n − r(T1) = n − r(T FM ) = n − r
A∗A
ET1 T M †T ∗ET1 ET1
= r(M ) − r[ A, B ],
A∗A
ET1
ET1
+ n − r(A) − r(ET1 ),
(6.19)
− r[ A∗A, ET1 ].
(6.20)
0
In
B A
A∗
0
+ r(M )
(6.21)
Applying Lemma 1.8 and simplifying by elementary operations of matrices that
r[ A∗A, ET1 ] = r(A∗AT1) + r(ET1 ) = r(AT FM ) + r(M ) − r[ A, B ]
= r(cid:20) A∗AT
M (cid:21) − r[ A, B ] = r
0 A
B A
A∗
0
20
− r[ A, B ] = r(A),
(6.22)
and by Lemma 1.8 and elementary operations of matrices that
− r(T FM )
(by (1.35))
0
0
0
0 −A∗AT EM
In
0
0
− r(cid:20) M
(cid:21) − r[ A, B ] + r(M )
0
(by (1.28))
T (cid:21) + r(M )
(by (1.20))
(by (1.22))
0
0
A∗BA
A∗BA
T M †T ∗
0
ET1
0
A∗A
0
0
0
T FM
0
A∗A
In
EM T
−EM T ∗A∗A
A∗A
In
0
0
0
A∗A
ET1 T M †T ∗ET1 ET1
0
0
A∗BA + A∗AT M †T ∗A∗A 0
0
In
0
i±
= i±
= i±
= i±(cid:20) A∗BA + A∗AT M †T ∗A∗A A∗AT EM
= i±(cid:20) A∗BA A∗AT
T ∗A∗A −M (cid:21) − r[ A, B ] + i±(M )
= i±
= i±
= i∓
0
0
A∗A −A∗
−A∗B
0
A∗A
−B −A
− r[ A, B ] + i±(M )
BA A
0
= i∓(N ) − r[ A, B ] + i±(M ).
0
B
A∗B
A∗
EM T ∗A∗A
0
0
0
−BA
−B
−A∗
A∗BA
0
0
− r[ A, B ] + i±(M )
0
−A
− r[ A, B ] + i±(M )
(by (1.20))
(by (1.20) and (1.21))
(6.23)
(6.24)
Substituting (6.21) -- (6.24) into (6.17) -- (6.20) yields (6.7) -- (6.10).
Applying (1.30) to (6.7) -- (6.10) yields (6.11) -- (6.17). Applying Lemma 1.6 to (6.7) -- (6.10) leads
to (a) -- (k). Under the condition B < 0, applying Lemma 1.6 to (6.11) -- (6.17) leads to (l) -- (u).
(cid:3)
The positive semi-definite least-squares Hermitian solution of (1.2) is defined to an X that
satisfies
the positive semi-definite least-rank Hermitian solution of (1.2) is defined to be a Y that satisfies
k B − AXA∗ kF = min s.t. X < 0;
(6.25)
The normal equations corresponding to (6.25) and (6.26) are given by
r( B − AY A∗ ) = min s.t. Y < 0.
A∗AXA∗A = A∗BA, X < 0,
ET1 Y ET1 = −ET1T M †T ∗ET1 , Y < 0.
(6.26)
(6.27)
(6.28)
Under the condition B < 0, (6.27) and (6.28) have solutions. In this case, it would be of interest
to characterize the following four inequalities
X ≻ Y < 0, X < Y < 0, Y ≻ X < 0, Y < X < 0.
(6.29)
As demonstrated in the previous sections, matrix ranks and inertias and their optimizations
problems are one of the most productive research field in matrix theory. The present author and
his collaborators paid great attention in recent years for the development of the theory on matrix
ranks and inertias, and established thousands of expansion formulas for calculating ranks and
inertias of matrices; see, e.g., [3, 4, 5, 6, 8, 9, 10, 11, 12, 14, 15, 16, 17, 18, 19, 20, 21, 22] for more
details. In addition, many follow-up papers by other people were also published on extensions and
applications of matrix rank and inertia formulas in different situations.
21
References
[1] Ben-Israel, A., Greville, T.N.E.: Generalized Inverses: Theory and Applications. Second ed., Springer,
New York, 2003
[2] Khatri, C.G., Mitra, S.K.: Hermitian and nonnegative definite solutions of linear matrix equations.
SIAM J. Appl. Math. 31, 579 -- 585 (1976)
[3] Liu, Y., Tian, Y.: Extremal ranks of submatrices in an Hermitian solution to the matrix equation
AXA∗ = B with applications. J. Appl. Math. Comput. 32, 289 -- 301 (2010)
[4] Liu, Y., Tian, Y.: A simultaneous decomposition of a matrix triplet with applications. Numer. Linear
Algebra Appl. 18, 69 -- 85 (2011)
[5] Liu, Y., Tian, Y.: Max-min problems on the ranks and inertias of the matrix expressions A − BX C ±
(BX C)∗ with applications. J. Optim. Theory Appl. 148, 593 -- 622 (2011)
[6] Liu, Y., Tian, Y., Takane, Y.: Ranks of Hermitian and skew-Hermitian solutions to the matrix
equation AXA∗ = B. Linear Algebra Appl. 431, 2359 -- 2372 (2009)
[7] Marsaglia, G., Styan, G.P.H.: Equalities and inequalities for ranks of matrices. Linear Multilinear
Algebra 2, 269 -- 292 (1974)
[8] Tian, Y.: On additive decompositions of solutions of the matrix equation AXB = C. Calcolo 47,
193 -- 209 (2010)
[9] Tian, Y.: Completing block Hermitian matrices with maximal and minimal ranks and inertias. Elec-
tron. J Linear Algebra 21, 124 -- 141 (2010)
[10] Tian, Y.: Equalities and inequalities for inertias of Hermitian matrices with applications. Linear
Algebra Appl. 433, 263 -- 296 (2010)
[11] Tian, Y.: Maximization and minimization of the rank and inertia of the Hermitian matrix expression
A − BX − (BX)∗ with applications. Linear Algebra Appl. 434, 2109 -- 2139 (2011)
[12] Tian, Y.: Expansion formulas for the inertias of Hermitian matrix polynomials and matrix pencils of
orthogonal projectors. J. Math. Anal. Appl. 376, 162 -- 186 (2011)
[13] Tian, Y.: Extremal ranks of a quadratic matrix expression with applications. Linear Multilinear
Algebra 59, 627 -- 644 (2011)
[14] Tian, Y.: Formulas for calculating the extremum ranks and inertias of a four-term quadratic matrix-
valued function and their applications. Linear Algebra Appl. 437, 835 -- 859 (2012)
[15] Tian, Y.: Solutions to 18 constrained optimization problems on the rank and inertia of the linear
matrix function A + BXB ∗. Math. Comput. Modelling 55, 955 -- 968 (2012)
[16] Tian, Y.: On an equality and four inequalities for generalized inverses of Hermitian matrices. Electron.
J. Linear Algebra 23, 11 -- 42 (2012)
[17] Tian, Y.: On additive decomposition of the Hermitian solution of the matrix equation AXA∗ = B.
Mediterr. J. Math. 9, 47 -- 60 (2012)
[18] Tian, Y.: Solving optimizations problems on ranks and inertias of some constrained nonlinear matrix
functions via an algebraic linearization method. Nonlinear Analysis A 75, 717 -- 734 (2012)
[19] Tian, Y.: On relations between least-squares and least-rank Hermitian solutions of the matrix equation
AXA∗ = B and their relations. Numer. Linear Algebra Appl., doi:10.1002/nla.829
[20] Tian, Y., Li, Y.: Distributions of eigenvalues and inertias of some block Hermitian matrices consisting
of orthogonal projectors. Linear Multilinear Algebra 60, 1027 -- 1069 (2012)
[21] Tian, Y., Liu, Y.: Extremal ranks of some symmetric matrix expressions with applications. SIAM J.
Matrix Anal. Appl. 28, 890 -- 905 (2006)
[22] Tian, Y., von Rosen, D.: Solving the matrix inequality AXB + (AXB)∗ > C. Math. Inequal. Appl.
15, 537 -- 548 (2012)
Yongge Tian
China Economics and Management Academy
Central University of Finance and Economics
Beijing 100081, China
e-mail: [email protected]
22
|
1508.07496 | 3 | 1508 | 2016-08-20T18:48:02 | On the utility of Robinson-Amitsur ultrafilters. II | [
"math.RA",
"math.LO"
] | We extend the result of arXiv:0911.5414 about embedding of ideal-determined algebraic systems into ultraproducts, to arbitrary algebraic systems, and to ultraproducts over $\kappa$-complete ultrafilters. We also discuss the scope of applicability of this result, and correct a mistake from the previous paper concerning homomorphisms from ultraproducts. | math.RA | math |
ON THE UTILITY OF ROBINSON -- AMITSUR ULTRAFILTERS. II
PASHA ZUSMANOVICH
ABSTRACT. We extend the result of the previous paper under the same title about embedding of ideal-
determined algebraic systems into ultraproducts, to arbitrary algebraic systems, and to ultraproducts over
k -complete ultrafilters. We also discuss the scope of applicability of this result, and correct a mistake from
the previous paper concerning homomorphisms from ultraproducts.
0. Introduction. In [Z], an old and simple trick, used by A. Robinson and S. Amitsur (see [A1, Proof
of Theorem 15] and [A2, Theorem 3]) to establish, in the context of ring theory, an embedding of
certain rings in ultraproducts, was generalized to algebraic systems with ideal-determined congruences.
Here we push it to the full generality, for arbitrary algebraic systems without any restrictions on their
congruences.
1. Recollection on congruences, ultrafilters, and ultraproducts. We refer to [C.B] and [CK] for the
rudiments of universal algebra and model theory we need.
defined on them, of, generally, various arity. The signature W
Let us recall some basic notions and fix notation. We consider the most general algebraic systems,
i.e. sets with a number of operations W
is
arbitrary, but fixed (so, in what follows all algebraic systems are supposed to be of signature W
). The set
of all congruences on a given algebraic system A forms a partially ordered set (actually, a lattice), and
the minimal element in this set is the trivial congruence, coinciding with the diagonal {(a, a) a ∈ A} in
A × A. The intersection of all congruences containing a given relation r , i.e. a set of pairs of elements
from A ×A, is called a congruence generated by r and is denoted by Con(r ). In the case of one-element
relation, i.e. a single pair (a, b) ∈ A ×A with a 6= b, we shorten this notation to Con(a, b) and speak about
principal congruences.
Given a cardinal k > 2, let us call a set S of sets k -complete, if the intersection of any nonempty set
of fewer than k elements of S belongs to S . If a set S of subsets of a set I satisfies a weaker condition
-- that the intersection of any nonempty set of fewer than k elements of S contains an element of S
-- then the set of subsets of I which are oversets of all such intersections is a k -complete filter on I
containing S (this is an obvious generalization of the standard and frequently employed fact that any
set of subsets satisfying the finite intersection property can be extended to a filter).
An algebraic system A is called k -subdirectly irreducible, if either of the following two equivalent
conditions is satisfied:
• The set of nontrivial congruences of A is k -complete.
• If A embeds in the direct product of < k algebraic systems (cid:213)
one of Bi's.
i∈I Bi, I < k , then A embeds in
Obviously, the condition of w -subdirect irreducibility, also called finite subdirect irreducibility, is equiv-
alent to n-subdirect irreducibility for any finite n > 2. Note that w -complete filters (ultrafilters) are just
the usual filters (ultrafilters).
Given a filter (respectively, ultrafilter) F on a set I, the quotient of the direct product of algebraic
systems (cid:213)
i∈I Ai by the congruence
q F = {(a, b) ∈ ((cid:213)
Ai) × ((cid:213)
i∈I
i∈I
Ai) {i ∈ I a(i) = b(i)} ∈ F }
Date: last revised August 14, 2016.
J. Algebra, to appear; arXiv:1508.07496.
1
ON THE UTILITY OF ROBINSON -- AMITSUR ULTRAFILTERS. II
2
is called filtered product (respectively, ultraproduct) of the corresponding algebraic systems, and is
denoted by (cid:213) F Ai.
As any k -complete ultrafilter on a set of cardinality < k
is principal, the condition of k -subdirect
irreducibility of an algebraic system A may be trivially reformulated as follows:
• If A embeds in the direct product of < k algebraic systems (cid:213)
the ultraproduct (cid:213) U Bi for some ultrafilter U on I.
i∈I Bi, I < k , then A embeds in
What is, perhaps, surprising, is that under a suitable set-theoretic assumption on the cardinal k , the latter
condition is equivalent to the same condition with an arbitrary index set I, without a restriction on its
cardinality. This equivalence is what constitutes the "Robinson -- Amitsur theorem" (Theorem 1 below).
2. Embeddings in direct products. The next proof follows the proof of Theorem 1.1 from [Z], with
ideals being replaced by congruences, and w being replaced by an arbitrary cardinal k . On this level of
generality, the proof is even simpler than an already simple proof from [Z].
Theorem 1 (ROBINSON -- AMITSUR). Let k be a cardinal > 2 and such that any k -complete filter can
be extended to a k -complete ultrafilter. Then for any algebraic system A the following are equivalent:
(i) A is k -subdirectly irreducible.
(ii) For any embedding f of A in the direct product (cid:213)
i∈I Bi of a set of algebraic systems {Bi}i∈I,
there is a k -complete ultrafilter U on the set I such that the composition of f with the canonical
homomorphism (cid:213)
i∈I Bi → (cid:213) U Bi, is an embedding.
Proof. (ii) ⇒ (i): take I < k and observe, as above, that then the ultrafilter U is principal.
(i) ⇒ (ii). For any two elements a, b ∈ A, define Sa,b = {i ∈ I a(i) 6= b(i)}, and let S = {Sa,b a, b ∈
A, a 6= b}. Let us verify that the intersection of < k elements of S contains an element of S .
Let a, b be two elements of A such that a 6= b. If for some i0 ∈ I we have a(i0) = b(i0), then (a, b)
belongs to the congruence of (cid:213)
i∈I Bi which is the kernel of the canonical projection Pri0 to Bi0, and by the
Third Isomorphism Theorem, belongs to the congruence Ker(Pri0) ∩ (A × A) of A, whence Con(a, b) ⊆
Ker(Pri0). The latter means that for any (c, d) ∈ Con(a, b), we have Sc,d ⊆ Sa,b.
Due to k -subdirect irreducibility of A, for any set of pairs of different elements {ai, bi} of A of
cardinality < k , the intersection Ti<k Con(ai, bi) is a nontrivial congruence. Pick an element (c, d),
c 6= d, from this intersection. Then by just proved we have Sc,d ⊆ Ti<k Sai,bi, as required.
Thus S can be extended to a k -complete filter, and then to a k -complete ultrafilter U on I. Factoring
i∈I Bi by the congruence q U , we get, again by the Third Isomorphism Theorem,
the embedding of A in (cid:213)
an embedding of the quotient A/(q U ∩ (A × A)) in the ultraproduct (cid:213) U Bi. But if (a, b) ∈ q U ∩ (A × A),
then, by definition, {i ∈ I a(i) = b(i)} ∈ U , and, since U is an ultrafilter, {i ∈ I a(i) 6= b(i)} /∈ U ,
and hence {i ∈ I a(i) 6= b(i)} /∈ S , what, in its turn, implies a = b. This shows that the congruence
q U ∩ (A × A) is trivial, and hence A embeds in (cid:213) U Bi.
(cid:3)
Coupling the case k = w of this theorem with the Birkhoff theorem about varieties of algebraic
systems, and some rudimentary model theory, we get:
Corollary (CRITERION FOR ABSENCE OF NONTRIVIAL IDENTITIES). Let A be a variety of algebraic
systems such that any free system in A is finitely subdirectly irreducible. Then for an algebraic system
A ∈ A the following three conditions are equivalent:
(i) A does not satisfy a nontrivial identity within A.
(ii) Any free system in A embeds in an ultrapower of A.
(iii) Any free system in A embeds in a filtered power of A.
Proof. (i) ⇒ (ii) is established exactly as in [Z, Corollaries 1.2 and 1.3]: by the Birkhoff theorem, a free
system in A embeds in the direct power of A. Apply Theorem 1, with k = w
.
(ii) ⇒ (iii): obvious.
(iii) ⇒ (i): follows from the fact that any nontrivial identity is a first-order Horn sentence, and Horn
(cid:3)
sentences are preserved under filtered products (see, e.g., [CK, Proposition 6.2.2]).
ON THE UTILITY OF ROBINSON -- AMITSUR ULTRAFILTERS. II
3
3. No way to move beyond w
. Discussion. Let us make a few remarks about the scope of applicability
of Theorem 1 and Corollary. By the Ło´s theorem, an algebraic system is elementarily equivalent to its
ultrapower, so if the condition of Corollary is met, it allows to reduce the question about the absence
of nontrivial identities in an algebraic system A to the question whether a system elementary equivalent
to A contains a free subsystem. This was demonstrated to be useful in some situations in [Z], but the
usefulness is severely restricted by the fact that any property of an algebraic system A we can use in
the process should be a first-order property. Most of the interesting properties of algebraic systems are
second- or higher-order.
Note that by the analog of Ło´s theorem for ultraproducts over k -complete ultrafilters, the ultrapower
AU of an algebraic system A is elementarily equivalent to A in the sense of certain higher-order logic
(denoted by Lk
in [D, §3.3]). But unfortunately, we cannot derive an analog
of Corollary to Theorem 1 for k > w
, at least in, arguably, the most interesting for applications cases --
groups and algebras over fields (the main protagonists of [Z]). The reason is that free systems in these
varieties are residually nilpotent (the intersection of all terms of the lower central series is trivial) and
hence are already not w 1-subdirectly irreducible.
in [CK, §4.2], and by Lkk
Moreover, it seems that no statement similar to Corollary, where the (usual) ultrapower construction
is replaced by whatever other construction providing us with elementary equivalence in the sense of
second- or higher-order logic of some sort, is possible. Indeed, any such logic L should be a priori weak
enough not to allow encode the fact that an algebraic system contains a free subsystem: otherwise, if in
the condition (ii) of Corollary we would be able to replace the ultrapower of A by another construction
elementary equivalent to A in the sense of L, this would imply that for any algebraic system the absence
of a nontrivial identity is equivalent to having a free subsystem -- an obviously false statement. While,
in general, free systems in arbitrary varieties are not necessary countable, they are, essentially, so in
all cases of interest (groups and algebras over fields): all identities in a given variety are captured by
free systems of no more than countable rank; such free groups are, obviously, countable, and such free
algebras are countable-dimensional over the base field. In the latter case, as identities of algebras do not
change under field extensions (except, possibly, the cases of finite base fields of "small" cardinality), we
may always assume the base field to be countable, and, hence, the whole free algebra to be countable.
To summarize: the hypothetical logic L should not allow to encode the fact that an algebraic system
contains a countable subsystem -- a very weak logic indeed, on the verge not to be qualified as the
"second-order" one!
What about the condition on k
in Theorem 1 (sometimes called strong compactness)? In the case
k = w
this amounts to the well-known fact that any filter can be extended to an ultrafilter (true by
the Zorn lemma). In the general case we are in deep waters of set theory. It seems that this is one
of the statements about which one does not make sense to speak if they are "true" or "false", but it is
rather a matter of which model of set theory is adopted. See [D, §3.3, Parts C,D] for this and many
other conditions of such sort, and relationships between them. Note also that the very existence of non-
principal k -complete ultrafilters for k > w
, being equivalent to the existence of measurable cardinals,
entails rather strange properties of the corresponding model(s) of set theory (see, again, [CK, §4.2] and
[D, §0.4]).
Note finally that it would be very interesting to extend arguments used in the proof of Theorem 1 to
the case of metric ultraproducts. This could provide an approach to various questions related to sofic
and hyperlinear groups.
4. Semigroups. Arguably, the most interesting class of algebraic systems whose congruences are not
ideal-determined, i.e. not covered by [Z], but by Theorem 1 and Corollary above, is the class of semi-
groups. Unfortunately, the Corollary is not applicable to the variety of all semigroups, as free semi-
groups of rank > 1 are not finitely subdirectly irreducible. For example, if x and y are among free gen-
erators of a free semigroup G, then the congruences of G generated by pairs (x, x2) and (y, y2) intersect
trivially. It seems to be interesting to investigate for which classes of semigroups (inverse semigroups?
ON THE UTILITY OF ROBINSON -- AMITSUR ULTRAFILTERS. II
4
Burnside varieties?) the Corollary would be applicable, and what good it could do. The same question
for the class of loops.
: (cid:213)
5. Homomorphisms from direct products. Finally, we take the opportunity to note a mistake in [Z].
Theorem 7.1 there contains a statement dual, in a sense, to Theorem 1: if there is a surjective homomor-
phism a
i∈I Bi → A, where A is finitely subdirectly irreducible, then A is a surjective homomorphic
image of an ultraproduct of the same family {Bi}. The proof of this theorem is in error. At the top of
p. 283 in the published version (at the bottom of p. 13 -- top of p. 14 in the arXiv version), it is, essen-
tially, claimed (keeping the notation and terminology of the original paper) that any equality in A of the
form a (u) = t(b1, . . ., bn,a ( f ), . . .,a ( f )), where u, f ∈ (cid:213)
i∈I Bi, b1, . . . , bn ∈ A, and t is an ideal term in
arguments occupied by a ( f )'s, can be lifted to (cid:213)
i∈I Bi
such that a (hi) = bi. There is no reason whatsoever for this to be true in general: for example, in the
simplest possible case t(x) = x (the ideal term just in one variable, without arguments occupied by bi's),
this claim amounts to saying that a
i∈I Bi: u = t(h1, . . ., hn, f , ..., f ) for some hi ∈ (cid:213)
is an isomorphism.
Luckily, we can slightly modify the statement, and supply it with a new proof. While the new state-
ment differs from the original one, it is more general for the most classes of algebraic systems of interest.
The proof uses an ingenious (and barely utilizing anything beyond the definition of ultrafilter) argument
from an array of recent papers by G. Bergman, some of them jointly with N. Nahlus (e.g., [BN] and
[G.B]), devoted to factoring homomorphisms from direct product of groups, algebras over fields, and
modules. The argument is valid for arbitrary algebraic systems, and we reproduce it here, throwing in
k -complete ultrafilters for good measure.
Theorem 2 (BERGMAN -- NAHLUS). For any cardinal k > 2, and any algebraic system A consisting of
more than one element, the following are equivalent:
(i) For any surjective homomorphism f of the direct product (cid:213)
such that f factors through the canonical projection (cid:213)
i∈I Bi, I < k , to A, there is i0 ∈ I
i∈I Bi to A, there is a k -complete
ultrafilter U on the indexing set I such that f factors through the canonical homomorphism
(ii) For any surjective homomorphism f of the direct product (cid:213)
i∈I Bi → Bi0.
i∈I Bi → (cid:213) U Bi.
Proof. (ii) ⇒ (i) is obvious (like in a similar situation in Theorem 1, take I < k and observe that U is
necessary principal).
(i) ⇒ (ii). Let U consist of subsets J ⊆ I such that f factors through a surjective homomorphism
from the (smaller) direct product of Bi's indexed over J, i.e. the diagram
Bi
i∈I
f ✲ A
✒
❄
Bi
i∈J
where the vertical arrow is the canonical projection, commutes. Utilizing the condition (i) with I = 2,
and reasoning as in [BN, Lemma 7 and Proposition 8] or [G.B, Lemma 1.2], we get that U is an
ultrafilter, and f factors through the canonical homomorphism (cid:213)
To prove that U is k -complete, it is sufficient to show that for any decomposition I = S j<k I j into
the union of pairwise disjoint sets I j,
i∈I Bi =
Bi for some j0 < k , i.e.
j<k (cid:16)(cid:213)
I j0 ∈ U , as required.
i∈I j Bi(cid:17), f factors through the canonical projection (cid:213)
j < k , at least one of them belongs to U . Since (cid:213)
i∈I Bi → (cid:213) U Bi.
i∈I Bi → (cid:213)
Note that in the case of groups and algebras over fields, the condition (i) of Theorem 2 with k = w
is
weaker than the finite subdirect irreducibility. This follows from [BN, Lemma 5] in the case of algebras
over fields, and in the case of groups the argument is repeated almost verbatim.
i∈I j0
(cid:3)
(cid:213)
(cid:213)
(cid:213)
(cid:213)
ON THE UTILITY OF ROBINSON -- AMITSUR ULTRAFILTERS. II
5
Note also that though Theorems 1 and 2 are, in a sense, dual to each other, this duality does not stretch
to the proofs: the proofs are different, and each proof seemingly cannot be adapted to the dual situation.
Acknowledgements. Thanks are due to Alexander Gein, Fares Maalouf, and Eugene Plotkin for very
useful remarks: Gein has dispelled my illusion about well-behavior of semigroup congruences, Maalouf
has turned my attention to a mistake in [Z], and Plotkin has insisted, a long time ago, that the setup of [Z]
can be generalized further (it took me several years to realize this -- obvious, as it looks now -- fact). The
financial support of the Regional Authority of the Moravian-Silesian Region (grant MSK 44/3316) and
of the Ministry of Education and Science of the Republic of Kazakhstan (grant 0828/GF4) is gratefully
acknowledged.
REFERENCES
[A1] S.A. Amitsur, Generalized polynomial identities and pivotal monomials, Trans. Amer. Math. Soc. 114 (1965), 210 --
226; reprinted in Selected Papers of S.A. Amitsur with Commentary, Part 1, AMS, 2001, 191 -- 208.
[A2]
, Prime rings having polynomial identities with arbitrary coefficients, Proc. London Math. Soc. 17 (1967),
470 -- 486; reprinted in Selected Papers of S.A. Amitsur with Commentary, Part 1, AMS, 2001, 417 -- 433.
[C.B] C. Bergman, Universal Algebra, CRC Press, 2012.
[G.B] G.M. Bergman, Families of ultrafilters, and homomorphisms on infinite direct product algebras, J. Symb. Logic 79
(2014), 223 -- 239; arXiv:1301.6383.
[BN]
and N. Nahlus, Homomorphisms on infinite direct product algebras, especially Lie algebras, J. Algebra 333
(2011), 67 -- 104; arXiv:0910.5181.
[CK] C.C. Chang and H.J. Keisler, Model Theory, Studies in Logic and the Foundations of Mathematics, Vol. 73, North-
Holland, 3rd ed., 1990; reprinted by Dover, 2012.
[D] M.A. Dickmann, Large Infinitary Languages, Studies in Logic and the Foundations of Mathematics, Vol. 83, North-
Holland, 1975.
P. Zusmanovich, On the utility of Robinson -- Amitsur ultrafilters, J. Algebra 388 (2013), 268 -- 286; arXiv:0911.5414v7.
[Z]
DEPARTMENT OF MATHEMATICS, UNIVERSITY OF OSTRAVA, OSTRAVA, CZECH REPUBLIC
E-mail address: [email protected]
|
1902.03396 | 1 | 1902 | 2019-02-09T09:37:44 | Commuting maps on certain incidence algebras | [
"math.RA",
"math.OA"
] | Let $\mathcal{R}$ be a $2$-torsion free commutative ring with unity, $X$ a locally finite pre-ordered set and $I(X,\mathcal{R})$ the incidence algebra of $X$ over $\mathcal{R}$. If $X$ consists of a finite number of connected components, in this paper we give a sufficient and necessary condition for each commuting map on $I(X,\mathcal{R})$ being proper. | math.RA | math | COMMUTING MAPS ON CERTAIN INCIDENCE ALGEBRAS
HONGYU JIA AND ZHANKUI XIAO
Abstract. Let R be a 2-torsion free commutative ring with unity, X a locally
finite pre-ordered set and I(X, R) the incidence algebra of X over R. If X
consists of a finite number of connected components, in this paper we give a
sufficient and necessary condition for each commuting map on I(X, R) being
proper.
9
1
0
2
b
e
F
9
]
.
A
R
h
t
a
m
[
1
v
6
9
3
3
0
.
2
0
9
1
:
v
i
X
r
a
1. Introduction
Let A be an associative algebra over R, a commutative ring with unity. Then
A has the Lie algebra structure under the Lie bracket [x, y] := xy − yx. An R-
linear map θ : A → A is called a commuting map if [θ(x), x] = 0 for all x ∈ A. A
commuting map θ of A is said to be proper if it can be written as
θ(x) = λx + µ(x),
∀x ∈ A,
where λ ∈ Z(A), the center of A, and µ is an R-linear map with image in Z(A). A
commuting map which is not proper will be called improper. The purpose of this
paper is to identify a class of algebras on which every commuting map is proper.
As far as we know, the first result about commuting maps is due to Posner. He
showed in [11] that the existence of a nonzero commuting derivation on a prime
algebra A implies the commutativity of A. One of the major promoter for studying
commuting maps is Bresar. He [3] initially related the commuting maps to some
Herstein's conjectures [7] which described the forms of Lie-type maps (for example,
Lie isomorphisms, Lie derivations) on associative simple or prime rings. We warmly
encourage the reader to read the well-written survey paper [4], in which the author
presented the development of the theory of commuting maps and their applications,
especially to Lie theory. More results related to commuting maps are considered in
[1, 2, 5, 6, 9, 10, 17] etc.
We now recall another notion, incidence algebras [12], with which we deal in
this paper. Let (X, 6) be a locally finite pre-ordered set (i.e., 6 is a reflexive and
transitive binary relation). Here the local finiteness means for any x 6 y there are
only finitely many elements z ∈ X satisfying x 6 z 6 y. The incidence algebra
I(X, R) of X over R is defined on the set of functions
I(X, R) := {f : X × X −→ R f (x, y) = 0 if x (cid:10) y}
2010 Mathematics Subject Classification. Primary 16W25, Secondary 15A78, 47L35.
Key words and phrases. commuting map, incidence algebra.
This work is partially supported by the NSF of Fujian Province (No. 2018J01002) and the
National Natural Science Foundation of China (No. 11301195).
1
2
HONGYU JIA AND ZHANKUI XIAO
with the natural R-module structure and the multiplication given by the convolu-
tion
(f g)(x, y) := X
x6z6y
f (x, z)g(z, y)
for all f, g ∈ I(X, R) and x, y ∈ X. It would be helpful to claim that the full matrix
algebra Mn(R), the upper (or lower) triangular matrix algebras Tn(R), and the
infinite triangular matrix algebras T∞(R) are examples of incidence algebras. The
incidence algebra of a partially ordered set was first considered by Ward [14] as a
generalized algebra of arithmetic functions. Rota and Stanley developed incidence
algebras as fundamental structures of enumerative combinatorial theory and the
allied areas of arithmetic function theory (see [13]).
Recently, the second author [16] studied the Herstein's Lie-type mapping research
program (see [4]) on incidence algebras and proved that every Jordan derivation of
I(X, R) degenerates to a derivation, provided that R is 2-torsion free. Since then
more and more Lie-type maps were considered on incidence algebras, see [8, 15, 18]
etc. Roughly speaking, all the known Lie-type maps of I(X, R) are proper or of
the standard form. On the other hand, it is well-known that the commuting maps
of the triangular matrix algebras Tn(R) and the full matrix algebra Mn(R) are
proper (see [6, 17]). Therefore, it is hopefully that every commuting map of the
incidence algebra I(X, R) is proper. Unfortunately, we can find a counter-example
2
◦.
as follows. Let X be the partially ordered set with Hasse diagram
Then
1
◦ ←
3
◦ →
I(X, R) =
a11
0
0
0
a13
a22 a23
0
a33
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
aij ∈ R
.
The reader can verify that the R-linear map θ defined by
θ :
a11
0
0
0
a13
a22 a23
0
a33
7−→
a11
0
0
0
a33
0
a13
0
a33
is a commuting map, but generally improper. Hence it is fascinating to study
the relationship between commuting maps and the structure of incidence algebras,
especially the combinatorial structure. This is our main motivation of this paper.
2. The finite rank case
From now on, we always assume that all the rings and algebras are 2-torsion
free. In this section, we study commuting maps of the incidence algebra I(X, R)
when X is a finite pre-ordered set. For reader's convenience, let's start with two
well-known results for general algebras.
Lemma 2.1. ([6, Proposition 2]) Let A, B be two R-algebras. Then A and B have
no improper commuting maps if and only if A ⊕ B has no improper commuting
maps.
Lemma 2.2. Let A be an R-algebra with an R-linear basis Y . Then an R-linear
map θ : A → A is a commuting map if and only if [θ(x), y] = [x, θ(y)] for all
x, y ∈ Y .
COMMUTING MAPS ON CERTAIN INCIDENCE ALGEBRAS
3
We now introduce some standard notations for the incidence algebra I(X, R).
The unity element δ of I(X, R) is given by δ(x, y) = δxy for x 6 y, where δxy ∈
{0, 1} is the Kronecker delta.
If x, y ∈ X with x 6 y, let exy be defined by
exy(u, v) = 1 if (u, v) = (x, y), and exy(u, v) = 0 otherwise. Then exyeuv = δyuexv
by the definition of convolution. Clearly the set B := {exy x 6 y} forms an
R-linear basis of I(X, R) when X is finite. For i 6 j and i 6= j, we write i < j or
j > i for short.
When X is finite, I(X, R) is also known as a digraph algebra. This means that
there is a directed graph with the vertex set X associated with I(X, R). This graph
contains all the self loops and the matrix unit exy corresponds to a directed edge
from y to x. We now define an equivalent relation on D := {(x, y) x < y}, the set
of all directed edges associated with the pre-ordered set X. Let ∼ be the relation
on X defined by x ∼ y if and only if x 6 y or y 6 x, in other words, there is at
least one directed edge between the vertices x and y. We also write y ∼ x for x ∼ y
by the symmetry. Let x1, x2, . . . , xn be n different vertices in X with n ≥ 2. We
say that {x1, x2, . . . , xn} forms a circle if
(i) x1 ∼ x2 when n = 2;
(ii) x1 ∼ x2, . . . , xi−1 ∼ xi, . . . , xn−1 ∼ xn, xn ∼ x1 when n ≥ 3.
Definition 2.3. For any two directed edges (x, y) and (u, v) in D, we define (x, y) ∼=
(u, v) if and only if there is a circle contains both x ∼ y and u ∼ v.
The reader can verify that the binary relation ∼= is in deed an equivalent relation
on the set D. We remark here that the n = 2 case does not satisfy the standard
notion, circle, in combinatorial theory, but it provides the reflexivity of the rela-
tion ∼=. The example appeared in the introduction, where (1, 3) ≇ (2, 3), and the
following Example 2.4 can help us get a more clear picture for the relation ∼=.
Example 2.4. Let X = {1, 2, 3, 4} with partially ordered relations (or arrows) 1 <
2, 2 < 3, 2 < 4. Then the associated set of directed edges is D = {(1, 2), (2, 3), (1, 3),
(2, 4), (1, 4)}. We have (2, 3) ∼= (2, 4), since 2 ∼ 3, 3 ∼ 1, 1 ∼ 4, 4 ∼ 2, and D forms
one equivalent class under the relation ∼=.
The main result of this section is as follows.
Theorem 2.5. Let R be a 2-torsion free commutative ring with unity and X be
finite. Then every commuting map θ of I(X, R) is proper if and only if the set of
all directed edges associated with each connected component forms one equivalent
class under the relation ∼=.
We only need to prove Theorem 2.5 when X is connected. In fact, assume that
X = Fi∈I Xi be the union of its distinct connected components, where I is a
finite index set. Let δi := Px∈Xi
exx. It follows from [12, Theorem 1.3.13] that
{δi i ∈ I} forms a complete set of central primitive idempotents. In other words,
I(X, R) = Li∈I δiI(X, R). Clearly δiI(X, R) ∼= I(Xi, R) for each i ∈ I. Hence we
only need to prove Theorem 2.5 when X is connected by Lemma 2.1.
From now on until the end of this section, we assume X is finite and connected.
Let θ : I(X, R) → I(X, R) be an arbitrary commuting map. We denote for all
i, j ∈ X with i 6 j
θ(eij ) = X
exy ∈B
C ij
xyexy.
We also make the convention C ij
xy = 0, if needed, for x (cid:10) y.
4
HONGYU JIA AND ZHANKUI XIAO
Lemma 2.6. The commuting map θ satisfies
θ(eii) = X
x∈X
θ(eij ) = X
x∈X
C ii
xxexx;
C ij
xxexx + C ij
ij eij,
if i 6= j.
(1)
(2)
Proof. Without loss of generality, we assume that X ≥ 2. Since [θ(eii), exx] =
[eii, θ(exx)] for any i, x ∈ X, multiplying this identity by exx on the left and by eyy
on the right, we have
Then left multiplication by exx on [θ(eii), eii] = 0 leads to
C ii
xy = 0, if i 6= x < y 6= i.
C ii
xi = 0, if x < i.
Similarly, right multiplication by eyy on [θ(eii), eii] = 0 leads to
C ii
iy = 0, if i < y.
Combining the identities (3), (4) and (5), we obtain
θ(eii) = X
x6y
xyexy = X
C ii
x∈X
xxexx + X
C ii
x<y
C ii
xyexy
(3)
(4)
(5)
= X
x∈X
= X
x∈X
xxexx + X
C ii
i6=x<y6=i
xyexy + X
C ii
x<i
xiexi + X
C ii
i<y
C ii
iyeiy
C ii
xxexx,
which proves (1).
Noting that [θ(eij), eyy] = [eij, θ(eyy)] for any i < j and y ∈ X, multiplying this
identity by exx on the left and by eyy on the right, we have
C ij
xy = 0, if i 6= x < y 6= j.
(6)
Taking the identity (1) into account, then left multiplication by exx on [θ(eij ), ejj ] =
[eij, θ(ejj )] leads to
Similarly, right multiplication by eyy on [θ(eij ), eii] = [eij, θ(eii)] leads to
C ij
xj = 0, if i 6= x < j.
C ij
iy = 0, if i < y 6= j.
Combining the identities (6), (7) and (8), we obtain
θ(eij ) = X
x6y
xyexy = X
C ij
x∈X
xxexx + X
C ij
x<y
C ij
xyexy
(7)
(8)
= X
x∈X
= X
x∈X
xxexx + X
C ij
i6=x<y6=j
xyexy + X
C ij
i6=x<j
xjexj + X
C ij
i<y6=j
C ij
iyeiy + C ij
ij eij
C ij
xxexx + C ij
ij eij,
which proves (2).
(cid:3)
COMMUTING MAPS ON CERTAIN INCIDENCE ALGEBRAS
5
Lemma 2.7. Let X be connected and θ : I(X, R) → I(X, R) be an R-linear map
satisfying the formulas (1) and (2). Then θ is a commuting map if and only if the
coefficients C ij
xy are subject to the following relations:
(R1) C il
(R2) C ki
(R3) C ii
(R4) C ij
(R5) C ij
il = C ii
ii − C ii
ll ,
ki = C ii
ii − C ii
kk,
kk = C ii
ll ,
xx = C ij
yy,
ij = C jl
jl ,
if i < l;
if k < i;
if k < l and k 6= i 6= l;
∀ x, y ∈ X;
if i < j < l.
Proof. By Lemma 2.2, we have θ is a commuting map if and only if [θ(eij ), ekl] =
[eij, θ(ekl)] for all i 6 j and k 6 l. There are four cases occurring depending on the
formulas (1) or (2).
Case 1. If i = j and k = l, then [θ(eii), ekk] = [eii, θ(ekk)] always holds.
Case 2. If i = j and k 6= l, then the formulas (1) and (2) imply that
(C ii
kk − C ii
ll )ekl = (δik − δil)C kl
kl ekl.
(9)
Noting that k 6= l, hence we can get (R1) (resp. (R2)) from the identity (9) by
setting i = k (resp. i = l). We can also get the relation (R3) from (9) by setting
k 6= i 6= l.
Case 3. If i 6= j and k = l, this case is symmetric to the Case 2.
Case 4. If i 6= j and k 6= l, it follows from the formula (2) that
(C ij
kk − C ij
ll )ekl + δjk(C ij
ij − C kl
kl )eil = (C kl
jj − C kl
ii )eij + δil(C ij
ij − C kl
kl )ekj .
When j 6= k and i 6= l, the identity (10) is equivalent to
C ij
kk = C ij
ll
for j 6= k < l 6= i.
(10)
(11)
When j 6= k and i = l (hence k < i < j), the identity (10) can be rewritten as
(C ij
ki )ekj , which gives
ii )eki = (C ki
jj − C ki
ij − C ki
kk − C ij
ii )eij + (C ij
kk = C ij
ii = C ki
C ki
C ij
jj
ii
for j 6= k < i;
for i < j 6= k;
(12)
(13)
(14)
(15)
For convenience, substituting the indices we here rewrite the identity (13) as
C ki
ki = C ij
ij
for j 6= k < i < j.
C ij
jj = C ij
ll
for j < l 6= i.
When j = k and i 6= l, by symmetry we can also get the identities (12)-(15).
When j = k and i = l (hence i < j < i), the identity (10) is equivalent to
(C ij
ji )ejj , which gives
ji )eii = (C ji
ii )eij + (C ij
ii )eji + (C ij
jj − C ji
ij − C ji
ij − C ji
jj − C ij
C ij
jj = C ij
ii
for j < i;
C ij
ij = C ji
ji
for i < j < i.
(16)
(17)
Combining (14) and (17), we obtain the relation (R5). The identities (11), (12),
(15) and (16) imply that
C ij
xx = C ij
yy for all x < y.
(18)
6
HONGYU JIA AND ZHANKUI XIAO
Since X is connected, for any x, y ∈ X there is a sequence x = x0, x1, . . . , xs = y
in X such that xi−1 covers or is covered by xi for 1 ≤ i ≤ s. Then a recursive
procedure, using (18), on the length s implies the desired relation (R4).
(cid:3)
Remark 2.8. The reader may find that the set of the relations (R1-R5) in Lemma
2.7 is not minimal. For example, the relations (R1), (R2) and (R5) can imply the
relation (R3). However, we present here the Lemma 2.7 for later use.
Proof of Theorem 2.5. Let θ be an arbitrary commuting map of the incidence alge-
bra I(X, R). Then θ has the form (1) and (2) in Lemma 2.6, where the coefficients
C ij
xy satisfy the relations (R1-R5) in Lemma 2.7. If X = 1, the incidence algebra
I(X, R) ∼= R, and hence Theorem 2.5 is clear. We now assume X ≥ 2 and first
study the sufficiency.
We claim that the relation (R3) can be strengthened as C ii
yy for x 6= i 6= y.
Since the set of all directed edges associated with X (is assumed connected) forms
one equivalent class under the relation ∼=, there is a circle contains both x and
y. Breaking the circle at x and y, there exists a sequence x = x1, x2, . . . , xs = y
of different vertices in X such that x1 ∼ x2, . . . , xi ∼ xi+1, . . . , xs−1 ∼ xs and
i /∈ {x1, x2, . . . , xs}. A recursive procedure, using (R3), on the length s implies the
desired claim.
xx = C ii
We then claim that C ij
il + C ii
ij + C ii
jj = C ii
ii = C il
C ij
kl for all i < j and k < l. By (R1), we have
ll for j 6= i 6= l. Hence
ij = C kl
ll and we just proved that C ii
C ij
for i < j and i < l.
jj = C ii
(19)
ij = C il
il
Similarly, we have
C ij
ij = C kj
(20)
i.e.,
Since X is connected, there is a path (forget the direction) form j to k,
there exists a series j = x1, x2, . . . , xs = k in X such that x1 ∼ x2, . . . , xi ∼
xi+1, . . . , xs−1 ∼ xs. Using (19), (20) and the relation (R5), an induction on the
index s implies the desired claim.
for i < j and k < j.
kj
For any i ∈ X, since X is connected and X ≥ 2, there is a vertex comparable
ij eij for
ki eii
kl for all i < j and
ij for
with i. Let's define an R-linear map L : I(X, R) → I(X, R) by L(eij) = C ij
i < j and L(eii) = C ij
if there exists a vertex k such that k < i. Note that C ij
k < l. The map L is well-defined and is of the form L(f ) = λf , where λ = C ij
i < j and f ∈ I(X, R).
ij eii if there exists a vertex j such that i < j, or L(eii) = C ki
ij = C kl
ij )eii. Combining the fact C ii
Let µ := θ−L. If there is a vertex j such that i < j, then µ(eii) = Px6=i C ii
ii − C ij
xxexx+
(C ii
yy for x 6= i 6= y and the relation
(R1), we get µ(eii) ∈ Z(I(X, R)), the center of I(X, R). Similarly, we can get
µ(eii) ∈ Z(I(X, R)) if there is a vertex k such that k < i. Finally the relation (R4)
implies µ(eij ) ∈ Z(I(X, R)) for all i < j. Hence µ is a central-valued linear map
and θ is proper.
xx = C ii
We now prove the necessity. By the hypothesis, θ is proper. Hence C ij
ij = C kl
kl
for all i < j and k < l. For any i ∈ X, since X is connected and X ≥ 2, there
ij = C ii
is a vertex comparable with i. Therefore, either C ii
xx for x 6= i < j
or C ii
xx for k < i 6= x. We can get C ii
ki = C ii
yy for x 6= i 6= y from
each case.
If the set of all directed edges associated with X does not form one
equivalent class under the relation ∼=, then there exist two vertices j and k such
ii − C ij
xx = C ii
ii − C ki
COMMUTING MAPS ON CERTAIN INCIDENCE ALGEBRAS
7
that they can not be contained in any circle. In this case, we must have X ≥ 3.
In fact, when X = 2, every commuting map of I(X, R) is proper and the set of all
directed edges associated with X forms one equivalent class. Fix one vertex i which
is different from j and k. The connectivity of X shows that there is a path (forget
the direction) form i to j (resp. a path from i to k). By (R3), we can assume j
and k are comparable to i without lose of generality. Noting that j and k do not
be contained in any circle, we can construct, by Lemmas 2.6 and 2.7, a commuting
map θ such that C ii
kk, a contradiction. We completes the proof of Theorem
2.5.
(cid:3)
jj 6= C ii
From Theorem 2.5, we immediately have the following corollaries.
Corollary 2.9. Let R be a 2-torsion free commutative ring with unity and Tn(R)
be the upper (or lower) triangular matrix algebra over R. Then every commuting
map of Tn(R) is proper.
Corollary 2.10. Let R be a 2-torsion free commutative ring with unity and Mn(R)
be the full matrix algebra over R. Then every commuting map of Mn(R) is proper.
3. The general case
In this section, we extend the Theorem 2.5 to the case of X being a locally
finite pre-ordered set. Let I(X, R) be the R-subspace of I(X, R) generated by the
elements exy with x 6 y. That means I(X, R) consists exactly of the functions
f ∈ I(X, R) which are nonzero only at a finite number of (x, y). Clearly I(X, R)
is a subalgebra of I(X, R). Hence I(X, R) becomes an I(X, R)-bimodule in the
natural manner. Let θ : I(X, R) → I(X, R) be a commuting map, i.e.
[θ(f ), f ] = 0
for all f ∈ I(X, R). Observe that Lemmas 2.6 and 2.7 remain valid, when we
replace the domain of θ by I(X, R). In fact, although the formal sums L(eij) =
Px6y C ij
xyexy are now infinite, multiplication by euv on the left or on the right
works as in the finite case.
Let's now recall some notations and results from [18]. For any f ∈ I(X, R) and
x 6 y, the restriction of f to {z ∈ X x 6 z 6 y} is defined by
x = X
f y
x6u6v6y
f (u, v)euv
Observe that the sum above is finite, and hence f y
and x 6 y, the observation
x ∈ I(X, R). For any f ∈ I(X, R)
exxf eyy = f (x, y)exy
(21)
will be frequently used.
Lemma 3.1. ([18, Lemma 3.2 (ii)]) For any f ∈ I(X, R) and x 6 y, we have
(f g)(x, y) = (f y
xg)(x, y) = (f gy
x)(x, y) = (f y
xgy
x)(x, y).
Lemma 3.2. Let θ be a commuting map of I(X, R) and x < y. Then
θ(f )(x, y) = θ(f y
x)(x, y).
8
HONGYU JIA AND ZHANKUI XIAO
Proof. It follows from (21) that
θ(f )(x, y) = [exx, θ(f )](x, y) = [θ(exx), f ](x, y)
= [θ(exx), f y
x](x, y) = θ(f y
x)(x, y),
where the third identity relies on Lemma 3.1.
(cid:3)
For the locally finite pre-ordered set X, we set D := {(x, y) x < y} and also
define the equivalent relation ∼= on D as those in Definition 2.3.
Theorem 3.3. Let θ be a commuting map of I(X, R). If D forms one equivalent
class under the relation ∼=, then θ is proper.
Proof. Restricting θ to I(X, R), denoting as θ I , we get that θ I : I(X, R) → I(X, R)
is a commuting map. Then θ I is proper by Theorem 2.5. By the hypothesis, X is
connected and hence Z(I(X, R)) ∼= R by [12, Corollary 1.3.15]. There exists λ ∈ R
such that θ I (f ) = λf + µ(f ), where f ∈ I(X, R) and µ : I(X, R) → Z(I(X, R)) is
an R-linear map.
Define µ : I(X, R) → I(X, R) by µ(f ) := θ(f ) − λf for all f ∈ I(X, R). We only
need to prove the commuting map µ is central-valued. Notice that µ(f ) = µ(f ) ∈
Z(I(X, R)) for all f ∈ I(X, R). For x < y and f ∈ I(X, R), Lemma 3.2 shows
µ(f )(x, y) = µ(f y
x)(x, y) = 0. We now prove µ(f )(x, x) = µ(f )(y, y)
for all f ∈ I(X, R) and x, y ∈ X. Since X is connected, we only need to prove
µ(f )(x, x) = µ(f )(y, y) for x < y. On one hand, by (21),
x)(x, y) = µ(f y
[exy, µ(f )](x, y) = µ(f )(y, y) − µ(f )(x, x).
On the other hand, by Lemma 3.1,
[exy, µ(f )](x, y) = [µ(exy), f ](x, y) = [µ(exy), f y
x)(x, x)
x)](x, y) = µ(f y
x)(y, y) − µ(f y
x](x, y)
= [exy, µ(f y
= µ(f y
x)(y, y) − µ(f y
x)(x, x) = 0.
Therefore, it follows from [12, Theorem 1.3.13] that µ is central-valued and hence
θ is proper.
(cid:3)
Corollary 3.4. Let R be a 2-torsion free commutative ring with unity. Let T∞(R)
be the ring of countable upper triangular R-matrices. Then every commuting map
of T∞(R) is proper.
It is obvious that Theorem 3.3 can be generalized to the case of X consisting of a
finite number of connected components. Hence the following conjecture is natural.
Conjecture 3.5. Let (X, 6) be a locally finite pre-ordered set and R be 2-torsion
free. If the set D associated with each connected component forms one equivalent
class under the relation ∼=, then every commuting map of I(X, R) is proper.
References
[1] M. Bresar, Centralizing mappings on von Neumann algebras, Proc. Amer. Math. Soc. 111
(1991), 501-510.
[2] M. Bresar, Centralizing mappings and derivations in prime rings, J. Algebra 156 (1993),
385-394.
[3] M. Bresar, Commuting traces of biadditive mappings, commutativity-preserving mappings and
Lie mappings, Trans. Amer. Math. Soc. 335 (1993), 525-546.
[4] M. Bresar, Commuting maps: a survey, Taiwanese J. Math. 8 (2004), 361-397.
COMMUTING MAPS ON CERTAIN INCIDENCE ALGEBRAS
9
[5] L. Chen and J. Cai, Commuting mappings on the Hochschild extension of an algebra, Bull.
Iran. Math. Soc. 44 (2018), 749-762.
[6] W.S. Cheung, Commuting maps of triangular algebras, J. London Math. Soc. 63 (2001), 117-
127.
[7] I.N. Herstein, Lie and Jordan structures in simple, associative rings, Bull. Amer. Math. Soc.
67 (1961), 517-531.
[8] M. Khrypchenko, Jordan derivations of finitary incidence rings, Linear Multilinear Algebra
64 (2016), 2104-2118.
[9] T.K. Lee, Derivations and centralizing mappings in prime rings, Taiwanese J. Math. 1 (1997),
333-342.
[10] T.K. Lee and T.C. Lee, Commuting additive mappings in semiprime rings, Bull. Inst. Math.
Acad. Sinica 24 (1996), 259-268.
[11] E.C. Posner, Derivations in prime rings, Proc. Amer. Math. Soc. 8 (1957), 1093-1100.
[12] E. Spiegel and C. O'Donnell, Incidence algebras, Monographs and Textbooks in Pure and
Applied Mathematics, vol.206, Marcel Dekker, New York, 1997.
[13] R. Stanley, Enumerative Combinatorics, vol. 1. With a foreword by Gian-Carlo Rota. Cam-
bridge Studies in Advanced Mathematics, vol. 49, Cambridge University Press, Cambridge, 1997.
[14] M. Ward, Arithmetic functions on rings, Ann. of Math. 38 (1937), 725-732.
[15] D.N. Wang and Z.K. Xiao, Lie triple derivations of incidence algebras, Comm. Algebra, to
apper.
[16] Z.K. Xiao, Jordan derivations of incidence algebras, Rocky Mountain J. Math. 45 (2015),
1357-1368.
[17] Z.K. Xiao and F. Wei, Commuting mappings of generalized matrix algebras, Linear Algebra
Appl. 433 (2010), 2178-2197.
[18] X. Zhang and M. Khrypchenko, Lie derivations of incidence algebras, Linear Algebra Appl.
513 (2017), 69-83.
Jia: School of Mathematical Sciences, Huaqiao University, Quanzhou, Fujian, 362021,
P. R. China
E-mail address: [email protected]
Xiao: Fujian Province University Key Laboratory of Computation Science, School
of Mathematical Sciences, Huaqiao University, Quanzhou, Fujian, 362021, P. R. China
E-mail address: [email protected]
|
1708.07758 | 1 | 1708 | 2017-08-24T17:37:35 | Degenerations of Jordan Superalgebras | [
"math.RA"
] | We describe degenerations of three-dimensional Jordan superalgebras over $\mathbb{C}$. In particular, we describe all irreducible components in the corresponding varieties. | math.RA | math |
Degenerations of Jordan Superalgebras
Mar´ıa Alejandra Alvareza, Isabel Hern´andezb, Ivan Kaygorodovc
a Departamento de Matem´aticas, Universidad de Antofagasta, Chile.
b CONACYT - Centro de Investigaci´on en Matem´aticas, A.C. Unidad M´erida, M´exico.
c Universidade Federal do ABC, CMCC, Santo Andr´e, Brazil.
E-mail addresses:
Mar´ıa Alejandra Alvarez ([email protected]),
Isabel Hern´andez ([email protected]),
Ivan Kaygorodov ([email protected]),
Abstract. We describe degenerations of three-dimensional Jordan superalgebras over C. In particular,
we describe all irreducible components in the corresponding varieties.
Keywords: Jordan superalgebra, orbit closure, degeneration, rigid superalgebra
1. INTRODUCTION
Contractions of Lie algebras are limiting processes between Lie algebras, which have been studied first
in physics [20, 34]. For example, classical mechanics is a limiting case of quantum mechanics as → 0,
described by a contraction of the Heisenberg-Weyl Lie algebras to the Abelian Lie algebra of the same
dimension. Description of contractions of low dimensional Lie algebras was given in [32]. The study of
contractions and graded contractions of binary algebras has a very big history (see, for example, [13,18,36]).
The study of graded contractions of Jordan algebras and Jordan superalgebras was iniciated in [23]. The
first attempt of the study of contractions of n-ary algebras stays in the variety of Filippov algebras [5].
In mathematics, often a more general definition of contraction is used, the so-called degeneration. De-
generations are related to deformations. Degenerations of algebras is an interesting subject, which was
studied in various papers (see, for example, [6–8, 10, 11, 16, 17, 29, 33]). In particular, there are many results
concerning degenerations of algebras of low dimensions in a variety defined by a set of identities. One
of the important problems in this direction is the description of the so-called rigid algebras [22]. These
algebras are of big interest, since the closures of their orbits under the action of generalized linear group
form irreducible components of a variety under consideration (with respect to the Zariski topology). There
are fewer works in which the full information about degenerations was found for some variety of alge-
bras. This problem was solved for two-dimensional pre-Lie algebras in [6], for two-dimensional Jordan
algebras in [2], for three-dimensional Novikov algebras in [7], for three-dimensional Jordan algebras [15],
for four-dimensional Lie algebras in [8], for nilpotent four-dimensional Jordan algebras [3], for nilpotent
four-dimensional Leibniz and Zinbiel algebras in [27], for nilpotent five- and six-dimensional Lie algebras
in [16, 33], for nilpotent five- and six-dimensional Malcev algebras in [26], and for all 2-dimensional alge-
bras [28]. On the same time, the study of degenerations of superalgebras and graded algebras was iniciated
in [4] for associative case and in [1] for Lie superalgebras.
1
2
2. DEFINITIONS AND NOTATION
2.1. Jordan superalgebras. Jordan algebras appeared as a tool for studies in quantum mechanic in the
paper of Jordan, von Neumann and Wigner [19]. A commutative algebra is called a Jordan algebra if it
satisfies the identity
(x2y)x = x2(yx).
The study of the structure theory and other properties of Jordan algebras was initiated by Albert. Jordan
algebras are related with some questions in differentional equations [35], superstring theory [12], analysis,
operator theory, geometry, mathematical biology, mathematical statistics and physics (see, the survey of
Iordanescu [21]).
Let G be the Grassmann algebra over F given by the generators 1, ξ1, . . . , ξn, . . . and the defining relations
ξ2
i = 0 and ξiξj = −ξjξi. The elements 1, ξi1ξi2 . . . ξik, i1 < i2 < . . . < ik, form a basis of the algebra G
over F. Denote by G0 and G1 the subspaces spanned by the products of even and odd lengths, respectively;
then G can be represented as the direct sum of these subspaces, G = G0 ⊕ G1. Here the relations GiGj ⊆
Gi+j(mod 2), i, j = 0, 1, hold. In other words, G is a Z2-graded algebra (or a superalgebra) over F. Suppose
now that A = A0 ⊕A1 is an arbitrary superalgebra over F. Consider the tensor product G⊗A of F-algebras.
The subalgebra
G(A) = G0 ⊗ A0 + G1 ⊗ A1
of G ⊗ A is referred to as the Grassmann envelope of the superalgebra A. Let Ω be a variety of algebras
over F. A superalgebra A = A0 ⊕ A1 is referred to as an Ω-superalgebra if its Grassmann envelope G(A)
is an algebra in Ω. In particular, A = A0 ⊕ A1 is referred to as a Jordan superalgebra if its Grassmann
envelope G(A) is a Jordan algebra. The study of Jordan superalgebras has very big history (for example,
see [9, 14, 24, 25, 30, 31]).
2.2. Degenerations. Given an (m, n)-dimensional vector superspace V = V0 ⊕ V1, the set
Hom(V ⊗ V, V ) = (Hom(V ⊗ V, V ))0 ⊕ (Hom(V ⊗ V, V ))1
is a vector superspace of dimension m3 + 3mn2. This space has a structure of the affine variety Cm3+3mn2
.
If we fix a basis {e1, . . . , em, f1, . . . , fn} of V , then any µ ∈ Hom(V ⊗ V, V ) is determined by m3 + 3mn2
structure constants αk
p,q ∈ C such that
i,j, βq
i,j, γq
i,j, δk
µ(ei ⊗ ej) =
m
X
k=1
αk
i,jek, µ(ei ⊗ fp) =
n
X
q=1
βq
i,pfq, µ(fp ⊗ ei) =
n
X
q=1
γq
p,ifq, µ(fp ⊗ fq) =
m
X
k=1
δk
p,qek.
A subset L(T ) of Hom(V ⊗ V, V ) is Zariski-closed if it can be defined by a set of polynomial equations T
in the variables αk
p,q (1 ≤ i, j, k ≤ m, 1 ≤ p, q ≤ n).
i,j, βq
i,p, γq
p,i, δk
Let Sm,n be the set of all superalgebras of dimension (m, n) defined by the family of polinomial super-
identities T , understood as a subset L(T ) of an affine variety Hom(V ⊗ V, V ). Then one can see that Sm,n
is a Zariski-closed subset of the variety Hom(V ⊗ V, V ). The group G = (Aut V )0 ≃ GL(V0) ⊕ GL(V1)
acts on Sm,n by conjugations:
(g ∗ µ)(x ⊗ y) = gµ(g−1x ⊗ g−1y)
for x, y ∈ V , µ ∈ L(T ) and g ∈ G.
Thus, Sm,n is decomposed into G-orbits that correspond to the isomorphism classes of superalgebras. Let
O(µ) denote the orbit of µ ∈ L(T ) under the action of G and O(µ) denote the Zariski closure of O(µ).
Let J, J ′ ∈ Sm,n and λ, µ ∈ L(T ) represent J and J ′ respectively. We say that λ degenerates to µ and
write λ → µ if µ ∈ O(λ). Note that in this case we have O(µ) ⊂ O(λ). Hence, the definition of a
degeneration does not depend on the choice of µ and λ, and we will right indistinctly J → J ′ instead of
λ → µ and O(J) instead of O(λ). If J 6∼= J ′, then the assertion J → J ′ is called a proper degeneration. We
write J 6→ J ′ if J ′ 6∈ O(J).
Let J be represented by λ ∈ L(T ). Then J is rigid in L(T ) if O(λ) is an open subset of L(T ). Recall
that a subset of a variety is called irreducible if it cannot be represented as a union of two non-trivial closed
subsets. A maximal irreducible closed subset of a variety is called irreducible component. In particular, J
3
is rigid in Sm,n iff O(λ) is an irreducible component of L(T ). It is well known that any affine variety can
be represented as a finite union of its irreducible components in a unique way. We denote by Rig(Sm,n) the
set of rigid superalgebras in Sm,n.
2.3. Principal notations. Let JSm,n be the set of all Jordan superalgebras of dimension (m, n). Let J be a
Jordan superalgebra with fixed basis {e1, . . . , em, f1, . . . fn}, defined by
m
eiej =
X
αk
ijek,
eifj =
k=1
We will use the following notation:
n
X
k=1
βk
ijfk,
fifj =
m
X
k=1
γk
ijek.
(1) a(J) is the Jordan superalgebra with the same underlying vector superspace than J , and defined by
n
fifj =
X
γk
ijek.
(2) J 1 = J , J r = J r−1J + J r−2J 2 + · · · + JJ r−1, and in every case J r = (J r)0 ⊕ (J r)1.
k=1
tr(L(x)i) · tr(L(y)j)
tr(L(x)i · L(y)j)
(3) ci,j =
is the Burde invariant, where L(x) is the left multiplication. This
invariant ci,j is defined as a quotient of two polynomials in the structure constants of J , for all
x, y ∈ J such that both polynomials are not zero and ci,j is independent of the choice of x, y.
3. METHODS
First of all, if J → J ′ and J 6∼= J ′, then dim Aut(J) < dim Aut(J ′), where Aut(J) is the space of
automorphisms of J . Secondly, if J → J ′ and J ′ → J ′′ then J → J ′′. If there is no J ′ such that J → J ′
and J ′ → J ′′ are proper degenerations, then the assertion J → J ′′ is called a primary degeneration. If
dim Aut(J) < dim Aut(J ′′) and there are no J ′ and J ′′′ such that J ′ → J , J ′′ → J ′′′, J ′
6→ J ′′′ and one
of the assertions J ′ → J and J ′′ → J ′′′ is a proper degeneration, then the assertion J 6→ J ′′ is called
a primary non-degeneration.
It suffices to prove only primary degenerations and non-degenerations to
describe degenerations in the variety under consideration. It is easy to see that any superalgebra degenerates
to the superalgebra with zero multiplication. From now on we will use this fact without mentioning it.
Let us describe the methods for proving primary non-degenerations. The main tool for this is the follow-
ing lemma.
Lemma 1. If J → J ′ then the following hold:
(1) dim(J r)i ≥ dim(J ′r)i, for i ∈ Z2;
(2) (J)0 → (J ′)0;
(3) a(J) → a(J ′);
(4) If the Burde invariant exist for J and J ′, then both superalgebras have the same Burde invariant;
(5) If J is associative then J ′ must be associative. In fact, if J satisfies a P.I. then J ′ must satisfy the
same P.I.
In the cases where all of these criteria can't be applied to prove J 6→ J ′, we will define R by a set of
polynomial equations and will give a basis of V , in which the structure constants of λ give a solution to all
these equations. We will omit everywhere the verification of the fact that R is stable under the action of
the subgroup of upper triangular matrices and of the fact that µ 6∈ R for any choice of a basis of V . These
verifications can be done by direct calculations.
Degenerations of Graded algebras. Let G be an abelian group and let V(F) be a variety of algebras
defined by a family of polynomial identities F. It is important to notice that degeneration on the G-graded
variety GV(F) is a more restrictive notion than degeneration on the variety V(F), In fact, consider A, A′ ∈
GV(F) such that A, A′ ∈ V(F), a degeneration between the algebras A and A′ may not give rise to a
degeneration between the G-graded algebras A and A′, since the matrices describing the basis changes in
GV(F) must preserve the G-graduation. Hence, we have the following result.
Lemma 2. Let A, A′ ∈ GV(F) ∩ V(F). If A 6→ A′ as algebras, then A 6→ A′ as G-graded algebras.
4
4. MAIN RESULT
In this section we describe all degenerations and non-degenerations of 3-dimensional Jordan superalge-
bras. Note that, there is only one (trivial) Jordan superalgebra of the type (0, 3); the variety of 3-dimensional
Jordan algebras (Jordan superalgebras of the type (3, 0)) has 19 non isomorphic algebras with non zero
multiplication, in particularly, 5 algebras are rigid. The full description of all degenerations and non-
degenerations of 3-dimensional Jordan algebras was given in [15]. The rest of the section is dedicated to
study of degenerations of Jordan superalgebras of types (1, 2) and (2, 1).
4.1. Jordan Superalgebras of type (1, 2).
4.1.1. Algebraic classification. As was noticed the algebraic classification of Jordan superalgebras of the
type (1, 2) was received in [30]. In the next table we give this classification with some additional useful
information about these superalgebras.
Table 1. (1, 2)-Dimensional Jordan superalgebras.
J multiplication tables
U s
1
S2
1
S3
1
S2
2
S3
2
S3
3
e1e1 = e1
e1e1 = e1, e1f1 = 1
2 f1
e1f1 = f2, f1f2 = e1
e1e1 = e1, e1f1 = f1
f1f2 = e1
e1f1 = f2
S3
4
S3
5
S3
6
S3
7
S3
8
e1e1 = e1, e1f1 = f1, e1f2 = 1
2 f2
e1e1 = e1, e1f1 = 1
2 f1, e1f2 = 1
e1e1 = e1, e1f1 = f1, e1f2 = f2
2 f2
e1e1 = e1, e1f1 = 1
2 f1, e1f2 = 1
2 f2, f1f2 = e1
e1e1 = e1, e1f1 = f1, e1f2 = f2, f1f2 = e1
dim Aut(J)
4
2
2
2
4
3
2
4
4
3
3
ci,j
1
2
6 ∃
2
6 ∃
6 ∃
(cid:16)2+( 1
(cid:16)1+2( 1
2 )j(cid:17)
2)j(cid:17)
2 )i(cid:17)(cid:16)2+( 1
2 )i+j(cid:17)
(cid:16)2+( 1
2 )i(cid:17)(cid:16)1+2( 1
2)i+j(cid:17)
(cid:16)1+2( 1
3
2 )i(cid:17)(cid:16)1+2( 1
2)i+j(cid:17)
(cid:16)1+2( 1
3
(cid:16)1+2( 1
2)j(cid:17)
type
associative
non-associative
non-associative
associative
associative
associative
non-associative
non-associative
associative
non-associative
non-associative
4.1.2. Degenerations.
Theorem 3. The graph of primary degenerations for Jordan superalgebras of dimension (1, 2) has the
following form:
S 3
4
S 3
3
S 3
6
S 2
2
U s
1
S 3
1
S 3
7
S 2
1
S 3
8
S 3
5
S 3
2
C1,2
Proof. We prove all required primary degenerations in Table 2 below. Recall that an associative superalge-
bra can only degenerate to an associative superalgebra. Let us consider the first degeneration S2
3 to
clarify this table. Write nonzero products in S2
1 → S3
1 in the basis Et
1, F t
1, F t
2:
Et
1Et
1 = tEt
1, Et
1F t
1 = te1(f1 − 2t−1f2) =
t
2
(f1 − 2t−1f2) + f2 =
t
2
1 + F t
F t
2.
It is easy to see that for t = 0 we obtain the multiplication table of S3
considered in the same way.
3 . The remaining degenerations can be
Table 2. Primary degenerations of Jordan superalgebras of dimension (1, 2).
degenerations
parametrized bases
5
3
3
S 2
S 2
S 3
S 3
S 3
S 3
S 3
S 3
S 3
1 → S 3
2 → S 3
1 → S 3
1 → S 3
4 → S 3
7 → S 3
7 → S 3
8 → S 3
8 → S 3
5
6
3
2
3
2
2
Et
Et
Et
Et
Et
Et
Et
Et
Et
1 = te1, F t
1 = te1, F t
F t
1 = e1,
F t
1 = e1,
1 = te1, F t
1 = te1, F t
F t
1 = e1,
1 = te1, F t
F t
1 = e1,
1 = f1 − 2t−1f2, F t
1 = f1 − t−1f2,
F t
F t
1 = tf1,
F t
1 = tf1,
1 = f1 − 2t−1f2, F t
F t
1 = tf1,
F t
1 = f1,
F t
1 = tf1,
F t
1 = tf1,
2 = f2
2 = f2
2 = tf2
2 = tf2
2 = f2
2 = f2
2 = tf2
2 = f2
2 = f2
The primary non-degenerations are proved in Table 3.
Table 3. Primary non-degenerations of Jordan superalgebras of dimension (1, 2).
2
non-degenerations
S 3
S 3
S 3
S 3
S 2
3 6→ S 3
3 6→ U s
7 6→ U s
4 6→ S 3
1 6→ S 3
1 , S 3
1 , S 3
7 , S 3
2 ; S 3
6 ; S 3
6 ; S 3
8 , U s
1 , S 3
4 6→ S 3
8 , U s
3 ; S 2
1 6→ S 3
8 6→ U s
5 , S 3
2 ; S 2
7 , S 3
1 , S 5
6; S 2
2 6→ S 3
2 ;
1 , S 3
6
5 , S 3
1 6→ S 3
1 , S 3
6
2 6→ U s
7 , S 3
8 , U s
1 , S 3
5 , S 3
6 ;
reason
dim(J 2)0 < dim(J ′2)0
J0 6→ J ′
0
ci,j
a(J) 6→ a(J ′)
✷
4.1.3. Irreducible components and rigid algebras. Using Theorem 3, we describe the irreducible compo-
nents and the rigid algebras in JS1,2.
Corollary 4. The irreducible components of JS1,2 are:
C1 = O(U s
C2 = O(S2
C3 = O(S3
C4 = O(S2
C5 = O(S3
C6 = O(S3
C7 = O(S3
1 ) = {U s
1) = {S2
1) = {S3
2) = {S2
4) = {S3
7) = {S3
8) = {S3
1 , C1,2} :
1 , S3
1 , S3
2 , S3
4 , S3
5 , S3
8 , S3
3 , C1,2};
3 , C1,2};
3 , S3
3 , C1,2};
2 , C1,2};
2 , S3
2 , C1,2};
6 , C1,2};
In particular, Rig(JS1,2) = {U s
1 , S2
1 , S3
1 , S2
2 , S3
4 , S3
7 , S3
8}.
4.2. Superalgebras of type (2, 1). In this section we describe all possible primary degenerations between
Jordan superalgebras of dimension (2, 1). First of all, notice that every Jordan superalgebra of dimension
(2, 1) has trivial odd product, so it can be considered as a Jordan algebra of dimension 3. However, the graph
of primary degenerations of Jordan superalgebras of dimension (2, 1) is not a subgraph of the primary
degenerations of de Jordan algebras of dimension 3.
In fact, there exist Jordan superalgebras, without
degenerations between them, such that they degenerate as Jordan algebras. Take for example S2
1 and
Bs
1 → Bs
1
as Jordan algebras. Notice that this basis change does not preserve the Z2-graduation. Moreover, we shall
prove that S2
1 (see Table 5), Taking the basis change: Et
1 = te2, it follows that S2
2 = f1 and F t
1 = e1, Et
2 ⊕ U s
2 ⊕ U s
1 as Jordan superalgebras.
1 6→ Bs
2 ⊕ U s
6
4.2.1. Algebraic classification. In the next table we provide the classification of (2, 1)-dimensional Jordan
superalgebras with some additional useful information about these superalgebras.
Table 4. (2, 1)-Dimensional Jordan superalgebras.
J
2U s
1
U s
1
Bs
1
Bs
2
Bs
3
S2
1 ⊕ U s
1
S2
1
S2
2 ⊕ U s
1
S2
2
S3
9
S3
10
S3
11
S3
12
S3
13
Multiplication tables
e1e1 = e1, e2e2 = e2
e1e1 = e1
e1e1 = e1, e1e2 = e2
e1e1 = e1, e1e2 = 1
2 e2
e1e1 = e2
e1e1 = e1, e2e2 = e2, e1f1 = 1
e1e1 = e1, e1f1 = 1
e1e1 = e1, e2e2 = e2, e1f1 = f1
e1e1 = e1, e1f1 = f1
e1e1 = e1, e1e2 = e2, e1f1 = 1
2 f1
e1e1 = e1, e1e2 = e2, e1f1 = f1
2 e2, e1f1 = 1
e1e1 = e1, e1e2 = 1
e1e1 = e1, e1e2 = 1
2 e2, e1f1 = f1
e1e1 = e1, e2e2 = e2, e1f1 = 1
2 f1
2 f1
2 f1
2 f1, e2f1 = 1
2 f1
dim Aut(J) Type
1
2
2
3
3
1
2
1
2
2
2
3
3
1
associative
associative
associative
non-associative
associative
non-associative
non-associative
associative
associative
non-associative
associative
non-associative
non-associative
non-associative
4.2.2. Degenerations.
Theorem 5. The graph of primary degenerations for Jordan superalgebras of dimension (2, 1) has the
following form:
S 2
1
⊕ U s
1
S 3
13
S 2
2
⊕ U s
1
2U s
1
S 3
9
S 2
1
S 3
10
S 3
12
U s
1
Bs
3
C2,1
S 2
2
Bs
1
S 3
11
Bs
2
Proof. We prove all required primary degenerations in Table 5 below. For clarifying this table was conside-
red an example in the proof of the theorem 3.
Table 5. Primary degenerations of Jordan superalgebras of dimension (2, 1).
7
parametrized bases
1 → U s
1 → S 2
1 → S 3
3
1
1
3
degenerations
2U s
2U s
U s
Bs
S 2
S 2
S 2
S 2
S 2
S 2
S 2
S 2
S 3
S 3
S 3
S 3
1 → U s
1 → Bs
1 → Bs
1 → Bs
1 ⊕ U s
1 ⊕ U s
1 ⊕ U s
1 → Bs
2 → Bs
2 ⊕ U s
2 ⊕ U s
2 ⊕ U s
9 → Bs
10 → Bs
13 → S 2
13 → S 3
10
3
3
3
3
1
1 → S 2
1 → U s
1 → S 3
Et
Et
Et
Et
1 Et
1 Et
9 Et
Et
Et
2 Et
1 Et
10 Et
Et
Et
Et
Et
1 = e1,
1 = e1 + e2,
1 = te1 + e2,
1 = te1 + e2,
1 = e2,
1 = e1,
1 = e1 + e2,
1 = te1 + e2,
1 = te1 + e2,
1 = e1,
1 = 1
1 = e1 + e2,
1 = te1 + e2,
1 = te1 − 1
1 = e1,
1 = e1 + e2,
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
2 e2, Et
Et
Et
F t
2 = te2,
F t
2 = te2,
2 = t2e1,
F t
2 = −t2e1,
F t
F t
2 = te1,
F t
2 = te2,
F t
2 = te2,
2 = −t2e1,
F t
2 = t2e1,
F t
F t
2 = te2,
F t
2 = e1,
F t
2 = te2,
F t
2 = te2,
2 = t2e1 − te2, F t
F t
2 = te2,
F t
2 = te2,
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
1 = f1
t e2,
Primary non-degenerations between 3-dimensional Jordan algebras are given in [15], we use this result
and Lemma 2 to prove some primary non-degenerations, (see table 6).
12, and S2
From [15] it follows that 2U s
1 → Bs
1 → S2
9 → S3
2 , S2
2, S3
2 ⊕ U s
1 → Bs
1 as Jordan algebras; we
shall prove that they do not degenerate as Jordan superalgebras.
1 as Jordan superalgebras, then there exists a parameterized basis
First, suppose that S2
1 → Bs
2 ⊕ U s
Et
1 = a(t)e1 + b(t)e2, Et
1 , such that for t = 0 we obtain Bs
1 = Et
2 ⊕ U 2
for S2
that a(0) = c(0) = 0. Now, since Et
since Et
S2
2 ⊕ U s
2Et
1 6→ Bs
1. For the remaining cases see table 6.
1Et
2 = 0 at t = 0, it follows that d(0) = 0, showing that Et
2 = c(t)e1 + d(t)e2, F t
1. Since Et
1 = a(t)F t
1, it follows
1 it follows that a(t) = 0 and b(t) = 1 for all t, Finally,
2 = 0, for all t. This proves that
1 = x(t)f1
2F t
1 and Et
1 = c(t)F t
1Et
1F t
Table 6. Primary non-degenerations of Jordan superalgebras of dimension (2, 1).
non-degenerations
2U s
S 2
S 3
2U s
S 2
1 , S 2
2; S 3
2, S 3
2; S 2
1 6→ Bs
1 6→ S 2
1 6→ Bs
13 6→ Bs
1 6→ Bs
2 , S 3
9 6→ S 3
11, S 3
1 6→ S 3
1, S 2
1 ⊕ U s
12
9 , S 3
10, S 3
12; S 2
11, S 3
1 ⊕ U s
12
1 6→ Bs
2, S 3
11, S 3
12;
11, S 3
12; S 3
9 6→ Bs
2 , S 3
10; S 2
2 ⊕ U s
2, S 3
1 6→ S 2
11;
1 , S 3
9; S 3
13 6→ U s
1 , Bs
1, S 2
2 , S 3
9
reason
dim(J r)1 < dim((J ′)r)1
J0 6→ J ′
0
J 6→ J ′ as Jordan algebras
✷
4.2.3. Irreducible components and rigid algebras. Using Theorem 5, we describe the irreducible compo-
nents and the rigid superalgebras in JS2,1.
Corollary 6. The irreducible components of JS2,1 are:
1 , Bs
1, Bs
3, C2,1};
3, S2
3, S2
1 ⊕ U s
2 ⊕ U s
1 , S2
1 , S2
1 , S3
2 , S3
9 , C2,1};
10, C2,1};
C1 = O(2U s
1 , U s
2, C2,1};
1 ) = {2U s
2) = {Bs
1 ⊕ U s
2 ⊕ U s
11) = {S3
12) = {S3
13) = {Bs
1 ) = {U s
1 ) = {U s
1 , Bs
1 , Bs
11, C2,1};
12, C2,1};
3, S2
1 , S3
C2 = O(Bs
C3 = O(S2
C4 = O(S2
C5 = O(S3
C6 = O(S3
C7 = O(S3
10, S3
13, C2,1}.
8
In particular, Rig(JS2,1) = {2U s
1 , Bs
2, S2
1 ⊕ U s
1 , S2
2 ⊕ U s
1 , S3
11, S3
12, S3
13}.
Acknowledgements. This work was started during the research stay of I. Kaygorodov at the Department
of Mathematics, parcially funded by the Coloquio de Matem´atica (CR 4430) of the University of Antofa-
gasta. The work was supported by RFBR 17-01-00258, by FOMIX-CONACYT YUC-2013-C14-221183
and 222870.
REFERENCES
[1] Alvarez M.A., Hern´andez I., On degenerations of Lie superalgebras, preprint.
[2] Ancochea Berm´udez J.M., Campoamor-Stursberg R., Garc´ıa Vergnolle L., S´anchez Hern´andez J., Con-
tractions d´algebres de Jordan en dimension 2, Journal of Algebra, 319 (2008), 6, 2395–2409.
[3] Ancochea Berm´udez J.M., Fres´an J., Margalef Bentabol J., Contractions of low-dimensional nilpotent
Jordan algebras, Communications in Algebra, 39 (2011), 3, 1139–1151.
[4] Armour A., Zhang Yi., Geometric classification of 4-dimensional superalgebras, Chapter Algebra,
Geometry and Mathematical Physics, 315 Springer Proceedings in Mathematics and Statistics, 85,
(2014), 291–323.
[5] de Azcarraga J., Izquierdo J., Picon M., Contractions of Filippov algebras, J. Math. Phys., 52 (2011),
1, 013516.
[6] Benes T., Burde D., Degenerations of pre-Lie algebras, J. Math. Phys., 50 (2009), 11, 112102, 9 pp.
[7] Benes T., Burde D., Classification of orbit closures in the variety of three-dimensional Novikov alge-
bras, J. Algebra Appl., 13 (2014), 2, 1350081, 33 pp.
[8] Burde D., Steinhoff C., Classification of orbit closures of 4-dimensional complex Lie algebras, Journal
of Algebra, 214 (1999), 2, 729–739.
[9] Cantarini N., Kac V., Classification of linearly compact simple Jordan and generalized Poisson super-
algebras, J. Algebra, 313 (2007), 100–124
[10] Casas J., Khudoyberdiyev A., Ladra M., Omirov B., On the degenerations of solvable Leibniz algebras,
Linear Algebra and its Applications, 439 (2013), 2, 472–487.
[11] Chouhy S., On geometric degenerations and Gerstenhaber formal deformations, arXiv:1708.02933
[12] Corrigan E., Hollowood T., The exceptional Jordan algebra and the superstring, Comm. Math. Phys.,
122 (1989), 3, 393–410.
[13] Couture M., Patera J., Sharp R. T., Winternitz P., Graded contractions of sl(3, C), J. Math. Phys., 32
(1991), 2310–2318.
[14] Elduque A., Laliena J., Sacrist´an S., Maximal subalgebras of Jordan superalgebras, J. of Pure and
Appl. Alg., 212 (2008), 11, 2461–2478.
[15] Gorshkov I., Kaygorodov I., Popov Yu., Degenerations of Jordan algebras, arXiv:1707.08836
[16] Grunewald F., O'Halloran J., Varieties of nilpotent Lie algebras of dimension less than six, J. Algebra,
112 (1988), 315–325.
[17] Grunewald F., O'Halloran J., A characterization of orbit closure and applications, J. Algebra, 116
(1988), 163–175.
[18] Hrivnak J., Novotny P., Patera J., Tolar J., Graded contractions of the Pauli graded sl(3, C), Linear
Algebra Appl., 418 (2006), 498–550.
[19] Jordan P., v. Neumann J., Wigner E., On an Algebraic Generalization of the Quantum Mechanical
Formalism, Annals of Mathematics (2), 35 (1934), 1, 29–64.
[20] Inonu E., Wigner E. P., On the contraction of groups and their representations, Proc. Natl. Acad. Sci.
U.S.A. 39 (1953), 510–524.
[21] Iordanescu R., Jordan structures in mathematics and physics, arXiv:1106.4415
[22] Ismailov N., Kaygorodov I., Volkov Yu., The geometric classification of Leibniz algebras,
arXiv:1705.04346
[23] Kashuba I., Patera J., Graded contractions of Jordan algebras and of their representations, Journal of
Physics A: Mathematical and General, 36 (2003), 50, 12453.
9
[24] Kaygorodov I., δ-derivations of simple finite-dimensional Jordan superalgebras, Algebra and Logic,
46 (2007), 5, 318–329.
[25] Kaygorodov I., δ-superderivations of semisimple finite-dimensional Jordan superalgebras, Mathemat-
ical Notes, 91 (2012), 2, 187–197.
[26] Kaygorodov I., Popov Yu., Volkov Yu., Degenerations of binary Lie and nilpotent Malcev algebras,
arXiv:1609.07392.
[27] Kaygorodov I., Popov Yu., Pozhidaev A., Volkov Yu., Degenerations of nilpotent Leibniz and Zinbiel
algebras, Linear and Multilinear algebra, 2018, arXiv:1611.06454
[28] Kaygorodov I., Volkov Yu., 2-dimensional algebras over algebraically closed fields, arXiv:1701.08233
[29] Lauret J., Degenerations of Lie algebras and geometry of Lie groups, Differential Geom. Appl., 18
(2003), 2, 177–194.
[30] Martin M.E.,
Shestakov
I., Classification
of
three-dimensional
Jordan
superalgebras,
arXiv:1708.01963
[31] Martinez C., Zelmanov E., Simple finite-dimensional Jordan superalgebras of prime characteristic, J.
Algebra, 236 (2001), 2, 575–629.
[32] Nesterenko M., Popovych R., Contractions of low-dimensional Lie algebras, J. Math. Phys., 47 (2006),
123515.
[33] Seeley C., Degenerations of 6-dimensional nilpotent Lie algebras over C, Comm. Algebra, 18 (1990),
3493–3505.
[34] Segal I. E., A class of operator algebras determined by groups, Duke Math. J., 18 (1951), 221–265.
[35] Svinolupov S., Jordan algebras and generalized Kortewegde Vries equations, Theoret. and Math.
Phys., 87 (1991), 3, 611–620.
[36] Weimar-Woods E., The three-dimensional real Lie algebras and their contractions, J. Math. Phys., 32
(1991), 2028–2033.
|
0805.2064 | 2 | 0805 | 2010-05-10T22:45:23 | KW-sections for exceptional type Vinberg's $\theta$-groups | [
"math.RA",
"math.GR"
] | Let $k$ be an algebraically closed field of characteristic not equal to 2 or 3, let $G$ be an almost simple algebraic group of type $F_4$, $G_2$ or $D_4$ and let $\theta$ be an automorphism of $G$ of finite order, coprime to the characteristic. In this paper we consider the $\theta$-group (in the sense of Vinberg) associated to these choices; we classify the positive rank automorphisms and give their Kac diagrams and we describe the little Weyl group in each case. As a result we show that all such $\theta$-groups have KW-sections, confirming a conjecture of Popov in these cases. | math.RA | math |
KW-sections for exceptional type Vinberg's θ-groups
Paul Levy
[email protected]
November 1, 2018
Abstract
Let k be an algebraically closed field of characteristic not equal to 2 or 3, let G be
an almost simple algebraic group of type F4, G2 or D4 and let θ be an automorphism
of G of finite order, coprime to the characteristic.
In this paper we consider the θ-
group (in the sense of Vinberg) associated to these choices; we classify the positive
rank automorphisms via Kac diagrams and we describe the little Weyl group in each
case. As a result we show that all θ-groups in types G2, F4 and D4 have KW-sections,
confirming a conjecture of Popov in these cases.
0
Introduction
Let G be a reductive algebraic group over the algebraically closed field k and let g = Lie(G).
Let θ be a semisimple automorphism of G of order m, let dθ be the differential of θ and let ζ
be a primitive m-th root of unity in k. (Thus if k is of positive characteristic p then p ∤ m.)
There is a direct sum decomposition
g = g(0) ⊕ . . . ⊕ g(m − 1) where g(i) = {x ∈ g dθ(x) = ζ ix}
This is a Z/mZ-grading of g, that is [g(i), g(j)] ⊂ g(i + j) (i, j ∈ Z/mZ). Let G(0) = (Gθ)◦.
Then G(0) is reductive, Lie(G(0)) = g(0) and G(0) stabilizes each component g(i). In [21],
Vinberg studied invariant-theoretic properties of the G(0)-representation g(1). The central
concept in [21] is that of a Cartan subspace, which is a subspace of g(1) which is maximal
subject to being commutative and consisting of semisimple elements. The principal results
of [21] (for k = C) are:
- any two Cartan subspaces of g(1) are G(0)-conjugate and any semisimple element of
g(1) is contained in a Cartan subspace.
- the G(0)-orbit through x ∈ g(1) is closed if and only if x is semisimple, and is unstable
(that is, its closure contains 0) if and only if x is nilpotent.
- let c be a Cartan subspace of g(1) and let Wc = NG(0)(c)/ZG(0)(c), the little Weyl group.
Then we have a version of the Chevalley restriction theorem: the embedding c ֒→ g(1)
induces an isomorphism k[g(1)]G(0) → k[c]Wc.
- Wc is a finite group generated by complex (often called pseudo-)reflections, hence k[c]Wc
is a polynomial ring.
In the case of an involution, the decomposition g = g(0) ⊕ g(1) is the symmetric space
decomposition, much studied since the seminal paper of Kostant and Rallis [10]. (Many of
1
the results of [10] were generalized to good positive characteristic by the author in [11].)
While the theory of θ-groups can in some ways be thought of as an extension of the theory
of symmetric spaces, there are certain differences of emphasis. Broadly speaking, one can
say that the results here on geometry and orbits are weaker than for symmetric spaces, but
the connection with groups generated by pseudoreflections is more interesting. Recall that
a KW-section is an affine linear subspace v ⊂ g(1) such that restricting to v induces an
isomorphism k[g(1)]G(0) → k[v]. A long-standing conjecture of Popov [16] is the existence of
a KW-section in g(1) for the action of G(0). In characteristic zero, this conjecture is known
to hold in the cases when G(0) is semisimple [14] and when g(1) contains a regular nilpotent
element of g (the 'N-regular' case) [15], both due to Panyushev.
In [12], the results of [21] were extended to the case where k has positive characteristic
p and G satisfies the standard hypotheses: (A) the derived subgroup G′ of G is simply-
connected, (B) p is good for G and (C) there exists a non-degenerate G-equivariant symmet-
ric bilinear form κ : g × g → k. (In fact, the first three results mentioned above hold for all
p > 2; the standard hypotheses are only required for the fourth.) Moreover, an analysis of
the little Weyl group and an extension of Panyushev's result on N-regular automorphisms
revealed that KW-sections exist for all classical graded Lie algebras in zero or odd positive
characteristic. This leaves Popov's conjecture open in the following cases: (i) G is of excep-
tional type, and (ii) G is simply-connected of type D4, p > 3 and θ is an outer automorphism
of G such that θ3 is inner. Following Vinberg, we refer to all such cases as exceptional type
θ-groups.
In characteristic zero the automorphisms of finite order were classified by Kac [8]. In Sect.
2 we show that Kac's classification is valid in characteristic p, restricting to automorphisms
of order coprime to p. (The extension of Kac's results to positive characteristic was already
outlined by Serre [18].) Subsequently, we determine (Sect. 4) the positive rank exceptional
θ-groups of types F4, G2 and D4 and describe (Sect. 5) the corresponding little Weyl groups.
(The rank of an automorphism is the dimension of a Cartan subspace, thus θ is of positive
rank if and only if the invariants are non-trivial.) Our method is a continuation of the
method used in [12] to determine the Weyl group for the classical graded Lie algebras in
positive characteristic. In particular, given an automorphism θ let T be a θ-stable maximal
torus such that Lie(T ) contains a Cartan subspace. Then θ = Int nw where nw ∈ NAut G(T ),
and w = nwT ∈ NAut G(T )/T is either of order m, or is trivial (in which case θ is of zero
rank). Using Carter's classification of conjugacy classes in the Weyl group, this approach
gives us a relatively straightforward means to determine the positive rank automorphisms
and their Weyl groups (see Tables 1-3). This classification allows us to show that all θ-groups
in types G2, F4 or D4 have KW-sections (see Thm. 5.11):
Theorem. Any θ-group of type G2, F4 or D4 admits a KW-section.
Together with [12], this leaves only the θ-groups of type E to deal with. These will be
dealt with in subsequent work with B. Gross, M. Reeder and J-K. Yu.
Notation.
Throughout, G will denote an almost simple (semisimple) algebraic group and g its Lie
algebra. If T is a maximal torus of G and α ∈ Φ(G, T ) then we denote the corresponding
coroot by α∨, which we can also consider as a cocharacter k× → T . If H is a subgroup of
G then NG(H) (resp. ZG(H)) will denote its normalizer (resp. centralizer) in G. Similar
notation will be used for elements of G and subalgebras or elements of g. The connected
component of an algebraic group H is denoted H ◦; the derived subgroup of a connected
2
algebraic group is denoted H ′. We denote by µs the cyclic group of order s. Throughout the
paper we will assume that m is coprime to the characteristic of k, if this is positive.
Acknowledgements.
I would like to thank Ross Lawther for directing me towards the paper of Carter. I have
also benefited from numerous helpful comments from Dmitri Panyushev, Oksana Yakimova,
Willem de Graaf, Jiu-Kang Yu, Benedict Gross, Mark Reeder and Gunter Malle.
1 Preparation
We continue with the basic set-up of the introduction. Here we recall some results and
definitions which will be necessary in what follows.
Let Φ be an irreducible root system with basis ∆ = {α1, . . . , αn}. Let α =Pn
i=1 miαi be
the longest root with respect to ∆. Recall that p is good for G if p > mi for all i, 1 ≤ i ≤ n;
otherwise p is bad. Specifically, p is bad if either: p = 2 and Φ is not of type A; p = 3 and Φ
is of exceptional type, or; p = 5 and Φ is of type E8. If Φ is a reducible root system, then p
is good for Φ if and only if p is good for each irreducible component of Φ; p is good for the
reductive group G if and only if it is good for the root system of G. We will sometimes refer
to the standard hypotheses on a group G: (A) that p = char k is good for G; (B) that G′ is
simply-connected; (C) that there exists a non-degenerate symmetric bilinear form g× g → k.
If G is simple and of exceptional type and p is good for G, then G is separably isogenous to
a group satisfying the standard hypotheses. (In particular, if G is of type F4 or G2 then (B)
and (C) are automatic as long as (A) is satisfied.)
Let T be any θ-stable torus in G.
If θ is an involution then it is not difficult to see
that there is a decomposition T = T+ · T−, where T+ = {t ∈ T θ(t) = t}◦ and T− = {t ∈
T θ(t) = t−1}◦. In [12] we introduced analogues of T± for an automorphism of arbitrary
finite order m (if the characteristic is positive, then we require p ∤ m). Let pd(x) be the
(monic) minimal polynomial of e2πi/d over Q. Since pd has coefficients in Z, we can (and
will) consider it as a polynomial in Fp[x] as well. If p ∤ d then pd(x) has no repeated roots in
i=0 aixi ∈ Z[x], let f (θ) denote the algebraic endomorphism
i=0 θi(tai). Then the map Z[x] → End T , f 7→ f (θ) is a homomorphism of
rings, where the addition in End T is pointwise multiplication of endomorphisms, and the
multiplication is composition.
k. For any polynomial f = Pn
T → T , t 7→ Qn
Lemma 1.1 ([12, Lemma 1.10]). For each dm, let Td = {t ∈ T pm/d(θ)(t) = e}◦. Then
i=0 t(i) where
T = Qdm Td and Td1 ∩ Td2 is finite for any distinct d1, d2. Moreover, t = Pm
t(i) = t ∩ g(i) and Lie(Td) =P(i,m)=d t(i).
Remark 1.2. An alternative description of the tori Td is given in [2]. Since θ is an alge-
braic automorphism of T , it induces an automorphism θ∗ of Y (T ) := Hom(k×, T ) given by
θ∗(λ)(t) = θ∗(λ(t)). Then set Y (Td) = ker pd(θ∗). This also allows us to define the tori Td
when m is divisible by p; however, in this case we do not have a direct sum decomposition
t = ⊕dm Lie(Td).
An immediate application of Lemma 1.1 is the following.
Lemma 1.3. Let T be a maximal torus of G and let w ∈ W = NG(T )/T be of order m
(p ∤ m) and have finitely many fixed points on T . Then any two representatives for w in
NG(T ) are T -conjugate, and have the same m-th power.
3
Proof. Suppose nw ∈ NG(T ) is such that w = nwT . We have T =Qim,i6=m Ti and therefore
tw(t)w2(t) . . . w−1(t) = 1 for all t ∈ T . Thus (nwt)m = nm
w for all t ∈ T . Moreover, since
Tm = {1}, the map T → T , t 7→ tw(t−1) is surjective. In particular, if t ∈ T then there
exists s ∈ T such that tnw = snws−1.
Let now θ be an automorphism of G of order m, let ζ be a primitive m-th root of 1 in
k and let g(i) = {x ∈ g dθ(x) = ζ ix}. A θ-stable torus in G is θ-split if T = T1, and
is θ-anisotropic if Tm = {e}, that is, if θ has finitely many fixed points on T . We recall
that a Cartan subspace of g(1) is a maximal subspace which is commutative and consists of
semisimple elements. We have the following results on Cartan subspaces.
Lemma 1.4. (a) Any two Cartan subspaces of g(1) are G(0)-conjugate, and any semisimple
element of g(1) is contained in a Cartan subspace.
(b) Let c be a Cartan subspace of g(1). There is a θ-split torus T1 in G such that
c ⊂ Lie(T1). We have dim T1 = dim c · ϕ(m), where ϕ(m) is the Euler number of m. (In
particular, dim(Lie(T1) ∩ g(i)) = dim c for any i coprime to m.) If p > 2, then T1 is unique.
(c) Let c be a Cartan subspace of g(1) and let Wc = NG(0)(c)/ZG(0)(c), the little Weyl
group. If char k 6= 2, then the embedding c ֒→ g(1) induces an isomorphism k[g(1)]G(0) →
k[c]Wc.
(d) If char k = 0 or if G satisfies the standard hypotheses, then Wc is generated by
pseudoreflections and k[c]Wc is a polynomial ring.
Proof. Part (a) is [21, Thm. 1], and [12, Thm. 2.5 and Cor. 2.6]. For (b), see [21, §3.1] and
[12, Lemma 2.7]. Part (c) was [21, Th. 7], [12, Thm. 2.20]. Finally, part (d) is Thm. 8 in
[21] and [12, Prop. 4.22].
We recall the following results of Steinberg [20, 7.5,9.16].
Lemma 1.5. (a) If θ is a semisimple automorphism of G then there is a θ-stable Borel
subgroup B of G and a θ-stable maximal torus T of B.
(b) Let π : G → G be the universal covering of G. Then there exists a unique automor-
phism θ of G such that π(θ(g)) = θ(π(g)) for all g ∈ G. Moreover, if θ is of order m then
so is θ.
Let G be a simple, simply-connected group and let g = Lie(G). Denote by Aut G
(resp. Aut g) the algebraic group of rational (resp. restricted Lie algebra) automorphisms
of G (resp. g). In characteristic zero it is well-known that all automorphisms of g arise as
differentials of automorphisms of G, and the corresponding map Aut G → Aut g is bijective.
This may fail to be true in small characteristic [6].
Lemma 1.6. If p > 2 then differentiation d : Aut G → Aut g is bijective.
Proof. If G = SL(n, k) then this was proved in [11, Lemma 1.4]. But if p > 2 then the
automorphism group of g as an abstract Lie algebra is isomorphic to Int G ⋊ D, where
D is the group of 'graph automorphisms' of the Dynkin diagram of G [6, II. Table 1].
Thus any automorphism of g is the differential of an automorphism of G, and the map
d : Aut G → Aut g is bijective.
4
2 Kac diagrams
Here we recall Kac's classification [8] of periodic automorphisms of simple Lie algebras in
characteristic zero, and show that the classification extends to positive characteristic (con-
sidering only those automorphisms of Lie(G) which arise as differentials of rational auto-
morphisms of G, and only automorphisms of order coprime to the characteristic). For more
details on the classification, we recommend chapters 6 to 8 of Kac's book [9]. Let G be almost
simple and simply-connected, let ∆ = {α1, . . . , αr} be a basis of the root system Φ(G, T )
and let ∆ = ∆ ∪ {α0} be a basis of the extended root system. Let {hα, eβ α ∈ ∆, β ∈ Φ} be
a Chevalley basis for G. Recall that Aut G is generated over Int G by finitely many (graph)
automorphisms, in particular Aut G/ Int G is finite. If θ is a (rational) automorphism of G
then the index of θ is the order of θ modulo the inner automorphisms.
For the time being, let k = C and let ζ = e2πi/m. To each automorphism σ of g of (finite)
order m, Kac associated an infinite-dimensional Lie algebra, the twisted loop algebra:
L(g, σ) =Xj∈Z
ti ⊗ gj ⊂ C[t, t−1] ⊗ g
where gj = {x ∈ g σ(x) = ζ jx}. Kac showed that twisted loop algebras are closely related to
the extended Dynkin diagram of g; the Dynkin diagrams for the "twisted" types s > 1 are
Theorem 2.1. a) If σ and θ are two periodic automorphisms of g of the same index, then
L(g, σ) ∼= L(g, θ). In particular, if σ is inner then L(g, σ) ∼= L(g, 1) = C[t, t−1] ⊗ g.
affine type Lie algebras. To explain this, letbg(s) denote the affine type Lie algebra associated
to g and s (for s the index of some automorphism of g): the Dynkin diagram of bg(1) is just
given below. It is well known that bz(s) = z(bg(s)) is one-dimensional and contained in the
derived subalgebra (bg(s))′ (which is of codimension 1 inbg(s)). Identify a basis of simple roots
forbg(1) with ∆. Let bΦ denote the root system ofbg(s).
b) L(g, σ) ∼= (bg(s))′/bz(s), where s is the index of σ.
c) The grading of L(g, σ) given by L(g, σ)i = ti ⊗ gi induces a grading of bg(s). There is
a Cartan subalgebra bh of bg(s), a basis Π = {β0, . . . , βl} of simple roots in bΦ and a sequence
(n0, . . . , nl) of non-negative integers such that the grading of bg(s) is the grading obtained by
putting bh in degree zero and bg(s)
d) In the notation of (c), the order of σ is m = s(Pl
the shortest positive imaginary root in bΦ.
e) Given a grading of bg(s) determined by a sequence (n0, . . . , nl) of non-negative in-
tegers as in (c) there is an induced Z/mZ-grading of g given by the isomorphism g ∼=
L(g, σ)/(tm, t−m)L(g, σ).
j=0 ajnj), where sδ = sPl
in degree ±ni.
±βi
j=0 ajβj is
Remark 2.2. a) Another way to express (c) is that the grading of ZΠ, considered as an
element of the affine apartment, is conjugate by an element of the affine Weyl group to a
point in either the fundamental alcove or its opposite.
b) It is easy to show that bz = {h ∈ bh hα, hi = 0 for all α ∈ bΦ}. Dually, the space
{χ ∈ bh∗ hχ, α∨i = 0 for all α ∈ bΦ} is one-dimensional, generated by δ.
δ = α0 +P aiαi, where P aiαi = α is the highest root of the (finite) root system Φ. For
s 6= 1 we mark the coefficients aj on the diagrams below.
If s = 1 then
The "twisted type" Dynkin diagrams are as follows (type A(2)
3
is given by D(2)
3 ):
5
A(2)
2n (n ≥ 2)
A(2)
2
❝ ✟❍ ❝
2
1
❡ ❍✟ ❡
2
1
A(2)
2n−1 (n ≥ 3)
❝ ✟❍ ❝
2
1
D(2)
n+1 (n ≥ 2)
❝ ✟❍ ❝
1
1
❝ ❍✟ ❝
2
2
❝
1
❝
❅
2
❅
❅
❝
1
❝ ❍✟ ❝
1
1
D(3)
4
E(2)
6
❝
1
❝
1
❝ ❍✟ ❝
1
2
❝
2
❝ ❍✟ ❝
2
3
❝
1
(The diagrams for A(2)
2n are not given in their usual orientation. This choice of orientation
is due to our 'type-free' choice of graph automorphism, see the discussion below.)
The information Π, (n0, . . . , nl) can be written in a diagram called a Kac diagram µ, which
is a copy of the Dynkin diagram of bg(s) with the coefficient ni written above the vertex βi.
The order of a Kac diagram µ is the sum s(a0n0 + . . .+ alnl). From now on, any Kac diagram
µ will be primitive, that is, it won't be possible to write µ as jµ′ for j > 1 and µ′ some other
Kac diagram. Then to each Kac diagram we can associate an automorphism (which we call
a Kac automorphism) of g of order m; Kac's theorem tells us that any automorphism of g
of order m is conjugate (possibly by an element of the outer automorphism group of g) to a
Kac automorphism of order m.
In the inner case it is easy to see what the Kac automorphism corresponding to µ is: it's
Ad t, where t ∈ T is such that αi(t) = ζ µ(αi) for 1 ≤ i ≤ r. (Here ζ = e2πi/m.) Then it is
immediate that α(t) = ζ −µ(α0) and that gt is generated by t and all g±αi, where µ(αi) = 0
(0 ≤ i ≤ r).
Example 2.3. We describe all (classes of) automorphisms of order 3 of a Lie algebra of type
F4. Here any automorphism is inner. The possible Kac diagrams of order 3 are 00100, 11000
and 10001. Thus there are three classes of automorphisms of order 3, with representatives
Int t1, Int t2, Int t3, t1, t2, t3 ∈ T satisfying:
(i) α1(t1) = α3(t1) = α4(t1) = 1, α2(t1) = ζ;
(ii) α2(t2) = α3(t2) = α4(t2) = 1, α1(t2) = ζ; and
(iii) α1(t3) = α2(t3) = α3(t3) = 1, α4(t3) = ζ.
We wish to generalise this construction to positive characteristic p. If p ∤ m then we can
do this by fixing a primitive m-th root ζ of unity, and continuing exactly as above. Here,
we need to specify that Int t is an automorphism of G with differential Ad t; while it is no
longer true in general that gt is generated by t and the subspaces gαi with µ(αi) = 0, it is
nevertheless true that the subgroup ZG(t)◦ is (reductive and is) generated by T and all root
subgroups U±αi (see, for example, [7, 26.3]) with µ(αi) = 0.
Now consider Kac diagrams of twisted type.
In order to extend Kac's classification
to positive characteristic, we will first give a precise description of the Kac automorphism
6
corresponding to each Kac diagram in characteristic zero, and then explain how to extend
this description to positive characteristic. For a given index s let γ be a graph automorphism
of the root system Φ which preserves the basis ∆, and by abuse of notation let γ also denote
the unique automorphism of g which satisfies γ(eα) = eγ(α) for α ∈ ±∆. (For each index
s there is a unique such graph automorphism of Φ unless g is of type D4 and s = 3; in
this case there are two such γ, which are mutually inverse and conjugate under the action
of the automorphism group of g. In this case let us fix the choice satisfying γ(α1) = α3.)
Let T (0) = (T γ)◦. For each α ∈ Φ let (α) be the set of all hγi-conjugates of α and let
4 ; α = α1 +α2 +α3 +2α4 +2α5 +α6 in type E(2)
g(α) =Pβ∈(α) gβ. Let α be a root which is longest subject to g(α) 6⊆ gγ. Specifically, we have:
1 αi in type D(2)
α = α for type A(2)
n ; α = α1+α2+α3
in type D(3)
6 . Let β0 be the hγi-conjugacy class
2n−1, D(2)
of −α and choose E±β0 ∈ g(∓α) as follows: in case A(2)
n
and E(2)
let Eβ0 = e−α + σ−1γ(e−α) + σγ2(e−α) and
6
E−β0 = eα +σγ(eα)+σ−1γ2(eα), where σ = e2πi/3. We label the node on the far left hand side
(as displayed in the diagrams above; on the top in type A(2)
2n−1) with β0. (As a character of
T (0), β0 is just (−α)T (0).) Note that [Eβ0, E−β0] = Hβ0 ∈ tγ satisfies [Hβ0, E±β0] = ±2E±β0.
The remaining nodes are all labelled by hγi-conjugacy classes in ∆. Specifically, we label
the nodes β1, . . . from left to right:
let E±β0 = e∓α − γ(e∓α); in case D(3)
4
2n−1; α =Pn−1
2n ; α =P2n−2
2n let E±β0 = e∓α; in cases A(2)
i=1 αi for type A(2)
- βi = {αi, α2n+1−i} for each i, 1 ≤ i ≤ n in type A(2)
- βi = {αi, α2n−i} for 1 ≤ i ≤ n − 1 (β1 appearing underneath β0) and βn = {αn} in type
2n , n ≥ 1,
A(2)
2n−1,
- βi = {αi} for 1 ≤ i ≤ n − 1 and βn = {αn, αn+1} in type D(2)
- β1 = {α1, α3, α4} and β2 = {α2} in type D(3)
4 ,
- β1 = {α1, α6}, β2 = {α3, α5}, β3 = {α4} and β4 = {α2} in type E6
n+1,
(2)
.
We can set E±βi = Pα∈βi
e±α for each i, 1 ≤ i ≤ n except for βn in A(2)
2n in which case
(p 6= 2 and) we set Eβn = eαn + eαn+1, E−βn = 2e−αn + 2e−αn+1. Note that (for 0 ≤ i ≤ n)
βi is a well-defined character on T (0) and Ad t(Eβi) = βi(t)Eβi for any t ∈ T (0). Moreover,
setting Hβi = [Eβi, E−βi] and letting hβi, βji be the Cartan numbers determined by the
twisted diagram, it is easy to establish the following commutation relations:
[Eβi, E−βj ] =(cid:26) 0
Hβi
if i 6= j,
if i = j,
[Hβi, E±βj ] = ±hβi, βjiE±βj , [E±βi, E±βj ] = 0 if hβi, βji = 0
and (ad Eβi)1−hβi,βji(Eβj ) = (ad E−βi)1−hβi,βji(E−βj ) = 0 for i 6= j. In addition, γ(Hβi) = Hβi
for all i and γ(E±βi) = E±βi for 1 ≤ i ≤ n. Thus, if I is a proper subset of Π then the
elements E±βi, βi ∈ I and tγ together generate a reductive subalgebra of g with root system
given by the subdiagram of the twisted affine diagram which contains exactly the vertices
corresponding to elements of I. With this set-up, let µ be a Kac diagram of twisted affine
i=0 µ(βi)mi, where mi are the coefficients indicated
on the twisted diagrams above.) There exists t ∈ T (0) such that βi(t) = ζ µ(βi) for 1 ≤ i ≤ n.
Let θµ = Ad t ◦ γ: then θµ is of order m, θµ(E±βi) = ζ ±µ(βi)E±βi for 0 ≤ i ≤ n and gθµ is the
reductive subalgebra of g generated by tγ and the E±βi such that µ(βi) = 0.
type, of order m. (The order of µ is sPn
We can also repeat the above construction in (coprime) positive characteristic, with the
following clarification. For each α ∈ Φ let Uα be the root subgroup of G corresponding to
α [7, 26.3]. There exists a unique isomorphism (of algebraic groups) ǫα : Ga → Uα such
that tǫα(x)t−1 = ǫα(α(t)x) for all t ∈ T , x ∈ k and (dǫα)0(1) = eα (see for example [7,
7
Thm. 26.3]). By abuse of notation, let γ denote the unique automorphism of G such that
γ(ǫα(x)) = ǫγ(α)(x) for α ∈ ±∆. Fixing a primitive m-th root of unity ζ, we can associate
an automorphism Int t ◦ γ of G to each Kac diagram µ (of order coprime to p). As before
we will say that the automorphism θµ constructed in this way is the Kac automorphism
associated to the Kac diagram µ. We note that γ is the Kac automorphism corresponding
orthogonal and the subgroups Uα, α ∈ βi commute, or βi = {αn, αn+1} in type A(2)
to µ with µ(βi) = (cid:26) 1 if i = 0,
i ≥ 1, if the roots in βi are orthogonal then let ǫ±βi : Ga → G, ǫ±βi(x) = Qα∈±βi
2n . For
ǫα(x) and
let U±βi = ǫ±βi(Ga) (unipotent subgroups of G). Then (dǫ±βi)0(1) = E±βi, where E±βi is
as defined in the paragraph above. For the case βn in type A(2)
2n we can construct ǫ±βn by
restricting to the subgroup of type A2 which contains U±αn and U±αn+1. Specifically, ǫβn(x) =
ǫαn(x/2)ǫαn+1(x)ǫαn(x/2) and ǫ−βn(x) = ǫ−αn(x)ǫ−αn+1(2x)ǫ−αn(x) define homomorphisms
ǫ±βn : Ga → Gγ such that (dǫ±βn)0(1) = E±βn, where E±βn are defined as in characteristic
zero. Finally, either:
Moreover, for βi ∈ Π, either all roots α ∈ βi are
0 otherwise.
- α is γ-stable, in which case we define ǫ±β0(x) = ǫ∓α(x);
- or s = 2 and γ(α) 6= α, in which case we set ǫ±β0(x) = ǫ∓α(x)γ(ǫ∓α(−x));
- or s = 3 and we set σ = ζ m/3, ǫβ0(x) = ǫ−α(x)γ(ǫ−α(σ−1x))γ−1(ǫ−α(σx)) and ǫ−β0(x) =
ǫα(x)γ(ǫα(σx))γ−1(ǫα(σ−1x)).
We have γ(ǫ±β0(x)) = ǫ±β0(−x) if s = 2, γ(ǫ±β0)(x) = ǫ±β0(σ±1x) if s = 3 and
(dǫ±β0)0(1) = E±β0, where E±β0 are defined as in characteristic zero. Let U±β0 = ǫ±β0(Ga).
Then each of the subgroups U±βi with µ(βi) = 0 is contained in G(0) = (Gθµ)◦. We recall [20,
8.1] that G(0) is reductive. Moreover, (T θµ)◦ = T (0) is regular in G by inspection, and hence
is a maximal torus of G(0). We claim that G(0) is generated by T (0) and the subgroups
U±βi with µ(βi) = 0. To see this, we remark first of all that dim g(0) is independent of the
characteristic. For if α ∈ Φ then let l(α) be the number of γ-conjugates of α (here either 1, 2
or 3). Then g(α) ∩g(0) is of dimension 1 if dθl(α)
(eα) = eα and of dimension 0 otherwise. Since
this description is clearly independent of the characteristic of k, it follows that the same can
be said of the dimension of g(0). Let ΦI be the root system with basis {βi ∈ Π : µ(βi) = 0}.
Then our assumption that p is coprime to the order of θµ implies that in all cases p is good
for ΦI except for one case in type D4
(the Kac diagram 100) with p = 2 and one case in
type E(2)
(the Kac diagram 10000) with p = 3. If p is good then E±βi, µ(βi) = 0 and tγ
6
generate a subalgebra of g which is isomorphic to the Lie algebra of a reductive group with
root system ΦI . Since the dimension is independent of the characteristic, this subalgebra
must be all of g(0). Now, since the subgroup of G generated by T (0) and the U±βi, µ(βi) is
contained in G(0), it must be reductive and have root system ΦI.
(3)
µ
4 and 10000 in type E(2)
This leaves only the Kac diagrams 100 in type D(3)
two classes: (i) those such that γ(eα) = eα (specifically ±α ∈ {α2,P4
6 ; in both cases
the corresponding Kac automorphism is just γ. In the first case the roots α ∈ Φ fall into
1 αi, α}), (ii) those
It is thus easy to see that dim g(α) ∩ gγ = 1 for all α and, setting
such that γ(α) 6= α.
Hβ1 = hα1 + hα3 + hα4, Hβ2 = hα2, E(α) = eα in case (i) and E(α) = eα + γ(eα) + γ−1(eα)
in case two, the set {Hβ1, Hβ2, E(α) α ∈ Φ} is a basis for gγ and satisfies the relations of a
Chevalley basis for a Lie algebra of type G2. (Hβ1 and Hβ2 are always linearly independent
since the centre of g is generated by hα1 + hα3 and hα1 + hα4.) Since G(0) contains T (0)
and the subgroups U±β1, U±β2, this shows that G(0) is generated by these subgroups and
is semisimple of type G2.
In the second case, even though p = 3 is not good for a root
8
system of type F4, it is nevertheless clear that the Lie subalgebra of g generated by tγ and
the E±βi, 1 ≤ i ≤ 4 is isomorphic to the Lie algebra of a semisimple group of type F4 (since
it contains all appropriate root subspaces for T (0)). Thus by the same argument as above
G(0) is generated by T (0) and the subgroups U±βi, 1 ≤ i ≤ 4.
We will now show that any automorphism of order m is Aut G-conjugate to a Kac auto-
morphism. For inner automorphisms, this result appeared in [18]. One approach would be
to construct the loop algebra L(g, σ) as in characteristic zero; but this would (presumably)
require the theory of Kac-Moody Lie algebras in positive characteristic. For inner automor-
phisms, a more direct proof is suggested by considering all Z/mZ-gradings of g, and reducing
the problem to the statement about the fundamental alcove for the affine Weyl group 2.2(a).
(This was the approach of [18], which also makes it possible to extend Kac's classification to
semisimple inner automorphisms of order p, that is, Z/pZ-gradings for which Lie(T ) ⊂ g(0).)
This argument could also be applied to the outer automorphisms by considering the appro-
priate Bruhat-Tits building; but here we avoid such technical machinery by expressing the
problem in terms of conjugacy classes in T (0).
Lemma 2.4. The number of elements of order m in a torus of rank n is independent of the
characteristic of the ground field, assuming this characteristic is either zero or coprime to
m.
Proof. Let T be a torus of rank n. There are mdim T elements t of T satisfying tm = 1.
Let φT (m) denote the number of elements of T of order m. Then φT (m) = mdimT −
Pdm,d6=m φT (d). Since φT (1) = 1, the lemma follows by induction.
Lemma 2.5. Let G be almost simple and of simply-connected or adjoint type. Then any
automorphism of G of finite order coprime to p is Aut G-conjugate to a Kac automorphism.
Proof. Let T be a maximal torus of G. Let Z(G) be the scheme-theoretic centre of G and
let G = G/Z(G) be the adjoint group of G. Clearly any automorphism of G induces an
automorphism of G. Moreover, the induced map Aut G → Aut G is injective by a standard
description of Aut G in terms of graph automorphisms and inner automorphisms. Thus we
may assume that G is of adjoint type. Now any semisimple inner automorphism of G is
conjugate to Int t for some t ∈ T . Let ¯W = NAut G(T )/T . Then two elements of T are
Aut G-conjugate if and only if they are ¯W -conjugate.
To each element t ∈ T of order m we associate the homomorphism λt : ZΦ → Z/mZ such
that α(t) = ζ λt(α). If µ is a Kac diagram of non-twisted type, then by the above discussion
the corresponding Kac automorphism of G is Int tµ, where λtµ(αi) = µ(αi) modulo m for
0 ≤ i ≤ n. The action of ¯W on Φ gives rise to an action on the set of homomorphisms
ZΦ → Z/mZ. Moreover, this action satisfies Z ¯W (t) = Z ¯W (λt) for t ∈ T of order m, and
thus Z ¯W (tµ) = Z ¯W (µ) for any Kac diagram µ.
It is possible for two distinct Kac diagrams to be conjugate by some element of W . In
our circumstances it is easier to describe when Kac automorphisms corresponding to Kac
diagrams µ1 and µ2 are conjugate by an element of ¯W : this holds if and only if there is an
isomorphism of diagrams µ1 → µ2 (that is, an isomorphism ψ of graphs which also satisfies
µ2(ψ(α)) = µ1(α) for each node α). Pick a set I of representatives for the ¯W -conjugacy
classes of Kac diagrams of order m. Then the number of automorphisms of the form Int t,
t ∈ T which are conjugate to a Kac automorphism of order m is Pµ∈I #( ¯W )/#(Z ¯W (µ)) =
Pµ∈I #( ¯W )/#(Z ¯W (tµ)). If char k = 0, then this sum is equal to the number of elements
9
of T of order m. But therefore it is equal to the number of elements of T of order m in
arbitrary characteristic.
Now, suppose θ is an outer automorphism of G. Since any semisimple automorphism
of G fixes some Borel subgroup of G and a maximal torus contained in it [20, 7.5], any
outer automorphism of G is conjugate to an automorphism of the form Int t ◦ γ, where
t ∈ T and γ is one of the graph automorphisms described above. Let T (0) = (T γ)◦ and
let T (1) = {t ∈ T γ(t) = t−1}◦. Then T is the almost direct product of T (0) and T (1) by
Lemma 1.1. Hence t = t0.t1 for some t0 ∈ T (0), t1 ∈ T (1). Thus there is some s ∈ T (1) such
that s2 = t1; then Int s−1 ◦ Int t ◦ γ ◦ Int s = Int ts−2 ◦ γ = Int t0 ◦ γ. Therefore any outer
automorphism of G is conjugate to an automorphism of the form Int t ◦ γ with t ∈ T (0).
We claim that Int t ◦ γ and Int s ◦ γ, for t, s ∈ T (0) are conjugate if and only if there
exists some g ∈ NGγ (T ) such that gtg−1 = s. Since G is adjoint, Int t ◦ γ and Int s ◦ γ are
conjugate if and only if there exists some x ∈ G such that xtγ(x−1) = s. Consider the Bruhat
decomposition x = unv of x, where u, v ∈ U and n ∈ NG(T ). We have xt = unt · (t−1vt) and
sγ(x) = (sγ(u)s−1) · sγ(n)γ(v), thus by uniqueness nt = sγ(n). In particular, nT ∈ W γ.
Now, a direct check shows that each element of W γ has a representative in Gγ (once again
using adjointness of G), thus s and t are NGγ (T )-conjugate.
It follows that the Int G-conjugacy classes of outer automorphisms of the form Int x◦γ are
in one-to-one correspondence with the W γ-conjugacy classes in T . Repeating the counting
argument used for inner automorphisms above, we deduce that any outer automorphism is
conjugate to a Kac automorphism.
From now on we will refer to the type of an outer automorphism of a simple group G
by specifying the type of its corresponding Kac algebra, but without the tilde. We note the
following corollary of the above result for inner automorphisms. In good characteristic this
was already known ([13, Prop. 30] or [17, Prop. 3.1]). Recall that a pseudo-Levi subgroup of
G is a subgroup of the form ZG(x)◦ for a semisimple element x of G. If I is a proper subset
of ∆ then we denote by LI the subgroup of G generated by T and all U±αi with αi ∈ I.
Corollary 2.6. The pseudo-Levi subgroups are the G-conjugates of subgroups of the form
LI , where I is a proper subset of the ∆ such that not all elements of ∆ \ I have coefficient
divisible by p in the expression for ν.
(If p is good then there is no restriction here. The groups LI which don't satisfy the
conditions of Cor. 2.6 are the centralizers of non-smooth diagonalizable subgroups [18].)
For later use we make the following observation.
Lemma 2.7. Suppose µ is a Kac diagram corresponding to a positive rank automorphism.
Then µ(α) ∈ {0, 1} for each α ∈ Π.
Proof. Suppose µ(α) > 1 for some α ∈ Π. Then Π′ = Π \ {α} is a union of Dynkin diagrams.
If µ corresponds to an inner automorphism then l = t ⊕Pβ∈ZΠ′ gβ is the Lie algebra of the
subgroup LΠ′ of G, in the notation introduced above. Moreover, the subspace spanned by all
gβ, where β is non-zero and has non-negative coefficients in the elements of Π′, is contained
in the Lie algebra uΠ′ of a maximal unipotent subgroup of LΠ′. In particular, g(1) ⊂ uΠ′
and thus does not contain any semisimple elements.
Suppose therefore that µ is an outer automorphism of G. If µ(βi) > 1 for some i then let
µ′ be the Kac diagram with µ′(βj) = δij. There is a Kac automorphism ψ of G corresponding
to µ′. Moreover, it follows from the description of Kac automorphisms that g(1) ⊂ gdψ. But
θGψ is a zero-rank (inner) automorphism by the same argument as above.
10
3 Carter's classification of conjugacy classes
Let W be a Weyl group with natural real representation V . Here we recall the set-up of the
classification, according to Carter [3] of the conjugacy classes in W . Any element w of W can
be expressed as a product w = w1w2, where w2
2 = 1 and {v ∈ V w1 · v = −v} ∩ {v ∈
V w2 · v = −v} = {0}. Moreover, any involution w′ ∈ W can be expressed as a product of l
reflections corresponding to orthogonal roots, where l = dim{v ∈ V w′(v) = −v}. Thus let
sα (i = 1, 2). Then this gives
an expression for w as a product of reflections corresponding to l1 + l2 = dim V − dim V w
roots, where li = dim{v ∈ V wi(v) = −v}, i = 1, 2. Associate a graph Γ to w with one
I1, I2 be subsets of Φ with #(Ii) = l(wi) such that wi =Qα∈Ii
node for each root in I1` I2 and hα, βihβ, αi edges between nodes corresponding to distinct
1 = w2
roots α, β ∈ I1 ∪ I2. The graph Γ constructed in this way is the admissible diagram
associated to w. Less formally, we will say that w has Carter type Γ. (In fact, it is possible
to have two different such expressions for w giving rise to two different diagrams, and on
the other hand, non-conjugate elements can have the same admissible diagram. The first
problem Carter solved simple by making a particular choice of admissible diagram for each
conjugacy class. For the second problem, which only occurs in types D and E, some classes
are marked with a single prime and others with a double prime. There are never more than
two classes for the same admissible diagram [3].) For example, if Γ is the Dynkin diagram
for W then w is a Coxeter element of W , while if Γ is the trivial graph then w is the
identity element. The irreducible admissible diagrams are classified and their characteristic
polynomials determined in [3, Tables 2-3]. We will say that an admissible diagram Γ has
order N if an element of the corresponding conjugacy class has order N. Let w be an element
of W with admissible diagram Γ. We denote by Φ1 the (closed) subsystem of Φ spanned by
the elements of Γ, and by Φ2 the set of roots α ∈ Φ which are orthogonal to all roots in Φ1.
In [3, Tables 8-12], Carter lists the conjugacy classes in the exceptional type Weyl groups,
classified by admissible diagrams, and gives the number of elements in each class (and thus
the number of elements in the centralizer of an element of a given class). Following Carter's
notation, we will use a tilde to denote a subsystem consisting of short roots.
Example 3.1. We give some examples to illustrate the above notions.
a) Let W = Sn, the symmetric group on n letters. Then the admissible diagram corre-
ponding to an m-cycle is a Coxeter graph of type Am−1. More generally, any element w of
W is a product of disjoint cycles, of length m1, . . . , mr say. Then the admissible diagram
for w is a disjoint union of Coxeter graphs Am1−1, . . . , Amr−1. We say that w is of type
Am1−1 × . . . × Amr−1. In this case Φ1 is isomorphic to the union Am1−1 ∪ . . . ∪ Amr−1 and Φ2
is isomorphic to An−m1−...−mr−1.
b) If W is a Weyl group of type Bn then any element w ∈ W can be decomposed as a
product of signed cycles: a positive m-cycle has order m, and a negative m-cycle has order
2m. The admissible diagram for a positive m-cycle is a Coxeter graph of type Am−1; the
admissible diagram for a negative m-cycle is a Coxeter graph of type Bm. If w is a positive
m-cycle then Φ2 is isomorphic to Bn−m unless m = 2, in which case Φ2 is isomorphic to
Bn−2 ∪ A1. If w is a negative m-cycle then Φ2 is isomorphic to Dn+1−m.
c) If W is a Weyl group of type Dn then W is contained in a Weyl group of type Bn.
Thus any element of W can also be expressed as a product of disjoint signed cycles. (On the
other hand, two elements of W may have the same signed cycle type but not be conjugate
in W .) A product w of two negative 2-cycles has admissible diagram labelled D4(a1).
We want to ensure that all of the relevant information from [3] can also be applied in
11
positive characteristic, including the information about the characteristic polynomials of
elements of W . Consider the 'natural' representation of W to be its representation in a
Cartan subalgebra of the corresponding complex simple Lie algebra. For w ∈ W , denote by
cw(t) the characteristic polynomial of w in this natural representation.
Lemma 3.2. Let char k = p > 0 and let G be a simple group over k with root system
Φ = Φ(G, T ), where T is a maximal torus of G. Then the minimal polynomial of w as an
automorphism of t = Lie(T ) is just the reduction modulo p of cw(t).
Proof. Given a basis {χ1, . . . , χn} of the lattice of cocharacters Y (T ) we can associate a
matrix A ∈ GL(n, Z) to w by: w(χi) = Pj Ajiχj. Since w(dχi(1)) = Pj Ajidχj(1) and
{dχ1(1), . . . , dχn(1)} is a basis of t, clearly the characteristic polynomial of w as an element
of GL(t) is just the reduction modulo p of the characteristic polynomial of A. On the other
hand, the characteristic polynomial of A is invariant under change of basis in Y (T ) ⊗Z Q,
and in particular is the same if one replaces G by its universal covering. But now one can
choose a basis for Y (T ) consisting of the α∨ where α is a simple root. Since the matrix thus
associated to w is clearly independent of the characteristic, its characteristic polynomial
must be equal to cw(t).
Lemma 3.3. Let G, T, t, Φ, W, w, Φ1, Φ2 be as above and suppose that w has order m, p ∤ m.
Let U = {t ∈ t w(t) = ζt}, where ζ is a primitive m-th root of unity in k. Then any element
of W (Φ2) acts trivially on U.
Proof. Since G is simply-connected, t is spanned by elements hα, with α in a basis of simple
roots. Let tΦ1 be the linear span of the hα with α ∈ Φ1 and let V = {t ∈ t dα(t) = 0 ∀α ∈
Φ1}. Then t = tΦ1 ⊕ V by consideration of dimensions, and clearly U ⊂ tΦ1. But if β ∈ Φ2
then sβ(hα) = hα for all α ∈ Φ1, thus any element of W (Φ2) acts trivially on U.
We remark that a suitable generalization of Lemma 3.3 holds for any orthogonal subsys-
tems Φ1, Φ2 and an element w ∈ W (Φ1).
4 Determination of positive rank automorphisms
In this section we give details of the calculations we use to determine the positive rank
automorphisms. Assume that G is simple, simply-connected and of exceptional type, and
that char k is either zero or a good prime for G. (If G is simple and of exceptional type, then
the assumption that p is good implies that G is separably isogenous to a group satisfying the
standard hypotheses.) In particular we assume that p > 3, and therefore that all elements
of a Weyl group of type G2 or F4 are semisimple. (We recall that W (G2) is a dihedral group
of order 12, and W (F4) has order 1152 = 27.32.) Let c be a Cartan subspace of g(1) and let
T1 be the (unique) maximal θ-split torus of G such that c ⊂ Lie(T1) (Lemma 1.4(b)). Recall
that ZG(c)◦ = ZG(T1) [12, Rk. 2.8(c)]; since G is simply-connected, ZG(c) is connected and
hence equals ZG(T1). Let T (0) be a maximal torus of ZG(0)(c)◦. Then ZG(T (0)) ∩ ZG(c)
is a θ-stable maximal torus of G [12, Lemma 4.1]. In this section we will fix c and T (0),
and set T = ZG(T (0)) ∩ ZG(c), t = Lie(T ). With this assumption on T , we make some
straightforward observations.
Lemma 4.1. Suppose θ is inner. Then θ = Int nw, where nw ∈ NG(T ). Moreover, either
w = nwT has order m, or nw ∈ T and θ is of zero rank.
12
Proof. Since T is θ-stable, the first statement is obvious. But if nw has order less than
m, then {t ∈ t dθ(t) = ζt} is trivial, and therefore by our choice of T , T (0) = T and
nw ∈ T .
In fact, we can make a more precise statement than this.
Lemma 4.2. Either nw ∈ T or w = nwT has Carter type Γ(1) × . . . × Γ(l), where the Γ(i),
1 ≤ i ≤ l are disjoint irreducible diagrams and at least one has order m.
Proof. Let Φ(i) be the smallest root subsystem of Φ containing all the elements of Γ(i), let
t(i) be the vector subspace of t spanned by all hα with α ∈ Φ(i) and let w = w(1) . . . w(l),
where w(i) is a product of reflections corresponding to the vertices of Γ(i). (Here we define
hα = dα∨(1).) Since the roots in Γ(1) ∪ . . . ∪ Γ(l) are linearly independent ([3, §4]), and since
G is simply-connected, it follows that t contains the direct sum t(1) ⊕ . . . ⊕ t(l). Clearly w
preserves each of the subspaces Φ(i) and acts on Φ(i) by w(i). Hence {t ∈ t w(t) = ζt} is
spanned by the subspaces {t ∈ t(i) w(i)(t) = ζt}. In particular, if none of the w(i) has order
m then c is trivial, hence so is w by our assumption on T (0).
Lemma 4.2 excludes certain possibilities for w, such as an element of type A1 × A2 (since
here m = 6, but no irreducible subdiagram has order 6). Later we will see that in fact each
of the Γ(i) must have order m. (This was already established for the classical case in [12].
For outer automorphisms the situation is more complicated.)
Lemma 4.3. Let G be simple and simply-connected group and let h be the Coxeter number
of G. Assume the characteristic of the ground field is coprime to h. If nw ∈ NG(T ) is such
that w = nwT is a Coxeter element in W , then Int nw is an automorphism of G of order h
and the corresponding Kac diagram is the diagram with 1 on every node.
r (αr(tr−1)) . . . α∨
w ∈ Z(G). Clearly t = nh
Proof. For the first part we just have to show that nh
w satisfies
w(t) = t. We claim that this implies αi(t) = 1 for each i, hence t = Z(G). To prove the
claim, we may assume that w = sαr . . . sα1 where {α1, . . . , αr} is a basis of simple roots
of Φ(G, T ). Let t0 = t, ti = sαi . . . sα1(t) for 1 ≤ i ≤ r. Then ti+1 = ti · α∨
i+1(αi(ti)).
Hence t = t · α∨
1 (α1(t)). Since G is simply-connected, this implies that
αr(tr−1) = . . . = α1(t) = 1. But then t = t1 = . . . = tr−1, thus αi(t) = 1 for 1 ≤ i ≤ r.
Therefore t ∈ Z(G). Now, since any two maximal tori of G are conjugate, any two Coxeter
elements in W are conjugate, and any two representatives of a Coxeter element w are T -
conjugate by Lemma 1.3, it follows that there is a unique conjugacy class C of automorphisms
of G which act as a Coxeter element on some maximal torus. We claim that the Kac
automorphism corresponding to the Kac diagram with 1 on every node is of positive rank.
Indeed, it follows from the fact that p ∤ h that G satisfies the standard hypotheses. Thus we
regular nilpotent element, let λ : k× → T be the unique cocharacter such that α(λ(t)) = t2
for α ∈ ∆ and let m be an Ad λ(k×)-stable subalgebra of u− such that g = u− ⊕ [e, g]. (Here
can construct a KW-section for the adjoint representation as follows: let e =Pα∈∆ eα be a
u− = Pα∈−Φ+ gα.) Then e + m is a KW-section and hence e + m is non-nilpotent for any
non-zero m ∈ m. Now, since m is Ad λ(k×)-stable, we must have e− α ∈ m, and therefore
c = e + e− α ∈ g(1) ∩ (e + m). Since c is not nilpotent, we deduce that the Kac automorphism
with all coefficients equal to 1 is positive rank. Since it is the only Kac automorphism of
order h with coefficients 1 or 0, the lemma follows.
13
We will say that an automorphism which is conjugate to Int nw, where nw ∈ NG(T )
represents a Coxeter element in W , is a Coxeter automorphism. By Lemma 4.3, the Coxeter
automorphisms form a single conjugacy class in Int G. The following lemma is presumably
known, but we could not find a clear reference in the literature.
Lemma 4.4. Let nw ∈ NG(T ) be such that nwT represents a Coxeter element in W . Assume
p is coprime to h as before. Then nh
w is as follows:
a) G = SL(n, k), then h = n and nh
b) G = Spin(2n + 1, k), then h = 2n and nh
w = I if n is odd, nh
w = −I if n is even.
w = 1 if n ≡ 0 or 3 modulo 4, while
nh
w = α∨
n(−1) otherwise.
c) G = Sp(2n, k), then h = 2n and nh
d) G = Spin(2n, k) then h = 2(n − 1) and nh
w = −I.
nh
w = α∨
n−1(−1)α∨
n(−1) otherwise.
w = 1 if n ≡ 0 or 1 modulo 4, while
e) G of type G2 (resp. F4, E6, E8) then h = 6 (resp. h = 12, h = 12, h = 30) and
nh
w = 1.
g) G of type E7, then h = 18 and nh
w = α∨
2 (−1)α∨
5 (−1)α∨
7 (−1).
automorphism is just Int t, where t = Qn−1
Int t, where t =Qn−2
Proof. By Lemma 4.3, nh
w ∈ Z(G) in each case. Let ζ be a primitive h-th root of unity.
Types A or C can be calculated directly. If G is of type G2, F4 or E8 then Z(G) is trivial.
If G is of type Bn, then we use Lemma 4.3: Int nw is conjugate to the Kac automorphism
which sends each eα, α ∈ ∆ to ζeα, where ζ is a primitive h-th root of unity. But this
n(ξn(n+1)/2), ξ a square-root
of ζ. Thus th is as described. The calculation for type Dn is similar: Int nw is conjugate to
n(ξn(n−1)/2) and thus th = 1 if and
only if n ≡ 0 or 1 modulo 4. If G is of type E6 then let γ be an outer automorphism of G.
Then γ permutes the two non-identity elements of Z(G). On the other hand, Int γ(nw) ∈ C
and thus γ(n12
w = 1. Finally, if G is of type E7 then w is of order 18 and
Int nw is conjugate to Int t, where t = α∨
7 (ξ27),
and thus n18
i (ζ i(n−1)−i(i−1)/2) · α∨
i (ζ in−i(i−1)/2) · α∨
n−1(ξn(n−1)/2)α∨
w , whence n12
w ) = n12
1 (ξ34)α∨
3 (ξ66)α∨
4 (ξ96)α∨
6 (ξ52)α∨
5 (ξ75)α∨
2 (ξ49)α∨
i=0 α∨
w = t18 = α∨
2 (−1)α∨
5 (−1)α∨
i=1 α∨
7 (−1).
Recall that an automorphism θ is N-regular if g(1) contains a regular nilpotent element
of g, and that an automorphism θ of order m is of maximal rank if dim c · ϕ(m) = rk G,
where ϕ denotes the Euler number of m.
Lemma 4.5. Let w be one of the Weyl group elements on the following list. Then all
representatives nw of w in NG(T ) are T -conjugate and, modulo the centre of G, have the
same order as w. The (unique) Kac diagram corresponding to Int nw, the order of w (and
thus of Int nw) and the rank of Int nw are as given below.
a) G of type G2: w of type G2, A2 or A1 × A1. The respective Kac diagrams are: 111,
011 and 010, the respective orders of w are 6, 3 and 2 and the respective ranks of Int nw are
1, 1 and 2.
b) G of type F4: w of type F4, F4(a1), D4(a1), A2 × A2 and A4
1. The corresponding Kac
diagrams are: 11111, 10101, 10100, 00100 and 01000, the respective orders of w are 12, 6,
4, 3 and 2 and the respective ranks are 1, 1, 2, 2, and 4.
Each such automorphism is N-regular.
Proof. Each w listed in the Lemma is of maximal rank and is a power of the Coxeter element
by inspection of [3, Table 3]. (In particular, the orders and ranks are as given in the lemma.)
Thus all representatives of w are T -conjugate by Lemma 1.3. It remains to check that the Kac
14
diagrams are as stated. For the Coxeter elements this is Lemma 4.4. Suppose on the other
hand that l divides the Coxeter number. Then the l-th power of a Coxeter automorphism is
conjugate to the automorphism which sends e±α (α simple) to ζ ±le±α, where ζ is a primitive
h-th root of unity (h the Coxeter number). (We note that since p is good it doesn't divide
h in type G2 or F4.) In particular, g(0) is conjugate to the subalgebra spanned by t and
all root subspaces gα where α has length divisible by l. Thus, inspecting the possibilities
for automorphisms of order h/l, it is straightforward to determine the Kac diagram in each
case.
Type A2 in G2: this is the square of the Coxeter element, and hence g(0) is conjugate to
t ⊕ g±(2α1+α2).
Type A1 × A1 in G2: this is the cube of a Coxeter element, and hence g(0) is conjugate
to t ⊕ g±(α1+α2) ⊕ g±(3α1+α2).
Type F4(a1) in F4: this is the square of a Coxeter element and hence g(0) is conjugate
to t ⊕ g±1221 ⊕ g±1122.
Type D4(a1) in F4: this is the cube of a Coxeter element and hence after conjugation
g(0) = t ⊕ g±1111 ⊕ g±1120 ⊕ g±0121 ⊕ g±1232. Here 1111 and 0121 generate a root system of
type A2, while 1120 is a long root, thus g(0) is of type A2 × A1.
Type A2 × A2 in F4: this is a Coxeter element raised to the fourth power, hence g(0) is
conjugate to t ⊕ g±1110 ⊕ g±0120 ⊕ g±0111 ⊕ g±1221 ⊕ g±1122 ⊕ g±1242. Here 1110, 0111 generate
a root subsystem of Φ of type A2 while 0120, 1122 generate a root system of type A2.
Type A4
1 in F4: this is the sixth power of a Coxeter element, and hence g(0) is conjugate
to the subalgebra spanned by t and all root spaces gα where α is of even length. But the
roots of even length have basis: 1100, 0011, 0110 and 1120, where the first three span a
subsystem of type C3 and 1120 is a long root.
It is now easily seen that the only possible Kac diagrams with coefficients 0 and 1 which
could correspond to these automorphisms are those stated in the lemma.
In type G2 the only remaining non-trivial conjugacy classes in W are involutions. More-
over, it is easy to see from Lemma 2.5 (or see, for example [19]) that there is a unique class
of involution of a simple group of type G2 and hence Lemma 4.5 gives us a complete list of
positive rank automorphisms. (See Table 1.)
In type F4, we are reduced to studying elements of the Weyl group of the following types:
A2, A2 (order 3), B2, B2 × A1, A3, A3 × A1 (order 4), C3, B3, C3 × A1, D4 (order 6), B4
(order 8).
Consider an admissible diagram Γ = Γ′∪Γ′′, where Γ′ is a union of irreducible subdiagrams
of order m, and Γ′′ is a union of irreducible subdiagrams of order less than m. As before
let Φ1 be the smallest root subsystem of Φ containing all of the roots in Γ′, and let Φ2 be
the set of roots in Φ which are orthogonal to all elements of Φ1. Let Li (i = 1, 2) be the
subgroup of G generated by all root subgroups Uα, α ∈ Φi and let S be the torus generated
by all α∨(k×), α ∈ Φ1. Clearly S is a maximal torus of L1 and ZG(S)′ = L2.
Let w = w′w′′ be an element in the conjugacy class corresponding to Γ, where w′ corre-
sponds to Γ′ and w′′ corresponds to Γ′′. Suppose θ = Int nw, where nwT = w ∈ W . Since T
and S are θ-stable, it follows that L1 and L2 are θ-stable. Moreover, wS = w′S and (Sw)◦
is trivial, thus there exists nw′ ∈ NL1(S) such that θS = Int nw′S. Since (Int n−1
w′ ◦ θ)L1
acts trivially on S, we can replace nw′ by an element of the form nw′s, s ∈ S such that
θL1 = Int nw′L1. Therefore nw = nw′g for some g ∈ ZG(L1). Recall that we assume that
T (0) is maximal in ZG(c)θ.
15
Lemma 4.6. w′′ is trivial.
Proof. We have g ∈ ZG(L1) = Z ′ · ZL2(L1) where Z ′ = {t ∈ T α(t) = 1 for all α ∈ Φ1}.
Note that (Z ′)◦ ·S = T , w′ acts trivially on Z ′ and has finitely many fixed points on S. Hence
S′ = (Z ′)◦ = (T w′)◦ and H = L2S′ = S′L2. Then Z ′ = S′Z(L1) and thus, after replacing
nw′ by an appropriate element of nw′Z(L1) if necessary, we may assume that g ∈ ZH(L1) =
ZL2(L1)S′. Let h ∈ L1 be such that s = hnw′h−1 ∈ S. Then hnwh−1 = shgh−1 = sg.
Moreover, s and g commute and (sg)m = 1, thus g is semisimple. We deduce that there
exists h′ ∈ L2 such that h′gh′−1 = s′ ∈ S′. Thus nw is G-conjugate to ss′ ∈ T . But therefore
nw is conjugate to h−1ss′h = nw′s′. By our assumption of the maximality of T (0), we must
have w = w′ in the first place.
This result eliminates the cases A3 × A1, C3 × A1 and B2 × A1. But it also eliminates the
conjugacy class of type D4 by the same argument, since D4 has an alternative admissible
diagram of type B3 × A1. (This can easily be seen in the Weyl group of type B4, in which
an element of type D4 is a product of a negative 3-cycle and a negative 1-cycle.)
To continue, we make a few observations in a similar vein to Lemma 4.5.
Lemma 4.7. There exists a unique conjugacy class of positive rank automorphism of order
8 in type F4. The corresponding Kac diagram is 11101.
Proof. Let w be an element of W of type B4. We observe that B4 is the unique class in
W of order 8 [3, Tables 3,8]. Moreover, any two representatives nw, nwt of w in NG(T )
w = (nwt)4 = 1 (Lemma 4.4), thus there is a unique
are T -conjugate (Lemma 1.3) and n4
conjugacy class of positive rank automorphisms of order 8.
It remains to check that the
Kac diagram is as indicated in the Lemma. Consider the B4-type subsystem of Φ with
basis { − α = α0, α1, α2, α3}. Then Int nw is conjugate to the automorphism of G which
sends eαi to ζeαi for i = 0, 1, 2, 3, where ζ is a primitive 8-th root of unity. But for such an
automorphism either eα4 7→ ζ −1eα4 and thus g(0) = t ⊕ g±(α3+α4), or eα4 7→ ζ 3eα4, in which
case g(0) = t ⊕ g±1221. Since 11101 is the only Kac diagram of order 8 which has coefficients
1 and 0 and has fixed point subalgebra of type A1, this proves the Lemma.
Lemma 4.8. There are two conjugacy classes of automorphisms of F4 of order 6 and rank
1. They have Kac diagrams 01010 and 11100 and each element of the first (resp. second)
class acts on some maximal torus as a Weyl group element of type C3 (resp. B3).
Proof. We note first of all that there are 4 classes of automorphisms of order 6 with Kac
coefficients 0 and 1: the Kac diagrams 10101, 11100, 01010 and 00011. In addition, we have
seen above that 10101 is the unique class of automorphism of order 6 and rank two. Thus
any automorphism of rank 6 and order 1 is conjugate to Int nw where w = nwT ∈ W is
either of type C3 or of type B3.
Let θ = Int t be a Kac automorphism of type 01010 and consider the root subsystem
Σ of Φ generated by α3 + α4, α2 + α3 and α1. Then Σ is of type C3 and if we let H
be the (Levi) subgroup of G generated by T and the subgroups U±α with α ∈ Σ then θ
stabilizes H and is easily seen to be a Coxeter automorphism of H ′ (since dθ(eα) = ζeα for
α = α3 + α4, α2 + α3, α1). Thus t is H-conjugate to some nw ∈ NG(T ), where w = nwT is of
type C3.
Let θ = Int t be the Kac automorphism for the Kac diagram 11100 and consider the root
subsystem Σ of Φ generated by α1, α2 and α3 + α4. Then Σ is of type B3 and dθ(eα) = ζeα
16
for α = α1, α2, α3 + α4. Thus if we let H be the Levi subgroup of G generated by T and all
α with α ∈ Σ then θH ′ is a Coxeter automorphism. In particular, t is H-conjugate to some
nw ∈ NG(T ), where w = nwT is of type B3.
It remains to show that a Kac automorphism of type 00011 is of zero rank. Here G(0)
is the pseudo-Levi subgroup generated by T and Uα with α = ±α2, ±α1, ±α. In particular,
dim G(0) = 16. We have dim g(1) = 5; thus it will suffice to show that there exists e ∈ g(1)
such that dim ZG(0)(e) = 11. Let e = eα3 + eα4. Then zg(0)(e) is spanned by gα with α =
±α, ±α1, ±1342, −α2, −(α1 + α2), 1242 and a two-dimensional subalgebra of t. In particular,
dim ZG(0)(e) ≤ 11, whence dim G(0) · e ≥ 5. It follows that G(0) · e is dense in g(1) and
therefore θ is of zero rank.
Corollary 4.9. There are two conjugacy classes of automorphisms of F4 of order 3 and rank
1. They have Kac diagrams 11000 and 10001 and each element of the first (resp. second)
class acts on some maximal torus of G as a Weyl group element of type A2 (resp. A2).
Proof. Apart from the (unique) rank 2 automorphism 00100, the two Kac automorphisms
given in the corollary are the only remaining Kac automorphisms of order 3. Let θ be the
square of the Kac automorphism with diagram 01010: then g(0) is isomorphic to so(7, k)⊕k,
with basis of simple roots {α2, 1120, α4}. Thus θ is conjugate to a Kac automorphism with
diagram 10001. The statement in this case now follows since θ must therefore have rank 1,
and the square of an element of type C3 is an element of type A2. Now let θ be the square
of the Kac automorphism with diagram 11100. Then g(0) is isomorphic to sp(6) ⊕ k, with
basis of simple roots {α4, α3, 1220}. Thus, by the same reasoning, the Kac automorphism
with diagram 11000 has rank 1 and corresponds to a Weyl group element of type A2.
Lemma 4.10. There is one conjugacy class of automorphisms in type F4 of order 4 and
rank 1. The corresponding Kac diagram is 01001 and is represented by an element of W of
type B2.
Proof. By Lemma 2.7 the Kac diagram corresponding to a positive rank automorphism
can only have 0s and 1s on the nodes. Moreover, we have already seen in Lemma 4.5(b)
that the Kac automorphism with diagram 10100 is the unique class of automorphisms of
order 4 and rank 2. Hence there are only two other possibilities: 01001 and 00010.
It
is easy to see that the Kac automorphism with diagram 01001 is of positive rank since
e1100 + e0011 + f1122 is semisimple. Hence it is of rank 1. (By considering the root subsystem
generated by 1100 and 0011, we can see that θ is conjugate to an automorphism of the form
Int nw, where w = nwT is an element of W of type B2.) It remains to prove that the Kac
automorphism with diagram 00010 is of zero rank (first proved by Vinberg [21, §9]).
In
this case g(0) ∼= sl(4, k) ⊕ sl(2, k) and g(1), as a g(0)-module, is isomorphic to the tensor
product of the natural representation for sl(4) and the natural representation for sl(2). In
particular, dim g(1) = 8. If we let e = e0010 + f1231 then it is easy to check that zg(0)(e) is
spanned by hα1, eα1, fα1, hα2 + hα3 + hα4, f α, f1342, f1242 + fα4, e1242 + eα4, f1100 and fα2.
Thus dim ZG(0)(e) ≤ 10 and therefore dim G(0) · e ≥ 8. Hence G(0) · e is dense in g(1), and
so g(1) is of zero rank.
This completes the determination of the positive rank automorphisms in type F4. We
will use these results on automorphisms in type F4 to study the positive rank triality au-
tomorphisms in type D4. In [21] all automorphisms of so(2n, C) of the form x 7→ gxg−1,
g ∈ O(2n, k) were studied; this was generalized to (odd) positive characteristic in [12], and
17
it was proved that all such θ-groups have a KW-section. If n ≥ 5 then all automorphisms
have this form; in type D4 this leaves only the automorphisms of type D(3)
4 . Recall that the
set of long roots in a root system of type F4 is a root system of type D4, and correspondingly
W (F4) contains W (D4) as a normal subgroup of index 6. Let G be a simple algebraic group
of type F4, let T be a maximal torus of G, let Φ = Φ( G, T ) and let Φl be the set of long
roots in Φ. If {α1, α2, α3, α4} is a basis of Φ then {α2, α1, α2 + 2α3, α2 + 2α3 + 2α4} is a
basis of Φl. Note that (α2 + 2α3)∨ = α∨
4 , thus
the subgroup of G generated by T and all Uα with α ∈ Φl is isomorphic to Spin(8, k) = G.
The subgroup H of G generated by G and N G(T ) normalizes G and the corresponding map
H → Aut G is surjective. (In fact H is isomorphic to the semidirect product of G by the
symmetric group S3.)
Lemma 4.11. All Kac automorphisms of type D(3)
4 which have Kac coefficients equal to 0
or 1 are of positive rank, with the exception of the Kac automorphism of order 9. These
diagrams are:
3 and (α2 + 2α3 + 2α4)∨ = α∨
2 + α∨
2 + α∨
3 + α∨
(a) 111, which has rank 1 and order 12 and acts as a Coxeter element of F4,
(b) 101, which has rank 2 and order 6 and corresponds to an element of type F4(a1),
(c) 001, which has rank 2 and order 3 and corresponds to an element of type A2 × A2,
(d) 010, which has rank 1 and order 6 and corresponds to an element of type C3,
(e) 100, which has rank 1 and order 3 and corresponds to an element of type A2 (this is
the automorphism γ constructed in §2).
Proof. Let G, G and T be as above. By assumption on T , any automorphism of G is of
the form Int nwG for some nw ∈ N G(T ). Recall by Lemma 2.7 that a Kac diagram which
corresponds to a positive rank automorphism satisfies h(βi) ∈ {0, 1} for i = 0, 1, 2. Moreover,
there are no elements of W (F4) of order 9 and hence the only Kac diagrams which can be of
positive rank are the five diagrams listed in the Lemma. Clearly there is exactly one such Kac
diagram of order 12. Since n12
w = 1 if nw ∈ N G(T ) represents a Coxeter element in W (F4),
there exists at least one automorphism of rank 1 and order 12 and hence this automorphism
has Kac diagram 111. Let us recall our description of the Kac automorphism corresponding
to this class. Let ζ be a primitive 12-th root of unity. Then θ is the automorphism satisfying:
e±α1 7→ ζ ±1e±α3, e±α2 7→ ζ ±1e±α2, e±α3 7→ ζ ±1e±α4, e±α4 7→ ζ ±1e±α1
With this clarification, we can determine which Kac diagrams correspond to θ2 and θ4.
(We note that θ2 and θ4 are necessarily of positive rank, since θ is.) Indeed, the fixed point
subspace for θ2 is spanned by h2, h1+h3+h4 and e±(α1+α2)+ζ ±2e±(α2+α3)−ζ ±4e±(α2+α4). Thus
θ2, which corresponds to an element of W (F4) of type F4(a2), has Kac diagram 101. Similarly,
the fixed point subspace for θ4 is spanned by h2, h1 + h3 + h4, e±α1 + ζ ±4e±α3 + ζ ∓4e±α4,
e±(α1+α2) + ζ ±2e±(α2+α3) − ζ ±4e±(α2+α4) and e±(α1+α2+α3) + e±(α2+α3+α4) + e±(α1+α2+α4). Thus
the dimension of the fixed-point space is 6, and hence the corresponding Kac diagram is 100.
On the other hand, let nw ∈ N G(T ) be such that w = nwT is an element of type C3
(resp. A2) and Int nw, as an automorphism of G, is conjugate to a Kac automorphism with
diagram 01010 (resp. 10001). Then Int nwG has order 6 (resp. 3) and has rank at least 1
(since Lie(T ) ∩ g(1) is non-trivial); but Int nw has rank 1 (as an automorphism of Lie( G)) by
Lemma 4.8 and Corollary 4.9. Thus Int nwG is a rank one automorphism, which must also
be of type D4
since elements of W (F4) of type C3 and A2 are of order 3 modulo W (D4).
Hence there exist automorphisms of order 3 and 6 and of rank 1, which must correspond to
the remaining two Kac diagrams as indicated in the diagram.
(3)
18
5 Calculation of the Weyl group
It remains to calculate the little Weyl group for each of the automorphisms of the previous
section. To begin we recall the following straightforward observation [12, Lemma 4.2]. We
maintain the assumptions on θ, c, T from the previous section.
Lemma 5.1. Let ¯W = W θ/ZW θ(c). Then ¯W acts on c and Wc is a subgroup of ¯W .
We will see that for all θ-groups of type G2, F4 and D(3)
4 , ¯W = Wc. This is not true in
general, see Rk. 5.10(c).
Lemma 5.2. Let Γ be a product of irreducible admissible diagrams of order m and let w ∈ W
be an element of type Γ. Let Φ1 be the smallest root subsystem of G containing all roots in Γ
and let Φ2 be the set of roots in Φ which are orthogonal to Φ1. Let t(1) = {t ∈ t w(t) = ζt}
and let W0 = {w ∈ ZW (w) wt(1) = 1t(1)}. Then W0 contains W (Φ2).
Proof. This is clear from Lemma 3.3.
From now on, fix w and let W1 (resp. W2) denote the subgroup of W generated by all
sα with α ∈ Φ1 (resp. Φ2). Then Lemma 5.2 shows that the order of ¯W (and hence of Wc)
divides #(ZW (w))/#(W2). Use of this straightforward observation will allow us to identify
Wc for all the cases which concern us. Let Ti, im be the subtori of T defined in Lemma 1.1.
We will need the following lemma ([12, Lemma 4.3]).
Lemma 5.3. Let T ′
m. Assuming G is simply-connected, Wc = ¯W .
T ′
m = Qi6=m Ti = {t−1θ(t) t ∈ T }. Suppose {t ∈ Tm = T (0) tm = 1} ⊂
Our first result is for maximal rank automorphisms.
Lemma 5.4. Suppose θ is a maximal rank automorphism. Then W0 is trivial. Thus ¯W =
Wc = W θ.
Proof. By assumption, G is simply-connected. But now ZG(c) is connected, and therefore
equals T . Thus ZNG(T )(c) = ZG(c) = T . This shows that W0 is trivial. Furthermore, T ′
m = T
by assumption on θ, thus any element of ¯W has a representative in G(0) by Lemma 5.3.
Finally, we have the following preparatory lemma, which appeared in [15] in characteristic
zero, and was generalised to positive characteristic in [12, Prop. 5.3]. Recall that a KW-
section for θ is an affine linear subvariety v ⊂ g(1) such that the restriction of the categorical
quotient π : g(1) → g(1)//G(0) to v is an isomorphism.
Lemma 5.5. Suppose θ is an N-regular automorphism. Then k[t]W → k[c]Wc is surjective.
In particular, if θ is inner then the degrees of the generators of k[c]Wc are simply those degrees
of the invariants of g which are divisible by m.
Moreover, θ admits a KW-section.
We have the following application of Lemma 5.4. We identify automorphisms by Weyl
group elements w: that is, w, where θ is conjugate to Int nw as described in Sect. 4. Let Gt
for the t-th group in the Shephard-Todd classification.
19
Lemma 5.6. For the maximal rank automorphisms, the Weyl group is as described in the
following list:
(a) Type G2. Coxeter element: Wc = µ6; A2: Wc = µ6; A1 × A1: Wc = W (G2).
(b) Type F4. Coxeter element: Wc = µ12; B4: Wc = µ8; F4(a1) or A2 × A2: Wc = G5;
D4(a1): Wc = G8; A4
1: Wc = W (F4).
(c) Type D(3)
4 . Type F4: Wc = µ4; F4(a1) or A2 × A2: Wc = G4.
Proof. For the rank 1 inner automorphisms, this follows on reading off the orders of conjugacy
classes (and hence centralizers) in [3]. Moreover, since the maximal rank involutions in type
G2 and F4 lie in the centre of the Weyl group, it remains only to check type D(3)
4 and classes
F4(a1), A2 × A2 and D4(a1) in type F4.
Let W be a Weyl group of type F4 and let Wl be the subgroup of W generated by all sα
with α a long root. Then Wl is isomorphic to the Weyl group of type D4, and is a normal
subgroup of W of index 6. For an element in class F4(a1) or A2 × A2, the centralizer in
W has order 72 by [3]. Moreover, the square of an element in class F4(a1) is an element in
class A2 × A2, thus the little Weyl group for these two cases is equal. There are six cosets
of Wl in W and the factor group is isomorphic to S3. It follows that if w is an element of
W which has order 3 modulo Wl then ZWl(w) is a normal subgroup of ZW (w) of index 3.
In particular, ZWl(w) has order 4 (resp. 24) if w is of Coxeter type (resp. of type F4(a1) or
A2 × A2). Thus it is clear that ZWl(w) ∼= µ4 in the case where w is a Coxeter element of
type F4.
We claim that if w is of type A2 × A2 then ZWl(w) is isomorphic to G4.
Indeed,
#(ZWl(w)) = 24, and thus the only other possibilities are G(6, 3, 2), G(12, 12, 2) or a product
of two cyclic groups [4]. (We require coprimeness of the characteristic here, which is auto-
matic since we assume char k = 0 or char k > 3.) But if w is the fourth power of a Coxeter
element w0 then w3
0 ∈ ZWl(w) is non-central and thus ZWl(w) cannot be commutative. If we
write w as w1w2 = w2w1, where w1 is an element of type A2 and w2 is an element of type
A2 then we can construct a basis {c1, c2} for c such that wi(cj) = ζ δij cj. Moreover, w1 ∈ Wl
and hence there exists an element of ZWl(w) with characteristic polynomial (t − ζ)(t − 1) as
an automorphism of c. Since there is no such element of G(6, 3, 2) or G(12, 12, 2), ZWl(w) is
isomorphic to G4. This proves the remaining cases in type D(3)
4 . But now ZW (w) is a non-
commutative pseudoreflection group of rank 2 which has polynomial generators of degree 6
and 12 by Lemma 5.5, and thus is either G(6, 1, 2), G(12, 4, 2) or G5. Since ZW (w) contains
G4 as a normal subgroup, it follows that it is isomorphic to G5.
For an element of type D4(a1), the centralizer in W has order 96 and has degrees 8 and
12 by Lemma 5.5. Thus the only possibilities for ZW (w) are G(12, 3, 2), G8, G13 or µ8 × µ12
[4]. Since Wl is a normal subgroup of W , ¯W contains a normal subgroup which is isomorphic
to G(4, 2, 2). This rules out µ8 × µ12 and G(12, 3, 2) since for example if ξ is a primitive
12-th root of unity
(cid:18)ξ2 0
ξ(cid:19)(cid:18)0 1
1 0(cid:19)(cid:18)ξ−2
0
0
0
ξ−1(cid:19) =(cid:18)ξ
1 0(cid:19)
0 ξ−1(cid:19)(cid:18)0 1
0
and hence G(4, 2, 2) is not normal in G(12, 3, 2). Moreover, by [5, p. 395], G13 contains no
reflections of order 4. Thus ¯W is equal to G8.
For w ∈ W , let Φ1 be the smallest root subsystem of Φ containing all roots corresponding
to vertices of the admissible diagram for w, let Φ2 be the set of roots in Φ which are orthogonal
20
to all elements of Φ1 and let Wi be the subgroup of W generated by all sα with α ∈ Φi
(i = 1, 2). Let L1 be the Levi subgroup of G generated by T and all Uα with α ∈ Φ1. (See
the discussion in the paragraph preceding Lemma 4.6.) Then L1 is θ-stable and c ⊂ Lie(L′
1),
hence one can also consider the little Weyl group in L1, which is naturally a subgroup of Wc.
Lemma 5.7. For the following automorphisms in type F4, ¯W = Wc is equal to the little
Weyl group one obtains on restricting to the subgroup L1. We have:
(a) Wc = µ6 if w is of type C3 or B3,
(b) Wc = µ4 if w is of type B2.
(c) Wc = µ2 if w is of type A1.
Proof. This is a straightforward observation of the orders of the centralizer and the subgroup
W2 (see [3, Table 8]). If w is of type C3 or B3 then Φ2 has a basis consisting of one element
and hence W2 = W (Φ2) has order 2. But the centralizer of w has order 12, thus the order
of ¯W divides 6. If w is of type B2 then Φ2 is isomorphic to B2 (hence W2 has order 8) and
the centralizer of w has order 32, thus the order of ¯W divides 4. Finally, if w is of type A1
then Φ2 = B3 and hence the order of W2 is 48. Since the order of the centralizer is 96 by
[3], this shows that ¯W = µ2.
We remark that in other types, Wc may not be equal to the group one obtains on re-
stricting to L1.
Lemma 5.8. If θ = Int nw is an automorphism in type F4 such that w is of type A2 or A2
then ¯W = Wc = µ6. In fact, there exists a θ-stable semisimple subgroup L of G of type B3
(if w is of type A2) or type C3 (if w is of type A2) such that c ⊂ Lie(L), each element of Wc
has a representative in L(0), and θL is N-regular.
Proof. This is immediate since if w is of type A2 (resp.
automorphism corresponding to a Weyl group element of type B3 (resp. C3).
A2) then θ is the square of an
Lemma 5.9. Let θ be an automorphism of type D(3)
Wc = ¯W = µ2.
4 with Kac diagram 010 or 100. Then
Proof. We noted in the proof of Lemma 5.6 that if w ∈ W = W (F4) has order 3 modulo
Wl = W (D4) then ZWl(w) has index 3 in ZW (w). Thus in both cases here the group ¯W
of Lemma 5.1 is isomorphic to µ2. It therefore remains only to prove that the non-trivial
element of ¯W has a representative in G(0).
Let w be an element of W of type C3. Then (T w)◦ is of dimension 1 and is generated by
the coroot of a long root element β. The idea here is that we can centralize by β∨(k×) and
obtain a Levi subgroup of G whose Lie algebra contains c, and which has a little Weyl group
isomorphic to µ2. It is easy to see that ZG(β∨(k×)) is of type A1 × A1 × A1. Moreover, w and
w2 permute the 3 subgroups of type A1 transitively. (If, for example, we take w = s2s3s4
in W (F4), then Φl has basis {α2, α1, α2 + 2α3, α2 + 2α3 + 2α4}; β = α, the longest root
in F4, and hence the roots in Φl which are orthogonal to β are ±α2, ±(α2 + 2α3) and
±(α2 + 2α3 + 2α4); furthermore, w(α2) = α2 + 2α3, w(α2 + 2α3) = α2 + 2α3 + 2α4 and
w(α2 + 2α3 + 2α4) = −α2.) Setting L = ZG(β∨(k×))′, it is thus easy to see that c ⊂ Lie(L)
and that NL(0)(c)/ZL(0)(c) ∼= µ2. (Here the Kac automorphism with diagram 010 restricts to
an automorphism of L of the form (g1, g2, g3) 7→ (tg−1
3 , g1, g2); the Kac automorphism with
diagram 100 is conjugate to the square of the Kac automorphism with diagram 010.) Thus
¯W = Wc = µ2 in either case.
21
Table 1: Positive rank automorphisms in type G2
Kac diagram m
6
3
2 A1 × A1
r
N-reg.
1
N-reg.
1
2 W (G2) N-reg.
111
011
010
Wc
µ6
µ6
w
G2
A2
L
Remark 5.10. (a) In the case where the ground field has characteristic zero and G(0) is
semisimple (arbitrary G), the rank and little Weyl group were determined by Vinberg in
[21]. It was shown in [21, Prop. 18] that G(0) is semisimple if and only if the corresponding
Kac diagram has exactly one non-zero entry, which is equal to 1. Our calculations for Wc
agree with [21] in these cases.
(b) Let Gθ
Z = {g ∈ G g−1θ(g) ∈ Z(G)} and let W Z
(c). Then W Z
c
is a subgroup of ¯W . Recall that θ is saturated if Wc = W Z
It is immediate that all
c .
automorphisms in type F4 and G2 are saturated since in both cases the centre is trivial.
Moreover, it is not difficult to show without using our classification that any automorphism
of type D(3)
4
is saturated.
c = NGθ
Z
(c)/ZGθ
Z
(c) There is a strong relationship between Vinberg's construction of Wc and work of Brou´e
and Malle constructing certain pseudo-reflection groups in finite groups of (exceptional) Lie
type [1]. In general, the group constructed by Brou´e and Malle corresponds to our ¯W . It is
possible for Wc to be a proper subgroup of ¯W . For example, if θ = Int nw in type E6, where
w is an element of type D4(a1), then Wc is either G8 or G(4, 1, 2).
(d) In Table 3 we have indicated the action of θL for the cases above using the automor-
phism τ : SL(2, k)3 → SL(2, k)3, (g1, g2, g3) 7→ (t(g3)−1, g1, g2).
Theorem 5.11. Any θ-group of type G2, F4 or D(3)
4 has a KW-section.
Proof. To prove this we will observe that there exists a θ-stable reductive subgroup L of
G such that c ⊂ Lie(L), NL(0)(c)/ZL(0)(c) = Wc and θL is N-regular. Then we can apply
Lemma 5.5. Indeed, L is simply the group L1 (see the discussion before Lemma 4.6) in all
cases except automorphisms of type A2 or A2 in type F4 or those of type C3 or A2 in type
D(3)
4 . In case A2 (resp. A2) in type F4 we can reduce to a group of type B3 (resp. C3) by
(the proof of) Lemma 5.8. In cases C3 and A2 in type D(3)
4 we can reduce to a subgroup of
G isomorphic to SL(2, k)3 as indicated in the proof of Lemma 5.9.
This establishes Popov's conjecture in all types other than type E.
It is possible to
use similar methods to solve the remaining cases, but the calculations required would make
the task very time-consuming. An alternative approach to the problem (for inner automor-
phisms) is to consider representatives nw of (suitable) elements w ∈ W , and to use com-
putational methods to determine the Kac automorphism corresponding to Int nw in each
case. We will return to this question in future work with Benedict Gross, Mark Reeder and
Jiu-Kang Yu.
References
[1] Michel Brou´e and Gunter Malle. Zyklotomische Heckealgebren. Ast´erisque, (212):119 --
189, 1993. Repr´esentations unipotentes g´en´eriques et blocs des groupes r´eductifs finis.
22
Table 2: Positive rank automorphisms in type F4
C3
B3
F4(a1)
w
F4
B4
Kac diagram m
12
8
6
6
6
4 D4(a1)
4
3 A2 × A2
3
3
2
2
11111
11101
10101
01010
11100
10100
01001
00100
11000
10001
01000
00001
A2
A2
A4
1
A1
B2
Wc
µ12
µ8
G5
µ6
µ6
G8
µ4
G5
µ6
µ6
r
1
1
2
1
1
2
1
2
1
1
4 W (F4)
1
µ2
L
N-reg.
Spin(9)
N-reg.
Sp(6)
Spin(7)
N-reg.
Spin(5)
N-reg.
Spin(7)
Sp(6)
N-reg.
θL
Coxeter
Coxeter
Coxeter
Coxeter
positive 3-cycle
positive 3-cycle
short SL(2)
Coxeter
w
F4
Table 3: Positive rank automorphisms in type D(3)
4
Kac diagram m
θL
12
6
6
3 A2 × A2
3
r Wc
1 µ12 N-reg.
2 G4 N-reg.
SL(2)3
1 µ2
2 G4 N-reg.
SL(2)3
1 µ2
111
101
010
001
100
F4(a1)
A2
C3
L
τ
τ 2
23
[2] Michel Brou´e, Gunter Malle, and Jean Michel. Generic blocks of finite reductive
groups. Ast´erisque, (212):7 -- 92, 1993. Repr´esentations unipotentes g´en´eriques et blocs
des groupes r´eductifs finis.
[3] R. W. Carter. Conjugacy classes in the Weyl group. Compositio Math., 25:1 -- 59, 1972.
[4] Allan Clark and John Ewing. The realization of polynomial algebras as cohomology
rings. Pacific J. Math., 50:425 -- 434, 1974.
[5] Arjeh M. Cohen. Finite complex reflection groups. Ann. Sci. ´Ecole Norm. Sup. (4),
9(3):379 -- 436, 1976.
[6] G. M. D. Hogeweij. Almost-classical Lie algebras. I, II. Nederl. Akad. Wetensch. Indag.
Math., 44(4):441 -- 452, 453 -- 460, 1982.
[7] J. Humphreys. Linear Algebraic Groups. Springer-Verlag, New York, 1975.
[8] V. G. Kac. Automorphisms of finite order of semisimple Lie algebras. Funkcional. Anal.
i Prilozen., 3(3):94 -- 96, 1969. English translation: Functional Anal. Appl. 3 (1969), 252 --
254.
[9] Victor G. Kac. Infinite-dimensional Lie algebras. Cambridge University Press, Cam-
bridge, third edition, 1990.
[10] B. Kostant and S. Rallis. Orbits and representations associated with symmetric spaces.
Amer. J. Math., 93:753 -- 809, 1971.
[11] P. Levy. Involutions of reductive Lie algebras in positive characteristic. Adv. Math.,
210(2):505 -- 559, 2007.
[12] P. Levy. Vinberg's θ-groups in positive characteristic and Kostant-Weierstrass slices.
arXiv:0712.1725, 2007.
[13] George J. McNinch and Eric Sommers. Component groups of unipotent centralizers in
good characteristic. J. Algebra, 260(1):323 -- 337, 2003. Special issue celebrating the 80th
birthday of Robert Steinberg.
[14] D. I. Panyushev. Regular elements in spaces of linear representations. II. Izv. Akad.
Nauk SSSR Ser. Mat., 49(5):979 -- 985, 1120, 1985. English translation: Math. USSR-Izv.
27 (1986) 279284.
[15] Dmitri I. Panyushev. On invariant theory of θ-groups. J. Algebra, 283(2):655 -- 670, 2005.
[16] V. L. Popov. Representations with a free module of covariants. Funkcional. Anal. i
Prilozen., 10(3):91 -- 92, 1976. English translation: Functional Anal. Appl. 10 (1976), no.
3, 242 -- 243 (1977).
[17] Alexander Premet. Nilpotent orbits in good characteristic and the Kempf-Rousseau
theory. J. Algebra, 260(1):338 -- 366, 2003. Special issue celebrating the 80th birthday of
Robert Steinberg.
[18] Jean-Pierre Serre. Coordonn´ees de Kac. In Oberwolfach Reports Vol. 3 issue 3, 2006.
24
[19] T. A. Springer. The classification of involutions of simple algebraic groups. J. Fac. Sci.
Univ. Tokyo Sect. IA Math., 34(3):655 -- 670, 1987.
[20] Robert Steinberg. Endomorphisms of linear algebraic groups. Memoirs of the American
Mathematical Society, No. 80. American Mathematical Society, Providence, R.I., 1968.
[21] `E. B. Vinberg. The Weyl group of a graded Lie algebra. Izv. Akad. Nauk SSSR Ser.
Mat., 40(3):488 -- 526, 709, 1976. English translation: Math. USSR-Izv. 10 (1976), 463 --
495 (1977).
25
|
1305.5139 | 5 | 1305 | 2015-02-23T19:11:24 | Rings That Are Morita Equivalent to Their Opposites | [
"math.RA"
] | We consider the following problem: Under what assumptions do one or more of the following are equivalent for a ring $R$: (A) $R$ is Morita equivalent to a ring with involution, (B) $R$ is Morita equivalent to a ring with an anti-automorphism, (C) $R$ is Morita equivalent to its opposite ring. The problem is motivated by a theorem of Saltman which roughly states that all conditions are equivalent for Azumaya algebras. Basing on the recent "general bilinear forms", we present a general machinery to attack the problem, and use it to show that (C)$\iff$(B) when $R$ is semilocal or $\mathbb{Q}$-finite. Further results of similar flavor are also obtained, for example: If $R$ is a semilocal ring such that $\mathrm{M}_{n}(R)$ has an involution, then $\mathrm{M}_{2}(R)$ has an involution, and under further mild assumptions, $R$ itself has an involution. In contrast to that, we demonstrate that (B) does not imply (A). Our methods also give a new perspective on the Knus-Parimala-Srinivas proof of Saltman's Theorem. Finally, we give a method to test Azumaya algebras of exponent $2$ for the existence of involutions, and use it to construct explicit examples of such algebras. | math.RA | math |
RINGS THAT ARE MORITA EQUIVALENT TO THEIR
OPPOSITES
URIYA A. FIRST∗
Abstract. We consider the following problem: Under what assumptions are
one or more of the following equivalent for a ring R: (A) R is Morita equivalent
to a ring with involution, (B) R is Morita equivalent to a ring with an anti-
automorphism, (C) R is Morita equivalent to its opposite ring. The problem
is motivated by a theorem of Saltman which roughly states that all conditions
are equivalent for Azumaya algebras. Based on the recent general bilinear
forms of [10], we present a general machinery to attack the problem, and
use it to show that (C) ⇐⇒ (B) when R is semilocal or Q-finite. Further
results of similar flavor are also obtained, for example: If R is a semilocal ring
such that Mn(R) has an involution, then M2(R) has an involution, and under
further mild assumptions, R itself has an involution.
In contrast to that,
we demonstrate that (B) 6=⇒(A). Our methods also give a new perspective
on the Knus-Parimala-Srinivas proof of Saltman's Theorem. Finally, we give a
method to test Azumaya algebras of exponent 2 for the existence of involutions,
and use it to construct explicit examples of such algebras.
1. Overview
Let R be a ring. This paper is concerned with the question of under what
assumptions are all or some of the following conditions equivalent:
(A) R is Morita equivalent to a ring with involution,
(B) R is Morita equivalent to a ring with an anti-automorphism,
(C) R is Morita equivalent to Rop (the opposite ring of R).
(Actually, we consider a slight refinement that takes into account the type of the
involution/anti-automorphism/Morita equivalence; see section 2.) Note that obvi-
ously (A)=⇒(B)=⇒(C), so one is interested in showing (B)=⇒(A) or (C)=⇒(B).
The motivation for the question comes from Azumaya algebras (we recall the
definition in section 9). Let C be a commutative ring and let A be an Azumaya
C-algebra . It was shown by Saltman [19, Th. 3.1] that:
(i) A is Brauer equivalent to an Azumaya algebra B with an involution of the
first kind if and only if A is Brauer equivalent to Aop.
(ii) If C/C0 is a Galois extension with Galois group G = {1, σ} (σ 6= 1), then
A is Brauer equivalent to an Azumaya algebra B with an involution whose
restriction to C is σ if and only if the corestriction algebra CorC/C0(A) =
(A ⊗ Aσ)G is split (i.e. trivial in the Brauer group of C0).
In case C is semilocal and connected, Saltman also showed that one can take B = A
in (i) and (ii). (The case where C is a field is an earlier result of Albert, e.g. see [1,
∗Einstein Institute of Mathematics, Hebrew University of Jerusalem
E-mail address: [email protected].
Date: November 3, 2018.
2010 Mathematics Subject Classification. 11E39, 16D90, 16H05, 16W10.
Key words and phrases. Morita equivalence, anti-automorphism, involution, bilinear form,
general bilinear form, Brauer group, Azumaya algebra, corestriction, semilocal ring.
This research was supported by an ERC grant #226135 and by the Lady Davis Fellowship
Trust.
1
2
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
Ths. 10.19 & 10.22].) Two Azumaya algebras are Brauer equivalent if and only if
they are Morita equivalent as C-algebras ([3, Cor. 17.2]), so (i) can be understood
as: (C)=⇒(A) for Azumaya algebras, provided the Morita equivalence is "of the first
kind". A simpler proof of Saltman's Theorem was later found by Knus, Parimala
and Srinivas ([13, §4]).
In this paper, we use general bilinear forms, introduced in [10], to give partial
answer to Saltman's problem. More precisely, we show that the conditions (A) and
(B) above can be phrased in terms of existence of certain bilinear forms, and use
this observation to give some positive and negative results, Saltman's Theorem in
particular.
We show that (C)=⇒(B) when R is semilocal or when dimQ(R ⊗Z Q) and
ker(R → R ⊗Z Q) are finite (i.e. when R is Q-finite). Further results of simi-
lar flavor are also obtained. These include:
(1) If R is semiperfect (e.g. artinian) and Morita equivalent to Rop, then S,
the basic ring that is Morita equivalent to R, has an anti-automorphism.
In addition, if R has an involution, then so does M2(S).
(2) Suppose R is semilocal. If Mn(R) has an involution, then so does M2(R),
and under mild assumptions (see Theorem 7.3 below), R itself has an in-
volution.
In the special case that R is a division ring, (2) implies that Mn(R) has an involution
if and only if R has an involution, a result obtained by Albert (e.g. [1, Th. 10.12])
when [R : Cent(R)] < ∞ and by Herstein (e.g. [11, Th. 1.2.2]) in the general case.
In contrast to the above, we show (B) 6=⇒(A) even when R is a finite dimensional
algebra over a field, and even when it has an anti-automorphism fixing the center
pointwise. Whether (C)=⇒(B) in general is still open (and we believe this is not
the case).
We continue by describing the proof of Saltman's Theorem by Knus, Parimala
and Srinivas ([13, §4]) from the perspective of our methods. Namely, we recover this
proof as an application of our characterization of (A) in terms of general bilinear
forms. This suppresses some of the computations of [13, §4]. We also recover
the sharpening of Saltman's Theorem in case the base ring is semilocal. Finally,
we apply our methods to construct non-trivial Azumaya algebras over Dedekind
domains satisfying the conditions of Saltman's Theorem, but not admitting an
involution. (However, these examples are not optimal; see section 11.)
The main problem of the paper was suggested to us by David Saltman himself
after we gave a talk about general bilinear forms at the 10th Brauer Group confer-
ence. The idea to apply the same methods for the construction of Azumaya algebras
of exponent 2 not admitting an involution was communicated to us by Asher Auel,
after posting a previous version of this work on the internet. We deeply thank both
of them for their contribution.
Sections 2 and 3 are preliminaries: Section 2 recalls the basics of Morita theory,
and section 3 recalls general bilinear forms. In section 4, we give a criterion in terms
of bilinear forms to when a ring is Morita equivalent to a ring with an involution
(resp. anti-automorphism). This criterion is the core of this paper and it is used
several times later. In sections 5 -- 7, we show that (C)=⇒(B) under certain finiteness
assumptions, as well as other results of the same flavor. Section 8 demonstrates
that (B) 6=⇒(A). The rest of the paper concerns applications to Azumaya algebras:
Section 9 recalls some facts about Azumaya algebras, in section 10 we show how
our methods can reproduce the proof of Saltman's Theorem given in [13], and
in section 11, we construct Azumaya algebras of exponent 2 without involution.
Finally, section 12 presents some questions that we were unable to answer.
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
3
Notation and conventions: Unless specified otherwise, all rings are assumed
to have a unity and ring homomorphisms are required to preserve it. Given a ring R,
denote its set of invertible elements by R×, its center by Cent(R), and its Jacobson
radical by Jac(R). The n × n matrices over R are denoted by Mn(R). The category
of right R-modules is denoted by Mod-R and the category of f.g. projective right
R-modules is denoted by proj-R. For a subset X ⊆ R, we let CentR(X) denote the
centralizer of X in R. If a module M can be viewed as a module over several rings,
we use MR (resp. RM ) to denote "M , considered as a right (resp. left) R-module".
Endomorphisms of left (right) modules are applied on the right (left). Throughout,
a semisimple ring means a semisimple artinian ring.
2. Morita Theory
In this section, we recall some facts from Morita Theory, and also give a refine-
ment of the main problem of the paper. See [15, §18] or [18, §4.1] for proofs and
further details.
2.1. Morita Theory. Let R be a ring. A right R-module M is called a generator
if every right R-module is an epimorphic image ofLi∈I M for I sufficiently large,
or equivalently, if RR is a summand of M n for some n ∈ N. The module M is called
a progenerator if M is a generator, finitely generated and projective. In this case,
we also call M a (right) R-progenerator.
Let S be another ring. An (S, R)-progenerator is an (S, R)-bimodule P such
that PR is a progenerator and S = End(PR) (i.e. every endomorphism of PR is of
the form p 7→ sp for unique s ∈ S). In this case, SP is also a progenerator and
R = End(SP ).
The rings R, S are said to be Morita equivalent, denoted R ∼Mor S, if the
categories Mod-R and Mod-S are equivalent. Morita's Theorems assert that:
(1) For all equivalences1 F : Mod-S → Mod-R there exists an (S, R)-progenera-
tor P such that F M is naturally isomorphic to M ⊗S P for all M ∈ Mod-S.
(2) Conversely, for any (S, R)-progenerator P the functor (−) ⊗S P : Mod-S →
Mod-R is an equivalence of categories.
(3) There is a one-to-one correspondence between equivalences of categories
F : Mod-S → Mod-R (considered up to natural isomorphism) and iso-
morphism classes of (S, R)-progenerators. The correspondence maps the
composition of two equivalences to the tensor product of the corresponding
progenerators.
Note that (3) allows us to "divide" progenerators. For example, if P is an (S, R)-
progenerator and Q is an (S′, R)-progenerator, then there exists an (S′, S)-pro-
generator V , unique up to isomorphism, such that Q ∼= V ⊗S P . Indeed, if F :
Mod-S → Mod-R and G : Mod-S′ → Mod-R correspond to P and Q, respectively,
then V is the progenerator corresponding to F −1G : Mod-S′ → Mod-S. (In fact,
one can take V = HomR(P, Q).)
Every (S, R)-progenerator P induces an isomorphism σP : Cent(R) → Cent(S)
given by σP (r) = s where s is the unique element of Cent(S) satisfying sp = pr
for all p ∈ P . As σP depends only on the isomorphism class of P , it follows
that any equivalence of categories F : Mod-S → Mod-R induces an isomorphism
σF : Cent(R) → Cent(S).
1 According to textbooks, an equivalence between two categories A and B consists of a quartet
(F, G, δ, ε) such that F : A → B and G : B → A are functors and δ : idA → GF and
ε : idB → F G are natural isomorphisms. We do not need this detailed description here and hence
we only specify F . In this case, the implicit functor G is determined up to natural isomorphism.
4
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
Let C be a commutative ring and assume R and S are C-algebras. We say that
R and S are Morita equivalent as C-algebras or over C, denoted R ∼Mor/C S, if
there exists an equivalence F : Mod-S → Mod-R such that σF (c · 1R) = c · 1S for
all c ∈ C. Equivalently, this means that there exists an (S, R)-progenerator P such
that cp = pc for all p ∈ P and c ∈ C.
If S is an arbitrary ring that is Morita equivalent to R and F : Mod-S → Mod-R
is any equivalence, then we can make S into a C-algebra by letting C act on S via
σF . In this setting, we have S ∼Mor/C R.
2.2. Types. To make the phrasing of some results in the next sections easier, we
now introduce types: Let C be a commutative ring and let R and S be central
C-algebras. (The algebra R is central if the map c 7→ c · 1R : C → Cent(R) is
an isomorphism.) Every equivalence of categories F : Mod-S → Mod-R gives rise
to an isomorphism σF : Cent(R) → Cent(S). As both Cent(R) and Cent(S) are
isomorphic to C, we can realize σF as an automorphism of C, which we call the
type of F .
(For example, when F is of type idC , R is Morita equivalent to S
as C-algebras.) Likewise, the type of an (S, R)-progenerator P is the type of the
equivalence induced by P . Namely, it is the unique automorphism σ of C satisfying
σ(c)p = pc for all p ∈ P , c ∈ C.
Let α be an automorphism or an anti-automorphism of R. The type of α is
defined to be its restriction to C = Cent(R). For example, an involution of R is of
the first kind (i.e. it fixes Cent(R) pointwise) if and only if its type is idC .
We now make an essential sharpening of the problem presented in section 1. Let
R be a ring, let C = Cent(R) and let σ ∈ Aut(C). We look for sufficient conditions
ensuring that some or all of the following are equivalent:
(A) R is Morita equivalent over C to a (necessarily central) C-algebra with
involution of type σ.
(B) R is Morita equivalent over C to a (necessarily central) C-algebra with an
anti-automorphism of type σ.
(C) R is Morita equivalent to Rop via equivalence of type σ (Rop is considered
as a C-algebra in the obvious way).
Again, (A)=⇒(B)=⇒(C),2 so we want to show that (B)=⇒(A) or (C)=⇒(B). Satl-
man's Theorem ([19, Th. 3.1]) for involutions of the first kind can now be phrased
as (C)=⇒(A) when R/C is Azumaya and σ = idC .
2.3. Progenerators and Scalar Extension. We proceed by recalling several
facts about the behavior of progenerators with respect to scalar extension.
Throughout, C is a commutative ring and R is a C-algebra. All tensor products
are taken over C. If σ is an automorphism of C, then Rσ denotes the C-algebra
obtained from R by letting C act via σ. Observe that for all M, N ∈ Mod-R,
HomR(M, N ) admits a (right) C-module structure given by (f c)m = (f m)c (f ∈
HomR(M, N ), c ∈ C, m ∈ M ). In addition, if R′ is another C-algebra and M ′ ∈
Mod-R, then M ⊗M ′ has a right R⊗R′-module structure given by (m⊗m′)(r⊗r′) =
mr ⊗ m′r′. We start with recalling a well-known fact:
Proposition 2.1. Let R′ be a C-algebra and let X, Y ∈ Mod-R, X ′, Y ′ ∈ Mod-R′.
Then there is a natural homomorphism of C-modules
HomR(X, Y ) ⊗ HomR′ (X ′, Y ′) → HomR⊗R′ (X ⊗ X ′, Y ⊗ Y ′)
2 To see that (B)=⇒(C), let S be a central C-algebra admitting an anti-automorphism α of
type σ, and assume that there is an (S, R)-progenerator P of type idC (i.e. S ∼Mor/C R). Let P ′
be the (Rop, S)-bimodule obtained from P by setting rop · p · s = sαpr. Then P ′ is an (Rop , S)-
progenerator, hence P ′ ⊗S P is an (Rop, R)-progenerator, and the latter is easily seen to have
type σ.
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
5
given by setting (f ⊗f ′)(x⊗x′) = f x⊗f ′x′. This homomorphism is an isomorphism
whenever XR, X ′
R′ are finitely generated projective.
Proof. See [8, pp. 14], for instance. Alternatively, one can easily check that the map
in the proposition is an isomorphism in case X = RR and X ′ = R′
R′ . Since the
map is natural (in the functorial sense), it is additive in X and X ′, hence it is an
isomorphism whenever X and X ′ are summands of Rn and R′m, respectively. (cid:3)
As a special case, we get:
Proposition 2.2. Let S be a C-algebra, let RS := R⊗S, and set XS = X ⊗S for all
X ∈ Mod-R. (Observe that XS is a right RS-module via (x ⊗ s)(r ⊗ s′) = xr ⊗ ss′.)
Then for all X, Y ∈ Mod-R, there is a natural homomorphism
HomR(X, Y ) ⊗ S → HomRS (XS, YS)
given by setting (f ⊗ s)(x ⊗ s′) = f x ⊗ ss′ for s, s′ ∈ S, x ∈ X, r ∈ R. This
homomorphism is an isomorphism when XR is finitely generated projective.
Proof. Take R′ = X ′ = Y ′ = S in Proposition 2.1.
(cid:3)
Proposition 2.3. Let S and D be C-algebras and let P be an (S, R)-progenerator
of type σ ∈ Aut(C). Put RD = R ⊗ D, Sσ
D = (Sσ) ⊗ D and PD = P ⊗ D, and
endow PD with an (Sσ
D, RD)-bimodule structure by setting (s ⊗ d)(p ⊗ d′)(r ⊗ d′′) =
(spr) ⊗ (dd′d′′). Then PD is an (Sσ
Proof. Since P has type σ, End(PR) ∼= Sσ as C-algebras. By Proposition 2.2,
EndRD (PD) = HomRD (PD, PD) ∼= HomR(P, P ) ⊗ D = Sσ ⊗ D = Sσ
D. It is routine
to verify that the action of Sσ
D on PD via endomorphisms is the action specified in
the proposition. Finally, that P is a right R-progenerator implies that PD is a right
RD-progenerator, hence PD is an (Sσ
D, RD)-progenerator. The type is verified by
straightforward computation.
(cid:3)
D, RD)-progenerator of type idD.
Proposition 2.4. Let P be an (S, R)-progenerator and let N and M be the prime
radicals (resp. Jacobson radicals) of R and S, respectively. View P := P/P N as
a right R := R/N -module. Then P N = M P , hence P = P/M P admits a left
S-module structure. Furthermore, P is an (S, R)-progenerator.
Proof. By [15, Pr. 18.44], there is an isomorphism between the lattice of R-ideals
and the lattice of (S, R)-submodules of P given by I 7→ P I. Similarly, the ideals of
S correspond to (S, R)-submodules of P via J 7→ JP , hence every ideal I ER admits
a unique ideal J E S such that JP = P I. The ideal J can also be described as
HomR(P, P I). This description implies that S/J = HomR(P, P )/ Hom(P, P I) ∼=
HomR(P, P/P I) ∼= EndR/I (P/P I). Thus, P/P I is an (S/J, R/I)-progenerator.
Choose I = N . Then by [15, Cor. 18.45] (resp. [15, Cor. 18.50]) J = M , so we are
done.
(cid:3)
3. General Bilinear Forms
General bilinear forms were introduced in [10]. In this section, we recall their
basics and record several facts to be needed later. When not specified, proofs can
be found in [10, §2]. Throughout, R is a (possibly non-commutative) ring.
Definition 3.1. A (right) double R-module is an additive group K together with
two operations ⊙0, ⊙1 : K × R → K such that K is a right R-module with respect
to each of ⊙0, ⊙1 and
(k ⊙0 a) ⊙1 b = (k ⊙1 b) ⊙0 a
∀ k ∈ K, a, b ∈ R .
We let Ki denote the R-module obtained by letting R act on K via ⊙i.
6
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
For two double R-modules K, K ′, we define Hom(K, K ′) = HomR(K0, K ′
0) ∩
HomR(K1, K ′
1). This makes the class of double R-modules into an abelian cate-
gory (which is isomorphic to Mod-(R ⊗Z R) and also to the category of (Rop, R)-
bimodules).3
An involution of a double R-module K is an additive bijective map θ : K → K
satisfying θ ◦ θ = idK and
(k ⊙i a)θ = kθ ⊙1−i a
∀a ∈ R, k ∈ K, i ∈ {0, 1} .
Definition 3.2. A (general) bilinear space over R is a triplet (M, b, K) such that
M ∈ Mod-R, K is a double R-module and b : M × M → K is a biadditive map
satisfying
b(xr, y) = b(x, y) ⊙0 r
and
b(x, yr) = b(x, y) ⊙1 r
for all x, y ∈ M and r ∈ R. In this case, b is called a (general) bilinear form (over
R). Let θ be an involution of K. The form b is called θ-symmetric if
b(x, y) = b(y, x)θ
∀x, y ∈ M .
See [10, §2] for various examples of general bilinear forms.
Fix a double R-module K and let i ∈ {0, 1}. The i-K-dual (or just i-dual) of an
R-module M is defined by
M [i] := HomR(M, K1−i) .
Note that M [i] admits a right R-module structure given by (f r)(m) = (f m) ⊙i r
(where f ∈ M [i], r ∈ R and m ∈ M ). In fact, M 7→ M [i] is a left-exact contravariant
functor from Mod-R to itself, which we denote by [i]. Also observe that R[i] ∼= Ki
via f ↔ f (1).
Let b : M × M → K be a (general) bilinear form. The left adjoint map and right
adjoint map of b are defined as follows:
Adℓ
b : M → M [0],
(Adℓ
bx)(y) = b(x, y) ,
Adr
b : M → M [1],
(Adr
bx)(y) = b(y, x) ,
where x, y ∈ M . It is straightforward to check that Adℓ
b are right R-linear.
We say that b is right (resp. left) regular if Adr
b) is bijective. If b is both
right and left regular, we say that b is regular. Left and right regularity are not
equivalent properties; see [10, Ex. 2.6].
b (resp. Adℓ
b and Adr
Assume b is regular. Then for every w ∈ EndR(M ) there exists a unique element
wα ∈ EndR(M ) such that
b(wx, y) = b(x, wαy)
∀x, y ∈ M .
The map w 7→ wα, denoted α, turns out to be an anti-automorphism of EndR(M )
which is called the (right) corresponding anti-automorphism of b. (The left corre-
sponding anti-automorphism of b is the inverse of α.) If b is θ-symmetric for some
involution θ : K → K, then α is easily seen to be an involution.
We say that two bilinear spaces (M, b, K), (M, b′, K ′) are similar if there is an
isomorphism f : K → K ′ such that b′ = f ◦ b. It is easy to see that in this case, b
and b′ have the same corresponding anti-automorphism, provided they are regular.
3 The usage of double R-modules, rather than (Rop , R)-modules or R ⊗Z R-modules, was more
convenient in [10], so we follow the notation of that paper.
In addition, the notion of double
module is more natural when considering bilinear forms as a special case of multilinear forms,
where the form takes values in a (right) multi-R-module.
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
7
Theorem 3.3 ([10, Th. 5.7]). Let M be a right R-generator. Then the map sending
a regular bilinear form on M to its corresponding anti-automorphism gives rise to
a one-to-one correspondence between the class of regular bilinear forms on M , con-
sidered up to similarity, and the anti-automorphisms of EndR(M ). Furthermore,
the correspondence maps the equivalence classes of forms which are θ-symmetric
with respect to some θ to the involutions of EndR(M ).
The theorem implies that every anti-automorphism α of EndR(M ), with M an
R-generator, is induced by some regular bilinear form on M , which is unique up
to similarity. We will denote this form by bα and the double R-module in which it
takes values by Kα. In case α is an involution, then bα is symmetric with respect
to some involution of Kα, which we denote by θα.
The objects Kα, bα and θα can be explicitly constructed as follows: Let W =
EndR(M ). Then M is a left W -module. Using α, we may view M as a right W -
module by defining m · w = wαm. Denote by M α the right W -module obtained.
Then Kα = Mα ⊗W M . We make Kα into a double R-module by defining (x ⊗
y) ⊙0 r = x ⊗ yr and (x ⊗ y) ⊙1 r = xr ⊗ y. The form bα is given by bα(x, y) = y ⊗ x,
and when α is an involution, θα : Kα → Kα is defined by (x ⊗ y)θα = y ⊗ x.
We will also need the following technical proposition.
Proposition 3.4 ([10, Lm. 7.7]). Fix a double R-module K. For M ∈ Mod-R,
define ΦM : M → M [1][0] by (ΦM x)f = f (x) for all x ∈ M and f ∈ M [1]. Then:
(i) {ΦM }M∈Mod-R is a natural transformation from idMod-R to [0][1] (i.e. for
all N, N ′ ∈ Mod-R and f ∈ HomR(N, N ′), one has f [1][0] ◦ ΦN = ΦN ′ ◦ f ).
(ii) Φ is additive (i.e. ΦN ⊕N ′ = ΦN ⊕ ΦN ′ for all N, N ′ ∈ Mod-R).
(iii) (Adr
(iv) R[1][0] can be identified with EndR(K1). Under that identification, (ΦRr)k =
b for every general bilinear form b : M × M → K.
b)[0] ◦ ΦM = Adℓ
k ⊙0 r for all r ∈ R and k ∈ K.
To finish, we recall that the orthogonal sum of two bilinear spaces (M, b, K) and
(M ′, b′, K) is defined to be (M ⊕ M ′, b ⊥ b′, K) where (b ⊥ b′)(x ⊕ x′, y ⊕ y′) =
b(x, y) + b′(x′, y′). The form b ⊥ b′ is right regular if and only if b and b′ are right
regular.
4. Double Progenerators
The observation which forms the basis of all results of this paper is the fact that
whether a ring R is equivalent to a ring with an anti-automorphism (resp. involu-
tion) can be phrased in terms of existence of certain bilinear forms (resp. double
R-modules). In this section, we state and prove this criterion (Proposition 4.3 and
Theorem 4.5).
Throughout, R is a ring and C = Cent(R). Recall from section 2 that for all
M, N ∈ Mod-R, HomR(M, N ) admits a (right) C-module structure. In particular,
EndR(M ) is a C-algebra.
Let K be a double R-module. Then K can be viewed as an (Rop, R)-bimodule
by setting aop · k · b = k ⊙0 b ⊙1 a. If this bimodule is an (Rop, R)-progenerator,
we say that K is a double R-progenerator. The type of a double R-progenerator
is the type of its corresponding (Rop, R)-module. Namely, it is the automorphism
σ ∈ Aut(C) satisfying k ⊙0 c = k ⊙1 σ(c) for all c ∈ Cent(R), k ∈ K.
It turns out that Theorem 3.3 is useful for producing double progenerators.
Lemma 4.1. Let M be an R-progenerator and let α be an anti-automorphism of
EndR(M ). Then Kα is a double R-progenerator. Viewing EndR(M ) as a C-algebra,
the type of Kα is the type of α.
8
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
Proof. Let W = EndR(M ). Recall from section 3 that Kα = M α ⊗W M . Since
MR is an R-progenerator, M is a (W, R)-progenerator. Endow M α with a left Rop-
module structure by setting ropm = mr. Then M α is an (Rop, W )-bimodule and,
moreover, it is easily seen to be an (Rop, W )-progenerator (because M is a (W, R)-
progenerator). By Morita theory (see section 2), this means that Kα = Mα ⊗W M
is an (Rop, R)-progenerator, as required. That the type of Kα is the type of α
follows by straightforward computation.
(cid:3)
Lemma 4.2. Fix a double R-progenerator K. Then:
(i) ΦM is an isomorphism for all M ∈ proj-R (see Proposition 3.4 for the
definition of Φ).
(ii) The (contravariant) functors [0], [1] : proj-R → proj-R are mutual inverses.
(iii) Let (M, b, K) be a bilinear space with M ∈ proj-R. Then b is right regular
if and only if b is left regular.
Proof. (i) Since K is a double R-progenerator, every automorphism of K1 is of
the form k 7→ k ⊙0 r for a unique r ∈ R. Hence, ΦR is an isomorphism by
Proposition 3.4(iv). By the additivity of Φ (Proposition 3.4(ii)), ΦRn is also an
isomorphism for all n. As every M ∈ proj-R admits an M ′ ∈ proj-R with M ⊕M ′ ∼=
Rn, it follows that ΦM is an isomorphism (since ΦM ⊕ ΦM ′ = ΦRn ).
(ii) Recall that R[i] ∼= Ki ∈ proj-R. Since the functor [i] is additive, it follows
that [i] takes proj-R into itself. By (i), Φ : idproj-R → [1][0] is a natural isomorphism.
For all M ∈ Mod-R, define ΨM : M → M [0][1] by (ΨM x)f = f (x), where f ∈ M [0]
and x ∈ M . Then a similar argument shows that Ψ : idproj-R → [0][1] is a natural
isomorphism. It follows that [0] and [1] are mutual inverses.
(iii) By Proposition 3.4(iii), (Adr
b is bijective
b is bijective. The other direction follows by symmetry. (Use the identity
(cid:3)
b, so by (i), that Adr
b)[0] ◦ ΦM = Adℓ
implies Adℓ
(Adℓ
b)[1] ◦ ΨM = Adr
b.)
Proposition 4.3. The ring R is Morita equivalent over C to a (necessarily central)
C-algebra with an anti-automorphism of type σ ∈ Aut(C) if and only if there exists
a regular bilinear space (M, b, K) such that M is an R-progenerator and K is a
double R-progenerator of type σ.
Proof. If (M, b, K) is a regular bilinear space, we may associated to b its correspond-
ing anti-automorphism α : EndR(M ) → EndR(M ). Since M is an R-progenerator,
EndR(M ) ∼Mor/C R (recall that we view EndR(M ) as a C-algebra). In addition,
for all x, y ∈ M and c ∈ C, we have b(x, y) ⊙1 σ(c) = b(x, y) ⊙0 c = b(cx, y) =
b(x, cαy) = b(x, y) ⊙1 cα, so since K1 is faithful (because it is a progenerator), the
type of α is σ.
Conversely, let M be an R-progenerator and assume EndR(M ) has an anti-
automorphism of type σ. Then by Theorem 3.3, bα : M × M → Kα is a regular
bilinear form and Kα is a double R-progenerator of type σ by Lemma 4.1.
(cid:3)
Lemma 4.4. Let K be a double R-module with involution θ. For every M ∈
Mod-R, define uθ,M : M [0] → M [1] by uθ,M (f ) = θ ◦ f . Then uθ,M is a natural
isomorphism of right R-modules.
Proof. By computation.
(cid:3)
The proof of following theorem demonstrates an idea that we wish to stress: One
can construct involutions on rings that are Morita equivalent to R by constructing
symmetric general bilinear forms. Note that all possible involutions are obtained in
this manner by Theorem 3.3. (However, this fails if we limit ourselves to standard
sesquilinear forms. Indeed, the base ring may not even have an anti-automorphism;
see [10, Exs. 2.4 & 2.7], or section 11.)
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
9
Theorem 4.5. The ring R is Morita equivalent over C to a C-algebra with an
involution of type σ ∈ Aut(C) if and only if there exists a double R-progenerator of
type σ admitting an involution.
Proof. Assume MR is an R-progenerator such that EndR(M ) has an involution of
type σ. Then, as in the proof of Proposition 4.3, Kα is a double R-progenerator of
type σ. Since α is an involution, Kα has an involution, namely θα.
Conversely, assume K is a double R-progenerator of type σ with an involution θ.
Let P be any right R-progenerator and let M = P ⊕ P [1]. Define b : M × M → K
by b(x ⊕ f, y ⊕ g) = gx + (f y)θ. The form b is clearly θ-symmetric. We claim that b
is regular. By Lemma 4.2(iii) (or since b is θ-symmetric), it is enough to show only
right regularity. Indeed, a straightforward computation shows that
(cid:21) ∈ HomR(P ⊕ P [1], P [1] ⊕ P [1][1]) .
Adr
b =(cid:20)
0
uθ,P [1] ◦ ΦP
idP [1]
0
As ΦP and uθ,P [1] are both bijective by Lemmas 4.2(i) and 4.4, respectively, so
is Adr
b. Let α : EndR(M ) → EndR(M ) be the anti-endomorphism associated
with b. Then α is an involution (since b is θ-symmetric), and, as in the proof of
Proposition 4.3, α is of type σ.
(cid:3)
Using very similar ideas, one can show the following variation, which states when
a particular ring S that is Morita equivalent to R over C has an anti-automorphism
or an involution of a given type. Compare with [19, Th. 4.2].
Proposition 4.6. Let M be a right R-progenerator. Then EndR(M ) has an anti-
automorphism (involution) if and only if there exists a double R-progenerator K of
type σ (admitting an involution θ) and a right regular (θ-symmetric) bilinear form
b : M × M → K.
5. Rings That Are Morita Equivalent to Their Opposites
Let R be a ring.
In this section, we use Proposition 4.3 to show that under
certain finiteness assumptions, R ∼Mor Rop implies that R is Morita equivalent to
a ring with an anti-automorphism (i.e. (C)=⇒(B)). Whether this holds in general
is still open. Henceforth, we freely consider (Rop, R)-progenerators as double R-
progenerators and vice versa (see section 4).
We begin with two lemmas whose purpose is to show that the functors [0] and
[1] commute with certain scalar extensions.
D , RD)-module as in Proposition 2.3. Then KD is an (Rop
Lemma 5.1. Let K be an (Rop, R)-progenerator of type σ ∈ Aut(Cent(R)), let
C ⊆ Cent(R) be a subring fixed pointwise by σ, and let D/C be a commutative ring
extension. For every P ∈ proj-R, let PD := P ⊗C D ∈ proj-RD. Make KD into an
(Rop
D , RD)-progenerator
and (P [1])D ∼= (PD)[1] for all P ∈ proj-R (the functor [1] is computed with respect
to K in the left hand side and with respect to KD in the right hand side).
Proof. That KD is an (Rop
that (Rop)σ
from Proposition 2.2.
D , RD)-progenerator follows from Proposition 2.3 (observe
D = (Rop)D = (RD)op since σ fixes C). That (P [1])D ∼= (PD)[1] follows
(cid:3)
Lemma 5.2. Let K be an (Rop, R)-progenerator and let N be the prime (resp.
Jacobson) radical of R. For all P ∈ proj-R, define P := P/P N ∼= P ⊗R (R/N ) ∈
[1]
proj-R/N . Then EndR(K) ∼= R
(the functor [1] is computed with respect to K in the left hand side and with respect
to K in the right hand side).
, R)-progenerator and P [1] ∼= P
, K is an (R
op
op
10
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
Proof. By Proposition 2.4, KN = N opK and K is an (R
, R)-progenerator. Let
P ∈ proj-R. We claim that P [1]N = HomR(P, (KN )0) (we consider KN as a double
R-module here). Indeed, by definition, P [1]N = HomR(P, K0)N ⊆ HomR(P, (K ⊙1
N )0) = HomR(P, (N opK)0) = HomR(P, (KN )0), so there is a natural inclusion
P [1]N ⊆ HomR(P, (KN )0). As this inclusion is additive in P , it is enough to verify
the equality for P = RR, which is routine. Now, we get
op
P [1] =
HomR(P, K0)
P [1]N
=
HomR(P, K0)
HomR(P, (KN )0)
∼= HomR(P, K 0) ∼= HomR(P , K0) = P
as required.
[1]
,
(cid:3)
Recall that a ring R is called Q-finite if both dimQ(R ⊗Z Q) and the cardinality
of ker(R → R ⊗Z Q) are finite. The following lemma is inspired by [3, Pr. 18.2].
Lemma 5.3. Assume R is Q-finite and let N be the prime radical of R. Then
N is nilpotent and R/N ∼= T × Λ where T is a semisimple finite ring and Λ is a
subring of a semisimple Q-algebra E such that ΛQ = E.
Proof. Let T = ker(R → R ⊗Z Q). Then T is an ideal of R. Consider T as a non-
unital ring and let J = Jac(T ) (i.e. J is the intersection of the kernels of all right
T -module morphisms T → M where M is a simple right T -module with M T = M ).
Arguing as in [3, Pr. 18.2], we see that J is also an R-ideal. Since J and all its
submodules are finite (as sets), J n = 0 for some n (because M J ⊆ N for any right
T -module M and any maximal submodule N ≤ M ). In particular, J ⊆ N .
Replacing R with R/J, we may assume J = 0. Now, T is semisimple and of finite
length, hence it has a unit e. As er, re ∈ T for all r ∈ R, we see that er = ere = re.
Thus, e ∈ Cent(R) and R ∼= T × (1 − e)R. As Λ := (1 − e)R is torsion-free, it is a
subring of E := Λ ⊗Z Q, which is a f.d. Q-algebra by assumption.
Let I = Jac(E) ∩ Λ. Then I is nilpotent, hence 0 × I ⊆ N . Replacing R by
R/(0 × I), we may assume E semisimple. We are thus finished if we show that the
prime radical of Λ, denoted N ′, is 0 (because then N = 0 × N ′ = 0). Indeed, by
[5, Th. 2.5], Λ is noetherian (here we need E to be semisimple). Thus, N ′ is nil,
hence so is N ′Q E E. But E is semisimple, so we must have N ′ ⊆ N ′Q = 0.
(cid:3)
Theorem 5.4. Assume R is semilocal or Q-finite and let K be an (Rop, R)-
progenerator. Then for every P ∈ proj-R there is n ∈ N such that P ∼= P [1]n
:=
P [1][1]...[1] (n times). When R is semilocal, n is independent of P .
Proof. Let N denote the Jacobson radical of R in case R is semilocal and the
prime radical of R otherwise. In the latter case, N is nilpotent by Lemma 5.3, so
N ⊆ Jac(R) in both cases.
Using the notation of Lemma 5.2, observe that every P ∈ proj-R is the projective
cover of P = P/P N . As projective covers are unique up to isomorphism, we have
P ∼= Q ⇐⇒ P ∼= Q for all P, Q ∈ proj-R. Therefore, using Lemma 5.2, we may
assume N = 0. Thus, R is semisimple or R = T × Λ with Λ ⊗Z Q semisimple and
T finite and semisimple.
Assume R is semisimple and let V1, . . . , Vt be a complete set of simple R-modules
up to isomorphism. By Lemma 4.2(ii), [1] permutes isomorphism classes of f.g.
projective R-modules, and since [1] is additive, it permutes the indecomposable
projective modules, namely, V1, . . . , Vt. Therefore, there is n ∈ N (say, n = t!) such
that V [1]n
∼= Vi for all i. As any P ∈ proj-R is a direct sum of simple modules, we
get P [1]n ∼= P . We also record that length(P [1]) = length(P ).
i
Now assume R = T × Λ as above, and set E := Λ ⊗Z Q and S := T × E.
Let σ ∈ Aut(Cent(R)) be the type of K and let e = (1T , 0) ∈ R. Then e is the
maximal torsion idempotent in Cent(R), hence σ(e) = e. Define C := Ze + Z(1 − e)
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
11
and D := Ze + Q(1 − e)+ ⊆ T × E = S. Then σC = idC and S ∼= R ⊗C D.
Let K ′ = K ⊗C D. Then by Lemma 5.1, K ′ is an (Sop, S)-progenerator and
(PS)[1] ∼= (P [1])S for all P ∈ proj-R (where PS := P ⊗R S).
For every P ∈ proj-R, define j(P ) = length(PS). Since S is semisimple, the
previous paragraphs imply that j(P [1]) = j(P ). Therefore, we are done if we show
that for all m ∈ N there are finitely many isomorphism classes of modules P with
j(P ) = m (for this implies that action of [1] have finite orbits on proj-R/ ∼=, which
yields the theorem). Indeed, thanks to [5, Th. 2.8], up to isomorphism, there are
finitely many Λ-modules of any given Z-rank, and this is easily seen to imply that
there are finitely many isomorphism classes of modules P ∈ proj-R with PS of a
given length, as required.
(cid:3)
Remark 5.5. It is also possible to prove the semilocal case of Theorem 5.4 by using
a result of Facchini and Herbera [9, Cor. 2.13] stating that every semilocal ring R
has only finitely many indecomposable projective modules up to isomorphism, and
every f.g. projective module is a direct sum of these modules.
Corollary 5.6. Let C = Cent(R), σ ∈ Aut(C) and assume R is semilocal or
Q-finite. Then R is Morita equivalent over C to a (central) C-algebra with an anti-
automorphism of type σ if and only if R is Morita equivalent to Rop (as rings) via
Morita equivalence of type σ.
Proof. We only check the nontrivial direction. Assume R is Morita equivalent to
Rop via Morita equivalence of type σ. Then there exists a double R-progenerator
K of type σ. By Theorem 5.4, there exists n ∈ N such that R[1]n ∼= RR. Let
. Then there is an isomorphism f : M → M [1]. This isomorphism
gives rise to a right regular bilinear space (M, b, K) with f = Adr
b, namely, b :
M × M → K is given by b(x, y) = (f y)x. By Lemma 4.2(iii), b is regular, so we
are done by Proposition 4.3.
(cid:3)
M =Ln−1
m=0 R[1]m
The proof of Corollary 5.6 cannot be applied to arbitrary rings since there are
double R-modules K for which M [1] ≇ M for all 0 6= M ∈ proj-R:
Example 5.7. Let F be a field and let R = lim−→ {M2(F )⊗n}n∈N. Then any f.g.
projective right R-module is obtained by scalar extension from a f.g. projective
module over M2(F )⊗n ֒→ R. Using this, it not hard (but tedious) to show that
the monoid (proj-R/ ∼=, ⊕) is isomorphic to (Z[ 1
2 ] ∩ [0, ∞), +). If Vn is the unique
simple projective right module over M2(F )⊗n, then Vn ⊗ R is mapped to 2−n.
Let T denote the transpose involution on M2(F ). Then bT = lim
involution of R. Let K = R2 ∈ proj-R. Then EndR(K) ∼= M2(R) ∼= R and using bT ,
we can identify EndR(K) with Rop, thus making P into an (Rop, R)-progenerator.
We claim that for M ∈ proj-R and n ∈ N, M [1] ∼= M implies M = 0. To see this,
let ϕ1 be the map obtained from [1] by identifying proj-R/ ∼= with Z[ 1
2 ] ∩ [0, ∞).
∼= RR. Therefore, ϕ1(x) = 1
Then ϕ1(2) = 1 because (R2
2 x for all
x ∈ Z[ 1
2 ] ∩ [0, ∞), which means that ϕ1(x) 6= x for all 0 6= x ∈ Z[ 1
{T ⊗n}n∈N is an
R)[1] ∼= K 2
1
−→
2 ] ∩ [0, ∞).
6. Semiperfect Rings
Let R be a semilocal ring that is Morita equivalent to its opposite ring. While
Corollary 5.6 implies that R is Morita equivalent to a ring with an anti-automorphism,
it does not provide any information about this ring. However, when R is semiper-
fect, we can actually point out a specific ring which is Morita equivalent to R and
has an anti-automorphism.
12
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
Recall that a ring R is semiperfect if R is semilocal and idempotents lift modulo
Jac(R) (e.g. if Jac(R) is nil). In this case, the map
P 7→ P/P Jac(R) : proj-R/ ∼= → proj-(R/ Jac(R))/ ∼=
i=1 P mi
i
i=1 P ni
proj-R can be written as P ∼=Lt
particular, RR ∼= Lt
is bijective (e.g. see [18, §2.9] or [4, Th. 2.1]). Thus, up to isomorphism, R admits
finitely many indecomposable projective right R-modules P1, . . . , Pt and any P ∈
i with n1, . . . , nt uniquely determined. In
for some (necessarily positive) m1, . . . , mt. The ring
R is called basic if m1 = · · · = mt = 1, namely, if RR is a sum of non-isomorphic
indecomposable projective modules. It is well-known that every semiperfect ring
R admits a basic ring S that is Morita equivalent to it over Cent(R), and S is
well-determined up to isomorphism as a Cent(R)-algebra. (Indeed, S and the map
Cent(R) → S can be determined from Mod-R; see [15, Prp. 18.37], the preceding
discussion, and [15, Rm. 18.43]. Once S is fixed, the Morita equivalence Mod-S →
Mod-R is not unique in general.) Explicitly, we can take S = EndR(M ), where
M = P1 ⊕ · · · ⊕ Pt. (For example, if R = Mn(L) with L a local ring, then S = L.)
Assume now that there is an (Rop, R)-progenerator K of type σ ∈ Aut(Cent(R)).
Then the functor [1] must permute the isomorphism classes of P1, . . . , Pt (because
they are the only indecomposable modules in proj-R) and hence stabilize M =
P1 ⊕ · · · ⊕ Pt. Therefore, as in the proof of Corollary 5.6, EndR(M ), the basic ring
which is Morita equivalent to R, has an anti-automorphism of type σ. We have
thus obtained the following proposition.
Proposition 6.1. Let R be a semiperfect ring and let S be a basic ring that is
Morita equivalent to R over Cent(R). Then R is Morita equivalent to Rop via an
equivalence of type σ ∈ Aut(Cent(R)) if and only if S has an anti-automorphism
of type σ.
Proposition 6.1 has a version for involutions in which the ring S is replaced with
M2(S).
Proposition 6.2. Let R be a semiperfect ring and let S be a basic ring that is
Morita equivalent to R over Cent(R). If R has an involution of type σ, then so
does M2(S).
Proof. Let α be an involution of R of type σ, and let K be the double R-module
obtained from R by setting k ⊙0 r = rαk and k ⊙1 r = kr (k, r ∈ R). Then K
is double R-progenerator of type σ admitting an involution θ := α. For any P ∈
proj-R, let bP denote the bilinear form b constructed in the proof of Theorem 4.5.
Then (bP , P ⊕ P [1], K) is an α-symmetric regular bilinear space. Let P1, . . . , Pt be
a complete list of indecomposable projective right R-modules up to isomorphism.
Then b := bP1 ⊥ · · · ⊥ bPt is a regular α-symmetric bilinear form defined over
M ⊕ M [1], where M = P1 ⊕ · · · ⊕ Pt. Therefore, EndR(M ⊕ M [1]) has an involution
of type σ, namely, the corresponding anti-automorphism of b. However, we have
seen above that M ∼= M [1], so EndR(M ⊕ M [1]) ∼= M2(EndR(M )) ∼= M2(S) as
Cent(R)-algebras.
(cid:3)
In fact, in many cases, the assumptions of Proposition 6.2 imply that S itself
has an involution. The general statement and its proof will be given in the next
section. However, in case S is a division ring or local with with 2 ∈ S×, there is a
significantly simpler proof, with which we finish this section.
Proposition 6.3. Let L be a local ring and let R = Mn(L). Assume 2 ∈ L× or
L is a division ring. Then R has an involution of type σ if and only if L has an
involution of type σ.
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
13
Proof. That R has an involution when L has an involution is obvious. Conversely,
let α be an involution of R and let K, θ be as in the proof of Proposition 6.2.
Let P be the unique indecomposable projective right R-module. Then necessarily
P ∼= P [1]. Fix an isomorphism f : P → P [1] and observe that the bilinear form
b(x, y) := (f y)x+ ((f x)y)θ (resp. b′(x, y) := (f y)x− ((f x)y)θ) is θ-symmetric (resp.
(−θ)-symmetric). In addition, Adr
Assume 2 ∈ L×. Then f −1◦Adr
b′ = 2 idP . Since the r.h.s. is invertible
and lies in the local ring EndR(P ) ∼= L, one of f −1 ◦ Adr
b′ must be
invertible, hence one of b, b′ is right regular. In any case, we get that L ∼= EndR(P )
has an involution of type σ, namely, the corresponding anti-automorphism of b or
b′.
b + Adr
b +f −1◦Adr
b, f −1 ◦ Adr
b′ = 2f .
When L is a division ring, P is simple, so b′ is regular if Adr
b′ = 0,
then b′ = 0, hence (f y)x = ((f x)y)θ for all x, y ∈ P . This means that the bilinear
form b′′(x, y) := (f y)x is θ-symmetric. As Adr
(cid:3)
b′ 6= 0. If Adr
b′′ = f , b′′ is regular.
Remark 6.4. In case L is a division ring, Proposition 6.3 follows from [11, Th. 1.2.2].
The case where L is also finite dimensional over its center was noted earlier by Al-
bert (e.g. see [1, Th. 10.12]).
7. Transferring Involutions
Motivated by Proposition 6.2, this section concerns the question of whether the
fact that the ring Mn(R) has an involution implies that R has an involution. We
shall provide a positive answer for "most" semilocal rings, but the question is still
open for general semilocal rings, and even for finite dimensional algebras over fields.
For non-semilocal rings, that Mn(R) has an involution does not imply that R has
an anti-automorphism; an example will be given in section 11.
Throughout, R is a ring, C = Cent(R), σ ∈ Aut(C), and n ∈ N is fixed.
We begin by introducing some notation. For an anti-endomorphism γ : R → R,
we define the standard double R-module of (R, γ) to be R endowed with the actions
k ⊙0 r = rγk and k ⊙1 r = kr. Observe that when γ is invertible, this yields a
double R-progenerator of same type as γ. We now recall the following theorem,
which follows from Th. 7.8 in [10] and the comment following the statement. To
make the exposition more self-contained, we present here an ad-hoc proof which is
slightly simpler than the proof given in [10], but not essentially different.
Theorem 7.1. Assume that up to isomorphism M = RR is the only right R-module
satisfying M n ∼= Rn (e.g. when R is semilocal). Then for every anti-automorphism
α : Mn(R) → Mn(R), there exists an anti-automorphism γ : R → R such that Kα
is isomorphic to the standard double R-module of (R, γ). Furthermore, there is an
inner automorphism ϕ ∈ Inn(Mn(R)) such that ϕ ◦ α = Tnγ where Tnγ is defined
by (rij )Tnγ = (rγ
ji).
Proof. Identify Mn(R) with EndR(Rn), and let bα : Rn × Rn → K := Kα be
the bilinear form induced by α. Since bα is regular, we have an isomorphism
bα : Rn → (Rn)[1] ∼= (R[1])n. By assumption, this means that RR ∼= R[1].
Adr
Observe that R[1] = HomR(R, K0) ∼= K1 via f 7→ f (1). Thus, we may identify
RR ∼= R[1] with K1, so now k ⊙1 r = kr for all k, r ∈ K = R. Define γ : R → R
by rγ = 1R ⊙0 r, and note that k ⊙0 r = 1 ⊙0 r ⊙1 k = rγ ⊙1 k = rγk. This easily
implies that γ is an anti-endomorphism of R, and that K is just the standard
double R-module of (R, γ). Since K is a double R-progenerator, we must have
Rop ∼= End(K1) where the isomorphism is given by rop 7→ [k 7→ k ⊙0 r]. As
k ⊙0 r = rγ k, this forces γ to be an anti-automorphism. Finally, we refer the reader
to the proof of [10, Th. 7.8(iv)] for the existence of ϕ.
(cid:3)
14
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
As a corollary, we get the following theorem, which can be regarded as a variation
of Propositions 6.2.
Theorem 7.2. Suppose that up to isomorphism M = RR is the only right R-module
satisfying M n ∼= Rn. If Mn(R) has an involution (resp. anti-automorphism) of type
σ, then M2(R) (resp. R) has an involution (resp. anti-automorphism) of type σ.
Proof. The case where Mn(R) has an anti-automorphism follows from Theorem 7.1.
Let α be an involution of Mn(R). By Theorem 3.3 and Lemma 4.1, Kα is a double
R-progenerator admitting an involution. Thus, by the proof of Theorem 4.5, applied
with P = RR, EndR(R ⊕ R[1]) has an involution of type σ. As in the proof of
Theorem 7.1, R[1] ∼= (Kα)1 ∼= RR, so EndR(R ⊕ R[1]) ∼= EndR(R2) ∼= M2(R) (as
C-algebras).
(cid:3)
We now show that if Mn(R) has an involution, then so does R, under further
assumptions on R. To make the statement more general, we introduce reduced
types.
Set R := R/ Jac(R) and C := Cent(R) (C may be larger than the image of
C = Cent(R) in R). Let S be a ring such that S := S/ Jac(S) is equipped with
a central C-algebra structure (e.g. Rop or Mn(R)). If P is an (S, R)-progenerator,
then by Proposition 2.4, P Jac(R) = Jac(S)P and P := P/P Jac(R) is an (S, R)-
progenerator. Thus, P induces an isomorphism Cent(R) → Cent(S). We may
view this isomorphism as an automorphism σ : C → C which we call the reduced
type of P . Specializing to the case S = Rop, we can define reduced types of
(Rop, R)-progenerators and double R-progenerators. In addition, if α is an anti-
automorphism of S, then the reduced type of α is the automorphism it induces on
Cent(S) ∼= C. It is easy to check that all previous results mentioning types are
also valid for reduced types, provided the following convention: If M is a right
R-progenerator and S = EndR(M ), then C-algebra structure of S is given by the
isomorphism C → Cent(S) induced by the (S, R)-progenerator M := M/M Jac(R).
Theorem 7.3. Let R be a semilocal ring such that Mn(R) has an involution of type
i=1 Mni (Di)
σ ∈ Aut(C) and reduced type σ ∈ Aut(C). Write R := R/ Jac(R) ∼=Qt
where Di is a division ring and ni ∈ N, and identify C = Cent(R) with Qi=1 Fi
where Fi = Cent(Di). We view each Fi as a non-unital subring of C. Assume that
there is 1 ≤ ℓ ≤ t such that for all i ∈ {1, . . . , t} \ {ℓ} we have
(1) Di is not a field, or ni is even, or σFi 6= idFi ,
and for the index ℓ we have
(2) Dℓ is not a field, or nℓ is even, or σFℓ 6= idFℓ, or 2 ∈ D×
ℓ .
Then R has an involution of type σ and reduced type σ.
Proof. Step 1. Note first that for an R-module M , M n ∼= Rn implies M ∼= RR.
Indeed, this is clear when R is semisimple, and we can reduce to this case by
replacing M with M := M/M Jac(R), as explained in the proof of Theorem 5.4.
Let α be an involution of Mn(R) as above. By Theorem 7.1, we may assume Kα
is the standard double R-module of (R, γ) for some anti-automorphism γ : R → R,
necessarily of type (resp. reduced type) σ (resp. σ). Furthermore, by Theorem 3.3,
Kα has an involution θ. Namely, θ : R → R is a map satisfying
(1)
(aγbc)θ = cγbθa
∀a, b, c ∈ R.
Suppose that θ fixes an invertible element u ∈ R×. We claim that the map
β : R → R defined by rβ = u−1rγu is an involution, clearly of the same type and
reduced type as γ. Indeed, observe that for all r ∈ R, (ur)θ = rγ uθ = uu−1rγu =
urβ. Therefore, ur = (ur)θθ = (urβ)θ = urββ, which implies r = rββ.
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
15
We are thus reduced to show that there exists u ∈ R× with uθ = u. Likewise,
we are also done if we can find u ∈ R× with uθ = −u; simply replace θ with −θ.
Step 2. Let J = Jac(R). Clearly J γ = J, hence γ induces an anti-automorphism
on R. We claim that J θ ⊆ J. Indeed, J θ is an ideal by (1) and the bijectivity of
γ, so is enough to show that 1 + J θ = (1θ + J)θ consists of left invertible elements.
Indeed, 1 = (1 · 1)θθ = (1 · 1θ)θ = (1θ)γ · 1θ, hence 1θ is left invertible, and a simple
variation shows that 1θ is also right invertible. Thus, 1θ + J consists of invertible
elements. Now, θ sends right invertible elements to left invertible elements since
sr = 1 implies rγsθ = (sr)θ = 1θ ∈ R×. Therefore, (1θ + J)θ consists of left
invertible elements, as required.
Let γ, θ be the maps induced by γ, θ on R = R/J. In the following steps, we
shall establish the existence of x ∈ R with x + xθ ∈ R× or x − xθ ∈ R×. This would
finish the proof since we can then take u of Step 1 to be x ± xθ. It is enough to
prove the existence of x for R, γ, θ and then lift it arbitrarily to R. Thus, we may
henceforth assume R is semisimple.
Step 3. Let C = Cent(R). Then Cγ = C. Moreover, for all c ∈ C,
c = cθθ = ((1c)θ)θ = (cγ1θ)θ = (1θcγ)θ = cγγ1θθ = cγγ ,
i = fi for all i.
so γC is an involution. Let {e1, . . . , et} be the primitive idempotents of C. As R
is semisimple, we may assume that eiRei = eiR = Mni(Di) (with Di, ni as in the
theorem). Now, γ permutes {e1, . . . , et} and the orbits have size 1 or 2. Choose
representatives {ei}i∈I for the orbits of γ such that 1 ∈ I, and let fi = ei + eγ
if
i
ei 6= eγ
and f γ
i and fi = ei otherwise. Then {fi}i∈I are pairwise orthogonal,Pi∈I fi = 1
particular, θ takes fiRfi = fiR into itself for all i. As R =Q fiR, it is enough to
and take x =Pi∈I x′
ℓ is a unit of fℓR. (Indeed, take x =Pi∈I xi if xℓ + xθ
ℓ ∈ (f1R)×.) We now split into cases.
i with ei 6= eγ
i . Then fiR = eiR × eγ
i ∈ (fiR)× for all i 6= ℓ, and at least one
ℓ ∈ (f1R)×
ℓ − x′θ
Step 4. Suppose fi = ei + eγ
Observe that if f ∈ R satisfies f γ = f , then (f Rf )θ = f γRθf γ−1
find xi, x′
of xℓ + xθ
i ∈ fiR such that xi + xθ
ℓ , x′
ℓ − x′θ
= f Rf . In
i − x′θ
i if x′
i , x′
i = (eiei)θ = eγ
i eθ
i ∈ eγ
i R. We claim that eθ
i is invertible in eγ
i R. Observe
i R. Indeed, we
that eθ
have
i = (eγ
i eγ
i )θθ = (eγγ
eγ
i = eγθθ
i = eγ−1θθ
eγ
i
i )θ = eγθγ
i eγθ
)θθ = (eγ−1θ
eγγθ
i = eγθγ
eγ−1γ−1
)θ = eγ−1γ−1θ
eθ
i ,
i
i
i
i
i
= (eγ−1
i
eγ−1
i
eγ−1θγ−1
i
= eθ
i eγ−1θγ−1
i
,
i is left and right invertible in eγ
hence eθ
repeatedly). Since ei is invertible in eiR, it follows that ei + eθ
invertible in fiR = eiR × eγ
i R (we used (1) and the fact that eγγ
i and ei − eθ
i R. We may therefore take xi = x′
i = ei.
i = ei
i are
Step 5. Assume that fi = ei, and let S := fiR = eiR = Mni (Di). By
Theorem 7.1, there is an anti-automorphism β : Di → Di and v ∈ S× such that
v−1sγv = sTni β for all s ∈ S. Let γ′ = Tniβ and define θ′ : S → S by sθ ′
=
v−1sθvγ−1
. It is straightforward to check that (aγ ′
bc)θ ′
= a for
all a, b, c ∈ S. Furthermore, for all x ∈ S, x ± xθ ′
= v−1(vx + (vx)θ). Therefore,
without loss of generality, we may replace γ, θ with γ′, θ′.
a and aθ ′θ ′
= cγ ′
bθ ′
Let {ejk} be the standard Di-basis of S = Mni (Di). Then γ = Tniβ implies
eγ
jk = ekj . Therefore, (ejj Sekk)θ = eγ
jj = ekkSejj, namely, θ takes ejkDi to
ekjDi. Now, if ni is even, we can take xi = x′
j=1 e(2j)(2j−1). Indeed, both
kkSθeγ−1
i =Pni/2
16
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
xi + xθ
i and x′
i − x′θ
i have matrices of the form
0 ∗
∗ 0
0 ∗
∗ 0
. . .
with all the ∗-s being nonzero, so they are invertible.
Next, embed D := Di in S by sending d ∈ D to the diagonal matrix with diagonal
entries (d, . . . , d). Since γ = Tniβ, the anti-automorphism β of D coincides with
γD. We claim that θ sends D into itself.
Indeed, as a subring of S, D is the
centralizer of {ejk}, and for all d ∈ D and 1 ≤ j, k ≤ ni, we have
ejkdθ = eγ
kjdθ = (dekj )θ = (ekj d)θ = dθeγ−1
i , x′
kj = dθejk .
i − x′θ
It is therefore enough to find xi, x′
i ∈ D×. Since D is a
division ring, the non-existence of xi or x′
i forces θ = ± idD, which in turn implies
β = γD = idD, because then ±d = dθ = (1d)θ = dγ1θ = ±dγ. This is impossible
in case Di is not a field or σ (which is the restriction of γ to Cent(R)) does not fix
Fi = Cent(S) = Cent(D) pointwise.
i ∈ D with xi + xθ
In case i 6= ℓ, either ni is even, or Di is not a field, or σ does not fix Fi element-
wise, so the existence of both xi and x′
i is guaranteed by the previous paragraphs.
In case i = ℓ, ni is odd, Di is a field, and σ fixes Fj element-wise, we must have
char Di 6= 2. Returning to the setting of the previous paragraph, this means that
for any 0 6= y ∈ D, at least one of y + yθ, y − yθ is nonzero (for their sum is 2y 6= 0).
Thus, either x1 or x′
(cid:3)
1 exists, as required.
Corollary 7.4. Let R be a semiperfect ring having an involution of type σ and
i=1 Mni (Di) as in Theorem 7.3 and
reduced type σ. Write R := R/ Jac(R) ∼= Qt
assume that there is 1 ≤ ℓ ≤ t such that for all i ∈ {1, . . . , t} \ {ℓ} we have
(1′) Di is not a field, or σFi 6= idFi ,
and for the index ℓ we have
(2′) Dℓ is not a field, or σFℓ 6= idFℓ, or 2 ∈ D×
ℓ .
Then the basic ring of R has an involution of type σ and reduced type σ.
Proof. Let S be the basic ring of R. By Proposition 6.2, M2(S) has an involution
of type σ and reduced type σ. Using the description S given in section 6 and
i=1 Di.
Therefore, S satisfies the conditions of Theorem 7.3, and hence S has an involution
of type σ and reduced type σ.
(cid:3)
Proposition 2.4, it is easy to see that S is the basic ring of R. That is, S ∼=Qt
Corollary 7.5. Let R be a finite-dimensional algebra over a field F . Suppose
Mn(R) admits an involution α sending F = F · 1R into itself and satisfying αF 6=
idF . Then R has an involution of the same type and reduced type as α. The same
applies to the basic F -algebra that is Morita equivalent to R.
Proof. Let a ∈ F be an element with aα − a 6= 0. Write C = Cent(R/ Jac(R)) =
F1 × · · · × Fn with each Fi a field, and let σ ∈ Aut(C) be the reduced type of α.
Pick some 0 6= x ∈ Fi and view it as an element of C.
If both σ(x) = x, and
σ(ax) = ax, then ax = σ(ax) = σ(a)σ(x) = aαx, which implies (a − aα)x = 0,
a contradiction. Thus, σ does not fix any of the fields Fi element-wise, so we are
done by Theorems 7.3 and Corollary 7.4.
(cid:3)
Remark 7.6. In general, if R, θ, γ are as in the proof of Theorem 7.3, then there
may not exist u ∈ R× with u = uθ or u = −uθ. For example, take R = F ×F with F
a field of characteristic not 2, let α = idR, and let θ be defined by (x, y)θ = (x, −y).
This implies that the proof of Theorem 7.3 cannot be applied to arbitrary semilocal
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
17
rings. In fact, we believe that there exist semilocal rings R without involution such
that M2(R) has an involution.
Remark 7.7. Given R, the datum of θ and γ satisfying (1) is equivalent to the
datum of an anti-structure, a notion introduced by Wall [22, pp. 244] to define
quadratic forms over rings. Recall that an anti-structure on R consists of an anti-
automorphism γ : R → R and an element v ∈ R× such that vγ = v−1 and
rγγ = vrv−1 for all r ∈ R. Fixing γ, there is a one-to-one correspondence between
the possible θ-s and v-s given by v = θ(1) and rθ = rγ v = vrγ−1
. Therefore, the
proof of Theorem 7.2 shows that if R has an anti-structure, then M2(R) has an
involution. Unfolding the construction, this involution is given by
(cid:20) a b
d (cid:21) 7→(cid:20)
c
dγ
v−1cθv−1 aγ−1 (cid:21) =(cid:20)
bθ
dγ
v−1cγ
bγv
aγ−1 (cid:21)
This trick was noted by several authors in the past (see [19, pp. 532], for instance).
The proof of theorem 4.5 can be viewed as a generalization of it.
This means that the question of whether there is a semilocal ring R without
involution such that M2(R) has an involution is equivalent to whether there is a
semilocal ring without an involution admitting an anti-structure. In fact, we were
unable to find in the literature an example of a general ring with these properties.
8. Counterexamples
We now demonstrate that there are "nice" rings which are not Morita equivalent
to rings with an involution (of any kind), but still admit an anti-automorphism. In
the first example, that automorphism fixes the center element-wise.
Example 8.1. Recall that a poset consists of a finite set I equipped with a partial
order which we denote by ≤. For a field F and a poset I, the incidence algebra
F (I) is defined to be the subalgebra of the I-indexed matrices over F spanned as
an F -vector space by the matrix units {eij i, j ∈ I, i ≤ j}. If I is not the disjoint
union of two non-comparable subsets, then Cent(F (I)) = F . In this case, we say I
is connected. The algebra F (I) is well-known to be basic.
Let R be a ring that is Morita equivalent to F (I). The poset I can be recov-
ered from R, up to isomorphism, as follows: Let E denote the set of primitive
idempotents in R. Then R× acts by conjugation on E. Define IR to be the set
of equivalence classes in E, and for i, j ∈ IR, let i ≤ j if and only if eRf 6= 0 for
some (and hence any) e ∈ i and f ∈ j. To see that IR is indeed isomorphic to I,
observe that there is an isomorphism between I and the isomorphism classes of in-
decomposable projective R-modules via i 7→ Pi := eR (e ∈ i), and i ≤ j if and only
if HomR(Pj, Pi) 6= 0. This shows that IR can be determined from Mod-R, hence
IR ∼= IF (I). Finally, IF (I) is easily seen to be isomorphic to I (via sending the class
of eii ∈ F (I) to i). The construction of IR implies that any anti-automorphism
(resp. involution) of R induces an anti-automorphism (resp. involution) on I ∼= IR.
Suppose now that I is connected, has an anti-automorphism, but admits no
involution. Then the previous paragraphs imply that Cent(F (I)) = F , and any
ring that is Morita equivalent to F (I) does not have an involution. On the other
hand, the anti-automorphism of I gives rise to an anti-automorphism of F (I) of
type idF . An example of such a poset was given in [21] by Scharlau (for other
18
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
purposes); I is the 12-element poset whose Hasse diagram is:
•
•
?⑧⑧⑧⑧⑧⑧⑧
•
•
❅❅❅❅❅❅❅
/ •
•
/ •
•
•
•
_❅❅❅❅❅❅❅
⑧⑧⑧⑧⑧⑧⑧
•
•
(Using Scharlau's words, it is "the simplest example I could find".) An anti-
automorphism of I is given by rotating the diagram ninety degrees clockwise.
Involutions and automorphisms of incidence algebras are well-understood in gen-
eral; we refer the reader to the survey [6] and the references therein.
Example 8.2. Various f.d. division algebras admitting an anti-automorphism but
no involution were constructed in [16]. By Proposition 6.3, none of these algebras
is Morita equivalent to a ring with involution. Note that by Albert's Theorem
([1, Th. 10.19]), anti-automorphisms of such division algebras cannot fix the center
pointwise.
9. Azumaya Algebras
The rest of this paper concerns Saltman's Theorem about Azumaya algebras
with involution (see section 1) and the construction of Azumaya algebras without
involution. As preparation, we now briefly recall Azumaya algebras and several
facts about them to be used later. We refer the reader to [20] and [8] for an
extensive discussion and proofs. Throughout, C is a commutative ring and, unless
specified otherwise, all tensor products are taken over C.
9.1. Azumaya Algebras. A C-algebra A is called Azumaya if A is a progenerator
as a C-module and the standard map Ψ : A ⊗ Aop → EndC (A) given by Ψ(a ⊗
bop)(x) = axb is an isomorphism. When C is a field, being Azumaya is equivalent
to being simple and central, so Azumaya algebras are a generalization of central
simple algebras.
Let A, B be Azumaya C-algebras. The following facts are well-known:
(1) Cent(A) = C and CentA⊗B(C ⊗ B) = A ⊗ C.
(2) A ⊗ B and Aop are Azumaya C-algebras.
(3) If ψ : C → C′ is a commutative ring homomorphism, then A ⊗ C′ is an
Azumaya C′-algebra (C′ is viewed as a C-algebra via ψ).
(4) For every C-progenerator P , EndC (P ) is an Azumaya C-algebra.
(5) If B is a subalgebra of A, then B′ := CentA(B) is Azumaya (over C),
B = CentA(B′) and B ⊗ B′ ∼= A via b ⊗ b′ 7→ bb′.
For every M ∈ proj-C, we define rank(M ) = rankC (M ) to be the function
Spec(C) → Z sending a prime ideal P to dimkP (M ⊗ kP ) where kP is the fraction
field of C/P . We write rank(M ) = n to denote that rank(M )(P ) = n for all
P ∈ Spec(C). For example, when C is a field, rank(M ) = dimC (M ). More
generally, rank(M ) is constant when Spec(C) is connected (as a topological space),
or equivalently, when C does not decompose as product of two nonzero rings. In
fact, we can always write C = C(M)
) is
constant. (This follows from the well-known facts that rank(M ) : Spec(C) → Z
is continuous, and Spec(C) is a compact topological space.) Lastly, observe that
such that rank(M ⊗ C(M)
i
1 × · · · × C(M)
t
/
?
O
O
_
o
o
/
O
O
o
o
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
19
if ψ : C → C′ is a commutative ring homomorphism and P ′ ∈ Spec(C′), then
rankC ′(M ⊗ C′)(P ′) = rankC (M )(P ) where P = ψ−1(P ′). This allows us to
extend scalars when computing ranks.
Proposition 9.1. Let A and B be Azumaya C-algebras and let ϕ : A → B be a
homorphism of C-algebras. Then ϕ is injective. If moreover rank(A) = rank(B),
then ϕ is an isomorphism.
Two Azumaya C-algebras A, B are said to be Brauer equivalent, denoted A ∼Br B,
if there are C-progenerators P, Q such that
A ⊗ EndC (P ) ∼= B ⊗ EndC (Q)
as C-algebras. The Brauer class of A, denoted [A], is the collection of Azumaya
algebras which are Brauer equivalent to A. The Brauer group of C, denoted Br(C),
is the set of all Brauer equivalence classes endowed with the group operation [A] ⊗
[B] = [A ⊗ B]. The unit element of Br(C) is the class of C, namely, the class of
Azumaya algebras which are isomorphic to EndC (P ) for some progenerator P . The
inverse of [A] is [Aop].
The following theorem, which is essentially due to Bass, presents an alternative
definition of the Brauer equivalence.
Theorem 9.2 (Bass). Let A, B be two Azumaya C-algebras. Then A ∼Br B ⇐⇒
A ∼Mor/C B, i.e. there exists an (A, B)-progenerator P such that cp = pc for all
p ∈ P and c ∈ C.
Proof. See [3, Cor. 17.2]. The assumption that Spec(C) is noetherian in [3] can be
ignored by [20, Th. 2.3]; see also [19, §1].
(cid:3)
The exponent of an Azumaya algebra A is its order in Br(C).
Theorem 9.3. Let A be an Azumaya C-algebra. Then rank(A) is a square. Fur-
thermore, if rank(A) divides n2, then A⊗n ∼Br C.
We will also need the following proposition.
Proposition 9.4. Let A, B be Azumaya C-algebras and let K be an (A, B)-
bimodule satisfying ck = kc for all c ∈ C. Then K is an (A, B)-progenerator if and
only if K is a (C, Aop ⊗ B)-progenerator with Aop ⊗ B acting via k(aop ⊗ b) = akb.
In particular, in this case, EndC (C K) ∼= Aop ⊗ B.
Proof. The "only if" part easily follows from CentAop⊗B(C ⊗ B) = Aop ⊗ C, so we
turn to prove the "if" part. Let D = A ⊗ Aop and consider A as a right D-module
via a(x ⊗ yop) = yax. Since C A is a progenerator and D ∼= End(C A) (because
A is Azumaya), A is a (C, D)-progenerator. By Proposition 2.3, K ⊗ Aop is an
(A ⊗ Aop, B ⊗ Aop)-progenerator and hence C A ⊗D (K ⊗ Aop) is a (C, B ⊗ Aop)-
progenerator. It is straightforward to check that C A ⊗D (K ⊗ Aop) ∼= K as (C, B ⊗
Aop)-bimodules via x ⊗D (k ⊗ yop) 7→ yxk and k 7→ 1 ⊗D (k ⊗ 1) in the other
direction, so we are done.
(cid:3)
9.2. Galois Extensions. Let G be a finite group of ring-automorphisms of C and
let C0 = CG be the ring of elements fixed by G. For every 2-cocycle f ∈ Z 2(G, C×),4
let ∆(C/C0, G, f ) = Lσ∈G Cuσ where {uσ}σ∈G are formal variables. We make
∆(C/C0, G, f ) into a ring by linearly extending the following relations
uσuτ = f (σ, τ )uστ ,
uσc = σ(c)uσ,
∀σ, τ ∈ G, c ∈ C .
4 Recall that a 2-cocycle of G taking values in C× is a function f : G × G → C× satisfying
σ(f (τ, η))f (σ, τ η) = f (σ, τ )f (στ, η) and f (1, σ) = f (σ, 1) = 1 for all σ, τ, η ∈ G. The 2-cocycles
with the operation of point-wise multiplication form an abelian group denoted Z 2(G, C×).
20
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
This makes ∆(C/C0, G, f ) into a C0-algebra whose unity is u1.
We say that C/C0 is a Galois extension with Galois group G if ∆(C/C0, G, 1) is
an Azumaya C0-algebra (1 denotes the trivial 2-cocycle f (σ, τ ) ≡ 1). In this case:
(1) ∆(C/C0, G, 1) ∼= EndC0 (C) viaPσ cσuσ 7→ [x 7→Pσ cσσ(x)] ∈ EndC0(C).
(2) The C0-algebra ∆(C/C0, G, f ) is Azumaya for all f ∈ Z 2(G, C×).
(3) rankC0(C) = G.
(4) For every commutative C0-algebra D, the extension C ⊗C0 D/C0 ⊗C0 D is
Galois with Galois group {g ⊗C0 idD g ∈ G}.
See [20, §6] for proofs. Azumaya algebras of the form ∆(C/C0, G, f ) are called
crossed products.
Proposition 9.5. Let C/C0 be a Galois extension with Galois group G. Assume
M is a C-module endowed with a G-action such that σ(mc) = σ(m)σ(c) for all
σ ∈ G, m ∈ M , c ∈ C. Then M G ⊗C0 C ∼= M via m ⊗ c 7→ mc, where
M G := {m ∈ M : σ(m) = m ∀σ ∈ G} .
In particular, rankC0(M ) = rankC0(C) rankC0(M G) = G · rankC0(M G). Further-
more, if M is an Azumaya C-algebra and G acts via ring automorphisms on M ,
then M G is an Azumaya C0-algebra.
Proof. See [20, Prps. 6.10 & 6.11].
(cid:3)
9.3. Corestriction. Let C/C0 be a Galois extension with Galois group G. For
every Azumaya C-algebra A and σ ∈ G, define Aσ to be the C-algebra obtained
from A by viewing A as a C-algebra via σ : C → A. The algebra Aσ is also
Azumaya, but it need not be Brauer-equivalent to A. Observe that the algebra
B :=Nσ∈G Aσ admits a G-action given by τ (Nσ aσ) =Nσ aτ −1σ and that action
satisfies σ(bc) = σ(b)σ(c) for all σ ∈ G, b ∈ B, c ∈ C. The corestriction of A (with
respect to C/C0) is defined to be
CorC/C0(A) = BG =(cid:16)Oσ∈G
Aσ(cid:17)G
.
By Proposition 9.5, CorC/C0(A) is an Azumaya C0-algebra. Moreover, it can be
shown that the corestriction induces a group homomorphism
CorC/C0
: Br(C) → Br(C0)
given by CorC/C0([A]) = [CorC/C0(A)]. We also note that the corestriction can be
defined for separable extensions C/C0; see [20, Ch. 8] for further details.
10. Saltman's Theorem
In this section, we show how to recover the Knus-Parimala-Srinivas proof of
Saltman's Theorem ([13, §4]) as an application of Theorem 4.5. In order to keep
the exposition as self-contained and fluent as possible, we will first give a proof of
Saltman's Theorem using Theorem 4.5, and then explain how this proof relates to
the proof in [13].
10.1. Saltman's Theorem. Let C be a commutative ring and let R be a C-
algebra. Recall that a Goldman element of R/C is an element g ∈ R ⊗C R such
that g2 = 1 and g(r ⊗ s) = (s ⊗ r)g for all r, s ∈ R.
Proposition 10.1. All Azumaya C-algebras have a Goldman element
Proof. See [14, p. 112] or [20, Pr. 5.1], for instance.
(cid:3)
Theorem 10.2 (Saltman). Let C be a commutative ring and let A be an Azumaya
C-algebra. Then:
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
21
(i) A is Brauer equivalent to an Azumaya algebra B with an involution of the
first kind if and only if A ∼Br Aop.
(ii) Let C/C0 be a Galois extension with Galois group {1, σ} (σ 6= 1). Then
A is Brauer equivalent to an Azumaya algebra B with an involution whose
restriction to C is σ if and only if CorC/C0(B) ∼Br C0.
Proof. (i) By Theorem 9.2, A ∼Br Aop if and only if there exists a double A-
progenerator of type idC (i.e. an (Aop, A)-progenerator of type idC ). Likewise, by
Theorems 4.5 and 9.2, A is Brauer equivalent to a C-algebra B with involution
of the first kind if and only if there exist a double A-progenerator of type idC
with involution. Therefore, it is enough to show that every double A-progenerator
K of type idC has an involution. Indeed, consider K as an A ⊗C A-module via
k · (a ⊗C a′) = k ⊙0 a ⊙1 a′ (a, a′ ∈ A) and let g be a Goldman element of A. It is
easy to check that k 7→ kg is an involution of K.
(ii) Throughout, D = A ⊗C Aσ and E = DG = CorC/C0(D). Recall that σ
extends to an automorphism of D, also denoted σ, given by σ(a ⊗ a′) = a′ ⊗ a. By
Theorems 4.5 and 9.2, it is enough to prove that CorC/C0(A) ∼Br C0 if and only if
then there exists a double A-progenerator of type σ admitting an involution.
Assume K is a double A-progenerator of type σ with involution θ. Then K can
be considered as an (Aop, Aσ)-progenerator of type idC , hence by Proposition 9.4,
End(C K) ∼= A ⊗ Aσ = D. Extend the left action of C on K to a left action
of ∆ := ∆(C/C0, G, 1) on K by letting uσ act as θ. Then it is easy to see that
End(∆K) ∼= DG = E. Viewing ∆op and E as subrings of End(C0K), this means E is
the centralizer of ∆op in End(C0 K). As all these algebras are Azumaya over C0, we
have ∆op⊗C0 E ∼= End(C0K). In Br(C0) this reads as [C0]⊗[E] = [∆op]⊗[E] = [C0],
so CorC/C0(A) = E ∼Br C0.
Conversely, assume E = CorC/C0(A) ∼Br C0. Then there exists a (C0, E)-
progenerator Q of type idC0. Let K = Q ⊗C0 C. Then by Proposition 2.2, K is a
(C0⊗C0 C, E⊗C0 C)-progenerator. Identify E⊗C0 C with D via (e⊗c) 7→ ec. Then K
is a (C, D)-progenerator, hence by Proposition 9.4, K is an (Aop, Aσ)-progenerator
via aop · k · a′ := k(a ⊗ a′). We may thus view K as a double A-progenerator of
type σ. Let θ = σ ⊗C0 idQ : K → K (so (c ⊗ q)θ = σ(c) ⊗ q). We claim that θ is
an involution of K. Indeed, θ2 = idK and for all c, c′ ∈ C, e ∈ E, q ∈ Q, we have
((c ⊗ q)(c′e))θ = (cc′ ⊗ qe)θ = σ(c)σ(c′) ⊗ qe = (σ(c) ⊗ q)(σ(c′)e) = (c ⊗ q)θ(σ(c′e))
(e = σ(e) since e ∈ E = DG). This implies that (kx)θ = kθσ(x) for all k ∈ K,
x ∈ D. Putting x = 1 ⊗ a and x = a ⊗ 1 yields that θ is an involution.
(cid:3)
Remark 10.3. (i) The proof of Theorem 10.2 also shows that a C-algebra R with
a Goldman element is Morita equivalent to its opposite over C if and only if it is
Morita equivalent as a C-algebra to an algebra with an involution of the first kind.
We could not find a non-Azumaya algebra admitting a Goldman element, though.
(ii) A slightly different proof of the "only if" part of Theorem 10.2(ii) that does
not use double progenerators appears in [19, pp. 531].
In order to explain the connection of the previous proof with the Knus-Parimala-
Srinivas proof of Saltman's Theorem ([13, §4]), let us backtrack our proof to see
what is the algebra B. According to Theorem 4.5, B = EndA(P ⊕ P [1]) with P
an arbitrary A-progenerator, and the involution of B is induced by the bilinear
form bP constructed in the proof of Theorem 4.5. The functor [1] is computed
using a double A-progenerator K with involution which is obtained in the proof of
Theorem 10.2. Let us choose P = AA. Then B ∼= EndA(A ⊕ K1) (since A[1] =
HomA(AA, K0) ∼= K1 via f 7→ f (1A)). This algebra with involution is essentially
the algebra with involution constructed in [13, §4]. However, in [13], the involution
of B is constructed directly, while here we have obtained it from a general bilinear
22
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
form on the right A-module A⊕K1, and hence suppressed some of the computations
of [13].
We also note that many proofs of special cases of Saltman's Theorem in the
literature, [13] and [19] in particular, involve the construction of an involutary
map θ taking some (Aop, A)-bimodule or an equivalent object K into itself and
satisfying (aopkb)θ = bopkθa (a, b ∈ A, k ∈ K). For example, consider the map
α in [19, pp. 532 -- 3, 537], the maps ψ and σP in [13, pp. 71 -- 2] and the map u of
[12, pp. 196]. This hints that these proofs can be effectively phrased using general
bilinear forms.
10.2. Rank Bounds. We proceed by showing that the algebra B in Saltman's
Theorem can be taken to have rank(B) = 4 rank(A) in part (i) and rank(B) =
4 min{rank(A), rank(Aσ)} in part (ii). This follows from the following technical
lemmas.
Lemma 10.4. Let A be an Azumaya C-algebra and let M, N ∈ proj-A. Then
rank(HomA(M, N )) =
rank(M ) · rank(N )
rank(A)
.
Proof. Let P ∈ Spec(C) and let F be an algebraic closure of kP , the fraction field
of C/P . Set AF = A ⊗C F , MF = M ⊗C F and NF = N ⊗C F . By Proposition 2.2,
we have HomAF (MF , NF ) ∼= HomA(M, N ) ⊗C F ∼= (HomA(M, N ) ⊗C kP ) ⊗kP F .
Therefore, it is enough to verify the lemma in case C is an algebraically closed
field, i.e. when C = F , in which case rankC (−) and dimF (−) coincide.
In this
case, we must have A ∼= Mn(F ) (as F -algebras) and M ∼= Ms×n(F ), N ∼= Mt×n(F )
(as A-modules) for suitable s, t, n ∈ N ∪ {0}. But then, dimF HomA(M, N ) =
dimF Mt×s(F ) = ts = dimF (N ) · dimF (M )/ dimF (A), as required.
(cid:3)
Lemma 10.5. Let A be an Azumaya C-algebra and let K be a double A-progenerator
of type σ ∈ Aut(C). Then
rank(K0) =prank(A) rank(Aσ)
and
rank(K1) =qrank(A) rank(Aσ−1 )
Proof. Consider K as an (Aop, A)-bimodule as explained in section 4. Since A is
of type σ, EndA(K0) = End(KA) ∼= (Aσ)op as C-algebras, hence rank(EndA(K0)) =
rank(Aσ). However, by Lemma 10.4, we have rank(EndA(K0)) = rank(K0)2/ rank(A).
Comparing both expressions yields the formula for rank(K0). The formula for
rank(K1) is shown in the same manner.
(cid:3)
Proposition 10.6. In Theorem 10.2(i) (resp. Theorem 10.2(ii)), the algebra B can
be chosen such that rank(B) = 4 rank(A) (resp. rank(B) = 4 min{rank(A), rank(Aσ)}).
Proof. Let σ be idC in the case of Theorem 10.2(i) and a nontrivial Galois automor-
phism of C/C0 in the case of Theorem 10.2(ii). In the comment after Remark 10.3
we saw that B can taken to be EndA(A ⊕ K1) with K a double A-progenerator of
type σ. In this case, by Lemmas 10.5 and 10.4, we have
rank(B) = (cid:16)rank(A) +prank(A) rank(Aσ)(cid:17)2
rank(A)
= 1 +s rank(Aσ)
rank(A) !2
rank(A) .
In the case of Theorem 10.2(i), this means rank(B) = 4 rank(A), so we are done.
For the case of Theorem 10.2(ii), it is enough to show that A is Brauer equivalent to
an Azumaya algebra A′ of rank min{rank(A), rank(Aσ)} (for then we can replace
A with A′ and get the desired rank).
Recall from section 9 that we can decompose C as a product C = C1 × · · · × Ct
such that rank(A ⊗C Ci) and rank(Aσ ⊗C Ci) are constant for all i. Let ei ∈ C
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
23
the standard way. It is easy to see that every C-algebra R factors as a product
be the idempotent satisfying Ci = eiC. We identify Spec(C) with S Spec(Ci) in
Qi eiR and R/C is Azumaya if and only if eiR/eiC is Azumaya for all i. Define
A′ = Qi A′
i = ei(Aσ)op
otherwise. Then A′ ∼Br A (since eiA′ ∼Br eiA for all i), and, by definition,
rankC (A′) = min{rankC (A), rankC (Aσ)}, as required.
(cid:3)
i where A′
i = eiA if rankCi(eiA) < rankCi(ei(Aσ)) and A′
10.3. The Semilocal Case. In [19, §4], Saltman also showed that when C is
semilocal and connected, one can take B = A in Theorem 10.2 (Albert's Theorems
follow as the special case where C is a field). Let us recover the proof using the
language and the tools of the previous sections:
Since C is connected, A has constant rank, say rank(A) = n2. Suppose the
assumptions of parts (i) or (ii) of Theorem 10.2 are satisfied, and set σ = idC for
part (i). By the proof of Theorem 10.2, there is a double A-module K of type
σ admitting an involution θ. We claim that K is the standard double A-module
of some anti-automorphism γ : A → A (see section 7).
Indeed, by assumption,
rank(A) = n2 = rank(Aσ), hence by Lemma 10.5, rank(K1) = rank(A). We now
claim that for P, Q ∈ proj-A, rank(P ) = rank(Q) implies P ∼= Q. By tensoring
with C/ Jac(C), we may assume C is a finite product of fields and A is a finite
product of central simple algebras, in which case the claim is routine. Therefore,
K1 ∼= AA, and arguing as in the proof of Theorem 7.1 yields the desired γ.
We can now apply the proof of Theorem 7.3 with R = A (which is semilocal)
to assert that A has an involution of type σ. However, we need to verify the
assumptions of the theorem, so let ni, Di and Fi be defined as in Theorem 7.3 for
R = A. Since rank(A) = n2, we have dimFi Mni(Di) = n2 for all i. Thus, when
n is even, for all i, either ni is even or Di is a division ring, as required. When n
is odd and σ 6= idC, set A = A/ Jac(A), C = C/ Jac(C) and C0 = C0/ Jac(C0).
It is a standard fact that A/C is Azumaya (so C = Cent(A)) and C/C0 is Galois
with Galois group {σ, id}, where σ is the map induced by σ on C. Noting that
(C) = 2, we see that σ cannot fix any Fi pointwise, so
C = F1 × · · · × Ft and rankC 0
again the assumptions of Theorem 7.3 hold. Finally, in case n is odd and σ = idC ,
we have A⊗2 ∼Br C (by assumption) and A⊗n ∼Br C (by Theorem 9.3, since
rank(A) = n2), which implies A ∼Br C, i.e. A ∼= EndC (P ) for some C-progenerator
P , necessarily of rank n. By what we have shown above, this implies P ∼= Cn, so
A ∼= EndC(P ) ∼= Mn(C) and A has the transpose involution.
We note that specializing in Theorem 7.3 to the case R is an Azumaya C-
algebra and C is connected saves many technicalities in the last steps of the proof.
In particular, one can use the Skolem-Noether Theorem rather than basing on
Theorem 7.1; cf. the argument given by Saltman in [19, pp. 538].
Remark 10.7. Assume C is semilocal or Q-finite and let τ be any automorphism
of C. Then Corollary 5.6 (together with Theorem 9.2) implies that A is Brauer
equivalent to an Azumaya algebra with an anti-automorphism of type τ if and only
if Aop ∼Br Aτ . The semilocal part of this claim is somewhat standard since it is
well-known that Aop ∼Br Aτ implies Aop ∼= Aτ in this case (e.g. see the previous
paragraphs). However, the Q-finite case seems to be unknown. When τ is a Galois
automorphism of order 2 (i.e. C/C{τ } is Galois with Galois group {1, τ }), the
condition Aop ∼Br Aτ is equivalent to A ⊗C Aτ ∼= ResC/C0(CorC/C0 (A)) ∼ C. In
this special case, it could be that the claim is true without any assumption on C.
11. Azumaya Algebras of Exponent 2 Without Involution
Let C be a commutative ring and let A be an Azumaya C-algebra with A ∼Br
Aop. Then Saltman's Theorem asserts that there is some algebra B ∼Br A with
24
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
involution of the first kind. However, it can be very tricky to construct examples of
algebras A with A ∼Br Aop that do not have an involution already. For example,
Saltman have shown in [19, §4] that this is impossible if C is semilocal or rank(A) =
4. In particular, A cannot be a tensor product of quaternion algebras. (The proof
of the semilocal case was recovered in subsection 10.3. In case rank(A) = 4, the
map a 7→ Tr(a) − a is an involution on A, where Tr(a) denotes the reduced trace
of a.) Saltman has pointed out in [19, §4] how to produce examples of Azumaya
algebras without involution that are trivial in Br(C). However, we have not seen
in the literature examples of non-trivial such algebras. Of course, one can always
take the product of a non-trivial quaternion algebra with a trivial Azumaya algebra
without involution, but this results in C being non-connected (i.e. C is a product of
two rings). One is therefore interested in examples in which C is connected. More
generally, one can ask whether the bound rank(B) ≤ 4 rank(A) of Proposition 10.6
is tight.
In this section, we shall use the results of the section 4 to explain how to detect
whether an Azumaya algebra admits an anti-automorphisms. This will be used to
construct an explicit example of a non-trivial Azumaya algebra over a Dedekind
domain C that does not have an involution. However, this example does not show
that the bound rank(B) ≤ 4 rank(A) is optimal, and it also has zero divisors.
(Even worse, we shall see that for C as in the example, and also for many other
low-dimensional domains, we can always choose B to be without zero divisors and
to satisfy rank(B) ≤ rank(A).)
We thank Asher Auel for pointing us to this problem.
Throughout, C is a commutative ring and σ ∈ Aut(C). Unless specified other-
wise, all tensor products are taken over C. Recall that the Picard group, Pic(C),
consists of the isomorphism classes of rank-1 projective C-modules, multiplication
being tensor product (over C). Equivalently, Pic(C) consists of isomorphism classes
of (C, C)-progenerators of type idC . In case C is a Dedekind domain (e.g. a number
ring), Pic(C) is just the ideal class group of C (e.g. see [17]).
Theorem 11.1. Let A be an Azumaya C-algebra such that Aop ∼Br Aσ, let K
be any (Aop, Aσ)-progenerator of type idC (cf. Theorem 9.2), which we also view
as a double A-progenerator (of type σ), let P be a right A-progenerator and set
P [1] = HomA(P, K0). Then the following conditions are equivalent:
(a) B := EndA(P ) has an anti-automorphism of type σ.
(b) There is [I] ∈ Pic(C) such that I ⊗ P ∼= P [1] as A-modules.
(c) EndA(P ) ∼= EndA(P [1]) as C-algebras.
In particular, if I ⊗ P ≇ P [1] for all [I] ∈ Pic(C), then EndA(P ) does not have an
anti-automorphism of type σ.
Proof. (a)=⇒(b): By Proposition 4.6, there exists a double A-progenerator K ′ of
type σ and a regular bilinear form b : P × P → K ′. View K ′ as a right A ⊗ Aσ-
module via k(a ⊗ a′) = k ⊙0 a ⊙1 a′ and view K as a right A ⊗ Aσ in the standard
way. By Proposition 9.4, both K and K ′ are (C, A⊗Aσ ) progenerators. Thus, there
exists a (C, C)-progenerator I such that K ′ = I ⊗C K. As both K and K ′ have type
σ ∈ Aut(C), I is of type idC and hence [I] ∈ Pic(C). Now, by Proposition 2.1, we
have P ∼= Hom(P, K ′
0) ∼= HomC⊗A(C ⊗ P, I ⊗ K0) = HomC (C, I) ⊗ HomA(P, K0) ∼=
I ⊗ P [1].
(b)=⇒(c): We have EndC (I) = C since I is a (C, C)-progenerator of type idC , so
by Proposition 2.1, EndA(P [1]) ∼= EndC⊗A(I ⊗ P, I ⊗ P ) ∼= EndC(I) ⊗ EndA(P ) ∼=
C ⊗ EndA(P ) ∼= EndA(P ). The isomorphism is easily seen to be a C-algebra
isomorphism.
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
25
(c)=⇒(a): Let ψ : EndA(P [1]) → EndA(P ) be an isomorphism. The map
w 7→ w[1] : EndA(P ) → EndA(P [1]) is an anti-homomorphism of rings (since [1]
is an additive contravariant functor). Moreover, it is an anti-isomorphism of rings
since [1] : proj-A → proj-A is a duality of categories (Lemma 4.2(ii)). It is also
straightforward to check that for c ∈ C ⊆ EndA(P ), one has c[1] = σ(c). Thus,
α : EndA(P ) → EndA(P ) defined by wα = ψ(w[1]) is an anti-automorphism of type
σ.
(cid:3)
Remark 11.2. By applying Theorem 11.1 with P = AA, we can test whether A
itself has an involution. In this case, P [1] is just K1 (via f 7→ f (1)), so we only
need to verify that I ⊗ A ≇ K1 as A-modules for all [I] ∈ Pic(C).
We shall now use Theorem 11.1 to construct an Azumaya algebra B over a
Dedekind domain C such that B ⊗ B ∼Br C, B ≁Br C and B does not have a
C-anti-automorphism. For a fractional ideal I of C, we shall write I ⊗n for the n-th
power of I as a fraction ideal, and I ⊕n for the C-module I ⊕ · · · ⊕ I (n times).
The ambiguous notation I n shall be avoided henceforth, except for writing I −1 to
denote the inverse of I. We shall make extensive usage of some well-known facts
about projective modules over Dedekind domain, such as: Every f.g. projective
module over a Dedekind domain C is a direct sum of fractional ideals, and for
i=1 Ji if and only if
i=1 Ii ∼= Ln
fractional ideals I1, . . . , In, J1, . . . , Jn, we have Ln
[Qi Ii] = [Qi Ji] in Pic(C). We refer the reader to [17] for proofs.
Example 11.3. Let C is a Dedekind domain admitting a non-trivial quaternion
Azumaya C-algebra A. By [19, Th. 4.1], A has an involution of the first kind α,
which gives rise to a double A-module K with K1 ∼= AA (see section 7). Viewing
A as a C-module, we can write AC ∼= C⊕3 ⊕ J with J a fractional ideal.
Suppose further that there is [L] ∈ Pic(C) such that [L⊗8] is not a 16-th power
in Pic(C). We claim that the Azumaya C-algebra
B := EndC (C⊕3 ⊕ L) ⊗ A,
does not have an anti-automorphism of type idC, despite clearly satisfying B ⊗
B ∼Br C and B ≁Br C. We show this by applying Theorem 11.1 with P =
(C⊕3 ⊕ L) ⊗ AA. (The module P is a right C ⊗ A-module which we view as a right
A-module in the standard way. We have EndA(P ) ∼= B by Proposition 2.1.) In
fact, we shall show that I ⊗ P ≇ P [1] as C-modules for all [I] ∈ Pic(C).
Indeed, observe that P [1] = HomA(P, K0) = HomC⊗A((C⊕3⊕L)⊗AA, C⊗K0) ∼=
HomC (C⊕3 ⊕ L, C) ⊗ HomA(AA, K0) ∼= (C⊕3 ⊕ L−1) ⊗ K1. Thus, we have
P [1] ∼= (C⊕3 ⊕ L−1) ⊗ K1 ∼= (C⊕3 ⊕ L−1) ⊗ (C⊕3 ⊕ J)
∼= C⊕9 ⊕ (L−1)⊕3 ⊕ J ⊕3 ⊕ L−1J ∼= C15 ⊕ (L−1)⊗4J ⊗4 .
(as C-modules). On the other hand, for [I] ∈ Pic(C),
I ⊗ P = I ⊗ (C⊕3 ⊕ L) ⊗ A ∼= I ⊗ (C⊕3 ⊕ L) ⊗ (C⊕3 ⊕ J)
∼= I ⊕9 ⊕ (IJ)⊕3 ⊕ (IL)⊕3 ⊕ (ILJ) ∼= C⊕15 ⊕ I ⊗16L⊗4J ⊗4 .
It follows that I⊗P ∼= P [1] as C-modules if and only if [I ⊗16L⊗4J ⊗4] = [(L−1)⊗4J ⊗4]
in Pic(C), or equivalently, if [(I −1)⊗16] = [L⊗8]. But this is impossible by the as-
sumption on L.
We remark that if L is chosen such that [L⊗16] = [C], then (C⊕3 ⊕L)⊕16 ∼= C⊕64.
In this case, M16(B) ∼= A ⊗ M16(C), so M16(B) has an involution of type idC . As
a result, one of the algebras B, M2(B), M4(B), M8(B) does not an involution of
type idC while its 2 × 2 matrix algebra does have such involution. In addition, if
Aut(C) = idC (e.g. if the fraction field of C has no nontrivial automorphisms), then
B has no anti-automorphisms at all.
26
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
Explicit choices of C, A and L satisfying all previous conditions (including
[L⊗16] = [C] and Aut(C) = idC) are the following: Let D be the integer ring
of K := Q[x x3 + x + 521 = 0], and take
C = D[2−1],
A = C[i, j i2 = j2 = −1, ij = −ji],
L = h113, x − 16i⊗3 .
That A is Azumaya follows from the general fact (left to the reader) that (a, b)2,C :=
C[i, j i2 = a, j2 = b, ij = −ji] is Azumaya for all a, b ∈ C×, provided 1
2 ∈ C.
The algebra A is non-trivial in Br(C) since K (and hence C) embeds in R, and
A ⊗C R ∼= (−1, −1)2,R is well-known to be a division ring. The class group of D
is a cyclic group of order 48 generated by [h113, x − 16i] (verified using SAGE), so
when viewed as an element of Pic(D), [h113, x − 16i⊗3] has order 16 and its eighth
power is not a 16-th power in Pic(D). Adjoining 1
2 to D does not affect these facts
by the following easy lemma (note that 2 is prime in D).
Lemma 11.4. Let D be a Dedekind domain, let 0 6= p ∈ D be a prime element of
D, and set Dp = D[ 1
p ]. Then the map [I] 7→ [IDp] : Pic(D) → Pic(Dp) (known as
the restriction map) is an isomorphism of groups.
Proof. That [I] 7→ [IDp] is a group homomorphism is straightforward. It is onto
since any ideal I E Dp is easily seen to satisfy (I ∩ D)Dp = I.
To show injectivity, observe that D has Krull-dimension 1, hence pD is maximal.
As a result, every ideal 0 6= J E D can be written as J = pnJ ′ (n ∈ N) with J ′
is coprime to p. The same applies to elements x ∈ D. Also observe that if J is
coprime to p and x ∈ D, the px ∈ J implies x ∈ J. Indeed, px ∈ J ∩ pD = pJ
(since J is coprime to p), hence x ∈ J.
Suppose now that J E D is such that I := JDp is principal. We need to show
that J is principal (i.e. that [J] is trivial in Pic(D)). Replacing J with p−nJ if
needed, we may assume that J is coprime to p. Let z be a generator of I. Again,
replacing z with pnz (n ∈ Z), we may assume that z ∈ J, and since J is coprime
to p, we may further divide z by p unitl it is coprime to p. Let y ∈ J. Then
y = z(pnw) for some n ∈ Z and w ∈ D coprime to p. This implies p−ny = zw, so
n cannot be negative (since the right hand side is not divisible by p). Therefore,
wpn ∈ D and y = zwpn ∈ zD. This means that J = zD, as required.
(cid:3)
Remark 11.5. Let C be a regular integral domain of Krull dimension at most
2, and let F be the fraction field of C. Assume further that every central simple
F -algebra of exponent 2 is Brauer equivalent to a quaternion algebra. For exam-
ple, this is the case when F is a global field by the Albert-Brauer-Hasse-Noether
Theorem, or when F has transcendence degree at most 2 over a separably closed
subfield by [7]. Then every Azumaya C-algebra A of exponent 2 admits an algebra
B ∈ [A] with involution and rankC (B) ≤ rankC (A). In particular, in this case, the
bound rankC(B) ≤ 4 rankC (A) of Proposition 10.6 is not tight. Indeed, by Saltman
[19, Th. 4.1], it is enough to show that B can be chosen to be a quaternion algebra.
As A′ := A ⊗C F has exponent 2, there exists by assumption a quaternion central
simple F -algebra B′ ∼Br A′. Now, by the proof of [2, Pr. 7.4], B′ contains an
Azumaya C-algebra B such that B′ = B ⊗C F . Finally, by [2, Th. 7.2], B ∼Br A,
so we are done.
12. Questions
We finish with several questions that we were unable to answer.
Question 1. Is there a ring R such that R ∼Mor Rop but R is not Morita equivalent
to a ring with an anti-automorphism (cf. section 1)?
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
27
Question 2. Is there a semilocal ring R such that R does not have an involution
despite M2(R) having an involution (cf. Theorem 7.3)?
Question 3. Is there a non-Azumaya algebra admitting a Goldman element (cf.
Remark 10.3(i))?
Question 4. Let C/C0 be a Galois extension with Galois group {1, σ}. Is it true
that all Azumaya C-algebras A with A ∼Br Aσ are Brauer equivalent to an Azumaya
algebra with an anti-automorphism whose restriction to C is σ (cf. Remark 10.7)?
Question 5. Is the bound rank(B) ≤ 4 rank(A) of Proposition 10.6 tight when the
base ring is connected (cf. Remark 11.5)?
Question 6. Are there Azumaya algebras of exponent 2 and without zero-divisors
that do not admit an involution of the first kind?
For our last question, recall that an involution α on a ring R is called hyperbolic
if there exists an idempotent e ∈ R such that e + eα = 1. In all proofs of Salt-
man's Theorem that we have encountered (e.g. [19], [13], and the proof given in
section 10), the constructed involution was hyperbolic. In contrast, Albert's The-
orem guarantees an involution which is non-hyperbolic (and in fact anisotropic);
indeed, we can take B to be a division ring, and any involution on a division ring
is non-hyperbolic. We therefore ask:
Question 7. Let A be an Azumaya C-algebra with A⊗A ∼Br C. Does there always
exist an Azumaya algebra B ∈ [A] having a non-hyperbolic involution of the first
kind?
References
[1] A. Adrian Albert. Structure of algebras. Revised printing. American Mathematical Society
Colloquium Publications, Vol. XXIV. American Mathematical Society, Providence, R.I., 1961.
[2] Maurice Auslander and Oscar Goldman. Maximal orders. Trans. Amer. Math. Soc., 97:1 -- 24,
1960.
[3] H. Bass. K-theory and stable algebra. Inst. Hautes ´Etudes Sci. Publ. Math., (22):5 -- 60, 1964.
[4] Hyman Bass. Finitistic dimension and a homological generalization of semi-primary rings.
Trans. Amer. Math. Soc., 95:466 -- 488, 1960.
[5] E. Bayer-Fluckiger, C. Kearton, and S. M. J. Wilson. Decomposition of modules, forms and
simple knots. J. Reine Angew. Math., 375/376:167 -- 183, 1987.
[6] Rosali Brusamarello and David W. Lewis. Automorphisms and involutions on incidence al-
gebras. Linear Multilinear Algebra, 59(11):1247 -- 1267, 2011.
[7] A. J. de Jong. The period-index problem for the Brauer group of an algebraic surface. Duke
Math. J., 123(1):71 -- 94, 2004.
[8] Frank DeMeyer and Edward Ingraham. Separable algebras over commutative rings. Lecture
Notes in Mathematics, Vol. 181. Springer-Verlag, Berlin, 1971.
[9] Alberto Facchini and Dolors Herbera. Projective modules over semilocal rings. In Algebra and
its applications (Athens, OH, 1999), volume 259 of Contemp. Math., pages 181 -- 198. Amer.
Math. Soc., Providence, RI, 2000.
[10] Uriya A. First. General bilinear forms. Israel Journal of Mathematics, 2014. Currently avail-
able via Springer Link.
[11] I. N. Herstein. Rings with involution. The University of Chicago Press, Chicago, Ill.-London,
1976. Chicago Lectures in Mathematics.
[12] Nathan Jacobson. Finite-dimensional division algebras over fields. Springer-Verlag, Berlin,
1996.
[13] M.-A. Knus, R. Parimala, and V. Srinivas. Azumaya algebras with involutions. J. Algebra,
130(1):65 -- 82, 1990.
[14] Max-Albert Knus and Manuel Ojanguren. Th´eorie de la descente et alg`ebres d'Azumaya.
Lecture Notes in Mathematics, Vol. 389. Springer-Verlag, Berlin, 1974.
[15] T. Y. Lam. Lectures on modules and rings, volume 189 of Graduate Texts in Mathematics.
Springer-Verlag, New York, 1999.
[16] P. J. Morandi, B. A. Sethuraman, and J.-P. Tignol. Division algebras with an anti-
automorphism but with no involution. Adv. Geom., 5(3):485 -- 495, 2005.
28
RINGS THAT ARE MORITA EQUIVALENT TO THEIR OPPOSITES
[17] I. Reiner. Maximal orders, volume 28 of London Mathematical Society Monographs. New
Series. The Clarendon Press Oxford University Press, Oxford, 2003. Corrected reprint of the
1975 original, With a foreword by M. J. Taylor.
[18] Louis H. Rowen. Ring theory. Vol. I, volume 127 of Pure and Applied Mathematics. Academic
Press Inc., Boston, MA, 1988.
[19] David J. Saltman. Azumaya algebras with involution. J. Algebra, 52(2):526 -- 539, 1978.
[20] David J. Saltman. Lectures on division algebras, volume 94 of CBMS Regional Conference
Series in Mathematics. Published by American Mathematical Society, Providence, RI, 1999.
[21] Winfried Scharlau. Automorphisms and involutions of incidence algebras. In Proceedings of
the International Conference on Representations of Algebras (Carleton Univ., Ottawa, Ont.,
1974), Paper No. 24, pages 11 pp. Carleton Math. Lecture Notes, No. 9, Ottawa, Ont., 1974.
Carleton Univ.
[22] C. T. C. Wall. On the axiomatic foundations of the theory of Hermitian forms. Proc. Cam-
bridge Philos. Soc., 67:243 -- 250, 1970.
|
1808.00907 | 1 | 1808 | 2018-08-01T12:46:26 | Degenerations of Leibniz and anticommutative algebras | [
"math.RA"
] | We describe all degenerations of three dimensional anticommutative algebras $\mathfrak{Acom}_3$ and of three dimensional Leibniz algebras $\mathfrak{Leib}_3$ over $\mathbb{C}.$ In particular, we describe all irreducible components and rigid algebras in the corresponding varieties | math.RA | math |
Degenerations of Leibniz and anticommutative algebras 1
Nurlan Ismailova,b, Ivan Kaygorodovc, Yury Volkovd
a Universidade de Sao Paulo, IME, Sao Paulo, Brazil.
b Institute of Mathematics and Mathematical Modeling, Almaty, Kazakhstan.
c Universidade Federal do ABC, CMCC, Santo Andr´e, Brazil.
d Saint Petersburg State University, Saint Petersburg, Russia.
E-mail addresses:
Nurlan Ismailov ([email protected]),
Ivan Kaygorodov ([email protected]),
Yury Volkov (wolf86 [email protected]).
Abstract. We describe all degenerations of three dimensional anticommutative algebras Acom3 and of
three dimensional Leibniz algebras Leib3 over C. In particular, we describe all irreducible components and
rigid algebras in the corresponding varieties.
Keywords:
degeneration, rigid algebra, orbits closure, anticommutative algebra, Leibniz algebra, Lie algebra
2010 Mathematics Subject Classification: 17A32, 14D06, 14L30
1. INTRODUCTION
Degenerations of algebras is an interesting subject, which was studied in various papers (see, for example,
[1 -- 3, 6 -- 9, 12, 15 -- 17, 20 -- 23, 25 -- 27, 29, 30]). In particular, there are many results concerning degenerations
of algebras of low dimensions in a variety defined by a set of identities. One of important problems in this
direction is the description of so-called rigid algebras. These algebras are of big interest, since the closures
of their orbits under the action of the generalized linear group form irreducible components of a variety
under consideration (with respect to the Zariski topology). For example, rigid algebras were classified in
the varieties of low dimensional unital associative, Lie, Jordan and Leibniz algebras [19]. There are fewer
works in which the full information about degenerations was found for some variety of algebras. This
problem was solved for two dimensional pre-Lie algebras in [6], for three dimensional Novikov algebras
in [7], for four dimensional Lie algebras in [9], for four dimensional Zinbiel algebras and nilpotent four-
dimensional Leibniz algebras in [21], for nilpotent five and six dimensional Lie algebras in [15, 30], for
nilpotent five and six dimensional Malcev algebras in [20], and for all two dimensional algebras in [22].
The most well known generalizations of Lie algebras are Leibniz, Malcev and binary Lie algebras. The
Leibniz algebras were introduced as a non-anticommutative generalization of Lie algebras. The study of
the structure theory and other properties of Leibniz algebras was initiated by Loday in [28]. An algebra A
is called a Leibniz algebra if it satisfies the identity
(xy)z = (xz)y + x(yz).
The classification of all three dimensional Leibniz algebras can be found in [29]. Malcev algebras and
binary Lie algebras are anticommutative. Gainov proved that there are no Malcev and binary Lie three
dimensional algebras except Lie algebras [14]. The description of all three dimensional anticommutative
algebras was given in [24] and the central extensions of three dimensional anticommutative algebras were
described in [10]. In this paper we consider anticommutative algebras as a generalization of Lie algebras.
Note that some steps towards a classification of all three dimensional algebras have been done in [5].
In this paper we give the full information about degenerations of three dimensional anticommutative and
Leibniz algebras over C. The vertices of this graph are the isomorphism classes of algebras in the variety
under consideration. An algebra A degenerates to an algebra B if and only if there is a path from the
1The work was supported by FAPESP 17/15437-6, 17/21429-6; AP05131123 "Cohomological and structural problems of
non-associative algebras"; RFBR 18-31-00001; the President's Program "Support of Young Russian Scientists" (grant MK-
1378.2017.1).
1
2
vertex corresponding to A to the vertex corresponding to B. We also describe rigid algebras and irreducible
components in the corresponding varieties.
2. DEFINITIONS AND NOTATION
All spaces in this paper are considered over C, and we write simply dim, Hom and ⊗ instead of dimC,
HomC and ⊗C. An algebra A is a set with a structure of a vector space and a binary operation that induces
a bilinear map from A × A to A.
Given an n dimensional vector space V , the set Hom(V ⊗ V, V ) ∼= V ∗ ⊗ V ∗ ⊗ V is a vector space of
. Indeed, let us fix a basis e1, . . . , en of
i,j ∈ C such that µ(ei ⊗ ej) =
dimension n3. This space has a structure of the affine variety Cn3
V . Then any µ ∈ Hom(V ⊗ V, V ) is determined by n3 structure constants ck
n
ck
i,jek. A subset of Hom(V ⊗V, V ) is Zariski-closed if it can be defined by a set of polynomial equations
Pk=1
in the variables ck
i,j (1 ≤ i, j, k ≤ n).
Let T be a set of polynomial identities. All algebra structures on V satisfying polynomial identities from
T form a Zariski-closed subset of the variety Hom(V ⊗ V, V ). We denote this subset by L(T ). The general
linear group GL(V ) acts on L(T ) by conjugations:
(g ∗ µ)(x ⊗ y) = gµ(g−1x ⊗ g−1y)
for x, y ∈ V , µ ∈ L(T ) ⊂ Hom(V ⊗ V, V ) and g ∈ GL(V ). Thus, L(T ) is decomposed into GL(V )-orbits
that correspond to the isomorphism classes of algebras. Let O(µ) denote the orbit of µ ∈ L(T ) under the
action of GL(V ) and O(µ) denote the Zariski closure of O(µ).
Let A and B be two n dimensional algebras satisfying identities from T and µ, λ ∈ L(T ) represent A
and B respectively. We say that A degenerates to B and write A → B if λ ∈ O(µ). Note that in this case
we have O(λ) ⊂ O(µ). Hence, the definition of a degeneration does not depend on the choice of µ and λ.
If A 6∼= B, then the assertion A → B is called a proper degeneration. We write A 6→ B if λ 6∈ O(µ).
Let A be represented by µ ∈ L(T ). Then A is rigid in L(T ) if O(µ) is an open subset of L(T ). Recall
that a subset of a variety is called irreducible if it cannot be represented as a union of two non-trivial closed
subsets. A maximal irreducible closed subset of a variety is called an irreducible component. It is well
known that any affine variety can be represented as a finite union of its irreducible components in a unique
way. The algebra A is rigid in L(T ) if and only if O(µ) is an irreducible component of L(T ).
We denote by AComn the variety of n dimensional anticommutative algebras and by Leibn the variety of
n dimensional Leibniz algebras.
We use the following notation:
(1) AnnL(A) = {a ∈ A xa = 0 for all x ∈ A} is the left annihilator of A;
(2) A(+2) is the space {xy + yx x, y ∈ A}.
Given spaces U and W , we write simply U > W instead of dim U > dim W .
3. METHODS
In the present work we use the methods applied to Lie algebras in [9, 15, 16, 30]. First of all, if A → B
and A 6∼= B, then Der(A) < Der(B), where Der(A) is the Lie algebra of derivations of A. We will
compute the dimensions of algebras of derivations and will check the assertion A → B only for such
A and B that Der(A) < Der(B). Secondly, if A → C and C → B then A → B. If there is no C
such that A → C and C → B are proper degenerations, then the assertion A → B is called a primary
degeneration. If Der(A) < Der(B) and there are no C and D such that C → A, B → D, C 6→ D
and one of the assertions C → A and B → D is a proper degeneration, then the assertion A 6→ B is
called a primary non-degeneration. It suffices to prove only primary degenerations and non-degenerations
to describe degenerations in the variety under consideration. It is easy to see that any algebra degenerates
to the algebra with zero multiplication. From now on we use this fact without mentioning it.
To prove primary degenerations, we will construct families of matrices parametrized by t. Namely, let A
and B be two algebras represented by the structures µ and λ from L(T ) respectively. Let e1, . . . , en be a
basis of V and ck
i (t) ∈ C
aj
i (t)ej (1 ≤ i ≤ n) form a basis of V for any t ∈ C∗, and the
i,j (1 ≤ i, j, k ≤ n) be the structure constants of λ in this basis. If there exist aj
(1 ≤ i, j ≤ n, t ∈ C∗) such that Et
i =
n
Pj=1
3
structure constants of µ in the basis Et
A → B. In this case Et
1, . . . , Et
n is called a parametrized basis for A → B.
1, . . . , Et
n are such polynomials ck
i,j(t) ∈ C[t] that ck
i,j(0) = ck
i,j, then
Note also the following fact. Let B(α) be a series of algebras parametrized by α ∈ C and e1, . . . , en
be a basis of V . Suppose also that, for any α ∈ C, the algebra B(α) can be represented by a structure
µ(α) ∈ L(T ) having structure constants ck
i,j(t) ∈ C[t] for all
1 ≤ i, j, k ≤ n. Let A be an algebra such that A → B(α) for α ∈ C\S, where S is a finite subset of C. Then
A → B(α) for all α ∈ C. Indeed, if λ ∈ L(T ) represents A, then we have µ(α) ∈ {µ(β)}β∈C\S ⊂ O(λ)
for any α ∈ C. Thus, to prove that A → B(α) for all α ∈ C we will construct degenerations that are valid
for all but finitely many α.
i,j(α) ∈ C in the basis e1, . . . , en, where ck
Let us describe the methods for proving primary non-degenerations. The main tool for this is the follow-
ing lemma.
Lemma 1 ( [15]). Let B be a Borel subgroup of GL(V ) and R ⊂ L(T ) be a B-stable closed subset. If
A → B and A can be represented by µ ∈ R then there is λ ∈ R that represents B.
In particular, it follows from Lemma 1 that A 6→ B in the following cases:
(1) AnnL(A) > AnnL(B);
(2) A(+2) < B(+2).
In the cases where all of these criteria cannot be applied to prove A 6→ B, we will define R by a set of
polynomial equations and will give a basis of V , in which the structure constants of µ give a solution to all
these equations. We will omit everywhere the verification of the fact that R is stable under the action of
the subgroup of upper triangular matrices and of the fact that λ 6∈ R for any choice of a basis of V . These
verifications can be done by direct calculations.
If the number of orbits under the action of GL(V ) on L(T ) is finite, then the graph of primary degener-
ations gives the whole picture. In particular, the description of rigid algebras and irreducible components
can be easily obtained. Since the variety Leib3 contains infinitely many non-isomorphic algebras, we have
to fulfill some additional work. Let A(∗) := {A(α)}α∈I be a set of algebras, and let B be another algebra.
Suppose that, for α ∈ I, A(α) is represented by the structure µ(α) ∈ L(T ) and B ∈ L(T ) is represented
by the structure λ. Then A(∗) → B means λ ∈ {O(µ(α))}α∈I, and A(∗) 6→ B means λ 6∈ {O(µ(α))}α∈I.
Let A(∗), B, µ(α) (α ∈ I) and λ be as above. To prove A(∗) → B it is enough to construct a family of
pairs (f (t), g(t)) parametrized by t ∈ C∗, where f (t) ∈ I and g(t) ∈ GL(V ). Namely, let e1, . . . , en be a
i : C∗ → C
basis of V and ck
aj
i (t)ej (1 ≤ i ≤ n) form a basis of V for any t ∈ C∗, and
i,j (1 ≤ i, j, k ≤ n) be the structure constants of λ in this basis. If we construct aj
(1 ≤ i, j ≤ n) and f : C∗ → I such that Et
i =
n
the structure constants of µf (t) in the basis Et
then A(∗) → B. In this case Et
for A(∗) → B respectively.
1, . . . , Et
1, . . . , Et
i,j,
n and f (t) are called a parametrized basis and a parametrized index
n are such polynomials ck
i,j(t) ∈ C[t] that ck
i,j(0) = ck
We now explain how to prove A(∗) 6→ B. Note that if dim Der(A(α)) > dim Der(B) for all α ∈ I
then A(∗) 6→ B. One can use also the following generalization of Lemma 1, whose proof is the same as the
proof of Lemma 1.
Lemma 2. Let B be a Borel subgroup of GL(V ) and R ⊂ L(T ) be a B-stable closed subset. If A(∗) → B
and for any α ∈ I the algebra A(α) can be represented by a structure µ(α) ∈ R, then there is λ ∈ R
representing B.
Pj=1
4
4. CLASSIFICATION AND DEGENERATIONS OF THREE DIMENSIONAL ANTICOMMUTATIVE ALGEBRAS
First we consider the variety ACom3. Let us fix the basis e1, e2, e3 of V . Any structure µ ∈ ACom3
i,j (1 ≤ i, j, k ≤ 3) is determined by the 3 × 3 matrix Aµ whose (i, j)-entry is
with structure constants ck
(−1)i−1cj
u,v, where (u, v) is a unique pair of numbers such that u, v ∈ {1, 2, 3}\{i} and u < v. Since C is an
algebraically closed field the structure λ belongs to O(µ) if and only if there is a nonsingular matrix X such
that Aλ = X T AµX by [24, Corollary 2.4]. Then the classification of three dimensional anticommutative
algebras modulo isomorphism can be obtained from the classification of bilinear forms modulo congruence
given in [18].
i,j, where dk
We denote by W some four dimensional space that contains V as a subspace and by e4 some fixed vector
of W such that W = V ⊕ Ce4. Let us now consider four dimensional algebras A such that A(A2) =
(A2)A = 0 and dim A2 ≤ 1. It is easy to see that such an algebra can be represented by a structure χ on
W such that χ(W, W ) ⊂ Ce4 and χ(W, e4) = χ(e4, W ) = 0. Such a structure is defined by the 3 × 3
matrix Bχ, whose (i, j)-entry is d4
i,j (1 ≤ i, j, k ≤ 4) are the structure constants of χ. It is
clear that two such structures χ and η lie in the same orbit if and only if there is a nonsingular matrix X
such that Bη = X T BχX. Now we put in correspondence to an anticommutative algebra structure µ on
V the structure χµ on W satisfying the properties above with Bχµ = Aµ. As it was explained above, we
get a bijection between orbits of ACom3 and isomorphism classes of four dimensional algebras A such
that A(A2) = (A2)A = 0 and dim A2 ≤ 1. Moreover, it is clear that if λ ∈ O(µ), then χλ ∈ O(χµ).
The converse assertion follows from [15, Proposition 1.7] and the fact that the set of structures χ on W
satisfying χ(W, W ) ⊂ Ce4 and χ(W, e4) = χ(e4, W ) = 0 is a closed subset stable under the action of
lower triangular matrices.
Thus, isomorphism classes and degenerations of three dimensional anticommutative algebras can be
transfered from the isomorphism classes and degenerations of four dimensional algebras A such that
A(A2) = (A2)A = 0 and dim A2 ≤ 1. The last mentioned problem is a part of the problems that were
solved in [13, 21]. Unfortunately both of the mentioned works have problems. Some degenerations are
missed in the paper [13]. All degenerations between algebras that we are interested in are described cor-
rectly in [21], but the classification used in this paper lost one algebra. In the current work we will use
the results of [21]. Let us first deduce the classification of three dimensional anticommutative algebras
using the last mentioned paper. We do this in Table A.1 below, where in the first column we put the names
of anticommutative three dimensional algebras, in the second column we put the corresponding names of
algebras from [21], in the third and fourth columns we put multiplication tables and dimensions of alge-
bras of derivations of the corresponding anticommutative algebras. We omit products of basic elements
whose values are zero or can be recovered from the anticommutativity and given products. Note that if
dim Der(µ) = k, then dim Der(χµ) = k + 4.
Here g1, g2, gα
3 and g4 are Lie algebras and A2 corresponds to the algebra missed in [21] that is denoted by
N0 in this paper. We have gα
1 if αβ = 1 and there are no other nontrivial isomorphisms
3
between the algebras in the table. All degenerations and non-degenerations between the algebras from the
column B that do not involve N0 are described in [21]. Thus, it remains to describe degenerations involving
A2.
3 and Aα
1
∼= Aβ
∼= gβ
Note that dim AnnL(N0) > dim AnnL(N3(0)) = dim AnnL(N10) and dim Der(N0) <
dim Der(N3(0)) = dim Der(N10), and hence there are no degenerations between N0, N3(0) and N10.
Note now that, for any α ∈ C we have
• the degeneration A2 → gα
3 given by the parametrized basis
1 = te3, Et
2 = te1, Et
3 = e1 + (α + t)e2 + e3;
Et
• the degeneration Aα
1 → A2 given by the parametrized basis
Et
1 = te2, Et
2 = −e1, Et
3 = αe1 − e2 + e3.
Thus, we get the following result.
Theorem 3. The graph of primary degenerations for ACom3 can be obtained from the graph given in Figure
1 below by deleting all vertices named with L.
Since ACom3 is isomorphic to C9 as an algebraic variety, it is irreducible and equals O(Aα
1 )
5. DEGENERATIONS OF THREE DIMENSIONAL LEIBNIZ ALGEBRAS
The classification of three dimensional non-Lie Leibniz algebras is presented in Table A.2 below.
Theorem 4. The graph of primary degenerations for Leib3 can be obtained from the graph given in Figure
1 below by deleting all vertices named with A.
5
Proof. We prove all the required primary degenerations in Table A.3 below. Let us consider the degenera-
tion Lβ
1 → L2 to clarify our formulas. Write nonzero products in Lβ
3Et
2 = βt2Et
2 = tEt
1, Et
1, Et
2Et
3Et
Et
1 in the basis Et
i :
3 = Et
1.
It is easy to see that for t = 0 we obtain the multiplication table of L2. The remaining degenerations can be
interpreted in the same way.
A part of non-degenerations is given in Table A.4 below. Whenever an algebra named by A with the basis
i,j (1 ≤ i, j, k ≤ 3) denote the structure constants of A in the given basis and
f1, f2, f3 appear in this table, ck
Ai (1 ≤ i ≤ 3) denotes the subspace of A generated by fi, . . . , f3.
3i = cj
ij (1 ≤ i, j, k ≤ 3) satisfy the conditions ck
In the rest of the proof we will use ideas from the proof of [19, Theorem 2]. All the remaining de-
generations involve only solvable non-nilpotent Leibniz algebras with a two dimensional nilpotent radical.
Moreover, each of them is represented in Table A.2 by a structure µ such that he1, e2i is the nilpotent radical
and the structure constants ck
ij = 0 if i, j ≤ 2 and k ≥ min(i, j)
and cj
i3 = 0 for any 1 ≤ i, j ≤ 2 such that j < i. During this proof we will call a structure with
three dimensional nilpotent radical that satisfies the described conditions a standard structure. Let us put
in correspondence to a standard structure µ the 4-tuple Sµ = (c1
32) ∈ C4. It is not difficult
to show that if Sµ = (a1, b1, a2, b2) and λ ∈ O(µ) is a standard structure, then there is some permutation
σ : {1, 2} → {1, 2} and some c ∈ C∗ such that Sλ = (caσ(1), cbσ(1), caσ(2), cbσ(2)). Suppose now that
{µs}s∈T is a set of standard structures, Sµs = (a1,s, b1,s, a2,s, b2,s), and the homogeneous linear polynomials
f1, . . . , fl ∈ C[x1, x2, x3, x4] are such that fr(a1,s, b1,s, a2,s, b2,s) = 0 for all s ∈ T and 1 ≤ r ≤ l. If λ
is a standard structure with Sλ = (a1, b1, a2, b2), then it easily follows from Lemma 2 that there is some
permutation σ : {1, 2} → {1, 2} and some c ∈ C∗ such that fr(caσ(1), cbσ(1), caσ(2), cbσ(2)) = 0 for all
1 ≤ r ≤ l. Thus, we get
31, c2
23, c2
13, c1
B 6→ L0
6 for B ∈ {Lα6=0
6
, L7}
and
B 6→ L1
6 for B ∈ {Lα
4 , L5, Lα6=1
6
, L9}.
✷
Corollary 5. The irreducible components of Leib3 are
3 }α∈C) = {g1, g2, gα
3 , g4, C3};
3, Lβ
C1 = O({gα
C2 = O(g4) = {g1, g−1
C3 = O({Lα
C4 = O(L5) = {L2, L3, L2
6 }α∈C) = {L0
C5 = O({Lα
4 }α∈C) = {g1, g0
4, L5, C3};
1, L2, Lα
3 , C3}α∈C;
1 , L2, L3, Lα
4 , L0
6, C3}α,β∈C;
6 , L7, L8, L9, C3}α∈C.
In particular, the set of rigid algebras in the variety Leib3 is formed by g4 and L5.
Proof. All degenerations and non-degenerations that do not follow directly from Theorem 4 follow from
Talble A.5.
✷
6
1
2
3
4
5
6
9
FIGURE 1: THE GRAPH OF PRIMARY DEGENERATIONS FOR
LIE, ANTICOMMUTATIVE AND LEIBNIZ THREE DIMENSIONAL ALGEBRAS.
Aα
1
α = 1
A2
α = −1
A3
g4
α = 1
α = −1
α = 0
gα
3
α = 1
g1
g2
L5
α = 2
Lα
4
Lα
6
L7
L9
L8
L0
6
α = 2
β = 1−α
(2−α)2
β = 0
β = 0
Lβ
1
L3
β = 1/4
L1
6
L2
C3
Appendix A. Tables
Table A.1. Classification of three dimensional anticommutative algebras.
A
g1
g2
gα
3
g4
Aα
1
A2
A3
B
NC2
1
NC
1
NC
3 if α = −1,
NC
N3(0)
N4 if α = −1, N5 if α = 1,
(1+α)2(cid:17) otherwise
2 (cid:16) α
N3(cid:16) 1+α
1−α i(cid:17) otherwise
N0
N10
Multiplication tables
e2e3 = e1
e1e3 = e1, e2e3 = e2
e1e3 = e1 + e2, e2e3 = αe2
e1e2 = e3, e1e3 = −e2, e2e3 = e1
e1e2 = e3, e1e3 = e1 + e3, e2e3 = αe2
e1e2 = e1, e2e3 = e2
e1e2 = e3, e1e3 = e1, e2e3 = e2
Table A.2. Classification of three dimensional Leibniz (non-Lie) algebras.
A
Lβ
1
L2
L3
Lα
4
L5
Lα
6
L7
L8
L9
[11]
2(a)
2(b)
2(c)
2(e), 2(f )
2(g)
2(d), 3(a)
3(b)
3(c)
3(d)
Sµ
( β, 0, 1, −1 )
2, 0, 1, −1 )
(
( α,
0 )
0 )
(
0, 1,
1, 0, 1,
(
0, 0, 1,
0 )
Multiplication tables
e2e2 = βe1, e3e2 = e1, e3e3 = e1
e3e3 = e1
e2e2 = e1, e3e3 = e1
e1e3 = αe1, e2e3 = e2, e3e2 = −e2, e3e3 = e1
e1e3 = 2e1, e2e2 = e1, e2e3 = e2, e3e2 = −e2, e3e3 = e1
e1e3 = αe1, e2e3 = e2
e1e3 = e1 + e2, e2e3 = e2
e1e3 = e2, e3e3 = e1
e1e3 = e2, e2e3 = e2, e3e3 = e1
Table A.3. Degenerations of Leibniz algebras of dimension 3.
7
Der
6
6
4
3
1
2
3
Der
4
5
4
3
2
2 + δα,0 + 2δα,1
2
3
2
Degenerations
Lβ
1 → L2
L1/4
1 → g1
L3 → L2
1−α
4
Lα6=2
(2−α)2
→ L
1
4 → g0
L0
3
L2
4 → L3
L5 → L2
4
L0
6 → L0
1
L1
6 → L2
Lα6=0,1
→ L8
L7 → L1
6
L7 → L8
L8 → L0
1
L9 → L0
6
L9 → L8
6
Parametrized bases
2−α e1 + (1−α+t)t
2−α
Et
Et
Et
Et
Et
Et
1 = e1,
1 = t3e1,
1 = e1,
Et
1 = t2e1,
1 = −t−1e1 + te2, Et
Et
1 = t2e1,
Et
1 = te1,
1 = t2e2,
Et
Et
1 = te1,
Et
1 = e1 + te2,
Et
1 = te1,
Et
1 = e1 + e2,
Et
1 = te2,
Et
1 = t−1e1,
1 = t2e1,
Et
2 = te2,
2 = −2te2 + te3,
2 = te2,
2 = t
2 = t−1e1,
2 = ite1 + e2 − ite3,
2 = te2,
2 = te3,
2 = e2,
2 = (α − 1)te1,
2 = e2,
2 = te2,
2 = te3,
2 = t−2e2,
2 = t3e2,
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
2 e1 + e3
2−α e2 + te3
3 = e3
3 = 2t2e2
3 = e3
3 = 1
3 = e3
3 = te3
3 = t−1
3 = te1 + te2 + te3
3 = e1 + te3
3 = α−1t−1e1 + e2 + te3
3 = e3
3 = t−1e1 + te3
3 = e1 + te3
3 = e3
3 = te3
e3, Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
Et
8
Table A.4. Non-degenerations of Leibniz algebras of dimension 3.
Non-degenerations
Reasons
Lα6=2
4
9 B, B ∈( L
β6= 1−α
(α−2)2
1
g1(α 6= 0), g2, gβ6=0
, L3,
3
L5
9 B, B ∈ {Lβ
1 , g1, g2}
A
9 B,
A ∈ {Lα
6 , L7, L9},
, L3,
B ∈(cid:26) Lβ6=0
4, g1, g2 (cid:27)
1
L1
) R =
R =
A(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
A(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
AnnL(A) > AnnL(B)
A = hf1, f2, f3i, c2
31 = −αc2
c3
A1A3 + A2A2 = 0, A3A1 ⊆ A3, A1A1 ⊆ A2
21 = (α − 1)c3
11 = 0, c2
21 = −c2
12, c3
12,
12,
Lα
4 ∈ O(R) (take f1 = e3, f2 = e2, f3 = e1), but B 6∈ O(R)
12, c3
A = hf1, f2, f3i, c3
A1A3 + A3A2 = 0, A3A1 + A2A2 ⊆ A3,
A1A2 + A2A1 ⊆ A2
21 = −2c2
31 = 2c2
21 = c3
12,
L5 ∈ O(R) (take f1 = e3, f2 = e2, f3 = e1), but B 6∈ O(R)
Table A.5. Orbit closures for some families of three dimensional Leibniz algebras.
L∗
4
L∗
6
L∗
6
L∗
1
L∗
4
L∗
4
L∗
6
Degenerations
→ L0
6
→ L7
→ L9
→ L3
Et
Et
Et
Et
Et
1 = e2,
1 = e1 + e2, Et
1 = e1 + te2, Et
1 = t4e1,
Et
Parametrized bases
Et
2 = e1,
Et
2 = te2,
2 = (1 − t)e1, Et
2 = t3e2,
Et
3 = te3
3 = e3
3 = e1 + e2 + te3
3 = t2e3
Indices
ǫ = t−1
ǫ = 1 − t
ǫ = t−1
ǫ = t−2
Non-degenerations
Reasons
9 B, B ∈ {L1
6, L8}
9 B, B ∈ {g2, gβ6=0
9 B, B ∈ {g1, g2, Lβ6=0
}
3
1
, L3, L1
4}
4 )(+2) < B(+2)
(Lα
AnnL(Lα
AnnL(Lα
4 ) > AnnL(B)
6 ) > AnnL(B)
REFERENCES
[1] Alvarez M.A., On rigid 2-step nilpotent Lie algebras, Algebra Colloquium, 25 (2018), 2, 349 -- 360.
[2] Alvarez M.A., The variety of 7-dimensional 2-step nilpotent Lie algebras, Symmetry, 10 (2018), 1, 26.
[3] Alvarez M.A., Hern´andez I., On degenerations of Lie superalgebras, arXiv:1802.08707
[4] Alvarez M.A., Hern´andez I., Kaygorodov I., Degenerations of Jordan superalgebras, arXiv:1708.07758
[5] Bekbaev U., Complete classification of a class of 3-dimensional algebras, arXiv:1710.10734
[6] Benes T., Burde D., Degenerations of pre-Lie algebras, Journal of Mathematical Physics, 50 (2009),
11, 112102, 9 pp.
[7] Benes T., Burde D., Classification of orbit closures in the variety of three dimensional Novikov alge-
bras, Journal of Algebra and Its Applications, 13 (2014), 2, 1350081, 33 pp.
[8] Burde D., Degenerations of nilpotent Lie algebras, Journal of Lie Theory, 9 (1999), 1, 193 -- 202.
[9] Burde D., Steinhoff C., Classification of orbit closures of 4 -- dimensional complex Lie algebras, Journal
of Algebra, 214 (1999), 2, 729 -- 739.
[10] Calder´on A., Fern´andez Ouaridi A., Kaygorodov I., The classification of n-dimensional anticom-
mutative algebras with (n − 3)-dimensional annihilator, Communications in Algebra, 47 (2019),
DOI:10.1080/00927872.2018.1468909
[11] Casas J., Insua M., Ladra M., Ladra S., An algorithm for the classification of 3-dimensional complex
Leibniz algebras, Linear Algebra and its Applications, 436 (2012), 9, 3747 -- 3756.
[12] Casas J., Khudoyberdiyev A., Ladra M., Omirov B., On the degenerations of solvable Leibniz algebras,
Linear Algebra and its Applications, 439 (2013), 2, 472 -- 487.
[13] Fialowski A., Penkava M., The moduli space of 4-dimensional nilpotent complex associative algebras,
Linear Algebra and its Applications, 457 (2014), 408 -- 427.
[14] Gainov A., Binary Lie algebras of lower ranks [Russian], Algebra i Logika Sem., 2 (1963), 4, 21 -- 40.
9
[15] Grunewald F., O'Halloran J., Varieties of nilpotent Lie algebras of dimension less than six, Journal of
Algebra, 112 (1988), 315 -- 325.
[16] Grunewald F., O'Halloran J., A Characterization of orbit closure and applications, Journal of Algebra,
116 (1988), 163 -- 175.
[17] Grunewald F., O'Halloran J., Deformations of Lie algebras, Journal of Algebra, 162 (1993), 1, 210 --
224.
[18] Horn R., Sergeichuk V., Canonical matrices of bilinear and sesquilinear forms, Linear Algebra and its
Applications, 428 (2008), 1, 193 -- 223.
[19] Ismailov N., Kaygorodov I., Volkov Yu., The geometric classification of Leibniz algebras, Interna-
tional Journal of Mathematics, 29 (2018), 5, 1850035.
[20] Kaygorodov I., Popov Yu., Volkov Yu., Degenerations of binary Lie and nilpotent Malcev algebras,
Communications in Algebra, 46 (2018), 11, 4929 -- 4941.
[21] Kaygorodov I., Popov Yu., Pozhidaev A., Volkov Yu., Degenerations of Zinbiel and nilpotent Leibniz
algebras, Linear and Multilinear Algebra, 66 (2018), 4, 704 -- 716.
[22] Kaygorodov I., Volkov Yu., The variety of 2-dimensional algebras over an algebraically closed field,
arXiv:1701.08233
[23] Kaygorodov I., Volkov Yu., The degeneration level classification of algebras, arXiv:1710.08943
[24] Kobayashi Yu., Shirayanagi K., Takahasi S., Tsukada M., Classification of three dimensional zeropo-
tent algebras over an algebraically closed field, Communications in Algebra, 45 (2017), 12, 5037 -- 5052.
[25] Khudoyberdiyev A., Omirov B., The classification of algebras of level one, Linear Algebra and its
Applications, 439 (2013), 11, 3460 -- 3463.
[26] Khudoyberdiyev A., Ladra M., Masutova K., Omirov B., Some irreducible components of the variety
of complex (n + 1)-dimensional Leibniz algebras, Journal of Geometry and Physics, 121 (2017), 228 --
246.
[27] Lauret J., Degenerations of Lie algebras and geometry of Lie groups, Differential Geometry and its
Applications, 18 (2003), 2, 177 -- 194.
[28] Loday J.-L., Pirashvili T., Universal enveloping algebras of Leibniz algebras and (co)homology, Math-
ematische Annalen, 296 (1993), 1, 139 -- 158.
[29] Rakhimov I., Mohd Atan K., On contractions and invariants of Leibniz algebras, Bulletin of the
Malaysian Mathematical Sciences Society, 35 (2012), 557 -- 565.
[30] Seeley C., Degenerations of 6-dimensional nilpotent Lie algebras over C, Communications in Algebra,
18 (1990), 3493 -- 3505.
|
1501.06808 | 3 | 1501 | 2016-06-30T16:47:04 | A prime ideal principle for two-sided ideals | [
"math.RA",
"math.AC"
] | Many classical ring-theoretic results state that an ideal that is maximal with respect to satisfying a special property must be prime. We present a "Prime Ideal Principle" that gives a uniform method of proving such facts, generalizing the Prime Ideal Principle for commutative rings due to T.Y. Lam and the author. Old and new "maximal implies prime" results are presented, with results touching on annihilator ideals, polynomial identity rings, the Artin-Rees property, Dedekind-finite rings, principal ideals generated by normal elements, strongly noetherian algebras, and just infinite algebras. | math.RA | math |
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
MANUEL L. REYES
Abstract. Many classical ring-theoretic results state that an ideal that is maximal with
respect to satisfying a special property must be prime. We present a "Prime Ideal Principle"
that gives a uniform method of proving such facts, generalizing the Prime Ideal Principle
for commutative rings due to T. Y. Lam and the author. Old and new "maximal implies
prime" results are presented, with results touching on annihilator ideals, polynomial identity
rings, the Artin-Rees property, Dedekind-finite rings, principal ideals generated by normal
elements, strongly noetherian algebras, and just infinite algebras.
1. Introduction
A well-known phenomenon in ring theory is that, for certain properties P, ideals of a ring
that are maximal with respect to satisfying P are prime. This is probably most familiar in
the setting of commutative algebra, as in [7, pp. 70–71, 84–85]. For instance, an ideal in any
commutative ring R is prime if it is maximal with respect to any of the following properties:
• being a proper subset of R,
• not being finitely generated,
• not being principal, or
• having empty intersection with a given multiplicative submonoid S ⊆ R.
The Prime Ideal Principle of [19] unifies the above results along with many other classical
"maximal implies prime" results in commutative algebra. The basic idea behind this unifi-
cation was to find certain properties of a family F of ideals in a fixed commutative ring that
guarantee that an ideal maximal with respect to not lying in F is prime. The most notable
such condition is that the family F be an Oka family; this means that, for any ideal I E R
and any element a ∈ R, the following implication holds:
(I, a), (I : a) ∈ F =⇒ I ∈ F .
The Prime Ideal Principle then states that, for any Oka family F in a commutative ring R,
an ideal I E R maximal with respect to I /∈ F is prime. All of the examples above were
recovered from the Prime Ideal Principle in [19] by demonstrating that a relevant family of
ideals is Oka.
A version of the Prime Ideal Principle for right ideals in noncommutative rings was estab-
lished in [26]. This was achieved by identifying an appropriate type of "prime right ideal"
Date: June 30, 2016.
2010 Mathematics Subject Classification. Primary: 16D25, 16N60; Secondary: 13A15, 16D20.
Key words and phrases. maximal implies prime, Prime Ideal Principle, Oka family of ideals.
This material is based upon work supported by the National Science Foundation under Grant No. DMS-
1407152.
1
2
MANUEL L. REYES
(these were called completely prime right ideals) and by finding a suitable definition of Oka
families of right ideals. However, to date there is no version of the Prime Ideal Principle for
two-sided ideals in noncommutative rings.
The purpose of this paper is to present such a theory for two-sided ideals in noncommuta-
tive rings, unifying many of the classical noncommutative "maximal implies prime" results
and presenting some new ones. In Section 2 we briefly introduce Oka families of (two-sided)
ideals in noncommutative rings and prove the corresponding Prime Ideal Principle: if F is
an Oka family of ideals in a noncommutative ring R and I is an ideal that is maximal with
respect to I /∈ F , then I is prime. We also introduce stronger conditions on families of ideals
called (P1), (P2), and (P3) that are noncommutative analogues of those defined in [19].
The heart of the paper lies in the remaining sections, which present many examples of
families satisfying these different properties with corresponding applications of the Prime
Ideal Principle. Section 3 gives examples of families satisfying the properties (P1), (P2), and
(P3). Section 4 considers families that satisfy the weaker r-Oka and Oka properties, many
of which are constructed from various classes of (bi)modules; it concludes with a discussion
of problems for future work.
Acknowledgments. I wish to thank George Bergman, T. Y. Lam, and Lance Small for
helpful conversations, suggestions, and references to the literature. I also thank the referee
for offering several clarifying remarks.
Conventions. Throughout this paper, all rings, modules, and ring homomorphisms are
assumed to be unital. Let R be a ring. We write I E R to denote that I is an ideal of R.
If X is a subset of R, we write (X) for the ideal of R generated by X; in case X = {x} is
a singleton we simply write (x) = ({x}). Furthermore, if J E R and a ∈ R, then we denote
(I, J) = (I ∪ J) = I + J and (I, a) = (I ∪ {a}) = I + (a). The notation MR and RM
respectively indicate that M is a right or left R-module.
2. The Prime Ideal Principle
The families of ideals defined below illustrate one possible way to generalize those ideal
families defined in [19] from commutative rings to noncommutative rings. For two ideals
I, J E R we denote J −1I = {x ∈ R : Jx ⊆ I} and IJ −1 = {x ∈ R : xJ ⊆ I}, both of which
are clearly ideals in R.
Definition 2.1. Let R be a ring and let F be a family of ideals of R with R ∈ F . We say
that F is an Oka family of ideals if, for any I E R and a ∈ R,
(2.2)
(I, a), (a)−1I, I(a)−1 ∈ F =⇒ I ∈ F .
We say that F is an r-Oka family of ideals if, for any I E R and a ∈ R, the condition
(I, a), (a)−1I ∈ F implies that I ∈ F . Furthermore, we define the following properties of the
family F (recall the standing assumption that R ∈ F ), where I, J E R are arbitrary ideals:
• Monoidal : I, J ∈ F =⇒ IJ ∈ F ;
• Semifilter : I ∈ F and J ⊇ I =⇒ J ∈ F ;
• (P1): F is a monoidal semifilter;
• (P2): F is monoidal, and for any J ∈ F , I 2 ⊆ J ⊆ I =⇒ I ∈ F ;
• (P3): For A, B ∈ F and I E R, AB ⊆ I ⊆ A ∩ B =⇒ I ∈ F ;
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
3
• Strongly r-Oka: I + J, J −1I ∈ F =⇒ I ∈ F ;
• Strongly Oka: I + J, J −1I, IJ −1 ∈ F =⇒ I ∈ F .
The r-Oka and strongly r-Oka familes defined above have obvious left-handed analogues,
which we refer to as (strongly) ℓ-Oka families. An alternative term would have been
"(strongly) right- and left-Oka" families. However, we have avoided this terminology since it
would have suggested that an Oka family is the same as a family that is both right and left
Oka. It seems likely that an Oka family need not be either r-Oka nor ℓ-Oka in full generality
(such as the families produced via Proposition 4.13), but we do not have an explicit example
to verify this.
Each of the properties of ideal families listed above is easily seen to pass to arbitrary
intersections of such families. Thus for each of these properties, the set of all ideal families
in R with that property forms a complete sublattice of the lattice of all families of ideals in
R.
The following result outlines the logical dependence between the many types of families
defined above.
Proposition 2.3. For a family F of ideals in a ring R, we have the following implications
of the properties defined above:
(P1) =⇒ (P2) =⇒ (P3) =⇒ strongly r-Oka =⇒ r-Oka
⇓
⇓
strongly Oka =⇒ Oka
Proof. The horizontal implications are proved in the commutative case in [19, Thm. 2.7],
and those arguments apply here with virtually no adjustment. The vertical implications are
obvious.
(cid:3)
We mention briefly that there are various equivalent characterizations of the (P2) and
(P3) conditions given in [19, §2] and [20, §2]. Though the statements and proofs are given
for commutative rings, they once again transfer easily to the noncommutative case. Note
that because (P2) ⇒ (P3) and because any (P3) family is monoidal, a family satisfying any
property (Pi) is monoidal.
Before proving the Prime Ideal Principle in this setting, we establish a variant of the usual
criterion for a ring to be prime. If R is a prime ring and a, b ∈ R \ {0} then one can conclude
that both aRb 6= 0 and bRa 6= 0. There is an inherent symmetry in the definition of a
prime ring in the sense that we can interchange the roles of a and b. What if R fails to be
prime? We know offhand that there exist a, b ∈ R \ {0} such that aRb = 0. But if R is not
commutative, we cannot necessarily conclude that bRa = 0. At first glance, the symmetry
seems to disappear in the failure of a ring to be prime. The following lemma rectifies this
situation.
Lemma 2.4. For any ring R 6= 0, the following are equivalent:
(1) R is a prime ring
(2) For a, b ∈ R \ {0}, both aRb 6= 0 and bRa 6= 0
(3) For a, b ∈ R \ {0}, either aRb 6= 0 or bRa 6= 0.
In particular if R is not a prime ring, then there exist 0 6= a, b ∈ R such that aRb = 0 = bRa.
4
MANUEL L. REYES
Proof. The implications (1) ⇒ (2) ⇒ (3) are immediate. To prove (3) ⇒ (1) we will verify
the contrapositive. Suppose that R is not a prime ring; we will produce a, b ∈ R \ {0} such
that aRb = bRa = 0. Because the ring is not prime, there exist x, y ∈ R \ {0} such that
xRy = 0. If yRx = 0 then we are finished. Otherwise there exists r ∈ R such that yrx 6= 0.
Setting a = b = yrx yields aRb = bRa = 0.
(cid:3)
We arrive at the following noncommutative version of the Prime Ideal Principle for two-
sided ideals in noncommutative rings. Given a family F of ideals in a ring R, its complement
is F ′ = {I E R : I /∈ F }. We use Max(F ′) to denote the set of ideals that are maximal in
the complement of F . Also, we let Spec(R) denote the set of prime ideals of R.
Theorem 2.5 (Prime Ideal Principle). For any Oka family F of ideals in a ring R, we have
Max(F ′) ⊆ Spec(R).
Proof. Let I ∈ Max(F ′), and assume for contradiction that I is not prime. By Lemma 2.4
there exist a, b ∈ R\I such that aRb, bRa ⊆ I. Then a ∈ (I, a)\I and b ∈ (a)−1I ∩I(a)−1 \I.
Maximality of I implies that the ideals (I, a), (a)−1I, and I(a)−1 lie in F . Because F is Oka,
this means that I ∈ F , a contradiction.
(cid:3)
The Prime Ideal Principle supplement of [19] was a streamlined way to apply the Prime
Ideal Principle in order to deduce that certain ideal-theoretic properties can be "tested" on
the set of primes. We also have a noncommutative analogue of this result.
Theorem 2.6. Let F be an Oka family such that every nonempty chain of ideals in F ′ (with
respect to inclusion) has an upper bound in F ′ (this holds, for example, if every ideal in F
is finitely generated).
(1) Let F0 be a semifilter of ideals in R. If every prime ideal in F0 belongs to F , then
F0 ⊆ F .
(2) For an ideal J E R, if all prime ideals containing J (resp. properly containing J)
belong to F , then all ideals containing J (resp. properly containing J) belong to F .
(3) If all prime ideals of R belong to F , then all ideals belong to F .
Proof. Both (3) and (2) are special cases of (1). We prove (1) by a contrapositive argument.
Let F0 be a semifilter such that F0 * F ; we will construct a prime ideal in F0 \ F . Let
I0 ∈ F0 \ F = F0 ∩ F ′. The assumption on chains in F ′ allows us to apply Zorn's lemma and
find I ∈ Max(F ′) such that I0 ⊆ I. The fact that F0 is a semifilter gives I ∈ F0. Because
I is maximal in the complement of F , the Prime Ideal Principle 2.5 implies that I is prime.
Thus I is a prime ideal in F0 \ F as desired.
(cid:3)
Remark 2.7. Another property of ideal families in commutative rings that was introduced
in [19] was that of an Ako family; a family F in a commutative ring R has the Ako property
if (I, a), (I, b) ∈ F implies (I, ab) ∈ F for any a, b ∈ R. One can readily produce noncom-
mutative versions of this property that satisfies the Prime Ideal Principle. For instance,
K. Goodearl proposed the following noncommutative version of the Ako property: given
I E R and a, b ∈ R,
(I, a), (I, b) ∈ F =⇒ (I, arb) ∈ F for some r ∈ R.
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
5
Another possibility would be (I, a), (I, b) ∈ F =⇒ I + (a)(b) ∈ F . One can readily verify
that a family F satisfying either of these properties will also have Max(F ′) ⊆ Spec(R).
Since the Ako property has seemed to be less useful than the Oka property already in the
commutative setting, we do not focus our current efforts on this property.
Before proceeding to examples of Oka families, we conclude this section with a discussion
of a prominent family of ideals that does not generally form an Oka family. One might
wonder why a Prime Ideal Principle was given for one-sided ideals before it was established
in the more classical setting of two-sided ideals. We will give a few examples to show that,
in a sense, the case of two-sided ideals is more difficult to handle. Specifically, we recall the
following classical result of Cohen [6, Thm. 2].
Cohen's Theorem. An ideal in a commutative ring that is maximal with respect to not
being finitely generated is prime. Consequently, if every prime ideal of a commutative ring
is finitely generated, then that ring is noetherian.
This result provided strong guidance for the creation of the Prime Ideal Principle in [19]
and its noncommutative analogue in [26]. The behavior of the family of finitely generated
directly motivated the definition of an Oka family of ideals in a commutative ring and for
the definition of a right Oka family in a noncommutative ring.
In [26, Thm. 3.8] we provided a version of Cohen's result above for right ideals of a
noncommutative ring. Specifically, after fixing an appropriate notion of an Oka family of
right ideals, we proved that the family of finitely generated right ideals of any ring is an Oka
family. As a consequence we showed that a right ideal of a ring that is maximal with respect
to not being finitely generated is a completely prime right ideal. (Further Cohen-type results
for right ideals in noncommutative rings were investigated in [27].)
A fundamental difference in the case of two-sided ideals of a noncommutative ring is that,
if R is not a commutative ring, then an ideal of R that is maximal with respect to not being
finitely generated need not be prime. This is illustrated in the following examples.
Let A and B be simple rings, and let M be an (A, B)-bimodule. The triangular matrix
ring
R = T (A, M, B) = (cid:18)A M
0 B(cid:19)
will have exactly two prime ideals:
0
0
P1 = (cid:18)A M
= (cid:18)0 M
P2 = (cid:18)0 M
= (cid:18)0 M
0 (cid:19) = R(cid:18)1 0
0 (cid:19) + R(cid:18)1 0
0 B(cid:19) = R(cid:18)0 0
0 (cid:19) + R(cid:18)0 0
0 0(cid:19) R
0 0(cid:19) = (cid:18)0 M
0 1(cid:19) R
0 1(cid:19) = (cid:18)0 M
0
0
0 (cid:19) +(cid:18)1 0
0 0(cid:19) R,
0
0 (cid:19) +(cid:18)0 0
0 1(cid:19) R.
6
MANUEL L. REYES
The expressions above make it evident that P1 and P2 are always finitely generated as ideals,
and that if M is finitely generated as a left A-module (respectively, as a right B-module),
then P1 and P2 will be finitely generated as left (respectively, right) ideals of R.
Example 2.8. Let K be a field, let MK be an infinite-dimensional vector space viewed as
a bimodule with the same left and right action, and consider the ring R = T (K, M, K) (so
A = B = K above). Then P1 and P2 are finitely generated as ideals, but the ideal I = ( 0 V
0 0 )
is not finitely generated.
Example 2.9. Let K/F be an infinite extension of fields. In the (K, K)-bimodule M =
K ⊗F K, let N be the sub-bimodule generated by {1 ⊗ α − α ⊗ 1 : α ∈ K}. This is a maximal
sub-bimodule since M/N ∼= K as (K, K)-bimodules, and one can readily check that M is
not finitely generated (by choice of K and F ).
In the ring R = T (K, M, K), the choice
of M and N guarantees that the ideal ( 0 N
0 0 ) is maximal with respect to not being finitely
generated as an ideal of R. But because R/I ∼= ( K K
0 K ), we see that I is not even semiprime.
Example 2.10. We thank Lance Small for suggesting this example. Let S be a simple
ring that is not right noetherian (such as the endomorphism ring of a countably infinite
dimensional vector space modulo its finite rank operators), let Z = Z(S) be its central
subfield, and set R = T (Z, S, S). Then the prime ideals P1 and P2 of R are both finitely
generated as right ideals, but for any right ideal J of S that is not finitely generated, the
corresponding ideal ( 0 J
0 0 ) of R is not finitely generated.
Remark 2.11. Thanks to the examples given above, we see that for a general ring R the
family Ffg of finitely generated ideals of R need not be an Oka family. For if Ffg were
Oka and all prime ideals of R were finitely generated, then by Theorem 2.6 all ideals of R
would be finitely generated, contrary to those examples. Similarly, the family of ideals that
are finitely generated as right ideals is not an Oka family. These counterexamples sharply
contrast the case of commutative rings [19, Proposition 3.16] and the case of Oka families of
right ideals in noncommutative rings [26, Proposition 3.7].
The examples above naturally suggest the following question: do there exist simple rings
A and B and an (A, B)-bimodule M such that M is finitely generated as both a left A-
module and a right B-module, but such that M has a sub-bimodule that is not finitely
generated? For given such data, the ring R = T (A, M, B) will have both prime ideals
finitely generated as left ideals and as right ideals, but will have an ideal that is not finitely
generated. Consequently, for this ring R the family of ideals that are finitely generated
both as left ideals and as right ideals would fail to be an Oka family. We suspect that such
(A, M, B) should exist, but we have yet to produce an explicit example.
3. Some families satisfying the properties (Pi)
In this section we provide examples of families satisfying the conditions (P1), (P2), and
(P3). Comparing this section with the next, one sees that many well-known examples of the
noncommutative "maximal implies prime" phenomenon come from such families. We begin
with some of these classical examples.
Recall that a subset S ⊆ R containing 1 is called an m-system if, for every a, b ∈ S, there
exists r ∈ R such that arb ∈ S. It is well-known (and easy to show) that an ideal I E R is
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
7
prime if and only if R \ I is an m-system. The notion of m-system generalizes the concept
of a multiplicative subset of R, more common in commutative algebra. Our first application
of the Prime Ideal Principle recovers a well-known generalization of the fact that an ideal in
a commutative ring maximal with respect to being disjoint from some fixed multiplicative
set is prime. The last part of this proposition is due to McCoy [22, Lem. 4].
Proposition 3.1. Let S ⊆ R be an m-system. Then the family F = {I E R : I ∩ S 6= ∅}
is (P1). Hence an ideal maximal with respect to being disjoint from S is prime, and every
(nonzero) ideal of R intersects S if and only if every (nonzero) prime ideal intersects S.
Proof. Because 1 ∈ S, we have R ∈ F . It is clear that F is a semifilter. To see that F is
monoidal, let A, B ∈ F and fix a ∈ A∩S and b ∈ B ∩S. Then for some r ∈ R, arb ∈ S ∩AB.
Hence AB ∈ F . The rest of the proposition follows from the Prime Ideal Principle 2.5. (cid:3)
Next we give an application of this result to the theory of algebras over a commutative
ring k. Let R be a k-algebra. We will say that an element r ∈ R is integral over k if, for all
r ∈ R, there exists a monic polynomial f (x) ∈ k[x] such that f (r) = 0 in R. The algebra
R is integral over k if every element of R is integral over k; if k is a field, notice that this is
equivalent to R being algebraic over k. The next result shows that, to check whether R is
integral over k, it suffices to check whether it is integral modulo each prime ideal. The fact
is well-known; an instance of its application in the noncommutative setting is [24, Thm. 1].
Proposition 3.2. Let R be an algebra over a commutative ring k. The family F = {I E R :
R/I is integral over k} is (P1). In particular, R is integral over k if and only if R/P is
integral over k for all (minimal) P ∈ Spec(R).
Proof. First, fix an element a ∈ R and define the multiplicative set Sa = {f (a) ∈ R : f (x) ∈
k[x] is monic}. Then by Proposition 3.1, the family Fa = {I E R : I ∩ Sa 6= ∅} is (P1).
Notice that Fa is precisely the family of ideals such that the image of a in R/I is integral
over k. Now notice that the family F in question is given by F = Ta∈R Fa. Because F
is an intersection of (P1) families, F itself is (P1). The last statement now follows from
Theorem 2.6.
(cid:3)
While on the subject of polynomials, we turn our attention to PI-algebras. We continue to
assume that k is a commutative ring. Let khx1, x2, . . .i denote the ring of polynomials over k
in noncommuting indeterminates xi. Suppose that R is a k-algebra and f (x1, . . . , xn) ∈ khxii.
Then we denote f (R) = {f (r1, . . . , rn) ∈ R : r1, . . . , rn ∈ R}. One says that f is monic if
one of its monomials of highest total degree has coefficient 1.
Lemma 3.3. If f, g ∈ khx1, x2, . . . i are monic, then f g is monic.
Proof. Suppose that f has degree m and g has degree n. Let
f = a1u1 + a2u2 + · · · + arur + (lower degree terms),
g = b1v1 + b2v2 + · · · + bsvs + (lower degree terms),
where the {ui} are distinct words of length m and the {vj} are distinct words of length n.
We may assume that a1 = b1 = 1 because f and g are monic.
8
MANUEL L. REYES
The highest degree terms of f g are the products aibjuivj. Moreover, the monomials uivj
are readily seen to be distinct. Hence f g has u1v1 as one of its leading monomials, with
coefficient a1b1 = 1, so that f g is monic.
(cid:3)
A k-algebra R is a polynomial identity algebra (or PI algebra) over k if f (R) = 0 for some
monic polynomial f ∈ khx1, x2, . . .i (see, for instance, [28, 6.1.2]). The last sentence of the
following is a folk result; for example, it is used implicitly in the proof of [1, Thm. 2].
Proposition 3.4. Let k be a commutative ring and let S ⊆ khx1, x2, . . .i be an m-system of
polynomials in noncommuting indeterminates. For a k-algebra R, the family F = {I E R :
R/I satisfies some f ∈ S} is (P1). Hence, an ideal maximal with respect to R/I not being a
PI-algebra over k is prime.
Proof. Notice that F = {I E R : f (R) ⊆ I for some f ∈ S}. This makes it clear that
F is a semifilter. To see that F is monoidal, suppose that A, B ∈ F . Choose f, g ∈ S
and h ∈ khx1, x2, . . .i such that f (R) ⊆ A and g(R) ⊆ B as well as f hg ∈ S. Then
(f hg)(R) ⊆ f (R)h(R)g(R) ⊆ AB while f hg ∈ S, proving that AB ∈ F .
For the last statement, we apply Theorem 2.5 to the family F taking S to be the set of all
monic polynomials in khx1, x2, . . .i, which is multiplicatively closed by the lemma above. (cid:3)
We recall another familiar result from commutative algebra. For a module MR 6= 0 over a
commutative ring R, a point annihilator is an ideal I E R of the form I = ann(m) for some
m ∈ M \ {0}. Then a maximal point annihilator of M is prime. But notice that, because R
is commutative, we have ann(m) = ann(mR) for all m ∈ M. So a maximal point annihilator
of M is just an ideal maximal among the ideals of the form ann(N) where 0 6= NR ⊆ M.
From this perspective, the result generalizes to modules over noncommutative rings: for a
module MR 6= 0 over a ring R, an ideal maximal among the annihilators ann(N) where
0 6= NR ⊆ M is prime.
This well-known fact is widely attributed to I. N. Herstein; for instance, see [14, Thm. 6]
(Lance Small has communicated to us that the first instance of this result is likely in [11].)
We recover it via the Prime Ideal Principle below.
Proposition 3.5. Fix a module MR 6= 0. Then the family F = {J E R : NJ = 0 for NR ⊆
M =⇒ N = 0} is (P1). A maximal annihilator of a nonzero submodule of M is prime.
Proof. It is straightforward to see that F is a semifilter. So suppose that A, B ∈ F and that
NR ⊆ M such that N(AB) = 0. Then (NA)B = 0 =⇒ NA = 0 =⇒ N = 0.
To see that a maximal annihilator is prime, we only need to note that the maximal
annihilators of nonzero submodules of M are precisely the elements of Max(F ′) and then
apply the Prime Ideal Principle 2.5.
(cid:3)
Applying this to the module RR gives the following result.
Proposition 3.6. In a nonzero ring R, the family F of ideals I that are faithful as left
R-modules is a (P1) family. If R satisfies the ascending chain condition (ACC) on ideals,
then R is a prime ring if and only if every nonzero prime ideal of R is left faithful.
Proof. Applying Proposition 3.5 to MR = RR, we see that the family F there corresponds
to the family of left-faithful ideals of R. Turning our attention to the final statement, notice
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
9
that the "only if" part holds without the ACC assumption. The "if" part then follows from
Theorem 2.6.
(cid:3)
As a variation on the theme of maximal annihilator ideals, there are also maximal "middle
annihilators." For ideals X, Y E R we use the notation ann(X, Y ) = {r ∈ R : XrY = 0},
which is certainly an ideal of R. We call such an ideal a middle annihilator of R if XY 6= 0.
This definition originated in unpublished work of Kaplansky (see [15, p. 782]), who proved
the following "maximal implies prime" result.
Proposition 3.7. For any ring R, the family F = {I E R : for all X, Y E R, XIY = 0 =⇒
XY = 0} is (P1). A maximal middle annihilator of R is prime.
Proof. Certainly F is a semifilter. Suppose that A, B ∈ F , and let X, Y E R such that
X(AB)Y = 0. Then (XA)BY = 0 with B ∈ F implies that XAY = 0. But then A ∈ F , so
we must have XY = 0. Thus AB ∈ F , proving that F is (P1). Because Max(F ′) is equal
to the set of maximal middle annihilators of R, the last statement follows from the Prime
Ideal Principle 2.5.
(cid:3)
Our next aim is to recover the well-known result that if the ideals in a ring R satisfy ACC,
then R has finitely many minimal primes, along with a noncommutative generalization of a
result of D. D. Anderson from [2] that any commutative ring in which the minimal primes
are finitely generated has finitely many minimal primes. This makes use of the following.
Lemma 3.8. Let I be a right ideal and J be an ideal of a ring R. Suppose that these are
generated as right ideals by X ⊆ I and Y ⊆ J. Then the product IJ is generated as a right
ideal by the set {xy : x ∈ X, y ∈ Y }. In particular, if I and J are finitely generated (resp.
principal) as right ideals, then IJ is finitely generated (resp. principal) as a right ideal.
Proof. The second claim clearly follows from the first. Given I and J as above, the fact that
J is a (left) ideal gives the second equality below:
xR(cid:17) J
xJ
IJ = (cid:16)Xx∈X
= Xx∈X
= Xx∈X
= X(x,y)∈X×Y
x(cid:16)Xy∈Y
xyR.
yR(cid:17)
(cid:3)
The following statement under hypothesis (1) is well known; for instance, see [10, Theo-
rem 3.4]. The version of Anderson's theorem under hypothesis (2) was first suggested to us
by T. Y. Lam.
Proposition 3.9. For any monoidal set S of ideals in R, the family F = {I E R : I ⊇
J for some J ∈ S} is (P1). Suppose that one of the following two conditions holds:
(1) R satisfies ACC on ideals;
(2) The minimal prime ideals of R are finitely generated as right ideals.
Then the zero ideal is a product of (minimal) prime ideals, and R has finitely many minimal
primes.
10
MANUEL L. REYES
Proof. It is clear that the family F is (P1). To prove the last statement, let S be the set of
all finite products of minimal prime ideals of R and let F be the (P1) family described in the
Proposition. Notice that Spec(R) ⊆ F because every prime contains a minimal prime. In
the next paragraph, we will show that every nonempty chain in F ′ has an upper bound in
F ′. Theorem 2.6 thus implies that F consists of all ideals. In particular 0 ∈ F , so that the
zero ideal is a product of minimal primes, say P1 · · · Pn = 0. For any minimal prime P E R,
we have P ⊇ 0 = P1 · · · Pn. So some Pi ⊆ P , giving P = Pi. Thus the minimal primes of R
are precisely P1, . . . , Pn.
It remains to verify that every nonempty chain in F ′ has an upper bound. This is trivial
in case (1) holds. If R satisfies (2) then Lemma 3.8 implies that every ideal in S is finitely
generated as a (right) ideal, whence the claim easily follows from Zorn's lemma.
(cid:3)
Another notion that gives rise to prime ideals is the Artin-Rees property. An ideal I E R
is said to have the right Artin-Rees property if, for any right ideal KR ⊆ R, there exists a
positive integer n such that K ∩ I n ⊆ KI. This property arises in the study of localization
in noetherian rings (for example, see [10, Ch. 13] or [21, §4.2]). The ring R is said to be a
right Artin-Rees ring if every ideal of R has the right Artin-Rees property.
(In commutative algebra, the Artin-Rees Lemma is a result about I-filtrations of finitely
generated R-modules, where R is commutative noetherian and I E R-see [7, Lem. 5.1]. In
particular, the lemma implies that every ideal of a commutative noetherian ring satisfies the
Artin-Rees property. Thus the family of ideals satisfying the right Artin-Rees property is
mostly of interest in the noncommutative case.)
Parts of the following result were previously observed in [29, §3]. However, the "maximal
implies prime" result seems to be new.
Proposition 3.10. The family of ideals with the right Artin-Rees property in a ring R
satisfies (P2). Thus an ideal of R maximal with respect to not having the right Artin-Rees
property is prime. In particular, if R satisfies ACC on ideals, then R is a right Artin-Rees
ring if and only if every prime ideal of R has the right Artin-Rees property.
Proof. Let F be the family of ideals in R with the right Artin-Rees property. Suppose that
J ∈ F and I E R with I 2 ⊆ J ⊆ I. Fix a right ideal KR ⊆ R and an integer n such that
K ∩ J n ⊆ KJ. Then K ∩ I 2n ⊆ K ∩ J n ⊆ KJ ⊆ KI, which proves that I ∈ F .
Next suppose that A, B ∈ F . Fix a right ideal KR ⊆ R, and choose integers m and n such
that K ∩ Am ⊆ KA and (KA) ∩ Bn ⊆ (KA)B. Then for N = max{m, n}, we have
K ∩ (AB)N = K ∩ (AB)m ∩ (AB)n
⊆ (K ∩ Am) ∩ Bn
⊆ (KA) ∩ Bn
⊆ KAB.
Thus AB ∈ F and F is monoidal. So F satisfies (P2). The last two statements follow from
Theorems 2.5 and 2.6.
(cid:3)
The next family that we consider is the family of ideal direct summands of R; that is, the
family of all ideals I E R such that R = I ⊕ J for some ideal J E R. It is well-known that
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
11
an ideal I E R is such a direct summand if and only if I = (e) for some central idempotent
e of R, in which case J = (1 − e) (see, for instance, [18, Ex. 1.7]). In fact, we will work in
the slightly more general setting of multiplicative sets of central idempotents.
Proposition 3.11. Let R be a ring and let S ⊆ R be a multiplicative set of central idem-
potents. Then the family F = {(e) E R : e ∈ S} satisfies (P2). In particular, the family
of ideal direct summands of R is (P3). An ideal maximal with respect to not being an ideal
direct summand is maximal. A ring R is a finite direct product of simple rings if and only if
every ideal of R is an ideal direct summand, if and only if all maximal ideals of R are ideal
direct summands.
Proof. To prove that F is monoidal, let e, e′ ∈ S so that (e) and (e′) are arbitrary ideals in
F . Then ee′ ∈ S, so we have (e)(e′) = (ee′) ∈ F . Also, if I E R with (e)2 ⊆ I ⊆ (e), then
(e)2 = (e2) = (e) implies that I = (e) ∈ F . Thus F satisfies (P2).
Taking S to be the multiplicative set of all central idempotents of R, the family F be-
comes the set of ideal direct summands of R. Now let M ∈ Max(F ′). By the Prime Ideal
Principle 2.5, M is prime. In the factor ring R/M, every ideal is generated by a central
idempotent. We want to prove that R/M is simple. Let 0 6= I E R/M with I = (e) for some
central idempotent e ∈ R/M. Then R/M = (e) ⊕ (1 − e), and (e)(1 − e) = (0). Because
R/M is prime and (e) 6= 0, we must have (1 − e) = 0. Thus R/M = (e) = I, proving that
R/M is simple.
The equivalence of the condition that all ideals are direct summands with the condition
that all maximal ideals are summands follows from Theorem 2.6 along with the fact that every
ideal in F is finitely generated. It remains to show that, if all ideals are direct summands,
then R is a finite direct product of simple rings (the converse being clear). If all ideals of R
are direct summands, then R is a finitely generated semisimple module over the enveloping
ring Re = R ⊗Z Rop. By the classical theory of semisimple modules (see [17, §2]), this means
that R is a direct sum of simple Re-submodules. But a simple Re-submodule is a minimal
ideal, and the desired equivalence easily follows.
(cid:3)
Next we investigate the ideals I E R such that R/I is Dedekind finite. (Recall that a
ring R is Dedekind finite, alternatively called directly finite or von Neumann finite, if ab = 1
implies ba = 1 for any a, b ∈ R.)
Proposition 3.12. The family F = {I E R : R/I is Dedekind finite} is (P3). An ideal
maximal with respect to R/I not being Dedekind finite is prime.
Proof. Let A, B ∈ F and I E R with AB ⊆ I ⊆ A ∩ B. We want to show that I ∈ F . This is
equivalent to showing that if a, b ∈ R with 1 − ab ∈ I, then 1 − ba ∈ I. If 1 − ab ∈ I ⊆ A ∩ B
then we must have 1 − ba ∈ A ∩ B because A, B ∈ F . But then 1 − 2ba + baba = (1 − ba)2 ∈
AB ⊆ I. Also notice that ba − baba = b(1 − ab)a ∈ I because 1 − ab ∈ I. Thus
1 − ba = (1 − 2ba + baba) + (ba − baba) ∈ I,
proving that I ∈ F . The last statement now follows from Theorem 2.5.
(cid:3)
Consideration of flat modules leads to the following example. Notice that the conclusion
here that the family is (P3) is stronger than the one given in [19, (5.7)] when restricted to
the commutative case.
12
MANUEL L. REYES
Proposition 3.13. Let R be a ring and set F = {I E R : (R/I)R is flat}. Then F is a (P3)
family. In particular, F is monoidal and an ideal P maximal with respect to (R/P )R not
being flat is a prime ideal.
Proof. Applying [16, (4.14)] to the exact sequence 0 → I → R → R/I → 0, we see that
(R/I)R is flat if and only if, for every left ideal RL ⊆ R one has I ∩ L ⊆ IL (so that
I ∩ L = IL). Thus
F = {I E R : I ∩ L ⊆ IL for all RL ⊆ R}.
Suppose that I, A, B E R with A, B ∈ F and AB ⊆ I ⊆ A ∩ B. Then I ∈ F because, for
every RL ⊆ R, we have
I ∩ L ⊆ (A ∩ B) ∩ L
= A ∩ (B ∩ L)
= A ∩ (BL)
= (AB)L
⊆ IL.
The last two statements follow from the the fact that (P3) families are monoidal and from
the Prime Ideal Principle 2.5.
(cid:3)
4. Oka families and families constructed from module categories
In this final section, we provide examples of families that satisfy the r-Oka and Oka
properties. Beginning with Proposition 4.7, these examples are produced from classes of
modules that satisfy certain properties. Some of those examples in fact satisfy the strongest
property (P1), but they are proved in this section as applications of the module-theoretic
methods to be introduced.
Kaplansky showed [13, p. 486] that a commutative ring R is a principal ideal ring (that is,
every ideal of R is principal) if and only if every prime ideal of R is principal. Here we give
a certain generalization of this result to noncommutative rings. (A different generalization
for one-sided ideals was given in [27, Theorem 5.11].) We will say that an element x ∈ R
is normal if xR = Rx, in which case we have xR = (x) = Rx. This is equivalent to saying
that both inclusions xR ⊆ (x) and Rx ⊆ (x) are equalities. The set of normal elements is
multiplicatively closed: if x, y ∈ R are normal, then xyR = xRy = Rxy.
Let us refine some of our notation. For a right ideal IR ⊆ R and an element x ∈ R, we
define x−1I = {y ∈ R : xy ∈ I} which is a priori only a right ideal of R. However, if I is
a two-sided ideal and x is normal, then one can easily verify that x−1I = (x)−1I, in which
case x−1I is an ideal.
Proposition 4.1. If S is a multiplicative set of normal elements in a ring R, then the set
F = {(s) : s ∈ S} is a strongly r-Oka family. Every ideal of a ring R is generated by a single
normal element if and only if every prime ideal in R is generated by a normal element.
Proof. Let I, J E R such that (I, J), J −1I ∈ F . Choose x, y ∈ S such that (I, J) = (x) and
J −1I = (y). We claim that I = (xy); because xy ∈ S we will therefore have I ∈ F . Noticing
that I ⊆ (I, J) = (x) = xR, it is easy to verify that I = x(x−1I). But x is normal, giving
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
13
the first of the following equalities: x−1I = (x)−1I = J −1I = (y). Once more, normality of
x gives the final equality in I = x(x−1I) = x(y) = (xy).
For the last statement, we apply Theorem 2.6 to the family F taking S to be the set of
(cid:3)
all normalizing elements.
In the special case of a right noetherian ring, for an ideal to be principally generated by
a normal element, it suffices to test separately that it is principal as a left and a right ideal.
The following may be a known folk result, but we became aware of it in a conversation with
Konstantin Ardakov on Mathematics Stack Exchange.1
Lemma 4.2. Let I be an ideal of a ring R such that I = aR = Rb for some a, b ∈ R.
Assume that one of the following hypotheses holds:
(1) R is right noetherian;
(2) R satisfies ACC on principal right ideals and aR is a projective right ideal (e.g., the
right annihilator of a is zero).
Then I = bR and b is normal.
Proof. Fix x, y ∈ R such that a = xb and b = ay. Because I is a left ideal, the function
λ(r) = xr restricts to a right R-module endomorphism of I = aR. This map is surjective
because a = xb = xay. In both cases (1) and (2), we will prove that λ is an automorphism.
Because λ(b) = a implies that the submodule bR ⊆ aR also has image equal to λ(bR) = aR,
injectivity of λ will in turn imply that bR = aR as desired.
Under hypothesis (1), the right ideal I is a noetherian (and therefore Hopfian) module,
whence this surjection is an automorphism. Now suppose (2) holds, and assume toward a
contradiction that λ has kernel 0 6= K1 ⊆ aR. We can inductively define a chain of right
ideals K1 ( K2 ( · · · by setting Kn/Kn−1 to be the (nonzero) kernel of the homomorphism
aR/Kn−1
∼= aR λ→ aR. The third isomorphism theorem thus yields short exact sequences
0 → Kn → aR → aR → 0
for all n. Because aR is projective, each of these sequences splits. Thus Kn, being a direct
summand of the principal right ideal aR, is itself principal. Therefore the strictly ascending
chain of principal right ideals Kn contradicts condition (2).
(cid:3)
Corollary 4.3. Let R be a ring such that either:
(1) R is right noetherian, or
(2) all principal right ideals of R are projective (e.g., R is a domain) and R satisfies
ACC on principal right ideals.
Then any multiplicative submonoid of the family
F = {I E R I = aR = Rb for some a, b ∈ R}
(including F itself ) is an r-Oka family, and every ideal of R is generated by a single normal
element if and only if every prime ideal of R is principal as both a left and right ideal.
1See http://math.stackexchange.com/q/627609
14
MANUEL L. REYES
Proof. By Lemma 3.8, the family F is closed under products. Next, Lemma 4.2 implies that
every element of F is generated by a single normal element. The result now follows from
Proposition 4.1.
(cid:3)
We also mention the following immediate corollary of Lemma 4.2.
Corollary 4.4. If R is a principal left and right ideal ring, then every ideal of R is generated
by a single normal element.
Now we turn our attention to invertible ideals. Let R be a subring of a ring Q. We will
say that an ideal I E R is invertible with respect to Q (or Q-invertible) if there exists an
(R, R)-subbimodule I ∗ ⊆ Q such that I · I ∗ = I ∗ · I = R. Certainly I ∗ is unique. We claim
that every invertible ideal in R is finitely generated as a right ideal. To see this, let I E R be
invertible and choose xi ∈ I and yi ∈ I ∗ such that Pn
i=1 xiyi = 1. Then for x ∈ I we have
x = P xi(yix) where each yix ∈ I ∗I = R. Thus I = P xiR is finitely generated as a right
ideal. A symmetric argument proves that an invertible ideal is also finitely generated as a
left ideal.
In the argument below, the reader is advised to take care not to confuse products J ∗I of
(R, R)-bimodules inside of Q with ideal quotients J −1I = {x ∈ R : xJ ⊆ I}.
Lemma 4.5. Let R ⊆ Q be as above, and let I, J E R where I ⊆ J and J is invertible. Then
I = J · (J −1I).
Proof. Notice that I ⊆ J implies that J ∗I ⊆ J ∗J = R is an ideal of R. Because J(J ∗I) = I,
it follows that J ∗I ⊆ J −1I. Left multiplying by J gives I = J(J ∗I) ⊆ J(J −1I) ⊆ I, and the
desired equality follows.
(cid:3)
Proposition 4.6. Let R be a subring of a ring Q. Then any monoidal family F of Q-
invertible ideals of R is a strongly r-Oka family. Hence an ideal maximal with respect to not
being invertible is prime. Every nonzero ideal of R is invertible if and only if every nonzero
prime ideal of R is invertible.
If every nonzero ideal of R is Q-invertible, then the following hold:
(1) R is a prime ring;
(2) If every nonzero ideal of Q has nonzero intersection with R, then Q is simple.
Proof. Let I, J E R such that (I, J), J −1I ∈ F . Noting that (I, J)−1I = J −1I, we may
apply Lemma 4.5 to find that I = (I, J) · (J −1I) ∈ F . The second statement follows from
Theorem 2.6 once we notice that the family of all Q-invertible ideals in R is monoidal and
every invertible ideal is finitely generated (as a right ideal).
Now assume that all nonzero ideals of R are Q-invertible. Then (1) follows from the Prime
Ideal Principle 2.5 because the zero ideal is maximal in the complement of the Oka family of
all invertible ideals. Now assume that the hypothesis of (2) holds. To see that Q is simple,
let K E Q be nonzero. Then I = K ∩ R is a nonzero ideal in R. Thus I is invertible and
consequently 1 ∈ R = II ∗ ⊆ KI ∗ ⊆ K. So K = Q, proving that Q is simple.
(cid:3)
Invertibility is typically studied in the context where Q is a ring of quotients. For instance,
let R be a right Ore ring (that is, a ring in which the set of non-zero-divisors is a right Ore
set [16, §10B]) and let Q be its classical right ring of quotients. (In this case, every right
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
15
ideal of Q has nonzero intersection with R, so the hypothesis of (B) above is satisfied.) If
every nonzero prime ideal of R is Q-invertible, the above implies that all nonzero ideals of
R are Q-invertible, that R is prime, and that Q is simple. When these conditions hold, Q
is (simple) artinian if and only if R has finite right uniform dimension. Indeed, the simple
ring Q is equal to its own classical right ring of quotients. As it is prime and nonsingular,
Goldie's theorem states that it is (right) artinian exactly when Q has finite right uniform
dimension. But Q has finite right uniform dimension if and only if R does (see [16, (10.35)]).
(Of course, the ring of quotients Q need not always be artinian if all ideals of R are Q-
invertible. For instance, consider that a ring which is a union of simple artinian subrings
is again simple and equal to its own classical ring of quotients, but may fail to be artinian.
Specifically we may take R = Q to be the direct limit of the matrix rings Rn = M2n(k)
for some division ring k, with Rn considered as a subring of Rn+1 via the natural map
Rn → Rn ⊗ I2 ⊆ Rn ⊗k M2(k) ∼= Rn+1. An explicit argument that Q is not artinian is given
in [17, p. 40].)
The remainder of the results in this section are based on families of ideals in a ring R that
are formed by considering those ideals I such that the (R, R)-bimodule R/I lies in a class C
of bimodules, always assumed to contain the zero module, satisfying certain properties. We
say that a class C of (R, R)-bimodules is closed under extensions if, for every short exact
sequence of bimodules 0 → L → M → N → 0, if L, N ∈ C then N ∈ C. (Taking L = 0,
note that this implies that if M ∼= N and N ∈ C then also M ∈ C.) We say that C is closed
under quotients if every homomorphic image of a bimodule in C is also in C.
The first application of this kind of reasoning gives a method to construct examples of
(P1) and r-Oka families. As one can tell from the hypotheses of (2) and (3) below, it applies
especially well to right noetherian rings. In the following, for a module M we use the term
direct sum power of M to mean a module that is a direct sum of copies of M, that is, a
module of the form M (X) = LX M for some set X.
Proposition 4.7. Let R be a ring and C a class of right R-modules containing the zero
module, and denote F = {I E R : R/I ∈ C}.
(1) If C is closed under extensions, quotients, and arbitrary direct sum powers, then F
satisfies (P1).
(2) Suppose that every ideal in F is finitely generated as a right ideal and that C is closed
under extensions and quotients. Then F satisfies (P1).
(3) Suppose that every principal (equivalently, each finitely generated) ideal in R is finitely
generated as a right ideal and that C is closed under extensions and quotients. Then
F is an r-Oka semifilter.
Proof. In each case, F is a semifilter because C is closed under quotients, and R ∈ F because
0 ∈ C. First suppose that C is as in (1). Let I, J ∈ F and consider the short exact sequence
of right R-modules
0 →
I
IJ
→
R
IJ
→
R
I
→ 0.
The right module I/IJ is annihilated by J, so for some sufficiently large set X we have a
surjection of LX R/J (which is in C by closure under direct sum powers) onto I/IJ (which
16
MANUEL L. REYES
we now see to be in C by closure under quotients). Because I/IJ and R/I belong to C,
which is closed under extensions, we have R/IJ ∈ C and thus IJ ∈ F .
Next suppose that F and C are as in (2), and fix I, J ∈ F . The argument that IJ ∈ F
is identical to that for (1), taking into account the following modificiations. Because of the
hypothesis on F , I ∈ F implies that I/IJ is finitely generated as a right R-module. Thus
the set X above can be made finite, and because C is closed under extensions we obtain that
(R/J)X lies in C.
Now under the hypothesis of (3), let I E R and a ∈ R be such that (I, a), (a)−1I ∈ F .
The assumption on R implies that (I, a)/I is a finitely generated right module in the short
exact sequence
0 →
(I, a)
I
→
R
I
→
R
(a)−1I
→ 0.
Further, the right annihilator of (I, a)/I is equal to (a)−1I. So we have a surjection of R-
modules (R/(a)−1I)n ։ (I, a)/I for some positive integer n, and the argument can proceed
as above to conclude that I ∈ F . Hence F is an r-Oka family.
(cid:3)
Example 4.8. Given an m-system S in a ring R, let C be the class of right R-modules M
such that ann(M) ∩ S 6= ∅. Then C is closed under extensions and quotients, because in
any exact sequence 0 → L → M → N → 0 of right R-modules, one has ann(N) ann(L) ⊆
ann(M) ⊆ ann(N). Also, ann(LX M) = ann(M) for any indexing set X. So C satisfies
the hypothesis of Proposition 4.7(1) and satisfies (P1). Because ann((R/I)R) = I for any
I E R, the family F = {I E R : R/I ∈ C} is equal to the set of all ideals that have nonempty
intersection with S. Thus we recover the same family that we considered in Proposition 3.1.
Example 4.9. Let R be a ring, and let S ⊆ R be a multiplicative subset. Let C be the class
of S-torsion modules: those modules in which every element is annihilated by an element
of S. This class is easily seen to be closed under extensions, quotients, and arbitrary direct
sums. It follows from Proposition 4.7(1) that the corresponding family
F = {I E R : R/I ∈ C} = {I E R : for all r ∈ R, r−1I ∩ S 6= ∅}
is (P1).
Two familiar facts in noncommutative ring theory state that (i) a ring is right artinian
if and only if it is right noetherian and all of its prime factors are right artinian, and (ii) a
left artinian ring is right artinian if and only if it is right noetherian. Here we improve these
criteria for a ring to be right artinian by relaxing the right noetherian hypothesis in each
case to the mere assumption that the two-sided ideals of the ring be finitely generated as
right ideals.
Proposition 4.10. If every principal ideal of a ring R is finitely generated as a right ideal,
then the family F = {I E R : R/I is right artinian} is r-Oka. A ring R is right artinian
if and only if R/P is right artinian for all P ∈ Spec(R) and every ideal in R is finitely
generated as a right ideal. A left artinian ring is right artinian if and only if all of its ideals
are finitely generated as right ideals.
Proof. The fact that F is r-Oka follows from Proposition 4.7(3) applied to the class C of
artinian right R-modules.
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
17
If R is right artinian, then certainly all of its prime factors are right artinian. Also R
is right noetherian, so its ideals are all finitely generated as right ideals. For the converse,
suppose that all of the prime factors of R are right artinian and that its ideals are finitely
generated as right ideals. By the first part of this proposition, the family F defined above is
strongly r-Oka. Because Spec(R) ⊆ F , the finiteness assumption on the ideals of R allows us
to conclude that all ideals lie in F , according to Theorem 2.6. In particular we have 0 ∈ F ,
and R is right artinian.
For the last statement, given a left artinian ring R whose ideals are finitely generated as
right ideals, we merely need to show that R/P is right artinian for all P ∈ Spec(R). For
each prime ideal P E R the factor ring R/P is left artinian and prime, hence semisimple and
right artinian.
(cid:3)
An even more general criterion for a left artinian ring to be right artinian is the result of
D. V. Huynh [12, Theorem 1] that a left artinian ring is right artinian if and only if J/(J 2+D)
is right finitely generated where J ER is the Jacobson radical of R and D ER is the maximal
divisible torsion ideal.
A similar consideration gives a criterion for a ring to be right noetherian. We do not
include the proof, which is identical to the proposition above. Notice that we do not have
a corresponding result telling when a left noetherian ring is right noetherian due to the
existence of prime left noetherian rings which are not right noetherian, such as a twisted
polynomial extension k[x; σ] of a field k by an endomorphism σ : k → k that is not surjective.
Proposition 4.11. If every principal ideal of a ring R is finitely generated as a right ideal,
then the family F = {I ER : R/I is right noetherian} is r-Oka. A ring R is right noetherian
if and only if R/P is right noetherian for all P ∈ Spec(R) and every ideal in R is finitely
generated as a right ideal.
The previous two results raise the question of whether the assumption on the ideals of R
can be further weakened. The following example shows that we cannot hope to do much
better than requiring every ideal to be right finitely genrated. Consider the ring R = (cid:0) R R
0 Q(cid:1).
One can show (using, for example, [17, (1.17)]) that the only ideals in R are
0, I = (cid:18)0 R
0 0(cid:19) , P1 = (cid:18)R R
0(cid:19) , P2 = (cid:18)0 R
0 Q(cid:19) , and R.
0
In particular we see that R has finitely many ideals and thus satisfies both ACC and DCC
∼= Q
on ideals. In addition, R is not prime since I 2 = 0, and Spec(R) = {P1, P2} with R/P1
∼= R. Notice that I is the smallest nonzero ideal of R and that R/I ∼= R × Q
and R/P2
is semisimple, hence right artinian. Let Fa (resp. Fn) denote the family of all ideals whose
factor rings are right artinian (resp.
right noetherian). Then we see that every nonzero
ideal belongs to both Fa and Fn, but the zero ideal is not prime. Then by the Prime Ideal
Principle neither Fa nor Fn can be an Oka family. Additionally R is left artinian but not
right artinian or right noetherian (see [17, (1.22)]), even though it has finitely many ideals
and all of its prime quotients are right noetherian.
Following Artin, Small, and Zhang [3], we say that an algebra R over a commutative
noetherian ring C is strongly right noetherian if, for every commutative noetherian ring D
containing C as a subring, the D-algebra R ⊗C D is right noetherian. (An example of Resco
18
MANUEL L. REYES
and Small [25] shows that a noetherian algebra over a field need not be strongly noetherian.
On the other hand, Bell has shown [5] that this condition is automatically satisfied for a
countably generated right noetherian algebra over an uncountable algebraically closed field.)
These algebras play an important role in noncommutative algebraic geometry, especially in
the theory of point modules [4].
Similar to the above examples, the strong right noetherian property provides another
r-Oka family.
Proposition 4.12. Let C be a commutative noetherian ring and let R be a C-algebra. If
every principal ideal of R is finitely generated as a right ideal, then the semifilter
F = {I E R : R/I is a strongly right noetherian C-algebra}
is r-Oka. If R right noetherian, then it is strongly right noetherian over C if and only if
R/P is strongly right noetherian for all (minimal) primes P ∈ Spec(R).
Proof. Let C be the class of all right R-modules M such that, for any commutative noetherian
overring D ⊇ C, the right R ⊗C D-module M ⊗C D is noetherian. (For brevity, we write
M D = M ⊗C D below.) The statement will directly follow from Proposition 4.7 if we show
that C is closed under extensions and quotients.
Fix a commutative noetherian overring D ⊇ C. Given a short exact sequence 0 → L →
M → N → 0 of R-modules where L, N ∈ C, right exactness of the tensor product gives an
exact sequence LD → M D → N D → 0 of RD-modules. Letting K ⊆ M D denote the kernel
of M D → N D, we obtain a short exact sequence of RD modules:
0 → K → M D → N D → 0.
But K Is a homomorphic image of the noetherian RD-module LD. So in the above, K
and N D are noetherian, implying that M D is noetherian. A similar argument using right
exactness shows that C is also closed under quotient modules.
(cid:3)
All of the examples of Oka families encountered up to this point actually satisfied stronger
properties than the Oka condition. The next proposition will allow us to construct some
Oka families without relying on any of the stronger conditions.
Proposition 4.13. Let k be a commutative ring, and let C be a class of k-modules containing
the zero module that is closed under quotients, extensions, and tensor products. Then for
any k-algebra R, the family F = {I E R : R/I ∈ C} is an Oka semifilter.
Proof. The family F is a semifilter because C is closed under quotients. To see that F is
Oka, let I E R and a ∈ R be such that (I, a), (a)−1I, I(a)−1 ∈ F . Consider the following
short exact sequence of (R, R)-bimodules:
0 →
(I, a)
I
→
R
I
→
R
(I, a)
→ 0.
We want to conclude that R/I ∈ C. Because R/(I, a) ∈ C and C is closed under extensions,
it suffices to show that (I, a)/I ∈ C. Consider the k-bilinear map R/I(a)−1 × R/(a)−1I →
(I, a)/I given by
(r + I(a)−1, s + (a)−1I) 7→ ras + I.
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
19
This induces a k-module homomorphism R/I(a)−1 ⊗k R/(a)−1I → (I, a)/I that is evidently
surjective. The fact that C is closed under tensor products and quotients then implies that
(I, a)/I ∈ C as desired.
(cid:3)
Part (2) of the following proposition gives a method of finding categories C satisfying the
hypothesis of Proposition 4.13.
Lemma 4.14. Let k be a ring and let C be a class of k-modules containing the zero module.
(1) If C is closed under quotients and extensions, then it is closed under finite sums of
modules in the following sense: if N1, . . . , Nm ∈ C are submodules of a single k-module
M, then P Ni ∈ C also.
(2) Assume that k is commutative. If C is closed under quotients and extensions and if
M, N ∈ C with N finitely generated, then M ⊗k N ∈ C. In particular, if every module
in C is finitely generated, then C is closed under tensor products.
Proof. (1) It suffices to prove the statement for a sum of two submodules, for then an
easy inductive argument extends the result to arbitrary finite sums. So suppose N1, N2 are
submodules of the k-module M. Then we have the following short exact sequence:
0 → N1 → N1 + N2 → (N1 + N2)/N1 → 0.
But (N1 + N2)/N1
exact sequence above gives N1 + N2 ∈ C as desired.
∼= N2/(N1 ∩ N2) ∈ C because C is closed under quotients. Thus the short
(2) Let x1, . . . , xn ∈ N form a generating set. Each of the submodules M ⊗ xi ⊆ M ⊗ N
is a homomorphic image of M (under the map − ⊗ xi : M → M ⊗ xi) and thus lies in C.
(cid:3)
i=1 M ⊗ xi ∈ C.
But then M ⊗ N = Pn
Now we take a look at some specific examples.
Example 4.15. Let R be an algebra over a commutative ring k. The family F of all ideals
I of R such that R/I is module-finite over k is an Oka semifilter (using Proposition 4.13
applied to the class C of finitely generated k-modules). Thus, an ideal I E R maximal with
respect to having R/I not module-finite over k is prime.
The result above is related to the study of "just infinite" algebras. An algebra R over a
commutative ring k is is said to be just infinite if R is not finitely generated as a k-module,
but every proper factor of R is module-finite over k. A result attributed to Lance Small (see
[28, Exercise 6.2.5], [23, Prop. 3.2], or [9, Lem. 2.1]) states that in case k is a field, an infinite
dimensional k-algebra has a just infinite factor ring (so that the family above has the Zorn
property) which is prime.
The "maximal implies prime" statement above recovers part of Small's result. (For an
alternative proof, see [8].) To recover the stronger result using the Zorn property, we require
the following Artin-Tate-style lemma. We follow the proof strategy of [23, Lemma 3.1].
Recall that a k-algebra is affine if it is finitely generated as a k-algebra.
Lemma 4.16. Let k be a commutative noetherian ring, and let R be an affine k-algebra. If
M is a finitely generated right R-module with a submodule L ⊆ M such that M/L is finitely
generated as a k-module, then L is also finitely generated as an R-module. (In particular, if
20
MANUEL L. REYES
I is a right ideal of R such that R/I finitely generated over k, then I is finitely generated as
a right ideal.)
Proof. Let {x1, . . . , xm} be k-algebra generating set of R and let {g1, . . . , gn} be a set of R-
module generators of M. Because M/L is finitely generated over k, there exist v1, . . . , vt ∈ M
such that, for V = P kvi, we have M = L + V .
Consequently there exist elements pα ∈ L and qβγ ∈ L such that each gα ∈ pα + V and
each vβxγ ∈ qβγ + V . Let L0 ⊆ L be the R-submodule generated by the finite set of the pα
and qβγ. It is straightforward to see that L0 + V is an R-submodule of M. By construction,
each gα ∈ L0 + V , so M = L0 + V .
Finally, if l ∈ L then l = l0 + v for l0 ∈ L0 ⊆ L and v ∈ V . But also v = l − l0 ∈ L. It
follows that L = L0 + (L ∩ V ). Because k is noetherian and V is finitely generated over k,
the k-submodule L ∩ V is also finitely generated over k. Thus L = L0 + (L ∩ V ) is finitely
generated as an R-module.
(cid:3)
Theorem 4.17. Let R be an affine algebra over a commutative noetherian ring k. The
family
F = {I E R : R/I is finitely generated as a k-module}
satisfies (P1). Thus any just infinite k-algebra is prime. Furthermore, an algebra R that is
not finitely generated as a k-module is just infinite over k if and only if R/P is module-finite
over k for all nonzero P ∈ Spec(R).
Proof. The previous lemma implies that every ideal in F is finitely generated as a right ideal.
Applying Proposition 4.7(2) to the class C of right R-modules that are finitely generated over
k, we conclude that F is a (P1) family. The final two claims respectively follow from the
Prime Ideal Principle 2.5 and Theorem 2.6.
(cid:3)
We present two more examples of Oka families derived from Proposition 4.13. In both
examples, k denotes a commutative ring.
Example 4.18. The class of noetherian k-modules is closed under quotients and extensions,
and consists of finitely generated modules. Then as above, the family of all I E R with R/I
noetherian over k is an Oka semifilter.
Example 4.19. The class of k-modules of finite (composition) length, which is the same as
the category of finitely generated artinian k-modules, is closed under quotients and extensions
and consists of finitely generated modules. So the family of I E R such that R/I is artinian
over k is an Oka semifilter.
We conclude with some thoughts on future directions in the study of the "maximal implies
prime" phenomenon for two-sided ideals in noncommutative rings. When comparing the
theories of Oka families for commutative rings [19] and for right ideals in noncommuative
rings [26] with the theory developed above, we notice one striking difference. In both of the
former two settings, it was shown that Oka families in a ring R are in bijective correspondence
with classes of cyclic R-modules that are closed under extensions (see [19, §4] and [26, §4]).
This allows one to take any module-theoretic property that is preserved by extensions and
produce a corresponding Oka family, a method which led to a number of novel examples of
"maximal implies prime" results in both cases.
A PRIME IDEAL PRINCIPLE FOR TWO-SIDED IDEALS
21
By contrast, we have not found any (bi)module-theoretic characterization of the Oka
families of Definition 2.1 leading to a correspondence between Oka families in a ring R and
certain classes of (bi)modules. While Propositions 4.7 and 4.13 provide sufficient conditions
for a class of bimodules to give rise to an Oka family, both of these methods seem severely
limited by the fact that they can only produce Oka semifilters of ideals. (Note that each of
those results involved a short exact sequence of (R, R)-bimodules the form
0 →
(I, a)
I
→
R
I
→
R
(I, a)
,
where I is an ideal and a is an element of the ring R. In each case, the hypotheses are imposed
with the intent of finding a description of (I, a)/I in terms of the bimodules R/I(a)−1 and
R/(a)−1I. The key noncommutative difficulty seems to be the lack of such a description in
full generality.)
Thus, in conclusion, we offer two general problems to stimulate further work in this direc-
tion.
(1) Is it possible to characterize Oka families of ideals in a noncommutative ring R
in terms of suitable classes of (R, R)-bimodules (perhaps using a more appropriate
definition than the one given in Definition 2.1)?
(2) In lieu of such a characterization, is there a related method that yields more examples
of Oka families that are not semifilters and that do not satisfy conditions (P1)–(P3)?
Even partial progress on these problems seems likely to lead to new "maximal implies prime"
results for ideals in noncommutative rings.
References
1. S. A. Amitsur and Lance W. Small, Algebras over infinite fields, revisited, Israel J. Math. 96 (1996),
no. part A, 23–25. MR 1432724 (97k:16025)
2. D. D. Anderson, A note on minimal prime ideals, Proc. Amer. Math. Soc. 122 (1994), no. 1, 13–14.
MR 1191864 (94k:13001)
3. M. Artin, L. W. Small, and J. J. Zhang, Generic flatness for strongly Noetherian algebras, J. Algebra
221 (1999), no. 2, 579–610. MR 1728399 (2001a:16006)
4. M. Artin and J. J. Zhang, Abstract Hilbert schemes, Algebr. Represent. Theory 4 (2001), no. 4, 305–394.
MR 1863391 (2002h:16046)
5. Jason P. Bell, Noetherian algebras over algebraically closed fields, J. Algebra 310 (2007), no. 1, 148–155.
MR 2307786 (2008f:16050)
6. I. S. Cohen, Commutative rings with restricted minimum condition, Duke Math. J. 17 (1950), 27–42.
MR 0033276 (11,413g)
7. David Eisenbud, Commutative Algebra: with a View Toward Algebraic Geometry, Graduate Texts in
Mathematics, vol. 150, Springer-Verlag, New York, 1995. MR 1322960 (97a:13001)
8. John Farina and Cayley Pendergrass-Rice, A few properties of just infinite algebras, Comm. Algebra 35
(2007), no. 5, 1703–1707. MR 2317639 (2008b:16030)
9. Daniel R. Farkas and Lance W. Small, Algebras which are nearly finite dimensional and their identities,
Israel J. Math. 127 (2002), 245–251. MR 1900701 (2003c:16033)
10. K. R. Goodearl and R. B. Warfield, Jr., An introduction to noncommutative Noetherian rings, second
ed., London Mathematical Society Student Texts, vol. 61, Cambridge University Press, Cambridge, 2004.
MR 2080008 (2005b:16001)
11. I. N. Herstein, Due risultati classici sugli anelli, Univ. e Politec. Torino Rend. Sem. Mat. 21 (1961/1962),
99–102. MR 0148689 (26 #6196)
22
MANUEL L. REYES
12. Dinh Van Huynh, A note on rings with chain conditions, Acta Math. Hungar. 51 (1988), no. 1-2, 65–70.
MR 934584 (89e:16024)
13. Irving Kaplansky, Elementary divisors and modules, Trans. Amer. Math. Soc. 66 (1949), 464–491.
MR 0031470 (11,155b)
14.
, Commutative Rings, revised ed., The University of Chicago Press, Chicago, Ill.-London, 1974.
MR 0345945 (49 #10674)
15. Gunter Krause, Middle annihilators in Noetherian rings, Comm. Algebra 8 (1980), no. 8, 781–791.
MR 566421 (81h:16025)
16. T. Y. Lam, Lectures on Modules and Rings, Graduate Texts in Mathematics, vol. 189, Springer-Verlag,
New York, 1999. MR 1653294 (99i:16001)
17.
18.
, A First Course in Noncommutative Rings, second ed., Graduate Texts in Mathematics, vol.
131, Springer-Verlag, New York, 2001. MR 1838439 (2002c:16001)
, Exercises in Classical Ring Theory, second ed., Problem Books in Mathematics, Springer-Verlag,
New York, 2003. MR 2003255 (2004g:16001)
19. T. Y. Lam and Manuel L. Reyes, A Prime Ideal Principle in commutative algebra, J. Algebra 319 (2008),
no. 7, 3006–3027. MR 2397420
20.
, Oka and Ako ideal families in commutative rings, Rings, modules and representations, Contemp.
Math., vol. 480, Amer. Math. Soc., Providence, RI, 2009, pp. 263–288. MR 2508156 (2010j:13008)
21. J. C. McConnell and J. C. Robson, Noncommutative Noetherian Rings, revised ed., Graduate Studies in
Mathematics, vol. 30, American Mathematical Society, Providence, RI, 2001, With the cooperation of
L. W. Small. MR 1811901 (2001i:16039)
22. Neal H. McCoy, Prime ideals in general rings, Amer. J. Math. 71 (1949), 823–833. MR 0032590 (11,311c)
23. D. S. Passman and W. V. Temple, Representations of the Gupta-Sidki group, Proc. Amer. Math. Soc.
124 (1996), no. 5, 1403–1410. MR 1307556 (96g:20009)
24. C. Procesi, The Burnside problem, J. Algebra 4 (1966), 421–425. MR 0212081 (35 #2956)
25. Richard Resco and L. W. Small, Affine Noetherian algebras and extensions of the base field, Bull. London
Math. Soc. 25 (1993), no. 6, 549–552. MR 1245080 (94m:16029)
26. Manuel L. Reyes, A one-sided prime ideal principle for noncommutative rings, J. Algebra Appl. 9 (2010),
no. 6, 877–919. MR 2748128 (2012a:16007)
27.
, Noncommutative generalizations of theorems of Cohen and Kaplansky, Algebr. Represent. The-
ory 15 (2012), no. 5, 933–975. MR 2969284
28. Louis H. Rowen, Ring Theory, Student Edition, Academic Press Inc., Boston, MA, 1991. MR 1095047
(94e:16001)
29. Boris Sirola, On noncommutative Noetherian schemes, J. Algebra 282 (2004), no. 2, 667–698.
MR 2101080 (2006b:16050)
Department of Mathematics, Bowdoin College, 8600 College Station, Brunswick, ME
04011–8486, USA
E-mail address: [email protected]
URL: http://www.bowdoin.edu/~reyes/
|
1307.0778 | 8 | 1307 | 2013-09-08T02:26:02 | Congruences of the fork extensions. I. The Congruence Extension Property | [
"math.RA"
] | For a slim, planar, semimodular lattice, G. Cz\'edli and E.\,T. Schmidt introduced the fork extension in 2012. In this note we prove that the fork extension has the Congruence Extension Property. This paper has been merged with Part II, under the title Congruences of fork extensions of slim semimodular lattices, see arXiv: 1307.8404 | math.RA | math |
CONGRUENCES OF THE FORK EXTENSIONS. I.
THE CONGRUENCE EXTENSION PROPERTY
G. GR ATZER
Abstract. For a slim, planar, semimodular lattice, G. Cz´edli and E. T. Schmidt
introduced the fork extension in 2012. In this note we prove that the fork ex-
tension has the Congruence Extension Property.
1. Introduction
Let L and K be lattices and let K be an extension of L. We say that the
Congruence Extension Property holds, if every congruence of L has an extension
to K.
Let L be a slim, planar, semimodular lattice, an SPS lattice. As in G. Cz´edli and
E. T. Schmidt [2], inserting a fork to L at the covering square S, firstly, replaces S
by a copy of S7 (see the lattice S7 in Figure 1).
Secondly, if there is a chain u ≺ v ≺ w such that the element v has just been
added and T = {x = u ∧ z, z, u, w = z ∨ u} is a covering square in the lattice L
(and so u ≺ v ≺ w is not on the boundary of L) but x ≺ z at the present stage of
the construction, then we insert a new element y such that x ≺ y ≺ z and y ≺ v.
Let L[S] denote the lattice we obtain when the procedure terminates. We say
that L[S] is obtained from L by inserting a fork to L at the covering square S. See
Figure 2 for an illustration.
Let L be an SPS lattice and S a covering square of L.
In this note we will
examine the connections between the congruence lattice of L and the congruence
lattice of L[S].
Date: July 2, 2013.
2010 Mathematics Subject Classification. Primary: 06C10. Secondary: 06B10.
Key words and phrases. semimodular lattice, fork extension, congruence.
Figure 1. The lattice S7
1
abcdeoi 2
G. GR ATZER
Figure 2. Inserting a fork at S
Theorem 1. Let L be a slim, semimodular, planar lattice. Let S be a covering
square of L. Then the extension of L to L[S] has the Congruence Extension Prop-
erty.
We will use the notations and concepts of lattice theory as in [3].
2. Congruences of finite lattices
The following lemma is trivial but very useful.
Lemma 2. Let L be a finite lattice. Let δ be an equivalence relation on L with
intervals as equivalence classes. Then δ is a congruence relation iff the following
condition and its dual hold:
If x is covered by y (cid:54)= z in L and x ≡ y (mod δ), then z ≡ y ∨ z (mod δ).
(C∨)
Proof. We want to prove that if x ≤ y and x ≡ y (mod δ), then x ∨ z ≡ y ∨
z (mod δ). The proof is a trivial induction first on length[x, y] and then on
length[x, x ∨ z].
(cid:3)
Let (C∧) denote the dual of (C∨).
3. The fork construction
Let L be an SSP lattice. Let S be a covering square of L.
We need some notation for the L[S] construction, see Figure 3.
We start the construction by adding the elements t, zl,1, and zr,1 so that the set
{o, zl,1, zr,1, al, ar, t, i} forms a sublattice S7.
Let al = xl,1, o = yl,1. If k is the largest number so that xl,k, yl,k, and zl,k have
already been defined, and
T = {yl,k+1 = xl,k+1 ∧ yl,k, xl,k+1, yl,k, xl,k = xl,k+1 ∨ yl,k}
is a covering square in L, then we add the element zl,k+1, so we get two new covering
squares {yl,k+1, xl,k+1, yl,k, xl,k} and {zl,k+1, yl,k+1, zl,k, yl,k}. We proceed similarly
on the right.
So L[S] is constructed by inserting the elements in the set
F [S] = {t, zl,1 (cid:31) ··· (cid:31) zl,n, zr,1 (cid:31) ··· (cid:31) zr,m}
S CONGRUENCES OF THE FORK EXTENSIONS. I. CEP
3
Figure 3. Notation for the fork construction
so that {o, zl,1, zr,1, al, ar, t, i} is a sublattice S7, moreover, xl,i (cid:31) zl,i (cid:31) yl,i for
i = 1, . . . , n, and xr,i (cid:31) zr,i (cid:31) yr,i for i = 1, . . . , m. The new elements are black
filled in Figure 3.
Lemma 3. Let L be an SSP lattice with the covering square S. Then L is a
sublattice of L[S]. Therefore, every element x of L[S] has an upper cover x and
a lower cover x in L. Specifically, zl,i = xl,i for i = 1, . . . , n, zr,i = xr,i for
i = 1, . . . , m, and t = i.
If x, y ∈ L[S] and x /∈ F [S], then
x ∨ y = x ∨ y,
and dually.
The following results are well known.
Lemma 4. Let L be an SSP lattice.
(i) An element of L has at most two covers.
(ii) Let a ∈ L. Let a cover the three elements x1, x2, and x3. Then the set
{x1, x2, x3} generates an S7 sublattice.
(iii) If the elements x1, x2, and x3 are adjacent, then the S7 sublattice of (ii) is a
cover-preserving sublattice.
4. Proving Theorem 1
We prove Theorem 1 (and somewhat more) in the next three lemmas.
Lemma 5. Let L be an SSP lattice. Let S = {0, al, ar, i} be a covering square of L.
Let o ≡ i (mod α). Then α extends to L[S] uniquely to a congruence α of L[S].
Proof. We define α as the partition:
π = { [u, v]L[S] [u, v]L is a congruence class of α}
zl,nzr,1zr,2zr,mzl,1zl,2txl,2xl,nxr,2xr,myl,2yl,nyr,mal=xl,1yl,1=yr,1ar=xr,1ioyr,2 4
G. GR ATZER
(cid:91)
To verify that π is indeed a partition of L[S], let
A =
( [u, v]L[S] [u, v]L is a congruence class of α ).
Clearly, L ⊆ A. Since o ≡ i (mod α), there is a congruence class [u, v]L containing
o and i. Hence [u, v]L[S] ∈ π. Therefore, t, zl,1, zr,1 ∈ A. For i = 1, . . . , n, there is a
congruence class [ui, vi]L containing xl,i and yl,i. Hence zl,i ∈ [ui, vi]L[S] ⊆ A and
symmetrically zl,i ∈ A for i = 1, . . . , m. This proves that A = L.
Next we show that the sets in π are pairwise disjoint.
Finally, we verify the substitution properties. Let a, b, c ∈ L[S] and a ≡ b
(mod π). Then there exist u, v ∈ L with u ≡ v (mod α) such that a, b ∈ [u, v]L[S].
There is also an interval [u(cid:48), v(cid:48)]L[S] ∈ π with c ∈ [u(cid:48), v(cid:48)]L[S]. Since u ≡ v (mod α)
and u(cid:48) ≡ v(cid:48) (mod α), it follows that u∨u(cid:48) ≡ v∨v(cid:48) (mod α), so there is a congruence
class [u(cid:48)(cid:48), v(cid:48)(cid:48)] of α containing u∨u(cid:48) and v∨v(cid:48). So u∨u(cid:48), v∨v(cid:48) ∈ [u(cid:48)(cid:48), v(cid:48)(cid:48)]L[S], verifying
the substitution property for joins. The dual proof verifies the substitution property
for meets.
(cid:3)
Lemma 6. Let L be an SSP lattice. Let S = {0, al, ar, i} be a covering square of L.
Let α be a congruence of L satisfying α(cid:101)S = 0S. Then α extends to L[S].
Proof. We are going to define the minimal congruence α of L[S] extending α.
For i = 1, . . . , n, define i as the smallest element in {1, . . . , n} satisfying xl,i ≡ xl,i
The uniqueness statement is obvious.
(mod α); let i be the largest one.
We define α as the partition:
(1)
πd ={ [u, v]L[S] [u, v]L is a congruence class of α}
∪ {t} ∪ { [zl,i, zl,i] i = 1, . . . , n} ∪ { [zr,i, zr,i] i = 1, . . . , m}.
Clearly, α is a partition. To see that α is a congruence, we use Lemma 2. To
verify (C∨), let x ≺ y, z ∈ L[S], y (cid:54)= z, and x ≡ y (mod α). We want to prove
that z ≡ z ∨ y (mod α). By (1), either x ≡ y (mod α) or x, y ∈ [zl,i, zl,i] for some
i = 1, . . . , n (or symmetrically).
If z ∈ L, then z ≡ z ∨ y (mod α) since α is a
First, let x ≡ y (mod α).
congruence, so by (1), z ≡ z ∨ y (mod α). If z /∈ L, that is, if z ∈ F [S], then t (cid:54)= i,
since i does not cover an element not in F [S]. So z = zl,i for some i = 1, . . . , n
(or symmetrically). Since x ≺ z = zl,i and x ∈ L, we get that x = zl,i, i > 1 and
x = yl,i−1. It easily follows that z ≡ y ∨ z (mod α).
Second, let x, y ∈ [zl,i, zl,i] for some i = 1, . . . , n (or symmetrically). Then
x = zl,j and y = zl,j−1 with zl,j ≡ zl,j−1 (mod α) by (1), and so z ≡ z ∨ y
(cid:3)
(mod α).
Lemma 7. Let L be an SSP lattice. Let S = {0, al, ar, i} be a covering square of L.
Let α be a congruence of L satisfying al ≡ i (mod α) and ar (cid:54)≡ i (mod α). Then
α extends to a unique congruence α of L[S].
Proof. Define
(2)
(3)
(4)
i/α = i/α ∪ { zl,i xl,i ≡ xl,1 (mod α)}
xl,i/α = xl,i/α ∪ { zl,k xl,i ≡ xl,k (mod α)}
xr,j/α = xr,j/α ∪ { zr,k xr,j ≡ xr,k (mod α)}
We define α as follows:
for i = 1, . . . , n;
for i, k = 2, . . . , n;
for j, k = 2, . . . , m.
CONGRUENCES OF THE FORK EXTENSIONS. I. CEP
5
Figure 4. Illustrating α for Lemma 7
Let x/α be as defined above for
x ∈ i/α ∪(cid:91)
( xl,i/α i = 2, . . . , n ) ∪(cid:91)
( xr,j/α j = 2, . . . , m );
let x/α = x/α otherwise.
We have to prove that α is a congruence.
The sets x/α, x ∈ L[S], are pairwise disjoint intervals, so they define an equiv-
alence relation α with intervals as equivalence classes.
We again use Lemma 2. To verify (C∨), let x ≺ y, z ∈ L[S], y (cid:54)= z, and x ≡ y
(mod α). We want to prove that z ≡ z ∨ y (mod α) There are four cases to
consider.
Case 1: x, y ∈ L. Then x ≡ y (mod α). If z ∈ L, then x ∨ z ≡ y ∨ z (mod α),
since α is a congruence of L, so x ∨ z ≡ y ∨ z (mod α). So we can assume that
z ∈ F [S].
Since x ≺ y in L, it follows that x, y ∈ id(o) or x, y ∈ fil(xl,n) ∪ fil(xr,m).
Case 1.1: x, y ∈ id(o) and z ∈ F [S]. Let z = zl,i (z = zr,i proceeds similarly).
Note that x ≤ xl,i and xl,i ∨ y = xl,i−1. Since x ≡ y (mod α), it follows that
Case 1.2: x, y ∈ fil(xl,n) ∪ fil(xr,m) and z ∈ F [S]. This cannot happen in view
Then z ∨ y = zl,i−1. We have to verify that zl,i ≡ zl,i−1 (mod α).
xl,i ≡ xl,i−1 (mod α). By (3), zl,i ≡ zl,i−1 (mod α), as claimed.
of x ≺ z.
Case 2: x ∈ L, y ∈ F [S].
Case 2.1: y = zl,i. This cannot happen because by (3) there is no x ∈ L,
x < zl,i in xl,i/α.
Case 2.2: y = zr,i. Then x = yr,i.
zl,nzr,1zr,2zr,mzl,1zl,2txl,2xl,nxr,2xr,myl,2yl,nyr,mal=xl,1yl,1=yr,1ar=xr,1ioyr,2al,2yl,3zl,3xl,3xl,4zl,4yl,4 G. GR ATZER
Case 2.2.1: i (cid:54)= 1. Then z = yr,i−1 and zr,i−1 ≡ y ∨ z = zr,i−1 (mod α)
6
by (4).
Case 2.3: y = t. This cannot happen because there is no x ∈ L with x ≺ t.
Case 2.2.2: i = 1, so y = yr,1. Then z = zl,1 and we proceed as in Case 2.2.1.
Case 3: x ∈ F [S], y ∈ L. Then x = zl,i and y = xl,i. Joining them with z, if we
get two distinct elements, they are of the same form, of the pair t, i, all of which
are congruence modulo α.
Case 4: x, y ∈ F [S]. Let x = zl,i and y = zl,i−1 (or symmetrically). Then
z = xl,i and the conclusion follows from (4).
(mod α). We distinguish the same four cases as for (C∧).
so x ∧ z ≡ y ∧ z (mod α). So we can assume that z ∈ F [S].
We have verified condition (C∨) of Lemma 2. The final step is to verify (C∧).
So let x cover y (cid:54)= z in L[S] and x ≡ y (mod α). We want to prove that z ≡ x∧z
Case 1: x, y ∈ L. Then x ≡ y (mod α). If z ∈ L, then x ∧ z ≡ y ∧ z (mod α),
Since z ∈ F [S] is covered by exactly one element z of L, so x = x, and y ∧ z, z ∈
Case 2: x ∈ L, y ∈ F [S], say y = zl,i. Then x = xl,i, so y = zl,i and z = xl,i+1.
Case 3: x ∈ F [S], y ∈ L. This cannot happen because there is no x ∈ F [S] and
Case 4: x, y ∈ F [S]. Let x = zl,i and y = zl,i+1 (or symmetrically). Then
(cid:3)
x/α, and the claim follows.
By (3), y ≡ y ∧ z (mod α), as claimed.
y ∈ L with x (cid:31) y and x ≡ y (mod α).
z = yl,i and the conclusion follows from the fact that yl,i ≡ yl,i+1 (mod α).
The uniqueness statement is obvious.
References
[1] G. Cz´edli and E. T. Schmidt, Slim semimodular lattices. I. A visual approach, Order 29 (2012),
481-497.
[2]
[3] G. Gratzer, Lattice Theory: Foundation. Birkhauser Verlag, Basel, 2011. xxix+613 pp. ISBN:
, Slim semimodular lattices. II. A description by patchwork systems, Order.
978-3-0348-0017-4.
Department of Mathematics, University of Manitoba, Winnipeg, MB R3T 2N2, Canada
E-mail address, G. Gratzer: [email protected]
URL, G. Gratzer: http://server.maths.umanitoba.ca/homepages/gratzer/
|
1106.2680 | 1 | 1106 | 2011-06-14T11:40:11 | $\delta$-superderivations of semisimple Jordan superalgebras | [
"math.RA",
"math.QA"
] | We described $\delta$-derivations and $\delta$-superderivations of simple and semisimple finite-dimensional Jordan superalgebras over algebraic closed fields with characteristic $p\neq2$. We constructed new examples of 1/2-derivations and 1/2-superderivations of simple Zelmanov's superalgebra $V_{1/2}(Z,D).$ | math.RA | math |
δ-SUPERDERIVATIONS OF SEMISIMPLE JORDAN SUPERALGEBRAS
Ivan Kaygorodov
Sobolev Inst. of Mathematics
Novosibirsk, Russia
[email protected]
Abstract:
We described δ-derivations and δ-superderivations of simple and semisimple finite-dimensional Jordan
superalgebras over algebraic closed fields with characteristic p 6= 2. We constructed new examples of 1
2 -
derivations and 1
2 -superderivations of simple Zelmanov's superalgebra V1/2(Z, D).
Key words: δ-(super)derivation, Jordan superalgebra.
В в е д е н и е
Антидифференцирования, то есть такие линейные отображения µ алгебры A, что
µ(xy) = −µ(x)y − xµ(y),
изучались в работах [1, 2]. Впоследствии, в работах В. Т. Филиппова было введено понятие δ-
дифференцирования, то есть такого линейного отображения φ алгебры A, что для фиксированного эле-
мента δ из основного поля и произвольных элементов x, y ∈ A верно
φ(xy) = δ(φ(x)y + xφ(y)).
Он рассматривал δ-дифференцирования первичных лиевых, альтернативных и мальцевских нелиевых ал-
гебр [3, 4, 5]. В дальнейшем, δ-дифференцирования йордановых алгебр и супералгебр, а также простых
супералгебр Ли рассматривались И. Б. Кайгородовым в работах [6, 7, 8, 9]; δ-супердифференцирования
супералгебр йордановой скобки рассматривались И. Б. Кайгородовым и В. Н. Желябиным в рабо-
те [10]; П. Зусманович рассматривал δ-супердифференцирования первичных супералгебр Ли в [11].
Отметим, что в [12] сделан подробный обзор, посвященный изучению δ-дифференцирований и δ-
супердифференцирований.
В данной работе дается полное описание δ-дифференцирований и δ-супердифференцирований простых
конечномерных неунитальных йордановых супералгебр над алгебраически замкнутым полем характери-
стики p > 2. В результате мы имеем полное описание δ-дифференцирований и δ-супердифференцирований
полупростых конечномерных йордановых супералгебр над алгебраически замкнутым полем характери-
стики отличной от 2. В частности показано, что полупростые конечномерные йордановы супералгебры над
алгебраически замкнутым полем характеристики нуль не имеют нетривиальных δ-дифференцирований и
δ-супердифференцирований.
§1 Основные факты и определения.
Пусть F (cid:22) поле характеристики p 6= 2. Алгебра A над полем F называется йордановой, если она
удовлетворяет тождествам
xy = yx,
(x2y)x = x2(yx).
Пусть G (cid:22) алгебра Грассмана над F , заданная образующими 1, ξ1, . . . , ξn, . . . и определяющими соотноше-
ниями: ξ2
i = 0, ξiξj = −ξjξi. Элементы 1, ξi1ξi2 . . . ξik , i1 < i2 < . . . < ik, образуют базис алгебры G над F .
Обозначим через G0 и G1 подпространства, порожденные соответственно произведениями четной и нечет-
ной длинны; тогда G представляется в виде прямой суммы этих подпространств: G = G0 ⊕ G1, при этом
справедливы соотношения GiGj ⊆ Gi+j(mod2), i, j = 0, 1. Иначе говоря, G является Z2-градуированной
алгеброй (или супералгеброй) над F .
Пусть теперь A = A0 ⊕ A1 (cid:22) произвольная супералгебра над F . Рассмотрим тензорное произведение
F -алгебр G ⊗ A. Его подалгебра
G(A) = G0 ⊗ A0 + G1 ⊗ A1
1
называется грассмановой оболочкой супералгебры A.
Пусть Ω (cid:22) некоторое многообразие алгебр над F . Супералгебра A = A0 ⊕ A1 называется Ω-
супералгеброй, если ее грассманова оболочка G(A) является алгеброй из Ω. В частности, A = A0 ⊕ A1
называется йордановой супералгеброй, если ее грассманова оболочка G(A) является йордановой алгеброй.
Далее для однородного элемента x супералгебры A = A0 ⊕ A1 будем считать p(x) = i, если x ∈ Ai, а для
элемента x ∈ A = A0 ⊕ A1 через xi обозначим проекцию на Ai.
Классификация простых конечномерных йордановых супералгебр над алгебраически замкнутым по-
лем характеристики ноль была приведена в работах И. Кантора и В. Каца [13, 14]. В последствии М.
Расином и Е. Зельмановым [15], были описаны простые конечномерные йордановы супералгебры с полу-
простой четной частью над алгебраически замкнутым полем характеристики отличной от 2. В работе Е.
Зельманова и К. Мартинез [16] была дана классификация простых унитальных конечномерных йордано-
вых супералгебр с неполупростой четной частью над алгебраически замкнутым полем характеристики
p > 2. Простые неунитальные конечномерных йордановы супералгебры над алгебраически замкнутым
полем характеристики p > 2 были описаны в работе Е. Зельманова [17]. Также, в этой работе была предъ-
явлена структура полупростых конечномерных йордановых супералгебр над алгебраически замкнутым
полем характеристики p 6= 2.
Приведем некоторые примеры йордановых супералгебр.
1.1. Супералгебра векторного типа J(Γ, D). Пусть Γ = Γ0 ⊕Γ1 (cid:22) ассоциативная суперкоммутатив-
ная унитальная супералгебра с ненулевым четным дифференцированием D. Рассмотрим J(Γ, D) = Γ ⊕ Γx
(cid:22) прямую сумму пространств, где Γx (cid:22) изоморфная копия пространства Γ. Тогда операция умножения
на J(Γ, D), которую мы обозначем через ·, определяется формулами
a · b = ab, a · bx = (ab)x, ax · b = (−1)p(b)(ab)x, ax · bx = (−1)p(b)(D(a)b − aD(b)),
где a, b однородные элементы из Γ и ab (cid:22) произведение в Γ.
В дальнейшем, нас будет интересовать случай Γ = B(m), где B(m) = F [a1, . . . , amap
i = 0] (cid:22) алгебра
усеченных многочленов от m четных переменных над полем характеристики p > 2.
1.2. Супералгебра Каца. Простая девятимерная супералгебра Каца K9 над полем характеристики
3 определяется следующим образом:
K9 = A ⊕ M, (K9)0 = A, (K9)1 = M ,
A = F e + F uz + F uw + F vz + F vw,
M = F z + F w + F u + F v .
Умножение задано условиями:
e2 = e, e (cid:22) единица алгебры A, em = 1
2 m, для любого m ∈ M,
[u, z] = uz, [u, w] = uw, [v, z] = vz, [v, w] = vw,
[z, w] = e, [u, z]w = −u, [v, z]w = −v, [u, z][v, w] = 2e,
а все остальные ненулевые произведения получаются из приведенных либо применением одной из косо-
симметрий z ↔ w, u ↔ v, либо применением одновременной подстановки z ↔ u, w ↔ v. Отметим, что
супералгебра K9 не является унитальной.
1.3. Супералгебра V1/2(Z, D). Пусть Z (cid:22) ассоциативно-коммутативная F -алгебра с единицей e и
дифференцированием D : Z → Z, удовлетворяющая двум условиям
i) Z не имеет собственных D-инвариантных идеалов,
ii) D обнуляет только элементы вида F e.
Рассмотрим Zx как изоморфную копию алгебры Z. Определим на векторном пространстве V (Z, D) =
Z + Zx структуру супералгебры. Положим A = Z и M = Zx (cid:22) соответственно четная и нечетная части.
Умножение · зададим следующим образом
a · b = ab, a · bx =
1
2
(ab)x, ax · bx = D(a)b − aD(b),
для произвольных элементов a, b ∈ Z. Полученную супералгебру будем обозначать как V1/2(Z, D).
Как было отмечено выше, для фиксированного элемента δ из основного поля, под δ-
дифференцированием супералгебры A мы понимаем линейное отображение φ : A → A, которое при
произвольных x, y ∈ A удовлетворяет условию
φ(xy) = δ(φ(x)y + xφ(y)).
2
Под центройдом Γ(A) супералгебры A мы будем понимать множество всех линейных отображений
χ : A → A, при произвольных a, b ∈ A удовлетворяющих условию
χ(ab) = χ(a)b = aχ(b).
Определение 1-дифференцирования совпадает с обычным определением дифференцирования; 0-
дифференцированием является произвольный эндоморфизм φ алгебры A такой, что φ(A2) = 0. Ясно,
что любой элемент центроида супералгебры является 1
2 -дифференцированием.
Ненулевое δ-дифференцирование φ будем считать нетривиальным δ-дифференцированием, если δ 6=
0, 1 и φ /∈ Γ(A).
Под суперпространством мы понимаем Z2-градуированное пространство. Рассматривая пространство
эндоморфизмов End(A) супералгебры A, мы можем задать Z2-градуировку, положив четными отображе-
ния φ, такие, что φ(Ai) ⊆ Ai, а под нечетными подразумевать отображения φ, такие что φ(Ai) ⊆ Ai+1.
Однородный элемент ψ суперпространства End(A) называется супердифференцированием, если для од-
нородных x, y ∈ A0 ∪ A1 выполненно
ψ(xy) = ψ(x)y + (−1)p(x)p(ψ)xψ(y).
Для фиксированного элемента δ ∈ F определим понятие δ-супердифференцирования супералгебры
A = A0 ⊕ A1. Однородное линейное отображение φ : A → A будем называть δ-супердифференцированием,
если для однородных x, y ∈ A0 ∪ A1 выполнено
φ(xy) = δ(φ(x)y + (−1)p(x)p(φ)xφ(y)).
Рассмотрим супералгебру Ли A = A0 ⊕ A1 с умножением [., .] и зафиксируем элемент x ∈ Ai. Тогда
Rx : y → [x, y] является супердифференцированием супералгебры A и его четность p(Rx) = i.
Под суперцентроидом Γs(A) супералгебры A мы будем понимать множество всех однородных линейных
отображений χ : A → A, для произвольных однородных элементов a, b удовлетворяющих условию
χ(ab) = χ(a)b = (−1)p(a)p(χ)aχ(b).
Определение 1-супердифференцирования совпадает с определением обычного супердифференцирова-
ния; 0-супердифференцированием является произвольный эндоморфизм φ супералгебры A такой, что
φ(A2) = 0. Ненулевое δ-супердифференцирование будем считать нетривиальным, если δ 6= 0, 1 и φ /∈ Γs(A).
Легко видеть, что четное δ-супердифференцирование будет являться δ-дифференцированием. Далее
мы будем этим пользоваться дополнительно не отмечая.
§2 δ-дифференцирования и δ-супердифференцирования простых конечномерных
йордановых супералгебр.
В данном параграфе мы рассмотрим δ-дифференцирования и δ-супердифференцирования неуниталь-
ных простых конечномерных йордановых супералгебр над алгебраически замкнутым полем характери-
стики p > 2, т.е. супералгебр K9 над полем характеристики 3 и супералгебры V1/2(Z, D) над полем
характеристики p > 2. В результате, учитывая результаты работ [6, 8], мы будем иметь полное описание
δ-дифференцирований и δ-супердифференцирований простых конечномерных йордановых супералгебр
над алгебраически замкнутым полем характеристики отличной от 2.
Далее через hx, yi обозначим векторное пространство, порожденное векторами x, y.
ЛЕММА 1. Супералгебра K9
супердифференцирований.
не
имеет нетривиальных
δ-дифференцирований
и
δ-
ДОКАЗАТЕЛЬСТВО. Пусть φ (cid:22) нетривиальное δ-дифференцирование супералгебры K9. Легко ви-
деть, что при m ∈ (K9)1 выполнено
Следовательно, учитывая φ(e)1u = 0 и φ(e)1v = 0, имеем φ(e)1 = 0.
0 = φ(me − em) = 2δmφ(e)1.
Рассмотрим ограничение φ на (K9)0. Заметим, что (K9)0 (cid:22) йорданова алгебра билинейной формы, по
2 получаем φ(x)0 = αx, x ∈ (K9)0, α ∈ F, в частности видим φ(e)0 = αe, а при
доказанному в [6], при δ = 1
δ 6= 1
2 имеем φ((K9)0)0 = 0.
3
Если δ 6= 1
2 и m ∈ (K9)1, то
φ(m) = 2φ(em) = 2δeφ(m) = 2δφ(m)0 + δφ(m)1.
Откуда получаем φ(m) = 0 и φ(m1m2) = δφ(m1)m2 + δm1φ(m2) = 0, для произвольных m1, m2 ∈ (K9)1,
что влечет тривиальность φ.
Если δ = 1
2 и m ∈ (K9)1, то 2φ(m) = 4φ(em) = 2eφ(m) + αm, что дает φ(m) = φ(m)0 + αm. Далее
можем считать, что α = 0 и φ(m) = φ(m)0. Заметим, что 0 = φ(m2) = φ(m)m, откуда
φ(w) = αwuw + βwvw, φ(z) = αzuz + βzvz,
φ(u) = αuuz + βuuw, φ(v) = αvvw + βvvz.
φ(uz) = −(φ(u)z + uφ(z)) = βzz − βuu,
Легко видеть, что
откуда
αuuz + βuuw = φ(u) = −φ((uz)w) = (βzz − βuu)w + uz(αwuw + βwvw) = βze − βuuw + 2βwe.
т.е. αu = βu = 0. Аналогично получаем αγ = βγ = 0, при γ ∈ {z, w, u, v}. Таким образом, φ (cid:22) тривиально.
Если φ (cid:22) нечетное δ-супердифференцирование, то легко получить, что
Также видим, что при δ 6= 1
2 верно
φ(e) = δφ(e)e + δeφ(e) = δφ(e) и φ(e) = 0.
φ(a) = φ(ea) = δ
2 φ(a) и φ(m) = 2φ(em) = 2δφ(m),
что нам дает тривиальность φ.
Пусть δ = 1
2 и для q ∈ {u, z, v, w} верно
φ(q) = χqe + αquz + βquw + γqvz + µqvw.
Теперь легко получить
φ(u) = −φ([u, z]w) = φ(uz)w + (uz)φ(w) = −(φ(u)z − uφ(z))w + (uz)φ(w),
φ(u) = φ([u, w]z) = −(φ(uw)z + (uw)φ(z)) = (φ(u)w − uφ(w))z − (uw)φ(z).
Ясно, что (φ(u)w − uφ(w))z − (uw)φ(z) ∈ ζµuvw + he, uw, uz, vwi, ζ = ±1, и −(φ(u)z − uφ(z))w + (uz)φ(w) ∈
he, uw, uz, vwi, то есть µu = 0. Аналогично получаем, что γu = 0. Проводя аналогичные рассуждения для
v, z, w, мы получми, что φ(q) = χqe + yqq, yq ∈ (K9)1.
Легко заметить, что
0 = 2φ(e) = 2φ([z, w]) = φ(z)w − zφ(w) = −χzw + (yzz)w + χwz − z(yww),
а также, что (yzz)w − z(yww) ∈ hu, vi, то есть χz = χw = 0. Аналогично вычисляя, имеем χu = χz = 0.
Далее заметим, что
0 = φ(e) = −φ([u, z][v, w]) = φ(uz)vw + uzφ(vw) =
1
2
((φ(u)z − uφ(z))vw + uz(φ(v)w − vφ(w))) = ζ1βuw + ζ2γzv + ζ3γvz + ζ4βwu,
где ζi принимает значения +1 или −1. Отсюда имеем, что βu = βw = γz = γv = 0.
Осталось отметить, что
0 = −φ(e) = −φ(uv) = φ(u)v − uφ(v) = (αuuz)v − u(µvvw),
что дает µv = αu = 0. Аналогично показываем, что φ(z) = φ(w) = 0. Таким образом, имеем φ (cid:22) триви-
ально. Лемма доказана.
ЛЕММА 2. Пусть V1/2(Z, D) супералгебра над полем характеристики 3, Z (cid:22) ее четная часть, ψ
(cid:22) дифференцирование алгебры Z, связанное c D соотношением ψ(a)D(b) = D(a)ψ(b) для любых a, b ∈ Z.
Тогда существует z ∈ Z, что D(z) (cid:22) обратим и ψ = cD при некотором c ∈ Z.
4
ДОКАЗАТЕЛЬСТВО. Легко заметить, что D(a)ψ = ψ(a)D. Мы покажем, что для некоторого z ∈ Z
элемент D(z) будет обратим, откуда будем иметь ψ = (D(z)−1ψ(z))D, то есть искомое.
Заметим, что D(a3) = 3D(a)a2 = 0. Отсюда, легко получить, что a3Z является D-инвариантным
идеалом в Z. Таким образом, либо a3Z = Z, либо a3Z = 0. Если для проивольного a ∈ Z верно a3Z = Z,
то Z (cid:22) поле. В противном случае, найдется элемент a ∈ Z, такой что a3 = 0. Рассмотрим R = {a ∈ Za3 =
0} 6= 0 (cid:22) идеал в Z, отличный от Z. Пусть I ✁ Z, тогда любой элемент a ∈ I не является обратимым,
и, следовательно, a3 не является обратимым, а также верно a3Z ✁ Z. Учитывая то, что a3Z является D-
инвариантным идеалом, мы получаем a3Z = 0, что влечет a3 = 0 и I ⊆ R. Поэтому, что R (cid:22) наибольший
идеал в Z. Легко понять, что D(R) * R. Иначе бы R был D-инвариантным идеалом и совпадал с Z.
Таким образом, для некоторого z ∈ R верно D(z) /∈ R. Следовательно, D(z) (cid:22) обратим и, автоматически,
доказывает условие леммы.
ЛЕММА 3. Пусть φ (cid:22) нечетное 1
2 -супердифференцирование
супералгебры V1/2(Z, D) над полем характеристики 3, заданное условиями φ(a) = ψ(a)x и φ(ax) = µ(a).
Тогда µ(a) = D(ψ(a)) + aµ(e) и ψ = cD при некотором c ∈ A.
2 -дифференцирование или нечетное 1
ДОКАЗАТЕЛЬСТВО. Заметим, что для произвольных a, b ∈ A выполняется
ψ(ab)x = φ(ab) = 2(φ(a) · b + a · φ(b)) = (ψ(a)b + aψ(b))x,
то есть, ψ (cid:22) дифференцирование A. Также верно, что
2µ(ab) = φ(a · bx) = 2(φ(a) · bx + a · φ(bx)) = 2(ψ(a)x · bx + aµ(b)) = 2D(ψ(a))b − 2ψ(a)D(b) + 2aµ(b),
то есть
что, при b = e, влечет
µ(ab) = D(ψ(a))b − ψ(a)D(b) + aµ(b),
Таким образом, при подстановке (2) в (1), мы имеем
µ(a) = D(ψ(a)) + aµ(e).
D(ψ(ab)) + abµ(e) = D(ψ(a))b − ψ(a)D(b) + aD(ψ(b)) + abµ(e),
что, путем прямых вычислений, нам дает
D(a)ψ(b) = ψ(a)D(b).
(1)
(2)
Откуда, по лемме 2, мы получаем, что ψ = cD, для некоторого c ∈ A. Лемма доказана.
ЛЕММА 4. Пусть φ (cid:22) нетривиальное δ-дифференцирование супералгебры V1/2(Z, D) над полем ха-
рактеристики p 6= 2. Тогда δ = 1
2 и φ(y) = (1 + p(y))z · y для фиксированного z ∈ A \ {F e}.
ДОКАЗАТЕЛЬСТВО. Легко заметить, что φ(e) = 2δφ(e)e = 2δφ(e)0 +δφ(e)1, т.е. либо δ 6= 1
2 и φ(e) = 0,
либо δ = 1
2 и φ(e) ∈ A.
Если mx ∈ M, то
φ(mx) = 2φ(e · mx) = 2δφ(e) · mx + 2δe · φ(mx) = δ(φ(e)m)x + 2δφ(mx)0 + δφ(mx)1.
Откуда видим, что при δ 6= 1
Для произвольного a ∈ A верно
2 будет φ(M ) = 0, а при δ = 1
2 получим φ(mx)1 = (φ(e)m)x.
φ(a) = φ(e · a) = δφ(e) · a + δe · φ(a) = δφ(e) · a + δφ(a)0 +
δ
2
φ(a)1.
Откуда, выполнено одно из условий:
2 , 2 и φ = 0;
1) δ 6= 1
2) δ = 2, p 6= 2, 3 и φ(a) ∈ M ;
3) δ = 1
4) δ = 1
2 , p 6= 2, 3 и φ(a) = φ(e)a;
2 и p = 3.
Покажем, что второй случай не дает нетривиальных 2-дифференцирований. Пусть φ (cid:22) 2-
дифференцирование и φ(a) = φax, тогда
0 = φ(a · x) = 2φ(a) · x = 2φax · x = 2D(φa),
5
откуда, пользуясь тем, что D обнуляет только элементы вида αe, имеем φa = αe.
Заметим, что
φa2 x = φ(a2) = 4φ(a) · a = 2(φaa)x,
т. е. φa = 0 при a ∈ A \ {F e}. Полученное дает тривиальность φ.
Рассмотрим третий случай, т.е. δ = 1
2 и p 6= 2, 3. Ясно, что отображение ψ заданное по правилу
ψ(a) = za, ψ(mx) = (zm)x, для z, a, m ∈ A
(3)
является 1
являться 1
2 -дифференцированием. Поэтому, отображение χ определенное как χ = φ − ψ также будет
2 -дифференцированием. Покажем, что χ = 0.
Легко видеть, что χ(A) = 0 и χ(mx) = φ(mx)0. Заметим, что
0 = 2χ(x · x) = 2χ(x) · x = χ(x)x,
т.е. χ(x) = 0. Следовательно, можем получить χ(ax) = χ(a) · x + a · χ(x) = 0. Откуда, χ (cid:22) тривиально и φ
имеет искомый вид.
Заметим, что для φ, определенного по правилу (3), верно
φ(ax · bx) = φ(D(a)b − aD(b)) = z(D(a)b − aD(b)) = (ax) · ((bz)x) + abD(z) = (ax) · φ(bx) + abD(z).
Откуда видим, что φ будет являться нетривиальным 1
2 -дифференцированием только если D(z) 6=
0, т.е. если z ∈ A \ {F e}. Ясно, что отображение, заданное полученным образом будет также 1
2 -
дифференцированием и над полем характеристики p = 3.
Рассмотрим четвертый случай, т.е. δ = 1
2 и p = 3. В силу того, что отображение, заданное формулами
(3), является 1
2 -дифференцированием данной супералгебры над полем характеристики 3, то мы можем
считать φ(A) ⊆ M, φ(M ) ⊆ A. Положим, что φ(a) = ψ(a)x и φ(ax) = µ(a). По лемме 3, мы получаем, что
ψ = cD, для некоторого c ∈ A.
Для того, чтобы отображение φ, заданное по правилу φ(a) = (cD(a))x, φ(ax) = D(cD(a)) + az для
2 -дифференцированием, нам необходимо проверить выполнение равенства
некоторых c, z ∈ A, являлось 1
Непосредственными вычислениями, имеем
φ(ax · bx) =
1
2
(φ(ax) · bx + ax · φ(bx)).
(cD(D(a)b − aD(b)))x = φ(ax · bx) =
1
2
откуда
(φ(ax) · bx + ax · φ(bx)) = ((D(cD(a)) + az)b)x + (a(D(cD(b)) + bz))x,
acD2(b) = D(c)D(a)b + aD(c)D(b) + 2azb,
что при b = a = e дает z = 0, а при b = e дает D(c)D(a) = 0. Учитывая, что по лемме 2 существует
w ∈ A, что D(w) является обратимым, то D(c) = 0 и c = αe, α ∈ F. Соответственно, получаем равенство
αaD2(b) = 0. Допустим, что α 6= 0, тогда D2(b) = 0. Отсюда получаем, что D(b) = βbe, βb ∈ F. Заметим,
что βb2e = D(b2) = 2D(b)b = 2βbb, то есть βb = 0, что дает D = 0 и противоречение с тем, что D ненулевое
отображение. Таким образом, мы имеем α = 0 и ψ = µ = 0. Лемма доказана.
ЛЕММА 5. Пусть φ (cid:22) нетривиальное нечетное δ-супердифференцирование супералгебры V1/2(Z, D)
над алгебраически замкнутым полем характеристики p 6= 2. Тогда δ = 1
1) если p 6= 3, то φ(A) = 0, φ(ax) = az для некоторого z ∈ A \ {0};
2) если p = 3, то φ(a) = (αD(a))x, φ(ax) = D(αD(a)) + az для некоторых z ∈ A, α ∈ F, таких что
элементы z и α одновременно не обращаются в нуль.
2 и
ДОКАЗАТЕЛЬСТВО. Ясно, что φ(A) ⊆ M и φ(M ) ⊆ A. Легко заметить, что
т.е. φ(e) = 0. Пусть a ∈ A, тогда
φ(e) = 2δφ(e) · e = δφ(e),
φ(a) = φ(e · a) = δeφ(a) =
δ
2
φ(a), φ(ax) = 2φ(e · ax) = 2δe · φ(ax) = 2δφ(ax).
Откуда, выполнено одно из условий:
1) δ = 1
2 , p = 3;
6
2) δ = 1
2 , p 6= 3;
3) δ = 2, p 6= 3;
4) φ = 0.
Третий случай дает φ(M ) = 0 и 0 = φ(a · x) = 2φ(a) · x, то есть φ(A) = 0.
Второй случай дает то φ(A) = 0 и φ(ax) = φa. Тогда
0 = 2φ(ax · bx) = φa · bx − ax · φb,
что влечет φab = aφb, т.е. φa = aφe. Для того, чтобы отображение φ определенное как φ(A) = 0, φ(ax) = za
являлось 1
2 -супердифференцированием, необходимо чтобы выполнялось условие
φ(a · bx) =
1
2
φ((ab)x) =
1
2
abz =
1
2
(0 · bx + a · (bz)) =
1
2
(φ(a) · bx + a · φ(bx)).
Откуда видим, что φ является 1
2 -супердифференцированием. Заметим, что
φ(x · x) = 0 6= −
1
2
zx = −x · φ(x),
т.е. φ является нетривиальным 1
2 -супердифференцированием.
Первый случай, то есть p = 3 и δ = 1
2 , рассмотрим более детально. Положим, что φ(a) = ψ(a)x и
φ(ax) = µ(a). По лемме 3, мы получаем, что ψ = cD, для некоторого c ∈ A.
Для того, чтобы отображение φ, заданное по правилу φ(a) = (cD(a))x, φ(ax) = D(cD(a))+ az для неко-
2 -супердифференцированием, нам необходимо проверить выполнение равенства
торых c, z ∈ A, являлось 1
φ(ax · bx) =
1
2
(φ(ax) · bx − ax · φ(bx)).
Непосредственными вычислениями, имеем
(cD(D(a)b − aD(b))x = φ(ax · bx) =
1
2
(φ(ax) · bx − ax · (bx)) = ((D(cD(a)) + az)b)x − (a(D(cD(b)) + bz))x,
что влечет D(c)(D(a)b − aD(b)) = 0, то есть, при b = e, легко получаем D(c)D(a) = 0. Согласно лемме 2,
существует такой элемент w ∈ A, что D(w) (cid:22) обратим, что влечет D(c) = 0 и по условиям определения
дифференцирования D для алгебры A, заключаем, что c = αe. Легко проверить, что отображение φ,
заданное по правилам
φ(a) = (αD(a))x, φ(ax) = D(αD(a)) + az
для z ∈ A, α ∈ F, является нечетным 1
то
2 -супердифференцированием. Ясно, что если α 6= 0 и a 6= βe, β ∈ F ,
φ(a · e) − a · φ(e) = φ(a) 6= 0,
то есть φ (cid:22) нетривиальное 1
2 -супердифференцирование. Лемма доказана.
Отметим, что леммы 4 и 5 дают примеры новых нетривиальных 1
супердифференцирований, не являющихся операторами правого умножения.
2 -дифференцирований и 1
2 -
Напомним, что в работе [10] были построены новые примеры нетривиальных 1
2 -дифференцирований
для простых унитальных супералгебр векторного типа J(Γ, D), построенных на супералгебрах с тривиаль-
ной нечетной часть. Оказалось, что каждый оператор правого умножения Rz, где D(z) 6= 0, z ∈ Γ являет-
ся нетривиальным 1
2 -дифференцированием, и, в свою очередь, все нетривиальные δ-дифференцирования
данных супералгебр исчерпываются операторами правого умножения данного вида.
ТЕОРЕМА 6. Пусть J (cid:22) простая конечномерная йорданова супералгебра над алгебраически за-
мкнутым полем характеристики p 6= 2, обладающая нетривиальным δ-дифференцированием или δ-
супердифференцированием. Тогда p > 2, δ = 1
2 и либо J = V1/2(Z, D), либо J = J(B(m), D).
ДОКАЗАТЕЛЬСТВО. Отметим, что δ-дифференцирования и δ-супердифференцирования простых ко-
нечномерных йордановых супералгебр над алгебраически замкнутым полем характеристики нуль пол-
ностью описаны в работах И. Б. Кайгородова [6, 8]. Там было показано отсутствие нетривиальных δ-
дифференцирований и δ-супердифференцирований для данных супералгебр. Согласно работе И. Б. Кай-
городова и В. Н. Желябина [10], нетривиальные δ-(супер)дифференцирования на простых унитальных
конечномерных йордановых супералгебрах над алгебраически замкнутым полем характеристики p > 2
7
возможны только в случае δ = 1
2 и супералгебры векторного типа J(B(m), D). Согласно работе Е. Зель-
манова [17] простые неунитальные конечномерные йордановы супералгебры над алгебраически замкну-
тым полем характеристики p > 2 исчерпываются супералгебрами вида K3, V1/2(Z, D), супералгеброй
K9 в характеристике 3. Для супералгебры Капланского K3 результаты об отсутствии нетривиальных δ-
дифференцирований и δ-супердифференцирований справедливы из работ И. Б. Кайгородова [6, 8] вне
зависимости от характеристики поля. Лемма 1 показывает, что супералгебра K9 не обладает нетриви-
альными δ-дифференцированиями и δ-супердифференцированиями. Леммы 4-5 дают примеры нетриви-
альных δ-дифференцирований и δ-супердифференцирований супералгебры V1/2(Z, D), которые возможны
только в случае δ = 1
2 . Учитывая приведенные пояснения, теорема доказана.
§3 δ-дифференцирования и δ-супердифференцирования полупростых конечномерных
йордановых супералгебр.
В
данном
параграфе
мы
δ-
супердифференцирований полупростых конечномерных йордановых супералгебр над алгебраически
замкнутым полем характеристики отличной от 2. В частности, мы покажем отсутствие нетривиаль-
ных δ-дифференцирований и δ-супердифференцирований на полупростых конечномерных йордановых
супералгебрах над алгебраически замкнутым полем характеристики нуль.
δ-дифференцирований
описание
получим
полное
и
Под супералгеброй с неделителем нуля, мы будем подразумевать супералгебру A для которой суще-
ствует элемент a, что из равенства ax = 0 для некоторого элемента x ∈ A, следует x = 0. Примером таких
супералгебр могут являться унитальные супералгебры, супералгебры K9, V1/2(Z, D) и др.
ЛЕММА 7. Пусть φ (cid:22) δ-дифференцирование или δ-супердифференцирование супералгебры A, причем
A = A1 ⊕ A2, где Ai (cid:22) супералгебры с однородным неделителем нуля. Тогда φ(Ai) ⊆ Ai.
ДОКАЗАТЕЛЬСТВО. Пусть ei (cid:22) однородный неделитель нуля супералгебры Ai, а xi произвольный
элемент Ai. Тогда
0 = φ(eixj) = δ(φ(ei)xj + (−1)p(ei)p(φ)eiφ(xj )) = δφ(ei)jxj + (−1)p(ei)p(φ)δeiφ(xj )i.
Откуда получаем eiφ(xj )i = 0 и φ(xj )i = 0, что влечет φ(Aj ) ⊆ Aj. Лемма доказана.
Пусть Ji неунитальная супералгебра, удовлетворяющая условиям:
i) алгебра (Ji)0 имеет единицу ei;
ii) для любого элемента z ∈ (Ji)1 верно 2ei · z = z.
Примерами таких супералгебр являются супералгебры K9, K3 и V1/2(Z, D).
Тогда справедлива
ЛЕММА 8. Супералгебра J = J1 ⊕ . . . ⊕ Jn + F · 1 не имеет нетривиальных δ-дифференцирований и
δ-супердифференцирований.
ДОКАЗАТЕЛЬСТВО. Пусть φ (cid:22) нетривиальное δ-дифференцирование или δ-супердифференциро-
вание. Согласно [6], унитальные супералгебры могут иметь нетривиальные δ-дифференцирования и δ-
супердифференцирования только при δ = 1
2 . Рассмотрим случай n = 1, то есть J = J1 + F · 1 и e (cid:22)
единица алгебры (J1)0, тогда φ(1) = α · 1 + j0 + j1, где ji ∈ (J1)i. Отметим, что
2αe + 2j0 +
1
2
j1 = 2(φ(1) · e) · e = φ(e) · e + e · φ(e) = 2φ(e) = 2φ(1) · e = 2αe + 2j0 + j1,
следовательно φ(1) = α · 1 + j0. Заметим, что для x ∈ (J1)1 верно
αx + j0 · x = φ(x) = 2φ(e · x) = φ(e) · x + e · φ(x) =
1
2
αx + j0 · x +
1
2
αx +
1
2
j0 · x,
откуда j0 = 0. Что влечет тривиальность φ в случае n = 1. В общем случае, рассматривая подсупералгебры
Ii = Ji +F ·1, получим тривиальность ограничения φ на этих подсупералгебрах, что влечет тривиальность
φ на всей супералгебре J. Лемма доказана.
Согласно работе Е. Земальнова [17], если J полупростая конечномерная йорданова супералгебра над
i=1(Ji1 ⊕. . .⊕Jiri +Ki ·1)⊕J1 ⊕. . .⊕Jt,
алгебраически замкнутым полем F характеристики p 6= 2, то J = Ls
8
где J1, . . . , Jt (cid:22) простые йордановы супералгебры, K1, . . . , Ks (cid:22) расширения поля F и Ji1, . . . , Jiri (cid:22) про-
стые неунитальные йордановы супералгебры над полем Ki. Воспользовавнись этим результатом, лем-
мами 7-8 и работами [6, 8], где было показано отсутствие нетривиальных δ-дифференцирований и δ-
супердифференцирований у простых конечномерных йордановых супералгебр над алгебраически замкну-
тым полем характеристики нуль, мы можем получить теорему
ТЕОРЕМА 9. Полупростая конечномерная йорданова супералгебра над алгебраически за-
мкнутым полем характеристики нуль не имеет нетривиальных δ-дифференцирований и δ-
супердифференцирований.
Пользуясь результатами работы Е. Зельманова [17], леммами 7-8 и теоремой 6 мы можем получить
следующую теорему
ТЕОРЕМА 10. Пусть полупростая конечномерная йорданова супералгебра J над алгебраиче-
ски замкнутым полем характеристики p > 2 имеет нетривиальное δ-дифференцирование или δ-
супердифференцирование. Тогда δ = 1
2 и J = J ∗ ⊕ J∗, где либо J∗ = J(B(m), D), либо J∗ = V1/2(Z, D).
В заключение, автор выражает благодарность В. Н. Желябину за внимание к работе и предложенное
доказательство леммы 2.
9
Список литературы
[1] Hopkins N. C., Generalizes Derivations of Nonassociative Algebras, Nova J. Math. Game Theory Algebra
5 (1996), №3, 215 -- 224.
[2] Филиппов В. Т., Об алгебрах Ли, удовлетворяющих тождеству 5-ой степени, Алгебра и логика, 34
(1995), №6, 681 -- 705.
[3] Филиппов В. Т., О δ-дифференцированиях алгебр Ли, Сиб. матем. ж. 39 (1998), №6, 1409 -- 1422.
[4] Филиппов В. Т., О δ-дифференцированиях первичных алгебр Ли, Сиб. матем. ж. 40 (1999), №1, 201 --
213.
[5] Филиппов В. Т., О δ-дифференцированиях первичных альтернативных и мальцевских алгебр, Ал-
гебра и Логика 39 (2000), №5, 618 -- 625.
[6] Кайгородов И. Б., О δ-дифференцированиях простых конечномерных йордановых супералгебр, Алгеб-
ра и логика 46 (2007), №5, 585 -- 605. [ http://arxiv.org/abs/1010.2419 ]
[7] Кайгородов И. Б., О δ-дифференцированиях классических супералгебр Ли, Сиб. матем. ж. 50 (2009),
№3, 547 -- 565. [ http://arxiv.org/abs/1010.2807 ]
[8] Кайгородов И. Б., О δ-супердифференцированиях простых конечномерных йордановых и лиевых су-
пералгебр, Алгебра и логика 49 (2010), №2, 195 -- 215. [ http://arxiv.org/abs/1010.2423 ]
[9] Кайгородов И. Б., Об обобщенном дубле Кантора, Вестник Самарского гос. университета 78 (2010),
№4, 42 -- 50. [ http://arxiv.org/abs/1101.5212 ]
[10] Желябин В. Н., Кайгородов И. Б., О δ-супердифференцированиях простых супералгебр йордановой
скобки, Алгебра и анализ 23 (2011), №4, 40 -- 58.
[11] Zusmanovich P., On δ-derivations of Lie algebras and superalgebras, J. of Algebra 324 (2010), №12, 3470 --
3486. [ http://arxiv.org/abs/0907.2034 ]
[12] Кайгородов И. Б., О δ-дифференцированиях алгебр и супералгебр, "Проблемы теоретической и при-
кладной математики Труды 41-ой Всероссийской молодежной школы-конференции, 29 -- 33. URL:
http://home.imm.uran.ru/digas/School-2010_A5.pdf
[13] Кантор И. Л., Йордановы и лиевы супералгебры, определяемые алгеброй Пуассона, Алгебра и анализ,
Томск, изд-во ТГУ (1989), 89 -- 126.
[14] Kac V. G., Сlassification of simple Z-graded Lie superalgebras and simple Jordan superalgebras, Comm.
Algebra 13 (1977), 1375 -- 1400.
[15] Racine M. L., Zelmanov E. I., Simple Jordan superalgebras with semisimple even part, J. of Algebra 270
(2003), №2, 374 -- 444.
[16] Martines C., Zelmanov E., Simple finite-dimensional Jordan superalgebras of prime characteristic, J. of
Algebra 236 (2001), №2, 575 -- 629.
[17] Zelmanov E., Semisimple finite dimensional Jordan superalgebras, (English) [A] Fong, Yuen (ed.) et al., Lie
algebras, rings and related topics. Papers of the 2nd Tainan-Moscow international algebra workshop '97,
Tainan, Taiwan, January 11 -- 17, 1997. Hong Kong: Springer (2000), 227 -- 243.
10
|
1611.01340 | 1 | 1611 | 2016-11-04T11:50:04 | Derivation Hom-Lie 2-algebras and non-abelian extensions of Hom-Lie algebras | [
"math.RA"
] | In this paper, we introduce the notion of a derivation of a Hom-Lie algebra and construct the corresponding strict Hom-Lie 2-algebra, which is called the derivation Hom-Lie 2-algebra. As applications, we study non-abelian extensions of Hom-Lie algebras. We show that iso- morphism classes of diagonal non-abelian extensions of a Hom-Lie algebra g by a Hom-Lie algebra h are in one-to-one correspondence with homotopy classes of morphisms from g to the derivation Hom-Lie 2-algebra DER(h). | math.RA | math |
Derivation Hom-Lie 2-algebras and non-abelian extensions of
Hom-Lie algebras ∗
Lina Song and Rong Tang
Department of Mathematics, Jilin University,
Changchun 130012, Jilin, China
Email: [email protected], [email protected]
Abstract
In this paper, we introduce the notion of a derivation of a Hom-Lie algebra and construct
the corresponding strict Hom-Lie 2-algebra, which is called the derivation Hom-Lie 2-algebra.
As applications, we study non-abelian extensions of Hom-Lie algebras. We show that iso-
morphism classes of diagonal non-abelian extensions of a Hom-Lie algebra g by a Hom-Lie
algebra h are in one-to-one correspondence with homotopy classes of morphisms from g to the
derivation Hom-Lie 2-algebra DER(h).
1
Introduction
The notion of a Hom-Lie algebra was introduced by Hartwig, Larsson, and Silvestrov in [9] as part
of a study of deformations of the Witt and the Virasoro algebras. In a Hom-Lie algebra, the Jacobi
identity is twisted by a linear map, called the Hom-Jacobi identity. Some q-deformations of the
Witt and the Virasoro algebras have the structure of a Hom-Lie algebra [9, 11]. Because of close
relation to discrete and deformed vector fields and differential calculus [9, 13, 14], more people
pay special attention to this algebraic structure. In particular, representations and deformations
of Hom-Lie algebras were studied in [3, 17, 19]; Extensions of Hom-Lie algebras were studied in
[6, 13]. Geometric generalization of Hom-Lie algebras was given in [16]; Quantization of Hom-Lie
algebras was studied in [22] and integration of Hom-Lie algebras was studied in [15].
Now higher categorical structures are very important due to connections with string theory [5].
One way to provide higher categorical structures is by categorifying existing mathematical concepts.
A Lie 2-algebra is the categorification of a Lie algebra [4]. The Jacobi identity in a Lie 2-algebra
is replaced by a natural transformation, called the Jacobiator, which also satisfies a coherence law
of its own. A very important example is the derivation Lie 2-algebra DER(k) = (k
−→ Der(k), l2)
associated to a Lie algebra k, where l2 is given by
ad
l2(D1, D2) = D1 ◦ D2 − D2 ◦ D1,
l2(D, u) = −l2(u, D) = D(u),
∀D, D1, D2 ∈ Der(k), u ∈ k.
0Keyword: Hom-Lie algebras, derivations, non-abelian extensions, Hom-Lie 2-algebras
0MSC: 17B40, 17B70, 18D35.
∗Research supported by NSFC (11471139) and NSF of Jilin Province (20140520054JH).
1
In [18], a non-abelian extension of a Lie algebra g by a Lie algebra h is explained by a Lie 2-algebra
morphism from g to the derivation Lie 2-algebra DER(h). The notion of a Hom-Lie 2-algebra,
which is the categorification of a Hom-Lie algebra, was given in [20].
In this paper, we give the notion of a derivation of a Hom-Lie algebra (g, [·, ·]g, φg). The set
of derivations Der(g) is a Hom-Lie subalgebra of the Hom-Lie algebra (gl(g), [·, ·]φg , Adφg) which
was given in [21]. We prove that the set of outer derivation Out(g) is exactly the first cohomology
group H1(g, ad) of the Hom-Lie algebra g with the coefficient in the adjoint representation. Then
we construct the derivation Hom-Lie 2-algebra DER(g). As applications, we study non-abelian
extensions of Hom-Lie algebras. Parallel to the case of Lie algebras, we characterize a diagonal
non-abelian extension of a Hom-Lie algebra g by a Hom-Lie algebra h using a Hom-Lie 2-algebra
morphism from g to the derivation Hom-Lie 2-algebra DER(h).
The paper is organized as follows.
In Section 2, we recall some basic notions of Hom-Lie
algebras, representations of Hom-Lie algebras and their cohomologies, Hom-Lie 2-algebras and
morphisms between Hom-Lie 2-algebras. In Section 3, we give the notion of a derivation of a Hom-
Lie algebra and construct the associated derivation Hom-Lie 2-algebra. In Section 4, we study
diagonal non-abelian extensions of Hom-Lie algebras and use Hom-Lie 2-algebra morphisms to
characterize them. In Section 5, we give a discussion on some other ways to characterize diagonal
non-abelian extensions of Hom-Lie algebras.
2 Preliminaries
In this section, we recall some basic notions of Hom-Lie algebras, representations of Hom-Lie
algebras and their cohomologies, Hom-Lie 2-algebras and morphisms between Hom-Lie 2-algebras.
Moreover, we give the notion of a 2-morphism between two morphisms of Hom-Lie 2-algebras.
2.1 Hom-Lie algebras and their representations
Definition 2.1.
(1) A (multiplicative) Hom-Lie algebra is a triple (g, [·, ·]g, φg) consisting of a
vector space g, a skew-symmetric bilinear map (bracket) [·, ·]g : ∧2g −→ g and a linear map
φg : g → g preserving the bracket, such that the following Hom-Jacobi identity with respect
to φg is satisfied:
[φg(x), [y, z]g]g + [φg(y), [z, x]g]g + [φg(z), [x, y]g]g = 0.
(1)
(2) A Hom-Lie algebra is called a regular Hom-Lie algebra if φg is an algebra automorphism.
In the sequel, we always assume that φg is an algebra automorphism.
Definition 2.2. A morphism of Hom-Lie algebras f : (g, [·, ·]g, φg) → (h, [·, ·]h, φh) is a linear
map f : g → h such that
f [x, y]g = [f (x), f (y)]h, ∀x, y ∈ g,
f ◦ φg = φh ◦ f.
(2)
(3)
Definition 2.3. A representation of a Hom-Lie algebra(g, [·, ·]g, φg) on a vector space V with
respect to β ∈ gl(V ) is a linear map ρ : g → gl(V ) such that for all x, y ∈ g, the following
equalities are satisfied:
ρ(φg(x)) ◦ β = β ◦ ρ(x),
ρ([x, y]g) ◦ β = ρ(φg(x)) ◦ ρ(y) − ρ(φg(y)) ◦ ρ(x).
(4)
(5)
2
We denote a representation by (ρ, V, β). For all x ∈ g, we define adx : g → g by
adx(y) = [x, y]g,
∀y ∈ g.
(6)
Then ad : g −→ gl(g) is a representation of the Hom-Lie algebra (g, [·, ·]g, φg) on g with respect to
φg, which is called the adjoint representation.
Let (ρ, V, β) be a representation. The cohomology of the Hom-Lie algebra (g, [·, ·]g, φg) with
the coefficient in V is the cohomology of the cochain complex C k(g, V ) = Hom(∧kg, V ) with the
coboundary operator d : C k(g, V ) → C k+1(g, V ) defined by
(df )(x1, · · · , xk+1) = Pk+1
g xi, φ−2
+Pi<j(−1)i+j βf ([φ−2
i=1 (−1)i+1ρ(xi)(f (φ−1
g xj ]g, φ−1
g x1, · · · ,
g x1, · · · ,
\
φ−1
g xi, · · · ,
\
φ−1
g xi, · · · , φ−1
\
φ−1
g xj , · · · , φ−1
g xk+1))
g xk+1).
The fact that d2 = 0 is proved in [7]. Denote by Z k(g; ρ) and Bk(g; ρ) the sets of k-cocycles and
k-coboundaries respectively. We define the k-th cohomolgy group Hk(g; ρ) to be Z k(g; ρ)/Bk(g; ρ).
Let (ad, g, φg) be the adjoint representation. For any 0-hom-cochain x ∈ g = C0(g, g), we have
(dx)(y) = [y, x]g, for all y ∈ g. Thus, we have dx = 0 if and only if x ∈ Cen(g), where Cen(g)
denotes the center of g. Therefore, we have
H0(g, ad) = Z 0(g, ad) = Cen(g).
We will analyze the first cohomology group after we introduce the notion of a derivation of a
Hom-Lie algebra.
Let V be a vector space, and β ∈ GL(V ). Define a skew-symmetric bilinear bracket operation
[·, ·]β : ∧2gl(V ) −→ gl(V ) by
[A, B]β = β ◦ A ◦ β−1 ◦ B ◦ β−1 − β ◦ B ◦ β−1 ◦ A ◦ β−1, ∀A, B ∈ gl(V ).
Denote by Adβ : gl(V ) → gl(V ) the adjoint action on gl(V ), i.e.
Adβ(A) = β ◦ A ◦ β−1.
(7)
(8)
Proposition 2.4. ([21, Proposition 4.1]) With the above notations, (gl(V ), [·, ·]β , Adβ) is a regular
Hom-Lie algebra.
This Hom-Lie algebra plays an important role in the representation theory of Hom-Lie algebras.
See [21] for more details.
2.2 Hom-Lie 2-algebras
Definition 2.5. ([20, Definition 3.6]) A Hom-Lie 2-algebra V consists of the following data:
• a complex of vector spaces V1
d−→ V0,
• bilinear maps l2 : Vi × Vj −→ Vi+j , where 0 ≤ i + j ≤ 1,
• two linear transformations φ0 ∈ gl(V0), φ1 ∈ gl(V1) satisfying φ0 ◦ d = d ◦ φ1,
• a skew-symmetric trilinear map l3 : V0 × V0 × V0 −→ V1 satisfying l3 ◦ φ⊗3
0 = φ1 ◦ l3, such
that for all w, x, y, z ∈ V0 and m, n ∈ V1, the following equalities are satisfied:
3
(a) l2(x, y) = −l2(y, x),
l2(x, m) = −l2(m, x),
(b) dl2(x, m) = l2(x, dm),
l2(dm, n) = l2(m, dn),
(c) φ0(l2(x, y)) = l2(φ0(x), φ0(y)), φ1(l2(x, m)) = l2(φ0(x), φ1(m)),
(d) dl3(x, y, z) = l2(φ0(x), l2(y, z)) + l2(φ0(y), l2(z, x)) + l2(φ0(z), l2(x, y)),
(e) l3(x, y, dm) = l2(φ0(x), l2(y, m)) + l2(φ0(y), l2(m, x)) + l2(φ1(m), l2(x, y)),
(f) l3(l2(w, x), φ0(y), φ0(z)) + l2(l3(w, x, z), φ2
+ l3(l2(w, z), φ0(x), φ0(y)) = l2(l3(w, x, y), φ2
+l3(φ0(w), l2(x, y), φ0(z))+l2(φ2
0(y)) + l3(φ0(w), l2(x, z), φ0(y))
0(z)) + l3(l2(w, y), φ0(x), φ0(z))
0(w), l3(x, y, z))+l2(l3(w, y, z), φ2
0(x))+l3(φ0(w), l2(y, z), φ0(x)).
We usually denote a Hom-Lie 2-algebra by (V1, V0, d, l2, l3, φ0, φ1) or simply by V. A Hom-Lie
2-algebra is called strict if l3 = 0. If φ0 and φ1 are identity maps, we obtain the notion of a Lie
2-algebra [4].
Definition 2.6. Let V and V ′ be Hom-Lie 2-algebras, a morphism f from V to V ′ consists of:
• a chain map f : V → V ′, which consists of linear maps f0 : V0 → V ′
0 and f1 : V1 → V ′
1
satisfying
and
f0 ◦ d = d′ ◦ f1,
f0 ◦ φ0 = φ′
0 ◦ f0, f1 ◦ φ1 = φ′
1 ◦ f1.
• a skew-symmetric bilinear map f2 : V0 × V0 → V ′
such that for all x, y, z ∈ V0 and m, n ∈ V1, we have
1 satisfying f2(φ0(x), φ0(y)) = φ′
1f2(x, y)
• df2(x, y) = f0(l2(x, y)) − l′
2(f0(x), f0(y)),
• f2(x, dm) = f1(l2(x, m)) − l′
2(f0(x), f1(m)),
• l′
2(f0(φ0(x)), f2(y, z)) + l′
= f2(l2(x, y), φ0(z)) + f2(l2(y, z), φ0(x)) + f2(l2(z, x), φ0(y)) + f1(l3(x, y, z)).
2(f0(φ0(z)), f2(x, y)) + l′
2(f0(φ0(y)), f2(z, x)) + l′
3(f0(x), f0(y), f0(z))
Definition 2.7.
(1) Let V and V ′ be Hom-Lie 2-algebras and let f, g : V → V ′ be morphisms
from V to V ′. A 2-morphism τ from f to g is a chain homotopy τ : f =⇒ g such that the
following equations hold for all x, y ∈ V0 :
φ′
1(τ (x)) = τ (φ0(x)),
f2(x, y) − g2(x, y) = l′
2(τ (x), f0(y)) + l′
2(g0(x), τ (y)) − τ (l2(x, y)).
(2) Two morphisms bewteen Hom-Lie 2-algebras are called homotopic if there is a 2-morphism
between them.
4
3 Derivations of a Hom-Lie algebra and the associated deriva-
tion Hom-Lie 2-algebra
In this section, we introduce the notion of a derivation of a Hom-Lie 2-algebra and give its coho-
mological characterization. Then we construct a strict Hom-Lie 2-algebra using derivations of a
Hom-Lie algebra, which is called the derivation Hom-Lie 2-algebra.
Definition 3.1. A linear map D : g → g is called a derivation of a Hom-Lie algebra (g, [·, ·]g, φg)
if
D[x, y]g = [φg(x), (Ad
φ−1
g
D)(y)]g + [(Ad
φ−1
g
D)(x), φg(y)]g, ∀x, y ∈ g.
(9)
Denote by Der(g) the set of derivations of the Hom-Lie algebra (g, [·, ·]g, φg).
Remark 3.2. The notion of an α-derivation was given in [19] under an extra condition D ◦ φg =
φg ◦ D. Under this condition, Adφg D = D, and it follows that our derivation is the same as
the one given in [19]. So our definition of a derivation of a Hom-Lie algebra is more general
than the one given in [19]. We will show that the set of derivations is a Hom-Lie subalgebra of
the Hom-Lie algebra (gl(g), [·, ·]φg , Adφg) given in Proposition 2.4 and the first cohomology group
of a Hom-Lie algebra with the coefficient in the adjoint representation is exactly the set of outer
derivations. These facts show that our definition of a derivation of a Hom-Lie algebra is more
meaningful. More importantly, we can construct a Hom-Lie 2-algebra using the derivations, by
which we characterize nonabelian extensions of Hom-Lie algebras.
Lemma 3.3. For all D ∈ Der(g), we have AdφgD ∈ Der(g).
Proof. By (8) and (9), we have
(AdφgD)[x, y]g = (φg ◦ D ◦ φ−1
g x, φ−1
g y]g
g D(y)]g + φg[φ−1
g )[x, y]g = (φg ◦ D)[φ−1
g D(x), y]g
(AdφgD)(cid:1)(y)]g + [(cid:0)Ad
φ−1
g
= φg[x, φ−1
= [φgx,(cid:0)Ad
φ−1
g
(AdφgD)(cid:1)(x), φgy]g.
Thus, Adφg D ∈ Der(g).
Consider the Hom-Lie bracket [·, ·]φg on gl(g) defined by (7), we have
Lemma 3.4. For all D, D′ ∈ Der(g), we have [D, D′]φg ∈ Der(g).
Proof. For all x, y ∈ g, by (7) and (9) we have
[D, D′]φg ([x, y]g) = (φg ◦ D ◦ φ−1
= (φg ◦ D ◦ φ−1
= (φg ◦ D ◦ φ−1
g ◦ D′ ◦ φ−1
g ◦ D′)([φ−1
g )([x, φ−1
g − φg ◦ D′ ◦ φ−1
g x, φ−1
g ◦ D ◦ φ−1
g y]g) − (φg ◦ D′ ◦ φ−1
g )([x, y]g)
g ◦ D)([φ−1
g (D′(y))]g) + (φg ◦ D ◦ φ−1
g )([φ−1
g x, φ−1
g (D′(x)), y]g)
g y]g)
−(φg ◦ D′ ◦ φ−1
g )([x, φ−1
g (D(y))]g) − (φg ◦ D′ ◦ φ−1
g )([φ−1
g (D(x)), y]g)
5
g (D′(x)), φ−1
g y]g)
g (D(x)), φ−1
g (y)]g)
= (φg ◦ D)([φ−1
g x, φ−2
g (D′(y))]g) + (φg ◦ D)([φ−2
−(φg ◦ D′)([φ−1
g (D(φ−1
g (D′(x)), φ−1
= φg[x, φ−1
+φg[φ−1
−φg[x, φ−1
−φg[φ−1
g (D′(φ−1
= [φg(x), (D ◦ φ−1
g (D′(φ−1
g (x), φ−2
g (D′(y))))]g + φg[φ−1
g (D(y))]g) − (φg ◦ D′)([φ−2
g (D(x)), φ−1
g (D′(y))]g
g (D(y))]g + φg[φ−1
g (D(y))))]g − φg[φ−1
g (D(x)))), y]g − φg[φ−1
g ◦ D′ − D′ ◦ φ−1
g ◦ D)(y)]g
g (D(φ−1
g (D′(x)), φ−1
g (D(x)), φ−1
g (D′(x)))), y]g
g (D(y))]g
g (D′(y))]g
+[(D ◦ φ−1
g ◦ D′ − D′ ◦ φ−1
g ◦ D)(x), φg(y)]g
= [φg(x), (Ad−1
φg
[D, D′]φg)(y)]g + [(Ad−1
φg
[D, D′]φg)(x), φg(y)]g.
Therefore, we have [D, D′]φg ∈ Der(g).
Proposition 3.5. With the above notations, (Der(g), [·, ·]g, Adφg) is a Hom-Lie algebra, which is
a sub-algebra of the Hom-Lie algebra (gl(g), [·, ·]φg , Adφg) given in Proposition 2.4.
Proof. By Lemma 3.3 and 3.4, (Der(g), [·, ·]g, Adφg) is a Hom-Lie subalgebra of the Hom-Lie algebra
(gl(g), [·, ·]φg , Adφg).
For all x ∈ g, adx is a derivation of the Hom-Lie algebra (g, [·, ·]g, φg), which we call an inner
derivation. This follows from
adx[y, z]g = [x, [y, z]g]g = [φg(φ−1
g x), [y, z]g]g
= [[φ−1
= [φ−1
= [(Ad
g x, y]g, φg(z)]g + [φg(y), [φ−1
g [x, φg(y)]g, φg(z)]g + [φg(y), φ−1
adx)(y), φg(z)]g + [φg(y), (Ad
g x, z]g]g
φ−1
g
g [x, φg(z)]g]g
adx)(z)]g .
φ−1
g
Denote by Inn(g) the set of inner derivations of the Hom-Lie algebra (g, [·, ·]g, φg), i.e.
Lemma 3.6. Let (g, [·, ·]g, φg) be a Hom-Lie algebra. For all x ∈ g and D ∈ Der(g), we have
Inn(g) = {adx x ∈ g}.
(10)
Adφg
adx = adφg(x),
[D, adx]φg = adD(x).
Therefore, Inn(g) is an ideal of the Hom-Lie algebra (Der(g), [·, ·]g, Adφg ).
Proof. For all x, y ∈ g, by (8), we have
(Adφg
adx)(y) = (φg ◦ adx ◦ φ−1
g )(y) = (φg ◦ adx)(φ−1
g (y)) = φg[x, φ−1
g (y))]g = [φg(x), y]g
= adφg(x)(y).
g ◦ adx ◦ φ−1
g [x, φ−1
g )(y) − (φg ◦ adx ◦ φ−1
g ◦ D ◦ φ−1
g (y)))
g )(D(φ−1
g )(y)
g (y)]g) − (φg ◦ adx ◦ φ−1
g (y)]g) − φg[x, φ−1
D)(φ−2
g (y))]g + [(Ad
g D(φ−1
g (y))]g
D)(φ−1
φ−1
g
g (x)), φ−1
g (y)]g(cid:17) − [φg(x), D(φ−1
g (y))]g
By (9), we have
[D, adx]φg(y) = (φg ◦ D ◦ φ−1
= (φg ◦ D)(φ−1
= φg(D[φ−1
= φg(cid:16)[x, (Ad
= [D(x), y]g
= adD(x)(y).
φ−1
g
g x, φ−2
6
Therefore, we have Adφg
adx = adφg(x) and [D, adx]φg = adD(x). The proof is finished.
Denote by Out(g) the set of out derivations of the Hom-Lie algebra (g, [·, ·]g, φg), i.e.
Out(g) = Der(g)/Inn(g).
(11)
Proposition 3.7. Let (g, [·, ·]g, φg) be a Hom-Lie algebra. We have
H1(g, ad) = Out(g).
Proof. For any f ∈ C1(g, g), we have
(df )(x1, x2) = [x1, f (φ−1
g x2)]g − [x2, f (φ−1
g x1)]g − φgf ([φ−2
g x1, φ−2
g x2]g).
Therefore, the set of 1-cocycles Z 1(g, ad) is given by
f ([φ−2
g x1, φ−2
g x2]g) = [φ−1
g x1, φ−1
g f (φ−1
g x2)]g + [φ−1
g f (φ−1
g x1), φ−1
= [φg(φ−2
g x1), (Ad
f )(φ−2
g x2)]g + [(Ad
φ−1
g
φ−1
g
Thus, we have Z 1(g, ad) = Der(g).
Furthermore, the set of 1-coboundaries B1(g, ad) is given by
dx = [·, x]g = ad−x,
g x2]g
f )(φ−2
g x1), φg(φ−2
g x2)]g.
for some x ∈ g. Therefore, we have B1(g, ad) = Inn(g), which implies that H1(g, ad) = Out(g).
At the end of this section, we construct a strict Hom-Lie 2-algebra using derivations of a Hom-
Lie algebra. We call this strict Hom-Lie 2-algebra the derivation Hom-Lie 2-algebra. It plays
an important role in our later study of nonabelian extensions of Hom-Lie algebras.
Let (h, [·, ·]h, φh) be a Hom-Lie algebra. Consider the complex h
ad
−→ Der(h), where h is of
degree 1 and Der(h) is of degree 0. Define l2 by
l2(D1, D2) = [D1, D2]φh , ∀D1, D2 ∈ Der(h),
l2(D, u) = −l2(u, D) = D(u), ∀D ∈ Der(h), u ∈ h.
(12)
(13)
Theorem 3.8. With the above notations, (h, Der(h), d = ad, l2, φ0 = Adφh , φ1 = φh) is a strict
Hom-Lie 2-algebra.
We denote this strict Hom-Lie 2-algebra by DER(h).
Proof. First we show that the condition φ0 ◦ d = d ◦ φ1 holds. It follows from
φ0(d(u))(v) = (Adφh(adu))(v) = (φh ◦ adu ◦ φ−1
h )(v) = [φh(u), v]h,
and
d(φ1(u))(v) = d(φh(u))(v) = adφh(u)(v) = [φh(u), v]h.
Condition (a) in Definition 2.5 holds obviously.
For all D ∈ Der(h) and u ∈ h, we have
d(l2(D, u))(v) = d(D(u))(v) = adD(u)(v) = [D(u), v]h.
7
By Lemma 3.6, we have
l2(D, du)(v) = [D, adu]φh(v) = [D(u), v]h = d(l2(D, u))(v).
(14)
For all u, v ∈ h, we have
l2(du, v) = l2(adu, v) = [u, v]h = −[v, u]h = −l2(adv, u) = l2(u, adv) = l2(u, dv).
(15)
By (14) and (15), we deduce that Condition (b) holds.
Conditions (c) and (d) follow from the fact that (Der(h), [·, ·]h, φ0 = Adφh) is a Hom-Lie algebra.
Condition (e) follows from
l2(φ0(D1), l2(D2, u)) + l2(φ0(D2), l2(u, D1)) + l2(φ1(u), l2(D1, D2))
= l2(Adφh(D1), D2(u)) + l2(Adφh(D2), −D1(u)) + l2(φh(u), [D1, D2]φh )
= (φh ◦ D1 ◦ φ−1
h )(D2(u)) − (φh ◦ D2 ◦ φ−1
h )(D1(u))
+l2(φh(u), φh ◦ D1 ◦ φ−1
h ◦ D2 ◦ φ−1
h − φh ◦ D2 ◦ φ−1
h ◦ D1 ◦ φ−1
h ) = 0.
Since l3 = 0, Condition (f) holds naturally. The proof is finished.
4 Non-abelian extensions of Hom-Lie algebras
In this section, we characterize diagonal non-abelian extensions of Hom-Lie algebras by Hom-Lie 2-
algebra morphisms from a Hom-Lie algebra (viewed as a trival Hom-Lie 2-algebra) to the derivation
Hom-Lie 2-algebra constructed in the last section.
Definition 4.1. A non-abelian extension of a Hom-Lie algebra (g, [·, ·]g, φg) by a Hom-Lie algebra
(h, [·, ·]h, φh) is a commutative diagram with rows being short exact sequence of Hom-Lie algebra
morphisms:
0 −−−−→ h
ι−−−−→ g
−−−−→ g −−−−→ 0
φhy
0 −−−−→ h
φgy
ι−−−−→ g
where (g, [·, ·]g, φg) is a Hom-Lie algebra.
p
p
φgy
−−−−→ g −−−−→ 0,
We can regard h as a subspace of g and φgh = φh. Thus, h is an invariant subspace of φg. We
say that an extension is diagonal if g has an invariant subspace X of φg such that h ⊕ X = g. In
general, g does not always have an invariant subspace X of φg such that h ⊕ X = g. For example,
the matrix representation of φg is a Jordan block. We only study diagonal non-abelian extensions
in the sequel.
Definition 4.2. Two extensions of g by h, (g1, [·, ·]g1 , φg1) and (g2, [·, ·]g2 , φg2 ), are said to be
isomorphic if there exists a Hom-Lie algebra morphism θ : g2 → g1 such that we have the following
commutative diagram:
0 −−−−→ h
ι2−−−−→ g2
p2−−−−→ g −−−−→ 0
(cid:13)(cid:13)(cid:13)
0 −−−−→ h
θy
ι1−−−−→ g1
(cid:13)(cid:13)(cid:13)
p1−−−−→ g −−−−→ 0.
8
Proposition 4.3. Let (g1, [·, ·]g1 , φg1 ) and (g2, [·, ·]g2 , φg2 ) be two isomorphic extensions of a Hom-
Lie algebra (g, [·, ·]g, φg) by a Hom-Lie algebra (h, [·, ·]h, φh). Then g1 is a diagonal non-abelian
extension if and only if g2 is also a diagonal non-abelian extension.
Proof. Let g2 be a diagonal non-abelian extension. Then it has an invariant subspace X of φg2
such that h ⊕ X = g2. Since θ is a Hom-Lie algebra morphism, for all u ∈ X, we have
φg1 (θu) = θ(φg2 u).
Therefore, θ(X) is an invariant subspace of φg1 . Moreover, we have h ⊕ θ(X) = g1. Thus, g1 is a
diagonal non-abelian extension.
A section of an extension g of g by h is a linear map s : g −→ g such that p ◦ i = Id.
Lemma 4.4. A Hom-Lie algebra (g, [·, ·]g, φg) is a diagonal non-abelian extension of a Hom-Lie
algebra (g, [·, ·]g, φg) by a Hom-Lie algebra (h, [·, ·]h, φh) if and only if there is a section s : g → g
such that
φg ◦ s = s ◦ φg.
(16)
A section s : g → g of g is called diagonal if (16) is satisfied.
Proof. Let g be a diagonal non-abelian extension of g by h. Then g has an invariant subspace X
of φg such that h ⊕ X = g. By the exactness, we have pX : X → g is a linear isomorphism. Thus
we have a section s = p−1
X : g → g, such that s(g) = X and p ◦ s = Id. Since p is a Hom-Lie
algebra morphism, we have
Thus, we have φg(s(x)) − s(φg(x)) ∈ h. Moreover, since X is an invariant subspace of φg, we have
p(cid:0)φg(s(x)) − s(φg(x))(cid:1) = 0,
∀x ∈ g.
which implies that
φg(s(x)) − s(φg(x)) ∈ X,
φg(s(x)) − s(φg(x)) ∈ h ∩ X = {0}.
Therefore, φg(s(x)) = sφg(x).
Conversely, let s : g −→ g be a section such that
φg(s(x)) = s(φg(x)),
∀x ∈ g.
Then s(g) is an invariant subspace of φg. By the exactness, we have h ⊕ s(g) = g. Hence, the
extension is diagonal.
Let (g, [·, ·]g, φg) be a diagonal extension of a Hom-Lie algebra (g, [·, ·]g, φg) by a Hom-Lie
algebra (h, [·, ·]h, φh) and s : g → g a diagonal section. Define linear maps ω : g ∧ g → h and
ρ : g → gl(h) respectively by
ω(x, y) = [s(x), s(y)]g − s[x, y]g, ∀x, y ∈ g,
ρx(u) = [s(x), u]g, ∀x ∈ g, u ∈ h.
(17)
(18)
9
Obviously, g is isomorphic to g ⊕ h as vector spaces. Transfer the Hom-Lie algebra structure on g
to that on g ⊕ h, we obtain a Hom-Lie algebra (g ⊕ h, [·, ·](ρ,ω), φ), where [·, ·](ρ,ω) and φ are given
by
[x + u, y + v](ρ,ω) = [x, y]g + ω(x, y) + ρx(v) − ρy(u) + [u, v]h,
φ(x + u) = φg(x) + φh(u).
(19)
(20)
The following proposition gives the conditions on ρ and ω such that (g⊕ h, [·, ·](ρ,ω), φ) is a Hom-Lie
algebra.
Proposition 4.5. With the above notations, (g ⊕ h, [·, ·](ρ,ω), φ) is a Hom-Lie algebra if and only
if ρ and ω satisfy the following equalities:
(21)
(22)
(23)
(24)
(25)
φh ◦ ω = ω ◦ φ⊗2
g ,
φh ◦ ρx = ρφg(x) ◦ φh,
ρx([u, v]h) = [φh(u), (Ad
[ρx, ρy]φh − ρ[x,y]g = adω(x,y),
φ−1
h
ρx)(v)]h + [(Ad
φ−1
h
ρx)(u), φh(v)]h,
ρφg(x)(ω(y, z)) + c.p. = ω([x, y]g, φg(z)) + c.p..
Proof. If (g ⊕ h, [·, ·](ρ,ω), φ) is a Hom-Lie algebra. By
φ([x + u, y + v](ρ,ω)) = [φ(x + u), φ(y + v)](ρ,ω),
we deduce that (21) and (22) holds. By (22) and
[[u, v](ρ,ω), φ(x)](ρ,ω) + [[v, x](ρ,ω), φ(u)](ρ,ω) + [[x, u](ρ,ω), φ(v)](ρ,ω) = 0,
we deduce that (23) holds. By (22) and
[[u, x](ρ,ω), φ(y)](ρ,ω) + [[x, y](ρ,ω), φ(u)](ρ,ω) + [[y, u](ρ,ω), φ(x)](ρ,ω) = 0,
we deduce that (24) holds. By
[[x, y](ρ,ω), φ(z)](ρ,ω) + [[y, z](ρ,ω), φ(x)](ρ,ω) + [[z, x](ρ,ω), φ(y)](ρ,ω) = 0,
we deduce that (25) holds.
Conversely, if (21)-(25) hold, it is straightforward to see that (g ⊕ h, [·, ·](ω,ρ), φ) is a Hom-Lie
algebra. The proof is finished.
Remark 4.6. Note that (23) implies that ρx ∈ Der(h).
Obviously, we have
Corollary 4.7. If ρ and ω satisfy Eqs. (21)-(25), the Hom-Lie algebra (g ⊕ h, [·, ·](ρ,ω), φ) is a
diagonal non-abelian extension of g by h.
For any diagonal nonabelian extensions, by choosing a diagonal section, it is isomorphic to
(g ⊕ h, [·, ·](ρ,ω), φ). Therefore, we only consider diagonal nonabelian extensions of the form (g ⊕
h, [·, ·](ρ,ω), φ) in the sequel.
10
Theorem 4.8. Let g = (g ⊕ h, [·, ·](ρ,ω), φ) be a diagonal non-abelian extension of g by h. Then,
(ρ, ω) give rise to a Hom-Lie 2-algebra morphism f = (f0, f1, f2) from g to the derivation Hom-Lie
2-algebra DER(h) given in Theorem 3.8, where f0, f1, f2 are given by
f0(x) = ρx,
f1 = 0,
f2(x, y) = −ω(x, y),
∀x, y ∈ g.
Conversely, for any morphism f = (f0, f1, f2) from g to DER(h), there is a diagonal non-abelian
extension (g ⊕ h, [·, ·](ρ,ω), φ) of g by h, where ρ and ω are given by
ρx = f0(x), ω(x, y) = −f2(x, y),
∀x, y ∈ g.
Proof. Let g = (g ⊕ h, [·, ·](ρ,ω), φ) be a diagonal non-abelian extension of g by h. By (23),
ρx ∈ Der(h) for all x ∈ g. By (21) and (22), we have
f0(φg(x)) = ρφg(x) = φh ◦ ρx ◦ φ−1
f2(φg(x), φg(y)) = −ω(φg(x), φg(y)) = φh(−ω(x, y)) = φh(f2(x, y)).
h = Adφh(ρx) = Adφh(f0(x)),
By (24), we have
df2(x, y) = ad−ω(x,y) = ρ[x,y]g − [ρx, ρx]φh = f0([x, y]g) − [f0(x), f0(y)]φh .
By (25), we have
l′
2(f0(φg(x)), f2(y, z)) + c.p. = ρφg(x)(−ω(x, y)) + c.p.
= −ω([x, y]g, φg(z)) + c.p.
= f2([x, y]g, φg(z)) + c.p..
Thus, (f0, f1, f2) is a morphism from g to the derivation Hom-Lie 2-algebra DER(h).
The converse part is easy to be checked. The proof is finished.
Theorem 4.9. The isomorphism classes of diagonal non-abelian extensions of g by h are in one-
to-one correspondence with the homotopy classes of Hom-Lie 2-algebra morphisms from g to the
derivation Hom-Lie 2-algebra DER(h).
Proof. Let (g ⊕ h, [·, ·](ω1,ρ1), φ) and (g ⊕ h, [·, ·](ω2,ρ2), φ) be diagonal non-abelian extensions of g
by h. By Theorem 4.8, we have two Hom-Lie 2-algebra morphisms from g to the strict Hom-Lie
2-algebra DER(h), given by
g = (g0 = ρ1, g1 = 0, g2 = −ω1),
f = (f0 = ρ2, f1 = 0, f2 = −ω2).
Assume that the two extensions are isomorphic. Then there is a Hom-Lie algebra morphism
θ : g ⊕ h → g ⊕ h, such that we have the following commutative diagram:
0 −−−−→ h
ι−−−−→ g ⊕ h(ρ2,ω2)
pr
−−−−→ g −−−−→ 0
(cid:13)(cid:13)(cid:13)
0 −−−−→ h
θy
ι−−−−→ g ⊕ h(ρ1,ω1)
(cid:13)(cid:13)(cid:13)
pr
−−−−→ g −−−−→ 0,
where pr is the projection. Since for all x ∈ g, pr(θ(x)) = x, we can assume that θ(x + u) =
x − ϕθ(x) + u for some linear map ϕθ : g → h. By θ ◦ φ = φ ◦ θ, we get
φh ◦ ϕθ = ϕθ ◦ φg.
(26)
11
By θ[x + u, y + v](ρ2,ω2) = [θ(x + u), θ(y + v)](ρ1,ω1), we get
ρ1
x − ρ2
x = adϕθ(x),
ω1(x, y) − ω2(x, y) = ρ2
x(ϕθ(y)) − ρ2
y(ϕθ(x)) + [ϕθ(x), ϕθ(y)]h − ϕθ([x, y]g).
(27)
(28)
We define the chain homotopy τ : f =⇒ g by τ (x) = ϕθ(x). Since
g0(x) − f0(x) = ρ1
x − ρ2
x = adϕθ(x) = d(τ (x)),
the above definition is well-defined. We go on to prove that τ is a 2-morphism from f to g. For
all x ∈ g, obviously, we have
φh(τ (x)) = φh(ϕθ(x)) = ϕθ(φg(x)) = τ (φg(x)).
By (28), we have
f2(x, y) − g2(x, y) = (cid:0) − ω2(x, y)(cid:1) − (cid:0) − ω1(x, y)(cid:1) = ω1(x, y) − ω2(x, y)
x(ϕθ(y)) − ρ2
x, ϕθ(y)) − l2(ρ2
y(ϕθ(x)) + [ϕθ(x), ϕθ(y)]h − ϕθ([x, y]g)
= ρ2
= l2(ρ2
= l2(f0(x), τ (y)) − l2(f0(y), τ (x)) + l2(d(τ (x)), τ (y)) − τ ([x, y]g)
= l2(f0(x) + d(τ (x)), τ (y)) + l2(τ (x), f0(y)) − τ ([x, y]g)
= l2(g0(x), τ (y)) + l2(τ (x), f0(y)) − τ ([x, y]g).
y, ϕθ(x)) + l2(adτ (x), τ (y)) − τ ([x, y]g)
Thus, τ is a 2-morphism from f to g, which implies that f and g are homotopic Hom-Lie 2-
algebra morphisms. Therefore, isomorphic diagonal non-abelian extensions of g by h correspond
to homotopic Hom-Lie 2-algebra morphisms from g to the derivation Hom-Lie 2-algebra DER(h).
Conversely, let g = (g0, g1 = 0, g2) and f = (f0, f1 = 0, f2) be homotopic morphisms from g
to the derivation Hom-Lie 2-algebra DER(h). By Theorem 4.8, we have two diagonal non-abelian
extensions (g⊕h, [·, ·](g0 ,−g2), φ) and (g⊕h, [·, ·](f0 ,−f2), φ) of g by h. Assume that τ is a 2-morphism
from f to g. We define θ : g ⊕ h → g ⊕ h by
θ(x + u) = x − τ (x) + u.
It is straightforward that θ is a Hom-Lie algebra morphism making the diagram in Definition 4.2
commutative. Therefore, (g ⊕ h, [·, ·](g0 ,−g2), φ) and (g ⊕ h, [·, ·](f0 ,−f2), φ) are isomorphic diagonal
extensions of g by h. Thus, homotopic Hom-Lie 2-algebra morphisms from g to the derivation
Hom-Lie 2-algebra DER(h) correspond to isomorphic diagonal non-abelian extensions of g by h.
This finishes the proof.
5 Outlook
There are several ways to study non-abelian extensions of Lie algebras. The authors classified
non-abelian extensions of Lie algebras using the second non-abelian cohomology group [12]. One
can also use outer derivations to study non-abelian extensions of Lie algebras and characterize the
existence of a non-abelian extension using a cohomological class [1, 2, 10]. Furthermore, one can
also use Maurer-Cartan elements to describe non-abelian extensions [8]. In the sequel, we give
a sketch on how to generalize these ways to study diagonal non-abelian extensions of Hom-Lie
12
algebras. In all, we think that the most nontrivial part is defining the notion of a derivation of
a Hom-Lie algebra and constructing the associated derivation Hom-Lie 2-algebra. Thus, we only
study diagonal non-abelian extensions of Hom-Lie algebras using the derivation Hom-Lie 2-algebra
in detail in this paper and omit details of other ways.
For non-abelian extensions of Hom-Lie algebras, we can define a non-abelian 2-cocycle using
Eqs. (21)-(25). Namely, a pair (ρ, ω) is called a nonabelian 2-cocycle if Eqs. (21)-(25) are satisfied.
We can further define an equivalence relation using Eqs. (26)-(28), i.e. two nonabelian 2-cocycles
(ρ1, ω1) and (ρ2, ω2) are equivalent if Eqs. (26)-(28) are satisfied. Then we obtain the second
nonabelian cohomology group, by which we can classify diagonal non-abelian extensions of Hom-
Lie algebras.
By (22)-(24), we deduce that the pair (ρ, ω) give rise to a Hom-Lie algebra morphism ¯ρ : g −→
Out(h), where ¯ρ = π ◦ ρ and π is the quotient map in the following exact sequence:
0 −→ Inn(h) −→ Der(h) π−→ Out(h) −→ 0.
(29)
Note that Der(h) is a non-abelian extension of Out(h) by Inn(h). We have another abelian extension
of Hom-Lie algebras:
0 −→ Cen(h) −→ h
ad
−→ Inn(h) −→ 0.
(30)
Under the assumption that both (29) and (30) are diagonal non-abelian extensions, for any Hom-Lie
algebra morphism ¯ρ : g −→ Out(h), the existence of a diagonal non-abelian extension that inducing
the morphism ¯ρ can be characterized by a condition on a cohomological class in H3(g, Cen(h)).
For a differential graded Hom-Lie algebra (DGHLA for short) (L, [·, ·]L, φL, d), the set M C(L)
of Maurer-Cartan elements is defined by
M C(L)
△
= {α ∈ L1 dα +
1
2
[α, α]L = 0, φLα = α}.
Let (g, [·, ·]g, φg) be a Hom-Lie algebra. Then (C(g, g) = ⊕kC k(g, g), [·, ·]φg , Adφg) is a graded
Hom-Lie algebra, where [·, ·]φg and Adφg are given by
[P, Q]φg = P ◦ Q − (−1)pqQ ◦ P, ∀P ∈ C p+1(g, g), Q ∈ C q+1(g, g),
(AdφgP )(x1, · · · , xp+1) = φg(P (φ−1
g xp+1)), ∀P ∈ C p+1(g, g),
g x1, · · · , φ−1
in which C k(g, g) = Hom(∧kg, g) and P ◦ Q ∈ C p+q+1(g, g) is defined by
P ◦ Q(x1, · · · , xp+q+1) =
X
(−1)σ
σ∈(q+1,p)−unshuffles
φgP(cid:16)φ−1
g Q(φ−1
g xσ(1), · · · , φ−1
g xσ(q+1)), φ−1
g xσ(q+2), · · · , φ−1
g xσ(p+q+1)(cid:17).
(31)
(32)
(33)
(34)
Furthermore (C(g, g), [·, ·]φg , Adφg, ∂) is a DGHLA, where ∂P = (−1)k+1[µg, P ]φg for all P ∈
C k(g, g) and µg is the Hom-Lie bracket on g, i.e. µg(x, y) = [x, y]g. See [7] for more details.
Let (g, [·, ·]g, φg) and (h, [·, ·]h, φh) be two Hom-Lie algebras. Let g ⊕ h be the Hom-Lie algebra
direct sum of g and h, where the bracket is defined by [x + u, y + v] = [x, y]g + [u, v]h, and the
algebra morphism φ is defined by φ(x + u) = φg(x) + φh(u). Then there is a DGHLA (C(g ⊕
h, g ⊕ h), [·, ·], Adφ, ∂), where ∂P = (−1)k+1[µg + µh, P ]φ for all P ∈ C k(g ⊕ h, g ⊕ h). Define
C k
>(g ⊕ h, h) ⊂ C k(g ⊕ h, h) by
C k(g ⊕ h, h) = C k
>(g ⊕ h, h) ⊕ C k(h, h).
13
Denote by C>(g⊕h, h) = ⊕kC k
>(g⊕h, h). It is straightforward to see that (C>(g⊕h, h), [·, ·]φ, Adφ, ∂)
is a sub-DGHLA of (C(g⊕h, g⊕h), [·, ·]φ, Adφ, ∂). We denote by (L, [·, ·]φ, Adφ, ∂) this sub-DGHLA,
where Lk = C k+1
>(g⊕h, h) = Hom(g, h) is abelian. Then it
is straightforward to check that (ρ, ω) satisfy Eqs. (21)-(25), i.e. (ρ, ω) give rise to a diagonal non-
abelian extension if and only if ρ + ω is a Maurer-Cartan element in the DGHLA (L, [·, ·]φ, Adφ, ∂).
> (g⊕h, h). Obviously, its degree 0 part C0
References
[1] D. Alekseevsky, P. W. Michor and W. Ruppert, Extensions of Lie algebras.
arXiv:math.DG/0005042.
[2] D. Alekseevsky, P. W. Michor and W. Ruppert, Extensions of super Lie algebras. J. Lie
Theory (2005) 15(1): 125-134.
[3] F. Ammar, Z. Ejbehi and A. Makhlouf, Cohomology and Deformations of Hom-algebras. J.
Lie Theory 21 (2011), no. 4, 813-836.
[4] J. Baez and A. S. Crans, Higher-Dimensional Algebra VI: Lie 2-Algebras, Theory and Appl.
Categ. 12 (2004), 492-528.
[5] J. C. Baez, A. E. Hoffnung, and C. L. Rogers, Categorified symplectic geometry and the
classical string, Comm. Math. Phys., 293(3) (2010), 701-725.
[6] J. M. Casas, M. A. Insua and N. Pacheco, On universal central extensions of Hom-Lie algebras.
Hacet. J. Math. Stat. 44 (2015), no. 2, 277-288.
[7] L. Cai and Y. Sheng, Hom-big brackets: theories and applications. SIGMA. 12 (2016), 014.
[8] Y. Frégier, Non-abelian cohomology of extensions of Lie algebras as Deligne groupoid. J.
Algebra. 398 (2014), 243 -- 257.
[9] J. Hartwig, D. Larsson and S. Silvestrov, Deformations of Lie algebras using σ-derivations.
J. Algebra 295 (2006), 314-361.
[10] N. K. Hermann, Non-abelian extensions of topological Lie algebras. Comm. Algebra. 34
(2006), no. 3, 991 -- 1041.
[11] N. Hu, q-Witt algebras, q-Lie algebras, q-holomorph structure and representations. Algebra
Colloq. 6 (1999), no. 1, 51-70.
[12] N. Inassaridze, E. Khmaladze and M. Ladra, Non-abelian cohomology and extensions of Lie
algebras. J. Lie Theory. 18 (2008), no. 2, 413 -- 432.
[13] D. Larsson and S. Silvestrov, Quasi-hom-Lie algebras, central extensions and 2-cocycle-like
identities. J. Algebra 288 (2005), 321-344.
[14] D. Larsson and S. Silvestrov, Quasi-Lie algebras. Contemp. Math. 391 (2005), 241-248.
[15] C. Laurent-Gengoux, A. Makhlouf and J. Teles, Universal algebra of a Hom-Lie algebra and
group-like elements. arXiv:1505.02439.
[16] C. Laurent-Gengoux and J. Teles, Hom-Lie algebroids. J. Geom. Phys. 68 (2013), 69-75.
14
[17] A. Makhlouf and S. Silvestrov, Notes on Formal Deformations of Hom-associative and Hom-
Lie Algebras. Forum Math. 22 (2010), no. 4, 715-739.
[18] Y. Sheng and C. Zhu, Integration of Lie 2-algebras and their morphisms. Lett. Math. Phys.
102 (2012), no. 2, 223 -- 244.
[19] Y. Sheng, Representations of Hom-Lie algebras. Algebr. Represent. Theory. 15 (2012), no. 6,
1081-1098.
[20] Y. Sheng and D. Chen, Hom-Lie 2-algebras. J. Algebra. 376 (2013), 174-195.
[21] Y. Sheng and Z. Xiong, On hom-Lie algebras. Linear Multilinear Algebra. 63 (2015), No. 12,
2379-2395.
[22] D. Yau, Hom-quantum groups: I. Quasi-triangular Hom-bialgebras. J. Phys. A 45 (2012),
no. 6, 065203, 23 pp.
15
|
1312.3756 | 1 | 1312 | 2013-12-13T10:17:08 | Arithmetical pseudo-valuations associated to Dubrovin valuation rings and prime divisors of bounded Krull domains | [
"math.RA"
] | We look at value functions of primes in simple Artinian rings and associate arithmetical pseudo-valuations to Dubrovin valuation rings which, in the Noetherian case, are $\mathbb{Z}$-valued. This allows a divisor theory for bounded Krull domains. | math.RA | math |
Arithmetical pseudo-valuations associated to Dubrovin
valuation rings and prime divisors of bounded Krull
domains
F. Van Oystaeyen and N. Verhulst
Universiteit Antwerpen
Department of mathematics & computer science
Middelheimlaan 1
2020 Antwerpen, Belgium
Abstract
We look at value functions of primes in simple Artinian rings and associate arith-
metical pseudo-valuations to Dubrovin valuation rings which, in the Noetherian case,
are Z-valued. This allows a divisor theory for bounded Krull domains.
1
Introduction
Primes in simple Artinian rings, introduced by Van Oystaeyen and Van Geel, replace
noncommutative valuation rings in skewfield. However, the value functions associated
to primes have only been studied in skewfields and arithmetical pseudo-valuations
have been studied only for orders containing an order with a commutative semigroup
of fractional ideals.
First we consider primes (R, P ) where R is a Goldie ring with a simple Artinian
quotient ring and such that R is invariant under inner A-automorphisms (this is a
property of valuation rings in skewfields). We obtain an arithmetical pseudo-valuation
(we will abbreviate this to apv) v ∶ F(R) → Γ where F(R) is the partially ordered
semigroup of fractional ideals of R and Γ is a totally ordered semigroup. We charac-
terize the rings for which Γ is a group and establish that if Γ is an Archimedean group,
R is a Dubrovin valuation. If ⋂ P n = 0, where P = J(R) is the Jacobson radical of R,
is added to the assumptions on R, then A is a skewfield and R is a valuation ring of A.
Next we look at Noetherian Dubrovin valuation rings and establish an apv v ∶
F(R) → Γ where Γ is totally ordered such that P = {a ∈ A v(RaR) > 0} and R =
{a ∈ A v(RaR) ≥ 0}. The case of non-Noetherian Dubrovin valuations allows P = P 2
and then v cannot exist (since v(P ) = 0 would follow and then v(I) = 0 for all I ∈ F(R)).
The condition ⋂ P n = 0 is enough to characterize Dubrovin valuation rings coming from
an apv in the well-described way. We may call these the valued Dubrovin rings. Our
results establish that Noetherian Dubrovin valuations define a Z-valued apv with all
1
the nice properties and so these are good generalizations of discrete valuation rings in
skewfields. In the final section we establish that bounded Krull domains have a divisor
theory based on those apvs.
2 Stable fractional primes
A prime in a ring R is a subring R′ with a prime ideal P ⊂ R′ such that for any x, y in
R, xR′y ⊆ P implies either x ∈ P or y ∈ P . A prime (P, R′) is called localized if for any
r ∈ R ∖ R′, there exist x, y ∈ R′ ∖ {0} with xry ∈ R′ ∖ P . A prime (P, R′) in a ring R is
called strict fractional if for any r ≠ 0 in R there exist x, y ∈ R′ with 0 ≠ xry ∈ R′, e.g.
any localized prime is strict fractional. We refer the interested reader to [1] for much
more information about primes.
Proposition 2.1. Let (P, R′) be a strict fractional prime in a simple Artinian ring A.
If R is any semisimple Artinian subring of A and R ≠ A, then R′ is not contained in
R.
Proof. Assume R′ ⊆ R and pick a ∈ A ∖ R. Since R is Noetherian, we may choose L
maximal for the property that Lay ⊆ R for some y ∈ R. Since R is semisimple Artinian,
we have R = L ⊕ U where U is a left ideal and uay ∉ R for every u ∈ U (otherwise
(L + Ru)ay ⊆ R entails u ∈ L which is a contradiction). There exist x′, y′ ∈ R′ with
0 ≠ x′uayy′ ∈ R′ ⊆ R. Since L is maximal for the property that Layy′ ⊆ R, it follows
that x′u ∈ L but x′u ∈ U , so x′u = 0 contradicting x′uayy′ ≠ 0.
A ring is said to be a Goldie ring is the set of regular elements satisfies the Ore
condition and S−1R is a semisimple Artinian ring.
Proposition 2.2. If (P, R′) is a strict fractional prime in a simple Artinian ring and
R′ is a Goldie ring, then S−1R′ = A where S is the set of regular elements in R′.
Proof. Let r ∈ S, then r is regular in A because if ru = 0 then u ∈ A ∖ R′ so there
exist x, y ∈ R′ with 0 ≠ xuy ∈ R′. Since S satisfies the Ore condition, there are x′ ∈ R′
and r′ ∈ S with r′x = x′r so r′xuy = x′ruy = 0 hence xuy = 0 which is a contradiction.
Since A is simple Artinian, r−1 ∈ A so S is invertible in A and R′ ↪ A extends to
S−1R′ ↪ A. Since S−1R′ is semisimple Artinian, the preceding proposition implies
that S−1R′ = A.
Remark 2.3. If R is a Goldie prime ring and I is an essential left ideal of R then I
is generated by the regular elements of I. (See [3].)
Consider a strict fractional prime (P, R) of a simple Artinian ring A with R = AP ,
i.e. R = {a ∈ A aP ⊆ P and P a ⊆ P }. We always assume that R is Goldie hence a
prime ring and an order of A (by proposition 2.2).
If P is invariant under innner
automorphisms of A, we say that (P, R) is an invariant prime of A.
Remark 2.4. If (R, P ) is an invariant prime in A, then R is invariant under inner
automorphisms of A.
Proof. Consider u ∈ A∗. For p ∈ P we have uRu−1p = uRu−1puu−1 and u−1pu ∈ P so
Ru−1pu ⊆ P and uRu−1p ⊆ uP u−1 ⊆ P . Hence uRu−1P ⊆ P which implies uRu−1 ⊆ R.
A similar reasoning gives P uRu−1 ⊆ R.
2
By a fractional R-ideal of A we mean an R-bimodule I ⊆ A such that I contains
a regular element of R and for some r ∈ R, rI ⊆ R. Observe that we may choose r
regular since R is an order. We will denote the set of fractional ideals of R by F(R).
Lemma 2.5. With assumptions as before, we have:
(1) If u is regular and uI ⊆ R then Iu ⊆ R. Also: uI ⊆ P if and only if Iu ⊆ P .
(2) If I, J ∈ F(R), then IJ ⊆ P if and only if J I ⊆ P .
(3) If I, J ∈ F(R) then IJ ⊆ R implies J I ⊆ R and vice versa. Moreover, if J ⊈ P
then I ⊆ R and if P ⊉ I ⊆ R then J ⊆ R.
Proof. (1) If uI ⊆ R, then Iu ⊆ u−1Ru = R. The other case is similar.
(2) If IJ ⊆ P then, since (P, R) is a prime, either I or J is in P , say I ⊆ P . Since I
is an ideal it is left essential so it is generated by regular elements. For every regular
element u ∈ I uJ ⊆ P yields J u ⊆ u−1P u = P , hence J I ⊆ P . The case I ⊈ P and J ⊆ P
is similar.
(3) Suppose I, J ∈ F(R) such that IJ ⊆ R.
If I, J ⊆ P there is nothing to prove
since then IJ ⊆ P and J I ⊆ P , so assume J ⊈ P (I ⊈ P is completely similar). From
P IJ ⊆ P we obtain then P I ⊆ P since (P, R) is a prime of A, so I ⊆ AP = R. Again, I
is generated by regular elements since it is left essential and for u ∈ I regular uJ ⊆ R
gives J u ⊆ u−1Ru = R hence J I ⊆ R.
Corollary 2.6. P is the unique maximal ideal of R.
Proof. Consider an ideal I ⊈ P and a regular element u of R which is in I but not in P
(this exists since I is generated by regular elements). Then Ru = RuR ≠ R so u−1 ∉ R.
From Ru−1RuR = R with RuR ⊈ P we obtain Ru−1R ⊆ R which is a contradiction.
Note that we really showed that every regular element in R ∖ P is invertible in R.
Corollary 2.7. If C(P ) = {x ∈ R x mod P regular in R~P } satisfies the Ore condi-
tion then it is invertible in R, i.e. QP (R) = R or R is local and P is the Jacobson
radical of R.
Proof. If C(P ) is an Ore set in the prime Goldie ring R which is also an order in a
simple Artinian ring A, then C(P ) consists of regular elements and since C(P ) ⊆ R ∖ P
it consists of invertible elements of R. Consequently, the localization of R at C(R) is
equal to R. It then follows that P is the Jacobson radical of R.
Proposition 2.8. If ⋂ P n = 0 then C(P ) satisifes the Ore condition.
Proof. We claim that 1 + P consists of units. Indeed, consider 1 + p with p ∈ P and
assume it is not regular, then r(1 + p) = 0 for some 0 ≠ r ∈ R. Then r = −rp yields
r ∈ ⋂ P n hence r = 0 which is a contradiction. If c ∈ C(P ) then c is regular in R~P . We
have that P + Rc is essential in R since it contains P hence it is generated by regular
elements. Since Ru = RuR for regular u, it follows that P + Rc is a two-sideed ideal of
R, hence P + Rc = R and Rc = R i.e. c is invertible. Then there is an u ∈ R with uc = 1
which means uc ∈ 1+ P . If rc = 0 then uru−1uc = 0 which would contradict the fact that
all elements of 1 + P are units. Consequently, C(P ) consists of R-regular elements. For
every r ∈ R and c ∈ C(P ) we have cr = crc−1c = r′c which gives the left Ore condition
and also rc = cc−1rc = cr′ which gives the right Ore condition. Therefore C(P ) is an
Ore set.
3
Corollary 2.9. If ⋂ P n = 0 then C(P ) is invertible in R and R is local with Jacobson
radical P .
Corollary 2.10. R~P is a skewfield.
Proof. If a ∈ R~P is not invertible then it is not regular (cfr. the proof of proposition
2.8), say as = 0. Let a = a mod P and s = s mod P , then sa ∈ P implies (Rs+P )a ⊆ P .
Furthermore, Rs + P is two-sided and it contains P strictly so Rs + P = R. This means
that Ra ⊆ P so a = 0.
Proposition 2.11. Under assumptions as before, the left ideals of R are totally ordered
and every finitely generated left ideal is generated by one regular element.
Proof. Suppose xy ∈ P with either x or y regular in A. We suppose without loss of
generality that x is regular, so it is invertible in A. We find xRy = xRx−1xy = Rxy ⊆ P
so since (R, P ) is prime, x or y must be in P . Consider now x regular (hence invertible)
in A ∖ R. Since x ∉ R, there must be a p ∈ P with xp ∉ P (or px ∉ P in which case we
argue similarly). Then we have Rx−1xp ⊆ P so x−1 ∈ P since it is A-regular. Consider
now a finitely generated left ideal I in R. By [3], it is generated by R-regular elements
so it is generated by a finite number of R-regular elements say I = Ru1 + ⋯+ Run. Since
R is Goldie, every R-regular element is A-regular, so by the preceding statements either
u1u−1
1 must be in R. Suppose the latter (again, in the other case we argue
similarly), then Ru2 = Ru2u−1
1 u1 ⊆ Ru1 which means that Ru1 + Ru2 = Ru1. By
induction we find that every finitely generated left ideal is principal and in fact even
principal for a regular element. This in turn implies that the finitely generated left
ideals are totally ordered by inclusion. Suppose now that I and J are left R-ideals with
J ⊈ I. There must be a regular x ∈ J ∖ I and for every y ∈ I we have either yx−1 ∈ R
which would imply y ∈ xR ⊆ J or xy−1 ∈ R but this is contradictory since it means
x ∈ Ry ⊆ I.
2 or u2u−1
Corollary 2.12. Since every left ideal generated by a regular element is two-sided,
every finitely generated left ideal is two-sided.
3 Arithmetical pseudo-valuations
An arithmetical pseudovaluation on R as before is a function v ∶ F(R) → Γ for some
partially ordered semigroup Γ such that:
(APV1) v(IJ) = v(I) + v(J);
(APV2) v(I + J) ≥ min {v(I), v(J)};
(APV3) v(R) = 0;
(APV4) I ⊆ J implies v(I) ≥ v(J).
For more information about airthmetical pseudo-valuations, we refer to [1] and [2].
Theorem 3.1. If (R, P ) is as before then there is an arithmetical pseudo-valuation v ∶
F(R) → Γ, where Γ is a totally ordered semigroup, such that P = {a ∈ A v(RaR) > 0}
and R = {a ∈ A v(RaR) ≥ 0}.
4
Proof. Observe that for any I, J ∈ F(R) we have IJ ∈ F(R) and I +J ∈ F(R), moreover
for every a ∈ A we have RaR ∈ F(R). Indeed, if a ∈ A then there is a regular u ∈ R
such that ua ∈ R since R is an order, then RuaR = RuRaR ⊆ RaR and as an R-ideal,
RuaR contains a regular element of R. If I and J are in F(R) then IJ contains a
regular element and if uI ⊆ R and vJ ⊆ R for regular u and v then J v ⊆ R so uIJ v ⊆ R
whence vuIJ ⊆ vRv−1 = R with vu regular. For I + J we have vu(I + J) ⊆ R + vuJ with
vuJ = vuv−1vJ ⊆ R since vuv−1 ∈ R.
For any I ∈ F(R) we define v(I) = (P ∶ I) = {a ∈ A aI ⊆ P } and since RaRI ⊆ P if
and only if IRaR ⊆ P this is also equal to v(I) = {a ∈ A Ia ⊆ P }. Note that v(I) ≠ {0}
because uI ⊆ R for some regular u ∈ R, hence 0 ≠ P u ⊆ v(I). We also have v(R) = P .
Put Γ = {v(I) I ∈ F(R)} and define a partial order ≤ by
v(I) ≤ v(J) ⇔ v(I) ⊆ v(J).
Note that if I ⊆ J then v(I) ≥ v(J). We claim that Γ is in fact totally ordered. Indeed,
if I, J ∈ F(R) such that v(I) ⊈ v(J) and v(J) ⊈ v(I) then there is an a ∈ A with aI ⊆ P
but aJ ⊈ P and a b ∈ A with bJ ⊆ P but bI ⊈ P . Since P is prime, aJ bI ⊈ P but
RbIaJ ⊆ RbP J ⊆ RbJ ⊆ P yields RaJ bI ⊆ P which is a contradiction in view of lemma
2.5.
We can define a (not necessarily commutative) operation + on Γ by putting v(I) +
v(J) = v(IJ). The unit for this operation is v(R). We now verify that + is well-defined.
Suppose v(I) = v(I ′) and v(J) = v(J ′) and consider x ∈ v(IJ), then RxRIJ ⊆ P so
RxRI ⊆ v(J) = v(J ′) or RxRIJ ′ ⊆ P . By the same lemma as before, IJ ′RxR ⊆ P
follows hence J ′RxR ⊆ v(I) = v(I ′) i.e.
I ′J ′RxR ⊆ P which implies x ∈ v(I ′J ′)
and consequently v(IJ) ⊆ v(I ′J ′). The other inclusion can be obtained by the same
argument if the roles of I, J and I ′, J ′ are interchanged.
We now check that this operation is compatible with ≤. Take some v(I) ≥ v(J)
and consider v(HI) and v(HJ). If q ∈ v(HJ) then qHJ ⊆ P so qH ⊆ v(J) ⊆ v(I)
which implies qHI ⊆ P so q ∈ v(HI). To prove that ≤ is also stable under right
multiplication, we consider q ∈ v(J H). Then qJ H ⊆ P or equivalently J Hq ⊆ P . By
lemma 2.5 HqJ ⊆ P follows so Hq ⊆ v(J) ⊆ v(I) hence IHq ⊆ P i.e. q ∈ v(IH).
If v(I) ≤ v(J) then aI ⊆ P yields a(I + J) ⊆ P since aJ ⊆ P , so v(I + J) ⊇ v(I) =
min {v(I), v(J)}. Together with the preceding, this implies that v is an arithmetical
pseudo-valuation. The only thing left to prove is that
R = {a ∈ A v(RaR) ≥ 0}
and P = {a ∈ A v(RaR) > 0} .
Suppose v(RaR) > 0 = v(R) = P , then there is some x ∈ v(RaR)∖P . Now xRaR ⊆ P
gives a ∈ P , so {a ∈ A v(RaR) > 0} ⊆ P .
If p ∈ P , then v(RpR) ⊇ R ⊋ P , hence
p ∈ {a ∈ A v(RaR) > 0} so P = {a ∈ A v(RaR) > 0}. If a ∈ A is such that v(RaR) = 0
and p ∈ P then v(RaRRpR) = v(RaR) + v(RpR) = v(RpR) > 0 so RaRRpR ⊆ P and
therefore ap ∈ P which implies a ∈ R since R = AP . On the other hand, if r ∈ R then
P RrR ⊆ P . Since RrR is generated by regular elements, it follows that r ∈ ∑ RuiR for
a finite set of regular ui. Consequently, since Γ is totally ordered, v(RrR) = v(RuiR)
where v(RuiR) has the minimal value among these regular elements. If v(RrR) < 0
then v(P ) ≤ v(P RrR) since P RrR ⊆ P and then
v(P ) ≤ v(P RrR) = v(P ) + v(RrR) ≤ v(P )
(1)
5
since v(RrR) < 0. This means that all ≤ in 1 are actually equalities and in fact
v(P ) = v(P ) + v(RuiR) = v(P RuiR) so if aP RuiR ⊆ P then also aP ⊆ P . By choosing
a = u−1
i we find u−1
i ⊆ P and consequently
u−1
∈ AP = R so v(RrR) = v(RuiR) = v(R) = 0. which contradicts v(RrR) < 0.
Consequently R = {a ∈ A v(RaR) ≥ 0}.
i P ⊆ P . In a similar fashion we find P u−1
i
Proposition 3.2. With R, P and A as before, Γ is a group if and only if for any
fractional R-ideal I there is a nonzero y ∈ R with yI ⊆ R but yI ⊈ P .
Proof. If Γ is a group and I ∈ F(R) then for some J ∈ F(R) we have v(I)+v(J) = 0 i.e.
v(IJ) = v(J I) = v(R). Consequently, IJ P ⊆ P ⊇ P IJ so IJ ⊆ AP = R. Since aIJ ⊆ P
iff aR ⊆ P we have IJ ⊈ P . Then we can choose a y ∈ J with Iy ⊆ R but Iy ⊈ P which
implies RIRy ⊆ R and RIRy ⊈ P .
Suppose now that there is some y with yI ⊆ R but yI ⊈ P . For any x ∈ v(RyRI)
we have RxRRyRI ⊆ P which implies RxR ⊆ P and consequently x ∈ v(R). From
RyRI ⊆ R we can deduce v(R) ⊆ v(RyRI) hence v(R) = v(RyRI) which means that
v(RyR) is the inverse of v(I).
Note that the second part of the proof of the preceding theorem guarantees that
every v(I) is also v(RaR) for some a ∈ A.
Lemma 3.3. If ⋂ P n = 0 and Γ is a group then R is a Dubrovin valuation.
Proof. Applying corollary 2.9 gives us that R~P is prime Goldie with invertible regular
elements, i.e. it is a simple Artinian ring. Consider q ∈ A ∖ R. There exists some y ∈ A
with RyRRqR ⊆ R but RyRRqR ⊈ P . Then there exists a z ∈ RyR with zq ∈ R ∖ P
and since RqRRyR ⊆ R but RqRRyR ⊈ P we can use a similar construction to find an
element z′ with qz′ ∈ R ∖ P .
Theorem 3.4. If Γ is a group and ⋂ P n = 0, then R is a valuation ring and A is a
skewfield.
Proof. If a ∈ R ∖ P then P + Ra is essential, two-sided and contains P so it is equal to
R. Then, for some r ∈ R and p ∈ P , we have 1 = p + ra. We have already seen (cfr.
proof of proposition 2.8) that 1 + P consist of units, so ra is a unit hence a is a unit.
If q ∈ A ∖ R there there is some y with yq ∈ R ∖ P , so yq is a unit of R hence q is a unit
of A. Finally, if some p ∈ P were not invertible, then Ap ⊆ P since no element in Ap
is a unit. Then we would have A(RpR) ⊆ P , but this would contain some regular u
which is invertible in A and Au ⊆ P would give a contradiction. This implies that A is
a skewfield and R is an invariant Dubrovin valuation on A, so it must be a valuation
ring.
Corollary 3.5. If Γ is an Archimedean group, then R is a valuation ring.
Proof. In view of the preceding proposition we only have to show that ⋂ P n = 0.
Suppose it is not, then I = ⋂ P n is a nonzero ideal. Pick 0 ≠ b ∈ I, then RbR ⊆ I is a
fractional ideal, hence there exists an ideal J ∈ F(R) with v(J) + v(RbR) = 0. Then
v(P n) + v(RbR) ≤ 0 for any n, so nv(P ) + v(RbR) ≤ 0. However, putting v(RbR) = γ,
there must be some n with nv(p) > −γ which is a contradiction.
6
Proposition 3.6. Let R be any order in a simple Artinian ring A and suppose that
v ∶ F(R) → Γ is an apv. Then:
(1) P = {a ∈ A v(RaR) > 0} defines a prime (P, AP ) with {a ∈ A v(RaR) ≥ 0} ⊆ AP .
(2) if v(I) = {a ∈ A aI ⊆ P } and Γ is a group, then {a ∈ A v(RaR) ≥ 0} = AP .
Proof. (1) For a, b ∈ P we have v(R(a + b)R) ≥ min {v(RaR), v(RbR)} which is strictly
positive, so a + b ∈ P . Clearly, P is an ideal in AP and R ⊆ AP since v(R) = 0 and
If a, a′ ∈ A are such
for all r ∈ R and p ∈ P we have v(RrRRpR) = v(RpR) > 0.
that aAP a′ ⊆ P then aRa′ ⊆ P hence v(RaRa′R) > 0. From v(RaR) + v(Ra′R) > 0
it follows that either v(RaR) > 0 or v(Ra′R) > 0, i.e. either a ∈ P or a′ ∈ P .
If
v(RaR) ≥ 0 for some a ∈ A then for all p ∈ P v(RaRpR)=v(RaR)+v(Rp)¿0
and
v(RpRaR)=v(RpR)+v(RaR)¿0 so a ∈ AP .
(2) Consider a ∈ AP . RaR is invertible in F(R) so there is a J ∈ F(R) with
v(RaR) + v(J) = 0 = v(J) + v(RaR), hence v(J aR) = v(RaJ) = 0.
If v(RaR) < 0
then v(J) > 0 or in other words J ⊆ P . But then a ∈ AP would give RaRJ ⊆ P
which implies v(RaRJ) > 0 which is a contradiction. Therefore v(RaR) ≥ 0 and
AP = {a ∈ A v(RaR) ≥ 0}.
4 Arithmetical pseudo-valuations on Dubrovin
valuations
For primes containing an order with commutative semigroup of fractional ideal, Van
Geel ([2]) introduced artithmetical pseudo-valuations, but this condition is very strong
and reduces the applicability in practice to maximal orders and Dubrovin valuations
in finite dimensional central simple algebras. For Dubrovin valuations on infinite di-
mensional csa the semigroup F(R) need not be commutative. The following facts are
known about Dubrovin valuations on simple Artinian rings (see [1]):
(D1) R is a (left and right) Goldie ring and a prime order of A;
(D2) (P, R) is a localized prime of R;
(D3) P = J(R) and it is the unique maximal ideal of R. Consequently, 1 + P consists
of units;
(D4) F(R) is linearly ordered;
(D5) finitely generated R-submodules of A are cyclic.
Proposition 4.1. For a Noetherian Dubrovin valuation R we have for all I, J ∈ F(R)
that IJ ⊆ P iff J I ⊆ P .
Proof. From IJ ⊆ P it follows that either I ⊆ P or J ⊆ P (by (D2)), assume without
loss of generality I ⊆ P . By (D4), if J I ⊈ P then P ⊊ J I so we have R ⊆ J I ⊆ J P ⊆ J
hence J = RJ ⊆ J IJ ⊆ J P ⊆ J which gives J = J P . Since R is an order, there is some
regular u ∈ R with uJ ⊆ R and since R is Noetherian uJ = ∑ aiR for a finite set of ai's
in uJ. Then also J = ∑ u−1aiR, so J is a finitely generated R-submodule of A. By
Nakayama's lemma J must be zero, which is a contradiction.
7
Corollary 4.2. If R is a Noetherian Dubrovin valuation then
v ∶ F(R) → Γ ∶ I ↦ (P ∶ I) = {a ∈ A aI ⊆ P }
is an arithmetical pseudo-valuation and Γ is a totally ordered group. Furthermore,
P = {a ∈ A v(RaR) > 0} and R = {a ∈ A v(RaR) ≥ 0}.
Proof. Using the preceding proposition instead of 2.5 we can repeat the proof of theo-
rem 3.1. The only thing we need to prove is that Γ is a group, so consider I ∈ F(R).
By a similar argument as in the proof of proposition 4.1 it is finitely generated as a
left R-ideal of A. By (D5) it is cyclic, in fact I = Ru for some regular u and thus
RuR = Ru. Since R is a Dubrovin valuation, there is some a ∈ A with ua ∈ R ∖ P . Then
Rua ⊆ R, Rua ⊈ P and Ia ⊆ R, Ia ⊈ P . We can similarly find a b such that bI ⊆ R but
bI ⊈ P . We can now repeat the last part of the proof of proposition 3.2 to conclude
that Γ is a group.
Remark 4.3. If R is a Dubrovin valuation where
v ∶ F(R) → Γ ∶ I ↦ {a ∈ A aI ⊆ P }
is a non-trivial arithmetical pseudo-valuation with values in a totally ordered group,
then IJ ⊆ P if and only if J I ⊆ P . Indeed, suppose IJ ⊆ P and J I ⊈ P then, as in
proposition 4.1 we find J P = J but then v(P ) = 0 which is impossible.
If R is non-Noetherian, then P = P 2 is possible in which case no nice apv can exist
since otherwise v(P ) = 2v(P ) which would imply v(P ) = 0. If we exclude this slightly
pathological case, a nice apv does exist.
Proposition 4.4. Let R be a Dubrovin valuation with ⋂ P n = 0, then there is an apv
as before.
Proof. If I, J ∈ F(R) with IJ ⊆ P but J I ⊈ P . The same argument as in proposition 4.1
leads to J = J P so J = J P n for any n. There is some regular u ∈ R with uJ ⊆ P hence
uJ = uJ P n ⊆ P n+1. But then uJ = 0 which implies J = 0 and this is a contradiction.
Now we can proceed as in corollary 4.2 to find an apv with values in a semigroup.
The only thing we need to prove is that Γ is a group. Lemma 1.5.4 in [1] says that
P = Rp = pR for some regular p ∈ P . Since P is principal as a left R-ideal, lemma 1.5.6
in the same source gives P P −1 = R = P −1P (here P −1 = {a ∈ A P aP ⊆ P }). Consider
now a fractional R-ideal I. Clearly, (R ∶ I)I ⊆ R. Suppose we also have (R ∶ I)I ⊆ P ,
then P −1(R ∶ I)I ⊆ R hence P −1(R ∶ I) ⊆ (R ∶ I) so P −1(R ∶ I)I ⊆ P . This means
(R ∶ I)I ⊆ P 2 and by repeating this process we find (R ∶ I)I ⊆ P n for any n, but
(R ∶ I)I ⊆ ⋂ P n = 0 which is a contradiction. Therefore, (R ∶ I)I ⊆ R but (R ∶ I)I ⊈ P ,
so there exists an a ∈ (R ∶ I) such that aI ⊆ R but aI ⊈ P .
The following characterizes Noetherian Dubrovin valuation rings within the class of
rank one Dubrovin valuations. The result may be known but we found no reference for
it in the literature. Recall that the rank of a Dubrovin valuation ring is the maximal
length of a chain of Goldie prime ideals in the ring. A rank 1 Dubrovin valuation ring
on a simple Artinian A is a maximal subring of A.
8
Proposition 4.5. For a Dubrovin valuation R on a simple Artinian ring A the fol-
lowing are equivalent:
(1) R is Noetherian.
(2) R has rank 1 and P ≠ P 2.
(3) R has rank 1 and ⋂ P n = 0.
Proof. (1) ⇒ (2) If R is Noetherian then all ideals (and R-ideals of A) are principal,
so P ≠ P 2. Suppose 0 ≠ Q is another prime ideal in P . Let P = Rp, then Q = Ip
for some non-trivial ideal I of R. Q = IP yields I ⊆ Q since Q is prime and P ⊈ Q.
Hence Q = IP ⊆ QP ⊆ Q implies Q = QP which implies Q = 0 by Nakayama's lemma.
(2) ⇒ (3) If P ≠ P 2, then ⋂ P n ≠ P which, by Lemma 1.5.15 in [1], gives ⋂ P n = 0.
(3) ⇒ (1) Since ⋂ P n = 0, P ≠ P 2. Since R is rank 1, R = Ol(I) = Or(I) for any R-ideal
I. By proposition 1.5.8 in [1], it follows that if I is not principal, then II −1 = P and
P = P 2 which s a contradiction.
5 Divisors of Bounded Krull orders
It is an order in a simple Artinian ring
We consider a prime Noetherian ring R.
A = Q(R), the classical ring of fractions.
If R is a maximal order then the set of
divisorial R-ideals of A (cfr. [1]) is a group and since R is Noetherian it is an Abelian
group generated by the maximal divisorial ideals and every maximal divisorial ideal is
a minimal prime ideal.
Recall that an order is an Asano order if every ideal I ≠ 0 of R is invertible and it
is a Dedekind order if it is an hereditary Asano order. If R is an Asano order satisfying
the ascending chain condition on ideals, then F(R) is the Abelian group generated
by maximal ideals and every maximal ideal is a minimal nonzero prime ideal. Any
bounded Noetherian order is a Dedekind order.
A semi-local order R in a simple Artinian A is a Noetherian Asano order if and
only if it is a principal ideal ring. If R is a Dubrovin valuation ring of A then R is a
maximal order if and only if rk(R) = 1 and R is Asano if and only if it is a principal
ideal ring, so a Noetherian Dubrovin valuation ring is a Noetherian maximal order and
an Asano order, i.e. a principal ideal ring.
Proposition 5.1. If R is a Noetherian Dubrovin valuation then the corresponding apv
takes values in Z.
Proof. Since R is a Noetherian Asano order, F(R) is generated by the maximal ideals
of R, but since P is the unique maximal ideal and the value group is necessarily
torsion-free, we have F(R) = Z.
Recall that an order R in a simple Artinian ring A is said to be a Krull order if it is
a maximal order and it is τ -Noetherian (see [1], definition 2.2.2). A Noetherian order
R in a simple Artinian A is a Krull order if and only if it is a maximal order.
Theorem 5.2. Let R be a prime Noetherian ring and an order in A = Q(R). Suppose
every minimal nonzero prime ideal is localizeable, Rp is a Dubrovin valuation ring for
every p ∈ X 1(R) = {minimal nonzero prime ideals of R} and R = ⋂ Rp, then R is a
bounded Krull domain.
9
Proof. Since R is Noetherian, every Rp is Noetherian too and since it is a Dubrovin
valuation it must be an Asano order hence a principal ideal ring. Since every Rp is a
maximal order, so is R. As a Noetherian maximal order, R is a Krull domain and by
theorem 2.2.16 in [1] every regular element is a non-unit in only finitely many of the
Rp's (for p maximal divisorial, i.e. p ∈ X 1(R)). Theorem 2.2.20 in the same source
gives the result.
Remark 5.3. If R is a bounded Krull domain then R = ⋂p∈X 1(R) Rp (cfr. [1] 2.2.18
& 2.2.19). Then every Rp is rank one Dubrovin valuation with P ≠ P 2 (2.2.16 in the
aforementioned source), so by proposition 4.5 it is Noetherian. This gives a correspon-
dence between p ∈ X 1(R) and apvs.
A divisor of a bounded Krull domain R is an element in the free Abelian group
Zp. To any I ∈ F(R) we can associate the divisor div(I) = ∑ vp(Ip)p where
ࣷp∈X 1(R)
Ip = RpI. This definition is justified by the following:
Proposition 5.4. Suppose Rp is Noetherian. If I is an R-ideal of A, then IP is an
RP -ideal of A.
Proof. Let u be regular in R with uI ⊆ R, then RuRI ⊆ R and RuR = Ru′ for some
regular u′ ∈ R. Then Rpu′I is the localization of Ru′I and it is an ideal of Rp. Rpu′ is
the localization of Ru′ so it is also an ideal of Rp, hence Rpu′I = Rpu′RpI = Rpu′Ip is
an ideal of Rp. Now U ′Ip ⊆ Rp, i.e. Ip is an Rp-ideal of A.
Observe that since any regular element is a non-unit in only finitely many localiza-
tions, div(I) contains only finitely many non-zero terms. Moreover, div(I) ≤ div(J)
is and only if vp(I) ≤ vp(J) for all p ∈ X 1(R). By putting I ∗ = ⋂p∈X 1(R) Ip we find
vp(I) = vp(I∗). For bounded Krull domains we can consider div ∶ F(R) → Div(R).
This is a group morphism of Abelian groups and it is order reversing in the sense that
I ⊆ J yields div(I) div(J). Further divisor theory requires a version of the aproxima-
tion theorems. This is work in progress.
References
[1] H. Marubayashi & F. Van Oystaeyen, General theory of primes, Springer LNM
2059, 2012
[2] J. Van Geel, Places and valuations in noncommutative ring theory, Lct. Notes in
Pure & Appl. Math. vol.71, M.Dekker, 1982
[3] J. C. McConnell & J. C. Robson, Noncommutative noetherian rings, Graduate
Studies in Mathematics Vol.30, Am. Math. Soc., ISBN 0-8218-2169-5, Providence
1987
[4] M.Chamaire, Anneaux de Krull non commutatifs, J. of Alg. 72, pp.210-222, 1981
[5] G.Maury & J.Raynaud, Ordres maximaux au sens de K.Asano, LNM vol.808,
Springer 1980
10
|
1210.5438 | 1 | 1210 | 2012-10-19T15:09:00 | Free resolutions of algebras | [
"math.RA",
"math.KT",
"math.RT"
] | Given an algebra A, presented by generators and relations, i.e. as a quotient of a tensor algebra by an ideal, we construct a free algebra resolution of A, i.e. a differential graded algebra which is quasi-isomorphic to A and which is itself a tensor algebra. The construction rests combinatorially on the set of bracketings that arise naturally in the description of a free contractible differential graded algebra with given generators. | math.RA | math |
Free resolutions of algebras
Joe Chuang
Alastair King
18 Oct 2012
Abstract
Given an algebra A, presented by generators and relations, i.e. as
a quotient of a tensor algebra by an ideal, we construct a free alge-
bra resolution of A, i.e. a differential graded algebra which is quasi-
isomorphic to A and which is itself a tensor algebra. The construction
rests combinatorially on the set of bracketings that arise naturally in
the description of a free contractible differential graded algebra with
given generators.
1
Introduction
Let S be any ring. We write ⊗ for ⊗S, and, for any S, S-bimodule V , we
write V n for V ⊗n. Further, we write
T +
S (V ) =Mn≥1
V n,
V (m) =Mn≥m
V n,
(1.1)
for the non-unital tensor algebra of ‘free words’ in V over S and for the ideal
of words of length at least m, respectively.
Our objective in this paper is to construct a free algebra resolution of
an arbitrary (non-unital) S-algebra A presented in terms of generators and
relations, that is,
A = T +
S (V ) /I,
for some S, S-bimodule V and ideal I. As a first step, in Section 2, we work
towards describing the case where I ⊆ V (2), i.e. the generators are ‘minimal’,
1
and I is homogeneous (Theorem 2.2). This case is controlled by the by-
now-familiar A∞ combinatorics of rooted trees or the corresponding set of
bracketings (see, for example, Figure 1). By extending this set of bracketings
in various ways, we find, in Sections 3 and 4, that the essential nature of the
proof becomes more transparent in the general case (Theorem 4.1).
As a typical motivating example, consider a quiver Q and field k. Set
S = kQ0, the semi-simple algebra spanned by the vertex idempotents, and
V = kQ1, the S, S-bimodule spanned by the arrows. Then the augmented
unital algebra S ⊕ T +
S (V ) is the path algebra kQ and S ⊕ A is a ‘quiver
algebra’, i.e.
is presented by a quiver with relations. Such augmentation
allows one to move easily between unital and non-unital S-algebras; we find
it notationally simpler to work in the slightly less familiar non-unital context
in this paper.
Giving a free resolution of A as an S-algebra is, almost tautologically, the
same as giving a ‘locally finite’ A∞-coalgebra over S
K• =Mn≥1
Kn,
where each Kn is an S, S-bimodule and such that A is quasi-isomorphic to
the differential graded (dg) algebra
Cobar K• =(cid:0)T +
S (K•[1]) , d(cid:1).
(1.2)
Here K•[1] denotes the shifted complex, defined by K[1]n = Kn+1. In general,
the quasi-isomorphism should be induced by some S, S-morphism
θ : K1 → A
such that θ(K1) generates A. In our case, since the generators V of A are
explicitly specified, we will suppose that K1 = V and that θ : V → A is the
specifying map. Then the degree zero term in Cobar K• is T +
S (V ) and the
quotient map T +
S (V ) → A will be the quasi-isomorphism.
The cobar construction here is ‘almost tautological’ in that the differential
d on T +
S (K•[1]) may be considered simply as an efficient way of encoding the
A∞-coalgebra structure on K•. More precisely, the A∞-coalgebra structure
consists of S, S-morphisms
∆k : Kn → Km1 ⊗ · · · ⊗ Kmk ,
2
for each k ≥ 1 andPk
Xr+s+t=n
i=1 mi = n − 2 + k, satisfying the conditions
(−1)r+st(cid:0)1⊗r ⊗ ∆s ⊗ 1⊗t(cid:1) ◦ ∆r+1+t = 0
for all n ≥ 1 (cf.
single condition that the endomorphism d of T +
[5, Definition 1.2.1.8]). These may be packaged into the
S (K•[1]), determined by
d = −[1]⊗n ◦ ∆n ◦ [−1] : K•[1] → K•[1]⊗n
(1.3)
and the (graded) Leibniz rule (see (3.2)), satisfies d2 = 0 (cf.
[5, §1.2]).
Note also that, in evaluating [1]⊗n, we use the standard Koszul sign rule; for
example
[1] ⊗ [1] := (−1)m : Km ⊗ Kn → Km ⊗ Kn.
Remark 1.1. The condition that K• is ‘locally finite’ is simply the require-
ment that (1.3) does define a differential on T +
S (K•[1]), which is defined as
an infinite direct sum, and not just on its completion, i.e. the corresponding
direct product.
In other words, for any element x ∈ K•, the coproducts
∆n(x) are non-zero for only finitely many n. It will be clear that all coalge-
bras we will encounter have this property and we will not explicitly mention
it again.
2 First examples
We begin by discussing the sort of construction we are looking for in the case
that I = V (2) and so A = T +
S (V ) /I is just V with trivial multiplication.
As a warm-up, we first observe that the classical candidate for K•, in this
case, is the free coassociative coalgebra generated by V ,
B•(V ) =Mn≥1
V n,
(2.1)
which is what the usual (unaugmented) bar construction yields. The comul-
tiplication ∆ : Bm+n → Bm ⊗ Bn is tautological, i.e. is the natural identifica-
tion τ : V m+n → V m ⊗ V n. This gives the A∞-coproduct ∆2, with all other
coproducts vanishing.
3
V π = V n,
for n =Xk
πk,
As one should expect, the dg algebra A• = Cobar B•(V ) is a free resolu-
tion of A. Indeed, we may explicitly write
A• =Mπ∈Π
V π,
(2.2)
where Π is the set of all finite sequences π = (π1, . . . , πr) of positive integers,
while
which is in homological degree d, i.e. is a summand of Ad, for
d = π =Xk
(πk − 1).
Multiplication in A• corresponds to concatenation of sequences, i.e. is given
by the tautological maps τ : V π ⊗ V η → V πη, while the non-trivial compo-
nents of the differential d : V π → V π′ are ±τ whenever π′ is obtained from π
p)
by splitting some term in the sequence into two. The sign is (−1)(π′
when the p-th term of π is split.
1,...,π′
To see explicitly that A• is quasi-isomorphic to the algebra A, we notice
first that, for each n ≥ 0, there is precisely one sequence of degree 0 summing
to n, namely π = (1, . . . , 1), and this yields A0 = T +
S (V ). On the other hand,
for each n ≥ 2, the part of A• consisting of summands equal to V n is given
by V n ⊗Z C aug
(σn−2)[1] the (shifted and augmented) chain complex of the
(n − 2)-simplex, which is exact, as required.
•
From the point-of-view of this paper, a more natural, but rather bigger,
candidate for K• is the free A∞-coalgebra generated by V ,
B∞
• (V ) = Mβ∈B[2]
V β.
(2.3)
Here B[2] is the set of all closed non-degenerate bracketings (or equivalently
those that correspond to ‘rooted trees’) and the bracketed tensor product
V β = V n if β has n inputs and is in homological degree d if β has d − 1 pairs
of brackets. Note that “closed” means that the whole expression is enclosed
in an outer bracket, while “non-degenerate” means that each inner pair of
brackets encloses at least two inputs. For example, one summand of B∞
4 (V )
would be
V β = [[V ⊗ V ] ⊗ [V ⊗ V ⊗ V ]] = V 5
4
for β = [[ • •][ • • • ]] or [[2][3]], a 3-fold bracketing of 5 inputs.
The differential ∆1 has non-zero components
(−1)m−1τ : V β → V β( bm),
where β(bm) is obtained from β by removing the mth internal left bracket
[ (counted from the left) together with its matching right bracket ]. Pre-
cisely one higher coproduct ∆k, for some k ≥ 2, is defined on each V β and
corresponds to removing the outer bracket and writing what is inside as a
concatenation of k closed bracketings, up to a sign. For example, there is a
component
∆4 : V [[2]2[3]] → V [2] ⊗ V ⊗ V ⊗ V [3].
For the general component
∆k : V [β1...βk] → V β1 ⊗ . . . ⊗ V βk
the sign we choose (see Remark 3.1 for an explanation) is
− (−1)Pk
i=1(k−i)di,
(2.4)
where di is the homological degree of V βi in B∞
• (V ).
By convention, the only 0-fold closed bracketing is 1 and so B∞
1 (V ) = V .
On the other hand, the 1-fold closed bracketings are [k], for each k ≥ 2,
and hence B∞
2 (V ) = V (2). As a further example, the possible bracketings of
4 inputs are 3, 2 or 1-fold, as listed in Figure 1 together with their corre-
sponding rooted trees. These are well-known to correspond to the 0,1 and 2
dimensional cells of a pentagon. Furthermore, the restriction of the differen-
tial on B∞
• corresponds to the coboundary map on the cochain complex of
the pentagon (with some care needed over signs).
More generally, writing B∞
n,k(V ) for the sum over all (n − 1)-fold bracket-
ings of k inputs, we have for k ≥ 2,
B∞
n,k(V ) ∼= V k ⊗Z C k−n(Sk−2),
(2.5)
where C •(Sm) is the cochain complex of the mth associahedron (or Stasheff
polytope, introduced in [7]). In addition, the differential on C •(Sk−2) induces
the differential on B∞
•,k(V ). Since the associahedra are all contractible, this
means that the homology of B∞
•,k(V ) is just V k in degree k. In other words,
5
[[[2]1]1] [1[[2]1]] [[1[2]]1] [1[1[2]]] [[2][2]]
[[3]1]
[[2]2]
[1[2]1]
[2[2]]
[1[3]]
[4]
Figure 1: Bracketings with four inputs.
the kernel of d restricted to B∞
map
k,k(V ) is isomorphic to Bk(V ) and the induced
η : B•(V ) → B∞
• (V ),
(2.6)
is a quasi-isomorphism of A∞ coalgebras (essentially because the A∞ operad
resolves the associative operad).
Using Corollary 6.3 from the Appendix, this is actually sufficient to prove
the following result, but we will also prove it directly as a special case of our
main result Theorem 4.1 in Section 4.
Theorem 2.1. Cobar B∞
• (V ) ≃ V .
More generally, suppose A is a graded algebra, generated by V in degree
1, so that A = T +
S (V ) /I, where
is a homogeneous ideal in V (2). Then, generalising (2.3), we define
Rn
I =Mn≥2
• (V, I) = Mβ∈B[2]
B∞
(V, I)β,
(2.7)
where (V, I)β is obtained from V β by replacing every occurrence of an inner
bracketed [V n] by [Rn]. The A∞ coalgebra structure on B∞
• (V, I) is defined,
as for B∞
• (V ), by tautological maps (with the appropriate sign), except for
6
the component of the differential corresponding to the removal of an inner
bracket. Such a component is induced by one of the two maps
V a ⊗ [Rn] ⊗ V b → Rn+a+b,
V a ⊗ [Rn] ⊗ V b → V n+a+b,
depending on whether the domain is enclosed by matching brackets [..] or
non-matching ones, e.g. ]..]. The first map exists because I is an ideal, while
the second is the composite of the first with the inclusion Rn+a+b ⊆ V n+a+b.
This inclusion also gives the degree 2 component ∆k : [Rk] → V k.
Another special case of Theorem 4.1 is then the following.
Theorem 2.2. Cobar B∞
• (V, I) ≃ A.
Example 2.3. We look more closely at the case when S is a separable k-
algebra, so that ⊗ is exact and hence B∞
• (V ). Now Theorem 2.2
implies (by Remark 4.3 and Proposition 5.1, or directly from Proposition 6.4)
that the degree k part of TorA
n (S, S) is isomorphic to the homology of d
at B∞
n,k(V, I), so this degree k part vanishes for k < n, since there are no
bracketings in this case. Since Extn
n (S, S), S), this also
starts in degree n (cf. [1, Lemma 2.1.2]).
A(S, S) ∼= HomS(TorA
• (V, I) ⊆ B∞
Observe further that all inner brackets in the bracketings that contribute
n,n(V ) are [2] (as for example in the top row of Figure 1). Hence, under
• (V ) of (2.6), the kernel of d restricted to
n,n(V, I) is identified with the familiar Koszul term Bn(V, R2) ⊆ Bn(V ) (cf.
to B∞
the embedding η : B•(V ) → B∞
B∞
[1, §2.6]), defined by
Bn(V, R2) = \p+q=n−2
V p ⊗ R2 ⊗ V q.
Thus, as is also well-known from Koszul theory (cf. [1, Thm 2.6.1]), the de-
gree n part of Torn(S, S) is isomorphic to Bn(V, R2). Note also that B•(V, R2)
is a subcoalgebra of B•(V ).
But now, if Extn
A(S, S), and hence TorA
n (S, S), is concentrated in degree
n, which is one characterisation of A being Koszul (cf. [1, Prop 2.1.3]), then
necessarily I = (R2) and we also deduce that B•(V, R2) ≃ B∞
• (V, I) and
hence, by Corollary 6.3, that Cobar B•(V, R2) ≃ A, which is the algebra
incarnation of the Koszul resolution for A.
7
3 A free contractible dg algebra
We now develop the machinery that will enable us to generalise the construc-
tions behind Theorems 2.1 and 2.2. In the process, we are naturally led to
consider bracketings of a slightly more general form than in Section 2.
For any ring S and any S, S-bimodule V , we can construct a free con-
tractible dg S-algebra
F• = F•(V ) =Mn≥0
Fn(V )
that is ‘freely’ generated by V in degree 0 and a contracting homotopy
h : F• → F• of degree 1, satisfying dh + hd = id and h2 = 0. The differential
d, of degree −1, is then determined recursively by the two conditions
d(hv) = v − h(dv),
d(v · w) = dv · w + (−1)deg vv · dw,
(3.1)
(3.2)
starting from dv = 0 for v ∈ V . Notice that (with these signs) we deduce
inductively that d2 = 0, by observing that
d2hv = dv − dhdv = dv − (dv − hd2v) = hd2v
and
d2(v · w) = d(cid:0)dv · w + (−1)deg vv · dw(cid:1)
= d2v · w + (−1)deg dvdv · dw + (−1)deg vdv · dw + v · d2w
= d2v · w + v · d2w,
since deg(dv) = (deg v) − 1.
We may describe F• more explicitly using an extended notion of bracketed
tensor products, similar to Section 2,
F•(V ) = Mβ∈B(1)
V β,
(3.3)
where, if β is a d-fold bracketing of n inputs, then V β is V n and is in homo-
logical degree d, i.e. is a summand of Fd. Note that this degree is different
from (2.3), because we are not describing an A∞-coalgebra here, but its cobar
construction directly.
8
We define the set B(1) of bracketings recursively as follows, noting that
each such bracketing is, in the first instance, a word in the symbols “•”
(representing an input), “[” and “]”, with the later two balanced in the
usual way of brackets:
(i) B(1) contains •,
(ii) if β ∈ B(1) and β 6= [α], for some α ∈ B(1), then [β] ∈ B(1),
(iii) if α, β ∈ B(1), then their concatenation αβ ∈ B(1)
Note that, in contrast to B[2] in (2.3), bracketings in B(1) may be open, i.e.
without an outer bracket, and degenerate, i.e. with only one input in an
inner bracket.
As before, we abbreviate a sequence of n uninterrupted •’s by “n”. Thus,
for example, [ • •][ • •[ • ]] becomes [2][2[1]]. In particular, there are
just two bracketings 1 and [1] with one input and eight with two inputs
2, [2], [1]1, 1[1], [[1]1], [1[1]], [1][1], [[1][1]].
Now the product in F• has non-zero components consisting of the tauto-
logical maps
τ : V α ⊗ V β → V αβ,
while the contracting homotopy h : F• → F• has non-zero components given
by the tautological maps
τ : V β → V [β],
for each β 6= [α],
(3.4)
that is, τ is the identity map V k → V k, where k is the number of inputs in
β (and in [β]).
On the other hand, the differential d : Fn → Fn−1 has n non-zero compo-
nents
(−1)m−1τ : V β → V β( bm),
for m = 1, . . . , n,
(3.5)
where β(bm) is obtained from β by removing the mth “[” from the left,
together with its matching “]”.
It is straightforward to check that this
does give a differential satisfying (3.1) and (3.2), noting that we count “[”s
from the left, because (3.2) is a ‘left’ Leibniz rule. Thus h is a contracting
homotopy and so F• is a contractible dg algebra.
9
Remark 3.1. Forgetting the differential d and the contracting homotopy h,
F• is the free graded S-algebra generated by
K•(V )[1] = V ⊕ h(F•) = Mβ∈B[1]
V β,
(3.6)
where B[1] ⊆ B(1) is the set of closed bracketings. Note that the convention
that “1” is a closed bracketing is precisely to get the initial summand V here.
In other words, F• = Cobar K•(V ). The differential and coproducts on
the A∞-coalgebra K•(V ) are determined by applying (1.3) in reverse and one
can check that they are given explicitly by the same rules as those of B∞
• (V )
in the paragraphs following (2.3). In particular, this explains the choice of
sign in (2.4).
4 Main theorem
S (V ) and, furthermore, that we have many other copies
S (V ) ‘embedded’ in F• for every pair of inner brackets. To make this
Note that F0(V ) = T +
of T +
explicit, we may introduce a new symbol [ ∗ ] with the meaning
V [∗] =Mm≥1
V [m] = [T +
S (V ) ]
and more generally
V α[∗]β =Mm≥1
V α[m]β = V α ⊗ [T +
S (V ) ] ⊗ V β,
for matching partial bracketings α, β. Thus we can define contracted sets B[∗]
of closed bracketings and B(∗) of open bracketings, in which the innermost
brackets are all [ ∗ ], and so that we can write
F•(V ) = Mβ∈B(∗)
V β,
K•(V ) = Mβ∈B[∗]
V β,
(4.1)
where the degrees of terms in the 2nd equation are shifted compared to (3.6).
Now, for any ideal I ⊆ T +
S (V ) we can define a new dg algebra
F•(V, I) = Mβ∈B(∗)
(V, I)β,
10
(4.2)
where (V, I)β is obtained from V β by replacing [T +
S (V ) ] by [I], for each
occurrence of [ ∗ ] in β. Because I is an ideal, there is a well-defined differ-
ential d : F•(V, I) → F•(V, I) given by the same rule (3.5) as the differential
d : F•(V ) → F•(V ). We can not quite define a contracting homotopy in the
same way, but simply because we have not replaced F0(V ) = T +
S (V ) by I.
S (V ) and hence
Indeed, the image of d : F1(V, I) → F0(V, I) is I ⊆ T +
there is a dg morphism
ρ : F•(V, I) → T +
S (V ) /I,
(4.3)
where the codomain here is just an algebra concentrated in degree 0.
Theorem 4.1. The map ρ in (4.3) is a quasi-isomorphism, that is, the dg
algebra F•(V, I) in (4.2) is a free resolution of the algebra A = T +
S (V ) /I.
Proof. Because ρ is surjective, we just want to show that ker ρ is contractible.
For this, we observe that ker ρ is the dg ideal
F ′
•(V, I) ⊆ F•(V, I)
(4.4)
obtained by also replacing F0(V ) by I. Then, the contracting homotopy h
can be defined on F ′
•(V, I) by the rule (3.4) as it was on F•(V ) and so we
can use it to deduce that F ′
•(V, I) is contractible, as required.
Remark 4.2. In some cases, e.g. when S is a separable k-algebra (so that
⊗ is exact), we note that F•(V, I) is a sub-dg-algebra of F•(V ), and so the
equations d2 = 0, for F•(V, I), and dh + hd = id, for F ′
•(V, I), follow by
restriction from F•(V ). However, in general, they follow rather because the
combinatorial structure of F•(V, I) is identical to that of F•(V ).
Remark 4.3. Just as F•(V ) = Cobar K•(V ), as in Remark 3.1, we have
F•(V, I) = Cobar K•(V, I), where
K•(V, I) = Mβ∈B[∗]
(V, I)β,
(4.5)
In particular, when I = V (2), we have K•(V, I) = B∞
with (V, I)β now in homological degree d when β has d − 1 pairs of brackets.
• (V ) from (2.3),
so that Theorem 4.1 yields Theorem 2.1. Further, for a homogeneous ideal
I ⊆ V (2), we have K•(V, I) = B∞
• (V, I) from (2.7), so that Theorem 4.1
yields Theorem 2.2.
11
5 Homological application
Here we restict to the case when S is a separable k-algebra over a field k
and V is an S, S-bimodule over k, i.e. the S, S-action on V factors through
S ⊗k Sop.
Note that S is a left and right A-module, with A acting trivially, and
we would expect (in good cases) that we could use an A∞-coalgebra K•
with Cobar K• ≃ A to compute the (positive) Tor-groups of S and hence, by
duality, its Ext groups. We observe that the case in hand, with K• = K•(V, I)
as in (4.5), is a good one.
Proposition 5.1. If A = T +
S (V ) /I, then we have an isomorphism
H(K•(V, I)) ∼= TorA
• (S, S)
of graded coalgebras.
Proof. The A∞-coalgebra K•(V ) is cocomplete; e.g. the filtration given by
K•(V )(i) = Mβ∈B[1],β≤i
V β
is admissible (see Appendix for definition). Since S is a separable k-algebra,
K•(V, I) is a A∞-sub-coalgebra of K•(V, I) and is therefore also cocomplete.
Since Theorem 4.1 tells us that F•(V, I) = Cobar K•(V, I) is a dg resolution
of A, we deduce from Proposition 6.4 that H(K•(V, I)) ∼= TorA(S, S) as
graded coalgebras.
It follows that HomS(K•(V, I), S) is an A∞-algebra whose cohomology is
A(S, S), in positive degrees.
Ext•
6 Appendix
Again we restrict to the case where S is a separable k-algebra over a field k.
We extend part of the bar-cobar formalism for dg algebras and cocomplete
dg coalgebras over S to cocomplete A∞-coalgebras, following [5], [3] and [4].
As a change of notation, we write Ω∞C for the cobar construction of an
A∞-coalgebra C. This coincides with the more classical cobar construction
ΩC, in the special case that C is a dg coalgebra. The bar construction of a
dg algebra A is denoted BA.
12
A filtration on an A∞-coalgebra C is a sequence C(0) ⊆ C(1) ⊆ . . . ⊆ C
of graded sub-bimodules such that for all n ≥ 1, the coproduct ∆n : C → C n
is compatible with the (induced) filtrations on C and C n. A filtration is
admissible if C(0) = 0 and C = ∪C(i). We say C is cocomplete if it supports
an admissible filtration; this is equivalent to the usual definition if C is a
dg coalgebra. See [6, §9.3, Remark] for the related notion of conilpotent
(curved) A∞-coalgebra.
A morphism f = (fi) : C ′ → C between A∞-coalgebras is a quasi-
isomorphism if f1 is a quasi-isomorphism of the underlying complexes of C ′
and C, and a weak equivalence if the morphism Ω∞f : Ω∞C ′ → Ω∞C of cobar
constructions is a quasi-isomorphism of dg algebras. A morphism f : C ′ → C
between cocomplete A∞-coalgebras is a filtered quasi-isomorphism if admis-
sible filtrations can be chosen on C ′ and on C so that the maps fi : C ′ → C i
are compatible with the (induced) filtrations on C ′ and C i, and the morphism
Gr(f1) : Gr(C) → Gr(C ′) induced by f1 is a quasi-isomorphism. Note that a
filtered quasi-isomorphism is a quasi-isomorphism.
For any dg algebra A, the classical bar-cobar resolution gives a canon-
ical morphism εA : Ω∞BA → A of dg algebras. Taking A = Ω∞C for
an A∞-coalgebra C, we obtain a corresponding map ǫC : BΩ∞C → C of
In fact, ǫC is a weak equivalence, because εA is a quasi-
A∞-coalgebras.
isomorphism (see, e.g.
[5, Lemme 1.3.2.3(b)]), but we can say more if C is
cocomplete.
Lemma 6.1.
1. If A → A′ is a quasi-isomorphism of dg algebras, then
the induced morphism BA → BA′ is a filtered quasi-isomorphism.
2. If C is a cocomplete A∞-coalgebra, then ǫC : BΩ∞C → C is a filtered
quasi-isomorphism.
Proof. The induced morphism BA → BA′ is a filtered quasi-isomorphism
with respect to the primitive filtrations of BA and BA′. For the second
statement, equip BΩ∞C with the admissible filtration induced by a given
admissible filtration on C. Then ǫC : BΩ∞C → C is filtered, and it suffices
to show that Gri((ǫC)1) : Gri(BΩ∞C) → Gri(C) is a quasi-isomorphism for
all i ≥ 1. Put W = C[1], so that
BΩ∞C = T +(cid:0)T +(W )[−1](cid:1)
13
Since the filtration on C is admissible, Gri(C j) = 0 if j > i. Equip
with the filtration
Gri (BΩ∞C) = Gri Mi1+...+ik≤i
Fl = Gri Mi+1−l≤i1+...+ik≤i
W i1[−1] ⊗ . . . ⊗ W ik[−1]!
W i1[−1] ⊗ . . . ⊗ W ik[−1]! ,
l ≥ 0.
Then Gri((ǫC)1) is a surjective map of complexes with kernel Fi−1, and, for
each 1 ≤ l ≤ i − 1, the subquotient complex
Fl/Fl−1 = Mi1+...+ik=i+1−l
W i1[−1] ⊗ . . . ⊗ W ik[−1]
is acyclic, with a contracting homotopy vanishing on components with i1 = 1
and given by isomorphisms
W i1[−1] ⊗ . . . ⊗ W ik[−1] → W [−1] ⊗ W i1−1[−1] ⊗ W i2[−1] ⊗ . . . ⊗ W ik[−1]
otherwise.
Lemma 6.2. Let f : C ′ → C be a morphism of cocomplete A∞-coalgebras.
1. If f is a filtered quasi-isomorphism, then it is a weak equivalence.
2. If f is a weak equivalence, then it is a quasi-isomorphism.
Proof. For the first part, the proof given by Lefevre [5, Lemma 1.3.2.2] for fil-
tered quasi-isomorphisms of cocomplete dg coalgebras goes through without
change. For the second, consider the commutative diagram
C ′
ǫC ′
f
/ C
ǫC
BΩ∞C ′
BΩ∞f
/ BΩ∞C
By Lemma 6.1, ǫC, ǫC ′ and BΩ∞f are quasi-isomorphisms. Hence f is a
quasi-isomorphism.
14
/
O
O
/
O
O
Corollary 6.3 (Keller [4]). Let f : C ′ → C be a morphism of A∞-coalgebras
which is also compatible with strictly positive gradings on C ′ and C. If f is
a quasi-isomorphism, then it is a weak equivalence.
Proof. The additional positive gradings give rise to admissible filtrations on
C ′ and C in an obvious way. If f is a quasi-isomorphism, then it is automat-
ically a filtered quasi-isomorphism, and then Lemma 6.2 applies.
Proposition 6.4. Suppose that C is a cocomplete A∞-algebra and we have
a quasi-isomorphism Ω∞C → A of dg algebras. Then C is weakly equivalent
to BA. In particular we have an isomorphism of graded coalgebras
H(C) ∼= TorA(S, S).
Proof. By Lemma 6.1, we have filtered quasi-isomorphisms BΩ∞C → BA
and BΩ∞C → C. In particular C and BA are weakly equivalent, and thus
quasi-isomorphic, by Lemma 6.2. Finally, recall that H(BA) ∼= TorA(S, S).
References
[1] A. Beilinson, V. Ginzburg, W. Soergel, Koszul duality patterns in rep-
resentation theory J.A.M.S. 9 (1996) 473–527.
[2] V. Hinich. Homological algebra of homotopy algebras, Comm. Algebra
25 (1997), no. 10, 3291–3323.
[3] B. Keller, Koszul duality and coderived categories (after K. Lef´evre),
notes of a 50-minute talk given at the Frobenius conference in Toru-n
(Poland) in September 2003.
[4] B. Keller, Notes on minimal models, available at the author’s homepage
http://www.math.jussieu.fr/~keller.
[5] K. Lef`evre, Sur le A∞-cat´egories, Th`ese, Universit´e Paris 7, 2003. Avail-
able at http://www.math.jussieu.fr/~lefevre/publ.html.
[6] L. Positselski. Two kinds of derived categories, Koszul duality and
comodule-contramodule correspondence, Mem. Amer. Math. Soc. 212
(2011), no. 996, vi+133 pp.
15
[7] J.D. Stasheff, Homotopy associativity of H-spaces. I, II. Trans. A.M.S.
108 (1963), 275–292; ibid. 293–312.
Coordinates:
JC: City University, London; [email protected]
AK: University of Bath; [email protected]
16
|
1607.06576 | 1 | 1607 | 2016-07-22T07:02:31 | Invariant theory of relatively free right-symmetric and Novikov algebras | [
"math.RA"
] | Algebras with the polynomial identity (x,y,z)=(x,z,y), where (x,y,z)=x(yz)-(xy)z is the associator, are called right-symmetric. Novikov algebras are right-symmetric algebras satisfying additionally the polynomial identity x(yz)=y(xz). We consider the d-generated free right-symmetric algebra F(R) and the free Novikov algebra F(N) over a filed K of characteristic 0. The general linear group GL(K,d) with its canonical action on the d-dimensional vector space with basis the generators of F(R) and F(N) acts on F(R) and F(N) as a group of linear automorphisms. For a subgroup G of GL(K,d) we study the algebras of G-invariants in F(R) and F(N). For a large class of groups G we show that these algebras of invariants are never finitely generated. The same result holds for any subvariety of the variety R of right-symmetric algebras which contains the subvariety L of left-nilpotent of class 3 algebras in R. | math.RA | math |
INVARIANT THEORY OF RELATIVELY FREE
RIGHT-SYMMETRIC AND NOVIKOV ALGEBRAS
VESSELIN DRENSKY
Dedicated to Askar Dzhumadil'daev on the occasion of his 60th birthday
Abstract. Algebras with the polynomial identity (x1, x2, x3) = (x1, x3, x2),
where (x1, x2, x3) = x1(x2x3) − (x1x2)x3 is the associator, are called right-
symmetric. Novikov algebras are right-symmetric algebras satisfying addi-
tionally the polynomial identity x1(x2x3) = x2(x1x3). We consider the free
right-symmetric algebra Fd(R) and the free Novikov algebra Fd(N) freely gen-
erated by Xd = {x1, . . . , xd} over a filed K of characteristic 0. The general
linear group GLd(K) with its canonical action on the d-dimensional vector
space KXd acts on Fd(R) and Fd(N) as a group of linear automorphisms. For
a subgroup G of GLd(K) we study the algebras of G-invariants Fd(R)G and
Fd(N)G. For a large class of groups G we show that the algebras Fd(R)G and
Fd(N)G are never finitely generated. The same result holds for any subvariety
of the variety R of right-symmetric algebras which contains the subvariety L
of left-nilpotent of class 3 algebras in R.
Introduction
In this paper we fix a field K of characteristic 0 and consider nonassociative
K-algebras. An algebra A is called right-symmetric if it satisfies the polynomial
identity
(x1, x2, x3) = (x1, x3, x2),
(1)
where (x1, x2, x3) = x1(x2x3) − (x1x2)x3 is the associator, i.e.,
(a1, a2, a3) = (a1, a3, a2) for all a1, a2, a3 ∈ A.
A right-symmetric algebra is Novikov if it satisfies additionally the polynomial
identity of left-commutativity
x1(x2x3) = x2(x1x3).
(2)
We denote by R and N the varieties of all right-symmetric algebras and all Novikov
algebras, respectively. For details on the history of right-symmetric and Novikov
algebras we refer to the introductions of the paper by Dzhumadil'daev and Lofwall
[20] and the recent preprint by Bokut, Chen, and Zhang [4]. The origins of the
right-symmetric algebras can be traced back till the paper by Cayley [6] in 1857.
Translated in modern language, Cayley mentioned an identity which implies the
2010 Mathematics Subject Classification. 17A36; 17A30; 17A50; 15A72.
Key words and phrases. Free right-symmetric algebras, free Novikov algebras, noncommutative
invariant theory.
Partially supported by Grant I02/18 "Computational and Combinatorial Methods in Algebra
and Applications" of the Bulgarian National Science Fund.
1
2
VESSELIN DRENSKY
right-symmetric identity for the associators and holds for the right-symmetric Witt
algebra in d variables
equipped with the multiplication
∂fi
∂xj(cid:19) ∂
∂xi
.
fi ∈ K[Xd])
W rsym
d
fi
∂
∂xi
=( d
Xi=1
∂xj(cid:19) =(cid:18)fj
∂xi(cid:19) ∗(cid:18)fj
∂
∂
(cid:18)fi
d
Cayley also considered the realization of W rsym
in terms of rooted trees. Later
right-symmetric algebras were studied under different names: Vinberg, Koszul,
Gerstenhaber, and pre-Lie algebras, see the references in [20]. The opposite algebras
of Novikov algebras (satisfying the left-symmetric identity for the associators and
right commutativity) appeared in the paper by Gel'fand and Dorfman [22]. There
the authors gave an algebraic approach to the notion of Hamiltonian operator in
finite-dimensional mechanics and the formal calculus of variations. Independently,
later Novikov algebras were rediscovered by Balinskii and Novikov in the study of
equations of hydrodynamics [2], see also the survey article by Novikov [31]. (Due
to the contributions in [22, 2] and [31] Bokut, Chen, and Zhang [4] suggest to call
these algebras Gel'fand-Dorfman-Novikov algebras. We shall continue to keep the
name Novikov algebras.) An example of a Novikov algebra is the right-symmetric
Witt algebra W rsym
in one variable. In a series of papers, see, e.g., [20, 18, 19]
Dzhumadil'daev, with coauthors or alone, has studied free right-symmetric and
free Novikov algebras, with applications to nonassociative algebras with polynomial
identities.
1
In commutative invariant theory one usually considers the general linear group
GLd(K) with its canonical action on the d-dimensional vector space Vd with basis
{e1, . . . , ed}. This induces an action on the polynomial algebra K[Xd] = K[x1, . . . , xd]
in d variables
g(f (v)) = f (g−1(v)),
g ∈ GLd(K), v ∈ Vd,
where the linear functions xi : Vd → K are defined by
xi(ej) = δij,
i, j = 1, . . . , d,
and δij is the Kronecker symbol. For our noncommutative considerations it is
more convenient to suppress one step and, replacing V with its dual space V ∗, to
assume that GLd(K) acts canonically on the vector space KXd with basis Xd =
{x1, . . . , xd}. Then, identifying the polynomial algebra K[Xd] with the symmetric
algebra of KXd, we extend diagonally this action of GLd(K) on K[Xd]:
g(f (Xd)) = g(f (x1, . . . , xd)) = f (g(x1), . . . , g(xd)),
(3)
g ∈ GLd(K), f (Xd) ∈ K[Xd].
In this way GLd(K) acts as the group of linear
automorphisms of K[Xd]. For a subgroup G of GLd(K) the algebra of G-invariants
is
K[Xd]G = {f ∈ K[Xd] g(f ) = f for all g ∈ G}.
This is a Z-graded vector space and its Hilbert (or Poincar´e) series is the formal
power series
H(K[Xd]G, z) =Xn≥0
dim(K[Xd]G)nzn,
INVARIANT THEORY OF RELATIVELY FREE ALGEBRAS
3
where (K[Xd]G)n is the homogeneous component of degree n in K[Xd]G. The
following are among the main problems related with the description of the algebra
K[Xd]G for different groups or classes of groups G. For details concerning also
computational and algorithmic problems see [9] or [35].
• Is the algebra K[Xd]G finitely generated? This problem was the main moti-
vation for the Hilbert 14th problem in his famous lecture "Mathematische
Probleme" given at the International Congress of Mathematicians held in
It is known that K[Xd]G is finitely generated for fi-
1900 in Paris [25].
nite groups (the theorem of Emmy Noether [30]), for reductive groups (the
Hilbert-Nagata theorem, see e.g., [11]), and for groups close to reductive
(see e.g., Grosshans [23] and Hadziev [24]). The first example of an algebra
of invariants K[Xd]G which is not finitely generated is due to Nagata [29].
• If K[Xd]G is finitely generated, describe it in terms of generators and defin-
In different degree of generality this problem is solved for
ing relations.
classes of groups. For example, the theorem of Emmy Noether [30] gives
that for finite groups the algebra K[Xd]G is generated by invariants of de-
gree ≤ G. Also for finite groups, the Chevalley-Shephard-Todd theorem
[7, 33] states that the algebra K[Xd]G is isomorphic to the polynomial alge-
bra in d variables (i.e., it is generated by a set of d algebraically independent
invariants) if and only if G is generated by pseudo-reflections.
• Calculate the Hilbert series H(K[Xd]G, z). For finite groups the answer is
given by the Molien formula [28]
H(K[Xd]G, z) =
1
G Xg∈G
1
det(1 − gz)
.
The analogue for reductive and close to them groups is the Molien-Weyl
integral formula [37], see also [38].
In noncommutative invariant theory one replaces the polynomial algebras K[Xd]
with other noncommutative or nonassociative algebras still keeping some of the typ-
ical features of polynomial algebras. One such feature is the universal property that
for an arbitrary commutative algebra A every mapping Xd → A is extended to a ho-
momorphism K[Xd] → A. In the noncommutative set-up the class of commutative
algebras is replaced by an arbitrary variety of algebras V and instead on K[Xd]G
one studies the algebra of G-invariants Fd(V)G of the d-generated relatively free
algebra Fd(V) in V, d ≥ 2. For a background see the surveys [21, 13]. Comparing
with commutative invariant theory, when K[Xd]G is finitely generated for all "nice"
groups, the main difference in the noncommutative case is that Fd(V)G is finitely
generated quite rarely. For a survey on invariants of finite groups G acting on rel-
atively free associative algebras see [21, 13] and [26]. For finite groups G 6= h1i and
varieties of Lie algebras Fd(V)G is finitely generated if and only if V is nilpotent,
see [5, 12].
Concerning the Hilbert series of Fd(V)G, for G finite there is an analogue of the
Molien formula, see Formanek [21]. Let
H(Fd(V), z1, . . . , zd) = Xni≥0
dim Fd(V)(n1,...,nd)zn1
1 · · · znd
d
be the Hilbert series of Fd(V) as a multigraded vector space. It is equal to the
generating function of the dimensions of the vector spaces Fd(V)(n1,...,nd) of the
4
VESSELIN DRENSKY
elements in Fd(V) which are homogeneous of degree ni in xi. If ξ1(g), . . . , ξd(g)
are the eigenvalues of g ∈ G, then the Hilbert series of the algebra of invariants
Fd(V)G is
H(Fd(V)G; z) =
H(Fd(V); ξ1(g)z, . . . , ξd(g)z).
1
G Xg∈G
There is also an analogue of the Molien-Weyl formula for the Hilbert series of
Fd(V)G which combines ideas of De Concini, Eisenbud, and Procesi [8] and Almkvist,
Dicks, and Formanek [1]. Evaluating the corresponding multiple integral one uses
the Hilbert series of Fd(V) instead of the Hilbert series of K[Xd]
H(K[Xd], z1, . . . , zd) =
1
1 − zi
.
d
Yi=1
We refer to [3] for other methods for computing the Hilbert series of Fd(V)G when
G is isomorphic to the special linear group SLm(K) or to the group U Tm(K) of
the m × m unitriangular matrices.
In this paper we study invariant theory of relatively free right-symmetric and
Novikov algebras. Let L be the variety of right-symmetric algebras which are left-
nilpotent of class 3, i.e., L is the subvariety of R satisfying the polynomial identity
x1(x2x3) = 0.
1 ]-module. If (A2
d)G is not finitely generated as a K[(Ad)G
(4)
For a large class of subgroups G of GLd(K), G 6= h1i, d > 1, we show that Fd(V)G
is not finitely generated for any variety V containing L. More precisely, let Ad =
K[Xd]+ be the algebra of polynomials without constant term and let (Ad)G
1 =
(KXd)G be the vector space of linear polynomials fixed by G. Clearly, (A2
d)G is a
K[(Ad)G
1 ]-module, then
Fd(V)G is not finitely generated for any V containing L. The class of such groups
G contains all finite groups. It contains also the classical and close to them groups
under some natural restrictions on the embedding into GLd(K). In particular, if
(Ad)G
d)G 6= 0, then Fd(V)G is not finitely generated. Results in the
same spirit hold if we replace the polynomial algebra K[Xd] with the free metabelian
Lie algebra Fd(A2) = Ld/L′′
d, where Ld is the free Lie algebra freely generated by
Xd and A2 is the variety of all metabelian (solvable of class 2) Lie algebras.
If
1 = 0 and dim Fd(A2)G = ∞, then again Fd(V)G is not finitely
(KXd)G = (Ad)G
generated.
1 = 0 and (A2
1. Preliminaries
We fix a field K of characteristic 0. All vector spaces and algebras will be over
K. Let
F (X) = K{X} = K{x1, x2, . . .}
be the (absolutely) free nonassociative algebra freely generated by the countable
set X = {x1, x2, . . .}. Recall that the polynomial f (x1, . . . , xm) ∈ K{X} is a
polynomial identity for the algebra A if f (a1, . . . , am) = 0 for all a1, . . . , am ∈ A.
The class of all algebras satisfying a given set U ⊂ K{X} of polynomial identities is
called the variety of associative algebras defined by the system U . If V is a variety,
then T (V) is the ideal of K{X} consisting of all polynomial identities of V. Let
Xd = {x1, . . . , xd} ⊂ X. Then the algebra
Fd(V) = K{x1, . . . , xd}/(K{x1, . . . , xd} ∩ T (V)) = K{Xd}/(K{Xd} ∩ T (V))
INVARIANT THEORY OF RELATIVELY FREE ALGEBRAS
5
is the relatively free algebra of rank d in V. We shall denote the generators of Fd(V)
with the same symbols Xd. The ideals K{Xd} ∩ T (V) of K{Xd} are preserved by
all endomorphisms ϕ of K{Xd}, i.e., ϕ(K{Xd} ∩ T (V)) ⊆ K{Xd} ∩ T (V).
In
particular, GLd(K)(K{Xd} ∩ T (V)) = K{Xd} ∩ T (V). Here the general linear
group GLd(K) acts canonically on the vector space KXd with basis Xd and this
action is extended diagonally on the whole Fd(V) as in (3). Hence Fd(V) has a
natural structure of a GLd(K)-module. For a background on representation theory
of GLd(K) see, e.g., [27, 38]. Since char(K) = 0, the algebra Fd(V) is a direct sum
of irreducible GLd(K)-modules and
Fd(V) =X mλ(V)Wd(λ),
where Wd(λ) is the irreducible polynomial GLd(K)-module corresponding to the
partition λ = (λ1, . . . , λd), λ1 ≥ · · · ≥ λd ≥ 0, and mλ(V) is the multiplicity of
Wd(λ) in the decomposition of Fd(V). Then the Hilbert series of Fd(V) is
H(Fd(V), z1, . . . , zd) =X mλ(V)Sλ(z1, . . . , zd),
where Sλ(z1, . . . , zd) is the Schur function corresponding to λ. Since the Schur
functions form a basis of the vector space K[Xd]Sn of symmetric polynomials in d
variables, the Hilbert series H(Fd(V), z1, . . . , zd) determines the GLd(K)-module
structure of Fd(V).
In the sequel we shall need some well known information for two relatively
free algebras: the polynomial algebra K[Xd] and the free metabelian Lie algebra
Fd(A2) = Ld/L′′
d.
Lemma 1.1. (i) The GLd(K)-module structure of the polynomial algebra K[Xd]
is
K[Xd] =Xn≥0
Wd(n).
(ii) The free metabelian Lie algebra Fd(A2) has a basis
{xi, [[· · · [xi1 , xi2 ], . . .], xin ] i, ij = 1, . . . , d, i1 > i2 ≤ · · · ≤ in}.
The GLd(K)-module structure of Fd(A2) is
Fd(A2) = Wd(1) +Xn≥2
Wd(n − 1, 1).
Part (i) of the lemma is well known. Part (ii) is also well known, see e.g., [34,
§52, pp. 274-276 of the English translation] for the basis of Fd(A2) and [17, the
proof of Lemma 2.5] for its GLd(K)-module structure.
The product of two Schur functions Sλ(z1, . . . , zd)Sµ(z1, . . . , zd) can be expressed
as a sum of Schur functions using the Littlewood-Richardson rule. A very special
case of this rule is the Branching theorem, when µ = (1). It states that
Sλ(z1, . . . , zd)S(1)(z1, . . . , zd) =X Sν(z1, . . . , zd),
where the sum runs on all partitions ν = (ν1, . . . , νd) obtained by adding 1 to one
of the components λi of λ = (λ1, . . . , λd).
In other words, the Young diagram
of ν is obtained by adding a box to the diagram of λ. Since the product of two
Schur functions corresponds to the tensor product of the corresponding irreducible
GLd(K)-modules, we obtain equivalently
(5)
Wd(λ) ⊗K Wd(1) =X Wd(ν),
(6)
6
VESSELIN DRENSKY
with the same summation on ν as in (5).
If G is a subgroup of GLd(K), then the GLd(K)-action on the irreducible
GLd(K)-module Wd(λ) induces a G-action on Wd(λ). Let Wd(λ)G be the vec-
tor space of the elements of Wd(λ) fixed by G, i.e., of the G-invariants of Wd(λ). If
W is a graded GLd(K)-module with polynomial homogeneous components,
mλ(k)Wd(λ),
then its Hilbert series is
W =Mk≥0
Wk, Wk =Xλ
H(W, z1, . . . , zd, z) =Xk≥0
Xni≥0
=Xk≥0Xλ
dim(Wk)(n1,...,nd)zn1
1 · · · znd
mλ(k)Sλ(z1, . . . , zd)zk.
(7)
(8)
d
zk
Lemma 1.2. Let W be a graded GLd(K)-module with polynomial homogeneous
components, as in (7), and let G be a subgroup of GLd(K). Then the Hilbert series
of the G-invariants of W
H(W G, z) =Xk≥0
dim W G
k zk
is determined from the Hilbert series (8) of W .
Proof. We follow the main ideas of the recent preprint [10] which contains more
applications in the spirit of the lemma. Since the dimension of Wd(λ)G depends on
Wd(λ) only, and the Schur functions Sλ(z1, . . . , zd) are in 1-1 correspondence with
the modules Wd(λ), we conclude that dim Wd(λ)G is a function of Sλ(z1, . . . , zd).
This immediately completes the proof because
H(W G, z) =Xk≥0 Xλ
mλ(k) dim Wd(λ)G! zk.
(cid:3)
The proof of the next statement can be found in [16, Proposition 4.2] in the case
of homomorphic images of the free associative algebra KhXdi. The proof in the
case below is exactly the same.
Proposition 1.3. Let I be an ideal of the relatively free algebra Fd(W) of the
variety W and let I be preserved under the GLd(K)-action on Fd(W). If G is a
subgroup of GLd(K), then every G-invariant of the factor algebra Fd(W)/I can be
lifted to a G-invariant of Fd(W), i.e., under the canonical homomorphism
π : Fd(W) → Fd(W)/I
Fd(W)G maps onto (Fd(W)/I)G. In particular, if V is a subvariety of the variety
W and π : Fd(W) → Fd(V), then
π(Fd(W)G) = Fd(V)G.
For more details on varieties of algebras (in the associative case) and the ap-
plications of representation theory of GLd(K) to PI-algebras we refer to the book
[14].
INVARIANT THEORY OF RELATIVELY FREE ALGEBRAS
7
2. The main result
Let L be the subvariety of the variety of right-symmetric algebras R defined
by the identity (4) of left-nilpotency of class 3. Since the identity (2) of left-
commutativity is a consequence of (4), L is also a subvariety of the variety N of
Novikov algebras. Working in L, the only nonzero products are left-normed. We
shall omit the parentheses and shall write a1a2 · · · an instead of (· · · (a1a2) · · · )an
and a1ak
.
2 instead of a1 a2 · · · a2
k times
{z }
Lemma 2.1. (i) The relatively free algebra Fd(L) has a basis
{xi1 xi2 · · · xin i1 = 1, . . . , d,
1 ≤ i2 ≤ · · · ≤ in ≤ d}.
(ii) The GLd(K)-module structure of Fd(L) is
Fd(L) = Wd(1) +Xn≥2
(Wd(n) + Wd(n − 1, 1)).
Proof. (i) Modulo the identity (4) the right-symmetric identity (1) reduces to
Hence L satisfies the identity
x1x2x3 = x1x3x2.
(9)
(10)
(11)
x1x2 · · · xn = x1xσ(2) · · · xσ(n),
σ ∈ Sn, σ(1) = 1,
and the algebra Fd(L) is spanned as a vector space on the elements (9). In order
to show that (9) is a basis of Fd(L) it is sufficient to construct an algebra A in L
which is generated by a1, . . . , ad and has a basis
{aian1
1 · · · and
d i1 = 1, . . . , d, nj ≥ 0}.
(12)
Since A is a homomorphic image of Fd(L), this would imply that (9) is a basis
of Fd(L). Consider the vector space A with basis (12) and define a multiplication
there by the rule
(aian1
1 · · · and
1 · · · and
d ) ∗ (aiam1
1
(aian1
d ) ∗ aj = aian1
1 · · · anj +1
j
· · · and
d ,
· · · amd
d ) = 0, if mj > 0 for some j.
Obviously A satisfies the identities (4) and (11), and hence belongs to L.
(ii) For n ≥ 2 we divide the basis elements from (9) in two groups. The first
group contains the monomials xi1 xi2 · · · xin with i1 ≤ i2 and the second group the
monomials with i1 > i2. Obviously, the monomials in the first group are in 1-1
correspondence with the monomials of degree ≥ 2 in K[Xd]. By Lemma 1.1 (ii),
the same holds for the monomials from the second group and the elements of degree
≥ 2 in Fd(A2). Hence the Hilbert series of Fd(L) is a sum of the Hilbert series of
the algebra of polynomials without constant term and the commutator ideal of the
Lie algebra Fd(A2). Now the proof follows from Lemma 1.1.
(cid:3)
The construction in the proof of Lemma 2.1 (i) suggests that the algebra Fd(L)
has the structure of a right K[Xd]-module with action defined by
(xpxn1
1 · · · xdnd) ◦ (xm1
1
Clearly, the ideal F 2
d (L) of the elements in Fd(L) without linear term is a K[Xd]-
submodule. We shall denote by (Ad)1 the vector space KXd and shall identify
K[Xd] and K[(Ad)1].
p = 1, . . . , d, nj, mj ≥ 0.
d ) = xpxn1+m1
· · · xnd+md
· · · xmd
1
,
d
8
VESSELIN DRENSKY
The following theorem and its consequences together with the examples in the
next section are the main results of the paper.
Theorem 2.2. Let V be a subvariety of the variety R of all right-symmetric al-
gebras and let V contain the variety L of left-nilpotent of class 3 algebras in R.
If G 6= h1i is a subgroup of GLd(K) such that the ideal F 2
d (L)G of the algebra of
invariants Fd(L)G is not finitely generated as a K[(Ad)G
1 ]-module, then the algebra
of G-invariants Fd(V)G is not finitely generated.
Proof. By Proposition 1.3 the canonical homomorphism Fd(V) → Fd(L) maps
Fd(V)G onto Fd(L)G and if Fd(V)G is finitely generated, the same is Fd(L)G.
Hence it is sufficient to show that Fd(L)G is not finitely generated. Therefore we
may work in Fd(L) and assume that Fd(L)G is finitely generated. As a vector
space Fd(L)G is a direct sum of the invariants of first degree (KXd)G = (Ad)G
1
and the invariants F 2
d (L)G without linear term. We may assume that Fd(L)G is
d (L)G. Since
generated by U = {u1, . . . , uk} ⊂ (Ad)G
Fd(L)GF 2
d (L)G = 0, the only nonzero products of the generators of Fd(L)G are
upui1 · · · uum, and wqui1 · · · uum, m ≥ 0. Hence KU = (Ad)G
1 ,
1 and W = {w1, . . . , wl} ⊂ F 2
F 2
d (L)G =
k
Xi=1
up1up2 ◦ K[(Ad)G
1 ] +
wq ◦ K[(Ad)G
1 ]
l
Xj=1
and F 2
d (L)G is a finitely generated K[(Ad)G
1 ]-module which is a contradiction. (cid:3)
Corollary 2.3. Let Ad = K[Xd]+ be the algebra of polynomials without constant
term and let G be a subgroup of GLd(K). If (A2
d)G is not finitely generated as a
K[(Ad)G
1 ]-module, then Fd(V)G is not finitely generated for any variety V contain-
ing L.
Proof. By the Branching theorem (6)
Wd(n − 1, 1) ⊗K Wd(1) = Wd(n, 1) ⊕ Wd(n − 1, 2) ⊕ Wd(n − 1, 1, 1).
(13)
Consider the GLd(K)-module decomposition of Fd(L) given in Lemma 2.1 (ii).
Since Fd(L)F 2
d (L) = 0, the only nonzero products Wd(λ)Wd(µ) with λ or µ equal
to (n − 1, 1), n ≥ 2, come from
Wd(n − 1, 1)Wd(1) = Wd(n − 1, 1)Fd(L) = Wd(n − 1, 1)(KXd) = Wd(n − 1, 1)(Ad)1.
This is a homomorphic image in Fd(L) of Wd(n − 1, 1) ⊗K Wd(1). By (13) we derive
that Wd(n − 1, 1)Fd(L) ⊂ Wd(n, 1). This implies that
is an ideal of Fd(L) and the GLd(K)-module structure of the factor algebra is
I =Xn≥2
Fd(L)/I =Xn≥1
Wd(n − 1, 1) ⊂ Fd(L)
Wd(n) ∼= Ad.
Hence the algebras Fd(L)/I and Ad have the same Hilbert series and by Lemma
1.2 the same holds for their algebras of invariants. Since (A2
d)G is not finitely gen-
erated as a K[(Ad)G
d (L)/I.
By Proposition 1.3, the K[(Ad)G
d (L) is not finitely generated and the
application of Theorem 2.2 completes the proof.
(cid:3)
1 ]-module, the same is true for the K[(Ad)G
1 ]-module F 2
1 ]-module F 2
INVARIANT THEORY OF RELATIVELY FREE ALGEBRAS
9
Corollary 2.4. Let Fd(A2) be the free metabelian Lie algebra and let G be a sub-
1 = 0 and dim Fd(A2)G = ∞, then Fd(V)G
group of GLd(K). If (KXd)G = (Ad)G
is not finitely generated for any variety V containing L.
1 = 0 we obtain that Fd(L)G = F 2
Proof. Since (KXd)G = (Ad)G
d (L)G. Hence the
algebra Fd(L)G is with trivial multiplication and the finite generation is equivalent
to the finite dimensionality. As a GLd(K)-module Fd(A2) is a homomorphic image
of Fd(L). Hence the vector space Fd(A2)G is a homomorphic image of Fd(L)G.
This implies that dim F 2
d (L)G = ∞, i.e., both the algebras Fd(L)G and Fd(V)G are
not finitely generated.
(cid:3)
Remark 2.5. In Corollary 2.4 we cannot remove directly the restriction (KXd)G =
0, as in Corollary 2.3, because the GLd(K)-submodule I =Pn≥2 Wd(n) of Fd(L)
is not an ideal. For example, one can show that Wd(2)(KXd) = Wd(3) ⊕ Wd(2, 1).
Hence we cannot use the property that the Lie algebra Fd(A2)G is not finitely
generated to show that the algebra Fd(L) is also not finitely generated. On the other
hand, we do not know examples of groups G when (KXd)G = 0, K[Xd]G = K, and
dim Fd(A2)G = ∞. Such an example would show that we may apply Corollary 2.4
when we cannot apply Corollary 2.3.
3. Examples
All examples in this section use the following statement which is a consequence
of Corollary 2.3.
Proposition 3.1. If for a subgroup G of GLd(K)
transcend.deg(K[Xd]G) > dim(KXd)G,
then the algebra Fd(V)G is not finitely generated for any variety V containing L.
Proof. Let t = transcend.deg(K[Xd]G). Since K[Xd]G is graded, we may choose
t algebraically independent homogeneous elements in AG
d = (K[Xd]+)G. If m =
dim(KXd)G, changing linearly the variables Xd we assume that (KXd)G has a basis
1 ] = K[Xm]. Since t > m, we obtain that ((Ad)2)G
Xm = {x1, . . . , xm} and K[(Ad)G
contains an element f (Xd), such that the system Xm ∪ {f (Xd)} is algebraically
independent. Hence the K[Xm]-module generated by the powers f k(Xd), k =
1, 2, . . ., is not finitely generated. Now the proof follows from Corollary 2.3.
(cid:3)
3.1. Finite groups.
Theorem 3.2. Let G be a finite subgroup of GLd(K) and G 6= h1i. Then the
algebra Fd(V)G is not finitely generated for any variety V containing L.
Proof. It is well known that for a finite group G
transcend.deg(K[Xd]G) = transcend.deg(K[Xd]) = d.
(14)
For self-containedness of the exposition, every element f (Xd) ∈ K[Xd] satisfies the
equation
uf (z) = Yg∈G
(z − g(f (Xd))) = zG − c1zG−1 + c2zG−2 − · · · ± cG
10
VESSELIN DRENSKY
where the coefficients ck are equal to the elementary symmetric polynomials in
{g(f (Xd)) g ∈ G}. Hence ck ∈ K[Xd]G and as a K[Xd]G-module K[Xd] is
generated by
1 · · · xad
xa1
d ,
0 ≤ ai < G.
The finite generation of the K[Xd]G-module K[Xd] implies (14) and the theorem
follows from Proposition 3.1.
(cid:3)
3.2. Reductive groups. If G ⊂ GLd(K) is a reductive group then there exists a
G-submodule W of KXd such that KXd = (KXd)G ⊕ W .
Proposition 3.3. In the above notation, if K[W ]G 6= K, then Fd(V)G is not
finitely generated for all V containing L.
Proof. Since the elements of K[W ] cannot be expressed as polynomials in (KXd)G,
the condition K[W ]G 6= K implies that
transcend.deg(K[Xd]G) = dim(KXd)G + transcend.deg(K[W ]G) > dim(KXd)G
and this completes the proof in virtue of Proposition 3.1.
(cid:3)
Example 3.4. For each k ≥ 1 there is a unique irreducible rational k-dimensional
SL2(K)-module Wk. Let the subgroup G of GLd(K) be isomorphic to SL2(K) and
KXd ∼= Wk1 ⊕ · · · ⊕ Wkp
It is well known that if k1 ≥ 3, then K[Wk1 ] contains
as an SL2(K)-module.
nontrivial SL2(K)-invariants. Similarly, K[W2 ⊕ W2]SL2(K) 6= K. Hence the only
cases when K[Xd]SL2(K) = K[(KXd)SL2(K)] are k1 = 2, k2 = · · · = kp = 1 when
K[Xd]SL2(K) = K[(KXd)SL2(K)] ∼= K[Xd−1] and k1 = · · · = kp = 1 with the trivial
action of SL2(K) on KXd (and the latter case is impossible because G ∼= SL2(K)
is a nontrivial subgroup of GLd(K)).
3.3. Weitzenbock derivations. A linear operator δ of an algebra A is a derivation
if
δ(uv) = δ(u)v + uδ(v),
u, v ∈ A.
If V is a variety of algebras, then every mapping δ : Xd → Fd(V) can be uniquely
extended to a derivation of Fd(V) which we shall denote by the same symbol δ.
If δ is a nilpotent linear operator on KXd, then the induced derivation is called
a Weitzenbock derivation. Weitzenbock [36] proved that in the case of polynomial
algebras the algebra of constants
K[Xd]δ = {f (Xd) ∈ K[Xd] δ(f (Xd)) = 0}
is finitely generated. Details on the algebra of constants K[Xd]δ can be found in the
book by Nowicki [32]. For varieties V of unitary associative algebras (and δ 6= 0)
the algebra Fd(V)δ is finitely generated if and only if V does not contain the algebra
T2(K) of 2 × 2 upper triangular matrices, see [15, 16]. Up to a change of the basis
of KXd the Weitzenbock derivation δ is determined by the Jordan normal form
J(δ) of the linear operator δ acting on KXd. Since δ acts nilpotently on KXd, the
matrix J(δ) consists of Jordan blocks with zero diagonals.
Proposition 3.5. If d > 2 and the Jordan normal form J(δ) of the Weitzenbock
derivation consists of less than d − 1 blocks, then the algebra Fd(V)δ is not finitely
generated for any variety V containing the variety L.
INVARIANT THEORY OF RELATIVELY FREE ALGEBRAS
11
Proof. Since αδ, α ∈ K, is nilpotent on KXd, it is a locally nilpotent deriva-
tion of Fd(V), i.e., for every f (Xd) ∈ Fd(V) there exists an n ≥ 1 such that
(αδ)n(f (Xd)) = 0. Hence
exp(αδ) = 1 +
αδ
1!
+
(αδ)2
2!
+ · · ·
is a well defined linear automorphism of Fd(V). It is well known that the group
{exp(αδ) α ∈ K}
is isomorphic to the unipotent group U T2(K) and
Fd(V)δ = Fd(V)U T2(K).
If the matrix J(δ) consists of p blocks, then the dimension of the vector space
(KXd)δ of the linear constants is equal to the number of the blocks p. Reading
carefully [32, Proposition 6.5.1, p. 65] we can see that
transcend.deg(K[Xd]δ) = d − 1
which is larger than p = dim(KXd)δ. Now the proof follows from Proposition 3.1
applied for U T2(K) ⊂ GLd(K).
(cid:3)
Remark 3.6. If in Proposition 3.5 the Jordan normal form of δ consists of d − 1
blocks, then the algebra of constants K[Xd]δ is generated by linear constants. In
this case we may assume that δ(x1) = x2 and δ(xi) = 0 for i = 2, . . . , d. It is easy
to see that Fd(L)δ is generated by x1x2 − x2x1, x2, . . . , xd. We do not know how
far can be lifted to Fd(V) the finite generation property of the algebra of constants
and do not have a description of the varieties V containing L such that the algebra
Fd(V)δ is finitely generated.
References
[1] G. Almkvist, W. Dicks, E. Formanek, Hilbert series of fixed free algebras and noncommutative
classical invariant theory, J. Algebra 93 (1985), 189-214.
[2] A.A. Balinskii, S.P. Novikov, Poisson brackets of hydrodynamic type, Frobenius algebras and
Lie algebras (Russian), Dokl. Akad. Nauk SSSR 283 (1985), No. 5, 1036-1039. Translation:
Sov. Math., Dokl. 32 (1985), 228-231.
[3] F. Benanti, S. Boumova, V. Drensky, G.K. Genov, P. Koev, Computing with rational sym-
metric functions and applications to invariant theory and PI-algebras, Serdica Math. J. 38
(2012), Nos 1-3, 137-188.
[4] L.A. Bokut, Y. Chen, Z. Zhang, On free Gelfand-Dorfman-Novikov-Poisson algebras and a
PBW theorem, arXiv:1604.06676v1 [math.RA].
[5] R.M. Bryant, On the fixed points of a finite group acting on a free Lie algebra, J. Lond.
Math. Soc., II. Ser. 43 (1991), No. 2, 215-224.
[6] A. Cayley, On the theory of analitical forms called trees, Phil. Mag. 13 (1857), 19-30. Col-
lected Math. Papers, University Press, Cambridge, vol. 3, 1890, 242-246.
[7] C. Chevalley, Invariants of finite groups generated by reflections, Amer. J. Math. 67 (1955),
778-782.
[8] C. De Concini, D. Eisenbud, C. Procesi, Young diagrams and determinantal varieties, Invent.
Math. 56 (1980), 129-165.
[9] H. Derksen, G. Kemper, Computational Invariant Theory, Invariant Theory and Algebraic
Transformation Groups, I. Encyclopaedia of Mathematical Sciences, 130, Springer-Verlag,
Berlin, 2002.
[10] M. Domokos, V. Drensky, Rationality of Hilbert series in noncommutative invariant theory,
arXiv:1512.06411v1 [math.RA].
[11] J.A. Dieudonn´e, J.B. Carrell, Invariant Theory, Old and New, Academic Press, New York-
London, 1971.
12
VESSELIN DRENSKY
[12] V. Drensky, Fixed algebras of residually nilpotent Lie algebras, Proc. Amer. Math. Soc. 120
(1994), 1021-1028.
[13] V. Drensky, Commutative and noncommutative invariant theory, Math. and Education in
Math., Proc. of the 24-th Spring Conf. of the Union of Bulgar. Mathematicians, Svishtov,
April 4-7, 1995, Sofia, 1995, 14-50.
[14] V. Drensky, Free Algebras and PI-Algebras, Springer-Verlag, Singapore, 1999.
[15] V. Drensky, Invariants of unipotent transformations acting on noetherian relatively free al-
gebras, Serdica Math. J. 30 (2004), 395-404.
[16] V. Drensky, C.K. Gupta, Constants of Weitzenbock derivations and invariants of unipotent
transformations acting on relatively free algebras, J. Algebra 292 (2005), 393-428.
[17] V. Drensky, A. Kasparian, Polynomial identities of eighth degree for 3×3 matrices, Annuaire
de l'Univ. de Sofia, Fac. de Math. et Mecan., Livre 1, Math. 77 (1983), 175-195.
[18] A.S. Dzhumadildaev, Codimension growth and non-Koszulity of Novikov operad, Commun.
Algebra 39 (2011), No. 8, 2943-2952.
[19] A.S. Dzhumadildaev, N.A. Ismailov, Sn- and GLn-module structures on free Novikov alge-
bras, J. Algebra 416 (2014), 287-313.
[20] A.S. Dzhumadil'daev, C. Lofwall, Trees, free right-symmetric algebras, free Novikov algebras
and identities, The Roos Festschrift volume, 1, Homology Homotopy Appl. 4 (2002), No. 2,
Part 1, 165-190.
[21] E. Formanek, Noncommutative invariant theory, Contemp. Math. 43, 1985, 87-119.
[22] I.M. Gel'fand, I.Ya. Dorfman, Hamiltonian operators and algebraic structures structures re-
lated to them (Russian), Funktsional. Anal. i Prilozhen. 13 (1979), No. 4, 13-30. Translation:
Funct. Anal. Appl. 13 (1980), 248-262.
[23] F. D. Grosshans, Algebraic Homogeneous Spaces and Invariant Theory, Lecture Notes in
Mathematics, 1673, Springer-Verlag, Berlin, 1997.
[24] Dz. Hadziev, Some questions in the theory of vector invariants (Russian), Mat. Sb. (NS)
72(114), (1967), 420-435. Translation: Math. USSR, Sb. 1 (1967), 383-396.
[25] D. Hilbert, Mathematische Probleme, Archiv f. Math. u. Phys. 1 (1901), 44-63, 213-
237; reprinted in "Gesammelte Abhandlungen, Band III, Analysis, Grundlagen der Math-
ematik, Physik, Verschiedenes, Lebensgeschichte", Zweite Auflage, Springer-Verlag, Berlin-
Heidelberg-New York, 1970, 290-329.
[26] O.G. Kharlampovich, M.V. Sapir, Algorithmic problems in varieties, Intern. J. Algebra and
Computation 5 (1995), 379-602.
[27] I.G. Macdonald, Symmetric Functions and Hall Polynomials, Oxford Univ. Press (Claren-
don), Oxford, 1979, Second Edition, 1995.
[28] T. Molien, Uber die Invarianten der linearen Substitutionsgruppen, Sitz. Konig Preuss. Akad.
Wiss. (1897), No. 52, 1152-1156.
[29] M. Nagata, On the 14th problem of Hilbert, Amer. J. Math. 81 (1959), 766-772.
[30] E. Noether, Der Endlichkeitssatz der Invarianten endlicher Gruppen, Math. Ann. 77 (1916),
89-92; reprinted in "Gesammelte Abhandlungen. Collected Papers", Springer-Verlag, Berlin-
Heidelberg-New York-Tokyo, 1983, 181-184.
[31] S.P. Novikov, Geometry of conservative systems of hydrodynamic type. The method of aver-
aging for field-theoretical systems (Russian), International conference on current problems in
algebra and analysis (Moscow-Leningrad, 1984), Uspekhi Mat. Nauk 40 (1985), No. 4(244),
79-89. Translation: Russ. Math. Surv. 40 (1985), No.4, 85-98.
[32] A. Nowicki, Polynomial Derivations and Their Rings of Constants, Uniwersytet Mikolaja
Kopernika, Torun, 1994. www-users.mat.umk.pl/anow/ps-dvi/pol-der.pdf
[33] G.C. Shephard, J.A. Todd, Finite unitary reflection groups, Canad. J. Math. 6 (1954), 274-
304.
[34] A.L. Shmel'kin, Free polynilpotent groups, Izv. Akad. Nauk SSSR Ser. Mat. 28 (1964), 91-
122. Translation: Amer. Math. Soc. Transl. Ser. 2, 55, Amer. Math. Soc., Providence, RI,
1966, 270-304.
[35] B. Sturmfels, Algorithms in Invariant Theory, Second Edition, Texts and Monographs in
Symbolic Computation, Springer-Wien-New York, Vienna, 2008.
[36] R. Weitzenbock, Uber die Invarianten von linearen Gruppen, Acta Math. 58 (1932), 231-293.
[37] H. Weyl, Zur Darstellungstheorie und Invariantenabzahlung der projektiven, der Komplex-
und der Drehungsgruppe, Acta Math. 48 (1926), 255-278; reprinted in "Gesammelte Abhand-
lungen", Band III, Springer-Verlag, Berlin - Heidelberg - New York, 1968, 1-25.
INVARIANT THEORY OF RELATIVELY FREE ALGEBRAS
13
[38] H. Weyl, The Classical Groups. Their Invariants and Representations, Princeton Univ. Press,
Princeton, N.J., 1946, New Edition: Princeton Landmarks in Mathematics, 1997.
Institute of Mathematics and Informatics, Bulgarian Academy of Sciences, 1113
Sofia, Bulgaria
E-mail address: [email protected]
|
1411.1382 | 2 | 1411 | 2016-03-29T09:21:20 | The ring of polynomials integral-valued over a finite set of integral elements | [
"math.RA"
] | Let $D$ be an integral domain with quotient field $K$ and $\Omega$ a finite subset of $D$. McQuillan proved that the ring ${\rm Int}(\Omega,D)$ of polynomials in $K[X]$ which are integer-valued over $\Omega$, that is, $f\in K[X]$ such that $f(\Omega)\subset D$, is a Pr\"ufer domain if and only if $D$ is Pr\"ufer. Under the further assumption that $D$ is integrally closed, we generalize his result by considering a finite set $S$ of a $D$-algebra $A$ which is finitely generated and torsion-free as a $D$-module, and the ring ${\rm Int}_K(S,A)$ of integer-valued polynomials over $S$, that is, polynomials over $K$ whose image over $S$ is contained in $A$. We show that the integral closure of ${\rm Int}_K(S,A)$ is equal to the contraction to $K[X]$ of ${\rm Int}(\Omega_S,D_F)$, for some finite subset $\Omega_S$ of integral elements over $D$ contained in an algebraic closure $\bar K$ of $K$, where $D_F$ is the integral closure of $D$ in $F=K(\Omega_S)$. Moreover, the integral closure of ${\rm Int}_K(S,A)$ is Pr\"ufer if and only if $D$ is Pr\"ufer. The result is obtained by means of the study of pullbacks of the form $D[X]+p(X)K[X]$, where $p(X)$ is a monic non-constant polynomial over $D$: we prove that the integral closure of such a pullback is equal to the ring of polynomials over $K$ which are integral-valued over the set of roots $\Omega_p$ of $p(X)$ in $\bar K$. | math.RA | math | J. Commut. Algebra 8 (2016), no. 1, 113-141. http://dx.doi.org/10.1216/JCA-2016-8-1-113.
The ring of polynomials integral-valued over a finite set of
6
1
0
2
r
a
M
9
2
integral elements
Giulio Peruginelli∗
July 24, 2018
Abstract
]
.
A
R
h
t
a
m
[
2
v
2
8
3
1
.
1
1
4
1
:
v
i
X
r
a
Let D be an integral domain with quotient field K and Ω a finite subset of D. McQuillan
proved that the ring Int(Ω, D) of polynomials in K[X] which are integer-valued over Ω, that
is, f ∈ K[X] such that f (Ω) ⊂ D, is a Prufer domain if and only if D is Prufer. Under the
further assumption that D is integrally closed, we generalize his result by considering a finite
set S of a D-algebra A which is finitely generated and torsion-free as a D-module, and the ring
IntK(S, A) of integer-valued polynomials over S, that is, polynomials over K whose image over
S is contained in A. We show that the integral closure of IntK (S, A) is equal to the contraction
to K[X] of Int(ΩS, DF ), for some finite subset ΩS of integral elements over D contained in an
algebraic closure K of K, where DF is the integral closure of D in F = K(ΩS). Moreover,
the integral closure of IntK (S, A) is Prufer if and only if D is Prufer. The result is obtained
by means of the study of pullbacks of the form D[X] + p(X)K[X], where p(X) is a monic
non-constant polynomial over D: we prove that the integral closure of such a pullback is equal
to the ring of polynomials over K which are integral-valued over the set of roots Ωp of p(X) in
K.
Keywords: Pullback, Integral closure, Integer-valued polynomial, Divided differences, Prufer ring. MSC
Classification codes: 13B25 (primary), 13F20, 13B22, 13F05 (secondary).
1
Introduction
Rings of integer-valued polynomials are a prominent source for providing examples of non-Noetherian Prufer
domains (see the book [6, Chapt. VI, p. 123]). Throughout this paper, D is an integral domain which is
not a field, and K is its quotient field. We denote by K a fixed algebraic closure of K and by D the integral
closure of D in K. We give the following definition, which generalizes the classical definition of the ring of
integer-valued polynomials over a subset ([6, Chapt. I.1, p. 3]).
Definition 1.1. Let R be an integral domain containing D. Let F be the quotient field of R (so that
K ⊆ F ). For a subset Ω of F we set
∗Department of Mathematics, University of Padova, Via Trieste, 63 35121 Padova, Italy. E-mail: gperu-
[email protected]
IntK(Ω, R) + {f ∈ K[X] f (Ω) ⊂ R},
1
which is the ring of polynomials in K[X] which map every element of Ω into R. If F = K we omit the
subscript K. Thus, Int(Ω, R) is a subring of K[X] (the coefficients of the relevant polynomials are in the
quotient field of R).
In the case of a finite subset Ω of D, McQuillan studied the algebraic structure of the corresponding
ring of integer-valued polynomials Int(Ω, D), describing the spectrum of such a ring and also its additive
structure ([13]). Using McQuillan's arguments, in [5] Boynton observed that Int(Ω, D) fits in a pullback
diagram. Here we generalize this class of rings by considering first a finite set Ω of integral elements in
D and polynomials in K[X] which preserve the integrality of the elements of Ω, that is, for each α in Ω,
f (α) is integral over D; according to the above definition, this ring is denoted by IntK (Ω, D). For example,
given a monic non-constant polynomial p ∈ D[X], let Ωp be the set of roots of p(X) in a splitting field.
Then the ring IntK(Ωp, D) is of the above kind, and it is not difficult to show that the ring IntK (Ω, D),
for a finite set Ω of D, can be reduced to this case. More generally, we consider a finite set S of integral
elements over D which do not necessarily lie in an algebraic extension of K, i.e.: S is contained in a
D-algebra A, which is finitely generated and torsion-free as a D-module (for example, a matrix algebra
or a quaternion algebra). We consider then polynomials in K[X] which map the elements of S into A:
IntK (S, A) = {f ∈ K[X] f (S) ⊂ A}. Note that A is not necessarily commutative and may contain
zero-divisors, and each of its elements satisfies a monic polynomial over D.
Given a monic polynomial p(X) in D[X], the study of the ring IntK (Ωp, D) we are going to do goes
through another kind of pullback ring. As for the rings IntK (Ω, D), the rings we introduce now are the
pullbacks of the canonical residue map K[X] ։
p(X)K[X] , thus
they are subrings of K[X] sharing with K[X] the ideal p(X)K[X].
p(X)K[X] with respect to some subring of
K[X]
K[X]
Definition 1.2. Let p(X) be a non-constant monic polynomial in D[X]. We consider the following subring
of K[X]:
D(p) + D[X] + p(X) · K[X] = {r(X) + p(X)q(X) r ∈ D[X], q ∈ K[X]}.
It is straightforward to verify that the elements of this set form a ring under the usual operation of sum
and product induced by the polynomial ring K[X]. In Lemma 1.6 we will show that a polynomial f (X) in
K[X] is in D(p) if and only if the remainder in the division of f (X) by p(X) is in D[X]. Note that the
principal ideal p(X) · K[X] of K[X] is also an ideal of D(p). We have then the following diagram:
D(p)
/ K[X]
D(p)
p(X)K[X]
/ K[X]
p(X)K[X]
so that D(p) is a pullback of K[X] (for a general reference about pullbacks see [9]). Examples of such
pullbacks appear in [5], and more widely in [16].
We see at once that D(p) is contained in IntK (Ωp, D). Also, IntK (Ωp, D) has the ideal p(X)K[X] in
common with K[X], so, like D(p), also IntK(Ωp, D) is a pullback ring. This point of view is clearly a
generalization of [5, Example 4.4 (1)], which we briefly recall below in section 1.2.
We give some motivation which led us to study the pullback rings D(p) = D[X] + p(X) · K[X]. In
[16] this kind of polynomial pullback arose as the ring of integer-valued polynomials over certain subsets
of matrices. Let Mn(D) be the D-algebra of n × n matrices with entries in D and let IntK (Mn(D)) =
{f ∈ K[X] f (Mn(D)) ⊂ Mn(D)}, the ring of integer-valued polynomials over Mn(D). Given a monic
polynomial p ∈ D[X] of degree n, we denote by M p
n(D) the set of matrices M in Mn(D) whose characteristic
polynomial is equal to p(X). We consider the overring of IntK(Mn(D)) made up by those polynomials which
are integer-valued over M p
n(D), namely:
IntK(M p
n(D), Mn(D)) = {f ∈ K[X] f (M p
n(D)) ⊂ Mn(D)}
2
/
/
This partition of Mn(D) into subsets of matrices having prescribed characteristic polynomial was used in
[16] to give a characterization of the polynomials of IntK (Mn(D)) in terms of their divided differences (see
[16, Theorem 4.1]). By [16, Lemma 2.2 & Remark 2.1] we have
IntK (M p
n(D), Mn(D)) = D(p).
In particular, the ring IntK (Mn(D)) is represented as an intersection of pullbacks D(p), as p(X) ranges
through the set of all the monic polynomials in D[X] of degree n ([16, Remarks 2.1 and 2.2]). In [17] the
authors address the following question, which generalizes the previous case: for a D-algebra A as above,
where D is integrally closed, we consider the ring IntK(A) = {f ∈ K[X] f (A) ⊂ A} of integer-valued
polynomials over A. Is IntK(A) equal to the intersection of pullbacks of the form D(p)? In general, we
have
D(µa) ⊆ IntK (A),
\a∈A
where, for a ∈ A, µa(X) denotes the minimal polynomial of a over K (by assumption on A and D,
µa ∈ D[X] and is monic). The conditions under which the previous containment is an equality are not
known.
Throughout the paper, given a monic polynomial p(X) in D[X], we denote by Ωp the multi-set of its
roots in K (we recall the notion of multi-set in section 2).
This work is organized as follows. In section 2 we recall a characterization for the polynomials in D(p)
in terms of their divided differences. We use this result to show that the ring Int{n}(Ω, D) of polynomials
whose divided differences of order less than or equal to n are integer-valued over a subset Ω of D can be
represented as an intersection of such pullbacks. This ring has been introduced by Bhargava in [1]; we recall
the definition in that section. In section 3 we prove the following theorem:
Theorem 1.3. Let p(X) a monic non-constant polynomial in D[X]. Then the integral closure of the ring
D(p) = D[X] + p(X)K[X] is the ring IntK (Ωp, D).
As a corollary, we show that the integral closure of Int{n}(Ω, D) is equal to the ring Int(Ω, D), in the
case of a finite subset Ω of D. For a general subset Ω of D, in the case where D has finite residue rings, an
argument from [17] gives the same conclusion. In section 4, we prove the main theorem:
Theorem 1.4. Assume D integrally closed and let Ω be a finite subset of D. Then the ring IntK (Ω, D) is
Prufer if and only if D is Prufer.
If Ω ⊂ D, then this is precisely the main result obtained by McQuillan. The crucial remark is that,
for a monic polynomial p(X) in D[X], IntK(Ωp, D) ⊆ IntF (Ωp, D) is an integral ring extension, where
F = K(Ωp) is the splitting field of p(X). It is not difficult to see that IntF (Ω, D) is equal to IntF (Ω, DF ),
where DF is the integral closure of D in F , and this is precisely the kind of ring considered by McQuillan.
We note that this is a partial answer to [17, Question 29], where we asked if IntK (Ω, D) is Prufer, when
Ω is a subset of integral elements of degree over K bounded by some positive integer n. If D is integrally
closed, we give also a criterion to establish when the pullback D(p) is integrally closed, that is, equal to
IntK (Ωp, D) (see Theorem 4.5). In particular, in the case of a Prufer domain D, this condition is satisfied
automatically if D(p) is integrally closed.
Finally, in the last section, we apply the previous results in the more general setting of a finite set S of
integral elements over D which do not necessarily lie in an algebraic extension of K.
Corollary 1.5. Assume D integrally closed and let S be a finite set of a torsion-free D-algebra A, which
is finitely generated as a D-module. Let ΩS be the set of roots in D of the minimal polynomials of s over
D, as s ranges through S. Then the integral closure of IntK (S, A) is IntK (ΩS, D).
3
1.1 Preliminary results
K[X]
In the case of a monic polynomial, the following lemma determines the quotient of D(p) by the ideal
p(X)K[X]. We denote by π : K[X] ։
p(X)K[X] the canonical residue map, which associates to a polynomial
f ∈ K[X] the residue class f (X) + p(X)K[X].
Lemma 1.6. Let p ∈ D[X] be a monic non-constant polynomial. Then D(p) is the pullback of
p(X)K[X] with respect to the canonical residue map π : K[X] ։
a pullback diagram (i.e.: D(p) = π−1( D[X]
p(X)D[X] ֒→
p(X)K[X] . In other words, the following is
K[X]
p(X)D[X] )):
K[X]
D[X]
D(p)
/ K[X]
D[X]
p(X)D[X]
/ K[X]
p(X)K[X]
In particular, a polynomial f ∈ K[X] belongs to D(p) if and only if the remainder in the division by p(X)
in K[X] belongs to D[X]. Equivalently, we have
D(p)
p(X) · K[X] ∼=
D[X]
p(X) · D[X]
.
p(X)D[X] embeds naturally into
p(X)D[X] is a free D-module of
In
Proof. Since p(X) is monic, we have two consequences. Firstly, D[t] ∼= D[X]
rank n = deg(p) with basis {1, t, . . . , tn−1}, where t is the residue class of X modulo p(X)D[X].
particular, every element r ∈ D[t] can be uniquely represented as r(t) = Pi=0,...,n−1 citi, with ci ∈ D.
Secondly, p(X) · K[X] ∩ D[X] = p(X) · D[X], so the image of the restriction of π to D[X] is isomorphic to
D[t]. Therefore, D[t] ∼= D[X]
p(X)K[X] ∼= K[t] (the class X (mod p(X)D[X]) is
mapped to X (mod p(X)K[X]), so without confusion we may denote them with the same letter t). Note
that K[t] is a free K-module of rank n with the same basis {1, t, . . . , tn−1}.
We consider now the composition of mappings D[X] ֒→ D(p) ։ D(p)/p(X)K[X]. By the second
consequence above and by the second isomorphism theorem we have the isomorphism of the claim. More
explicitly, given f ∈ K[X], there exist (uniquely determined) a quotient q ∈ K[X] and a remainder r ∈ K[X]
(with either r = 0 or deg(r) < deg(p)) such that f (X) = r(X) + q(X)p(X). Hence, if r(X) = Pi ciX i,
then π(f ) = π(r) = r(t) = Pi citi ∈ K[t]. From the algebraic structure of D[t] we deduce that r(t) is in
D[t] if and only if the remainder r(X) is in D[X]. This condition in turn is equivalent to f ∈ D(p).
K[X]
Lemma 1.7. Let p, q ∈ D[X] be monic polynomials. Then
D(p) is contained in D(q) ⇔ p(X) is divisible by q(X).
In particular, D(p) = D(q) ⇔ p(X) = q(X).
Proof. One direction is easy. Conversely, suppose D(p) ⊆ D(q) so that p(X) = r(X) + q(X)k(X), for some
r ∈ D[X], r = 0 or deg(r) < deg(q), k ∈ K[X]. If r 6= 0, let c ∈ K \ D be such that c · r(X) is not in D[X].
Then c · p is in D(p) but it is not in D(q), contradiction. Notice that k(X) has to be in D[X] (see also [12,
Lemma]).
The following two cases, linear and irreducible polynomial, are given as an example and to further
illustrate the connection between polynomial pullbacks and rings of integer-valued polynomials.
4
/
/
1.2 Linear case
In the linear case the connection between the polynomial pullbacks and ring of integer-valued polynomials
over finite sets becomes evident. Suppose p(X) = X − a ∈ D[X]. Then the remainder of the division of a
polynomial f ∈ K[X] by X − a is the value of f (X) at a. Hence,
D(p) = D + (X − a) · K[X] = Int({a}, D)
It is well-known (see for example [6, Proposition IV.4.1]) that Int({a}, D) is integrally closed if and only
if D is. It is easy to see that the integral closure of Int({a}, D) is Int({a}, D′), where D′ is the integral
closure of D in K (notice that Int({a}, D′) = D′ + (X − a)K[X] is a pullback). More generally, we recall
the following result.
Lemma 1.8. Let E ⊂ K be a finite set. Then the integral closure of Int(E, D) is Int(E, D′), where D′ is
the integral closure of D in K.
Proof. By [6, Proposition IV.4.1], Int(E, D′) is integrally closed. Conversely, take f ∈ Int(E, D′). Then
for each a ∈ E, there exists a monic polynomial mf (a) ∈ D[X] such that mf (a)(f (a)) = 0. We consider the
monic polynomial of D[X] equal to the product of the mf (a)(X)'s, as a ranges through E. Then m(f (X))
is in Int(E, D), because for each a ∈ E we have m(f (a)) = 0 ∈ D. This gives a monic equation for f (X)
over Int(E, D).
Remark 1.9. We recall now the following observation made in [5]. Under the assumptions of Lemma
i=1 K with respect to the canonical mapping π : K[X] ։
that p(X)K[X] is an ideal of Int(E, D), because every polynomial of K[X] which is divisible by p(X) is
zero on E. In particular, we have the following isomorphism of D-modules
i=1 K, where p(X) =Qa∈E(X − a). The map π is given by f (X) 7→ (f (a))a∈E. Notice also
K[X]
1.8, Int(E, D) is the pullback of Qm
p(X)K[X] ∼= Qm
i=1 D ⊂ Qm
Int(E, D)
p(X)K[X] ∼=
D
m
Yi=1
1.3
Irreducible polynomial case
We suppose now that D is integrally closed and p(X) is a monic irreducible polynomial in D[X] of degree
n > 0. It is easy to see (see for example [12] or [2, Proposition 11, Chapt. V]) that p(X) is irreducible in
K[X], so that p ∈ D[X] is also prime and D[X]/(p(X)) ∼= D[α], where α is a root of p(X) in K. The next
proposition follows by [15, Prop. 3.1] (which is proved in the case D = Z). We sketch the proof for the
sake of the reader, giving emphasis to the relevant points.
Proposition 1.10. Let p ∈ D[X] be a monic and irreducible polynomial, with set of roots Ωp ⊂ K. Let
F = K(Ωp) be the splitting field of p(X) over K and DF the integral closure of D in F . For each α ∈ Ωp
we set
Sα + IntK ({α}, Dα),
Then, for each α ∈ Ωp, Sα = IntK (Ωp, DF ) and this ring is the integral closure of D(p). Moreover,
where Dα is the integral closure of D in K(α) ⊆ F .
D(p) is integrally closed if and only if Dα = D[α], for some (hence all) α ∈ Ωp.
Proof. Using a Galois-invariance argument it is easy to show that the ring Sα does not depend on the choice
of the root α of p(X) and is equal to IntK(Ωp, DF ). We observe that Sα = {f ∈ K[X] f (α) is integral over D}.
Then for a polynomial f ∈ Sα and for every conjugate α′ of α over K, f (α′) is integral over D as well.
Since D[α], for α ∈ Ωp, is a free D-module of rank n, we can show that
D(p) = {f ∈ K[X] f (α) ∈ D[α]} = IntK({α}, D[α]).
5
Finally, using a pullback diagram argument, since Dα is the integral closure of D[α] in K(α), we deduce
that IntK(Ωp, DF ) is the integral closure of D(p) (see [15, Proposition 3.1] for the details).
In particular, the proposition shows that all the subrings Int({α}, DF ) ⊂ F [X], for α ∈ Ωp, contracts in
K[X] to the same ring Sα. Notice also that IntK (Ωp, D) is equal to IntK(Ωp, DF ), where DF is the integral
closure of D in the splitting field F = K(Ωp) of p(X) over K.
2 Pullbacks and divided differences
In this section we recall a result of [16] which characterizes a polynomial f (X) in a pullback D(p) =
D[X] + p(X)· K[X] in terms of a finite set of conditions on the evaluation of the divided differences of f (X)
at the roots of p(X) in K. We use this result to show that the ring of integer-valued polynomials whose
divided differences are also integer-valued can be represented as an intersection of such pullbacks.
Given a polynomial f ∈ K[X], the divided differences of f (X) are defined recursively as follows:
Φ0(f )(X0) +f (X0)
Φ1(f )(X0, X1) +
. . .
f (X0) − f (X1)
X0 − X1
Φk(f )(X0, . . . , Xk) +
Φk−1(f )(X0, . . . , Xk−1) − Φk−1(f )(X0, . . . , Xk−2, Xk)
Xk−1 − Xk
For each k ∈ N, Φk(f ) is a symmetric polynomial over K in k + 1 variables (see [7], [16], [18] and [19] for
the main properties of the divided differences of a polynomial). We recall here that, given a finite sequence
of elements a0, . . . , an of a commutative ring R, and a polynomial f ∈ R[X] of degree ≤ n we have the
following expansion due to Newton:
f (X) = f (a0) + Φ1(f )(a0, a1)(X − a0) + Φ2(f )(a0, a1, a2)(X − a0)(X − a1) + . . .
+Φn(f )(a0, . . . , an)(X − a0) · . . . · (X − an−1)
(2.1)
Since in general a polynomial may not have distinct roots, we need to recall the following definition.
Definition 2.1. A multi-set is a collection of elements Ω in which elements may occur multiple times. The
number of times an element occurs is called its multiplicity in the multi-set. The cardinality of a multi-set
Ω is defined as the number of elements of Ω, each of them counted with multiplicity. The underlying set of
Ω is the (proper) set containing the distinct elements in Ω.
A multi-set Ω1 is a sub-multi-set of a multi-set Ω2 if every element α of Ω1 of multiplicity n1 belongs to Ω2
with multiplicity n2 ≥ n1.
Remark 2.2. Let Ω be a multi-set of cardinality n and let S be the underlying set of Ω. The choice of an
ordering on the elements of Ω corresponds to a n-tuple in Sn (we have thus n! choices). Conversely, given
an n-tuple s in Sn, where S is a set, if we do not consider the order its components, we have a multi-set Ω
of cardinality n.
Remark 2.3. A particular ring of integer-valued polynomials involving divided differences has been intro-
duced by Bhargava in [1]. Given a subset S of D and n ∈ N, we consider those polynomials f (X) in K[X]
whose k-th divided difference Φk(f ) is integer-valued on S for all k ∈ {0, . . . , n}, namely:
Int{n}(S, D) + {f ∈ K[X] ∀0 ≤ k ≤ n, Φk(f )(Sk+1) ⊂ D}.
6
For n = 0 we recover the ring Int(S, D), which contains Int{n}(S, D) for all n ∈ N.
Given f ∈ Int{n}(S, D) and k ∈ {0, . . . , n} we have:
∀(a1, . . . , ak+1) ∈ Sk+1, Φk(f )(a1, . . . , ak+1) ∈ D.
(*)
Since Φk(f ) is a symmetric polynomial in k+1 variables, for all permutations σ ∈ Sk+1 we have Φk(f )(a1, . . . , ak+1) =
Φk(f )(aσ(1), . . . , aσ(k+1)). Hence, we may disregard the order of the components of the chosen (k + 1)-tuple.
If we consider a multi-set Ω of cardinality k + 1 formed by elements of S, we may define Φk(f )(Ω) as the
value of Φk(f ) at one of the (k + 1)-tuple associated to Ω. Thus we choose an ordering of Ω and, by above,
the value Φk(f )(Ω) does not depend on the chosen ordering. Notice that Ω is not necessarily a sub-multi-set
of S. We only require that the underlying set of Ω is contained in S. For example, if S = {1, 2, 3} and
k = 1, we have {1, 1}, {1, 3} and {2, 2} as possible choices for Ω.
We may rephrase the above property (*) by saying that for all multi-sets Ω of cardinality k + 1 such
that the underlying set Ω′ is contained in S, we have Φk(f )(Ω) ∈ D.
Notation. We fix now the notation for the rest of this section.
- p(X) is a monic non-constant polynomial in D[X] of degree n.
- Ωp = {α1, . . . , αn} is the multi-set of roots of p(X) in K (the αi's are integral over D).
- F = K(α1, . . . , αn) the splitting field of p(X).
- DF the integral closure of D in F .
Given f ∈ F [X], whenever we expand f ∈ F [X] as in (2.1) in terms of the roots Ωp of p(X), we implicitly
assume that an order of Ωp has been fixed (so we choose one of the n! associated n-tuples). Changing the
order of Ωp will give a different expansion.
We need now the following preliminary lemma: the divided differences of a polynomial p(X) are zero
when they are evaluated at the roots of the polynomial p(X) itself.
Lemma 2.4. For every sub-multi-set Ω of Ωp of cardinality k + 1, k < n − 1, we have Φk(p)(Ω) = 0, and
Φn−1(p)(Ωp) = 1. Equivalently, we have:
Φk(p)(α1, . . . , αk+1) =(cid:26) 0,
1,
if
if
0 ≤ k < n
k = n
for any possible choice of an ordering for Ωp.
Proof. We fix an ordering for Ωp. We consider the Newton expansion of p(X) over F with respect to Ωp up to
the order n (p(X) is split over F ). The coefficients of this expansion are exactly {Φk(p)(α1, . . . , αk+1)}0≤k≤n,
where for k = n we have the leading coefficient of p(X) which is 1. Since p(X) is divisible by itself, all the
other coefficients in this expansion are zero. Obviously, the result does not depend on the chosen ordering
for Ωp.
Lemma 2.5. Let f ∈ K[X] and let r ∈ K[X] be the unique remainder in the division of f (X) by p(X) in
K[X]. If r 6= 0, let m < n be the degree of r(X). Then over F [X] we have
r(X) = f (α1) + Φ1(f )(α1, α2) · (X − α1) + . . . + Φm(f )(α1, . . . , αm+1)
m
Yi=1
(X − αi)
(2.2)
which is the Newton expansion of r(X) with respect to Ωp = {α1, . . . , αn}.
7
Proof. If f (X) = q(X)p(X) + r(X), by linearity of the divided difference operator, we have Φk(f ) =
Φk(r) + Φk(p · q), for all k ∈ N. Moreover, by the so-called Leibniz rule for divided differences (see for
example [18]), we have Φk(p · q) = Pi=0,...,k Φi(p)Φk−i(q) (we omit the variables). By Lemma 2.4, for
0 ≤ k < n, we get that
Φk(f )(α1, . . . , αk+1) = Φk(r)(α1, . . . , αk+1).
(2.3)
Notice that, for k = m the above value is the leading coefficient of r(X), and for m < k < n it is zero.
Because of the last formula, r(X) has the desired expansion over F [X].
By means of Lemma 2.4 and Lemma 2.5 we give a new proof of [16, Proposition 4.1], which says that
a polynomial f (X) of K[X] is in D(p) if and only if the divided differences of f (X) up to the order n − 1
are integral on every sub-multi-set of the multi-set Ωp of the roots of p(X).
Proposition 2.6. Let D be an integrally closed domain with quotient field K. Let f ∈ K[X] and p ∈ D[X]
monic of degree n. Let Ωp = {α1, . . . , αn} be the multi-set of roots of p(X) in a splitting field F over K.
Then the following are equivalent:
i) f ∈ D(p).
ii) for all 0 ≤ k < n, Φk(f )(α1, . . . , αk+1) ∈ D[α1, . . . , αk+1].
iii) for all 0 ≤ k < n, Φk(f )(α1, . . . , αk+1) ∈ DF .
Proof. If i) holds, let f (X) = r(X) + p(X)q(X), for some q ∈ K[X], r ∈ D[X], deg(r) < n or r = 0. In par-
ticular, the divided differences of r(X) are polynomials with coefficients in D. By (2.3), Φk(f )(α1, . . . , αk+1) =
Φk(r)(α1, . . . , αk+1) ∈ D[α1, . . . , αk+1], for all the relevant k's. Hence, i)⇒ ii).
Obviously ii)⇒ iii), since the roots of p(X) are integral over D, so that D[α1, . . . , αk+1] ⊆ DF .
Suppose now that iii) holds. We have to prove that the remainder r(X) of the Euclidean division
in K[X] of f (X) by p(X) is in D[X]. Let m < n be the degree of r(X). Consider the Newton ex-
pansion of r(X) with respect to Ωp over F [X] as in Lemma 2.5 (see (2.2)). By assumption, the coef-
ficients {Φk(f )(α1, . . . , αk+1)}k=0,...,m of this expansion are in DF . The leading coefficient of r(X) is
equal to Φm(f )(α1, . . . , αm+1), so that it is in DF ∩ K = D (we use here the assumption that D is inte-
grally closed). The coefficient cm−1 of the term X m−1 of r(X) is Φm−1(f )(α1, . . . , αm) ± (Pi=1,...,m αi) ·
Φm(f )(α1, . . . , αm+1) which is in DF , so cm−1 is in K ∩ DF = D. If we continue in this way we prove that
r(X) is in D[X], which gives i).
Remark 2.7. If we choose another ordering on the multi-set Ωp of roots of p(X) we have other conditions
of integrality on the values of the divided differences of a polynomial f ∈ D(p) at the vectors of elements
in Ωp. Since condition i) of Proposition 2.6 does not depend on the order we choose on Ωp, the above
conditions are also equivalent to this one:
ii') for all 0 ≤ k < n, and for every sub-multi-set Ω of Ωp of cardinality k + 1, Φk(f )(Ω) ∈ D[Ω],
that is, Φk(f ) is integral-valued on Ω: Φk(f )(Ω) ∈ DF (see also [16, Proposition 4.1 & Remark 4.1]).
Note that, if p ∈ D[X] is a monic polynomial of degree n which is split over D, that is, p(X) =
Qn
i=1(X − ai), ai ∈ D, then condition i) and ii) are equivalent without the assumption that D is integrally
closed (this follows immediately from the formula (2.2)).
for all
0 ≤ k < n, Φk(f )(a1, . . . , ak+1) ∈ D. We have thus in this case found again the result of [7, Proposition
11] (see also [16, Lemma 2.2 & Remark 2.1]).
In particular, condition ii) becomes:
Now we give the link between the ring of integer-valued polynomials whose divided differences are also
integer-valued introduced by Bhargava and the polynomial pullbacks D(p) we are working with.
We observe first that, if p ∈ D[X] is a monic polynomial of degree n which is split over D (i.e.: the set
of roots Ωp is contained in D), then Int{n−1}(Ωp, DF ) may be strictly contained in D(p).
8
Example 2.8. Let n = 2, Ω = {1, 3} ⊂ Z and p(X) = (X − 1)(X − 3). Let f (X) = p(X)/3 ∈ Z(p). We
have that Φ1(f )(1, 1) = −2/3, so that f /∈ Int{1}(Ω, Z).
Indeed, by Proposition 2.6, given any f ∈ Z(p), Φ1(f ) is integer-valued over {(1, 3), (3, 1)} ( Ω2.
We need to introduce another notation before the next theorem.
Notation. Let Ω be a subset of D and let n be a positive integer. We denote by Pn(Ω) the set of monic
polynomials q(X) over D of degree n whose set of roots is contained in Ω (so, in particular, they are split
over D).
Theorem 2.9. Let Ω ⊆ D and n ∈ N. Then
Int{n−1}(Ω, D) = \q∈Pn(Ω)
D(q).
Proof. (⊆). Let f ∈ Int{n−1}(Ω, D) and let q ∈ Pn(Ω). Since for all 0 ≤ k < n we have Φk(f )(Ωk+1) ⊂ D,
then for each sub-multi-set {a1, . . . , ak+1} of Ωq of cardinality k + 1 we have Φk(f )(a1, . . . , ak+1) ∈ D. Then
by Proposition 2.6 (see also Remark 2.7), we have that f ∈ D(q).
(⊇). Let f ∈ D(q), for all q ∈ Pn(Ω). Let k ∈ {0, . . . , n− 1} and let (a1, . . . , ak+1) ∈ Ωk+1. We consider
a polynomial q ∈ Pn(Ω) such that the multi-set {a1, . . . , ak+1} is a sub-multi-set of the multi-set of roots Ωq
(that is, Qk+1
i=1 (X − ai) divides q(X)). Then by Proposition 2.6, condition ii), Φk(f )(a1, . . . , ak+1) ∈ D (see
also Remark 2.7, condition ii')). Since (a1, . . . , ak+1) was chosen arbitrarily, f (X) is in Int{n−1}(Ω, D).
In the Example 2.8 above, we have that f (X) = q(X)
is a polynomial in P2(Ω) = {(X − 1)(X − 3), (X − 1)2, (X − 3)2}.
3 − 2
3 (X − 1) is not in Z(q), where q(X) = (X − 1)2
Remark 2.10. By [16, Lemma 5.1], given a monic polynomial p ∈ D[X] of degree n which is split over
D, the pullback ring D(p) is equal to IntK (T p
n (D) is the set of n × n triangular
matrices with characteristic polynomial equal to p(X). In particular, we have this representation for the
ring of integer-valued polynomials over the algebra of n × n triangular matrices over D:
n (D), Mn(D)), where T p
IntK (Tn(D)) = \p∈P s
n(D)
D(p)
(2.4)
where P s
n(D) is the set of monic polynomials over D of degree n which are split over D; as we mentioned
in the introduction, a similar result holds for IntK(Mn(D)), see [16]. We note that this gives a positive
answer to [17, Question 31] for the algebra Tn(D). Similarly, given any subset P of P s
n(D), the intersection
of the pullbacks D(p) as p(X) ranges through P is the ring of polynomials which are integer-valued over
the set of triangular matrices whose characteristic polynomial belongs to P. By Theorem 2.9, this ring is
equal to Int{n−1}(Ω, D), where Ω ⊆ D is the set of roots of the polynomials in P.
In the case Ω = D, [7, Theorem 16] proves that Int{n−1}(D) = IntK (Tn(D)), which by (2.4) is also
equal to the intersection of the pullbacks D(p), as p(X) ranges through P s
n(D). Therefore, Theorem 2.9
generalizes this result to any subset Ω of D.
3
Integral closure of polynomial pullbacks
Remark 3.1. Let Ω ⊂ K be a finite set. Let F = K(Ω) and let DF be the integral closure of D in F .
By [6, Proposition IV.4.1], Int(Ω, DF ) is integrally closed. Hence, IntK (Ω, DF ) = Int(Ω, DF ) ∩ K[X] is
integrally closed, too. The same remark was used in [17, Proposition 7].
9
Lemma 3.2. Let D be an integrally closed domain. Let p ∈ D[X] be a non-constant polynomial of degree
n and Ωp ⊂ K the multi-set of its roots. Let f ∈ K[X] be integral-valued over Ωp, that is, f ∈ IntK (Ωp, D).
Then the polynomial
is in D[X]. Moreover, P (f (X)) is divisible by p(X) in K[X].
P (X) = Pf,p(X) + Yα∈Ωp
(X − f (α))
Proof. Notice that P (X) has degree n, because the product is over the elements of the multi-set Ωp. We
set g(X) +
(X − α), where lc(p) is the leading coefficient of p(X). The polynomial g(X) is
in K[X] and is monic.
p(X)
lc(p) = Qα∈Ωp
Let M ∈ Mn(K) be a matrix with characteristic polynomial equal to g(X) (e.g., the companion matrix
of g(X)). The multi-set of eigenvalues of M over K is exactly Ωp. Notice that f (M ) is in Mn(K), so its
characteristic polynomial is in K[X]. By [3, Chap. VII, Proposition 10] (considering everything over K)
the characteristic polynomial of f (M ) is precisely P (X). In particular, the set of eigenvalues of f (M ) is
f (Ωp) = {f (α) α ∈ Ωp}, which, by assumption on f (X), is contained in D. Hence, the coefficients of
P (X) are integral over D (being the elementary symmetric functions of the roots), and since D is integrally
closed they are in D.
For the last statement, notice that for each α ∈ Ωp, X − α divides f (X)− f (α) over F = K(Ωp). Hence,
(f (X) − f (α)) over F . Since both polynomials are in
p(X) = Qα∈Ωp
K[X], one divides the other over K, as we wanted.
(X − α) divides P (f (X)) = Qα∈Ωp
We prove now Theorem 1.3 of the Introduction. For the sake of the reader we repeat here the statement.
Theorem 3.3. Let p ∈ D[X] be a monic non-constant polynomial and let Ωp ⊂ K be the multi-set of its
roots. Then the integral closure of D(p) is IntK (Ωp, D).
Notice that by definition IntK (Ωp, D) = IntK(Ω′p, D), where Ω′p is the underlying set of Ωp, the set of
distinct roots of p(X).
Proof. Remember that IntK (Ωp, D) is integrally closed by the Remark at the beginning of this section. If
D′ is the integral closure of D in its quotient field K, then D(p) ⊆ D′(p) is an integral ring extension,
because D[X] ⊆ D′[X] is. Since D(p) ⊆ D′(p) ⊆ IntK(Ωp, D) (because p(X) is monic, so Ωp is contained
in D), without loss of generality we may assume that D is integrally closed (that is, D = D′). To prove
the statement, it suffices to prove that D(p) ⊆ IntK (Ωp, D) is an integral ring extension.
Let f ∈ IntK(Ωp, D) and consider P (X) defined as in Lemma 3.2. Then P (X) is a monic polynomial
in D[X] such that P (f (X)) is divisible by p(X) over K. Hence, P (f (X)) is in D(p), and this gives a monic
integral equation for f (X) over the pullback ring D(p).
We prove now that the ring of polynomials in K[X] whose divided differences of order up to n are
integer-valued over a finite subset Ω of D has integral closure equal to the ring of polynomials which are
integer-valued over Ω.
Corollary 3.4. Let D be an integrally closed domain. Let Ω ⊂ D be a finite set and let n ∈ N. Then the
integral closure of Int{n}(Ω, D) is Int(Ω, D).
Proof. As in the proof of Theorem 3.3, it is sufficient to show that any f ∈ Int(Ω, D) satisfies a monic
equation over the ring Int{n}(Ω, D).
By Theorem 2.9, Int{n}(Ω, D) is equal to the intersection of the pullbacks D(p), as p(X) ranges through
the finite family Pn+1(Ω) of monic polynomials over D of degree n + 1 whose set of roots is contained in
10
Ω. We consider the subset P of Pn+1(Ω) of those polynomials of the form q(X) = (X − a)n+1, for a ∈ Ω.
For each of them we consider the polynomial Pf,q ∈ D[X] as defined in Lemma 3.2. Therefore
Q(X) + Yq∈P
Pf,q(X)
is a monic polynomial in D[X] such that Q(f (X)) is in p(X)K[X] for each p ∈ Pn+1(Ω).
In fact, let
If a ∈ Ω is a root of p(X) of multiplicity e ≤ deg(p) = n + 1, then (X − a)e divides
p ∈ Pn+1(Ω).
(f (X) − f (a))n+1 over K. Notice that the latter is a factor of Q(f (X)). Since this holds for every root of
p(X), then p(X) divides Q(f (X)) over K, that is, Q(f (X)) ∈ pK[X] ⊂ D(p). Since this holds for every
p ∈ Pn+1(Ω), this concludes the proof of the Corollary.
Remark 3.5. If Ω ⊆ D is an infinite set and D has finite residue rings (that is, D/dD is a finite ring for
every non-zero d ∈ D), reasoning as in [17] by means of the pullback representation of Int{n}(Ω, D) given
by Theorem 2.9, the same result of Corollary 3.4 holds. For Ω = D, the result was given in [17, Corollary
17], where it is proved that the integral closure of IntK (Tn+1(D)) is Int(D). Note that, by [7, Theorem 16],
the former ring is equal to Int{n}(D) (see Remark 2.10).
4 Prufer rings of integral-valued polynomials
The next lemma, though easy, is a crucial step to establish when IntK (Ω, D) is a Prufer domain, for a finite
set Ω of integral elements over D.
Lemma 4.1. Let p ∈ D[X] be a monic non-constant polynomial and K ⊆ F be an algebraic extension. Let
DF the integral closure of D in F . Then D(p) ⊆ DF (p) is an integral ring extension.
Proof. We use the well-known fact that the integral closure of D[X] in F [X] is DF [X] ([2, Proposition 13,
Chapt. V]). Hence, given f (X) = r(X)+p(X)q(X) ∈ DF (p), for some r ∈ DF [X] (r = 0 or deg(r) < deg(p))
and q ∈ F [X], the polynomial r(X) is integral over D[X], so in particular it is also integral also over D(p).
We show now that h(X) = p(X)q(X) ∈ p(X) · F [X] is integral over D(p).
It is easy to see that, if Ψq(T, X) is the minimal polynomial of q(X) over K[X], then the minimal
polynomial of h(X) over K[X] is given by Ψh(T, X) = pn · Ψq( T
p , X), which is a monic polynomial in T
over D. Notice that the coefficients of Ψh(T, X)− T n are in p(X)· K[X], so that Ψh(T, X) ∈ D(p)[T ]. This
proves our assertion.
We prove now Theorem 1.4 of the Introduction.
Theorem 4.2. Assume that D is integrally closed and let Ω be a finite subset of D. Then IntK (Ω, D) is
Prufer if and only if D is Prufer.
Proof. Given f (X) in IntK(Ω, D) and α ∈ Ω, f (X) is integral-valued over all the conjugates of α over K
(see Proposition 1.10). Hence, without loss of generality, we can assume that Ω is equal to the set of roots
Ωp of a monic polynomial p(X) over D (more precisely, p(X) is the product of all the minimal polynomials
of the elements of Ω, without repetitions).
Let F = K(Ωp) be the splitting field of p(X) over D and let DF be the integral closure of D in F . By
assumption, Ωp ⊂ DF . Remember that IntK (Ωp, D) = IntK (Ωp, DF ) (see the remarks after Proposition
1.10). By the result of McQuillan ([13, Corollary 7]), Int(Ωp, DF ) is a Prufer domain if and only if DF is a
11
Prufer domain. Since D is integrally closed, by [10, Theorem 22.3 & 22.4] D is Prufer if and only if DF is
Prufer. We have the following diagram:
DF (p)
/ Int(Ωp, DF )
/ F [X]
D(p)
/ IntK (Ωp, DF )
/ K[X]
By Theorem 3.3, D(p) ⊆ IntK (Ωp, DF ) and DF (p) ⊆ Int(Ωp, DF ) are integral ring extensions. Hence, by
Lemma 4.1, Int(Ωp, DF ) is integral over IntK(Ωp, DF ). Moreover, since the former ring is integrally closed,
it is the integral closure of the latter ring in F [X]. Finally, we have these equivalences:
D Prufer ⇔ DF Prufer ⇔ Int(Ωp, DF ) Prufer ⇔ IntK(Ωp, DF ) Prufer
where the last equivalence follows again by [10, Theorem 22.3 & 22.4] (IntK (Ωp, DF ) is integrally closed by
Remark 3.1).
As we recalled in the introduction, the intersection of the polynomial pullbacks D(p) arises in many
different contexts, especially those concerning rings of integer-valued polynomials over algebras. In section
2 we saw that the ring of integer-valued polynomials whose divided differences are also integer-valued can
be represented as an intersection of such pullbacks. We now investigate more deeply how these pullbacks
intersect with each other. As a corollary, we obtain a criterion for a pullback D(p) to be integrally closed.
At the beginning of Section 1.3 we recalled that a monic irreducible polynomial over an integrally closed
domain D is still irreducible over the quotient field K. Moreover, a monic polynomial p ∈ D[X] can be
uniquely factored into monic irreducible polynomials over D (see [12]; this is a sort of Gauss' Lemma for
monic polynomials over an integrally closed domain). Therefore, given a monic polynomial p(X) in D[X],
we have p(X) =Qi qi(X), where qi(X) are powers of monic irreducible polynomials in D[X]. In particular,
the qi(X)'s are pairwise coprime in K[X] (but they may not be coprime over D, see below). A polynomial
p(X) is square-free exactly when each qi(X) is irreducible. Notice that p(X)K[X] is an ideal of each
pullback D(qi), for all i. In particular, it is an ideal of the intersection of the rings D(qi).
The next proposition is a generalization of Lemma 1.6. Recall that two ideals I, J of a commutative
ring R are coprime if I + J = R (see [2, Chapt. 2, p. 53]). For this statement we do not require D to be
integrally closed. Given q1, q2 ∈ D[X], we simply say that q1(X) and q2(X) are coprime (over D) if the
corresponding principal ideals q1(X)D[X] and q2(X)D[X] are coprime.
Proposition 4.3. Let p ∈ D[X] be a monic polynomial. Let p(X) = Qi qi(X) be a factorization into
monic polynomials over D which are pairwise coprime when they are considered over K. Then
Ti D(qi)
p(X)K[X] ∼=Yi
D[X]
qi(X)D[X]
.
Moreover, D(p) =Ti D(qi) if and only if {qi(X)}i are pairwise coprime over D.
Note that two polynomials q1, q2 ∈ D[X] may be coprime over K without being coprime over D: for
example, q1(X) = X and q2(X) = X − 2 over Z. However, under this condition, it is easy to verify that
q1(X)D[X] ∩ q2(X)D[X] = q1(X)q2(X)D[X].
Proof. It is sufficient to notice that Ti D(qi) is the pullback of
qi(X)K[X] ∼=
p(X)K[X]
K[X]
K[X]
Yi
D[X]
qi(X)D[X] ⊂Yi
12
/
/
O
O
/
O
O
/
O
O
with respect to the canonical residue mapping
that is:
Indeed, by definition we have
π : K[X] ։
K[X]
p(X)K[X]
,
D[X]
qi(X)D[X]
π−1(Yi
D(qi).
) =\i
π−1 Yi
D[X]
qi(X)D[X]! = {f ∈ K[X] f
(mod qi(X)K[X]) ∈
D[X]
qi(X)D[X]
,∀i}.
Since each qi(X) is monic, by Lemma 1.6 this is equivalent to the fact that the remainder of the division of
f (X) by qi(X) is in D[X], that is, f (X) is in D(qi), hence the statement regarding the isomorphism. We
have then the following pullback diagram:
D(p)
D[X]
p(X)D[X]
/ Ti D(qi)
D[X]
qi(X)D[X]
/ Qi
where the vertical arrows are the quotient map modulo the common ideal p(X)K[X]. Note that the bottom
horizontal arrow is injective by the remark above before the proof. Then D(p) = Ti D(qi) if and only if
D[X]/p(X)D[X] and Qi D[X]/qi(X)D[X] are isomorphic. By the converse of the Chinese Remainder
Theorem (see [2, Chapt. 2, §1, Proposition 5]) this holds if and only if the principal ideals qi(X)D[X] are
pairwise coprime.
Recall that, given two polynomials p1, p2 ∈ D[X], the principal ideals pi(X)D[X], i = 1, 2, are coprime
if and only if the resultant Res(p1, p2) is a unit of D if and only if p1, p2 have no common root modulo any
maximal ideal M ⊂ D. Notice that p1(X), p2(X) are coprime in K[X] if and only if Res(p1, p2) 6= 0.
The next proposition is a generalization of Remark 1.9: given a monic non-constant square-free polyno-
mial p(X) in D[X], we determine the quotient ring of IntK(Ωp, D) modulo the principal ideal p(X)K[X].
Note that in the case Ωp ⊂ D, we have IntK (Ωp, D) = Int(Ωp, D) and we are in the case already treated
(essentially by McQuillan).
IntK(Ωp, D)
p(X)K[X] ∼=
DKi
Proposition 4.4. Let p ∈ D[X] be a monic non-constant polynomial which is square-free, say p(X) =
Qk
i=1 pi(X), where pi(X), for i = 1, . . . , k, are monic, distinct and irreducible polynomials over D. Then
k
Yi=1
where DKi is the integral closure of D in the field Ki ∼= K[X]
Proof. For each i = 1, . . . , k, we set Ki + K[X]/(pi(X)) ∼= K[αi], which is a finite field extension of K, where
αi is a (fixed) root of pi(X). Let also DKi be the integral closure of D in Ki, for i = 1, . . . , k. By assumption
on the pi(X)'s, D[X]/(pi(X)D[X]) ∼= D[αi] ⊂ K[αi]. Note that IntK(Ωp, D) = IntK ({α1, . . . , αk}, D): if
f ∈ K[X] is integral-valued on αi then it is integral-valued on every conjugate root of α of αi, that is on
the set of roots Ωpi (see also Proposition 1.10).
As we remarked in the introduction, the rings IntK (Ωp, D) ⊂ K[X] have the ideal p(X)K[X] in common,
p(X)K[X] . The
so that IntK(Ωp, D) is a pullback with respect to the canonical residue map π : K[X] ։
pi(X)K[X] , for each i = 1, . . . , k.
K[X]
13
/
/
polynomial ring K[X] is mapped to K[X]/(p(X)) ∼=Qk
so that a polynomial f ∈ K[X] is mapped to (f (α1), . . . , f (αk)).
definition,
In the same way as in Proposition 4.3 we have just to prove that IntK(Ωp, D) = π−1(Qk
i=1 K[αi] by the map which sends X to (α1, . . . , αk),
i=1 DKi ). By
π−1 k
Yi=1
DKi! = {f ∈ K[X] f (αi) ∈ DKi ,∀i = 1, . . . , k}
so that a polynomial f (X) is in this ring if and only if it is integral-valued on every αi, that is, f ∈
IntK ({α1, . . . , αk}, D).
An equivalent statement of Proposition 4.4 is the following:
let Ω be a finite subset of D and let
p ∈ D[X] be the product of the minimal polynomials p1(X), . . . , pk(X) of the elements of Ω, without
repetitions. Then the quotient of IntK(Ω, D) modulo p(X)K[X] is isomorphic to Qk
i=1 DKi , where DKi is
as in the statement of Proposition 4.4. We notice that a proof of Theorem 4.2 follows also in another way
by [5, Theorem 4.3], due to Proposition 4.4.
Theorem 4.5. Let p ∈ D[X] be a monic non-constant polynomial. Suppose that p(X) =Qk=1,...,k pi(X)ei
is the unique factorization of p(X) into powers of monic irreducible polynomials in D[X], ei ≥ 1. Then
D(p) is integrally closed if and only if the following conditions are satisfied:
i) p(X) is squarefree (i.e.: ei = 1 for all i).
ii) for each i = 1, . . . , k, D[X]/(pi(X)) ∼= DKi , where the latter is the integral closure of D in the field
Ki ∼= K[X]/(pi(X)).
iii) Res(pi, pj ) ∈ D∗ for each i 6= j.
If D is a Prufer domain, D(p) is integrally closed if and only if D(p) is a Prufer domain, and in that case
D(p) = IntK (Ωp, D).
greater than 1. Let q(X) =Qk
Proof. Suppose that D(p) is integrally closed. If p(X) is not squarefree, then some exponent ei is strictly
i=1 pi(X) be the square-free part of p(X). By assumption, q(X) 6= p(X) and
q(X) divides p(X). So by Lemma 1.7, D(p) ( D(q). Since q(X) has the same set of roots of p(X), D(q) is
contained in IntK (Ωp, DF ). Hence, D(p) cannot be equal to IntK (Ωp, DF ) which is in contradiction with
Theorem 3.3. Then p(X) is square-free.
By Propositions 4.3 and 4.4 (we retain the same notation of those Propositions) we have the following
i=1 IntK (Ωpi , D) = IntK(Ωp, D)), where the vertical
diagram of pullbacks (notice that Ωp =Sk
lines are the reduction map modulo p(X)K[X]:
/ IntK (Ωp, D)
/ K[X]
(4.1)
D(p)
D[X]
p(X)D[X]
i=1 Ωpi andTk
/ Ti D(pi)
/ Qi D[αi]
/ Qi K[αi]
/ Qi DKi
i=1 D(pi) and Tk
i=1 D(pi) = IntK (Ωp, D).
Obviously, D(p) is integrally closed if and only if D(p) =Tk
Since D(p) = Tk
the above diagram, Tk
i=1 D(pi) = IntK (Ωp, D) if and only if
condition ii).
i=1 D(pi), by Proposition 4.3 this condition is equivalent to condition iii). Looking at
pi(X)D[X] = DKi for all i = 1, . . . , k, which is
D[X]
Conversely, suppose conditions i), ii) and iii) hold. Then looking at the above pullback diagram again,
we have that D(p) is equal to IntK(Ωp, D), hence, by Theorem 3.3, D(p) is integrally closed.
Suppose now D is a Prufer domain. If D(p) = IntK (Ωp, D) then D(p) is a Prufer domain by Theorem
4.2. Conversely, if D(p) is Prufer then it is integrally closed. The very last assertion follows at once by
Theorem 3.3.
14
/
/
/
/
/
/
In the next examples we show that the theorem does not hold if we remove one of the conditions.
Example 4.6. Let p1(X) = X 2 + 1, p2(X) = X 2 − 2 ∈ Z[X] and p(X) = p1(X)p2(X). The resultant
Res(p1, p2) is equal to 9. Moreover, K1 = Q(i) ⊃ OK1 = Z[X]/(p1(X)) and K2 = Q(√2) ⊃ OK2 =
Z[X]/(p2(X)).
Then Z(p1) ∩ Z(p2) = IntQ(Ωp, Z) (see the proof of Theorem 4.5 and the diagram (4.1)) but Z(p1 · p2) =
Z(p) ( Z(p1) ∩ Z(p2) (Proposition 4.3). Notice that Z(p1), Z(p2) are integrally closed: Z(pi) = IntQ(Ωpi , Z)
for i = 1, 2, but Z(p) is not integrally closed. Here, condition iii) of Theorem 4.5 is not satisfied.
Example 4.7. p1(X) = X 2 − 5, p2(X) = X 2 − 6. The resultant of p1(X) and p2(X) is equal to 1. Then
K1 = Q(√5) ⊃ OK1 = Z[ 1+√5
Then Z(p1) ∩ Z(p2) ( IntQ(Ωp, Z) but Z(p) = Z(p1) ∩ Z(p2). Hence, Z(p) is not integrally closed, because
condition ii) of Theorem 4.5 is not satisfied.
] ) Z[√5] ∼= Z[X]/(p1(X)) and K2 = Q(√6) ⊃ OK2 = Z[X]/(p2(X)).
2
Corollary 4.8. Let p ∈ D[X] be a monic polynomial over D which is split in D. Then D(p) is integrally
closed if and only if the discriminant of p(X) is a unit in D.
Notice that if the latter condition holds, in particular p(X) is separable, that is, it has no repeated roots.
We denote by ∆(p) the discriminant of p(X).
Proof. Let Ωp = {α1, . . . , αn} ⊂ D be the multi-set of roots of p(X). By Theorem 3.3 the integral closure
of D(p) is Int(Ωp, D) =Ti D(X − αi).
It is enough to observe that ∆(p) =Qi<j (αi − αj)2 and that if pi(X) = X − αi, for i = 1, . . . , n, then
Res(pi, pj) = ±(αj − αi). Then by Theorem 4.5 we conclude.
Remark 4.9. The statement is false if we do not assume that p(X) is split over D. For example, let
D = Z and p(X) = X 2 − 2. Then Z(p) is integrally closed by Proposition 1.10 (see also Theorem 4.5),
because Z[√2] is the ring of integers OK of K = Q(√2), so Z(p) = IntQ({±√2}, OK ). However, ∆(p) = 8.
This implies that the pullback OK (p) ⊂ K[X] is not integrally closed: the polynomial f (X) = X−√2
is in
2√2
Int({±√2}, OK ) and not in OK (p), and by Theorem 3.3 f (X) is integral over OK (p).
D[X]
D[X]
K[X]
K[X]
Remark 4.10. We can prove Theorem 3.3 by means of a pullback diagram argument. By Lemma 1.6 and
Proposition 4.4, looking at the diagram (4.1), by [8, Lemma 1.1.4 (8)], IntK (Ωp, D) is the integral closure
p(X)K[X] is
p(X)D[X] (see the proof of [14, Theorem 10.15]). Hence, by [10, Proposition 2.7],
of D(p), since Qi DKi is the integral closure of
p(X)K[X] . Indeed, it is known that
the total quotient ring of
p(X)K[X] is the total quotient ring of every subring containing
p(X)D[X] in
K[X]
[2, Proposition 9, Chapt. V], Qk
over D, it follows that Qk
p(X)D[X] , and in particular of Qi D[αi]. By
i=1 DKi is the integral closure of D in Qk
i=1 D[αi] in Qk
i=1 DKi is the integral closure of Qk
i=1 K[αi]. Since each αi is integral
i=1 K[αi].
D[X]
5 General case of a finite set of integral elements over D
We show in this section how to apply the previous results to the more general setting mentioned in the
introduction, namely when the finite set of integral elements over D is not necessarily contained in an
algebraic extension of K. We recall the assumptions we mentioned in the introduction.
For simplicity, we assume that D is integrally closed. Let A be a D-algebra, possibly non-commutative
and with zero-divisors, which is finitely generated and torsion-free as a D-module. Note that every element
a of A is integral over D. Let µa(X) be the minimal polynomial of a over D, which is not necessarily
irreducible. To be precise, µa(X) is the monic generator of the ideal of K[X] of those polynomials which
are zero on a. Since D is supposed integrally closed and a is integral over D, µa ∈ D[X] (so that µa(X)
15
is also the generator of the ideal of D[X] of those polynomials which are zero at a). For short, we set
Ωa = Ωµa , the set of roots in D of µa(X). We may evaluate polynomials of K[X] at the elements of A in
the extended K-algebra B = A ⊗D K (note that, by assumption, K and A embed into B). Given a subset
S of A, we consider the ring of integer-valued polynomials over S:
IntK (S, A) = {f ∈ K[X] f (S) ⊂ A}.
For S = A, we have the ring IntK (A, A) = IntK (A) of integer-valued polynomials over A. For more details
about this setting we refer to [17]. As in [17], we consider polynomials over K whose evaluation at the
elements of S are not necessarily in A, but are still integral over D. For this reason, we call them integral-
valued polynomials over S, since they preserve the integrality of the elements of S. We retain the notation
introduced in [17].
Definition 5.1. Let K[S] be the K-subalgebra of B = A ⊗D K generated by K and the elements of S.
Let also S′ be the subset of K[S] of those elements which are integral over D. We set
which we call integral-valued polynomials over S.
IntK (S, S′) = {f ∈ K[X] f (S) ⊂ S′}
Note that in general S′ does not form a ring, if A is non-commutative (even if S is a ring; for example,
consider the case A = Mn(D)). Nevertheless, IntK(S, S′) does form a ring by the argument given in [17,
Proposition 6]: in order to show that IntK(S, S′) is closed under addition and multiplication, it is sufficient
to consider what happens point-wise and use the fact that for each s ∈ S, K[s] is a commutative K-algebra.
We note that the ring IntK (S, S′) is equal to the ring of polynomials in K[X] such that f (s) (which a priori
is in K[s] ⊆ B) is integral over D for each s ∈ S. Clearly, IntK(S, A) ⊆ IntK (S, S′), because every element
of A is integral over D. The key result which links the ring of integral-valued polynomials IntK (S, S′) to a
previous ring of integral-valued polynomials over a subset Ω of D is the following.
Theorem 5.2. [17, Theorem 9] Let S be a subset of A and set ΩS =Ss∈S Ωs ⊂ D. Then
IntK(S, S′) = IntK (ΩS, D).
Proof. For the sake of the reader we give the proof. Since 1 ∈ D ⊂ B, we may embed B into the
endomorphism ring EndK (B), via the map given by multiplication on the left by b ∈ B. In particular, A
is a sub-D-algebra of EndK(B), and for s ∈ S, Ωs is the set of eigenvalues (in K) of s considered as a
K-endomorphism of B. Since A is finitely generated as a D-module, by [3, Chapt. VII, §. 5, Proposition
10], for any polynomial f ∈ K[X], f (Ωs) = {f (α) α ∈ Ωs} is the set of eigenvalues of f (s), so that in our
notation f (Ωs) = Ωf (s). Given f ∈ K[X] and s ∈ S, f (s) is integral over D if and only if the elements of
Ωf (s) = f (Ωs) are integral over D (because D is integrally closed). The claim is then proved.
We are ready to give the proof of the last main result of the paper, see Corollary 1.5 of the Introduction.
IntK (S, A) is IntK (ΩS, D).
Corollary 5.3. Let S be a finite subset of A and ΩS = Ss∈S Ωs ⊂ D. Then the integral closure of
Proof. Let p(X) =Qs∈S µs(X) ∈ D[X]. By above, we have the following inclusions:
D(p) ⊆ IntK(S, A) ⊆ IntK (S, S′) = IntK(ΩS, D)
and the claim follows by Theorem 3.3.
16
Note that by Theorem 4.2, the ring IntK (S, A) has Prufer integral closure if and only if D is Prufer.
Acknowledgements.
We wish to thank the referee for his/her valuable suggestions which improved the quality of the paper.
The author was supported by the Austrian Science Foundation (FWF), Project Number P23245-N18,
INDAM and also University of Roma 3. The paper was prepared during a visit at the Department of
Mathematics and Physics of the University of Roma 3. The author wishes to thanks Francesca Tartarone
for the hospitality.
References
[1] M. Bhargava. On P-orderings, Rings of Integer-Valued Polynomials and Ultrametric Analysis, Journal
of the Amer. Math. Soc. 22 (4) (2009), 963-993.
[2] N. Bourbaki. Commutative Algebra. Hermann, Paris; Addison-Wesley Publishing Co., Reading, Mass.
(1972)
[3] N. Bourbaki: Algebra. Translated from the French by P. M. Cohn and J. Howie. Elements of Mathe-
matics (Berlin). Springer-Verlag, Berlin, 1990. vii+461 pp.
[4] J. G. Boynton, S. Sather-Wagstaff. Regular pullbacks. Progress in commutative algebra 2, 145-169,
Walter de Gruyter, Berlin, 2012.
[5] J. G. Boynton. Pullbacks of arithmetical rings. Comm. Algebra 35 (2007), no. 9, 2671-2684.
[6] J.-P. Cahen, J.-L. Chabert, Integer-valued polynomials, Amer. Math. Soc. Surveys and Monographs,
48, Providence, 1997.
[7] S. Evrard, Y. Fares, K. Johnson. Integer valued polynomials on lower triangular integer matrices,
Monats. fur Math. 170 (2013), no. 2, 147-160.
[8] M. Fontana, J. A. Huckaba, I. J. Papick. Prufer domains. Monographs and Textbooks in Pure and
Applied Mathematics, 203. Marcel Dekker, Inc., New York, 1997.
[9] S. Gabelli, E. Houston. Ideal Theory in Pullbacks, Non-Noetherian commutative ring theory, 199-227,
Math. Appl., 520, Kluwer Acad. Publ., Dordrecht, 2000.
[10] R. Gilmer. Multiplicative ideal theory, Corrected reprint of the 1972 edition. Queen's University,
Kingston, ON, 1992.
[11] M. Griffin. Prufer rings with zero divisors. J. Reine Angew. Math. 239/240 1969 55-67.
[12] S. McAdam. Unique factorization of monic polynomials. Comm. Algebra 29 (2001), no. 10, 4341-4343.
[13] D. L. McQuillan. Rings of integer-valued polynomials determined by finite sets. Proc. Roy. Irish Acad.
Sect. A 85 (1985), no. 2, 177-184.
[14] M. Nagata. Local rings. Interscience Tracts in Pure and Applied Mathematics, No. 13 Interscience
Publishers a division of John Wiley & Sons New York-London 1962.
[15] G. Peruginelli. Integral-valued polynomials over sets of algebraic integers of bounded degree, J. Number
Theory 137 (2014) 241-255.
[16] G. Peruginelli. Integer-valued polynomials over matrices and divided differences, Monatsh. Math. 173
(2014), no. 4, 559-571.
17
[17] G. Peruginelli, N. Werner. Integral closure of rings of integer-valued polynomials on algebras, in "Com-
mutative Algebra: Recent Advances in Commutative Rings, Integer-Valued Polynomials, and Polyno-
mial Functions", M. Fontana, S. Frisch and S. Glaz (editors), Springer 2014, pp. 293-305.
[18] J. F. Steffensen. Note on divided differences, Danske Vid. Selsk. Math.-Fys. Medd. 17,(1939). no. 3.
[19] B. L. van der Waerden. Algebra. Vol. I. Springer-Verlag, New York, 1991.
18
|
1904.05968 | 2 | 1904 | 2019-09-23T14:19:19 | Every quasitrivial n-ary semigroup is reducible to a semigroup | [
"math.RA",
"cs.DM",
"math.CO"
] | We show that every quasitrivial n-ary semigroup is reducible to a binary semigroup, and we provide necessary and sufficient conditions for such a reduction to be unique. These results are then refined in the case of symmetric n-ary semigroups. We also explicitly determine the sizes of these classes when the semigroups are defined on finite sets. As a byproduct of these enumerations, we obtain several new integer sequences. | math.RA | math |
EVERY QUASITRIVIAL n-ARY SEMIGROUP IS REDUCIBLE
TO A SEMIGROUP
MIGUEL COUCEIRO AND JIMMY DEVILLET
Abstract. We show that every quasitrivial n-ary semigroup is reducible to
a binary semigroup, and we provide necessary and sufficient conditions for
such a reduction to be unique. These results are then refined in the case of
symmetric n-ary semigroups. We also explicitly determine the sizes of these
classes when the semigroups are defined on finite sets. As a byproduct of these
enumerations, we obtain several new integer sequences.
1. Introduction
Let X be a nonempty set and let n ≥ 2 be an integer.
In this paper we are
interested in n-ary operations F ∶ X n → X that are associative, namely that satisfy
the following system of identities
F (x1, . . . , xi−1, F (xi, . . . , xi+n−1), xi+n, . . . , x2n−1)
= F (x1, . . . , xi, F (xi+1, . . . , xi+n), xi+n+1, . . . , x2n−1)
for all x1, . . . , x2n−1 ∈ X and all 1 ≤ i ≤ n − 1. This generalisation of associativity
was originally proposed by Dornte [4] and studied by Post [10] in the framework of
n-ary groups and their reductions. An operation F ∶ X n → X is said to be reducible
to a binary operation (resp. ternary operation) if it can be written as a composition
of a binary (resp. ternary) associative operation (see Definition 2.1).
Recently, the study of reducibility criteria for n-ary semigroups , that is, a set
X endowed with an associative n-ary operation, gained an increasing interest (see,
e.g., [1,5 -- 7]). In particular, Dudek and Mukhin [5] provided necessary and sufficient
conditions under which an n-ary associative operation is reducible to a binary
associative operation. Indeed, they proved (see [5, Theorem 1]) that an associative
operation F ∶ X n → X is reducible to an associative binary operation if and only if
one can adjoin a neutral element e to X for F , that is, there is an n-ary associative
extension F ∶ (X ⋃{e})n → X ⋃{e} of F such that e is a neutral element for F and
F X n = F . In this case, a binary reduction Ge of F can be defined by
Ge(x, y) = F (x, e, . . . , e
´¹¹¹¹¹¹¹¹¹¸¹¹¹¹¹¹¹¹¹¶
n−2 times
, y) x, y ∈ X.
Date: September 24, 2019.
1991 Mathematics Subject Classification. 05A15, 20N15; 16B99, 20M14.
Key words and phrases. quasitrivial polyadic semigroup, reducibility, symmetry, enumeration.
The second author is supported by the Luxembourg National Research Fund under the project
PRIDE 15/10949314/GSM.
1
2
M. COUCEIRO AND J. DEVILLET
Recently, Ackerman [1] also investigated reducibility criteria for n-ary associative
operations that are quasitrivial, that is, operations that preserve all unary relations:
for every x1, . . . , xn ∈ X, F (x1, . . . , xn) ∈ {x1, . . . , xn}.
Remark 1.1. Quasitrivial operations are also called conservative operations [11].
This property has been extensively used in the classification of constraint satisfac-
tion problems into complexity classes (see, e.g, [2] and the references therein).
The following result reassembles Corollaries 3.14 and 3.15, and Theorem 3.18
of [1].
Theorem 1.2. Let F ∶ X n → X be an associative and quasitrivial operation.
(a) If n is even, then F is reducible to an associative and quasitrivial binary
operation G∶ X 2 → X.
(b) If n is odd, then F is reducible to an associative and quasitrivial ternary
operation H∶ X 3 → X.
(c) If n = 3 and F is not reducible to an associative binary operation G∶ X 2 → X,
then there exist a1, a2 ∈ X with a1 ≠ a2 such that
● F (X∖{a1,a2})3 is reducible to an associative binary operation.
● a1 and a2 are neutral elements for F .
From Theorem 1.2 (c) it would follow that if an associative and quasitrivial
operation F ∶ X n → X is not reducible to an associative binary operation G∶ X 2 → X,
then n is odd and there exist distinct a1, a2 ∈ X that are neutral elements for F .
However, Theorem 1.2 (c) supposes the existence of a ternary associative and
quasitrivial operation H∶ X 3 → X that is not reducible to an associative binary
operation, and Ackerman did not provide any example of such an operation.
In this paper we show that there is no associative and quasitrivial n-ary opera-
tion that is not reducible to an associative binary operation (Corollary 2.3). Hence,
for any associative and quasitrivial operation F ∶ X n → X one can adjoin a neutral
element to X. Now this raises the question of whether such a binary reduction is
unique and whether it is quasitrivial. We show that both of these properties are
equivalent to the existence of at most one neutral element for the n-ary associative
and quasitrivial operation (Theorem 3.9). Since an n-ary associative and quasitriv-
ial operation has at most one neutral element when n is even or at most two when
n is odd (Proposition 3.7), in the case when X is finite, we also provide several
enumeration results (Propositions 3.15 and 3.17) that explicitly determine the sizes
of the corresponding classes of associative and quasitrival n-ary operations in terms
of the size of the underlying set X. As a by-product, these enumeration results
led to several integer sequences that were previously unknown in the Sloane's On-
Line Encyclopedia of Integer Sequences (OEIS, see [12]). These results are further
refined in the case of symmetric operations (Theorem 4.6).
2. Motivating results
In this section we recall some basic definitions and present some motivating
In particular, we show that every associative and quasitrivial operation
results.
F ∶ X n → X is reducible to an associative binary operation (Corollary 2.3).
Throughout this paper let k ≥ 1 and x ∈ X. We use the shorthand notation
[k] ={1, . . . , k} and n ⋅ x = x, . . . , x (n times), and we denote the set of all constant
X ={(n ⋅ y) y ∈ X}. Also, we denote the size of any set S by
S .
n-tuples over X by ∆n
QUASITRIVIAL n-ARY SEMIGROUPS
3
Recall that a neutral element for F ∶ X n → X is an element eF ∈ X such that
F((i − 1) ⋅ eF , x,(n − i) ⋅ eF) = x
for all x ∈ X and all i ∈ [n]. When the meaning is clear from the context, we may
drop the index F and denote a neutral element for F by e.
Definition 2.1 (see [1, 5]). Let G∶ X 2 → X, and H∶ X 3 → X be associative opera-
tions.
(1) An operation F ∶ X n → X is said to be reducible to G if F(x1, . . . , xn) =
Gn−1(x1, . . . , xn) for all x1, . . . , xn ∈ X, where G1 = G and
Gm(x1, . . . , xm+1) = Gm−1(x1, . . . , xm−1, G(xm, xm+1)),
for each integer 2 ≤ m ≤ n−1. In this case, G is said to be a binary reduction
of F .
(2) Similarly, F is said to be reducible to H if n is odd and F(x1, . . . , xn) =
Hn−3(x1, . . . , xn) for all x1, . . . , xn ∈ X, where H0 = H and
Hm(x1, . . . , xm+3) = Hm−2(x1, . . . , xm, H(xm+1, xm+2, xm+3)),
for each even integer 2 ≤ m ≤ n − 3. In this case, H is said to be a ternary
reduction of F .
As we will see, every associative and quasitrivial operation F ∶ X n → X is re-
ducible to an associative binary operation. To show this, we will make use of the
follwing auxiliary result.
Lemma 2.2 (see [5, Lemma 1]). If F ∶ X n → X is associative and has a neutral
element e ∈ X, then F is reducible to the associative operation Ge∶ X 2 → X defined
by
(2.1)
Ge(x, y) = F(x,(n − 2) ⋅ e, y), for every x, y ∈ X.
The following corollary follows from Theorem 1.2 and Lemma 2.2.
Corollary 2.3. Every associative and quasitrivial operation F ∶ X n → X is reducible
to an associative binary operation.
Theorem 1.2(c) states that a ternary associative and quasitrivial operation F ∶ X 3 →
X must have two neutral elements, whenever it is not reducible to a binary oper-
ation. In particular, we can show that two distinct elements a1, a2 ∈ X are neutral
elements for F if and only if they are neutral elements for the restriction F {a1,a2}3
of F to {a1, a2}3. Indeed, the condition is obviously necessary, while its sufficiency
follows from the Lemma 2.4 below.
Lemma 2.4. Let H∶ X 3 → X be an associative and quasitrivial operation.
H(a1, a1, x) = H(x, a1, a1) = x = H(x, a2, a2) = H(a2, a2, x),
(a) If a1, a2 ∈ X are two distinct neutral elements for H {a1,a2}3 , then
(b) If a1, a2 ∈ X are two distinct neutral elements for H {a1,a2}3 , then both a1
(a) Let x ∈ X. We only show that H(a1, a1, x) = x, since the other
{a1, a2}. So let x ∈ X ∖ {a1, a2} and, for a contradiction, suppose that
equalities can be shown similarly. Clearly, the equality holds when x ∈
and a2 are neutral elements for H.
x ∈ X.
Proof.
4
M. COUCEIRO AND J. DEVILLET
H(a1, a1, x) = a1. By the associativity and quasitriviality of H, we then
have
a1 = H(a1, a1, x) = H(a1, H(a1, a2, a2), x)
= H(H(a1, a1, a2), a2, x) = H(a2, a2, x) ∈{a2, x},
which contradicts the fact that a1, a2 and x are pairwise distinct.
(b) Suppose to the contrary that a1 is not a neutral element for H (the other
case can be dealt with similarly). By Lemma 2.4(a) we have that H(a1, a1, y) =
H(y, a1, a1) = y for all y ∈ X. By assumption, there exists x ∈ X ∖{a1, a2}
such that H(a1, x, a1) = a1. We have two cases to consider.
● If H(a2, x, a2) = x, then by Lemma 2.4(a) we have that
H(x, a2, a1) = H(H(x, a1, a1), a2, a1) = H(x, a1, H(a1, a2, a1))
= H(x, a1, a2) = H(H(a1, a1, x), a1, a2)
= H(a1, H(a1, x, a1), a2) = H(a1, a1, a2) = a2.
Also, by Lemma 2.4(a) we have that
which contradicts the fact that x ≠ a1.
x = H(x, a1, a1) = H(H(a2, x, a2), a1, a1)
= H(a2, H(x, a2, a1), a1) = H(a2, a2, a1) = a1,
● If H(a2, x, a2) = a2, then by Lemma 2.4(a) we have that
H(x, x, a2) = H(x, H(a2, a2, x), a2)
= H(x, a2, H(a2, x, a2)) = H(x, a2, a2) = x,
and
H(a1, x, x) = H(a1, H(x, a1, a1), x)
= H(H(a1, x, a1), a1, x) = H(a1, a1, x) = x.
By Lemma 2.4(a) we also have that
x = H(x, a2, a2) = H(H(a1, x, x), a2, a2)
= H(a1, H(x, x, a2), a2) = H(a1, x, a2)
= H(a1, H(x, a1, a1), a2) = H(H(a1, x, a1), a1, a2)
= H(a1, a1, a2) = a2,
which contradicts the fact that x ≠ a2.
(cid:3)
We now present some geometric considerations of quasitrivial operations. The
preimage of an element x ∈ X under an operation F ∶ X n → X is denoted by F −1[x].
When X is finite, namely X = [k], we also define the preimage sequence of F as
the nondecreasing k-element sequence of the numbers F −1[x] , x ∈[k]. We denote
this sequence by F −1 .
Recall that the kernel of an operation F ∶[k]n → [k] is the equivalence relation
Ker(F) = {{x, y} F(x) = F(y)}. The contour plot of F ∶[k]n → [k] is the undi-
rected graph CF = ([k]n, E), where E is the non-reflexive part of Ker(F), that is,
E = {{x, y} x ≠ y and F(x) = F(y)}. We say that two tuples x, y ∈ [k]n are
F -connected (or simply connected ) if {x, y} ∈ Ker(F).
F(n ⋅ x) = x for all x ∈ X.
By an idempotent operation, we mean an operation F ∶ X n → X that satisfies
QUASITRIVIAL n-ARY SEMIGROUPS
5
Lemma 2.5. An operation F ∶[k]n →[k] is quasitrivial if and only if it is idempo-
tent and each (x1, ..., xn) ∈[k]n ∖ ∆n
Proof. Clearly, F is quasitrivial if and only if it is idempotent and for any(x1, ..., xn) ∈
[k]n ∖ ∆n
[k] there exists i ∈{1, ..., n} such that F(x1, ..., xn) = xi = F(n ⋅ xi).
[k] is connected to some (n ⋅ x) ∈ ∆n
[k].
(cid:3)
In the sequel we shall make use of the following two lemmas.
Remark 2.8. A neutral element need not be unique when n ≥ 3 (for instance,
one component equal to x. By Lemma 2.6, we conclude that there are exactly
(cid:3)
Lemma 2.6. For each x ∈ [k], the number of tuples (x1, . . . , xn) ∈ [k]n with at
least one component equal to x is given by kn −(k − 1)n.
Proof. Let x ∈[k]. The set of tuples in [k]n with at least one component equal to x
is the set [k]n ∖([k] ∖{x})n, and its cardinality is kn −(k − 1)n since ([k] ∖{x})n ⊆
[k]n.
Lemma 2.7. Let F ∶[k]n →[k] be a quasitrivial operation. Then, for each x ∈[k],
we have F −1[x] ≤ kn −(k − 1)n.
Proof. Let x ∈ [k]. Since F ∶[k]n → [k] is quasitrivial, it follows from Lemma 2.5
that the point (n ⋅ x) is at most connected to all (x1, ..., xn) ∈ [k]n with at least
kn −(k − 1)n such points.
Recall that an element z ∈ X is said to be an annihilator for F if F(x1, ..., xn) = z,
whenever (x1, ..., xn) ∈ X n has at least one component equal to z.
F(x1, x2, x3) ≡ x1 + x2 + x3 (mod 2) on X = Z2). However, if an annihilator exists,
Proposition 2.9. Let F ∶[k]n → [k] be a quasitrivial operation and let z ∈ [k].
Then z is an annihilator if and only if F −1[z] = kn −(k − 1)n.
Proof. (Necessity) If z is an annihilator, then we know that F(i ⋅ z, xi+1, ..., xn) = z
for all i ∈[n], all xi+1, ..., xn ∈ [k] and all permutations of (i ⋅ z, xi+1, ..., xn). Thus,
by Lemma 2.6 (n ⋅ z) is connected to kn −(k − 1)n points. Finally, using Lemma
2.7 we get F −1[z] = kn −(k − 1)n.
(Sufficiency) If F −1[z] = kn −(k − 1)n, then by Lemmas 2.5 and 2.6 we have
that (n ⋅ z) is connected to the kn −(k − 1)n points (x1, ..., xn) ∈ [k]n containing
at least one component equal to z. Thus, we have F(i ⋅ z, xi+1, ..., xn) = z for all
i ∈ [n], all xi+1, ..., xn ∈ [k] and all permutations of (i ⋅ z, xi+1, ..., xn), which shows
Remark 2.10. By Proposition 2.9, if F ∶[k]n → [k] is quasitrivial, then each ele-
ment x such that F −1[x] = kn −(k − 1)n is unique.
that z is an annihilator.
then it is unique.
(cid:3)
(cid:3)
3. Criteria for unique reductions and some enumeration results
In this section we show that an associative and quasitrivial operation F ∶ X n → X
is uniquely reducible to an associative and quasitrivial binary operation if and only
if F has at most one neutral element (Theorem 3.9). We also enumerate the class of
associative and quasitrivial n-ary operations, which leads to a previously unknown
sequence in the OEIS [12] (Proposition 3.17). Let us first recall a useful result
from [6].
6
M. COUCEIRO AND J. DEVILLET
Lemma 3.1 (see [6, Proposition 3.5]). Assume that the operation F ∶ X n → X is
associative and reducible to associative binary operations G∶ X 2 → X and G′∶ X 2 →
X. If G and G′ are idempotent or have the same neutral element, then G = G′.
From Lemma 3.1, we immediately get a necessary and sufficient condition that
guarantees unique reductions for associative operation that have a neutral element.
Corollary 3.2. Let F ∶ X n → X be an associative operation that is reducible to as-
sociative binary operations G∶ X 2 → X and G′∶ X 2 → X that have neutral elements.
Then, G = G′ if and only if G and G′ have the same neutral element.
Using Lemma 2.2, Corollary 3.2, and observing that
(i) a binary associative operation has at most one neutral element,
(ii) the neutral element of a binary reduction G∶ X 2 → X of an associative
operation F ∶ X n → X is also a neutral element for F , and
(iii) if e is a neutral element for an associative operation F ∶ X n → X and G∶ X 2 →
X is a reduction of F , then Gn−2((n−1) ⋅e) (see Definition 2.1) is the neutral
element for G,
we can generalise Corollary 3.2 as follows.
Proposition 3.3. Let F ∶ X n → X be an associative operation, and let EF be the
set of its neutral elements and RF of its binary reductions. If EF ≠ ∅, then for any
G ∈ RF , there exists e ∈ EF such that G = Ge. Moreover, the mapping σ∶ EF → RF
defined by σ(e) = Ge is a bijection. In particular, e is the unique neutral element
for F if and only if Ge is the unique binary reduction of F .
As we will see in Proposition 3.7, the size of EF , and thus of RF , is at most 2
whenever F is quasitrivial.
X ,
Let Q2
1(X) denote the class of associative and quasitrivial operations G∶ X 2 → X
1(X) denote the class of associative
that have exactly one neutral element, and let A2
operations G∶ X 2 → X that have exactly one neutral element eG ∈ X and that satisfy
the following conditions:
Note that Q2
Q2
1(X) = X X 2
1(X) = A2
1(X). Actually, we have that G ∈ Q2
● G(x, x) ∈{eG, x} for all x ∈ X,
● G(x, y) ∈{x, y} for all (x, y) ∈ X 2 ∖ ∆2
● If there exists x ∈ X ∖{eG} such that G(x, x) = eG, then x is unique and
we have G(x, y) = G(y, x) = y for all y ∈ X ∖{x, eG}.
when X = 1. Also, it is not difficult to see that
1(X) and
1(X) ⊆ A2
G−1[e] = 1, where e is the neutral element for G. A characterization of the class
1(X) can be found
Recall that two groupoids (X, G) and (Y, G′) are said to be isomorphic, and
we denote it by (X, G) ≃ (Y, G′), if there exists a bijection φ∶ X → Y such that
φ(G(x, y)) = G′(φ(x), φ(y)) for every x, y ∈ X. The following straightforward
1(X) gives rise to a
proposition states, in particular, that any G ∈ A2
semigroup which has a unique 2-element subsemigroup isomorphic to the additive
semigroup on Z2.
of associative and quasitrivial binary operations as well as Q2
in [3, Theorem 2.1, Fact 2.4].
1(X) if and only if G ∈ A2
1(X) ∖ Q2
Proposition 3.4. Let G∶ X 2 → X be an operation. Then G ∈ A2
if and only if there exists a unique pair (x, y) ∈ X 2 ∖ ∆2
X such that the following
1(X) ∖ Q2
1(X)
conditions hold
QUASITRIVIAL n-ARY SEMIGROUPS
7
(a) ({x, y}, G {x,y}2) ≃(Z2, +),
(b) G (X∖{x,y})2 is associative and quasitrivial, and
(c) every z ∈ X ∖{x, y} is an annihilator for G {x,y,z}2 .
Proposition 3.5. Let F ∶ X n → X be an associative and quasitrivial operation.
Suppose that e ∈ X is a neutral element for F .
(a) If n is even, then F is reducible to an operation G ∈ Q2
(b) If n is odd, then F is reducible to the operation Ge ∈ A2
1(X).
1(X).
Proof. (a) By Theorem 1.2(a) we have that F is reducible to an associative and
quasitrivial binary operation G∶ X 2 → X. Finally, we observe that Gn−2((n − 1) ⋅ e)
is the neutral element for G.
associativity of Ge, we have that
(b) By Lemma 2.2 we have that F is reducible to an associative operation
Ge∶ X 2 → X of the form (2.1) and that e is also a neutral element for Ge. Since
If
F is quasitrivial, it follows from (2.1) that Ge(x, x) ∈ {x, e} for all x ∈ X.
X = 2, then the proof is complete. So suppose that X > 2 and let us show that
Ge(x, y) ∈ {x, y} for all (x, y) ∈ X 2 ∖ ∆2
X . Since e is a neutral element for Ge, we
have that Ge(x, e) = Ge(e, x) = x for all x ∈ X ∖{e}. So suppose to the contrary
that there are distinct x, y ∈ X ∖{e} such that Ge(x, y)~∈{x, y}. As Ge is a reduc-
tion of F and F is quasitrivial, we must have Ge(x, y) = e. But then, using the
y = Ge(e, y) = Ge(Ge(x, y), y) = Ge(x, Ge(y, y)) ∈{Ge(x, y), Ge(x, e)} ={e, x},
Now, suppose that there exists x ∈ X ∖ {e} such that Ge(x, x) = e and let
y ∈ X ∖{x, e}. Since
we must have Ge(x, y) = y. Similarly, we can show that Ge(y, x) = y.
to the contrary that there exists x′ ∈ X ∖{x, e} such that Ge(x′, x′) = e. Since x, x′
y = Ge(e, y) = Ge(Ge(x, x), y) = Ge(x, Ge(x, y)),
To complete the proof, we only need to show that such an x is unique. Suppose
which contradicts the fact that x, y and e are pairwise distinct.
and e are pairwise distinct and
and
x′ = Ge(e, x′) = Ge(Ge(x, x), x′) = Ge(x, Ge(x, x′)),
x = Ge(x, e) = Ge(x, Ge(x′, x′)) = Ge(Ge(x, x′), x′),
we must have x = Ge(x, x′) = x′, which yields the desired contradiction.
We observe that the associative operation F ∶ Zn
2 → Z2 defined by
(cid:3)
F(x1, . . . , xn) ≡
n
Q
i=1
xi (mod 2),
x1, . . . , xn ∈ Z2,
has 2 neutral elements, namely 0 and 1, when n is odd. Moreover, it is quasitrivial if
and only if n is odd. This also illustrates the fact that an associative and quasitrivial
n-ary operation that has 2 neutral elements does not necessarily have a quasitrivial
reduction. Indeed, when n is odd, G(x1, x2) ≡ x1 + x2 (mod 2) and G′(x1, x2) ≡
x1 + x2 + 1 (mod 2) on X = Z2 are two distinct reductions of F but neither is
1(X), then it is quasitrivial. The following proposition provides a necessary
G ∈ Q2
and sufficient condition for F to be quasitrivial when G ∈ A2
Clearly, if an associative operation F ∶ X n → X is reducible to an operation
quasitrivial.
1(X) ∖ Q2
1(X).
8
M. COUCEIRO AND J. DEVILLET
1(X) ∖ Q2
Proposition 3.6. Let F ∶ X n → X be an associative operation. Suppose that F is
reducible to an operation G ∈ A2
if n is odd.
1(X). Then, F is quasitrivial if and only
Proof. To show that the condition is necessary, let x ∈ X ∖{e} such that G(x, x) = e.
If n is even, then F(n ⋅ x) = G n
1(X),
clearly have that F(n ⋅ x) = x for all x ∈ X such that G(x, x) = x.
Let y ∈ X ∖{e} such that G(y, y) = e. Since n is odd, we have that
⋅ G(y, y) = G(y, e) = y.
⋅ G(x, x)) = e, contradicting quasitriviality.
1(X)∖Q2
and thus we only need to show that F is idempotent. Since F is reducible to G, we
So let us prove that the condition is also sufficient. Note that G ∈ A2
F(n ⋅ y) = G y, G n−1
2 −1 n − 1
2
2 −1( n
2
Hence, F is idempotent and the proof is now complete.
(cid:3)
It is not difficult to see that the operation F ∶ Zn
n−1 → Zn−1 defined by
F(x1, . . . , xn) ≡
n
Q
i=1
xi (mod (n − 1)),
x1, . . . , xn ∈ Zn−1,
is associative, idempotent, symmetric (that is, F(x1, . . . , xn) is invariant under any
permutation of x1, . . . , xn), and has n − 1 neutral elements. However, this number
is much smaller for quasitrivial operations.
Proposition 3.7. Let F ∶ X n → X be an associative and quasitrivial operation.
(a) If n is even, then F has at most one neutral element.
(b) If n is odd, then F has at most two neutral elements.
Proof. (a) By Theorem 1.2(a) we have that F is reducible to an associative and
quasitrivial binary operation G∶ X 2 → X. Suppose that e1, e2 ∈ X are two neutral
elements for F . Since G is quasitrivial we have
e2 = F((n − 1) ⋅ e1, e2) = G(Gn−2((n − 1) ⋅ e1), e2)
= G(e1, e2) = G(e1, Gn−2((n − 1) ⋅ e2)) = F(e1,(n − 1) ⋅ e2) = e1.
Hence, F has at most one neutral element.
(b) By Theorem 1.2(b) we have that F is reducible to an associative and quasi-
trivial ternary operation H∶ X 3 → X. For a contradiction, suppose that e1, e2, e3 ∈
X are three distinct neutral elements for F . Since H is quasitrivial, it is not difficult
to see that e1, e2, and e3 are neutral elements for H. Also, by Proposition 3.5(b)
we have that H is reducible to the operations Ge1 , Ge2 , Ge3 ∈ A2
we have
1(X). In particular,
Ge1(e2, e3) = Ge1(Ge1(e1, e2), e3)
= H(e1, e2, e3) = Ge2(Ge2(e1, e2), e3) = Ge2(e1, e3)
and
H(e1, e2, e3) = Ge3(e1, Ge3(e2, e3)) = Ge3(e1, e2).
Hence, H(e1, e2, e3) ∈ {e2, e3} ∩{e1, e3} ∩{e1, e2}, which shows that e1, e2, e3 are
not pairwise distinct, and thus yielding the desired contradiction.
(cid:3)
Corollary 3.8. Let F ∶ X n → X be an operation and let e1 and e2 be distinct
elements of X. Then F is associative, quasitrivial, and has exactly the two neutral
QUASITRIVIAL n-ARY SEMIGROUPS
9
elements e1 and e2 if and only if n is odd and F is reducible to exactly the two
operations Ge1 , Ge2 ∈ A2
1(X) ∖ Q2
1(X).
Proof. (Necessity) This follows from Propositions 3.3, 3.5, and 3.7 together with
the observation that Ge1(e2, e2) = e1 and Ge2(e1, e1) = e2.
(Sufficiency) This follows from Propositions 3.3 and 3.6.
(cid:3)
We can now state and prove the main result of this section.
Theorem 3.9. Let F ∶ X n → X be an associative and quasitrivial operation. The
following assertions are equivalent.
(i) Any binary reduction of F is idempotent.
(ii) Any binary reduction of F is quasitrivial.
(iii) F has at most one binary reduction.
(iv) F has at most one neutral element.
(v) F((n − 1) ⋅ x, y) = F(x,(n − 1) ⋅ y) for any x, y ∈ X.
Proof. The implications (i) ⇒((ii) and (v)) and (v) ⇒ (iv) are straightforward. By
Proposition 3.7 and Corollary 3.8 we also have the implications ((ii) or (iii)) ⇒ (iv).
Hence, to complete the proof, it suffices to show that (iv) ⇒ ((i) and (iii)). First,
we prove that (iv) ⇒ (i). We consider the two possible cases.
If F has a unique neutral element e, then by Proposition 3.3 G = Ge is the unique
reduction of F with neutral element e. For the sake of a contradiction, suppose
that G is not idempotent. By Proposition 3.5 we then have that n is odd and
G ∈ A2
1(X) ∖ Q2
So let x ∈ X ∖{e} such that G(x, x) ≠ x. Since G = Ge, we must have G(x, x) = e.
It is not difficult to see that F(y,(n − 1) ⋅ x) = y = F((n − 1) ⋅ x, y) for all y ∈ X.
Now, if there is i ∈{2, . . . , n − 1} such that
1(X).
F((i − 1) ⋅ x, e,(n − i) ⋅ x) = x,
then we have that i − 1 and n − i are both even or both odd (since n is odd), and
thus
x = F((i − 1) ⋅ x, e,(n − i) ⋅ x) ∈{G2(x, e, x), G2(e, e, e)} ={e},
which contradicts our assumption that x ≠ e. Hence, we have F((i − 1) ⋅ x, e,(n −
i) ⋅ x) = e for all i ∈{1, . . . , n}.
Now, if X = 2, then the proof is complete since e and x are both neutral elements
for F , which contradicts our assumption. So suppose that X > 2.
Since e is the unique neutral element for F , there exist y ∈ X ∖ {e, x} and
i ∈{2, . . . , n − 1} such that
F((i − 1) ⋅ x, y,(n − i) ⋅ x) = x.
Again by the fact that n is odd, i − 1 and n − i are both even or both odd, and thus
x = F((i − 1) ⋅ x, y,(n − i) ⋅ x) ∈{G2(x, y, x), G2(e, y, e)} ={G2(x, y, x), y}.
Since x ≠ y, we thus have that G2(x, y, x) = x. But then
e = G(x, x) = G(x, G2(x, y, x))
= G(G(x, x), G(y, x)) = G(e, G(y, x)) = G(y, x) ∈{x, y},
which contradicts our assumption that x, y, and e are pairwise distinct.
Now, suppose that F has no neutral element and that G is a reduction of F that
is not idempotent. So let x ∈ X such that G(x, x) ≠ x, and let y ∈ X ∖{x, G(x, x)}.
10
M. COUCEIRO AND J. DEVILLET
By the quasitriviality of F we have F((n − 1) ⋅ x, y) ∈ {x, y}. On the other hand,
by the quasitriviality (and hence idempotency) of F and the associativity of G we
have
F((n − 1) ⋅ x, y) = F(F(n ⋅ x),(n − 2) ⋅ x, y)
= G(Gn−2(Gn−1(n ⋅ x),(n − 2) ⋅ x), y)
= G(G2n−3((2n − 2) ⋅ x), y)
= G(Gn−2((n − 1) ⋅ G(x, x)), y)
= F((n − 1) ⋅ G(x, x), y) ∈{G(x, x), y}.
Since x, G(x, x), and y are pairwise distinct, it follows that F((n − 1) ⋅ x, y) = y,
which implies that G(Gn−2((n − 1) ⋅ x), y) = y. Similarly, we can show that
G(y, Gn−2((n − 1) ⋅ x)) = y.
Also, it is not difficult to see that
G(Gn−2((n − 1) ⋅ x), G(x, x)) = G(x, x) = G(G(x, x), Gn−2((n − 1) ⋅ x)).
Furthermore, since F is idempotent and reducible to G, we also have that
G(Gn−2((n − 1) ⋅ x), x) = x = G(x, Gn−2((n − 1) ⋅ x)).
Thus Gn−2((n − 1) ⋅ x) is a neutral element for G and therefore a neutral element
for F , which contradicts our assumption that F has no neutral element.
As both cases yield a contradiction, we conclude that G must be idempotent. The
implication (iv) ⇒ (iii) is an immediate consequence of the implication (iv) ⇒ (i)
together with Lemma 3.1. Thus, the proof of Theorem 3.9 is now complete.
(cid:3)
Remark 3.10. We observe that an alternative necessary and sufficient condition
for the quasitriviality of a binary reduction of an n-ary quasitrivial semigroup has
also been provided in [1, Corollary 3.16].
Theorem 3.9 together with Corollary 2.3 imply the following result.
Corollary 3.11. Let F ∶ X n → X be an operation. Then F is associative, qua-
sitrivial, and has at most one neutral element if and only if it is reducible to an
associative and quasitrivial operation G∶ X 2 → X.
In this case, G is defined by
G(x, y) = F(x,(n − 1) ⋅ y).
Recall that a weak ordering on X is a binary relation ≲ on X that is total and
transitive (see, e.g., [8] p. 14). We denote the symmetric part of ≲ by ∼. Also, a
weakly ordered set, an element a ∈ X is said to be maximal for ≲ if x ≲ a for all
total ordering on X is a weak ordering on X that is antisymmetric. If (X, ≲) is a
x ∈ X. We denote the set of maximal elements of X for ≲ by M≲(X).
Given a weak ordering ≲ on X, the n-ary maximum operation on X for ≲ is the
partial symmetric n-ary operation maxn
≲ defined on
X n ∖{(x1, . . . , xn) ∈ X n ∶ M≲({x1, . . . , xn}) ≥ 2}
≲(x1, . . . , xn) = xi where i ∈ [n] is such that xj ≲ xi for all j ∈ [n]. If ≲
by maxn
reduces to a total ordering, then clearly the operation maxn
≲ is defined everywhere
on X n. Also, the projection operations π1∶ X n → X and πn∶ X n → X are respectively
defined by π1(x1, . . . , xn) = x1 and πn(x1, . . . , xn) = xn for all x1, . . . , xn ∈ X.
QUASITRIVIAL n-ARY SEMIGROUPS
11
Corollary 3.11 together with [9, Theorem 1] and [3, Corollary 2.3] imply the
following characterization of the class of quasitrivial n-ary semigroups with at most
one neutral element.
Theorem 3.12. Let F ∶ X n → X be an operation. Then F is associative, quasitriv-
ial, and has at most one neutral element if and only if there exists a weak ordering
≲ on X and a binary reduction G∶ X 2 → X of F such that
G A×B = ⎧⎪⎪⎨⎪⎪⎩
π1 A×B or π2 A×B,
max2
≲ A×B ,
if A = B,
otherwise,
∀A, B ∈ X~ ∼.
(3.1)
(3.2)
Moreover, when X =[k], then the weak ordering ≲ is uniquely defined as follows:
x ≲ y ⇔ G−1[x] ≤ G−1[y] ,
x, y ∈[k].
Now, let us illustrate Theorem 3.12 for binary operations by means of their
contour plots. We can always represent the contour plot of any operation G∶[k]2 →
[k] by fixing a total ordering on [k]. In Figure 1 (left), we represent the contour
plot of an operation G∶ X 2 → X using the usual total ordering ≤ on X ={1, 2, 3, 4}.
To simplify the representation of the connected components, we omit edges that
can be obtained by transitivity. It is not difficult to see that G is quasitrivial. To
check whether G is associative, by Theorem 3.12, it suffices to show that G is of
the form (3.1) where the weak ordering ≲ is defined on X by (3.2). In Figure 1
(right) we represent the contour plot of G using the weak ordering ≲ on X defined
by (3.2). We observe that G is of the form (3.1) for ≲ and thus by Theorem 3.12 it
is associative.
✻
4
3
2
1
r
r
r
r
✄
r
r
r
✂
r
r
r
r
r
r
r
r
✁
r
✻
1
3
4
2
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
✲
✲
1
2
3
4
2
4
3
1
Figure 1. An associative and quasitrivial binary operation G on
X = {1, 2, 3, 4} whose values on (1, 1),(2, 2),(3, 3) and (4, 4) are
1, 2, 3 and 4, respectively.
1, . . . , x′
Let ≤ be a total ordering on X. An operation F ∶ X n → X is said to be ≤-
for which an operation G ∈ A2
the binary addition modulo 2.
associative binary operations G∶ X 2 → X are ≤-preserving for any total ordering on
n), whenever xi ≤ x′
preserving if F(x1, . . . , xn) ≤ F(x′
i for all i ∈ [n]. Some
X (e.g., G(x, y) = x for all x, y ∈ X). However, there is no total ordering ≤ on X
1(X) is ≤-preserving. A typical example is
1(X), then there is no total
Proposition 3.13. Suppose X ≥ 2. If G ∈ A2
Proof. Let e ∈ X be the neutral element for G and let x ∈ X ∖ {e} such that
G(x, x) = e. Suppose to the contrary that there exists a total ordering ≤ on X such
ordering ≤ on X that is preserved by G.
1(X) ∖ Q2
1(X) ∖ Q2
12
M. COUCEIRO AND J. DEVILLET
that G is ≤-preserving. If x < e, then e = G(x, x) ≤ G(x, e) = x, which contradicts
our assumption. The case x > e yields a similar contradiction.
(cid:3)
Remark 3.14. It is not difficult to see that any ≤-preserving operation F ∶ X n → X
has at most one neutral element. Therefore, by Corollary 2.3 and Theorem 3.9 we
conclude that any associative, quasitrivial, and ≤-preserving operation F ∶ X n → X
is reducible to an associative, quasitrivial, and ≤-preserving operation G∶ X 2 → X.
For a characterization of the class of associative, quasitrivial, and ≤-preserving
operations G∶ X 2 → X, see [3, Theorem 4.5].
We now provide several enumeration results that give the sizes of the classes of
associative and quasitrivial operations that were considered above when X = [k].
Recall that for any integers 0 ≤ ℓ ≤ k, the Stirling number of the second kind k
ℓ is
defined by
k
ℓ = 1
ℓ!
ℓ
Q
i=0(−1)ℓ−iℓ
i ik.
1(k) the cardinality of Q2
1([k]). By definition, we have a2
by q2
of A2
In [3] the authors solved several
enumeration problems concerning associative and quasitrivial binary operations.
For any integer k ≥ 0, let q2(k) (resp. qn(k)) denote the number of associative and
quasitrivial binary (resp. n-ary) operations on [k]. For any integer k ≥ 1, we denote
1(k) the cardinality
In particular, they computed q2(k) (see [3, Theorem 4.1]) as well as q2
1(k) (see
A292932(k) and A292933(k). The following result summarizes [3, Theorem 4.1]
[3, Proposition 4.2]). These sequences were also introduced in the OEIS [12] as
1([k]). Also, we denote by a2
1(1) = 1.
and [3, Proposition 4.2].
Proposition 3.15. For any integer k ≥ 0, we have the closed-form expression
q2(k) =
k
Q
i=0
2i
k−i
Q
ℓ=0(−1)ℓk
ℓk − ℓ
i (i + ℓ)! ,
k ≥ 0,
1([k]) ∖ Q2
Proposition 3.16. For any integer k ≥ 2, we have a2
Proof. We already have that Q2
an operation G ∈ A2
where q2(0) = q2(1) = 1. Moreover, for any integer k ≥ 1, we have q2
1)q2(k − 2).
1(k) = k q2(k−1).
1(k) = kq2(k − 1) + k(k −
1([k]) ⊆ A2
1([k]). Now, let us show how to construct
1([k]). There are k ways to choose the element x ∈[k]
such that G(x, x) = e and G(x, y) = G(y, x) = y for all y ∈ [k] ∖ {x, e}. Then
1([k] ∖ {x}),
we observe that the restriction of G to ([k] ∖ {x})2 belongs to Q2
1(k − 1) possible choices to construct this restriction. This shows
1(k) + kq2
1(k − 1). Finally, by Proposition 3.15 we conclude that
1(k) = kq2(k − 1) + k(k − 1)q2(k − 2).
0(k)) denote the number of associative and
2(k) denote the number
2(1) = 0. The following proposition provides
no neutral element) on [k]. Also, for any integer k ≥ 1, let qn
[k]. Clearly, qn(1) = qn
of associative and quasitrivial n-ary operations that have two neutral elements on
quasitrivial n-ary operations that have exactly one neutral element (resp. that have
so we have q2
that a2
a2
explicit forms of the latter sequences. Table 1 below provides the first few values of
all the previously considered sequences. In view of Corollary 3.8, we only consider
the case where n is odd for qn
1(1) = 1 and qn
1(k) (resp. qn
For any integer k ≥ 1 let qn
1(k) = q2
(cid:3)
2(k) and qn(k).
QUASITRIVIAL n-ARY SEMIGROUPS
13
Proposition 3.17. For any integer k ≥ 1 we have qn
q2(k) − q2
1(k). Also, for any integer k ≥ 2 we have
1(k) = q2
1(k) and qn
0(k) =
and
qn(k) =
if n is even
if n is odd.
qn
2(k) =
0
k
2q2(k − 2)
q2(k)
q2(k) +k
2q2(k − 2)
if n is even
if n is odd.
Proof. By Theorem 3.9 we have that the number of associative and quasitrivial
n-ary operations that have exactly one neutral element (resp. that have no neutral
element) on [k] is exactly the number of associative and quasitrivial binary opera-
tion on [k] that have a neutral element (resp. that have no neutral element). This
1(k)). Also, if n is even, then by Theorem
1.2(a) and Proposition 3.7(a) we conclude that qn(k) = q2(k) and qn
1(k) (resp. q2(k) − q2
Now, suppose that n is odd. By Corollary 3.8 and Propositions 3.15 and 3.16
number is given by q2
2(k) = 0.
we have that qn
that qn(k) = qn
2(k) = a2
0(k) + qn
1(k)−q2
1 (k)
2
1(k) + qn
=k
2q2(k − 2). Finally, by Proposition 3.7 we have
2(k) = q2(k) +k
2q2(k − 2).
(cid:3)
k
1
2
3
4
5
6
q2(k)
q2
1(k)
qn
0(k)
qn
2(k)
qn(k)
a2
1(k)
1
4
18
128
1 090
11 232
OEIS A292932 A292933 A308352 A308354 A308362 A308351
1
4
20
138
1 182
12 166
1
5
23
162
1 382
14 236
0
2
8
58
492
5 074
1
2
12
80
690
7 092
0
1
3
24
200
2 070
Table 1. First few values of q2(k), q2
1(k), qn
0(k), qn
2(k), qn(k) and a2
1(k)
4. Symmetric operations
In this section we refine our previous results to the subclass of associative and
quasitrivial operations that are symmetric, and present further enumeration results
accordingly.
We first recall and establish some auxiliary results.
Fact 4.1. Suppose that F ∶ X n → X is associative and surjective. If it is reducible
to an associative operation G∶ X 2 → X, then G is surjective.
Lemma 4.2 (see [6, Lemma 3.6]). Suppose that F ∶ X n → X is associative, sym-
metric, and reducible to an associative and surjective operation G∶ X 2 → X. Then
G is symmetric.
Proposition 4.3. If F ∶ X n → X is associative, quasitrivial, and symmetric, then
it is reducible to an associative, surjective, and symmetric operation G∶ X 2 → X.
Moreover, if X =[k], then F has a neutral element.
14
M. COUCEIRO AND J. DEVILLET
Proof. By Corollary 2.3, F is reducible to an associative operation G∶ X 2 → X. By
Fact 4.1 and Lemma 4.2, it follows that G is surjective and symmetric.
For the moreover part, we only have two cases to consider.
● If G is quasitrivial, then by [3, Theorem 3.3] it follows that G has a neutral
element, and thus F also has a neutral element.
● If G is not quasitrivial, then by Proposition 3.7 and Theorem 3.9 F has in
(cid:3)
fact two neutral elements.
Proposition 4.4 (see [1, Corollary 4.10]). An operation F ∶ X n → X is associative,
quasitrivial, symmetric, and reducible to an associative and quasitrivial operation
G∶ X 2 → X if and only if there exists a total ordering ⪯ on X such that F = maxn
⪯ .
Proposition 4.5. A quasitrivial operation F ∶[k]n →[k] is associative, symmetric,
and reducible to an associative and quasitrivial operation G∶[k]2 → [k] if and only
if F −1 =(1, 2n − 1, . . . , kn −(k − 1)n).
Proof. (Necessity) Since G is quasitrivial, it is surjective and hence by Lemma 4.2
it is symmetric. Thus, by Proposition 4.4 there exists a total ordering ⪯ on X
⪯ , which has an
⪯(x, y) for all x, y ∈ [k]. Hence F = maxn
such that G(x, y) = max2
annihilator, and the proof of the necessity then follows by Proposition 2.9.
(Sufficiency) We proceed by induction on k. The result clearly holds for k = 1.
Suppose that it holds for some k ≥ 1 and let us show that it still holds for k + 1.
Assume that F ∶[k + 1]n →[k + 1] is quasitrivial and that
F −1 =(1, 2n − 1, . . . ,(k + 1)n − kn).
Let ⪯ be the total ordering on [k + 1] defined by
F −1(x) ≤ F −1(y) ,
x ⪯ y if and only if
⪯
and let z = maxk+1
(1, . . . , k + 1). Clearly, F ′ = F ([k+1]∖{z})n is quasitrivial and
F ′−1 = (1, 2n − 1, . . . , kn −(k − 1)n). By induction hypothesis we have that F ′ =
⪯′ , where ⪯′ is the restriction of ⪯ to [k +1]∖{z}. By Proposition 2.9, F −1[z] =
(k + 1)n − kn and thus F = maxn
maxn
⪯ .
(cid:3)
We can now state and prove the main result of this section.
Theorem 4.6. Let F ∶ X n → X be an associative, quasitrivial, symmetric operation.
The following assertions are equivalent.
(i) F is reducible to an associative and quasitrivial operation G∶ X 2 → X.
(ii) There exists a total ordering ⪯ on X such that F is ⪯-preserving.
(iii) There exists a total ordering ⪯ on X such that F = maxn
⪯ .
Moreover, when X = [k], each of the assertions (i) − (iii) is equivalent to each of
the following assertions.
(iv) F has exactly one neutral element.
(v) F −1 =(1, 2n − 1, . . . , kn −(k − 1)n).
Furthermore, the total ordering ⪯ considered in assertions (ii) and (iii) is uniquely
defined as follows:
(4.1)
x ⪯ y
if and only if
G−1[x] ≤ G−1[y] ,
x, y ∈[k].
Moreover, there are k! operations satisfying any of the conditions (i) − (v).
QUASITRIVIAL n-ARY SEMIGROUPS
15
Proof. (i) ⇒ (iii). This follows from Proposition 4.4.
(iii) ⇒ (ii). Obvious.
(ii) ⇒ (i). By Corollary 2.3 we have that F is reducible to an associative opera-
tion G∶ X 2 → X. Suppose to the contrary that G is not quasitrivial. From Theorem
3.9 and Proposition 3.7, it then follows that F has two neutral elements e1, e2 ∈ X,
which contradicts Remark 3.14.
(i) ⇔ (v). This follows from Proposition 4.5.
(i) ⇒ (iv). This follows from Theorem 3.9 and Proposition 4.3.
(iv) ⇒ (i). This follows from Lemma 2.2 and Theorem 3.9.
The rest of the statement follows from [3, Theorem 3.3].
(cid:3)
X using the usual total ordering ≤ on X = {1, 2, 3, 4}.
Now, let us illustrate Theorem 4.6 for binary operations by means of their contour
plots. In Figure 2 (left), we represent the contour plot of an operation G∶ X 2 →
In Figure 2 (right) we
represent the contour plot of G using the total ordering ⪯ on X defined by (4.1).
We then observe that G = max2
⪯, which shows by Theorem 4.6 that G is associative,
quasitrivial, and symmetric.
✻
4
3
2
1
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
✻
1
4
2
3
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
✲
✲
1
2
3
4
3
2
4
1
Figure 2. An associative, quasitrivial, and symmetric binary op-
eration G on X = {1, 2, 3, 4} whose values on (1, 1),(2, 2),(3, 3)
and (4, 4) are 1, 2, 3 and 4, respectively.
Based on this example, we illustrate a simple test to check whether an operation
ordering and if F is the maximum operation for ≾.
We denote the class of associative, quasitrivial, symmetric operations G∶ X 2 → X
that have a neutral element e ∈ X by QS2
class of symmetric operations G∶ X 2 → X that belong to A2
to see that QS2
F ∶[k]n → [k] is associative, quasitrivial, symmetric, and has exactly one neutral
element. First, construct the unique weak ordering ≾ on [k] from the preimage
sequence F −1 , namely, x ≾ y if F −1[x] ≤ F −1[y] . Then, check if ≾ is a total
1(X) the
1(X). It is not difficult
1(X) if and only if G ∈ AS2
1(X)
and G−1[e] = 1, where e is the neutral element for G.
For each integer k ≥ 2, let qsn(k) denote the number of associative, quasitrivial,
and symmetric n-ary operations on[k]. Also, denote by as2
1(k) the size of AS2
1([k]).
1([k]) = k!. Also, it is easy
From Theorems 3.9 and 4.6 it follows that qs2(k) = QS2
1(2) = 4. The remaining terms of the sequence are given in the
1(X). Also, we denote by AS2
1(X) ⊆ AS2
1(X). In fact, G ∈ QS2
to check that as2
following proposition.
Proposition 4.7. For every integer k ≥ 3, as2
Proof. As observed QS2
AS2
1([k]) ⊆ AS2
1([k]) ∖ QS2
1([k]). So let us enumerate the operations in
1([k]). There are k ways to choose the element x ∈ [k] such that
1(k) = qs2(k) + kqs2(k − 1) = 2k!.
16
M. COUCEIRO AND J. DEVILLET
For any integer k ≥ 2 let qsn
such restrictions. Thus as2
then follows that as2
1(k) = k! + k(k − 1)! = 2k!.
G(x, x) = e and G(x, y) = G(y, x) = y for all y ∈[k]∖{x, e}. Moreover, the restriction
1([k] ∖{x}), and we have qs2(k − 1) possible
of G to ([k] ∖{x})2 belongs to QS2
1(k) = qs2(k) + kqs2(k − 1). By Theorems 3.9 and 4.6 it
1(k) denote the number of associative, quasitrivial,
and symmetric n-ary operations that have exactly one neutral element on [k].
2(k) denote the number of associative, quasitrivial, and symmetric n-
ary operations that have two neutral elements on [k].
2(k) =
2 , and qsn(k) = 3k!
Proposition 4.8. For each integer k ≥ 2, qsn
k!
1(k) = qs2(k) = k!. Moreover, qsn
Also, let qsn
2 .
(cid:3)
Proof. By Theorems 4.6 and 3.9 and Lemma 4.2 we have that the number of as-
sociative, quasitrivial, and symmetric n-ary operations that have exactly one neu-
tral element on [k] is exactly the number of associative, quasitrivial, and sym-
metric binary operations on [k]. By Theorems 3.9 and 4.6 this number is given
by qs2(k) = k!. Also, by Corollary 3.8, Proposition 4.7, and Theorems 3.9 and
qn(k) = qsn
2 and by Proposition 3.7 we have that
(cid:3)
2(k) = as2
2(k) = 3k!
1(k) + qsn
4.6, we have that qn
= k!
1(k)−qs2(k)
2
2 .
Remark 4.9. Recall that an operation F ∶ X n → X is said to be bisymmetric if
F(F(r1), . . . , F(rn)) = F(F(c1), . . . , F(cn))
for all n × n matrices [c1 ⋯ cn] = [r1 ⋯ rn]T ∈ X n×n. In [6, Corollary 4.9] it was
shown that associativity and bisymmetry are equivalent for operations F ∶ X n →
X that are quasitrivial and symmetric. Thus, we can replace associativity with
bisymmetry in Theorem 4.6.
5. Conclusion
In this paper we proved that any quasitrivial n-ary semigroup is reducible to
a semigroup. Furthermore, we showed that a quasitrivial n-ary semigroup is re-
ducible to a unique quasitrivial semigroup if and only if it has at most one neutral
element. Moreover, we characterized the class of quasitrivial (and symmetric) n-ary
semigroups that have at most one neutral element. Finally, when the underlying
set is finite, this work led to four new integer sequences in the Sloane's OEIS [12],
namely, A308351, A308352, A308354, and A308362.
Note however that there exist idempotent n-ary semigroups that are not re-
ducible to a semigroup (for instance, consider the idempotent associative operation
F ∶ R3 → R defined by F(x, y, z) = x − y + z for all x, y, z ∈ R). This naturally asks
for necessary and sufficient conditions under which an idempotent n-ary semigroup
is reducible to a semigroup. This and other related questions constitute topics for
future research.
Acknowledgements
Both authors would like to thank Jean-Luc Marichal and the anonymous referee
for their useful comments and insightful remarks that helped improving the current
paper.
QUASITRIVIAL n-ARY SEMIGROUPS
17
References
[1] Ackerman, N.L.: A characterization of quasitrivial n-semigroups. Algebra Universalis (in
press)
[2] Bulatov, A.A.: Conservative constraint satisfaction re-revisited. Journal of Computer and
System Sciences 82, 347 -- 356 (2016)
[3] Couceiro, M., Devillet, J., Marichal, J.-L.: Quasitrivial semigroups: characterizations and
enumerations. Semigroup Forum 98, 472 -- 498 (2019)
[4] Dornte, W.: Untersuchungen uber einen verallgemeinerten Gruppenbegriff. Math. Z. 29, 1 -- 19
(1928)
[5] Dudek, W.A., Mukhin, V.V.: On n-ary semigroups with adjoint neutral element. Quasigroups
and Related Systems 14, 163 -- 168 (2006)
[6] Devillet, J., Kiss, G., Marichal, J.-L.: Characterizations of quasitrivial symmetric nonde-
creasing associative operations. Semigroup Forum 98, 154 -- 171 (2019)
[7] Kiss, G., Somlai, G.: Associative idempotent nondecreasing functions are reducible. Semi-
group Forum 98, 140 -- 153 (2019)
[8] Krantz, D.H., Luce, R.D., Suppes, P., Tverskyand, A.: Foundations of measurement, vol. 1.
Academic Press, New York (1971)
[9] Langer, H.: The free algebra in the variety generated by quasi-trivial semigroups. Semigroup
Forum 20, 151 -- 156 (1980)
[10] Post, E.L.: Polyadic groups. Trans. Amer. Math. Soc. 48, 208 -- 350 (1940)
[11] Pouzet, M., Rosenberg, I.G., Stone, M.G.: A projection property. Algebra Universalis 36,
159 -- 184 (1996)
[12] Sloane, N.J.A.: The On-Line Encyclopedia of Integer Sequences. http://www.oeis.org
Universit´e de Lorraine, CNRS, Inria, LORIA, F-54000 Nancy, France
E-mail address: miguel.couceiro[at]{loria,inria}.fr
Mathematics Research Unit, University of Luxembourg, Maison du Nombre, 6, avenue
de la Fonte, L-4364 Esch-sur-Alzette, Luxembourg
E-mail address: jimmy.devillet[at]uni.lu
|
1002.0278 | 1 | 1002 | 2010-02-01T16:13:45 | Minimal Prime Ideals of Ore Extensions over Commutative Dedekind Domains | [
"math.RA"
] | Let R = D[x;\sigma;\delta] be an Ore extension over a commutative Dedekind domain D, where \sigma is an automorphism on D. In the case \delta = 0 Marubayashi et. al. already investigated the class of minimal prime ideals in term of their contraction on the coefficient ring D. In this note we extend this result to a general case \delta not 0. | math.RA | math |
Minimal Prime Ideals of Ore
Extensions over Commutative
Dedekind Domains
Amir Kamal Amir1, Pudji Astuti2, Intan Muchtadi-Alamsyah2
1Algebra Research Division, Faculty of Mathematics and Natural Sciences,
Institut Teknologi Bandung (ITB), Jl. Ganesha 10 Bandung 40132, Indonesia;
Department of Mathematics, Hasanuddin University, Makassar 90245, Indonesia.
2Algebra Research Division, Faculty of Mathematics and Natural Sciences,
[email protected]
Institut Teknologi Bandung (ITB), Jl. Ganesha 10 Bandung 40132, Indonesia.
[email protected]; [email protected]
Abstract
Let R = D[x; σ, δ] be an Ore extension over a commutative Dedekind
domain D, where σ is an automorphism on D.
In the case δ = 0
Marubayashi et. al. already investigated the class of minimal prime
ideals in term of their contraction on the coefficient ring D. In this
note we extend this result to a general case δ 6= 0.
Keywords: minimal prime, Ore extension, derivation.
1
Introduction
This paper studies minimal prime ideals of an Ore extension over a commuta-
tive Dedekind domain. Ore extensions are widely used as the underlying rings
of various linear systems investigated in the area Algebraic system theory.
These systems may represent systems coming from mathematical physics, ap-
plied mathematics and engineering sciences which can be described by means
of systems of ordinary or partial differential equations, difference equations,
differential time-delay equations, etc. If these systems are linear, they can
be defined by means of matrices with entries in non-commutative algebras of
functional operators such as the ring of differential operators, shift operators,
time-delay operators, etc. An important class of such algebras is called Ore
extensions (Ore Algebras).
The structure of prime ideals of various kind of Ore extensions have been
investigated during the last few years. In [7], [8] primes of Ore extensions
over commutative noetherian rings were considered.
In [2], [3], and [11],
prime ideals of Ore extensions of derivation type were described. These
result recently were exploited in [9] to investigate properties of minimal prime
rings of Ore extensions of derivation type in term of their contraction on the
coefficient ring. In this note we extend this result to a general Ore extension
of automorphism type .
2 Ore Extension
We recall some definitions, notations, and more or less well known facts
concerning. A (left) skew derivation on a ring D is a pair (σ, δ) where σ is
a ring endomorphism of D and δ is a (left) σ-derivation on D; that is,
an additive map from D to itself such that δ(ab) = σ(a)δ(b) + δ(a)b for all
a, b ∈ D. For (σ, δ) any skew derivation on a ring D, we obtain
δ(am) =
m−1
X
i=0
σ(a)iδ(a)am−1−i
for all a ∈ D and m = 1, 2, · · · . (See [4, Lemma 1.1])
Definition 1 Let D be a ring with identity 1 and (σ, δ) be a (left) skew
derivation on the ring D. The Ore Extension over D with respect to the
skew derivation (σ, δ) is the ring consisting of all polynomials over D with
an indeterminate x denoted by:
D[x; σ, δ] = { f (x) = anxn + · · · + a0 ai ∈ D }
2
satisfying the following equation, for all a ∈ D
xa = σ(a)x + δ(a).
The notations D[x; σ] and D[x; δ] stand for the particular Ore extensions
where respectively δ = 0 dan σ the identity map. For the case δ = 0,
Marubayashi et. al. [9] studied the factor rings of D[x; σ] over minimal prime
ideals where D is a commutative Dedekind domain. In order to extend their
results to general cases, this paper investigates the class of minimal prime
ideals in D[x; σ, δ].
The Ore extension D[x; σ, δ] is a free left D-module with basis 1, x, x2, · · ·
To abbreviate the assertion, the symbol R stands for the Ore extension
D[x; σ, δ] constructed from a ring D and a skew derivation (σ, δ) on D.
The degree of a nonzero element f ∈ R is defined in the obvious fashion.
Since the standard form for elements of R is with left-hand coefficients, the
leading coefficient of f is fn if
f = f0 + f1x + · · · + fn−1xn−1 + fnxn
with all fi ∈ D and fn 6= 0. If σ is an automorphism, f can also be written
with right-hand coefficients, but then its xn-coefficient is σ−n(fn). While a
general formula for xna where a ∈ D and n ∈ N is too involved to be of much
use, an easy induction establishes that
xna = δn(a) + ax + · · · + an−1xn−1 + σn(a)xn
for some a1, · · · , an−1 ∈ D.
In preparation for our analysis of the types of ideals occured when prime
ideals of an Ore extension D[x; σ, δ] are contracted to the coefficient ring D,
we consider σ-prime, δ-prime, and (σ, δ)-prime ideals of D.
Definition 2 Let Σ be a set of map from the ring D to itself. A Σ-ideal of D
is any ideal I of D such that α(I) ⊆ I for all α ∈ Σ. A Σ-prime ideal is any
proper Σ-ideal I such that whenever J, K are Σ-ideals satisfying JK ⊆ I,
then either J ⊆ I or K ⊆ I.
3
In the context of a ring D equipped with a skew derivation (σ, δ), we shall
make use of the above definition in the cases Σ = {σ}, Σ = {δ} or Σ =
{σ, δ}; and simplify the prefix Σ to respectively σ, δ, or (σ, δ). Concerning
the contraction of prime ideals in an Ore Extension to its coefficient ring,
Goodearl [4] obtained the following theorem which will be of use later.
Theorem 1 Let R = D[x; σ, δ] where D is a commutative Dedekind domain
and σ is an automorphism. If p is any ideal of D which is (σ, δ)-prime, then
p = P ∩ R for some prime ideal P of R and more specially pR ∈ Spec(R)
where Spec(R) denotes the set of all Prime ideal in R.
3 Minimal Prime Ideals of Ore Extensions
Throughout this section, let D be a commutative Dedekind domain and
R = D[x; σ, δ] be the Ore extension over D, for (σ, δ) is a skew derivation,
σ 6= 1 is an automorphism of D and δ 6= 0.
Marubayashi et. al.
[9] already investigated the class of minimal prime
ideals in R where δ = 0. For the case δ 6= 0 but it is inner, Goodearl [4]
showed the existance of an isomorphism between D[x; σ, δ] and D[y; σ] as
described in the following. The σ-derivative δ is called inner if there exists
an element a ∈ D such that δ(b) = ab − σ(b)a for all b ∈ D.
Theorem 2 Let D[x; σ, δ] be an Ore extension where σ 6= 1 and δ 6= 1. If δ
is an inner σ-derivation, i.e, there exists a ∈ D such that δ(b) = ab − σ(b)a
for all b ∈ D, then D[x; σ, δ] and D[y; σ], where y = x − a, are isomorph.
Hence, by combining Theorem 2 and the class of minimal prime ideals in
Ore extensions R obtained in [9] we can derive the class of minimal prime
ideals in Ore extension R = D[x; σ, δ] for σ and δ being respectively an
automorphism and inner as the following. Notation Spec0(R) stands for the
set of all prime ideals in R = D[x; σ, δ] having zero intersection with D.
4
Theorem 3 Let R = D[x; σ, δ] be an Ore extension.
(1) The set
{p[x; σ, δ], P p is a σ − prime ideal of Dand P ∈ Spec0(R) with P 6= (0)}
consists of all minimal prime ideals of R.
(2) Let P ∈ Spec(R) with P 6= (0). Then P is invertible if and only if it is
a minimal prime ideal of R.
Now we shall investigate the class of minimal prime ideals for general
δ 6= 0. For this general case, we need the following lemma.
Lemma 1 If P = p[x; σ, δ] is a minimal prime ideal of R where p is a (σ, δ)-
prime ideal of D, then p is a minimal (σ, δ)-prime ideal of D.
Proof. Assume that p is a (σ, δ)-prime ideal of D but is not minimal (σ, δ)-
prime. Let q be a (σ, δ)-prime ideal of D such that q p and q 6= (0). Then
applying [2, Theorem 3.1] we have qR ∈ Spec(R). So, qR pR = P . This
is a contradiction because P is a minimal prime ideal. (cid:3)
Theorem 4 Let P be a prime ideal of R and P ∩ D = p 6= (0). Then P is
a minimal prime ideal of R if and only if either P = p[x; σ, δ] where p is a
minimal (σ, δ)-prime ideal of D or (0) is the largest (σ, δ)-ideal of D in p.
Proof. Let P be a minimal prime ideal of R and P ∩ D = p 6= (0). Since
p 6= (0), then there are two cases [4, Theorem 3.1]; namely, either p is a
(σ, δ)-prime ideal of D or p is a prime ideal of D and σ(p) 6= p.
If p is a (σ, δ)-prime ideal of D, then pR ∈ Spec(R), by [4, Theorem 3.1].
So, pR = P because pR ⊆ P and P is a minimal prime ideal. On the other
hand pR = p[x; σ, δ]. From here, we get P = p[x; σ, δ] and p is a minimal
(σ, δ)-prime ideal of D, by Lemma 1.
Suppose p is a prime ideal of D and σ(p) 6= p. Let m be the largest
(σ, δ)-ideal contained in p and assume that m 6= (0). Then by primeness of
p it can be shown that m is a (σ, δ)-prime ideal of D. So, mR is a prime
ideal of R by [4, Proposition 3.3]. On the other hand, since σ(p) 6= p, we
5
have m p. So, mR pR ⊆ P , i.e, P is not a minimal prime. This is a
contradiction. So, (0) is the largest (σ, δ)-ideal of D in p.
Conversely, let P = p[x; σ, δ], where p is a minimal (σ, δ)-prime ideal
of D. Then, according to [4, Theorem 3.3], P = p[x; σ, δ] is a prime ideal
of R. Let Q be a prime ideal of R where Q ⊆ P . Set q = Q ∩ D, then
q = Q ∩ D ⊆ P ∩ D = p. Applying [4, Theorem 3.1], we have two cases;
namely, either q is a (σ, δ)-prime ideal of D or q is a prime ideal of D and
σ(q) 6= q.
For the first case, suppose q is a (σ, δ)-prime ideal of D. Then q = p
because q ⊆ p and p is a minimal (σ, δ)-prime ideal of D. So, P = p[x; σ, δ] =
q[x; σ, δ] ⊆ Q. This implies P = Q. For the other case, if q is a prime ideal
of D, then q = p because D is a Dedekind domain. So, P = p[x; σ, δ] =
q[x; σ, δ] ⊆ Q. This implies P = Q. Thus P is a minimal prime ideal of R.
Let (0) be the largest (σ, δ)-ideal of D in p. Let Q be a prime nonzero
ideal of R satisfying Q ⊆ P . Set q = Q ∩ D, then q = Q ∩ D ⊆ P ∩ D = p.
Similar to the above explanation, we have two cases; namely, either q is a
(σ, δ)-prime ideal of D or q is a prime ideal of D and σ(q) 6= q but the first
case will not happen. If it is so, then, because of (0) being the largest (σ, δ)-
ideal of D in p, q = (0) implying a contradiction Q ∩ D = 0 (see [5, Lemma
2.19]). Thus q is a prime ideal of D with σ(q) 6= q. As a result, q = p. So
Q ∩ D = P ∩ D, which, according to [5, Proposition 3.5], implies Q = P .
Thus P is the minimal prime ideal of R. (cid:3)
4 Concluding Remark
In this paper we identify all minimal prime ideals in an Ore extension over
a Dedekind domain according to their contraction on the coefficient ring.
Studying the results in Marubayashi et. al [9] it is expected that this identi-
fication can be used to study the structure of the corresponding factor rings
which is currently under investigation.
6
Acknowledgement
This research was supported by ITB according to Surat Perjanjian Pelak-
sanaan Riset Nomor: 265/K01.7/PL/2009, 6 February 2009.
References
[1] A. D. Bell, When are all prime ideals in an Ore extension Goldie?,
Comm. Algebra 13(8) (1985), 1743-1762.
[2] W. Chin, Prime ideals in differential operator rings and crossed product
of infinite groups, J. Algebra 106 (1987), 78-104.
[3] M. Ferrero, J. Matczuk, Prieme ideals in skew polinomial rings of deriva-
tion type, Comm. Algebra 18 (3) (1990), 689-710.
[4] K.R. Goodearl, Prieme ideals in skew polinomial ring and quantized
Weyl algebras, J. of Algebra 150, (1992), 324-377
[5] K.R. Goodearl, R.B. Warfield, JR An Introduction to Noncommuta-
tive Noetherian rings, London Mathematical Society Student Text, 16
(1989).
[6] R.S. Irving, Prime ideals of Ore extension over commutative rings , J.
Algebra 56 (1979), 315-342
[7] R.S. Irving, Prime ideals of Ore extension over commutative rings II ,
J. Algebra 58 (1979), 399-423.
[8] A. Leroy, J. Matczuk, The extended centeroid and X-inner automor-
phism of Ore extensions, J. Algebra 145 (1992), 143-177.
[9] H. Marubayashi, A.K. Amir, Y. Wang, Prime factor rings of skew poly-
nomial rings over a commutative Dedekind domain, (submitted).
[10] J.C. McConnell and J.C. Robson, Noncommutative Noetherian Rings,
Wiley-Interscience, New York, (1987).
7
[11] D.S. Passman, Prime ideals in enveloping rings, Trans. Amer. Math.
Soc. 302(2) (1987), 535-560.
8
|
1305.0380 | 1 | 1305 | 2013-05-02T09:39:38 | Some remarks on non-commutative principal ideal rings | [
"math.RA",
"math-ph",
"math-ph",
"nlin.SI"
] | We prove some algebraic results on the ring of matrix differential operators over a differential field in the generality of non-commutative principal ideal rings. These results are used in the theory of non-local Poisson structures. | math.RA | math | SOME REMARKS ON NON-COMMUTATIVE PRINCIPAL IDEAL RINGS
SYLVAIN CARPENTIER, ALBERTO DE SOLE, VICTOR G. KAC
Abstract: We prove some algebraic results on the ring of matrix differential operators over a differential field in
the generality of non-commutative principal ideal rings. These results are used in the theory of non-local Poisson
structures.
R´esum´e: Nous d´emontrons quelques r´esultats alg´ebriques sur l'anneau des matrices op´erateurs diff´erentiels sur un
corp diff´erentiel dans la g´en´eralit´e des anneaux non-commutatives principaux. Ces r´esultats sont utilis´es dans la
th´eorie des structures de Poisson non-locales.
3
1
0
2
y
a
M
2
]
.
A
R
h
t
a
m
[
1. Introduction
In our previous two papers [CDSK12a, CDSK12b] we established some algebraic properties of the ring of ma-
trix differential operators over a differential field. The problems naturally arose in the study of non-local Poisson
structures [DSK12a, DSK12b].
Eventually we realized that the proofs of [CDSK12b] can be simplified, so that our results hold in the full generality
of left and right principal ideal rings.
The new result which is not contained in our previous paper is Theorem 3.3, which is used in the theory of
non-local Lenard-Magri scheme in [DSK12b].
We wish to thank Toby Stafford and Lance Small for very useful correspondence. In particular, Toby Stafford
provided us a proof of Theorem 2.2.
1
v
0
8
3
0
.
5
0
3
1
:
v
i
X
r
a
2. General facts about principal ideal rings
Let R be a unital associative (not necessarily commutative) ring. Recall that a left (resp. right ) ideal of R is an
additive subgroup I Ă R such that RI " I (resp. IR " I). The left (resp. right) principal ideal generated by a P R
is, by definition, Ra (resp. aR).
Throughout the paper, we assume that the ring R is both a left and a right principal ideal ring, meaning that
every left ideal of R and every right ideal of R is principal.
Example 2.1. Let K be a differential field with a derivation B, and let KrBs be the ring of differential operators over
K. It is well known that KrBs is a left and right principal ideal domain, see e.g. [CDSK12a]. Let R " MatℓℓpKrBsq
be the ring of ℓ ℓ matrices with coefficients in KrBs. By Theorem 2.2(a) below, the ring R is a left and right
principal ideal ring as well. Note also that Kℓ is naturally a left R-module.
Given an element a P R, an element d P R is called a right (resp. left ) divisor of a if a " a1d (resp. a " da1) for
some a1 P R. An element m P R is called a left (resp. right ) multiple of a if m " qa (resp. m " aq) for some q P R.
Given elements a, b P R, their right (resp. left ) greatest common divisor is the generator d of the left (resp. right )
ideal generated by a and b: Ra ` Rb " Rd (resp. aR ` bR " dR). It is uniquely defined up to multiplication by an
invertible element. It follows that d is a right (resp. left) divisor of both a and b, and we have the Bezout identity
d " ua ` vb (resp. d " au ` bv) for some u, v P R.
Similarly, the left (resp. right ) least common multiple of a and b is an element m P R, defined, uniquely up to
multiplication by an invertible element, as the generator of the intersection of the left (resp. right) principal ideals
generated by a and by b: Rm " Ra X Rb (resp. mR " aR X bR)..
We say that a and b are right (resp.
left ) greatest common divisor is 1 (or
invertible), namely if the left (resp. right ) ideal that they generate is the whole ring R: Ra ` Rb " R (resp.
aR ` bR " R). Clearly, this happens if and only if we have the Bezout identity ua ` vb " 1 (resp. au ` bv " 1) for
some u, v P R.
left ) coprime if their right (resp.
An element k P R is called a right (resp.
left ) zero divisor if there exists k1 P Rzt0u such that k1k " 0 (resp.
kk1 " 0). Note that, if d is a right (resp. left) divisor of a, and d is a left (resp. right) zero divisor, then so is a. In
particular, if either a or b is not a left (resp. right) zero divisor, then their right (resp. left) greatest common divisor
d is also not a left (resp. right) zero divisor. A non-zero element b P R is called regular if it is neither a left nor a
right zero divisor.
1
2
SYLVAIN CARPENTIER, ALBERTO DE SOLE, VICTOR G. KAC
The following results summarize some important properties of principal ideal rings that will be used in this paper.
Since a principal ideal ring is obviously Noetherian, one can use the powerful theory of non-commutative Noetherian
rings (see [MR01]).
Theorem 2.2. Let R be a left and right principal ideal ring. Then:
(a) The ring MatℓℓpRq of ℓ ℓ matrices with entries in R is a left and right principal ideal ring.
(b) The sets of left and right zero divisors of R coincide. Hence, an element of R is regular if and only if it is not
a left (or a right) zero divisor.
(c) The set of regular elements of R satisfies the left (resp. right) Ore property: for a, b P R with b regular, there
exist a1, b1 P R, with b1 regular, such that ba1 " ab1 (resp. a1b " b1a).
(d) There exists the ring of fractions QpRq containing R, consisting of left fractions ab´1 (or, equivalently, right
fractions b´1a), with a, b P R and b regular.
(e) Given a, b P R with b regular, there exists q P R such that a ` qb (resp. a ` bq) is regular.
(f ) Suppose that the ring R contains a central regular element r P R such that r´1 is regular too. Given a1, a2, b1, b2 P
R with b1, b2 regular, there exists q P R such that a1 ` qb1 and a2 ` qb2 (resp. a1 ` b1q and a2 ` b2q) are both
regular.
Proof (by J.T. Stafford). Statement (a) is in [MR01, Prop.3.4.10]. For part (b) [MR01, Cor.4.1.9] shows that R is a
direct sum R " A ' B of an Artinian ring A and a Noetherian semiprime ring B. Obviously the regular elements of
A are just the units. By Goldie's Theorem the right regular elements of B are the same as the left regular elements,
i.e. the regular elements (see [MR01, Prop.2.3.4 and 2.3.5]). Since an element pa, bq P R " A ' B is regular if and
only if a and b are both regular the same conclusion holds for R. This proves (b) The equivalence of (c) and (d)
is Ore's Theorem [MR01, Thm.2.1.12]. Part (c) then follows from Goldie's Theorem. It is routine to see that the
regular elements of A ' B form an Ore set if this is true for both A and B. Of course this result is vacuously true
for A while Goldie's Theorem does it for B. Part (e) follows from [SS82, Cor.2.5], and part (f) is in [Sta12].
(cid:3)
Remark 2.3. As T. Stafford pointed out, the ring R " Z{2Z does not satisfy the property in part (f).
Remark 2.4. From the above theorem we immediately get the following simple observations.
(a) By Theorem 2.2(b) we have that if a " bc, then a is regular if and only if b and c are regular. In particular, any
left or right divisor of a regular element is regular.
(b) If b is regular and a arbitrary, then we can write their right (resp.
left) least common multiple as ab1 " ba1
with b1 regular. This follows from the Ore property in Theorem 2.2(c). Indeed, let I " tb1 P R ab1 P bRu. It is
clearly a right ideal of R. Hence, I " b1R for some b1. Clearly, m " ab1 is the right least common multiple of a
and b. By the Ore property, there exists a regular element rb P I. Hence, rb " b1c, and therefore b1 is regular too.
(c) It follows from the above observation that, if a and b are regular, so is their right (resp.
left) least common
multiple.
(d) If a " a1d, b " b1d, (resp. a " da1, b " db1), and a1 and b1 are right (resp. left) coprime, then d is the right
left) greatest common divisor of a and b. Indeed, by the Bezout identity we have ua1 ` vb1 " 1 (resp.
(resp.
a1u ` b1v " 1), which implies ua ` vb " d (resp. au ` bv " d). But then Rd " Ra ` Rb (resp. dR " aR ` bR),
proving the claim.
(e) Conversely, if a " a1d, b " b1d, (resp. a " da1, b " db1), and d is the right (resp. left) greatest common divisor
of a and b, then, assuming that d is regular, we get that a1 and b1 are right (resp. left) coprime. Indeed, by the
Bezout identity we have d " ua ` vb " pua1 ` vb1qd (resp. d " au ` bv " dpa1u ` b1vq), and since by assumption
d is regular it follows that ua1 ` vb1 " 1.
3. Some arithmetic properties of principal ideal rings
1 P QpRq (resp. f " b´1a " b´1
Theorem 3.1. Let R be a left and right principal ideal ring and let QpRq be its ring of fractions. Let f " ab´1 "
a1b´1
1 a1 P QpRq), with a, a1, b, b1 P R and b, b1 regular, and assume that a1 and b1 are
right (resp. left) coprime. Then there exists a regular element q P R such that a " a1q and b " b1q (resp. a " qa1
and b " qb1).
Proof. By assumption a1 and b1 are right coprime, hence we have the Bezout identity ua1 ` vb1 " 1, for some
u, v P R. Let q " ua ` vb. We have
b1q " b1pua ` vbq " b1puab´1 ` vqb " b1pua1b´1
1 ` vqb
" b1pua1 ` vb1qb´1
1 b " b ,
and
a1q " a1b´1
1 b1q " a1b´1
1 b " ab´1b " a .
SOME REMARKS ON NON-COMMUTATIVE PRINCIPAL IDEAL RINGS
Finally, q is regular since q " b´1
1 b is invertible in QpRq.
3
(cid:3)
Corollary 3.2. For every f P QpRq there is a "minimal" right (resp. left) fractional decomposition f " ab´1 (resp.
f " b´1a) with a, b right (resp. left) coprime. Any other right (resp. left) fractional decomposition is obtained from
it by simultaneous multiplication of a and b on the right (resp. left) by some regular element q P R.
Proof. It follows immediately from Remark 2.4(d) and Theorem 3.1.
(cid:3)
Theorem 3.3. Let R be a left and right principal ideal ring, and let V be a left module over R. Assume that the
ring R contains a central regular element r P R such that r ´ 1 is regular too. Let a, b P R, with b regular, be left
coprime. Let m " ab1 " ba1 be their right least common multiple. Then, for every x, y P V such that ax " by, there
exists z P V such that x " b1z and y " a1z. In particular, aV X bV " mV .
Proof. We first reduce to the case when a is regular. Indeed, let, by Theorem 2.2(e), q P R be such that a ` bq is
regular. Then it is immediate to check that the right least common multiple of a` bq and b is pa` bqqb1 " bpa1 ` qb1q.
Moreover, since by assumption ax " by, we have pa ` bqqx " bpy ` qxq. Therefore, assuming that the theorem holds
for regular a, there exists z P V such that x " b1z and y ` qx " pa1 ` qb1qz, which implies y " a1z, proving the
claim.
Next, let us prove the theorem under the assumption that both a and b are regular. Since m " ab1 " ba1 is the
right least common multiple of a and b, it follows that a1 and b1 are right coprime, and therefore we have the Bezout
identity
(3.1)
ub1 ` va1 " 1 ,
for some u, v P R. After replacing u by u ` qa and v by v ´ qb, equation (3.1) still holds. Hence, by Theorem 2.2(f),
we can assume, without loss of generality, that u and v are both regular. Moreover, by Remark 2.4(c), since by
assumption both a and b are regular, their right least common multiple is regular too, and therefore a1 and b1 are
regular too. Multiplying in QpRq both sides of equation (3.1) on the left by u´1 and on the right by a´1
1 , we get
(3.2)
and similarly, multiplying (3.1) on the left by v´1 and on the right by b´1
a´1b " pa1uq´1p1 ´ a1vq ,
1 , we get
(3.3)
b´1a " pb1vq´1p1 ´ b1uq .
Since, by assumption, a and b are left coprime, both fractions a´1b and b´1a are in their minimal fractional decom-
position. Hence, by equations (3.2) and (3.3), there exist p, q P R such that
(3.4)
(3.5)
1 ´ a1v " pb , a1u " pa ,
1 ´ b1u " qa , b1v " qb .
Applying the first equation in (3.4) to y P V and using the assumption ax " by and the second equation of (3.4),
we get
y " a1vy ` pby " a1vy ` pax " a1pvy ` uxq ,
and, similarly, applying the first equation in (3.5) to x P V and using the second equation of (3.5), we get
x " b1ux ` qax " b1ux ` qby " b1pux ` vyq .
Hence, the statement holds with z " ux ` vy.
(cid:3)
If V is a left R-module and a P R, we denote Ker a " x P V ax " 0(.
Remark 3.4. If d is the right greatest common divisor of a and b in R, then Ker a X Ker b " Ker d. Indeed, by the
Bezout identity we have b1a ` a1b " d. Therefore Ker a X Ker b Ă Ker d. The reverse inclusion is obvious.
Corollary 3.5. Let R be as in Theorem 3.3, and let V be a left R-module. Let σ : R Ñ R be an anti-automorphism
of the ring R. Let a, b P R, with b regular, be right coprime, and suppose that σpaqb " ǫσpbqa, for some invertible
central element ǫ P R. Let x, y P V be such that σpaqx " ǫσpbqy. Then there exists z P V such that x " bz and
y " az.
Proof. First, since b is regular and σ is an anti-automorphism, σpbq is regular too. Moreover, since by assumption a
and b are right coprime and σ is an anti-automorphism, it follows that σpaq and σpbq are left coprime.
We claim that the left least common multiple of a and b is equal to the right least common multiple of σpaq and
σpbq, and it is given by m " σpaqb " ǫσpbqa. Indeed, clearly m is a common right multiple of σpaq and σpbq. It is
therefore a right multiple of the minimal one: m1 " σpaqb1 " σpbqa1. Namely, there exists q P R such that b " b1q
and a " ǫ´1a1q. But by assumption a and b are right coprime. Hence, q must be invertible, proving that m is
(3.6)
La,b " !bx ' ax x P V) Ă V ' V .
4
SYLVAIN CARPENTIER, ALBERTO DE SOLE, VICTOR G. KAC
the right least common multiple of σpaq and σpbq. The same argument proves that m is also the left least common
multiple of a and b.
We can now apply Theorem 3.3 to σpaq and σpbq, to deduce that there exists z P V such that x " bz and ǫy " ǫaz,
(cid:3)
hence y " az.
As in [CDSK12b], Corollary 3.5 implies the following maximal isotropicity property important for the theory of
Dirac structures, [Dor93, DSK12a].
Corollary 3.6. Let R be as in Theorem 3.3, and let V be a left R-module and let p , q : V V Ñ A be a non-
degenerate symmetric bi-additive pairing on V with values in an abelian group A. Let be an anti-involution of R
such that pax, yq " px, ayq for all a P R and x, y P V . Extend the pairing p , q to a pairing x y on V ' V with
values in A, given by
xx1 ' x2 y1 ' y2y " px1, y2q ` px2, y1q ,
for every x1, x2, y1, y2 P V . Given two elements a, b P R, we consider the following additive subgroup of V ' V :
(a) The subgroup La,b Ă V ' V is isotropic with respect to the pairing x y if and only if ab ` ba acts as 0 on V .
(b) If b is regular, a and b are right coprime, and ab ` ba " 0, then the subgroup La,b Ă V ' V is maximal
isotropic.
Proof. Part (a) is straightforward and part (b) follows immediately from Corollary 3.5 with σpaq " a and ǫ "
´1.
(cid:3)
Corollary 3.7. Let R be as in Theorem 3.3, and let V be a left R-module. Let a, b P R, with b regular, be left
coprime. Let m " ab1 " ba1 be their right least common multiple. Then Ker b " a1pKer b1q.
Proof. If k1 P Ker b1, then bpa1k1q " ab1k1 " 0. Therefore, a1pKer b1q Ă Ker b. We need to prove the opposite
inclusion. If k P Ker b, we have a0 " 0 " bk. Hence, by Theorem 3.3, there exists z P V such that 0 " b1z and
k " a1z. Namely, k P a1pKer b1q.
(cid:3)
Remark 3.8. In the ring R " Matℓℓ KrBs of ℓ ℓ matrix differential operators over a differential field K, the above
Corollary 3.7 implies that if b´1a " a1b´1
is a rational matrix pseudodifferential operator in its minimal left and
right fractional decompositions, then degpbq " degpb1q (where degpbq is the degree of the Dieudonn´e determinant
of b). Indeed, the fractional decomposition b´1a being minimal means that a and b are left coprime. Hence, by
Corollary 3.7 we have that dimpKer bq " dimpa1 Ker b1q in any differential field extension of K. Moreover, the
fractional decomposition a1b´1
1 being minimal means that Ker a1 X Ker b1 " 0 in any differential field extension of
K. The claim follows by the fact that deg b is equal to the dimension of Ker b in the linear closure of K [CDSK12b].
1
References
[CDSK12a] S. Carpentier, A. De Sole, and V.G. Kac, Some algebraic properties of matrix differential operators and their Dieudonn´e
determinant, J. Math. Phys. 53 (2012), 063501
[CDSK12b] S. Carpentier, A. De Sole, and V.G. Kac, Rational matrix pseudodifferential operators, preprint (2012) arXiv:1206.4165.
[DSK12a] A. De Sole, and V.G. Kac, Non-local Poisson structures and applications to the theory of integrable systems I, preprint (2012)
arXiv:1210.1688
[DSK12b] A. De Sole, and V.G. Kac, Non-local Poisson structures and applications to the theory of integrable systems II, preprint (2012)
arXiv:1211.2391
[Dor93] I.Ya. Dorfman, Dirac structures and integrability of nonlinear evolution equations, Nonlinear Sci. Th. Appl., Wiley & Sons, New
York, 1993.
[MR01] J.C. McConnell and J.C. Robson, Non-commutative Noetherian rings, Graduate Studies in Mathematics, v.30, American Math-
ematical Society, Providence, RI, 2001.
[SS82] L.W. Small, J.T. Stafford, Regularity of zero divisors, Proc. London Math. Soc. (3) 44 (1982), no. 3, 405 -- 419.
[Sta12] J.T. Stafford, private communication (2012).
Ecole Normale Superieure, Paris, France, e-mail: [email protected]
Dept. of Math., University of Rome 1, Italy e-mail: [email protected]
Dept. of Math., MIT, USA e-mail: [email protected]
|
1012.1849 | 3 | 1012 | 2018-08-10T06:43:23 | Isotopes of Hurwitz algebras | [
"math.RA"
] | We study the class of all algebras that are isotopic to a Hurwitz algebra. Isomorphism classes of such algebras are shown to correspond to orbits of a certain group action. A complete, geometrically intuitive description of the category of isotopes of Hamilton's quaternions is given. As an application, we demonstrate how some results concerning the classification of finite-dimensional composition algebras can be deduced from our general results. | math.RA | math |
ISOTOPES OF HURWITZ ALGEBRAS
ERIK DARP O
Abstract. We study the class of all algebras that are isotopic to a Hurwitz algebra. In the
general case, isomorphism classes of such algebras are shown to correspond to orbits of a certain
group action. Our main focus is the isotopes of Hamilton's quaternions, for which a more de-
tailed, and geometrically intuitive description of the isomorphism classes is given. As an applic-
ation, we also demonstrate how some results concerning the classification of finite-dimensional
composition algebras can be deduced from our general results.
1. Introduction
Let k be a field, and V a vector space over k. We shall say that a quadratic form q : V → k is
non-degenerate if the associated bilinear form hx, yi = q(x+ y)− q(x)− q(y) is non-degenerate (i.e.,
hx, V i = 0 only if x = 0). A composition algebra is a non-zero (not necessarily associative) algebra
A over a field k, equipped with a non-degenerate quadratic form n : A → k such that n(ab) =
n(a)n(b) for all a, b ∈ A. The form n is usually called the norm of A. A composition algebra
possessesing an identity element is called a Hurwitz algebra. Every Hurwitz algebra has dimension
one, two, four or eight (thus, in particular, it is finite dimensional), and can be constructed via
an iterative method known as the Cayley-Dickson process.1 The facts about Hurwitz algebras
referred in this section are described in detail in [15], Chapter VIII.
Two algebras A and B over a field k are said to be isotopic if there exist invertible linear maps
α, β, γ : A → B such that γ(xy) = α(x)β(y) for all x, y ∈ A. Clearly, isotopy is an equivalence
relation amongst k-algebras. If A and B are isotopic then there exist α, β ∈ GL(A) such that the
algebra (A, ◦), with multiplication x ◦ y = α(x)β(y), is isomorphic to B. The algebra (A, ◦) is
called the principal isotope of A determined by α and β, and is denoted by Aα,β.
Several important classes of non-associative algebras can be constructed by isotopy from the
Hurwitz algebras. Examples include:
(1) All finite-dimensional composition algebras [14, p. 957]. This includes in particular all
finite-dimensional absolute valued algebras, which are precisely the finite-dimensional com-
position algebras over R whose norm is anisotropic (i.e., n(x) = 0 only if x = 0). However,
there exist composition algebras without identity element that are infinite dimensional,
and thus not isotopic to any Hurwitz algebra; see e.g. [4, 8].
(2) All division algebras of dimension two over a field of characteristic different from two [17].
(3) All finite-dimensional division algebras A that are not isotopic to an associative algebra,
and satisfy the following property: for all non-zero a ∈ A there exists an element b ∈ A
such that b(ax) = x for all x ∈ A. [6]
The purpose of the present article is to describe, in a uniform way, all isotopes of Hurwitz
algebras. Generalising ideas that have earlier been used in more specialised situations (for example
in [3, 17]), we give a general description of all algebras isotopic to a Hurwitz algebra, encompassing
also the case of the ground field k having characteristic two. Our principal result is a comprehensive
geometric description of all isotopes of Hamilton's quaternion algebra H -- a class of real division
algebras that has not been studied before.
Key words and phrases. Hurwitz algebra, isotope, quaternion algebra, octonion algebra.
MSC 2010: 17A60 (17A35, 17A75).
1Some authors have used a weaker notion of non-degeneracy for the norm n, requiring that n(x + y) = n(y) for
all y implies x = 0. This definition gives rise to additional unital composition algebras over fields k of characteristic
two, in form of purely inseparable field extensions of k [14]. If char k 6= 2, the two definitions are equivalent.
1
2
ERIK DARP O
Isotopes of Hurwitz algebras of dimension two is a special case of Petersson's [17] description of
all 2-dimensional algebras, which is also based on the concept of isotopy. There is a large number
of articles on finite-dimensional composition algebras and absolute valued algebras (some examples
are [2, 3, 8, 14, 18]), most of which use the concept of isotopy as a central tool.
This article is organised as follows. In Section 2, a general description is given of the category
of isotopes of a Hurwitz algebra A. A more elaborate study of the case where (A, n) is a Euclidean
space is given in Section 3, bringing about the announced description of isotopes of H. Finally,
Section 4 treats composition algebras, showing how a description of these can be deduced from
the results in Section 2.
From here on, let k denote a field. All algebras are, unless otherwise stated, assumed to be finite
dimensional over k. Every element a of an algebra A determines linear endomorphisms La and Ra
of A, defined by La(x) = ax and Ra(x) = xa respectively. For a principal isotope Aα,β = (A, ◦)
of A, we use the notation L◦
a(x) = x ◦ a. An algebra A is said to be a division
algebra if A 6= 0 and La and Ra are bijective for all non-zero a ∈ A. Moreover, A is alternative if
the identities x2y = x(xy) and xy2 = (xy)y hold for all x, y ∈ A.
a(x) = a ◦ x and R◦
All Hurwitz algebras are alternative. Two-dimensional Hurwitz algebras are quadratic ´etale
algebras, i.e., either separable field extensions of k or isomorphic to k × k. The Hurwitz algebras
of dimension four are all quaternion algebras, that is, all four-dimensional central simple associ-
ative algebras. Eight-dimensional Hurwitz algebras are precisely the central simple alternative
algebras that are not associative [21] (these are called octonion algebras). A Hurwitz algebra A is
commutative if and only if dim A 6 2, and associative if and only if dim A 6 4.
Any element x in a Hurwitz algebra A = (A, n) satisfies x2 = hx, 1ix − n(x)1. Hence, the norm
n is uniquely determined by the algebra structure of A, and every algebra morphism of Hurwitz
algebras that respects the identity element also preserves the norm. Every Hurwitz algebra of
dimension at least 2 has a unique non-trivial involution2 κ : A → A, x 7→ ¯x satisfying x + ¯x ∈ k1
and x¯x = ¯xx = n(x)1 for all x ∈ A. Moreover, two Hurwitz algebras are isomorphic if and only
if their respective norms are equivalent (this was first proved in [12] in characteristic different
from two). The quadratic forms occurring as norms of Hurwitz algebras are precisely the m-fold
Pfister forms over k, for m ∈ {0, 1, 2, 3}. If A is a Hurwitz algebra and a ∈ A, then La and Ra
are invertible if and only if n(a) 6= 0. This is also equivalent to the existence of an inverse a−1 of
a in A: since a¯a = ¯aa = n(a)1, we have a−1 = n(a)−1¯a if n(a) 6= 0. Moreover, L−1
a = La−1 and
R−1
a = Ra−1 in this case. The invertible elements of any alternative algebra A form a so-calles
Moufang loop under multiplication (the concept was introduced by Moufang in [16] under the
name quasi-group), denoted by A∗.
For any algebra A, the nucleus is defined as N (A) = {a ∈ A (xy)z = x(yz) for a ∈ {x, y, z}}.
The nucleus is an associative subalgebra of A. If A is a Hurwitz algebra, then a ∈ N (A) if and
only if (xa)y = x(ay) for all x, y ∈ A. If dim A 6 4 then A is associative and thus N (A) = A; the
nucleus of an eight-dimensional Hurwitz algebra is k1.
A similitude of a non-zero quadratic space V = (V, q) is an invertible linear map ϕ : V → V
such that q(ϕ(x)) = µ(ϕ)q(x) for all x ∈ V , where µ(ϕ) ∈ k is a scalar independent of x. The
If l = dim V is even, then det(ϕ) = ±µ(ϕ)l/2 [15,
element µ(ϕ) is called the multiplier of ϕ.
12A]. If char k 6= 2, a similitude ϕ satisfying det(ϕ) = µ(ϕ)l/2 are said to be proper.
In the
characteristic two case, a similitude ϕ : V → V is defined to be proper if its Dickson invariant
(see [15, 12.12]) is zero. The group of all similitudes of V is denoted by GO(V, q), or GO(V ) for
short. The proper similitudes form a normal subgroup GO+(V ) ⊂ GO(V ) of index two. Elements
in GO(V ) r GO+(V ) are called improper similitudes. The map µ : GO(V ) → k∗, ϕ 7→ µ(ϕ) is
a group homomorphism, the kernel of which is the orthogonal group O(V ). We write O+(V ) =
O(V ) ∩ GO+(V ), or SO(V ) = O+(V ) in case k = R and (V, q) is a Euclidean space.
For any Hurwitz algebra A = (A, n), define G(A) to be the set of all ϕ ∈ GO(A) for which there
exist ϕ1, ϕ2 ∈ GO(A) such that ϕ(xy) = ϕ1(x)ϕ2(x) for all x, y ∈ A. Now, by what is known as
2By an involution is meant a self-inverse isomorphism A → Aop.
ISOTOPES OF HURWITZ ALGEBRAS
3
the principle of triality for Hurwitz algebras [19, 3.2],
G(A) =(GO+(A)
GO(A)
if dim A > 4,
if dim A 6 2.
We call ϕ1 and ϕ2 triality components of ϕ. It is not difficult to see that any other pair of triality
components of ϕ is of the form (R−1
ϕ2(1) ϕ and
ϕ2 = L−1
w ϕ1, Lw ϕ2) for some w ∈ N (A)∗, and that ϕ1 = R−1
ϕ1(1) ϕ. If A is associative, then
ϕ(xy) = ϕ1(x)ϕ2(y) = (ϕ(x)ϕ2(1)−1)(ϕ1(1)−1ϕ(y)) = ϕ(x)ϕ(1)−1ϕ(y) .
Given a Hurwitz algebra A, triality components ϕ1, ϕ2 of any ϕ ∈ G(A) are again elements of G(A).
Moreover, (ϕ1)−1 and (ϕ2)−1 are a pair of triality components of ϕ−1. We write PG(A) = G(A)/
(k∗I).
A groupoid is a category in which every morphism is an isomorphism. Every action of a group
G on a set X gives rise to a groupoid with object class X, and morphisms x → y being the set
of group elements g ∈ G satisfying g · x = y. We call this the groupoid of the G-action on X.
Given a vector space V over k, and a quadratic form q : V → k, set PGL(V ) = GL(V )/(k∗I),
PGO(V, q) = GO(V, q)/(k∗I) and PGO+(V, q) = GO+(V, q)/(k∗I). The orbit/coset (of some
group action/normal subgroup) represented by an element in a set/group will be denoted using
square brackets around the element in question. For example, if α ∈ GL(A), the element in
PGL(A) represented by α is written as [α]. If V is a Euclidean space then Pds(V ) denotes the
set of positive definite symmetric endomorphisms of V , and SPds(V ) = Pds(V ) ∩ SL(A). The
set of isomorphisms from an object A to an object B in a category C is denoted by Iso(A, B) =
IsoC (A, B). Throughout, C2 denotes the cyclic group of order two, generated by the canonical
involution in a Hurwitz algebra of dimension at least two: C2 = hκi = {I, κ}.
2. General description
The following lemma is essentially due to Albert [1, Theorem 7], for general (not necessarily
associative) algebras. The formulation has been sharpened slightly and adapted to our situtation.
Lemma 1. Let A be a Hurwitz algebra, and α, β ∈ GL(A). The isotope Aα,β is unital if and only
if α = Ra, β = Lb for some a, b ∈ A∗. The identity element in ARa,Lb is (ab)−1.
e = IA. Now e ◦ x = α(e)β(x), that is, L◦
e = Lα(e) β, so L◦
Proof. The isotope Aα,β = (A, ◦) is unital if and only if there exists an element e ∈ A such that
L◦
e = R◦
e = IA if and only if Lα(e) is
invertible β = L−1
α(e) =
Lα(e)−1 . Similarly, R◦
α(e). Now Lα(e) being invertible means that α(e) ∈ A∗, and thus β = L−1
e = Rβ(e) α equals the identity map if and only if α = R−1
β(e) = Rβ(e)−1 .
It readily verified that (ab)−1 ◦ x = x = x ◦ (ab) in ARa,Lb.
(cid:3)
Proposition 2. Let A be a Hurwitz algebra. Any isotope B of A that has an identity element is
again a Hurwitz algebra, isomorphic to A, and nB = nA(1B)−1nA.
Proof. By Lemma 1, any principal isotope B = (A, ◦) of A that is unital has the form B = ARc,Ld .
Defining nB(x) = nA(cd)nA(x) for all x ∈ B, we have
nB(x ◦ y) = nB((xc)(dy)) = nA(cd)nA((xc)(dy))
= nA(cd)nA(x)nA(cd)nA(y) = nB(x)nB(y)
so B is a Hurwitz algebra. Moreover, Lcd : (B, nB) → (A, nA) is an isometry. Being isometric
as quadratic spaces, A and B are isomorphic algebras. Since 1B = (cd)−1, it is clear that nB =
nA(cd)nA = nA(1B)−1nA.
(cid:3)
From Proposition 2 follows the simple but important observation that, if two isotopes Aα,β and
Bγ,δ of Hurwitz algebras A and B are isomorphic to each other, then A ≃ B.
Corollary 3. Let A be a Hurwitz algebra, and α, β, γ, δ ∈ GL(A). Any isomorphism ϕ : Aα,β →
Aγ,δ is a similitude of (A, nA) with multiplier nA(ϕ(1A)).
4
ERIK DARP O
Proof. Let ϕ : Aα,β → Aγ,δ be an isomorphism. It is straightforward to verify that ϕ is also an
isomorphism A → B, where B = Aγϕα−1ϕ,δϕβ−1ϕ Thus, in particular, ϕ is an orthogonal map
(A, nA) → (B, nB), and ϕ(1A) = 1B. By Proposition 2, nB = nA(1B)−1nA, so nA(ϕ(x)) =
nA(1B)nB(ϕ(x)) = nA(ϕ(1A))nA(x).
(cid:3)
The following proposition describes all isomorphisms between isotopes of Hurwitz algebras. It
generalises the description of eight-dimensional absolute-valued algebras given in [3, Theorem 4.3].
Proposition 4. Let A = (A, n) be a Hurwitz algebra and α, β, γ, δ ∈ GL(A). A map ϕ : Aα,β →
Aγ,δ is an isomorphism if and only if ϕ ∈ G(A) and there exist triality componenets ϕ1, ϕ2 of ϕ
such that
(γ = ϕ1αϕ−1,
δ = ϕ2βϕ−1.
Remark 5. Note that, since any pair (ϕ1, ϕ2) of triality components of ϕ satisfy ϕ1 = R−1
ϕ2 = L−1
ϕ1(1), we could also have written
ϕ2(1),
(γ = ϕ1αϕ−1 = R−1
δ = ϕ2βϕ−1 = L−1
ϕ2(1A) ϕαϕ−1,
ϕ1(1A) ϕβϕ−1
in the proposition above.
Proof. A direct computation shows that if ϕ ∈ G(A) and γ = ϕ1αϕ−1 and δ = ϕ2βϕ−1, then
ϕ ∈ Iso(Aα,β, Aγ,δ).
For the other direction, assume that ϕ ∈ Iso(Aα,β, Aγ,δ). This means that ϕ(α(x)β(y)) =
γϕ(x) · δϕ(y) for all x, y ∈ A, and setting z = α(x), w = β(y) gives
ϕ(zw) = γϕα−1(z) · δϕβ−1(w).
Inserting w = 1A into this equation yields ϕ(z) = γϕα−1(z)·δϕβ−1(1A), so γϕα−1 = R−1
δϕβ−1(1A) ϕ.
By Corollary 3, we know that ϕ is a similitude with multiplier nA(ϕ(1A)), so it follows that γϕα−1
nA(δϕβ−1(1A)) . Similarly, δϕβ−1 is a similitude with multiplier
is a similitude with multiplier
nA(γϕα−1(1A)) . Setting ϕ1 = γϕα−1 and ϕ2 = δϕβ−1 gives ϕ(zw) = ϕ1(z)ϕ2(w) with ϕ1, ϕ2 ∈
GO(A), which means that ϕ ∈ G(A). Moreover, γ = ϕ1αϕ−1 and δ = ϕ2βϕ−1, as required.
(cid:3)
nA(ϕ(1A))
nA(ϕ(1A))
We record the following observations for future use. The first statement is a consequence of the
fact, referred to in the introduction, that x ∈ N (A) if and only if (ax)b = a(xb) for all a, b ∈ A;
the second follows from Proposition 4.
Lemma 6. Let A be a Hurwitz algebra, α, β, γ, δ ∈ GL(A) and ρ ∈ k∗.
(1) Aα,β = Aγ,δ if and only if α = R−1
w γ, β = Lw δ for some w ∈ N (A)∗.
(2) The homothety hρ(x) = ρx on A defines an isomorphism AρI,I → A.
The set T = {([ϕ0], [ϕ1], [ϕ2]) ∈ PG(A)3 ϕ0(xy) = ϕ1(x)ϕ2(y) , ∀x, y ∈ A} is a group under
component-wise multiplication. The kernel of the group epimorphism
π0 : T → PG(A), (ϕ0, ϕ1, ϕ2) 7→ ϕ0
is
hence π0 induces an isomorphism T /N → PG(A). Now T acts on PGL(A)2 by
N = {([IA], [R−1
w ], [Lw]) w ∈ N (A)∗} ≃ N (A)∗/(k∗1) ,
(1)
We denote by XA = PGL(A)2/N the orbit set of the induced N -action on PGL(A)2. Then (1)
induces an action of PG(A) ≃ T /N on XA, given by
([ϕ0], [ϕ1], [ϕ2]) · ([α], [β]) = ([ϕ1αϕ−1
0 ], [ϕ2βϕ−1
0 ]) .
(2)
ϕ · [α, β] =(cid:2)ϕ1αϕ−1, ϕ2βϕ−1(cid:3)
for [α, β] ∈ XA and ϕ ∈ PG(A). Here ϕ1, ϕ2 are any pair of triality components of ϕ, that is,
(ϕ, ϕ1, ϕ2) ∈ π−1
0 (ϕ) ⊂ T .
ISOTOPES OF HURWITZ ALGEBRAS
5
We use the following notational convention: if γ, δ ∈ GL(A), then the element of XA represented
by ([γ], [δ]) ∈ PGL(A) is denoted by [γ, δ] (rather than [[γ], [δ]]).
Let X (A) = PG(A) XA be the groupoid of the action (2). For any Hurwitz algebra A, let I (A)
denote the category of principal isotopes of A, and I (A) the groupoid obtained from I (A) by
removing all non-isomorphisms between the objects. Note that if A is a division algebra then so
are all its isotopes, and thus any non-zero morphism in I (A) is an isomorphism in this case.
The essence of our findings so far is summarised in the following theorem.
Theorem 7. For any Hurwitz algebra A, the categories
equivalence FA : I (A) → X (A) is given by FA(Aα,β ) = [α, β] and FA(ϕ) = [ϕ].
Proof. Let (α, β), (γ, δ) ∈ GL(A)2. If Aα,β = Aγ,δ then, by Lemma 6,
I (A) and X (A) are equivalent. An
(γ, δ) = (R−1
w α, Lw β) = (IA, R−1
w , Lw) · (α, β)
for some w ∈ N (A)∗, and hence the elements ([α], [β]) and ([γ], [δ]) in PGL(A)2 represent the
same object in XA = PGL(A)2/N . Proposition 4 guarantees that [ϕ] · [α, β] = [γ, δ] whenever
ϕ ∈ Iso(Aα,β , Aγ,δ). This shows that FA is well defined. Clearly, FA is surjective on objects,
hence dense as a functor.
If ϕ, ψ ∈ Iso(Aα,β, Aγ,δ) and FA(ϕ) = FA(ψ), then ϕ = ρψ for some ρ ∈ k∗, so ρIA = ϕψ−1 ∈
Aut(Aα,β). This implies ρ = 1 and ϕ = ψ ; hence FA is faithful. Fullness is clear from the
construction.
(cid:3)
Remark 8.
(1) If k is a Euclidean field3 (e.g., k = R), then the group PGO(A, n) is canon-
ically isomorphic to O(A, n)/(±I), via composition of inclusion and quotient projection:
O(A, n)/(±I) ֒→ GO(A, n)/(±I) ։ GO(A, n)/(k∗I) = PGO(A, n). This also induces an
isomorphism O+(A, n)/(±I) → PGO+(A, n).
(2) The set XA = PGL(A)2/N may also be viewed as the orbit set PGL(A)2/N (A)∗ of the
N (A)∗-action on PGL(A)2 given by w · ([α], [β]) =(cid:0)[R−1
(3) Assume that A is associative. The maps πi : T → PG(A), (ϕ0, ϕ1, ϕ2) 7→ ϕi, i = 1, 2,
are group epimorphisms, and since N is normal in T , the images π1(N ) = {[Ra] a ∈
A∗} ⊂ PG(A) and π2(N ) = {[La] a ∈ A∗} ⊂ PG(A) of N under π1 and π2 are normal
subgroups of PG(A). It follows that RA∗ = {Ra a ∈ A∗} and LA∗ = {La a ∈ A∗} are
normal subgroups of G(A).
w α], [Lw β](cid:1).
(4) If dim A = 8, since N (A) = k1, we have XA = PGL(A)2.
If A is associative, then the group G(A) has a particularly nice form.
Proposition 9. If A is an associative Hurwitz algebra, then G(A) can be written as a semi-direct
product in the following ways:
In particular, LA∗ and RA∗ are normal subgroups of G(A).
G(A) = LA∗ ⋊ Aut(A) = RA∗ ⋊ Aut(A) .
Proof. As already noted in Remark 8(2), LA∗ and RA∗ are normal subgroups of G(A). Since any
automorphism of A must fix the identity element 1A in A, we have La ∈ Aut(A) for a ∈ A∗ if and
only if a = 1. Thus LA∗ ∩ Aut(A) = {I}. The inclusion Aut(A) ⊂ G(A) is obvious. If ϕ ∈ G(A),
then
L−1
ϕ(1) ϕ(xy) = L−1
ϕ(1) ϕ ∈ Aut(A). This implies G(A) = LA∗ Aut(A), which shows that G(A) = LA∗ ⋊ Aut(A).
(cid:3)
so L−1
The identity G(A) = RA∗ ⋊ Aut(A) is proved similarly.
ϕ(1)(cid:0)ϕ(x)ϕ(1)−1ϕ(y)(cid:1) = L−1
ϕ(1) ϕ(x) · Lϕ(1)−1 ϕ(y) = L−1
ϕ(1) ϕ(x) · L−1
ϕ(1) ϕ(y),
Note that if dim A = 2 then Aut(A) = {I, κ} = C2. If A is a quaternion algebra then it is
a ∈ A∗}.
b = Lab Rb−1 . This leads to the following
central simple, and the Skolem-Noether Theorem [13, p. 222] gives Aut(A) = {La R−1
a
Hence every ϕ ∈ GO+(A) can be written as ϕ = La Lb R−1
result.
3A field k is called Euclidean if k∗2 forms an ordering of k.
6
ERIK DARP O
Corollary 10. Every proper similitude of a quaterion algebra A has the form La Rb for some
a, b ∈ A∗. The kernel of the group epimorphism A∗ × (A∗)op → GO+(A), (a, b) 7→ La Rb is
{(ρ, ρ−1) ρ ∈ k∗} ⊂ A∗ × (A∗)op.
Proof. From the argument in the paragraph preceding Corollary 10, it is clear that every ϕ ∈
GO+(A) can be written as La Rb for some a, b ∈ A∗. For the second assertion, assume that
La Rb = IA, i.e., that axb = x for all x ∈ A. Setting x = a−1 gives b = a−1, and now La Ra−1 = IA
if and only if a ∈ k∗1.
(cid:3)
If A is an associative Hurwitz algebra, then the map La ψ ∈ LA∗ ⋊ Aut(A) = G(A) has a pair
of triality components (La ψ, ψ). Thus the action (2) of [La ψ] ∈ PG(A) on [α, β] ∈ XA can be
written as
(3)
[La ψ] · [α, β] =(cid:2)La ψαψ−1 L−1
a , ψβψ−1 L−1
a (cid:3) .
For dim A = 2, this description of the groupoid X (A) is equivalent to the isomorphism criterion
1.12 in [17], applied to isotopes of quadratic ´etale algebras.
We conclude this section by introducing a numerical, easily computed isomorphism invariant.
For l a positive integer, let k∗l = {ρl ∈ k∗ ρ ∈ k∗} ⊂ k∗.
Proposition 11. Let A be a Hurwitz algebra of dimension 2l > 4, and α, β ∈ GL(A). Then the
pair ([det(α)], [det(β)]) ∈(cid:0)k∗/k∗l(cid:1)2
is an isomorphism invariant for Aα,β.
Proof. Assuming the existence of an isomorphism ϕ : Aα,β → Aγ,δ, we need to show that
det(α) det(γ)−1 ∈ k∗l and det(β) det(δ)−1 ∈ k∗l. For any x ∈ A, we have ϕ Lα(x) β = Lγϕ(x) δϕ
and hence ϕ Lα(x) βϕ−1 = Lγϕ(x) δ, so
det(cid:0)Lα(x)(cid:1) det(β) = det(cid:0)ϕ Lα(x) βϕ−1(cid:1) = det(cid:0)Lγϕ(x) δ(cid:1) = det(Lγϕ(x)) det(δ).
(4)
Since the maps La and Ra are proper similitudes with multiplier nA(a) for any a ∈ A∗, the
equation (4) implies that nA(α(x))l det(β) = nA(γϕ(x))l det(δ) whenever α(x) ∈ A∗. Hence,
det(β) det(δ)−1 ∈ k∗l. Similarly, the identity ϕ Rβ(y) αϕ−1 = Rδϕ(y) γ implies that det(α) det(γ)−1 ∈
k∗l.
(cid:3)
Let A and l be as in Proposition 11. For i, j ∈ k∗/k∗l, setting
X (A)i,j =n[α, β] ∈ X (A) ([det(α)], [det(β)]) = (i, j) ∈(cid:0)k∗/k∗l(cid:1)2o ⊂ X (A) ,
the groupoid X (A) can be written as a coproduct
(5)
X (A) = ai,j∈ k∗/k∗l
X (A)i,j .
Hence, any subcategory A ⊂ X (A) can be classified by classifying each of the subcategories
Ai,j = A ∩ X (A)i,j ⊂ X (A)i,j .
As for the real ground field, [R∗ : R∗l] = 2 for any even integer l > 2, the two cosets being
represented by 1 and −1, and the quotient projection R∗ → R∗/R∗l is given by ρ 7→ [sign(ρ)]. If
A is either H or O, α, β ∈ GL(A) and Aα,β = (A, ◦), then det(L◦
x) = det(Lα(x)) det(β). Since
det(Lα(x)) = n(α(x))(dim A)/2 > 0 for any x 6= 0, it follows that sign(det(L◦
x)) = sign(det(β)), and
similarly sign(det(R◦
here coincides with the "double sign" decomposition for real division algebras, introduced in [5].
x)) = sign(det(α)). This means that the decomposition X (A) =`i,j∈{−1,1}
X (A)i,j
3. Isotopes of Hamilton's quaternions
In this section , we give a more detailed account for isotopes of real Hurwitz algebras whose
underlying quadratic space is Euclidean. Such a Hurwitz algebra is isomorphic to either R, C, H,
or O, and the isotopes are precisely the real division algebras that are isotopic to a Hurwitz algebra.
The isotopes of C are all the two-dimensional real division algebras, and their classification has
been described in [7, 11]. While our goal is to study the isotopes of H, some arguments hold also in
the 8-dimensional case. Thus, let A be either H or O. Our main tool will be polar decomposition
of linear maps.
ISOTOPES OF HURWITZ ALGEBRAS
7
Any α ∈ GL(A) can be written as α = α′λ, where det(α′) = det(α) > 0 and λ ∈ C2 = {I, κ}.
Polar decomposition now yields α′ = ζδ, with ζ ∈ SO(A) and δ ∈ Pds(A). Hence α = ζδλ, and
this decomposition is unique [10, §14].
Passing to the projective setting, the above implies that every α ∈ PGL(A) uniquely determines
ζ ∈ SO(A)/(±I), δ ∈ SPds(A) = Pds(A) ∩ SL(A) and λ ∈ C2 such that α = [ζδλ]. As noted
in Remark 8, SO(A)/(±I) ≃ PGO+(A), and ζ can indeed be viewed as an element in PGO+(A)
instead of SO(A)/(±I). The cyclic group C2 acts on PGO+(A) by [ϕ]λ = [λϕλ] (λ ∈ C2, ϕ ∈
a ] in PGO+(A). We write [ϕ]−λ
GO+(A)). Note that [La]κ = [R¯a] = [R−1
a ] and [Ra]κ = [L¯a] = [L−1
for (cid:0)[ϕ]−1(cid:1)λ
=(cid:0)[ϕ]λ(cid:1)−1
.
Let α, β ∈ PGL(A), α = [ζδλ] and β = [ηǫµ], where ζ, η ∈ PGO+(A), δ, ǫ ∈ SPds(A) and
λ, µ ∈ C2. The action (2) of [ϕ] ∈ PGO+(A) on [α, β] ∈ XA is given by
[ϕ] · [α, β] =(cid:2)ϕ1ζδλϕ−1 , ϕ2ηǫµϕ−1(cid:3) =(cid:2)ϕ1ζϕ−λ(ϕλδϕ−λ)λ , ϕ2ηϕ−µ(ϕµǫϕ−µ)µ(cid:3) ,
and ϕ1ζϕ−λ, ϕ2ηϕ−µ ∈ PGO+(A) , ϕλδϕ−λ, ϕµǫϕ−µ ∈ SPds(A).
The group N (A)∗ acts on PGO+(A)2 by w · (ζ, η) = ([Rw]−1ζ , [Lw]η), and we denote the orbit
set of this action by PGO+(A)2/N (A)∗. The argument in the preceding paragraph shows that
the PGO+(A)-action on PGO+(A)2/N (A)∗ × SPds(A)2 × C2
2 by
(6)
is equivalent to the PGO+(A)-action (2) on XA, via the map
[ϕ] · ([ζ, η], (δ, ǫ), (λ, µ)) =(cid:0)(cid:2)ϕ1ζϕ−λ, ϕ2ηϕ−µ(cid:3) , (ϕλδϕ−λ, ϕµǫϕ−µ), (λ, µ)(cid:1)
m : PGO+(A)2/N (A)∗ × SPds(A)2 × C2
2 → XA , ([ζ, η], (δ, ǫ), (λ, µ)) 7→ [ζδλ, ηǫµ]
(i.e., ϕm = mϕ for all ϕ ∈ PGO+(A)). Hence, in particular, the groupoid
of the action (6) is isomorphic to X (A):
Y (A) = PGO+(A)(cid:0) PGO+(A)2/N (A)∗ × SPds(A)2 × C2
2(cid:1)
Proposition 12. The functor GA : Y (A) → X (A) defined by GA(y) = m(y) for y ∈ Y (A), and
GA(ϕ) = ϕ for morphisms ϕ ∈ PGO+(A), is an isomorphism of categories.
From here on we focus exclusively on the four-dimensional case, in which A ≃ H. Corollary 10
implies that every element ζ ∈ PGO+(H) has the form ζ = [La Rb] for some a, b ∈ H∗. Since H is
associative, it follows that if ϕ = [La Rb] then ϕ1 = [La], ϕ2 = [Rb] are triality components of ϕ.
Recall that N (H)∗ = H∗ acts on PGO+(H)2 by c · (ζ, η) = (cid:0)[R−1
for all a, b ∈ H, the following identities hold in PGO+(H)2/N (H)∗ = PGO+(H)2/H∗:
]ζ, [Lc]η(cid:1). Since La Rb = Rb La
c
[La Rb, Lc Rd] = c−1 · [La Rb, Lc Rd] = [Rc La Rb, Rd] = [La Rbc, Rd] .
Hence every element [ζ, η] ∈ PGO+(H)2/H∗ can be written on the form [ζ, η] = [La Rb, Rc], with
a, b, c ∈ H∗. Now the action of ϕ = [Ls Rt] ∈ PGO+(H) (s, t ∈ H∗) on ([ζ, η], (δ, ǫ), (λ, µ)) ∈ Y (H)
is given by
(7) ϕ · ([ζ, η], (δ, ǫ), (λ, µ)) =(cid:16)hLs La Rb L−λ
For (i, j) ∈ {−1, 1}2, set Y (H)i,j = G −1
Y (H)(−1)i,(−1)j . This gives the decomposition
s R−λ
t
, Rt Rc L−µ
s R−µ
A (X (H)i,j ) ⊂ Y (H). Note that ([ζ, η], (δ, ǫ), (κi, κj)) ∈
t i , (cid:0)ϕλδϕ−λ, ϕµǫϕ−µ(cid:1) , (λ, µ)(cid:17) .
Y (H) = ai,j∈{−1,1}
Y (H)i,j ,
and the action of ϕ = Ls Rt on each of the cofactors Y (H)i,j can be studied separately.
For ([ζ, η], (δ, ǫ), (I, I)) ∈ X (H)1,1, we have
ϕ · ([ζ, η], (δ, ǫ), (I, I)) =(cid:0)(cid:2)Ls La Rb L−1
s Rb R−1
s R−1
t Ls La L−1
, Rt Rc L−1
s R−1
s , Rt Rc R−1
t
In particular, if s = 1, t = b, so that ϕ = Rb, then
=(cid:0)(cid:2)R−1
t (cid:3) , (cid:0)ϕδϕ−1, ϕǫϕ−1(cid:1) , (I, I)(cid:1)
t (cid:3) , (cid:0)ϕδϕ−1, ϕǫϕ−1(cid:1) , (I, I)(cid:1) .
ϕ · ([ζ, η], (δ, ǫ), (I, I)) =(cid:0)[La, Rbcb−1 ] , (cid:0)Rb δ R−1
b
, Rb ǫ R−1
b (cid:1) , (I, I)(cid:1) .
8
ERIK DARP O
Hence every orbit in Y (H)1,1 contains an element of the form ([La, Rb], (δ, ǫ), (I, I)). Moreover,
the action of ϕ = [Ls Rt] on such an element is given by
s R−1
t
, Rt Rb L−1
s R−1
(8)
ϕ · ([La, Rb], (δ, ǫ), (I, I)) =(cid:0)(cid:2)Ls La L−1
which has the form ([Lc, Rd], (δ, ǫ), (I, I)) if and only if t = s−1, that is, ϕ = [Ls R−1
s ].
=(cid:0)(cid:2)R(st)−1 Lsas−1 , Rt−1bt(cid:3) , (ϕδϕ−1, ϕǫϕ−1), (I, I)(cid:1)
Denoting by Z = H∗/R∗(cid:0)(H∗/R∗)2 × SPds(H)2(cid:1) the groupoid of the action
s · ((a, b), (δ, ǫ)) = ((sas−1, sbs−1), (csδc−1
s , csǫc−1
s )) ,
t (cid:3) , (ϕδϕ−1, ϕǫϕ−1), (I, I)(cid:1)
the above implies that the functor H1,1 : Z → Y (H)1,1 defined by
H1,1((a, b), (δ, ǫ)) = ([La, Rb], (δ, ǫ), (I, I))
on objects, and
Hi,j (s) = [Ls R−1
s ]
on morphisms, is an equivalence of categories.
Next, consider ([ζ, η], (δ, ǫ), (κ, I)) ∈ Y (H)−1,1. In this case,
s R−1
ϕ · ([ζ, η], (δ, ǫ), (κ, I)) =(cid:0)(cid:2)Ls La Rb L−κ
=(cid:0)(cid:2)Ls La Lt R−1
Setting s = 1 and t = a−1 gives ϕ = R−1
a and
t
, Rt Rc L−1
s R−κ
s Rb Rs, Rt Rc R−1
t (cid:3) , (cid:0)ϕκδϕ−κ, ϕǫϕ−1(cid:1) , (κ, I)(cid:1)
t (cid:3) , (cid:0)ϕκδϕ−κ, ϕǫϕ−1(cid:1) , (κ, I)(cid:1)
[Ra]−1 · ([ζ, η], (δ, ǫ), (κ, I)) =(cid:0)[Rb, Raca−1 ] , (cid:0)La δ L−1
so the orbit of ([ζ, η], (δ, ǫ), (κ, I)) contains an element of the form ([Ra, Rb], (δ′, ǫ′), (κ, I)). Similarly
to the previous case with Y (H)1,1, the group elements ϕ ∈ PGO+(H) stabilising this form, so
that ϕ · ([Ra, Rb], (δ, ǫ), (κ, I)) = ([Rc, Rd], (δ′, ǫ′), (κ, I)) for some c, d ∈ H, are precisely those of
the form ϕ = [Ls R−1
s ], s ∈ H∗. Note that if ϕ = [Ls R−1
s ] then ϕκ = ϕ, hence
a ǫ Ra(cid:1) , (κ, I)(cid:1)
a , R−1
ϕ · ([Ra, Rb], (δ, ǫ), (κ, I)) =(cid:0)[Rsas−1 , Rsbs−1 ], (ϕδϕ−1, ϕǫϕ−1), (κ, I)(cid:1) .
Again, this shows that there is an equivalence of categories H−1,1 : Z → Y (H)−1,1 is given by
H−1,1(s) = [Ls R−1
s ] on morphisms and H−1,1((a, b), (δ, ǫ)) = ([Ra, Rb], (δ, ǫ), (κ, I)) on objects.
Similar computations can be made for Y (H)1,−1 and Y (H)−1,−1, in each case resulting in an
equivalence with the category Z . The results are summarised in Theorem 13 below.
For a ∈ H∗, set ca = [La R−1
a ] ∈ PGO+(H). Note that the kernel of the morphism H∗ →
PGO+(H), a 7→ ca is R∗1.
Theorem 13. Let (i, j) ∈ {−1, 1}2. Each of the subcategories Y (H)i,j of Y (H) is equivalent to
the groupoid Z = H∗/R∗(cid:0)(H∗/R∗)2 × SPds(H)2(cid:1) of the action
s · ((a, b), (δ, ǫ)) = ((sas−1, sbs−1), (csδc−1
s , csǫc−1
s )) .
An equivalence Hi,j : Z → Y (H)i,j is given by Hi,j (s) = cs for morphisms s ∈ H∗/R∗ and
H1,1((a, b), (δ, ǫ)) = ([La, Rb], (δ, ǫ), (I, I)) , H−1,1((a, b), (δ, ǫ)) = ([Ra, Rb], (δ, ǫ), (κ, I)) ,
H1,−1((a, b), (δ, ǫ)) = ([La, Lb], (δ, ǫ), (I, κ)) , H−1,−1((a, b), (δ, ǫ)) = ([La, Rb], (δ, ǫ), (κ, κ)) .
Remark 14. Theorem 13 shows, in particular, that X (H)i,j ≃ X (H)i′,j ′ for all (i, j), (i′, j′) ∈
{−1, 1}2.
The image of the monomorphism H∗/R∗ → PGO+(H), s 7→ cs is {ϕ ∈ PGO+(H) ϕ(R∗1) =
R∗1}, which can be identified with SO1(H) = {ϕ ∈ SO(H) ϕ(1) = 1} ≃ SO(1⊥
H ). Thus the
map f : H∗/R∗ → SO1(H), s 7→ cs is an isomorphism, inducing an action of SO1(H) on the
object set (H∗/R∗)2 × SPds(H)2 of Z , given by ϕ · ((a, b), (δ, ǫ)) = f −1(ϕ) · ((a, b), (δ, ǫ)) =
((ϕ(a), ϕ(b)), (ϕδϕ−1, ϕǫϕ−1)).
This allows for the following geometric interpretation of the category Z . Elements in H∗/R∗
are viewed as lines through the origin in H (that is, elements in the real projective space P(H)),
and δ ∈ SPds(A) is identified with the three-dimensional hyper-ellipsoid Eδ = {x ∈ H hx, δ(x)i =
ISOTOPES OF HURWITZ ALGEBRAS
9
1} ⊂ H. The set E = {Eδ δ ∈ SPds(H)} consists of all hyper-ellipsoids centered in the origin
and satisfying λ1λ2λ3λ4 = 1, where λi ∈ R>0, i = 1, 2, 3, 4 are the lengths of the principal axes
(which lie along the eigenspaces of δ). Moreover, H is identified with R4 in the natural way, and
1⊥
H ⊂ H with V = span{e2, e3, e4} ⊂ R4. Now objects in Z can be viewed as configurations
in R4 consisting of two lines a, b ∈ P(R4), and two hyper-ellipsoids Eδ, Eǫ ∈ E. A morphism
(a, b, Eδ, Eǫ) → (a′, b′, Eδ′ , Eǫ′) between two such configurations is an element ϕ ∈ SO(V ) ⊂
SO(R4) transforming one configuration to the other: (ϕ(a), ϕ(b), ϕ(Eδ ), ϕ(Eǫ)) = (a′, b′, Eδ′ , Eǫ′ ).
The above interpretation gives a picture of the category of isotopes of H in elementary geometric
terms, making it easy to read of whether or not two different isotopes of H are isomorphic. To
extract from this description a complete and irredundant list of representatives for the set of
isomorphism classes of isotopes of H, one would need to find a normal form for the described
configurations of lines and hyper-ellipsoids under the natural action of SO(V ). While this task
is in principle not difficult, it is by its nature quite technical, and we shall not pursue it further
here. We note, however, that the object set of the groupoid Z is a 24-dimensional manifold, with
isomorphism classes being orbits under a continuous action of a 3-dimensional real Lie group.
4. Composition algebras
In this section, as an example, we show how the theory of composition algebras fits into the
general setting of Section 2. This leads to a short and easy proof of results by Stampfli-Rollier
[20] on the structure four-dimensional composition algebras. As a bonus, our approach works also
in characteristic two.
Lemma 15. Let A = (A, nA) be a Hurwitz algebra, and α, β ∈ GL(A). The isotope Aα,β is a
composition algebra if and only if α, β ∈ GO(A, nA).
Proof. If α, β are similitudes, then Aα,β is a composition algebra with respect to the norm n =
µ(α)µ(β)nA. Conversely, suppose that Aα,β = (A, ◦) is a composition algebra with norm n. Let
a ∈ A be an element such that n(a) 6= 0, then b = n(a)−1(a ◦ a) has norm 1, and thus the
linear endomorphisms L◦
b of Aα,β are orthogonal with respect to n. Now the isotope
b )−1 is a Hurwitz algebra with norm n and identity element b ◦ b.4 It follows
B = (Aα,β)(R◦
that B is isotopic to A and thus, by Proposition 2, n = ρnA for some ρ ∈ k∗. Hence
b and R◦
b )−1,(L◦
nA(α(x))nA(β(y)) = nA(α(x)β(y)) = nA(x ◦ y) = ρ−1n(x ◦ y) = ρ−1n(x)n(y) = ρnA(x)nA(y)
for all x, y ∈ A. Inserting y = 1A into this equation gives nA(α(x)) = ρnA(β(1))−1nA(x), so α
is a similitude with muliplier ρnA(β(1))−1 with respect to nA. Similarly, β is a similitude with
multiplier ρnA(α(1))−1.
(cid:3)
Given a Hurwitz algebra A, let X c(A) ⊂ X (A) be the full subcategory formed by all images
[α, β] = FA(Aα,β) where Aα,β is a composition algebra. Lemma 15 implies that X c(A) =
PGO(A, nA)2/N ⊂ X (A).
For dim A > 2, the property of being a proper respectively improper similitude is retained
by the factors α, β of [α, β] ∈ X c(A) under the action of PG(A). This means that for each
(i, j) ∈ {−1, 1}2, the subset
(where we use the notational convention PGO±1(A) = PGO±(A)) is invariant under this action,
and the category X c(A) decomposes as
X c(i,j)(A) =(cid:0)PGOi(A) × PGOj(A)(cid:1) /N (A)∗ ⊂ X c(A)
X c(A) = ai,j∈{−1,1}
X c(i,j)(A) .
This refines the decomposition (5) for dim A > 4: X c(i,j)(A) ⊂ X c(A)i,j for all i, j ∈ {−1, 1}. If
−1 6∈ kl, l = (dim A)/2, then X c(i,j)(A) = X c(A)i,j , but whereas the four sets X c(i,j)(A) are
always distinct, X c(A)i,j = X c(A) for all i, j ∈ {−1, 1} in case −1 ∈ kl.
4This argument, by which any finite-dimensional composition algebra is shown to be isotopic to a Hurwitz
algebra with the same norm, is sometimes referred to as Kaplansky's trick, after [14, p. 957].
10
ERIK DARP O
If (A, nA) is a Euclidean space, the groups PGO(A) and PGO+(A) are canonically identified
with O(A)/(±IA) and SO(A)/(±IA) respectively. This, together with Theorem 7, gives a descrip-
tion of all finite-dimensional absolute valued algebras equivalent to Theorem 4.3 in [3].
We proceed to describe all composition algebras isotopic to a fixed quaternion algebra A =
(A, n) over k. Every [α, β] ∈ X c(A) can be written as [α, β] = [ζλ, ηµ] with ζ, η ∈ GO+(A) and
λ, µ ∈ C2. Corollary 10 now gives ζ = [La Rb], η = [Lc Rd] for some a, b, c, d ∈ A∗, and since
N (A) = A, we have [α, β] = [La Rb λ, Lc Rd µ] = [La Rbc λ, Rd] in XA. Thus, analogously with the
Euclidean case treated in Section 3, [α, β] can be written as [α, β] = [La Rb λ, Rc µ]. Moreover,
one reads off that [α, β] ∈ X c(det(λ),det(µ))(A).
Let [α, β] = [La Rb, Rc] ∈ X c(1,1)(A). Now [Rb] · [α, β] = [La, Rb−1cb], so every PGO+(A)-orbit
in X c(1,1)(A) contains an element of the form [Lb, Rb]. Similarly, every orbit in X c(−1,1)(A)
contains an element of the form [Ra κ, Rb], every orbit in X c(1,−1)(A) contains some [La, Lb κ],
and every orbit in X c(−1,−1)(A) contains an element of the form [La κ, Rb κ]. By computations
similar to (8), one proves that ϕ ∈ PGO+(A) stabilises each of these forms if and only if ϕ = cs
for some s ∈ A∗, and that
cs · [La, Rb] = [Ls La L−1
s , R−1
s Rb Rs] = [Lsas−1 , Rsbs−1 ] ,
s Ra Rs κ, R−1
s Rb Rs] = [Rsas−1 κ, Rsbs−1 ] ,
cs · [Ra κ, Rb] = [R−1
cs · [La, Lb κ] = [Ls La L−1
cs · [La κ, Rb κ] = [Ls La L−1
s , Ls Lb L−1
s κ, R−1
s κ] = [Lsas−1 , Lsbs−1 κ] ,
s Rb Rs κ] = [Lsas−1 κ, Rsbs−1 κ] .
This proves the following result, which gives an isomorphism criterion for all composition algebras
isotopic to A.
Proposition 16. Let A be a quaternion algebra. Each of the categories X c(i,j)(A) is equivalent
to the groupoid Z c(A) = A∗/k∗ (A∗/k∗)2 of the action s · (a, b) = (sas−1, sbs−1). Equivalences
H c
i,j : Z c(A) → X c(A)i,j are given by H c
i,j (s) = cs for morphisms, and
H c
1,1(a, b) = [La, Rb] ,
1,−1(a, b) = [La, Lb] ,
H c
H c
−1,1(a, b) = [Ra, Rb] ,
−1,−1(a, b) = [La, Rb] .
H c
A characterisation of all four-dimensional composition algebras over k for char k 6= 2, similar
to Proposition 16, is given by Stampfli-Rollier in Section 3 -- 4 of [20]. Her exposition also contains
explicit isomorphism criteria in terms of the parameters a, b. The Euclidean case, comprising all
composition algebras isotopic to H, that is, all four-dimensional absolute valued algebras, has also
been described by Ram´ırez ´Alvarez [18]. Forsberg, in his master thesis [9], refined that description
to give an explicit cross-section for the isomorphism classes of these algebras.
References
[1] A. A. Albert. Non-associative algebras. I. Fundamental concepts and isotopy. Ann. of Math. (2), 43:685 -- 707,
1942.
[2] Seidon Alsaody. An approach to finite-dimensional real division composition algebras through reflections. Bull.
Sci. Math., 139(4):357 -- 399, 2015.
[3] A. Calder´on, A. Kaidi, C. Mart´ın, A. Morales, M. Ram´ırez, and A. Rochdi. Finite-dimensional absolute-valued
algebras. Israel J. Math., 184:193 -- 220, 2011.
[4] J. A. Cuenca Mira. On one-sided division infinite-dimensional normed real algebras. Publ. Mat., 36(2A):485 --
488 (1993), 1992.
[5] E. Darpo and E. Dieterich. The double sign of a real division algebra of finite dimension greater than one.
Math. Nachr., 285(13):1635 -- 1642, 2012.
[6] Erik Darpo and Jos´e Maria P´erez Izquierdo. Autotopies and quasigroup identities: New aspects of non-
associative division algebras. Forum Math., 27(5):2691 -- 2745, 2015.
[7] E. Dieterich. Classification, automorphism groups and categorical structure of the two-dimensional real division
algebras. J. Algebra Appl., 4(5):517 -- 538, 2005.
[8] A. Elduque and H. C. Myung. On flexible composition algebras. Comm. Algebra, 21(7):2481 -- 2505, 1993.
[9] L. Forsberg. Four-dimensional absolute valued algebras. Master's thesis, Uppsala University, 2009. U.U.D.M.
Project Report 2009:9.
[10] F. R. Gantmacher. The theory of matrices. Vol. 1. Translated by K. A. Hirsch. Chelsea Publishing Co., New
York, 1959.
ISOTOPES OF HURWITZ ALGEBRAS
11
[11] M. Hubner and H. P. Petersson. Two-dimensional real division algebras revisited. Beitrage Algebra Geom.,
45(1):29 -- 36, 2004.
[12] N. Jacobson. Composition algebras and their automorphisms. Rend. Circ. Mat. Palermo, 7(2):55 -- 80, 1958.
[13] N. Jacobson. Basic Algebra II. W.H. Freeman and Company, second edition, 1985.
[14] I. Kaplansky. Infinite-dimensional quadratic forms admitting composition. Proc. Amer. Math. Soc., 4:956 -- 960,
1953.
[15] M.-A. Knus, A. Merkurjev, M. Rost, and J.-P. Tignol. The book of involutions, volume 44 of American
Mathematical Society Colloquium Publications. American Mathematical Society, Providence, RI, 1998.
[16] R. Moufang. Zur Struktur von Alternativkorpern. Math. Ann., 110(1):416 -- 430, 1935.
[17] H. P. Petersson. The classification of two-dimensional nonassociative algebras. Results Math., 37(1-2):120 -- 154,
2000.
[18] M. I. Ram´ırez ´Alvarez. On four-dimensional absolute-valued algebras. In Proceedings of the International
Conference on Jordan Structures (M´alaga, 1997), pages 169 -- 173, M´alaga, 1999. Univ. M´alaga.
[19] T. A. Springer and F. D. Veldkamp. Octonions, Jordan algebras and exceptional groups. Springer Monographs
in Mathematics. Springer-Verlag, Berlin, 2000.
[20] C. Stampfli-Rollier. 4-dimensionale Quasikompositionsalgebren. Arch. Math. (Basel), 40(6):516 -- 525, 1983.
[21] M. Zorn. Theorie der alternativen Ringe. Abh. Math. Sem. Hamburg, 8:123 -- 147, 1931.
Graduate School of Mathematics, Nagoya University, Furo-cho, Chikusa-ku, Nagoya, Japan
|
1803.10209 | 1 | 1803 | 2018-03-27T17:48:44 | A graded pullback structure of Leavitt path algebras of trimmable graphs | [
"math.RA",
"math.OA"
] | Motivated by recent results in graph C*-algebras concerning an equivariant pushout structure of the Vaksman-Soibelman quantum odd spheres, we introduce a class of graphs called trimmable. Then we show that the Leavitt path algebra of a trimmable graph is graded-isomorphic to a pullback algebra of simpler Leavitt path algebras and their tensor products. | math.RA | math |
A GRADED PULLBACK STRUCTURE OF
LEAVITT PATH ALGEBRAS OF TRIMMABLE GRAPHS
PIOTR M. HAJAC, ATABEY KAYGUN, AND MARIUSZ TOBOLSKI
Abstract. Motivated by recent results in graph C*-algebras concerning an equivariant
pushout structure of the Vaksman-Soibelman quantum odd spheres, we introduce a class
of graphs called trimmable. Then we show that the Leavitt path algebra of a trimmable
graph is graded-isomorphic to a pullback algebra of simpler Leavitt path algebras and
their tensor products.
Introduction
The goal of this paper is to introduce and apply the concept of a trimmable graph.
We begin by recalling the fundamental concepts of path algebras [5] and Leavitt path
algebras [7, 2, 1, 3]. Then we define a trimmable graph, and prove our main result: There
is a Z-graded algebra isomorphism from the Leavitt path algebra of a trimmable graph to
an appropriate pullback algebra. The graph C*-algebraic version of this result is proven
in [6], where it was used to analyze the generators of K-groups of quantum complex
projective spaces.
1. Leavitt path algebras
Definition 1.1. A graph Q is a quadruple (Q0, Q1, s, t) consisting of the set of vertices Q0,
the set of edges Q1, and the source and target maps s, t : Q1 → Q0 assigning to each edge
its source and target vertex respectively.
0, Q(cid:48)
1, s(cid:48), t(cid:48)) is a sub-graph of a graph Q = (Q0, Q1, s, t) iff
We say that a graph Q(cid:48) = (Q(cid:48)
1 ⊆ Q1 , and the source and target maps s(cid:48) and t(cid:48) are respective restrictions-
0 ⊆ Q0 , Q(cid:48)
Q(cid:48)
corestrictions of the source and target maps s and t. Furthermore, we say that two edges
are composable if the end of one of them is the beginning of the other. Now we can define
a path in a graph as a sequence of composable edges. The length of a path is the number
of edges it consists of, infinity included. We treat vertices as zero-length paths that begin
and end in themselves.
Definition 1.2. Let k be a field and Q a graph. The path algebra kQ is the k-algebra
whose underlying vector space has as its basis the set of all finite-length paths Path(Q).
The product is given by the composition of paths when the end of one path matches the
beginning of the other path. The product is defined to be zero otherwise.
1
2
P. M. HAJAC, A. KAYGUN, AND M. TOBOLSKI
One can check that the path algebra kQ is unital if and only if the set of vertices Q0 is
finite. Then the unit is the sum of all vertices. It is also straightforward to verify that
kQ is N-graded by the path length.
we create a new set Q∗
To define a Leavitt path algebra, we need ghost edges. For any graph Q = (Q0, Q1, s, t),
1 := {x∗ x ∈ Q1} and call its elements ghost edges. Now, the
source and the target maps for the extended graph (cid:98)Q :=(cid:0)Q0, Q1
1, s, t(cid:1) are defined
(cid:96) Q∗
as follows:
s(x) := s(x),
s(x∗) := t(x),
t(x) := t(x),
t(x∗) := s(x).
(1.1)
Definition 1.3. Let k be a field and Q a graph. The Leavitt path algebra Lk(Q) of a
graph Q is the path algebra of the extended graph (cid:98)Q divided by the ideal generated by the
relations:
(1) For all edges xi, xj ∈ Q1 , we have x∗
(2) For every vertex v ∈ Q0 whose preimage s−1(v) is not empty and finite, we have
i xj = δijt(xi).
(cid:88)
xx∗ = v.
x∈s−1(v)
1, subject to relations:
In other words, the Leavitt path algebra Lk(Q) of a graph Q is the universal k-algebra
generated by the elements v ∈ Q0 , x ∈ Q1 , x∗ ∈ Q∗
(L1) vivj = δijvi for all vi, vj ∈ Q0,
(L2) s(x)x = xt(x) = x for all x ∈ Q1,
(L3) t(x)x∗ = x∗s(x) = x∗ for all x∗ ∈ Q∗
1,
i xj = δijt(xi) for all xi, xj ∈ Q1, and
(L4) x∗
x∈s−1(v) xx∗ = v for all v ∈ Q0 such that s−1(v) is finite and nonempty.
(L5) (cid:80)
Furthermore, note that the N-grading of the path algebra k(cid:98)Q induces a Z-grading of
the Leavitt path algebra Lk(Q) by counting the length of any ghost edge as −1 (see [2,
Lemma 1.7]). Let us recall now the Graded Uniqueness Theorem [9, Theorem 4.8] that
shows the importance of this grading. We will need it in the next section.
Theorem 1.4 ([9]). Let Q be an arbitrary graph and k be any field. If A is a Z-graded
ring, and f : Lk(E) → A is a graded ring homomorphism with f (v) (cid:54)= 0 for every vertex
v ∈ Q0, then f is injective.
Recall that a vertex v ∈ Q0 is called a sink if s−1(v) = ∅. Next, let x = x1x2 . . . xn be
a path in Q. If the length of x is at least 1, and if s(x) = v = t(x) ∈ Q0, we say that x
is a closed path based at v. If in addition s(xi) (cid:54)= s(xj) for every i (cid:54)= j, then x is called a
cycle based at v.
Next, if v ∈ Q0 is either a sink or a base of a cycle of length 1 (a loop), then a singleton
set {v} ⊆ Q0 is a basic example of a hereditary subset of Q0 [3, Definition 2.0.5 (i)], and
it follows from Corollary 2.4.13 (i) in [3] (cf. [9, Theorem 5.7 (2)]) that
Lk(Q)/I(v) ∼= Lk(Q \ {v}).
(1.2)
LEAVITT PATH ALGEBRAS OF TRIMMABLE GRAPHS
3
Here the graph Q \ {v} is obtained by removing from Q the vertex v and every edge that
ends in v. In other words,
(Q \ {v})0 = Q0 \ {v}
and
(Q \ {v})1 = {x ∈ Q1 : t(x) (cid:54)= v}.
(1.3)
By Lemma 2.4.1 in [3] (cf. [9, Lemma 5.6]), we also know that
(cid:40) n(cid:88)
i=1
I(v) =
kixiy∗
i
(cid:12)(cid:12)(cid:12) n ≥ 1, xi , yi ∈ Path(Q), t(xi) = s(y∗
i ) = v
(cid:41)
.
(1.4)
2. Trimmable graphs
We are now ready for the main definition of the paper. Merely to focus attention, we
assume henceforth that graphs are finite, i.e. that the set of vertices and the set of edges
are both finite.
Definition 2.1. Let Q be a finite graph consisting of a sub-graph Q(cid:48) emitting at least one
edge to an external vertex v0 whose only outgoing edge x0 is a loop. We call such a graph
(Q(cid:48), v0)-trimmable iff all edges from Q(cid:48) to v0 begin in a vertex emitting an edge that ends
inside Q(cid:48).
In symbols, a trimmable graph is described as follows:
0 ∪ {v0},
Q0 = Q(cid:48)
s−1(v0) = {x0},
∀ v ∈ s(cid:0)t−1(v0) \ {x0}(cid:1) :
v0 (cid:54)∈ Q(cid:48)
t(x0) = v0 ,
0 , Q1 = Q(cid:48)
1 ∪ t−1(v0),
t−1(v0) \ {x0} (cid:54)= ∅,
s−1(v) \ t−1(v0) (cid:54)= ∅.
(2.1)
(2.2)
(2.3)
The condition for a trimmable graph guarantees the fact that when we remove the dis-
tinguised vertex v0, the resulting graph does not have new sinks. One can imagine a
(Q(cid:48), v0)-trimmable graph like this:
The following graph is a simple example of a trimmable graph:
.
v2
v1
/ v0
.
(2.4)
O
O
/
t
t
4
P. M. HAJAC, A. KAYGUN, AND M. TOBOLSKI
Note that the direction of the edge joining vertices v1 and v2 is important as the following
graph is no longer trimmable:
v2
v1
/ v0
.
(2.5)
We need the trimmability condition to guarantee the existence of maps given in the
lemma below. This lemma is an algebraic incarnation of a graph-C*-algebraic lemma
proved in [6]. The only difference between their proofs is that instead of using the Gauge
Uniqueness Theorem [9, Theorem 4.8] we use the Graded Uniqueness Theorem [4, Theo-
rem 2.3] (Theorem 1.4).
0, v0)-trimmable graph. Denote by Q(cid:48)(cid:48) the sub-graph of Q
Lemma 2.2. Let Q be a (Q(cid:48)
obtained by removing the edge x0. The following formulas define homomorphisms of al-
gebras:
(1) π1 : Lk(Q) → Lk(Q(cid:48)),
(2) π2 : Lk(Q(cid:48)(cid:48)) → Lk(Q(cid:48)),
π1(α) =
π2(α) =
0 ∪ Q(cid:48)
1 ∪ Q(cid:48)∗
1 ,
if α ∈ Q(cid:48)
otherwise.
0 ∪ Q(cid:48)
1 ∪ Q(cid:48)∗
1 ,
if α ∈ Q(cid:48)
otherwise.
0
α
α
0
(3) f : Lk(Q) → Lk(Q(cid:48)(cid:48)) ⊗ k[u, u−1],
α ⊗ 1
v0 ⊗ u
v0 ⊗ u−1
α ⊗ u
α ⊗ u−1
f (α) =
(4) δ : Lk(Q(cid:48)) → Lk(Q(cid:48)) ⊗ k[u, u−1],
α ⊗ 1
α ⊗ u
α ⊗ u−1
δ(α) =
if α ∈ Q0,
if α = x0,
if α = x∗
0,
if α ∈ Q1 \ {x0},
1 \ {x∗
if α ∈ Q∗
0}.
if α ∈ Q(cid:48)
0,
if α ∈ Q(cid:48)
1,
if α ∈ (Q(cid:48)
1)∗.
These morphisms are Z-graded for the standard grading on Lk(Q), Lk(Q(cid:48)), Lk(Q(cid:48)(cid:48)), and
the gradings on Lk(Q(cid:48))⊗ k[u, u−1] and Lk(Q(cid:48)(cid:48))⊗ k[u, u−1] given by the standard grading of
the rightmost tensorand. Furthermore, π1 and π2 are surjective, and f and δ are injective.
/
t
t
LEAVITT PATH ALGEBRAS OF TRIMMABLE GRAPHS
5
3. A graded pullback structure
To prove the theorem of the paper, we need a general lemma along the lines of [8,
Proposition 3.1]. We omit its routine proof.
Lemma 3.1. Let A1, A2, B and P be abelian groups. A commutative diagram of group
homomorphisms
A1
p1
q1
P
B
p2
q2
A2
is a pullback diagram if and only if the following conditions hold:
ker(p1) ∩ ker(p2) = {0},
q−1
1 (q2(A2)) = p1(P ),
p2(ker(p1)) = ker(q2).
(3.1)
(3.2)
(3.3)
Recall that to prove that an algebra P is a pullback algebra, one can proceed as fol-
lows. The first step is to establish the existence of a commutative diagram of algebra
homomorphisms as above. This implies that p1 and p2 define an algebra homomorphism
p into the pullback algebra of A1 and A2 over B. Then one only needs to prove that the
three conditions of Lemma 3.1 are satisfied to conclude that p is an isomorphism. Note
that (3.1) is equivalent to the injectivity of p, whereas the conjunction of (3.2) and (3.3)
is equivalent to the surjectivity of p.
Much as Lemma 2.2, the theorem of the paper is an algebraic incarnation of a graph-C*-
algebraic theorem proved in [6]. This time, the only difference between their proofs is that
instead of using [4, Lemma 3.1] we use [3, Lemma 2.4.1] (see (1.4), cf. [9, Lemma 5.6]).
Theorem 3.2. Let π1, π2, f and δ be as in Lemma 2.2. Then the commutative diagram
Lk(Q)
π1
f
Lk(Q(cid:48))
Lk(Q(cid:48)(cid:48)) ⊗ k[u, u−1]
δ
Lk(Q(cid:48)) ⊗ k[u, u−1]
π2⊗id
of graded algebra homomorphisms is a pullback diagram.
Representing pictorially the Leavitt path algebras by their respective graphs, the above
diagram becomes:
~
~
~
~
w
w
)
)
'
'
u
u
6
P. M. HAJAC, A. KAYGUN, AND M. TOBOLSKI
π1
δ
⊗
f
π2⊗id
⊗
.
Note that the only non-standard map in this diagram is f . It can be described verbally
by the assignment
vertex (cid:55)−→ vertex ⊗ 1,
v0-emitted edge (cid:55)−→ v0 ⊗ u,
other edge (cid:55)−→ other edge ⊗ u.
Acknowledgement
The authors are very grateful to Francesco D'Andrea for sharing his mathematical insight
with us, and for designing the pictures of graphs. It is also a pleasure to thank Søren
Eilers, Adam Skalski and Wojciech Szyma´nski for discussions. This work was partially
supported by NCN grant UMO-2015/19/B/ST1/03098.
References
[1] G. Abrams, Leavitt path algebras: the first decade, Bull. Math. Sci. 5 (1), 55 -- 120, 2015.
[2] G. Abrams, G. Aranda Pino, The Leavitt path algebra of a graph, J. Algebra 293 (2), 319 -- 334, 2005.
[3] G. Abrams, P. Ara, M. Siles Molina, Leavitt path algebras. Lecture Notes in Mathematics, 2191.
Springer, London, 2017.
[4] A. an Huef, I. Raeburn, The ideal structure of Cuntz-Krieger algebras, Ergodic Theory Dynam. Sys.
17, 611 -- 624, 1997.
[5] I. Assem, D. Simson, A. Skowro´nski, Elements of the representation theory of associative algebras.
Vol. 1. Techniques of representation theory. London Mathematical Society Student Texts, 65. Cam-
bridge University Press, Cambridge, 2006.
{
{
'
'
$
$
w
w
LEAVITT PATH ALGEBRAS OF TRIMMABLE GRAPHS
7
[6] Francesco D'Andrea, Piotr M. Hajac, Mariusz Tobolski, An equivariant pullback structure of
trimmable graph C*-algebras, arXiv:1712.08010.
[7] W. G. Leavitt, The module type of a ring. Trans. Am. Math. Soc. 103, 113 -- 130, 1962.
[8] G. K. Pedersen, Pullback and pushout constructions in C∗-algebra theory, J. Funct. Anal. 167 (2),
243 -- 344, 1999.
[9] M. Tomforde, Uniqueness theorems and ideal structure for Leavitt path algebras, J. Algebra 318,
270 -- 299, 2007.
(P. M. Hajac) Instytut Matematyczny, Polska Akademia Nauk, ul. ´Sniadeckich 8, Warszawa,
00-656 Poland
E-mail address: [email protected]
(A. Kaygun) Istanbul Technical University, Istanbul, Turkey.
E-mail address: [email protected]
(M. Tobolski) Instytut Matematyczny, Polska Akademia Nauk, ul. ´Sniadeckich 8, Warszawa,
00-656 Poland
E-mail address: [email protected]
|
1506.00736 | 2 | 1506 | 2016-09-23T15:26:57 | On The Kantor Product | [
"math.RA"
] | We study the algebra of bilinear multiplications of an $n$-dimensional vector space. In particular, we study the Kantor product of some well-known (associative, Lie, alternative, Novikov and some other) multiplications. | math.RA | math |
On The Kantor Product
Ivan Kaygorodov
Universidade Federal do ABC, CMCC, Santo Andr´e, Brazil.
Abstract. We study the algebra of bilinear multiplications of an n-dimensional vector space. In particu-
lar, we study the Kantor product of some well-known (associative, Lie, alternative, Novikov and some other)
multiplications.
1. Introduction
Kantor introduced the class of conservative algebras in [11]. This class includes some well-known algebras:
associative, Jordan, Lie, Leibniz and Zinbiel [13]. In the theory of conservative algebras of great importance is
the conservative algebra U (n) [12]. In the theory of Lie algebras U (n) plays a role analogous to the role of gln.
The space of the algebra U (n) is the space of all bilinear multiplications on the n-dimensional space Vn. To
define the operation of multiplication [ , ] in the algebra U (n) we fix a vector u ∈ Vn and for two multiplications
A, B ∈ U (n) and two elements x, y ∈ Vn we set
x ∗ y = [A, B](x, y) = A(u, B(x, y)) − B(A(u, x), y) − B(x, A(u, y)).
(1)
Some properties of the algebra U (2) were studied in [13, 14]. We say that the product of two multiplications
on n-dimensional vector space defined by (1), is the left Kantor product of these multiplications. In a similar
way we can define the right Kantor product and obtain similar results. We assume that the Kantor product is
the left Kantor product. The Kantor product of a multiplication · by itself is the Kantor square of · and it is
denoted by [·,·]. It gives us a map K from any variety V of algebras to some class K(V ).
The Kantor square of a multiplication · can be rewritten (see [11]) as the product of the left multiplication
Lu and the multiplication ·, as [Lu,·], where
[Lu,·](x, y) = u · (x · y) − (u · x) · y − x · (u · y) = [·,·](x, y).
The multiplication [Lu,·] plays an important role in the definition of a (left) conservative algebras [2, 11]. We
recall that an algebra A with a multiplication · is called a (left) conservative algebra if and only if there exist a
new multiplication ∗ such that
[La, [Lb,·]] = −[La∗b,·].
The main aim of this paper is to study the properties of the Kantor product of multiplications. One of the
central questions studied in this paper is the following:
Question. What identities does the class of algebras K(V ) satisfy if we know the identities of V ?
We give some particularly answer of this question for associative, (anti)commutative, Perm, Lie, Leibniz,
Zinbiel, left-commutative, bicommutative, Novikov, alternative, quasi-associative and quasi-alternative algebras;
we also describe the Kantor product of multiplications in associative dialgebras, duplicial, dual duplicial, As(2),
Poisson, generalized Poisson and Novikov-Poisson algebras. Finally, we study the Kantor square in some special
cases; in particular, the associative algebras with identities, nilpotent and right-nilpotent algebras, associative
algebras isomorphic to its Kantor square; and discuss the coincidence of derivations and automorphisms of the
algebra and its Kantor square. Here we can to formulate
Open problem. Is K(V ) a variety of algebras for some variety V ?
2. The Kantor square
In this section we leave technical and trivial proofs of lemmas. We are using the standard notation:
(a, b, c)∗ = (a ∗ b) ∗ c − a ∗ (b ∗ c), (a, b, c) = (ab)c − a(bc);
[a, b]∗ = a ∗ b − b ∗ a, [a, b] = ab − ba;
1
2
2.1. Associative algebras. The variety of associative algebras is defined by the identity
(cid:8)a,b,c [f (a, b, c)] = f (a, b, c) + f (b, c, a) + f (c, a, b).
(xy)z = x(yz).
Lemma 1. Let (A;·) be an associative algebra. Then (A; [·,·]) is an associative algebra.
2.2. (Anti)commutative algebras. The variety of (anti)commutative algebras is defined by the identity
where ǫ = 1 in the commutative case and ǫ = −1 in the anticommutative case.
Lemma 2. Let (A;·) be an (anti)commutative algebra. Then (A; [·,·]) is an (anti)commutative algebra.
xy = ǫyx,
2.3. Perm algebras. The variety of Perm algebras (see, for example, [3]) is defined by the identity
Lemma 3. Let (A;·) be a Perm algebra. Then (A; [·,·]) is a Perm algebra.
(xy)z = x(yz) = x(zy).
2.4. Lie algebras. The variety of Lie algebras is defined by the identities
xy = −yx, (xy)z + (yz)x + (zx)y = 0.
Lemma 4. Let (A;·) be a Lie algebra. Then [·,·] = 0.
2.5. Leibniz algebras. The variety of (left) Leibniz algebras (see, for example, [6]) is defined by the identity
Lemma 5. Let (A;·) be a (left) Leibniz algebra. Then [·,·] = 0.
x(yz) = (xy)z + y(xz).
2.6. Left-commutative algebras. The variety of left-commutative algebras (see, for example, [16]) includes
commutative-associative, bicommutative, Novikov, Zinbiel algebras and some other. This variety is defined by
the identity
x(yz) = y(xz).
Lemma 6. Let (A;·) be a left-commutative algebra. Then (A; [·,·]) is a left-commutative algebra.
2.7. Bicommutative algebras. The variety of bicommutative algebras (see, for example, [8]) is defined by the
identities
x(yz) = y(xz), (xy)z = (xz)y.
Lemma 7. Let (A;·) be a bicommutative algebra. Then (A; [·,·]) is an associative-commutative algebra.
2.8. Zinbiel algebras. The variety of (left) Zinbiel algebras (see, for example, [7]) is defined by the identity
Lemma 8. Let (A;·) be a (left) Zinbiel algebra. Then (A; [·,·]) is a (left) Zinbiel algebra.
x(yz) = (xy)z + (yx)z.
3
2.9. Novikov algebras. The variety of (left) Novikov algebras (see, for example, [9]) is defined by the identities
Lemma 9. Let (A;·) be a (left) Novikov algebra. Then (A; [·,·]) is a (left) Novikov algebra.
x(yz) = y(xz), (x, y, z) = (x, z, y).
2.10. Alternative algebras. The variety of alternative algebras (see, for example, [15]) is defined by the
identities
x2y = x(xy), xy2 = (xy)y.
(2)
It is also well known (see, for example, [22]) that an alternative algebra is flexible:
and satisfies the Moufang identities:
(xy)x = x(yx);
and the following identities hold:
x(yzy) = ((xy)z)y, (yzy)x = y(z(yx)), (xy)(zx) = x(yz);
(x, y, z) = −(y, x, z), (x, y, z) = −(x, z, y),
(x, xy, z) = (x, y, z)x, (x, yx, z) = x(x, y, z).
The main example of a non-associative alternative algebra is a Cayley -- Dickson algebra C [22]. Let F be
a field of characteristic 6= 2. It is an algebra C with the basis e0 = 1, e1, . . . , e7 and the following multiplication
table:
(3)
(4)
e3
e2
e1
1
αe2
e3
α · 1
e1
−βe1
β · 1
−e3
e2
βe1 −αβ · 1
e3 −αe2
−e6
−e7
−e5
e4
e5 −αe4 −e7
−αe6
βe5
e6
e7
−βe4
e7
αe6 −βe5
αβe4
e6
−e7
βe4
−βe5
−γe2
γe3
e4
e5
e5
αe4
e6
e7
e7
αe6
γ · 1 −γe1
γe1 −αγ · 1
γe2
γe3 −αγe2
e7
−αe6
βe5
−αβe4
−γe3
−αγe2
−γe3 −βγ · 1 −βγ · e1
αβγ · 1
βγe1
Theorem 10. Let (A;·) be an alternative algebra. Then (A; [·,·]) is a flexible algebra. Furthermore,
1) (A; [·,·]) is an alternative algebra if and only if A satisfies the identity
(x, u, (x, u, y)) = 0;
2) (A; [·,·]) is a noncommutative Jordan algebra if and only if A satisfies the identity
[LuLxLuLx, RuRx] = [Lxuxu, Rux];
3) (A; [·,·]) is a Jordan algebra if (A,·) is a commutative alternative algebra;
4) (C; [·,·]) is alternative for a Cayley -- Dickson algebra C, if and only if u = u0 · 1.
Proof. It is easy to see that
Now, we can see
a ∗ b = u(ab) − (ua)b − a(ub) = (au)b − 2a(ub).
(x ∗ y) ∗ x − x ∗ (y ∗ x) = ((xu)y − 2x(uy)) ∗ x − x ∗ ((yu)x − 2y(ux)) =
(((xu)y)u)x − 2((x(uy))u)x − 2((xu)y)(ux) + 4(x(uy))(ux)−
(xu)((yu)x) + 2(xu)(y(ux)) + 2x(u((yu)x)) − 4x(u(y(ux))) =
x(uyu)x − 2((x, u, y)u)x − 2x(uyu)x − 2x(uyu)x + 2((xu)y, u, x) + 4x(uyu)x−
x(uyu)x − 2(xy)(y, u, x) + 2x(uyu)x + 2x(uyu)x + 2x(u(y, u, x)) − 4x(uyu)x =
2[((xu)y, u, x) − (x, u, x(uy)) − ((xy)(y, u, x) − x(u(y, u, x)))] =
2[((x, u, y), u, x) − ((x, u, y), u, x)] = 0.
4
It follows that (A,∗) is a flexible algebra.
1) It is easy to see that a flexible algebra is alternative if and only if it satisfies the first identity from (2).
We have
(x ∗ x) ∗ y − x ∗ (x ∗ y) =
−(xuxu)y + 2(xux)(uy) − (xu)((xu)y) + 2x(u((xu)y)) + 2(xu)(x(uy)) − 4x(u(x(uy))) =
−2(xu)2y − 2x(u(x(uy))) + 2x(u((xu)y)) + 2(xu)(x(uy)) = (x, u, (x, u, y)).
Now, the multiplication [·,·] is alternative if and only if (x, u, (x, u, y)) = 0.
Jordan identity: (x2, y, x) = 0.
2) Note that an algebra B is a non-commutative Jordan algebra if and only if it is flexible and it satisfies the
Obviously,
((x ∗ x) ∗ y) ∗ x − (x ∗ x) ∗ (y ∗ x) =
(−(xuxu)y + 2(xux)(uy)) ∗ x + (xux) ∗ ((yu)x − 2y(ux)) =
−(((xuxu)y)u)x + 2((xuxu)y)(ux) + 2(((xux)(uy))u)x − 4((xux)(uy))(ux)+
(xuxu)((yu)x) − 2(xux)(u((yu)x)) − 2(xuxu)(y(ux)) + 4(xux)(u(y(ux))) =
2(((xuxu)y)(ux) − (xuxu)(y(ux)) + ((x(u(x(uy))))u)x − x(u(x(u((yu)x)))).
Now, [·,·] is a noncommutative Jordan multiplication if and only if
[LxLuLxLu, RuRx] = [Lxuxu, Rux].
3) It is easy to see that if A is a commutative alternative algebra then we have
((xuxu)y)(ux) − (xuxu)(y(ux)) + (((xux)(uy))u)x − (xux)(u((yu)x)) =
(xu)((xu)((xu)y))) − (xu)((xu)((xu)y))) + x(u(x(u(x(uy))))) − x(u(x(u(x(uy))))) = 0.
4) If (C,∗) is an alternative algebra for every u then A satisfies (4). Note that for the elements ei1 , ei2, ei3 ,
It follows that [·,·] is non-commutative Jordan and from Theorem 2 we infer that [·,·] is Jordan.
where eik eil 6= ǫeim (where ǫ is some element from the ground field), we have
(ei1 , ei2, (ei1 , ei2 , ei3)) = 2(ei1 )2(ei2 )2ei3 .
Such triple (i1, i2, i3) we call a g-triple. It is easy to see that if (i, j, k) is not a g-triple then the subalgebra
generated by ei, ej, ek is a two-generated subalgebra, and by the Artin theorem this subalgebra is associative,
i. e., (ei, ej, (ei, ej, ek)) = 0. Now, for the element u = u0 · 1 + u1e1 + . . . + u7e7 we have (ei, u, (ei, u, ej)) = 0 if
and only if
(ukek)2 = 0. It is equivalent to the following system
P
k, (i,j,k) is a g-triple
αu2
1
1 +βu2
αu2
2
1 +βu2
αu2
2
αu2
1
+βu2
+βu2
−αβu2
3 +γu2
4
+γu2
4
−αβu2
2 −αβu2
2 −αβu2
3
3 +γu2
3
+γu2
−αβu2
−αβu2
4 −αβu2
4 −αβu2
5
5
−βγu2
5 −βγu2
6
6
+αβγu2
+αβγu2
= 0,
= 0,
7 = 0,
7 = 0,
= 0,
7 = 0,
7 = 0.
−βγu2
5 −βγu2
6 +αβγu2
6 +αβγu2
Calculating, we obtain u1 = √βγu7, u6 = √−αu7, u2 = u3 = u4 = u5 = 0.
we can find that u7 = 0 and u = u0 · 1.
The theorem is proved.
Now, from the relation (4) by simple calculations (for example, for x = e1 +e2, y = e1 and x = e2 +e6, y = e6)
2.11. Quasi-associative algebras. Quasi-associative algebras (see, for example, [4]) is defined by the identities
5
(x, y, z) + (y, z, x) + (z, x, y) = 0,
(x, y, z) = α[y, [x, z]],
where α is a fixed element in the ground field F. It is known [4] that an algebra (A,·) is quasi-associative if and
only if there exist an associative algebra A with the new multiplication, such that for some λ ∈ F :
Lemma 11. Let (A;·) be a quasi-associative algebra. Then (A; [·,·]) is a quasi-associative algebra.
x · y = λxy + (1 − λ)yx.
2.12. Quasi-alternative algebras. Quasi-alternative algebras (see, for example, [4]) is defined by the identities
(x, y, x) = 0,
(x, x, y) = α[x, [x, y]],
where α is a fixed element from the ground field F. It is known [4] that an algebra (A,·) is a quasi-alternative
algebra if and only there exist an alternative algebra A with new multiplication, such that for some λ ∈ F :
Lemma 12. Let (A;·) be a quasi-alternative algebra. Then (A; [·,·]) is a flexible algebra.
x · y = λxy + (1 − λ)yx.
2.13. Associative dialgebras. The variety of associative dialgebras (see, for example, [20]) is defined by the
identities
(x ⊢ y) ⊢ z = (x ⊣ y) ⊢ z, x ⊣ (y ⊢ z) = x ⊣ (y ⊣ z),
(x ⊢ y) ⊢ z = x ⊢ (y ⊢ z), (x ⊣ y) ⊣ z = x ⊣ (y ⊣ z), (x ⊢ y) ⊣ z = x ⊢ (y ⊣ z).
Lemma 13. Let (A;⊢,⊣) be an associative dialgebra. Then (A; [⊢,⊣]) is an associative algebra.
2.14. Duplicial algebras. The variety of duplicial algebras (see, for example, [18]) is defined by the identities
(x ≺ y) ≺ z = x ≺ (y ≺ z),
(x ≻ y) ≺ z = x ≻ (y ≺ z),
(x ≻ y) ≻ z = x ≻ (y ≻ z).
Lemma 14. Let (A;≺,≻) be a duplicial algebra. Then (A; [≻,≺]) is an associative algebra.
2.15. Dual duplicial algebras. The variety of dual duplicial algebras (see, for example, [23]) is defined by the
identities
(x ≺ y) ≺ z = x ≺ (y ≺ z), (x ≻ y) ≺ z = x ≻ (y ≺ z), (x ≻ y) ≻ z = x ≻ (y ≻ z),
x ≺ (y ≻ z) = (x ≺ y) ≻ z = 0.
Lemma 15. Let (A;≺,≻) be a dual duplicial algebra. Then [≻,≺] = 0 and (A; [≺,≻]) is a 2-nilpotent algebra.
2.16. As(2)-algebras. The variety of As(2)-algebras (see, for example, [23]) is defined by the identities
(x ◦ y) · z = x ◦ (y · z), (x · y) ◦ z = x · (y ◦ z),
(x ◦ y) ◦ z = x ◦ (y ◦ z), (x · y) · z = x · (y · z).
Lemma 16. Let (A;·,◦) be a As(2)-algebra. Then (A; [·,◦]) and (A; [◦,·]) are associative algebras.
2.17. Commutative tridendriform algebra. The variety of commutative tridendriform algebras (see, for
example, [17]) is defined by the identities
6
x · y = y · x, (x · y) · z = x · (y · z),
(x ≺ y) ≺ z = x ≺ (y ≺ z) + x ≺ (z ≺ y),
(x · y) ≺ z = x · (y ≺ z).
Lemma 17. Let (A;·,≺) be a commutative tridendriform algebra. Then (A; [≺,·]) is a commutative algebra
and (A; [·,≺]) is a right Zinbiel algebra.
2.18. Poisson algebras. The variety of Poisson algebras (see, for example, [19]) is defined by the identities
xy = yx, (xy)z = x(yz),{xy, z} = {x, z}y + x{y, z},
{x, y} = −{y, x},{{x, y}, z} + {{y, z}, x} + {{z, x}, y} = 0.
Theorem 18. Let (A;·,{,}) be a Poisson algebra. Then [{,},·] = 0 and (A; [·,{,}]) is a Lie algebra.
Proof. In the first case, we have
In the second case,
a ∗ b = {u, ab} − {u, a}b − a{u, b} = 0.
and
a ∗ b = u{a, b} − {ua, b} − {a, ub} = −a{u, b} − b{a, u} − u{a, b} = −b ∗ a.
(a ∗ b) ∗ c + (b ∗ c) ∗ a + (c ∗ a) ∗ b =
−(a{u, b} + {a, u}b + {a, b}u) ∗ c
−(b{u, c} + {b, u}c + {b, c}u) ∗ a
−(c{u, a} + {c, u}a + {c, a}u) ∗ b =
a{u, b}{u, c} + {a{u, b}, u}c + {a{u, b}, c}u+
{a, u}b{u, c} + {{a, u}b, u}c + {{a, u}b, c}u+
{a, b}u{u, c} + {{a, b}u, u}c + {{a, b}u, c}u+
b{u, c}{u, a} + {b{u, c}, u}a + {b{u, c}, a}u+
{b, u}c{u, a} + {{b, u}c, u}a + {{b, u}c, a}u+
{b, c}u{u, a} + {{b, c}u, u}a + {{b, c}u, a}u+
c{u, a}{u, b} + {c{u, a}, u}b + {c{u, a}, b}u+
{c, u}a{u, b} + {{c, u}a, u}b + {{c, u}a, b}u+
{c, a}u{u, b} + {{c, a}u, u}b + {{c, a}u, b}u =
a{u, b}{u, c} + {u, b}{a, u}c + ca{{u, b}, u} + au{{u, b}, c} + {u, b}{a, c}u+
{a, u}b{u, c} + {{a, u}, u}bc + {a, u}{b, u}c + bu{{a, u}, c} + {a, u}{b, c}u+
{a, b}u{u, c} + {{a, b}, u}uc + {a, b}{u, c}u + uu{{a, b}, c}+
b{u, c}{u, a} + {b, u}{u, c}a + ba{{u, c}, u} + {{u, c}, a}bu + {u, c}{b, a}u+
{b, u}c{u, a} + {b, u}{c, u}a + ca{{b, u}, u} + {b, u}{c, a}u + {{b, u}, a}cu+
{b, c}u{a, u} + {{b, c}, u}ua + {b, c}{u, a}u + uu{{b, c}, a}+
c{u, a}{u, b} + {c, u}{u, a}b + cb{{u, a}, u} + {c, b}{u, a}u + cu{{u, a}, b}+
{c, u}a{u, b} + {{c, u}, u}ab + {c, u}{a, u}b + {c, u}{a, b}u + {{c, u}, b}au+
{c, a}u{u, b} + {{c, a}, u}ub + {c, a}{u, b}u + uu{{c, a}, b} =
(a{u, b}{u, c} + {c, u}a{u, b}) + ({u, b}{a, u}c + {a, u}{b, u}c)+
(ca{{u, b}, u} + ca{{b, u}, u}) + ({u, b}{a, c}u + {c, a}{u, b}u)+
7
({a, u}b{u, c} + b{u, c}{u, a}) + ({{a, u}, u}bc + cb{{u, a}, u})+
({a, u}{b, c}u + {b, c}u{a, u}) + ({a, b}u{u, c} + {c, u}{a, b}u)+
({a, b}{u, c}u + {b, u}{u, c}a) + (ba{{u, c}, u} + {{c, u}, u}ab)+
({b, u}c{u, a} + {c, u}{u, a}b) + ({b, u}{c, a}u + {c, a}u{u, b})+
({b, c}{u, a}u + {c, b}{u, a}u) + ({c, u}{u, a}b + {c, u}{a, u}b)+
[au{{u, b}, c} + au{{b, c}, u} + au{{c, u}, b}]+
[bu{{a, u}, c} + bu{{u, c}, a} + bu{{c, a}, u}]+
[cu{{a, b}, u} + cu{{u, a}, b} + cu{{b, u}, a}]+
[uu{{a, b}, c} + uu{{b, c}, a} + uu{{c, a}, b}] = 0.
The theorem is proved.
2.19. Generalized Poisson algebras. The variety of unital generalized Poisson algebras (see, for example,
[1]) is defined by the identities
xy = yx, (xy)z = x(yz),{xy, z} = {x, z}y + x{y, z} + D(z)xy, D(x) = {1, x},
{x, y} = −{y, x},{{x, y}, z} + {{y, z}, x} + {{z, x}, y} = 0.
Theorem 19. Let (A;·,{,}) be a generalized Poisson algebra. Then (A; [{,},·]) is an associative-commutative
algebra, and (A; [·,{,}]) is a Lie algebra.
Proof. In the first case, we have
and
In the second case,
a ∗ b = {u, ab} − {u, a}b − a{u, b} = −D(u)ab = b ∗ a
(a ∗ b) ∗ c = D(u)2abc = a ∗ (b ∗ c).
a ∗ b = u{a, b} − {ua, b} − {a, ub} =
Here we use the proof of Theorem 18. It is easy to see that
−a{u, b} − b{a, u} − u{a, b} − D(b)ua + D(a)ub = −b ∗ a.
(cid:8)a,b,c [
(cid:8)a,b,c [
(a ∗ b) ∗ c] =(cid:8)a,b,c [({b, u}a + {b, a}u + {u, a}b − D(b)ua + D(a)ub) ∗ c] =
{c, u}{b, u}a + {c, b}{b, u}a + {c,{b, u}}au − {b, u}D(c)au + {u, a}{b, u}c +
{u,{b, u}}ac − {b, u}D(u)ac − {b, u}D(c)au + {b, u}D(a)cu + D({b, u})acu +
{c, u}{b, a}u + {c, u}{b, a}u + {c,{b, a}}u2 − {b, a}D(c)u2 + {u,{b, a}}cu −
{b, a}D(u)cu − {b, a}D(c)u2 + {b, a}D(u)cu + D({b, a})cu2 + {c, u}{u, a}b +
{c, b}{u, a}u + {c,{u, a}}bu − {u, a}D(c)bu + {u,{u, a}}bc + {u, b}{u, a}c −
{u, a}D(u)bc − {u, a}D(c)bu + {u, a}D(b)cu + D({u, a})bcu − {c, u}D(b)au −
{c, D(b)}au2 − {c, u}D(b)ua − {c, a}D(b)u2 + 2D(b)D(c)au2 − {u, a}D(b)cu −
{u, D(b)}acu + 2D(b)D(u)acu + D(b)D(c)au2 − D(D(b))acu2 − D(b)D(u)acu −
D(a)D(b)acu2 + {c, u}D(a)bu + {c, D(a)}bu2 + {c, u}D(a)bu + {c, b}D(a)u2 −
2D(a)D(c)bu2 + {u, D(a)}bcu + {u, b}D(a)cu − 2D(a)D(u)bcu − D(a)D(c)bu2 +
D(D(a))bcu2 + D(a)D(u)bcu + D(a)D(b)cu2].
By the proof of Theorem 18, we can conclude that the sum of all elements without D is zero. Now, we have
8
(cid:8)a,b,c [
(cid:8)a,b,c [
(a ∗ b) ∗ c] =
−{b, u}D(c)au − {b, u}D(u)ac − {b, u}D(c)au + {b, u}D(a)cu + D({b, u})acu −
{b, a}D(c)u2 − {b, a}D(u)cu − {b, a}D(c)u2 + {b, a}D(u)cu + D({b, a})cu2 −
{u, a}D(c)bu − {u, a}D(u)bc − {u, a}D(c)bu + {u, a}D(b)cu + D({u, a})bcu −
{c, u}D(b)au − {c, D(b)}au2 − {c, u}D(b)ua − {c, a}D(b)u2 + 2D(b)D(c)au2 −
{u, D(b)}acu + 2D(b)D(u)acu + D(b)D(c)au2 − D(D(b))acu2 − D(b)D(u)acu −
{u, a}D(b)cu − D(a)D(b)acu2 + {c, u}D(a)bu + {c, D(a)}bu2 + {c, u}D(a)bu +
{c, b}D(a)u2 − 2D(a)D(c)bu2 + {u, D(a)}bcu + {u, b}D(a)cu − 2D(a)D(u)bcu −
D(a)D(c)bu2 + D(D(a))bcu2 + D(a)D(u)bcu + D(a)D(b)cu2].
It is easy to see that
(cid:8)a,b,c
(cid:8)a,b,c
(cid:8)a,b,c
(cid:8)a,b,c
(cid:8)a,b,c
(cid:8)a,b,c
[D(D(a))bcu2 − D(D(b))acu2] = 0,
[D({b, u})acu + D({u, a})bcu] = 0,
[{u, D(a)}bcu − {u, D(b)}acu] = 0,
[D({b, a})cu2 − {c, D(b)au2 + {c, D(a)}bu2] = 0,
[−2{b, a}D(c)u2 − {c, a}D(b)u2 + {c, b}D(a)u2] = 0,
[{b, u}D(c)au + {u, a}D(c)bu + {c, u}D(b)au + {u, c}D(a)bu] = 0.
Obviously, (cid:8)a,b,c [(a ∗ b) ∗ c] = 0 and [·,{,}] is a Lie algebra. The theorem is proved.
2.20. Novikov-Poisson algebras. The variety of left Novikov-Poisson algebras is defined by the identities
xy = yx, (xy)z = x(yz),
x ◦ (y ◦ z) = y ◦ (x ◦ z), (x, y, z)◦ = (x, z, y)◦,
x ◦ (yz) = (x ◦ y)z, (xy) ◦ z − x(y ◦ z) = (xz) ◦ y − x(z ◦ y).
Theorem 20. Let (A;·,◦) be a left Novikov-Poisson algebra. Then (A; [·,◦]) is a left Novikov algebra and
(A; [◦,·]) is an associative-commutative algebra.
Proof. Firstly, we have
Hence,
and
Secondly,
Therefore,
a ∗ b = u(a ◦ b) − (ua) ◦ b − a ◦ (ub) = −(ua) ◦ b.
a ∗ (b ∗ c) = (ua) ◦ ((ub) ◦ c) = (ub) ◦ ((ua) ◦ c) = b ∗ (a ∗ c),
(a, b, c)∗ = (a ∗ b) ∗ c − a ∗ (b ∗ c) = (u((ua) ◦ b)) ◦ c − (ua) ◦ ((ub) ◦ c) =
((ua) ◦ (ub)) ◦ c − (ua) ◦ ((ub) ◦ c) = (ua, ub, c)◦ = (ua, c, ub)◦ =
= ((ua) ◦ c) ◦ (ub) − (ua) ◦ (c ◦ (ub)) =
u(ua, c, b)◦ = u(ua, b, c)◦ = (a ∗ c) ∗ b − a ∗ (c ∗ b) = (a, c, b)∗.
a ∗ b = u ◦ (ab) − (u ◦ a)b − a(u ◦ b) = −u ◦ (ab) = b ∗ a.
The theorem is proved.
(a ∗ b) ∗ c = u ◦ ((u ◦ (ab))c) = u ◦ (u ◦ (abc)) = u ◦ (a(u ◦ (bc))) = a ∗ (b ∗ c).
Similarly, the variety of right Novikov-Poisson algebras may be defined (see, for example, [21]). It is easy to
prove the following theorem:
9
Theorem 21. Let (A;·,◦) be a right Novikov-Poisson algebra. Then (A; [·,◦]) is a right Novikov algebra and
(A; [◦,·]) is a commutative algebra.
3. The Kantor square of algebras of special type.
Here we study some special cases of the Kantor square. For an algebra A := (A;·) its the Kantor square
(A; [·,·]) we denote by (A,∗). We denote the Kantor square for a fixed element u by (A,∗u). We consider the
relations between the ideals in A and (A,∗), the relations between an associative algebra A with polynomial
identity and its the Kantor square. Moreover, the relations between the nilpotency and right nilpotency in A
and (A,∗) are investigated.
3.1. Ideals in the Kantor product.
Theorem 22. Let I be an ideal of A. Then (I,∗) is an ideal of (A,∗), but the converse statement is not true
in general.
Proof. It is easy to see that if i ∈ I and a ∈ A then
i ∗ a = u(ia) − (ui)a − i(ua) ∈ I and a ∗ i = u(ai) − (ua)i − a(ui) ∈ I.
It follows that I is an ideal of (A,∗).
Conversely, we can consider the trivial case, where for an algebra A has zero Kantor square (for example,
Lie or Leibniz algebra) and every subspace of A is an ideal of (A,∗). For the non-trivial case (nonzero Kantor
product), we can consider the following associative algebra: A1 ⊕ A2 is the direct sum of the matrix algebras
of order 2. Here, if ei is the unit of Ai then the subspace generated by A1 and e2 is an ideal of (A,∗e1 ), but is
not an ideal of A1 ⊕ A2. The theorem is proved.
3.2. Associative algebras with polynomial identity. Given a polynomial f in n variables, we define
f∗(x1, . . . , xn) as the value of f in (A,∗), where x1, . . . , xn are some elements in A.
Theorem 23. Let (A;·) be an associative algebra that satisfies the polynomial identity f (x1, . . . , xn). Then
there exists an identity g such that A and (A,∗) satisfy g.
Proof.
It is easy to see that if A satisfies the identity f (x1, . . . , xn) then A satisfies the identity
g(x1, . . . , xn, z) = f (x1, . . . , xn)z. By Theorem 1, the multiplication in algebra (A,∗) is defined by x∗ y = −xuy.
Now, we can calculate the element g∗(x1, . . . , xn, z) in (A,∗). Obviously, it is (−1)nf (x1u, . . . , xnu)z which
amounts to zero in A. It follows that (A,∗) satisfies the identity g. The theorem is proved.
One of the most popular identity in the associative algebras is the standard polynomial identity of degree n:
sn(x1, . . . , xn) = X
σ∈Sn
(−1)σxσ(1) . . . xσ(n).
Theorem 24. Let (A;·) be an associative algebra that satisfies sn. Then (A,∗) satisfies sn+1.
Proof. It is easy to see that the standard polynomial of degree n + 1 may be written as
sn+1(x1, . . . , xn+1) =
X
( X
n
i=1
σ ∈ Sn+1,
σ(n + 1) = i
(−1)σxσ(1) . . . xσ(n)xσ(n+1)=i) =
=
n
X
i=1
(ǫisn(x1, . . . , xi, . . . , xn+1)xi), ǫi = ±1.
Now, by the proof of Theorem 23, (A,∗) satisfies the standard polynomial identity of degree n + 1. The theorem
is proved.
10
3.3. Nilpotent algebras. For the nilpotent algebras, we can prove the following theorem.
Lemma 25. Let (A;·) be a nilpotent algebra of nilpotency index n. Then (A,∗) is a nilpotent algebra of
nilpotency index ≤ [n/2] + 1.
Proof. Obviously, every product of the form x1 ∗ x2 ∗ . . . ∗ xt (with some order of brackets) is a sum of
multiplications of the form y1y2 ··· y2t−1 (with some order of brackets). Now, it is easy to see that (A,∗) is
nilpotent and its index of nilpotency ≤ [n/2] + 1. The Lemma is proved.
3.4. Right-nilpotent algebras. An algebra A is called right-nilpotent (or left-nilpotent) of nilpotency index
n if it satisfies the identity
Curiously, an analogue of the Theorem 25 is not true for the right-nilpotent algebras.
(. . . (x1x2) . . .)xn = 0
( or x1(. . . (xn−1xn) . . .) = 0).
Theorem 26. There exists a right nilpotent algebra (A;·) such that (A,∗u) is not right nilpotent, but (A,∗u)
is solvable.
Proof. An algebra A is right alternative if the following identity holds in A:
(x, y, y) = 0.
It is interesting fact that in contrast to the algebras of many well-studied classes (Jordan, alternative, Lie and
so on) a right nilpotent right alternative algebra need not be non-nilpotent. The corresponding example of a
five-dimensional right nilpotent but not nilpotent algebra belongs to Dorofeev [5]. Its basis is {a, b, c, d, e}, and
the multiplication is given by (zero products of basis vectors are omitted)
It is easy to see that
ab = −ba = ae = −ea = db = −bd = −c, ac = d, bc = e.
Obviously, c = (. . . (c ∗a b) ∗a . . .) ∗a b 6= 0, and (A,∗a) is not right-nilpotent. It is easy to see that A2 ⊆ hc, d, ei
and A ∗a A ⊆ hc, d, ei, but (A ∗a A) ∗a (A ∗a A) = 0, and (A,∗a) is solvable.
The theorem is proved.
c ∗a b = a(cb) − (ac)b − c(ab) = c.
3.5. Derivations. Remember that a linear mapping D of an algebra A is called a derivation if it satisfies the
relation D(xy) = D(x)y + xD(y). By [12], an element a of an algebra A is called a Jacobi element if if satisfies
the relation a(xy) = (ax)y + x(ay). All Jacobi elements of A form a vector space, which is called the Jacobi
space of A.
Lemma 27. Let D be a derivation of both A and (A,∗). Then
1) If A has zero Jacobi space, then D = 0;
2) If D is invertible, then A is a left Leibniz algebra and (A,∗) is a zero algebra. In particularly, if A is a
finite-dimensional algebra over a field of zero characteristic, then A is nilpotent.
Proof. 1). By simple calculations, from D(x ∗ y) = D(x) ∗ y + x ∗ D(y), we have
D(u)(xy) = (D(u)x)y + x(D(u)y).
By the definition of the Jacobi space, we have D = 0.
2). By invertibility of mapping D and arbitrarity of element u, we infer that A is a left Leibniz algebra. By
the Lemma 5, we imply that (A,∗) is zero algebra.
ting an invertible derivation is nilpotent. The Lemma is proved.
In [10] it was proved that a finite-dimensional Leibniz algebra over a field of characteristic zero which admit-
11
3.6. Automorphisms. Remember that an invertible linear mapping φ of an algebra is called an automorphism
if it satisfies the relation φ(xy) = φ(x)φ(y).
Lemma 28. Let φ be an automorphism of both A and (A,∗). If A is an algebra with zero Jacobi space, then φ
is the identity mapping.
Proof. By simple calculations from φ(x ∗ y) = φ(x) ∗ φ(y), we have
(u − φ(u))(xy) = ((u − φ(u))x)y + x((u − φ(u))y).
By the definition of the Jacobi space, we have that φ = id. The Lemma is proved.
3.7. Isomorphic Kantor squares. Here we talk about the situation where algebra A and its Kantor square
are isomorphic.
Theorem 29. Let A be a finite-dimensional associative algebra. Then A is isomorphic to (A,∗), if and only if
A is a skew field.
Proof. Let fu is an isomorphism between algebras A and (A,∗u) and fu(xy) = fu(x)∗fu(y) = −fu(x)ufu(y).
If in A there are two elements u and v with zero product, for x = f −1
u (v) we have
fu(fx(ab)) = −fu(fx(a)xfx(b)) = fu(fx(a))uvufu(fx(b)) = 0.
Now, if there is a zero divisor, then the algebra A has zero multiplication. It is Well known that every finite-
dimensional algebra without zero divisors is a skew field.
On the other side, for some fixed nonzero element u from a skew field A we define fu(a) = −au−1. It is an
isomorphism between algebras A and (A,∗u) for every nonzero element u. The theorem is proved.
Acknowledgments.
Ivan Shestakov for the idea of this work, Prof. Dr.
Alexandre Pojidaev and Yury Popov for some correction of the language, Dr. Yury Volkov for the idea of the
proof of the Theorem 29.
I am grateful to Prof. Dr.
References
[1] Cantarini N., Kac V., Classification of linearly compact simple Jordan and generalized Poisson superalgebras, J. Algebra, 313
(2007), 100 -- 124.
[2] Cantarini N., Kac V., Classification of linearly compact simple rigid superalgebras, Int. Mat. Res. Not., 17 (2010), 3341 -- 3393.
[3] Chapoton F., Un endofoncteur de la categorie des operades, Dialgebras and related operads, 105 -- 110, Lect. Notes in Math.,
1763, Springer, 2001.
[4] Dedkov A., Some properties of quasi-associative and quasi-alternative algebras, Siberian Math. J., 30 (1989), 3, 479 -- 483.
[5] Dorofeev G. V., On nilpotency of right alternative rings, Algebra and Logic, 9 (1970), 3, 302 -- 305.
[6] Dzhumadil'daev A., Abdykassymova S., Leibniz algebras in characteristic p, C. R. Acad. Sci. Paris Ser. I Math., 332 (2001),
12, 1047 -- 1052.
[7] Dzhumadil'daev A., Zinbiel algebras over a q-commutator, J. Math. Sci. (N. Y.), 144 (2007), 2, 3909 -- 3925.
[8] Dzhumadil'daev A., Ismailov N., A., Tulenbaev K., Free bicommutative algebras, Serdica Math. J., 37 (2011), 1, 25 -- 44.
[9] Dzhumadil'daev A., Ismailov N., Sn- and GLn-module structures on free Novikov algebras, Journal of Algebra, 416 (2014),
287 -- 313.
[10] Fialowski A., Khudoyberdiyev A., Omirov B., A characterization of nilpotent Leibniz algebras, Algebras and Repr. Theory, 16
(2013), 5, 1489 -- 1505.
[11] Kantor I., Certain generalizations of Jordan algebras, Trudy Sem. Vektor. Tenzor. Anal., 16 (1972), 407 -- 499. [in Russian]
[12] Kantor I., The universal conservative algebra, Siberian Math. J., 31 (1990), 3, 388 -- 395.
[13] Kaygorodov I., Lopatin A., Popov Yu., Conservative algebras of 2-dimensional algebras, Linear Algebra and its Applications,
486 (2015), 255 -- 274.
[14] Kaygorodov I., Volkov Yu., Conservative algebras of 2-dimensional algebras, II, Comm. Alg. (2017).
[15] Kaygorodov I., Popov Yu., Alternative algebras admitting derivations with invertible values and invertible derivations,
Izvestiya: Mathematics, 78 (2014), 5, 922 -- 936.
[16] Kuzmin E., Shestakov I., Nonassociative structures, Itogi Nauki i Tekhniki: Sovremennye Problemy Mat.: Fundamental'nye
Napravleniya, vol. 57, VINITI, Moscow 1990, pp. 179 -- 266; English transl., A.I. Kostrikin and I. R. Shafarevich (eds.), Encyclopae-
dia Math. Sci., vol. 57, Algebra VI, Springer-Verlag, London -- New York 1995, pp. 199 -- 280.
[17] Loday J.-L., On the algebra of quasi-shuffles, Manuscripta mathematica, 123(1) (2007), 79 -- 93.
[18] Loday J.-L., Generalized bialgebras and triples of operads, Asterisque, 320 (2008), x+116 p.
[19] Makar-Limanov L., Shestakov I., Polynomial and Poisson dependence in free Poisson algebras and free Poisson fields, J.
Algebra, 349 (2012), 372 -- 379.
[20] Pozhidaev A. P., Dialgebras and related triple systems, Sib. Math. J., 49 (2008), 4, 696 -- 708
[21] Zakharov A., Novikov-Poisson algebras and superalgebras of Jordan brackets, Comm. Algebra, 42 (2014), 5, 2285 -- 2298.
[22] Zhevlakov K., Slinko A., Shestakov I., Shirshov A., Rings that are nearly associative, Pure and Applied Mathematics, 104.
Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], New York-London, 1982. xi+371 pp.
[23] Zinbiel G. W., Encyclopedia of types of algebras 2010, Operads and universal algebra, 217 -- 297, Nankai Ser. Pure Appl. Math.
Theoret. Phys., 9, World Sci. Publ., Hackensack, NJ, 2012.
12
|
1912.08594 | 1 | 1912 | 2019-12-16T08:41:39 | Derivation of Five-Dimensional Lotka-Voltera Algebra | [
"math.RA"
] | Lotka-Volterra (LV) algebras are generally applied in solving biological problems and in examining the interactions among neighboring individuals. With reference to the methods applied by Gutierrez-Fernandez and Garcia in \cite{17}. this study examines the derivations of five-dimensional LV algebras. Before explicitly describing its derivation space, anti-symmetric matrices are used to define the LV algebra A. | math.RA | math |
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA
ALGEBRA
AHMAD ALARAFEEN
IZZAT QARALLEH AZHANA AHMAD
Abstract. Lotka-Volterra (LV) algebras are generally applied in solving bio-
logical problems and in examining the interactions among neighboring individu-
als. With reference to the methods applied by Gutierrez-Fernandez and Garcia
in [17]. this study examines the derivations of five-dimensional LV algebras. Be-
fore explicitly describing its derivation space, anti-symmetric matrices are used
to define the LV algebra A.
Mathematics Subject Classification: 17D92, 17D99, 39A70, 47H10.
Key words: Evolution algebra; Volterra quadratic stochastic operator; Nilpotent;
Isomorphism; Derivation.
1. Introduction
Many studies have examined non-associative algebras and their different classes,
including baric, evolution, Bernstein, train, and stochastic algebras. While each
of these classes has its own unique definition, they are all referred to as genetic
algebras. For instance, the genetics of abstract algebras were investigated in [20]
by referring to their formal language.
Bernstein [21], [20] introduced the concept of population genetics in their study
of evolution operators , which are often described by using quadratic stochastic
operators (see [13]).
Introduced in 1981 by Itoh [10], Lotka-Volterra (LV) algebras are formulations
that are used to describe solutions and singularities and are associated with qua-
dratic differential equations [10,17]. They have been widely applied in different
mathematical fields,
including dynamic systems, Markov chains, graph theory,
and population genetics [10, 12].
Normal Bernstein algebras, which are LV algebras associated with a zero anti-
symmetric matrix, are essential in solving the classical Bernstein problem [4, 5, 8,
14]. Many studies have described the structure of LV algebras [1 -- 3, 6, 7, 10, 16, 17,
24 -- 27]. For instance, Holgate described how the derivations of genetic algebras,
Date: Received: xxxxxx; Revised: yyyyyy; Accepted: zzzzzz.
∗ Corresponding author.
1
2
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
such as LV algebras, can be interpreted in biological terms [9]. The derivations
of LV algebras have been classified up to three dimensions in [7] and up to three
and four dimensions in [3]. Meanwhile, [17], proposed several novel approaches for
examining LV algebras. For example, digraphs can be attached to LV algebras to
classify their derivations of into four dimensions at most without requiring much
effort and while taking into account the special cases in arbitrary dimensions.
By using the techniques introduced in [17], this paper attempts to classify the
derivations of LV algebras up to five dimensions.
2. Preliminaries
This section describes all the definitions and theorems used in this paper to
achieve the aforementioned objective.
Definition 2.1. A commutative (not necessarily associative) algebra A over a
field F which characteristic is not equal to 2 is called LV algebra if the natural
basis B = {e1, ..., en} is admitted such that eiej = αijei + αjiej, where αij and αji
are in field A, whereas aii = 1
2 and αij + αji = 1 for all i, j = 1, 2, ..., n.
The structural constants of LV-algebra A are formulated in matrix form as
(αij)n
i,j=1 with respect to natural basis B.
Remark 2.2. Each LV algebra A is baric in the sense that it admits a non-trivial
homomorphism ω : A → F, which is defined by ω (Pi λiei) = Pi λi for all λi ∈ F.
Definition 2.3. Let (A, ◦) be an algebra with a linear mapping D : A → A such
that
D(u ◦ v) = D(u) ◦ v + u ◦ D(v)
for all u, v ∈ A , derivation on algebra (A, ◦) D ≡ 0 is clearly a derivation, which
we call a trivial one.
We obtain the following notations on the basis of those described in [17]. The
hyper-plane unit and barideal of A are defined as H = {x ∈ Aω(x) = 1} and
N = kerω = {x ∈ Aω(x) = 0}, respectively.
supp(x) denotes the support of x = Pi λiei with respect to natural basis B, and is
induced when the corresponding coefficient is nonzero, that is i ∈ supp(x) if and
only if λi 6= 0. For simplification, we denote {1, 2, ..., n} by S. For each index k,
we define Sk as the set {i ∈ Seiek = (ei + ek)/2}, where k ∈ S. In this case, we
have, k ∈ Sk, and index i ∈ S belongs to Sk if and only if k ∈ Si. Meanwhile, if
x1, ..., xk ∈ A, then hx1, ..., xki is the vector subspace of A spanned by x1, ..., xk.
Let I be a non-empty subset of S, and U= heii ∈ Ii. For each element x =
Pn
k=1 λkek in A, the U-component of x is defined by [x]U = Pi∈I λiei. If U = {ei},
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
3
then we denote [x]U by [x]i. L(A) represents the set of all linear mappings from
A to A. In the following, D is a derivation of A, and λtk ∈ F such that
D(ek) =
n
X
t=1
λtket, k = 1, n.
(2.1)
Proposition 2.4.
[17] The following statements hold true:
(1) If D ∈ Der(A), then ImD ⊂ N and supp(D(ek)) ⊂ S for all k ∈ S.
(2) If Sk = 1 for all k ∈ S, then Der(A) = {0}.
(3) If i ∈ Sk, then supp(D(ek)) ⊆ Si.
(4) supp(D(ek)) ⊆ ∩u∈Sk Si, that is, if i, j ∈ supp(D(ek)), then
eiej = (ei+ej)
2
and eiek = (ei+ek)
2
.
We define Nk = ∩i∈Sk Si for k = 1, 2, ..., n., where index j belongs to Nk if and
only if eiej = (ei + ej)/2 for all i ∈ Sk. We observe that supp(D(ek)) ⊆ Nk for all
j ∈ S.
Theorem 2.5.
[17] Let e1e2 = αe1 + α′e2 where, α 6= 1
2, and define
I = N1 ∩ N2,
J = N1 \ N2,
K = N2 \ N1
U = < eii ∈ I >
V = < ejj ∈ J >
W = < ekk ∈ K >
Then,
D(e1) = u + v,
and D(e2) = u + w.
For some u ∈ U, we have v ∈ V and w ∈ W, where
ω(u) = ω(v) = ω(w) = 0.
Moreover,
eje2 = αej + α′e2,
for all j ∈ supp(v) and k ∈ supp(w).
and
e1ek = αe1 + α′ek
Corollary 2.6.
[17] The following statements hold true:
(i) Let A be similar to the previous theorem. If I ≤ 1, then u = 0. If J ≤ 1,
then v = 0. If K ≤ 1, then w = 0.
(ii) Let e1e2 = αe1 + α′e2, where α 6= 1
i ∈ N1 ∩ N2.
2. Therefore, [D(e1)]i = [D(e2)]i for all
4
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
(iii) If i belongs to Nk, then we have Ni ⊆ Nk.
(iv) If i ∈ Nk and k ∈ Ni, then Ni = Nk.
Attached to the LV algebra are constantly a graph and a weight graph that
are connected and directed. Therefore, LV algebras will usually be presented in
graphical form throughout the rest of this paper. The following properties are
satisfied by graph Γ(A, B) = (S, E) :
(1) (i, j) ∈ E if and only if (i, j) /∈ E for all i, j ∈ S,i 6= j, and
(2) (i, i) /∈ E for all i ∈ S.
The weight graph attached to the LV algebra A relative to natural basis is
represented by the triple Γτ (A, ϕ) = (S, E, τ ) where τ : E −→ F is given by
τ (i, j) = αij. A unique LV algebra is defined up to isomorphism by a weight
′, E ′, τ ) when S = S
′.
graph. An LV algebra has equivalent graphs (S, E, τ ) and (S
For every pair (i, j) ∈ E, :if (i, j) ∈ E ′, then v(i,j) = ρ(i, j) or v(j,i) = 1 − ρ(i, j).
If and only if two graphs are equivalent, then they can be are attached to the same
LV algebra relative to the same natural basis. We define those edges whose weights
are equal and not equal to 1
2 , as :
Meanwhile, we define those edges that are not equal to6= 1
2 as:
The degree of symmetry of an algebra may be assessed by using its weight graph.
Therefore, we also take into consideration the subgraphs attached to algebra A.
The following implications hold (the elements of the basis are identified by the
graphs on the basis of their subindices):
G.1: If α 6= β, then 3 6∈ Supp(D(e2)).
1
β
α
3
2
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
5
G.2: [D(e2)]1 = [D(e3)]1
1
α
2
3
Proof. G.1: Suppose that 3 ∈ Supp(D(e2)), e1e2 = αe1 + α′e2 for α 6= 1
2, we can
obtain 3 ∈ N2 and 3 6∈ N1, . Therefore, we have 3 ∈ N2 \ (N1 ∩ N2), according to
the preivous theorem e1e3 = αe1 + α′e3.
G.2: Suppose that G.2 holds. If 1 /∈ supp(D(e2))∩ supp(D(e3)), then [D(e2)]1 =
[D(e3)]1 because both vanish. Now let 1 /∈ supp(D(e2)) ∪ supp(D(e3)), and
1 ∈ supp(D(e3)). We observe that 1 /∈ N3 \ (N3 ∩ N2) , and according to Theorem
1, we obtain [e2ek]k = (1 − α)ek for all k N3 (N3 ∩ N2) ∩ supp(D(e3)). Therefore,
1 ∈ N2 ∩ N3 according to the results of Corollary 2.6(ii), where [D(e2)]1 = [D(e3)]1.
(cid:3)
Remark 2.7. fully describes the derivation of 3D LV algebra, whereas [7], presents
a new technique for describing the 3D and 4D LV algebra. To describe the deriva-
tion of 5D LV algebra, we use the technique proposed in [17] .
3. Main Results
Let A be a 5D LV algebra over a field whose characteristic is unequal to 2,
and let D is a derivation of A. With reference to a suitable natural basis, A is
represented by a weighted graph γ(A) , which denotes the number of edges whose
weight is not equal to 1
2. If γ(A) = 0, then Der(A) represents the linear mapping
f from A to A such that Im(f ) ⊂ N. We present below some cases of 5D γ(A) .
Theorem 3.1. Let A be a 5D LV algebra, and in the following table summarizes
its derivations:
6
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
Name Graph of
Der(A)
M1
M2
M3
M4
M5
2
3
2
3
2
3
2
3
2
3
1
5
1
5
1
α
5
1
γ
5
1
α
5
4
4
α
4
α
α
4
α
α
4
{f ∈ L(A)Im(f ) ⊂ N}
{f ∈ L(A)Im(f ) ⊂< e3 − e4, e3 − e5 > and f (e1) =
f (e2)}
{f ∈ L(A)Im(f ) ⊂< e2 − e4 > and f (e3) = f (e5) =
f (e1)}
{f ∈ L(A)e1, e5, e4 ∈ ker(f ) and Im(f ) ⊂< e2 − e3 >}
{f ∈ L(A)Im(f) ⊂< e2 − e3, e2 − e5 > and e1, e4 ∈
ker(f)}
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
7
Name Graph of
Der(A)
M6
M7
M8
M9
M10
1
α
5
1
γ
5
1
5
1
β
5
1
α
5
α
γ
α
2
3
2
3
2
3
2
α
3
2
3
α
4
α
β
4
α
4
α
β
4
α
α
4
{f ∈ L(A)Im(f ) ⊂< e2 − e5 > and e1, e3, e4 ∈ ker(f)}
{f ∈ L(A)Im(f ) ⊂< e3 − e4 > and f (e1) = f (e2) =
f (e5)}
{f ∈ L(A)e1, e2, e3 ∈ ker(f ) and Im(f ) ⊂< e4 − e5 >}
{f ∈ L(A)f (e1), f (e3) ∈< e1 − e3 >, f (e2), f (e5) ∈<
e2 − e5 > and e4 ∈ ker(f )}
{f ∈ L(A)e1 ∈ ker(f ) and Im(f ) ⊂< e2 − e3, e2 −
e4, e2 − e5 > }
8
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
Name Graph of
Der(A)
M11
M12
M13
M14
M15
2
3
2
3
2
3
2
3
2
3
δ
δ
δ
α
1
α
5
1
α
5
1
α
5
1
5
1
α
5
α
α
4
α
δ
4
α
γ
4
α
4
α
4
{ f ∈ L(A)Im(f ) ⊂< e2 - e3, e2 − e5 > and e4, e1 ∈
ker(f ) }
{f ∈ L(A)f (e2), f (e5) ∈< e2 − e5 >, f (e3), f (e4) ∈<
e3 − e4 > and e1 ∈ ker(f )}
{f ∈ L(A)Im(f ) ⊂< e2 − e5 > and e1, e4, e3 ∈ ker(f )}
{f ∈ L(A)Im(f ) ⊂< e2 − e4 > and e1 ∈ ker(f )}
{f ∈ L(A)Im(f ) ⊂< e2 − e5 > and e1, e3, e4 ∈ ker(f )}
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
9
Name Graph of
Der(A)
M16
M17
M18
M19
M20
1
α
5
1
α
5
1
α
5
1
α
5
1
5
α
δ
2
3
2
3
2
3
2
3
2
3
δ
β
α
4
α
4
α
4
α
4
4
{f ∈ L(A)Im(f ) ⊂< e2 − e5 > and e1, e3, e4 ∈ ker(f )}
{f ∈ L(A)Im(f ) ⊂< e2 − e5 > and e1, e3, e4 ∈ ker(f )}
{f ∈ L(A)Im(f ) ⊂< e2 − e5 > and e1, e3, e4 ∈ ker(f )}
{f ∈ L(A)Im(f ) ⊂< e1 − e3 > and e2, e4, e5 ∈ ker(f )}
{f ∈ L(A)Im(f ) ⊂< e1 − e3 > and e2, e5, e4 ∈ ker(f )}
10
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
δ
α
α
δ
4
2
3
2
α
3
1
5
1
5
M21
α
4
M22
{f ∈ L(A)f (e1), f (e3) ∈< e1 − e3 > and f (e2), f (e5) ∈< e2 − e5 >}
{f ∈ L(A)Im(f ) ⊂< e2 − e4 > and e1, e3, e5 ∈ ker(f )}
Proof. Depending on the order of the set, the proof is divided into the following
cases:
Case 1: If γ(A) = 1, then A can be graphed as follows up to equivalence and
isomorphism :
2
3
1
5
4
In this graph, we have S1 = {1, 3, 4, 5}, S2 = {2, 3, 4, 5} = S5, and Sk = S ,
where k = 3, 4, 5. From these we obtain N1 = {1, 3, 4, 5}, and Nk = {3, 4, 5},
where k = 3, 4, 5, and N2 = {2, 3, 4, 5}. If we have
N1(cid:31)(N1 ∩ N2) = N2(cid:31)(N1 ∩ N2) = 1,
then we obtain,
D(e1), D(e2) ∈< ekk ∈ N1 ∩ N2 > ∩N = < e3 − e4, e3 − e5 >.
Meanwhile, if we have Nk = {3, 4, 5}, then we obtain,
D(e3), D(e4), D(e5) ∈< ekk ∈ N3 > ∩N = < e3 − e4, e3 − e5 >.
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
11
We find from G.2 that [D(e1)]k = [D(e2)]k, where k = 3, 4, 5. In this case,
[D(e1)] = [D(e2)].
Case 2: If γ(A) = 2, then A an be graphed as follows up to equivalence and iso-
morphism:
2
3
1
5
4
1
α
5
2
3
β
4
where in the first graph, S1 = {1, 3, 4, 5}, S2 = {2, 3, 4, 5}, S3 = {1, 2, 3, 5},
S5 = S, and S4 = {1, 2, 4, 5}, where k = 3, 4, 5. From these, we obtain
N1 = {1, 5}, N2 = {2, 5}, N3 = {3, 5}, N5 = {5}, and N5 = {4, 5}. In this
case, D(e5) = 0 given that
Nk(cid:31)(N1 ∩ Nk) = N1(cid:31)(N1 ∩ Nk) = Nk ∩ N1 = 1; . Therefore,
D(ek) = 0, where k = 2, 3, 4, and
Der(A) = 0.
As for the second case, we have S1 = {1, 2, 3, 4}, S2 = S, S3 = {1, 2, 3, 4},
S4 = S, and S5 = {2, 4, 5}, and we obtain NK = {1, 2, 3, 4}, where k = 1,
3. Given that Nr = {2, 4} for r = 2, 4 ,N5 = {2, 4, 5}; we obtain D(e2),
D(e4) ∈< e2 − e4 >. Following G.2, we have:
[D(e1)]4 = [D(e5)]4
[D(e1)]2 = [D(e5)]2 ⇒ D(e5) = D(e1).
Similarly, we can obtain D(e3) = D(e5). Therefore, D(e3) = D(e5) =
D(e1).
, and Der(A) = 0 when α 6= β, Der(A) = 0. Following G.1, if
α = β, then the f ∈ L(A) set is a deviation if and only if
f (e1), f (e3), f (e5) ∈< eii ∈ N1 ∩ N3 ∩ N5 > ∩N =< e2 − e4 > .
12
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
Case 3:
If γ(A) = 3, then A can be graphed up to equivalence and isomorphism as
follows:
1
γ
5
1
5
2
3
2
β
3
α
β
4
α
γ
4
1
γ
5
1
α
5
2
3
2
3
α
β
4
β
γ
4
In the first case, S1 = {1, 4}, S2 = {2, 3, 4, 5} = S5, S3 = {2, 3, 4, 5}, and
S4 = S, from which we obtain N1 = {1, 4}, and Nk = {2, 3, 4, 5}, where k
= 2, 3, 5, and N4 = {4}. Therefore, D(e4) = 0. Given that
N1 ∩ Nk = N1(cid:31)(N1 ∩ Nk) = 1 for k = 2, 3, 5, we have
D(e1) = 0. According to G.1, if α 6= β 6= γ α = β 6= γ then we have
Der(A) = 0. In this case, Der(A) is treated as the set of all f ∈ L(A)
where e5, e1, e4 ∈ ker(f ) and
f (e2), f (e3) ∈< e2 − e3 > .
Both α 6= β = γ and α = γ 6= β are analogous cases. Specifically, if
α = β = γ, then Der(A) can be easily proven as the set of all f ∈ L(A),
such as e1, e4 ∈ ker(f), that is,
Im(f) ⊆< e2, e3, e5 > ∩N =< e2 − e3, e2 − e5 > .
In the second case, we have S1 = { 1, 3, 4 }, S2 = {2, 3, 4}, S3 = S, S4 = S,
and S5 = {3, 4, 5}, such that N1 = {1, 3, 4}, N2 = {2, 3, 4}, and Nk = {3,
4} for k = 3, 4. N5 = {5, 3, 4}. From Nr = {3, 4},where r ={ 3, 4}, we
obtain D(e3), D(e4) ∈< e3 − e4 >. Given that
N1(cid:31)(N1 ∩ Nk) = Nk(cid:31)(N1 ∩ Nk) = 1,
where k = 2, 5, we have
D(e1), D(e2), D(e5) ∈< ek : k ∈ N1 ∩ Nk > ∩N =< e3 − e4 >
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
13
Following G.2 we have
[D(e1)]4 = [D(e2)]4 = [D(e5)]4 and [D(e1)]3 = [D(e2)]3 = [D(e5)]3.
Therefore,
D(e1) = D(e2) = D(e5).
In the third case, we have S1 = {1, 3, 4, 5 }, S2 = { 2, 4, 5 }, S3 = {
1, 3, 5 }, S4 = { 1, 2, 4, 5 }, and S5 = S, and we obtain N1 = { 1, 5},
N2 = {2, 4, 5}, N3 = {1, 3, 5}, N4 = {4, 5}, and N5 = {5}. Therefore,
D(e5) = 0. Given that
N1(cid:31)(N1 ∩ N2) = N1 ∩ N2 = 1,
we have D(ek) = 0, and k = 1. Similarly, on the basis of e4 and e3, we
obtain D(e4) = 0. Following G.2 we obtain [D(e2)]5 = [D(e3)]5, which
leads to D(e3) = D(e2). We also obtain [D(e2)]4 = [D(e1)]4, where,
D(e2) = D(e1) = 0, and D(e3) = D(e4) = 0, where Der(A) = 0.
In the fourth case, we have S1 = {1, 3, 4}, S2 = {2, 3, 4, 5}, S3 = {1, 2, 4, 5},
S4 = {1, 2, 4, 5}, and S5 = {2, 3, 4, 5}, and we obtain N1 = {1} and
Nk = {2, 5}, where k = 2, 5, N3 = {3}, and N4 = {4}. Therefore,
D(er) = 0, where r = 1, 3, 4. According to G.1, However, if α 6= β,
then Der(A) = 0. If α = β then Der(A) represents the set of all f ∈ L(A),
where er ∈ ker(f ) for r = 1, 3, 4, such that
f(e2), f(e5) ∈< e2, e5 > ∩N =< e2 − e5 >.
Case 4: If γ(A) = 4, then A can be graphed up to equivalence and isomorphism as
follows:
2
3
1
5
4
1
5
2
3
β
α
4
14
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
2
β
3
1
δ
5
α
γ
4
2
3
1
β
5
δ
α
γ
4
In the first case, we have S1 = {1, 3, 4, 5}, S2 = {2, 4, 5}, S3 = {1, 3, 5},
S4 = {1, 2, 4}, and S5 = {1, 2, 3, 5}, and we obtain Nk = {k}, for k = 1, 2,
4, 5 and N3 = {1, 3, 5}. Therefore, D(ek) = 0, k=1, 2, 4, 5. Following G.2,
we have [D(e3)]5 = [D(e2)]5, where D(e3) = D(e2) = 0, and Der(A) = 0.
In the second case, we have S1 = {1, 3, 4, 5}, S2 = {2, 4, 5}, S3 = {1, 3},
S4 = {1, 2, 4, 5}, and S5 = {1, 2, 4, 5}, and we obtain N1 = {1}, N2 =
{2, 4, 5}, N3 = {1, 3}, N4 = {4, 5}, and N5 = {4, 5}, where D(e1) = 0. Fol-
lowing G.2 , we have [D(e1)]s = [D(e2)]s = 0 (where s = 4, 5). [D(e3)]1 =
[D(e4)]1, and [D(e3)] = [D(e4)] , and [D(e3)] = [D(e5)]. If α 6= β, then we
have Der(A) = 0, but if α 6= β, then we have e2, e1 ∈ ker(A). Therefore,
f (e4), f (e5) ∈ N ∩ N4 =< e4 − e5 >.
In the third case, we have S1 = {1, 3, 4}, S2 = {2, 4, 5}, S3 = {1, 3, 4},
S4 = S, and S5 = {2, 4, 5}, and Nr = {1, 3, 4}, (where r=1, 3) Nr = {2, 4, 5}
(for r= 2, 5). N4 = {4}. We obtain D(e4) = 0. Following G.1, if either
α 6= β or γ 6= δ, then D(e1) = D(e3) = 0, α 6= δ or γ 6= β , and
D(e2), D(e5) = 0 . We rearrange the induces if necessary and assume
without loss of generality that if α = β 6= γ = δ, then f ∈ L(A) is a
derivation of A if and only if e4 ∈ ker(f ) and f(e1), f(e3) ∈< e1 − e3 >, and
f(e5),f(e2) ∈< e2 − e5 >.
In the fourth case, we have S1 = {1}, and Sk = {2, 3, 4, 5}, where k = 2,
3, 4, 5. We subsequently obtain N1 = 1, D(e1) = 0, and Nk = {2, 3, 4, 5},
and k = 2, 3, 4, 5. The following cases are derived accordingly:
1)By letting α = β = γ = δ, Der(A) can be easily proven to be a set of all
f ∈ L(A) such that e1 ∈ ker(f ) and
Im(f) ⊂< e2, e3, e4, e5 > ∩N =< e2 − e3, e2 − e4, e2 − e5 > .
2)By letting α = β = γ 6= δ, we acquire G.1 forces e4, e1 ∈ ker(f ) and
D(e2), D(e3), D(e5) ∈ U ∩ N =< e2 − e3, e2 − e5 > .
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
15
Cases α 6= β = γ = δ, α = γ = δ 6= β, and α = γ = β 6= δ are all analogous.
3) By letting α = β 6= γ = δ, Der(A) can be easily proven to be a set of
all f ∈ L(A) such that f (e5), f(e2) ∈< e2 − e5 >, f(e3), f(e4) ∈< e3 − e4 >
and e1 ∈ ker(f ). Cases α = β 6= γ = δ, α = γ 6= β = δ, α = δ 6= γ = β are
all analogous.
4) By letting α = β 6= γ 6= δ, we have e1, e4, e3 ∈ ker(f), f(e2), f(e5) ∈<
e2 − e5 >. Cases α 6= β 6= γ = δ is analogous.
Case 5:
If γ(A) = 5, A can be graphed up to equivalence and isomorphism as
follows:
2
3
2
3
β
α
4
α
4
1
5
1
γ
5
2
3
1
γ
5
1
β
5
α
4
α
4
2
3
2
3
1
5
4
In the first case, we have S1 = {1}, S2 = {2, 3, 4, 5}, S3 = {2, 3, 4}, S4 =
{2, 3, 4, 5}, and S5 = {2, 4, 5}, and we obtain N1 = {1}, N2 = {2, 4},
N3 = {2, 3, 4}, N4 = {2, 4}, and N5 = {2, 4, 5}. Therefore, D(e1) = 0 and,
N3(cid:31)(N5 ∩ N3) = N5(cid:31)(N3 ∩ N5) = 1.
16
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
In this case, D(e3), D(e5) ∈ U ∩ N =< e2 − e4 >. Following G.1, if α 6= β,
then e2, e4 ∈ ker(f).whereas if α = β, then f ∈ L(A) is considered a deriva-
tion of A if and only if e1 ∈ ker(f ) and f(e2), f(e4) ∈ U ∩ N =< e2 − e4 > .
In the second case, we have S1 = {1, 3, 4}, S2 = {2, 4, 5}, S3 = {1, 3, 5},
S4 = {1, 2, 4}, and S5 = {2, 3, 5}. and, we obtain Nk = {k}, k=1, 2, 3, 4,
5. Therefore, Der(A) = 0.
In the third case, we have S1 = {1, 4}, S2 = S5 = {2, 4, 5}, S3 = {3, 4},
S4 = {2, 4, 5}. and we obtain N1 = {1, 4}, Nk = {2, 4, 5}, (where k = 2, 5)
N3 = {3, 4}, and N4 = {4}. In this case, D(e4) = 0. Given that
N1 ∩ N3 = N3 \ (N1 ∩ N3) = N1 \ (N3 ∩ N1) = 1,
we have D(e1), D(e3) = 0. If α 6= γ, then Der(A)=0, whereas if α = γ,
then the set of all f ∈ L(A), similar that e1, e3, e4 ∈ ker(f ) and
f(e2), f(e5) ∈< e2, e5 > ∩N =< e2 − e5 >
.
In the fourth case, we have S1 = {1, 3}, S2 = {2, 4, 5}, S3 = {1, 3, 4}, S4 =
{2, 3, 4, 5}, and S5 = {2, 4, 5} and we obtain N1 = {1, 3}, N2 = {2, 4, 5},
N3 = {3}, N4 = {4}, and N5 = {2, 4, 5}. Therefore, D(e3), D(e4) = 0. By
using the previous theorem, we obtain
N1(cid:31)(N3 ∩ N1) = N3 ∩ N1 = 1.
and show that, D(e1) = 0. Following G.1, if α 6= β, then Der(A) = 0,
whereas if α = β, then the set of all f ∈ L(A) similar e1, e3, e4 ∈ ker(f )
and
f(e2), f(e5) ∈< e2, e5 > ∩N =< e2 − e5 >
In the fifth case, we have S1 = {1, 3}, S2 = {2, 5}, S3 = {1, 3, 4, 5}, S4 =
{3, 4, 5}, and S5 = {2, 3, 4, 5} and we obtain N1 = {1, 3}, N2 = {2, 5},
N3 = {3}, N4 = {3, 4, 5}, and N5 = {5}. Therefore, D(ek) = 0 for k = 3,
4, and
N2(cid:31)(N2 ∩ N4) = N2 ∩ N4 = 1.
In this case, D(e2) = 0. Following G.2, we obtain [D(e2)]5 = [D(e4)]5 = 0
, which suggests that D(e4) = 0. We also obtain D(e1) = 0, thereby sug-
gesting that Der(A)=0.
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
17
Case 6:
If γ(A) = 6, then A can be graphed up to equivalence and isomorphism as
follows:
2
3
2
3
4
4
1
5
1
5
1
3
1
β
5
1
β
5
α
4
α
4
2
3
2
3
2
5
4
In the first case, we have S1 = {1, 3}, S2 = {2, 4, 5}, S3 = {1, 3, 5}, S4 =
{2, 4}, and S5 = {2, 3, 5}, and we obtain N1 = {1, 3}, N2 = {2}, N3 = {3},
N4 = {2, 4}, and N5 = {5}. Therefore, D(e2), D(e3), D(e5) = 0. Following
G.2,
[D(e5)]2 = [D(e4)]2 = 0,
which suggests that, D(e4) = 0, D(e1) = 0 and Der(A) = 0.
In the second case, we have S1 = {1}, S2 = {2, 4, 5}, S3 = {3, 4}, S4 =
{2, 3, 4, 5}, and S5 = {2, 4, 5} and we obtain N1 = {1}, N2 = {2, 4, 5}, N3 =
{3, 4}, N4 = {4}, and N5 = {2, 4, 5}. Therefore, acquire D(e1), D(e4) = 0.
Following the previous theorem,
N3(cid:31)(N2 ∩ N3) = N2 ∩ N3 = 1,
which means that, D(e3) = 0. With reference to the same theorem, we let
Nk = {2, 4, 5} for k = 2, 5. In this case, if α = β, then the f ∈ L(A) set
18
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
can be treated as a derivation if and only if ei ∈ ker(f ), where i = 1, 3, 4
and f(e2), f(e5) ∈< e2 − e5 >, and if α 6= β. Therefore, Der(A) = 0.
In the third case, we have S1 = {1, 4}, S2 = {2, 4}, S3 = {3, 4}, S1 = S,
and S5 = {4, 5}, and we obtain N1 = {1, 4}, N2 = {2, 4}, N3 = {3, 4},
N4 = {4}, and N5 = {4, 5}. Therefore, D(e4) = 0. Given that
N1(cid:31)(N1 ∩ Nk) = Nk(cid:31)(N1 ∩ Nk) = N1 ∩ Nk = 1.
we obtain D(ek) = 0, where k = 1, 2, 3, 5. In this case, Der(A) = 0.
In the fourth case, we have S1 = {1, 3}, S2 = {2, 5}, S3 = {1, 3}, S4 =
{3, 4, 5}, and S5 = {2, 4, 5}, so that N1 = {1, 3}, N2 = N5 = {2, 5},
N3 = {3}, and N4 = {4}. Therefore, D(ek) = 0 , where k = 3, 4. Following
G.2, we obtain
[D(e1)]3 = [D(e4)]3 ⇒ D(e1) = 0.
If α 6= β, then Der(A) = 0, whereas if α = β, then Der(A) repre-
sents the set of all f ∈ L(A), such as er ∈ ker(f ), where r = 1, 3, 4.
f(e2), f(e5) ∈< e2 − e5 >.
In the fifth case, we have S1 = {1, 3, 5}, S2 = {2, 3, 5}, S3 = {1, 2, 3},
S4 = {4}, and S5 = {1, 2, 5} and we obtain, Nk = {k}, where k = 1, 2, 3,
4, 5. In this case, Der(A) = 0.
Case 7:
If γ(A) = 7, then A can be graphed up to equivalence and isomorphism as
follows:
2
3
1
5
4
2
3
1
β
5
α
4
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
19
1
5
2
3
4
In the first case, we have S1 = {1}, S2 = {2, 4, 5}, S3 = {3, 5}, S4 = {2, 4},
and S5 = {2, 3, 5} and we obtain, Nk = {k}, for k = 1, 2, N3 = {3, 5},
N4 = {2, 4}, and N5 = {2, 5}. Given that
N4(cid:31)(N5 ∩ N4) = N4 ∩ N5 = N4(cid:31)(N3 ∩ N5) = 1.
we obtain D(e4), D(e5) = 0. Following G.2,
[D(e2)]5 = [D(e3)]5 ⇒ D(e3) = 0.
In this case, Der(A) = 0.
In the second case, we have S1 = {1, 3}, S2 = {2, 4, 5}, S3 = {1, 3},
S4 = {2, 4}, and S5 = {2, 5} and we obtain N1 = {1, 3},N2 = {2},
N3 = {1, 3}, N4 = {4}, and N5 = {2, 5}. Therefore, D(e4), D(e2) = 0.
Given that
[D(e5)]2 = [D(e4)]2 ⇒ D(e5) = 0.
we have D(e5) = 0. If α 6= β, then Der(A) = 0 whereas if α = β, then
Der(A) suggests that the set of all f ∈ L(A) is a derivation of A if and
only if e2, e4, e5 ∈ ker(f ) and f(e1), f(e3) ∈< e1 − e3 >.
In the third case, we have S1 = {1, 4}, S2 = {2, 4}, S3 = {3}, S4 =
{1, 2, 4, 5}, and S5 = {4, 5}, and we obtain N1 = {1, 4}, N2 = {2, 4},
N3 = {3}, N4 = {4}, and N5 = {4, 5}. Therefore, D(e4), D(e3) = 0. Given
that
N1(cid:31)(N1 ∩ Nk) = Nk(cid:31)(N1 ∩ Nk) = Nk ∩ N1 = 1.
we have D(ek) = 0, where k = 1, 2, 5. In this case, Der(A) = 0.
20
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
Case 8:
If γ(A) = 8, then A can be graphed up to equivalence and isomorphism as
follows:
2
3
1
5
4
2
3
1
5
γ
β
α
δ
4
In the first case, we have Sr = {r}, (where r = 1, 3,) S2 = {2, 4, 5},
S4 = {2, 4} and S5 = {2, 5}, and we obtain Nk = {k}, (where k = 1, 2,
3)N4 = {2, 4} and N5 = {2, 5}. Therefore, D(ek) = 0 where k = 1, 2, 3.
Given that
N4(cid:31)(N5 ∩ N4) = N4 ∩ N5 = N4(cid:31)(N4 ∩ N5) = 1.
we have D(er) = 0, r = 4, 5. In this case, Der(A) = 0.
In the second case , we have S1 = {1, 3}, S2 = {2, 5}, S3 = {1, 3}, S4 = {4},
and S5 = {2, 5}, and we obtain, Nk = {1, 3}, (k = 1, 3) Nr = {2, 4}, (r
= 2, 5) and N4 = {4}. Therefore, D(e4) = 0. Following G.1, if α 6= γ
or β 6= δ, then D(e2), D(e5) = 0. Similarly, if α 6= β or γ 6= δ, we have
D(e1) = D(e3) = 0. In this case, Der(A) = 0.
If α 6= β, then f ∈ L(A) can be treated as a derivation if and only if
e1, e3 ∈ ker(f) and f(e2), f(e5) ∈< e2 − e5 >. Similarly, if α 6= γ, then
e2, e5 ∈ ker(f) and f(e1), f(e3) ∈< e1 − e3 >.
If α = β 6= γ = δ, then
f ∈ L(A) can be viewed as a derivation of A if and only if e4 ∈ ker(f ) and
f(e1), f(e3) ∈< e1 − e3 > and f(e2), f(e5) ∈< e2 − e5 >.
Case 9: If γ(A) = 9 then A up to equivalence and isomorphism as follows:
1
5
2
α
3
β
4
DERIVATION OF FIVE-DIMENSIONAL LOTKA-VOLTERA ALGEBRA
21
In this case, we have Sk = {k} (k = 1, 3, 5) and Sr = {2, 4}, r = 2, 4, and we
obtain, Nk = {k} (k = 1, 5, 3)Nk = {2, 4} (k = 2, 4), Therefore, D(ek) = 0,
where k= 1, 3, 5. Following G.1, if α 6= β, then e4, e2 ∈ ker(f). Therefore,
Der(A) = 0, and if α = β, then f ∈ L(A) can be treated as a derivation of
A if and only if es ∈ ker(f ), where s= 1, 3, 5, and f(e2), f(e4) ∈< e2 − e4 >.
(cid:3)
References
[1] H. Boujemaa, M. Rachidi and A. Micali, Sur les algA"bres de Lotka-Volterra, Algebras
Groups Geom. 10 (1993) 169-180
[2] H. Boujemaa, M. Rachidi and A. Micali, Automorphisms in Lotka-Volterra algebras, Alge-
bras Groups Geom. 18(2001) 203-222.
[3] H. Boujemaa, M. Rachidi and A. Micali, D"erivations dans les algA"bres de Lotka- Volterra,
Algebras Groups Geom. 2 (2004) 471-487.
[4] Juan C. Gutierrez Fernandez, Solution of the Bernstein problem in the non-regular case, J.
Algebra 223 (2000) 109-132.
[5] Juan C. Gutierrez Fernandez, Nuclear Bernstein algebras with a stochastic basis, J. Algebra
217 ( 1999) 300-311.
[6] uan C. Gutierrez Fernandez and Claudia I. Garcia, Antisymetric matrices and LotkaVolterra
algebras, https://arxiv.org/abs/1801.08888
[7] R. Ganikhodzhaev,
F. Mukhamedov, A. Pirnapasov
and
netic Volterra
https://doi.org/10.1080/00927872.2017.1347663.
algebras
and
their
derivations,
Comm.
I. Qaralleh, Ge-
(2017)
in Algebra
[8] P. Holgate, Genetic algebras satisfying Bernstein?s stationary principle, J. London Math.
Soc. 9 (1975) 621-624.
[9] P. Holgate, The interpretation of derivations in genetic algebras, Linear Algebra Appl. 85
(1987) 75-79
[10] Y. Itoh, Nonassociative algebra and Lotka-Volterra equation with ternary interaction, Non-
linear Anal. 5 (1981) 53-56.
[11] R.V. Kadison, Local derivations, Journal of Algebra 130 (1990) 494-509
[12] M. Kimura, On the change of population fitness by natural selection, Heredity 12 (1958)
145-167
[13] Y. Lyubich, Mathemmatical structures in population genetics, (Springer-Verlag, 1983).
[14] K. Mather, Selection through competition, Heredity 24 (1969) 529-540
[15] S. I. Yoon, Automorphisms of Lotka-Volterra algebras, Comm Korean Math. Soc. 12 (1997)
45-50
[16] S. I. Yoon, Idempotent elements in the Lotka-Volterra algebra., Comm Korean Math. Soc.
10 (1995) 123-13.
[17] Juan C. Gutierrez Fernandez and Claudia I. Garcia,Derivations of Lotka-Volterra alge-
bras,Sao Paulo Journal of Mathematical Sciences(2018)
[18] Worz-Busekros, A. (1980). Algebras in Genetics. Lect. Notes in Biomathematics. Vol. 36.
Berlin: Springer-Verlag.
[19] Reed, M. L. (1997). Algebraic structure of genetic inheritance. Bull. Amer. Math. Soc.(N.S.)
34:107130.
[20] Etherington, I. M. H. (1939). Genetic algebras. Proc. Roy. Soc. Edinburgh 59:242258
22
AHMAD ALARAFEEN
IZZAT QARALLEH
AZHANA AHMAD
[21] Bernstein, S. N. (1923). Principe de stationarit et gnralisation de la loi de Mendel. Comptes
Rendus Acad. Sci. Paris 177:581584.
[22] Bernstein, S. N. (1924). Solution of a mathematical problem connected with the theory of
heredity. Ann. Sci. de lUkraine 1:83114 (Russian).
[23] Mukhamedov, Farrukh, and Izzat Qaralleh. "On derivations of genetic algebras." Journal
of Physics: Conference Series. Vol. 553. No. 1. IOP Publishing, 2014.
[24] Mukhamedov, Farrukh, et al. "On Genetic and Evolution Algebras in Dimension Three."
International Journal of Algebra 10.7 (2016): 327-334.
[25] Qaralleh, Izzat, and Farrukh Mukhamedov. "Volterra Evolution Algebras and Their
Graphs." arXiv preprint arXiv:1904.10305 (2019).
[26] Qaralleh, Izzat. "On Genetic and Evolution Algebras." Journal of Physics: Conference
Series. Vol. 819. No. 1. IOP Publishing, 2017.
[27] Ladra, M., and U. A. Rozikov. "Evolution algebra of a bisexual population." Journal of
Algebra 378 (2013): 153-172.
Ahmad Alarafeen, School of Mathematical Sciences, , Universiti Sains Malaysia,
11800 USM, Penang, Malaysia
E-mail address: [email protected]
Izzat Qaralleh, Department of Mathematics, Faculty of Science, Tafila Tech-
nical University, Tafila, Jordan
E-mail address: izzat [email protected]
AZHANA AHMAD, School of Mathematical Sciences, , Universiti Sains Malaysia,
11800 USM, Penang, Malaysia
E-mail address: [email protected]
|
1211.5763 | 1 | 1211 | 2012-11-25T13:34:29 | On Artinian Rings with Restricted Class of Injectivity Domains | [
"math.RA"
] | In a recent paper of Alahmadi, Alkan and Lopez-Permouth, a ring R is defined to have no (simple) middle class if the injectivity domain of any (simple) R-module is the smallest or largest possible. Er, Lopez-Permouth and Sokmez use this idea of restricting the class of injectivity domains to classify rings, and give a partial characterization of rings with no middle class. In this work, we continue the study of the property of having no (simple) middle class. We give a structural description of right Artinian right nonsingular rings with no right middle class. We also give a characterization of right Artinian rings that are not SI to have no middle class, which gives rise to a full characterization of rings with no middle class. Furthermore, we show that commutative rings with no middle class are those Artinian rings which decompose into a sum of a semisimple ring and a ring of composition length two. Also, Artinian rings with no simple middle class are characterized. We demonstrate our results with several examples. | math.RA | math |
On Artinian Rings with Restricted Class of Injectivity Domains
Pınar Aydogdu∗
and
Bulent Sara¸c
Department of Mathematics, Hacettepe University, Beytepe 06800 Ankara / Turkey
Abstract
In a recent paper of Alahmadi, Alkan and L´opez -- Permouth, a ring R is defined to have no
(simple) middle class if the injectivity domain of any (simple) R -- module is the smallest or largest
possible. Er, L´opez -- Permouth and Sokmez use this idea of restricting the class of injectivity
domains to classify rings, and give a partial characterization of rings with no middle class. In
this work, we continue the study of the property of having no (simple) middle class. We give a
structural description of right Artinian right nonsingular rings with no right middle class. We
also give a characterization of right Artinian rings that are not SI to have no middle class, which
gives rise to a full characterization of rings with no middle class. Furthermore, we show that
commutative rings with no middle class are those Artinian rings which decompose into a sum of a
semisimple ring and a ring of composition length two. Also, Artinian rings with no simple middle
class are characterized. We demonstrate our results with several examples.
1
Introduction
Throughout this paper, our rings will be associative rings with identity, and modules will be unital
right modules, unless otherwise stated. For any ring R, Mod -- R will denote the category of all right
R -- modules.
Let R be a ring. Recall that an R -- module M is injective relative to an R -- module N (or, M is N --
injective) if, for any submodule K of N, any R -- homomorphism f : K → M extends to some member
of HomR(N, M ). It is evident that every module is injective relative to semisimple modules. Thus,
for any R -- module M, the injectivity domain In−1(M ) = {N ∈ Mod -- R : M is N -- injective} of M
contains all semisimple right R -- modules. In [1], Alahmadi, Alkan and L´opez -- Permouth initiated the
study of poor modules, namely modules whose injectivity domains consist only of semisimple modules
in Mod -- R. They consider rings over which every right module is either injective or poor, and refer
such rings as having no right middle class. The study of rings with no middle class has a growing
interest in recent years (see [1], [3], [6], and [9]).
In [3], Er, L´opez -- Permouth and Sokmez studied the rings with no right middle class and gave
a partial characterization of such rings. The following two theorems summarize the results on rings
with no right middle class obtained in [3]. To simplify the statements, we assume that the ring R is
not semisimple Artinian. All statements can be made to fit that possibility by setting T = 0.
Theorem 1 Let R be a right SI -- ring. Then R has no right middle class if and only if R ∼= S ⊕ T,
where S is semisimple Artinian and T is either
(i) Morita equivalent to a right P CI -- domain or
(ii) an indecomposable ring with homogenenous essential right socle satifying one of the following
equivalent conditions (where Q is the maximal right quotient ring of T ):
(a) Non-semisimple quasi -- injective right T -- modules are injective.
(b) Proper essential submodules of QT are poor.
(c) For any submodule A of QT containing Soc(TT ) properly, QA = Q.
Those rings of type (ii) are either right Artinian or right V -- rings and have a unique simple singular
right T -- module up to isomorphism.
Theorem 2 Let R be a ring with no right middle class which is not right SI. Then R ∼= S ⊕ T,
where S is semisimple Artinian and T is an indecomposable right Artinian ring satisfying the following
conditions:
∗Corresponding author : [email protected] (P. Aydogdu)
1
(i) soc(TT ) = Z(TT ) = J(T ),
(ii) T has homogeneous right socle, and
(iii) there is a unique non -- injective simple right T -- module up to isomorphism.
In this case T is either a QF -- ring with J(T )2 = 0, or poor as a right module. Conversely, if T is
a QF -- ring with homogeneous right socle and J(T )2 = 0, then T has no right middle class.
Note that the authors of [3] could not reverse this implication to show that the conditions (i) -- (iii)
in Theorem 2 above are sufficient as well as necessary. As a matter of fact, we show in our work that
there exist rings satisfying conditions (i) -- (iii) in Theorem 2 above which are poor as a right module
over itself and do have right middle class (see Examples 2.22 and 2.23). We also give a complete
characterization of non -- SI -- rings with no right middle class (see Theorem 2.16).
A characterization of right Artinian rings with no right middle class was given in Corollary 3.2 of [9].
Using that result in conjunction with those of [3], we have the following two complete characterizations:
Theorem 3 Let R be any ring. Then R has no right middle class if and only if R ∼= S ⊕ T, where S
is semisimple Artinian and T satisfies one of the following conditions:
(i) T is Morita equivalent to a right P CI -- domain, or
(ii) T is a right SI right V -- ring with the following properties:
(a) T has essential homogeneous right socle and
(b) for any submodule A of QT which does not contain the right socle of T properly, QA = Q,
where Q is the maximal right quotient ring of T, or
(iii) T is a right Artinian ring whose Jacobson radical properly contains no nonzero ideals.
Theorem 4 Let R be any ring. Then R has no right middle class if and only if R ∼= S ⊕ T, where S
is semisimple Artinian and T satisfies one of the following conditions:
(i) T is Morita equivalent to a right P CI -- domain, or
(ii) (a) T is a right Artinian ring or a right SI right V -- ring with homogeneous essential right
socle, and
(b) every nonsemisimple quasi -- injective right T -- module is injective.
In the process of studying these rings, various necessary or sufficient conditions are presented in
[3]. For instance,
(P1) R has homogeneous right socle, and
(P2) there is a unique simple singular right R -- module up to isomorphism
are necessary conditions for a nonsemisimple indecomposable right SI -- ring R to have no right middle
class (see [3, Theorem 2]). Likewise, in [3, Proposition 6], it is shown that right Artinian right SI -- rings
with homogeneous right socle anda unique local module of length two up tp isomorphism must have
no right middle class. We show here that (P1) and (P2) are not sufficient while the condition that
the ring has a unique local module of length two up to isomorphism is not necessary (see Examples
2.9(i) and 2.13).
It is shown, in [3, Corollary 5], that if R is an indecomposable right nonsingular right Artinian
ring with no right middle class, then R is isomorphic to a formal triangular matrix ring of the form
0
D
0
of Morita equivalences, we see that such rings simplifies to formal triangular matrix rings of the form
(cid:18) S
A S′ (cid:19) , where S and S′ are simple Artinian rings and A is an S′ -- S -- bimodule. Using the theory
(cid:18)
Mn×1(D) D′ (cid:19) , where D is a division ring and D′ is a division subring of Mn(D) for some positive
integer n (Theorem 2.1). We also prove that certain conditions on D′ characterizes these triangular
rings to have no right middle class which yields a general characterization for right nonsingular right
Artinian rings to have no right middle class (Theorem 2.5). This result also enables us to produce
many interesting examples of right nonsingular right Artinian rings with no right middle class.
It is also known from [3, Corollary 6] that if R is an indecomposable right Artinian ring with no
right middle class which is not right SI, then R is isomorphic to a formal triangular matrix ring of
0
X
the form (cid:18) Mn(A)
B (cid:19) , where A is a (nonsemisimple) local right Artinian ring, B is a semisimple
Artinian ring, and X is a B -- Mn(A) -- bimodule. As a matter of fact, a right Artinian ring which is not
right SI has no right middle class if and only if R ∼= S ⊕ Mn(A) where S is semisimple Artinian and
A is a local right Artinian ring whose Jacobson radical properly contains no nonzero ideals (Theorem
2.16).
2
In Section 3, we restrict our attention to only simple modules, and consider rings whose simple
right modules are either injective or poor. Such rings are said to have no simple middle class (see [1]).
We give necessary and sufficient conditions for a right Artinian ring to have no simple middle class.
The last section of our paper is concerned with the property of having no (simple) middle class in
the commutative setting. We give a complete description of commutative rings with no middle class.
In particular, we see that a commutative ring with no middle class is Artinian. We conclude our work
with a characterization of commutative Noetherian rings to have no simple middle class.
Recall that a ring is said to be a right V -- ring if every simple right module is injective. As a
generalization of right V -- rings, right GV -- rings were introduced by Ramamurthi and Rangaswamy
in [10]. A ring is called right GV if every simple singular module is injective, or equivalently, every
simple module is either injective or projective. We call a ring right SI if every singular right module
is injective (see[2]). Note that semilocal right GV -- rings are right SI.
If M is an R -- module, then E(M ), J(M ), Z(M ) and Soc(M ) will respectively denote the injective
hull, Jacobson radical, the singular submodule and the socle of M. We will use the notations ≤ and
≤e in order to indicate submodules and essential submodules, respectively. For a module with a
composition series, cl(M ) stands for the composition length of M. The ring of n × n matrices over a
ring R will be denoted by Mn(R). The notation A[i, j] will be used to indicate the (i, j) -- th entry of
a matrix A. We use eij to designate the standard matrix unit of Mn(R) with 1 in the (i, j) -- th entry
and zeros elsewhere. For any unexplained terminology, we refer the reader to [4] and [8].
2 Artinian Rings with No Middle Class
In [3, Corollary 5], Er, L´opez -- Permouth, and Sokmez proved that if R is an indecomposable
right nonsingular right Artinian ring with no right middle class, then R is isomorphic to a formal
triangular matrix ring of the form (cid:18) S
A S′ (cid:19) , where S and S′ are simple Artinian rings and A
is an S′ -- S -- bimodule. With the following theorem, we see that to determine when such rings have
no right middle class, it is enough to consider formal triangular matrix rings of the much simpler
0
form (cid:18)
positive integer n
D
Mn×1(D) D′ (cid:19) , where D is a division ring and D′ is a division subring of Mn(D) for some
0
Theorem 2.1 If R is a right Artinian right SI ring satisfying the properties (P 1) and (P 2), then it
is Morita equivalent to a formal triangular matrix ring of the form
(cid:18)
D
Mn×1(D) D′ (cid:19) ,
0
where D is a division ring and D′ is a division subring of Mn(D).
Proof. Since R is a right Artinian ring, there exists a complete set {e1, . . . , el, f1, . . . , fm} of local
orthogonal idempotents such that the eiR's are simple and fiR's are nonsemisimple local. Then
R = e1R ⊕ ··· ⊕ elR ⊕ f1R ⊕ ··· ⊕ fmR. By (P 2), fiR ∼= fjR for all i, j. Moreover, since eiR is
nonsingular, there is no nonzero R -- homomorphism from fjR into eiR, and so eiRfj = 0 for all i, j. As
R is right SI, R/Soc(RR) is semisimple Thus, J(R) ≤ Soc(RR). This gives that J(fiR) = Soc(fiR).
We set e = e1 + f1. Then eR = e1R ⊕ f1R. We shall prove that e is a full idempotent of R,
i.e., ReR = R. It is clear that e1, f1 ∈ ReR. Since R has homogeneous right socle, eiRe1R 6= 0
for all i = 1, . . . , l. Then ei ∈ eiR = eiRe1R ≤ ReR for all i = 1, . . . , l. Now, assume that
fk /∈ fkRf1R. Then fkRf1R ≤ Soc(fkR). Since R has homogeneous right socle and eiRfj = 0
for all i, j, fkRf1Rf1 = 0. Let u : f1R → fkR be an isomorphism of right R -- modules. Then
u(f1Rf1Rf1) = u(f1)Rf1Rf1 ≤ fkRf1Rf1 = 0, and so f1 ∈ f1Rf1Rf1 = 0, a contradiction. It follows
that fk ∈ fkRf1R ≤ ReR for all k = 1, . . . , m. Therefore, ReR = R.
Now let α be an endomorphism on E(f1R), and let α′ be the restriction of α to Soc(f1R). Suppose
Soc(f1R) = B1 ⊕ ··· ⊕ Bn, where B1, . . . , Bn are simple right R -- modules isomorphic to e1R, and let
vk : Bk → e1R be an isomorphism. Set αk = ikv−1
k and βk = vkπk, where ik is the natural embedding
of Bk into Soc(f1R), and πk the natural projection of Soc(f1R) onto Bk. Then the correspondence
α
ϕ
←→ (βiα′αj)n×n
3
between End(E(f1R)) and Mn(End(e1R)) gives a ring isomorphism. Moreover, since E(f1R) is
nonsingular and E(f1R)/Soc(f1R) is singular, this correspondence yields an embedding of End(f1R)
into Mn(End(e1R)). We denote ϕ(End(f1R)) by D′. Note that D′ is a division ring in view of the
proof of [3, Corollary 5]. Now it is routine to check that the mapping
eRe ∼=(cid:18)
0
End(e1R)
Hom(e1R, f1R) End(f1R) (cid:19) −→ (cid:18)
u2 u3 (cid:19) 7−→ (cid:18)
(cid:18) u1
0
0
D
Mn×1(D) D′ (cid:19) ,
(βku2)n×1 ϕ(u3) (cid:19) ,
u1
0
where D denotes End(e1R), is an isomorphism of rings. This completes the proof.
From now on, we will denote the formal triangular matrix ring
(cid:18)
D
Mn×1(D) D′ (cid:19)
0
by (D, Dn, D′).
Let R = (D, Dn, D′), where D is a division ring and D′ is a division subring of Mn(D). Then
R = (D, 0, 0) ⊕ (0, Dn, D′), where (D, 0, 0) is a simple right ideal and (0, Dn, D′) is a local right ideal
with the maximal submodule (0, Dn, 0).
Define (D, Mn(D))i as the set of ordered pairs (a, A), where a ∈ D, A ∈ Mn(D), and a scalar
multiplication
(a, A)(a1, (xk), A1) = (aa1 + A(i)(xk), AA1),
where (a1, (xk), A1) ∈ R and A(i) denotes the ith row of A. It is easy to see that (D, D′)i is a right
R -- submodule of (D, Mn(D))i. Also, if Bi = ⊕j6=iejD, where {e1, . . . , en} is the natural basis for Dn
over D, then (D, D′)i is isomorphic to
.
(0, Dn, D′)
(0, Bi, 0)
Lemma 2.2 u : (D, D′)i → (D, Mn(D))j is a nonzero R -- homomorphism if and only if there exist
d0 ∈ D and A0 ∈ Mn(D) such that A0[j, k] = δkid0 (k = 1, . . . , n) and u(d, A) = (d0d, A0A) for all
(d, A) ∈ (D, D′)i.
Proof. Straightforward.
Lemma 2.3 [3, Lemma 1] The property of having no (simple) middle class is inherited by factor
rings.
Lemma 2.4 Suppose a ring R = S ⊕ T is a direct sum of two rings S and T, where S is semisimple.
Then R has no (simple) middle class if and only if T has no (simple) middle class.
Proof. Let M be an N -- injective right R -- module, where N is cyclic and nonsemisimple. Then N is
isomorphic to a direct sum S/A ⊕ T /B of right R -- modules for some right ideals A and B contained
in S and T , respectively. Note that T /B is not semisimple (as both R -- and T -- modules). Since
M = M S ⊕ M T, M T is (T /B) -- injective as both R -- and T -- modules. By assumption, M T is an
injective right T -- module. However, it is not difficult to see that it is also injective as an R -- module.
We may also show, in a similar way, that M S is an injective right R -- module. This gives that M is an
injective R -- module. Thus we established the sufficiency part. The necessity is obvious by the above
lemma.
Let S be a subring of a ring R and u a unit in R. Obviously, uSu−1 is a subring of R isomorphic
to S as a ring. We call uSu−1 a conjugate ring of S in R.
Theorem 2.5 Let R be a right nonsingular right Artinian ring. Then R has no right middle class
if and only if R ∼= S ⊕ T, where S is a semisimple Artinian ring and T is zero or Morita equivalent
to a formal triangular matrix ring of the form (D, Dn, D′) where D is a division ring and D′ is a
division subring of Mn(D) such that for each conjugate ring U of D′ in Mn(D), the set of i -- th rows
of elements in U span Dn as a left D -- space for every i = 1, . . . , n.
4
Proof. Assume first that R has no right middle class. By [3, Theorem 2], R ∼= S ⊕ T, where S is
a semisimple Artinian ring and T is zero or a right Artinian right SI ring satisfying the properties
(P 1) and (P 2). Suppose T is not zero. Then by Theorem 2.1, T is Morita equivalent to a formal
triangular matrix ring of the form (D, Dn, D′), where D is a division ring and D′ is a division subring
of Mn(D). Since the property of having no right middle class is a Morita invariant property (as
remarked, for example, in [3] before Proposition 5), the ring (D, Dn, D′) has no right middle class.
j=1Deij )1 for all i = 1, . . . , n. By Lemma 2.2, (D, De1kD′)1 is (D, D′)k --
Note that (D, 0)1 ≤e (D,⊕n
injective as a right R -- module for all k = 1, . . . , n. Since R has no right middle class and (D, D′)k
is nonsemisimple, (D, De1kD′)1 must be injective for all k = 1, . . . , n. On the other hand, we have
(D, 0)1 ≤ (D, De1kD′)1 ≤ (D,⊕n
j=1De1j for all k = 1, . . . , n.
It therefore follows that the set of the k -- th rows of elements of D′ span Dn as a left D -- space for all
k = 1, . . . , n. Now let U = uD′u−1 for some unit u in Mn(D). Obviously, R ∼= (1, 0, u)R(1, 0, u−1) =
(D, Dn, U ). Then (D, Dn, U ) has no right middle class. Repeating the above arguments, we complete
the proof of the necessity part.
j=1De1j)1, which gives that De1kD′ = ⊕n
For the sufficiency, it is enough, by Lemma 2.4, to show that if, for each conjugate ring U of D′ in
Mn(D), the set of i -- th rows of elements in U span Dn as a left D -- space for every i = 1, . . . , n, then
the ring (D, Dn, D′) has no right middle class. Assume the contrary, i.e., assume that (D, Dn, D′)
has a right middle class. Note that the maximal right quotient ring of R is Q = Mn+1(D). By [3,
Proposition 8], there exists X ≤ QR which contains the right socle of (D, Dn, D′) properly such that
QX 6= Q. Since the right socle of the ring (D, Dn, D′) is (D, Dn, 0), there exists a nonzero right
D′ -- submodule Y of M(n+1)×n(D) such that
X =(cid:18) D
Dn
Y (cid:19) .
One can observe that if, for each i = 1, . . . , n, there exists an element of Y (depending on i) whose
i -- th column has a nonzero entry, then QX = Q. Thus, there exists j such that the j -- th column of
each element of Y is zero. Without loss of generality, we may assume j = 1. Then there exist elements
d1, . . . , dn−1 of D which are not all zero such that d1A[2, 1] +··· + dn−1A[n, 1] = 0 for all A ∈ D′. We
may choose dn−1 6= 0. Let B = e11 + (Pn
i=2 eii + di−1eni). Obviously, B is invertible in Mn(D), and
all elements of BD′B−1 have zero in the (n, 1) -- th entry. It follows that the n -- th rows of elements of
BD′B−1 cannot span Dn. This completes the proof.
Corollary 2.6 (i) Let D be a division ring and D′ a division subring of D. Then the ring(cid:18) D 0
D D′ (cid:19)
has no right middle class.
(ii) Let D be a division ring and D′ a division subring of M2(D). Then the ring R = (D, D2, D′)
has no right middle class if and only if the set of the ith rows of elements of D′ span D2 for i = 1, 2.
Proof. (i) Clear by Theorem 2.5. (ii) If we take n = 2 in the proof of Theorem 2.5, then we deduce
that when R has right middle class, there exists a nonzero element d ∈ D such that dA[2, 1] = 0 for
all A ∈ D′. This contradicts the fact that the set of the second rows of elements of D′ span D2.
It is shown in [3, Proposition 6] that if R is a right Artinian right SI -- ring with homogeneous right
socle and a unique local module of length two up to isomorphism, then R has no right middle class.
Example 2.9 below shows that the converse of this fact is not true in general. Before the example, we
need the following proposition.
Proposition 2.7 Let D be a division ring and D′ be a division subring of Mn(D). Let R = (D, Dn, D′)
and let R be the set of the first rows of all elements of D′. Then R has a unique local right R -- module
of length two up to isomorphism if and only if S(xi)∈R D(xi) = Dn.
Proof. Let M be a local right R -- module of length two. Then there exists an epimorphism g : R → M ,
and since M is local, we must have g(0, Dn, D′) = M . It follows that there exists a maximal submodule
A of (0, Dn, 0) such that A = Ker(g). Thus the unique simple submodule of M , say N , is isomorphic
to (D, 0, 0).
j=1 De1j)1 is injective relative to (D, D′)i for each i = 1, . . . , n, by Lemma 2.2.
j=1 De1j)1
j=1 De1j)1 is the injective hull of the local right R --
module (D, e11D′)1. There is an isomorphism from N to (D, 0)1 which extends to a homomorphism f
Observe that (D,Ln
Since R can be embedded into the sum (D, 0, 0)⊕(D, D′)1⊕···⊕(D, D′)n, we get that (D,Ln
is an injective R -- module. It follows that (D,Ln
5
j=1 De1j)1.
j=1 De1j)1 is nonsingular and M/N is singular, f must be
a monomorphism. It therefore follows that every local right R -- module of length two can be embedded
j=1 De1j)1. Since (D,Ln
from M into (D,Ln
into (D,Ln
Let M ′ be a local submodule of (D,Ln
right D′ -- subspace X of Ln
This shows that any local right R -- submodule of length two in (D,Ln
(D, (Pn
isomorphism from (D, (Pn
j=1 De1j)1 of length two. Then M ′ = (D, X)1 for a
j=1 De1j. Since M ′ is local, X must be one -- dimensional as a D′ -- space.
j=1 De1j)1 is of the form
j=1 e1jdj)D′)1 for some d1, . . . , dn ∈ D. Moreover, one can also prove that there exists an
j=1 e1jdj)D′)1 onto (D, e11D′)1 if and only if there exists a nonzero d ∈ D
i=1 if
i=1 ∈ R such that (di) = d(xi)n
such that d(di)n
and only if (di)n
i=1 ∈ R if and only if there exists d ∈ D and (xi)n
i=1 ∈S(xi)∈R D(xi). Now the result follows.
Remark 2.8 It can be easily seen from Corollary 2.6 (ii) that, for a division ring D and a division
subring D′ of M2(D), the ring
D
D
D
0
D′
has right middle class if and only if either all elements of D′ are lower triangular matrices or all
elements of D′ are upper triangular matrices. Suppose, in particular, that all elements of D′ are
upper triangular matrices. Then for every [aij ] ∈ D′, a22 is uniquely determined. Thus, we have a
mapping from D′ to D such that [aij] 7→ a22 for all [aij ] ∈ D′ which is a ring monomorphism, that is,
D′ can be embedded into D as a ring.
Example 2.9 (i) If D is a division ring and D′ is a division subring of M2(D) consisting only of
lower triangular matrices, then the ring (D, D2, D′) is a right Artinian right SI ring which satisfies the
properties (P 1) and (P 2). However, by Remark 2.8, (D, D2, D′) has right middle class. For instance,
if we let δ be a derivation on D and consider the division subring
D′ =(cid:26)(cid:18) a δ(a)
a (cid:19) a ∈ D(cid:27)
0
of M2(D), then the ring (D, D2, D′) has right middle class. Thus the converse of Theorem 2 of [3] is
not true, in general.
(ii) Let D = Z3 and w =(cid:18) 1
1
2
1 (cid:19) . Observe that D′ = {0, 1, w, w2, . . . , w7} is a field. By Remark
2.8, the ring (D, D2, D′) has no right middle class.
Example 2.10 Let p be a prime integer, F ≤ F1 a field extension, and K a division subring of Mp(F )
which properly contains the field of scalar matrices in Mp(F ). Then the ring
R = F1
1 K !
F p
0
If, in particular, we take F = Q, then any division subring of Mp(Q)
has no right middle class.
contains all scalar matrices. It follows that R has no right middle class for any division subring K of
Mp(Q) which is not the field of scalar matrices.
Proof. We first claim that for any i = 1, 2, . . . , p, the i -- th rows of all elements of K span F p as
an F -- space. To see this, let A be an element of K which is not a scalar matrix and let mA(x)
be the minimal polynomial of A over F. Let mA(x) = u(x)v(x), where u(x), v(x) ∈ F [x]. Then
0 = mA(A) = u(A)v(A), which implies that one of the determinants det(u(A)) or det(v(A)) is zero.
Assume that det(u(A)) = 0. Since K contains all scalar matrices over F, we have u(A) ∈ K. But K is
a division ring which means that every nonzero matrix in K has nonzero determinant. This gives that
u(A) = 0. Since deg(u(x)) ≤ deg(mA(x)) and mA(x) is the monic polynomial of least degree which
assumes A as a root, we must have u(x) and mA(x) are associates. It follows that mA(x) is irreducible
over F. Then the characteristic polynomial cA(x) of A over F is a power of mA(x). This implies that
deg(mA(x)) divides deg(cA(x)) = p. Since p is prime, deg(mA(x)) is either 1 or p. If deg(mA(x)) = 1,
then A is similar to a scalar matrix, B say. In other words, there exists a p × p invertible matrix
6
P over F such that P −1AP = B. Thus A = P BP −1 = BP P −1 = B, a contradiction. Therefore
mA(x) ∈ F [x] is an irreducible polynomial of degree p.
Now we shall show that for any i = 1, 2, . . . , p, the i -- th rows of the matrices I, A, . . . , Ap−1 span
F p as an F -- space. In order to prove this, without loss of generality, we may choose i = 1. Assume
the contrary, i. e., the first rows of the matrices I, A, . . . , Ap−1 do not span F p. Then the first rows of
these matrices should be linearly dependent. So, there exist scalar matrices c0, c1, . . . , cp−1 over F, not
all zero, such that the first row of the matrix C = c0 + c1A + ··· + cp−1Ap−1 is zero. This gives that
det(C) = 0. Since C lies in K, we must have C = 0. But then A happens to be a root of a polynomial
over F of degree at most p − 1, a contradiction. Consequently, for any i = 1, 2, . . . , p, the i -- th rows
of all elements of K span F p as an F -- space. This gives that the F -- space spanned by the i -- th rows
of elements of K contains the standard basis which is also contained in the F1 -- space spanned by the
i -- th rows of elements of K. Therefore, for any i = 1, 2, . . . , p, the i -- th rows of all elements of K span
F p
1 as an F1 -- space. This fact is true for any conjugate of K in Mp(F ) because, just as K, it also
properly contains the field of scalar matrices over F. The proof is complete by Theorem 2.5.
Remark 2.11 Let p be a prime number and F be a field. Let P (x) be an irreducible polynomial over
F of degree p (if exists) and let A ∈ Mp(F ) be such that P (A) = 0 (one can use the companion
matrix of P (x) from linear algebra to find such A). Then the set K = {c0 + c1A + ··· + cp−1Ap−1 :
c0, c1, . . . , cp−1 ∈ F} is a field isomorphic to F [x]/(P (x)) and properly contains the field of scalar
matrices over F. Indeed, we can show that all division subrings K of Mp(F ) properly containing the
field of scalar matrices are of this form: Let A be an element of K which is not a scalar matrix and
let B be any nonzero element of K. Since the set of the first rows of I, A, . . . , Ap−1 is a basis for F p,
we must have, by the same reasoning used in the proof of Example 2.10, there exist scalar matrices
c0, c1, . . . , cp−1, cp over F, not all zero, such that c0 + c1A + ··· + cp−1Ap−1 + cpB = 0. Here, clearly,
cp 6= 0. It follows that B is a linear combination of the powers Ai of A, where i = 0, 1, . . . , p − 1.
Therefore K = {c0 + c1A + ··· + cp−1Ap−1 : c0, c1, . . . , cp−1 ∈ F}.
Example 2.12 Let F be a field and P (x) be an irreducible polynomial over F of prime degree. Then
the ring
has no right middle class.
F
F [x]/(P (x)) F [x]/(P (x)) !
0
Proof. Let deg(P (x)) = p. Set K = {c0 + c1A + ··· + cp−1Ap−1 : c0, c1, . . . , cp−1 ∈ F}, where A is
the companion matrix of P (x) over F. It is routine to check that the mapping
F
F [x]/(P (x)) F [x]/(P (x)) ! −→ F
F p K ! ,
0
0
is a ring isomorphism. The result follows from Example 2.10 and Remark 2.11.
Example 2.13 The ring
R =
C
C
C
0
C
, where C =(cid:26)(cid:18) a −b
b
a (cid:19) a, b ∈ R(cid:27) ∼= C,
has no right middle class. However, R has at least two nonisomorphic local right R -- modules of length
two by Proposition 2.7. Therefore, the converse of [3, Proposition 6] is not true, in general.
Note that if a ring of the form (D, Dn, D′), where D and D′ are as above, has no right middle
class, then the property of having no right middle class of (D, Dn, D′) remains unaltered when we
7
by
a
p−1
Pi=0
cixi
0
p−1
dixi
Pi=0
7−→
a
c0
c1
...
cp−1
0
diAi
p−1
Pi=0
replace D by any division ring containing D. However, such a replacement may result in an increased
number of local modules of length two. On the other hand, as the following theorem shows, for the
ring R = (D, Dn, D′), it is necessary that certain local R -- modules of length two are isomorphic, which
is also sufficient when n = 2.
Lemma 2.14 Let 1 ≤ i, j ≤ n. Then (D, D′)i ∼= (D, D′)j if and only if D′ contains an element A
such that A[j, k] = δikc for some nonzero c ∈ D.
Proof. The proof is straightforward by Lemma 2.2.
Theorem 2.15 Let D be a division ring and D′ a division subring of Mn(D). If the ring
R =(cid:18)
D
Mn×1(D) D′ (cid:19)
0
has no right middle class, then there exists a conjugate D′′ of D′ in Mn(D) such that (D, D′′)i ∼=
(D, D′′)j for each i, j = 1, . . . , n as right (D, Dn, D′′) -- modules. In particular, if n = 2, then R has no
right middle class if and only if (D, D′)1 ∼= (D, D′)2.
Proof. Since the ring R has no right middle class, by Theorem 2.5, D′ has an element A such that
A[1, n] 6= 0. By standard techniques of linear algebra, it is not difficult to see that there exists a unit
y1 in Mn(D) such that B[1, i] = 0 for all i = 1, . . . , n − 1 and B[1, n] = A[1, n], where B = y1Ay−1
1 .
Now, pick an element A′ of D′ such that A′[1, n − 1] 6= 0. Then there exists a unit y2 ∈ Mn(D) such
that B′[1, i] = 0 for 1 ≤ i ≤ n, i 6= n− 1, and B′[1, n− 1] = A′[1, n− 1]. Notice that conjugating B by
y2 does not effect the current form of B, i.e., y2By−1
is still a matrix whose (1, n) -- th entry is nonzero
and (1, i) -- th entry is zero for every i = 1, . . . , n − 1. If we continue in this fashion, we may find a unit
y in Mn(D) such that the ring yD′y−1 contains elements A2, . . . , An, where Ai[1, k] = δikci for some
0 6= ci ∈ D for all i = 2, . . . , n. Now the first part of the theorem follows by Lemma 2.14. Moreover,
the remaining part also follows from the first part together with Corollary 2.6.
2
Our aim in concluding this section is to complete our investigation of rings which have no right
middle class. From Theorem 1 given in the introductory part, we know exactly what it means for a
right SI -- ring to have no right middle class . With the following theorem, we determine how precisely
rings with no right middle class which are not right SI look like.
Theorem 2.16 Let R be a ring which is not right SI. Then R has no right middle class if and only
if R ∼= S ⊕ Mn(A), where S is a semisimple Artinian ring, n is a positive integer, and A is either zero
or a local right Artinian ring whose Jacobson radical properly contains no nonzero ideals.
Proof. The sufficiency follows from [9, Corollary 2.14]. For the necessity, suppose that R is a ring with
no right middle class which is not right SI. By [3, Theorem 2], R ∼= S ⊕ T, where S is a semisimple
Artinian ring and T is zero or it is a ring as in Theorem 2(iii) of [3]. If T = 0, then we are done. Let T
be nonzero. Then by [3, Corollary 6], T ∼=(cid:18) Mn(A)
B (cid:19) , where A is a (nonsemisimple) local right
Artinian ring, B is a semisimple Artinian ring, and X is a B -- Mn(A) -- bimodule. Note that T has no
right middle class, too. As J(A) 6= 0, by [9, Corollary 2.14], we must have X = 0 and J(A) properly
contains no nonzero ideals. Since T is indecomposable, we must also have B = 0. This completes the
proof.
X
0
Corollary 2.17 Let R be a right Noetherian ring. Then R has no right middle class if and only if
R ∼= S ⊕ T, where S is a semisimple Artinian ring and T is zero or it is Morita equivalent to one of
the following rings:
(i) a right P CI -- domain, or
(ii) a formal triangular matrix ring of the form (D, Dn, D′) where D is a division ring and D′ is
a division subring of Mn(D) such that for each conjugate ring U of D′ in Mn(D), the set of i -- th rows
of elements in U span Dn as a left D -- space for every i = 1, . . . , n, or
(iii) a local right Artinian ring whose Jacobson radical properly contains no nonzero ideals.
Proof. The sufficiency follows from [3, Proposition 5], Theorem 2.5, and Theorem 2.16. For the
necessity, assume that R is a right Noetherian ring which has no right middle class. If R is right
Artinian, then by [3, Theorem 2], Theorem 2.5, and 2.16, R is Morita equivalent to a ring which
belongs to the class of rings in (ii) or (iii). If R is not right Artinian, then, by [3, Theorem 2], it is
either Morita equivalent to a right P CI -- domain or a V -- ring with essential socle. Since R is right
Noetherian, in the latter case R is semisimple Artinian. This completes the proof.
8
Corollary 2.18 Let R be any ring. Then R has no right middle class if and only if R ∼= S ⊕ T where
S is a semisimple Artinian ring and T is zero or it satisfies one of the following conditions:
(i) T is Morita equivalent to a right P CI -- domain, or
(ii) T ∼= Mn(A), where n is a positive integer, and A is a local right Artinian ring whose Jacobson
radical properly contains no nonzero ideals, or
(iii) T is an indecomposable right SI -- ring with homogeneous essential right socle which satisfies
one the following equivalent conditions (where Q is the maximal right quotient ring of T ):
(a) Non-semisimple quasi -- injective right T -- modules are injective.
(b) Proper essential submodules of QT are poor.
(c) For any submodule A of QT containing Soc(TT ) properly, QA = Q.
Proof. It follows from [3, Theorem 2 and Proposition 8] and Theorem 2.16.
Remark 2.19 Let R be a nonsemisimple right Artinian ring with no right middle class. For the sake
of simplicity, we assume that R is indecomposable. We know that either R is a QF -- ring with Jacobson
radical square zero or RR is poor. If R is a QF -- ring, then it is not right SI (because a nonsemisimple
right SI -- ring with no right middle class cannot be right self -- injective), and so, by Theorem 2.16,
R ∼= Mn(A) for some local right Artinian ring whose Jacobson radical properly contains no nonzero
ideals. Then A is also QF. If I is a nonzero right ideal of R, then J = r.ann(l.ann(I)) = I. This
gives that cl(RR) = 2, i.e., R is a right chain ring with right composition length two. However, there
are Artinian rings with no right (and left) middle class which are poor as a right (and left) module
although they have right (or left) composition length two (see Example 2.20).
Example 2.20 Let F be the field Q(√2) and α : F −→ Q defined by α(a + b√2) = a. Let R = F × F
as additive abelian group and define the multiplication on R as
(u, v)(w, z) = (uw, uz + vα(w)).
Then R is a noncommutative local Artinian ring with cl(RR) = 3 and cl(RR) = 2. Since cl(RR) = 2,
J(R) properly contains no nonzero ideals of R. It follows that R has no right (and left) middle class.
Also, R does not satisfy the double annihilator condition for right ideals. Then R cannot be QF, i.e.,
both RR and RR are poor.
Obviously, a ring with no right middle class is either right self -- injective or poor as a right module
over itself. It is known, from [3, Proposition 9], that the condition of Theorem 2(iii) in [3] are sufficient
if the ring is taken to be right self -- injective. The next two examples illustrate that these conditions
are not sufficient in general even if the ring is poor. In particular, these examples also indicate that
there are rings satisfying the conditions of Theorem 2(iii) in [3] which are not of the form Mn(A),
where A is as in Theorem 2.16. We first need the following lemma.
Lemma 2.21 Let R be a right semiartinian ring. Then RR is poor if and only if R is not injective
relative to a local right R -- module of length two.
Proof. The necessity is obvious. For the sufficiency, suppose that R is not injective relative to a local
right R -- module of length two. Let R be M -- injective. Without loss of generality, we may choose M
cyclic. Assume that M is not semisimple. Then there exists a local subfactor N of M of length two
since M is semiartinian. This gives that R is N -- injective, a contradiction.
Example 2.22 Let F be a field and V a finite dimensional vector space over F of dimension greater
than 1. Let R = (cid:26)(cid:18) a 0
v a (cid:19) : a ∈ F, v ∈ V(cid:27) . Then R is a commutative local Artinian ring which
satisfies the conditions of Theorem 2(iii) in [3]. However, R has right middle class by [9, Corollary
2.14]. Note too that R is poor as a module over itself. Indeed, if L is a local module of length two
with simple submodule S, then there is an isomorphism from S into R which cannot be extended to a
homomorphism from L. Otherwise, R would contain a local module of length two, which is impossible
since dimF (V ) > 1. Then R is not L -- injective. Hence, by Lemma 2.21, R is poor.
9
Example 2.23 Let R =(cid:18) Z/4Z
Z/2Z Z/2Z (cid:19) . Then
0
(1) Soc(RR) = J(R) = Z(RR),
(2) R has essential homogeneous right socle,
(3) there is a unique noninjective simple right R -- module up to isomorphism,
(4) RR is poor, and
(5) R has right middle class.
Proof. (1) It follows from [5, Proposition 4.2] that Soc(RR) = J(R) = Z(RR) =(cid:18) 2Z/4Z 0
Z/2Z 0 (cid:19)
(2) Since R is right Artinian, Soc(RR) ≤e RR. On the other hand, Soc(RR) = (cid:18) 2Z/4Z 0
(cid:18) 0
Z/2Z 0 (cid:19) , where simple summands are isomorphic.
(3) Since R/J(R) ∼= S ⊕ S′, where
0
0
0 (cid:19) ⊕
S =(cid:18) Z/2Z 0
0
0
0 (cid:19) , and S′ = (cid:18) 0
Z/2Z Z/2Z (cid:19)
(cid:18) 0
Z/2Z 0 (cid:19)
0
,
any simple right R -- module is isomorphic to either S or S′. As Soc(RR) ≤e RR, SR cannot be injective.
To establish (3), we shall show that S′
R is injective. Note that all proper essential right ideals of R are
0
0
0
0
0
0
I1 =(cid:18) 2Z/4Z 0
Z/2Z 0 (cid:19) , I2 =(cid:18) Z/4Z 0
Z/2Z 0 (cid:19) , and I3 =(cid:18) 2Z/4Z
Z/2Z Z/2Z (cid:19) .
If f : I1 −→ S is a nonzero homomorphism of R -- modules, then f (I1) = S. But I1(cid:18) 0 0
0 1 (cid:19) = 0 while
S(cid:18) 0
1 (cid:19) 6= 0, a contradiction. Thus HomR(I1, S) = 0. Similarly, HomR(I2, S) = 0. Now, let f :
I3 −→ S be a nonzero homomorphism. Observe that there are only two maximal right R -- submodule
of I3 : Soc(RR) and M =(cid:18) 0
Z/2Z Z/2Z (cid:19) . Since I3/M ≇ S, we must have Ker(f ) = Soc(RR). It
follows that there exists a unique nonzero homomorphism f : I3 −→ S, which can be extended to a
homomorphism R −→ S. Therefore, SR is injective.
(4) Note that there are two local right R -- modules of length two up to isomorphism: X =
(cid:18) Z/4Z 0
Z/2Z Z/2Z (cid:19) . Indeed, we can decompose RR as R = X ⊕ Y. Notice
that both X and Y are local right R -- modules of length two. Now let M be a local right R -- module
of length two. Then there exists an epimorphism f : R −→ M. Since M is local, f (X) = M or
f (Y ) = M. This gives that X ∼= M or Y ∼= M.
x 0 (cid:19) ,
where 0 6= x ∈ Z/2Z, is an R -- homomorphism which does not extend to a homomorphism f : X −→ R.
Indeed, if f(cid:18) 1
0 (cid:19) −→ R defined by(cid:18) 2
0 (cid:19) =(cid:18) 2a
R is not X -- injective because the map(cid:18) 2Z/4Z 0
0 (cid:19) and Y = (cid:18) 0
b (cid:19) , then f(cid:18) 2
0 (cid:19) =(cid:18) a 0
0 (cid:19) 7−→(cid:18) 2
1 0 (cid:19) 7−→
Now we claim that R is not Y -- injective. To see this, let f :(cid:18) 0
1 0 (cid:19) . Then f is a homomorphism which cannot extend to a homomorphism g : Y −→ R. Indeed,
(cid:18) 2 0
if g(cid:18) 0
0 (cid:19) =
0 (cid:19) and
(cid:18) a 0
(cid:18) 0 0
b (cid:19) , then a = y = 0 since(cid:18) 0
1 (cid:19) =(cid:18) 0
1 (cid:19)(cid:18) 1
b (cid:19) . But we also have(cid:18) 0
1 (cid:19) =(cid:18) a 0
0 (cid:19) . Thus, g(cid:18) 0
b (cid:19)(cid:18) 0
0 (cid:19) = 0 and(cid:18) a 0
1 (cid:19)(cid:18) 0
b (cid:19)(cid:18) 1
0 (cid:19) =(cid:18) 0
0 (cid:19) , a contradiction.
0 (cid:19) =(cid:18) 0
0
0
2b (cid:19) , a contradiction.
Z/2Z 0 (cid:19) −→ R with(cid:18) 0 0
1
0
y
0
y
0
0
b
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
y
0
0
y
By Lemma 2.21, RR is poor.
(5) By [9, Corollary 2.14], R has right middle class.
10
3 Artinian Rings With No Simple Middle Class
Notice that right V -- rings have automatically no simple middle class. In the theorem below, we
consider right GV -- rings without simple middle class whose proof uses almost the same arguments as
those used in the proof of [3, Lemma 8]. Before giving the theorem, we need the following lemma.
Lemma 3.1 Suppose that R is not a right V -- ring. Then R is a right GV -- ring with no simple middle
class if and only if R has a simple projective poor module.
Proof. It follows from [1, Corollaries 4.4 and 4.5].
As a direct consequence of Lemma 3.1, we have the following lemma.
Lemma 3.2 Suppose that a ring R has no simple middle class. Then R is a right GV -- ring or every
simple projective right module is injective.
The proof of the following theorem uses the notion of orthogonal modules. Recall that two modules
are said to be orthogonal if they have no nonzero isomorphic submodules.
Theorem 3.3 Suppose that R is a right nonsingular right GV -- ring which is not a right V -- ring. If
R has no simple middle class, then there is a ring decomposition R = S ⊕ T , where S is semisimple
Artinian and the right socle of T is nonzero poor homogeneous. If, further, RR has finite uniform
dimension, then the converse also holds.
Proof. Claim 1 : Soc(RR) does not contain a direct sum of two infinitely generated orthogonal
submodules.
Assume that A and B are infinitely generated orthogonal submodules of Soc(RR). Then A and
B are noninjective modules. Let f be a homomorphism from E(B) into E(A). We will show that
Ker(f ) ≤e E(B). Let X be a nonzero submodule of E(B). Then X ∩ B 6= 0. If we assume that
Ker(f ) ∩ X = 0, then we get 0 6= X ∩ B ∼= f (X ∩ B). But this contradicts the fact that A and B are
orthogonal. Hence, Ker(f ) ∩ X 6= 0. It follows that E(B)/Ker(f ) ∼= Im(f ) is singular, and hence
Im(f ) = 0. Thus, A is E(B) -- injective. Since A is poor by [1, Corollary 4.5], B is injective, which is
a contradiction.
Claim 2 : One of the two nonisomorphic simple right ideals is injective.
If S1 and S2 are two nonisomorphic noninjective simple right ideals, then S1 is E(S2) -- injective
which implies by assumption that either E(S2) is semisimple or S1 is injective. This gives that either
S1 or S2 is injective.
By the same technique above, one can observe that a simple right ideal which is orthogonal to an
infinitely generated semisimple right ideal is injective. Thus, Soc(RR) can have only finitely many
homogeneous components. Let H1, H2, . . . , Hn be the homogeneous components of Soc(RR). Notice
that all the Hi's will have to be injective except possibly for at most one of them. If n = 1 and H1 is
noninjective, then Soc(RR) is homogeneous and poor, and so we are done. Now suppose n > 1, H1 is
either noninjective or zero, and H2, . . . , Hn are injective. Set S = H2 ⊕ ··· Hn. Then R = S ⊕ T for
some right ideal T .
Claim 3 : R = S ⊕ T is a ring direct sum, where Soc(TT ) is nonzero poor homogeneous.
Obviously, T S = 0. If ST 6= 0, then there exists s ∈ S such that sT 6= 0. But Soc(TT ) ≤ annr(s)
and ST ≤ S. We will show that X = annr(s) ∩ T ≤e T . Let I ≤ T such that X ∩ I = 0. Define
f : I → R, x 7→ sx. Clearly I ∼= Im(f ) = sI, and hence I = 0. Therefore, sT ∼= T /(annr(s) ∩ T )
is zero, which gives that ST = 0. Thus, we obtain a ring decomposition R = S ⊕ T , where S is
semisimple Artinian and Soc(TT ) ∼= H1 is homogeneous, and poor if H1 is nonzero. It is routine to
check that T is a right GV -- ring, too. If Soc(TT ) is zero, then T has to be a right V -- ring. But this
leads to the fact that R is a right V -- ring, a contradiction.
For the last statement, suppose that RR has finite uniform dimension and Soc(TT ) is poor homo-
geneous. Then Soc(TT ) is a direct sum of finitely many isomorphic simple right ideals. This gives
that simple right ideals of T are poor, by assumption. On the other hand T is also a GV -- ring. Since
a simple module is either projective or singular and a simple projective right T -- module is isomorphic
to a simple right ideal of T, we get that T has no simple middle class. Now the result follows from
Lemma 2.4.
11
Corollary 3.4 Let R be a nonsemisimple ring with no simple middle class. If R is a right semiar-
tinian right GV -- ring which is not a right V -- ring, then we have a ring decomposition R = S ⊕ T ,
where S is semisimple Artinian and Soc(TT ) is poor homogeneous.
Proof. Since right semiartinian right GV -- rings are nonsingular the result follows from Theorem 3.3.
Corollary 3.5 [1, Theorem 4.7] Let R be a semiperfect right GV -- ring. If R has no simple middle
class, then we have a ring decomposition R = S ⊕ T , where S is semisimple Artinian and T is a
semiperfect ring with homogeneous projective and poor right socle.
Proof. Since semilocal right GV -- rings are right semiartinian, the proof follows from Corollary 3.4.
Proposition 3.6 Let R be a right semiartinian ring with a singular right socle. If R has no simple
middle class, then R is an indecomposable ring with unique noninjective simple R -- module up to
isomorphism and Soc(RR) is homogeneous.
Proof. Since Soc(RR) 6= 0, R is not a GV -- ring.
If a simple right ideal is injective, then it is
projective. But this is a contradiction since simple right ideals are singular. Consequently, every
simple right ideal of R is noninjective. Let S be a simple right ideal and M be any noninjective
simple singular right R -- module. Since R is right semiartinian, there exists N ≤ E(M ) such that M
is maximal in N . Since S is poor, there exists a homomorphism f : N → E(S) such that f (N ) * S.
Then S ( f (N ). Since N has composition length two whereas f (N ) has at least two, f must be
monic. Thus, S ∼= M . Noninjective simple modules are not projective because of Lemma 3.2, whence
they are singular. Thus, R has a unique noninjective simple module up to isomorphism. Moreover,
since Soc(RR) is a direct sum of noninjective simple singular right ideals, it is homogeneous. Also, it
is clear that a semiartinian ring with homogeneous socle is indecomposable.
Theorem 3.7 If R is a right Artinian nonsemisimple ring with no simple middle class, then R has
a ring decomposition R = S ⊕ T , where S is semisimple Artinian and Soc(TT ) is poor homogeneous.
Moreover, Soc(RR) is either projective or singular.
Proof. If R is a right GV -- ring, then we are done with by Theorem 3.3. Suppose that R is not a right
GV -- ring. Then every simple projective module is injective by Lemma 3.2. Write Soc(RR) = P ⊕ N ,
where P is the sum of all simple projective right ideals of R.Then P is injective. Therefore, R = P ⊕K
for some right ideal K. Since N does not have any simple projective summand, P is an ideal of R. Now
we will show that R = P ⊕ K is a ring direct sum. Obviously, KP = 0. Assume that P K 6= 0. Then
there exists p ∈ P such that pK 6= 0. But Soc(K) ⊆ annr(p). We claim that X = annr(p)∩ K ≤e K.
Let I ≤ K such that X ∩ I = 0. Define f : I → R such that x 7→ px for all x ∈ I. Then
I ∼= Im(f ) = pI, and hence I = 0. Note that P is nonsingular since it is semisimple projective. It
follows that pK ∼= K/(annr(p) ∩ K) is both singular and nonsingular, and hence P K = 0. Thus,
R = P ⊕ K is a ring direct sum. Then K ∼= R/P has no simple middle class by Lemma 2.3. K is
nonzero because R is nonsemisimple. Hence, Soc(K) is nonzero, too. Also, Soc(K) ∼= N is singular.
By Proposition 3.6, K is an indecomposable ring with singular poor homogeneous right socle.
Theorem 3.8 Let R be a right Artinian ring. Then R has no simple middle class if and only if there
is a ring decomposition R = S ⊕ T where S is semisimple Artinian and T is zero or has one of the
following properties:
(1) T is a right SI -- ring with homogeneous right socle.
(2) T has a unique noninjective simple right R -- module up to isomorphism, and the right socle of
T is (homogeneous) singular.
Proof. (⇒) It follows from Lemma 3.2, Theorem 3.3, Proposition 3.6 and Theorem 3.7.
(⇐) Let T be a nonzero ring which is not a V -- ring and assume that it satisfies (1). Since T is
right Artinian, we have a decomposition T = e1T ⊕ . . . ⊕ enT ⊕ f1T ⊕ . . . ⊕ fkT , where ei and fj
form a complete set of local orthogonal idempotents, eiT are isomorphic simple right ideals, and fjT
are nonsimple local T -- modules. Since T is right SI, the simple modules of the form fjT /fjJ are
injective, where J denotes the Jacobson radical of T . Therefore, if a right module does not contain
an isomorphic copy of eiT , then it is semisimple. Now assume that eiT is M -- injective, where M is a
12
cyclic right module. Then we have a decomposition M = A1 ⊕ . . . ⊕ Ap ⊕ B1 ⊕ . . . ⊕ Bq, where Ak
and Bt are indecomposable modules such that the Ak's do not contain an isomorphic copy of eiT and
the Bt's contain an isomorphic copy of eiT . By the above argument, A1 ⊕ . . . ⊕ Ap is semisimple. On
the other hand, eiT is Bt -- injective. One can observe that B1 ⊕ . . . ⊕ Bq is semisimple, too. Hence,
eiT is poor.
Now assume that T satisfies (2). By assumption, T has no simple projective module. Then we
have a decomposition T = f1T ⊕ . . . ⊕ fmT , where fiT are nonsimple local modules. Let fiT /fiJ be
a noninjective module for some i. Assume that fiT /fiJ is M -- injective for a cyclic module M . We
can write M = A1 ⊕ . . . ⊕ Ap ⊕ B1 ⊕ . . . ⊕ Bq, where Ak and Bt are indecomposable modules such
that the Ak's do not contain an isomorphic copy of fiT /fiJ and the Bt's contain an isomorphic copy
of fiT /fiJ. Because T has a unique noninjective simple module up to isomorphism, A1 ⊕ . . . ⊕ Ap
is semisimple. B1 ⊕ . . . ⊕ Bq is also semisimple since fiT /fiJ is Bt -- injective for each t = 1,··· , q.
Hence, T has no simple middle class. Now, the theorem follows from Lemma 2.4.
Following [1], we call a ring R simple -- destitute if every simple right R -- module is poor. Notice
that, just as V -- rings, simple -- destitute rings also constitute a natural subclass of rings with no simple
middle class. In [1, Theorem 5.2], it is proved that if a right Artinian ring R has only one simple
module up to isomorphism, then R is simple -- destitute. Now we establish the converse of this theorem
as follows.
Corollary 3.9 Assume that R is a right Artinian ring. R is simple -- destitute if and only if either R
is semisimple or R has a unique simple module up to isomorphism.
Proof. (⇐) It follows from [1, Theorem 5.2].
It follows from [1,
Theorem 5.3] that Soc(RR) is singular. Then R is an indecomposable ring by Proposition 3.6. Since
R is neither a right V -- ring nor a right SI -- ring, we get the desired result by Theorem 3.8.
(⇒) If R is semisimple, then we are done. Suppose R is not semisimple.
We see, in [1] and [3], that the ring S =(cid:18) F 0
F F (cid:19) ,where F is a field, is of a particular interest.
In [1], it is shown that S has no simple middle class. In [3], Er et al. proved that S has, indeed, no
right middle class. It is also proved, in [3], that a QF -- ring R with J(R)2 = 0 and homogeneous right
socle has no right middle class. In the following theorem, we give a more general result by replacing
QF with Artinian serial. Note that the class of Artinian serial rings contains that of both QF -- rings
of above type and rings in the form of S.
Theorem 3.10 If R is an Artinian serial ring with J(R)2 = 0 and homogeneous right socle, then R
has no (simple) middle class.
Proof. Since R is Artinian serial, we can write R = ⊕n
i=1eiR, where ei's are local idempotents and
eiR's are uniserial. Suppose ekR is not simple for some k = 1, . . . , n. Since ekJ(R) is the unique
maximal submodule of ekR, ekJ(R) = l.annekR(J(R)) = Soc(ekR). Moreover, ekR is an injective
R -- module by [2, 13.5, p.124]. It follows that, for each i = 1, . . . , n, eiR is either a simple module
or an injective local module of length two. Now let etR and et′R be nonsimple. By homogeneity of
the right socle, we have Soc(etR) ∼= Soc(et′ R). Then the injectivity of et′ R yields an isomorphism
between etR and et′R. Thus the nonsimple eiR's are all isomorphic to each other.
Now let M be a (simple) module. Assume M is N -- injective, where N is cyclic. Since R is an
Artinian serial ring, by [4, Theorem 5.6], N = ⊕r
k=1Nk, where Nk's are cyclic uniserial. If N is not
semisimple, then there exists t such that Nt is not simple. Since Nt is cyclic and local, Nt ∼= ejR for
some j = 1, . . . , n. This gives that M is ejR -- injective. Also, M is injective relative to any eiR which
is simple. It follows that M is R -- injective, i.e., it is injective. This completes the proof.
Corollary 3.11 Let R be an indecomposable Artinian serial ring. Then R has no right middle class
if and only if J(R)2 = 0 and R has homogeneous right socle.
Theorem 3.10 shows that, for a nonsemisimple Artinian serial ring R with J(R)2 = 0 and homo-
geneous right socle, R has no right middle class if and only if R has no simple middle class. However,
one can find an Artinian serial ring with homogeneous right socle and no simple middle class which
has right middle class, as the following example illustrates.
Example 3.12 Let R = Z/8Z. Then R is an Artinian chain ring with no simple middle class by
Corollary 3.9. However, R has right middle class since J 2(R) 6= 0.
13
4 Commutative Rings
In this section, we focus on commutative rings and investigate the property of having no (simple)
middle class. We see that commutative rings with no middle class are precisely those Artinian rings
which decompose into a sum of a semisimple ring and a ring of composition length two. We start with
the following lemma.
Lemma 4.1 [5, Exercise 17, Ch. 1, Sec. B] Let R be a commutative Noetherian ring, and let P, Q
be prime ideals of R. Then P ⊆ Q if and only if HomR(E(R/P ), E(R/Q)) 6= 0.
Proposition 4.2 If R is a commutative Noetherian ring with no middle class, then R is Artinian.
Proof. Let R be a commutative Noetherian ring with no middle class. We shall complete the proof
by showing that Krull dimension of R is zero, i.e., every prime ideal of R is maximal. If R is a V -- ring,
then there is nothing to prove. So, assume that R is not a V -- ring. Then there exists a maximal ideal
P of R such that R/P is not injective. Hence E(R/P ) is not semisimple. Let Q be any prime ideal
of R such that Q 6= P. Then by above lemma, HomR(E(R/P ), E(R/Q)) = 0, and so R/Q is injective
relative to E(R/P ). Since R has no middle class and E(R/P ) is nonsemisimple, R/Q is injective.
Thus R/Q is a self -- injective domain, which implies that Q is a maximal ideal. This completes the
proof.
Theorem 4.3 A commutative ring R has no middle class if and only if there is a ring decomposition
R = S ⊕ T, where S is a semisimple Artinian ring, and T is zero or a local ring whose maximal ideal
is minimal.
Proof. Suppose first that R has no middle class. Then, by [3, Theorem 2], there is a ring decompo-
sition R = S ⊕ T where S is semisimple Artinian and T is either zero or fits in one of the following
three cases:
Case I : T is Morita equivalent to a right PCI domain T ′. In this case, since T ′ is right Noetherian,
so is T . Then by Proposition 4.2, T is Artinian. Thus T ′ is an Artinian domain, and hence a simple
ring. This gives that T is also a simple ring. Since T is commutative, it is a field.
Case II : T is an indecomposable SI -- ring which is either Artinian or a V -- ring. Assume first that
T is Artinian. Then T is a finite product of local rings. Thus indecomposability gives that T is a
commutative local Artinian ring. Suppose that T is not a field. Then there is a minimal nonzero ideal
A of T. Notice that A ∼= T /J(T ). But since T is an SI -- ring and T /J(T ) is singular as a T -- module, A is
injective. Then A is a direct summand of T, which contradicts the indecomposability of T. Therefore
T is a field. Now let T be a V -- ring. We may assume that T is not Noetherian. Then by [3, Lemma 5],
T is semiartinian. This gives that soc(T ) 6= 0. Let A be a nonzero minimal ideal of T. Since J(T ) = 0,
there exists a maximal ideal M which does not contain A. Then A ⊕ M = T. It follows that A = T
and hence T is a field.
Case III : T is an indecomposable Artinian ring with soc(T ) = J(T ). Note that, just as with Case
II, R is a local ring. It, therefore, follows from [9, Corollary 2.14] that T is a ring whose maximal
ideal J(T ) is minimal.
Conversely, if T is a commutative local ring whose maximal ideal is simple, then, clearly, T
has a unique (up to isomorphism) local module of length two (which is, indeed, T itself), and has
homogeneous soc(T ) = J(T ). Thus, by [3, Proposition 7], T has no middle class. Now the result
follows from Lemma 2.4.
We give the following immediate consequences of the above theorem.
Corollary 4.4 Any commutative ring with no middle class is Artinian.
Corollary 4.5 A commutative ring R is a local ring whose (unique) maximal ideal is minimal if and
only if
(i) R is indecomposable Artinian,
(ii) soc(R) = J(R), and
(iii) R has no middle class.
Now we turn our attention to commutative Noetherian rings with no simple middle class although
they need not be Artinian as the following lemma shows.
14
Lemma 4.6 A commutative local ring (not necessarily Noetherian) has no simple middle class.
Proof. Let R be a commutative local ring with the unique maximal ideal M. Let R/M be (R/I) --
injective for some proper ideal I of R. Then R/I is a V -- ring since its unique simple module is injective.
This gives that I = M, and so R/M is a poor module. This completes the proof.
Theorem 4.7 Let R be a commutative Noetherian ring. Then R has no simple middle class if and
only if there is a ring decomposition R = S ⊕ T where S is semisimple Artinian and T is a local ring.
Proof. The sufficiency follows easily from Lemma 2.4 together with the above lemma. For the
necessity, let R be a commutative Noetherian ring with no simple middle class. Suppose R is not
semisimple Artinian. Then R is not a V -- ring, and so there exists a maximal ideal M of R such that
R/M is a poor R -- module. Let P be a prime ideal of R with P * M. Then R/M is E(R/P) -- injective
by Lemma 4.1. Then E(R/P) is semisimple, i.e., E(R/P) = R/P and P is a maximal ideal of R.
Since R/P is injective, by [8, Theorem 3.71], RP is a field. This, in particular, gives that P contains
no prime ideals properly, and that Pk = P for every positive integer k. Since R is Noetherian, there
exist minimal prime ideals P1, . . . , Pn of R such that Pt1
n = 0. If R is local, then we are
done. So, suppose R is not local. If Pi is contained in M for every i = 1, . . . , n, then for any maximal
ideal P of R, Pt1
n = 0 ⊆ P which yields Pj ⊆ P for some j. Then we must have, by
above arguments, Pj = P = M, a contradiction. Thus we may arrange the Pi's in such a way that
P1, . . . , Ps are not contained in M but Ps+1, . . . , Pn are, for some s < n. It follows that P1, . . . , Ps, M
is the complete list of all maximal ideals of R, and that P1 ∩ . . .∩ Ps ∩ P
n = 0. It is also
easy to see that (P1 ∩ . . .∩ Ps)⊕ (P
n ) = R. Notice that P
n is a semisimple
Artinian ring isomorphic to R/(P1 ∩ . . . ∩ Ps) whereas P1 ∩ . . . ∩ Ps is a local ring isomorphic to
R/(Pts+1
s+1 ∩ . . .∩ Ptn
n ). This completes the proof.
1 ∩ . . . ∩ Ptn
1 ∩ . . . ∩ Ptn
s+1 ∩ . . . ∩ Ptn
ts+1
ts+1
ts+1
s+1 ∩ . . .∩ Ptn
s+1 ∩ . . .∩ Ptn
Acknowledgement: The authors would like to express their gratitude to Professor Sergio R.
L´opez-Permouth and the anonymous referee for their invaluable comments and suggestions which
improved the presentation of this work.
References
[1] A. N. Alahmadi, M. Alkan, S. R. L´opez -- Permouth, Poor Modules: The opposite of injectivity,
Glasgow Math. J. 52A (2010) 7 -- 17.
[2] N. V. Dung, D. V. Huynh, P. F. Smith, R. Wisbauer, Extending modules. With the collaboration
of J. Clark and N. Vanaja. Pitman Research Notes in Mathematics Series, 313. Longman Scientific
& Technical, Harlow; copublished in the United States with John Wiley & Sons, Inc., New York,
1994.
[3] N. Er, S. R. L´opez -- Permouth, N. Sokmez, Rings whose modules have maximal or minimal injec-
tivity domains, J. Algebra 330(1) (2011) 404 -- 417.
[4] A. Facchini, Module theory. Endomorphism rings and direct sum decompositions in some classes
of modules. Progress in Mathematics, 167. Birkhauser Verlag, Basel, 1998.
[5] K. R. Goodearl, Ring theory. Nonsingular rings and modules. Pure and Applied Mathematics,
No. 33. Marcel Dekker, Inc., New York-Basel, 1976.
[6] C. Holston, S. R. L´opez -- Permouth, N. Orhan Erta¸s, Rings whose modules have maximal or
minimal projectivity domain, J. Pure Appl. Algebra 216(3) (2012) 480 -- 494.
[7] T. Y. Lam, A first course in noncommutative rings. Graduate Texts in Mathematics, 131.
Springer-Verlag, New York, 1991.
[8] T. Y. Lam, Lectures on Modules and Rings. New York: Springer -- Verlag, 1998
[9] S. R. L´opez-Permouth, J. E. Simental, Characterizing rings in terms of the extent of the injectivity
and projectivity of their modules, J. Algebra 362 (2012) 56 -- 69.
[10] V.S. Ramamurthi, K.M Rangaswamy,Generalized V-Rings, Math. Scand. 31 (1972) 69 -- 77.
15
|
1210.0958 | 1 | 1210 | 2012-10-03T02:03:03 | Spectrum nonincreasing maps on matrices | [
"math.RA"
] | Maps $\Phi$ which do not increase the spectrum on complex matrices in a sense that $\Sp(\Phi(A)-\Phi(B))\subseteq\Sp(A-B)$ are classified. | math.RA | math |
SPECTRUM NONINCREASING MAPS ON MATRICES
GREGOR DOLINAR, JINCHUAN HOU, BOJAN KUZMA, AND XIAOFEI QI
Abstract. Maps Φ which do not increase the spectrum on complex
matrices in a sense that Sp(Φ(A) − Φ(B)) ⊆ Sp(A − B) are classified.
1. Introduction
Let Mn(C) be the set of all n × n matrices over the complex field C,
and let Sp(X) be the spectrum of X ∈ Mn(C). In [9], Marcus and Moyls
proved that every linear map Φ : Mn(C) → Mn(C) preserving eigenvalues
(counting multiplicities) is either an isomorphism or an anti-isomorphism.
Furthermore, by using their result, one can show that every linear map
Φ : Mn(C) → Mn(C) preserving spectrum of matrices (that is, Sp(Φ(A)) =
Sp(A) for all A ∈ Mn(C)) also has the standard form, that is, it is an
isomorphism or an anti-isomorphism.
This result has been generalized in different directions. Instead of matrix
algebras, the algebras of all bounded linear operators on a complex Banach
space were considered, see for example [2, 3, 8, 11] and the references therein.
Also, instead of linear or additive preservers, general preservers (without lin-
earity and additivity assumption) of spectrum on Mn(C) were considered.
Baribeau and Ransford [4] proved that a spectrum preserving C1 diffeomor-
phism from an open subset of Mn(C) into Mn(C) has the standard form.
Mrcun showed in [10] that, if Φ : Mn(C) → Mn(C) is a Lipschitz map with
Φ(0) = 0 such that Sp(Φ(A) − Φ(B)) ⊆ Sp(A − B) for all A, B ∈ Mn(C)
then Φ has the standard form. Costara in [7] improved the above result by
relaxing Lipschitzian property to continuity. Recently, the continuity of the
map was replaced by surjectivity. Namely, in [5], Bendaoud, Douimi and
Sarih proved that a surjective map Φ : Mn(C) → Mn(C) satisfying Φ(0) = 0
and Sp(Φ(A) − Φ(B)) ⊆ Sp(A − B) for all A, B ∈ Mn(C) has the standard
form. We should mention here that the condition Φ(0) = 0 is a harmless
normalization: If Ψ is any map with Sp(Ψ(A) − Ψ(B)) ⊆ Sp(A − B), then
Φ(X) := Ψ(X) − Ψ(0) also satisfies this property.
2010 Mathematics Subject Classification. 15A04, 15A18.
Key words and phrases. Spectrum, Matrix algebras, General preservers.
This work was partially supported by a joint Slovene-Chinese Grant BI-CN/11-13/9-2,
National Natural Science Foundation of China (11171249, 11101250, 11201329) and a
grant from International Cooperation Program in Sciences and Technology of Shanxi
(2011081039).
1
2
GREGOR DOLINAR, JINCHUAN HOU, BOJAN KUZMA, AND XIAOFEI QI
It is our aim to prove the following generalization of [7, Theorem 1] and
[5, Theorem 1.3], in which the maps considered are neither continuous nor
surjective.
Theorem 1.1. Let n ≥ 2. Suppose that Φ : Mn(C) → Mn(C) is a map with
Φ(0) = 0 and
Sp(Φ(A) − Φ(B)) ⊆ Sp(A − B)
for all A, B ∈ Mn(C).
Then there exists an invertible matrix S ∈ Mn(C) such that either Φ(A) =
SAS−1 for all A ∈ Mn(C) or Φ(A) = SAtS−1 for all A ∈ Mn(C), where At
denotes the transpose of A.
Remark 1.2. It was shown by Costara [7] that the maps which satisfy
Sp(Φ(A) − Φ(B)) ⊇ Sp(A − B) are also linear and of a standard form.
2. Structural features of bases of matrix algebras
In this section, some features of bases of Mn(C) will be given, which are
useful for proving our main result.
Z, the ring of integers, if the only possibility that Pn
numbers z1, . . . , zn ∈ Z is z1 = z2 = · · · = zn = 0.
Recall that complex numbers α1, . . . , αn are linearly independent over
i=1 ziαi = 0 for some
Lemma 2.1. Let α1(t), . . . , αn(t) be n linearly independent analytic func-
tions in t, defined in a neighborhood of a closed unit disc ∆. Then the set
of all parameters t0 ∈ ∆ such that the complex numbers α1(t0), . . . , αn(t0)
are linearly dependent over Z is at most countable.
Proof. An analytic function either vanishes identically or it has only finitely
many zeros in compact subsets. Since α1(t), . . . , αn(t) are linearly indepen-
dent, none of them can be identically zero, hence each of them has only
finitely many zeros in ∆. Let t1, . . . , tk be all the zeros of the product
α1(t)α2(t) · · · αn(t) and let t0 ∈ ∆ \ {t1, . . . , tk}. It is straightforward that
there exist integers z1, . . . , zn ∈ Z, with zn 6= 0, such that
z1α1(t0) + · · · + znαn(t0) = 0
if and only if there exist rational numbers pi ∈ Q (in fact, pi = zi
zn
that the meromorphic functions bαi(t) := αi(t)
αn(t) satisfy
) such
p1cα1(t0) + · · · + pn−1[αn−1(t0) + 1 = 0.
By linear independency, the meromorphic function αp1,...,pn−1 : t 7→ p1cα1(t)+
· · · + pn−1[αn−1(t) + 1 is nonzero. Thus it has at most finitely many zeros on
compact subsets of ∆\{t1, . . . , tk}, and so it has at most countably many ze-
ros in ∆. Since there are countably many functions αp1,...,pn−1 as p1, . . . , pn−1
varies along rational numbers, we get at most Q × Q = Q, i.e., at most
countably many points which annihilate one among the functions αp1,...,pn−1.
Adding also {t1, . . . , tk}, we see that there are at most countably many points
SPECTRUM NONINCREASING MAPS
3
t ∈ ∆, for which the complex numbers α1(t), . . . , αn(t) are linearly depen-
dent with zn 6= 0. If zn = 0 we are seeking for linear dependence of scalars
α1(t), . . . , αn−1(t). By the same argument as before, there are at most count-
ably many such t's, under additional hypothesis that zn−1 6= 0. Proceeding
inductively backwards, there are at most countably many parameters t ∈ ∆
for which α1(t), . . . , αn(t) are linearly dependent.
(cid:3)
2 ) , ( 2
To formulate the next technical lemma, we introduce the following no-
tation: given the set Ω ⊆ C of cardinality n, let ~Ω ⊆ Cn be the set of n!
column vectors in Cn such that the set of their components, relative to a
standard basis, equals Ω. For example, if Ω = {1, 2}, then ~Ω = {( 1
1 )}.
Lemma 2.2. There exist n sets Ω1, . . . , Ωn ⊆ C, each of cardinality n, such
that n vectors x1, . . . , xn are linearly independent for any choice of xi ∈ ~Ωi,
i = 1, . . . , n.
Proof. It suffices to prove that there exist n column vectors x1, . . . , xn ∈ Cn
such that (i) each of them has pairwise distinct components in a standard
basis, and (ii) for arbitrary permutation matrices P1, . . . , Pn, the vectors
P1x1, . . . , Pnxn are linearly independent. Let x1 be an arbitrary column
vector with pairwise distinct components. Assume x1, . . . , xk, k < n, are col-
umn vectors with the properties from the lemma. Then Lin{P1x1, . . . , Pkxk}
(the linear span of vectors P1x1, . . . , Pkxk) with P1, . . . , Pk permutation ma-
trices is a k-dimensional linear plane, and there are at most finitely many
such distinct k-dimensional subspaces of Cn for all possible choices of per-
mutation matrices. Since the union of such k-dimensional subspaces is a
closed proper subset of the whole space Cn, there exists a vector xk+1 ∈ Cn
which does not belong to any of these subspaces and has pairwise distinct
components. Let P1, . . . , Pk, Pk+1 be arbitrary permutation matrices.
If
Pk+1xk+1 ∈ Lin{P1x1, . . . , Pkxk}, then xk+1 ∈ P −1
k+1 Lin{P1x1, . . . , Pkxk} =
{P −1
k+1Pi, i = 1, . . . , k, are again per-
mutation matrices, we obtain a contradiction. Thus P1x1, . . . , Pk+1xk+1 are
linearly independent. By induction on n, the lemma is true.
(cid:3)
k+1Pkxk}, and since P −1
k+1P1x1, . . . , P −1
Let K denotes the set of all matrices in Mn(C) which have n pairwise
distinct eigenvalues, linearly independent over Z. For example, since π =
is not algebraic, the diagonal matrix diag(1, π, π2, . . . , πn−1) ∈ K.
3·14 . . .
We will use the following properties of the set K.
Lemma 2.3. Let B1, . . . , Bm ∈ K be a finite sequence of matrices in Mn(C).
Then, the set of matrices X ∈ K such that Bk −X ∈ K for each k = 1, . . . , m
is dense in Mn(C).
Proof. Choose any A ∈ Mn(C) and let ε > 0. We will show that there exists
X ∈ K such that kA − Xk < ε and Bk − X ∈ K for each k = 1, . . . , m.
Let B0 = 0. It is easy to see that the set Dn of matrices with n distinct
eigenvalues is dense and open in Mn(C). Hence we may find A0 ∈ Dn
arbitrarily close to A. Since Dn is open, each neighborhood of A0, which
4
GREGOR DOLINAR, JINCHUAN HOU, BOJAN KUZMA, AND XIAOFEI QI
is small enough, contains only matrices from Dn. So we may find A1 ∈ Dn
arbitrarily close to A0 such that B1 −A1 ∈ Dn. Proceeding recursively, there
exists A ∈ Mn(C) such that kA − Ak < ε and that, moreover, A − Bk ∈ Dn
for each k = 0, . . . , m. Without loss of generality we write in the sequel A
instead of A, that is, we assume that A, A − Bk are all in Dn, k = 1, . . . , m.
Note that, for each fixed k, 0 ≤ k ≤ m, by Lemma 2.2, there exist n
matrices Ck,1, . . . , Ck,n such that the spectral sets Sp(Bk − Ck,i) have the
properties stated in Lemma 2.2. Then, for such Ck,i, it is easily checked that
there exists a polynomial A(x) such that A(0) = A and A((m+n)k+i) = Ck,i
for each (k, i) ∈ M := {0, . . . , m} × {1, . . . , n}.
We claim that, for each fixed k, 0 ≤ k ≤ m, Sp(Bk − A(x)) consists of n
linearly independent functions which are analytic in a neighborhood of x =
0. In fact, the spectrum of analytic perturbation Bk − A(x) of Bk consists of
functions which are locally analytic outside a closed discrete set of branching
points (see, e.g., [1, Theorem 3.4.25]). Let b1, . . . , br be all the real branching
points of modulus smaller or equal to (m+n)m+n. Note that for j ∈ {0, (m+
n)k + i : (k, i) ∈ M}, A(j) = A if j = 0, and A(j) = Ck,i otherwise, and so
# Sp(Bk − A(j)) = n. Hence none of bi equals some j ∈ {0, (m + n)k + i :
(k, i) ∈ M}. Choose a piecewise linear path α : [0, (m + n)m + n] → C
which avoids and does not encircle branching points and passes through j =
(m + n)k + i, (k, i) ∈ M. Then the spectral points λk,1(α(s)), . . . , λk,n(α(s))
of Sp(Bk − A(α(s))) are continuous functions of s ∈ [0, n + m(m + n)]. By
the construction of Ck,i, for any choice of x1 ∈ ~Sp(Bk − Ck,1), . . . , xn ∈
~Sp(Bk − Ck,n), the vectors x1, . . . , xn are linearly independent. Then it
easily follows that the n functions λk,1(α(s)), . . . , λk,n(α(s)) are also linearly
independent for each fixed k. Hence, λk,1(x), . . . , λk,n(x) of Sp(Bk − A(x))
are linearly independent analytic functions in a neighborhood of the curve
given by path α. Moreover, since # Sp(A(0)) = # Sp(A) = n, x = 0 is
not a branching point. Thus, also the restrictions of λk,1(x), . . . , λk,n(x)
to a neighborhood of x = 0 are linearly independent and distinct analytic
functions since linear independence is checked by nonvanishing of analytic
function, i.e. Wronskian.
Finally, by Lemma 2.1, Bk − A(x) ∈ K for every k ∈ {0, . . . , m} and each
x outside a countable subset of C. Hence, there exists x arbitrarily close
to 0 such that Bk − A(x) ∈ K for every k ∈ {0, . . . , m}. Since A(0) = A,
we can find x such that X = A(x) is close to original matrix A, and that
Bk − X ∈ K for each k. In particular, B0 = 0 implies that also X ∈ K. The
proof is complete.
(cid:3)
In particular, Lemma 2.3 implies that the set K is dense in Mn(C) and
hence it contains a basis of Mn(C).
The following proposition is the main result of this section.
It gives a
interesting structural feature of basis of Mn(C), and is crucial for our proof
of Theorem 1.1.
SPECTRUM NONINCREASING MAPS
5
Proposition 2.4. If B1, . . . , Bn2 ∈ K is a basis in Mn(C), then there exists
a basis C1, . . . , Cn2 in Mn(C) such that Ci ∈ K and Bi − Cj ∈ K for every
i, j ∈ {1, . . . , n2}.
Proof. The matrices C1, . . . , Cn2 form a basis if and only if the n2×n2 matrix
of their coefficients with respect to the standard basis Eij of Mn(C) (ordered
lexicographically) is invertible. Hence, starting with the basis Eij (whose
matrix of coefficients is the n2 × n2 identity matrix), the small perturbation
of Eij is again a basis for Mn(C). We now use the fact that the set
K′ = {C ∈ K : C − Bi ∈ K, i = 1, 2, . . . , n2}
is dense in Mn(C) (Lemma 2.3). By the density of K′, we can find matrices
C11, C12, . . . , Cnn ∈ K with Cij arbitrarily close to Eij, such that Cij −
Bk ∈ K for each i, j, k. Since they are close to basis Eij, the matrices
C11, C12, . . . , Cnn are again a basis for Mn(C).
(cid:3)
3. Proof of Theorem 1.1
In this section we give a proof of our main result. Throughout we al-
ways assume that Φ : Mn(C) → Mn(C) is a map satisfying Φ(0) = 0 and
Sp(Φ(A) − Φ(B)) ⊆ Sp(A − B) for all A, B ∈ Mn(C). First let us state four
facts which were already used in a paper by Costara [7]. As usual, Tr(X)
denotes the trace of a matrix X.
Lemma 3.1. For every X ∈ Mn(C), we have Tr(Φ(X)) = Tr(X).
Proof. The proof is the same as the proof of Equation (8) in [7, pp. 2675 --
2676]. We omit the details.
(cid:3)
Lemma 3.2. If A − B ∈ K, then Sp(Φ(A) − Φ(B)) = Sp(A − B), counted
with multiplicities. In particular, A ∈ K implies Sp Φ(A) = Sp A.
Proof. The first claim follows by [7, Lemma 5], Lemma 3.1 and the linearity
of trace. The last claim follows by inserting B = 0.
(cid:3)
For any X ∈ Mn(C), let S2(X) be the second symmetric function in eigen-
values of a matrix X (i.e., the coefficient of xn−2 in characteristic polynomial
p(x) = det(xI − X)).
Lemma 3.3. For every X ∈ Mn(C), we have (Tr X)2 = Tr(X 2) + 2S2(X).
Proof. A straightforward calculation. Also see [7, Eq.(10)].
(cid:3)
Lemma 3.4. If A, B, (A − B) ∈ K, then Tr(AB) = Tr(Φ(A)Φ(B)).
Proof. For any A, B ∈ K such that A − B ∈ K, by Lemma 3.3, we have
Tr(cid:0)(A − B)2(cid:1) = (cid:0)Tr(A − B)(cid:1)2 − 2S2(A − B).
Since A − B ∈ K, Lemma 3.2 implies S2(A − B) = S2(Φ(A) − Φ(B)) and
Tr(A − B) = Tr(Φ(A) − Φ(B)). It follows that
(1)
Tr(cid:0)(A − B)2(cid:1) = (cid:0)Tr(Φ(A) − Φ(B))2(cid:1).
6
GREGOR DOLINAR, JINCHUAN HOU, BOJAN KUZMA, AND XIAOFEI QI
Moreover, note that A, B ∈ K. Likewise, A, B ∈ K implies Tr(A2) =
Tr(Φ(A)2) and Tr(B2) = Tr(Φ(B)2). Hence, linearizing (1) gives
(2)
Tr(AB) = Tr(Φ(A)Φ(B))
whenever A, B, A − B ∈ K.
(cid:3)
Proof of Theorem 1.1. Given a matrix X = (xij) ∈ Mn, let us introduce its
row vector RX := (x11, x12, . . . , x1n, x21, . . . , x2n, . . . , xnn) and its column
vector CX := (cid:0)x11, x21, . . . , xn1, x12, . . . , xn2, . . . , xnn(cid:1)t
. It is elementary that
Tr(XY ) = RX CY , and hence we may rewrite (2) into
(3)
RΦ(A)CΦ(B) = RACB
whenever A, B, A − B ∈ K (see also Chan, Li, and Sze [6]).
Now, for any A ∈ K, by Proposition 2.4, we can find a basis B1, . . . , Bn2 ∈
K such that A − Bi ∈ K. Using Bi in place of B in (3), we obtain a system
of n2 linear equations
RΦ(A)CΦ(Bi) = RACBi ,
i = 1, 2, . . . , n2.
Introducing two n2 × n2 matrices U := [CΦ(B1)CΦ(B2) . . . CΦ(Bn2 )] and C :=
[CB1 . . . CBn2 ], this system can be rewritten into
(4)
RΦ(A)U = RAC.
The identity holds for each A ∈ K satisfying A − Bi ∈ K. By Proposition
2.4 again, there exists another basis A1, . . . , An2 ∈ K such that Bi − Aj ∈ K
for every i, j ∈ {1, 2, . . . , n2}. Using Aj in place of A in (4), the identity (4)
can be rewritten into a matrix equation
VU = RC,
where R is an n2 ×n2 matrix with j-th row equal to RAj , and V is an n2 ×n2
matrix with j-th row RΦ(Aj). Since A1, . . . , An2 is a basis, the matrix R is
invertible. Likewise, since B1, . . . , Bn2 is a basis, C is invertible. This implies
invertibility of U . In particular, (4) yields
RΦ(A) = RACU −1 = RAW;
(W = CU −1)
for all A ∈ K with A − Bi ∈ K.
Set K′ = {A ∈ K : A − Bi ∈ K, i = 1, 2, . . . , n2}. By Lemma 2.3, the
set K′ is dense in Mn(C). Therefore we get
RΦ(X) = RXW,
∀ X ∈ K′.
Recall that W = CU −1 is an invertible n2 × n2 matrix. Now, define a linear
bijection Ψ : Mn(C) → Mn(C) by
RΨ(X) = RX W,
∀ X ∈ Mn(C).
The map Ψ coincides with Φ on a dense subset K′. Moreover, Lemma 3.2 im-
plies that Sp(Φ(K)) = Sp(K) for every K ∈ K′. Hence we have Sp(Ψ(K)) =
Sp(K) for K ∈ K′. It follows from the continuity of Ψ and of the spectral
function Sp : Mn(C) → C that Ψ is a linear bijection satisfying Sp(Ψ(X)) =
SPECTRUM NONINCREASING MAPS
7
Sp(X) for every X ∈ Mn(C). Then, by Marcus-Moyls [9], there exists an
invertible S ∈ Mn(C) such that either Ψ(X) = SXS−1 for all X ∈ Mn(C)
or Ψ(X) = SX tS−1 for all X ∈ Mn(C). This gives that either
or
Φ(K) = SKS−1
for all K ∈ K′,
Φ(K) = SK tS−1
for all K ∈ K′.
Clearly, neither the hypothesis nor the end result changes if we replace Φ
by the map X 7→ S−1Φ(X)S or by the map X 7→ (S−1Φ(X)S)t. So, with
no loss of generality, we can assume that
Φ(K) = K for all K ∈ K′.
We assert that Φ(X) = X for every X ∈ Mn(C). To show this, write
Y := Φ(X). By the assumption on Φ, for any K ∈ K′, we have
Sp(K − Y ) = Sp(Φ(K) − Φ(X)) ⊆ Sp(K − X).
Since K′ is dense in Mn(C) and as spectral function is continuous, we derive
Sp(A − Y ) ⊆ Sp(A − X)
for all A ∈ Mn(C).
For any A ∈ Mn(C), let B = A − Y . Then the above relation yields
(5)
Sp(B) ⊆ Sp(B + (Y − X))
for all B ∈ Mn(C).
We will show that Y = X. Assume on the contrary that Y − X 6= 0. In
addition, for the sake of convenience, we may also assume that Y − X is
already in its Jordan form. If Y −X is not a nilpotent matrix, then inserting
B = −(Y −X) in (5) yields a contradiction {0} 6= Sp(Y −X) ⊆ Sp(0) = {0}.
If Y −X = Jn1 ⊕· · ·⊕Jnk is a nonzero nilpotent matrix in its Jordan form, let
)t in (5), where we tacitly assume that A0 = I for
B = (J n1−1
any A ∈ Mn(C) including the zero matrix. Since Y − X 6= 0, there exists at
least one block of size ≥ 2. Clearly then, B is not invertible, but (Y −X)+B
is of full rank. This contradicts the fact that Sp(B) ⊆ Sp(B +Y −X). Hence
Y = X, that is, Φ(X) = X for each X ∈ Mn(C).
(cid:3)
n1 ⊕ · · · ⊕ J nk−1
nk
References
[1] B. Aupetit, A Primer on Spectral Theory, Springer-Verlag, New York, 1991.
[2] B. Aupetit, Spectrum-preserving linear mappings between Banach algebras or Jordan-
Banach algebras, J. London Math. Soc.(2) 62 (2000), 917 -- 924.
[3] B. Aupetit, H. du T. Mouton, Spectrum preserving linear mappings in Banach alge-
bras, Studia math. 109 (1994), 91 -- 100.
[4] L. Baribeau, T. Ransford, Non-linear spectrum-preserving maps, Bull. London Math.
Soc. 32 (2000), no. 1, 8 -- 14.
[5] M. Bendaoud, M. Douimi, M. Sarih, Maps on matrices preserving local spectra, Linear
and Multilinear Algebra (2012), DOI:10.1080/03081087.2012.716429.
[6] J.T. Chan, C.K. Li, N.S. Sze, Mappings preserving spectra of products of matrices,
Proc. Amer. Math. Soc. 135 (2007) 977 -- 986.
[7] C. Costara, Maps on matrices that preserve the spectrum, Linear Algebra Appl. 435
(2011), 2674 -- 2680.
8
GREGOR DOLINAR, JINCHUAN HOU, BOJAN KUZMA, AND XIAOFEI QI
[8] A.A. Jafarian, A.R. Sourour, Spectrum-preserving linear maps, J. Funct. Anal. 66
(1986), 255 -- 261.
[9] M. Marcus, B.N. Moyls, Linear transformations on algebras of matrices, Canad. J.
Math. 11 (1959), 61 -- 66.
[10] J. Mrcun, Lipschitz spectrum preserving mappings on algebras of matrices, Linear
Algebra Appl. 215 (1995), 113 -- 120.
[11] M. Omladic, P. Semrl, Spectrum-preserving additive maps, Linear Algebra Appl. 153
(1991), 67 -- 72.
(Gregor Dolinar) Faculty of Electrical Engineering, University of Ljubl-
jana, Trzaska cesta 25, SI-1000 Ljubljana, Slovenia
E-mail address: [email protected]
(Jinchuan Hou) School of Mathematics, Taiyuan University of Technology,
Taiyuan 030024, P. R. China
E-mail address: [email protected]
(Bojan Kuzma) 1University of Primorska, FAMNIT, Glagoljaska 8, SI-6000
Koper, and 2Institute of Mathematics, Physics and Mechanics, Department of
Mathematics, Jadranska 19, SI-1000 Ljubljana, Slovenia.
E-mail address: [email protected]
(Xiaofei Qi) School of Mathematical science, Shanxi University, Taiyuan
030006, P. R. China
E-mail address: [email protected]
|
1404.6263 | 2 | 1404 | 2015-04-01T19:01:01 | Effect algebras are the Eilenberg-Moore category for the Kalmbach monad | [
"math.RA"
] | The Kalmbach monad is the monad that arises from the free-forgetful adjunction between bounded posets and orthomodular posets. We prove that the category of effect algebras is isomorphic to the Eilenberg-Moore category for the Kalmbach monad. | math.RA | math |
EFFECT ALGEBRAS ARE THE EILENBERG-MOORE
CATEGORY FOR THE KALMBACH MONAD
GEJZA JEN CA
Abstract. The Kalmbach monad is the monad that arises from the free-
forgetful adjunction between bounded posets and orthomodular posets. We
prove that the category of effect algebras is isomorphic to the Eilenberg-Moore
category for the Kalmbach monad.
In [6], Kalmbach proved the following theorem.
1. Introduction
Theorem 1. Every bounded lattice L can be embedded into an orthomodular lattice
K(L).
The proof of the theorem is constructive, K(L) is known under the name Kalm-
bach extension or Kalmbach embedding.
In [10], Mayet and Navara proved that
Theorem 1 can be generalized: every bounded poset P can be embedded in an
orthomodular poset K(P ). In fact, as proved by Harding in [5], this K is then left
adjoint to the forgetful functor from orthomodular posets to bounded posets. This
adjunction gives rise to a monad on the category of bounded posets, which we call
the Kalmbach monad.
For every monad (T, η, µ) on a category C, there is a standard notion Eilenberg-
Moore category CT (sometimes called the category of algebras or the category of
modules for T ). The category CT comes equipped with a canonical adjunction
between C and CT and this adjunction gives rise to the original monad T on C.
In the present paper we prove that the Eilenberg-Moore category for the Kalm-
bach monad is isomorphic to the category of effect algebras.
2. Preliminaries
We assume familiarity with basics of category theory, see [9, 1] for reference.
2.1. Bounded posets. A bounded poset is a structure (P, ≤, 0, 1) such that ≤ is a
partial order on P , 0, 1 ∈ P are the bottom and top elements of (P, ≤), respectively.
Let P1, P2 be bounded posets. A map φ : P1 → P2 is a morphism of bounded
posets if and only if it satisfies the following conditions.
• φ(1) = 1 and φ(0) = 0.
• φ is isotone.
The category of bounded posets is denoted by BPos.
1991 Mathematics Subject Classification. Primary: 03G12, Secondary: 06F20, 81P10.
Key words and phrases. effect algebra, Kalmbach extension, orthomodular poset, monad.
This research is supported by grants VEGA G-1/0297/11,G-2/0059/12 of MS SR, Slovakia
and by the Slovak Research and Development Agency under the contracts APVV-0073-10, APVV-
0178-11.
1
2
GEJZA JEN CA
2.2. Effect algebras. An effect algebra is a partial algebra (E; ⊕, 0, 1) with a bi-
nary partial operation ⊕ and two nullary operations 0, 1 satisfying the following
conditions.
(E1) If a ⊕ b is defined, then b ⊕ a is defined and a ⊕ b = b ⊕ a.
(E2) If a ⊕ b and (a ⊕ b) ⊕ c are defined, then b ⊕ c and a ⊕ (b ⊕ c) are defined
and (a ⊕ b) ⊕ c = a ⊕ (b ⊕ c).
(E3) For every a ∈ E there is a unique a′ ∈ E such that a ⊕ a′ exists and
a ⊕ a′ = 1.
(E4) If a ⊕ 1 is defined, then a = 0.
Effect algebras were introduced by Foulis and Bennett in their paper [3].
In an effect algebra E, we write a ≤ b if and only if there is c ∈ E such that
a ⊕ c = b. It is easy to check that for every effect algebra E, ≤ is a partial order on
E. Moreover, it is possible to introduce a new partial operation ⊖; b ⊖ a is defined
if and only if a ≤ b and then a ⊕ (b ⊖ a) = b. It can be proved that, in an effect
algebra, a ⊕ b is defined if and only if a ≤ b′ if and only if b ≤ a′. In an effect
algebra, we write a ⊥ b if and only if a ⊕ b exists.
Let E1, E2 be effect algebras. A map φ : E1 → E2 is called a morphism of effect
algebras if and only if it satisfies the following conditions.
• φ(1) = 1.
• If a ⊥ b, then φ(a) ⊥ φ(b) and φ(a ⊕ b) = φ(a) ⊕ φ(b).
The category of effect algebras is denoted by EA. There is an evident forgetful
functor U : EA → BPos.
2.3. D-posets. In their paper [8], Chovanec and Kopka introduced a structure
called D-poset. Their definition is an abstract algebraic version the D-poset of fuzzy
sets, introduced by Kopka in the paper [7].
A D-poset is a system (P ; ≤, ⊖, 0, 1) consisting of a partially ordered set P
bounded by 0 and 1 with a partial binary operation ⊖ satisfying the following
conditions.
(D1) b ⊖ a is defined if and only if a ≤ b.
(D2) If a ≤ b, then b ⊖ a ≤ b and b ⊖ (b ⊖ a) = a.
(D3) If a ≤ b ≤ c, then c ⊖ b ≤ c ⊖ a and (c ⊖ a) ⊖ (c ⊖ b) = b ⊖ a.
Let D1, D2 be D-posets. A map φ : D1 → D2 is called a morphism of D-posets
if and only if it satisfies the following conditions.
• φ(1) = 1.
• If a ≤ b, then φ(a) ≤ φ(b) and φ(b ⊖ a) = φ(b) ⊖ φ(a).
The category of D-posets is denoted by DP.
There is a natural, one-to-one correspondence between D-posets and effect al-
gebras. Every effect algebra satisfies the conditions (D1)-(D3). When given a
D-poset (P ; ≤, ⊖, 0, 1), one can construct an effect algebra (P ; ⊕, 0, 1): the do-
main of ⊕ is given by the rule a ⊥ b if and only if a ≤ 1 ⊖ b and we then have
a ⊕ b = 1 ⊖(cid:0)(1 ⊖ a) ⊖ b(cid:1). The resulting structure is then an effect algebra with
the same ⊖ as the original D-poset. It is easy to see that this correspondence is, in
fact, an isomorphism of categories DP and EA.
Another equivalent structure was introduced by Giuntini and Greuling in [4].
We refer to [2] for more information on effect algebras and related topics.
THE KALMBACH MONAD
3
The following lemma collects some well-known properties connecting the ⊕, ⊖
and ′ operations in effect algebras (or D-posets). Complete proofs can be found, for
example, in Chapter 1 of [2]. We shall use these facts without an explicit reference.
Lemma 2.
(a) a ≤ b′ iff b ≤ a′ iff a ⊕ b exists and then
(a ⊕ b)′ = a′ ⊖ b = b′ ⊖ a.
(b) a ≤ b iff a ⊕ b′ exists and then
(b ⊖ a)′ = a ⊕ b′
(c) a ≤ c ⊖ b iff b ≤ c ⊖ a iff a ⊕ b ≤ c and then
c ⊖ (a ⊕ b) = (c ⊖ a) ⊖ b = (c ⊖ b) ⊖ a
(d) a ≤ b ≤ c iff a ⊕ (c ⊖ b) exists and then
c ⊖ (b ⊖ a) = a ⊕ (c ⊖ b).
2.4. Orthomodular posets. An orthomodular poset is a structure (A, ≤,′ , 0, 1)
such that (A, ≤, 0, 1) is a bounded poset and ′ is a unary operation (called ortho-
complementation) satisfying the following conditions.
• x ≤ y implies y′ ≤ x′.
• x′′ = x.
• x ∧ x′ = 0.
• If x ≤ y′, then x ∨ y exists.
• If x ≤ y, then x ∨ (x ∨ y′)′ = y.
If x ≤ y′, we say that x, y are orthogonal.
Let A1, A2 be orthomodular posets. A map φ : A1 → A2 is called a morphism
of orthomodular posets if and only if it satisfies the following conditions.
• φ(1) = 1.
• If a ≤ b′, then φ(a) ≤ φ(b)′ and φ(a ∨ b) = φ(a) ∨ φ(b).
Alternatively, we may define a morphism of orthomodular posets as order pre-
serving, preserving the orthocomplementation, and preserving joins of orthogonal
elements.
The category of orthomodular posets is denoted by OMP. An orthomodular
lattice is an orthomodular poset that is a lattice. We remark that the usual category
of orthomodular lattices, with morphisms preserving joins and meets is not a full
subcategory of OMP.
If A is an orthomodular poset, then we may introduce a partial ⊕ operation on
A by the following rule: x ⊕ y exists iff x ≤ y′ and then x ⊕ y := x ∨ y. The resulting
structure is then an effect algebra. This gives us the object part of an evident full
and faithful functor OMP → EA.
2.5. Kalmbach construction. If C = {x1, . . . , xn} is a finite chain in a poset P ,
we write C = [x1 < · · · < xn] to indicate the partial order.
Definition 3. [6] Let P be a bounded poset, write
K(P ) = {C : C is a finite chain in P with even number of elements}
Define a partial order on K(P ) by the following rule:
[x1 < x2 < · · · < x2n−1 < x2n] ≤ [y1 < y2 < · · · < y2k−1 < y2k]
4
GEJZA JEN CA
if for every 1 ≤ i ≤ n there is 1 ≤ j ≤ k such that
y2j−1 ≤ x2i−1 < x2i ≤ y2j.
Define a unary operation C 7→ C ⊥ on K(P ) to be the symmetric difference with
the set {0, 1}.
Originally, Kalmbach considered the construction only for lattices.
If P is a
′, 0, 1) is
bounded lattice, then K(P ) is a lattice as well. Moreover, (K(P ), ∧, ∨,
an orthomodular lattice. However, as observed by Harding in [5], K is not an
object part of a functor from the category of bounded lattices to the category of
orthomodular lattices.
On the positive side, for any bounded poset P , K(P ) is an orthomodular poset
(see [10]) and K can be made to a functor K : BPos → OMP.
Indeed, let
f : P → Q be a morphism in BPos and define K(f ) : K(P ) → K(Q) by the
following rule.
For an arrow f : P → Q in BPos, write
K(f )([x1 < x2 < · · · < x2n−1 < x2n]) =
{y ∈ Q : card({1 ≤ i ≤ n : f (xi) = y}) is odd}.
A more elegant way how to write the same rule is
K(f )([x1 < x2 < · · · < x2n−1 < x2n]) = ∆2n
i=1{f (xi)},
where ∆ is the symmetric difference of sets.
Then K is a functor. Moreover, as proved by Harding in [5], K is left-adjoint
to the forgetful functor U : OMP → BPos. Since every functor has (up to
isomorphism) at most one adjoint, this can be viewed as an alternative definition
of the Kalmbach construction.
The unit of the K ⊣ U adjunction is the natural transformation η : idBPos →
U K, given by the rule
ηP (a) =([0 < a] a > 0
a = 0
∅
and the counit of the adjunction is the natural transformation ǫ : KU → idOMP
given by the rule
ǫL([x1 < · · · < x2n]) = (x⊥
1 ∧ x2) ∨ · · · ∨ (x⊥
2n−1 ∧ x2n).
3. Kalmbach monad
Let C be a category. A monad on C can be defined as a monoid in the strict
monoidal category of endofunctors of C. Explicitly, a monad on C is a triple
(T, η, µ), where T : C → C is an endofunctor of C and η, µ are natural transfor-
mations η : idC → T , µ : T 2 → T satisfying the equations µ ◦ T µ = µ ◦ µT and
µ ◦ T η = µ ◦ ηT = 1T .
Every adjoint pair of functors F : C ⇄ D : G, with F being left adjoint, gives
rise to a monad (F G, η, GǫF ) on C.
Let (T, η, µ) be a monad on a category C. Recall, that the Eilenberg-Moore
category for (T, η, µ) is a category (denoted by CT ), such that objects (called al-
gebras for that monad) of CT are pairs (A, α), where α : T (A) → A, such that the
diagrams
(1)
(2)
THE KALMBACH MONAD
5
A
ηA /
T (A)
❉
❉
❉
❉
1A
❉
❉
❉
❉
!❉
α
A
T (α)
T 2(A)
T (A)
µA
α
T (A)
α
/ A
commute. A morphism of algebras h : (A1, α1) → (A2, α2) is a C-morphism such
that the diagram
T (A)
α1
T (h)
T (B)
α2
A
h
/ B
commutes.
Consider now the adjunction K ⊣ U between the categories, BPos and OMP
from the preceding section. This adjunction gives rise to a monad on BPos, which
we will denote (T, η, µ). Explicitly, T = U ◦ K, η remains the same and µ = U ǫK
turns out to be given by the following rule
where [C1 < C2 < · · · < C2n] is a chain of even length of chains of even length,
that means, an element of T 2(P ), and ∆ is the symmetric difference of sets.
µP(cid:0)[C1 < C2 < · · · < C2n](cid:1) = ∆2n
i=1Ci,
Theorem 4. The category of effect algebras is isomorphic to the Eilenberg-Moore
category for the Kalmbach monad.
Proof. From now on, let U be the forgetful functor U : EA → BPos. Let us define
a functor G : EA → BPosT . For an effect algebra A, define mA : T (U (A)) → U (A)
by the rule
mA(cid:0)[x1 < x2 < . . . x2n−1 < x2n](cid:1) = (x2 ⊖ x1) ⊕ · · · ⊕ (x2n ⊖ x2n−1).
We claim that G(A) = (U (A), mA) is an algebra for the Kalmbach monad. We
need to prove that the diagrams (1) and (2) commute. Clearly, for every x ∈ U (A),
(mA ◦ ηU(A))(x) = mA([0 < x]) = (x ⊖ 0) = x,
and we see that the triangle diagram (1) commutes. Consider now the square di-
agram (2): the elements of T 2(U (A)) are chains of chains of elements of U (A);
let [C1 < C2 < · · · < C2n] ∈ T 2(U (A)). Note that Ci < Cj implies that
mA(Ci) < mA(Cj ), so the elements of the sequence (mA(C1), . . . , mA(C2n)) are
pairwise distinct. Therefore,
(cid:0)mA ◦ T (mA)(cid:1)(cid:0)[C1 < C2 < · · · < C2n](cid:1) =
mA(cid:0)(cid:2)mA(C1) < mA(C2) < · · · < mA(C2n)(cid:3)(cid:1) =
(cid:0)mA(C2) ⊖ mA(C1)(cid:1) ⊕ · · · ⊕(cid:0)mA(C2n) ⊖ mA(C2n−1)(cid:1)
/
!
/
/
/
/
/
/
6
GEJZA JEN CA
Note that, if C < D in T (U (A)), then C∆D < D, mA(C) < mA(D) and
mA(D) ⊖ mA(C) = mA(C∆D). Using these facts,
(cid:0)mA ◦ µU(A)(cid:1)(cid:0)[C1 < C2 < · · · < C2n](cid:1) = mA(C1∆C2∆ . . . ∆C2n) =
mA(cid:0)(C1∆C2∆ . . . ∆C2n−1)∆C2n(cid:1) =
mA(C2n) ⊖ mA(C1∆C2∆ . . . ∆C2n−1) =
mA(C2n) ⊖(cid:0)mA(C2n−1) ⊖ mA(C1∆C2∆ . . . ∆C2n−2)(cid:1) =
(cid:0)mA(C2n) ⊖ mA(C2n−1)(cid:1) ⊕ mA(C1∆C2∆ . . . ∆C2n−2).
The desired equality now follows by a simple induction.
If f : A → B is a morphism of effect algebras, we define G(f ) = U (f ). We need
to prove that the diagram
T U(f )
T U (A)
T U (A)
mA
mB
U (A)
/ U (B)
U(f )
commutes. After some simple steps, this reduces to the following equality in B:
(3)
(cid:0)f (x2) ⊖ f (x1)(cid:1) ⊕ · · · ⊕(cid:0)f (x2n) ⊖ f (x2n−1)(cid:1) = mB(cid:0)∆n
for each [x1 < · · · < x2n] ∈ T (U (A)).
i=1{f (xi)}(cid:1),
Let us define an auxiliary function k : T (U (A)) → N: for C = [x1 < x2 < · · · <
x2n] ∈ T (U (A)), k(C) is the number of equal consecutive pairs in the sequence
(cid:0)f (x1), . . . , f (xn)(cid:1), that means, k(C) is the cardinality of the set {i : f (xi) =
To prove the equality (3), we use induction with respect to k(C). If k(C) = 0,
f (xi+1)}.
then the equality (3) clearly holds.
If k(C) > 0, then let us pick some i with f (xi) = f (xi+1). Then {f (xi)}∆{f (xi+1)} =
∅ and we may skip them on the right hand side of (3).
If i is odd, then f (xi+1) ⊖ f (xi) = 0 and we may delete that term from the
left-hand side of (3). If i is even, then
and we may simplify the left-hand side of (3) accordingly.
(cid:0)f (xi) ⊖ f (xi−1)(cid:1) ⊕(cid:0)f (xi+2) ⊖ f (xi+1)(cid:1) = f (xi+2) ⊖ f (xi−1)
So (3) is true if and only if it is true for the chain C − {xi, xi+1}. Clearly,
k(C − {xi, xi+1}) = k(C) − 1 and we have completed the induction step.
Let (A, α) be an algebra for the Kalmbach monad. Let us define a partial
operation ⊖ on the bounded poset A given by this rule: b ⊖ a is defined if and only
if a ≤ b and
b ⊖ a =(0
a = b
α([a < b]) a < b
We claim that E(A, α) = (A, ≤, ⊖, 0, 1) is then a D-poset, hence an effect algebra.
The axiom (D1) follows by definition.
Before we prove the other two axioms, let us note that for all a ∈ A, a ⊖ 0 = a.
Indeed, if 0 < a then the triangle diagram (1) implies that a = α([0 < a]) = a ⊖ 0
and for a = 0 we obtain a ⊖ 0 = 0 ⊖ 0 = 0 by definition of ⊖.
To prove (D2), let a, b ∈ A be such that a ≤ b.
/
/
/
THE KALMBACH MONAD
7
Let us prove that b ⊖ a ≤ b. If a < b, then [a < b] ≤ [0 < b] in the poset T (A)
and
If a = b then b ⊖ a = 0 ≤ b.
b ⊖ a = α(cid:0)[a < b](cid:1) ≤ α(cid:0)[0 < b](cid:1) = b ⊖ 0 = b.
Let us prove that b ⊖ (b ⊖ a) = a There are three possible cases.
(D2.1) Suppose that 0 < a < b. Then, [a < b] < [0 < b] in T (A) and hence (cid:2)[a <
b] < [0 < b](cid:3) ∈ T 2(A). Suppose that α(cid:0)[a < b](cid:1) = α(cid:0)[0 < b](cid:1). From the
commutativity of the square (2) we obtain
(cid:2)[a < b] < [0 < b](cid:3) ✤
µA
❴
T (α)
∅
❴
α
[0 < a] ✤
α
/ α(cid:0)[0 < a](cid:1) = 0
However, 0 < a = α(cid:0)[0 < a](cid:1) = 0 is false and we have proved that α(cid:0)[a <
b](cid:1) < α(cid:0)[0 < b](cid:1). Chasing the element (cid:2)[a < b] < [0 < b](cid:3) around the square
(cid:2)[a < b] < [0 < b](cid:3) ✤
(cid:2)α(cid:0)[a < b](cid:1) < α(cid:0)[0 < b](cid:1)(cid:3)
T (α)
❴
µA
❴
α
[0 < a] ✤
α
/ α(cid:0)[0 < a](cid:1) = α(cid:16)(cid:2)α(cid:0)[a < b](cid:1) < α(cid:0)[0 < b](cid:1)(cid:3)(cid:17)
gives us the equality in the bottom right corner, meaning that b ⊖ (b ⊖ a) = a.
(D2.2) Suppose that 0 = a. We already know that b ⊖ 0 = b and we may compute
b ⊖ (b ⊖ a) = b ⊖ (b ⊖ 0) = b ⊖ b = 0 = a.
(D2.3) Suppose that a = b. Then
b ⊖ (b ⊖ a) = b ⊖ 0 = b = a.
To prove (D3), let a, b, c ∈ A be such that a ≤ b ≤ c.
Let us prove that c ⊖ b ≤ c ⊖ a. If a = b, there is nothing to prove. If b = c, then
b ⊖ c = 0 ≤ c ⊖ a. Assume that a < b < c. Then [b < c] < [a < c] and
Let us prove that b ⊖ a = (c ⊖ a) ⊖ (c ⊖ b).
c ⊖ b = α(cid:0)[b < c](cid:1) ≤ α(cid:0)[a < c](cid:1) = c ⊖ a.
(D3.1) Suppose that a < b < c and assume that α([b < c]) = α([a < c]). The square
(cid:2)[b < c] < [a < c](cid:3) ✤
µA
❴
T (α)
∅
❴
α
[a < b] ✤
α
/ α(cid:0)[a < b](cid:1) = 0
gives us α(cid:0)[a < b](cid:1) = 0, so b ⊖ a = 0. However, using only the properties of ⊖
we already proved,
b = b ⊖ 0 = b ⊖ (b ⊖ a) = a < b,
/
/
/
/
/
/
/
/
/
8
GEJZA JEN CA
which is false. Thus, assuming α(cid:0)[b < c](cid:1) < α(cid:0)[a < c](cid:1) the square
(cid:2)[b < c] < [a < c](cid:3) ✤
(cid:2)α(cid:0)[b < c](cid:1) < α(cid:0)[a < c](cid:1)(cid:3)
T (α)
❴
α
❴
µA
[a < b] ✤
α
/ α(cid:0)[a < b](cid:1) = α(cid:16)(cid:2)α(cid:0)[b < c](cid:1) < α(cid:0)[a < c](cid:1)(cid:3)(cid:17)
gives us the equality in the bottom right corner meaning that
b ⊖ a = (c ⊖ a) ⊖ (c ⊖ b).
(D3.2) Suppose that b = c. Then
(c ⊖ a) ⊖ (c ⊖ b) = (c ⊖ a) ⊖ 0 = c ⊖ a = b ⊖ a.
(D3.3) If a = b, then there is nothing to prove.
If is now easy to check that an arrow h : (A, α) → (B, β) in BPosT is, at
the same time, a morphism of D-posets (and thus, a morphism of effect algebras)
E(A, α) → E(B, β). Indeed, h(0) = 0 and h(1) = 1. If a < b and h(a) < h(b), then
h(b ⊖ a) = h(cid:0)α([a < b])(cid:1) = β(cid:0)T (h)([a < b])(cid:1) = β(cid:0)[h(a) < h(b)](cid:1) = h(b) ⊖ h(a).
If a < b and h(a) = h(b) then
h(b ⊖ a) = h(α[a < b]) = β(cid:0)T (h)([a < b])(cid:1) = β(∅) = 0 = h(b) ⊖ h(a).
Therefore E is a functor from BPosT to EA.
It remains to prove that E, G are mutually inverse functors. Let A be an effect
algebra. We claim that EG(A) = A. The underlying poset of EG(A) and A is the
same. For all a < b,
b ⊖EG(A) a = mA([a < b]) = b ⊖A a,
hence EG(A) = A.
algebras EG(f ) = f , since both E and G preserve the underlying poset maps.
It is obvious that for every morphism f : A → B of effect
Let (A, α) be an algebra for the Kalmbach monad. We claim that (A, α) =
GE(A, α), that means, for all [x1 < · · · < x2k] ∈ T (A),
(4)
α([x1 < x2 < · · · < x2k−1 < x2k]) = (x2 ⊖ x1) ⊕ · · · ⊕ (x2k ⊖ x2k−1),
where the ⊕, ⊖ on the right-hand side are taken in E(A, α).
To prove (4) we need an auxiliary claim:
for every C ∈ T (A) and an upper
bound u of C with u /∈ C, α(C∆[0 < u]) = α([α(C) < u]). This is easily seen by
chasing the element [C < [0 < u]] ∈ T 2(A) around the square (2).
Clearly, equality (4) is true for k = 0. Suppose it is valid for some k = n ∈ N.
Then, for k = n + 1, equality (4) is then equivalent to
α([x1 < x2 < · · · < x2n−1 < x2n < x2n+1 < x2n+2]) =
α([x1 < x2 < · · · < x2n−1 < x2n]) ⊕ (x2n+2 ⊖ x2n+1)
/
/
/
THE KALMBACH MONAD
9
Put C = [x1 < x2 < · · · < x2n−1 < x2n]. Using the definition of ⊖ in E(A, α) and
applying the auxiliary claim twice we obtain
α(C) ⊕ (x2n+2 ⊖ x2n+1) =
x2n+2 ⊖(cid:0)x2n+1 ⊖ α(C)(cid:1) =
x2n+2 ⊖(cid:16)α(cid:0)(cid:2)α(C) < x2n+1(cid:3)(cid:1)(cid:17) =
x2n+2 ⊖(cid:0)α(cid:0)C∆[0 < x2n+1](cid:1)(cid:1) =
α(cid:16)(cid:2)α(cid:0)C∆[0 < x2n+1](cid:1) < x2n+2(cid:3)(cid:17) =
α(cid:0)(cid:2)C∆[0 < x2n+1]∆[0 < x2n+2](cid:3)(cid:1) =
α(cid:0)(cid:2)C∆[x2n+1 < x2n+2](cid:3)(cid:1) =
α(cid:0)[x1 < x2 < · · · < x2n−1 < x2n < x2n+1 < x2n+2](cid:1)
It is obvious that for every morphism f in BPosT , GE(f ) = f .
References
(cid:3)
[1] Awodey, S.: Category theory. No. 49 in Oxford Logic Guides. Oxford University Press (2006)
[2] Dvurecenskij, A., Pulmannov´a, S.: New Trends in Quantum Structures. Kluwer, Dordrecht
and Ister Science, Bratislava (2000)
[3] Foulis, D., Bennett, M.: Effect algebras and unsharp quantum logics. Found. Phys. 24, 1325 --
1346 (1994)
[4] Giuntini, R., Greuling, H.: Toward a formal language for unsharp properties. Found. Phys.
19, 931 -- 945 (1989)
[5] Harding, J.: Remarks on concrete orthomodular lattices. International Journal of Theoretical
Physics 43(10), 2149 -- 2168 (2004)
[6] Kalmbach, G.: Orthomodular lattices do not satisfy any special lattice equation. Archiv der
Mathematik 28(1), 7 -- 8 (1977)
[7] Kopka, F.: D-posets of fuzzy sets. Tatra Mt. Math. Publ. 1, 83 -- 87 (1992)
[8] Kopka, F., Chovanec, F.: D-posets. Math. Slovaca 44, 21 -- 34 (1994)
[9] Mac Lane, S.: Categories for the Working Mathematician. No. 5 in Graduate Texts in Math-
ematics. Springer-Verlag (1971)
[10] Mayet, R., Navara, M.: Classes of logics representable as kernels of measures. Contributions
to General Algebra 9, 241 -- 248 (1995)
Department of Mathematics and Descriptive Geometry, Faculty of Civil Engineer-
ing, Slovak Technical University, Radlinsk´eho 11, Bratislava 813 68, Slovak Republic
E-mail address: [email protected]
|
1506.07203 | 2 | 1506 | 2016-04-01T14:07:19 | Range-compatible homomorphisms on spaces of symmetric or alternating matrices | [
"math.RA"
] | Let $U$ and $V$ be finite-dimensional vector spaces over an arbitrary field $\mathbb{K}$, and $\mathcal{S}$ be a linear subspace of the space $\mathcal{L}(U,V)$ of all linear maps from $U$ to $V$. A map $F : \mathcal{S} \rightarrow V$ is called range-compatible when it satisfies $F(s) \in \mathrm{im}(s)$ for all $s \in \mathcal{S}$. Among the range-compatible maps are the so-called local ones, that is the maps of the form $s \mapsto s(x)$ for a fixed vector $x$ of $U$.
In recent works, we have classified the range-compatible group homomorphisms on $\mathcal{S}$ when the codimension of $\mathcal{S}$ in $\mathcal{L}(U,V)$ is small. In the present article, we study the special case when $\mathcal{S}$ is a linear subspace of the space $S_n(\mathbb{K})$ of all $n$ by $n$ symmetric matrices: we prove that if the codimension of $\mathcal{S}$ in $S_n(\mathbb{K})$ is less than or equal to $n-2$, then every range-compatible homomorphism on $\mathcal{S}$ is local provided that $\mathbb{K}$ does not have characteristic $2$. With the same assumption on the codimension of $\mathcal{S}$, we also classify the range-compatible homomorphisms on $\mathcal{S}$ when $\mathbb{K}$ has characteristic $2$. Finally, we prove that if $\mathcal{S}$ is a linear subspace of the space $A_n(\mathbb{K})$ of all $n$ by $n$ alternating matrices with entries in $\mathbb{K}$, and the codimension of $\mathcal{S}$ is less than or equal to $n-3$, then every range-compatible homomorphism on $\mathcal{S}$ is local. | math.RA | math | Range-compatible homomorphisms on spaces of
symmetric or alternating matrices
Cl´ement de Seguins Pazzis∗†
July 23, 2021
Abstract
Let U and V be finite-dimensional vector spaces over an arbitrary field
K, and S be a linear subspace of the space L(U, V ) of all linear maps from
U to V . A map F : S → V is called range-compatible when it satisfies
F (s) ∈ Im s for all s ∈ S. Among the range-compatible maps are the so-
called local ones, that is the maps of the form s 7→ s(x) for a fixed vector
x of U .
In recent works, we have classified the range-compatible group homo-
morphisms on S when the codimension of S in L(U, V ) is small. In the
present article, we study the special case when S is a linear subspace of the
space Sn(K) of all n by n symmetric matrices: we prove that if the codimen-
sion of S in Sn(K) is less than or equal to n−2, then every range-compatible
homomorphism on S is local provided that K does not have characteristic
2. With the same assumption on the codimension of S, we also classify
the range-compatible homomorphisms on S when K has characteristic 2.
Finally, we prove that if S is a linear subspace of the space An(K) of all
n by n alternating matrices with entries in K, and the codimension of S is
less than or equal to n − 3, then every range-compatible homomorphism on
S is local.
AMS Classification: 15A03, 15A30
Keywords: Range-compatible map, operator space, symmetric matrices, alter-
nating matrices.
∗Universit´e de Versailles Saint-Quentin-en-Yvelines, Laboratoire de Math´ematiques de Ver-
sailles, 45 avenue des Etats-Unis, 78035 Versailles cedex, France
†e-mail address: [email protected]
1
6
1
0
2
r
p
A
1
]
.
A
R
h
t
a
m
[
2
v
3
0
2
7
0
.
6
0
5
1
:
v
i
X
r
a
1
Introduction
1.1 Notation
Throughout the article, all the vector spaces are assumed to be finite-dimensional
and over a given arbitrary field K. Given vector spaces U and V , we denote by
L(U, V ) the space of all linear maps from U to V . Given abelian groups A and B,
we denote by Hom(A, B) the group of all homomorphisms from A to B. Group
homomorphisms will be simply called homomorphisms in this article.
Given non-negative integers n and p, we denote by Mn,p(K) the space of all
matrices with n rows, p columns and entries in K. For (i, j) ∈ [[1, n]] × [[1, p]], we
denote by Ei,j the matrix of Mn,p(K) in which all the entries equal zero except
the one at the (i, j)-spot, which equals 1. Given finite-dimensional vector spaces
U and V , equipped with respective bases B and C, and given a linear map
u : U → V , we denote by MB,C(u) the matrix of u in the bases B and C.
We set Mn(K) := Mn,n(K) and we denote by Sn(K) the subspace of Mn(K)
consisting of its symmetric matrices, and by An(K) the subspace of all n by n
alternating matrices with entries in K. Recall that a square matrix M ∈ Mn(K)
is called alternating whenever X T M X = 0 for all X ∈ Kn, which is equivalent
to having M skew-symmetric with all diagonal entries zero.
Given a square matrix M = (mi,j) ∈ Mn(K), we denote its diagonal vector
by
Given a vector X = (xi)1≤i≤n ∈ Kn and a map α : K → K, we denote by
X α := (α(xi))1≤i≤n the vector of Kn deduced from X by applying α entry-wise.
Let m, n, p, q be non-negative integers. Given respective subsets A and B of
Mm,n(K) and Mp,q(K), one sets
A ⊕ B :=(cid:26)(cid:20)A 0
0 B(cid:21) A ∈ A, B ∈ B(cid:27),
which is a subset of Mm+p,n+q(K). Note that if A and B are respective linear
subspaces of Sn(K) and Sp(K) (respectively, of An(K) and Ap(K)), then A ⊕ B is
a linear subspace of Sn+p(K) (respectively, of An+p(K)) with dimension dim A +
dim B.
2
∆(M ) :=
m1,1
...
mn,n
∈ Kn.
1.2 The problem
Let U and V be vector spaces, and S be a subset of L(U, V ). A map F : S → V
is called range-compatible whenever
∀s ∈ S, F (s) ∈ Im s.
It is called local when there exists a vector x ∈ U such that
∀s ∈ S, F (s) = s(x).
Clearly, every local map is range-compatible, but the converse is not true, even
for linear maps.
The study of range-compatible maps is a very recent topic in linear algebra.
It is mainly motivated by its connection with spaces of bounded rank matrices
and from the fact that it is a reformulation of the concept of algebraic reflexivity.
The basic result in the theory of range-compatible maps is the following one.
Theorem 1.1. Every range-compatible linear map on L(U, V ) is local.
This result was independently discovered by several authors (it is implicit in
[2], for instance). In [8], it was extended as follows.
Theorem 1.2. Let S be a linear subspace of L(U, V ) such that codim S ≤
dim V − 2. Then, every range-compatible linear map on S is local.
Then, in [5], this result was also extended to group homomorphisms, for
which the upper-bound dim V − 2 turns out to be optimal1.
Theorem 1.3. Let S be a linear subspace of L(U, V ) such that codim S ≤
dim V − 2. Then, every range-compatible homomorphism on S is local.
Theorem 1.2 is a major key in a generalization [8] of an important result
of Atkinson and Lloyd [1] on the classification of large spaces of matrices with
rank bounded above, and it was also used in a study of invertibility preservers
on large spaces of square matrices [9]. Further work on range-compatibility has
appeared recently: we refer to [5, 6, 7] for the latest developments of this topic.
1The upper-bound dim V − 2 is not optimal for linear maps, see Theorem 1.2 of [5] for the
optimal bound.
3
In this article, we shall focus on the special case when S is a linear subspace
of Sn(K) or An(K). Of course, S is then naturally identified with a linear sub-
space of L(Kn, Kn) through the canonical basis of Kn. The main motivation
for considering such subspaces is the potential application to the study of large
spaces of symmetric or alternating matrices with bounded rank: in further re-
search, the results of this article will be used to generalize theorems of Radwan
and Loewy [3, 4] on the structure of such spaces.
The determination of the range-compatible homomorphisms on Sn(K) was
achieved in [5]. For the characteristic 2 case, we need some additional terminol-
ogy before we can state the result.
Definition 1.1. Assume that K has characteristic 2. Let V and V ′ be vector
spaces over K. A root-linear map f : V → V ′ is a group homomorphism such
that
∀(λ, x) ∈ K × V, f (λ2x) = λf (x).
Note that if K has characteristic 2 and more than 2 elements, then only the
zero map from V to V ′ is both linear and root-linear. However, if #K = 2
root-linear maps coincide with linear maps.
Theorem 1.4 (Theorem 3.1 of [5]). Let n ≥ 2 be an integer.
(a) If K has characteristic not 2, then every range-compatible homomorphism
on Sn(K) is local.
(b) If K has characteristic 2, then the range-compatible homomorphisms on
Sn(K) are the sums of the local maps and the maps of the form M 7→ ∆(M )α
where α is a root-linear map from K to K.
(c) If K has more than 2 elements, then every range-compatible linear map on
Sn(K) is local.
1.3 Main results
Definition 1.2. Let S be a linear subspace of Sn(K).
A range-compatible homomorphism on S is called standard when it can be
extended into a range-compatible homomorphism on the full space Sn(K).
Thus, if K has characteristic not 2 and n ≥ 2, then standard range-compatible
homomorphisms are just local maps. If K has characteristic 2, then a range-
compatible homomorphism on S is standard if and only if it is the sum of a local
map and of a map of the form M 7→ ∆(M )α for some root-linear form α on K.
4
Now, we can state our main result on range-compatible homomorphisms over
spaces of symmetric matrices.
Theorem 1.5. Let S be a linear subspace of Sn(K) such that codim S ≤ n − 2.
Then, every range-compatible homomorphism on S is standard.
If #K > 2, only the zero map is both linear and root-linear. Thus, Theorem
1.5 yields:
Corollary 1.6. Let S be a linear subspace of Sn(K) such that codim S ≤ n − 2
and #K > 2. Then, every range-compatible linear map on S is local.
The following counter-example demonstrates that the upper-bound n − 2 is
optimal. For n ≥ 2, consider the subspace S := S1(K) ⊕ Sn−1(K) of Sn(K). Note
that S has codimension n − 1 in Sn(K). If there exists a non-linear homomor-
phism ϕ : K → K (which is the case if and only if K is not a prime field) then
the mapping
is obviously range-compatible but we check that it is not standard. Indeed, as
F is non-linear, if it is standard then K has characteristic 2 and more than 2
elements and there is a root-linear form α on K and linear forms β1, . . . , βn on
[0](n−1)×1(cid:21)
F : (mi,j) ∈ S 7→(cid:20) ϕ(m1,1)
+ ∆(M )α. Judging from the last entry
β1(M )
in F (M ), we would find that α is both root-linear and linear, and hence α = 0;
then we would see from the first entry of F (M ) that ϕ is linear, contradicting
our assumptions.
S such that F : M = (mi,j) 7→
...
βn(M )
Remark 1. An open problem remains to find the optimal upper-bound on the
codimension on S for all range-compatible linear maps on S to be local, provided
that K has more than 2 elements of course. We conjecture that a reasonable
upper-bound should be 2 dim V − 4, except for some very small fields for which
special cases might arise.
In any case, the following example shows that the
upper-bound should be less than or equal to 2 dim V − 4.
Consider the space U2 of all matrices of the form (cid:20)a b
and the map F : (mi,j) ∈ U 7→ (cid:20)m1,1
b 0(cid:21) with (a, b) ∈ K2,
0 (cid:21). The map F is non-local. However, F
5
is range-compatible because whenever M ∈ U2 r {0} satisfies F (M ) 6= 0, either
F (M ) is the first column of M or else M has rank 2 and hence its range contains
F (M ). For all n ≥ 2, we extend this counter-example as follows: we consider
the space S := U2 ⊕ Sn−2(K) and, for all M = (cid:20) A [0]
[0] B(cid:21) with A ∈ U2 and
B ∈ Sn−2(K), we set
[0](n−2)×1(cid:21) .
eF (M ) :=(cid:20) F (A)
Then, it is easy to check that eF is linear, range-compatible and non-local,
whereas codim S = 2n − 3.
Now, we turn to our main result on range-compatible homomorphisms on
spaces of alternating matrices.
Theorem 1.7. Let S be a linear subspace of An(K) such that codim S ≤ n − 3.
Then, every range-compatible homomorphism on S is local.
The following example shows that the upper-bound n−3 is optimal. Consider
indeed the subspace S consisting of all the alternating matrices A = (ai,j) such
that ai,1 = 0 for all i ∈ [[3, n]]. If K is not a prime field, we can choose a non-linear
group endomorphism ϕ of K, and hence
A ∈ S 7→
0
ϕ(a2,1)
[0](n−2)×1
is obviously range-compatible, but it is non-linear and hence non-local.
As in the symmetric case, the upper bound n − 3 in Theorem 1.7 does not
seem to be optimal if we restrict the study to range-compatible linear maps.
The following examples suggest that the optimal upper-bound in this situation
could be 2n − 6 if #K > 2, and 2n − 7 if #K = 2, unless n ≤ 3 in which case no
provision on the codimension of S appears necessary.
Example 2. Assume that n ≥ 4. Consider the space U of all matrices of M2(K) of
the form(cid:20)a b
0 a(cid:21) with (a, b) ∈ K2. As in Remark 1, one checks that(cid:20)a b
0 a(cid:21) 7→(cid:20)b
0(cid:21)
is range-compatible. Now, consider the space S of all matrices of the form
[0]2×2
A
[0](n−4)×2
−AT
B
C
6
[0]2×(n−4)
−C T
D
in which A ∈ U , B ∈ A2(K), C ∈ Mn−4,2(K) and D ∈ An−4(K). One checks
that S is a linear subspace of An(K) with codimension 2n − 5. The linear map
is range-compatible (as F (M ) is always a linear combination of the first two
columns of M ), but one checks that it is non-local.
Example 3. Assume that n ≥ 4 and #K = 2. We know from Theorem 1.4 that
c(cid:21) is range-compatible. Now, consider the space S of all matrices of
c(cid:21) 7→(cid:20)a
(cid:20)a b
b
the form
in which A ∈ S2(K), B ∈ A2(K), C ∈ Mn−4,2(K) and D ∈ An−4(K). One checks
that S is a linear subspace of An(K) with codimension 2n − 6. However, the
linear map
F : (mi,j) ∈ S 7→
0
0
m3,2
[0](n−3)×1
A
[0]2×2
[0](n−4)×2
−AT
B
C
F : (mi,j) ∈ S 7→
[0]2×(n−4)
−C T
D
0
0
m3,1
m4,2
[0](n−4)×1
is range-compatible (as F (M ) is always a linear combination of the first two
columns of M ), but one checks that it is non-local.
As an easy consequence of Theorem 1.7, we obtain the following result.
Theorem 1.8. Let n be an arbitrary non-negative integer. Every range-compatible
linear map on An(K) is local.
Let us immediately show how this result follows from Theorem 1.7.
Proof of Theorem 1.8. For n ≥ 3, the result is a direct consequence of Theorem
1.7. On the other hand, in a linear subspace S = Ks1 of L(U, V ) with dimension
at most 1, every-range compatible linear map is local:
indeed, given a linear
map F : S → V that is range-compatible, we have some x ∈ U such that
7
F (s1) = s1(x), and as the linear maps F and s 7→ s(x) coincide on a generating
subset of the vector space S, they are equal. For n ≤ 2, the claimed result thus
follows from that general basic point.
1.4 Strategy and structure of the article
In this article, we will follow the main method from [5], which consists in per-
forming inductive proofs over the dimension of the target space V , by using
quotient spaces. Here, a major problem appears: if S is a subspace of Sn(K) or
of An(K), "moding out" a non-zero subspace V0 of V yields an operator space
S mod V0 which cannot be represented by a space of symmetric of alternating
matrices, simply because the source and target spaces of the operators in S no
longer have the same dimension!
In order to rescue the quotient space tech-
nique, it is necessary to enlarge the theorem as to encompass not only subspaces
of Sn(K), but also subspaces of Sn(K)` Mn,p(K) for arbitrary integers n ≥ 2
and p ≥ 0 (and ditto for alternating matrices instead of symmetric ones). To
avoid the need of constantly going from matrix spaces to operator spaces, we
shall reframe the results in terms of subspaces of self-adjoint or alternating oper-
ators from U to the dual space of one of its subspaces V . Then, the generalized
versions of Theorems 1.5 and 1.7 can be proved by induction on the dimension
of V (with fixed U ), and Theorems 1.5 and 1.7 follow from the special case when
U = V .
The rest of the article is organized as follows. In Section 2, we give a quick re-
view of the main techniques for the study of range-compatible homomorphisms,
that is the quotient space technique and the splitting technique. We also intro-
duce some notation on duality that is used throughout the text. Sections 3 and
4 are devoted to the study of range-compatible homomorphisms, respectively on
spaces of self-adjoint operators and on spaces of alternating operators. These
two sections are logically independent from one another, but they are essentially
based upon similar methods. In the self-adjoint case, the main difficulty comes
from the case of fields with characteristic 2, whereas for alternating operators
the proof is common to all fields, but substantially more technical.
2 Quotient and splitting space techniques
2.1 Quotient space techniques
We recall the following notation and result from [5].
8
Notation 2.1. Let S be a linear subspace of L(U, V ) and V0 be a linear subspace
of V . Denote by π : V → V /V0 the canonical projection. Then, we set
which is a linear subspace of L(U, V /V0).
S mod V0 :=(cid:8)π ◦ s s ∈ S(cid:9),
Lemma 2.1. Let F : S → V be a range-compatible homomorphism. Then, there
is a unique range-compatible homomorphism F mod V0 : S mod V0 → V /V0 such
that
∀s ∈ S, (F mod V0)(π ◦ s) = π(F (s)),
i.e. such that the following diagram is commutative
S
F
s7→π◦s
V
π
S mod V0
F mod V0
/ V /V0.
When V0 = Ky for some non-zero vector y, we simply write S mod y instead
of S mod Ky, and F mod y instead of F mod Ky.
In terms of matrices, the case when V0 is a hyperplane reads as follows.
Corollary 2.2. Let S be a linear subspace of Mn,p(K) and F : S → Kn
be a range-compatible homomorphism. Then, there are group homomorphisms
F1, . . . , Fn such that
∀M ∈ S, F (M ) =
F1(R1(M ))
...
Fn(Rn(M ))
where R1(M ), . . . , Rn(M ) denote the rows of M .
2.2 On duality
Notation 2.2. Given a vector space V , the dual space of V (i.e. the set of all
linear forms on V ) is denoted by V ⋆.
9
/
/
/
Definition 2.3. Let V be a linear subspace of the vector space U . We denote
by V o the linear subspace of U ⋆ consisting of the linear forms on U that vanish
everywhere on V .
Let W be a linear subspace of U ⋆. We denote by oW the linear subspace of
U consisting of the vectors that annihilate all the linear forms in W .
In any case, the letter "o" stands for "orthogonal".
Remember that if U is finite-dimensional then o(V o) = V for every linear
subspace V of U , and (oW )o = W for every linear subspace W of U ⋆. Moreover,
for every linear subspace W of U ⋆, we have a natural isomorphism
U ⋆/W
≃−→ (oW )⋆.
In the rest of the article, we shall systematically use this isomorphism to identify
U ⋆/W with (oW )⋆. In particular, given a linear form f ⋆ ∈ U ⋆, the space U ⋆/Kf ⋆
is naturally identified with (Ker f ⋆)⋆.
We finish with some terminology on dual bases.
Definition 2.4. Let U be a finite-dimensional vector space and V be a linear
subspace of U , with respective dimensions p and n. A compatible pair of
bases of U and V ⋆ is a pair (B1, B⋆
2) in which B1 = (e1, . . . , ep) is a basis of U
such that (e1, . . . , en) is a basis of V whose dual basis is B⋆
2.
2.3 The splitting technique
Notation 2.5. Let n, p, q be non-negative integers, and A and B be respective
subsets of Mn,p(K) and Mn,q(K). We set
Aa B :=n(cid:2)A B(cid:3) (A, B) ∈ A × Bo.
The following easy lemma was established in [5].
Lemma 2.3 (Splitting Lemma). Let n, p, q be non-negative integers, and A and
B be linear subspaces, respectively, of Mn,p(K) and Mn,q(K).
Given maps f : A → Kn and g : B → Kn, set
fa g :(cid:2)A B(cid:3) ∈ Aa B 7−→ f (A) + g(B).
Then:
10
where f ∈ Hom(A, Kn) and g ∈ Hom(B, Kn).
(a) The homomorphisms from A` B to Kn are the maps of the form f` g,
(b) Given f ∈ Hom(A, Kn) and g ∈ Hom(B, Kn), the homomorphism f` g
is range-compatible (respectively, local) if and only if f and g are range-
compatible (respectively, local).
3 Range-compatible homomorphisms on large spaces
of symmetric matrices
3.1 Spaces of self-adjoint operators
The main obstacle in trying to prove Theorem 1.5 by induction is that, given a
vector y ∈ Kn r {0}, the operator space S mod y can no longer be represented
by a space of symmetric matrices. In order to circumvent that problem, we shall
both broaden the result and frame it in terms of operators rather than in terms
of matrices.
Definition 3.1. Let U be a vector space and V be a linear subspace of U . A
linear map f : U → V ⋆ is called self-adjoint when
∀(x, y) ∈ V 2,
f (x)[y] = f (y)[x],
i.e. when the bilinear form (x, y) ∈ V 2 7→ f (x)[y] is symmetric. The set of all
self-adjoint linear maps from U to V ⋆ is denoted by Ls(U, V ⋆). One checks that
it is a linear subspace of L(U, V ⋆).
Let (B1, B⋆
2) be a compatible pair of bases of U and U ⋆. A linear map from U
to U ⋆ is self-adjoint if and only if its matrix in B1 and B⋆
2 is symmetric. Hence,
we have a vector space isomorphism f ∈ Ls(U, U ⋆) 7−→ MB1,B⋆
2(f ) ∈ Sn(K),
where n := dim U . More generally, given a compatible pair (B1, B⋆
2) of bases
of U and V ⋆, a linear map f : U → V ⋆ is self-adjoint if and only if its matrix
in the bases B1 and B⋆
p := dim U .
Remark 4. Let W be a linear subspace of V ⋆. Then, through the canonical iden-
tification between V ⋆/W and (oW )⋆, the space Ls(U, V ⋆) mod W corresponds to
Ls(U, (oW )⋆).
2 belongs to Sn(K)` Mn,p−n(K), where n := dim V and
In particular, Ls(U, V ⋆) corresponds to Ls(U, U ⋆) mod V o.
11
Note these identifications have no impact on the problems under scrutiny,
which follows from the following general observation: given a linear subspace
S of L(U1, U2) (where U1 and U2 are arbitrary vector spaces) , a vector space
2 and a group homomorphism F : S → U2, the linear
isomorphism ϕ : U2
subspace S ′ := {ϕ ◦ s s ∈ S} of L(U1, U ′
2) is isomorphic to S through Φ :
s 7→ ϕ ◦ s, and F ′ := ϕ ◦ F ◦ Φ−1 is a group homomorphism from S ′ to U ′
2;
one checks that F is range-compatible (respectively, local) if and only if F ′ is
range-compatible (respectively, local).
≃→ U ′
Combining the splitting lemma and Theorems 1.2 and 1.4, we can easily de-
scribe the general form of range-compatible homomorphisms on Sn(K)` Mn,p(K)
for arbitrary integers n ≥ 2 and p ≥ 0. The interpretation in terms of operators
then reads as follows.
Proposition 3.1. Let U be a finite-dimensional vector space and V be a linear
subspace of U with dim V ≥ 2. Let (cid:0)(e1, . . . , ep), (e⋆
pair of bases of U and V ⋆.
1, . . . , e⋆
n)(cid:1) be a compatible
(a) If K does not have characteristic 2 then every range-compatible homomor-
phism on Ls(U, V ⋆) is local.
(b) If K has characteristic 2 then every range-compatible homomorphism on
Ls(U, V ⋆) splits in a unique fashion as the sum of a local map and of s 7→
α(s(ek)[ek]) e⋆
k for some root-linear form α on K.
nPk=1
As a corollary, we obtain the following result.
Corollary 3.2. Let U be a finite-dimensional vector space and V be a linear
subspace of U with dim V ≥ 2. Then, every range-compatible homomorphism on
Ls(U, V ⋆) can be written as G mod V o for some range-compatible homomorphism
G on Ls(U, U ⋆).
Proof. Let F be a range-compatible homomorphism on Ls(U, V ⋆). If F is local,
we have F : s 7→ s(x) for some x ∈ U and then it is clear that F = G mod V o
for G : s ∈ Ls(U, U ⋆) 7→ s(x).
Now, assume that K has characteristic 2. Let (e⋆
n) be a basis of V ⋆,
with pre-dual basis (e1, . . . , en) of V , which we extend into a basis (e1, . . . , ep)
of U . Denote by (f ⋆
p ) the dual basis of (e1, . . . , ep). One sees that
n+1, . . . , f ⋆
f ⋆
p belong to V o, whereas (f ⋆
1 , . . . , f ⋆
i )V = e⋆
i for all i ∈ [[1, n]].
1, . . . , e⋆
12
nPk=1
pPk=1
By Proposition 3.1, the map F splits as F1 + F2 with some local map F1 :
k for some root-linear form α on K.
s 7→ s(x) and with F2 : s 7→
α(s(ek)[ek]) e⋆
Setting G2 : s ∈ Ls(U, U ⋆) 7→
k , we check that G2 mod V o = F2,
whereas G1 mod V o = F1 for G1 : s ∈ Ls(U, U ⋆) 7→ s(x). Hence, F = (G1 +
G2) mod V o and G1 + G2 is range-compatible.
α(s(ek)[ek]) f ⋆
Now, we can extend the definition of standard maps as follows.
Definition 3.2. Let S be a linear subspace of Ls(U, V ⋆). A range-compatible
homomorphism on S is called standard when it can be extended into a range-
compatible homomorphism on the full space Ls(U, V ⋆).
In particular, when K has characteristic not 2 the standard range-compatible
homomorphisms are simply the local ones. Note that in any case the set of
all standard range-compatible homomorphisms on S is a linear subspace of
Hom(S, V ⋆).
We are now ready to state our main result on range-compatible homomor-
phisms on spaces of self-adjoint operators.
Theorem 3.3. Let U be a finite-dimensional vector space, V be a linear subspace
of U , and S be a linear subspace of Ls(U, V ⋆). Assume that codim S ≤ dim V −2
(here, the codimension of S is meant to be the one of S in the space Ls(U, V ⋆),
not in the full space L(U, V ⋆)). Then, every range-compatible homomorphism
on S is standard.
The special case U = V is simply Theorem 1.5.
The proof will require the following additional notation and remarks. Let S
be a linear subspace of Ls(U, V ⋆). To S, we can attach two operator spaces:
• The space
Sr :=(cid:8)fV f ∈ S(cid:9),
which is a linear subspace of Ls(V, V ⋆) (the subscript "r" stands for "re-
stricted");
• And the space of all f ∈ S that vanish everywhere on V , which we denote
by Sm and naturally reinterpret as a linear subspace of L(U/V, V ⋆) (the
subscript "m" stands for "modulo").
13
By the rank theorem, we have
dim S = dim Sr + dim Sm,
and on the other hand
dim Ls(U, V ⋆) = dim Ls(V, V ⋆) + dim L(U/V, V ⋆).
Let (B1, B⋆
2) be a compatible pair of bases of U and V ⋆, and denote by
M the matrix space representing S in them. Write B1 = (e1, . . . , ep) and set
B′
1 := (en+1, . . . , ep), which are bases of V and U/V ,
respectively. Every M ∈ M splits as
1 := (e1, . . . , en) and B′′
M =(cid:2)S(M ) R(M )(cid:3) with S(M ) ∈ Sn(K) and R(M ) ∈ Mn,p−n(K).
Then, the matrix space S(M) represents Sr in the bases B′
1. On the
other hand, if we denote by T the subspace of M consisting of all its matrices
M such that S(M ) = 0, then R(T ) represents Sm in the bases B′′
1 and B⋆
1 and B⋆
1.
3.2 Basic lemmas
Lemma 3.4. Let U be a finite-dimensional vector space with dim U ≥ 3. Let
W be a proper linear subspace of Ls(U, U ⋆). Then, there exist non-collinear
vectors f ⋆
1 and
no operator with range Kf ⋆
2 .
2 of U ⋆ such that W contains no operator with range Kf ⋆
1 and f ⋆
Proof. Assume on the contrary that there exists a non-zero linear form f ⋆ ∈ U ⋆
such that W contains a rank 1 operator with range g⋆ for all g⋆ ∈ U ⋆ r Kf ⋆.
We can find a basis B of U such that the coordinates of f ⋆ in the dual basis
B⋆ are all equal to 1. Denote by M the matrix space that represents W in the
bases B and B⋆. Then, M should be a proper linear subspace of Sn(K) where
n := dim U . Set X0 := (cid:2)1 · · ·
1(cid:3)T ∈ Kn. For all X ∈ Kn r KX0, we know
that M contains aXX T for some a ∈ K r {0}, and hence M actually contains
XX T for all such X. In particular, as n ≥ 3, we see that M contains Ei,i for
all i ∈ [[1, n]], and it also contains Ei,i + Ej,j + Ei,j + Ej,i for all i and j in
[[1, n]]2 with i 6= j. Hence, it contains Ei,j + Ej,i for all i and j in [[1, n]]2 with
i 6= j. By linearly combining those special matrices, we obtain that M = Sn(K),
contradicting the fact that M is a proper subspace of Sn(K).
Hence, for all f ⋆ ∈ U ⋆, there exists g⋆ ∈ U ⋆ r Kf ⋆ such that W contains no
1 ∈ U ⋆ r {0} such
operator with range Kg⋆. Starting from f ⋆ = 0, we obtain f ⋆
14
that W contains no operator with range Kf ⋆
more to find some f ⋆
Kf ⋆
2 .
2 ∈ U ⋆ r Kf ⋆
1 , and then we apply the result once
1 such that W contains no operator with range
Lemma 3.5. Let U be a vector space, V be a linear subspace of U , and S
be a linear subspace of Ls(U, V ⋆). Assume that codim S ≤ dim V − 2 and
dim V ≥ 3. Then, there are non-collinear vectors f ⋆ and g⋆ in V ⋆ such that
codim(S mod f ⋆) ≤ dim V − 3 and codim(S mod g⋆) ≤ dim V − 3.
Proof. Let f ⋆ be a non-zero vector of V ⋆. We say that f ⋆ is S-good if codim(S mod f ⋆) ≤
dim V − 3. Thus, f ⋆ is not S-good if and only if codim S = dim V − 2 and S
contains every rank 1 operator of Ls(U, V ⋆) with range Kf ⋆. Assume that we
1 , · · · , f ⋆
cannot find non-collinear S-good vectors. Then, we can find a basis (f ⋆
n)
of V ⋆ in which no vector is S-good. In particular, for all i ∈ [[1, n]], the space S
contains every operator whose kernel includes V and whose range is included in
Kf ⋆
i . Summing such operators yields that Sm = L(U/V, V ⋆), whence
codim Sr = codim S = dim V − 2 > 0.
Applying Lemma 3.4, we deduce that there are non-collinear vectors f ⋆ and g⋆
of V ⋆ such that Sr contains no operator whose range equals either Kf ⋆ or Kg⋆.
Then, f ⋆ and g⋆ are both S-good, which completes the proof.
Lemma 3.6. Let S be a linear subspace of Ls(U, V ⋆) such that codim S ≤
dim V −2 and dim V ≥ 3. Let F : S → V ⋆ be a range-compatible homomorphism.
Assume that there exist non-collinear vectors f ⋆
2 in V ⋆ such that both
maps F mod f ⋆
1 and f ⋆
2 are local. Then, F is local.
1 and F mod f ⋆
Proof. Subtracting a local map from F , we lose no generality in assuming that
F mod f ⋆
1 = 0, in which case we have a vector x ∈ U such that
and
∀s ∈ S, F (s) = s(x) mod Kf ⋆
2 ,
∀s ∈ S, F (s) ∈ Kf ⋆
1 .
In particular, we note that
∀s ∈ S,
s(x) ∈ span(f ⋆
1 , f ⋆
2 ).
Now, we show that x 6= 0 would lead to a contradiction. We distinguish
between three cases.
15
Case 1: x 6∈ V .
We choose a basis B = (e1, . . . , en) of V , whose dual basis we denote by B⋆.
dimensional) space of coordinates of the vectors of span(f ⋆
from the above that M is a subspace of M′
However, dim M ≥ dim M′, whence M = M′. Denote by G : M → Kn
the map on matrices that corresponds to F in the above bases. Then, the
range-compatible homomorphism G′ : P → Kn deduced from restricting G to
Then, (e1, . . . , en, x) can be extended into a basis eB of U . We denote by M
the space of matrices representing S in eB and B⋆, and we denote by P the (2-
:= Sn(K)` P` Mn,p−n−1(K).
{0}` P`{0} has rank 1! This is absurd as G′, being local by Theorem 1.3,
should be a scalar multiple of the identity of P .
2 ) in B⋆. It follows
1 , f ⋆
1 , f ⋆
1 , f ⋆
2 ) and x 6= 0.
Case 2: x ∈ o span(f ⋆
In that case, we choose two vectors e1, e2 of V such that f ⋆
i (ej ) = δi,j for all
(i, j) ∈ [[1, 2]]2, we note that x 6∈ span(e1, e2), and we extend (e1, e2) into a basis
B = (e1, . . . , en) of V such that en = x and all the vectors e3, . . . , en belong
to o span(f ⋆
n). Finally, we
2 ). Then, the dual basis of V ⋆ reads (f ⋆
2 , . . . , f ⋆
of Sn(K) consisting of the matrices (si,j) such that si,n = 0 for all i ∈ [[3, n]].
extend B into a basis eB of U . In the bases eB and B⋆, the matrix space M that
corresponds to S is included in N` Mn,p−n(K), where N denotes the subspace
Obviously, N` Mn,p−n(K) has codimension n−2 in Sn(K)` Mn,p−n(K), whence
M = N` Mn,p−n(K). Moreover, the map that corresponds to F in those bases
1 , f ⋆
is
G : (mi,j) ∈ M 7−→(cid:20) m1,n
[0](n−1)×1(cid:21) .
Define M = (mi,j) ∈ M by mi,j = 1 whenever i ≤ 2 or j ≤ 2, and mi,j = 0
otherwise. Then, G(M ) =(cid:20)
M is spanned by (cid:2)1 1 0 · · ·
1
[0](n−1)×1(cid:21), but one easily checks that the range of
, and hence it is
0(cid:3)T
and (cid:2)0 0 1 · · ·
1(cid:3)T
obvious that this range does not contain G(M ).
2 ).
1 , f ⋆
1 , f ⋆
Case 3: x ∈ V r o span(f ⋆
We can extend (f ⋆
i (x) = 0
for all i ∈ [[3, n]]. We denote by (e1, . . . , en) its dual basis of V , which we extend
2 ) into a basis B⋆ = (f ⋆
n) of V ⋆ such that f ⋆
1 , . . . , f ⋆
are scalars α and β such that x = α e1 + β e2. Now, denote by M the space of
into a basis eB = (e1, . . . , ep) of U . Note that x ∈ span(e1, e2), and hence there
matrices representing S in eB and B⋆. Then, M is included in N1` Mn,p−n(K),
where N1 denotes the subspace of Sn(K) consisting of all the matrices (si,j) such
16
that α si,1 + β si,2 = 0 for all i ∈ [[3, n]].
As codim S ≤ n − 2, we see that dim(N1` Mn,p−n(K)) ≤ dim M, and hence
M = N1` Mn,p−n(K). Then, the map that corresponds to F through the above
bases reads
and hence it is range-compatible. It follows from the splitting lemma that
[0](n−1)×1 (cid:21) ,
G : (si,j) 7→(cid:20)α s1,1 + β s1,2
[0](n−1)×1 (cid:21)
G′ : (si,j) ∈ N1 7→(cid:20)α s1,1 + β s1,2
is also range-compatible.
Applying G′ to any matrix of the form (cid:20)
S
[0](n−2)×2
[0]2×(n−2)
[0](n−2)×(n−2)(cid:21), with
S ∈ S2(K), yields that the linear map
(cid:20)a b
c(cid:21) ∈ S2(K) 7→(cid:20)αa + βb
0
b
(cid:21)
is range-compatible. Yet, since (α, β) 6= (0, 0) it is obvious that this map is
non-local. Hence, by Theorem 1.4 we have #K = 2. Applying the above map to
(cid:20)1 1
1 1(cid:21) then yields α + β = 0, whence α = β = 1 (since (α, β) 6= (0, 0)). A final
contradiction then comes by applying G′ to the matrix M = (mi,j) ∈ N1 defined
by m1,j = m2,j = mj,1 = mj,2 = 1 for all j ∈ [[3, n]], m1,2 = 1, m1,1 = m2,2 = 0
and mi,j = 0 for all (i, j) ∈ [[3, n]]2. One checks that the range of M is spanned
by the vectors
and
and that it does not include(cid:20)
1
[0](n−1)×1(cid:21) = G′(M ).
0
1
1
...
1
1
0
1
...
1
This contradicts the fact that G′ is range-compatible.
In any of the above cases, we have found a contradiction. The only remaining
possibility is that x = 0. Then, F is the local map s 7→ s(0).
3.3 Proof of Theorem 3.3 for fields with characteristic not 2
Here, we assume that the characteristic of K is not 2.
17
We prove Theorem 3.3 by induction on the dimension of V . The result is vac-
uous if dim V ≤ 1. If dim V = 2, then the result is obvious because the assump-
tions tell us that S = Ls(U, V ⋆). Assume now that dim V ≥ 3. By Lemma 3.5,
we can pick non-collinear vectors f ⋆
1 ) ≤
dim V − 3 and codim(S mod f ⋆
2 ) ≤ dim V − 3. However, we have a canoni-
cal identification of S mod f ⋆
i )⋆), and
dim Ker f ⋆
i is local. Lemma
3.6 then yields that F is local. This completes the proof.
i = dim V −1. Thus, by induction, each map F mod f ⋆
i with a linear subspace of Ls(U, (Ker f ⋆
2 in V ⋆ such that codim(S mod f ⋆
1 and f ⋆
3.4 Proof of Theorem 3.3 for fields with characteristic 2 and
more than 2 elements
Here, we assume that K has characteristic 2 and more than 2 elements. Again, we
prove the result by induction on the dimension of V . The case when dim V ≤ 2 is
either vacuous or trivial. Assuming that dim V ≥ 3 and considering an arbitrary
range-compatible homomorphism F : S → V ⋆, we proceed as in the above proof
2 in V ⋆ such that both maps F mod f ⋆
to find linearly independent forms f ⋆
1
1 equals F ′ mod f ⋆
and F mod f ⋆
1 for some standard
map F ′ on S. Thus, replacing F with F − F ′, we see that no generality is lost
in assuming that
1 and f ⋆
2 are standard. Then, F mod f ⋆
∀s ∈ S, F (s) ∈ Kf ⋆
1 .
Claim 1. The map F mod f ⋆
2 is local.
2 is non-local. Then, F mod f ⋆
Proof. Assume on the contrary that F mod f ⋆
2 is
the sum of a local map x 7→ s(x), for some vector x ∈ U , and of a non-zero
root-linear map. Let us extend (f ⋆
2 , . . . ) of V ⋆, and then
consider the dual basis (e1, . . . , en) of V , which we extend into a basis (e1, . . . , ep)
of U . Denote by M the matrix space associated with S in the bases (e1, . . . , ep)
n), and by G : M → Kn the map on matrices that corresponds to
and (f ⋆
F . One deduces that there is a non-zero root-linear form α on K together with
linear form β1, β3, . . . , βn on M such that
2 ) into a basis (f ⋆
1 , . . . , f ⋆
1 , f ⋆
1 , f ⋆
G : M = (mi,j) 7−→
α(m1,1) + β1(M )
?
α(m3,3) + β3(M )
...
α(mn,n) + βn(M )
18
On the other hand, as F mod f ⋆
1 = 0, it turns out that
∀i ∈ [[3, n]], ∀M ∈ S, α(mi,i) = βi(M ).
As K has more than 2 elements, only the zero map is both root-linear and linear,
and hence β3 = · · · = βn = 0 and, as α is non-zero, we deduce that
∀i ∈ [[3, n]], ∀M ∈ S, mi,i = 0.
Denote by T the subspace of Sn(K) consisting of all the matrices M = (mi,j) ∈
Sn(K) such that ∀i ∈ [[3, n]], mi,i = 0. Then, M ⊂ T ` Mn,p−n(K) and
dim(T ` Mn,p−n(K)) ≤ dim S, whence M = T ` Mn,p−n(K).
Denote by H : S2(K) → K2 the homomorphism such that
∀A ∈ S2(K), G(cid:18)(cid:20)
A
[0](n−2)×2
[0]2×(p−2)
[0](n−2)×(p−2)(cid:21)(cid:19) =(cid:20) H(A)
[0](n−2)×1(cid:21) .
As G is range-compatible, so is H. Note that we have a linear form θ on S2(K)
such that
Then, by Theorem 1.4, the map H equals
0
H : A = (ai,j) 7→(cid:20)α(a1,1) + θ(A)
(cid:21) .
A = (ai,j) 7→(cid:20)α′(a1,1) + θ1(A)
α′(a2,2) + θ2(A)(cid:21)
for some root-linear form α′ on K and some linear forms θ1 and θ2 on S2(K).
As above we would deduce from looking at the first row that α′ = α, and from
looking at the second row that α′ = 0. This leads to α = 0, contradicting earlier
results.
Thus, both maps F mod f ⋆
2 are local. By Lemma 3.6, the map
F is local. This completes the inductive step. Hence, Theorem 3.3 is now proved
for all fields with more than 2 elements.
1 and F mod f ⋆
3.5 Proof of Theorem 3.3 for fields with 2 elements
Here, we use a similar strategy as in the previous two proofs, but now we need
a stronger version of Lemma 3.5. In its statement, we shall use the following
19
notation:
T3(K) :=(
a b
c
b d 0
c 0 e
(a, b, c, d, e) ∈ K5) ⊂ S3(K).
Note that T3(K) is a linear hyperplane of S3(K).
Lemma 3.7. Assume that #K = 2. Let U be a finite-dimensional vector space,
V be a linear subspace of U , and S be a linear subspace of Ls(U, V ⋆). Assume
that codim S ≤ dim V − 2 and dim V ≥ 3. Then:
(a) Either there exist distinct non-zero vectors f ⋆
i ) ≤ dim V − 3 for all i ∈ [[1, 3]];
codim(S mod f ⋆
1 , f ⋆
2 and f ⋆
3 in V ⋆ such that
(b) Or dim V = 3 and S is represented, in a choice of bases of U and V ⋆, by
T3(K)` M3,p−3(K), where p := dim U .
1 , . . . , f ⋆
Proof. We use a similar line of reasoning as in the proof of Lemma 3.5. Assume
first that Sr = Ls(V, V ⋆). If S = Ls(U, V ⋆), we simply pick three distinct vectors
in V ⋆r{0}. If S ( Ls(U, V ⋆), then Sm ( L(U/V, V ⋆), and hence there is no basis
n) of V ⋆ for which, for all i ∈ [[1, n]], the linear subspace Sm contains
(f ⋆
every operator of L(U/V, V ⋆) with range Kf ⋆
i ; hence, there is a linear hyperplane
H of V ⋆ such that, for all f ⋆ ∈ V ⋆ r H, the space S does not contain every
operator of Ls(U, V ⋆) with range Kf ⋆, leading to codim(S mod f ⋆) ≤ dim V − 3.
In any case, condition (a) is easily seen to hold.
1 , f ⋆
1 and f ⋆
In the rest of the proof, we assume that Sr ( Ls(V, V ⋆) and that condition (a)
2 in V ⋆ r {0} such that,
does not hold. Then, there are distinct vectors f ⋆
2 }, one has codim(S mod f ⋆) > dim V −3. For any such
for all f ⋆ ∈ V ⋆ r{0, f ⋆
f ⋆, we deduce that S contains all the operators of Ls(U, V ⋆) with range Kf ⋆.
Then, as those linear forms span V ⋆, we learn that Sm = L(U/V, V ⋆). On the
other hand, as there is at least one such linear form we find codim S = dim V − 2
and hence codim Sr = dim V − 2. As f ⋆
2 are non-collinear (as they are
n) of V ⋆
distinct and non-zero, while #K = 2), there is a basis B⋆ = (e⋆
e⋆
k. Consider the predual basis B of V , and
denote by M the matrix space representing Sr is the bases B and B⋆. Note that
M has codimension n − 2 in Sn(K). Moreover, we know that, for every X ∈ Kn
in which f ⋆
k and f ⋆
e⋆
nPk=2
nPk=1
1 and f ⋆
1 =
2 =
1, . . . , e⋆
20
with the possible exception of X1 :=
0
1
...
1
contain XX T . As in the proof of Lemma 3.4, this shows:
and X2 :=
1
...
1
, the space M should
• That M contains Ei,i for all i ∈ [[1, n]];
• That M contains Ei,j + Ei,j for all (i, j) ∈ [[1, n]]2 such that i 6= j, unless
n = 3 in which case we can only assert that M contains E1,2 + E2,1 and
E1,3 + E3,1.
If n > 3, then we deduce that M = Sn(K), which contradicts the fact that
codim M = n−2. Therefore, n = 3, and we see from the above that T3(K) ⊂ M.
However dim T3(K) = 5 = dim M, whence M = T3(K). Let us extend B into a
basis eB of U . Then, as Sm = L(U/V, V ⋆), we conclude that in the bases eB and
B⋆ the matrix space T3(K)` M3,p−3(K) represents S.
Now, let us assume that #K = 2. We prove Theorem 3.3 by induction on the
dimension of V . Again, the case dim V ≤ 2 is either vacuous or trivial. Assume
now that dim V ≥ 3.
Claim 2. There are distinct non-zero forms f ⋆
map F mod f ⋆
i
is standard.
1 , f ⋆
2 , f ⋆
3 in V ⋆ such that each
1
1
1
0
1
1
Proof. If condition (a) from Lemma 3.7 holds, then the result is an obvious
consequence of the induction hypothesis. Assume now that S is represented, in a
1
0
0
2 and f ⋆
1 , f ⋆
compatible pair (B, C⋆) of bases of U and V ⋆, by T3(K)` M3,p−3(K). Set X1 :=
, X2 :=
and X3 :=
and note that X1, X2, X3 are distinct non-zero
3 is represented by the matrix space K2` D1` D2` M2,p−3(K),
vectors of K3. Denote by f ⋆
3 the corresponding vectors of V ⋆ (through
B⋆). Then, as T3(K) contains no rank 1 matrix with range either KX1 or KX2,
we see that codim(S mod f ⋆
2 ) = 0. On the other hand, we see
that S mod f ⋆
where D1 and D2 respectively denote the 1-dimensional linear subspaces K×{0}
and {0} × K of K2. Combining Theorem 1.1 with the splitting lemma, we
find that every range-compatible linear map on S mod f ⋆
3 is local. Thus, as
here #K = 2, every range-compatible homomorphism on S mod f ⋆
3 is local. In
particular, F mod f ⋆
i
is standard for all i ∈ {1, 2, 3}.
1 ) = codim(S mod f ⋆
21
2 = G2 mod f ⋆
1 and F mod f ⋆
2 are both non-local. We know that F mod f ⋆
We are now about to conclude the proof. Let us pick distinct non-zero forms
2 , f ⋆
f ⋆
1 , f ⋆
3 satisfying the conclusion of Claim 2. Without loss of generality, we
can assume that F mod f ⋆
2 are both local or both non-local. In
the first case, Lemma 3.6 yields that F is local. Assume now that F mod f ⋆
1
and F mod f ⋆
1 and
F mod f ⋆
2 for some standard range-compatible homomorphisms G1
and G2 on S. Neither G1 nor G2 is local. Moreover, the identity is the sole non-
zero root-linear map from K to K. Hence, from the general form of standard
maps that is given in Proposition 3.1, we deduce that G1 − G2 is local. Thus,
1 and F ′ mod f ⋆
replacing F with F ′ = F − G1, we see that both maps F ′ mod f ⋆
2
are local. It follows again from Lemma 3.6 that F ′ is local, and we conclude
that F = F ′ + G1 is standard.
1 = G1 mod f ⋆
Thus, the inductive step is completed for fields with two elements. Theorem
3.3 is now established, and Theorem 1.5 follows from it.
4 Range-compatible homomorphisms on large spaces
of alternating matrices
4.1 Main results
As in the symmetric case, it is necessary, in order to obtain Theorem 1.7, to
consider a general version that applies to a larger class of operator spaces.
Definition 4.1. Let U be a finite-dimensional vector space and V be a linear
subspace of U . A linear map f : U → V ⋆ is called alternating when
∀x ∈ V, f (x)[x] = 0,
i.e. when the bilinear form (x, y) ∈ V 2 7→ f (x)[y] is alternating. The set of all
alternating linear maps from U to V ⋆ is denoted by La(U, V ⋆). One checks that
it is a linear subspace of L(U, V ⋆).
For every compatible pair (B1, B⋆
2) of bases of U and V ⋆, a map from
2 belongs to
U to V ⋆ is alternating if and only if its matrix in B1 and B⋆
An(K)` Mn,p−n(K), where n := dim V and p := dim U .
Remark 5. Let W be a linear subspace of V ⋆. Through the canonical identifi-
cation between V ⋆/W and (oW )⋆, the space La(U, V ⋆) mod W corresponds to
La(U, (oW )⋆).
22
In particular, La(U, V ⋆) corresponds to La(U, U ⋆) mod V o.
As in Remark 4, such identifications have no impact on the specific problem
we are studying.
Let S be a linear subspace of La(U, V ⋆). To S, we naturally attach two
operator spaces: the space
Sr :=(cid:8)fV f ∈ S(cid:9),
which is a linear subspace of La(V, V ⋆), and the space of all f ∈ S such that f
vanishes everywhere on V , which we denote by Sm and naturally identify with
a linear subspace of L(U/V, V ⋆). By the rank theorem, we have
dim S = dim Sr + dim Sm
whereas
dim La(U, V ⋆) = dim La(V, V ⋆) + dim L(U/V, V ⋆).
2) be a compatible pair of bases of U and V ⋆, and denote by M
Let (B1, B⋆
the matrix space representing S in them. Write B1 = (e1, . . . , ep) and set
B′
1 := (en+1, . . . , ep), which are bases of V and U/V ,
respectively. Every M ∈ M splits as
1 := (e1, . . . , en) and B′′
M =(cid:2)S(M ) R(M )(cid:3) with S(M ) ∈ An(K) and R(M ) ∈ Mn,p−n(K).
Then, the matrix space S(M) represents Sr in the bases B′
2. On the
other hand, if we denote by T the subspace of S consisting of all its matrices M
such that S(M ) = 0, then R(T ) represents Sm in the bases B′′
1 and B⋆
1 and B⋆
2.
The proper generalization of Theorem 1.7 to spaces of alternating linear
operators is the following one.
Theorem 4.1. Let U be a finite-dimensional vector space and V be a lin-
ear subspace of U . Let S be a linear subspace of La(U, V ⋆) and assume that2
codim S ≤ dim V − 2 and codim Sr ≤ dim V − 3. Then, every range-compatible
homomorphism on S is local.
Again, the proof of Theorem 4.1 will be done by induction on the dimension
of V . The case when dim V = 3 is dealt with in Section 4.2, and the inductive
step is performed in Section 4.3.
2 Here, codim S and codim Sr stand, respectively, for the codimension of S in La(U, V ⋆)
and for the one of Sr in La(V, V ⋆).
23
4.2 The case dim V = 3
Here, we let U be a finite-dimensional vector space and V be a linear subspace
of U with dimension 3. Let S be a linear subspace of La(U, V ⋆) such that
codim Sr ≤ dim V − 3 and codim S ≤ dim V − 2. Then, Sr = La(V, V ⋆) and Sm
has codimension at most 1 in L(U/V, V ⋆). Set p := dim U .
Notation 4.2. For any non-negative integer r, we denote by slr(K) the space
of all trace zero matrices in Mr(K).
Let us consider the bilinear form
b : (s, t) ∈ L(U/V, V ⋆) × L(V ⋆, U/V ) 7→ tr(v ◦ u).
The orthogonal complement of Sm with respect to b is then a linear subspace
S ⊥
m of L(V ⋆, U/V ) with dimension at most 1, and its orthogonal complement for
b is Sm. There are two main cases to consider:
• Case 1: S ⊥
m = {0}.
Then, S = La(U, V ⋆).
m contains a non-zero operator, say t, with rank de-
• Case 2: S ⊥
noted by r.
We can find bases (e⋆
in U ) in which t is represented by(cid:20)
of the form (cid:20) N
[?](3−r)×r
1, e⋆
2, e⋆
3) and (f4, . . . , fp) of V ⋆ and U/V (with f4, . . . , fp
Ir
[0](p−3−r)×r
[0]r×(3−r)
[0](p−3−r)×(3−r)(cid:21). Then,
in those bases, the space Sm is represented by the set of all matrices
[?]r×(p−3−r)
[?](3−r)×(p−3−r)(cid:21) with N ∈ slr(K).
If we de-
note by (e1, e2, e3) the predual basis of (e⋆
3), we find that B =
(e1, e2, e3, f4, . . . , fp) is a basis of U . Finally, as Sr = La(V, V ⋆), we con-
clude that there is a linear map f : A3(K) → Mr(K) such that S is repre-
sented by the space of all matrices of the form
2, e⋆
1, e⋆
(cid:2)A [0]3×(p−3)(cid:3) +(cid:20) [0]r×3
[0](3−r)×3
with A ∈ A3(K) and N ∈ slr(K).
N + f (A)
[?](3−r)×r
[?]r×(p−3−r)
[?](3−r)×(p−3−r)(cid:21)
Case 2 helps motivate the following notation.
24
Notation 4.3. Let r ∈ [[0, 3]] and f : A3(K) → Mr(K) be a linear map. We
denote by Mf the space of all 3 by 3 + r matrices of the form
(cid:2)A [0]3×r(cid:3) +(cid:20) [0]r×3
[0](3−r)×3
with A ∈ A3(K) and N ∈ slr(K).
N + f (A)
[?](3−r)×r(cid:21)
Note that Mf = A3(K) if r = 0. To sum up, we have proved that there
exists an integer r ∈ [[0, 3]], a linear map f : A3(K) → Mr(K) and a compatible
pair of bases in which S is represented by the matrix space Mf` M3,p−3−r(K).
However, we know from Theorem 1.3 that every range-compatible homomor-
phism on M3,p−3−r(K) is local. Thus, the splitting lemma shows that in order
to conclude, it remains to prove the following lemma.
Lemma 4.2. Let r ∈ [[0, 3]] and f : A3(K) → Mr(K) be a linear map. Then,
every range-compatible homomorphism on Mf is local.
To prove Lemma 4.2, we start with the case r = 0.
Lemma 4.3. Every range-compatible homomorphism on A3(K) is local.
Proof of Lemma 4.3. Let F : A3(K) → K3 be a range-compatible homomor-
phism. By Corollary 2.2, there are group endomorphisms u, g, h, i, j, k of K such
that
0 −a
a
−b
b
0 −c
c
0
F :
7→
−u(a) + g(b)
h(a) − i(c)
−j(b) + k(c)
.
For all (a, b, c) ∈ K3r{(0, 0, 0)}, the range of the alternating matrix
is the orthogonal complement of its kernel, which is spanned by
to
0 −a
a
−b
b
0 −c
c
0
c
b
a
. This leads
∀(a, b, c) ∈ K3, −u(a)c + g(b)c + h(a)b − i(c)b − j(b)a + k(c)a = 0.
(1)
Fixing b = 0 and c = 1 and varying a, we obtain that u is linear. Similarly, one
shows that g, h, i, j, k are all linear. Then, (1) reads
∀(a, b, c) ∈ K3, (cid:0)−u(1) + k(1)(cid:1)ac +(cid:0)g(1) − i(1)(cid:1)bc +(cid:0)h(1) − j(1)(cid:1)ab = 0.
25
On the left-hand side of this identity, we have a polynomial of degree at most 1 in
each variable, and hence all its coefficients are zero. This leads to u(1) = k(1),
g(1) = i(1) and h(1) = j(1). Therefore, F is the local map A 7→ AX where
X :=
h(1)
u(1)
g(1)
.
Now, we prove Lemma 4.2.
Proof of Lemma 4.2. The case r = 0 has already been dealt with in Lemma 4.3.
Assume now that r ≥ 1. Let F : Mf → K3 be a range-compatible homo-
morphism.
Denote by T the subspace of Mf consisting of its matrices with all first three
columns equal to 0. Then, we see that T = {0}` N for some subspace N
of M3,r(K) with codimension 1. By the splitting lemma and Theorem 1.3, the
restriction of F to T must be local. Thus, subtracting a well-chosen local map
from F , no generality is lost in assuming that F vanishes everywhere on T ,
a condition that will be assumed to hold throughout the rest of the proof. It
follows that there is a homomorphism
G : A3(K) → K3
such that, for all M ∈ Mf , we have F (M ) = G(K(M )), where
M =(cid:2)K(M )
[?]3×r(cid:3) with K(M ) ∈ A3(K).
We shall prove that G is range-compatible. To do so, we distinguish between
three cases.
Case a: r = 3.
Let A ∈ A3(K) r {0}. We choose a non-zero column C of A. The subspace
{CX T X ∈ K3} of M3(K) is not included in sl3(K), and hence we can find a
rank 1 matrix B ∈ M3(K) r sl3(K) whose range equals KC. Then, M3(K) =
KB ⊕ sl3(K), and hence we can write f (A) = λ B + N for some λ ∈ K and
some N ∈ sl3(K). Thus, Mf contains M :=(cid:2)A λB(cid:3), whose range equals Im A.
Then, G(A) = F (M ) ∈ Im M , and hence G(A) ∈ Im A.
Case b: r = 2.
Let A ∈ A3(K) r {0}. We prove that there exists a matrix M ∈ Mf such
26
G(A) ∈ Im A.
that M =(cid:2)A [?]3×2(cid:3) and Im M = Im A, which, as in Case a, will entail that
One of the columns of A, say the i-th, can be written as (cid:20)Y
?(cid:21) with Y ∈
find a row matrix L ∈ M1,2(K) such that(cid:20)λ B
L (cid:21) has all its columns collinear to
K2 r {0}. As in Case a, we find B ∈ M2(K) r sl2(K) with range KY . Then,
f (A) = λ B + N for some λ ∈ K and some N ∈ sl2(K). It follows that we can
the i-th column of A. Then, the matrix
(cid:2)A [0]3×2(cid:3) +(cid:20)[0]2×3 λ B
L (cid:21)
[0]1×3
belongs to Mf and its range equals Im A, which proves our claim.
Case c: r = 1.
Using Corollary 2.2, we find endomorphisms u, g, h, i, j, k, θ of (K, +) such that
−u(a) + g(b) + θ(c)
0 −a
a
−b
?
0 −c d
e
c
F :
.
M ∈ Mf of the form M = (cid:2)A [?]3×1(cid:3) such that Im M = Im A:
2-dimensional affine subspaces
7→
+ ({0} × K2) and Im A of K3 are non-
Let A ∈ A3(K) be with non-zero first row. Again, we see that there exists some
indeed, the
parallel, and hence they have a common point. Thus, G(A) = F (M ) ∈ Im A.
h(a) − i(c)
−j(b) + k(c)
b
0
f (A)
0
0
As in the proof of Lemma 4.3, this yields
∀(a, b, c) ∈ K3, (a, b) 6= (0, 0) ⇒ −u(a)c+g(b)c+h(a)b−i(c)b−j(b)a+k(c)a+θ(c)c = 0.
Next, we prove that θ = 0. We distinguish between two cases:
• Assume that #K > 2. Let c ∈ Kr{0}. Taking b = 0, we find that −u(a)c+
k(c)a + θ(c)c = 0 for all a ∈ K r {0}. Thus, the group endomorphism
χ : a ∈ K 7→ −u(a)c + k(c)a is constant on K r {0} with sole value −θ(c)c.
However, since K has more than 2 elements, we can find x and y in K r {0}
such that x + y 6= 0, which leads to χ(x) + χ(y) = χ(x + y), and hence
θ(c)c = 0. Therefore, θ(c) = 0.
27
• Assume that #K = 2. Then, u, g, h, i, j, k, θ are all linear, and hence the
quadratic form
(a, b, c) 7→ (−u(1) + k(1))ac + (g(1) − i(1))bc + (h(1) − j(1))ab + θ(1)c2 = 0
vanishes everywhere on K3 r span(cid:0)(0, 0, 1)(cid:1). By Lemma 5.2 of [5], q = 0,
and in particular θ(1) = 0.
Therefore, in any case θ = 0. Moreover, if (a, b) = (0, 0) it is obvious that
−u(a)c + g(b)c + h(a)b − i(c)b − j(b)a + k(c)a = 0.
Hence,
∀(a, b, c) ∈ K3, −u(a)c + g(b)c + h(a)b − i(c)b − j(b)a + k(c)a = 0,
which shows that G is range-compatible (see the proof of Lemma 4.3).
Thus, in any case we have shown that G is range-compatible. Lemma 4.3
then yields a vector Y ∈ K3 such that G : A 7→ AY , and we conclude that F is
the local map M 7→ M X for X :=(cid:20) Y
[0]r×1(cid:21).
This completes the proof in the case when dim V = 3.
4.3 The case dim V > 3
Here, we perform the inductive step of the proof of Theorem 4.1. Set n := dim V ,
p := dim U and assume that dim V > 3.
Step 1. There exist non-collinear forms f ⋆
n − 3 and codim(S mod f ⋆
i )r ≤ n − 4 for all i ∈ {1, 2}.
1 and f ⋆
2 in V ⋆ such that codim(S mod f ⋆
i ) ≤
Proof. Let f ⋆ ∈ V ⋆ r {0}. Denote by S f ⋆
the subspace of all the operators
s ∈ Sr such that (x, y) 7→ s(x)[y] vanishes everywhere on Ker f ⋆ × Ker f ⋆. Then,
by the rank theorem, we have
r
codim(S mod f ⋆)r = codim Sr +(cid:0)dim S f ⋆
r − dim V + 1(cid:1).
Thus, in order to have
codim(S mod f ⋆)r ≤ n − 4,
28
i
1 , . . . , f ⋆
n−1 in V ⋆ such that dim S f ⋆
it suffices that dim S f ⋆
r < n−1. Assume that there are n−1 linearly independent
vectors f ⋆
r = dim V − 1 for all i ∈ [[1, n]].
n) of V ⋆, whose
Then, we extend this family into a basis B⋆ = (f ⋆
corresponding basis of V is denoted by B. Denote by M the matrix space
associated with Sr in the bases B and B⋆. The assumptions show that, for
all (i, j) ∈ [[1, n]]2 such that i > j, the space M contains Ei,j − Ej,i (since
f ⋆
j
dim S
r = dim V − 1). By linearly combining those matrices, we find that
M = An(K), which means that Sr = La(V, V ⋆). In turn, this shows that for all
f ⋆ ∈ V ⋆ r {0}, we have (S mod f ⋆)r = La(cid:0)Ker f ⋆, (Ker f ⋆)⋆(cid:1) and hence
1 , . . . , f ⋆
n−1, f ⋆
codim(S mod f ⋆)r = 0 ≤ n − 4.
Thus, in any case, we have found a linear subspace P of codimension 2 in V ⋆
such that
∀f ⋆ ∈ V ⋆ r P, codim(S mod f ⋆)r ≤ n − 4.
Let f ∈ V ⋆ r {0}. If we denote by S f ⋆
m the space of all linear maps s ∈ Sm
such that Im s ⊂ Kf ⋆, the rank theorem yields
codim(S mod f ⋆) = codim S +(cid:0)dim(S f ⋆
m ) − dim(U/V )(cid:1).
Indeed, the rank of an alternating linear map from V to V ⋆ cannot equal 1, and
hence if s ∈ S satisfies Im s ⊂ Kf ⋆ then s vanishes everywhere on V , to the
effect that s ∈ Sm.
Thus, in order to have
codim(S mod f ⋆) ≤ n − 3,
it suffices that dim S f ⋆
V ⋆ such that dim S f ⋆
Sm = L(U/V, V ⋆), and hence
i
m < dim(U/V ).
n) of
m = dim(U/V ) for all i ∈ [[1, n]], then it would follow that
If there existed a basis (f ⋆
1 , . . . , f ⋆
codim S = codim Sr ≤ n − 3.
In that case, we would obtain
codim(S mod f ⋆) ≤ codim S ≤ n − 3
whatever the choice of f ⋆ in V ⋆ r {0}.
29
This yields a linear hyperplane H of V ⋆ such that
∀f ⋆ ∈ V ⋆ r H, codim(S mod f ⋆) ≤ n − 3.
Hence, for every f ⋆ ∈ V ⋆ r (P ∪ H), the space S mod f ⋆ satisfies the conditions
of Theorem 4.1. However, the union of two proper linear subspaces of V ⋆ is
always a proper subset of V ⋆: this successively yields some f ⋆
1 ∈ V r (P ∪ H),
and then we can pick f ⋆
2 are linearly
independent and both spaces S mod f ⋆
2 satisfy the requirements of
Theorem 4.1.
2 in V r ((P + Kf ⋆
1 ) ∪ H), so that f ⋆
1 and S mod f ⋆
1 and f ⋆
By induction, we obtain that both range-compatible homomorphisms F mod f ⋆
1
and F mod f ⋆
2 are local, yielding vectors x1 and x2 in U such that
∀s ∈ S, F (s) = s(x1) mod Kf ⋆
1
and F (s) = s(x2) mod Kf ⋆
2 .
Replacing F with s 7→ F (s) − s(x1), we reduce the situation further to the point
where there exists x ∈ U such that
∀s ∈ S, F (s) ∈ Kf ⋆
1
and F (s) = s(x) mod Kf ⋆
2 .
(2)
Notice then that
∀s ∈ S,
s(x) ∈ span(f ⋆
1 , f ⋆
2 ).
Step 2. The vector x belongs to V .
2 ) into a basis B⋆
Proof. Assume on the contrary that x 6∈ V . Let us extend (f ⋆
of V ⋆, whose predual basis of V is denoted by B = (e1, . . . , en), and let us extend
1 , f ⋆
(e1, . . . , en, x) into a basis eB of U . Denote by M the matrix space representing S
ineB and B⋆. Then, if we denote by P the subspace K2×{0} of Kn, we see that M
is a linear subspace of An(K)` P` Mn,p−n−1(K), where p := dim U . However,
that subspace has codimension n − 2 in An(K)` Mn,p−n(K), whereas M has
codimension at most n − 2 in it. Therefore, M = An(K)` P` Mn,p−n−1(K).
[0](n−1)×1(cid:21). Then, we have a range-compatible
matrix M = (mi,j) of M to (cid:20) m1,n+1
Denote by G : M → Kn the range-compatible homomorphism that is as-
sociated with F in the above bases. Condition (2) shows that G maps every
homomorphism H : P → Kn such that, for all Y ∈ P ,
G(cid:16)(cid:2)[0]n×n Y [0]n×(p−n−1)(cid:3)(cid:17) = H(Y ),
30
and we see from the above form of G that Im H = K×{0}. However, by Theorem
1.3 the map H should be a scalar multiple of the identity, which is false since P
has dimension 2.
Step 3. The vector x belongs to Ker f ⋆
1 ∩ Ker f ⋆
2 .
1, g⋆
1 , f ⋆
2) of span(f ⋆
Proof. Assume on the contrary that x 6∈ o span(f ⋆
2 ). Then, we can find a
basis (g⋆
2(x) = 0. We extend
this basis into a basis B⋆ = (g⋆
i (x) = 0 for all
i ∈ [[2, n]], and we consider the corresponding basis B of V . Obviously, x is the
first vector of B. It then turns out that if we denote by Mr the matrix space
associated with Sr in the bases B and B⋆, this space is included in An(K) and
1(x) = 1 and g⋆
n) of V ⋆ in which g⋆
2 ) such that g⋆
2, . . . , g⋆
1 , f ⋆
1, g⋆
every matrix in Mr has its first column of the form
0
?
[0](n−2)×1
. Thus, we see
that Mr has codimension at least n − 2 in An(K), contradicting the assumption
that codim Sr ≤ n − 3.
Step 4. One has x = 0.
Proof. Assume on the contrary that x 6= 0. Let us consider a linear form f ⋆
0 on
2 ) into a basis B∗ in which
V such that f0(x) = 1. Then, we extend (f ⋆
the last n − 3 linear forms annihilate x, and we denote by B the corresponding
0 , f ⋆
1 , f ⋆
associated with S in those bases. For M ∈ M we write
basis of V , which we extend into a basis eB of U . Note that x is the first vector
of B, and hence also the first one of eB. Let us denote by M the matrix space
M =(cid:2)A(M ) B(M )(cid:3) with A(M ) ∈ An(K) and B(M ) ∈ Mn,p−n(K).
Then, we know that the first column of every matrix of M has the form
vector
[0](n−3)×1
Denote by G : M → Kn the range-compatible homomorphism that corre-
sponds to F in the above bases. Then, G maps every matrix (mi,j) of M to the
Denote by N the linear subspace of An(K) consisting of its matrices (mi,j)
such that mi,1 = 0 for all i ∈ [[4, n]]. Then, codim N = n − 3. However,
.
[0](n−2)×1
m2,1
0
?
?
0
.
31
codim A(M) = codim Sr ≤ n − 3, whence A(M) = N and codim Sr = n − 3. It
follows that we can find a linear mapping
K : A3(K) −→ M3,p−n(K)
such that, for all C ∈ A3(K), the space M contains a matrix of the form
C
[0]3×(n−3)
[0](n−3)×3
[0](n−3)×(n−3)
K(C)
[?](n−3)×(p−n)(cid:21) .
Now, denote by H the subspace of all matrices M ∈ M such that A(M ) = 0.
Then codim B(H) + codim Sr = codim S, whence codim B(H) ≤ 1. Thus, B(H)
includes a linear hyperplane of Mn,p−n(K).
For M ∈ H, denote by B′(M ) the sub-matrix of B(M ) obtained by deleting
all the rows starting from the fourth one, so that B′(H) is a linear subspace
of M3,p−n(K) with codimension at most 1. Thus, for all C ∈ A3(K) and N ′ ∈
B′(H), the space M contains a matrix of the form
C
[0]3×(n−3)
[0](n−3)×3
[0](n−3)×(n−3)
K(C) + N ′
[?](n−3)×(p−n)(cid:21) .
Let us finally denote by M′ the space of all matrices
(cid:2)C K(C) + N ′(cid:3) with C ∈ A3(K) and N ′ ∈ B′(H).
Applying Lemma 2.1 to the subspace V0 = {0} × Kn−3 of Kn, we obtain that
(cid:20)
(cid:20)
G′ : (mi,j) ∈ M′ 7→
m2,1
0
0
is a range-compatible homomorphism. On the other hand, M′ represents a space
of alternating operators that satisfies the assumptions of Theorem 4.1. From
the 3-dimensional case treated in Section 4.2, we deduce that G′ is a local map.
This would yield a vector X ∈ K3+p−n such that dim M′X = 1. However, if
X ∈ K3 × {0} we would have dim M′X = 2 or dim M′X = 0; on the other hand,
if X 6∈ (K3 × {0}), we see that dim M′X ≥ 2 because B′(H) has codimension
at most 1 in M3,p−n(K). Thus, we have a contradiction in any case, and we
conclude that x = 0.
We conclude that, for all s ∈ S, the vector F (s) belongs to both Kf ⋆
1 and
Kf ⋆
2 , and hence it equals 0. Thus, F is the local map s 7→ s(0), which completes
the inductive step. Thus, Theorem 4.1 is now established, and Theorem 1.7
follows as it is a reformulation of the special case when U = V .
32
References
[1] M. D. Atkinson, S. Lloyd, Large spaces of matrices of bounded rank. Quart.
J. Math. Oxford (2) 31 (1980), 253 -- 262.
[2] J. Dieudonn´e, Sur une g´en´eralisation du groupe orthogonal `a quatre vari-
ables. Arch. Math. 1 (1948), 282 -- 287.
[3] R. Loewy, Large spaces of symmetric matrices of bounded rank are decom-
posable. Linear Multilinear Algebra 48 (2001), 355 -- 382.
[4] R. Loewy, N. Radwan, Spaces of symmetric matrices of bounded rank. Linear
Algebra Appl. 197-198 (1994), 189 -- 215.
[5] C. de Seguins Pazzis, Range-compatible homomorphisms on matrix spaces.
Linear Algebra Appl. 484 (2015) 237 -- 289.
[6] C. de Seguins Pazzis, Range-compatible homomorphisms over the field with
two elements. Preprint, arXiv: http://arxiv.org/abs/1407.4077
[7] C. de Seguins Pazzis, Quasi-range-compatible affine maps on large operator
spaces. Linear Multilinear Algebra 64 (2016) 1056 -- 1085.
[8] C. de Seguins Pazzis, The classification of large spaces of matrices with
bounded rank. Israel J. Math. 208 (2015) 219 -- 259
[9] C. de Seguins Pazzis, The linear preservers of non-singularity in a large space
of matrices. Linear Algebra Appl. 436 (2012), 3507 -- 3530.
33
|
1004.1660 | 1 | 1004 | 2010-04-09T22:56:13 | Representation Theory of Finite Semigroups over Semirings | [
"math.RA",
"math.RT"
] | We develop the representation theory of a finite semigroup over an arbitrary commutative semiring with unit, in particular classifying the irreducible and minimal representations. The results for an arbitrary semiring are as good as the results for a field. Special attention is paid to the boolean semiring, where we also characterize the simple representations and introduce the beginnings of a character theory. | math.RA | math |
REPRESENTATION THEORY OF FINITE SEMIGROUPS
OVER SEMIRINGS
ZUR IZHAKIAN, JOHN RHODES, AND BENJAMIN STEINBERG
Abstract. We develop the representation theory of a finite semigroup
over an arbitrary commutative semiring with unit, in particular clas-
sifying the irreducible and minimal representations. The results for an
arbitrary semiring are as good as the results for a field. Special attention
is paid to the boolean semiring, where we also characterize the simple
representations and introduce the beginnings of a character theory.
1. Introduction
This paper is about the representation theory of finite semigroups over
commutative semirings with unit. See [33] for a modern presentation of finite
semigroup theory, including a theory of semirings influenced by a semigroup
perspective.
The importance of semirings in mathematics and theoretical computer
science was first recognized by Schutzenberger [38] in his theory of weighted
automata and rational power series [3, 36]. Conway also heavily employed
semirings in his approach to automata theory [9]. See [26 -- 29] for further ap-
plications of semirings to theoretical computer science. Recently, idempotent
semirings have entered into mainstream mathematics because they are at the
heart of the relatively new subject of tropical geometry [10, 12, 17, 21, 35].
They also play a role in the notion of characteristic one being developed by
Connes et al. [7, 8].
Before turning to semirings, we discuss the classical case of fields. Work
of Clifford [4, 5], Munn [24, 25] and Ponizovskiı [30] parameterized the ir-
reducible representations of a finite semigroup over a field in terms of the
irreducible representations of its maximal subgroups. The reader is referred
to [6, Chapter 5] for a full account of this work. Explicit constructions of the
irreducible representations were later obtained independently by Rhodes and
Zalcstein [34] (which was written in 1968) and by Lallement and Petrich [20]
in terms of the Schutzenberger representation by monomial matrices [37];
Date: August 4, 2018.
2010 Mathematics Subject Classification. Primary: 16G10, 20M30, 20M25; Secondary:
16Y60, 14T05.
Key words and phrases. Representations, semirings, semigroups.
The first author was supported in part by the ISF (Ref. No. 448/09).
The third author was supported in part by NSERC.
1
2
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
see also [23]. All of these approaches make use of Rees's Theorem [31] char-
acterizing 0-simple semigroups up to isomorphism and Wedderburn theory.
This viewpoint makes it hard to see how to generalize things to rings, let
alone semirings. So it was once believed that we knew everything about
representations over the field of complex numbers and nothing about repre-
sentations over, say, the boolean semiring.
The purpose of this paper is to show that we essentially know just as much
about boolean representations as we do about complex ones. In particular,
we construct all irreducible representations of a finite semigroup S over any
commutative semiring with unit modulo the case of groups. In the case of the
boolean semiring we identify the congruence associated to the direct sum of
all irreducible representations with one of the congruences from the semilocal
theory [33, Chapter 4], generalizing the case of representations over fields [1,
32]. In addition to classifying irreducible representations, we also describe
all the minimal representations. For the case of boolean representations we
are also able to classify the simple representations by showing that they are
dual to the minimal ones. In the boolean setting, we propose a notion of
characters.
Here we follow the coordinate-free approach from [11], which was used to
handle arbitrary commutative rings, in order to deal with semirings. Basic
facts about semirings can be found in [13, 16, 18] or [33, Chapters 8 and 9].
2. Preliminaries
We collect here some basic definitions and facts concerning finite semi-
groups that can be found in any of [6, 19, 33]. Let S be a (fixed) finite
semigroup. If e is an idempotent, then eSe is a monoid with identity e; its
group of units Ge is called the maximal subgroup of S at e. Two idempotents
e, f are said to be isomorphic if there exist x ∈ eSf and x′ ∈ f Se such that
xx′ = e, x′x = f . In this case one can show that eSe is isomorphic to f Sf
as monoids and hence Ge ∼= Gf ; see Fact 2.1.
If s ∈ S, then J(s) = S1sS1 is the principal (two-sided) ideal generated
by s (here S1 means S with an adjoined identity). Following Green [14], two
elements s, t of a semigroup S are said to be J -equivalent if J(s) = J(t).
In this case one writes s J t. In fact, there is a preorder on S given by
s ≤J t if J(s) ⊆ J(t). This preorder induces an ordering on J -classes in
the usual way.
Fact 2.1. In a finite semigroup, idempotents e, f are isomorphic if and only
if e J f , that is, SeS = Sf S.
An element s of a semigroup S is said to be (von Neumann) regular if
s = sts for some t ∈ S. Each idempotent is of course regular.
Fact 2.2. Let S be a finite semigroup and J a J -class of S. Then the
following are equivalent:
(1) J contains an idempotent;
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
3
(2) J contains a regular element;
(3) all elements of J are regular;
(4) J 2 ∩ J 6= ∅.
A J -class satisfying the equivalent conditions in Fact 2.2 is called a
regular J -class. The poset of regular J -classes is denoted U (S). The
following standard result from finite semigroup theory will play a role in
this paper.
Fact 2.3. Let S be a finite semigroup and J a regular J -class. Let e ∈ J
be an idempotent. Then eSe ∩ J = Ge.
Let J be a J -class of S. Set J 6 ↑ = {s ∈ S J * J(s)}; it is the ideal of
all elements of S that are not J -above some (equals any) element of J.
The reader is referred to [33, Chapter 9] for basic notions concerning
modules1 over semirings. Let us say that a congruence on a module over a
semiring is proper if it has more than one equivalence class and that it is
trivial if the associated partition is into singletons. Fix a base commutative
semiring ring k with unit.
Definition 2.4. If A is a k-algebra, not necessarily unital, then a right
A-module M with M A 6= 0 is said to be:
(1) simple if it has only trivial quotient modules;
(2) minimal if it contains no proper non-zero submodules;
(3) irreducible if it is both simple and minimal.
Over a ring k, all three notions coincide. Observe that minimality amounts
to asking that, for all 0 6= m ∈ M , the cyclic module mA is M .
The category of (right) A-modules will be denoted mod-A. We adopt here
the convention that if A is unital, then by mod-A we mean the category of
unitary A-modules. The reader should verify that all functors considered in
this paper respect this convention.
Definition 2.5 (Radical and socle). Let us denote by rad(M ) the intersec-
tion of all maximal congruences on a module M ; it is called the radical of
M . The submodule of M generated by all minimal submodules is denoted
Soc(M ) and is called the socle of M .
We state and prove Schur's lemma in the semiring context.
Lemma 2.6 (Schur). Let M and N be A-modules. Then each non-zero
homomorphism ϕ : M → N is:
(1) injective if M is simple;
(2) surjective if N is minimal;
(3) an isomorphism if M is simple and N is minimal.
Proof. Suppose that ϕ 6= 0. Then ϕ(M ) is a non-zero submodule of N and
ker ϕ is a proper congruence on M , from which the result is trivial.
(cid:3)
1We note that some authors use the term semimodule for what we call a module.
4
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
The class of minimal modules is closed under quotients.
Proposition 2.7. Let M be a minimal A-module. Then any non-zero quo-
tient module of M is also minimal.
Proof. Suppose that M is minimal and ϕ : M → N is a surjective A-module
homomorphism. Let N ′ be a non-zero submodule of N . Then ϕ−1(N ) is a
non-zero submodule of M and so M = ϕ−1(N ), whence N = N ′. Thus N
is minimal.
(cid:3)
3. Construction of the irreducible and minimal modules
Fix a finite semigroup S and a commutative semiring with unit k. The
If M is a kS-module, then
semigroup algebra kS need not be unital.
AnnS(M ) = {s ∈ S M s = 0} is an ideal of S. The following definition,
due to Munn [25], is crucial to semigroup representation theory.
Definition 3.1 (Apex). A regular J -class J is said to be the apex of a
kS-module M if AnnS(M ) = J 6 ↑.
It is easy to see that M has apex J if and only if J is the unique ≤J -
minimal J -class that does not annihilate M . The notion of apex is closely
related to that of lifting J -classes [33, Lemma 4.6.10]. For further back-
ground see [33, Chapter 4.6].
Fix an idempotent transversal E = {eJ J ∈ U (S)} to the set U (S)
of regular J -classes and set GJ = GeJ . Let AJ = kS/kJ 6 ↑. Notice that
the category of kS-modules with apex J can be identified with the full
subcategory of mod-AJ whose objects are modules M such that M eJ 6= 0.
Our first goal is to show that every minimal module has an apex. This
result is due independently to Munn and Ponizovskiı [24, 25, 30] in the case
of fields.
Theorem 3.2. Let M be a minimal kS-module. Then M has an apex.
Proof. Because M kS 6= 0, there is a ≤J -minimal J -class J such that
J * AnnS(M ). Let I = S1JS1; of course, I is an ideal of S. Since I \ J
annihilates M by minimality of J, it follows 0 6= M kJ = M kI. From the
fact that I is an ideal of S, we may deduce that M kI is a kS-submodule
and so by minimality
M = M kI = M kJ.
(3.1)
Therefore, since JJ 6 ↑ ⊆ I \ J ⊆ AnnS(M ), it follows from (3.1) that J 6 ↑ =
AnnS(M ). Now if J is not regular, then Fact 2.2 implies J 2 ⊆ I \ J and
hence J annihilates M by (3.1), a contradiction. Thus J is regular and is
an apex for M .
(cid:3)
Now we wish to establish a bijection between irreducible kS-modules with
apex J and irreducible kGJ -modules. This relies on a semiring analogue of
a well-known result of Green [15, Chapter 6]. Let A be a k-algebra and e
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
5
an idempotent of A. Then eA is an eAe-A-bimodule and Ae is an A-eAe-
bimodule. Hence we have a restriction functor Res : mod-A → mod-eAe and
induction/coinduction functors Ind, Coind : mod-eAe → mod-A given by
Ind(M ) = M ⊗eAe eA, Res(M ) = M e and Coind(M ) = HomeAe(Ae, M ).
Moreover, Ind is right exact, Res is exact, Coind is left exact and Ind and
Coind are the left and right adjoints of Res, respectively (where a functor is
left exact if it preserves finite limits and right exact if it preserves finite col-
imits). This follows from observing that HomA(eA, M ) = M e = M ⊗A Ae
and the usual adjunction between hom and the tensor product. Further-
more, it is well known that the unit of the first adjunction gives a natural
isomorphism M ∼= Ind(M )e while the counit of the second gives a natural
isomorphism Coind(M )e ∼= M .
We need the following lemma relating congruences on M and M e.
Proposition 3.3. Let M be an A-module and let e ∈ A be an idempotent.
Then every congruence R on M e extends to a congruence R′ on M by setting
m R′ m′ if and only if mae R m′ae for all a ∈ A. Moreover, R′ is the largest
congruence whose restriction to M e is contained in R.
Proof. Clearly R′ is k-module congruence. To see that it is an A-module
congruence, suppose b ∈ A and m R′ m′. Then for any a ∈ A, we have
m(bae) R m′(bae) and so mb R′ m′b as required. Finally, observe that if
m, m′ ∈ M e and m R′ m′, then m = mee R m′ee = m′. Conversely, if
m, m′ ∈ M e are R-related and a ∈ A, then mae = meae R m′eae = m′ae
and so m R′ m′. Thus R′ extends R.
Finally, suppose that R′′ is some other congruence on M whose restriction
to M e is contained in R and suppose m R′′ m′. Then if a ∈ A one has
mae R m′ae and so m R′ m′ completing the proof.
(cid:3)
Let M be an A-module and define
N (M ) = {(m, m′) mae = m′ae, ∀a ∈ A}
and L(M ) = M eA. It is easily verified using Proposition 3.3 that N (M )
is the largest congruence on M whose restriction to M e is trivial. On the
other hand, L(M ) is the smallest A-submodule of M with L(M )e = M e.
The congruence class of 0 under N (M ) is
K(M ) = {m ∈ M mae = 0, ∀a ∈ A}.
It is also the largest submodule of M annihilated by e. The constructions
K, L, N are functorial.
Observe that if V is an eAe-module, then
L(Ind(V )) = Ind(V )eA = V ⊗eAe eAeA = V ⊗eAe eA = Ind(V ).
(3.2)
Alternatively, one can observe that Ind(V )eA satisfies the same universal
property as Ind(V ).
The analogue of our next result for rings can be found in [15, Chapter
6.2]. The proof for semirings is essentially a modification. We remind the
6
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
reader that if M is an A-module and N is a submodule, then M/N is the
quotient of M by the congruence given by m ≡ m′ mod N if and only if
there exist n, n′ ∈ N with m + n = m′ + n′; it is the largest congruence
identifying all elements of N with 0, cf. [33, Chapter 9].
Lemma 3.4. Let A be a k-algebra and e an idempotent of A.
(1) If M is an irreducible (minimal, simple) A-module, then M e = 0 or
M e is an irreducible (minimal, simple) eAe-module.
(2) If V is a minimal eAe-module, then the unique maximal A-submodule
of Ind(V ) is K(Ind(V )). Moreover, the minimal A-modules M with
M e ∼= V are (up to isomorphism) the quotients of Ind(V ) by con-
gruences in between congruence modulo K(Ind(V )) and N (Ind(V )).
(3) If V is a simple eAe-module, then Ind(V ) has unique maximal con-
gruence N (Ind(V )). Moreover, if V is irreducible, then the quo-
tient Ind(V )/N (Ind(V )) is the unique irreducible A-module M with
M e ∼= V .
(4) If V is a minimal eAe-module, then Coind(V ) has unique minimal
A-submodule L(Coind(V )). Furthermore, if V is irreducible, then
L(Coind(V )) is the unique irreducible A-module M with M e ∼= V .
Consequently, restriction yields a bijection between irreducible A-modules
that are not annihilated by e and irreducible eAe-modules.
Proof. To prove (1), assume M e 6= 0. Suppose first that M is minimal
and let m ∈ M e be non-zero. Then meA = mA = M , so meAe = M e.
Thus M e is minimal. Next suppose that M is simple and let R be a proper
congruence on M e. Then R′ from Proposition 3.3 is a proper congruence
on M and hence trivial. But then R, which is the restriction of R′ to M e,
is trivial. Thus M e is simple. The irreducible case follows by combining the
minimal and simple cases.
Next we turn to (2). Suppose that L is a proper submodule of Ind(V ).
Then Le is an eAe-submodule of Ind(V )e ∼= V . By minimality of V , either
Le = {0}, and so L ⊆ K(Ind(V )), or Le = Ind(V )e.
In the latter case
we have L ⊇ LeA = Ind(V )eA = Ind(V ), where the last equality uses
(3.2). Thus K(Ind(V )) is the unique maximal proper submodule. It follows
that Ind(V )/K(Ind(V )) is minimal. Next observe that since N (Ind(V ))
has trivial restriction to Ind(V )e, Proposition 2.7 now implies that, for any
congruence R between congruence modulo K(Ind(V )) and N (Ind(V )), one
has that Ind(V )/R is minimal and [Ind(V )/R]e ∼= V . Finally, suppose that
M is a minimal A-module with M e ∼= V . Then the adjunction yields a
non-zero homomorphism ϕ : Ind(V ) → M which is surjective by minimality
and restricts to an isomorphism of eAe-modules from Ind(V )e → M e. In
particular, it follows that the congruence ker ϕ is injective on Ind(V )e and
hence contained in N (Ind(V )). Also, as K(Ind(V ))e = 0 and M e 6= 0,
minimality implies ϕ(K(Ind(V ))) = 0. This establishes (2).
Now we turn to (3). Let V be a simple eAe-module.
It is immediate
from (3.2) that any proper congruence on Ind(V ) must restrict to a proper
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
7
congruence on Ind(V )e ∼= V . But V is simple and so the only proper con-
gruence on V is the trivial one. Proposition 3.3 now implies that N (Ind(V ))
is the unique maximal congruence on Ind(V ). Assume furthermore that V
is irreducible. Then Ind(V )/N (Ind(V )) is minimal, and hence irreducible,
and [Ind(V )/N (Ind(V ))]e ∼= V by (2). Uniqueness follows also from (2),
since if M is irreducible with M e ∼= V , then in particular M is minimal and
so isomorphic to Ind(V )/R for a certain congruence R ⊆ N (Ind(V )). But
simplicity of M implies that R = N (Ind(V )). This completes the proof of
(3).
Finally, to prove (4) first observe that if M is any non-zero A-submodule
of Coind(V ), then M e 6= 0. Indeed, suppose M e = 0 and let ϕ ∈ M . Then,
for any x in Ae, we have ϕ(x) = (ϕxe)(e) = 0 since ϕxe ∈ M e = 0. It follows
that M = 0. Since Coind(V )e ∼= V is a minimal eAe-module, it now follows
that if M is a non-zero A-submodule of Coind(V ), then M e = Coind(V )e
and hence
L(Coind(V )) = Coind(V )eA ⊆ M eA ⊆ M.
This establishes that L(Coind(V )) is the unique minimal A-submodule.
Suppose now that V is irreducible. We must show that any proper congru-
ence on L(Coind(V )) is trivial. Since Coind(V )e ∼= V is irreducible, it fol-
lows that any proper congruence on L(Coind(V )) is trivial on L(Coind(V ))e
and so by Proposition 3.3 it suffices to show that N (L(Coind(V ))) is the
trivial congruence. So suppose ϕ, ψ ∈ L(Coind(V )) are equivalent and let
a ∈ A. Then
ϕ(ae) = (ϕae)(e) = (ψae)(e) = ψ(ae)
Since L(Coind(V ))e = Coind(V )eAe = Coind(V )e ∼= V ,
and hence ϕ = ψ. This completes the proof L(Coind(V)) that is irreducible.
it just re-
mains to prove uniqueness. Suppose M is an irreducible A-module with
M e ∼= V . Then the existence of a non-zero element of HomeAe(M e, V ) ∼=
HomA(M, Coind(V )) implies that M admits a non-zero homomorphism to
Coind(V ). Hence M is isomorphic to an irreducible (and hence minimal) A-
submodule of Coind(V ). But L(Coind(V )) is the unique minimal submodule
of Coind(V ) and so M ∼= L(Coind(V )), as required.
(cid:3)
We may now etablish an analogue of the Clifford-Munn-Ponizovskiı the-
orem for semirings.
Theorem 3.5. Let S be a finite semigroup, k a commutative semiring with
unit and E = {eJ J ∈ U (S)} be an idempotent transversal to the set U (S)
of regular J -classes of S. Let GJ be the maximal subgroup GeJ . Define
functors IndJ , CoindJ : mod-kGJ → mod-kS by
IndJ (V ) = V ⊗kGJ eJ (kS/kJ 6 ↑)
CoindJ (V ) = HomkGJ ((kS/kJ 6 ↑)eJ , V ).
Then:
8
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
(1) If M is an irreducible (minimal) kS-module with apex J, then M eJ
is an irreducible (minimal) kGJ -module;
(2) If V is an irreducible kGJ -module, then
N = {(u, v) ∈ IndJ (V ) uae = vae, ∀a ∈ kS/kJ 6 ↑},
is the unique maximal congruence on IndJ (V ) and IndJ (V )/N is the
unique irreducible kS-module M with apex J such that M eJ ∼= V ;
(3) If V is an irreducible kGJ -module, then the unique minimal kS-
submodule of CoindJ (V ) is CoindJ (V )eJ kS, which moreover is the
unique irreducible kS-module M with apex J such that M eJ ∼= V .
(4) If V is a minimal kGJ -module, then
K = {m ∈ IndJ (V ) mae = 0, ∀a ∈ kS/kJ 6 ↑}
is the unique maximal kS-submodule of IndJ (V ). Moreover, the min-
imal kS-modules M with apex J such that M eJ ∼= V are up to iso-
morphism the quotients IndJ (V )/R with (retaining the notation of
(2)) R ⊆ N and K contained in a single class of R.
Consequently, there is a bijection between the irreducible kS-modules and
the irreducible kGJ -modules, where J runs over U (S).
Proof. Theorem 3.2 implies that every minimal kS-module M has an apex.
Again setting AJ = kS/kJ 6 ↑ for a regular J -class J, we know that irre-
ducible kS-modules with apex J are in bijection with irreducible AJ -modules
M such that M eJ 6= 0. It follows directly from Fact 2.3 that
eJ AJ eJ = keJ SeJ /keJ J 6 ↑eJ = kGJ .
Lemma 3.4 then yields that irreducible AJ -modules not annihilated by eJ ,
that is, irreducible kS-modules with apex J, are in bijection with irreducible
kGJ -modules in the prescribed manner. Similarly, the minimal kS-modules
are as advertised by application of Lemma 3.4.
(cid:3)
3.1. A construction in coordinates. Let us relate the above construction
of the irreducible modules to the explicit ones found in [20,34] for fields. All
the facts about finite semigroups used in this discussion can be found in the
appendix of [33] or in [19]. According to Green [14], two elements s, t of a
semigroup are said to be R-equivalent if they generate the same principal
right ideal. Dually s, t are said to be L -equivalent if they generate the same
principal left ideal.
Once more let S be a finite semigroup, k a commutative semiring with
unit and E = {eJ J ∈ U (S)} an idempotent transversal to the set U (S)
of regular J -classes of S. Let GJ be the maximal subgroup GeJ . We use
LJ and RJ for the L - and R-classes of eJ , respectively.
Here we follow [22] to give a concrete description of the irreducible kS-
modules. If V is an irreducible kGJ -module, we use eV for the corresponding
irreducible kS-module. The reader should recall the definition of radical and
socle from Definition 2.5.
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
9
Proposition 3.6. Let V be an irreducible kGJ -module. Then there is a
natural isomorphism HomkS(IndJ (V ), CoindJ (V )) ∼= HomkGJ (V, V ) 6= 0.
Moreover, if ϕ ∈ HomkS(IndJ (V ), CoindJ (V )) is non-zero, then we have
ker ϕ = rad(IndJ (V )) and Im ϕ = eV = Soc(CoindJ (V )).
Proof. First note that since IndJ (V ), CoindJ (V ) are kS/kJ 6 ↑-modules, the
adjunction yields
HomkGJ (V, V ) = HomkGJ (IndJ (V )eJ , V ) ∼= HomkS(IndJ (V ), CoindJ (V )).
Suppose now that ϕ : IndJ (V ) → CoindJ (V ) is a non-zero homomorphism.
Because IndJ (V )eJ kS = IndJ (V ) by construction, it follows that
ϕ(IndJ (V )) = ϕ(IndJ (V ))eJ kS ⊆ CoindJ (V )eJ kS = eV .
Since ϕ 6= 0,
As IndJ (V ) has a unique maximal congruence, we conclude that ker ϕ =
rad(IndJ (V )).
(cid:3)
it follows by irreducibility of eV , that ϕ(IndJ (V )) = eV .
Observe that as k-modules, kLJ = (kS/kJ 6 ↑)eJ and kRJ = eJ kS/kJ 6 ↑
by stability. Moreover, the corresponding kGJ -kS-bimodule structure on
kRJ is induced by left multiplication by elements of GJ and by the right
Schutzenberger representation of S on RJ [6, 19, 33] (i.e., the action of S on
RJ by partial functions obtained via restriction of the regular action). To
simplify notation, we will use kRJ and kLJ for the rest of this section. Then
we have
IndJ (V ) = V ⊗kGJ kRJ
CoindJ (V ) = HomkGJ (kLJ , V ).
Multiplication in the semigroup induces a non-zero homomorphism
CJ : kRJ ⊗kS kLJ ∼= eJ kS/kJ 6 ↑ ⊗kS (kS/kJ 6 ↑)eJ → eJ (kS/kJ 6 ↑)eJ ∼= kGJ
which moreover is a map of kGJ -bimodules.
Let T ⊆ RJ be a complete set of representatives of the L -classes of J and
T ′ ⊆ LJ be a complete set of representatives of the R-classes of J. Then
GJ acts freely on the left of RJ and T is a transversal for the orbits and
dually T ′ is a transversal for the orbits of the free action of GJ on the right
of LJ , see [33, Appendix A]. Thus kRJ is a free left kGJ -module with basis
T and kLJ is a free right kGJ -module with basis T ′.
It is instructive to
verify that the associated matrix representation over kGJ of S on kRJ is
the classical right Schutzenberger representation by row monomial matrices
and the representation of S on kLJ is the left Schutzenberger representation
by column monomial matrices [6,19,33,34]. Hence if ℓJ = T and rJ = T ′,
then as kGJ -modules we have kRJ ∼= kGℓJ
J . Thus CJ is
the bilinear form given by the ℓJ × rJ -matrix (also denoted CJ ) with
J and kLJ ∼= kGrJ
(3.3)
(CJ )ba =(λbρa λbρa ∈ J
otherwise
0
10
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
where λb ∈ T represents the L -class b and ρa ∈ T ′ represents the R-class a.
Note that (CJ )ba ∈ GJ ∪{0} by stability and CJ is just the usual sandwich (or
structure) matrix of the J -class J coming from the Green-Rees structure
theory [6,19,33]. The reader may take (3.3) as the definition of the sandwich
matrix if he/she so desires.
Suppose now that V is a kGJ -module. We can consider the induced map
V ⊗ CJ : V ⊗kGJ kRJ ⊗kS kLJ → V ⊗kGJ kGJ ∼= V
which moreover is non-zero as v ⊗ eJ ⊗ eJ 7→ v. From the isomorphism
HomkGJ (V ⊗kGJ kRJ ⊗kS kLJ , V ) ∼= HomkS(V ⊗kGJ kRJ , HomkGJ (kLJ , V ))
we obtain a corresponding non-zero kS-linear map
V ⊗ CJ : IndJ (V ) → CoindJ (V )
(abusing notation).
Putting together the above discussion with Proposition 3.6, we obtain the
following result.
Theorem 3.7. Let V be an irreducible kGJ -module. Then the irreducible
kS-module corresponding to V is the image of the morphism
V ⊗ CJ : IndJ (V ) → CoindJ (V )
where CJ is the sandwich matrix for J, i.e., it is the k-span of the rows of
V ⊗ CJ .
Remark 3.8. Note that since kRJ and kLJ are free kGJ -modules with bases
T and T ′ respectively, as k-modules, we have IndJ (V ) = V ℓJ and IndJ (V ) =
V rJ ; in particular, these two functors are exact. Moreover, one can easily
compute that V ⊗ CJ is given via right multiplication by CJ where we view
elements of V ℓJ and V rJ as row vectors with entries in V .
A semigroup S is called generalized group mapping if it has a distinguished
(0-)minimal ideal I on which it acts faithfully on both the left and right [33,
Chapter 4]. The distinguished ideal is unique and regular. The following
result generalizes a result of Rhodes and Zalcstein for fields [34]. The original
proof uses Wedderburn theory, while our proof uses the description of the
irreducible modules. Recall that a module M is faithful for S if ms = mt
for all m ∈ M implies s = t.
Proposition 3.9. Let S be a finite semigroup and suppose that the irre-
ducible kS-module M is faithful. Then S is generalized group mapping with
distinguished ideal J or J ∪ {0} where J is the apex of M .
Proof. We just handle the case that S has a zero element 0, as the other
case is easier. Also, by Schur's lemma, 0 must act either as the identity (in
which case S is trivial) or as the zero endomorphism of M . Suppose that we
are in the latter case. By definition of an apex, it is clear that I = J ∪ {0} is
an ideal. Let e ∈ E(J) and put V = M e. Then M is a quotient of IndJ (V )
and a submodule of CoindJ (V ). It then follows that S acts faithfully on
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
11
IndJ (V ) and CoindJ (V ). Now if s, t act the same on the right of I, then
they will act the same on the right of Re and hence on IndJ (V ). Thus the
action of S on the right of I is faithful. Similarly, if s, t act the same on the
left of I, then they act the same on the left of Le and hence on CoindJ (V ).
This completes the proof that S is generalized group mapping.
(cid:3)
4. The case of idempotent semirings
It turns out that we could have restricted our attention to two cases
for irreducible modules: rings (already handled in [11]) and idempotent
semirings, as the following observation shows, cf. [40].
Proposition 4.1. Let S be a finite semigroup and M a simple kS-module.
Then the additive structure of M is either an abelian group or a join semi-
lattice with minimum.
Proof. Define a relation on M by m ≡ n if there exists j, k ∈ N so that
jm ∈ n + M and kn ∈ m + M . It is straightforward to verify that ≡ is
a congruence. Suppose first that ≡ is trivial and let m ∈ M . Then since
m ≡ m + m, it follows that m + m = m and so M is a join semilattice with
minimum. If ≡ is not proper, then m ≡ 0 and hence 0 ∈ M + m and so M
is an additive group.
(cid:3)
Let us now consider the join semilattice case. It turns out that in this
case, every representation of a finite group is trivial and so there is exactly
on irreducible module over an idempotent semiring associated to any regular
J -class of a finite semigroup.
Proposition 4.2. Let G be a finite group and k a commutative semiring
with unit. Suppose that M is a minimal kG-module with idempotent addi-
tion. Then there is a quotient semiring k′ of k so that M = k′ with trivial
action by G.
Proof. By minimality, it follows that G acts by automorphisms of M . Let
0 6= m0 ∈ M and put m = Pg∈G m0g. Then m is evidentally fixed by G
and so the k-span of m is a kG-submodule and hence is M by minimality.
If k′ is the faithful quotient of k acting on M , then M = k′ with the trivial
G-action.
(cid:3)
In particular, if k is a congruence-free commutative idempotent semiring
with unit (e.g., the boolean semiring B), then the trivial action of G on k is
the only irreducible kG-module for G a group. Recall that a finite semigroup
is aperiodic if all its maximal subgroups are trivial.
Corollary 4.3. Let S be a finite semigroup with a faithful irreducible kS-
module M whose addition is idempotent. Then S is generalized group map-
ping with aperiodic distinguished minimal ideal.
Proof. By Proposition 3.9, S is generalized group mapping. Let J be the
apex of M and let e ∈ J be an idempotent. Then M e is an irreducible
12
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
kGe-module, and hence has trivial action of Ge by Proposition 4.2. By
faithfulness, we conclude Ge is trivial. Since Ge is the maximal subgroup of
the distinguished ideal of S, this completes the proof.
(cid:3)
The irreducible BS-modules admit the following description.
Theorem 4.4. Let S be a finite semigroup. Then the minimal and irre-
ducible BS-modules are obtained as follows. Fix a regular J -class J of S
with set A of R-classes and B of L -classes. Let C : B × A → B be the
matrix with Cba = 1 if and only if the H -class a ∩ b contains an idempotent.
Let BB be the free module on B and consider the natural right action of S on
BB. Define a congruence ≡ on BB by putting m ≡ n if ms = 0 ⇐⇒ ns = 0
for all s ∈ S. Then:
(1) The congruence ≡ is the unique maximal proper congruence on B;
(2) The module BB is minimal, as is every proper quotient of B;
(3) BB/≡ is irreducible and all irreducible modules BS-modules are of
this form for some J -class.
Alternatively, BB/≡ can be identified with the B-span of the rows of the
matrix C.
Proof. First observe that by Proposition 4.2, the only minimal (and hence
irreducible) module for a finite group G over B is B equipped with the trivial
action of G. Thus each J -class provides a unique irreducible BS-module
MJ , coming from the trivial representation of the maximal subgroup.
Let us observe that C is the tensor product of the structure matrix of
J with the trivial representation of the maximal subgroup of G. It follows
that MJ can be identified with the B-span of the rows of C by Theorem 3.7.
This theorem also implies that MJ ∼= BB/ker C. Let m, n ∈ BB. Then
first observe that mC is determined by which entries are 0. Now (mC)a =
Pb∈B mbCba and hence is 0 if and only if m is annihilated by the R-class
It now follows that mC = nC if and only if, for all s ∈ J, one has
a.
ms = 0 ⇐⇒ ns = 0.
Next observe that since J is the apex of MJ , it follows easily by minimality
of MJ that if 0 6= m ∈ MJ , then mJ 6= 0. Thus if ms 6= 0 with s ∈ S, we
can find x ∈ J so that msx 6= 0. In particular, it follows by the definition of
an apex that sx ∈ J. We conclude that m ≡ n if and only if ms = 0 ⇐⇒
ns = 0 for all s ∈ J. This completes the proof of the (3) of the proposition.
The general theory implies that BB has a unique maximal congruence and
hence it must be ≡.
It remains to prove (2). First notice that by the general theory, the unique
maximal submodule of BB is the congruence class of 0 under the unique
maximal congruence. Therefore, the unique maximal submodule of BB is
the set of all vectors annihilated by C. But C is a boolean matrix with no
zero rows or columns. Hence, no non-zero vector in BB is annihilated by C.
Thus BB is minimal and hence so are all its quotients by Proposition 2.7. (cid:3)
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
13
In [33, Chapter 4.6], it is shown that, for each regular J -class J of a finite
semigroup S, there is a unique congruence ≡J on S such that S/≡J is gener-
alized group mapping with aperiodic distinguished ideal and J 6 ↑ maps to 0.
The resulting quotient is denoted AGGMJ (S) and the quotient map is written
ΓJ : S → AGGMJ (S). Consequently, AGGMJ (S) must then be isomorphic
to the image of S under the irreducible representation S → EndB(MJ ) con-
structed in the above proof. In [33, Chapter 4] (see also [39]), it is shown that
the intersection over all regular J -classes J of the congruences ≡J is the
largest J ′-congruence on S. Recall that a congruence ≡ is a J ′-congruence
if s ≡ t and s, t regular implies s J t. Thus we have proved:
Theorem 4.5. The largest J ′-congruence on a finite semigroup is the con-
gruence associated to the direct sum of all irreducible representations of S
over B.
The analogous theorems for fields of characteristic 0 and p can be found
in [1, 32].
A semigroup S is called a local group if eSe is a group for each idempotent
e ∈ S. The collection of finite local groups is denoted LG and is a pseudova-
riety, i.e., is closed under finite products, subsemigroups and homomorphic
images. If V is a pseudovariety, then the Mal'cev product LG (cid:13)m V consists
of all finite semigroups S admitting a homomorphism ϕ : S → T with T ∈ V
and ϕ−1(e) ∈ LG for each idempotent e of T . Theorem 4.5 in conjunction
with [33, Theorem 4.6.50] yields our next result.
Corollary 4.6. Let V be a pseudovariety of semigroups and let S be a finite
semigroup. Then S ∈ LG (cid:13)m V if and only if the image of S under every
irreducible representation over B belongs to V.
4.1. Duality and simple boolean modules. If M is a finite join semi-
lattice with identity, then it is automatically a complete lattice. The dual
semilattice M op is M with the reverse ordering. It can be identified with
the collection of functionals f : M → B with pointwise operations. Indeed,
it is well known [33, Chapter 9] that the functionals are given by choosing
m ∈ M and defining
ϕm(n) =(0 n ≤ m
1 n (cid:2) m
and so m ≤ k if and only if ϕk ≤ ϕm.
From now on we identify M op with the space of functionals. It will be
convenient to use the following boolean analogue of the Stone-Weierstrass
theorem.
Proposition 4.7. Let M be a finite join semilattice with identity. Then
a subsemilattice with identity N ⊆ M op is equal to M op if and only if it
separates the points of M .
Proof. By the above construction of the functionals, it is clear that M op
separates the points of M . Suppose that N is a proper subsemilattice of M
14
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
with identity and let
M ′ = {m ϕm ∈ N }.
Then M ′ is a proper meet-semilattice containing the top of M . Let m be a
maximal element of M that does not belong to M ′. Then the set of elements
strictly above m is non-empty and belongs to M ′. It therefore has a meet m′
in M ′, which must be the unique cover of m. Suppose that f ∈ N . We claim
f (m) = f (m′). Indeed, if f = ϕk with k ∈ M ′, then we have two cases. If
m ≤ k, then since m < k we have m′ ≤ k and so f (m) = 0 = f (m′). If
m (cid:2) k, then certainly m′ (cid:2) k and so f (m) = 1 = f (m′). It follows that N
does not separate points of M .
(cid:3)
Suppose now that M is a finite right-BS module for a finite semigroup S.
Then M op is naturally a left BS-module by putting sf (m) = f (ms).
Theorem 4.8 (Duality). Let S be a finite semigroup and let M be a finite
right BS-module. Then M is simple (minimal) if and only if M op is minimal
(simple).
Proof. Since (M op)op ∼= M , it suffices to show that M is simple if and only if
M op is minimal. Suppose first that M is simple and let N be a non-zero BS-
submodule of M op. Define a congruence ≡ on M by m ≡ m′ if f (m) = f (m′)
for all f ∈ N . This is a congruence because m ≡ m′ and s ∈ S implies that,
for all f ∈ N , we have f (ms) = sf (m) = sf (m′) = f (m′s) as sf ∈ N . Thus
ms ≡ m′s and so ≡ is a congruence. Since N contains a non-zero functional
f , it follows that ≡ is a proper congruence. Thus by simplicity of M , it
follows that ≡ is the trivial congruence. Thus the elements of N separate
points and so N = M op by Proposition 4.7. It follows that M op is minimal.
Conversely, suppose that M op is minimal and let ≡ be a non-trivial con-
gruence on M . Let N = {f ∈ M op ≡ ⊆ ker f }. Then N is a BS-submodule
of M op. Indeed, trivially 0 ∈ N . If f, g ∈ N and m ≡ m′, then (f + g)(m) =
f (m) + g(m) = f (m′) + g(m′) = (f + g)(m′). Finally, if f ∈ N , s ∈ S and
m ≡ m′, then ms ≡ m′s and so sf (m) = f (ms) = f (m′s) = sf (m′). Thus
sf ∈ N . Since M op separates points and ≡ is non-trivial, we conclude that
N 6= M op and hence N = 0 by minimality. Suppose ≡ is proper. Then
M/≡ 6= 0 and has enough functionals to separate points. Hence there is a
non-zero functional f on M with ≡ ⊆ ker f , a contradiction. Thus ≡ is not
proper. This completes the proof that M is simple.
(cid:3)
It follows that in principle, we can construct the simple BS-modules by du-
alizing the minimal left BS-modules, which are constructed in Theorem 4.4.
4.2. Characters. In this subsection we propose a definition of the character
of a boolean representation and also the notion of a generalized character.
These results are preliminary and will be expanded on in a future paper.
Let M be a finite lattice, which we view as a B-module via its join (and
hence we utilize additive notation for the join).
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
15
Definition 4.9 (sji). An element m ∈ M is said to be strictly join irreducible
(or sji) if m 6= 0 and m = m1 + m2 implies m = m1 or m = m2. See [33,
Chapter 6.1.2].
Let us denote by min M the set of sji elements: it is the the unique minimal
spanning set of M as a B-module. In particular, M is a free B-module if
and only if min M is a basis. Let us say that a decomposition
m = Xx∈min M
cxx
(4.1)
is irredundant if changing any coefficient cx from a 1 to a 0 results in a strictly
smaller element of M . In other words, an irredundant decomposition is one
of the form m = Px∈X x with X ⊆ min M and such that for no proper
subset Y ( X is m = Py∈Y y.
If M is free, then any decomposition of
an element m is irredundant and in particular x ∈ min M appears in an
irredundant decomposition of m if and only if x ≤ m.
Example 4.10. Let M be the span of B = {(1, 0, 0), (1, 1, 0), (0, 1, 1)} in B3.
Then min M = B. Now one checks
(1, 1, 1) = (1, 0, 0) + (0, 1, 1)
(1, 1, 1) = (1, 1, 0) + (0, 1, 1)
and these are all the irredundant decompositions of (1, 1, 1).
Suppose that S is a finite semigroup and M is a module for S over a field
k. Let χ be the character of M . Then if B is a basis for M , one has that
where
χ(s) =Xb∈B
bs = Xb′∈B
cbb
cbb′b′.
In other words, it sums over all b ∈ B the multiplicity of b in the decom-
position of bs with respect to the basis B (where the multiplicity is taken
in the field k). The problem with extending this idea to boolean case is
the non-uniqueness of irredundant decompositions. So instead we try and
minimize over all decompositions. With this in mind, we proceed to define
the character of a BS-module in two steps.
Definition 4.11 (Character of a boolean matrix representation). Suppose
that ϕ : S → Mn(B) is a matrix representation. Let B be the basis of Bn.
Then define
χϕ(s) = {b ∈ Bbs ≥ b} .
Equivalently, χϕ(s) is the trace tr(ϕ(s)) where we view ϕ(s) as a zero-one
matrix over C. If M is the corresponding BS-module, then we also use the
notation χM for its character.
16
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
Recall [2] that a boolean representation ϕ : S → Mn(B) is unambiguous if
the product ϕ(s)ϕ(t), viewed as matrices over C, coincides with ϕ(st). Un-
ambiguous representations play a key role in the theory of rational codes [2].
We shall say that a BS-module M is unambiguous if it is a free B-module
and the corresponding matrix representation is unambiguous. For instance,
if S acts on a finite set X, then the BS-module BX is unambiguous. From
the definition, it is immediate that the character of an unambiguous module
is a complex character of the semigroup.
Next suppose that M is a finite BS-module (with S a finite semigroup).
Then there is a natural surjective B-linear map π : B(min M ) → M induced
by the identity map on min M . Associated to each set-theoretic section
σ : M → B(min M ) of π with σmin M = 1min M is a BS-module structure Mσ
on B(min M ) defined by sx = σ(sx) for s ∈ S and x ∈ min M . Moreover,
notice that if S has an identity 1 and M is unitary, then 1x = σ(1x) =
σ(x) = x and so Mσ is unitary. In any case, one has that π : Mσ → M is a
surjective morphism of BS-modules.
Definition 4.12 (Min character). Let S be a finite semigroup and M a
finite BS-module. Then the min character χM : S → N is defined by
χM (s) = min{χMσ (s)}
where σ runs over all set-theoretic sections of π : B(min M ) → M with
σmin M = 1min M .
In more concrete terms, to compute χM (s), one first fixes an irredundant
decomposition of each element of M . Then one counts how many x ∈ min X
appear in the chosen irredundant decomposition of xs. Then one minimizes
this quantity over all choices made. Of course, if M is a BS-module that
is free as a B-module, then clearly the two notions of the character of M
coincide.
Example 4.13. Let M be the span of B = {(1, 0, 0), (1, 1, 0), (0, 1, 1)} in B3.
The remaining elements of M are (0, 0, 0) and (1, 1, 1). As observed earlier
min M = B. Any admissible section of π must be the identity on all elements
of M except (1, 1, 1), which has three lifts:
(1, 1, 1) = (1, 0, 0) + (0, 1, 1)
(1, 1, 1) = (1, 1, 0) + (0, 1, 1)
(1, 1, 1) = (1, 0, 0) + (1, 1, 0) + (0, 1, 1)
and so there are three corresponding sections σ1, σ2, σ3. Let S be the two-
element semilattice {1, e} and let S act on M by having 1 act as the identity
and e act via the map sending all non-zero elements to (1, 1, 1) (and of course
preserving zero). Then M is a BS-module. Let us compute χM (e). It is easy
to compute χMσ1 (e) = 2, χMσ2 (e) = 2 and χMσ3 (e) = 3. Thus χM (e) = 2.
Notice that σ3 corresponds to a redundant decomposition and hence has to
be eliminated when taking the minimum.
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
17
Notice that if we were working over C, then B would be a basis for C and
if e took each basis vector to (1, 1, 1), then since (1, 1, 1) = (1, 0, 0) + (0, 1, 1)
we would also get 2 as being the character value on e.
Next we discuss the notion of a generalized character. Let M be a finite
BS-module and min M ⊆ D ⊆ M \ {0}. Then we define the generalized
character ψM,D : S → N by
ψM,D(s) = {m ∈ D ms = m} .
For example, if S acts on the finite set X, then the generalized character
ψBX,X is the complex character of the CS-module CX and also coincides
with the min character introduced above. Let us define the generalized
character spectrum cspec(M ) to be the set of all min M ⊆ D ⊆ M \ {0}
such that ψM,D is a complex character of S.
Proposition 4.14. Let M be a non-zero finite BS-module. Let min M ⊆
D ⊆ M \ {0} be a subset such that d ∈ D implies ds ∈ D ∪ {0} for all s ∈ S.
Then D ∈ cspec(M ). In particular, M \ {0} ∈ cspec(M ).
Proof. Let us write for the moment θ for the zero of M and put D′ = D∪{θ}.
Then S acts by total functions on D′. Thus CD′ is a finite dimensional
CS-module in a natural way and Cθ is a CS-submodule. Consider the CS-
module V = CD/Cθ. It is easy to see that ψM,D is precisely the character
of V . This completes the proof.
(cid:3)
Recall that if S is a semigroup and R is an R-class the Schutzenberger
representation of S on R is the actions of S on R by partial transformations
given by r · s = rs if rs ∈ R and otherwise is undefined (where r ∈ R and
s ∈ S). One can turn CR into a CS-module as follows. First let S act on
R ∪ {(cid:3)} by sending all undefined products to (cid:3) and demanding (cid:3)s = (cid:3) for
all s ∈ S. Then consider the CS-module C(R ∪ {(cid:3)})/C(cid:3) (which as a vector
space is isomorphic to CR).
Theorem 4.15. Let S be a finite semigroup and MJ an irreducible BS-
module where we retain the notation of Theorem 4.4. Then there is subset
min M ⊆ D ⊆ M \ {0} such that the generalized character ψM,D is the
complex character of S obtained by lifting the right Schutzenberger repre-
sentation of AGGMJ (S) on an R-class of its distinguished J -class via the
projection ΓJ : S → AGGMJ (S).
Proof. Without loss of generality, we may assume that the action of S on
MJ is faithful and hence S = AGGMJ (S). In this case, since J is aperiodic,
we may identify the action of S on B by partial transformations with the
Schutzenberger representation. Let [b] denote the equivalence class of b ∈ B
under the congruence ≡ on B from Theorem 4.4. Then D = {[b] b ∈ B}
satisfies the conditions of Proposition 4.14. If we can show that the map
b 7→ [b] is injective, then the proof of Proposition 4.14 will imply that ψMJ ,D
is the complex character of the Schutzenberger representation. Suppose that
18
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
[b] = [b′] with b 6= b′. Let a be an R-class of J and choose s ∈ a ∩ b and
s′ ∈ a ∩ b′. Then, for x ∈ B, one has [x]s = [b] if Cxa = 1 and is otherwise
0 and [x]s′ = [b′] if Cxa = 1 and is otherwise 0. Thus s and t act the same
on MJ and so the action of S on MJ is not faithful, a contradiction. This
completes the proof.
(cid:3)
5. Density
This section relates our work with that of Zumbragel [40]. In particular,
we look at irreducible modules for semirings and discuss an application to
semigroups. Of course, any irreducible module for a semiring R is an irre-
ducible module for its underlying multiplicative semigroup, but the converse
is false. So in principle if R is a finite semiring, then we can use our results
to understand its irreducible representations as a semiring. One just has to
determine which irreducible representations preserve the additive structure.
However, since we are not assuming in general that the semirings in question
are finite and also because proofs can become shorter and coordinate-free
by taking advantage of the additive structure, we do not treat the represen-
tation theory of semirings as a special case of the representation theory of
semigroups.
1, . . . , m′
1, . . . , m′
A semiring R is called primitive if it has a faithful irreducible module M .
Let D be a division ring and M a left vector space over D. Then a subring
R ⊆ EndD(M ) is said to be dense if, for any pair of k-tuples (m1, . . . , mk)
and (m′
k) of linearly independent elements of M , there is an ele-
ment r ∈ R with (m1r, . . . , mnr) = (m′
k). If R is a primitive ring
with faithful irreducible module M , then EndR(M ) is a division ring D by
Schur's lemma and so R ⊆ EndD(M ). Jacobson's density theorem shows
that R must be a dense subring. Zumbragal [40] proved an analogous re-
sult for finite primitive semirings with idempotent addition, although he
stated it only under the hypothesis that R is congruence-simple. In fact,
for finite semirings with idempotent addition, being primitive implies being
congruence-simple. For finite rings, primitivity is also equivalent to simplic-
ity since a finite primitive ring R will have a faithful finite irreducible module
M and hence D above will be a finite division ring and thus a field by a the-
orem of Wedderburn. But then R will be a finite dimensional algebra over
a field with a faithful irreducible module and hence by Wedderburn-Artin
theory is simple.
Because Zumbragal does not state his result [40] in full generality, we
reproduce it for the reader's convenience. In what follows we assume that
M is a join semilattice with minimum and maximum. For example any finite
join semilattice with minimum also has a maximum. We shall always denote
the maximum by ∞. Following Zumbragal [40], given a, b ∈ M , define
xea,b =(0 x ≤ a
else.
b
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
19
One can check that the ea,b ∈ End(M ) and that {ea,b a, b ∈ M } ∪ {0} is a
subsemigroup E since
ea,bec,d =(0
ea,d
b ≤ c
b (cid:2) c.
It is shown below that M is an irreducible module for E. Zumbragal defines
R ⊆ End(M ) to be dense if it contains all the ea,b with a, b ∈ M .
Proposition 5.1. Let M be a join semilattice with minimum and let E be as
above. Then M is an irreducible BE-module. Thus any dense subsemiring
of End(M ) is primitive.
Indeed, if 0 6= m, n ∈ M then
Proof. First observe that M is minimal.
me0,n = n. Next suppose that ≡ is a non-trivial semilattice congruence on
M . To show that ≡ is not proper, suppose m ≡ n with m 6= 0. Then without
loss of generality we may assume m (cid:2) n. Let a ∈ M . Then mem,a = 0 and
men,a = a. Thus 0 ≡ a, completing the proof since a was arbitrary.
(cid:3)
The following is [40, Theorem 2.3].
Theorem 5.2 (Zumbragal). Let M be a join semilattice with minimum and
let R ⊆ End(M ) be dense. Then R is congruence-simple.
The next result is a strengthening of the statement of [40, Proposition
3.13], but the proof is the same. The reader should compare with Theo-
rem 4.4.
Theorem 5.3. Let R be an idempotent semiring and M a faithful irreducible
R-module with maximum. Then R is a dense subring of End(M ).
Proof. Define a congruence on M by m ≡ n if mr = 0 ⇐⇒ nr = 0 for all
r ∈ R. This is easily verified to be a congruence. Moreover, it cannot be
universal since m ≡ 0 implies mR = 0. Thus it is the trivial congruence.
Let Im denote the annihilator in R of m. It is a right ideal.
First note that e∞,b = 0 and so belongs to R. So fix ∞ 6= a ∈ M . We
first show ea,∞ ∈ R. Suppose x (cid:2) a. We claim Ia does not annihilate x.
Indeed, if Ia annihilates x, then since (x+ a)r = 0 if and only if xr = 0 = ar,
it follows that x + a ≡ a and so x ≤ x + a = a. Now xIa is then a non-
zero submodule of M and so xIa = M . Thus we can find rx ∈ Ia so that
xrx = ∞. Let
s =Xx(cid:2)a
rx ∈ Ia.
Then if x ≤ a, one has xs ≤ as = 0, whereas if x (cid:2) a, then xs ≥ xrx = ∞
and so s = ea,∞.
Now let ∞ 6= b ∈ M . Then by irreducibility, ∞R = M and so b = ∞r for
some r ∈ R. Then, ea,∞r = ea,∞r = ea,b and so R is dense, as required. (cid:3)
As a consequence, we obtain the following result of [40] (wherein the
equivalence with the first item is not stated explicitly).
20
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
Corollary 5.4 (Zumbragal). Let R be a finite idempotent semiring with
R2 6= 0. Then the following are equivalent:
(1) R is primitive;
(2) R is isomorphic to a dense subsemiring of End(M ) for some finite
semilattice M with minimum;
(3) R is congruence-free.
Proof. The implication 2 implies 3 is [40, Theorem 2.3], whereas 3 implies
1 is [40, Proposition 3.10]. The final implication follows from Theorem 5.3
and the observation that any irreducible module for a finite idempotent
semiring must be finite and idempotent (since M = mR for any non-zero
element m ∈ M ).
(cid:3)
It follows that if S is a finite semigroup, k is a finite idempotent semir-
ing and M is an irreducible kS-module, then the span of the image of the
representation ϕ : S → Endk(M ) is dense. Zumbragal showed [40, Proposi-
tion 4.9, Remark 4.10] that if M is finite, then End(M ) contains no proper
dense subsemirings if and only if M is a distributive lattice. In particular,
this applies when M is a finitely generated free B-module since then M is
isomorphic to the power set of its basis. We summarize this discussion in
the next corollary.
Corollary 5.5. Suppose S is a finite semigroup and let ϕ : S → End(M )
be an irreducible representation where M is a join semilattice with identity.
Then ϕ(S) spans a dense subsemiring of End(M ). In particular, if M is
a distributive lattice, then ϕ(S) spans End(M ). Consequently, given an
irreducible representation ϕ : S → Mn(B), the image of ϕ spans Mn(B).
The last part of the corollary can also be deduced in a straightforward
way from Theorem 4.4. The key step is to note that the span of the rows
of C is a free B-module if and only if C has an identity submatrix of the
appropriate rank.
References
[1] J. Almeida, S. Margolis, B. Steinberg, and M. Volkov. Representation theory of finite
semigroups, semigroup radicals and formal language theory. Trans. Amer. Math. Soc.,
361(3):1429 -- 1461, 2009.
[2] J. Berstel, D. Perrin, and C. Reutenauer. Codes and automata, volume 129 of Encyclo-
pedia of Mathematics and its Applications. Cambridge University Press, Cambridge,
2010.
[3] J. Berstel and C. Reutenauer. Rational series and their languages, volume 12 of
EATCS Monographs on Theoretical Computer Science. Springer-Verlag, Berlin, 1988.
[4] A. H. Clifford. Matrix representations of completely simple semigroups. Amer. J.
Math., 64:327 -- 342, 1942.
[5] A. H. Clifford. Basic representations of completely simple semigroups. Amer. J.
Math., 82:430 -- 434, 1960.
[6] A. H. Clifford and G. B. Preston. The algebraic theory of semigroups. Vol. I. Math-
ematical Surveys, No. 7. American Mathematical Society, Providence, R.I., 1961.
REPRESENTATION THEORY OF FINITE SEMIGROUPS OVER SEMIRINGS
21
[7] A. Connes and C. Consani. Characteristic one, entropy and the absolute point. Tech-
nical Report arXiv:0911.3537, arXiv, 2009.
[8] A. Connes, C. Consani, and M. Marcolli. Fun with F1. J. Number Theory,
129(6):1532 -- 1561, 2009.
[9] J. H. Conway. Regular algebra and finite machines. Chapman & Hall, London, 1971.
[10] M. Develin, F. Santos, and B. Sturmfels. On the rank of a tropical matrix. In Combi-
natorial and computational geometry, volume 52 of Math. Sci. Res. Inst. Publ., pages
213 -- 242. Cambridge Univ. Press, Cambridge, 2005.
[11] O. Ganyushkin, V. Mazorchuk, and B. Steinberg. On the irreducible representations
of a finite semigroup. Proc. Amer. Math. Soc., 137(11):3585 -- 3592, 2009.
[12] A. Gathmann. Tropical algebraic geometry. Jahresber. Deutsch. Math.-Verein.,
108(1):3 -- 32, 2006.
[13] K. G lazek. A guide to the literature on semirings and their applications in mathemat-
ics and information sciences. Kluwer Academic Publishers, Dordrecht, 2002. With
complete bibliography.
[14] J. A. Green. On the structure of semigroups. Ann. of Math. (2), 54:163 -- 172, 1951.
[15] J. A. Green. Polynomial representations of GLn, volume 830 of Lecture Notes in
Mathematics. Springer-Verlag, Berlin, 1980.
[16] U. Hebisch and H. J. Weinert. Semirings: algebraic theory and applications in com-
puter science, volume 5 of Series in Algebra. World Scientific Publishing Co. Inc.,
River Edge, NJ, 1998. Translated from the 1993 German original.
[17] I. Itenberg, G. Mikhalkin, and E. Shustin. Tropical algebraic geometry, volume 35 of
Oberwolfach Seminars. Birkhauser Verlag, Basel, second edition, 2009.
[18] M. Kilp, U. Knauer, and A. V. Mikhalev. Monoids, acts and categories, volume 29 of
de Gruyter Expositions in Mathematics. Walter de Gruyter & Co., Berlin, 2000. With
applications to wreath products and graphs, A handbook for students and researchers.
[19] K. Krohn, J. Rhodes, and B. Tilson. Algebraic theory of machines, languages, and
semigroups. Edited by Michael A. Arbib. With a major contribution by Kenneth
Krohn and John L. Rhodes. Academic Press, New York, 1968. Chapters 1, 5 -- 9.
[20] G. Lallement and M. Petrich. Irreducible matrix representations of finite semigroups.
Trans. Amer. Math. Soc., 139:393 -- 412, 1969.
[21] G. L. Litvinov. The Maslov dequantization, and idempotent and tropical mathemat-
ics: a brief introduction. Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov.
(POMI), 326(Teor. Predst. Din. Sist. Komb. i Algoritm. Metody. 13):145 -- 182, 282,
2005.
[22] S. W. Margolis and B. Steinberg. The quiver of an algebra associated to the Mantaci-
Reutenauer descent algebra and the homology of regular semigroups. Algebr. Repre-
sent. Theory, to appear.
[23] D. B. McAlister. Characters of finite semigroups. J. Algebra, 22:183 -- 200, 1972.
[24] W. D. Munn. On semigroup algebras. Proc. Cambridge Philos. Soc., 51:1 -- 15, 1955.
[25] W. D. Munn. Matrix representations of semigroups. Proc. Cambrdige Philos. Soc.,
53:5 -- 12, 1957.
[26] L. Pol´ak. Syntactic semiring of a language (extended abstract). In Mathematical foun-
dations of computer science, 2001 (Mari´ansk´e L´azne), volume 2136 of Lecture Notes
in Comput. Sci., pages 611 -- 620. Springer, Berlin, 2001.
[27] L. Pol´ak. Syntactic semiring and language equations. In Implementation and appli-
cation of automata, volume 2608 of Lecture Notes in Comput. Sci., pages 182 -- 193.
Springer, Berlin, 2003.
[28] L. Pol´ak. Syntactic semiring and universal automaton. In Developments in language
theory, volume 2710 of Lecture Notes in Comput. Sci., pages 411 -- 422. Springer,
Berlin, 2003.
[29] L. Pol´ak. A classification of rational languages by semilattice-ordered monoids. Arch.
Math. (Brno), 40(4):395 -- 406, 2004.
22
Z. IZHAKIAN, J. RHODES, AND B. STEINBERG
[30] I. S. Ponizovskiı. On matrix representations of associative systems. Mat. Sb. N.S.,
38(80):241 -- 260, 1956.
[31] D. Rees. On semi-groups. Proc. Cambridge Philos. Soc., 36:387 -- 400, 1940.
[32] J. Rhodes. Characters and complexity of finite semigroups. J Combinatorial Theory,
6:67 -- 85, 1969.
[33] J. Rhodes and B. Steinberg. The q-theory of finite semigroups. Springer Monographs
in Mathematics. Springer, New York, 2009.
[34] J. Rhodes and Y. Zalcstein. Elementary representation and character theory of fi-
nite semigroups and its application. In Monoids and semigroups with applications
(Berkeley, CA, 1989), pages 334 -- 367. World Sci. Publ., River Edge, NJ, 1991.
[35] J. Richter-Gebert, B. Sturmfels, and T. Theobald. First steps in tropical geometry. In
Idempotent mathematics and mathematical physics, volume 377 of Contemp. Math.,
pages 289 -- 317. Amer. Math. Soc., Providence, RI, 2005.
[36] A. Salomaa and M. Soittola. Automata-theoretic aspects of formal power series.
Springer-Verlag, New York, 1978. Texts and Monographs in Computer Science.
[37] M.-P. Schutzenberger. Sur la repr´esentation monomiale des demi-groupes. C. R. Acad.
Sci. Paris, 246:865 -- 867, 1958.
[38] M. P. Schutzenberger. On the definition of a family of automata. Information and
Control, 4:245 -- 270, 1961.
[39] B. R. Tilson. Appendix to "Algebraic theory of finite semigroups". On the p-length
of p-solvable semigroups: Preliminary results. In K. Folley, editor, Semigroups (Proc.
Sympos., Wayne State Univ., Detroit, Mich., 1968), pages 163 -- 208. Academic Press,
New York, 1969.
[40] J. Zumbragel. Classification of finite congruence-simple semirings with zero. J. Alge-
bra Appl., 7(3):363 -- 377, 2008.
Department of Mathematics, Bar-Ilan University, Ramat-Gan 52900, Israel
E-mail address: [email protected]
Department of Mathematics, University of California, Berkeley, Berke-
ley, CA 94720-3840, USA
E-mail address: [email protected]
E-mail address: [email protected]
Carleton University, 1125 Colonel By Drive, Ottawa, Ontario K1S 5B6,
Canada
E-mail address: [email protected]
|
1903.09063 | 1 | 1903 | 2019-03-21T15:32:35 | Noncyclic Division Algebras over Fields of Brauer Dimension One | [
"math.RA"
] | Let $K$ be a complete discretely valued field of rank one, with residue field $\Q_p$. It is well known that period equals index in $\Br(K)$. We prove that when $p=2$ there exist noncyclic $K$-division algebras of every $2$-power degree divisible by four. Otherwise, every $K$-division algebra is cyclic. | math.RA | math |
NONCYCLIC DIVISION ALGEBRAS OVER FIELDS OF
BRAUER DIMENSION ONE
ERIC BRUSSEL
Abstract. Let K be a complete discretely valued field of rank one, with
residue field Qp. It is well known that period equals index in Br(K). We prove
that when p = 2 there exist noncyclic K-division algebras of every 2-power
degree divisible by four. Otherwise, every K-division algebra is cyclic.
1. Introduction
Recent research has focused on formulating and computing arithmetic invariants in the
Brauer group, over fields of cohomological dimension greater than 1. Two of the most
prominent measures are Brauer dimension and cyclic length. Famous fields of Brauer
dimension one include local fields and global fields, and both have cyclic length one. Be-
cause of the local-global principle and the degree-criterion for splitting over local fields,
one proves both "cyclicity" and "period equals index" for global fields at the same time.
Which raises the question: If K is a field of Brauer dimension one, is it like a global field
in this respect, that every K-division algebra is cyclic? We show that the answer is no.
A field K has Brauer dimension one if the Brauer group Br(K) is not trivial, and every
element in Br(K) of period n is represented by a central simple algebra of degree n, for
any n. In other words, every class has equal period and index. K has cyclic length one
if every class of period n is represented by a cyclic central simple algebra of degree n, for
any n. All K-division algebras are then cyclic crossed products.
When K contains the n-th roots of unity, the theorem of Merkurjev-Suslin [19] guaran-
tees that nBr(K) is generated by cyclic classes of degree n, and the cyclic length is called
the symbol length. This has been studied extensively; see [2] and [6] for a discussion of
generation by cyclic classes, and [25] for fundamental results on symbol length. Cyclic
length is interesting not only as a measure of arithmetic complexity, but also because of
the central open problem in the area of finite-dimensional division algebras, which is to
determine whether every division algebra of prime degree is cyclic.
Recent work on Brauer dimension includes [3], [5], [9], [10], [11], [12], [16], [17], [20],
[21], and [22]. Two-dimensional local fields, such as Qp((t)) and Fp((s))((t)), were shown
to have Brauer dimension one in [9], and Fp((s))((t)) was shown to have cyclic length one
in [5]. The Brauer dimension of the function field of a p-adic curve was proved to be two
in [22], and shown to have cyclic length two in [8]. See [13], [18], and [14] for known results
on the relationship between Brauer dimension and symbol length.
In this paper we show that the field K = Q2((t)) has Brauer dimension one and cyclic
length two. This is the first example of a field K whose Brauer dimension is proved to be
less than its cyclic length. The testifying noncyclic division algebras are of every 2-power
degree divisible by four. Noncyclic division algebras of equal period and degree themselves
are not new, and were constructed in the early 1930's by Albert, but over fields of Brauer
dimension at least three [1, Lemma 10]. Our examples exist in a very small space, in the
1
2
ERIC BRUSSEL
sense that they seem to depend on the noncyclicity of the cyclotomic extension Q2(µ8)/Q2,
they all have small cyclic splitting fields, and all Q2((t))-division algebras of odd degree
are cyclic, as are all Qp((t))-division algebras for odd p.
2. Notation and Background
2.1. For Fields of Characteristic Zero. Let K be a field of characteristic zero. Let
X(K) denote the character group, and Br(K) the Brauer group. If δ ∈ Br(K), the period
per(δ) is the order in Br(K), and the index ind(δ) is the smallest degree of a splitting
field, equal to the degree of the K-division algebra underlying δ. If δ ∈ Br(K) and E/K
is a field extension, write δE for the restriction from K to E.
If ψ ∈ X(K) has order
n, let κ(ψ)/K denote the cyclic field extension of degree n fixed by ker(ψ). If E/K is a
finite field extension, write ψE for the restriction, then κ(ψE) is the corresponding cyclic
extension of E, which is the compositum of κ(ψ) and E inside some algebraic closure.
Let µn and µ denote the roots of unity groups, with µn(K) and µ(K) the corresponding
roots of unity in K. Let ζn denote a primitive n-th root of unity, and put i = ζ4. By
Hilbert 90 we have an isomorphism K ×/K ×n ≃ H1(K, µn); let (a)n ∈ H1(K, µn) denote
either the class of a (mod K ×n), or the cocycle GK → µn defined by σ 7→ (a1/n)σ−1.
Since −1 ∈ K and µ2 is canonically identified with {±1}, we may identify (a)2 with the
character of order 2. We will cite Albert's norm criterion, which says a character ψ extends
to one of order 2ψ if and only if −1 ∈ N(κ(ψ)/K). This appears in [4].
Let (a, b) ∈ Br(K) denote the standard quaternion class, for a, b ∈ K, and put (a, b) = 1
or (a, b) = −1, depending on whether the class is trivial or nontrivial. A cyclic class of
degree n is a Brauer class of the form θ ∪ (b)n where θ ∈ X(K), b ∈ K ×, and θ = n. It
then represents a cyclic central simple algebra of degree n. If ind((θ, b)) = θ, the division
algebra underlying (θ, b) is a cyclic crossed product, with maximal subfield κ(θ). By the
norm criterion for cyclic classes, (a, b) = 1 if and only if b ∈ N(K(√a)/K), and (θ, b) = 0
if and only if b ∈ N(κ(θ)/K).
For n ∈ N, the n-cyclic length of K is zero if Br(K) has no elements of period n,
and otherwise it is the minimum number of cyclic classes of degree n needed to express
a general element of Br(K) of period n. We use ∞ if there is no such expression. The
cyclic length of K is zero if Br(K) = {0}, and otherwise it is the supremum of the n-cyclic
lengths for all n. The n-Brauer dimension of K is zero if Br(K) has no elements of period
n, and otherwise it is the smallest number d such that ind(δ) divides nd for all δ ∈ Br(K)
of period n. The Brauer dimension of K is zero if Br(K) = {0}, and otherwise it is the
supremum of the n-Brauer dimensions over all n. Since a class of period n and n-cyclic
length d has index at most nd, the n-Brauer dimension is bounded by the n-cyclic length.
2.2. For Complete Discretely-Valued Fields. Suppose K is complete with respect to
a discrete rank-one valuation, with residue field k of characteristic zero, valuation ring OK,
n = Hom(µn, Q/Z) be the Pontrjagin dual, then H0(µn, Q/Z) =
and uniformizer t. Let µ∗
(µ∗
k is a subgroup of µ∗
n of order µn(k), and we have a split residue sequence
n)G
0 −→ nX(k) −→ nX(K) −→ (µ∗
n)Gk −→ 0
Thus every χ ∈ nX(K) is expressible as a sum χ = ψ + ω ∗ ◦ (t)n for ψ defined over k and
ω ∗ : µn → Q/Z of some order e dividing µn(k). If L = κ(eχ) then χL = ω ∗ ◦ (vt)n for
some v ∈ L by Kummer Theory, and then κ(χ) = L( e√vt) has t-ramification degree e.
The split residue sequence
0 −→ Br(k) −→ Br(K) −→ X(k) −→ 0
NONCYCLIC DIVISION ALGEBRAS
3
lets us express every δ ∈ Br(K) in the form δ = α + (θ, t), where α ∈ Br(k) and θ ∈ X(k)
are identified with their inflations to K. We call this the Witt decomposition (with respect
to t). The splitting criterion states that δ is split by a finite t-unramified field extension
E/k if and only if E contains κ(θ) and splits α, and we have an index formula (see e.g.
[7])
ind(δ) = θind(ακ(θ))
2.3. For Local Fields. Let k be a local field of characteristic zero, and let ζm be a
generator of µ(k). The reciprocity map ρ : k× → Gab
−→ (k×)∗
k induces an isomorphism
ρ∗ : X(k)
∼
by Pontrjagin duality. For each θ ∈ X(k) we set θ∗ = ρ∗(θ).
If k has residue field Fq, Wedderburn's Theorem Br(Fq) = 0 transforms the residue
sequence into an isomorphism from Br(k) to X(Fq), and its composition with the evaluation
at Frobenius yields the invariant map
Then
invk : Br(k)
∼
−→ Q/Z
invk(θ, b) = θ∗(b)
by [24, Ch.XIV, Prop.3] and Tate Duality. In particular, θ∗(b) equals the order of b in
k×/N(κ(θ)/k).
The decomposition k× ≃ µ(k) × Z × ZN
p induces a split exact sequence
0 −→ (Z × ZN
p )∗ −→ (k×)∗ −→ µ(k)∗ −→ 0
Since (Z × ZN
p )∗ is divisible, it follows that a character θ ∈ X(k) of prime-power order
extends to a character of any (larger) prime-power order if and only if θ∗(ζm) = 0, if and
only if ζm ∈ N(κ(θ)/k).
3. Results
Proposition 3.1. Let K be a complete discretely valued field of rank one with residue
field k = Qp. Then K has cyclic length at most two, and Brauer dimension one.
Proof. This is well known. The Witt decomposition shows nBr(K) is generated by n-cyclic
classes, and its n-cyclic length is at most two. Let δ = α + (θ, t) ∈ Br(K). By Local Class
Field Theory (LCFT), ind(ακ(θ)) = per(α)/gcd{per(α),θ}, hence ind(δ) = per(δ) =
lcm{θ, per(α)} by the index formula. Therefore K has Brauer dimension one.
Theorem 3.2. Let K be a complete discretely valued field with residue field Q2, Then
there exist noncyclic K-division algebras of every 2-power degree divisible by 4.
(cid:3)
Proof. Set k = Q2, and let t be a uniformizer for K. Since k does not contain a primitive
fourth root of unity, any character of 2-power order over K has residue of order dividing
2, and the corresponding cyclic extension has t-ramification degree dividing 2. Similarly,
any tamely ramified cyclic field extension of k of 2-power degree has p-ramification degree
dividing 2. Let α ∈ Br(k) be an element of order 2n divisible by 4, and let D be the
K-division algebra underlying the Brauer class
δ = α + (−1, t)
Then ind(D) = per(α) = 2n. D has no t-unramified cyclic maximal subfields: Such
a field extension would have degree 2n and contain k(i), by the splitting criterion, but
(−1,−1) = −1 by computation of the Hilbert symbol ([23, III.1.2, Theorem 1]), hence −1
is not a norm from k(i), hence k(i) is not contained in a cyclic extension of degree 4 by
4
ERIC BRUSSEL
Albert's norm criterion. Therefore no such subfield exists. It follows that D is cyclic if
and only if it has a cyclic maximal subfield of t-ramification degree 2. Such fields have the
form
E = L(√vt)
L . Let ¯L denote the residue
where L/K is of degree 2n−1 and defined over k, and v ∈ O×
field of L. Since the group of principal units of OL is divisible, we may assume v ∈ ¯L×.
Let χ ∈ X(K) be a character for E/K, and let
Then κ(ψ) = ¯L(√v) is a cyclic extension of k of degree 2n. Conversely if ¯L(√v)/k is cyclic
of degree 2n, then E = L(√vt) is cyclic over K. Therefore for our purposes ¯L/k could be
any cyclic extension of degree 2n−1 that extends to a cyclic extension of degree 2n.
ψ = χ − (t)2
Let τ = √vt. Compute
δE = αE + (−1, v−1τ 2)E = αE + (−1, v−1)E
Since E/L is totally t-ramified, and αE and (−1, v−1)E are defined over ¯L, the index of
this class is the same as that of α ¯L + (−1, v) ¯L (see [7, Lemma 1]). Therefore
ind(δE) = ind(α ¯L + (−1, v) ¯L)
Since [ ¯L : k] = 2n−1, ind(α ¯L) = 2 by the local theory. Since Br( ¯L) ≃ Q/Z, any two
classes of order 2 are inverse to each other. Therefore D is cyclic via E/K if and only if
(−1, v) ¯L = −1.
Claim: (−1, v) ¯L = 1. It follows immediately that δE is nonzero, and since E/K is arbitrary,
that D is noncyclic. By LCFT the corestriction from Br( ¯L) to Br(k) is injective, and since
−1 ∈ k, cor ¯L/k(−1, v) ¯L = (−1, N ¯L/k(v)) by the cup product formula for corestriction.
Therefore to prove the claim it suffices to show (−1, N ¯L/k(v)) = 1.
The claim applies to all possible ¯L/k, namely all ¯L/k of degree 2n−1 that extend to
cyclic extensions of degree 2n, and we must test all ¯L and v ∈ ¯L such that ¯L(√v)/k is
cyclic. We first show that if the claim is true on a given ¯L for one such v then it is true
for all v. Suppose κ(ψ) = ¯L(√v) is cyclic as above. If ψ′ ∈ X(k) is a character of order 2n
such that 2ψ′ = 2ψ, then ψ′ = ψ + (w)2 for some w ∈ k, and since ψ ¯L = (v)2, ψ′
¯L = (vw)2.
The corresponding χ′ = ψ′ + (t)2 ∈ X(K) yields the cyclic extension E ′ = L(√vwt), and
δE ′ = (α + (−1, t))E ′ = αE ′ + (−1, vw)E ′ . Since E ′/L is totally ramified, the index of δE ′
can again be computed over ¯L, and since w ∈ k, cor ¯L/k(−1, vw) ¯L = (−1, N ¯L/k(v)w2n−1
) =
(−1, N ¯L/k(v)). Thus if the claim is true for ¯L and v ∈ ¯L then (−1, N ¯L/k(v)) = 1 for any v.
We first prove the claim for n = 2, i.e., D of degree 4. Then [ ¯L : k] = 2. We have
H1(Q2, µ2) = {(±1)2, (±2)2, (±5)2, (±10)2}
by [23, II.3.3, Corollary], hence Q2 has 7 nontrivial quadratic extensions. By the computa-
tion of the Hilbert symbol ([23, III.1.2, Theorem 1]), (−1, 2) = (−1, 5) = (−1, 10) = 1 and
(−1,−1) = −1, and (−1, −2) = (−1,−5) = (−1, −10) = −1. Therefore the candidates for
¯L are Q2(√2), Q2(√5), or Q2(√10), by Albert's criterion. Now:
(i) If ¯L = Q2(√2), let v = 2 + √2. Then N ¯L/Q2 (v) = 2 = NQ2(i)/Q2 (1 + i).
(ii) If ¯L = Q2(√5), let v = 5 + 2√5. Then N ¯L/Q2 (v) = 5 = NQ2(i)/Q2 (2 + i).
(iii) If ¯L = Q2(√10), let v = 10 + √10. Then N ¯L/Q2 (v) = 90 = NQ2(i)/Q2 (9 + 3i).
It is easy to check, using [15, VI, Exercise 8] for example, that each ¯L(√v)/Q2 is cyclic
of degree 4. We conclude that (−1, N ¯L/Q2 (v)) = 1, hence (−1, v) ¯L = 1, for all quadratic
extensions ¯L/Q2 extendable to degree-4 cyclic extensions. Therefore D is noncyclic.
NONCYCLIC DIVISION ALGEBRAS
5
For the general case, we will reduce to the n = 2 case, as follows. Let ¯L′/k be the
quadratic subextension of ¯L, and let v′ = N ¯L/ ¯L′ (v). Then v′ is a product of the Galois
conjugates of v over ¯L′, each of which can replace v as an element whose square root
generates ¯L(√v), since ¯L(√v) is a splitting field over ¯L′. Since √v′ is the product of these
square roots, it is contained in ¯L, hence ¯L′(√v′)/k is a subextension of ¯L/k.
n)G
The element v′ is not a square in ¯L′, hence ¯L′(√v′)/k is cyclic of degree four. To
see this, let ω ∗ ∈ (µ∗
k be the element of order 2, and let res and cor be restriction
and corestriction for ¯L/ ¯L′. We have (v′)2, ¯L′ = cor ¯L/ ¯L′ ((v)2, ¯L), since the corestriction is
induced by the norm, hence cor(ω ∗ ◦ (v)2, ¯L) = ω ∗ ◦ (v′)2, ¯L′ . Since res(ψ ¯L′ ) = ω ∗ ◦ (v)2, ¯L,
and cor ◦ res is multiplication by [ ¯L : ¯L′] = 2n−2, we conclude ω ∗ ◦ (v′)2, ¯L′ = 2n−2ψ ¯L′ is a
character of order 2, hence v′ is not a square in ¯L′.
Since ¯L′(√v′)/k is cyclic of degree 4, (−1, v′) = (−1, N ¯L′/k(v′)) = 1 by the n = 2 case
of the claim. But since v′ = N ¯L/ ¯L′ (v), we have N ¯L′/k(v′) = N ¯L/k(v), so (−1, N ¯L/k(v)) = 1.
This proves the claim in general, and we conclude that D is noncyclic in general.
(cid:3)
Remark 3.3. It is not hard to show that the noncyclic division algebra of degree 2n
constructed above has a cyclic splitting field of degree 2n+1. In fact, set ζj = ζ2j , and
let ¯Ej, j ≥ 0, be the subfield of Q2(ζj+2) fixed by the automorphism ζj+2 7→ ζ −1
j+2. Then
¯Ej/Q2 is cyclic of degree 2j , and ¯Ej = Q2(ηj ), where ηj = ζj+2 + ζ −1
j+2. The sequence
{ηj : j = 0, . . . ,} is recursively definable by η0 = 0 and η2
j+1 = 2+ηj , and it is easy to prove
by induction that N ¯Ej+1/Q2 (ηj+1) = 2 for j ≥ 1. Now E = En(√ηnt) is a cyclic extension
of K of degree 2n+1, where En = ¯EnK. Since N ¯En/Q2 (ηn) = 2 = NQ2(i)/Q2 (1 + i), we
compute (−1, ηn) = 1, and since ¯En splits α, we easily verify that E splits D.
Theorem 3.4. Let K be a complete discretely valued field with residue field Qp, and let
D be a K-division algebra. Suppose either ind(D) is odd, or p is odd. Then D is cyclic.
Proof. Assume the hypotheses. Set k = Qp, and let m = µ(k). Let D be a K-division
algebra of degree n, with Brauer class δ = α + (θ, t) for α ∈ Br(k) and θ ∈ X(k).
If
θ = 0 then D is defined over the local field k, and is cyclic by local field theory. Therefore
suppose θ 6= 0. We may assume n = ℓa for some prime ℓ and a ∈ N by prime-power
decomposition. If κ(θ) extends to a (t-unramified) cyclic extension E/k of degree n, then
E contains κ(θ) and splits α, hence D is cyclic via E by the splitting principle. Therefore
we assume for the remainder of the proof that 1 < θ < n, and θ does not extend to a
character of order n. Since ind(δ) = n and θ < n, it follows that ind(α) = n, hence
invk(α) = i/n for some i : gcd(i, n) = 1.
Since it does not extend to a cyclic extension of degree n, κ(θ)/k is p-ramified, and
e = θ∗(ζm) is the smallest number such that e θ is extendable to a character of any
larger ℓ-power order, hence e is the order of ζm in k×/N(κ(θ)/k), by Section 2. Evidently
e : 1 < e ≤ θ.
By hypothesis, at least one of the numbers ℓ, p is odd, and our assumption on the non-
extendability of θ implies that ℓ 6= p: For if ℓ = p, then p is odd, so K contains no p-th
roots of unity, and the p-primary subgroup X(k)(p) is p-divisible by LCFT, contradicting
the non-extendability of θ. Therefore ℓ 6= p.
Suppose ℓ 6= p, and if ℓ = 2 then p ≡ 1 (mod 4). We will construct a cyclic maximal
subfield of D. In this case κ(θ)/k is tame, e is its ramification degree, and κ(eθ) is its
maximal p-unramified subextension. Since ℓ divides e divides m, and since ζm generates
µ(k), ζm is not an ℓ-th power in k×. If ℓ = 2 then ζm is not in −4k×4 since p ≡ 1 (mod 4).
Therefore xn − ζm is irreducible by [15, VI.9], with root ζmn, and the field ¯L = k(ζmn/e)
is cyclic and p-unramified of degree n/e over k. Since ζm has order e modulo N(κ(θ)/k),
we can find a j : gcd(j, e) = 1 so that invk(θ, ζ j
m) = i/e, where i is as above.
6
Let
ERIC BRUSSEL
v =(ζ j
mn/e
ζ j(m/2+1)
mn/e
if ℓ is odd
if ℓ = 2
m: When ℓ is odd, v is a root of xn/e − ζ j
Since the exponent is prime-to-ℓ in each case, k(v) = k(ζmn/e). This v is designed to have
norm ζ j
m. When ℓ = 2,
v is a root of xn/e − ζ j(m/2+1)
=
(−1)1+jm/2ζ j
, and since 4m, we compute Nk(v)/k(v) = −ζ j(m/2+1)
Put ¯L = k(v), L = K(v). Then ¯L( e√v) = k(ζmn) is cyclic of degree n over k. Let τ be
m, so Nk(v)/k(v) = ζ j
m
m
m = ζ j
m.
a root of xe − vt ∈ L[t]. Then we claim
E = L(τ )
is a cyclic maximal subfield of D.
It is cyclic over K as the cyclic extension of a sum
φ + ω ∗ ◦ (t)e as in Section 2.2, where φ ∈ X(k) is such that κ(φ) = k(ζmn)/k and φ ¯L =
ω ∗ ◦ (v)e. It remains to show E is a maximal subfield of D. Since t = τ e/v,
δE = αE + (θE, τ e) + (θE, 1/v)
Since ¯L/k is p-unramified, and since κ(eθ) is the maximal p-unramified subextension of
κ(θ)/k, θE = θ ¯L = e, so (θE, τ e) = 0, and since E/L is t-totally ramified, ind(δE) =
ind(α ¯L + (θ ¯L, 1/v)).
Since [ ¯L : k] = n/e and invk(α) = i/n, inv ¯L(α ¯L) = i/e by the local theory. Since
cor ¯L/k((θ ¯L, 1/v)) = (θ, N ¯L/k(1/v)) = (θ, ζ −j
m )
by the cup product formula, and since the corestriction is the identity on invariants by
the local theory, we have inv ¯L(θ ¯L, 1/v) = invk(θ, ζ −j
m ) = −i/e. Therefore inv ¯L(α ¯L +
(θ ¯L, 1/v)) = i/e − i/e = 0, hence δE = 0. Since ind(δ) = [E : K] = n, we conclude E is a
maximal subfield of D, so D is cyclic.
It remains to prove it for ℓ = 2 and p ≡ 3 (mod 4). Set ind(D) = n = 2r. If r = 1 then
D is already known to be cyclic, so assume r ≥ 2. We have ind(α) = 2r and 1 < θ < 2r,
and θ does not extend to a character of order 2r, hence κ(θ) is p-ramified. Since p ≡ 3
(mod 4), all characters of 2-power order have residue of order at most 2 by the sequence
in Section 2.2, so κ(θ)/k has ramification degree 2, and θ does not extend to a character
of larger 2-power degree. Therefore −1 6∈ N(κ(θ)/k) by Albert's criterion.
Let s = v2(p2 − 1) be the (additive) 2-value. Then ζ2s is the largest 2-power root of
unity in k(i). Let v = ζ2s+r−2 , and set ¯L = k(v), L = K(v), and
E = L(√vt)
By the local theory, ¯L/k is cyclic unramified of degree 2r−1, θ ¯L has order 2, and N ¯L/k(i)(v) =
−ζ2s . The Galois conjugate of −ζ2s over k is −ζ p
2s , the image under Frobenius, so
N ¯L/k(v) = ζ p+1
. We compute v2(p + 1) = s − 1 since p ≡ 3 (mod 4), therefore ζ p+1
has order 2, therefore N ¯L/k(v) = −1.
2s
2s
An argument similar to the previous case shows E/K is cyclic of degree 2r, and
ind(δE) = ind(α ¯L + (θ, v) ¯L)
Since N ¯L/k(v) = −1 is not a norm from κ(θ), ind((θ, v) ¯L) = ind((θ, N ¯L/k(v))) 6= 1, and
since θ ¯L has order 2, ind((θ, v) ¯L) = 2. Since [ ¯L : k] = 2r−1 we compute ind(α ¯L) = 2 by
the local theory, and since in Br(k) any two classes of order 2 are inverse to each other, we
conclude α ¯L + (θ, v) ¯L = 0. Therefore ind(δE) = 1, so D is split by E, hence D is cyclic.
This completes the proof.
(cid:3)
NONCYCLIC DIVISION ALGEBRAS
7
Corollary 3.5. Let K be a complete discretely valued field with residue field k = Qp.
Then the 2r-cyclic length of K is two if p = 2 and 4 divides 2r; the n-cyclic length of K
is one if p = 2 and n is odd or n = 2; and the cyclic length of K is one if p 6= 2.
Proof. Since all division algebras over the fields Q2 and Qp are cyclic, the Witt decom-
position of 2r Br(K) shows the 2r-cyclic length is at most two. Theorem 3.2 shows the
2r-cyclic length is at least two for Q2 and r ≥ 2. If 2r = 2 then the Witt decomposition
shows that in all cases there is a quadratic splitting field, hence the cyclic length is one.
The rest is Theorem 3.4.
(cid:3)
The symbol length does not detect the examples of Theorem 3.2, because symbol length
is only defined in circumstances where the examples don't exist:
Theorem 3.6. Let K be a complete, discretely valued field with residue field a local field k
of characteristic zero, and suppose k contains µn. Then every K-division algebra of degree
n is cyclic.
Proof. Let t be a uniformizer for K. As above, the Brauer class of a division algebra D of
degree n has the form δ = α + (θ, t), where α ∈ Br(k) and θ ∈ X(k) have order dividing
n. Since k contains µn, θ extends to a character ψ of order n by Kummer theory, and
the t-unramified cyclic extension κ(ψ)/K then contains κ(θ) and splits α, hence it splits
δ. Therefore D is cyclic.
(cid:3)
Example 3.7. Let K = k((t)) for k = Q3(i), and let n be a power of a prime ℓ. We show
how to prove every K-division algebra is cyclic. We have µ(k) = 8, and so
H1(k, µn) ≃ (Z/n × Z/gcd(n, 8)
Z/n × Z/n × Z/n
if ℓ 6= 3
if ℓ = 3
Now arguments identical to those in the proof of Theorem 3.4 show that every K-division
algebra is cyclic.
References
[1] A. Albert. Non-cyclic division algebras of degree and exponent four. Trans. Amer. Math. Soc.,
35:112 -- 121, 1933.
[2] S. Amitsur. Division algebras. A survey. In Algebraists' homage: papers in ring theory and
related topics (New Haven, Conn., 1981), volume 13 of Contemp. Math., pages 3 -- 26. Amer.
Math. Soc., Providence, R.I., 1982.
[3] B. Antieau. Cohomological obstruction theory for Brauer classes and the period-index prob-
lem. J. K-Theory, 8(3):419 -- 435, 2011.
Amer. Math. Soc., 313(2):843 -- 851, 1989.
[4] J. K. Arason, B. Fein, M. Schacher, and J. Sonn. Cyclic extensions of K(√
−1)/K. Trans.
[5] R. Aravire and B. Jacob. p-algebras over maximally complete fields. In K-theory and algebraic
geometry: connections with quadratic forms and division algebras (Santa Barbara, CA, 1992),
volume 58 of Proc. Sympos. Pure Math., pages 27 -- 49. Amer. Math. Soc., Providence, RI, 1995.
With an appendix by Jean-Pierre Tignol.
[6] A. Auel, E. Brussel, S. Garibaldi, and U. Vishne. Open problems on central simple algebras.
Transform. Groups, 16(1):219 -- 264, March 2011.
[7] E. Brussel. Noncrossed products and nonabelian crossed products over Q(t) and Q((t)). Amer.
J. Math., 117(2):377 -- 393, 1995.
[8] E. Brussel, K. McKinnie, and E. Tengan. Cyclic length in the tame Brauer group of the
function field of a p-adic curve. Amer. J. Math., 138(2):251 -- 286, 2016.
[9] I. D. Chipchakov. Henselian discrete valued stable fields. ArXiv e-prints, Feb. 2018.
[10] P. L. Clark. The period-index problem in WC-groups IV: a local transition theorem. J. Théor.
Nombres Bordeaux, 22(3):583 -- 606, 2010.
8
ERIC BRUSSEL
[11] A. de Jong. The period-index problem for the Brauer group of an algebraic surface. Duke
Math. J., 123(1):71 -- 94, 2004.
[12] D. Harbater, J. Hartmann, and D. Krashen. Applications of patching to quadratic forms and
central simple algebras. Invent. Math., 178:231 -- 263, 2009.
[13] B. Kahn. Comparison of some field invariants. J. Algebra, 232(2):485 -- 492, 2000.
[14] D. Krashen. Period and index, symbol lengths, and generic splittings in Galois cohomology.
Bull. Lond. Math. Soc., 48(6):985 -- 1000, 2016.
[15] S. Lang. Algebra, volume 211 of Graduate Texts in Mathematics. Springer-Verlag, New York,
third edition, 2002.
[16] M. Lieblich. Period and index in the Brauer group of an arithmetic surface. J. Reine Angew.
Math., 659:1 -- 41, 2011. With an appendix by Daniel Krashen.
[17] M. Lieblich. The period-index problem for fields of transcendence degree 2. Ann. of Math.
(2), 182(2):391 -- 427, 2015.
[18] E. Matzri. Symbol length in the Brauer group of a field. Trans. Amer. Math. Soc., 368(1):413 --
427, 2016.
[19] A. Merkurjev and A. Suslin. K-cohomology of Severi-Brauer varieties and the norm residue
homomorphism. Math. USSR Izv., 21(2):307 -- 340, 1983.
[20] R. Parimala and V. Suresh. The u-invariant of the function fields of p-adic curves. Ann. of
Math. (2), 172(2):1391 -- 1405, 2010.
[21] R. Parimala and V. Suresh. Period-index and u-invariant questions for function fields over
complete discretely valued fields. Invent. Math., 197(1):215 -- 235, 2014.
[22] D. Saltman. Division algebras over p-adic curves. J. Ramanujan Math. Soc., 12:25 -- 47, 1997.
[23] J.-P. Serre. A Course in Arithmetic. Springer-Verlag, New York-Heidelberg, 1973. Translated
from the French, Graduate Texts in Mathematics, No. 7.
[24] J.-P. Serre. Local Fields, volume 67 of Graduate Texts in Mathematics. Springer, New York-
Berlin, 1979.
[25] J.-P. Tignol. On the length of decompositions of central simple algebras in tensor products
of symbols. In Methods in ring theory (Antwerp, 1983), volume 129 of NATO Adv. Sci. Inst.
Ser. C Math. Phys. Sci., pages 505 -- 516. Reidel, Dordrecht, 1984.
Department of Mathematics, California Polytechnic State University, San Luis
Obispo, CA 93405, USA
E-mail address: [email protected]
|
1811.10584 | 1 | 1811 | 2018-11-26T18:41:25 | Dimension theory in iterated local skew power series rings | [
"math.RA",
"math.QA"
] | By analysing the structure of the associated graded ring with respect to certain filtrations, we deduce a number of good properties of iterated local skew power series rings over appropriate base rings. In particular, we calculate the Krull dimension, prime length and global dimension in well-behaved cases, obtaining lower bounds to complement the upper bounds obtained by Venjakob and Wang. | math.RA | math | Dimension theory in iterated local skew power series rings
William Woods
November 27, 2018
Abstract
By analysing the structure of the associated graded ring with respect to certain filtrations, we deduce a
number of good properties of iterated local skew power series rings over appropriate base rings. In particular,
we calculate the Krull dimension, prime length and global dimension in well-behaved cases, obtaining lower
bounds to complement the upper bounds obtained by Venjakob and Wang.
8
1
0
2
v
o
N
6
2
]
.
A
R
h
t
a
m
[
1
v
4
8
5
0
1
.
1
1
8
1
:
v
i
X
r
a
Contents
1 Preliminaries
1.1 Definitions and first results
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Local skew power series extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Filtered rings
2.1 Constructing skew derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Filtrations and their associated graded rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Lifting properties from the graded ring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
3
4
5
5
6
7
3 Rigid iterated extensions
8
8
3.1 Quotients by stable ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
3.2 Skew power series presentations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
3.4 Non-examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4 Dimension theory
13
4.1 Krull dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Classical Krull dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3 Global dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1
Introduction
The main objects of study of this paper are iterated skew power series rings
R = k[[x1; σ1, δ1]] . . . [[xn; σn, δn]]
for suitable choices of k and (σi, δi). By this, we mean that R0 = k, and we iteratively define for each 1 ≤ i ≤ n
((σi, δi) is a skew derivation on Ri−1,
Ri is the skew power series ring Ri−1[[xi; σi, δi]],
(0.1)
and finally set R = Rn. Full definitions of these objects are given below.
Skew power series extensions B = A[[x; σ, δ]] were first introduced in [18] and [20], with a view to using them
to study Iwasawa algebras: it was noted in [18] that certain Iwasawa algebras can be written as iterated skew
power series extensions over a base field (usually Fp) or complete discrete valuation ring (usually Zp).
Modelled on the notion of a skew polynomial extension A[x; σ, δ], a skew power series extension A[[x; σ, δ]] is a
ring which is isomorphic to A[[x]] as topological left (or right) A-modules, with multiplication given by
xa = σ(a)x + δ(a)
(0.2)
for all a ∈ A. Here, σ is an automorphism of A, and δ is a σ-derivation of A, i.e. a linear map δ : A → A
satisfying δ(ab) = δ(a)b + σ(a)δ(b). (We say that (σ, δ) is a skew derivation for short.)
Unlike the case of skew polynomial extensions, it is not the case that an arbitrary skew derivation (σ, δ) on
A gives rise to a well-defined ring A[[x; σ, δ]], due to possible convergence issues. For this reason, we work
throughout the paper with a special class of skew power series extensions whose question of existence is already
well understood. We denote this class of rings by SPS(A), and call them local skew power series extensions:
we require A to have a unique maximal ideal m, with respect to which it is complete (and separated), and we
stipulate that σ(m) ⊆ m, δ(R) ⊆ m and δ(m) ⊆ m2. These are known to exist, and while other classes have been
studied (see e.g. [2] and [9, §5], dealing with the rather different case in which δ is assumed locally nilpotent ),
the local case encompasses the case of interest in the Iwasawa context.
There is a growing literature on simple power series extensions S = R[[x; σ, δ]]: see, for instance, [19] and [13]
for important milestones in this theory. However, very little work has been done to treat the iterated case so far.
The most basic noncommutative case, that of q-commutative power series rings (completed quantum planes),
has been studied in [14]. A few properties of general iterated local skew power series rings are established
in [21] (though see the remarks before Theorem D below for a caveat). The author is not aware of any other
developments in the theory.
As above, we restrict our work to the local case, and denote the class of n-fold iterated local skew power series
extensions of A by SPSn(A) (Definition 1.10 below). Even under rather general conditions like this, we can
understand much of the dimension theory of these rings: our first result, along these lines, complements the
work of Wang [21].
Theorem A. Let k be a field, and R ∈ SPSn(k). Then Kdim(R) = gldim(R) = n.
This result follows from Theorem 4.3, Theorem 4.4 and Corollary 4.10.
The process of iteration is often rather badly behaved. Indeed, in the notation of (0.1): when n ≥ i ≥ 2, there is
no guarantee that (σi, δi) restricts to a skew derivation of Rj for any 0 ≤ j ≤ i− 2. Hence we identify a subclass
of SPSn(A), which we denote RSPSn(A), and whose elements we call rigid, whose elements are well-behaved
in the above sense. See Definition 3.8 for the precise definition.
This appears to be a fairly restrictive condition: we expect that RSPSn(A) is in fact a much smaller class than
SPSn(A) except in trivial cases when n and A are very small. But, in fact many well-known rings do lie in
RSPSn(A), including q-commutative skew power series rings (Example 3.11), Iwasawa algebras of supersoluble
uniform groups (Examples 3.12 -- 3.13), and various other completed quantum algebras (Examples 3.14 -- 3.15).
We are interested in the prime ideal structure of iterated skew power series rings R ∈ SPSn(A). As a first step
towards a more general theory, we calculate the classical Krull dimension (prime length), as defined in [15], of
elements of RSPSn(A), and show that they obtain the upper bound given by the Krull dimension, similarly
to pure automorphic extensions, which are well studied and in some senses easier to understand. This perhaps
indicates that, in understanding the theory of prime ideals of iterated non-pure-automorphic skew power series
rings, a good first step will be to consider the rigid ones.
2
Theorem B. Let k be a division ring, and R ∈ SPSn(k). If R ∈ RSPSn(k), or R has pure automorphic type,
then clKdim(R) = n.
The proof of Theorem B is contained within Theorem 4.4 and Proposition 4.6.
So far we have not said much about the natural filtration on R, but this will be crucial for understanding its
ideals in future work. If R is an iterated local skew power series ring with unique maximal ideal m, we can
consider R as a complete m-adically filtered ring. We compute its associated graded ring as follows:
Theorem C. Let (R, m) be a complete local ring, (σ, δ) a local skew derivation on R, and S = R[[x; σ, δ]]
a local skew power series extension of R with unique maximal ideal n. Then there exists a skew deriva-
tion (σ, δ) of grm(R) such that the inclusion of graded rings grm(R) → grn(S) extends to an isomorphism
(grm(R))[X; σ, δ] ∼= grn(S) upon mapping X to gr(x). Moreover, if δ(m) ⊆ m3, then δ = 0.
The proof is given in Lemma 2.3.
We give Example 2.5 to show that δ can indeed be nonzero: this corrects a small error in [20, Proposition
2.9], [21, 2.3(iii)] and [13, 3.10(i)], and a similar error can be found in [18, Lemma 1.4(iv)]. This phenomenon
also arises in practice when studying the Iwasawa algebras of certain non-uniform 2-valuable groups. However,
we can get rid of δ by modifying the filtration, which we prove as Lemma 2.6:
Theorem D. Let R and S be as in Theorem C. Then there exists a filtration f of S, cofinal with the n-adic filtra-
tion, such that the inclusion of graded rings grf (R) → grf (S) extends to an isomorphism (grf (R))[X; σ] ∼= grf (S)
upon mapping X to gr(x).
With this in mind, we note in §2.3 that the results of [21] hold even in cases such as Example 2.5.
Finally, we note some basic structural results that seem to be absent from the literature.
Theorem E. Let R and S be as in Theorem C, and let n be the maximal ideal of S. Let f be either the n-adic
filtration or the filtration of Theorem D. Then:
(i) If grf (R) is prime, then grf (S) is prime, and hence S is prime.
(ii) If grf (R) is a maximal order, then grf (S) is a maximal order, and hence S is a maximal order.
(iii) The natural inclusion map R → S is faithfully flat.
Part (i) is proved as Proposition 2.7, part (ii) as Lemma 2.9 and part (iii) as Lemma 1.11.
Notation and conventions
We often talk about topological rings: unless specified, the filtration on a local ring R with unique maximal
ideal m is the m-adic filtration. All filtrations considered are assumed to be discrete, positive and separated.
When we need to specify a filtration explicitly, we will usually denote it as a function f : R → N ∪ {∞}.
When the derivation δ is zero, we will write the skew polynomial ring R[x; σ, δ] (resp. the skew power series
ring R[[x; σ, δ]]) as R[x; σ] (resp. R[[x; σ]]). Such rings are said to have (pure) automorphic type.
Acknowledgements
I am very grateful to K. A. Brown for some interesting discussions and his extensive comments on an early
draft of this paper. I also gratefully acknowledge a helpful discussion with Adam Jones about soluble Iwasawa
algebras; Example 3.13(ii) and Non-example 3.17 are in large part due to him.
1 Preliminaries
1.1 Definitions and first results
Write A for the left R-module of left formal power series
ri ∈ R)
A =( ∞
Xi=0
rixi (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
3
over R. Suppose the multiplication in R, together with the multiplication rule (0.2), extends to a well-defined
right R-module structure on A. Then there is a unique R-bimodule map A ⊗R
A → A such that r ⊗ s 7→ rs
for all r, s ∈ R and xi ⊗ xj 7→ xi+j ; and A becomes a ring under this multiplication. In this case, we write
R[[x; σ, δ]] := A for this ring. Note that R[[x; σ, δ]] contains R, and is a skew power series ring over R as
described above by definition.
When A has a right R-module structure in this way, we will say as shorthand simply that the ring R[[x; σ, δ]]
exists, and otherwise that R[[x; σ, δ]] does not exist.
The following computational lemma will be useful.
Lemma 1.1. Fix a positive integer n. Then, for a ∈ R, we have
xna = Xm∈Mn
m(σ, δ)(a)xe(m)
inside the skew polynomial ring R[x; σ, δ] or (if it exists) the skew power series ring R[[x; σ, δ]]. Here, Mn is the
set of formal (noncommutative) monomials m = m(X, Y ) of degree n in the variables X and Y , and e(m) is
the total degree of X in the monomial m.
Proof. When n = 1, this is just the multiplication rule (0.2) in S. For n > 1, this follows by an easy induction:
note that all elements of Mn are either of the form Xm(X, Y ) or of the form Y m(X, Y ) for some m ∈ Mn−1.
1.2 Local skew power series extensions
There are several variant notions of "locality" in the literature; for the purposes of this paper, we fix the
following definition.
Definition 1.2. Let R be a ring. Then R is local if it has a unique maximal two-sided ideal m, and R is scalar
local ring if additionally m is maximal among right (and hence also left ) ideals. (It is easy to see that scalar
local implies local, but that this implication is not reversible.)
An alternative formulation of this definition, which is sometimes useful, goes as follows. Let R be a ring with
unique maximal ideal m. Then R is local (resp. scalar local ) if R/m is simple artinian (resp. a division ring).
We will sometimes say that (R, m) is a local (resp. scalar local ) ring, as shorthand for "R is a local (resp. scalar
local) ring with unique maximal ideal m".
The residue ring of a local ring (R, m) is defined to be the quotient R/m.
Remark 1.3. Let R be a ring with unique maximal left ideal m. Then it is well known that m is also the unique
maximal right ideal, and is hence equal to the Jacobson radical of R.
The applicability of our results is guaranteed by the lemma below.
Definition 1.4. Let (R, m) be a complete local ring. We say that (σ, δ) is a skew derivation of (R, m) (or just
R) if σ is an automorphism of R and δ is a (left) σ-derivation of R. We say that the skew derivation (σ, δ) is
local if σ(m) ⊆ m, δ(R) ⊆ m and δ(m) ⊆ m2.
Lemma 1.5. [20, Lemma 2.1] Let (R, m) be a complete local ring, and (σ, δ) a local skew derivation of R. Then
the ring R[[x; σ, δ]] exists.
We do not deal with the question of existence any further in this paper; all rings written are implicitly assumed
to exist.
Definition 1.6. Let (R, m) be a local ring. We will say that the ring S is a local skew power series ring over
R, or a local skew power series extension of R, if it satisfies the following properties:
(i) S is a skew power series ring over R with respect to some automorphism σ and some σ-derivation δ, say
S = R[[x; σ, δ]];
(ii) σ(m) ⊆ m;
(iii) δ(R) ⊆ m and δ(m) ⊆ m2.
Remark 1.7. Condition (iii) implies that δ(mn) ⊆ mn+1 for all n ∈ N. Hence, taken together, conditions (ii)
and (iii) are simply the natural requirement that the multiplication data σ, δ of R should interact "nicely" with
the m-adic topology on R, i.e. σ should be a homeomorphism, and δ should be topologically nilpotent.
4
As the hypotheses above are cumbersome to repeat, we set up some shorthand notation that will remain in
force for the rest of the paper.
Definition 1.8. Let (R, m) be a complete local ring. Then we will write S ∈ SPS(R) to mean that S is a
local skew power series ring R[[x; σ, δ]] over R. If σ and δ are A-linear (for some subring A ⊆ R), we may write
S ∈ SPSA(R) to emphasise this fact. We may also choose to specify the maximal ideals of R or S explicitly,
e.g. by writing (S, n) ∈ SPS(R, m) to mean that m and n are the maximal ideals of R and S respectively.
We justify the name "local skew power series ring":
Lemma 1.9. Let (R, m) be a complete local ring, and let S = R[[x; σ, δ]] ∈ SPS(R).
S/n ∼= R/m of residue rings.
(i) n = m + xR is the unique maximal ideal of S, and the natural inclusion R ⊆ S induces an isomorphism
(ii) (S, n) is local. Moreover, (S, n) is scalar local if and only if (R, m) is scalar local.
(iii) S is n-adically complete.
Proof. (i) and (iii) follow from [20, §2]. (ii) follows from (i) and the remark immediately after Definition 1.2.
Definition 1.10. Let (R, m) be a complete local ring. Then, for each positive integer n, we will inductively
define the notation S ∈ SPSn(R) as follows: S ∈ SPS1(R) if and only if S ∈ SPS(R); and, for n > 1, we write
S ∈ SPSn(R) if and only if S ∈ SPS(T ) for some T ∈ SPSn−1(R).
It will be convenient to have several explicit ways of writing a ring S ∈ SPSn(R). We will use either of the
following equivalent forms interchangeably:
• S = R[[x1; σ1, δ1]] . . . [[xn; σn, δn]],
• S = R[[x; σ, δ]] when the number n is understood; and, in this case, x = (x1, . . . , xn), σ = (σ1, . . . , σn),
and δ = (δ1, . . . , δn).
A(R) if all the σi and δi are A-linear.
As before, we may write S ∈ SPSn
In this way, the expression SPSn
series extensions of R".
If S ∈ SPSn(R), the number n is called the rank of S (over R). (Later, we will see that this number can often
be recovered from the global dimensions or Krull dimensions of R and S.)
A(R) may be read as "the class of n-fold iterated A-linear local skew power
Lemma 1.11. S is a faithfully flat R-module.
Proof. As a left R-module, we have S ∼= Q∞
i=0 Rxi, a direct product of flat R-modules, which is flat by [7,
Theorem 2.1], as R is noetherian. We may now conclude that S is faithfully flat as a left R-module using [15,
Proposition 7.2.3]: given any proper right ideal I of R, we have IS ⊆ mS ⊆ n 6= S. (The same holds for S as a
right module by a symmetric argument.)
2 Filtered rings
2.1 Constructing skew derivations
In the rare case when we wish to construct a local skew derivation on a ring R, the following lemma gives us a
method of doing so.
Let k be a fixed local ring, and R ∈ SPSn
automorphism of R (necessarily preserving m).
Lemma 2.1. Given any choice of b1, . . . , bn ∈ m2, the assignment d(xi) = bi for all 1 ≤ i ≤ n extends to a
unique local k-linear τ -derivation d of R.
k (k), say R = k[[x; σ, δ]] with maximal ideal m. Let τ be a fixed
Proof. First, as the elements x1, . . . , xn are k-linearly independent, the mapping xi 7→ bi (for all 1 ≤ i ≤ n)
extends uniquely to a k-linear map d : kx1 ⊕ ··· ⊕ kxn → m2.
To extend d to a k-module homomorphism R → R, it will suffice to define d(m) for each ordered monomial m
in the elements x1, . . . , xn. We do this inductively on the degree of the monomial as follows. If the monomial
m is of degree 1, then m ∈ {x1, . . . , xn}, and the value of d(m) is already known. Proceeding inductively, let
5
n > 1: if m is a monomial of degree n, then it may be written uniquely as xjm′, where m′ is a monomial of
degree n − 1 in xj , xj+1, . . . , xn. Then we will define d(m) := d(xj )m′ + τ (xj )d(m′).
To show that this is indeed a τ -derivation, we must show that the value of d(m) is well-defined regardless of
how it is computed. More precisely: let m be an ordered monomial of total degree s,
m = xj1 xj2 . . . xjs ,
where j1 ≤ j2 ≤ ··· ≤ js. Then, for each 1 ≤ r < s, the monomial m can be written as the product m = prqr,
where pr = xj1 xj2 . . . xjr and qr = xjr+1 xjr+2 . . . xjs . Then we may define:
dr(m) = d(pr)qr + τ (pr)d(qr).
Note that d(m) was defined to be d1(m). In this notation, we must show that d1(m) = ··· = ds−1(m).
We will do this by induction on s. There is nothing to check when s = 2. Now suppose that, for some N , we
have established d1(m) = ··· = dt−1(m) for all monomials m of total degree 2 ≤ t ≤ N . Take a monomial m of
total degree N + 1: in the above notation, we will write it as m = xj1 xj2 . . . xjN +1 = prqr for each 1 ≤ r ≤ N .
Fix 1 ≤ r ≤ N − 1. Then m = prqr = pr+1qr+1, where qr = xr+1qr+1 and pr+1 = prxr+1, i.e.
So, exactly as above, we may calculate
m = pr xr+1qr+1
qr+1.
= prxr+1
pr+1
{z }
}
qr
{z
dr(m) = d(pr)qr + τ (pr)d(qr)
= d(pr)xr+1qr+1 + τ (pr)d(xr+1qr+1)
= d(pr)xr+1qr+1 + τ (pr)[d(xr+1)qr+1 + τ (xr+1)d(qr+1)]
= [d(pr)xr+1 + τ (pr)d(xr+1)]qr+1 + τ (prxr+1)d(qr+1)
= d(prxr+1)qr+1 + τ (prxr+1)d(qr+1)
= d(pr+1)qr+1 + τ (pr+1)d(qr+1)
= dr+1(m).
Hence these di are all equal, and in particular are equal to d. This shows that d is a τ -derivation as required.
2.2 Filtrations and their associated graded rings
Let k be a complete local ring and R ∈ SPSn(k). Then R is also a complete local ring, say with maximal
ideal m. Studying the m-adic filtration and its associated graded ring is the key to understanding many basic
properties of R.
and can be recovered as n = dimk(m/m2).
Remarks 2.2.
(i) When k is a field, it is easy to see that the rank n of (R, m) ∈ SPSn(k) over a field is uniquely defined,
(ii) Suppose A is a central subring of R and S ∈ SPSn(R). Then S ∈ SPSn
A(R) if and only if A is a central
subring of S (under the natural inclusion A ⊆ R ⊆ S).
(iii) Let (R, m) be a complete local ring, and (S, n) ∈ SPSn(R). Then the restriction of the n-adic valuation
to R is the m-adic valuation, i.e. R ∩ ni = mi. In particular, m is generated in n-adic degree 1.
(iv) Let (R, m) be a complete local ring, and (S, n) ∈ SPSn(R, m). We will always write gr(R) for the graded
ring associated to the m-adic filtration, and likewise gr(S) for the graded ring associated to the n-adic
filtration.
We now calculate this associated graded ring inductively.
Let R be a complete local ring with maximal ideal m, take a local skew derivation (σ, δ) of R, and form
S = R[[x; σ, δ]] ∈ SPS(R) with maximal ideal n. Note that σ and δ induce linear endomorphisms of the graded
ring gr(R), of degrees 0 and 1 respectively, as follows: for all x ∈ mλ \ mλ+1, where λ ∈ N, we have
• σ(x + mλ+1) = σ(x) + mλ+1,
• δ(x + mλ+1) = δ(x) + mλ+2.
It is easy to check that σ is in fact a graded automorphism.
6
Lemma 2.3. The unique homomorphism of graded rings (gr(R))[z; σ, δ] → gr(S), extending the natural inclu-
sion map gr(R) → gr(S) and sending z to x + m2, is an isomorphism of graded rings.
Proof. Clear.
Remark 2.4. It is easy to see that δ = 0 if
δ(v(x)) > v(x) + 1
(2.1)
for all nonzero x ∈ R, where v is the m-adic filtration: v(x) = λ if x ∈ mλ \ mλ+1. We give an example for
which δ 6= 0.
Example 2.5. Take R = k[[x]] with skew derivation σ = id and δ(x) = x2: this is indeed a skew derivation by
e.g. 2.1. Form S = R[[y; id, δ]]. It is easy to see that the graded ring has nonzero δ. Set X = x + n2 and
Y = y + n2 inside gr(S): then the multiplication in gr(S) is determined by the rule
Y X = XY + X 2,
i.e. gr(S) ∼= k[X][Y ; id, δ], where δ(X) = X 2.
Retain the above notation, and continue to write v for the m-adic filtration on R.
Lemma 2.6. The function f : S → R ∪ {∞} defined by
f ∞
Xi=0
rixi! = inf{2v(ri) + i}
is a Zariskian ring filtration on S, and the associated graded ring is
grf (S) ∼= (grf (R))[z; σ].
Proof. Write r = P∞
f (xisj) = f (σi(sj)xi) and that f (xisj − σi(sj )xi) > f (xisj). The claims of the lemma now follow easily.
j=0 sjxj, for ri, sj ∈ R. An application of Lemma 1.1 shows that
i=0 rixi and s = P∞
2.3 Lifting properties from the graded ring
Set f1 to be the n-adic filtration on S, and f2 to be the filtration obtained in Lemma 2.6.
A few preliminary results are given for the rings S ∈ SPSn(R), for R a complete local ring, in [21, Corollary
2.9]:
(a) If grf1 (R) is a domain (resp. right noetherian, resp. Auslander regular), then so is grf1 (S), and hence so
is S.
(b) If grf1 (R) is right noetherian, then r.Kdim(S) ≤ r.Kdim(grf1 (R))+n and r.gldim(S) ≤ r.gldim(grf1(R))+n.
The proofs in [21] were given under the erroneous assumption that δ = 0 (in the notation of Lemma 2.3), but
they remain true with essentially identical proofs even after removing this assumption. (We note that they also
hold on replacing f1 by f2, of course.)
Throughout this subsection, we fix either f = f1 or f = f2, giving a filtration f on R and S. The expressions
gr(R) and gr(S) will always implicitly mean grf (R) and grf (S) respectively.
We record some further important properties that lift from the graded ring.
Proposition 2.7. Let R be a complete local ring, and S ∈ SPSn(R). If gr(R) is prime, then gr(S) is prime,
and hence S is prime.
Proof. If gr(R) is prime, then gr(S) is prime by Remark 2.2(iv) along with iterated application of [15, Theorem
1.2.9(iii)]. Hence S is prime by [11, II, Lemma 3.2.7].
Corollary 2.8. Let R be a complete local ring, and S ∈ SPSn(R). Suppose that gr(R) is Auslander regular.
Then S is AS-Gorenstein.
Proof. S is Auslander regular by [21, 2.3(iii)], and so in particular S is Auslander-Gorenstein. As S is scalar
local by Lemma 1.9(ii), we deduce from [6, Lemma 4.3] that S is AS-Gorenstein.
7
Lemma 2.9. Let R be a complete local ring such that gr(R) is a maximal order, and take S ∈ SPSn(R). Then
S is a maximal order.
Proof. This follows from [11, Ch. II, §3, Theorem 11] alongside the above, similarly to the proof of [11, Ch. II,
§3, Corollary 12].
3 Rigid iterated extensions
Let (R0, m0) be a complete local ring, and let S ∈ SPSn(R0) for n ≥ 2. That is, for each 1 ≤ i ≤ n, we may
inductively find Ri ∈ SPS(Ri−1), say Ri = Ri−1[[xi; σi, δi]] with maximal ideal mi, such that S = Rn. For
convenience, write σ = σn, δ = δn.
We noted in Remark 1.7 that we wanted each Ri to interact "nicely" with the topology of Ri−1, i.e. (σi, δi)
should be a local skew derivation of (Ri−1, mi−1) (see Definition 1.6(ii -- iii)). Unfortunately, naively iterating
this procedure as in Definition 1.10, we may end up in a situation like the following. Writing the above chain
of local rings explicitly,
(R0, m0) (cid:12) (R1, m1) (cid:12) ··· (cid:12) (Rn−1, mn−1) (cid:12) S,
and (σ, δ) is a local skew derivation of Rn−1 as desired, but we are not guaranteed that it restricts to a local
skew derivation (σRi , δRi ) of Ri for any 0 ≤ i ≤ n − 2, as the following example shows.
Example 3.1. Let k be a field. Let R0 = k[[Y ]] (cid:12) k[[Y, Z]] = R1, so that R0 has maximal ideal m0 = (Y ). Take
S = R[[X; σ, δ]], where σ is the k-linear automorphism of R given by σ(Y ) = Z, σ(Z) = −Y , and δ = 0. Then
S ∈ SPS2
This turns out to be a very natural stipulation to make when performing this iterative construction, and this
motivates the definitions we make in this section.
k(R0), but σ(m0) 6⊆ m0.
3.1 Quotients by stable ideals
Recall the following basic definition (cf. [13, 3.13]).
Definition 3.2. Let R be a ring and I an ideal. If (σ, δ) is a skew derivation of R, then I is said to be a
(σ, δ)-ideal if σ(I) ⊆ I and δ(I) ⊆ I.
This is a useful class of ideals because of results such as the following lemma:
Lemma 3.3. Let R be a complete local ring, and S = R[[x; σ, δ]] ∈ SPS(R). Suppose that I is a (σ, δ)-ideal
of R. Then IS = SI is a two-sided ideal, and S/IS ∼= (R/I)[[x; σ, δ]] ∈ SPS(R/I).
Proof. Note that [13, Setup 3.1(4)] is satisfied by Remark 2.2(iv). Hence we may apply [13, 3.13(ii) and Lemma
3.14(iv)].
We slightly extend this notion as follows.
Definition 3.4. Let R be a ring and I an ideal. Let σ = (σ1, . . . , σr) and δ = (δ1, . . . , δr), where each (σi, δi)
is a skew derivation of R. We will say that I is a (σ, δ)-ideal if it is a (σi, δi)-ideal for each 1 ≤ i ≤ r.
Proposition 3.5. Let R be a complete local ring, and S = R[[x; σ, δ]] ∈ SPSn(R). Suppose that I is a
(σ, δ)-ideal of R. Then IS = SI is a two-sided ideal, and S/IS ∼= (R/I)[[x; σ, δ]] ∈ SPSn(R/I).
Proof. Both claims follow from recursive application of Lemma 3.3.
3.2 Skew power series presentations
Definition 3.6. Fix a complete local ring (R0, m0) and a ring R ∈ SPSn(R0). A presentation for R (over R0)
is a sequence of rings
(3.1)
where Ri ∈ SPSdi(Ri−1) for each 1 ≤ i ≤ ℓ, and d1 + ··· + dℓ = n. The number ℓ is called the length of the
presentation.
R0 (cid:12) R1 (cid:12) ··· (cid:12) Rℓ−1 (cid:12) Rℓ = R,
8
Remark 3.7. Note that, by definition, each di ≥ 1, so the length ℓ of any presentation (3.1) must be bounded
above by the rank n of R over R0.
Definition 3.8. With R ∈ SPSn(R0) as above, denote R = R0[[x; σ, δ]]. We will say that the presentation
(3.1) is stable if the maximal ideal of Ri is a (σj, δj)-ideal for each pair (i, j) satisfying 0 ≤ i < j ≤ ℓ.
Definition 3.9. We say that R ∈ SPSn(R0) is rigid if it admits a stable presentation of length n, i.e. a
saturated presentation which is also stable, and we will denote this by R ∈ RSPSn(R0). (As before, we will
write RSPSn
Proposition 3.10. Let R0 be a complete local ring with residue ring k, and take some R ∈ RSPSn(R0), say
with stable presentation (R0, m0) (cid:12) (R1, m1) (cid:12) ··· (cid:12) (Rn, mn) = (R, m). Then R/miR ∈ RSPSn−i(k).
Proof. When i = 0, we may apply Proposition 3.5, and the result is immediate.
A(R0) ∩ RSPSn(R0).)
A(R0) = SPSn
When i > 0, set R′
R as an element of RSPSn−i(R′
where R′
t = Rt+i and m
0 = Ri, so that R′
′
0 is a complete local ring with residue ring k by Lemma 1.9. We may view
n−i) = R,
n−i, m
1, m
t = mt+i. Now the result follows from the previous case.
0), with stable presentation (R′
0) (cid:12) (R′
0, m
′
′
1) (cid:12) ··· (cid:12) (R′
′
3.3 Examples
It turns out that many interesting iterated local skew power series extensions arising in nature are indeed rigid.
Examples 3.11 -- 3.13 below were important motivating examples for this paper, and will be referred to later.
Example 3.11. q-commutative power series rings. Let k be a field, and q ∈ Mn(k×) a multiplicatively
antisymmetric matrix (i.e. qijqji = 1 for all i and j). Then the q-commutative power series ring is the ring
R = kq[[x1, . . . , xn]],
defined as follows: R ∼= k[[x1, . . . , xn]] as topological k-modules; and the multiplication is given by the n2
relations xixj = qij xjxi. It is easy to see that R ∈ RSPSn
For the following two examples, we assume familiarity with the notion of uniform groups [8] and completed
group rings (also known as Iwasawa algebras) [12].
k (k).
Example 3.12. Nilpotent Iwasawa algebras. Let G be a nilpotent uniform group. G admits a series
{1} = G0 ≤ G1 ≤ ··· ≤ Gn = G
of closed subgroups of G such that Gi−1 is normal in G and Gi/Gi−1 ∼= Zp for all 1 ≤ i ≤ n: an appropriate
refinement of the isolated lower central series, defined in [22, Definition 3.7], will suffice. Hence G is supersoluble,
as defined below. Now it will follow from Example 3.13 below that its appropriate Iwasawa algebras are rigid.
Example 3.13. Supersoluble Iwasawa algebras. Let G be a supersoluble uniform group, i.e. there exists a
sequence
of closed subgroups of G such that Gi−1 is normal in G and Gi/Gi−1 ∼= Zp for all 1 ≤ i ≤ n. Fix such a
sequence. Let k be either the ring of integers of a finite extension of Qp or a finite field extension of Fp. Now,
as in [20, Example 2.3], kG ∈ SPSn
k (k) with presentation
{1} = G0 ≤ G1 ≤ ··· ≤ Gn = G
k = kG0 ≤ kG1 ≤ ··· ≤ kGn = kG,
and the condition that Gi−1 should be normal in G ensures that this presentation is stable. Hence kG ∈ RSPSn
We demonstrate the existence of some supersoluble, non-nilpotent uniform groups of small rank below. In both
cases, uniformity can be easily checked using [8, Theorem 4.5].
(i) Let Γ1 and Γ2 be copies of Zp, and take the continuous group homomorphism ρ : Γ2 → Aut(Γ1) ∼= Z×
p
p . Form the semidirect product G = Γ1 ⋊ Γ2. Then G is a
sending a generator of Γ2 to 1 + p ∈ Z×
supersoluble but non-nilpotent uniform group of dimension 2. (cf. [20, Example 2.2])
k (k).
(ii) Take p to be a prime congruent to 1 mod 4, so that i := √−1 ∈ k. Let A = hy, zi ∼= Z2
p and B = hxi ∼= Zp,
and fix the left action of B on A, say ρ : B → Aut(A), defined by ρ(x)(y) = yzp and ρ(x)(z) = zy−p.
Then G = B ⋊ A is easily checked to be a soluble, non-nilpotent uniform group of dimension 3, and the
chain of normal subgroups
shows that G is in fact supersoluble. (Compare Non-example 3.17.)
1 ≤ hyzii ≤ B ≤ A
9
The next two examples are not crucial for the current paper, but we include them to illustrate the wide
applicability of Theorems B and D.
Example 3.14. Completed quantised k-algebras. Let k be a field, Γ = (γij) ∈ Mn(k×) a multiplicatively
antisymmetric matrix, and P = (p1, . . . , pn) ∈ (k×)n, Q = (q1, . . . , qn) ∈ (k×)n two vectors with pi 6= qi for
all 1 ≤ i ≤ n. Horton's algebra R = K P,Q
n,Γ (k), a simultaneous generalisation of quantum symplectic space and
quantum Euclidean 2n-space, can be written as an iterated skew polynomial ring,
k[x1][y1; τ1][x2; σ2][y2; τ2, δ2] . . . [xn; σn][yn; τn, δn],
where the σi (for 2 ≤ i ≤ n) and τi (for 1 ≤ i ≤ n) are k-linear automorphisms, and each δi (for 2 (cid:12) i (cid:12) n) is a k-
linear τi-derivation. In [21, §3.2], it is proved that the I-adic completion R of R, where I = (x1, y1, x2, y2, . . . , xn, yn),
is an iterated skew power series extension of k:
R = k[[x1]][[y1; τ1]][[x2; σ2]][[y2; τ2, δ2]] . . . [[xn; σn]][[yn; τn, δn]] ∈ SPS2n
k (k).
(3.2)
We do not spell out the relations in full: see [10, Proposition 3.5] for details.
purposes, to know the following:
It is only necessary, for our
• σi(xj ), τi(xj ), δi(xj ) are scalar multiples of xj for all j < i;
• σi(yj), τi(yj), δi(yj) are scalar multiples of yj for all j < i;
• τi(xi) is a scalar multiple of xi for all i;
• δi(xi) is a k-linear combination of the elements ylxl for all l < i.
k (k).
It now follows by an easy calculation that the saturated presentation associated to (3.2) is stable, and hence
that R ∈ RSPS2n
Example 3.15. Completed quantum matrix algebras. Let k be a field, λ ∈ k× a scalar, and p = (pij) ∈ Mn(k×)
a multiplicatively antisymmetric matrix (i.e. pijpji = 1 for all i, j). Then the multiparameter quantum n × n
matrix algebra R = Oλ,p(Mn(k)) can be defined (see e.g. [3, Definition I.2.2]) as a skew polynomial ring in n2
variables labelled Xi,j for each 1 ≤ i, j ≤ n, in which k is central. Again, we do not spell out the relations in
full, but we note:
• Xl,mXi,j is a linear combination of Xi,jXl,m and Xi,mXl,j when l > i and m > j;
• Xl,mXi,j is a scalar multiple of Xi,jXl,m whenever either l ≤ i or j ≤ m.
In [21, §3.2] it is proved that the I-adic completion R of R, where I is the ideal generated by the n2 variables
Xi,j for 1 ≤ i, j ≤ n, is an iterated skew power series extension of k satisfying the same relations: R ∈ SPSn2
k (k).
But the "obvious" saturated presentations, e.g. those associated to
k[[X1,1]][[X1,2]][[X1,3]] . . . [[Xn,n]]
(where we have omitted the skew derivations for readability) are usually not stable. (For a counterexample,
take R to be the usual quantum 2 × 2 matrix algebra Oq(M2(k)), given by n = 2, p =(cid:16) 1 q
for any q ∈ k×. See [3, Definition I.1.7]) for the relations in this case.)
We fix this by adjoining the variables in the following order:
q−1 1(cid:17), and λ = q−2
• at the 0th stage, adjoin the "antidiagonal" elements Xi,j satisfying i + j − (n + 1) = 0, in any order;
• at the 1st stage, adjoin those Xi,j satisfying i + j − (n + 1) = 1;
• at the 2nd stage, adjoin those Xi,j satisfying i + j − (n + 1) = 2;
and so on, until finally all variables have been adjoined at the end of the (n − 1)th stage.
Diagrammatically:
⊛
. . .
⊛
⊛
⊛
0th stage
. . .
⊛ ∗
⊛ ∗ ⊛
⊛ ∗ ⊛
∗ ⊛
1st stage
. . .
. . .
. . .
. . .
⊛ ∗
∗
∗
⊛ ∗
∗
∗
∗
∗ ⊛
∗
∗ ⊛
2nd stage
. . .
. . .
. . .
. . .
It is easy to verify that such a presentation of R is stable. The only case in which there is anything to prove is
when l > i and m > j: multiplying Xl,m by Xi,j results in a term involving the variables Xi,m and Xl,j, and we
must check that each of the variables Xi,m and Xl,j has been adjoined before we adjoin both of Xl,m and Xi,j.
10
More precisely, let M = max{i + j, l + m} − (n + 1) be the first stage after which both Xl,m and Xi,j have
been adjoined, and likewise let N = max{i + m, l + j} − (n + 1) be the first stage after which both Xi,m and
Xl,j have been adjoined. Then it is easy to see that the inequalities l > i and m > j imply
i + j < i + m < l + m,
i + j < l + j < l + m,
and hence N < M .
The existence of such a stable presentation shows that R ∈ RSPSn2
k (k).
3.4 Non-examples
Showing that an extension is not rigid appears to involve lots of tedious calculation, but we give two examples
which seem of interest to the theory.
Non-example 3.16. An extension of pure automorphic type. Let k be a field such that √−1 6∈ k, and let
S = k[[Y, Z]][[X; σ]], where σ is a k-linear automorphism of the commutative power series ring k[[Y, Z]] defined
by σ(Y ) = Z, σ(Z) = −Y : that is, ZX = XY and Y X = −XZ inside S.
By construction, S ∈ SPS3
Suppose that S is rigid, and denote its maximal ideal by n. Then there must exist a stable presentation
k(k).
k (cid:12) (R1, m1) (cid:12) (R2, m2) (cid:12) (S, n)
for S. We will show that no such stable presentation can exist. Write I = m1S and J = m2S: then, by Lemma
1.9(i), I = sS and J = sS + tS with s, t ∈ n, and s, t 6∈ n2 by Remark 2.2(iii).
For computation purposes, we will write
• s = aX + bY + cZ + ε,
• t = αX + βY + γZ + ε′,
where ε, ε′ ∈ n2, and a, b, c, α, β, γ ∈ k are constants, which must satisfy (a, b, c) 6= (0, 0, 0) 6= (α, β, γ) by
Remark 2.2(iii).
As S is rigid, I = m1S must in fact be a two-sided ideal. This places some restrictions on possible choices for
s, which we now compute.
Henceforth, we work in S/n3 for ease of computation. We have
Xs ≡ aX 2 + bXY + cXZ,
sX ≡ aX 2 + cXY − bXZ mod n
3,
and so
Xs − sX = (b − c)XY + (b + c)XZ mod n
3.
But Xs− sX ∈ I = sS, so we must have (b− c)XY + (c + b)XZ ≡ sα mod n3 for some α ∈ S. As the left-hand
side belongs to n2, we must have α ∈ n, and so εα ∈ n3. Writing therefore α ≡ dX X + dY Y + dZ Z mod n2 for
some dX , dY , dZ ∈ k, we see that
(b − c)XY + (b + c)XZ ≡ (aX + bY + cZ)(dX X + dY Y + dZ Z) mod n
3.
Multiplying out the right-hand side:
(b − c)XY + (b + c)XZ ≡ adX X 2 + (adY + cdX )XY + (adZ − bdX )XZ
+ bdY Y 2 + (bdZ + cdY )Y Z + cdZ Z 2
mod n
3.
Now, equating the coefficients of each monomial on both sides, some tedious case-checking shows that the only
solution to this congruence is b = c = dX = dY = dZ = 0.
Hence we have s = aX + ε, so that a 6= 0.
Now, as S is rigid, J = m2S must also be a two-sided ideal, and so we calculate the restrictions that this places
on t.
As R1 ∈ SPSk(k), it must be a commutative power series ring over k, so that R1 ∼= k[[s]]. Then R2 = k[[s]][[t; τ, δ]]
for some local skew derivation (τ, δ) of k[[s]], and an easy calculation shows that there must be a unit η ∈ k[[s]]×
11
such that τ (s) = ηs, and there must be an element θ ∈ k[[s]] such that δ(s) = θs2. (Note that η 6≡ 0 mod n2.)
Hence we have τ (s) ≡ ηs ≡ ηaX mod n2 and δ(s) ≡ θs2 ≡ a2θX 2 mod n3, and so
(αX + βY + γZ)ηaX ≡ ηaX(αX + βY + γZ) + a2θX 2 mod n
3.
Multiplying out again:
from which we see immediately that β = γ = θ = 0, and hence t = αX + ε′.
a2θX 2 + ηa(β − γ)XY + ηa(β + γ)XZ ≡ 0 mod n
3,
This implies that the images of s and t in n/n2, and hence also in m2/m2
dimk(m2/m2
dimk(m2/m2
2) ≤ 1. But R2 ∈ SPS2
2) = 2. This is a contradiction.
2, are linearly dependent, and so
k(k) by construction, and Remark 2.2(i) tells us that we should have
Non-example 3.17. A soluble Iwasawa algebra.
This example is similar to the previous in many ways. Compare also Example 3.13(ii).
Fix a prime p congruent to 3 mod 4, so that √−1 6∈ Fp. Let A = hy, zi ∼= Z2
semidirect product G = B ⋊ A as in Example 3.13(ii). Construct its Fp-Iwasawa algebra S = FpG.
Writing X = x − 1, Y = y − 1, Z = z − 1, we may easily calculate (see e.g. [20, Example 2.3]) that
p and B = hxi ∼= Zp, and form the
S = Fp[[Y, Z]][[X; σ, δ]],
where σ(Y ) = Y +(1+Y )Z p, σ(Z) = Z +(−Y +Y 2−Y 3 +. . . )p(1+Z), and δ = σ−id. Clearly, S ∈ SPS3
Suppose that S is rigid, and denote its maximal ideal by n. Then there must exist a stable presentation
Fp(Fp).
Fp (cid:12) (R1, m1) (cid:12) (R2, m2) (cid:12) S
for S: in particular, by Proposition 3.10, S must admit a quotient S/m1S ∈ RSPS2
this leads to a contradiction.
As in Non-example 3.16, take I = m1S = sS, and write s = aX + bY + cZ + ε, where ε ∈ n2, and a, b, c ∈ Fp.
By rigidity of S, I must be a two-sided ideal; this time, we calculate restrictions on s by working in S/np+1.
We have
Fp(Fp). We will show that
σ(Y ) ≡ Y + Z p,
σ(Z) ≡ Z − Y p, mod n
p+1,
and so we must have
Xs − sX ≡ bZ p − cY p mod n
p+1.
But Xs − sX ∈ I = sS, so we must have bZ p − cY p ≡ sα mod np+1 for some α ∈ S. As the left-hand side
belongs to np, we must have α ∈ np−1, and so εα ∈ np+1. So this equation becomes
dγX γ1Y γ2 Z γ3! mod n
bZ p − cY p ≡ (aX + bY + cZ) Xγ
p+1,
for some choices of dγ ∈ Fp, where this sum ranges over all γ = (γ1, γ2, γ3) with p − 1 ≤ γ1 + γ2 + γ3 ≤ p.
Suppose that, for our given a, b, c, we have a solution {dγ} to this congruence. Then, multiplying out the
right-hand side, and writing e1 = (1, 0, 0), e2 = (0, 1, 0), e3 = (0, 0, 1):
bZ p − cY p ≡Xγ ′
(adγ ′−e1 + bdγ ′−e2 + cdγ ′−e3 ) X γ ′
1Y γ ′
2 Z γ ′
3 mod n
p+1,
where for convenience we set dγ = 0 if any of the γi is equal to −1. We may eliminate any term in the sum
of total degree not equal to p, as all nonzero monomials appearing on the left hand side have degree p, so this
sum may be taken to range over all γ ′ = (γ ′
3 = p; furthermore, as there are no terms
in X γ ′
3 on the left hand side except those with γ1 = 0, we may also eliminate all monomials of nonzero
degree in X on the right. We can now rewrite this congruence as
3) with γ ′
2 + γ ′
1 + γ ′
1, γ ′
2, γ ′
1Y γ ′
2 Z γ ′
bZ p − cY p ≡
p
Xi=0(cid:0)bd(0,i−1,p−i) + cd(0,i,p−i−1)(cid:1) Y iZ p−i mod n
p+1.
It is easy to see that, if b 6= 0, then c 6= 0, and vice-versa. We assume for contradiction that b 6= 0 6= c: we will
show that, in this case, the above congruence cannot hold for any choice of {dγ}.
12
Indeed, equating monomial coefficients on the left and right hand sides:
b = cd(0,0,p−1)
0 = bd(0,m−1,p−m) + cd(0,m,p−m−1)
−c = bd(0,p−1,0)
(i = 0)
(i = m : 1 ≤ m ≤ p − 1)
(i = p).
On the one hand, multiplying the equations labelled (i = 0) and (i = p) together, we get
−bc = bcd(0,0,p−1)d(0,p−1,0),
i.e. d(0,0,p−1)d(0,p−1,0) = −1. On the other hand, multiplying the equation labelled (i = m) by bp−mcm and
rearranging for each 1 ≤ m ≤ p − 1, we get
substituting each one into the next, we eventually get
bp−m+1cmd(0,m−1,p−m) = −bp−mcm+1d(0,m,p−m−1);
bpcd(0,0,p−1) = bcpd(0,p−1,0)
as p is odd; and since b, c ∈ Fp, we have bp = b and cp = c. This tells us that d(0,0,p−1) = d(0,p−1,0), and denoting
this common value by d, we have shown that we must have d2 = −1 ∈ Fp, which is a contradiction.
Hence we have shown that b = c = 0, and so s = aX + ε. But now
sY − Y s ≡ aZ p mod n
p+1,
and a very similar (but easier) calculation shows that we must have a = 0. Hence s = ε ∈ n2: that is, m1 is
generated in n-adic degree ≥ 2. This contradicts Remark 2.2(iii).
4 Dimension theory
Many of the results in this section can be slightly extended; but we will not always strive for full generality, and
often work over a field or a division ring for simplicity.
4.1 Krull dimension
Let k be a division ring, and (R, m) ∈ SPSn(k). Remark 2.2(v)(c) implies that Kdim(R) ≤ n: in this subsection,
we show that this is always an equality.
For any ring R, write Ir(R) for the lattice of right ideals of R.
The following is the result corresponding to [13, 3.14(ii)] in the case when I is not a (σ, δ)-ideal.
Proposition 4.1. Let R be a complete local ring, S = R[[x; σ, δ]] ∈ SPS(R), and let I be a right ideal of R.
Set I[[x; σ, δ]] := {P aixi : ai ∈ I}. Then I[[x; σ, δ]] is a right ideal of S. Moreover, the map θ : Ir(R) → Ir(S)
sending I to I[[x; σ, δ]] is a strictly increasing poset map.
Proof. I[[x; σ, δ]] is clearly an additive subgroup of S.
Take r ∈ R and a =P∞
i=0 aixi ∈ I[[x; σ, δ]]. Then we may evaluate ar inside S:
ar =
=
∞
∞
aixir
Xi=0
Xi=0 Xm∈Mi
aim(σ, δ)(r)xe(m)
in the notation of Lemma 1.1. Now, m(σ, δ)(r) ∈ R, so aim(σ, δ)(r) ∈ I, and hence ar ∈ I[[x; σ, δ]].
We need to check that this gives a well-defined right S-action on I[[x; σ, δ]] -- in other words, that the right
actions of the elements xr and σ(r)x + δ(r) agree for all r ∈ R. But this is already true a fortiori, as the right
action of S on I[[x; σ, δ]] is just induced by multiplication inside the ring S.
Finally, it is clear that, if I1 ≤ I2, then θ(I1) ≤ θ(I2); and, if J = θ(I), then we may recover I as J/Jx. This
shows that θ is a strict map of posets.
13
Lemma 4.2. Let (R, m) be a complete local ring with residue ring k. Let Y ⊆ X be adjacent right ideals of R:
then, as right S-modules, θ(X)/θ(Y ) is isomorphic to S/mS ∼= k[[x; σ]], a skew power series ring of automorphic
type.
Proof. Let θ continue to denote the map Ir(R) → Ir(S) defined in Proposition 4.1.
Y ⊂ X are adjacent if and only if X/Y is a simple right module. In particular, the annihilator ann(X/Y )R
must be the unique maximal right ideal m of R, and so we must have X/Y ∼= R/m ∼= k.
The proposed isomorphism is obvious on the level of abelian groups, and it is easy to see that right multiplication
by x ∈ S is the same as right multiplication by x ∈ k[[x; σ]]. It remains to check the R-action. Recall, from
Lemma 1.1, that
∞
Xi=0
aixi! r =
∞
Xi=0 Xm∈Mi
aim(σ, δ)(r)xe(m)
for all ai ∈ X, r ∈ R. But, for a given m ∈ Mi, if m(σ, δ) contains an instance of δ (i.e.
m(σ, δ)(r) ∈ m, and hence aim(σ, δ)(r) ∈ X m ⊆ Y . This implies that
if e(m) < i), then
∞
Xi=0
aixi! r =
∞
Xi=0
aiσi(r)xi
inside X/Y , as required.
Theorem 4.3. Let k be a division ring, and S ∈ SPSn(k). Then Kdim(S) = Kdim(gr(S)) = n.
Proof. The proof of this theorem closely follows some of the methods of [1]. We will calculate the right Krull
dimension of S, but the calculation of the left Krull dimension is identical. Let θ continue to denote the map
Ir(R) → Ir(S) of Proposition 4.1.
When n = 0, we have S = gr(S) = k, and there is nothing to prove. We proceed by induction on the rank of S.
Let R ∈ SPSn−1(k), with S = R[[x; σ, δ]] ∈ SPS(R). By the inductive hypothesis, we know that
Kdim(R) = Kdim(gr(R)) = n − 1.
We also know, by [21, Corollary 2.9(ii)], that gr(S) ∼= gr(R)[x; σ], a skew polynomial ring of automorphic type,
and so Kdim(gr(S)) = Kdim(gr(R)) + 1 = n by [15, Proposition 6.5.4(i)].
Now, given arbitrary adjacent right ideals Y ⊆ X of R, we know already from Lemma 4.2 that θ(X)/θ(Y ) ∼= k[[x; σ]]
as right S-modules: in particular, θ(X)/θ(Y ) is not artinian as a right S-module, so we must have KdimS(θ(X)/θ(Y )) ≥ 1.
Now applying [1, Lemma 2.3], we see that n = Kdim(R) + 1 ≤ Kdim(S).
Together with the inequality of [21, Corollary 2.9(iv)], we see that
n ≤ Kdim(S) ≤ Kdim(gr(R)) + 1 = n,
and the result follows.
4.2 Classical Krull dimension
Theorem 4.4. Let k be a division ring, and R ∈ RSPSn(k). Then clKdim(R) = n.
Proof. It is well known that clKdim(R) ≤ Kdim(R) (e.g. [15, Lemma 6.4.5]), and here Kdim(R) = n by Theorem
4.3. To obtain a lower bound, consider the chain of ideals {miR}n
i=0 given in Proposition 3.10. This chain has
length n, and again by Proposition 3.10, the quotient rings are iterated local skew power series rings over k,
and are hence prime by Proposition 2.7. Hence clKdim(R) ≥ n.
Remark 4.5. Rigidity is a sufficient, but not a necessary, condition to have clKdim(R) = Kdim(R), as shown
by the following proposition together with Non-example 3.16.
Proposition 4.6. Let k be a division ring, and let R ∈ SPSn(k) have pure automorphic type. Then
clKdim(R) = n.
14
Proof. Suppose R = k[[x1; σ1]][[x2; σ2]] . . . [[xn; σn]]. Note that (xn) is a two-sided ideal of R, and
The result follows by induction on n.
R/(xn) ∼= k[[x1; σ1]][[x2; σ2]] . . . [[xn−1; σn−1]].
It is likely that, for many R ∈ SPSn(k), we have clKdim(R) < n. Of particular interest is the case where
k = Fp or k = Zp, and R = kG is the completed group algebra of a soluble p-valuable group G over k. Based
on the corresponding result for polycyclic group algebras [17, Theorem H1], we can ask the following question:
Question 4.7. Is clKdim(FpG) equal to the plinth length of G?
By Example 3.13, this is true in the case when G is supersoluble.
4.3 Global dimension
Theorem 4.8. Let R be a complete local ring and S ∈ SPS(R). Then Exti
groups for all i and all left S-modules M that are finitely generated as left R-modules.
R(M, R) ∼= Exti+1
S (M, S) as abelian
Proof. This follows from a similar argument to [16], after minor adjustments -- see e.g. [18, Theorem 3.1].
Corollary 4.9. Let k be a field, R ∈ SPSn(k), and S = R[[x; σ, δ]] ∈ SPS(R). Then gldim(S) = gldim(R) + 1.
Proof. We have augmentation maps R → k and S → k, so by [4, Corollary 3.7], we have gldim(R) = projdimR(k)
and gldim(S) = projdimS(k).
Suppose that projdimR(k) = n < ∞. Then by [5, VI, Ex. 9], we have Extn
4.8, Extn+1
Now assume (for contradiction) that projdimS(k) ≥ n+2. By the same argument, this gives us Extn+2
and hence Extn+1
projdimS(k) = n + 1 = projdimR(k) + 1.
(k, S) 6= 0,
R (k, R) 6= 0. But this implies that projdimR(k) ≥ n + 1, a contradiction. So we must have
(k, S) 6= 0, so that projdimS(k) ≥ n + 1.
S
R(k, R) 6= 0, and hence by Theorem
S
Finally, as projdimk(k) = 0, it is easy to see by induction on the rank of R (as an iterated local skew power
series ring over k) that we always have projdimR(k) < ∞.
Corollary 4.10. Let k be a field and R ∈ SPSn(k). Then gldim(R) = n.
References
[1] K. Ardakov. Krull dimension of Iwasawa algebras. J. Algebra, 280:190 -- 206, 2004.
[2] Jeffrey Bergen and Piotr Grzeszczuk. Skew power series rings of derivation type. Journal of Algebra and
its Applications, 10(6):1383 -- 1399, 2011.
[3] K.A. Brown and K.R. Goodearl. Lectures on Algebraic Quantum Groups. Advanced courses in mathematics
-- CRM Barcelona. Birkhauser, 2002.
[4] A. Brumer. Pseudocompact algebras, profinite groups and class formations. Journal of Algebra, 4:442 -- 470,
1966.
[5] Henri Cartan and Samuel Eilenberg. Homological Algebra. Princeton University Press, 1956.
[6] D. Chan, Q.-S. Wu, and J.J. Zhang. Pre-balanced dualizing complexes. Israel J. Math., 132(1):285 -- 314,
2002.
[7] S.U. Chase. Direct products of modules. Trans. Amer. Math. Soc., 97(3):457 -- 473, 1960.
[8] J.D. Dixon, M.P.F. du Sautoy, A. Mann, and D. Segal. Analytic Pro-p Groups. Cambridge University
Press, 1999.
[9] Be'eri Greenfeld, Agata Smoktunowicz, and Micha l Ziembowski. On radicals of Ore etensions and related
questions. Preprint, arXiv: 1702.08103v2, 2017.
[10] K.L. Horton. The prime and primitive spectra of multiparameter quantum sympectic and Euclidean spaces.
Comm. Alg., 31(10):4713 -- 4743, 2003.
15
[11] Li Huishi and Freddy van Oystaeyen. Zariskian Filtrations. Springer Science+Business Media, 1996.
[12] Michel Lazard. Groupes analytiques p-adiques. Publications Math´ematiques de l'IH ´ES, 26:5 -- 219, 1965.
[13] Edward S. Letzter. Prime ideals of noetherian skew power series rings. Israel J. Math., 192:67 -- 81, 2012.
[14] Edward S. Letzter and Linhong Wang. Prime ideals of q-commutative power series rings. Algebr. Represent.
Theor., 14:1003 -- 1023, 2011.
[15] J.C. McConnell and J.C. Robson. Noncommutative Noetherian Rings. American Mathematical Society,
2001.
[16] George S. Rinehart and Alex Rosenberg. The global dimensions of Ore extensions and Weyl algebras.
Algebra, Topology, and Category Theory, pages 169 -- 180, 0.
[17] J.E. Roseblade. Prime ideals in group rings of polycyclic groups. Proc. London Math. Soc., 36:385 -- 447,
1978.
[18] Peter Schneider and Otmar Venjakob. On the codimension of modules over skew power series rings with
applications to Iwasawa algebras. J. Pure Appl. Algebra, 204:349 -- 367, 2005.
[19] Peter Schneider and Otmar Venjakob. Localisations and completions of skew power series rings. Am. J.
Math., 132(1):1 -- 36, 2010.
[20] Otmar Venjakob. A noncommutative Weierstrass preparation theorem and applications to Iwasawa theory.
J. Reine Angew. Math., 559:153 -- 191, 2003.
[21] Linhong Wang. Completions of quantum coordinate rings. Proc. Amer. Math. Soc., 137(3):911 -- 919, 2009.
[22] William Woods. On the structure of virtually nilpotent compact p-adic analytic groups. J. Group Theory,
21(1):165 -- 188, 2018.
16
|
1112.6245 | 3 | 1112 | 2013-03-08T14:26:55 | Graded polynomial identities, group actions, and exponential growth of Lie algebras | [
"math.RA"
] | Consider a finite dimensional Lie algebra L with an action of a finite group G over a field of characteristic 0. We prove the analog of Amitsur's conjecture on asymptotic behavior for codimensions of polynomial G-identities of L. As a consequence, we prove the analog of Amitsur's conjecture for graded codimensions of any finite dimensional Lie algebra graded by a finite Abelian group. | math.RA | math |
GRADED POLYNOMIAL IDENTITIES, GROUP ACTIONS, AND
EXPONENTIAL GROWTH OF LIE ALGEBRAS
A. S. GORDIENKO
Abstract. Consider a finite dimensional Lie algebra L with an action of a finite group G
over a field of characteristic 0. We prove the analog of Amitsur's conjecture on asymptotic
behavior for codimensions of polynomial G-identities of L. As a consequence, we prove the
analog of Amitsur's conjecture for graded codimensions of any finite dimensional Lie algebra
graded by a finite Abelian group.
1. Introduction
In the 1980's, a conjecture about the asymptotic behaviour of codimensions of ordinary
polynomial identities was made by S.A. Amitsur. Amitsur's conjecture was proved in 1999
by A. Giambruno and M.V. Zaicev [1, Theorem 6.5.2] for associative algebras, in 2002 by
M.V. Zaicev [2] for finite dimensional Lie algebras, and in 2011 by A. Giambruno, I.P. Shes-
takov, M.V. Zaicev for finite dimensional Jordan and alternative algebras [3]. In 2011 the
author proved its analog for polynomial identities of finite dimensional representations of
Lie algebras [4]. Alongside with ordinary polynomial identities of algebras, graded polyno-
mial identities [5, 6] and G-identities are important too [7, 8]. Therefore the question arises
whether the conjecture holds for graded and G-codimensions. E. Aljadeff, A. Giambruno,
and D. La Mattina proved [9, 10] the analog of Amitsur's conjecture for codimensions of
graded polynomial identities of associative algebras graded by a finite Abelian group (or,
equivalently, for codimensions of G-identities where G is a finite Abelian group).
This article is concerned with graded codimensions (Theorem 1) and G-codimensions
(Theorem 2) of Lie algebras.
1 , x(g)
Sg∈G X (g), X (g) = {x(g)
1.1. Graded polynomial identities and their codimensions. Let G be an Abelian
group. Denote by L(X gr) the free G-graded Lie algebra on the countable set X gr =
2 , . . .}, over a field F of characteristic 0, i.e. the algebra of Lie
polynomials in variables from X gr. The indeterminates from X (g) are said to be homogeneous
of degree g. The G-degree of a monomial [x(g1)
] ∈ L(X gr) (all long commutators in
the article are left-normed) is defined to be g1g2 . . . gt, as opposed to its total degree, which
is defined to be t. Denote by L(X gr)(g) the subspace of the algebra L(X gr) spanned by all
the monomials having G-degree g. Notice that [L(X gr)(g), L(X gr)(h)] ⊆ L(X gr)(gh), for every
g, h ∈ G. It follows that
, . . . , x(gt)
i1
it
L(X gr) =Mg∈G
L(X gr)(g)
is a G-grading. Let f = f (x(g1)
identity of a G-graded Lie algebra L =Lg∈G L(g) and write f ≡ 0 if f (a(g1)
2010 Mathematics Subject Classification. Primary 17B01; Secondary 17B70, 17B40, 20C30.
Key words and phrases. Lie algebra, polynomial identity, grading, codimension, cocharacter, symmetric
it ) ∈ L(X gr). We say that f is a graded polynomial
it ) = 0
, . . . , x(gt)
, . . . , a(gt)
i1
i1
group, Young diagram, group action.
Supported by post doctoral fellowship from Atlantic Association for Research in Mathematical Sciences
(AARMS), Atlantic Algebra Centre (AAC), Memorial University of Newfoundland (MUN), and Natural
Sciences and Engineering Research Council of Canada (NSERC).
1
2
A. S. GORDIENKO
npcgr
npcn(L).
for all a(gj )
ij ∈ L(gj ), 1 6 j 6 t. The set Idgr(L) of graded polynomial identities of L is a
graded ideal of L(X gr). The case of ordinary polynomial identities is included for the trivial
group G = {e}.
Example 1. Let G = Z2 = {¯0, ¯1}, gl2(F ) = gl2(F )(¯0) ⊕ gl2(F )(¯1) where gl2(F )(¯0) =
(cid:18) F 0
0 F (cid:19) and gl2(F )(¯1) =(cid:18) 0 F
F 0 (cid:19). Then [x(¯0), y(¯0)] ∈ Idgr(gl2(F )).
Let Sn be the nth symmetric group, n ∈ N, and
σ(2), . . . , x(gn)
n := h[x(g1)
V gr
σ(1), x(g2)
σ(n)] gi ∈ G, σ ∈ SniF .
The non-negative integer cgr
polynomial identities or the nth graded codimension of L.
V gr
n (L) := dim(cid:16)
V gr
n
n ∩Idgr(L)(cid:17) is called the nth codimension of graded
The analog of Amitsur's conjecture for graded codimensions can be formulated as follows.
Conjecture. There exists PIexpgr(L) := lim
n→∞
n (L) ∈ Z+.
Remark. I.B. Volichenko [11] gave an example of an infinite dimensional Lie algebra L with
a non-trivial polynomial identity for which the growth of codimensions cn(L) of ordinary
polynomial identities is overexponential. M.V. Zaicev and S.P. Mishchenko [12, 13] gave an
example of an infinite dimensional Lie algebra L with a non-trivial polynomial identity such
that there exists fractional PIexp(L) := lim
n→∞
Theorem 1. Let L be a finite dimensional non-nilpotent Lie algebra over a field F of char-
acteristic 0, graded by a finite Abelian group G. Then there exist constants C1, C2 > 0,
r1, r2 ∈ R, d ∈ N such that C1nr1dn 6 cgr
n (L) 6 C2nr2dn for all n ∈ N.
Corollary. The above analog of Amitsur's conjecture holds for such codimensions.
Remark. If L is nilpotent, i.e.
cgr
n (L) = 0 for all n > p.
[x1, . . . , xp] ≡ 0 for some p ∈ N, then V gr
n ⊆ Idgr(L) and
Theorem 1 will be obtained as a consequence of Theorem 2 in Subsection 1.3.
1.2. Polynomial G-identities and their codimensions. Analogously, one can consider
polynomial G-identities for any group G. We use the exponential notation for the action of
a group and its group algebra. We say that a Lie algebra L is a Lie algebra with G-action or
a Lie G-algebra if there is a fixed linear representation G → GL(L) such that [a, b]g = [ag, bg]
for all a, b ∈ L and g ∈ G. Denote by L(XG) the free Lie algebra over F with free formal
generators xg
for h ∈ G. Let X := {x1, x2, x3, . . .} where
xj := x1
j , 1 ∈ G. Then L(XG) becomes the free G-algebra with free generators xj, j ∈ N.
Let L be a Lie G-algebra over F . A polynomial f (x1, . . . , xn) ∈ L(XG) is a G-identity of
L if f (a1, . . . , an) = 0 for all ai ∈ L. The set IdG(L) of all G-identities of L is an ideal in
L(XG) invariant under G-action.
j , j ∈ N, g ∈ G. Define (xg
j )h := xhg
j
Example 2. Consider ψ ∈ Aut(gl2(F )) defined by the formula
(cid:18) a b
c d (cid:19)ψ
:=(cid:18) a −b
d (cid:19) .
−c
Then [x + xψ, y + yψ] ∈ IdG(gl2(F )) where G = hψi ∼= Z2.
Denote by V G
n the space of all multilinear G-polynomials in x1, . . . , xn, i.e.
n = h[xg1
V G
σ(1), xg2
σ(2), . . . , xgn
σ(n)] gi ∈ G, σ ∈ SniF .
GRADED POLYNOMIAL IDENTITIES
3
Then the number cG
G-identities or the nth G-codimension of L.
n (L) := dim(cid:16)
V G
V G
n
n ∩IdG(L)(cid:17) is called the nth codimension of polynomial
Remark. As in the case of associative algebras [1, Lemma 10.1.3], we have
cn(L) 6 cG
n (L) 6 Gncn(L).
Here cn(L) = c{e}
n (L) are ordinary codimensions.
Also we have the following upper bound:
Lemma 1. Let L be a finite dimensional Lie algebra with G-action over any field F and let
G be any group. Then cG
n (L) 6 (dim L)n+1.
Proof. Consider G-polynomials as n-linear maps from L to L. Then we have a natural map
n → HomF (L⊗n; L) with the kernel V G
V G
n ∩ IdG(L) that leads to the embedding
V G
n
n ∩ IdG(L)
V G
֒→ HomF (L⊗n; L).
Thus
cG
n (L) = dim(cid:18)
V G
n
n ∩ IdG(L)(cid:19) 6 dim HomF (L⊗n; L) = (dim L)n+1.
V G
The analog of Amitsur's conjecture for G-codimensions can be formulated as follows.
(cid:3)
Conjecture. There exists PIexpG(L) := lim
n→∞
n (L) ∈ Z+.
npcG
Theorem 2. Let L be a finite dimensional non-nilpotent Lie algebra over a field F of char-
acteristic 0. Suppose a finite group G not necessarily Abelian acts on L. Then there exist
constants C1, C2 > 0, r1, r2 ∈ R, d ∈ N such that C1nr1dn 6 cG
n (L) 6 C2nr2dn for all n ∈ N.
Corollary. The above analog of Amitsur's conjecture holds for such codimensions.
Remark. If L is nilpotent, i.e. [x1, . . . , xp] ≡ 0 for some p ∈ N, then, by the Jacobi identity,
n ⊆ IdG(L) and cG
V G
n (L) = 0 for all n > p.
Remark. The theorem is still true if we allow G to act not only by automorphisms, but by
anti-automorphisms too, i.e. if G = G0 ∪ G1 such that [a, b]g = [ag, bg] for all a, b ∈ L, g ∈ G0
and [a, b]g = [bg, ag] for all a, b ∈ L, g ∈ G1. Indeed, we can replace G with G = G0 ∪ (−G1)
where [a, b]−g = −[a, b]g = −[bg, ag] = [a−g, b−g] for all (−g) ∈ (−G1). Then G acts on
L by automorphisms only. Moreover, n-linear functions from L to L that correspond to
polynomials from P G
n , are the same. Thus
n and P G
cG
n (L) = dim(cid:18)
V G
n
n ∩ IdG(L)(cid:19) = dim
V G
V G
n
V G
n ∩ Id
G(L)! = c G
n (L)
has the desired asymptotics.
Theorem 2 is proved in Sections 4 -- 6.
4
A. S. GORDIENKO
1.3. Duality between group gradings and group actions. If F is an algebraically
closed field of characteristic 0 and G is finite Abelian, there exists a well known duality
defines a G-grading on L.
of this duality to polynomial identities can be found, e.g., in [1, Chapters 3 and 10].
between G-gradings and bG-actions where bG = Hom(G, F ∗) ∼= G. Details of the application
A character ψ ∈ bG acts on L in the natural way: (ag)ψ = ψ(g)ag for all g ∈ G and
ag ∈ L(g). Conversely, if L is a bG-algebra, then L(g) = {a ∈ L aψ = ψ(g)a for all ψ ∈ bG}
Note that if G is finite Abelian, then L(X gr) is a free bG-algebra with free generators
yj = Pg∈G x(g)
. Thus there exists an isomorphism ε : L(XbG) → L(X gr) defined by
ε(xj) =Pg∈G x(g)
, that preserves bG-action and G-grading. The isomorphism has the prop-
GPψ(ψ(g))−1ψ is one of the minimal idempotents of FbG
Lemma 2. Let L be a G-graded Lie algebra where G is a finite Abelian group. Consider the
erty ε((xj)eg) = x(g)
defined above.
j where eg := 1
j
j
corresponding bG-action on L. Then
(1) ε(Id bG(L)) = Idgr(L);
(2) c bG
n (L) = cgr
n (L).
n ) = V gr
n .
Proof. The first assertion is evident. The second assertion follows from the first one and the
equality ε(V bG
(cid:3)
Remark. Note that Z2-grading in Example 1 corresponds to Z2-action in Example 2.
Proof of Theorem 1. Codimensions do not change upon an extension of the base field. The
proof is analogous to the cases of ordinary codimensions of associative [1, Theorem 4.1.9]
and Lie algebras [2, Section 2]. Thus without loss of generality we may assume F to be
algebraically closed.
In virtue of Lemma 2, Theorem 1 is an immediate consequence of
Theorem 2.
(cid:3)
1.4. Formula for the PI-exponent. Theorem 2 is formulated for an arbitrary field F of
characteristic 0, but without loss of generality we may assume that F is algebraically closed.
Fix a Levi decomposition L = B ⊕ R where B is a maximal semisimple subalgebra of
L and R is the solvable radical of L. Note that R is invariant under G-action. By [14,
Theorem 1, Remark 3], we can choose B invariant under G-action too.
We say that M is an L-module with G-action if M is both left L- and F G-module, and
(a · v)g = ag · vg for all a ∈ L, v ∈ M and g ∈ G. There is a natural G-action on EndF (M)
defined by ψgm = (ψmg−1)g, m ∈ M, g ∈ G, ψ ∈ EndF (M). Note that L → gl(M) is a
homomorphism of F G-modules. Such module M is irreducible if for any G- and L-invariant
subspace M1 ⊆ M we have either M1 = 0 or M1 = M. Each G-invariant ideal in L can be
regarded as a left L-module with G-action under the adjoint representation of L.
Consider G-invariant ideals I1, I2, . . . , Ir, J1, J2, . . . , Jr, r ∈ Z+, of the algebra L such that
Jk ⊆ Ik, satisfying the conditions
(1) Ik/Jk is an irreducible L-module with G-action;
(2) for any G-invariant B-submodules Tk such that Ik = Jk ⊕ Tk, there exist numbers
qi > 0 such that
Let M be an L-module. Denote by Ann M its annihilator in L. Let
q1
{z }
d(L) := max(cid:18)dim
[[T1, L, . . . , L
], [T2, L, . . . , L
]] 6= 0.
], . . . , [Tr, L, . . . , L
qr
q2
{z }
{z }
Ann(I1/J1) ∩ · · · ∩ Ann(Ir/Jr)(cid:19)
L
GRADED POLYNOMIAL IDENTITIES
5
where the maximum is found among all r ∈ Z+ and all I1, . . . , Ir, J1, . . . , Jr satisfying
Conditions 1 -- 2. We claim that PIexpG(L) = d(L) and prove Theorem 2 for d = d(L).
1.5. Examples. Now we give several examples.
Example 3. Let L be a finite dimensional G-simple Lie algebra over an algebraically closed
field F of characteristic 0 where G is a finite group. Then there exist C > 0 and r ∈ R such
that Cnr(dim L)n 6 cG
n (L) 6 (dim L)n+1.
Proof. The upper bound follows from Lemma 1. Consider G-invariant L-modules I1 = L
and J1 = 0. Then I1/J1 is an irreducible L-module, Ann(I1/J1) = 0 since a G-simple algebra
has zero center, and dim(L/ Ann(I1/J1)) = dim L. Thus d(L) > dim L and by Theorem 2
we obtain the lower bound.
(cid:3)
Example 4. Let L be a finite dimensional simple G-graded Lie algebra over an algebraically
closed field F of characteristic 0 where G is a finite Abelian group. Then there exist C > 0
and r ∈ R such that Cnr(dim L)n 6 cgr
n (L) 6 (dim L)n+1.
Proof. This follows from Example 3 and Lemma 2.
(cid:3)
Example 5. Let L be a finite dimensional Lie algebra with G-action over any field F of
characteristic 0 such that PIexpG(L) 6 2 where G is a finite group. Then L is solvable.
Proof. It is sufficient to prove the statement for an algebraically closed field F . (See the
remark before Theorem 2.) Consider the G-invariant Levi decomposition L = B ⊕ R. If B 6=
0, there exists a G-simple Lie subalgebra B1 ⊆ L, dim B1 > 3 and PIexpG(L) = d(L) > 3
by Example 3. We get a contradiction. Hence L = R is a solvable algebra.
(cid:3)
Analogously, we derive Example 6 from Example 4.
Example 6. Let L be a finite dimensional G-graded Lie algebra over any field F of charac-
teristic 0 such that PIexpgr(L) 6 2 where G is a finite Abelian group. Then L is solvable.
Example 7. Let L = B1 ⊕ . . . ⊕ Bs be a finite dimensional semisimple Lie G-algebra over
an algebraically closed field F of characteristic 0 where G is a finite group and Bi are G-
minimal ideals. Let d := max16i6s dim Bi. Then there exist C1, C2 > 0 and r1, r2 ∈ R such
that C1nr1dn 6 cG
n (L) 6 C2nr2dn.
Proof. Note that if I is a G-simple ideal of L, then [I, L] 6= 0 and hence [I, Bi] 6= 0 for some
1 6 i 6 s. However [I, Bi] ⊆ Bi ∩ I is a G-invariant ideal. Thus I = Bi. And if I is a
G-invariant ideal of L, then it is semisimple and each of its simple components coincides with
one of Bi. Thus if I ⊆ J are G-invariant ideals of L and I/J is irreducible, then I = Bi ⊕ J
for some 1 6 i 6 s and dim(L/ Ann(I/J)) = dim Bi. Suppose I1, . . . , Ir, J1, . . . , Jr satisfy
Conditions 1 -- 2. Let Ik = Bik ⊕ Jk, 1 6 k 6 r. Then
[[Bi1, L, . . . , L], [Bi2, L, . . . , L], . . . , [Bir , L, . . . , L]] 6= 0
for some number of copies of L. Hence i1 = . . . = ir and
dim
L
Ann(I1/J1) ∩ · · · ∩ Ann(Ir/Jr)
= dim Bi1.
Therefore, d(L) = max16i6s dim Bi and the result follows from Theorem 2.
(cid:3)
Example 8. Let L = B1 ⊕ . . . ⊕ Bs be a finite dimensional semisimple G-graded Lie algebra
over an algebraically closed field F of characteristic 0 where G is a finite Abelian group and
Bi are minimal graded ideals. Let d := max16i6s dim Bi. Then there exist C1, C2 > 0 and
r1, r2 ∈ R such that C1nr1dn 6 cgr
n (L) 6 C2nr2dn.
6
A. S. GORDIENKO
Proof. This follows from Example 7 and Lemma 2.
(cid:3)
Example 9. Let m ∈ N, G ⊆ Sm and Oi be the orbits of G-action on
Denote
{1, 2, . . . , m} =
Oi.
sai=1
d := max
16i6s
Oi.
Let L be the Lie algebra over any field F of characteristic 0 with basis a1, . . . , am, b1, . . . , bm,
dim L = 2m, and multiplication defined by formulas [ai, aj] = [bi, bj] = 0 and
[ai, bj] =(cid:26) bj
if
0 if
i = j,
i 6= j.
Suppose G acts on L as follows: (ai)σ = aσ(i) and (bj)σ = bσ(j) for σ ∈ G. Then there exist
C1, C2 > 0 and r1, r2 ∈ R such that
C1nr1dn 6 cG
n (L) 6 C2nr2dn.
In particular, if
G = hτ i ∼= Zm = Z/(mZ) = {¯0, ¯1, . . . , m − 1}
where τ = (1 2 3 . . . m) (a cycle), then
C1nr1mn 6 cG
n (L) 6 C2nr2mn.
However, cn(L) = n − 1 for all n ∈ N.
Proof. If K ⊇ F is a larger field, then K ⊗F L is defined by the same formulas as L.
Since cG
n (L) = cG,K
(K ⊗F L) (see the remark before Theorem 2), we may assume F to be
algebraically closed.
n
Let Bi := hbj j ∈ OiiF , 1 6 i 6 s. Suppose I is a G-invariant ideal of L. If bi ∈ I,
then bσ(i) = (bi)σ ∈ I for all σ ∈ G. Thus if i ∈ Oj, then bk ∈ I for all k ∈ Oj. Let
mPi=1
c :=
(αiai + βibi) ∈ I for some αi, βi ∈ F . Then βibi = [ai, c] ∈ I for all 1 6 i 6 m too.
Therefore, I = A0 ⊕ Bi1 ⊕ . . . ⊕ Bik for some 1 6 ij 6 s and A0 ⊆ ha1, . . . , amiF .
If I, J ⊆ L are G-invariant ideals, then J ⊆ J + [L, L] ∩ I ⊆ I is a G-invariant ideal too.
Suppose I/J is irreducible. Then either [L, L] ∩ I ⊆ J and Ann(I/J) = L or I ⊆ J + [L, L]
where [L, L] = hb1, . . . , bmiF . Thus Ann(I/J) 6= L implies J = A0 ⊕ Bi1 ⊕ . . . ⊕ Bik and
I = Bℓ ⊕ J for some 1 6 ℓ 6 s. In this case dim(L/ Ann(I/J)) = Oℓ.
Note that if I1 = Bi1 ⊕ J1 and I2 = Bi2 ⊕ J2, then
[[Bi1, L, . . . , L], [Bi2, L, . . . , L]] = 0.
Thus I1, . . . , Ir, J1, . . . , Jr can satisfy Conditions 1 -- 2 only if r = 1. Hence
d(L) = max
16i6s
Oℓ
and by Theorem 2 we obtain the bounds.
Consider the ordinary polynomial identities. Using the Jacobi identity, any monomial in
Vn can be rewritten as a linear combination of left-normed commutators [x1, xj, xi3, . . . , xin].
Since the polynomial identity
[[x, y], [z, t]] ≡ 0
2 6 j 6 n, are linearly independent modulo Id(L). Indeed, ifPn
holds in L, we may assume that i3 < i4 < . . . < in. Note that fj = [x1, xj, xi3, . . . , xin],
k=2 αkfk ≡ 0, αk ∈ F , then
we substitute xj = b1 and xi = a1 for i 6= j. Only fj does not vanish. Hence αj = 0 and
cn(L) = n − 1.
(cid:3)
Example 10. Let m ∈ N, L = L¯k∈Zm
GRADED POLYNOMIAL IDENTITIES
7
L(¯k) be the Zm-graded Lie algebra with L(¯k) = hc¯k, d¯kiF ,
dim L(¯k) = 2, multiplication [c¯ı, c¯] = [d¯ı, d¯] = 0 and [c¯ı, d¯] = d¯ı+¯ where F is any field of
characteristic 0. Then there exist C1, C2 > 0 and r1, r2 ∈ R such that
Proof. Again, we may assume F to be algebraically closed. Let ζ ∈ F be an mth primitive
C1nr1mn 6 cgr
n (L) 6 C2nr2mn.
root of 1. Then bG = {ψ0, . . . , ψm−1} for G = Zm where ψℓ(¯) := ζ ℓj. We can identify
the algebras from Example 9 and Example 10 by formulas c¯ = Pm
Pm
k=1 ζ −jkak and d¯ =
k=1 ζ −jkbk. The Zm-grading and hτ i-action correspond to each other since (c¯)τ ℓ = ζ ℓjc¯ =
n (L) and the bounds follow
(cid:3)
ψℓ(¯)c¯ and (d¯)τ ℓ = ζ ℓjd¯ = ψℓ(¯)d¯. By Lemma 2, cgr
from Example 9.
n (L) = chτ i
V G
n
V G
1.6. Sn-cocharacters. One of the main tools in the investigation of polynomial identities
is provided by the representation theory of symmetric groups. The symmetric group Sn acts
n ∩IdG(L) by permuting the variables. Irreducible F Sn-modules are described
on the space
by partitions λ = (λ1, . . . , λs) ⊢ n and their Young diagrams Dλ. The character χG
n (L) of
n ∩IdG(L) is called the nth cocharacter of polynomial G-identities of L. We
the F Sn-module
can rewrite it as a sum χG
eTλ = aTλbTλ and e∗
(sign σ)σ, be the
∼= F Se∗
Young symmetrizers corresponding to a Young tableau Tλ. Then M(λ) = F SeTλ
Tλ
is an irreducible F Sn-module corresponding to the partition λ ⊢ n. We refer the reader
to [1, 17, 18] for an account of Sn-representations and their applications to polynomial
identities.
n (L) = Pλ⊢n m(L, G, λ)χ(λ) of irreducible characters χ(λ). Let
Tλ = bTλaTλ where aTλ = Pπ∈RTλ
π and bTλ = Pσ∈CTλ
V G
n
V G
Our proof of Theorem 2 follows the outline of the proof by M.V. Zaicev [2]. However, in
many cases we need to apply new ideas.
In Section 2 we discuss modules with G-action over Lie G-algebras, their annihilators and
complete reducibility.
In Section 3 we prove that m(L, G, λ) is polynomially bounded. In Section 4 we prove
that if m(L, G, λ) 6= 0, then the corresponding Young diagram Dλ has at most d long rows.
This implies the upper bound.
In Section 5 we consider faithful irreducible L0-modules with G-action where L0 is a re-
ductive Lie G-algebra. For an arbitrary k ∈ N, we construct an associative G-polynomial
that is alternating in 2k sets, each consisting of dim L0 variables. This polynomial is not an
identity of the corresponding representation of L0. In Section 6 we choose reductive alge-
bras and faithful irreducible modules with G-action, and glue the corresponding alternating
polynomials. This allows us to find λ ⊢ n with m(L, G, λ) 6= 0 such that dim M(λ) has the
desired asymptotic behavior and the lower bound is proved.
2. Lie algebras and modules with G-action
We need several auxiliary lemmas. First, the Weyl theorem [15, Theorem 6.3] on complete
reducibility of representations can be easily extended to the case of Lie algebras with G-
action.
Lemma 3. Let M be a finite dimensional module with G-action over a Lie G-algebra L0.
Suppose M is a completely reducible L0-module disregarding the G-action. Then M is com-
pletely reducible L0-module with G-action.
Corollary. If M is a finite dimensional module with G-action over a semisimple Lie G-
algebra B0, then M is a completely reducible module with G-action.
8
A. S. GORDIENKO
Proof of Lemma 3. Suppose M1 ⊆ M is a G-invariant L0-submodule of M. Then it is
sufficient to prove that there exists a G-invariant L0-submodule M2 ⊆ M such that M =
M1 ⊕ M2.
Since M is completely reducible, there exists an L0-homomorphism π : M → M1 such that
π(v) = v for all v ∈ M1. Consider a homomorphism π : M → M1, π(v) = 1
Then π(v) = v for all v ∈ M1 too and for all a ∈ L0, h ∈ G we have
GPg∈G π(vg−1)g.
π(a · v) =
π((a · v)g−1
)g =
π(ag−1
· vg−1
)g =
a · π(vg−1
)g = a · π(v),
π(vh) =
π((vh)g−1
)g =
π(v(h−1g)−1
)h(h−1g) =
1
GXg∈G
GXg∈G
1
1
GXg∈G
GXg∈G
1
where g′ = h−1g. Thus we can take M2 = ker π.
1
GXg∈G
GXg′∈G
1
(π(vg′−1
)g′
)h = π(v)h
(cid:3)
Note that [L, R] ⊆ N by [16, Proposition 2.1.7] where N is the nilpotent radical, which is
a G-invariant ideal.
Lemma 4. There exists a G-invariant subspace S ⊆ R such that R = S ⊕ N is the direct
sum of subspaces and [B, S] = 0.
Proof. Note that R is a B-submodule under the adjoint representation of B on L. Applying
the corollary of Lemma 3 to N ⊆ R, we obtain a G-invariant complementary subspace S ⊆ R
such that [B, S] ⊆ S. Thus [B, S] ⊆ S ∩ [L, R] ⊆ S ∩ N = 0.
(cid:3)
Therefore, L = B ⊕ S ⊕ N (direct sum of subspaces).
Let M be an L-module and let T be a subspace of L. Denote AnnT M := (Ann M) ∩ T .
Lemma 5 is a G-invariant analog of [2, Lemma 4].
Lemma 5. Let J ⊆ I ⊆ L be G-invariant ideals such that I/J is an irreducible L-module
with G-action. Then
(1) AnnB(I/J) and AnnS(I/J) are G-invariant subspaces of L;
(2) Ann(I/J) = AnnB(I/J) ⊕ AnnS(I/J) ⊕ N.
Proof. Since I/J is a module with G-action, Ann(I/J), AnnB(I/J), and AnnS(I/J) are
G-invariant. Moreover [N, I] ⊆ J since N is a nilpotent ideal and I/J is a composition
factor of the adjoint representation. Hence N ⊆ Ann(I/J). In order to prove the lemma, it
is sufficient to show that if b + s ∈ Ann(I/J), b ∈ B, s ∈ S, then b, s ∈ Ann(I/J). Denote
ϕ : L → gl(I/J). Then ϕ(b) + ϕ(s) = 0 and
[ϕ(b), ϕ(B)] = [−ϕ(s), ϕ(B)] = 0.
Hence ϕ(b) belongs to the center of ϕ(B) and ϕ(b) = ϕ(s) = 0 since ϕ(B) is semisimple.
Thus b, s ∈ Ann(I/J) and the lemma is proved.
(cid:3)
Lemma 6. Let L0 = B0 ⊕ R0 be a finite dimensional reductive Lie algebra with G-action,
B0 be a maximal semisimple G-subalgebra, and R0 be the center of L0. Let M be a finite
dimensional irreducible L0-module with G-action. Then
(1) M = M1 ⊕ . . . ⊕ Mq for some L0-submodules Mi, 1 6 i 6 q;
(2) elements of R0 act on each Mi by scalar operators;
(3) for every 1 6 i 6 q and g ∈ G there exists such 1 6 j 6 q that M g
i = Mj and this
action of G on the set {M1, . . . , Mq} is transitive.
Proof. Denote by ϕ the homomorphism L0 → gl(M). Then ϕ is a homomorphism of G-
representations. We claim that ϕ(R0) consist of semisimple operators. Let r1, . . . , rt be a
basis in R0. Consider the Jordan decomposition ϕ(ri) = r′
i is semisimple,
i where each r′
i + r′′
GRADED POLYNOMIAL IDENTITIES
9
each r′′
i is nilpotent, and both are polynomials of ϕ(ri) without a constant term [15, Section
4.2]. Since each ϕ(ri) commutes with all operators ϕ(a), a ∈ L0, the elements (r′′
i )g, 1 6
i 6 t, g ∈ G, generate a nilpotent G-invariant associative ideal K in the enveloping algebra
A ⊆ EndF (M) of the Lie algebra ϕ(L0). Suppose KM 6= 0. Then for some κ ∈ N we have
Kκ+1M = 0, but KκM 6= 0. Note that KκM is a non-zero G-invariant L0-submodule. Thus
KκM = M and KM = Kκ+1M = 0. Since K ⊆ EndF (M), we obtain K = 0.
Therefore ϕ(ri) = r′
i are commuting semisimple operators. They have a common basis of
eigenvectors. Hence we can choose subspaces Mi, 1 6 i 6 q, q ∈ N, such that
and each Mi is the intersection of eigenspaces of ϕ(ri). Note that [ϕ(ri), ϕ(x)] = 0 for all
x ∈ L0. Thus Mi are L0-submodules and Propositions 1 and 2 are proved.
M = M1 ⊕ . . . ⊕ Mq,
For every Mi we can define a linear function αi : R0 → F such that ϕ(r)m = αi(r)m for
all r ∈ R0 and m ∈ Mi. Then Mi =Tr∈R0
M g
i = \r∈R0
ker(ϕ(rg) − αi(r) · 1) = \r∈R0
ker(ϕ(r) − αi(r) · 1) and
ker(ϕ(r) − αi(rg−1
) · 1)
where r = rg. Therefore, M g
irreducible with respect to L0- and G-action that implies Proposition 3.
i must coincide with Mj for some 1 6 j 6 q. The module M is
(cid:3)
Lemma 7. Let W be a finite dimensional L-module with G-action. Let ϕ : L → gl(W )
be the corresponding homomorphism. Denote by A the associative subalgebra of EndF (W )
generated by the operators from ϕ(L) and G. Then ϕ([L, R]) ⊆ J(A) where J(A) is the
Jacobson radical of A.
Proof. Let W = W0 ⊇ W1 ⊇ W2 ⊇ . . . ⊇ Wt = {0} be a composition chain in W of
not necessarily G-invariant L-submodules. Then each Wi/Wi+1 is an irreducible L-module.
Denote the corresponding homomorphism by ϕi : L → gl(Wi/Wi+1). Then by E. Cartan's
theorem [16, Proposition 1.4.11], ϕi(L) is semisimple or the direct sum of a semisimple ideal
and the center of gl(Wi/Wi+1). Thus ϕi([L, L]) is semisimple and ϕi([L, L] ∩ R) = 0. Since
[L, R] ⊆ [L, L] ∩ R, we have ϕi([L, R]) = 0 and [L, R]Wi ⊆ Wi+1. Denote by ρ : G →
GL(W ) the homomorphism corresponding to G-action. The associative G-invariant ideal of
A generated by ϕ([L, R]) is nilpotent since for any ai ∈ ϕ([L, R]), bij ∈ ϕ(L), gij ∈ G we
have
a1(cid:0)ρ(g10)b11ρ(g11) . . . ρ(g1,s1−1)b1,s1ρ(g1,s1)(cid:1)a2 . . .
at−1(cid:0)ρ(gt−1,0)bt−1,1ρ(gt−1,1) . . . ρ(gt−1,st−1−1)bt−1,st−1ρ(gt−1,st−1)(cid:1)at =
a1(cid:0)bg10
t−1,st−1(cid:1)agt
g′
g′
t−1,st−1
t−1,1
t−1,1 . . . b
1,s1 (cid:1)ag2
t−1(cid:0)b
g′
1,s1
11 . . . b
2 . . . agt−1
t ρ(gt+1) = 0
where gi, g′
g ∈ G, b ∈ ϕ(L), w ∈ W . Thus ϕ([L, R]) ⊆ J(A).
ij ∈ G are products of gij obtained using the property ρ(g)bw = bgρ(g)w where
(cid:3)
3. Multiplicities of irreducible characters in χG
n (L)
The aim of the section is to prove
Theorem 3. Let L be a finite dimensional Lie G-algebra over a field F of characteristic 0
where G is a finite group. Then there exist constants C > 0, r ∈ N such that
for all n ∈ N.
m(L, G, λ) 6 Cnr
Xλ⊢n
for some C3 > 0 and r3 ∈ N.
Let G1 ⊆ G2 be finite groups and W1, W2 be F G1- and F G2-modules respectively. Then
we denote F G2-module F G2 ⊗F G1 W1 by W1 ↑ G2. Here G2 acts on the first component.
Let W2 ↓ G1 be W2 with G2-action restricted to G1. We use analogous notation for the
characters.
Denote by length(M) the number of irreducible components of a module M.
Consider the diagonal embedding ϕ : Sn → SnG,
2
n
. . .
. . . σ(n) n + σ(1) n + σ(2)
n + 1
n + 2
2n
. . .
. . . n + σ(n)
. . .
. . . (cid:19) .
ϕ(σ) :=(cid:18) 1
σ(1) σ(2)
Then we have
Lemma 8.
10
A. S. GORDIENKO
Remark. Cocharacters do not change upon an extension of the base field F (the proof is
completely analogous to [1, Theorem 4.1.9]), so we may assume F to be algebraically closed.
In [19, Theorem 13 (b)] A. Berele, using the duality between Sn- and GLm(F )-
cocharacters [20, 21], showed that such sequence for an associative algebra with an action
of a Hopf algebra is polynomially bounded. One may repeat those steps for Lie G-algebras
and prove Theorem 3 in that way. However we provide an alternative proof based only on
Sn-characters.
Let {e} be the trivial group, Vn := V {e}
n
, χn(L) := χ{e}
n (L), m(L, λ) := m(L, {e}, λ),
Id(L) := Id{e}(L). Then, by [22, Theorem 3.1],
m(L, λ) 6 C3nr3
(1)
Xλ⊢n
m(L, G, λ) = length(cid:18)
V G
n
n ∩ IdG(L)(cid:19) 6 length(cid:18)(cid:18)
V G
VnG
VnG ∩ Id(L)(cid:19) ↓ ϕ(Sn)(cid:19) .
Xλ⊢n
Proof. Consider Sn-isomorphism π : (VnG ↓ ϕ(Sn)) → V G
where G = {g1, g2, . . . , gG}, 1 6 t 6 n. Note that π(VnG ∩ Id(L)) ⊆ V G
F Sn-module
n ∩IdG(L) is a homomorphic image of F Sn-module(cid:16) VnG
n defined by π(xn(i−1)+t) = xgi
n ∩ IdG(L). Thus
(cid:3)
VnG∩Id(L)(cid:17) ↓ ϕ(Sn).
Hence it is sufficient to prove that length(cid:16)(cid:16) VnG
VnG∩Id(L)(cid:17) ↓ ϕ(Sn)(cid:17) is polynomially bounded.
However, we start with the study of the restriction on the larger subgroup
V G
n
V G
t
S{1, . . . , n} × S{n + 1, . . . , 2n} × . . . × S{n(G − 1), . . . , nG} ⊆ SnG
that we denote by (Sn)G.
This is a particular case of a more general situation. Let m = m1 + . . . + mt, mi ∈ N.
Then we have a natural embedding Sm1 × . . . × Smt ֒→ Sm. Irreducible representations of
Sm1 × . . . × Smt are isomorphic to M(λ(1))♯ . . . ♯M(λ(t)) where λ(i) ⊢ mi. Here
M(λ(1))♯ . . . ♯M(λ(t)) ∼= M(λ(1)) ⊗ . . . ⊗ M(λ(t))
as a vector space and Smi acts on M(λ(i)). Denote by χ(λ(1))♯ . . . ♯χ(λ(t)) the character of
M(λ(1))♯ . . . ♯M(λ(t)).
Analogously, χ(λ(1))b⊗ . . .b⊗ χ(λ(t)) is the character of F Sm-module
M(λ(1))b⊗ . . .b⊗ M(λ(t)) := (M(λ(1))♯ . . . ♯M(λ(t))) ↑ Sm.
Note that if m1 = . . . = mt = k, one can define the inner tensor product, i.e.
M(λ(1)) ⊗ . . . ⊗ M(λ(t))
with the diagonal Sk-action. The character of this F Sk-module equals χ(λ(1)) . . . χ(λ(t)).
GRADED POLYNOMIAL IDENTITIES
11
Recall that irreducible characters of any finite group G0 are orthonormal with respect to
the scalar product (χ, ψ) = 1
χ(g−1)ψ(g).
Denote by λT the transpose partition of λ ⊢ n. Then λT
1 equals the height of the first
G0Pg∈G0
1
1 .
Proof. By Frobenius reciprocity,
6 λT
1 for all 1 6 i 6 t and λT
column of Dλ.
Lemma 9. Let h, t ∈ N. There exist C4 > 0, r4 ∈ N such that for all λ ⊢ m, λ(1) ⊢ m1, . . . ,
λ(t) ⊢ mt, where Dλ lie in the strip of height h, i.e. λT
1 6 h, and m1 + m2 + . . . + mt = m,
we have
If λ ⊢ m, λ(1) ⊢ m1, . . . , λ(t) ⊢ mt, m1 + m2 + . . . + mt = m, and
Now we prove the lemma by induction on t. The case t = 1 is trivial. Suppose
1 6Pt
i=1(cid:0)λ(i)(cid:1)T
(cid:16)χ(λ), χ(λ(1))b⊗ . . .b⊗ χ(λ(t))(cid:17).
(cid:16)χ(λ) ↓ (Sm1 × . . . × Smt), χ(λ(1))♯ . . . ♯χ(λ(t))(cid:17) =(cid:16)χ(λ), χ(λ(1))b⊗ . . .b⊗ χ(λ(t))(cid:17) 6 C4mr4.
(cid:16)χ(λ) ↓ (Sm1 × . . . × Smt), χ(λ(1))♯ . . . ♯χ(λ(t))(cid:17) =(cid:16)χ(λ), χ(λ(1))b⊗ . . .b⊗ χ(λ(t))(cid:17) 6= 0
then(cid:0)λ(i)(cid:1)T
(cid:16)χ(λ) ↓ (Sm1 × . . . × Smt), χ(λ(1))♯ . . . ♯χ(λ(t))(cid:17) =(cid:16)χ(λ), (χ(λ(1))♯ . . . ♯χ(λ(t))) ↑ Sm(cid:17) =
(cid:16)χ(µ), χ(λ(1))b⊗ . . .b⊗ χ(λ(t−1))(cid:17) is polynomially bounded for every µ ⊢ (m1 + . . . + mt−1)
(cid:16)χ(λ), χ(λ(1))b⊗ . . .b⊗ χ(λ(t))(cid:17) =(cid:16)χ(λ), (cid:0)χ(λ(1))b⊗ . . .b⊗ χ(λ(t−1))(cid:1)b⊗ χ(λ(t))(cid:17) =
In order to determine the multiplicity of χ(λ) in χ(µ) b⊗ χ(λ(t)), we are using the Little-
if(cid:0)λ(t)(cid:1)T
strip of height h. Then we may consider only the case(cid:0)λ(t)(cid:1)T
the number of variants to add the boxes from a row is bounded by mh. Since(cid:0)λ(t)(cid:1)T
wood -- Richardson rule (see the algorithm in [23, Corollary 2.8.14]). We cannot obtain Dλ
1 . Suppose the Young diagram Dλ lies in the
1 6 h. Each time
6 h,
the second multiplier in (2) is bounded by (mh)h = mh2. The number of diagrams in the
strip of height h is bounded by mh. Thus the number of terms in (2) is bounded by mh.
Together with the inductive assumption this yields the lemma.
(cid:3)
Xµ⊢(m1+...+mt−1)(cid:16)χ(µ), χ(λ(1))b⊗ . . .b⊗ χ(λ(t−1))(cid:17)(cid:16)χ(λ), χ(µ)b⊗ χ(λ(t))(cid:17).
1 >(cid:0)λ(t)(cid:1)T
1 > λT
1 or µT
1 > λT
1 , or λT
with µT
1 6 h. We have
6 h and µT
1
1 + µT
(2)
1
VnG
VnG ∩ Id(L)(cid:19) ↓ (Sn)G(cid:19) 6 C5nr5
1 > dim L for some 1 6 i 6 G, then M(λ(1))♯ . . . ♯M(λ(G))
Lemma 10. There exist C5 > 0, r5 ∈ N such that
length(cid:18)(cid:18)
for all n ∈ N. Moreover, if(cid:0)λ(i)(cid:1)T
M(λ(1))♯ . . . ♯M(λ(G)) in(cid:16) VnG
does not appear in the decomposition.
Proof. Fix an G-tuple of partitions (λ(1), . . . , λ(G)), λ(i) ⊢ n. Then the multiplicity of
VnG∩Id(L)(cid:17) ↓ (Sn)G equals
(cid:16)χ(λ(1))♯ . . . ♯χ(λ(G)), χnG(L) ↓ (Sn)G(cid:17) =
12
A. S. GORDIENKO
Xλ⊢nG(cid:16)χ(λ(1))♯ . . . ♯χ(λ(G)), χ(λ) ↓ (Sn)G(cid:17)m(L, λ).
(3)
By [22, Lemma 3.4] (or Lemma 14 for G = hei), m(L, λ) = 0 for all λ ⊢ nG with
1 > dim L. Thus Lemma 9 implies that for all M(λ(1))♯ . . . ♯M(λ(G)) that appear in
λT
VnG∩Id(L)(cid:17) ↓ (Sn)G, the Young diagrams Dλ(i) lie in the strip of height (dim L). Thus
(cid:16) VnG
the number of different (λ(1), . . . , λ(G)) that appear in the decomposition of(cid:16) VnG
VnG∩Id(L)(cid:17) ↓
(Sn)G is bounded by n(dim L)G. Together with (1), (3), and Lemma 9, this yields the
lemma.
(cid:3)
Lemma 11. Let h, k ∈ N. There exist C6 > 0, r6 ∈ N such that for the inner tensor product
M(λ) ⊗ M(µ) of any F Sn-modules M(λ) and M(µ), λ, µ ⊢ n, λT
1 6 k, we have
1 6 h, µT
lengthSn(M(λ) ⊗ M(µ)) 6 C6nr6
and(cid:16)χ(λ)χ(µ), χ(ν)(cid:17) = 0 for any ν ⊢ n with νT
1 > hk.
Proof. Let Tµ be any Young tableau of the shape µ. Denote by IRTµ the one-dimensional
trivial representation of the Young subgroup (i.e. the row stabilizer) RTµ. Then
F SnaTµ
∼= IRTµ ↑ Sn
(see [24, Section 4.3]). By [25, Theorem 38.5],
M(λ) ⊗ (IRTµ ↑ Sn) ∼= ((M(λ) ↓ RTµ) ⊗ IRTµ) ↑ Sn.
Thus
M(λ) ⊗ M(µ) ∼= M(λ) ⊗ F Sne∗
M(λ) ⊗ (IRTµ ↑ Sn) ∼= ((M(λ) ↓ RTµ) ⊗ IRTµ) ↑ Sn
∼=
Tµ = M(λ) ⊗ F SnbTµaTµ ⊆ M(λ) ⊗ F SnaTµ
∼= (M(λ) ↓ RTµ) ↑ Sn.
1 6 k, κ(i) ⊢ µi,(cid:0)κ(i)(cid:1)T
1
Note that length(M(λ) ↓ RTµ) is polynomially bounded by Lemma 9 and M(λ) ↓ RTµ is a
sum of M(κ(1))♯ . . . ♯M(κ(s)), s = µT
6 h. Thus (M(λ) ↓ RTµ) ↑ Sn
Lemma 12. There exist C7 > 0, r7 ∈ N satisfying the following properties.
is a sum of M(κ(1))b⊗ . . .b⊗ M(κ(s)). Applying Lemma 9 again, we obtain the lemma. (cid:3)
(λ(1), . . . , λ(G)) is an G-tuple of partitions λ(i) ⊢ n where (cid:0)λ(i)(cid:1)T
lengthSn(cid:0)M(λ(1)) ⊗ . . . ⊗ M(λ(G))(cid:1) 6 C7nr7.
If
6 dim L for all
1 6 i 6 G, then
Proof. Note that
1
M(λ(1)) ⊗ . . . ⊗ M(λ(t)) = (M(λ(1)) ⊗ . . . ⊗ M(λ(t−1))) ⊗ M(λ(t)).
Using induction on t and applying Lemma 11 with h = (dim L)t−1 and k = dim L, we obtain
the lemma.
(cid:3)
Proof of Theorem 3. The theorem is an immediate consequence of Lemmas 8, 10, and 12. (cid:3)
Fix a composition chain of G-invariant ideals
4. Upper bound
L = L0 % L1 % L2 % . . . % N % . . . % Lθ−1 % Lθ = {0}.
Let ht a := maxa∈Lk k for a ∈ L.
Remark. If d = d(L) = 0, then L = Ann(Li−1/Li) for all 1 6 i 6 θ and [a1, a2, . . . , an] = 0
for all ai ∈ L and n > θ + 1. Thus cG
n (L) = 0 for all n > θ + 1. Therefore we assume d > 0.
GRADED POLYNOMIAL IDENTITIES
13
Let Y := {y11, y12, . . . , y1j1; y21, y22, . . . , y2j2; . . . ; ym1, ym2, . . . , ymjm}, Y1,
. . . , Yq, and
{z1, . . . , zm} be subsets of {x1, x2, . . . , xn} such that Yi ⊆ Y , Yi = d + 1, Yi ∩ Yj = ∅
for i 6= j, Y ∩ {z1, . . . , zm} = ∅, ji > 0. Denote
fm,q := Alt1 . . . Altq[[zg1
1 , yg11
11 , yg12
12 , . . . , y
2 , yg21
21 , yg22
22 , . . . , y
g2j2
2j2 ], . . . ,
g1j1
1j1 ], [zg2
m2 , . . . , ygmjm
mjm ]]
[zgm
m , ygm1
m1 , ygm2
where Alti is the operator of alternation on the variables of Yi, gi, gij ∈ G.
Let ϕ : L(XG) → L be a G-homomorphism induced by some substitution
{x1, x2, . . . , xn} → L. We say that ϕ is proper for fm,q if ϕ(z1) ∈ N ∪ B ∪ S, ϕ(zi) ∈ N for
2 6 i 6 m, and ϕ(yik) ∈ B ∪ S for 1 6 i 6 m, 1 6 k 6 ji.
Lemma 13. Let ϕ be a proper homomorphism for fm,q. Then ϕ(fm,q) can be rewritten as
a sum of ψ(fm+1,q′) where ψ is a proper homomorphism for fm+1,q′, q′ > q − (dim L)m − 2.
(Y ′, Y ′
m+1 may be different for different terms.)
1, . . . , z′
i , z′
Proof. Let αi := ht ϕ(zi). We will use induction on Pm
that αi 6 θ 6 dim L. Denote Ii := Lαi, Ji := Lαi+1.
First, consider the case when I1, . . . , Im, J1, . . . , Jm do not satisfy Conditions 1 -- 2. In this
case we can choose G-invariant B-submodules Ti, Ii = Ti ⊕ Ji, such that
i=1 αi. (The sum will grow.) Note
], [T2, L, . . . , L
], . . . , [Tm, L, . . . , L
]] = 0
(4)
q2
{z }
qm
{z }
for all qi > 0. Rewrite ϕ(zi) = a′
i > ht ϕ(zi). Since fm,q
is multilinear, we can rewrite ϕ(fm,q) as a sum of similar terms ϕ(fm,q) where ϕ(zi) equals
either a′
i ∈ Ti, equals 0. For the other terms
i . By (4), the term where all ϕ(zi) = a′
i ∈ Ji. Note that ht a′′
i ∈ Ti, a′′
i or a′′
i , a′
q1
[[T1, L, . . . , L
i + a′′
{z }
i=1 ht ϕ(zi) >Pm
i=1 ht ϕ(zi).
ϕ(fm,q) we havePm
Thus without lost of generality we may assume that I1, . . . , Im, J1, . . . , Jm satisfy Condi-
tions 1 -- 2. In this case, dim(Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm)) > dim(L) − d. In virtue of
Lemma 5,
Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm) = B ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm)⊕
S ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm) ⊕ N.
Choose a basis in B that includes a basis of B ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm) and a basis
in S that includes the basis of S ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm). Since fm,q is multilinear,
we may assume that only basis elements are substituted for ykℓ. Note that fm,q is alternating
in Yi. Hence, if ϕ(fm,q) 6= 0, then for every 1 6 i 6 q there exists yjk ∈ Yi such that either
or
ϕ(yjk) ∈ B ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm)
ϕ(yjk) ∈ S ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm).
Consider the case when ϕ(ykj) ∈ B ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm) for some ykj. By the
corollary from Lemma 3, we can choose G-invariant B-submodules Tk such that Ik = Tk ⊕Jk.
We may assume that ϕ(zk) ∈ Tk since elements of Jk have greater heights. Therefore
[ϕ(zgk
k ), a] = 0.
Moreover, B ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm) is a G-invariant ideal of B and [B, S] = 0.
Thus, applying Jacobi's identity several times, we obtain
k ), a] ∈ Tk ∩ Jk for all a ∈ B ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm). Hence [ϕ(zgk
Expanding the alternations, we get ϕ(fm,q) = 0.
ϕ([zgk
k , ygk1
k1 , . . . , y
gkjk
kjk
]) = 0.
14
A. S. GORDIENKO
Consider the case when ϕ(ykℓ) ∈ S ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm) for some ykℓ ∈ Yq.
Expand the alternation Altq in fm,q and rewrite fm,q as a sum of
2 , yg21
fm,q−1 := Alt1 . . . Altq−1[[zg1
1 , yg11
m , ygm1
[zgm
11 , yg12
m1 , ygm2
12 , . . . , y
m2 , . . . , ygmjm
g1j1
1j1 ], [zg2
mjm ]].
21 , yg22
22 , . . . , y
g2j2
2j2 ], . . . ,
The operator Altq may change indices, however we keep the notation ykℓ for the variable
with the property ϕ(ykℓ) ∈ S ∩ Ann(I1/J1) ∩ . . . ∩ Ann(Im/Jm). Now the alternation does
not affect ykℓ. Note that
[zgk
k , ygk1
k1 , . . . , ygkℓ
kℓ , . . . , y
gkjk
kjk
] = [zgk
k , ygkℓ
kℓ , ygk1
k1 , . . . , y
gkjk
kjk
]+
[zgk
k , ygk1
k1 , . . . , ygk,β−1
k,β−1 , [ygkβ
kβ , ygkℓ
kℓ ], ygk,β+1
k,β+1 , . . . , ygk,ℓ−1
k,ℓ−1 , ygk,ℓ+1
k,ℓ+1 , . . . , y
gkjk
kjk
].
ℓ−1Xβ=1
k , ygkℓ
kℓ ] with z′
k and define ϕ′(z′
In the first term we replace [zgk
kℓ ]), ϕ′(x) :=
ϕ(x) for other variables x. Then ht ϕ′(z′
k) > ht ϕ(zk) and we can use the inductive as-
If ykβ ∈ Yj for some j, then we expand the alternation Altj in this term in
sumption.
fm,q−1. If ϕ(ykβ) ∈ B, then the term is zero. If ϕ(ykβ) ∈ S, then ϕ([ygkβ
kℓ ]) ∈ N. We
replace [ygkβ
kℓ ]),
ψ(x) := ϕ(x) for other variables x. Applying Jacobi's identity several times, we obtain the
polynomial of the desired form. In each inductive step we reduce q no more than by 1 and
the maximal number of inductive steps equals (dim L)m. This finishes the proof.
(cid:3)
kℓ ] with an additional variable z′
m+1 and define ψ(z′
m+1) := ϕ([ygkβ
k) := ϕ([zgk
kβ , ygkℓ
kβ , ygkℓ
kβ , ygkℓ
k , ygkℓ
Since N is a nilpotent ideal, N p = 0 for some p ∈ N.
Lemma 14. If λ = (λ1, . . . , λs) ⊢ n and λd+1 > p((dim L)p + 3) or λdim L+1 > 0, then
m(L, G, λ) = 0.
Proof. It is sufficient to prove that e∗
λ ⊢ n, with λd+1 > p((dim L)p + 3) or λdim L+1 > 0.
Tλf ∈ IdG(L) for every f ∈ V G
n and a Young tableau Tλ,
Fix some basis of L that is a union of bases of B, S, and N. Since polynomials are
multilinear, it is sufficient to substitute only basis elements. Note that e∗
Tλ = bTλaTλ and bTλ
alternates the variables of each column of Tλ. Hence if we make a substitution and e∗
Tλf does
not vanish, then this implies that different basis elements are substituted for the variables of
each column. But if λdim L+1 > 0, then the length of the first column is greater than dim L.
Therefore, e∗
Tλf ∈ IdG(L).
Consider the case λd+1 > p((dim L)p + 3). Let ϕ be a substitution of basis elements for
the variables x1, . . . , xn. Then e∗
Tλf can be rewritten as a sum of polynomials fm,q where
1 6 m 6 p, q > p((dim L)p + 2), and zi, 2 6 i 6 m, are replaced with elements of N.
(For different terms fm,q, numbers m and q, variables zi, yij, and sets Yi can be different.)
Indeed, we expand symmetrization on all variables and alternation on the variables replaced
with elements from N. If we have no variables replaced with elements from N, then we take
m = 1, rewrite the polynomial f as a sum of long commutators, in each long commutator
expand the alternation on the set that includes one of the variables in the inner commutator,
and denote that variable by z1. Suppose we have variables replaced with elements from N.
We denote them by zk. Then, using Jacobi's identity, we can put one of such variables inside
a long commutator and group all the variables, replaced with elements from B ∪ S, around
zk such that each zk is inside the corresponding long commutator.
Applying Lemma 13 many times, we increase m. The ideal N is nilpotent and ϕ(fp+1,q) = 0
for every q and a proper homomorphism ϕ. Reducing q no more than by p((dim L)p + 2),
we obtain ϕ(e∗
(cid:3)
Tλf ) = 0.
Now we can prove
GRADED POLYNOMIAL IDENTITIES
15
Theorem 4. If d > 0, then there exist constants C2 > 0, r2 ∈ R such that cG
for all n ∈ N. In the case d = 0, the algebra L is nilpotent.
n (L) 6 C2nr2dn
Proof. Lemma 14 and [1, Lemmas 6.2.4, 6.2.5] imply
Xm(L,G,λ)6=0
dim M(λ) 6 C8nr8dn
for some constants C8, r8 > 0. Together with Theorem 3 this implies the upper bound. (cid:3)
In this section we prove auxiliary propositions needed to obtain the lower bound.
5. Alternating polynomials
Lemma 15. Let α1, α2, . . . , αq, β1, . . . , βq ∈ F , 1 6 k 6 q, αi 6= 0 for 1 6 i < k, αk = 0,
and βk 6= 0. Then there exists such γ ∈ F that αi + γβi 6= 0 for all 1 6 i 6 k.
Proof. It is sufficient to choose γ /∈ n− α1
infinite.
β1
, . . . , − αk−1
βk−1
, 0o.
It is possible to do since F is
(cid:3)
Let F hXGi be the free associative algebra over F with free formal generators xg
j )h = xhg
g ∈ G. Define (xg
with free generators xj = x1
multilinear G-polynomials in variables x1, . . . , xn. In other words,
j , j ∈ N,
for h ∈ G. Then F hXGi becomes the free associative G-algebra
j , j ∈ N, 1 ∈ G. Denote by P G
n , n ∈ N, the subspace of associative
j
P G
n =( Xσ∈Sn, g1,...,gn∈G
ασ,g1,...,gn xg1
σ(1)xg2
σ(2) . . . xgn
σ(n)(cid:12)(cid:12)(cid:12)(cid:12) ασ,g1,...,gn ∈ F) .
Lemma 16. Let L0 = B0 ⊕ R0 be a reductive Lie algebra with G-action, B0 be a maximal
semisimple G-subalgebra, and R0 be the center of L0 with a basis r1, r2, . . . , rt. Let M be
a faithful finite dimensional irreducible L0-module with G-action. Denote the correspond-
ing representation L0 → gl(M) by ϕ. Then there exists such alternating in x1, x2, . . . , xt
polynomial f ∈ P G
t
that f (ϕ(r1), . . . , ϕ(rt)) is a nondegenerate operator on M.
Proof. By Lemma 6, M = M1 ⊕ . . . ⊕ Mq where Mj are L0-submodules and ri acts on each
Mj as a scalar operator. Note that it is sufficient to prove that for each j there exists such
alternating in x1, x2, . . . , xt polynomial fj ∈ P G
that fj(ϕ(r1), . . . , ϕ(rt)) multiplies each
t
element of Mj by a nonzero scalar. Indeed, in this case Lemma 15 implies the existence of
such f = γ1f1 + . . . + γqfq, γi ∈ F , that f (ϕ(r1), . . . , ϕ(rt)) acts on each Mi as a nonzero
scalar.
Lemma 6, Proposition 3, we can choose such gi ∈ G that M gi
Denote by pi ∈ EndF (M) the projection on Mi along Lk6=i Mk. Fix 1 6 j 6 q. By
Consider fj := Pσ∈Sq (sign σ)xg1
i = pj.
σ(q). Note that either pg1
σ(q) = 0 or
σ(1)pg2
σ(q) = pj if and
σ(i) = Mj. Hence σ(i) = i.
pg1
σ(1)pg2
only if σ(i) = i for all 1 6 i 6 q. Indeed, pgi
This implies that fj(p1, . . . , pq) acts as an identical map on Mj.
σ(q) = pk for some 1 6 k 6 s. Now we prove that pg1
σ(i) = pj if and only if M gi
i = Mj, 1 6 i 6 q. Then pgi
σ(1)pg2
σ(2) . . . pgq
σ(2) . . . xgq
σ(2) . . . pgq
σ(2) . . . pgq
σ(1)xg2
We can choose it+1, . . . , iq such that ϕ(r1), ϕ(r2), . . . , ϕ(rt), pit+1, . . . , piq form a basis in
hp1, . . . , pqiF . Then fj(ϕ(r1), ϕ(r2), . . . , ϕ(rt), pit+1, . . . , piq) acts as a nonzero scalar on Mj.
If t = q, then we define fj = fj. Suppose t < q. Since the projections commute, we can
rewrite
fj(ϕ(r1), ϕ(r2), . . . , ϕ(rt), pit+1, . . . , piq ) =
fi(ϕ(r1), ϕ(r2), . . . , ϕ(rt))pi
qXi=1
16
A. S. GORDIENKO
where fi ∈ P G
as a nonzero scalar operator. We define fj := fj.
t are alternating in x1, x2, . . . , xt. Hence fj(ϕ(r1), ϕ(r2), . . . , ϕ(rt)) acts on Mj
(cid:3)
Let L0 be a Lie algebra with G-action, M be L0-module with G-action, ϕ : L0 → gl(M)
be the corresponding representation. A polynomial f (x1, . . . , xn) ∈ F hXGi is a G-identity
of ϕ if f (ϕ(a1), . . . , ϕ(an)) = 0 for all ai ∈ L0. The set IdG(ϕ) of all G-identities of ϕ is a
two-sided ideal in F hXGi invariant under G-action.
Lemma 17 is an analog of [3, Lemma 1].
Lemma 17. Let L0 be a Lie algebra with G-action, M be a faithful finite dimensional irre-
ducible L0-module with G-action, and ϕ : L0 → gl(M) be the corresponding representation.
n \ IdG(ϕ) alternating in {x1, . . . , xℓ}
Then for some n ∈ N there exists a polynomial f ∈ P G
and in {y1, . . . , yℓ} ⊆ {xℓ+1, . . . , xn} where ℓ = dim L0.
Proof. Since M is irreducible, by the density theorem, EndF (M) ∼= Mq(F ) is generated by
operators from G and ϕ(L0). Here q := dim M. Consider Regev's polynomial
(sign(στ ))xσ(1) yτ (1) xσ(2)xσ(3)xσ(4) yτ (2)yτ (3)yτ (4) . . .
f (x1, . . . , xq2; y1, . . . , yq2) := Xσ∈Sq ,
τ ∈Sq
xσ(q2−2q+2) . . . xσ(q2) yτ (q2−2q+2) . . . yτ (q2).
This is a central polynomial [1, Theorem 5.7.4] for Mk(F ), i.e. f is not a polynomial identity
for Mq(F ) and its values belong to the center of Mq(F ).
Let a1, . . . , aℓ be a basis of L0. Denote by ρ the representation G → GL(M). Note that if
we have the product of elements of ϕ(L0) and ρ(G), we can always move the elements from
ρ(G) to the right, using ρ(g)a = agρ(g) for g ∈ G and a ∈ ϕ(L0). Then ϕ(a1), . . . , ϕ(aℓ),
(cid:0)ϕ (ai11) . . . ϕ(cid:0)ai1,m1(cid:1)(cid:1) ρ(g1), . . . ,(cid:0)ϕ(cid:0)air,1(cid:1) . . . ϕ(cid:0)air,mr(cid:1)(cid:1) ρ(gr), is a basis of EndF (M) for ap-
g ∈ G, in fq to the right and rewrite f asPg∈G fg ρ(g) where each fg ∈ P G
propriate ijk ∈ {1, 2, . . . , ℓ}, gj ∈ G, since EndF (M) is generated by operators from G and
j,mj ρ(gj) in f and
ϕ(L0). We replace xℓ+j with zj1zj2 . . . zj,mj ρ(gj) and yℓ+j with z′
denote the expression obtained by f . Using ρ(g)a = agρ(g) again, we can move all ρ(g),
is an
alternating in x1, . . . , xℓ and in y1, . . . , yℓ polynomial. Note that f becomes a nonzero scalar
operator on M under the substitution xi = yi = ϕ(ai) for 1 6 i 6 ℓ and zjk = z′
jk = ϕ(aijk )
for 1 6 j 6 r, 1 6 k 6 mj. Thus fg /∈ IdG(ϕ) for some g ∈ G and we can take f = fg.
(cid:3)
j2 . . . z′
2ℓ+2 Pr
j1z′
j=1 mj
Let kℓ 6 n where k, ℓ, n ∈ N are some numbers. Denote by QG
n the sub-
space spanned by all polynomials that are alternating in k disjoint subsets of variables
{xi
ℓ,k,n ⊆ P G
1, . . . , xi
Theorem 5 is an analog of [3, Theorem 1].
ℓ} ⊆ {x1, x2, . . . , xn}, 1 6 i 6 k.
Theorem 5. Let L0 = B0 ⊕R0 be a reductive Lie algebra with G-action over an algebraically
closed field F of characteristic 0, B0 be a maximal semisimple G-subalgebra, R0 be the center
of L0, and dim L0 = ℓ. Let M be a faithful finite dimensional irreducible L0-module with
G-action. Denote the corresponding representation L0 → gl(M) by ϕ. Then there exists
T ∈ Z+ such that for any k ∈ N there exists f ∈ QG
Proof. Let f1 = f1(x1, . . . , xℓ, y1, . . . , yℓ, z1, . . . , zT ) be the polynomial from Lemma 17 al-
ℓ,2,2ℓ+T \ IdG(ϕ), we may assume that
ternating in x1, . . . , xℓ and in y1, . . . , yℓ. Since f1 ∈ QG
k > 1. Note that
ℓ,2k,2kℓ+T \ IdG(ϕ).
f (1)
1 (u1, v1, x1, . . . , xℓ, y1, . . . , yℓ, z1, . . . , zT ) :=
f1(x1, . . . , [u1, [v1, xi]], . . . , xℓ, y1, . . . , yℓ, z1, . . . , zT )
ℓXi=1
GRADED POLYNOMIAL IDENTITIES
17
is alternating in x1, . . . , xℓ and in y1, . . . , yℓ and
f (1)
1 (¯u1, ¯v1, ¯x1, . . . , ¯xℓ, ¯y1, . . . , ¯yℓ, ¯z1, . . . , ¯zT ) =
tr(adϕ(L0) ¯u1 adϕ(L0) ¯v1)f1(¯x1, ¯x2, . . . , ¯xℓ, ¯y1, . . . , ¯yℓ, ¯z1, . . . , ¯zT )
for any substitution of elements from ϕ(L0) since we may assume ¯x1, . . . , ¯xℓ to be different
basis elements. Here (ad a)b = [a, b].
Let
f (j)
1 (u1, . . . , uj, v1, . . . , vj, x1, . . . , xℓ, y1, . . . , yℓ, z1, . . . , zT ) :=
f (j−1)
1
(u1, . . . , uj−1, v1, . . . , vj−1, x1, . . . , [uj, [vj, xi]], . . . , xℓ, y1, . . . , yℓ, z1, . . . , zT ),
ℓXi=1
2 6 j 6 s, s = dim B. Note that if we substitute an element from ϕ(R0) for ui or vi, then
f (j)
1 vanish since R0 is the center of L0. Again,
f (j)
1 (¯u1, . . . , ¯uj, ¯v1, . . . , ¯vj, ¯x1, . . . , ¯xℓ, ¯y1, . . . , ¯yℓ, ¯z1, . . . , ¯zT ) =
tr(adϕ(L0) ¯u1 adϕ(L0) ¯v1) tr(adϕ(L0) ¯u2 adϕ(L0) ¯v2) . . . tr(adϕ(L0) ¯uj adϕ(L0) ¯vj)·
· f1(¯x1, ¯x2, . . . , ¯xℓ, ¯y1, . . . , ¯yℓ, ¯z1, . . . , ¯zT )
(5)
Let h be the polynomial from Lemma 16. We define
Xσ,τ ∈Sℓ
f2(u1, . . . , uℓ, v1, . . . , vℓ, x1, . . . , xℓ, y1, . . . , yℓ, z1, . . . , zT ) :=
sign(στ )f (s)
1 (uσ(1), . . . , uσ(s), vτ (1), . . . , vτ (s), x1, . . . , xℓ, y1, . . . , yℓ, z1, . . . , zT )
·h(uσ(s+1), . . . , uσ(ℓ))h(vτ (s+1), . . . , vτ (ℓ)).
Then f2 ∈ QG
ℓ,4,4ℓ+T . Suppose a1, . . . , as ∈ ϕ(B0) and as+1, . . . , aℓ ∈ ϕ(R0) form a basis of
ϕ(L0). Consider a substitution xi = yi = ui = vi = ai, 1 6 i 6 ℓ. Suppose that the values
zj = ¯zj, 1 6 j 6 T , are chosen in such a way that f1(a1, . . . , aℓ, a1, . . . , aℓ, ¯z1, . . . , ¯zT ) 6= 0.
We claim that f2 does not vanish either. Indeed,
f2(a1, . . . , aℓ, a1, . . . , aℓ, a1, . . . , aℓ, a1, . . . , aℓ, ¯z1, . . . , ¯zT ) =
sign(στ )f (s)
1 (aσ(1), . . . , aσ(s), aτ (1), . . . , aτ (s), a1, . . . , aℓ, a1, . . . , aℓ, ¯z1, . . . , ¯zT )
sign(στ )f (s)
·h(aσ(s+1), . . . , aσ(ℓ))h(aτ (s+1), . . . , aτ (ℓ)) =
1 (aσ(1), . . . , aσ(s), aτ (1), . . . , aτ (s), a1, . . . , aℓ, a1, . . . , aℓ, ¯z1, . . . , ¯zT )! ·
Xπ,ω∈S{s+1,...,ℓ}
sign(πω)h(aπ(s+1), . . . , aπ(ℓ))h(aω(s+1), . . . , aω(ℓ))
since aj, s < j 6 ℓ, belong to the center of ϕ(L0) and f (s)
vanishes if we substitute such ai
for ui or vi. Here S{s + 1, . . . , ℓ} is the symmetric group on {s + 1, . . . , ℓ}. Note that h is
alternating. Using (5), we obtain
j
Xσ,τ ∈Sℓ
Xσ,τ ∈Ss
Xσ,τ ∈Ss
f2(a1, . . . , aℓ, a1, . . . , aℓ, a1, . . . , aℓ, a1, . . . , aℓ, ¯z1, . . . , ¯zT ) =
sign(στ ) tr(adϕ(L0) aσ(1) adϕ(L0) aτ (1)) . . . tr(adϕ(L0) aσ(s) adϕ(L0) aτ (s))! ·
f1(a1, . . . , aℓ, a1, . . . , aℓ, ¯z1, . . . , ¯zT )((ℓ − s)!)2 (h(as+1, . . . , aℓ))2 .
18
A. S. GORDIENKO
sign(στ ) tr(adϕ(L0) aσ(1) adϕ(L0) aτ (1)) . . . tr(adϕ(L0) aσ(s) adϕ(L0) aτ (s)) =
Note that
Xσ,τ ∈Ss
Xσ,τ ∈Ss
Xσ,τ ′∈Ss
sign(στ ) tr(adϕ(L0) a1 adϕ(L0) aτ σ−1(1)) . . . tr(adϕ(L0) as adϕ(L0) aτ σ−1(s))
(τ ′=τ σ−1)
=
sign(τ ′) tr(adϕ(L0) a1 adϕ(L0) aτ ′(1)) . . . tr(adϕ(L0) as adϕ(L0) aτ ′(s)) =
s! det(tr(adϕ(L0) ai adϕ(L0) aj))s
i,j=1 = s! det(tr(adϕ(B0) ai adϕ(B0) aj))s
i,j=1 6= 0
since the Killing form tr(ad x ad y) of the semisimple Lie algebra ϕ(B0) is nondegenerate.
Thus
f2(a1, . . . , aℓ, a1, . . . , aℓ, a1, . . . , aℓ, a1, . . . , aℓ, ¯z1, . . . , ¯zT ) 6= 0.
Note that if f1 is alternating in some of z1, . . . , zT , the polynomial f2 is alternating in
those variables too. Thus if we apply the same procedure to f2 instead of f1, we obtain
f3 ∈ QG
ℓ,6,6ℓ+T . Analogously, we define f4 using f3, f5 using f4, etc. Eventually, we obtain
f = fk ∈ QG
(cid:3)
ℓ,2k,2kℓ+T \ IdG(ϕ).
By the definition of d = d(L), there exist G-invariant ideals I1, I2, . . . , Ir, J1, J2, . . . , Jr,
r ∈ Z+, of the algebra L, satisfying Conditions 1 -- 2, Jk ⊆ Ik, such that
6. Lower bound
d = dim
L
Ann(I1/J1) ∩ · · · ∩ Ann(Ir/Jr)
.
We consider the case d > 0.
Without loss of generality we may assume that
Ann(Ik/Jk) 6=
r\k=1
Ann(Ik/Jk)
r\k=1,
k6=ℓ
for all 1 6 ℓ 6 r. In particular, L has nonzero action on each Ik/Jk.
Our aim is to present a partition λ ⊢ n with m(L, G, λ) 6= 0 such that dim M(λ) has the
desired asymptotic behavior. We will glue alternating polynomials constructed in Theorem 5
for faithful irreducible modules over reductive algebras. In order to do this, we have to choose
the reductive algebras.
Lemma 18. There exist G-invariant ideals B1, . . . , Br in B and G-invariant subspaces
R1, . . . , Rr ⊆ S (some of Ri and Bj may be zero) such that
dim(Bk ⊕ Rk) = d;
(1) B1 + . . . + Br = B1 ⊕ . . . ⊕ Br;
(2) R1 + . . . + Rr = R1 ⊕ . . . ⊕ Rr;
(3)
rPk=1
(5) Ik/Jk is an irreducible(cid:16)Pr
(6) BiIk/Jk = RiIk/Jk = 0 for i > k.
(4) Ik/Jk is a faithful (Bk ⊕ Rk ⊕ N)/N-module;
i=1(Bi ⊕ Ri) ⊕ N(cid:17) /N-module with G-action;
Proof. Consider Nℓ :=
ℓTk=1
GRADED POLYNOMIAL IDENTITIES
19
Ann(Ik/Jk), 1 6 ℓ 6 r, N0 = L. Note that Nℓ are G-invariant.
Since B is semisimple, we can choose such G-invariant ideals Bℓ that Nℓ−1∩B = Bℓ⊕(Nℓ∩B).
Also we can choose such G-invariant subspaces Rℓ that Nℓ−1 ∩ S = Rℓ ⊕ (Nℓ ∩ S). Hence
Properties 1, 2, 6 hold.
By Lemma 5, Nk = (Nk ∩ B) ⊕ (Nk ∩ S) ⊕ N. Thus Property 4 holds. Furthermore,
Nℓ−1 = Bℓ ⊕ (Nℓ ∩ B) ⊕ Rℓ ⊕ (Nℓ ∩ S) ⊕ N = (Bℓ ⊕ Rℓ) ⊕ Nℓ
(direct sum of subspaces). Hence L =(cid:16)Lr
too.
i=1(Bi ⊕ Ri)(cid:17) ⊕ Nr, and Properties 3 and 5 hold
(cid:3)
Let A be the associative subalgebra in EndF (L) generated by operators from ad L and G.
Then J(A)p = 0 for some p ∈ N. Denote by A2 a subalgebra of EndF (L) generated by ad L
only. Let aℓ1, . . . , aℓ,kℓ be a basis of Rℓ.
Lemma 19. There exist decompositions ad aij = cij + dij, 1 6 i 6 r, 1 6 j 6 ki, such
that cij ∈ A acts as a diagonalizable operator on L, dij ∈ J(A), elements cij commute with
each other, and cij and dij are polynomials in ad aij. Moreover, Rℓ := hcℓ1, . . . , cℓ,kℓiF are
G-invariant subspaces in A.
Proof. Consider the solvable G-invariant Lie algebra (ad R) + J(A).
In virtue of the Lie
theorem, there exists a basis in L in which all the operators from (ad R) + J(A) have upper
triangular matrices. Denote the corresponding embedding A ֒→ Mm(F ) by ψ. Here m :=
dim L.
Let A1 be the associative algebra generated by ad aij, 1 6 i 6 r, 1 6 j 6 ki. This algebra
is G-invariant since for every fixed i the elements aij, 1 6 j 6 ki, form a basis of the G-
invariant subspace Ri. By the G-invariant Wedderburn -- Malcev theorem [14, Theorem 1,
Remark 1], A1 = A1 ⊕ J(A1) (direct sum of subspaces) where A1 is a G-invariant semisimple
subalgebra of A1. Since ψ(ad R) ⊆ tm(F ), we have ψ(A1) ⊆ UTm(F ). Here UTm(F ) is the
associative algebra of upper triangular matrices m × m. There is a decomposition
where
UTm(F ) = F e11 ⊕ F e22 ⊕ · · · ⊕ F emm ⊕ N
N := heij 1 6 i < j 6 miF
is a nilpotent ideal. Thus there is no subalgebras in A1 isomorphic to M2(F ) and A1 =
F e1 ⊕ · · · ⊕ F et for some idempotents ei ∈ A1. Denote for every aij its component in
J(A1) by dij and its component in F e1 ⊕ · · · ⊕ F et by cij. Note that ei are commuting
diagonalizable operators. Thus they have a common basis of eigenvectors in L and cij are
commuting diagonalizable operators too. Moreover
ad ag
ij = cg
ij + dg
ij ∈ had aiℓ 1 6 ℓ 6 kiiF ⊆ hciℓ 1 6 ℓ 6 kiiF ⊕ hdiℓ 1 6 ℓ 6 kiiF
for all g ∈ G. Thus Ri is G-invariant.
We claim that the space J(A1)+J(A) generates a nilpotent G-invariant ideal I in A. First,
ψ(J(A1)), ψ(J(A)) ⊆ UTm(F ) and consist of nilpotent elements. Thus the corresponding
matrices have zero diagonal elements and ψ(J(A1)), ψ(J(A)) ⊆ N. Denote Nk := heij
i + k 6 jiF ⊆ N . Then
N = N1 % N2 % . . . % Nm−1 % Nm = {0}.
Let ht N a := k if ψ(a) ∈ Nk, ψ(a) /∈ Nk+1.
20
A. S. GORDIENKO
Recall that (J(A))p = 0. We claim that I m+p = 0. Let ρ : G → GL(L) be the G-action on
L. Using the property
(6)
where a ∈ A2, g ∈ G, we obtain that the space I m+p is a span of h1j1h2j2 . . . jm+phm+p+1ρ(g)
where jk ∈ J(A1)∪J(A), hk ∈ A2 ∪{1}, g ∈ G. If at least p elements jk belong to J(A), then
the product equals 0. Thus we may assume that at least m elements jk belong to J(A1).
ρ(g)a = agρ(g)
expressed as a sum of j1j2 . . . jαj′
Let ji ∈ J(A1), hi ∈ A2 ∪ {1}. We prove by induction on ℓ that j1h1j2h2 . . . hℓ−1jℓ can be
i ∈ J(A), a ∈ A2 ∪ {1}, and
i > ℓ. Indeed, suppose that j1h1j2h2 . . . hℓ−2jℓ−1 can be expressed as a sum of
i > ℓ − 1.
βa where ji ∈ J(A1), j′
i=1 ht N j′
2 . . . j′
i=1 ht N j′
j1j2 . . . jγj′
κa where ji ∈ J(A1), j′
Then j1h1j2h2 . . . jℓ−1hℓ−1jℓ is a sum of
2 . . . j′
1j′
i ∈ J(A), a ∈ A2 ∪ {1}, and γ +Pκ
α +Pβ
1j′
j1j2 . . . jγj′
1j′
2 . . . j′
κahℓ−1jℓ = j1j2 . . . jγj′
1j′
2 . . . j′
κ[ahℓ−1, jℓ] + j1j2 . . . jγj′
1j′
2 . . . j′
κjℓ(ahℓ−1).
Note that, in virtue of the Jacobi identity and Lemma 7, [ahℓ−1, jℓ] ∈ J(A). Thus it is
sufficient to consider only the second term. However
j1j2 . . . jγj′
1j′
2 . . . j′
κjℓ(ahℓ−1) = j1j2 . . . jγjℓj′
1j′
2 . . . j′
κ(ahℓ−1)+
j1j2 . . . jγj′
1j′
2 . . . j′
i−1[j′
i, jℓ]j′
i+1 . . . j′
κ(ahℓ−1).
κXi=1
Since [j′
fore,
i, jℓ] ∈ J(A) and ht N [j′
i, jℓ] > 1 + ht N j′
i, all the terms have the desired form. There-
I m+p = 0, and
j1h1j2h2 . . . jm−1hm−1jm ∈ ψ−1( Nm) = {0},
J(A) ⊆ J(A1) + J(A) ⊆ I ⊆ J(A).
In particular, dij ∈ J(A1) ⊆ J(A).
(cid:3)
Denote
B := rMi=1
ad Bi! ⊕ hcij 1 6 i 6 r, 1 6 j 6 kiiF ,
B0 := (ad B) ⊕ hcij 1 6 i 6 r, 1 6 j 6 kiiF ⊆ A.
Lemma 20. The space L is a completely reducible B0-module with G-action. Moreover, L
is a completely reducible (ad Bk) ⊕ Rk-module with G-action for any 1 6 k 6 r.
Proof. By Lemma 3, it is sufficient to show that L is a completely reducible B0-module and
a completely reducible (ad Bk) ⊕ Rk-module disregarding the G-action. The elements cij are
diagonalizable on L and commute. Therefore, an eigenspace of any cij is invariant under the
i=1 Wi where Wi are intersections of
eigenspaces of ckℓ and elements ckℓ act as scalar operators on Wi. In virtue of Lemmas 4, 19,
and the Jacobi identity, [cij, ad B] = 0. Thus Wi are B-submodules and L is a completely
reducible B0-module and (ad Bk) ⊕ Rk-module since B and Bk are semisimple.
(cid:3)
Lemma 21. There exist complementary subspaces Ik = Tk ⊕ Jk such that
action of other ckℓ. Using induction, we split L = Lα
(1) Tk is a B-submodule and an irreducible B-submodule with G-action;
(2) Tk is a completely reducible faithful (ad Bk) ⊕ Rk-module with G-action;
(3)
dim((ad Bk) ⊕ Rk) = d;
rPk=1
(4) Bi Tk = Ri Tk = 0 for i > k.
GRADED POLYNOMIAL IDENTITIES
21
Proof. By Lemma 20, L is a completely reducible B0-module with G-action. Therefore, for
every Jk we can choose a complementary G-invariant B0-submodules Tk in Ik. Then Tk are
both B- and B-submodules.
Note that (ad aij)w = cijw for all w ∈ Ik/Jk since Ik/Jk is an irreducible A-module and
J(A) Ik/Jk = 0. Hence, by Lemma 18, Ik/Jk is a faithful (ad Bk) ⊕ Rk-module, Ri Ik/Jk =
0 for i > k and the elements cij are linearly independent. Moreover, by Property 5 of
i=1(Bi ⊕ Ri) ⊕ N(cid:17) /N-module with G-action. However
Lemma 18, Ik/Jk is an irreducible(cid:16)Pr
(cid:16)Pr
i=1(Bi ⊕ Ri) ⊕ N(cid:17) /N acts on Ik/Jk by the same operators as B. Thus Tk
∼= Ik/Jk
is an irreducible B-module with G-action. Property 1 is proved. By Lemma 20, L is a
completely reducible (ad Bk) ⊕ Rk-module with G-action for any 1 6 k 6 r. Using the
∼= Ik/Jk, we obtain Properties 2 and 4 from the remarks above. Property 3
isomorphism Tk
is a consequence of Property 3 of Lemma 18.
(cid:3)
Lemma 22. For all 1 6 k 6 r we have
Tk = Tk1 ⊕ Tk2 ⊕ · · · ⊕ Tkm
where Tkj are faithful irreducible (ad Bk)⊕Rk-submodules with G-action, m ∈ N, 1 6 j 6 m.
Proof. By Lemma 21, Property 2, Tk = Tk1⊕Tk2⊕· · ·⊕Tkm for some irreducible (ad Bk)⊕Rk-
submodules with G-action. Suppose Tkj is not faithful for some 1 6 j 6 m. Hence bTkj = 0
for some b ∈ (ad Bk) ⊕ Rk, b 6= 0. Note that B = ((ad Bk) ⊕ Rk) ⊕ Bk where
Bk :=Mi6=k
(ad Bi) ⊕Mi6=k
Ri
generated by operators from Bk. Then
and [(ad Bk) ⊕ Rk, Bk] = 0. Denote by bBk the associative subalgebra of EndF ( Tk) with 1
andPa∈ bBk
aTkj ⊇ Tkj is a G-invariant B-submodule of Tk since
[(ad Bk) ⊕ Rk, bBk] = 0
aTkj
Xa∈ bBk
kj = Xa∈ bBk
= Xa∈ bBk
for all g ∈ G. Thus Tk =Pa∈ bBk
baTkj = Xa∈ bBk
b Tk = Xa∈ bBk
aTkj and
agT g
agTkj = Xa′∈ bBk
a(bTkj) = 0.
a′Tkj
g
We get a contradiction with faithfulness of Tk.
(cid:3)
By Condition 2 of the definition of d, there exist numbers q1, . . . , qr ∈ Z+ such that
[[ T1, L, . . . , L
], [ T2, L, . . . , L
] . . . , [ Tr, L, . . . , L
]] 6= 0
jrk! Tr, L, . . . , L
{z }
qr
]] 6= 0
Choose ni ∈ Z+ with the maximal
q2
q1
ni such that
{z }
{z }
rPi=1
], [ n2Yk=1
j2k! T2, L, . . . , L
{z }
q2
qr
{z }
] . . . , [ nrYk=1
[[ n1Yk=1
j1k! T1, L, . . . , L
{z }
q1
k=1 jik. Then ji ∈ J(A) ∪ {1} and
], . . . , [jr Tr, L, . . . , L
], [j2 T2, L, . . . , L
]] 6= 0,
22
A. S. GORDIENKO
[[j1 T1, L, . . . , L
for some jik ∈ J(A). Let ji :=Qni
{z }
{z }
for all j ∈ J(A) and 1 6 k 6 r.
[[j1 T1, L, . . . , L
but
q1
q1
q2
{z }
{z }
qk
qr
{z }
{z }
qr
qr
{z }
], . . . , [jk(j Tk), L, . . . , L
], . . . , [jr Tr, L, . . . , L
]] = 0
(7)
In virtue of Lemma 22, for every k we can choose a faithful irreducible (ad Bk) ⊕ Rk-
submodule with G-action Tk ⊆ Tk such that
] . . . , [jrTr, L, . . . , L
]] 6= 0.
(8)
Lemma 23. Let ψ : Lr
i=1((ad Bi) ⊕ Ri) be the linear isomorphism defined
by formulas ψ(b) = ad b for all b ∈ Bi and ψ(aiℓ) = ciℓ, 1 6 ℓ 6 kℓ. Let fi be multilinear
associative G-polynomials, h(i)
i=1 Bi ⊕ Ri, ¯ti ∈ Ti, ¯uik ∈ L, be some elements.
Then
[[j1T1, L, . . . , L
q2
q1
], [j2T2, L, . . . , L
{z }
{z }
i=1(Bi ⊕ Ri) →Lr
ni ∈Lr
1 , . . . , h(i)
[[j1f1(ad h(1)
1 , . . . , ad h(1)
nr )¯tr, ¯ur1, . . . , ¯urqr]] =
n1 )¯t1, ¯u11, . . . , ¯u1q1], . . . , [jrfr(ad h(r)
[[j1f1(ψ(h(1)
[jrfr(ψ(h(r)
1 , . . . , ad h(r)
n1 ))¯t1, ¯u11, . . . , ¯u1q1], . . . ,
nr ))¯tr, ¯ur1, . . . , ¯urqr]].
1 ), . . . , ψ(h(1)
1 ), . . . , ψ(h(r)
In other words, we can replace ad aiℓ with ciℓ and the result does not change.
Proof. We rewrite ad aiℓ = ciℓ + diℓ = ψ(ai) + diℓ and use the multilinearity of fi. By (7),
terms with diℓ vanish.
(cid:3)
Denote by A3 ⊆ EndF (L) the linear span of products of operators from ad L and G such
that each product contains at least one element from ad L.
Lemma 24. J(A) ⊆ A3.
Proof. Note that A3 is a G-invariant two-sided ideal of A and A3 + A3 = A where
A3 ⊆ EndF (L) is the associative subalgebra generated by operators from G. Thus
∼= A3/( A3 ∩ A3) is a semisimple algebra since A3 is a homomorphic image of the
A/A3
semisimple group algebra F G. Thus J(A) ⊆ A3.
(cid:3)
Lemma 25. If d 6= 0, then there exist a number n0 ∈ N such that for every n > n0 there
exist disjoint subsets X1, . . . , X2k ⊆ {x1, . . . , xn}, k :=(cid:2) n−n0
n \ IdG(L) alternating in the variables of each set Xj.
2d (cid:3), X1 = . . . = X2k = d and
a polynomial f ∈ V G
Proof. Denote by ϕi : (ad Bi) ⊕ Ri → gl(Ti) the representation corresponding to the action of
(ad Bi) ⊕ Ri on Ti. In virtue of Theorem 5, there exist constants mi ∈ Z+ such that for any
di,2k,2kdi+mi\ IdG(ϕi), di := dim((ad Bi) ⊕ Ri),
k there exist multilinear polynomials fi ∈ QG
alternating in the variables from disjoint sets X (i)
, 1 6 ℓ 6 2k, X (i)
= di.
ℓ
ℓ
In virtue of (8),
[[j1¯t1, ¯u11, . . . , ¯u1,q1], [j2¯t2, ¯u21, . . . , ¯u2,q2], . . . , [jr¯tr, ¯ur1, . . . , ¯ur,qr]] 6= 0,
for some ¯uiℓ ∈ L and ¯ti ∈ Ti. All ji ∈ J(A) ∪ {1} are polynomials in elements from G and
ad L. Denote by m the maximal degree of them.
Recall that each Ti is a faithful irreducible (ad Bi) ⊕ Ri-module with G-action. Therefore
by the density theorem, EndF (Ti) is generated by operators from G and (ad Bi) ⊕ Ri. Note
that EndF (Ti) ∼= Mdim Ti(F ). Thus every matrix unit e(i)
jℓ ∈ Mdim Ti(F ) can be represented as
GRADED POLYNOMIAL IDENTITIES
23
a polynomial in operators from G and (ad Bi) ⊕ Ri. Choose such polynomials for all i and
all matrix units. Denote by m0 the maximal degree of those polynomials.
Let n0 := r(2m0 + m + 1) +Pr
fi /∈ IdG(ϕi), there exist ¯xi1, . . . , ¯xi,2kdi+mi ∈ (ad Bi)⊕Ri such that fi(¯xi1, . . . , ¯xi,2kdi+mi) 6= 0.
Hence
i=1(mi + qi). Now we choose fi for k = (cid:2) n−n0
2d (cid:3). Since
e(i)
ℓiℓi
fi(¯xi1, . . . , ¯xi,2kdi+mi)e(i)
sisi 6= 0
for some matrix units e(i)
ℓiℓi
, e(i)
sisi ∈ EndF (Ti), 1 6 ℓi, si 6 dim Ti. Thus
dimTiXℓ=1
e(i)
ℓℓi
fi(¯xi1, . . . , ¯xi,2kdi+mi)e(i)
siℓ
is a nonzero scalar operator in EndF (Ti).
Hence
e(1)
ℓℓ1f1(¯x11, . . . , ¯x1,2kd1+m1)e(1)
e(r)
ℓℓr fr(¯xr1, . . . , ¯xr,2kdr+mr )e(r)
s1ℓ! ¯t1, ¯u11, . . . , ¯u1q1], . . . ,
srℓ! ¯tr, ¯ur1, . . . , ¯urqr]] 6= 0.
[[j1 dim T1Xℓ=1
[jr dim TrXℓ=1
i=1 X (i)
ℓ
. Let Altℓ be the operator of alternation in the variables from Xℓ.
Denote Xℓ :=Sr
Consider
Then
f (x11, . . . , x1,2kd1+m1, . . . , xr1, . . . , xr,2kdr+mr ) :=
Alt1 Alt2 . . . Alt2k[[j1 dim T1Xℓ=1
e(1)
ℓℓ1f1(x11, . . . , x1,2kd1+m1)e(1)
s1ℓ! ¯t1, ¯u11, . . . , ¯u1q1], . . . ,
[jr dim TrXℓ=1
e(r)
ℓℓr fr(xr1, . . . , xr,2kdr+mr )e(r)
srℓ! ¯tr, ¯ur1, . . . , ¯urqr]].
(d1!)2k . . . (dr!)2k[[j1 dim T1Xℓ=1
[jr dim TrXℓ=1
f (¯x11, . . . , ¯x1,2kd1+m1, . . . , ¯xr1, . . . , ¯xr,2kdr+mr ) =
e(1)
ℓℓ1f1(¯x11, . . . , ¯x1,2kd1+m1)e(1)
s1ℓ! ¯t1, ¯u11, . . . , ¯u1q1], . . . ,
e(r)
ℓℓr fr(¯xr1, . . . , ¯xr,2kdr+mr )e(r)
srℓ! ¯tr, ¯ur1, . . . , ¯urqr]] 6= 0.
since fi are alternating in each X (i)
ℓ and, by Lemma 21, ((ad Bi) ⊕ Ri) Tℓ = 0 for i > ℓ. Now
ℓj as polynomials in elements of (ad Bi) ⊕ Ri and G. Using linearity of f in e(i)
we rewrite e(i)
ℓj ,
we can replace e(i)
ℓj with the products of elements from (ad Bi)⊕Ri and G, and the expression
will not vanish for some choice of the products. Using (6), we can move all ρ(g) to the right.
By Lemma 23, we can replace all elements from (ad Bi) ⊕ Ri with elements from Bi ⊕ Ri
i=1((ad Bi) ⊕ Ri)
the corresponding linear isomorphism. Now we rewrite ji as polynomials in elements ad L
and G. Since f is linear in ji, we can replace ji with one of the monomials, i.e. with the
product of elements from ad L and G. Using (6), we again move all ρ(g) to the right. Then
we replace the elements from ad L with new variables, and
and the expression will be still nonzero. Denote by ψ : Lr
i=1(Bi ⊕ Ri) →Lr
f := Alt1 Alt2 . . . Alt2k(cid:20)hhy11, [y12, . . . [y1α1,hz11, [z12, . . . , [z1β1,
24
A. S. GORDIENKO
(f1(ad x11, . . . , ad x1,2kd1+m1))g1[w11, [w12, . . . , [w1γ1, th1
hhyr1, [yr2, . . . , [yrαr,hzr1, [zr2, . . . , [zrβr,
(fr(ad xr1, . . . , ad xr,2kdr+mr ))gr[wr1, [wr2, . . . , [wrγr, thr
1 ] . . .i, u11, . . . , u1q1i, . . . ,
r ] . . .i, ur1, . . . , urqri(cid:21)
for some 0 6 αi 6 m,
under the substitution ti = ¯ti, uiℓ = ¯uiℓ, xiℓ = ψ−1(¯xiℓ), yiℓ = ¯yiℓ, ziℓ = ¯ziℓ, wiℓ = ¯wiℓ.
0 6 βi, γi 6 m0,
¯yiℓ, ¯ziℓ, ¯wiℓ ∈ L does not vanish
gi, hi ∈ G,
we take f := f . Suppose n > n. Let b ∈ (ad B1) ⊕ R1, b 6= 0. Then e(1)
i=1(mi + qi + αi + βi + γi) 6 n. If n = n, then
6= 0 for
¯t1 = µ¯t1, µ ∈ F \{0}. Hence f does not
jj be(1)
ℓℓ
¯t1; ti = ¯ti for 2 6 i 6 r; uiℓ = ¯uiℓ,
ℓs )(cid:17)n−n
(e(1)
sj be(1)
ℓs )(cid:17)n−n
s=1
(e(1)
sj be(1)
Note that f ∈ V G
n , n := 2kd + r +Pr
some 1 6 j, ℓ 6 dim T1 and(cid:16)Pdim T1
vanish under the substitution t1 =(cid:16)Pdim T1
can rewrite (cid:16)Pdim T1
xiℓ = ψ−1(¯xiℓ), yiℓ = ¯yiℓ, ziℓ = ¯ziℓ, wiℓ = ¯wiℓ.
By Lemma 24,
s=1
b ∈ J(A) ⊕ ad(B1 ⊕ R1) ⊆ A3
and using (6) we
[¯v1, [¯v2, [. . . , [¯vq, ¯tg
substitution t1 = [¯v1, [¯v2, [. . . , [¯vq, ¯tg
2 6 i 6 r; uiℓ = ¯uiℓ, xiℓ = ψ−1(¯xiℓ), yiℓ = ¯yiℓ, ziℓ = ¯ziℓ, wiℓ = ¯wiℓ. Therefore,
elements
1] . . .], q > n − n, ¯vi ∈ L, g ∈ G. Hence f does not vanish under a
1] . . .] for some q > n − n, ¯vi ∈ L, g ∈ G; ti = ¯ti for
¯t1 as a sum of
(e(1)
sj be(1)
s=1
ℓs )(cid:17)n−n
f := Alt1 Alt2 . . . Alt2k(cid:20)hhy11, [y12, . . . [y1α1,hz11, [z12, . . . , [z1β1,
(f1(ad x11, . . . , ad x1,2kd1+m1))g1[w11, [w12, . . . , [w1γ1,
(f2(ad x21, . . . , ad x2,2kd2+m2))g2[w21, [w22, . . . , [w2γ2, th2
(cid:2)vh1
1 , [vh1
n−n, th1
2 , [. . . , [vh1
1 ] . . .(cid:3) . . .i, u11, . . . , u1q1i,
hhy21, [y22, . . . [y2α2,hz21, [z22, . . . , [z2β2,
. . . ,hhyr1, [yr2, . . . , [yrαr ,hzr1, [zr2, . . . , [zrβr ,
(fr(ad xr1, . . . , ad xr,2kdr+mr ))gr[wr1, [wr2, . . . , [wrγr, thr
2 ] . . .i, u21, . . . , u2q2i,
r ] . . .i, ur1, . . . , urqri(cid:21)
does not vanish under
[¯vn−n+1, [¯vn−n+2, [. . . , [¯vq, ¯tg
ziℓ = ¯ziℓ, wiℓ = ¯wiℓ. Note that f ∈ V G
the substitution vℓ = ¯vℓ, 1 6 ℓ 6 n − n,
t1 =
1] . . .]; ti = ¯ti for 2 6 i 6 r; uiℓ = ¯uiℓ, xiℓ = ψ−1(¯xiℓ), yiℓ = ¯yiℓ,
(cid:3)
n and satisfies all the conditions of the lemma.
Lemma 26. Let k, n0 be the numbers from Lemma 25. Then for every n > n0 there exists a
partition λ = (λ1, . . . , λs) ⊢ n, λi > 2k − C for every 1 6 i 6 d, with m(L, G, λ) 6= 0. Here
C := p((dim L)p + 3)((dim L) − d) where p ∈ N is such number that N p = 0.
Proof. Consider the polynomial f from Lemma 25.
IdG(L) for some tableau Tλ of the desired shape λ. It is known that F Sn =Lλ,Tλ
Pλ,Tλ
Tλf /∈
F Sne∗
Tλ
where the summation runs over the set of all standard tableax Tλ, λ ⊢ n. Thus F Snf =
Tλf /∈ IdG(L) for some λ ⊢ n. We claim that λ is of the desired
shape. It is sufficient to prove that λd > 2k − C, since λi > λd for every 1 6 i 6 d. Each row
of Tλ includes numbers of no more than one variable from each Xi, since e∗
= bTλaTλ and
Tλ
Tλf 6⊆ IdG(L) and e∗
It is sufficient to prove that e∗
F Sne∗
aTλ is symmetrizing the variables of each row. ThusPd−1
In virtue of Lemma 14,Pd
{z
}
d
Proof of Theorem 1. The Young diagram Dλ from Lemma 26 contains the rectangular subdi-
agram Dµ, µ = (2k − C, . . . , 2k − C
). The branching rule for Sn implies that if we consider
GRADED POLYNOMIAL IDENTITIES
25
i=1 λi > n − C. Therefore λd > 2k − C.
i=1 λi 6 2k(d−1)+(n−2kd) = n−2k.
(cid:3)
the restriction of Sn-action on M(λ) to Sn−1, then M(λ) becomes the direct sum of all
non-isomorphic F Sn−1-modules M(ν), ν ⊢ (n − 1), where each Dν is obtained from Dλ by
deleting one box. In particular, dim M(ν) 6 dim M(λ). Applying the rule (n − d(2k − C))
times, we obtain dim M(µ) 6 dim M(λ). By the hook formula,
where hij is the length of the hook with edge in (i, j). By Stirling formula,
dim M(µ) =
(d(2k − C))!
Qi,j hij
cG
n (L) > dim M(λ) > dim M(µ) >
(d(2k − C))!
((2k − C + d)!)d ∼
p2πd(2k − C)(cid:16) d(2k−C)
(cid:16)p2π(2k − C + d)(cid:0) 2k−C+d
(cid:17)d(2k−C)
(cid:1)2k−C+d(cid:17)d ∼ C9kr9d2kd
e
e
for some constants C9 > 0, r9 ∈ Q, as k → ∞. Since k =(cid:2) n−n0
The upper bound has been proved in Theorem 4.
2d (cid:3), this gives the lower bound.
(cid:3)
I am grateful to Yuri Bahturin and Mikhail Kotchetov for helpful discussions.
7. Acknowledgements
References
[1] Giambruno, A., Zaicev, M. V. Polynomial identities and asymptotic methods. AMS Mathematical Sur-
veys and Monographs Vol. 122, Providence, R.I., 2005.
[2] Zaitsev, M. V. Integrality of exponents of growth of identities of finite-dimensional Lie algebras. Izv.
Math. 66 (2002), 463 -- 487.
[3] Giambruno, A., Shestakov, I.P., Zaicev, M. V. Finite-dimensional non-associative algebras and codimen-
sion growth. Adv. Appl. Math. 47 (2011), 125 -- 139.
[4] Gordienko, A. S. Codimensions of polynomial identities of representations of Lie algebras. Proc. Amer.
Math. Soc. (to appear).
[5] Bahturin, Yu. A., Zaicev, M. V. Identities of graded algebras. J. Algebra, 205 (1998), 1 -- 12.
[6] Bahturin, Yu. A., Zaicev, M. V. Identities of graded algebras and codimension growth. Trans. Amer.
Math. Soc. 356:10 (2004), 3939 -- 3950.
[7] Bakhturin, Yu. A., Zaıtsev, M. V., Sehgal, S. K. G-identities of non-associative algebras. Sbornik: Math-
ematics, 190:11 (1999), 1559 -- 1570.
[8] Bahturin, Yu. A., Giambruno, A., Zaicev, M. V. G-identities on associative algebras. Proc. Amer. Math.
Soc., 127:1 (1999), 63 -- 69.
[9] Aljadeff, E., Giambruno, A., La Mattina, D. Graded polynomial identities and exponential growth. J.
reine angew. Math., 650 (2011), 83 -- 100.
[10] Giambruno, A., La Mattina, D. Graded polynomial identities and codimensions: computing the expo-
nential growth. Adv. Math., 225 (2010), 859 -- 881.
[11] Volichenko, I. B. Varieties of Lie algebras with identity [[X1, X2, X3], [X4, X5, X6]] = 0 over a field of
characteristic zero. (Russian) Sibirsk. Mat. Zh. 25:3 (1984), 40 -- 54.
[12] Zaicev, M. V., Mishchenko, S. P. An example of a variety of Lie algebras with a fractional exponent. J.
Math. Sci. (New York), 93:6 (1999), 977 -- 982.
[13] Mishchenko, S.P., Verevkin, A.B., Zaitsev, M.V. A sufficient condition for coincidence of lower and
upper exponents of the variety of linear algebras. Mosc. Univ. Math. Bull., 66:2 (2011), 86 -- 89.
[14] Taft, E.J. Invariant Wedderburn factors. Illinois J. Math., 1 (1957), 565 -- 573.
26
A. S. GORDIENKO
[15] Humphreys, J. E. Introduction to Lie algebras and representation theory. New-York, Springer-Verlag,
1978.
[16] Goto, M., Grosshans, F. Semisimple Lie algebras. Marcel Dekker, New York and Basel, 1978.
[17] Bakhturin, Yu. A. Identical relations in Lie algebras. VNU Science Press, Utrecht, 1987.
[18] Drensky, V. S. Free algebras and PI-algebras: graduate course in algebra. Singapore, Springer-Verlag,
2000.
[19] Berele, A. Cocharacter sequences for algebras with Hopf algebra actions. J. Alg., 185 (1996), 869 -- 885.
[20] Drensky, V.S. Representations of the symmetric group and varieties of linear algebras. Mathematics of
the USSR -- Sbornik, 43:1 (1982), 85 -- 101.
[21] Berele, A. Homogeneous polynomial identities. Israel J. Math, 42:3 (1982), 258 -- 272.
[22] Giambruno, A., Regev, A., Zaicev, M. V. Simple and semisimple Lie algebras and codimension growth.
Trans. Amer. Math. Soc., 352:4 (2000), 1935 -- 1946.
[23] James, G., Kerber, A. The representation theory of the symmetric group. Massachusetts, Addison-
Wesley, 1981.
[24] Fulton, W., Harris, J. Representation theory: a first course. New York , Springer-Verlag, 1991.
[25] Curtis, C. W., Reiner, I. Representation theory of finite groups and associative algebras. Pure and
Applied Mathematics, Vol. XI, John Wiley & Sons, New York-London-Sydney, 1962.
Memorial University of Newfoundland, St. John's, NL, Canada
E-mail address: [email protected]
|
1601.07827 | 1 | 1601 | 2016-01-28T16:47:12 | A non-abelian Hom-Leibniz tensor product and applications | [
"math.RA"
] | The notion of non-abelian Hom-Leibniz tensor product is introduced and some properties are established. This tensor product is used in the description of the universal ($\alpha$-)central extensions of Hom-Leibniz algebras. We also give its application to the Hochschild homology of Hom-associative algebras. | math.RA | math |
A non-abelian Hom-Leibniz tensor product and applications
J. M. Casas(1), E. Khmaladze(2) and N. Pacheco Rego(3)
(1)Dpto. Matem´atica Aplicada I, Univ. de Vigo, 36005 Pontevedra, Spain
e-mail address: [email protected]
(2)A. Razmadze Math. Inst. of I. Javakhishvili Tbilisi State University,
Tamarashvili Str. 6, 0177 Tbilisi, Georgia
e-mail address: [email protected]
(3)IPCA, Dpto. de Ciencias, Campus do IPCA, Lugar do Aldao
4750-810 Vila Frescainha, S. Martinho, Barcelos, Portugal
e-mail address: [email protected]
Abstract
The notion of non-abelian Hom-Leibniz tensor product is introduced
and some properties are established. This tensor product is used in the
description of the universal (α-)central extensions of Hom-Leibniz algebras.
We also give its application to the Hochschild homology of Hom-associative
algebras.
A. M. S. Subject Class. (2010): 17A30, 17B55, 17B60, 18G35, 18G60
Key words: Hom-Leibniz algebra, non-abelian tensor product, universal
(α)-central extension, Hom-associative algebra, Hochschild homology.
1
Introduction
Since the invention of Hom-Lie algebras as the algebraic model of deformed Lie
algebras coming from twisted discretizations of vectors fields by Hartwig, Larsson
and Silvestrov in [10], many papers appeared dealing with Hom-type generaliza-
tions of various algebraic structures (see for instance [1, 2, 3, 6, 7, 14, 15, 16,
17, 18, 19] and related references given therein).
In particular, Makhlouf and
Silvestrov introduced the notion of Hom-Leibniz algebras in [14], which simulta-
neously is a non-commutative generalization of Hom-Lie algebras and Hom-type
generalization of Leibniz algebras. In this generalized framework, it is natural to
seek for possible extensions of results in the categories of Lie or Leibniz algebras
to the categories of Hom-Lie or Hom-Leibniz algebras.
1
Recently, in [4], we developed the non-abelian Hom-Lie tensor product, ex-
tending the non-abelian Lie tensor product by Ellis [8] from Lie to Hom-Lie alge-
bras. It has applications in universal (α-)central extensions of Hom-Lie algebras
and cyclic homology of Hom-associative algebras.
In this paper we have chosen to work with Hom-Leibniz algebras. This work
is a continuation and non-commutative generalization of the work already begun
in [4]. Thus, we introduce a non-abelian Hom-Leibniz tensor product, extend-
ing the non-abelian Leibniz tensor product by Gnedbaye [9], which itself is the
Leibniz algebra (non-commutative) version of the non-abelian Lie tensor product
[8, 11]. Then we investigate properties of the non-abelian Hom-Leibniz tensor
product (Section 3) and give its applications to universal (α-)central extensions
of Hom-Leibniz algebras (Section 4) and Hochschild homology of Hom-associative
algebras (Section 5).
One observes that not all results can be transferred from Leibniz to Hom-
Leibniz algebras. For example, results on universal central extensions of Leib-
niz algebras can not be extended directly to Hom-Leibniz algebras because the
category of Hom-Leibniz algebras doesn't satisfy the so-called universal central
extension condition [5], which means that the composition of central extensions
is not central in general [3]. By this reason, the notion of α-central extension
of Hom-Leibniz algebras is introduced in [3], and classical results are divided
between universal central and universal α-central extensions of Hom-Leibniz al-
gebras (see Theorem 4.3). Further, Hom-type version of Gnedbaye's result relat-
ing Hochschild and Milnor type Hochschild homology of associative algebras [9],
doesn't hold for all Hom-associative algebras and requires an additional condition
(see Theorem 5.7).
Notations
Throughout this paper we fix K as a ground field. Vector spaces are considered
over K and linear maps are K-linear maps. We write ⊗ for the tensor product ⊗K
over K. For any vector space (resp. Hom-Leibniz algebra) L, a subspace (resp. a
two-sided ideal) L′ and x ∈ L we write x to denote the coset x + L′. We denote
by Lie and Lb the categories of Lie and Leibniz algebras, respectively.
2 Preliminaries on Hom-Leibniz algebras
2.1 Basic definitions
In this section we review some terminology on Hom-Leibniz algebras and recall
notions used in the paper. We also introduce notions of actions and semi-direct
product of Hom-Leibniz algebras.
2
Definition 2.1. [14] A Hom-Leibniz algebra is a triple (L, [−, −], αL) consisting
of a vector space L, a bilinear map [−, −] : L × L → L, called bracket operation,
and a linear map αL : L → L satisfying:
[αL(x), [y, z]] = [[x, y], αL(z)] − [[x, z], αL(y)] (Hom-Leibniz identity)
(2.1)
for all x, y, z ∈ L.
In the whole paper we only deal with (the so-called multiplicative [19]) Hom-
Leibniz algebras (L, [−, −], αL) such that αL preserves the bracket operation, that
is, αL[x, y] = [αL(x), αL(y)], for all x, y ∈ L.
Example 2.2.
a) Taking α = id in Definition 2.1, we obtain the definition of a Leibniz algebra
[12]. Hence any Leibniz algebra L can be considered as a Hom-Leibniz
algebra with αL = id.
b) Any Leibniz algebra L can be considered as a Hom-Leibniz algebra with
αL = 0. In fact, any vector space L endowed with any bracket operation is
a Hom-Leibniz algebra with αL = 0.
c) Any Hom-Lie algebra [10] is a Hom-Leibniz algebra whose bracket operation
satisfies the skew-symmetry condition.
d) Any Hom-dialgebra [18] (D, ⊣, ⊢, αD) becomes a Hom-Leibniz algebra (D,
[−, −], αD) with the bracket given by [x, y] = x ⊣ y − y ⊢ x, for all x, y ∈ D.
e) Let (L, [−, −]) be a Leibniz algebra and αL : L → L a Leibniz algebra
endomorphism. Define [−, −]α : L ⊗ L → L by [x, y]α = [α(x), α(y)], for
all x, y ∈ L. Then (L, [−, −]α, αL) is a Hom-Leibniz algebra.
f ) Any Hom-vector space (V, αV ), (i.e. V is a vector space and αV : V → V
is a linear map) together with the trivial bracket [−, −] (i.e. [x, y] = 0 for
all x, y ∈ V ) is a Hom-Leibniz algebra, called abelian Hom-Leibniz algebra.
g) The two-dimensional C-vector space L with basis {a1, a2}, endowed with the
bracket operation [a2, a2] = a1 and zero elsewhere, and the endomorphism
αL given by the matrix (cid:18) 1 1
0 1 (cid:19) is a non-Hom-Lie Hom-Leibniz algebra.
In the sequel we shall use the shortened notation (L, αL) for (L, [−, −], αL).
Definition 2.3. A homomorphism of Hom-Leibniz algebras f : (L, αL) → (L′, αL′)
is a linear map f : L → L′ such that
for all x, y ∈ L.
f ([x, y]) = [f (x), f (y)],
f ◦ αL = αL′ ◦ f,
3
Hom-Leibniz algebras and their homomorphisms form a category, which we
denote by HomLb. Example 2.2 a), b) and c) say that there are full embedding
functors
I0, I1 : Lb −→ HomLb,
I0(L) = (L, 0), I1(L) = (L, id)
and
inc : HomLie −→ HomLb,
(L, αL) 7→ (L, αL),
where HomLie denotes the category of Hom-Lie algebras. Note that the restric-
tions of I0 and I1 to the category of Lie algebras are full embeddings
0, I ′
I ′
1 : Lie −→ HomLie,
0(L) = (L, 0), I ′
I ′
1(L) = (L, id).
Definition 2.4. A Hom-Leibniz subalgebra (H, αH) of a Hom-Leibniz algebra
(L, αL) consists of a vector subspace H of L, which is closed under the bracket,
together with the linear map αH : H → H being the restriction of αL on H. In
such a case we may write αL for αH.
A Hom-Leibniz subalgebra (H, αH) of (L, αL) is said to be a two-sided Hom-
ideal if [x, y], [y, x] ∈ H, for all x ∈ H, y ∈ L.
If (H, αH) is a two-sided Hom-ideal of (L, αL), then the quotient vector space
L/H together with the endomorphism αL : L/H → L/H induced by αL, naturally
inherits a structure of Hom-Leibniz algebra, and it is called the quotient Hom-
Leibniz algebra.
The commutator of two-sided Hom-ideals (H, αL) and (K, αL) of a Hom-
Leibniz algebra (L, αL), denoted by ([H, K], αL), is the Hom-Leibniz subalgebra
of (L, αL) spanned by the brackets [h, k] and [k, h] for all h ∈ H, k ∈ K.
The following lemma can be readily checked.
Lemma 2.5. Let (H, αH ) and (K, αK) be two-sided Hom-ideals of a Hom-Leibniz
algebra (L, αL). The following statements hold:
a) (H ∩ K, αL) and (H + K, αL) are two-sided Hom-ideals of (L, αL);
b) [H, K] ⊆ H ∩ K;
c) If αL is surjective, then ([H, K], αL) is a two-sided Hom-ideal of (L, αL);
d) ([H, K], αL) is a two-sided Hom-ideal of (H, αH) and (K, αK). In particu-
lar, ([L, L], αL) is a two-sided Hom-ideal of (L, αL);
e) (αL(L), αL) is a Hom-Leibniz subalgebra of (L, αL);
f ) If H, K ⊆ αL(L), then ([H, K], αL) is a two-sided Hom-ideal of (αL(L), αL).
Definition 2.6. Let (L, αL) be a Hom-Leibniz algebra. The subspace Z(L) =
{x ∈ L [x, y] = 0 = [y, x], for all y ∈ L} of L is said to be the center of (L, αL).
4
Note that if αL : L → L is a surjective homomorphism, then (Z(L), αL) is a
two-sided Hom-ideal of L.
Corollary 2.7. Any Hom-Leibniz algebra (L, αL) gives rise to a Hom-Lie algebra
(LLie, αL), which is obtained as the quotient of L by the relation [x, x] = 0, x ∈ L.
Here αL is induced by αL. This defines a functor (−)Lie : HomLb −→ HomLie.
Moreover, the canonical epimorphism (L, αL) ։ (LLie, αL) is universal among all
homomorphisms from (L, αL) to a Hom-Lie algebra, implying that the functor
(−)Lie is left adjoint to the inclusion functor inc : HomLie −→ HomLb. This ad-
junction is a natural extension of the well-known adjunction Lie ⊥
Lb
between
inc
the categories of Lie and Leibniz algebras, in the sense that the following inner
and outer diagrams
(−)Lie
Lie
I ′
i
HomLie
(−)Lie
⊥
inc
inc
⊤
(−)Lie
Lb
Ii
HomLb
are commutative for i = 0, 1.
2.2 Hom-Leibniz actions and semi-direct product
Definition 2.8. Let (L, αL) and (M, αM ) be Hom-Leibniz algebras. A Hom-
Leibniz action of (L, αL) on (M, αM ) is a couple of linear maps, L ⊗ M → M,
x ⊗ m 7→ xm and M ⊗ L → M, m ⊗ x 7→ mx, satisfying the following identities:
a) αM (m)[x,y] = (mx)αL(y) − (my)αL(x) ,
b) [x,y]αM (m) = (xm)αL(y) − αL(x) (my) ,
c) αL(x) (ym) = −αL(x) (my) ,
d) αL(x) [m, m′] = [xm, αM (m′)] − [xm′, αM (m)] ,
e) [m, m′] αL(x) = [mx, αM (m′)] + [αM (m) , m′ x] ,
f ) [αM (m) , xm′] = − [αM (m) , m′ x] ,
g) αM (xm) = αL(x)αM (m) ,
h) αM (mx) = αM (m)αL(x) ,
5
/
/
o
o
/
/
o
o
/
/
o
o
for all x, y ∈ L and m, m′ ∈ M.
The action is called trivial if xm = 0 = mx, for all x ∈ L, m ∈ M.
Example 2.9.
a) A Hom-representation of a Hom-Leibniz algebra (L, αL) is a Hom-vector
space (M, αM ) equipped with two linear operations, L ⊗ M → M, x ⊗ m 7→
xm and M ⊗ L → M, m ⊗ x 7→ mx, satisfying the axioms a), b), c), g),
h) of Definition 2.8. Therefore, representations over a Hom-Leibniz algebra
(L, αL) are abelian Leibniz algebras enriched with Hom-Leibniz actions of
(L, αL).
b) Any action of a Leibniz algebra L on another Leibniz algebra M (see e.g.
[13]) gives a Hom-Leibniz action of (L, idL) on (M, idM ).
c) Let (K, αK) be a Hom-Leibniz subalgebra of a Hom-Leibniz algebra (L, αL)
and (H, αH) be a two-sided Hom-ideal of (L, αL). There exists a Hom-
Leibniz action of (K, αK) on (H, αH) given by the bracket in (L, αL) .
d) Let 0 → (M, αM ) i→ (K, αK) π→ (L, αL) → 0 be a split short exact sequence
of Hom-Leibniz algebras, i. e. there exists a homomorphism of Hom-Leibniz
algebras s : (L, αL) → (K, αK) such that π ◦ s = idL. Then there is a
Hom-Leibniz action of (L, αL) on (M, αM ) defined in the standard way:
xm = i−1[s(x), i(m)] and mx = i−1[i(m), s(x)] for all x ∈ L, m ∈ M.
Definition 2.10. Let (M, αM ) and (L, αL) be Hom-Leibniz algebras together
with a Hom-Leibniz action of (L, αL) on (M, αM ). Their semi-direct product
(M ⋊ L, α⋊) is the Hom-Leibniz algebra with the underlying vector space M ⊕ L,
endomorphism α⋊ : M ⋊ L → M ⋊ L given by α⋊ (m, l) = (αM (m) , αL (l)) and
bracket
[(m1, l1) , (m2, l2)] = (cid:16)[m1, m2] + αL(l1)m2 + mαL(l2)
1
, [l1, l2](cid:17) .
Let (M, αM ) and (L, αL) be Hom-Leibniz algebras with a Hom-Leibniz action
of (L, αL) on (M, αM ). Then we have the following short exact sequence of Hom-
Leibniz algebras
0 → (M, αM ) i→ (M ⋊ L, α⋊) π→ (L, αL) → 0 ,
(2.2)
where i : M → M ⋊ L, i(m) = (m, 0), and π : M ⋊ L → L, π (m, x) = x.
Moreover, this sequence splits by the Hom-Leibniz homomorphism L → M ⋊ L,
x 7→ (0, x).
6
2.3 Homology of Hom-Leibniz algebras
In this subsection we recall from [3] the construction of homology vector spaces
of a Hom-Leibniz algebra with coefficients in a Hom-co-representation.
Definition 2.11. A Hom-co-representation of a Hom-Leibniz algebra (L, αL) is a
Hom-vector space (M, αM ) together with two linear maps L ⊗ M → M, x ⊗ m 7→
xm and M ⊗ L → M, m ⊗ x = mx, satisfying the following identities:
a) [x,y]αM (m) = αL(x)(ym) − αL(y)(xm),
b) αM (m)[x,y] = (ym)αL(x) − αL(y)(mx),
c) (mx)αL(y) = −αL(y)(mx),
d) αM (xm) = αL(x)αM (m),
e) αM (mx) = αM (m)αL(x),
for any x, y ∈ L and m ∈ M
Example 2.12.
a) Let M be a co-representation of a Leibniz algebra L [13], then (M, idM ) is
a Hom-co-representation of the Hom-Leibniz algebra (L, idL).
b) The underlying Hom-vector space of a Hom-Leibniz algebra (L, αL) has a
Hom-co-representation structure given by xy = −[y, x] and yx = [y, x],
x, y ∈ L.
Let (L, αL) be a Hom-Leibniz algebra and (M, αM ) be a Hom-co-representation
⋆ (L, M) of (L, αL) with coefficients in (M, αM ) is
of (L, αL). The homology HLα
defined to be the homology of the chain complex (CLα
⋆ (L, M), d⋆), where
CLα
n(L, M) := M ⊗ L⊗n, n ≥ 0
and dn : CLα
n(L, M) → CLα
n−1(L, M), n ≥ 1, is the linear map given by
dn(m ⊗ x1 ⊗ · · · ⊗ xn) = mx1 ⊗ αL(x2) ⊗ · · · ⊗ αL(xn)
n
+
Xi=2
+ X1≤i<j≤n
(−1)i xim ⊗ αL(x1) ⊗ · · · ⊗ \αL(xi) ⊗ · · · ⊗ αL(xn)
(−1)j+1αM (m) ⊗ αL(x1) ⊗ · · · ⊗ αL(xi−1) ⊗ [xi, xj] ⊗ αL(xi+1)
⊗ · · · ⊗ \αL(xj) ⊗ · · · ⊗ αL(xn),
where the notation \αL(xi) indicates that the variable αL(xi) is omitted. Hence
HLα
n(L, M) := Hn(CLα
⋆ (L, M), d⋆), n ≥ 0.
7
Direct calculations show that HLα
0 (L, M) = M/M L, where M L = {mx m ∈
M, x ∈ L}, and if (M, αM ) is a trivial Hom-co-representation of (L, αL), that is
mx = xm = 0, then HLα
1 (L, M) = (M ⊗ L) / (αM (M) ⊗ [L, L]). In particular, if
M = K then HLα
n(L, K)
and it is called homology with trivial coefficients.
1 (L, K) = L/[L, L]. Later on we write HLα
n(L) for HLα
3 Non-abelian Hom-Leibniz tensor product
Let (M, αM ) and (N, αN ) be Hom-Leibniz algebras acting on each other. We
denote by M ∗ N the vector space spanned by all symbols m ∗ n, n ∗ m and
subject to the following relations:
λ(m ∗ n) = (λm) ∗ n = m ∗ (λn),
λ(n ∗ m) = b(λn) ∗ m = n ∗ (λm),
(m + m′) ∗ n = m ∗ n + m′ ∗ n,
m ∗ (n + n′) = m ∗ n + m ∗ n′,
(n + n′) ∗ m = n ∗ m + n′ ∗ m,
n ∗ (m + m′) = n ∗ m + n ∗ m′,
αM (m) ∗ [n, n′] = mn ∗ αN (n′) − mn′ ∗ αN (n),
αN (n) ∗ [m, m′] = nm ∗ αM (m′) − nm′ ∗ αM (m),
[m, m′] ∗ αN (n) = mn ∗ αM (m′) − αM (m) ∗ nm′,
[n, n′] ∗ αM (m) = nm ∗ αN (n′) − αN (n) ∗ mn′,
αM (m) ∗ m′n = −αM (m) ∗ nm′,
αN (n) ∗ n′m = −αN (n) ∗ mn′,
mn ∗ m′n′ = mn ∗ m′n′
,
mn ∗ n′m′
= mn ∗ n′m′,
nm ∗ m′n′ = nm ∗ m′n′
,
nm ∗ n′m′
= nm ∗ n′m′,
(3.1)
for all λ ∈ K, m, m′ ∈ M, n, n′ ∈ N.
We claim that (M ∗ N, αM ∗N ) is a Hom-vector space, where αM ∗N is the linear
map induced by αM and αN , i.e.
αM ∗N (m ∗ n) = αM (m) ∗ αN (n) , αM ∗N (n ∗ m) = αN (n) ∗ αM (m) .
Indeed, it can be checked readily that α∗ preserves all the relations in (3.1).
To be able to introduce the non-abelian Hom-Leibniz tensor product, we need
to assume that the actions are compatibly in the following sense.
Definition 3.1. Let (M, αM ) and (N, αN ) be Hom-Leibniz algebras with Hom-
8
Leibniz actions on each other. The actions are said to be compatible if
(mn)m′ = [mn, m′],
(nm)m′ = [nm, m′],
m(m′
n) = [m, m′ n],
m(nm′
) = [m,n m′],
(nm)n′ = b[nm, n′],
(mn)n′ = [mn, n′],
n(n′
m) = [n, n′ m],
n(mn′
) = [n,m n′],
(3.2)
for all m, m′ ∈ M and n, n′ ∈ N.
Example 3.2. If (H, αH) and (H ′, αH ′) both are Hom-ideals of a Hom-Leibniz
algebra (L, αL), then the Hom-Leibniz actions of (H, αH) and (H ′, αH ′) on each
other, considered in Example 2.9 c), are compatible.
Now we have the following property:
Proposition 3.3. Let (M, αM ) and (N, αN ) be Hom-Leibniz algebras acting com-
patibly on each other, then the Hom-vector space (M ∗ N, αM ∗N ) endowed with
the following bracket operation
[m ∗ n, m′ ∗ n′] = mn ∗m′ n′,
[m ∗ n, n′ ∗ m′] = mn ∗ n′m′
,
[n ∗ m, m′ ∗ n′] = nm ∗m′ n′,
[n ∗ m, n′ ∗ m′] = nm ∗ n′m′
,
(3.3)
is a Hom-Leibniz algebra.
Proof. Routine calculations show that the bracket given by (3.3) is compatible
with the defining relations in (3.1) of M ∗ N and can be extended from generators
to any elements. The verification of the Hom-Leibniz identity (2.1) is straight-
forward by using compatibility conditions in (3.2). Finally, it follows directly by
definition of αM ∗N that it preserves the bracket given by (3.3).
Definition 3.4. The above Hom-Leibniz algebra structure on (M ∗ N, αM ∗N ) is
called the non-abelian Hom-Leibniz tensor product of the Hom-Leibniz algebras
(M, αM ) and (N, αN ).
Remark 3.5. If αM = idM and αN = idN , then M ∗ N coincides with the non-
abelian tensor product of Leibniz algebras introduced in [9].
Remark 3.6. Let (M, αM ) and (N, αN ) be Hom-Lie algebras. One can readily
check that the following assertions hold:
a) Any Hom-Lie action of (M, αM ) on (N, αN ), M ⊗ N → N, m ⊗ n 7→ mn
(see [4, Definition 1.7] for the definition), gives a Hom-Leibniz action of
(M, αM ) on (N, αN ) by letting nm = −mn for all m ∈ M and n ∈ N.
9
b) For compatible Hom-Lie actions (see [4, Definition 2.1]) of (M, αM ) and
(N, αN ) on each other, the induced Hom-Leibniz actions are also compatible.
c) If (M, αM ) and (N, αN ) act compatibly on each other, then there is an
epimorphism of Hom-Leibniz algebras (M ∗ N, αM ∗N ) ։ (M ⋆ N, αM ⋆N )
defined on generators by m ∗ n 7→ m ⋆ n and n ∗ m 7→ −m ⋆ n, where ⋆
denotes the non-abelian Hom-Lie tensor product (see [4, Definition 2.4]).
Sometimes the non-abelian Hom-Leibniz tensor product can be described as
the tensor product of vector spaces. In particular, we have the following:
Proposition 3.7. If the Hom-Leibniz algebras (M, αM ) and (N, αN ) act trivially
on each other and both αM , αN are epimorphisms, then there is an isomorphism
of abelian Hom-Leibniz algebras
(M ∗ N, αM ∗N ) ∼= (cid:0)(M ab ⊗ N ab) ⊕ (N ab ⊗ M ab), α⊕(cid:1),
where M ab = M/[M, M], N ab = N/[N, N] and α⊕ denotes the linear self-map of
(M ab ⊗ N ab) ⊕ (N ab ⊗ M ab) induced by αM and αN .
Proof. Since the actions are trivial, then relations (3.3) enables us to see that
(M ∗ N, αM ∗N ) is an abelian Hom-Leibniz algebra.
Since αM and αN are epimorphisms, the defining relations (3.1) of the non-
abelian tensor product say precisely that the vector space M ∗ N is the quotient
of (M ⊗ N)⊕ (N ⊗ M) by the relations
m ∗ [n, n′] = [m, m′] ∗ n = [n, n′] ∗ m = n ∗ [m, m′] = 0
for all m, m′ ∈ M, n, n′ ∈ N. The later is isomorphic to (M ab⊗N ab)⊕(N ab⊗M ab)
and this isomorphism commutes with the endomorphisms α⊕ and αM ∗N .
The non-abelian Hom-Leibniz tensor product is functorial in the following
if f : (M, αM ) → (M ′, αM ′) and g : (N, αN ) → (N ′, αN ′) are homomor-
sense:
phisms of Hom-Leibniz algebras together with compatible actions of (M, αM )
(resp. (M ′, αM ′)) and (N, αN ) (resp. (N ′, αN ′)) on each other such that f , g
preserve these actions, that is
f (nm) = g(n)f (m),
g(mn) = f (m)g(n),
f (mn) = f (m)g(n),
g(nm) = g(n)f (m),
for all m ∈ M, n ∈ N, then we have a homomorphism of Hom-Leibniz algebras
f ∗ g : (M ∗ N, αM ∗N ) → (M ′ ∗ N ′, αM ′∗N ′)
defined by (f ∗ g)(m ∗ n) = f (m) ∗ g(n), (f ∗ g)(n ∗ m) = g(n) ∗ f (m).
10
g
→ (M3, αM3) → 0 be a
Proposition 3.8. Let 0 → (M1, αM1)
short exact sequence of Hom-Leibniz algebras. Let (N, αN ) be a Hom-Leibniz
algebra together with compatible Hom-Leibniz actions of (N, αN ) and (Mi, αMi)
(i = 1, 2, 3) on each other and f , g preserve these actions. Then there is an exact
sequence of Hom-Leibniz algebras
f
→ (M2, αM2)
(M1 ∗ N, αM1∗N )
f ∗idN−→ (M2 ∗ N, αM2∗N )
g∗idN−→ (M3 ∗ N, αM3∗N ) −→ 0.
Proof. Clearly g ∗ idN is an epimorphism and Im (f ∗ idN ) ⊆ Ker (g ∗ idN ). Since
Im (f ∗ idN ) is generated by all elements of the form f (m1) ∗ n, n ∗ f (m1), with
m1 ∈ M1, n ∈ N, it is a two-sided Hom-ideal in (M2 ∗ N, αM2∗N ) because of the
following equalities:
1 ) ∗ m2n′ ∈ Im (f ∗ idN ) ,
[f (m1) ∗ n, m2 ∗ n′] = f (m1)n ∗ m2n′ = f (m n
[f (m1) ∗ n, n′ ∗ m2] = f (m1)n ∗ n′ m2 = f (m n
1 ) ∗ n′ m2 ∈ Im (f ∗ idN ) ,
[n ∗ f (m1), m2 ∗ n′] = nf (m1) ∗ m2n′ = f (nm1) ∗ m2n′ ∈ Im (f ∗ idN ) ,
[n ∗ f (m1), n′ ∗ m2] = nf (m1) ∗ n′ m2 = f (nm1) ∗ n′ m2 ∈ Im (f ∗ idN ) ,
[m2 ∗ n′, f (m1) ∗ n] = m n′
∗ f (m1n) ∈ Im (f ∗ idN ) ,
[m2 ∗ n′, n ∗ f (m1)] = m n′
m1(cid:1) ∈ Im (f ∗ idN ) ,
2 ∗ f (cid:0)n
[n′ ∗ m2, f (m1) ∗ n] = n′m2 ∗f (m1) n = n′m2 ∗ f (m1n) ∈ Im (f ∗ idN ) ,
[n′ ∗ m2, n ∗ f (m1)] = n′m2 ∗ nf (m1) = n′m2 ∗ f (cid:0)n
m1(cid:1) ∈ Im (f ∗ idN ) ,
∗f (m1) n = m n′
2
∗ nf (m1) = mn′
αM2∗N (f (m1) ∗ n) = f (αM2 (m1)) ∗ αN (n) ∈ Im (f ∗ idN ) ,
αM2∗N (n ∗ f (m1)) = f (αN (n) ∗ αM2 (m1)) ∈ Im (f ∗ idN ) ,
2
2
for any m1 ∈ M1, m2 ∈ M2, n, n′ ∈ N. Thus, g ∗ idN induces a homomorphism
of Hom-Leibniz algebras
((M2 ∗ N) /Im (f ∗ idN ) , αM2∗N )−→(M3 ∗ N, αM3∗N ),
given on generators by m2 ∗ n 7→ g(m2) ∗ n and n ∗ m2 7→ n ∗ g(m2), which is an
isomorphism with the inverse map
(M3 ∗ N, αM3∗N ) −→ ((M2 ∗ N) /Im (f ∗ idN ) , αM2∗N )
defined by m3 ∗ n 7→ m2 ∗ n and n ∗ m3 = n ∗ m2, where m2 ∈ M2 such that
g(m2) = m3. Then the required exactness follows.
Proposition 3.9. If (M, αM ) is a two-sided Hom-ideal of a Hom-Leibniz algebra
(L, αL), then there is an exact sequence of Hom-Leibniz algebras
(cid:0)(M ∗ L) ⋊ (L ∗ M), α⋊(cid:1) σ−→ (L ∗ L, αL∗L)
τ−→ (L/M ∗ L/M, αL/M ∗L/M ) −→ 0.
where the Hom-Leibniz action of L ∗ M on M ∗ L is naturally given by the bracket
in L ∗ M.
11
Proof. First we note that τ is the functorial homomorphism induced by projec-
tion (L, αL) ։ (cid:0)L/M, αL/M(cid:1) and clearly it is an epimorphism.
Let σ′ : (M ∗ L, αM ∗L) → (L ∗ L, αL∗L), σ′′ : (L ∗ M, αL∗M ) → (L ∗ L, αL∗L)
be the functorial homomorphisms induced by the inclusion (M, αM ) ֒→ (L, αL)
and the identity map (L, αL) → (L, αL). Then σ : ((M ∗ L) ⋊ (L ∗ M) , α⋊) →
(L ∗ L, αL∗L), is defined by σ (x, y) = σ′(x) + αL∗L ◦ σ′′(y) for all x ∈ M ∗ L and
y ∈ L ∗ M, i. e.
σ((m1 ∗ l1), (l2 ∗ m2)) = m1 ∗ l1 + αL(l2) ∗ αM (m2),
σ((l1 ∗ m1), (l2 ∗ m2)) = l1 ∗ m1 + αL(l2) ∗ αM (m2),
σ((m1 ∗ l1), (m2 ∗ l2)) = m1 ∗ l1 + αM (m2) ∗ αL(l2),
σ((l1 ∗ m1), (m2 ∗ l2)) = l1 ∗ m1 + αM (m2) ∗ αL(l2),
for all m1, m2 ∈ M and l1, l2 ∈ L.
The exactness can be checked in the same way as in the proof of Proposition
3.8 and we omit it.
Proposition 3.10. Let (M, αM ) and (N, αN ) be Hom-Leibniz algebras with com-
patible actions on each other.
a) There are homomorphisms of Hom-Leibniz algebras
ψ1 : (M ∗ N, αM ∗N ) → (M, αM ), ψ1(m ∗ n) = mn, ψ1(n ∗ m) = nm,
ψ2 : (M ∗ N, αM ∗N ) → (N, αN ), ψ2(m ∗ n) = mn, ψ2(n ∗ m) = nm.
b) There is a Hom-Leibniz action of (M, αM ) (resp. (N, αN ) ) on the non-
abelian tensor product (M ∗ N, αM ∗N ) given, for all m, m′ ∈ M, n, n′ ∈ N,
by
m′(m ∗ n) = [m′, m] ∗ αN (n) −m′ n ∗ αM (m),
m′(n ∗ m) = m′n ∗ αM (m) − [m′, m] ∗ αN (n),
(m ∗ n)m′ = [m, m′] ∗ αN (n) + αM (m) ∗ nm′,
(n ∗ m)m′ = nm′ ∗ αM (m) + αN (n) ∗ [m, m′],
(cid:0) resp. n′(m ∗ n) = n′m ∗ αN (n) − [n′, n] ∗ αM (m),
n′(n ∗ m) = [n′, n] ∗ αM (m) −n m ∗ αN (n′),
(m ∗ n)n′ = mn′ ∗ αN (n) + αM (m) ∗ [n, n′] ,
(n ∗ m)n′ = [n, n′] ∗ αM (m) + αN (n) ∗ mn′ (cid:1).
d) Ker(ψ1) and Ker(ψ2) both are contained in the center of (M ∗ N, αM ∗N ).
e) The induced Hom-Leibniz action of Im(ψ1) (resp. Im(ψ2)) on Ker(ψ1) (resp.
Ker(ψ2)) is trivial.
f ) ψ1 and ψ2 satisfy the following properties for all m, m′ ∈ M, n, n′ ∈ N:
12
i) ψ1(m′(m ∗ n)) = [αM (m′), ψ1(m ∗ n)],
ii) ψ1((m ∗ n)m′) = [ψ1(m ∗ n), αM (m′)],
iii) ψ1(m′(n ∗ m)) = [αM (m′), ψ1(n ∗ m)],
iv) ψ1((n ∗ m)m′) = [ψ1(n ∗ m), αM (m′)],
v) ψ2(n′(m ∗ n)) = [αN (n′), ψ2(m ∗ n)],
vi) ψ2((m ∗ n)n′) = [ψ2(m ∗ n), αN (n′)],
vii) ψ2(n′(n ∗ m)) = [αN (n′), ψ2(n ∗ m)],
viii) ψ2((n ∗ m)n′) = [ψ2(n ∗ m), αN (n′)],
ix) ψ1(m∗n)(m′ ∗ n′) = [αM ∗N (m ∗ n), m′ ∗ n′] = ψ2(m∗n)(m′ ∗ n′),
x) ψ1(m∗n)(n′ ∗ m′) = [αM ∗N (m ∗ n), n′ ∗ m′] = ψ2(m∗n)(n′ ∗ m′),
xi) ψ1(n∗m)(m′ ∗ n′) = [αM ∗N (n ∗ m), m′ ∗ n′] = ψ2(n∗m)(m′ ∗ n′),
xii) ψ1(n∗m)(n′ ∗ m′) = [αM ∗N (n ∗ m), n′ ∗ m′] = ψ2(n∗m)(n′ ∗ m′),
xiii) (m′ ∗ n′)ψ1(m∗n) = [m′ ∗ n′, αM ∗N (m ∗ n)] = (m′ ∗ n′)ψ2(m∗n),
xiv) (n′ ∗ m′)ψ1(m∗n) = [n′ ∗ m′, αM ∗N (m ∗ n)] = (n′ ∗ m′)ψ2(m∗n),
xv) (m′ ∗ n′)ψ1(n∗m) = [m′ ∗ n′, αM ∗N (n ∗ m)] = (m′ ∗ n′)ψ2(n∗m),
xvi) (n′ ∗ m′)ψ1(n∗m) = [n′ ∗ m′, αM ∗N (n ∗ m)] = (n′ ∗ m′)ψ2(n∗m).
Proof. The proof requires routine combinations of equations in Definitions 2.8
and 3.1 with the defining relations (3.1) and (3.3) of the non-abelian Hom-Leibniz
tensor product.
Remark 3.11.
a) If αM = idM and αN = idN then the statement f ) of Lemma 3.10 says that
both ψ1 and ψ2 are crossed modules of Leibniz algebras (see [9]).
b) If M = N and the action of M on itself is given by the bracket in M, then
ψ1 = ψ2 and they are defined on generators by m ∗ m′ 7→ [m, m′]. In such
a case we write ψM for ψ1.
4 Application in universal (α-)central exten-
sions of Hom-Leibniz algebras
In this section we complement by new results the investigation of universal central
extensions of Hom-Leibniz algebras done in [3]. We also describe universal (α)-
central extensions via non-abelian Hom-Leibniz tensor product.
13
π
Definition 4.1. An epimorphism of Hom-Leibniz algebras (K, αK)
։ (L, αL),
with Ker(π) = (M, αM ), is called a central (resp. α-central) extension of (L, αL)
if [M, K] = 0 = [K, M], i.e. M ⊆ Z(K) (resp. [αM (M), K] = 0 = [K, αM (M)],
i.e. αM (M) ⊆ Z(K)).
Such a central extension is called universal central (resp. α-central) exten-
π
sion if, for every central (resp. α-central) extension (K ′, αK ′)
։ (L, αL) of
(L, αL) there exists one and only one homomorphism of Hom-Leibniz algebras
h : (K, αK) → (K ′, αK ′) such that π′ ◦ h = π.
Remark 4.2. Obviously every central extension is an α-central extension and
these notions coincide when αM = idM . On the other hand, every universal α-
central extension is a universal central extension and these notions coincide when
αM = idM . Let us also observe that if a universal (α-)central extension exists then
it is unique up to isomorphism.
The category HomLb is an example of a semi-abelian category which does not
satisfy universal central extension condition in the sense of [5], that is, the com-
position of central extensions of Hom-Leibniz algebras is not central in general,
but it is α-central extension (see Theorem 4.3 (a) below). This fact does not
allow complete generalization of classical results to Hom-Leibniz algebras and
the well-known properties of universal central extensions are divided between
universal central and universal α-central extensions of Hom-Leibniz algebras. In
particular, the assertions in Theorem 4.3 below are proved in [3].
To state the following theorem we need to recall that a Hom-Lie algebra
(L, αL) is said to be perfect if L = [L, L].
Theorem 4.3.
a) Let (K, αK)
π
։ (L, αL) and (F, αF )
ρ
։ (K, αK) be central extensions of
Hom-Leibniz algebras and (K, αK) be a perfect Hom-Leibniz algebra. Then
the composition extension (F, αF )
π◦ρ
։ (L, αL) is an α-central extension.
b) Let (K, αK)
π
։ (L, αL) and (K ′, αK ′)
π′
։ (L, αL) be two central extensions
of a Hom-Leibniz algebra (L, αL). If (K, αK) is perfect, then there exists at
most one homomorphism of Hom-Leibniz algebras f : (K, αK) → (K ′, αK ′)
such that π′ ◦ f = π.
c) Let (K, αK)
π′
π
։ (L, αL) be a central extension and (K ′, αK ′)
։ (L, αL)
be an α-central extension of a Hom-Leibniz algebra (L, αL).
If (K, αK)
is α-perfect, then there exists at most one homomorphism of Hom-Leibniz
algebras f : (K, αK) → (K ′, αK ′) such that π′ ◦ f = π.
d) If (K, αK)
π
։ (L, αL) is a universal α-central extension, then (K, αK) is a
perfect Hom-Leibniz algebra and every central extension of (K, αK) splits.
14
e) If (K, αK) is a perfect Hom-Leibniz algebra and every central extension of
π
։ (L, αL) is a universal
(K, αK) splits, then any central extension (K, αK)
central extension.
f ) A Hom-Leibniz algebra (L, αL) admits a universal central extension if and
only if (L, αL) is perfect. Furthermore, the kernel of the universal central
extension is canonically isomorphic to the second homology HLα
2 (L).
g) An α-perfect Hom-Leibniz algebra admits a universal α-central extension.
h) If (K, αK)
HLα
2 (K) = 0.
π
։ (L, αL) is a universal α-central extension, then HLα
1 (K) =
i) If HLα
1 (K) = HLα
2 (K) = 0, then any central extension (K, αK)
π
։ (L, αL)
is a universal central extension.
Given a Hom-Leibniz algebra (L, αL) we have the epimorphism of Hom-
Leibniz algebras
ψL : (L ∗ L, αL∗L) ։ ([L, L], αL), ψL(l ∗ l′) = [l, l′],
(cf. Remark 3.11 b)) which is a central extension of ([L, L], αL) because of the
equalities (3.3). Moreover, we have
Theorem 4.4. If (L, αL) is a perfect Hom-Leibniz algebra, then the central ex-
tension (L ∗ L, αL∗L)
ψL
։ (L, αL) is the universal central extension of (L, αL).
φ
Proof. Let (C, αC)
։ (L, αL) be a central extension of (L, αL). Since Ker (φ) ⊆
Z (C), we get a well-defined homomorphism of Hom-Leibniz algebras f : (L ∗ L,
αL∗L) → (C, αC) given on generators by f (l ∗ l′) = [cl, cl′], where cl and cl′ are
any elements in φ−1(l) and φ−1(l′), respectively. Obviously φ ◦ f = ψL and
f ◦ αL∗L = αC ◦ f , having in mind that αC(cl) ∈ φ−1(αL(l)) for all l ∈ L. Since
L is perfect, then by equalities (3.3), so is L ∗ L. Hence the homomorphism f is
unique by Theorem 4.3 b).
Remark 4.5. If the Hom-Leibniz algebra (L, αL) is perfect, by Theorem 4.3 e)
we have HLα
ψL→ L).
2 (L) ≈ Ker(L ∗ L
Theorem 4.6. Let (M, αM ) be a two-sided Hom-ideal of a perfect Hom-Leibniz
algebra (L, αL). Then there is an exact sequence of vector spaces
Ker(L ∗ M
ψ2−→ L) → HLα
2 (L) → HLα
2 (L/M) → M/[L, M] → 0.
15
Proof. By Proposition 3.9 there is a commutative diagram of Hom-Leibniz alge-
bras with exact rows
(cid:0)(M ∗ L) ⋊ (L ∗ M ), α⋊(cid:1)
ψ
0
/ (M, αM )
σ /
(L ∗ L, αL∗L)
τ
(L/M ∗ L/M, αL/M ∗L/M )
ψL
/ (L, αL)
ψL/M
π
/ (L/M, αL/M )
0
/ 0
where
and
ψ((m1 ∗ l1), (l2 ∗ m2)) = [m1, l1] + [αL(l2), αM (m2)],
ψ((l1 ∗ m1), (l2 ∗ m2)) = [l1, m1] + [αL(l2), αM (m2)],
ψ((m1 ∗ l1), (m2 ∗ l2)) = [m1, l1] + [αM (m2), αL(l2)],
ψ((l1 ∗ m1), (m2 ∗ l2)) = [l1, m1] + [αM (m2), αL(l2)],
ψL(l1 ∗ l2) = [l1, l2],
ψL/M (l1 ∗ l2) = [l1, l2],
for all m1, m2 ∈ M, l1, l2 ∈ L. Then, forgetting the Hom-Leibniz algebra struc-
tures, by using the Snake Lemma for the same diagram of vector spaces, the
assertion follows from Remark 4.5 and the fact that there is a surjective map
Ker(ψ) → Ker(ψ2).
Now suppose (L, αL) is a perfect Hom-Lie algebra. By [4, Theorem 3.4 and
Remark 3.5] we know that its universal central extension in the category Hom-Lie
is
(L ⋆ L, αL⋆L)
uL
։ (L, αL),
(4.1)
and Ker(uL) ≈ H α
2 (L), where ⋆ denotes the non-abelian Hom-Lie tensor product
and H α
2 (L) is the second homology with trivial coefficients of the Hom-Lie algebra
(L, αL). Furthermore, since any perfect Hom-Lie algebra is a perfect Hom-Leibniz
algebra as well, by Theorem 4.3 f ), (L, αL) admits universal central extension in
the category HomLb, which is described via the non-abelian Hom-Leibniz tensor
product in Theorem 4.4 and Remark 4.5. Then, by the universal property of (L ∗
ψL
L, αL∗L)
։ (L, αL) and considering (4.1) as a central extension of Hom-Leibniz
algebras, we get a homomorphism of Hom-Leibniz algebras ξ : (L∗L, αL∗L) → (L⋆
L, αL⋆L), ξ(l ∗ l′) = l ⋆ l′, such that ψL = uL ◦ ξ. Note that ξ is the epimorphism
described in Remark 3.6 c). Now the proof of the following proposition is obvious.
Proposition 4.7. Let (L, αL) be a perfect Hom-Lie algebra, then the non-abelian
Hom-Lie and Hom-Leibniz tensor squares are isomorphic, (L ⋆ L, αL⋆L) ≈ (L ∗ L,
αL∗L), if and only if the second Hom-Lie and Hom-Leibniz homologies with trivial
coefficients are isomorphic, H α
2 (L) ≈ HLα
2 (L).
Now we deal with the universal α-central extension of an α-perfect Hom-
Leibniz algebra. We need the following notion.
16
/
/
/
/
/
/
/
/
/
Definition 4.8. [3] A Hom-Leibniz algebra (L, αL) is said to be α-perfect if
Remark 4.9.
L = [αL(L), αL(L)]
a) When αL = id the notions of perfect and α-perfect Hom-Leibniz algebras
are the same.
b) Any α-perfect Hom-Lie algebra is an α-perfect Hom-Leibniz algebra.
c) Any α-perfect Hom-Leibniz algebra is perfect. Nevertheless the converse is
not true in general. An example of this is given in [4, Remark 3.8 b)].
d) If (L, αL) is α-perfect, then L = αL(L), i.e. αL is surjective.
It is shown in [3, Theorem 5.5 (b)] that an α-perfect Hom-Leibniz algebra
admits a universal α-central extension. Now we give its description via the non-
abelian Hom-Leibniz tensor product.
Theorem 4.10. Let (L, αL) be an α-perfect Hom-Leibniz algebra. Then the
epimorphism ψ : (cid:0)αL(L) ∗ αL(L), ααL(L)∗αL(L)(cid:1) ։ (L, αL), ψ (αL (l) ∗ αL (l′)) =
[αL (l) , αL (l′)], is its universal α-central extension.
Proof. Obviously ψ is a central extension. For every α-central extension (C, αC)
φ
։
(L, αL) there is a homomorphism f : (cid:0)αL(L) ∗ αL(L), ααL(L)∗αL(L)(cid:1) → (C, αC)
given on generators by f (αL (l) ∗ αL (l′)) = [αC (cl) , αC (cl′)], where cl and cl′
are any elements in φ−1(l) and φ−1(l′), respectively.
f is well-defined since
αC (Ker (φ)) ⊆ Z (C). Obviously φ ◦ f = ψ and f ◦ ααL(L)∗αL(L) = αC ◦ f ,
having in mind that αC(cl) ∈ φ−1(αL(l)) and α2
L(l)) for all l ∈ L.
Since L is α-perfect, then by equality (3.3), so is αL(L) ∗ αL(L). Hence the
homomorphism f is unique by Theorem 4.3 c).
C(cl) ∈ φ−1(α2
Lb
An alternative construction of the universal α-central extension (uce
α (L), α)
of an α-perfect Hom-Leibniz algebra (L, αL) is given in [3, Theorem 5.5 (b)]. Let
us recall that uce
α (L) is the quotient of αL(L) ⊗ αL(L) by the vector subspace
spanned by all elements −[x1, x2] ⊗ αL(x3) + [x1, x3] ⊗ αL(x2) + αL(x1) ⊗ [x2, x3],
where x1, x2, x3 ∈ L, the bracket of uce
α (L) is given on generators by
Lb
Lb
[αL(x1) ⊗ αL(x2), αL(y1) ⊗ αL(y2)] = [αL(x1), αL(x2)] ⊗ [αL(y1), αL(y2)],
x1, x2, y1, y2 ∈ L, and α : uce
Lb
α (L) → uce
α (L) is induced by αL.
Lb
Since the universal α-central extension of an α-perfect Hom-Leibniz algebra
(L, αL) is unique up to isomorphisms, then there is an isomorphism of Hom-
Leibniz algebras
(cid:0)α(L) ∗ α(L), αα(L)∗α(L)(cid:1) ≈ (uce
Lb
α (L), α).
17
uct (cid:0)α(L) ⋆ α(L), αα(L)⋆α(L)(cid:1) in [4, Theorem 3.11].
Now suppose that (L, αL) is an α-perfect Hom-Lie algebra. By [3, Theo-
rem 5.5 (a)] it admits a universal α-central extension in the category HomLie of
Hom-Lie algebras, which is described as the non-abelian Hom-Lie tensor prod-
(L, αL) admits a universal
α-central extension in the category HomLb described in Theorem 4.10 via non-
abelian Hom-Leibniz tensor product. Then [3, Proposition 5.6] provides the fol-
lowing relationship between the non-abelian Hom-Lie and Hom-Leibniz tensor
products:
(cid:0)α(L) ⋆ α(L), αα(L)⋆α(L)(cid:1) ≈ (cid:0)(α(L) ∗ α(L))Lie , αα(L)∗α(L)(cid:1) .
versal α-central extension in the category HomLb of Hom-Leibniz algebras is
Moreover, (cid:0)α(L) ⋆ α(L), αα(L)⋆α(L)(cid:1) is an α-perfect Hom-Lie algebra and its uni-
isomorphic to (cid:0)α(L) ∗ α(L), αα(L)∗α(L)(cid:1).
5 Application in Hochschild homology of Hom-
associative algebras
Definition 5.1. A (multiplicative) Hom-associative algebra (A, αA) is a vector
space A together with linear maps αA : A → A and A ⊗ A → A, a ⊗ b 7→ ab, such
that
αA(a)(bc) = (ab)αA(c),
αA(ab) = αA(a)αA(b)
for all a, b, c ∈ A.
Any Hom-associative algebra (A, αA) can be viewed as a Hom-Leibniz (in
fact Hom-Lie) algebra with the bracket defined by [a, b] := ab − ba, a, b ∈ A (see
Example 5.2 in [19]).
Given a Hom-associative algebra (A, αA), we denote Lα(A) := (A⊗A)/Im(b3),
where b3 : A⊗A⊗A → A⊗A, b3 (a ⊗ b ⊗ c) = ab⊗αA(c)−αA(a)⊗bc+ca⊗αA(b),
is the boundary map for the Hochschild complex of the Hom-associative algebra
(A, αA) constructed in [19]. Then we have the following short exact sequence of
vector spaces
0 −→ HH α
1 (A)
i−→ Lα(A)
φ
−→ [A, A] −→ 0 ,
(5.1)
where HH α
1 (A) denotes the first Hochschild homology of (A, αA) [19], [A, A] is
the subspace of A spanned by the elements ab − ba and φ(a ⊠ b) = ab − ba,
a, b ∈ A. Here a ⊠ b stands for the equivalence class of a ⊗ b.
Suppose αA : Lα(A) → Lα(A) is the linear map induced by αA, i.e. αA(a ⊠ b) =
αA(a) ⊠ αA(b). Then there is a Hom-Leibniz algebra structure on (Lα(A), αA)
given by the bracket
[a ⊠ b, a′ ⊠ b′] := (ab − ba) ⊠ (a′b′ − b′a′) = [a, b] ⊠ [a′, b′],
18
for all a, b, a′, b′ ∈ A. Indeed, it is immediate that
[a, b] ⊠ αA(c) − αA(a) ⊠ [b, c] + [c, a] ⊠ αA(b) = 0,
(5.2)
then the Hom-Leibniz identity (2.1) easily follows.
Remark 5.2. (Lα(A), αA) is the quotient of the non-abelian Hom-Leibniz tensor
product (A ∗ A, αA∗A) by the two-sided Hom-ideal generated by the elements of
the form ab ∗ αA(c) − αA(a) ∗ bc + ca ∗ αA(b), for all a, b, c ∈ A.
Definition 5.3. [4] We say that a Hom-associative algebra (A, αA) satisfies the
α-identity condition if
[A, Im(αA − idA)] = 0 ,
(5.3)
where [A, Im(αA − idA)] = 0 is the subspace of A spanned by all elements ab − ba,
with a ∈ A and b ∈ Im(αA − idA).
Examples of Hom-associative algebras satisfying the α-identity condition can
be found in [4].
Proposition 5.4. Let (A, αA) be a Hom-associative algebra.
a) There is a Hom-Leibniz action of (A, αA) on (Lα(A), αA) given, for all
a′ ∈ A, a ⊠ b ∈ Lα(A), by
a′
(a ⊠ b) = [a′, a] ⊠ αA (b) − [a′, b] ⊠ αA (a) ,
(a ⊠ b)a′
= [a, a′] ⊠ αA (b) + αA (a) ⊠ [b, a′] ,
and a Hom-Leibniz action of (Lα(A), αA) on (A, αA) given by
(a⊠b)a′ = [[a, b] , a′] ,
a′(a⊠b) = [a′, [a, b]] .
Moreover, these actions are compatible if (A, αA) satisfies the α-identity
condition (5.3).
b) There is a short exact sequence of Hom-Leibniz algebras
0 → (HH α
1 (A), αA) i→ (Lα(A), αA)
φ
→ ([A, A], αA) → 0
where (HH α
1 (A), αA) is considered as an abelian Hom-Leibniz algebra, αA
(resp. αA) is the restriction of αA (resp. αA) and φ(a⊠b) = [a, b] = ab−ba.
c) The induced Hom-Leibniz action of (A, αA) on (HH α
1 (A), αA) is trivial.
Moreover, i always preserves the actions and, if (A, αA) satisfies the α-
identity condition (5.3), then φ also preserves the actions.
19
Proof. a) By using formula (5.2) and Remark 5.2, one can readily check that the
actions are well-defined Hom-Leibniz actions.
It is also easy to see that these
actions are compatible under condition (5.3). For example, we have
(a′
(a⊠b))a′′ = [a′,a]⊠αA(b)a′′ − [a′,b]⊠αA(a)a′′ = (cid:2)[[a′, a], αA(b)] − [[a′, b], αA((a))], a′′(cid:3)
(2.1)
= (cid:2)[αA(a′), [a, b]], a′′(cid:3) (5.3)
= (cid:2)[a′, [a, b]], a′′(cid:3) = [a′ a⊠b, a′′].
b) It is an immediate consequence of (5.1) and the definitions above.
c) Since a′ (a ⊠ b) = αA(a′) ⊠ [a, b] = 0 when a ⊠ b ∈ Ker(φ), we deduce that
the action of (A, αA) on (HH α
1 (A), αA) is trivial and, by this reason, i preserves
the actions of (A, αA). Now, if (A, αA) satisfies the α-identity condition (5.3),
then we have
(a ⊠ b)(cid:1) = (cid:2)αA(a′), [a, b](cid:3) (5.3)
φ(cid:0)a′
φ(cid:0)(a ⊠ b)a′(cid:1) = (cid:2)[a, b], αA(a′)(cid:3) (5.3)
= (cid:2)a′, [a, b](cid:3) = a′
= (cid:2)[a, b], a′(cid:3) = φ(a ⊠ b)a′
φ(a ⊠ b),
,
for all a′, a, b ∈ A, implying that φ also preserves the actions.
Remark 5.5. The statements in Proposition 5.4 when αA = idA recover the
corresponding results established in [9, Section 7] for Leibniz algebras.
In the following definition we introduce the Hom-version of the first Milnor-
type Hochschield homology of associative algebras (see e.g. [9]).
Definition 5.6. Let (A, αA) be a Hom-associative algebra. The first Milnor-type
Hochschield homology HH M
1 (A, αA) is the quotient of the vector space A ⊗ A by
the relations
ab ⊗ αA(c) − αA(a) ⊗ bc + ca ⊗ αA(b) = 0,
αA(a) ⊗ bc − αA(a) ⊗ cb = 0,
ab ⊗ αA(c) − ba ⊗ αA(c) = 0,
for any a, b, c ∈ A.
Theorem 5.7. Let (A, αA) be a Hom-associative (non-commutative) algebra sat-
isfying the α-identity condition (5.3). Then there is an exact sequence of vector
spaces
A ∗ HH α
1 (A) → Ker(A ∗ Lα(A) → Lα(A)) → Ker(A ∗ [A, A] → [A, A])
1 (A, αA) → [A, A]/[A, [A, A]] → 0.
1 (A) → HH M
→ HH α
20
Proof. By using Proposition 5.4 and Proposition 3.8 we have the commutative
diagram of Hom-Leibniz algebras (written without α endomorphisms)
A ∗ HH α
1 (A)
A ∗ Lα(A)
A ∗ [A, A]
0
ψ2
ψ2
ψ3
0
/ HH α
1 (A)
/ Lα(A)
φ
/ [A, A]
/ 0,
where ψ homomorphisms are defined as in Proposition 3.10. ψ2(a ∗ (x ⊗ y)) =
a(x⊗y), ψ2((x⊗y)∗a) = (x⊗y)a and ψ3(a∗[x, y]) = [αA(a), [x, y]], ψ2((x⊗y)∗a) =
[[x, y], αA(a)].
Thus Coker(ψ3) = [A, A]/[A, [A, A]], Coker(ψ2) = HH M
1 (A) and Ker(ψ2) = A ∗ HH α
1 (A, αA), Coker(ψ2) =
1 (A). Then the assertion is a consequence of
HH α
the Snake Lemma.
Remark 5.8.
a) Let (A, αA) be a Hom-associative (non-commutative) algebra satisfying the
α-identity condition (5.3). If in addition αA is an epimorphism, then the
term A ∗ HH α
1 (A) in the exact sequence of Theorem 5.7 can be replaced by
Proposition 3.7. In Particular, if αA = idA, the exact sequence in Theorem
5.7 coincides with that of [9, Theorem 7.4].
1 (A) ⊗A/[A, A](cid:1) since they are isomorphic by
(cid:0)A/[A, A] ⊗HH α
1 (A)(cid:1)⊕(cid:0)HH α
b) From the exact sequence in Theorem 5.7 and Theorem 4.4 immediately fol-
lows that the vector spaces HH α
1 (A, αA) are isomorphic when
the Hom-associative algebra (A, αA) is perfect as Hom-Leibniz algebra and
HLα
1 (A) and HH M
2 (A) = 0.
1 (A) and HH M
c) HH α
1 (A, αA) are also isomorphic for any commutative Hom-
associative algebra (A, αA).
Acknowledgements
First and second authors were supported by Ministerio de Econom´ıa y Competi-
tividad (Spain) (European FEDER support included), grant MTM2013-43687-P.
The second author was supported by Xunta de Galicia, grants EM2013/016 and
GRC2013-045 (European FEDER support included), and Shota Rustaveli Na-
tional Science Foundation, grant FR/189/5-113/14.
References
[1] F. Ammar, S. Mabrouk and A. Makhlouf, Representations and cohomology
of n-ary multiplicative Hom-Nambu-Lie algebras, J. Geom. Phys. 61 (10)
(2011), 1898 -- 1913.
21
/
/
/
/
/
/
/
/
/
/
[2] S. Benayadi and A. Makhlouf, Hom-Lie algebras with symmetric invariant
nondegenerate bilinear forms, J. Geom. Phys. 76 (2014), 38 -- 60.
[3] J. M. Casas, M. A. Insua and N. Pacheco Rego, On universal central exten-
sions of Hom-Leibniz algebras, J. Algebra Appl. 13 (8) (2014), 1450053 (22
pp.).
[4] J. M. Casas, E. Khmaladze and N. Pacheco Rego, A non-abelian tensor prod-
uct of Hom-Lie algebras, Bull. Malaysian Math. Sci. Soc (2016), to appear.
[5] J. M. Casas and T. Van der Linden, Universal central extensions in semi-
abelian categories, Appl. Categor. Struct. 22 (1) (2014), 253 -- 268.
[6] D. Chen and Y. Sheng, Hom-Lie 2-algebras, J. Algebra 376 (2013), 174 -- 195.
[7] Y. S. Chen and Y. C. Su, (Co)homology and universal central extensions of
Hom-Leibniz algebras, Acta Math. Sin. (Engl. Ser.) 27 (5) (2011), 813 -- 830.
[8] G. Ellis, A non-abelian tensor product of Lie algebras, Glasgow Math. J. 33
(1991), 101 -- 120.
[9] A. V. Gnedbaye, A non-abelian tensor product of Leibniz algebras, Ann. Inst.
Fourier, Grenoble 49 (4) (1999), 1149 -- 1177.
[10] J. T. Hartwing, D. Larson and S. D. Silvestrov, Deformations of Lie algebras
using σ-derivations, J. Algebra 295 (2006), 314 -- 361.
[11] N. Inassaridze, E. Khmaladze and M. Ladra, Non-abelian homology of Lie
algebras, Glasgow Math. J. 46 (2004), 417 -- 429.
[12] J.-L. Loday, Une version non commutative des alg`ebres de Lie: les alg`ebres
de Leibniz, L'Enseignement Math´ematique 39 (1993), 269 -- 292.
[13] J.-L. Loday and T. Pirashvili, Universal enveloping algebras of Leibniz alge-
bras and (co)homology, Math. Ann. 296 (1993), 139 -- 158.
[14] A. Makhlouf and S. Silvestrov, Hom-algebra structures, J. Gen. Lie Theory
Appl. 2 (2) (2008), 51 -- 64.
[15] A. Makhlouf and S. Silvestrov, Hom-Lie admissible Hom-coalgebras and
Hom-Hopf algebras, Generalized Lie theory in mathematics, physics and
beyond, (Springer, Berlin 2009) 189 -- 206.
[16] A. Makhlouf and S. Silvestrov, Notes on 1-parameter formal deformations of
Hom-associative and Hom-Lie algebras, Forum Math. 22 (4) (2010), 715 -- 739.
[17] Y. Sheng, Representations of Hom-Lie algebras, Algebr. Represent. Theory
15 (6) (2012), 1081 -- 1098.
22
[18] D. Yau, Enveloping algebras of Hom-Lie algebras, J. Gen. Lie Theory Appl.
2 (2) (2008), 95 -- 108.
[19] D. Yau, Hom-algebras and homology, J. Lie Theory 19 (2) (2009), 409 -- 421.
23
|
1307.7891 | 1 | 1307 | 2013-07-30T09:50:03 | Symmetric Powers of Hyperbolic Forms and of Trace Forms on Symbol Algebras | [
"math.RA",
"math.RT"
] | Let $K$ be a field with characteristic different from 2 and let $S$ be a symbol algebra over $K$. We compute the symmetric powers of hyperbolic quadratic forms over $K$. Also, we compute the symmetric powers of the quadratic trace form of $S$. In both cases we apply a generalised form of the Vandermonde convolution in the course of the computations. | math.RA | math |
SYMMETRIC POWERS OF HYPERBOLIC FORMS
AND OF TRACE FORMS ON SYMBOL ALGEBRAS
RONAN FLATLEY
Abstract. Let K be a field with characteristic different from 2 and let S be a symbol
algebra over K. We compute the symmetric powers of hyperbolic quadratic forms
over K. Also, we compute the symmetric powers of the quadratic trace form of S . In
both cases we apply a generalised form of the Vandermonde convolution in the course
of the computations.
1. Introduction
We assume throughout this paper that K is a field with characteristic different from
2. In [3] we computed the exterior powers of hyperbolic forms over K and used these
results to compute the exterior powers of the quadratic trace form of symbol algebras
over K. Here we are interested in finding corresponding results for the symmetric
powers of such forms.
In [8], McGarraghy investigated symmetric powers of classes of symmetric bilin-
ear forms in the Witt-Grothendieck ring of K, deriving their basic properties as well
as computing several of their classical invariants such as determinant, signature with
respect to an ordering and Hasse invariant. Their relation to exterior powers was also
given. For a given symmetric bilinear form ϕ over K there was a distinction drawn
between factorial symmetric powers Skϕ and non-factorial symmetric powers S kϕ.
Factorial symmetric powers were seen to have the advantages of being computable
for any given symmetric bilinear form and of being independent of basis. However,
effectively they are useful only in characteristic 0. On the other hand, non-factorial
symmetric powers are independent of characteristic but require the form to be diag-
onalised, i.e., a basis needs to be chosen.
In this paper we deal with non-factorial
symmetric powers of forms, hereafter referred to simply as symmetric powers. Since
char(K) , 2, we use the one-to-one correspondence between symmetric bilinear forms
and quadratic forms.
Notation and terminology is borrowed from Lam [7] and Scharlau [9]. A diago-
nalised quadratic form over K with coefficients a1, . . . , am ∈ K× is denoted by ha1, . . . , ami.
The hyperbolic plane h1, −1i is denoted by H. If ϕ and ψ are forms over K then ϕ ≃ ψ
means that these forms are isometric. The tensor product of the 1-dimensional form
Date: September 25, 2018.
2010 Mathematics Subject Classification. Primary: 11E81; Secondary: 16K20.
Key words and phrases. hyperbolic form, symmetric power, symbol algebra, trace form.
1
2
RONAN FLATLEY
hni with ϕ is denoted hniϕ to be distinguished from n copies of ϕ which is written as
n × ϕ.
2. Preliminaries
2.1. Exterior powers. Bourbaki defined the concept of exterior power of a symmetric
bilinear form in [1, IX, §1, (37)]:
Definition 2.1. Let V be a vector space of dimension m over K. Let ϕ : V × V → K
be a non-singular bilinear form and let k be a positive integer, k ≤ m. We define the
k-fold exterior power of ϕ,
Λkϕ : ΛkV × ΛkV −→ K
by
Λkϕ(x1 ∧ · · · ∧ xk, y1 ∧ · · · ∧ yk) = det(cid:0)ϕ(xi, y j)(cid:1)1≤i, j≤k.
The bilinear extension of this form defines Λkϕ everywhere on ΛkV × ΛkV. We define
Λ0ϕ := h1i. Also, for k > m, we define Λkϕ := 0, the zero form.
We use the following two results on exterior powers of hyperbolic forms:
Proposition 2.2. [3, Proposition 5.6] Let φ ≃ h × H where h ∈ N and k odd with
1 ≤ k ≤ 2h − 1. Then
Λkφ ≃
1
2 2h
k ! × H.
Proposition 2.3. [3, Proposition 5.7] Let φ ≃ h×H where h ∈ N, k = 2ℓ and 0 ≤ ℓ ≤ h.
Then
Λkφ = Λ2ℓφ ≃ h
ℓ! × h(−1)ℓi ⊥
1
2 2h
2ℓ! − h
ℓ!! × H.
2.2. Symmetric powers.
Definition 2.4. [8, adapted statement of Proposition 3.5] Let V be a vector space of
dimension m over K. Let ϕ be a symmetric bilinear form over K with ϕ ≃ ha1, . . . , ami.
Let k be a non-negative integer. We define the k-fold symmetric power of ϕ,
by
S kϕ : S kV × S kV → K
S kϕ = ⊥
+···+kiℓ
1≤i1<···<iℓ≤m
=k
ki1
haki1
i1
. . . akiℓ
iℓ
i.
Remark 2.5. The form S kϕ has dimension (cid:16)m+k−1
k
(cid:17).
Remark 2.6. Clearly, S 0ϕ = h1i and S 1ϕ = ϕ.
Remark 2.7. We note that S k preserves isometries, i.e., ϕ ≃ ψ ⇒ S kϕ ≃ S kψ. In
general the converse is not true; consider a field where h1i ; h−1i. Let ϕ = h1, 1i ⊥ H
and ψ = h−1, −1i ⊥ H. Then ϕ ; ψ but S 2ϕ ≃ S 2ψ = 4 × h1i ⊥ 3 × H.
SYMMETRIC POWERS OF HYPERBOLIC FORMS AND OF TRACE FORMS ON SYMBOL ALGEBRAS3
Remark 2.8. The k-fold exterior power is always a subform of the k-fold symmetric
power. Therefore, Λkϕ isotropic ⇒ S kϕ isotropic.
Lemma 2.9. S k(m × h1i) = (cid:16)m+k−1
k
(cid:17) × h1i and S k(m × h−1i) = (cid:16)m+k−1
k
(cid:17) × h(−1)ki.
The next two propositions give properties that are very useful for the computations
in Sections 3 and 4.
Proposition 2.10. [6, p. 47] Let ϕ and ψ be symmetric bilinear forms over K and let k
be a positive integer. Then
S k(ϕ ⊥ ψ) = ⊥
i+ j=k
S iϕ ⊗ S jψ .
Proposition 2.11. [8, Proposition 4.4] Let ϕ be an n-dimensional symmetric bilinear
form over K and let k be a positive integer. Then
S kϕ =
⌈k/2⌉
i=0 n + i − 1
⊥
i
! × Λk−2iϕ .
3. Symmetric powers of hyperbolic forms
For each dimension of even parity, there exists exactly one hyperbolic form up to
isometry. In other words, if ϕ is a hyperbolic form of dimension 2h, then ϕ ≃ h × H.
In this section we compute the symmetric powers of these forms. For computational
convenience, we often do not state the specific number of hyperbolic planes. Instead
we shall use the word "Hyp" to represent the number of hyperbolic planes required in
order to give full dimension to the particular symmetric power in question.
Proposition 3.1. Let k be any positive odd integer, h ∈ N. Then
S k(h × H) =
Proof.
S k(h × H) = S k(h × h1i ⊥ h × h−1i)
1
2 2h + k − 1
k
! × H.
= ⊥
i+ j=k
= ⊥
i+ j=k
j even
S i(h × h1i) ⊗ S j(h × h−1i)
S i(h × h1i) ⊗ S j(h × h−1i) ⊥ ⊥
i+ j=k
j odd
S i(h × h1i) ⊗ S j(h × h−1i)
[by Proposition 2.10]
= ⊥
i+ j=k
j even
h + i − 1
i
! × h1i! ⊗ h + j − 1
j
! × h1i! ⊥ ⊥
i+ j=k
j odd
h + i − 1
i
! × h1i! ⊗ h + j − 1
j
! × h−1i!
4
RONAN FLATLEY
[by Lemma 2.9]
h + i − 1
i
h + i − 1
i
= Xi+ j=k
j even
= Xi+ j=k
j even
! h + j − 1
j
! h + j − 1
j
! × h1i ⊥ Xi+ j=k
! × h1i ⊥ Xi+ j=k
j odd
i even
h + i − 1
i
! h + j − 1
j
! × h−1i
h + j − 1
j
! h + i − 1
i
! × h−1i .
From the expression above we observe that there is the same number of summands h1i
as summands h−1i. Given that the dimension of S k(h × H) is(cid:16)2h+k−1
k
(cid:17), we conclude that
S k(h × H) =
1
2 2h + k − 1
k
! × H.
We now seek to compute S k(h × H) for k an arbitrary even natural number. First, we
need two lemmata:
Lemma 3.2.
r
Xj=0
(−1) j 2p + j − 1
j
r − j! = (−1)r p + r − 1
! p
r
! .
Proof. By the Vandermonde convolution (see [4] for example), for any integer p and
any positive integer r,
r
j ! p
By applying the "minus transformation" (cid:16)−q
Xj=0 −2p
r − j! = −p
r ! .
i (cid:17) = (−1)i(cid:16)q+i−1
required identity.
i
(cid:17) to both sides, we get the
Lemma 3.3. [Pascal's rule]
For any r ∈ N, 1 ≤ s ≤ r, we have that
r − 1
s
! + r − 1
s − 1! = r
s! .
Remark 3.4. In several of the remaining proofs we use the following property of
binomial coefficients, which is easy to derive:
Proposition 3.5. Let k be any non-negative even integer, k = 2ℓ, h ∈ N. Then
r
s! =
r
s r − 1
s − 1! .
S k(h × H) = S 2ℓ(h × H) = h + ℓ − 1
ℓ
! × h1i ⊥
1
2 2h + 2ℓ − 1
2ℓ
! − h + ℓ − 1
ℓ
!! × H.
SYMMETRIC POWERS OF HYPERBOLIC FORMS AND OF TRACE FORMS ON SYMBOL ALGEBRAS5
Proof. Using the formula
⌈k/2⌉
S kφ =
i=0 n + i − 1
⊥
i
! × Λk−2iϕ
[Proposition 2.11]
we get
S k(h × H) = S 2ℓ(h × H)
=
=
ℓ
i
i=0 2h + i − 1
⊥
i=0 2h + i − 1
⊥
i
ℓ
! × Λ2ℓ−2i(h × H)
! × h
ℓ − i! × h(−1)ℓ−ii ⊥
1
2 2h
2(ℓ − i)! − h
ℓ − i!! × H!
[by Proposition 2.3
ℓ
ℓ
i
i
ℓ
=
=
i=0 2h + i − 1
! h
ℓ − i! × h(−1)ℓ−ii ⊥ Hyp
⊥
i=0 2h + i − 1
! h
ℓ − i! × h(−1)i−ℓi ⊥ Hyp
⊥
(−1)i 2h + i − 1
! h
ℓ − i! × h(−1)−ℓi ⊥ Hyp
Xi=0
= (−1)ℓ h + ℓ − 1
= h + ℓ − 1
! × h(−1)−ℓi ⊥ Hyp
! × h1i ⊥ Hyp .
=
ℓ
ℓ
i
[by Lemma 3.2]
4. Symmetric powers of trace forms on symbol algebras
Let n be an arbitrary positive integer and let K contain a primitive n-th root of unity
ω. Let a, b ∈ K× and let S be the algebra over K generated by elements x and y where
xn = a,
yn = b
and
yx = ωxy.
We call this algebra a symbol algebra (see [5, Chapter 1, §2]) and denote it as (a, b; n, K, ω).
In [2, §11], Draxl shows it to be a central simple algebra over K of degree n.
Let A be a central simple algebra of degree n over a field K of characteristic different
from 2. We write TA : A → K for the quadratic trace form
for z ∈ A,
TA(z) = TrdA(z2)
where TrdA is the reduced trace of A.
In [3], we proved the following:
6
RONAN FLATLEY
Proposition 4.1. [3, Propositions 2.1, 3.1] For S = (a, b; n, K, ω), we have
(i) TS ≃ h(−1)n/2i ⊥
n2 − 1
2
× H
(ii) TS ≃ hnih1, a, b, (−1)n/2abi ⊥
n2 − 4
2
× H
for n odd.
for n even.
Proposition 4.2. Let n be odd and k ≥ 0. Then
S kTS =
Proof. Let k be odd. Then
2
k−1
2
(cid:16) n2 +k−2
(cid:16) n2 +k−1
2
k
2
2 i ⊥ Hyp,
(cid:17) × h(−1) n−1
(cid:17) × h1i ⊥ Hyp,
if k is odd;
if k is even.
[by Proposition 4.1]
[by Proposition 2.10]
S i(cid:16)h(−1) n−1
2 i(cid:17) ⊗ S j n2 − 1
2
× H!
[by Proposition 3.1]
[by Lemma 2.9]
S kTS = S k h(−1) n−1
2 i ⊥
= ⊥
i+ j=k
= ⊥
j even
S i(cid:16)h(−1) n−1
S i(cid:16)h(−1) n−1
S i(cid:16)h(−1) n−1
(n−1)i
2
2
2
n2 − 1
j odd
× H!
2 i(cid:17) ⊗ S j n2 − 1
2 i(cid:17) ⊗ S j n2 − 1
2 i(cid:17) ⊗ S j n2 − 1
2 i ⊗ S j n2 − 1
2 i ⊗
2
n2−1+ j−2
× H!
× H! ⊥ ⊥
× H! ⊥ Hyp
× H!! ⊥ Hyp
! × h1i ⊥
2
j
2
2
(n−1)
j even
= ⊥
j even h(−1)
= ⊥
j even
= ⊥
h(−1)
n2 − 1 + j − 2
j
! −
n2−1+ j−2
2
j
2
!
× H
⊥ Hyp
[by Proposition 3.5]
2
j
2
=
k−1
2
= Xj even
Xj′=0 n2−1
= n2 +k−2
2
k−1
2
2
n2+ j−3
! × h(−1) n−1
2 i ⊥ Hyp
+ j′ − 1
j′
! × h(−1) n−1
2 i ⊥ Hyp
! × h(−1) n−1
2 i ⊥ Hyp .
[changing the index, j′ = j/2]
[by repeated use of Lemma 3.3]
SYMMETRIC POWERS OF HYPERBOLIC FORMS AND OF TRACE FORMS ON SYMBOL ALGEBRAS7
Now let k be even. Then
S kTS = S k h(−1) n−1
n2 − 1
2
j even
j even
2
2
i+ j=k
2 i ⊥
= ⊥
= ⊥
× H!
2 i(cid:17) ⊗ S j n2 − 1
S i(cid:16)h(−1) n−1
2 i(cid:17) ⊗ S j n2 − 1
S i(cid:16)h(−1) n−1
2 i(cid:17) ⊗ S j n2 − 1
= ⊥
S i(cid:16)h(−1) n−1
2 i ⊗ S j n2 − 1
j even h(−1) (n−1)i
= ⊥
! × h1i ⊥
j even
h1i ⊗
= ⊥
= Xj even
Xj′ =0 n2−1
= n2+k−1
! × h1i ⊥ Hyp .
! × h1i ⊥ Hyp
! × h1i ⊥ Hyp
+ j′ − 1
j′
n2−1+ j−2
n2 + j−3
k
2
=
2
j
2
2
2
j
2
2
k
2
2
2
j odd
× H!
× H! ⊥ ⊥
× H! ⊥ Hyp
× H!! ⊥ Hyp
n2 − 1 + j − 2
j
[by Proposition 4.1]
[by Proposition 2.10]
S i(cid:16)h(−1) n−1
2 i(cid:17) ⊗ S j n2 − 1
2
× H!
[by Proposition 3.1]
[by Lemma 2.9]
! −
n2−1+ j−2
2
j
2
!
× H
[by Proposition 3.5]
⊥ Hyp
[changing the index, j′ = j/2]
[by repeated use of Lemma 3.3]
Proposition 4.3. Let n be even, k odd. We write TS ≃ qS ⊥ m × H where qS ≃
hnih1, a, b, (−1) n
2 abi and m = n2−4
S kTS =
Proof.
2m+2k+4
2 . Then for k > 1,
(cid:16)(2m+k+3)/2
(k−3)/2 (cid:17) × qS ⊥ Hyp,
k−1 (cid:16)(2m+k+3)/2
(k−3)/2 (cid:17) × qS ⊥ Hyp,
k−1
2m+6
if n ≡ 0
if n ≡ 2
(mod 4);
(mod 4).
S kTS = S k (qS ⊥ m × H)
= ⊥
i+ j=k
= ⊥
j even
S iqS ⊗ S j (m × H)
[by Proposition 2.10]
S iqS ⊗ S j (m × H) ⊥ ⊥
j odd
S iqS ⊗ S j (m × H)
8
RONAN FLATLEY
= ⊥
j even
S iqS ⊗ S j (m × H) ⊥ Hyp
! × h1i ⊥
m + j
2 − 1
j
2
j even
j even
= ⊥
S iqS ⊗
S iqS ⊗
= ⊥
= Xj even m + j
= Xj even m + j
= Xj even m + j
= Xj even m + j
= Xj even m + j
= Xj even m + j
Xj′=0 m + j′ − 1
2 − 1
j
2
2 − 1
j
2
2 − 1
j
2
2 − 1
j
2
2 − 1
j
2
2 − 1
j
2
k−1
2
j′
=
i−1
2
(i+1)/2
2 − 1
j
2
m + j
! × h1i
! × S iqS ⊥ Hyp
! ×
ℓ=0 3 + ℓ
⊥
3 + i−1
! ×
3 + i−1
! ×
! 3 + i−1
! 3 + k−1
! 3 + k−1
k−1
2 − j′
2 −
j
2
2 − j′
k−1
2 −
i−1
2
i−1
2
j
2
2
2
[by Proposition 3.1]
1
2
2m + j − 1
j
! − m + j
2 − 1
j
2
[by Proposition 3.5]
!
× H
⊥ Hyp
⊥ Hyp
2
2
ℓ ! × Λi−2ℓqS
! × Λ1qS ⊥ 3 + i−3
! × qS ⊥ 3 + i−3
! × qS ⊥ Xj even m + j
i−3
2
i−3
2
2
2 − 1
j
2
! × qS ⊥ Xj even m + j
! × qS ⊥ Xj even m + j′ − 1
j′
! × Λ3qS
! ×(cid:16)h(−1)n/2i ⊗ qS(cid:17)
! 3 + i−3
2 − 1
j
2
k−3
2 −
2 −
j
2
! 3 + k−3
! 3 + k−3
k−3
2 − j′
i−3
2
2
j
2
2 − j′
⊥ Hyp
! ×(cid:16)h(−1)n/2i ⊗ qS(cid:17) ⊥ Hyp
! ×(cid:16)h(−1)n/2i ⊗ qS(cid:17) ⊥ Hyp
! ×(cid:16)h(−1)n/2i ⊗ qS(cid:17) ⊥ Hyp
⊥ Hyp
[by Proposition 2.11]
⊥ Hyp [since for larger k, Λk is zero]
[changing the index, j′ = j/2]
= m + 3 + k−1
2
k−1
2
Now if n ≡ 0 (mod 4)
S kTS =
=
m + 3 + k−1
2
k−1
2
2m + 2k + 4
k − 1
! × qS ⊥ m + 3 + k−3
2
k−3
2
! ×(cid:16)h(−1)n/2i ⊗ qS(cid:17) ⊥ Hyp .
!
! × qS ⊥ Hyp .
! + m + 3 + k−3
2m+k+3
× qS ⊥ Hyp
k−3
2
2
2
k−3
2
[by Remark 3.4]
[by [4, Equation (3)]]
SYMMETRIC POWERS OF HYPERBOLIC FORMS AND OF TRACE FORMS ON SYMBOL ALGEBRAS9
On the other hand, if n ≡ 2 (mod 4)
S kTS =
=
2
k−1
2
m + 3 + k−1
k − 1 2m+k+3
! − m + 3 + k−3
! × qS ⊥ Hyp .
k−3
2
2m + 6
2
2
k−3
2
× qS ⊥ Hyp
!
[by Remark 3.4]
Proposition 4.4. Let n and k be even. We write TS ≃ qS ⊥ m × H where qS ≃
hnih1, a, b, (−1) n
2 . Then for k > 1,
2 abi and m = n2−4
(cid:16)1 + (m+3+k/2)(m+2+k/2)
(cid:16) (m+3+k/2)(m+2+k/2)
(k/2)(k/2−1)
(k/2)(k/2−1)
S kTS =
Proof.
S kTS = S k (qS ⊥ m × H)
(cid:17)(cid:16)m+1+k/2
− 1(cid:17)(cid:16)m+1+k/2
k/2−2 (cid:17) × h1i ⊥ Hyp,
k/2−2 (cid:17) × h1i ⊥ Hyp,
if n ≡ 0 (mod 4);
if n ≡ 2 (mod 4).
S iqS ⊗ S j (m × H)
[by Proposition 2.10]
S iqS ⊗ S j (m × H) ⊥ ⊥
j odd
S iqS ⊗ S j (m × H)
S iqS ⊗ S j (m × H) ⊥ Hyp
! × h1i ⊥
m + j
2 − 1
j
2
[by Proposition 3.1]
1
2
2m + j − 1
j
! − m + j
2 − 1
j
2
[by Proposition 3.5]
⊥ Hyp
!
× H
= ⊥
i+ j=k
= ⊥
j even
= ⊥
j even
j even
j even
= ⊥
S iqS ⊗
S iqS ⊗
= ⊥
= Xj even m + j
= Xj even m + j
= Xj even m + j
⊥ Hyp
2 − 1
j
2
m + j
! × h1i
! × S iqS ⊥ Hyp
! ×
ℓ=0 3 + ℓ
⊥
3 + i
! ×
2 − 1
j
2
2 − 1
j
2
2 − 1
j
2
i/2
i
2
2
ℓ ! × Λi−2ℓqS
! × Λ0qS ⊥ 3 + i−2
2
i−2
2
⊥ Hyp
[by Proposition 2.11]
! × Λ2qS ⊥ 3 + i−4
2
i−4
2
⊥ Hyp
! × Λ4qS
10
RONAN FLATLEY
[since for larger k, Λk is zero]
! × h1i ⊥ 3 + i−2
2
i−2
2
! × ha, b, abi ⊗ h1, (−1)n/2i ⊥ 3 + i−4
2
i−4
2
! × h(−1)n/2i
= Xj even m + j
2 − 1
j
2
⊥ Hyp
= Xj even m + j
2 − 1
j
2
2
i
2
3 + i
! ×
! 3 + k− j
2
2
2
k− j−4
2
k− j−2
! 3 + k− j−2
2 − 1
j
2
k− j
2
2 − 1
j
2
! × h1i ⊥ Xj even m + j
! 3 + k− j−4
! 3 + k/2 − j′
! × h(−1)n/2i ⊥ Hyp
Xj′ =0 m + j′ − 1
! × h1i ⊥
! 3 + k/2 − 2 − j′
! × h(−1)n/2i ⊥ Hyp
k/2 − 2 − j′
k/2 − j′
k/2−1
j′
2
⊥ Xj even m + j
Xj=0 m + j′ − 1
k/2
j′
=
⊥
k/2−2
Xj′ =0 m + j′ − 1
j′
! × ha, b, abi ⊗ h1, (−1)n/2i
! 3 + k/2 − 1 − j′
k/2 − 1 − j′
! × ha, b, abi ⊗ h1, (−1)n/2i
[changing the index, j′ = j/2]
= m + 3 + k/2
k/2
! × h1i ⊥ m + 2 + k/2
k/2 − 1 ! × ha, b, abi ⊗ h1, (−1)n/2i
⊥ m + 1 + k/2
k/2 − 2 ! × h(−1)n/2i ⊥ Hyp .
[by [4, Equation (3)]]
We now examine separately the cases of n ≡ 0 (mod 4) and n ≡ 2 (mod 4). Suppose
n ≡ 0 (mod 4). In this case we have (−1)n/2 = 1 and so
S kTS = m + 3 + k/2
! + m + 1 + k/2
k/2 − 2 !! × h1i ⊥ 2 m + 2 + k/2
k/2 − 1 ! × ha, b, abi ⊥ Hyp
k/2
(m + 3 + k/2)(m + 2 + k/2)
(k/2)(k/2 − 1)
(m + 3 + k/2)(m + 2 + k/2)
(k/2)(k/2 − 1)
= 1 +
= 1 +
! m + 1 + k/2
! m + 1 + k/2
k/2 − 2 ! × h1i ⊥ 2 m + 2 + k/2
k/2 − 2 ! × h1i ⊥ Hyp .
k/2 − 1 ! × ha, b, abi ⊥ Hyp
Now suppose n ≡ 2 (mod 4). In this case we have (−1)n/2 = −1 and so
S kTS = m + 3 + k/2
= m + 3 + k/2
k/2 − 2 ! × h−1i ⊥ Hyp
! × h1i ⊥ m + 1 + k/2
! − m + 1 + k/2
k/2 − 2 !! × h1i ⊥ Hyp
k/2
k/2
SYMMETRIC POWERS OF HYPERBOLIC FORMS AND OF TRACE FORMS ON SYMBOL ALGEBRAS11
= (m + 3 + k/2)(m + 2 + k/2)
(k/2)(k/2 − 1)
− 1! m + 1 + k/2
k/2 − 2 ! × h1i ⊥ Hyp .
5. Summary of results
We give here a summary of the above results on symmetric powers alongside the
corresponding results on exterior powers computed in [3, Section 5].
Let ϕ ≃ h × H for arbitrary h ∈ N. Then
k
odd
Λkϕ
Hyp
S kϕ
Hyp
even h
k
2! × h(−1) k
2 i ⊥ Hyp h + k
2 − 1
k
2
! × h1i ⊥ Hyp
Let S = (a, b; n, K, ω) be a symbol algebra of degree n as described in Section 4.
Let ϕ = TS = hnih1, a, b, (−1) n
2 abi and m = n2−4
2 . Then
n
odd
odd
k
odd
even
0 (mod 4)
odd
0 (mod 4) even
2 (mod 4)
odd
2 (mod 4) even
Λkϕ
S kϕ
2 i ⊥ Hyp
2 i ⊥ Hyp
4 iqS ⊥ Hyp
n2−1
2
k
2
k
2
k−1
k−1
2
2 (cid:17) × h(−1) k−1
(cid:16)
(cid:16) n2−1
2 (cid:17) × h(−1) k
m + 1
! × h(−1) n(k−1)
n2
2 ! × h1i ⊥ Hyp
! × h(−1) n(k−1)
2 ! × h(−1) k
2 ! × h(−1) k+2
m + 1
n2(cid:17) n2
(cid:16)1 − 2k
n2 − 1(cid:17) n2
(cid:16) 2k
k−1
2
2
k
2
k
4 iqS ⊥ Hyp
2 i ⊥ Hyp, k ≤ n2
2
2 i ⊥ Hyp, k > n2
2
n2 +k−2
2
k
2
k−1
(cid:16)
2
k−1
2
2m+2k+4
2 i ⊥ Hyp
(cid:17) × h(−1) n−1
(cid:16) n2 +k−1
(cid:17) × h1i ⊥ Hyp
(cid:16) 2m+k+3
(cid:16)1 + (m+3+k/2)(m+2+k/2)
k−1 (cid:16) 2m+k+3
(cid:17) × qS ⊥ Hyp
(cid:17)(cid:16) 2m+2+k
(cid:17) × qS ⊥ Hyp
(k/2)(k/2−1)
2
k−3
2
2
k−4
2
2
k−3
2
2m+6
(cid:17) × h1i ⊥ Hyp
(cid:16) (m+3+k/2)(m+2+k/2)
(k/2)(k/2−1)
− 1(cid:17)(cid:16) 2m+2+k
2
k−4
2
(cid:17) × h1i ⊥ Hyp
12
RONAN FLATLEY
References
[1] Bourbaki, N., ´El´ements de Math´ematique, Livre II, Alg`ebre, Hermann, Paris, 1959.
[2] Draxl, P. K., Skew Fields, Cambridge Univ. Press, Cambridge, 1983.
[3] Flatley, R., Trace Forms of Symbol Algebras, Algebra Colloq. 19:spec01 (2012) 1117-1124.
[4] Gould, H. W., Some generalizations of Vandermonde's Convolution, Amer. Math. Monthly 63:2
(1956) 84-91.
[5] Knus, M.-A., Merkurjev, A., Rost, M., Tignol, J.-P., The Book of Involutions, Amer. Math. Soc.
Colloq. Publ., Providence, RI, 1998.
[6] Knutson, D. Rings and the Representation Theory of the Symmetric Group, volume 308 of Lecture
Notes in Mathematics. Springer-Verlag, Berlin, 1973.
[7] Lam, T. Y., Introduction to Quadratic Forms over Fields, Grad. Stud. Math., Providence, RI, 2004.
[8] McGarraghy, S., Symmetric powers of symmetric bilinear forms, Algebra Colloq. 12:1 (2005)
41-57.
[9] Scharlau, W., Quadratic and Hermitian Forms, Springer, Berlin, 1985.
Department of Mathematics and Computer Studies, Mary Immaculate College, Limerick, Ireland
E-mail address: [email protected]
|
1701.02169 | 1 | 1701 | 2017-01-09T13:33:00 | Orthogonal involutions and totally singular quadratic forms in characteristic two | [
"math.RA"
] | We associate to every central simple algebra with involution of orthogonal type in characteristic two a totally singular quadratic form which reflects certain anisotropy properties of the involution. It is shown that this quadratic form can be used to classify totally decomposable algebras with orthogonal involution. Also, using this form, a criterion is obtained for an orthogonal involution on a split algebra to be conjugated to the transpose involution. | math.RA | math |
Orthogonal involutions and totally singular
quadratic forms in characteristic two
A.-H. Nokhodkar
Abstract
We associate to every central simple algebra with involution of ortho-
gonal type in characteristic two a totally singular quadratic form which
reflects certain anisotropy properties of the involution. It is shown that
this quadratic form can be used to classify totally decomposable algebras
with orthogonal involution. Also, using this form, a criterion is obtained
for an orthogonal involution on a split algebra to be conjugated to the
transpose involution.
Mathematics Subject Classification: 16W10, 16K20, 11E39, 11E04.
1
Introduction
The theory of algebras with involution is closely related to the theory of bilinear
forms. Assigning to every symmetric or anti-symmetric regular bilinear form
on a finite-dimensional vector space V its adjoint involution induces a one-
to-one correspondence between the similarity classes of bilinear forms on V
and involutions of the first kind on the endomorphism algebra EndF (V ) (see
[5, p. 1]). Under this correspondence several basic properties of involutions
reflect analogous properties of bilinear forms. For example a symmetric or anti-
symmetric bilinear form is isotropic (resp. anisotropic, hyperbolic) if and only
if its adjoint involution is isotropic (resp. anisotropic, hyperbolic).
Let (A, σ) be a totally decomposable algebra with orthogonal involution
over a field F of characteristic two.
In [3], a bilinear Pfister form Pf(A, σ)
was associated to (A, σ), which determines the isotropy behaviour of (A, σ).
It was shown that for every splitting field K of A, the involution σK on AK
becomes adjoint to the bilinear form Pf(A, σ)K (see [3, (7.5)]). Also, according
to [7, (6.5)] this invariant can be used to determine the conjugacy class of the
involution σ on A.
In this work we associate to every algebra with orthogonal involution (A, σ)
in characteristic two a pair (S(A, σ), qσ) where S(A, σ) is a subalgebra of A and
qσ is a totally singular quadratic form on S(A, σ). Using this pair, we find in
Theorem 3.7 some criteria for σ to be direct (a stronger notion than anisotropy
defined in [2]). As a consequence, several sufficient conditions for anisotropy
of σ are obtained in Corollary 3.10. Further, it is shown in Theorem 3.14
that qσ can classify the transpose involution on a split algebra. For the case
where (A, σ) is totally decomposable, using the quadratic space (S(A, σ), qσ)
we will complement some results of [3] and [7] in Theorem 4.1 and state several
necessary and sufficient conditions for σ to be anisotropic. Finally, we shall
1
see in Theorem 4.6 and Theorem 4.10 that the form qσ can be used to find
a classification of totally decomposable algebras with orthogonal involution in
characteristic two.
2 Preliminaries
In this paper, F denote a field of characteristic two.
Let V be a vector space of finite dimension over F . A bilinear form b :
V × V → F is called isotropic if b(v, v) = 0 for some 0 6= v ∈ V . Otherwise,
b is called anisotropic. We say that b represents α ∈ F if b(v, v) = α for some
nonzero vector v ∈ V . The set of all elements represented by b is denoted by
D(b). If K/F is a field extension, we denote by bK the scalar extension of b
i=1 αixiyi is denoted
i=1h1, αii is called a bilinear Pfister form
and is denoted by hhα1, · · · , αnii. If b is a bilinear Pfister form over F , there
exists a bilinear form b′, called the pure subform of b, such that b ≃ h1i ⊥ b′.
As observed in [1, p. 16] the pure subform of b is uniquely determined, up
to isometry. According to [4, (6.5)], the form b is anisotropic if and only if
dimF 2 Q(b) = 2n, where Q(b) = D(b) ∪ {0}.
to K. For α1, · · · , αn ∈ F the diagonal bilinear form Pn
by hα1, · · · , αni. Also, the form Nn
i=1 αix2
i
by hα1, · · · , αniq.
A quadratic form on V is a map q : V → F such that (i) q(αv) = α2q(v)
for α ∈ F and v ∈ V ; (ii) the map bq : V × V → F given by b(v, w) =
q(v + w) − q(v) − q(w) is an F -bilinear form. A quadratic form q is called
isotropic if q(v) = 0 for some 0 6= v ∈ V and anisotropic otherwise. For a
quadratic space (V, q) over F we use the notation D(q) = {q(v) 0 6= v ∈ V }
and Q(q) = D(q) ∪ {0}. The scalar extension of q to an extension K of F is
denoted by qK. A quadratic form q is called totally singular if bq is the zero map.
is denoted
For α1, · · · , αn ∈ F the totally singular quadratic form Pn
Let A be a central simple algebra over F and let σ be an involution on
A, i.e., an anti-automorphism of A of order two. We say that σ is of the first
kind if it restricts to the identity map on F . For a symmetric bilinear space
(V, b) over F we denote by σb the adjoint involution of EndF (V ) with respect
to b (see [5, p. 1]). An involution σ of the first kind on A is called symplectic
if it becomes adjoint to an alternating bilinear form over a splitting field of
A (i.e., a bilinear form b with D(b) = 0). Otherwise, σ is called orthogonal.
According to [5, (2.6)], σ is orthogonal if and only if 1 /∈ Alt(A, σ), where
Alt(A, σ) = {x − σ(x) x ∈ A}. The discriminant of an orthogonal involution
σ is denoted by disc σ (see [5, (7.1)]). An involution σ is called isotropic if
σ(x)x = 0 for some nonzero element x ∈ A. Otherwise, σ is called anisotropic.
We will frequently use the following result. Recall that if u is a unit in a
central simple algebra A, the inner automorphism of A induced by u is defined
as Int(u)(x) = uxu−1 for x ∈ A.
Proposition 2.1. Let α1, · · · , αn ∈ F × and let u ∈ Mn(F ) be the diagonal ma-
trix diag(α1, · · · , αn). Consider the involution σ = Int(u) ◦ t on Mn(F ), where t
is the transpose involution. If (V, b) is the diagonal bilinear space hα1, · · · , αni,
then (Mn(F ), σ) ≃ (EndF (V ), σb).
Proof. See [5, pp. 13-14].
2
3 The alternator form
For a central simple algebra with orthogonal involution (A, σ) over F we use
the following notation:
S(A, σ) = {x ∈ A σ(x)x ∈ F ⊕ Alt(A, σ)}.
In other words, x ∈ S(A, σ) if and only if there exists a unique element α ∈ F
such that σ(x)x + α ∈ Alt(A, σ). We denote the element α by qσ(x). Hence,
qσ : S(A, σ) → F is a map satisfying
σ(x)x + qσ(x) ∈ Alt(A, σ)
for x ∈ S(A, σ).
Lemma 3.1. Let (A, σ) be a central simple algebra with involution over F . If
x ∈ Alt(A, σ), then σ(y)xy ∈ Alt(A, σ) for every y ∈ A.
Proof. Write x = z − σ(z) for some z ∈ A. Then
σ(y)xy = σ(y)(z − σ(z))y = σ(y)zy − σ(σ(y)zy) ∈ Alt(A, σ).
Lemma 3.2. Let (A, σ) be a central simple algebra with orthogonal involution
over F . Then
(i) S(A, σ) is a (unitary) F -subalgebra of A.
(ii) qσ(λx + y) = λ2qσ(x) + qσ(y) and qσ(xy) = qσ(x)qσ(y) for every λ ∈ F
and x, y ∈ S(A, σ).
Proof. For every λ ∈ F we have σ(λ)λ + λ2 = 0 ∈ Alt(A, σ), so F ⊆ S(A, σ).
Let x, y ∈ S(A, σ) and λ ∈ F . Set α = qσ(x) ∈ F and β = qσ(y) ∈ F . Then
σ(λx+ y)(λx+ y)+ λ2α+ β = λ2(σ(x)x+ α)+ (σ(y)y + β)+ σ(λx)y − σ(σ(λx)y).
Hence, σ(λx + y)(λx + y) + λ2α + β ∈ Alt(A, σ), which implies that λx + y ∈
S(A, σ) and qσ(λx+y) = λ2α+β = λ2qσ(x)+qσ(y). Similarly, using Lemma 3.1
we have
σ(xy)(xy) + αβ = σ(y)σ(x)xy + αβ = σ(y)(σ(x)x + α)y + ασ(y)y + αβ
= σ(y)(σ(x)x + α)y + α(σ(y)y + β) ∈ Alt(A, σ).
Hence, xy ∈ S(A, σ) and qσ(xy) = αβ = qσ(x)qσ(y), proving the result.
Corollary 3.3. Let (A, σ) be a central simple algebra with orthogonal involution
over F . Then qσ is a totally singular quadratic form on S(A, σ).
Proof. The result follows from Lemma 3.2 (ii).
Definition 3.4. Let (A, σ) be a central simple algebras with orthogonal invo-
lution over F . We call S(A, σ) the alternator subalgebra of (A, σ). We also call
the quadratic form qσ the alternator form of (A, σ).
Lemma 3.5. Let (A, σ) and (A′, σ′) be two central simple algebras with ortho-
gonal involution over F . If f : (A, σ) ∼−→ (A′, σ′) is an isomorphism of algebras
with involution, the restriction of f to S(A, σ) defines an isometry (S(A, σ), qσ)
∼−→ (S(A′, σ′), qσ′ ).
3
Proof. For every x ∈ S(A, σ) we have σ(x)x + qσ(x) ∈ Alt(A, σ), which implies
that f (σ(x)x + qσ(x)) ∈ Alt(A′, σ′). Hence, σ′(f (x))f (x) + qσ(x) ∈ Alt(A′, σ′),
i.e., f (x) ∈ S(A′, σ′) and qσ′ (f (x)) = qσ(x).
The following definition was given in [2].
Definition 3.6. An involution σ on a central simple algebra A is called direct
if for every x ∈ A the condition σ(x)x ∈ Alt(A, σ) implies that x = 0.
Theorem 3.7. For an orthogonal involution σ on a central simple F -algebra
A the following conditions are equivalent:
(1) σ is direct.
(2) qσ is anisotropic.
(3) S(A, σ) is a field.
Moreover, if these conditions hold, then x2 ∈ F for all x ∈ S(A, σ).
Proof. The implication (1) ⇒ (2) is evident.
(2) ⇒ (3) : Since S(A, σ) is a subalgebra of A, it suffices to show that (i)
S(A, σ) contains no zero devisor; (ii) x−1 ∈ S(A, σ) for every nonzero element
x ∈ S(A, σ); and (iii) S(A, σ) is commutative.
Suppose that xy = 0 for some 0 6= x ∈ S(A, σ) and y ∈ A. Set α = qσ(x),
so that σ(x)x + α ∈ Alt(A, σ). By Lemma 3.1 we have
ασ(y)y = 0 + ασ(y)y = σ(y)σ(x)xy + ασ(y)y = σ(y)(σ(x)x + α)y ∈ Alt(A, σ).
Since qσ is anisotropic we have α 6= 0. Hence, y ∈ S(A, σ) and qσ(y) = 0.
Again, the anisotropy of qσ implies that y = 0, proving (i).
To prove (ii) let 0 6= x ∈ S(A, σ) and α = qσ(x) 6= 0, so that σ(x)x + α ∈
Alt(A, σ). By Lemma 3.1 we have σ(x−1)(σ(x)x + α)x−1 ∈ Alt(A, σ), i.e.,
ασ(x−1)(x−1) + 1 ∈ Alt(A, σ). It follows that σ(x−1)(x−1) + α−1 ∈ Alt(A, σ),
so x−1 ∈ S(A, σ).
Finally, if x, y ∈ S(A, σ), then Lemma 3.2 implies that
qσ(xy + yx) = qσ(xy) + qσ(yx) = qσ(x)qσ(y) + qσ(y)qσ(x) = 0.
Since qσ is anisotropic we get xy = yx, proving (iii).
(3) ⇒ (1) : Suppose that σ(x)x ∈ Alt(A, σ) for some x ∈ A. Then x ∈
S(A, σ) and qσ(x) = 0. If x 6= 0 then x is a unit, because S(A, σ) is a field. Since
σ(x)x ∈ Alt(A, σ), Lemma 3.1 implies that 1 = σ(x−1)σ(x)xx−1 ∈ Alt(A, σ).
This contradicts the orthogonality of σ.
To prove the last statement of the result, let x ∈ S(A, σ) and set α = qσ(x).
By 3.2 we have qσ(x2) = α2 = qσ(α), so qσ(x2 + α) = 0. As qσ is anisotropic,
we get x2 = α ∈ F .
Lemma 3.8. ([5, (2.26)]) Let (A, σ) be a central simple algebra with orthogonal
involution over F . Then the set Sym(A, σ) generates A as an (associative)
F -algebra.
4
Proof. If degF A > 2, the result follows from [5, (2.26)]. Suppose that degF A =
2. Let B ⊆ A be the subalgebra of A generated by Sym(A, σ). Using the idea of
the proof of [5, (2.26)], it is enough to show that BK = AK for some extension
K of F . Choose an extension K/F with (A, σ)K ≃ (M2(K), t). The conclusion
now easily follows by identifying Sym(A, σ)K with the set of matrices of the
form (cid:0) a b
b c(cid:1), where a, b, c ∈ K.
Proposition 3.9. Let (A, σ) be a central simple algebra with orthogonal invo-
lution over F . If S(A, σ) ⊆ Sym(A, σ), then σ is direct.
Proof. Suppose that σ(x)x ∈ Alt(A, σ) for some x ∈ A. Then x ∈ S(A, σ),
which implies that σ(x) = x. We claim that Sym(A, σ) ⊆ CA(x), where CA(x)
is the centralizer of x in A. Let y ∈ Sym(A, σ). By Lemma 3.1 we have
σ(xy)xy = σ(y)σ(x)xy ∈ Alt(A, σ),
hence xy ∈ S(A, σ) ⊆ Sym(A, σ). It follows that yx = σ(y)σ(x) = σ(xy) = xy,
because the elements x, y and xy are all symmetric. This proves the claim.
By Lemma 3.8, Sym(A, σ) generates A as an F -algebra. Hence, CA(x) = A,
i.e., x ∈ F . Since σ is orthogonal, the condition σ(x)x ∈ Alt(A, σ) implies that
x = 0, so σ is direct.
Since every direct involution is anisotropic, one can use Theorem 3.7 and
Proposition 3.9 to find some sufficient conditions for anisotropy of orthogonal
involutions:
Corollary 3.10. Let (A, σ) be a central simple algebra with orthogonal involu-
tion over F . If any of these conditions is satisfied, then σ is anisotropic: (i) qσ
is anisotropic. (ii) S(A, σ) is a field. (iii) S(A, σ) ⊆ Sym(A, σ).
Lemma 3.11. Let (V, b) be a symmetric non-alternating bilinear space over F .
(i) If x ∈ EndF (V ), then b(x(v), x(v)) = qσb (x)b(v, v) for every v ∈ V .
(ii) If b represents 1, then D(qσb ) ⊆ D(b).
Proof. Let α = qσb (x), so that σb(x)x + α ∈ Alt(EndF (V ), σb). Write σb(x)x =
y − σb(y) − α for some y ∈ EndF (V ). For every v ∈ V we have
b(x(v), x(v)) = b((σb(x)x)(v), v) = b((y − σb(y) − α)(v), v)
= b(y(v), v) − b(σb(y)(v), v) − αb(v, v)
= b(y(v), v) − b(v, y(v)) − αb(v, v) = αb(v, v).
This proves the first part. The second part follows by applying (i) to a vector
v ∈ V with b(v, v) = 1.
Remark 3.12. The converse of Proposition 3.9 does not hold in general. To
construct a counter-example, let hhα, βii be an anisotropic bilinear Pfister form
over F and let u be the diagonal matrix diag(1, α, β, αβ + 1). Consider the
involution σ = Int(u) ◦ t on M4(F ). By Proposition 2.1 we have (M4(F ), σ) ≃
(EndF (V ), σb), where (V, b) is the diagonal bilinear space h1, α, β, αβ +1i. Since
5
b is anisotropic, the form qσ is also anisotropic by Lemma 3.11 (ii). Hence, σ
is direct by Theorem 3.7. Let
0
0 (αβ)−1 β−1
0
1
0 1+(αβ)−1
0
0
0 α(1+αβ)−1
(1+αβ)−1
0
0
0
x =
∈ M4(F ).
Computations shows that (σ(x)x + β−1I4)u ∈ Alt(M4(F ), t), where I4 is the
4 × 4 identity matrix. By [5, (2.7)] we have σ(x)x + β−1I4 ∈ Alt(M4(F ), σ), so
x ∈ S(M4(F ), σ). On the other hand
xu = 0 β−1
0 α+β−1 0 0! /∈ Sym(M4(F ), t),
1 0
0 α
0 1
0
1
0
0
which implies that x /∈ Sym(M4(F ), σ), thanks to [5, (2.7)].
x ∈ S(M4(F ), σ) \ Sym(M4(F ), σ).
It follows that
Lemma 3.13. Let t be the transpose involution on Mn(F ). Then
(i) Q(qt) = F 2.
(ii) qt ≃ h1iq ⊥ (n2 − n) × h0iq.
Proof. (i) Clearly, we have F 2 ⊆ Q(qt). The converse inclusion follows from
Lemma 3.11 (ii) and the fact that the transpose involution is the adjoint invo-
lution of the bilinear form n × h1i (see Proposition 2.1).
(ii) Using the first part and the isometry h1, 1iq ≃ h1, 0iq we have qt ≃
h1iq ⊥ k × h0iq for some non-negative integer k. We claim that k = n2 − n.
Let W = {x ∈ S(Mn(F ), t) qt(x) = 0}. Then W is a k-dimensional vector
space over F . A matrix x = (xij ) ∈ Mn(F ) belongs to W if and only if
xtx ∈ Alt(Mn(F ), t). Note that Alt(Mn(F ), t) is the set of symmetric matrices
ij = 0 for
with zero diagonal. Hence, xtx ∈ Alt(Mn(F ), t) if and only if Pn
j = 1, · · · , n, or equivalently
i=1 x2
i=1 xij = 0,
for j = 1, · · · , n.
(1)
Hence, W is the set of answers of the homogeneous system of linear equations
(1) with n2 unknowns xij , i, j = 1, · · · , n. This set is a vector space of dimension
n2 − n over F , hence k = n2 − n.
Pn
Theorem 3.14. Let (A, σ) be a central simple algebra of degree n with ortho-
gonal involution over F . If A splits, then (A, σ) ≃ (Mn(F ), t) if and only if
qσ ≃ h1iq ⊥ (n2 − n) × h0iq.
Proof. The 'only if' implication follows from Lemma 3.5 and Lemma 3.13.
To prove the converse, we may identify (A, σ) = (EndF (V ), σb), where (V, b)
is a symmetric non-alternating bilinear space over F . By [6, (2.1)] there exist
unique integers m, k and a1, · · · , am ∈ F such that
b ≃ ban ⊥ M(a1) ⊥ · · · ⊥ M(am) ⊥ k × H,
(2)
where ban is an anisotropic bilinear form, M(ai) = hai, aii and H is the hyper-
bolic plane. Let
U = {x ∈ S(EndF (V ), σb) qσb (x) = 0} and W = {v ∈ V b(v, v) = 0}.
6
Then U and W are two vector spaces over F . We have dimF U = n2−n, because
qσ ≃ h1iq ⊥ (n2 − n) × h0iq. Also, by [6, (2.1)] we have dimF W = 2k + m.
On the other hand Lemma 3.11 (i) implies that x(V ) ⊆ W for every x ∈ U ,
i.e., U ⊆ HomF (V, W ). By dimension count we get dimF W > n − 1. However,
we have W 6= V , because b is not alternating. Hence, dimF W = n − 1, i.e.,
2k + m = n − 1. By (2) we have n = dimF ban + 2k + 2m, which implies
that dimF ban + m = 1.
It follows that either ban is trivial and m = 1 or
dimF ban = 1 and m = 0. Hence, either b ≃ hα, αi ⊥ k × H or b ≃ hαi ⊥ k × H
for some α ∈ F ×. Using the isometry hαi ⊥ H ≃ hα, α, αi in [4, (1.16)], we get
b ≃ n × hαi. Hence, (A, σ) ≃ (Mn(F ), t) by Proposition 2.1.
The next example shows that for n > 2 there exists a central simple algebra
with orthogonal involution (A, σ) of degree n over F such that qσ = h1iq. Note
that for every algebra with involution (A, σ) over F we have F ⊆ S(A, σ).
Hence, the form h1iq is always a subform of the alternator form qσ.
Example 3.15. For α1, · · · , αn ∈ F × let u be the diagonal n × n matrix
diag(α1, · · · , αn) and consider the involution σ = Int(u) ◦ t on Mn(F ). Let
b = hα1, · · · , αni. By Proposition 2.1 we have (M4(F ), σ) ≃ (EndF (V ), σb),
where V is an underlying vector space of b. Let x = (xij ) ∈ Mn(F ). We first
claim that x ∈ S(Mn(F ), σ) with qσ(x) = λ ∈ F if and only if
α−1
1 x2
1i + · · · + α−1
n x2
ni = α−1
i λ for i = 1, · · · , n.
(3)
By definition x ∈ S(Mn(F ), σ) (with qσ(x) = λ) if and only if
y := σ(x)x + λIn ∈ Alt(Mn(F ), σ).
By [5, (2.4)] we have Alt(Mn(F ), σ) = u · Alt(Mn(F ), t), so y ∈ Alt(Mn(F ), σ)
if and only if u−1y ∈ Alt(Mn(F ), t). Since y ∈ Sym(Mn(F ), σ) we have u−1y ∈
Sym(Mn(F ), t) by [5, (2.4)]. Hence, x ∈ S(Mn(F ), σ) if and only if u−1y has
zero diagonal. Write u−1y = (yij ) for some yij ∈ F , 1 6 i, j 6 n. As y =
uxtu−1x + λIn, computation shows that
n x2
i λ for i = 1, · · · , n.
1i + · · · + α−1
ni + α−1
yii = α−1
1 x2
We have yii = 0 if and only if (3) is satisfied, as claimed.
Now, if n > 2 and hhα1, · · · , αnii is an anisotropic bilinear Pfister form, then
nn and xij = 0 for i 6= j. It follows that
(3) implies that λ = x2
λ ∈ F 2 and x is a scalar (matrix). Hence, S(A, σ) = F and qσ = h1iq.
11 = · · · = x2
Remark 3.16. Example 3.15 shows that the alternator form does not necessar-
ily classify orthogonal involutions on a given central simple algebra. Also, using
Example 3.15 one can show that the inclusion Q(qσb ) ⊆ Q(b) in Lemma 3.11
Indeed, with the notation of Example 3.15 set b′ = α1 · b.
could be strict.
Then b′ represents 1 and (EndF (V ), σb) ≃ (EndF (V ), σb′ ) (see [5, p. 1]). By
Lemma 3.5 we have Q(qσb′ ) = Q(qσb ) = F 2. Since b is anisotropic we have
α2 /∈ F 2, hence α2 ∈ Q(b′) \ Q(qσb′ ).
4 Applications to totally decomposable involu-
tions
A quaternion algebra over F is a central simple algebra of degree 2. An algebra
with involution (A, σ) over F is called totally decomposable if it decomposes
7
Nn
into tensor products of quaternion F -algebras with involution. Let (A, σ) ≃
i=1(Qi, σi) be a totally decomposable algebra with orthogonal involution over
F . By [5, (2.23)] every σi is an orthogonal involution. Write disc σi = αiF ×2 ∈
F ×/F ×2 for some αi ∈ F ×, i = 1, · · · , n. As in [3] we denote the bilinear Pfister
form hhα1, · · · , αnii by Pf(A, σ). By [3, (7.5)], Pf(A, σ) is independent of the
decomposition of (A, σ). Also, as observed in [7] there exists a unique, up to
isomorphism, F -algebra Φ(A, σ) ⊆ F ⊕ Alt(A, σ) of dimension 2n satisfying: (i)
x2 ∈ F for x ∈ Φ(A, σ); (ii) the centralizer of Φ(A, σ) in A coincides with Φ(A, σ)
itself; (iii) Φ(A, σ) is generated, as an F -algebra by n elements. According to
[7, (5.5)] the algebra Φ(A, σ) may be considered as an underlying vector space
of Pf(A, σ) such that
Pf(A, σ)(x, x) = x2
for x ∈ Φ(A, σ).
Finally, let x ∈ Φ(A, σ) and set α = x2 ∈ F . Then σ(x)x + α = x2 + α = 0 ∈
Alt(A, σ). Hence, Φ(A, σ) ⊆ S(A, σ) and
qσ(x) = Pf(A, σ)(x, x) = x2
for x ∈ Φ(A, σ).
The following theorem complements some results in [3] and [7].
Theorem 4.1. For a totally decomposable algebra with orthogonal involution
(A, σ) over F the following conditions are equivalent: (1) σ is anisotropic. (2)
σ is direct. (3) Pf(A, σ) is anisotropic. (4) qσ is anisotropic. (5) Φ(A, σ) is a
field. (6) S(A, σ) is a field. (7) Φ(A, σ) = S(A, σ). (8) S(A, σ) ⊆ Sym(A, σ).
In particular, if these conditions hold, the algebra Φ(A, σ) is uniquely deter-
mined.
Proof. The equivalences (1) ⇔ (2) ⇔ (3) can be found in [3] (see [3, (6.1), (6.2)
and (7.5)]). The equivalences (2) ⇔ (4) ⇔ (6) are proved in Theorem 3.7 and
(1) ⇔ (5) follows from [7, (6.6)] and [3, (6.2)]. Suppose that is S(A, σ) is a
field. Since Φ(A, σ) ⊆ S(A, σ) and Φ(A, σ) is maximal commutative, we get
Φ(A, σ) = S(A, σ). This proves (6) ⇒ (7). The implication (7) ⇒ (8) is evident
and (8) ⇒ (2) follows from Proposition 3.9.
Lemma 4.2. Let K/F be a separable quadratic extension and let b be a sym-
metric bilinear form over F . Then D(bK) ∩ F = D(b).
Proof. Clearly, we have D(b) ⊆ D(bK) ∩ F . Suppose that α ∈ D(bK) ∩ F . Let
V be an underlying vector space of b. Write K = F (η) for some η ∈ K with
δ := η2 + η ∈ F . Then α = bK(u ⊗ 1 + v ⊗ η, u ⊗ 1 + v ⊗ η) for some u, v ∈ V .
Thus
α = (b(u, u) + δb(v, v)) + η(b(u, v) + b(v, u) + b(v, v))
= (b(u, u) + δb(v, v)) + ηb(v, v).
Since α ∈ F we get b(v, v) = 0. Hence, α = b(u, u) ∈ D(b), proving the
result.
Proposition 4.3. If (A, σ) is a totally decomposable algebra with orthogonal
involution over F , then D(qσ) = D(Pf(A, σ)).
8
Nn
Proof. Let b = Pf(A, σ). Since b(x, x) = x2 = qσ(x) for every x ∈ Φ(A, σ) ⊆
S(A, σ) we have D(b) ⊆ D(qσ). To prove the converse inclusion, let (A, σ) ≃
i=1(Qi, σi) be a decomposition of (A, σ) into quaternion algebras with invo-
lution. For i = 0, · · · , n, define a field Ki inductively as follows: set K0 = F
and suppose that Ki is defined. If Ki splits A, set Ki+1 = Ki. Otherwise, let
r be the minimal number for which Ki does not split Qr. Then Qr ⊗F Ki is a
division algebra over Ki. Let Ki+1 be a maximal separable subfield of Qr ⊗ Ki.
Note that for i = 0, · · · , n − 1, Ki may be identified with a subfield of Ki+1.
Also, either Ki+1 = Ki or Ki+1/Ki is a separable quadratic extension. Set
L := Kn, so that AL splits.
By [3, (7.5)], we may identify (A, σ)L = (EndL(V ), σbL ). Using Lemma 3.11
we get D(qσL ) ⊆ D(bL). If x ∈ S(A, σ) and α = qσ(x), then x ⊗ 1 ∈ S((A, σ)L)
and qσL (x ⊗ 1) = α ⊗ 1. Hence, by identifying F ⊗ F ⊆ A ⊗ F with F we have
D(qσ) ⊆ D(qσL ). It follows that D(qσ) ⊆ D(bL). By 4.2 and induction on
n we have D(bL) ∩ F = D(b) ⊆ D(qσL ). Hence, D(qσ) ⊆ D(b), proving the
result.
Lemma 4.4. Let b and b′ be two isotropic bilinear n-fold Pfister forms over F .
Then b ≃ b′ if and only if Q(b) = Q(b′).
Proof. The 'only if'
implication is evident. To prove the converse, choose
positive integers r and r′ and anisotropic bilinear Pfister forms c and c′ over
F such that b ≃ hh1iir ⊗ c and b′ ≃ hh1iir′
⊗ c′, where hh1iis is the s-fold
Pfister form hh1, · · · , 1ii (see [1, p. 909]). Since c and c′ are anisotropic we
have dimF 2 Q(c) = 2n−r and dimF 2 Q(c′) = 2n−r′
. As Q(b) = Q(c) and
Q(b′) = Q(c′), the assumption implies that Q(c) = Q(c′), hence r = r′. The
conclusion now follows from [1, (A.8)].
Corollary 4.5. Let (A, σ) and (A′, σ′) be totally decomposable algebras with
isotropic orthogonal involution over F . If qσ ≃ qσ′ then Pf(A, σ) ≃ Pf(A′, σ′).
Proof. The result follows from Proposition 4.3 and Lemma 4.4.
Theorem 4.6. Let (A, σ) and (A′, σ′) be totally decomposable algebras with
isotropic orthogonal involution over F . Then (A, σ) ≃ (A′, σ′) if and only if
A ≃ A′ and qσ ≃ qσ′ .
Proof. The 'only if' implication follows from Lemma 3.5 and the converse follows
from Corollary 4.5 and [7, (6.5)].
Notation 4.7. Let (A, σ) be a totally decomposable algebra with anisotropic
orthogonal involution over F . By Theorem 4.1 we have S(A, σ) = Φ(A, σ) ⊆
F ⊕ Alt(A, σ). We denote the set S(A, σ) ∩ Alt(A, σ) by S ′(A, σ). We also
denote by q′
σ the restriction of qσ to S ′(A, σ).
Note that we have S(A, σ) = F ⊕ S ′(A, σ) and qσ ≃ h1iq ⊥ q′
σ, because
qσ(1) = 1. Also, as observed in [7, p. 223] one has an orthogonal decomposition
σ(x) = x2 =
Φ(A, σ) = F ⊥ S ′(A, σ) with respect to Pf(A, σ). It follows that q′
b′(x, x) for x ∈ S ′(A, σ), where b′ is the pure subform of Pf(A, σ). Hence,
we have the following result (recall that for a symmetric bilinear space (V, b)
over F , there exists a unique totally singular quadratic form ϕb on V given by
ϕb(x) = b(x, x)).
9
Lemma 4.8. Let (A, σ) be a totally decomposable algebra with anisotropic or-
thogonal involution over F and let b = Pf(A, σ). Then qσ = ϕb and q′
σ = ϕb′ ,
where b′ is the pure subform of b.
Let (A, σ) be a totally decomposable algebra of degree 2n with orthogonal
involution over F . Then there exists a set {v1, · · · , vn} consisting of units such
that Φ(A, σ) ≃ F [v1, · · · , vn] and vi1 · · · vis ∈ Alt(A, σ) for every 1 6 s 6 n
and 1 6 i1 < · · · < is 6 n (see [7, (5.1)] for more details). As in [7] we
call {v1, · · · , vn} a set of alternating generators of Φ(A, σ). According to [7,
(5.3) and (5.5)], if {v1, · · · , vn} is a set of alternating generators of Φ(A, σ) and
αi = v2
i ∈ F × for i = 1, · · · , n, then Pf(A, σ) ≃ hhα1, · · · , αnii.
Proposition 4.9. Let (A, σ) and (A′, σ′) be totally decomposable algebras with
anisotropic orthogonal involution over F . Then q′
σ′ if and only if Pf(A, σ)
≃ Pf(A′, σ′).
σ ≃ q′
Proof. The 'if' implication follows from Lemma 4.8. To prove the converse, let
degF A = 2n and let {x1, · · · , xn} be a set of alternating generators of Φ(A, σ).
Set αi = x2
i ∈ F ×, so that Pf(A, σ) ≃ hhα1, · · · , αnii. By dimension count the
set
{xi1 · · · xis 1 6 s 6 n and 1 6 i1 < · · · < is 6 n},
is a basis of S ′(A, σ) over F . Let f : (S ′(A, σ), q′
isometry and set x′
i = f (xi) ∈ S ′(A′, σ′), i = 1, · · · , n. Then
σ) ∼−→ (S ′(A′, σ′), q′
σ′ ) be an
x′2
i = q′
σ′ (x′
i) = q′
σ(xi) = αi
for i = 1, · · · , n.
(4)
We claim that f (xi1 · · · xis ) = x′
Since S(A′, σ′) is an F -algebra we have x′
i1 · · · x′
is for 1 6 s 6 n and 1 6 i1 < · · · < is 6 n.
i1 · · · x′
is ∈ S(A′, σ′). We also have
qσ′ (f (xi1 · · · xis ) + x′
i1 · · · x′
is ) = qσ(xi1 · · · xis ) + qσ′ (x′
is )
i1 · · · x′
is )2.
= (xi1 · · · xis )2 + (x′
i1 · · · x′
(5)
Since S(A, σ) = Φ(A, σ), S(A, σ) is commutative. Similarly, S(A′, σ′) is also
commutative. Hence, using (5) and (4) we get
qσ′ (f (xi1 · · · xis ) + x′
i1 · · · x′
is ) = x2
i1 · · · x2
is + x′2
i1 · · · x′2
is = 0.
As qσ′ is anisotropic we get f (xi1 · · · xis ) = x′
is , proving the claim. In
is ∈ S ′(A′, σ′) ⊆ Alt(A′, σ′) for 1 6 s 6 n and
particular, we have x′
1 6 i1 < · · · < is 6 n. Hence, {x′
n} is a set of alternating generators
of Φ(A′, σ′). The relation (4) now implies that Pf(A′, σ′) ≃ hhα1, · · · , αnii ≃
Pf(A, σ).
1, · · · , x′
i1 · · · x′
i1 · · · x′
Using Lemma 3.5, [7, (6.5)] and Proposition 4.9 we have the following ana-
logue of Theorem 4.6.
Theorem 4.10. Let (A, σ) and (A′, σ′) be totally decomposable algebras with
anisotropic orthogonal involution over F . Then (A, σ) ≃ (A′, σ′) if and only if
A ≃ A′ and q′
σ ≃ q′
σ′ .
10
References
[1] J. Arason, R. Baeza, Relations in I n and I nW
q in characteristic 2. J. Algebra 314
(2007), no. 2, 895 -- 911.
[2] A. Dolphin, Decomposition of algebras with involution in characteristic 2. J. Pure Appl.
Algebra 217 (2013), no. 9, 1620 -- 1633.
[3] A. Dolphin, Orthogonal Pfister involutions in characteristic two. J. Pure Appl. Algebra
218 (2014), no. 10, 1900 -- 1915.
[4] R. Elman, N. Karpenko, A. Merkurjev, The algebraic and geometric theory of quadratic
forms. American Mathematical Society Colloquium Publications, 56. American Math-
ematical Society, Providence, RI, 2008.
[5] M.-A. Knus, A. S. Merkurjev, M. Rost, J.-P. Tignol, The book of involutions. American
Mathematical Society Colloquium Publications, 44. American Mathematical Society,
Providence, RI, 1998.
[6] A. Laghribi, P. Mammone. Hyper-isotropy of bilinear forms in characteristic 2. Con-
temporary Mathematics, 493 (2009), 249-269.
[7] M. G. Mahmoudi, A.-H. Nokhodkar, On totally decomposable algebras with involution
in characteristic two. J. Algebra 451 (2016), 208 -- 231.
A.-H. Nokhodkar, [email protected],
Department of Pure Mathematics, Faculty of Science, University of Kashan, P. O.
Box 87317-51167, Kashan, Iran.
11
|
1705.04501 | 3 | 1705 | 2018-11-05T18:49:09 | Uniqueness of the von Neumann continuous factor | [
"math.RA",
"math.OA"
] | For a division ring $D$, denote by $\mathcal M_D$ the $D$-ring obtained as the completion of the direct limit $\varinjlim_n M_{2^n}(D)$ with respect to the metric induced by its unique rank function. We prove that, for any ultramatricial $D$-ring $\mathcal B$ and any non-discrete extremal pseudo-rank function $N$ on $\mathcal B$, there is an isomorphism of $D$-rings $\overline{\mathcal B} \cong \mathcal M_D$, where $\overline{\mathcal B}$ stands for the completion of $\mathcal B$ with respect to the pseudo-metric induced by $N$. This generalizes a result of von Neumann. We also show a corresponding uniqueness result for $*$-algebras over fields $F$ with positive definite involution, where the algebra $\mathcal M_F$ is endowed with its natural involution coming from the $*$-transpose involution on each of the factors $M_{2^n}(F)$. | math.RA | math |
UNIQUENESS OF THE VON NEUMANN CONTINUOUS FACTOR
PERE ARA AND JOAN CLARAMUNT
Abstract. For a division ring D, denote by MD the D-ring obtained as the completion of
the direct limit lim−→n
M2n(D) with respect to the metric induced by its unique rank function.
We prove that, for any ultramatricial D-ring B and any non-discrete extremal pseudo-rank
function N on B, there is an isomorphism of D-rings B ∼= MD, where B stands for the
completion of B with respect to the pseudo-metric induced by N . This generalizes a result of
von Neumann. We also show a corresponding uniqueness result for ∗-algebras over fields F
with positive definite involution, where the algebra MF is endowed with its natural involution
coming from the ∗-transpose involution on each of the factors M2n(F ).
1. Introduction
Murray and von Neumann showed in [13, Theorem XII] a uniqueness result for approx-
imately finite von Neuman algebra factors of type II1. This unique factor R is called the
hyperfinite II1-factor and plays a very important role in the theory of von Neumann algebras.
It was shown later by Alain Connes [6] that the factor R is characterized (among II1-factors)
by various other properties, such as injectivity (in the operator space sense), semidiscrete-
ness or Property P. It is in particular known (e.g.
[18, Theorem 3.8.2]) that, for an infinite
countable discrete group G whose non-trivial conjugacy classes are all infinite, the group von
Neumann algebra N (G) is isomorphic to R if and only if G is an amenable group. (The
groups with the above property on the conjugacy classes are termed ICC-groups.)
For a field K, the direct limit lim−→n
Von Neumann also considered a purely algebraic analogue of the above situation, as follows.
M2n(K) with respect to the block diagonal embeddings
0 x(cid:19) is a (von Neumann) regular ring, which admits a unique rank function (see
x 7→ (cid:18)x 0
M2n(K), denoted here
below for the definition of rank function). The completion of lim−→n
by MK, with respect to the induced rank metric, is a complete regular ring with a unique
rank function, which is a continuous factor, i.e., it is a right and left self-injective ring and
the set of values of the rank function fills the unit interval [0, 1]. There are recent evidences
[7, 8, 9] that the factor MK could play a role in algebra which is similar to the role played
by the unique hyperfinite factor R in the theory of operator algebras. In particular, Elek
has shown in [8] that, if Γ = Z2 ≀ Z is the lamplighter group, then the continuous factor c(Γ)
obtained by taking the rank completion of the ∗-regular closure of C[Γ] in the ∗-algebra U(Γ)
of unbounded operators affiliated to N (Γ), is isomorphic to MC.
Date: November 6, 2018.
2000 Mathematics Subject Classification. Primary 16E50; Secondary 16D70.
Key words and phrases. rank function, von Neumann regular ring, completion, factor, ultramatricial.
Both authors were partially supported by DGI-MINECO-FEDER through the grants MTM2014-53644-P
and MTM2017-83487-P.
1
2
PERE ARA AND JOAN CLARAMUNT
This raises the question of what uniqueness properties the von Neumann factor MK has,
and whether we can formulate similar characterizations to those in the seminal paper by
Connes [6]. As von Neumann had already shown (and was published later by Halperin [16]),
MK is isomorphic to the factor obtained from any factor sequence (pi), that is,
MK
∼= lim−→ Mpi(K),
where (pi) is a sequence of positive integers converging to ∞ and such that pi divides pi+1
for all i. Here the completion is taken with respect to the unique rank function on the direct
limit.
The purpose of this paper is to obtain stronger uniqueness properties of the factor MK.
Specifically, we show that if B is an ultramatricial K-algebra and N is a non-discrete extremal
pseudo-rank function on B, then the completion of B with respect to N is necessarily isomor-
phic to MK. We also derive a characterization of the factor MK by a local approximation
property (see Theorem 2.2). This will be used in [2] to generalize Elek's result to arbitrary
fields K of characteristic 6= 2, using a concrete approximation of the group algebra K[Γ] by
matricial algebras. It is also worth to mention that, as a consequence of our result and [14,
Theorem 2.8], one obtains that the center of an algebra Q satisfying properties (ii) or (iii) in
Theorem 2.2 is the base field K.
Gabor Elek and Andrei Jaikin-Zapirain have recently raised the question of whether, for
any subfield k of C closed under complex conjugation, and any countable amenable ICC-
group G, the rank completion c(k[G]) of the ∗-regular closure of k[G] in U(G) is either of the
form Mn(D) or of the form MD := D ⊗k Mk, where D is a division ring with center k. In
view of this question, it is natural to obtain uniqueness results in the slightly more general
setting of D-rings over a division ring D, and also in the setting of rings with involution,
since in the above situation, the algebras have a natural involution which is essential even to
define the corresponding completions. We address these questions in the final two sections.
2. von Neumann's continuous factor
A ring R is said to be (von Neumann) regular in case for each x ∈ R there exists y ∈ R
such that x = xyx. We refer the reader to [15] for the general theory of regular rings.
We recall the definition of a pseudo-rank function on a general unital ring.
Definition 2.1. A pseudo-rank function on a unital ring R is a function N : R → [0, 1]
satisfying the following properties:
(1) N(1) = 1.
(2) N(a + b) ≤ N(a) + N(b) for all a, b ∈ R.
(3) N(ab) ≤ N(a), N(b) for all a, b ∈ R.
(4) If e, f ∈ R are orthogonal idempotents, then N(e + f ) = N(e) + N(f ).
A rank function on R is a pseudo-rank function N such that N(x) = 0 implies x = 0.
VON NEUMANN FACTOR
3
Any pseudo-rank function N on a ring R induces a pseudo-metric by d(x, y) = N(x − y) for
x, y ∈ R. If, in addition, R is regular, then the completion of R with respect to this pseudo-
metric is again a regular ring R, and R is complete with respect to the unique extension N
of N to a rank function on R ([15, Theorem 19.6]). The space of pseudo-rank functions P(R)
on a regular ring R is a Choquet simplex ([15, Theorem 17.5]), and the completion R of R
with respect to N ∈ P(R) is a simple ring if and only if N is an extreme point in P(R) ([15,
Theorem 19.14]).
For a field K, a matricial K-algebra is a K-algebra which is isomorphic to an algebra of
the form
Mn(1)(K) × Mn(2)(K) × · · · × Mn(k)(K)
for some positive integers n(1), n(2), . . . , n(k). An ultramatricial K-algebra is an algebra
An of a sequence of matricial K-algebras An and
which is isomorphic to a direct limit lim−→n
unital algebra homomorphisms ϕn : An → An+1, see [15, Chapter 15].
Let K be a field. Write M = MK for the rank completion of the direct limit lim−→n
M2n(K)
with respect to its unique rank function. Von Neumann proved a uniqueness property for
M. We are going to extend it to ultramatricial algebras. The proof follows the steps in the
paper by Halperin [16] (based on von Neumann's proof), but the proof is considerably more
involved. Indeed, we will obtain a uniqueness result for the class of continuous factors which
have a local matricial structure.
By a continuous factor we understand a simple, regular, (right and left) self-injective ring
Q of type IIf (see [15, Chapter 10] for the definition of the types and for the structure
theory of regular self-injective rings). It follows from [15, Corollary 21.14] that Q admits a
unique rank function, denoted here by NQ, and that Q is complete in the NQ-metric. Also, it
follows easily from the structure theory of regular self-injective rings ([15, Chapter 10]) that
NQ(Q) = [0, 1].
The adjective "continuous" used here refers to the fact that NQ takes a "continuous" set
of values, in contrast with the algebra of finite matrices, where the rank function takes only
a finite number of values. Note however that any regular self-injective ring R is a right (left)
continuous regular ring, in the technical sense that the lattice of principal right (left) ideals
is continuous, see [15]. The latter property will play no explicit role in the present paper.
We will show the following result:
Theorem 2.2. Let Q be a continuous factor, and assume there exists a dense subalgebra
(with respect to the NQ-metric topology) Q0 ⊆ Q of countable K-dimension. The following
are equivalent:
(i) Q ∼= MK.
(ii) Q ∼= B for a certain ultramatricial K-algebra B, where the completion of B is taken
with respect to the metric induced by an extremal pseudo-rank function on B.
(iii) For every ε > 0 and x1, ..., xn ∈ Q, there exists a matricial K-subalgebra A of Q and
elements y1, ..., yn ∈ A such that
NQ(xi − yi) < ε
(i = 1, . . . , k).
4
PERE ARA AND JOAN CLARAMUNT
(i) =⇒ (ii) =⇒ (iii) is clear. For the proof of the implication (iii) =⇒ (i) we will use a
method similar to the one used in [16]. However the technical complications are much higher
here.
We first prove a lemma, and show the implication (iii) =⇒ (i) assuming that the hypotheses
of the lemma are satisfied. After this is done, we will show how to construct (using (iii)) the
sequences, algebras, and homomorphisms appearing in this lemma.
Given a factor sequence (pi), the natural block-diagonal unital embeddings Mpi(K) →
Mpi+1(K) will be denoted by γi+1,i. If j > i, the map γj,i : Mpi(K) → Mpj (K) will denote the
composition γj,i = γj,j−1◦· · ·◦γi+1,i, and the map γ∞,i : Mpi(K) → lim−→n
Mpn(K) will stand for
the canonical map into the direct limit. By [16], there is an isomorphism MK
Mpn(K),
where the completion is taken with respect to the unique rank function on the direct limit.
We henceforth will identify M = MK with the algebra lim−→n
Notation 2.3. Let (X, d) be a metric space, Y a subset of X and ε > 0. For A ⊆ X, we
write A ⊆ε Y in case each element of A can be approximated by an element of Y up to ε,
that is, for each a ∈ A there exists y ∈ Y such that d(a, y) < ε.
∼= lim−→n
Mpn(K).
Lemma 2.4. Let Q be a continuous factor with unique rank function NQ. Assume there
exists a dense subalgebra Q0 of Q of countable dimension, and let {xn}n be a K-basis of Q0.
Let θ be a real number such that 0 < θ < 1. Assume further that we have constructed
two strictly increasing sequences (qi) and (pi) of natural numbers such that pi divides pi+1,
satisfying
1 >
and
p1
q1
> · · · >
>
pi
qi
pi+1
qi+1
> · · · > θ,
lim
i→∞
= θ,
pi
qi
pi+1
qi+1
− θ <
1
2(cid:16)pi
qi
− θ(cid:17),
for i ≥ 0, where we set p0 = q0 = 1. Moreover, suppose that there exists a sequence of
positive numbers εi < δi := pi
2δi−1 < 2−i for all i ≥ 1),
qi
and matricial subalgebras Ai ⊆ Q together with algebra homomorphisms ρi : Mpi(K) → Q
satisfying the following properties:
− θ for each i (in which case δi < 1
(i) NQ(ρi(1)) = pi
qi
(ii) For each i and each x ∈ ρi(1)Aiρi(1), there exists y ∈ Mpi+1(K) such that
for all i,
(iii) For each z ∈ Mpi(K), we have
NQ(x − ρi+1(y)) < δi ,
NQ(ρi(z) − ρi+1(γi+1,i(z))) < δi ,
(iv) span{x1, ..., xi} ⊆εi Ai (that is, we can approximate every element of span{x1, ..., xi}
by an element of Ai up to εi in rank).
Then there exists an isomorphism ψ : M → eQe, with e ∈ Q an idempotent such that
NQ(e) = θ.
Proof. For a given positive integer i, and for z ∈ Mpi(K), we consider the sequence in Q
VON NEUMANN FACTOR
5
{ρj(γj,i(z))}j≥i.
It is a simple computation to show, using (iii), that for h > j ≥ i we have
(2.1)
NQ(ρh(γh,i(z)) − ρj(γj,i(z))) < δj + · · · + δh−1 .
As a consequence, we obtain
NQ(ρh(γh,i(z)) − ρj(γj,i(z))) < 2−j + · · · + 2−h+1 < 2−j+1,
and the sequence is Cauchy; so we can define ψi : Mpi(K) → Q by
ψi(z) = lim
j
ρj(γj,i(z)) ∈ Q.
Note that ψi+1(γi+1,i(z)) = ψi(z), so the maps {ψi}i give a well-defined algebra homomor-
phism ψ : lim−→i
Mpi(K) → Q, defined by ψ(γ∞,i(z)) = ψi(z) for z ∈ Mpi(K).
Observe that NQ(ρi(z)) = pi
qi
Npi(z) = pi
qi
NM(γ∞,i(z)) for z ∈ Mpi(K), where Npi denotes
the unique rank function on Mpi(K). Therefore
NQ(ψ(γ∞,i(z))) = lim
j
NQ(ρj(γj,i(z))) = lim
j
pj
qj
NM(γ∞,i(z)) = θ · NM(γ∞,i(z))
It follows that NQ(ψ(x)) = θ · NM(x) for every x ∈ lim−→i
Mpi(K), and thus ψ can be extended
to a unital algebra homomorphism ψ : M → eQe, where e := ψ(1) = lim−→i
ρi(1), which
satisfies the identity NQ(ψ(z)) = θ · NM(z) for all z ∈ M. In particular, NQ(e) = θ. Clearly,
ψ is injective.
It remains to show that ψ is surjective onto eQe. Let x ∈ Q, and fix η > 0. Take i large
,
δi <
,
NQ(e − ρi(1)) <
η
5
η
5
enough so that
εi <
η
10
and such that there exists an element ex ∈ span{x1, ..., xi} satisfying NQ(x −ex) < η
By (iv), there exists y ∈ Ai so that NQ(ex − y) < η
NQ(x − y) <
10 .
10 ; hence
η
5
.
We thus have
NQ(exe − ρi(1)yρi(1)) ≤ NQ(exe − exρi(1)) + NQ(exρi(1) − eyρi(1))
+ NQ(eyρi(1) − ρi(1)yρi(1)) <
3
5
η.
On the other hand, it follows from (ii) that there exists z ∈ Mpi+1(K) such that
NQ(ρi(1)yρi(1) − ρi+1(z)) < δi <
η
5
.
Also, for i + 1 < h, we get from (2.1):
NQ(ρh(γh,i+1(z)) − ρi+1(z)) < δi+1 + · · · + δh−1 < δi+1(cid:16)1 + · · · + 2−h+i+2(cid:17) < 2δi+1 < δi ,
6
PERE ARA AND JOAN CLARAMUNT
and so letting h → ∞ leads to
NQ(ψ(γ∞,i+1(z)) − ρi+1(z)) ≤ δi <
η
5
.
Using the above inequalities, we obtain
NQ(exe − ψ(γ∞,i+1(z))) ≤ NQ(exe − ρi(1)yρi(1)) + NQ(ρi(1)yρi(1) − ρi+1(z))
+ NQ(ρi+1(z) − ψ(γ∞,i+1(z))) ≤
3
5
η +
1
5
η +
1
5
η = η.
By choosing a decreasing sequence ηn → 0, it follows that there is a sequence {wn}n of
elements in M such that limn ψ(wn) = exe. Since NM(wi − wj) = θ−1 · NQ(ψ(wi) − ψ(wj))
for all i, j, it follows that {wn}n is a Cauchy sequence in M, and hence convergent to w ∈ M
satisfying ψ(w) = exe. This shows that ψ is surjective.
(cid:3)
We now show how Theorem 2.2 follows from Lemma 2.4, assuming we are able to show
that the hypotheses of that lemma are satisfied. This indeed follows as in [16]. Take θ = 1/2
and apply Lemma 2.4 to obtain an isomorphism ψ : M → eQe, where NQ(e) = 1/2. Since
eQ ∼= (1 − e)Q as right Q-modules (by [15, Corollary 9.16]), we get an isomorphism of K-
algebras Q ∼= M2(eQe). Hence, we obtain an isomorphism M ∼= M2(M) ∼= M2(eQe) ∼= Q.
It remains to show that the hypotheses of Lemma 2.4 are satisfied. We need a preliminary
lemma, which might be of independent interest.
Lemma 2.5. Let p be a positive integer. Then there exists a constant K(p), depending only
on p, such that for any field K, for any ε > 0, for any pair A ⊆ B, where B is a unital K-
algebra and A is a unital regular subalgebra of B, for any pseudo-rank function N on B, and
for every K-algebra homomorphism ρ : Mp(K) → B such that {ρ(eij) i, j = 1, . . . , p} ⊆ε A
with respect to the N-metric, where eij denote the canonical matrix units in Mp(K), there
exists a K-algebra homomorphism ψ : Mp(K) → A such that
N(ρ(eij) − ψ(eij)) < K(p)ε
for 1 ≤, i, j ≤ p.
Proof. We proceed by induction on p. Let p = 1, and let K, ε, A, B, N and ρ : K → B be
as in the statement. Then ρ(1) is an idempotent in B and, by assumption, there is x ∈ A
such that N(ρ(1) − x) < ε. By [15, Lemma 19.3], there exists an idempotent g ∈ A such
x − g ∈ A(x − x2) and it follows that N(ρ(1) − g) < 4ε. Therefore we can take K(1) = 4.
Now assume that p ≥ 2 and that there is a constant K(p − 1) satisfying the property
corresponding to p − 1. Let K, ε, A, B, N and ρ : Mp(K) → B be as in the statement. We
identify Mp−1(K) with the subalgebra of Mp(K) generated by eij with 1 ≤ i, j ≤ p − 1.
By the induction hypothesis, there is a set of (p − 1) × (p − 1) matrix units xij ∈ A (so
that xijxkl = δjkxil for 1 ≤ i, j, k, l ≤ p − 1) satisfying N(ρ(eij) − xij) < K(p − 1)ε for all
1 ≤ i, j ≤ p − 1. By hypothesis, there are z1p, zp1 ∈ A such that N(ρ(e1p) − z1p) < ε and
N(ρ(ep1) − zp1) < ε. Our first task is to modify z1p and zp1 in order to obtain new elements
z′
1p and z′
(2.2)
for 1 ≤ i ≤ p − 1,
p1 such that
z′
1pxi1 = 0 = x1iz′
p1
VON NEUMANN FACTOR
7
with suitable bounds on the ranks. To get the desired elements, we proceed by induction on
i. We will only prove the result for the position (1, p). The element in the position (p, 1) is
built in a similar way. For i = 1, we use that A is regular to obtain an idempotent g1 ∈ A
such that
Note that
z1px11A = g1A.
N(g1) = N(z1px11) = N(z1px11 − ρ(e1p)ρ(e11)) < ε + K(p − 1)ε = (K(p − 1) + 1)ε.
Now take z(1)
1p := (1 − g1)z1p. We get that z(1)
1p x11 = 0 and that
N(z(1)
1p − ρ(e1p)) ≤ N(z1p − ρ(e1p)) + N(g1) < (K(p − 1) + 2)ε.
Iterating this process, we get, for each 1 ≤ i ≤ p−1, an element z(i)
for 1 ≤ j ≤ i and
1p in A such that z(i)
1p xj1 = 0
N(z(i)
1p − ρ(e1p)) < ((2i − 1)K(p − 1) + 2i)ε.
Therefore, taking z′
elements z′
1p, z′
p1 ∈ A satisfying (2.2) and such that
1p = z(p−1)
1p
and the element z′
p1 := z(p−1)
p1
built in a similar fashion, we get
(2.3)
max{N(z′
1p − ρ(e1p)), N(z′
p1 − ρ(ep1))} < ((2p−1 − 1)K(p − 1) + 2p−1)ε.
The next step is to convert z′
p1 into mutually quasi-inverse elements in A. Indeed, we
will replace in addition our original elements x1i and xi1, for 1 ≤ i ≤ p − 1, in order to get
a coherent family of "partial matrix units" y1j, yj1 for 1 ≤ j ≤ p. For this we will use [15,
Lemma 19.3] and its proof. Consider the element x′
p1x11 ∈ A, and note that
11 := x11z′
1p and z′
1pz′
N(x′
11 − ρ(e11)) ≤ 2N(x11 − ρ(e11)) + N(z′
1p − ρ(e1p)) + N(z′
p1 − ρ(ep1))
< (K(p − 1) + 1)2pε,
where we have used the bound given by the induction hypothesis and (2.3). Therefore, we
get
N(x′
11 − (x′
11)2) < 3(K(p − 1) + 1)2pε.
11A ∩ Ax′
11 − g ∈ A(x′
11g = g, so that we get gz′
Now using [15, Lemma 19.3] and its proof, we can find an idempotent g ∈ A such that
g ∈ x′
It follows that g ≤ x11 and
that gx′
p1g = g. Set y11 = g ≤ x11, y1i = gx1i and yi1 = xi1g
for i = 2, . . . , p − 1. Also, set y1p = gz′
p1g. Then we have y1iyj1 = δijy11 for
1 ≤ i, j ≤ p. We have
1p and yp1 = z′
11 − (x′
1pz′
11)2) and x′
11g = g.
11, x′
N(g − ρ(e11)) ≤ N(g − x′
11) + N(x′
11 − ρ(e11))
< 3(K(p − 1) + 1)2pε + (K(p − 1) + 1)2pε = (K(p − 1) + 1)2p+2ε.
8
PERE ARA AND JOAN CLARAMUNT
Using this inequality, we obtain
N(y1p − ρ(e1p)) = N(gz′
1p − ρ(e11)ρ(e1p)) ≤ N(g − ρ(e11)) + N(z′
1p − ρ(e1p))
< (K(p − 1) + 1)2p+2ε + ((2p−1 − 1)K(p − 1) + 2p−1)ε
= ((2p+2 + 2p−1 − 1)K(p − 1) + 2p+2 + 2p−1)ε.
Similar computations give that max{N(y1i − ρ(e1i)), N(yi1 − ρ(ei1)) : i = 1, . . . , p} < ((2p+2 +
2p−1 − 1)K(p − 1) + 2p+2 + 2p−1)ε. Finally, put yij = yi1y1j. We obtain that {yij} is a
complete system of p×p matrix units in A, so that we can define a K-algebra homomorphism
ψ : Mp(K) → A such that ψ(eij) = yij. Moreover we have N(ρ(eij) − ψ(eij)) < K(p)ε, where
K(p) := (2p+3 + 2p − 2)K(p − 1) + (2p+3 + 2p). This concludes the proof.
(cid:3)
We now show that that the hypotheses of Lemma 2.4 are satisfied (assuming condition (iii)
in Theorem 2.2). This is obtained from the next lemma by applying induction (starting with
p0 = q0 = 1 and A0 = K).
Lemma 2.6. Let Q be a continuous factor with unique rank function NQ. Assume there
exists a dense subalgebra Q0 of Q of countable dimension, and let {xn}n be a K-basis of
Q0. Assume that Q satisfies condition (iii) in Theorem 2.2, and let θ be a real number such
that 0 < θ < 1. Let p be a positive integer such that there exist an algebra homomorphism
ρ : Mp(K) → Q, a matricial subalgebra A ⊆ Q, a positive integer m, and ε > 0 such that
q > θ for some positive integer q.
(a) NQ(ρ(1)) = p
(b) {ρ(eij) i, j = 1, . . . , p} ⊆ε A, and span{x1, ..., xm} ⊆ε A.
(c) ε <
q − θ(cid:17), where K(p) is the constant introduced in Lemma 2.5.
48K(p)p2(cid:16) p
1
Then there exist positive integers p′, g, q′, with p′ = gp, a real number ε′ > 0, an algebra
homomorphism ρ′ : Mp′(K) → Q and a matricial subalgebra A′ ⊆ Q such that the following
conditions hold:
(1) NQ(ρ′(1)) = p′/q′.
(2)
0 <
p′
q′ − θ <
1
2(cid:16)p
q
− θ(cid:17).
(3) For each x ∈ ρ(1)Aρ(1) there exists y ∈ Mp′(K) such that
NQ(x − ρ′(y)) <
p
q
− θ.
(4) For each z ∈ Mp(K), we have
NQ(ρ(z) − ρ′(γ(z))) <
p
q
− θ,
where γ : Mp(K) → Mp′(K) = Mp(K)⊗Mg(K) is the canonical unital homomorphism
sending z to z ⊗ 1g.
(5) {ρ′(e′
ij) i, j = 1, . . . , p′} ⊆ε′ A′, and span{x1, ..., xm, xm+1} ⊆ε′ A′, where {e′
ij i, j =
1, . . . , p′} denote the canonical matrix units in Mp′(K).
VON NEUMANN FACTOR
9
(6) ε′ <
1
48K(p′)p′2(cid:16) p′
q′ − θ(cid:17).
Proof. We denote by eij, for 1 ≤ i, j ≤ p, the canonical matrix units in Mp(K). Set f ′ :=
ρ(e11), which is an idempotent in Q with NQ(f ′) = 1/q (because NQ(ρ(1)) = p/q). By
(b) and Lemma 2.5, there exists a K-algebra homomorphism ψ : Mp(K) → A such that
N(ρ(eij) − ψ(eij)) < K(p)ε for 1 ≤ i, j ≤ p. Set f = ψ(e11) ∈ A and observe that
(2.4)
NQ(f − f ′) < K(p)ε.
Since A is matricial, we can write f = f1 + · · · + fk, where f1, f2, . . . , fk are nonzero mutually
orthogonal idempotents belonging to different simple factors of A. We can now consider, for
each 1 ≤ i ≤ k, a set of matrix units {f (i)
is a
minimal idempotent in the simple factor to which fi belongs, and such that
: 1 ≤ j, l ≤ ri} inside fiAfi such that each f (i)
jj
jl
for i = 1, . . . , k. Note that
riXj=1
f (i)
jj = fi
NQ(f ) =
kXi=1
riNQ(f (i)
11 )
so that by [15, Lemma 19.1(e)] and (2.4), we get
(2.5)
1
q
−
kXi=1
(cid:12)(cid:12)(cid:12)
riNQ(f (i)
11 )(cid:12)(cid:12)(cid:12) = NQ(f ′) − NQ(f ) ≤ NQ(f − f ′) < K(p)ε
We now approximate each real number NQ(f (i)
set
11 ) by a rational number pi/qi. Concretely, we
1
i=1 ri)(cid:16) p
q
− θ(cid:17) ,
δ :=
48p(Pk
and take positive integers pi, qi so that 0 < NQ(f (i)
we may assume that qi = q′ for i = 1, . . . , k. Let α′ be such that 1/α′ = Pk
observe that, by using (2.5), we have
11 )−pi/qi < δ. Taking common denominator,
i=1 ripi/q′, and
p/q − p/α′ ≤ ph(cid:12)(cid:12)(cid:12)
1
q
−
kXi=1
riNQ(f (i)
11 )(cid:12)(cid:12)(cid:12) +
ri(cid:17)δ <
< K(p)pǫ + p(cid:16) kXi=1
kXi=1
ri(cid:16)NQ(f (i)
11 ) −
pi
q′(cid:17)i
1
48p
(p/q − θ) +
1
48
(p/q − θ) ≤
1
24
(p/q − θ).
So in particular
(2.6)
p/q − p/α′ <
1
8
(p/q − θ)
10
PERE ARA AND JOAN CLARAMUNT
Now take
Then Pk
λi =
α′piri
q′
and
εi =
λi
pri(cid:16)p
q
− θ(cid:17).
i=1 λi = 1. Moreover λi, εi (and of course pi/q′), i = 1, . . . , k, do not depend on
replacing pi and q′ by piN and q′N respectively, for any integer N ≥ 1, so we can assume
that pi and q′ are arbitrarily large. Taking pi large enough, we see that we can find non-
negative integers p′
(2.7)
i, for 1 ≤ i ≤ k, such that
p′
pi
i
q′ −
q′ <
for i = 1 . . . , k. Indeed, using (2.6), we can see that
− θ(cid:17)(cid:17) <
8 (cid:16) α′
p(cid:16) p
5εi
8
5εiq′
5pi
=
<
8
q
3εi
4
5pi
8
8
7
=
5
7
pi < pi .
We can choose q′ big enough so that εiq′
( 5
8εiq′, 3
4εiq′). Since 5εiq′
Now, using that Pk
8 < pi we can find a non-negative integer p′
i=1 λi = 1, we get
5
8p(cid:16)p
q
− θ(cid:17) <
1
α′ −(cid:16) kXi=1
ri
p′
i
q′(cid:17) <
3
4p(cid:16)p
q
− θ(cid:17).
8 > 1 and thus there is an integer in the open interval
i such that (2.7) holds.
Hence, setting g =Pk
i=1 rip′
Hence, using (2.6), we get
(2.8)
i and p′ = pg, we get
p′
q′ <
p
α′ −
5
q
8(cid:16) p
2(cid:16)p
1
q
− θ(cid:17) <
− θ(cid:17) <
p
q
−
p′
q′ <
q
3
4(cid:16) p
8(cid:16)p
7
q
− θ(cid:17).
− θ(cid:17).
Now, since Q is continuous, there exists an idempotent e in Q such that NQ(e) = 1/q′, and
since p′
11 ) and Q is simple and injective, we get
i/q′ < NQ(f (i)
for i = 1, . . . , k. We may (and will) assume that e ≤ f (1)
of matrix units
11 . Therefore we can build a system
i · e . f (i)
p′
11
{h(i1,i2)
(j1,j2),(u1,u2) : 1 ≤ i1, i2 ≤ k, 1 ≤ j1 ≤ ri1, 1 ≤ j2 ≤ ri2, 1 ≤ u1 ≤ p′
i1, 1 ≤ u2 ≤ p′
i2},
with
(j1,j2),(u1,u2)h(i′
h(i1,i2)
(1,1),(1,1), {h(i,i)
(j ′
such that e = h(1,1)
f (i)
11 Qf (i)
11 for all i, and
1,i′
2)
2),(u′
1,j ′
2) = δi2,i′
1,u′
1 · δj2,j ′
1 · δu2,u′
1 · h(i1,i′
2)
2),(u1,u′
(j1,j ′
2) ,
(1,1),(u1,u2) : 1 ≤ u1, u2 ≤ p′
i} is a system of matrix units inside
h(i,i)
(j1,j2),(u1,u2) = f (i)
j1,1h(i,i)
(1,1),(u1,u2)f (i)
1,j2
VON NEUMANN FACTOR
11
for i = 1, . . . , k, 1 ≤ j1, j2 ≤ ri and 1 ≤ u1, u2 ≤ p′
we proceed as follows. First we construct a family
i. To build such a system of matrix units,
{h(1,i)
(1,1),(1,u), h(i,1)
(1,1),(u,1) : 1 ≤ u ≤ p′
i}
so that
h(1,i)
(1,1),(1,u)h(i,1)
(1,1),(u,1) = h(1,1)
(1,1),(1,1) = e,
eh(1,i)
(1,1),(1,u) = h(1,i)
(1,1),(1,u),
h(i,1)
(1,1),(u,1)e = h(i,1)
(1,1),(u,1),
(1,1),(u,u)
:= h(i,1)
(1,1),(u,1)h(1,i)
(1,1),(1,u) are pairwise orthogonal idempotents inside
and such that h(i,i)
f (i)
11 Qf (i)
11 for all i. Then, define, for 1 ≤ i ≤ k, 1 ≤ j ≤ ri and 1 ≤ u ≤ p′
i,
(1,1),(1,u) · f (i)
1,j .
h(1,i)
(1,j),(1,u) = h(1,i)
h(i,1)
(j,1),(u,1) = f (i)
j,1 · h(i,1)
(1,1),(u,1),
Finally, set, for 1 ≤ i1, i2 ≤ k, 1 ≤ j1 ≤ ri1, 1 ≤ j2 ≤ ri2, 1 ≤ u1 ≤ p′
i1, 1 ≤ u2 ≤ p′
i2,
h(i1,i2)
(j1,j2),(u1,u2) = h(i1,1)
(j1,1),(u1,1) · h(1,i2)
(1,j2),(1,u2).
It is straightforward to verify that the family {h(i1,i2)
(j1,j2),(u1,u2)} satisfies the required properties.
(j1,j2),(u1,u2)} is a system of g × g-matrix
units inside f Qf , and we can now define an algebra homomorphism ρ′ : Mp′(K) = Mp(K) ⊗
Mg(K) → Q by the rule
Recalling that g = Pk
i, we get that {h(i1,i2)
i=1 rip′
ρ′(eij ⊗ e(i1,i2)
(j1,j2),(u1,u2)) = ψ(ei1)h(i1,i2)
(j1,j2),(u1,u2)ψ(e1j) ,
where {e(i1,i2)
(j1,j2),(u1,u2)} is a complete system of matrix units in Mg(K).
It remains to verify properties (1)-(7).
(1) We have
NQ(ρ′(1)) = pgNQ(e) =
p′
q′ ,
as desired.
(2) This follows from (2.8).
(3) Let x ∈ ρ(1)Aρ(1). Then we can write
x = ρ(1)x′ρ(1) =
pXa,b=1
ρ(ea1)f ′ρ(e1a)x′ρ(eb1)f ′ρ(e1b) ,
where x′ ∈ A. Now by (b) we can approximate each ρ(e1a), ρ(eb1) by an element of A:
NQ(ρ(e1a) − x1a) < ε
for x1a ∈ A
NQ(ρ(eb1) − xb1) < ε
for xb1 ∈ A
Thus we can consider the element
pXa,b=1
ex =
ψ(ea1) f x1ax′xb1f
ψ(e1b) ∈ A ,
xab∈f Af
{z
}
12
PERE ARA AND JOAN CLARAMUNT
so that NQ(x−ex) < p2(4K(p) + 2)ε < 5K(p)p2ε (using that K(p) ≥ 4 for the last inequality).
Now we can write each xab in the form
xab =
for some scalars λ(a, b)(i)
jl ∈ K.
Take
kXi=1
riXj,l=1
λ(a, b)(i)
jl f (i)
jl
λ(a, b)(i)
jl(cid:16) p′
iXu=1
e(i,i)
(j,l),(u,u)(cid:17)(cid:17) ∈ Mp′(K).
jl f (i)
j1 and dbil = f (i)
1l ψ(e1b), we have
h(i,i)
(1,1),(u,u)(cid:17)dbil(cid:17)
(1,1),(u,u)(cid:17)(cid:17)(cid:17)
h(i,i)
p′
i
p′
i
11 −
11 ) −
i=1 rip′
q′
riNQ(f (i)
cbil(cid:16)f (i)
riXl=1
kXi=1
iXu=1
11 ) − NQ(cid:16) p′
ri(cid:16)NQ(f (i)
iXu=1
q′(cid:17)(cid:17)
ri(cid:16)NQ(f (i)
11 )(cid:17) − pPk
qi +(cid:16)p
− θ(cid:17)
8(cid:16)p
48(cid:16) p
− θ(cid:17) ≤
− θ(cid:17) +
11 ) −
riNQ(f (i)
−
1
5
7
q
q
q
p′
q′(cid:17)
e(i,i)
(j,j),(u,u) = γ(z).
q
43
48(cid:16)p
48(cid:16)p
− θ(cid:17)
− θ(cid:17) =
43
q
p
q
− θ.
Writing cbil =Pp
y =
riXj,l=1
j=1 ψ(ea1)λ(a, b)(i)
eab ⊗(cid:16) kXi=1
pXa,b=1
a=1Pri
NQ(ex − ρ′(y)) = NQ(cid:16) pXb=1
≤ p(cid:16) kXi=1
= p(cid:16) kXi=1
= p(cid:16) kXi=1
= ph kXi=1
Using (2.5) and (2.8), we get
Putting everything together,
q
q
1
7
≤
NQ(ex − ρ′(y)) < K(p)εp +
48p(cid:16) p
48(cid:16)p
8(cid:16)p
− θ(cid:17) +
− θ(cid:17) <
(4) Suppose now that x = ρ(z) for z = Pp
riXj=1
iXu=1
NQ(x − ρ′(y)) < 5K(p)p2ε +
kXi=1
y = z ⊗
43
q
p′
can see that, using the same construction of y given in (3), we obtain
a,b=1 µabeab ∈ Mp(K), with µab ∈ K. Then we
VON NEUMANN FACTOR
13
To conclude the proof, just take ε′ > 0 satisfying ε′ <
(iii) in Theorem 2.2, consider a matricial subalgebra A′ such that {ρ′(e′
A′ and span{x1, ..., xm, xm+1} ⊆ε′ A′.
1
48K(p′)p′2(cid:16) p′
q′ −θ(cid:17) and, using condition
ij) i, j = 1, . . . , p′} ⊆ε′
(cid:3)
3. D-rings
We now consider a generalization of Theorem 2.2 to D-rings, where D is a division ring.
We have not found a reasonable analogue of the local condition (iii) in this setting, but we
are able to extend condition (ii).
The reason we consider this generalization is the question raised by Elek and Jaikin-
Zapirain of whether the completion of the ∗-regular closure of the group algebra of a countable
amenable ICC-group is isomorphic to either Mn(D), n ≥ 1, or to MD, for some division ring
D. Here MD is the completion of lim−→n
M2n(D) with respect to its unique rank function.
Throughout this section, D will denote a division ring, and K will stand for the center of
D.
A D-ring is a unital ring R together with a unital ring homomorphism ι : D → R. A
morphism of D-rings R1 → R2 is a ring homomorphism ϕ : R1 → R2 such that ι2 = ϕ ◦ ι1.
A matricial D-ring is a D-ring A which is isomorphic as a D-ring to a finite direct product
Mn1(D) × · · · × Mnr (D), where the structure of D-ring of the latter is the canonical one.
A D-ring A is an ultramatricial D-ring if it is isomorphic as D-ring to a direct limit of a
sequence (An, ϕn) of matricial D-rings An and D-ring homomorphisms ϕn : An → An+1.
We start with a simple lemma.
Lemma 3.1. There is a unique rank function S on the (possibly non regular) simple D-ring
M2n(D) is dense in D ⊗K MK with respect to
D ⊗K MK, and D ⊗K (lim−→n
the S-metric.
M2n(K)) ∼= lim−→n
Proof. We denote by NMK the unique rank function on MK.
The ring D ⊗K MK is simple by [5, Corollary 7.1.3]. Let x = Pk
M2n(K) be such that NMK (xi − yi) < ε
and ε > 0. Let yi ∈ lim−→n
Then, for any rank function S on D ⊗K MK, we have
i=1 di ⊗ xi ∈ D ⊗K MK
i=1 di ⊗ yi.
k , and set y :=Pk
S(x − y) ≤
kXi=1
S(1 ⊗ (xi − yi)) =
kXi=1
NMK (xi − yi) < ε.
Since there is a unique rank function on lim−→n
unique rank function S on D ⊗K MK, and that D ⊗K (lim−→n
with respect to the S-metric.
M2n(D), this shows at once that there is a
M2n(K)) is dense in D ⊗K MK
(cid:3)
Theorem 3.2. Let A be an ultramatricial D-ring, and let N be an extremal pseudo-rank
function on A such that the completion Q of A with respect to N is a continuous factor.
Then there is an isomorphism of D-rings Q ∼= MD.
14
PERE ARA AND JOAN CLARAMUNT
Proof. We can assume that A = lim−→n
map ϕn : An → An+1 is an injective morphism of D-rings.
(An, ϕn), where each An is a matricial D-ring and each
Write Bn = CAn(D) for the centralizer of D in An. Then Bn is a matricial K-algebra and,
∼= D ⊗K Bn. Moreover, we have ϕn(Bn) ⊆ Bn+1
(Bn, (ϕn)Bn) is an ultramatricial
since K is the center of D, we have An
for all n ≥ 1, and A ∼= D ⊗K B, where B = CA(D) = lim−→n
K-algebra.
Now, it is not hard to show that the restriction map S → SB defines an affine homeomor-
phism P(A) ∼= P(B). Consequently, the restriction NB of N to B is an extremal pseudo-rank
function on B. Moreover, since N(A) = N(B), it follows that N(B) is a dense subset of
the unit interval, which implies that the completion R of B in the NB-metric is a continu-
ous factor over K. Now it follows from Theorem 2.2 that there is a K-algebra isomorphism
ψ′ : MK → R, which induces an isomorphism of D-rings
ψ := idD ⊗ ψ′ : D ⊗K MK → D ⊗K R.
Since A ⊆ D ⊗K R = ψ(D ⊗K MK) ⊆ Q, it follows that ψ(D ⊗K MK) is dense in Q. By
M2n(K))) is dense in ψ(D ⊗K MK) with respect to the restriction
Lemma 3.1, ψ(D ⊗K (lim−→n
M2n(K))) is dense in Q. Hence, the restriction of ψ to
of NQ to it, therefore ψ(D ⊗K (lim−→n
D ⊗K (lim−→n
M2n(D) gives a rank-preserving isomorphism of D-rings from
M2n(D) onto a dense D-subring of Q, and thus it can be uniquely extended to an
lim−→n
isomorphism from MD onto Q.
(cid:3)
M2n(K)) ∼= lim−→n
4. Fields with involution
In this section, we will consider the corresponding problem for ∗-algebras. Again, the
motivation comes from the theory of group algebras. If K is a subfield of C closed under
complex conjugation, and G is a countable discrete group, then there is a natural involution
on the group algebra K[G], and the completion of the ∗-regular closure of K[G] in U(G)
is a ∗-regular ring containing K[G] as a ∗-subalgebra.
It would be thus desirable to find
conditions under which this completion is ∗-isomorphic to MK, where MK is endowed with
M2n(K), which is in turn obtained by
the involution induced from the involution on lim−→n
endowing each algebra M2n(K) with the conjugate-transpose involution.
We recall some facts about ∗-regular rings and their completions (see for instance [1, 3]). A
∗-regular ring is a regular ring endowed with a proper involution, that is, an involution ∗ such
that x∗x = 0 implies x = 0. The involution is called positive definite in case the condition
nXi=1
x∗
i xi = 0 =⇒ xi = 0
(i = 1, . . . , n)
holds for each positive integer n. If R is a ∗-regular ring with positive definite involution,
then Mn(R), endowed with the ∗-transpose involution, is a ∗-regular ring.
We will work with ∗-algebras over a field with positive definite involution (F, ∗). The
involution on Mn(F ) will always be the ∗-transpose involution.
VON NEUMANN FACTOR
15
The ∗-algebra A is standard matricial if A = Mn(1)(F ) × · · · × Mn(r)(F ) for some positive
integers n(1), . . . , n(r). A standard map between two standard matricial ∗-algebras A, B is a
block-diagonal ∗-homomorphism A → B (see [1, p. 232]). A standard ultramatricial ∗-algebra
Φ3−→ · · · of standard matricial ∗-algebras An
is a direct limit of a sequence A1
and standard maps Φn : An → An+1. An ultramatricial ∗-algebra is a ∗-algebra which is ∗-
isomorphic to the direct limit of a sequence of standard matricial ∗-algebras An and ∗-algebra
maps Φn : An → An+1. Let A be a ∗-algebra which is ∗-isomorphic to a standard matricial
algebra, through a ∗-isomorphism
Φ2−→ A3
Φ1−→ A2
γ : A → Mn(1)(F ) × · · · × Mn(r)(F ).
Then we say that a projection (i.e., a self-adjoint idempotent) p in A is standard (with respect
to γ) in case, for each i = 1, 2, . . . , r, the i-th component γ(p)i of γ(p) is a diagonal projection
in Mn(i)(F ).
Two idempotents e, f ∈ R are equivalent, written e ∼ f , if there are x ∈ eRf and y ∈ f Re
If e, f are projections of a ∗-ring R, then we say that e is
such that e = xy, f = yx.
∗-equivalent to f , written e
∗∼ f , in case there is x ∈ eRf such that e = xx∗ and f = x∗x.
If R is a ∗-regular ring and x ∈ R, then there exist unique projections LP(x) and RP(x),
called the left and the right projections of x, such that xR = LP(x)R and Rx = R · RP(x).
Moreover, with e = LP(x) and f = RP(x), there exists a unique element y ∈ R, the relative
inverse of x, such that xy = e and yx = f . We will denote the relative inverse of x by x.
A ∗-regular ring R satisfies the condition LP
∗∼ RP(x) holds for each
x ∈ R. Observe that R satisfies LP ∗∼ RP if and only if equivalent projections of R are
∗-equivalent ([1, Lemma 1.1]). In general this condition is not satisfied for a ∗-regular ring,
but many ∗-regular rings satisfy it. It is worth to mention that for a field F with positive
∗∼ RP for all n ≥ 1 if and only if F is ∗-Pythagorean
definite involution, Mn(F ) satisfies LP
[12, Theorem 4.9] (see also [11, Theorem 4.5], [1, Theorem 1.12]).
∗∼ RP in case LP(x)
The following result is relevant for our purposes:
Theorem 4.1. [1, Theorem 3.5] Let (F, ∗) be a field with positive definite involution, let A
be a standard ultramatricial ∗-algebra, and let N be a pseudo-rank function on A. Then the
type II part of the N-completion of A is a ∗-regular ring satisfying LP
∗∼ RP.
As a consequence of this result, the ∗-algebra MF always satisfies LP ∗∼ RP, independently
of whether the field F is ∗-Pythagorean or not.
We collect, for the convenience of the reader, some properties of a pseudo-rank function on
a ∗-regular ring. For an element r of a ∗-regular ring R we denote by r the relative inverse
of r.
Lemma 4.2. Let N be a pseudo-rank function on a ∗-regular ring R. The following hold:
(a) The involution is isometric, that is, N(r∗) = N(r) for each r ∈ R.
(b) N(r − s) ≤ 3N(r − s) for all r, s ∈ R.
(c) N(LP(r) − LP(s)) ≤ 4N(r − s), and N(RP(r) − RP(s)) ≤ 4N(r − s) for all r, s ∈ R.
16
PERE ARA AND JOAN CLARAMUNT
(d) Suppose that e1, e2, f1, f2 are projections in R such that f1
i ≤ ei such that e′
1
for i = 1, 2. Then there exist projections e′
for i = 1, 2.
∗∼ f2 and N(ei − fi) ≤ ε
∗∼ e′
i) ≤ 5ε
2 and N(ei − e′
Proof. (a) See the proof of Proposition 1 in [10] or [17, Proposition 6.11].
(b) In [4, p. 310], it is shown that N(r − s) ≤ 19N(r − s), and the authors comment that
K. R. Goodearl has reduced 19 to 5. Here we show that indeed it can be reduced to 3.
Let e = rr, f = rr, g = ss and h = ss. Clearly r∗r + (1 − f ) and ss∗ + (1 − g) are invertible
in R and so
(4.1)
N(r − s) = N((r∗r + (1 − f ))(r − s)(ss∗ + (1 − g))).
On the other hand, we have
N((r∗r + (1 − f ))(r − s)(ss∗ + (1 − g))) = N(r∗ss∗ − r∗rs∗ + r∗(1 − g) − (1 − f )s∗)
≤ N(r∗ss∗ − r∗rs∗) + N(r∗(1 − g) − (1 − f )s∗)
≤ N(s − r) + N(r∗ − s∗) + N(f s∗ − r∗g)
= 2N(s − r) + N(f (s∗ − r∗)g) ≤ 3N(r − s).
By (4.1), we get N(r − s) ≤ 3N(r − s).
(c) Using (b), we get
N(RP(r) − RP(s)) = N(rr − ss) ≤ N((r − s)r) + N(s(r − s))
≤ 3N(r − s) + N(r − s) = 4N(r − s).
The proof for LP is similar.
(d) We follow the idea in [1, proof of Lemma 2.6]. Let w ∈ f1Rf2 be a partial isometry
such that f1 = ww∗ and f2 = w∗w. Consider the self-adjoint element a = e1 − e1ww∗e1 and
set p1 := LP(a) = RP(a) ≤ e1. Then
N(p1) = N(a) = N(e1 − e1f1e1) ≤ ε.
1 := e1 − p1. Then N(e1 − p′
1) ≤ ε and, since p′
1ap′
1 = 0, we have p′
1 = w′(w′)∗, where
Set p′
w′ := p′
1w.
Now observe that (w′)∗w′ = w∗p′
1w ≤ w∗w = f2. Consider the elements
b = e2 − e2(w′)∗w′e2,
e′′
2 = LP(b) = RP(b).
We have e′′
2 ≤ e2 and
N(e′′
2) = N(b) = N(e2 − e2(w′)∗w′e2) ≤ N(e2 − (w′)∗w′)
≤ N(e2 − f2) + N(w∗w − w∗p′
≤ ε + N(f1 − p′
1)
≤ ε + N(f1 − e1) + N(e1 − p′
1w)
1) ≤ 3ε.
Set e′
N(e2 − e′
2 = e2 − e′′
2) = N(e′′
2. As before, we have e′
2) ≤ 3ε. Write e′
1) = N(e1 − p′
1 = w′′(w′′)∗. Then e′
1) + N(p′
1 − e′
N(e1 − e′
1 ≤ p′
1 ≤ e1, e′
1
∗∼ e′
1) ≤ ε + N(w′f2(w′)∗ − w′e′
2, and
2(w′)∗)
VON NEUMANN FACTOR
17
2 = (w′′)∗w′′, where w′′ = w′e′
2 = p′
1we′
2, and
≤ ε + N(f2 − e2) + N(e2 − e′
2) ≤ 5ε.
(cid:3)
Lemma 4.3. Let R be a ∗-regular ring, and assume that R is complete with respect to a rank
function N. Then R satisfies LP ∗∼ RP if and only if, given equivalent projections p, q ∈ R
and ε > 0 there exist subprojections p′ ≤ p and q′ ≤ q such that p′ ∗∼ q′ and N(p − p′) < ε,
N(q − q′) < ε.
Proof. The "only if" direction follows trivially from [1, Lemma 1.1].
Set p′ := p − (Pn
∗∼ qi for i = 1, . . . , n, and N(p − (Pn
To show the "if" direction, suppose that p and q are equivalent projections of R. Assume
we have built, for some n ≥ 1, orthogonal projections p1, . . . , pn ≤ p and q1, . . . , qn ≤ q
i=1 qi)) < 2−n.
such that pi
i=1 qi). Then p′ ∼ q′ by [15, Theorems 19.7 and
∗∼ qn+1
4.14], so that there are subprojections pn+1 ≤ p′ and qn+1 ≤ q′ such that pn+1
and N(p′ − pn+1) < 2−n−1 and N(q′ − qn+1) < 2−n−1. Therefore we can build sequences
∗∼ qn, and
{pn} and {qn} of orthogonal subprojections of p and q respectively such that pn
i=1 qi)) < 2−n for all n ≥ 1. Let wn ∈ pnRqn be partial
i=1 pi)) < 2−n, N(q − (Pn
i=1 pi) and q′ = q − (Pn
i=1 pi)) < 2−n, N(q − (Pn
isometries such that pn = wnw∗
N(p − (Pn
n and qn = w∗
nwn. Then
N(wn) ≤ N(pn) ≤ N(cid:16)p − (
n−1Xi=1
pi)(cid:17) < 2−n+1 ,
and it follows that the sequence {Pn
i=1 wi}n converges to a partial isometry w ∈ pRq such
that p = ww∗ and q = w∗w. Hence R satisfies condition LP ∗∼ RP (by [1, Lemma 1.1]). (cid:3)
In order to state the local condition in our main result of this section, we need the following
somewhat technical definition.
Definition 4.4. Let R be a unital ∗-regular ring with pseudo-rank function N, and let A be
a unital ∗-subalgebra which is ∗-isomorphic to a standard matricial ∗-algebra. We say that a
projection p ∈ A is hereditarily quasi-standard if
(1) p is ∗-equivalent in A to a standard projection of A, and,
(2) for each subprojection p′ ≤ p, p′ ∈ A, and each ε > 0 there exists a unital ∗-subalgebra
A′ of R and a projection p′′ ∈ A′ satisfying the following properties:
(a) A′ is ∗-isomorphic to a standard matricial ∗-algebra,
(b) p′′ is ∗-equivalent in A′ to a standard projection of A′,
(c) p′′ ≤ p′ and N(p′ − p′′) < ε, and
(d) A ⊆ A′.
18
PERE ARA AND JOAN CLARAMUNT
We can now state the following analogue of Theorem 2.2. By a continuous ∗-factor over F
we mean a ∗-regular ring Q which is a ∗-algebra over F , and which is a continuous factor in
the sense of Section 2.
Theorem 4.5. Let (F, ∗) be a field with positive definite involution. Let Q be a continuous ∗-
factor over F , and assume that there exists a dense subalgebra (with respect to the NQ-metric
topology) Q0 ⊆ Q of countable F -dimension. The following are equivalent:
(i) Q ∼= MF as ∗-algebras.
(ii) Q is isomorphic as a ∗-algebra to B for a certain standard ultramatricial ∗-algebra B,
where the completion of B is taken with respect to the metric induced by an extremal
pseudo-rank function on B.
(iii) For every ε > 0, elements x1, ..., xn ∈ Q, and projections p1, p2 ∈ Q, there exist a ∗-
subalgebra A of Q, which is ∗-isomorphic to a standard matricial ∗-algebra, elements
y1, ..., yn ∈ A, and hereditarily quasi-standard projections q1, q2 ∈ A such that
N(pj − qj) < ε,
(j = 1, 2),
and
N(xi − yi) < ε
(i = 1, . . . , n).
Proof. Clearly, (i) =⇒ (ii).
(Bn, Φn) as a direct limit of a sequence of standard matricial
(ii) =⇒ (iii). Write B = lim−→n
∗-algebras Bn and standard maps Φn : Bn → Bn+1. Write Φji : Bi → Bj for the composition
maps Φj−1 ◦ · · · ◦ Φi, for i < j, and write θi : Bi → B ∼= Q for the canonical map. We identify
Q with B.
We will show that the desired ∗-subalgebra A satisfying the required conditions is of the
form θj(Bj). Since those algebras form an increasing sequence, whose union is dense in the
NQ-metric topology, we see that it is enough to deal only with the projections, and indeed
that it is enough to deal with a single projection. Let p be a projection in Q and let ε > 0.
Now there is some i ≥ 1 and an element x ∈ Bi such that N(p − θi(x)) < ε/8. Write
p1 := LP(x) ∈ Bi. By Lemma 4.2(c), we have
N(p − θi(p1)) = N(LP(p) − θi(LP(x))) = N(LP(p) − LP(θi(x))) ≤ 4N(p − θi(x)) <
ε
2
,
so that N(p − θi(p1)) < ε/2.
There exists a standard projection g in Bi such that p1 ∼ g in Bi. By the proof of [1,
1 ≤ Φji(p1), g′ ≤
Theorem 3.5], there are j > i and projections p′
Φji(g), g′ is a standard projection, p′
1)) < ε/2 and
1
N(θi(g) − θj(g′)) < ε/2. Therefore, p′
1 is ∗-equivalent to a standard projection in Bj, and
moreover
∗∼ g′, and moreover N(θi(p1) − θj(p′
1, g′ ∈ Bj such that p′
N(p − θj(p′
1)) ≤ N(p − θi(p1)) + N(θi(p1) − θj(p′
1)) < ε/2 + ε/2 = ε.
Now take A := θj(Bj) and q := θj(p′
1). Clearly, property (1) in Definition 4.4 is satisfied. To
show property (2), take a subprojection θj(p′) of q = θj(p′
1), where p′ is a subprojection of
p′
1, and δ > 0. Then we use the same argument as above but now applied to the projection
p′ of Bj and to δ > 0. We obtain k ≥ j and a projection θk(p′′) in the ∗-subalgebra θk(Bk)
VON NEUMANN FACTOR
19
such that the pair (θk(p′), θk(Bk)) satisfies properties (a)-(d) in Definition 4.4 (with ε replaced
with δ).
∗∼ RP. Let p1, p2 be equivalent projections of
(iii) =⇒ (i). We first show that Q satisfies LP
Q and ε > 0. Choose x ∈ p1Qp2 and y ∈ p2Qp1 such that p1 = xy and p2 = yx. Observe
that necessarily y = x, the relative inverse of x in Q. By (iii), there exists a ∗-subalgebra
A of Q, which is ∗-isomorphic to a standard matricial ∗-algebra, projections q1, q2 ∈ A such
∗∼ ei in A, i = 1, 2, for some standard projections e1, e2 ∈ A, and
that N(pi − qi) < ε, and qi
an element x1 ∈ A such that N(x − x1) < ε. Now set x′
1 := q1x1q2 ∈ A, and note that
N(x − x′
1) ≤ N(p1xp2 − q1xq2) + N(q1xq2 − q1x1q2) < 2ε + ε = 3ε.
It follows from Lemma 4.2(c) that, with q′′
1 := LP(x′
1) ∈ A and q′′
2 := RP(x′
1) ∈ A, we have
Moreover, we have q′′
1 = LP(x′
1) = q′′
2 . In addition, we get
N(pi − q′′
i ) < 12ε,
1) ∼ RP(x′
q′′
i ≤ qi,
(i = 1, 2).
N(qi − q′′
i ) ≤ N(qi − pi) + N(pi − q′′
i ) < ε + 12ε = 13ε.
∗∼ ei in A, we obtain in particular projections e′′
1 ∼ e′′
Write η = 13ε. Since qi
2
(in A) and N(ei − e′′
i ) < η. Now A is a standard matricial ∗-algebra, and the restriction of
N to A is a convex combination of the normalized rank functions on the different simple
components of A, so the above information enables us to build standard projections e′
i ≤ ei
i) < η for i = 1, 2. This in turn gives us projections q′
such that e′
i ≤ qi
1
∗∼ ei) such that q′
(through the ∗-equivalences qi
2 and N(qi − q′
The last step is to transfer these to p1, p2. For this, observe that
i ≤ ei such that e′′
i) < η for i = 1, 2.
2, and N(ei − e′
∗∼ e′
∗∼ q′
1
N(pi − q′
i) ≤ N(pi − qi) + N(qi − q′
i) < ε + η.
Since moreover q′
projections p′
Lemma 4.3 to conclude that Q satisfies LP
i ≤ pi such that p′
1
1 and q′
2 are ∗-equivalent, it follows from Lemma 4.2(d) that there exist
i) < 5(ε + η) = 70ε. Now we can apply
2 and N(pi − p′
∗∼ p′
∗∼ RP.
Now (i) is shown by using the same method employed in Section 2. We only need to prove a
variant of Lemma 2.4 with ∗-algebra homomorphisms ρi : Mpi(F ) → Q instead of just algebra
homomorphisms. For this, a new version of Lemmas 2.5 and 2.6 is required, as follows:
Lemma 4.6. Let p be a positive integer. Then there exists a constant K ∗(p), depending
only on p, such that for any field with involution F , for any ε > 0, for any pair A ⊆ B,
where B is a unital ∗-algebra over F , and A is a unital ∗-regular subalgebra of B, for any
pseudo-rank function N on B such that N(b∗) = N(b) for all b ∈ B, and for every ∗-algebra
homomorphism ρ : Mp(F ) → B such that {ρ(eij) i, j = 1, . . . , p} ⊆ε A with respect to the
N-metric, where eij denote the canonical matrix units in Mp(F ), there exists a ∗-algebra
homomorphism ψ : Mp(F ) → A such that
N(ρ(eij) − ψ(eij)) < K ∗(p)ε
for 1 ≤, i, j ≤ p.
If, in addition, we are given a projection f ∈ A such that N(ρ(e11) − f ) < ε, then the map ψ
can be built with the additional property that ψ(e11) ≤ f .
20
PERE ARA AND JOAN CLARAMUNT
Proof of Lemma 4.6. The proof follows the same steps as the proof of Lemma 2.5. There is
only an additional degree of approximation due to the fact that we need projections instead
of idempotents. Proceeding by induction on p, just as in the proof of Lemma 2.5, we start
with ∗-matrix units {xij} for 1 ≤ i, j ≤ p − 1, so that xji = x∗
ij for all i, j, and we have to
define new elements y1i, for i = 1, . . . , p, so that the family yij = y∗
1iy1j, 1 ≤ i, j ≤ p, is the
desired new family of ∗-matrix units. To this end, one only needs to replace the idempotent
g found in that proof by the projection LP(g). Using Lemma 4.2, one can easily control the
corresponding ranks.
The last part is proven by the same kind of induction, starting with ψ(1) = f for the case
(cid:3)
p = 1.
Lemma 4.7. Assume that Q satisfies condition (iii) in Theorem 4.5. Let θ be a real number
such that 0 < θ < 1 and let {xn}n be a K-basis of Q0. Let p be a positive integer such that
there exist a ∗-algebra homomorphism ρ : Mp(F ) → Q, a ∗-subalgebra A ⊆ Q, which is ∗-
isomorphic to a standard matricial ∗-algebra, a hereditarily quasi-standard projection g ∈ A,
a positive integer m, and ε > 0 such that
q > θ for some positive integer q.
(a) NQ(ρ(1)) = p
(b) NQ(ρ(e11) − g) < ε, where eij are the canonical matrix units of Mp(F ).
(c) {ρ(eij) i, j = 1, . . . , p} ⊆ε A, and span{x1, ..., xm} ⊆ε A.
(d) ε <
1
48K ∗(p)p2(cid:16) p
q − θ(cid:17).
Then there exist positive integers p′, t, q′, and a real number ε′ > 0 with p′ = tp, a ∗-algebra
homomorphism ρ′ : Mp′(F ) → Q, a ∗-subalgebra A′ ⊆ Q, which is ∗-isomorphic to a stan-
dard matricial ∗-algebra, and a hereditarily quasi-standard projection g′ ∈ A′, such that the
following conditions hold:
(1) NQ(ρ′(1)) = p′/q′.
(2)
0 <
p′
q′ − θ <
1
2(cid:16)p
q
− θ(cid:17).
(3) For each x ∈ ρ(1)Aρ(1) there exists y ∈ Mp′(F ) such that
NQ(x − ρ′(y)) <
p
q
− θ.
(4) For each z ∈ Mp(F ), we have
NQ(ρ(z) − ρ′(γ(z))) <
p
q
− θ,
where γ : Mp(F ) → Mp′(F ) = Mp(F )⊗Mt(F ) is the canonical unital ∗-homomorphism
sending z to z ⊗ 1t.
11) − g′) < ε′, where e′
ij are the canonical matrix units of Mp′(F ).
ij) i, j = 1, . . . , p′} ⊆ε′ A′, and span{x1, ..., xm, xm+1} ⊆ε′ A′
(5) NQ(ρ′(e′
(6) {ρ′(e′
(7) ε′ <
48K ∗(p′)p′2(cid:16) p′
q′ − θ(cid:17).
1
VON NEUMANN FACTOR
21
Proof of Lemma 4.7. The proof is very similar to the proof of Lemma 2.6. We only indicate
the points where the proof has to be modified.
We denote by eij, for 1 ≤ i, j ≤ p, the canonical matrix units in Mp(F ). Note that
e∗
ij = eji for all i, j. Set f ′ := ρ(e11), which is a projection in Q with NQ(f ′) = 1/q. By
hypothesis, there is a hereditarily quasi-standard projection g in the ∗-subalgebra A such that
NQ(f ′ − g) < ε. Now by Lemma 4.6 there exists a ∗-algebra homomorphism ψ : Mp(F ) → A
such that ψ(e11) ≤ g and NQ(ρ(eij) − ψ(eij)) < K ∗(p)ε for all i, j. Now by condition (2) in
Definition 4.4, there exists another ∗-subalgebra A′ of Q, which is ∗-isomorphic to a standard
matricial ∗-algebra and contains A, and a projection f ∈ A′, which is ∗-equivalent in A′ to a
standard projection of A′, such that f ≤ ψ(e11) and NQ(ψ(e11) − f ) < K ∗(p)ε − µ, where
µ = max{NQ(ρ(eij) − ψ(eij)) : i, j = 1, . . . , p}.
Now, setting ψ′(eij) = ψ(ei1)f ψ(e1j), we obtain that ψ′ is a ∗-algebra homomorphism from
Mp(F ) to A′, and that
NQ(ρ(eij) − ψ′(eij)) < µ + (K ∗(p)ε − µ) = K ∗(p)ε,
so that, after changing notation, we may assume that f = ψ(e11), and that f is ∗-equivalent
in A to a standard projection of A.
Since A is a standard matricial ∗-algebra, we can write f = f1 +· · ·+fk, where f1, f2, . . . , fk
are nonzero mutually orthogonal projections belonging to different simple factors of A. Since
f is ∗-equivalent in A to a standard projection, there exists, for each 1 ≤ i ≤ k, a set of
matrix units {f (i)
is a minimal projection in
the simple factor to which fi belongs, such that
: 1 ≤ j, l ≤ ri} inside fiAfi such that each f (i)
jj
jl
riXj=1
f (i)
jj = fi
for i = 1, . . . , k and moreover (f (i)
jl )∗ = f (i)
lj
for all i, j, l.
Now the proof follows the same steps as the one of Lemma 2.6. The idempotent e built in
that proof can be replaced now by a projection and, since Q satisfies LP ∗∼ RP, we have that
i · e is ∗-equivalent to a subprojection of f (i)
p′
for
all i, j, l, one builds a system of matrix units inside f Qf
11 . Using this and the fact that (f (i)
jl )∗ = f (i)
lj
{h(i1,i2)
(j1,j2),(u1,u2) : 1 ≤ i1, i2 ≤ k, 1 ≤ j1 ≤ ri1, 1 ≤ j2 ≤ ri2, 1 ≤ u1 ≤ p′
i1, 1 ≤ u2 ≤ p′
i2},
satisfying all the conditions stated in the proof of 2.6, and in addition
(h(i1,i2)
(j1,j2),(u1,u2))∗ = h(i2,i1)
(j2,j1),(u2,u1)
for all allowable indices.
We can now define a ∗-algebra homomorphism ρ′ : Mp′(K) = Mp(K) ⊗ Mt(K) → Q by the
rule
ρ′(eij ⊗ e(i1,i2)
(j1,j2),(u1,u2)ψ(e1j) ,
(j1,j2),(u1,u2)} is a complete system of ∗-matrix units in Mt(K).
(j1,j2),(u1,u2)) = ψ(ei1)h(i1,i2)
where {e(i1,i2)
22
PERE ARA AND JOAN CLARAMUNT
The verification of properties (1)-(7) is done in the same way, using condition (iii) to show
(cid:3)
that conditions (5) and (6) are satisfied.
Lemma 4.7 enables us to build the sequence of ∗-algebra homomorphisms ρi : Mpi(F ) → Q
satisfying the properties stated in Lemma 2.4, and the same proof gives a ∗-isomorphism from
M to Q, as desired.
(cid:3)
In case the base field with involution (F, ∗) is *-Pythagorean, we can derive a result which
is completely analogous to Theorem 2.2, as follows.
Corollary 4.8. Let (F, ∗) be a *-Pythagorean field with positive definite involution. Let Q be
a continuous ∗-factor over F , and assume that there exists a dense subalgebra (with respect
to the NQ-metric topology) Q0 ⊆ Q of countable F -dimension. The following are equivalent:
(i) Q ∼= MF as ∗-algebras.
(ii) Q is isomorphic as a ∗-algebra to B for a certain ultramatricial ∗-algebra B, where the
completion of B is taken with respect to the metric induced by an extremal pseudo-rank
function on B.
(iii) For every ε > 0 and elements x1, ..., xn ∈ Q, there exist a matricial ∗-subalgebra A of
Q, and elements y1, ..., yn ∈ A such that
NQ(xi − yi) < ε
(i = 1, . . . , k).
Proof. This follows from Theorem 4.5, by using the fact that Mn(F ) satisfies LP ∗∼ RP for all
∗∼ RP for
n ([12, Theorem 4.9]) and [1, Proposition 3.3]. Note that, since Mn(F ) satisfies LP
all n, every projection of a standard matricial ∗-algebra is hereditarily quasi-standard.
(cid:3)
Acknowledgments
The authors would like to thank Andrei Jaikin-Zapirain and Kevin O'Meara for their
useful suggestions. The authors also thank the referee for his/her very careful reading of the
manuscript and for his/her many suggestions.
References
[1] P. Ara, Matrix rings over *-regular rings and pseudo-rank functions, Pacific J. Math. 129 (1987), 209 --
241.
[2] P. Ara, J. Claramunt, Approximating the group algebra of the lamplighter by finite-dimensional algebras.
In preparation.
[3] S. K. Berberian, Baer ∗-rings. Die Grundlehren der mathematischen Wissenschaften, Band 195. Springer-
Verlag, New York-Berlin, 1972.
[4] B. Blackadar, D. Handelman, Dimension functions and traces on C*-algebras, J. Funct. Anal. 45 (1982),
297 -- 340.
[5] P. M. Cohn, Algebra. Vol. 3, Second edition, John Wiley & Sons, Ltd., Chichester, 1991.
[6] A. Connes, Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1, Ann. of Math. (2) 104 (1976),
73 -- 115.
[7] G. Elek, Connes embeddings and von Neumann regular closures of amenable group algebras, Trans. Amer.
Math. Soc. 365 (2013), 3019 -- 3039.
VON NEUMANN FACTOR
23
[8] G. Elek, Lamplighter groups and von Neumann continuous regular rings, Proc. Amer. Math. Soc. 144
(2016), 2871 -- 2883.
[9] G. Elek, Infinite dimensional representations of finite dimensional algebras and amenablility, Math.
Annalen 369 (2017), 397 -- 439.
[10] D. Handelman, Completions of rank rings, Canad. Math. Bull. 20 (1977), 199 -- 205.
[11] D. Handelman, Coordinatization applied to finite Baer *-rings, Trans. Amer. Math. Soc. 235 (1978),
1 -- 34.
[12] D. Handelman, Rings with involution as partially ordered abelian groups, Rocky Mountain J. Math. 11
(1981), 337 -- 381.
[13] F. J. Murray, J. von Neumann, On rings of operators. IV, Ann. of Math. (2) 44 (1943), 716 -- 808.
[14] K. R. Goodearl, Centers of regular self-injective rings, Pacific J. Math. 76 (1978), 381 -- 395.
[15] K. R. Goodearl, "Von Neumann Regular Rings", Pitman, London 1979; Second Ed., Krieger, Malabar,
Fl., 1991.
[16] I. Halperin, Von Neumann's manuscript on inductive limits of regular rings, Canad. J. Math. 20 (1968),
477 -- 483.
[17] A. Jaikin-Zapirain, L2-Betti numbers and their analogues in positive characteristic, to appear in Pro-
ceedings of Groups St. Andrews 2017.
[18] A. M. Sinclair, R. R. Smith, Finite von Neumann algebras and masas, London Mathematical Society
Lecture Note Series, 351. Cambridge University Press, Cambridge, 2008.
Departament de Matem`atiques, Universitat Aut`onoma de Barcelona, 08193 Bellaterra
(Barcelona), Spain.
E-mail address: [email protected]
Departament de Matem`atiques, Universitat Aut`onoma de Barcelona, 08193 Bellaterra
(Barcelona), Spain.
E-mail address: [email protected]
|
1610.03740 | 2 | 1610 | 2017-04-25T02:06:53 | Extensions of almost faithful prime ideals in virtually nilpotent mod-$p$ Iwasawa algebras | [
"math.RA",
"math.RT"
] | Let $G$ be a nilpotent-by-finite compact $p$-adic analytic group for some $p>2$, and $H = \mathbf{FN}_p(G)$ its finite-by-(nilpotent $p$-valuable) radical. Fix a finite field $k$ of characteristic $p$, and write $kG$ for the completed group ring of $G$ over $k$. We show that almost faithful $G$-stable prime ideals $P$ of $kH$ extend to prime ideals $PkG$ of $kG$. | math.RA | math | Extensions of almost faithful prime ideals in
virtually nilpotent mod-p Iwasawa algebras
r
p
A
5
2
]
.
A
R
h
t
a
m
[
2
v
0
4
7
3
0
.
0
1
6
1
:
v
i
X
r
a
William Woods
November 5, 2018
Abstract
Let G be a nilpotent-by-finite compact p-adic analytic group for some
p > 2, and H = FN
p(G) its finite-by-(nilpotent p-valuable) radical. Fix
a finite field k of characteristic p, and write kG for the completed group
ring of G over k. We show that almost faithful G-stable prime ideals P
of kH extend to prime ideals P kG of kG.
2010 Mathematics Subject Classification: 16S34, 16D25, 16S35, 16W60.
1
Contents
1 p-valuations and crossed products
1.1 Preliminaries on p-valuations
. . . . . . . . . . . . . . . . . . . .
1.2 Ordered bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Separating a free abelian quotient . . . . . . . . . . . . . . . . . .
1.4
. . . . . . . . .
Invariance under the action of a crossed product
2 A graded ring
2.1 Generalities on ring filtrations . . . . . . . . . . . . . . . . . . . .
2.2 Constructing a suitable valuation . . . . . . . . . . . . . . . . . .
2.3 Automorphisms trivial on a free abelian quotient . . . . . . . . .
3 Extending prime ideals from FNp(G)
. . . . . . . . . . . . . . . . . . . . . . .
3.1 X-inner automorphisms
3.2 Properties of FNp(G)
. . . . . . . . . . . . . . . . . . . . . . . .
3.3 The extension theorem . . . . . . . . . . . . . . . . . . . . . . . .
6
6
8
10
12
16
16
18
24
27
27
28
30
2
Introduction
Let G be a nilpotent-by-finite compact p-adic analytic group and k a finite field
of characteristic p.
Recall the characteristic open subgroup H = FNp(G), the finite-by-(nilpotent
p-valuable) radical of G, defined in [10, Theorem C]. This plays an important
role in the structure of the group G: for instance, see the structure theorem [10,
Theorem D].
In this paper, we demonstrate a connection between certain prime ideals of kH
and those of kG. The main result of this paper is:
Theorem A. Fix some prime p > 2. Let G be a nilpotent-by-finite compact
p-adic analytic group, H = FNp(G), and k a finite field of characteristic p. Let
P be an almost faithful, G-stable prime ideal of kH. Then P kG is a prime ideal
of kG.
The proof (given in Propositions 3.7 and 3.8) comprises several technical ele-
ments, which we outline below.
First, let G be a nilpotent-by-finite compact p-adic analytic group with finite
radical ∆+ = 1 [10, Definition 1.2], and let H = FNp(G). Note that H is p-
valuable [6, III, 2.1.2], and that G acts on the set of p-valuations of H as follows:
if α is a p-valuation on H and g ∈ G, then we may define a new p-valuation g · α
on H by
g · α(x) = α(g−1xg).
(In fact, we do this in a slightly more general case, but the details are identical.
See Lemma 1.17 for the setup.)
Recall the definition of an isolated orbital (closed) subgroup L of H from [10,
Definition 1.4], and that normal subgroups are automatically orbital. We show
in Definition 1.2 that, if ω is a p-valuation on H and L is a closed isolated normal
subgroup of H, then ω induces a quotient p-valuation Ω on H/L. We also define
the (t, p)-filtration (actually a p-valuation) on a free abelian pro-p group A of
finite rank in Definition 1.4: this is a particularly "uniform" p-valuation on A,
analogous to the p-adic valuation vp on Zp.
It is now easy to show the following.
Theorem B. With the above notation: let L be a proper closed isolated normal
subgroup of H containing the commutator subgroup [H, H]. Then there exists
a p-valuation ω on H with the following properties:
(i) ω is G-invariant,
(ii) there exists a real number t > (p − 1)−1 such that ωL > t, and the
quotient p-valuation induced by ω on G/L is the (t, p)-filtration.
3
Continue to take G to be a nilpotent-by-finite compact p-adic analytic group
with ∆+ = 1, and H = FNp(G). Let p be a prime, k a field of characteristic
p, and P a faithful prime ideal of kH.
It is shown in [2, Theorem 8.4] that
P = pkH for some prime ideal p of kZ, where Z is the centre of H; and,
furthermore, in [2, proof of Theorem 8.6], that there exist an integer e and a
ring filtration f on kH/P such that
grf (kH/P ) ∼= (grv(kZ/p))[Y1, . . . , Ye]
(†)
where v = f kZ/p is a valuation, and grf (kZ/p) is a commutative domain. The
valuation f is partly constructed using the p-valuation ω on the group H.
Our next theorem is an extension of this result. Write f1 for the ring filtration
constructed above. Suppose now that P is G-stable so that we may consider
the ring kG/P kG, and fix a crossed product decomposition (kH/P ) ∗ F of this
ring, where F = G/H.
Write Q′ for the classical ring of quotients of kZ/p. The filtration f1 restricts
to a valuation on kZ/p, which extends naturally to a valuation on Q′, which
we will call v1. F acts on the set of valuations of Q′, and v1 has some orbit
{v1, . . . , vs}.
Write Q for a certain partial ring of quotients of kH/P containing Q′. We may
naturally form Q ∗ F as an overring of (kH/P ) ∗ F .
Theorem C. In the above notation: there exists a filtration f on Q ∗ F such
that
(i) gr f (Q∗F ) ∼= gr f (Q)∗F , where the right-hand side is some crossed product,
(ii) gr f (Q) ∼=Ls
i=1 grfi (Q),
(iii) grfi (Q) ∼= (grvi Q′)[Y1, . . . , Ye] for all 1 ≤ i ≤ s,
where s and e are determined as in (†), and the action of F in the crossed
product of (i) permutes the s summands in the decomposition of (ii) transitively
by conjugation.
We combine Theorems B and C as follows.
Theorem C, of course, only invokes (†) in the case when Q 6= Q′, so we suppose
that we are in this case, which occurs precisely when H is non-abelian. Take
L to be the smallest closed isolated normal subgroup of H containing both the
isolated derived subgroup H ′ [10, Theorem B] and the centre Z of H. Now L is
a proper subgroup by Lemma 3.5, and we will choose ω for L as in Theorem B.
We may arrange it in (†) and Theorem C so that, for some l ≤ e, the elements
Y1, . . . , Yl correspond to a Zp-module basis x1, . . . , xl for H/L; and here the
value of the filtration f can be understood in terms of the p-valuation ω. We
show that:
4
Theorem D. Take an automorphism σ of H. Suppose that the induced auto-
morphism on gr f (Q ∗ F ) fixes each of the valuations v1, . . . , vs and fixes each
of the elements Y1, . . . , Yl. Then the induced automorphism on H/L (which
can be seen as a matrix Mσ ∈ GLl(Zp)) lies in the first congruence subgroup of
GLl(Zp), i.e. it takes the form Mσ ∈ 1+pX for some X ∈ Ml(Zp). In particular,
when p > 2, σ has finite order if and only if it is the identity automorphism.
A special case of Theorem A, in which ∆+ = 1, is now deduced from Theo-
rem D via a long but elementary argument about X-inner automorphisms: see
Definition 3.1 and Proposition 3.7 for details.
The case when ∆+ 6= 1 now follows as a consequence of the "untwisting" results
of [11, Theorems B and C], which allow us to understand the prime ideals of
kH, along with the conjugation action of G, in terms of the corresponding
information for k′[[H/∆+]] (for various finite field extensions k′/k). Now, as
∆+(G/∆+) = 1 and H/∆+ = FNp(G/∆+), we are back in the previous case.
See Proposition 3.8 for details.
5
1 p-valuations and crossed products
1.1 Preliminaries on p-valuations
Definition 1.1. Recall from [6, III, 2.1.2] that a p-valuation on a group G is a
function ω : G → R ∪ {∞} satisfying:
• ω(xy−1) ≥ min{ω(x), ω(y)} for all x, y ∈ G
• ω([x, y]) ≥ ω(x) + ω(y) for all x, y ∈ G
• ω(x) = ∞ if and only if x = 1
• ω(x) > 1
p−1 for all x ∈ G
• ω(xp) = ω(x) + 1 for all x ∈ G.
Throughout this paper we will often be considering several p-valuations admit-
ted by a group G, so to clarify we may refer to G together with a p-valuation ω
as the p-valued group (G, ω) (though when the p-valuation in question is clear
from context, we will simply write G).
Given a p-valuation ω on a group G, we may write
Gω,λ := Gλ := ω−1([λ, ∞]),
Gω,λ+ := Gλ+ := ω−1((λ, ∞])
and define the graded group
grωG :=Mλ∈R
Gλ/Gλ+ .
Then each element 1 6= x ∈ G has a principal symbol
grω(x) := xGµ+ ∈ Gµ/Gµ+ ≤ grωG,
where µ is defined such that µ = ω(x).
Remark. Let (G, ω) be a p-valued group, and N an arbitrary subgroup of G.
Then (N, ωN ) is p-valued. Moreover, if G has finite rank [6], then so does N ;
and if G is complete with respect to ω and N is a closed subgroup of G, then
N is complete with respect to ωN .
Definition 1.2. Given an arbitrary complete p-valued group (G, ω) of finite
rank, and a closed isolated normal subgroup K (i.e. a closed normal subgroup
K such that G/K is torsion-free), we may define the quotient p-valuation Ω
induced by ω on G/K as follows:
Ω(gK) = sup
k∈K
{ω(gk)}.
6
This is defined by Lazard, but the definition is spread across several results, so we
collate them here for convenience. The definition in the case of filtered modules
is [6, I, 2.1.7], and is modified to the case of filtered groups in [6, the remark
after II, 1.1.4.1]. The specialisation from filtered groups to p-saturable groups
is done in [6, III, 3.3.2.4], where it is proved that Ω is indeed still a p-valuation
on G/K; and the general case is stated in [6, III, 3.1.7.6], and eventually proved
in [6, IV, 3.4.2].
As a partial inverse to the above process of passing to a quotient p-valuation, we
prove the following general result about "lifting" p-valuations from torsion-free
quotients.
Theorem 1.3. Let G be a complete p-valued group of finite rank, and N a
closed isolated orbital (hence normal) subgroup of G. Suppose we are given two
functions
α, β : G → R ∪ {∞},
such that α is a p-valuation on G, and β factors through a p-valuation on G/N ,
i.e.
β : G/N → R ∪ {∞}.
Then ω = inf{α, β} is a p-valuation on G.
Proof. α and β are both filtrations on G (in the sense of [6, II, 1.1.1]), and so
by [6, II, 1.2.10], ω is also a filtration. Following [6, III, 2.1.2], for ω to be a
p-valuation, we need to check the following three conditions:
(i) ω(x) < ∞ for all x ∈ G, x 6= 1.
This follows from the fact that α is a p-valuation, and hence α(x) < ∞
for all x ∈ G, x 6= 1.
(ii) ω(x) > (p − 1)−1 for all x ∈ G.
This follows from the fact that α(x) > (p − 1)−1 and β(x) > (p − 1)−1 for
all x ∈ G by definition.
(iii) ω(xp) = ω(x) + 1 for all x ∈ G.
Take any x ∈ G. As α is a p-valuation, we have by definition that
α(xp) = α(x) + 1.
If x ∈ N , this alone is enough to establish the condition, as ωN = αN
(since β(x) = ∞).
Suppose instead that x ∈ G \ N . Then, as N is assumed isolated orbital
in G, we also have xp ∈ G \ N , so by definition of β we have
β(xp) = β((xN )p) = β(xN ) + 1 = β(x) + 1,
with the middle equality coming from the fact that β is a p-valuation.
Now it is clear that ω(xp) = ω(x) + 1 by definition of ω.
7
Finally, the following function will be crucial. (It is in fact a p-valuation, but
we delay the proof of this fact until Lemma 1.6.)
Definition 1.4. Let A be a free abelian pro-p group of rank d > 0 (here written
multiplicatively). Choose a real number t > (p − 1)−1. Then the (t, p)-filtration
on A is the function ω : A → R ∪ {∞} defined by
ω(x) = t + n,
where n is the non-negative integer such that x ∈ Apn
ω(1) = ∞.)
\ Apn−1
. (By convention,
1.2 Ordered bases
Definition 1.5. Recall from [2, 4.2] that an ordered basis for a p-valued group
(G, ω) is a set {g1, . . . , ge} of elements of G such that every element x ∈ G can
be uniquely written as the (ordered) product
for some λi ∈ Zp, and
x = Y1≤i≤e
gλi
i
ω(x) = inf
1≤i≤e
{ω(gi) + vp(λi)},
where vp is the usual p-adic valuation on Zp. (Note that an ordered basis for
(G, ω) need not be an ordered basis for (G, ω′) for another p-valuation ω′.)
As in [2], we will often write
gλ := Y1≤i≤e
gλi
i
as shorthand, where λ ∈ Ze
p.
We now show that the function given in Definition 1.4 is indeed a p-valuation,
and demonstrate some of its properties.
Lemma 1.6. Let A and t be as in Definition 1.4.
(i) The (t, p)-filtration ω is a p-valuation on A.
(ii) Suppose we are given a Zp-module basis B = {a1, . . . , ad} for A, and a
p-valuation α on A satisfying α(a1) = · · · = α(ad) = t. Then α is the
(t, p)-filtration on A, and B is an ordered basis for (A, α).
(iii) The (t, p)-filtration ω is completely invariant under automorphisms of A,
i.e. the subgroups Aω,λ and Aω,λ+ are characteristic in A.
Proof.
8
(i) This is a trivial check from the definition [6, III, 2.1.2].
(ii) By [6, III, 2.2.4], we see that
α(aλ1
1 . . . aλd
d ) = t + inf
1≤i≤d
{vp(λi)},
which is precisely the (t, p)-filtration.
(iii) The subgroups Apn
are clearly characteristic in A.
Remark. The (t, p)-filtration as defined here is equivalent to the definition given
in [6, II, 3.2.1] for free abelian pro-p groups of finite rank.
Recall from [10, Definitions 1.1 and 1.4] that a closed subgroup H of a profinite
group G is (G-)orbital if it has finitely many G-conjugates, and isolated orbital
if any G-orbital H ′ (cid:13) H satisfies [H ′ : H] = ∞.
The following is a general property of ordered bases.
Lemma 1.7. Let (G, ω) be a complete p-valued group of finite rank, and N
a closed isolated normal subgroup of G. Then there exist sets BN ⊆ BG such
that BN is an ordered basis for (N, ωN ) and BG is an ordered basis for (G, ω).
Proof. This was established in [2, proof of Lemma 8.5(a)].
Remark. It may be helpful to think of this as follows:
BG =(cid:8) x1, . . . , xr
}
{z
BG/N
BN
{z
, xr+1, . . . , xs
(cid:9),
}
where BG/N = BG \BN is in fact some appropriate preimage in G of any ordered
basis for (G/N, Ω), where Ω is the quotient p-valuation.
Lemma 1.8. Let (G, α) be a complete p-valued group of finite rank, and N a
closed isolated orbital (hence normal) subgroup of G. Take also a p-valuation β
on G/N . Suppose we are given sets
BG =(cid:8) x1, . . . , xr
{z
}
BG/N
, xr+1, . . . , xs
BN
{z
}
(cid:9),
such that
• BN is an ordered basis for (N, αN ),
• BG is an ordered basis for (G, α), and
• the image in G/N of BG/N is an ordered basis for (G/N, β).
In the notation of Theorem 1.3, write β for the composite of G → G/N with β,
and form the p-valuation ω = inf{α, β} for G.
Then BG is an ordered basis for (G, ω).
9
Proof. We need only check that
ω(xλ) = inf
1≤i≤s
{ω(xi) + vp(λi)}
for any λ ∈ Zs
p. But we have by definition that
α(xλ) = inf
1≤i≤s
{α(xi) + vp(λi)},
β(xλ) = inf
1≤i≤r
{β(xi) + vp(λi)},
and the result follows trivially.
1.3 Separating a free abelian quotient
Results in later sections will require the existence of a p-valuation on an appro-
priate group G satisfying a certain technical property, which we can now finally
state:
Definition 1.9. Let (G, ω) be a complete p-valued group of finite rank, and
L a closed isolated normal subgroup containing [G, G] (and hence containing
the isolated derived subgroup G′, which was defined and written as G(1) in [10,
Theorem B]). We will say that ω satisfies property (AL) if there is an ordered
basis {gd+1, . . . , ge} for (L, ωL), contained in an ordered basis {g1, . . . , ge} for
(G, ω) (e.g. constructed by Lemma 1.7), such that the following hold:
and, for all ℓ ∈ L,
ω(g1) = · · · = ω(gd),
ω(g1) < ω(ℓ). )
(AL)
Remark. In the notation of the above definition, suppose that ω satisfies ω(g1) < ω(ℓ)
for all ℓ ∈ L. Then, by our earlier remarks, we note that the condition (AL) is
equivalent to the statement that the quotient p-valuation induced by ω on G/L
is the (t, p)-filtration for t := ω(g1).
Definition 1.10. Following [6, III, 2.1.2], we will say that a group G is p-
valuable if there exists a p-valuation ω for G, and G is complete with respect to
ω and has finite rank.
Lemma 1.11. Let G be a nilpotent p-valuable group, and L a closed isolated
normal subgroup containing G′. Then there exists some p-valuation ω for G
satisfying (AL).
Proof. Let α be a p-valuation on G. Take an ordered basis {gd+1, . . . , ge} for
(L, αL) and extend it to an ordered basis {g1, . . . , ge} for (G, α) by Lemma 1.7.
Fix a number t satisfying
(p − 1)−1 < t ≤ inf
1≤i≤e
α(gi).
10
Applying Theorem 1.3 with N = L and β the (t, p)-filtration on G/L, we see
that ω = inf{α, β} is a p-valuation for G; and by Lemma 1.8, {g1, . . . , ge} is
still an ordered basis for (G, ω), so we can check easily that ω satisfies (AL) by
construction.
Remark. In fact, by analysing the construction in Theorem 1.3, we can see that
we have shown something stronger: that any p-valuation α may be refined to
such an ω satisfying ωL = αL. But we will not use this fact in this paper.
Remark. If ω satisfies (AL) as above, write t := ω(g1). Then, for any automor-
phism σ of G and any 1 ≤ i ≤ d, we have
ω(σ(gi)) = t.
This follows from Lemma 1.6(iii). Indeed, by construction, we have Gt = G,
and Gt+ = Gp · L, an open normal subgroup; and when L is characteristic, Gt+
is characteristic.
Now let G be a p-valuable group with fixed p-valuation ω, and let σ ∈ Aut(G).
In this subsection and the next, we seek to establish conditions under which a
given automorphism σ of G will preserve the "dominant" part of certain elements
x ∈ G (with respect to ω). That is, we are looking for a condition under which
Clearly it is necessary and sufficient that the following holds:
grω(σ(x)) = grω(x).
ω(σ(x)x−1) > ω(x).
(1.1)
The results of this paper rely on our ability to invoke the following technical
result.
Theorem 1.12. Let G be a p-valuable group, and let L be a proper closed
isolated orbital (hence normal) subgroup containing [G, G], so that we have an
isomorphism ϕ : G/L → Zd
p for some d ≥ 1. Write q : G → G/L for the natural
quotient map.
Choose a Zp-basis {e1, . . . , ed} for Zd
p. For each 1 ≤ i ≤ d, fix an element gi ∈ G
with ϕ◦ q(gi) = ei. Fix an automorphism σ of G preserving L, so that σ induces
an automorphism σ of G/L, and hence an automorphism σ = ϕ ◦ σ ◦ ϕ−1 of Zd
p.
Let Mσ be the matrix of σ with respect to the basis {e1, . . . , ed}.
Suppose there exists some p-valuation ω on G with the following properties:
(i) (1.1) holds for all x ∈ {g1, . . . , gd},
(ii) ω(g1) = · · · = ω(gd)(= t, say),
(iii) ω(ℓ) > t for all ℓ ∈ L.
Then Mσ − 1 ∈ pMd(Zp).
11
Remark. Conditions (ii) and (iii) are just the statement that ω satisfies (AL).
Proof. Define the function Ω : Zd
p → R ∪ {∞} by
Ω ◦ ϕ(gL) = sup
ℓ∈L
{ω(gℓ)}.
By the remarks made in Definition 1.2, Ω is in fact a p-valuation.
By assumption (iii), we see that, for each 1 ≤ i ≤ d and any ℓ ∈ L, we have
ω(gi) = ω(giℓ), so that
Ω(ei) = Ω ◦ ϕ(giL) = sup
ℓ∈L
{ω(giℓ)} = ω(gi),
so by assumption (ii), Ω(ei) = t. Hence, by Lemma 1.6(ii), Ω must be the
(t, p)-filtration on Zd
p. Now, by assumption (i), we have
Ω(σ(x) − x) > t
for all x ∈ {e1, . . . , ed}, and hence, as Ω − t takes integer values (by Definition
1.4),
Ω(σ(x) − x) ≥ t + 1,
and so σ(x) − x ∈ pZd
prove.
p for each x ∈ {e1, . . . , ed}, which is what we wanted to
Remark. With Lemma 1.6 in mind, we note the following: suppose ω satis-
fies hypothesis (iii) of Theorem 1.12. Then hypothesis (ii) is equivalent to the
statement that the quotient filtration induced by ω on G/L is actually the
(t, p)-filtration on G/L.
1.4
Invariance under the action of a crossed product
Definition 1.13. Let R be a ring, and fix a subgroup G ≤ R×; let F be a
group. Fix a crossed product
S = R ∗
hσ,τ i
F.
Consider the following properties that this crossed product may satisfy:
The image σ(F ) normalises G, i.e. xσ(f ) ∈ G for all x ∈ G, f ∈ F .
The image τ (F, F ) normalises G.
The image τ (F, F ) is a subset of G.
(NG)
(N′
G)
(PG)
In the case when G is p-valuable, consider the set of p-valuations of G. Then
Aut(G) acts on this set as follows:
(ϕ · ω)(x) = ω(xϕ).
12
When S satisfies (N′
G), τ (F, F ) ⊆ G, so we get a map ρ : τ (F, F ) → Inn(G)
(with elements of G acting by conjugation), so we will also consider the following
property:
Every p-valuation ω of G is invariant under elements of τ (F, F ).
(QG)
Lemma 1.14. In the notation above:
(i) If S satisfies (NG), then S satisfies (N′
(ii) If S satisfies (PG), then S satisfies (N′
G).
G).
(iii) If S satisfies (PG), then S satisfies (QG).
Proof.
(i) Note that ρ ◦ τ (x, y) = σ(xy)−1σ(x)σ(y).
(ii) Obvious.
(iii) By (ii), we see that S satisfies (N′
G), so it makes sense to consider (QG).
Let ω be a p-valuation of G, and take t ∈ τ (F, F ). As S satisfies (PG), we
actually have t ∈ G. Then, for any x ∈ G, we have
(t · ω)(x) = ω(xt)
= ω(t−1xt)
= ω(x · [x, t])
≥ min{ω(x), ω([x, t])} = ω(x),
and so (by symmetry) ω(t−1xt) = ω′(x).
Definition 1.15. Recall, from [11, Definition 5.4], that if we have a fixed crossed
product
S = R ∗
hσ,τ i
F
(1.2)
and a 2-cocycle
then we may define the ring
α ∈ Z 2
σ(F, Z(R×)),
Sα = R ∗
hσ,τ αi
F,
the 2-cocycle twist (of R, by α, with respect to the decomposition (1.2)).
Lemma 1.16. Continuing with the notation above,
(i) S satisfies (NG) if and only if Sα satisfies (NG).
13
(ii) S satisfies (QG) if and only if Sα satisfies (QG).
Proof.
(i) Trivial from Definitions 1.13 and 1.15.
(ii) As α(F, F ) ⊆ Z(R)×, conjugation by α is the identity automorphism on
G.
These properties will be interesting to us later as they will allow us to invoke
the following lemma:
Lemma 1.17. If S satisfies (NG), then, given any g ∈ F and p-valuation ω on
G, the function g · ω given by
(g · ω)(x) = ω(xσ(g))
is again a p-valuation on G. If, further, S satisfies (QG), then this is a group
action of F on the set of p-valuations of G.
Proof. If x ∈ G, then xσ(g) ∈ G because S satisfies (NG), so it makes sense to
consider ω(xσ(g)). The definition above does indeed give a group action when
S satisfies (QG), as, for all g, h ∈ F ,
(g · (h · ω))(x) = h · ω(xσ(g))
= ω(xσ(g)σ(h))
= ω(xσ(gh)τ (g,h))
= ω(xσ(gh))
= (gh · ω)(x).
by (QG)
The following lemma will allow us to prove the existence of a sufficiently "nice"
p-valuation.
Lemma 1.18. Suppose S satisfies (NG) and (QG), so that σ induces an action
of F on the set of p-valuations on G as in the above lemma. Let ω be a p-
If the F -orbit of ω is finite, then ω′(x) = inf g∈F (g · ω)(x)
valuation on G.
defines an F -invariant p-valuation on G.
Furthermore, if L is a closed isolated characteristic subgroup of G containing
G′, and ω satisfies (AL) (as in Definition 1.9), then ω′ satisfies (AL).
Proof. The function ω′ satisfies condition [6, III, 2.1.2.2], since the F -orbit of
ω is finite, and is hence a p-valuation that is F -stable by the remark in [6, III,
2.1.2].
Suppose ω satisfies (AL). That is, for some t > (p − 1)−1, ω induces the
(t, p)-filtration on G/L, and ω(ℓ) > t for all ℓ ∈ L. But, given any g ∈ F ,
14
clearly g · ω still induces the (t, p)-filtration on G/L by Lemma 1.6(iii), and
(g · ω)(ℓ) = ω(ℓσ(g)) > t, since ℓσ(g) ∈ L as L is characteristic. Taking the
infimum over the finitely many distinct g · ω, g ∈ F , shows that ω′ also satisfies
(AL).
Recall the finite radical ∆+ = ∆+(G) from [10, Definition 1.2].
Definition 1.19. Let G be an arbitrary compact p-adic analytic group with
∆+ = 1, H an open normal subgroup of G, F = G/H, and P a faithful G-
stable ideal of kH. Recall from [11, Definition 5.11] that the crossed product
decomposition
kG/P kG = kH/P ∗
hσ,τ i
F
is standard if the basis F is a subset of the image of the map G ֒→ (kG/P kG)×.
Lemma 1.20. Suppose that kG/P kG = kH/P ∗
hσ,τ i
F is a standard crossed
product decomposition. Take any α ∈ Z 2
tral 2-cocycle twist
σ(F, Z((kH/P )×)), and form the cen-
(kG/P kG)α := kH/P ∗
hσ,τ αi
F
with respect to this decomposition [11, Definition 5.4].
Consider H as a subgroup of (kH/P )×: then conjugation by elements of G
inside ((kG/P kG)α)× induces a group action of F on the set of p-valuations of
H, as in Lemma 1.17.
Remark. As the crossed product notation suggests, this lemma simply says that
the action of F on H, via σ, is unchanged after applying (−)α.
Proof. As the decomposition is standard, kG/P kG trivially satisfies both (NH )
(as H is normal in G) and (PH ). By Lemma 1.14(iii), kG/P kG also satisfies
(QH ). Now Lemma 1.16 shows that (kG/P kG)α also satisfies (NH ) and (QH ), so
that σ induces a group action of F on the p-valuations of H inside (kG/P kG)α
by Lemma 1.17.
Let L be a closed isolated characteristic subgroup of H containing [H, H].
Corollary 1.21. With notation as above, we can find an F -stable p-valuation
ω on H satisfying (AL).
Proof. This now follows immediately from Lemmas 1.11 and 1.18.
Proof of Theorem B. This follows from Corollary 1.21.
15
2 A graded ring
2.1 Generalities on ring filtrations
Definition 2.1. Recall that a filtration v on the ring R is a function v : R → R∪{∞}
satisfying, for all x, y ∈ R,
• v(x + y) ≥ min{v(x), v(y)},
• v(xy) ≥ v(x) + v(y),
• v(0) = ∞, v(1) = 0.
If in addition we have v(xy) = v(x) + v(y) for all x, y ∈ R, then v is a valuation
on R.
First, a basic property of ring filtrations.
Lemma 2.2. Suppose v is a filtration on R which takes non-negative values,
i.e. v(R) ⊆ [0, ∞], and let u ∈ R×. Then v(ux) = v(xu) = v(x) for all x ∈ R.
Proof. By the definition of v, we have 0 = v(1) = v(uu−1) ≥ v(u) + v(u−1). As
v(u) ≥ 0 and v(u−1) ≥ 0, we must have v(u) = 0 = v(u−1). Then
v(x) = v(u−1ux) ≥ v(u−1) + v(ux) = v(ux) ≥ v(u) + v(x) = v(x),
from which we see that v(x) = v(ux); and by a symmetric argument, we also
have v(xu) = v(x).
We will fix the following notation for this subsection.
Notation. Let G be a p-valuable group equipped with the fixed p-valuation
ω, and k a field of characteristic p. Take an ordered basis {g1, . . . , gd} for G,
and write bi = g1 − 1 ∈ kG for all 1 ≤ i ≤ d. As in [2], we make the following
definitions:
• for each α ∈ Nd, bα means the (ordered) product bα1
p, gα means the (ordered) product gα1
• for each α ∈ Zd
1 . . . bαd
1 . . . gαd
d ∈ kG,
d ∈ G,
• for each α ∈ Nd, hα, ω(g)i means
αiω(gi),
d
Xi=1
• the canonical ring homomorphism Zp → k will sometimes be left implicit,
but will be denoted by ι when necessary for clarity.
Definition 2.3. With notation as above, let w be the valuation on kG defined
in [2, 6.2], given by
Xα∈Nd
λαbα 7→ inf
α∈Nd(cid:8)hα, ω(g)i(cid:12)(cid:12) λα 6= 0(cid:9).
16
Note that, in light of this formula [2, Corollary 6.2(b)], and by the construc-
tion [6, III, 2.3.3] of w, it is clear that the value of w is in fact independent
of the ordered basis chosen.
In particular, if ϕ is an automorphism of G,
then {gϕ
d } is another ordered basis of G; hence if ω is ϕ-stable (in the
sense that ω(gϕ) = ω(g) for all g ∈ G), then w is ϕ-stable (in the sense that
w(x ϕ) = w(x) for all x ∈ kG, where ϕ here denotes the natural extension of ϕ
to kG, obtained by the universal property [11, Lemma 2.2]).
1 , . . . , gϕ
We will need the following result:
Lemma 2.4. Let
b = b0 + b1p + b2p2 + · · · ∈ Zp,
n = n0 + n1p + n2p2 + · · · + nsps ∈ N,
where all bi, ni ∈ {0, 1, . . . , p − 1}. Then
(cid:18)b
n(cid:19) ≡
s
Yi=0(cid:18)bi
ni(cid:19) mod p.
Proof. See e.g. [1, Theorem].
Corollary 2.5. Let b ∈ Zp, n ∈ N. If
vp(cid:18)(cid:18)b
n(cid:19)(cid:19) = 0,
(2.1)
then vp(b) ≤ vp(n). Further, for fixed b ∈ Zp,
Proof. From Lemma 2.4 above, we can see that
n(cid:19)(cid:19) = 0(cid:27) = pvp(b).
vp(cid:18)(cid:18)b
inf(cid:26)n ∈ N(cid:12)(cid:12)(cid:12)(cid:12)
(cid:18)b
n(cid:19) ≡ 0 mod p
if and only if, for some 0 ≤ i ≤ s,
ni(cid:19) = 0,
(cid:18)bi
which happens if and only if one of the pairs (bi, ni) for 0 ≤ i ≤ s has bi = 0 6= ni.
Hence, to ensure that this does not happen, we must have vp(b) ≤ vp(n). It is
clear from Lemma 2.4 that n = pvp(b) satisfies (2.1), and is the least n ∈ N with
vp(b) ≤ vp(n).
Theorem 2.6. Take any x ∈ G, and t = inf ω(G). Then w(x − 1) > t implies
ω(x) > t.
17
Proof. Write x = gα. In order to show that ω(gα) > t, it suffices to show that
ω(gj) + vp(αj) > t for each j (as there are only finitely many), and hence that
vp(αj ) ≥ 1 for all j such that ω(gj) = t. This is equivalent to the claim that
pvp(αj ) > 1, which we will write as pvp(αj )ω(gj) > t for all j with ω(gj) = t.
Let β(j) be the d-tuple with ith entry δij pvp(αj ). Then, of course,
hβ(j), ω(g)i = pvp(αj )ω(gj),
and by Corollary 2.5, we have
(cid:18) α
β(j)(cid:19) 6≡ 0 mod p.
Now suppose that w(gα − 1) > t. We perform binomial expansion in kG to see
that
(1 + bi)αi − 1
gα − 1 = Y1≤i≤d
ι(cid:18)α
β(cid:19)bβ − 1
= Xβ∈Nd
ι(cid:18)α
β(cid:19)bβ,
=Xβ6=0
so that
(ordered product)
β 6= 0,(cid:18)α
β(cid:19) 6≡ 0 mod p(cid:27) .
w(gα − 1) = inf(cid:26)hβ, ω(g)i(cid:12)(cid:12)(cid:12)(cid:12)
So in particular, for all j satisfying ω(gj) = t, we have
t < w(gα − 1) ≤ hβ(j), ω(g)i = pvp(αj )ω(gj),
which is what we wanted to prove.
2.2 Constructing a suitable valuation
Let H be a nilpotent p-valuable group with centre Z. If k is a field of charac-
teristic p, and p is a faithful prime ideal of kZ, then by [2, Theorem 8.4], the
ideal P := pkH is again a faithful prime ideal of kH.
We will fix the following notation for this subsection.
Notation. Let G be a nilpotent-by-finite compact p-adic analytic group, with
∆+ = 1, and let H = FNp(G) [10, Definition 5.3], here a nilpotent p-valuable
radical, so that ∆ = Z := Z(H) [10, proof of Lemma 1.2.3(iii)]. We will also
write F = G/H.
18
Define Q′ = Q(kZ/p), the (classical) field of fractions of the (commutative)
domain kZ/p, and Q = Q′ ⊗
kH, a tensor product of kZ-algebras, which (as
kZ
P = pkH) we may naturally identify with the (right) localisation of kH/P with
respect to (kZ/p) \ {0} -- a subring of the Goldie ring of quotients Q(kH/P ).
Suppose further that the prime ideal p ✁ kZ is invariant under conjugation by
elements of G.
Choose a crossed product decomposition
kG/P kG = kH/P ∗
hσ,τ i
F
which is standard in the sense of the notation of Corollary 1.21. Choose also
σ(F, Z((kH/P )×)), and form as in [11, Definition 5.4] the central
any α ∈ Z 2
2-cocycle twist
(kG/P kG)α = kH/P ∗
hσ,τ αi
F.
Now the (right) divisor set (kZ/p) \ {0} is G-stable by assumption, so by [8,
Lemma 37.7], we may define the partial quotient ring
R := Q ∗
hσ,τ αi
F.
(2.2)
Our aim in this subsection is to construct an appropriate filtration f on the ring
R. We will build this up in stages, following [2]. First, we define a finite set of
valuations on Q′.
Definition 2.7. In [2, Theorem 7.3], Ardakov defines a valuation on Q(kH/P );
let v1 be the restriction of this valuation to Q′, so that v1(x + p) ≥ w(x) for all
x ∈ kZ (where w is as in Definition 2.3).
Lemma 2.8. σ induces a group action of F on the set of valuations of Q′.
Proof. Let u be a valuation of Q′. G acts on the set of valuations of Q′ as
follows:
(g · u)(x) = u(g−1xg).
Clearly, if g ∈ H, then g−1xg = x (as x ∈ Q(kZ/p) where Z is the centre of
H). Hence H lies in the kernel of this action, and we get an action of F on the
set of valuations. By our choice of F as a subset of the image of G, this is the
same as σ.
Write {v1, . . . , vs} for the F -orbit of v1.
Lemma 2.9. The valuations v1, . . . , vs are independent.
Proof. The vi are all non-trivial valuations with value groups equal to subgroups
of R by definition. Hence, by [3, VI.4, Proposition 7], they have height 1.
19
They are also pairwise inequivalent:
indeed, suppose vi is equivalent to g · vi
for some g ∈ F . Then by [3, VI.3, Proposition 3], there exists a positive real
number λ with vi = λ(g · vi), and so vi = λn(gn · vi) (as the actions of λ and
g commute) for all n. But F is a finite group: so, taking n = o(g), we get
vi = λnvi. As vi is non-trivial, we must have that λn = 1, and so λ = 1. So
we may conclude, from [3, VI.4, Proposition 6(c)], that the valuations v1, . . . , vs
are independent.
Definition 2.10. Let v be the filtration of Q′ defined by
v(x) = inf
1≤i≤s
vi(x)
for each x ∈ Q′.
Lemma 2.11. grvQ′ ∼=
Proof. The natural map
grvi Q′.
s
Mi=1
Q′
v,λ →
s
Mi=1
Q′
vi,λ/Q′
vi,λ+
Q′
vi,λ+ = Q′
v,λ+ , giving an injective map grvQ′ →
grvi Q′.
s
Mi=1
clearly has kernel
s
\i=1
The surjectivity of this map now follows from the Approximation Theorem [3,
VI.7.2, Th´eor`eme 1], as the vi are independent by Lemma 2.9.
Next, we will extend the vi and v from Q′ to Q, as in the proof of [2, 8.6].
Notation. Continue with the notation above. Now, H is p-valuable, and by
Lemma 1.20, F acts on the set of p-valuations of H; hence, by Lemma 1.18 (or
Corollary 1.21), we may choose a p-valuation ω which is F -stable. Fix such an
ω, and construct the valuation w on kH from it as defined in Definition 2.3.
Let {ye+1, . . . , yd} be an ordered basis for Z, and extend it to an ordered basis
{y1, . . . , yd} for H as in Lemma 1.7 (noting that Z is a closed isolated normal
subgroup of H by [2, Lemma 8.4(a)]). For each 1 ≤ j ≤ e, set cj = yj − 1 inside
the ring kH/P .
Recall from [2, 8.5] that elements of Q may be written uniquely as
rγ cγ ,
Xγ∈Ne
where rγ ∈ Q′ and cγ := cγ1
module.
1 . . . cγe
e , so that Q ⊆ Q′[[c1, . . . , ce]] as a left Q′-
20
Definition 2.12. For each 1 ≤ i ≤ s, as in [2, proof of Theorem 8.6], we will
define the valuation fi : Q → R ∪ {∞} by
rγ cγ
fi
Xγ∈Ne
γ∈Ne
= inf
{vi(rγ) + w(cγ )}.
(We remark here a slight abuse of notation: the domain of w is kH, and so
w(cγ ) must be understood to mean w(bγ ), where bj = yj − 1 inside the ring
kH for each 1 ≤ j ≤ e. That is, bj is the "obvious" lift of cj from kH/P to kH.
This relies on the assumption that P is faithful.)
Note in particular that fiQ′ = vi, and grfi Q is a commutative domain, again
by [2, proof of Theorem 8.6].
Lemma 2.13. σ induces a group action of F on the set of valuations of Q.
Proof. Let u be a valuation of Q. Again, G acts on the set of valuations of Q
by (g · u)(x) = u(g−1xg). Now, any n ∈ H can be considered as an element of
Q×, so that
(n · u)(x) = u(n−1xn) = u(n−1) + u(x) + u(n) = u(x).
In the following lemma, we crucially use the fact that ω has been chosen to be
F -stable.
Lemma 2.14. f1, . . . , fs is the F -orbit of f1.
Proof. Take some g ∈ F and some 1 ≤ i, j ≤ s such that vj = g · vi. We
will first show that, for all x ∈ Q, we have fj(x) ≤ g · fi(x).
Indeed, as
fjQ′ = vj = g · vi = g · fiQ′ , and both fj and g · fi are valuations, it will suffice
to show that (w(ck) =)fj(ck) ≤ g · fi(ck) for each 1 ≤ k ≤ e.
Fix some 1 ≤ k ≤ e. Write yg
k = zyα for some α ∈ Ze
p and z ∈ Z, so that
cg
k = yg
k − 1 = zyα − 1
(1 + ci)αi − 1!
= (z − 1) + z e
Yi=1
β(cid:19)cβ
= (z − 1) + z
ι(cid:18)α
Xβ6=0
,
(g · fi)(ck) = inf(cid:26)vi(z − 1), w(cβ )(cid:12)(cid:12)(cid:12)(cid:12)
ι(cid:18)α
β(cid:19) 6= 0(cid:27)
≥ inf(cid:26)w(z − 1), w(cβ )(cid:12)(cid:12)(cid:12)(cid:12)
ι(cid:18)α
β(cid:19) 6= 0(cid:27)
= w(cg
k),
and hence
21
(ordered product)
by Definition 2.12
by Definition 2.7
with this final equality following from [2, Lemma 8.5(b)]. But now, as ω has
been chosen to be G-stable, w is also G-stable (see the remark in Definition
2.3), so that w(cg
k) = w(ck).
Now, we have shown that, if vj = g · vi on Q′, then fj ≤ g · fi on Q.
Similarly, we have vi = g−1·vj on Q′, so fi ≤ g−1·fj on Q. But fi(x) ≤ fj(xg−1
)
for all x ∈ Q is equivalent to fi(yg) ≤ fj(y) for all y ∈ Q (by setting x = yg).
Hence we have fi = g · fj on Q, and we are done.
As in Definition 2.10:
Definition 2.15. Let f be the filtration of Q defined by
f (x) = inf
1≤i≤s
fi(x)
for each x ∈ Q.
We now verify that the relationship between f and v is the same as that between
the fi and the vi (Definition 2.12).
Lemma 2.16. Take any x ∈ Q, and write it in standard form as
rγ cγ.
x = Xγ∈Ne
Then we have
f (x) = inf
γ∈Ne
{v(rγ) + w(cγ )}.
Proof. Immediate from Definitions 2.10, 2.12 and 2.15.
Now we can extend Lemma 2.11 to Q:
Lemma 2.17. grf Q ∼=
s
Mi=1
grfi Q.
Proof. As in the proof of Lemma 2.11, we get an injective map
grf Q →
s
Mi=1
grfi Q.
The proof of [2, 8.6] gives a map
and isomorphisms
(grv(kZ/p))[Y1, . . . , Ye] → grf (kH/P )
(grvi(kZ/p))[Y1, . . . , Ye] ∼= grfi (kH/P )
22
for each 1 ≤ i ≤ s, in each case mapping Yj to gr(cj) for each 1 ≤ j ≤ e.
Now, gr kH is a gr-free [5, §I.4.1, p. 28] gr kZ-module with respect to f and
each fi, and each of these filtrations is discrete on kH by construction (see [2,
Corollary 6.2 and proof of Theorem 7.3]), so by [5, I.6.2(3)], kH is a filt-free kZ-
module with respect to f and each fi; and by [5, I.6.14], these maps extend to a
map (grvQ′)[Y1, . . . , Ye] → grf Q and isomorphisms (grvi Q′)[Y1, . . . , Ye] ∼= grfiQ
for each i.
Applying Lemma 2.14 to each 1 ≤ i ≤ s, we get isomorphisms
(grvi Q′)[Y1, . . . , Ye] → grfi Q,
which give a commutative diagram
(grvQ′)[Y1, . . . , Ye]
∼= /
s
Mi=1
(grvi Q′)[Y1, . . . , Ye]
grf Q
∼=
grfiQ.
s
Mi=1
Hence clearly all maps in this diagram are isomorphisms.
Now we return to the ring R = Q ∗ F defined in (3.2).
Definition 2.18. We can extend the filtration f on Q to an F -stable fil-
tration on R by giving elements of the basis F value 0. That is, writing
r=1 grxr
F = {g1, . . . , gm}, any element of Q ∗ F can be expressed uniquely asPm
for some xr ∈ Q: the assignment
Q ∗ F → R ∪ {∞}
m
Xr=1
grxr 7→ inf
1≤r≤mnf (xr)o
is clearly a filtration on Q∗F whose restriction to Q is just f . We will temporar-
ily refer to this filtration as f , though later we will drop the hat and simply call
it f .
Note that, for any real number λ,
(Q ∗ F ) f ,λ =
(Q ∗ F ) f ,λ+ =
gi(Qf,λ),
gi(Qf,λ+ ),
m
m
Mi=1
Mi=1
23
/
/
/
so that
gr f (Q ∗ F ) =Mλ∈R m
Mi=1
gi Mλ∈R
Mi=1
=
m
gi(Qf,λ/Qf,λ+)!
(Qf,λ/Qf,λ+)! =
m
Mi=1
gi(cid:0)grf (Q)(cid:1) .
That is, given the data of a crossed product Q ∗ F as in (3.2), we may view
gr f (Q ∗ F ) as grf (Q) ∗ F in a natural way.
We will finally record this as:
Lemma 2.19.
grf (Q ∗ F ) = grf (Q) ∗ F ∼= s
Mi=1
∼= s
Mi=1
grfiQ! ∗ F
(grvi Q′)[Y1, . . . , Ye]! ∗ F,
where each grfi Q (or equivalently each grvi Q′) is a domain (see Definition 2.12).
F permutes the fi (or equivalently the vi) transitively by conjugation.
Proof of Theorem C. This is Lemma 2.19.
2.3 Automorphisms trivial on a free abelian quotient
We will fix the following notation for this subsection.
Notation. Let H be a nilpotent but non-abelian p-valuable group with centre
Z. Write H ′ for its isolated derived subgroup [10, Theorem B]. Suppose we are
given a closed isolated proper characteristic subgroup L of H which contains
H ′ and Z. (We will show that such an L always exists in Lemma 3.5.) Fix a
p-valuation ω on H satisfying (AL) (which is possible by Corollary 1.21).
Let {gm+1, . . . , gn} be an ordered basis for Z. Using Lemma 1.7 twice, extend
this to an ordered basis {gl+1, . . . , gn} for L, and then extend this to an ordered
basis {g1, . . . , gn} for H. Diagrammatically:
, gl+1, . . . , gm
, gm+1, . . . , gn
BH =(cid:8) g1, . . . , gl
{z
}
BH/L
BL/Z
{z
}
BZ
{z
(cid:9),
}
extending the notation of the remark after Lemma 1.7 in the obvious way. Here,
0 < l ≤ m < n, corresponding to the chain of subgroups 1 (cid:12) Z ≤ L (cid:12) H.
Let k be a field of characteristic p. As before, let p be a faithful prime ideal of kZ,
so that P := pkH is a faithful prime ideal of kH, and write bj = gj − 1 ∈ kH/P
for all 1 ≤ j ≤ m.
24
In this subsection, we will write:
• for each α ∈ Nm, bα means the (ordered) product bα1
p , gα means the (ordered) product gα1
• for each α ∈ Zm
1 . . . bαm
1 . . . gαm
m ∈ H,
m ∈ kH/P ,
• for each α ∈ Nm, hα, ω(g)i means
αiω(gi).
m
Xi=1
x = zgα
Note the use of m rather than n in each case. This means that every element
x ∈ H may be written uniquely as
for some α ∈ Zm
uniquely as
p and z ∈ Z; and every element y ∈ kH/P may be written
y = Xγ∈Nm
rγ bγ
for some elements rγ ∈ kZ/p.
Recall the definitions of the filtrations w on kH (Definition 2.3), v on kZ/p
(Definition 2.10) and f on kH/P (Definition 2.15). We will continue to abuse
notation slightly for w, as in Definition 2.12.
Recall also that, as we have chosen ω to satisfy (AL), we have that
w(b1) = · · · = w(bl) < w(br)
for all r > l.
Let σ be an automorphism of H, and suppose that, when naturally extended to
an automorphism of kH, it satisfies σ(P ) = P . Hence we will consider σ as an
automorphism of kH/P , preserving the subgroup H ⊆ (kH/P )×.
Corollary 2.20. With the above notation, fix 1 ≤ i ≤ l. If f (σ(bi)−bi) > f (bi),
then w(σ(bi) − bi) > w(bi).
Proof. Write in standard form
σ(bi) − bi = Xγ∈Nm
rγ bγ,
for some rγ ∈ kZ, and suppose that f (σ(bi) − bi) > f (bi). That is, by Lemma
2.16,
v(rγ ) + w(bγ ) > w(bi)
for each fixed γ ∈ Nm.
We will show that w(rγ ) + w(bγ) > w(bi) for each γ. We deal with two cases.
Case 1: w(bγ ) > w(bi). Then, as w takes non-negative values on kH, we are
already done.
25
Case 2: w(bγ ) ≤ w(bi). Then, by (AL), we have either w(rγ ) > w(bi) or
w(rγ ) = 0. In the former case, we are done automatically, so assume we are in
the latter case and w(rγ ) = 0. Then, by [2, 6.2], rγ must be a unit in kZ, and
so f (rγ) = 0 by Lemma 2.2, a contradiction.
Hence w(rγ ) + w(bγ ) > w(bi) for all γ ∈ Nm. But, as w is discrete by [2, 6.2],
we may now take the infimum over all γ ∈ Nm, and the inequality remains
strict.
Let σ be an automorphism of H, and recall that H/L is a free abelian pro-
p group of rank l. Choose a basis e1, . . . , el for Zl
p; then the map giL 7→ ei
for 1 ≤ i ≤ l is an isomorphism j : H/L → Zl
p. As L is characteristic in
H by assumption, σ induces an automorphism of H/L, which gives a matrix
Mσ ∈ GLl(Zp) under this isomorphism.
Write
for the quotient p-valuation on H/L induced by ω, i.e.
ω : H/L → R ∪ {∞}
ω(xL) = sup
ℓ∈L
{ω(xℓ)}
-- note that this is just the (t, p)-filtration (Definition 1.4), as we have chosen ω
to satisfy (AL). Then write
Ω : Zl
p → R ∪ {∞}
for the map Ω = ω ◦ j−1, the (t, p)-filtration on Zl
the isomorphism j.
p corresponding to ω under
Remark. If x ∈ Zl
(t, p)-filtration.
p has Ω(x) ≥ t + 1, then x ∈ pZl
p, by the definition of the
We will write Γ(1) = 1 + pGLl(Zp) for the first congruence subgroup of GLl(Zp),
the open subgroup of GLl(Zp) whose elements are congruent to the identity
element modulo p.
Corollary 2.21. With the above notation, if f (σ(bi) − bi) > f (bi) for all
1 ≤ i ≤ l, then Mσ ∈ Γ(1).
Proof. We have, for all 1 ≤ i ≤ l,
f (σ(bi) − bi) > f (bi) =⇒ w(σ(bi) − bi) > w(bi)
by Corollary 2.20,
=⇒ ω(σ(gi)g−1
i
) > ω(gi)
by Theorem 2.6,
-- which is condition (1.1). Now we may invoke Theorem 1.12.
Corollary 2.22. Suppose now further that σ is an automorphism of H of finite
order. If p > 2 and f (σ(bi) − bi) > f (bi) for all 1 ≤ i ≤ l, then σ induces the
identity automorphism on H/L.
26
Proof. We have shown that Mσ ∈ Γ(1), which is a p-valuable (hence torsion-
free) group for p > 2 by [4, Theorem 5.2]; and if σ has finite order, then Mσ
must have finite order. So Mσ is the identity map.
Proof of Theorem D. This now follows from Corollaries 2.21 and 2.22.
Remark. When p = 2, Γ(1) is no longer p-valuable.
Example 2.23. Let p = 2, and let
H = hx, y, z [x, y] = z, [x, z] = 1, [y, z] = 1i
be the (2-valuable) Z2-Heisenberg group. Let σ be the automorphism sending
x to x−1, y to y−1 and z to z. Take L = hzi, and P = 0.
Write X = x − 1 ∈ kH/P , and likewise Y = y − 1 and Z = z − 1. Now,
σ(X) = σ(x) − 1 = x−1 − 1 = (1 + X)−1 − 1 = −X + X 2 − X 3 + . . . ,
and so σ(X)−X = X 2−X 3+. . . (as char k = 2). Hence f (σ(X)−X) = f (X 2) > f (X);
but
0
Mσ =(cid:18)−1
0 −1(cid:19) ∈ Γ(1, GL2(Z2)),
and in particular Mσ 6= 1.
3 Extending prime ideals from FNp(G)
3.1 X-inner automorphisms
Definition 3.1. We recall the notation of [11, §5]: given R a ring, G a group
and a fixed crossed product S of R by G, we will sometimes write the structure
explicitly as
S = R ∗
hσ,τ i
G,
where σ : G → Aut(R) is the action and τ : G × G → R× the twisting.
Furthermore, we say that an automorphism ϕ ∈ Aut(R) is X-inner if there exist
nonzero elements a, b, c, d ∈ R such that, for all x ∈ R,
axb = cxϕd,
where xϕ denotes the image of x under ϕ. Write Xinn(R) for the subgroup of
Aut(R) consisting of X-inner automorphisms; and, given a crossed product as
in the previous paragraph, we will write XinnS(R; G) = σ−1(σ(G) ∩ Xinn(R)).
Lemma 3.2. R a prime ring and R∗G a crossed product. Let Ginn := XinnR∗G(R; G).
(i) If σ ∈ Aut(R) is X-inner, then σ is trivial on the centre of R.
27
(ii) If H is a subgroup of G containing Ginn, and R ∗ H is a prime ring, then
R ∗ G is a prime ring.
Proof.
(i) This follows from the description of X-inner automorphisms of R as re-
strictions of inner automorphisms of the Martindale symmetric ring of quo-
tients Qs(R), and the fact that Z(R) stays central in Qs(R): see [8, §12]
for details.
(ii) This follows from [8, Corollary 12.6]: if I is a nonzero ideal of R ∗ G, then
I ∩ R ∗ Ginn is nonzero, and hence I ∩ R ∗ H is nonzero.
3.2 Properties of FNp(G)
We prove here some important facts about the group FNp(G) (defined in [10,
Theorem C]).
Lemma 3.3. Let G be a nilpotent-by-finite compact p-adic analytic group with
∆+ = 1. Let H = FNp(G), and write
K := KG(H) = {x ∈ G [H, x] ≤ H ′},
where H ′ denotes the isolated derived subgroup of H [10, Theorem B]. Then
K = H.
Proof. Firstly, note that K clearly contains H, by definition of H ′.
Secondly, suppose that H is p-saturable. By the same argument as in [10,
Lemma 4.3], K acts nilpotently on H, and so K acts nilpotently on the Lie
algebra h associated to H under Lazard's isomorphism of categories [6]. That
is, we get a group representation Ad : K → Aut(h), and (Ad(k) − 1)(hi) ⊆ hi+1
for each k ∈ K and each i. (Here, hi denotes the ith term in the lower central
series for h.)
Choosing a basis for h adapted to the flag
h ) h2 ) · · · ) hr = 0,
we see that Ad is a representation of K for which Ad(k) − 1 is strictly upper
triangular for each k ∈ K; in other words, Ad : K → U, where U is a closed
subgroup of some GLn(Zp) consisting of unipotent upper-triangular matrices.
Hence the image Ad(K) is nilpotent and torsion-free.
Furthermore, ker Ad is the subgroup of K consisting of those elements k which
centralise h, and therefore centralise H. This clearly contains Z(H). On the
other hand, if k centralises H, then k is centralised by H, an open subgroup of
G, and so k must be contained in ∆. But ∆ = Z(H) by [10, Lemma 5.1(ii)]
28
Hence K is a central extension of two nilpotent, torsion-free compact p-adic
analytic groups of finite rank, and so is such a group itself; hence K is nilpotent
p-valuable by [10, Lemma 2.3], and so must be contained in H by definition of
FNp(G).
Now suppose H is not p-saturable, and fix a p-valuation on H. Conjugation
by k ∈ K induces the trivial automorphism on H/H ′, so by [6] it does also
on Sat (H/H ′), which is naturally isomorphic to Sat H/(Sat H)′ by [10, Lemma
3.2]. This shows that K ⊆ KG(Sat H). But now, writing h for the Lie algebra
associated to Sat H, the same argument as above, mutatis mutandis, shows that
KG(Sat H) = H.
Some properties.
Lemma 3.4. Let G be a compact p-adic analytic group with ∆+ = 1, and
write H = FNp(G). If H is not abelian, then H/Z = FNp(G/Z).
Proof. H/Z is a nilpotent p-valuable open normal subgroup of G/Z, so must be
contained within FNp(G/Z). Conversely, the preimage in G of FNp(G/Z) is
a central extension of Z by FNp(G/Z), two nilpotent and torsion-free groups,
and hence is nilpotent and torsion-free, so must be p-valuable by [10, Lemma
2.3], which shows that it must be contained within H.
Recall that, if J is a closed isolated subgroup of H, then there exists a unique
smallest isolated orbital subgroup of H containing J, which we call its isolator,
and denote iH (J), as in [10, Definition 1.6].
The (closed, isolated orbital, characteristic) subgroup iH (H ′Z) of H = FNp(G)
will be crucial throughout this section, so we record some results.
Lemma 3.5. Let H be a nilpotent p-valuable group. If H is not abelian, then
H 6= iH (H ′Z).
Proof. Suppose first that H is p-saturated, and write h and z for the Lie algebras
of H and Z respectively under Lazard's correspondence [6]. If h = h2z (writing
h2 for the second term in the lower central series of h), then by applying [h, −] to
both sides, we see that h2 = h3. But as h is nilpotent, this implies that h2 = 0,
so that h is abelian, a contradiction.
When H is not p-saturated: note that iH (H ′Z) = Sat(H ′Z) ∩ H, by [10,
Lemma 3.1], and so that Sat(H/iH (H ′Z)) ∼= Sat(H)/Sat(H ′Z) by Lemma [10,
Lemma 3.2]. Hence H/iH(H ′Z) has the same (in particular non-zero) rank as
Sat(H)/Sat(H ′Z).
Lemma 3.6. Let G be a nilpotent-by-finite compact p-adic analytic group with
∆+ = 1. Let H = FNp(G), and assume that H is not abelian. Write
M := MG(H) = {x ∈ G [H, x] ≤ iH (H ′Z)},
29
where H ′ denotes the isolated derived subgroup of H, and Z the centre of H.
Then M = H.
Proof. Clearly Z ≤ M . We will calculate M/Z.
First, note that iH (H ′Z)/Z is an isolated normal subgroup of H/Z, as the
quotient is isomorphic to H/iH(H ′Z), which is torsion-free. Also, as iH (H ′Z)
contains H ′Z and hence [H, H]Z as an open subgroup, clearly iH (H ′Z)/Z con-
tains [H, H]Z/Z as an open subgroup, so that iH (H ′Z)/Z ≤ iH/Z ([H, H]Z/Z).
Now, [H/Z, H/Z] = [H, H]Z/Z as abstract groups, so by taking their closures
followed by their (H/Z)-isolators, we see that
so that
(H/Z)′ = iH/Z(cid:0)[H, H]Z/Z(cid:1) = iH/Z(cid:0)[H, H]Z/Z(cid:1),
iH (H ′Z)/Z = (H/Z)′.
But x ∈ M if and only if [H, x] ≤ iH (H ′Z), which is equivalent to [H/Z, xZ] ≤ (H/Z)′,
or in other words xZ ∈ KG/Z(H/Z) = H/Z by Lemma 3.3. So M/Z = H/Z,
and hence M = H.
3.3 The extension theorem
Proposition 3.7. Fix a prime p > 2 and a finite field k of characteristic p. Let
G be a nilpotent-by-finite compact p-adic analytic group with ∆+ = 1. Suppose
H = FNp(G), and write F = G/H. Let P be a G-stable, faithful prime ideal
of kH. Let (kG)α be a central 2-cocycle twist of kG with respect to a standard
(Definition 1.19) decomposition
kG = kH ∗
hσ,τ i
F,
for some α ∈ Z 2
prime ideal of (kG)α.
σ(F, Z((kH)×)), as in [11, Theorem 5.12]. Then P (kG)α is a
Proof. First, we note that the claim that P (kG)α is a prime ideal of (kG)α is
equivalent to the claim that
(kG)α/P (kG)α = kH/P ∗
hσ,τ αi
F
is a prime ring.
Case 1. Suppose that G centralises Z.
If H is abelian, so that H = Z, then every g ∈ G is centralised by Z, an open
subgroup of G. Hence g ∈ ∆, i.e. G = ∆. But, by [10, Theorem D], ∆ ≤ H,
and so we have G = H and there is nothing to prove.
30
So suppose henceforth that Z (cid:12) H, and write L := iH (H ′Z), so that, by Lemma
3.5, we have L (cid:12) H. As the decomposition of kG is standard, we may view F
as a subset of G.
The idea behind the proof is as follows. We will construct a crossed product
R ∗ F ′, where R is a certain commutative domain and F ′ is a certain subgroup
of F , with the following property: if R ∗ F ′ is a prime ring, then P (kG)α is a
prime ideal. Then, by using the well-understood structure of R, we will show
that the action of F ′ on R is X-outer (in the sense of Definition 3.1), so that
R ∗ F ′ is a prime ring.
By Corollary 1.21, we can see that H admits an F -stable p-valuation ω satisfying
(AL). Hence, in the notation of §2.1, we may define the filtration w from ω as
in Definition 2.3. Furthermore, we write
Q′ = Q(kZ/P ∩ kZ),
Q = Q′ ⊗
kZ
kN,
as in §2.2; and we endow Q with the F -orbit of filtrations fi (1 ≤ i ≤ s) and
the filtration f of Definitions 2.12 and 2.15, defined in terms of the filtration w
above.
By [7, 2.1.16(vii)], in order to show that the crossed product
kH/P ∗
hσ,τ αi
F
is a prime ring, it suffices to show that the related crossed product
Q ∗
hσ,τ αi
F
(3.1)
(3.2)
is prime, where this crossed product is defined in §2.2. Then, by [5, II.3.2.7], it
suffices to show that
grf (Q ∗ F )
is prime. Details of this graded ring are given in Lemma 2.19:
note that
grf (Q ∗ F ) ∼= s
Mi=1
grfi Q! ∗ F.
(3.3)
in particular,
Now, as noted in Definition 2.12, each grfiQ is a commutative domain, and by
construction, F permutes the summands grfi Q transitively. So by [8, Corollary
14.8] it suffices to show that
grf1 Q ∗ F ′
(3.4)
is prime, where F ′ = StabF (f1).
31
Notation. We set up notation in order to be able to apply the results of §2.3.
Let {ym+1, . . . , yn} be an ordered basis for Z, which we extend to an ordered
basis {yl+1, . . . , yn} for L, which we extend to an ordered basis {y1, . . . , yn} for
H. Set bi = yi − 1 ∈ kH/P , and let Yi = grf1 (bi) for all 1 ≤ i ≤ m. Then
The ring on the right-hand side inherits a crossed product structure
(3.5)
grf1 Q ∼=(cid:0)grv1 Q′(cid:1) [Y1, . . . , Ym].
(cid:0)grv1 Q′(cid:1) [Y1, . . . , Ym] ∗ F ′.
from (3.4). Writing R :=(cid:0)grv1 Q′(cid:1) [Y1, . . . , Ym], we have now shown, by passing
along the chain
(3.5) → (3.4) → (3.3) → (3.2) → (3.1),
that we need only show that R ∗ F ′ is prime.
Write F ′
crossed product (3.5), i.e.
inn for the subgroup of F ′ acting on R by X-inner automorphisms in the
inn = XinnR∗F ′ (R; F ′)
F ′
in the notation of Definition 3.1. By the obvious abuse of notation, we will
denote this action as the map of sets gr σ : F ′ → Aut(R).
Take some g ∈ F ′. If gr σ(g) acts non-trivially on R, then as R is commutative,
we have g 6∈ F ′
inn
is prime, we may restrict our attention to those g ∈ F ′ that act trivially on R.
In particular, such a g ∈ F ′ must centralise each Yi. But
inn. Hence, as by Lemma 3.2(ii) we need only show that R ∗ F ′
gr σ(g)(Yi) = Yi ⇔ f (σ(g)(bi) − bi) > f (bi).
Now we see (as p > 2) from Corollary 2.22 that σ(g) induces the identity
automorphism on H/L, and hence from Lemma 3.6 that g ∈ H. That is, F ′
is the trivial group, so that R ∗ F ′
inn = R is automatically prime.
inn
Case 2. Suppose some x ∈ F does not centralise Z. Write Finn for the subgroup
of F acting by X-inner automorphisms on kH/P in the crossed product (3.1),
i.e.
Finn := Xinn(kG)α/P (kG)α (kH/P ; F ).
Then, by Lemma 3.2(i), x 6∈ Finn; so Finn is contained in CF (Z), and we need
only prove that the sub-crossed product (kH/P ) ∗ CF (Z) is prime by Lemma
3.2(ii). This reduces the problem to Case 1.
Proposition 3.8. Let G be a nilpotent-by-finite compact p-adic analytic group,
and k a finite field of characteristic p > 2. Let H = FNp(G), and write
F = G/H. Let P be a G-stable, almost faithful prime ideal of kH. Then P kG
is prime.
32
Proof. We assume familiarity with [11, Lemma 1.1], and adopt the notation
of [11, Notation 1.2] for this proof.
Let e ∈ cpik∆+
of kG if and only if f · P kG is prime in f · kG.
(P ), and write fH = eH , f = eG. Then P kG is a prime ideal
Write H1 = StabH (e) and G1 = StabG(e). Then, by the Matrix Units Lemma
[11, Lemma 6.1], we get an isomorphism
f · kG ∼= Ms(e · kG1)
for some s, under which the ideal f · P kG is mapped to Ms(e · P1kG1), where
P1 is the preimage in kH1 of e · P · e. It is easy to see that P1 is prime in kH1;
indeed, applying the Matrix Units Lemma to kH, we get
fH · kH ∼= Ms′ (e · kH1),
under which fH · P 7→ Ms′(e · P1), so that P1 is prime by Morita equivalence
(see e.g. [11, Lemma 1.7]). We also know from [11, Lemma 6.6] (or the remark
after [11, Lemma 6.2]) that
P † = \h∈H
(P †
1 )h.
Now, writing q to denote the natural map G → G/∆+,
q(cid:18)(cid:16)P †
1 ∩ ∆(cid:17)h(cid:19) = q(cid:16)P †
1 ∩ ∆(cid:17)
for all h ∈ H, as q(∆) = Z(q(H)) by definition of H (see [10, Lemma 5.1(ii)]);
and so
q(cid:0)P † ∩ ∆(cid:1) = q(cid:16)P †
1(cid:17) is a normal subgroup of the nilpotent group q (H1). Hence, as the
1(cid:17) with the centre q(∆) of q(H) is trivial, we must have that
1 ∩ ∆(cid:17) = q(1).
1 ≤ ∆+(H1) = ∆+.
1(cid:17) is trivial also [9, 5.2.1]. That is, P †
But q(cid:16)P †
intersection of q(cid:16)P †
q(cid:16)P †
Now, in order to show that Ms(e · P1kG1) is prime, we may equivalently (by
Morita equivalence) show that e · P1kG1 is prime. By [11, Theorem 5.12], we
get an isomorphism
e · kG1 ∼= Mt(cid:0)(k′[[G1/∆+]])α(cid:1),
for some integer t, some finite field extension k′/k, and a central 2-cocycle twist
(see above or [11, Definition 5.4]) of k′[[G1/∆+]] with respect to a standard
crossed product decomposition
k′[[G1/∆+]] = k′[[H1/∆+]] ∗
hσ,τ i
(G1/H1)
33
given by some
α ∈ Z 2
σ(cid:16)G1/H1, Z(cid:16)(cid:0)k′[[H1/∆+]](cid:1)×(cid:17)(cid:17) .
Writing the image of e · P1 as Mt(p) for some ideal p ∈ k′[[H1/∆+]], we see
by (see above or [11, Theorem C]) that p is a faithful, (G1/∆+)-stable prime
ideal of k′[[H1/∆+]]. It now remains only to show that the extension of p to
k′[[G1/∆+]] is prime; but this now follows from Proposition 3.7.
Proof of Theorem A. This follows from Proposition 3.8.
References
[1] R.C. Alperin. p-adic binomial coefficients MOD P . The American Mathe-
matical Monthly, 92(8):576 -- 578, 1985.
[2] K. Ardakov. Prime ideals in nilpotent Iwasawa algebras. Inventiones math-
ematicae, 190(2):439 -- 503, 2012.
[3] N. Bourbaki. Commutative Algebra. Springer, 1998.
[4] J.D. Dixon, M.P.F. du Sautoy, A. Mann and D. Segal. Analytic Pro-p
Groups. Cambridge University Press, 1999.
[5] Li Huishi and Freddy van Oystaeyen. Zariskian Filtrations. Springer Sci-
ence+Business Media, 1996.
[6] Michel Lazard. Groupes analytiques p-adiques. Publications Math´ematiques
de l'IH ´ES, 26:5 -- 219, 1965.
[7] J.C. McConnell and J.C. Robson. Noncommutative Noetherian Rings.
American Mathematical Society, 2001.
[8] Donald S. Passman. Infinite Crossed Products. Academic Press Inc., 1989.
[9] Derek J.S. Robinson. A course in the theory of groups. Springer-Verlag
New York, 1996.
[10] W. Woods. On the structure of virtually nilpotent compact p-adic analytic
groups. Preprint, arXiv:1608.03137 [math.GR], 2016. To appear in J.
Group Theory.
[11] W. Woods. Maximal prime homomorphic images of mod-p iwasawa alge-
bras. Preprint, arXiv:1608.07755 [math.RT], 2016.
34
|
1408.3580 | 2 | 1408 | 2015-01-19T19:50:52 | Extensions of simple modules over Leavitt path algebras | [
"math.RA"
] | Let $E$ be a directed graph, $K$ any field, and let $L_K(E)$ denote the Leavitt path algebra of $E$ with coefficients in $K$. For each rational infinite path $c^\infty$ of $E$ we explicitly construct a projective resolution of the corresponding Chen simple left $L_K(E)$-module $V_{[c^\infty]}$. Further, when $E$ is row-finite, for each irrational infinite path $p$ of $E$ we explicitly construct a projective resolution of the corresponding Chen simple left $L_K(E)$-module $V_{[p]}$. For Chen simple modules $S,T$ we describe ${\rm Ext}_{L_K(E)}^1(S,T)$ by presenting an explicit $K$-basis. For any graph $E$ containing at least one cycle, this description guarantees the existence of indecomposable left $L_K(E)$-modules of any prescribed finite length. | math.RA | math |
EXTENSIONS OF SIMPLE MODULES
OVER LEAVITT PATH ALGEBRAS
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
Dedicated to Alberto Facchini on the occasion of his sixtieth birthday.
Abstract. Let E be a directed graph, K any field, and let LK(E) denote the Leavitt path
algebra of E with coefficients in K. For each rational infinite path c∞ of E we explicitly
construct a projective resolution of the corresponding Chen simple left LK(E)-module V[c∞].
Further, when E is row-finite, for each irrational infinite path p of E we explicitly construct
a projective resolution of the corresponding Chen simple left LK(E)-module V[p]. For Chen
simple modules S, T we describe Ext1
LK (E)(S, T ) by presenting an explicit K-basis. For any
graph E containing at least one cycle, this description guarantees the existence of indecom-
posable left LK(E)-modules of any prescribed finite length.
1. Introduction
Given any directed graph E and field K, one may construct the Leavitt path algebra of
E with coefficients in K (denoted LK(E)), as first described in [2] and [3]. Since their in-
troduction, various structural properties of the algebras LK(E) have been discovered, with a
significant number of the results in the subject taking on the following form: LK(E) has some
specified algebraic property if and only if E has some specified graph-theoretic property. (The
structure of the field K often plays no role in results of this type.) A few (of many) examples
of such results include a description of those Leavitt path algebras which are simple; purely
infinite simple; finite dimensional; prime; primitive; exchange; etc.
Although there are graphs for which the structure of corresponding Leavitt path algebra is
relatively pedestrian (e.g., is a direct sum of matrix rings either over K, or over the Laurent
polynomial algebra K[x, x−1], or some combination thereof), the less-mundane examples of
Leavitt path algebras exhibit somewhat exotic behavior. For instance, the prototypical Leav-
itt path algebra A = LK(Rn) (n ≥ 2), which arises from the graph Rn having one vertex and
n loops, has the property that A ∼= An as left (or right) A-modules. Analogous "super de-
composability" properties are also found in other important classes of Leavitt path algebras.
Key words and phrases. Leavitt path algebra, Chen simple module.
2010 AMS Subject Classification: 16S99 (primary)
∗corresponding author [email protected] 7192553182
.
The first author is partially supported by a Simons Foundation Collaboration Grants for Mathematicians
Award #208941. The second and third authors are supported by Progetto di Eccellenza Fondazione Cariparo
"Algebraic structures and their applications: Abelian and derived categories, algebraic entropy and repre-
sentation of algebras" and Progetto di Ateneo "Categorie Differenziali Graduate" CPDA105885. Part of this
work was carried out during a visit of the first author to the Universit`a degli Studi di Padova. The first
author is pleased to take this opportunity to again express his thanks to the host institution, and its faculty,
for its warm hospitality and support by the Visiting Scientist 2012-13 grant of the Universit`a degli Studi di
Padova.
1
2
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
These types of structural properties lead to a dearth (if not outright absence) of indecom-
posable one-sided LK(E)-ideals, which subsequently makes the search for simple (and, more
generally, indecomposable) modules over Leavitt path algebras somewhat of a challenge.
For a graph E, an infinite path in E is a sequence of edges e1e2e3 · · · , for which s(ei+1) =
r(ei) for all i ∈ N. In [6], Chen produces, for each infinite path p in E, a simple left LK(E)-
module V[p]. Further, Chen describes, for each sink vertex w of E, a simple left LK(E)-module
Nw.
In Section 2 we produce explicit projective resolutions for Chen simple modules. As a
result, we will see in Theorem 2.8 that V[c∞] is finitely presented for any closed path c.
Further, in Theorem 2.20 we give necessary and sufficient conditions on a row-finite graph E
which ensure that V[p] is not finitely presented for an irrational infinite path p. In Section 3 we
describe the extension groups Ext1(S, T ) corresponding to any pair of Chen simple modules
S, T . Using some general results about uniserial modules over hereditary rings, we conclude
by showing (Corollary 3.25) how our description of Ext1(S, S) guarantees the existence of
indecomposable LK(E)-modules of any prescribed finite length.
We set some notation. A (directed) graph E = (E0, E1, s, r) consists of a vertex set E0,
an edge set E1, and source and range functions s, r : E1 → E0. For v ∈ E0, the set of edges
{e ∈ E1 s(e) = v} is denoted s−1(v). E is called finite in case both E0 and E1 are finite sets.
E is called row-finite in case s−1(v) is finite for every v ∈ E0. A path α in E is a sequence
e1e2 · · · en of edges in E for which r(ei) = s(ei+1) for all 1 ≤ i ≤ n − 1. We say that such α
has length n, and we write s(α) = s(e1) and r(α) = r(en). We view each vertex v ∈ E0 as a
path of length 0, and denote v = s(v) = r(v). We denote the set of paths in E by Path(E).
A path σ = e1e2 · · · en in E is closed in case r(en) = s(e1). Following [6] (but not standard
in the literature), a closed path σ is called simple in case σ 6= βm for any closed path β and
integer m ≥ 2. A sink in E is a vertex w ∈ E0 for which the set s−1(w) is empty, while an
infinite emitter in E is a vertex u ∈ E0 for which the set s−1(u) is infinite.
For any field K and graph E the Leavitt path algebra LK(E) has been the focus of sus-
tained investigation since 2004. We give here a basic description of LK(E); for additional
information, see [2] or [1]. Let K be a field, and let E = (E0, E1, s, r) be a directed graph
with vertex set E0 and edge set E1. The Leavitt path K-algebra LK(E) of E with coefficients
in K is the K-algebra generated by a set {v v ∈ E0}, together with a set of symbols
{e, e∗ e ∈ E1}, which satisfy the following relations:
vu = δv,uv for all v, u ∈ E0,
s(e)e = er(e) = e for all e ∈ E1,
r(e)e∗ = e∗s(e) = e∗ for all e ∈ E1,
(V)
(E1)
(E2)
(CK1) e∗e′ = δe,e′r(e) for all e, e′ ∈ E1, and
the "double graph" of E, gotten by adding to E an edge e∗ in a reversed direction for each
(CK2) v =P{e∈E1s(e)=v} ee∗ for every v ∈ E0 for which 0 < s−1(v) < ∞.
An alternate description of LK(E) may be given as follows. For any graph E let bE denote
edge e ∈ E1. Then LK(E) is the usual path algebra KbE, modulo the ideal generated by the
Pv∈E0 v. Every element of LK(E) may be written as Pn
It is easy to show that LK(E) is unital if and only if E0 is finite; in this case, 1LK (E) =
i , where ki is a nonzero
If α ∈ Path(E) then we may
relations (CK1) and (CK2).
element of K, and each of the αi and βi are paths in E.
i=1 kiαiβ∗
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
3
view α ∈ LK(E), and will often refer to such α as a real path in LK(E); analogously, for
β = e1e2 · · · en ∈ Path(E) we often refer to the element β∗ = e∗
1 of LK(E) as a ghost
path in LK(E). The map KE → LK(E) given by the K-linear extension of α 7→ α (for
α ∈ Path(E)) is an injection of K-algebras by [1, Corollary 1.5.12].
n · · · e∗
2e∗
The ideas presented in the following few paragraphs come from [6]; however, some of the
notation we use here differs from that used in [6], in order to make our presentation more
notationally consistent with the general body of literature regarding Leavitt path algebras.
Let p be an infinite path in E; that is, p is a sequence e1e2e3 · · · , where ei ∈ E1 for all
i ∈ N, and for which s(ei+1) = r(ei) for all i ∈ N. We emphasize that while the phrase
infinite path in E might seem to suggest otherwise, an infinite path in E is not an element of
Path(E), nor may it be interpreted as an element of the path algebra KE nor of the Leavitt
path algebra LK(E). (Such a path is sometimes called a left-infinite path in the literature.)
We denote the set of infinite paths in E by E∞.
For p = e1e2e3 · · · ∈ E∞ and n ∈ N we denote by τ≤n(p), or often more efficiently by pn,
the (finite) path e1e2 · · · en, while we denote by τ>n(p) the infinite path en+1en+2 · · · . We note
that τ≤n(p) is an element of Path(E) (and thus may be viewed as an element of LK(E)), and
that p is the concatenation p = τ≤n(p) · τ>n(p).
Let c be a closed path in E. Then the path ccc · · · is an infinite path in E, which we denote
by c∞. We call an infinite path of the form c∞ a cyclic infinite path. For c a closed path in
E let d be the simple closed path in E for which c = dn. Then c∞ = d∞ as elements of E∞.
If p and q are infinite paths in E, we say that p and q are tail equivalent (written p ∼ q)
in case there exist integers m, n for which τ>m(p) = τ>n(q); intuitively, p ∼ q in case p and q
eventually become the same infinite path. Clearly ∼ is an equivalence relation on E∞, and
we let [p] denote the ∼ equivalence class of the infinite path p.
The infinite path p is called rational in case p ∼ c∞ for some closed path c. By a previous
observation, we may assume without loss of generality that such c is a simple closed path. In
particular, for any closed path c we have that c∞ is rational. If p ∈ E∞ is not rational we
say p is irrational.
Example 1.1. Let R2 denote the "rose with two petals" graph
e
•v
f
.
Then q = ef ef f ef f f ef f f f e · · · is an irrational infinite path in R∞
2 . Indeed, it is easy to
show that there are uncountably many distinct irrational infinite paths in R∞
2 . We note
additionally that there are infinitely many simple closed paths in Path(R2), for instance, any
path of the form ef i for i ∈ Z+.
Let M be a left LK(E)-module. For each m ∈ M we define the LK(E)-homomorphism
ρm : LK(E) → M, given by ρm(r) = rm. The restriction of the right-multiplication map
ρm may also be viewed as an LK(E)-homomorphism from any left ideal I of LK(E) into M.
When I = LK(E)v for some vertex v of E, we will denote ρm simply by ρm.
Following [6], for any infinite path p in E we construct a simple left LK(E)-module V[p], as
follows.
Definition 1.2. Let p be an infinite path in the graph E, and let K be any field. Let V[p]
denote the K-vector space having basis [p], that is, having basis consisting of distinct elements
6
6
h
h
4
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
of E∞ which are tail-equivalent to p. For any v ∈ E0, e ∈ E1, and q = f1f2f3 · · · ∈ [p], define
e∗ · q =(τ>1(q)
0
if e = f1
otherwise.
v · q =(q
if v = s(f1)
0 otherwise,
e · q =(eq
0
if r(e) = s(f1)
otherwise,
and
Then the K-linear extension to all of V[p] of this action endows V[p] with the structure of a
left LK(E)-module.
Theorem 1.3. ([6, Theorem 3.3]). Let E be any directed graph and K any field. Let p ∈ E∞.
Then the left LK(E)-module V[p] described in Definition 1.2 is simple. Moreover, if p, q ∈ E∞,
∼= V[q] as left LK(E)-modules if and only if p ∼ q, which happens precisely when
then V[p]
V[p] = V[q].
We will refer to a module of the form V[p] as presented in Theorem 1.3 as a Chen simple
module.
For any sink w in a graph E, Chen in [6] presents a construction, similar in flavor to the
one given in Definition 1.2, of a simple left LK(E)-module Nw. He then shows that Nw
is isomorphic as a left LK(E)-module to the left ideal LK(E)w of LK(E) generated by w.
Observe that, for any sink w, the ideal LK(E)w is spanned by the paths in E ending in w.
Moreover for any i ∈ Z+, we get that wi = w and thus we can consider w = w∞ as an element
in E∞. For these reasons, for any sink w of E, we refer to Nw = LK(E)w as a Chen simple
module and, for consistency, we denote Nw by V[w∞].
Remark 1.4. By invoking a powerful result of Bergman, it was established in [3, Theorem
3.5] that, when E is row-finite, then LK(E) is hereditary, i.e., every left ideal of LK(E)
is projective. This presumably could make the search for projective resolutions of various
LK(E)-modules somewhat easier, in that the projectivity of left ideals is already a given.
However, much of the strength of our results lies in our explicit description of the kernels of
germane maps; for instance, it is these explicit descriptions which will allow us to analyze
the Ext1 groups of the Chen simple modules.
A significant majority of the structural properties of a Leavitt path algebras LK(E) do
not rely on the specific structure of the field K. The results contained in this article are no
exceptions. So while each of the statements of the results made herein should also contain the
explicit hypothesis "Let K be any field", we suppress this statement throughout for efficiency.
For a field K, K × denotes the nonzero elements of K.
2. Projective resolutions of Chen simple modules over LK(E)
The goal of this section is to present an explicit description of a projective resolution
of S, where S is a Chen simple module over the Leavitt path algebra LK(E). Such an
explicit description will provide a strengthening of some previously established results (see
[4, Proposition 4.1]), as well as provide the necessary foundation for subsequent results. As
we shall see, the description of projective resolutions, as well as the description of the Ext1
groups, of Chen simple modules will proceed based on which of the following three types
describes the module:
∼= LK(E)w where w is a sink,
(1) V[w∞]
(2) V[c∞] where c is a simple closed path;
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
5
(3) V[q] where q is an irrational infinite path.
Let v be any vertex in E. Since v is an idempotent, the left ideal LK(E)v is a projective
left LK(E)-module. Therefore projective resolutions of Chen simple modules of type 1 are
easy:
Proposition 2.1 (Type (1)). Let w be a sink in E. Then the Chen simple left module V[w∞]
is projective.
Proof. We have V[w∞]
∼= LK(E)w as left LK(E)-modules by [6].
(cid:3)
We now begin the process of describing projective resolutions of Chen simple modules of
the second type, namely, of the form V[c∞] for c a simple closed path.
Notations: Let c = e1e2 · · · et be a simple closed path in E, with v = s(e1) = r(et).
(1) For 0 ≤ i ≤ t we define ci := e1e2 · · · ei and di := ei+1ei+2 · · · et (where c0 = v = dt
and ct = c = d0). Then clearly c = cidi for each 0 ≤ i ≤ t.
(2) For n ≥ 0 we let c−n denote (c∗)n, and let c0 denote v = s(c).
(3) An element µ of LK(E) is said to be a standard form monomial in case there exist
α, β ∈ Path(E) for which µ = αβ∗. We denote the set of standard form monomials in
LK(E) by S. For µ = αβ∗ a standard form monomial we define r(µ) := r(β∗) = s(β);
that is, r(µ) is the unique element v ∈ E0 for which µv = µ. Define
S1(c) := {µ ∈ S µ · cN = 0 in LK(E) for some N ∈ N}, and S2(c) := S \ S1(c).
Although S1(c) and S2(c) depend on c, we will often simply write S1 and S2 for these
sets.
By analyzing the form of monomials in LK(E), we get the following description of the
elements of S1 and S2.
Lemma 2.2. Let 0 6= µ ∈ S. If c is a sink then µ ∈ S1 if and only if r(µ) 6= c. If c is a
closed path e1e2 · · · et, then µ ∈ S1 if and only if µ is of one of the following two forms:
(1) r(µ) 6= s(c) (i.e., µ · s(c) = 0 in LK(E)), or
(2) µ = µ′f ∗c∗
i (c∗)n for some n, i ∈ Z+, some µ′ ∈ S, and some f ∈ E1 for which
s(f ) = s(ei+1) but f 6= ei+1.
Consequently, 0 6= µ ∈ S2 if and only if µ = αc∗
of non-negative integers n, i.
i (c∗)n for some path α in E, and some pair
Lemma 2.3. Let c be a closed path in the graph E, and let v = s(c).
(1) For any z ∈ Z we have cz − v ∈ LK(E)(c − v).
(2) Suppose µ ∈ S1(c). Then µ ∈ (Pu∈E0\{v} LK(E)u)S LK(E)(c − v).
Proof. (1) If z = 0 we have s(c) − v = 0 = 0(c − v). For z > 0 we have cz − v = (cz−1 + cz−2 +
. . . + c + v)(c − v). For z < 0 we have cz − v = −cz(c−z − v), which is in LK(E)(c − v) by the
previous case.
(2) Suppose µ ∈ S1(c).
suppose r(µ) = v, and that µ · cN = 0 for some N ∈ N. But r(µ) = v gives µv = µ, so that
with the hypothesis µ = −µ(cN − v), which gives that µ ∈ LK(E)(cN − v) ⊆ LK(E)(c − v)
by the previous paragraph.
(cid:3)
If r(µ) 6= v then µ ∈ Pu∈E0\{v} LK(E)u. On the other hand,
6
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
Remark 2.4. Let p = e1e2 · · · be an infinite path in E. If p = τ>r(p) for some r > 0, then
p is a rational path of the form p = c∞, where c is the closed path e1e2 · · · er. This follows
from the observation that p = τ>r(p) implies p = τ>ir(p) for all i ∈ N.
Lemma 2.5. Let c be a simple closed path e1e2 · · · et in E with s(c) = r(c) = v. Suppose α
and β are paths in E for which 0 6= α · c∞ = β · c∞ in V[c∞]. Then there exists N ∈ Z+ for
which α = βcN or β = αcN .
Consequently, α · c∞ = β · c∞ in V[c∞] implies α − β ∈ LK(E)(c − v).
Proof. Assume α = f1f2 · · · fℓ and β = g1g2 · · · gm, where the fi and gj are edges in E. If
ℓ = m, from α · c∞ = β · c∞ we get α = β. So assume m > ℓ; we have
with m − ℓ = k + n × t, k ≤ t. If k < t, from α · c∞ = β · c∞ we get
g1 · · · gℓgℓ+1 · · · gm = f1 · · · fℓcne1e2 · · · ek
c∞ = ck+1 · · · ct · c∞ = τ>k(c∞);
then by Remark 2.4 we would have c∞ = (e1 · · · ek)∞, a contradiction since c is simple.
Therefore k = t and β = αcn+1. The case m < ℓ is identical.
For the second statement, note that 0 6= α · c∞ = β · c∞ gives r(α) = r(β); denote this
common vertex by v. So if α = βcn then α − β = β(cn − v), which is in LK(E)(c − v) by
Lemma 2.3(1). The case β = αcn is identical.
(cid:3)
Proposition 2.6. Let E be any graph. Let c be a simple closed path in E, and let v de-
note s(c) = r(c). Let ρc∞ : LK(E)v → V[c∞] and ρc∞ : LK(E) → V[c∞] denote the right
multiplication by c∞. Then
Ker(ρc∞) = LK(E)(c − v)
and Ker(ρc∞) = Xu∈E0\{v}
LK(E)u ⊕ LK(E)(c − v).
Proof. Since (c − v) · c∞ = c∞ − c∞ = 0 in V[c∞], we get LK(E)(c − v) ⊆ Ker(ρc∞). We now
proceed to show that Ker(ρc∞) ⊆ LK(E)(c − v). For notational convenience we denote the
left ideal LK(E)(c − v) of LK(E) by J.
So let λ ∈ Ker(ρc∞), and write
λ = Xµ∈M
kµµ
where M ⊆ S is some finite set of distinct standard form monomials in LK(E), and kµ ∈ K ×.
By Lemma 2.3 we may assume that M ⊆ S2; that is, by Lemma 2.2, we may assume that,
for each µ ∈ M, µ = αµc∗
iµ(c∗)nµ for some path αµ, some 0 ≤ iµ ≤ t, and some nµ ≥ 0.
iµ(c∗)nµ. By hypothesis λ · c∞ = 0 in V[c∞], so that
So we have λ =Pµ∈M kµαµc∗
Xµ∈M
Xµ∈M
But (c∗)n · c∞ = c∞ in V[c∞] for any n ∈ Z. So
kµαµc∗
iµ(c∗)nµ · c∞ = 0 in V[c∞].
kµαµc∗
iµ · c∞ = 0 in V[c∞].
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
7
Also, c∗
i · c∞ = di · c∞ in V[c∞] for any 0 ≤ i ≤ t. So
Now define
kµαµdiµ · c∞ = 0 in V[c∞].
Xµ∈M
λ′ = Xµ∈M
kµαµdiµ.
Then the previous equation gives that λ′ ∈ Ker(ρc∞).
We claim that λ ∈ J if and only if λ′ ∈ J. To show this, we show that λ = λ′ as elements
i = di in LK(E)/J; this follows immediately from the
of LK(E)/J. We note first that c∗
observation that di − c∗
i = c∗
i (c − v) ∈ J. But then in LK(E)/J we have
kµαµc∗
iµ(c∗)nµ
kµαµc∗
iµ(c∗)nµ
kµαµc∗
iµv
by Lemma 2.3(1)
kµαµc∗
iµ
kµαµdiµ
by the above note
kµαµdiµ
λ = Xµ∈M
= Xµ∈M
= Xµ∈M
= Xµ∈M
= Xµ∈M
= Xµ∈M
= λ′.
By Lemma 2.5, if µ ∼ µ′ then αµdiµ = αµ′diµ′ in LK(E)/J; we denote this common element
Thus in order to show that λ ∈ J, it suffices to show that λ′ ∈ J, i.e., that λ′ = 0 in LK(E)/J.
t=1Mt in such
a way that µ ∼ µ′ ∈ Mt (for some t) if and only if αµdiµ · c∞ = αµ′diµ′ · c∞ in V[c∞].
But λ′ ∈ Ker(ρc∞), i.e.,Pµ∈M kµαµdiµ · c∞ = 0 in V[c∞]. Now partition M = ⊔ℓ
NowPµ∈M kµαµdiµ · c∞ = 0 gives
of LK(E)/J by xt.
ℓXt=1 Xµ∈Mt
kµαµdiµ · c∞ = 0,
Pµ∈Mt kµ = 0 for each 1 ≤ t ≤ ℓ. But then
kµαµdiµ =
kµαµdiµ
ℓXt=1 Xµ∈Mt
=
λ′ = Xµ∈M
ℓXt=1 Xµ∈Mt
ℓXt=1
(Xµ∈Mt
=
kµxt
kµαµdiµ =
ℓXt=1 Xµ∈Mt
ℓXt=1
easily from LK(E) =Pu∈E0\{v} LK(E)u ⊕ LK(E)v.
kµ)xt =
(0)xt = 0,
8
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
which by the linear independence of sets of distinct elements of the form α · c∞ in V[c∞] gives
which establishes that Ker(ρc∞) ⊆ LK(E)(c − v), as desired. The claim about ρc∞ follows
(cid:3)
Lemma 2.7. Let E be any graph. Let c be a simple closed path in E based at the vertex v,
and let r ∈ LK(E)v. Then r(c − v) = 0 in LK(E) if and only if r = 0. In particular, the map
ρc−v : LK(E)v → LK(E)(c − v)
is an isomorphism of left LK(E)-modules.
Furthermore, if E is a finite graph, then the map
is an isomorphism of left LK(E)-modules.
ρc−1 : LK(E) → LK(E)(c − 1)
i=1 kiαiβi
Proof. Let r ∈ LK(E)v. If r(c − v) = 0 then rc = rv = r, which recursively gives rcj = r for
∗, where the αi and βi are in Path(E). We note that,
for any m ∈ N, if β ∈ Path(E) has length at most m, then β∗ has the property that β∗cm
is either 0 or an element of Path(E) in LK(E). Now let N be the maximum length of the
paths in the set {β1, β2, . . . , βn}. Then the above discussion shows that rcN is an element of
i=1 kiγi, where γi ∈ Path(E) for 1 ≤ i ≤ n; that is, rcN ∈ KE. But
rcN = r, so that r ∈ KE. However, the equation rc = r (i.e., r(c − v) = 0) has only the zero
solution in KE by a degree argument. So r = 0.
any j ≥ 1. Now write r =Pn
LK(E)v of the formPn
The second statement is established in an almost identical manner.
(cid:3)
We now have all the tools to describe a projective resolution for the modules V[c∞] where c
is a simple closed path, thus completing the study of the second type of Chen simple module.
Theorem 2.8 (Type(2)). Let E be any graph. Let c be a simple closed path in E, with
v = s(c). Then the Chen simple module V[c∞] is finitely presented.
Indeed, a projective
resolution of V[c∞] is given by
0
/ LK(E)v
ρc−v
/ LK(E)v
ρc∞
/ V[c∞]
/ 0 .
If E is a finite graph, an alternate projective resolution of V[c∞] is given by
0
/ LK(E)
ρc−1
/ LK(E)
ρc∞
/ V[c∞]
/ 0 .
/
/
/
/
/
/
/
/
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
9
Proof. V[c∞] is a simple left LK(E)-module by Theorem 1.3, and c∞ = vc∞ is a nonzero
element in V[c∞]. So the map ρc∞ : LK(E)v → V[c∞] is surjective. By Proposition 2.6 we have
Ker(ρc∞) = LK(E)(c − v). We get the first short exact sequence since by Lemma 2.7 the map
ρc−v : LK(E)v → LK(E)(c − v)
is an isomorphism of left LK(E)-modules. Moreover since v is idempotent, LK(E)v is a
projective left LK(E)-module.
Assume now that E is a finite graph. Let us see that Ker(ρc∞) = LK(E)(c − 1). Since
Ker(ρc∞) clearly contains u for any u 6= v ∈ E0, we have c − 1 = c −Pu∈E0 u = (c − v) −
Pu6=v u ∈ Ker(ρc∞). But for any u 6= v = s(c) we have uc = 0, so that u = −u(c − 1) ∈
LK(E)(c − 1). Since c − v = v(c − 1) ∈ LK(E)(c − 1), using Proposition 2.6 we have
shown that each of the generators of Ker(ρc∞) is in LK(E)(c − 1). But by Lemma 2.7,
ρc−1 : LK(E) → LK(E)(c − 1) is an isomorphism of left LK(E)-modules, thus establishing
the result.
(cid:3)
Corollary 2.9. Let E be any graph. Let c be a simple closed path in E, with v = s(c). Then
the Chen simple module V[c∞] has projective dimension 1.
Proof. From Theorem 2.8 we get the exact sequence
0
/ LK(E)v
ρc−v
/ LK(E)v
ρc∞
/ V[c∞]
/ 0 .
Since v is an idempotent in LK(E), the left module LK(E)v is projective and hence V[c∞]
has projective dimension ≤ 1. The left module V[c∞] is not projective, otherwise the above
sequence splits and LK(E)v would contain a direct summand isomorphic to V[c∞]; in particular
LK(E)v would contain a nonzero element α (the element corresponding to c∞) such that
cα = α and hence cnα = α for each n ∈ N. This is impossible by a degree argument.
(cid:3)
Before we present a projective resolution of the third type of Chen simple module, we study,
in the situation where E is row-finite, right multiplication by any of the monomial generators
of V[c∞] for c a simple closed path or a sink. We first introduce some notation which will be
useful throughout the remainder of the section.
Definition 2.10. Let E be any graph. Let β = e1e2 · · · en be a path in E. For each 1 ≤ i ≤ n
let βi denote e1e2 · · · ei. For each 0 ≤ i ≤ n − 1 let
Xi(β) = {f ∈ E1 s(f ) = s(ei+1), and f 6= ei+1}.
The elements of Xi(β) are called the exits of β at s(ei+1). Note that, for a given i, it is
possible that Xi(β) = ∅. For each i ≥ 0 let Ji(β) be the left ideal of LK(E) defined by setting
Ji(β) = Xf ∈Xi(β)
LK(E)f ∗β∗
i .
(So possibly Ji(β) = {0}, precisely when Xi(β) = ∅.) When the path β is clear from context,
we may denote Xi(β) (resp., Ji(β)) by Xi (resp., Ji).
Now let p = e1e2e3 · · · ∈ E∞ be an infinite path in E. Let p0 denote s(e1), and for each
i ≥ 0 let pi+1 denote τ≤i+1(p) = e1e2 · · · ei+1. For each i ≥ 0 we define
Xi(p) := Xi(pi+1), and Ji(p) := Ji(pi+1).
/
/
/
/
10
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
Definition 2.11. Let E be any graph. Let β = e1e2 · · · en be a path in E for which no vertex
of β is an infinite emitter. For 0 ≤ i ≤ n − 1 let
Fi(β) = Xf ∈Xi(β)
f f ∗ ∈ LK(E).
Note that this sum is finite by the hypothesis on β. (We interpret Fi(β) as 0 in case Xi(β) =
∅.) In particular, by the (CK2) relation we have
for 0 ≤ i ≤ n − 1, and by (CK1) that Fi(β)ei+1 = 0.
s(ei+1) − Fi(β) = ei+1e∗
i+1
Lemma 2.12. Let E be any graph. Let α = e1e2 · · · en be a path in E for which no vertex of
α is an infinite emitter. Let αi denote e1e2 · · · ei for each 1 ≤ i ≤ n (so in particular α = αn).
Suppose q, x ∈ LK(E) satisfy the equation qα = x in LK(E). Then
n−1 + · · · + qα1F1(α)α∗
q = xα∗ + qαn−1Fn−1(α)α∗
1 + qF0(α).
Proof. Multiply both sides of the equation qα = x by e∗
n, to get
xe∗
n = qαe∗
n = qe1 · · · en−1ene∗
n = qe1 · · · en−1(s(en) − Fn−1(α)).
Multiplying the final term and switching sides, this gives
qe1 · · · en−1 = xe∗
n + qe1 · · · en−1Fn−1(α).
Multiplying now both sides of this displayed equation on the right by e∗
in the same way, we easily get
ne∗
n−1 + qe1 · · · en−1Fn−1(α)e∗
qe1 · · · en−2 = xe∗
n−1 + qe1 · · · en−2Fn−2(α).
n−1, and proceeding
Continuing in this way, after n steps we reach
q = xe∗
ne∗
n−1 · · · e∗
1 + qe1 · · · en−1Fn−1(α)e∗
n−1 · · · e∗
1 + · · · + qe1F1(α)e∗
1 + qF0(α)
as desired.
(cid:3)
Of course, if c is a simple closed path or a sink, any nonzero element of the Chen simple
module V[c∞] generates V[c∞]; this is in particular true of any "monomial" element αc∞,
where α is a path in E for which r(α) = s(c). We describe here the projective resolution
corresponding to such elements, in case E is row-finite.
Theorem 2.13 (Types (1) & (2)). Let E be any graph. Let c be a simple closed path or a
sink in E, with v = s(c). Let α = e1e2 · · · en be any path in E for which no vertex of α is
an infinite emitter, and for which r(α) = r(en) = v. Let u denote s(α) = s(e1). Then the
following is a projective resolution of the Chen simple LK(E)-module V[c∞]:
0
/ LK(E)(αcα∗ − u)
/ LK(E)u
ραc∞
/ V[c∞]
/ 0 .
Proof. Since uα = α, we have that ραc∞(u) = αc∞ is a nonzero element of the Chen sim-
ple module V[c∞], so that ραc∞ is surjective. So we need only establish that Ker(ραc∞) =
LK(E)(αcα∗ − u). Since ραc∞(αcα∗ − u) = (αcα∗ − u)αc∞ = αcc∞ − αc∞ = 0, it remains
only to show that Ker(ραc∞) ⊆ LK(E)(αcα∗ − u).
/
/
/
/
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
11
So let q ∈ Ker(ραc∞); specifically, qαc∞ = 0. But then qα ∈ Ker(ρc∞), which, by Theorem
2.8, is precisely LK(E)(c − v). So
qα = r(c − v)
for some r ∈ LK(E). By Lemma 2.12, we have
q = r(c − v)α∗ + qαn−1Fn−1(α)α∗
n−1 + · · · + qα1F1(α)α∗
1 + qF0(α).
Using this representation of q, it suffices to show that each of the summands on the right
hand side is an element of LK(E)(αcα∗ − u). Since easily we get (c − v)α∗ = α∗(αcα∗ − u),
we have that r(c − v)α∗ ∈ LK(E)(αcα∗ − u). But for each 0 ≤ i ≤ n − 1 we have Fi(α)α∗
i α =
Fi(α)ei+1 · · · en = 0 (using the observation made in Definition 2.11). Using this, we see
that qαiFi(α)α∗
i ∈ LK(E)(αcα∗ − u) for each
0 ≤ i ≤ n − 1, thus completing the proof.
(cid:3)
i (αcα∗ − u), so that qαiFi(α)α∗
i = −qαiFi(α)α∗
We now describe a projective resolution of the third type of Chen simple module, namely,
one corresponding to an irrational infinite path. Whereas a Chen simple corresponding to
a rational path is always finitely presented, we will see that the determination of the finite-
presentedness of a Chen simple corresponding to an irrational infinite path will depend on
the structure of the graph itself.
Lemma 2.14. Let E be any graph. Let p be an irrational infinite path in E with s(p) = v,
and let ρp : LK(E)v → V[p] be the map r 7→ rp. Let x ∈ Ker(ρp). Then there exists nx ∈ N
such that xτ≤nx(p) = 0 in LK(E). In other words, if xp = 0 in V[p], then xpnx = 0 in LK(E)
for some finite initial segment pnx of p.
i=1 kiαiβ∗
i ∈ Ker(ρp), where αi, βi ∈ Path(E). Denote by N the maximum
Proof. Let x =Pm
length of the βi, i = 1, ..., m. We have
ρp(x) =
kiαiρp(β∗
i ) =
kiαiρτ>N (p)(β∗
i τ≤N (p)).
mXi=1
Since the length of each βi is less than or equal to N, ti := β∗
path. Therefore
i τ≤N (p) is either zero or a real
m′Xℓ=1
m′Xℓ=1
0 = ρp(x) =
kiαiρτ>N (p)(ti) = ρτ>N (p)(
kiαiti) = ρτ>N (p)(
hℓγℓ) =
hℓγℓτ>N (p),
where the γℓ (1 ≤ ℓ ≤ m′) are distinct elements of the form αiti in Path(E), and hℓ ∈ K.
Since p is irrational and γℓ (1 ≤ ℓ ≤ m′) are distinct paths, we claim that the infinite paths
γℓτ>N (p) (1 ≤ ℓ ≤ m′) are distinct elements of V[p], as follows. Assume to the contrary that
γiτ>N (p) = γjτ>N (p) for some i 6= j; necessarily γi and γj have distinct lengths si and sj.
Assume si − sj = s > 0; then
γiτ>N (p) = γjκiτ>N (p) = γjτ>N (p),
and hence κiτ>N (p) = τ>N (p), where κi is a suitable element of Path(E) having length s.
Therefore τ>N (p) = τ>s(τ>N (p)) = τ>s+N (p). But this property implies by Remark 2.4 that
p is rational, contrary to hypothesis. Thus the γℓτ>N (p) (1 ≤ ℓ ≤ m′) are distinct infinite
paths.
mXi=1
mXi=1
mXi=1
12
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
Consequently, the set {γℓτ>N (p) 1 ≤ ℓ ≤ m′} is linearly independent over K, so the
ℓ=1 hℓγℓτ>N (p) yields that hℓ = 0 for each 1 ≤ ℓ ≤ m′.
previously displayed equation 0 =Pm′
Therefore
xτ≤N (p) =
as desired.
mXi=1
kiαiβ∗
i τ≤N (p) =
kiαiti =
mXi=1
m′Xℓ=1
hℓγℓ = 0,
(cid:3)
Lemma 2.15. Let E be any graph. Suppose β is a path of length n in E for which no vertex
of β is an infinite emitter, and for which s(β) = v. For each 0 ≤ i ≤ n − 1 let Ji(β) be the left
ideal of LK(E) given in Definition 2.10. If x ∈ LK(E)v has xβ = 0, then x ∈Pn−1
Proof. Write β = e1e2 · · · en. For each 1 ≤ i ≤ n let βi = e1e2 · · · ei. So β = βn, and thus
by hypothesis we are assuming that xβn = 0. Then using the (CK2) relation at the vertices
s(e1), s(e2), · · · , s(en) in order (this is possible by the hypothesis on β), and interpreting
empty sums as 0, we get
i=0 Ji(β).
= · · ·
f f ∗ + β1β∗
x = xv = x( Xf ∈X0(β)
= x( Xf ∈X0(β)
= x( Xf ∈X0(β)
= x( Xf ∈X0(β)
= x( Xf ∈X0(β)
= x( Xf ∈X0(β)
x = Xf ∈X0(β)
f f ∗) + xβ1( Xf ∈X1(β)
f f ∗) + xβ1( Xf ∈X1(β)
f f ∗) + xβ1( Xf ∈X1(β)
f f ∗) + xβ1( Xf ∈X1(β)
f f ∗) + xβ1( Xf ∈X1(β)
(xf )f ∗ + Xf ∈X1(β)
f f ∗) + xβ1r(β1)β∗
1
1 ) = x( Xf ∈X0(β)
f f ∗ + e2e∗
2)β∗
1
f f ∗)β∗
1 + xβ2r(β2)β∗
2
f f ∗)β∗
f f ∗)β∗
f f ∗)β∗
1 + · · · + xβn−1( Xf ∈Xn−1(β)
1 + · · · + xβn−1( Xf ∈Xn−1(β)
1 + · · · + xβn−1( Xf ∈Xn−1(β)
1 + · · · + Xf ∈Xn−1(β)
f f ∗)β∗
n−1 + xβn−1ene∗
nβ∗
n−1
f f ∗)β∗
n−1 + xβnβ∗
n
f f ∗)β∗
n−1 + 0,
with the final statement following from the hypothesis that xβ = xβn = 0. Thus
(xβ1f )f ∗β∗
(xβn−1f )f ∗β∗
n−1 ∈
Ji(β).
n−1Xi=0
(cid:3)
Lemma 2.16. Let E be a finite graph, and let p = e1e2 · · · ∈ E∞ be an irrational infinite
path in E. Then Xi(p) is nonempty for infinitely many i ∈ Z+. Consequently, in this case,
Ji(p) is nonzero for infinitely many i ∈ Z+.
Proof. Suppose to the contrary that there exists N ∈ N for which Xi(p) = ∅ for all i ≥ N.
Since E0 is finite, there exist t, t′ ≥ N, t < t′, for which s(et) = s(et′). But Xt(p) = ∅ then
gives et = et′, and in a similar manner yields et+ℓ = et′+ℓ for all ℓ ∈ Z+. If d denotes the
closed path etet+1 · · · et′−1, then we get p ∼ d∞, the desired contradiction.
(cid:3)
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
13
We note that Lemma 2.16 is not necessarily true without the finiteness hypothesis on the
graph. For instance, let MN be the graph
and let p ∈ M ∞
N be the irrational infinite path e1e2 · · · . Then Xi(p) = ∅ for all i ≥ 0.
•
e1
/ •
e2
/ •
e3 /
/ · · ·
Corollary 2.17. Let E be any graph. Let p ∈ E∞ be an irrational infinite path in E for
which no vertex of p is an infinite emitter, and for which s(p) = v. Let ρp : LK(E)v → V[p]
be the map r 7→ rp. For each i ≥ 0 let Ji(p) be the left ideal of LK(E) given in Definition
2.10. Then
Ker(ρp) =M∞
i=0
Ji(p).
Proof. Clearly, for every i ≥ 0, each element of Ji(p) is in Ker(ρp). Now suppose x ∈ LK(E)
has xp = 0 in V[p]. By Lemma 2.14, xβ = 0 where β = pn = τ≤n(p) for some n ∈ N. Then
i=0Ji(p).
i=0 ri = 0 in LK(E), where ri ∈ Ji(p) for 0 ≤ i ≤ n. By construction,
n, so that
n = rn for all rn ∈ Jn(p). Thus multiplying both sides of the proposed equation
n gives rn = 0. Using this same idea iteratively, we get ri = 0
(cid:3)
Lemma 2.15, together with the definition of Ji(p) for p ∈ E∞, gives that Ker(ρp) =P∞
Now suppose Pn
Pn
ripn = 0 for all i < n. On the other hand, for any f ∈ Xn(p), f ∗p∗
rnpnp∗
for all 0 ≤ i ≤ n, so that the sum is indeed direct.
i=0 ri = 0 on the right by pnp∗
We note that Corollary 2.17 is not necessarily true without the finite emitter hypothesis
npnp∗
n = f ∗p∗
on the vertices of p. For instance, let F be the graph
•w
∞
•v
e1
/ •
e2
/ •
e3 /
/ · · ·
where there are infinitely many edges {fi i ∈ Z+} from v to w. Let p be the irrational infinite
path e1e2 · · · , and let ρp : LK(E)v → V[p] as usual. Then easily x = v − e1e∗
1 ∈ Ker(ρp).
1 plus
i=0Ji(p), since otherwise this would yield that v is a finite sum of e1e∗
i for ri ∈ LK(E), which cannot happen as v is an infinite emitter.
However, x /∈P∞
terms of the form rifif ∗
Remark 2.18. Corollary 2.17 shows that if E is row-finite and p is an irrational infinite
path, then Ker(ρp) is generated by those ghost paths of LK(E) which annihilate some (finite)
initial path of p. Effectively, this is the main difference between the rational and irrational
cases; in the rational case, where p = d∞ and s(p) = v, there are additional elements in
Ker(ρp) which are not of this form, namely, elements of the form r(d − v) where r ∈ LK(E).
Lemma 2.19. Let p be an irrational infinite path in an arbitrary graph E. For f ∈ E1 let
vf denote the vertex r(f ). Then, for each i ≥ 0,
Ji(p) = Mf ∈Xi(p)
LK(E)f ∗p∗
i
LK(E)vf .
∼= Mf ∈Xi(p)
as left LK(E)-modules. In particular, each Ji(p) is a projective left LK(E)-module.
Proof. By definition Ji(p) =Pf ∈Xi(p) LK(E)f ∗p∗
0 =Pf ∈Xi(p) rf f ∗p∗
for each f ∈ Xi(p). Take g ∈ Xi(p); by multiplying 0 =Pf ∈Xi(p) rf f ∗p∗
i . We claim the sum is direct. So suppose
i , with rf ∈ LK(E) for each f ∈ Xi(p). Without loss we may assume that
each expression rf f ∗ is nonzero, so that we may further assume without loss that rf vf = rf
i on the right by pig,
and using the (CK1) relation, we get 0 = rgg∗g = rg · vg = rg. Thus the sum is direct, so that
/
/
k
s
/
/
14
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
i . But for g ∈ Xi(p) it is easy to show that LK(E)g∗p∗
Ji(p) = ⊕f ∈Xi(p)LK(E)f ∗p∗
by the map x 7→ xpig.
Theorem 2.20 (Type (3)). Let E be any graph. Let p ∈ E∞ be an irrational infinite path in
E for which no vertex of p is an infinite emitter. Then the Chen simple LK(E)-module V[p]
is finitely presented if and only if Xi(p) is nonempty only for finitely many i ∈ Z+.
∼= LK(E)vg,
(cid:3)
i
In particular, if E is a finite graph, then V[p] is not finitely presented.
Proof. Let v denote s(p). We consider the exact sequence
0
/ Ker(ρp)
/ LK(E)v
ρp
/ V[p]
/ 0 .
By Corollary 2.17 we have that Ker(ρp) = ⊕∞
i=0Ji(p). Furthermore, each Ji(p) is projective
by Lemma 2.19, so the given exact sequence is a projective resolution of V[p]. Therefore V[p]
is finitely presented if and only if Ji(p) is nonzero only for finitely many i ∈ Z+, i.e. Xi(p) is
nonempty only for finitely many i ∈ Z+.
For the particular case, when E is finite then by Lemma 2.16 Ji(p) is nonzero for infinitely
(cid:3)
many i.
Corollary 2.21. If E is a finite graph, and p ∈ E∞ is an irrational infinite path in E, then
the Chen simple LK(E)-module V[p] has projective dimension 1.
Proof. From Corollary 2.17 we get the exact sequence
0
i=0Ji(p)
/ LK(E)v
ρp
/ V[p]
/ 0 .
/L∞
Since v is an idempotent in LK(E), the left module LK(E)v is projective; by Lemma 2.19
i=0Ji(p) is projective and hence V[p] has projective dimension ≤ 1. Since E is finite,
i=0Ji(p) is not finitely generated. Then
i=0Ji(p) would be a not finitely generated
(cid:3)
also L∞
Ji(p) is not zero for infinitely many i and hence L∞
the left module V[p] is not projective, otherwiseL∞
direct summand of a cyclic module: contradiction.
Remark 2.22. Let MN be the graph
•
e1
/ •
e2
/ •
e3 /
/ · · ·
considered previously, and let p ∈ M ∞
N be the irrational infinite path e1e2e3 · · · . Then Xi(p) =
∅ for all i ≥ 0. So by Corollary 2.17, the Chen simple module V[p] is isomorphic to LK(E)v,
and hence it is projective.
Example 2.23. We reconsider the graph R2 and irrational infinite path q = ef ef f ef f f e · · · ∈
R∞
2 described in Example 1.1. Then, as R2 is finite, Theorem 2.20 yields that the Chen simple
module V[q] is not finitely presented.
Remark 2.24. We note that Theorems 2.8 and 2.20 strengthen and sharpen [4, Proposition
4.1], most notably because we have been able to explicitly describe a projective resolution of
each of the Chen simple modules.
In [5, Theorem 4.12] it is shown that for any graph E, for any vertex v ∈ E0, L(E)v is a
simple left ideal if and only if v is a line point, i.e.
in the full subgraph of E generated by
{u ∈ E0 there is a path from v to u} there are no cycles, and there are no vertices which
emit more than one edge. Our results allow us to recover [5, Theorem 4.12], as follows.
/
/
/
/
/
/
/
/
/
/
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
15
Corollary 2.25. Let E be any graph. Let u ∈ E0. Then LK(E)u is simple if and only if u
is a line point.
Proof. There are three possibilities:
(1) there is a path α ∈ Path(E) with s(α) = u and for which r(α) = w is a sink in E;
(2) there is a path α ∈ Path(E) with s(α) = u and for which r(α) = v is the source of a
simple closed path c;
(3) there is an infinite irrational path q for which s(q) = u.
If in α (cases 1 and 2) or in q (case 3) there is an infinite emitter x, then LK(E)x is not a
simple submodule of LK(E)v (see [5, Lemma 4.3]). Therefore we can assume that α (cases 1
and 2) and q (case 3) have no infinite emitter.
Cases 1 and 2. By Theorem 2.13 LK(E)u is a simple module if and only if αcα∗ = u, where
c is either a simple closed path or a sink. If c is a simple closed path, by a degree argument
αcα∗ is not a vertex. If c is a sink and α = e1 · · · eℓ then αcα∗ = αα∗ = e1 · · · eℓe∗
1 is
equal to u if and only if eie∗
if and only if s(ei) is the source of
only one edge, i.e. u is a line point.
i = s(ei) for i = 1, ..., ℓ, i.e.
ℓ · · · e∗
Case 3. By Corollary 2.17, LK(E)u is simple if and only ifL∞
is equivalent to p having no exits, i.e. u is a line point.
i=0 Ji(p) = 0 and the latter
(cid:3)
3. Extensions of Chen simple modules
In this section we use the results of Section 2 to describe Ext1
LK (E)(S, T ), where S and
T are Chen simple modules over the Leavitt path algebra LK(E) corresponding to a finite
graph E. As a consequence, this will allow us to (among other things) construct classes of
indecomposable non-simple LK(E)-modules.
We give here a short review of Ext1; see e.g.
[9] for more information. Let R be a ring,
and let M, N be left R-modules. Suppose
0
/ Q
µ
/ P
f
/ M
/ 0
is a short exact sequence with P projective. Then there is an exact sequence of abelian groups
HomR(P, N)
µ∗
/ HomR(Q, N)
∆f
/ Ext1
R(M, N)
/ 0,
where µ∗(ϕ) = ϕ ◦ µ for ϕ ∈ HomR(P, N), and ∆f is the "connecting morphism". If µ is
viewed as an inclusion of submodules, then µ∗(ϕ) = ϕQ, the restriction of ϕ to Q. Exactness
yields that Ext1
R(M, N) 6= 0 if and
only if there exists a non-splitting short exact sequence
R(M, N) = 0 if and only µ∗ is surjective. Moreover, Ext1
0 → N → L → M → 0,
i.e., L is a non-trivial extension of N by M. For instance if M, N are simple left R-modules,
then Ext1
R(M, N) 6= 0 if and only if there exist indecomposable left R-modules of length 2
which are extensions of N by M. Finally, observe that if R is a K-algebra over a field K,
then the abelian group Ext1
R(M, N) has a natural structure of K-vector space for any left
R-module M and N.
/
/
/
/
/
/
/
16
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
We outline our approach. There are three types of Chen simple modules: those of the
form V[w∞] for a sink w; of the form V[c∞] for a simple closed path c; and of the form V[p]
for an irrational infinite path p. Let T denote any Chen simple module.
In Lemma 3.1
we make the (trivial) observation that Ext1
LK (E)(V[w∞], T ) = 0; in Theorem 3.13 we describe
Ext1
LK (E)(V[p], T ). We recall that we are
assuming w = w∞ ∈ E∞ for any sink w.
LK (E)(V[c∞], T ); and in Theorem 3.21 we describe Ext1
Lemma 3.1. (Type (1)) Let E be any graph. Let w be a sink in E, and let T denote any
left LK(E)-module. Then Ext1
LK (E)(V[w∞], T ) = 0, i.e. any extension of V[w∞] by T splits.
Proof. This follows immediately from the fact that V[w]
module (see Proposition 2.1).
∼= LK(E)w is a projective LK(E)-
(cid:3)
Definition 3.2. Let T be a Chen simple module. Denote by U(T ) the set
U(T ) := {v ∈ E0 vT 6= {0}} = {v ∈ E0 there exists t ∈ T with vt 6= 0}.
Remark 3.3. Let T be a Chen simple module and let q = e1e2 · · · ∈ E∞ such that T = V[q].
Then U(T ) consists of those vertices v for which there is a path α ∈ Path(E) having s(α) = v
and r(α) = s(ei) for some i ≥ 1. Equivalently, a vertex v ∈ U(T ) if and only if there is an
infinite path tail-equivalent to q starting from v. Hence U(T ) is a feature of the Chen simple
module T that can be read directly from the graph E.
Definitions 3.4. Let E be any graph and let d be a simple closed path in E.
For any p ∈ E∞ we say that p is divisible by d if p = dp′ for some p′ ∈ E∞.
For any q ∈ E∞, we define the set
L(d,q) := {p ∈ E∞ p ∼ q, s(p) = s(d), and p is not divisible by d} ⊆ V[q],
where V[q] is the Chen simple LK(E)-module generated by q.
An infinite path p is divisible by a simple closed path d ∈ E if and only if d = t≤ℓ(p), where
ℓ is the length of d. The set L(d,q) consists of those infinite paths which start at s(d), and
which eventually equal some tail of q, but do not start out by traversing the closed path d.
Observe that the subset L(d,q) of V[q] does not depend on q but only on the equivalence class
[q]. Let T = V[q]; if q is not tail equivalent to d∞, then there exists q′ ∼ q such that d 6 q′ and
hence T has a generator not divisible by d.
Remark 3.5. Let d be a simple closed path in E and q ∈ E∞.
1) Suppose q is not tail equivalent to d∞ and consider the Chen simple module V[q]; we can
assume without loss of generality that q is not divisible by d. The set L(d,q) is not empty if
and only if s(d) belongs to U(V[q]); in such a case any 0 6= t ∈ V[q] for which s(d)t = t is a
linear combination of infinite paths tail equivalent to q whose sources coincide with s(d). In
particular, taking in account the divisibility by d of these infinite paths, t can be written in
a unique way as
t = t0 + dt1 + d2t2 + · · · + dsts,
where the ti are K-linear combinations of elements in L(d,q) and ts 6= 0. We call s ≥ 0 the
d-degree of t and we denote it by degd(t).
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
17
(2) Suppose q = d∞. Then L(d,d∞) 6= ∅ if and only if there exists a cycle c 6= d with
s(c) = s(d). Any 0 6= t ∈ V[d∞] for which s(d)t = t can be written in a unique way as
t = kd∞ + t0 + dt1 + d2t2 + · · · + dsts,
where the ti ∈ V[d∞] are K-linear combinations of elements in L(d,d∞) and ts 6= 0. We call
s ≥ 0 the d-degree of t and we denote it by degd(t).
In particular, any 0 6= t ∈ L(d,q) has d-degree equal to 0. We emphasize that, in case
q = d∞, the d-degree of the element d∞ of V[d∞] is zero too: degd(d∞) = 0. The d-degree is
not defined on 0.
Example 3.6. We revisit the graph R2 given by
e
•v
f
.
Consider the simple closed path e and the rational infinite path f ∞. Then L(e,f ∞) = {p ∈
R∞
p ∼ f ∞ and p is not divisible by e} ⊆ V[f ∞] contains, for instance, the infinite paths
2
{f iejf ∞ i ≥ 1, j ≥ 0}. (There are additional elements of L(e,f ∞), for instance, f ef ef ∞.)
Moreover, consider an element of the form ejf ief ∞ ∈ V[f ∞], with i ≥ 1 and j ≥ 0. Then
dege(ejf ief ∞) = j.
On the other hand, L(f,f ∞) = {p ∈ R∞
2
p ∼ f ∞ and p is not divisible by f } contains
the infinite paths {eif ∞ i ≥ 1}. Note that the element f ∞ of Vf ∞ is defined to have
degf (f ∞) = 0.
Recall that LK(E) is a ring wit unity if and only if E is finite.
Lemma 3.7. Let E be a finite graph. Let d be a simple closed path and q ∈ E∞. Let t ∈ V[q],
and consider the equation in the variable X
(d − 1)X = t.
The equation admits a solution in V[q] if one of the following holds:
(1) s(d)t = 0;
(2) t = dnp − p for some p ∈ L(d,q) and n ≥ 0.
Proof. (1) is easy, since if s(d)t = 0 then dt = 0, and hence X = −t is a solution. (2) is
i=0 dip is a
(cid:3)
i=0 dip = dnp − p = t, and hence X =Pn−1
nearly as easy, since we have (d − 1)Pn−1
solution.
Lemma 3.8. Let E be a finite graph. Let d be a simple closed path and let q ∈ E∞. Assume
either q = d∞ or q is a generator of V[q] not divisible by d. Let 0 6= t ∈ V[q], and consider the
equation
(d − 1)X = t.
Assume t = dnt′ for some n ≥ 0 and some 0 6= t′ ∈ V[q] for which s(d)t′ = t′ and degd(t′) = 0.
Then the equation has no solution in V[q]. In particular:
(1) the equation (d − 1)X = t has no solution in V[q] whenever t ∈ L(d,q), and
(2) the equation (d − 1)X = d∞ has no solution in V[d∞].
6
6
h
h
18
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
Proof. Let v = s(d). Since vd = d and t = dnt′, we get vt = t. So if x is a solution of
(d − 1)X = t, then we would have v(d − 1)x = t, so that (d − 1)vx = t; thus we may assume
without loss that vx = x. Hence the equation yields
and, since t 6= 0, necessarily then x 6= 0. Let degd(x) = s and write
x = kd∞ + x0 + dx1 + · · · + dsxs,
x = dx − t,
where the xi's are linear combination of elements in L(d,q), and k = 0 in case q 6= d∞. Then,
using d · d∞ = d∞, we get
t = dnt′ = (d − 1)x = dx − x
= kd∞ − kd∞ − x0 + d(x0 − x1) + · · · + ds(xs−1 − xs) + ds+1xs
= −x0 + d(x0 − x1) + · · · + ds(xs−1 − xs) + ds+1xs.
We claim that this is impossible. Set x−1 = 0 = xs+1. By the uniqueness of the decomposition
in Remark 3.5, since degd(t′) = 0, one gets t′ = xn−1 − xn and xi−1 − xi = 0 for any i 6= n,
−1 ≤ n ≤ s + 1. Then we have 0 = x0 = x1 = · · · = xn−1 and t′ = −xn = · · · − xs+1 = 0,
contradiction.
(cid:3)
Assume E is a finite graph and d a simple closed path in E. In order to compute the groups
LK (E)(V[d∞], T ) for any Chen simple module T , we can consider the projective resolution
Ext1
of V[d∞]
0 → LK(E)
ρ(d−1)→ LK(E)
ρd∞→ V[d∞] → 0
ensured by Theorem 2.8.
Lemma 3.9. Let E be a finite graph. Let d be a simple closed path in E and let T be a Chen
simple module. Consider the exact sequence
HomLK (E)(LK(E), T )
ρ(d−1)∗
/ HomLK (E)(LK(E), T )
π
/ Ext1
LK (E)(V[d∞], T )
/ 0
where ρ(d−1)∗ (φ) = φ ◦ ρd−1, and π is the connecting homomorphism. Then
π(ρt) = 0 if and only if the equation (d − 1)X = t has a solution in T.
Consequently, Ext1
t ∈ T .
LK (E)(V[d∞], T ) = 0 if and only if (d − 1)X = t has a solution in T for every
Proof. By exactness it follows that π(ρt) = 0 if and only if there exists x ∈ T such that
ρt = ρ(d−1)∗(ρx) = ρx ◦ ρ(d−1) = ρ(d−1)x
i.e. if and only if the equation (d − 1)X = t has a solution in T .
The final statement follows directly from the exactness of the displayed sequence.
(cid:3)
Theorem 3.10. (Type (2)) Let E be a finite graph. Let d be a simple closed path in E and
let T be a Chen simple module. Then the following are equivalent:
(1) Ext1
(2) s(d) ∈ U(T )
LK (E)(V[d∞], T ) 6= 0.
/
/
/
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
19
Proof. (1) ⇒ (2) If s(d)T = 0, then for any t ∈ T we have s(d)t = 0, so by Lemma 3.7(1) the
equation (d − 1)X = t admits a solution for any t ∈ T . Applying Lemma 3.9(2), we get that
Ext1
LK (E)(V[d∞], T ) = 0.
(2) ⇒ (1) First assume T 6= V[d∞]. As observed in Remark 3.5, T admits a generator
q not divisible by d and L(d,q) is not empty. Let p ∈ L(d,q). By Lemma 3.8, the equation
(d − 1)X = p has no solution in V[q] and so, by Lemma 3.9, π(ρp) 6= 0.
On the other hand, suppose T = V[d∞]. By Lemma 3.8, the equation (d − 1)X = d∞ has
no solution in V[d∞], and so, again invoking Lemma 3.9, π(ρd∞) 6= 0.
In either case we have established the existence of a nonzero element in Ext1
LK (E)(V[d∞], T ).
(cid:3)
Corollary 3.11. Let E be a finite graph. For any simple closed path d, Ext1
0.
LK (E)(V[d∞], V[d∞]) 6=
Example 3.12. We again revisit the graph R2:
e
•v
f
.
Let q be any element of R∞
Since clearly Condition (2) of Theorem 3.10 is satisfied for V[q], we get that Ext1
0.
2 . Let d be any (of the infinitely many) simple closed paths in R2.
LK (R2)(V[d∞], V[q]) 6=
✷
Having now established conditions which ensure that there exist nontrivial extensions of
the Chen simple module T by the simple module V[d∞], we now give a more explicit description
of the number of such extensions.
Proposition 3.13. Let E be any finite graph. Let d be a simple closed path in E and let T
be a Chen simple module. Assume q ∈ E∞ such that T = V[q].
(1) Suppose T 6= V[d∞]. Then dimK Ext1
(2) On the other hand, dimK Ext1
LK (E)(V[d∞], T ) = L(d,q).
LK (E)(V[d∞], V[d∞]) = L(d,d∞) + 1.
Proof. We consider the exact sequence
HomLK (E)(LK(E), V[q])
ρ(d−1)∗
/ HomLK (E)(LK(E), V[q])
π
/ Ext1
LK (E)(V[d∞], V[q])
/ 0 .
(1) Without loss of generality we can assume q is not divisible by d. By Remark 3.5(1)
and Theorem 3.10, if L(d,q) = ∅ then Ext1
LK (E)(V[d∞], T ) = 0. Otherwise, by Lemmas 3.8
and 3.9, π(ρp) 6= 0 for any path p ∈ L(d,q). We claim that the set {π(ρp) p ∈ L(d,q)} is a
basis for the vector space Ext1
LK (E)(V[d∞], T ). In order to show that {π(ρp) p ∈ L(d,q)} is
K-linearly independent, suppose there is a K-linear combination 0 = k1π(ρp1)+· · ·+knπ(ρpn)
in Ext1
LK (E)(V[d∞], V[q]), with pi ∈ L(d,q). Let t = k1p1 + · · · + knpn in V[q] so that π(ρt) = 0
in Ext1
LK (E)(V[d∞], V[q]). Thus, applying Lemma 3.9, we get that the equation (d − 1)X = t
has a solution in V[q]. If t 6= 0, since s(d)t = t and pi ∈ L(d,q) we get degd(t) = 0, which is a
contradiction by Lemma 3.8. Hence t = 0 and by the linear independence of {p1, . . . , pn} in
V[q] we get that k1 = · · · = kn = 0. So {π(ρp) p ∈ L(d,q)} is K-linearly independent.
We now show that {π(ρp) p ∈ L(d,q)} spans Ext1
Lemma 3.9(1) it suffices to show that any π(ρt) ∈ Ext1
LK (E)(V[d∞], T ). As π is surjective, by
LK (E)(V[d∞], T ) is a K-linear combination
6
6
h
h
/
/
/
20
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
′′
t
j=1 kjp′′
j with s(p′′
i=1 kip′
i with s(p′
i) = s(d) and
j ) 6= s(d). By Lemma 3.7(1), the equation (d − 1)X = t has
solution in T = V[q] if and only if (d − 1)X = t′ has solution in V[q], so we can assume
without loss of generality that s(d)t = t. Hence t = t0 + dt1 + d2t2 + · · · + dsts, and so
j=1 kijuij, for some
of elements from this set. Write t = t′ + t′′ where t′ = Pmt′
t′′ = Pm
ρt = ρt0 + ρdt1 + ρd2t2 + · · · + ρdsts, where each ti is of the form ti = Pmi
uij ∈ L(d, q). Thus π(ρt) =Pm0
j=1 k1jπ(ρdu1j ) + · · · +Pj=1 msksjπ(ρdsusj ).
Observe that, by Lemmas 3.7(2) and 3.9, we get π(ρdnu − ρu) = 0 for any u ∈ L(d,q) and any
n ∈ N, so π(ρdnu) = π(ρu) for any n ∈ N. Hence {π(ρu) u ∈ L(d,q)} is a set of generators for
Ext1
j=1 k0jπ(ρu0j ) +Pm1
LK (E)(V[d∞], T ).
(2) Let us show that {π(ρp) p ∈ L(d,d∞)} ∪ {π(ρd∞)} is a basis for Ext1
LK (E)(V[d∞], V[d∞]).
First observe that, by Lemmas 3.8 and 3.9(2), π(ρd∞) 6= 0 and π(ρp) 6= 0 for any p ∈ L(d,d∞).
Arguing as in part (1) we claim that {π(ρp) p ∈ L(d,d∞)} ∪ {π(ρd∞)} is a linearly indepen-
dent set in Ext1
LK (E)(V[d∞], V[d∞]). Indeed, consider a K-linear combination 0 = k0π(ρd∞) +
k1π(ρp1) + · · · + knπ(ρpn) in Ext1
LK (E)(V[d∞], V[d∞]). Define y = k0d∞ + k1p1 + · · · + knpn ∈ V[d∞]
so that π(ρy) = 0 and hence, by Lemma 3.9(2), the equation (d − 1)X = y has a solution in
V[d∞]. Note that if y 6= 0 then degd(y) = 0 (whether or not k0 = 0) since each pi ∈ L(d,d∞),
which is a contradiction by Lemma 3.8. So y = 0, which yields that each ki (0 ≤ i ≤ n) is 0.
Since any t in V[d∞] with s(d)t = t is of the form t = kd∞ + t0 + dt1 + d2t2 + · · · + dsts,
using the same arguments as in part (1) it can be easily be shown that the set {π(ρp) p ∈
L(d,d∞)} ∪ {π(ρd∞)} spans Ext1
(cid:3)
Lemma 3.14. Let d be a simple closed path in the finite graph E.
LK (E)(V[d∞], V[d∞]).
(1) If q ∈ E∞ is irrational and L(d,q) 6= ∅, then L(d,q) is infinite.
(2) If L(d,d∞) 6= ∅, then L(d,d∞) is infinite.
Proof. (1) Let q = e1e2 · · · for ei ∈ E1. First notice that, for any w ∈ E0, if w = r(ei) for
some i > 0, then we can assume without loss of generality that w = r(ej) for infinitely many
j > 0 (as otherwise, since E1 is finite, we can replace q with q′ ∈ E∞ for which q ∼ q′ and
w 6∈ (q′)0).
Consider now an element p ∈ L(d,q). Then p = βq0 and q = γq0 for some q0 ∈ [q] and
β, γ ∈ Path(E) and β not divisible by d. Consider w = r(β) = s(q0). Then, by the previous
assumption, there exists a set of infinite and distinct truncations {τ>nk(q) k ∈ N} such that,
for each k ∈ N, q = γkwτ>nk(q) for some γk ∈ Path(E). Since q is irrational, by Remark 2.4
the infinite paths in the set {τ>nk(q) k ∈ N} are distinct. Hence there are infinitely many
distinct elements βτ>nk(q) in L(d,q), which establishes (1).
(2) If L(d,d∞) 6= ∅, then there is at least one simple closed path c for which s(c) = s(d) and
c 6= d. Then we easily get that each of the distinct paths {cid∞ i ∈ N} is tail equivalent to
d∞, which gives that L(d,d∞) is infinite.
(cid:3)
Corollary 3.15. Let d be a simple closed path in E and T a Chen simple module.
dimK Ext1
Proof. It follows directly from Lemma 3.14 and Proposition 3.13.
LK (E)(V[d∞], T ) is finite, then T = V[c∞] for a simple closed path c.
If
(cid:3)
Example 3.16. For each n ∈ N, consider the graph
En =
d
•v
(n)
/ •w
f
,
6
6
/
h
h
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
21
where the symbol (n) indicates that there are n edges {e1, . . . , en} for which s(ei) = v and
r(ei) = w. Then in En we have L(d,f ∞) = {e1f ∞, . . . , enf ∞}, so that L(d,f ∞) = n. By
Proposition 3.13(1) we conclude that dimK Ext1
LK (En)(V[d∞], V[f ∞]) = n.
Example 3.17. Consider the graph
R1 = •v
d
,
for which LK(R1) ∼= K[x, x−1]. Then L(d,d∞) is empty, hence by Proposition 3.13(2) we
conclude that dimK Ext1
LK (R1)(V[d∞], V[d∞]) = 1.
Having given a complete analysis of Ext1
LK (E)(V[d∞], T ) for any simple closed path d of E
and any Chen simple module T , we now analyze Ext1
LK (E)(V[p], T ) for any irrational infinite
path p and any Chen simple module T . The projective resolution of V[p] we are going to use
is the one introduced in the proof of Theorem 2.20:
0
/ Ker(ρp)
/ LK(E)v
ρp
/ V[p]
/ 0 ,
where Ker(ρp) = ⊕∞
i=0Ji(p) as in Corollary 2.17 and v = s(p).
Remark 3.18. Let E be a finite graph and u ∈ LK(E). For each left LK(E)-module M, any
morphism φ ∈ HomLK (E)(LK(E)u, M) is the right product by the element φ(u) of M. If u is
an idempotent, then it is HomLK (E)(LK(E)u, M) ∼= uM as abelian groups, by means of the
isomorphism φ 7→ φ(u) = uφ(u).
In order to state the analog of Theorem 3.10 for the irrational case we need the following
notation. Let p = e1e2 · · · ∈ E∞ be an irrational infinite path in E. For each i ≥ 0 let Xi(p)
denote the set {f ∈ E1 s(f ) = s(ei+1) and f 6= ei+1} as presented in Definition 2.10, and
define
r(Xi(p)) := {w ∈ E0 w = r(f ) for some f ∈ Xi(p)}.
Finally, for any p ∈ E∞ and any i ≥ 0 recall from Definitions 2.10 that the left LK(E)-ideal
Ji(p) is
Ji(p) = Xf ∈Xi(p)
LK(E)f ∗p∗
i ,
Ji(p) = Mf ∈Xi(p)
where pi denotes the truncation τ≤i(p). As proved in Lemma 2.19, if p is irrational, then
LK(E)f ∗p∗
i
LK(E)r(f ).
∼= Mf ∈Xi(p)
Lemma 3.19. Let p be an irrational infinite path in the finite graph E. Let T denote a
Chen simple module and let t ∈ T . Then there exists a positive integer N = N(t) for which
(Ji(p))t = 0 for all i ≥ N.
Proof. Assume q ∈ E∞ (q can be rational, irrational or a sink) and let α ∈ Path(E) with
length(α) = n. Observe that, for any i ≥ n, one has p∗
i αq = 0 unless pi = ατ≤i−n(q). So if
p∗
i αq 6= 0 for all i ∈ N, we can conclude p = αq. Finally notice that, if there exists N ∈ N
such that p∗
i αq = 0 for any i ≥ N.
N αq = 0, then p∗
(Case 1.) Let T 6= V[p] and let q ∈ E∞, necessarily not tail-equivalent to p, such that
u=1 kuαuτ>iu(q), where
T = V[q]. Consider an element t ∈ T . Then t can be written as t =Ps
h
h
/
/
/
/
22
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
the αu's are in Path(E) and the iu's are in N, and hence αuτ>iu(q) 6= p for any u = 1, . . . , s.
Since any element in (Ji(p))t is a finite sum of expressions of the form kuf ∗p∗
i αuτ>iu(q), and
since αuτ>iu(q) 6= p, by the previous observations we can choose an N = N(t) sufficiently
large such that (Ji(p))t = 0 for any i ≥ N.
written as t = Ps
(Case 2.) On the other hand, let T = V[p] and consider an element t ∈ T . Then t can be
u=1 kuqu, where the qu's are tail equivalent to p. Let f ∈ Xi(p), i ≥ 0; if
i p = f ∗τ>i(p) = 0, by construction of f ∗. If qu 6= p, there exists an integer
iuτ≤iu(q) = 0. Then, by the initial observation, if
(cid:3)
qu = p then f ∗p∗
iu > 0 such that τ≤iu(q) 6= piu and hence p∗
N = Nt = max{iu : u = 1, ..., s}, we conclude that (Ji(p))t = 0 for any i ≥ N.
Lemma 3.20. Let E be a finite graph, and let p be an infinite irrational path in E. Let T
be a Chen simple LK(E)-module. Then HomLK (E)(Ji(p), T ) 6= 0 if and only if r(f ) ∈ U(T )
for some f ∈ Xi(p).
Proof. By standard ring theory, we have the following isomorphisms of abelian groups
HomLK (E)(Ji(p), T ) ∼= HomLK (E)(⊕f ∈Xi(p)LK(E)r(f ), T )
∼= ⊕f ∈Xi(p)HomLK (E)(LK(E)r(f ), T ) ∼= ⊕f ∈Xi(p)r(f )T,
where the second isomorphism holds because Xi(p) is finite, and the final one by Remark 3.18
because each r(f ) is idempotent.
(cid:3)
Theorem 3.21. (Type (3)) Let p be an irrational infinite path in the finite graph E and let
T be any Chen simple LK(E)-module. Then Ext1
LK (E)(V[p], T ) 6= 0 if and only if r(Xi(p)) ∩
U(T ) 6= ∅ for infinitely many i ≥ 0. In such a situation, dimK(Ext1
LK (E)(V[p], T )) is infinite.
Proof. (⇒) Suppose r(Xi(p)) ∩ U(T ) 6= ∅ for at most finitely many i ≥ 0. We seek to show
that every element of HomLK (E)(Ker(ρp), T ) arises as right multiplication by an element of
T . We have HomLK (E)(Ker(ρp), T ) = HomLK (E)(⊕i≥0Ji(p), T ) ∼= Qi≥0 HomLK (E)(Ji(p), T ),
which by Lemma 3.20 and hypothesis equalsQN
i=0 HomLK (E)(Ji(p), T ) for some N ∈ N. For
is an idempotent generator of LK(E)f ∗p∗
i ;
i : f ∈ Xi(p), i ≥ 0} is a set of orthogonal idempotents in LK(E). Every el-
i ); then every
i ).
each i ≥ 0 and f ∈ Xi(p), the element pif f ∗p∗
i
moreover {pif f ∗p∗
ement ϕ of HomLK (E)(LK(E)pif f ∗p∗
element ψ of HomLK (E)(⊕N
So Ext1
LK (E)(V[p], T ) = 0.
(⇐) Conversely, let us see that r(Xi(p)) ∩ U(T ) 6= ∅ for infinitely many i ≥ 0 implies that
there is an element of HomLK (E)(Ker(ρp), T ) which does not arise as a right multiplication
by an element of T . By Lemma 3.20 there exists an increasing sequence (in)n∈N of natural
numbers such that HomLK (E)(Ji(p), T ) 6= 0 if and only if i = in for a suitable n ∈ N. Let
{φi ∈ HomLK (E)(Ji(p), T ) : i ∈ N} be a family of morphisms such that φin 6= 0 for each
n ∈ N. Then
i , T ) is the right multiplication by ϕ(pif f ∗p∗
i=0Ji(p), T ) is the right multiplication by ψ(PN
i=0Pf ∈Xi(p) pif f ∗p∗
ϕ =Yi∈N
φi ∈ HomLK (E)(⊕i∈NJi(p), T )
is a morphism which is not, by Lemma 3.19, right multiplication by element of T .
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
23
To establish the final statement, consider an increasing sequence (in)n∈N of natural numbers
and a family {φin ∈ HomLK (E)(Jin(p), T ) : n ∈ N} of nonzero morphisms. Define for each
prime z ∈ N the morphism Ψz ∈ HomLK (E)(Ker(ρp), T ) as follows: for each j ≥ 0,
Ψz(Jij (p)) = ψij (Jij (p)) if z divides j, while Ψz(Jℓ(p)) = 0 otherwise.
Observe that π(Ψz) 6= 0, since Ψz has infinitely many nonzero components. Finally, {π(Ψz) z ∈
N, z prime} is a set of linearly independent elements of Ext1
LK (E)(V[p], T )), as follows. Let F
be a finite subset of primes in N, and assumePz∈F kzπ(Ψz) = 0; then π(Pz∈F kzΨz) = 0 and
therefore Pz∈F kzΨz is a right multiplication by an element t of T . By Lemma 3.19 there
exists a positive integer N = N(t) for which (Ji(p))t = 0 for all i ≥ N. For each prime z ∈ F ,
let mz a natural number greater than N; then
0 = (Jzmz (p))t =Xz∈F
kzΨz(Jzmz (p)) = kzΨz(Jzmz (p))
and hence kz = 0. Hence dimK(Ext1
LK (E)(V[p], T )) is infinite.
(cid:3)
We emphasize the fact that Theorems 3.10 and 3.21 allow us to compute the dimension of the
Ext1-groups between two Chen simple modules completely and solely in terms of properties
of the graph E.
Example 3.22. We again revisit the graph R2 and irrational infinite path q = ef ef f ef f f e · · ·
of Example 1.1. Let T = V[q]. Since clearly U(T ) = {v} and r(Xi) = {v} for all i ∈ N as
well, Theorem 3.21 yields that dimK(Ext1
LK (E)(V[q], V[q])) is infinite.
Example 3.23. With the statement of Theorem 3.21 as motivation, we give examples of
graphs En, an irrational infinite path p in En, and Chen simple LK(En)-modules T having
r(Xi(p)) ∩ U(T ) 6= ∅ for only finitely many i ∈ Z+. For n ∈ N consider the graph En given
by
•v1
e1
•v2
e2
· · ·
/ •vn
en /
s❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣
u❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥
❉
❉
❉
❉
❉
❉
❉
!❉
•w
en+1
•vn+1
•v
e
f
h
g
Let p denote the irrational infinite path e1e2 · · · enen+1ef ef f ef f f e · · · . Let T1 be the Chen
simple LK(E)-module V[g∞], and let T2 denote the Chen simple LK(E)-module V[q] correspond-
ing to the irrational infinite path q = ghghhghhhg · · · . Then for j = 1, 2, r(Xi(p)) ∩ U(Tj) is
nonempty (indeed, equals {w}) precisely when 0 ≤ i ≤ n.
Consequently, by Theorem 3.21, Ext1
LK (E)(V[p], Tj) = {0} for j = 1, 2.
We conclude the article by demonstrating the existence of indecomposable LK(E)-modules
of prescribed finite length, in case E is a finite graph which contains cycles. Recall that
a module M is called uniserial in case the lattice of submodules of M is totally ordered.
In particular, any uniserial module is indecomposable. Moreover, the radical Rad(M) of a
uniserial module M is the unique maximal submodule of M, hence M/ Rad M is simple.
/
/
!
/
/
/
/
u
/
/
s
R
R
j
j
4
4
24
GENE ABRAMS∗, FRANCESCA MANTESE, AND ALBERTO TONOLO
Lemma 3.24. [8, Lemma 16.1 with Proposition 16.2] Let R be any unital ring. Let U be a
uniserial left R-module of finite length, and X a simple left R-module. Consider the morphism
ψ : Ext1
R(X, U) is uniserial if and only
if it does not belong to Ker ψ.
R(X, U/ Rad U). An extension in Ext1
R(X, U) → Ext1
In particular, if R is hereditary, there exists a uniserial extension of U by X if and only if
R(X, U/ Rad U) 6= 0.
Ext1
As observed in Remark 1.4, LK(E) is hereditary for any row-finite graph E. So Lemma
3.24 gives the following:
Corollary 3.25. Let E be a finite graph. If S is a Chen simple LK(E)-module such that
LK (E)(S, S) 6= 0 and L is a uniserial LK(E)-module such that L/ Rad(L) ∼= S, then there
Ext1
exists a uniserial LK(E)-module M which is an extension of L by S.
In particular, for any n ∈ N there exists a uniserial LK(E)-module of length n, all of whose
composition factors are isomorphic to S.
Proof. The first statement follows directly from Lemma 3.24 and the hereditariness of LK(E).
In order to show the existence of uniserial modules of arbitrary length, first observe that any
non-zero element of the abelian group Ext1
LK (E)(S, S) 6= 0 corresponds to an indecomposable
uniserial module L2 of length 2, with Rad(L2) ∼= S and L2/Rad(L2) ∼= S. Then, by applying
the first statement, there exists a uniserial module L3 of length 3 which is an extension of L2
by S. Since Rad(L3) ∼= L2 and hence L3/ Rad(L3) ∼= S, we can proceed by induction.
(cid:3)
Observe that if E contains a simple closed path d, by Corollary 3.11 the module S = V[d∞]
satisfies Ext1
LK (E)(S, S) 6= 0 and hence Corollary 3.25 applies.
References
[1] G. Abrams, P. Ara, M. Siles Molina. "Leavitt path algebras". Lecture Notes Series in Mathematics,
Springer Verlag, to appear.
[2] G. Abrams, G. Aranda Pino, The Leavitt path algebra of a graph, J. Algebra 293(2) (2005), 319 - 334.
[3] P. Ara, M.A. Moreno, E. Pardo, Non-stable K-theory for graph algebras, Algebr. Represent. Theor. 10(2)
(2007), 157 - 178.
[4] P. Ara, K.M. Rangaswamy, Finitely presented simple modules over Leavitt path algebras, J. Algebra 417
(2014), 333 - 352.
[5] G. Aranda Pino, D. Martin Barquero, C. Martin Gonzalez, M. Siles Molina, Socle theory for leavitt path
algebras of arbitrary graphs, Rev. Mat. Iberoamericana 26 (2010), no. 2, 611 - 638.
[6] X.W. Chen, Irreducible representations of Leavitt path algebras, Forum Math., to appear. arXiv:
1108.3726.
[7] W. Leavitt, The module type of a ring, Trans. Amer. Math. Soc. 103(1) (1962), 113 - 130.
[8] L.S. Levy, J. C. Robson. "Hereditary Noetherian prime rings and idealizers." Mathematical Surveys and
Monographs Volume 174. American Mathematical Society, Providence, USA. 2011. ISBN 978-0-8218-
5350-4.
[9] C. Weibel. "An introduction to homological algebra." Cambridge Studies in Advanced Mathematics
(Book 38). Cambridge University Press, Cambridge, UK. ISBN 978-0-5215-5987-4.
EXTENSIONS OF SIMPLE MODULES OVER LEAVITT PATH ALGEBRAS
25
Department of Mathematics, University of Colorado, Colorado Springs, CO 80918 U.S.A.
E-mail address: [email protected]
Dipartimento di Informatica, Universit`a degli Studi di Verona, I-37134 Verona, Italy
E-mail address: [email protected]
Dipartimento Matematica, Universit`a degli Studi di Padova, I-35121, Padova, Italy
E-mail address: [email protected]
|
1407.3978 | 2 | 1407 | 2015-11-03T03:39:21 | Some structure theories of Leibniz triple systems | [
"math.RA"
] | In this paper, we investigate the Leibniz triple system $T$ and its universal Leibniz envelope $U(T)$. The involutive automorphism of $U(T)$ determining $T$ is introduced, which gives a characterization of the $\Z_2$-grading of $U(T)$. We give the relationship between the solvable radical $R(T)$ of $T$ and $Rad(U(T))$, the solvable radical of $U(T)$. Further, Levi's theorem for Leibniz triple systems is obtained. Moreover, the relationship between the nilpotent radical of $T$ and that of $U(T)$ is studied. Finally, we introduce the notion
of representations of a Leibniz triple system. | math.RA | math |
Some structure theories of Leibniz triple
systems
Yao Ma1, Liangyun Chen2∗
1School of Mathematical Sciences, University of Science and Technology of China,
2School of Mathematics and Statistics, Northeast Normal University,
Hefei 230026, CHINA
Changchun 130024, CHINA
Abstract
In this paper, we investigate the Leibniz triple system T and its universal Leibniz
envelope U (T ). The involutive automorphism of U (T ) determining T is introduced,
which gives a characterization of the Z2-grading of U (T ). We give the relationship
between the solvable radical R(T ) of T and Rad(U (T )), the solvable radical of
U (T ). Further, Levi's theorem for Leibniz triple systems is obtained. Moreover, the
relationship between the nilpotent radical of T and that of U (T ) is studied. Finally,
we introduce the notion of representations of a Leibniz triple system.
Key words: Leibniz triple system, Lie triple system, Levi's theorem, solvable
radical, nilpotent radical
MSC(2010): 17A32, 17A40, 17A60, 17A65
1
Introduction
The notion of Leibniz algebras was introduced by Loday [14], which is a "nonan-
tisymmetric" generalization of Lie algebras. A Leibniz algebra is a vector space equipped
with a bilinear bracket satisfying the (right) Leibniz identity
[[x, y], z] = [[x, z], y] + [x, [y, z]].
(1.1)
A Leibniz algebra whose bracket is antisymmetric is a Lie algebra. In fact, such algebras
were considered by Bloh in 1965, who called them D-algebras [6]. Later, Loday introduced
* Corresponding author(L. Chen): [email protected].
Supported by NNSF of China (Nos. 11171055 and 11471090) and China Postdoctoral Science Foun-
dation(No. 159347).
1
this class of algebras to search an "obstruction" to the periodicity in algebraic K-theory.
So far many results of Leibniz algebras have been studied, including extending important
theorems in Lie algebras: there are analogs of Lie's theorem, Engel's theorem, Levi's
theorem and Cartan's criterion for Leibniz algebras [1 -- 5, 8 -- 10, 15, 17].
Leibniz triple systems were introduced by Bremner and S´anchez-Ortega [7]. Leib-
niz triple systems are defined in a functorial manner using the Kolesnikov-Pozhidaev
algorithm, which takes the defining identities for a variety of algebras and produces the
defining identities for the corresponding variety of dialgebras [12]. As an application of
this algorithm, it is showed that associative dialgebras and Leibniz algebras can be ob-
tained from associative and Lie algebras, respectively. In [7], Leibniz triple systems were
obtained by applying the Kolesnikov-Pozhidaev algorithm to Lie triple systems. Further-
more, Leibniz triple systems are related to Leibniz algebras in the same way that Lie
triple systems related to Lie algebras. So it is natural to prove analogs of results from the
theory of Lie triple systems to Leibniz triple systems.
This paper proceeds as follows. Section 2 devotes to some basic facts about a Leibniz
triple system T . In Section 3, we give the involutive automorphism of U(T ) determining
T , where U(T ) is the universal Leibniz envelope of T , using which to investigate the
connection between automorphisms of T and those of U(T ), and to describe the Z2-
graded subspace of U(T ). In Section 4, the solvable radical of a Leibniz triple system
and the definition of a semisimple Leibniz triple system are introduced.
In Section 5,
Levi's theorem is extended to the case of Leibniz triple systems. In Section 6, the notion
of nilpotent ideals of T is introduced. Section 7 is devoted to the representations of a
Leibniz triple system.
In the sequel, all vector spaces are finite-dimensional and defined over a fixed but
arbitrarily chosen field K of characteristic 0.
2 Preliminaries
Definition 2.1. [13, 16] A vector space T together with a trilinear map (a, b, c) 7→ [abc]
is called a Lie triple system if
[aac] = 0,
[abc] + [bca] + [cab] = 0,
[ab[cde]] = [[abc]de] + [c[abd]e] + [cd[abe]],
for all a, b, c, d, e ∈ T .
Definition 2.2. [7] A Leibniz triple system (LeibTS for short) is a vector space T
with a trilinear operation {·, ·, ·} : T × T × T → T satisfying
{a{bcd}e} = {{abc}de} − {{acb}de} − {{adb}ce} + {{adc}be},
{ab{cde}} = {{abc}de} − {{abd}ce} − {{abe}cd} + {{abe}dc}.
(2.2)
(2.3)
2
A Lie triple system gives a LeibTS with the same ternary product. If L is a Leibniz
algebra with product [·, ·], then L becomes a LeibTS by putting {xyz} = [[xy]z]. The
following examples of LeibTSs are induced by Leibniz algebras.
Example 2.3. (1) Let T = Khx, yi with multiplications given by {xyy} = {yyy} = x,
and the rest are 0. Then T is a LeibTS.
(2) Let T = Khx, y, zi with multiplications given by {xzz} = {yzz} = {zzz} = x,
and the rest are 0. Then T is a LeibTS.
(3) Let T = Khx, a, b, c, di with the following multiplications
{axx} = −{xax} = a,
{bax} = {cxx} = c,
{abx} = {xab} = {axb} = {xcx} = {xdx} = {dxx} = d,
and the rest are 0. Then T is a LeibTS.
More examples refer to [7].
Proposition 2.4. [7] Let T be a LeibTS. Then the following identities hold.
{a{bcd}e} + {a{cbd}e} = 0, {ab{cde}} + {ab{dce}} = 0,
(2.4)
{a{bcd}e}+ {a{cdb}e}+ {a{dbc}e} = 0, {ab{cde}}+ {ab{dec}}+ {ab{ecd}} = 0, (2.5)
{{cde}ba} − {{cde}ab} − {{cba}de} + {{cab}de} − {c{abd}e} − {cd{abe}} = 0.
(2.6)
Remark 2.5. Any two of Equations (2.2), (2.3) and (2.6) induce the rest one.
Theorem 2.6. [7] Suppose that T is a LeibTS with product {xyz}. The universal
Leibniz envelope of T is the quotient algebra U(T ) = A(T )/I(T ) with A(T ) the free
Leibniz algebra and I(T ) the ideal generated by the elements [[x, y], z] − {xyz}. Then
U(T ) = T ⊕ (T ⊗ T ) together with product [·, ·] satisfies
[a, b] = a ⊗ b,
[a ⊗ b, c] = {abc},
[a, b ⊗ c] = {abc} − {acb},
[a ⊗ b, c ⊗ d] = {abc} ⊗ d − {abd} ⊗ c.
It is straightforward to verify that U(T ) is a Leibniz algebra and T ⊗ T is a subalgebra of
U(T ).
Generally speaking, U(T ) is not a Lie algebra even if T is a Lie triple system, so we
try to add a condition.
Proposition 2.7. Let T be a Lie triple system and U(T ) be defined as in Theorem 2.6.
then U(T ) is a Lie algebra.
Denote by K(T ) = {P a ⊗ b ∈ T ⊗ T [P a ⊗ b, T ] = [T,P a ⊗ b] = 0}. If K(T ) = 0,
Proof. We need only to show that for all x = a +P b ⊗ c ∈ U(T ), [x, x] = 0, since
U(T ) is already a Leibniz algebra by Theorem 2.6.
3
Note that T is a Lie triple system. Take a, b, c ∈ T , then one has
[a ⊗ a, b] = {aab} = 0,
[b, a ⊗ a] = {baa} − {baa} = 0,
which imply [a, a] = a ⊗ a ∈ K(T ) = 0. It follows from (2.3) and the definition of Lie
triple systems that
[[a ⊗ b, a ⊗ b], c] = [{aba} ⊗ b − {abb} ⊗ a, c] = {{aba}bc} − {{abb}ac}
={ab{abc}} + {{abc}ab} − {{abc}ba} = {ab{abc}} + {{abc}ab} + {b{abc}a} = 0,
and
[c, [a ⊗ b, a ⊗ b]] = [c, {aba} ⊗ b − {abb} ⊗ a]
={c{aba}b} − {cb{aba}} − {c{abb}a} + {ca{abb}}
={b{aba}c} + {{abb}ac} = {ab{bac}} − {ba{abc}} = 0,
that is, [a ⊗ b, a ⊗ b] ∈ K(T ) = 0.
i=1 bi ⊗ ci ∈ U(T ), then
n
n
bi ⊗ ci, a +
bi ⊗ ci#
Xi=1
Now take a +Pn
"a +
Xi=1
Xi=1
=a ⊗ a +
n
=Xi<j
({bicia} + {abici} − {acibi}) +
([bi ⊗ ci, bj ⊗ cj] + [bj ⊗ cj, bi ⊗ ci]).
n
[bi ⊗ ci, bj ⊗ cj]
Xi,j=1
Note that for all d ∈ T , [d, [bi ⊗ ci, bj ⊗ cj] + [bj ⊗ cj, bi ⊗ ci]] = 0 by using (1.1), and
[[bi ⊗ ci, bj ⊗ cj], d] = [{bicibj} ⊗ cj − {bicicj} ⊗ bj, d]
={{bicibj}cjd} − {{bicicj}bjd} = {bici{bjcjd}} − {bjcj{bicid}}.
Then [[bi ⊗ ci, bj ⊗ cj] + [bj ⊗ cj, bi ⊗ ci], d] = 0, and so
[bi ⊗ ci, bj ⊗ cj] + [bj ⊗ cj, bi ⊗ ci] ∈ K(T ) = 0.
Thus [x, x] = 0.
Definition 2.8. Let I be a subspace of a LeibTS T . Then I is called a subsystem of T ,
if {III} ⊆ I; I is called an ideal of T , if {IT T } + {T IT } + {T T I} ⊆ I.
Let I be an ideal of a LeibTS T and define some notations as follows.
i(I) = I ⊗ T + T ⊗ I,
j(I) =nX a ⊗ b ∈ T ⊗ T (cid:12)(cid:12)(cid:12) hX a ⊗ b, Ti +hT,X a ⊗ bi ⊆ Io ,
I(I) = I + i(I), J (I) = I + j(I).
It is easy to see that K(T ) = j(0) = J (0). The notion of K(T ) will be used in Section 4,
too.
4
Theorem 2.9. If I is an ideal of a LeibTS T , then
(1) i(I) ⊆ j(I), I(I) ⊆ J (I).
(2) i(I) and j(I) are ideals of T ⊗ T .
(3) I(I) and J (I) are ideals of U(T ).
Proof.
(1) follows from
[i(I), T ] = [I ⊗ T + T ⊗ I, T ] ⊆ {IT T } + {T IT } ⊆ I
and
[T, i(I)] = [T, I ⊗ T + T ⊗ I] ⊆ {T IT } + {T T I} ⊆ I.
(2) i(I) is an ideal of T ⊗ T since
[i(I), T ⊗ T ] = [I ⊗ T + T ⊗ I, T ⊗ T ] ⊆ {IT T } ⊗ T + {T IT } ⊗ T ⊆ I ⊗ T ⊆ i(I)
and
[T ⊗ T, i(I)] = [T ⊗ T, I ⊗ T +T ⊗ I] ⊆ {T T I} ⊗ T +{T T T } ⊗ I ⊆ I ⊗ T +T ⊗ I = i(I).
It is sufficient to prove that for all c ∈ T ,
To show that j(I) is an ideal of T ⊗ T , we take P t ⊗ s ∈ T ⊗ T and P a ⊗ b ∈ j(I).
hhX a ⊗ b,X t ⊗ si , ci ∈ I,
hhX t ⊗ s,X a ⊗ bi , ci ∈ I,
hc,hX a ⊗ b,X t ⊗ sii ∈ I,
hc,hX t ⊗ s,X a ⊗ bii ∈ I.
In fact, it follows from (2.3) that
hhX a ⊗ b,X t ⊗ si , ci =X([{abt} ⊗ s − {abs} ⊗ t, c])
=X({{abt}sc} − {{abs}tc}) =X({[a ⊗ b, t]sc} − {[a ⊗ b, s]tc}) ∈ I,
hc,hX a ⊗ b,X t ⊗ sii =X[c, {abt} ⊗ s − {abs} ⊗ t]
=X({c{abt}s} − {cs{abt}} − {c{abs}t} + {ct{abs}})
=X({c[a ⊗ b, t]s} − {cs[a ⊗ b, t]} − {c[a ⊗ b, s]t} + {ct[a ⊗ b, s]}) ∈ I,
hhX t ⊗ s,X a ⊗ bi , ci =X([{tsa} ⊗ b − {tsb} ⊗ a, c])
=X({{tsa}bc} − {{tsb}ac}) =X({ts{abc}} + {{tsc}ab} − {{tsc}ba})
=X({ts[a ⊗ b, c]} + [{tsc}, a ⊗ b]) ∈ I.
5
and
Moreover, by (1.1), [c, [P t ⊗ s,P a ⊗ b]] = − [c, [P a ⊗ b,P t ⊗ s]] ∈ I.
(3) It is routine to prove that I(I) is an ideal of U(T ) by the facts of I being an ideal
of T and i(I) being an ideal of T ⊗ T . J (I) is an ideal of U(T ) since
[J (I), U(T )] =[I + j(I), T + T ⊗ T ] = I ⊗ T + [j(I), T ] + [I, T ⊗ T ] + [j(I), T ⊗ T ]
⊆i(I) + I + j(I) = I + j(I) = J (I),
and
[U(T ), J (I)] =[T + T ⊗ T, I + j(I)] = T ⊗ I + [T ⊗ T, I] + [T, j(I)] + [T ⊗ T, j(I)]
⊆i(I) + I + j(I) = J (I).
Thus the proof is completed.
3 The involutive automorphism of U (T ) determining
T
Theorem 2.6 says that U(T ) is a Z2-graded Leibniz algebra with U(T )¯0 = T ⊗ T and
U(T )¯1 = T . Then every Leibniz triple system T is the ¯1 component of a Z2-graded Leibniz
algebra. Moreover, for a Z2-graded Leibniz algebra L = L¯0 ⊕ L¯1, L¯1 has the structure of
a Leibniz triple system.
The Z2-grading of U(T ) can be characterized by an involutive automorphism. In what
follows we will use this involutive automorphism of U(T ) to investigate the connection
between automorphisms of T and those of U(T ), and describe the Z2-graded subspace of
U(T ).
Theorem 3.1. Let T be a LeibTS and U(T ) the universal Leibniz envelope of T . Then
there is a unique involutive automorphism θ : U(T ) → U(T ) such that
T = {x ∈ U(T ) θ(x) = −x}.
that θ is linear and θ2 = id. Note that
Proof. For a +P b ⊗ c ∈ U(T ), define θ (a +P b ⊗ c) = −a +P b ⊗ c. It is clear
hθ(cid:16)a +X b ⊗ c(cid:17) , θ(cid:16)a′ +X b′ ⊗ c′(cid:17)i =h−a +X b ⊗ c, −a′ +X b′ ⊗ c′i
=a ⊗ a′ −X{bca′} −X{ab′c′} +X{ac′b′} +X{bcb′} ⊗ c′ −X{bcc′} ⊗ b′,
and
θ(cid:16)ha +X b ⊗ c, a′ +X b′ ⊗ c′i(cid:17)
=θ(cid:16)a ⊗ a′ +X{bca′} +X{ab′c′} −X{ac′b′} +X{bcb′} ⊗ c′ −X{bcc′} ⊗ b′(cid:17)
6
=a ⊗ a′ −X{bca′} −X{ab′c′} +X{ac′b′} +X{bcb′} ⊗ c′ −X{bcc′} ⊗ b′.
Then θ is an involutive automorphism of U(T ) and T = {x ∈ U(T ) θ(x) = −x}.
If φ is another involutive automorphism of U(T ) with T = {x ∈ U(T ) φ(x) = −x},
then θT = φT = −idT and
θ(cid:16)a +X b ⊗ c(cid:17)
=θ(cid:16)a +X[b, c](cid:17) = θ(a) +X[θ(b), θ(c)] = φ(a) +X[φ(b), φ(c)] = φ(cid:16)a +X[b, c](cid:17)
=φ(cid:16)a +X b ⊗ c(cid:17) ,
which implies θ = φ, so θ is unique. θ is called the involutive automorphism of U(T )
determining T .
Proposition 3.2. Suppose that T is a LeibTS and f is an automorphism of T . Then
there is a unique automorphism f of U(T ) such that f = f T and f θ = θ f , where θ is the
involutive automorphism of U(T ) determining T .
clear that f is well defined, f T = f and f is a linear isomorphism. Moreover, we have
Proof. Define f : U(T ) → U(T ) by f (a +P b ⊗ c) = f (a) +P f (b) ⊗ f (c). It is
f(cid:16)ha +X b ⊗ c, a′ +X b′ ⊗ c′i(cid:17)
= f(cid:16)a ⊗ a′ +X{bca′} +X{ab′c′} −X{ac′b′} +X{bcb′} ⊗ c′ −X{bcc′} ⊗ b′(cid:17)
=f (a) ⊗ f (a′) +X{f (b)f (c)f (a′)} +X{f (a)f (b′)f (c′)} −X{f (a)f (c′)f (b′)}
+X{f (b)f (c)f (b′)} ⊗ f (c′) −X{f (b)f (c)f (c′)} ⊗ f (b′)
=hf (a) +X f (b) ⊗ f (c), f (a′) +X f (b′) ⊗ f (c′)i
=h f(cid:16)a +X b ⊗ c(cid:17) , f(cid:16)a′ +X b′ ⊗ c′(cid:17)i ,
and
f θ(cid:16)a +X b ⊗ c(cid:17) = f(cid:16)−a +X b ⊗ c(cid:17) = −f (a) +X f (b) ⊗ f (c)
=θ(cid:16)f (a) +X f (b) ⊗ f (c)(cid:17) = θ f(cid:16)a +X b ⊗ c(cid:17) .
Hence f is an automorphism and f θ = θ f . f is unique since U(T ) is generated by T .
Proposition 3.3. If f is an automorphism of U(T ) satisfying f θ = θf , then f (T ) ⊆ T
and f T is an automorphism of T .
Proof. For all a ∈ T , f θ(a) = f (−a) = −f (a). Since f θ = θf , it follows that
It is clear that f T is an
θf (a) = −f (a), which implies f (a) ∈ T by Theorem 3.1.
injective endomorphism of T . Note that
f (T ⊗ T ) = f ([T, T ]) = [f (T ), f (T )] = f (T ) ⊗ f (T ) ⊆ T ⊗ T,
and f is surjective, so f T is surjective. Thus f T is an automorphism of T .
7
Note that U(T ) is a Z2-graded Leibniz algebra with U(T )¯0 = T ⊗ T and U(T )¯1 = T .
Then a subspace V of U(T ) is Z2-graded if and only if V = V ∩ T + V ∩ (T ⊗ T ). Suppose
that is, V = V ∩ T + V ∩ (T ⊗ T ). Conversely, if V is a subspace of U(T ) such that
that V is a θ-invariant subspace of U(T ). Then for a +P b ⊗ c ∈ V , θ (a +P b ⊗ c) =
−a +P b ⊗ c ∈ V . Since chK = 0, it follows a ∈ V ∩ T and P b ⊗ c ∈ V ∩ (T ⊗ T ),
V = V ∩ T + V ∩ (T ⊗ T ), then a +P b ⊗ c ∈ V implies that a ∈ V and P b ⊗ c ∈ V , so
θ (a +P b ⊗ c) = −a +P b ⊗ c ∈ V and V is a θ-invariant subspace of U(T ).
Therefore, V being a θ-invariant subspace of U(T ) is equivalent to that V is a Z2-
graded subspace of U(T ).
Theorem 3.4. Let V be a θ-invariant subspace of U(T ). If V is an ideal of U(T ), then
V ∩ T is an ideal of T and V ∩ (T ⊗ T ) is an ideal of T ⊗ T such that
i(V ∩ T ) ⊆ V ∩ (T ⊗ T ) ⊆ j(V ∩ T ),
I(V ∩ T ) ⊆ V ⊆ J (V ∩ T ).
Proof. Denote I = V ∩ T and M = V ∩ (T ⊗ T ). Then V = I + M. Using the
condition that V is an ideal of U(T ), one has
[V, U(T )] = [I + M, T + T ⊗ T ] = I ⊗ T + [M, T ] + [I, [T, T ]] + [M, T ⊗ T ] ⊆ I + M,
[U(T ), V ] = [T + T ⊗ T, I + M] = T ⊗ I + [[T, T ], I] + [T, M] + [T ⊗ T, M] ⊆ I + M.
It follows that
I ⊗ T + T ⊗ I + [M, T ⊗ T ] + [T ⊗ T, M] ⊆ M,
[M, T ] + [I, [T, T ]] + [[T, T ], I] + [T, M] ⊆ I.
Then i(I) ⊆ M ⊆ j(I) and M is an ideal of T ⊗ T . Note that
{T T I} = [[T, T ], I] ⊆ I,
{T IT } = [[T, I], T ] = [T ⊗ I, T ] ⊆ [M, T ] ⊆ I,
{IT T } = [[I, T ], T ] = [I ⊗ T, T ] ⊆ [M, T ] ⊆ I.
Thus I is an ideal of T and I(I) ⊆ I + M ⊆ J (I).
4 The solvability of a Leibniz triple system
Given an arbitrary Leibniz algebra L, the derived sequence of an ideal I of L is defined
as
Suppose that I is an ideal of a LeibTS T . Let I [0] = I and for n ≥ 1,
I (0) = I, I (n+1) = [I (n), I (n)],
∀n ≥ 1.
I [n+1] = {T I [n]I [n]} + {I [n]T I [n]} + {I [n]I [n]T }.
8
Proposition 4.1. If I is an ideal of a LeibTS T , then I [n] is also an ideal of T and
I ⊇ I [1] ⊇ · · · ⊇ I [n] ⊇ · · · , ∀n ∈ N.
Proof. First we show {I [1]T T } ⊆ I [1]. By (2.2) and (2.3), it follows that
{{T II}T T } ⊆ {T I{IT T }}+{{T IT }IT }+{{T IT }T I} ⊆ {T II}+{IIT }+{IT I} ⊆ I [1],
{{IT I}T T } ⊆ {IT {IT T }} + {{IT T }IT } + {{IT T }T I} ⊆ {IT I} + {IIT } ⊆ I [1],
{{IIT }T T } ⊆ {{IT I}T T } + {{IT T }IT } + {I{IT T }T } ⊆ I [1] + {IIT } = I [1],
which imply {I [1]T T } ⊆ I [1]. Similarly, one can prove {T I [1]T } ⊆ I [1] and {T T I [1]} ⊆ I [1].
So I [1] is an ideal of T . Suppose that I [k] is an ideal of T . Then I [k+1] = (I [k])[1] is an ideal
of T . Hence I [n] is an ideal of T for all n ∈ N. It is clear that I ⊇ I [1] ⊇ · · · ⊇ I [n] ⊇ · · · ,
∀n ∈ N.
Definition 4.2. An ideal I of a LeibTS T is called solvable, if there is a positive integer
k such that I [k] = 0.
Theorem 4.3. Let I be an ideal of a LeibTS T . Then the following statements are
equivalent.
(1) I is a solvable ideal of T .
(2) I(I) is a solvable ideal of U(T ).
(3) J (I) is a solvable ideal of U(T ).
Proof.
(1)⇒(2). We will use the induction to show
I(I)(2n) ⊆ I [n] + I [n] ⊗ T.
Consider the base step n = 1. We obtain
I(I)(1) = [I(I), I(I)] = [I + i(I), I + i(I)] = I ⊗ I + [i(I), I] + [I, i(I)] + [i(I), i(I)].
Note that
[i(I), I] = {IT I} + {T II} ⊆ I [1],
[i(I), i(I)] = [I ⊗ T + T ⊗ I, I ⊗ T + T ⊗ I]
[I, i(I)] ⊆ {IIT } + {IT I} ⊆ I [1],
⊆{IT I} ⊗ T + {IT T } ⊗ I + {T II} ⊗ T + {T IT } ⊗ I ⊆ I [1] ⊗ T + I ⊗ I.
Hence I(I)(1) ⊆ I [1] + I [1] ⊗ T + I ⊗ I. So
I(I)(2) = [I(I)(1), I(I)(1)] ⊆ [I [1] + I [1] ⊗ T + I ⊗ I, I [1] + I [1] ⊗ T + I ⊗ I]
⊆I [1] ⊗ I [1] + {I [1]T I [1]} + {III [1]} + {I [1]I [1]T } + {I [1]T I [1]} ⊗ T + {I [1]T T } ⊗ I [1]
+ {III [1]} ⊗ T + {IIT } ⊗ I [1] + {I [1]II} + {I [1]T I} ⊗ I + {III} ⊗ I
9
⊆I [1] + I [1] ⊗ T.
Suppose I(I)(2k) ⊆ I [k] + I [k] ⊗ T . Then
I(I)(2k+1) =[I(I)(2k), I(I)(2k)] ⊆ [I [k] + I [k] ⊗ T, I [k] + I [k] ⊗ T ]
⊆I [k] ⊗ I [k] + {I [k]T I [k]} + {I [k]I [k]T } + {I [k]T I [k]} ⊗ T + {I [k]T T } ⊗ I [k]
⊆I [k+1] + I [k] ⊗ I [k] + I [k+1] ⊗ T.
Hence
I(I)(2k+2) = [I(I)(2k+1), I(I)(2k+1)]
⊆[I [k+1] + I [k] ⊗ I [k] + I [k+1] ⊗ T, I [k+1] + I [k] ⊗ I [k] + I [k+1] ⊗ T ]
⊆I [k+1] ⊗ I [k+1] + {I [k]I [k]I [k+1]} + {I [k+1]T I [k+1]} + {I [k+1]I [k]I [k]} +{I [k]I [k]I [k]} ⊗ I [k]
+ {I [k+1]T I [k]} ⊗ I [k] + {I [k+1]I [k+1]T } +{I [k]I [k]I [k+1]} ⊗ T
+{I [k]I [k]T } ⊗ I [k+1] +{I [k+1]T I [k+1]} ⊗ T +{I [k+1]T T } ⊗ I [k+1]
⊆I [k+1] + I [k+1] ⊗ T.
Thus I(I)(2n) ⊆ I [n] + I [n] ⊗ T by induction, which implies that I(I) is solvable.
(2)⇒(1). It is sufficient to prove I [n] ⊆ I(I)(n) for all n ∈ N. In fact,
I [1] ={T II} + {IT I} + {IIT } = [[T, I], I] + [[I, T ], I] + [[I, I], T ]
⊆[[T, I], I] + [[I, T ], I] + [I, [I, T ]] ⊆ [i(I), I] + [I, i(I)] ⊆ [I(I), I(I)] = I(I)(1).
Suppose I [k] ⊆ I(I)(k). Then
I [k+1] ={T I [k]I [k]} + {I [k]T I [k]} + {I [k]I [k]T }
⊆[[T, I [k]], I [k]] + [[I [k], T ], I [k]] + [I [k], [I [k], T ]]
⊆[[T, I(I)(k)], I(I)(k)] + [[I(I)(k), T ], I(I)(k)] + [I(I)(k), [I(I)(k), T ]]
⊆[I(I)(k), I(I)(k)] = I(I)(k+1).
Hence I [n] ⊆ I(I)(n) follows from the induction. So (1)⇔(2).
(3)⇒(2). It is clear since I(I) ⊆ J (I).
(2)⇒(3). Note that
[J (I), U(T )] = [I + j(I), T + T ⊗ T ]
⊆I ⊗ T + [j(I), T ] + {IT T } + [j(I), T ] ⊗ T
⊆I ⊗ T + I ⊆ I(I).
Then J (I)(1) = [J (I), J (I)] ⊆ [J (I), U(T )] ⊆ I(I). Suppose J (I)(k) ⊆ I(I)(k−1). Then
J (I)(k+1) = [J (I)(k), J (I)(k)] ⊆ [I(I)(k−1), I(I)(k−1)] = I(I)(k).
It follows that J (I)(n+1) ⊆ I(I)(n) by induction and so the solvability of I(I) implies the
solvability of J (I). Thus (2)⇔(3).
Therefore, the above three items are equivalent and the proof is completed.
10
Corollary 4.4. Let V = I + M be a θ-invariant ideal of U(T ). Then I is a solvable ideal
of T if and only if V is a solvable ideal of U(T ). In particular, T is a solvable LeibTS if
and only if U(T ) is a solvable Leibniz algebra.
Proposition 4.5. Suppose that I and J are solvable ideals of a LeibTS T . Then I + J
is a solvable ideal of T .
Proof.
It is clear that I + J is an ideal of T . Now we use the induction to prove
(I + J)[n] ⊆ I [n] + J [n] + I ∩ J. The base step n = 1 is true, for
(I + J)[1] = {T, I + J, I + J} + {I + J, T, I + J} + {I + J, I + J, T } ⊆ I [1] + J [1] + I ∩ J.
Suppose (I + J)[k] ⊆ I [k] + J [k] + I ∩ J. Then
(I + J)[k+1] = {T (I + J)[k](I + J)[k]} + {(I + J)[k]T (I + J)[k]} + {(I + J)[k](I + J)[k]T }
⊆{T, I [k] + J [k] + I ∩ J, I [k] + J [k] + I ∩ J} + {I [k] + J [k] + I ∩ J, T, I [k] + J [k] + I ∩ J}
+ {I [k] + J [k] + I ∩ J, I [k] + J [k] + I ∩ J, T }
⊆I [k+1] + J [k+1] + I ∩ J.
So there is n1 ∈ N such that (I + J)[n1] ⊆ I ∩ J. Note that I ∩ J is a solvable ideal of T ,
which implies I + J is solvable.
Definition 4.6. Let R(T ) denote the sum of all solvable ideals in T . Then R(T ) is the
unique maximal solvable ideal by Proposition 4.5, called the solvable radical of T .
For a Leibniz algebra L, there is a corresponding Lie algebra L/Ker(L), where Ker(L)
is a two-side ideal of L spanned by {[x, x] x ∈ L}. Similarly, for a LeibTS T , we define
Ker(T ) to be the subspace of T spanned by all elements in T of the form {abc} − {acb} +
{bca}, for all a, b, c ∈ T .
Theorem 4.7. Let T be a LeibTS and U(T ) be its universal Leibniz envelope. Then
(1) Ker(T ) is an ideal of T satisfying {T T Ker(T )} = {T Ker(T )T } = 0.
(2) T /Ker(T ) is a Lie triple system. Moreover, T is a Lie triple system if and only if
Ker(T ) = 0.
(3) Ker(T ) ⊆ R(T ).
(4) Ker(U(T )) ∩ T = Ker(T ).
(5) I(Ker(T )) ⊆ Ker(U(T )) ⊆ J (Ker(T )).
Proof.
(1) For {abc} − {acb} + {bca} ∈ Ker(T ) and d, e ∈ T , by (2.4) and (2.5), we
have
{d, e, {abc} − {acb} + {bca}} = {d, e, {abc} + {cab} + {bca}} = 0,
11
{d, {abc} − {acb} + {bca}, e} = {d, {abc} + {cab} + {bca}, e} = 0,
which imply {T T Ker(T )} = {T Ker(T )T } = 0. Using (2.2) and (2.3), we have
{{abc} − {acb}, d, e} ={{abc}de} − {{acb}de} = {a{bcd}e} + {{adb}ce} − {{adc}be}
={a{bcd}e} + {ad{bce}} + {{ade}bc} − {{ade}cb},
and
Note that
and
{{bca}de} ={bc{ade}} + {{bcd}ae} + {{bce}ad} − {{bce}da}.
{{ade}bc} − {{ade}cb} + {bc{ade}} ∈ Ker(T ),
{{bce}ad} − {{bce}da} + {ad{bce}} ∈ Ker(T ),
{a{bcd}e} + {{bcd}ae}
=({a{bcd}e} − {ae{bcd}} + {{bcd}ea}) + ({{bcd}ae} − {{bcd}ea} + {ae{bcd}})
∈Ker(T ).
Therefore, {{abc} − {acb} + {bca}, d, e} ∈ Ker(T ) and Ker(T ) is an ideal of T .
(2) Since Ker(T ) is an ideal of T , T /Ker(T ) is also a LeibTS with the product
[¯a¯b¯c] = {abc}. [¯a¯a¯b] = ¯0 comes from {aab} = {aab} − {aba} + {aba} ∈ Ker(T ), and then
[¯a¯b¯c] + [¯b¯c¯a] + [¯c¯a¯b] = [¯a¯b¯c] + [¯b¯c¯a] − [¯a¯c¯b] = {abc} − {acb} + {bca} = ¯0.
Moreover,
[¯a¯b[¯c ¯d¯e]] ={ab{cde}} = {{abc}de} − {{abd}ce} − {{abe}cd} + {{abe}dc}
={{abc}de} + {c{abd}e} + {cd{abe}} = [[¯a¯b¯c] ¯d¯e] + [¯c[¯a¯b ¯d]¯e] + [¯c ¯d[¯a¯b¯e]].
Hence T /Ker(T ) is a Lie triple system. It is clear T is a Lie triple system if and only if
Ker(T ) = 0.
(3) Since
Ker(T )[1] = {T Ker(T )Ker(T )} + {Ker(T )T Ker(T )} + {Ker(T )Ker(T )T } = 0,
Ker(T ) is a solvable ideal of T and so Ker(T ) ⊆ R(T ).
(4) Ker(T ) ⊆ Ker(U(T )) ∩T follows from {abc} −{acb} + {bca} = [a, b⊗c] + [b⊗c, a]
= [a + b ⊗ c, a + b ⊗ c] − [a, a] − [b ⊗ c, b ⊗ c] ∈ Ker(U(T )).
Conversely, for all x = [a +P bi ⊗ ci, a +P bi ⊗ ci] ∈ Ker(U(T )), suppose x = d + m
with d ∈ T, m ∈ T ⊗ T . Then d = P({bicia} + {abici} − {acibi}) ∈ Ker(T ), that is,
Ker(U(T )) ∩ T ⊆ Ker(T ).
(5) Note that Ker(U(T )) is an ideal of U(T ) and Ker(U(T )) ∩ T = Ker(T ).
It
follows from Theorem 3.4 that I(Ker(T )) ⊆ Ker(U(T )) ⊆ J (Ker(T )).
12
Let L be a Leibniz algebra and Rad(L) the solvable radical of L. L is called semisim-
ple if Rad(L) = Ker(L). Likewise, we introduce the definition of a semisimple LeibTS.
Definition 4.8. A LeibTS T is said to be semisimple, if R(T ) = Ker(T ).
Proposition 4.9. Let T be a LeibTS. Then the following statements hold.
(1) T /R(T ) is a semisimple Lie triple system.
(2) Let I be an ideal of T . If T /I is a semisimple LeibTS, then R(T ) ⊆ Ker(T ) + I.
(3) If I is an ideal of T and T /I is a semisimple Lie triple system, then R(T ) ⊆ I.
Proof.
(1) T /R(T ) is clearly a Lie triple system since Ker(T ) ⊆ R(T ). If I/R(T )
is a solvable ideal of T /R(T ), then I [k]/R(T ) = (I/R(T ))[k] = ¯0, for some k ∈ N. Then
I [k] ⊆ R(T ), it follows that I is a solvable ideal of T , so I = R(T ). Hence the only solvable
ideal of T /R(T ) is ¯0 and T /R(T ) is semisimple.
(2) Since T /I is semisimple, R(T /I) = Ker(T /I). Note that R(T ) + I/I is solvable
in T /I and Ker(T /I) = Ker(T ) + I/I. Then
R(T ) + I/I ⊆ R(T /I) = Ker(T /I) = Ker(T ) + I/I,
which implies R(T ) ⊆ Ker(T ) + I.
(3) If T /I is a Lie triple system, then Ker(T /I) = 0, so Ker(T ) ⊆ I, and hence
R(T ) ⊆ I by using (2).
Theorem 4.10. R(T ) = Rad(U(T )) ∩T and Rad(U(T )) = I(R(T )) + K(T ) = J (R(T )).
Proof. Suppose that θ is the involutive automorphism of U(T ) determining T .
its homomorphic image θ(Rad(U(T ))) is solvable, then
Since Rad(U(T )) is solvable,
θ(Rad(U(T ))) ⊆ Rad(U(T )) and Rad(U(T )) is θ-invariant. So
Rad(U(T )) = Rad(U(T )) ∩ T + Rad(U(T )) ∩ (T ⊗ T ),
I(Rad(U(T )) ∩ T ) ⊆ Rad(U(T )) ⊆ J (Rad(U(T )) ∩ T ).
Then I(Rad(U(T ))∩T ) is solvable. By Theorem 4.3, Rad(U(T ))∩T and J (Rad(U(T ))∩
T ) is solvable, so J (Rad(U(T )) ∩ T ) = Rad(U(T )).
Since Rad(U(T ))∩T is solvable, Rad(U(T ))∩T ⊆ R(T ), which implies Rad(U(T )) =
J (Rad(U(T )) ∩ T ) ⊆ J (R(T )). But J (R(T )) is solvable by the solvability of R(T ),
then J (R(T )) ⊆ Rad(U(T )), hence J (R(T )) = Rad(U(T )) = J (Rad(U(T )) ∩ T ). So
R(T ) = Rad(U(T )) ∩ T and Rad(U(T )) = J (R(T )).
Consider the factor Leibniz algebra U(T ) = U(T )/I(R(T )). Since U(T ) = T + T ⊗ T
= T + T ⊗ T = U(T ) and T = T + I(R(T ))/I(R(T )) ∼= T /(T ∩ I(R(T ))) = T /R(T ), it
follows that T is a semisimple Lie triple system, so R(T ) = 0. Hence
Rad(U(T )) ⊆ Rad(U(T )) = Rad(U(T )) = J (R(T )) = J (0) = J (0) = K(T ),
13
which implies Rad(U(T )) ⊆ K(T ) + I(R(T )). On the other hand, note that K(T ) =
J (0) ⊆ J (R(T )) and I(R(T )) ⊆ J (R(T )), so Rad(U(T )) = I(R(T ))+ K(T ) = J (R(T )).
Theorem 4.11. If U(T ) is a semisimple Leibniz algebra, then T is a semisimple LeibTS;
if T is a semisimple LeibTS with K(T ) = 0, then U(T ) is a semisimple Leibniz algebra.
Proof.
If U(T ) is semisimple, then Rad(U(T )) = Ker(U(T )), so
R(T ) = Rad(U(T )) ∩ T = Ker(U(T )) ∩ T = Ker(T ),
that is, T is semisimple.
If T is semisimple, then R(T ) = Ker(T ), which implies Rad(U(T )) = I(Ker(T )) =
J (Ker(T )) = Ker(U(T )). Hence U(T ) is semisimple.
5 Levi's theorem for Leibniz triple systems
Definition 5.1. [11] A vector space V is called a module for a Lie triple system T if the
vector space direct sum T ∔ V is itself a Lie triple system such that (1) T is a subsystem,
(2) [abc] lies in V if any one of a, b, c is in V , (3) [abc] = 0 if any two of a, b, c are in V .
Definition 5.2. [18] Let T be a Lie triple system and V a vector space. V is called a
T -module if there exists a bilinear map δ : T ×T → End(V ) such that for all a, b, c, d ∈ T ,
δ(c, d)δ(a, b) − δ(b, d)δ(a, c) − δ(a, [bcd]) + D(b, c)δ(a, d) = 0,
δ(c, d)D(a, b) − D(a, b)δ(c, d) + δ([abc], d) + δ(c, [abd]) = 0,
where D(a, b) = δ(b, a) − δ(a, b).
It is not difficult to prove that the above two definitions about the module for Lie
triple systems are equivalent.
Theorem 5.3. [11] If T is a finite-dimensional semisimple Lie triple system over a field
of characteristic 0, then every finite-dimensional module V is completely reducible.
Theorem 5.4 (Levi's theorem for Lie triple systems). [13] If T is a Lie triple system,
then there is a semisimple subsystem S such that T = R(T )∔S, where R(T ) is the solvable
radical of T .
Theorem 5.5 (Levi's theorem for Leibniz triple systems). Let T be a LeibTS and R(T )
its solvable radical. Then there is a semisimple subsystem S of T such that T = S ∔R(T ).
In particular, S is a semisimple Lie triple system.
14
Proof. By Theorem 4.7, T = T /Ker(T ) is a Lie triple system. Then there are a
semisimple subsystem S and the solvable radical R of T such that T = S ∔ R from Levi's
theorem for Lie triple systems. Let π : T → T be the canonical projection. Then π(R(T ))
is solvable since R(T ) is solvable, hence π(R(T )) ⊆ R.
Conversely, denote by π−1(cid:0)R(cid:1) the inverse image of R. Then π−1(cid:0)R(cid:1) is a subsystem
of T . Since Kerπ = Ker(T ), π−1(cid:0)R(cid:1) /Ker(T ) ∼= R. Note that Ker(T ) and R are
both solvable, which implies that π−1(cid:0)R(cid:1) is solvable, then π−1(cid:0)R(cid:1) ⊆ R(T ), and so
R ⊆ π(R(T )). Thus π(R(T )) = R and R(T ) = π−1(cid:0)R(cid:1).
Denote by F = π−1(cid:0)S(cid:1) the inverse image of S. Then F is a subsystem of T and
S = F/Ker(T ). Since T = T /Ker(T ) and R = R(T )/Ker(T ), it follows that
T /Ker(T ) = F/Ker(T ) ∔ R(T )/Ker(T ).
If a ∈ F ∩ R(T ), then ¯a ∈ (F/Ker(T )) ∩ (R(T )/Ker(T )) = 0, which implies a ∈ Ker(T ),
hence F ∩ R(T ) = Ker(T ). So
T = F + R(T ), F ∩ R(T ) = Ker(T ).
Define a bilinear map δ : S × S → End(F ) by δ( ¯s1, ¯s2)(a) = {as1s2}. Then {as1s2} ∈
2, then
If ¯s1 = ¯s′
F since S = F/Ker(T ) and F is a subsystem of T .
s1 − s′
1 and ¯s2 = ¯s′
2 ∈ Ker(T ), it follows from {T T Ker(T )} = {T Ker(T )T } = 0 that
1, s2 − s′
{as1s2} − {as′
1s′
2} ={as1s2} − {as′
1s2} + {as′
={a(s1 − s′
1)s2} + {as′
1(s2 − s′
1s2} − {as′
1s′
2)} = 0.
2}
Hence δ is well defined. Let D( ¯s1, ¯s2)(a) = δ( ¯s2, ¯s1)(a) − δ( ¯s1, ¯s2)(a) = {as2s1} − {as1s2}.
Then by (2.3) and (2.6),
δ( ¯s3, ¯s4)δ( ¯s1, ¯s2)(a) − δ( ¯s2, ¯s4)δ( ¯s1, ¯s3)(a) − δ( ¯s1, {s2s3s4})(a) + D( ¯s2, ¯s3)δ( ¯s1, ¯s4)(a)
={{as1s2}s3s4} − {{as1s3}s2s4} − {as1{s2s3s4}} + {{as1s4}s3s2} − {{as1s4}s2s3}
=0,
δ( ¯s3, ¯s4)D( ¯s1, ¯s2)(a) − D( ¯s1, ¯s2)δ( ¯s3, ¯s4)(a) + δ({s1s2s3}, s4)(a) + δ( ¯s3, {s1s2s4})(a)
={{as2s1}s3s4} − {{as1s2}s3s4} − {{as3s4}s2s1} + {{as3s4}s1s2}
+ {a{s1s2s3}s4} + {as3{s1s2s4}}
=0.
Thus F is an S-module. By Theorem 5.3, F is a completely reducible S-module since S
is a semisimple Lie triple system. Note that Ker(T ) is an ideal of F . Then Ker(T ) is
a submodule of F . Hence by the completely reducibility of F , there is a complementary
submodule S of F such that F = S ∔Ker(T ). Let σ : F → S be the canonical projection.
Then Kerσ = Ker(T ) and S ∼= F/Ker(T ) is a semisimple Lie triple system such that
T = F + R(T ) = S + Ker(T ) + R(T ) = S + R(T ).
15
Similarly, R(T ) is an S-module and there is a complementary submodule R′ of R(T )
such that R(T ) = R′ ∔ Ker(T ). If a ∈ S ∩ R(T ), then a = a′ + t, for some a′ ∈ R′
and t ∈ Ker(T ). Then a′ ∈ S ⊆ F , which implies a′ ∈ R(T ) ∩ F = Ker(T ). Moreover,
a′ ∈ R′ ∩ Ker(T ) = 0. Hence a = t ∈ Ker(T ) ∩ S = 0, i.e., R(T ) ∩ S = 0. Therefore,
T = S ∔ R(T ) with S both a Lie triple system and a semisimple subsystem of T .
6 The nilpotency of Leibniz triple systems
For a Leibniz algebra L, the sequence of ideals
L ⊇ L1 ⊇ L2 ⊇ · · · where Ln+1 = [LnL]
is called the central lower sequence of L.
Given an arbitrary LeibTS T , the central lower sequence of an ideal I of T is defined
as I 0 = I and for n ≥ 1,
I n+1 = {I nIT } + {I nT I} + {II nT } + {IT I n} + {T II n} + {T I nI}.
In particular, T 0 = T and T n+1 = {T nT T } + {T T nT } + {T T T n}.
In [3], the authors defined another sequence for a Leibniz algebra L as L<0> = L and
L<n+1> = [L<0>L<n>] + [L<1>L<n−1>] + · · · + [L<n−1>L<1>] + [L<n>L<0>].
It is proved that L<n> = Ln for all n ∈ N. Similarly, we could simplify the notation of
T n.
Proposition 6.1. For a LeibTS T , we define another sequence:
T (0) = T, T (n+1) = {T (n)T T }.
Then T (n) = T n, for all n ∈ N.
Proof. The proof is by induction on n. The base step n = 0 is true, for T (0) = T 0 =
T . Suppose T (k) = T k, ∀k ≤ n. Note that
T n+1 = {T nT T } + {T T nT } + {T T T n} = {T (n)T T } + {T T (n)T } + {T T T (n)}.
So it is sufficient to prove
{T T (n)T } + {T T T (n)} ⊆ {T (n)T T } = T (n+1).
In fact, the relation {T (i)T (j)T } + {T (i)T T (j)} ⊆ T (i+j+1) holds, for all i, j ∈ N. We
prove it by induction on j for each i. If j = 0, then
{T (i)T (0)T } + {T (i)T T (0)} = {T (i)T T } = {T (i+1)T T }, ∀i ∈ N.
16
Suppose {T (i)T (j)T } + {T (i)T T (j)} ⊆ T (i+j+1), ∀i ∈ N. Then
{T (i)T (j+1)T } + {T (i)T T (j+1)}
={T (i){T (j)T T }T } + {T (i)T {T (j)T T }}
⊆{{T (i)T (j)T }T T } + {{T (i)T T (j)}T T } + {{T (i)T T }T (j)T }
+ {{T (i)T T (j)}T T } + {{T (i)T T }T (j)T } + {{T (i)T T }T T (j)}
⊆{T (i+j+1)T T } + {T (i+1)T (j)T } + {T (i+1)T T (j)}
=T (i+j+2).
Fix i = 0, and it follows {T T (n)T } + {T T T (n)} ⊆ T (n+1), which completes the proof.
Hence the notions of T n and T (n) coincide.
Definition 6.2. An ideal I of a LeibTS T is called nilpotent, if there exists n ∈ N such
that I n = 0.
Proposition 6.3. If I is an ideal of a LeibTS T , then I n is also an ideal of T and
I n+1 ⊆ I n, ∀n ∈ N.
Proof. We proceed by induction. It is clear when n = 0. Suppose that I k is an ideal
of T . Then
{I k+1T T } = {{I kT I} + {I kIT } + {IT I k} + {II kT } + {T I kI} + {T II k}, T, T }.
By (2.2) and (2.3), it follows that
{{I kT I}T T } ⊆ {I kT {IT T }} + {{I kT T }IT } + {{I kT T }T I} ⊆ I k+1,
{{I kIT }T T } ⊆ {I k{IT T }T } + {{I kT I}T T } + {{I kT T }IT } ⊆ I k+1,
{{IT I k}T T } ⊆ {IT {I kT T }} + {{IT T }I kT } + {{IT T }T I k} ⊆ I k+1,
{{II kT }T T } ⊆ {I{I kT T }T } + {{IT I k}T T } + {{IT T }I kT } ⊆ I k+1,
{{T I kI}T T } ⊆ {T I k{IT T }} + {{T I kT }IT } + {{T I kT }T I} ⊆ I k+1,
{{T II k}T T } ⊆ {T I{I kT T }} + {{T IT }I kT } + {{T IT }T I k} ⊆ I k+1,
which imply {I k+1T T } ⊆ I k+1. Similarly, one can prove
{T I k+1T } ⊆ I k+1
and {T T I k+1} ⊆ I k+1.
So I k+1 is an ideal of T . Hence I n is an ideal of T and I n+1 ⊆ I n, ∀n ∈ N.
Proposition 6.4. If V is an ideal of U(T ), then V ∩ T is an ideal of T such that
(V ∩ T )n ⊆ V n ∩ T for all n ∈ N.
17
Proof. Since V is an ideal of a Leibniz algebra U(T ), V 0 = V and V n = [V n−1, V ].
Note that V is an ideal of U(T ) if and only if [V, U(T )] + [U(T ), V ] ⊆ V . Then
{(V ∩ T )T T } = [[V ∩ T, T ], T ] ⊆ [[V, U(T )], U(T )] ⊆ V,
{T (V ∩ T )T } = [[T, V ∩ T ], T ] ⊆ [[U(T ), V ], U(T )] ⊆ V,
{T T (V ∩ T )} = [[T, T ], V ∩ T ] ⊆ [U(T ), V ] ⊆ V,
{(V ∩ T )T T } + {T (V ∩ T )T } + {T T (V ∩ T )} ⊆ {T T T } ⊆ T.
That is,
{(V ∩ T )T T } + {T (V ∩ T )T } + {T T (V ∩ T )} ⊆ V ∩ T.
So V ∩ T is an ideal of T .
Now we use the induction to prove (V ∩ T )n ⊆ V n ∩ T for all n ∈ N. The base step
n = 0 is clearly true. Suppose (V ∩ T )k ⊆ V k ∩ T . Then
(V ∩ T )k+1
={(V ∩ T )k(V ∩ T )T } + {(V ∩ T )kT (V ∩ T )} + {(V ∩ T )(V ∩ T )kT }
+ {(V ∩ T )T (V ∩ T )k} + {T (V ∩ T )(V ∩ T )k} + {T (V ∩ T )k(V ∩ T )}
⊆{(V k ∩ T )(V ∩ T )T } + {(V k ∩ T )T (V ∩ T )} + {(V ∩ T )(V k ∩ T )T }
+ {(V ∩ T )T (V k ∩ T )} + {T (V ∩ T )(V k ∩ T )} + {T (V k ∩ T )(V ∩ T )}
⊆[[V k, V ], T ] + [[V k, T ], V ] + [[V, V k], T ] + [[V, T ], V k] + [[T, V ], V k] + [[T, V k], V ]
⊆[[V k, T ], V ] + [V k, [V, T ]] + [[V, T ], V k] + [V, [V k, T ]] + [V, V k] + [V k, V ]
⊆[V k, V ] + [V, V k] = V k+1.
Then (V ∩ T )k+1 ⊆ V k+1 ∩ T . Therefore, (V ∩ T )n ⊆ V n ∩ T for all n ∈ N.
Lemma 6.5. Let I be an ideal of a LeibTS T . Then
(1) I n ⊆ I(I)n, for all n ∈ N.
(2) I(I)2n−1 ⊆ I n + I n−1 ⊗ I + I n ⊗ T and I(I)2n ⊆ I n + I n ⊗ T , ∀n ≥ 1.
Proof.
(1) It follows from Proposition 6.4.
(2) We still use the induction. It is straightforward that
I(I)1 =[I(I), I(I)] = [I + I ⊗ T + T ⊗ I, I + I ⊗ T + T ⊗ I]
⊆I ⊗I +{IIT }+{IT I}+{IT I}⊗T +{IT T }⊗I +{T II}+{T II}⊗T +{T IT }⊗I
⊆I 1 + I ⊗ I + I 1 ⊗ T
and
I(I)2 =[I(I)1, I(I)] ⊆ [I 1 + I ⊗ I + I 1 ⊗ T, I + I ⊗ T + T ⊗ I]
⊆I 1⊗I +{I 1IT }+{I 1T I}+{III}+{III}⊗T +{IIT }⊗I +{I 1T I}⊗T +{I 1T T }⊗I
18
⊆I 1 + I 1 ⊗ T.
Suppose I(I)2k−1 ⊆ I k + I k−1 ⊗ I + I k ⊗ T and I(I)2k ⊆ I k + I k ⊗ T . Then
I(I)2k+1 =[I(I)2k, I(I)] ⊆ [I k + I k ⊗ T, I + I ⊗ T + T ⊗ I]
⊆I k ⊗I +{I kIT }+{I kT I}+{I kT I}⊗T +{I kT T }⊗I
⊆I k+1 + I k ⊗ I + I k+1 ⊗ T
and
I(I)2k+2 =[I(I)2k+1, I(I)] ⊆ [I k+1 + I k ⊗ I + I k+1 ⊗ T, I + I ⊗ T + T ⊗ I]
⊆I k+1 ⊗ I + {I k+1IT } + {I k+1T I} + {I kII} + {I kII} ⊗ T + {I kIT } ⊗ I
+ {I k+1T I} ⊗ T + {I k+1T T } ⊗ I
⊆I k+1 + I k+1 ⊗ T.
The proof is completed.
Lemma 6.6. Let I and J be two ideals of a LeibTS T . Then
(I + J)n ⊆ I n + J n +
n−1
Xk=0
I k ∩ J n−1−k,
∀n ≥ 1.
Proof.
If n = 1, then
(I + J)1 = {(I + J)(I + J)T } + {(I + J)T (I + J)} + {T (I + J)(I + J)} ⊆ I 1 + J 1 + I ∩ J.
Suppose (I + J)m ⊆ I m + J m +Pm−1
k=0 I k ∩ J m−1−k. Then
(I + J)m+1 ={(I + J)m(I + J)T } + {(I + J)mT (I + J)} + {(I + J)(I + J)mT }
+ {(I + J)T (I + J)m} + {T (I + J)(I + J)m} + {T (I + J)m(I + J)}
Note that
m−1
{(I + J)m(I + J)T } ⊆ {I m + J m +
I k ∩ J m−1−k, I + J, T }
m−1
Xk=0
Xk=0
I k+1 ∩ J m−1−k +
I k ∩ J m−k
m−1
Xk=0
⊆I m+1 + I m ∩ J + I ∩ J m + J m+1 +
=I m+1 + J m+1 +
I k ∩ J m−k.
m
Xk=0
One can prove that other five items belong to I m+1 + J m+1 +Pm
same way. Hence the lemma follows from the induction.
k=0 I k ∩ J m−k in the
19
By Lemmas 6.5 and 6.6, it is clear that the following theorem holds.
Theorem 6.7. Let T be a LeibTS. Then
(1) I is a nilpotent ideal of T if and only if I(I) is a nilpotent ideal of U(T ). In
particular, T is a nilpotent LeibTS if and only if U(T ) is a nilpotent Leibniz algebra.
(2) If I and J are nilpotent ideals of T , then I + J is a nilpotent ideal of T .
Definition 6.8. Let N(T ) denote the sum of all nilpotent ideals in T . Then N(T ) is the
unique maximal nilpotent ideal by Theorem 6.7, called the nilpotent radical of T .
Theorem 6.9. Suppose Nil(U(T )) is the nilpotent radical of U(T ). Then N(T ) =
Nil(U(T )) ∩ T .
Proof. Since Nil(U(T )) is a nilpotent ideal of U(T ), Nil(U(T )) ∩ T is a nilpotent
ideal of T by Proposition 6.4. If I is a nilpotent ideal of T , then I(I) a nilpotent ideal
of U(T ) by Theorem 6.7, hence I(I) ⊆ Nil(U(T )). So I = I(I) ∩ T ⊆ Nil(U(T )) ∩ T .
Therefore, Nil(U(T )) ∩ T is a maximal nilpotent ideal of T , that is, Nil(U(T )) ∩ T =
N(T ).
Lemma 6.10. [3] If L is a solvable Leibniz algebra, then [L, L] is nilpotent.
Proposition 6.11. Suppose T is a LeibTS. If T is nilpotent, then T is solvable; if T is
solvable, then T 1 = {T T T } is nilpotent.
Proof.
It is clear that T [n] ⊆ T n. So the nilpotency of T implies the solvability of
T . Conversely, if T is solvable, then U(T ) is solvable by Corollary 4.4, so [U(T ), U(T )] is
nilpotent, hence [U(T ), U(T )] ⊆ Nil(U(T )). Note that
T 1 = {T T T } = [[T, T ], T ] ⊆ [U(T ), U(T )] ⊆ Nil(U(T )).
Then T 1 ⊆ Nil(U(T )) ∩ T = N(T ). So T 1 is nilpotent.
7 The representation of Leibniz triple systems
Definition 7.1. Let T be a LeibTS and V a vector space. V is called a T -module, if
T ∔ V is a LeibTS such that (1) T is a subsystem, (2) {abc} ∈ V if any one of a, b, c is
in V , (3) {abc} = 0 if any two of a, b, c are in V .
Definition 7.2. Let T be a LeibTS and V a vector space. Suppose l, m, r : T × T →
End(V ) are bilinear maps such that
l(a, {bcd}) = l({abc}, d) − l({acb}, d) − l({adb}, c) + l({adc}, b),
m(a, d)l(b, c) = m({abc}, d) − m({acb}, d) − r(c, d)m(a, b) + r(b, d)m(a, c),
m(a, d)m(b, c) = r(c, d)l(a, b) − r(c, d)m(a, b) − m({acb}, d) + r(b, d)l(a, c),
(7.7)
(7.8)
(7.9)
20
m(a, d)r(b, c) = r(c, d)m(a, b) − r(c, d)l(a, b) − r(b, d)l(a, c) + m({acb}, d),
r({abc}, d) = r(c, d)r(a, b) − r(c, d)r(b, a) − r(b, d)r(c, a) + r(a, d)r(c, b),
l(a, b)l(c, d) = l({abc}, d) − l({abd}, c) − r(c, d)l(a, b) + r(d, c)l(a, b),
l(a, b)m(c, d) = m({abc}, d) − r(c, d)l(a, b) − l({abd}, c) + m({abd}, c),
l(a, b)r(c, d) = r(c, d)l(a, b) − m({abc}, d) − m({abd}, c) + l({abd}, c),
m(a, {bcd}) = r(c, d)m(a, b) − r(b, d)m(a, c) − r(b, c)m(a, d) + r(c, b)m(a, d),
r(a, {bcd}) = r(c, d)r(a, b) − r(b, d)r(a, c) − r(b, c)r(a, d) + r(c, b)r(a, d).
(7.10)
(7.11)
(7.12)
(7.13)
(7.14)
(7.15)
(7.16)
Then (r, m, l) is called a representation of T on V .
Remark 7.3. If V is a T -module and define l(a, b)(v) = {abv}, m(a, b)(v) = {avb},
r(a, b)(v) = {vab}, then (r, m, l) is a representation of T . Conversely, if (r, m, l) is
a representation of T , then V is a T -module by setting {abv} = l(a, b)(v), {avb} =
m(a, b)(v) and {vab} = r(a, b)(v). So the notions of a T -module and a representation of
T are equivalent.
Proposition 7.4. Suppose that T is a LeibTS and (r, m, l) is a representation of T on
V .
(1) m(a, b)k+1 = (−1)km(a, b)r(a, b)k, ∀k ∈ N. So if r(a, b) is nilpotent, then m(a, b)
is nilpotent.
(2) Set R(a, b) = r(a, b) − r(b, a). Then l(a, b)k+1 = (−1)kl(a, b)R(a, b)k, ∀k ∈ N. So
if R(a, b) is nilpotent, then l(a, b) is nilpotent.
Proof.
(1) By (7.9) and (7.10), m(a, b)m(c, d) = −m(a, b)r(c, d). Then
m(a, b)k+1 = −m(a, b)kr(a, b) = · · · = (−1)km(a, b)r(a, b)k, ∀k ∈ N.
(2) By (7.14),
l(a, b)r(c, d) = r(c, d)l(a, b) − m({abc}, d) − m({abd}, c) + l({abd}, c),
l(a, b)r(d, c) = r(d, c)l(a, b) − m({abd}, c) − m({abc}, d) + l({abc}, d).
Then
l(a, b)R(c, d) = −l({abc}, d) + l({abd}, c) + r(c, d)l(a, b) − r(d, c)l(a, b) = −l(a, b)l(c, d),
where the last identity is by (7.12). Hence
l(a, b)k+1 = −l(a, b)kR(a, b) = · · · = (−1)kl(a, b)R(a, b)k, ∀k ∈ N.
The proof is completed.
Denote by l(T, T ) = {P l(a, b) a, b ∈ T }, m(T, T ) = {P m(a, b) a, b ∈ T },
r(T, T ) = {P r(a, b) a, b ∈ T } and R(T, T ) = {P R(a, b) a, b ∈ T }. Then l(T, T ),
21
m(T, T ), r(T, T ) and R(T, T ) are subspaces of End(V ). Note that for all a, b, c, d ∈ T , by
(7.11) and (7.16),
[R(a, b), R(c, d)] = R(a, b)R(c, d) − R(c, d)R(a, b)
=(r(a, b) − r(b, a))(r(c, d) − r(d, c)) − (r(c, d) − r(d, c))(r(a, b) − r(b, a))
=r(a, b)r(c, d) − r(b, a)r(c, d) − r(a, b)r(d, c) + r(b, a)r(d, c)
− r(c, d)r(a, b) + r(d, c)r(a, b) + r(c, d)r(b, a) − r(d, c)r(b, a)
=r(c, {dab}) + r(d, b)r(c, a) + r(d, a)r(c, b) − r(a, d)r(c, b)
− r(c, {dba}) − r(d, a)r(c, b) − r(d, b)r(c, a) + r(b, d)r(c, a)
− r(d, {cab}) − r(c, b)r(d, a) − r(c, a)r(d, b) + r(a, c)r(d, b)
+ r(d, {cba}) + r(c, a)r(d, b) + r(c, b)r(d, a) − r(b, c)r(d, a)
− r({cba}, d) + r(a, d)r(c, b) − r(a, d)r(b, c) − r(b, d)r(a, c)
+ r({dba}, c) − r(a, c)r(d, b) + r(a, c)r(b, d) + r(b, c)r(a, d)
+ r({cab}, d) − r(b, d)r(c, a) + r(b, d)r(a, c) + r(a, d)r(b, c)
− r({dab}, c) + r(b, c)r(d, a) − r(b, c)r(a, d) − r(a, c)r(b, d)
=R(c, {dab}) − R(c, {dba}) − R(d, {cab}) + R(d, {cba}).
So R(T, T ) is a subalgebra of gl(V ).
Proposition 7.5. l(T, T ), m(T, T ) and r(T, T ) are all adjoint R(T, T )-module.
Proof.
It is sufficient to prove that [R(T, T ), l(T, T )] ⊆ l(T, T ), [R(T, T ), m(T, T )] ⊆
m(T, T ) and [R(T, T ), r(T, T )] ⊆ r(T, T ).
It follows from (7.14) that
[r(c, d), l(a, b)] = m({abc}, d) + m({abd}, c) − l({abd}, c),
[r(d, c), l(a, b)] = m({abd}, c) + m({abc}, d) − l({abc}, d).
Then [R(c, d), l(a, b)] = l({abc}, d) − l({abd}, c) ∈ l(T, T ).
It follows from (7.10) that
m(a, d)r(b, c) = r(c, d)m(a, b) − r(c, d)l(a, b) − r(b, d)l(a, c) + m({acb}, d),
m(a, d)r(c, b) = r(b, d)m(a, c) − r(b, d)l(a, c) − r(c, d)l(a, b) + m({abc}, d).
Then m(a, d)R(b, c) = r(c, d)m(a, b) − r(b, d)m(a, c) + m({acb}, d) − m({abc}, d). Note
that by (7.15),
R(b, c)m(a, d) = r(c, d)m(a, b) − r(b, d)m(a, c) − m(a, {bcd}).
Hence [R(b, c), m(a, d)] = −m(a, {bcd}) − m({acb}, d) + m({abc}, d) ∈ m(T, T ).
It follows from (7.11) and (7.16) that
r(c, d)R(a, b) = r({abc}, d) + r(b, d)r(c, a) − r(a, d)r(c, b),
R(a, b)r(c, d) = −r(c, {abd}) + r(b, d)r(c, a) − r(a, d)r(c, b).
Then [R(a, b), r(c, d)] = −r(c, {abd}) − r({abc}, d) ∈ r(T, T ).
22
Set
ZT (V ) = {x ∈ T {xT V } = {xV T } = {T xV } = {V xT } = {V T x} = {T V x} = 0}.
If T is a Lie triple system, then it is clear that ZT (V ) is an ideal of T . But it is not true
if T is a non-Lie LeibTS.
In fact, it is easy to prove that
{T V {ZT (V )T T }} = {T V {T ZT (V )T }} = {T V {T T ZT (V )}} = 0,
{V T {ZT (V )T T }} = {V T {T ZT (V )T }} = {V T {T T ZT (V )}} = 0,
{T {ZT (V )T T }V } = {T {T ZT (V )T }V } = {T {T T ZT (V )}V } = 0,
{V {ZT (V )T T }T } = {V {T ZT (V )T }T } = {V {T T ZT (V )}T } = 0,
{{T T ZT (V )}T V } = {{T ZT (V )T }T V } = {{T T ZT (V )}V T } = {{T ZT (V )T }V T } = 0.
However, we fail to show {{ZT (V )T T }T V } = 0 or {{ZT (V )T T }V T } = 0. Therefore, one
can not say that ZT (V ) is an ideal of T .
Theorem 7.6. Suppose that T is a LeibTS and V is an irreducible T -module. If ZT (V )
is an ideal of T , then T /ZT (V ) is a Lie triple system, and either l(T, T ) = m(T, T ) = 0
or
l(x, y) = −l(y, x) = R(x, y),
r(x, y) = −m(x, y),
∀x, y ∈ T.
Proof. Let T be a split extension of V by T . Then T = T ∔ V and V is also a
T -module. Note that
Z T (V ) = {x ∈ T {x T V } = {xV T } = { T xV } = {V x T } = {V T x} = { T V x} = 0}
= {x ∈ T {xT V } = {xV T } = {T xV } = {V xT } = {V T x} = {T V x} = 0}
= ZT (V ) ∔ V.
Consider the following two cases.
1. ZT (V ) = 0. Since V is an irreducible T -module, V is a minimal ideal of T .
Note that Ker( T ) is an ideal of T . Then either V ⊆ Ker( T ) or V ∩ Ker( T ) = 0.
Then {V T Ker( T )} = {V Ker( T ) T } = { T V Ker( T )} = {Ker( T )V T } = { T Ker( T )V } =
{Ker( T ) T V } = 0 by Theorem 4.7. Hence Ker( T ) ⊆ Z T (V ) = V , and so Ker(T ) ⊆
Ker( T ) ⊆ V . Then Ker(T ) ⊆ V ∩ T = 0 and T /V ∼= T is a Lie triple system. It follows
from Ker( T ) ⊆ V and the minimality of V that Ker( T ) = 0 or Ker( T ) = V .
If Ker( T ) = V , then { T T V } = { T V T } = 0, which implies {T T V } = {T V T } = 0,
that is, l(T, T ) = m(T, T ) = 0.
If Ker( T ) = 0, then T is a Lie triple system. Hence l(x, y) = −l(y, x) = R(x, y) and
r(x, y) = −m(x, y), for all x, y ∈ T .
2. ZT (V ) 6= 0. Let ¯T = T /ZT (V ). Then ¯T is a LeibTS and V is a ¯T -module
satisfying Z ¯T (V ) = 0. Note that V is also an irreducible ¯T -module. Hence ¯T is a Lie
triple system.
23
Theorem 7.7. Let T be a LeibTS and V 6= 0 a finite-dimensional T -module. Suppose
that ZT (V0) is an ideal of T for every submodule V0 of V . If R(x, y) is nilpotent for all
x, y ∈ T , then l(x, y) is nilpotent and there exists 0 6= v ∈ V such that
R(x, y)(v) = l(x, y)(v) = 0,
∀x, y ∈ T.
Proof. Since V is finite-dimensional, there exists a sequence of submodules
V = V0 ⊃ V1 ⊃ · · · ⊃ Vn−1 ⊃ Vn = 0
such that each factor Vk/Vk+1 (k = 0, 1, · · · , n − 1) is an irreducible T -module. Then
V ∼= V0/V1 ∔ V1/V2 ∔ · · · ∔ Vn−2/Vn−1 ∔ Vn−1/Vn.
Note that R(T, T ) is a subalgebra of gl(V ) and by the Engel's theorem for Lie
algebras, there exists 0 6= v ∈ V such that R(x, y)(v) = 0, ∀x, y ∈ T . Suppose
v = v1 + · · · + vn−1, where vi ∈ Vi/Vi+1. Since Vi/Vi+1 is an irreducible T -module, it
follows that l(x, y)Vi/Vi+1 = 0 or l(x, y)Vi/Vi+1 = −R(y, x)Vi/Vi+1. In either case we have
l(x, y)(vi) = 0, so l(x, y)(v) = 0, ∀x, y ∈ T .
Remark 7.8. If ZT (I) is an ideal of T for every ideal I of T , then by Theorem 7.7, there
exits a basis of a finite-dimensional LeibTS T relative to which the matrix of l(x, y) is
strictly upper triangular for all x, y ∈ T . Then {T T T n} = 0 for some n ∈ N. But one
can not say that T is nilpotent. Hence the Engel's theorem for a LeibTS could not be
proved in this way.
References
[1] S. Albeverio, S. Ayupov, B. Omirov, (2005), On nilpotent and simple Leibniz alge-
bras. Comm. Algebra 33(1), 159-172.
[2] S. Albeverio, S. Ayupov, B. Omirov, (2006), Cartan subalgebras, weight spaces, and
criterion of solvability of finite dimensional Leibniz algebras. Rev. Mat. Complut.
19(1), 183-195.
[3] S. Ayupov, B. Omirov, (1998), On Leibniz algebras. Algebra and operator theory,
Kluwer Acad. Publ., Dordrecht, 1-12.
[4] D. Barnes, (2012), On Engel's theorem for Leibniz algebras. Comm. Algebra 40(4),
1388-1389.
[5] D. Barnes, (2012), On Levi's theorem for Leibniz algebras. Bull. Aust. Math. Soc.
86(2), 184-185.
[6] A. Bolh, (1965), On a generalization of the concept of Lie algebra. (Russian) Dokl.
Akad. Nauk SSSR 165, 471-473.
24
[7] M. Bremner, J. S´anchez-Ortega, (2014), Leibniz triple systems. Commun. Contemp.
Math. 16(1), 1350051, 19 pp.
[8] I. Demir, K. Misra, E. Stitzinger, (2014), On some structures of Leibniz algebras.
Contemp. Math. 623, Amer. Math. Soc., Providence, RI.
[9] S. G´omez-Vidal, S. Khudoyberdiyev, B. Omirov, (2014), Some remarks on semisimple
Leibniz algebras. J. Algebra 410, 526-540.
[10] V. Gorbatsevich, (2013), On some basic properties of Leibniz algebras. arXiv: 1302.
3345.
[11] B. Harris, (1961), Cohomology of Lie triple systems and Lie algebras with involution.
Trans. Amer. Math. Soc. 98, 148-162.
[12] P. Kolesnikov, (2008), Varieties of dialgebras, and conformal algebras. (Russian)
Sibirsk. Mat. Zh. 49(2), 322-339; translation in Sib. Math. J. 49(2), 257-272.
[13] W. Lister, (1952), A structure theory of Lie triple systems. Trans. Amer. Math. Soc.
72, 217-242.
[14] J. Loday, (1993), Une version non commutative des alg`ebres de Lie: les alg`ebres de
Leibniz. (French) Enseign. Math. 39(3-4), 269-293.
[15] J. Loday, T. Pirashvili, (1993), Universal enveloping algebras of Leibniz algebras and
(co)homology. Math. Ann. 296(1), 139-158.
[16] K. Meyberg, (1972), Lectures on algebras and triple systems. The University of
Virginia, Charlottesville, Va. Notes on a course of lectures given during the academic
year 1971-1972.
[17] A. Patsourakos, (2007), On nilpotent properties of Leibniz algebras. Comm. Algebra
35(12), 3828-3834.
[18] K. Yamaguti, (1960), On the cohomology space of Lie triple system. Kumamoto J.
Sci. Ser. A 5, 44-52.
25
|
1211.6323 | 1 | 1211 | 2012-11-27T15:10:47 | Ring coproducts embedded in power-series rings | [
"math.RA",
"math.GR"
] | Let $R$ be a ring (associative, with 1), and let $R<< a,b>>$ denote the power-series $R$-ring in two non-commuting, $R$-centralizing variables, $a$ and $b$. Let $A$ be an $R$-subring of $R<< a>>$ and $B$ be an $R$-subring of $R<< b>>$, and let $\alpha$ denote the natural map $A \amalg_R B \to R<< a,b>>$. This article describes some situations where $\alpha$ is injective and some where it is not.
We prove that if $A$ is a right Ore localization of $R[a]$ and $B$ is a right Ore localization of $R[b]$, then $\alpha$ is injective. For example, the group ring over $R$ of the free group on $\{1+a, 1+b\}$ is $R[ (1+a)^{\pm 1}] \amalg_R R[ (1+b)^{\pm 1}]$, which then embeds in $R<< a,b>>$. We thus recover a celebrated result of R H Fox, via a proof simpler than those previously known.
We show that $\alpha$ is injective if $R$ is \textit{$\Pi$-semihereditary}, that is, every finitely generated, torsionless, right $R$-module is projective.
The article concludes with some results contributed by G M Bergman that describe situations where $\alpha$ is not injective. He shows that if $R$ is commutative and $\text{w.gl.dim\,} R \ge 2$, then there exist examples where the map $\alpha' \colon A \amalg_R B \to R<< a>>\amalg_R R<< b>>$ is not injective, and hence neither is $\alpha$. It follows from a result of K R Goodearl that when $R$ is a commutative, countable, non-self-injective, von Neumann regular ring, the map $\alpha"\colon R<< a>>\amalg_R R<< b>> \to R<< a,b>>$ is not injective. Bergman gives procedures for constructing other examples where $\alpha"$ is not injective. | math.RA | math |
Ring coproducts embedded in power-series rings
Pere Ara and Warren Dicks
June 8, 2021
Abstract
Let R be a ring (associative, with 1), and let Rhha, bii denote the power-
series R-ring in two non-commuting, R-centralizing variables, a and b. Let
A be an R-subring of Rhhaii and B be an R-subring of Rhhbii, and let α denote
the natural map A ∐R B → Rhha, bii. This article describes some situations
where α is injective and some where it is not.
We prove that if A is a right Ore localization of R[a] and B is a right Ore
localization of R[b], then α is injective. For example, the group ring over R
of the free group on {1+ a, 1+ b} is R[(1+ a)±1] ∐R R[(1+ b)±1], which then
embeds in Rhha, bii. We thus recover a celebrated result of R. H. Fox, via a
proof simpler than those previously known.
We show that α is injective if R is Π-semihereditary, that is, every finitely
generated, torsionless, right R-module is projective. (This concept was first
studied by M. F. Jones, who showed that it is left-right symmetric.
It fol-
lows from a result of I. I. Sahaev that if w.gl.dim R 6 1 and R embeds in
a skew field, then R is Π-semihereditary. Also, it follows from a result of
V. C. Cateforis that if R is right semihereditary and right self-injective, then
R is Π-semihereditary.)
The arguments and results extend easily from two variables to any set of
variables.
The article concludes with some results contributed by G. M. Bergman
that describe situations where α is not injective. He shows that if R
is commutative and w.gl.dim R > 2, then there exist examples where the
map α′ : A ∐R B → Rhhaii ∐R Rhhbii is not injective, and hence neither is
α.
It follows from a result of K. R. Goodearl that when R is a commu-
tative, countable, non-self-injective, von Neumann regular ring, the map
α′′ : Rhhaii ∐R Rhhbii → Rhha, bii is not injective. Bergman gives procedures
for constructing other examples where α′′ is not injective.
2010 Mathematics Subject Classification. Primary: 16S10; Secondary: 20C07, 20E05.
Key words. Ring coproduct, free-group group ring, power series, Ore localization.
1 Summary of the results
Some of the terms that we use in this section will be explained in more detail in
subsequent sections.
Throughout, we fix a ring R (associative, with 1). By an R-ring S we mean a
ring S together with a specified ring homomorphism R → S. We write U(R) for
the multiplicative group of units of R. We shall be considering the following.
1.1 Hypotheses. Let Rhha, bii denote the power-series R-ring in two non-com-
muting, R-centralizing variables, a and b. Let A be an R-subring of Rhhaii and
B be an R-subring of Rhhbii. Let α denote the natural map A ∐R B → Rhha, bii.
1
2
Ring coproducts embedded in power-series rings
This article describes some situations where α is injective and some where it is
not.
In Section 2, we fix some of the notation, and we use a result of P. M. Cohn
to decompose the map from the domain of α to the image of α as a direct sum of
R-bimodule maps.
In Section 3, in Proposition 3.1, we shall see that if A is a right Ore localization
of R[a] and B is a right Ore localization of R[b], then α is injective.
1.2 Example. Let h := 1+ a, H := hhi 6 U(Rhhaii), and let RH denote the
group ring of H over R.
It is not difficult to see that we may view R[a] as
R[h] ⊆ RH ⊆ Rhhaii and that RH is a central Ore localization of R[h]. The anal-
ogous statements hold for k := 1+ b and K := hki 6 U(Rhhbii). By Proposition 3.1
below, the map α : RH ∐R RK → Rhha, bii is injective. Here, RH ∐R RK may be
identified with the group ring over R of the group F freely generated by two symbols
1+ a, 1+ b. We thus recover the celebrated result given in 1953 by R. H. Fox [11],
that RF embeds in Rhha, bii. Although Fox considers only the case R = Z, his
argument works for any ring. The proof given here is simpler than the previously
known proofs, such as the original proof reviewed recently in [1, pp. 366 -- 371], and
the proof by G. M. Bergman [8, pp. 528 -- 529] of the case where R is a field.
(One consequence of Fox's result is that if R 6= {0}, then the group F embeds
in U(Rhha, bii). W. Magnus [21] had already given a short proof of the R = Z
case of this consequence in 1935, and deduced much useful information about free
groups. Some of this information eventually led to the discovery, in 1946, of the
important fact that F is an orderable group. Much later, Bergman [3] remarked
that orderability follows directly from Magnus' embedding, since Zhha, bii can easily
be given a ring ordering whose positive cone contains F .)
In Section 4, we review some definitions and results from ring theory. We
say that R is Π-semihereditary if every finitely generated, right R-submodule of
a direct product of copies of R is projective.
(Completely reducible rings are
clearly Π-semihereditary, since all their modules are projective. It is also clear that
Π-semihereditary rings are right semihereditary and, hence, satisfy w.gl.dim R 6 1.)
This concept was studied first by M. F. Jones [16] who showed that it is left-right
symmetric. Earlier, K. R. Goodearl [12] had observed that a result of V. C. Cate-
foris [4] shows that if R is right semihereditary and right self-injective, then
R is Π-semihereditary; see Theorem 4.17 below.
It follows from a result of
I. I. Sahaev [24] that if w.gl.dim R 6 1 and R embeds in a skew field, then R is
Π-semihereditary; see Corollary 4.15(i) below. We recall these and other examples
of Π-semihereditary rings.
In Section 5, we write α as a composite A ∐R B α′
→ Rhha, bii.
Then, in Corollary 5.4, we show that α′, α′′, and α are injective if R is Π-semihered-
itary, or, more generally, if every finitely generated right R-submodule of Rhhaii is
projective. The examples of the preceding section then apply.
→ Rhhaii ∐R Rhhbii α′′
The results of Sections 3 and 5 extend from two variables to any set of variables;
the arguments are very similar but the notation is more complicated. We leave the
details to the interested reader.
Section 6 contains results contributed by Bergman that describe situations where
α is not injective. Examples of non-injectivity of α can arise from non-injectivity ei-
ther of α′ or of α′′. (Examples of the former sort clearly lead to non-injectivity of α,
while an example of the latter sort does so on taking A = Rhhaii, B = Rhhbii.)
Pere Ara and Warren Dicks
3
Examples are given of both sorts. He obtains non-injectivity of α′ for any com-
mutative ring R with w.gl.dim R > 2, e.g. R = Q[x, y]. A result of Goodearl [12]
yields non-injectivity of α′′ whenever R is a commutative, countable, non-self-in-
jective, von Neumann regular ring, e.g. R = Q[ei : i ∈ Z]/(eiej − δi,jei : i, j ∈ Z).
Constructions are described that generalize the latter example.
2 The bimodule structure of ring coproducts
2.1 Notation. We denote by N the set of finite cardinals, {0, 1, 2, 3, . . .}.
For any R-bimodules M1 and M2, we will denote the R-bimodule M1 ⊗R M2 by
M1 ⊗ M2 if R is understood.
For any R-bimodule M and any n ∈ N, we recursively define the R-bimodule
tensor power M ⊗n by the formulas M ⊗0 := R and M ⊗(n+1) := M ⊗n ⊗ M .
2.2 Notation. Suppose that Hypotheses 1.1 hold.
Let {a, b}∗ denote the free monoid on {a, b}. We shall usually write elements of
f (w)·w, where f : w 7→ f (w) is an arbitrary function
Rhha, bii as formal sums Pw∈{a,b}∗
from {a, b}∗ to R. The set {w ∈ {a, b}∗ : f (w) 6= 0} is called the {a, b}∗-support of
Pw∈{a,b}∗
f (w)·w.
We view R ⊆ R[a] ⊆ Rhhaii ⊆ Rhha, bii and R ⊆ R[b] ⊆ Rhhbii ⊆ Rhha, bii. Let
a := A ∩ aRhhaii and b := B ∩ bRhhbii. Then a is a two-sided ideal of A, and, as
R-bimodules, A = R ⊕ a. Analogous statements hold for B.
The ring coproduct of A and B amalgamating the two copies of R will be denoted
A ∐R B, or by A ∐ B if R is understood. Now A ∐ B = (R ⊕ a) ∐ (R ⊕ b). As
noted by P. M. Cohn [7, pp. 60 -- 61], it is not difficult to show that there is then an
expansion as a direct sum of R-bimodules
A ∐ B = R ⊕ a ⊕ b ⊕ (a⊗b) ⊕ (b⊗a) ⊕ (a⊗b⊗a) ⊕ (b⊗a⊗b) ⊕ · · · .
We may write this in the form
(1) A ∐ B = R ⊕ a ⊕ b ⊕ Ln>1
(cid:16)(a⊗b)⊗n ⊕ (b⊗a)⊗n ⊕ ((a⊗b)⊗n⊗a) ⊕ ((b⊗a)⊗n⊗b)(cid:17).
When α is applied to (1), the direct sums are replaced with sums, and the
tensor-product multiplications within A ∐ B are replaced with the multiplications
of Rhha, bii. Thus, we may write
(2)
α(A ∐ B) = R + a + b + Pn>1
(cid:16)(ab)n + (ba)n + ((ab)na) + ((ba)nb)(cid:17).
The {a, b}∗-support of a lies in {ai : i > 1}, and the {a, b}∗-support of b lies in
{bj : j > 1}.
It follows that any two summands appearing in (2) have disjoint
{a, b}∗-supports. Hence, we recover directness in the summation, and may write
α(A ∐ B) = R ⊕ a ⊕ b ⊕ Ln>1
(cid:16)(ab)n ⊕ (ba)n ⊕ ((ab)na) ⊕ ((ba)nb)(cid:17).
For any finite alternating string w formed from the letters a and b, let us write αw
for the map from the corresponding alternating tensor product of factors a and b in
A∐B to the corresponding alternating product in Rhha, bii. Thus, as an R-bimodule
map, A ∐ B → α(A ∐ B) decomposes as the direct sum of the three identity maps
R
αb−→ b, and the further R-bimodule maps, n > 1,
α∅−−→ R, a
αa−−→ a, b
(a⊗b)⊗n
(a⊗b)⊗n ⊗ a
α(ab)n
−−−−→ (ab)n,
α(ab)n a
−−−−−→ (ab)na,
(b⊗a)⊗n
(b⊗a)⊗n ⊗ b
α(ba)n
−−−−→ (ba)n,
α(ba)n b
−−−−−→ (ba)nb.
4
Ring coproducts embedded in power-series rings
Thus α is injective if and only if each of the four maps is injective for all n > 1.
3 Ore localizations
We can now give a simple condition that ensures that α is injective.
3.1 Proposition. If Hypotheses 1.1 hold and, moreover, A is a right Ore localiza-
tion of R[a] and B is a right Ore localization of R[b], then α is injective.
Proof. We use Notation 2.2. We shall show below that for any n > 1, if α(ab)n is
injective, then α(ab)na is injective. The analogous arguments will clearly work for
the other three cases required.
Suppose then that (a⊗b)⊗n
α(ab)n
−−−−→ (ab)n ⊆ Rhha, bii is injective. By considering
{a, b}∗-supports, we see that, for each d > 0, (a ⊗ b)⊗n⊗ Rad → (ab)nad is then
injective. Again by considering {a, b}∗-supports, we see that the map
(cid:16)(ab)nad(cid:17) ⊆ Rhha, bii
Ld>0
(cid:16)(a⊗b)⊗n ⊗ (Rad)(cid:17) → Ld>0
is injective; that is, the map
(a⊗b)⊗n ⊗ (R[a]) → (ab)n(R[a]) ⊆ Rhha, bii
is injective.
Since A is a right Ore localization of R[a], there exists some subset U of U(A)
such that the family (R[a]u : u ∈ U ) forms a directed system under inclusion and
the direct limit of the system is A. For each u ∈ U , we have a commutative diagram
(a⊗b)⊗n ⊗ (R[a]) −−−−→ (ab)n(R[a])
u
y
(a⊗b)⊗n ⊗ (R[a]u) −−−−→ (ab)n(R[a]u)
u
y
in which the maps on the top and the two sides are bijective. Hence, the map on
the bottom,
(a⊗b)⊗n ⊗ (R[a]u) → (ab)n(R[a]u) ⊆ Rhha, bii,
is injective. On taking the directed union over u ∈ U , we find that the map
(a ⊗ b)⊗n ⊗ A → (ab)nA ⊆ Rhha, bii
is injective. By considering {a, b}∗-supports, we see that the latter map is the direct
sum of the two maps α(ab)n and α(ab)na. Hence, α(ab)na is injective.
The other three inductive implications can be proved in the same way. The
injectivity of α then follows by induction.
3.2 Remarks. In Example 1.2, we noted that one case of Proposition 3.1 is the
result of Fox [11] that the natural map RF → Rhha, bii is injective, where F is the
free group on two symbols 1+ a, 1+ b. Fox's proof introduced R-linear endomor-
phisms of RF that are now called Fox derivatives; see [8, Example 7, p. 55]. To
indicate one important connection between the above proof of Proposition 3.1 and
Fox's original argument, we remark that right multiplication by a gives a bijective
map
(a⊗b)⊗n ⊗ (Rad) → (a⊗b)⊗n ⊗ (Rad+1),
and the inverse map has the same action as that of the left Fox derivative on RF
with respect to 1+ a relative to the free generating set {1+ a, 1+ b}.
The original motivation of this article was to show that the proof of the embed-
ding became more transparent when more emphasis was placed on the ring-theoretic
viewpoint.
Pere Ara and Warren Dicks
5
4 Review of Π-semihereditary rings
In this section, we review some of the history of what we call Π-semihereditary rings.
This will provide us with some examples to which the results of the next section
apply. We shall recall the proofs since they are all fairly short, scattered over the
literature, and not always available in the formulation that will be required. No
new results are claimed, but we have tried to give proofs that minimize background
requirements.
4.1 Definitions. For each of the definitions given below, the dual left-right defini-
tion will also be understood.
Let M be a left R-module.
(1) Let X and Y be sets.
For m, n ∈ N, we let mY n denote the set of m × n matrices with coordinates
in Y . Absence of an exponent is to be read as an exponent 1.
Let QX
Y denote the set whose elements are the functions X → Y , x 7→ yx,
represented as families (yx : x ∈ X). We shall sometimes write QX
think of the elements intuitively as '1 × X matrices'. Similarly, we sometimes write
QX
Y as XY , and think of the elements intuitively as 'X × 1 matrices'.
In a natural way, QX
M is a left R-module. By
Y as Y X , and
the multiplication map (QX
R-module map determined by (rx : x ∈ X) ⊗ m 7→ (rxm : x ∈ X). We shall some-
times let (XR) ⊗ M → (XR)M ⊆ XM denote the underlying additive map; here,
XR has a right R-module structure and XM has an abelian group structure. When
M is an R-bimodule, the right R-action is to be maintained in all cases.
M , we shall mean the left
R is an R-bimodule and QX
R)M ⊆ QX
R) ⊗ M → (QX
When M = QY
we have an R-bimodule map
R, we let µX,Y denote the foregoing multiplication map. Here,
(QX
R) ⊗ (QY
R)
µX,Y−−−→ QX×Y
R,
(rx : x ∈ X) ⊗ (sy : y ∈ Y ) 7→ (rxsy : (x, y) ∈ X × Y ).
We shall sometimes denote the underlying additive map by XR ⊗ RY µX,Y−−−→ XRY ,
where we think of the elements of XRY imprecisely as 'X × Y matrices'.
(2) We say that M is torsionless if there exists an injective left R-module map
from M to a direct product of copies of R. Clearly, submodules of torsionless
modules are torsionless.
(3) One says that M is coherent if every finitely generated submodule is finitely
presented. Clearly, submodules of coherent modules are coherent.
The ring R is left coherent if R is coherent as left R-module, that is, every finitely
generated left ideal of R is finitely presented as left R-module.
The ring R is left Π-coherent if, for every set Y , RY is coherent as left R-module,
or, equivalently, every finitely generated, torsionless, left R-module is finitely pre-
sented. This property was studied first by Jones [16, p. 103]. Clearly, left Π-coherent
rings are left coherent.
(4) We say that M is semihereditary if every finitely generated submodule is
projective. Clearly, semihereditary modules are coherent, and submodules of semi-
hereditary modules are semihereditary.
The ring R is left semihereditary if R is semihereditary as left R-module, that
is, every finitely generated left ideal of R is projective as left R-module.
6
Ring coproducts embedded in power-series rings
We now come to the property of interest to us, which was studied first by
Jones [16], who showed that it is left-right symmetric; see Theorem 4.11 below.
We say that the ring R is Π-semihereditary if, for every set X, XR is semiheredi-
tary as right R-module, or, equivalently, every finitely generated, torsionless right
R-module is projective. Prior to Jones' work, Goodearl [12] had mentioned that
right self-injective, von Neumann regular rings have this property; see Theorem 4.17
below.
Clearly, Π-semihereditary rings are right semihereditary and right Π-coherent.
(5) We say that M is flat if, for all injective right R-module maps K → F , the
induced map K ⊗R M → F ⊗R M is injective, or, equivalently, for all finitely gen-
erated right ideals I of R, the multiplication map I ⊗R M → IM ⊆ M = R ⊗R M
is injective. See [23, Theorem 3.53]. Notice that it is therefore sufficent to consider
inclusion maps K → F where F is a finitely generated, free, right R-module and K
is a finitely generated R-submodule of F .
If every finitely generated submodule of M is flat, then M is flat. More generally,
if M is the direct limit of a directed system of flat modules, then M is flat. See [23,
Theorem 3.47].
If M is flat, then every left R-summand of M is flat. See [23, Theorem 3.45].
We write w.gl.dim R 6 1 if every finitely generated left ideal of R is flat as left
R-module, or, equivalently, every submodule of every flat left R-module is flat, or
equivalently every submodule of every flat right R-module is flat; see [23, Theo-
rem 9.24]. Otherwise, we write w.gl.dim R > 2.
(6) For any infinite cardinal κ, we say that R is left κ-Noetherian if every left
ideal of R is generated by κ (or fewer) elements. For example, if R 6 ℵ0, then R
is left ℵ0-Noetherian. Also, if R > ℵ0, then R is left R-Noetherian.
The next result is a variant of [6, Theorem 2.4] suitable for our purposes.
4.2 Theorem. (Cohn, 1959) For any flat left R-module F and left R-submodule K
of F , the following are equivalent.
(a) F/K is left R-flat.
(b) For all ℓ, m ∈ N and y ∈ ℓRm, (y(mF )) ∩ (ℓK) ⊆ y(mK) in ℓF .
Proof. If ℓ, m ∈ N and y ∈ ℓRm, then ℓR may be considered an arbitrary finitely
generated, free, right R-module and y(mR) an arbitrary finitely generated, right
R-submodule of ℓR. Thus, the left R-module M := F/K is flat if and only if the
natural map
(3)
(y(mR)) ⊗R M → (ℓR) ⊗R M
is injective for all ℓ, m, y.
Suppose that ℓ, m, and y are given.
We have an exact sequence
(y(mR)) ⊗R K → (y(mR)) ⊗R F → (y(mR)) ⊗R M → 0,
(4)
and we shall re-interpret each term. Since F is left R-flat, (y(mR)) ⊗R F embeds
in (ℓR) ⊗R F = ℓF , and the image is y(mF ). Hence, we may make the identification
(y(mR)) ⊗R F = y(mF ). The image of (y(mR)) ⊗R K in (y(mR)) ⊗R F = y(mF )
is then y(mK). Using (4), we now find that we may make the identification
(y(mR)) ⊗R M = (y(mF ))/(y(mK)).
In the special case where y is an identity matrix, we have the (obvious) iden-
tification (ℓR) ⊗R M = (ℓF )/(ℓK). We now find that we may identify (3) with the
natural map (y(mF ))/(y(mK)) → (ℓF )/(ℓK). Then (a) ⇔(b) follows.
The next two results are taken from [5, Proposition 2.2 and Corollary], attributed
to O. E. Villamayor.
Pere Ara and Warren Dicks
7
4.3 Theorem. (Villamayor, 1960) For any n ∈ N and left R-submodule K of Rn,
the following are equivalent.
(a) Rn/K is left R-flat.
(b) For all ℓ ∈ N and x ∈ ℓK ⊆ ℓRn, x ∈ x(nK).
Proof. The case F = Rn of Theorem 4.2 allows us to rewrite (a) as
(a′) For all ℓ, m ∈ N and y ∈ ℓRm, (y(mRn)) ∩ (ℓK) ⊆ y(mK) in ℓRn.
(a′)⇒(b). Suppose x ∈ ℓK. Taking m = n and y = x in (a′) we obtain
(x(nRn)) ∩ (ℓK) ⊆ x(nK) in ℓRn. Then x = xIn ∈ x(nRn) ∩ (ℓK) ⊆ x(nK).
(b)⇒(a′). Suppose y ∈ ℓRm. Consider any x ∈ (y(mRn)) ∩ (ℓK). We have
x ∈ ℓK, and then, by (b), x ∈ (x)(nK) ⊆ (y(mRn))(nK) ⊆ y(mK).
4.4 Theorem. (Villamayor, 1960) For any left R-module M , the following are
equivalent.
(a) M is finitely generated and projective.
(b) M is finitely presented and flat.
Proof. (a)⇒(b) is clear.
(b)⇒(a). Since M is a finitely presented left R-module, there exist n ∈ N and
a finitely generated left R-submodule K of Rn such that M = Rn/K. Since K
is finitely generated, there exist ℓ ∈ N and x ∈ ℓK ⊆ ℓRn such that K = (Rℓ)x.
By Theorem 4.3 (a)⇒(b), since M is R-flat, x ∈ x(nK), that is, there exists some
e ∈ nK such that x = x·e. Now the left R-module map Rn → K, y 7→ y·e, acts as
the identity on K because it does so on the rows of x, which form a generating set
of K. This proves that Rn/K is projective.
The following result is a variant of [20, Satz 2] which is the same as [9, Propo-
sition 2.1], attributed to M. Auslander.
4.5 Theorem. (Lenzing, 1969 & Auslander, 1970) For any finitely generated left
R-module M , the following are equivalent.
(a) M is finitely presented.
(b) For every set X, the multiplication map (XR) ⊗R M → XM is injective.
(c) For some infinite set X such that R is left X-Noetherian, the multiplication
map (XR) ⊗R M → XM is injective.
Proof. (a)⇒ (b). By (a), there exist m, n ∈ N and an exact sequence of left
R-modules Rm → Rn → M → 0. We then have a commutative diagram with exact
rows:
(XR) ⊗R (Rm) −−−−→ (XR) ⊗R (Rn) −−−−→ (XR) ⊗R M −−−−→ 0
µX,m(cid:13)(cid:13)(cid:13)
XRm
µX,n(cid:13)(cid:13)(cid:13)
XRn
−−−−→
y
XM
−−−−→
−−−−→ 0.
Hence, (XR) ⊗R M → XM is injective (and surjective).
(b)⇒ (c) is clear.
(c)⇒ (a). By hypothesis, there exist n ∈ N and an exact sequence of left
R-modules 0 → K → Rn → M → 0, and it suffices to show that K is finitely
generated. We then have a commutative diagram with exact rows:
(XR) ⊗R K −−−−→ (XR) ⊗R (Rn) −−−−→ (XR) ⊗R M −−−−→ 0
y
XK
µX,n(cid:13)(cid:13)(cid:13)
XRn
−−−−→
(cid:13)(cid:13)(cid:13)
XM
−−−−→
0 −−−−→
−−−−→ 0,
8
Ring coproducts embedded in power-series rings
where (XR) ⊗R M → XM is injective by (c), and is here easily seen to be surjec-
tive. Hence, (XR) ⊗R K → XK is surjective, that is, (XR)K = XK. Since R is left
X-Noetherian by (c), K is X-generated; hence, there exists some generating fam-
ily (kx : x ∈ X) which may be viewed as an element of XK, and it then has the form
Pm
i=1 rx,ik′
i.
Hence, K = Px∈X Rkx ⊆ Pm
i, since XK = (XR)K. Then, for each x ∈ X, kx = Pm
i ⊆ K. Thus, K is finitely generated.
i=1(rx,i : x ∈ X)k′
i=1 Rk′
The next two results are taken from [5, Theorems 2.1 and 4.1].
4.6 Theorem. (Chase, 1960) The following are equivalent.
(a) R is left coherent, i.e. every finitely generated left ideal is finitely presented.
(b) For every finitely generated left ideal I of R and every set X, the multiplication
map (XR) ⊗R I → (XR)I ⊆ XI ⊆ XR = (XR) ⊗R R is injective.
(c) For every set X, XR is right R-flat.
Proof (after Lenzing [20, Section 4]). (a)
Theorem 4.5
⇐======⇒ (b)
clear⇐==⇒ (c).
4.7 Theorem. (Chase, 1960) The following are equivalent.
(a) R is left semihereditary, i.e. every finitely generated left ideal is projective.
(b) w.gl.dim R 6 1 and R is left coherent.
(c) w.gl.dim R 6 1 and, for every set X, XR is right R-flat.
(d) Every torsionless right R-module is flat.
(e) Every finitely generated, torsionless, right R-module is flat.
Proof. (a) Theorem 4.4
⇐======⇒ (b) Theorem 4.6
⇐======⇒ (c)
clear⇐==⇒ (d)
clear⇐==⇒ (e).
Goodearl [12, Theorem 1] generalized Theorem 4.5 to the non-finitely generated
case by showing that, for any left R-module M , all of the multiplication maps
(XR) ⊗R M → XM are injective if and only if, for each finitely generated submodule
N of M , the inclusion map N → M factors through a finitely presented module. We
record the special case we shall be using and its special proof. Recall the meaning
of µX,Y from Definitions 4.1(1).
4.8 Theorem. (Goodearl, 1972) For any non-empty set Y , the following are
equivalent.
(a) RY is coherent as left R-module.
(b) R is a left coherent ring, and, for every set X, µX,Y is injective.
(c) R is a left coherent ring, and, for some infinite set X such that R is left
X-Noetherian, µX,Y is injective.
Proof. Let M := RY , and let M denote the directed family of finitely generated left
R-submodules of M .
(a)⇒(b).
Since Y is non-empty, R embeds in M , and, hence, R is co-
Let N ∈ M.
herent as left R-module, that is, R is a left coherent ring.
By (a), N is finitely presented. By Theorem 4.5 (a)⇒(b), the multiplication map
(XR) ⊗ N → XN ⊆ XM is injective. On taking the direct limit of the system of
µX,Y−−−→ XM
injective maps (XR) ⊗ N → XM over all N ∈ M, we see that (XR) ⊗ M
is injective, as desired.
(b)⇒(c) is clear.
(c)⇒(a) (after Herbera-Trlifaj [13, Corollary 2.11] ). Let N ∈ M and consider
the natural commutative diagram
Pere Ara and Warren Dicks
9
(XR) ⊗ N −−−−→ XN
y
(XR) ⊗ M
y
µX,Y−−−−→ XM.
By Theorem 4.6 (a)⇒(c), XR is right R-flat, and, hence, the left map is injective.
By (c), the bottom map is injective. Hence, the top map is injective. By Theorem 4.5
(c)⇒(a), N is finitely presented. Thus, M is coherent as left R-module.
The following result is a variant of [2, (4.5)] that is suitable for our purposes.
4.9 Theorem. (Bass, 1960) Let Y be an infinite set such that R is right
Y -Noetherian. If M is a finitely generated, torsionless, left R-module, then M
embeds in RY . Hence, R is a left Π-coherent ring if and only if RY is a coherent
left R-module.
Proof. Since M is a finitely generated left R-module, on dualizing we find that the
right R-module M ∗ := HomR(M, RR) embeds in a finitely generated, free, right
R-module. Since R is right Y -Noetherian, M ∗ is Y -generated. On dualizing we
find that the left R-module M ∗∗ := HomR(M ∗, RR) embeds in RY . Since M is
torsionless, the double-dual map M → M ∗∗ is injective, and M embeds in RY .
4.10 Corollary. The following are equivalent.
(a) R is a left Π-coherent ring, i.e. every finitely generated, torsionless, left module
is finitely presented.
(b) R is a left coherent ring and, for all sets X and Y, µX,Y is injective.
(c) R is a left coherent ring and, for some infinite sets X and Y such that R is left
X-Noetherian and right Y -Noetherian, µX,Y is injective.
Proof. This follows from Theorems 4.8 and 4.9.
We now recall the aspect of [16, Theorem 2.11] that is of interest to us.
4.11 Theorem. (Jones, 1982) The following are equivalent.
(a) R is left and right semihereditary, and, for all sets X, Y , µX,Y is injective.
(b) R is left and right coherent, w.gl.dim R 6 1, and, for all sets X, Y , µX,Y is
injective.
(c) R is left coherent, w.gl.dim R 6 1, and R is right Π-coherent.
(d) Every finitely generated, torsionless, right R-module is flat and finitely pre-
sented.
(e) R is Π-semihereditary, i.e. every finitely generated, torsionless, right R-module
is projective.
These conditions are also equivalent to their left-right duals.
Proof. (a) Theorem 4.7
⇐======⇒ (b)
(a) is left-right symmetric.
Corollary 4.10
⇐=======⇒ (c) Theorem 4.7
⇐======⇒ (d) Theorem 4.4
⇐======⇒ (e), and
In the remainder of this section, we recall some types of rings that are
Π-semihereditary. We first present a result from [24].
4.12 Theorem. (Sahaev, 1965) For each n ∈ N and each y ∈ nRn, let
In− y. The following are equivalent.
(a) There exists a finitely generated, non-finitely presented, flat, left R-module M .
(b) There exist n ∈ N and a function N → nRn, k 7→ yk, such that, for each
•yk; equivalently, ykyk+1 = yk and yk+1yk 6= yk+1;
•yk+1 = 0n 6= yk+1
•y denote
k ∈ N, yk
equivalently,
•yk
•yk+1 = •yk+1, and
•yk+1
•yk 6= •yk.
10
Ring coproducts embedded in power-series rings
Proof. (a)⇒ (b). Since M is finitely generated and not finitely presented, there exist
n ∈ N and a non-finitely generated left R-submodule K of Rn such that M = Rn/K.
We view Rn as an (R, nRn)-bimodule.
We shall recursively construct a sequence (yk : k ∈ N) in nK ⊆ nRn.
Set
y0 := 0n ∈ nK. We may now suppose that k ∈ N and yk ∈ nK. Since K is not
finitely generated, it is easy to see that there exists some xk ∈ n+1K ⊆ n+1Rn
such that (Rn)yk ⊂ (Rn+1)xk ⊆ K. By Theorem 4.3 (a)⇒(b), since M is
flat, xk ∈ xk(nK), that is, there exists some yk+1 ∈ nK such that xk = xkyk+1.
•yk+1 = {0}, and we see that
Hence, xk
yk
•yk+1 = 0n.
Also, (Rn)yk+1yk ⊆ (Rn)yk ⊂ (Rn+1)xk = (Rn+1)xkyk+1 ⊆ (Rn)yk+1, and we
•yk+1 = 0n. Now, (Rn)yk
•yk+1 ⊆ (Rn+1)xk
see that yk+1yk 6= yk+1. Thus, yk+1
•yk 6= 0n.
This completes the recursive construction of the sequence, and (b) holds.
(b)⇒ (a). Let F denote the (R, nRn)-bimodule Rn. Then (F yk : k ∈ N) is a
family of left R-submodules of F . For each k ∈ N, F yk = F ykyk+1 ⊆ F yk+1; also
•yk+1, and, hence, F yk+2 6= F yk. Thus, (F yk : k ∈ N ) is a
F yk+2
non-stationary, ascending chain of left R-submodules of F .
•yk+1 6= {0} = F yk
Set K := S∞
k=0(F yk) and M := F/K. Then M is a finitely generated,
non-finitely presented, left R-module. To prove (a), it suffices to show that M
is left R-flat, and here we may apply Theorem 4.3 (b)⇒ (a). Let ℓ ∈ N and x ∈ ℓK;
it remains to show that x ∈ x(nK). As x ∈ ℓK = S∞
k=0(ℓ(F yk)), there exists some
k ∈ N such that x ∈ ℓ(F yk). Then x •yk+1 = 0. Since each row of yk+1 lies in K,
yk+1 ∈ nK. Now, x = xyk+1 ∈ x(nK), as desired.
4.13 Definitions. (1) We shall say that R is a left Sahaev ring if every finitely
generated, flat, left R-module is finitely presented, or, equivalently, projective, by
Theorem 4.4. The term 'left S-ring' is sometimes used in the literature.
Every subring of a left Sahaev ring is again a left Sahaev ring, since these are
the rings for which the condition in Theorem 4.12(b) fails. In particular, rings that
embed in skew fields are left and right Sahaev rings.
It is clear from the foregoing definition that left Noetherian rings are left Sahaev
rings. Both Sahaev [24] and S. Jøndrup [15, Lemma 1.4] discuss connections between
the condition in Theorem 4.12(b) and various chain conditions.
(2) A commutative Prufer domain is a semihereditary, commutative domain.
A commutative B´ezout domain is a commutative domain in which every finitely
generated ideal is principal.
(3) One says that R is a semifir if every finitely generated left ideal of R is free of
unique rank, as left R-module, and similarly on the right. It is sufficient to assume
the condition on the left; see [8, Theorem I.1.1].
We now present [15, Corollary 2].
4.14 Theorem. (Jøndrup, 1971) The following hold.
(i) If R is right semihereditary and not left Sahaev, then R is not right Sahaev.
(ii) If w.gl.dim R 6 1 and R is right Sahaev, then R is left Sahaev and Π-semi-
hereditary.
Proof. (i) We apply Theorem 4.12 (a)⇒(b), and we change y0 to 0n. We shall speak
of the kernel and image of each element of nRn viewed as an R-endomorphism of
nR acting on the left by matrix multiplication.
Let k ∈ N. Since yk+2
•yk+1, Ker(yk+2) 6⊇ Im( •yk+1) ⊆ Ker(yk).
Hence, Ker(yk+2) 6= Ker(yk). Clearly •yk+1 acts as the identity on Ker(yk+1),
and we see that Ker(yk+1) ⊆ Im( •yk+1) ⊆ Ker(yk). Hence, Ker(y2k+2) ⊂ Ker(y2k).
Since R is right semihereditary, Im(y2k+2) is a projective R-submodule of nR, and,
•yk+1 6= 0n = yk
Pere Ara and Warren Dicks
11
hence, Ker(y2k+2) is an R-summand of nR, and, therefore, of Ker(y2k). Thus,
Ker(y2k) = Ker(y2k+2) ⊕ Lk for some Lk 6= {0}. Then,
k
nR = Ker(y0) = Ker(y2k+2) ⊕ (
Li).
Li=0
Each term of the directed system (nR/(Lk
taking the direct limit, we see that nR/(L∞
presented, and flat. Hence, R is not right Sahaev.
i=0 Li) : k ∈ N) is then R-projective. On
i=0 Li) is finitely generated, not finitely
(ii) As w.gl.dim R 6 1, every finitely generated right ideal of R is flat, and hence
projective, since R is right Sahaev. Thus, R is right semihereditary. By the contra-
positive of (i), R is left Sahaev. Hence, R is left semihereditary. By Theorem 4.7
(a)⇒(e), every finitely generated, torsionless, right R-module is flat, and hence
projective, since R is right Sahaev.
4.15 Corollary. The following hold.
(i) If w.gl.dim R 6 1 (e.g. R is left or right (semi)hereditary) and R is left or right
Sahaev (e.g. R is left or right Noetherian or R embeds in a skew field ), then
R is Π-semihereditary.
(ii) If R is a commutative principal ideal domain, or more generally, a commuta-
tive B´ezout domain, or, even more generally, a commutative Prufer domain,
then R is Π-semihereditary.
Proof. (i) is Theorem 4.14(ii) and its left-right dual, while (ii) gives some cases of (i)
where R is semihereditary and embeds in a field.
The following is [14, Theorem 2B], attributed to Cohn.
4.16 Theorem. (Cohn, 1969) Semifirs are left and right semihereditary, left and
right Sahaev rings. Hence, semifirs are Π-semihereditary.
Proof (after Jøndrup [15, Corollary 3]). It is clear that semifirs are left and right
semihereditary. From the proof of Theorem 4.14(i), it may be seen that if R is right
semihereditary and not left Sahaev, then, for some n ∈ N, nR is a direct sum of n+1
non-zero right R-submodules; this cannot happen if R is a semifir. Hence semifirs
are left (and right) Sahaev rings, and Corollary 4.15(i) applies.
We now turn to the final topic of this section, self-injective rings; for background
on injective modules, see [23, pp. 65 -- 68].
At the beginning of the proof of [12, Theorem 2 (c)⇒(a)], Goodearl notes the
consequence of [4, Theorem 2.1 (a)⇒(b)] that, over a right self-injective, von Neu-
mann regular ring, every torsionless right module is projective. This seems to have
been one of the very earliest occasions on which the projectivity of finitely generated
torsionless modules was considered.
4.17 Theorem. (Cateforis, 1969) If R is right semihereditary and right self-injec-
tive, then R is Π-semihereditary.
Proof. By definition, R is Π-semihereditary if each finitely generated, torsionless,
right R-module M is projective. Since M is finitely generated, we may write
M = F/K where F is a finitely generated, free, right R-module, and K is a submod-
ule of F . In particular, F is right R-injective. Let L denote the set of submodules
of F whose intersection with K is {0}. Clearly, L 6= ∅. By Zorn's lemma, L con-
tains some ⊆-maximal element L. Then F ⊇ K⊕ L. We then have injective right
R-module maps K → F/L, k 7→ k+L, and K → F , k 7→ k. Since F is right
R-injective, the domain of the latter map can be transported along the former map
to give a right R-module map φ : F/L → F such that, for all k ∈ K, φ(k+L) = k.
12
Ring coproducts embedded in power-series rings
Define φ′ : F → F by f 7→ φ(f +L). Then Ker(φ′) ⊇ L, and, for all k ∈ K,
φ′(k) = k. If Im(φ′) = K, then F/K is projective, as desired. Thus, it suffices to
assume that Im(φ′) ⊃ K, and obtain a contradiction.
Define φ′′ : F → M by f 7→ φ′(f )+K ∈ F/K = M . Then there exists some
m ∈ Im(φ′′)−{0}. Thus, F ⊃ Ker(φ′′) ⊇ K⊕L.
Since M is torsionless, we may compose φ′′ with a map from M to R which
does not vanish on m, and thus obtain a map ψ : F → R with F ⊃ Ker(ψ) ⊇ K⊕L.
Then Im(ψ) is a finitely generated right ideal of R. Since R is right semihereditary,
Im(ψ) is projective. Hence, F = Ker(ψ)⊕L′ for some L′ 6= {0}. Now F ⊇ K⊕L⊕L′
and this contradicts the ⊆-maximality of L in L, as desired.
4.18 Remarks. Recall that if R is von Neumann regular, then every finitely pre-
sented right R-module is projective by Theorem 4.4, hence every finitely generated
right ideal of R is a summand of R, and hence R is right semihereditary.
Recall also that if R is right semihereditary and right self-injective, then every
finitely generated right ideal is projective, hence injective, hence a summand of R,
and R is von Neumann regular.
We shall be using [22, Corollary], which is as follows.
4.19 Theorem. (Osofsky, 1964) If R is right self-injective and right hereditary,
then R is completely reducible.
Proof. Suppose not. We shall obtain a contradiction. Let U denote the set of right
ideals of R that are not finitely generated. By Remarks 4.18, R is von Neumann
regular; since R is not completely reducible, U 6= ∅. By Zorn's lemma, U contains
some ⊆-maximal element I.
Since R is right hereditary, I is right R-projective. Since R is von Neumann
regular, results of Kaplansky show that I is a direct sum of infinitely many nonzero
right ideals of R; in the case where I is ℵ0-generated, this follows from [18, Proof
of Lemma 1], and in the case where I is not ℵ0-generated, it follows from [17,
Theorem 1]. There then exists some decomposition I = I ′ ⊕ I ′′ with I ′, I ′′ ∈ U.
We shall now show that there exists some right ideal I ′ of R with the properties
that I ′ ⊆ I ′, I ′ 6∈ U, and I ′ ∩ I ′′ = {0}; in Remarks 4.20, we shall give an alter-
native argument which does not require background on injective hulls. Recall that
Eckmann and Schopf [10] showed that, since R is right self-injective, I ′ has a right
R-injective hull I ′ in R which is a maximal right R-essential extension of I ′ in R;
see [23, proof of Theorem 3.30 (iii)⇒(i)]. Then I ′ is a right R-summand of R, and,
hence, I ′ is right R-cyclic. Thus, I ′ 6∈ U. Also, since I ′ ∩ I ′′ = {0}, it follows that
I ′ ∩ I ′′ = {0}, as claimed.
Since I ′
6∈ U, we have I 6= I ′ and I ⊂ I ′. Since I ′′ ∈ U, it can be seen that
I ′ ⊕ I ′′ ∈ U. Now I = I ′ ⊕ I ′′ ⊂ I ′ ⊕ I ′′. This contradicts the maximality of I in U
and completes the proof.
4.20 Remarks. Here is an alternative proof of the existence of an I ′. Let
M := R/I ′, and consider the double-dual map φ : M → M ∗∗ as in the proof of The-
orem 4.9. There exists a right ideal I ′ of R such that Ker(φ) = I ′/I ′, and then
Im(φ) ≃ R/I ′. By Theorem 4.17, R is right Π-coherent.
It is clear that Im(φ)
is a finitely generated, torsionless, right R-module, and, hence, Im(φ) is finitely
presented. Thus, I ′ is finitely generated, and, hence, I ′ 6∈ U. We shall now show
that I ′ ∩ I ′′ = {0}. Consider any x ∈ I ′′. Then xR = eR for some idempotent
e ∈ R, since R is von Neumann regular. Since eR ⊆ I ′′, we see that eR ∩ I ′ = {0},
and, hence, eR embeds in R/I ′ = M . Since R is right self-injective, eR is right
R-injective. Thus, eR becomes a summand of M which, as eR is projective,
Pere Ara and Warren Dicks
13
must then embed in M ∗∗. We now see that eR ∩ I ′ = {0}.
I ′ ∩ I ′′ = {0}, as claimed.
It then follows that
The following combines [12, Theorem 2 (a)⇒(c)] and the commutative case of
Theorem 4.19.
4.21 Theorem. Let R be a commutative, Π-coherent, von Neumann regular ring.
(i) (Goodearl, 1972) R is self-injective.
(ii) (Osofsky, 1964) If R is ℵ0-Noetherian or, more generally, hereditary, then R
is completely reducible.
Proof. (i) (based loosely on Kobayashi [19, Theorem 2]) By Baer's criterion [23,
Theorem 3.20], to show that R is self-injective, it suffices to show that, for each ideal
I of R, each R-module map φ : I → R is given by multiplication by some element
of R.
Now π1 : R2 → R, (x, y) 7→ x, and ψ : R2 → RI, (x, y) 7→ (φ(x·i) − y·i : i ∈ I),
are R-module maps. Then Im(ψ) is a finitely generated, torsionless R-module.
Since R is Π-coherent, Ker(ψ) is a finitely generated R-submodule of R2. Hence,
π1(Ker(ψ)) is a finitely generated ideal of R. Since R is von Neumann regular, there
exists some idempotent e ∈ R such that π1(Ker(ψ)) = R·e. For each x ∈ R, we now
see that x ∈ R·e if and only if there exists some y ∈ R such that, for each i ∈ I,
φ(x·i) = y·i, that is, φ(i·x) = i·y, since R is commutative.
We apply the 'if' part taking x ∈ I and y = φ(x). For each i ∈ I, we have
φ(i·x) = i·φ(x) = i·y. Hence, x ∈ R·e. Thus, I ⊆ R·e and i·e = i for all i ∈ I.
We now apply the 'only if' part taking x = e. There exists some y ∈ R such
that, for each i ∈ I, φ(i·e) = i·y, that is, φ(i) = i·y, as desired.
(ii) Recall that a right ℵ0-Noetherian, von Neumann regular ring is right hered-
itary, by [18, Lemma 1]. Now (ii) follows from (i) and Theorem 4.19.
4.22 Remarks. If w.gl.dim R > 2, then R is not Π-semihereditary.
If w.gl.dim R = 1 and R is right Sahaev, then R is left Sahaev and Π-semihered-
itary, by Theorem 4.14(ii). This class includes all the semifirs that are not skew
fields, by Theorem 4.16. It also includes all the Prufer domains that are not fields.
If w.gl.dim R = 0 and R is left or right self-injective, then R is Π-semihereditary,
by Theorem 4.17. This class includes all the right semihereditary, right self-injective
rings.
It also includes all the commutative Π-semihereditary rings which have
w.gl.dim R = 0, by Theorem 4.21(i). It also includes all the completely reducible
rings.
If w.gl.dim R = 0 and R is right Sahaev, then clearly every finitely generated
right R-module is projective, and R is then completely reducible. Such rings are
included in the previous case.
Thus, we have two disjoint classes of examples of Π-semihereditary rings.
5
If R is Π-semihereditary, then α is injective
For the remainder of the article, we fix the following.
5.1 Notation. For each n ∈ N, we have a multiplication map that will be denoted
(QN
R)⊗n βn−−→ QNn
R.
Whenever Hypotheses 1.1 hold, we have a natural factorization of α that will be
denoted A ∐R B α′
−→ Rhhaii ∐R Rhhbii α′′
−−→ Rhha, bii.
14
Ring coproducts embedded in power-series rings
5.2 Proposition. If Hypotheses 1.1 hold and A, B, and QN
R-flat, then A ∐R B α′
Proof. Clearly, the R-bimodule map QN
is bijective. Similarly for bRhhbii.
−→ Rhhaii ∐R Rhhbii is injective.
R → aRhhaii, (ri : i ∈ N) 7→ Pi>0(riai+1),
R are left and right
We see that aRhhaii and bRhhbii are left and right R-flat.
We use Notation 2.2. Now, A = R ⊕ a and B = R ⊕ b; hence, a and b are left
and right R-flat.
Let n ∈ N. For 1 6 i 6 n, let Mi denote one of aRhhaii, a, bRhhbii, or b. It
follows by induction that the R-bimodule M1 ⊗ M2 ⊗ · · · ⊗ Mn is left and right
R-flat. It then follows by induction that each of the natural maps
(a ⊗ b)⊗n → (aRhhaii ⊗ bRhhbii)⊗n,
(b ⊗ a)⊗n → (bRhhbii ⊗ aRhhaii)⊗n,
(a ⊗ b)⊗n ⊗ a → (aRhhaii ⊗ bRhhbii)⊗n ⊗ aRhhaii,
(b ⊗ a)⊗n ⊗ b → (bRhhbii ⊗ aRhhaii)⊗n ⊗ bRhhbii,
is injective. By using the standard R-bimodule decompositions of A ∐ B and
Rhhaii ∐ Rhhbii, we see that, as an R-bimodule map, α′ is a direct sum of these
injective maps, and hence is injective.
5.3 Proposition. Suppose that Hypotheses 1.1 hold.
(i) Rhhaii ∐R Rhhbii α′′
−−→ Rhha, bii is injective if and only if, for each n ∈ N,
(QN
R)⊗n βn−−→ QNn
R is injective.
R is left or right R-coherent, then α′′ is injective.
(ii) If R is left or right coherent and β2 is injective, then α′′ is injective.
(iii) If QN
Proof. (i) We may assume that A = Rhhaii, B = Rhhbii and α = α′′. We use
Notation 2.2. Let n ∈ N. Consider the map (a ⊗ b)⊗n → (ab)n ⊆ Rhha, bii.
The R-bimodule isomorphisms QN
R → b give an embedding
(QN
ding QN2n
the map (a ⊗ b)⊗n → \(ab)n is a copy of β2n. We now see that α is a direct sum
of R-bimodule maps that consist of one copy of β0 and two copies of βn for each
n > 1. In particular, α is injective if and only if all the βn, n ∈ N, are injective.
R → Rhha, bii with image \(ab)n, the closure of (ab)n in Rhha, bii. Thus
R)⊗(2n) → Rhhaii ∐ Rhhbii with image (a ⊗ b)⊗n. They also give an embed-
R → a and QN
(ii) By symmetry, we may assume that R is left coherent.
We shall show by induction that, for each n ∈ N, βn is injective. It is clear that
β0 is injective. Suppose that n > 0, and that βn is injective. Notice that βn+1 may
be factored as
(Q
R)⊗βn
N
β′
R)⊗n)
(QN
R) ⊗ ((QN
−−−−−−→ (QN
R) ⊗ ( QNn
Since R is left coherent, by Theorem 4.6 (a)⇒(c), QN
R) ⊗ βn is injective. Since there exists a bijection between Nn and N, we see that
(QN
β′
2 may be identified with β2 and is then injective. Hence, the composition βn+1 is
injective. This completes the inductive argument. By (i), α′′ is injective.
R is right R-flat, and, hence,
R)
2−→ QNn+1
R.
(iii) By symmetry, we may assume that QN
R is left R-coherent. By Theo-
rem 4.8 (a)⇒(b), R is left coherent and β2 is injective. By (ii), α′′ is injective.
Pere Ara and Warren Dicks
15
R is a left or right semihereditary R-module, then α is injective.
5.4 Corollary. With Hypotheses 1.1, the following hold.
(i) If QN
(ii) If R is a Π-semihereditary ring, then α is injective.
(iii) If R is right semihereditary (e. g. w.gl.dim R = 0) and R is right self-injective,
then α is injective (and w.gl.dim R = 0).
(iv) If w.gl.dim R 6 1 (e. g. R is left or right semihereditary) and R is left or right
Sahaev, then α is injective (and R is left and right semihereditary and left and
right Sahaev).
Proof. (i) By symmetry, we may assume that QN
Proposition 5.3(iii), α′′ is injective.
R is left R-semihereditary. By
Notice that QN
R is left R-coherent and that every left R-submodule of QN
R is
flat. In particular, R is left coherent and w. gl. dim R 6 1. By Theorem 4.6 (a)⇒(c),
QN
R, Rhhaii,
Rhhbii, A, and B, are left and right R-flat. By Proposition 5.2, α′ is injective.
R is right R-flat, and then all its right R-submodules are flat. Hence, QN
Hence, α is injective.
(ii) follows from (i).
(iii) follows from (ii) and Theorem 4.17.
(iv) follows from (ii) and Corollary 4.15(i).
In the next section, we shall see results of Bergman and Goodearl that show
that α can fail to be injective whenever R is a commutative ring such that either
w. gl. dim R > 2 or R is ℵ0-Noetherian, non-self-injective, and von Neumann regular.
6 Non-injectivity phenomena
In this section, we give examples contributed by G. M. Bergman where Hypothe-
ses 1.1 hold and α is not injective. Throughout, we use the factorization of α as
A ∐R B α′
−→ Rhhaii ∐R Rhhbii α′′
−−→ Rhha, bii.
We begin with a general example where R[a] ⊆ A ⊆ R[a, (1+ a)−1] and
R[b] ⊆ B ⊆ R[b, (1+ b)−1], but α′ is not injective.
6.1 Proposition. (Bergman) Suppose that R has two-sided ideals I, J such that
the multiplication map I ⊗ J → IJ is not injective, e.g. R = Z/p2Z, p prime, or
R = Q[x, y]. Then, with a suitable choice of A and B, Hypotheses 1.1 hold and α′,
α are not injective.
Proof. For R = Z/p2Z, p prime, we may take I = J = pR, where p ⊗ p 7→ 0.
For R = Q[x, y], we may take I = J = xR + yR, where (x ⊗ y)−(y ⊗ x) 7→ 0.
Let h := 1+ a and H := hhi 6 U(Rhhaii). Then R[h] = R[a], and we view
R[h] ⊆ RH ⊆ Rhhaii. We have R-bimodule decompositions
RH = Li∈Z
hiR = R[h] ⊕ Li6−1
(hiR).
In RH, take A to be the subring R[h]⊕ Li6−1
(hiI). Since h−1 = 1−h−1a, R+ aA = A.
Let k := 1+ b, K := hki 6 U(Rhhbii). We view R[b] = R[k] ⊆ RK ⊆ Rhhbii,
and we have an R-bimodule decomposition
RK = R[k] ⊕ Li6−1
(Rki).
16
Ring coproducts embedded in power-series rings
In RK, take B to be the subring R[k]⊕ Li6−1
(Jki). Since k−1 = 1−k−1b, R+ bB = B.
We use Notation 2.2. Here,
a = aA = aR[h] ⊕ Li6−1
(hiaI),
and
b = bB = bR[k] ⊕ Li6−1
(Jbki).
In particular, h−1aI is an R-bimodule direct summand of a, and Jbk−1 is an
R-bimodule direct summand of b. Hence (h−1aI) ⊗ (Jbk−1) may be viewed as
an R-bimodule direct summand of a ⊗ b, and also of A ∐ B. This summand does
not map injectively to (h−1aI)(Jbk−1) in (RH) ∐ (RK). It follows that α′ is not
injective in this case. Hence, α is not injective.
Recall that w.gl.dim R > 2 if and only if there exists a right ideal I of R that is
not right R-flat; recall also that I is not right R-flat if and only if there exists a left
ideal J of R such that the multiplication map I ⊗R J → IJ is not injective. The
next result then follows from Proposition 6.1.
6.2 Corollary. Suppose that R is commutative and that w.gl.dim R > 2. Then with
a suitable choice of A and B, Hypotheses 1.1 hold and α is not injective.
6.3 Remarks. Suppose that R is a commutative Sahaev ring and that Hypothe-
ses 1.1 hold.
By Corollary 6.2, if α is injective for all choices of A and B, then w.gl.dim R 6 1.
Conversely, by Corollaries 5.4 and 4.15 (i), if w.gl.dim R 6 1, then α is injective.
Thus, among the commutative Sahaev rings R, the ones for which α is always
injective are the ones with w.gl.dim R 6 1, that is, the semihereditary ones.
Hence, among the commutative domains R, the ones for which α is always
injective are the commutative Prufer domains.
Also, among the commutative Noetherian rings R, the ones for which α is always
injective are the hereditary ones.
We now give examples where Hypotheses 1.1 hold and β2 is not injective, and,
hence, by Proposition 5.3(i), α′′ is not injective. These are then examples where
A = Rhhaii and B = Rhhbii in Hypotheses 1.1 and α is not injective. We remark
that if we had been working in the setting of commutative rings of power series,
then α′′ would have looked like β2.
The following result gives a method for constructing examples where β2 is not
injective.
6.4 Proposition. (Bergman) Let Q be a field. For i = 1, 2, let Ri be a commutative,
augmented Q-ring whose augmentation ideal Ii is not finitely generated, and let Xi
be an infinite set such that Ii is Xi-generated. Let R := (R1 ⊗Q R2)/(I1 ⊗Q I2), a
commutative Q-ring. Then µX1,X2 : (X1R) ⊗R (RX2) → X1RX2 is not injective.
In detail, for i = 1, 2, any family mi := (ri,ji : ji ∈ Xi) that generates Ii as
R, and then m1 ⊗ m2 is a non-zero
ideal of Ri may be viewed as an element of QXi
element of Ker(µX1,X2 ).
Proof. Notice that R1 ⊗Q R2 = (Q ⊕ I1) ⊗Q (Q ⊕ I2) = Q ⊕ I1 ⊕ I2 ⊕ (I1 ⊗Q I2),
R = Q ⊕ I1 ⊕ I2, R1 and R2 are subrings of R, I1 and I2 are ideals of R with
product zero, and R is a commutative augmented Q-ring with augmentation ideal
I := I1 ⊕ I2.
Here, µX1,X2 carries m1 ⊗ m2 to (r1,j1 r2,j2 : (j1, j2) ∈ X1 × X2), which is zero, by
our definition of R. Thus, it remains to show that m1 ⊗ m2 6= 0 in (X1R) ⊗R (RX2 ),
and we shall do this by finding a quotient R-module in which the image of m1 ⊗ m2
is not zero.
Pere Ara and Warren Dicks
17
Let M1 := X1(R/I2), a quotient R-module of X1R with trivial I2-action. Let
M2 := (R/I1)X2 , a quotient R-module of RX2 with trivial I1-action. For i = 1, 2, set
M i := Mi/(MiI) = Mi/(MiIi), a quotient R-module of QXi
R with trivial I-action,
and let mi denote the image of mi in M i. Then we have a map
(X1R) ⊗R (RX2 ) → M 1 ⊗R M 2 = M 1 ⊗Q M 2,
with m1 ⊗R m2 mapping to m1 ⊗Q m2. It suffices to show that m1 ⊗Q m2 6= 0, and,
since Q is a field, it suffices to show that m1 6= 0 and m2 6= 0. By symmetry it
suffices to assume that m1 = 0 and obtain a contradiction, as follows.
The natural bijection R1 → R/I2 induces a bijection X1R1 → M1, and it carries
(X1R1)I1 to M1I1. Since m1 ∈ X1R1 and m1 = 0, we have m1 ∈ (X1R1)I1. Hence,
there is some expression m1 = Pℓ
k=1 vkrk with each vk ∈ X1R1 and each rk ∈ I1.
For each j1 ∈ X1, the j1st coordinate of m1 is r1,j1 , and, hence, r1,j1 lies in the
ideal of R1 generated by {rk : 1 6 k 6 ℓ}, a finite subset of I1. It follows that I1 is
a finitely generated ideal of R1. This is a contradiction, as desired.
The following expands on the last two paragraphs of [12].
6.5 Proposition. If R is a commutative, ℵ0-Noetherian, von Neumann regular
ring that is not completely reducible, then the map α′′ : Rhhaii ∐ Rhhbii → Rhha, bii
is not injective.
Proof. By Theorem 4.21(ii), R is not Π-coherent. By the contrapositive of
Corollary 4.10 (c)⇒(a), β2 is not injective. By Proposition 5.3(i), α′′ is not in-
jective.
6.6 Example. If Q is a field and R = Q[ej : j ∈ N]/(ekeℓ − δk,ℓek : k, ℓ ∈ N), then
α′′ is not injective, since R satisfies the hypotheses of Proposition 6.5; the countable
case was first seen in the last two paragraphs of [12]. Let us give an alternative
proof by showing that this R has the form considered in Proposition 6.4 with X1
and X2 countable. Partition N into two infinite subsets X1 and X2. For i = 1, 2,
let mi := {ej : j ∈ Xi}, let Ii denote the ideal of R generated by mi, and let Ri
denote the Q-subring of R generated by mi. Then we may make the identifica-
tion R = (R1 ⊗Q R2)/(I1 ⊗Q I2), and, by Proposition 6.4, m1 ⊗ m2 represents a
non-zero element of Ker(µX1,X2 ). We may identify µX1,X2 with β2.
Here is a result similar to Proposition 6.4, which uses a different, rather curious
argument to prove that an appropriate element of Ker µX,X is not zero.
6.7 Proposition. (Bergman) Let Q be a non-zero commutative ring, X an infinite
set, I the free Q-module on X, R the commutative augmented Q-ring Q ⊕ I with
I 2 = {0}, and m the element (x : x ∈ X) of QX
R. Then m⊗m represents a non-zero
element of Ker(µX,X ).
Proof. It is clear that m ⊗R m is an element of Ker(µX,X ), and it remains to show
that m ⊗R m 6= 0 in (XR) ⊗R (RX ); we shall do this by finding a quotient R-module
in which the image of m ⊗R m is not zero.
Define R-bimodules P := X(R ⊗Q R)X and P := (R/I) ⊗R P ⊗R (R/I), and
make the identification P = P/(IP + P I). We have natural R-bimodule maps
(XR) ⊗Q (RX ) → X(R ⊗Q R)X = P → P .
Since I acts trivially on P on the left and the right, the R-bimodule map
(XR) ⊗Q (RX ) → P is R-centralizing, and, hence, factors through the universal
18
Ring coproducts embedded in power-series rings
R-centralizing quotient (XR) ⊗R (RX ) of the domain. Under the resulting map
(XR) ⊗R (RX ) → P , our element m ⊗R m is mapped to p + IP + P I where
p := (x ⊗Q y : (x, y) ∈ X × X) ∈ P . Thus to show that m ⊗R m 6= 0, it suf-
fices to show that p 6∈ IP + P I.
For every x, y ∈ X and t ∈ R ⊗Q R, we shall write cx⊗y(t) ∈ Q for the coefficient
of x ⊗ y in the expression for t with respect to the Q-basis (X ∪ {1}) ⊗ (X ∪ {1})
of R ⊗Q R.
Consider any q = (qx,y : (x, y) ∈ X × X) ∈ P . For each x0 ∈ X,
it can be
seen that x0P = X(x0Q ⊗Q R)X . Hence, if q ∈ x0P , then, for all (x, y) ∈ X × X,
cx⊗y(qx,y) = 0 if x 6= x0. Similar statements hold for P x0. If q ∈ IP + P I, then
there exist finite subsets X0, Y0 ⊆ X such that q ∈ Px0∈X0
P y0.
Since X is infinite, there exist x ∈ X−X0 and y ∈ X−Y0, and then cx⊗y(qx,y) = 0.
Since Q 6= {0}, cx⊗y(qx,y) 6= 1. Hence, q 6= p. This shows that m ⊗R m 6= 0, as
desired.
x0P + Py0∈Y0
In the previous two propositions, we have created elements of Ker(µX1,X2 ) of
the form m1 ⊗ m2, and this requires the existence of zero-divisors in R. By using
elements of Ker(µX1 ,X2) of the form m1 ⊗ m2 − m3 ⊗ m4, we can avoid having
zero-divisors in R, as follows.
6.8 Proposition. (Bergman) Let Q be a field, and let R be the subring of Q[x, y]
generated by {xiy : i ∈ N}. In particular, R is a commutative domain. In QN
R, let
m1 := m2 := (x2i+1y : i ∈ N), m3 := (x2iy : i ∈ N), m4 := (x2i+2y : i ∈ N).
Then m1 ⊗ m2 − m3 ⊗ m4 lies in Ker(β2) − {0}.
Proof. For i, j ∈ N, both m1 ⊗ m2 and m3 ⊗ m4 have x2i+2j+2y2 as their (i, j)th
component. Thus m1 ⊗ m2 − m3 ⊗ m4 ∈ Ker(β2).
Let I denote the ideal of R generated by {x2iy, x2iy2 : i ∈ N}, and let R′ := R/I.
Then R′ is the commutative augmented Q-ring with augmentation ideal of square
R′, let m := (x2i+1y : i ∈ N). By
zero and Q-basis X := {x2i+1y : i ∈ N}. In QN
R)⊗2 → (QN
Proposition 6.7, in (QN
carries m1⊗m2−m3⊗m4 to (m⊗m)−(0⊗0) 6= 0. Hence, m1⊗m2−m3⊗m4 6= 0.
R′)⊗2, m ⊗ m 6= 0. The natural map (QN
R′)⊗2
Acknowledgments
The first-named author was partially supported by DGI MICIIN MTM2011-28992-C02-01,
and by the Comissionat per Universitats i Recerca de la Generalitat de Catalunya.
The second-named author was partially supported by Spain's Ministerio de Ciencia e Inno-
vaci´on through Project MTM2011-25955.
We are very grateful to George Bergman for generously contributing the examples of Section 6
and suggesting many improvements of various earlier versions.
We are also grateful to Dolors Herbera for providing much useful information for Section 4.
References
[1] Pere Ara and Warren Dicks, Universal localizations embedded in power-series rings, Forum
Math. 19 (2007), 365 -- 378.
[2] Hyman Bass, Finitistic dimension and a homological generalization of semi-primary rings,
Trans. Amer. Math. Soc. 95 (1960), 466 -- 488.
[3] George M. Bergman, Ordering coproducts of groups and semigroups, J. Algebra 133 (1990),
313 -- 339.
[4] Vasily C. Cateforis, On regular self-injective rings, Pacific J. Math. 30 (1969), 39 -- 45.
[5] Stephen U. Chase, Direct products of modules, Trans. Amer. Math. Soc. 97 (1960), 457 -- 473.
[6] P. M. Cohn, On the free product of associative rings, Math. Z. 71 1959, 380 -- 398.
Pere Ara and Warren Dicks
19
[7] P. M. Cohn, Free ideal rings, J. Algebra 1 (1964), 47 -- 69.
[8] P. M. Cohn, Free rings and their relations. Second edition, LMS Monographs 19, Academic
Press, London, 1985. xxii+588 pp.
[9] S. H. Cox and R. L. Pendleton,
Rings for which certain flat modules are projective,
Trans. Amer. Math. Soc. 150 (1970), 139 -- 156.
[10] B. Eckmann and A. Schopf, Uber injektive Moduln, Arch. Math. 4 (1953), 75 -- 78.
[11] Ralph H. Fox, Free differential calculus. I Derivation in the free group ring, Ann. of Math.
(2) 57 (1953), 547 -- 560.
[12] K. R. Goodearl, Distributing tensor product over direct product, Pacific J. Math. 43 (1972),
107 -- 110.
[13] Dolors Herbera and Jan Trlifaj, Almost
free modules and Mittag-Leffler conditions,
Adv. Math. 229 (2012), 3436 -- 3467.
[14] C. U. Jensen, Some cardinality questions for flat modules and coherence, J. Algebra 12
(1969), 231 -- 241.
[15] S. Jøndrup, On finitely generated flat modules III , Math. Scand. 29 (1971), 206 -- 210.
[16] Marsha Finkel Jones, Flatness and f -projectivity of torsion-free modules and injective mod-
ules, pp. 94 -- 116 in: Advances in noncommutative ring theory. Proc. Twelfth George H. Hud-
son Symp., SUNY, Plattsburgh NY, 1981, (ed. Patrick J. Fleury), Lecture Notes in Math.
951, Springer-Verlag, Berlin, 1982. iii+142 pp.
[17] Irving Kaplansky, Projective modules, Ann. of Math. 68 (1958), 372 -- 377.
[18] Irving Kaplansky, On the dimension of modules and algebras, X A right hereditary ring which
is not left hereditary, Nagoya Math. J. 13 (1958), 85 -- 88.
[19] Shigeru Kobayashi, A note on regular self-injective rings, Osaka J. Math. 21 (1984), 679 -- 682.
[20] Helmut Lenzing, Endlich prasentierbare Moduln, Arch. Math. 20 (1969), 262 -- 266.
[21] Wilhelm Magnus, Beziehungen zwischen Gruppen und Idealen in einem speziellen Ring,
Math. Ann. 111 (1935), 259 -- 280.
[22] B. L. Osofsky, Rings all of whose finitely generated modules are injective, Pacific J. Math. 14
(1964), 645 -- 650.
[23] Joseph J. Rotman, An introduction to homological algebra, Pure and applied mathematics 89,
Academic Press, New York, 1979. xi+376 pp.
[24] I. I. Sahaev, The projectivity of finitely generated flat modules, Mat. Z. 6 (1965), 564 -- 573.
Departament de Matem`atiques, Universitat Aut`onoma de Barcelona, E-08193 Bel-
laterra (Barcelona), Spain
E-mail addresses: [email protected], [email protected].
Home page: http://mat.uab.cat/~dicks/
|
1005.2373 | 1 | 1005 | 2010-05-13T16:47:18 | A Hom-associative analogue of n-ary Hom-Nambu algebras | [
"math.RA",
"math-ph",
"math-ph"
] | It is shown that every n-ary totally Hom-associative algebra with equal twisting maps yields an n-ary Hom-Nambu algebra via an n-ary version of the commutator bracket. The class of n-ary totally Hom-associative algebras is shown to be closed under twisting by self-weak morphisms. Every multiplicative n-ary totally Hom-associative algebra yields a sequence of multiplicative totally Hom-associative algebras of exponentially higher arities. Under suitable conditions, an n-ary totally Hom-associative algebra gives an (n-k)-ary totally Hom-associative algebra. | math.RA | math |
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
DONALD YAU
Abstract. It is shown that every n-ary totally Hom-associative algebra with equal twisting maps
yields an n-ary Hom-Nambu algebra via an n-ary version of the commutator bracket. The class
of n-ary totally Hom-associative algebras is shown to be closed under twisting by self-weak mor-
phisms. Every multiplicative n-ary totally Hom-associative algebra yields a sequence of multiplica-
tive totally Hom-associative algebras of exponentially higher arities. Under suitable conditions,
an n-ary totally Hom-associative algebra gives an (n − k)-ary totally Hom-associative algebra.
1. Introduction
Nambu algebras play an important role in physics. For example, Nambu mechanics [16, 20]
involves n-ary Nambu algebras [8, 18]. In these n-ary Nambu algebras, the n-ary compositions are
derivations, a property called the n-ary Nambu identity (a.k.a. Filippov identity), which generalizes
the Jacobi identity. Lie triple systems [10, 11], which are ternary Nambu algebras with two further
properties, can be used to solve the Yang-Baxter equation [17]. An n-ary Nambu algebra whose
product is anti-symmetric is called an n-ary Nambu-Lie algebra (a.k.a. Filippov algebra). Ternary
Nambu-Lie algebras appear in the work of Bagger and Lambert [6] and many others on M-theory.
Other applications of n-ary Nambu algebras in physics are discussed in, e.g., [4, 17, 18].
Generalizations of n-ary Nambu(-Lie) algebras, called n-ary Hom-Nambu(-Lie) algebras, were
introduced by Ataguema, Makhlouf, and Silvestrov in [3] . In an n-ary Hom-Nambu(-Lie) algebra,
the n-ary Nambu identity is replaced by the n-ary Hom-Nambu identity (see Definition 4.4), which
involves n − 1 linear twisting maps. These twisting maps can be thought of as additional degrees
of freedom or as deformation parameters. An n-ary Nambu(-Lie) algebra can be regarded as an
n-ary Hom-Nambu(-Lie) algebra in which the twisting maps are all equal to the identity map. Hom-
Nambu(-Lie) algebras are interesting even when the underlying algebras are Nambu(-Lie) algebras.
In fact, certain ternary Nambu-Lie algebras can be regarded as ternary Hom-Nambu-Lie algebras
with non-identity twisting maps [3].
Binary Hom-Nambu-Lie algebras are called Hom-Lie algebras, which originated in [9] in the study
of q-deformations of the Witt and the Virasoro algebras. The binary Hom-Nambu identity in this
case is called the Hom-Jacobi identity [14]. The associative counterparts of Hom-Lie algebras are
Hom-associative algebra [14]. Hom-Lie algebras are to Hom-associative algebras as Lie algebras are
to associative algebras [14, 21].
Let us recall some properties of n-ary Hom-Nambu(-Lie) algebras.
It is shown in [3] that n-
ary Nambu(-Lie) algebras can be twisted along self-morphisms to yield n-ary Hom-Nambu(-Lie)
algebras. This twisting construction of n-ary Hom-Nambu(-Lie) algebras is a generalization of a
result about G-Hom-associative algebras due to the author [22].
It is shown in [1] that ternary
Date: November 3, 2018.
2000 Mathematics Subject Classification. 17A40, 17A42, 17B81.
Key words and phrases. Hom-Nambu algebra, totally Hom-associative algebra, n-commutator bracket.
1
2
DONALD YAU
Virasoro-Witt algebras can be q-deformed into ternary Hom-Nambu-Lie algebras. It is shown in [2]
that a ternary Hom-Nambu-Lie algebra can be obtained from a Hom-Lie algebra together with a
compatible linear map and a trace function.
Further properties of n-ary Hom-Nambu(-Lie) algebras were established by the author in [24, 25].
A unique feature of Hom-type algebras is that they are closed under twisting by suitably defined self-
morphisms. In particular, it is shown in [24] that the category of n-ary Hom-Nambu(-Lie) algebras
is closed under twisting by self-weak morphisms. Starting with an n-ary Nambu(-Lie) algebra, this
closure property reduces to the twisting construction for n-ary Hom-Nambu(-Lie) algebras in [3].
Moreover, it is proved in [24] that every multiplicative n-ary Hom-Nambu algebra yields a sequence
of Hom-Nambu algebras of exponentially higher arities. It is also shown in [24] that, under suitable
conditions, an n-ary Hom-Nambu(-Lie) algebra reduces to an (n− k)-ary Hom-Nambu(-Lie) algebra.
Hom-Jordan and Hom-Lie triple systems were defined in [24] as Hom-type generalizations of
Jordan and Lie triple systems [10, 11, 15]. A Hom-Lie triple system is automatically a ternary Hom-
Nambu algebra, but it is usually not a ternary Hom-Nambu-Lie algebra because its ternary product
is not assumed to be anti-symmetric. It is proved in [24] that Hom-Lie triple systems, and hence
ternary Hom-Nambu algebras, can be obtained from Hom-Jordan triple systems, ternary totally
Hom-associative algebras [3], multiplicative Hom-Lie algebras, and Hom-associative algebras.
Furthermore, it is proved in [25] that multiplicative Hom-Jordan algebras [23] have underlying
Hom-Jordan triple systems, and hence also ternary Hom-Nambu algebras. Combined with results
from [24], this implies that every multiplicative Hom-Jordan algebra gives rise to a sequence of
Hom-Nambu algebras of arities 2k+1 + 1. As in the classical case, a major source of Hom-Jordan
algebras is the class of Hom-alternative algebras, which were defined in [13]. It is proved in [23] that
multiplicative Hom-alternative algebras are Hom-Jordan admissible. Therefore, every multiplicative
Hom-alternative algebra also gives rise to a sequence of Hom-Nambu algebras of arities 2k+1 + 1.
Finally, the class of n-ary Hom-Nambu-Lie algebras is extended to the class of n-ary Hom-Maltsev
algebras in [25].
The main purpose of this paper is to study a Hom-associative analogue of n-ary Hom-Nambu
algebras. The basic motivation is that the n-ary Hom-Nambu identity is an n-ary version of the
Hom-Jacobi identity. As is well-known, the commutator bracket of an associative algebra satisfies
the Jacobi identity. Likewise, the commutator bracket of a Hom-associative algebra satisfies the
Hom-Jacobi identity [14]. Therefore, it is natural to ask the following question.
Is there an n-ary version of a Hom-associative algebra that gives rise to an n-ary
Hom-Nambu algebra via an n-ary version of the commutator bracket?
One main result of this paper is an affirmative answer to this question when the twisting maps
are equal (Theorem 4.5). The relevant Hom-associative type objects are the n-ary totally Hom-
associative algebras defined in [3], which generalize n-ary totally associative algebras. The relevant
commutator bracket is what we call the n-commutator bracket (Definition 4.3), which involves 2n−1
terms. Restricting to the special case where all the twisting maps are equal to the identity map,
this result implies that every n-ary totally associative algebra yields an n-ary Nambu algebra via
the n-commutator bracket.
Low-dimensional cases of the n-commutator bracket have been used elsewhere. In particular, the
2-commutator bracket is the usual commutator. The 3-commutator bracket was used by the author
in [24] to show that a ternary totally Hom-associative algebra with equal twisting maps yields a
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
3
ternary Hom-Nambu algebra. We should point out that our n-commutator bracket is not the same
as the totally anti-symmetrized n-ary commutator in [5, 7]. Moreover, the n-commutator bracket
is not anti-symmetric when n ≥ 3. Therefore, the n-ary Hom-Nambu algebras arising from n-ary
totally Hom-associative algebras via the n-commutator bracket are usually not n-ary Hom-Nambu-
Lie algebras.
A description of the rest of this paper follows.
In section 2 we observe that the class of n-ary totally Hom-associative algebras is closed under
twisting by self-weak morphisms (Theorem 2.5). The corresponding closure property for n-ary Hom-
Nambu(-Lie) and n-ary Hom-Maltsev algebras can be found in [24] and [25], respectively. A special
case of Theorem 2.5 says that each multiplicative n-ary totally Hom-associative algebra gives rise
to a sequence of multiplicative n-ary totally Hom-associative algebras by twisting along its own
twisting map (Corollary 2.7). We obtain Theorem 3.6 in [3] as another special case of Theorem
2.5.
It says that n-ary totally associative algebras can be twisted along self-morphisms to yield
multiplicative n-ary totally Hom-associative algebras (Corollary 2.8). Section 2 ends with several
examples of n-ary totally Hom-associative algebras.
In section 3 we study how totally Hom-associative algebras of different arities are related. First, we
show that every multiplicative n-ary totally Hom-associative algebra yields a sequence of multiplica-
tive totally Hom-associative algebras of exponentially higher arities, namely, 2k(n − 1) + 1 for k ≥ 0
(Corollary 3.2). The transition from Hom-associative algebras to ternary totally Hom-associative
algebras was proved in [24]. One major difference between the n = 2 case in [24] and the n ≥ 3 case
here is that the latter requires multiplicativity while the former does not. Second, we show that
under suitable conditions an n-ary totally Hom-associative algebra reduces to an (n − k)-ary totally
Hom-associative algebra (Corollary 3.5). The corresponding results for Hom-Nambu algebras can
be found in [24].
In section 4 we define the n-commutator and show that every n-ary totally Hom-associative alge-
bra with equal twisting maps yields an n-ary Hom-Nambu algebra via the n-commutator (Theorem
4.5). The special case of this result when n = 2 is Proposition 1.6 in [14]. The special case when
n = 3 is Corollary 4.3 in [24]. In these low dimensional cases, due to the relatively small number of
terms involved, the binary and ternary Hom-Nambu identities can be shown by a direct computation
with all the terms written out. In the general case, a more systematic argument is needed because
of the large number of terms in the Hom-Nambu identity. In fact, the n-ary Hom-Nambu identity
for the n-commutator bracket involves 22n−2(n + 1) terms.
2. Totally Hom-associative algebras
In this section we observe that the class of n-ary totally Hom-associative algebras is closed under
twisting by self-weak morphisms. Some examples of n-ary totally Hom-associative algebras are then
given.
2.1. Conventions. Throughout this paper we work over a fixed field k of characteristic 0. If V is
a k-module and f : V → V is a linear map, then f n denotes the composition of n copies of f with
f 0 = Id. For i ≤ j, elements xi, . . . , xj ∈ V and maps f, fk, . . . , fl : V → V with j − i = l − k, we
4
DONALD YAU
adopt the abbreviations
xi,j = (xi, xi+1, . . . , xj ),
f (xi,j) = (f (xi), f (xi+1), . . . , f (xj)),
(2.1.1)
fk,l(xi,j) = (fk(xi), fk+1(xi+1), . . . , fl(xj)).
For i > j, the symbols xi,j, f (xi,j), and fk,l(xi,j ) denote the empty sequence. For a bilinear map
µ : V ⊗2 → V , we often write µ(x, y) as the juxtaposition xy.
Let us begin with the following basic definitions.
Definition 2.2. Let n ≥ 2 be an integer.
(1) An n-ary Hom-algebra (V, (, ..., ), α) with α = (α1, . . . , αn−1) [3] consists of a k-module
V , an n-linear map (, ..., ) : V ⊗n → V , and linear maps αi : V → V for i = 1, . . . , n − 1,
called the twisting maps.
(2) An n-ary Hom-algebra (V, (, ..., ), α) is said to be multiplicative if (i) the twisting maps
are all equal, i.e., α1 = · · · = αn−1 = α, and (ii) α ◦ (, ..., ) = (, ..., ) ◦ α⊗n.
(3) A weak morphism f : V → U of n-ary Hom-algebras is a linear map of the underlying
k-modules such that f ◦ (, ..., )V = (, ..., )U ◦ f ⊗n. A morphism of n-ary Hom-algebras is a
weak morphism such that f ◦ (αi)V = (αi)U ◦ f for i = 1, . . . n − 1 [3].
For an n-ary Hom-algebra V and elements x1, . . . , xn ∈ V , using the abbreviations in (2.1.1), the
n-ary product (x1, . . . , xn) will often be denoted by (x1,n) below. We sometimes omit the commas
in the n-ary product (, ..., ).
An n-ary Hom-algebra V in which all the twisting maps are equal, as in the multiplicative case,
will be denoted by (V, (, ..., ), α), where α is the common value of the twisting maps. An n-ary
algebra in the usual sense is a k-module V with an n-linear map (, ..., ) : V ⊗n → V . We consider
an n-ary algebra (V, (, ..., )) also as an n-ary Hom-algebra (V, (, ..., ), Id) in which all n − 1 twisting
maps are the identity map. Also, in this case a weak morphism is the same thing as a morphism,
which agrees with the usual definition of a morphism of n-ary algebras.
Let us now recall the definition of an n-ary totally Hom-associative algebra from [3].
Definition 2.3. Let (A, (, ..., ), α) be an n-ary Hom-algebra.
(1) For i ∈ {1, . . . , n − 1} define the ith Hom-associator asi
A : A⊗2n−1 → A to be the (2n − 1)-
linear map
asi
A(a1,2n−1) = (α1,i−1(a1,i−1), (ai,i+n−1), αi,n−1(ai+n,2n−1))
− (α1,i(a1,i), (ai+1,i+n), αi+1,n−1(ai+n+1,2n−1))
for a1, . . . , a2n−1 ∈ A.
(2) An n-ary totally Hom-associative algebra is an n-ary Hom-algebra A that satisfies
total Hom-associativity
for all i ∈ {1, . . . , n − 1}.
asi
A = 0
An n-ary totally Hom-associative algebra with αi = Id for all i is called an n-ary totally
associative algebra. In this case, asi
A = 0 is referred to as total associativity.
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
5
When n = 2 total Hom-associativity means
((a1a2)α(a2)) = (α(a1)(a2a3)),
which is the defining identity for Hom-associative algebras [14]. When n = 3 total Hom-associativity
means
((a1a2a3), α1(a4), α2(a5)) = (α1(a1), (a2a3a4), α2(a5)) = (α1(a1), α2(a2), (a3a4a5)).
(2.3.1)
In particular, if both twisting maps α1 and α2 are equal to the identity map, then (2.3.1) is the
defining identity for ternary rings [12].
To see that the category of n-ary totally Hom-associative algebras is closed under twisting by
self-weak morphisms, we need the following observations.
Lemma 2.4. Let (A, (, ..., ), α) be an n-ary Hom-algebra and β : A → A be a weak morphism.
Consider the n-ary Hom-algebra
Aβ = (A, (, ..., )β = β(, ..., ), βα = (βα1, . . . , βαn−1)).
(2.4.1)
Then the following statements hold.
(1) β2asi
(2) If A is multiplicative and βα = αβ, then Aβ is also multiplicative.
for i ∈ {1, . . . , n − 1}.
A = asi
Aβ
Proof. Both assertions are immediate from the definitions.
(cid:3)
The desired closure property is now an immediate consequence of Lemma 2.4.
Theorem 2.5. Let (A, (, ..., ), α) be an n-ary totally Hom-associative algebra and β : A → A be a
weak morphism. Then Aβ in (2.4.1) is also an n-ary totally Hom-associative algebra. Moreover, if
A is multiplicative and βα = αβ, then Aβ is also multiplicative.
Let us now discuss some special cases of Theorem 2.5. If A is a multiplicative n-ary Hom-algebra,
then its twisting map is a morphism (and hence a weak morphism) on A. Therefore, we have the
following special case of Theorem 2.5.
Corollary 2.6. Let (A, (, ..., ), α) be a multiplicative n-ary totally Hom-associative algebra. Then
is also a multiplicative n-ary totally Hom-associative algebra.
Aα = (A, (, ..., )α = α(, ..., ), α2)
Iterating Corollary 2.6 we obtain the following result, which says that every multiplicative n-ary
totally Hom-associative algebra gives rise to a sequence of derived n-ary totally Hom-associative
algebras.
Corollary 2.7. Let (A, (, ..., ), α) be a multiplicative n-ary totally Hom-associative algebra. Then
is also a multiplicative n-ary totally Hom-associative algebra for each k ≥ 0.
Ak = (A, (, ..., )k = α2k −1(, ..., ), α2k
)
On the other hand, if we set αi = Id for all i in Theorem 2.5, then we obtain the following
twisting result, which is Theorem 3.6 in [3]. It says that n-ary totally Hom-associative algebras can
be obtained from n-ary totally associative algebras and their morphisms.
6
DONALD YAU
Corollary 2.8. Let (A, (, ..., )) be an n-ary totally associative algebra and β : A → A be a morphism.
Then
is a multiplicative n-ary totally Hom-associative algebra.
Aβ = (A, (, ..., )β = β(, ..., ), β)
The rest of this section contains examples of n-ary totally Hom-associative algebras.
Example 2.9. Let A be an associative algebra, f : A → A be an algebra morphism, and ζ ∈ k be
a primitive nth root of unity. Then the ζ-eigenspace of f ,
is an (n + 1)-ary totally associative algebra under the (n + 1)-ary product
A(f, ζ) = {a ∈ A : f (a) = ζa},
(a1, . . . , an+1) = a1 · · · an+1.
(2.9.1)
(2.9.2)
Let α : A → A be an algebra morphism such that αf = f α. Then α restricts to a morphism
of (n + 1)-ary totally associative algebras on A(f, ζ). By Corollary 2.8 there is a multiplicative
(n + 1)-ary totally Hom-associative algebra
where
for all ai ∈ A.
A(f, ζ)α = (A(f, ζ), (, ..., )α, α),
(a1, . . . , an+1)α = α(a1 · · · an+1)
(cid:3)
Example 2.10. Let A be the associative algebra over k consisting of polynomials in r ≥ 2 associative
variables X1, . . . , Xr with 0 constant term. Fix an integer n ≥ 2. Let An denote the submodule
of A spanned by the homogeneous polynomials of degrees 1 (mod n). Note that An is actually an
eigenspace A(f, ζ) as in (2.9.1), where ζ ∈ k is a primitive nth root of unity and f : A → A is
determined by
f (Xi) = ζXi
for all i. In any case, An is an (n + 1)-ary totally associative algebra with the (n + 1)-ary product
in (2.9.2). When n = 2 the ternary totally associative algebra A3 is an example in [12] (p.47).
For each i ∈ {1, . . . , r} let mi ≥ 1 be an integer with mi ≡ 1 (mod n). Then the map α : A → A
determined by
α(Xi) = X mi
i
for all i is an algebra morphism that commutes with f . As in Example 2.9 there is a multiplicative
(n + 1)-ary totally Hom-associative algebra (An)α = (An, (, ..., )α, α).
Note that (An, (, ..., )α) is not totally associative because
((X1 · · · X1
)α X2 · · · X2
)α = X
1(n+1)
m2
1
X m2n
2
,
{z
n
n+1
{z
}
}
}
n
{z
{z
n
}
whereas
They are not equal, provided m1 > 1 or m2 > 1.
(X1 · · · X1
(X1 X2 · · · X2
)α)α = X m1(m1+n)
1
X
2n
m2
2
.
(cid:3)
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
7
Example 2.11. Fix an integer n ≥ 2, and let V1, . . . , Vn be k-modules. Consider the direct sum
A =
Hom(Vi, Vi+1),
n
Mi=1
where Vn+1 ≡ V1. A typical element in A is written as ⊕fi, where fi ∈ Hom(Vi, Vi+1) for i ∈
{1, . . . , n}. One can visualize the element ⊕fi ∈ A as the braid diagram
· · ·
V1
2
2
2
2
2
2
2
2
2
2
2
2
2
V2
tiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiii
V3
2
2
2
2
2
2
2
2
2
2
2
2
2
V4
V2
2
2
2
2
2
2
2
2
2
2
2
2
2
V3
Vn−1
4
4
4
4
4
4
4
4
4
4
4
4
4
4
Vn
· · ·
Vn
V1
on n strands. Define an (n + 1)-ary product on A by
i , . . . , ⊕f n+1
i
where
Fi = f n+1
· · · f n−i+2
1
f n−i+1
n
· · · f 2
i+1f 1
i
(cid:0)⊕f 1
i
i
{z
(cid:1) = ⊕Fi,
{z
n−i+1
}
}
for each i. Pictorially, the (n + 1)-ary product ⊕Fi is represented by the vertical composition of
n + 1 braid diagrams. With this (n + 1)-ary product, A becomes an (n + 1)-ary totally associative
algebra. When n = 2 the ternary totally associative algebra A is an example in [12] (p.46).
For each i ∈ {1, . . . , n}, let γi : Vi → Vi be a linear automorphism. Define the map α : A → A by
α(⊕fi) = ⊕γ−1
i+1fiγi,
where γn+1 ≡ γ1. Then α is an automorphism of (n + 1)-ary totally associative algebras. By
Corollary 2.8 there is a multiplicative n-ary totally Hom-associative algebra
Aα = (A, (, ..., )α = α(, ..., ), α).
Moreover, (A, (, ..., )α) is in general not an n-ary totally associative algebra. For example, suppose
i ∈ Hom(Vi, Vi+1) for i ∈ {1, . . . , n} and f n+1
f i
∈ Hom(V1, V2). Then
1
((f 1
1 , . . . , f n
n , f n+1
1
)α, f 2
2 , . . . , f n+1
1
)α = γ−1
2 f n+1
1
· · · f 2
2 γ−1
2 F1γ2
1 ,
whereas
n , (f n+1
They are not equal in general.
1 , . . . , f n
(f 1
1
, f 2
2 , . . . , f n+1
1
)α)α = γ−2
2 f n+1
1
· · · f 2
2 f n+1
1
γ1f n
n · · · f 1
1 γ1.
(cid:3)
3. Totally Hom-associative algebras of different arities
There are two main results in this section. The first main result (Theorem 3.1) says that every
multiplicative n-ary totally Hom-associative algebra yields a multiplicative (2n−1)-ary totally Hom-
associative algebra. The second main result (Theorem 3.4) says that under suitable conditions an
n-ary totally Hom-associative algebra reduces to an (n − 1)-ary totally Hom-associative algebra.
Both of these results can be iterated.
Here is the first main result of this section. The Hom-Nambu analogue is discussed in [24]. Recall
the abbreviations in (2.1.1).
t
8
DONALD YAU
Theorem 3.1. Let (A, (, ..., ), α) be a multiplicative n-ary totally Hom-associative algebra. Then
A1 = (A, (, ..., )(1), α2)
is a multiplicative (2n − 1)-ary totally Hom-associative algebra, where
(a1,2n−1)(1) = ((a1,n), α(an+1,2n−1))
for all ai ∈ A.
Proof. It is clear that A1 is a multiplicative (2n − 1)-ary Hom-algebra. We must show that its jth
Hom-associator asj
A1 (Definition 2.3) is equal to 0, that is,
(cid:16)α2(a1,j−1), (aj,j+2n−2)(1), α2(aj+2n−1,4n−3)(cid:17)(1)
=(cid:16)α2(a1,j), (aj+1,j+2n−1)(1), α2(aj+2n,4n−3)(cid:17)(1)
(3.1.1)
for all j ∈ {1, . . . , 2n − 2}. The condition (3.1.1) is divided into three cases: (1) j ≤ n − 1, (2) j = n,
and (3) j ≥ n + 1. Note that the case j ≥ n + 1 does not occur if n = 2. The three cases are proved
similarly, so we only provide the details for case (3), which is the only case where multiplicativity is
used. With j ≥ n + 1, using the multiplicativity and total Hom-associativity of A, we compute the
left-hand side of (3.1.1) as follows:
(cid:16)α2(a1,j−1), (aj,j+2n−2)(1), α2(aj+2n−1,4n−3)(cid:17)(1)
=(cid:0)(α2(a1,n)), α3(an+1,j−1), α ((aj,j+n−1), α(aj+n,j+2n−2)) , α3(aj+2n−1,4n−3)(cid:1)
=(cid:0)(α2(a1,n)), α3(an+1,j−1), α (α(aj), (aj+1,j+n), α(aj+n+1,j+2n−2)) , α3(aj+2n−1,4n−3)(cid:1)
=(cid:0)α2((a1,n)), α3(an+1,j−1),(cid:0)α2(aj), α(aj+1,j+n), α2(aj+n+1,j+2n−2)(cid:1) , α3(aj+2n−1,4n−3)(cid:1)
=(cid:0)α2((a1,n)), α3(an+1,j),(cid:0)α(aj+1,j+n), α2(aj+n+1,j+2n−1)(cid:1) , α3(aj+2n,4n−3)(cid:1)
=(cid:0)(α2(a1,n)), α3(an+1,j), α((aj+1,j+n), α(aj+n+1,j+2n−1)), α3(aj+2n,4n−3)(cid:1)
The last expression above is equal to the right-hand side of (3.1.1), as desired.
(cid:3)
Applying Theorem 3.1 repeatedly, we obtain the following result. It says that every multiplica-
tive n-ary totally Hom-associative algebra gives rise to a sequence of multiplicative totally Hom-
associative algebras of exponentially higher arities.
Corollary 3.2. Let (A, (, ..., ), α) be a multiplicative n-ary totally Hom-associative algebra. Define
the (2k(n − 1) + 1)-ary product (, ..., )(k) inductively by setting (, ..., )(0) = (, ..., ) and
(cid:0)a1,2k(n−1)+1(cid:1)(k)
for k ≥ 1. Then
=(cid:16)(cid:0)a1,2k−1(n−1)+1(cid:1)(k−1)
, α2k−1(cid:0)a2k−1(n−1)+2,2k(n−1)+1(cid:1)(cid:17)(k−1)
Ak = (A, (, ..., )(k), α2k
)
is a multiplicative (2k(n − 1) + 1)-ary totally Hom-associative algebra for each k ≥ 0.
For example, when k = 2 we have:
(a1,4n−3)(2) =(cid:16)(a1,2n−1)(1) , α2(a2n,4n−3)(cid:17)(1)
=(cid:0)(cid:0)((a1,n), α(an+1,2n−1)) , α2(a2n,3n−2)(cid:1) , α3(a3n−1,4n−3)(cid:1) .
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
9
When k = 3, writing x = (a1,4n−3)(2), we have:
(a1,8n−7)(3) =(cid:0)x, α4(a4n−2,8n−7)(cid:1)(2)
=(cid:0)(cid:0)(cid:0)(x, α4(a4n−2,5n−4)), α5(a5n−3,6n−5)(cid:1) , α6(a6n−4,7n−6)(cid:1) , α7(a7n−5,8n−7)(cid:1) .
In general, (, ..., )(k) involves an iterated composition of 2k copies of the n-ary product (, ..., ) and
n − 1 copies of αi for each i ∈ {1, . . . , 2k − 1}.
Restricting to the case α = Id in Corollary 3.2, we obtain the following construction result for
higher arity totally associative algebras.
Corollary 3.3. Let (A, (, ..., )) be an n-ary totally associative algebra. Define the (2k(n − 1) + 1)-ary
product (, ..., )(k) inductively by setting (, ..., )(0) = (, ..., ) and
(cid:0)a1,2k(n−1)+1(cid:1)(k)
for k ≥ 1. Then
=(cid:16)(cid:0)a1,2k−1(n−1)+1(cid:1)(k−1)
, a2k−1(n−1)+2,2k(n−1)+1(cid:17)(k−1)
Ak = (A, (, ..., )(k))
is a multiplicative (2k(n − 1) + 1)-ary totally associative algebra for each k ≥ 0.
The following result is the second main result of this section. It gives sufficient conditions under
which an n-ary totally Hom-associative algebra reduces to an (n − 1)-ary totally Hom-associative
algebra. This result is the totally Hom-associative analogue of results due to Pozhidaev [19] and
Filippov [8] about n-ary Maltsev algebras and Nambu-Lie algebras, respectively. The n-ary Hom-
Nambu(-Lie) and Hom-Maltsev analogues can be found in [24] and [25], respectively.
Theorem 3.4. Let (A, (, ..., ), α) be an n-ary totally Hom-associative algebra with n ≥ 3. Suppose
a ∈ A satisfies
(1) αn−1(a) = a, and
(2) (x1,n−1, a) = (x1,n−2, a, xn−1) for all xi ∈ A.
Then
is an (n − 1)-ary totally Hom-associative algebra, where
A1 = (A, (, ..., )′, α′)
(x1,n−1)′ = (x1,n−1, a) and α′ = (α1, . . . , αn−2)
for all xi ∈ A. Moreover, if A is multiplicative, then so is A1.
Proof. The multiplicativity assertion is clear. We must show that for j ∈ {1, . . . , n − 2}, the jth
Hom-associator asj
A1 is equal to 0, that is,
(α1,j−1(x1,j−1), (xj,j+n−2)′, αj,n−2(xj+n−1,2n−3))
′
= (α1,j(x1,j ), (xj+1,j+n−1)′, αj+1,n−2(xj+n,2n−3))
′
.
(3.4.1)
Using the assumptions on a and the total Hom-associativity of A, we compute the left-hand side of
(3.4.1) as follows:
(α1,j−1(x1,j−1), (xj,j+n−2)′, αj,n−2(xj+n−1,2n−3))
′
= (α1,j−1(x1,j−1), (xj,j+n−2, a), αj,n−2(xj+n−1,2n−3), αn−1(a))
= (α1,j(x1,j), (xj+1,j+n−2, a, xj+n−1), αj+1,n−2(xj+n,2n−3), αn−1(a))
= (α1,j(x1,j), (xj+1,j+n−1, a), αj+1,n−2(xj+n,2n−3), a) .
10
DONALD YAU
The last expression above is equal to the right-hand side of (3.4.1), as desired.
(cid:3)
Note that in Theorem 3.4 the first assumption about a is automatically satisfied if αn−1 is the
identity map on A. On the other hand, the second assumption on a is automatically satisfied if the
n-ary product (, ..., ) is commutative in the last two variables.
Applying Theorem 3.4 repeatedly, we obtain the following result. It gives sufficient conditions un-
der which an n-ary totally Hom-associative algebra reduces to an (n−k)-ary totally Hom-associative
algebra.
Corollary 3.5. Let (A, (, ..., ), α) be an n-ary totally Hom-associative algebra with n ≥ 3. Suppose
there exist a1, . . . , ak ∈ A for some k ≤ n − 2 such that
(1) αn−i(ai) = ai for all i ∈ {1, . . . , k}, and
(2) (x1,n−i, ai, ai−1, . . . , a1) = (x1,n−i−1, ai, xn−i, ai−1, . . . , a1) for all i ∈ {1, . . . , k} and xj ∈ A.
Then
is an (n − k)-ary totally Hom-associative algebra, where
Ak = (A, (, ..., )k, (α1, . . . , αn−1−k))
(x1,n−k)k = (x1,n−k, ak, ak−1, . . . , a1)
for all xj ∈ A. Moreover, if A is multiplicative, then so is Ak.
4. From totally Hom-associative algebras to Hom-Nambu algebras
The main result of this section says that an n-ary totally Hom-associative algebra with equal
twisting maps yields an n-ary Hom-Nambu algebra via the n-commutator bracket.
Let us first define the n-commutator words, which generalize the two terms in the usual commu-
tator bracket.
Definition 4.1. Let X1, X2, . . . be non-commuting variables. For n ≥ 2 define the set Wn of
n-commutator words inductively as
W2 = {X1X2, −X2X1}
Wn = {zXn, −Xnz : z ∈ Wn−1}
and
for n > 2.
For example,
and W4 consists of the 4-commutator words
W3 = {X1X2X3, −X2X1X3, −X3X1X2, X3X2X1}
X1X2X3X4, −X2X1X3X4, −X3X1X2X4, X3X2X1X4,
− X4X1X2X3, X4X2X1X3, X4X3X1X2, −X4X3X2X1.
In general, Wn consists of 2n−1 n-commutator words. Every n-commutator word gives a self-map
on the n-fold tensor product as follows.
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
11
Definition 4.2. Let V be a k-module and w = ±Xi1 · · · Xin ∈ Wn. Define the map w : V ⊗n → V ⊗n
by
w(v1, . . . , vn) = ±vi1 ⊗ · · · ⊗ vin
for all vj ∈ V .
Using these maps defined by the n-commutator words, we can now define the n-commutator
bracket.
Definition 4.3. Let (A, (, ..., ), α) be an n-ary Hom-algebra. Define the n-commutator bracket
[, ... , ] : A⊗n → A⊗n by
(w(a1, . . . , an))
(4.3.1)
[a1, . . . , an] = Xw∈Wn
for all aj ∈ A.
For example, the 2-commutator bracket is the usual commutator bracket:
[a1, a2] = (a1a2) − (a2a1).
The 3-commutator bracket is the sum
[a1,3] = (a1a2a3) − (a2a1a3) − (a3a1a2) + (a3a2a1),
which was first used in Corollary 4.3 in [24].
Let us now recall the definition of an n-ary Hom-Nambu algebra from [3].
Definition 4.4. Let (V, [, ... , ], α) be an n-ary Hom-algebra.
(1) The n-ary Hom-Jacobian of V is the (2n − 1)-linear map J n
V : V ⊗2n−1 → V defined as
(using the shorthand in (2.1.1))
J n
V (x1,n−1; y1,n) = [α1,n−1(x1,n−1), [y1,n]]
−
[α1,i−1(y1,i−1), [x1,n−1, yi], αi,n−1(yi+1,n)]
(4.4.1)
n
Xi=1
for x1, . . . , xn−1, y1, . . . , yn ∈ V .
(2) An n-ary Hom-Nambu algebra is an n-ary Hom-algebra V that satisfies
called the n-ary Hom-Nambu identity.
J n
V = 0,
When the twisting maps are all equal to the identity map, n-ary Hom-Nambu algebras are the
V is called the
V = 0 is called the n-ary Nambu identity.
usual n-ary Nambu algebras [4, 8, 16, 18]. In this case, the n-ary Hom-Jacobian J n
n-ary Jacobian, and the n-ary Hom-Nambu identity J n
We are now ready for the main result of this section.
Theorem 4.5. Let (A, (, ..., ), α) be an n-ary totally Hom-associative algebra with equal twisting
maps. Then
N (A) = (A, [, ... , ], α)
is an n-ary Hom-Nambu algebra, where [, ... , ] is the n-commutator bracket in (4.3.1). Moreover, if
A is multiplicative, then so is N (A).
12
DONALD YAU
Since the n-commutator bracket does not involve the twisting maps, we have the following special
case of Theorem 4.5
Corollary 4.6. Let (A, (, ..., )) be an n-ary totally associative algebra with equal twisting maps.
Then
is an n-ary Nambu algebra, where [, ... , ] is the n-commutator bracket.
N (A) = (A, [, ... , ])
For the proof of Theorem 4.5, we need to prove the n-ary Hom-Nambu identity J n
N (A) = 0. In
the following Lemmas, we first compute the various terms in the n-ary Hom-Jacobian J n
N (A) (4.4.1).
The n terms in the sum in (4.4.1) are considered in two cases, i = n (Lemma 4.8) and 1 ≤ i ≤ n − 1
(Lemma 4.10).
Let us compute the first term in the n-ary Hom-Jacobian J n
N (A).
Lemma 4.7. With the hypotheses of Theorem 4.5, we have
[α(x1,n−1), [y1,n]] = Xz,z ′∈Wn−1
− Xz,z ′∈Wn−1
− Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
(z′(α(x1,n−1)), (z(y1,n−1), yn))
(z′(α(x1,n−1)), (yn, z(y1,n−1)))
((z(y1,n−1), yn), z′(α(x1,n−1)))
((yn, z(y1,n−1)), z′(α(x1,n−1)))
(4.7.1)
for all xj , yl ∈ A.
Proof. From the definitions of the n-commutator bracket (4.3.1) and the n-commutator words (Def-
inition 4.1), we have
(w(a1,n))
[a1,n] = Xw∈Wn
= Xz∈Wn−1
(z(a1,n−1), an) − (an, z(a1,n−1)).
(4.7.2)
for all aj ∈ A. The Lemma is obtained by using (4.7.2) on both [y1,n] and [α(x1,n−1), ·].
(cid:3)
In Lemma 4.7 we did not use the total Hom-associativity of A. In the next result, we compute
the i = n term in the sum in the n-ary Hom-Jacobian J n
N (A).
Lemma 4.8. With the hypotheses of Theorem 4.5, we have
−[α(y1,n−1), [x1,n−1, yn]] = − Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
− Xz,z ′∈Wn−1
(z(α(y1,n−1)), (z′(x1,n−1), yn))
((z(y1,n−1), yn), z′(α(x1,n−1)))
(z′(α(x1,n−1)), (yn, z(y1,n−1)))
(4.8.1)
((yn, z′(x1,n−1)), z(α(y1,n−1)))
for all xj , yl ∈ A.
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
13
Proof. Using (4.7.2) we have
−[α(y1,n−1), [x1,n−1, yn]] = − Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
− Xz,z ′∈Wn−1
(z(α(y1,n−1)), (z′(x1,n−1), yn))
(z(α(y1,n−1)), (yn, z′(x1,n−1)))
((z′(x1,n−1), yn), z(α(y1,n−1)))
(4.8.2)
((yn, z′(x1,n−1)), z(α(y1,n−1))) .
The first (resp., fourth) sums in (4.8.1) and (4.8.2) are equal. The second (resp., third) sums in
(4.8.1) and (4.8.2) are equal by the total Hom-associativity of A.
(cid:3)
Combining Lemmas 4.7 and 4.8, we obtain the following result.
Lemma 4.9. With the hypotheses of Theorem 4.5, we have
[α(x1,n−1), [y1,n]] − [α(y1,n−1), [x1,n−1, yn]]
= Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
− Xz,z ′∈Wn−1
− Xz,z ′∈Wn−1
(z′(α(x1,n−1)), (z(y1,n−1), yn))
((yn, z(y1,n−1)), z′(α(x1,n−1)))
(z(α(y1,n−1)), (z′(x1,n−1), yn))
((yn, z′(x1,n−1)), z(α(y1,n−1)))
A1
{z
Bn1
A4
{z
{z
{z
Bn4
}
}
}
}
for all xj , yl ∈ A.
Proof. This follows from Lemmas 4.7 and 4.8 because the second (resp., third) sum in (4.7.1) is
equal to the third (resp., second) sum in (4.8.1) with the opposite sign.
(cid:3)
The labels A1, A4, Bn1, and Bn4 will be used below.
Next we compute the ith term (1 ≤ i ≤ n − 1) in the sum in the n-ary Hom-Jacobian J n
N (A).
Lemma 4.10. With the hypotheses of Theorem 4.5, for i ∈ {1, . . . , n − 1} we have
−[α(y1,i−1), [x1,n−1, yi], α(yi+1,n)]
= − Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
+ Xz,z ′∈Wn−1
− Xz,z ′∈Wn−1
(z(α(y1,i−1), (z′(x1,n−1), yi), α(yi+1,n−1)), α(yn))
(α(yn), z(α(y1,i−1), (z′(x1,n−1), yi), α(yi+1,n−1)))
(z(α(y1,i−1), (yi, z′(x1,n−1)), α(yi+1,n−1)), α(yn))
(α(yn), z(α(y1,i−1), (yi, z′(x1,n−1)), α(yi+1,n−1)))
Bi1
{z
{z
{z
{z
Bi2
Bi3
Bi4
}
}
}
}
14
DONALD YAU
for all xj , yl ∈ A.
Proof. Just use (4.7.2) twice, as in Lemma 4.7.
(cid:3)
Using Lemmas 4.9 and 4.10 we can now give the proof of Theorem 4.5.
Proof of Theorem 4.5. Since the multiplicativity assertion is clear, it remains to establish the n-ary
Hom-Nambu identity J n
N (A) = 0. By Lemmas 4.9 and 4.10 the n-ary Hom-Jacobian J n
N (A) is
J n
N (A) = Xz,z ′∈Wn−1 A1 + A4 − Bn1 − Bn4 +
(−Bi1 + Bi2 + Bi3 − Bi4)! .
n−1
Xi=1
(4.10.1)
The total Hom-associativity of A implies the following six types of cancellation. For i ∈ {1, . . . , n−1}
we have:
z=±Xi ···∈Wn−1
Xz ′∈Wn−1
Xz ′∈Wn−1
z=±···Xi ∈Wn−1
(A1 − Bi1) = 0,
(−Bn1 + Bi3) = 0,
z=±···Xi∈Wn−1
Xz ′∈Wn−1
Xz ′∈Wn−1
z=±Xi ···∈Wn−1
(A4 − Bi4) = 0,
(−Bn4 + Bi2) = 0.
For i 6= j in {1, . . . , n − 1} we have:
Xz ′∈Wn−1
z=±···Xj Xi ···∈Wn−1
(−Bi1 + Bj3) = 0,
Xz ′∈Wn−1
z=±···Xj Xi ···∈Wn−1
(Bi2 − Bj4) = 0.
Here z = ±Xi · · · ∈ Wn−1 (resp., z = ± · · · Xi ∈ Wn−1) means that z is an (n − 1)-commutator
word starting (resp., ending) with Xi. Likewise, z = ± · · · XjXi · · · ∈ Wn−1 means that z is an
(n − 1)-commutator word in which Xi is immediately preceded by Xj. Using the expression (4.10.1)
for J n
N (A), these six types of cancellation imply that N (A) satisfies the n-ary Hom-Nambu identity
J n
N (A) = 0.
(cid:3)
References
[1] F. Ammar, A. Makhlouf,
and S. Silvestrov, Ternary q-Virasoro-Witt Hom-Nambu-Lie algebras,
arXiv:1002.4116v1.
[2] J. Arnlind, A. Makhlouf, and S. Silvestrov, Ternary Hom-Nambu-Lie algebras induced by Hom-Lie algebras, to
appear in J. Math. Phys., arXiv:0912.0116.
[3] H. Ataguema, A. Makhlouf, and S. Silvestrov, Generalization of n-ary Nambu algebras and beyond, J. Math.
Phys. 50, no. 8 (2009), 083501.
[4] J.A. de Azc´arraga and J.M. Izquierdo, n-ary algebras: A review with applications, arXiv:1005.1028v1.
[5] J.A. de Azc´arraga and and J.C. P´erez-Bueno, Higher-order simple Lie algebras, Comm. Math. Phys. 184 (1997)
669-681.
[6] J. Bagger and N. Lambert, Gauge symmetry and supersymmetry of multiple M2-branes, Phys. Rev. D 77 (2008),
no. 6, 065008, 6 pp.
[7] M.R. Bremner and L.A. Peresi, Ternary analogues of Lie and Malcev algebras, Linear Alg. Appl. 414 (2006)
1-18.
[8] V.T. Filippov, n-Lie algebras, Sib. Mat. Zh. 26 (1985) 126-140.
[9] J.T. Hartwig, D. Larsson, and S.D. Silvestrov, Deformations of Lie algebras using σ-derivations, J. Algebra 295
(2006) 314-361.
[10] N. Jacobson, Lie and Jordan triple systems, Amer. J. Math. 71 (1949) 149-170.
[11] W.G. Lister, A structure theory of Lie triple systems, Trans. Amer. Math. Soc. 72 (1952) 217-242.
[12] W.G. Lister, Ternary rings, Trans. Amer. Math. Soc 154 (1971) 37-55.
[13] A. Makhlouf, Hom-alternative algebras and Hom-Jordan algebras, arXiv:0909.0326.
A HOM-ASSOCIATIVE ANALOGUE OF n-ARY HOM-NAMBU ALGEBRAS
15
[14] A. Makhlouf and S. Silvestrov, Hom-algebra structures, J. Gen. Lie Theory Appl. 2 (2008) 51-64.
[15] K. Meyberg, Jordan-Tripelsysteme und die Koecher-Konstruktion von Lie-Algebren, Math. Z. 115 (1970) 58-78.
[16] Y. Nambu, Generalized Hamiltonian dynamics, Phys. Rev. D 7 (1973), 2405-2412.
[17] S. Okubo, Introduction to octonion and other non-associative algebras in physics, Cambridge Univ. Press,
Cambridge, UK, 1995.
[18] N.G. Pletnev, Filippov-Nambu n-algebras relevant to physics, Sib. ´Elektron. Mat. Izv. 6 (2009) 272-311.
[19] A.P. Pozhidaev, n-ary Mal'tsev algebras, Alg. Logic. 40 (2001) 170-182.
[20] L. Takhtajan, On foundations of generalized Nambu mechanics, Comm. Math. Phys. 160 (1994) 295-315.
[21] D. Yau, Enveloping algebras of Hom-Lie algebras, J. Gen. Lie Theory Appl. 2 (2008) 95-108.
[22] D. Yau, Hom-algebras and homology, J. Lie Theory 19 (2009) 409-421.
[23] D. Yau, Hom-Maltsev, Hom-alternative, and Hom-Jordan algebras, arXiv:1002.3944.
[24] D. Yau, On n-ary Hom-Nambu and Hom-Nambu-Lie algebras, arXiv:1004.2080.
[25] D. Yau, On n-ary Hom-Nambu and Hom-Maltsev algebras, arXiv:1004.4795.
Department of Mathematics, The Ohio State University at Newark, 1179 University Drive, Newark,
OH 43055, USA
E-mail address: [email protected]
|
1207.3487 | 1 | 1207 | 2012-07-15T09:48:31 | Categorical notions of layered tropical algebra and geometry | [
"math.RA",
"math.CT"
] | This paper supplements [17], showing that categorically the layered theory is the same as the theory of ordered monoids (e.g. the max-plus algebra) used in tropical mathematics. A layered theory is developed in the context of categories, together with a "tropicalization functor" which permits us to pass from usual algebraic geometry to the tropical world. We consider tropical varieties from this categorical viewpoint, with emphasis on polynomial functions and their roots. | math.RA | math |
CATEGORICAL NOTIONS OF
LAYERED TROPICAL ALGEBRA AND GEOMETRY
ZUR IZHAKIAN, MANFRED KNEBUSCH, AND LOUIS ROWEN
Abstract. This paper supplements [17], showing that categorically the layered theory is the same as
the theory of ordered monoids (e.g. the max-plus algebra) used in tropical mathematics. A layered
theory is developed in the context of categories, together with a "tropicalization functor" which permits
us to pass from usual algebraic geometry to the tropical world. We consider tropical varieties from this
categorical viewpoint, with emphasis on polynomial functions and their roots.
1. Introduction
Tropical geometry has led to considerable mathematical success in degenerating various combinatoric
questions. At the algebraic level, the degeneration often has led to the max-plus algebra, but in certain
ways this is too crude a process to preserve many important algebraic properties. Over the last few
years, the theory of supertropical algebras has been developed in a series of papers including [16], [22],
[23], [24], and [25], in which classical notions of commutative algebra pertaining to algebraic varieties,
matrices, and valuations, carry over intrinsically to the "tropical" world. This degeneration still is too
crude to handle several issues, such as multiple roots of polynomials. A more refined structure, called
L-layered domains, was introduced in [17] together with its basic traits, in order to be able to preserve
more algebraic properties.
Recent years have seen considerable progress in the algebraic perspective of tropical geometry. Notably,
building on work of Bieri-Groves [5], Berkovich [3] and Payne [36] have shown how to view the analyti-
fication of an affine variety algebraically, in terms of valuations and multiplicative seminorms extending
valuations.
This paper is part of a series including [17] and [20]. In [17] we showed by example how the layered
structure can cope with algebraic aspects of tropical mathematics that are inaccessible to less refined
structures; some of these examples are reviewed here for the reader's convenience.
Our main purpose in this paper is to provide a more formal, unified foundation for further study,
for both the algebraic and geometric aspects. Since category theory pervades modern mathematics so
thoroughly, one feels obligated to describe the theory in categorical terms, and indeed this language
provides valuable information as to how the theory should progress, thereby throwing further light on
tropical geometry. We aim to understand those categories arising from algebraic considerations, focusing
on those algebraic aspects of the theory that reflect most directly on tropical geometry, largely via
a Zariski-type correspondence. To describe these categories in full detail would involve an inordinate
amount of technical detail, so we often make simplifying assumptions when they do not impact on the
tropical applications. Even so, each aspect of the theory involves its corresponding categories, and so
Date: April 3, 2019.
2010 Mathematics Subject Classification. Primary 06F20, 11C08, 12K10, 14T05, 14T99, 16Y60; Secondary 06F25,
16D25.
Key words and phrases. Tropical categories, tropical algebra, tropical geometry, valued monoids, valuation, tropicaliza-
tion, Zariski topology.
This research of the first and third authors is supported by the Israel Science Foundation (grant No. 448/09).
The research of the first author has been supported by the Oberwolfach Leibniz Fellows Programme (OWLF), Mathe-
matisches Forschungsinstitut Oberwolfach, Germany.
The second author was supported in part by the Gelbart Institute at Bar-Ilan University, the Minerva Foundation at
Tel-Aviv University, the Mathematics Dept. of Bar-Ilan University, and the Emmy Noether Institute. Research on this
paper was carried out by the three authors in the Resarch in Pairs program of the MFO in Oberwohlfach.
We thank Steve Shnider and Erez Sheiner for explaining the proof of Payne's theorem in our tropical seminar.
1
2
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
there are many categories to be described here. Although the language involves some technicalities, we
try to keep it to a minimum, leaving subtler matters to [20]. Another related paper is [19], which delves
into considerable detail in the supertropical setting, for which we generalize parts to the layered setting,
To obtain the appropriate functors, we need first to make categories of the classical "algebraic world"
and the "tropical world." Informally, the classical "algebraic world" is described by the categories asso-
ciated to classical algebraic geometry, often over the complex numbers C.
A deep connection between tropical geometry and valuation theory is already implicit in [5], and it is
convenient to work over algebraically closed fields with valuation. Thus, as the algebraic aspect of the
tropical theory has developed, C has been replaced by the field of Puiseux series, an algebraically closed
field endowed with a (nonarchimedean) valuation, whose target is an ordered group, so it makes sense to
work with ordered groups, or, slightly more generally, ordered monoids.
There has been considerable recent interest in developing algebraic geometry over arbitrary monoids
[4, 8], and we shall draw on their work. One theme of this paper is how the assumption of an order on
the monoid enriches the theory. The ordered monoid most commonly used in the tropical world is the
"max-plus" algebra M (or its dual, the "min-plus" algebra), cf. [1], [2], [14], and [30]. Our first main
result (Proposition 3.11), which sets the flavor for the paper, is that the category of ordered cancellative
monoids is isomorphic to the category of bipotent semirings (without a zero element). Any ordered
monoid can be viewed as a semiring, where multiplication is the given monoid operation and addition is
defined by taking a + b to be max{a, b} in Proposition 3.11. The universal of the appropriate forgetful
functor is constructed in this context, in Proposition 3.13.
Since the underlying algebraic structures now are semirings, we switch to the language of semirings in
order to be able to adapt concepts from ring theory and module theory, such as polynomials and matrices.
We find it more convenient to work in the category of semirings†, defined as semirings not necessarily
having a zero element, for the following reasons:
• The duality given in Proposition 3.12 holds for semirings† but not for semirings;
• Proofs are usually neater for semirings†, since the zero element 0 of a semiring needs special
treatment;
• Many important examples (such as Laurent series and tori) are defined over semirings† but
not over semirings (and in particular, Bieri-Groves' main theorem [5, Theorem A] is given for
multiplicative groups);
• Once we get started with the layered theory, it is more natural to utilize a 0-layer (an ideal
comprised of several elements) rather than a single element 0; anyway, one can recover the
element 0 by inserting it into the 0-layer.
One might counter that various critical aspects of geometry such as intersection theory (which involve
curves such as xy = 0) require a zero element. This turns out to be less important in the tropical theory
since the zero element, −∞, already is artificial, and can be dealt with at the appropriate time.
To describe tropicalization categorically, we utilize the category ValField describing fields with valua-
tion, or, slightly more generally, the category ValDom describing integral domains with valuations. (In
the sequel [20] to this paper, we proceed still further, with valued rings.) The theory is applicable to fields
with valuation, in particular to the Puiseux series field. Intuitively, the corresponding tropical category
just reformulates the valuation, where the operations are taken from the target of the valuation. Our
category ValMon (cf. §4) is described in the language of monoids, in order to permit other tropicalization
techniques.
At this point, let us stress that one principal role of the tropical algebra is to provide an intrinsic
algebraic setting for studying valuations in ValField and their extensions, as described in §6.2, via Maslov
dequantization [29] or the degeneration of "amoebas" [11], [32], and [40]. In a few words, one takes the
power valuation to pass from the Puiseux series field to Q, viewed as the max-plus algebra. This is
formalized in Remark 4.9 as the functor Fval from the category of valued monoids to the category of
ordered monoids (or, equivalently, bipotent semirings).
Unfortunately, the algebraic theory of bipotent semirings† is too weak to provide much information
without additional structure. Accordingly, the algebra M was extended to extended tropical arith-
metic [15] which evolved into the supertropical domain [22] and then to the layered domain† R(L,G)
CATEGORICAL LAYERED MATHEMATICS
3
of an ordered monoid G with respect to an indexing semiring† L, called the the sorting set, cf. [17, Def-
inition 3.5]. L-layered domains† become max-plus algebras when L is {1} and become supertropical
domains when L is {1,∞}.
The general L-layered theory, set forth in §5 and §6, has many advantages over the other theories, as
shown in [17], because it enables us to distinguish among different ghost levels. This is really a linguistic
distinction, as is explained in the next paragraph. Nevertheless, there is a definite advantage in making
use of the tools available in the language of layered semirings.
Whereas the supertropical domain enables us to distinguish multiple tropical roots (say in the poly-
nomial f (λ) = (λ + 3)2) from single roots, it does not say anything about the multiplicity of the corner
root 3. Thus, it would not enable us to tell intrinsically whether 3 is a root or a pole of the function (λ+3)j
(λ+3)k ,
whereas questions of this sort are answered at once in the layered structure. More sophisticated geometric
examples are given in Example 7.24.
For the reader's convenience let us point also to several applications of the layered structure from [17]:
• [17, Theorem 8.25] The ν-multiplicativity of the resultant of tropical polynomials (in one inde-
terminate).
• [17, Theorem 8.33] The multiplicativity of the resultant of products of primary tropical polyno-
mials (in one indeterminate).
• [17, Theorem 9.8] The computation of the layered discriminant of a tropical polynomial.
• [17, Example 10.6] Multiplicity of roots of tropical polynomials, by means of [17, Equation (9.1)].
• [17, Example 10.8] Unique factorization in many cases, as well as integration being defined.
Thus, we rely heavily on Construction 5.1 in order to pass back and forth between cancellative ordered
monoids and L-layered domains†, and this should be considered the main thrust of the layering procedure.
Note that Construction 5.1 is formulated for semirings† without 0, in order to avoid complications. The
more general theory is given in [20].
Intuitively, to obtain the appropriate layered category one might expect to take morphisms to be
semiring† homomorphisms that preserve the layers, and these indeed play a key role to be described below.
The category of main interest for the tropical algebraic theory is the category ULayBidom† of uniform L-
layered bi-domains†, which by Theorem 6.3 is isomorphic to the category OMon+ of cancellative ordered
monoids, under the natural functor that restricts a uniform L-layered bi-domain† to its submonoid of
tangible elements. In this way, we begin to see how identifications of categories help guide us in developing
the theory.
This leads us to a delicate side issue. Although the ordered monoids of interest in the tropical theory
are cancellative, such as the real max-plus algebra or, more generally, any ordered group, homomorphic
images of cancellative monoids need not be cancellative. Thus, for a rich algebraic theory, we need a way
of passing from arbitrary ordered monoids to layered semirings†. Unfortunately the naive generalization
of Construction 5.1 is not a semiring† since distributivity fails! In order not to go too far afield in this
paper, we stick with cancellative monoids, and consider noncancellative monoids in [20].
In §6.2, we get to the functor FLTrop : ValMon+ → ULayBidom†, which describes the passage to the
layered tropical world. FLTrop is applied to fields with valuation, in particular the field of Puiseux series,
and enables one to translate equality to the "surpassing relation" described in [17, §3.2]. In full generality
the functor FLTrop involves subtleties discussed in [20].
The functor FLTrop is not faithful, since it only measures the action of the given valuation, and does not
enable us to distinguish among elements having the same value. Thus in §6.3 we also introduce briefly the
exploded tropicalization functor utilized in [31] and another functor FLTrop;unit which retains extra
information given in Proposition 6.10, such as is contained in "coamoebas." Borrowing from classical
valuation theory, we describe the exploded tropicalization functor in terms of the associated graded
algebra, noting that in the case of the field of Puiseux series, the components of the associated graded
algebra can be identified with the base field.
Although monoids recently have been seen to provide much of the underpinning for algebraic geometry,
cf. [4, 8, 33, 34] for example, classical algebraic geometry relies for a large part on roots of polynomials,
which can be understood more easily using semirings†. Our approach to tropical geometry is to define
affine varieties as sets of "ghost roots" of polynomials.
4
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
As is well known, and discussed in detail in [22], in contrast to the classical situation for polynomials
over algebras over an infinite field, different tropical polynomials over a semiring† R often take on the
same values identically, viewed as functions. Furthermore, in max-plus situations one often wants to use
variants such as Laurent polynomials (involving λ−1 as well as λ) or polynomials with rational exponents,
or even more generally one could talk in the language of the lattice of characters and its dual lattice,
cf. [35, §2.2]. Also, as in classical algebraic geometry, we often want to limit the domain of definition
to a given subset of R(n) such as an algebraic variety. Thus, we work directly with functions from
a set S to a layered domain, denoted Fun(S, R), or, more specifically, polynomially defined functions,
denoted Pol(S, R) or Laurent polynomially defined functions, denoted Laur(S, R). In Proposition 7.8 we
check that passing to the monoid of functions from S to an ordered monoid M and then translating to
semirings† yields the same categorical theory as moving first to a bipotent semiring† and then passing to
its function semiring†.
Thus, taking S ⊂ R(n), we redefine polynomials and monomials over R intrinsically as functions from S
to R, leading to an analog of the Zariski topology in Definition 9.2. This enables us to define a coordinate
semiring† via Definition 9.5.
Our view of tropical geometry relies largely on Kapranov's Theorem, as extended by [36], which
describes the 1:1 correspondence between roots of polynomials and corner roots of their tropicalizations.
This process is understood categorically in terms of the supertropical structure, in §8. Ironically, although
the Kapranov-Payne Theorem can be stated in the language of the tropicalization functor FLTrop, the
exploded tropicalization functor is needed (at least implicitly) in order to carry out the proof.
1.1. Overview of the major categories and functors in this paper.
In summary, let us review the main algebraic categories and their uses.
• The category ValField (resp. ValMon) describing fields with valuation (resp. integral domains)
with valuation. This is the ultimate arena of investigation, but its theory often is very difficult,
thereby historically giving rise to tropical mathematics, which can be thought of as a degeneration
of ValField.
• The category OMon+ of cancellative ordered monoids, which is isomorphic to the category Bipot†
of bipotent semirings. This is the traditional algebraic category underpinning tropical mathemat-
ics, but is too coarse a degeneration for many algebraic arguments, and often requires returning
to ValField in proofs.
• The category ULayBidom† of uniform L-layered bi-domains†. For L = {1} this is just the
category OMon+. For L = {1,∞} or L = {0, 1,∞} we get the supertropical theory, which suffices
in linear algebra for the investigation of nonsingular matrices, bases, characteristic polynomials,
and related notions. In order to discuss multiple roots of polynomials and singularity of curves,
one needs to take L containing N.
• The category of exploded L-layered bi-domains†. This is used by Sheiner and Shnider for the
proof of the Kapranov-Payne theorem, as well as other deep results in the theory.
Here is a diagram of the categories under discussion in this paper, and the main functors connecting
them.
ValMon+
Fval
/ OMon+
FOMon
◆
◆
◆
◆
◆
Flay
◆
FunMon(S,
◆
◆
◆
◆
◆
◆
◆
◆
'◆
❖
'❖
ULayBidom†
/ Bipot† ⊂ Semir†
)
FunSemir† (S,
)
OMon+
/ Semir†
❖
❖
❖
❖
❖
❖
❖
FLTrop
❖
❖
❖
❖
❖
❖
❖
❖
❖
❖
❖
❖
❖
FLTrop;exp
❖
❖
❖
❖
❖
❖
❖
'❖
S
ValDom
S
ValField
ULayBidom† × Ring
'
/
'
o
o
/
'
/
O
O
CATEGORICAL LAYERED MATHEMATICS
5
Fval, which can be viewed as the customary tropicalization procedure, formalizes the order valuation
on Puiseux series, which in most recent research has replaced the logarithm as the means of tropicalizing
a variety.
geometry to the layered tropical world of this paper.
putting tools of semiring theory (such as polynomials) at our disposal.
FLTrop, perhaps the most important functor in our theory, takes us from the classical world of algebraic
FOMon is the functor that enables us to pass from ordered monoids to bipotent semirings, thereby
Flay is the functor that enables us to "layer" an ordered monoid, and thus pass to the layered theory.
FLTrop;exp is the "exploded" functor, which preserves the leading coefficient of the original polynomial
when tropicalizing, and thus permits Payne's generalization of Kapranov's theorem (and its application
to tropical varieties).
The Fun functors take us to semirings† of functions, thereby enabling us to treat polynomials (as
functions).
At the conclusion of this paper, we consider how the layered category LayBidom† enables us to define
corner varieties, and we relate the algebraic and geometric categories along the classical lines of algebraic
geometry, obtaining a Zariski-type correspondence in Proposition 9.8.
There also is a functor FLTrop;unit : ValDom → ULayBidom† ×ValMon(1) which we did not put into
the diagram, whose justification is given in the discussion after Proposition 6.10.
We start with the category Mon of monoids and their monoid homomorphisms, viewed in the context
2. Background
of universal algebras, cf. Jacobson [27, §2].
Definition 2.1. A semigroup is a set with an associative operation, usually written multiplicatively as ·.
A monoid M := (M,· ) is a semigroup with a unit element 1M. A semigroup M is (left) cancellative
with respect to a subset S if for every a1, a2,∈ M, b ∈ S,
b · a1 = b · a2
implies a1 = a2.
M is cancellative if M is cancellative with respect to itself.
An element a of M is absorbing if ab = ba = a for all b ∈ M. Usually the absorbing element (if it
exists) is denoted as the zero element 0M, but it could also be identified with −∞. A semigroup M is
pointed if it has an absorbing element 0M. A pointed semigroup M is cancellative if M is cancellative
with respect to M \ {0M}. A subset a ⊂ M is a left (right) semigroup ideal if Ma ⊂ a (aM ⊂ a).
The semigroups (as well as monoids) in this paper are presumed commutative, so left semigroup ideals
are semigroup ideals.
bm = a.
φ(a1a2) = φ(a1)φ(a2),
A semigroup M := (M,· ) is divisible if for every a ∈ M and m ∈ N there is b ∈ M such that
A semigroup homomorphism is a map φ : M → M′ satisfying
∀a1, a2 ∈ M.
(When dealing with pointed semigroups, we also require that φ(0M) = 0′
M
is a semigroup homomorphism φ : M → M′ also satisfying φ(1M) = 1M′ .
2.1. Semirings without zero. We ultimately work in the environment of semirings (or, more precisely,
semirings without a zero element, which we call a semiring†. A standard general reference for the
structure of semirings is [12]; also cf. [9]. Thus, a semiring† (R, +,· , 1R) is a set R equipped with two
binary operations + and · , called addition and multiplication, together with a unit element 1R such
that:
.) A monoid homomorphism
(1) (R, +) is an Abelian semigroup;
(2) (R,· , 1R) is a monoid with unit element 1R;
(3) Multiplication distributes over addition on both sides.
When the multiplicative monoid (R,· , 1R) is cancellative, we say that (R, + , · , 1R) is a domain†; when
(R,· , 1R) is also an Abelian group, we say that (R, + , · , 1R) is a semifield†. As customary, N denotes
the positive natural numbers, which is a cancellative domain†.
6
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Definition 2.2. A homomorphism ϕ : R → R′ between two semirings† is defined as a homomorphism
of multiplicative monoids that also preserves addition, to wit,
A semiring† isomorphism is a semiring† homomorphism that is 1:1 and onto.
ϕ(a + b) = ϕ(a) + ϕ(b)
for all a, b ∈ R.
Thus, we have the category Semir† of semirings† and their homomorphisms. This is closely related to
the category Semir of semirings and semiring homomorphisms, especially since the semirings R of interest
in the tropical theory, besides being multiplicatively cancellative, have the property that a+b 6= 0R unless
a = b = 0R; in other words, R \ {0R} is closed under addition.
Remark 2.3. Any semiring† R can be embedded in a semiring R ∪ {0} by formally adjoining a zero
element 0 satisfying 0 + a = a + 0 = a and 0· a = a· 0 = 0, ∀a ∈ R ∪{0}. Conversely, if R is a semiring
such that R \ {0R} is closed under multiplication and addition, then R \ {0R} is a semiring†.
Proposition 2.4. The category Semir† is isomorphic to a subcategory of the category Semir.
Proof. We just apply Remark 2.3, noting that any semiring† homomorphism ϕ : R → R′ can be extended
to a semiring homomorphism ϕ : R ∪ {0R} → R′ ∪ {0R′} by putting ϕ(0R) = 0R′ .
(cid:3)
An ideal a of a semiring† R is defined to be a sub-semigroup of (R, +) which is also a (multiplicative)
semigroup ideal of (R,· , 1R). (Clearly, when R has a zero element 0R, then 0R ∈ a.)
Example 2.5. If R is a semiring, then {0R} is an ideal of the semiring R ∪ {0R} of Remark 2.3.
The tropical theory is closely involved with certain kinds of semirings†.
Definition 2.6. A semiring† R is idempotent if
a + a = a
∀a ∈ R;
R is bipotent if
∀a, b ∈ R.
The max-plus algebra is the prototype of a bipotent semiring†.
a + b ∈ {a, b}
2.2. Congruences. Unfortunately, kernels, such an important feature of category theory, play virtually
In ring theory, the kernel ϕ−1(0R′ ) of any onto
no role in the general structure theory of semirings.
homomorphism ϕ : R → R′ is an ideal a of R, and furthermore one can recover R′ as isomorphic to R/a.
This is not the case with semirings†. Ideals do not play such a powerful role in the structure theory of
semirings†, since the construction R/a is problematic for an arbitrary ideal a (the difficulty arising from
the fact that distinct cosets need not be disjoint).
Instead, one needs to consider more generally equivalence relations preserving the semiring† operations.
From the general theory of universal algebra, one defines a congruence Ω of an algebraic structure A
to be an equivalence relation ≡ which preserves all the relevant operations and relations; we call ≡ the
underlying equivalence of Ω. Equivalently, a congruence Ω is a sub-semiring† of A × A that contains
the diagonal
diag(A) := {(a, a) : a ∈ A}
as described in Jacobson [27, §2]. In other words, writing the underlying equivalence relation as a ≡ b
whenever (a, b) ∈ Ω, we require that ≡ preserves all the relevant operations and relations.
Remark 2.7. We recall some key results of [27, §2]:
• Given a congruence Ω of an algebraic structure A, one can endow the set
A/Ω := {[a] : a ∈ A}
of equivalence classes with the same (well-defined) algebraic structure, and the map a 7→ [a]
defines an onto homomorphism A → A/Ω. (For this reason, Berkovich [4] calls them "ideals,"
but this terminology conflicts with some of the literature, and we prefer to reserve the usage of
"ideal" for the usual connotation.)
CATEGORICAL LAYERED MATHEMATICS
7
• For any homomorphism ϕ : A → A′, one can define a congruence Ω on A by saying that a ≡ b
iff ϕ(a) = ϕ(b). Then ϕ induces a 1:1 homomorphism
eϕ : A/≡ → A′,
via eϕ([a]) = ϕ(a).
We repeat the definition of congruence in each specific case that we need. Thus, a congruence Ω on
a semigroup M is an equivalence relation that preserves multiplication, in the sense that if a1 ≡ b1 and
a2 ≡ b2, then a1a2 ≡ b1b2. In this case, the set of equivalence classes M/Ω becomes a semigroup under
the operation
[a][b] = [ab],
and there is a natural semigroup homomorphism given by a 7→ [a]. When M is a monoid, this becomes a
monoid homomorphism, since [1M ] is the multiplicative unit of M/Ω. When M is pointed, then M/Ω
is also pointed, with absorbing element [0M].
Here is another instance of a congruence that comes up in the passage from arbitrary monoids to
cancellative monoids.
Example 2.8. Given an equivalence relation ≡ on a semigroup M, and a sub-semigroup S of M, we
define the equivalence ≡S given by b1 ≡S b2 if b1s ≡ b2s for some s ∈ S. When ≡ defines a congruence Ω,
then ≡S also defines a congruence ΩS. This congruence then identifies b1 and b2, thereby eliminating
instances of non-cancellation, and is a useful tool.
2.2.1. Congruences over semirings†. The congruence Ω is a semiring† congruence on a semiring† R iff
a1 ≡ a2 and b1 ≡ b2
imply (cid:26) a1 + b1 ≡ a2 + b2,
≡ a2b2.
a1b1
(2.1)
Lemma 2.9. To verify the conditions in (2.1) for commutative semirings†, it is enough to assume b1 = b2
and show for all a1, a2, and b in R:
a1 ≡ a2
implies a1 + b ≡ a2 + b;
implies a1b ≡ a2b.
Proof. a1 + b1 ≡ a2 + b1 ≡ a2 + b2. Likewise, a1b1 ≡ a2b1 ≡ a2b2.
a1 ≡ a2
(2.2)
(2.3)
(cid:3)
It often turns out that (2.2) enables us to obtain (2.3). On the other hand, in the case of semifields†,
multiplicative cosets are more easily described than additive cosets, as is described in detail in [13]. To
wit, let N := {a ∈ R : a ≡ 1R}. For any a ∈ R we have ab1 ≡ ab2 iff b1b−1
We write diag(M) for {(a, a) : a ∈ M}. As Berkovich [4] points out, any semigroup ideal a of
a semigroup M gives rise to the congruence (a × a) ∪ diag(M), which corresponds to the Rees factor
semigroup, and the analogous statement holds for monoids. A wrinkle emerges when we move to bipotent
semirings†, since (a× a)∪ diag(M) need not be closed under addition. Thus, the applications are limited,
and are discussed in [20].
2 ∈ N.
Definition 2.10. An identity f = g of a semiring† R is an elementary sentence f (x1, . . . , xm) =
g(x1, . . . , xm) that holds for all x1, . . . , xm in R.
Remark 2.11. Suppose we want to force a semiring† R to satisfy a particular identity, in the sense
that we want a semiring† ¯R in which f = g is an identity, together with a surjective homomorphism
ϕ : R → ¯R satisfying the universal property that any homomorphism of R to a semiring† satisfying the
identity f = g factors through ϕ.
Intuitively, one must mod out the relation f = g by putting f (a1, . . . , am) ≡ g(a1, . . . , am) for all ai
in R. For semirings†, in view of Lemma 2.9, since we are dealing with congruences, we must mod out
the equivalence relation obtained by putting f (a1, . . . , am) + c ≡ g(a1, . . . , am) + c and f (a1, . . . , am)c ≡
g(a1, . . . , am)c for all c and ai in R.
8
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Example 2.12. Consider the additive idempotence identity x + x = x. We attain this by imposing
the equivalence relation given by a + a ≡ a,∀a ∈ R. The congruence that it generates must also satisfy
the relation a + a + c ≡ a + c,∀a ∈ R.
But then we also get (2.3), since ab + ab = (a + a)b and
(a + a + c)b = ab + ab + cb ≡ ab + cb = (a + c)b.
Thus, (2.2) already defines the congruence. (This observation is to be elaborated shortly.)
Note that additive idempotence implies all identities of the form x + x + ··· + x = x.
Whereas in ring theory the equivalence class [0R] determines the congruence Ω, this is no longer the
case for semirings, and we need to consider all the classes {[a] : a ∈ R}. This is another reason that we
do not require the element 0 in a semiring†, for it has lost much of its significance. Nevertheless, ideals
do play a role in the layered algebraic theory, pursued in a different paper [21].
Lemma 2.13. The bipotent semirings† comprise a full subcategory Bipot† of Semir†.
Proof. If ϕ : R → R′ is a semiring† homomorphism and a + b ∈ {a, b}, then
ϕ(a) + ϕ(b) = ϕ(a + b) ∈ {ϕ(a), ϕ(b)}.
(cid:3)
2.3. Hom and Adjoint functors. We need to use some well-known facts about categories.
Definition 2.14. For any category C and some given object A in C, we recall the well-known covariant
functor
which sends an object B in C to Hom(A, B), and which sends the morphism φ : B → B′ to the morphism
Hom( , φ) : Hom(A, B) → Hom(A, B′) given by f 7→ φf for f : A → B.
Likewise, given an object B in D, we define the contravariant functor
Hom(A,
) : C → Set,
which sends an object A to Hom(A, B), and which sends the morphism ϕ : A′ → A to the morphism
Hom(ϕ,
) : Hom(A′, B) → Hom(A, B) given by f 7→ f ϕ for f : A → B.
Hom(
, B) : D → Set
Recall that a functor F : C → D is a left adjoint to H : D → C (and H is a right adjoint to F ) if
there is a canonical identification Ψ : Hom(F (A), B) → Hom(A,H(B)) for all objects A of C and B of D,
for which the following diagrams are always commutative for all morphisms ϕ : A → A′ and φ : B → B′:
Hom(F (A), B)
Ψ
Hom(A,H(B))
Hom(F (A), B)
Ψ
Hom(F (ϕ),
)
/ Hom(F (A′), B)
Ψ
Hom(ϕ,
)
Hom( ,φ)
/ Hom(A′,H(B)),
/ Hom(F (A), B′)
Ψ
It is well-known that any left adjoint functor is unique up to isomorphism.
Hom(A,H(B))
Hom( ,H(φ))
/ Hom(A,H(B′)).
2.4. Universals. Recall from [27, §1.7 and §1.8] that the adjoint functor of a functor F : C → D
is obtained by identifying the appropriate universal U of F together with the canonical morphisms
ι : D → F (U (D)), for objects D in D, satisfying the property that for any morphism f : D → F (C) in D
and object C in C, there is a morphism g : U (D) → C in C such that
F (g ◦ ι) = f.
The example used in [8] is for the forgetful functor from K-algebras to monoids; its universal is the
monoid algebra K[M] of a monoid M.
/
/
/
/
endowed with addition
[M] =(cid:26)Xa∈S
Xa∈S
a + Xa∈S′
a S ⊂ M(cid:27)
a = Xa∈S∪S′
a ,
CATEGORICAL LAYERED MATHEMATICS
9
Example 2.15. We define the forgetful functor Semir† → Mon, by forgetting addition. The appropriate
universal in this case is the monoid semiring† N[M], defined analogously to the monoid algebra.
Definition 2.16. [M] denotes the semiring† obtained by taking N[M] modulo the additive idempotence
congruence of Example 2.12. Explicitly, [M] is comprised of formal sums of distinct elements of the
monoid M, i.e.,
and multiplication is obtained from the original multiplication in M, extended distributively.
Example 2.17. Since additive idempotence defines an identity, one has the category of additively idem-
potent semirings†; the forgetful functor to Mon now has the universal [M]. The customary way to view
tropical mathematics is by means of the max-plus semiring†, which is additively idempotent.
3. Pre-ordered semigroups, monoids, and semirings
Recall that a partial pre-order is a transitive relation (≤); it is called a partial order if it is
A partial pre-order is called a preorder if any two elements are comparable. A (total) order is a
antisymmetric, i.e., a ≤ b and b ≤ a imply a = b. We write a < b when a ≤ b but a 6= b.
partial order which is also a preorder.
3.1. Pre-ordered semigroups. We work with pre-ordered semigroups in this paper. The natural defi-
nition in terms of universal algebra is the following:
Definition 3.1. A semigroup M := (M,· ) (or a monoid M := (M,· , 1M)) is partially pre-ordered
(resp. partially ordered, pre-ordered, ordered) if it has a partial pre-order ≤ (resp. partial order,
pre-order, order) such that
b ≤ c
(3.1)
We denote an ordered semigroup by (M,· ,≤). Thus, totally ordered semigroups satisfy the following
(3.2)
implies ab ≤ ac and
∀a ∈ M.
ba ≤ ca,
property:
a max{b, c} = max{ab, ac},
∀a, b, c ∈ M.
We say that the relation (≤) is strict if
b < c
and ba < ca,
implies ab < ac
(3.3)
Note 3.2. This definition requires that all elements of (M,· ,≤) are positive or 0, an implicit assumption
made throughout this paper, to be discussed after Definition 3.8.
Lemma 3.3. A total order (≤) on a semigroup M is strict iff the semigroup M is cancellative.
Proof. (⇒) : Suppose ab = ac. By symmetry, we may assume that b ≤ c. But if b < c then ab < ac, a
contradiction, so we conclude b = c.
∀a ∈ M.
(⇐) : If b < c then ab 6= ac, implying ab < ac.
Let us construct the appropriate categories.
(cid:3)
Definition 3.4. An order-preserving semigroup homomorphism is a homomorphism φ : M → M′
satisfying the condition (where ≤ denotes the partial order on the appropriate semigroup):
(3.4)
PPreOMon, PreOMon, POMon, POMon+, OMon, and OMon+ denote the respective categories of
partially pre-ordered, pre-ordered, partially ordered, cancellative partially ordered, ordered, and cancella-
tive ordered monoids, whose morphisms are the order-preserving homomorphisms.
implies φ(a) ≤ φ(b),
∀a, b ∈ M.
a ≤ b
By definition, OMon+ is a full subcategory both of OMon and of POMon+, each of which is a full
subcategory of POMon, which is a full subcategory of PPreOMon.
10
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Remark 3.5. The forgetful functor from the category POMon to the category PPreOMon is obtained by
viewing any partially ordered monoid M naturally in PPreOMon.
We also can go in the other direction.
Remark 3.6. For the class of partially pre-ordered semigroups, our congruences Ω also satisfy the prop-
erty that if a1 ≤ a2 and bi ≡ ai, then b1 ≤ b2. In this case, M/Ω inherits the partial pre-order given by
[a] ≤ [b] iff a ≤ b.
Proposition 3.7. There is a retraction F : PPreOMon → POMon to the forgetful functor of Remark 3.5.
Namely, we take the congruence Ω on a pre-ordered monoid M given by a ≡ b when a ≤ b and b ≤ a,
and define F (M) := M/Ω.
Proof. It is easy to see that ≡ is an equivalence relation that preserves the operation and the order, so is
an ordered monoid congruence, and thus induces a partial order on M/Ω according to Remark 3.6.
We claim that any order-preserving homomorphism φ : M → M′ induces an order-preserving ho-
momorphism eφ : M/Ω → M′/Ω. Indeed, if a ≡ b, then a ≤ b and b ≤ a, implying φ(a) ≤ φ(b) and
φ(b) ≤ φ(a), yielding φ(a) ≡ φ(b).
The functor F is a retraction to the forgetful functor, since it acts trivially on any total ordered
monoid.
(cid:3)
3.2. Pre-ordered semirings†.
Definition 3.8. We say that a semiring† R is pre-ordered (resp. partially ordered, ordered) if it
has a pre-order ≥ (resp. partial order, order) with respect to which both the monoid (R,· , 1R) and the
semigroup (R, +) satisfy Condition (3.1) of Definition 3.1.
In other words, both multiplication and addition preserve the pre-order. There is a delicate issue in this
definition. In the rational numbers, viewed as a multiplicative monoid, we have 1 < 2 but (−1)1 > (−1)2.
This difficulty is dealt with in [17], in which we define the order in terms of a cone of "positive" elements.
Definition 3.8 is the special case in which all elements of R are positive or 0, and is reasonable for tropical
mathematics since the "zero" element (when it is included) is minimal. We use Definition 3.8 here because
it is more appropriate to our categorical treatment.
A semiring† R has the infinite element ∞ if
for some a ∈ R.
Recall from [17, Corollary 2.15] that if R has a unique infinite element, then
∞ + a = ∞ = ∞ · a = a · ∞,
(3.5)
∞ + a = ∞ = ∞ · a = a · ∞,
∀a < ∞.
Nonzero positive elements of an ordered semiring† need not be finite, and we could have several infinite
elements (as can be seen easily by means of ordinals). We do not deal with such issues in this paper, and
assume there is at most one infinite element ∞.
The following observation is implicit in [13, Theorem 4.2].
Proposition 3.9. There is a natural functor Semir† → PPreOMon, where we define the preorder on a
semiring† R given by
a ≤ b
iff
a = b
or
b = a + c
for some c ∈ R.
This functor always yields the trivial partial preorder on rings, since then b = a + (b− a). The situation
is quite different for the semirings† arising in tropical mathematics, because of bipotence.
Proposition 3.10. Suppose R is an idempotent semiring†.
(i) a ≤ b iff a + b = b.
(ii) ≤ is a partial order, which is total when R is bipotent.
Proof. (i): (⇐) Take c = b.
(⇒) Suppose a + c = b. Then
a + b = a + (a + c) = (a + a) + c = a + c = b.
CATEGORICAL LAYERED MATHEMATICS
11
(ii): Transitivity follows because a + b = b and b + c = c imply
a + c = a + (b + c) = (a + b) + c = b + c = c.
It remains to prove antisymmetry. Suppose a ≤ b and b ≤ a. Then, in view of (i), b = a + b = a. (cid:3)
We are ready for a key identification of categories.
Proposition 3.11. There is a faithful functor FOMon : OMon → Semir†, whose image is Bipot†.
Proof. Given any totally ordered monoid (M,· ,≥, 1M) we define a + b to be max{a, b}. Then (M, +) is
a semigroup, since
Finally, this gives rise to a semiring†, since, by Equation (3.2),
(a + b) + c = max{a, b, c} = a + (b + c).
Any order-preserving monoid homomorphism ϕ : M → M′ is a semiring† homomorphism, since for a ≤ b
we have
(a + b)c = max{a, b}c = max{ac, bc} = ab + ac.
ϕ(a + b) = ϕ(b) = ϕ(a) + ϕ(b).
Conversely, given a bipotent semiring† R, the relation ≤ of Proposition 3.9 is a total order by Propo-
sition 3.10. Furthermore, ≤ is preserved under multiplication, since b ≤ c implies b + c = c and thus
ab+ac = a(b+c) = ac, yielding ab ≤ ac. Any semiring† homomorphism ϕ : R → R′ is an order-preserving
monoid homomorphism, for if a ≤ b, then
ϕ(a) + ϕ(b) = ϕ(a + b) = ϕ(b),
implying ϕ(a) ≤ ϕ(b).
(cid:3)
Note that we have just reconstructed the max-plus algebra. We will rephrase this result in the layered
setting, as Theorem 6.3.
Proposition 3.11 enables one to pass back and forth between categories of totally ordered monoids and
bipotent semirings. The first category enables us to exploit techniques from valuation theory, whereas
the second enables us to introduce concepts from ring theory such as polynomials, modules, matrices,
and homology theory.
Proposition 3.12. There is a functor Bipot† → Bipot† sending a semiring† to its dual bipotent
semiring† obtained as the same multiplicative monoid, but reversing the bipotence in addition; i.e., if
originally a + b = a, now we put a + b = b.
Proof. This is seen readily by defining the pre-order given by a ≥ b iff a + b = a; then the dual bipotent
semiring† corresponds to the reverse pre-order, and any homomorphism preserves the (reverse) order. (cid:3)
For example, the dual semiring† of (R, max, +, 0) is (R, min, +, 0). (The number 0 is really the unit
element 1R.)
3.3. The universal for the Frobenius property. Usually one works with commutative, totally ordered
monoids and semirings†. In this case, recall the well-known Frobenius property, cf. [22, Remark 1.1]:
(a + b)m = am + bm
(3.6)
for any m ∈ N.
These are identities of R, so we could define the Frobenius monoid semiring† [M] of an arbitrary
monoid M, in which, in view of Remark 2.11, we impose on [M] (defined in Definition 2.16) the relations
a(cid:19)m
(cid:18)Xa∈S
+ c =(cid:18)Xa∈S
am(cid:19) + c,
a(cid:19)m
(cid:18)Xa∈S
c =Xa∈S
amc,
for S ⊂ M finite.
we have
a(cid:19)m
dm +(cid:18)Xa∈S
Note that when R is divisible, these relations are formal consequences of (3.6) since writing c = dm
=(cid:18)d +Xa∈S
a(cid:19)m
= dm +Xa∈S
am;
dm(cid:18)Xa∈S
a(cid:19)m
=(cid:18)dXa∈S
a(cid:19)m
= dmXa∈S
am.
12
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
The Frobenius monoid semiring† [M] satisfies the following universal property:
Proposition 3.13. Suppose ϕ : M → M′ is a monoid homomorphism, where the monoid M′ is totally
ordered. Viewing M′ as a bipotent semiring† via Proposition 3.11, we have a natural homomorphism
bϕ : [M] → M′
bϕ(cid:18)Xa∈S
a(cid:19) =Xa∈S
ϕ(a),
given by
satisfying the universal property that ϕ factors as
ϕ : M
/ [M]
bϕ
/ M′.
satisfies the Frobenius property.
Proof. The map given by Pa∈S a 7→ Pa∈S ϕ(a) is the desired semiring† homomorphism, since M′
On the other hand, the same argument shows that [M] itself is not ordered as a monoid, since one
can provide any ordered monoid M with two orders, one order making a + b = a and the reverse order
making a + b = b, as shown in Remark 3.12. The point here is that the Frobenius property, being
an algebra identity, permits the definition of a universal. Furthermore, one could define the category
of semirings† satisfying the Frobenius property, which comprises a full subcategory of Semir† . Although
Proposition 3.13 indicates that this is the "correct" category in which to conduct much of the investigation
in tropical algebraic geometry, we forego further consideration of this category in this paper.
(cid:3)
4. Integral domains and monoids with valuation
We turn to the main notion of "tropicalization." As indicated in the introduction, we need to consider
integral domains W with valuation v : W \ {0W} → G, having cancellative (ordered) value monoid G. In
valuation theory it is customary to write the operation of the value monoid G as addition, and to utilize
the axiom
Note that we can replace the valuation v by v′ := −v to get the dual equation
v(a + b) ≥ min{v(a), v(b)}.
v′(a + b) ≤ max{v′(a), v′(b)}.
(4.1)
We adjust the notation of valuation theory to fit in with the algebraic language of semirings†. Thus,
from now on in this paper, we use multiplicative notation, written G := (G,· ,≥, 1G), for the value
monoid G with unit element 1G, which can be viewed as a semiring† via Proposition 2.13, and use (4.1)
for the valuation axiom, since it fits in better with the semiring† approach. (But several authors, such as
Sturmfels and his school, have used the min-plus algebra instead, in order to forego taking the negative.)
Definition 4.1. The algebra of Puiseux series K over an algebraically closed field K is the field of series
of the form
p := Xτ∈T
cτ tτ ,
cτ ∈ K,
(4.2)
with T ⊂ R well-ordered (from below).
Sometimes one takes T ⊂ Q; any totally ordered field will do. We will take R in this paper. For any
field F , we write K× for K \ {0}. The tropical connection is that the max-plus algebra appears as the
target of the valuation
given by sending p(t) 6= 0K to the negative of the lowest exponent of its monomials having nonzero
coefficient;
Val : K× −→ R
Val(p) := − min{τ ∈ T : cτ 6= 0K}.
(4.3)
/
/
CATEGORICAL LAYERED MATHEMATICS
13
Remark 4.2. There is a natural multiplicative map π : K× → K×, sending a Puiseux series p =
Pτ∈T cτ tτ to cVal(p).
This gives an extra important piece of information, since for any two Puiseux series p, q we must have
Val(p + q) = max{Val(p), Val(q)} unless Val(p) = Val(q) and π(p) = π(q), in which case Val(p + q) <
max{Val(p), Val(q)}. In this way, π measures how much bipotence is lost with respect to Val .
4.1. Valued monoids. We view the previous observations in a somewhat more general setting.
Definition 4.3. A monoid M = (M,· , 1M) is m-valued with respect to a totally ordered monoid G :=
(G,· ,≥, 1G) if there is an onto monoid homomorphism v : M → G. This set-up is notated as the
triple (M,G, v).
Note 4.4. The hypothesis that v is onto can always be attained by replacing G by v(M) if necessary.
Given a field with valuation, or more generally, an integral domain W with valuation v : W \{0W} −→
G, we take M = \{0M}, a cancellative submonoid of W , to obtain the triple (M,G, v) as in Definition
4.3. When W is an arbitrary commutative ring with valuation, we must assume that the monoid G is
pointed, and take the triple (W,G, v).
Example 4.5. Another major example of an m-valued monoid is M = (C×,· , 1), G = (R≥0, +,≥, 0),
and v : C× → R≥0, given by v : z 7→ logt(z), where t is a given positive parameter. This leads us to the
theory of complex amoebas, cf. Passare [32].
The category of m-valued monoids is quite intricate, since the morphisms should include all maps
which "transmit" one m-valuation to another, as defined in [28]. In order to simplify this aspect of the
theory, we restrict ourselves to a subcategory, but consider the general version in [20].
Definition 4.6. ValMon is the category of m-valued monoids whose objects are triples (M,G, v) as in
Definition 4.3, for which a morphism
(4.4)
is comprised of a pair (φM, φG) of a monoid homomorphism φM : M → M′, as well as an order-
preserving monoid homomorphism φG : G → G′, satisfying the compatibility condition
φ : (M,G, v) −→ (M′,G′, v′)
ValMon+ is the full subcategory of ValMon in which the target monoid G is cancellative.
Thus, we have the categories ValField (resp. ValDom) whose objects are fields (resp. integral domains)
with valuations to cancellative monoids, and whose morphisms are ring homomorphisms which restrict
to morphisms in ValMon+,and each has its respectful forgetful functor to ValMon+.
Remark 4.7. If (M,G, v) is a triple, then v induces a pre-order ≤ on M, given by a ≤ b iff v(a) ≤ v(b)
in G.
Lemma 4.8. There is a fully faithful functor of categories ValMon → PreOMon .
Proof. The functor is given by Remark 4.7.
Conversely, given any monoid M with pre-order, we define G := M/≡ as in Proposition 3.7. Then we
define v : M → G by a 7→ [a]; clearly (M,G, v) is a triple. One sees easily that the morphisms match. (cid:3)
On the other hand, as we have observed, the main idea of tropicalization is the following observation:
Remark 4.9. The valuation itself provides a forgetful functor Fval : ValMon+ → OMon+, where we
remember only the target monoid G from the triple (M,G, v) .
Remark 4.10. Let us recall some valuation theory, which we can state in terms of an integral domain W
with valuation (W,G, v). The valuation ring R (resp. valuation ideal p) is the set of elements of W
having value ≥ 0 (resp. > 0); the residue domain W is R/p.
The residue domain is an integral domain. When W is a field F , the residue domain ¯F is also a field.
For example, the valuation ideal p of K of Definition 4.1 is the set of Puiseux series having value > 0, and
the residue field can be identified with K.
Here is another example, to illustrate some subtler aspects of the definitions.
v′(φM(a)) = φG(v(a)),
∀a ∈ M.
(4.5)
14
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Example 4.11.
(i) The integral domain C[λ1, λ2] has the natural valuation to Z×Z, ordered via the lexicographic order
where v(λ1) = (1, 0) and v(λ2) = (0, 1). On the other hand, there is the valuation v : C[λ1, λ2] → Z
given by v(Pi,j αi,jλi
2) = k for that smallest k = i + j such that αi,j 6= 0. (In other words,
v(λ1) = v(λ2) = 1.) The identity map (1C[λ1,λ2], 1Z×Z) (where we replace the valuation v by v)
is not a morphism in ValMon+ since it is not order-preserving. v(λ1) = (1, 0) > (0, 2) = v(λ2
2)
whereas v(λ1) = 1 < 2 = v(λ2
1λj
2).
(ii) In (i), we take a different order on Z × Z, where two pairs are ordered first by the sum of their
coordinates and then only secondarily via the lexicographic order. Now the identity map (where
we replace the valuation v by v) is a morphism in ValMon+ since it is order-preserving. Note
however that it is not strictly order-preserving, since v(λ1) = (1, 0) > (0, 1) = v(λ2) whereas
v(λ1) = 1 = v(λ2).
5. The layered structure
We are ready to bring our leading player. In this section we describe the algebraic category in whose
context we may formulate all the algebraic structure we need (including matrices and polynomials) for
the layered theory. To simplify notation and avoid technical complications, we work with a semiring† L
without a zero element, even though information is lost; the full theory is given in [20]. Much of the
layered theory stems from the following fundamental construction from [17], which is inspired by [2].
Construction 5.1. R := R(L,G) is defined set-theoretically as L×G, where we denote the element (ℓ, a)
as [ℓ]a and, for k, ℓ ∈ L, a, b ∈ G, we define multiplication componentwise, i.e.,
Addition is given by the usual rules:
[k]a ·
[ℓ]b = [kℓ]ab .
[k]a + [ℓ]b =
[k]a
[ℓ]b
[k+ℓ]a
if a > b,
if a < b,
if a = b.
(5.1)
(5.2)
We define Rℓ := {ℓ} × G, for each ℓ ∈ L. Namely R = Sℓ∈LRℓ.
This is to be our prototype of a layered bi-domain†, and should be borne in mind throughout the
sequel. Nevertheless, one should also consider the possibility that the monoid G is non-cancellative, in
which case, as noted in [17], Construction 5.1 fails to satisfy distributivity and thus is not a semiring†.
This difficulty can be resolved, but the ensuing category becomes rather technical, so we defer it to [20].
5.1. Layered pre-domains†. We axiomatize Construction 5.1 in order to place it in its categorical
framework.
Definition 5.2. Suppose (L,≥) is a partially pre-ordered semiring†. An L-layered pre-domain†
R := (R, L, (νm,ℓ)),
is a semiring† R, together with a partition {Rℓ : ℓ ∈ L} into disjoint subsets Rℓ ⊂ R, called layers, such
that
Rℓ,
(5.3)
and a family of sort transition maps
such that
for every ℓ ∈ L, and
R :=
[ℓ∈L
νm,ℓ : Rℓ → Rm,
νℓ,ℓ = idRℓ
∀m ≥ ℓ > 0,
whenever both sides are defined, satisfying the following axioms A1 -- A4, and B.
νm,ℓ ◦ νℓ,k = νm,k,
∀m ≥ ℓ ≥ k,
CATEGORICAL LAYERED MATHEMATICS
15
We say that any element a of Rk has layer k (k ∈ L). We write a ∼=ν b for b ∈ Rℓ, whenever
νm,k(a) = νm,ℓ(b) in Rm for some m ≥ k, ℓ. (This notation is used generically: we write a ∼=ν b even
when the sort transition maps νm,ℓ are notated differently.)
Similarly, in line with Remark 3.9, we write a ≤ν b if νm,k(a) + νm,ℓ(b) = νm,ℓ(b) in Rm for some
The axioms are as follows:
m ≥ k, ℓ.
A1. 1R ∈ R1.
A2. If a ∈ Rk and b ∈ Rℓ, then ab ∈ Rkℓ.
A3. The product in R is compatible with sort transition maps: Suppose
a ∈ Rk and b ∈ Rℓ, with m ≥ k and m′ ≥ ℓ. Then
νm,k(a) · νm′,ℓ(b) = νmm′,kℓ(ab).
A4. νℓ,k(a) + νℓ′,k(a) = νℓ+ℓ′,k(a) for all a ∈ Rk and all ℓ, ℓ′ ≥ k.
B. (Supertropicality) Suppose a ∈ Rk, b ∈ Rℓ, and a ∼=ν b. Then
a + b ∈ Rk+ℓ with a + b ∼=ν a.
If moreover k = ∞, then a + b = a.
L is called the sorting semiring† of the L-layered pre-domain† R =Sℓ∈L Rℓ.
For convenience, we assume in the sequel that L = L≥0, i.e., all nonzero elements of L are positive.
Often L is N or N+.
Remark 5.3. The L-layered pre-domain† R has the special layer R1, which is a multiplicative monoid,
called the monoid of tangible elements, and acts with the obvious monoid action (given by multiplica-
tion) on each layer Rk of R.
Thus, in one sense, R extends its monoid of tangible elements. Although we have given up bipotence,
and Axiom B provides us the slightly weaker notion of ν-bipotence, which says that a + b ∼=ν a or
a + b ∼=ν b for all a, b ∈ R.
Definition 5.4. An L-layered bi-domain† is a ν-bipotent L-layered pre-domain† which is (multiplica-
tively) cancellative. (We use the prefix "bi" in this paper to stress the ν-bipotence.)
An L-layered bi-domain† R is called an L-layered bi-semifield† if (R1,· ) is an Abelian group.
Note that according to this definition, an L-layered bi-semifield† need not be a semifield† unless L
itself also is a multiplicative group (and thus a semifield†). When R is an L-layered bi-semifield†, the
action of Remark 5.3 is simply transitive, in the sense that for any a, b ∈ Rℓ there is a unique element
r ∈ R1 for which ar = b.
Definition 5.5. We write R>ℓ (resp. R≥ℓ ) for Sk>ℓ Rk (resp. Sk≥ℓ Rk).
Definition 5.6. The layer R1 of an L-layered pre-domain† R is called the monoid of tangible ele-
ments.
We are interested in the case that R1 generates R.
Lemma 5.7. If M is any submonoid of a layered pre-domain† R := (R, L, (νm,ℓ)), then the additive
sub-semigroup M of R generated by M is also a layered pre-domain†.
Proof. M is a semiring†, by distributivity. Axiom A1 is given, and the other axioms follow a fortiori. (cid:3)
Definition 5.8. The tangibly generated sub-semiring† is the sub-semiring† generated by R1; if this
is R, we say that R is tangibly generated.
Remark 5.9. Several initial observations are in order.
(i) The layered structure resembles that of a graded algebra, with two major differences: On the one
hand, the condition that R is the disjoint union of its layers is considerably stronger than the
usual condition that R is the direct sum of its components; on the other hand, Axioms A4 and B
show that the layers are not quite closed under addition.
16
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
(ii) This paper is mostly about L-layered bi-domains† (in particular, L-layered bi-semifields†). How-
ever, since ν-bipotence does not hold for polynomials, one considers the more general L-layered
pre-domains† when studying polynomial semirings†.
(iii) For each ℓ ∈ L we introduce the sets
Rm
R≥ℓ := [m≥ℓ
and
R>ℓ := [m>ℓ
Rm.
Many of our current examples satisfy L = L≥1 and thus R = R≥1. When R is an L-layered
bi-domain†, we claim that R≥1 is an L≥1-layered sub-bi-domain† of R, and R≥k and R>k are
semiring† ideals of R≥1, for each k ∈ L≥1. Indeed, this is an easy verification of the axioms,
mostly from Axiom A2.
(iv) Given any L-layered bi-domain† R and any multiplicative submonoid M of R≥1, we want to
define the L-layered sub-bi-domain† of R generated by M. First we take
M′ := {νℓ,k(a) : ℓ ≥ k ∈ L, a ∈ M ∩ Rk},
which is a submonoid closed under the transition maps. Then we take
M′′ := M′ ∪ {a + b : a, b ∈ M′ with a ∼=ν b}.
This is closed under all the relevant operations, so is the desired L-layered bi-domain†. Note that
the second stage is unnecessary for a = b, in view of Axiom A4.
(v) Although ubiquitous in the definition, the sort transition maps get in the way of computations,
and it is convenient to define the elements
eℓ := νℓ,1(1R)
(ℓ ≥ 1).
(5.4)
If a ∈ Rk, ℓ ∈ L, and ℓ ≥ 1, we conclude by Axiom A3 that
νℓ·k,k(a) = νℓ·k,1·k(a · 1R) = νℓ,1(1R) · νk,k(a) = νℓ,1(1R) · a = eℓa.
Thus the sort transition map νℓ·k,k means multiplication by eℓ.
Note that ek + eℓ = ek+ℓ by Axiom A4.
The element eℓ is a (multiplicative) idempotent of R iff ℓ2 = ℓ in L. In particular, e1 and e∞
(when ∞ ∈ L) are idempotents of R.
Let us introduce the sorting map s : R → L, which sends every element a ∈ Rℓ to its sort index ℓ,
and we view the semiring† R as an object fibered by s over the sorting semiring† L.
Remark 5.10. Axioms A1 and A2 yield the conditions
s(1R) = 1L,
s(ab) = s(a) s(b),
∀a, b ∈ R.
(5.5)
Also, Axiom A4 yields s(a + a) = s(a) + s(a) = 2s(a), thereby motivating us to view addition of an
element with itself as doubling the layer. Applying ν-bipotence to Axiom B shows that
s(a + b) ∈ {s(a), s(b), s(a) + s(b)}.
To emphasize the sorting map, as well as the order on L, we sometimes write (R, L, s, P, (νm,ℓ)) for a
given L-layered bi-domain† R with sort transition maps (νm,ℓ : m ≥ ℓ) and their accompanying sorting
map s : R → L.
5.2. Uniform L-Layered bi-domains†.
sort transition maps νm,ℓ are all bijective.
There are two main examples coming from tropical mathematics.
(a) Let R = R(L,G) (corresponding to the "naive" tropical geometry). By Construction 5.1, the
(b) Suppose K is the field of Puiseux series {f := Pu∈Q αuλu : f has well-ordered support} over a
given field F . Then we have the m-valuation v : K → Q taking any Puiseux series f to the lowest
real number u in its support.
We incorporate K into the structure of R, by putting Rℓ to be a copy of K for ℓ ≤ 1 and Rℓ
to be a copy of G for ℓ > 1. We take the νm,ℓ to be v whenever m > 1 ≥ ℓ, and the identity
otherwise. In this way, v is preserved within the structure of R.
CATEGORICAL LAYERED MATHEMATICS
17
We focus on the first case, since one can reduce to it anyway via the equivalence given below in
Definition 5.19 (which takes us from the usual algebraic world to the tropical world).
Definition 5.11. An L-layered pre-domain† R is uniform if νℓ,k is 1:1 and onto for each ℓ > k.
Example 5.12. Example 5.1 is a uniform L-layered bi-domain†, when the monoid G is cancellative.
Let us see how the layered theory simplifies for uniform L-layered bi-domains†, enabling us to remove
the sort transition maps νℓ,k from the picture.
Lemma 5.13. Any element a ∈ Rℓ can be written uniquely as eℓa1 = νℓ,1(a1) for a1 ∈ R1.
Proof. Existence and uniqueness of a1 are clear since νℓ,1 is presumed to be 1:1. The last assertion follows
from Axiom A3.
(cid:3)
Proposition 5.14. In a uniform L-layered bi-domain†, if a ∼=ν b for a ∈ Rk and b ∈ Rℓ with ℓ ≥ k then
b = νℓ,k(a). In particular, if a ∼=ν b for a, b ∈ Rℓ, then a = b.
Proof. An immediate application of Lemma 5.13.
Now we can remove the sort transition maps from the definition, when we write R =Sℓ∈L eℓR1.
Proposition 5.15. If (L,· ) is a multiplicative group, then one has νℓ,k(ak) ∼=ν emak where ℓ = mk.
In a uniform L-layered bi-domain† (for L arbitrary), the transition map νm,ℓ is given by eℓa1 7→ ema1.
Proof. If ℓ = k there is nothing to prove, so we assume that ℓ > k are non-negative, and write ℓ = k + p
for p ∈ L. Then ℓ = k(1 + pk−1), and m = 1 + pk−1. Now emak ∈ Rℓ, and
(cid:3)
emak ∼=ν ak ∼=ν νℓ,k(ak).
The second assertion now is clear.
(cid:3)
Thus the sort transition maps have been replaced by multiplication by the em. Note that ν-bipotence
and Axiom B′ could then be used as the definition for addition in R, and we summarize our reductions:
Proposition 5.16. A uniform L-layered bi-domain† can be described as the semiring†
R :=
Rℓ,
[ℓ∈L
where each Rℓ = eℓR1, (R1,· ) is a monoid, and there is a 1:1 correspondence R1 → Rℓ given by a 7→ eℓa
for each a ∈ R1.
Proposition 5.17. In the uniform case, axioms A2 and A3 can be replaced by the respective axioms:
A2′. If a = eka1 ∈ Rk and b = eℓb1 ∈ Rℓ, for a1, b1 ∈ R1, then
A3′. eℓek = eℓk for all k, ℓ ∈ L.
ab = (a1b1)ekℓ.
Furthermore, Axiom A4 now is equivalent to Axiom B, which we can reformulate as:
B′. If a = eka1 and b = eℓa1 (so that a ∼=ν b), then a + b = ek+ℓa1.
The operations in R are given by Axioms A2 ′, A3 ′, B ′, and ν-bipotence.
Proof. Axiom A3 follows from the observation that eℓa1ekb1 = eℓk(a1b1); when a1, b1 ∈ R1 then a1b1 ∈
R1.
(cid:3)
5.3. Reduction to the uniform case. In one sense, we can reduce the general case of an L-layered
pre-domain† R to the uniform case. First we cut down on superfluous elements. Note that if νk,1 are
onto for all k ≥ 1, then all the νℓ,k are onto for all ℓ ≥ k. Indeed, if a ∈ Rℓ then writing a = νℓ,1(a1) we
have
a = νℓ,k(νk,1(a1)).
Remark 5.18. Suppose L = L≥1. For any L-layered pre-domain† R := (R, L, (νm,ℓ)), if we replace Rk
by νk,1(R1) for each k ∈ L, we get an L-layered bi-domain† for which all the νℓ,k are onto.
18
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Having reduced many situations to the case for which all the νℓ,k are onto, we can get a uniform
L-layered pre-domain† by specifying when two elements are "interchangeable" in the algebraic structure.
Definition 5.19. Define the equivalence relation
a ≡ b
when
s(a) = s(b) and a ∼=ν b.
In view of Proposition 5.14, this relation is trivial in case R is a uniform L-layered bi-domain†.
Proposition 5.20. The binary relation <ν on an L-layered pre-domain† R induces a pre-order on the
semiring† of equivalence classes R/≡. Furthermore, if a ≡ b, then ac ≡ bc and a + c ≡ b + c for all c ∈ R.
Thus, ≡ is a congruence.
Proof. The first assertion is immediate. For the second assertion, s(ac) = s(a)s(c) = s(b)s(c) = s(bc)
and ac ∼=ν bc, proving ac ≡ bc.
Next, we consider addition. If a >ν c, then
If a <ν c, then a + c = c = b + c. If a ∼=ν c, then
a + c = a ≡ b = b + c.
s(a + c) = s(a) + s(c) = s(b) + s(c) = s(b + c),
and a + c ∼=ν a ∼=ν b ∼=ν b + c.
Let us summarize.
(cid:3)
Corollary 5.21. When the transition maps νℓ,k are onto, one can reduce to uniform L-layered bi-
domains†, by means of the equivalence relation ≡ of Definition 5.19.
Proof. Any ν-equivalent elements having the same sort are identified. Then Proposition 5.20 shows that
R/≡ is an L-layered bi-domain†, under the natural induced layering, and the transition maps on R/≡
clearly are bijective.
(cid:3)
5.3.1. Ghosts and the surpassing relation. We want a layered version of ghosts.
Definition 5.22. An ℓ-ghost sort is an element ℓ + k ∈ L for some 0 6= k ∈ L. An element b ∈ R is
an ℓ-ghost (for given ℓ ∈ L) if s(b) ∈ L is an ℓ-ghost sort. A ghost element of R is a 1-ghost.
Thus, the relation (≥) on L satisfies
m ≥ ℓ when
Definition 5.23. The L-surpassing relation L
m = ℓ
or
m is an ℓ-ghost sort.
= is given by
Here is a key relation in the theory, even though it does not play a major role in this discussion.
= b
a L
iff either
a = b + c
a = b,
a ∼=ν b
with c an s(b)-ghost,
with a an s(b)-ghost.
(5.6)
It follows that if a L
Remark 5.24. If a L
operations.
= b, then a + b is s(b)-ghost. When a 6= b, this means a ≥ν b and a is an s(b)-ghost.
= respects the semiring†
= bc. Thus L
= b, then clearly a + c L
= b + c and ac L
Remark 5.25. If a >ν b, then (a + b)m = am. Hence, the Frobenius property (a + b)m = am + bm is
satisfied in an L-layered pre-domain† whenever a 6∼=ν b. We always have (a + b)m L
= am + bm.
CATEGORICAL LAYERED MATHEMATICS
19
5.4. Layered homomorphisms. In line with the philosophy of this paper, we would like to introduce
the category of L-layered pre-domains†. This entails finding the correct definition of morphism. We start
with the natural definition from the context of domains†. Although this definition is good enough for
the purposes of this paper, a more sophisticated analysis would require us to to consider the notion of
"supervaluation" from [16], and how this relates to morphisms that preserve the layers. Here we take
the morphisms in this category to be semiring† homomorphisms which respect the order on the sorting
semiring L:
Definition 5.26. A layered homomorphism of L-layered pre-domains† is a map
(5.7)
such that ρ : L → L′ is a semiring† homomorphism, together with a semiring† homomorphism ϕ : R → R′
such that
Φ := (ϕ, ρ) : (R, L, s, (νm,ℓ)) → (R′, L′, s′, (ν′m′,ℓ′))
M1. If ϕ(a) /∈ R′0, then s′(ϕ(a)) ≥ ρ(s(a)).
M2. If a ∼=ν b, then ϕ(a) ∼=ν ϕ(b). (This is taken in the context of the
ν′m′,ℓ′.)
We always denote Φ = (ϕ, ρ) as Φ : R → R′ when unambiguous. In most of the following examples,
the sorting semirings† L and L′ are the same. Accordingly, we call the layered homomorphism Φ an
L-homomorphism when L = L′ and ρ = 1L.
Proposition 5.27. Any layered homomorphism ϕ preserves ν, in the following sense:
If a ≥ν b, then ϕ(a) ≥ν′ ϕ(b).
Proof. ϕ(a) ∼=ν′ ϕ(a + b) = ϕ(a) + ϕ(b), implying ϕ(a) ≥ν′ ϕ(b).
Proposition 5.28. Suppose ϕ : R → R′ is a layered homomorphism, and R is tangibly generated
(cf. Definition 5.8). Then ϕ is determined by its restriction to R1, via the formula
(cid:3)
ϕ(a + b) = ϕ(a) + ϕ(b),
∀a, b ∈ R1.
Proof. It is enough to check sums, in view of Lemma 5.7. We get the action of ϕ on all of R since R1
generates R.
(cid:3)
Remark 5.29. The definition given here of layered homomorphism is too strict for some applications.
One can weaken the definition of layered homomorphism by utilizing the surpassing relation, requiring
merely that ϕ(a) + ϕ(b) L
= ϕ(a + b), but various technical difficulties arise, so we defer the study of this
category to [20].
Before continuing, let us see how this definition encompasses various prior tropical situations.
Example 5.30. We assume throughout that R is an L-layered pre-domain†.
(i) In the max-plus situation, when L = {1}, ρ must be the identity, and Φ is just a semiring†
homomorphism.
(ii) In the "standard supertropical situation," when L = {1,∞}, Φ must send the ghost layer R∞
to R∞. If a ⊳ R, one could take R′1 := R \ a and R′
∞ := R∞ ∪ a. The identity map is clearly a
layered homomorphism; its application "expands the ghost ideal" to a, thereby taking the place of
a semiring homomorphism to the factor semiring†.
(iii) Generalizing (ii), we obtain layered homomorphisms by modifying the layering. We say a re-
sorting map of a uniform L-layered pre-domain† R is a map s′ : R → L satisfying the following
properties:
(a) s′(1R) = 1,
(b) s′(R1) ⊆ L≥1,
(c) s′(ab) = s′(a)s′(b), ∀a, b ∈ R1,
20
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
(d) s′(eℓa) = ℓs′(a), ∀a ∈ R1.
Then the following properties also are satisfied:
(a) s′(eℓ) = ℓ for all ℓ ∈ L.
(b) s′(ab) ≥ s′(a)s′(b) for all a, b ∈ R.
(c) s′(a) ≥ s(a), ∀a ∈ R.
To see this, take a ∈ Rk and b ∈ Rℓ, and write a = eka1 and b = eℓb1 for a1, b1 ∈ R1. Then
s′(eℓ) = s′(eℓ1R) = ℓs′(1R) = ℓ · 1 = ℓ
Taking c1 = a1b1 ∈ R1, we have
s′(ab) = s′(ekeℓc1) = kℓs′(c1) = kℓs′(a1)s′(b1) = s′(a)s′(b),
and s′(a) = s′(eka1) = ks′(a1) ≥ k.
phisms.
(iv) The natural injections R≥1 → R and {Sℓ Rℓ : ℓ ∈ N} → R are all examples of layered homomor-
(v) The L-truncation map of [17, §3] is a layered homomorphism.
(vi) Suppose R is a layered pre-domain†. We adjoin ∞ to L, and take R∞ and ν∞,k to be the direct
limit of the Rk and νℓ,k, and write ν for the various ν∞,k. An element in a ∈ R1 is ν-non-
cancellative if ab ∼=ν ac for suitable b, c, where b 6∼=ν c. We define the map ϕ : R → R which is
the identity on ν-cancellative elements but ϕ(a) = aν for all ν-non-cancellative elements a ∈ R.
In particular, ϕ(R)1 is comprised precisely of the ν-cancellative tangible elements.
We claim that ϕ is a homomorphism. If ab is ν-cancellative this is clear, so we may assume
that a is ν-non-cancellative. Then
ϕ(ab) = (ab)ν = aνϕ(b) = ϕ(a)ϕ(b).
Certainly ϕ(a + b) = ϕ(a) + ϕ(b) by bipotence unless a ∼=ν b, in which case
ϕ(a + b) = ϕ(aν ) = aν = ϕ(a) + ϕ(b).
Furthermore, ϕ(R) is a layered pre-domain† which is ν-cancellative with respect to ϕ(R)1.
Indeed, if ϕ(a)ϕ(b) ∼=ν ϕ(a)ϕ(c) with ϕ(a) ∈ ϕ(R)1, then ϕ(b) ∼=ν ϕ(c).
Note that this is not the same example used in [19].
Our main example for future use is to be given in Remark 7.10.
6. The layered categories and the corresponding tropicalization functors
Having assembled the basic concepts, we are finally ready for the layered tropical categories. Our
objective in this section is to introduce the functor that passes from the "classical algebraic world" of
integral domains with valuation to the "layered world," taking the cue from [22, Definition 2.1], which
we recall and restate more formally.
Here are our first main layered categories, starting with the more encompassing and proceeding to the
specific. In each case the morphisms are the relevant layered homomorphisms.
Definition 6.1.
(a) LayPreD† is the category whose objects are layered pre-domains†.
(b) LayBidom† is the full subcategory of LayPreD† whose objects are layered bi-domains†.
(c) ULayBidom† is the full subcategory of LayPreD† whose objects are uniform layered bi-domains†.
CATEGORICAL LAYERED MATHEMATICS
21
6.1. Identifications of categories of monoids and layered pre-domains†.
Remark 6.2. We define the forgetful functor ULayBidom† → OMon+ given by sending any uniform
L-layered bi-domain† R := R(L,G) to R1.
We want retracts for this forgetful functor. By Proposition 5.28, any layered homomorphism cor-
responds to a homomorphism of the underlying monoid of tangible elements, thereby indicating an
identification between categories arising from the construction of layered bi-domains† from pre-ordered
monoids.
Theorem 6.3. There is a faithful layering functor
Flay : OMon+ −→ ULayBidom†,
given by sending G to R(L,G), and sending the ordered homomorphism ϕ : G → G′ to the layered
homomorphism R(L,G) → R(L,G′) induced by ϕ. The functor Flay is a left retract of the forgetful
functor of Remark 6.2.
Proof. The image of a cancellative ordered monoid G is a layered bi-domain†, in view of [17, Proposi-
tion 2.3], and one sees easily that Flayϕ is a layered morphism since, for a ≥ν b,
Flayϕ( [k]a + [ℓ]b ) ∼=ν Flayϕ( [k]a ) ∼=ν ϕ( [k]a ) + ϕ( [ℓ]b ),
and s′(Flayϕ( [k]a + [ℓ]b )) ≥ k.
morphisms by applying the forgetful functor of Remark 6.2.
Also, the morphisms match. The functor Flay is faithful, since one recovers the original objects and
(cid:3)
More subtly, at times we want to forget the order on our monoids, to apply the theory of [8]. Even so,
we have a universal construction with respect to "universal characteristic."
Example 6.4. Given a cancellative monoid G and a partially ordered semiring L, define the semiring†
UL(G) as follows:
Each element of UL(G) is a formal sum of elements of G, each supplied with its layer, i.e., has the
form
Addition is given by the rule
(cid:26)Xa∈S
[ℓa]a + Xa∈S′
[ℓa]a : S ⊂ G(cid:27).
a]a = Xa∈S∪S′
[ℓ′
Xa∈S
[ℓa+ℓ′
a]a .
(6.1)
(6.2)
Here we formally define ℓa + ℓ′a to be ℓa (resp. ℓ′a) if a /∈ S′ (resp. if a /∈ S).
Multiplication is given by
extended via distributivity.
[k]a ·
[ℓ]b = [kℓ](ab) ,
We want UL(G) to be the universal to the forgetful functor of Corollary 6.3. This is "almost" true,
with a slight hitch arising from (6.1).
Proposition 6.5. Given any monoid bijection ϕ : M → G where G := (G,· ,≤, 1G) is a totally ordered
monoid, viewed as a bipotent semiring† as in Proposition 3.11, there is a natural homomorphism
given by
The composite
bϕ : UL(M) → R(L,G)
bϕ(cid:18)Xa∈S
[ℓa]a(cid:19) =Xa∈S
[ℓa]ϕ(a1) .
M
ϕ
/ G
/ R(L,G).
/
/
22
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
also factors naturally as
M
/ (UL(M),· )
bϕ
/ R(L,G).
In case L = {1} (the max-plus setting) or L = {1,∞} (the standard supertropical setting), the previous
assertion holds more generally for any monoid homomorphism ϕ : M → G.
Proof. The multiplication rules match, so the verifications follow formally, cf. Remark 5.25. The last
assertion is true because in these particular situations the Frobenius property is an identity, holding even
when a ∼=ν b.
(cid:3)
Remark 6.6. Since the Frobenius property is an identity, one could just mod it out from our construction
of UL(M) utilizing Remark 2.11, and thus get a universal with respect to satisfying the Frobenius property.
6.2. The layered tropicalization functor. Having our categories in place, we can get to the heart of
tropicalization.
Definition 6.7. Given a semiring† L, the L-tropicalization functor
FLTrop : ValMon+ −→ ULayBidom†
from the category of valued monoids (with cancellation in the target) to the category of uniform layered
bi-domains† is defined as follows: FLTrop : (M,G, v) 7→ R(L,G) and FLTrop : φ 7→ αφ, where given a
morphism φ : (M,G, v) → (M′,G′, v′) we define αφ : R(L,G) → R(L,G′), by
cf. Formula (4.4).
αφ( [ℓ]a ) := [ℓ]φ(a) ,
a ∈ G,
We also consider FLTrop as acting on individual elements of M, whereby
FLTrop(a) = [1]v(a) .
(6.3)
(6.4)
This is indeed a functor, in view of [17, Theorem 4.9].
Note that the tropicalization functor FLTrop factors as ValMon+ → OMon+ → ULayBidom†.
6.3. More comprehensive layered tropicalization functors. The basic layered tropicalization func-
tor only recognizes the image in G, so loses much information about the original monoid M. In analogy
to [31], in order to preserve information, we can encode extra information, motivated by the residue field
in valuation theory.
6.3.1. The unit tropicalization functor.
Definition 6.8. Given a monoid M := (M,· , 1M) with m-valuation v : M → G, we define its unit
submonoid
the submonoid of M on which the restriction of v is the trivial valuation.
M(1) := {a ∈ M : v(a) = 1G},
When M is a group, then M(1) also is a group. In particular, the category ValMon(1) of unit monoids
with m-valuation is a full subcategory of the category ValMon.
Example 6.9. In Example 4.5, C(1) is the complex unit circle.
The following observation is now clear.
Proposition 6.10. There is a functor
FLTrop;unit : ValMon+ −→ ULayBidom† ×ValMon(1),
given as follows: FLTrop;unit((M,G, v)) = (R(L,G),M(1)) and FLTrop;unit(φ) = (αφ, φM(1) ), where the
morphism αφ : R(L,G) → R(L′,G′) is given by Equation (6.3).
Proof. We piece together the two functors.
(cid:3)
/
/
CATEGORICAL LAYERED MATHEMATICS
23
This functor could be interpreted as separating the m-valuation v into two components, corresponding
to the value monoid and the residue domain. Tropicalization in its original form involved taking the
logarithm of the absolute value of reiθ, which is just log r. Thus, the argument eiθ is lost, and researchers
dealt with that separately. Since these all have absolute value 1, it seems appropriate in the valuation-
theory analog to have FLTrop;unit at our disposal.
A more direct approach in the terminology of Remark 4.2: Given two Puiseux series p, q ∈ K with
Val(p) = Val(q), we see that V al(p) = V al(q) iff V al(pq−1) = V al(1) = 1K, i.e., pq−1 − 1 is in the
valuation ideal of the valuation Val. Thus, Proposition 6.10 gives us a way of understanding FLTrop;unit
in terms of Val. Namely, we check whether two Puiseux series have the same lowest order exponent, and
then can check whether their lowest coefficients are the same by means of the residue field.
Remark 6.11. Suppose W is an arbitrary integral domain with valuation v : W \ {0W} → G, with
valuation ring R and residue domain ¯W . Take the unit submonoid W1 of W , cf. Definition 6.8. Clearly
W1 = {r ∈ R : r + p = 1 + p}.
When W is a field, W1 is a multiplicative subgroup of W which could be thought of as the "first
congruence subgroup" in valuation theory. Then, for b 6= 0, aW1 = bW1 iff v(a) = v(b) and 1 − ab−1 ∈ p,
which relates to the condition of the previous paragraph.
6.3.2. The exploded tropicalization functor. One could preserve more information, according to Parker [31],
who introduced "exploded" tropical mathematics, and Payne [36]. This entails taking the leading coeffi-
cient of Puiseux series.
E. Sheiner introduced a related structure R(K, R) on Puiseux series, in which he uses the residue
field K as the sorting set. Define the map K → R(K, R) by p 7→ [α]v(p) where α is the coefficient of the
lowest monomial of the Puiseux series p. This map, generalizing the Kapranov map, keeps track of the
"leading coefficient" of the Puiseux series p in terms of when the image of p has layer 0.
From this perspective, the 0K layer represents the "corner ghosts." Thus, Sheiner has "exploded" the
notion of valuation, and it is not difficult to define the "exploded functor" and transfer the statement
and proof of Payne [36] to this context, to be indicated in §8. Let us describe this procedure in algebraic
terms, which means working in the associated graded algebra.
Definition 6.12. Given a valued monoid v : M → G, and g ∈ G, we write M≥g for the M-module
{a ∈ M : v(a) ≥ g}, and M>g for its submodule {a ∈ M : v(a) > g}.
When M is the multiplicative monoid of an integral domain W , we can define the associated graded
algebra
gr(W ) :=Ma∈G
W≥g/W>g,
where operations are given by
(a + W>g)(b + W>h) = ab + W>gh
(a + W>g) + (b + W>h) = a + b + W>gh.
Remark 6.13. It is well known that the associated graded algebra is an algebra, with the natural valua-
tion v induced by v, i.e., v(a+ W>g) = v(a). When the valuation v is discrete, each component W≥g/W>g
is (multiplicatively) isomorphic to ¯W .
Let us interpret "explosion" with respect to Puiseux series. For any real number α, the component
K≥α/K>α can be identified with Ktα, which as a module is isomorphic to K, by means of taking the
coefficient of the monomial of lowest order in a Puiseux series.
Definition 6.14. Notation as in Remark 6.11, define the exploded layered domain† R( ¯W ,G). In other
words, we sort the elements according to ¯W , with multiplication following the given multiplication in G
and addition given by the following rules:
[a]x + [b]y =
[a]x
[b]y
[a+b]x
if x > y,
if x < y,
if x = y.
(6.5)
24
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Remark 6.15. Note that addition here is the classical addition induced from the integral domain W , so
although this structure has a tropical aroma, it does preserve some of the original algebraic structure of
the residue domain ¯W .
Proposition 6.16. There is a functor
FLTrop;exp : ValDom −→ ULayBidom† × Ring,
given as follows: FLTrop;exp((W,G, v)) = (R(L,G), ¯W ) and FLTrop;exp(φ) = (αφ, φ), where the morphism
αφ : R(L,G) → R(L′,G′) is given by Equation (6.3) and φ is the induced map on the residue domains.
Proof. As in Proposition 6.10, we piece together the two functors.
(cid:3)
To preserve even more information, one could sort instead with gr(W ).
7. The function category
We assume throughout that R is an L-layered domain†. In the next section we describe layered varieties
in terms of corner roots of ideals of polynomials over R. Thus, we need some preliminaries about the
polynomial semiring† over a layered bi-domain†; this is no longer bipotent.
7.1. The layered function monoid and domain†. As noted in the introduction, one significant
difference between the tropical theory and "classical" algebra is that different tropical polynomials can
agree as functions (whereas for algebras over an infinite field, any two distinct polynomials are also
distinct as functions). The clearest way of coping with this phenomenon is to treat polynomials directly
as functions from some subset of R(n) to an extension of R, and this enables us to unify various other
constructions related to polynomials.
Definition 7.1. For any set S and monoid M, Fun(S,M) denotes the set of functions from S to M.
Remark 7.2.
(i) Fun(S,M) becomes a monoid, under pointwise multiplication, i.e.,
(f g)(a) = f (a)g(a),
∀f, g ∈ Fun(S, R),
∀a ∈ S.
(7.1)
(ii) If the monoid M is partially ordered, then Fun(S,M) is also partially ordered, with respect to
taking f ≥ g when f (a) ≥ g(a) for all a ∈ S.
When moreover R is a semiring†, Fun(S, R) also becomes a semiring†, under pointwise addition, i.e.,
(f + g)(a) = f (a) + g(a),
∀a ∈ S,
(7.2)
cf. [17, Definition 5.1].
Lemma 7.3.
(i) If a monoid M is cancellative, then the function monoid Fun(S,M) is cancellative.
(ii) If a layered pre-domain† R is cancellative, then the function semiring† Fun(S, R) is also a can-
cellative layered pre-domain† (but not bipotent!).
(iii) If a semiring† R satisfies the Frobenius property (3.6), then Fun(S, R) also satisfies the Frobenius
property.
Proof. (i): By pointwise verification. For cancellation, note that if f g = f h, then f (a)g(a) = f (a)h(a)
for all a ∈ S, implying g(a) = h(a) and thus g = h.
(ii): Same verification as in (i).
(iii): For the Frobenius property,
(f + g)n(a) = ((f + g)(a))n = f (a)n + g(a)n,
for all positive n ∈ N.
∀f, g ∈ Fun(S, R),
(7.3)
(cid:3)
There is a natural semiring† injection R → Fun(S, R), given by viewing r ∈ R as the constant
function fr given by fr(a) = r, ∀a ∈ S. In this way, we view R as a sub-semiring† of Fun(S, R). At first,
we take S to be R(n). Later we will take S to be a given "layered variety." More generally, following
Payne [35, §2.2], one could take the set S to be the lattice of characters of an algebraic torus.
CATEGORICAL LAYERED MATHEMATICS
25
7.2. Functorial properties of the function monoid and semiring†. We categorize the discussion
of Section 7.1. First we define the function and polynomial categories.
Definition 7.4. F := FunMon(S,
) is the functor from Mon to Mon given by M 7→ Fun(S,M) for
objects, and such that for any morphism ϕ : M → M′, we define F ϕ : Fun(S,M) → Fun(S,M′) to be
given by
The functor F := FunSemir† (S,
) : Semir† → Semir† is given by R 7→ Fun(S, R) for objects, and
again such that for any morphism ϕ : R → R′, F ϕ : Fun(S, R) → Fun(S, R′) is given by
F ϕ(f )(a) = ϕ(f (a)).
F ϕ(f )(a) = ϕ(f (a)).
Lemma 7.5. FunMon(S,
a functor from OMon+ to POMon+.
) and FunSemir† (S,
) are functors. Furthermore, FunMon(S,
) restricts to
Proof. The verifications are straightforward, in view of Remark 7.2(ii) and Lemma 7.3.
Definition 7.6. We denote the respective images of OMon+ and Semir† under the functors FunMon(S,
and FunSemir† (S,
and Semir†.
)
)as Fun(S, OMon+) and Fun(S, Semir†), which are respective subcategories of POMon+
(cid:3)
Now Proposition 3.11 says:
Proposition 7.7. There is a faithful functor
F(S,OMon+) : Fun(S, OMon+) −→ Fun(S, Semir†),
induced by the functor FOMon of Proposition 3.11, as described in the proof.
Proof. We define F(S,OMon)(Fun(S,M)) = Fun(S,M) (viewing M as a semiring) and, for any monoid
homomorphism ϕ : M → M′, F(S,OMon)(ϕ) : f 7→ ϕ ◦ f. This is clearly a functor, and is faithful since M
is embedded into Fun(S,M).
Proposition 7.8. The functors F(S,OMon)(S,
tion 3.11, in the sense that
) commute with FOMon of Proposi-
) and FunSemir† (S,
(cid:3)
F(S,OMon) FunMon(S,
) = FunSemir† (S,
) FunOMon .
Proof. Letting R be the semiring† of Proposition 3.11, we have F(S,OMon)FMon(S, R)(M) = Fun(S, R) =
FunSemir† (S, R) FunOMon(M).
Lemma 7.9. Construction 5.1 is functorial, in the sense that
(cid:3)
Fun(S, R(L,G)) ≈ R(Fun(S, L), Fun(S,G)).
Proof. Any f ∈ Fun(S, R(L,G)) is given by f (a) = [k]b for suitable k ∈ L and b ∈ G; we define fL and
fG by fL(a) = k and fG(a) = b. Now f 7→ [fL]fG defines a semiring† homomorphism
Conversely, given fL ∈ Fun(S, L) and fG ∈ Fun(S,G) we define f ∈ Fun(S, R(L,G)) by putting
Fun(S, R(L,G)) → R(Fun(S, L), Fun(S,G)).
f (a) = [fL(a)]fG(a) .
One sees that the sorts are preserved.
(cid:3)
7.3. Sorting the function semiring†.
Remark 7.10. If S′ ⊆ S, there is a natural semiring† homomorphism
Fun(S, R) → Fun(S′, R)
given by f 7→ fS ′ . In particular, for S′ = {a}, we have the evaluation homomorphism at a.
One main interest in the layered theory is the nature of these homomorphisms. To understand them,
we need to introduce the appropriate sorting function.
26
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Remark 7.11. When L is a partially ordered semiring†, Fun(S, L) is also a semiring† (whose unit
element is the constant function 1), which by Remark 7.2(iii) is partially ordered by the relation:
f ≤S g
if
f (a) ≤ g(a)
for all a ∈ S.
When L is directed from above, this partial order also is directed from above, since f (a), g(a) ∈ L are
bounded by max{f (a), g(a)}.
If R is L-layered, then Fun(S, R) inherits the layered structure from R pointwise with respect to
Fun(S, L), in the following sense taken from [17, Remark 5.3]:
Definition 7.12. The L-layering map of a function f ∈ Fun(S, R) is the map ϑf : S → L given by
In the layered theory, we only consider functions that are ν-compatible, in the sense that if a ∼=ν a′,
For a set I ⊂ Fun(S, R) we define
ϑf (a) := s(f (a)),
a ∈ S.
ϑI(a) := min{ϑf (a) : f ∈ I}.
then f (a) ∼=ν f (a′).
Example 7.13. ϑ{1R} is the given sorting map on R.
Example 7.14. Take R = R(N, R). Assume that S = R(2)
logarithmic notation; e.g., 1 := 0 is the multiplicative unit, and 2 · 3 = 5.
1 + λ2 + 0 for k ∈ N, and a = (a1, a2) ∈ S.
(i) Take fk = λk
ϑfk (a) =
3
2
1
for a1 = a2 = 0;
for a1 = 0 > a2
otherwise.
or
a1 = 0 > a2
or
ak
1 = a2 > 0;
1 = R(2). The examples are written in
(ii) Take I = {fk : k ∈ N}, a = (a1, a2) ∈ S. In view of (i),
ϑI (a) =
3
2
1
for a1 = a2 = 0;
for a1 = 0 > a2
otherwise.
or
a1 = 0 > a2;
Thus, the 2-layer is the union of two perpendicular rays.
(iii) Take I = {λ1 +2, λ1 +3}. The layering map ϑI restricted to the tangible elements is identically 1,
the same as that of a tangible constant, although the ideal generated by I does not contain any
constants.
Nevertheless, we can distinguish between I and tangible constants, by assuming that S contains
elements of R having layer > 1. For example, ϑI( [2]4 ) = 2 whereas ϑf for a tangible constant
function f is identically 1.
sending f 7→ ϑf .
As noted in [17], we layer the semiring† Fun(S, R) with respect to the sorting semiring† Fun(S, L), by
Given f, g ∈ Fun(S, L), write k = ϑf and ℓ = ϑg. When ℓ > k we define the transition map
by
νℓ,k : Fun(S, R)k → Fun(S, R)ℓ
νℓ,k(f ) : a → νℓ(a),k(a)(f (a)),
∀a ∈ S.
Lemma 7.15. If R is a layered pre-domain† with partial pre-order ≥ν, then we can extend ∼=ν and ≥ν
respectively to an equivalence and a partial pre-order on Fun(S, R) as follows:
(i) f ∼=ν g iff f (a) ∼=ν g(a),
(ii) f ≥ν g iff f (a) ≥ν g(a),
∀a ∈ S;
∀a ∈ S.
Proof. An easy point-by-point verification.
We usually start with a given layered domain† R, and then apply Lemma 7.15. This rather general
framework encompasses some very useful concepts, including polynomials, Laurent polynomials, etc.
(cid:3)
CATEGORICAL LAYERED MATHEMATICS
27
7.4. Polynomials. We want to understand tropical algebraic geometry in terms of roots of polynomials.
Specifically, we work in the sub-semiring† of Fun(S, R)≥1 (for S ⊆ R(n)) defined by formulas in the
elementary language under consideration, which we call polynomial functions. Thus, in the usual
language of semirings†, R[Λ] := R[λ1, . . . , λn] denotes the usual polynomials, whose image in Fun(S, R)≥1
we denote as Pol(S, R). If we adjoin the symbol −1 (for multiplicative inverse), then Laur(S, R) denotes
the image of the Laurent polynomials R[Λ, Λ−1] := R[λ1, λ−1
n ]. If our language includes the
symbol m√ , i.e., if we are working over a divisible monoid, then we would consider polynomials with
1
rational powers, which are well-defined in view of Equation (3.6); although this case is important, we do
not treat it here explicitly because of the extra notation involved.
. . . , λn, λ−1
Thus, we are working in the full subcategories Pol(S, OMon+) and Laur(S, OMon+) of Fun(S, OMon+),
and the full subcategories Pol(S, Semir†) and Laur(S, Semir†) of Fun(S, Semir†). The functor of Proposi-
tion 7.7 restricts to faithful functors Pol(S, OMon) → Pol(S, Semir†) and Laur(S, OMon) → Laur(S, Semir†).
The difficulty with treating polynomials (as well as Laurent polynomials) as functions could be that
two polynomial functions may agree on R but differ on some extension semiring† of R. Fortunately,
semifield† of fractions of R, if two polynomial functions differ on some extension of R, then they already
in [17, Theorem 5.33 and Corollary 5.34] we saw that taking eR to be the 1-divisible closure of the bi-
differ on eR. Thus, it suffices to look at Pol(S, eR) and Laur(S, eR). Strictly speaking, this was proved only
for the specific construction used in [17], so to work with layered bi-domains† we need to generalize the
construction of 1-divisible closure to L-layered bi-domains†.
Example 7.16 (1-localization). If R is an L-layered bi-domain†, then taking any multiplicative sub-
monoid S of R1, we can form the localization S−1R as a monoid, and define addition via
a
u
+
b
v
=
av + bu
uv
u ) = s(a). There is a
, which is injective since R1 is cancellative.
Taking S = R1, we call S−1R the L-layered bi-semifield† of fractions of R; this construction shows
for a, b ∈ R, u, v ∈ S. S−1R becomes an L-layered bi-domain† when we define s( a
natural layered homomorphism R → S−1R given by a 7→ a
that any uniform L-layered bi-domain† can be embedded into a uniform L-layered bi-semifield†.
Example 7.17 (ν-divisible closure). We say that an L-layered bi-domain† R is ν-divisible if for each
a ∈ R and n ∈ N there is b ∈ R such that bn ≡ a under the equivalence of Definition 5.19. Note that if
s(a) = ℓ then s(b) = n√ℓ. This implies that L must be closed under taking n-th roots for each n. Assuming
that L is a group satisfying this condition, one can construct the ν-divisible closure, sketched as follows:
1R
Step 1: Given a ∈ Rℓ, adjoin a formal element b ∈ R n√ℓ, and consider all formal sums
(7.4)
f (b) :=Xi
αibi :
αi ∈ R.
i ai.)
∈ L.
(Pi αibi is to be considered as the n-th root of Pi αn
Define Rb to be the set of all elements of the form (7.4), where any α ∈ R is identified with
αb0. We can define the sorting map s : Rb → L via
s(αi) n√ℓ
i
s(f (b)) =Xi
We define ∼=ν on Rb (notation as in (7.4)) by saying fb ∼=ν Pt′
In particular, fb ∼=ν c for c ∈ R if Pi αn
i ai ∼=ν cn. Likewise, we write fb >ν f′b := Pt′
i ai >ν α′j
αn
Now we can define addition on Rb so as to be ν-bipotent, where for ν-equivalent elements
we define f (b) + g(b) to be their formal sum (combining coefficients of the same powers of b);
multiplication is then defined in the obvious way, via distributivity over addition. Now Rb is an
L-layered bi-domain†, in view of Proposition 5.15.
j=0 α′jbj ifPi αn
naj.
i ai ∼=ν Pj α′j
naj.
j=0 α′jbi if
Step 2: Using Step 1 as an inductive step, one can construct the ν-divisible closure by means
of Zorn's Lemma, analogously to the well-known construction of the algebraic closure, cf. [37,
Theorem 4.88].
28
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Example 7.18 (Completion). One can construct the completion of any L-layered domain† R as follows:
We define ν-Cauchy sequences in R to be those sequences (ai) := {a1, a2, . . .} which become Cauchy
sequences modulo ∼=ν, but which satisfy the extra property that there exists an m (depending on the
sequence) for which s(ai) = s(ai+1), ∀i ≥ m. This permits us to define the sort of the ν-Cauchy sequence
to be s(am). Then we define the null ν-Cauchy sequences in R to be those sequences (ai) := {a1, a2, . . .}
which become null Cauchy sequences modulo ∼=ν , and the completion bR to be the factor group.
We also extend our given pre-order ν to ν-Cauchy sequences by saying that (ai) ∼=ν (bi) if (aib−1
) is
a null ν-Cauchy sequence, and, for (ai) 6∼=ν (bi), we say (ai) >ν (bi) when there is m such that ai >ν bi
for all i > m. The completion bR becomes an L-layered bi-domain† under the natural operations, i.e.,
componentwise multiplication of ν-Cauchy sequences, and addition given by the usual rule that
(In the last line, we arranged for the layers to be added when the ν-Cauchy sequences are ν-equivalent.)
(ai)
(bi)
νs(ai)+s(bi),s(ai+bi)(ai + bi)
(ai) + (bi) =
It is easy to verify ν-bipotence for bR.
Proposition 7.19. Suppose there is an embedding ϕ : R → F ′ of a uniform L-layered domain† R into a
1-divisible, uniform L-layered bi-semifield† F ′, and let F be the 1-divisible closure of the bi-semifield† of
fractions of R. Then F ′ is an extension of F . If F ′ is complete with respect to the ν-pre-order, then we
can take F ′ to be an extension of the completion of F .
if (ai) >ν (bi),
if (ai) <ν (bi),
if (ai) ∼=ν (bi).
These constructions are universal, in the following sense:
i
(7.5)
Proof. This is standard, so we just outline the argument. First we embed the L-layered bi-semifield† of
a1 → ϕ(b)
fractions of R into F ′, by sending b
, then c1b = a1d, implying
ϕ(c1b) = ϕ(a1d), and thus ϕ(b)
ϕ(c1) . Thus, we may assume that R is an L-layered bi-semifield†. Now
we define the map F → F ′ by sending m√a → m√ϕ(a), for each a ∈ F1. This is easily checked to be a
ϕ(a1) . This map is 1:1, since if b
ϕ(a1) = ϕ(d)
= d
c1
a1
well-defined, 1:1 layered homomorphism.
In case F ′ is complete, then we can embed the completion of F into F ′. (The completion of a 1-
divisible bi-semifield† is 1-divisible, since taking roots of a ν-Cauchy sequence in F1 yields a ν-Cauchy
sequence.)
(cid:3)
Corollary 7.20. Any L-layered homomorphism Φ : R → R′ of uniform L-layered bi-domains† extends
uniquely to an L-layered homomorphism eΦ : eR → eR′.
Proof. Each of the constructions of bi-semifield† of fractions, 1-divisible closure, and completion are
universal, so applying them in turn yields us a unique ordered monoid homomorphism from the tangible
component R1 to R′1, which readily extends (uniquely) to all of R since eℓ 7→ e′ℓ.
(cid:3)
7.5. Roots and layered varieties. In order to understand affine layered geometry, we need to know
more about the affine layered algebraic sets. We fix the sorting semiring† L for convenience, although one
could also let L vary, and think of S as a subset of R(n), where R is an L-layered bi-domain†. Standard
tropical geometry can be recaptured by taking S ⊆ R(n)
1 .
Definition 7.21. Suppose f, g ∈ Pol(S, R). We say that f dominates g at a ∈ S if f (a) ≥ν g(a).
Write f = Pi fi as a sum of monomials. The dominant part fa of f at a is the sum of all those fi
dominating f at a, i.e., for which fi(a) ∼=ν f (a). We write fX for the restriction of f to a nonempty
subset X of S.
Definition 7.22. An element a ∈ S is a corner root of f = Pi fi if f (a) is s(fi(a))-ghost for every
monomial fi of f . (Thus, fa contains at least two fi.) The (affine) corner locus of f ∈ Pol(S, R) with
respect to the set S is
Zcorn(f ;S) := {a ∈ S : a is a corner root of f}.
We write Zcorn(f ) for Zcorn(f ; R(n)).
CATEGORICAL LAYERED MATHEMATICS
29
An element a ∈ S is a cluster root of f if f (a) = fi(a) is 1-ghost for some monomial fi of fa.
(Thus, fa is comprised of a single monomial fi.) The combined ghost locus with respect to the set S
is
The (affine) corner algebraic set and the (affine) algebraic set of a subset A ⊆ Pol(S, R) with
Zcomb(f ;S) := {a ∈ S : a is a cluster or corner root}.
respect to the set S, are respectively
Zcorn(A;S) := \f∈A
Zcorn(f ;S),
Zcomb(A;S) := \f∈A
Zcomb(f ;S).
When S is unambiguous (usually R(n)), we write Zcorn(A) and Zcomb(A) for Zcorn(A;S) and Zcomb(A;S)
respectively.
Example 7.23. Here are basic examples of affine layered algebraic sets.
(i) We view R(n) as Zcomb(∅). Note that also R(n) = Zcomb({a}) for any "ghost" constant a ∈ R>1.
(ii) The empty set is an algebraic set: ∅ = Z({a}) for any a ∈ R1.
(iii) A single point a = (a1, . . . , an) ∈ S, where S ⊂ R(n)
1 , is a corner algebraic set:
(iv) The familiar tropical line in the affine plane is Zcorn(f ;S) where f is linear of the form
a = Z({λ1 + a1, . . . , λn + an}).
αλ1 + βλ2 + γ,
with α, β, γ ∈ R1, and S = R(2)
restricted to R(2)
1
"balancing condition" of tropical geometry.
1 . On the other hand, for I as in Example 7.14, Zcorn(I;S)
is the union of two perpendicular rays, and does not satisfy the celebrated
Example 7.24. A more sophisticated example: Whereas in the standard supertropical theory we have
(x + y + z)(xy + xz + yz) = (x + y)(x + z)(y + z),
they differ in the layered theory, since xyz has layer 3 in the left side but only layer 2 in the right side.
Thus the layered theory permits greater refinement in reducing tropical varieties.
8. The tropicalization functor on polynomials and their roots
The tropicalization map FLTrop of §6.2, Equation (6.4), extends readily to polynomials, i.e., to the
functor bFLTrop : Pol(S, ValMon+) → Pol(S, LayBidom†), where we define
n =Xi
n (cid:19) =Xi
FLTrop(ai)λi1
1 ··· λin
n ,
aiλi1
1 ··· λin
bFLTrop(cid:18)Xi
1 ··· λin
[1]v(ai) λi1
for i = (i1, . . . , in), (and analogously for morphisms).
If a ∈ F (n) is a root of f ∈ F [Λ], then clearly v(a) is a corner root of bFLTrop(f ). We are interested in
the opposite direction. One of the key results of tropical mathematics is Kapranov's theorem [14], which
says that for any polynomial f (λ1, . . . , λn), any corner root of the tropicalization of f has a pre-image
which is a root of f . This assertion also works for finite sets of polynomials, and thus for ideals, in view
of [36]. Our objective in this section is to understand this result in terms of the appropriate layered
categories.
Remark 8.1. Let A := F [λ1, . . . , λn]. Then the Puiseux series valuation V al extends naturally to a
map V al : A → R[λ1, . . . , λn], where each λi is fixed. If I is an ideal of A, then Φ(I) is an ideal of
Φ(A), so this "tropicalization" process sends ideals of algebras to semiring† ideals, and transfers many
properties from the "classical algebraic" world to the "tropical" world. One property which it does not
preserve is generation of ideals. For example, two different polynomials f, g of I might have the same
leading monomial and the same tropicalization, and then Φ(f + (−g)) cannot be described in terms of
Φ(f ) and Φ(g) = Φ(f ). One can bypass this particular difficulty by using Grobner bases (since they are
comprised of polynomials of different lowest orders), but the necessity of choosing the "right" generators
raises serious issues in tropical geometry. Fortunately, this concern is not critical in the current paper,
since we do not require generators for studying the relevant categories.
30
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
i
Remark 8.2. We start with a triple (F,G, v), where F for example may be the algebra of Puiseux series
over C, G = (R, +), and v : F → G the valuation Val. Any point (α1, . . . , αn) ∈ F (n) can be considered
as a valuation v extending v, where v(λi) = αi. This can be extended to the group G generated by the λi
and λ−1
in the ring of Laurent series over F . But if p is a prime ideal of F [Λ] := F [λ1, . . . , λn], then the
natural image of G in the field of fractions K of F [Λ]/p is a group ¯G, and Bieri-Groves [5] describe the
possible extensions of v to F [ ¯G]. Namely, ¯G, being a finitely generated Abelian group, can be written as
the direct sum of a free Abelian group of some rank m and a torsion group T . We let F [ ¯G] denote the
F -subalgebra of K generated by ¯G. After extending the valuation v to the free Abelian group, one sees by
an exercise of Bourbaki [6] that further extensions to F [ ¯G] correspond to corner roots of the polynomials
of p.
This is explained in the proof of [5, Theorem A], and can be explained tropically in terms of the proof
of Bourbaki's exercise:
If f (a) = 0 then two of the monomials of FLTrop(f ) must be equal and dominant when evaluated at a,
say αai = βaj, so one can extend v to a valuation v on F [a] given by v(a) = v(βα−1)
.
This discussion could be formulated in the language of [16], [18], [19], as elaborated in [20, Remark 6.6].
Definition 8.3. As in Remark 6.11, suppose F is an arbitrary field with valuation v : F → G, having
valuation ring R and associated graded algebra gr(F ). For any f ∈ F [λ], we define ¯f to be its natural
image in gr(F )[λ]. For a ⊳ F [Λ], we define the exploded tropicalization ¯a of a to be
i−j
An element a := (a1, . . . , an) of gr(F ) is a graded root of a polynomial ¯f ∈ gr(F )[λ] if
¯f (a1 + F>s(a1), . . . , an + F>s(an)) = 0
{ ¯f : f ∈ a}.
in gr(F ). (Intuitively, s(f (a1, . . . , an)) is larger than expected.)
We take F to be a Henselian field with respect to a valuation v whose residue field is algebraically
closed. For example, we could take F = K, the field of Puiseux series over C. We have two areas of
interest when studying Puiseux series -- the semifield† (which corresponds to the value group) and the
residue field, which is a copy of C. We can combine these using the 'exploded' structure of Definition 6.14.
Given a polynomial f ∈ gr(F )[Λ], we define its corner exploded roots to be
(cid:8)Graded roots a = (a1, . . . , an) ∈ F (n)
1
of f : s(f (a1, . . . , an)) = 0(cid:9),
cf. Remark 6.15. The corner exploded variety of an ideal a of gr(F ) is the set of common corner
exploded roots of the polynomials of a.
The standard valuation-theoretic proofs of Kapranov's theorem show that any corner root x of FLTrop(f )
is the tropicalization of a point in the variety Z defined by f . In other words, x lifts to an exploded root
of f . Payne's theorem [36] can be stated as follows:
Suppose X is an affine variety defined by a proper ideal a of F [Λ], and a is an graded root of the
exploded tropicalization ¯a of a. Then the preimage of any point defined by ¯a, if nonempty, is Zariski
dense in X. This is the algebraic essence of Parker's 'exploded' approach.
9. The category of affine layered geometry
Our goal in this section is to connect affine layered geometry to a category which can be studied by
means of standard algebraic techniques. This ties in with the algebraic categories of the previous sections,
by means of the coordinate semiring†, which is to be studied more thoroughly in a subsequent paper.
Throughout, let F denote a layered bi-semifield†.
9.1. The Zariski topology. We want to mimic the classical Zariski theory as far as we can, starting
with our layered bi-semifield† F and describing a topology on S, a given subset of F (n).
Actually, there are several natural topologies on S.
Definition 9.1. Suppose f =Pi fi is written as a sum of monomials in Pol(S, R). The fi-component
of f is
Dfi (S) := {a ∈ S : fi(a) = f (a)}.
CATEGORICAL LAYERED MATHEMATICS
31
Any root of f in a component Dfi must be a cluster root.
Remark 9.2. In [17, Definition 6.5], we defined the L-layered component topology to have as its
sub-base the components of polynomials of Pol(S, R). Note that different components of a polynomial are
disjoint, so open sets here are not necessarily dense in the component topology. Thus, although it provides
useful information, the component topology is too fine to permit us to develop tropical algebraic geometry
along classical lines. We rectify the situation by defining the principal corner open sets to be
D(f ;S) = S \ Zcorn(f ) = [i∈I
Dfi,
where f =Pi∈I fi is written as a sum of monomials in Pol(S, R). Put another way,
D(f ;S) = {a ∈ S : s(f (a)) = s(fi(a)) for some monomial fi of f},
The principal corner open sets form a base for a topology on S, which we call the (L-layered) corner
Analogously, one could respectively take cluster roots and use Zcomb(f ;S) in place of corner roots and
Zcorn(f ;S) to define the combined Zariski topology, whose closed sets are the algebraic sets. This is
a somewhat finer topology, but the corner Zariski topology provides a closer analog to the usual notions
of tropical geometry, so we will use that.
Zariski topology, whose closed sets are affine corner algebraic sets.
Lemma 9.3. The intersection of two principal corner open sets contains a nonempty principal corner
open set.
Proof. If f (a) = fi(a) on Df,i and g(a) = gj(a) on Dg,j, then clearly f g(a) = fi(a)gj(a) on Df,i ∩ Dg,j
and nowhere else.
(cid:3)
Proposition 9.4. All open sets in the (L-layered) corner Zariski topology are dense.
Proof. Immediate from the lemma.
(cid:3)
9.2. The coordinate semiring†. We can return to algebra via the coordinate semiring†, just as in
classical algebraic geometry.
Definition 9.5. The coordinate semiring† of an affine layered algebraic set X ⊆ S, denoted F [X], is
the natural image of the semiring† Pol(X, F ). The Laurent coordinate semiring† F (X) is the natural
image of Laur(X, F ); its elements are called the regular functions of the algebraic set.
Pol(X, F ) can be identified with classes of polynomials over X whose representatives are polynomials
having no inessential monomials, cf. [17, Definition 5.5]. We say that a function in F [X] (resp. F (X)) is
tangible if it can be written as a tangible polynomial (resp. Laurent polynomial), i.e., having coefficients
only in F1.
Remark 9.6. When X ⊂ Y we have a natural map F [Y ] → F [X] obtained by restricting the domain of
the function from Y to X.
Coordinate semirings† correspond naturally to congruences on Pol(S, F ) in the following manner.
Definition 9.7. A nonempty subset X ⊆ S defines the congruence of X on Fun(S, F ), denoted ΩX ,
whose underlying equivalence ≡X is given by
f ≡X g
iff
f (a) = g(a) for every a ∈ X.
Conversely, given a congruence Ω on Fun(S, F ), define the variety of the congruence
V (Ω) := {a ∈ S : f (a) = g(a), ∀(f, g) ∈ Ω} ⊆ S.
It is readily checked that
ΩX∪Y = ΩX ∩ ΩY ,
for any X, Y ⊂ S.
9.3. Zariski correspondences. We have various correspondences between varieties and the algebraic
structure.
32
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
9.3.1. The Zariski correspondence with ideals. Inspired by the layered Nullstellensatz given in [17, The-
orem 6.14], the naive approach would be to define the corner ideal Icorn(S) of a set S to be
{f ∈ Pol(S, F ) : a is a corner root of f,
∀a ∈ S},
and Icomb(S) to be
{f ∈ Pol(S, F ) : a is in the combined ghost locus of f,
∀a ∈ S}.
This approach misses the mark, somewhat. On the one hand, different congruences can define the
same ideal which is the pre-image of 0. On the other hand, there are "too many" ideals, in the sense that
not every ideal defines a variety, and the correct algebraic approach is to utilize congruences rather than
ideals. Furthermore, we need somehow to filter out those varieties obtained by degenerate intersections
of hypersurfaces; this is treated in a later paper under preparation.
9.3.2. The Zariski correspondence with congruences. As just noted, it makes more sense to deal with
congruences instead of ideals. We have the usual straightforward but important inverse Zariski corre-
spondence:
Proposition 9.8. If Ω1 ⊇ Ω2, then V (Ω1) ⊆ V (Ω2). Conversely, if Y ⊇ X, then ΩY ⊆ ΩX . Conse-
quently,
V (ΩV (Ω)) = V (Ω) and ΩV (ΩX ) = ΩX .
It follows that there is a 1:1 correspondence between congruences of varieties and varieties of congruences,
given by X 7→ ΩX and Ω 7→ V (Ω). Furthermore, the coordinate semiring† satisfies
F [V (Ω)] ∼= Pol(S, F )/Ω.
Proof. The inverse correspondence is immediate, and the next assertion is immediate. The 1:1 corre-
spondence is then formal. To see the last assertion, note that two polynomials in f, g are identified in
F [V (Ω)] iff they agree on V (Ω), which by definition is the point set on which every pair (f, g) ∈ Ω agree;
namely, f and g are identified in Pol(S, F )/Ω.
(cid:3)
By the proposition, one sees that for any nonempty subset X ⊆ S we have
F [X] ∼= Pol(S, F )/ΩX .
(9.1)
Definition 9.9. A morphism of affine layered algebraic sets Φ : X → Y is a continuous function that
preserves (i.e., pulls back) regular functions, in the sense that if U is an open subset of Y and ψ ∈ F (U ),
then ψ ◦ Φ ∈ F (Φ−1(U )).
L -TropAff is the category whose objects are the affine layered algebraic sets X ⊂ F (n) and whose
morphisms Φ : X → Y are morphisms of layered affine algebraic sets .
Proposition 9.10. Any morphism Φ : X → Y of affine layered algebraic set gives rise to a natural
algebra homomorphism Φ∗ : F (Φ(U )) → F (U ), by ψ 7→ Φ∗(ψ), where Φ∗(ψ)(a) = ψ(Φ(a)), for every
a ∈ U .
Proof. Φ∗(ψ + ϕ) = Φ∗(ψ) + Φ∗(ϕ) and Φ∗(ψϕ) = Φ∗(ψ)Φ∗(ϕ).
(cid:3)
We conclude by introducing the functor linking the algebraic and geometric (affine) categories.
Definition 9.11. FL -Coord is the contravariant functor from L -TropAff to ULayBidom† given by sending
an affine layered algebraic set X to its coordinate semiring† F [X], and any morphism Φ : X → Y of
affine layered algebraic sets to the layered semiring† homomorphism Φ∗ : F [Y ] → F [X], i.e., f 7→ fΦ
where fΦ(a) = f (Φ(a)).
Many subtleties lie behind this definition; for example, which affine layered varieties correspond to
the coordinate semirings† of tropical varieties satisfying the balancing condition? This question is to be
treated in a subsequent paper.
CATEGORICAL LAYERED MATHEMATICS
33
References
[1] M. Akian, R. Bapat, and S. Gaubert. Max-plus algebra, In: Hogben, L., Brualdi, R., Greenbaum, A., Mathias, R. (eds.)
Handbook of Linear Algebra. Chapman and Hall, London, 2006.
[2] M. Akian, S. Gaubert, and A. Guterman. Linear independence over tropical semirings and beyond. In Tropical and
Idempotent Mathematics, G.L. Litvinov and S.N. Sergeev, (eds.), Contemp. Math., 495:1 -- 38, 2009.
[3] V. Berkovich. Spectral theory and analytic geometry over non-Archimedean fields. Mathematical surveys and mono-
graphs 33, Amer. Math. Soc., 1990
[4] V. Berkovich. Lecture at Lodz joint meeting of the Polish and Israel Mathematical Societies, July, 2011
[5] R. Bieri and R. Groves. The geometry of the set of characters induced by valuations. J. fur die Reine und angevandte
Mathematik, 374:168 -- 195, 1984.
[6] N. Bourbaki. Commutative Algebra. Paris and Reading, 1972.
[7] A. Connes, C. Consani. Characteristsic 1, entropy, and the absolute point. preprint at arXiv:math.0911.3537, 2009.
[8] G. Cortinas, C. Haesemeyer, M. Walker, and C. Weibel. Toric varieties, monoid schemes, and descent, preprint, 2010.
[9] A.A. Costa. Sur la theorie g´en´erale des demi-anneaux. Publ. Math. Decebren 10 (1963), 14 -- 29.
[10] A. Dickenstein, E. M. Feichtner, and B. Sturmfels. Tropical discriminants. J. Amer. Math. Soc., 20:1111 -- 1133., 2007.
[11] M. Einsiedler, M. Kapranov, and D. Lind. Non-Archimedean amoebas and tropical varieties. J. Reine Angew. Math.,
601:139 -- 157, 2006.
[12] J. Golan. The theory of semirings with applications in mathematics and theoretical computer science, volume 54.
Longman Sci & Tech., 1992.
[13] H.C. Hutchins and H.J. Weinert. Homomorphisms and kernels of semifields Periodica Mathematica Hungaria,
21(2):113 -- 152, 1990.
[14] I. Itenberg, G. Mikhalkin, and E. Shustin. Tropical algebraic geometry, vol. 35. Birkhauser, 2007. Oberwolfach seminars.
[15] Z. Izhakian. Tropical arithmetic and matrix algebra. Commun. in Algebra, 37(4):1445 -- 1468, 2009.
[16] Z. Izhakian, M. Knebusch, and L. Rowen. Supertropical semirings and supervaluations. J. Pure and Appl. Alg.,
215(10):2431 -- 2463, 2011.
[17] Z. Izhakian, M. Knebusch, and L. Rowen. Layered tropical mathematics, preprint at arXiv:0912.1398, 2011.
[18] Z. Izhakian, M. Knebusch, and L. Rowen. Dominance and transmissions in supertropical valuation theory. Comm. in
Alg, to appear. (Preprint at arXiv:1102.1520.)
[19] Z. Izhakian, M. Knebusch, and L. Rowen. Supertropical monoids: Basics, canonical factorization, and lifting ghosts to
tangibles, preprint at arXiv:1108.1880, 2011.
[20] Z. Izhakian, M. Knebusch, and L. Rowen. Categories of layered semirings, preprint, 2012.
[21] Z. Izhakian and L. Rowen. Ideals of polynomial semirings in tropical mathematics. J. Alg. and its Appl., to appear.
(Preprint at arXiv:1111.6253.)
[22] Z. Izhakian and L. Rowen. Supertropical algebra. Adv. in Math. 324(8):1860 -- 1886, 2010.
[23] Z. Izhakian and L. Rowen. Supertropical matrix algebra. Israel J. Math. 182(1):383 -- 424, 2011.
[24] Z. Izhakian and L. Rowen. Supertropical matrix agebra II: solving tropical equations. Israel J. Math., 186(1):69-97,
2011.
[25] Z. Izhakian and L. Rowen. Supertropical resultants. J. Alg., 324(8):1860 -- 1886, 2010.
[26] Z. Izhakian and E. Shustin. Idempotent semigroups and tropical sets. J. EMS, 14:489 -- 520, 2012.
[27] N. Jacobson. Basic Algebra II. Freeman, 1980.
[28] M. Knebusch and D. Zhang, Manis Valuations and Prufer Extensions. I. A New Chapter in Commutative Algebra.
Lecture Notes in Mathematics, 1791, Springer-Verlag, Berlin, 2002.
[29] G. Litvinov. The Maslov dequantization, idempotent and tropical mathematics: a very brief introduction. J. of Math.
Sciences, 140(3):426 -- 444, 2007.
[30] D. Maclagan and B. Sturmfels. Tropical Geometry. Preprint, 2009.
[31] B. Parker. Exploded fibrations, preprint at arXiv: 0705.2408v1, 2007.
[32] M. Passare and A. Tsikh. Amoebas: their spines and their contours. Contemp. Math. 377:275 -- 288, 2005.
[33] A. Patchkoria. On derived functors of semimodule-valued functors. Proc. A. Razmadze Math. Inst. (in Russian) 83:60 --
75,1986.
[34] A. Patchkoria, Extension of semimodules and the Takahashi functor ExtA(C, A). Homology, Homotopy, and Applica-
tions, 5:387 -- 406, 2003.
[35] S. Payne. Fibers of tropicalizations, Arch. Math., 2010 Correction: preprint at arXiv:0705.1732v2 [math.AG], 2012.
[36] S. Payne. Analytification is the limit of all tropicalizations, preprint at arXiv: 0806.1916v3 [math.AG], 2009.
[37] L. Rowen. Graduate algebra: Commutative view. Graduate Studies in Mathematics 73, 2006. American Mathematical
Society 91, 2006.
[38] L. Rowen. Graduate algebra: Noncommutative view. Graduate Studies in Mathematics 73, 2006. American Mathemat-
ical Society, 2006.
[39] E. Sheiner and S. Shnider. An exploded-layered version of Payne's generalization of Kapranov's theorem. preprint,
2012.
[40] E. Shustin. Patchworking singular algebraic curves, non-Archimedean amoebas and enumerative geometry. Algebra i
Analiz, 17(2):170-214, 2005.
[41] M.Takahashi. Extension of semimodules I. Math.Sem.Notes Kobe. Univer. 10:563 -- 592, 1982.
[42] J. Worthington. A bialgebra approach to automata and formal language theory. Proceedings of the 2009 International
Symposium on Logical Foundations of Computer Science, Proceeding LFCS 09, 451 -- 467, 2009.
34
Z. IZHAKIAN, M. KNEBUSCH, AND L. ROWEN
Department of Mathematics, Bar-Ilan University, Ramat-Gan 52900, Israel
E-mail address: [email protected]
Department of Mathematics, University of Regensburg, Regensburg, Germany
E-mail address: [email protected]
Department of Mathematics, Bar-Ilan University, Ramat-Gan 52900, Israel
E-mail address: [email protected]
|
1107.2076 | 1 | 1107 | 2011-07-11T17:24:44 | Finite Modules over $\Bbb Z[t,t^{-1}]$ | [
"math.RA",
"math.AT"
] | Let $\Lambda=\Bbb Z[t,t^{-1}]$ be the ring of Laurent polynomials over $\Bbb Z$. We classify all $\Lambda$-modules $M$ with $|M|=p^n$, where $p$ is a primes and $n\le 4$. Consequently, we have a classification of Alexander quandles of order $p^n$ for $n\le 4$. | math.RA | math |
FINITE MODULES OVER Z[t, t−1]
XIANG-DONG HOU
Abstract. Let Λ = Z[t, t−1] be the ring of Laurent polynomials over Z. We
classify all Λ-modules M with M = pn, where p is a primes and n ≤ 4.
Consequently, we have a classification of Alexander quandles of order pn for
n ≤ 4.
1. Introduction
Let Λ = Z[t, t−1] be the ring of Laurent polynomials over Z, which is also the
group ring over Z of the infinite cyclic group. Each Λ-module is uniquely determined
by a pair (M, α), where M is an abelian group and α ∈ AutZ(M ). The resulting
Λ-module, denoted by Mα, is M with a scalar multiplication defined by tx = α(x),
x ∈ M . If two Λ-modules Mα and Nβ are isomorphic, where α ∈ AutZ(M ) and
β ∈ AutZ(N ), then M ∼= N as abelian groups. Moreover, for α, β ∈ AutZ(M ),
Mα ∼= Mβ if and only if α and β are conjugate in AutZ(M ). Thus, to classify Λ-
modules with an underlying abelian group M is to determine the conjugacy classes
of AutZ(M ).
Our interest in finite Λ-modules comes from topology. In knot theory, a quandle
is defined to be a set of Q equipped with a binary operation ∗ such that for all
x, y, z ∈ Q,
(i) x ∗ x = x,
(ii) ( ) ∗ y is a permutation of Q,
(iii) (x ∗ y) ∗ z = (x ∗ z) ∗ (y ∗ z).
Finite quandles are used to color knots; the number of colorings of a knot K by
a finite quandle Q is an invariant of K which allows us to distinguish inequivalent
knots effectively [1, 2].
An Alexander quandle is a Λ-module M with a quandle operation defined by
x∗y = tx+(1−t)y, x, y ∈ M . The following theorem is of fundamental importance.
Theorem 1.1 ([8]). Two finite Alexander quandles M and N are isomorphic if
and only if M = N and the Λ-modules (1 − t)M and (1 − t)N are isomorphic.
Therefore, the classification of finite Alexander quandles is essentially the clas-
sification of finite Λ-modules.
The classification of finite Λ-modules can be reduced to that of Λ-modules of
order pn, where p is a prime; see section 2. The same is true for the classification
of finite Alexander quandles; see section 4. Finite Alexander quandles have been
classified for orders up to 15 in [8] and for order 16 in [6, 7]. Also known is the
classification of connected Alexander quandles of order p2 [3, 8]. (A finite Alexander
quandle is called connected if 1 − t ∈ AutZ(M ).) The purpose of the present paper
2000 Mathematics Subject Classification. 16S34, 20K01, 57M27.
Key words and phrases. Alexander quandle, finite module, knot, group ring.
1
2
XIANG-DONG HOU
is to classify all Λ-modules and Alexander quandles of order pn, n ≤ 4. The
details of the classification are given in Table 1 in the appendix. For a snapshot,
there are 5p4 − 2p3 − 2p − 1 nonisomorphic Λ-modules of order p4 and there are
5p4 − 6p3 + p2 − 6p − 1 nonisomorphic Alexander quandles of order p4.
In section 2, we show that every finite Λ-module has a unique decomposition
where each direct summand M has the following properties:
(i) M = pn for some prime p and integer n > 0.
(ii) When treating t as an element of EndZp (M/pM ), the minimal polynomial
of t is a power of some irreducible f ∈ Zp[X].
In section 3, we classify such Λ-modules M of order pn with n ≤ 4. In section 4, we
derive the classification of Alexander quandles of order pn, n ≤ 4, from the results
of section 3. For this purpose, we prove the following fact which is of interest in
its own right: Given a finite Λ-module N and an integer l > 0, a necessary and
sufficient condition for the existence of a Λ-module M ⊃ N such that (1 − t)M = N
and M/N = l is N/(1 − t)N (cid:12)(cid:12) l.
In our notation, the letter t is reserved for the element t ∈ Λ = Z[t, t−1]. The
group of units of a ring R is denoted by R×; the set of all m × n matrices over R
is denoted by Mm×n(R).
2. Decomposition of Finite Λ-Modules
Let M be a finite Λ-module. For each prime p, let
Mp = {x ∈ M : pnx = 0 for some n ≥ 0}.
It is quite obvious that
(2.1)
Mp.
M =Mp
Moreover, two finite Λ-modules M and N are isomorphic if and only if Mp ∼= Np
for all primes p.
Theorem 2.1. Let M be a finite Λ-module with M = pn. For each irreducible
f ∈ Zp[X], let f ∈ Z[X] be a lift of f and define
Mf = {x ∈ M : f (t)mx = 0 for some m ≥ 0}.
(2.2)
Then
(2.3)
Mf ,
M =Mf
where f runs over all irreducible polynomials in Zp[X]. Moreover, if N is another
finite Λ-module whose order is a power of p, then M ∼= N if and only if Mf ∼= Nf
for all irreducible f ∈ Zp[X].
Note. Mf depends only on f but not on f . Also, Mf = 0 unless f divides the
minimal polynomial of t (viewed as an element of EndZp (M/pM )).
Proof of Theorem 2.1. 1◦ Let the minimal polynomial of t (∈ EndZp (M/pM )) be
f e1
1 · · · f ek
k , where f1, . . . , fk ∈ Zp[X] are distinct irreducibles and e1, . . . , ek are
positive integers. We claim that
(2.4)
M = M1≤i≤k
Mfi .
FINITE MODULES OVER Z[t, t−1]
3
We first show that P1≤i≤k Mfi
(cid:0)P1≤j≤k, j6=i Mfj(cid:1). Then there exists m > 0 such that fi(t)mx = 0 and
(cid:0)Q1≤j≤k, j6=i fj(t)(cid:1)m
x = 0. Since gcd(cid:0)fi,Q1≤j≤k, j6=i fj(cid:1) = 1, there exist u, v ∈
is a direct sum. Assume that x ∈ Mfi ∩
Zp[X] such that
uf m
i + v(cid:16) Y1≤j≤k
j6=i
fj(cid:17)m
= 1.
Let u, v ∈ Z[X] be arbitrary lifts of u, v, respectively. Then
m
ufi
+ v(cid:16) Y1≤j≤k
j6=i
fj(cid:17)m
≡ 1 (mod p).
u(t)fi(t)m + v(t)(cid:16) Y1≤j≤k
j6=i
fj(t)(cid:17)m
∈ AutZ(M ).
hu(t)fi(t)m + v(t)(cid:16) Y1≤j≤k
j6=i
fj(t)(cid:17)mix = 0,
Therefore
Since
we have x = 0.
Now we prove that M =P1≤i≤k Mfi . There exist u1, . . . , uk ∈ Zp[X] such that
X1≤i≤k
ui(cid:16) Y1≤j≤k
j6=i
f ej
j (cid:17)n
= 1.
Let ui ∈ Z[X] be a lift of ui and let Fi =Q1≤j≤k, j6=i fj
uiF n
i ≡ 1
(mod p).
X1≤i≤k
ej . Then
Thus P1≤i≤k ui(t)Fi(t)n ∈ AutZ(M ). It follows that
Fi(t)nM.
(2.5)
M = X1≤i≤k
Since (cid:0)Q1≤j≤k fj(t)ej(cid:1)M ⊂ pM , we have
fi(t)einFi(t)nM =(cid:16) Y1≤j≤k
fj(t)ej(cid:17)n
M ⊂ pnM = 0.
Thus Fi(t)nM ⊂ Mfi. Then it follows from (2.5) that M =P1≤i≤k Mfi .
2◦ Let N be another finite Λ-module whose order is a power of p.
If there
is a Λ-module isomorphism φ : M → N , then for each irreducible f ∈ Zp[X],
φMf : Mf → Nf is an isomorphism. Conversely, if Mf ∼= Nf for all irreducible
f ∈ Zp[X], then by (2.3), M ∼= N .
(cid:3)
4
XIANG-DONG HOU
3. Classification of Λ-Modules of Order pn, n ≤ 4
3.1. The automorphism group of a finite abelian group.
Let p be a prime and let m ≥ n > 0 be integers. Elements of Zpm can be viewed
as elements of Zpn via the homomorphism
Zpm
−→
Zpn
a + pmZ 7−→ a + pnZ,
a ∈ Z.
Likewise, elements of pm−nZpn can be viewed as elements of Zpm via the embedding
pm−nZpn
pm−n(a + pnZ)
−→
7−→ pm−na + pmZ,
Zpm
a ∈ Z.
We shall adopt these conventions hereafter.
Let M = Zn1
pe1 × · · · × Znk
pek , where ni > 0 and e1 > · · · > ek > 0. Elements of
EndZ(M ) are of the form
σA : M −→
M
x1
...
xk
7−→ A
,
xi ∈ Zni
pei ,
x1
...
xk
A11
A21
...
Ak1
A =
pe1−e2 A12
A22
...
Ak2
· · ·
· · ·
· · ·
pe1−ek A1k
pe2−e1 A2k
...
Akk
,
where
(3.1)
and Aij ∈ Mni×nj (Zpei ). Let M(M ) denote the set of all matrices of the form (3.1).
Then
M(M ) −→ EndZ(M )
A
7−→
σA
is a ring isomorphism. Let GL(M ) denote the group of units of M(M ). Of course
GL(M ) ∼= AutZ(M ) under the above isomorphism. It is known [4, 9] (and also easy
to prove) that
GL(M ) = {A : A is of the form (3.1) with Aii ∈ GL(ni, Zpei ), 1 ≤ i ≤ k}.
The modulo p reduction from GL(M ) to GL(n1 + · · · + nk, Zp) is denoted by ( ).
For each (monic) irreducible f ∈ Zp[X] with f 6= X, define
GL(M )f = {A ∈ GL(M ) : the minimal polynomial of
A ∈ GL(n1 + · · · + nk, Zp) is a power of f }.
If λ(i) = (λi1, λi2, . . . ) is a partition of the integer ni/ deg f , 1 ≤ i ≤ k, we define
GL(M )λ(1) ...λ(k)
f
= {A as in (3.1) : the elementary divisors of Aii
are f λi1 , f λi2 , . . . , 1 ≤ i ≤ k}.
In this setting, our goal is to determine the GL(M )-conjugacy classes in GL(M )f .
We will proceed according to the structure of (M, +).
FINITE MODULES OVER Z[t, t−1]
5
3.2. (M, +) = Zpe .
In this case we must have f = X − a, a ∈ Z×
p . The conjugacy classes in GL(M )f
are represented by
3.3. (M, +) = Zn
p .
[b],
b ∈ Zpe , b ≡ a (mod p).
In this case we must have deg f n. The conjugacy classes in GL(M )f are
represented by the rational canonical forms in GL(n, Zp) with elementary divisors
f λ1, f λ2 , . . . , where λ1 ≥ λ2 ≥ · · · > 0 is a partition of n/ deg f .
3.4. (M, +) = Zpe × Zp, e > 1.
In this case, deg f = 1.
Theorem 3.1. Assume (M, +) = Zpe × Zp, e > 1, and f = X − a, a ∈ Z×
conjugacy classes in GL(M )f are represented by the following matrices:
p . The
(i) (cid:20)b
(ii) (cid:20)b
(iii) (cid:20)b
0
b(cid:21) +(cid:20)pe−1α 0
0(cid:21),
b(cid:21) +(cid:20)0 0
1 0(cid:21),
b(cid:21) +(cid:20)0
0 (cid:21),
pe−1
γ
0 < b < pe−1, b ≡ a (mod p), α ∈ Zp.
0 < b < pe−1, b ≡ a (mod p).
0 < b < pe−1, b ≡ a (mod p), γ ∈ Zp.
Proof. Elements of GL(M )f are of the form
A(b, α, β, γ) :=(cid:20)b
0
0
b(cid:21) +(cid:20)pe−1α pe−1β
0 (cid:21) ,
γ
where 0 < b < pe−1, b ≡ a (mod p), α, β, γ ∈ Zp. Let A(b, α, β, γ), A(b, α′, β′, γ′) ∈
GL(M )f and
The above equation can be written as a matrix equation over Zp:
The equation P A(b, α, β, γ) = A(b, α′, β′, γ′)P is equivalent to
x ∈ Z×
pe , w ∈ Z×
p , y, z ∈ Zp.
z
0
wγ
P =(cid:20)x pe−1y
w (cid:21) ∈ GL(M ),
(cid:20)pe−1(xα + yγ) pe−1xβ
(cid:20)xα + yγ xβ
(cid:20)x y
0 w(cid:21)(cid:20)β α
wγ
0
(cid:21) .
0
γ′x
(cid:21) =(cid:20)pe−1(α′x + β′z) pe−1β′w
0 (cid:21) =(cid:20)α′x + β′z β′w
0 (cid:21) ,
γ(cid:21) =(cid:20)β′ α′
γ′(cid:21)(cid:20)w z
0 x(cid:21) .
γ′x
0
equivalently,
So, A(b, α, β, γ) and A(b, α′, β′, γ′) are conjugate if and only if there exist x, y ∈ Z×
p
and y, z ∈ Zp such that
(3.2)
(cid:20)x y
0 w(cid:21)(cid:20)β α
γ(cid:21)(cid:20)w−1
0
z
0
0
x−1(cid:21) =(cid:20)β′ α′
γ′(cid:21) .
0 γ ′i ∈ M, say A ∼ A′
0 γi , A′ = h β ′ α′
p and y, z ∈ Zp. It is easy to see that the
Let M = {h β α
0 γi : α, β, γ ∈ Zp}. For A = h β α
if (3.2) is satisfied for some x, w ∈ Z×
∼ equivalence classes in M are represented by
6
XIANG-DONG HOU
0
(i) (cid:20)0 α
0(cid:21) , α ∈ Zp,
(ii) (cid:20)0
1(cid:21),
(iii) (cid:20)1
γ(cid:21) ,
γ ∈ Zp.
These matrices correspond to the representatives of the conjugacy classes in GL(M )f
stated in the theorem.
(cid:3)
3.5. (M, +) = Z2
p2 .
In this case, deg f = 1 or 2.
Theorem 3.2. Assume (M, +) = Z2
p2 .
(i) Let f = X − a, a ∈ Z×
p . Then the conjugacy classes in GL(M )(1,1)
f
are
represented by the following matrices:
(i.1) (cid:20)b
(i.2) (cid:20)b
(i.3) (cid:20)b
b(cid:21) + p(cid:20)α
δ(cid:21) ,
b(cid:21) + p(cid:20)α 1
α(cid:21) ,
b(cid:21) + p(cid:20) 0
1
−b0 −b1(cid:21) ,
Zp[X] irreducible.
(ii) Let f = X − a, a ∈ Z×
sented by
0 < b < p, b ≡ a (mod p), 0 ≤ α ≤ δ < p.
0 < b < p, b ≡ a (mod p), α ∈ Zp.
0 < b < p, b ≡ a (mod p), X 2 + b1X + b0 ∈
p . Then the conjugacy class in GL(M )(2)
f
are repre-
(cid:20)b
0
1
b(cid:21) + p(cid:20)α 0
0(cid:21) ,
γ
0 < b < p, b ≡ a (mod p), α, γ ∈ Zp.
(iii) Let f = X 2 + a1X + a0 ∈ Zp[X] be irreducible. Then the conjugacy classes
in GL(M )f are represented by
1
−b0 −b1(cid:21) + p(cid:20)α β
0(cid:21) ,
(cid:20) 0
0
0 ≤ b0, b1 < p, b0 ≡ a0
(mod p),
b1 ≡ a1
(mod p), α, β ∈ Zp.
Proof. We remind the reader that in the proof, our notation is local in each of the
three cases.
(i) Elements of GL(M )(1,1)
f
are of the form
γ
are represented by
0 < b < p, b ≡ a (mod p), α, β, γ, δ ∈ Zp.
b(cid:21) + p(cid:20)α β
δ(cid:21) ,
Clearly, A(b, α, β, γ, δ) and A(b, α′, β′, γ′, δ′) are conjugate in GL(M ) if and only if
A(b, α, β, γ, δ) :=(cid:20)b
γ ′ δ′i are conjugate in M2×2(Zp). The conjugacy classes in M2×2(Zp)
γ δi and h α′ β ′
h α β
(1) (cid:20)α
δ(cid:21) ,
(2) (cid:20)α 1
α(cid:21) , α ∈ Zp,
0 ≤ α ≤ δ < p,
FINITE MODULES OVER Z[t, t−1]
7
(3) (cid:20) 0
−b0 −b1(cid:21) , X 2 + b1X + b0 ∈ Zp[X] irreducible.
1
These correspond to the matrices in (i.1) -- (i.3).
(ii) Elements of GL(M )(2)
f
are conjugate to matrices of the form
A(b, α, β, γ, δ) :=(cid:20)b
0
1
b(cid:21) + p(cid:20)α β
δ(cid:21) ,
γ
0 < b < p, b ≡ a (mod p), α, β, γ, δ ∈ Zp.
Assume that P ∈ GL(M ) = GL(2, Z
p2 ) such that
(3.3)
Then over Zp,
P A(b, α, β, γ, δ) = A(b, α′, β′, γ′, δ′)P.
which implies that P = [ c d
0 c ], c ∈ Z×
p , d ∈ Zp. So
P(cid:20)b
0
1
b(cid:21) =(cid:20)b
0
1
b(cid:21) P ,
P =(cid:20)c d
c(cid:21) + p(cid:20)x y
z w(cid:21) ,
0
x, y, z, w ∈ Zp.
Now (3.3) becomes
p(cid:20)c d
c(cid:21)(cid:20)α β
δ(cid:21) + p(cid:20)x y
z w(cid:21)(cid:20)b
γ
0
0
Over Zp, this becomes
(cid:20)c d
c(cid:21)(cid:20)α β
δ(cid:21) +(cid:20)x y
z w(cid:21)(cid:20)b
γ
0
0
which can be simplified as
1
b(cid:21) = p(cid:20)α′ β′
δ′(cid:21)(cid:20)c d
c(cid:21) + p(cid:20)b
γ′
0
0
1
b(cid:21)(cid:20)x y
z w(cid:21) .
1
b(cid:21) =(cid:20)α′ β′
δ′(cid:21)(cid:20)c d
c(cid:21) +(cid:20)b
γ′
0
0
1
b(cid:21)(cid:20)x y
z w(cid:21) ,
(cid:20)c d
c(cid:21)(cid:20)α β
δ(cid:21) =(cid:20)α′ β′
δ′(cid:21)(cid:20)c d
c(cid:21) +(cid:20)z w − x
−z (cid:21) .
γ′
γ
0
0
0
Thus, A(b, α, β, γ, δ) and A(b, α′, β′, γ′, δ′) are conjugate if and only if there exist
d, z, w ∈ Zp such that
(3.4)
0
γ
(cid:20)1 d
1(cid:21)(cid:20)α β
γ δi, A′ = h α′ β ′
δ(cid:21)(cid:20)1 −d
1 (cid:21) +(cid:20)z w
0 −z(cid:21) =(cid:20)α′ β′
δ′(cid:21) .
γ′
0
For A = h α β
some d, z, w ∈ Zp. It is easy to see that the ∼ equivalence classes in M2×2(Zp) are
represented by
γ ′ δ′i ∈ M2×2(Zp), say A ∼ A′ if (3.4) is satisfied for
which correspond to the matrices in (ii).
(cid:20)α 0
0(cid:21) , α, β ∈ Zp,
γ
(iii) Elements of GL(M )f are conjugate to matrices of the form
A(f, α, β, γ, δ) :=(cid:20) 0
−b0 −b1(cid:21) + p(cid:20)α β
δ(cid:21) , 0 ≤ b0, b1 < p, b0 ≡ a0
γ
1
(mod p),
b1 ≡ a1
(mod p), α, β, γ, δ ∈ Zp.
Assume that P ∈ GL(2, Z
p2 ) such that
(3.5)
P A(f, α, β, γ, δ) = A(f, α′, β′, γ′, δ′)P.
8
XIANG-DONG HOU
Then over Zp,
which implies that
1
−b0 −b1(cid:21) =(cid:20) 0
P (cid:20) 0
−b0 −b1(cid:21) P ,
1
P = uI + v(cid:20) 0
−b0 −b1(cid:21) + p(cid:20)x y
z w(cid:21) ,
1
0 ≤ u, v < p, (u, v) 6= (0, 0), x, y, z, w ∈ Zp.
Now (3.5) becomes the following equation over Zp:
(3.6)
(cid:16)uI + v(cid:20) 0
−b0 −b1(cid:21)(cid:17)(cid:20)α β
γ
1
1
[5, §4.4] and has a basis
The space (cid:8)(cid:2) 0
−b0 −b1(cid:3) X − X(cid:2) 0
−b0 −b1(cid:21)(cid:20)1 0
(cid:20) 0
−b0 −b1(cid:21)(cid:20)0
(cid:20) 0
1
1
1
1
γ′
δ(cid:21) = (cid:20)α′ β′
+(cid:20) 0
δ′(cid:21)(cid:16)uI + v(cid:20) 0
−b0 −b1(cid:21)(cid:20)x y
−b0 −b1(cid:21)(cid:17)
z w(cid:21) −(cid:20)x y
−b0 −b1(cid:21) .
−b0 −b1(cid:3) : X ∈ M2×2(Zp)(cid:9) has dimension 2 over Zp
z w(cid:21)(cid:20) 0
1
1
1
1
0 0(cid:21) −(cid:20)1 0
0(cid:21) −(cid:20)0
0 0(cid:21)(cid:20) 0
0(cid:21)(cid:20) 0
−b0 −b1(cid:21) =(cid:20) 0
−b0 −b1(cid:21) =(cid:20)b0
0 (cid:21) ,
0 −b0(cid:21) .
−b0
−1
b1
0
1
1
0
So (3.6) can be written as
(3.7)
(cid:16)uI + v(cid:20) 0
−b0 −b1(cid:21)(cid:17)(cid:20)α β
γ
1
γ′
δ(cid:21) = (cid:20)α′ β′
+(cid:20)−d
−b0 −b1(cid:21)(cid:17)
δ′(cid:21)(cid:16)uI + v(cid:20) 0
(cid:21) ,
− d b1
b0
d
c, d ∈ Zp.
c
b0
1
c
Thus A(f, α, β, γ, δ) and A(f, α′, β′, γ′, δ′) are conjugate if and only if there exist
0 ≤ u, v ≤ v < p, (u, v) 6= (0, 0), and c, d ∈ Zp such that
(3.8)
1
γ
c
b0
(cid:16)uI+v(cid:20) 0
For A = h α β
−b0 −b1(cid:21)(cid:17)(cid:16)(cid:20)α β
=(cid:20)α′ β′
δ(cid:21)−(cid:20)−d
δ′(cid:21) .
γ ′ δ′i ∈ M2×2(Zp), say A ∼ A′ if (3.8) is satisfied for
γ δi , A′ = h α′ β ′
−b0 −b1(cid:21)(cid:17)−1
(cid:21)(cid:17)(cid:16)uI+v(cid:20) 0
some 0 ≤ u, v < p, (u, v) 6= (0, 0), and c, d ∈ Zp. It remains to show that the ∼
equivalence classes in M2×2(Zp) are represented by
− d b1
b0
d
γ′
1
c
(cid:20)α β
0(cid:21) , α, β ∈ Zp.
0
First, choose u = 1, v = 0, c = −γ, d = −δ. Then the left side of (3.8) becomes
(cid:2) α′ β ′
0 0 (cid:3) for some α′, β′ ∈ Zp.
Next, assume that (3.8) holds with γ = δ = γ′ = δ′ = 0. We want to show that
(α, β) = (α′, β′). Taking the traces of the two sides of (3.8), we have α = α′. Now
(3.7) with α = α′ and γ = δ = γ′ = δ′ = 0 gives
uβ
−vb0α −vb0β(cid:21) =(cid:20)αu − β′vb0 αv − β′(u − vb1)
(cid:20) uα
0
0
(cid:21) +(cid:20)−d
c
c
b0
− d b1
b0
d
(cid:21) .
FINITE MODULES OVER Z[t, t−1]
9
Taking the traces, we have vb0β = vb0β′. If v 6= 0, then β = β′. If v = 0, then
c = d = 0, which implies uβ = uβ′. Since u 6= 0, we also have β = β′.
(cid:3)
3.6. (M, +) = Z
p2 × Z2
p.
In this case deg f = 1.
Theorem 3.3. Assume (M, +) = Z
p2 × Z2
p and f = X − a, a ∈ Z×
p .
(i) The conjugacy classes in GL(M )(1)(1,1)
f
matrices:
are represented by the following
0 < b < p, b ≡ a (mod p), α ∈ Zp.
0 < b < p, b ≡ a (mod p).
0 < b < p, b ≡ a (mod p), η ∈ Zp.
0 < b < p, b ≡ a (mod p).
0 < b < p, b ≡ a (mod p), α ∈ Zp.
0 < b < p, b ≡ a (mod p).
0 < b < p, b ≡ a (mod p), η ∈ Zp.
b
b
b
b
(i.1)
(i.2)
(i.3)
(i.4)
trices:
b
b
b
(ii.1)
(ii.2)
(ii.3)
b
b
b
b
b 1
b 1
b 1
pα 0
0
0
0
0
0
0
0 0
0 0
0
,
0 0
,
0
,
0 0
p 0
0
0
0
,
0 p 0
1
0 0
0
pα 0
0
0
0
0
0 0
0 0
1 0
0
0
0
0
0
,
0
,
0
,
p 0
0
0
0
0
1
0
0
0
η
b
+
b
+
+
b
b
+
b
+
b
+
b
+
+
b
0
0
η
b
(ii) The conjugacy classes in GL(M )(1)(2)
f
are represented by the following ma-
Proof. (i) Elements of GL(M )(1)(1,1)
are of the form
f
b
A(b, α, . . . , η) :=
pα pβ pγ
δ
0
η
0
0
0
,
Assume that P ∈ GL(M ) such that
(3.9)
Write
P A(b, α, . . . , η) = A(b, α′, . . . , η′)P.
P =(cid:20)x pY
Z W(cid:21) ,
0 < b < p, b ≡ a (mod p), α, . . . , η ∈ Zp.
10
XIANG-DONG HOU
where x ∈ Z×
p2 , Y ∈ M1×2(Zp), Z ∈ M2×1(Zp), W ∈ GL(2, Zp). Then (3.9) becomes
η(cid:21) px[β, γ]
pxα + pY (cid:20)δ
W (cid:20)δ
η(cid:21)
Over Zp, this becomes
0
0
α
=
=
η(cid:21)
=
(cid:20)δ
η(cid:21)
(cid:20)W −1
(cid:20)δ
η(cid:21)
(cid:20)δ
η(cid:21)
, A′ =
(cid:20)δ
[β, γ]
α
α
α
0
0
η(cid:21) x[β, γ]
xα + Y (cid:20)δ
W (cid:20)δ
η(cid:21)
0 W(cid:21)
(cid:20)x Y
[β, γ]
0
0
[β, γ]
0 W(cid:21)
(cid:20)x Y
M =n
pα′x + p[β′, γ′]z
(cid:20)δ′
η′(cid:21) x
α′x + [β′, γ′]z
(cid:20)δ′
η′(cid:21) x
p[β′, γ′]W
0
.
[β′, γ′]W
0
.
[β′, γ′]
0
0
α′
η′(cid:21)
(cid:20)W Z
x(cid:21) .
(cid:20)δ′
x−1(cid:21) =
η′(cid:21)
(cid:20)δ′
[β′, γ′]
α′
Z
0
: α, . . . , η ∈ Zpo.
η′(cid:21)
(cid:20)δ′
∈ M,
[β′, γ′]
α′
0
(3.10)
Let
For
It is more convenient to write the above equation as
So, A(b, α, . . . , η) and A(b, α′, . . . , η′) are conjugate if and only if
for some x ∈ Z×
p , W ∈ GL(2, Zp), Y ∈ M1×2(Zp), Z ∈ M2×1(Zp).
say A ∼ A′ if (3.10) is satisfied. It is easy to see that the ∼ equivalence classes in
M are represented by
0
α
[β, γ]
A =
0(cid:21)
(cid:20)0
, α ∈ Zp,
0(cid:21)
(cid:20)1
,
η(cid:21)
(cid:20)0
0(cid:21)
(cid:20)1
η ∈ Zp,
.
,
[0 0]
[0 0]
[1 0]
[1 0]
(1)
(2)
(3)
(4)
FINITE MODULES OVER Z[t, t−1]
11
These correspond to the matrices in (i.1) -- (i.4).
(ii) Elements of GL(M )(1)(1,1)
f
are conjugate for matrices of the form
b
A(b, α, . . . , η) :=
b 1
b
+
pα pβ pγ
δ
0
η
0
0
0
Assume that P ∈ GL(M ) such that
, 0 < b < p, b ≡ a (mod p), α, . . . , η ∈ Zp.
P A(b, α, . . . , η) = A(b, α′, . . . , η′)P.
P =(cid:20)x pY
Z W(cid:21) ,
where x ∈ Z×
p2 , Y ∈ M1×2(Zp), Z ∈ M2×1(Zp), W ∈ GL(2, Zp). Since W commutes
0 c ]. Equation (3.11) is equivalent to
(3.11)
Write
with (cid:2) b 1
pα′x + p[β′, γ′]Z p[β′, γ′]W
0
.
Over Zp, this becomes
η(cid:21) px[β, γ]
0
pxα + pY (cid:20)δ
W (cid:20)δ
η(cid:21)
+
=
0
1
0
0
1
0(cid:21) Z (cid:20)0
b(cid:3), we have W = [ c d
+
0 pY (cid:20)0 1
0(cid:21)
0 W (cid:20)0
0(cid:21)
=
0(cid:21) W
(cid:20)0 1
+
xα+Y(cid:20)δ
η(cid:21) x[β, γ]
W(cid:20)δ
η(cid:21)
+
Y (cid:20)0 1
0(cid:21)
η(cid:21)
(cid:20)δ
[β, γ]
α
0
0
0
0
(cid:20)δ′
η′(cid:21) x
0(cid:21)Z 0
+
(cid:20)0 1
=
0(cid:21) Z
0
1
−(cid:20)0
0 Y(cid:20)0 1
0(cid:21)
0
0
0 W(cid:21)
(cid:20)x Y
which can be written as
α′x+[β′, γ′]Z [β′, γ′]W
(cid:20)δ′
η′(cid:21) x
0
,
[β′, γ′]
0
α′
η′(cid:21)
(cid:20)W Z
x(cid:21) .
(cid:20)δ′
0
So, A(b, α, . . . , η) and A(b, α′, . . . , η′) are conjugate if and only if
(3.12)
[β, γ] α
0 W(cid:21)
(cid:20)x Y
0
η(cid:21)
"W −1 Z
(cid:20)δ
0
x−1# +
Y(cid:20)0 1
0(cid:21)W −1 Y (cid:20)0 1
0(cid:21) Z
(cid:20)0 1
0(cid:21) W Z
0
=
for some x ∈ Z×
p , Y ∈ M1×2(Zp), Z ∈ M2×1(Zp), W = [ c d
c ] ∈ GL(2, Zp).
[β′, γ′] α′
0
η′(cid:21)
(cid:20)δ′
Let
[β, γ]
0
M =n
α
η(cid:21)
(cid:20)δ
: α, . . . , η ∈ Zpo.
12
For
XIANG-DONG HOU
[β, γ]
0
A =
α
η(cid:21)
(cid:20)δ
, A′ =
[β′, γ′]
0
α′
η′(cid:21)
(cid:20)δ′
∈ M,
say A ∼ A′ if (3.12) is satisfied. It remains to determine the representatives of the
∼ equivalence classes in M.
In (3.12), we may assume W = [ 1 d
respectively. Let Y = [y1, y2], Z = [ z1
(3.13)
1 ] by replacing x, Y, W, Z with 1
z2 ]. Then (3.12) becomes
c x, 1
c Y, 1
c W, cZ,
xβ −xβd+xγ +y1 α + xβz1 + xγz2+y1δx−1 +y2ηx−1 +y1z2
0
0
δx−1 + dηx−1 + z2
ηx−1
0
0
=
β′ γ′ α′
0 δ′
0
0
0 η′
.
We claim that the ∼ equivalence classes in M are represented by
[0 0]
[0 0]
[1 0]
(1)
(2)
(3)
0
α
0(cid:21)
(cid:20)0
, α ∈ Zp,
1(cid:21)
(cid:20)0
,
η(cid:21)
(cid:20)0
,
η ∈ Zp.
0
The proof of (ii) will be complete when this claim is proved.
First, it is clear that matrices from different families of (1) -- (3) are not ∼ equiv-
alent. So it remains to show that every A ∈ M can be brought into one of the
"canonical forms" in (1) -- (3) through ∼ equivalence, and the entries in the canon-
ical form are uniquely determined by A.
Let
First assume β = 0 and η = 0. Then
α
[β, γ]
A =
η(cid:21)
(cid:20)δ
LHS of (3.13) =
(xγ + y1 = 0,
δx−1 + z2 = 0,
0
[0 0]
∈ M.
α′
0(cid:21)
(cid:20)0
α′ = α + γδ.
if and only if
(3.14)
and
(3.15)
System (3.14) has a solution (x, y1, z2) = (1, −γ, −δ). Equation (3.15) shows that
α′ is uniquely determined by A.
FINITE MODULES OVER Z[t, t−1]
13
Next, assume β = 0 and η 6= 0. Then
if and only if
(3.16)
[0 0]
LHS of (3.13) =
0
1(cid:21)
(cid:20)0
xγ + y1 = 0,
α + xγz2 + y1δx−1 + y2ηx−1 + y1z2 = 0,
δx−1 + dηx−1 + z2 = 0,
ηx−1 = 1.
System (3.16) has a solution (d, x, y1, y2, z2). (Let d be arbitrary. Solve for x form
the last equation, y1 from the first equation, z2 from the third equation, and y2
from the second equation.)
Finally, assume β 6= 0. Then
[1 0]
LHS of (3.13) =
0
η′(cid:21)
(cid:20) 0
xβ = 1,
−xβd + xγ + y1 = 0,
α + xβz1 + xγz2 + y1δx−1 + y2ηx−1 + y1z2 = 0,
δx−1 + dηx−1 + z2 = 0,
η′ = ηβ.
if and only if
(3.17)
and
(3.18)
System (3.17) has a solution (d, x, y1, y2, z1, z2). (Let y1 and y2 be arbitrary. Solve
for x from the first equation, d from the second equation, z2 from the last equa-
tion, and z1 from the third equation.) Equation (3.18) shows that η′ is uniquely
determined by A.
(cid:3)
3.7. Classification of Λ-modules of order pn, n ≤ 4.
The classification of Λ-modules of order pn, n ≤ 4, is obtained by combining
the results in 3.2 -- 3.6 and using Theorem 2.1. A complete description of the
classification is given in Table 1 in the appendix. From Table 1 we find that the
number of nonisomorphic Λ-modules of order pn is
(3.19)
1
p − 1
2p2 − p − 1
3p3 − 2p2 − 1
5p4 − 2p3 − 2p − 1
if n = 0,
if n = 1,
if n = 2,
if n = 3,
if n = 4.
14
XIANG-DONG HOU
4. Finite Alexander Quandles
By Theorem 1.1, the classification of finite Alexander quandles M is the same
as the classification of finite Λ-modules of the form (1 − t)M . First, the following
question has to be answered: Given a finite Λ-module N and an integer l > 0, does
there exist a Λ-module M ⊃ N such that (1 − t)M = N and M/N = l? Assume
that such an M exists. Since
M/(1 − t)M 1−t−→ (1 − t)M/(1 − t)2M = N/(1 − t)N
is an onto Λ-map, we have N/(1 − t)N (cid:12)(cid:12) l. We will see in Theorem 4.3 that
N/(1 − t)N (cid:12)(cid:12) l is also a sufficient condition for the existence of M .
Lemma 4.1. Let N ⊂ M be abelian groups and let α : N → N and α : M → N be
Z-maps such that αN = α. If 1 − α ∈ Aut(N ), then 1 − α ∈ Aut(M ).
Proof. We first show that 1 − α is 1-1. Let x ∈ ker(1 − α). Then
0 = α(0) = α(x − α(x)) = α(x) − α(α(x)) = (1 − α)(α(x)).
Since 1 − α ∈ Aut(N ), we have α(x) = 0. Thus x = [(1 − α) + α](x) = 0.
Now we show that 1 − α : M → M is onto. Let y ∈ M . Since α(y) ∈ N
and 1 − α ∈ Aut(N ), there exists x ∈ N such that (1 − α)(x) = α(y). Then
y = y − α(y) + x − α(x) = (1 − α)(y + x).
(cid:3)
Theorem 4.2. Let N be a finite abelian group and α ∈ EndZ(N ). Then there exist
a finite abelian group M ⊃ N with M/N = N/α(N ) and an onto homomorphism
α : M → N such that αN = α.
Proof. We may assume that N is a finite abelian p-group. If α(N ) = N , there is
nothing to prove. So assume α(N ) 6= N . Let N/α(N ) = pk and rank N = r (the
number of cyclic summands in a decomposition of N ). We will inductively construct
finite abelian p-groups N = M0 ⊂ M1 ⊂ · · · ⊂ Mk and Z-maps αi : Mi → N ,
0 ≤ i ≤ k, such that α0 = α, Mi+1/Mi = αi+1(Mi+1)/αi(Mi) = p, rank Mi = r,
and αi+1Mi = αi. Then M = Mk and α = αk have the desired property.
Mk
........................................
...........
...
........................................
...........
Mi+1
⊂
...........
....................................................................
Mi
........................................
...........
...
........................................
...........
N = M0
αi+1
............................................................................................................................................................................................................................................................................................................................................. ...........
........................................................................................................................................................................................................................................................................................ ...........
............................................................................................................................................................................................................................................... ...........
αi
α
N
FINITE MODULES OVER Z[t, t−1]
15
Let 0 ≤ i < k and assume that Mi and αi have been constructed. We now
construct Mi+1 and αi+1.
We claim that the mapping Mi/pMi → N/pN induced by αi is not 1-1. Oth-
erwise, since Mi/pMi = pr = N/pN , the mapping is also onto. Then for each
x ∈ N , there exist y0 ∈ Mi and x0 ∈ N such that
x = αi(y0) + px0.
In the same way, x0 = αi(y1) + px1 for some y1 ∈ Mi and x1 ∈ N . Continuing this
way, we can write
x = αi(y0 + py1 + · · · + pnyn) + pn+1xn,
y0, . . . , yn ∈ Mi, xn ∈ N.
Choose n large enough such that pn+1xn = 0. Then x = αi(y0 + py1 + · · · + pnyn) ∈
αi(Mi). So N = αi(Mi), which is a contradiction since αi(Mi)/α(N ) = pi < pk =
N/α(N ).
By the above claim, there exists a ∈ Mi \ pMi such that αi(a) ∈ pN . Write
αi(a) = pb for some b ∈ N .
Case 1. Assume b /∈ αi(Mi). Write Mi = Zpe1 × · · · × Zper , e1 ≥ · · · ≥ er > 0.
pes−1
pes−1 ,
Since a ∈ Mi \ pMi, we may assume a = (pw, 1, 0), where w ∈ Zpe1 × · · · × Z
pes+1 × B where A = Zpe1 × · · · × Z
for some 0 ≤ s < r. Let Mi+1 = A × Z
B = Z
pes+1 × · · · × Zper . Define Z-maps
ι : Mi = A × Zpes × B −→ Mi+1 = A × Z
pes +1 × B
(x, y, z)
7−→
(x, py, z)
αi+1 : Mi+1 = A × Z
pes +1 × B −→
N
(x, y, z)
7−→ αi(x, 0, z) + y(b − αi(w, 0, 0))
Then ι is 1-1, αi+1ι = αi, and Mi+1/Mi = p = αi+1(Mi+1)/αi(Mi).
Case 2. Assume b ∈ αi(Mi), say, b = αi(c), c ∈ Mi. Then αi(a) = pαi(c).
Let a1 = a − pc. Then a1 ∈ Mi \ pMi and αi(a1) = 0. Choose b′ ∈ N \ αi(Mi)
such that pb′ ∈ αi(Mi). Write pb′ = αi(d), d ∈ Mi. Choose ǫ = 0 or 1 such that
a′ := d + ǫa1 /∈ pMi. Then αi(a′) = pb′. Now we are in Case 1 with a′, b′ in place
of a, b, respectively.
(cid:3)
Theorem 4.3. Let N be a finite Λ-module with N/(1 − t)N = l. Then there
exists a finite Λ-module M ⊃ N with M/N = l and (1 − t)M = N .
Proof. Let α = 1−t ∈ EndZ(N ). By Theorem 4.2, there exist a finite abelian group
M ⊃ N and an onto Z-map α : M → N such that αN = α and M/N = N/α(N ).
Since 1−α ∈ AutZ(N ), by Lemma 4.1, 1−α ∈ AutZ(M ). Make M into a Λ-module
by defining
tx = (1 − α)(x),
x ∈ M.
Then ΛN is a submodule of ΛM and (1 − t)M = α(M ) = N .
(cid:3)
Corollary 4.4. Let p be a prime and n a positive integer. Let Mpn be a complete
set of nonisomorphic Λ-modules N such that N = pi, (1 − t)N = pj, 2i − j ≤ n.
For each N ∈ Mpn , let MN ⊃ N be a Λ-module with MN = pn and (1 − t)MN =
N . (The existence of MN is MN is guaranteed by Theorem 4.3.) Then {(MN , ∗) :
N ∈ Mpn } is a complete set of nonisomorphic Alexander quandles of order pn.
16
XIANG-DONG HOU
Proof. Given a Λ-module N with N = pi and (1 − t)N = pj, it follows from
Theorem 4.3 that n ≥ 2i − j is a necessary and sufficient condition on n for which
there exists a Λ-module M ⊃ N with M = pn and (1 − t)M = N . Now the
conclusion in the corollary follows from Theorem 1.1.
(cid:3)
The Λ-modules in Mpn , n ≤ 4, are contained in Table 1; to save space, we will
not enumerate these modules separately. From Table 1 we find that the number of
nonisomorphic Alexander quandles of order pn (n ≤ 4) is
(4.1)
(4.2)
Remark.
The number of nonisomorphic connected Alexander quandles of order pn (n ≤ 4),
also from Table 1, is
if n = 0,
1
if n = 1,
p − 1
2p2 − 2p − 1
if n = 2,
3p3 − 4p2 + p − 3
if n = 3,
5p4 − 6p3 + p2 − 6p − 1 if n = 4.
if n = 0,
1
if n = 1,
p − 2
2p2 − 3p − 1
if n = 2,
3p3 − 6p2 + p
if n = 3,
5p4 − 9p3 + p2 − 2p + 1 if n = 4.
(i) (4.1) agrees with the numbers of nonisomorphic Alexander quandles of order
≤ 15 in [8].
(ii) (4.2) with n = 2 agrees with the results of [3, 8]; (4.2) with pn = 24 agrees
with the number in [7].
(iii) [7] stated that the number of nonisomorphic Alexander quandles of order
16 is 24. Our result ((4.1) with pn = 24) is 23. It appears that the two
Alexander quandles in [7] with Im(Id − φ) = Z4 ⊕ Z2 (notation of [7]) are
isomorphic. Professor W. E. Clark computed the numbers of nonisomorphic
quandles of order pn < 28 using a computer program; his results (with
n ≤ 4) agree with (4.1).
Appendix: Table
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4
n = 0
(M, +) matrix of t
0
[0]
n = 1
(M, +) matrix of t
Zp
[b], b ∈ Z×
p
(1 − t)M number
1
1
total
1
(1 − t)M number
p0 if b = 1
p1 if b 6= 1
p − 2
1
total
p − 1
FINITE MODULES OVER Z[t, t−1]
17
Acknowledgment
I thank Professor W. Edwin Clark for sharing his computational results on the
numbers of nonisomorphic Alexander quandles.
References
[1] S. Carter, S. Kamada, M. Saito, Surfaces in 4-Space, Encyclopaedia of Mathematical Sciences,
142. Low-Dimensional Topology, III. Springer-Verlag, Berlin, 2004.
[2] W. E. Clark, M. Elhamdadi, X. Hou, M. Saito, T. Yeatman Connected quandles associated
with pointed abelian groups, preprint.
[3] M. Grana, Indecomposable racks of order p2, Beitrage Algebra Geom. 45 (2004), 665 - 676.
[4] C. J. Hillar and D. L. Rhea, Automorphisms of finite abelian groups, Amer. Math. Monthly
114 (2007), 917 - 923.
[5] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis, Cambridge University Press,
Cambridge, 1991.
[6] G. Murillo and S. Nelson, Alexander quandles of order 16, J. Knot Theory Ramifications 17
(2008), 273 - 278.
[7] G. Murillo and S. Nelson, Erratum: Alexander quandles of order 16, J. Knot Theory Rami-
fications 18 (2009), 727.
[8] S. Nelson, Classification of finite Alexander quandles, Topology Proc. 27 (2003), 245 - 258.
[9] A. Ranum, The group of classes of congruent matrices with application to the group of
isomorphisms of any abelian group, Trans. Amer. Math. Soc. 8 (1907), 71 - 91.
Department of Mathematics and Statistics, University of South Florida, Tampa, FL
33620
E-mail address: [email protected]
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4 (continued)
1
8
matrix of t
(1 − t)M
number
total
n = 2
(M, +)
Z
p2
[b], b ∈ Z×
p2
-
I
X
A
N
G
D
O
N
G
H
O
U
Z2
p
Z2
p
Z2
p
(cid:20)b
c(cid:21) , 0 < b ≤ c < p
(cid:20)b 1
b(cid:21) , b ∈ Z×
−b0 −b1(cid:21) , X 2 + b1X + b0 ∈ Zp[X] irr p2
(cid:20) 0
1
p
p0 if b = 1
p1 if b 6= 1, b ≡ 1 (p)
p2 if b 6≡ 1 (p)
p0 if b = c = 1
p1 if b = 1 < c
p2 if b > 1
p1 if b = 1
p2 if b 6= 1
1
p − 1
(p − 2)p
1
p − 2
(cid:0)p−1
2 (cid:1)
1
p − 2
p2 − p
(cid:0)p
2(cid:1)
p − 1
1
2 (p2 − p)
1
2 (p2 − p)
n = 3
(M, +)
matrix of t
(1 − t)M
Z
p3
[b], b ∈ Z×
p3
Z
p2 × Zp (cid:20)b
c(cid:21) , b ∈ Z×
p2 , c ∈ Z×
p
p0 if b = 1
p1 if b 6= 1, b ≡ 1 (p2)
p2 if b 6≡ 1 (p2), b ≡ 1 (p)
p3 if b 6≡ 1 (p)
p0 if b = 1, c = 1
p1 if b 6= 1, b ≡ 1 (p), c = 1
or b = 1, c 6= 1
p2 if b 6≡ 1 (p), c = 1
or b 6= 1, b ≡ 1 (p), c 6= 1
p3 if b 6≡ 1 (p), c 6= 1
number
1
p − 1
p(p − 1)
p2(p − 2)
1
2p − 3
total
p3 − p2
p(p − 1)2
(p − 2)(2p − 1)
p(p − 2)2
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4 (continued)
n = 3 (continued)
(M, +)
matrix of t
(1 − t)M
number
p1 if b = 1
p3 if b 6= 1
p1 if b = 1, γ = 0
p2 if b = 1, γ 6= 0
p3 if b 6= 1
p0 if b = c = d = 1
p1 if b = c = 1 < d
p2 if b = 1 < c
p3 if b > 1
p1 if b = c = 1
p2 if b = 1, c 6= 1
or b 6= 1, c = 1
p3 if b 6= 1 c 6= 1
p2 if b = 1
p3 if b 6= 1
1
p − 2
1
p − 1
p(p − 2)
1
p − 2
(cid:0)p−1
2 (cid:1)
(cid:0)p
3(cid:1)
1
2(p − 2)
(p − 2)2
1
p − 2
total
p − 1
p(p − 1)
(cid:0)p+1
3 (cid:1)
(p − 1)2
p − 1
F
I
N
I
T
E
M
O
D
U
L
E
S
O
V
E
R
Z
[
t
,
t
−
1
]
1
9
Z
p2 × Zp
Z
p2 × Zp
Z3
p
Z3
p
Z3
p
Z3
p
Z3
p
b
γ
1
(cid:20)b
(cid:20)b
0
0
0
0
c
p
, b, c ∈ Z×
p
, 0 < b ≤ c ≤ d < p
b(cid:21), 0 < b < p
b(cid:21), 0 < b < p, γ ∈ Zp
d
c
b
, b ∈ Z×
c
, X 2 + b1X + b0 ∈ Zp[X] irr, c ∈ Z×
0
1
1
0
1
p
b 1
b
b 1
b
1
−b0 −b1 −b2
−b0 −b1
, X 3 + b2X 2 + b1X + b0 ∈ Zp[X] irr p3
1
3 (p3 − p)
1
3 (p3 − p)
p
p2 if c = 1
p3 if c 6= 1
1
2 (p2 − p)
1
2 (p2 − p)(p − 2)
1
2 p(p − 1)2
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4 (continued)
n = 4
(M, +)
matrix of t
Z
p4
[b], b ∈ Z×
p4
Z
p3 × Zp
(cid:20)b
c(cid:21), b ∈ Z×
p3 , c ∈ Z×
p
Z
p3 × Zp
Z
p3 × Zp
0
1
b(cid:21), 0 < b < p2, b 6≡ 0 (p)
(cid:20)b
(cid:20)b p2
γ b(cid:21), 0 < b < p2, b 6≡ 0 (p), γ ∈ Zp
Z2
p2
Z2
p2
(cid:20)b
(cid:20)b
0
c(cid:21), 0 < b ≤ c < p2
p
b(cid:21) , b ∈ Z×
p2
(1 − t)M
p0 if b = 1
p1 if b 6= 1, b ≡ 1 (p3)
p2 if b 6≡ 1 (p3), b ≡ 1 (p2)
p3 if b 6≡ 1 (p2), b ≡ 1 (p)
p4 if b 6≡ 1 (p)
p0 if b = 1, c = 1
p1 if b 6= 1, b ≡ 1 (p2), c = 1
or b = 1, c 6= 1
p2 if b 6≡ 1 (p2), b ≡ 1 (p), c = 1
or b 6= 1, b ≡ 1 (p2), c 6= 1
p3 if b 6≡ 1 (p), c = 1
or b 6≡ 1 (p2), b ≡ 1 (p), c 6= 1
p4 if b 6≡ 1 (p), c 6= 1
p1 if b = 1
p2 if b 6= 1, b ≡ 1 (p)
p4 if b 6≡ 1 (p)
p1 if b = 1, γ = 0
p2 if b 6= 1, b ≡ 1 (p)
or b = 1, γ 6= 0
p4 if b 6≡ 1 (p)
p0 if b = c = 1
p1 if b = 1 < c, c ≡ 1 (p)
p2 if b = 1, c 6≡ 1 (p)
or b, c 6= 1, b, c ≡ 1 (p)
p3 if {b, c} = {b1, c1}, b1 6≡ 1 (p), c1 6= 1, c1 ≡ 1 (p)
p4 if b, c 6≡ 1 (p)
p1 if b = 1
p2 if b 6= 1, b ≡ 1 (p)
p4 if b 6≡ 1 (p)
2
0
total
p3(p − 1)
number
1
p − 1
p(p − 1)
p2(p − 1)
p3(p − 2)
1
2p − 3
-
I
X
A
N
G
D
O
N
G
H
O
U
2(p − 1)2
p2(p − 1)2
p(p − 2)(2p − 1)
p2(p − 2)2
1
p − 1
p(p − 2)
1
p2 − 1
p2(p − 2)
1
p − 1
p(p − 1)(p − 2)
p(p − 2) +(cid:0)p
2(cid:1)
(cid:0)p(p−2)+1
(cid:1)
2
1
p − 1
p(p − 2)
p(p − 1)
p2(p − 1)
2
(cid:0)p(p−1)+1
(cid:1)
p(p − 1)
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4 (continued)
n = 4 (continued)
(M, +)
matrix of t
(1 − t)M
Z2
p2
Z2
p2
Z2
p2
0 < b < p,
X 2 + b1X + b0 ∈ Zp[X] irr
p
−pb0
b − pb1(cid:21),
(cid:20) b
(cid:20)b + pα 1
b(cid:21), 0 < b < p, α, γ ∈ Zp
−b1 (cid:21),
(cid:20) pα 1 + pβ
−b0
pγ
α, β ∈ Zp, 0 ≤ b0, b1 < p
X 2 + b1X + b0 ∈ Zp[X] irr
Z
p2 × Z2
b
p
c
, b ∈ Z×
p2 , 0 < c ≤ d < p
d
Z
p2 × Z2
Z
p2 × Z2
Z
p2 × Z2
b
1
0
b
0
η
b
1
0
p
p
p
0 0
b
0
0
p 0
b
0
0
p 0
b
0
0
p
, 0 < b < p, η ∈ Zp
c
, 0 < b < p, c ∈ Z×
b
b
, 0 < b < p
p2 if b = 1
p4 if b 6= 1
p2 if b = 1, γ = 0
p3 if b = 1, γ 6= 0
p4 if b 6= 1
p4
p0 if b = 1, c = d = 1
p1 if b 6= 1, b ≡ 1 (p), c = d = 1
or b = 1, c = 1 < d
p2 if b 6≡ 1 (p), c = d = 1
or b 6= 1, b ≡ 1 (p), c = 1 < d
or b = 1, c > 1
p3 if b 6≡ 1 (p), c = 1 < d
or b 6= 1, b ≡ 1 (p), c > 1
p4 if b 6≡ 1 (p), c > 1
p1 if b = 1, c = 1
p2 if b = 1, c 6= 1
p3 if b 6= 1, c = 1
p4 if b 6= 1, c 6= 1
p1 if b = 1, η = 0
p2 if b = 1, η 6= 0
p4 if b 6= 1
p2 if b = 1
p4 if b 6= 1
number
1
2 (p2 − p)
1
2 (p2 − p)(p − 2)
p
p(p − 1)
p2(p − 2)
total
1
2 p(p − 1)2
p2(p − 1)
1
2 p2(p2 − p)
1
2 p2(p2 − p)
1
2p − 3
1
2 (p − 2)(5p − 3)
1
2 p2(p − 1)2
1
2 (p − 2)(3p2 − 6p + 1)
1
2 p(p − 1)(p − 2)2
1
F
I
N
I
T
E
M
O
D
U
L
E
S
O
V
E
R
Z
[
t
,
t
−
1
]
p − 2
p − 2
(p − 2)2
1
p − 1
p(p − 2)
1
p − 2
(p − 1)2
p(p − 1)
p − 1
2
1
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4 (continued)
n = 4 (continued)
(M, +)
matrix of t
Z
p2 × Z2
c
, b ∈ Z×
p2 , c ∈ Z×
p
Z
p2 × Z2
, 0 < b < p
Z
p2 × Z2
, 0 < b < p, η ∈ Zp
b
b
0
1
b
0
η
b
b
b
p
p
p
1
c
b
b
0 0
b
1
0
p 0
b
1
0
c
1
b
1
b
c
c
d
1
c
Z4
p
Z4
p
Z4
p
, 0 < b ≤ c ≤ d ≤ e < p
d
e
, b ∈ Z×
p , 0 < c ≤ d < p
, 0 < b ≤ c < p
2
2
-
I
X
A
N
G
D
O
N
G
H
O
U
(1 − t)M
p1 if b = 1, c = 1
p2 if b 6= 1, b ≡ 1 (p), c = 1
or b = 1, c 6= 1
p3 if b 6≡ 1 (p), c = 1
or b 6= 1, b ≡ 1 (p), c 6= 1
p4 if b 6≡ 1 (p), c 6= 1
number
1
2p − 3
(p − 2)(2p − 1)
p(p − 2)2
p2 if b = 1
p4 if b 6= 1
p2 if b = 1, η = 0
p3 if b = 1, η 6= 0
p4 if b 6= 1
p0 if b = c = d = e = 1
p1 if b = c = d = 1 < e
p2 if b = c = 1 < d
p3 if b = 1 < c
p4 if b > 1
p1 if b = 1, c = d = 1
p2 if b 6= 1, c = d = 1
or b = 1, c = 1 < d
p3 if b 6= 1, c = 1 < d
or b = 1, c > 1
p4 if b 6= 1, c > 1
p2 if b = c = 1
p3 if b = 1 < c
p4 if b > 1
1
p − 2
1
p − 1
p(p − 2)
1
p − 2
(cid:0)p−1
2 (cid:1)
(cid:0)p
3(cid:1)
(cid:0)p+1
4 (cid:1)
1
2(p − 2)
1
2 (p − 2)(3p − 5)
1
2 (p − 1)(p − 2)2
1
p − 2
(cid:0)p−1
2 (cid:1)
total
p(p − 1)2
p − 1
p(p − 1)
(cid:0)p+2
4 (cid:1)
1
2 p(p − 1)2
(cid:0)p
2(cid:1)
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4 (continued)
n = 4 (continued)
(M, +)
matrix of t
Z4
p
Z4
p
Z4
p
Z4
p
Z4
p
c
1
b
b 1
b
b 1
b
1
b
1
b
0
1
−b0 −b1
0
1
−b0 −b1
, b, c ∈ Z×
p
, b ∈ Z×
p
0
1
−c0 −c1
1
0
0
0
1
1
−b0 −b1
0
0
0
1
0
0
0
1
0
0
0
1
−b0 −b1 −b2 −b3
(*) ≤ is a total order in Z2
p
,
X 2 + b1X + b0, X 2 + c1X + c0 ∈ Zp[X] irr
(b0, b1) ≤ (c0, c1) (*)
, X 2 + b1X + b0 ∈ Zp[X] irr
, X 4 + b3X 3 + b2X 2 + b1X + b0 ∈ Zp[X] irr
(1 − t)M
p2 if b = c = 1
p3 if b = 1, c 6= 1
or b 6= 1, c = 1
p4 if b 6= 1, c 6= 1
p3 if b = 1
p4 if b 6= 1
number
total
1
2(p − 2)
(p − 2)2
1
p − 2
(p − 1)2
p − 1
p4
p4
p4
1
2 (p2−p)+1
2
(cid:1)
(cid:0)
(cid:0)
1
2 (p2−p)+1
2
(cid:1)
1
2 (p2 − p)
1
2 (p2 − p)
1
4 (p4 − p2)
1
4 (p4 − p2)
F
I
N
I
T
E
M
O
D
U
L
E
S
O
V
E
R
Z
[
t
,
t
−
1
]
2
3
Table 1. Nonisomorphic Λ-modules of order pn, n ≤ 4 (continued)
2
4
n = 4 (continued)
(M, +)
matrix of t
, 0 < b < p, γ ∈ Zp, c ∈ Z×
p , b 6≡ c (p)
p
b
c
1
0
0
0
0
1
−b0 −b1 −b2
,
X 3 +b2X 2+b1X +b0 ∈ Zp[X] irr,
c ∈ Z×
p
Z
p2 × Z2
b
γ
b
p
p
Z4
p
Z4
p
Z4
p
c
d
0
1
−c0 −c1
0
1
−b0 −b1
b 1
b
0
1
−c0 −c1
Z
p2 × Z2
, b ∈ Z×
p2 , X 2 + c1X + c0 ∈ Zp[X] irr
,
X 2 + b1X + b0 ∈ Zp[X] irr,
0 < c ≤ d < p
c
, b ∈ Z×
p , X 2 + c1X + c0 ∈ Zp[X] irr
(1 − t)M
p2 if b = 1, γ = 0, c 6= 1
p3 if b 6= 1, c = 1
or b = 1, γ 6= 0, c 6= 1
p4 if b 6= 1, c 6= 1
p3 if c = 1
p4 if c 6= 1
p2 if b = 1
p3 if b 6= 1, b ≡ 1 (p)
p4 if b 6≡ 1 (p)
p2 if c = d = 1
p3 if c = 1 < d
p4 if c > 1
p3 if b = 1
p4 if b 6= 1
number
p − 2
total
(p − 2)(2p − 1)
p(p − 1)(p − 2)
p(p − 2)(p − 3)
1
3 (p3 − p)
1
3 (p3 − p)(p − 2)
1
2 (p2 − p)
1
2 (p2 − p)(p − 1)
1
2 (p2 − p)p(p − 2)
1
2 (p2 − p)
1
2 (p2 − p)(p − 2)
1
2 (p2 − p)(cid:0)p−1
2 (cid:1)
2 (p2 − p)
1
1
2 (p2 − p)(p − 2)
1
3 (p3 − p)(p − 1)
1
2 p2(p − 1)2
1
4 p2(p − 1)2
1
2 p(p − 1)2
-
I
X
A
N
G
D
O
N
G
H
O
U
|
1504.06013 | 2 | 1504 | 2015-05-05T04:53:05 | A note on standard equivalences | [
"math.RA",
"math.RT"
] | We prove that any derived equivalence between triangular algebras is standard, that is, it is isomorphic to the derived tensor functor given by a two-sided tilting complex. | math.RA | math |
A NOTE ON STANDARD EQUIVALENCES
XIAO-WU CHEN
Abstract. We prove that any derived equivalence between triangular algebras
is standard, that is, it is isomorphic to the derived tensor functor given by a
two-sided tilting complex.
1. Introduction
Let k be a field. We will require that all categories and functors are k-linear.
Let A be a finite dimensional k-algebra. We denote by A-mod the category of finite
dimensional left A-modules and by Db(A-mod) its bounded derived category.
Let B be another finite dimensional k-algebra. We will require that k acts cen-
trally on any B-A-bimodule. Recall that a two-sided tilting complex is a bounded
complex X of B-A-bimodules such that the derived tensor functor gives an equiv-
alence X ⊗L
A − : Db(A-mod) → Db(B-mod).
A triangle equivalence F : Db(A-mod) → Db(B-mod) is said to be standard if it
is isomorphic, as a triangle functor, to X ⊗L
A − for some two-sided tilting complex
X. It is an open question whether all triangle equivalences are standard. The aim
of this note is to answer this question affirmatively in a special case.
Recall that the algebra A is triangular provided that the Ext-quiver of A has no
oriented cycles. There are explicit examples of algebras A and B, which are derived
equivalent such that A is triangular, but B is not; consult the top of [2, p.21]. It
makes sense to have the following notion: an algebra A is derived-triangular if it is
derived equivalent to a triangular algebra.
Theorem 1.1. Let A be a derived-triangular algebra. Then any triangle equivalence
F : Db(A-mod) → Db(B-mod) is standard.
We observe that a derived-triangular algebra has finite global dimension. The
converse is not true in general. Indeed, let A be a non-triangular algebra with two
simple modules which has finite global dimension. Then A is not derived-triangular.
Indeed, any triangular algebra B which is derived equivalent to A has two simple
modules and thus is hereditary. This forces that the algebra A is triangular, yielding
a contradiction.
We recall that a piecewise hereditary algebra is triangular. In particular, The-
orem 1.1 implies that the assumption on the standardness of the auto-equivalence
in [3, Section 4] is superfluous.
The proof of Theorem 1.1 is a rather immediate application of [1, Theorem
4.7], which characterizes certain triangle functors between the bounded homotopy
categories of Orlov categories. Here, we observe that the category of projective
modules over a triangular algebra is naturally an Orlov category.
We refer to [4, 5] for unexplained notions in the representation theory of algebras.
Date: August 26, 2018.
2010 Mathematics Subject Classification. 13D09, 16G10, 18E30.
Key words and phrases. tilting complex, standard equivalence, Orlov category, homotopy cat-
egory, triangle functor.
E-mail: [email protected].
1
2
XIAO-WU CHEN
2. The bounded homotopy category of an Orlov category
Let A be a k-linear additive category, which is Hom-finite and has split idempo-
tents. Here, the Hom-finitess means that all the Hom spaces are finite dimensional.
It follows that A is a Krull-Schmidt category.
We denote by Ind A a complete set of representatives of indecomposable ob-
jects in A. The category A is called bricky if the endomorphism algebra of each
indecomposable object is a division algebra.
We slightly generalize [1, Definition 4.1]. A bricky category A is called an Orlov
category provided that there is a degree function deg : Ind A → Z with the following
property: for any indecomposable objects P, P ′ having HomA(P, P ′) 6= 0, we have
that P ≃ P ′ or deg(P ) > deg(P ′). An object X in A is homogeneous of degree n, if
it is isomorphic to a finite direct sum of indecomposables of degree n. An additive
functor F : A → A is homogeneous if it sends homogeneous objects to homogenous
objects and preserves their degrees.
Let A be a finite dimensional k-algebra. We denote by {S1, S2, · · · , Sn} a com-
plete set of representatives of simple A-modules. Denote by Pi the projective cover
of Si. We recall that the Ext-quiver QA of A is defined as follows. The vertex set
of QA equals {1, 2, · · · , n}, and there is a unique arrow from i to j provided that
Ext1
A(Si, Sj) 6= 0. The algebra A is triangular provided that QA has no oriented
cycles.
Let A be a triangular algebra. We denote by Q0
A inductively, such that a vertex i belongs to Qd
A the set of sources in QA. Here, a
vertex is a source if there is no arrow ending at it. For each d ≥ 1, we define the set
Qd
A if and only if any arrow ending
at i necessarily starts at S0≤m≤d−1 Qm
A ⊆ · · · and
that Sd≥0 Qd
A = {1, 2, · · · , n}. We mention that this construction can be found in
[4, p.42].
A . It follows that Q0
A ⊆ Q1
A ⊆ Q2
We denote by A-proj the category of finite dimensional projective A-modules.
Then {P1, P2, · · · , Pn} is a complete set of representatives of indecomposables in
A and i /∈ Qd−1
A-proj. For each 1 ≤ i ≤ n, we define deg(Pi) = d such that i ∈ Qd
A .
The following example of an Orlov category seems to be well known.
Lemma 2.1. Let A be a triangular algebra. Then A-proj is an Orlov category
with the above degree function. Moreover, any equivalence F : A-proj → A-proj is
homogeneous.
Proof. Since A is triangular,
it is well known that EndA(Pi) is isomorphic to
EndA(Si), which is a division algebra. Then A-proj is bricky. We recall that
for i 6= j with HomA(Pi, Pj) 6= 0, there is a path from j to i in QA. From the very
A, we have a ∈ Qd−1
construction, we infer that for an arrow α : a → b with b ∈ Qd
A .
Then we are done by the following consequence: if there is a path from j to i in
QA, then deg(Pj ) < deg(Pi).
For the final statement, we observe that the equivalence F extends to an auto-
equivalence on A-mod, and thus induces an automorphism of QA. The automor-
phism preserves the subsets Qd
A. Consequently, the equivalence F preserves degrees,
and is homogeneous.
(cid:3)
Let A be a k-linear additive category as above. We denote by Kb(A) the ho-
motopy category of bounded complexes in A. Here, a complex X is visualized as
dn−1
X→ X n dn
· · · → X n−1
X = 0.
The translation functor on Kb(A) is denoted by [1], whose n-th power is denoted
by [n].
X→ X n+1 → · · · , where the differentials satisfy dn
X ◦ dn−1
We view an object A in A as a stalk complex concentrated at degree zero, which
is still denoted by A. In this way, we identify A as a full subcategory of Kb(A).
A NOTE ON STANDARD EQUIVALENCES
3
We are interested in triangle functors on Kb(A). We recall that a triangle functor
(F, θ) consists of an additive functor F : Kb(A) → Kb(A) and a natural isomor-
phism θ : [1]F → F [1], which preserves triangles. More precisely, for any triangle
X → Y → Z h→ X[1] in Kb(A), the sequence F X → F Y → F Z
−→ (F X)[1] is
a triangle. We refer to θ as the commutating isomorphism for F . A natural trans-
formation between triangle functors is required to respect the two commutating
isomorphisms.
θX ◦F (h)
For a triangle functor (F, θ), the commutating isomorphism θ is trivial if [1]F =
F [1] and θ is the identity transformation. In this case, we suppress θ and write F
for the triangle functor.
Any additive functor F : A → A gives rise to a triangle functor Kb(F ) : Kb(A) →
Kb(A), which acts on complexes componentwise. The commuting isomorphism for
Kb(F ) is trivial. Similarly, any natural transformation η : F → F ′ extends to a
natural transformation Kb(η) : Kb(F ) → Kb(F ′) between triangle functors.
The following fundamental result is due to [1, Theorem 4.7].
Proposition 2.2. Let A be an Orlov category, and let (F, θ) : Kb(A) → Kb(A) be
a triangle functor such that F (A) ⊆ A. We assume further that F A : A → A is
homogeneous. Let F1, F2 : A → A be two homogeneous functors.
(1) There is a unique natural isomorphism (F, θ) → Kb(F A) of triangle func-
tors, which is the identity on the full subcategory A.
(2) Any natural transformation Kb(F1) → Kb(F2) of triangle functors is of the
form Kb(η) for a unique natural transformation η : F1 → F2.
Proof. The existence of the natural isomorphism in (1) is due to [1, Theorem 4.7];
compare [1, Remark 4.8]. The uniqueness follows from the commutative diagram
(4.10) and Lemma 4.5(2) in [1], by induction on the support of a complex in the
sense of [1, Subsection 4.1]. Here, we emphasize that the commutating isomor-
phism θ is used in the construction of the natural isomorphism on stalk complexes;
compare the second paragraph in [1, p.1541].
The statement (2) follows from the uniqueness part of (1).
(cid:3)
Recall that Db(A-mod) denotes the bounded derived category of A-mod. We
identify A-mod as the full subcategory of Db(A-mod) formed by stalk complexes
concentrated at degree zero. We denote by H n(X) the n-th cohomology of a com-
plex X.
The following observation is immediate.
Lemma 2.3. Let A be a finite dimensional algebra, and let F : Db(A-mod) →
Db(A-mod) be a triangle equivalence with F (A) ≃ A. Then we have F (A-mod) =
A-mod, and thus the restricted equivalence F A-mod : A-mod → A-mod.
Proof. We use the canonical isomorphisms H n(X) ≃ HomDb(A-mod)(A[−n], X). It
follows that both F and its quasi-inverse send stalk complexes to stalk complexes.
Then we are done.
(cid:3)
We assume that we are given an equivalence F : A-mod → A-mod with F (A) ≃
A. Then there is an algebra automorphism σ : A → A such that F is isomorphic to
σA1 ⊗A −. Here, the A-bimodule σA1 is given by the regular right A-module, where
the left A-module is twisted by σ. This bimodule is invertible and thus viewed as
a two-sided tilting complex. We refer to [5, Section 6.5] for details on two-sided
tilting complexes and standard equivalences.
We now combine the above results.
4
XIAO-WU CHEN
Proposition 2.4. Let A be a triangular algebra, and let (F, θ) : Db(A-mod) →
Db(A-mod) be a triangle equivalence with F (A) ≃ A. We recall the algebra auto-
morphism σ given by the restricted equivalence F A-mod, and the A-bimodule σA1.
Then there is a natural isomorphism (F, θ) → σA1 ⊗L
A − of triangle functors. In
particular, the triangle equivalence (F, θ) is standard.
Proof. Since the algebra A is triangular, it has finite global dimension. The natural
functor Kb(A-proj) → Db(A-mod) is a triangle equivalence. We identify these
two categories. Therefore, the triangle functor (F, θ) : Kb(A-proj) → Kb(A-proj)
restricts to an equivalence F A-proj, which is isomorphic to σA1 ⊗A −. By Lemma
2.1, the statements in Proposition 2.2 apply in our situation. Consequently, we
have an isomorphism between (F, θ) and Kb(σA1 ⊗A −). Then we are done.
(cid:3)
3. The proof of Theorem 1.1
We now prove Theorem 1.1.
In what follows, for simplicity, when writing a
triangle functor we suppress its commutating isomorphism.
We first assume that the algebra A is triangular. The complex F (A) is a one-
sided tilting complex. By [5, Theorem 6.4.1], there is a two-sided tilting complex
X of B-A-bimodules with an isomorphism X → F (A) in Db(B-mod). Denote by
A − : Db(A-mod) → Db(B-mod).
G a quasi-inverse of the standard equivalence X ⊗L
Then the triangle functor GF : Db(A-mod) → Db(A-mod) satisfies GF (A) ≃ A.
Proposition 2.4 implies that GF is standard, and thus F is isomorphic to the
composition of X ⊗L
A − and a standard equivalence. Then we are done in this
case by the well-known fact that the composition of two standard equivalences is
standard.
In general, let A be derived-triangular. Assume that A′ is a triangular algebra
which is derived equivalent to A. By [5, Proposition 6.5.5], there is a standard
equivalence F ′ : Db(A′-mod) → Db(A-mod). The above argument implies that
the composition F F ′ is standard. Recall from [5, Proposition 6.5.6] that a quasi-
inverse F ′−1 of F ′ is standard. We are done by observing that F is isomorphic to
the composition (F F ′)F ′−1, a composition of two standard equivalences.
Acknowledgements We thank Martin Kalck for pointing out the example in
[2]. The author is supported by National Natural Science Foundation of China
(No. 11201446), NCET-12-0507, and the Fundamental Research Funds for the
Central Universities.
References
[1] P.N. Achar, and S. Riche, Koszul duality and semisimplicity of Frobenius, Ann. Inst.
Fourier 63(4) (2013), 1511 -- 1612.
[2] G. Bonbinski, C. Geiss, and A. Skowronski, Classification of derived discrete algebras,
Cent. Euro. J. Math. 2(1) (2004), 19 -- 49.
[3] B. Keller, On triangulated orbit categories, Doc. Math. 10 (2005), 551 -- 581.
[4] C.M. Ringel, Tame Algebras and Integral Quadratic Forms, Lect. Notes Math. 1099,
Springer-Verlag, Berlin Heidelberg New York Tokyo, 1984.
[5] A. Zimmermann, Representation Theory, A Homological Algebra Point of View, Springer,
International Publishing Switzerland, 2014.
Xiao-Wu Chen
Key Laboratory of Wu Wen-Tsun Mathematics, Chinese Academy of Sciences
School of Mathematical Sciences, University of Science and Technology of China
No. 96 Jinzhai Road, Hefei, 230026, Anhui, P.R. China.
URL: http://home.ustc.edu.cn/∼ xwchen.
|
1602.08731 | 1 | 1602 | 2016-02-28T15:39:03 | Yetter-Drinfeld modules for weak Hom-Hopf algebras | [
"math.RA",
"math-ph",
"math-ph"
] | The aim of this paper is to define and study Yetter-Drinfeld modules over weak Hom-Hopf algebras. We show that the category ${}_H{\cal WYD}^H$ of Yetter-Drinfeld modules with bijective structure maps over weak Hom-Hopf algebras is a rigid category and a braided monoidal category, and obtain a new solution of quantum Hom-Yang-Baxter equation. It turns out that, If $H$ is quasitriangular (respectively, coquasitriangular)weak Hom-Hopf algebras, the category of modules (respectively, comodules) with bijective structure maps over $H$ is a braided monoidal subcategory of the category ${}_H{\cal WYD}^H$ of Yetter-Drinfeld modules over weak Hom-Hopf algebras. | math.RA | math |
Yetter-Drinfeld modules for weak Hom-Hopf
algebras
Shuangjian Guo1, Yizheng Li1, Shengxiang Wang2
1. School of Mathematics and Statistics, Guizhou University of Finance and Economics
Guiyang, 550025, P. R. China
Email: [email protected], Email: [email protected]
2. School of Mathematics and Finance, Chuzhou University
Chuzhou 239000, P. R. China
E-mail: [email protected]
ABSTRACT
The aim of this paper is to define and study Yetter-Drinfeld modules
over weak Hom-Hopf algebras. We show that the category HWYDH of
Yetter-Drinfeld modules with bijective structure maps over weak Hom-
Hopf algebras is a rigid category and a braided monoidal category, and
obtain a new solution of quantum Hom-Yang-Baxter equation. It turns
out that, If H is quasitriangular (respectively, coquasitriangular)weak
Hom-Hopf algebras, the category of modules (respectively, comodules)
with bijective structure maps over H is a braided monoidal subcategory
of the category H WYDH of Yetter-Drinfeld modules over weak Hom-
Hopf algebras.
Key words: Yetter-Drinfeld module, braided monoidal category,
(co)quasitriangular, weak-Hom type entwined-module.
Mathematics Subject Classification: 16W30, 16T15.
0. Introduction
The first examples of Hom-type algebras were related to q-deformations of Witt and
Virasoro algebras, which play an important role in Physics, mainly in conformal field
theory. The q-deformations of Witt and Virasoro algebras are obtained when the deriva-
tion is replaced by a σ-derivation. It was observed in the pioneering works (See [5]-[8]).
Motivated by these examples and their generalization, Hartwig, Larsson and Silvestrov
introduced the Hom-Lie algebras when they concerned about the q-deformations of Witt
and Virasoro algebras in [4]. In a Hom-Lie algebra, the Jacobi identity is replaced by the
1
so called Hom-Jacobi identity via an homomorphism. Hom-associative algebras, the cor-
responding structure of associative algebras, were introduced by Makhlouf and Silvestrov
in [12]. The associativity of the Hom-algebra is twisted by an endomorphism (here we call
it the Hom structure map). The generalized notions, Hom-bialgebras, Hom-Hopf alge-
bras were developed in [13], [14]. Caenepeel and Goyvaerts studied in [2] Hom-bialgebras
and Hom-Hopf algebras from a categorical view point, and called them monoidal Hom-
bialgebras and monoidal Hom-Hopf algebras respectively, which are slightly different from
the above Hom-bialgebras and Hom-Hopf algebras. Thus a monoidal Hom-bialgebra is
Hom-bialgebra if and only if the Hom-structure map α satisfies α2 = id. Yau introduced
Quasitriangular Hom-bialgebras in [18]), which provided a solution of the quantum Hom-
Yang-Baxter euqation, a twisted version of the quantum Yang-Baxter equation called the
Hom-Yang-Baxter equation in [19]. Zhang and Wang introduced weak Hom-Hopf algebra
H, which is generalization of both Hom-Hopf algebras and weak Hopf algebras, and dis-
cussed the category Rep(H) (resp. Corep(H)) of Hom-modules (resp. Hom-comodules)
with bijective Hom-structure maps, they proved that if H is a (co)quasitrialgular weak
Hom-bialgebra (resp. ribbon weak Hom-Hopf algebra), then Rep(H) (resp. Corep(H)) is
a braided monoidal category (resp. ribbon category) in [16].
Makhlouf and Panaite defined and studied Yetter-Drinfeld modules over Hom bialge-
bras, a generalized version of bialgebras obtained by modifying the algebra and coalgebra
structures by a homomorphism. Yetter-Drinfeld modules over a Hom bialgebra with bi-
jective structure map provide solutions of the Hom-Yang-Baxter equation in [10]. It is
well known that the Yetter-Drinfeld modules category of a (weak) Hopf algebra is a rigid
monoidal category, and is braided. Does this result remain true in a weak Hom-Hopf
algebra? How the corresponding results appear under the condition that the associativity
and coassociativity are twisted by an endomorphism? Is there any relation between this
Yetter-Drinfeld modules category and module category or comodule category? This is the
motivation of the present article.
In order to investigate these questions, we introduce
the definition of Yetter-Drinfeld modules over weak Hom-Hopf algebras, which is general-
ization of both weak Yetter-Drinfeld modules introduced by [3] or [15] and Hom-Yetter-
Drinfeld modules introduced by [10] or [11], and consider that when the Yetter-Drinfeld
modules category of a weak Hom-Hopf algebra is is a rigid monoidal category, and is
braided.
To make sure that the Yetter-Drinfeld modules category of a weak Hom-Hopf algebra
H is a monoidal category, we need H is unital and counital, and the Hom structure maps
over H are all bijective maps.
The paper is organized as follows. In Section 2, we recall now several concepts and
results, fixing thus the terminology to be used in the rest of the paper.
In Section 3, we introduce the definition of Yetter-Drinfeld modules over weak Hom-
Hopf algebras and show the category HWYDH of Yetter-Drinfeld modules is a monoidal
category and a rigid category.
In Section 4, we show that the category HWYDH of Yetter-Drinfeld modules is a
braided monoidal category and obtain a new solution of quantum Hom-Yang-Baxter equa-
tion. It turns out that, if H is a quasitriangular weak Hom-Hopf algebra, the category
of left H-modules with bijective structure maps is a braided monoidal subcategory of the
category H WYDH of Yetter-Drinfeld modules.
In Section 5, we find another braided monoidal category structure on the category
2
HWYDH of Yetter-Drinfeld modules , with the property that if H is a coquasitriangular
weak Hom-Hopf algebra, then HWYDH contains the category of right H-comodules with
bijective structure maps as a braided monoidal category.
2. Preliminaries
Throughout the paper, we let k be a fixed field and all algebras are supposed to be
over k. For the comultiplication ∆ of a vector space C, we use the Sweedler-Heyneman's
notation:
∆(c) = c1 ⊗ c2,
for any c ∈ C. τ means the flip map τ (a ⊗ b) = b ⊗ a. When we say a "Hom-algebra" or
a "Hom-coalgebra", we mean the unital Hom-algebra and counital Hom-coalgebra.
In this section, we will review several definitions and notations related to weak Hom-
Hopf algebras and rigid categories.
2.1. Hom-algebras and Hom-coalgebras.
Recall from [12] that a Hom-associative algebra is a quadruple (A, µ, η, αA), in which A
is a linear space, αA : A → A, µ : A⊗ A → A and η : k → A are linear maps, with notation
µ(a ⊗ b) = ab and η(1k) = 1A, satisfying the following conditions, for all a, b, c ∈ A:
(1) αA(ab) = αA(a)α(b);
(2) αA(a)(bc) = (ab)αA(c);
(3) αA(1A) = 1A;
(4) 1Aa = a1A = αA(a).
A morphism f : A → B of Hom-algebras is a linear map such that αB ◦ f = f ◦ αA,
f (1A) = 1B and µB ◦ (f ⊗ f ) = f ◦ µA.
Let A be a Hom-algebra. Recall that a left A-module is a triple (M, αM , θM ), where
M is a k-space, αM : M → M and θM : A ⊗ M → M are linear maps with notation
θM (a ⊗ m) = a · m, satisfying the following conditions, for all a, b ∈ A, m ∈ M :
(1) α(a · m) = α(a) · αM (m);
(2) α(a) · (b · m) = (ab) · αM (m);
(3) 1A · m = αM (m).
A morphism f : M → N of A-modules is a linear map such that αN ◦ f = f ◦ αM and
θN ◦ (idA ⊗ f ) = f ◦ θM .
Recall from [14] that a Hom-coassociative coalgebra is a quadruple (C, ∆, ε, αC ), in
which C is a linear space, α : C → C, ∆ : C → C ⊗ C and ε : C → k are linear maps,
with notation ∆(c) = c1 ⊗ c2, satisfying the following conditions for all c ∈ C:
(1) ∆(αC(c)) = αC(c1)α(c2);
(2) αC(c1) ⊗ ∆(c2) = ∆(c1) ⊗ αC(c2);
(3) ε ◦ αC = ε;
(4) ε(c1)c2 = c1ε(c2) = αC(c).
A morphism f : C → D of Hom-coalgebras is a linear map such that αD ◦ f = f ◦ αC,
εC = εD ◦ f and ∆D ◦ f = (f ⊗ f ) ◦ ∆C.
3
Let C be a Hom-coalgebra. Recall that a right C-comodule is a triple (M, αM , ρM ),
where M is a k-space, αM : M → M and ρM : M → M ⊗ C are linear maps with notation
ρM (m) = m(0) ⊗ m(1), satisfying the following conditions for all m ∈ M :
(1) ρM (αM (m)) = αM (m(0)) ⊗ αC(m(1));
(2) αM (m(0)) ⊗ ∆(m(1)) = ρM (m(0)) ⊗ αC(m(1));
(3) ε(m1)m0 = αM (m).
A morphism f : M → N of C-comodules is a linear map such that αN ◦ f = f ◦ αM
and ρN ◦ f = (idC ⊗ f ) ◦ ρM .
Recall from [16] that a weak Hom-bialgebra is a sextuple H = (H, αH , µ, η, ∆, ε) if
(H, αH ) is both a Hom-algebra and a Hom-coalgebra, satisfying the following identities
for any a, b, c ∈ H:
(1) ∆(ab) = ∆(a)∆(b);
(2) ε((ab)c) = ε(ab1)ε(b2c), ε(a(bc)) = ε(ab2)ε(b1c);
(3) (∆ ⊗ idH )∆(1H ) = 11 ⊗ 121′
Recall from [16] that a Weak Hom-Hopf algebra is a septuple (H, µ, η, ∆, ε, S, αH ),
in which (H, αH ) is a weak Hom-bialgebra, if H endowed with a k-linear map S (the
antipode), such that for any h, g ∈ H, the following conditions hold:
(idH ⊗ ∆)∆(1H ) = 11 ⊗ 1′
112 ⊗ 1′
2.
1 ⊗ 1′
2,
(1) S ◦ αH = αH ◦ S;
(2) h1S(h2) = εt(h), S(h1)h2 = εs(h);
(3) S(hg) = S(g)S(h), S(1H ) = 1H ;
(4) ∆(S(h)) = S(h2) ⊗ S(h1), ε ◦ S = ε.
Let (H, αH ) be a weak Hom-bialgebra. Define linear maps εt and εs by the formula
εt(h) = ε(11h)12,
εs(h) = 11ε(h12),
for any h ∈ H, where εt, εs are called the target and source counital maps. We adopt the
notations Ht = εt(H) and Hs = εs(H) for their images.
Similarly, we define the linear maps bεt and bεs by the formula
bεs(h) = 11ε(12h),
bεt(h) = ε(h11)12,
for any h ∈ H. Their images are denoted by cHt = bεt(H) and cHs = bεs(H).
2.2. Duality and rigid categories.
Recall from [9] that let (C, ⊗, I, a, l, r) be a monoidal category. V ∈ C, a left dual of V is
a triple (V ∗, evV , coevV ), where V ∗ is an object, evV : V ∗ ⊗V → I and coevV : I → V ⊗V ∗
are morphisms in C, satisfying
rV ◦ (idV ⊗ evV ) ◦ aV,V ∗,V ◦ (coevV ⊗ idV ) ◦ l−1
V = idV ,
and
lV ∗ ◦ (evV ⊗ idV ∗) ◦ a−1
V ∗,V,V ∗ ◦ (idV ∗ ⊗ coevV ) ◦ r−1
V ∗ = idV ∗.
Similarly, a right dual of V is a triple (∗V, eevV , gcoevV ), where ∗V is an object, eevV :
V ⊗ ∗V → I and gcoevV : I → ∗V ⊗ V are morphisms in C, satisfying
r∗V ◦ (idV ⊗ eevV ) ◦ a∗V,V,∗V ◦ (gcoevV ⊗ id∗V ) ◦ l−1
∗V = id∗V ,
4
and
lV ◦ (eevV ⊗ idV ) ◦ a−1
V,∗V,V ◦ (idV ⊗ gcoevV ) ◦ r−1
V = idV .
If each object in C admits a left dual (resp. a right dual, both a left dual and a right
dual), then C is called a left rigid category (resp. a right rigid category, a rigid category).
3. Left-right Yetter-Drinfeld modules over a weak Hom-Hopf
algebra
Definition 3.1. Let (H, αH ) be a weak Hom-Hopf algebra. A Yetter-Drinfeld module
over H is a vector space (M, αM ), such that M is a unital left H-module (with notation
h ⊗ m 7→ h · m) and a counital right H-comodule (with notation m ⊗ h 7→ m(0) ⊗ m(1))
with the following compatibility condition:
(h · m)(0) ⊗ (h · m)(1) = α−1
H (h21) · m(0) ⊗ [α−2(h22)α−1
H (m(1))]S−1(h1),
(3. 1)
for all h ∈ H and m ∈ M . We denote by H WYDH the category of Yetter-Drinfeld
modules, morphisms being the H-linear H-colinear maps.
Proposition 3.2. one has that(3.1) is equivalent to the following equations
αH(h1) · m(0) ⊗ α2
ρ(m) = m(0) ⊗ m(1) ∈ M ⊗t H , (11 ⊗ 12) · (M ⊗ H)
H(h1)
H(h2)αH (m(1)) = (h2 · m)(0) ⊗ (h2 · m)(1)α2
(3. 2)
(3. 3)
Proof (3.1) =⇒ (3.2), (3.3). We have
m(0) ⊗ m(1)
H (121) · α−1
1 · (12 · α−1
1 · (12 · α−1
1 · (12 · α−1
= α−1
= 1′
= 1′
= 1′
H (m(1))]S−1(11)
M (m(0)) ⊗ [α−2
M (m(0))) ⊗ [α−2
M (m(0))) ⊗ α−1
M (m(0))) ⊗ 1′
H (122)α−2
H (1′
H (1′
2[α−2
2)α−2
2)[α−2
H (m(1))S−1(11)].
H (m(1))]S−1(11)
H (m(1))S−1(α−1
H (11))]
5
Then we do a calculation as follows:
H (m(1)))S−1(h21)]α2
H (m(1)))S−1(h12)]αH (h11)
H (h1)
H (h1)
(h2 · m)(0) ⊗ (h2 · m)(1)α2
H (h221) · m(0) ⊗ [(α−2
H (h22)α−1
H (h222)α−1
= α−1
= h21 · m(0) ⊗ [(α−1
= h21 · m(0) ⊗ [(h22m(1))]S−1(h12)h11
= h21 · m(0) ⊗ [h22m(1)]S−1(11)ε(h112)
= h12 · m(0) ⊗ [αH(h2)m(1)]S−1(11)ε(α−1
= α−1
= α−1
= α−1
= h112 · m(0) ⊗ [h2m(1)]S−1(11)
H (m(1))]]S−1(11)
2α−1
= h11′
H (h2)[[13α−1
= αH(h1)[12 · α−1
= αH(h1) · m(0) ⊗ α2
H (h11)12)
2 · m(0) ⊗ [αH (h2)m(1)]S−1(11)ε(α−2
2 · m(0) ⊗ [αH (h2)m(1)]S−1(11)ε(α−1
2 · m(0) ⊗ [αH (h2)m(1)]S−1(11)ε(α−1
H (h12)1′
H (h12)1′
H (h12)1′
112 · m(0) ⊗ [αH (h2)[1′
M (m(0))] ⊗ α2
H (h2)αH (m(1)).
H (m(1))]S−1(α−1
H (11))]
H (h11)α−1
H (h11)[α−1
H (h11)[1′
H (1′
H (1′
112])
1)12)
1)α−1
H (12)])
For (3.2), (3.3) =⇒ (3.1), we have
H (h22) · α−2
H (h21) · m(0) ⊗ [α−2
α−1
M (α−4
= α2
H (h2) · m)(0) ⊗ (α−2
= (α−1
H (h2) · m)(0) ⊗ (α−2
= (α−1
2α−3
2α−2
H (h2) · m)(0) ⊗ (1′
= (1′
H (h2) · m)(0) ⊗ (13α−3
= (13α−2
H (h) · m)(0) ⊗ (12α−2
= (12α−1
= (12 · (α−1
= 11 · α−1
= (h · m)(0) ⊗ (h · m)(1).
M (h · m)(0) ⊗ 12α−1
H (h) · α−1
H (h22)α−1
H (m(1))]S−1(h1)
M (m))(0) ⊗ (α−4
H (h2) · α−1
H (h2) · α−1
H (h22) · α−2
M (m))(1)α−2
M (m))(1)11ε(12α−2
M (m))(1)α2
H(α−4
H (h12))S−1(α−2
H (h21))S−1(h1)
H (h11))
H (h1))
1α−3
H (h1))
H (h1))
H (h2) · α−1
H (h2) · α−1
H (h2) · α−1
H (m))(0) ⊗ (12 · (α−2
M (m))(1)11ε(121′
M (m))(1)11ε(12α−2
M (m))(1)11
H (h2) · α−2
M (m)))(1)11
H (h · m)(1)
Definition 3.3. Let (H, αH ) be a weak Hom-Hopf algebra. Left-right weak-Hom type
entwining structure is a triple (A, C, ψ), where (A, αA) is a Hom-algebra and (C, αC ) is a
Hom-coalgebra with a linear map ψ : A⊗C → A⊗C such that ψ◦(αA⊗αC) = (αA⊗αC)◦ψ
satisfying the following conditions:
αA(ψa) ⊗ ∆(cψ) = αA(a)ϕψ) ⊗ (c(1)
ψ(c ⊗ 1A) = ε(cψ
ψ(ab) ⊗ αC(cψ) = ψaϕb ⊗ αC(c)ψϕ,
1 )ψ1A ⊗ c2,
ψ ⊗ c(2)
ϕ),
ε(cψ)ψa = ε(cψ)a(ψ1A).
(3. 4)
(3. 5)
(3. 6)
(3. 7)
6
Over a weak-Hom type entwining structure (A, C, ψ), a left-right weak-Hom type entwined
modules (M, αM ) is both a right C-comodule and a left A-module such that
ρM (a · m) = αA(ψa) · m(0) · ⊗αC(mψ
(1)),
(3. 8)
for all a ∈ A and m ∈ M . AWM(ψ)C will denote the category of left-right weak-Hom
type entwined modules and morphisms between them.
Proposition 3.4. Let (H, αH ) be a weak Hom-Hopf algebra. Define φ : H ⊗ H →
H (c))S−1(a1) for all h, g ∈ H, and
In fact, for any
H ⊗ H given by φ(a ⊗ c) = α−1
so HWM(ψ)H is the category of weak-Hom type entwined modules.
(M, µ) ∈ HWM(ψ)H , one has compatible condition
H (a21) ⊗ (α−2
H (a22)α−1
ρM (a · m) = α−1
H (a21) · m(0) ⊗ (α−2
H (a22)α−1
H (m(1)))S−1(a1).
Proof. We need to prove that (3.4-3.7) hold. First, it is straightforward to check (3.4)
and (3.6). In what follows, we only verify (3.5) and (3.7). In fact, for all a, b, c ∈ H, we
have
c2 ⊗ ε(cψ
1 )ψ1A = c2 ⊗ ε(α−2
1S−1(11))121′′
1
1S−1(11))121′′
1
112)S(11)1′′
1
112)S(11)1′′
1
H (c1))]1′
H (c1)1′
H (c1))S−1(11))αH (12)
2[e11α−2
2e11]α−1
2 α−1
H (c1)]1′
2 α−1
2)ε(1′
2)ε(1′
2)ε(1′
2)ε(1′
H (c1)]1′
112)S(11)1′′
1
H (13)α−1
H (c2) ⊗ ε(1′′
H (c2) ⊗ ε([1′′
H (c2) ⊗ ε([1′′
H (c2)1′
H (c)1′
H (c)12 ⊗ S(11)1′′
1
H (c)S−1(11) ⊗ 121′
1
H (c)S−1(11) ⊗ 12
= e12α−1
= e12α−1
3α−1
= 1′′
3α−2
= 1′′
2α−1
= 1′′
2α−1
= 1′′
2α−1
= 1′
= 13α−1
= cψ ⊗ ψ1A.
3 ⊗ ε([1′′
2 ⊗ ε(1′
As for (3.7), we compute:
H (c)12))ε(11S−1(a1))αH (a21)
H (c)S−1(11)))ε(a112)αH (a21)
H (a11)12)αH (a12)
2(α−1
2(α−1
H (c))S−1(a1))αH (a21)
H (c))12)ε(11S−1(a1))αH (a21)
H (c)12))ε(11S−1(a1))αH (a21)
1)ε(1′
1)ε(1′
2(α−1
2(α−1
21′
1])ε(1′
H (a22)α−1
ε(cψ)ψa = ε((α−2
H (a22)α−1
= ε((α−2
H (a22)(α−1
= ε(α−1
= ε(α−1
H (a22)1′
= ε(α−1
H (a22)1′
1)ε(1′
= ε(a21′
= ε(a21′
1)ε(1′
= ε(α−1(a2)1′′
21′
= ε(a2[1′′
2(α−1
H (c)S−1(11)))αH (a1)1′
= ε(1′
2(α−1
H (c)S−1(11)))a1[1′
= ε(1′
112]
H (c)S−1(11)))a112
= ε(13(α−1
= ε(cψ)a(ψ1A).
H (c)S−1(11)))ε(α−1
H (c)S−1(11)))ε(a1(1′′
1)ε(1′
2(α−1
2(α−1
H (c)S−1(11)))a11′′
H (c)S−1(11)))a11′′
112
112
112
112))a211′′
2
7
Proposition 3.5. Let (H, αH ) be a weak Hom-Hopf algebra, for any (M, αM ), (N, αN ) ∈
HWYDH , and define the linear map
BM,N : M ⊗ N → N ⊗ M,
BM,N (m ⊗ n) = n(0) ⊗ α−1
H (n(1)) · m.
Then, we have (αN ⊗ αM ) ◦ BM,N = BM,N ◦ (αM ⊗ αN ) and, if (P, αP ) ∈ HWYDH, the
maps B−,− satisfy the Hom-Yang-Baxter equation:
(αP ⊗ BM,N ) ◦ (BM,P ) ⊗ (αM ⊗ BN,P )
= (BN,P ⊗ αM ) ◦ (αN ⊗ BN,P ) ◦ ((BM,P ⊗ αM ).
Proof. The proof is similar to Proposition 3.4 in [10].
lemma 3.6. Let(H, αH ) be a weak Hom-Hopf algebra, then Hs is the unit object in
HWYDH with the action: for any h ∈ H, x ∈ Hs,
and αHs = α.
h · x = bεs(h)(hx), ρ(x) = x1 ⊗ x2,
Proof. The proof is similar to [15].
lemma 3.7. Let (H, αH ) be a weak Hom-Hopf algebra, the left and right unit con-
straints lM : Hs ⊗t M → M and rM : M ⊗t Hs → M and their inverses are given by the
formulas
lM (x ⊗ m) = S(x) · α−2
rM (m ⊗ x) = x · α−2
M (m),
M (m),
l−1
M (m) = 1H ⊗ αM (m),
r−1
M (m) = ε(13)εs(12) · αM (m) ⊗ 11.
Proof. It is easy to see that lM is natural isomorphisms in HWYDH . We only check that
M lM (x ⊗t m) = l−1
l−1
M (S(x) · α−2
M (m)) = 1H ⊗t S(x) · α−1
M (m)
= εt(11) ⊗ (12S(x)) · m
= εt(bεs(S(x))1) ⊗ bεs(x)2 · m
= εt(11bεs(S(x))) ⊗ 12 · m
= εt(11x) ⊗ 12 · m = x ⊗t m,
and
lM l−1
M (m) = lM (1H ⊗t αM (m)) = 1H · α−2
M (αM (m)) = m,
which implies l−1
M is the inverse of lM .
HWYDH .
Similarly, we can check that rM is a natural isomorphism with the inverse r−1
M in
Theorem 3.8. Let (H, αH ) be a weak Hom-Hopf algebra. Then (H WYDH , ⊗t, Hs)
is a monoidal category.
Proof. Firstly, for any (M, αM ), (N, αN ), (P, αP ) ∈ HWYDH, define an associativity
constraint by
aM,N,P ((m ⊗t n) ⊗t p) = α−1
M (m) ⊗t (n ⊗t αP (p)), m ∈ M, n ∈ N, p ∈ P.
8
Obviously that a is natural and satisfies aM,N,P ◦ (αM ⊗ (αN ⊗ αP )) = ((αM ⊗ αN ) ⊗ αP ) ◦
aM,N,P . For any h ∈ H, since
aM,N,P (h · ((m ⊗t n) ⊗t p))
= α−1
M (h11 · m) ⊗t (h12 · n ⊗t αP (h2 · p))
= h1 · α−1
M (m) ⊗t (h21 · n ⊗t h22 · αP (p))
= h · (α−1
M (m) ⊗t (n ⊗t αP (p)))
= h · (aM,N,P ((m ⊗t n) ⊗t p)),
aM,N,P is H-linear.
Next we will check that aM,N,P is H-colinear.
(aM,N,P ⊗ idH ) ◦ ρ(M ⊗N ) ⊗P ((m ⊗ n) ⊗ p)
= (aM,N,P ⊗ idH )((m ⊗ n)(0) ⊗ p(0) ⊗ α−2
= aM,N,P ((m(0) ⊗ n(0)) ⊗ p(0)) ⊗ α−2
M (m(0)) ⊗ (n(0) ⊗ αP (p(0))) ⊗ α−2
= α−1
H (p(1)α−2
H (p(1))α−4
H (n(1)m(1)))
H (n(1)m(1)),
H (p(1)(m ⊗ n)(1)))
ρ(M ⊗N ) ⊗P ◦ aM,N,P ((m ⊗ n) ⊗ p)
M (m) ⊗ (n ⊗ αP (p)))
M (m)(0) ⊗ (n ⊗ αP (p))(0) ⊗ α−2
M (m)(0) ⊗ (n(0) ⊗ αP (p)(0)) ⊗ α−2
M (m)(0) ⊗ (n(0) ⊗ αP (p)(0)) ⊗ [α−3
M (m(0)) ⊗ (n(0) ⊗ αP (p(0))) ⊗ [α−2
M (m(0)) ⊗ (n(0) ⊗ αP (p(0))) ⊗ α−2
= ρ(M ⊗N ) ⊗P (α−1
= α−1
= α−1
= α−1
= α−1
= α−1
H ((n ⊗ αP (p))(1)α−1
M (m)(1))
H (αH(p)(1)n(1))α−1
H (α−2
H (p)(1)α−4
H (p)(1)[α−4
H (p(1))α−4
H (n(1))]α−3
H (n(1))α−4
H (n(1)m(1)).
M (m)(1)
M (m)(1)]
M (m)(1))
And aM,N,P is a bijection because of αM , αP are all bijective maps. Thus a is a natural
isomorphism in HWYDH.
Secondly, it is also a direct check to prove that a satisfies the Pentagon Axiom.
At last, we will check the Triangle Axiom. In fact, for any x ∈ Hs, we have
(rM ⊗ idN )((m ⊗t x) ⊗t n)
M (m) ⊗t n
M (m)) ⊗t 12 · n
= x · α−2
= 11 · (x · α−2
= 11x · α−1
= 11 · α−1
= 11 · α−1
= α−1
= α−1
= (idM ⊗ lN )aM,Ht,N ((m ⊗t x) ⊗t n).
M (m) ⊗t 12 · n
M (m) ⊗t 12S(x) · n
M (m) ⊗t 12 · (S(x) · α−1
M (m) ⊗ S(x) · α−1
M (m) ⊗t S(x) · α−2
N (n)
N (αN (n))
N (n))
Let (H, αH ) be a weak Hom-Hopf algebra with a bijective antipode S. Consider the
f.d. of HWYDH whose objects are finite-dimensional. Using the
full subcategory H WYDH
antipode S of H, we can provide WYD(H)f.d. with a duality.
9
For any (M, αM ) ∈ HWYDH
f.d., we set ∗M = Hom(M, k), with the action and the
coaction of H on M ∗ given by
(h·f )(m) = f (S(α−1
H (h))·α−2
M (m)) and f(0)(m)⊗f(1) = f (α−2
M (m(0)))⊗S−1(α−1
H (m(1))).
Similarly, for any (M, αM ) ∈ WYD(H)f.d., we set ∗M = Hom(M, k), with the action
and the coaction of H on M ∗ given by
(h·f )(m) = f (S−1(α−1
H (h))·α−2
M (m)) and f(0)(m)⊗f(1) = f (α−2
M (m(0)))⊗S(α−1
H (m(1))).
Theorem 3.9. The category H WYDH
Proof. Define the maps
f.d. is a rigid category.
coevM : Ht → M ⊗t M ∗,
x 7→ X x · (ei ⊗t αM ∗(ei)),
where ei and ei are bases of M and M ∗, respectively, dual to each other, and
evM : M ∗ ⊗t M → Ht,
f ⊗t m 7→ f (11 · m)12.
Firstly, we will prove that M ∗ is indeed an object in HWYDH, and α∗M is given by
α∗M (f )(m) = f (α−1
M (m)),
f ∈ ∗M, m ∈ M.
We have
(h · f )(0)(m) ⊗ (h · f )(1) = (h · f )(α−2
= f (S(α−1
M (m(0))) ⊗ S−1(α−1
H (m(1)))
H (h)) · α−4(m(0)))) ⊗ S−1(α−1
H (m(1))),
10
(α−1
H (h21) · f(0))(m) ⊗ (α−2
H (h22)α−1
H (f(1)))S−1(h1)
H (f(1)))S−1(h1)
H (S(α−1
H (h21)) · α−2
H (m(1))))S−1(h1)
H (h221))]
H (h22)α−1
H (h22)α−2
H (h22)
H (h2111))])S−1(h1)
H (h222)S−1(α−4
M (m)) ⊗ (α−2
M (m))(0)) ⊗ (α−2
M (m)(0)) ⊗ (α−2
H (S−1(m(1)))α−6
M (m)(0))) ⊗ ([α−4
H (h211))])S−1(h1)
M (m)(0))) ⊗ (11ε(12h22)
H (h211))])S−1(h1)
M (m)(0))) ⊗ (11ε(12h222)(α−4
M (m)(0))) ⊗ (11ε(12h222)(α−4
M (m)(0))) ⊗ (11ε(12h22)(α−3
2h22)(α−3
M (m)(0))) ⊗ (1′
2h2)(α−2
2h2)(α−3
M (m)(0))) ⊗ (1′
M (m)(0))) ⊗ (1′
1ε(1′
1ε(1′
1ε(1′
(α−4
(α−4
= f (S(α−4
H (h21)) · α−2
= f(0)(S(α−2
= f (S(α−4
= f (S(α−5
[S−1(α−4
= f (S(α−4
H (h21)) · α−4
H (h2112)) · α−4
H (h212))(α−5
H (h212)) · α−4
H (S−1(m(1)))α−4
H (h212)) · α−2
H (S−1(m(1)))α−4
H (h221)) · α−4
H (h221)) · α−4
H (h21)) · α−4
H (h21)) · α−4
2h12)) · α−4
H (1′′
H (12h1)) · α−4
H (1′′
1h)) · α−4
H (121′
H (h))S(12) · α−4
H (h))(12 · α−5
H (h)) · α−4
= f (S(α−4
= f (S(α−4
= f (S(α−3
= f (S(α−3
= f (S(α−4
= f (S(α−3
= f (S(α−4
= f (S(α−3
= f (S(α−2
= f (S(α−1
= f (S(α−1
2)(α−3
M (m)(0))) ⊗ (S−1(1′
M (m)(0))) ⊗ (S−1(13α−3
M (m)(0))) ⊗ (S−1(13α−3
H (m(1))).
M (m)(0)) ⊗ S−1(α−1
H (h21))])S−1(h1)
H (h21))])S−1(h1)
H (h12))S−1(α−2
H (h11))]
1h11)
H (S−1(m(1)))α−3
H (S−1(m(1)))α−3
H (S−1(m(1)))[α−2
H (S−1(m(1)))11ε(12h1)
H (S−1(m(1)))11ε(121′′
H (S−1(m(1)))11
2h2)(α−3
2)ε(1′
H (S−1(m(1)))11
H (S−1(m(1)))S−1(11))
11′′
H (S−1(m(1)))S−1(11))
1h1))S(12) · α−4
M (m)(0))) ⊗ (S−1(1′
H (S−1(m(1)))11
Which means that
(h · f )(0) ⊗ (h · f )(1) = (α−1
H (h21) · f(0)) ⊗ (α−2
H (h22)α−1
H (f(1)))S−1(h1).
We have known that Hs ∈ HWYDH, with left H-module structure h · z = bεs(hz) and
right H-comodule structure ρ(x) = x1 ⊗ x2, for all x ∈ Hs. Next, we will check evM and
coevM are morphisms in HWYDH. For any h ∈ H, m ∈ M , f ∈ M ∗, we compute
evM (h · (f ⊗t m)) = (h1 · f )(11 · (h2 · m))12
2α−3
H (h1)) · ((11 · (1′
H (h1))(12α−2
H (h)) · m)12
1α−2
= f (S(1′
= f ((S(11α−2
= f (εs(11α−2
= f (εs(α−2
= f (11 · m)εt(h12) = h · (evM (f ⊗t m)),
H (h1)) · v)εt(α−2
H (h2))) · ·α−1
H (h1))
H (h2))) · α−1
M (m))13
M (m)))12
and on one hand,
ρ ◦ evM (f ⊗ m) = ρ(12)f (11 · m)
= f (11 · m)S(1′
1) ⊗ 1′
2S−1(12),
11
on the other hand
(evM ⊗ id) ◦ ρ(f ⊗ m) = (evM ⊗ id)(f(0) ⊗ m(0)) ⊗ α−2
H (m(1)f(1))
= f(0)(11 · m(0))12 ⊗ α−2
= f (12 · α−2
M (m(0)(0)))(1′
= f (12 · α−1
M (m)(0))(1′
= f (12 · α−1
M (m)(0))(1′
= f (11 · m)1′
= f (11 · m)S(1′
2 ⊗ S−1(12)S−1(1′
1)
2S−1(12).
1) ⊗ 1′
H (m(1)f(1))
2) ⊗ α−2
H (m(1))S−1(1′
H (m(1)2)S−1(1′
2) ⊗ α−3
2) ⊗ S−1(εt(m(1)))S−1(1′
113α−3
113α−3
1))
H (m(0)(1))S−1(11))
H (m(1)1)S−1(11))
Thus evM is H-linear and H-colinear. And it is easy to get that evM ◦ (αM ∗ ⊗ αM ) =
αH ◦ evM , hence evM ∈ HWYDH.
Next we have
and
coevM (h · x)(m) = X εt(hx) · (ei ⊗ αM ∗(ei))(m)
= εt(hx) · α−2
M (m)),
h · coevM (x)(m) = (α−1
= (α−1
= εt(α−2
H (h1)x1) · αM (ei) ⊗t ((α−1
H (h2)α−1
H (h1)x1) · (S(α−2
H (h2)x2) · α2
H (x2)) · α−3
M (m))
M ∗(ei))(m)
H (h)α−1
H (x)) · α−2
M (m) = εt(hx) · α−2
M (m),
hence coevV is H-linear, it is not hard to check that coevV is H-colinear and is left to the
reader. Obviously that coevM ◦ αH = (αM ⊗ αM ∗) ◦ coevM , thus coevM ∈ H WYDH.
Secondly, we consider
(rM ◦ (idM ⊗ evM ) ◦ aM,M ∗,M ◦ (coevM ⊗ idM ) ◦ l−1
M )(m)
= (rM ◦ (idM ⊗ evM ) ◦ aM,M ∗,M )((11 · ei ⊗t 12 · αM ∗ (ei)) ⊗t αM (m))
= rM (ei ⊗t 12 · α2
M (m)))
= bεs(12) · (11 · α2
and
M ∗(ei)(11 · α2
M (m)) = m,
(lM ∗ ◦ (evM ⊗ idM ∗ ) ◦ a−1
= (lM ∗ ◦ (evM ⊗ idM ∗ ) ◦ a−1
= lV ∗(α2
= f (11 · (S(12) · α−2
M (m))) = f (m).
M ∗ (f )(11 · αM (ei))12 ⊗t αM ∗(ei))(m)
M ∗,M,M ∗ ◦ (idM ∗ ⊗ coevM ) ◦ r−1
M ∗,M,M ∗)(αM ∗ (f ) ⊗t (αM (ei) ⊗t α2
M ∗)(f )(m)
M ∗(ei)))(m)
Thus H WYDH
f.d. is a left rigid category.
Similarly, we define the following maps
and
x 7→ x · (X α∗M (ei) ⊗t ei),
∗M → Ht, m ⊗t f 7→ f (S−1(11) · m)12.
gcoevM : Ht → ∗M ⊗t M,
eevM : M ⊗t
We can show that (∗M, eevM , gcoevM ) is a right dual of M . Thus H WYDH
category.
f.d. is a right rigid
12
4. A braided monoidal category HWYDH I
Proposition 4.1. Let (H, αH ) be a weak Hom-Hopf algebra. For any (M, αM ), (N, αN ) ∈
HWYDH , then M ⊗t N = 11M ⊗ 12N ∈ H WYDH with structures:
h · (m ⊗t n) = h1 · m ⊗t h2 · n,
m ⊗t n 7→ (m ⊗t n)(0) ⊗t (m ⊗t n)(1) = (m(0) ⊗t n(0)) ⊗ α−2
H (n(1)m(1)).
Proof. Obviously M ⊗t N is a left H-module and a right H-comodule. We check now the
compatibility condition. We compute:
(h · (m ⊗ n))(0) ⊗ (h · (m ⊗ n))(1)
H ((h2 · n)(1)(h1 · m)(1))
H (((α−2
H (h222)α−1
H (n1))S−1(h21))
= (h1 · m ⊗ h2 · n)(0) ⊗ (h1 · m ⊗ h2 · n)(1)
= ((h1 · m)(0) ⊗ (h2 · n)(0) ⊗ α−2
= (α−1
((α−2
H (h121) · m(0) ⊗ α−1
H (h122)α−1
= (h12 · m(0) ⊗ α−1
H (h221) · n(0)) ⊗ α−2
H (m(1)))S−1(h11)))
H (h212) · n(0)) ⊗ α−2
H (h2111))(α−1
H (h212) · n(0)) ⊗ α−2
((S−1(α−3
H (h2112))α−3
= (h12 · m(0) ⊗ α−1
(12ε(α−3
H (h211)11)(α−1
H (m(1))S−1(α−1
= (h12 · m(0) ⊗ α−1
(12ε(α−2
H (h21)11)(α−1
H (h221) · n(0)) ⊗ α−2
H (m(1))S−1(α−1
H ((h22n1)
H (m(1))S−1(α−1
H ((h22n1)
H (h11))))
H ((h222n1)
H (h11))))
H (h11))))
= (α−1
H (h211) · m(0) ⊗ α−1
H (h212)11)(α−1
H (h221) · n(0)) ⊗ α−2
H (m(1))S−1(h1)))
(12ε(α−2
H ((h222n1)
= (α−2
H (h211)1′
(12ε(α−2
H (h212)1′
= (α−2
H (h211)1′
1 · m(0) ⊗ α−1
211)(α−1
1 · m(0) ⊗ α−1
211)(α−1
H (h221) · n(0)) ⊗ α−2
H (m(1))S−1(h1)))
H (h221) · n(0)) ⊗ α−2
H (m(1))S−1(h1)))
(12ε(α−2
H (h212)1′
H ((h222n1)
H ((h222n1)
= (α−2
H (h211)1′
211 · m(0) ⊗ α−1
H (m(1))S−1(α−1(h1)1′
(12(α−1
H (h21) · (m ⊗ n)(0) ⊗ α−2
= α−1
1)))
H (h22)α−1
H (m ⊗ n)(1)S−1(h1).
H (h212) · n(0)) ⊗ α−2
H ((αH (h22)n1)
Hence M ⊗t N ∈ HWYDH.
Proposition 4.2. Let (M, αM ), (N, αN ) ∈ HWYDH. Define the map
cM,N : M ⊗t N → M ⊗t M, cM,N (m ⊗ n) = α−1
N (n(0)) ⊗ α−1
M (α−1
H (n(1)) · m).
Then cM,N is H-linear H-colinear and satisfies the conditions (for (P, αP ) ∈ HWYDH)
P,M,N ◦ cM ⊗N,P ◦ a−1
a−1
M,N,P = (cM,P ⊗ idN ) ◦ a−1
M,P,N ◦ (idM ⊗ cN,P ),
(4. 1)
13
aN,P,M ◦ cM,N ⊗P ◦ aM,N,P = (idN ⊗ cM,P ) ◦ aN,M,P ◦ (cM,N ⊗ idP ).
(4. 2)
Proof. First, we prove that cM,N is H-linear, we compute:
cM,N (h · (m ⊗ n))
M (α−1
H (h221) · n(0)) ⊗ α−1
= cM,N (h1 · m ⊗ h2 · n)
N ((h2 · n)(0)) ⊗ α−1
= α−1
N (α−1
= α−1
N (h21 · n(0)) ⊗ α−1
= α−1
N (h21 · n(0)) ⊗ α−1
= α−1
N (h21 · n(0)) ⊗ α−1
= α−1
N (h21 · n(0)) ⊗ α−1
= α−1
N (α−1
= α−1
N (h112 · n(0)) ⊗ α−1
= α−1
= h1 · (12 · α−2
= h1 · α−1
= h · cM,N (m ⊗ n).
N (n(0)) ⊗ α−1
H (h12)13 · n(0)) ⊗ α−1
M ([α−2
N (n(0))) ⊗ α−1
M (h2)α−2
M (α−1
M ([α−2
M ([α−2
M ([α−2
H (n(1))]S−1(h21)) · (h1 · m))
H (h11) · m))
H (h11) · m))
H (n(1))]S−1(h12)) · (α−1
H (h12)) · (α−1
H (h12))α−2
H (h11)] · αM (m))
H (h222)α−1
H ([α−2
H (h22)α−1
H ((h2 · n)(1))) · (h1 · m))
M (α−1
H ([α−1
H (h22)α−2
H (h22)α−2
H (h22)α−2
M ([α−2
H (h2)13α−2
H (h2)[13α−4
H (n(1)) · m
H (n(1))]S−1(α−1
H (n(1))][S−1(α−2
H (n(1))][S−1(11)ε(α−2
H (h2)14α−2
H (n(1))]S−1(11) · αM (m))
H (n(1))S−1(11)] · m
H (h112))] · αM (m))
H (n(1))][S−1(11)ε(α−2
H (h1112))] · αM (m))
Next we prove that cM,N is H-colinear.
ρN ⊗M cM,N (m ⊗ n)
H (n(1)) · m))
H (n(1)) · α−1
H (n(1)21) · α−1
M (m))(0) ⊗ α−2
M (m(0)) ⊗ α−2
H ((α−2
H ((α−4
H (n(1)) · α−1
H (n(1)22)α−2
M (m))(1)α−1
H (m(1))
N (n(0)(1)))
= α−1
M (α−1
N (n(0)) ⊗ α−1
= ρN ⊗M (α−1
= α−1
= α−1
N (n(0)(1)))
N (n(0)(0)) ⊗ (α−2
N (n(0)(0)) ⊗ α−3
S−1(n(1)1))α−1
N (n(0)(0)) ⊗ α−2
[S−1(α−3
= n(0) ⊗ α−3
= α−1
= α−1
= α−1
= (cM,N ⊗ idH )ρM ⊗N (m ⊗ n).
N (n(0)(0)) ⊗ α−3
N (n(0)(0)) ⊗ α−2
N (n(0)(0)) ⊗ α−2
H (n(0)(1)2))α−3
H (n(1)1)12 · α−1
H (n(1)1) · α−1
M (m(0)) ⊗ α−2
H ((α−2
H (n(1)2)α−1
H (m(1)))
N (n(0)(1)(1))])
M (m(0)) ⊗ α−2
H ((α−3
H (n(1)2)13α−1
H (m(1)))S−1(11))
H (n(0)(1))12 · α−1
H (n(0)(1))(12 · α−2
H (n(0)(1)) · α−1
M (m(0)) ⊗ α−2
M (m(0))) ⊗ α−2
M (m(0)) ⊗ α−2
H (n(1)m(1))
H ((α−2
H (n(1))13α−1
H (m(1)))S−1(11))
H (n(1)[13α−2
H (m(1))]S−1(11))
As for (4.2), for any m ∈ M, n ∈ N and p ∈ P , we have
(aN,P,M ◦ cM,N ⊗P ◦ aM,N,P )((m ⊗t n) ⊗t p)
N (n(0)) ⊗ p(0) ⊗ α−4
N (n(0)) ⊗ (p(0) ⊗ α−3
H (αH (p(1))n(1)) · α−2
= aN,P,M (α−1
= α−2
= (idN ⊗t (cM,P ))(α−2
M (m) ⊗ αP (p))
= ((idN ⊗t cM,P ) ◦ aN,M,P ◦ (cM,N ⊗t idP ))((m ⊗t n) ⊗t p),
H (αH (p(1))n(1)) · α−1
H (n(1)) · α−1
N (n(0)) ⊗ (α−2
M (m))
M (m))
14
we can check that (4.1) in the similar way.
Lemma 4.3. cM,N is bijective with inverse
M,N (n ⊗ m) = α−1
c−1
M (α−1
H (S(n(1))) · m) ⊗ α−1
N (n(0)).
Proof. First, we prove that cM,N ◦ c−1
M,N = id. For any m ∈ M and n ∈ N , we have
H (S(n(1))) · m) ⊗ α−1
M (m) ⊗ α−1
N (n(0)))
N (n(0)))
M,N (n ⊗ m)
M (α−1
H (S(n(1))) · α−1
cM,N c−1
= cM,N (α−1
= cM,N (α−2
N (n(0)(0)) ⊗ α−3
= α−2
N (n(0)(0)) ⊗ [α−4
= α−2
N (n(1)1)α−4
N (n(0)) ⊗ [α−4
= α−1
N (n(0)) ⊗ 12ε(11n(1)) · α−1
= α−1
1α−2
= 1′
= 11α−1
= n ⊗ m.
N (n(0)) ⊗ 12ε(111′
N (n) ⊗ 12 · α−1
M (m)
N (n(0)(1)) · [α−3
N (n(0)(1))α−3
H (S(n(1))) · α−2
H (S(n(1)))] · α−1
M (m)
M (m)]
M (m)
H (S(n(1)2))] · α−1
M (m)
2n(1)) · α−1
M (m)
Then, we note that the following relation holds, for all m ∈ M ,
m(0) ⊗ m(1)
H (121) · α−1
1 · (12 · α−1
1 · (12 · α−1
1 · (12 · α−1
= α−1
= 1′
= 1′
= 1′
= 12 · α−1
H (m(1))]S−1(11)
M (m(0)) ⊗ [α−2
M (m(0))) ⊗ [α−2
M (m(0))) ⊗ α−1
M (m(0))) ⊗ 1′
H (122)α−2
H (1′
H (1′
2[α−2
2)α−2
2)[α−2
H (m(1))S−1(11)]
H (m(1))]S−1(11)
H (m(1))S−1(α−1
H (11))]
M (m(0)) ⊗ α−2
H (m(1))S−1(11).
15
Finally, we check that c−1
M,N ◦ cM,N = id. For any m ∈ M and n ∈ N , we have
c−1
M,N cM,N (m ⊗ n)
N (n(0)) ⊗ α−1
M,N (α−1
= c−1
H (S(n(0)(1))) · [α−3
= α−3
H (S(n(0)(1)))α−3
= [α−4
H (S(n(1)1))α−4
= [α−4
M (m) ⊗ α−1
= εs(n(1)) · α−1
= 11S((S−1εs(n(1)))) · α−1
= S(12)S((S−1εs(n(1)))) · α−1
= S((S−1εs(n(1)))12) · α−1
= S((S−1εs(n(1)))S(11)) · α−1
= S2(11) · α−1
= S2(11) · α−1
= m ⊗ n.
H (n(1)) · m))
M (α−1
H (n(1)) · α−2
H (n(1))] · α−1
H (n(1)2)] · α−1
N (n(0))
M (m) ⊗ 12 · α−1
M (m)] ⊗ α−2
M (m) ⊗ α−2
M (m) ⊗ α−1
N (n(0)(0))
N (n(0)(0))
N (n(0))
N (n(0))
M (m) ⊗ S(11) · α−1
N (n(0))
M (m) ⊗ S(11) · α−1
N (n(0))
M (m) ⊗ S2(12) · α−1
N (n(0))
N (n(0))
M (m) ⊗ S2(12S−1εs(n(1))) · α−1
M (m) ⊗ S2(12) · α−1
N (n)
Theorem 4.4. HWYDH is a braided monoidal category.
We can make now the connection between Yetter-Drinfeld modules over weak Hom-
Hopf algebras and modules over quasitriangular weak Hom-Hopf algebras.
Definition 4.5.[16] Let (H, α) be a weak Hom-bialgebra. If there exists R = R(1) ⊗
R(2) ∈ ∆op(1)(H ⊗k H)∆(1), such that the following conditions hold:
(1) (α ⊗ α)R = R;
(2) R∆(h) = ∆op(h)R;
(3) there exists R ∈ ∆(1)(H ⊗k H)∆op(1), such that RR = ∆op(1), RR = ∆(1);
(4) α(R(1)) ⊗ R(2)
(5) R(1)
2 = α−1(r(1)R(1)) ⊗ R(2) ⊗ r(2);
2 ⊗ α(R(2)) = r(1) ⊗ R(1) ⊗ α−1(r(2)R(2)),
1 ⊗ R(2)
1 ⊗ R(1)
where h ∈ H, r = R = R(1) ⊗ R(2) = r(1) ⊗ r(2), then R is called an R-matrix of H, R is
called the weak inverse of R. (H, R) is called a quasitriangular weak Hom-bialgebra.
Proposition 4.6. Let (H, αH , R) be a quasitriangular weak Hom-Hopf algebra, then
we have
(i) Let (M, αM ) be a left H-module with action H ⊗ M → M, h ⊗t m 7→ h · m. Define
the linear map ρM : M → M ⊗t H, ρM (m) = m(0) ⊗t m(1) := R(2) · m ⊗t αH (R(1)). Then
(M, αM ) with these structures is a Yetter-Drinfeld module over H.
(ii) Let (N, αN ) be another left H-module with action H ⊗t N → N, h ⊗t n 7→ h · n,
regarded as a Yetter-Drinfeld module as in (i), via the map ρN : N → N ⊗ H, ρN (n) =
n(0) ⊗ n(1) := r(2) · m ⊗t αH(r(1)). We regard (M ⊗t N, αM ⊗ αN ) as a left H-module via
the standard action h · (m ⊗ n) = h1 · m ⊗t h2 · n and then we regard (M ⊗t N, αM ⊗ αN ) as
a Yetter-Drinfeld module as in (i). Then this Yetter-Drinfeld module (M ⊗t N, αM ⊗ αN )
coincides with the Yetter-Drinfeld module M ⊗t N defined as in Proposition 4.1.
Proof. First we have to prove that (M, αM ) is a right H-comodule; ρ(α(m)) =
16
αM (m(0)) ⊗ α(m(1)) is easy and left to the reader, we check
(αM ⊗ ∆)ρM (m) = αM (R(2) · m) ⊗ ∆(αH (R(1)))
= αH(R(2)) · αM (m) ⊗ αH (R(1)
= r(2)R(2) · αM (m) ⊗ α2
= αH(r(2)) · (R(2) · m) ⊗ α2
= r(2) · (R(2) · m) ⊗ αH (r(1)) ⊗ α2
= ρM (R(2) · m) ⊗ α2
= (ρM ⊗ αM )ρM (m).
H(R(1))
1 ) ⊗ αH (R(1)
2 )
H(R(1))
H(r(1)) ⊗ α2
H(r(1)) ⊗ α2
H(R(1))
H(R(1))
Now we check the Yetter-Drinfeld condition (3.3):
(h2 · m)(0) ⊗ (h2 · m)(1)α2
H(h1) = R(2) · (h2 · m) ⊗ αH (R(1))α2
= αH (R(2)) · (h2 · m) ⊗ α2
= (R(2)h2) · αM (m) ⊗ α2
= (h1R(2)) · αM (m) ⊗ α2
= αH (h1) · (R(2) · m) ⊗ α2
= αH (h1) · m(0) ⊗ α2
H (h2)αH(m(1)).
H (h1)
H (h1)
H (R(1))α2
H (R(1)h1)
H (h2R(1))
H (h2)α2
H (R(1))
(ii) We only need to prove that the two comodule structures on M ⊗t N coincide, that
is, for all m ∈ M and n ∈ N ,
m(0) ⊗ n(0) ⊗ α−2
H (n(1)m(1)) = R(2) · (m ⊗ n) ⊗ αH(R(1)),
that is
R(2) · m ⊗ r(2) · n ⊗ α−2
H (αH (r(2))αH (R(2))) = R(2)
1
· m ⊗ R(2)
2
· n ⊗ αH (R(1)),
which is equivalent to
αH (R(2)) · m ⊗ αH(r(2)) · n ⊗ r(2)R(2) = R(2)
1
· m ⊗ R(2)
2
· n ⊗ αH(R(1)).
Proposition 4.7. Let (H, αH , R) be a quasitriangular weak Hom-Hopf algebra. De-
note by Rep(H) the category whose objects are left H-modules and whose morphisms
are H-linear maps. Then Rep(H) is a braided monoidal subcategory of H WYDH, with
tensor product defined as in Proposition 4.1, associativity constraints defined by the for-
mula aM,N,P ((m ⊗n) ⊗p) = α−1
M (m) ⊗(n ⊗αP (p)) for any M, N, P ∈ Rep(H), and braid-
N (n) ⊗t R(1) · α−1
ing cM,N : M ⊗t M → N ⊗t M, m ⊗t n 7→ R(2) · α−1
M (m), with in-
· α−1
· α−1
verse c−1
M (m) ⊗t R
N (n), for any
(M, αM ), (N, αN ) ∈ Rep(H).
M,N : N ⊗t M → M ⊗t M, n ⊗t m 7→ R
(1)
(2)
5. A braided monoidal category HWYDH II
17
Modules over quasitriangular weak Hom-Hopf algebras become Yetter-Drinfeld mod-
ules over weak Hom-Hopf algebras are proved in Section 4. Similarly, comodules over
coquasitriangular weak Hom-Hopf algebras become Yetter-Drinfeld modules over weak
Hom-Hopf algebras; inspired by this, we can introduce a second braided monoidal cate-
gory structure on H WYDH. We include these facts here for completeness. Each of the
next results is the analogue of a result in Section 4; their proofs are similar to those of
their analogues and are left to the reader.
Proposition 5.1. Let (H, αH ) be a weak Hom-Hopf algebra.
(i) Let (M, αM ), (N, αN ) ∈ H WYDH, with notation as above, and the tensor product
M ⊗N is obtained by
M ⊗N = {m ⊗n = m0 ⊗k n0ε(m1n1) m ∈ M, n ∈ N },
with structures:
h · (m ⊗n) = α−2
m ⊗n 7→ (m ⊗n)(0) ⊗ (m ⊗n)(1) = (m(0) ⊗n(0)) ⊗n(1)m(1).
H (h1) · m ⊗α−2
H (h2) · n,
(ii) H WYDH is a braided monoidal category, with tensor product ⊗ as in (i) and asso-
ciativity constraints aM,N,P and quasi-braiding cM,N defined as follows: for any (M, αM ),
(N, αN ), (P, αP ) ∈ HWYDH, define an associativity constraint by
aM,N,P ((m ⊗n) ⊗p) = α−1
M (m) ⊗(n ⊗αP (p)), m ∈ M, n ∈ N, p ∈ P,
cM,N : M ⊗N → M ⊗M, cM,N (m ⊗n) = α−1
N (n(0)) ⊗α−1
M (α−1
H (n(1)) · m).
with inverse
M,N (n ⊗m) = α−1
c−1
M (α−1
H (S(n(1))) · m) ⊗α−1
N (n(0))
.
Definition 5.2.[16] Let (H, α) be a weak Hom-bialgebra, if there is a linear map
σ : H ⊗k H → k, such that the following conditions hold:
(1) σ(a, b) = ε(b1a1)σ(a2, b2)ε(a3b3);
(2) σ(a1, b1)a2b2 = b1a1σ(a2, b2);
(3) there exists σ′ ∈ homk(H ⊗ H, k), such that σ(a1, b1)σ′(a2, b2) = ε(ab),
and σ′(a1, b1)σ(a2, b2) = ε(ba);
(4) σ(α(a), α(b)) = σ(a, b);
(5) σ(α(a), bc) = σ(a1, α(c))σ(a2, α(b));
(6) σ(ab, α(c)) = σ(α(a), c1)σ(α(b), c2),
where a, b, c ∈ H, then σ is called an coquasitriangular form of H, σ′ is called the weak
convolution inverse of σ. (H, σ) is called a coquasitriangular weak Hom-bialgebra.
Proposition 5.3. Let (H, αH , σ) be a coquasitriangular weak Hom-Hopf algebra,
then we have
(i) Let (M, αM ) be a left H-comodule with coaction M → M ⊗H, ρM (m) = m(0) ⊗m(1).
Define the linear map H ⊗ M → M, h ⊗m 7→ h · m := σ(αH (h), m(1))m(0). Then (M, αM )
with these structures is a Yetter-Drinfeld module over H.
(ii) Let (N, αN ) be another left H-comodule with coaction N → N ⊗H, ρN (n) =
n(0) ⊗n(1). Define the linear map H ⊗ N → N, h ⊗n 7→ h · n := σ(αH (h), n(1))n(0). We
18
regard (M ⊗N, αM ⊗ αN ) as a left H-module via the standard action h·(m ⊗n) = α−2
H (h1)·
m ⊗α−2
H (h2)· n and then we regard (M ⊗N, αM ⊗ αN ) as a Yetter-Drinfeld module as in (i).
Then this Yetter-Drinfeld module (M ⊗N, αM ⊗ αN ) coincides with the Yetter-Drinfeld
module M ⊗N defined as in Proposition 5.1.
Theorem 5.4. Let (H, αH , σ) be a coquasitriangular weak Hom-Hopf algebra. Denote
by Corep(H) the category whose objects are right H-comodules (M, αM ) and morphisms
are morphisms of right H-comodules. Then Corep(H) is a braided monoidal subcategory
of H WYDH, with tensor product defined as in Proposition 5.1, associativity constraints
defined by the formula aM,N,P ((m ⊗n) ⊗p) = α−1
M (m) ⊗(n ⊗αP (p)) for any M, N, P ∈
Corep(H), and braiding cM,N : M ⊗N → N ⊗M, m ⊗n 7→ α−1
M (m0)σ(m1, n1),
with inverse c−1M,N : N ⊗M → M ⊗N, n ⊗m 7→ α−1
N (n0)σ′(m1, n1), for any
(M, αM ), (N, αN ) ∈ Corep(H).
M (m0) ⊗α−1
M (n0) ⊗α−1
ACKNOWLEDGEMENT
The work is supported by Project Funded by China Postdoctoral Science Foundation
(No. 2015M580508), the Fund of Science and Technology Department of Guizhou Province
(No. 2014GZ81365) and the Program for Science and Technology Innovation Talents in
Education Department of Guizhou Province(No. KY[2015]481).
References
[1] G. Bohm, F. Nill, and K. Szlach´anyi, Weak Hopf algebras I: Integral theory and C ∗-
structure, J. Algebra 221 (1999), 385-438.
[2] S. Caenepeel, and I. Goyvaerts, Monoidal Hom-Hopf algebras, Comm. Algebra 39
(2011), 2216-2240.
[3] S. Caenepeel, D. G. Wang, and Y. Yan, Yetter-Drinfeld modules over weak bialgebras,
Ann. Univ. Ferrara Sez. VII (N.S.) 51 (2005), 69-98.
[4] J.T. Hartwig, D. Larsson, and S.D. Silvestrov, Deformation of Lie algebras using
σ-derivations, J. Algebra 295 (2006), 314-361.
[5] M. Chaichian, D. Ellinas and Z. Popowicz, Quantum conformal algebra with central
extension, Phys. Lett. B 248(1-2) (1990), 95-99.
[6] M. Chaichian, A. P. Isaev, J. Lukierski, Z. Popowi, and P. Presnajder, q-deformations
of Virasoro algebra and conformal dimensions, Phys. Lett. B 262(1) (1991), 32-38.
[7] M. Chaichian, P. Kulish, and J. Lukierski, q-deformed Jacobi identity, q-oscillators
and q-deformed infinite-dimensional algebras, Phys. Lett. B 237(3-4) (1990), 401-406.
[8] N. Hu, q-Witt algebras, q-Lie algebras, q-holomorph structure and representations,
Algebra Colloq. 6(1)(1999), 51-70.
[9] C. Kassel, Quantum Groups, Springer-Verlag, New York, 1995.
[10] A. Makhlouf, and F. Panaite, Yetter-Drinfeld modules for Hom-bialgebras, J. Math.
Phys. 55(1) (2014), 013501.
19
[11] A. Makhlouf, and F. Panaite, Hom-L-R-smash products, Hom-diagonal crossed prod-
ucts and the Drinfeld double of a Hom-Hopf algebra, J. Algebra 441(2015), 314-343.
[12] A. Makhlouf, and S. D. Silvestrov, Hom-algebra stuctures, J. Gen. Lie Theory Appl.
2 (2008), 51-64.
[13] A. Makhlouf, and S. D. Silvestrov, Hom-Lie admissible Hom-coalgebras and Hom-
Hopf algebras, J. Gen. Lie Theory in Mathematics, Physics and beyond. Springer-
Verlag, Berlin, 2009, pp. 189-206.
[14] A. Makhlouf, and S. D. Silvestrov, Hom-algebras and Hom-coalgebras, J. Algebra
Appl. 9(2010), 553-589.
[15] A. Nenciu, The center construction for weak Hopf algebras, Tsukuba J. Math. 26
(2002), 189-2004.
[16] X. H. Zhang, and S. H. Wang, Weak Hom-Hopf algebras and their (co)representations,
Journal of Geometry and Physics 94(2014), 50-71.
[17] D. Yau, Hom-bialgebras and comodule algebras, Int. Electron. J. Algebra 8 (2010),
45-64.
[18] D. Yau, The Hom-Yang-Baxter equation, Hom-Lie algebras, and quasi-triangular bial-
gebras, J. Phys. A. 42(16) (2010), 165202.
[19] D. Yau, The Hom-Yang-Baxter equation and Hom-Lie algebras, J. Math. Phys.
52(2011), 053502.
20
|
1511.02323 | 1 | 1511 | 2015-11-07T08:53:30 | Centers of Cuntz-Krieger $C^*$-algebras | [
"math.RA"
] | For a finite directed graph $\Gamma$ we determine the center of the Cuntz-Krieger $C^*$-algebra $CK(\Gamma).$ | math.RA | math |
CENTERS OF CUNTZ-KRIEGER C ∗-ALGEBRAS
ADEL ALAHMADI AND HAMED ALSULAMI
Abstract. For a finite directed graph Γ we determine the center of the Cuntz-
Krieger C ∗-algebra CK(Γ).
1. Definitions and Main Results
Let Γ = (V, E, s, r) be a directed graph that consists of a set V of vertices and a
set E of edges, and two maps r : E → V and s : E → V identifying the range and
the source of each edge. A graph is finite if both sets V and E are finite. A graph
is row-finite if s−1(v) < ∞ for an arbitrary vertex v ∈ V.
Definition 1. If Γ is a row-finite graph, the Cuntz-Krieger C ∗-algebra CK(Γ) is
the universal C ∗-algebra generated ( as a C ∗-algebra ) by V ∪ E and satisfying the
relations (1) s(e)e = er(e) = e for all e ∈ E, (2) e∗f = δe,f r(f ) for all e, f ∈ E, (3)
v = Xe∈s−1(v)
ee∗ whenever s−1(v) 6= ∅.
The discrete C-subalgebra of CK(Γ) generated by V, E, E ∗ is isomorphic to the
Leavitt path algebra L(Γ) of the graph Γ ( see [T2] ). We will identify L(Γ) with
its image in CK(Γ). Clearly L(Γ) is dense in CK(Γ) in the C ∗- topology.
A path is a finite sequence p = e1 · · · en of edges with r(ei) = s(ei+1) for 1 ≤ i ≤
n − 1. We consider the vertices to be paths of length zero. We let P ath(Γ) denote
the set of all paths in the graph Γ and extend the maps r, s to P ath(Γ) as follows:
for p = e1 · · · en we set s(p) = s(e1), r(p) = r(en). For v ∈ V viewed as a path we
set s(v) = r(v) = v. A vertex w is a descendant of a vertex v if there exists a path
p ∈ P ath(Γ) such that s(p) = v, r(p) = w.
A cycle is a path C = e1 · · · en, n ≥ 1 such that s(e1) = r(en) and all vertices
s(e1), . . . , s(en) are distinct. An edge e ∈ E is called an exit from the cycle C if
s(e) ∈ {s(e1), · · · , s(en)}, but e /∈ {e1, · · · , en}. A cycle without an exit is called a
Key words and phrases. Cuntz-Krieger algebra, Leavitt path algebra.
1
2
ADEL ALAHMADI AND HAMED ALSULAMI
N E-cycle. A subset W ⊂ V is hereditary if all descendants of an arbitrary vertex
w ∈ W also lie in W. For two nonempty subsets W1, W2 ⊂ V let E(W1, W2) denote
the set of edges {e ∈ E s(e) ∈ W1, r(e) ∈ W2}.
For a hereditary subset W ⊂ V the C ∗-subalgebra of CK(Γ) generated by
W, E(W, W ) is isomorphic to the Cuntz-Krieger algebra of the graph
(W, E(W, W ), sE(W,W ), rE(W,W )), ( see [BPRS], [BHRS] ). We will denote it as
CK(W ).
Example 1. Let the graph Γ be a cycle, V = {v1, · · · , vn}, E = {e1, · · · , en},
s(ei) = vi, 1 ≤ i ≤ n; r(ei) = vi+1 for 1 ≤ i ≤ n − 1, r(en) = v1. The algebra
CK(Γ) in this case is isomorphic to the matrix algebra Mn(T ), where T is the C ∗-
algebra of continuous functions on the unit circle. The center of CK(Γ) is generated
by the element e1 · · · en + e2e3 · · · e1 + · · · + ene1 · · · e1 and is isomorphic T.
For an arbitrary row-finite graph Γ if a path C = e1 · · · en is a N E-cycle with
the hereditary set of vertices V (C) = {s(e1), · · · , s(en)} then we denote CK(C) =
CK(V (C)) ∼= Mn(T ). The center Z(C) of the subalgebra CK(C) of CK(Γ) is
generated by the element
z(C) = e1 · · · en + e2e3 · · · ene1 + · · · + ene1 · · · en−1.
We will need some definitions and some results from [AA1], [AA2].
Definition 2. Let W ⊂ V be a nonempty subset. We say that a path p =
e1 · · · en, ei ∈ E, is an arrival path in W if r(p) ∈ W, and {s(e1), · · · , s(en)} * W.
In other words, r(p) is the first vertex on p that lies in W. In particular, every vertex
w ∈ W, viewed as a path of zero length, is an arrival path in W. Let Arr(W ) be
the set of all arrival paths in W.
Definition 3. A hereditary set W ⊆ V is called finitary if Arr(W ) < ∞.
If W is a hereditary finitary subset of V then e(W ) = Xp∈Arr(W )
idempotent in L(Γ) and, hence in CK(Γ), see [AA1],[AA2]. If C is a N E-cycle in Γ
pp∗ is a central
with the hereditary set of vertices V (C) then we will denote Arr(C) = Arr(V (C)).
If the set V (C) is finitary then we will say that the cycle is finitary. In this case
pzp∗ lies in the center of L(Γ)
for an arbitrary element z ∈ Z(C) the sum Xp∈Arr(C)
and CK(Γ), see [AA2].
CENTERS OF CUNTZ-KRIEGER C ∗-ALGEBRAS
3
Theorem 1. Let Γ be a finite graph. The center Z(CK(Γ)) is spanned by:
(i) central idempotents e(W ), where W runs over all nonempty hereditary fini-
tary subsets of V ;
(ii) subspaces { Xp∈Arr(C)
cycles of Γ.
pzp∗ z ∈ Z(C)}, where C runs over all finitary N E-
Corollary 1. Z(CK(Γ)) is the closure of Z(L(Γ)).
Corollary 2. The center Z(CK(Γ)) is isomorphic to a finite direct sum C ⊕ · · · ⊕
C ⊕ T ⊕ · · · ⊕ T.
In [CGBMGSMSH] it was shown that the center of a prime Cuntz-Krieger C ∗-
algebra is equal to C · 1 ( for a finite graph ) or to (0) ( for infinite graph ).
We remark that for two distinct hereditary subsets W1 $ W2 the central idem-
potents e(W1), e(W2) may be equal.
To make the statement of the Theorem more precise we will consider annihilator
hereditary subsets.
Let W be a nonempty subset of V. Consider the subset W ⊥ = {v ∈ V
the vertex v does not have descendants in W }.
For empty set we let ∅⊥ = V. It is easy to see that W ⊥ is always a hereditary
subset of V. If W1 ⊆ W2 ⊆ V then W ⊥
1 ⊇ W ⊥
2 , (W ⊥)⊥ ⊇ W, ((W ⊥)⊥)⊥ = W ⊥.
Definition 4. We will refer to W ⊥, W ⊆ V as annihilator hereditary subsets.
In [AA1] it was proved that (W ⊥)⊥ is the largest hereditary subset of V such
that every vertex in it has a descendant in W and that e(W ) = e((W ⊥)⊥).
Hence in the part (i) of the Theorem we can let W run over nonempty finitary
hereditary annihilator subsets of V.
In fact there is a 1-1 correspondence ( and a Boolean algebra isomorphism )
between finitary hereditary annihilator subsets of V and central idempotents of
L(Γ) and CK(Γ).
Example 2. Let Γ =
Thus there are no proper finitary hereditary subsets and Z(L(Γ)) = C · 1.
v2 . The set {v2} is hereditary, but not finitary.
v1
4
ADEL ALAHMADI AND HAMED ALSULAMI
v1
v2
v3
v4
. The set {v2, v3, v4} is hereditary and finitary,
Example 3. Let Γ =
but ({v2, v3, v4}⊥)⊥ = V. Thus there are no proper finitary hereditary annihilator
subsets and again Z(L(Γ)) = C · 1.
v1
v2
v3
v5
Example 4. Let Γ =
subsets are {v5} and {v2, v3, v4}. Hence Z(L(Γ)) = T ⊕ C.
. The only finitary hereditary annihilator
v4
2. Closed Ideals
Let W be a hereditary subset of V. The ideal I(W ) of the Leavitt path algebra
generated by the set W is the C-span of the set {pq∗ p, q ∈ P ath(Γ), r(p) = r(q) ∈
W }. The closure I(W ) of the ideal W is the closed ideal of the C ∗-algebra CK(Γ)
generated by the set W.
In this section we will use induction on V to prove that for a proper hereditary
subset W ⊂ V central elements from Z(CK(Γ)) ∩ I(W ) are of the form predicted
by the Theorem.
Given paths p, q ∈ P ath(Γ) we say q is a continuation of the path p if there exists
a path p′ ∈ P ath(Γ) such that q = pp′. In this case the path p is called a beginning
of the path q. We will often use the following well known fact: if p, q ∈ P ath(Γ),
p∗q 6= 0, then one of the paths p, q is a continuation of the other one.
Remark that if W is a hereditary subset of V and p, q ∈ Arr(W ), p 6= q, then
none of the paths p, q is a continuation of the other one.
For a path p ∈ P ath(Γ) we consider the idempotent ep = pp∗.
Lemma 1. Let W be a nonempty hereditary subset of V let a ∈ I(W ), let ǫ > 0.
Then the set {p ∈ Arr(W ) kepak ≥ ǫ} is finite.
CENTERS OF CUNTZ-KRIEGER C ∗-ALGEBRAS
5
Proof. The ideal I(W ) generated by the set W in the Leavitt path algebra L(Γ) is
pbp,qq∗, bp,q ∈ L(W ),
dense in I(W ). Choose an element b ∈ I(W ), b = Xp,q∈Arr(W )
such that ka − bk < ǫ. If p′ ∈ Arr(W ) is an arrival path in W that is different
from all the paths involved in the decomposition of the element b then p′∗b = 0 and
ep′b = 0.
Now
kep′ak = kep′(a − b)k ≤ kep′kka − bk < ǫ,
which proves the Lemma.
(cid:3)
Lemma 2. Let a ∈ CK(Γ), v ∈ V, a ∈ vCK(Γ)v; p ∈ P ath(Γ), r(p) = v. Then
kpap∗k = kak.
Proof. We have
kpap∗k ≤ kpk · kak · kp∗k = kak.
On the other hand, a = p∗(pap∗)p,
kak ≤ kp∗k · kpap∗k · kpk = kpap∗k,
which proves the Lemma.
(cid:3)
Lemma 3. Let W be a nonempty hereditary subset of V let z ∈ Z(CK(Γ) ∩ I(W ).
For an arbitrary vertex w ∈ W if wz 6= 0 then the set {p ∈ Arr(W ) r(p) = w} is
finite.
Proof. Let p ∈ Arr(W ), r(p) = w, zw 6= 0. We have epz = pp∗z = pzp∗ =
p(wzw)p∗. By Lemma 2 kepzk = kwzk > 0. Now it remains to refer to Lemma
1.
(cid:3)
where C runs over finitary N E-cycles of Γ.
pzp∗ z ∈ Z(C)},
Let eZ denote the sum of all subspaces Ce(W ), where W runs over nonempty
finitary hereditary subsets of V, and all subspaces { Xp∈Arr(C)
Our aim is to show that Z(CK(Γ)) = eZ.
Let's use induction on the number of vertices. In other words, let's assume that
for a graph with < V vertices the assertion of the Theorem is true.
Lemma 4. Let W be a proper hereditary subset of V. Then Z(CK(Γ) ∩ I(W ) ⊆ eZ.
6
ADEL ALAHMADI AND HAMED ALSULAMI
Proof. Let 0 6= z ∈ Z(CK(Γ) ∩ I(W ). Consider the element z0 = z(Xw∈W
Z(CK(W )). If z0 = 0 then zW = (0), zI(W ) = (0), z2 = 0, which contradicts
semiprimeness of CK(Γ) ( see [BPRS],[BHRS]).
w) ∈
By the induction assumption there exist disjoint hereditary finitary ( in W )
cycles C1, · · · , Cr and hereditary finitary (again in W ) subsets W1, · · · , Wk ⊂ W
such that
z0 =
rXi=1
αi Xp∈ArrW (Ci)
paip∗+
kXj=1
βj Xq∈ArrW (Wj )
qq∗ ; αi, βj ∈ C, ai ∈ Z(Ci).
The notations ArrW (Ci), ArrW (Wj ) are used to stress that arrival paths are
considered in the graph (W, E(W, W )).
The fact that the hereditary finitary subsets V (Ci), Wj can be assumed disjoint
follows from the description of the Boolean algebra of finitary hereditary subsets in
[AA2].
If αi 6= 0 then for arbitrary vertex w ∈ V (Ci) we have z0w = zw 6= 0. Hence by
Lemma 3 there are only finitely many paths p ∈ Arr(W ) such that r(p) = w. Hence
V (Ci) is a finitary subset of V. Similarly, if βj 6= 0 then Wj is a finitary subset in
V.
βje(Wj) ∈ Z(CK(Γ)).
w) = (0), (z −z′)I(W ) =
Consider the central element
z′ =
αi Xp∈Arr(Ci)
rXi=1
w) = z0 = z(Xw∈W
kXj=1
paip∗ +
w). Hence, (z −z′)(Xw∈W
We have z′(Xw∈W
(0), (z − z′)2 = 0. Again by semiprimeness of CK(Γ) we conclude that z = z′, which
proves the Lemma.
(cid:3)
3. Proof of the Theorem
Definition 5. A vertex v ∈ V is called a sink if s−1(v) = ∅.
Definition 6. A hereditary subset W ⊂ V is called saturated if for an arbitrary
non-sink vertex v ∈ V the inclusion r(s−1(v)) ⊆ W implies v ∈ W.
Definition 7. If W is hereditary subset then we define the saturation of W to be the
smallest saturated hereditary subsetcW that contains W. In this case I(W ) = I(cW ).
CENTERS OF CUNTZ-KRIEGER C ∗-ALGEBRAS
7
Definition 8. If W is a hereditary saturated subset of V then the graph Γ/W =
(V \ W, E(V, V \ W )) is called the factor graph of Γ modulo W.
We have CK(Γ)/I(W ) ∼= CK(Γ/W ) ( see[T1] ). In [AAP], [T2] it was proved
that the following 3 statements are equivalent:
1) the Cuntz-Krieger C ∗-algebra CK(Γ) is simple,
2) the Leavitt path algebra L(Γ) is simple,
3) (i) V does not have proper hereditary saturated subsets, (ii) every cycle
has an exit.
We call a graph satisfying the condition 3) simple.
The following lemma is well known. Still we prove it for the sake of completeness.
Lemma 5. Let Γ be a graph such that V does not have proper hereditary subsets.
Then Γ is either simple or a cycle.
Proof. If Γ is not simple then Γ contains a N E-cycle C. The set of vertices V (C)
is hereditary subset of V. In view of our assumption V (C) = V, which proves the
Lemma.
(cid:3)
Let U = {α ∈ C α = 1} be the unit circle in C. Let E ′ be a subset of the set E
of edges. For an arbitrary α ∈ U the mapping gE ′(α) such that gE ′(α) : v 7→ v, v ∈
V ; gE ′(α) : e 7→ αe, e∗ 7→ αe, e ∈ E ′; gE ′(α) : e 7→ e, e∗ 7→ e∗, e ∈ E \ E ′, extends
to an automorphism gE ′(α) of the C ∗-algebra CK(Γ). Denote GE ′ = {gE ′(α), α ∈
U} ≤ Aut(CK(Γ)). The group GE is called the gauge group of the C ∗-algebra
CK(Γ). An ideal of CK(Γ) is called gauge invariant if it is invariant with respect
to the group GE.
In [BPRS], [BHRS] it is proved that a nonzero closed gauge invariant ideal of
CK(Γ) has a nonempty intersection with V.
Lemma 6. Let the graph Γ be a cycle, Γ = (V, E), V = {v1, · · · , vd}, E = {e1, · · · , ed},
s(ei) = vi for 1 ≤ i ≤ d; r(ei) = vi+1 for 1 ≤ i ≤ d − 1, r(ed) = v1. Then the central
elements from Z(CK(Γ)) that are fixed by all gE(α), α ∈ U, are scalars.
Proof. The center of CK(Γ) is isomorphic to the algebra of continuous function T =
{f : U → C}, the corresponding action of GE on T is (gE(α)f )(u) = f (αdu).Now,
if f (u) = f (αdu) for all α, u ∈ U then f is a constant function, which proves the
Lemma.
(cid:3)
8
ADEL ALAHMADI AND HAMED ALSULAMI
Lemma 7. Let W be a hereditary saturated subset of V such that Γ/W is a cycle,
E(V \W, W ) 6= ∅, Γ/W = {v1, · · · , vd}. Then dXi=1
viCK(Γ)vi!∩Z(CK(Γ)) = (0).
Proof. Consider the set
W ′ = {w ∈ W E(V \ W, W )CK(Γ)w = (0)}.
The set W ′ is hereditary and saturated. Moreover, dXi=1
viCK(Γ)vi!∩I(W ′) = (0).
Indeed, we only need to notice that if p ∈ P ath(Γ), s(p) = vi and r(p) ∈ W ′ then
p = 0. Let p = e1 · · · en, ei ∈ E. At least one edge ej, 1 ≤ j ≤ n, lies in E(V \W, W ).
This implies the claim. Factoring out I(W ′) we can assume that W ′ = ∅.
viCK(Γ)vi! ∩ Z(CK(Γ)). For an arbitrary edge e ∈ E(V \
Let 0 6= z ∈ dXi=1
W, W ) we have ze = ez. Hence ze = zer(e) = ezr(e) = 0. Consider the ideal
J = {a ∈ CK(Γ) E(V \ W, W )CK(Γ)a = aCK(Γ)E(V \ W, W ) = (0)}
of the algebra CK(Γ). The element z lies in J. The ideal J is gauge invariant. Hence
by [BPRS], [BHRS] J ∩ V 6= ∅. A vertex vi, 1 ≤ i ≤ d, can not lie in J because
viCK(Γ)E(V \ W, W ) 6= (0). On the other hand J ∩ W ⊆ W ′ = ∅, a contradiction,
which proves the Lemma.
(cid:3)
Lemma 8. Let W be a hereditary saturated subset of V such that Γ/W is a cycle
and E(V \ W, W ) 6= ∅. Then Z(CK(Γ)) ⊆ C · 1CK(Γ) + I(W ).
Proof. As in the Lemma 7 we assume that V \ W = {v1, · · · , vd}, E ′ = E \
E(V, W ) = {e1, · · · , ed}; s(ei) = vi, 1 ≤ i ≤ d, r(ei) = vi+1 for 1 ≤ i ≤ d − 1,
r(ed) = v1. Consider the action of the group GE ′ on CK(Γ). Let z ∈ Z(CK(Γ)),
dXi=1
vizvi, b = Xw∈W
z = a + b, a =
wzw. Since every element from GE ′ fixes CK(W )
dXi=1
it follows that g(b) = b for all g ∈ GE ′. Now g(z) = g(a) + b, g(z) − z = g(a) − a ∈
viCK(Γ)vi! ∩ Z(CK(Γ)) = (0) by Lemma 7. we proved that an arbitrary
element of Z(CK(Γ)) is fixed by GE ′ . The ideal I(W ) is invariant with respect to
GE ′. Hence the group GE ′ acts on CK(Γ)/I(W ) ∼= CK(Γ/W ) as the full gauge
group of automorphisms. The image of the central element z in Z(CK(Γ/W ))
is fixed by GΓ/W . Hence by Lemma 6 it is scalar. This finishes the proof of the
Lemma.
(cid:3)
CENTERS OF CUNTZ-KRIEGER C ∗-ALGEBRAS
9
Proof of Theorem 1. If V does not contain proper hereditary subsets then by Lemma
5 Γ is either simple or a cycle. If the C ∗-algebra CK(Γ) is simple then Z(CK(Γ)) =
C · 1 ( see [D] ) and the assertion of the Theorem is clearly true.
If C is a cycle then Z(CK(Γ)) ∼= T and the assertion of the Theorem is again
true.
Let now W be a maximal proper hereditary subset of V. The saturation cW is
equal to V or the set W is saturated. In the first case CK(Γ) = I(W ) and it suffices
to refer to Lemma 4. Suppose now that the set W is saturated. The graph Γ/W
does not have proper hereditary subsets. Again by Lemma 5 the graph Γ/W is
either simple or a cycle. If Γ/W is simple then Z(CK(Γ/W )) = C · 1, which implies
Z(CK(Γ) ⊆ C · 1 + I(W ), which together with Lemma 4 implies the Theorem. If
Γ/W is a cycle then by Lemma 8 we again have Z(CK(Γ) ⊆ C · 1 + I(W ) and the
Theorem follows.
(cid:3)
Acknowledgement
This project was funded by the Deanship of Scientific Research (DSR), King
Abdulaziz University, under Grant No. (27-130-36-HiCi). The authors, therefore,
acknowledge technical and financial support of KAU.
References
[A] G. Abrams, Leavitt path algebras: the first decade, Bull. Math. Sci. (2015) 5:59 -- 120.
[AAP] G. Abrams, G. Aranda Pino, The Leavitt path algebra of a graph, J. Algebra 293 (2005),
319-334.
[AMP] P. Ara, M.A.Moreno, E.Pardo, Nonstable k-theory for graph algebras, Algebra Represent.
Theory 10 (2007), 157-178.
[AA1] A. Alahmadi, H. Alsulami, Completions of Leavitt path algebras, to appear in the Bull. Math.
Sciences.
[AA2] A. Alahmadi, H. Alsulami, Centers of Leavitt path algebras and their completions,
arXiv:1507.07439 [math.RA].
[BPRS] T. Bates, D. Pask, I. Raeburn, W. Szymanski, C ∗-algebras of row-finie graphs, New York J.
Math. 6( 2000), 307-324.
[BHRS] T. Bates, J.H.Hong, I. Raeburn, W. Szymanski, The ideal structure of C ∗-algebras of ifinite
graphs, Illinois J. Math. 46( 2002), 1159-1176.
[C] J. Cuntz, Simple C ∗-algebras generated by isometries, Comm. Math. Phys. 57(1977), 173-185.
[CGBMGSMSH] M.G. Corrales Garcia,D.M.Barquero, C. Martin Gonzalez, M.Siles Molina, J.F.
Solanilla Hernandez, Centers of path algebras, Cohn and Leavitt path algebras, Bull. Malay.
Math. Sci. So. DOI 10.1007/s4084001502141
[D] J. Dixmier, C ∗-algebras, North-Holland, 1977.
10
ADEL ALAHMADI AND HAMED ALSULAMI
[T1] M. Tomforde, Structure of graph C*-algebras and generalizations, Chapter in the book"Graph
Algebras: Bridging the Gap between Algebra and Analysis", Servicio de Publications de la
Universidas de Malaga, Malaga, Spain, 2006
[T2] M. Tomforde,, Uniqueness theorems and ideal structure for Leavitt path algebras, J. Algebra
318 (2007), 270-299.
Department of Mathematics, Faculty of Science, King Abdulaziz University, P.O.Box
80203, Jeddah, 21589, Saudi Arabia
E-mail address: [email protected]
Department of Mathematics, Faculty of Science, King Abdulaziz University, P.O.Box
80203, Jeddah, 21589, Saudi Arabia
E-mail address: [email protected]
|
1101.4415 | 1 | 1101 | 2011-01-23T23:21:28 | On some embedding of recursively presented Lie algebras | [
"math.RA"
] | In this paper we prove that every recursively presented Lie algebra over a field which is a finite extention of its simple subfield can be embedded in a recursively presented Lie algebra defined by relations which are equalities of (nonassociative) words of generators and a+b=c, where a,b,c are generators. | math.RA | math |
On some embedding of recursively presented Lie algebras
E.Chibrikov
1
Introduction
In 1961 G.Higman [7] proved an important Embedding Theorem which states that every recursively
presented group can be embedded in a finitely presented group. Recall that a group (or an algebra) is
called recursively presented if it can be given by a finite set of generators and a recursively enumerable
set of defining relations. If a group (or an algebra) can be given by finite sets of generators and defining
relations it is called finitely presented. As a corollary to this theorem G.Higman proved the existence of
a universal finitely presented group containing every finitely presented group as a subgroup. In fact, its
finitely generated subgroups are exactly the finitely generated recursively presented groups.
In [2] V.Ya.Belyaev proved an analog of Higman's theorem for associative algebras over a field which is
a finite extention of its simple subfield. The proof was based on his theorem stating that every recursively
presented associative algebra over a field as above can be embedded in a recursively presented associative
algebra with defining relations which are equalities of words of generators and α + β = γ, where α, β, γ
are generators. This result allowed Belyaev to apply V.Murskii analog of Higman's theorem [10] for
semigroups to obtain a Higman embedding for associative algebras. In recent paper [1] Y.Bahturin and
A.Olshanskii showed that such embedding can be performed distortion-free. The idea of transition from
algebras to semigroups was also used by G.P.Kukin in [8] (see also [5]).
This paper appears as a byproduct of the author's joint attemps with Prof. Y.Bahturin to prove Lie
algebra analog of Higman's Theorem. In particular, Y.Bahturin suggested to prove that any recursively
presented Lie algebra can be embedded in a Lie algebra given by Lie relations of the type mentioned
above. In this paper we show that this is indeed true. Namely, every recursively presented Lie algebra
over a field which is a finite extention of its simple subfield can be embedded in a recursively presented
Lie algebra defined by relations which are equalities of (nonassociative) words of generators and α+β = γ
(α, β, γ are generators). To prove this embedding we use Grobner-Shirshov basis theory for Lie algebras.
A short review of this theory as well as some properties of Lyndon-Shirshov words including a new proof
of Kukin's Lemma (see [5], Lemma 2.11.15) are given in the section 2. Note that an existence of Higman's
embedding for Lie algebras is still an open problem (see [9]). It is worth to mention a result by L.A.Bokut
[3] that for every recursively enumerable set M of positive integers, the Lie algebra
LM = Lie < a, b, c [abnc] = 0, n ∈ M >,
where [abnc] = [[. . . [[ab]b] . . . b]c], can be embedded into a finitely presented Lie algebra.
1
This paper was written when the author was a postdoctoral fellow at Memorial University of New-
foundland, supported by a PDF grant by Atlantic Association for Research in Mathematical Sciences as
well as Discovery grants of Drs. Y.Bahturin, M.Kotchetov and M.Parmenter. It gives me a pleasure to
thank them for the support and encouragement.
2
Some definitions and results
Let X = {xii ∈ I} be a linearly ordered set, k be a field and Liek(X) be the free Lie algebra over k
generated by X. Define hXi to be the free monoid of all associative words in X (including the empty
word 1). We use two linear orderings of hXi:
(i) (lexicographical order) u < 1 for every nonempty word u, and, by induction, u < v if u = xiu′,
v = xj v′ and either xi < xj or xi = xj and u′ < v′;
(ii) (deg-lex order) u ≺ v if u < v, or u = v and u < v, where w denotes the length of w.
We regard Liek(X) as the subspace of the free associative algebra khXi which is generated as an alge-
bra by X under the Lie bracketing [xy] = xy − yx. Given f ∈ khXi, denote by ¯f the maximal associative
word of f under the ordering (ii); f is a monic if f = ¯f +P αivi, where αi ∈ k, vi ∈ hXi and vi ≺ ¯f .
Definition 2.1 ([13], see also [15], [11]) An associative word w is an associative Lyndon -- Shirshov word
if, for arbitrary nonempty u and v such that w = uv, we have w > vu.
We will use the following properties of associative Lyndon-Shirshov words (see [13], [15] [4], [5], [12]):
(1) if w is an associative Lyndon-Shirshov word, then an arbitrary proper prefix of w cannot be
a suffix of w;
(2) if w = uv is an associative Lyndon-Shirshov word, where u, v 6= 1, then w > v;
(3) if u, v are associative Lyndon-Shirshov words and u > v, then uv is also an associative
Lyndon-Shirshov word;
(4) an arbitrary associative word w can be uniquely represented as
where c1, . . . , cn are associative Lyndon-Shirshov words and c1 ≤ c2 ≤ . . . ≤ cn.
w = c1c2 . . . cn,
Definition 2.2 A nonassociative word [u] is a Lyndon-Shirshov word if
(1) u is an associative Lyndon-Shirshov word;
(2) if [u] = [[u1][u2]], then [u1] and [u2] are Lyndon-Shirshov words (from (1) it then follows that
u1 > u2);
(3) if [u] = [[[u11][u12]][u2]], then u12 ≤ u2.
Put [x] = x for x ∈ X and put by induction [w] = [[u][v]] for an associative Lyndon-Shirshov word
w, where v is the longest proper associative Lyndon-Shirshov end of w (then u is also an associative
2
Lyndon-Shirshov word). Then [w] is a (nonassociative) Lyndon-Shirshov word. The main property of
[w] is
[w] = w.
It was shown in [13], [6] (see also [15] [5], [12]) that the set of all Lyndon-Shirshov words in the al-
phabet X forms a linear basis of Liek(X). This implies that if f ∈ Liek(X) then ¯f is an associative
Lyndon-Shirshov word.
Lemma (Shirshov [13], see also [15]) Suppose that w = aub, where w, u are associative Lyndon-
Shirshov words. Then a pair of related brackets in the nonassociative Lyndon-Shirshov word [w] is given
by [uc], i.e.
[w] = [a[uc]d],
where [uc] is a nonassociative Lyndon-Shirshov word and b = cd. Represent c in the form
c = c1c2 . . . ck,
where c1 ≤ c2 ≤ . . . ≤ ck are associative Lyndon-Shirshov words. Then replacing [uc] by [. . . [[u][c1]] . . . [ck]]
we obtain the word [w]u such that
[w]u = w.
The word [w]u is called the special bracketing or the special Shirshov bracketing of w relative to u.
We provide some additional information about bracketings in nonassociative Lyndon-Shirshov words.
Lemma 2.1 Take some associative word w represented as
w = c1c2 . . . cn,
where c1 ≤ c2 ≤ . . . ≤ cn are associative Lyndon-Shirshov words and an associative Lyndon-Shirshov
word v which is a subword of w. Then v is a subword (possibly not proper) of one of the words c1,
c2,. . .,cn.
Proof. Assume that v is not a subword of any of the words c1, . . . , cn. Then
v = c′′
i ci+1 . . . c′
t
for some 1 ≤ i < t ≤ n, where ci = c′
ct = c′
associative Lyndon-Shirshov words implies that c′′
property 2 also yields v > c′
t.
i , ct = c′
ic′′
t. Since a proper prefix is bigger than the word v, we have v < c′′
i and ct ≤ c′
i ≤ ci. Since ci ≤ . . . ≤ ct it follows that v < c′
tc′′
t and c′′
i 6= 1, c′
t 6= 1. It is possible that ci = c′′
i and/or
t. Property 2 of
t, but
3
Corollary 2.1 Suppose that w = aub, where w and u are associative Lyndon-Shirshov words. Then
a pair of related brackets in the nonassociative Lyndon-Shirshov word [w] is given by [uc], i.e.
where [uc] is a nonassociative Lyndon-Shirshov word and b = cd. Represent c in the form
[w] = [a[uc]d],
c = c1c2 . . . ck,
where c1 ≤ c2 ≤ . . . ≤ ck are associative Lyndon-Shirshov words. Then [uc] = [u[c1][c2] . . . [ck]], that is
[w] = [a[u[c1] . . . [ck]]d].
Proof. Shirshov's Lemma implies that [w] = [a[uc]d], where b = cd (possibly c = 1). Represent
the word c as c = c1c2 . . . ck, where c1 ≤ c2 ≤ . . . ≤ ck are associative Lyndon-Shirshov words. Let us
consider the Lyndon-Shirshov word [uc] = [uc1c2 . . . ck]. By applying Shirshov's Lemma and Lemma 2.1
to the words c1, c2,. . . ,ck we deduce that [uc] = [u[c1][c2] . . . [ck]].
The following Lemma is a generalization of the Shirshov Lemma and for the first time was noticed by
G.Kukin (see [5], Lemma 2.11.15). To prove this Lemma we use a different approach.
Lemma 2.2 (Kukin) Suppose that w = aubvd, where w, u, v are associative Lyndon-Shirshov words.
Then there is some bracketing
in the word w such that
[w]u,v = [a[u]b[v]d]
[w]u,v = w.
Proof. Shirshov's Lemma implies that [w] = [a[uc]p], where [uc] is a nonassociative Lyndon-Shirshov
word and cp = bvd. If v is a subword of p, then [w] = [a[uc]q[vs]l] by the Shirshov Lemma, where [vs] is a
Lyndon-Shirshov word and p = qvsl. Write c = c1c2 . . . cn and s = s1s2 . . . sm, where c1 ≤ c2 ≤ . . . ≤ cn
and s1 ≤ s2 ≤ . . . ≤ sm are associative Lyndon-Shirshov words. Then we define
where
and
[w]u,v = [a [[u][c1] . . . [cn]] q [[v][s1] . . . [sm]] l],
[[u][c1] . . . [cn]] = [. . . [[u][c1]] . . . [cn]]
[[v][s1] . . . [sm]] = [. . . [[v][s1]] . . . [sm]]
are Shirshov's special bracketings in the words uc and vs relative to u and v respectively (see Shirshov's
Lemma). Therefore, we get
[w]u,v = w.
4
Assume now that the word v is not a subword of p. Shirshov's Lemma implies that [w] = [aub[vs]q],
where [vs] is a nonassociative Lyndon-Shirshov word and sq = d. Therefore, [vs] is a nonassociative
subword of [uc], hence v is a subword of c.
Represent the word c as c = c1c2 . . . cn, where c1 ≤ c2 ≤ . . . ≤ cn are associative Lyndon-Shirshov
words. By Lemma 2.1 we deduce that v is a subword of ct for some t. If v = ct, then we define
[w]u,v = [w]u = [a [[u] [c1] . . . [ct] . . . [cn]] p],
i.e. [w]u,v is the special bracketing in w relative to u. It is obvious that
[w]u,v = w.
Now, we suppose that v is a proper subword of ct. Corollary 2.1 gives us the equality:
[w] = [a [u [c1] . . . [ct] . . . [cn] ] p].
t and c′′
Consequently, vs is a subword of ct, that is ct = c′
s1 ≤ s2 ≤ . . . ≤ sm are associative Lyndon-Shirshov words. Then we define
t for some c′
tvsc′′
t . Write s = s1s2 . . . sm, where
[w]u,v = [a [[u] [c1] . . . [ct]v . . . [cn]] p],
t [[v][s1] . . . [sm]] c′′
where [ct]v = [c′
t ] is the Shirshov special bracketing in ct relative to v. The word [w]u,v
results from [[u] [c1] . . . [ct] . . . [cn]], which is the Shirshov special bracketing in uc1 . . . ct . . . cn relative to
u, by replacing [ct] by [ct]v. It is clear that
[w]u,v = w.
Remark 2.1 It is not difficult to see that Lemma 2.2 can be easily generalized to an arbitrary number
of associative Lyndon-Shirshov subwords of w.
Let f be a monic Lie polynomial such that ¯f = h is a subword of w, i.e. w = u ¯f v. Denote by [uf v]f
the Lie polynomial obtained from [uhv]h by replacing [h] by f . The polynomial [uf v]f has the property
[uf v]f = w.
Definition 2.3 Take monic Lie polynomials f and g and a word w such that w = ¯f u = v¯g, where
u, v ∈ hXi and ¯f + ¯g > w. The intersection composition of f and g relative to w is defined by
(f, g)w = [f u]f − [vg]g.
Definition 2.4 Take monic Lie polynomials f and g and a word w such that w = ¯f = u¯gv, where
u, v ∈ hXi. The inclusion composition of f and g relative to w is defined by
(f, g)w = f − [ugv]g.
5
The main properties of compositions are: (f, g)w ∈ Id(f, g) and (f, g)w ≺ w.
Definition 2.5 ([3]) Given a set S of monic Lie polynomials, the composition (f, g)w of f and g is called
trivial relative to S if (f, g)w = Pi αi[uisivi]si, where αi ∈ k, ui, vi ∈ hXi, si ∈ S and ui¯sivi ≺ w.
Definition 2.6 Let S be a set of monic Lie polynomials. Then S is a Grobner-Shirshov basis if every
composition of any two elements of S is trivial relative to S.
Remark 2.2 If f ∈ Liek(X), then ¯f is an associative Lyndon-Shirshov word and by property 1 of
associative Lyndon-Shirshov words, there is no composition (f, f )w for any w ∈ hXi. Therefore, S = {f }
is a Grobner-Shirshov basis for every f ∈ Liek(X).
A Lyndon-Shirshov word [w] is called S-reduced if w 6= u¯sv for any s ∈ S and u, v ∈ hXi. A set Y is
called well-ordered if any chain of elements w1 > . . . > wk > . . . of hY i is stabilized. An important result
is Shirshov's Composition Lemma. We recall it as stated in [3]:
Composition Lemma Let X be a well-ordered set. If S is a Grobner-Shirshov basis and f ∈ Id(S),
then ¯f = u¯sv for some s ∈ S and u, v ∈ hXi.
It has an important corollary:
Composition-Diamond Lemma Suppose that X is a well-ordered set. Then S is a Grobner-
Shirshov basis if and only if the set of all Lyndon-Shirshov S-reduced words is a linear basis for Liek(X)/Id(S) =
Liek(XS).
Let f and g be elements of Liek(X). In some applications the following question appears:
which conditions on f and g give rise to the equality
where Id(h) denotes the ideal of Liek(X) generated by h?
As a corollary to Kukin's Lemma we obtain the following sufficient condition for the equality (1) to
Id(f )\ Id(g) = Id(f ) · Id(g),
(1)
hold.
Proposition 2.1 If there is a linear order of X such that there are no compositions (f, g)w and (g, f )w
for any w ∈ hXi, then Id(f )T Id(g) = Id(f ) · Id(g).
Proof. If h ∈ Id(f )T Id(g), then
h = Xi
αi[aif bi] = Xj
βj[cjgdj],
6
where αi, βj ∈ k and [...] are some bracketings. The Composition-Diamond Lemma (see also Remark 2.2)
and the hypotheses of the proposition imply that the maximal associative word of the polynomial h is
Let us consider
¯h = u1 ¯f u2¯gu3.
h − [u1f u2gu3]f,g,
where [u1f u2gu3]f,g is obtained from [u1 ¯f u2¯gu3] ¯f ,¯g by replacing ¯[f ] and ¯[g] by f and g respectively.
Lemma 2.2 implies that
Since
and
[u1f u2gu3]f,g = ¯h.
h − [u1f u2gu3]f,g ∈ Id(f )\ Id(g)
h − [u1f u2gu3]f,g < ¯h,
the claim of the proposition follows by induction.
3 Definition of the Functions ni(j) and s(i, j)
The idea of these functions belongs to V.Ya.Belyaev. We introduce them slightly modified.
Suppose that φ(i, j) is a recursive function. We define two functions
ni(j) : N × N → N and s(i, j) : N × N → N
as follow:
Let i > j and n2(1) be a natural number which is greater than or equal to max{2, φ(2, 1).}. Then we
define ni(j) as an arbitrary natural number such that ni(j) ≥ max{ni(j − 1) + 1, φ(i, j)} for j ≥ 2 and
ni(1) ≥ max{ni−1(i − 2) + 1, φ(i, 1)} for j = 1. For i ≤ j we set ni(j) = nj+1(i).
Let i < j. We define s(i, j) = s(j, i) = nj(i).
It is not difficult to notice that the functions satisfy the following properties:
ni(j) ≥ i for ∀j and s(i, j) ≥ φ(i, j).
4 Encoding of Addition and Multiplication
Now, let X = {x, y, z}, k be a field which is a finite extension of its simple subfield and let Liek(X) be
the free Lie algebra over k generated by X. If w = [..[[x1, x2], x3], . . . , xn] is a left-normed word, we will
write w =(cid:2)(cid:2)(cid:2)x1x2 . . . xn(cid:3)(cid:3)(cid:3)l. Denote by I the ideal of Liek(X) generated by the following relations:
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l, i > j; i, j = 1, 2, . . .
(2)
7
(cid:2)(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l,(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l(cid:3) −(cid:2)(cid:2)(cid:2)xyψ(i,j)z(cid:3)(cid:3)(cid:3)l, i > j; i, j = 1, 2, . . .
δt(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyj z(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D1,
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D2,
where φ, ψ are recursive functions, D1, D2 are recursively denumerable sets, δt (t = 1, 2, . . . , n) are
generators of the field k.
Let X1 = X S{u, α, β, γ} and Liek(X1) be the free Lie algebra over the field k generated by X1.
Denote by I1 the ideal of Liek(X1) generated by the relations:
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xuni(j)ǫijz(cid:3)(cid:3)(cid:3)l; i, j = 1, 2, . . . ; ǫij = (cid:26) α,
β,
i + j is even ,
i + j is odd ,
(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xus(i,j)γz(cid:3)(cid:3)(cid:3)l, i 6= j; i, j = 1, 2, . . .
α + β − γ,
(cid:2)(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l,(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l(cid:3) −(cid:2)(cid:2)(cid:2)xyψ(i,j)z(cid:3)(cid:3)(cid:3)l, i > j; i, j = 1, 2, . . .
δt(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyj z(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D1,
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D2,
where φ, ψ are the recursive functions, D1, D2 are the recursively denumerable sets, δt (t = 1, 2, . . . , n)
are the generators of the field k and ni(j), s(i, j) are functions defined in 1.1 relative to φ.
Lemma 4.1 I ⊆ I1.
Proof. It suffices to show that the relation (2) belogns to I1. Let i < j, then in Liek(X1)/I1 we have
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l =(cid:2)(cid:2)(cid:2)xuni(j−1)ǫi(j−1)z(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xunj (i)ǫijz(cid:3)(cid:3)(cid:3)l.
Since ni(j − 1) = nj(i) we obtain that
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l =(cid:2)(cid:2)(cid:2)xus(i,j)(ǫi(j−1) + ǫij)z(cid:3)(cid:3)(cid:3)l =(cid:2)(cid:2)(cid:2)xus(i,j)(α + β)z(cid:3)(cid:3)(cid:3)l =(cid:2)(cid:2)(cid:2)xus(i,j)γz(cid:3)(cid:3)(cid:3)l =(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l.
Thus, the relation (cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l belogns to I1.
Let x > z > α > β > γ > u > y. Denote by R a set of relations the union of which and (2)-(5)
is closed with respect to the operation of composition, that is the union is a Grobner-Shirshov basis. It
follows from the definition of composition that R ⊆ I and the set R consists of Lie polynomials in(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l.
Lemma 4.2 The following set S forms a Grobner-Shirshov basis:
(cid:2)(cid:2)(cid:2)xuni(j)ǫijz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l; i, j = 1, 2, . . . ; ǫij = (cid:26) α,
β,
i + j is even ,
i + j is odd ,
(12)
8
(cid:2)(cid:2)(cid:2)xus(i,j)γz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l, i 6= j; i, j = 1, 2, . . .
α + β − γ,
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l, i > j; i, j = 1, 2, . . .
(cid:2)(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l,(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l(cid:3) −(cid:2)(cid:2)(cid:2)xyψ(i,j)z(cid:3)(cid:3)(cid:3)l, i < j; i, j = 1, 2, . . .
δt(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyj z(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D1,
(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D2,
R
(13)
(14)
(15)
(16)
(17)
(18)
(19)
where φ, ψ are the recursive functions, D1, D2 are the recursively denumerable sets, δt (t = 1, 2, . . . , n)
are the generators of the field k.
Proof. Note that all the words of the relations (12)-(19) are nonassociative Lyndon-Shirshov words.
Since ni(j) ≥ i for ∀j and s(i, j) ≥ φ(i, j), the leading words of the relations (12)-(14) contain α, β
or γ. Therefore, the relations (12)-(14) do not form compositions with the relations (15)-(19). Because
the relations (15)-(19) are closed under composition by the definition of R, we have only one composition
to verify that it is trivial, namely, the inclusion composition of (12) and (14).
Let f =(cid:2)(cid:2)(cid:2)xuni(j)αz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l, where i + j is even , and g = α + β − γ. Then w = xuni(j)αz and
(f, g)w = f −(cid:2)(cid:2)(cid:2)xuni(j)(α + β − γ)z(cid:3)(cid:3)(cid:3)l = −(cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xuni(j)βz(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xuni(j)γz(cid:3)(cid:3)(cid:3)l.
If i ≤ j, then ni(j) = nj+1(i) and
(f, g)w = −((cid:2)(cid:2)(cid:2)xunj+1(i)βz(cid:3)(cid:3)(cid:3)l−(cid:2)(cid:2)(cid:2)xyj+1z(cid:3)(cid:3)(cid:3)l)+((cid:2)(cid:2)(cid:2)xus(i,j+1)γz(cid:3)(cid:3)(cid:3)l−(cid:2)(cid:2)(cid:2)xyφ(i,j+1)z(cid:3)(cid:3)(cid:3)l)−((cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l+(cid:2)(cid:2)(cid:2)xyj+1z(cid:3)(cid:3)(cid:3)l−(cid:2)(cid:2)(cid:2)xyφ(i,j+1)z(cid:3)(cid:3)(cid:3)l).
If i > j, then nj(i − 1) = ni(j) and
(f, g)w = −((cid:2)(cid:2)(cid:2)xunj (i−1)βz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l) + ((cid:2)(cid:2)(cid:2)xus(i,j)γz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l) − ((cid:2)(cid:2)(cid:2)xyiz(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)xyjz(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)xyφ(i,j)z(cid:3)(cid:3)(cid:3)l).
Therefore, (f, g)w is trivial relative to S.
Proposition 4.1 Liek(X)/I is embedded into Liek(X1)/I1.
Proof. It is sufficient to prove that Liek(X)T I1 = I. The inclusion I ⊆ Liek(X)T I1 follows from
Lemma 4.1. Suppose that f ∈ Liek(X)T I1. Composition Lemma for Lie algebras implies that in the free
f = asb for some words a, b ∈ hX1i and s ∈ S (recall that I1 = Id(S)). But
associative algebra khX1i
f ∈ Liek(X), hence, f ∈ hXi and, in particular, s ∈ hXi. Since the relations (12)-(14) contain α, β or γ
in their leading words, s is a relation of (15)-(19) and, therefore, s ∈ I. We have f − [asb]s ∈ Liek(X)T I1
and f − [asb]s < f . By performing the same calculations at some step we obtain that
f − [asb]s − [a1s1b1]s1 − ... − [akskbk]sk = 0,
where s, s1, ..., sk ∈ I and a, b, a1, b1, ..., ak, bk ∈ hXi. Thus, f ∈ I.
9
5 Proof of Embedding
Let Y = {a, b, c} with a > c > b and Liek(Y ) be the free Lie algebra generated by Y over a field k which
is a finite extention of its simple subfield. Note that the set of left-normed elements
T = {(cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l i = 1, 2, 3, . . .}
generates a free Lie subalgebra and constitutes a set of free generators. Indeed, the words (cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l are
Lyndon-Shirshov words and by ordering T as (cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l > (cid:2)(cid:2)(cid:2)abjc(cid:3)(cid:3)(cid:3)l if and only if i < j, we obtain that every
nonassocitive Lyndon-Shirshov word in the alphabet T is also a Lyndon-Shirshov word in the alphabet
Y . It implies that Lyndon-Shirshov words in the alphabet T are linearly independent in Liek(Y ), thus,
T is a set of free generators.
Let L be a finitely generated Lie algebra over the field k with a recursively denumerable set of defining
relations. Then L can be represented as L = {a1, a2, a3, . . .}, where each element of L is repeated at least
twice. There exists an embedding Φ : L −→ Liek(Y )/J such that Φ(ai) = (cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l and the ideal J of
Liek(Y ) is generated by the following relations:
(cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l +(cid:2)(cid:2)(cid:2)abjc(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)abφ(i,j)c(cid:3)(cid:3)(cid:3)l, i > j; i, j = 1, 2, . . .
(cid:2)(cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l,(cid:2)(cid:2)(cid:2)abjc(cid:3)(cid:3)(cid:3)l(cid:3) −(cid:2)(cid:2)(cid:2)abψ(i,j)c(cid:3)(cid:3)(cid:3)l, i > j; i, j = 1, 2, . . .
δt(cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)abjc(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D1,
(cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l −(cid:2)(cid:2)(cid:2)abjc(cid:3)(cid:3)(cid:3)l, (i, j) ∈ D2,
(20)
(21)
(22)
(23)
where φ, ψ are recursive functions, D1, D2 are recursively denumerable sets, δt (t = 1, 2, . . . , n) are
generators of the field k.
Indeed, the Lie algebra Liek(T )/J ′, where J ′ is an ideal of Liek(T ) generated by (20)-(23), is iso-
morphic to L. We have to show that Liek(T )T J = J ′. The inclusion J ′ ⊆ Liek(T )T J is obvious. Let
f ∈ Liek(T )T J. By Composition Lemma we have f = asb, where a, b ∈ hY i and s is a composition
obtained from the relations (20)-(23). Since s is a Lie polynomial in (cid:2)(cid:2)(cid:2)abic(cid:3)(cid:3)(cid:3)l (hence, s ∈ J ′) and f ∈ khT i
we can write
f = abi1 c . . . abit c s abit+1 c . . . abik c.
Therefore,
f − [(cid:2)(cid:2)(cid:2)abi1 c(cid:3)(cid:3)(cid:3)l . . .(cid:2)(cid:2)(cid:2)abit c(cid:3)(cid:3)(cid:3)l s (cid:2)(cid:2)(cid:2)abit+1c(cid:3)(cid:3)(cid:3)l . . .(cid:2)(cid:2)(cid:2)abik c(cid:3)(cid:3)(cid:3)l]s ∈ Liek(T )\ J and
f − [(cid:2)(cid:2)(cid:2)abi1 c(cid:3)(cid:3)(cid:3)l . . .(cid:2)(cid:2)(cid:2)abitc(cid:3)(cid:3)(cid:3)l s (cid:2)(cid:2)(cid:2)abit+1 c(cid:3)(cid:3)(cid:3)l . . .(cid:2)(cid:2)(cid:2)abik c(cid:3)(cid:3)(cid:3)l]s < f ,
and the proof follows from the induction on the degree of f .
Now, Proposition 4.1 implies the following
Theorem 5.1 A recursively presented Lie algebra over a field which is a finite extention of its simple
subfield can be embedded into a recursively presented Lie algebra defined by relations which are equalities
of (nonassociative) words of generators and α + β − γ, where α, β, γ are generators.
10
References
[1] Y.Bahturin, A.Olshanskii Filtrations and distortions in infinite-dimensional algebras, J.Algebra,
to appear; doi:10.1016/j.jalgebra.2010.09.019.
[2] V.Ya.Belyaev Subrings of finitely presented associative rings, Algera i Logica, 17 (1978), 627-638.
[3] L.A.Bokut Unsolvability of the equality problem and subalgebras of finitely presented Lie algebras.
Math. USSR Izvestia 6 (1972), 1153 -- 1199.
[4] L.A.Bokut, P.S.Kolesnikov Grobner-Shirshov bases: from their incipiency to the present. Journal of
Mathematical Sciences, 116 (2003), 1, 2894 -- 2916.
[5] L.A.Bokut, G.P.Kukin Algorithmic and combinatorial algebra. Mathematics and its Applications,
Kluwer Academic Publishers Group, Dordrecht, (1994).
[6] K.T.Chen, R.H.Fox, R.C.Lyndon Free differential calculus.IV: The quotient groups of the lower
central series. Ann. Math. 68 (1958), 2, 81 -- 95.
[7] G.Higman Subgroups of finitely presented groups, Proc. Royal Soc. London (Series A), 262 (1961),
455-475.
[8] G.P.Kukin On the equality problem for Lie algebras, Sib. Mat. Zh., 18, No. 5, 1194-1197 (1977).
[9] O. G. Kharlampovich, M. V. Sapir Algorithmic problems in varieties, Inter. J.Algebra Compt., 5,
no.4, 5 (1995) 379 -- 602.
[10] V.L.Murskii Isomorphic embeddability of a semigroup with an enumerable set of defining relations
into a finitely presented semigroup, Mat.Zemetki, 1(2) (1967), 217-224.
[11] R.C.Lyndon On Burnside's problem I. Trans. Amer. Math. Soc. 77 (1954), 202 -- 215.
[12] C.Reutenauer Free Lie algebras. London Mathematical Society Monographs. New Series, 7. Oxford
Science Publications. The Clarendon Press, Oxford University Press, New York, (1993).
[13] A.I.Shirshov On free Lie rings. Mat. Sb., 45 (1958), 2, 113 -- 122.
[14] A.I.Shirshov Sertain algorithmic problems for Lie algebras. Sibirsk. Mat. Z. 3 (1962), 292 -- 296 (Trans-
lation in ACM SIGSAM Bull 33 (1999), 2, 3 -- 6).
[15] Selected works of A.I.Shirshov, Birkhauser, 2009.
11
|
1103.4721 | 1 | 1103 | 2011-03-24T10:33:03 | Automorphisms and derivations of Leibniz algebras | [
"math.RA"
] | The present work is devoted to the extension of some general properties of automorphisms and derivations which are known for Lie algebras to finite dimensional complex Leibniz algebras. The analogues of the Jordan-Chevalley decomposition for derivations and the multiplicative decomposition for automorphisms of finite dimensional complex Leibniz algebras are obtained. | math.RA | math |
AUTOMORPHISMS AND DERIVATIONS OF LEIBNIZ ALGEBRAS
M. LADRA, I. M. RIKHSIBOEV, AND R. M. TURDIBAEV
Abstract. The present work is devoted to the extension of some general properties of
automorphisms and derivations which are known for Lie algebras to finite dimensional
complex Leibniz algebras. The analogues of the Jordan-Chevalley decomposition for
derivations and the multiplicative decomposition for automorphisms of finite dimen-
sional complex Leibniz algebras are obtained.
1. Introduction
Leibniz algebras were first introduced by Loday in [10, 11] as a non-antisymmetric
version of Lie algebras. Many results of Lie algebras were also established in Leibniz
algebras. Since the study of the properties of derivations and automorphisms of Lie
algebra play an essential role in the theory of Lie algebras, the question naturally arises
whether the corresponding results can be extended to the more general framework of
the Leibniz algebras.
In this work we consider some general properties of derivations and automorphisms of
Leibniz algebras. We extend some results obtained for derivations and automorphisms
of Lie algebras in [4, 7] to the case of Leibniz algebras. Among them we prove the ana-
logue of the Jordan-Chevalley decomposition, which expresses a derivation of a Leibniz
algebra as the sum of its commuting semisimple and nilpotent parts. Similar results
were established in [4] and [6] for Lie algebras. If the linear operator is invertible, then
the Jordan–Chevalley decomposition expresses it as a product of commuting semisimple
and unipotent operators. Gantmacher [4], in the theory of Lie algebras, proved that any
automorphism of Lie algebras decomposes into the product of commuting semisimple
automorphism and exponent of a nilpotent derivation. In this work we verify that the
same results hold in Leibniz algebras.
In 1955, Jacobson [7] proved that every Lie algebra over a field of characteristic zero
admitting a nonsingular derivation is nilpotent. The problem whether the converse
of this statement is correct remained open until that an example of a nilpotent Lie
algebra in which every derivation is nilpotent (and hence, singular) was constructed in
[3]. Nilpotent Lie algebras with this property were named characteristically nilpotent
Lie algebras. In [9] it was proved that every irreducible component of the variety of
complex filiform Lie algebras of dimension greater than 7 contains a Zariski open set
consisting of characteristically nilpotent Lie algebras. Note that among the nilpotent
Lie algebras of dimension less than 7, characteristically nilpotent Lie algebras do not
occur due to the classification given in [5].
In this paper we prove that a finite dimensional complex Leibniz algebra admitting
a non-degenerate derivation is nilpotent. Similar to the Lie case, the inverse of this
statement does not hold. The notion of characteristically nilpotent Leibniz algebra is
2010 Mathematics Subject Classification. 17A32, 17A36, 17B40.
Key words and phrases. Leibniz algebra; Lie algebra; derivation; automorphism; nilpotent Leibniz
algebra.
1
2
M. LADRA, I. M. RIKHSIBOEV, AND R. M. TURDIBAEV
defined similarly as in the Lie case. It was established in [13] that characteristically
nilpotent non-Lie filiform Leibniz algebras occur starting with dimension 5.
An advance of Theorems 3.8 and 3.10 was presented in [14]. We also have found in
the literature the reference [2], where some results here proved are showed with another
techniques.
In the present paper, all vector spaces and algebras are considered over the field of
the complex numbers C. We will denote by C i
k the binomial coefficient (cid:0)k
i(cid:1).
2. Preliminaries
In this section we present some known notions and results concerning Leibniz algebras
that we use further in this work.
Definition 2.1. An algebra L over a field F is called a Leibniz algebra if for any
x, y, z ∈ L the Leibniz identity
is satisfied, where [−, −] is the multiplication in L.
[[x, y], z] = [[x, z], y] + [x, [y, z]]
In other words, the right multiplication operator [−, z] by any element z is a derivation
(see [10]).
Any Lie algebra is a Leibniz algebra, and conversely any Leibniz algebra L is a Lie
algebra if [x, x] = 0 for all x ∈ L. Moreover, if Lann = ideal < [x, x] x ∈ L >, then the
factor algebra L/Lann is a Lie algebra.
For a Leibniz algebra L consider the following derived and lower central series:
(i)
(ii)
L(1) = L,
L1 = L,
L(n+1) = [L(n), L(n)],
Ln+1 = [Ln, L],
n > 1;
n > 1.
Definition 2.2. An algebra L is called solvable (nilpotent) if there exists s ∈ N (k ∈ N,
respectively) such that L(s) = 0 (Lk = 0, respectively).
The following theorem from linear algebra characterizes the decomposition of a vector
space into a direct sum of characteristic subspaces.
Theorem 2.3 ([12]). Let A be a linear transformation of vector space V . Then V
decomposes into the direct sum of characteristic subspaces V = Vλ1 ⊕ Vλ2 ⊕ · · · ⊕ Vλk
with respect to A, where Vλi = {x ∈ V (A − λiI)k(x) = 0 for some k ∈ N} and
λi, 1 ≤ i ≤ k, are eigenvalues of A.
The following proposition gives the (additive) Jordan–Chevalley decomposition of an
endomorphism.
Proposition 2.4 ([6]). Let V be a finite dimensional vector space over C, x ∈ End(V ).
(i) There exist unique xd, xn ∈ End(V ) satisfying the conditions: x = xd + xn, xd
is diagonalizable, xn is nilpotent, xd and xn commute.
(ii) There exist polynomials p(t), q(t) ∈ C[t], without constant term, such that xd =
p(x), xn = q(x). In particular, xd and xn commute with any endomorphism
commuting with x.
(iii) If A ⊆ B ⊆ V are subspaces and x maps B in A, then xd and xn also map B
in A.
In Leibniz algebras a derivation is defined as usual.
AUTOMORPHISMS AND DERIVATIONS OF LEIBNIZ ALGEBRAS
3
Definition 2.5. A linear operator d : L → L is called a derivation of L if
d([x, y]) = [d(x), y] + [x, d(y)]
for any x, y ∈ L.
For an arbitrary element x ∈ L, we consider the right multiplication operator
Rx : L → L defined by Rx(z) = [z, x]. Right multiplication operators are derivations
of the algebra L. The set R(L) = {Rx x ∈ L} is a Lie algebra with respect to the
commutator and the following identity holds:
RxRy − RyRx = R[y,x] .
(2.1)
Definition 2.6 ([8]). A subset S of an associative algebra A over a field F is called
weakly closed if for every pair (a, b) ∈ S × S an element γ(a, b) ∈ F is defined such that
ab + γ(a, b) ba ∈ S.
Further, we need a result concerning the weakly closed sets.
Theorem 2.7 ([8]). Let S be a weakly closed subset of the associative algebra A of
linear transformations of a finite-dimensional vector space V over F . Assume every
W ∈ S is nilpotent, that is, W k = 0 for some positive integer k. Then the enveloping
associative algebra S ∗ of S is nilpotent.
The classical Engel's theorem from Lie algebras has the following analogue in Leibniz
algebras.
Theorem 2.8 ([1], Engel's theorem). A Leibniz algebra L is nilpotent if and only if Rx
is nilpotent for any x ∈ L.
Definition 2.9. The set AnnR(L) = {x ∈ L [L, x] = 0} of a Leibniz algebra L is called
the right annihilator of L.
One can show that AnnR(L) is an ideal of L.
For a Leibniz algebra L, let H be a maximal solvable ideal in the sense that H
contains any solvable ideal of L. Since the sum of solvable ideals is again a solvable
ideal (see [1]), this implies the existence of a unique maximal solvable ideal, which is
said to be the radical of L.
Similarly, let K be a maximal nilpotent ideal of Leibniz algebra L. Since the sum
of nilpotent ideals is a nilpotent ideal (see [1]), this implies the existence of a unique
maximal nilpotent ideal, which is said to be the nilradical of L. Notice that, the
nilradical does not possess the properties of the radical in the sense of Kurosh.
3. Main Result
This section is devoted to the extension of known results for Lie algebras on auto-
morphisms and derivations to Leibniz algebras.
Lemma 3.1. Let L be a finite dimensional Leibniz algebra with a derivation d defined
on it and L = Lρ1 ⊕ · · · ⊕ Lρs be a decomposition of L into characteristic spaces with
respect to d. Then for any α, β ∈ Spec(d) we have
[Lα, Lβ] ⊆(cid:26) Lα+β
0
if α + β is an eigenvalue of d
if α + β is not an eigenvalue of d .
Proof. First observe that (d − (α + β)I)([x, y]) = [d(x), y] + [x, d(y)] − (α + β)[x, y] =
[(d − αI)(x), y] + [x, (d − βI)(y)]. Now assume that
(d − (α + β)I)k([x, y]) =
k
Xi=0
C i
k[(d − αI)i(x), (d − βI)k−i(y)]
(3.1)
4
M. LADRA, I. M. RIKHSIBOEV, AND R. M. TURDIBAEV
for some k > 1. Then
(d − (α + β)I)k+1([x, y]) = (d − (α + β)I) k
Xi=0
Xi=0
k[(d − αI)i+1(x), (d − βI)k−i(y)] +
Xi=0
C i
=
k
k
C i
k[(d − αI)i(x), (d − βI)k−i(y)]!
C i
k[(d − αI)i(x), (d − βI)k−i+1(y)]
= [(d − αI)k+1(x), (y)] +
C i
k[(d − αI)i+1(x), (d − βI)k+1−(i+1)(y)]
k−1
Xi=0
C i
k[(d − αI)i(x), (d − βI)k+1−i(y)] + [x, (d − βI)k+1(y)]
+
k
Xi=1
k
Xi=1
Xi=1
k
= [(d−αI)k+1(x), y]+
(C i−1
k +C i
k)[(d−αI)i(x), (d−βI)k+1−i(y)]+[x, (d−βI)k+1(y)]
= [(d − αI)k+1(x), y] +
C i
k+1[(d − αI)i(x), (d − βI)k+1−i(y)] + [x, (d − βI)k+1(y)]
=
k+1
Xi=0
C i
k+1[(d − αI)i(x), (d − βI)k+1−i(y)] .
Hence (3.1) holds for any k ∈ N.
Consider x ∈ Lα, y ∈ Lβ. Then there exist natural numbers p, q such that (d −
αI)p(x) = 0 and (d − βI)q(y) = 0. In (3.1) taking k = p + q we have that (cid:0)d − (α +
β)I(cid:1)k([x, y]) = 0 which completes the proof of the statement of the lemma.
Let d be a derivation of a Leibniz algebra L. From the definition of derivation
it is straightforward that ker d is a subalgebra. Moreover, by Lemma 3.1 we have
[L0, L0] ⊆ L0 and hence L0 is also a subalgebra of L.
(cid:3)
The following theorem is a generalization of the analogous result in the theory of Lie
algebras established in [4].
Theorem 3.2. Let D be a derivation of a Leibniz algebra L. Then there exists a unique
diagonalizable derivation D0 and a unique nilpotent derivation T such that D = D0 + T
and D0T = T D0.
Proof. Let L = Lρ1 ⊕ · · · ⊕ Lρs be a decomposition of L into characteristic spaces with
respect to d. Let us define a linear operator D0 : L → L as D0(x) = ρix for x ∈ Lρi.
Then D0 is obviously diagonalizable and D0D = DD0.
Now we show that D0 is a derivation of L.
By Lemma 3.1 if x ∈ Lρi, y ∈ Lρj we obtain [x, y] ∈ Lρi+ρj if ρi + ρj is an eigenvalue
and [x, y] = 0 otherwise. If ρi + ρj is an eigenvalue of D, then we obtain
D0([x, y]) = (ρi + ρj)[x, y] ,
[D0(x), y] + [x, D0(y)] = [ρix, y] + [x, ρjy] = (ρi + ρj)[x, y] .
So D0([x, y]) = [D0(x), y] + [x, D0(y)].
If ρi + ρj is not an eigenvalue, then [x, y] = 0 and again we obtain D0([x, y]) = 0 and
[D0(x), y] + [x, D0(y)] = (ρi + ρj)[x, y] = 0. Hence, D0 is a derivation.
AUTOMORPHISMS AND DERIVATIONS OF LEIBNIZ ALGEBRAS
5
Now denote by T = D − D0. Obviously, T is a derivation of L and T is nilpotent.
Moreover, T commutes with D0.
The uniqueness of such decomposition follows from Proposition 2.4.
(cid:3)
In order to obtain a similar result for automorphisms of Leibniz algebras we need the
following lemma.
Lemma 3.3. Let P be a nilpotent transformation of Leibniz algebra L such that P + I
is an automorphism. Then
P k([x, y]) =
k
Xi=0
i
Xj=0
for all k ∈ N.
C i
kC j
i [P k−j(x), P k−i+j(y)]
(3.2)
Proof. Let us denote Q = P + I. Since Q is an automorphism we obtain
P ([x, y]) = (Q − I)([x, y]) = [Q(x), Q(y)] − [x, y]
= [Q(x) − x, Q(y) − y] + [Q(x) − x, y] + [x, Q(y) − y]
= [P (x), P (y)] + ([P (x), y] + [x, P (y)]) =
C i
1C j
i [P 1−j(x), P 1−i+j(y)] .
1
Xi=0
i
Xj=0
Now assume that (3.2) holds for some natural k > 1. Then
P k+1([x, y]) =
k
i
=
C i
k
C j
k
Xi=0
i
Xj=0
Consider
C i
kC j
Xj=0
i P ([P k−j(x), P k−i+j(y)])
Xi=0
i (cid:0)[P k−j+1(x), P k−i+j+1(y)] + [P k−j+1(x), P k−i+j(y)] + [P k−j(x), P k−i+j+1(y)](cid:1) .
C j
i [P k−j+1(x), P k−i+j(y)] +
i
Xj=0
i
Xj=0
C j
i [P k−j(x), P k−i+j+1(y)]
= C 0
i [P k+1(x), P k−i(y)] +
i [P k+1−j(x), P k−i+j(y)]
+ C j−1
i
i [P k−i(x), P k+1(y)]
i
Xj=1(cid:0)C j
[P k+1−j(x), P k−i+j(y)](cid:1) + C i
(cid:1) [P k+1−j(x), P k+1−(i+1)+j(y)] + C i+1
i+1[P k+1(x), P k+1−(i+1)(y)]
= C 0
+
i
Xj=1(cid:0)C j
i + C j−1
i
Using the fact C j
i + C j−1
i = C j
i+1 we obtain
i+1 [P k+1−(i+1)(x), P k+1(y)] .
C j
i [P k−j+1(x), P k−i+j(y)] +
i
Xj=0
i
Xj=0
C j
i [P k−j(x), P k−i+j+1(y)]
=
i+1
Xj=0
C j
i+1[P k+1−j(x), P k+1−(i+1)+j(y)] .
6
Now
M. LADRA, I. M. RIKHSIBOEV, AND R. M. TURDIBAEV
P k+1([x, y]) =
k
Xi=0
i
C i
Xj=0
Xi=0
+
k
i+1
Xj=0
kC j
i [P k−j+1(x), P k−i+j+1(y)]
C i
kC j
i+1[P k+1−j(x), P k+1−(i+1)+j(y)]
k−1
Xi=0
i+1
Xj=0
= [P k+1(x), P k+1(y)] +
C i+1
k C j
i+1[P k−j+1(x), P k+1−(i+1)+j(y)]
+
k−1
Xi=0
i+1
Xj=0
C i
kC j
i+1[P k+1−j(x), P k+1−(i+1)+j(y)] +
C j
k+1[P k+1−j(x), P j(y)]
k+1
Xj=0
= [P k+1(x), P k+1(y)] +
k + C i
i+1[P k−j+1(x), P k+1−(i+1)+j(y)]
k−1
Xi=0
i+1
Xj=0(cid:0)C i+1
k(cid:1) C j
C j
k+1[P k+1−j(x), P j(y)] = [P k+1(x), P k+1(y)]
+
k+1
Xj=0
+
k−1
Xi=0
i+1
Xj=0
C i+1
k+1C j
i+1[P k−j+1(x), P k+1−(i+1)+j(y)] +
C j
k+1[P k+1−j(x), P j(y)]
k+1
Xj=0
= [P k+1(x), P k+1(y)] +
C i
k+1C j
i [P k−j+1(x), P k+1−i+j(y)]
k
Xi=1
i
Xj=0
+
k+1
Xj=0
C j
k+1[P k+1−j(x), P j(y)] =
k+1
Xi=0
i
Xj=0
C i
k+1C j
i [P k−j+1(x), P k+1−i+j(y)] .
Thus, (3.2) is proved.
(cid:3)
The next lemma presents the similar result for automorphisms of Leibniz algebras
as Lemma 3.1 does for derivations. Notice that, it also generalizes the result for Lie
algebras given in [4].
Lemma 3.4. Let L be a finite dimensional Leibniz algebra and L = Lρ1 ⊕ · · · ⊕ Lρs
be a decomposition of L into characteristic spaces with respect to an automorphism A.
Then for any α, β ∈ Spec(A) we have
[Lα, Lβ] ⊆(cid:26) Lαβ
0
if αβ is an eigenvalue of A
if αβ is not an eigenvalue of A .
Proof. First observe that
(A − αβI)([x, y]) = [A(x), A(y)] − αβ[x, y]
= [(A − αI)(x), (A − βI)(y)] + [(A − αI)(x), βy] + [αx, (A − βI)(y)] .
Similarly to the proof of Lemma 3.3 one can establish by induction
(A − αβI)k([x, y]) =
k
Xj=0
j
Xi=0
αiβj−iC j
kC i
j[(A − αI)k−i(x), (A − βI)k−j+i(y)] .
(3.3)
AUTOMORPHISMS AND DERIVATIONS OF LEIBNIZ ALGEBRAS
7
Now let x ∈ Lα and y ∈ Lβ. Then there exist natural numbers p, q such that
In (3.3) taking k = p + q we have that
(cid:3)
(A − αI)p(x) = 0 and (A − βI)q(y) = 0.
(A − αβI)k([x, y]) = 0 which completes the proof of the lemma.
Below, we establish a technical lemma and a corollary in order to obtain a similar
result to Theorem 3.2 for automorphisms of Leibniz algebra.
Lemma 3.5. For any polynomial P of degree less than n, where n ∈ N, the following
equality holds:
(−1)iC i
nP (i) = 0 .
n
Xi=0
Proof. Since deg P (x) < n, applying Lagrange interpolation formula to the points
qk(x)P (k), where qk(x) =
1
=
(−1)n−1 · (−1)kC k
n .
n−1
xk = k, 0 ≤ k ≤ n − 1 we obtain P (x) =
Xk=0
x(x − 1) · · ·(cid:0)x − (k − 1)(cid:1) ·(cid:0)x − (k + 1)(cid:1) · · ·(cid:0)x − (n − 1)(cid:1)
k(k − 1) · · · 1 · (−1)(−2) · · ·(cid:0) − (n − 1 − k)(cid:1)
n(n − 1) · · ·(cid:0)n − (k − 1)(cid:1) ·(cid:0)n − (k + 1)(cid:1) · · ·(cid:0)n − (n − 1)(cid:1)
k(k − 1) · · · 1 · (−1)(−2) · · ·(cid:0) − (n − 1 − k)(cid:1)
qk(n) =
Now
n!
.
=
(−1)n−1−kk!(n − k)!
Thus P (n) =
qk(n)P (k) =
1
(−1)n−1
(−1)kC k
n · P (k).
n−1
n−1
Xk=0
Xk=0
n−1
Xk=0
n
Xi=0
Hence, 0 =
(−1)kC k
n · P (k) + (−1)nC n
n P (n) =
(−1)iC i
nP (i).
(cid:3)
Corollary 3.6. Let n, m be non-negative integers such that n < m. Then
n
Xi=0
(−1)i
m − i
C i
nC n
m−i =(cid:26) 1
m
0
if n = 0
otherwise .
Proof. Let n > 1 and consider the polynomial
P (x) =
1
n!
(m − 1 − x)(m − 2 − x) · · · (m − (n − 1) − x) =
1
m − x
· C n
m−x
of degree n − 1.
By Lemma 3.5 we obtain
0 =
(−1)iC i
nP (i) =
n
Xi=0
(−1)i
m − i
C i
nC n
m−i .
n
Xi=0
For n = 0, 1, simple calculations verify the statement of the corollary.
(cid:3)
The following result shows that the analogous one established for Lie algebras [4] is
also valid for Leibniz algebras.
Theorem 3.7. Let A be an automorphism of a Leibniz algebra. Then there exists a
unique diagonalizable automorphism A0 and a unique nilpotent derivation T such that
A = A0 exp(T ) and A0T = T A0.
8
M. LADRA, I. M. RIKHSIBOEV, AND R. M. TURDIBAEV
Proof. Let L = Lρ1 ⊕ · · · ⊕ Lρs be a decomposition of a Leibniz algebra L into charac-
teristic spaces with respect to A.
Let us define a linear operator A0 : L → L as A0(x) = ρix for x ∈ Lρi. Then
A0 is obviously diagonalizable and A0A = AA0. Notice that if x ∈ Lρi, y ∈ Lρj
then [A0(x), A0(y)]) = ρiρj[x, y] and by Lemma 3.4 we have [x, y] ∈ Lρiρj . Therefore,
A0([x, y]) = ρiρj[x, y], which implies that A0 is an automorphism.
Let us denote by Q = A−1
0 A. Then A = A0Q and A0Q = QA0. Also note that
Spec(Q) = {1}.
Consider P = Q − I. Obviously, P is nilpotent and hence log Q = log(I + P ) =
P − 1
2P 2 + · · · + (−1)n−1
n P n + · · · diverges.
Since P is nilpotent, log Q is also a nilpotent transformation. We will prove that
log(I + P ) is a derivation, i.e.,
∞
Xk=1
(−1)k−1
k
P k([x, y]) =h ∞
Xk=1
(−1)k−1
k
P k(x), yi +hx,
∞
Xk=1
(−1)k−1
k
P k(y)i .
(3.4)
By Lemma 3.3, formula (3.4) is valid for P . Putting C i
k = C k−i
k
and substituting
r = k − i, we obtain P k([x, y]) =
k
Xr=0
k−r
Xj=0
C r
kC j
k−r[P k−j(x), P r+j(y)].
Now denote by Bk,r =
C r
kC j
k−r[P k−j(x), P j+r(y)] for all 0 ≤ r ≤ k.
k−r
Xj=0
Then P k([x, y]) = Bk,0 + Bk,1 + · · · + Bk,k.
Therefore,
(−1)k−1
k
P k([x, y]) =
∞
Xk=1
(−1)k−1
k
∞
Xk=1
(Bk,0 + Bk,1 + · · · + Bk,k)
B2m+1,0 −
2m + 1
1
2m
B2m,1 + · · · +
(−1)m
m + 1
B2m,0 −
1
2m − 1
B2m−1,1 + · · · +
(−1)m
m
∞
∞
=
−
2m
Xm=0 1
Xm=1 1
Xm=0 m
Xt=0
Xm=0 m
Xt=0
=
∞
∞
=
Bm+1,m!
Bm,m!
B2m−t,t!
(−1)t
2m + 1 − t
2m + 1 − t
2m+1−2t
(−1)t
B2m+1−t,t! −
Xj=0
Xm=1 m
Xt=0
(−1)t
2m − t
Xj=0
2m+1−tC j
2m−2t
C t
∞
−
∞
(−1)t
2m − t
Xm=1 m
Xt=0
2m+1−2t[P 2m+1−t−j(x), P j+t(y)]!
C t
2m−tC j
2m−2t[P 2m−t−j(x), P j+t(y)]!
AUTOMORPHISMS AND DERIVATIONS OF LEIBNIZ ALGEBRAS
9
=
∞
Xm=0 1
2m + 1
[P 2m+1(x), y] +
· [P 2m+1−s(x), P s(y)] +
+
2m−1
Xs=1 s
Xt=0
(−1)t
2m − t
C t
2m+1−2t!
1
2m
C t
2m+1−tC s−t
(−1)t
2m + 1 − t
Xs=1 s
Xt=0
[x, P 2m+1(y)](cid:19) −
Xm=1(cid:18) 1
2m−2t![P 2m−s(x), P s(y)] +
2m−tC s−t
2m
∞
2m + 1
[P 2m(x), y]
1
2m
[x, P 2m(y)]! .
Now since C t
2m+1−tC s−t
2m+1−2t = C t
sC s
2m+1−t and C t
2m−tC s−t
2m−2t = C t
sC s
2m−t we obtain
s
Xt=0
(−1)t
2m + 1 − t
C t
2m+1−tC s−t
2m+1−2t =
(−1)t
2m − t
C t
2m−tC s−t
2m−2t =
s
Xt=0
(−1)t
2m + 1 − t
C t
sC s
2m+1−t ,
(−1)t
2m − t
C t
sC s
2m−t .
s
s
Xt=0
Xt=0
However, by Corollary 3.6 the last sums are zero for all 1 ≤ s ≤ 2m (1 ≤ s ≤ 2m − 1,
respectively). Hence,
(−1)k−1
k
∞
Xk=1
P k([x, y]) =
(−1)n−1
n
∞
Xn=1
([P n(x), y] + [x, P n(y)])
= [
∞
Xn=1
(−1)n−1
n
P n(x), y] + [x,
(−1)n−1
n
P n(y)]
∞
Xn=1
and (3.4) is proved.
Thus, T = log Q is a nilpotent derivation of L and A = A0 exp(T ), A0T = T A0. Now
since exp(T ) − I is nilpotent, we obtain the additive Jordan–Chevalley decomposition
A = A0 + A0(exp(T ) − I) of A. Therefore, by Proposition 2.4 A0, and as consequence
T , are determined uniquely.
(cid:3)
The following theorems generalize the results from the theory of Lie algebras [7] to
Leibniz algebras.
Theorem 3.8. Let L be a finite-dimensional complex Leibniz algebra which admits a
non-degenerated derivation. Then L is a nilpotent algebra.
Proof. Let d be a non-singular derivation of a Leibniz algebra L and L = Lρ1 ⊕ Lρ2 ⊕
· · · ⊕ Lρk be a decomposition of L into characteristic spaces with respect to d.
Let α, β ∈ Spec(d). Then by Lemma 3.1 we have [. . . [[Lα, Lβ], Lβ], . . . , Lβ]
⊆ Lα+kβ.
k−times
{z
}
Since for sufficiently large k ∈ N we have α + kβ 6∈ Spec(d), and by Lemma 3.1 we
obtain [. . . [[Lα, Lβ], Lβ], . . . , Lβ] = 0.
nilpotent.
Thus, for x ∈ Lβ any right multiplication operator Rx is nilpotent, and due to the
i=1 R(Lρi) is
fact that α, β were taken arbitrarily, it follows that every operator fromSk
Now from identity (2.1) and Lemma 3.1 it follows thatSk
i=1 R(Lρi) is a weakly closed
set of an associative algebra R(L). Hence, by Theorem 2.7 it follows that every operator
from R(L) is nilpotent.
Now by Theorem 2.8 we obtain the result, i.e., L is nilpotent.
(cid:3)
10
M. LADRA, I. M. RIKHSIBOEV, AND R. M. TURDIBAEV
Remark 3.9. The following family L(β) = he1, . . . , eni of characteristically nilpotent
Leibniz algebras, i.e. algebras in which every derivation is nilpotent, with the following
multiplication
[e0, e0] = e2,
[e0, e1] = α3e3 + α4e4 + · · · + αn−1en−1 + θen ,
[ei, e1] = α3ei+2 + α4ei+3 + · · · + αn+1−ien
[ei, e0] = ei+1
(1 ≤ i ≤ n − 1) ,
(1 ≤ i ≤ n − 2) ,
where (α3, . . . , αn, θ ∈ C) and αiαj 6= 0 for some 3 ≤ i 6= j ≤ n, was constructed in
[13]. This implies that the statement of Theorem 3.8 in the opposite direction does not
hold.
Theorem 3.10. Let L be a finite dimensional complex Leibniz such that it admits an
automorphism of prime order with no fixed-points. Then L is a nilpotent algebra.
Proof. Let A be an automorphism of Leibniz algebra L with the properties given in the
statement of the theorem. Since A has no fixed points then 1 is not an eigenvalue of A.
Let L = Lρ1 ⊕ Lρ2 ⊕ · · · ⊕ Lρk be a decomposition of L into characteristic spaces with
respect to A. From the condition that A is an automorphism of prime order we obtain
that the spectrum of A consists of primitive p−th roots of the unity. Therefore, for any
α, β ∈ Spec(A) there exists k ∈ N such that αβk = 1 6∈ Spec(A). Hence, by Lemma 3.4
we obtain
[. . . [[Lα, Lβ], Lβ], . . . , Lβ]
⊆ Lαβk = 0 .
k−times
{z
}
Thus, for x ∈ Lβ any right multiplication operator Rx is nilpotent, and similarly as
(cid:3)
in the proof of Theorem 3.8 we obtain that L is nilpotent.
Let D be a derivation of a Leibniz algebra L such that D commutes with any in-
ner derivation. Then D(L) ⊆ AnnR(L).
Indeed, since D commutes with any right
multiplication operator we have [D(x), y] = (Ry ◦ D)(x) = (D ◦ Ry)(x) = D([x, y]) =
[D(x), y] + [x, D(y)] which implies [x, D(y)] = 0 for any x, y ∈ L. Thus, [L, D(L)] = 0
and D(L) ⊆ AnnR(L).
Lemma 3.11. Let J be an ideal of Leibniz algebra L and D be a derivation given on
L. Then J + D(J) is also an ideal of L.
Proof. Since for any x ∈ J, y ∈ L we have
[y, D(x)] = D([x, y]) − [D(x), y] ∈ D([J, L]) + [J, L] ⊆ D(J) + J ,
and so [L, D(J)] ⊆ D(J) + J. Therefore, [L, J + D(J)] ⊆ J + D(J).
Similarly, since for any x ∈ J, y ∈ L we have
[D(x), y] = D([x, y]) − [x, D(y)] ∈ D([J, L]) + [J, L] ⊆ D(J) + J ,
and so [D(J), L] ⊆ D(J) + J. Therefore, [J + D(J), L] ⊆ J + D(J). This implies that
J + D(J) is an ideal of L.
(cid:3)
Theorem 3.12. Let J be the solvable radical of a Leibniz algebra L and D be a deriva-
tion. Then D(J) ⊆ J.
Proof. By Lemma 3.11 it follows that J + D(J) is an ideal of Leibniz algebra L. We
have
(J + D(J))(2) = [J + D(J), J + D(J)] ⊆ J + [D(J), D(J)] ⊆ J + D2(J (2)) .
AUTOMORPHISMS AND DERIVATIONS OF LEIBNIZ ALGEBRAS
11
Now assume that
for some natural k > 1.
Then
(J + D(J))(k) ⊆ J + D2k−1
(J (k))
(3.5)
(J + D(J))(k+1) = [(J + D(J))(k), (J + D(J))(k)]
⊆ [J + D2k−1
(J (k)), J + D2k−1
(J (k))] ⊆ J + [D2k−1
(J (k)), D2k−1
(J (k))]
⊆ J + D2k−1+2k−1
([J (k), J (k)]) = J + D2k
(J (k+1)) .
Hence, (3.5) is verified.
Let J (m) = 0. Then (J + D(J))(m) ⊆ J + D2m−1
(J (m)) = J. Now (J + D(J))(2m−1) =
(cid:0)(J + D(J))(m)(cid:1)(m)
⊆ J (m) = 0.
Hence, J + D(J) is a solvable ideal of Leibniz algebra L. Since J is the solvable
(cid:3)
radical of L, it follows that J + D(J) ⊆ J and therefore, D(J) ⊆ J.
Remark 3.13. In Theorem 3.12 if J is the nilradical, analogous arguments establish the
invariance of J with respect to any derivation of L.
It is not difficult to verify that a derivation in a Leibniz algebra induces a derivation
in the corresponding Lie quotient algebra. However, the following example shows that
the inverse is not necessarily true, i.e., not every derivation in the Lie quotient algebra
can be extended to a derivation of the Leibniz algebra.
Example 3.14. Consider a Leibniz algebra L = he1, . . . , em, f1, . . . , fmi with the following
multiplication
[ei, ei] = fi, 1 ≤ i ≤ m
[e1, ei] = fi, 1 ≤ i ≤ m .
Then Lann = hf1, . . . , fmi and L/Lann is an abelian Lie algebra. Therefore, any linear
operator in L/Lann is a derivation.
Now consider an arbitrary derivation d : L → L.
Since [ep, e1] = 0 for p > 1, we have that 0 = d([ep, e1]) = [d(ep), e1] + [ep, d(e1)].
If d(ep) = d1pe1 +· · ·+dmpem +c1pf1 +· · ·+cmpfm then [d(ep), e1] = d1p[e1, e1] = d1pf1.
Now if d(e1) = d11e1 + · · · + dm1em + c11f1 + · · · + cm1fm then [ep, d(e1)] = dp1[ep, ep] =
dp1fp. Hence we obtain a condition d1pf1 + dp1fp = 0 which implies d1p = dp1 = 0 for
all 2 ≤ p ≤ m. Therefore, not every derivation of L/Lann can be extended to L.
Acknowledgments
The first author was supported by MICINN grant MTM 2009-14464-C02 (European
FEDER support included) and by Xunta de Galicia grant Incite 09207215 PR.
References
[1] Sh.A. Ayupov, B.A. Omirov, On Leibniz algebras. Algebra and operator theory, Proceedings of
the Colloquium in Tashkent 1997. Kluwer Academic Publishers. 1998, p. 1–12.
[2] L. Bosko, J.T. Hird, A. Mcalister, N. Schwartz, K. Stagg, Jacobson's refinement of Engel's theorem
for Leibniz algebras, arXiv:1101.2438v1, 2001.
[3] J. Dixmier, W.G. Lister, Derivations of nilpotent Lie algebras, Proc. Amer. Math. Soc. 8 (1957),
155–158.
[4] F. Gantmacher, Canonical representation of automorphisms of a complex semi-simple Lie group,
Rec. Math. (Moscou) 5(47) (1939), 101–146.
12
M. LADRA, I. M. RIKHSIBOEV, AND R. M. TURDIBAEV
[5] M. Goze, Yu.B. Khakimdjanov, Nilpotent Lie algebras. Mathematics and its Applications 361,
Kluwer Academic Publishers, Dordrecht, 1996.
[6] J.E. Humphreys, Introduction to Lie algebras and representation theory, GTM Vol. 9, Springer-
Verlag, New York-Berlin, 1972.
[7] N. Jacobson, A note on automorphisms and derivations of Lie algebras, Proc. Amer. Math. Soc.
6 (1955), 281–283.
[8] N. Jacobson, Lie algebras, Interscience Publishers, John Wiley & Sons, New York-London 1962.
[9] Yu.B. Khakimdjanov, Characteristically nilpotent Lie algebras, Math. USSR Sbornik 70 (1) 1991,
65–78.
[10] J.-L. Loday, Cyclic homology, Grundl. Math. Wiss. Bd. 301, Springer-Verlag, Berlin, 1992.
[11] J.-L. Loday, Une version non commutative des alg`ebres de Lie: les alg`ebres de Leibniz, Enseign.
Math. (2) 39 (1993), 269–293.
[12] A.I. Mal'cev, Foundations of linear algebra, W. H. Freeman & Co., San Francisco, Calif.-London
1963.
[13] B.A. Omirov, On derivations of filiform Leibniz algebras, Math. Notes 77 (5-6) (2005), 677–685.
[14] I.M. Rikhsiboev, On the nilpotence of complex Leibniz algebras (Russian), Uzbek. Mat. Zh. 2,
2005, 58-62.
Manuel Ladra: Department of Algebra, University of Santiago de Compostela, 15782 San-
tiago de Compostela, Spain
[email protected]
Ikrom Rikhsiboev: Malaysian Institute of Industrial Technology (MITEC), University
Kuala Lumpur (UniKL), 81750 Masai, Johor Darul Takzim, Malaysia
[email protected]
Rustam Turdibaev: Department of Mathematics, National University of Uzbekistan, Vuz-
gorogok, 27, 100174 Tashkent, Uzbekistan
[email protected]
|
1605.01970 | 1 | 1605 | 2016-05-05T10:32:50 | Pseudo Principal Pivot Transform: The Group Inverse Case | [
"math.RA"
] | In this short note, we prove a formula for the group inverse of a block matrix and consider the pseudo principal pivot transform expressed in terms of group inverses. Extensions of the usual principal pivot transform, where the usual inverse is replaced by the Moore-Penrose inverse, were considered in the literature. The objective here is to derive the corresponding formulae for the group generalized inverse. These are expected to be useful in our future work. | math.RA | math | Pseudo Principal Pivot Transform: The Group Inverse Case
Kavita Bisht
Department of Mathematics
Indian Institute of Technology Madras
Chennai 600 036, India
and
K.C. Sivakumar
Department of Mathematics
Indian Institute of Technology Madras
Chennai 600 036, India.
Abstract
In this short note, we prove a formula for the group inverse of a block matrix and consider
the pseudo principal pivot transform expressed in terms of group inverses. Extensions of
the usual principal pivot transform, where the usual inverse is replaced by the Moore-
Penrose inverse, were considered in the literature. The objective here is to derive the
corresponding formulae for the group generalized inverse. These are expected to be useful
in our future work.
Keywords: Group inverse, pseudo Schur complement, pseudo principal pivot trans-
form.
6
1
0
2
y
a
M
5
]
.
A
R
h
t
a
m
[
1
v
0
7
9
1
0
.
5
0
6
1
:
v
i
X
r
a
1
Introduction and Preliminaries
Let A ∈ Cn×n. If there exists X ∈ Cn×n satisfying the equations AXA = A, XAX = X
and AX = XA, then such an X can be shown to be unique. This unique solution is
called the group inverse of A and is denoted by A#. The nomenclature group inverse
arises from the fact that the positive powers of A and A# together with the idempotent
matrix AA# as the identity element, form an abelian group under matrix multiplication
and was thus named by I.Erdelyi in 1967. The group inverse of a matrix A ∈ Cn×n need
not always exist. It is well known that A# exists if and only if R(A) ∩ N(A) = {0}. In
other words, A# exists if and only if the index of A equals 1. Let us recall that the index
of a square matrix A is the smallest positive integer k such that rank(Ak) = rank(Ak+1).
If A is invertible, then the index is defined to be zero. An easy dimensionality argument
can be used to show that the index exists for any square matrix. Thus, A# exists if and
only if rank(A) = rank(A2). For more details and other properties of the group inverse,
we refer to the book [1].
Let A, B, C and D ∈ Cn×n and M =
A B
C D
. Suppose that A−1 exists. Then
the matrix D − CA−1B is called the Schur complement of A in M. Next, let D−1 exist.
Then the matrix A − BD−1C is called the associated Schur complement of D in M. The
usefulness of the Schur complement is well documented in numerous texts on matrices and
numerical analysis (see for instance, [7]). Let us turn our attention to the principal pivot
transform. Tucker introduced this notion in his studies on linear programming problems.
Let M be the block matrix, defined as above. Suppose that A−1 exists. Then the principal
pivot transform of M relative to A is defined by the matrix
A1
−A−1B
CA−1 D − CA−1B
.
Note that the bottom right block of the principal pivot transform is the Schur complement
of A in M, establishging a relationship between these notions. We refer to the recent work
[6] for some interesting properties of the principal pivot transform.
Extensions of the formulae for the Schur complement [3] and the principal pivot trans-
form [5], where the usual inverse is replaced by the Moore-Penrose inverse were studied
rather long ago. We refer to these as the pseudo Schur complement and pseudo princi-
2
pal pivot transform, respectively. In [2], some of the results of [6] and a few inheritance
properties were proved, for the pseudo Schur complement and the pseudo principal pivot
transform. In the present work, we consider the case of the group inverse, especially for
the pseudo principal pivot transform. First, we prove a formula for the group inverse of
a block matrix in the presence of some assumptions and then derive the basic properties
of the principal pivot transform.
2 The Pseudo Principal Pivot Transform in terms of
the Group Inverse.
We begin by proving a formula for the group inverse of partitioned matrices.
Theorem 2.1. Let A, B, C and D ∈ Cn×n. Suppose that A# exists. Set K = D − CA#B
and M =
A B
C D
. Suppose that K # exists. Then R(C ∗) ⊆ R(A∗), R(B) ⊆ R(A),
R(C) ⊆ R(K) and R(B ∗) ⊆ R(K ∗) if and only if
M # =
A# + A#BK #CA# −A#BK #
−K #CA#
K #
.
Proof. First, we observe that CA#A = C (since R(C ∗) ⊆ R(A∗)) and BK #K = B (since
R(B ∗) ⊆ R(K ∗)). We then have
DK #K = (K + CA#B)K #K
= K + CA#BK #K
= K + CA#B
= D.
Also,
A#B + A#BK #CA#B − A#BK #D = A#B + A#BK #(CA#B − D)
= A#B − A#BK #K
= 0.
3
Set
X =
A# + A#BK #CA# −A#BK #
−K #CA#
K #
.
From the earlier calculations one has
and
XM =
=
M X =
=
A# + A#BK #CA# −A#BK #
K #
A#A
−K #CA#
0
0 K #K
A B
C D
A# + A#BK #CA# −A#BK #
−K #CA#
K #
A B
C D
AA#
0
0 KK #
,
where we have used the facts that AA#B = B (since R(B) ⊆ R(A)) and C = KK #C
(since R(C) ⊆ R(K)). So, XM = M X. Also,
M XM =
=
= M.
A BK #K
C DK #K
A B
C D
Further,
XM X =
=
= X.
A#A
0
0 K #K
A# + A#BK #CA# −A#BK #
K #
A# + A#BK #CA# −A#BK #
−K #CA#
−K #CA#
K #
This completes the proof of the necessity part.
4
Conversely, suppose that
M # =
A# + A#BK #CA# −A#BK #
−K #CA#
K #
.
Then,
M M # =
AA# + AA#BK #CA# − BK #CA# −AA#BK # + BK #
CA# + CA#BK #CA − DK #CA# −CA#BK # + DK #
.
Equating the top left blocks of M M #M and M, we then have
AA#A + AA#BK #CA#A − BK #CA#A − AA#BK #C + BK #C = A.
This simplifies to
which in turn yields
A + (I − AA#)BK #C(I − A#A) = A,
(I − AA#)BK #C(I − A#A) = 0.
(1)
By equating the top right blocks of M M #M and M, we get
AA#B + AA#BK #CA#B − BK #CA#B − AA#BK #D + BK #D = B,
so that one has
This reduces to
AA#B + (I − AA#)BK #K = B.
(I − AA#)B(I − K #K) = 0.
Again, equating the bottom left blocks of M M #M and M,
CA#A + CA#BK #CA#A − DK #CA#A − CA#BK #C + DK #C = C,
we then have
(I − KK #)C(I − A#A) = 0.
(2)
(3)
We also have M M # = M #M and so, on equating the bottom left blocks of M M # and
M #M, we have
K #C(I − A#A) = (I − KK #)CA#.
(4)
5
Premultiplying (4) by K, we get
KK #C(I − A#A) = 0
and thus from equation (3), one has CA#A = C. So, R(C ∗) ⊆ R(A∗). Postmultiplying
(4) by A and using the fact that CA#A = C, we get
Thus R(C) ⊆ R(K). Now, equating top right blocks of M M # and M #M, one has
(I − KK #)C = 0.
(I − AA#)BK # = A#B(I − KK #).
(5)
Postmultiplying equation (5) by K, we get
(I − AA#)BK #K = 0
and from equation (2), we get AA#B = B. Thus R(B) ⊆ R(A). Finally, premultiplying
(5) by A and using the fact that AA#B = B,
Thus R(B ∗) ⊆ R(K ∗).
B(I − K #K) = 0.
Next, we state a complementary result, whose proof is similar to the proof of Theorem
2.1. Note that this result uses the pseudo Schur complement L = A − BD#C, which will
be called the complementary Schur complement.
Theorem 2.2. Let M =
A B
C D
as above. Suppose that D# exists and L = A −
BD#C. Suppose that L# exists. Then R(B ∗) ⊆ R(D∗), R(C) ⊆ R(D), R(B) ⊆ R(L)
and R(C ∗) ⊆ R(L∗) if and only if
M # =
L#
−L#BD#
−D#CL# D# + D#C #BD#
.
Next, we define the principal pivot transform in terms of the group inverse. As men-
tioned earlier, principal pivot transform involving Moore-penrose inverse of the block
matrices was studied in [5].
6
Definition 2.1. Let M be defined as above and A# exist. Then the pseudo principal pivot
transform of M relative to A is defined by
P := pppt(M, A)# =
A# −A#B
CA#
K
,
where K = D − CA#B. Next, suppose that D# exists. The complementary pseudo
principal pivot transform of M relative to D is defined by
Q := cpppt(M, D)# =
L
BD#
−D#C D#
,
where L = A − BD#C.
Both the operations of pseudo principal transforms are involutions, in the presence of
certain assumptions, as we prove next.
Lemma 2.1. Let M be defined as above. Let A# and D# exist.
(i) Suppose that R(B) ⊆ R(A) and R(C ∗) ⊆ R(A∗). Then pppt(cid:0)P, A#(cid:1)# = M.
(ii) Suppose that R(C) ⊆ R(D) and R(B ∗) ⊆ R(D∗). Then cpppt(cid:0)Q, D#(cid:1)# = M.
Proof. (i): Set W = A#, X = −A#B and Y = CA#. Then P =
. So,
W X
Y K
pppt(P, A#)# = pppt(P, W )# =
=
=
= M.
W #
−W #X
Y W # K − Y W #X
A
−A(−A#B)
CA#A D − CA#B + CA#AA#B
A B
C D
(ii): The proof is similar to part (i).
Finally, we derive the domain-range exchange property. Domain-range exchange prop-
erty in terms of Moore-Penrose inverse of the matrix is proved in [2]. This property is
well known in the nonsingular case [6].
7
Lemma 2.2. Let M be defined as earlier. Let A# and D# exist.
(i) Suppose that R(B) ⊆ R(A) and R(C ∗) ⊆ R(A∗). Then M and P = pppt(M, A)# are
related by the formula:
=
M
x1
x2
AA#y1
y2
if and only if P
y1
x2
=
A#Ax1
y2
.
(ii) Suppose that R(C) ⊆ R(D) and R(B ∗) ⊆ R(D∗). Then M and Q = cpppt(M, D)#
are related by the formula:
M
x1
x2
=
y1
DD#y2
if and only if Q
y1
x2
=
x1
D#Dy2
.
Proof. We prove (i). The proof for (ii) is similar. Suppose that
Then
and
M
x1
x2
=
AA#y1
y2
.
Ax1 + Bx2 = AA#y1
Cx1 + Dx2 = y2.
Premultipling the first equation by A# (and rearranging) we get
A#y1 − A#Bx2 = A#Ax1.
Premultiplying this equation by C, we then have
CA#y1 − CA#Bx2 = CA#Ax1 = Cx1.
So, CA#y1 + Kx2 = CA#y1 + Dx2 − CA#Bx2 = Cx1 + Dx2 = y2. Thus,
P
y1
x2
=
A#y1 − A#Bx2
CA#y1 + Kx2
=
A#Ax1
y2
.
Conversely, let
Then
P
y1
x2
=
A#Ax1
y2
.
A#y1 − A#Bx2 = A#Ax1
8
and
CA#y1 + (D − CA#B)x2 = y2.
Premultiplying the first equation by A, we have AA#y1 − Bx2 = Ax1 so that Ax1 + Bx2 =
AA#y1. Again, premultiplying the first equation by C, we get CA#y1 − CA#Bx2 = Cx1.
Hence, using the second equation we have, Cx1 + Dx2 = CA#y1 − CA#Bx2 + Dx2 = y2,
proving that
M
x1
x2
=
AA#y1
y2
.
References
[1] A. Ben-Israel and T.N.E. Greville, Generalized Inverses: Theory and Applications:
Springer-Verlag, New York, 2003.
[2] K. Bisht, G. Ravindran and K.C. Sivakumar, Pseudo Schur complements, pseudo prin-
cipal pivot transforms and their inheritance properties, Electron. J. Linear Algebra,
30, (2015), 455-477.
[3] D. Carlson, E.V. Haynsworth and T.L. Markham, A generalization of the Schur com-
plement by means of the Moore-Penrose inverse, SIAM J. Appl. Math, 26 (1974)
169-175.
[4] Kavita Bisht and K.C. Sivakumar, Pseudo Schur Complements, Pseudo Principal
Pivot Transforms and Their Inheritance Properties, (April 2015) arXiv:1504.04527, 8
pp.
[5] A.R. Meenakshi, Principal pivot transforms of an EP matrix, C.R. Math. Rep. Acad.
Sci. Canada, 8 (1986) 121-126.
[6] M. Tsatsomeros, Principal pivot transforms: Properties and applications, Lin. Alg.
Appl., 307 (2000), 151-165.
[7] F. Zhang, The Schur Complement and Its Applications, Springer, New York, 2005.
9
|
1101.3736 | 3 | 1101 | 2011-05-24T12:11:12 | On tropical dualities in cluster algebras | [
"math.RA"
] | We study two families of integer vectors playing a crucial part in the structural theory of cluster algebras: the $\gg$-vectors parameterizing cluster variables, and the $\cc$-vectors parameterizing the coefficients. We prove two identities relating these vectors to each other. The proofs depend on the sign-coherence assumption for $\cc$-vectors that still remains unproved in full generality. | math.RA | math | ON TROPICAL DUALITIES IN CLUSTER ALGEBRAS
TOMOKI NAKANISHI AND ANDREI ZELEVINSKY
Abstract. We study two families of integer vectors playing a crucial part in the
structural theory of cluster algebras: the g-vectors parameterizing cluster variables,
and the c-vectors parameterizing the coefficients. We prove two identities relating
these vectors to each other. The proofs depend on the sign-coherence assumption
for c-vectors that still remains unproved in full generality.
1
1
0
2
y
a
M
4
2
]
.
A
R
h
t
a
m
[
3
v
6
3
7
3
.
1
0
1
1
:
v
i
X
r
a
1. Introduction
In the theory of cluster algebras, the central role is played by two families of el-
ements obtained by certain algebraic recurrences: cluster variables and coefficients
(V. Fock and A. Goncharov refer to them as a-coordinates, and x-coordinates, respec-
tively). In this paper we study certain tropical limits of these elements: two families
of integer vectors (g-vectors and c-vectors, respectively) introduced in [5]. Despite
the fact that these two families are given by very different piecewise-polynomial re-
currences, it turns our that they are closely related to each other. We refer to the
identities involving g-vectors and c-vectors as tropical dualities. The main result of
this paper is Theorem 1.2 containing two such dualities.
We start by briefly recalling the formalism of cluster algebras (more details can
be found in [5]). In the heart of this formalism there are several discrete dynami-
cal systems on a n-regular tree given by birational recurrences and their "tropical"
versions. More precisely, we fix a positive integer n, and denote by Tn an n-regular
tree whose edges are labeled by the numbers 1, . . . , n, so that the n edges emanating
k−−− t′ to indicate that vertices
from each vertex receive different labels. We write t
t, t′ ∈ Tn are joined by an edge labeled by k.
We need a little algebraic preparation. We call a semifield an abelian multiplicative
group P, supplied with the addition operation ⊗, which is commutative, associative,
and distributive with respect to multiplication. Let ZP denote the integer group ring
of (the multiplicative group of) P (note that its definitionignores the addition ⊕). It
is easy to show that (the multiplicative group of) P is torsion-free (see the remark
before Definition 5.3 in [4]), hence ZP is a domain, hence it has the field of fractions
Q(P).
We call an ambient field a field F isomorphic to the field of rational functions in
n indeterminates with the coefficients from Q(P).
Fix a semifield P and an ambient field F . Following [5], we call a (labeled) seed a
triple (x, y, B), where
Date: January 19, 2011; revised April 13, 2011 and May 24, 2011.
Research of A. Z. supported by the NSF grant DMS-0801187.
1
2
TOMOKI NAKANISHI AND ANDREI ZELEVINSKY
• B = (bij) is an n×n integer matrix which is skew-symmetrizable, that is, its
transpose BT is equal to −DBD−1 for some diagonal matrix D with positive
integer diagonal entries d1, . . . , dn;
• y = (y1, . . . , yn) is an n-tuple of elements of P, and
• x = (x1, . . . , xn) is an n-tuple of elements of F forming a free generating set,
that is, being algebraically independent and such that F = Q(P)(x1, . . . , xn).
We refer to B as the exchange matrix of a seed, y as the coefficient tuple, and x
as the cluster.
Throughout the paper we use the notation [b]+ = max(b, 0). For k = 1, . . . , n, the
seed mutation µk transforms (x, y, B) into the labeled seed µk(x, y, B) = (x′, y′, B ′)
defined as follows:
• The entries of the exchange matrix B ′ = (b′
ij) are given by
b′
ij =(−bij
bij + [bik]+[bkj]+ − [−bik]+[−bkj]+ otherwise.
if i = k or j = k;
• The coefficient tuple y′ = (y′
1, . . . , y′
n) is given by
y′
j =(y−1
k
yjy
1, . . . , x′
[bkj]+
(yk ⊕ 1)−bkj
k
n) is given by x′
if j = k;
if j 6= k.
• The cluster x′ = (x′
is determined by the exchange relation
j = xj for j 6= k, whereas x′
k ∈ F
(1.1)
(1.2)
(1.3)
x′
k =
yk Q x[bik]+
i
(yk ⊕ 1)xk
+ Q x[−bik]+
i
.
It is easy to see that B ′ is skew-symmetrizable (with the same choice of D),
implying that (x′, y′, B ′) is indeed a seed. Furthermore, the seed mutation µk is
involutive, that is, it transforms (x′, y′, B ′) into the original seed (x, y, B). We
shall also use the notation B ′ = µk(B) (resp. (y′, B ′) = µk(y, B)) and call the
transformation B 7→ B ′ the matrix mutation (resp. (y, B) 7→ (y′, B ′) the Y -seed
mutation).
A seed pattern is an assignment of a seed (xt, yt, Bt) to every vertex t ∈ Tn, such
k−−− t′ are obtained from each
that the seeds assigned to the endpoints of any edge t
other by the seed mutation µk. We write:
(1.4)
xt = (x1;t , . . . , xn;t) , yt = (y1;t , . . . , yn;t) , Bt = (bij;t) .
We also refer to a family (yt, Bt)t∈Tn as a Y -seed pattern, and a family (Bt)t∈Tn as
an exchange matrix pattern.
It is convenient to fix a vertex (root) t0 ∈ Tn. Then a seed pattern is uniquely
determined by a seed at t0, which can be chosen arbitrarily. We will just write this
initial seed as (x, y, B), with
(1.5)
x = (x1 , . . . , xn) , y = (y1 , . . . , yn) , B = (bij) .
The cluster algebra associated with a seed (x, y, B) (or rather with the correspond-
ing seed pattern) is the ZP-subalgebra of the ambient field F generated by all cluster
ON TROPICAL DUALITIES IN CLUSTER ALGEBRAS
3
variables xj;t. One of the central problems in the theory of cluster algebras is to find
explicit expressions for all xj;t and yj;t as rational functions in x1, . . . , xn, y1, . . . , yn.
A step in this direction was made in [5], where the following was proved.
Proposition 1.1. [5, Proposition 3.13, Corollary 6.3] Every pair (B; t0) gives rise
to a family of polynomials Fj;t = F B;t0
j;t ∈ Z[u1, . . . , un] and two families of integer
vectors cj;t = cB;t0
j;t = (g1j;t, . . . , gnj;t) ∈ Zn
(where j ∈ {1, . . . , } and t ∈ Tn) with the following properties:
j;t = (c1j;t, . . . , cnj;t) ∈ Zn and gj;t = gB;t0
(1) Each Fj;t is not divisible by any ui, and can be expressed as a ratio of two poly-
nomials in u1, . . . , un with positive integer coefficients, thus can be evaluated
in every semifield P.
(2) For any j and t, we have
(1.6)
yj;t = y
c1j;t
1
· · · ycnj;t
(3) For any j and t, we have
Fi;tP(y1, . . . , yn)bij;t .
n Yi
(1.7)
g1j;t
xj;t = x
1
· · · xgnj;t
n
Fj;tF (y1, . . . , yn)
Fj;tP(y1, . . . , yn)
,
where the elements yj are given by
bij
x
i
.
yj = yjYi
Following [5], we refer to gj;t as the g-vector of a cluster variable xj;t. It is in-
structive to view the g-vectors as some kind of discrete (or tropical ) limits of cluster
variables. In fact, as explained in [5, Remark 7.15], these vectors are expected to
provide a parametrization of cluster variables (and more generally, cluster monomi-
als) by "Langlands dual tropical Y -seed patterns" in accordance with a conjecture
by V. Fock and A. Goncharov [3, Conjecture 5.1]. Some properties of g-vectors
and F -polynomials were established in [5] but several basic properties still remain
conjectural.
Comparing (1.6) and (1.7), it is natural to view the c-vector cj;t as a tropical limit
of a coefficient yj;t. These vectors also made their appearance in [5], although not
under this name, and only as a tool for studying g-vectors and F -polynomials. In this
note we focus on the properties of c-vectors and their relationships with g-vectors.
The following (unfortunately, still conjectural) sign-coherence property of c-vectors
is crucial for our analysis:
(1.8) Each vector cj;t has either all entries nonnegative or all entries nonpositive.
As shown in [5, Proposition 5.6], (1.8) is equivalent to the following conjecture made
in [5, Conjecture 5.4]:
(1.9)
Each polynomial Fj;t(u1, . . . , un) has constant term 1.
This conjecture (hence the property (1.8)) was proved in [2] for the case of skew-
symmetric exchange matrices, using quivers with potentials and their representations
(two different proofs were recently given in [6, 8]).
4
TOMOKI NAKANISHI AND ANDREI ZELEVINSKY
We denote by C B;t0
) the integer matrix with columns c1;t, . . . , cn;t
(resp. with columns g1;t, . . . , gn;t), where B is the exchange matrix at t0. In partic-
ular, we have
(resp. GB;t0
t
t
(1.10)
C B;t0
t0 = GB;t0
t0 = I
(the identity matrix).
The main results of this paper are the following two identities.
Theorem 1.2. Under the assumption (1.8), for any skew-symmetrizable exchange
matrix B, and any t0, t ∈ Tn, we have
(1.11)
and
(GB;t0
t
)T = (C −BT ;t0
t
)−1,
(1.12)
where t 7→ Bt is the exchange matrix pattern on Tn such that Bt0 = B, and BT
stands for the transpose matrix of B.
C B;t0
t = (C −Bt;t
)−1,
t0
Note that in the case when B is skew-symmetric, (1.11) takes the form (GB;t0
)T =
t
(C B;t0
)−1. In this case it was proved in [7, Proposition 4.1].
t
One can show that identities (1.11) and (1.12) imply most of the conjectures made
in [5] and recast in [2], namely Conjectures 1.1 - 1.4 and 1.6 from [2]. For instance,
by combining them, we obtain the equality
(1.13)
(GB;t0
t
)T = C
t ;t
BT
t0
,
which can be shown to imply [2, Conjecture 1.6]. More details will be given in
Section 4.
In contrast with [2], the proofs of (1.11) and (1.12) given below do not use any
categorical interpretation of g- or c-vectors but are based on the analysis of recur-
rences satisfied by them. To describe these recurrences we need to develop some
matrix formalism.
We extend the notation [b]+ = max(0, b) to matrices, writing [B]+ for the matrix
obtained from B by applying the operation b 7→ [b]+ to all entries of B. For a matrix
index k, we denote by B •k the matrix obtained from B by replacing all entries outside
of the k-th column with zeros; the matrix Bk• is defined similarly using the k-th row
instead of the column. Note that the operations B 7→ [B]+ and B 7→ B •k commute
with each other, making the notation [B]•k
+ ) unambiguous.
+ (and [B]k•
Using this formalism, we can rephrase the sign-coherence condition (1.8) for a
matrix C = C B;t0
(1.14) For every j, there exists the sign εj(C) = ±1 such that [−εj(C)C]•j
as follows:
t
+ = 0.
Let Jk denote the diagonal matrix obtained from the identity matrix by replacing
the (k, k)-entry with −1. We deduce (1.11) from the following proposition.
Proposition 1.3. Suppose t
ℓ−−− t′ in Tn, and let C = C B;t0
t
, G = GB;t0
t
, C ′ = C B;t0
t′
,
and G′ = GB;t0
t′
(1.15)
. Then, under the assumption that C satisfies (1.14), we have
C ′ = C(Jℓ + [εℓ(C)Bt]ℓ•
+ ), G′ = G(Jℓ + [−εℓ(C)Bt]•ℓ
+ ) .
ON TROPICAL DUALITIES IN CLUSTER ALGEBRAS
5
As for (1.12), we prove it together with the following proposition that provides
"at the opposite end." We still assume
a recurrence relation for the matrices C B;t0
that every matrix of the form C = C B;t0
t
satisfies (1.8) or equivalently (1.14).
t
Proposition 1.4. Suppose t0
abbreviate Ct = C B;t0
, and C ′
t
t = C B1;t1
t
. Then we have
k−−− t1 in Tn, and let B1 = µk(B). For t ∈ Tn,
(1.16)
C ′
t = (Jk + [−εk(C −Bt;t
t0
)B]k•
+ )Ct .
Proposition 1.3 and the identity (1.11) will be proved in Section 2, while Propo-
sition 1.4 and (1.12) will be proved in Section 3. In the last section we give some
corollaries.
2. Proofs of Proposition 1.3 and identity (1.11)
Proof of Proposition 1.3. We start with the following identity involving the function
[b]+ = max(b, 0): for any real numbers b and c, and a sign ε = ±1, we have
(2.1)
[c]+[b]+ − [−c]+[−b]+ = c[εb]+ + b[−εc]+ .
This is a consequence of an obvious identity
(2.2)
Indeed, we have
[b]+ − [−b]+ = b .
c[b]+ + b[−c]+ = ([c]+ − [−c]+)[b]+ + ([b]+ − [−b]+)[−c]+ = [c]+[b]+ − [−c]+[−b]+ ,
proving (2.1) for ε = 1; the case ε = −1 is proved similarly.
The first equality in (1.15) was essentially shown in [1, (3.2)] (in a different nota-
tion). For the convenience of the reader, we reproduce the argument here. Recall
from [5, Remarks 3.2, 3.14] that the relationship between C = C B;t0
and C ′ = C B;t0
can be described as follows: if Bt is a 2n × n integer matrix with the top n × n block
, then Bt′ is obtained from Bt by the matrix mutation
Bt and the bottom block C B;t0
µℓ given by the same formula as in (1.1). Thus, we get
t′
t
t
(2.3)
c′
ij =(−cij
cij + [ciℓ]+[bℓj;t]+ − [−ciℓ]+[−bℓj;t]+ otherwise.
if j = ℓ;
Using (2.1) and rewriting the resulting formula in the matrix form, we get
(2.4)
C ′ = C(Jℓ + [εBt]ℓ•
+ ) + [−εC]•ℓ
+ Bt
for any choice of the sign ε. Setting ε = εℓ(C) and remembering (1.14), we see that
the second term in (2.4) disappears, implying the desired equality.
To prove the second equality in (1.15), we rewrite [5, (6.12)-(6.13)] in the matrix
form as follows:
(2.5)
G′ = G(Jℓ + [−εBt]•ℓ
+ ) − B[−εC]•ℓ
+
(again for any choice of the sign ε). Again setting ε = εℓ(C) makes the second term
in (2.5) disappear, completing the proof of Proposition 1.3.
(cid:3)
6
TOMOKI NAKANISHI AND ANDREI ZELEVINSKY
Proof of identity (1.11). We prove (1.11) by induction on the distance between t and
t0 in the tree Tn. For t = t0, the claim follows from (1.10). It remains to show that
ℓ−−− t′.
if (1.11) holds for some t ∈ Tn then it also holds for t′ ∈ Tn such that t
This implication is a consequence of (1.15) combined with the following three simple
observations:
• For every integer square matrix B and an index ℓ such that bℓℓ = 0, we have
(2.6)
(Jℓ + B •ℓ)T = Jℓ + (BT )
ℓ•
,
(Jℓ + Bℓ•)−1 = Jℓ + Bℓ• .
• The replacement of the initial exchange matrix Bt0 = B with −BT leads to
the replacement of each Bt with −BT
t , and we have
(2.7)
C −BT ;t0
t
= DC B;t0
t D−1,
where D is a diagonal n × n matrix with positive integer entries such that
−BT = DBD−1.
• In particular, we have εℓ(C −BT ;t0
t
) = εℓ(C B;t0
t
) for every matrix index ℓ.
(cid:3)
Remark 2.1. As in Proposition 1.3, let us abbreviate C = C B;t0
. In
the equality (2.5), the fact that two choices of the sign give the same answer implies
the following identity for G (shown already in [5, (6.14)]):
, and G = GB;t0
t
t
(2.8)
GBt = BC .
Combining this identity with (1.11) and (2.7), we obtain the following:
(2.9)
DBt = C T · DB · C ,
where the diagonal matrix D is as in (2.7). This shows that the exchange matrix Bt
at every vertex t ∈ Tn (that is, the top part of the 2n × n matrix Bt) is determined
by the initial exchange matrix B and the matrix C which is the bottom part of Bt.
3. Proofs of Proposition 1.4 and identity (1.12)
We prove (1.16) and (1.12) by a simultaneous induction. To prepare the ground
for it, we note that, under the notation and assumptions of Proposition 1.4, we have
Bk•
). In view
of the second equality in (2.6), this allows us to interchange t0 and t1 in (1.16). Thus,
it suffices to prove (1.16) under the assumption that
1 = −Bk•, and (C −Bt;t
)•k, hence εk(C −Bt;t
) = −εk(C −Bt;t
)•k = −(C −Bt;t
t1
t0
t1
t0
(3.1)
t0 belongs to the (unique) path between t and t1 in Tn.
In other words, (3.1) means that d(t, t1) = d(t, t0) + 1, where d(t, t0) denotes the
distance between t and t0 in Tn.
For d ≥ 0, we denote by (Id) and (IId) the following two statements:
(Id) The equality (1.12) holds for d(t, t0) = d.
(IId) The equality (1.16) holds under the assumption (3.1) whenever d(t, t0) = d.
ON TROPICAL DUALITIES IN CLUSTER ALGEBRAS
7
We will prove both (Id) and (IId) simultaneously by induction on d.
For d = 0 we have t = t0, and both (1.12) and (1.16) are immediate from the
definitions.
It remains to prove the implications (Id) & (IId) =⇒ (Id+1) and (Id) &(IId) =⇒
(IId+1).
Proof of the implication (Id) & (IId) =⇒ (Id+1). We need to show that if (1.12) and
(1.16) hold for some t, t0, and t1 satisfying (3.1), then (1.12) also holds if t0 is replaced
with t1, and B is replaced by B1 = µk(B) (note that this replacement leaves Bt
unchanged by the definition of an exchange matrix pattern). Let us abbreviate
C = C B;t0
. To prove the desired equality
t
C B1;t1
t
= (C −Bt;t
t1
)−1
we use (1.16), the second equality in (2.6), the inductive assumption C −1 = C −Bt;t
and the first equality in (1.15) to get
t0
,
(C B1;t1
t
)−1 = ((Jk + [−εk(C −Bt;t
t0
)B]k•
= C −1(Jk + [−εk(C −Bt;t
= C −Bt;t
(Jk + [εk(C −Bt;t
= C −Bt;t
t0
t0
t0
t1
+ ) C)−1
)B]k•
+ )
) · (−B)]k•
+ )
(to see that the last equality is indeed an instance of (1.15), note that the matrix
mutation commutes with the operation B 7→ −B, thus the exchange matrix pattern
that assigns −Bt to t, also assigns −B to t0).
(cid:3)
Proof of the implication (Id) & (IId) =⇒ (IId+1). This proof is more involved than
the previous one. Let us first summarize our assumptions, and what has to be proven.
ℓ−−− t′,
Suppose d(t0, t) = d in Tn, and let t1, t′ ∈ Tn be such that t0
k−−− t1, t
and t and t0 belong to the unique path between t′ and t1. Thus we have d(t1, t) =
d(t0, t′) = d + 1, and d(t1, t′) = d + 2. The assumption (Id) is just the equality (1.12).
Since we have already shown that (Id) and (IId) imply (Id+1), we can also assume
the equalities
(3.2)
C B1;t1
t
= (C −Bt;t
t1
)−1, C B;t0
t′ = (C
−Bt′ ;t′
t0
)−1 .
(3.3)
Now the assumption (IId) gives us the equalities
= (Jk + [−εk(C −Bt;t
t0
= (Jℓ + [εℓ(C B;t0
C B1;t1
t
−Bt′ ;t′
t0
C
t
)Bt]ℓ•
,
)B]k•
t
+ )C B;t0
+ )C −Bt;t
t0
,
while our goal is to prove that
(3.4)
C B1;t1
t′
= (Jk + [−εk(C
−Bt′ ;t′
t0
)B]k•
+ )C B;t0
t′
.
To prove (3.4), we invoke the following equalities which are instances of (1.15):
(3.5) C B;t0
t′ = C B;t0
t
(Jℓ + [εℓ(C B;t0
t
)Bt]ℓ•
+ ), C B1;t1
t′
= C B1;t1
t
(Jℓ + [εℓ(C B1;t1
t
)Bt]ℓ•
+ ) .
8
TOMOKI NAKANISHI AND ANDREI ZELEVINSKY
The following observation is immediate from the first equality in (3.3):
(3.6) The transformation C B;t0
t
7→ C B1;t1
t
affects only the entries in the k-th row.
Now we consider separately the following two mutually exclusive cases:
Case 1: the matrix C B;t0
Case 2: the only non-zero entry in the ℓ-th column of C B;t0
has a non-zero entry (i, ℓ) for some i 6= k.
is the (k, ℓ) entry;
equivalently, the ℓth column of C B;t0
is of the form εek, where e1, . . . , en is the
standard basis of Zn. Note that in this case the coefficient ε is equal to ±1, since
both C B;t0
and its inverse C −Bt;t
are integer matrices.
t
t
t
t
t0
First we deal with Case 1. In view of (3.6), we have εℓ(C B1;t1
) (here
we use the assumption (1.14)!). Combining (3.5) and (3.3), we see that the desired
equality (3.4) follows from the equality
) = εℓ(C B;t0
t
t
(3.7)
εk(C
−Bt′ ;t′
t0
) = εk(C −Bt;t
t0
) .
Applying the same argument as above to the transformation C −Bt;t
given
by the second equality in (3.3), we see that it is enough to show that C −Bt;t
satisfies
the analog of Case 1, namely has a non-zero entry (i, k) for some i 6= ℓ. Now recall
that by our assumption (Id), the matrices C B;t0
are inverses of each other.
Thus, it is enough to show the following statement from linear algebra:
and C −Bt;t
7→ C
−Bt′ ;t′
t0
t0
t0
t0
t
(3.8)
For any indices k and ℓ, an invertible matrix C has a non-zero entry
(i, ℓ) for some i 6= k if and only if C −1 has a non-zero entry (i, k) for some i 6= ℓ.
The easiest way to prove (3.8) is to observe that it becomes obvious after replacing
the equivalent statements in question by their negations:
(3.9)
For any k and ℓ, the ℓth column of C is of the form εek
with ε = ±1 if and only if the kth column of C −1 is εeℓ.
This completes the proof of (3.4) in Case 1.
Now assume that we are in Case 2, that is, in view of (3.9), the ℓth column
is εeℓ for some ε = ±1.
is εek, while the kth column of (C B;t0
)−1 = C −Bt;t
t
t0
of C B;t0
In particular, we have
t
εℓ(C B;t0
t
) = εk(C −Bt;t
t0
) = ε .
Using the two equalities in (3.3), we conclude that the ℓth column of C B1;t1
while the kth column of C
is −εeℓ, hence
t
is −εek,
−Bt′ ;t′
t0
εℓ(C B1;t1
t
) = εk(C
−Bt′ ;t′
t0
) = −ε .
Combining (3.5) and (3.3), we can rewrite the desired equality (3.4) as follows:
(Jk + [−εB]k•
+ )C B;t0
t
(Jℓ + [−εBt]ℓ•
+ ) = (Jk + [εB]k•
+ )C B;t0
t
(Jℓ + [εBt]ℓ•
+ ) ,
or equivalently (in view of (2.6)) as
(Jk + [εB]k•
+ )(Jk + [−εB]k•
+ )C B;t0
t = C B;t0
t
(Jℓ + [εBt]ℓ•
+ )(Jℓ + [−εBt]ℓ•
+ ) .
Performing the matrix multiplication and remembering (2.2), we can simplify the
last equality to
(I + εBk•)C B;t0
t = C B;t0
t
(I + εBℓ•
t ) ,
ON TROPICAL DUALITIES IN CLUSTER ALGEBRAS
9
or even further to
(3.10)
Bk•C B;t0
t = C B;t0
t Bℓ•
t
.
To prove (3.10) we recall the identity (1.11), and in particular the matrix GB;t0
t
that appears there. To simplify the notation, we abbreviate
GB;t0
t = G, C B;t0
t = C ;
as shown in the proof of (1.11) given above, it can be rewritten as follows:
(3.11)
GT = (C −BT ;t0
t
)−1 = DC −1D−1 ,
where D is the diagonal matrix with positive diagonal entries such that −BT =
DBD−1.
k εeℓ. However, we
know that both GT and its inverse C −BT ;t0
are integer matrices. Therefore, dℓ = dk,
and we have
In particular, the kth column of GT is equal to dℓd−1
t
(3.12)
the kth row of G is equal to εeT
ℓ .
Now we recall the equality (2.8). Computing the entries in the kth row on both sides
of (2.8) and using (3.12), we get, for any j = 1, . . . , n:
εbℓj;t =Xp
bkpcpj .
Rewriting this in the matrix form yields (3.10), thus completing the proofs of Propo-
sition 1.4 and identity (1.12).
(cid:3)
4. Some corollaries
In this section we show that the assumption (1.8) and the identities (1.11) and
(1.12) imply most of the conjectures made in [5] and recast in [2], namely Conjec-
tures 1.1 - 1.4 and 1.6 from [2]. For the convenience of the reader we reproduce their
statements.
Conjecture 4.1.
(i) Each polynomial F B;t0
j;t
has constant term 1.
(ii) Each polynomial F B;t0
has a unique monomial of maximal degree. Further-
more, this monomial has coefficient 1, and it is divisible by all the other
occurring monomials.
j;t
(iii) For every t ∈ Tn, the vectors gB;t0
1;t
n;t are sign-coherent, i.e., for any
i = 1, . . . , n, the i-th components of all these vectors are either all nonnega-
tive, or all nonpositive.
, . . . , gB;t0
(iv) For every t ∈ Tn, the vectors gB;t0
1;t
form a Z-basis of the lattice Zn.
k−−− t1 be two adjacent vertices in Tn, and let B ′ = µk(B). Then,
, . . . , gB;t0
n;t
(v) Let t0
for any t ∈ Tn and j = 1, . . . , n, the g-vectors gB;t0
gB ′;t1
j;t = (g′
n) are related as follows:
1, . . . , g′
j;t = (g1, . . . , gn) and
(4.1)
g′
i =(−gk
gi + [bik]+gk − bik min(gk, 0)
if i = k;
if i 6= k.
10
TOMOKI NAKANISHI AND ANDREI ZELEVINSKY
Proposition 4.2. The assumption (1.8) implies all the statements in Conjecture 4.1.
Proof. The equivalence of (i) and (ii) was shown in [5, Section 5] (see Conjectures 5.4,
5.5 and Proposition 5.3 there). The fact that (1.8) implies conjectures (i) and (ii)
was shown in [5, Proposition 5.6] (note that (1.8) appears as condition (ii′) in the
proof).
Part (iii) is immediate from (1.8) and (1.13).
Part (iv) can be restated by saying that the matrix GB;t0
t
invertible over Z. But this is immediate from (1.11).
with columns gB;t0
j;t
is
It remains to prove (v). Replacing min(gk, 0) with −[−gk]+, and using (2.1), we
rewrite the desired equality (4.1) in the matrix form as follows:
(4.2)
GB ′;t1
t
= (Jk + [εB]•k
+ )GB;t0
t + B[−εGB;t0
t
]k•
+ ,
for any choice of sign ε = ±1. Taking transpose matrices on both sides of (4.2) and
using (1.13), we can rewrite (4.2) as
(4.3)
t1 = C BT
C BT
t ;t
t ;t
t0
(Jk + [εBT ]k•
+ ) + [−εC BT
t ;t
t0
]•k
+ BT .
In view of (1.14), choosing ε = εk(C
t ;t
BT
t0
), we see that (4.3) takes the form
t ;t
BT
t1 = C
BT
t0
t ;t
C
(4.4)
which is just the first equality in (1.15) (with (t, t′) replaced by (t0, t1), and (B; t0)
replaced by (BT
t ; t)). This proves (4.2) and completes the proof of Proposition 4.2.
(cid:3)
(Jk + [εk(C
+ ) ,
t ;t
BT
t0
) BT ]k•
References
[1] A. Berenstein, S. Fomin, A. Zelevinsky, Cluster algebras III: Upper bounds and double Bruhat
cells, Duke Math. J. 126 (2005), 1 -- 52.
[2] H. Derksen, J. Weyman, A. Zelevinsky, Quivers with potentials and their representations II:
Applications to cluster algebras, J. Amer. Math. Soc. 23 (2010), no. 3, 749 -- 790.
[3] V. V. Fock, A. B. Goncharov, Cluster ensembles, quantization and the dilogarithm, Annales
Sci. de l' ´Ecole Norm. Sup. (4) 42 (2009), 865-930.
[4] S. Fomin, A. Zelevinsky, Cluster algebras I: Foundations, J. Amer. Math. Soc. 15 (2002),
497 -- 529.
[5] S. Fomin, A. Zelevinsky, Cluster algebras IV: Coefficients, Comp. Math. 143 (2007), 112 -- 164.
[6] K. Nagao, Donaldson-Thomas theory and cluster algebras, arXiv:1002.4884.
[7] T. Nakanishi, Periodicities in cluster algebras and dilogarithm identities, arXiv:1006.0632.
[8] P. Plamondon, Cluster algebras via cluster categories with infinite-dimensional morphism
spaces, arXiv:1004.0830.
Graduate School of Mathematics, Nagoya University, Nagoya, 464-8604, Japan
E-mail address: [email protected]
Department of Mathematics, Northeastern University, Boston, MA 02115, USA
E-mail address: [email protected]
|
1001.0191 | 2 | 1001 | 2012-11-30T22:35:27 | Group gradings on restricted Cartan type Lie algebras | [
"math.RA"
] | For a given abelian group G, we classify the isomorphism classes of G-gradings on the simple restricted Lie algebras of types W(m;1) and S(m;1) (m>=2), in terms of numerical and group-theoretical invariants. Our main tool is automorphism group schemes, which we determine for the simple restricted Lie algebras of types S(m;1) and H(m;1). The ground field is assumed to be algebraically closed of characteristic p>3. | math.RA | math | GROUP GRADINGS ON RESTRICTED CARTAN TYPE
LIE ALGEBRAS
YURI BAHTURIN AND MIKHAIL KOCHETOV
Abstract. For a given abelian group G, we classify the isomorphism classes of
G-gradings on the simple restricted Lie algebras of types W (m; 1) and S(m; 1)
(m ≥ 2), in terms of numerical and group-theoretical invariants. Our main tool
is automorphism group schemes, which we determine for the simple restricted
Lie algebras of types S(m; 1) and H(m; 1). The ground field is assumed to be
algebraically closed of characteristic p > 3.
]
.
A
R
h
t
a
m
[
2
v
1
9
1
0
.
1
0
0
1
:
v
i
X
r
a
1. Introduction
Let U be an algebra (not necessarily associative) over a field F and let G be a
group, written multiplicatively.
Definition 1.1. A G-grading on U is a vector space decomposition
U =Mg∈G
Ug
such that
UgUh ⊂ Ugh for all g, h ∈ G.
Ug is called the homogeneous component of degree g. The support of the G-grading
is the set
{g ∈ G Ug 6= 0}.
If U is finite-dimensional, then, replacing G with the subgroup generated by the
support of the grading, we may assume without loss of generality that G is finitely
generated.
ψ(Ug) = U ′
g for all g ∈ G,
are isomorphic if there exists an algebra automorphism ψ : U → U such that
Definition 1.2. We say that two G-gradings, U =Lg∈G Ug and U =Lg∈G U ′
i.e., U =Lg∈G Ug and U =Lg∈G U ′
Since the same vector space decomposition can be regarded as a G-grading for
different groups G, one can also define equivalence of gradings, which is a weaker
relation than isomorphism -- see e.g. [6] for a discussion. Fine gradings (i.e., those
that cannot be refined) as well as their universal groups are of particular interest.
We are interested in the problem of finding all possible group gradings on finite-
dimensional simple Lie algebras. If L is a simple Lie algebra, then it is known that
g are isomorphic as G-graded algebras.
g,
Keywords: graded algebra, simple Lie algebra, grading, Cartan type Lie algebra.
2000 Mathematics Subject Classification: Primary 17B70; Secondary 17B60.
The first author acknowledges support by by NSERC grant # 227060-04. The second author
acknowledges support by NSERC Discovery Grant # 341792-07.
1
2
BAHTURIN AND KOCHETOV
the support of any G-grading on L generates an abelian group. Hence, in this paper
we will always assume that G is abelian. We will also assume that the ground field
F is algebraically closed.
For an arbitrary abelian group G, the classification of G-gradings (up to isomor-
phism) is known for almost all classical simple Lie algebras over an algebraically
closed field of characteristic 0 or p > 2. The classification of fine gradings (up to
equivalence) is also known. The reader may refer to [2, 5] and references therein.
Our goal is to carry out the same classification for simple restricted Lie algebras
of Cartan type. In the present paper, we achieve this goal for Witt algebras and
for special algebras (see definitions in Section 2) in characterisitc p > 3: Theorem
4.13 with Corollary 4.14 and Theorem 4.17 with Corollary 4.18, respectively.
In a number of cases, a fruitful approach to the classification of gradings by
abelian groups on an algebra U is to use another algebra, R, that shares with U the
automorphism group scheme (see Section 3) and whose gradings are easier to study.
This approach often requires equipping R with some additional structure. For
classical simple Lie algebras of series A, B, C and D, R is the matrix algebra Mn(F),
possibly equipped with an antiautomorphism. For Lie algebras of Cartan type, R is
the "coordinate algebra" O(m; n) (see Definition 2.1), which, in the restricted case,
is just the truncated polynomial algebra O(m; 1) = F[x1, . . . , xm]/(xp
m). It
is not difficult to classify G-gradings on the algebra O(m; 1) -- see Theorem 4.8
and Corollary 4.11. Since it is known that O(m; 1) has the same automorphism
group scheme as the Witt algebra W (m; 1), this immediately gives the classification
of gradings for the latter.
In Section 3, we show that the automorphism group
scheme of the special algebra S(m; 1)(1), m ≥ 3 -- respectively, Hamiltonian algebra
H(m; 1)(2), m = 2r -- is isomorphic to the stabilizer of the differential form ωS --
respectively, ωH -- in the automorphism group scheme of O(m; 1); see Theorems 3.2
and 3.5, respectively. As a consequence, all gradings on S(m; 1)(1) and H(m; 1)(2)
come from gradings on O(m; 1). However, this does not yet give a classification
of gradings on S(m; 1)(1) and H(m; 1)(2) up to isomorphism, because one must
make sure that the isomorphism preserves the appropriate differential form -- see
Corollary 4.2. With some more work, we obtain a classification of gradings on
S(m; 1)(1), m ≥ 3, and S(2; 1)(2) = H(2; 1)(2). The classification for H(2r; 1)(2)
with r > 1 remains open.
1, . . . , xp
The paper is structured as follows. In Section 2, we recall the definitions and
basic facts regarding Lie algebras of Cartan type. In Section 3, we briefly recall
background information on automorphism group schemes and determine them for
S(m; 1)(1) and H(m; 1)(2). In Section 4, we obtain the classification of gradings for
Witt and special algebras (using the results on automorphism group schemes).
2. Cartan type Lie algebras
We start by briefly recalling the definitions and relevant properties of Cartan
type Lie algebras. We will use [11] as a standard reference. Fix m ≥ 1 and
n = (n1, . . . , nm) where ni ≥ 1. Set
Z(m;n) := {α ∈ Zm 0 ≤ αi < pni for i = 1, . . . , m}.
The elements of Z(m;n) will be called multi-indices and denoted by Greek letters α,
β, γ. For α = (α1, . . . , αm), we set
α = α1 + · · · + αm.
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
3
We will denote by 1 the multi-index that has 1 in all positions and by εi the multi-
index that has 1 in position i and zeros elsewhere.
Let F be a field of characteristic p > 0.
Definition 2.1. The algebra O = O(m; n) over F is a commutative associative
algebra with a basis {x(α) α ∈ Z(m,n)} where multiplication given by
x(α)x(β) =(cid:18)α + β
α (cid:19)x(α+β) where(cid:18)α + β
α (cid:19) =
αi (cid:19).
mYi=1(cid:18)αi + βi
m) by identifying xi with x(εi):
If n = 1, then O ∼= F[x1, . . . , xm]/(xp
1, . . . , xp
1
x(α) =
α1! · · · αm!
xα1
1 · · · xαm
m .
The algebra O has a canonical Z-grading O =Lℓ≥0 Oℓ defined by declaring the
degree of x(α) to be α. The associated filtration will be denoted by
O(ℓ) :=Mj≥ℓ
Oj.
Note that M := O(1) is the unique maximal ideal of O.
We will now define the following graded Cartan type Lie algebras: Witt, special
and Hamiltonian. The contact algebras and generalized (i.e., non-graded) Cartan
type Lie algebras will not be considered in this paper.
Definition 2.2. Define a linear map ∂i : O → O by ∂ix(α) = x(α−εi) where the
right-hand side is understood to be zero if αi = 0. Then ∂i is a derivation of O.
The Witt algebra W = W (m; n) is the subalgebra of Der(O) that consists of all
operators of the form
f1∂1 + · · · + fm∂m where fi ∈ O.
The canonical Z-grading of O induces a Z-grading on End (O). Since W is a
graded subspace of End (O), it inherits the Z-grading: W =Lℓ≥−1 Wℓ. We will
denote the associated filtration by W(ℓ).
The de Rahm complex Ω0
d−→ . . . is defined as follows: Ω0 = O,
Ω1 = Hom O(W, O), and Ωk = (Ω1)∧k for k ≥ 2. The map d : Ω0 → Ω1 is defined
by (df )(D) = D(f ) for all f ∈ O and D ∈ W . The remaining maps d : Ωk → Ωk+1
are defined in the usual way: d (f dxi1 ∧ · · · ∧ dxik ) = df ∧ dxi1 ∧ · · · ∧ dxik .
d−→ Ω2
d−→ Ω1
Any element D ∈ W acts on Ω1 = Hom O(W, O) by setting
D(ω)(E) = D(ω(E)) − ω([D, E]) for all ω ∈ Ω1 and E ∈ W.
This action turns all Ωk = (Ω1)∧k into W -modules. Of course, all Ωk also have
canonical Z-gradings and associated filtrations.
We will need the following differential forms to define the special and Hamiltonian
algebras:
ωS := dx1 ∧ dx2 ∧ · · · ∧ dxm ∈ Ωm,
ωH := dx1 ∧ dxr+1 + dx2 ∧ dxr+2 + · · · + dxr ∧ dx2r ∈ Ω2
(m ≥ 2);
(m = 2r).
Definition 2.3. The special algebra S = S(m; n) is the stabilizer of ωS in W (m; n):
S = {D ∈ W D(ωS) = 0}.
4
BAHTURIN AND KOCHETOV
The Hamiltonian algebra H = H(m; n) is the stabilizer of ωH in W (m; n):
H = {D ∈ W D(ωH ) = 0}.
Note that in the case m = 2, we have ωS = ωH and hence S = H. It is well-
known that W (m; n) is simple unless p = 2 and m = 1. The algebras S(m; n) and
H(m; n) are not simple, but the first derived algebra S(m; n)(1) (m ≥ 3) and the
second derived algebra H(m; n)(2) are simple.
The Lie algebras W (m; n), S(m; n) and H(m; n) are restrictable -- i.e., admit
p-maps making them restricted Lie algebras -- if and only if n = 1. From now on,
we will assume that this is the case. Since W , S and H have trivial center, their
p-maps are unique. Also, in this case W = Der(O) and hence the p-map is just the
p-th power in the associative algebra End (O). The algebras S and H are restricted
subalgebras of W .
3. Automorphism group schemes
Let O = O(m; 1). Any automorphism µ of the algebra O gives rise to an auto-
morphism Ad (µ) of W given by Ad (µ)(D) = µ ◦ D ◦ µ−1. Then we can define the
action of µ on Ω1 = Hom O(W, O) by setting µ(ω)(D) = µ(ω(Ad (µ−1)(D))) for
all ω ∈ Ω1 and D ∈ W . This action turns all Ωk = (Ω1)∧k into Aut(O)-modules.
Clearly, these actions can still be defined in the same way if we extend the scalars
from the base field F to any commutative associative F-algebra K, i.e., replace O
with O(K) := O ⊗ K, W with W (K) := W ⊗ K and Ωk with Ωk(K) := Ωk ⊗ K.
Recall the automorphism group scheme of a finite-dimensional F-algebra U (see
e.g. [14] for background on affine group schemes). As a functor, the (affine) group
scheme Aut(U ) is defined by setting Aut(U )(K) = AutK (U ⊗K) for any commuta-
tive associative F-algebra K. From the above discussion it follows that we have mor-
phisms of group schemes Ad : Aut(O) → Aut(W ) and also Aut(O) → GL(Ωk).
(We identify a smooth algebraic group scheme such as GL(V ) with the correspond-
ing algebraic group.) Note that, since the p-map of W ⊗ K is uniquely determined,
the automorphism group scheme of W as a Lie algebra is the same as its au-
tomorphism group scheme as a restricted Lie algebra. Note also that the maps
d : Ωk → Ωk+1 are Aut(O)-equivariant.
The algebraic group scheme Aut(U ) contains the algebraic group Aut(U ) as the
largest smooth subgroupscheme. The tangent Lie algebra of Aut(U ) is Der(U ), so
Aut(U ) is smooth if and only if Der(U ) equals the tangent Lie algebra of the group
Aut(U ). The automorphism group schemes of simple Cartan type Lie algebras,
unlike those of the classical simple Lie algebras, are not smooth. Indeed, the tangent
Lie algebra of Aut(O) is W(1), which is a proper subalgebra of W = Der(O), so
Aut(O) is not smooth. In view of the following theorem, we see that Aut(W ) is
not smooth.
Theorem 3.1 ([13]). Let O = O(m; 1) and W = W (m; 1). Assume p > 3. Then
the morphism Ad : Aut(O) → Aut(W ) is an isomorphism of group schemes. (cid:3)
The automorphism group scheme of the general W (m; n) is also known -- see
[15] (p > 2, with small exceptions in the case p = 3) and [9, 10] (any p, with
small exceptions in the cases p = 2 and p = 3). In particular, Theorem 3.1 holds
for p = 3 if m ≥ 2 and for p = 2 if m ≥ 3.
In this section we will establish
analogues of Theorem 3.1 for the simple algebras S(m; 1)(1) and H(m; 1)(2). We
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
5
will follow the approach of [13]. Suppose Φ : G → H is a morphism of algebraic
group schemes. Let Gred and Hred be the largest smooth subgroupschemes, which
will be regarded as algebraic groups.
In order for Φ to be an isomorphism, the
following two conditions are necessary:
A) the restriction Φ : Gred → Hred is a bijection;
B) the tangent map Lie(Φ) : Lie(G) → Lie(H) is a bijection.
However, unless G is known to be smooth (i.e., G = Gred), these two conditions
are not sufficient for Φ to be an isomorphism. In general, one has to show that
the associated map of distribution algebras eΦ : G → H is surjective. (The two
conditions above imply that eΦ : G → H is injective.)
Recall that the distribution algebra G of an algebraic group scheme G is a con-
nected cocommutative Hopf algebra (see e.g. [8] for background on Hopf algebras),
with the space of primitive elements Prim(G) = Lie(G) of finite dimension. Hence
Lie(G) has a descending chain of restricted Lie subalgebras (see e.g. [12] or [4, II,
§3, No. 2]):
Lie(G) = Lie0(G) ⊃ Lie1(G) ⊃ Lie2(G) ⊃ . . . ,
defined by Liek(G) := Liek(G) = Vk(G) ∩ Prim(G), where V : G → G is the Ver-
schiebung operator (see e.g. [12, Theorem 1] or [4, II, §2, No. 7]). The intersection
of this chain is Lie(Gred), which can be identified with the tangent algebra of the
algebraic group Gred.
Recall that a sequence of elements 1 = 0h, 1h, . . . , nh in a connected cocom-
mutative Hopf algebra G is called a sequence of divided powers (lying over 1h) if
ih ⊗ j−ih for all j = 1, . . . , n. It follows that 1h ∈ Prim(G) and
that ε(jh) = 0 for all j = 1, . . . , n. It is easy to see that Vk(pk
h) = 1h for any
pk ≤ n. Hence, if there exists a sequence of divided powers of length pk lying over
h ∈ Prim(G), then h ∈ Liek(G). The converse is also true [12, Theorem 2].
∆(jh) = Pj
i=0
Now we come back to the problem of proving that a morphism Φ : G → H of
algebraic group schemes is an isomorphism. Assuming that Φ satisfies the above
H. Regarding H as a Hopf subalgebra in the distribution algebra of GL(V ) for a
suitable space V , we can apply [4, II, §3, No. 2, Corollary 1] to conclude that
two conditions, we need to show that the Hopf subalgebra eΦ(G) ⊂ H in fact equals
eΦ(G) = H if and only if
C) Lie(Φ) maps Liek(G) onto Liek(H) for all k.
Here we are interested in the case G = Aut(O), where O = O(m; 1), and its
subgroupschemes AutS(O) := StabG(hωSi) and AutH (O) := StabG(hωH i). We
have
Lie(Aut(O)) = Der(O) = W ;
Aut(O)red = Aut(O),
AutS(O)red = StabAut(O)(hωSi),
Lie(AutS(O)) = StabW (hωSi) =: CS;
AutH (O)red = StabAut(O)(hωH i), Lie(AutH (O)) = StabW (hωH i) =: CH.
We will denote StabAut(O)(hωSi) by AutS(O) and StabAut(O)(hωH i) by AutH (O)
for brevity.
Assume p > 3.
It is known that the morphism Ad : Aut(O) → Aut(W )
as well as its restrictions AutS(O) → Aut(S(1)) (m ≥ 3) and AutH (O) →
Aut(H (2)) (m = 2r) induce bijections Aut(O) → Aut(W ), AutS(O) → Aut(S(1))
and AutH (O) → Aut(H (2)) -- see e.g. [11, Theorem 7.3.2]. Also, the tangent map
ad : W → Der(W ) as well as its restrictions CS → Der(S(1)) and CH → Der(H (2))
6
BAHTURIN AND KOCHETOV
are bijective -- see e.g. [11, Theorem 7.1.2]. Thus, the above conditions A) and B)
are satisfied for the morphisms Aut(O) → Aut(W ), AutS(O) → Aut(S(1)) and
AutH (O) → Aut(H (2)).
By [1, Lemma 3.5, 2], Liek(Aut(O)) = StabW (M) = W(0) for all k > 0. Hence,
for k > 0, we have
Liek(AutS(O)) ⊂ Liek(Aut(O)) ∩ CS = CS(0),
Liek(AutH (O)) ⊂ Liek(Aut(O)) ∩ CH = CH(0).
On the other hand, Lie(AutS(O)) = CS(0) and Lie(AutH (O)) = CH(0). It follows
that, in fact,
(1)
Liek(AutS(O)) = CS(0) and Liek(AutH (O)) = CH(0) for all k > 0.
By [1, Lemma 3.5, 4], Liek(Aut(W )) = ad (W(0)) for all k > 0, so condition
C) is satisfied for the morphism Aut(O) → Aut(W ). This is how Theorem 3.1 is
proved in [13]. We are now ready to prove our analogues for S and H.
Theorem 3.2. Let O = O(m; 1), m ≥ 3, and S(1) = S(m; 1)(1). Let
AutS(O) = StabAut(O)(hωSi).
Assume p > 3. Then the morphism Ad : AutS(O) → Aut(S(1)) is an isomorphism
of group schemes.
Proof. By the above discussion, we have to prove that condition C) is satisfied for
Ad : AutS(O) → Aut(S(1)). Let G be the distribution algebra of Aut(S(1)). In
view of (1), it suffices to show that Lie1(G) ⊂ ad (CS(0)). In other words, we have
to verify, for any D ∈ CS, that if D /∈ CS(0), then ad D /∈ Lie1(G). We can write
D = λ1∂1 + · · · + λm∂m + D0 where D0 ∈ CS(0) and the scalars λ1, . . . λm are not
all zero. Now, D0 ∈ Lie1(AutS(O)) implies ad D0 ∈ Lie1(G), so it suffices to prove
that ad (λ1∂1 + · · · + λm∂m) /∈ Lie1(G). Applying an automorphism of O induced
by a suitable linear transformation on the space Span {∂1, . . . , ∂m}, we may assume
without loss of generality that D = ∂1.
By way of contradiction, assume that ad ∂1 ∈ Lie1(G). Then there exists a
sequence of divided powers 1 = 0h, 1h, . . . , ph in G such that 1h = ad ∂1.
As pointed out in the proof of [1, Lemma 3.5, 4], it follows from [12, Lemma 7]
that we may assume without loss of generality that kh = 1
k! (1h)k for k = 0, . . . , p−1.
The distribution algebra G acts canonically on S(1), so we have a homomorphism
η : G → End (S(1)). The restriction of η to Prim(G) = Der(S(1)) is the identity
map. Let kδ = η(kh) for k = 0, . . . , p. Then we have
(2)
kδ =
1
k!
(ad ∂1)k for k = 0, . . . , p − 1,
and, since S(1) is a G-module algebra,
(3)
kδ([X, Y ]) =
kXj=0(cid:2)jδ(X), k−j δ(Y )(cid:3) for all k = 0, . . . , p and X, Y ∈ S(1).
Note that the action of G on S(1) extends canonically to the universal enveloping
algebra U (S(1)) and, since the p-map of S(1) is uniquely determined, the G-action
passes on to the restricted enveloping algebra u(S(1)). By abuse of notation, we
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
7
will use kδ to denote the action of kh on u(S(1)) as well as on S(1). We note for
future reference that
(4)
kδ(XY ) =
(jδ)(X) (k−j δ)(Y ) for all k = 0, . . . , p and X, Y ∈ u(S(1)).
kXj=0
Observe that if we replace pδ by pδ + ξ, where ξ is any derivation of S(1), then
equations (3) and (4) will still hold (with the same 0δ, . . . , p−1δ). We will use this
observation to simplify the operator pδ.
Let zi = 1 + xi, i = 1, . . . , m. For each multi-index α ∈ Z(m;1), we set
zα = zα1
1 · · · zαm
m .
Since zp
i = 1 for all i, we may regard the components of α as elements of the cyclic
group Zp when dealing with zα. It is this property that will make the basis {zα}
of O = O(m; 1) more convenient for us than the standard basis {xα}.
Recall [11, Section 4.2] that S(1) = S(m; 1)(1) is spanned by the elements of the
form
where f ∈ O and 1 ≤ i < j ≤ m. Hence, S(1) is spanned by the elements
Di,j(f ) := ∂j(f )∂i − ∂i(f )∂j
Di,j(zα) = αjzα−εj ∂i − αizα−εi∂j.
For the calculations we are about to carry out, we will need the following:
Lemma 3.3. For any 1 ≤ i < j ≤ m and α, β ∈ Z(m;1), the commutator
is given by the following expressions:
(cid:2)D1,2(zα), Di,j (zβ)(cid:3)
(5)
(6)
(7)
(8)
− (α1β2 − α2β1)Di,j(zα+β−ε1−ε2 ) if 2 < i < j;
− (α1(β2 − 1) − α2β1)D2,j(zα+β−ε1−ε2 ) if 2 = i < j;
−α1βjD1,2(zα+β−ε1−εj ) + α2β1D1,j(zα+β−ε1−ε2 ) − α1β1D2,j(zα+β−2ε1)
if i = 1, j > 2;
− (α1β2 − α2β1)D1,2(zα+β−ε1−ε2 ) if i = 1, j = 2.
Proof. The verification of the above expressions is straightforward. Here we will
verify (7), which is somewhat special, and leave the rest to the reader. We have,
for j > 2,
(cid:2)D1,2(zα), D1,j(zβ)(cid:3) =(cid:2)α2zα−ε2∂1 − α1zα−ε1∂2, βjzβ−εj ∂1 − β1zβ−ε1∂j(cid:3)
= α2βjβ1zα+β−ε1−ε2−εj ∂1 − α2βjα1zα+β−ε1−ε2−εj ∂1
+ α2β1αjzα+β−ε1−ε2−εj ∂1 − α2β1(β1 − 1)zα+β−2ε1−ε2 ∂j
+ α1βj(α1 − 1)zα+β−2ε1−εj ∂2 − α1βjβ2zα+β−ε1−ε2−εj ∂1
+ α1β1β2zα+β−2ε1−ε2 ∂j − α1β1αjzα+β−2ε1−εj ∂2
=(α2βjβ1 − α2βjα1 + α2β1αj − α1βjβ2)zα+β−ε1−ε2−εj ∂1
+ α1(βj(α1 − 1) − β1αj)zα+β−2ε1−εj ∂2
+ β1(−α2(β1 − 1) + α1β2)zα+β−2ε1−ε2 ∂j.
8
BAHTURIN AND KOCHETOV
Comparing the above with
D1,2(zα+β−ε1−εj ) =(α2 + β2)zα+β−ε1−ε2−εj ∂1 − (α1 + β1 − 1)zα+β−2ε1−εj ∂2,
D1,j(zα+β−ε1−ε2 ) =(αj + βj)zα+β−ε1−ε2−εj ∂1 − (α1 + β1 − 1)zα+β−2ε1−ε2 ∂j,
D2,j(zα+β−2ε1) =(αj + βj)zα+β−2ε1−εj ∂2 − (α2 + β2)zα+β−2ε1−ε2 ∂j,
one readily sees that (7) holds.
Another useful fact is the following:
(cid:3)
(9)
[∂ℓ, Di,j(f )] = Di,j(∂ℓ(f )) for all i, j, ℓ = 1, . . . , m and f ∈ O.
The remaining part of the proof of Theorem 3.2 will be divided into four steps.
Step 1: Without loss of generality, we may assume
(10)
pδ (D1,2(z1z2)) = 0.
Substituting α = ε1 + ε2 into expressions (5) through (8), we see that each
nonzero element Di,j(zβ) is an eigenvector for the operator ad D1,2(z1z2), with
eigenvalue λ(i, j, β) = β1 − β2 − 1, β1 − β2 or β1 − β2 + 1, depending on i, j. (For the
case i = 1 and j > 2, one has to combine the first and the third terms in expression
(7), which gives −(β2 + 1)D1,j(zβ).) Since λ(i, j, β) is in the field GFp, we have
D1,2(z1z2) = D1,2(z1z2)p. Applying the operator pδ to both sides and using (4),
we obtain:
pδ (D1,2(z1z2)) = Xi1+···+ip=p
(i1 δ) (D1,2(z1z2)) · · · (ip δ) (D1,2(z1z2))
Taking into account (2) and (9), we see that kδ (D1,2(z1z2)) = 0 for 1 < k < p.
Hence,
(11)
pδ (D1,2(z1z2)) =(cid:0)1δ (D1,2(z1z2))(cid:1)p
+
D1,2(z1z2)k(pδ) (D1,2(z1z2)) D1,2(z1z2)p−k−1.
p−1Xk=0
p−1Xk=0
Since 1δ (D1,2(z1z2)) = [∂1, D1,2(z1z2)] = D1,2(z2) = ∂1 and ∂p
1 = 0, the first term
on the right-hand side of (11) vanishes. The second term can be rewritten using
the identity
X k Y X p−k−1 = (ad X)p−1(Y ),
where X = D1,2(z1z2) and Y = pδ(X). Thus, (11) yields
Y = (ad D1,2(z1z2))p−1 (Y ).
It follows that Y can be written as a linear combination of those Di,j(zβ) for which
the eigenvalue λ(i, j, β) is nonzero:
pδ (D1,2(z1z2)) = Xi,j,β : λ(i,j,β)6=0
σi,j
β Di,j(zβ).
It remains to replace pδ with
pδ +Xi,j,β
σi,j
β
λ(i, j, β)
ad Di,j(zβ),
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
9
to complete Step 1.
Step 2: Without loss of generality, we may assume that, in addition to (10), we
pδ(∂1) = Xi,j,β : β1=p−1
τ i,j
β Di,j(zβ)
also have
(12)
for some scalars τ i,j
β .
pδ(∂1) =
(13)
Now replace pδ with
By (9), we have [∂1, D1,2(z1z2)] = ∂1. Applying the operator pδ to both sides
and using (3), (2) and (10), we obtain:
Hence [D1,2(z1z2), Y ] = −Y where Y = pδ(∂1). It follows that Y can be written as
a linear combination of those Di,j(zβ) for which the eigenvalue λ(i, j, β) is −1:
pXk=0(cid:2)kδ(∂1), p−kδ(D1,2(z1z2))(cid:3) = [pδ(∂1), D1,2(z1z2)] .
pδ(∂1) = Xi,j,β : λ(i,j,β)=−1
pδ + Xi,j,β : β16=p−1
ad Di,j(zβ+ε1).
τ i,j
β Di,j(zβ).
τ i,j
β
β1 + 1
Using (9), one readily sees that, for the new pδ, we have (12), because all terms
with β1 6= p − 1 in the right-hand side of (13) will cancel out. It remains to check
that we still have (10) for the new pδ.
In other words, we have to check that
[D1,2(z1z2), Di,j (zβ+ε1)] = 0 for all i, j, β with τ i,j
6= 0. But this is clear, because
β
λ(i, j, β + ε1) = λ(i, j, β) + 1 and thus we have λ(i, j, β + ε1) = 0 for all i, j, β that
occur in the right-hand side of (13). Step 2 is complete.
Step 3: For any element X = f1∂1 + · · · + fm∂m ∈ W , define
pr1(X) := f1 ∈ O.
Assume (10) and (12). Then, for any k = 1, . . . , p − 1, pr1(cid:0)pδ(D1,2(zk
linear combination of zγ with 0 ≤ γ1 < k.
1 z2))(cid:1) is a
We proceed by induction on k. The basis for k = 1 follows from (10). Now
suppose the claim holds for some k ≥ 1. By (9), we have
z2)] = (k + 1)D1,2(zk
[∂1, D1,2(zk+1
1
1 z2).
Applying the operator pδ to both sides, we obtain
(14)
pδ([∂1, D1,2(zk+1
1
z2)]) = (k + 1)(pδ)(D1,2(zk
1 z2)).
Using (3),(2) and (12), the left-hand side of (14) becomes
Xi,j,β : β1=p−1
Setting Y = pδ(D1,2(zk+1
1
z2)), we can rewrite (14) as follows:
(15) [∂1, Y ] = (k + 1)(pδ)(D1,2(zk
τ i,j
β Di,j(zβ), D1,2(zk+1
1
1
z2) +(cid:2)∂1, pδ(D1,2(zk+1
β (cid:2)D1,2(zk+1
τ i,j
1
z2))(cid:3) .
z2), Di,j(zβ)(cid:3) .
1 z2)) + Xi,j,β : β1=p−1
10
BAHTURIN AND KOCHETOV
Our goal is to show that monomials zγ with γ1 ≥ k + 1 do not occur in f1 :=
pr1(Y ). Since pr1[∂1, Y ] = ∂1f1, it suffices to show that elements zγ∂1 with γ1 ≥ k
do not occur in the right-hand side of (15), when it is regarded as an element of
W . The induction hypothesis tells us that such elements do not occur in the first
term of the right-hand side of (15). We are going to prove the same for the second
term.
In the case 2 < i < j, we have by (5):
Hence, no elements zγ∂1 occur here.
In the case 2 = i < j, we have by (6):
1
(cid:2)D1,2(zk+1
(cid:2)D1,2(zk+1
z2), Di,j (zβ)(cid:3) = (β1 − β2(k + 1))Di,j(zβ+kε1 ).
z2), Di,j(zβ)(cid:3) = (β1 − (β2 − 1)(k + 1))D2,j(zβ+kε1 ).
1
Again, no elements zγ∂1 occur.
In the case i = 1 and 2 < j, we can write(cid:2)D1,2(zk+1
β1 = p − 1, as follows:
1
z2), Di,j(zβ)(cid:3), using (7) and
−(k + 1)βjD1,2(zβ+kε1+ε2−εj ) − D1,j(zβ+kε1 ) + (k + 1)D2,j(zβ+(k−1)ε1+ε2 ).
Elements zγ∂1 occur only in the first two summands, and we have γ = β + kε1 − εj
in either case. Therefore, γ1 = β1 + k = k − 1 mod p (recall that we may take the
exponents of z modulo p).
In the case i = 1,j = 2, we have by β1 = p − 1 and (8):
(cid:2)D1,2(zk+1
1
z2), D1,2(zβ)(cid:3) = −((k + 1)β2 + 1)D1,2(zβ+kε1 ).
Hence, elements zγ∂1 occur with γ = β + kε1 − ε2. Once again, γ1 = k − 1 mod p.
The inductive proof of Step 3 is complete.
Step 4: We can finally obtain a contradiction. By (8), we have
hD1,2(zp−1
1
z2), D1,2(z2
1z2)i = 3D1,2(z2) = 3∂1.
Applying pδ and taking into account (3), (2) and (12), we obtain:
(16)
τ i,j
β Di,j(zβ)
1
3 Xi,j,β : β1=p−1
=hpδ(D1,2(zp−1
+(cid:20)D1,2(cid:18)
+(cid:20)D1,2(cid:18)
1
1
(p − 1)!
(p − 2)!
z2)), D1,2(z2
z2), pδ(D1,2(z2
1
1z2)i +hD1,2(zp−1
z2)(cid:19) , D1,2(cid:18) 1
z2)(cid:19) , D1,2(cid:18) 1
1!
2!
(zp−1
1
(zp−1
1
∂1(z2
1 (z2
∂2
1z2)(cid:19)(cid:21)
1z2)(cid:19)(cid:21) .
∂p−1
1
∂p−2
1
1z2))i
One readily verifies that the sum of the third and fourth terms in the right-hand
side of (16) is 3∂1. Let us consider the first and the second terms. Denote
X := pδ(D1,2(zp−1
1
z2)) and Y := pδ(D1,2(z2
1z2)).
By Step 3, we know that pr1(X) is a linear combination of zγ with 0 ≤ γ1 ≤ p − 2
and pr1(Y ) is a linear combination of zγ with 0 ≤ γ1 ≤ 1.
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
11
1z2) = z2
Since D1,2(z2
on z1 and hence all terms with ∂1 in the commutator [X, D1,2(z2
terms with ∂1 in X. In other words,
1∂1 − 2z1z2∂2, we see that the coefficient of ∂1 depends only
1z2)] come from
pr1[X, D1,2(z2
1z2)] = pr1[pr1(X)∂1, D1,2(z2
1z2)].
Since
[zγ∂1, D1,2(z2
1z2)] = (2 − γ1 + 2γ2)zγ+ε1∂1 − 2zγ+ε2∂2,
we conclude that pr1[X, D1,2(z2
1z2)] is a linear combination of monomials zγ+ε1
with 0 ≤ γ1 ≤ p − 2. Therefore, elements zα∂1 with α1 = 0 do not occur in the
first term in the right-hand side of (16).
∂1 + zp−2
z2), Y ] = pr1[D1,2(zp−1
z2) = zp−1
pr1[D1,2(zp−1
z2∂2, we also see that
Since D1,2(zp−1
z2), pr1(Y )∂1].
1
1
1
1
1
Since
[D1,2(zp−1
1
z2), zγ∂1] = (1 + γ1 + γ2)zγ+(p−2)ε1∂1 + 2zγ+(p−3)ε1+ε2 ∂2,
we conclude that pr1[D1,2(zp−1
z2), Y ] is a linear combination of monomials zγ+(p−2)ε1
with 0 ≤ γ1 ≤ 1. Therefore, elements zα∂1 with α1 = 0 do not occur in the second
term in the right-hand side of (16).
1
Finally, all elements zα∂1 that occur in the left-hand side of (16) have α1 = p−1.
Summarizing our analysis, we obtain
3∂1 = 0,
which is a contradiction, since p > 3. The proof of Theorem 3.2 is complete.
(cid:3)
Corollary 3.4. Under the assumptions of Theorem 3.2,
Liek(Aut(S(1))) = ad (CS(0)) for all k > 0.
Theorem 3.5. Let O = O(m; 1), m = 2r, and H (2) = H(m; 1)(2). Let
AutH (O) = StabAut(O)(hωH i).
Assume p > 3. Then the morphism Ad : AutH (O) → Aut(H (2)) is an isomor-
phism of group schemes.
Proof. Let G be the distribution algebra of Aut(H (2)). As in the proof of Theorem
3.2, we have to show that Lie1(G) ⊂ ad (CH(0)) -- i.e., for any D ∈ CH, if D /∈
CH(0), then ad D /∈ Lie1(G). Again, we can write D = λ1∂1+· · ·+λm∂m+D0 where
D0 ∈ CH(0) and the scalars λ1, . . . λm are not all zero. Since D0 ∈ Lie1(AutH (O))
implies ad D0 ∈ Lie1(G), it suffices to prove that ad (λ1∂1 + · · · + λm∂m) /∈ Lie1(G).
Applying an automorphism of O induced by a suitable symplectic transformation
on the space Span {∂1, . . . , ∂m}, we may assume without loss of generality that
D = ∂1.
By way of contradiction, assume that ad ∂1 ∈ Lie1(G). Then there exists a
sequence of divided powers 1 = 0h, 1h, . . . , ph in G such that 1h = ad ∂1. We may
assume without loss of generality that kh = 1
k! (1h)k for k = 0, . . . , p − 1.
The distribution algebra G acts canonically on H (2), so we have a homomorphism
η : G → End (H (2)). The restriction of η to Prim(G) = Der(H (2)) is the identity
map. Let kδ = η(kh) for k = 0, . . . , p. Then we have
(17)
kδ =
1
k!
(ad ∂1)k for k = 0, . . . , p − 1,
12
BAHTURIN AND KOCHETOV
and, since H (2) is a G-module algebra,
(18)
kδ([X, Y ]) =
kXj=0
[jδ(X), k−j δ(Y )] for all k = 0, . . . , p and X, Y ∈ H (2).
For the extended action of G on the restricted enveloping algebra u(H (2)), we have
(19)
kδ(XY ) =
kXj=0
(jδ)(X) (k−j δ)(Y ) for all k = 0, . . . , p and X, Y ∈ u(H (2)).
We may replace pδ by pδ + ξ, where ξ is any derivation of H (2), without affecting
equations (18) and (19). We will use this observation to simplify the operator pδ.
As in [11, Section 4.2], we define
σ(i) :=(cid:26) 1
−1
if i = 1, . . . , r;
if i = r + 1, . . . , 2r;
and
i′ := i + σ(i)r.
Also we define a map DH : O → H as follows:
DH (f ) :=
2rXi=1
σ(i)∂i(f )∂i′ .
Note that the kernel of DH is F1.
Let zi = 1 + xi, i = 1, . . . , 2r. For each multi-index α ∈ Z(2r;1), we set
zα = zα1
1 · · · zα2r
2r .
We may regard the components of α as elements of the cyclic group Zp when dealing
with zα. For the calculations we are about to carry out, we will need the following
formula:
(20)
(21)
In particular,
[DH (zα), DH (zβ)] = DH(cid:0)DH (zα)(zβ)(cid:1) for all α, β ∈ Z(2r;1).
[∂ℓ, DH(zβ)] = DH(cid:0)∂ℓ(zβ)(cid:1) for all ℓ = 1, . . . , 2r and β ∈ Z(2r;1).
Let τ = (p − 1, . . . , p − 1) ∈ Z(2r;1). It follows from [11, Section 4.2] that
{DH(zα) 0 < α < τ } is a basis of H (2).
The remaining part of the proof of Theorem 3.5 will be divided into four steps,
which are similar to the steps in the proof of Theorem 3.2.
Step 1: Without loss of generality, we may assume
(22)
pδ (DH(z1zr+1)) = 0.
Substituting α = ε1 +ε2 into equation (20), we see that DH (zβ) is an eigenvector
for the operator ad DH (z1zr+1), with eigenvalue βr+1 − β1. Since βr+1 − β1 is in
the field GFp, we have DH (z1zr+1) = DH (z1zr+1)p. Applying the operator pδ to
both sides and using (19), we obtain:
pδ (DH(z1zr+1)) = Xi1+···+ip=p
(i1 δ) (DH (z1zr+1)) · · · (ip δ) (DH (z1zr+1))
p−1Xk=0
p−1Xk=0
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
13
Taking into account (17) and (21), we see that kδ (DH (z1zr+1)) = 0 for 1 < k < p.
Hence,
(23)
pδ (DH (z1zr+1)) =(cid:0)1δ (DH (z1zr+1))(cid:1)p
+
DH (z1zr+1)k(pδ) (DH (z1zr+1)) DH (z1zr+1)p−k−1.
Since 1δ (DH(z1zr+1)) = [∂1, DH(z1zr+1)] = DH(zr+1) = −∂1 and ∂p
1 = 0, the first
term on the right-hand side of (23) vanishes. The second term can be rewritten
using the identity
X k Y X p−k−1 = (ad X)p−1(Y ),
where X = DH(z1zr+1) and Y = pδ(X). Thus, (23) yields
Y = (ad DH (z1zr+1))p−1 (Y ).
It follows that Y is a linear combination of those DH (zβ) for which the eigenvalue
βr+1 − β1 is nonzero:
It remains to replace pδ with
pδ (DH (z1zr+1)) = Xβ : βr+16=β1
σβDH (zβ).
pδ +Xβ
σβ
βr+1 − β1
ad DH (zβ),
Step 2: Without loss of generality, we may assume that, in addition to (22), we
to complete Step 1.
also have
(24)
for some scalars τβ.
pδ(∂1) = Xβ : β1=p−1
τβDH (zβ)
By (21), we have [∂1, DH(z1zr+1)] = −∂1. Applying the operator pδ to both
sides and using (18), (17) and (22), we obtain:
−pδ(∂1) =
pXk=0(cid:2)kδ(∂1), p−kδ(DH (z1zr+1))(cid:3) = [pδ(∂1), DH (z1zr+1)] .
Hence [DH(z1zr+1), Y ] = Y where Y = pδ(∂1). It follows that Y is a linear combi-
nation of those DH(zβ) for which the eigenvalue βr+1 − β1 is 1:
(25)
Now replace pδ with
pδ(∂1) = Xβ : βr+1−β1=1
pδ + Xβ : β16=p−1
β1 + 1
τβ
τβDH (zβ).
ad DH(zβ+ε1).
Using (21), one readily sees that, for the new pδ, we have (24), because all terms
with β1 6= p − 1 in the right-hand side of (25) will cancel out. It remains to check
that we still have (22) for the new pδ.
In other words, we have to check that
[DH(z1zr+1), DH (zβ+ε1)] = 0 for all β with τβ 6= 0. But this is clear, because we
14
BAHTURIN AND KOCHETOV
have βr+1 − (β1 + 1) = 0 for all β that occur in the right-hand side of (25). Step 2
is complete.
Step 3: Assume (22) and (24). Then, for any k = 1, . . . , p − 1, pδ(DH (zk
1 zr+1))
is a linear combination of DH (zγ) with 0 ≤ γ1 < k.
We proceed by induction on k. The basis for k = 1 follows from (22). Now
suppose the claim holds for some k ≥ 1. By (21), we have
[∂1, DH(zk+1
1
zr+1)] = (k + 1)DH(zk
1 zr+1).
Applying the operator pδ to both sides and taking into account (18) and (17), we
obtain
(26) [pδ(∂1), DH (zk+1
1
zr+1)]+[∂1, pδ(DH (zk+1
1
zr+1))] = (k+1)(pδ)(DH (zk
1 zr+1)).
Writing
(27)
pδ(DH (zk+1
1
zr+1)) = X0<γ<τ
σγDH (zγ)
and taking into account (24), (20) and (21), we can rewrite the left-hand side of
(26) as follows:
τβ[DH (zβ), DH (zk+1
1
σγ[∂1, DH (zγ)]
zr+1)] +Xγ
Xβ : β1=p−1
= − Xβ : β1=p−1
= − Xβ : β1=p−1
= − Xβ : β1=p−1
1
τβDH (((k + 1)zk
1 zr+1∂r+1 − zk+1
∂1)(zβ)) +Xγ
τβDH ((k + 1)βr+1zβ+kε1 − β1zβ+kε1) +Xγ
τβ((k + 1)βr+1 + 1)DH (zβ+kε1 ) +Xγ
σγγ1DH (zγ−ε1).
σγDH (∂1(zγ))
σγγ1DH (zγ−ε1)
Setting τ ′
β := τβ((k + 1)βr+1 + 1), we can now rewrite (26) as follows:
(28) Xγ
σγγ1DH (zγ−ε1) = (k + 1)(pδ)(DH (zk
1 zr+1)) + Xβ : β1=p−1
βDH (zβ+kε1).
τ ′
Since the induction hypothesis applies to the first term in the right-hand side of
(28), we see that all elements DH (zα) that occur in the right-hand side have 0 ≤
α1 ≤ k − 1. Comparing this with the left-hand side, we conclude that for any γ
with σγ 6= 0 we have either γ1 = 0 or 0 ≤ γ1 − 1 ≤ k − 1. Hence the elements
DH(zγ) that occur in the right-hand side of (27) have 0 ≤ γ1 ≤ k.
The inductive proof of Step 3 is complete.
Step 4: We can finally obtain a contradiction. By (20), we have
[DH(z2
1zr+1), DH (zp−1
1
zr+1)] = 3DH(zr+1) = −3∂1.
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
15
Applying pδ and taking into account (18) and (17), we obtain:
−3(pδ)(∂1) =[pδ(DH (z2
1zr+1)), DH (zp−1
1
zr+1)] + [DH (z2
1zr+1), pδ(DH (zp−1
1
+(cid:20)DH(cid:18) 1
+(cid:20)DH(cid:18) 1
1!
∂1(z2
∂2
1(z2
1zr+1)(cid:19) , DH(cid:18)
1zr+1)(cid:19) , DH(cid:18)
1
(p − 1)!
1
(p − 2)!
=[pδ(DH (z2
2!
1zr+1)), DH (zp−1
+ 3[DH(z1zr+1), DH (zr+1)]
1zr+1)), DH (zp−1
=[pδ(DH (z2
1
1
∂p−1
1
(zp−1
1
∂p−2
1
(zp−1
1
)zr+1(cid:19)(cid:21)
)zr+1(cid:19)(cid:21)
zr+1)] + [DH (z2
1zr+1), pδ(DH (zp−1
1
zr+1)] + [DH (z2
1zr+1), pδ(DH (zp−1
1
zr+1))]
zr+1))]
zr+1))]
− 3∂1.
So we have
(29)
3∂1 =[pδ(DH (z2
1zr+1)), DH (zp−1
1
+ 3(pδ)(∂1).
By Step 3, we know that
zr+1)] + [DH (z2
1zr+1), pδ(DH (zp−1
1
zr+1))]
pδ(DH (z2
pδ(DH (zp−1
1
1zr+1)) = Xα : 0≤α1≤1
zr+1)) = Xα : 0≤α1≤p−2
σ(2)
α DH(zα);
σ(p−1)
α DH (zα).
Hence, by (20) we obtain:
[pδ(DH (z2
1zr+1)), DH (zp−1
1
zr+1)]
α [DH (zα), DH(zp−1
σ(2)
1
zr+1)]
α DH ((−zp−2
σ(2)
1
zr+1∂r+1 − zp−1
1
∂1)(zα))
α (αr+1 + α1)DH (zα−2ε1 ),
σ(2)
= Xα : 0≤α1≤1
= − Xα : 0≤α1≤1
= Xα : 0≤α1≤1
= Xα : 0≤α1≤p−2
= Xα : 0≤α1≤p−2
= Xα : 0≤α1≤p−2
and, similarly,
[DH (z2
1zr+1), pδ(DH (zp−1
1
zr+1))]
1zr+1), DH (zα)]
σ(p−1)
α
[DH (z2
α DH((2z1zr+1∂r+1 − z2
σ(p−1)
1∂1)(zα))
σ(p−1)
α
(2αr+1 − α1)DH (zα+ε1 ).
σ(2)
α (αr+1 + α1)DH (zα−2ε1)
σ(p−1)
α
(2αr+1 − α1)DH (zα+ε1 )
3∂1 = Xα : 0≤α1≤1
+ Xα : 0≤α1≤p−2
+ 3 Xβ : β1=p−1
τβDH (zβ).
16
BAHTURIN AND KOCHETOV
Using the above calculations and (24), we can rewrite (29) as follows:
This equation is impossible, because p > 3 and none of the sums in the right-hand
side involves DH (zr+1). The proof of Theorem 3.5 is complete.
(cid:3)
Corollary 3.6. Under the assumptions of Theorem 3.5,
Liek(Aut(H (2))) = ad (CH(0)) for all k > 0.
4. Group gradings
Let G be an abelian group. In this section we will give a classification of G-
gradings on the algebra O = O(m; 1) and on the simple Lie algebras W (m; 1),
S(m; 1)(1) (m ≥ 3) and S(2; 1)(2) = H(2; 1)(2).
End (O). It is easy to see that W = Der(O) is a graded subspace, so it inherits a
Given a G-grading ΓO : O = Lg∈G Og, we obtain an induced grading on
G-grading, which will be denoted by ΓW : W =Lg∈G Wg. The spaces Ωk also re-
ceive G-gradings in a natural way, and one can verify that the maps d : Ωk → Ωk+1
respect the G-gradings. (The canonical Z-gradings of W and Ωk are induced by
the canonical Z-grading of O in this manner.) However, S = S(m; 1) (respectively,
H = H(m; 1)) is not a graded subspace of W , in general. It is certainly a graded
subspace if we assume that ωS (respectively, ωH) is a homogeneous element with
respect to the G-grading on Ωm (respectively, Ω2).
Definition 4.1. We will say that a G-grading ΓO : O =Lg∈G Og is S-admissible
(respectively, H-admissible) of degree g0 ∈ G if the form ωS (respectively, ωH ) is a
homogeneous element of degree g0.
If ΓO is S-admissible (respectively, H-admissible), then we will denote the in-
duced G-gradings on S (respectively, H) and its derived algebra(s) by ΓS (respec-
tively, ΓH ).
We will now recall the connection between group gradings on an algebra and
certain subgroupschemes of its automorphism group scheme. Let U be an algebra.
For any group G, a G-grading on U is equivalent to a structure of an FG-comodule
algebra (see e.g.
[8]). Assuming U finite-dimensional and G finitely generated
abelian, we can regard this comodule structure as a morphism of algebraic group
schemes GD → Aut(U ) where GD is the Cartier dual of G, i.e., the group scheme
represented by the commutative Hopf algebra FG. Two G-gradings are isomorphic
if and only if the corresponding morphisms GD → Aut(U ) are conjugate by an
automorphism of U .
If char F = 0, then GD = bG, the algebraic group of multiplicative characters of
G, and Aut(U ) = Aut(U ), the algebraic group of automorphisms. The image of bG
in Aut(U ) is a quasitorus, i.e., a diagonalizable algebraic group. The G-grading on
U is, of course, the eigenspace decomposition of U with respect to this quasitorus.
Hence, group gradings on U correspond to quasitori in Aut(U ).
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
17
Here we are interested in the case char F = p > 0. Then we can write G =
p ,
Gp′ × Gp where Gp′ has no p-torsion and Gp is a p-group. Hence GD =dGp′ × GD
p is finite and connected. The algebraic group dGp′
where dGp′
(which is equal to bG) is a quasitorus, and it acts by automorphisms of U as follows:
is smooth and GD
χ ∗ X = χ(g)X for all X ∈ Ug and g ∈ G.
If Gp is an elementary p-group, then the distribution algebra of GD
p is the restricted
enveloping algebra u(T ) where T is the group of additive characters of Gp, regarded
as an abelian restricted Lie algebra. If {a1, . . . , as} is a basis of Gp (as a vector
space over the field GFp), then the dual basis {t1, . . . , ts} of T has the property
(ti)p = ti for all i. Therefore, T is a torus in the sense of restricted Lie algebras. It
acts by derivations of U as follows:
t ∗ X = t(g)X for all X ∈ Ug and g ∈ G.
If Gp is not elementary, then the distribution algebra of GD
p is not generated by
primitive elements and hence its action on U does not reduce to derivations. Re-
gardless of what the case may be, the image of GD in Aut(U ) is a diagonalizable
subgroupscheme. In some sense, the G-grading on U is its eigenspace decomposition
(see e.g. [14]).
Now, Theorems 3.1 (where, as was pointed out, one can include the cases p = 2
and p = 3), 3.2 and 3.5 give us the following corollary.
Corollary 4.2. Let G be an abelian group. Let L be one of the following simple
Lie algebras: W = W (m; 1) (m ≥ 3 if p = 2 and m ≥ 2 if p = 3), S(1) = S(m; 1)(1)
(m ≥ 3 and p > 3) or H (2) = H(m; 1)(2) (m = 2r and p > 3). Then any G-grading
on L is induced by a G-grading on O = O(m; 1). More precisely,
1) The correspondence ΓO 7→ ΓW is a bijection between the G-gradings on O
and the G-gradings on W . It induces a bijection between the isomorphism
classes of these gradings.
2) The correspondence ΓO 7→ ΓS is a bijection between the S-admissible G-
gradings on O and the G-gradings on S(1). It induces a bijection between
the isomorphism classes of G-gradings on S(1) and the AutS(O)-orbits of
the S-admissible G-gradings on O.
3) The correspondence ΓO 7→ ΓH is a bijection between the H-admissible G-
gradings on O and the G-gradings on H (2). It induces a bijection between
the isomorphism classes of G-gradings on H (2) and the AutH (O)-orbits of
the H-admissible G-gradings on O.
Proof. Let Γ : L = Lg∈G Lg be a G-grading. Replacing G with the subgroup
generated by the support, we may assume that G is finitely generated. We also
obtain that the corresponding morphism GD → Aut(L) is a closed imbedding.
Consider the case L = W . Due to the isomorphism Ad : Aut(O) → Aut(L),
we obtain a closed imbedding GD → Aut(O), which corresponds to a G-grading
ΓO : O = Lg∈G Og (whose support also generates G, since otherwise we would
not have a closed imbedding). The induced G-grading ΓW on W is obtained by
inducing the FG-comodule structure from O to End (O) and then restricting to L,
which agrees with how Ad : Aut(O) → Aut(L) is defined. Therefore, Γ = ΓW .
18
BAHTURIN AND KOCHETOV
In the case L = S(1), we obtain a closed imbedding GD → AutS(O), so the
subspace hωSi of Ωm is GD-invariant, i.e., hωSi is an FG-subcomodule. Hence ωS
is a homogeneous element in the corresponding G-grading ΓO : O =Lg∈G Og.
The proof in the case L = H (2) is similar.
(cid:3)
Remark 4.3. It follows from the proof that the supports of the gradings ΓO, ΓW ,
ΓS and ΓH generate the same subgroup in G.
We will now describe all possible G-gradings on O = O(m; 1).
Proposition 4.4. Let O = O(m; 1) and let M be its unique maximal ideal. Let G
be an abelian group and let O =Lg∈G Og be a G-grading.
1) There exist elements y1, . . . , ym of M and 0 ≤ s ≤ m such that the elements
1 + y1, . . . , 1 + ys, ys+1, . . . , ym are G-homogeneous and {y1, . . . , ym} is a
basis of M modulo M2.
2) Let P = {g ∈ G Og 6⊂ M}. Then P is an elementary p-subgroup of G.
3) Let {b1, . . . , bs} be a basis of P . Then the elements y1, . . . , ym can be chosen
in such a way that the degree of 1 + yi is bi, for all i = 1, . . . , s.
Proof. 1) Pick a basis for O consisiting of G-homogeneous elements and select a
subset {f1, . . . , fm} of this basis that is linearly independent modulo F1⊕M2. Order
the elements fi so that f1, . . . , fs have a nonzero constant term and fs+1, . . . , fm
belong to M. Rescale f1, . . . , fs so that the constant term is 1. Let yi = fi − 1
for i = 1, . . . , s and yi = fi for i = s + 1, . . . , m. Then y1, . . . , ym is a basis of M
modulo M2.
2) Clearly, e ∈ P . If a, b ∈ P , then there exist elements u ∈ Oa and v ∈ Ob that
are not in M. Then the element uv ∈ Oab is not in M, so ab ∈ P . Also, since up is
a nonzero scalar, we have ap = e. It follows that P is an elementary p-subgroup.
3) Any element of O can be uniquely written as a (truncated) polynomial in the
variables 1 + y1, . . . , 1 + ys, ys+1, . . . , ym. Hence, for any g ∈ G,
(30) Og = Span {(1 + y1)j1 · · · (1 + ys)js yjs+1
m = g},
where a1, . . . , am ∈ G are the degrees of 1+y1, . . . , 1+ys, ys+1, . . . , ym, respectively.
It follows that a1, . . . , as generate P . Suppose they do not form a basis of P -- say,
as = aℓ1
m 0 ≤ ji < p, aj1
s+1 · · · yjm
1 · · · ajm
1 · · · aℓs−1
s−1 . Set eyi = yi for i 6= s and
eys := 1 + ys − (1 + y1)ℓ1 · · · (1 + ys−1)ℓs−1 .
eys = ys − (ℓ1y1 + · · · + ℓs−1ys−1)
Then 1 +ey1, . . . , 1 +eys−1,eys, . . . ,eym are homogeneous of degrees a1, . . . , am, respec-
tively. Also, eys ∈ M and
so ey1, . . . ,eym still form a basis of M modulo M2. We have decreased s by 1.
P . Finally, if {b1, . . . , bs} is another basis of P , we can write bj =Qs
Repeating this process as necessary, we may assume that {a1, . . . , as} is a basis of
, where
(ℓij) is a non-degenerate matrix with entries in the field GFp. Set
(mod M2),
i=1 aℓij
i
(1 + yi)ℓij − 1 for j = 1, . . . , s,
sYi=1
eyj :=
and eyj = yj for j = s + 1, . . . , m. Then ey1, . . . ,eym form a basis of M modulo M2,
and 1 +eyj is homogeneous of degree bj, j = 1, . . . , s.
(cid:3)
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
19
Remark 4.5. Without loss of generality, assume that G is generated by the support
of the grading O =Lg∈G Og. Let Q be the image of GD under the corresponding
closed imbedding GD → Aut(O). Let H = StabAut(O)(M). (In fact, H = Aut(O),
regarded as the largest smooth subgroupscheme of Aut(O).) Let Q0 = Q ∩ H.
Then P is the subgroup of G corresponding to the Hopf ideal of FG defining the
subgroupscheme Q0 of Q.
Proof. Let I0 be the Hopf ideal defining the subgroupscheme Q0 and let G0 be the
Og of the
This coarsening corresponds to the subgroupscheme Q0 ⊂ Q. Since Q0 stabilizes
(Og ∩ M). Hence Og ⊂ M for g 6= e
and Oe = F1 ⊕ (Oe ∩ M). Hence Og ⊂ M for all g /∈ G0, which proves P ⊂ G0.
To prove that P = G0, consider the Hopf ideal I of FG corresponding to P . Then
corresponding subgroup of G. Consider the coarsening O = Lg∈G/G0
G-grading induced by the natural homomorphism G → G/G0, i.e., Og =Lg∈g Og.
the subspace M ⊂ O, we have M =Lg∈G/G0
I ⊂ I0. The subgroupscheme eQ of Q defined by I acts trivially on each Og with
g ∈ P . It follows that eQ stabilizes M. Hence eQ ⊂ Q0 and I ⊃ I0.
The description of G-gradings on O(m; 1) resembles the description of G-gradings
on the matrix algebra Mn(F) -- see e.g.
[2, 5]. Namely, Proposition 4.4 shows
that the G-graded algebra O(m; 1) is isomorphic to the tensor product O(s; 1) ⊗
O(m − s; 1) where the first factor has a division grading (in the sense that each
homogeneous component is spanned by an invertible element) and the second factor
has an elementary grading (in the sense that it is induced by a grading of the
underlying vector space M/M2). The isomorphism in question is, of course, the one
defined by y1 7→ x1 ⊗1, . . . , ys 7→ xs ⊗1 and ys+1 7→ 1⊗x1, . . . , ym 7→ 1⊗xm−s. The
first factor, O(s; 1), is isomorphic to the group algebra FP as a G-graded algebra
(where FP has the standard P -grading, which is regarded as a G-grading).
(cid:3)
To state the classification of G-gradings on O up to isomorphism, we introduce
some notation.
Definition 4.6. Let P ⊂ G be an elementary p-subgroup of rank s, 0 ≤ s ≤ m.
Let t = m − s and let
γ = (g1, . . . , gt) ∈ Gt.
We endow the algebra O = O(m; 1) with a G-grading as follows. Select a basis
{b1, . . . , bs} for P and declare the degrees of 1 + x1, . . . , 1 + xs to be b1, . . . , bs,
respectively. Declare the degrees of xs+1, . . . , xm to be g1, . . . , gt, respectively. We
will denote the resulting G-grading on O by ΓO(G, b1, . . . , bs, g1, . . . , gt). Since the
gradings corresponding to different choices of basis for P are isomorphic to each
other, we will also denote this grading (abusing notation) by ΓO(G, P, γ).
Theorem 4.8. Let F be an algebraically closed field of characteristic p > 0. Let G
Definition 4.7. Let γ,eγ ∈ Gt. We will write γ ∼eγ if there exists a permutation
π of the set {1, . . . , t} such that egi ≡ gπ(i) (mod P ) for all i = 1, . . . , t.
be an abelian group. Let O =Lg∈G Og be a grading on the algebra O = O(m; 1)
Two G-gradings, ΓO(G, P, γ) and ΓO(G, eP ,eγ), are isomorphic if and only if P = eP
and γ ∼eγ (Definition 4.7).
over F. Then the grading is isomorphic to some ΓO(G, P, γ) as in Definition 4.6.
20
BAHTURIN AND KOCHETOV
1
(31)
· · · bℓis
defined by
s , then the automorphism of O defined by
Proof. Let y1, . . . , ym be as in Proposition 4.4. Let g1, . . . , gt ∈ G be the degrees
of ys+1, . . . , ym, respectively. Then the automorphism of O defined by yi 7→ xi,
xi 7→ xi for i = 1, . . . , s and xs+i 7→ xs+π(i) for i = 1, . . . , t,
i = 1, . . . , m, sends the grading O =Lg∈G Og to ΓO(G, b1, . . . , bs, g1, . . . , gt).
If egi = gπ(i), i = 1, . . . , t, for some permutation π, then the automorphism of O
sends ΓO(G, P,eγ) to ΓO(G, P, γ).
Ifegi = gibℓi1
sends ΓO(G, P,eγ) to ΓO(G, P, γ).
Hence, if γ ∼eγ as in Definition 4.7, then ΓO(G, P,eγ) is isomorphic to ΓO(G, P, γ).
invariants of the G-graded algebra O = Lg∈G Og. This is obvious for P , since
grading, O =Lg∈G Og, induced by the natural homomorphism G → G. It follows
from the definition of P that M is a G-graded subspace of O. Consequently, M2 is
also a G-graded subspace, and the quotient V := M/M2 inherits a G-grading:
P = {g ∈ G Og 6⊂ M}. Let G = G/P and consider the coarsening of the G-
It remains to show that the subgroup P and the equivalence class of γ are
xi 7→ xi for i = 1, . . . , s and xs+i 7→ xs+i
(1 + xj)ℓij for i = 1, . . . , t,
sYj=1
(32)
V = Va1 ⊕ · · · ⊕ Vaℓ.
and k1, . . . , kℓ are invariants of the G-graded algebra O = Lg∈G Og.
Let ki = dim Vai if ai 6= e and ki = dim Vai − s if ai = e. Clearly, a1, . . . , aℓ
If the G-
grading on O is ΓO(G, P, γ), then, up to a permutation, g1P = . . . = gk1 P = a1,
gk1+1P = . . . = gk1+k2 P = a2, and so on.
(cid:3)
Remark 4.9. Instead of using γ = (g1, . . . , gt) where some of the cosets giP may
be equal to each other, one can take multiplicities,
κ = (k1, . . . , kℓ) where ki are positive integers,
with κ := k1 + · · · + kℓ = t,
γ = (g1, . . . , gℓ) where gi ∈ G are such that g−1
i gj /∈ P for all i 6= j,
and write
ΓO(G, P, κ, γ) = ΓO(G, P, g1, . . . , g1
, . . . , gℓ, . . . , gℓ
same number of components ℓ and there exists a permutation π of the set {1, . . . , ℓ}
{z
Then ΓO(G, P, κ, γ) is isomorphic to ΓO(G, eP ,eκ,eγ) if and only if κ andeκ have the
such thateki = kπ(i) and egi ≡ gπ(i) (mod P ) for all i = 1, . . . , ℓ.
Definition 4.10. Fix 0 ≤ s ≤ m. For a multi-index α ∈ Zm, let
k1 times
kℓ times
{z
}
}
).
α := (α1 + pZ, . . . , αs + pZ, αs+1, . . . , αm) ∈ Zs
p × Zm−s.
Define a Zs
p × Zm−s-grading on O = O(m; 1) by declaring the degree of 1 + xi,
i = 1, . . . , s, and the degree of xi, i = s + 1, . . . , m, to be εi. This is the
grading ΓO(G, P, γ) where G = Zs
p, and
γ = (εs+1, . . . εm). We will denote this grading by ΓO(s).
p × Zm−s (written additively), P = Zs
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
21
Corollary 4.11. Let O = O(m; 1). Then, up to equivalence, there are exactly
m + 1 fine gradings of O. They are ΓO(s), s = 0, . . . , m. The universal group of
ΓO(s) is Zs
p × Zm−s.
Proof. All homogeneous components of ΓO(s) are 1-dimensional, so it is a fine
grading. All relations in the grading group Zs
p × Zm−s come from the fact that
0 6= (Og)p ⊂ Oe for certain elements g. Hence Zs
p × Zm−s is the universal group of
ΓO(s).
For any abelian group G and a p-subgroup P ⊂ G with a basis {b1, . . . , bs}, any
p × Zm−s-grading
G-grading ΓO(G, b1, . . . , bs, g1, . . . , gm−s) is induced from the Zs
ΓO(s) by the homomorphism Zs
p × Zm−s → G defined by
εi 7→ bi for i = 1, . . . , s, and εi 7→ gi−s for i = s + 1, . . . , m.
It follows that, up to equivalence, there are no other fine gradings. The grad-
ings ΓO(s) are pair-wise non-equivalent, because their universal groups are non-
isomorphic.
(cid:3)
Definition 4.12. The G-grading induced by ΓO(G, b1, . . . , bs, g1, . . . , gt) (Defini-
tion 4.6) on the Lie algebra W will be denoted by ΓW (G, b1, . . . , bs, g1, . . . , gt) or
ΓW (G, P, γ). Explicitly, we declare the degree of each element
(1 + x1)α1 · · · (1 + xs)αs xαs+1
s+1 · · · xαm
m ∂i where α ∈ Z(m;1), 1 ≤ i ≤ m,
to be
bα1−δi,1
1
· · · bαs−δi,s
s
gαs+1−δi,s+1
1
· · · gαm−δi,m
t
,
where δi,j is the Kronecker delta.
(Definition 4.10) will be denoted by ΓW (s).
In particular, the gradings induced by ΓO(s)
The following is a generalization of a result in [3] (see also [11, Corollary 7.5.2])
on maximal tori of the restricted Lie algebra W , which corresponds to the case
when G is an elementary p-group.
Theorem 4.13. Let F be an algebraically closed field of characteristic p > 0. Let G
be an abelian group. Let W = W (m; 1) over F. Assume m ≥ 3 if p = 2 and m ≥ 2
if p = 3. Then any grading W =Lg∈G Wg is isomorphic to some ΓW (G, P, γ) as
in Definition 4.12. Two G-gradings, ΓW (G, P, γ) and ΓW (G, eP ,eγ), are isomorphic
if and only if P = eP and γ ∼eγ (Definition 4.7).
Proof. A combination of Theorem 4.8 and Corollary 4.2.
(cid:3)
Corollary 4.14. Let W = W (m; 1). Assume m ≥ 3 if p = 2 and m ≥ 2 if p = 3.
Then, up to equivalence, there are exactly m + 1 fine gradings of W . They are
ΓW (s), s = 0, . . . , m. The universal group of ΓW (s) is Zs
(cid:3)
p × Zm−s.
We now turn to special algebras.
Proposition 4.15. In the notation of Proposition 4.4, assume that O =Lg∈G Og
is an S-admissible G-grading of degree g0. Then the elements y1, . . . , ym can be cho-
sen in such a way that the degrees a1, . . . , am ∈ G of 1 + y1, . . . , 1 + ys, ys+1, . . . , ym
(respectively) satisfy the equation g0 = a1 · · · am.
22
BAHTURIN AND KOCHETOV
Proof. Choose elements y1, . . . , ym as in Proposition 4.4. Let a1, . . . , am ∈ G be
the degrees of the elements 1 + y1, . . . , 1 + ys, ys+1, . . . , ym, respectively. We are
going to adjust y1, . . . , ym to make a1, . . . , am satisfy the above equation.
The form dy1 ∧ · · · ∧ dym is G-homogeneous of degree a0 := a1 · · · am. On the
other hand, we have
dy1 ∧ · · · ∧ dym = f ωS where f = det(∂jyi).
Since ωS is G-homogeneous of degree g0, we conclude that f is G-homogeneous of
degree a0g−1
0 . Since f /∈ M, we have a0g−1
0 ∈ P .
First consider the case s = m. Then a0 ∈ P and thus g0 ∈ P . Also, the G-
grading in this case is the eigenspace decomposition of O with respect to a torus
T ⊂ Der(O) = W , where T is isomorphic to the group of additive characters of P ,
so T has rank s = m. If g0 = e, then ωS is T -invariant, so T ⊂ StabW (ωS) = S,
which is a contradiction, because the toral rank of S = S(m; 1) is less than m (in
fact, it is m − 1). Therefore, in this case we necessarily have g0 6= e. It follows
that there exists a basis {b1, . . . , bm} of P such that g0 = b1 · · · bm. By Proposition
· · · a−ℓs
s
1
1 · · · aℓs
0 = aℓ1
are as desired.
Now assume that s < m. Write a0g−1
degree bi, i = 1, . . . , m. The proof in this case is complete.
4.4, we can replace y1, . . . , ym with ey1, . . . ,eym so that 1 +eyi is G-homogeneous of
s . Set eyi = yi for i < m and
Then eym is G-homogeneous of degreeeam = ama−ℓ1
and hence ey1, . . . ,eym
eym = ym(1 + y1)−ℓ1 · · · (1 + ys)−ℓs.
Recall that in Definition 4.6 of ΓO(G, P, γ), we had to choose a basis {b1, . . . , bs}
for P . The isomorphism class, i.e., the Aut(O)-orbit, of the grading does not depend
on this choice. Clearly, the grading is S-admissible of degree g0 = b1 · · · bsg1 . . . gt
and hence it induces a G-grading on the Lie algebra S and its derived subalgebras.
Let L = S(m; 1)(1) if m ≥ 3 and L = S(m; 1)(2) if m = 2. Since g0 is AutS(O)-
invariant, the induced gradings on L corresponding to different values of g0 are
not isomorphic. Conversely, suppose {eb1, . . . ,ebs} is another basis of P such that
eb1 · · ·ebs = b1 · · · bs (i.e., this basis leads to the same value of g0). Write ebj =
Qs
i where (αij ) is a non-degenerate matrix with entries in the field GFp. Set
i=1 bαij
(cid:3)
(33)
(1 + xi)αij − 1 for j = 1, . . . , s,
sYi=1
exj :=
(34)
andexj = xj for j = s + 1, . . . , m. Thenex1, . . . ,exm form a basis of M modulo M2,
and 1 +exj is homogeneous of degreeebj, j = 1, . . . , s. One readily computes that
sYi=1
det(∂jexi) = det(αij)
Noweb1 · · ·ebs = b1 · · · bs means that Ps
j=1 αij = 1 for all i, so det(∂jexj) is in F.
Therefore, the automorphism of O defined by xi 7→ exi, i = 1, . . . , m, belongs to
the subgroup AutS(O). We have proved that two G-gradings on L arising from the
same data P and γ, but different choices of basis for P , are isomorphic if and only
if they have the same value of g0. This justifies the following:
(1 + xi)−1+Ps
j=1 αij .
Definition 4.16. Let P and γ be as in Definition 4.6. Let g0 ∈ G be such that
g0g−1
1
· · · g−1
t ∈ P \ {e}.
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
23
Select a basis {b1, . . . , bs} for P such that g0 = b1 · · · bsg1 · · · gt. The G-grading in-
duced by ΓO(G, b1, . . . , bt, g1, . . . , gs) on the Lie algebra S and its derived subalge-
bras will be denoted by ΓS(G, b1, . . . , bt, g1, . . . , gs) or ΓS(G, P, γ, g0). In particular,
the Zs
p × Zm−s-grading induced by ΓO(s) (Definition 4.10, with {ε1, . . . , εs} as a
basis for Zs
p), will be denoted by ΓS(s).
The following is a generalization of a result in [3] (see also [11, Theorem 7.5.5])
on maximal tori of the restricted Lie algebra CS, which corresponds to the case
when G is an elementary p-group.
Theorem 4.17. Let F be an algebraically closed field of characteristic p > 3. Let G
be an abelian group. Let L = S(m; 1)(1) if m ≥ 3 and L = S(m; 1)(2) = H(m; 1)(2) if
to some ΓS(G, P, γ, g0) as in Definition 4.16. Two G-gradings, ΓS(G, P, γ, g0) and
m = 2 (a simple Lie algebra over F). Then any grading L =Lg∈G Lg is isomorphic
ΓS(G,eP ,eγ,eg0), are isomorphic if and only if P = eP , γ ∼eγ (Definition 4.7) and
g0 =eg0.
Proof. First we show that the grading L =Lg∈G Lg is isomorphic to some grading
O. Let Γ′ : O =Lg∈G O′
that induces the grading L =Lg∈G Lg. As usual, let P = {g ∈ G O′
ΓS(G, P, γ, g0). We can apply Corollary 4.2 to translate this problem to the algebra
g be the S-admissible grading on O, of some degree g0 ∈ G,
g 6⊂ M} and
let s be the rank of P . By Proposition 4.15, there exist elements y1, . . . , ym ∈ M
that form a basis of M mod M2 and such that 1 + yi, i ≤ s, and yi, i > s, are
G-homogeneous of some degrees ai, i = 1, . . . , m, where {a1, . . . , as} is a basis
of P and g0 = a1 · · · am. We want to show that there exists an automorphism
in AutS(O) that sends Γ′ to the grading ΓO = ΓO(G, a1, . . . , as, as+1, . . . , am).
Denote the latter grading by O = Lg∈G Og. Let µ be the automorphism of O
defined by yi 7→ xi, i = 1, . . . , m. Then µ sends Γ′ to ΓO, but µ may not belong to
AutS(O). Write µ(ωS) = f ωS for some f ∈ O. Now µ(ωS) has degree g0 relative
to the grading induced on Ωm by ΓO, ωS has degree a1 · · · am relative to the said
grading, and g0 = a1 · · · am, so we conclude that f has degree e relative to ΓO. If
s = m, this implies that f is in F and hence µ ∈ AutS(O), completing the proof.
So we assume s < m.
Now we follow the idea of the proof of [11, Proposition 7.5.4], which is due to
[7]. Observe that
µ(ωS) =µ (d(x1dx2 ∧ · · · ∧ dxm))
=d (µ(x1)dµ(x2) ∧ · · · ∧ dµ(xm))
=d mXi=1
= mXi=1
(−1)i−1hidx1 ∧ · · · ∧ dxi−1 ∧ dxi+1 ∧ · · · ∧ dxm!
∂ihi! ωS,
i=1 hi∂i ∈ W . Since µ(ωS) = f ωS, we have
i=1 ∂ihi = f . One can immediately verify that div(Wg) ⊂ Og for all
where h1, . . . , hm ∈ O. Set E := Pm
div(E) =Pm
g ∈ G, where ΓW : W = Lg∈G Wg is the grading induced on W by ΓO. (Also,
this is a consequence of the fact that div : W → O is Aut(O)-equivariant.) Since
f ∈ Oe, replacing E with its G-homogeneous component of degree e will not affect
the equation div(E) = f , so we will assume that E ∈ We.
24
BAHTURIN AND KOCHETOV
Define a Z-grading on O by declaring the degree of x1, . . . , xs (or, equivalently,
1 + x1, . . . , 1 + xs) to be 0 and the degree of xs+1, . . . , xm to be 1. This Z-grading is
compatible with the G-grading ΓO in the sense that the homogeneous components of
one grading are graded subspaces of O relative to the other grading. We will denote
the filtration associated to this Z-grading by O{ℓ}, ℓ = 0, 1, 2, . . ., to distinguish it
from the filtration O(ℓ) associated to the canonical Z-grading.
Write f =Pk≥0 fk, where fk has degree k in the Z-grading and degree e in the
G-grading. Observe that the constant term of f is equal to the constant term of
f0, so f0 is an invertible element of O. Let τ1 be the automorphism of O defined
by
τ1(xi) = xi for i < m and τ1(xm) = f −1
0 xm.
Since f0 has degree e in the G-grading, τ1 preserves ΓO, i.e., τ1(Og) = Og for all
g ∈ G. We also have τ1(O{ℓ}) = O{ℓ} for all ℓ. Since xm has degree 1 in the
Z-grading, it does not occur in f0. Hence τ1(f0) = f0 and we can compute:
(τ1 ◦ µ)(ωS) =τ1(f ωS) = τ1(f )τ1(ωS)
whereeh =Pk≥1 fk and h = f −1
the G-grading ΓO and has the following property:
0 ωS = (1 + h)ωS,
=(f0 + τ1(eh))f −1
0 τ1(eh). Note that h ∈ O{1}.
Claim: For any ℓ = 1, 2, . . . there exists an automorphism τℓ of O that preserves
(35)
(τℓ ◦ µ)(ωS) = (1 + h)ωS where h ∈ O{ℓ}.
We proceed by induction on ℓ. The basis for ℓ = 1 was proved above. Assume
(35) holds for some ℓ ≥ 1 and τℓ. Since τℓ preserves ΓO, we have 1 + h ∈ Oe and
defined by
induced from our Z-grading of O and degree e in the G-grading. Since div preserves
degree e in the G-grading. As was shown above, there exists E ∈ We such that
eτ (xi) = xi − Eℓ(xi), i = 1, . . . , m.
eτ (ωS) = (1 − div(Eℓ) + ef )ωS = (1 − hℓ + ef )ωS
hence h ∈ Oe. Write h = Pk≥ℓ hk where hk has degree k in the Z-grading and
div(E) = 1 + h. Write E = Pk≥−1 Ek where Ek has degree k in the Z-grading
the Z-grading, we have divEk = hk for k ≥ 1. Let eτ be the automorphism of O
Since Eℓ ∈ We, the automorphism eτ preserves the G-grading ΓO. We also have
eτ (f ) = f (mod O{k+1}) for all f ∈ O{k} and
for some ef ∈ O{2ℓ}. Hence we can compute:
=(1 + hℓ + bf )(1 − hℓ + ef )ωS = (1 +eh)ωS,
where bf ∈ O{ℓ+1} and eh = −h2
ℓ + bf (1 − hℓ) + ef (1 + hℓ + bf ) ∈ O{ℓ+1}. Setting
τℓ+1 =eτ ◦ τℓ, we complete the induction step.
Seteµ = τℓ ◦ µ for ℓ = (p − 1)(m − s) + 1. Theneµ sends Γ′ to ΓO and belongs to
AutS(O), since eµ(ωS) = ωS. We have proved the first assertion of the theorem.
that, if γ ∼ eγ and b1 · · · bsg1 · · · gt = g0 = b1 · · · bsg1 · · · gt where {b1, . . . , bs} and
{eb1, . . . ,ebs} are bases of P as in Definition 4.16, then ΓO(G, b1, . . . , bs, g1, . . . , gt)
Now, the subgroup P and the equivalence class of γ = (g1, . . . , gt) are invari-
ants of the G-graded algebra O, and g0 is AutS(O)-invariant. It remains to show
(eτ ◦ τℓ ◦ µ)(ωS) =eτ ((1 + h)ωS) =eτ (1 + h)eτ (ωS)
j=1 αij!
j=1 αij +Pt
sYj=1
tYi=1
j=1 ℓji! ωS.
(1 + xj )ℓij ωS
j=1 αij +Pt
j=1 ℓji
Ps
i
b
,
On the other hand,
(1 + xi)−1+Ps
(1 + xi)−1+Ps
µ(ωS) = det(αij ) sYi=1
= det(αij ) sYi=1
eb1 · · ·ebseg1 · · ·egt = g1 · · · gt
tXj=1
sXj=1
sYi=1
GRADINGS ON RESTRICTED CARTAN TYPE LIE ALGEBRAS
25
and ΓO(G, b1, . . . , bs, g1, . . . , gt) are in the same AutS(O)-orbit. Clearly, the au-
tomorphism (31) of O, determined by a permutation π of {1, . . . , t}, belongs to
i=1 bαij
AutS(O). So it suffices to consider the caseegi ≡ gi (mod P ). Writeebj =Qs
where (αij ) is a non-degenerate matrix with entries in the field GFp. Also writeegi =
giQs
xj 7→exj , j ≤ s, and xj 7→ xj , j > s, whereexj are as in (33), and the automorphism
defined by (32), sends ΓO(G,eb1, . . . ,ebs,eg1, . . . ,egt) to ΓO(G, b1, . . . , bs, g1, . . . gt).
, i = 1, . . . , t. Then the composition µ of the automorphism defined by
Now, (34) implies that
j=1 bℓij
j
i
so the equalityeb1 · · ·ebseg1 · · ·egt = g0 = b1 · · · bsg1 · · · gt implies that
ℓji = 1 for all i
αij +
and hence µ ∈ AutS(O).
(cid:3)
Corollary 4.18. Under the assumptions of Theorem 4.17, there are, up to equiva-
lence, exactly m+1 fine gradings of L. They are ΓS(s), s = 0, . . . , m. The universal
group of ΓS(s) is Zs
(cid:3)
p × Zm−s.
References
[1] Allen, H. P.; Sweedler, M. E. A theory of linear descent based upon Hopf algebraic techniques.
J. Algebra, 12 (1969), 242 -- 294.
[2] Bahturin, Y.; Kochetov, M. Classification of group gradings on simple Lie algebras of types
A, B, C and D. arXiv:0908.0906v1 [math.RA]
[3] Demuskin, S. P. Cartan subalgebras of the simple Lie p-algebras Wn and Sn. Sibirsk. Mat.
Z., 11 (1970), 310 -- 325 [Russian]; Siberian Math. J., 11 (1970), 233 -- 245 [English translation].
[4] Dieudonn´e, J. Introduction to the theory of formal groups. Pure and Applied Mathematics,
20, Marcel Dekker, Inc., New York, 1973.
[5] Elduque, A. Fine gradings on simple classical Lie algebras. arXiv:0906.0655v1 [math.RA]
[6] Kochetov, M. Gradings on finite-dimensional simple Lie algebras. Acta Appl. Math. 108
(2009), 101 -- 127.
[7] Kuznetsov, M.I.; Yakovlev, V.A. Elementary proof of Demuskin's theorem on tori in special
Lie p-algebras of Cartan type. Comm. Algebra 25 (1997), 3979 -- 3983.
[8] Montgomery, S. Hopf Algebras and their Actions on Rings. CBMS Regional Conference
Series in Mathematics, 82, American Math. Soc., Providence, RI, 1993.
[9] Skryabin, S. Modular Lie algebras of Cartan type over algebraically non-closed fields, I and
II. Comm. Algebra, 19 (1991), no. 6, 1629 -- 1741 and Comm. Algebra 23 (1995), no. 4,
1403 -- 1453.
[10] Skryabin, S. On the automorphism group schemes of Lie algebras of Witt type. Comm.
Algebra, 29 (2001), no. 9, 4047 -- 4077.
[11] Strade, H. Simple Lie algebras over fields of positive characteristic. I. Structure theory. De
Gruyter Expositions in Mathematics, 38, Walter de Gruyter & Co., Berlin, 2004.
26
BAHTURIN AND KOCHETOV
[12] Sweedler, M. E. Hopf algebras with one grouplike element. Trans. Amer. Math. Soc., 127
(1967), 515 -- 526.
[13] Waterhouse, W. C. Automorphism schemes and forms of Witt Lie algebras. J. Algebra, 17
(1971), 34 -- 40.
[14] Waterhouse, W. C. Introduction to affine group schemes. Graduate Texts in Mathematics,
66, Springer-Verlag, New York-Berlin, 1979.
[15] Waterhouse, W. C. Automorphisms and twisted forms of generalized Witt Lie algebras.
Trans. Amer. Math. Soc., 327 (1991), no. 1, 185 -- 200.
Department of Mathematics and Statistics, Memorial University of Newfoundland,
St. John's, NL, A1C5S7, Canada
E-mail address: [email protected]
Department of Mathematics and Statistics, Memorial University of Newfoundland,
St. John's, NL, A1C5S7, Canada
E-mail address: [email protected]
|
1906.06213 | 1 | 1906 | 2019-05-26T23:15:07 | Solvable extensions of the naturally graded quasi-filiform Leibniz algebra of second type $\mathcal{L}^4$ | [
"math.RA"
] | For a sequence of the naturally graded quasi-filiform Leibniz algebra of second type $\mathcal{L}^4$ introduced by Camacho, G\'{o}mez, Gonz\'{a}lez and Omirov, all the possible right solvable indecomposable extensions over the field $\Bbb C$ are constructed. | math.RA | math | SOLVABLE EXTENSIONS OF THE NATURALLY GRADED
QUASI-FILIFORM LEIBNIZ ALGEBRA OF SECOND TYPE L 4
Department of Mathematics, Adrian College, 110 S Madison St., Adrian, MI 49221,
A. Shabanskaya
USA
[email protected]
Abstract. For a sequence of the naturally graded quasi-filiform Leibniz algebra of
second type L 4 introduced by Camacho, Gómez, González and Omirov, all the possible
right solvable indecomposable extensions over the field C are constructed.
AMS Subject Classification: 17A30, 17A32, 17A36, 17A60, 17B30
Keywords: Leibniz algebra, solvability, nilpotency, nilradical, nil-independence, deriva-
tion.
9
1
0
2
y
a
M
6
2
]
.
A
R
h
t
a
m
[
1
v
3
1
2
6
0
.
6
0
9
1
:
v
i
X
r
a
1
2
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
1. Introduction
Leibniz algebras were discovered by Bloch in 1965 [4] who called them D− algebras.
Later on they were considered by Loday and Cuvier [10, 15, 16, 17] as a non-antisymmetric
analogue of Lie algebras. It makes every Lie algebra be a Leibniz algebra, but the converse
is not true. Exactly Loday named them Leibniz algebras after Gottfried Wilhelm Leibniz.
Since then many analogs of important theorems in Lie theory were found to be true
for Leibniz algebras, such as the analogue of Levi's theorem which was proved by Barnes
[3]. He showed that any finite-dimensional complex Leibniz algebra is decomposed into a
semidirect sum of the solvable radical and a semisimple Lie algebra. Therefore the biggest
challenge in the classification problem of finite-dimensional complex Leibniz algebras is
to study the solvable part. And to classify solvable Leibniz algebras, we need nilpotent
Leibniz algebras as their nilradicals, same as in the case of Lie algebras [18].
Every Leibniz algebra satisfies a generalized version of the Jacobi identity called the
Leibniz identity. There are two Leibniz identities: the left and the right. We call Leibniz
algebras right Leibniz algebras if they satisfy the right Leibniz identity and left, if they
satisfy the left. A left Leibniz algebra is not necessarily a right Leibniz algebra [11].
Leibniz algebras inherit an important property of Lie algebras which is that the right
(left) multiplication operator of a right (left) Leibniz algebra is a derivation [9]. Besides
the algebra of right (left) multiplication operators is endowed with a structure of a Lie
algebra by means of the commutator [9]. Also the quotient algebra by the two-sided ideal
generated by the square elements of a Leibniz algebra is a Lie algebra [19], where such
ideal is the minimal, abelian and in the case of non-Lie Leibniz algebras it is always non
trivial.
It is possible to find solvable Leibniz algebras in any finite dimension working with the
sequence of nilpotent Leibniz algebras in any finite dimension and their "nil-independent"
derivations. This method for Lie algebras is based on what was shown by Mubarakzyanov
in [18]: the dimension of the complimentary vector space to the nilradical does not exceed
the number of nil-independent derivations of the nilradical. This result was extended to
Leibniz algebras by Casas, Ladra, Omirov and Karimjanov [8] with the help of [2]. Be-
sides, similarly to the case of Lie algebras, for a solvable Leibniz algebra L we also have
the inequality dim nil(L) ≥ 1
2 dim L [18]. There is the following work performed over
the field of characteristic zero using this method: Casas, Ladra, Omirov and Karimjanov
classified solvable Leibniz algebras with null-filiform nilradical [8]; Omirov and his col-
leagues Casas, Khudoyberdiyev, Ladra, Karimjanov, Camacho and Masutova classified
solvable Leibniz algebras whose nilradicals are a direct sum of null-filiform algebras [13],
naturally graded filiform [9, 14], triangular [12] and finally filiform [7]. Bosko-Dunbar,
Dunbar, Hird and Stagg attempted to classify left solvable Leibniz algebras with Heisen-
berg nilradical [5]. Left and right solvable extensions of R18 [1], L 1, L 2 and L 3 [6] over
the field of real numbers were found by Shabanskaya in [22, 23, 24].
The starting point of the present article is a naturally graded quasi-filiform non Lie
Leibniz algebra of the second type L 4, (n ≥ 4) in the notation of [6]. This algebra is left
and right at the same time and an associative when n = 4.
Naturally graded quasi-filiform Leibniz algebras in any finite dimension over C were
studied by Camacho, Gómez, González, Omirov [6]. They found six such algebras of the
first type, where two of them depend on a parameter and eight algebras of the second
type with one of them depending on a parameter.
This paper continues a work on finding all solvable extensions of quasi-filiform Leib-
niz algebras over the field of complex numbers. For a sequence L 4 such extensions of
codimension at most two are possible.
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
3
The paper is organized as follows: in Section 2 we give some basic definitions, in Section
3 we show what involves constructing solvable Leibniz algebras with a given nilradical.
In Section 4 we describe the nilpotent sequence L 4 and give the summary of the results
stated in theorems, which could be found in the remaining Section 5.
As regards notation, we use he1, e2, ..., eri to denote the r-dimensional subspace gener-
ated by e1, e2, ..., er, where r ∈ N. Besides g and l are used to denote solvable right and
left Leibniz algebras, respectively.
Throughout the paper all the algebras are finite dimensional over the field of complex
numbers and if the bracket is not given, then it is assumed to be zero, except the brackets
for the nilradical, which most of the time are not given (see Remark 5.1) to save space.
In the tables an ordered triple is a shorthand notation for a derivation property of the
multiplication operators, which is either Rz ([x, y]) = [Rz(x), y]+[x, Rz(y)] or Lz ([x, y]) =
[Lz(x), y] + [x, Lz(y)]. We also assign Ren+1 := R and Len+1 := L.
We use Maple software to compute the Leibniz identity, the "absorption" (see [20, 21]
and Section 3.2), the change of basis for solvable Leibniz algebras in some particular
dimensions, which are generalized and proved in an arbitrary finite dimension.
2. Preliminaries
We give some Basic definitions encountered working with Leibniz algebras.
Definition 2.1.
1. A vector space L over a field F with a bilinear operation
[−,−] : L → L is called a Leibniz algebra if for any x, y, z ∈ L the Leibniz
identity
[[x, y], z] = [[x, z], y] + [x, [y, z]]
holds. This Leibniz identity is known as the right Leibniz identity and we call L
in this case a right Leibniz algebra.
2. There exists the version corresponding to the left Leibniz identity
[[x, y], z] = [x, [y, z]] − [y, [x, z]],
and a Leibniz algebra L is called a left Leibniz algebra.
Remark 2.1. In addition, if L satisfies [x, x] = 0 for every x ∈ L, then it is a Lie
algebra. Therefore every Lie algebra is a Leibniz algebra, but the converse is not true.
Definition 2.2. The two-sided ideal C(L) = {x ∈ L : [x, y] = [y, x] = 0} is said to be
the center of L.
Definition 2.3. A linear map d : L → L of a Leibniz algebra L is a derivation if for all
x, y ∈ L
d([x, y]) = [d(x), y] + [x, d(y)].
If L is a right Leibniz algebra and x ∈ L, then the right multiplication operator
Rx : L → L defined as Rx(y) = [y, x], y ∈ L is a derivation (for a left Leibniz algebra
L with x ∈ L, the left multiplication operator Lx : L → L, Lx(y) = [x, y], y ∈ L is a
derivation).
4
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Any right Leibniz algebra L is associated with the algebra of right multiplications
R(L) = {Rx x ∈ L} endowed with the structure of a Lie algebra by means of the
commutator [Rx, Ry] = RxRy − RyRx = R[y,x], which defines an antihomomorphism
between L and R(L).
For a left Leibniz algebra L, the corresponding algebra of left multiplications L(L) =
{Lx x ∈ L} is endowed with the structure of a Lie algebra by means of the commutator
as well [Lx, Ly] = LxLy − LyLx = L[x,y]. In this case we have a homomorphism between
L and L(L).
Definition 2.4. Let d1, d2, ..., dn be derivations of a Leibniz algebra L. The derivations
d1, d2, ..., dn are said to be "nil-independent" if α1d1 + α2d2 + α3d3 + . . . + αndn is not
nilpotent for any scalars α1, α2, ..., αn ∈ F, otherwise they are "nil-dependent".
Definition 2.5. For a given Leibniz algebra L, we define the sequence of two-sided ideals
as follows:
L0 = L, Lk+1 = [Lk, L], (k ≥ 0)
L(0) = L, L(k+1) = [L(k), L(k)], (k ≥ 0),
which are the lower central series and the derived series of L, respectively.
A Leibniz algebra L is said to be nilpotent (solvable) if there exists m ∈ N such that
Lm = 0 (L(m) = 0). The minimal such number m is said to be the index of nilpotency
(solvability).
Definition 2.6. A Leibniz algebra L is called a quasi-filiform if Ln−3 6= 0 and Ln−2 = 0,
where dim(L) = n.
3. Constructing solvable Leibniz algebras with a given nilradical
Every solvable Leibniz algebra L contains a unique maximal nilpotent ideal called the
nilradical and denoted nil(L) such that dim nil(L) ≥ 1
2 dim(L) [18]. Let us consider the
problem of constructing solvable Leibniz algebras L with a given nilradical N = nil(L).
Suppose {e1, e2, e3, e4, ..., en} is a basis for the nilradical and {en+1, ..., ep} is a basis for a
subspace complementary to the nilradical.
If L is a solvable Leibniz algebra [2], then
(3.0.1)
and we have the following structure equations
(3.0.2)
where 1 ≤ i, j, k, m ≤ n and n + 1 ≤ a, b ≤ p.
3.1. Solvable right Leibniz algebras
ijek, [ea, ei] = Ak
[ei, ej] = C k
[L, L] ⊆ N
aiek, [ei, ea] = Ak
iaek, [ea, eb] = Bk
abek,
Calculations show that satisfying the right Leibniz identity is equivalent to the following
ijAm
biAm
ak, Ak
ak, Ak
iaC m
aiAm
ajC m
aiAm
ki +C k
kb+Ak
aiC m
abC m
kj = Ak
ki = Ak
conditions:
(3.1.1) Ak
ijAm
(3.1.2) Bk
abC m
Then the entries of the matrices Aa, which are (Ak
obtained from all the possible Leibniz identities between the triples {ea, ei, ej}.
1 ≤ a ≤ p) is nilpotent i.e. the matrices Aa must be "nil-independent" [8, 18].
ka +Ak
jaC m
ik ,
ki +Ak
kaAk
ib.
i )a, must satisfy the equations (3.1.1)
Since N is the nilradical of L, no nontrivial linear combination of the matrices Aa, (n+
kj +Ak
ia−Am
kj = C k
kb = Bk
ijAm
abC m
ka = Ak
ik = Am
ajC m
ibAm
ik , C k
ak, Bk
iaC m
kbAk
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
5
Let us now consider the right multiplication operator Rea and restrict it to N, (n + 1 ≤
a ≤ p). We shall get outer derivations of the nilradical N = nil(L) [8]. Then finding the
matrices Aa is the same as finding outer derivations Rea of N. Further the commutators
[Reb, Rea] = R[ea,eb], (n + 1 ≤ a, b ≤ p) due to (3.0.1) consist of inner derivations of N. So
those commutators give the structure constants Bk
ab as shown in the last equation of (3.1.2)
but only up to the elements in the center of the nilradical N, because if ei, (1 ≤ i ≤ n) is
in the center of N, then (Rei)N
is an inner derivation of the nilradical.
= 0, where (Rei)N
3.2. Solvable left Leibniz algebras
jaC m
iaAm
ki +C k
kb+Ak
aiC m
abC m
jk = Ak
ik = Ak
Satisfying the left Leibniz identity, we have
jk = C k
ak = Bk
jiAm
(3.2.1) Ak
(3.2.2) Bk
abC m
and the entries of the matrices Aa, which are (Aa)k
Similarly Aa = (Lea)N
[Lea, Leb] = L[ea,eb] give the structure constants Bk
center of the nilradical N.
ajC m
ik ,
bkAk
ai,
i , must satisfy the equations (3.2.1).
, (n + 1 ≤ a ≤ p) are outer derivations and the commutators
ab, but only up to the elements in the
kj +Ak
bi−Am
ak = Ak
ki = Am
ka +Ak
ik +Ak
jaC m
aiAm
ik , C k
kb, Bk
ijAm
abC m
jiAm
ibAm
ak, Ak
ak, Ak
iaC m
ibAm
aiC m
akAk
Once the left or right Leibniz identities are satisfied in the most general possible way
and the outer derivations are found:
(i) We can carry out the technique of "absorption" [20, 21], which means we can sim-
plify a solvable Leibniz algebra without affecting the nilradical in (3.0.2) applying
the transformation
n
i = ei, (1 ≤ i ≤ n), e′
e′
a = ea +
dkek, (n + 1 ≤ a ≤ p).
Xk=1
(ii) We change basis without affecting the nilradical in (3.0.2) to remove all the pos-
sible parameters and simplify the algebra.
4. The nilpotent sequence L 4
In L 4, (n ≥ 4) the positive integer n denotes the dimension of the algebra. The center
in the
of this algebra is C(L 4) = he2, eni. L 4 can be described explicitly as follows:
basis {e1, e2, e3, e4, . . . , en} it has only the following non-zero brackets:
(4.0.1)
[e1, e1] = e2, [ei, e1] = ei+1, (3 ≤ i ≤ n − 1), [e1, e3] = 2e2 − e4, [e3, e3] = e2,
[e1, ej] = −ej+1, (4 ≤ j ≤ n − 1, n ≥ 4).
DS = [n, n − 2, 0], LS = [n, n − 2, n − 4, n − 5, n − 6, ..., 0], (n ≥ 4).
The dimensions of the ideals in the characteristic series are
A quasi-filiform Leibniz algebra L 4 was introduced by Camacho, Gómez, González
and Omirov in [6]. This algebra is served as the nilradical for the left and right solvable
indecomposable extensions we construct in this paper.
It is shown below that solvable right (left) Leibniz algebras with the nilradical L 4 only
exist for dim g = n + 1 and dim g = n + 2(dim l = n + 1 and dim l = n + 2).
Right solvable extensions with a codimension one nilradical L 4 are found following
the steps in Theorems 5.1.1, 5.1.2 and 5.1.3 with the main result summarized in The-
orem 5.1.3, where it is shown there are eight such algebras: gn+1,i, (1 ≤ i ≤ 4, n ≥ 4),
g5,5, g5,6, g5,7 and g5,8. It is noticed that gn+1,1 is left as well when a = 0, g5,5 is left when
b = −1 and gn+1,4, g5,7, g5,8 are right and left Leibniz algebras at the same time. There are
four solvable indecomposable right Leibniz algebras with a codimension two nilradical:
gn+2,1, (n ≥ 5), g6,2, g6,3 and g6,4 stated in Theorem 5.1.4, where none of them is left.
We follow the steps in Theorems 5.2.1, 5.2.2 and 5.2.3 to find codimension one left
solvable extensions. We notice in Theorem 5.2.3, that we have eight of them as well:
ln+1,1, ln+1,2, ln+1,3, gn+1,4, l5,5, l5,6, g5,7 and g5,8, such that ln+1,1 is right when a = 0 and
6
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
l5,5 is right when b = −1. We find four solvable indecomposable left Leibniz algebras with
a codimension two nilradical as well: ln+2,1, (n ≥ 5), l6,2, l6,3 and l6,4 stated in Theorem
5.2.4, where none of them is right.
5. Classification of solvable indecomposable Leibniz algebras with a
nilradical L 4
Our goal in this Section is to find all possible right and left solvable indecomposable
extensions of the nilpotent Leibniz algebra L 4, which serves as the nilradical of the
extended algebra.
Remark 5.1. It is assumed throughout this section that solvable indecomposable right
and left Leibniz algebras have the nilradical L 4; however, most of the time, the brackets
of the nilradical will be omitted.
5.1. Solvable indecomposable right Leibniz algebras with a nilradical L 4
5.1.1. Codimension one solvable extensions of L 4
The nilpotent Leibniz algebra L 4 is defined in (4.0.1). Suppose {en+1} is in the
complementary subspace to the nilradical L 4 and g is the corresponding solvable right
Leibniz algebra. Since [g, g] ⊆ L 4, we have the following:
Xs=1
[e1, e1] = e2, [ei, e1] = ei+1, (3 ≤ i ≤ n − 1), [e1, e3] = 2e2 − e4, [e3, e3] = e2,
[e1, ej] = −ej+1, (4 ≤ j ≤ n − 1), [er, en+1] =
bs,kes,
(1 ≤ k ≤ n, 1 ≤ r ≤ n + 1, n ≥ 4).
Theorem 5.1.1. We set a1,1 := a and a3,3 := b in (5.1.1). To satisfy the right Leibniz
(5.1.1)
as,res, [en+1, ek] =
Xs=1
n
n
identity, there are the following cases based on the conditions involving parameters, each
gives a continuous family of solvable Leibniz algebras:
(1) If a1,3 = 0, b 6= −a, a 6= 0, b 6= 0, (n = 4) or b 6= (3 − n)a, a 6= 0, b 6= 0, (n ≥ 5),
then we have the following brackets for the algebra:
n
ak−1,3ek, [ei, en+1] = ((i − 3)a + b) ei+
Xk=4
ak,1ek, [e2, en+1] = 2be2, [e3, en+1] = a2,3e2 + be3+
n
n
n
Xk=5
[e1, en+1] = ae1 + A2,1e2 + (b − a)e3 +
Xk=4
Xk=i+1
[en+1, e3] = (a2,3 + 2a4,3)e2 − be3 −
Xk=i+1
[en+1, ei] = ((3 − i)a − b) ei −
ak,3ek, [e4, en+1] = (b − a)e2 + (a + b)e4 +
Xk=4
ak−i+3,3ek, [en+1, en+1] = a2,n+1e2, [en+1, e1] = −ae1 + b2,1e2 + (a − b)e3 −
Xk=5
ak,3ek, [en+1, e4] = (a + b)(e2 − e4) −
Xk=4
ak−i+3,3ek, (5 ≤ i ≤ n),
(2b − a)b2,1 − 2b · a4,1 + 2(a − b)(a2,3 + a4,3)
where A2,1 :=
n
n
.
ak,1ek,
n
ak−1,3ek,
n
a
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
7
ak−i+3,3ek, [en+1, en+1] = a2,n+1e2 + an,n+1en,
n
n
n
n
Xk=4
ak,1ek, [e2, en+1] = 2(3 − n)ae2,
Xk=4
ak−1,3ek, [ei, en+1] = (i − n)aei +
(2) If a1,3 = 0, b := −a, a 6= 0, (n = 4) or b := (3 − n)a, a 6= 0, (n ≥ 5), then
[e1, en+1] = ae1 + A2,1e2 + (2 − n)ae3 +
[e3, en+1] = a2,3e2 + (3 − n)ae3 +
ak,3ek, [e4, en+1] = (2 − n)ae2 + (4 − n)ae4+
Xk=5
Xk=i+1
Xk=4
[en+1, e1] = −ae1 + b2,1e2 + (n − 2)ae3 −
(n − 3)ae3 −
[en+1, ei] = (n − i)aei −
where A2,1 := (5 − 2n)b2,1 + 2(n − 3)a4,1 + 2(n − 2)(a2,3 + a4,3).
(3) If a1,3 = 0, a = 0, b 6= 0, (n = 4) or a = 0 and b 6= 0, (n ≥ 5), then
[e1, en+1] = a2,1e2 + be3 +
ak,3ek, [en+1, e4] = (4 − n)ae2 + (n − 4)ae4 −
ak,1ek, [e2, en+1] = 2be2, [e3, en+1] = a2,3e2 + be3 +
ak−i+3,3ek, (5 ≤ i ≤ n − 1),
ak,1ek, [en+1, e3] = (a2,3 + 2a4,3)e2+
Xk=i+1
ak−1,3ek,
Xk=5
Xk=4
n
n
n
n
n
Xk=4
ak,3ek,
n
n
Xk=4
Xk=5
[e4, en+1] = b(e2 + e4) +
ak−1,3ek, [ei, en+1] = bei +
n
Xk=i+1
ak−i+3,3ek,
n
n
n
Xk=4
Xk=4
ak,3ek, [en+1, e4] = b(e2 − e4) −
[en+1, en+1] = a2,n+1e2, [en+1, e1] = (a2,3 + a4,1 + a4,3) e2 − be3 −
[en+1, e3] = (a2,3 + 2a4,3)e2 − be3 −
[en+1, ei] = −bei −
(4) If a1,3 = 0, b = 0, a 6= 0, (n = 4) or b = 0, a 6= 0, (n ≥ 5), then
[e1, en+1] = ae1 + (a2,3 − b2,1 + b2,3) e2 − ae3 +
[e4, en+1] = −ae2 + ae4 +
ak−i+3,3ek, (5 ≤ i ≤ n).
Xk=i+1
Xk=4
n
n
n
ak,1ek,
n
Xk=5
ak−1,3ek,
Xk=5
ak−1,3ek, [ei, en+1] = (i − 3)aei +
Xk=4
n
n
ak−i+3,3ek,
Xk=i+1
ak,1ek, [en+1, e3] = b2,3e2−
n
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −ae1 + b2,1e2 + ae3 −
Xk=4
(5 ≤ i ≤ n).
ak,3ek, [en+1, e4] = ae2 − ae4 −
Xk=5
In the remaining cases a1,3 := c.
ak−1,3ek, [en+1, ei] = (3 − i)aei −
n
Xk=i+1
ak−i+3,3ek,
ak,1ek, [e3, en+1] = a2,3e2 +
n
Xk=4
ak,3ek,
Re5 = "
a
0
c
A2,1 2(b+c) a2,3 2c+b−a
b+c−a
a4,1
0
0
b
A4,3
2a
0
a+b # .
0
8
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(5) If b 6= −a, a 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = ae1 + A2,1e2 + (b + c − a)e3 + a4,1e4, [e2, e5] = 2(b + c)e2, [e3, e5] = ce1 + a2,3e2+
be3 + A4,3e4, [e4, e5] = (2c + b − a)e2 + (a + b)e4, [e5, e5] = a2,5e2, [e5, e1] = −ae1 + b2,1e2+
(a − b − c)e3 − a4,1e4, [e5, e3] = −ce1 + b2,3e2 − be3 − A4,3e4, [e5, e4] = (a + b)e2 − (a + b)e4,
where A2,1 := a2,3 − b2,1 + b2,3 −
A4,3 :=
(c − a)a2,3 + (a + c)b2,3 − 2c(b2,1 − a4,1)
(b + c)(a2,3 − 2b2,1 + 2a4,1 + b2,3)
and
a
,
(6) If b := −a, a 6= 0, a 6= c, c 6= 0, then
[e1, e5] = ae1 + A2,1e2 + (c − 2a)e3 + a4,1e4, [e2, e5] = 2(c − a)e2, [e3, e5] = ce1 + a2,3e2−
ae3 + A4,3e4, [e4, e5] = 2(c − a)e2, [e5, e5] = a2,5e2 + a4,5e4, [e5, e1] = −ae1 + b2,1e2+
(2a − c)e3 − a4,1e4, [e5, e3] = −ce1 + b2,3e2 + ae3 − A4,3e4,
where A2,1 := 2a2,3 − 3b2,1 + 2b2,3 + 2a4,1 −
A4,3 :=
(c − a)a2,3 + (a + c)b2,3 − 2c(b2,1 − a4,1)
c(a2,3 − 2b2,1 + 2a4,1 + b2,3)
and
a
,
0
c
Re5 = " a
A2,1 2(c−a) a2,3 2(c−a)
c−2a
a4,1
−a
A4,3
0
0
0
0
# .
2a
0
c
a
0
0
0
0 −c
(7) If b := −c, c 6= 0, a 6= c, a 6= 0, then
[e1, e5] = ae1 + (a2,3 − b2,1 + b2,3) e2 − ae3 + a4,1e4, [e3, e5] = ce1 + a2,3e2 − ce3 + a4,3e4,
[e4, e5] = (c − a)e2 + (a − c)e4, [e5, e5] = a2,5e2, [e5, e1] = −ae1 + b2,1e2 + ae3 − a4,1e4,
[e5, e3] = −ce1 + b2,3e2 + ce3 − a4,3e4, [e5, e4] = (a − c)e2 + (c − a)e4,
0 a4,3 a−c# .
(8) If c := a, b := −a, a 6= 0, then1
[e1, e5] = ae1 + (a2,3 − b2,1 + b2,3) e2 − ae3 + a4,1e4, [e3, e5] = ae1 + a2,3e2 − ae3 + a4,3e4,
[e5, e5] = a2,5e2 + a4,5e4, [e5, e1] = −ae1 + b2,1e2 + ae3 − (a4,1 − a4,3 − b4,3) e4,
[e5, e3] = −ae1 + b2,3e2 + ae3 + b4,3e4,
Re5 = "
a2,3−b2,1+b2,3 0 a2,3 c−a
−a
a4,1
−a
a4,1
Re5 = "
0 a 0
a2,3−b2,1+b2,3 0 a2,3 0
0 −a 0
0 a4,3 0# .
(9) If a = 0, b = 0, c 6= 0, then
[e1, e5] = (3b2,1 − 4a4,1 − 2a2,3 − 2a4,3) e2 + ce3 + a4,1e4, [e2, e5] = 2ce2, [e3, e5] = ce1+
a2,3e2 + a4,3e4, [e4, e5] = 2ce2, [e5, e5] = a2,5e2 + a4,5e4, [e5, e1] = b2,1e2 − ce3 − a4,1e4,
[e5, e3] = −ce1 + (2b2,1 − 2a4,1 − a2,3) e2 − a4,3e4,
a
Re5 = "
0
0
3b2,1−4a4,1−2a2,3−2a4,3 2c a2,3 2c
0
0 a4,3 0 # .
a4,1
0
0
0
c
c
1The outer derivation Re5 is nilpotent, so we do not consider this case any further.
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
9
(10) If a = 0, b 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = a2,1e2 + (b + c)e3 + a4,1e4, [e2, e5] = 2(b + c)e2, [e3, e5] = ce1 + a2,3e2 + be3+
A4,3e4, [e4, e5] = (2c + b)e2 + be4, [e5, e5] = a2,5e2, [e5, e1] = a4,1 +
(b + c)e3 − a4,1e4, [e5, e3] = −ce1 + b2,3e2 − be3 − A4,3e4, [e5, e4] = b(e2 − e4),
where A4,3 :=
(3c + 2b)b2,3 − (2b + c)a2,3 − 2c(a2,1 + a4,1)
! e2−
a2,3 + b2,3
2
,
4(b + c)
0
0
0
0
0
c
b
A4,3
Re5 = " 0
a2,1 2(b+c) a2,3 2c+b
b+c
a4,1
b # .
(11) If a = 0, b := −c, c 6= 0, then
[e1, e5] = a2,1e2 + a4,1e4, [e3, e5] = ce1 + a2,3e2 − ce3 + a4,3e4, [e4, e5] = ce2 − ce4,
[e5, e5] = a2,5e2, [e5, e1] = (a2,3 + b2,3 − a2,1) e2 − a4,1e4, [e5, e3] = −ce1 + b2,3e2 + ce3−
a4,3e4, [e5, e4] = −ce2 + ce4,
a4,1 0 a4,3 −c# .
Re5 = " 0
0
a2,1 0 a2,3 c
0
0 −c
0
c
0
Proof.
(1) For (n ≥ 5) the proof is given in Table 1. For (n = 4) we work out the
following identities: 1., 3. − 5., 7., 8., 10., 12., 13.(or 15.), 16. − 19.
(2) For (n ≥ 5), we apply the same identities given in Table 1, except 17. For (n = 4),
the identities are as follows: 1., 3. − 5., 7., 8., 10., 12., 15., 16. − 19.
(3) For (n ≥ 5), same identities except 18. and 19. applying instead Re3 ([en+1, en+1]) =
[Re3(en+1), en+1] + [en+1, Re3(en+1)] and R[en+1, e1] = [R(en+1), e1] + [en+1, R(e1)].
For (n = 4), same identities as in case (1), except 18. and 19. applying two
identities given above.
(4) Same identities as in case (1).
(5) We apply the following identities: 1., 3.−5., 7., 8., 10., 12., 13., 16., 17., Re3 ([e5, e5]) =
[Re3(e5), e5] + [e5, Re3(e5)], 18., 19.
(6) We apply the same identities as in (5), except 17. and 18.
(7) Same as in (5), except 19.
(8) We apply the same identities as in (5), except 17., 18. and 19.
(9) Same identities as in (5), except 17. and 18.
(10) We have the identities: 1., 3. − 5., 7., 8., 10., 12., 13., 16., 17., 19., Re5 ([e5, e1]) =
[Re5(e5), e1] + [e5, Re5(e1)], Re3 ([e5, e5]) = [Re3(e5), e5] + [e5, Re3(e5)].
(11) Same identities as in (5), except 19.
(cid:3)
10
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Table 1. Right Leibniz identities in case (1) in Theorem 5.1.1, (n ≥ 5).
Result
a1,2 = 0, a3,1 := 1
2 a2,2 − a, ak,2 = 0, (3 ≤ k ≤ n) =⇒ [e1, en+1] = ae1 + a2,1e2 +
Steps Ordered triple
1.
2.
3.
4.
5.
6.
7.
8.
9.
R[e1, e1]
R[e3, ei]
R[e3, en]
R[e3, e3]
R[e1, e3]
R[e1, ei]
Re1 ([en+1, e1])
Re1 ([e3, en+1])
Re1 ([e1, en+1])
10.
Re3 ([e1, en+1])
11.
Rei ([e1, en+1])
12.
13.
14.
Ren ([e1, en+1])
Re1 ([en+1, e3])
Re1 ([en+1, ei])
15.
Re1 ([en+1, en−1])
16.
R[e1, en+1]
k=4 ak,iek, (4 ≤ i ≤ n − 1).
k=4 ak,1ek, [e2, en+1] = a2,2e2.
2 a2,2 − a(cid:1) e3 +Pn
(cid:0) 1
a1,3 = a1,i = a3,i = 0 =⇒ [e3, en+1] = a2,3e2 + be3 +Pn
a2,ie2 +Pn
a1,n = a3,n = 0 =⇒ [en, en+1] = a2,ne2 +Pn
a2,2 := 2b =⇒ [e1, en+1] = ae1 + a2,1e2 + (b − a)e3 +Pn
(a + b)e4 +Pn
k=5 ak−1,3ek.
k=4 ak,nek.
k=4 ak,1ek, [e2, en+1] = 2be2.
a2,4 := b − a, a4,4 := a + b, ak,4 := ak−1,3, (5 ≤ k ≤ n) =⇒ [e4, en+1] = (b − a)e2 +
a2,i+1 = a4,i+1 = 0, ai+1,i+1 := (i − 2)a + b, ak,i+1 := ak−1,i, (5 ≤ k ≤ n, k 6=
k=4 ak,3ek, [ei, en+1] =
i + 1, 4 ≤ i ≤ n − 1), where i is fixed =⇒ [ej , en+1] = ((j − 3)a + b) ej +
bk−1,1 := −ak−1,1, (5 ≤ k ≤ n, n ≥ 5) =⇒ [en+1, e1] = −ae1 + b2,1e2 + (a − b)e3 −
k=j+1 ak−j+3,3ek, (5 ≤ j ≤ n).
[en+1, e2] = 0 =⇒ bk,2 = 0, (1 ≤ k ≤ n).
Pn
b1,1 := −a, b3,1 := a − b =⇒ [en+1, e1] = −ae1 + b2,1e2 + (a − b)e3 +Pn
Pn−1
Pn−1
b3,i = 0 =⇒ b1,i = 0; bi,i
k=4 ak,1ek + bn,1en.
k=4 ak,3ek + bn,3en.
b1,3 = 0, b3,3 := −b, bk−1,3 := −ak−1,3, (5 ≤ k ≤ n) =⇒ [en+1, e3] = b2,3e2 − be3 −
:= (3 − i)a − b, bk−1,i = 0, (5 ≤ k ≤ i), bk−1,i
:=
k=4 bk,1ek.
−ak−i+2,3, (i+2 ≤ k ≤ n), where i is fixed. =⇒ [en+1, ei] = b2,ie2+((3 − i)a − b) ei−
k=i+1 ak−i+3,3ek + bn,ien, (4 ≤ i ≤ n − 1).
b3,n = 0 =⇒ b1,n = 0; bk−1,n = 0, (5 ≤ k ≤ n) =⇒ [en+1, en] = b2,ne2 + bn,nen.
k=5 ak−1,3ek.
b2,i+1 = 0, bn,i+1 := −an−i+2,3, (4 ≤ i ≤ n − 2) =⇒ [en+1, ej] = ((3 − j)a − b) ej −
Pn−1
b2,4 := a + b, bn,4 := −an−1,3 =⇒ [en+1, e4] = (a + b)e2 − (a + b)e4 −Pn
Pn
14., [en+1, ei] = ((3 − i)a − b) ei −Pn
a3,n+1 = 0 =⇒ a1,n+1 = 0; ak−1,n+1 = 0, (5 ≤ k ≤ n) =⇒ [en+1, en+1] =
b2,n = 0, bn,n := (3 − n)a − b =⇒ [en+1, en] = ((3 − n)a − b) en. Combining with
k=j+1 ak−j+3,3ek, (5 ≤ j ≤ n − 1).
k=i+1 ak−i+3,3ek, (5 ≤ i ≤ n).
a2,n+1e2 + an,n+1en.
17.
18.
R[en+1, en+1]
an,n+1 = 0, (b 6= (3 − n)a) =⇒ [en+1, en+1] = a2,n+1e2.
R[en+1, e3]
bn,3 := −an,3, (a 6= 0), b2,3 := a2,3 + 2a4,3, (b 6= 0) =⇒ [en+1, e3] = (a2,3 + 2a4,3) e2 −
19.
Re1 ([en+1, en+1])
bn,1 := −an,1, (a 6= 0), A2,1 := (2b−a)b2,1−2b·a4,1+2(a−b)(a2,3+a4,3)
=⇒ [en+1, e1] =
a
k=4 ak,1ek, [e1, en+1] = ae1 + A2,1e2 + (b − a)e3 +
k=4 ak,3ek.
be3 −Pn
−ae1 + b2,1e2 + (a − b)e3 − Pn
Pn
k=4 ak,1ek.
Theorem 5.1.2. Applying the technique of "absorption" (see Section 3.2), we can further
simplify the algebras in each of the cases in Theorem 5.1.1 as follows:
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
11
(1) If a1,3 = 0, b 6= −a, a 6= 0, b 6= 0, (n = 4) or b 6= (3 − n)a, a 6= 0, b 6= 0, (n ≥ 5),
then we have the following brackets for the algebra:
n
n
[e1, en+1] = ae1 + A2,1e2 + (b − a)e3, [e2, en+1] = 2be2, [e3, en+1] = a2,3e2 + be3 +
[e4, en+1] = (b − a)e2 + (a + b)e4 +
Xk=i+2
[en+1, e4] = (a + b) (e2 − e4) −
ak−i+3,3ek, [en+1, e1] = −ae1 + b2,1e2 + (a − b)e3, [en+1, e3] = a2,3e2 − be3 −
Xk=i+2
ak−1,3ek, [en+1, ei] = ((3 − i)a − b) ei −
ak−1,3ek, [ei, en+1] = ((i − 3)a + b) ei+
Xk=6
n
n
Xk=6
(2b − a)b2,1 + 2(a − b)a2,3
(5 ≤ i ≤ n); where A2,1 :=
(2) If a1,3 = 0, b := −a, a 6= 0, (n = 4) or b := (3 − n)a, a 6= 0, (n ≥ 5), then the
a
.
n
Xk=5
n
Xk=5
ak,3ek,
ak,3ek,
ak−i+3,3ek,
brackets for the algebra are
n
[e1, en+1] = ae1 + A2,1e2 + (2 − n)ae3, [e2, en+1] = 2(3 − n)ae2, [e3, en+1] = a2,3e2+
(3 − n)ae3 +
[ei, en+1] = (i − n)aei +
ak,3ek, [e4, en+1] = (2 − n)ae2 + (4 − n)ae4 +
ak−i+3,3ek, [en+1, en+1] = an,n+1en,
ak−1,3ek,
Xk=6
Xk=5
n
n
Xk=i+2
n
Xk=5
[en+1, e1] = −ae1 + b2,1e2 + (n − 2)ae3, [en+1, e3] = a2,3e2 + (n − 3)ae3 −
Xk=i+2
[en+1, e4] = (4 − n)a(e2 − e4) −
(5 ≤ i ≤ n); where A2,1 := (5 − 2n)b2,1 + 2(n − 2)a2,3.
(3) If a1,3 = 0, a = 0, b 6= 0, (n = 4) or a = 0 and b 6= 0, (n ≥ 5), then
Xk=5
[e1, en+1] = a2,1e2 + be3, [e2, en+1] = 2be2, [e3, en+1] = a2,3e2 + be3 +
ak−1,3ek, [en+1, ei] = (n − i)aei −
Xk=6
n
n
n
n
ak,3ek,
[e4, en+1] = b (e2 + e4) +
Xk=6
ak−1,3ek, [ei, en+1] = bei +
n
ak−i+3,3ek,
Xk=i+2
n
ak,3ek,
ak−i+3,3ek,
n
ak,3ek,
Xk=6
Xk=5
ak−1,3ek, [en+1, ei] = −bei −
[en+1, e1] = a2,3e2 − be3, [en+1, e3] = a2,3e2 − be3 −
[en+1, e4] = b (e2 − e4) −
(4) If a1,3 = 0, b = 0, a 6= 0, (n = 4) or b = 0, a 6= 0, (n ≥ 5), then
Xk=5
[e1, en+1] = ae1 + (a2,3 − b2,1 + b2,3) e2 − ae3, [e3, en+1] = a2,3e2 +
Xk=i+2
[e4, en+1] = −a(e2 − e4) +
ak−1,3ek, [ei, en+1] = (i − 3)aei +
Xk=i+2
Xk=6
n
n
n
n
ak−i+3,3ek, (5 ≤ i ≤ n).
ak,3ek,
ak−i+3,3ek,
n
Xk=5
ak,3ek,
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −ae1 + b2,1e2 + ae3, [en+1, e3] = b2,3e2 −
[en+1, e4] = a(e2 − e4) −
ak−1,3ek, [en+1, ei] = (3 − i)aei −
n
n
Xk=6
Xk=i+2
ak−i+3,3ek, (5 ≤ i ≤ n).
In the remaining cases a1,3 := c.
12
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(5) If b 6= −a, a 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = ae1 + A2,1e2 + (b + c − a)e3, [e2, e5] = 2(b + c)e2, [e3, e5] = ce1 + a2,3e2 + be3,
[e4, e5] = (2c + b − a)e2 + (a + b)e4, [e5, e1] = −ae1 + B2,1e2 + (a − b − c)e3,
[e5, e3] = −ce1 + B2,3e2 − be3, [e5, e4] = (a + b)(e2 − e4),
where A2,1 :=
(3a − 2b − 3c)a2,3 + (a − 2b − 3c)b2,3
(a − c)a2,3 − (a + c)b2,3
, B2,1 :=
,
2a
2a
B2,3 :=
(a − c)a2,3 − c · b2,3
,
a
Re5 = "
a
0
b+c−a
0
0
0
A2,1 2(b+c) a2,3 2c+b−a
c
b
0
0
a+b # .
0
(6) If b := −a, a 6= 0, a 6= c, c 6= 0, then
[e1, e5] = ae1 + A2,1e2 + (c − 2a)e3, [e2, e5] = 2(c − a)e2, [e3, e5] = ce1 + a2,3e2 − ae3,
[e4, e5] = 2(c − a)e2, [e5, e5] = a4,5e4, [e5, e1] = −ae1 + B2,1e2 + (2a − c)e3,
(5a − 3c)a2,3 + (3a − 3c)b2,3
[e5, e3] = −ce1 + B2,3e2 + ae3; where A2,1 :=
B2,1 :=
2a
(a − c)a2,3 − c · b2,3
,
, B2,3 :=
,
a
(a − c)a2,3 − (a + c)b2,3
# .
A2,1 2(c−a) a2,3 2(c−a)
c−2a
2a
0
0
c
0
0
0
0
0
−a
0
(7) If b := −c, c 6= 0, a 6= c, a 6= 0, then
Re5 = " a
[e1, e5] = ae1 + (a2,3 − b2,1 + b2,3) e2 − ae3, [e3, e5] = ce1 + a2,3e2 − ce3,
[e4, e5] = (c − a) (e2 − e4) , [e5, e5] = a2,5e2, [e5, e1] = −ae1 + b2,1e2 + ae3,
[e5, e3] = −ce1 + b2,3e2 + ce3, [e5, e4] = (a − c) (e2 − e4) ,
Re5 = "
a2,3−b2,1+b2,3 0 a2,3 c−a
0 a−c# .
0 −c
0
−a
0
0
0
0
a
c
(8) If a = 0, b = 0, c 6= 0, then
[e1, e5] = (3b2,1 − 2a2,3) e2 + ce3, [e2, e5] = 2ce2, [e3, e5] = ce1 + a2,3e2, [e4, e5] = 2ce2,
[e5, e5] = a4,5e4, [e5, e1] = b2,1e2 − ce3, [e5, e3] = −ce1 + (2b2,1 − a2,3) e2,
Re5 = "
0
0
0
3b2,1−2a2,3 2c a2,3 2c
0
c
0
0
0
0
0 # .
c
0
(9) If a = 0, b 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = A2,1e2 + (b + c)e3, [e2, e5] = 2(b + c)e2, [e3, e5] = ce1 + a2,3e2 + be3,
[e4, e5] = (2c + b)e2 + be4, [e5, e1] = B2,1e2 − (b + c)e3, [e5, e3] = −ce1 + B2,3e2 − be3,
[e5, e4] = b(e2 − e4); where A2,1 :=
B2,1 :=
(2b + c)a2,3 − (2b + 3c)b2,3
(2b + c)a2,3 − c · b2,3
, B2,3 :=
4(b + c)
,
,
2(b + c)
(4b + 3c)a2,3 − c · b2,3
b # .
A2,1 2(b+c) a2,3 2c+b
b+c
4(b + c)
c
Re5 = " 0
0
0
0
0
0
b
0
0
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
13
(10) If a = 0, b := −c, c 6= 0, then
[e1, e5] = a2,1e2, [e3, e5] = ce1 + a2,3e2 − ce3, [e4, e5] = c(e2 − e4), [e5, e5] = a2,5e2,
[e5, e1] = (a2,3 + b2,3 − a2,1) e2, [e5, e3] = −ce1 + b2,3e2 + ce3, [e5, e4] = −c(e2 − e4),
Re5 = " 0
0
a2,1 0 a2,3 c
0
0 −c# .
0 −c
0
0
0
0
c
Proof.
(1) The right (a derivation) and left (not a derivation) multiplication operators
restricted to the nilradical are given below:
Ren+1 =
·
·
·
a5,3
a4,3
ai,1
ai,3
ai−1,3 ai−2,3 ···
(i−3)a+b ···
2a+b
...
...
···
···
···
···
···
a4,3
...
...
0
0
0
0
0
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
...
...
...
0
...
...
...
0
...
...
...
0
0
an−1,1 0 an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3 (n−4)a+b
an,1
0
0
0
0
0 a2,3+2a4,3
0
0 −a4,3
0 −a5,3
an,3 an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3
···
···
···
···
···
−a−b
−a4,3 −2a−b
···
···
···
···
···
0
0
0
0
0
0
0
0
0
a+b
−b
0
−a
b2,1
a−b
−a4,1
−a5,1
a4,3
0
0
0
0
0
(n−3)a+b
0
0
0
0
0
Len+1 =
·
·
·
−a5,3
−a4,3
−ai,1
−ai,3
−ai−1,3 −ai−2,3 ···
−a4,3
(3−i)a−b ···
...
...
...
...
...
...
...
...
...
...
0
...
...
...
0
,
0
0
0
0
0
...
...
...
0
0
.
...
...
a
0
0
A2,1 2b a2,3
b−a
a4,1
a5,1
a4,3
a5,3
0
0
0
b
0
b−a
0
a+b
a4,3
...
...
0
...
...
0
...
...
...
...
...
...
...
...
...
...
0
0
0
0
...
...
...
...
−an−1,1 0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 (4−n)a−b
−an,1
0 −an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3
(3−n)a−b
• The transformation e′
k = ek, (1 ≤ k ≤ n), e′
n+1 = en+1 − a4,3e1 removes a4,3
in Ren+1 and −a4,3 in Len+1 from the (i, i − 1)st positions, where (4 ≤ i ≤ n),
but it affects other entries as well, such as the entry in the (2, 1)st position in
Ren+1 and Len+1, which we change to A2,1 − a4,3 and b2,1 − a4,3, respectively.
It also changes the entry in the (2, 3)rd position in Len+1 to a2,3. At the same
time, it affects the coefficient in front of e2 in the bracket [en+1, en+1], which
we change back to a2,n+1.
i = ei, (1 ≤ i ≤ n), e′
• Applying the transformation e′
k=3 ak+1,1ek,
we remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries in the (k +
1, 1)st positions, where (3 ≤ k ≤ n − 1). It changes the entry in the (2, 1)st
position in Ren+1 to A2,1 + 2a4,1 − a4,3, the entries in the (2, 3)rd positions in
Ren+1 and Len+1 to a2,3 + a4,1. It also affects the coefficient in front of e2 in
n+1 = en+1+Pn−1
[en+1, en+1], which we rename back by a2,n+1. We assign a2,3 + a4,1 := a2,3 and
b2,1 − a4,3 := b2,1. Then A2,1 + 2a4,1 − a4,3 := (2b−a)b2,1+2(a−b)a2,3
.
a
14
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
• The transformation e′
j = ej, (1 ≤ j ≤ n), e′
n+1 = en+1 − a2,n+1
2b e2 removes the
coefficient a2,n+1 in front of e2 in [en+1, en+1] and we prove the result.
(2) The right (a derivation) and left (not a derivation) multiplication operators re-
stricted to the nilradical are as follows:
0
0
0
0
2(3−n)a
(3−n)a
(2−n)a
(4−n)a
(2−n)a
A2,1
a2,3
0
0
0
0
a
Ren+1 =
·
·
·
a5,3
a4,3
a4,3
a5,3
...
...
a4,3
(5−n)a
...
...
···
a4,3
...
...
...
...
...
...
···
···
···
···
···
0
0
0
0
0
···
···
···
···
···
0
0
0
0
0
0 0
0 0
0 0
0 0
0 0
ai,3
ai−1,3 ai−2,3
(i−n)a ···
0 0
an,3
an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3 −a 0
an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3 a4,3 0
0
0
0
0
0
(n−4)a
−a4,3
···
···
···
···
···
···
···
···
···
···
(n−5)a
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
−a
b2,1
0
0 a2,3+2a4,3 (4−n)a
0
(n−2)a 0
−a4,1
−a5,1
(n−3)a
0 −a4,3
0 −a5,3
·
·
·
−ai,1
−ai,3
−ai−1,3 −ai−2,3 ···
−a4,3
(n−i)a
···
...
...
−a5,3
−a4,3
...
...
...
...
...
...
,
...
...
...
...
...
...
.
0
0
0
0
0
...
...
...
0
...
...
...
0
...
...
...
0
...
...
...
0
...
...
...
...
...
...
0
0
0
0
...
...
0
0
0
...
...
a4,1
a5,1
ai,1
...
...
an−1,1
an,1
...
...
0
...
...
Len+1 =
−an−1,1 0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3
0
0 −an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 0
−an,1
a
• The transformation e′
k = ek, (1 ≤ k ≤ n), e′
n+1 = en+1 − a4,3e1 removes a4,3
in Ren+1 and −a4,3 in Len+1 from the (i, i − 1)st positions, where (4 ≤ i ≤ n),
but it affects other entries as well, such as the entry in the (2, 1)st position in
Ren+1 and Len+1, which we change to A2,1 − a4,3 and b2,1 − a4,3, respectively.
It also changes the entry in the (2, 3)rd position in Len+1 to a2,3. At the same
time, it affects the coefficient in front of e2 in the bracket [en+1, en+1], which
we change back to a2,n+1.
i = ei, (1 ≤ i ≤ n), e′
• Then we apply the transformation e′
n+1 = en+1 +
Pn−1
k=3 ak+1,1ek to remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries
in the (k + 1, 1)st positions, where (3 ≤ k ≤ n−1). It changes the entry in the
(2, 1)st position in Ren+1 to A2,1+2a4,1−a4,3, the entries in the (2, 3)rd positions
in Ren+1 and Len+1 to a2,3 + a4,1. It also affects the coefficient in front of e2
in [en+1, en+1], which we rename back by a2,n+1. We assign a2,3 + a4,1 := a2,3
and b2,1 − a4,3 := b2,1. Then A2,1 + 2a4,1 − a4,3 := (5 − 2n)b2,1 + 2(n − 2)a2,3,
which we set to be A2,1.
• Applying the transformation e′
2(n−3)a e2, we
remove the coefficient a2,n+1 in front of e2 in [en+1, en+1] and prove the result.
j = ej, (1 ≤ j ≤ n), e′
n+1 = en+1 + a2,n+1
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
15
(3) The right (a derivation) and left (not a derivation) multiplication operators re-
stricted to the nilradical are
0
a2,1
b
a4,1
a5,1
0
0
2b a2,3
0
0
0
a4,3
a5,3
b
0
b
0
b
a4,3
0
0
0
0
b
···
···
···
···
···
0
0
0
0
0
Ren+1 =
·
·
·
a5,3
a4,3
ai,1
ai,3
ai−1,3 ai−2,3 ···
...
...
...
...
a4,3
...
...
...
...
...
b
a2,3+a4,1+a4,3 0 a2,3+2a4,3
an,1
0
an−1,1 0 an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3
0
an,3 an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3 a4,3 b
0
0
0
0
0
0
0
0
0
0
−b
0
b
0
−b
···
···
···
···
···
···
···
···
···
···
0
0
0
0
0
−a4,3
b
−b
−a4,1
−a5,1
−b
0
0 −a4,3
0 −a5,3
Len+1 =
·
·
·
−a5,3
−a4,3
−ai,1
−ai,3
−ai−1,3 −ai−2,3 ···
−a4,3
−b
···
...
...
...
...
...
...
0
...
...
...
...
...
...
0
...
...
0
0
...
...
...
...
...
...
...
...
···
···
···
···
···
···
...
...
...
0
0
0
0
0
0 0
0 0
0 0
0 0
0 0
...
...
...
0
...
...
...
0 0
,
0
0
0
0
0
0
0
0
0
0
...
...
...
...
...
...
0
.
...
...
...
0
...
...
...
0
−an−1,1
−an,1
0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 −b
0
0 −an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 −b
• The transformation e′
k = ek, (1 ≤ k ≤ n), e′
n+1 = en+1 − a4,3e1 removes a4,3
in Ren+1 and −a4,3 in Len+1 from the (i, i − 1)st positions, where (4 ≤ i ≤ n),
but it affects other entries as well, such as the entry in the (2, 1)st position in
Ren+1 and Len+1, which we change to a2,1 − a4,3 and a2,3 + a4,1, respectively.
It also changes the entry in the (2, 3)rd position in Len+1 to a2,3. At the same
time, it affects the coefficient in front of e2 in the bracket [en+1, en+1], which
we change back to a2,n+1.
• Applying the transformation e′
i = ei, (1 ≤ i ≤ n), e′
n+1 = en+1+Pn−1
k=3 ak+1,1ek,
we remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries in the (k +
1, 1)st positions, where (3 ≤ k ≤ n − 1). It changes the entry in the (2, 1)st
position in Ren+1 to a2,1 + 2a4,1 − a4,3, the entries in the (2, 3)rd positions in
Ren+1 and Len+1 to a2,3 + a4,1. It also affects the coefficient in front of e2 in
[en+1, en+1], which we rename back by a2,n+1. Then we assign a2,1 + 2a4,1 −
a4,3 := a2,1 and a2,3 + a4,1 := a2,3.
• The transformation e′
j = ej, (1 ≤ j ≤ n), e′
n+1 = en+1 − a2,n+1
2b e2 removes the
coefficient a2,n+1 in front of e2 in [en+1, en+1] and we prove the result.
···
···
···
···
···
0
0
0
0
0
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
a
0
0
a2,3−b2,1+b2,3 0 a2,3
−a
a4,1
a5,1
0
0
0 a4,3
0 a5,3
·
·
·
...
...
...
...
0
−a
0
a
a4,3
...
...
0
0
0
0
2a
...
...
a5,3
a4,3
...
...
a4,3
...
...
ai,1
0 ai,3
ai−1,3 ai−2,3 ···
(i−3)a ···
Ren+1 =
an−1,1
0
b2,3
an,1
0
0
0
0 −a4,3
0 −a5,3
0
−a
b2,1
a
−a4,1
−a5,1
0 an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3 (n−4)a
0 an,3 an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3
0
0
0
0
0
a4,3
0
0
0
0
0
0
a
0
−a
···
···
···
···
···
···
···
···
···
···
0
0
0
0
0
0
0
0
0
−2a
−a4,3
Len+1 =
·
·
·
−a5,3
−a4,3
...
...
...
...
...
...
...
...
...
...
...
...
−ai,1
0 −ai,3 −ai−1,3 −ai−2,3 ···
−a4,3
(3−i)a
···
...
...
...
...
...
...
...
...
...
...
...
...
...
0
...
...
...
0
...
...
...
0
...
...
...
0
...
...
...
0
0
(n−3)a
0
0
0
0
0
...
...
...
0
0
,
.
16
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(4) The right (a derivation) and left (not a derivation) multiplication operators re-
stricted to the nilradical are given below:
−an−1,1 0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 (4−n)a
−an,1
0 −an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 (3−n)a
• The transformation e′
k = ek, (1 ≤ k ≤ n), e′
n+1 = en+1 − a4,3e1 removes a4,3 in
Ren+1 and −a4,3 in Len+1 from the entries in the (i, i − 1)st positions, where
(4 ≤ i ≤ n), but it affects other entries as well, such as the entry in the
(2, 1)st position in Ren+1 and Len+1, which we change to a2,3 − b2,1 − a4,3 + b2,3
and b2,1 − a4,3, respectively. It also changes the entry in the (2, 3)rd position
in Len+1 to b2,3 − 2a4,3. At the same time, it affects the coefficient in front of
e2 in the bracket [en+1, en+1], which we change back to a2,n+1.
• Applying the transformation e′
i = ei, (1 ≤ i ≤ n), e′
n+1 = en+1+Pn−1
k=3 ak+1,1ek,
we remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries in the (k +
1, 1)st positions, where (3 ≤ k ≤ n − 1). It changes the entry in the (2, 1)st
position in Ren+1 to 2a4,1 + a2,3 − b2,1 − a4,3 + b2,3, the entries in the (2, 3)rd
positions in Ren+1 and Len+1 to a2,3 +a4,1 and b2,3 +a4,1−2a4,3, respectively. It
also affects the coefficient in front of e2 in [en+1, en+1], which we rename back
by a2,n+1. We assign a2,3 +a4,1 := a2,3, b2,1−a4,3 := b2,1 and b2,3 +a4,1−2a4,3 :=
b2,3. Then 2a4,1 + a2,3 − b2,1 − a4,3 + b2,3 := a2,3 − b2,1 + b2,3.
(5) We apply the transformation e′
2a4,1A4,3 + b2,1A4,3 − A2
assign a2,3 + a4,1 := a2,3 and b2,3 − 2b2,1 + a4,1 := b2,3.
4,3 − a2,3a4,1 − a2
i = ei, (1 ≤ i ≤ 4), e′
5 = e5−A4,3e1+ 1
2(b+c) (A2,1A4,3+
4,1 − a4,1b2,3 − a2,5)e2 + a4,1e3 and then we
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
17
(6) The transformation we apply is e′
i = ei, (1 ≤ i ≤ 4), e′
5 = e5−A4,3e1+ 1
2(c−a) (A2,1A4,3+
2a4,1A4,3 + b2,1A4,3 − A2
a2,3 + a4,1 := a2,3 and b2,3 − 2b2,1 + a4,1 := b2,3.
4,3 − a2,3a4,1 − a2
4,1 − a4,1b2,3 − a2,5)e2 + a4,1e3. We assign
(7) We apply the transformation e′
5 = e5 − a4,3e1 + a4,1e3 and
rename the coefficient in front of e2 in [e5, e5] back by a2,5. Then we assign a2,3 +
a4,1 := a2,3, b2,1 − a4,3 := b2,1 and b2,3 + a4,1 − 2a4,3 := b2,3.
i = ei, (1 ≤ i ≤ 4), e′
(8) We apply e′
2a4,1b2,1 + 3a2
b2,1 − a4,3 := b2,1.
i = ei, (1 ≤ i ≤ 4), e′
4,1 + 2a4,1a4,3 +
4,3 − 4a4,3b2,1 + a2,5)e2 + a4,1e3. We assign a2,3 + a4,1 := a2,3 and
5 = e5 − a4,3e1 − 1
2c(2a2,3a4,3 − a2
(9) The transformation is e′
2a2,3a4,1 + a2,3A4,3− 2a2
Then we assign a2,3 + a4,1 := a2,3 and b2,3 − 2a2,1 − 3a4,1 := b2,3.
i = ei, (1 ≤ i ≤ 4), e′
4,1 + 6a4,1A4,3− 2a4,1b2,3 − 2A2
5 = e5 − A4,3e1 + 1
4(b+c) (2a2,1A4,3 −
4,3 + b2,3A4,3− 2a2,5)e2 + a4,1e3.
(10) We apply the transformation e′
5 = e5 − a4,3e1 + a4,1e3 and
rename the coefficient in front of e2 in [e5, e5] back by a2,5. We assign a2,3 + a4,1 :=
a2,3, a2,1 + 2a4,1 − a4,3 := a2,1 and b2,3 + a4,1 − 2a4,3 := b2,3.
i = ei, (1 ≤ i ≤ 4), e′
(cid:3)
18
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Theorem 5.1.3. There are eight solvable indecomposable right Leibniz algebras up to
isomorphism with a codimension one nilradical L 4, (n ≥ 4), which are given below:
(i) gn+1,1 : [e1, en+1] = e1 + (a − 1)e3, [e2, en+1] = 2ae2, [e3, en+1] = ae3,
[e4, en+1] = (a − 1)e2 + (a + 1)e4, [ei, en+1] = (a + i − 3) ei, [en+1, e1] = −e1 − (a − 1)e3,
[en+1, e3] = −ae3, [en+1, e4] = (a + 1) (e2 − e4) , [en+1, ei] = (3 − i − a) ei, (5 ≤ i ≤ n),
(ii) gn+1,2 : [e1, en+1] = e1 + (2 − n)e3, [e2, en+1] = 2(3 − n)e2, [e3, en+1] = (3 − n)e3,
[e4, en+1] = (2 − n)e2 + (4 − n)e4, [ei, en+1] = (i − n)ei, [en+1, en+1] = en,
[en+1, e1] = −e1 + (n − 2)e3, [en+1, e3] = (n − 3)e3, [en+1, e4] = (4 − n) (e2 − e4) ,
[en+1, ei] = (n − i)ei, (5 ≤ i ≤ n),
(iii) gn+1,3 : [e1, en+1] = e3, [e2, en+1] = 2e2, [e4, en+1] = e2, [ei, en+1] = ei + ǫei+2+
n
n
bk−i−2ek,
Xk=i+3
bk−i−2ek, [en+1, e1] = −e3, [en+1, e4] = e2, [en+1, ei] = −ei − ǫei+2 −
Xk=i+3
(ǫ = 0, 1, 3 ≤ i ≤ n),
(iv) gn+1,4 : [e1, en+1] = e1 − e3, [e3, en+1] = de2, [e4, en+1] = e4 − e2, [ei, en+1] = (i − 3)ei,
[en+1, en+1] = ǫe2, [en+1, e1] = −e1 + (d + f ) e2 + e3, [en+1, e3] = f e2, [en+1, e4] = e2 − e4,
[en+1, ei] = (3 − i)ei, (5 ≤ i ≤ n; ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0),
(v) g5,5 : [e1, e5] = ae1 + (b − a + 1)e3, [e2, e5] = 2(b + 1)e2, [e3, e5] = e1 + be3,
[e4, e5] = (b − a + 2)e2 + (a + b)e4, [e5, e1] = −ae1 + (a − b − 1)e3, [e5, e3] = −e1 − be3,
[e5, e4] = (a + b)(e2 − e4), (if b = −1, then a 6= 1),
(vi) g5,6 : [e1, e5] = ae1 + (1 − 2a)e3, [e2, e5] = 2(1 − a)e2, [e3, e5] = e1 − ae3,
[e4, e5] = 2(1 − a)e2, [e5, e5] = e4, [e5, e1] = −ae1 + (2a − 1)e3, [e5, e3] = ae3 − e1, (a 6= 1),
(vii) g5,7 : [e1, e5] = a(e1 − e3), [e3, e5] = e1 + f e2 − e3, [e4, e5] = (1 − a) (e2 − e4) ,
[e5, e5] = ǫe2, [e5, e1] = −ae1 + (d + f ) e2 + ae3, [e5, e3] = −e1 + de2 + e3,
[e5, e4] = (a − 1) (e2 − e4) , (ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0; a 6= 1),
(viii) g5,8 : [e1, e5] = ce2, [e3, e5] = e1 − e3, [e4, e5] = e2 − e4, [e5, e5] = ǫe2,
[e5, e1] = (c + d) e2, [e5, e3] = −e1 + (d + 2c) e2 + e3, [e5, e4] = e4 − e2, (c 6= 0, ǫ = 0, 1).
Proof. One applies the change of basis transformations keeping the nilradical L 4 given
in (4.0.1) unchanged.
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
19
(1) We have the right (a derivation) and the left (not a derivation) multiplication
a
0
operators restricted to the nilradical given below:
0
0
0
0
0
0
A2,1 2b a2,3
b−a 0
0
0
0
···
···
···
···
···
3a+b ···
0
0
0
0
0
2a+b
0
0
0
0
b−a
a+b
a5,3
0
0
0
b
0
0
0
0
0
0
·
a5,3
Ren+1 =
0
...
·
...
0
...
0
0
0
·
...
·
...
0 an−2,3 an−3,3
0 an−1,3 an−2,3
0
an,3 an−1,3
0
0
−a 0
b2,1 0
a−b 0
0
0 −a5,3
0
a2,3
−b
0
0
0
0
·
a+b
0
−a−b
−a5,3
·
...
0
...
0
...
0
0
0
0 −an−2,3 −an−3,3
0 −an−1,3 −an−2,3
0 −an,3 −an−1,3
0
...
···
···
···
0
0
0
0
0
0
...
...
···
···
···
0
0
0
0
0
...
···
···
···
0
−2a−b
−3a−b
...
...
a5,3
...
...
...
0 (n−5)a+b
·
·
a5,3
·
0
a5,3
(n−4)a+b
0
···
···
···
···
···
···
0
0
0
0
0
0
...
...
...
...
(5−n)a−b
...
0
0
0
0
0
0
0
...
...
0
,
0
0
0
0
0
0
...
...
0
0
(n−3)a+b
0
0
0
0
0
0
...
...
0
0
0
0
0
0
0
0
...
...
0
0
.
Len+1 =
··· −a5,3
···
···
1 = e1, e′
·
· −a5,3
0
(4−n)a−b
·
−a5,3
(3−n)a−b
2 = e2, e′
i = ei− ak−i+3,3
• We apply the transformation e′
(k−i)a ek, (3 ≤ i ≤ n−
2, i+2 ≤ k ≤ n, n ≥ 5), e′
j = ej, (n−1 ≤ j ≤ n+1), where k is fixed, renaming
all the affected entries back. This transformation removes a5,3, a6,3, ..., an,3 in
Ren+1 and −a5,3,−a6,3, ...,−an,3 in Len+1. Besides it introduces the entries in
the (5, 1)st, (6, 1)st, ..., (n, 1)st positions in Ren+1 and Len+1, which we set to
be a5,1, a6,1, ..., an,1 and −a5,1,−a6,1, ...,−an,1, respectively.
(I) Suppose b 6= a
2 .
• The transformation e′
2 = e2, e′
a2,3
b e2, e′
A2,1 and b2,1 from the (2, 1)st positions in Ren+1 and Len+1, respectively. It
b
(cid:17) e2, e′
a−2b (cid:16)A2,1 + (b−a)a2,3
n+1 = en+1−a5,1e2 +Pn−1
1 = e1 + 1
i = ei, (4 ≤ i ≤ n, n ≥ 4), e′
3 = e3 −
k=4 ak+1,1ek removes
also removes a2,3 from the (2, 3)rd positions in Ren+1 and Len+1 as well as ak+1,1
in Ren+1 and −ak+1,1 in Len+1 from the entries in the (k + 1, 1)st positions,
where (4 ≤ k ≤ n − 1).
Remark 5.1.1. If n = 4, then a5,1 = 0 and the same is in case (II).
• Then we scale a to unity applying the transformation e′
i = ei, (1 ≤ i ≤ n, n ≥
a by b, we obtain a continuous family of Leibniz
a . Renaming b
n+1 = en+1
4), e′
20
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(5.1.2)
algebras:
[e1, en+1] = e1 + (b − 1)e3, [e2, en+1] = 2be2, [e3, en+1] = be3,
[e4, en+1] = (b − 1)e2 + (b + 1)e4, [ei, en+1] = (b + i − 3) ei,
[en+1, e1] = −e1 − (b − 1)e3, [en+1, e3] = −be3, [en+1, e4] = (b + 1)e2−
(b + 1)e4, [en+1, ei] = (3 − i − b) ei, (5 ≤ i ≤ n),
(b 6= 0, b 6=
, b 6= 3 − n).
1
2
(II) Suppose b := a
• The transformation e′
2 . We have that A2,1 = a2,3.
1 = e1− b2,1+a2,3
e2, e′
a
i ≤ n, n ≥ 4), e′
(2, 1)st, (2, 3)rd positions in Ren+1 and from the (2, 3)rd position in Len+1. It
n+1 = en+1 − a5,1e2 + Pn−1
2 = e2, e′
3 = e3− 2a2,3
i = ei, (4 ≤
k=4 ak+1,1ek removes a2,3 from the
a e2, e′
also removes b2,1 from the (2, 1)st position in Len+1 as well as ak+1,1 in Ren+1
and −ak+1,1 in Len+1 from the entries in the (k + 1, 1)st positions, where
(4 ≤ k ≤ n − 1).
• To scale a to unity, we apply the transformation e′
i = ei, (1 ≤ i ≤ n, n ≥
and obtain a limiting case of (5.1.2) with b = 1
2 given below:
e2
+
2
, [e2, en+1] = e2, [e3, en+1] =
, [e4, en+1] = −
3e4
2
e3
2
,
2(cid:19) ei, [en+1, e1] = −e1 +
(e2 − e4) , [en+1, ei] = (cid:18)5
, [en+1, e3] = −
2 − i(cid:19) ei, (5 ≤ i ≤ n).
e3
2
3
2
e3
2
,
a
4), e′
n+1 = en+1
e3
[e1, en+1] = e1 −
2
5
[ei, en+1] = (cid:18)i −
[en+1, e4] =
(2) We have the right (a derivation) and the left (not a derivation) multiplication
a
operators restricted to the nilradical are as follows:
···
···
···
···
···
(6−n)a ···
2(3−n)a
(4−n)a
(3−n)a
(5−n)a
(2−n)a
(2−n)a
0
0
0
0
0
0
0
0
0
0
0
0
0
A2,1
0
0
0
a2,3
a5,3
0
0
0
0
0
0
0
·
a5,3
Ren+1
0
...
0
0
0
0
...
0
0
0
0
0
−a
b2,1
·
...
·
...
an−2,3 an−3,3
an−1,3 an−2,3
an−1,3
0
an,3
0
a2,3
(4−n)a
(n−2)a 0 (n−3)a
0
0
0
0
0
0
0 −a5,3
0
·
(n−4)a
−a5,3
0
(n−5)a
...
···
···
···
0
0
0
0
0
...
···
···
···
0
0
0
0
0
(n−6)a
...
...
a5,3
·
·
···
···
···
···
···
···
...
...
−a5,3
Len+1 =
0
...
0
0
0
0
...
·
...
·
...
0 −an−2,3 −an−3,3
0 −an−1,3 −an−2,3
0 −an,3 −an−1,3
0
...
···
···
···
.
...
...
...
...
...
...
...
· −a5,3
·
0
2a
0
0 0
a 0
· −a5,3 0 0
0
...
···
···
···
0
0
0
0
0
0
0 0
0 0
0 0
0 0
0 0
0 0
...
...
...
...
...
...
...
a5,3
·
0 −2a 0 0
0 −a 0
a5,3 0 0
0 0
0 0
0 0
0 0
0 0
0 0
0
0
0
0
0
0
,
• We apply the transformation e′
2, i+2 ≤ k ≤ n, n ≥ 5), e′
1 = e1, e′
2 = e2, e′
(k−i)a ek, (3 ≤ i ≤ n−
j = ej, (n−1 ≤ j ≤ n+1), where k is fixed, renaming
i = ei− ak−i+3,3
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
21
all the affected entries back. This transformation removes a5,3, a6,3, ..., an,3 in
Ren+1 and −a5,3,−a6,3, ...,−an,3 in Len+1. Besides it introduces the entries in
the (5, 1)st, (6, 1)st, ..., (n, 1)st positions in Ren+1 and Len+1, which we set to
be a5,1, a6,1, ..., an,1 and −a5,1,−a6,1, ...,−an,1, respectively.
n−3 (cid:17) e2, e′
(2n−5)a (cid:16)A2,1 + (n−2)a2,3
a2,3
(n−3)a e2, e′
removes A2,1 and b2,1 from the (2, 1)st positions in Ren+1 and Len+1, respec-
n+1 = en+1 − a5,1e2 + Pn−1
i = ei, (4 ≤ i ≤ n, n ≥ 4), e′
• The transformation e′
1 = e1 + 1
k=4 ak+1,1ek
2 = e2, e′
3 = e3 +
It also removes a2,3 from the (2, 3)rd positions in Ren+1 and Len+1;
tively.
ak+1,1 and −ak+1,1 from the entries in the (k + 1, 1)st positions in Ren+1 and
Len+1, respectively, where (4 ≤ k ≤ n − 1). (See Remark 5.1.1).
• To scale a to unity, we apply the transformation e′
i = ei, (1 ≤ i ≤ n), e′
n+1 =
en+1
a
renaming the coefficient an,n+1
a2
in front of en in [en+1, en+1] back by an,n+1.
We obtain a Leibniz algebra:
[e1, en+1] = e1 + (2 − n)e3, [e2, en+1] = 2(3 − n)e2, [e3, en+1] = (3 − n)e3,
[e4, en+1] = (2 − n)e2 + (4 − n)e4, [ei, en+1] = (i − n)ei, [en+1, en+1] = an,n+1en,
[en+1, e1] = −e1 + (n − 2)e3, [en+1, e3] = (n − 3)e3, [en+1, e4] = (4 − n) (e2 − e4) ,
[en+1, ei] = (n − i)ei, (5 ≤ i ≤ n).
If an,n+1 = 0, then we have a limiting case of (5.1.2) with b = 3− n. If an,n+1 6= 0,
then an,n+1 = reiφ and we apply the transformation e′
n−2 ej, (1 ≤ j ≤
n+1 = en+1 to scale an,n+1 to 1. We have the
j = (cid:16)reiφ(cid:17)
2), e′
k−2
j
n−2 ek, (3 ≤ k ≤ n), e′
k = (cid:16)reiφ(cid:17)
algebra gn+1,2 given below:
[e1, en+1] = e1 + (2 − n)e3, [e2, en+1] = 2(3 − n)e2, [e3, en+1] = (3 − n)e3,
[e4, en+1] = (2 − n)e2 + (4 − n)e4, [ei, en+1] = (i − n)ei, [en+1, en+1] = en,
[en+1, e1] = −e1 + (n − 2)e3, [en+1, e3] = (n − 3)e3, [en+1, e4] = (4 − n) (e2 − e4) ,
[en+1, ei] = (n − i)ei, (5 ≤ i ≤ n).
(3) We have the right (a derivation) and the left (not a derivation) multiplication
operators restricted to the nilradical are as follows:
Ren+1 =
0
0
0
a2,1 2b a2,3
b
0
0
0
0
0
0
0
b
0
a5,3
·
0
...
·
...
0
...
0
0
0
0
b
0
b
0
·
...
a5,3
0
0
0
0
b
0
0
0
0
0
0
b
···
···
···
···
···
···
... 0
...
...
...
0 an−2,3 an−3,3 ···
0 an−1,3 an−2,3 ···
0
an,3 an−1,3 ···
··· a5,3
···
···
·
·
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
...
...
...
...
...
...
b
0
0 0
b 0
a5,3 0 b
,
...
0
a5,3
·
22
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Len+1 =
0
0
0
a2,3 0
−b 0
0
0
0 −a5,3
0
0
0
a2,3
−b
0
·
0
b
0
−b
0
−a5,3
0
0
0
0
0
0
0
0
−b
0
0 −b
···
···
···
···
···
···
... 0
...
...
...
...
·
...
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0 −an,3 −an−1,3 ···
··· −a5,3
···
···
·
· −a5,3
0
...
·
...
0
...
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
...
...
...
.
0
0
−b
0
0 −b 0
· −a5,3 0 −b
2 = e2, e′
e2, e′
• Applying the transformation e′
i =
ei, (4 ≤ i ≤ n + 1), we remove a2,1 from the (2, 1)st position in Ren+1 and a2,3
from the (2, 1)st, the (2, 3)rd positions in Len+1 and from the (2, 3)rd position
1 = e1− a2,1+a2,3
3 = e3− a2,3
b e2, e′
2b
in Ren+1 keeping other entries unchanged.
• To scale b to unity, we apply the transformation e′
i = ei, (1 ≤ i ≤ n), e′
n+1 =
en+1
b
. Then we rename a5,3
b , a6,3
b , ..., an,3
b by a5,3, a6,3, ..., an,3, respectively. We
obtain a Leibniz algebra
[e1, en+1] = e3, [e2, en+1] = 2e2, [e4, en+1] = e2, [ei, en+1] = ei +
n
n
Xk=i+2
ak−i+3,3ek,
[en+1, e1] = −e3, [en+1, e4] = e2, [en+1, ei] = −ei −
ak−i+3,3ek, (3 ≤ i ≤ n).
If a5,3 6= 0, (n ≥ 5), then a5,3 = reiφ and applying the transformation e′
(cid:16)reiφ(cid:17)
2 ek, (3 ≤ k ≤ n), e′
2 ej, (1 ≤ j ≤ 2), e′
k = (cid:16)reiφ(cid:17)
scale a5,3 to 1. We also rename all the affected entries back and then we
n+1 = en+1, we
j =
Xk=i+2
k−2
j
rename a6,3, ..., an,3 by b1, ..., bn−5, respectively. We combine with the case
when a5,3 = 0 and obtain a Leibniz algebra gn+1,3 given below:
n
[e1, en+1] = e3, [e2, en+1] = 2e2, [e4, en+1] = e2, [ei, en+1] = ei + ǫei+2 +
bk−i−2ek,
Xk=i+3
bk−i−2ek,
[en+1, e1] = −e3, [en+1, e4] = e2, [en+1, ei] = −ei − ǫei+2 −
(ǫ = 0, 1, 3 ≤ i ≤ n).
Remark 5.1.2. If n = 4, then ǫ = 0.
n
Xk=i+3
(4) We have the right (a derivation) and the left (not a derivation) multiplication
operators restricted to the nilradical are as follows:
Ren+1 =
a
0
0
a2,3−b2,1+b2,3 0 a2,3
−a
0
0
0
0
0
0
0
0 a5,3
0
·
0
−a
0
a
0
a5,3
·
...
0
0
0
0
2a
0
···
0
···
0
···
0
···
0
0
···
3a ···
... 0
...
...
...
0
...
·
...
0
...
0
0
0
0 an−2,3 an−3,3 ···
0 an−1,3 an−2,3 ···
0 an,3 an−1,3 ···
··· a5,3
···
···
·
·
0
0
0
0
0
0
...
...
(n−5)a
0
a5,3
...
0
a5,3
·
0
0
0
0
0
0
...
...
0
(n−4)a
0
0
0
0
0
0
...
...
0
0
0
(n−3)a
,
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
23
Len+1 =
0
b2,3
−a 0
b2,1 0
a 0
0 0
0 0 −a5,3
0 0
0
0
·
0
a
0
−a
0
0
0
0
0
0
0
0
0
−2a 0
−a5,3
0 −3a
···
···
···
···
···
···
0 0
...
...
·
...
·
...
0 0 −an−2,3 −an−3,3
0 0 −an−1,3 −an−2,3
0 0 −an,3 −an−1,3
...
···
···
···
...
...
0
...
··· −a5,3
···
···
·
0
0
0
0
0
0
...
...
(5−n)a
...
0
0
0
0
0
0
0
...
...
0
0
0
0
0
0
0
...
...
0
0
.
· −a5,3
0
(4−n)a
·
−a5,3
0
(3−n)a
2 = e2, e′
1 = e1, e′
i = ei− ak−i+3,3
• We apply the transformation e′
(k−i)a ek, (3 ≤ i ≤ n−
2, i+2 ≤ k ≤ n, n ≥ 5), e′
j = ej, (n−1 ≤ j ≤ n+1), where k is fixed, renaming
all the affected entries back. This transformation removes a5,3, a6,3, ..., an,3 in
Ren+1 and −a5,3,−a6,3, ...,−an,3 in Len+1. Besides it introduces the entries in
the (5, 1)st, (6, 1)st, ..., (n, 1)st positions in Ren+1 and Len+1, which we set to
be a5,1, a6,1, ..., an,1 and −a5,1,−a6,1, ...,−an,1, respectively.
• The transformation e′
n+1 =
en+1 + Pn−1
k=4 ak+1,1ek removes a2,3 − b2,1 + b2,3 from the (2, 1)st position in
Ren+1. It changes the entry in the (2, 1)st position in Len+1 to a2,3 + b2,3. It
also removes ak+1,1 and −ak+1,1 from the entries in the (k + 1, 1)st positions,
where (4 ≤ k ≤ n − 1) in Ren+1 and Len+1, respectively.
i = ei, (2 ≤ i ≤ n, n ≥ 4), e′
1 = e1+ a2,3−b2,1+b2,3
e2, e′
a
• We assign a2,3 := d and b2,3 := f and then we scale a to unity applying
a , f
by d, f and a2,n+1, respectively, we obtain a right and left Leibniz
i = ei, (1 ≤ i ≤ n, n ≥ 4), e′
the transformation e′
a . Renaming d
n+1 = en+1
and a2,n+1
a
a2
algebra:
[e1, en+1] = e1 − e3, [e3, en+1] = de2, [e4, en+1] = e4 − e2, [ei, en+1] = (i − 3)ei,
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −e1 + (d + f ) e2 + e3, [en+1, e3] = f e2,
[en+1, e4] = e2 − e4, [en+1, ei] = (3 − i)ei, (5 ≤ i ≤ n),
which is a limiting case of (5.1.2) with b = 0, when d = f = a2,n+1 = 0.
Altogether (5.1.2) and all its limiting cases after replacing b with a give us a
Leibniz algebra gn+1,1 given below:
[e1, en+1] = e1 + (a − 1)e3, [e2, en+1] = 2ae2, [e3, en+1] = ae3,
[e4, en+1] = (a − 1)e2 + (a + 1)e4, [ei, en+1] = (a + i − 3) ei, [en+1, e1] = −e1−
(a − 1)e3, [en+1, e3] = −ae3, [en+1, e4] = (a + 1) (e2 − e4) , [en+1, ei] = (3 − i − a) ei,
(5 ≤ i ≤ n).
24
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
It remains to consider a continuous family of Leibniz algebras given below
and scale any nonzero entries as much as possible.
[e1, en+1] = e1 − e3, [e3, en+1] = de2, [e4, en+1] = e4 − e2, [ei, en+1] = (i − 3)ei,
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −e1 + (d + f ) e2 + e3, [en+1, e3] = f e2,
2,n+1 + d2 + f 2 6= 0, 5 ≤ i ≤ n),
[en+1, e4] = e2 − e4, [en+1, ei] = (3 − i)ei, (a2
j
If a2,n+1 6= 0, then a2,n+1 = reiφ and applying the transformation e′
(cid:16)reiφ(cid:17)
2 ek, (3 ≤ k ≤ n, n ≥ 4), e′
2 ej, (1 ≤ j ≤ 2), e′
k = (cid:16)reiφ(cid:17)
scale it to 1. We also rename all the affected entries back. Then we combine
n+1 = en+1, we
j =
k−2
with the case when a2,n+1 = 0 and obtain a right and left Leibniz algebra
gn+1,4 given below:
[e1, en+1] = e1 − e3, [e3, en+1] = de2, [e4, en+1] = e4 − e2, [ei, en+1] = (i − 3)ei,
[en+1, en+1] = ǫe2, [en+1, e1] = −e1 + (d + f ) e2 + e3, [en+1, e3] = f e2,
[en+1, e4] = e2 − e4, [en+1, ei] = (3 − i)ei,
(5 ≤ i ≤ n; ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0).
(5) Applying the transformation e′
(c−2a)a2,3+c·b2,3
2a(b+c)
e2, e′
4 = e4, e′
5 = e5
1 = e1 + (b+2c−2a)a2,3+(b+2c)b2,3
c and renaming a
c and b
2a(b+c)
e2, e′
2 = e2, e′
3 = e3 +
c by a and b, respectively,
we obtain a continuous family of Leibniz algebras given below:
(5.1.3)
[e1, e5] = ae1 + (b − a + 1)e3, [e2, e5] = 2(b + 1)e2, [e3, e5] = e1 + be3,
[e4, e5] = (b − a + 2)e2 + (a + b)e4, [e5, e1] = −ae1 + (a − b − 1)e3,
[e5, e3] = −e1 − be3, [e5, e4] = (a + b)(e2 − e4), (b 6= −a, a 6= 0, b 6= −1).
(6) We apply the transformation e′
(2a−c)a2,3−c·b2,3
2a(a−c)
e2, e′
4 = e4, e′
5 = e5
1 = e1 + (3a−2c)a2,3+(a−2c)b2,3
c and rename a
c and a4,5
2a(a−c)
e2, e′
2 = e2, e′
3 = e3 +
c2 by a and a4,5, respectively,
to obtain a Leibniz algebra given below:
(5.1.4)
[e1, e5] = ae1 + (1 − 2a)e3, [e2, e5] = 2(1 − a)e2, [e3, e5] = e1 − ae3,
[e4, e5] = 2(1 − a)e2, [e5, e5] = a4,5e4, [e5, e1] = −ae1 + (2a − 1)e3,
[e5, e3] = ae3 − e1, (a 6= 0, a 6= 1),
which is a limiting case of (5.1.3) with b := −a if a4,5 = 0. If a4,5 6= 0, then
a4,5 = reiφ. To scale a4,5 to 1, we apply the transformation e′
2 =
1 = √rei φ
2 e1, e′
reiφe2, e′
3 = √rei φ
2 e3, e′
4 = reiφe4, e′
5 = e5 and obtain a Leibniz algebra given
(5.1.5)
below:
[e1, e5] = ae1 + (1 − 2a)e3, [e2, e5] = 2(1 − a)e2, [e3, e5] = e1 − ae3,
[e4, e5] = 2(1 − a)e2, [e5, e5] = e4, [e5, e1] = −ae1 + (2a − 1)e3,
[e5, e3] = −e1 + ae3, (a 6= 0, a 6= 1).
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
25
(7) We apply the transformation e′
assign d := a·b2,3+c(a2,3−b2,1+b2,3)
a
family of Leibniz algebras:
1 = e1 + a2,3−b2,1+b2,3
e2, e′
, f := a·a2,3−c(a2,3−b2,1+b2,3)
a
i = ei, (2 ≤ i ≤ 5) and
to obtain a continuous
a
[e1, e5] = a(e1 − e3), [e3, e5] = ce1 + f e2 − ce3, [e4, e5] = (c − a)(e2 − e4), [e5, e5] = a2,5e2,
[e5, e1] = −ae1 + (d + f ) e2 + ae3, [e5, e3] = −ce1 + de2 + ce3, [e5, e4] = (a − c)(e2 − e4).
c renaming
Then we continue with the transformation e′
5 = e5
i = ei, (1 ≤ i ≤ 4), e′
(5.1.6)
(5.1.7)
d
c , f
c and a2,5
c , a
c2 by d, f, a and a2,5, respectively, to obtain a Leibniz algebra:
[e1, e5] = a(e1 − e3), [e3, e5] = e1 + f e2 − e3, [e4, e5] = (1 − a) (e2 − e4) ,
[e5, e5] = a2,5e2, [e5, e1] = −ae1 + (d + f ) e2 + ae3, [e5, e3] = −e1 + de2 + e3,
[e5, e4] = (a − 1) (e2 − e4) , (a 6= 0, a 6= 1).
If d = f = a2,5 = 0, then we have a limiting case of (5.1.3) with b = −1. If
a2,5 6= 0, then we apply the same transformation as we did to scale a4,5 to 1. We
also rename all the affected entries back. Then we combine with the case when
a2,5 = 0 and obtain a Leibniz algebra given below:
[e1, e5] = a(e1 − e3), [e3, e5] = e1 + f e2 − e3, [e4, e5] = (1 − a) (e2 − e4) ,
[e5, e5] = ǫe2, [e5, e1] = −ae1 + (d + f ) e2 + ae3, [e5, e3] = −e1 + de2 + e3,
[e5, e4] = (a − 1) (e2 − e4) , (ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0),
(a 6= 0, a 6= 1).
Remark 5.1.3. We notice by applying the transformation e′
1 = e1 + f e2, e′
i =
ei, (2 ≤ i ≤ 5) that this algebra (5.1.7) is isomorphic to
[e1, e5] = ae1 − af e2 − ae3, [e3, e5] = e1 − e3, [e4, e5] = (1 − a) (e2 − e4) , [e5, e5] = ǫe2,
[e5, e1] = −ae1 + (af + d + f ) e2 + ae3, [e5, e3] = −e1 + (d + f )e2 + e3,
[e5, e4] = (a − 1) (e2 − e4) , (ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0), (a 6= 0, a 6= 1).
3 = e3 − b2,1
(8) Applying the transformation e′
c2 back by a4,5, we obtain a Leibniz algebra:
c and renaming a4,5
1 = e1 + a2,3−2b2,1
2 = e2, e′
c e2, e′
5 = e5
e2, e′
e4, e′
c
4 =
[e1, e5] = e3, [e2, e5] = 2e2, [e3, e5] = e1, [e4, e5] = 2e2, [e5, e5] = a4,5e4, [e5, e1] = −e3,
[e5, e3] = −e1,
which is a limiting case of (5.1.4) with a = 0 and at the same time a limiting
case of (5.1.3) with b := −a if a4,5 = 0. Therefore we only consider the case when
a4,5 6= 0. One applies the transformation below (5.1.4) to scale a4,5 to 1 and we
have a limiting case of (5.1.5) with a = 0. Altogether (5.1.5) and all its limiting
26
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
cases give us the algebra g5,6 in full generality:
[e1, e5] = ae1 + (1 − 2a)e3, [e2, e5] = 2(1 − a)e2, [e3, e5] = e1 − ae3, [e4, e5] = 2(1 − a)e2,
[e5, e5] = e4, [e5, e1] = −ae1 + (2a − 1)e3, [e5, e3] = −e1 + ae3,
where a 6= 1, otherwise nilpotent.
(9) If a = 0, b 6= 0, b 6= −c, c 6= 0, (n = 4), then we apply the transformation e′
e1 + (b+2c)b2,3−(3b+2c)a2,3
(b+c)2
e2, e′
2 = e2, e′
3 = e3 − B2,1
b+c e2, e′
4 = e4, e′
5 = e5
by b and obtain a Leibniz algebra:
[e1, e5] = (b + 1)e3, [e2, e5] = 2(b + 1)e2, [e3, e5] = e1 + be3, [e4, e5] = (b + 2)e2 + be4,
[e5, e1] = (−b − 1)e3, [e5, e3] = −e1 − be3, [e5, e4] = b(e2 − e4), (b 6= 0, b 6= −1),
which is a limiting case of (5.1.3) with a = 0. Altogether (5.1.3) and all its limiting
cases give us the algebra g5,5 in full generality:
1 =
c . We rename b
c
[e1, e5] = ae1 + (b − a + 1)e3, [e2, e5] = 2(b + 1)e2, [e3, e5] = e1 + be3,
[e4, e5] = (b − a + 2)e2 + (a + b)e4, [e5, e1] = −ae1 + (a − b − 1)e3, [e5, e3] = −e1 − be3,
[e5, e4] = (a + b)(e2 − e4),
where if b = −1, then a 6= 1, otherwise nilpotent.
c e2, e′
(10) We apply the transformation e′
1 = e1 + a2,3
c , b2,3
c and a2,5
c2 by a2,1, a2,3, b2,3 and a2,5, respectively, to obtain a
i = ei, (2 ≤ i ≤ 4), e′
5 = e5
c and
rename a2,1
c , a2,3
Leibniz algebra:
[e1, e5] = a2,1e2, [e3, e5] = e1 − e3, [e4, e5] = e2 − e4, [e5, e5] = a2,5e2,
[e5, e1] = (a2,3 + b2,3 − a2,1) e2, [e5, e3] = −e1 + (a2,3 + b2,3) e2 + e3, [e5, e4] = e4 − e2.
We scale a2,5 to 1 if nonzero. Combining with the case when a2,5 = 0 and renaming
all the affected entries back, we have the algebra:
[e1, e5] = a2,1e2, [e3, e5] = e1 − e3, [e4, e5] = e2 − e4, [e5, e5] = ǫe2,
[e5, e1] = (a2,3 + b2,3 − a2,1) e2, [e5, e3] = −e1 + (a2,3 + b2,3) e2 + e3, [e5, e4] = e4 − e2,
(ǫ = 0, 1).
If a2,1 = 0, then we obtain a limiting case of (5.1.6) and (5.1.7) with a = 0
according to Remark 5.1.3. Altogether (5.1.7) and all its limiting cases, give us
the algebra g5,7 in full generality:
[e1, e5] = a(e1 − e3), [e3, e5] = e1 + f e2 − e3, [e4, e5] = (1 − a) (e2 − e4) , [e5, e5] = ǫe2,
[e5, e1] = −ae1 + (d + f ) e2 + ae3, [e5, e3] = −e1 + de2 + e3, [e5, e4] = (a − 1) (e2 − e4) ,
(ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0),
where a 6= 1, otherwise nilpotent.
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
27
If c := a2,1 6= 0 and we assign d := a2,3 + b2,3 − 2a2,1, then we obtain a Leibniz
algebra g5,8 given below:
[e1, e5] = ce2, [e3, e5] = e1 − e3, [e4, e5] = e2 − e4, [e5, e5] = ǫe2, [e5, e1] = (c + d) e2,
[e5, e3] = −e1 + (d + 2c) e2 + e3, [e5, e4] = e4 − e2, (c 6= 0, ǫ = 0, 1).
(cid:3)
5.1.2. Codimension two and three solvable extensions of L 4
The non-zero inner derivations of L 4, (n ≥ 4) are given by
, Rei = −Ei+1,1 =
(4 ≤ i ≤ n−1),
0 0 0 ··· 0
0 0 0 ··· 0
0 0 0 ··· 0
0 0 0 ··· 0
−1 0 0 ··· 0
...
...
...
...
...
...
· 0 0 ··· 0
...
...
Re1 =
0 0 0 0 ··· 0 0 0
1 0 0 0 ··· 0 0 0
0 0 0 0 ··· 0 0 0
0 0 1 0 ··· 0 0 0
0 0 0 1 ··· 0 0 0
...
...
...
...
...
...
...
...
0 0 0 0 ··· 1 0 0
0 0 0 0 ··· 0 1 0
, Re3 =
0 0 0 0 ··· 0
2 0 1 0 ··· 0
0 0 0 0 ··· 0
−1 0 0 0 ··· 0
0 0 0 0 ··· 0
...
...
...
...
0 0 0 0 ··· 0
...
where Ei+1,1 is the n× n matrix that has 1 in the (i + 1, 1)st position and all other entries
are zero.
Remark 5.1.4. If we have four or more outer derivations, then they are "nil-dependent"(see
Section 2.). Therefore the solvable algebras we are constructing might be of codimension
at most three. (As Section 5.1.2.3 shows, we have at most two outer derivations.)
General approach to find right solvable Leibniz algebras with a codimension two
nilradical L 4.2
(i) We consider R[er,es] = [Res, Rer] = Res
Res, (n + 1 ≤ r ≤ n + 2, 1 ≤ s ≤
k=3 ckRek, because e2 and en are in the center
n + 2) and compare with c1Re1 +Pn−1
of L 4, (n ≥ 4) defined in (4.0.1) to find all the unknown commutators.
(ii) We write down [er, es], (n + 1 ≤ r ≤ n + 2, 1 ≤ s ≤ n + 2) including a linear
combination of e2 and en as well. We add the brackets of the nilradical L 4 and
outer derivations Ren+1 and Ren+2.
Rer − Rer
(iii) One satisfies the right Leibniz identity: [[er, es], et] = [[er, et], es] + [er, [es, et]] or,
equivalently, Ret ([er, es]) = [Ret(er), es] + [er, Ret(es)], (1 ≤ r, s, t ≤ n + 2) for all
the brackets obtained in step (ii).
(iv) Then we carry out the technique of "absorption" (see Section 3.2) to remove some
parameters to simplify the algebra.
(v) We apply the change of basis transformations without affecting the nilradical to
remove as many parameters as possible.
5.1.2.1. Codimension two solvable extensions of L 4, (n = 4)
There are the following cases based on the conditions involving parameters a, b and c:
(1) b 6= −a, a 6= 0, b 6= −c,
(2) b := −a, a 6= 0, a 6= c,
(3) b := −c, a 6= c, a 6= 0,
(4) a = 0, b 6= 0, b 6= −c.
2When we work with left Leibniz algebras, we first change the right multiplication operator to the left
everywhere and the right Leibniz identity to the left Leibniz identity in step (iii). We also interchange s
and r in the very left in steps (i) and (ii) as well.
28
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
1
b
1
1
a
0
a1
b1
c1
a2
b2
c2
=
=
(1) (a) One could set
Therefore the vector space of outer derivations as 4 × 4 matrices is as follows:
Re5 =
and
2 b2,3 2a a2,3 a − 1
a + 1
1
−b · α2,3 − (1 + b)β2,3 2(b + 1) α2,3
b
0
a − 1
1
2 a2,3 + 1−2a
a
0
3−2a
0
0
0
0
b
0
0
0
0
0
0
0
1
0
,
, (a 6= −1, a 6= 0, b 6= −1).
(i) Considering R[e5,e6], we obtain that a := 1 and α2,3 := (cid:16)b + 3
it follows that β2,3 := b2,3
2 . Since b 6= −1,
.
0
b + 1
2(cid:17) a2,3 + b2,3
b + 1
Re6 =
2 −(cid:16)b + 1
1
2(cid:17) a2,3 and R[e5,e6] = 0. As a result,
Re5 =
0
0
2 a2,3 0
0
0
0
2
a2,3−b2,3
2
0
0
1
0
0
,
2(cid:17) a2,3 + b2,3
1
2
b
0
0
0
0
0
b + 1
0
b + 1
, (b 6= −1).
Re6 =
a2,3
2 −(cid:16)b + 1
2(cid:17) b2,3 2(b + 1) (cid:16)b + 3
1
b
0
Further, we find the following commutators:
(ii) We include a linear combination of e2 and e4:
( R[e5,e1] = −Re1, R[e5,e2] = 0, R[e5,e3] = −Re3, R[e5,ei] = 0, R[e6,e1] = −Re1 − bRe3,
R[e6,e2] = 0, R[e6,e3] = −Re1 − bRe3, R[e6,ei] = 0, (b 6= −1, 4 ≤ i ≤ 6).
[e5, e1] = −e1 + c2,1e2 + c4,1e4, [e5, e2] = c2,2e2 + c4,2e4, [e5, e3] = c2,3e2 − e3 + c4,3e4,
[e5, ei] = c2,ie2 + c4,ie4, [e6, e1] = −e1 + d2,1e2 − be3 + d4,1e4, [e6, e2] = d2,2e2 + d4,2e4,
[e6, e3] = −e1 + d2,3e2 − be3 + d4,3e4, [e6, ei] = d2,ie2 + d4,ie4, (b 6= −1, 4 ≤ i ≤ 6).
Besides we have the brackets from L 4 and from outer derivations Re5 and Re6 as well.
(iii) We satisfy the right Leibniz identity, which is shown in Table 2.
We obtain that Re5 and Re6 restricted to the nilradical do not change, but the remaining
brackets are as follows:
a2,3 − b2,3
1
2
1
b2,3
b2,3
e2, [e5, e3] = a2,3e2 − e3, [e5, e4] = 2 (e2 − e4) , [e5, e5] = c2,5e2,
2 ! e2 − be3,
[e5, e1] = −e1 +
[e5, e6] = c2,6e2 + ((b + 1)c2,5 − c2,6) e4, [e6, e1] = −e1 + (cid:18)b +
[e6, e3] = −e1 + (cid:18)b +
[e6, e5] = d2,5e2 + (c2,6 − (b + 1)c2,5) e4, [e6, e6] = (cid:16)(b + 1) (c2,6 + d2,5) − (b + 1)2c2,5(cid:17) e2,
(b 6= −1).
2(cid:19) a2,3 −
2 ! e2 − be3, [e6, e4] = (b + 1) (e2 − e4) ,
2(cid:19) a2,3 −
Altogether the nilradical L 4 (4.0.1), the outer derivations Re5 and Re6 written in the
bracket notation and the remaining brackets given above define a continuous family of
Leibniz algebras depending on the parameters.
(iv)&(v) We apply the following transformation: e′
6 = e6 − d2,5
e3 − a2,3e2, e′
1 = e1 + b2,3−a2,3
2 c2,5 − c2,6
5 = e5 − c2,5
2 = e2, e′
4 = e4, e′
2 e2, e′
e2, e′
3 =
2 (cid:17) e4 and obtain a
2 e2 + (cid:16) b+1
2
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
29
Table 2. Right Leibniz identities in case (1) (a) with a nilradical L 4, (n = 4).
Steps Ordered triple
Result
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
Re1 ([e5, e1])
Re1 ([e6, e1])
Re3 ([e5, e1])
Re3 ([e6, e1])
Re5 ([e5, e5])
Re6 ([e5, e6])
Re5 ([e5, e3])
Re6 ([e5, e3])
Re1 ([e5, e6])
Re3 ([e5, e6])
Re5 ([e5, e6])
Re5 ([e6, e3])
Re6 ([e6, e3])
Re6 ([e6, e5])
[e5, e2] = 0 =⇒ c2,2 = c4,2 = 0.
[e6, e2] = 0 =⇒ d2,2 = d4,2 = 0.
c2,4 := 2, c4,4 := −2 =⇒ [e5, e4] = 2 (e2 − e4) .
d2,4 := 1 + b, d4,4 := −1 − b =⇒ [e6, e4] = (b + 1) (e2 − e4) .
c4,5 = 0 =⇒ [e5, e5] = c2,5e2.
d4,6 = 0 =⇒ [e6, e6] = d2,6e2.
c2,3 := a2,3, c4,3 = 0 =⇒ [e5, e3] = a2,3e2 − e3.
c2,1 := a2,3−b2,3
d4,1 = 0 =⇒ [e6, e1] = −e1 + d2,1e2 − be3.
, c4,1 = 0 =⇒ [e5, e1] = −e1 + a2,3−b2,3
2
2
e2.
d4,3 = 0 =⇒ [e6, e3] = −e1 + d2,3e2 − be3.
d4,5
:= −c4,6 =⇒ c4,6
:= (b + 1)c2,5 − c2,6 =⇒ [e5, e6] = c2,6e2 +
((b + 1)c2,5 − c2,6) e4, [e6, e5] = d2,5e2 + (c2,6 − (b + 1)c2,5) e4
b2,3
2 =⇒ [e6, e3] = −e1 +(cid:16)(cid:0)b + 1
2 =⇒ [e6, e1] = −e1 +(cid:16)(cid:0)b + 1
2(cid:1) a2,3 −
2(cid:1) a2,3 − b2,3
d2,3 :=(cid:0)b + 1
d2,1 :=(cid:0)b + 1
(cid:0)(b + 1) (c2,6 + d2,5) − (b + 1)2c2,5(cid:1) e2.
(b + 1) (c2,6 + d2,5) − (b + 1)2c2,5
d2,6
:=
b2,3
2(cid:1) a2,3 −
2(cid:1) a2,3 − b2,3
=⇒
2 (cid:17) e2 − be3.
2 (cid:17) e2 − be3.
[e6, e6]
=
Leibniz algebra g6,2 given below:
Remark 5.1.5. We notice that if b = −1, then the outer derivation Re6 is nilpotent.
(1) (b) We set
vector space of outer derivations as 4 × 4 matrices is as follows:
[e1, e5] = e1, [e2, e5] = 2e2, [e3, e5] = e3, [e4, e5] = 2e4, [e5, e1] = −e1, [e5, e3] = −e3,
[e5, e4] = 2 (e2 − e4) , [e1, e6] = e1 + be3, [e2, e6] = 2(b + 1)e2, [e3, e6] = e1 + be3,
[e4, e6] = (b + 1) (e2 + e4) , [e6, e1] = −e1 − be3, [e6, e3] = −e1 − be3,
[e6, e4] = (b + 1) (e2 − e4) , (b 6= −1).
a2
and
=
b2
c2
− 1+3c
2 a2,3 − 3+3c
2 b2,3 2(c + 2) a2,3 2c + 1
=
, (c 6= −2, d 6= −1). Therefore the
a1
b1
c1
1
1
d
1
0
0
c
,
0
1
2
c
c + 1
Re5 =
Re6 =
a2,3 := (cid:16) 3
2 c + 2(cid:17) α2,3 + c
(cid:16) 3
2 α2,3 −(cid:16) 3
2 + c(cid:17) β2,3 − 1
2 α2,3 − c−2
2 α2,3 − 1+3d
d
0
1−3d
1
.
d
0
0
0
2
0
0
0
0
3
0
2 β2,3 2(d + 1) α2,3 2d
0
2
2(cid:17) b2,3 = 0.
2 β2,3,
2 c + 1
2 c + 3
2(cid:17) a2,3 −(cid:16) 3
2 β2,3 and R[e5,e6] = 0. As a result,
1
0
There are the following two cases:
(I) If c 6= −1, then b2,3 := − 3c+2
(i) Considering R[e5,e6], we obtain that d = 0 and we have the system of equations:
Re5 =
α2,3
2 − 2c+3
c + 1
2 β2,3 2(c + 2) (cid:16) 3c
2 + 2(cid:17) α2,3 + c
2 β2,3 2c + 1
,
0
0
0
c
2
0
1
0
0
0
3
30
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
1
α2,3−β2,3
2
0
0
Further, we find the following:
Re6 =
0
0
0
2 α2,3 0
0
0
0
2
1
0
, (c 6= −2).
( R[e5,e1] = −Re1 − (c + 1)Re3, R[e5,e2] = 0, R[e5,e3] = −cRe1 − 2Re3, R[e5,ei] = 0,
R[e6,e1] = −Re1, R[e6,e2] = 0, R[e6,e3] = −Re3, R[e6,ei] = 0, (c 6= −2, 4 ≤ i ≤ 6).
(II) If c = −1, then a2,3 := α2,3−β2,3
0
−1
Re5 =
2 α2,3−β2,3
2
2
0
0
0
, Re6 =
0
0
2 α2,3 0
0
0
2
0
and R[e5,e6] = 0. It follows that
and we have the following commutators:
0
−1
0
3
α2,3−β2,3
α2,3−β2,3
2
0
0
2
0
0
1
0
1
0
1
2
( R[e5,e1] = −Re1, R[e5,e2] = 0, R[e5,e3] = Re1 − 2Re3, R[e5,ei] = 0,
R[e6,e1] = −Re1, R[e6,e2] = 0, R[e6,e3] = −Re3, R[e6,ei] = 0, (4 ≤ i ≤ 6).
(ii) We combine cases (I) and (II) together, include a linear combination of e2 and e4,
and have the following brackets:
[e5, e1] = −e1 + a2,1e2 − (c + 1)e3 + a4,1e4, [e5, e2] = a2,2e2 + a4,2e4, [e5, e3] = −ce1+
a2,3e2 − 2e3 + a4,3e4, [e5, ei] = a2,ie2 + a4,ie4, [e6, e1] = −e1 + b2,1e2 + b4,1e4,
[e6, e2] = b2,2e2 + b4,2e4, [e6, e3] = b2,3e2 − e3 + b4,3e4, [e6, ei] = b2,ie2 + b4,ie4,
(c 6= −2, 4 ≤ i ≤ 6).
Besides we have the brackets from L 4 and from outer derivations Re5 and Re6 as well.
(iii) To satisfy the right Leibniz identity, we refer to the identities given in Table 2.
as much as possible. As a result, the identities we apply are the following: 1. − 6., 8.,
Re6([e5, e1]) = [Re6(e5), e1]+[e5, Re6(e1)], 9.−11., Re6 ([e6, e1]) = [Re6(e6), e1]+[e6, Re6(e1)],
13. and 14. We obtain that Re5 and Re6 restricted to the nilradical do not change, but
the remaining brackets are the following:
3
β2,3
2(cid:19) α2,3 −
+ 2(cid:19) α2,3 −
[e5, e1] = −e1 + (cid:18)c +
2 ! e2 − (c + 1)e3,
[e5, e3] = −ce1 +(cid:18)(cid:18) c
β2,3(cid:19) e2 − 2e3, [e5, e4] = 3 (e2 − e4) ,
α2,3 − β2,3
[e5, e5] = (c + 2) (a2,6 + a4,6) e2, [e5, e6] = a2,6e2 + a4,6e4, [e6, e1] = −e1 +
[e6, e3] = α2,3e2 − e3, [e6, e4] = 2 (e2 − e4) , [e6, e5] = ((c + 2)b2,6 + a4,6) e2 − a4,6e4,
[e6, e6] = b2,6e2, (c 6= −2).
c
2
2
2
e2,
Altogether the nilradical L 4 (4.0.1), the outer derivations Re5 and Re6 written in the
bracket notation and the remaining brackets given above define a continuous family of
Leibniz algebras depending on the parameters.
(iv)&(v) We apply the transformation: e′
6 = e6 − b2,6
e4, e′
5 = e5 − a2,6
[e1, e5] = e1 + (c + 1)e3, [e2, e5] = 2(c + 2)e2, [e3, e5] = ce1 + 2e3, [e4, e5] = (2c + 1)e2 + 3e4,
[e5, e1] = −e1 − (c + 1)e3, [e5, e3] = −ce1 − 2e3, [e5, e4] = 3 (e2 − e4) , [e1, e6] = e1,
[e2, e6] = 2e2, [e3, e6] = e3, [e4, e6] = 2e4, [e6, e1] = −e1, [e6, e3] = −e3, [e6, e4] = 2 (e2 − e4) ,
(c 6= −2).
1 = e1− α2,3−β2,3
2 e2 and obtain a Leibniz algebra g6,3:
3 = e3 − α2,3e2, e′
2 e2 − a4,6
2 = e2, e′
2 e4, e′
e2, e′
4 =
2
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
31
a1
b1
c1
a2
b2
c2
a
1
0
=
(1) (c) We set
vector space of outer derivations as 4 × 4 matrices is as follows:
b
Re5 =
0
0
2 α2,3 2 − b
0
0
0
b
(i) Considering R[e5,e6], we obtain that a := 1 and we have the system of equations:
and
0
2 a2,3 1 − a
0
a + 1
0
=
, Re6 =
, (a, b 6= 0, a 6= −1). Therefore the
(3b−3)α2,3+(b−3)β2,3
(3a−2)a2,3+(a−2)b2,3
1 − b
0
1 − a
b
0
1
1
0
0
0
0
1
0
0
0
2a
a
2b
.
( α2,3 := 3a2,3+b2,3
b2a2,3 + (b2 − 2b)b2,3 + (3 − 3b)α2,3 + (3 − b)β2,3 = 0.
,
2
There are the following two cases:
2
0
1
a2,3−b2,3
(I) If b 6= 3, then β2,3 := (2b−3)a2,3+(2b−1)b2,3
, Re6 =
Re5 =
0
0
2 a2,3 0
0
0
0
2
Further, we find the commutators:
2
0
0
1
0
b·a2,3+(b−2)b2,3
2
1 − b
0
and R[e5,e6] = 0. Consequently,
b
1
0
2 3a2,3+b2,3
0
0
2
0
0
0
2 − b
0
b
, (b 6= 0).
( R[e5,e1] = −Re1, R[e5,e2] = 0, R[e5,e3] = −Re3, R[e5,ei] = 0, R[e6,e1] = −bRe1 + (b − 1)Re3,
R[e6,e2] = 0, R[e6,e3] = −Re1, R[e6,ei] = 0, (b 6= 0, 4 ≤ i ≤ 6).
(II) If b := 3, then R[e5,e6] = 0 and Re5, Re6 are as follows:
0
2 3a2,3+b2,3
0
0
0
0
2 a2,3 0
0
0
0
2
0
−1
0
3
−2
0
3a2,3+b2,3
a2,3−b2,3
2
0
0
2
0
0
1
0
0
3
1
1
.
2
, Re6 =
Re5 =
We have the following commutators:
( R[e5,e1] = −Re1, R[e5,e2] = 0, R[e5,e3] = −Re3, R[e5,ei] = 0, R[e6,e1] = −3Re1 + 2Re3,
R[e6,e2] = 0, R[e6,e3] = −Re1, R[e6,ei] = 0, (4 ≤ i ≤ 6).
(ii) We combine cases (I) and (II) together and include a linear combination of e2 and
e4 :
[e5, e1] = −e1 + c2,1e2 + c4,1e4, [e5, e2] = c2,2e2 + c4,2e4, [e5, e3] = c2,3e2 − e3 + c4,3e4,
[e5, ei] = c2,ie2 + c4,ie4, [e6, e1] = −be1 + d2,1e2 + (b − 1)e3 + d4,1e4, [e6, e2] = d2,2e2 + d4,2e4,
[e6, e3] = −e1 + d2,3e2 + d4,3e4, [e6, ei] = d2,ie2 + d4,ie4, (b 6= 0, 4 ≤ i ≤ 6).
We also have the brackets from L 4 and from outer derivations Re5 and Re6 as well.
(iii) To satisfy the right Leibniz identity, we apply the same identities as in Table 2.
We have that Re5 and Re6 restricted to the nilradical do not change, but the remaining
brackets are the following:
a2,3 − b2,3
2
[e5, e1] = −e1 +
[e5, e6] = c2,6e2 + (c2,5 − c2,6) e4, [e6, e1] = −be1 + 2 − b
[e6, e3] = −e1 +
[e6, e6] = (d2,5 − c2,5 + c2,6) e2, (b 6= 0).
e2, [e5, e3] = a2,3e2 − e3, [e5, e4] = 2 (e2 − e4) , [e5, e5] = c2,5e2,
b2,3! e2 + (b − 1)e3,
e2, [e6, e4] = b (e2 − e4) , [e6, e5] = d2,5e2 + (c2,6 − c2,5) e4,
Altogether the nilradical L 4 (4.0.1), the outer derivations Re5 and Re6 written in the
bracket notation and the remaining brackets given above define a continuous family of
Leibniz algebras depending on the parameters.
a2,3 − b2,3
a2,3 −
b
2
2
2
32
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
1 = e1 + b2,3−a2,3
2
e2, e′
2 = e2, e′
(iv)&(v) We apply the transformation: e′
e4, e′
2 e2, e′
2 e2 + c2,5−c2,6
6 = e6 − d2,5
5 = e5 − c2,5
[e1, e5] = e1, [e2, e5] = 2e2, [e3, e5] = e3, [e4, e5] = 2e4, [e5, e1] = −e1, [e5, e3] = −e3,
[e5, e4] = 2 (e2 − e4) , [e1, e6] = be1 + (1 − b)e3, [e2, e6] = 2e2, [e3, e6] = e1,
[e4, e6] = (2 − b)e2 + be4, [e6, e1] = −be1 + (b − 1)e3, [e6, e3] = −e1, [e6, e4] = b (e2 − e4) ,
(b 6= 0).
e4 and obtain a Leibniz algebra g6,4:
(2) In this case we consider R[e5,e6] and compare with c1Re1 + c3Re3, where
2
3 = e3 − a2,3e2, e′
4 =
a
0
c
0
Re5 =
Re6 =
(5a−3c)a2,3+(3a−3c)b2,3
2a
c − 2a
0
α
(5α−3γ)α2,3+(3α−3γ)β2,3
2α
γ − 2α
0
2(c − a) a2,3 2(c − a)
,
0
0
0
0
0
−a
0
γ
−α
0
2(γ − α) α2,3 2(γ − α)
.
0
0
0
0
0
c ! and
(3) Considering R[e5,e6] and comparing with c1Re1 + c3Re3, where
We obtain that aγ − αc = 0, where a, α 6= 0, which is impossible, because a
γ ! are supposed to be linearly independent.
α
Re5 =
0 −γ
0
0
we have to have that aγ − αc = 0, where a, α 6= 0, which is impossible.
(4) Finally we consider R[e5,e6] and compare with c1Re1 + c3Re3 as well. We obtain that
bγ − βc = 0, where b, β 6= 0, which is impossible.
α2,3 − β2,1 + β2,3 0 α2,3 γ − α
α − γ
c
a2,3 − b2,1 + b2,3 0 a2,3
0 −c
0
0
, Re6 =
0
c − a
0
a − c
−α
0
−a
0
α
0
0
0
0
a
γ
,
5.1.2.2. Codimension two solvable extensions of L 4, (n ≥ 5)
By taking a linear combination of Ren+1 and Ren+2 and keeping in mind that no nontrivial
linear combination of the matrices Ren+1 and Ren+2 can be a nilpotent matrix, one could
β ! = 1
1 ! . Therefore the vector space of outer derivations
b ! = 1
2 ! and α
set a
as n × n matrices is as follows:
1
3b2,1−2a2,3 4 a2,3
0
0
1
0
0
0
2
0
0
0
0 a5,3
0
·
Ren+1 =
0
1
0
3
0
a5,3
0
0
0
0
4
0
0
0
0
0
0
5
···
···
···
···
···
···
... 0
...
...
...
·
...
0
...
·
...
0
...
0
0
0
0 an−2,3 an−3,3 ···
0 an−1,3 an−2,3 ···
0 an,3 an−1,3 ···
··· a5,3
···
···
·
·
...
a5,3
·
0 n−3
0 n−2
a5,3
0 n−1
0
0
0
0
0
0
...
...
0
0
0
0
0
0
...
...
0
0
0
0
0
0
0
...
...
0
0
,
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
33
1
0
0
β2,1 2 α2,3
0
0
0
0
1
0
0
0
0 α5,3
0
·
0
0
0
2
0
α5,3
0
0
0
0
3
0
0
0
0
0
0
4
···
···
···
···
···
···
Ren+2 =
0
...
·
...
... 0
...
...
...
...
·
...
0 αn−2,3 αn−3,3 ···
0 αn−1,3 αn−2,3 ···
0 αn,3 αn−1,3 ···
··· α5,3
···
···
·
0 n−4
· α5,3
0 n−3
· α5,3
0 n−2
0
0
0
0
0
0
...
...
0
0
0
0
0
0
...
...
0
0
0
0
0
0
0
...
...
0
0
.
0
...
0
0
0
0
...
0
0
0
1
b2,1−
0
0
0
0
(i) Considering R[en+1,en+2], which is the same as [Ren+2, Ren+1], we deduce that αi,3 :=
ai,3, (5 ≤ i ≤ n), α2,3 := a2,3
2 . As a result, the outer derivation Ren+2
changes as follows:
2 and β2,1 := b2,1 − a2,3
a2,3
0
2
Ren+2 =
a2,3
0
2
0
0
0 a5,3
0
2
1
0
·
0
...
·
...
0
0
0
2
0
·
...
a5,3
0
0
0
0
3
0
0
0
0
0
0
4
···
···
···
···
···
···
... 0
...
...
...
0 an−2,3 an−3,3 ···
0 an−1,3 an−2,3 ···
0 an,3 an−1,3 ···
··· a5,3
···
···
·
·
0
0
0
0
0
0
...
...
0
0
0
0
0
0
...
...
0
0
0
0
0
0
0
...
...
0
0
.
...
a5,3
·
0 n−4
0 n−3
a5,3
0 n−2
Altogether we find the following commutators:
R[en+1,e1] = −Re1 − Re3, R[en+1,e2] = 0, R[en+1,ei] = (1 − i)Rei −
R[en+1,ej] = 0, (n ≤ j ≤ n + 1), R[en+1,en+2] = −
ak+1,3Rek, R[en+2,e1] = −Re1,
R[en+2,e2] = 0, R[en+2,ei] = (2 − i)Rei −
n−1
n−1
Xk=i+2
ak−i+3,3Rek, (3 ≤ i ≤ n − 1),
R[en+2,en] = 0, R[en+2,en+1] =
ak+1,3Rek, R[en+2,en+2] = 0.
n−1
Xk=4
n−1
Xk=i+2
ak−i+3,3Rek,
Xk=4
(ii) We include a linear combination of e2 and en :
[en+1, e1] = −e1 + c2,1e2 − e3 + cn,1en, [en+1, e2] = c2,2e2 + cn,2en, [en+1, ei] = c2,ie2+
(1 − i)ei −
ak−i+3,3ek + cn,ien, [en+1, ej] = c2,je2 + cn,jen, (n ≤ j ≤ n + 1),
n−1
Xk=i+2
[en+1, en+2] = c2,n+2e2 −
ak+1,3ek + cn,n+2en, [en+2, e1] = −e1 + d2,1e2 + dn,1en,
n−1
Xk=4
[en+2, e2] = d2,2e2 + dn,2en, [en+2, ei] = d2,ie2 + (2 − i)ei −
ak−i+3,3ek + dn,ien,
n−1
Xk=i+2
(3 ≤ i ≤ n − 1), [en+2, en] = d2,ne2 + dn,nen, [en+2, en+1] = d2,n+1e2 +
dn,n+1en, [en+2, en+2] = d2,n+2e2 + dn,n+2en.
n−1
Xk=4
ak+1,3ek+
Besides we have the brackets from L 4 and from outer derivations Ren+1 and Ren+2 as
well.
34
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(iii) We satisfy the right Leibniz identity shown in Table 3. We notice that Ren+1 and
Ren+2 restricted to the nilradical do not change, but the remaining brackets are as follows:
n
Xk=5
ak,3ek, [en+1, e4] = 3 (e2 − e4) −
n
n
Xk=i+2
ak−1,3ek, [en+1, ei] = (1 − i)ei −
[en+1, e1] = −e1 + b2,1e2 − e3, [en+1, e3] = a2,3e2 − 2e3 −
Xk=6
ak+1,3ek + cn,n+2en, [en+2, e1] = −e1 +(cid:18)b2,1 −
Xk=4
[en+1, en+2] = c2,n+2e2 −
Xk=5
[en+2, e3] =
e2 − e3 −
ak,3ek, [en+2, e4] = 2 (e2 − e4) −
Xk=i+2
ak−i+3,3ek, (5 ≤ i ≤ n), [en+2, en+1] = d2,n+1e2 +
ak−i+3,3ek, [en+1, en+1] = (2c2,n+2 − 2a5,3) e2,
2 (cid:19) e2,
ak−1,3ek,
a2,3
2
Xk=6
a2,3
n−1
n
n
n
n−1
Xk=4
ak+1,3ek−
[en+2, ei] = (2 − i)ei −
cn,n+2en, [en+2, en+2] =
d2,n+1 + a5,3
e2.
2
Table 3. Right Leibniz identities in the codimension two nilradical
L 4, (n ≥ 5).
Steps Ordered triple
Result
k=6 ak−1,3ek.
c2,i+1 = 0, cn,i+1 := −an−i+2,3, (4 ≤ i ≤ n − 2), where a4,3 = 0 =⇒ [en+1, ej ] =
c2,n = 0, cn,n := 1 − n =⇒ [en+1, en] = (1 − n) en. Altogether with 4., [en+1, ei] =
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
d2,4 := 2, dn,4 := −an−1,3, where a4,3 = 0 =⇒ [en+2, e4] = 2(e2 − e4) −
d2,i+1 = 0, dn,i+1 := −an−i+2,3, (4 ≤ i ≤ n − 2), where a4,3 = 0 =⇒ [en+2, ej] =
Re1 ([en+1, e1])
[en+1, e2] = 0 =⇒ c2,2 = cn,2 = 0.
k=6 ak−1,3ek.
[en+2, e2] = 0 =⇒ d2,2 = dn,2 = 0.
k=i+2 ak−i+3,3ek, (5 ≤ i ≤ n).
Re1 ([en+2, e1])
Re3 ([en+1, e1])
Rei ([en+1, e1])
Ren−1 ([en+1, e1])
Re3 ([en+2, e1])
Rei ([en+2, e1])
k=j+2 ak−j+3,3ek, (5 ≤ j ≤ n − 1).
c2,4 := 3, cn,4 := −an−1,3, where a4,3 = 0 =⇒ [en+1, e4] = 3(e2−e4)−Pn
(1 − j) ej −Pn
(1 − i) ei −Pn
Pn
(2 − j) ej −Pn
(2 − i) ei −Pn
c2,3 := a2,3, cn,3 := −an,3 =⇒ [en+1, e3] = a2,3e2 − 2e3 −Pn
2 e2 − e3 −Pn
2 , dn,1 = 0 =⇒ [en+2, e1] = −e1 +(cid:0)b2,1 −
2 , dn,3 := −an,3 =⇒ [en+2, e3] = a2,3
c2,1 := b2,1, cn,1 = 0 =⇒ [en+1, e1] = −e1 + b2,1e2 − e3.
Ren+1 ([en+1, en+1])cn,n+1 = 0 =⇒ [en+1, en+1] = c2,n+1e2.
Ren+2 ([en+1, en+2])dn,n+2 = 0 =⇒ [en+2, en+2] = d2,n+2e2.
Ren+1 ([en+1, e3])
Ren+2 ([en+2, e3]) d2,3 := a2,3
Ren+2 ([en+1, e1])
Ren+2 ([en+2, e1]) d2,1 := b2,1 −
Ren+1 ([en+1, en+2])c2,n+1
:= 2c2,n+2 − 2a5,3, dn,n+1
k=j+2 ak−j+3,3ek, (5 ≤ j ≤ n − 1).
k=i+2 ak−i+3,3ek, (5 ≤ i ≤ n).
k=5 ak,3ek.
k=5 ak,3ek.
2 (cid:1) e2.
:= −cn,n+2
=⇒
a2,3
a2,3
Ren+2 ([en+2, en+1])d2,n+2 := d2,n+1+a5,3
(2c2,n+2 − 2a5,3) e2, [en+2, en+1] = d2,n+1e2 +Pn−1
=⇒ [en+2, en+2] = d2,n+1+a5,3
2
2
e2.
k=4 ak+1,3ek − cn,n+2en.
Ren−1 ([en+2, e1]) d2,n = 0, dn,n := 2 − n =⇒ [en+2, en] = (2 − n) en. Combining with 7., [en+2, ei] =
[en+1, en+1] =
Altogether the nilradical L 4 (4.0.1), the outer derivations Ren+1 and Ren+2 and the re-
maining brackets given above define a continuous family of solvable right Leibniz algebras
depending on the parameters. Then we apply the technique of "absorption" according to
step (iv).
• First we apply the transformation e′
c2,n+2
e2, e′
n+2 = en+2 − d2,n+1+a5,3
4
2
i = ei, (1 ≤ i ≤ n, n ≥ 5), e′
n+1 = en+1 −
e2 to remove the coefficients c2,n+2 and d2,n+1+a5,3
2
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
35
2
a5,3
a5,3
2 e4, e′
n+1 = en+1 + a5,3
i = ei, (1 ≤ i ≤ n, n ≥ 5), e′
in front of e2 in [en+1, en+2] and [en+2, en+2], respectively. This transformation
changes the coefficient in front of e2 in [en+1, en+1] and [en+2, en+1] to −2a5,3 and
−a5,3, respectively.
• Then we apply e′
n+2 = en+2 to re-
move the coefficients −a5,3 and −2a5,3 in front of e2 in [en+2, en+1] and [en+1, en+1],
respectively. Besides this transformation removes a5,3 and −a5,3 in front of e4 in
[en+2, en+1] and [en+1, en+2], respectively, and changes the coefficients in front of
ek, (6 ≤ k ≤ n − 1) in [en+2, en+1] and [en+1, en+2] to ak+1,3 − a5,3
2 ak−1,3 and
2 ak−1,3 − ak+1,3, respectively. It also affects the coefficients in front en, (n ≥ 6)
in [en+1, en+2] and [en+2, en+1], which we rename back by cn,n+2 and −cn,n+2, re-
spectively. The following entries are introduced by the transformation: − a5,3
2 and
in the (5, 1)st, (n ≥ 5) position in Ren+1 and Len+1, respectively.
• Applying the transformation e′
j = ej, (1 ≤ j ≤ n + 1, n ≥ 6), e′
Pn−1
2 a5,3ak−1,3−Pk
(6 ≤ k ≤ n − 1, n ≥ 7), we remove the coefficients a6,3 and −a6,3 in front of e5 in
[en+2, en+1] and [en+1, en+2], respectively. We also remove ak+1,3 − a5,3
2 ak−1,3 and
a5,3
2 ak−1,3 − ak+1,3 in front of ek, (6 ≤ k ≤ n − 1) in [en+2, en+1] and [en+1, en+2],
respectively. This transformation introduces a6,3
in the (6, 1)st, (n ≥ 6)
position as well as Ak+1,3
in the (k + 1, 1)st, (6 ≤ k ≤ n− 1) in Ren+2 and
Len+2, respectively. It also affects the coefficients in front en, (n ≥ 7) in [en+1, en+2]
and [en+2, en+1], which we rename back by cn,n+2 and −cn,n+2, respectively.
i = ei, (1 ≤ i ≤ n + 1, n ≥ 5), e′
n+2 =
en+2 + cn,n+2
n−1 en, we remove cn,n+2 and −cn,n+2 in front of en in [en+1, en+2] and
[en+2, en+1], respectively, without affecting other entries. We obtain that Ren+1
and Ren+2 are as follows:
k−1 ek, where A6,3 := a6,3 and Ak+1,3 := ak+1,3− 1
• Finally applying the transformation e′
n+2 = en+2 −
4 and − a6,3
k−1 and Ak+1,3
1−k
Ak+1,3
k=5
i−3
i=7
4
Ai−1,3ak−i+5,3
,
Ren+2 =
a2,3
2
1
b2,1−
0
0
0
a6,3
4
A7,3
5
A8,3
6
...
An−2,3
n−4
An−1,3
n−3
An,3
n−2
0 an−2,3 an−3,3 ···
0 an−1,3 an−2,3 ···
0 an,3 an−1,3 ···
···
···
···
··· a5,3
0 n−4
···
···
·
·
a5,3
0 n−3
·
a5,3
0 n−2
1
0
0
3b2,1−2a2,3 4 a2,3
1
0
a5,3
2
0
−
2
0
0
0
0 a5,3
0
·
Ren+1 =
0
1
0
3
0
·
...
a5,3
0
0
0
0
4
0
0
0
0
0
0
5
···
···
···
···
···
···
... 0
...
...
...
0
...
0
0
0
0
a2,3
0
2
0
0
0 a5,3
0
2
1
0
·
0
...
·
...
0 an−2,3 an−3,3 ···
0 an−1,3 an−2,3 ···
0 an,3 an−1,3 ···
0
0
0
0
0
0
0
0
0
3
0
0
0
0
0
0
0
0
2
0
a5,3
0
4
0
0
5
a5,3
0
0
...
·
·
...
·
·
...
·
a5,3 0
...
...
a5,3
·
0
0
0
0
0
0
...
...
...
·
·
··· a5,3
···
···
···
···
···
···
···
···
...
...
...
,
0 n−3
0 n−2
0 n−1
0
0
0
0
0
0
...
...
a5,3
0
0
0
0
0
0
...
...
...
0
0
0
0
0
0
0
...
...
0
0
0
0
0
0
0
...
...
...
0
0
0
0
0
0
0
0
...
...
0
0
, (n ≥ 5).
36
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
The remaining brackets are given below:
a5,3
2
(v) Finally we apply the following two change of basis transformations:
Bk−i+3,3
n
ak,3ek,
ak−i+3,3ek,
n
n = en, e′
n
n
a2,3
2 = e2, e′
n
a2,3
2
n
Bk−2,3aj−k+5,3
k=7
k−5
ek, [en+2, e3] =
n
Xk=6
Ak,3
k − 2
3 = e3 − a2,3
2 e2, e′
n+1 = en+1, e′
Xk=5
ak,3ek,
e2 − e3 −
ak−i+3,3ek, (5 ≤ i ≤ n).
Xk=i+2
i = ei − Pn
Xk=5
e5, [en+1, e3] = a2,3e2 − 2e3 −
Xk=i+2
[en+1, e1] = −e1 + b2,1e2 − e3 +
Xk=6
[en+1, e4] = 3 (e2 − e4) −
ak−1,3ek, [en+1, ei] = (1 − i)ei −
[en+2, e1] = −e1 +(cid:18)b2,1 −
2 (cid:19) e2 −
Xk=6
[en+2, e4] = 2 (e2 − e4) −
ak−1,3ek, [en+2, ei] = (2 − i)ei −
1 = e1 − (cid:16)b2,1 − a2,3
2 (cid:17) e2, e′
• e′
n−1 = en−1, e′
(3 ≤ i ≤ n − 2), e′
aj,3 −Pj
, (5 ≤ j ≤ n).
This transformation removes a5,3, a6,3, ..., an,3 in Ren+1, Ren+2 and −a5,3,−a6,3, ...,−an,3
in Len+1, Len+2. It also removes − a5,3
from the (5, 1)st positions in Ren+1
and Len+1, respectively. Besides it removes 3b2,1 − 2a2,3 and b2,1 from the (2, 1)st
positions in Ren+1 and Len+1, respectively, as well as b2,1− a2,3
from the (2, 1)st po-
sitions in Ren+2 and Len+2. The transformation also removes a2,3 from the (2, 3)rd
positions in Ren+1, Len+1 and a2,3
from the same positions in Ren+2, Len+2. This
transformation introduces the entries in the (i, 1)st positions in Ren+1 and Len+1,
which we set to be ai,1 and −ai,1, (6 ≤ i ≤ n), respectively. The transforma-
tion affects the entries in the (j, 1)st, (8 ≤ j ≤ n) positions in Ren+2 and Len+2,
but we rename all the entries in the (i, 1)st positions by i−3
i−2 ai,1 in Ren+2 and by
3−i
i−2 ai,1, (6 ≤ i ≤ n) in Len+2.
n+2 = en+2 +Pn−1
i−2 ai,1 and 3−i
k−1ak+1,1ek, we remove ai,1 in Ren+1 and −ai,1 in Len+1 as
• Finally applying the transformation e′
ek,
n+2 = en+2, where Bj,3 :=
k=i+2
k−i
n+1 = en+1+Pn−1
i−2 ai,1, (6 ≤ i ≤ n) in Ren+2 and Len+2, respectively.
k = ek, (1 ≤ k ≤ n), e′
2 and a5,3
2
k=5 ak+1,1ek,
e′
well as i−3
k−2
k=5
2
2
We obtain a Leibniz algebra gn+2,1 given below:
[e1, en+1] = e1 + e3, [e2, en+1] = 4e2, [e3, en+1] = 2e3, [e4, en+1] = e2 + 3e4,
[ei, en+1] = (i − 1)ei, [en+1, e1] = −e1 − e3, [en+1, e3] = −2e3, [en+1, e4] = 3(e2 − e4),
[en+1, ei] = (1 − i)ei, [e1, en+2] = e1, [e2, en+2] = 2e2, [ej, en+2] = (j − 2)ej, (3 ≤ j ≤ n),
[en+2, e1] = −e1, [en+2, e3] = −e3, [en+2, e4] = 2(e2 − e4), [en+2, ei] = (2 − i)ei, (5 ≤ i ≤ n).
We summarize a result in the following theorem:
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
37
Theorem 5.1.4. There are four solvable indecomposable right Leibniz algebras up to
isomorphism with a codimension two nilradical L 4, (n ≥ 4), which are given below:
(i) gn+2,1 : [e1, en+1] = e1 + e3, [e2, en+1] = 4e2, [e3, en+1] = 2e3, [e4, en+1] = e2 + 3e4,
[ei, en+1] = (i − 1)ei, [en+1, e1] = −e1 − e3, [en+1, e3] = −2e3, [en+1, e4] = 3(e2 − e4),
[en+1, ei] = (1 − i)ei, [e1, en+2] = e1, [e2, en+2] = 2e2, [ej, en+2] = (j − 2)ej, (3 ≤ j ≤ n),
[en+2, e1] = −e1, [en+2, e3] = −e3, [en+2, e4] = 2(e2 − e4), [en+2, ei] = (2 − i)ei, (5 ≤ i ≤ n),
(n ≥ 5),
(ii) g6,2 : [e1, e5] = e1, [e2, e5] = 2e2, [e3, e5] = e3, [e4, e5] = 2e4, [e5, e1] = −e1, [e5, e3] = −e3,
[e5, e4] = 2 (e2 − e4) , [e1, e6] = e1 + be3, [e2, e6] = 2(b + 1)e2, [e3, e6] = e1 + be3,
[e4, e6] = (b + 1) (e2 + e4) , [e6, e1] = −e1 − be3, [e6, e3] = −e1 − be3, [e6, e4] = (b + 1) (e2 − e4) ,
(b 6= −1),
(iii) g6,3 : [e1, e5] = e1 + (c + 1)e3, [e2, e5] = 2(c + 2)e2, [e3, e5] = ce1 + 2e3,
[e4, e5] = (2c + 1)e2 + 3e4, [e5, e1] = −e1 − (c + 1)e3, [e5, e3] = −ce1 − 2e3, [e5, e4] = 3 (e2 − e4) ,
[e1, e6] = e1, [e2, e6] = 2e2, [e3, e6] = e3, [e4, e6] = 2e4, [e6, e1] = −e1, [e6, e3] = −e3,
[e6, e4] = 2 (e2 − e4) , (c 6= −2),
(iv) g6,4 : [e1, e5] = e1, [e2, e5] = 2e2, [e3, e5] = e3, [e4, e5] = 2e4, [e5, e1] = −e1, [e5, e3] = −e3,
[e5, e4] = 2 (e2 − e4) , [e1, e6] = be1 + (1 − b)e3, [e2, e6] = 2e2, [e3, e6] = e1, [e4, e6] = (2 − b)e2+
be4, [e6, e1] = −be1 + (b − 1)e3, [e6, e3] = −e1, [e6, e4] = b (e2 − e4) , (b 6= 0).
5.1.2.3. Codimension three solvable extensions of L 4, (n = 4)
Considering R[e5,e6], R[e5,e7] and R[e6,e7] and comparing with c1Re1 + c3Re3, where
Rei =
ai
0
ci
0
(3ai−2bi−3ci)ai2,3+(ai−2bi−3ci)bi2,3
bi + ci − ai
2ai
0
2(bi + ci) ai2,3 2ci + bi − ai
0
0
bi
0
0
ai + bi
, (5 ≤ i ≤ 7),
a1
b1
c1
we have to have, respectively, that
a1c2 − a2c1 − b1c2 + b2c1 = 0, a1c3 − a3c1 − b1c3 + b3c1 = 0, a2c3 − a3c2 − b2c3 + b3c2 = 0.
However, it means that
are linearly dependent: one could
see that considering det
a1 (a2c3 − a3c2) − a2 (a1c3 − a3c1) + a3 (a1c2 − a2c1) = 0.
5.2. Solvable indecomposable left Leibniz algebras with a nilradical L 4
= a1 (b2c3 − b3c2)−a2 (b1c3 − b3c1)+a3 (b1c2 − b2c1) =
and
a1 a2 a3
b1
b3
c3
c1
b2
c2
,
a2
b2
c2
a3
b3
c3
5.2.1. Codimension one solvable extensions of L 4
Classification follows the same steps in theorems with the same cases per step.
38
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Theorem 5.2.1. We set a1,1 := a and a3,3 := b in (5.1.1). To satisfy the left Leibniz
identity, there are the following cases based on the conditions involving parameters, each
gives a continuous family of solvable Leibniz algebras:
(1) If a1,3 = 0, b 6= −a, a 6= 0, b 6= 0, (n = 4) or b 6= (3 − n)a, a 6= 0, b 6= 0, (n ≥ 5),
then we have the following brackets for the algebra:
n
n
[e1, en+1] = ae1 + a2,1e2 + (b − a)e3 +
[e4, en+1] = (a + b)(e4 − e2) +
n
Xk=4
Xk=5
ak,1ek, [e3, en+1] = a2,3e2 + be3 +
ak−1,3ek, [ei, en+1] = ((i − 3)a + b) ei +
ak,3ek,
ak−i+3,3ek,
n
Xk=4
Xk=i+1
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −ae1 + B2,1e2 + (a − b)e3 −
Xk=4
[en+1, e2] = −2be2, [en+1, e3] = (a2,3 + 2a4,3)e2 − be3 −
Xk=i+1
(a + b)e4 −
ak−1,3ek, [en+1, ei] = ((3 − i)a − b) ei −
(2b − a)a2,1 + 2b · a4,1 + 2(a − b)(a2,3 + a4,3)
where B2,1 :=
Xk=5
n
n
n
,
n
Xk=4
ak,1ek,
ak,3ek, [en+1, e4] = (a − b)e2−
ak−i+3,3ek, (5 ≤ i ≤ n),
Len+1 =
−ai,1
−ai,3
−ai−1,3 −ai−2,3
−a4,3
(3−i)a−b ···
0
0
−a
B2,1 −2b a2,3+2a4,3
a−b
−a4,1
−a5,1
−a4,3
−a5,3
0
0
0
−b
·
·
·
...
...
0
...
...
...
...
0
a−b
0
...
...
−a−b
−a4,3 −2a−b
−a5,3
−a4,3
a
0
0
0
0
...
...
···
···
···
···
···
0
0
0
0
0
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
···
...
...
...
...
...
...
...
...
0
...
...
...
0
...
...
...
0
0
(3−n)a−b
.
−an−1,1
−an,1
0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 (4−n)a−b
0
−an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3
(2) If a1,3 = 0, b := −a, a 6= 0, (n = 4) or b := (3 − n)a, a 6= 0, (n ≥ 5), then the
brackets for the algebra are
n
n
n
n
n
Xk=4
ak,1ek, [e3, en+1] = a2,3e2 + (3 − n)ae3+
Xk=5
ak−1,3ek, [ei, en+1] = (i − n) aei+
ak,3ek, [e4, en+1] = (n − 4)a(e2 − e4) +
ak−i+3,3ek, [en+1, en+1] = a2,n+1e2 + an,n+1en, [en+1, e1] = −ae1 + B2,1e2+
[e1, en+1] = ae1 + a2,1e2 − (n − 2)ae3 +
Xk=4
Xk=i+1
(n − 2)ae3 −
(n − 3)ae3 −
[en+1, ei] = (n − i) aei −
where B2,1 := (5 − 2n)a2,1 + (6 − 2n)a4,1 + 2(n − 2)(a2,3 + a4,3),
ak,1ek, [en+1, e2] = (2n − 6)ae2, [en+1, e3] = (a2,3 + 2a4,3)e2+
ak,3ek, [en+1, e4] = (n − 2)ae2 + (n − 4)ae4 −
ak−i+3,3ek, (5 ≤ i ≤ n),
Xk=4
Xk=4
Xk=i+1
ak−1,3ek,
Xk=5
n
n
n
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
39
Len+1 =
0
0
0
(2n−6)a a2,3+2a4,3 (n−2)a
(n−3)a
−a4,3
−a5,3
0
(n−4)a
−a4,3
0
0
0
0
(n−5)a
···
···
···
···
···
0
0
0
0
0
−a5,3
−a4,3
...
...
...
...
...
...
···
−a4,3
...
...
·
...
...
...
...
...
−ai,1
−ai,3
−ai−1,3 −ai−2,3
(n−i)a
−a
B2,1
(n−2)a
−a4,1
−a5,1
·
...
...
···
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
0
...
...
...
0
.
...
...
...
0
−an−1,1
−an,1
−an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3
0
−an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 0
a
(3) If a1,3 = 0, a = 0, b 6= 0, (n = 4) or a = 0 and b 6= 0, (n ≥ 5), then
[e1, en+1] = (a2,3 + a4,3 − a4,1) e2 + be3 +
[e4, en+1] = b (e4 − e2) +
ak−1,3ek, [ei, en+1] = bei +
Xk=4
n
n
n
ak,1ek, [e3, en+1] = a2,3e2 + be3 +
n
Xk=4
ak,3ek,
n
Xk=4
ak−i+3,3ek,
Xk=i+1
ak,1ek, [en+1, e2] = −2be2,
Xk=5
n
ak,3ek, [en+1, e4] = −b (e2 + e4) −
ak−1,3ek,
Xk=5
0
...
...
[en+1, en+1] = a2,n+1e2, [en+1, e1] = b2,1e2 − be3 −
[en+1, e3] = (a2,3 + 2a4,3)e2 − be3 −
[en+1, ei] = −bei −
ak−i+3,3ek, (5 ≤ i ≤ n),
Xk=4
n
n
Xk=i+1
0
b2,1 −2b a2,3+2a4,3
−b
−b
−a4,1
−a5,1
−a4,3
−a5,3
0
−b
0
−b
−a4,3
0
0
0
0
−b
···
···
···
···
···
0
0
0
0
0
Len+1 =
·
·
·
−a5,3
−a4,3
−ai,1
−ai,3
−ai−1,3 −ai−2,3
...
...
...
...
...
...
···
−a4,3
...
...
...
...
...
−b
0
...
...
0
0
0
·
0
...
...
0
0
0
0
0
0
...
...
···
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
0
...
...
...
0
.
...
...
...
0
−an−1,1
−an,1
0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 −b
0
0
−an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 −b
(4) If a1,3 = 0, b = 0, a 6= 0, (n = 4) or b = 0, a 6= 0, (n ≥ 5), then
Xk=4
ak,1ek, [e3, en+1] = a2,3e2 +
[e1, en+1] = ae1 + a2,1e2 − ae3 +
Xk=i+1
[e4, en+1] = a(e4 − e2) +
ak−1,3ek, [ei, en+1] = (i − 3)aei +
Xk=5
Xk=4
n
n
n
n
ak,3ek,
ak−i+3,3ek,
n
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −ae1 + (a2,3 − a2,1 + b2,3) e2 + ae3 −
Xk=4
[en+1, e3] = b2,3e2 −
[en+1, ei] = (3 − i)aei −
ak,3ek, [en+1, e4] = a(e2 − e4) −
Xk=i+1
ak−i+3,3ek, (5 ≤ i ≤ n),
ak−1,3ek,
Xk=5
n
n
n
Xk=4
ak,1ek,
40
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Len+1 =
0
−a
0
a2,3−a2,1+b2,3 0
0
0 −a4,3
0 −a5,3
−a4,1
−a5,1
b2,3
0
a
0
a
0
−a
0
0
0
0
−a4,3
−2a
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
·
·
·
−a5,3
−a4,3
...
...
...
...
...
...
...
...
...
...
...
...
···
−a4,3
...
...
−ai,1
0 −ai,3 −ai−1,3 −ai−2,3
(3−i)a
...
...
...
...
...
...
0
...
...
...
0
...
...
...
0
···
···
···
···
···
···
−an−1,1
−an,1
0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 (4−n)a
0 −an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 (3−n)a
0
In the remaining cases a1,3 := c.
.
(5) If b 6= −a, a 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = ae1 + a2,1e2 + (b + c − a)e3 + a4,1e4, [e3, e5] = ce1 + a2,3e2 + be3 + A4,3e4,
[e4, e5] = (a + b)(e4 − e2), [e5, e5] = a2,5e2, [e5, e1] = −ae1 + B2,1e2 + (a − b − c)e3 − a4,1e4,
[e5, e2] = −2(b + c)e2, [e5, e3] = −ce1 + b2,3e2 − be3 − A4,3e4, [e5, e4] = (a − 2c − b)e2−
(a + b)e4; B2,1 :=
(2b − a + 2c)a2,1 + 2(b + c)a4,1 + (a − b − c)(a2,3 + b2,3)
and
a
2c(a2,1 + a4,1) − (a + c)a2,3 + (a − c)b2,3
,
A4,3 :=
0
−a
B2,1 −2(b+c)
Le5 =
.
(6) If b := −a, a 6= 0, a 6= c, c 6= 0, then
−A4,3 −a−b
a−b−c
−a4,1
0
0
0
2a
−c
b2,3 a−2c−b
−b
0
[e1, e5] = ae1 + a2,1e2 + (c − 2a)e3 + a4,1e4, [e3, e5] = ce1 + a2,3e2 − ae3 + A4,3e4,
[e5, e5] = a2,5e2 + a4,5e4, [e5, e1] = −ae1 + B2,1e2 + (2a − c)e3 − a4,1e4,
[e5, e2] = 2(a − c)e2, [e5, e3] = −ce1 + b2,3e2 + ae3 − A4,3e4, [e5, e4] = 2(a − c)e2,
where B2,1 :=
(2c − 3a)a2,1 + 2(c − a)a4,1 + (2a − c)(a2,3 + b2,3)
and
A4,3 :=
2c(a2,1 + a4,1) − (a + c)a2,3 + (a − c)b2,3
,
a
0
Le5 = " −a
B2,1 2(a−c)
2a−c
−a4,1
0
0
−c
b2,3
a
−A4,3
2a
0
2(a−c)
0
0
# .
−a
0 −c
(7) If b := −c, c 6= 0, a 6= c, a 6= 0, then
[e1, e5] = ae1 + a2,1e2 − ae3 + a4,1e4, [e3, e5] = ce1 + a2,3e2 − ce3 + a4,3e4,
[e4, e5] = (c − a)(e2 − e4), [e5, e5] = a2,5e2, [e5, e1] = −ae1 + (a2,3 − a2,1 + b2,3) e2 + ae3−
a4,1e4, [e5, e3] = −ce1 + b2,3e2 + ce3 − a4,3e4, [e5, e4] = (c − a)(e4 − e2),
a2,3−a2,1+b2,3 0 b2,3 a−c
Le5 = "
0 −a4,3 c−a# .
(8) If c := a, b := −a, a 6= 0, then3
[e1, e5] = ae1 + a2,1e2 − ae3 + (a4,3 + b4,3 − b4,1) e4, [e3, e5] = ae1 + a2,3e2 − ae3 + a4,3e4,
[e5, e5] = a2,5e2 + a4,5e4, [e5, e1] = −ae1 + (a2,3 − a2,1 + b2,3) e2 + ae3 + b4,1e4,
[e5, e3] = −ae1 + b2,3e2 + ae3 + b4,3e4,
3The outer derivation Le5 is nilpotent, so we remove this case from further consideration.
−a4,1
0
0
0
a
c
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
41
−a
Le5 = "
0 −a 0
a2,3−a2,1+b2,3 0 b2,3 0
0 a 0
0 b4,3 0# .
(9) If a = 0, b = 0, c 6= 0, then
a
b4,1
[e1, e5] = a2,1e2 + ce3 + a4,1e4, [e3, e5] = ce1 + (2a2,1 + 2a4,1 − b2,3) e2 + a4,3e4,
[e5, e5] = a2,5e2 + a4,5e4, [e5, e1] = (3a2,1 + 4a4,1 − 2b2,3 + 2a4,3) e2 − ce3 − a4,1e4,
[e5, e2] = −2ce2, [e5, e3] = −ce1 + b2,3e2 − a4,3e4, [e5, e4] = −2ce2,
Le5 = "
3a2,1+4a4,1−2b2,3+2a4,3 −2c b2,3 −2c
0 # .
0
0 −a4,3
−a4,1
−c
−c
0
0
0
0
0
2
− a4,1! e2 + (b + c)e3 + a4,1e4, [e3, e5] = ce1 + a2,3e2 + be3 + A4,3e4,
(10) If a = 0, b 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = a2,3 + b2,3
[e4, e5] = b(e4 − e2), [e5, e5] = a2,5e2, [e5, e1] = b2,1e2 − (b + c)e3 − a4,1e4,
[e5, e2] = −2(b + c)e2, [e5, e3] = −ce1 + b2,3e2 − be3 − A4,3e4, [e5, e4] = −(2c + b)e2 − be4,
where A4,3 :=
(2b + c)b2,3 − (2b + 3c)a2,3 + 2c(b2,1 − a4,1)
,
Le5 =
0
0
b2,1 −2(b+c)
−b−c
−a4,1
0
0
0
−c
b2,3 −2c−b
−b
0
−A4,3 −b
4(b + c)
.
(11) If a = 0, b := −c, c 6= 0, then
[e1, e5] = (a2,3 + b2,3 − b2,1) e2 + a4,1e4, [e3, e5] = ce1 + a2,3e2 − ce3 + a4,3e4,
[e4, e5] = c(e2 − e4), [e5, e5] = a2,5e2, [e5, e1] = b2,1e2 − a4,1e4, [e5, e3] = −ce1 + b2,3e2+
ce3 − a4,3e4, [e5, e4] = c(e4 − e2),
0 −c
Le5 = " 0
0
b2,1 0 b2,3 −c
0
−a4,1 0 −a4,3 c # .
0
0
c
Proof.
(1) The proof is off-loaded to Table 4, when (n ≥ 5). For (n = 4), we recalcu-
late the applicable identities, which are 1., 2., 4., 6. − 8., 10. − 15.
(2) We repeat case (1), except the identity 13.
(3) If (n ≥ 5), then we apply the identities given in Table 4, except 14. and 15. apply-
ing instead: Le3[en+1, en+1] = [Le3(en+1), en+1]+[en+1, Le3(en+1)] and L[e1, en+1] =
[L(e1), en+1] + [e1, L(en+1)]. For (n = 4), we apply the same identities as in case
(1) except 14. and 15. applying two identities given above.
(4) We apply the same identities as in case (1).
(5) We apply the following identities: 1., 2., 4., 6.−8., 10.−14., Le3[e5, e5] = [Le3(e5), e5]+
[e5, Le3(e5)], 15.
(6) We apply the same identities as in case (2) for (n = 4).
(7) We apply the same identities as in case (1) for (n = 4).
(8) We apply the same identities as in case (1) for (n = 4), except 13. and 15.
42
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(9) We apply the same identities as in case (2) for (n = 4).
(10) The same identities as in case (5).
(11) The same identities as in case (5), except the identity 15.
(cid:3)
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
43
Table 4. Left Leibniz identities in case (1) in Theorem 5.2.1, (n ≥ 5).
Steps Ordered triple
Result
1.
2.
Le1 ([e1, en+1])
[e2, en+1] = 0 =⇒ ak,2 = 0, (1 ≤ k ≤ n).
Le1 ([e3, en+1])
a1,4 = 0, ak,4 := ak−1,3, (5 ≤ k ≤ n), a3,1 := a2,4 + a1,3 + 2b, a3,4 = 0, a4,4 := a + b
=⇒ [e1, en+1] = ae1 + a2,1e2 + (a2,4 + a1,3 + 2b)e3 +Pn
(a + b)e4 +Pn
k=5 ak−1,3ek.
a1,i+1 = a3,i+1 = 0 and we had that a1,4 = a3,4 = 0 =⇒ a2,i+1 = a4,i+1 =
k=4 ak,1ek, [e4, en+1] = a2,4e2 +
0; ai+1,i+1 := (i − 2)a + b, ak,i+1 := ak−1,i, (5 ≤ k ≤ n, k 6= i + 1, 4 ≤ i ≤ n − 1)
=⇒ [ej , en+1] = ((j − 3)a + b) ej +Pn
k=j+1 ak−j+3,3ek, (5 ≤ j ≤ n).
b1,2 = b3,2 = 0, bk,2 = 0, (5 ≤ k ≤ n), b3,3 := 2a1,3 + b + b2,2, b4,2 := a1,3 + b1,3 =⇒
[en+1, e2] = b2,2e2 + (a1,3 + b1,3)e4, [en+1, e3] = b1,3e1 + b2,3e2 + (2a1,3 + b + b2,2)e3 +
=⇒ [e1, en+1] = ae1 + a2,1e2 + (a2,4 +
k=4 ak,3ek, [en+1, e2] = b2,2e2 +
k=4 bk,3ek, [en+1, ei] = b2,ie2 +
3.
Le1 ([ei, en+1])
4.
Le3 ([en+1, e3])
5.
Le3 ([en+1, ei])
6.
7.
8.
9.
Le3 ([en+1, en])
L[e3, e3]
Le1 ([en+1, e3])
Le1 ([en+1, ei])
10.
Le3 ([en+1, e1])
11.
Le1 ([en+1, e1])
b1,3 = 0 =⇒ b2,2 := −2b =⇒ [en+1, e2] = −2be2, [en+1, e3] = b2,3e2 − be3 +
b2,4 := a2,4 + 2a, b4,4 := −a − b, bk,4 := bk−1,3, (5 ≤ k ≤ n) =⇒ [en+1, e4] =
b2,i+1 = b4,i+1 = 0, bi+1,i+1 := (2 − i)a − b, bk,i+1 := bk−1,i, (5 ≤ k ≤ n, k 6= i + 1, 4 ≤
k=4 bk,3ek.
k=4 bk,3ek.
k=4 bk,nek.
b1,i = b3,i = 0, a1,3 = 0
k=4 bk,iek, (4 ≤ i ≤ n − 1).
Pn
k=4 ak,1ek, [e3, en+1] = a2,3e2 + be3 +Pn
2b)e3 +Pn
b1,3e4, [en+1, e3] = b1,3e1 + b2,3e2 + (b + b2,2)e3 +Pn
Pn
b1,n = b3,n = 0 =⇒ [en+1, en] = b2,ne2 +Pn
Pn
(a2,4 + 2a) e2 − (a + b) e4 +Pn
i ≤ n − 1) =⇒ [en+1, ej] = ((3 − j)a − b) ej +Pn
[en+1, e1] = −ae1 + b2,1e2 + (a2,4 + 2a)e3 +Pn
Pn−1
k=4 ak,3ek+bn,3en, [en+1, e4] = (a2,4+2a)e2−(a+b)e4−Pn
((3 − i)a − b) ei −Pn
(b − a)e3 +Pn
−ae1 + b2,1e2 + (a − b)e3 −Pn−1
Pn
k=4 ak,1ek, [e4, en+1] = (a + b)(e4 − e2) +Pn
k=i+1 ak−i+3,3ek, (5 ≤ i ≤ n).
k=5 ak−1,3ek.
k=5 bk−1,3ek.
b1,1
:= −a, b3,1
:= a2,4 + 2a, bk−1,3
:= −ak−1,3, (5 ≤ k ≤ n)
=⇒
k=j+1 bk−j+3,3ek, (5 ≤ j ≤ n).
k=4 bk,1ek, [en+1, e3] = b2,3e2 − be3 −
k=5 ak−1,3ek, [en+1, ei] =
a2,4 := −a − b, bk−1,1 := −ak−1,1, (5 ≤ k ≤ n) =⇒ [e1, en+1] = ae1 + a2,1e2 +
k=4 ak,1ek + bn,1en, [en+1, e4] = (a − b)e2 − (a + b)e4 −
k=5 ak−1,3ek, [en+1, e1] =
L[en+1, e1]
L[en+1, en+1]
12.
13.
14.
a1,n+1 = 0, ak,n+1 = 0, (3 ≤ k ≤ n − 1) =⇒ [en+1, en+1] = a2,n+1e2 + an,n+1en.
an,n+1 = 0, (because b 6= (3 − n)a) =⇒ [en+1, en+1] = a2,n+1e2.
L[e3, en+1]
bn,3 := −an,3, (because a 6= 0), b2,3 := a2,3 + 2a4,3, (because b 6= 0) =⇒ [en+1, e3] =
15.
Le1 ([en+1, en+1])
bn,1 := −an,1, B2,1 := (2b−a)a2,1+2b·a4,1+2(a−b)(a2,3+a4,3)
, (because a 6= 0) =⇒
k=4 ak,3ek.
(a2,3 + 2a4,3)e2 − be3 −Pn
[en+1, e1] = −ae1 + B2,1e2 + (a − b)e3 −Pn
a
k=4 ak,1ek.
Theorem 5.2.2. Applying the technique of "absorption" (see Section 3.2), we can further
simplify the algebras in each of the cases in Theorem 5.2.1 as follows:
44
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Len+1 =
(1) If a1,3 = 0, b 6= −a, a 6= 0, b 6= 0, (n = 4) or b 6= (3 − n)a, a 6= 0, b 6= 0, (n ≥ 5),
then we have the following brackets for the algebra:
n
Xk=5
n
ak,3ek,
Xk=6
ak−1,3ek, [ei, en+1] = ((i − 3)a + b) ei +
[e1, en+1] = ae1 + a2,1e2 + (b − a)e3, [e3, en+1] = a2,3e2 + be3 +
[e4, en+1] = (a + b)(e4 − e2) +
[en+1, e1] = −ae1 + B2,1e2 + (a − b)e3, [en+1, e2] = −2be2, [en+1, e3] = a2,3e2 − be3−
Xk=5
Xk=i+2
ak,3ek, [en+1, e4] = (a − b)e2 − (a + b)e4 −
ak−i+3,3ek, (5 ≤ i ≤ n); where B2,1 :=
Xk=6
(2b − a)a2,1 + 2(a − b)a2,3
ak−1,3ek, [en+1, ei] = ((3 − i)a − b) ei−
Xk=i+2
a
n
n
n
n
,
ak−i+3,3ek,
0
0
0
0
0
0
...
...
0
0
.
0
−a
B2,1 −2b
a−b
0
a2,3
−b
0
−a5,3
·
·
...
0
a−b
0
−a−b
−a5,3
·
...
0
0
0
0
0
...
0
0
0
0
...
0
0
0
0
−2a−b
0
0
0
0
0
...
···
···
···
0
0
0
0
0
−3a−b
0
...
···
···
···
···
···
···
...
...
0
0
0
0
0
0
0
0
0
0
0
0
...
...
(5−n)a−b
...
0
...
...
0
0 −an−2,3 −an−3,3
0 −an−1,3 −an−2,3
0
−an,3 −an−1,3
··· −a5,3
···
···
·
· −a5,3
0
(4−n)a−b
·
−a5,3
0
(3−n)a−b
(2) If a1,3 = 0, b := −a, a 6= 0, (n = 4) or b := (3 − n)a, a 6= 0, (n ≥ 5), then the
brackets for the algebra are as follows:
n
n
ak,3ek,
Xk=6
ak−1,3ek, [ei, en+1] = (i − n) aei +
[e1, en+1] = ae1 + a2,1e2 − (n − 2)ae3, [e3, en+1] = a2,3e2 + (3 − n)ae3 +
[e4, en+1] = (n − 4)a(e2 − e4) +
[en+1, en+1] = an,n+1en, [en+1, e1] = −ae1 + B2,1e2 + (n − 2)ae3, [en+1, e2] = (2n − 6)ae2,
[en+1, e3] = a2,3e2 + (n − 3)ae3 −
ak,3ek, [en+1, e4] = (n − 2)ae2 + (n − 4)ae4−
Xk=6
Xk=i+2
where B2,1 := (5 − 2n)a2,1 + 2(n − 2)a2,3,
Xk=5
ak−1,3ek, [en+1, ei] = (n − i) aei −
ak−i+3,3ek, (5 ≤ i ≤ n),
Xk=5
Xk=i+2
ak−i+3,3ek,
n
n
n
n
Len+1 =
(n−2)a
0
0
0
0
0
0
0
0
...
0
0
0
0
...
0
0
0
−a
B2,1
0
0
0
(2n−6)a
a2,3
(n−2)a
(n−3)a
0
0
(n−4)a
−a5,3
·
0
(n−5)a
−a5,3
0
0
0
0
0
0
0
0
0
···
···
···
···
···
···
·
...
·
...
−an−2,3 −an−3,3
−an−1,3 −an−2,3
−an,3 −an−1,3
0
...
···
···
···
(n−6)a
0
...
···
···
···
...
...
−a5,3
...
· −a5,3
·
0
2a
0
0 0
a 0
· −a5,3 0 0
0
0
0
0
0
0
...
...
0 0
0 0
0 0
0 0
0 0
0 0
...
...
...
...
.
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
45
(3) If a1,3 = 0, a = 0, b 6= 0, (n = 4) or a = 0 and b 6= 0, (n ≥ 5), then
[e1, en+1] = a2,3e2 + be3, [e3, en+1] = a2,3e2 + be3 +
n
ak,3ek, [e4, en+1] = b (e4 − e2) +
Xk=5
n
n
n
ak−1,3ek, [ei, en+1] = bei +
Xk=6
[en+1, e3] = a2,3e2 − be3 −
Xk=i+2
[en+1, ei] = −bei −
n
0
ak−i+3,3ek, [en+1, e1] = b2,1e2 − be3, [en+1, e2] = −2be2,
Xk=i+2
ak,3ek, [en+1, e4] = −b (e2 + e4) −
Xk=5
ak−i+3,3ek, (5 ≤ i ≤ n),
n
Xk=6
ak−1,3ek,
0
b2,1 −2b
−b
0
0
0
0
0
0
0
0
a2,3
−b
0
−a5,3
·
0
−b
0
−b
0
−a5,3
0
0
0
0
0
0
0
0
−b
0
0 −b
···
···
···
···
···
···
0
...
·
...
... 0
...
...
...
...
·
...
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0
−an,3 −an−1,3 ···
··· −a5,3
···
···
·
· −a5,3
0
0
−b
0
0 −b 0
· −a5,3 0 −b
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
...
...
...
.
Len+1 =
0
...
0
0
0
n
n
ak−1,3ek, [ei, en+1] = (i − 3)aei +
(4) If a1,3 = 0, b = 0, a 6= 0, (n = 4) or b = 0, a 6= 0, (n ≥ 5), then
[e1, en+1] = ae1 + a2,1e2 − ae3, [e3, en+1] = a2,3e2 +
Xk=6
[en+1, e1] = −ae1 + (a2,3 − a2,1 + b2,3) e2 + ae3, [en+1, e3] = b2,3e2 −
Xk=i+2
[en+1, e4] = a(e2 − e4) −
ak−1,3ek, [en+1, ei] = (3 − i)aei −
Xk=i+2
Xk=6
Xk=5
n
n
n
ak−i+3,3ek, [en+1, en+1] = a2,n+1e2,
n
ak,3ek,
Xk=5
ak−i+3,3ek, (5 ≤ i ≤ n),
ak,3ek, [e4, en+1] = a(e4 − e2)+
Len+1 =
0
−a
0
a2,3−a2,1+b2,3 0
0
0
0 −a5,3
0
a
0
0
0
b2,3
0
0
·
0
...
·
...
0
...
0
0
0
In the remaining cases a1,3 := c.
0
a
0
−a
0
·
...
0
0
0
0
0
0
0
0
−2a 0
...
0
...
···
···
···
···
···
···
...
...
−a5,3
0 −3a
0
0
0
0
0
0
...
...
(5−n)a
...
0
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0 −an,3 −an−1,3 ···
··· −a5,3
···
···
·
· −a5,3
0
(4−n)a
·
−a5,3
0
(3−n)a
0
0
0
0
0
0
...
...
0
0
.
(5) If b 6= −a, a 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = ae1 + A2,1e2 + (b + c − a)e3, [e3, e5] = ce1 + a2,3e2 + be3, [e4, e5] = (a + b)e4−
(a + b)e2, [e5, e1] = −ae1 + B2,1e2 + (a − b − c)e3, [e5, e2] = −2(b + c)e2, [e5, e3] = −ce1+
(a + c)a2,3 + (c − a)b2,3
B2,3e2 − be3, [e5, e4] = (a − 2c − b)e2 − (a + b)e4; where A2,1 :=
,
B2,1 :=
(3a − 2b − c)a2,3 + (a − 2b − c)b2,3
2a
(a + c)a2,3 + c · b2,3
,
and B2,3 :=
0
0
0
0
0
0
...
...
0
a
Le5 = " −a
0
B2,1 −2(b+c) B2,3 a−2c−b
a−b−c
0
2a
−c
−b
0
0
0
0
0
−a−b # .
46
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(6) If b := −a, a 6= 0, a 6= c, c 6= 0, then
[e1, e5] = ae1 + A2,1e2 + (c − 2a)e3, [e3, e5] = ce1 + a2,3e2 − ae3, [e5, e5] = a4,5e4,
[e5, e1] = −ae1 + B2,1e2 + (2a − c)e3, [e5, e2] = 2(a − c)e2, [e5, e3] = −ce1 + B2,3e2 + ae3,
(a + c)a2,3 + c · b2,3
[e5, e4] = 2(a − c)e2; where A2,1 :=
,
(5a − c)a2,3 + (3a − c)b2,3
B2,1 :=
,
# .
(a + c)a2,3 + (c − a)b2,3
Le5 = " −a
B2,1 2(a−c) B2,3 2(a−c)
2a−c
, B2,3 :=
2a
2a
−c
a
0
0
0
0
a
0
0
0
0
(7) If b := −c, c 6= 0, a 6= c, a 6= 0, then
[e1, e5] = ae1 + a2,1e2 − ae3, [e3, e5] = ce1 + a2,3e2 − ce3, [e4, e5] = (c − a)(e2 − e4),
[e5, e5] = a2,5e2, [e5, e1] = −ae1 + (a2,3 − a2,1 + b2,3) e2 + ae3, [e5, e3] = −ce1 + b2,3e2 + ce3,
[e5, e4] = (c − a)(e4 − e2),
Le5 = "
a2,3−a2,1+b2,3 0 b2,3 a−c
−a
a
0
0
0 −c
0 0 c−a# .
0 c
0
(8) If a = 0, b = 0, c 6= 0, then
[e1, e5] = a2,1e2 + ce3, [e3, e5] = ce1 + (2a2,1 − b2,3) e2, [e5, e5] = a4,5e4,
[e5, e1] = (3a2,1 − 2b2,3) e2 − ce3, [e5, e2] = −2ce2, [e5, e3] = −ce1 + b2,3e2, [e5, e4] = −2ce2,
Le5 = "
0
−c
0
0 −c
0
0
0
0
0
0 # .
0
3a2,1−2b2,3 −2c b2,3 −2c
(9) If a = 0, b 6= 0, b 6= −c, c 6= 0, then
[e1, e5] = A2,1e2 + (b + c)e3, [e3, e5] = ce1 + a2,3e2 + be3, [e4, e5] = b(e4 − e2),
[e5, e1] = B2,1e2 − (b + c)e3, [e5, e2] = −2(b + c)e2, [e5, e3] = −ce1 + B2,3e2 − be3,
[e5, e4] = −(2c + b)e2 − be4; where A2,1 :=
B2,1 :=
4(b + c)
(2b + 3c)a2,3 + c · b2,3
(4b + 5c)a2,3 + c · b2,3
and B2,3 :=
,
,
2(b + c)
(2b + 3c)a2,3 − (2b + c)b2,3
−b # .
(10) If a = 0, b := −c, c 6= 0, then
B2,1 −2(b+c) B2,3 −2c−b
−b−c
Le5 = " 0
4(b + c)
−c
0
−b
0
0
0
0
0
0
[e1, e5] = (a2,3 + b2,3 − b2,1) e2, [e3, e5] = ce1 + a2,3e2 − ce3, [e4, e5] = c(e2 − e4),
[e5, e5] = a2,5e2, [e5, e1] = b2,1e2, [e5, e3] = −ce1 + b2,3e2 + ce3, [e5, e4] = c(e4 − e2),
Le5 = " 0 0 −c 0
b2,1 0 b2,3 −c
0
c # .
0 0 c
0 0 0
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
47
Proof.
(1) The right (not a derivation) and the left (a derivation) multiplication op-
erators restricted to the nilradical are given below:
Ren+1 =
0
0
0 a2,3 −a−b
0
a
a2,1
b−a 0
a4,1
a5,1
0 a4,3
0 a5,3
0
a+b
a4,3
2a+b
·
·
·
a5,3
a4,3
b
...
...
...
...
0
0
0
0
...
...
···
···
···
···
···
a4,3
...
...
...
...
0
0
0
0
0
...
...
...
ai,1
0 ai,3
ai−1,3 ai−2,3 ···
(i−3)a+b ···
...
...
...
0
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
an−1,1 0 an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3 (n−4)a+b
0 an,3 an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3
an,1
0
0
−a
B2,1 −2b a2,3+2a4,3
a−b
−a4,1
−a5,1
−a4,3
−a5,3
0
0
0
−b
−a−b
−a4,3 −2a−b
···
···
···
···
···
0
0
0
0
0
···
···
···
···
···
a4,3
0
0
0
0
0
(n−3)a+b
0
0
0
0
0
Len+1 =
·
·
·
−a5,3
−a4,3
...
...
...
...
...
...
...
...
...
...
...
−ai,1
−ai,3
−ai−1,3 −ai−2,3 ···
−a4,3
(3−i)a−b ···
,
0
0
0
0
0
...
...
...
0
0
...
...
...
0
0
...
...
...
0
...
...
...
0
...
...
...
0
.
...
...
0
...
...
...
...
0
a−b
0
...
...
0
0
0
0
...
...
−an−1,1
−an,1
0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 (4−n)a−b
0
−an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3
(3−n)a−b
k = ek, (1 ≤ k ≤ n), e′
• We start with applying the transformation e′
n+1 =
en+1 − a4,3e1 to remove a4,3 in Ren+1 and −a4,3 in Len+1 from the (i, i − 1)st
positions, where (4 ≤ i ≤ n), but it affects other entries as well, such as the
entry in the (2, 1)st position in Ren+1 and Len+1 that we change to a2,1 − a4,3
and B2,1 − a4,3, respectively. It also changes the entry in the (2, 3)rd position
in Len+1 to a2,3. At the same time, it affects the coefficient in front of e2 in
the bracket [en+1, en+1], which we change back to a2,n+1.
i = ei, (1 ≤ i ≤ n), e′
• Then we apply the transformation e′
n+1 = en+1 +
Pn−1
k=3 ak+1,1ek to remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries
in the (k + 1, 1)st positions, where (3 ≤ k ≤ n − 1). The transformation
changes the entry in the (2, 1)st position in Ren+1 to 2a4,1 + a2,1 − a4,3, the
entries in the (2, 3)rd positions in Ren+1 and Len+1 to a2,3 + a4,1. It also affects
the coefficient in front of e2 in [en+1, en+1], which we rename back by a2,n+1.
We assign 2a4,1 + a2,1 − a4,3 := a2,1 and a2,3 + a4,1 := a2,3. Then B2,1 − a4,3 :=
(2b−a)a2,1+2(a−b)a2,3
.
a
• Applying the transformation e′
j = ej, (1 ≤ j ≤ n), e′
n+1 = en+1 + a2,n+1
2b e2, we
remove the coefficient a2,n+1 in front of e2 in [en+1, en+1].
48
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(2) The right (not a derivation) and left (a derivation) multiplication operators re-
stricted to the nilradical are as follows:
a
a2,1
0
0
0
a2,3
0
(n−4)a
(2−n)a 0 (3−n)a
0
a4,1
a5,1
0
0
a4,3
a5,3
·
·
·
(4−n)a
a4,3
(5−n)a
a5,3
a4,3
0
0
0
0
...
...
···
···
···
···
···
0
0
0
0
0
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
···
a4,3
...
...
...
...
...
...
...
...
0
...
...
...
0
ai,1
ai,3
ai−1,3 ai−2,3
(i−n)a ···
Ren+1 =
...
...
0
...
...
0
0 an,3
an,1
0
(2n−6)a a2,3+2a4,3 (n−2)a
an−1,1 0 an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3 −a 0
an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3 a4,3 0
0
0
0
0
0
(n−3)a
−a4,3
−a5,3
(n−4)a
−a4,3
···
···
···
···
···
···
···
···
···
···
(n−5)a
0
0
0
0
0
0
0
0
0
0
0
0
0
0
−a
B2,1
(n−2)a
−a4,1
−a5,1
Len+1 =
·
·
·
−a5,3
−a4,3
−ai,1
−ai,3
−ai−1,3 −ai−2,3 ···
−a4,3
(n−i)a
···
...
...
...
...
...
...
...
...
...
,
0
0
0
0
0
0
0
0
0
0
.
...
...
...
0
...
...
...
0
...
...
...
0
...
...
...
0
...
...
...
...
...
...
...
...
...
...
0
...
...
0
0
−an−1,1
−an,1
−an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3
0
−an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 0
a
• We apply the transformation e′
k = ek, (1 ≤ k ≤ n), e′
n+1 = en+1 − a4,3e1
to remove a4,3 in Ren+1 and −a4,3 in Len+1 from the (i, i − 1)st positions,
where (4 ≤ i ≤ n), but it affects other entries as well, such as the entry
in the (2, 1)st position in Ren+1 and Len+1 that we change to a2,1 − a4,3 and
B2,1 − a4,3, respectively. It also changes the entry in the (2, 3)rd position in
Len+1 to a2,3. At the same time, it affects the coefficient in front of e2 in the
bracket [en+1, en+1], which we change back to a2,n+1.
i = ei, (1 ≤ i ≤ n), e′
• Then we apply the transformation e′
n+1 = en+1 +
Pn−1
k=3 ak+1,1ek to remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries
in the (k + 1, 1)st positions, where (3 ≤ k ≤ n−1). It changes the entry in the
(2, 1)st position in Ren+1 to 2a4,1+a2,1−a4,3, the entries in the (2, 3)rd positions
in Ren+1 and Len+1 to a2,3 + a4,1. It also affects the coefficient in front of e2 in
[en+1, en+1], which we rename back by a2,n+1. We assign 2a4,1 + a2,1 − a4,3 :=
a2,1 and a2,3 + a4,1 := a2,3. Then B2,1 − a4,3 := (5 − 2n)a2,1 + 2(n − 2)a2,3.
n+1 = en+1 − a2,n+1
• Finally the transformation e′
j = ej, (1 ≤ j ≤ n), e′
2(n−3)a e2
removes the coefficient a2,n+1 in front of e2 in [en+1, en+1].
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
49
(3) The right (not a derivation) and left (a derivation) multiplication operators re-
stricted to the nilradical are given below:
Ren+1 =
a5,3
a4,3
···
···
···
···
···
0
0
0
0
0
0
0
0
a2,3+a4,3−a4,1 0 a2,3
b
a4,1
a5,1
b
0
0 a4,3
0 a5,3
·
·
·
...
...
...
...
...
...
0
−b
0
b
a4,3
...
...
0
0
0
0
b
...
...
ai,1
0 ai,3
ai−1,3 ai−2,3 ···
...
...
a4,3
...
...
···
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
0
...
...
...
0
an−1,1
an,1
0
0
b2,1 −2b a2,3+2a4,3
−b
−b
−a4,1
−a5,1
−a4,3
−a5,3
0 an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3
0
0 an,3 an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3 a4,3 b
0
0
0
0
0
0
0
0
0
−b
0
−b
0
−b
···
···
···
···
···
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
−a4,3
b
Len+1 =
·
·
·
−a5,3
−a4,3
...
...
...
...
...
...
...
...
...
...
−ai,1
−ai,3
−ai−1,3 −ai−2,3 ···
−a4,3
−b
···
...
...
...
0
...
...
...
b
...
...
...
0
...
...
0
0
0
0
...
...
,
0
0
0
0
0
...
...
...
0
...
...
...
0
...
...
...
0
.
−an−1,1
−an,1
0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 −b
0
0
−an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 −b
k = ek, (1 ≤ k ≤ n), e′
• Applying the transformation e′
n+1 = en+1 − a4,3e1, we
remove a4,3 in Ren+1 and −a4,3 in Len+1 from the (i, i − 1)st positions, where
(4 ≤ i ≤ n), but the transformation affects other entries, such as the entry
in the (2, 1)st position in Ren+1 and Len+1, that we change to a2,3 − a4,1 and
b2,1 − a4,3, respectively. It also changes the entry in the (2, 3)rd position in
Len+1 to a2,3. At the same time, it affects the coefficient in front of e2 in the
bracket [en+1, en+1], which we change back to a2,n+1.
• Then we apply the transformation e′
n+1 = en+1 +
Pn−1
k=3 ak+1,1ek to remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries
in the (k + 1, 1)st positions, where (3 ≤ k ≤ n − 1). This transformation
changes the entry in the (2, 1)st position in Ren+1 as well as the entries in
i = ei, (1 ≤ i ≤ n), e′
the (2, 3)rd positions in Ren+1 and Len+1 to a2,3 + a4,1. It also affects the
coefficient in front of e2 in [en+1, en+1], which we rename back by a2,n+1. We
assign a2,3 + a4,1 := a2,3 and b2,1 − a4,3 := b2,1.
• Finally applying the transformation e′
j = ej, (1 ≤ j ≤ n), e′
n+1 = en+1 +
a2,n+1
2b e2, we remove the coefficient a2,n+1 in front of e2 in [en+1, en+1].
50
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(4) The right (not a derivation) and left (a derivation) multiplication operators re-
...
...
...
...
...
...
...
...
...
...
...
...
...
...
0
a
0
−a
...
...
,
0
0
0
0
0
...
...
...
0
...
...
...
0
0
...
...
...
0
a4,3
0
0
0
0
0
···
···
···
···
···
(n−3)a
0
0
0
0
0
...
...
...
0
.
stricted to the nilradical are as follows:
a
a2,1
−a
a4,1
a5,1
0
0
0 a2,3
0
0 a4,3
0 a5,3
0
0
−a
0
a
a4,3
···
···
···
···
···
0
0
0
0
0
···
···
···
···
···
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Ren+1 =
·
·
·
a5,3
a4,3
ai,1
0 ai,3
ai−1,3 ai−2,3 ···
(i−3)a ···
...
...
a4,3
...
...
...
...
...
...
...
...
0
...
...
...
0
an−1,1 0 an−1,3 an−2,3 an−3,3 ··· an−i+3,3 an−i+2,3 ··· a4,3 (n−4)a
an,1
0 an,3 an−1,3 an−2,3 ··· an−i+4,3 an−i+3,3 ··· a5,3
0
0
0
0
2a
...
...
0
−a
0
a2,3−a2,1+b2,3 0
0
0 −a4,3
0 −a5,3
−a4,1
−a5,1
b2,3
0
a
−a4,3
−2a
Len+1 =
·
·
·
−a5,3
−a4,3
0
0
0
0
...
...
...
...
···
···
···
···
···
...
...
0
0
0
0
0
...
...
...
−ai,1
0 −ai,3 −ai−1,3 −ai−2,3 ···
−a4,3
(3−i)a
···
−an−1,1
−an,1
0 −an−1,3 −an−2,3 −an−3,3 ··· −an−i+3,3 −an−i+2,3 ··· −a4,3 (4−n)a
0 −an,3 −an−1,3 −an−2,3 ··· −an−i+4,3 −an−i+3,3 ··· −a5,3 −a4,3 (3−n)a
0
k = ek, (1 ≤ k ≤ n), e′
• We continue with the transformation e′
n+1 = en+1 −
a4,3e1 to remove a4,3 in Ren+1 and −a4,3 in Len+1 from the (i, i− 1)st positions,
where (4 ≤ i ≤ n), but other entries are affected as well, such as the entry
in the (2, 1)st position in Ren+1 and Len+1, that we change to a2,1 − a4,3 and
a2,3−a2,1−a4,3 +b2,3, respectively. The transformation also changes the entry
in the (2, 3)rd position in Len+1 to b2,3 − 2a4,3 and affects the coefficient in
front of e2 in the bracket [en+1, en+1], which we change back to a2,n+1.
• Applying the transformation e′
i = ei, (1 ≤ i ≤ n), e′
n+1 = en+1+Pn−1
k=3 ak+1,1ek,
we remove ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries in the (k +
1, 1)st positions, where (3 ≤ k ≤ n − 1). This transformation changes the
entry in the (2, 1)st position in Ren+1 to 2a4,1 + a2,1 − a4,3, the entries in
the (2, 3)rd positions in Ren+1 and Len+1 to a2,3 + a4,1 and a4,1 + b2,3 − 2a4,3,
respectively. It also affects the coefficient in front of e2 in [en+1, en+1], that we
rename back by a2,n+1. We assign 2a4,1 + a2,1 − a4,3 := a2,1, a2,3 + a4,1 := a2,3
and a4,1 + b2,3 − 2a4,3 := b2,3.
(5) We apply the transformation e′
a2,1A4,3 + 2a4,1A4,3− A2
a2,3 + a4,1 := a2,3 and b2,3 − 2a2,1 − 3a4,1 := b2,3.
4,3− a2,3a4,1− a2
i = ei, (1 ≤ i ≤ 4), e′
5 = e5−A4,3e1− 1
2(b+c) (B2,1A4,3+
4,1− a4,1b2,3− a2,5)e2 + a4,1e3. Then we assign
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
51
(6) One applies the transformation e′
i = ei, (1 ≤ i ≤ 4), e′
5 = e5−A4,3e1+ 1
2(a−c) (B2,1A4,3+
a2,1A4,3 + 2a4,1A4,3− A2
a2,3 + a4,1 := a2,3 and b2,3 − 2a2,1 − 3a4,1 := b2,3.
4,3− a2,3a4,1− a2
4,1− a4,1b2,3− a2,5)e2 + a4,1e3. Then we assign
(7) We apply the transformation e′
5 = e5 − a4,3e1 + a4,1e3 and
rename the coefficient in front of e2 in [e5, e5] back by a2,5. Then we assign a2,1 +
2a4,1 − a4,3 := a2,1, a2,3 + a4,1 := a2,3 and b2,3 + a4,1 − 2a4,3 := b2,3.
i = ei, (1 ≤ i ≤ 4), e′
(8) The transformation is as follows: e′
i = ei, (1 ≤ i ≤ 4), e′
5 = e5 − a4,3e1 +
4,3 + 2a4,3b2,3 + a2,5)e2 + a4,1e3. We
1
2c(2a2,1a4,1 − 4a2,1a4,3 + 3a2
assign a2,1 + 2a4,1 − a4,3 := a2,1 and b2,3 + a4,1 − 2a4,3 := b2,3.
5 = e5−A4,3e1+ 1
i = ei, (1 ≤ i ≤ 4), e′
4,1 − 6a4,1a4,3 − a2
(9) The transformation is e′
4,3−a2,3A4,3−
4,1 + 2a4,1b2,3 + 2a2,5)e2 + a4,1e3. We
4(b+c) (2A2
2a4,1A4,3 − 2b2,1A4,3 − b2,3A4,3 + 2a2,3a4,1 + 2a2
assign a2,3 + a4,1 := a2,3 and b2,3 + a4,1 − 2b2,1 := b2,3.
i = ei, (1 ≤ i ≤ 4), e′
(10) We apply the transformation e′
5 = e5 − a4,3e1 + a4,1e3 and
rename the coefficient in front of e2 in [e5, e5] back by a2,5. We assign a2,3 + a4,1 :=
a2,3, b2,1 − a4,3 := b2,1 and b2,3 + a4,1 − 2a4,3 := b2,3.
(cid:3)
52
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Theorem 5.2.3. There are eight solvable indecomposable left Leibniz algebras up to iso-
morphism with a codimension one nilradical L 4, (n ≥ 4), which are given below:
(i) ln+1,1 : [e1, en+1] = e1 + (a − 1)e3, [e3, en+1] = ae3, [e4, en+1] = (a + 1)(e4 − e2),
[ei, en+1] = (a + i − 3) ei, [en+1, e1] = −e1 − (a − 1)e3, [en+1, e2] = −2ae2, [en+1, e3] = −ae3,
[en+1, e4] = (1 − a)e2 − (a + 1)e4, [en+1, ei] = (3 − i − a) ei, (5 ≤ i ≤ n),
(ii) ln+1,2 : [e1, en+1] = e1 + (2 − n)e3, [e3, en+1] = (3 − n)e3, [e4, en+1] = (n − 4)(e2 − e4),
[ei, en+1] = (i − n) ei, [en+1, en+1] = en, [en+1, e1] = −e1 + (n − 2)e3, [en+1, e2] = (2n − 6)e2,
[en+1, e3] = (n − 3)e3, [en+1, e4] = (n − 2)e2 + (n − 4)e4, [en+1, ei] = (n − i) ei,
(5 ≤ i ≤ n),
(iii) ln+1,3 : [e1, en+1] = e3, [e4, en+1] = −e2, [ei, en+1] = ei + ǫei+2 +
bk−i−2ek,
n
Xk=i+3
n
Xk=i+3
bk−i−2ek,
[en+1, e1] = −e3, [en+1, e2] = −2e2, [en+1, e4] = −e2, [en+1, ei] = −ei − ǫei+2 −
(ǫ = 0, 1, 3 ≤ i ≤ n),
(iv) gn+1,4 : [e1, en+1] = e1 − e3, [e3, en+1] = f e2, [e4, en+1] = −e2 + e4, [ei, en+1] = (i − 3)ei,
[en+1, en+1] = ǫe2, [en+1, e1] = −e1 + (d + f ) e2 + e3, [en+1, e3] = de2,
[en+1, e4] = e2 − e4, [en+1, ei] = (3 − i)ei, (5 ≤ i ≤ n; ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0),
(v) l5,5 : [e1, e5] = ae1 + (b − a + 1)e3, [e3, e5] = e1 + be3, [e4, e5] = (a + b)(e4 − e2),
[e5, e1] = −ae1 + (a − b − 1)e3, [e5, e2] = −2(b + 1)e2, [e5, e3] = −e1 − be3,
[e5, e4] = (a − b − 2)e2 − (a + b)e4, (if b = −1, then a 6= 1),
(vi) l5,6 : [e1, e5] = ae1 + (1 − 2a)e3, [e3, e5] = e1 − ae3, [e5, e5] = e4, [e5, e1] = −ae1+
(2a − 1)e3, [e5, e2] = 2(a − 1)e2, [e5, e3] = ae3 − e1, [e5, e4] = 2(a − 1)e2, (a 6= 1),
(vii) g5,7 : [e1, e5] = a(e1 − e3), [e3, e5] = e1 + f e2 − e3, [e4, e5] = (1 − a) (e2 − e4) ,
[e5, e5] = ǫe2, [e5, e1] = −ae1 + (d + f ) e2 + ae3, [e5, e3] = −e1 + de2 + e3,
[e5, e4] = (a − 1) (e2 − e4) , (ǫ = 0, 1; if ǫ = 0, then d2 + f 2 6= 0; a 6= 1),
(viii) g5,8 : [e1, e5] = ce2, [e3, e5] = e1 − e3, [e4, e5] = e2 − e4, [e5, e5] = ǫe2, [e5, e1] = (c + d) e2,
[e5, e3] = −e1 + (d + 2c) e2 + e3, [e5, e4] = e4 − e2, (c 6= 0, ǫ = 0, 1).
Proof. One applies the change of basis transformations keeping the nilradical L 4 given
in (4.0.1) unchanged.
(1) We have the right (not a derivation) and the left (a derivation) multiplication
operators restricted to the nilradical are as follows:
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
53
Ren+1 =
Len+1 =
a 0
0
0
a2,1 0 a2,3 −a−b
b−a 0
0
0 a5,3
0
a+b
0
0
0
a5,3
b
0
0
0
·
·
...
0 an−2,3 an−3,3
0 an−1,3 an−2,3
0 an,3 an−1,3
0
...
0
...
·
...
0
0
0
−a
B2,1 −2b
a−b
0
0
a2,3
−b
0
−a5,3
·
0
0
0
0
2a+b
0
...
···
···
···
0
a−b
0
−a−b
−a5,3
·
...
·
...
0 −an−2,3 −an−3,3
0 −an−1,3 −an−2,3
0 −an,3 −an−1,3
0
0
0
0
0
...
0
0
0
0
...
0
0
0
,
0
0
0
0
0
0
...
...
0
(n−4)a+b
0
0
0
0
0
0
...
...
0
0
0
(n−3)a+b
0
0
0
0
0
0
...
...
(5−n)a−b
...
0
0
0
0
0
0
0
...
...
0
0 (n−5)a+b
0
0
0
0
0
···
···
···
···
···
3a+b ···
0
0
0
0
0
0
...
...
a5,3
·
·
0
...
···
···
···
0
0
0
0
0
...
...
a5,3
·
0
0
0
0
0
0
...
...
...
0
a5,3
···
···
···
···
···
···
...
...
···
··· −a5,3
···
···
···
···
1 = e1, e′
·
0
−2a−b
−3a−b
0
0
0
0
0
0
...
...
0
0
.
· −a5,3
0
(4−n)a−b
·
−a5,3
0
(3−n)a−b
2 = e2, e′
i = ei− ak−i+3,3
• We apply the transformation e′
(k−i)a ek, (3 ≤ i ≤ n−
2, i+2 ≤ k ≤ n, n ≥ 5), e′
j = ej, (n−1 ≤ j ≤ n+1), where k is fixed, renaming
all the affected entries back. This transformation removes a5,3, a6,3, ..., an,3 in
Ren+1 and −a5,3,−a6,3, ...,−an,3 in Len+1. Besides it introduces the entries in
the (5, 1)st, (6, 1)st, ..., (n, 1)st positions in Ren+1 and Len+1, which we set to
be a5,1, a6,1, ..., an,1 and −a5,1,−a6,1, ...,−an,1, respectively.
(I) Suppose b 6= a
2 .
• Applying the transformation e′
2 = e2, e′
3 =
1 = e1 + 1
i = ei, (4 ≤ i ≤ n, n ≥ 4), e′
b
2b−a (cid:16)B2,1 + (b−a)a2,3
(cid:17) e2, e′
n+1 = en+1 − a5,1e2 +Pn−1
we remove a2,1 and B2,1 from the (2, 1)st positions in Ren+1 and Len+1, re-
e3 + a2,3
b e2, e′
k=4 ak+1,1ek,
spectively. This transformation also removes a2,3 from the (2, 3)rd positions
in Ren+1 and Len+1 as well as ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the
entries in the (k + 1, 1)st positions, where (4 ≤ k ≤ n − 1). (For n = 4, we
direct to the Remark 5.1.1.)
• Then we scale a to unity applying the transformation e′
4), e′
i = ei, (1 ≤ i ≤ n, n ≥
a by b, we obtain a continuous family of Leibniz
a . Renaming b
n+1 = en+1
(5.2.1)
algebras:
[e1, en+1] = e1 + (b − 1)e3, [e3, en+1] = be3, [e4, en+1] = (b + 1)(e4 − e2),
[ei, en+1] = (b + i − 3) ei, [en+1, e1] = −e1 − (b − 1)e3,
[en+1, e2] = −2be2, [en+1, e3] = −be3, [en+1, e4] = (1 − b)e2 − (b + 1)e4,
[en+1, ei] = (3 − i − b) ei, (5 ≤ i ≤ n), (b 6= 0, b 6=
, b 6= 3 − n).
1
2
54
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
(II) Suppose b := a
2 . We have that B2,1 = a2,3.
• We apply the transformation e′
1 = e1+ a2,1+a2,3
a
e2, e′
2 = e2, e′
3 = e3+ 2a2,3
a e2, e′
i =
ei, (4 ≤ i ≤ n, n ≥ 4), e′
the (2, 1)st, (2, 3)rd positions in Len+1 and from the (2, 3)rd position in Ren+1.
n+1 = en+1−a5,1e2 +Pn−1
k=4 ak+1,1ek to remove a2,3 from
This transformation also removes a2,1 from the (2, 1)st position in Ren+1 as
well as ak+1,1 and −ak+1,1 from the entries in the (k + 1, 1)st positions in
Ren+1 and Len+1, respectively, where (4 ≤ k ≤ n − 1). (For n = 4, we refer to
Remark 5.1.1.)
• To scale a to unity, we apply the transformation e′
4), e′
n+1 = en+1
and obtain a limiting case of (5.2.1) with b = 1
i = ei, (1 ≤ i ≤ n, n ≥
2 given below:
2(cid:19) ei,
5
(e4 − e2), [ei, en+1] = (cid:18)i −
e2
2 −
, [en+1, e4] =
e3
2
3e4
2
,
e3
[e1, en+1] = e1 −
2
[en+1, e1] = −e1 +
[en+1, ei] = (cid:18)5
a
, [e3, en+1] =
e3
2
2 − i(cid:19) ei, (5 ≤ i ≤ n).
e3
2
, [e4, en+1] =
3
2
, [en+1, e2] = −e2, [en+1, e3] = −
operators restricted to the nilradical are as follows:
(2) We have the right (not a derivation) and the left (a derivation) multiplication
0
0
0
0
0
0
,
0 0
0 0
0 0
0 0
0 0
0 0
...
...
...
...
...
...
.
0
0
0
0
0
0
0 0
0 0
0 0
0 0
0 0
0 0
...
...
...
...
...
...
a5,3
0 −2a 0 0
0 −a 0
a5,3 0 0
·
a
a2,1
0
0
0
a2,3
0
(n−4)a
(2−n)a 0 (3−n)a
0
0
0
0
0
0
0
0
a5,3
·
(4−n)a
0
a5,3
0
0
0
0
0
0
0
0
0
···
···
···
···
···
(6−n)a ···
(5−n)a
0
...
···
···
···
0
0
...
0
0
0
0
...
·
...
·
...
0 an−2,3 an−3,3
0 an−1,3 an−2,3
0 an,3
an−1,3
0
0
(2n−6)a
a2,3
(n−2)a
(n−3)a
0
0
(n−4)a
...
...
a5,3
·
·
0
0
0
0
0
...
···
···
···
···
···
···
0
...
···
···
···
0
0
0
0
0
...
−a5,3
·
0
(n−5)a
−a5,3
(n−6)a
−a
B2,1
(n−2)a
0
0
0
0
0
0
0
Len+1 =
Ren+1
0
...
0
0
0
0
...
0
0
0
·
...
·
...
−an−2,3 −an−3,3
−an−1,3 −an−2,3
−an,3 −an−1,3
···
···
···
1 = e1, e′
...
...
−a5,3
...
· −a5,3
·
0
2a
0
0 0
a 0
· −a5,3 0 0
0
...
···
···
···
2 = e2, e′
i = ei− ak−i+3,3
• We apply the transformation e′
(k−i)a ek, (3 ≤ i ≤ n−
2, i+2 ≤ k ≤ n, n ≥ 5), e′
j = ej, (n−1 ≤ j ≤ n+1), where k is fixed, renaming
all the affected entries back. This transformation removes a5,3, a6,3, ..., an,3 in
Ren+1 and −a5,3,−a6,3, ...,−an,3 in Len+1. Moreover it introduces the entries
in the (5, 1)st, (6, 1)st, ..., (n, 1)st positions in Ren+1 and Len+1, which we set
to be a5,1, a6,1, ..., an,1 and −a5,1,−a6,1, ...,−an,1, respectively.
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
55
• Applying the transformation e′
e3 + a2,3
1 = e1+ 1
i = ei, (4 ≤ i ≤ n, n ≥ 4), e′
(3−n)a e2, e′
k=4 ak+1,1ek,
we remove a2,1 and B2,1 from the (2, 1)st positions in Ren+1 and Len+1, respec-
(5−2n)a (cid:16)B2,1 + (2−n)a2,3
3−n (cid:17) e2, e′
n+1 = en+1 − a5,1e2 +Pn−1
2 = e2, e′
3 =
tively. We also remove a2,3 from the (2, 3)rd positions in Ren+1 and Len+1 as
well as ak+1,1 in Ren+1 and −ak+1,1 in Len+1 from the entries in the (k + 1, 1)st
positions, where (4 ≤ k ≤ n − 1). (See Remark 5.1.1.)
• To scale a to unity, we apply the transformation e′
i = ei, (1 ≤ i ≤ n), e′
n+1 =
en+1
a
renaming the coefficient an,n+1
a2
in front of en in [en+1, en+1] back by an,n+1.
We obtain a Leibniz algebra given below:
[e1, en+1] = e1 + (2 − n)e3, [e3, en+1] = (3 − n)e3, [e4, en+1] = (n − 4)(e2 − e4),
[ei, en+1] = (i − n) ei, [en+1, en+1] = an,n+1en, [en+1, e1] = −e1 + (n − 2)e3,
[en+1, e2] = (2n − 6)e2, [en+1, e3] = (n − 3)e3, [en+1, e4] = (n − 2)e2 + (n − 4)e4,
[en+1, ei] = (n − i) ei, (5 ≤ i ≤ n),
If an,n+1 = 0, then we have a limiting case of (5.2.1) with b = 3− n. If an,n+1 6= 0,
then we scale it to 1. It gives us the algebra ln+1,2.
(3) We have the right (not a derivation) and the left (a derivation) multiplication
operators restricted to the nilradical are as follows:
Ren+1 =
0
0
...
0 0
0
a2,3 0 a2,3
b
0
0
b
0 0
0 0 a5,3
0 0
·
0 0
...
...
·
...
0
−b
0
b
0
a5,3
·
...
0 0 an−2,3 an−3,3 ···
0 0 an−1,3 an−2,3 ···
0 0 an,3 an−1,3 ···
0
0
0
0
b
0
0
0
0
0
0
b
···
···
···
···
···
···
... 0
...
...
...
·
·
··· a5,3
···
···
0
0
0
0
0
0
0
0
0
−b
0 −b
···
···
···
···
···
···
... 0
...
...
...
Len+1 =
0
−b
0
−b
0
·
...
−a5,3
0
b2,1 −2b
−b
0
0
0
0
0
0
0
0
a2,3
−b
0
−a5,3
·
·
...
0
...
0
0
0
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0 −an,3 −an−1,3 ···
··· −a5,3
···
···
·
· −a5,3
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
0 0 0
...
...
...
...
...
...
b
0
0 0
b 0
a5,3 0 b
...
0
a5,3
·
0
0
0
0
0
0
,
0
0
0
0
0
0
...
...
...
...
0
0
0
0
0
0
...
...
...
0
0
−b
0
0 −b 0
· −a5,3 0 −b
e2, e′
2 = e2, e′
.
• Applying the transformation e′
i =
ei, (4 ≤ i ≤ n + 1), we remove b2,1 from the (2, 1)st position in Len+1 and a2,3
from the (2, 1)st position in Ren+1 and from the (2, 3)rd positions in Ren+1 and
1 = e1 + b2,1+a2,3
3 = e3 + a2,3
b e2, e′
2b
Len+1 keeping other entries unchanged.
56
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
• To scale b to unity, we apply the transformation e′
i = ei, (1 ≤ i ≤ n), e′
n+1 =
en+1
b
. Then we rename a5,3
b , a6,3
b , ..., an,3
b by a5,3, a6,3, ..., an,3, respectively. We
obtain a Leibniz algebra
n
[e1, en+1] = e3, [e4, en+1] = −e2, [ei, en+1] = ei +
Xk=i+2
[en+1, e2] = −2e2, [en+1, e4] = −e2, [en+1, ei] = −ei −
n
ak−i+3,3ek, [en+1, e1] = −e3,
Xk=i+2
ak−i+3,3ek, (3 ≤ i ≤ n),
If a5,3 6= 0, (n ≥ 5), then we scale it to 1. We also rename all the affected
entries back and then we rename a6,3, ..., an,3 by b1, ..., bn−5, respectively. We
combine with the case when a5,3 = 0 and obtain a Leibniz algebra ln+1,3.
Remark 5.2.1. If n = 4, then ǫ = 0.
(4) We have the right (not a derivation) and the left (a derivation) multiplication
operators restricted to the nilradical are as follows:
Ren+1 =
a 0
0
a2,1 0 a2,3
−a 0
0 0
0 0 a5,3
0 0
0
0
·
0 0
...
...
·
...
0
−a
0
a
0
a5,3
·
...
0
0
0
0
2a
0
···
0
···
0
···
0
···
0
0
···
3a ···
... 0
...
...
...
0 0 an−2,3 an−3,3 ···
0 0 an−1,3 an−2,3 ···
0 0 an,3 an−1,3 ···
0
0
0
0
0
0
...
...
(n−5)a
0
a5,3
...
0
a5,3
·
0
0
0
0
0
0
...
...
0
(n−4)a
0
0
0
0
0
0
...
...
0
0
0
(n−3)a
,
·
·
··· a5,3
···
···
0
0
0
0
0
0
0
0
−2a 0
...
0
...
···
···
···
···
···
···
...
...
0
a
0
−a
0
·
...
−a5,3
0 −3a
0
...
·
...
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0 −an,3 −an−1,3 ···
··· −a5,3
···
···
1 = e1, e′
·
0
0
0
0
0
0
...
...
(5−n)a
...
0
0
0
0
0
0
0
...
...
0
0
0
0
0
0
0
0
...
...
0
0
.
· −a5,3
0
(4−n)a
·
−a5,3
(3−n)a
−a
0
a2,3−a2,1+b2,3 0
0
0
0 −a5,3
0
0
b2,3
0
0
a
0
0
0
·
Len+1 =
0
...
0
0
0
2 = e2, e′
i = ei− ak−i+3,3
• We apply the transformation e′
(k−i)a ek, (3 ≤ i ≤ n−
2, i+2 ≤ k ≤ n, n ≥ 5), e′
j = ej, (n−1 ≤ j ≤ n+1), where k is fixed renaming
all the affected entries back. This transformation removes a5,3, a6,3, ..., an,3 in
Ren+1 and −a5,3,−a6,3, ...,−an,3 in Len+1. Besides it introduces the entries in
the (5, 1)st, (6, 1)st, ..., (n, 1)st positions in Ren+1 and Len+1, which we set to
be a5,1, a6,1, ..., an,1 and −a5,1,−a6,1, ...,−an,1, respectively.
• Applying the transformation e′
i = ei, (2 ≤ i ≤ n, n ≥
k=4 ak+1,1ek, we remove a2,1 from the (2, 1)st position
in Ren+1. This transformation changes the entry in the (2, 1)st position in
n+1 = en+1 + Pn−1
1 = e1 + a2,1
a e2, e′
4), e′
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
57
Len+1 to a2,3 + b2,3. It also removes ak+1,1 and −ak+1,1 from the entries in the
(k + 1, 1)st positions, where (4 ≤ k ≤ n− 1) in Ren+1 and Len+1, respectively.4
• We assign a2,3 := d and b2,3 := f and then we scale a to unity applying the
a and
a . Renaming d
transformation e′
i = ei, (1 ≤ i ≤ n, n ≥ 4), e′
n+1 = en+1
a , f
by d, f and a2,n+1, respectively, we obtain a Leibniz algebra, which is
a2,n+1
a2
right and left at the same time and a limiting case of (5.2.1) with b = 0, when
d = f = a2,n+1 = 0:
[e1, en+1] = e1 − e3, [e3, en+1] = de2, [e4, en+1] = e4 − e2, [ei, en+1] = (i − 3)ei,
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −e1 + (d + f ) e2 + e3, [en+1, e3] = f e2,
[en+1, e4] = e2 − e4, [en+1, ei] = (3 − i)ei, (5 ≤ i ≤ n).
Altogether (5.2.1) and all its limiting cases after replacing b with a give us a
Leibniz algebra ln+1,1. It remains to consider a continuous family of Leibniz
algebras given below and scale any nonzero entries as much as possible.
[e1, en+1] = e1 − e3, [e3, en+1] = de2, [e4, en+1] = e4 − e2, [ei, en+1] = (i − 3)ei,
[en+1, en+1] = a2,n+1e2, [en+1, e1] = −e1 + (d + f ) e2 + e3, [en+1, e3] = f e2,
[en+1, e4] = e2 − e4, [en+1, ei] = (3 − i)ei, (a2
2,n+1 + d2 + f 2 6= 0, 5 ≤ i ≤ n),
If a2,n+1 6= 0, then we scale it to 1. We also rename all the affected entries
back. Then we combine with the case when a2,n+1 = 0 and obtain a right
and left Leibniz algebra gn+1,4.
(5) Applying the transformation e′
1 = e1 + (2a−b)a2,3−b·b2,3
2a(b+c)
e2, e′
2 = e2, e′
3 = e3 +
(2a+c)a2,3+c·b2,3
2a(b+c)
4 = e4, e′
e′
e2,
5 = e5
c and renaming a
c and b
c by a and b, respectively, we obtain a
continuous family of Leibniz algebras given below:
(5.2.2)
[e1, e5] = ae1 + (b − a + 1)e3, [e3, e5] = e1 + be3, [e4, e5] = (a + b)(e4 − e2),
[e5, e1] = −ae1 + (a − b − 1)e3, [e5, e2] = −2(b + 1)e2, [e5, e3] = −e1 − be3,
[e5, e4] = (a − b − 2)e2 − (a + b)e4, (b 6= −a, a 6= 0, b 6= −1).
(6) We apply the transformation e′
e′
4 = e4, e′
5 = e5
c and rename a
1 = e1+ 3a2,3+b2,3
c and a4,5
2(c−a) e2, e′
c2 by a and a4,5, respectively, to obtain a
3 = e3+ (2a+c)a2,3+c·b2,3
2 = e2, e′
2a(c−a)
e2,
4Except this transformation it is the same as case (4) for the right Leibniz algebras.
58
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Leibniz algebra given below:
(5.2.3)
(5.2.4)
[e1, e5] = ae1 + (1 − 2a)e3, [e3, e5] = e1 − ae3, [e5, e5] = a4,5e4,
[e5, e1] = −ae1 + (2a − 1)e3, [e5, e2] = 2(a − 1)e2, [e5, e3] = ae3 − e1,
[e5, e4] = 2(a − 1)e2, (a 6= 0, a 6= 1),
which is a limiting case of (5.2.2) with b := −a when a4,5 = 0. If a4,5 6= 0, then we
scale it to 1 and obtain a continuous family of Leibniz algebras:
[e1, e5] = ae1 + (1 − 2a)e3, [e3, e5] = e1 − ae3, [e5, e5] = e4,
[e5, e1] = −ae1 + (2a − 1)e3, [e5, e2] = 2(a − 1)e2, [e5, e3] = ae3 − e1,
[e5, e4] = 2(a − 1)e2, (a 6= 0, a 6= 1).
(7) We apply the transformation e′
d := a·b2,3+c·a2,1
a
, f := a·a2,3−c·a2,1
a
Theorem 5.1.3.
1 = e1 + a2,1
i = ei, (2 ≤ i ≤ 5) and assign
. Then this case becomes the same as case (7) of
a e2, e′
(8) Applying the transformation e′
1 = e1 + 2a2,1−b2,3
c
e2, e′
2 = e2, e′
3 = e3 + a2,1
c e2, e′
4 =
e4, e′
5 = e5
c and renaming a4,5
c2 back by a4,5, we obtain a Leibniz algebra:
[e1, e5] = e3, [e3, e5] = e1, [e5, e5] = a4,5e4, [e5, e1] = −e3, [e5, e2] = −2e2, [e5, e3] = −e1,
[e5, e4] = −2e2,
which is a limiting case of (5.2.3) with a = 0. If a4,5 6= 0, then we scale a4,5 to
1 and obtain a limiting case of (5.2.4) with a = 0. Altogether (5.2.4) and all its
limiting cases give us the algebra l5,6.
(9) One applies the transformation e′
1 = e1 + (3b+4c)a2,3−b·b2,3
4(b+c)2
e2, e′
2 = e2, e′
3 = e3 +
(4b+5c)a2,3+c·b2,3
4(b+c)2
given below:
e2, e′
4 = e4, e′
5 = e5
c . Renaming b
c by b, we obtain a Leibniz algebra
[e1, e5] = (b + 1)e3, [e3, e5] = e1 + be3, [e4, e5] = b(e4 − e2), [e5, e1] = (−b − 1)e3,
[e5, e2] = −2(b + 1)e2, [e5, e3] = −e1 − be3, [e5, e4] = −(b + 2)e2 − be4, (b 6= 0, b 6= −1),
which is a limiting case of (5.2.2) with a = 0. Altogether (5.2.2) and all its limiting
cases give us the algebra l5,5.
(10) We apply the transformation e′
1 = e1 + a2,3
c e2, e′
i = ei, (2 ≤ i ≤ 4), e′
5 = e5
c and
c , b2,3
rename b2,1
c , a2,3
c and a2,5
c2 by b2,1, b2,3, a2,3 and a2,5, respectively. Then we assign
a2,1 := a2,3 + b2,3 − b2,1 and this case becomes the same as case (10) of Theorem
5.1.3.
(cid:3)
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
59
5.2.2. Codimension two and three solvable extensions of L 4
The non-zero inner derivations of L 4, (n ≥ 4) are given by
Le1 =
···
0
0 0 0
···
0
1 0 2
···
0 0 0
0
0 0 −1 0
···
0 0 0 −1 ···
...
...
...
0 0 0
0 0 0
...
0
0
...
0
0
0
0
0
...
0 0
0 0
0 0
0 0
0 0
...
...
··· −1 0 0
···
0 −1 0
, Le3 =
0 0 0 0 ··· 0
0 0 1 0 ··· 0
0 0 0 0 ··· 0
1 0 0 0 ··· 0
0 0 0 0 ··· 0
...
...
...
...
0 0 0 0 ··· 0
...
0 0 0 ··· 0
0 0 0 ··· 0
0 0 0 ··· 0
0 0 0 ··· 0
1 0 0 ··· 0
...
...
...
...
...
...
· 0 0 ··· 0
...
...
, Lei = Ei+1,1 =
(4 ≤ i ≤ n−1),
where Ei+1,1 is the n× n matrix that has 1 in the (i + 1, 1)st position and all other entries
are zero. According to Remark 5.1.4, we have at most two outer derivations.
5.2.2.1. Codimension two solvable extensions of L 4, (n = 4)
We consider the same cases as in Section 5.1.2.1 and follow the General approach given
in Section 5.1.2.
, (a 6= −1, a 6= 0, b 6= −1).
1
a
0
a1
b1
c1
=
(1) (a) One could set
Therefore the vector space of outer derivations as 4 × 4 matrices is as follows:
Le5 =
=
0
1 − a
2 b2,3 −2a a2,3
−a
0 −1 − a
−1
1 − a
and
2 a2,3 + 1−2a
a2
b2
c2
1
b
1
3−2a
0
0
0
0
0
0
,
.
0
0
0
0
0
As a result,
−1
−b
0
−1
−b
0
2(cid:17) a2,3− b2,3
2 − b(cid:17) a2,3 + 3
(1 − b)α2,3 − b · β2,3 −2(b + 1) 2α2,3 + β2,3 −1 − b
−1 − b
Le6 =
2 and it follows that β2,3 := (cid:16) 1
(i) Considering L[e5,e6], we obtain that a := 1. Since b 6= −1, we have that α2,3 :=
(cid:16)b + 1
2 b2,3. It implies that L[e5,e6] = 0.
0
0
0
−2 a2,3
0
0 −1
0
0 −2
0
−1
2(cid:17) a2,3 + b2,3
−b
0
( L[e1,e5] = Le1, L[e2,e5] = 0, L[e3,e5] = Le3, L[ei,e5] = 0, L[e1,e6] = Le1 + bLe3,
L[e2,e6] = 0, L[e3,e6] = Le1 + bLe3 , L[ei,e6] = 0, (b 6= −1, 4 ≤ i ≤ 6).
2 −1 − b
−1 − b
2(cid:17) b2,3 −2(b + 1) (cid:16)b + 3
−1
2 −(cid:16)b + 1
−b
0
Le5 =
Further, we find the following commutators:
, (b 6= −1)5.
−1
2
0
0
Le6 =
a2,3−b2,3
0
0
a2,3
0
0
0
,
(ii) We include a linear combination of e2 and e4:
[e1, e5] = e1 + c2,1e2 + c4,1e4, [e2, e5] = c2,2e2 + c4,2e4, [e3, e5] = c2,3e2 + e3 + c4,3e4,
[ei, e5] = c2,ie2 + c4,ie4, [e1, e6] = e1 + d2,1e2 + be3 + d4,1e4, [e2, e6] = d2,2e2 + d4,2e4,
[e3, e6] = e1 + d2,3e2 + be3 + d4,3e4, [ei, e6] = d2,ie2 + d4,ie4, (b 6= −1, 4 ≤ i ≤ 6).
Besides we have the brackets from L 4 and from outer derivations Le5 and Le6 as well.
(iii) We satisfy the right Leibniz identity shown in Table 5.
5We notice that Le6 is nilpotent if b = −1.
60
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Table 5. Left Leibniz identities in case (1) (a) with a nilradical L 4, (n = 4).
Steps Ordered triple
Result
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
Le1 ([e1, e5])
Le1 ([e1, e6])
Le1 ([e3, e5])
Le1 ([e3, e6])
Le5 ([e5, e5])
Le6 ([e6, e6])
Le3 ([e5, e5])
Le1 ([e5, e5])
Le1 ([e5, e6])
Le5 ([e3, e6])
Le5 ([e6, e5])
[e2, e5] = 0 =⇒ c2,2 = c4,2 = 0.
[e2, e6] = 0 =⇒ d2,2 = d4,2 = 0.
c2,4 := −2, c4,4 := 2 =⇒ [e4, e5] = 2 (e4 − e2) .
d2,4 := −b − 1, d4,4 := b + 1 =⇒ [e4, e6] = (b + 1) (e4 − e2) .
c4,5 = 0 =⇒ [e5, e5] = c2,5e2.
b 6= −1 =⇒ d4,6 = 0 =⇒ [e6, e6] = d2,6e2.
c2,3 := a2,3, c4,3 = 0 =⇒ [e3, e5] = a2,3e2 + e3.
c2,1 := a2,3−b2,3
, c4,1 = 0 =⇒ [e1, e5] = e1 + a2,3−b2,3
e2.
2
2
d2,1 :=(cid:0)b + 1
d2,3 :=(cid:0)b + 1
2(cid:1) a2,3 −
2(cid:1) a2,3 −
d4,5
b2,3
b2,3
2 , d4,1 = 0 =⇒ [e1, e6] = e1 +(cid:16)(cid:0)b + 1
2 , d4,3 = 0 =⇒ [e3, e6] = e1 +(cid:16)(cid:0)b + 1
2(cid:1) a2,3 −
2(cid:1) a2,3 −
b2,3
b2,3
2 (cid:17) e2 + be3.
2 (cid:17) e2 + be3.
:= −c4,6 =⇒ c4,6
:= (b + 1)c2,5 − c2,6 =⇒ [e5, e6] = d2,5e2 +
12.
Le5 ([e6, e6])
d2,6
:=
(b + 1) (c2,6 + d2,5) − (b + 1)2c2,5
=⇒
[e6, e6]
=
(c2,6 − (b + 1)c2,5) e4, [e6, e5] = c2,6e2 + ((b + 1)c2,5 − c2,6) e4
(cid:0)(b + 1) (c2,6 + d2,5) − (b + 1)2c2,5(cid:1) e2.
We obtain that Le5 and Le6 restricted to the nilradical do not change, but the remaining
brackets are as follows:
a2,3 − b2,3
1
1
2
b2,3
b2,3
[e1, e5] = e1 +
e2, [e3, e5] = a2,3e2 + e3, [e4, e5] = 2 (e4 − e2) , [e5, e5] = c2,5e2,
2 ! e2 + be3,
[e6, e5] = c2,6e2 + ((b + 1)c2,5 − c2,6) e4, [e1, e6] = e1 + (cid:18)b +
[e3, e6] = e1 + (cid:18)b +
[e5, e6] = d2,5e2 + (c2,6 − (b + 1)c2,5) e4, [e6, e6] = (cid:16)(b + 1) (c2,6 + d2,5) − (b + 1)2c2,5(cid:17) e2,
(b 6= −1).
2(cid:19) a2,3 −
2 ! e2 + be3, [e4, e6] = (b + 1) (e4 − e2) ,
2(cid:19) a2,3 −
Altogether the nilradical L 4 (4.0.1), the outer derivations Le5 and Le6 and the remaining
brackets given above define a continuous family of Leibniz algebras depending on the
parameters.
(iv)&(v) We apply the following transformation: e′
e3 + a2,3e2, e′
algebra l6,2 given below:
3 =
e4 and obtain a Leibniz
2 e2 + c2,6−(b+1)c2,5
1 = e1 + a2,3−b2,3
6 = e6 + d2,5
5 = e5 + c2,5
2 = e2, e′
4 = e4, e′
2 e2, e′
e2, e′
2
2
[e1, e5] = e1, [e3, e5] = e3, [e4, e5] = 2(e4 − e2), [e5, e1] = −e1, [e5, e2] = −2e2, [e5, e3] = −e3,
[e5, e4] = −2e4, [e1, e6] = e1 + be3, [e3, e6] = e1 + be3, [e4, e6] = (b + 1) (e4 − e2) ,
[e6, e1] = −e1 − be3, [e6, e2] = −2(b + 1)e2, [e6, e3] = −e1 − be3, [e6, e4] = −(b + 1) (e2 + e4) ,
(b 6= −1).
Remark 5.2.2. We notice that if b = −1 in the algebra, then the outer derivation Le6 is
nilpotent.
(1) (b) We set
vector space of outer derivations as 4 × 4 matrices is as follows:
Le5 =
=
−1
− c+1
2 a2,3 − c+3
2 b2,3 −2(c + 2)
−c − 1
=
(c + 1)a2,3 + c · b2,3 −1 − 2c
, (c 6= −2, d 6= −1). Therefore the
and
−c
−2
0
0
−3
a1
b1
c1
a2
b2
c2
1
1
d
1
2
c
0
0
0
0
0
,
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
61
(i) Considering L[e5,e6], we obtain that d = 0 and we have the system of equations:
.
There are the following two cases:
(I) If c 6= −1, then a2,3 := c+4
0
1−d
0
0
−1
−d
0
−d
−1
0
2 α2,3 − d+1
2 β2,3 −2(d + 1)
Le6 =
( b2,3 := (cid:16) c
(2c + 3) β2,3 − α2,3 − (c + 1) a2,3 − (c + 3) b2,3 = 0.
0
(d + 1)α2,3 + d · β2,3 −2d
0
−2
2 + 1(cid:17) α2,3,
2 + 1(cid:17) β2,3 −(cid:16) c
2 α2,3 − c
2 β2,3 and L[e5,e6] = 0. Consequently,
0
2 β2,3 −2(c + 2)
2 β2,3 −2c − 1
Le5 =
2 α2,3 + c
0
−3
3c+4
α2,3
0
,
0
0
0
0
−2 α2,3
0 −1
0
and we find the commutators given below:
−1
2
0
0
0
0
0
0 −2
α2,3−β2,3
0
−1
2 − 2c+3
−c − 1
Le6 =
−c
−2
0
, (c 6= −2)
2
1
α2,3−β2,3
(II) If c = −1, then b2,3 := β2,3−α2,3
0
−2 α2,3−β2,3
0
−2
0
0
( L[e1,e5] = Le1 + (c + 1)Le3, L[e2,e5] = 0, L[e3,e5] = cLe1 + 2Le3, L[ei,e5] = 0,
L[e1,e6] = Le1, L[e2,e6] = 0, L[e3,e6] = Le3, L[ei,e6] = 0, (c 6= −2, 4 ≤ i ≤ 6).
Le5 =
( L[e1,e5] = Le1, L[e2,e5] = 0, L[e3,e5] = 2Le3 − Le1, L[ei,e5] = 0,
L[e1,e6] = Le1, L[e2,e6] = 0, L[e3,e6] = Le3, L[ei,e6] = 0, (4 ≤ i ≤ 6).
and L[e5,e6] = 0. As a result,
0
0
−2 α2,3
0 −1
0
, Le6 =
and we have the following commutators:
0
0
0
0 −2
0
1
0
−3
−1
2
0
0
−1
2
0
0
α2,3−β2,3
2
(ii) We combine cases (I) and (II), include a linear combination of e2 and e4, and obtain
the following:
[e1, e5] = e1 + a2,1e2 + (c + 1)e3 + a4,1e4, [e2, e5] = a2,2e2 + a4,2e4, [e3, e5] = ce1 + a2,3e2+
2e3 + a4,3e4, [ei, e5] = a2,ie2 + a4,ie4, [e1, e6] = e1 + b2,1e2 + b4,1e4, [e2, e6] = b2,2e2 + b4,2e4,
[e3, e6] = b2,3e2 + e3 + b4,3e4, [ei, e6] = b2,ie2 + b4,ie4, (c 6= −2, 4 ≤ i ≤ 6).
Besides we have the brackets from L 4 and from outer derivations Le5 and Le6 as well.
(iii) To satisfy the left Leibniz identity, we refer to the identities given in Table 5. as much
as possible. The identities we apply are the following: 1.−6., Le3([e6, e6]) = [Le3(e6), e6]+
[e6, Le3(e6)], Le1 ([e6, e6]) = [Le1(e6), e6] + [e6, Le1(e6)], Le1 ([e6, e5]) = [Le1(e6), e5] +
[e6, Le1(e5)], Le3 ([e6, e5]) = [Le3(e6), e5] + [e6, Le3(e5)], 11. and 12. We have that Le5 and
Le6 restricted to the nilradical do not change, but the remaining brackets are as follows:
[e1, e5] = e1 + (cid:18)c +
[e3, e5] = ce1 +(cid:18)(cid:18) c
2
3
2(cid:19) α2,3 −
+ 2(cid:19) α2,3 −
β2,3
2 ! e2 + (c + 1)e3,
β2,3(cid:19) e2 + 2e3, [e4, e5] = 3 (e4 − e2) ,
c
2
α2,3 − β2,3
[e5, e5] = (c + 2) (a2,6 + a4,6) e2, [e6, e5] = a2,6e2 + a4,6e4, [e1, e6] = e1 +
[e3, e6] = α2,3e2 + e3, [e4, e6] = 2 (e4 − e2) , [e5, e6] = ((c + 2)b2,6 + a4,6) e2 − a4,6e4,
[e6, e6] = b2,6e2, (c 6= −2).
Altogether the nilradical L 4 (4.0.1), the outer derivations Le5 and Le6 and the remaining
brackets given above define a continuous family of Leibniz algebras.
e2,
2
62
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
1 = e1 + α2,3−β2,3
2
e2, e′
2 = e2, e′
3 = e3 + α2,3e2, e′
4 =
(iv)&(v) Applying the transformation: e′
6 = e6 + b2,6
e4, e′
a
1
0
a1
b1
c1
2 e4, e′
2 e2 + a4,6
2 e2, we obtain a Leibniz algebra l6,3:
5 = e5 + a2,6
[e1, e5] = e1 + (c + 1)e3, [e3, e5] = ce1 + 2e3, [e4, e5] = 3 (e4 − e2) , [e5, e1] = −e1 − (c + 1)e3,
[e5, e2] = −2(c + 2)e2, [e5, e3] = −ce1 − 2e3, [e5, e4] = (−2c − 1)e2 − 3e4, [e1, e6] = e1,
[e3, e6] = e3, [e4, e6] = 2 (e4 − e2) , [e6, e1] = −e1, [e6, e2] = −2e2, [e6, e3] = −e3,
[e6, e4] = −2e4, (c 6= −2).
=
Le5 =
=
(1) (c) We set
vector space of outer derivations as 4 × 4 matrices is as follows:
0
0
0
a − 1
−2 a2,3
0 −1
0 −a − 1
0
−1
b
0
0
and
−a
a − 1
−b
b − 1
0
0
−2 (b+1)α2,3+β2,3
0
0
, (a, b 6= 0, a 6= −1). Therefore the
(3a−2)a2,3 +(a−2)b2,3
a2
b2
c2
b
0
1
0
0
2a
.
,
(i) Considering L[e5,e6], we obtain that a := 1 and we have the system of equations:
0
b − 2
0
−b
2b
(3b−1)α2,3+(b−1)β2,3
Le6 =
( β2,3 := 3b·a2,3+b·b2,3
(b − 3)α2,3 + b−3
2
− (b + 1)α2,3,
2 (b2,3 − a2,3) = 0.
There are the following two cases:
and L[e5,e6] = 0. As
a2,3−b2,3
a result,
−1
2
0
0
(I) If b 6= 3, then α2,3 := a2,3−b2,3
2
2
=⇒ β2,3 := (2b−1)a2,3+(2b+1)b2,3
0
−1
−2 3a2,3+b2,3
2
0
0
0
0
−b
b − 1
0
b·a2,3+(b−2)b2,3
2
, (b 6= 0).
0
b − 2
0
−b
Further, we find the following:
0
0
−2 a2,3
0 −1
0
0
0
0
0 −2
, Le6 =
Le5 =
( L[e1,e5] = Le1, L[e2,e5] = 0, L[e3,e5] = Le3, L[ei,e5] = 0, L[e1,e6] = bLe1 + (1 − b)Le3,
L[e2,e6] = 0, L[e3,e6] = Le1, L[ei,e6] = 0, (b 6= 0, 4 ≤ i ≤ 6).
(II) If b := 3, then β2,3 := 9a2,3+3b2,3
− 4α2,3 and L[e5,e6] = 0 as well as Le5, Le6 are as
, Le6 =
( L[e1,e5] = Le1, L[e2,e5] = 0, L[e3,e5] = Le3, L[ei,e5] = 0, L[e1,e6] = 3Le1 − 2Le3,
L[e2,e6] = 0, L[e3,e6] = Le1, L[ei,e6] = 0, (4 ≤ i ≤ 6).
0
−1
−2 3a2,3+b2,3
2
0
0
0
0
Le5 =
0
0
0
0 −2
0
0
−2 a2,3
0 −1
0
We have the following commutators:
(ii) We combine cases (I) and (II) and include a linear combination of e2 and e4 :
0
1
0
−3
−3
2
2
0
−1
2
0
0
follows:
3a2,3+b2,3
a2,3−b2,3
.
2
[e1, e5] = e1 + c2,1e2 + c4,1e4, [e2, e5] = c2,2e2 + c4,2e4, [e3, e5] = c2,3e2 + e3 + c4,3e4,
[ei, e5] = c2,ie2 + c4,ie4, [e1, e6] = be1 + d2,1e2 + (1 − b)e3 + d4,1e4, [e2, e6] = d2,2e2+
d4,2e4, [e3, e6] = e1 + d2,3e2 + d4,3e4, [ei, e6] = d2,ie2 + d4,ie4, (b 6= 0, 4 ≤ i ≤ 6).
Besides we have the brackets from L 4 and from outer derivations Le5 and Le6 as well.
(iii) To satisfy the left Leibniz identity, we mostly apply the identities given in Table
5: 1. − 5., Le6 ([e6, e5]) = [Le6(e6), e5] + [e6, Le6(e5)], 7. − 12. We have that Le5 and Le6
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
63
restricted to the nilradical do not change, but the remaining brackets are the following:
e2, [e3, e5] = a2,3e2 + e3, [e4, e5] = 2 (e4 − e2) , [e5, e5] = c2,5e2,
a2,3 − b2,3
[e1, e5] = e1 +
2
(2 − b)a2,3 − b · b2,3
[e6, e5] = c2,6e2 + (c2,5 − c2,6) e4, [e1, e6] = be1 +
[e3, e6] = e1 +
[e6, e6] = (d2,5 − c2,5 + c2,6) e2, (b 6= 0).
e2 + (1 − b)e3,
e2, [e4, e6] = b (e4 − e2) , [e5, e6] = d2,5e2 + (c2,6 − c2,5) e4,
a2,3 − b2,3
2
2
Altogether the nilradical L 4 (4.0.1), the outer derivations Le5 and Le6 and the remaining
brackets given above define a continuous family of Leibniz algebras depending on the
parameters.
(iv)&(v) We apply the transformation: e′
e4, e′
5 = e5 + c2,5
[e1, e5] = e1, [e3, e5] = e3, [e4, e5] = 2 (e4 − e2) , [e5, e1] = −e1, [e5, e2] = −2e2, [e5, e3] = −e3,
[e5, e4] = −2e4, [e1, e6] = be1 + (1 − b)e3, [e3, e6] = e1, [e4, e6] = b (e4 − e2) ,
[e6, e1] = −be1 + (b − 1)e3, [e6, e2] = −2e2, [e6, e3] = −e1, [e6, e4] = (b − 2)e2 − be4, (b 6= 0).
e4 and obtain a Leibniz algebra l6,4:
1 = e1 + a2,3−b2,3
3 = e3 + a2,3e2, e′
2 e2 + c2,6−c2,5
6 = e6 + d2,5
2 = e2, e′
2 e2, e′
e2, e′
4 =
2
2
5.2.2.2. Codimension two solvable extensions of L 4, (n ≥ 5)
One could set a
derivations as n × n matrices is as follows:
2 ! and α
β ! = 1
1 ! . Therefore the vector space of outer
b ! = 1
−1
0
0
3a2,1−2a2,3 −4
0
0
0 −a5,3
0
−1
0
0
0
a2,3
−2
0
·
Len+1 =
0
...
0
...
·
...
0
0
0
−1
0
α2,1 −2
0
0
0 −α5,3
0
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0 −an,3 −an−1,3 ···
0
0
0
0
0
0
0
0
−3 0
0 −4
0
0
0
−2
0
α2,3
−1
0
0
0
0
0
−α5,3
0
·
0
...
·
...
... 0
...
·
...
···
···
···
···
···
···
...
...
Len+2 =
0
−1
0
−3
0
−a5,3
·
...
0
0
0
0
0
0
0
0
−4 0
0 −5
... 0
...
···
···
···
···
···
···
...
...
...
0
··· −a5,3
···
···
·
· −a5,3
0
0
0
0
0
0
...
...
3−n
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
0
0
2−n 0
0
1−n
,
...
...
.
· −a5,3
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
...
4−n
0
...
0
0
0
0
0
0
0
...
...
0
...
0
0
0
0
...
0
0
0
0 −αn−2,3 −αn−3,3 ···
0 −αn−1,3 −αn−2,3 ···
0 −αn,3 −αn−1,3 ···
··· −α5,3
···
···
·
· −α5,3
0
0
3−n 0
· −α5,3
0
2−n
(i) Considering L[en+1,en+2] and comparing withPn
i ≤ n), α2,3 := a2,3
as follows:
2 and α2,1 := a2,1 − a2,3
i=1 ciLei, we deduce that αi,3 := ai,3, (5 ≤
2 . As a result, the outer derivation Len+2 changes
0
−1
a2,1−
0
0
0
0
a2,3
0
a2,3
2 −2
0
0
0 −a5,3
0
2
−1
0
·
0
...
·
...
0
0
0
−2
0
·
...
−a5,3
0
0
0
0
0
0
0
0
−3 0
0 −4
... 0
...
···
···
···
···
···
···
...
...
Len+2 =
0
0
0
0
0
0
...
...
4−n
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
...
0
0
3−n 0
...
0
.
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0 −an,3 −an−1,3 ···
··· −a5,3
···
···
·
· −a5,3
· −a5,3
0
2−n
64
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
Altogether we find the following commutators:
L[e1,en+1] = Le1 + Le3, L[e2,en+1] = 0, L[ei,en+1] = (i − 1)Lei +
n−1
n−1
Xk=i+2
ak−i+3,3Lek,
L[ej ,en+1] = 0, (n ≤ j ≤ n + 1), L[en+2,en+1] =
Xk=i+2
L[e2,en+2] = 0, L[ei,en+2] = (i − 2)Lei +
n−1
ak+1,3Lek, L[e1,en+2] = Le1,
Xk=4
ak−i+3,3Lek, (3 ≤ i ≤ n − 1),
L[en,en+2] = 0, L[en+1,en+2] = −
ak+1,3Lek , L[en+2,en+2] = 0.
(ii) We include a linear combination of e2 and en :
n−1
Xk=4
[e1, en+1] = e1 + c2,1e2 + e3 + cn,1en, [e2, en+1] = c2,2e2 + cn,2en, [ei, en+1] = c2,ie2+
(i − 1)ei +
n−1
Xk=i+2
ak−i+3,3ek + cn,ien, [ej, en+1] = c2,je2 + cn,jen, (n ≤ j ≤ n + 1),
n−1
Xk=4
[en+2, en+1] = c2,n+2e2 +
ak+1,3ek + cn,n+2en, [e1, en+2] = e1 + d2,1e2 + dn,1en,
n−1
Xk=i+2
[e2, en+2] = d2,2e2 + dn,2en, [ei, en+2] = d2,ie2 + (i − 2)ei +
ak−i+3,3ek + dn,ien,
(3 ≤ i ≤ n − 1), [en, en+2] = d2,ne2 + dn,nen, [en+1, en+2] = d2,n+1e2 −
dn,n+1en, [en+2, en+2] = d2,n+2e2 + dn,n+2en.
n−1
Xk=4
ak+1,3ek+
Besides we have the brackets from L 4 and from outer derivations Len+1 and Len+2 as
well.
(iii) We satisfy the right Leibniz identity shown in Table 6. We notice that Len+1 and
Len+2 restricted to the nilradical do not change, but the remaining brackets are as follows:
[e1, en+1] = e1 + a2,1e2 + e3, [e3, en+1] = a2,3e2 + 2e3 +
n
n
ak−1,3ek, [ei, en+1] = (i − 1)ei +
Xk=i+2
Xk=6
n
ak,3ek, [e4, en+1] = 3 (e4 − e2) +
Xk=5
ak−i+3,3ek, [en+1, en+1] = (2c2,n+2 + 2a5,3) e2,
[en+2, en+1] = c2,n+2e2 +
[e3, en+2] =
e2 + e3 +
a2,3
2
n−1
n
ak+1,3ek + cn,n+2en, [e1, en+2] = e1 +(cid:18)a2,1 −
Xk=4
Xk=5
ak,3ek, [e4, en+2] = 2 (e4 − e2) +
Xk=i+2
ak−i+3,3ek, (5 ≤ i ≤ n), [en+1, en+2] = d2,n+1e2 −
d2,n+1 − a5,3
ak−1,3ek,
Xk=6
n
n
n−1
Xk=4
a2,3
2 (cid:19) e2,
[ei, en+2] = (i − 2)ei +
cn,n+2en, [en+2, en+2] =
Altogether the nilradical L 4 (4.0.1), the outer derivations Len+1 and Len+2 and the
remaining brackets given above define a continuous family of the solvable left Leibniz
algebras. Then we apply the technique of "absorption" according to step (iv).
e2.
2
ak+1,3ek−
2
c2,n+2
e2, e′
n+2 = en+2 + d2,n+1−a5,3
• We start with the transformation e′
n+1 = en+1 +
e2. This transformation removes the coefficients
c2,n+2 and d2,n+1−a5,3
in front of e2 in [en+2, en+1] and [en+2, en+2], respectively, and
changes the coefficient in front of e2 in [en+1, en+1] and [en+1, en+2] to 2a5,3 and
a5,3, correspondingly.
i = ei, (1 ≤ i ≤ n, n ≥ 5), e′
2
4
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
65
Table 6. Left Leibniz identities in the codimension two nilradical
L 4, (n ≥ 5).
Steps Ordered triple
Result
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
c2,n = 0, cn,n := n − 1 =⇒ [en, en+1] = (n − 1) en. Altogether with 4., [ei, en+1] =
d2,4 := −2, dn,4 := an−1,3, where a4,3 = 0 =⇒ [e4, en+2] = 2(e4 − e2) +
d2,i+1 = 0, dn,i+1 := an−i+2,3, (4 ≤ i ≤ n − 2), where a4,3 = 0 =⇒ [ej, en+2] =
k=6 ak−1,3ek.
c2,i+1 = 0, cn,i+1 := an−i+2,3, (4 ≤ i ≤ n − 2), where a4,3 = 0 =⇒ [ej , en+1] =
Le1 ([e1, en+1])
[e2, en+1] = 0 =⇒ c2,2 = cn,2 = 0.
k=6 ak−1,3ek.
[e2, en+2] = 0 =⇒ d2,2 = dn,2 = 0.
k=i+2 ak−i+3,3ek, (5 ≤ i ≤ n).
k=j+2 ak−j+3,3ek, (5 ≤ j ≤ n − 1).
Le1 ([e1, en+2])
Le1 ([e3, en+1])
Le1 ([ei, en+1])
Le1 ([en−1, en+1])
Le1 ([e3, en+2])
Le1 ([ei, en+2])
c2,4 := −3, cn,4 := an−1,3, where a4,3 = 0 =⇒ [e4, en+1] = 3(e4−e2)+Pn
(j − 1) ej +Pn
(i − 1) ei +Pn
Pn
(j − 2) ej +Pn
(i − 2) ei +Pn
c2,3 := a2,3, cn,3 := an,3 =⇒ [e3, en+1] = a2,3e2 + 2e3 +Pn
2 e2 + e3 +Pn
2 , dn,1 = 0 =⇒ [e1, en+2] = e1 +(cid:0)a2,1 − a2,3
2 , dn,3 := an,3 =⇒ [e3, en+2] = a2,3
c2,1 := a2,1, cn,1 = 0 =⇒ [e1, en+1] = e1 + a2,1e2 + e3.
Len+1 ([en+1, en+1])cn,n+1 = 0 =⇒ [en+1, en+1] = c2,n+1e2.
Len+2 ([en+2, en+2])dn,n+2 = 0 =⇒ [en+2, en+2] = d2,n+2e2.
Le3 ([en+1, en+1])
Le3 ([en+2, en+2]) d2,3 := a2,3
Le1 ([en+1, en+1])
Le1 ([en+2, en+2]) d2,1 := a2,1 − a2,3
Len+1 ([en+2, en+1])c2,n+1
k=j+2 ak−j+3,3ek, (5 ≤ j ≤ n − 1).
k=i+2 ak−i+3,3ek, (5 ≤ i ≤ n).
k=5 ak,3ek.
k=5 ak,3ek.
:= 2c2,n+2 + 2a5,3, dn,n+1
2 (cid:1) e2.
=⇒
:= −cn,n+2
Len+1 ([en+2, en+2])d2,n+2 := d2,n+1−a5,3
(2c2,n+2 + 2a5,3) e2, [en+1, en+2] = d2,n+1e2 −Pn−1
=⇒ [en+2, en+2] = d2,n+1−a5,3
2
2
e2.
[en+1, en+1] =
k=4 ak+1,3ek − cn,n+2en.
Le1 ([en−1, en+2]) d2,n = 0, dn,n := n − 2 =⇒ [en, en+2] = (n − 2) en. Combining with 7., [ei, en+2] =
2 ak−1,3 and a5,3
i = ei, (1 ≤ i ≤ n, n ≥ 5), e′
• Then we apply the transformation e′
n+1 = en+1 +
a5,3
2 e4, e′
n+2 = en+2 to remove the coefficients a5,3 and 2a5,3 in front of e2 in
[en+1, en+2] and [en+1, en+1], respectively. At the same time this transformation
removes a5,3 and −a5,3 in front of e4 in [en+2, en+1] and [en+1, en+2], respectively,
and changes the coefficients in front of ek, (6 ≤ k ≤ n − 1) in [en+2, en+1] and
[en+1, en+2] to ak+1,3 − a5,3
2 ak−1,3 − ak+1,3, respectively. It also af-
fects the coefficients in front en, (n ≥ 6) in [en+2, en+1] and [en+1, en+2], which we
rename back by cn,n+2 and −cn,n+2, respectively. The following entries are intro-
duced by the transformation: − a5,3
in the (5, 1)st, (n ≥ 5) position in
Ren+1 and Len+1, respectively.
• Applying the transformation e′
n+2 = en+2 −
Pn−1
(6 ≤ k ≤ n − 1, n ≥ 7), we remove the coefficients a6,3 and −a6,3 in front of
e5 in [en+2, en+1] and [en+1, en+2], respectively. Besides the transformation re-
moves ak+1,3 − a5,3
2 ak−1,3 − ak+1,3 in front of ek, (6 ≤ k ≤ n − 1)
in [en+2, en+1] and [en+1, en+2], respectively. This transformation introduces a6,3
4
and Ak+1,3
and − a6,3
in the
1−k
(k + 1, 1)st, (6 ≤ k ≤ n − 1) position in Ren+2 and Len+2, respectively.
It also
affects the coefficients in front en, (n ≥ 7) in [en+2, en+1] and [en+1, en+2], which
we rename back by cn,n+2 and −cn,n+2, respectively.
j = ej, (1 ≤ j ≤ n + 1, n ≥ 6), e′
2 a5,3ak−1,3−Pk
in the (6, 1)st, (n ≥ 6) position as well as Ak+1,3
k−1 ek, where A6,3 := a6,3 and Ak+1,3 := ak+1,3− 1
2 ak−1,3 and a5,3
2 and a5,3
Ak+1,3
k−1
k=5
i−3
i=7
4
2
Ai−1,3ak−i+5,3
,
66
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
• Finally applying the transformation e′
en+2 − cn,n+2
[en+1, en+2], respectively, without affecting other entries.
n+2 =
n−1 en, we remove cn,n+2 and −cn,n+2 in front of en in [en+2, en+1] and
i = ei, (1 ≤ i ≤ n + 1, n ≥ 5), e′
Len+2 =
−1
We obtain that Len+1 and Len+2 are as follows:
···
···
···
···
···
···
0
3a2,1−2a2,3 −4
0
0
0 −a5,3
0
0
0
0
0
0
0
0
0
−4 0
0 −5
0
−1
0
−3
0
a2,3
−2
0
−1
0
a5,3
2
0
−a5,3
0
·
Len+1 =
0
...
0
0
0
0
...
·
...
... 0
...
...
...
·
...
...
0
0 −an−2,3 −an−3,3 ···
0 −an−1,3 −an−2,3 ···
0 −an,3 −an−1,3 ···
··· −a5,3
···
···
·
· −a5,3
0
0
0
0
0
0
...
...
3−n
0
0
0
0
0
0
0
0
0
0
0
0
0
...
...
...
...
0
0
2−n 0
0
a2,3
−1
a2,1−
0
0
0
a6,3
0
a2,3
2 −2
0
0
0 −a5,3
0
2
−1
0
−
·
4
0
0
0
−2
0
−a5,3
0
0
0
0
−3
0
0
0
0
0
0
−4
0
0
0
0
0
0
A7,3
5
A8,3
−
−
6
...
An−2,3
4−n
An−1,3
3−n
An,3
2−n
0
0
...
·
·
...
·
·
...
−a5,3
0 −5
· −a5,3 0
...
0 −an−2,3 −an−3,3
0 −an−1,3 −an−2,3
···
···
··· −a5,3
···
···
···
···
···
···
···
···
···
...
...
1−n
0
0
0
0
0
· −a5,3
0
0
0
0
0
0
0
...
...
...
4−n
0
0
0
0
0
0
...
...
...
0
...
0
0
...
...
...
0
···
···
· −a5,3
0
3−n 0
·
· −a5,3
0
2−n
n
,
, (n ≥ 5).
···
The remaining brackets are given below:
0 −an,3 −an−1,3
a5,3
2
(v) Finally we apply the following two change of basis transformations:
n
n
n
n
n
n
a2,3
ak,3ek,
ak,3ek,
Xk=6
Xk=5
a2,3
2
e2 + e3 +
ak−i+3,3ek,
ek, [e3, en+2] =
Ak,3
k − 2
Xk=5
Xk=i+2
e5, [e3, en+1] = a2,3e2 + 2e3 +
ak−i+3,3ek, (5 ≤ i ≤ n).
[e1, en+1] = e1 + a2,1e2 + e3 −
Xk=6
[e4, en+1] = 3 (e4 − e2) +
ak−1,3ek, [ei, en+1] = (i − 1)ei +
[e1, en+2] = e1 +(cid:18)a2,1 −
2 (cid:19) e2 +
Xk=6
[e4, en+2] = 2 (e4 − e2) +
ak−1,3ek, [ei, en+2] = (i − 2)ei +
2 (cid:17) e2, e′
1 = e1 + (cid:16)a2,1 − a2,3
• e′
(3 ≤ i ≤ n − 2), e′
n−1 = en−1, e′
aj,3 −Pj
, (5 ≤ j ≤ n).
This transformation removes a5,3, a6,3, ..., an,3 in Ren+1, Ren+2 and −a5,3,−a6,3, ...,−an,3
in Len+1, Len+2. It also removes − a5,3
from the (5, 1)st positions in Ren+1
and Len+1, respectively. Besides it removes a2,1 and 3a2,1 − 2a2,3 from the (2, 1)st
positions in Ren+1 and Len+1, respectively, as well as a2,1− a2,3
from the (2, 1)st po-
sitions in Ren+2 and Len+2. The transformation also removes a2,3 from the (2, 3)rd
positions in Ren+1, Len+1 and a2,3
from the same positions in Ren+2, Len+2. It intro-
duces the entries in the (i, 1)st positions in Ren+1 and Len+1, that we set to be ai,1
and −ai,1, (6 ≤ i ≤ n), respectively. The transformation affects the entries in the
(j, 1)st, (8 ≤ j ≤ n) positions in Ren+2 and Len+2, but we rename all the entries in
the (i, 1)st positions by i−3
Xk=i+2
i = ei − Pn
ek,
n+2 = en+2, where Bj,3 :=
2 e2, e′
n+1 = en+1, e′
3 = e3 + a2,3
2 and a5,3
i−2 ai,1 in Ren+2 and by 3−i
2 = e2, e′
n = en, e′
Bk−2,3aj−k+5,3
i−2 ai,1, (6 ≤ i ≤ n) in Len+2.
Bk−i+3,3
k=i+2
k−5
k=7
k−i
2
2
2
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
67
k−2
k=5
and 3−i
We obtain a Leibniz algebra ln+2,1 given below:
k = ek, (1 ≤ k ≤ n), e′
n+1 = en+1+Pn−1
i−2 ai,1, (6 ≤ i ≤ n) in Ren+2 and Len+2, respectively.
k=5 ak+1,1ek, e′
k−1 ak+1,1ek, we remove ai,1 in Ren+1 and −ai,1 in Len+1 as well as i−3
• Applying the transformation e′
en+2+Pn−1
[e1, en+1] = e1 + e3, [e3, en+1] = 2e3, [e4, en+1] = 3(e4 − e2), [ei, en+1] = (i − 1)ei,
[en+1, e1] = −e1 − e3, [en+1, e2] = −4e2, [en+1, e3] = −2e3, [en+1, e4] = −e2 − 3e4,
[en+1, ei] = (1 − i)ei, [e1, en+2] = e1, [e3, en+2] = e3, [e4, en+2] = 2(e4 − e2),
[ei, en+2] = (i − 2)ei, (5 ≤ i ≤ n, n ≥ 5), [en+2, e1] = −e1, [en+2, e2] = −2e2,
[en+2, ej] = (2 − j)ej, (3 ≤ j ≤ n).
n+2 =
i−2 ai,1
We summarize a result in the following theorem:
Theorem 5.2.4. There are four solvable indecomposable left Leibniz algebras up to iso-
morphism with a codimension two nilradical L 4, (n ≥ 4), which are given below:
(i) ln+2,1 : [e1, en+1] = e1 + e3, [e3, en+1] = 2e3, [e4, en+1] = 3(e4 − e2), [ei, en+1] = (i − 1)ei,
[en+1, e1] = −e1 − e3, [en+1, e2] = −4e2, [en+1, e3] = −2e3, [en+1, e4] = −e2 − 3e4,
[en+1, ei] = (1 − i)ei, [e1, en+2] = e1, [e3, en+2] = e3, [e4, en+2] = 2(e4 − e2),
[ei, en+2] = (i − 2)ei, (5 ≤ i ≤ n, n ≥ 5), [en+2, e1] = −e1, [en+2, e2] = −2e2,
[en+2, ej] = (2 − j)ej, (3 ≤ j ≤ n),
(ii) l6,2 : [e1, e5] = e1, [e3, e5] = e3, [e4, e5] = 2(e4 − e2), [e5, e1] = −e1, [e5, e2] = −2e2,
[e5, e3] = −e3, [e5, e4] = −2e4, [e1, e6] = e1 + be3, [e3, e6] = e1 + be3, [e4, e6] = (b + 1)e4−
(b + 1)e2, [e6, e1] = −e1 − be3, [e6, e2] = −2(b + 1)e2, [e6, e3] = −e1 − be3,
[e6, e4] = −(b + 1) (e2 + e4) , (b 6= −1),
(iii) l6,3 : [e1, e5] = e1 + (c + 1)e3, [e3, e5] = ce1 + 2e3, [e4, e5] = 3 (e4 − e2) , [e5, e1] = −e1−
(c + 1)e3, [e5, e2] = −2(c + 2)e2, [e5, e3] = −ce1 − 2e3, [e5, e4] = (−2c − 1)e2 − 3e4,
[e1, e6] = e1, [e3, e6] = e3, [e4, e6] = 2 (e4 − e2) , [e6, e1] = −e1, [e6, e2] = −2e2, [e6, e3] = −e3,
[e6, e4] = −2e4, (c 6= −2),
(iv) l6,4 : [e1, e5] = e1, [e3, e5] = e3, [e4, e5] = 2 (e4 − e2) , [e5, e1] = −e1, [e5, e2] = −2e2,
[e5, e3] = −e3, [e5, e4] = −2e4, [e1, e6] = be1 + (1 − b)e3, [e3, e6] = e1, [e4, e6] = b (e4 − e2) ,
[e6, e1] = −be1 + (b − 1)e3, [e6, e2] = −2e2, [e6, e3] = −e1, [e6, e4] = (b − 2)e2 − be4, (b 6= 0).
68
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
References
[1] Albeverio, S., Omirov, B. A., Rakhimov, I. S. (2006). Classification of 4-dimensional nilpotent
complex Leibniz algebras. Extracta Math. 21 (3):197 -- 210.
[2] Ayupov, Sh. A., Omirov, B. A. (1998). On Leibniz algebras. Algebra and operator theory. Kluwer
Acad. Publ., Dordrecht, pp. 1 -- 12.
[3] Barnes, D. W. (2012). On Levi's theorem for Leibniz algebras. Bull. Aust. Math. Soc. 86(2):184 -- 185.
[4] Bloch, A. M. (1965). On a generalization of the concept of Lie algebra. Dokl. Akad. Nauk SSSR
18(3):471 -- 473.
[5] Bosko-Dunbar, L., Dunbar, J. D., Hird, J. T., Stagg, K. (2015). Solvable Leibniz algebras with
Heisenberg nilradical. Comm. Algebra. 43(6):2272 -- 2281.
[6] Camacho, L. M., Gómez, J. R., González, A. J., Omirov, B. A. (2009). Naturally graded quasi-
filiform Leibniz algebras. J. Symbolic Comput. 44(5):527 -- 539.
[7] Camacho, L. M., Omirov, B. A., Masutova, K. K. (2016). Solvable Leibniz Algebras with Filiform
Nilradical. Bull. Malays. Math. Sci. Soc. 39(1):283 -- 303.
[8] Casas, J. M., Ladra, M., Omirov, B. A., Karimjanov, I. A. (2013). Classification of solvable Leibniz
algebras with null-filiform nilradical. Linear Multilinear Algebra. 61(6):758 -- 774.
[9] Casas, J. M., Ladra, M., Omirov, B. A., Karimjanov, I. A. (2013). Classification of solvable Leibniz
algebras with naturally graded filiform nilradical. Linear Algebra Appl. 438(7):2973 -- 3000.
[10] Cuvier, C. (1994). Algèbres de Leibnitz: définitions, propriétés. Ann. Sci. École Norm. Sup. (4)
27(1):1 -- 45.
[11] Demir, I., Misra, K. C., Stitzinger, E. (2014). On some structures of Leibniz algebras. Recent
advances in representation theory, quantum groups, algebraic geometry, and related topics. Contemp.
Math. 623:41 -- 54.
[12] Karimjanov, I. A., Khudoyberdiyev, A. K., Omirov, B. A. (2015). Solvable Leibniz algebras with
triangular nilradicals. Linear Algebra Appl. 466:530 -- 546.
[13] Khudoyberdiyev, A. K., Ladra, M., Omirov, B. A. (2014). On solvable Leibniz algebras whose
nilradical is a direct sum of null-filiform algebras. Linear Multilinear Algebra. 62(9):1220 -- 1239.
[14] Ladra, M., Masutova, K. K., Omirov, B. A. (2016). Corrigendum to "Classification of solvable
Leibniz algebras with naturally graded filiform nilradical" [Linear Algebra Appl. 438 (7) (2013)
2973-3000]. Linear Algebra Appl. 507:513 -- 517.
[15] Loday, J.-L. (1992). Cyclic homology. Grundl. Math. Wiss. 301. Springer, Berlin.
[16] Loday, J.-L. (1993). Une version non commutative des algèbres de Lie:
les algèbres de Leibniz.
Enseign. Math. (2) 39(3-4):269 -- 293.
[17] Loday J.-L., Pirashvili T.,
(1993). Universal enveloping algebras of Leibniz algebras and
(co)homology. Math. Ann. 296(1):139 -- 158.
SOLVABLE EXTENSIONS OF QUASI-FILIFORM ALGEBRAS
69
[18] Mubarakzyanov, G. M. (1963). On solvable Lie algebras. Izv. Vysshikh Uchebn. Zavedenii Mat.
1(32):114 -- 123.
[19] Omirov, B. A., Rakhimov, I. S., Turdibaev, R. M. (2013). On description of Leibniz algebras corre-
sponding to sl2. Algebr. Represent. Theory. 16(5):1507 -- 1519.
[20] Shabanskaya, A. (2011). Classification of Six Dimensional Solvable Indecomposable Lie Algebras
with a codimension one nilradical over R. (Doctoral dissertation, the University of Toledo).
[21] Shabanskaya, A., Thompson, G. (2013). Six-dimensional Lie algebras with a five-dimensional nil-
radical. Journal of Lie Theory. 23(2):313 -- 355.
[22] Shabanskaya, A. (2017). Right and left solvable extensions of an associative Leibniz algebra. Comm.
Algebra. 45(6):2633 -- 2661.
[23] Shabanskaya, A. (2017). Solvable extensions of naturally graded quasi-filiform Leibniz algebras of
second type L 1 and L 3. Comm. Algebra. 45(10):4492 -- 4520.
[24] Shabanskaya, A. (2018). Solvable extensions of the naturally graded quasi-filiform Leibniz algebra
of second type L 2. Comm. Algebra. 46(11):5008 -- 5033.
|
1207.0298 | 4 | 1207 | 2013-02-19T17:21:22 | Echoes in correlated neural systems | [
"q-bio.NC",
"cond-mat.dis-nn",
"cond-mat.stat-mech"
] | Correlations are employed in modern physics to explain microscopic and macroscopic phenomena, like the fractional quantum Hall effect and the Mott insulator state in high temperature superconductors and ultracold atoms. Simultaneously probed neurons in the intact brain reveal correlations between their activity, an important measure to study information processing in the brain that also influences macroscopic signals of neural activity, like the electro encephalogram (EEG). Networks of spiking neurons differ from most physical systems: The interaction between elements is directed, time delayed, mediated by short pulses, and each neuron receives events from thousands of neurons. Even the stationary state of the network cannot be described by equilibrium statistical mechanics. Here we develop a quantitative theory of pairwise correlations in finite sized random networks of spiking neurons. We derive explicit analytic expressions for the population averaged cross correlation functions. Our theory explains why the intuitive mean field description fails, how the echo of single action potentials causes an apparent lag of inhibition with respect to excitation, and how the size of the network can be scaled while maintaining its dynamical state. Finally, we derive a new criterion for the emergence of collective oscillations from the spectrum of the time-evolution propagator. | q-bio.NC | q-bio |
Echoes in correlated neural systems
M Helias1, T Tetzlaff1 and M Diesmann1,2,3
1Institute of Neuroscience and Medicine (INM-6) and
Institute for Advanced Simulation (IAS-6)
Julich Research Centre and JARA, Julich, Germany
2RIKEN Brain Science Institute, Wako, Saitama, Japan
3Medical Faculty, RWTH Aachen University, Germany
E-mail: [email protected]
Abstract. Correlations are employed in modern physics to explain microscopic
and macroscopic phenomena,
like the fractional quantum Hall effect and the
Mott insulator state in high temperature superconductors and ultracold atoms.
Simultaneously probed neurons in the intact brain reveal correlations between their
activity, an important measure to study information processing in the brain that
also influences macroscopic signals of neural activity, like the electro encephalogram
(EEG). Networks of spiking neurons differ from most physical systems: The
interaction between elements is directed, time delayed, mediated by short pulses,
and each neuron receives events from thousands of neurons. Even the stationary
state of the network cannot be described by equilibrium statistical mechanics.
Here we develop a quantitative theory of pairwise correlations in finite sized
random networks of spiking neurons. We derive explicit analytic expressions for
the population averaged cross correlation functions. Our theory explains why the
intuitive mean field description fails, how the echo of single action potentials causes
an apparent lag of inhibition with respect to excitation, and how the size of the
network can be scaled while maintaining its dynamical state. Finally, we derive a
new criterion for the emergence of collective oscillations from the spectrum of the
time-evolution propagator.
Keywords:
integrate-and-fire model
spiking neural networks,
correlations, non-equilibrium dynamics,
PACS numbers: 87.18.Sn, 87.19.lc, 87.19.lj, 87.19.lm, 05.10.Gg, 05.70.Ln, 82.40.Bj
AMS classification scheme numbers:
Echoes in correlated neural systems
2
Submitted to: New J. Phys.
Echoes in correlated neural systems
3
1. Introduction
Correlations are an established feature of neural activity [1] and evidence increases
that they can provide insights into the information processing in the brain [2].
The temporal relationship between the activity of pairs of neurons is described by
correlation functions. Their shape has early been related to the direct coupling
between neurons and to the common input shared by pairs of neurons. On the one
hand, correlations may limit the signal-to-noise ratio of population rate signals [3],
on the other hand they have been shown to increase the amount of information
available to unbiased observers [4]. Furthermore, synchronous neural activity has
been proposed to bind elementary representations into more complex objects [5]
and experimental evidence for such a correlation code is provided by task related
modulation of synchrony in primary visual cortex [6] and in motor cortex [7].
The small magnitude [8] of pairwise correlations in the asynchronous irregular
state [9] of cortex has recently been related to the balance between excitation and
inhibition in local networks [10, 11] and inhibitory feedback was identified as a
general mechanism of decorrelation [12]. However, a quantitative theory explaining
the temporal shape of correlation functions in recurrently impulse coupled networks
of excitatory and inhibitory cells remained elusive.
Assuming random connectivity with identical numbers and strengths of
incoming synapses per neuron, as illustrated in Figure 1, suggests by mean field
arguments [13, 14] that the resulting activity of two arbitrarily selected neurons
and hence the power spectra of activities averaged over excitatory or inhibitory
neurons should be the same. Direct simulations, however, exhibit different power
spectra for these sub-populations [15]. A similar argument holds for the covariance
cff between the two neurons: If the covariance c between any pair of inputs is known,
the covariance between their outgoing activity cff is fully determined [15–24]. By self-
consistency, as both neurons belong to the same recurrent network, one concludes
that cff = c. In particular the covariance averaged over excitatory pairs should be
identical to the corresponding average over inhibitory pairs, which is in contrast
to direct simulation (Figure 1b).
In this work, we elucidate why this mean field
argument for covariances fails and derive a self-consistency equation for pairwise
covariances in recurrent random networks which explains the differences in the power
spectra and covariances.
Theories for pairwise covariances have been derived for binary neuron models
Echoes in correlated neural systems
a
)b
.
.
u
a
(
n
o
i
t
l
a
e
r
r
o
c
4
cee
cii
s
s
o
r
c
−10
0
t(ms)
10
Figure 1.
Self-consistency argument fails for covariances in a homogeneous
recurrent random network. (a) Each neuron (black circles) receives input from the
same number of randomly chosen excitatory (e) and inhibitory (i) neurons in the
network, so the input statistics of all neurons is the same. The covariance c within
the network determines the covariance between the inputs to a pair of neurons
and hence the covariance cff of their outputs. Self-consistency seems to require
cff = c = cee = cii. (b) Covariance functions averaged over pairs of excitatory (cee)
and over pairs of inhibitory (cii) integrate-and-fire model neurons are different in
a direct simulation. Other parameters are given in Appendix E.
[11, 25] and for excitatory stochastic point process models [26]. However, the lack of
either inhibition or delayed pulsed interaction limits the explanatory power of these
models. A theory for networks of leaky integrate-and-fire (LIF) model neurons [27]
is required, because this model has been shown to well approximate the properties
of mammalian pyramidal neurons [28] and novel experimental techniques allow to
reliably assess the temporal structure of correlations in cortex [29]. Moreover, the
relative timing of action potentials is the basis for models of synaptic plasticity
[30], underlying learning in biological neural networks. Analytical methods to treat
population fluctuations in spiking networks are well advanced [31] and efficient hybrid
analytical-numerical schemes exist to describe pairwise covariances [32]. Here we
present an analytically solvable theory of pairwise covariances in random networks
of spiking leaky integrate-and-fire model neurons with delayed pulsed interaction in
the asynchronous irregular regime.
2. Results
We consider recurrent random networks of N excitatory and γN inhibitory leaky
integrate-and-fire model neurons receiving pulsed input (spikes) from other neurons
Echoes in correlated neural systems
5
in the network.
Each neuron has K = pN incoming excitatory synapses
independently and randomly drawn from the pool of excitatory neurons, and
γK = γpN inhibitory synapses (homogeneous Erdos-R´enyi random network with
fixed in-degree). An impulse at time t arrives at the target neuron after the synaptic
delay d and elicits a synaptic current Ii that decays with time constant τs and causes
a response in the membrane potential Vi (with time constant τm) proportional to the
synaptic efficacy J (excitatory) or −gJ (inhibitory), respectively. The coupled set
of differential equations governing the subthreshold dynamics of a single neuron i is
[33]
τm
τs
dVi
dt
dIi
dt
= − Vi + Ii(t)
Xj=1,j
= − Ii + τm
N
Jijsj(t − d),
(1)
where the membrane resistance was absorbed into Jij. If Vi reaches the threshold Vθ
at time point ti
k the neuron emits an action potential and the membrane potential is
reset to Vr, where it is clamped for the refractory time τr. The spiking activity
of neuron i is described by this sequence of action potentials, the spike train
si(t) =Pk δ(t − ti
k).
The activity of a given neuron i depends on the history of the other neurons’
activities s(t) = (s1(t), . . . , sN (t))T in the network, so formally we can consider
the spike train of neuron i as a functional of all other spike trains. The time
averaged covariance matrix expresses these interrelations and is defined as ¯c(τ ) =
(cid:10)s(t + τ )sT (t)(cid:11)t − rrT , where r = hsit is the vector of time averaged firing rates.
The diagonal contains the autocovariance functions (diagonal matrix a(τ )) which
are dominated by a δ-peak at zero time lag and for τ 6= 0 exhibit a continuous
shape mostly determined by refractoriness, the inability of the neuron to fire spikes
in short succession due to the voltage reset, as shown in Figure 2d. The off-diagonal
elements contain the cross covariance functions c(τ ) that originate from interactions.
We therefore decompose the covariance matrix into ¯c = a + c. A basic property of
covariance matrices is the symmetry ¯c(τ ) = ¯cT (−τ ), so we only need to consider
τ > 0 and obtain the solution for τ < 0 by symmetry. Each spike at time t′ can
influence the other neurons at time t > t′. Formally we express this influence of the
history up to time t on a particular neuron i in the network as si(t,{s(t′)t′ < t}),
In the asynchronous state of the
which is a functional of all spike times until t.
Echoes in correlated neural systems
6
network [9] and for small synaptic amplitudes Jij a single spike of neuron j causes
only a small perturbation of si. The other inputs to neuron i effectively act as
additional noise. We may therefore perform a linear approximation of j’s direct
influence on neuron i and average over the realizations of the remaining inputs s\sj,
as illustrated in Figure 2a. This linearization and the superposition principle lead to
the convolution equation
hsi(t)sjis\sj = ri +Z t
−∞
= ri + [hij ∗ (sj − rj)](t),
hij(t, t′)(sj(t′) − rj) dt′
(2)
the functional
δsj(t′)Es\sj
where we define as the linear response kernel hij(t, t′) = D δsi(t)
derivative of si(t) with respect to sj(t′), formally defined in Appendix A. The kernel
hij quantifies the effect of a single spike at time t′ of neuron j on the expected
density of spikes of neuron i at time t by a direct synaptic connection from neuron
j to neuron i, earlier introduced as the “synaptic transmission curve” [34]. This
density vanishes for t < t′ due to causality. For the stationary network state studied
here, the kernel further only depends on the time difference τ = t−t′. In a consistent
linear approximation there are no higher order terms, so the effects of two inputs
sj and sk superimpose linearly. In general, the response of a cell is typically supra-
linear in the number of synchronously arriving excitatory spikes [34–36]. If, however,
the network state to be described is sufficiently asynchronous, as is the case here,
a linear approximation is adequate. Using this linear expansion and the definition
of the covariance matrix, the off-diagonal elements of the covariance matrix fulfill a
linear convolution equation
c(τ ) = [h ∗ (a + c)](τ )
for τ > 0.
(3)
The equation is by construction valid for τ > 0 and needs to be solved simultaneously
obeying the symmetry condition c(t) = c(−t)T . Up to linear order in the interaction,
equation (3) is the correct replacement of the intuitive self-consistency argument
sketched in Figure 1.
In order to relate the kernel hij to the leaky integrate-and-fire model, we
employ earlier results based on Fokker-Planck theory [33]. For small synaptic
amplitudes J ≪ Vθ − Vr and weak pairwise covariances the summed synaptic input
τmPN
j=1 Jijsj(t − d) can be approximated as a Gaussian white noise with mean
µi = τmPj Jijrj and variance σ2
ijrj for neurons firing with rates rj and
i = τmPj J 2
Echoes in correlated neural systems
7
i ), taking into account the dependence of µi and σ2
Poisson statistics. For short synaptic time constants τs ≪ τm the stationary firing
rate ri(µi, σ2
i ) (A.1) depends on these two moments [33]. In a homogeneous random
recurrent network the input to each neuron is statistically the same, so the stationary
rate ri = r is identical for all neurons. It is determined by the self-consistent solution
of r(µi, σ2
i on the rate r itself [9].
0 hij(t) dt of the response kernel is equivalent to the DC
susceptibility wij = ∂ri
[37] and has earlier been termed “asynchronous gain” [34].
∂rj
The approximation is second order in the synaptic amplitude wij = αJij + βJ 2
ij. The
first order term originates from the dependence of µi on rj, the second order term
stems from the dependence of σ2
The integral wij = R ∞
i on rj (A.3).
Throughout this work we choose the working point of the neurons in the network
such that the firing of the cells is driven by strong fluctuations and the mean
membrane potential is close to threshold, in order to have irregular firing. This is
achieved by appropriate choices of the external driving Poisson sources, as described
in Appendix E. For the small networks considered here, it is realistic to assume that
about 50 percent of the inputs to a neuron come from outside the local network [38].
Figure 2b shows the deflection of the firing rate from baseline caused by an impulse
in the input averaged over many repetitions. For sufficiently strong fluctuations σi,
the quadratic term in Jij is negligible, as seen from Figure 2c. For smaller membrane
potential fluctuations σi we expect the linear approximation to be less accurate.
The kernel shows exponential relaxation with an effective time constant τe that
depends on the working point (µi, σi) and the parameters of the neuron, which is
obtained by a least squares fit of a single exponential to the simulated response in
Figure 2b for one particular amplitude Jij. We therefore approximate the response
hij(t) as
wijh(t) = Θ(t − d)
wij
τe
e− t−d
τe ,
(4)
where d is the synaptic delay and Θ the Heaviside step function.
In experiments covariance functions are typically averaged over statistically
equivalent pairs of neurons. Such averages are important, because they determine
the behavior of the network on the macroscopic scale of populations of neurons. We
therefore aim at a corresponding effective theory. Assuming identical dynamics (1),
all neurons have, to good approximation, the same autocovariance function a(t) and
response kernel h(t). So each incoming excitatory impulse causes a response w h(t),
an inhibitory impulse −gw h(t). We define the covariance function averaged over all
Echoes in correlated neural systems
8
a
)c
1
−
0
1
(
j
i
w
1
0
l
e
n
r
e
k
−1
l
a
r
g
e
n
t
i
b
)
s
/
1
(
)
t
(
j
i
h
10
0
l
e
n
r
e
k
−10
−10
0
t (ms)
10
d
2
s
/
1
)
t
(
a
.
v
o
c
t
o
u
a
1000
0
−5
0
Jij (mV)
5
−20
0
20
t (ms)
Figure 2. Mapping the integrate-and-fire dynamics to a linear coupling kernel. (a)
The kernel hij determines the transient effect of an incoming impulse at time point
tj (black arrow) on the density si(t) of outgoing action potentials, averaged over
realizations of the stochastic activity of the remaining inputs (indicated as red and
blue triangles). (b) Response kernel (4) (green) compared to direct simulation for
an impulse of amplitude J = 1 mV (black dots) and J = −1 mV (gray dots). Time
constant τe = 4.07 ms determined by a least squares fit to a single exponential. The
background activity causes a mean µi = 15 mV and fluctuations σi = 10 mV. (c)
Linear and quadratic dependence (A.3) of the integral response wij on Jij (dark
gray curve) and linear term alone (light gray line). (d) Autocovariance function of
the spike train with a δ peak at t = 0 and covariance trough due to refractoriness.
pairs of excitatory neurons as cee(τ ) = 1
N 2 Pi,j∈E,i6=j cij(τ ), (setting N(N − 1) ≃ N 2
for N ≫ 1), where E denotes the set of all excitatory neurons. The pairings cei, cie,
Inserting equation (3) into the average cee(τ ),
and cii are defined analogously.
the first term proportional to the autocovariance a(t) only contributes if neuron
j projects to neuron i. For fixed i, there are K such indices j, so the first term
yields Pi,j∈E,i6=j hij ∗ aj = NKw h ∗ a. The second sum Pi,j∈E,i6=j hik ∗ ckj can
be decomposed into a sum over all intermediate excitatory neurons k ∈ E and
over all inhibitory neurons k ∈ I projecting to neuron i. Replacing the individual
Echoes in correlated neural systems
9
covariances by their population average, cee and cie, respectively, and considering the
number of connections and their amplitude we obtain wNK h ∗ (cee − γg cie), with
γ = Ni/Ne and g = wi/we the relative number of inhibitory neurons and the relative
linearized inhibitory synaptic amplitude. Similar relations hold for the remaining
three averages, so we arrive at a two-by-two convolution matrix equation for the
pairwise averaged covariances for τ > 0
c(τ )
(5)
= [h ∗ Mc] (τ ) + Q [h ∗ a] (τ )
1 −γg ! , Q =
cii(τ ) ! .
with M = Kw 1 −γg
and c(τ ) = cee(τ ) cei(τ )
cie(τ )
Kw
1 −g ! ,
N 1 −g
The convolution equation only holds for positive time lags τ . For negative
time lags it is determined by the symmetry c(−τ ) = cT (τ ). The solution
of this equation can be obtained by an extension of the method used in [26]
employing Wiener-Hopf theory [39] to the cross spectrum C(ω) = R ∞
−∞ c(t)e−iωt dt
in frequency domain, as shown in Appendix B (here capital
letters denote the
Fourier transform of the respective lower case letters). With the definition
of the propagator P(ω) = (1 − MH(ω))−1 the cross spectrum takes the form
C(ω) = P(ω)(cid:2)D+(ω)Q + D+(−ω)QT − A(ω)H(ω)2QMT(cid:3) PT (−ω), where we split
the term H(ω) A(ω) = D+(ω) + D−(ω) so that d+(τ ) and d−(τ ) vanish for times
τ < 0 and τ > 0, respectively. For the averaged cross spectrum, the matrix Q is
defined in (5), the non-averaged cross spectrum can be recovered as a special case
setting Q = M = W, because the convolution equations (3) and (5) have the same
structure and symmetries. If all eigenvalues of MH(ω) have an absolute value smaller
than unity, the propagator P(ω) can be expanded into a geometric series in which
the n-th term contains only interactions via n steps in the connectivity graph. This
expansion has been used to obtain the contribution of different motifs to the integral
of covariance functions [40] and their temporal shape [41].
In the following, we neglect the continuous part of the autocovariance function,
setting a(t) = r δ(t), because the δ-peak is typically dominant. With this
replacement we mainly neglect the trough around 0 due to the relative refractoriness.
An estimate of the error can be obtained considering the respective weights of the
Echoes in correlated neural systems
10
delta peak (r) and the through. From the relation R ∞
−∞ a(t) dt = rCV2 [42, 43] the
integral weight of the trough follows as r(CV2 − 1), which is small for irregular spike
trains with a coefficient of variation CV close to unity.
For an arbitrary causal kernel h it follows that D+(ω) = r H(ω), so the cross
spectrum takes the form
C(ω) = r
Kw
N 1 −g
1 −g ! U(iω) + c.c. trans.
1 1 ! U(iω)2
N 1 1
(Kw)2
+ r(1 + g2γ)
(6)
1
with U(z) =
and L
H −1(−iz) − L
= Kw(1 − γg).
The limit ω → 0 corresponds to the time integral of the cross covariance function,
approximating the count covariance for long time bins [2, 8]. With A(0) = r, the
integral correlation coefficient averaged over neuron pairs fulfills the equation
C(0)
A(0)
=
Kw
N
1
1 − L 2
1 − g −2g !
1 − g
1 1 ! ,
(1 − L)2 1 1
1 + g2γ
+
(Kw)2
N
(7)
which has previously been derived from a noise-driven linear rate dynamics [12].
The quantity L plays a key role here:
It determines the feedback magnitude of
in-phase fluctuations of the excitatory and inhibitory population. Stability of the
average firing rate requires this feedback to be sufficiently small [9], i.e. L < 1,
indicated by the pole in equation (7) at L = 1. Typically cortical networks are in
the balanced regime [9, 13, 14], i.e. L < 0. For such inhibition dominated networks,
the denominator in equation (7) is larger than unity, indicating a suppression of
covariances [12]. As shown in Figure 3a, the prediction (7) agrees well with the results
of simulations of leaky integrate-and-fire networks for a broad range of network sizes
N.
Previous works have investigated neural networks in the thermodynamic limit
N → ∞ [11, 13], scaling the synaptic amplitudes J ∝ 1/√N in order to arrive
Echoes in correlated neural systems
a
100
)
0
(
C
10−1
10−2
10−3
Cee(0)
Cei(0)
Cii(0)
b
)
2
−
0
1
(
i
e
c
i
r
e
r
0
N N
5
0
11
N = 104
N = 2 104
N = 5 104
104
log(N)
105
−20
0
t (ms)
20
Figure 3. Shape invariance of covariances with network scale. Synapses scaled
as J ∝ 1/N with network size N to conserve the population feedback L = const.
(a) Integral correlation coefficient averaged over different pairs of neurons and
theory (7) confirming the ∝ 1/N dependence. (b) Rescaled covariance functions
N cei averaged over excitatory-inhibitory pairs of neurons for different network sizes
(color coded) and theory (14) (black).
at analytical results. Such a scaling increases the feedback on the network level
L ∝ √N and therefore changes the collective network state. Equation (7) provides
an alternative criterion to scale the synapses while keeping the dynamics comparable:
The two terms in equation (7) depend differently on the feedback L. In order to
maintain their ratio, we need to keep the population feedback L constant. The
synaptic amplitude J (approximately ∝ w (A.3)) hence needs to scale as J ∝ 1/N.
In addition, the response kernel h of each single neuron must remain unchanged,
requiring the same working point, characterized by the mean µi and fluctuations σi
in the input into each cell. Constant mean directly follows from L = const., but
the variance due to local input from other neurons in the network decreases as 1/N.
To compensate, we supply each neuron with an additional external uncorrelated
balanced noise whose variance appropriately increases with N (as described in detail
in Appendix E). Figure 3b shows that the shape of the covariance functions is
invariant over a large range of network sizes N, in particular the apparent time
lag of inhibition behind excitation observed as the asymmetry in Figure 3b does not
vanish in the limit N → ∞. The magnitude of the covariance decreases as 1/N as
expected from equation (7), because Kw = const.
Global properties of the network dynamics can be inferred by considering the
spectrum of equation (6), those complex frequencies zk at which the expression has
a pole due to the function U(z). These poles are resonant modes of the network,
Echoes in correlated neural systems
12
d = 5.0
d = 3.0
d = 1.0
d = 0.5
−20
0
−10
damping R(τez)
−5
0
5
0
−5
a
)
z
e
τ
(
I
.
q
e
r
f
.
c
s
o
c
e
e
c
2e
r
0.1
0.0
−0.1
500
0
cee
b
0
−2
L
−4
exp. decaying
damped osc.
osc.
1/4
1/2
fcrit.d
5
10
τe/d
d
ae
)
t
(
e
a
2e
r
0
−10
0
10
t(ms)
−1
−40 −20
0
20
40
t(ms)
Figure 4. Phase diagram determined by spectral analysis. Throughout all panels
colors correspond to delays as given in a. (a) Each dot in the inset represents a
pole zk (8) for delay d = 1 ms (two rightmost poles appear as one point). The two
rightmost poles change with delay d. At d = 0.753 ms (gray St. Andrew’s cross) (9)
the poles become a conjugate pair, at d = 6.88 ms (black crosses) (11) both poles
have a zero real part, causing oscillations (Hopf bifurcation). (b) Right: Phase
diagram spanned by τe/d and feedback L. Onset of oscillations below the black
curve (11) (Hopf bifurcation, black crosses in a), damped oscillations below the gray
curve (9) (gray cross in a). Left: Oscillation frequency (11) at the Hopf bifurcation.
(c) Averaged cross covariance between excitatory neurons and theory (14) (black).
Simulated data averaged over 106 neuron pairs for 100 s. (d) Autocovariance of
excitatory neurons (δ-peak not shown) averaged over 2500 neurons for 100 s.
where the real part ℜ(zk) denotes the damping of the mode, and the imaginary part
ℑ(zk) is the oscillation frequency. A pole appears whenever zk is a single root of
U −1(zk) = H −1(−izk) − L = 0. With the Fourier representation H(ω) = e−iωd
of
the response kernel (4), the poles correspond to the spectrum of the delay differential
equation τe
dt (t) = −y(t) + Ly(t − d), (cf. [44]) which describes the evolution of the
population averaged activity. As shown in Appendix C, the location of the poles
can be expressed by the branches k of the Lambert W function, the solution of
1+iωτe
dy
Echoes in correlated neural systems
WkeWk = x [45], as
zk = −
1
τe
+
1
d
Wk(L
d
τe
e
d
τe ) k ∈ N0.
13
(8)
The spectrum only depends on the population feedback L, the delay d and the
effective time constant τe of the neural response kernel. This explains why keeping
L constant while scaling the network in Figure 3 yields shape invariant covariance
functions. A typical spectrum is shown in Figure 4a as an inset, where each dot
marks one of the poles zk (8). The two principal branches of Wk are the modes
with the largest real part ℜ(zk), and hence with the least damping, dominating the
network dynamics. The remaining branches appear as conjugate pairs and their
real parts are more negative, corresponding to stronger damping. Investigating the
location of the principal branches therefore enables us to classify the dynamics in the
network. Their dependence on the delay is shown in Figure 4a as a parametric plot in
d. The point at which the two real principal solutions turn into a complex conjugate
pair marks a transition from purely exponentially decaying dynamics to damped
oscillations. This happens at sufficiently strong negative coupling L or sufficiently
long delay d, precisely when the argument of Wk is smaller than −e−1 [45], leading
to the condition
L < −
τe
d
e− d
τe
−1.
(9)
The gray cross marks this point in Figure 4a, the gray curve shows the corresponding
relation of feedback and delay in the phase diagram Figure 4b.
In the region
below the curve, the dominant mode of fluctuations in the network is thus damped
oscillatory, whereas above the curve fluctuations are relaxing exponentially in time.
For sufficiently long delay d the principal poles may assume positive real values,
leading to ongoing oscillations, a Hopf bifurcation. The condition under which
this happens can be derived from H −1(ωcrit.) = L, as detailed in Appendix C.
Equating the absolute values on both sides leads to the condition ωcrit.τe = √L2 − 1:
oscillations can only be sustained, if the negative population feedback is sufficiently
strong L < −1. The oscillation frequency increases the stronger the negative
feedback. The condition for the phases leads to the critical delay required for the
onset of oscillations (see Appendix C for details)
.
(10)
dcrit.
τe
=
π − arctan(√L2 − 1)
√L2 − 1
Echoes in correlated neural systems
14
This relation is shown as the black curve in the phase diagram Figure 4b. The
oscillatory frequency on the bifurcation line, at the onset of oscillations can be
expressed as
2πfcrit.d = ωcrit.d = π − arctan(√L2 − 1),
(11)
which is shown in the left sub-panel of the phase diagram Figure 4b. Consequently,
the oscillation frequency fcrit. at the onset is between (4dcrit.)−1 and (2dcrit.)−1 ,
depending on the strength of the feedback L, approaching fcrit. = (4dcrit.)−1 at the
onset with increasing negative feedback.
Changing the synaptic delay homogeneously for all synapses in the network
allows us to observe the transition of the network from exponentially damped,
to oscillatory damped, and finally to oscillatory dynamics. For a short delay of
d = 0.5 ms the dynamics is dominated by the single real pole near −1/τe (brown
dot in Figure 4a) and the covariance function is exponentially decaying (Figure 4c).
Increasing the delay to d = 1 ms the principal poles split into a complex conjugate
pair as the delay crosses the gray curve in Figure 4b so that side troughs become
visible in the covariance function in Figure 4c. Further increasing the delay, the
network approaches the point of oscillatory instability, where a Hopf bifurcation
occurs, marked by black crosses in Figure 4a and the black curve in Figure 4b.
The damping of oscillations decreases as the system approaches the bifurcation
(Figure 4c). The structure of the auto covariance function of single spike trains
(Figure 4d) is dominated by the dip due to refractoriness of the neuron after
reset. At pronounced network oscillations, the autocorrelation function shows a
corresponding modulation. The neglect of the oscillating continuous part of the
autocorrelation function in the theory does apparently not have a pronounced effect
on the obtained cross correlation functions, evidenced by the good agreement between
direct simulation and theory for d = 5 ms. For weaker oscillations, e.g. at d = 3 ms,
the coherence time of the oscillations is typically shorter than the width of the dip in
the autocorrelation. The phase coherence of the oscillation is then shorter than the
typical inter-spike-interval and single neuron spike trains are irregular. This state is
known as synchronous irregular activity [9], where the activity of a single neuron is
irregular, but collectively the neurons participate in a global oscillation.
If the network is not in the oscillatory state, all modes are damped in time,
all poles zk of the function U(z) appearing in equation (6) lie in the left
i.e.
complex half plane, ℜ(zk) < 0. Hence, the dynamics is stable, and we can expect
Echoes in correlated neural systems
15
a
c
b
echo e spike
echo i spike
cff
d
0.05
e
e
c
2e
r
0.00
cee
0
d
2d
3d
t(ms)
e
0.05
e
i
c
i
r
e
r
0.00
15
20
−20
0
20
cie
f
0.05
i
i
c
i
r
i
r
0.00
cii
−20
0
t (ms)
20
−20
0
t(ms)
20
Figure 5. Composition of covariance functions. (a) Network echo caused by
a spike sent at time t = 0 sets in after one synaptic delay (d = 3 ms): red
curve in c, inhibitory spike, blue curve in c, excitatory spike. (b) Correlated
inputs from the network to a pair of neurons (black circles) cause covariance cff
between their outputs (green curve in c). (d)-(f ) Covariance functions averaged
over pairs of neurons (black dots, d: excitatory pairs, e: excitatory-inhibitory
inhibitory pairs) and theory (14) (gray underlying curves). Inset shows
pairs, f :
the two components from c that are added up.
Echoes in correlated neural systems
16
to obtain a unique solution for the covariance as an observable of this stable
dynamics. We perform the Fourier transform to time domain using the residue
theorem u(t) = 1
U(z) ezt dz = Pzk∈Et Res(U, zk) ezkt, where the integration
path Et proceeds along the imaginary axis from −i∞ to i∞ and is then closed in
infinity in the left half plane (for t > d) or the right half plane (for 0 < t < d) to
ensure convergence, resulting in (see Appendix D for the detailed calculation)
2πiHEt
u(t) =
∞
Xk=0
1
(1 + zkτe) d + τe
Θ(t − d)ezk(t−d).
(12)
The back transform of V (ω) def= U(ω)2 proceeds along similar lines and results in
v(t) =
∞
Xk=0
1
(1 + zkτe) d + τe
ezkt
(1 − zkτe) − Lezkd .
(13)
The population averaged covariance functions in the time domain then follow from
equation (6) for t > 0 as
c(t) = r
Kw
N 1 −g
1 −g ! u(t)
(1 + g2γ) 1 1
1 1 ! v(t),
+ r
(Kw)2
N
(14)
which is the central result of our work. Figure 5 shows the comparison of the theory
(14) with the covariance functions obtained by direct simulation. The analytical
expression unveils that the covariance functions are composed of two components:
the first line in equation (14) has heterogeneous matrix elements and hence depends
Its origin is illustrated in Figure 5a: If
on the neuron types under consideration.
one of the neurons emits a spike, as indicated, this impulse travels along its axon
and reaches the target neurons after one synaptic delay d. Depending on the type
of the source neuron, the impulse excites (synaptic amplitude J) or inhibits (−gJ)
its targets.
Its effect is therefore visible in the pairwise covariance function as a
positive (blue) or negative (red) deflection, respectively in Figure 5c. This deflection
not only contains the direct synaptic effect, but also infinitely many reverberations
Echoes in correlated neural systems
17
of the network, seen formally in equation (12). This expression is not proportional
to the kernel h(t) directly, but is rather a series including the whole spectrum of the
network. The shape of the spike echo consequently shows onsets of reverberations at
integer multiples of the synaptic delay (Figure 5c), being transmitted over multiple
synaptic connections. The contribution of the second line in equation (14) follows the
intuitive argument illustrated in Figure 5b. The incoming activity from the network
to a pair of neurons is correlated. As the input statistics is the same for each neuron,
this contribution is identical for any pair of neurons (green curve in Figure 5c). The
sum of both components results in the covariance functions shown in Figure 5d-
f. The same analytical solution is shown for different delays in Figure 4c showing
good agreement with direct simulation. For different sizes of simulated networks
in Figure 3b-d the analytical expression (14) explains why the spike echo does not
become negligible in the thermodynamic limit N → ∞: for fixed population feedback
L, both contributions in equation (14) scale as 1/N, so the relative contribution of
the echo stays the same. This also explains the apparent paradox (see Figure 1), that
covariance functions in recurrent networks not only depend on the input statistics,
but in addition the spike feedback causes a reverberating echo. The power spectrum
of population-averages is dominated by pairwise covariances, explaining the different
spectra observed in the excitatory and inhibitory population activity [15]. Scaling
the network such as to keep the marginal statistics of single neurons constant,
J ∝ w ∝ 1/√N [11, 13] changes the spectrum (8), because the feedback increases as
L ∝ √N which can ultimately lead to oscillations as shown in Figure 4c.
3. Discussion
The present work qualitatively explains certain features of the correlation structure
of simultaneously recorded synaptic currents of two cells in vivo. Novel experimental
techniques are able to separate contributions of excitatory and inhibitory inputs [29].
We calculate such covariances in a random network and show that the covariance
between synaptic impulses decomposes into a linear combination of the covariance
of the spiking activity and the autocovariance functions (see caption of Figure 6).
Each synaptic impulse has a certain time course, here modeled as a first-order low-
pass filter with time constant τs = 2 ms (see (1)). The covariances between these
filtered currents are shown in Figure 6. Their temporal structure resembles those
measured in cortex in vivo [29, their Figure 1e,f]: covariances between afferents of
Echoes in correlated neural systems
18
cIeIe
cIiIi
cIeIi
cIiIe
)
2
V
m
(
I
I
c
0.1
0.0
−0.1
−10
0
10
t (ms)
Figure 6. Covariance between synaptic currents of a pair of neurons in a recurrent
random network in analogy to in vivo experiments [29]. Covariance of excitatory
contributions (blue, cIeIe = q ∗ (J 2pKae + J 2K 2cee)), analogously between
inhibitory contributions (red), between excitatory and inhibitory contribution
(green, cIeIi = −q ∗ (gJ 2γK 2cei)), and CIiIe analogously (brown). Currents filter
spiking input by an exponential kernel with time constant τs = 2 ms (1), leading
to the filtering of the covariances by q(t) = τ 2
2τs
e−t/τs.
m
the same type are monophasic and positive, while the covariances between excitatory
and inhibitory afferents are biphasic and mostly negative. The lag reported between
inhibitory and excitatory activity [29, their Figure 2b], which was also observed in
binary random networks [11, 25], is explained by the echo of the spike contributing
to the covariance function. In contrast to previous work, we take the delayed and
pulsed synaptic interaction into account. Without delays and with binary neurons
[11, 25] the echo appears as a time lag of inhibition with respect to excitation.
Measurements of membrane potential fluctuations during quiet wakefulness in
the barrel cortex of mice [46] showed that correlations between inhibitory neurons are
typically narrower than those between two excitatory neurons [46, their Figure 4A, 5B
and Figure 5C,E]. These results qualitatively agree with our theory for covariances
between the spiking activity, because fluctuations of the membrane potential are
uniformly transferred to fluctuations of the instantaneous firing intensity. The direct
measures of spiking activity [46, their Figure 6] confirm the asymmetric correlation
between excitation and inhibition. The low correlation between excitatory neurons
reported in that study may partly be due to the unnormalized, firing rate dependent
measure and the low rates of excitatory neurons.
The qualitative features of the cross correlation functions, namely their different
widths and their asymmetry for excitatory-inhibitory pairs, are generic and agree
Echoes in correlated neural systems
19
with experimental results. They are fully explained by the decomposition into an
echo term and a term corresponding to the feed-forward transmission of correlation.
This decomposition merely relies on the fact that the autocorrelation of a spike train
has a delta peak and that a spike triggers a response in the target cell with positive
or negative sign for excitatory and inhibitory connections, respectively. Hence, we
expect that these features are robust and survive also for more realistic network
models with many heterogeneous subpopulations. For weakly correlated fluctuations,
if the input to each cell is sufficiently noisy, a linear approximation of the neuronal
response provides a viable first order approximation also for non-linear neuron
dynamics, as shown here. We suspect that a deeper reason why such a linearization
is possible is the inherent decorrelation [12] by negative feedback in networks in the
inhibition dominated regime. The decorrelation keeps population fluctuations small
and hence prevents strong excursions that would exceed the validity of the linear
approximation.
Oscillations in the γ range (25 − 100 Hz) are ubiquitous in population measures
of neural activity in humans, and have earlier been explained in networks of leaky
integrate-and-fire model neurons [9] by the Hopf bifurcation induced by delayed
negative feedback. For the regime of high noise we here uncover a simpler analytical
condition for the onset (10) and frequency (11) of fast global oscillations. For lower
noise, deviations of the non-linear leaky integrate-and-fire dynamics from the linear
theory presented here are expected.
A traditional motivation to scale the network size to infinity is to obtain an
analytic solution for an otherwise hard problem. Biologically realistic networks have
a finite size. In the present work we have determined the correlation structure for
such finite sized networks analytically. We present the scaling of the correlation
functions in Figure 3 to relate our work to previous results that applied scaling
arguments to obtain the correlation structure [11]. The latter work investigated
networks of binary neurons without conduction delays and assumed a scaling of
the synaptic amplitudes J ∝ 1/√N. Such a scaling increases the overall feedback
strength L ∝ pNw ∝ √N. This has two consequences: Firstly, as seen from (7), the
relative contributions of the spike echo and the feed forward term change with L and
hence with network size. Therefore the shape of correlation functions depends on the
network size. Secondly, the overall dynamic regime of the network is affected by the
scaling. This can be seen from Figure 4b. For non-zero synaptic conduction delays,
the network eventually becomes oscillatory if the network size exceeds a certain value.
Echoes in correlated neural systems
20
This happens precisely at the point where L crosses the bifurcation line shown in
Figure 4b. We therefore propose an alternative scaling J ∝ 1/N here for which we
show that it preserves the shape of correlation functions and the overall network state.
Only the magnitude of the cross correlation functions decreases ∝ 1/N. However,
a caveat of this scaling is that while it preserves the global network properties, it
affects the working point of each individual neuron, because the fluctuations due
to the local synaptic input decrease ∝ 1/√N. We alleviated this shortcoming by
supplying each neuron with additional, uncorrelated noise.
in contrast,
Our results are based on a simplified network model composed of two
homogeneous (excitatory and inhibitory) subpopulations of leaky integrate-and-
fire neurons. The real cortex,
is highly heterogeneous in several
respects: Its layered structure with layer-specific neuron and connection properties
(e.g.
time constants, spike thresholds, synaptic weights, in-degrees) requires the
distinction of more than two subpopulations. Even within each subpopulation and
for each combination of subpopulations, the neuron and connection parameters are
not constant but broadly distributed. Further, a plethora of cortical neurons, in
particular various types of interneurons, exhibit a much richer dynamical repertoire
(e.g. resonating behavior) than the leaky integrate-and-fire neuron model (regular
spiking integrator).
In principle, the mathematical framework presented in this
article can be extended to networks composed of n (n > 2) heterogeneous
subpopulations of different neuron types. This would require to account for
subpopulation-specific working points (e.g. due to layer-specific firing rates; see [47])
and for the effect of parameter distributions [48, 49] and the single-neuron dynamics
on the effective linearized subpopulation responses. This extended theory would
result in an n-dimensional linear algebraic equation for the subpopulation-averaged
cross spectra, similar to (6) where n = 2. For n > 2, this equation most likely needs
to be solved numerically.
A fundamental assumption of the presented theory is that the network states
show irregular single neuron dynamics. This requirement arises from the analytical
description replacing spike trains by spike densities and a stochastic realization
of spikes. Regular spike trains are outside the scope of such a description.
Moreover, the approximation of the neuronal response to linear order is only a viable
approach in sufficiently asynchronous network states with low correlations. States
with stronger correlations, such as observed in convergent-divergent feed-forward
structures [34, 50], require an explicit treatment of the non-linear response [36].
Echoes in correlated neural systems
21
From a physics viewpoint, neuronal networks unite several interesting properties.
They do not reach thermodynamic equilibrium even in the stationary state, as
detailed balance does not hold for all pairs of states of the system.
In detailed
balance, the rate of transition from one state to another is equal to the rate of the
reverse transition. For a leaky integrate-and-fire neuron the state of the neuron
is uniquely determined by its membrane voltage.
In the stationary state neurons
fire with a constant rate, so there is a continuous flux of the neurons’ voltage
from reset up to threshold.
Imagining a discretization of the voltage axis we see
that a pair of adjacent voltage-intervals between reset and threshold is more often
traversed from lower to higher voltage than in the reverse direction, so obviously
detailed balance does not hold even for a single neuron in the stationary state.
Moreover, the interaction between pairs of neurons is directed, delayed, pulsed, and
depends on the flux of the sending neuron’s state variable at threshold. In contrast,
pairwise interactions frequently studied in physics, like the Coulomb interaction or
exchange interaction, can be expressed by a pair potential and are thus symmetric
(undirected), instantaneous, and depend directly on the state variables (e.g. spatial
coordinates or spins) of the pair of interacting particles. Non-equilibrium systems
are at the heart of ubiquitous transport phenomena, like heat or electric conduction.
Understanding fluctuations in such a system marks the starting point to infer
macroscopic properties by the assertion of the fluctuation-dissipation theorem that
connects microscopic fluctuations to macroscopic transport properties. Despite
the non-equilibrium dynamics and the non-conservative pairwise interaction,
in
this manuscript we develop a simple analytical framework merely based on linear
perturbation theory that explains time dependent covariance functions of the activity
of pairs of integrate-and-fire model neurons in a recurrent random network. Formally
our approach resembles the step from the kinetic Ising model near equilibrium to its
non-equilibrium counterpart, the network of binary neurons [25]. A difference is the
spiking interaction considered in our work, which led us to the describe each neuron
in terms of the flux over threshold (spike train) rather than by its state variables
(membrane voltage and synaptic current). In this respect, we follow the established
mean-field approach for spiking neuronal systems [9, 14]. However, while this mean
field approach proceeds by assuming vanishing correlations to obtain the dynamics of
the ensemble averaged activity, we here derive and solve the self-consistency equation
for the pairwise averaged covariances of the microscopic system.
The typical time scale of covariance functions found here coincides with the time
Echoes in correlated neural systems
22
window of biological synaptic plasticity rules [30], so that non-trivial interactions
of dynamics and structure are expected.
It is our hope that the novel capability
to resolve the temporal structure of covariances in spiking networks presented here
proves useful as a formal framework to further advance the theory of these correlated
non-equilibrium systems and in particular serves as a further stepping stone in the
endeavor to understand how learning on the system level is implemented by the
interplay of neuronal dynamics and synaptic plasticity.
Acknowledgments
We are thankful to Birgit Kriener, Sonja Grun and George Gerstein for discussions
and exploratory simulations in fall 2005 that led to the observation of asymmetric
covariance functions. Partially supported by the Helmholtz Association: HASB and
portfolio theme SMHB, the Next-Generation Supercomputer Project of MEXT, EU
Grant 15879 (FACETS), EU Grant 269921 (BrainScaleS). All network simulations
carried out with NEST (http://www.nest-initiative.org).
Appendix A. Response kernel of the LIF model
The response kernel kernel hij needs to be related to the dynamics of the neuron
model (1). Here we present an approximation of this kernel which is sufficiently
accurate to allow quantitative predictions, but yet simple enough to enable an
analytical solution for the correlation structure.
If the synaptic time constant is
short τs ≪ τm, the synaptic amplitude J can be thought of as the amplitude of
the jump in the membrane potential V caused upon arrival of an incoming impulse.
If correlations between incoming spike trains are sufficiently small, the first and
second moments of the summed impulses τmPj Jijsj(t − d) are µi = τmPj Jijrj
and σ2
ijrj, respectively, if the inputs’ statistics can be approximated by
Poisson processes of rate rj each. For small J and a high total rate, the system of
differential equations (1) is hence approximated by a stochastic differential equation
driven by a unit variance Gaussian white noise ξ
i = τmPj J 2
τm
τs
dVi
dt
dIi
dt
= − Vi + Is,i(t)
= − Ii + µi + σi√τmξ(t).
Echoes in correlated neural systems
The stationary firing rate in this limit is given by [33]
23
(A.1)
r−1
i
f (y)
= τr + τm√π (F (yθ) − F (yr))
(1 + erf(y)) F (y) =Z y
= ey2
Vθ,r − µi
+
α
f (y) dy
α = √2ζ(
1
2
),
2r τs
τm
with yθ,r =
σi
with Riemann’s zeta function ζ. The rate ri is the density of action potentials per
time. The response of the firing density of the neuron i at time point t with respect
to a point-like deflection of the afferent input sj at time point t′ defines the response
kernel as the functional derivative
δsj(t′)(cid:29)s\sj
(cid:28) δsi(t)
= lim
ǫ→0
1
ǫ hsi(t,{s(τ ) + ǫδ(τ − t′)ejτ < t})
− si(t,{s(τ )τ < t})is\sj
def= hij(t − t′) = wij h(t − t′).
Here we used the homogeneity, namely the identical input statistics of each neuron
i, leading to the same temporal shape h(t) independent of i and the stationarity, so
that the kernel only depends on the time difference t − t′. We choose h(t) to have
unit integral and define wij as the integral of the kernel. We determine the temporal
integral of the kernel as
hij(t) dt
(A.2)
wij = Z ∞
−∞
=
∂ri
∂rj
.
The second equality holds because the integral of the impulse response equals the
step response [51]. Further, a step in the density sj corresponds to a step of rj. Up to
linear approximation the effect of the step in the rate rj on the rate ri can be expressed
by the derivative considering the perturbation of the mean µi and the variance σ2
i
∂r−1
upon change of rj. Using equation (A.1) we note that by chain rule ∂ri
.
i
∂rj
∂rj
The latter derivative follows as ∂r−1
terms yields ∂r−1
∂yA
with yA = A−µi
σi
= √πτmf (yA) with yA ∈ {yθ, yr}. The second derivative evaluates
2p τs
. The first derivative in both
τmJij − A−µi
(1 + A−µi
σi
= −τm
= −r2
= ∂r−1
i
∂yθ
= − 1
σi
to ∂yA
∂ri
τmJ 2
ij
2σi
∂yθ
∂rj
+ ∂r−1
i
∂yr
∂yr
∂rj
+ α
Jij
2σi
). So
Jij
σi
i
∂rj
σ2
i
τm
i
i
Echoes in correlated neural systems
together we obtain
= αJij + βJ 2
wij
ij
with α = √π(τmri)2 1
σi
and β = √π(τmri)2(cid:18)f (yθ)
(f (yθ) − f (yr))
Vθ − µi
2σ3
i − f (yr)
i (cid:19) .
Vr − µi
2σ3
24
(A.3)
Appendix B. Cross spectral matrix in frequency domain
The autocovariance A(ω)
in the frequency domain (F (ω) = F [f ](ω) =
R ∞
−∞ f (t)e−iωt dt) has two different terms. The first term is a constant r due to the
spiking with rate r resulting from the delta peak rδ(t) in time domain. The second
term is the continuous function ac(t), for example due to refractoriness of the neuron.
Further follows from a(t) = a(−t) that A(ω) = A(−ω). For τ > 0 the covariance
matrix fulfills the linear convolution equation (5). As this equation only holds for the
positive half of the time axis, we cannot just apply the Fourier transform to obtain
the solution. For negative time lags τ < 0 the covariance matrix is determined
by the symmetry c(τ ) = cT (−τ ). Here we closely follow [26] and employ Wiener-
Hopf theory [39] to derive an equation for the cross spectral matrix in the frequency
domain that has the desired symmetry and solves (5) simultaneously. To this end
we introduce the auxiliary matrix b(τ ) = (h ∗ (Mc)) (τ ) + Q(h ∗ a)(τ ) − c(τ ) for
−∞ < τ < ∞. Obviously, b(τ ) = 0 for τ > 0. Since the defining equation
for b holds on the whole time axis, we may apply the Fourier transform to obtain
B(ω) = H(ω) (MC(ω) + QA(ω)) − C(ω). Solving for C
C(ω) = (1 − MH(ω))−1 (H(ω)A(ω)Q − B(ω))
(B.1)
and using the symmetry C(ω) = CT (−ω) we obtain the equation
(H(ω)A(ω)Q − B(ω))(cid:0)1 − MT H(−ω)(cid:1)
= (1 − MH(ω))(cid:0)H(−ω)A(−ω)QT − BT (−ω)(cid:1) .
We observe that QMT = MQT is symmetric and with A(ω) = A(−ω) the term
proportional to H(ω)2 cancels on both sides, remaining with
(B.2)
H(ω)A(ω)Q + (1 − MH(ω)) BT (−ω)
= H(−ω)A(ω)QT + B(ω)(cid:0)1 − MT H(−ω)(cid:1) .
Echoes in correlated neural systems
25
We next introduce D(ω) = H(ω)A(ω) which we split into D(ω) = D+(ω) + D−(ω),
chosen such that d+(t) (in time domain) vanishes for t < 0 and d−(t) vanishes
for t > 0. Consequently the Fourier transforms of both terms may have poles in
distinct complex half-planes: D+(ω) may only have poles in the upper half plane
ℑ(ω) > 0 and the function vanishes for limω→∞,ℑ(ω)<0 D+(ω) = 0, following from
the definition of the Fourier integral. For D−(ω) the half planes are reversed. The
analytical properties of H(ω) are thus similar to those of D+(ω), those of B(ω) are
similar to D−(ω). We sort the terms in (B.2) such that the left hand side only
contains terms that vanish at infinity in the lower half plane ℑ(ω) < 0, the right
hand side those that vanish in infinity in the upper half plane ℑ(ω) > 0
(B.3)
D+(ω)Q − D−(−ω)QT + (1 − MH(ω)) BT (−ω)
= D+(−ω)QT − D−(ω)Q + B(ω)(cid:0)1 − MT H(−ω)(cid:1) .
The left hand side consequently is analytic for ℑ(ω) < 0 the right hand side is analytic
for ℑ(ω) > 0, so (B.3) defines a function that is analytic on the whole complex plane
and that vanishes at the border for ω → ∞. Hence by Liouville’s theorem it is 0
and we can solve the right hand side of (B.3) for B
B(ω) = (cid:0)D−(ω)Q − D+(−ω)QT(cid:1)(cid:0)1 − MT H(−ω)(cid:1)−1
Inserted into (B.1) this yields with the definition P(ω) = (1 − MH(ω))−1
.
C(ω) = P(ω)(cid:0)H(ω)A(ω)Q −(cid:0)D−(ω)Q − D+(−ω)QT(cid:1) PT (−ω)(cid:1)
= P(ω)(cid:0)(D+(ω) + D−(ω)) Q(1 − MT H(−ω))
−D−(ω)Q + D+(−ω)QT(cid:1) PT (−ω)
= P(ω)(cid:0)D+(ω)Q + D+(−ω)QT − A(ω)H(ω)2QMT(cid:1) PT (−ω).
The latter expression can be brought to the form
C(ω) = D+(ω)P(ω)Q + D+(−ω)QT PT (−ω)
+(cid:0)D+(ω)H(−ω) + D+(−ω)H(ω) − A(ω)H(ω)2(cid:1) × (B.4)
× P(ω)MQT PT (−ω),
which has the advantage that the first two terms have poles in distinct half planes
ℑ(ω) > 0, and ℑ(ω) < 0, respectively. This means these terms only contribute for
Echoes in correlated neural systems
26
positive and negative times, respectively, the last term contributes for positive and
negative times.
Appendix C. Spectrum of the propagator
With the Fourier representation H(ω) = e−iωd
of the delayed exponential kernel (4)
1+iωτe
the averaged cross spectrum (6) contains the two functions U(z) and V (z) = U(z)2
defined on the complex frequency plane z = iω. These functions may exhibit
poles. The function U has a pole zk whenever the denominator has a single root
H −1(−iz) − L = 0 which amounts to the condition (1 + zkτe)ezkd = L. These
complex frequencies can be expressed by the Lambert W function, the solution of
W eW = x [45], by
as
(1 + zkτe)ezkd
= L
(
d
τe
+ zkd)e
d
τe
+zkd = L
d
τe
e
d
τe
zk = −
1
τe
+
1
d
Wk(L
d
τe ),
e
d
τe
(C.1)
leading to (8). The Lambert Wk(x) function has infinitely many branches k [45].
The principal branch has two real solutions, if x > −e−1. The remaining branches
appear in conjugate pairs. For x < −e−1 the principal solutions turn into a complex
conjugate pair. This happens at sufficiently strong negative coupling L or long delays
d
d
τe
e
L
L
or
d
τe < − e−1
τe
d
< −
d
τe
e− d
τe
−1
> W0(−
1
Le
).
The principal poles may assume positive real values, leading to oscillations. The
condition under which this happens can be derived from (C.1). At the point of
transition the pole can be written as z = iωcrit.; it is a solution to (1+iωcrit.τe)eiωcrit.d =
Echoes in correlated neural systems
27
L. In order for this equation to be fulfilled, the absolute value and the phase
must be identical on both sides. The equation for the absolute value requires
1 + (ωcrit.τe)2 = L2. This means there are only oscillatory solutions, if the magnitude
of the feedback exceeds unity L < −1. Since the poles come in conjugate pairs, we
can assume w.l.o.g. that ωcrit. > 0. The condition for the absolute value hence reads
ωcrit.τe = √L2 − 1.
(C.2)
This is the frequency of oscillation at the onset of the Hopf bifurcation. For strong
feedback L ≫ 1 the frequency increases linearly with the magnitude of the feedback.
The condition for the agreement of the phase angles reads ∠(1+iωcrit.τe)+ωcrit.d = 0,
ℜ(1+iωcrit.τe) = − ℑ(eiωcrit.d)
so ℑ(1+iωcrit.τe)
ℜ(eiωcrit.d) = tan ωcrit.d, which leads to tan ωcrit.d = −ωcrit.τe. This
equation has a solution in π
2 ≤ ωcrit.d ≤ π. In the limit of vanishing delay d → 0 the
frequency goes to infinity, as the solution converges to ωcrit.d = π
2 . This corresponds
√L2 − 1) = −√L2 − 1,
4d . Inserting (C.2) leads to tan( d
to the frequency fcrit. = 1
which can be solved for the critical delay
τe
d
τe
=
π − arctan(√L2 − 1)
√L2 − 1
(C.3)
where we took care that the argument of the tangent is in [ π
2 , π]. So with (C.2) and
(C.3) the oscillatory frequency at the transition can be related to the synaptic delay
as
2πfcrit.d = ωd = π − arctan(√L2 − 1).
Appendix D. Back transform by residue theorem
In the non-oscillatory state all poles zk (8) have a negative real part. The function
U(z) = ((1 + zτe)ezd − L)−1 in (6) then has all poles in the left complex half plane,
ℜ(zk) < 0 ∀k. We perform the Fourier back transform
u(t) =
=
−∞
1
2πZ ∞
2πiIEt
1
U(iω) eiωt dω
U(z) ezt dz,
(D.1)
replacing the integration path by a closed contour Et following the imaginary axis
from −i∞ to i∞. In order to ensure convergence of the integral, for t < d we need
Echoes in correlated neural systems
28
ℜ(z) > 0, so we close Et<d in infinity within the right half-plane, where the integrand
vanishes. Since there are no poles in the right half-plane, for t < d the path Et<d
does not enclose any poles, so u(t) = 0. For t ≥ d the path Et≥d must be closed in
the left half-plane to ensure convergence of (D.1), so the residue theorem yields
u(t) = Θ(t − d) Xzk∈Et>d
Res(U, zk) ezkt.
(D.2)
The residue can be calculated by linearizing the denominator of U(zk + z) around zk
(1 + (zk + z)τe)e(zk+z)d − L = (1 + zkτe)ezkdezd − L + zτeezkdezd
= L(ezd − 1) + zτeezkd(1 + zd) + O(z2)
= z(Ld + τeezkd) + O(z2),
which yields
Res(U, zk) = lim
z→0
zU(zk + z)
z
=
1
Ld + τeezkd
= lim
z→0
=
z(Ld + τeezkd)
e−zkd
,
(1 + zkτe)d + τe
where we used (C.1) in the last step. The poles of V (z) = U(z)U(−z) are located
in both half-planes, consequently v(t) is nonzero on the whole time axis. Here we
only calculate v(t) for positive times t > 0, because it follows for negative times by
symmetry v(−t) = v(t). The path has to be closed in the left half-plane, where the
poles zk have the residues
Res(V, zk) = Res(U, zk)U(−zk) =
So applying (D.2) the functions u and v are
1
(1 + zkτe) d + τe
1
(1 − zkτe) − Lezkd .
u(t) =
v(t) =
=
∞
∞
Xk=0
Xk=0
Xk=0
∞
(1 + zkτe) d + τe
Θ(t − d)ezk(t−d)
Res(V, zk)(cid:0)Θ(t)ezkt + Θ(−t)e−zkt(cid:1)
1
1
(1 + zkτe) d + τe
1
(1 − zkτe) − Lezkd ezkt.
(D.3)
Echoes in correlated neural systems
29
The amplitude of the modes decrease with k. For all figures in the manuscript we
truncated the series after k = 30.
Appendix E. Simulation parameters used for figures
All network simulations were performed using NEST [52]. The parameters of the
leaky integrate-and-fire neuron model (1) throughout this work are τm = 20 ms,
τs = 2 ms, τr = 2 ms, Vθ = 15 mV, Vr = 0 mV. All simulations are performed
with precise spike timing and time stepping of 0.1 ms [53]. Figure 1 and Figure 3-
Figure 6 of the main text all consider recurrent random networks of N excitatory and
γN inhibitory leaky integrate-and-fire model neurons receiving input from randomly
drawn neurons in the network and external excitatory and inhibitory Poisson input,
so that the first and second moments are µi = 15 mV and σ2
i = 10 mV, respectively.
Unless stated explicitly, we use N(1 + γ) = 10000 neurons except in Figure 3 where
the number of neurons is given in the legend. Each neuron has K = pN incoming
excitatory synapses with synaptic amplitude J independently and randomly drawn
from the pool of excitatory neurons, and γK = γpN inhibitory synapses with
amplitude −gJ (homogeneous Erdos-R´enyi random network with fixed in-degree),
realizing a connection probability p = 0.1. Cross covariance functions are measured
throughout as the covariance between two disjoint populations of 1000 neurons each
taken from the indicated populations in the network. Correlation functions are
evaluated with a time resolution of 0.1 ms.
In Figure 1 we use a synaptic delay d = 1 ms.
In Figure 3 we keep the feedback of the population rate constant L = Kw(1 −
γg) = const. Increasing the size of the network N the synaptic amplitude J (which
is proportional to w in linear approximation) needs to scale as J = J0N0/N, where
we chose J0 = 0.1 mV and N0 = 10000 here. The variance caused by local input
from the network then decreases with increasing network size ∝ 1/N, while the local
mean is constant because L = const. Each cell receives in addition uncorrelated
external balanced Poisson input, adjusted to keep the mean µi = 15 mV, and
fluctuations σi = 10 mV constant. This is achieved by choosing the rates of
the external excitatory (re,ext., amplitude Jext. = 0.1 mV) and inhibitory (ri,ext.,
amplitude −gJext.) inputs as
re,ext.
= re,0 + rbal
ri,ext. = rbal/g
(E.1)
REFERENCES
with re,0 =
µi − µloc.
Jext.τm
and rbal =
σ2
i − σ2
τmJ 2
loc. − τmre,0J 2
ext.(1 + g2)
ext.
30
,
loc. = τmrKJ 2(1 + γg2) are the mean and
where µloc. = τmrKJ(1 − γg) and σ2
variance due to local input from other neurons of the network firing with rate r.
From L = const. follows that also Kw = const., so that (7) predicts a scaling of
the magnitude of the covariance functions in proportion to 1/N. Other network
parameters are d = 3 ms and g = 5. The firing rate in the network is r = 23.6 Hz.
In Figure 4 and Figure 5 we use N = 104, g = 6, J = 0.1 mV, and the delay d
as described in the captions, the remaining parameters are as in Figure 3.
In Figure 6 we use N = 104, J = 0.1 mV, g = 5, and d = 2 ms and the
remaining parameters as in Figure 3. We obtain the filtered synaptic currents by
filtering the spike trains with an exponential filter of time constant τs = 2 ms. This
results in an effective filter for the cross covariances of q(t) = τ 2
e−t/τs. The different
m
2τs
contributions shown are cIeIe = q∗(J 2pKae +J 2K 2cee), cIiIi = q∗(J 2pKai +J 2K 2cii),
cIeIi = −q ∗ (gJ 2γK 2cei), and cIiIe = −q ∗ (gJ 2γK 2cie), where ∗ denotes the
convolution.
For Figure 2, we simulate two populations of N = 1000 neurons each. Each
neuron receives independent background activity from Poisson processes and in
addition input from a common Poisson process with rate rc = 25 Hz causing in
population 1 a positive synaptic amplitude of J and for population 2 a negative
synaptic amplitude J (J is given on the x-axis). The synaptic amplitude of the
background inputs is Je = 0.1 mV for an excitatory impulse and Ji = −0.5 mV for
an inhibitory impulse. The rates of the excitatory and inhibitory background inputs
are chosen so that the first and second moments µi = τm(Jere + Jiri + J rc) = 15 mV
and σ2
c ) = 10 mV are independent of J. The spikes produced
by each population are triggered to the arrival of an impulse in the common input
and averaged over a duration of 10 s to obtain the impulse response.
i = τm(J 2
i ri + J r2
e re + J 2
References
[1] Gerstein, G. L., & Perkel, D. H. (1969). Simultaneously recorded trains of action
potentials: analysis and functional interpretation. Science 881 (164), 828–830.
[2] Cohen, M. R., & Kohn, A. (2011). Measuring and interpreting neuronal
correlations. Nat. Rev. Neurosci. 14 (7), 811–819. doi:10.1038/nn.2842.
REFERENCES
31
[3] Zohary, E., Shadlen, M. N., & Newsome, W. T. (1994). Correlated neuronal
discharge rate and its implications for psychophysical performance. Nature 370,
140–143.
[4] Abbott, L. F., & Dayan, P. (1999). The effect of correlated variability on the
accuracy of a population code. Neural Comput. 11, 91–101.
[5] von der Malsburg, C. (1986). Am I thinking assemblies? In G. Palm & A. Aertsen
(Eds.), Brain Theory, pp. 161–176. Berlin: Springer-Verlag.
[6] Maldonado, P., Babul, C., Singer, W., Rodriguez, E., Berger, D., & Grun, S.
(2008). Synchronization of neuronal responses in primary visual cortex of monkeys
viewing natural images. J. Neurophysiol. 100 (3), 1523–1532.
[7] Kilavik, B. E., Roux, S., Ponce-Alvarez, A., Confais, J., Gruen, S., & Riehle, A.
(2009). Long-term modifications in motor cortical dynamics induced by intensive
practice. J. Neurosci. 29, 12653–12663.
[8] Ecker, A. S., Berens, P., Keliris, G. A., Bethge, M., & Logothetis, N. K. (2010).
Decorrelated neuronal firing in cortical microcircuits. Science 327 (5965), 584–587.
[9] Brunel, N. (2000). Dynamics of sparsely connected networks of excitatory and
inhibitory spiking neurons. J. Comput. Neurosci. 8 (3), 183–208.
[10] Hertz, J. (2010). Cross-correlations in high-conductance states of a model
cortical network. Neural Comput. 22, 427–447.
[11] Renart, A., De La Rocha, J., Bartho, P., Hollender, L., Parga, N., Reyes, A.,
& Harris, K. D. (2010). The asynchronous state in cortical cicuits. Science 327,
587–590.
[12] Tetzlaff, T., Helias, M., Einevoll, G., & Diesmann, M. (2012). Decorrelation
of neural-network activity by inhibitory feedback. PLoS Comput. Biol. 8 (8),
e1002596.
[13] van Vreeswijk, C., & Sompolinsky, H. (1996). Chaos in neuronal networks with
balanced excitatory and inhibitory activity. Science 274, 1724–1726.
[14] Amit, D. J., & Brunel, N. (1997). Model of global spontaneous activity and local
structured activity during delay periods in the cerebral cortex. Cereb. Cortex 7,
237–252.
[15] Kriener, B., Tetzlaff, T., Aertsen, A., Diesmann, M., & Rotter, S. (2008).
Correlations and population dynamics in cortical networks. Neural Comput. 20,
2185–2226.
REFERENCES
32
[16] Shadlen, M. N., & Newsome, W. T. (1998). The variable discharge of cortical
neurons: Implications for connectivity, computation, and information coding. J.
Neurosci. 18 (10), 3870–3896.
[17] Stroeve, S., & Gielen, S. (2001). Correlation between uncoupled conductance-
based integrate-and-fire neurons due to common and synchronous presynaptic
firing. Neural Comput. 13 (9), 2005–2029.
[18] Tetzlaff, T., Buschermohle, M., Geisel, T., & Diesmann, M. (2003). The spread
of rate and correlation in stationary cortical networks. Neurocomputing 52–54,
949–954.
[19] Moreno-Bote, R., & Parga, N. (2006). Auto- and crosscorrelograms for the
spike response of leaky integrate-and-fire neurons with slow synapses. Phys. Rev.
Lett. 96, 028101.
[20] De la Rocha, J., Doiron, B., Shea-Brown, E., Kresimir, J., & Reyes, A. (2007).
Correlation between neural spike trains increases with firing rate. Nature 448 (16),
802–807.
[21] Shea-Brown, E., Josic, K., de la Rocha, J., & Doiron, B. (2008). Correlation
basic properties and
and synchrony transfer in integrate-and-fire neurons:
consequences for coding. Phys. Rev. Lett. 100, 108102.
[22] Burak, Y., Lewallen, S., & Sompolinsky, H. (2009).
Stimulus-dependent
correlations in threshold-crossing spiking neurons. Neural Comput. 21, 2269–2308.
[23] Rosenbaum, R., & Josic, K. (2011). Mechanisms that modulate the transfer of
spiking correlations. Neural Comput. 23 (5), 1261–1305.
[24] Tchumatchenko, T., Malyshev, A., Geisel, T., Volgushev, M., & Wolf, F. (2010).
Correlations and synchrony in threshold neuron models. Phys. Rev. Lett. 104,
058102.
[25] Ginzburg, I., & Sompolinsky, H. (1994). Theory of correlations in stochastic
neural networks. Phys. Rev. E 50 (4), 3171–3191.
[26] Hawkes, A. (1971). Point spectra of some mutually exciting point process. J.
R. Statist. Soc. Ser. B 33 (3), 438–443.
[27] Stein, R. B. (1965). A theoretical analysis of neuronal variability. Biophys. J. 5,
173–194.
[28] Rauch, A., La Camera, G., Luscher, H., Senn, W., & Fusi, S. (2003). Neocortical
REFERENCES
33
pyramidal cells respond as integrate-and-fire neurons to in vivo like input currents.
J. Neurophysiol. 90, 1598–1612.
[29] Okun, M., & Lampl, I. (2008).
Instantaneous correlation of excitation and
inhibition during sensory-evoked activities. Nat. Neurosci. 11 (5), 535–537.
[30] Morrison, A., Diesmann, M., & Gerstner, W. (2008). Phenomenological models
of synaptic plasticity based on spike-timing. Biol. Cybern. 98, 459–478.
[31] Cai, D., Tao, L., Shakarayev, M. S., Rangan, A. V., McLaughlin, D. W., &
Kovacic, G. (2012). The role of fluctuations in coarse-grained descriptions of
neuronal networks. Comm. Math. 10 (1), 307–354.
[32] Cai, D., Tao, L., & McLaughlin, D. W. (2004). An embedded network approach
for scale-up of fluctuation-driven systems with preservation of spike information.
Proc. Natl. Acad. Sci. USA 101 (39), 14288–14293.
[33] Fourcaud, N., & Brunel, N. (2002). Dynamics of the firing probability of noisy
integrate-and-fire neurons. Neural Comput. 14, 2057–2110.
[34] Abeles, M. (1991). Corticonics: Neural Circuits of the Cerebral Cortex (1st
ed.). Cambridge: Cambridge University Press.
[35] Abeles, M. (1982). Role of cortical neuron: integrator or coincidence detector?
Israel J. Med. Sci. 18, 83–92.
[36] Goedeke, S., & Diesmann, M. (2008). The mechanism of synchronization in
feed-forward neuronal networks. New J. Phys. 10, 015007.
[37] Helias, M., Deger, M., Rotter, S., & Diesmann, M. (2010).
Instantaneous
non-linear processing by pulse-coupled threshold units. PLoS Comput Biol 6 (9),
e1000929.
[38] Stepanyants, A., Martinez, L. M., Ferecsk´o, A. S., & Kisv´arday, Z. F. (2009).
The fractions of short- and long-range connections in the visual cortex. Proc. Nat.
Acad. Sci. USA 106 (9), 3555–3560.
[39] Hazewinkel, M. (Ed.) (2002). Encyclopaedia of Mathematics. Springer.
[40] Pernice, V., Staude, B., Cardanobile, S., & Rotter, S. (2011). How structure
determines correlations in neuronal networks. PLoS Comput. Biol. 7 (5), e1002059.
[41] Trousdale, J., Hu, Y., Shea-Brown, E., & Josic, K. (2012).
Impact of
network structure and cellular response on spike time correlations. PLoS Comput.
Biol. 8 (3), e1002408.
REFERENCES
34
[42] Rieke, F., Warland, D., de Ruyter van Steveninck, R., & Bialek, W. (1997).
Spikes: Exploring the Neural Code. Cambridge, MA: The MIT Press.
[43] Moreno-Bote, R., Renart, A., & Parga, N. (2008). Theory of input spike auto-
and cross-correlations and their effect on the response of spiking neurons. Neural
Comput. 20, 1651–1705.
[44] Guillouzic, S., L’Heureux, I., & Longtin, A. (1999). Small delay approximation
of stochastic delay differential equations. Phys. Rev. E 59, 3970–3982.
[45] Corless, R. M., Gonnet, G. H., Hare, D. E. G., Jeffrey, D. J., & Knuth, D. E.
(1996). On the lambert w function. Advances in Computational Mathematics 5,
329–359.
[46] Gentet, L., Avermann, M., Matyas, F., Staiger, J. F., & Petersen, C. C. (2010).
Membrane potential dynamics of GABAergic neurons in the barrel cortex of
behaving mice. Neuron 65, 422–435.
[47] Potjans, T. C., & Diesmann, M. (2012). The cell-type specific cortical
microcircuit: Relating structure and activity in a full-scale spiking network model.
Cerebral Cortex , doi: 10.1093/cercor/bhs358.
[48] Tetzlaff, T., Einevoll, G. T., & Diesmann, M. (2009). Synchronization and
rate dynamics in embedded synfire chains: effect of network heterogeneity and
feedback. BMC Neuroscience 10 (Suppl I), P258.
[49] Roxin, A. (2011). The role of degree distribution in shaping the dynamics in
networks of sparsely connected spiking neurons. Front. Comput. Neurosci. 5 (8).
[50] Diesmann, M., Gewaltig, M.-O., & Aertsen, A. (1999). Stable propagation of
synchronous spiking in cortical neural networks. Nature 402 (6761), 529–533.
[51] Oppenheim, A., & Wilsky, A. (1996). Systems and signals. Prentice Hall.
[52] Gewaltig, M.-O., & Diesmann, M. (2007). NEST (NEural Simulation Tool).
Scholarpedia 2 (4), 1430.
[53] Hanuschkin, A., Kunkel, S., Helias, M., Morrison, A., & Diesmann, M. (2010).
A general and efficient method for incorporating precise spike times in globally
time-driven simulations. Front. Neuroinform. 4, 113.
|
1806.07365 | 2 | 1806 | 2018-07-03T14:45:06 | State equation from the spectral structure of human brain activity | [
"q-bio.NC",
"cond-mat.dis-nn"
] | Neural electromagnetic (EM) signals recorded non-invasively from individual human subjects vary in complexity and magnitude. Nonetheless, variation in neural activity has been difficult to quantify and interpret, due to complex, broad-band features in the frequency domain. Studying signals recorded with magnetoencephalography (MEG) from healthy young adult subjects while in resting and active states, a systematic framework inspired by thermodynamics is applied to neural EM signals. Despite considerable inter-subject variation in terms of spectral entropy and energy across time epochs, data support the existence of a robust and linear relationship defining an effective state equation, with higher energy and lower entropy in the resting state compared to active, consistently across subjects. Mechanisms underlying the emergence of relationships between empirically measured effective state functions are further investigated using a model network of coupled oscillators, suggesting an interplay between noise and coupling strength can account for coherent variation of empirically observed quantities. Taken together, the results show macroscopic neural observables follow a robust, non-trivial conservation rule for energy modulation and information generation. | q-bio.NC | q-bio |
State equation from the spectral structure of human brain activity
Trang-Anh Nghiem,1, 2, ∗ Jean-Marc Lina,3, 4, ∗ Matteo di Volo,1, 2 Cristiano
Capone,1, 2 Alan C. Evans,5, 6 Alain Destexhe,1, 2, † and Jennifer S. Goldman1, 2, 3, 5, †
1European Institute for Theoretical Neuroscience,
74 Rue du Faubourg Saint-Antoine, Paris, 75012 France
2Unit of Neuroscience, Information, and Complexity, Centre National de la Recherche Scientifique,
5Ludmer Centre for Neuroinformatics and Mental Health, Montreal Neurological Institute and Hospital,
1 Avenue de la Terrasse, Gif-sur-Yvette 91190 France
3Centre de recherches mathématiques, Université de Montréal, 2920,
Chemin de la tour, bur. 5357 Montréal, Québec, H3T 1J4, Canada
4Département de génie électrique, École de technologie supérieure,
1100 rue Notre-Dame Ouest, Montréal, Québec, H3C 1K3 Canada
3801 Rue University, Montréal, Quebec, H3A 2B4 Canada
6Department of Biomedical Engineering, McGill University,
3775 Rue University, 316, Montréal, Quebec, H3A 2B4 Canada
Neural electromagnetic (EM) signals recorded
non-invasively from individual human subjects
vary in complexity and magnitude. Nonetheless,
variation in neural activity has been difficult to
quantify and interpret, due to complex, broad-
band features in the frequency domain. Study-
ing signals recorded with magnetoencephalogra-
phy (MEG) from healthy young adult subjects
while in resting and active states, a systematic
framework inspired by thermodynamics is ap-
plied to neural EM signals. Despite consider-
able inter-subject variation in terms of spectral
entropy and energy across time epochs, data sup-
port the existence of a robust and linear rela-
tionship defining an effective state equation, with
higher energy and lower entropy in the resting
state compared to active, consistently across sub-
jects. Mechanisms underlying the emergence of
relationships between empirically measured effec-
tive state functions are further investigated us-
ing a model network of coupled oscillators, sug-
gesting an interplay between noise and coupling
strength can account for coherent variation of em-
pirically observed quantities. Taken together, the
results show macroscopic neural observables fol-
low a robust, non-trivial conservation rule for en-
ergy modulation and information generation.
Brain activity varies between states, e.g.
in uncon-
sciousness vs consciousness, but a formal macroscopic
description remains poorly understood. Spectral analy-
sis of neural electromagnetic (EM) signals indicates that
fluctuations occur simultaneously at all measurable fre-
quencies, along a 1/fβ distribution, above which resonant
modes occur at different frequencies [1–3]. While neural
power spectra display broadband features, they are con-
ventionally investigated by ad hoc partition of the fre-
quency domain. Results of band-limited analyses have
demonstrated that the power of bands shifts between
brain states. The shift in power within a frequency band
Lower spectral energy and higher spectral
FIG. 1.
entropy in active versus resting brain state, for one
human subject. (a) Magentoencephalography (MEG) sig-
nals (X) from human subjects in resting and active states are
used to compute (b) power spectra ( X) for each subject (φ),
state (τ), sensor (µ), and epoch (). (c) Spectral energy (Eq.
1) and entropy (Eq. 2) are derived. Panel c shows the energy
and entropy across epochs for one subject in resting and ac-
tive states. Between brain states, a significant shift in energy
and entropy (Mann-Whitney U test, p-values reported) can
be seen in histograms (c) as well as a change in slope and in-
tercept of the negative linear energy-entropy relation (best fit
line shown, Pearson's r reported) as shown in the scatter plot.
The inset labeled δS , δU implies a non-tautological relation
between effective energy and entropy as measured empirically
(see text).
is thought at least in part to reflect changes in the level of
abc010203040rest38.5[fT2/s2]activesignal[fT]timefrequencypowerentropyenergyx10-23******p << 0.001p << 0.001r = -0.89 p << 0.001r = -0.92p << 0.001 synchronous neuronal activity at the time scale of that
band. Indeed, vigilant behavioral states are commonly
called 'desynchronized' whereas unconscious states 'syn-
chronized' [1–3].
Macroscopic, EM brain signals measured outside the
human head reflect the coordinated activity of tens of
thousands of neurons [4]. Analogies to statistical physics
have helped to understand how microscopic properties
of neurons and their interactions lead to neural popula-
tion behavior at the mesoscale [5–12] across brain states
[13, 14]. Since the use of methodology from statistical
physics has uncovered descriptions of individual neurons'
activity and resulting population behaviour at meso-
scopic scales, macroscopically observed changes in signal
complexity and magnitude could be addressed in anal-
ogy to thermodynamics. In fact, for large-scale, whole-
brain recordings, spectral entropy measures reportedly
vary with levels of consciousness [15–18]. In the present
paper, brain states are described as a function of global
state variables, derived from the spectral structure of dif-
ferent human subjects' brain activity.
Neural signals from each human subject in each brain
state may be dissimilar, but the organization of spectra
could follow general principles characterized by a state
equation. To test this possibility, a statistical analysis of
broadband signals is performed using magnetic fields ob-
tained by magnetoencephalography (MEG) from healthy,
adult, human subjects in two brain states - resting state
with eyes open and fixated (rest) and while performing
an N-back visual working memory task (active) [19, 20].
Specifically, the magnetic field X, measured for each
subject φ = {1 . . . 77}, brain state τ = {rest, active},
sensor µ = {1 . . . 248}, and time epoch = {1 . . . M}
where M is the total number of epochs (Fig.1) is ana-
lyzed.
In figure 1a, the time courses of fields for such
Xφ,τ,µ,(t) in rest (τ = r, Fig.1a, top) and active (τ = a,
Fig.1a, bottom) states is shown. The corresponding
power spectra Xφ,τ,µ,(f ) are computed for each epoch
(Fig.1b).
Observables between resting and active states may
present different statistical properties. In order to test
this hypothesis, a spectral energy measure Uφ,τ, is in-
troduced for each subject in each brain state and each
time epoch, summing over sensors :
Xφ,τ,µ,(f )
(1)
Uφ,τ, ≡(cid:88)
µ,f
Similarly, a spectral entropy measure Sφ,τ, is defined,
for each subject in each brain state and epoch, summing
over sensors, as
P (f ) ln P (f ),
(2)
Sφ,τ, ≡ −(cid:88)
µ,f
(cid:80)
Xφ,τ,µ,(f )
Xφ,τ,µ,(f )
f
where, P (f )
:=
describes the relative
2
FIG. 2. Variation of the spectral energy-entropy re-
lation between brain states is robust across subjects.
Mean spectral entropy (a), mean spectral energy (b), slope
(c), and intercept (d) (Eq. 3) are shown in active state against
rest (one point per subject, identity line shown). In the rest-
ing state, spectral energy is consistently larger, and entropy
smaller than in the active state (Wilcoxon signed-rank test).
The energy-entropy relation has a steeper negative slope (c)
and a higher intercept (d) in resting than active state.
heights of power spectral frequency bins, from each sub-
ject φ in each brain state τ.
A negative relation between energy and entropy across
time epochs is found (Fig.1c): when energy decreases,
entropy increases, suggesting a linear relationship of the
following form:
Uφ,τ, = Tφ,τ Sφ,τ, + Wφ,τ
(3)
where Tφ,τ and Wφ,τ are respectively the slope and in-
tercept of the entropy-energy relation across time epochs,
for each subject in each brain state. Tφ,τ and Wφ,τ dif-
fer between brain states; in the resting state, energy is
higher, entropy is lower, and their relationship follows a
steeper (more negative slope) line with a higher-intercept
than in the active state (Fig. 1c).
Such a clear linear relationship between the state vari-
ables Sφ,τ and Uφ,τ raises naturally the question of
whether it could be a direct consequence of how the quan-
tities are defined. This possibility is tested through ma-
nipulation of the signals where the entropy can be mod-
ified (δS) by adding zero-mean noise to the power spec-
trum, with no consequent effect on energy as reported
in the inset of Fig.1c. Conversely, multiplying the full
spectrum by a constant results in a change of the energy
(δU) without altering the entropy. The results support
the hypothesis that the observed linear relationship be-
tween energy and entropy could reflect fundamental rules
that define the way neural activity is organized.
Despite important inter-subject variability in the
SrSaUaUrTaTrWrWaentropyenergyslopeinterceptabcdx 10x 10x 10x 10x 10x 10-23-23-23-23-22-22p << 0.001p << 0.001p << 0.001p << 0.001 means over epochs of Uφ,τ, and Sφ,τ,, as well as Tφ,τ ,
and Wφ,τ , it is verified that the active-rest differences are
strongly consistent across subjects, evidence for robust-
ness of the introduced framework (Fig. 2).
Toward a better understanding of mechanisms govern-
ing fluctuating oscillations in neural assemblies between
brain states, a classical model of coupled oscillators (Ku-
ramoto model) is next studied [21, 22]. N oscillators of
phase θi, i ∈ 1, ..., N are arranged over a fully connected
network, each one evolving according to the following
equation:
θi = ωi + K
sin(θj − θi) + σηi,
(4)
(cid:88)
j(cid:54)=i
where K is the coupling strength. The bare frequencies
ωi are extracted from a bimodal distribution (sum of two
Gaussians). A zero-mean white i.i.d. noise term ηi of
standard deviation σ is received by each oscillator.
From simulated network activity, a collective time-
varying variable R, to compare with Xφ,τ,µ,, is defined:
R = r
where
1 − cos(Θ)
2
(cid:88)
j
eiθj .
reiΘ =
,
(5)
(6)
The parameter r(t) captures the synchronization between
the oscillators - indeed, when the phases are equally dis-
tributed over [0, 2π], r(t) = 0. Conversly, when the sys-
tem is completely synchronous, all oscillators have the
same phase, so that r(t) = 1. The variable Θ(t) repre-
sents the time evolution of the system and describes the
frequency of synchronous events.
To investigate further the behavior of the effective
energy-entropy relation found in human data, the dy-
namics of the Kuramoto model is investigated by chang-
ing the coupling parameter K and the amplitude of noise
σ, which are known to strongly modulate synchrony and
complexity in coupled oscillator models [23]. Varying K,
a transition from asynchronous active-like to synchronous
rest-like dynamical regime (Fig.3a-c) occurs. To mimic
the observed variation between time epochs, the ampli-
tude of noise σ can be understood as varying in time
(Fig.3d-f).
Increasing noise amplitude σ results in less
synchronous activity, consistent with previous reports, as
well as decreased R [23], reduced energy, and increased
entropy (Fig. 3). Altogether, the results indicate that a
change in brain state can be modeled by varying coupling
strength, while variation across time can be simulated by
fluctuations in noise amplitude.
If changes in coupling contribute to empirically ob-
served shifts in neural state variables as predicted by the
model, one may expect to find alterations in the phase
synchrony in the data. In order to test this prediction,
3
FIG. 3. Phase coupling and noise modulate spectral
energy and entropy. A parameter exploration of the Ku-
ramoto model is performed in order to explain variation of
state variables in experiments. Change in coupling strength
K between oscillators causes a shift in spectral state variables.
(a) Phase of each oscillator throughout time. (b) R (Eq. 5).
(c) Power spectra of R. (d-e) To reproduce variation across
epochs, noise amplitude is modulated - the higher the noise,
the larger the spectral entropy and the lower the spectral en-
ergy. (f) Energy and entropy as defined therefore vary along
a negative linear relation, whose slope becomes steeper and
intercept increases with coupling strength.
the Phase-Lag Index (PLI), a measure of synchroniza-
tion across time series is applied to human data [24]. By
definition, the PLI is larger in more phase-locked states.
Indeed, in resting compared to active state data, a sig-
nificant trend toward enhanced phase synchrony is found
(Fig.4).
A potential interpretation of the robust relation be-
tween these effective state variables could reflect the fact
that energy is used to generate information in brain ac-
tivity. It is important to note that the energy quantity
proposed here is an indirect measure of neural activity
and may be non-trivially related to actual internal en-
ergy. We also find evidence for a change in coupling,
due to the consistent change in intercept and slope of the
energy-entropy relation. Coupling regulates the sensitiv-
0.00.51.01.52.02.53.05.05.56.06.57.00.00.10.20.30.40.50.60.70400.00.51.01.52.02.53.05.05.56.06.57.00.00.10.20.30.40.50.6123456Time (s)Time (s)Frequency (rad/s)Frequency (rad/s)PowerPowerPopulation activity ROscillatorsTime (s)Phase5.07.0Time (s)5.07.00.00.20.40.000.100.200.30K = 2K = 8CouplingstrengthCouplingstrengthOscillators8004080white noise amplitude1.21.41.61.82.02.22.42.620406080100120K = 2K = 2.5K = 3K = 3.5K = 4K = 4.5K = 5K = 5.5K = 6K = 6.5K = 7K = 7.5K = 8entropyenergyenergyentropy0.250.500.751.00204060801001200.250.500.751.001.21.41.61.82.02.22.42.62.8abcdef 4
with global electromagnetic measurements, it would pave
the way to more formal, scale-integrated descriptions of
brain activity in healthy and pathological states.
ACKNOWLEDGEMENTS
Data were provided by the Human Connectome
Project, WU-Minn Consortium (Principal Investigators:
David Van Essen and Kamil Ugurbil; 1U54MH091657)
funded by the 16 NIH Institutes and Centers that support
the NIH Blueprint for Neuroscience Research; and by the
McDonnell Center for Systems Neuroscience at Wash-
ington University. This project was supported by the
European Union (Human Brain Project H2020-720270
and H2020-785907). JSG was supported by the Lud-
mer Centre for Neuroinformatics and Mental Health, the
Molson Neuroengineering Award, The NSERC-CREATE
Training Program in Neuroengineering, and the Ann
and Richard Sievers Foundation for Innovation in Neu-
roscience. JML was supported by an NSERC Discovery
grant. The authors thank Gleb Bezgin, Bruno Cessac,
Alexis Garcia, Peter Grutter, George Kostopoulos, Naj
Mahani, Danko Nikolic, Bartosz Teleńczuk, and Núria
Tort-Colet for useful discussion.
AUTHOR CONTRIBUTIONS
T.-A. N. wrote data analysis codes, investigated the
Kuramoto model, contributed to interpreting the results,
and writing the manuscript. J.-M. L. contributed to writ-
ing the analysis codes and the manuscript. M. d. V.
contributed to interpreting the results, investigating the
Kuramoto model, and writing the manuscript. C. C.
investigated the Kuramoto model, contributed to inter-
preting the results and editing the manuscript. A. C. E.
provided resources and supervision of the work. A. D.
designed research, provided supervision, and edited the
manuscript. J. S. G. conceived of the question, designed
research, analyzed data, interpreted results, and wrote
the paper.
COMPETING INTERESTS
The authors declare no competing interests.
REFERENCES
∗ These authors contributed equally.
FIG. 4. Phase Lag Index (PLI) is larger in resting
than active state, consistent with stronger coupling
during rest. The PLI is computed across frequency bins
in each epoch in rest and task for each subject. (a-b) The
mean PLI across epochs for each subject and state is reported,
showing the PLI is significantly larger during the resting than
active state (paired Student's T-test). (c) Summary diagram
of the results. The data support a model in which variation
of effective energy and entropy between brain states relies on
noise fluctuations and a shift in coupling strength.
ity of networks to noise, which can account for the slope
of the aforementioned relation. Indeed, with strong cou-
pling, perturbations propagate more efficiently through-
out the network, causing changes in the collective ac-
tivity and affecting the resulting energy. Conversely, in
a weakly-coupled network, perturbations do not easily
spread across the system. In other words, the 'gas-like'
regime found in active brains may be useful for the poten-
tial to generate information versus the 'liquid-like' state
in which strong coupling promotes a low entropy state
with less informational capacity. This provides a new in-
terpretation of previously reported changes in functional
connectivity across brain states [25–29].
Lastly, in light of statistical physics inspired models
of neuronal activity [5–14], our results suggest a strat-
egy to bridge the gap from macroscopic system signals
studied here to microscopic cellular level activity.
In-
deed, biophysical models of spiking neural networks could
be used to investigate how the mechanisms uncovered
with the Kuramoto model translate at the microscopic
scale. For instance, biophysical processes that control
the variability of noisy inputs or the network's tendency
to synchronize could be studied. This may also provide
insight on how the spectral state variables defined here
relate to entropy and energy of micro-states in the sense
of statistical physics and spin glass models. While such
an approach demands scale-integrated data from a suf-
ficiently large population of single neurons simultaneous
0.0390.0400.041Phase Lag Index, rest0.0380.0390.0400.041Phase Lag Index, active0.0380.0390.0400.0410.00.10.20.3restactivePhase Lag Indexabcrestactivep << 0.001 5
1 (2008).
[27] M. E. Raichle, A. M. MacLeod, A. Z. Snyder, W. J. Pow-
ers, D. A. Gusnard, and G. L. Shulman, Proceedings of
the National Academy of Sciences 98, 676 (2001).
[28] M. Massimini, F. Ferrarelli, R. Huber, S. K. Esser,
H. Singh, and G. Tononi, Science 309, 2228 (2005).
[29] R. G. Bettinardi, N. Tort-Colet, M. Ruiz-Mejias, M. V.
and G. Deco, Neuroimage 114, 185
Sanchez-Vives,
(2015).
† These authors contributed equally.
[1] P. L. Nunez and R. Srinivasan, Electric fields of the brain:
the neurophysics of EEG (Oxford University Press, USA,
2006).
[2] E. Niedermeyer and F. L. da Silva, Electroencephalogra-
phy: basic principles, clinical applications, and related
fields (Lippincott Williams & Wilkins, 2005).
[3] G. Buzsaki, Rhythms of the Brain (Oxford University
Press, 2006).
[4] S. Baillet, Nature Neuroscience 20, 327 (2017).
[5] E. Schneidman, M. J. Berry, R. Segev, and W. Bialek,
Nature 440, 1007 (2006).
[6] S. Cocco, S. Leibler, and R. Monasson, Proceedings of
the National Academy of Sciences 106, 14058 (2009).
[7] O. Marre, S. El Boustani, Y. Frégnac, and A. Destexhe,
Physical review letters 102, 138101 (2009).
[8] C. Capone, C. Filosa, G. Gigante, F. Ricci-Tersenghi,
and P. Del Giudice, PloS one 10, e0118412 (2015).
[9] U. Ferrari, Physical Review E 94, 023301 (2016).
[10] M. Okun, N. A. Steinmetz, L. Cossell, M. F. Iacaruso,
H. Ko, P. Barthó, T. Moore, S. B. Hofer, T. D. Mrsic-
Flogel, M. Carandini, et al., Nature 521, 511 (2015).
[11] C. Gardella, O. Marre, and T. Mora, eneuro 3, ENEURO
(2016).
[12] C. Capone, G. Gigante,
and P. Del Giudice, arXiv
preprint arXiv:1804.00153 (2018).
[13] G. Tavoni, U. Ferrari, F. P. Battaglia, S. Cocco, and
R. Monasson, Network Neuroscience 1, 275 (2017).
[14] T.-A. Nghiem, B. Telenczuk, O. Marre, A. Destexhe, and
U. Ferrari, arXiv preprint arXiv:1801.01853 (2018).
[15] J. D. Sitt, J.-R. King,
I. El Karoui, B. Rohaut,
F. Faugeras, A. Gramfort, L. Cohen, M. Sigman, S. De-
haene, and L. Naccache, Brain 137, 2258 (2014).
[16] S. Sarasso, M. Boly, M. Napolitani, O. Gosseries,
V. Charland-Verville, S. Casarotto, M. Rosanova, A. G.
Casali, J.-F. Brichant, P. Boveroux, et al., Current Biol-
ogy 25, 3099 (2015).
[17] I. D. McKay, L. J. Voss, J. W. Sleigh, J. P. Barnard, and
E. K. Johannsen, Anesthesia & Analgesia 102, 91 (2006).
[18] A. L. Vanluchene, H. Vereecke, O. Thas, E. P. Mortier,
and M. M. Struys, The Journal of the
S. L. Shafer,
American Society of Anesthesiologists 101, 34 (2004).
[19] L. J. Larson-Prior, R. Oostenveld, S. Della Penna,
G. Michalareas, F. Prior, A. Babajani-Feremi, J.-M.
Schoffelen, L. Marzetti, F. de Pasquale, F. Di Pompeo,
et al., Neuroimage 80, 190 (2013).
[20] D. C. Van Essen, K. Ugurbil, E. Auerbach, D. Barch,
T. Behrens, R. Bucholz, A. Chang, L. Chen, M. Corbetta,
S. W. Curtiss, et al., Neuroimage 62, 2222 (2012).
[21] Y. Kuramoto, in International symposium on mathemat-
ical problems in theoretical physics (Springer, 1975) pp.
420–422.
[22] M. Breakspear, S. Heitmann, and A. Daffertshofer, Fron-
tiers in human neuroscience 4, 190 (2010).
[23] A. Pikovsky, M. Rosenblum, J. Kurths, and J. Kurths,
Synchronization: a universal concept in nonlinear sci-
ences, Vol. 12 (Cambridge university press, 2003).
[24] S. Silva Pereira, R. Hindriks, S. Mühlberg, E. Maris,
F. van Ede, A. Griffa, P. Hagmann, and G. Deco, Brain
Connectivity 7, 541 (2017).
[25] B. Biswal, F. Zerrin Yetkin, V. M. Haughton, and J. S.
Hyde, Magnetic resonance in medicine 34, 537 (1995).
[26] R. L. Buckner, J. R. Andrews-Hanna, and D. L. Schac-
ter, Annals of the New York Academy of Sciences 1124,
SUPPLEMENTARY INFORMATION
Data and preprocessing
Once again, summing over all bins and all sensors, the
entropy can be derived following Eq. 2.
6
Kuramoto model
(cid:110)
(cid:104)− (ω − ω0)2
(cid:105)
(cid:104)− (ω − ω1)2
(cid:105)(cid:111)
2∆2
In all our simulations, we consider an all-to-all network
of N = 100 oscillators. The bare frequencies of each oscil-
lator ωi are extracted from a distribution of the following
form:
1√
8π∆
exp
2∆2
+exp
p(ω) =
,
(7)
where ∆ = 1 Hz, ω0 = 5 Hz, and ω1 = 10 Hz. We
verified that the general picture reported in the main
text is not affected by the specific choice of these three
parameters. The simulations were run for 15 minutes
of simulated time (to match with the experimental data
and to ensure a robust estimation of observables), using
an Euler integration method with a time step of 5 ms.
The energy and entropy are computed on power spectra
estimated on windows of 2 s, as with the experiments. A
transient time of 100 s was discarded in all simulations.
Phase Lag Index (PLI)
The PLI was computed for each subject and time win-
dow, summing over all sensors. The frequency range of
interest [4,100] Hz was divided into ten equally wide bins.
Within each frequency bin, a time series is obtained by
filtering the signal with a Butterworth band-pass filter.
The Hilbert transform is then employed to extract the
phase ψk of the time series in each frequency bin k. From
there, the PLI, given by
PLI ≡ < sign(ψk) >,
(8)
is computed, where < · > denotes averaging over time
[24]. One may note that the PLI takes values between 0
(random phase relations) and 1 (perfect phase locking).
In this work we report the mean PLI over all time epochs
for each subject in each brain state.
Magnetoencephalography (MEG) data were acquired,
preprocessed, anonymized, and distributed by the Hu-
man Connectome Project (HCP). Healthy adult human
subjects were imaged in a magnetically shielded room
using a whole head MAGNES 3600 (4D Neuroimaging,
San Diego, CA) system at the Saint Louis University
(SLU) medical campus. HCP protocols collected rest-
ing state (rest) and visual N-back working memory (ac-
tive) data from the same subjects. The MEG has 248
magnetometer channels. Electrooculography, electrocar-
diography, and electromyography recordings are all syn-
chronized with the MEG. Data were acquired with a
bandwidth of 400Hz, DC high pass filter,
in a con-
tinuous acquisition mode, with a 2034.5 Hz sampling
rate. MEG data were preprocessed by the HCP. Bio-
logical and environmental artifacts, including 60 Hz ar-
tifacts, bad channels, and bad segments were removed
first automatically by an independent component anal-
ysis (ICA)-based, publically available, custom Fieldtrip-
scripted pipeline and further quality controlled by hu-
man verification. Channel-level preprocessing pipelines
also downsampled data to 508.63Hz. N-back data were
split into groups of trials corresponding to task design,
and divided into epochs time-locked to the onset of the
stimulus (TIM) to yield Matlab structures containing
epoched, preprocessed data. Further documentation de-
scribing data collection and preprocessing is available [19]
and www.humanconnectome.org/documentation.
Power spectrum
Subjects with data in both N-back memory and resting
states (N = 77) were included in the analyses. Per sub-
ject average recording time is: 14.06±0.55 min for resting
state and 12.35±0.67 min for working memory. Respect-
ing constraints introduced by epoch time and Nyquist ar-
tifacts, power spectra between [4,100] Hz were obtained
with Welch windowing method per Xφ,τ,µ,, indexing 2
sec MEG time series from each subject φ, brain state τ
(active or rest), sensor µ, and epoch .
Spectral energy and entropy
Each power spectrum Xφ,τ,µ,(f ) was used to construct
a histogram with bin size 0.1 Hz. Summing over the his-
togram bins of the power spectra of all sensors, the en-
ergy Uφ,τ, is obtained (Eq. 1). By normalizing each
histogram, one obtains the frequency probability distri-
bution P (f ) for each sensor, epoch, state, and subject.
|
1312.2950 | 1 | 1312 | 2013-12-10T13:33:18 | Revisiting the two-mass model of the vocal folds | [
"q-bio.NC"
] | Realistic mathematical modeling of voice production has been recently boosted by applications to different fields like bioprosthetics, quality speech synthesis and pathological diagnosis. In this work, we revisit a two-mass model of the vocal folds that includes accurate fluid mechanics for the air passage through the folds and nonlinear properties of the tissue. We present the bifurcation diagram for such a system, focusing on the dynamical properties of two regimes of interest: the onset of oscillations and the normal phonation regime. We also show theoretical support to the nonlinear nature of the elastic properties of the folds tissue by comparing theoretical isofrequency curves with reported experimental data. | q-bio.NC | q-bio |
Papers in Physics, vol. 5, art. 050004 (2013)
Received: 2 March 2013, Accepted: 5 June 2013
Edited by: G. Mindlin
Licence: Creative Commons Attribution 3.0
DOI: http://dx.doi.org/10.4279/PIP.050004
www.papersinphysics.org
ISSN 1852-4249
Revisiting the two-mass model of the vocal folds
M. F. Assaneo,1∗ M. A. Trevisan1†
Realistic mathematical modeling of voice production has been recently boosted by ap-
plications to different fields like bioprosthetics, quality speech synthesis and pathological
diagnosis. In this work, we revisit a two-mass model of the vocal folds that includes accu-
rate fluid mechanics for the air passage through the folds and nonlinear properties of the
tissue. We present the bifurcation diagram for such a system, focusing on the dynamical
properties of two regimes of interest: the onset of oscillations and the normal phonation
regime. We also show theoretical support to the nonlinear nature of the elastic properties
of the folds tissue by comparing theoretical isofrequency curves with reported experimental
data.
I.
Introduction
In the last decades, a lot of effort was devoted to de-
velop a mathematical model for voice production.
The first steps were made by Ishizaka and Flana-
gan [1], approximating each vocal fold by two cou-
pled oscillators, which provide the basis of the well
known two-mass model. This simple model repro-
duces many essential features of the voice produc-
tion, like the onset of self sustained oscillation of
the folds and the shape of the glottal pulses.
Early analytical treatments were restricted to
small amplitude oscillations, allowing a dimen-
sional reduction of the problem.
In particular, a
two dimensional approximation known as the flap-
ping model was widely adopted by the scientific
community, based on the assumption of a transver-
sal wave propagating along the vocal folds [2, 3].
Moreover, this model was also used to successfully
∗E-mail: [email protected]
†E-mail: [email protected]
1 Laboratorio de Sistemas Din´amicos, Depto. de F´ısica,
FCEN, Universidad de Buenos Aires. Pabell´on I, Ciudad
Universitaria, 1428EGA Buenos Aires, Argentina.
explain most of the features present in birdsong
[4, 5].
Faithful modeling of the vocal folds has recently
found new challenges: realistic articulatory speech
synthesis [6 -- 8], diagnosis of pathological behavior
of the folds [9, 10] and bioprosthetic applications
[11]. Within this framework, the 4-dimensional
two-mass model was revisited and modified. Two
main improvements are worth noting: a realistic
description of the vocal fold collision [13,14] and an
accurate fluid mechanical description of the glottal
flow, allowing a proper treatment of the hydrody-
namical force acting on the folds [8, 15].
In this work, we revisit the two-mass model de-
veloped by Lucero and Koenig [7]. This choice rep-
resents a good compromise between mathematical
simplicity and diversity of physical phenomena act-
ing on the vocal folds, including the main mechani-
cal and fluid effects that are partially found in other
models [13, 15].
It was also successfully used to
reproduce experimental temporal patterns of glot-
tal airflow. Here, we extend the analytical study
of this system: we present a bifurcation diagram,
explore the dynamical aspects of the oscillations
at the onset and normal phonation and study the
050004-1
Papers in Physics, vol. 5, art. 050004 (2013) / M. F. Assaneo et al.
isofrequency curves of the model.
This work is organized as follows: in the second
section, we describe the model.
In the third sec-
tion, we present the bifurcation diagram, compare
our solutions with those of the flapping model ap-
proximation and analyze the isofrecuency curves.
In the fourth and last section, we discuss our re-
sults.
force, m is the mass and kc the coupling stiffness.
The horizontal displacement from the rest position
x0 is represented by x.
We use a cubic polynomial for the restitution
term [Eq. (2)], adapted from [1, 7]. The term with
a derivable step-like function Θ [Eq. (3)] accounts
for the increase in the stiffness introduced by the
collision of the folds. The restitution force reads
II. The model
Each vocal fold is modeled as two coupled damped
oscillators, as sketched in Fig. 1.
Ki(xi) = kixi(1 + 100xi
2)
(2)
+ Θ(cid:18) xi + x0
x0 (cid:19) 3ki(xi + x0)[1 + 500(xi + x0)2],
with
Θ(x) =(cid:26)
0 if x ≤ 0
if x > 0
x2
8 10−4+x2
,
(3)
where x0 is the rest position of the folds.
For the damping force, we have adapted the ex-
pression proposed in [7], making it derivable, arriv-
ing at the following equation:
Bi(xi) =
(cid:20)1 + Θ(cid:18) xi + x0
x0 (cid:19) 1
ǫi(cid:21) ri(1 + 850xi
(4)
2)yi,
where ri = 2ǫi√kimi, and ǫi is the damping ratio.
In order to describe the hydrodynamic force that
the airflow exerts on the vocal folds, we have
adopted the standard assumption of small inertia of
the glottal air column and the model of the bound-
ary layer developed in [7, 11, 15]. This model as-
sumes a one-dimensional, quasi-steady incompress-
ible airflow from the trachea to a separation point.
At this point, the flow separates from the tissue
surface to form a free jet where the turbulence dis-
sipates the airflow energy. It has been experimen-
tally shown that the position of this point depends
on the glottal profile. As described in [15], the
separation point located at the glottal exit shifts
down to the boundary between masses m1 and m2
when the folds profile becomes more divergent than
a threshold [Eq. (7)].
Figure 1: Sketch of the two-mass model of the vo-
cal folds. Each fold is represented by masses m1
and m2 coupled to each other by a restitution force
kc and to the laryngeal walls by K1 and K2 (and
dampings B1 and B2), respectively. The displace-
ment of each mass from the resting position x0 is
represented by x1 and x2. The different aerody-
namic pressures P acting on the folds are described
in the text.
Assuming symmetry with respect to the saggital
plane, the left and right mass systems are identical
(Fig. 1) and the equation of motion for each mass
reads
xi = yi
1
mi
yi =
[fi − Ki(xi) − Bi(xi, yi) − kc(xi − xj)] ,
(1)
for i, j = 1 or 2 for lower and upper masses, re-
spectively. K and B represent the restitution and
damping of the folds tissue, f the hydrodynamic
Viscous losses are modeled according to a bi-
dimensional Poiseuille flow [Eqs. (6) and (7)]. The
equations for the pressure inside the glottis are
050004-2
Papers in Physics, vol. 5, art. 050004 (2013) / M. F. Assaneo et al.
,
ρu2
g
2a2
1
12µugd1l2
g
a3
1
g
Pin = Ps +
P12 = Pin −
P21 =( 12µug d2l2
a3
2
Pout = 0.
,
(5)
(6)
+ Pout
if a2 > ksa1
0 if a2 ≤ ksa1
,
(7)
(8)
As sketched in Fig. 1, the pressures exerted by
the airflow are: Pin at the entrance of the glottis,
P12 at the upper edge of m1, P21 at the lower edge
of m2, Pout at the entrance of the vocal tract and
Ps the subglottal pressure.
The width of the folds (in the plane normal to
Fig. 1) is lg; d1 and d2 are the lengths of the lower
and upper masses, respectively. ai are the cross-
sections of the glottis, ai = 2lg(xi + x0); µ and ρ
are the viscosity and density coefficient of the air;
ug is the airflow inside the glottis, and ks = 1.2 is an
experimental coefficient. We also assume no losses
at the glottal entrance [Eq. (5)], and zero pressure
at the entrance of the vocal tract [Eq. (8)].
The hydrodynamic force acting on each mass
reads:
2
Pin +P12
f1 =(cid:26) d1lgPs
f2 =
d2lgPs
0
P21+Pout
2
if x1 ≤ −x0 or x2 ≤ −x0
in other case
(9)
if x1 > −x0 and x2 ≤ −x0
if x1 ≤ −x0
in other case
(10)
Following [1, 7, 10], these functions represent
opening, partial closure and total closure of the
glottis. Throughout this work, piecewise functions
P21, f1 and f2 are modeled using the derivable step-
like function Θ defined in Eq. (3).
III. Analysis of the model
i. Bifurcation diagram
The main anatomical parameters that can be ac-
tively controlled during the vocalizations are the
subglottal pressure Ps and the folds tension con-
trolled by the laryngeal muscles. In particular, the
action of the thyroarytenoid and the cricothyroid
muscles control the thickness and the stiffness of
folds. Following [1], this effect is modeled by a pa-
rameter Q that scales the mechanic properties of
the folds by a cord-tension parameter: kc = Qkc0,
ki = Qki0 and mi = mi0
Q . We therefore performed a
bifurcation diagram using these two standard con-
trol parameters Ps and Q.
Five main regions of different dynamic solutions
are shown in Fig. 2. At low pressure values (re-
gion I), the system presents a stable fixed point.
Reaching region II, the fixed point becomes un-
stable and there appears an attracting limit cycle.
At the interface between regions I and II, three
bifurcations occur in a narrow range of subglot-
tal pressure (Fig. 3, left panel), all along the Q
axis. The right panel of Fig. 3 shows the oscilla-
tion amplitude of x2. At point A, oscillations are
born in a supercritical Hopf bifurcation. The am-
plitude grows continuously for increasing Ps until
point B, where it jumps to the upper branch. If the
pressure is then decreased, the oscillations persist
even for lower pressure values than the onset in A.
When point C is reached, the oscillations suddenly
stop and the system returns to the rest position.
This onset-offset oscillation hysteresis was already
reported experimentally in [12].
The branch AB depends on the viscosity. De-
creasing µ, points A and B approach to each other
until they collide at µ = 0, recovering the result re-
ported in [3, 10, 14], where the oscillations occur as
the combination of a subcritical Hopf bifurcation
and a cyclic fold bifurcation.
On the other hand, the branch BC depends on
the separation point of the jet formation. In par-
ticular, for increasing ks, the folds become stiffer
and the separation point moves upwards toward the
output of the glottis. From a dynamical point of
view, points C and B approach to each other until
they collapse. In this case, the oscillations are born
at a supercritical Hopf bifurcation and the system
presents no hysteresis, as in the standard flapping
model [17].
Regions II and III of Fig. 2 are separated by
a saddle-repulsor bifurcation. Although this bifur-
cation does not represent a qualitative dynamical
change for the oscillating folds, its effects are rele-
vant when the complete mechanism of voiced sound
050004-3
Papers in Physics, vol. 5, art. 050004 (2013) / M. F. Assaneo et al.
Figure 3: Hysteresis at the oscillation onset-offset.
Left panel: zoom of the interface between regions I
and II. The blue and green lines represent folds of
cycles (saddle-node bifurcations in the map). The
red line is a supercritical Hopf bifurcation. Right
panel: the oscillation amplitude of x2 as a function
of the subglottal pressure Ps, at Q = 1.71. The
continuation of periodic solutions was realized with
the AUTO software package [20].
Figure 2: Bifurcation diagram in the plane of sub-
glottal pressure and fold tension (Q,Ps). The in-
sets are two-dimensional projections of the flow on
the (v1,x1) plane, the red crosses represent unsta-
ble fixed points and the dotted lines unstable limit
cycles. Normal voice occurs at (Q, Ps) ∼ (1, 800).
The color code represents the linear correlation be-
tween (x1 − x2) and (y1 + y2): from dark red for
R = 1 to dark blue for R = 0.6. This diagram
was developed with the help of AUTO continuation
software [20]. The rest of the parameters were fixed
at m1 = 0.125 g, m2 = 0.025 g, k10 = 80 N/m,
k20 = 8 N/m, kc = 25 N/m, ǫ1 = 0.1, ǫ2 = 0.6,
lg = 1.4 cm, d1 = 0.25 cm, d2 = 0.05 cm and
x0 = 0.02 cm.
production is considered. Voiced sounds are gener-
ated as the airflow disturbance produced by the
oscillation of the vocal folds is injected into the se-
ries of cavities extending from the laryngeal exit to
the mouth, a non-uniform tube known as the vocal
tract. The disturbance travels back and forth along
the vocal tract, that acts as a filter for the origi-
nal signal, enhancing the frequencies of the source
that fall near the vocal tract resonances. Voiced
sounds are in fact perceived and classified accord-
ing to these resonances, as in the case of vowels [18].
Consequently, one central aspect in the generation
of voiced sounds is the production of a spectrally
rich signal at the sound source level.
Interestingly, normal phonation occurs in the re-
gion near the appearance of the saddle-repulsor bi-
furcation. Although this bifurcation does not al-
ter the dynamical regime of the system or its time
scales, we have observed that part of the limit cy-
cle approaches the stable manifold of the new fixed
point (as displayed in Fig. 4), therefore changing
its shape. This deformation is not restricted to the
appearance of the new fixed point but rather oc-
curs in a coarse region around the boundary be-
tween II and III, as the flux changes smoothly in
a vicinity of the bifurcation. In order to illustrate
this effect, we use the spectral content index SCI
[21], an indicator of the spectral richness of a sig-
nal: SCI = Pk Akfk/(Pk Akf0), where Ak is the
Fourier amplitude of the frequency fk and f0 is
the fundamental frequency. As the pressure is in-
creased, the SCI of x1(t) increases (upper right
panel of Fig. 4), observing a boost in the vicinity of
the saddle-repulsor bifurcation that stabilizes after
the saddle point is generated.
Thus, the appearance of this bifurcation near the
region of normal phonation could indicate a possi-
ble mechanism to further enhance the spectral rich-
ness of the sound source, on which the production
of voiced sounds ultimately relies.
050004-4
Papers in Physics, vol. 5, art. 050004 (2013) / M. F. Assaneo et al.
rium x0, and τ is the time that the surface wave
takes to travel half the way from bottom to top.
Equation (11) can be rewritten as (x1 − x2) =
τ (y1 + y2). We use this expression to quantify the
difference between the oscillations obtained with
the two-mass model solutions and the ones gener-
ated with the flapping approximation, computing
the linear correlation coefficient between (x1 − x2)
and (y1 + y2). As expected, the correlation coeffi-
cient R decreases for increasing Ps or decreasing Q.
In the region near normal phonation, the approx-
imation is still relatively good, with R ∼ 0.8. As
expected, the approximation is better for increasing
x0, since the effect of colliding folds is not included
in the flapping model.
ii.
Isofrequency curves
One basic perceptual property of the voice is the
pitch, identified with the fundamental frequency f0
of the vocal folds oscillation. The production of
different pitch contours is central to language, as
they affect the semantic content of speech, carry-
ing accent and intonation information. Although
experimental data on pitch control is scarce, it was
reported that it is actively controlled by the laryn-
geal muscles and the subglottal pressure. In par-
ticular, when the vocalis or interarytenoid muscle
activity is inactive, a raise of the subglottal pres-
sure produces an upraising of the pitch [16].
Compatible with these experimental results, we
performed a theoretical analysis using Ps as a sin-
gle control parameter for pitch. In the upper pan-
els of Fig. 5, we show isofrequency curves in the
range of normal speech for our model of Eqs. (1)
to (10). Following the ideas developed in [22] for
the avian case, we compare the behavior of the
fundamental frequency with respect to pressure Ps
in the two most usual cases presented in the lit-
the cubic [1, 7] and the linear [10, 14]
erature:
restitutions.
In the lower panels of Fig. 5, we
show the isofrequency curves that result from re-
placing the cubic restitution by a linear restitution
Ki(xi) = kixi + Θ( xi+x0
)3ki(xi + x0).
x0
Although the curves f0(Ps) are not affected by
the type of restitution at the very beginning of os-
cillations, the changes become evident for higher
values of Ps, with positive slopes for the cubic case
and negative for the linear case. This result sug-
gests that a nonlinear cubic restitution force is a
Figure 4: A projection of the limit cycle for x1 and
the stable manifold of the saddle point, for param-
eters consistent with normal phonatory conditions,
(Q, Ps) = (1, 850) (region III). Left inset: projec-
tion in the 3-dimensional space (y1, x1, x2). Right
inset: Spectral content index of x1(t) as a function
of Ps for a fixed value of Q = 0.95. In green, the
value at which the saddle-repulsor bifurcation takes
place.
In the boundary between regions III and IV, one
of the unstable points created in the saddle-repulsor
bifurcation undergoes a subcritical Hopf bifurca-
tion, changing stability as an unstable limit cycle
is created [19]. Finally, entering region V, the sta-
ble and the unstable cycles collide and disappear in
a fold of cycles where no oscillatory regimes exist.
In Fig. 2, we also display a color map that quan-
tifies the difference between the solutions of the
model and the flapping approximation. The flap-
ping model is a two dimensional model that, instead
of two masses per fold, assumes a wave propagating
along a linear profile of the folds, i.e., the displace-
ment of the upper edge of the folds is delayed 2τ
with respect to the lower. The cross sectional areas
at glottal entry and exit (a1 and a2) are approxi-
mated, in terms of the position of the midpoint of
the folds, by
(cid:26) a1 = 2lg(x0 + x + τ x)
a2 = 2lg(x0 + x − τ x)
,
(11)
where x is the midpoint displacement from equilib-
050004-5
Papers in Physics, vol. 5, art. 050004 (2013) / M. F. Assaneo et al.
respect to the oscillation onset, we showed how jets
and viscous losses intervene in the hysteresis phe-
nomenon.
Many different models for the restitution prop-
erties of the tissue have been used across the liter-
ature, including linear and cubic functional forms.
Yet, its specific role was not reported. Here we
showed that the experimental relationship between
subglottal pressure and pitch is fulfilled by a cubic
term.
Acknowledgements - This work was partially
funded by UBA and CONICET.
[1] K Ishizaka, J L Flanagan, Synthesis of voiced
sounds from a two-mass model of the vocal
cords, Bell Syst. Tech. J. 51, 1233 (1972).
[2] I R Titze, The physics of smallamplitude os-
cillation of the vocal folds, J. Acoust. Soc. Am.
83, 1536 (1988).
[3] M A Trevisan, M C Eguia, G Mindlin, Nonlin-
ear aspects of analysis and synthesis of speech
time series data, Phys. Rev. E 63, 026216
(2001).
[4] Y S Perl, E M Arneodo, A Amador, F Goller,
G B Mindlin, Reconstruction of physiological
instructions from Zebra finch song, Phys. Rev.
E 84, 051909 (2011).
[5] E M Arneodo, Y S Perl, F Goller, G B Mindlin,
Prosthetic avian vocal organ controlled by a
freely behaving bird based on a low dimensional
model of the biomechanical periphery, PLoS
Comput. Biol. 8, e1002546 (2012).
[6] B H Story, I R Titze Voice simulation with a
bodycover model of the vocal folds, J. Acoust.
Soc. Am. 97, 1249 (1995).
[7] J C Lucero, L Koening Simulations of tempo-
ral patterns of oral airflow in men and women
using a two-mass model of the vocal folds un-
der dynamic control, J. Acoust. Soc. Am. 117,
1362 (2005).
[8] X Pelorson, X Vescovi, C Castelli, E
Hirschberg, A Wijnands, A P J Bailliet, H M
Figure 5: Relationship between pitch and restitu-
tion forces. Left panels:
isofrequency curves in
the plane (Q,Ps). Right panels: Curves f0(Ps)
for Q=0.9, Q=0.925 and Q=0.95.
In the upper
panels, we used the model with the cubic nonlin-
ear restitution of Eq. (2). In the lower panels, we
show the curves obtained with a linear restitution,
Ki(xi) = kixi + Θ( xi+x0
)3ki(xi + x0).
x0
good model for the elastic properties of the oscil-
lating tissue.
IV. Conclusions
In this paper, we have analyzed a complete two-
mass model of the vocal folds integrating collisions,
nonlinear restitution and dissipative forces for the
tissue and jets and viscous losses of the air-stream.
In a framework of growing interest for detailed
modeling of voice production, the aspects studied
here contribute to understanding the role of the
different physical terms in different dynamical be-
haviors.
We calculated the bifurcation diagram, focusing
in two regimes: the oscillation onset and normal
phonation. Near the parameters of normal phona-
tion, a saddle repulsor bifurcation takes place that
modifies the shape of the limit cycle, contributing
to the spectral richness of the glottal flow, which is
central to the production of voiced sounds. With
050004-6
Papers in Physics, vol. 5, art. 050004 (2013) / M. F. Assaneo et al.
A Hirschberg, Description of the flow through
in-vitro models of the glottis during phonation.
Application to voiced sounds synthesis, Acta
Acust. 82, 358 (1996).
[9] M E Smith, G S Berke, B R Gerratt, Laryn-
geal paralyses: Theoretical considerations and
effects on laryngeal vibration, J. Speech Hear.
Res. 35, 545 (1992).
[10] I Steinecke, H Herzel Bifurcations in an asym-
metric vocalfold model, J. Acoust. Soc. Am.
97, 1874 (1995).
[11] N J C Lous, G C J Hofmans, R N J Veldhuis, A
Hirschberg, A symmetrical two-mass vocal-fold
model coupled to vocal tract and trachea, with
application to prosthesis design, Acta Acust.
United Ac. 84, 1135 (1998).
[12] T Baer, Vocal fold physiology, University of
Tokyo Press, Tokyo, (1981).
[13] T Ikeda, Y Matsuzak, T Aomatsu, A nu-
merical analysis of phonation using a two-
dimensional flexible channel model of the vocal
folds, J. Biomech. Eng. 123, 571 (2001).
[14] J C Lucero, Dynamics of the two-mass model
of the vocal folds: Equilibria, bifurcations, and
oscillation region, J. Acoust. Soc. Am. 94,
3104 (1993).
[15] X Pelorson, A Hirschberg, R R van Hassel, A
P J Wijnands, Y Auregan, Theoretical and ex-
perimental study of quasisteadyflow separation
within the glottis during phonation. Applica-
tion to a modified twomass model, J. Acoust.
Soc. Am. 96, 3416 (1994).
[16] T Baer, Reflex activation of laryngeal muscles
by sudden induced subglottal pressure changes,
J. Acoust. Soc. Am. 65, 1271 (1979).
[17] J C Lucero, A theoretical study of the hystere-
sis phenomenon at vocal fold oscillation onset-
offset, J. Acoust. Soc. Am. 105, 423 (1999).
[18] I Titze, Principles of voice production, Pren-
tice Hall, (1994).
[19] J Guckenheimer, P Holmes, Nonlinear oscil-
lations, dynamical systems and bifurcations of
vector fields, Springer, (1983).
[20] E Doedel, AUTO: Software for continuation
and bifurcation problems in ordinary differen-
tial equations, AUTO User Manual, (1986).
[21] J Sitt, A Amador, F Goller, G B Mindin, Dy-
namical origin of spectrally rich vocalizations
in birdsong, Phys. Rev. E 78, 011905 (2008).
[22] A Amador, F Goller, G B Mindlin, Frequency
modulation during song in a suboscine does
not require vocal muscles, J. Neurophysiol. 99,
2383 (2008).
050004-7
|
1106.5678 | 1 | 1106 | 2011-06-28T14:16:30 | The cell-type specific connectivity of the local cortical network explains prominent features of neuronal activity | [
"q-bio.NC"
] | In the past decade, the cell-type specific connectivity and activity of local cortical networks have been characterized experimentally to some detail. In parallel, modeling has been established as a tool to relate network structure to activity dynamics. While the available connectivity maps have been used in various computational studies, prominent features of the simulated activity such as the spontaneous firing rates do not match the experimental findings. Here, we show that the inconsistency arises from the incompleteness of the connectivity maps. Our comparison of the most comprehensive maps (Thomson et al., 2002; Binzegger et al., 2004) reveals their main discrepancies: the lateral sampling range and the specific selection of target cells. Taking them into account, we compile an integrated connectivity map and analyze the unified map by simulations of a full scale model of the local layered cortical network. The simulated spontaneous activity is asynchronous irregular and the cell-type specific spontaneous firing rates are in agreement with in vivo recordings in awake animals, including the low rate of layer 2/3 excitatory cells. Similarly, the activation patterns evoked by transient thalamic inputs reproduce recent in vivo measurements. The correspondence of simulation results and experiments rests on the consideration of specific target type selection and thereby on the integration of a large body of the available connectivity data. The cell-type specific hierarchical input structure and the combination of feed-forward and feedback connections reveal how the interplay of excitation and inhibition shapes the spontaneous and evoked activity of the local cortical network. | q-bio.NC | q-bio | The cell-type specific connectivity of the
local cortical network explains prominent
features of neuronal activity
Tobias C. Potjans1,2,3, Markus Diesmann1,2,4
1
1
0
2
n
u
J
8
2
]
.
C
N
o
i
b
-
q
[
1
v
8
7
6
5
.
6
0
1
1
:
v
i
X
r
a
1Institute of Neuroscience and Medicine (INM-6)
Computational and Systems Neuroscience
Research Center Juelich, 52425 Juelich, Germany
2Brain and Neural Systems Team
RIKEN Computational Science Research Program
Wako-shi, Saitama 351-0198, Japan
3Faculty of Biology III
Albert-Ludwigs-University Freiburg
Schaenzlestrasse 1, 79104 Freiburg, Germany
4RIKEN Brain Science Institute
Wako-shi, Saitama 351-0198, Japan
Abstract
In the past decade, the cell-type specific connectivity and activity of local cor-
tical networks have been characterized experimentally to some detail. In parallel,
modeling has been established as a tool to relate network structure to activity
dynamics. While the available connectivity maps have been used in various com-
putational studies, prominent features of the simulated activity such as the spon-
taneous firing rates do not match the experimental findings. Here, we show that
the inconsistency arises from the incompleteness of the connectivity maps. Our
comparison of the most comprehensive maps (Thomson et al., 2002; Binzegger
et al., 2004) reveals their main discrepancies: the lateral sampling range and the
specific selection of target cells. Taking them into account, we compile an inte-
grated connectivity map and analyze the unified map by simulations of a full scale
model of the local layered cortical network. The simulated spontaneous activity
is asynchronous irregular and the cell-type specific spontaneous firing rates are in
agreement with in vivo recordings in awake animals, including the low rate of layer
2/3 excitatory cells. Similarly, the activation patterns evoked by transient thalamic
inputs reproduce recent in vivo measurements. The correspondence of simulation
results and experiments rests on the consideration of specific target type selec-
tion and thereby on the integration of a large body of the available connectivity
data. The cell-type specific hierarchical input structure and the combination of
feed-forward and feedback connections reveal how the interplay of excitation and
inhibition shapes the spontaneous and evoked activity of the local cortical network.
Potjans & Diesmann, Connectivity explains cell-type specific activity
Introduction
The local cortical network is considered a building block of brain function and
over the last century, the hypothesis that the interactions of neurons within the
microcircuit are governed by the cell-type specific connectivity has been refined
(see e.g. Douglas & Martin (2007a,b) for reviews). In the last decade, progress in
methodology enabled the compilation of comprehensive connectivity maps: Thom-
son et al. (2002) used electrophysiological recordings to estimate the connection
probabilities between various cell types in layers 2/3, 4 and 5 in slices of rat and cat
neocortex. Shortly thereafter, Binzegger et al. (2004) applied a modified version
of Peters' rule (Braitenberg & Schuz, 1998) to derive the cell-type specific distri-
bution of synapses from morphological reconstructions of in vivo labeled cells from
area 17 of the cat. These approaches entail partly contradicting results (Thom-
son & Lamy, 2007), but the compatibility of these maps and the influence of the
different methodologies have not been quantitatively assessed.
Due to recent advances in in vivo electrophysiology and two-photon optical
imaging, the structural data can now be contrasted with observations of the dy-
namics. Characteristic features are the cell-type specific firing rates during ongoing
activity in awake animals: low pyramidal neuron firing rates below 1Hz are reported
in layer 2/3 (L2/3) and highest rates in L5 (e.g. Greenberg et al., 2008; de Kock
& Sakmann, 2009).
Early network models already incorporate basic anatomy and electrophysiology
such as the separation of excitatory and inhibitory cell types and a sparse, seem-
ingly random connectivity (Amit & Brunel, 1997; van Vreeswijk & Sompolinsky,
1996). The balance of excitation and inhibition explains the asynchronous irregu-
lar (AI) spiking activity and the large membrane potential fluctuations observed in
vivo. Models eventually incorporated multiple cell types to capture layer-specific
connections (e.g. Hill & Tononi, 2005; Traub et al., 2005) and employ data based
connectivity maps (Haeusler & Maass, 2007; Haeusler et al., 2009; Heinzle et al.,
2007; Izhikevich & Edelman, 2008; Binzegger et al., 2009). However, no study to
date reported cell-type specific firing rates consistent with the experimental obser-
vations. A priori it is unclear whether this is due to a misinterpretation of the raw
connectivity data or due to further model assumptions. Since the studies comprise
networks based on point-neuron as well as on multi-compartment neuron models,
the mismatch in fundamental characteristics like stationary firing rates is unlikely
to be caused by a lack in complexity of the network elements but rather by the
incompleteness of the connectivity map.
In the present study we quantify the discrepancies between the electrophysio-
logical and anatomical connectivity maps to systematically compile an integrated
map. To resolve conflicts, we additionally incorporate insights from photostim-
ulation (Dantzker & Callaway, 2000; Zarrinpar & Callaway, 2006) and electron
microscopy (McGuire et al., 1984) studies, reporting the specific selection of in-
terneurons by a subset of inter-layer projections. We then check the consistency
of structure and activity by means of full scale simulations of the local cortical
network.
1
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 1: Model definition. Layers 2/3, 4, 5 and 6 are each represented by
an excitatory (triangles) and an inhibitory (circles) population of model neurons.
The number of neurons in a population is chosen according to Binzegger et al.
(2004) based on the countings of Beaulieu & Colonnier (1983); Gabbott & Somogyi
(1986). Input to the populations is represented by thalamo-cortical input targeting
layers 4 and 6 and background input to all populations. Excitatory (black) and
inhibitory (gray) connections with connection probabilities > 0.04 are shown. The
model size corresponds to the cortical network below 1 mm2 surface.
2
Potjans & Diesmann, Connectivity explains cell-type specific activity
Materials and Methods
The network model (Fig. 1) represents four layers of cortex, L2/3, L4, L5 and
L6, each consisting of two populations of excitatory (e) and inhibitory (i) neurons.
Throughout the article, we use the term connection with reference to populations,
defined by the pre- and postsynaptic layers and neuron types. The term projection
is used for the two connections of a single presynaptic population to both popula-
tions of a target layer. A connectivity map is defined by the sixty-four connection
probabilities between the eight considered cell types.
Connectivity data
For the anatomical map (a), Binzegger et al. (2004) provide the relative number of
synapses participating in a connection and the total absolute number of synapses,
depending on pre- and postsynaptic type, of area 17 (supplementary Table 6).
The product of these measures gives the absolute number of synapses K for any
connection. To calculate the corresponding connection probabilities Ca we assume
that the synapses are randomly distributed allowing multiple contacts between any
two neurons. With N pre(post) being the number of neurons in the presynaptic
(postsynaptic) population:
(cid:18)
(cid:19)K
Ca = 1 −
The often used expression
1 −
1
N preN post
Ca = K/N preN post
.
(1)
(2)
is the corresponding first-order Taylor series approximation and valid for small
K/(N preN post) (see Supplemental Material). The original published data consti-
tute our raw connectivity map (Binzegger et al. (2004), their figure 12). Consistent
with other modeling work (Izhikevich & Edelman, 2008), we construct an improved
("modified") anatomical map by assigning the unassigned symmetric (inhibitory)
synapses, originating from a potential underestimation of interneuronal connectiv-
ity, to within-layer projections originating from local interneurons (Binzegger et al.,
2004). The derived connection probabilities are inversely proportional to the con-
sidered surface area πr2 (Fig. 2B). This can be easily understood when considering
the approximation eq. (2): the numbers of neurons N and synapses K increase
linear with the surface area and therefore Ca ∝ πr2/(πr2)2 = 1/πr2. The prod-
uct of the connection probability and the surface area is constant for large areas
(Fig. 2B). Hence, we use throughout this article the area-corrected connection
probability Ca = limr→∞ Caπr2 for all numerical values of the anatomical map.
The physiological hit rate estimates (Thomson et al. (2002), their Table 1)
provide the physiological map (p). We combine multiple independently measured
hit rates for the same connection by a weighted sum
Cp =(cid:80)
i RiQi/(cid:80)
j Qj
,
(3)
where Ri and Qi are the hit rate and the number of tested pairs in the ith ex-
periment, respectively.
In accordance with Haeusler & Maass (2007), we set the
probabilities of the L2/3i to L5e and of the L4i to L2/3i connections to 0.2. While
3
Potjans & Diesmann, Connectivity explains cell-type specific activity
these data constitute the raw map, we incorporate additional hit rate estimates
(Table 1) to create an improved (modified) physiological connectivity map. The
numerical values of all connectivity maps are listed in supplementary Table 6.
We classify all connections into two main groups:
recurrent intra-layer or
"within-layer" connections and connections between different layers or "inter-layer"
connections.
Lateral connectivity model
We use a Gaussian model for describing the lateral connection probability profile
(cid:18)−r2
(cid:19)
2σ2
C(r) = C0 exp
.
(4)
with r being the lateral distance and C0 and σ specifying the peak connec-
tion probability (zero lateral distance) and the lateral spread of connections, re-
spectively. Thereby, we assume that the underlying lateral connectivity is the
same for both connectivity maps, i.e.
that C0 and σ are universal. Further-
more, we assume that the experimental data correspond to a random sampling
of connections within a cylinder of a fixed sampling radius ra > rp and corre-
sponding mean connection probabilities Ca/p = 1/(πr2
0 C(r)rdrdϕ =
2πC0σ2/(πr2
a/p/2σ2)]. These expressions for the mean connection
probabilities of the two connectivity maps allow us to determine the two unknown
parameters C0 and σ. We obtain in closed form (see Supplemental Material)
a/p)[1− exp(−r2
(cid:82) 2π
0
a/p)(cid:82) ra/p
(cid:19)(cid:21)−1/2
(cid:20)
(cid:18)
−2 ln
1 − πr2
pCp
Ca
σ = rp
and
(5)
(6)
C0 =
Ca
2πσ2 .
In principle, this approach may be applied to an individual connection. However, the
sampling radius is the same for all connections of a map. Therefore we determine
σ and C0 only for the global mean connection probabilities of the two maps which
provides robustness against uncertainties in the probability estimate of a particular
connection.
The simulations use a laterally uniform connectivity profile, i.e. the connectivity
between two neurons is exclusively determined by the cell-types and not their
location in space. The mean connection probability of the model Cm depends on
m) and the parameters of the
the size of the network (e.g.
lateral model, but by applying eqs. (5) and (6) it can also be expressed in terms
of the experimentally accessible parameters Cp, rp and Ca
the surface area πr2
rm(cid:90)
2π(cid:90)
0
0
Cm =
1
πr2
m
C(r)rdrdϕ
=
=
(cid:104)
C0σ2[1 − exp(−r2
2
r2
m
Ca
πr2
m
1 − (1 − πr2
p
m/(2σ2))]
¯Cp/ ¯Ca)r2
m/r2
p
(cid:105)
.
(7)
4
Potjans & Diesmann, Connectivity explains cell-type specific activity
where ¯Cp, and ¯Ca specify the global means. To arrive at the individual connection
probabilities at a given model size it is sufficient to multiply a connectivity map by
the ratio of Cm and the global mean of the map ¯Ca/p.
Table 1: Modified physiological connectivity map. The map
incorporates hit rate estimates of several studies. The se-
lection of studies is based on the comprehensive review by
Thomson & Lamy (2007) including all data where hit rate
and the number of tested pairs can be extracted such that
eq. (3) is applicable. The table lists, from left to right, con-
nection specifier, number of existing connections (product
of hit rate and number of tested pairs), number of tested
pairs and the publication from which the data are extracted.
A star (*) indicates that the number of tested pairs is not
explicitly given, but estimated from the stated accuracy of
connection probability. Furthermore, we use the following
additional data on within-layer connections that is not re-
ported separately. Thomson et al. (1996); Thomson & Lamy
(2007): i to e in L2/3, L4 and L5 21 connected of 93 tested
pairs; Ali et al. (2007): i to e in L2/3, L4 and L5 30 con-
nected of 90 tested pairs, e to i in L2/3, L4 and L5 21 con-
nected of 48 tested pairs; i to e in L2/3 and L4 9 connected
of 69 tested pairs and e to i in L2/3 and L4 21 connected of
147 tested pairs. The reported numbers of connected and
tested pairs are uniformly distributed to the reported layers.
connection
L2/3e→L2/3e
L2/3e→L2/3i
L2/3i→L2/3e
L2/3i→L2/3i
L4e→L4e
existing tested
65
8
8
3
48
32
22
63
24
6
22
151
19
29
7
17
136
26
19
2
2
10
22
247
81
32
36
549
305
112
760
110
25
107
243
40
100
25
107
243
39
39
2
8
139
528
publication
Thomson et al. (2002) (rat)
Thomson et al. (2002) (cat)
Bannister & Thomson (2007) (rat)
Bannister & Thomson (2007) (cat)
Mason et al. (1991)
Kapfer et al. (2007)
Yoshimura et al. (2005)
Holmgren et al. (2003)
Ren et al. (2007)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (rat)
Holmgren et al. (2003)
Kapfer et al. (2007) (FS)
Kapfer et al. (2007)(SOM)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (rat)
Holmgren et al. (2003)
Kapfer et al. (2007) (FS)
Kapfer et al. (2007)(SOM)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (rat)
Bannister & Thomson (2007) (cat)
Bannister & Thomson (2007) (rat)
5
Potjans & Diesmann, Connectivity explains cell-type specific activity
connection
L4e→L4i
L4i→L4e
L4i→L4i
L5e→L5e
L5e→L5i
L5i→L5e
L5i→L5i
L6e→L6e
L6e→L6i
L2/3e→L4e
L2/3e→L4i
L2/3i→L4e
L2/3e→L5e
L4e→L2/3e
L4e→L2/3i
L4i→L2/3e
existing tested
4
131
11
25
8
3
11
74
36
10
10
83
26
4
3
15
50
218
29
148
26
173
19
6
7
9
3
56
4
8
5
0
0
1
7
0
0
2
16
25
247
10
7
7
64
1
3
6
23
655
89
234
42
21
154
172
63
64
40
190
74
42
6
163
500
1655
206
1233
260
1450
190
79
73
73
5
1512
204
38
21
25
70
12
37
29
10
2
29
259
1324
50
25
70
640
10
31
12
publication
Thomson et al. (2002)
Feldmeyer et al. (2005)
Beierlein et al. (2003)
Maffei et al. (2004)
Thomson et al. (2002)
Ali et al. (2007) (nonFS)
Ali et al. (2007) (FS)
Beierlein et al. (2003) (FS)
Beierlein et al. (2003) (LTS)
Ali et al. (2007) (nonFS)
Ali et al. (2007) (FS)
Beierlein et al. (2003) (FS)
Beierlein et al. (2003) (LTS)
Thomson et al. (2002)
Thomson et al. (2002)
Thomson et al. (2002)
Markram et al. (1997)
Le Be & Markram (2006)
Le Be & Markram (2006)
Wang et al. (2006)
Wang et al. (2006)
Silberberg & Markram (2007)
Thomson (1997); Thomson & Deuchars (1997)*
Thomson et al. (1995)
Thomson et al. (2002)
Thomson et al. (2002)
Thomson et al. (2002)
Mercer et al. (2005)
Beierlein & Connors (2002)
West et al. (2006) (cat)
West et al. (2006) (rat)
Thomson et al. (2002) (rat)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (rat)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (rat)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (rat)
Thomson & Bannister (1998)
Kampa et al. (2006)
Yoshimura et al. (2005)
Thomson et al. (2002) (rat)
Thomson et al. (2002) (cat)
Feldmeyer et al. (2002, 2005)
Thomson et al. (2002) (rat)
Thomson et al. (2002) (cat)
Thomson et al. (2002) (rat)
6
Potjans & Diesmann, Connectivity explains cell-type specific activity
connection
L4e→L5e
L5e→L2/3e
existing tested
10
12
1
3
37
86
29
259
publication
Thomson et al. (2002) (cat)
Feldmeyer et al. (2005)*
Thomson et al. (2002)
Thomson & Bannister (1998)
Connectivity data analysis
To compare the connectivity maps we define recurrence strength as the ratio of
the mean within-layer and the mean inter-layer connection probabilities and loop
strength as the ratio of the mean connection probability of the cortical feed-forward
loop (Gilbert (1983), L4 to L2/3 to L5 to L6 and back to L4) and the mean
connection probability of all other inter-layer connections. For a fair comparison
we base recurrence and loop strength only on connections for which estimates are
available in both data sets. Therefore, layers 2/3, 4 and 5 are included but not
layer 6.
A measure with higher resolution is the scaling factor ζ which compares the
connection probabilities of individual connections provided both connectivity maps
assign non-zero probabilities.
ζ =
max(Ca/ ¯Ca, Cp/ ¯Cp)
min(Ca/ ¯Ca, Cp/ ¯Cp)
(8)
is independent of model size because Cm cancels from the expression (see Supple-
mental Material).
Furthermore, to quantify the specificity of connections we introduce the target
specificity
T =
C post=e − C post=i
C post=e + C post=i .
(9)
as the normalized difference of the connection probabilities constituting a projec-
tion.
Consistent modifications of target specificity
In order to construct a consistent integrated connectivity map, it is necessary to
modify the target specificity of certain projections (see Results), i.e. connection
probabilities are modified to meet a given target specificity value. However, we
constrain these modifications by demanding consistency with the underlying ex-
perimental data. For the anatomical data, this underlying measure is the number
of synapses participating in a projection. For the physiological data, it is the mea-
sured connection probability of one of the two connections forming the projection
(typically the second connection has not been quantified).
Modifying the connection probabilities while conserving the total number of
synapses of a projection requires a redistribution of synapses across the target
neurons (see supplementary Fig. 13). To that end, we determine the fraction
of synapses targeting excitatory neurons ∆ as a function of the requested tar-
get specificity constrained by the total number of synapses and the sizes of the
presynaptic and the two postsynaptic populations. The main complication is that
target specificity is defined in terms of connection probabilities, see eq. (9), and
the relation of connection probability and the number of synapses is non-linear, see
7
Potjans & Diesmann, Connectivity explains cell-type specific activity
eq. (1) (for the exact expression see Supplemental Material). Nevertheless, in the
first-order Taylor series expansion of C, eq. (2), the relation is linear and substi-
tuting C post=e = ∆K/(N preN post=e) and C post=i = (1 − ∆)K/(N preN post=i) in
eq. (9) we find
∆ =
(1 + T )N post=e
(1 − T )N post=i + (1 + T )N post=e .
The modifications of the physiological data are straightforward because in all
cases considered here only one connection probability is experimentally given so
that we can estimate the unknown value based on the definition eq. (9):
(cid:18)1 − T
(cid:19)+(−)1
1 + T
C post=i(e) =
C post=e(i).
Compilation of the integrated connectivity map
We compile the integrated connectivity map algorithmically. The procedure re-
quires as input the anatomical and the physiological connectivity maps as well as
the model size and information about desired modifications of the target specifici-
ties. The procedure automatically estimates the parameters of the Gaussian model
and, based on these, the mean model connectivity. Subsequently the procedure
separately modifies the target specificities of both scaled maps as instructed and
finally averages the two maps (see also Supplemental Material).
Layer-specific external input
inputs:
We distinguish three types of layer-specific external
thalamic afferents
as reconstructed by Binzegger et al. (2004), "gray-matter" external inputs, i.e.
intrinsic non-local inputs entering the local network through the gray matter, which
we estimate based on the properties of axonal clusters (Binzegger et al., 2007),
and "white-matter" external inputs, which include all inputs not covered by the
previous types. The number of the latter is estimated based on the difference of
the number of synapses the different cell types receive according to Binzegger et al.
(2004) and the total synapse count in a given layer (Beaulieu & Colonnier, 1985).
The thalamic afferents are included in the anatomical connectivity map (Binzeg-
ger et al., 2004). We extract the gray-matter inputs from the information on bou-
ton distributions in 3-dimensional space described in Binzegger et al. (2007). The
authors find that boutons of all cell types form multiple clusters and the article
provides the lateral distance of cluster centers and the corresponding somata. We
interpret a cluster to be non-local if the lateral distance to the soma is larger than
≈ 0.56 mm (corresponding to a local network surface area of 1 mm2). By addi-
tionally using data on the relative sizes of different cluster types, we estimate the
proportion of intrinsic gray-matter connections that originate outside of the local
network. Thereupon, we use the estimated proportion of gray-matter inputs and
the number of local connections in our network to calculate the absolute number
of gray-matter inputs.
In this way we construct an estimate of the gray-matter
external inputs that is consistent with both, the axonal structure in Binzegger et al.
(2007) and the structure of our model. The detailed procedure is described in the
Supplemental Material.
The white-matter inputs are estimated based on the comparison of the abso-
lute number of synapses obtained in Binzegger et al. (2004) which only contains
8
Potjans & Diesmann, Connectivity explains cell-type specific activity
the contributions from local, thalamic and gray-matter synapses, with those in
Beaulieu & Colonnier (1985) containing all synapses. The difference has been
termed the "dark matter" of cortex (Binzegger et al., 2004) and, in case of the
excitatory synapses, is usually interpreted as white-matter external inputs. The
explicit numbers are published in Izhikevich & Edelman (2008, their figure 9 of the
supplemental material) at subcellular resolution. As our model is based on point
neurons, we sum over all contributions to a given cell type and average across the
cell types that are collapsed to a single population in our model. Thereby, our esti-
mates take neuronal morphology into account. The resulting counts for the three
external input types as well as the total number of external inputs to the excitatory
populations are given in supplementary Table 7. Since long-range projections tar-
get excitatory as well as inhibitory neurons (Johnson & Burkhalter, 1996; Gonchar
& Burkhalter, 2003), we choose target specificity values for external inputs to be
comparable to recurrent connections, resulting in similar total numbers of external
inputs to inhibitory neurons (Table 3).
The cell-type specific external inputs are used as a reference parametrization
of the model. To comprehensively assess the dependence of the activity on exter-
nal inputs, we conduct simulations in which we choose the external inputs to a
population randomly with uniform distribution: The inputs to the excitatory popu-
lations are constrained by the values from the reference and the layer-independent
parametrization (compare supplementary Fig. 18). Those to inhibitory populations
are chosen such that the target specificity of the external input to a given layer is
between 0 and 0.1 (to L2/3, L4 and L5) and to L6, due to the high number of
external inputs to excitatory cells in this layer, between 0 and 0.2.
Network simulations
The network is defined by eight neuronal populations representing the excitatory
and inhibitory cells in the four layers 2/3, 4, 5 and 6. The populations consist
of current-based leaky integrate-and-fire model neurons with exponential synaptic
currents (see Supplemental Material) and are randomly connected with connection
probabilities according to the integrated connectivity map we derive in this arti-
cle. Every population receives Poissonian background spike trains (Amit & Brunel,
1997; Brunel, 2000a); the firing rates of these inputs are composed by the number
of external inputs a neuron in a particular population receives and the background
spike rate contributed by each synapse. Synaptic strengths and synaptic time con-
stants of all connections are chosen such that an average excitatory postsynaptic
potential (EPSP) has an amplitude of 0.15 mV with a rise time of 1.6 ms and
a width of 8.8 ms mimicking the in vivo situation (Fetz et al., 1991).
Inhibitory
strengths are negative and increased by a factor g compared to the excitatory ones.
Delays in the network are chosen independent of the layer, with excitatory delays
being on average around twice as long as inhibitory delays but the exact ratio
is uncritical. To introduce heterogeneity into the network, we draw the synaptic
strengths and delays from Gaussian distributions (prohibiting a change of sign of
the strengths and constricting delays to be positive and multiples of the computa-
tion step size). The network structure and a complete list of parameters and their
values for the reference network model are systematically described according to
Nordlie et al. (2009) in Tables 2 and 3.
In some cases, we explicitly model the thalamic input to L4 and L6 by a thala-
mic population of 902 neurons (Peters & Payne, 1993) that emit Poissonian spike
trains at a given rate in some prescribed time interval. These relay cells are ran-
9
Potjans & Diesmann, Connectivity explains cell-type specific activity
A: Model summary
Populations
Topology
Connectivity
Neuron model
nine; eight cortical populations and one thalamic population
--
random connections
cortex: leaky integrate-and-fire, fixed voltage threshold, fixed absolute
refractory period (voltage clamp), thalamus: fixed-rate Poisson
Synapse model exponential-shaped postsynaptic currents
Plasticity
Input
Measurements spike activity, membrane potentials
B: Populations
--
cortex: independent fixed-rate Poisson spike trains
Type
Elements
Cortical network iaf neurons
Th
Poisson
Number of populations Population size
N (type-specific)
eight, two per layer
one
Nth
C: Connectivity
Type
Weights
Delays
random connections with independently chosen pre- and postsynaptic
neurons; see Table 3 for probabilities
fixed, drawn from Gaussian distribution
fixed, drawn from Gaussian distribution multiples of sim. stepsize
D: Neuron and synapse model
Name
Type
Subthreshold
dynamics
Spiking
+ I(t)
Cm
τm
iaf neuron
leaky integrate-and-fire, exponential-shaped synaptic current inputs
dt = − V
dV
V (t) = Vreset
Isyn(t) = w e−t/τsyn
If V (t−) < θ ∧ V (t+) ≥ θ
1. set t∗ = t, 2. emit spike with time stamp t∗
if (t > t∗ + τref)
else
Type
Background
Target
iaf neurons
E: Input
Description
independent Poisson spikes (see Table 3)
F: Measurements
Spiking activity and membrane potentials from a subset of neurons in every population
Table 2: Model description after Nordlie et al. (2009).
10
Potjans & Diesmann, Connectivity explains cell-type specific activity
L2/3e L2/3i L4e
Name
Population size N 20683 5834 21915 5479
External inputs kext 1600 1500 2100 1900
Background rate νbg 8 Hz
L4i
B+E: Populations and Inputs
L5i
1065
1900
L5e
4850
2000
L6i
L6e
14395 2948
2900
Th
902
2100 --
L4i
L6e
L5i
0.008
0.0
0.004
0.0
0.0003 0.045
0.106
0.0
0.020
0.373
0.316
0.009
0.040
0.020
0.008
0.066
C: Connectivity
from
L5e
0.032
0.075
0.007
0.003
0.083
0.060
0.057
0.028
Description
excitatory synaptic strengths
relative inhibitory synaptic strength
excitatory synaptic transmission delays
inhibitory synaptic transmission delays
L2/3e L2/3i L4e
0.101 0.169 0.044 0.082
0.135 0.137 0.032 0.052
0.008 0.006 0.050 0.135
0.069 0.003 0.079 0.160
0.100 0.062 0.051 0.006
0.055 0.027 0.026 0.002
0.016 0.007 0.021 0.017
0.036 0.001 0.003 0.001
Value
87.8 ± 8.8 pA
−4
1.5 ± 0.75 ms
0.8 ± 0.4 ms
L6i
0.0
0.0
0.0
0.0
0.0
0.0
0.225
0.144
Th
0.0
0.0
0.0983
0.0619
0.0
0.0
0.0512
0.0196
L2/3e
L2/3i
L4e
L4i
L5e
L5i
L6e
L6i
to
Name
w ± δw
g
de ± δde
di ± δdi
Name
τm
τref
τsyn
Cm
Vreset
θ
νth
Table 3: Parameter specification. The categories refer to the model description
in Table 2.
D: Neuron Model
Description
membrane time constant
absolute refractory period
postsynaptic current time constant
membrane capacity
reset potential
fixed firing threshold
thalamic firing rate during input period
Value
10 ms
2 ms
0.5 ms
250 pF
−65 mV
−50 mV
15 Hz
11
Potjans & Diesmann, Connectivity explains cell-type specific activity
domly connected to the cortex with the cell-type specific connection probabilities
according to Binzegger et al. (2004), see Table 3.
To instantiate the network model we randomly draw for every synapse the pre-
and the postsynaptic neuron. In contrast to the often used convergent and diver-
gent connectivity schemes (Eppler et al., 2009), this procedure results in binomially
distributed numbers of incoming and outgoing synapses. In practice, we could first
calculate the total number of synapses forming a connection by inverting eq. (1)
and then successively create the synapses. In a distributed simulation setup, how-
ever, this procedure is inefficient because the neurons are distributed over multiple
processes. Although a synapse is only created if the postsynaptic neuron is local to
the process (Morrison et al., 2005) the full algorithm would have to be carried out
on each process. We solve this problem by calculating a priori how many synapses
will be created locally on each process, exploiting that the distribution of synapses
over processes is multinomial. Subsequently we apply the serial algorithm on every
machine to the local synapses only: The presynaptic neuron is drawn from all neu-
rons in the presynaptic population and the postsynaptic cell on a given process is
drawn only from the neurons located on this process (compare supplementary Fig.
16). While the first step is serial, but efficient for the number of processes we typ-
ically employ, the second step is fully parallelized. The procedure is detailed in the
Supplemental Material. A reference implementation of this algorithm will be made
available in the NEST simulation tool (http://www.nest-initiative.org) as
RandomPopulationConnect.
All simulations are carried out with the NEST simulation tool (Gewaltig &
Diesmann, 2007) using a grid constrained solver and a computation step size
h = 0.1 ms on a compute cluster with 24 nodes each equipped with 2 quad core
AMD Opteron 2834 processors and interconnected by a 24 port Voltaire InfiniBand
switch ISR9024D-M. 48 cores simulate the network of around 80,000 neurons and
0.3 billion synapses in close to real-time (Djurfeldt et al., 2010).
Results
Comparison of connectivity maps
The anatomical and the physiological connectivity maps exhibit a similar struc-
ture (Fig. 2): Recurrent within-layer connections are all non-zero with connection
probabilities tending to decrease from superficial to deeper layers, excepting the
physiological interneuron to interneuron connections which are, however, subject
to poor statistics (Table 1). The inter-layer connections can be subdivided into
connections with similar probabilities as within-layer connections and those with
values close or equal to zero. Consequently, the recurrence strength, quantifying
the relative strength of recurrent within-layer connections, is greater than one and
statistically indistinguishable for the two maps (z-test, P > 0.1, Fig. 2C). Similarly,
the loop strength, comparing the connectivity of the feed-forward loop to all other
inter-layer connections, is indistinguishable (z-test, P > 0.05, Fig. 2D). However,
50% of the scaling factors of inter-layer connections are large (L2/3e to L4i, L5e
to L2/3e, L4i to L2/3e and L4i to L2/3i) for both, the raw and the modified data.
Lateral connectivity
In contrast to the largely consistent relative measures (Fig. 2C and D), absolute
connection probability values differ (supplementary Table 6). We hypothesize that
12
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 2: Properties of connectivity maps. (A) Connection probabilities accord-
ing to the anatomical (top panel, circles) and physiological connectivity map (cen-
ter panel, squares) and corresponding scaling factors (bottom panel, triangles).
Both, raw (closed markers) and modified (open markers) are shown. The data are
horizontally arranged according to their classification as within-layer (black) and
inter-layer (gray) connections. For a given pair of pre- and post-synaptic layer,
the data are arranged from left to right according to connection types: excita-
tory to excitatory, excitatory to inhibitory, inhibitory to excitatory and inhibitory
to inhibitory. L5i to L5e/i outside of the displayed range, see supplementary Ta-
ble 6. Error bars are minimal statistical errors (see Supplemental Material). (B)
Sampling radius dependence of anatomical connection probability (solid: eq. (1),
dashed: eq. (2)) and of the product of connection probability and area (dotted)
in double-logarithmic representation. (C) Anatomical and physiological recurrence
strength. (D) Anatomical and physiological loop strength. Error bars in C and
D are based on the minimal statistical error estimates of connection probabilities
using error propagation.
13
0.00.10.2anatomicalconnection probabilitiesconnectionprobabilityABCD0.00.20.40.6physiological15101520scaling factorL2/3 to L2/3 L4 to L4 L5 to L5 L6 to L6 L2/3 to L4 L2/3 to L5 L2/3 to L6 L4 to L2/3 L4 to L5 L4 to L6 L5 to L2/3 L5 to L4 L5 to L6 L6 to L2/3 L6 to L4 L6 to L5anat.phys.024recurrence strengthanat.phys.024loop strength0.10.51.0sampling radius [mm]0.010.11.0 Potjans & Diesmann, Connectivity explains cell-type specific activity
the differences in the mean connection probabilities are explained by differences in
the methodology applied to obtain the connectivity maps: Physiological recordings
in slices are usually restricted to a maximal lateral distance of the somata of around
100 µm (as reported in Thomson & Morris (2002) for the raw physiological map).
The anatomical data, in contrast, are based on reconstructed axons and dendrites,
with axons extending up to 4mm (Binzegger et al., 2007), in general beyond 1mm.
When providing absolute numbers, Binzegger et al. (2004) refer to the surface area
of cat area 17 (399 mm2).
We employ a Gaussian lateral connectivity model (eq. (4)), similar to the one
by Hellwig (2000); Buzas et al. (2006) to account for the different experimental
sampling radii (see Materials and Methods). We assume the model to reflect the
in vivo connectivity structure and the experimental connectivity maps to charac-
terize samples of this structure. The anatomical measurement is interpreted as
an unconstrained sampling over the complete lateral connectivity structure while
the physiological measurement corresponds to a local measure sampling from the
center region of the Gaussian (Fig. 3A).
The model parameters, peak connection probability and lateral spread, are
determined based on the mean connection probabilities of the two maps and the
physiological sampling radius (eqs. (5) and (6), Fig. 3). The estimated lateral
spread is consistent with data from rat and cat primary visual cortex based on the
potential connectivity method: Hellwig (2000), caption to his figure 7, reports a
lateral spread of 150 to 310 µm, Stepanyants et al. (2008), their figure 7, find a
spread of around 200 µm of main projections in input and output maps. Also the
overall connectivity level of 0.138 for nearby neurons with a distance of 100µm is in
good agreement with the extensively used estimate of 0.1 provided by Braitenberg
& Schuz (1998). These consistencies indicate that our underlying assumption
-- anatomical and physiological experiments sample independently from the same
lateral connectivity profile -- is valid.
Average model connectivity Our network model does not incorporate a lateral
connectivity structure but randomly connected populations. We use the Gaussian
lateral connectivity model exclusively to determine the average connection proba-
bility for a given network size (eq. (7), Fig. 3D). The connectivity of small networks
(up to about 7,000 neurons) is largely determined by the physiological connectivity
and the one of large networks (above 100,000 neurons) by the anatomical con-
nection probability (decaying quadratically, see Fig. 2B). For intermediate network
sizes, eq. (7) interpolates between these two extremes according to the Gaussian
lateral connectivity profile. Fig. 3E shows the average synaptic convergence as a
function of the network size. It reveals that only for large networks the majority of
local synapses is represented in the model: for example a network of around 80,000
neurons comprises more than 85% of all local synapses. In contrast, a network of
e.g. 10,000 neurons represents only around 20% of the local connectivity. There-
fore, we select our network to correspond to 1 mm2 of cortical surface (77,169
neurons).
According to our analysis the maximal average number of synapses per neuron
is about 5,000. This number is consistent with the data of Binzegger et al.(2004,
see their figure 11A). In Fig. 3 we also display the connectivities and convergences
of a selection of other modeling studies on the local cortical network (some data
representing local networks embedded in larger networks). Independent of model
size most studies use a connection probability around 0.1 which is largely consistent
14
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 3: Lateral connectivity model. (A) Two dimensional Gaussian with two
cylinders indicating the lateral sampling of the anatomical (gray outer cylinder) and
the physiological (black inner cylinder) experiments. (B) Estimated peak amplitude
and (C) lateral spread of the connectivity model based on mean connectivity of
the anatomical and physiological raw and modified maps. (D) Average connection
probability and (E) average synaptic convergence (average number of synaptic
inputs per neuron) of the layered network model as a function of the network
size expressed in number of neurons. The dashed horizontal line marks 85% of
the maximal synaptic convergence in the local network. Black diamonds show
the data used in our simulations, further markers indicate other published cortical
network models: Haeusler & Maass (2007) (light gray square), Izhikevich (2006)
(light gray triangle), Izhikevich et al. (2004) (light gray circle, embedded local
network is defined by the area receiving connections from a single long-range axon),
Lundqvist et al. (2006); Djurfeldt et al. (2008) (light gray diamond, local network
represented by one hypercolumn), Vogels & Abbott (2005); Vogels et al. (2005)
(dark gray circle), Sussillo et al. (2007) (dark gray square), Brunel (2000a) (dark
gray triangle), Kriener et al. (2008) (dark gray diamond), Kumar et al. (2008)
(black circles), Morrison et al. (2007) (black square).
15
Potjans & Diesmann, Connectivity explains cell-type specific activity
with our results. Only for large networks of 50,000 neurons and more, this con-
nection probability is above our estimate. Two studies use a significantly smaller
connection probabilities of 0.02, arguing that this value interpolates between high
local and low distal connectivity. Although the absolute numbers differ from our
estimates, the reasoning is the same as for our model. In all but two cases (Mor-
rison et al., 2007; Kumar et al., 2008) the models' convergence is maximally 20%
of the anatomical estimate.
Randomness and specificity
A central assumption of the anatomical connectivity map is randomness,
i.e.
synapses are established independent of the pre- and postsynaptic cell type. Never-
theless, the target specificity estimates of the anatomical connectivity map (circles
in Fig. 4) are > 0.2, reflecting a preferential selection of excitatory targets. The
bias is introduced by the application of Peters' rule, which assumes synaptic densi-
ties on dendrites to be independent of cell-type, to bouton densities and dendritic
lengths. Furthermore, some projections exhibit very high values > 0.5 because
only the dendrites of excitatory cells, but barely those of inhibitory cells, reach into
the cloud of presynaptic axonal elements.
The target specificity values of the physiological map contrast with the anatomi-
cal findings. Most values (squares in Fig. 4) are consistently smaller and show larger
variability than the anatomical estimates. In addition, excitatory within-layer con-
nections of the raw connectivity map univocally select their targets independent
of the postsynaptic type. Several projections in the physiological map, however,
connect exclusively to excitatory neurons due to incomplete sampling of inhibitory
subtypes. This highlights that a straightforward application of the physiological
connectivity map in simulations results in artifacts due to missing feed-forward
inhibition. Finally, the projection from L2/3e to L4 specifically targets inhibitory,
but not excitatory cells (see also Table 1).
The specific selection of inhibitory targets of the L2/3e to L4 projection cannot
be explained by differences in the overlap of the excitatory and inhibitory dendrites
with the excitatory axons and is therefore beyond the scope of anatomical studies
relying on the statistics of neuronal morphology such as Peters' rule (Binzegger
et al., 2004) and also potential connectivity (Stepanyants et al., 2008). The speci-
ficity in the inter-layer circuitry explains the large scaling factor of this projection
(Fig. 2A).
We identify three additional candidates of specific target type selection: L5e to
L2/3 (Dantzker & Callaway (2000); Thomson & Bannister (1998); Thomson et al.
(2002); Lefort et al. (2009)) L2/3e to L6 (Zarrinpar & Callaway, 2006) and L6e
to L4 (McGuire et al., 1984), see Supplemental Material for details on the choice
of these candidates. The structure of these projections has not been quantified
comprehensively in paired recordings and evidence suggests that the target type
selection is less strict than for the L2/3e to L4 projection. We tentatively assume
them to have intermediate values (triangles in Fig. 4). Two of these projections
(L2/3e to L4 and L5e to L2/3) are inverse to the feed-forward loop. Thomson
& Morris (2002); Thomson et al. (2002) argued that the specific target type
selection plays a distinct functional role because the inhibition-specific ("i-specific")
feedback projections may prevent reverberant excitation involving L2/3, L4 and L5
and enhance the propagation of synchronous thalamic inputs.
We utilize the information on i-specific feedback and the anatomical estimates
to remove for some projections the methodological biases by consistently mod-
16
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 4: Anatomical (circles) and physiological (squares) target specificity es-
timates based on the modified connectivity maps. +1(−1) indicates exclusive
selection of excitatory (inhibitory) targets, 0 random selection. Triangles show
additional candidates of inhibition-specific projections discovered in photostimula-
tion or electron microscopy studies. The shaded area highlights the estimates of
within-layer projections and inter-layer projections with target specificity values be-
tween 0 and 0.5 ("non-specific" connections). Within these bounds the anatomical
and physiological estimates are 0.32 ± 0.07 and −0.17 ± 0.28, respectively. The
largest statistically well constrained values from a single laboratory are exhibited
by the projections L2/3e to L2/3, L4e to L4 and L5e to L5 (compare supplemen-
tary Table 6) of the raw physiological map. The target specificity of these data is
−0.01 ± 0.03. The four highlighted data points are candidates of specific target
type selection. For every pair pre- and post-synaptic layer the figure shows data
both pre-synaptic neuron types (left: excitatory, right: inhibitory).
17
-1.0-0.50.00.51.0target specificityL2/3 to L2/3 L4 to L4 L5 to L5 L6 to L6 L2/3 to L4 L2/3 to L5 L2/3 to L6 L4 to L2/3 L4 to L5 L4 to L6 L5 to L2/3 L5 to L4 L5 to L6 L6 to L2/3 L6 to L4 L6 to L5 Potjans & Diesmann, Connectivity explains cell-type specific activity
projection
L2/3e to L4
L5e to L2/3
L2/3e to L6
L6e to L4
L2/3e to L5
L4e to L5
L2/3i to L5
L4i to L2/3
data source
T
-0.8 Thomson et al. 2002
-0.4 Dantzker & Callaway 2000
-0.4 Zarrinpar & Callaway 2006
-0.4 McGuire et al. 1984
0.29 Binzegger et al. 2004
0.32 Binzegger et al. 2004
0.4 Binzegger et al. 2004
0.23 Binzegger et al. 2004
Table 4: Amendment candidates for target specificity. The rows describe the
projections whose target specificities are modified during the compilation of the
integrated connectivity map. For each projection the second column states the
target specificity value T after the amendment, the third the publication on which
the modification is based. The top four rows are the candidate projections for
a preference of inhibitory targets.
In these cases, no quantitative estimates are
known. T = −0.8 is set if the literature provides a strong indication and T = −0.4
for a comparably weak indication (compare Supplemental Material). The T -values
of the bottom four rows are based on the anatomical map and provide estimates of
the previously not measured connections to inhibitory neurons for the physiological
map.
ifying the respective target specificities (see Materials and Methods and Table
4). Thereby, we estimate previously not measured physiological connection prob-
abilities and introduce the specific selection of targets into the anatomical map.
The latter constitutes an effective redistribution of synapses and corresponds to a
refinement of Peter's rule (supplementary Fig. 13).
Integrated connectivity map
Based on the information gathered in the previous sections we now compile an inte-
grated connectivity map. Our algorithmic compilation (see Materials and Methods)
accounts for the differences in the lateral sampling of the two methods and corrects
for methodological shortcomings expressed in the target specificity of projections
(Table 4) by incorporating photostimulation and electron microscopy data.
The resulting connection probabilities are shown in the upper panel of Fig. 5.
Visual inspection shows that the overall structure of the original connectivity maps
is preserved. For a consistency check, we calculate the scaling factors between
our integrated map and the recently published data for excitatory to excitatory
connections of mouse C2 barrel column (Lefort et al., 2009).
In this comparison
we combine the data of Lefort et al. on L2 and L3 to L2/3 and on L5A and L5B
to L5 according to eq. (3). The scaling factors (see Fig. 5, center panel) are low,
indicating a good agreement of the connectivity maps. The main outlier is the
recurrent L4e to L4e connection, which exhibits the highest connection probability
in the study of Lefort et al. (2009) but rather low values in the physiological
and especially the anatomical map. The lower panel of Fig. 5 shows the target
specificity values of the integrated connectivity map. Within-layer projections select
their targets randomly and inter-layer projections inherit mostly the properties of
the anatomical map, except for the four candidate i-specific projections.
The cell-type specific convergences and divergences (Fig. 6) show that the in-
tegrated connectivity map reflects prominent features of local cortical connectivity:
18
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 5: Properties of the integrated connectivity map. Layout of graphs as in
Fig. 2 and Fig. 4 respectively. The top panel shows the connection probabilities of
the model (L5i to L5e/i outside the displayed range, see Table 3). The center panel
shows the scaling factors between the integrated map, restricted to excitatory to
excitatory connections, and the recently published map of Lefort et al. (2009).
The bottom panel shows the target specificity of the integrated connectivity map
(within-layer connections: −0.02 ± 0.17, non-specific inter-layer connections, see
Fig. 4: 0.33 ± 0.08).
19
-1.0-0.50.00.51.0target specificityscaling factorconnection probabilityL2/3 to L2/3 L4 to L4 L5 to L5 L6 to L6 L2/3 to L4 L2/3 to L5 L2/3 to L6 L4 to L2/3 L4 to L5 L4 to L6 L5 to L2/3 L5 to L4 L5 to L6 L6 to L2/3 L6 to L4 L6 to L50.00.10.2123 Potjans & Diesmann, Connectivity explains cell-type specific activity
Except for L5e, convergence is dominated by within-layer connections (consistent
with e.g. Douglas & Martin, 1991, 2004). Furthermore, the strongest inter-layer
excitatory to excitatory divergences correspond to the feed-forward loop from L4
to L2/3 to L5 to L6 and back to L4 (Gilbert, 1983; Gilbert & Wiesel, 1983).
The excitatory to inhibitory divergence is dominated by the i-specific feedback
connections.
By comparing the average convergence and divergence to the neuronal densities
of the different cell populations (supplementary Fig. 14) we find that neurons in
the local microcircuit sample most excitatory inputs from L2/3 and fewest from
L5 and L6. In contrast, local outputs within the microcircuit project preferentially
to L5.
External inputs The model consists in total of about 217 million excitatory
and 82 million inhibitory synapses. The inhibitory synapse count is consistent with
Beaulieu & Colonnier (1985) (64 ± 21 million) while the number of excitatory
synapses is lower than their estimate (339± 43 million) presumably reflecting that
a fraction of all excitatory synapses originates outside of the local network. The
ratio of local synapses (total number of synapses in our network model) and all
synapses (according to the countings of Beaulieu & Colonnier (1985)) is 0.74,
similar to the ratio reported by Binzegger et al. (2004), but see Stepanyants et al.
(2009).
The layer-specific external input structure distinguishes between thalamic af-
ferents, gray-matter and white-matter inputs (see Materials and Methods). The
resulting total number of external inputs per neuron (Table 3) is lowest in L2/3,
intermediate in L4 and L5 and highest in L6. This relative structure is largely
dominated by the white-matter inputs. In contrast, most gray-matter long-range
inputs form synapses on L2/3 neurons and fewest form synapses in L6. Given the
presently available data, we cannot exclude that some synapses treated here as
external input actually represent a pathway in the local microcircuit still awaiting
comprehensive experimental assessment.
Spontaneous layer-specific activity
The simulated spontaneous spiking activity of all cell types corresponds to the
asynchronous irregular activity state observed in mono-layered balanced random
network models (Amit & Brunel, 1997; Brunel, 2000a). Fig. 7 (A-D) shows the
ongoing spontaneous spiking activity of all populations and the corresponding firing
rates, irregularity and synchrony. The activity varies significantly across cell types.
L2/3e and L6e exhibit the lowest firing rates with a mean below or close to 1 Hz.
L4e cells fire more rapidly at around 4 Hz and L5e cells at more than 7 Hz. In all
layers, inhibitory firing exceeds excitatory rates. The boxplots furthermore visualize
that the firing rates of single neurons can differ substantially. For instance in
L2/3e, several neurons fire at more than 5 Hz while the majority of neurons is
rather quiescent emitting less than one spike per second. This effect is due to the
binomially distributed convergence.
Single unit activity is irregular; the mean of the single unit coefficients of
variation of the inter-spike intervals is greater than 0.8. The population activity
is largely asynchronous, but exhibits fast oscillations with low amplitude similar
to balanced random networks (e.g. Brunel, 2000a). We assess the synchrony of
every population's multi-unit activity by the variability of the spike count histogram
(Fig. 7D). At the given firing rates and bin width, the synchrony of the spiking
20
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 6: Cell-type specific convergence (A) and divergence (B) of the integrated
connectivity map. The histograms display blocks of data for the four different
connection types between excitatory (e) and inhibitory (i) neurons (as indicated).
For a neuron in the layer specified on the horizontal axis, the individual bar segments
show the absolute number of synapses the neuron receives from a source layer
(convergence) or establishes in a target layer (divergence). Bar segments are
arranged according to the physical location of the layers in the cortex (from top
to bottom: L2/3, L4, L5, L6). Hatched bars represent within-layer connections.
Lightness of gray increases from superficial to deeper layers. Gray horizontal lines
indicate the convergence and divergence of a balanced random network with the
same total number of neurons and synapses as the layered model.
21
010002000300040005000convergencee to eAi to ee to ii to i010002000300040005000divergenceL2/3L4L5L6e to eBL2/3L4L5L6e to iL2/3L4L5L6i to eL2/3L4L5L6i to i Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 7: Spontaneous cell-type specific activity.
(A) Raster plot of spiking
activity recorded for 400 ms of biological time of layers 2/3, 4, 5 and 6 (from
top to bottom; black: excitatory, gray: inhibitory). Relative number of displayed
spike trains corresponds to the relative number of neurons in the network (total
of 1862 shown). (B-D) Statistics of the spiking activity of all eight populations
in the network based on 1000 spike trains recorded for 60 s (B and C) and 5 s
(D) for every population. (B) Boxplot (Tukey, 1977) of single-unit firing rates.
Crosses show outliers, stars indicate the mean firing rate of the population. (C)
Irregularity of single-unit spike trains quantified by the coefficient of variation of
the inter-spike intervals. (D) Synchrony of multi-unit spiking activity quantified by
the variance of the spike count histogram (bin width 3 ms) divided by its mean.
(E-H) Histograms of the excitatory and inhibitory population firing rates for L2/3
(E), L4 (F), L5 (G) and L6 (H) for randomly drawn external inputs (100 trials, see
Materials and Methods for details).
22
100200300400time [ms]L6L5L4L2/3ABCDEFGH0510152025firing rates [Hz]L6L5L4L2/30.00.20.40.60.81.0irregularityL6L5L4L2/302468synchronyL6L5L4L2/3051015firing rates [Hz]L6 (a.u.)0510152025firing rates [Hz]L5 (a.u.)0510firing rates [Hz]L4 (a.u.)0510firing rates [Hz]L2/3 (a.u.) Potjans & Diesmann, Connectivity explains cell-type specific activity
activity is highest in L5e and lowest in L6. The synchrony of the membrane
potential traces (Golomb, 2007, supplementary Fig. 20) confirms that the activity
is asynchronous.
The observed activity features of the network do not depend on the specific
structure of the external inputs: replacing the Poissonian background by a constant
current (DC) to all neurons (supplementary Fig. 17) or applying layer-independent
Poissonian background inputs (supplementary Fig. 18) yields similar results. Ran-
dom selection of the number of external inputs (see Materials and Methods) also
confirms the previous findings: Fig. 7 (E-H) shows the histograms of the population
firing rates in the different layers for excitatory and inhibitory cells, respectively.
The distribution of excitatory population firing rates in L2/3 and L6 are persistently
low. L5e activity exhibits the highest firing rates and also the largest variability.
L4e and the inhibitory populations vary only slightly with mean firing rates similar
to the reference parametrization (Fig. 7B). Apparently, the local microcircuitry
reconfigures the firing rate distributions for different input situations while con-
serving general features like the low rate regime in L2/3e and L6e. In 85% of the
simulations, L2/3e and L6e fire at a lower rate than L4e and simultaneously L5e
exhibits the highest firing rate. Also, we observe the inhibitory firing rate within a
given layer to be higher than the excitatory rate in 83% of all cases.
Comparison to in vivo activity Table 5 contrasts the experimentally observed
firing rates in individual layers with our simulation results. Experimentally, the
spontaneous activity of L2/3e pyramids has been studied extensively. Consistent
over species and areas the firing rate is below 1Hz, in good quantitative agreement
with the model. The L6e firing rates are similar to the model values, although the
experimental data base is more sparse. The L4e and L5e firing rates of rat primary
somato-sensory cortex are lower than in the model, but L5e consistently shows the
highest spontaneous activity, also in rat auditory cortex. The activity of cortico-
tectal cells in L5 (and L4) of various cortices in the rabbit is slightly higher. Several
studies provide data on putative interneurons (Swadlow, 1988, 1989, 1991, 1994;
Sakata & Harris, 2009), demonstrating that inhibitory activity is typically higher
than that of excitatory cells. The statistics of single neuron spike trains in our
model show great variations due to the random connectivity which also imposes
a variance in the convergence of inputs. Therefore, "neighboring" neurons, i.e.
neurons with statistically identical connectivity, can exhibit very different firing
rates, which has also been observed experimentally (e.g. Swadlow, 1988; Heimel
et al., 2005; de Kock & Sakmann, 2009).
Stability of the network activity The low-rate asynchronous irregular (AI)
firing regime has been considered to be the ground state of cortical activity (Amit
& Brunel, 1997). For the balanced random network model, the region of stability
of the AI state is bounded by the relative strength of inhibitory synapses (> 4)
and sufficiently high background rates (Brunel, 2000a). Fig. 8 characterizes the AI
state of the layered network. As in the balanced random network model, the layered
network requires sufficient external input and sufficiently strong inhibitory coupling
to enter the AI state.
Increasing the background rate predominantly affects the
activity of L4e, while an increase of the relative inhibitory synaptic strength de-
creases mostly the activity of L5e cells. Consequently, the order of excitatory firing
rates (smallest in L2/3 and L6, highest in L5) is largely preserved except for large
relative inhibitory synaptic strengths (combined with large background rates).
23
Potjans & Diesmann, Connectivity explains cell-type specific activity
i
d
e
n
a
t
b
o
s
e
t
a
r
g
n
i
r
fi
n
a
e
m
d
e
v
l
o
s
e
r
-
r
e
y
a
l
e
h
t
w
o
h
s
s
n
m
u
o
c
l
l
a
c
i
r
e
m
u
N
.
s
e
t
a
r
g
n
i
r
fi
c
fi
i
c
e
p
s
e
p
y
t
-
l
l
e
c
d
e
t
a
u
m
l
i
s
d
n
a
d
e
r
u
s
a
e
m
y
l
l
a
t
n
e
m
i
r
e
p
x
E
:
5
e
l
b
a
T
s
a
e
r
a
d
e
t
a
g
i
t
s
e
v
n
i
e
h
T
.
)
n
o
i
t
a
i
v
e
d
d
r
a
d
n
a
t
s
e
h
t
y
l
l
a
n
o
i
t
i
d
d
a
s
e
t
a
c
i
d
n
i
w
o
r
t
s
a
l
(
y
d
u
t
s
g
n
i
l
e
d
o
m
t
n
e
s
e
r
p
e
h
t
n
i
d
n
a
i
s
g
n
d
r
o
c
e
r
l
a
m
n
a
i
e
k
a
w
a
o
v
i
v
n
i
n
i
d
e
t
a
g
i
t
s
e
v
n
i
s
a
e
r
a
r
u
o
f
e
h
T
.
x
e
t
r
o
c
y
r
o
t
i
d
u
a
y
r
a
m
i
r
p
:
1
A
,
x
e
t
r
o
c
r
o
t
o
m
y
r
a
m
i
r
p
:
1
M
,
x
e
t
r
o
c
l
a
u
s
i
v
y
r
a
m
i
r
p
:
1
V
,
x
e
t
r
o
c
y
r
o
s
n
e
s
-
o
t
a
m
o
s
y
r
a
m
i
r
p
:
1
S
e
r
a
5
L
,
s
e
i
d
u
t
s
r
u
o
f
e
s
e
h
t
n
I
.
y
l
e
v
i
t
c
e
p
s
e
r
,
1
M
d
n
a
)
x
e
t
r
o
c
y
r
o
s
n
e
s
-
o
t
a
m
o
s
y
r
a
d
n
o
c
e
s
(
2
S
,
1
S
,
1
V
e
r
a
)
4
9
9
1
,
1
9
9
1
,
9
8
9
1
,
8
8
9
1
(
l
w
o
d
a
w
S
y
b
t
i
b
b
a
r
e
h
t
n
i
)
8
0
0
2
(
n
e
s
r
e
t
e
P
&
t
e
l
u
o
P
;
)
6
0
0
2
(
n
e
s
r
e
t
e
P
&
t
e
h
c
o
r
C
)
4
9
9
1
,
1
9
9
1
,
9
8
9
1
,
8
8
9
1
(
w
o
d
a
w
S
l
)
5
0
0
2
(
.
l
a
t
e
l
e
m
i
e
H
)
9
0
0
2
(
n
n
a
m
k
a
S
&
k
c
o
K
e
d
)
9
0
0
2
(
s
i
r
r
a
H
&
a
t
a
k
a
S
)
8
0
0
2
(
.
l
a
t
e
g
r
e
b
n
e
e
r
G
)
6
0
0
2
(
.
l
a
t
e
e
e
L
5
.
0
--
1
<
4
7
.
0
--
--
--
--
--
--
3
-
2
4
-
3
7
-
4
7
.
1
4
.
1
--
--
5
3
.
0
--
--
--
1
6
.
0
4
4
.
0
6
3
.
0
3
.
0
1
<
1
<
9
1
.
0
1
S
1
V
1
M
1
S
1
A
e
s
u
o
M
t
a
R
t
a
R
t
a
R
t
a
R
s
a
e
r
a
4
t
i
b
b
a
R
1
V
l
e
r
r
i
u
q
S
H
-
E
7
B
7
.
g
F
i
.
g
F
i
9
.
0
±
6
5
.
0
9
0
.
1
1
.
6
±
1
1
9
5
.
7
1
.
1
±
8
.
4
5
4
.
4
8
.
0
±
1
1
.
1
6
8
.
0
e
c
n
e
r
e
f
e
r
m
o
d
n
a
r
l
e
d
o
M
l
e
d
o
M
]
z
H
[
s
e
t
a
r
g
n
i
r
i
F
e
c
r
u
o
s
a
t
a
d
e
6
L
e
5
L
e
4
L
e
3
/
2
L
a
e
r
A
s
e
i
c
e
p
S
24
e
h
t
o
t
r
e
f
e
r
s
t
l
u
s
e
r
l
e
d
o
m
e
h
T
i
.
s
n
a
d
e
m
e
h
t
e
r
a
l
e
r
r
i
u
q
s
y
a
r
g
e
h
t
r
o
f
a
t
a
d
e
h
T
.
4
L
n
i
d
e
t
a
c
o
l
o
s
l
a
y
l
t
r
a
p
e
r
a
t
a
h
t
s
l
l
e
c
l
a
t
c
e
t
-
o
c
i
t
r
o
c
o
t
s
d
n
o
p
s
e
r
r
o
c
.
s
t
u
p
n
i
l
a
n
r
e
t
x
e
g
n
i
y
r
a
v
r
o
f
s
e
t
a
r
n
o
i
t
a
u
p
o
p
l
e
h
t
f
o
d
t
s
o
t
d
n
a
n
o
i
t
a
z
i
r
t
e
m
a
r
a
p
e
c
n
e
r
e
f
e
r
±
n
a
e
m
e
h
t
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 8: Dependence of network activity on the external background firing rate
and the relative inhibitory synaptic strength. White stars mark the reference pa-
rameter set. (A) Population firing rates of excitatory populations in layers 2/3
(squares), 4 (diamonds), 5 (circles) and 6 (triangles), lightness increases with cor-
tical depth, as a function of the background rate with fixed relative inhibitory
synaptic strength (4). (B) AIness%, the percentage of populations with firing rate
below 30 Hz, irregularity between 0.7 and 1.2, and synchrony below 8 (data col-
lected for 5 s per simulation), as a function of the background rate and the relative
inhibitory synaptic strength. Labeled black iso-lines indicate areas where 50%,
75% and 100% of all populations fire asynchronously and irregularly at low rate.
Dashed iso-lines confine the area where the firing rates are ordered in accordance
to Table 5. (C) Population firing rates of excitatory populations as a function of
the relative inhibitory synaptic strength at fixed background rate (8 Hz) (markers
as in A).
25
246810relative inh. synaptic strength468101214background rate [Hz]50%50%75%75%100%0%50%100%AIness%0481216firing rates [Hz]01020firing rates [Hz]CBA Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 9: Relevance of target specificity for network stability. (A) Population firing
rates (top panel) and synchrony (bottom panel), both in logarithmic representation,
as a function of the target specificity of candidate projections L2/3e to L4 (negative
values twice as large), L2/3e to L6, L5e to L2/3 and L6e to L4. A target specificity
of 0 reflects random connectivity, gray-shaded area marks range of target specificity
of other inter-layer connections (0.33 ± 0.08). (B) Population firing rates of the
model with target specificity of candidate projections of +0.4 as a function of
the asymmetry of excitatory to inhibitory and excitatory to excitatory synaptic
strengths (defined as (wie − wee)/(wie + wee), with fixed mean excitatory synaptic
strength). Stars indicate firing rates of the reference model with target specificity
of candidate projections of −0.4.
(C) As B, but as a function of the relative
inhibitory synaptic strength. All markers as in Fig. 8A.
Role of i-specific projections for the stability of the AI state In the
following we conduct a series of simulation experiments where we alter the target
type selection of the i-specific projections. The modifications affect exclusively the
target specificity of these projections (upper four rows in Table 4). Technically,
we compile for each parameter set a new connectivity map algorithmically (see
Materials and Methods) so that the experimental data are fully respected. Fig. 9A
shows that the firing rates and the synchrony of the excitatory populations first
increase exponentially and then superexponentially when the target specificity of
the four candidate projections approach random connectivity and then the level of
other inter-layer connections, before the spike rates saturate for target specificity
values above 0.2. The increase in synchronization precedes the firing rate increases.
For L6e, we observe two outliers at T = 0.2 and 0.25 where the activity in this layer
corresponds to the low-rate AI state. This is likely due to the strong within-layer
inhibitory feedback in L6 and the high amount of random external inputs to L6e.
The general trend for L6 is nevertheless the same as for the other layers.
The modification of target specificity not only changes the local microcircuit
at the level of specific cell types but also the overall level of excitation in the net-
work. Therefore, we simulate further control networks which globally correct for
the change in the level of excitation: we increase for all connections, not only for
the i-specific candidates, the synaptic strength of excitatory to inhibitory connec-
tions and simultaneously reduce the synaptic strength of excitatory to excitatory
connections. Fig. 9B shows that a sufficiently large asymmetry of excitatory synap-
tic strengths with respect to target cell type counterbalances the overexcitation.
However, the natural order of the excitatory activity levels of the cortical layers
26
100101102firing rate [Hz]-0.4-0.20.00.20.4target specificity100101102synchrony0.00.040.080.12asymmetry of exc. synaptic strengths100101102firing rate [Hz]48121620relative inh. synaptic strength100101102firing rate [Hz]ABC Potjans & Diesmann, Connectivity explains cell-type specific activity
(Table 5) is partly inverted.
This stabilization procedure uses different excitatory synaptic strengths ac-
cording to the target cell type and thereby introduces an additional parameter.
Therefore, we also investigate whether the network can alternatively be stabilized
by changing an already existing parameter, the relative inhibitory synaptic strength.
We find that only implausibly large values (> 15) lead to a stable low-rate AI state
(Fig. 9C) and that also in this case the order of firing rates is partly inverted.
In summary, a layered network model that is parametrized equivalent to the
balanced random network model requires the inclusion of i-specific feedback con-
nections in order to exhibit asynchronous irregular spiking activity. The alternative
stabilization of the network by global changes of parameters results in a distribution
of layer-specific firing rates in conflict with experimental observations.
Propagation of transient thalamic inputs
Confronted with a transient thalamic input the layered network model responds
with a stereotypic propagation of activity through the different layers. Fig. 10A
shows an exemplary spike raster of the model after a short-lasting increase of
thalamic firing rate. The cell-type specific activity pattern is consistent over 100
different network and input instantiations (Fig. 10B).
Amplitude and timing of cell-type specific responses L2/3e, L4e and L5e
emit a comparable amount of additional spikes in response to the stimulus while
L6e shows a comparatively sparse response (Fig. 10B). In total, only a minority of
all neurons in the network is activated by the thalamic stimulus (Fig. 10E). Only in
L5e a large fraction of all cells emits additional spikes in response to the stimulus.
In relation to the ongoing activity, we find that the input layers (L4 and L6) exhibit
a similar dynamical gain which is much lower than the gain of the output layers
L5 and especially L2/3.
The response is initiated in the input layers and then propagates to layers L2/3
and L5 (see Fig. 10A, B). The latency of activation (defined as the maximum of
the excitatory spike count histogram) is shortest in L4 followed by L6 and L5, and
finally L2/3. The early onset of activation of L5, not after but rather synchronously
with L2/3, is in contrast to the expectation according to the classical notion of the
feed-forward loop from L4 to L2/3 to L5 (Gilbert, 1983) but in agreement with the
experimental activity data (Sakata & Harris, 2009, compare also Mitzdorf, 1985).
The feed-forward connection from L2/3 to L5 causes the prolonged response in L5
that is reflected in the second peak in the spike count. L6, being already slightly
excited by the thalamic input, receives feed-forward input from L5 which triggers
a sparse response during the ramp-up phase in L5. The activity is back to baseline
in all layers before the last thalamic spikes arrive at cortex. Deactivation is ordered
similar to activation, starting in L4, followed by L2/3 and L6, and finally L5.
Interplay of excitation and inhibition in the propagation of inputs Not
only the excitatory populations but also the inhibitory populations show a distinct
activation pattern (see Fig. 10A, gray dots, and B, gray lines).
Initially, as the
interneurons in any layer receive the same, albeit slightly weaker, feed-forward
inputs, the inhibitory response resembles the excitatory response. However, the
inhibitory populations show, in contrast to the respective excitatory populations,
two pronounced activation peaks in L2/3, L4 and L6. As shown above, these
27
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 10: Response to transient thalamic input. Thalamic firing rates increase
step-like by 15 Hz for a duration of 10 ms; 0 ms corresponds to the onset of tran-
sient input, gray bars show the arrival of thalamic spikes at cortical neurons taking
the mean delay (for simplicity the same as within the network, 1.5 ms) into ac-
count. (A) Cell-type specific spiking activity of the network with i-specific projec-
tions. Markers as in Fig. 7A. (B) Corresponding cell-type specific population spike
counts averaged over 100 instantiations of network and input (excitatory popula-
tions in black, inhibitory in gray). Spike counts are calculated with a bin-width
of 0.5 ms with the number of recorded neurons L3e: 500, L3i: 141, L4e: 529,
L4i: 132, L5e: 117, L5i: 25, L6e: 347, L6i: 71 (numbers correspond to rela-
tive population sizes). (C) Spiking activity of a control network without i-specific
projections (target specificity of candidate projections of +0.4) stabilized by an
asymmetry of excitatory synaptic strengths of 0.095. (D) Corresponding cell-type
specific population spike counts of excitatory populations averaged over 100 in-
stantiations. Filled area shows the difference of the data in D and B. (E) Average
percentage of cells in the excitatory populations that are activated by thalamic
stimulation in relation to the ongoing activity (according to the data in B). (F)
Dynamical gain, defined as the firing rate during the stimulus presentation divided
by the spontaneous firing rate, of excitatory populations (according to data in B).
28
5 spikesL2/3L4L5L601020time [ms]10 spikesL2/3eL4eL5eL6eBDEF020406080100% activated cellsL2/3eL4eL5eL6e0510152025dynamical gainL6L5L4L2/3A010time [ms]L6L5L4L2/3C Potjans & Diesmann, Connectivity explains cell-type specific activity
layers receive i-specific feedback connections. Apparently, the cell-type specific
connectivity structure results in a complex interplay of excitation and inhibition not
explained by within-layer recurrent inhibition and thereby shapes the propagating
response to a transient thalamic input. Specifically we observe that the activation
of L2/3 entails the increased activity of inhibitory interneurons in L4, thereby
stopping the excitatory activity in this layer. A similar effect is observed in the next
step of the feed-forward loop between L2/3 and L5. Activation of L2/3 prolongs
the excitatory activity in L5 as stated above.
In contrast, the early activation of
L5 results in a sharpening of the response in L2/3 by the additional activation of
L2/3 interneurons. The connection from L6e to L4i adds to the sharpening of the
L4 response and the rather weak i-specific connection from L2/3 to L6 plays a role
in preventing a stronger ramp-up of excitatory activity in L6.
Role of i-specific feedback connections for the propagation of inputs
These observations suggest that the i-specific feedback connections control the
duration and the amplitude of the response to thalamic inputs. To further eluci-
date their role for the propagation of input-related activity, we study the response
of a network without i-specific feedback connections but similar spontaneous ac-
tivity. The small differences in the firing rates (compare Fig. 9B) are reflected in
the initial response of the network to the thalamic input (first bumps in the spike
count histograms in Fig. 11D) which exhibits a steeper increase in L2/3e and L4e
and is weaker in L5e. After this initial phase, the response is drastically different
and shows reverberating activation of the different layers lingering well beyond the
offset of the input. The response is much stronger and shows oscillatory compo-
nents by reciprocal activation of the different layers. The effect stays the same,
albeit smaller, for a control network with the candidate projections having a target
specificity value of +0.2 which is well below the range for non-specific inter-layer
connections.
Discussion
Modeling approach
The present work extends the balanced random network model (Amit & Brunel,
1997; Brunel, 2000a; van Vreeswijk & Sompolinsky, 1996, 1998) to multiple layers
with realistic connection probabilities. Despite their reduced structure the mono-
layered models exhibit qualitatively consistent activity dynamics and the classical
analysis of them has guided our research. Our neuron and synapse model as well as
the random connectivity scheme do not differ qualitatively from the earlier work.
The model size is selected sufficiently large to incorporate the majority of all local
synapses. The network structure, one excitatory and one inhibitory population in
each layer, represents the minimal laminar extension of the mono-layered models
and also the minimal set of cell types typically distinguished in experiments. Data
resolving the connectivity at a finer scale (e.g. Mercer et al., 2005; West et al.,
2006) are combined to match the more coarse resolution of our model. Previous
multi-layered models partly use the same approach (e.g. Haeusler et al., 2009) and
partly more detailed cell type classifications (e.g. Traub et al., 2005; Izhikevich &
Edelman, 2008). However, quantitative connectivity data are not yet widely and
consistently available on a finer level of detail so that it remains unclear which
dynamical consequences any further separation of cell types implies.
29
Potjans & Diesmann, Connectivity explains cell-type specific activity
Integrated connectivity map
The success of the dynamical analysis in the second part of the study relies on
our finding that the two connectivity maps are consistent when one considers the
differences in methodology. The compiled connectivity map accounts for these,
but it nevertheless mingles data not only from multiple laboratories but also from
different cortical areas and species. This choice is not exclusively motivated by
canonicity in structure (e.g. Douglas et al., 1989; Nelson, 2002; Groh et al., 2010)
and activity (Table 5), but driven by the incompleteness of data for a specific area
and species. The density of neurons in our model is based on cat area 17 and
the density of synapses is consistent with cat data (Beaulieu & Colonnier, 1985).
The most crucial difference between our map and the original data based maps
is the target specific structure which removes the fundamental bias introduced by
undersampling of inhibitory cells (electrophysiology) or by the application of Peters'
rule (anatomy).
The integrated connectivity map is consistent with the most prominent features
of the cortical microcircuit: the recurrence of connections (Douglas & Martin,
1991) and the feed-forward loop from L4 to L2/3 to L5 to L6 to L4 (Gilbert,
1983). The excitatory sub-circuit is largely consistent with the recently published
excitatory map of the C2 barrel column of the mouse (Lefort et al., 2009).
In
addition, the circuit exhibits a distinct feedback structure with projections targeting
predominantly interneurons from L2/3e to L4 and from L5e to L2/3. Thomson
et al. (2002); Thomson & Morris (2002) recognized the potential of selective
feedback projections for stabilizing the activity and increasing the sensitivity for
time-dependent signaling, but were not able to test this hypothesis. Furthermore,
reports on the projections from L6e to L4 and the rather weak projection from
L2/3e to L6 indicate specific selection of inhibitory target cells. Our selection of
the projection from L6e to L4 is based on the electron microscopy study by McGuire
et al. (1984), providing sparse and partial data. In their discussion, Ahmed et al.
(1994) present an alternative interpretation of these data proposing that many
synapses originally assigned to inhibitory targets are potentially on excitatory cells.
Still they acknowledge that the relative number of synapses targeting interneurons
is very high in comparison to other excitatory projections. The ongoing discussion
on this projection has recently been reviewed by Thomson (2010).
Here, we quantified the dynamical relevance of these i-specific connections for
spontaneous and evoked activity. Especially the two feedback projections (L2/3e to
L4 and L5e to L2/3) are crucial for stability and reliable input propagation because
of the relatively small structural and dynamical
impact of L6 on other layers.
Extending the idea of i-specific feedback, it is a conceivable but untested hypothesis
that L6e (especially cortico-thalamic cells) targets predominantly interneurons in
L5, too.
Model parametrization
Only a subset of the required model parameters are known experimentally. Here,
we focus on cell-type specific structural parameters:
the numbers of neurons,
the numbers of external inputs and the connection probabilities between neurons.
Other parameters like neuronal parameters, synaptic strengths and delays are se-
lected independent of the cell type. This choice enables us to expose the dynamical
consequences of the structure of the local microcircuit unaffected by additional cell-
type specificity (e.g. Br´emaud et al., 2007; Lefort et al., 2009). The consistency
30
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 11: Sketch of the flow of activity following transient thalamic input. Black
and gray ellipses represent excited and inhibited activity states, respectively. Dark
gray dashed arrows indicate input and output of the local network. The black
arrows represent the feed-forward loop projections L4 to L2/3 to L5 to L6. The
gray arrows correspond to the activation of the candidate i-specific connections
(L2/3e to L4 and to L6, L5e to L2/3 and L6e to L4).
of the activity in our model with the experimental data suggests that the (static)
connectivity structure plays a dominant role in shaping the neuronal activity and
that it is not required to model complex neuronal features such as morphology to
reproduce the experimental findings discussed here.
In future, modeling of more
complex neuronal dynamics will eventually reveal where spatially extended neuron
models are required to understand the experimentally observed phenomena.
An exception to the cell-type independent parametrization is the increased
synaptic strength for the connection from L4e to L2/3e. Although this change
affects the ongoing activity only marginally (supplementary Fig. 19), it is impor-
tant for a successful transmission of activity from L4 to L2/3 following thalamic
stimulation. The modification is motivated by the large discrepancies of the L4i
to L2/3 projection in the anatomical and physiological maps and the difference in
the relative convergence of excitatory inputs from L2/3 and L4 to L2/3 pyramids
between our model and Feldmeyer et al. (2006). To ultimately resolve this issue,
it might be necessary to incorporate additional specificity (Yoshimura et al., 2005;
Sarid et al., 2007; Fares & Stepanyants, 2009) and, despite a plethora of studies on
the L4 to L2/3 connections, potentially additional experiments especially regarding
the inhibitory projection.
Cell-type specific activity
The application of the integrated connectivity map and the parametrization accord-
ing to balanced random network models (Brunel, 2000a) results in asynchronous
irregular activity without specific tuning. This activity state is stable over a wide
range of parameters regarding e.g. external inputs, synaptic strengths and delays.
The model predicts cell-type specific firing rates in agreement with data from
awake animals (Table 5).
In particular, the connectivity map captures the low
excitatory firing rates in L2/3 and L6 quantitatively although mono-layered models
hardly show stable activity at these low levels of activity (Sussillo et al., 2007). The
particular ordering of firing rates depends, however, on the inclusion of i-specific
31
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 12: Differences in input structure for the excitatory cell types L2/3e (A),
L4e (B), L5e (C) and L6e (D) (large triangles). The illustrations show the strongest
pathways of direct (first-order, thick arrows) and indirect (second-order, thin ar-
rows) excitation (black) and inhibition (gray) of a given population (see Supple-
mental Material for further details). Triangles represent excitatory and circles
inhibitory populations.
feedback connections. The lacking consistency of the firing rates in previous models
is presumably due to the incompleteness of the previously used connectivity maps.
Confronted with a transient thalamic input, the model shows a concise prop-
agation of activity from the input layers to the output layers (compare Miller,
1996). The propagation pattern (Fig. 11) is shaped by the interaction of excita-
tion and inhibition between the different layers and promotes a temporal neural
code (Thomson & Morris, 2002).
Relation of structure and activity
The differences in the cell-type specific input structure (Fig. 12) shed light on the
mechanisms underlying the observed activity features. For example, the low firing
rates of L2/3e and L6e neurons have different structural origins: L2/3e effectively
integrates, next to the excitatory inputs from L2/3e and L4e, inhibitory inputs from
all layers. In contrast, L6e interacts largely with the recurrent within-L6 network;
other inputs predominantly pass through L5e and rather modulate L6e activity. L4e
is also dominated by within-layer connectivity and the i-specific inputs from L2/3
and L6 modulate its activity and temporally structure the response to transient
stimuli. L5 consists of relatively few neurons and correspondingly forms rather
few recurrent within-layer inputs. Furthermore, L5e integrates the highest number
of first-order inputs, but does not receive i-specific projections. As a result, L5e
reacts with a wide range of firing rates to changes in the external input, especially
in comparison to the other layers. This, together with the control that L5 exerts
on L2/3 by the i-specific feedback, puts it in a special position to integrate and
amplify information.
Inhibitory cells receive predominantly excitatory inter-layer
inputs (supplementary Fig. 15) and therefore exhibit elevated firing rates.
Conclusion
The connectivity structure of the local cortical network shapes the cell-type specific
activity and defines functional roles of the different layers. Any fine-scale connec-
tivity structure (e.g. Song et al., 2005; Yoshimura et al., 2005; Kampa et al., 2006)
32
Potjans & Diesmann, Connectivity explains cell-type specific activity
faces the constraints imposed by the connectivity implemented in this model. The
presented framework can be continuously refined as new data become available
and extends the available mathematical methods to infer synaptic connectivity
from neuronal morphology. Based on the currently available data it reproduces
prominent activity features, suggesting that these arise predominantly from the
network structure, not single cell properties. The results predict distinct activity
patterns of interneurons and highlight the need to uncover the target specificity of
projections in future experiments.
Acknowledgements: We are grateful to Kathleen S. Rockland and Dirk Feld-
meyer for fruitful discussions about the structure of the local cortical network,
and to our colleagues in the NEST Initiative for continued support. We also
acknowledge Ad Aertsen for his contributions in the initial phase of the project.
Partially supported by the Helmholtz Alliance on Systems Biology, EU Grant 15879
(FACETS), EU Grant 269921 (BrainScaleS), DIP F1.2, BMBF Grant 01GQ0420
to BCCN Freiburg, and the Next-Generation Supercomputer Project of MEXT
(Japan).
33
Potjans & Diesmann, Connectivity explains cell-type specific activity
References
Ahmed, B., Anderson, J., Douglas, R., Martin, K., & Nelson, J. (1994). Polyneu-
innervation of spiny stellate neurons in cat visual cortex. J. Compar.
ronal
Neurol. 341 (1), 39 -- 49.
Ali, A. B., Bannister, A. P., & Thomson, A. M. (2007). Robust correlations
between action potential duration and the properties of synaptic connections in
layer4 interneurones in neocortical slices from juvenile rats and adult rat and
cat. J. Physiol. (Lond.) 580 (1), 149 -- 169.
Amit, D. J., & Brunel, N. (1997). Model of global spontaneous activity and local
structured activity during delay periods in the cerebral cortex. Cereb. Cortex 7,
237 -- 252.
Bannister, A. P., & Thomson, A. M. (2007). Dynamic properties of excitatory
synaptic connections involving layer 4 pyramidal cells in adult rat and cat neo-
cortex. Cereb. Cortex 17 (9), 2190 -- 2203.
Barbour, D. L., & Callaway, E. M. (2008). Excitatory local connections of super-
ficial neurons in rat auditory cortex. J. Neurosci. 28 (44), 11174 -- 11185.
Beaulieu, C., & Colonnier, M. (1983). The number of neurons in the different
laminae of the binocular and monocular regions of area 17 in the cat. J. Compar.
Neurol. 217, 337 -- 344.
Beaulieu, C., & Colonnier, M. (1985). A laminar analysis of the number of round-
asymmetrical and flat-symmetrical synapses on spines, dendritic trunks, and cell
bodies in area 17 of the cat. J. Compar. Neurol. 231, 180 -- 189.
Beierlein, M., & Connors, B. W. (2002). Short-term dynamics of thalamocortical
and intracortical synapses onto layer 6 neurons in neocortex. J. Neurophysiol. 88,
1924 -- 1932.
Beierlein, M., Gibson, J. R., & Connors, B. W. (2003). Two dynamically distinct
inhibitory networks in layer 4 of the neocortex. J. Neurophysiol. 90, 2987 -- 3000.
Binzegger, T., Douglas, R. J., & Martin, K. A. C. (2004). A quantitative map of
the circuit of cat primary visual cortex. J. Neurosci. 39 (24), 8441 -- 8453.
Binzegger, T., Douglas, R. J., & Martin, K. A. C. (2007). Stereotypical bouton
clustering of individual neurons in cat primary visual cortex. J Neurosci 27 (45),
12242 -- 12254.
Binzegger, T., Douglas, R. J., & Martin, K. A. C. (2009). Topology and dynamics
of the canonical circuit of cat V1. Neural Networks 22 (8), 1071 -- 1078.
Braitenberg, V., & Schuz, A. (1998). Cortex: Statistics and Geometry of Neuronal
Connectivity (2nd ed.). Berlin: Springer-Verlag.
Br´emaud, A., West, D. C., & Thomson, A. M. (2007). Binomial parameters differ
across neocortical layers and with different classes of connections in adult rat
and cat neocortex. Proc. Nat. Acad. Sci. USA 104 (35), 14134 -- 14139.
Briggs, F. (2010). Organizing principles of cortical layer 6. Frontiers in Neural
Circuits 4 (3). doi: 10.3389/neuro.04.003.2010.
34
Potjans & Diesmann, Connectivity explains cell-type specific activity
Brunel, N. (2000a). Dynamics of sparsely connected networks of excitatory and
inhibitory spiking neurons. J. Comput. Neurosci. 8 (3), 183 -- 208.
Brunel, N. (2000b). Dynamics of sparsely connected networks of excitatory and
inhibitory spiking neurons. J. Comput. Neurosci. 8 (3), 183 -- 208.
Buzas, P., Kovacs, K., Ferecsko, A. S. Budd, J. M. L., & Eysel, U. T.Kisvarday,
Z. F. (2006). Model-based analysis of excitatory lateral connections in the visual
cortex. J. Compar. Neurol. 499, 861 -- 881.
Crochet, S., & Petersen, C. C. H. (2006). Correlating whisker behavior with
membrane potential in barrel cortex of awake mice. Nat. Neurosci. 9 (5), 608 --
610.
Dantzker, J. L., & Callaway, E. M. (2000). Laminar sources of synaptic input
to cortical inhibitory interneurons and pyramidal neurons. Nat. Neurosci. 3 (7),
701 -- 707.
Davis, C. S. (1993). The computer generation of multinomial random variates.
Computational Statistics & Data Analysis 16, 205 -- 217.
de Kock, C. P. J., & Sakmann, B. (2009). Spiking in primary somatosensory cortex
during natural whisking in awake head-restrained rats is cell-type specific. Proc.
Natl. Acad. Sci. USA 106 (38), 16446 -- 16450.
Djurfeldt, M., Hjorth, J., Eppler, J. M., Dudani, N., Helias, M., Potjans, T. C.,
Bhalla, U. S., Diesmann, M., Hellgren Kotaleski, J., & Ekeberg, O. (2010). Run-
time interoperability between neuronal network simulators based on the MUSIC
framework. Neuroinformatics 8, 43 -- 60.
Djurfeldt, M., Lundqvist, M., Johansson, C., Rehn, M., Ekeberg, O., & Lansner,
A. (2008). Brain-scale simulation of the neocortex on the IBM Blue Gene/L
supercomputer. IBM Journal of Research and Development 52 (1/2), 31 -- 41.
Douglas, R., & Martin, K. (2007a). Mapping the matrix: the ways to neocortex.
Neuron 56, 226 -- 238.
Douglas, R. J., Martin, K. A., & Whitteridge, D. (1989). A canonical microcircuit
for neocortex. Neural Comput. 1, 480 -- 488.
Douglas, R. J., & Martin, K. A. C. (1991). A functional microcircuit for cat visual
cortex. J. Physiol. (Lond) 440, 735 -- 769.
Douglas, R. J., & Martin, K. A. C. (2004). Neuronal circuits of the neocortex.
Annu. Rev. Neurosci. 27, 419 -- 451.
Douglas, R. J., & Martin, K. A. C. (2007b). The butterfly and the loom. Brain
Res. Bull. 55, 314 -- 328.
Elhanany, E., & White, E. L. (1990).
Intrinsic circuitry: synapses involving the
local axon collaterals of corticocortical projection neurons in the mouse primary
somatosensory cortex. J. Compar. Neurol. 291, 43 -- 54.
Eppler, J. M., Helias, M., Muller, E., Diesmann, M., & Gewaltig, M. (2009).
PyNEST: a convenient interface to the NEST simulator. Front. Neuroinform. 2,
12.
35
Potjans & Diesmann, Connectivity explains cell-type specific activity
Fares, T., & Stepanyants, A. (2009). Cooperative synapse formation in the neo-
cortex. Proc. Natl. Acad. Sci. USA 106 (38), 16463 -- 16468.
Feldmeyer, D., Lubke, J., & Sakmann, B. (2006). Efficacy and connectivity of
intracolumnar pairs of layer 2/3 pyramidal cells in the barrel cortex of juvenile
rats. J. Physiol. (Lond.) 575, 583 -- 602.
Feldmeyer, D., Lubke, J., Silver, R. A., & Sakmann, B. (2002). Synaptic connec-
tions between layer 4 spiny neurone-layer 2/3 pyramidal cell pairs in juvenile rat
barrel cortex: physiology and anatomy of interlaminar signalling within a cortical
column. J. Physiol. (Lond.) 538 (3), 803 -- 822.
Feldmeyer, D., Roth, A., & Sakmann, B. (2005). Monosynaptic connections be-
tween pairs of spiny stellate cells in layer 4 and pyramidal cells in layer 5a indicate
that lemniscal and paralemniscal afferent pathways converge in the infragranular
somatosensory cortex. J. Neurosci. 25 (13), 3423 -- 3431.
Fetz, E., Toyama, K., & Smith, W. (1991). Synaptic interactions between cortical
In A. Peters (Ed.), Cerebral Cortex, Volume 9, Chapter 1, pp. 1 -- 47.
neurons.
New York and London: Plenum Press.
Gabbott, P. L., & Somogyi, P. (1986). Quantitative distribution of GABA-
immunoreactive neurons in the visual cortex (area 17) of the cat. Exp. Brain
Res. 61 (2), 323 -- 331.
Gewaltig, M.-O., & Diesmann, M. (2007). NEST (Neural Simulation Tool). Schol-
arpedia 2 (4), 1430.
Gilbert, C. D. (1983). Microcircuitry of the visual cortex. Annu. Rev. Neurosci. 6,
217 -- 247.
Gilbert, C. D., & Wiesel, T. N. (1983). Clustered intrinsic connections in cat visual
cortex. J. Neurosci. 5, 1116 -- 33.
Golomb, D. (2007). Neuronal synchrony measures. Scholarpedia 2 (1), 1347.
Gonchar, Y., & Burkhalter, A. (2003). Distinct GABAergic targets of feedforward
and feedback connections between lower and higher areas of rat visual cortex.
J. Neurosci. 23 (34), 10904 -- 10912.
Greenberg, D. S., Houweling, A. R., & Kerr, J. N. D. (2008). Population imaging of
ongoing neuronal activity in the visual cortex of awake rats. Nat. Neurosci. 11 (7),
749 -- 751.
Groh, A., Meyer, H. S., Schmidt, E. F., Heintz, N., Sakmann, B., & Krieger, P.
(2010). Cell-type specific properties of pyramidal neurons in neocortex underlying
a layout that is modifiable depending on the cortical area. Cereb. Cortex 20 (4),
826 -- 836.
Haeusler, S., & Maass, W. (2007). A statistical analysis of information-processing
properties of lamina-specific cortical microcircuit models. Cereb Cortex 17 (1),
149 -- 162.
Haeusler, S., Schuch, K., & Maass, W. (2009). Motif distribution, dynamical prop-
erties, and computational performance of two data-based cortical microcircuit
templates. J. Physiol. (Paris) 103 (1 -- 2), 73 -- 87.
36
Potjans & Diesmann, Connectivity explains cell-type specific activity
Heimel, J. A., Van Hooser, S. D., & Nelson, S. B. (2005). Laminar organiza-
tion of response properties in primary visual cortex of the gray squirrel (sciurus
carolinensis). J. Neurophysiol. 94, 3538 -- 3554.
Heinzle, J., Hepp, K., & Martin, K. A. C. (2007). A microcircuit model of the
frontal eye fields. J. Neurosci. 27 (35), 9341 -- 9353.
Hellwig, B. (2000). A quantitative analysis of the local connectivity between pyra-
midal neurons in layers 2/3 of the rat visual cortex. Biol. Cybern. 2 (82), 111 -- 121.
Hill, S., & Tononi, G. (2005). Modeling sleep and wakefulness in the thalamocor-
tical system. J Neurophysiol 93 (3), 1671 -- 1698. Comparative Study.
Holmgren, C., Harkany, T., Svennenfors, B., & Zilberter, Y. (2003). Pyramidal cell
communication within local networks in layer 2/3 of rat neocortex. J. Physiol.
(Lond.) 1 (551), 139 -- 153.
Izhikevich, E. M. (2006). Polychronization: Computation with spikes. Neural
Comput. 18, 245 -- 282.
Izhikevich, E. M., & Edelman, G. M. (2008). Large-scale model of mammalian
thalamocortical systems. Proc. Natl. Acad. Sci. USA 105 (9), 3593 -- 3598.
Izhikevich, E. M., Gally, J. A., & Edelman, G. M. (2004). Spike-timing dynamics
of neuronal groups. Cereb. Cortex 14, 933 -- 944.
Johnson, R. R., & Burkhalter, A. (1996). Microcircuitry of forward and feedback
connections within rat visual cortex. J. Compar. Neurol. 368, 383 -- 398.
Kampa, B. M., Letzkus, J. J., & Stuart, G. J. (2006). Cortical feed-forward net-
works for binding different streams of sensory information. Nat. Neurosci. 9 (12),
1472 -- 1473.
Kapfer, C., Glickfeld, L., Atallah, B., & Scanziani, M. (2007). Supralinear increase
of recurrent inhibition during sparse activity in the somatosensory cortex. Nat.
Neurosci. 10 (6), 743 -- 753.
Kriener, B., Tetzlaff, T., Aertsen, A., Diesmann, M., & Rotter, S. (2008). Cor-
relations and population dynamics in cortical networks. Neural Comput. 20,
2185 -- 2226.
Kumar, A., Schrader, S., Aertsen, A., & Rotter, S. (2008). The high-conductance
state of cortical networks. Neural Comput. 20 (1), 1 -- 43.
Le Be, J.-V., & Markram, H. (2006). Spontaneous and evoked synaptic rewiring
in the neonatal neocortex. Proc. Natl. Acad. Sci. USA 103 (35), 13214 -- 13219.
Lee, A., Manns, I., Sakmann, B., & Brecht, M. (2006). Whole-cell recordings in
freely moving rats. Neuron 51, 399 -- 407.
Lefort, S., Tomm, C., Sarria, J.-C. F., & Petersen, C. C. H. (2009). The excita-
tory neuronal network of the C2 barrel column in mouse primary somatosensory
cortex. Neuron 61, 301 -- 316.
37
Potjans & Diesmann, Connectivity explains cell-type specific activity
Lundqvist, M., Rehn, M., Djurfeldt, M., & Lansner, A. (2006). Attractor dy-
namics in a modular network model of neocortex. Network: Comput. Neural
Systems 17 (3), 253 -- 276.
Maffei, A., Nelson, S. B., & Turrigiano, G. G. (2004). Selective reconfiguration of
layer 4 visual circuitry by visual deprivation. Nat. Neurosci. 7 (12), 1353 -- 1359.
Markram, H., Lubke, J., Frotscher, M., Roth, A., & Sakmann, B. (1997). Phys-
iology and anatomy of synaptic connections between thick tufted pyramidal
neurons in the developing rat neocortex. J. Physiol. (Lond) 500 (2), 409 -- 440.
Martin, K. A. C., & Whitteridge, D. (1984). Form, function, and intracortical
projections of spiny neurons in the striate visual cortex of the cat. J. Physiol.
(Lond.) 353, 463 -- 504.
Mason, A., Nicoll, A., & Stratford, K. (1991). Synaptic transmission between
individual pyramidal neurons of the rat visual cortex in vitro. J. Neurosci. 11 (1),
72 -- 84.
McGuire, B. A., Hornung, J.-P., Gilbert, C. D., & Wiesel, T. N. (1984). Patterns
of synaptic input to layer 4 of cat striate cortex. J. Neurosci. 4 (12), 3021 -- 3033.
Mercer, A., West, D. C., Morris, O. T., Kirchhecker, S., Kerkhoff, J. E., & Thom-
son, A. M. (2005). Excitatory connections made by presynaptic cortico-cortical
pyramidal cells in layer 6 of the neocortex. Cereb. Cortex 15, 1485 -- 1496.
Miller, R. (1996). Neural assemblies and laminar interactions in the cerebral cortex.
Biol. Cybern. 75 (3), 253 -- 261.
Mitzdorf, U. (1985). Current source-density method and application in cat cere-
Investigation of evoked potentials and eeg phenomena. Physiol.
bral cortex:
Rev. 65 (1), 37 -- 100.
Morrison, A., Aertsen, A., & Diesmann, M. (2007). Spike-timing dependent plas-
ticity in balanced random networks. Neural Comput. 19, 1437 -- 1467.
Morrison, A., Mehring, C., Geisel, T., Aertsen, A., & Diesmann, M. (2005). Ad-
vancing the boundaries of high connectivity network simulation with distributed
computing. Neural Comput. 17 (8), 1776 -- 1801.
Nelson, S. B. (2002). Cortical microcircuits: Diverse or canonical? Neuron 36,
19 -- 27.
Nordlie, E., Gewaltig, M.-O., & Plesser, H. E. (2009). Towards reproducible de-
scriptions of neuronal network models. PLoS Comput Biol 5 (8), e1000456.
Peters, A., & Payne, B. R. (1993). Numerical relationships between geniculocor-
tical afferents and pyramidal cell modules in cat primary visual cortex. Cereb.
Cortex 3 (1), 69 -- 78.
Poulet, J., & Petersen, C. (2008). Internal brain state regulates membrane poten-
tial synchrony in barrel cortex of behaving mice. Nature 454 (7206), 881 -- 5.
Ren, M., Yoshimura, Y., Takada, N., Horibe, S., & Komatsu, Y. (2007). Special-
ized inhibitory synaptic actions between nearby neocortical pyramidal neurons.
Science 316, 758 -- 761.
38
Potjans & Diesmann, Connectivity explains cell-type specific activity
Sakata, S., & Harris, K. D. (2009). Laminar structure of spontaneous and sensory-
evoked population activity in auditory cortex. Neuron 64, 404 -- 418.
Sarid, L., Bruno, R., Sakmann, B., Segev, I., & Feldmeyer, D. (2007). Modeling a
layer 4-to-layer 2/3 module of a single column in rat neocortex: Interweaving in
vitro and in vivo experimental observations. Proc. Natl. Acad. Sci. USA 104 (41),
16353 -- 16358.
Schubert, D., Kotter, R., Luhmann, H. J., & Staiger, J. F. (2006). Morphology,
electrophysiology and functional input connectivity of pyramidal neurons charac-
terizes a genuine layer va in the primary somatosensory cortex. Cereb. Cortex 16,
223 -- 236.
Silberberg, G., & Markram, H. (2007). Disynaptic inhibition between neocortical
pyramidal cells mediated by martinotti cells. Neuron 53, 735 -- 746.
Song, S., Per, S., Reigl, M., Nelson, S., & Chklovskii, D. (2005). Highly nonrandom
features of synaptic connectivity in local cortical circuits. Public Library of
Science, Biology 3 (3), 0507 -- 0519.
Stepanyants, A., Hirsch, J., Martinez, L. M., Kisvarday, Z. F., Ferecsko, A. S.,
& Chklovskii, D. B. (2008). Local potential connectivity in cat primary visual
cortex. Cereb. Cortex 18 (1), 13 -- 28.
Stepanyants, A., Martinez, L. M., Ferecsk´o, A. S., & Kisv´arday, Z. F. (2009). The
fractions of short- and long-range connections in the visual cortex. Proc. Nat.
Acad. Sci. USA 106 (9), 3555 -- 3560.
Sussillo, D., Toyoizumi, T., & Maass, W. (2007). Self-tuning of neural circuits
through short-term synaptic plasticity. J. Neurophysiol. 97, 4079 -- 4095.
Swadlow, H. A. (1988). Efferent neurons and suspected interneurons in binocular
visual cortex of the awake rabbit: Receptive fields and binocular properties. J.
Neurophysiol. 59 (4), 1162 -- 1187.
Swadlow, H. A. (1989). Efferent neurons and suspected interneurons in s-1 vib-
rissa cortex of the awake rabbit: Receptive fields and axonal properties. J.
Neurophysiol. 62 (1), 288 -- 308.
Swadlow, H. A. (1991). Efferent neurons and suspected interneurons in second
somatosensory cortex of the awake rabbit: Receptive fields and axonal properties.
J. Neurophysiol. 66 (4), 1392 -- 1409.
Swadlow, H. A. (1994). Efferent neurons and suspected interneurons in motor
cortex of the awake rabbit: Axonal properties, sensory receptive fields, and
subthreshold synaptic inputs. J. Neurophysiol. 71 (2), 437 -- 453.
Thomson, A. M. (1997). Activity-dependent properties of synaptic transmission
at two classes of connections made by rat neocortical pyramidal axons in vitro.
J. Physiol. (Lond.) 502 (1), 131 -- 147.
Thomson, A. M. (2010). Neocortical
layer 6, a review.
Frontiers in Neu-
roanatomy 4 (13). doi: 10.3389/fnana.2010.00013.
39
Potjans & Diesmann, Connectivity explains cell-type specific activity
Thomson, A. M., & Bannister, A. P. (1998). Postsynaptic pyramidal target se-
lection by descending layer iii pyramidal axons: dual intracellular recordings and
biocytin filling in slices of rat neocortex. Neuroscience 84 (3), 669 -- 683.
Thomson, A. M., Bannister, A. P., Mercer, A., & Morris, O. T. (2002). Target and
temporal pattern selection at neocortical synapses. Philosophical Transactions
of the Royal Society London, Series B 357, 1781 -- 1791.
Thomson, A. M., & Deuchars, J. (1997). Synaptic interactions in neocortical local
circuits: Dual intracellular recordings in vitro. Cereb. Cortex 7, 510 -- 522.
Thomson, A. M., & Lamy, C. (2007). Functional maps of neocortical local circuitry.
Frontiers in Neuroscience 1 (1), 19 -- 42.
Thomson, A. M., & Morris, O. T. (2002). Selectivity in the inter-laminar connec-
tions made by neocortical neurones. J. Neurocytol. 31, 239 -- 246.
Thomson, A. M., West, D. C., & Deuchars, J. (1995). Properties of single axon
excitatory postsynaptic potentials elicited in spiny interneurons by action po-
tentials in pyramidal neurons in slices of rat neocortex. Neuroscience 69 (3),
727 -- 738.
Thomson, A. M., West, D. C., Hahn, J., & Deuchars, J. (1996). Single axon
ipsps elicited in pyramidal cells by three classes of interneurones in slices of rat
neocortex. J. Physiol. (Lond.) 496 (1), 81 -- 102.
Thomson, A. M., West, D. C., Wang, Y., & Bannister, A. (2002). Synaptic
connections and small circuits involving excitatory and inhibitory neurons in layer
2-5 of adult rat and cat neocortex: Triple intracellular recordings and biocytin
labelling in vitro. Cereb. Cortex 12, 936 -- 953.
Traub, R. D., Contreras, D., Cunningham, M. O., Murray, H., LeBeau, F. E. N.,
Roopun, A., Bibbig, A., Wilent, W. B., Higley, M. J., & Whittington, M. A.
(2005). Single-column thalamocortical network model exhibiting gamma oscilla-
tions, sleep spindles, and epileptogenic bursts. J Neurophysiol 93 (4), 2194 -- 2232.
Comparative Study.
Tukey, J. W. (1977). Exploratory data analysis. Addison-Wesley.
van Vreeswijk, C., & Sompolinsky, H. (1996). Chaos in neuronal networks with
balanced excitatory and inhibitory activity. Science 274, 1724 -- 1726.
van Vreeswijk, C., & Sompolinsky, H. (1998). Chaotic balanced state in a model
of cortical circuits. Neural Comput. 10, 1321 -- 1371.
Vogels, T. P., & Abbott, L. F. (2005). Signal propagation and logic gating in
networks of integrate-and-fire neurons. J. Neurosci. 25 (46), 10786 -- 10795.
Vogels, T. P., Rajan, K., & Abbott, L. F. (2005). Neural network dynamics. Annu.
Rev. Neurosci. 28, 357 -- 376.
Wang, Y., Markram, H., Goodman, P. H., Berger, T. K., Ma, J., & Goldman-Rakic,
P. S. (2006). Heterogeneity in the pyramidal network of the medial prefrontal
cortex. Nat. Neurosci. 9 (4), 534 -- 542.
40
Potjans & Diesmann, Connectivity explains cell-type specific activity
West, D. C., Mercer, A., Kirchhecker, S., Morris, O. T., & Thomson, A. M. (2006).
Layer 6 cortico-thalamic pyramidal cells preferentially innervate interneurons and
generate facilitating epsps. Cereb. Cortex 16, 200 -- 211.
White, E. L., & Keller, A. (1987).
Intrinsic circuitry involving the local axon
collaterals of corticothalamic projection cells in mouse smi cortex. J. Compar.
Neurol. 262, 13 -- 26.
Yoshimura, Y., Dantzker, J., & Callaway, E. (2005). Excitatory cortical neurons
form fine-scale functional networks. Nature 433 (24), 868 -- 873.
Zarrinpar, A., & Callaway, E. M. (2006). Local connections to specific types of
layer 6 neurons in the rat visual cortex. J. Neurophysiol. 95, 1751 -- 1761.
41
K(cid:89)
Potjans & Diesmann, Connectivity explains cell-type specific activity
Supplemental Material
Connection probabilities
Derivation of Eq. (1) The connection probability CN pre,N post(K) of two neu-
rons from populations of sizes N pre and N post which are randomly connected (with
uniform probability) by K synapses can be derived as follows: the probability that
a particular pair of neurons is connected is equal to one minus the probability that
it is not connected, i.e.
CN pre,N post(K) = 1 − qN pre,N post(K).
Since the K synapses are created independently, qN pre,N post(K) is the probability
of not being connected by any of the synapses
qN pre,N post(K) =
qN pre,N post(1) = [qN pre,N post(1)]K.
i=1
The probability of not being connected by a particular synapse is one minus the
probability of being so
hence
qN pre,N post(1) = 1 − CN pre,N post(1)
= 1 − 1/N preN post,
CN pre,N post(K) = 1 −
(cid:18)
1 −
(cid:19)K
.
1
N preN post
Taylor-series expansion of Eq. (1) The expression for the connection prob-
ability can be expanded in a Taylor series using the binomial series
(1 + x)K = 1 + Kx +
K(K − 1)
2
x2 + ...
Although conversion is guaranteed for x < 1, we require K 2x2 → 0 in order to
write (1 + x)K = 1 + Kx. With x = −1/N preN post we find the approximation
CN pre,N post(K) = 1 −
1 −
1
(cid:18)
(cid:19)K
N preN post
= 1 − (1 + x)K
= 1 − (1 + Kx)
= −Kx
K
=
N preN post
valid for small K/N preN post.
Errors of experimental connection probabilities A lower bound for the sta-
tistical error of Ca results from the experimental standard deviations of the neuron
count by Beaulieu & Colonnier (1983) and the consideration of error propagation.
The physiological search for connected neurons corresponds to a random sam-
pling test, so that the statistical error of the number of connected neuron pairs
42
Potjans & Diesmann, Connectivity explains cell-type specific activity
δ(RQ) is given by the standard deviation of the binomial distribution. Therefore,
the statistical error of the connection probability is δCp =(cid:112)Cp(1 − Cp)/Q (given
Q > 0). For untested connections (i. e. (cid:80)
j Qj = 0), we set δCp = 0.01.
Both estimates are minimal statistical errors. The estimate for the anatomical
data ignores errors in the estimation of the number of synapses as well as errors due
to the limited number of reconstructed axons and dendrites. Both error estimates
do not account for any potential systematic error e.g. due to sampling biases. The
minimal statistical errors are sufficient to show that recurrence strength and loop
strength of the two connectivity maps (see Fig. 2) are statistically indistinguishable.
Gaussian lateral connectivity model
The anatomical and physiological mean connection probabilities Ca/p correspond
to the integration of the lateral connectivity profile C(r) over the corresponding
sampling radius ra/p, i.e. Ca/p = 1/(πr2
0 C(r)rdrdϕ, yielding the two
equations
0
a/p)(cid:82) ra/p
(cid:82) 2π
Ca = 2πC0σ2/(πr2
Cp = 2πC0σ2/(πr2
a)[1 − exp(−r2
p)[1 − exp(−r2
a/2σ2)]
p/2σ2)].
(10)
(11)
Assuming that the parameters of the lateral connectivity profile C0 and σ are
identical in both equations, we find
πr2
aCa
1 − exp(−r2
a/2σ2)
=
πr2
pCp
1 − exp(−r2
p/2σ2)
.
This equation can be solved numerically for σ. For ra (cid:29) σ we have exp(−r2
0 and therefore
a/2σ2) =
which can be solved for σ:
σ = rp
exp(−r2
pCp
aCa
p/2σ2) = 1 − πr2
(cid:20)
(cid:18)
πr2
(cid:19)(cid:21)−1/2
−2 ln
1 − πr2
πr2
pCp
aCa
C0 can now be found by solving eq. (10) (or analogously eq. (11)):
(cid:20)
(cid:18)
C0 =
πr2
aCa
2πσ2
1 − exp
− r2
a
2σ2
.
.
(cid:19)(cid:21)−1
In practice, we are interested in the limit ra → ∞, arriving at eq. (5) for σ and,
using limra→∞ exp(−r2
a/2σ2) = 0, eq. (6) for C0.
The resulting parameters depend on the sampling radius of the physiological
experiments, i.e. the lateral distance of the somata of cells in the paired recordings.
For the raw data set, this value is provided (100 µm, Thomson & Morris, 2002).
The modified physiological map, however, combines different experiments with po-
tentially different sampling radii. Here, we find that increasing the sampling radius
increases the zero-distance connection probability monotonically but decreases the
lateral spread. Nevertheless, the effect is small: the estimates of both parameters
change by less than 8% when altering the sampling radius from 50 to 150 µm.
43
Potjans & Diesmann, Connectivity explains cell-type specific activity
Scaling factor
The scaling factor ζ is used to compare measurements of connection probabilities
of individual connections. In order to first remove global differences that arise from
the differences in the lateral sampling, we scale the connection probabilities Ca/p of
a connectivity map such that the new mean of the map corresponds to a previously
a/p ← Ca/pCm/ ¯Ca/p, where ¯Ca/p denotes
determined model connectivity Cm: C(cid:48)
the global mean of the original connectivity map. The ratio of the scaled individual
connection probabilities is 1 if two maps differ with respect to the lateral sampling
but are otherwise in perfect agreement. In order to exclusively measure the quality
of the agreement, we remove the information which connectivity map provides the
larger estimate by calculating the ratio of the larger of the two estimates divided
by the smaller one
ζ =
max(C(cid:48)
min(C(cid:48)
a, C(cid:48)
p)
a, C(cid:48)
p)
if C(cid:48)
a/p > 0,
restricting the measure to ζ > 1.
Instead of scaling to a previously determined
model connectivity, it is equivalent to compare the normalized connection proba-
bilities
ζ =
=
=
a, C(cid:48)
p)
a, C(cid:48)
p)
max(C(cid:48)
min(C(cid:48)
max(CaCm/ ¯Ca, CpCm/ ¯Cp)
min(CaCm/ ¯Ca, CpCm/ ¯Cp)
max(Ca/ ¯Ca, Cp/ ¯Cp)
min(Ca/ ¯Ca, Cp/ ¯Cp)
.
This can be seen by substituting the mathematical definitons of max(x, y) and
min(x, y), x + y ± x − y.
Consistent modification of target specificity: exact expression for redis-
tribution of synapses
The redistribution of synapses is a method to modify the connection probabilities
forming a projection while conserving the total number of synapses. The connection
probability is a non-linear function of the number of synapses (see Eq. (1)) and
therefore the proportion of synapses that contacts excitatory targets ∆ depends
on the number of neurons in the presynaptic and the two postsynaptic populations
(N pre, N post=e and N post=i) as well as on the total number of synapses K and
the desired target specificity T . In Materials and Methods, we provide the solution
for ∆ assuming the linear approximation C = K/(N preN post). The exact value is
found by numerically solving
2T =
(cid:18)
−
1 −
(cid:18)
(cid:19)(1−∆)K
(cid:19)∆K
1
N post=iN pre
1 −
1
N post=eN pre
(1 + T )
(1 − T )
(12)
for ∆. In practice, we use this variant in order to prevent inaccuracies due to the
approximation if K/(N preN post) is not sufficiently small.
44
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 13: Redistribution of synapses with conserved total number of synapses
according to eq. (12). The cartoons depict a situation with N pre = 10 presy-
naptic neurons (bottom triangles), N post=e = 16 excitatory (top triangles) and
N post=i = 4 inhibitory (circles) postsynaptic neurons and K = 20 synapses form-
ing the projection of the presynaptic population to the postsynaptic layer. The
arrow thickness corresponds to the indicated number of synapses that select ex-
citatory (Ke) and inhibitory (Ki) targets. (A) Due to the larger dendritic length
of the postsynaptic excitatory neurons, the application of Peters' rule results in
a preferential targeting of excitatory neurons with a target specificity of at least
T = +0.2 (Fig. 4). (B) Random selection of postsynaptic neurons as it is typically
applied in mono-layered network models (e.g. Brunel (2000b)). (C) and (D) show
preferental selection of inhibitory targets with a target specificity of T = −0.2 and
T = −0.5, respectively.
45
Potjans & Diesmann, Connectivity explains cell-type specific activity
Selection of additional candidates for specific target type selection
Electrophysiological recordings (Thomson et al., 2002) demonstrated that the pro-
jection from L2/3e to L4 exclusively targets interneurons. We are not aware of
further studies that investigate the inputs from L2/3 to L4 inhibitory cells. How-
ever, no further study in visual or somatosensory areas found connections from
L2/3e to L4e, but connections in the inverse direction were reported from L4e to
L2/3e (see Table 5). Only in auditory cortex, projections from L2/3e to L4e have
been reported to pyramids and star pyramids (Barbour & Callaway, 2008).
The projection from L5e to L2/3 has a large scaling factor (Fig. 2) and evidence
exists from several studies that this projection targets specifically inhibitory neurons
in L2/3: Strongest evidence comes from a photostimulation study (Dantzker &
Callaway, 2000) reporting that L2/3e cells receive few inputs from L5e while a
subset of L2/3 inhibitory neurons receives the majority of its inputs from this layer.
For the inputs from deep L5, only the input to this type of interneurons is significant
at all. Yet, no paired recordings study reports measurements of the L5e to L2/3i
connection, but several studies report an asymmetry of the feed-forward L2/3e
to L5e and the inverse L5e to L2/3e connection (Thomson & Bannister, 1998;
Thomson et al., 2002: ten-fold higher feed-forward connectivity, Lefort et al., 2009:
3.6-fold higher feed-forward connectivity).
In contrast, the anatomical estimate
provides similar probabilities for both connections and, in addition to Binzegger
et al. (2004) which is based on a single reconstructed L5e axon projecting to L2/3,
several studies found L5e cells predominantly projecting to L2/3 (e.g. Martin
& Whitteridge, 1984; Stepanyants et al., 2008). These discrepancies between
physiological and anatomical estimates can partly be related to undersampling of
cells projecting to L2/3 in physiological recordings and also to stronger cutting of
ascending axons in slice experiments: Stepanyants et al. (2009) estimated that only
60% of all potential connections can be expected to be intact for the excitatory
L5 to L2/3 connection in paired recordings in slices. A refinement of our model
that takes into account multiple excitatory L5 populations (e.g. L5A and L5B, see
Schubert et al., 2006) can potentially resolve some of the connectivity observed
on a finer scale. Taken together, this evidence qualifies the L5e to L2/3 projection
as an i-specific candidate projection.
Zarrinpar & Callaway (2006) investigated the layer-specific input to L6 pyra-
mids and interneurons in photostimulation experiments in rat primary visual cortex.
Although the inputs from L2/3 to L6 are found to be sparse, they are significantly
preferring interneuronal targets: in only very few cases (less than 10%) significant
excitatory inputs from L2/3 to L6 pyramids are found, but in around 70% of all
trials to interneurons (see their figure 5).
McGuire et al. (1984) find in an electron microscopy (EM) study that asym-
metric (excitatory) synapses originating in L6 preferentially target the shafts of
dendrites in L4. We include this projection as a candidate for specific target type
selection inspite of the few samples underlying this study and the indirect assess-
ment of connectivity without explicit identification of the postsynaptic cell. This
interpretation of the EM data is partly supported by a similar case in the literature:
White & Keller (1987) and Elhanany & White (1990) applied EM to study the
connections of L6 cortico-cortical (CC) and cortico-thalamic (CT) cells and found
that CT cells target predominantly dendritic shafts (as the L6e to L4 projection
inMcGuire et al. (1984)) while CC cells target spines as typical for excitatory cells
(Braitenberg & Schuz, 1998). Later, paired recordings (Mercer et al., 2005; West
et al., 2006) showed that this ultrastructural feature indeed reflects specific target
46
Potjans & Diesmann, Connectivity explains cell-type specific activity
type selection of interneurons. Since primarily CT cells project to L4 (Briggs, 2010)
it is possible that these cells show the same target type selection in this layer as in
L6. Although the evidence is rather sparse (see also Discussion), we include also
this projection in the list of i-specific candidates.
For the simulations, our reference parametrization assumes that the target
specificity of the i-specific projections is −0.8 for the projection from L2/3e to L4
and −0.4 for the three other candidate projections.
Algorithmic compilation of integrated connectivity map
(i) Input:
(a) Ca = 1 −(cid:0)1 −
i RiQi/((cid:80)
(b) Cp =(cid:80)
1
(cid:1)K
N preN post
j Qj)
(c) define rm (here 1 mm2)
(d) define amendment candidates for target specificity T (Table 2)
(ii) Scale connection matrices
(cid:104)−2 ln
(cid:16)
1 − Cp
Ca
(cid:17)(cid:105)−1/2
(a) σ = rp
(b) C0 = Ca
2πσ2
(c) ¯Cm = 2
r2
m
(d) C a/p ← Ca/p
C0σ2[1 − exp(−r2
m/(2σ2))]
¯Cm/ ¯Ca/p
(iii) Redistribute synapses of anatomical map
(a) solve eq. (12) for ∆
(b) C post=e
a
← 1 −(cid:0)1 −
←(cid:0) 1−T
1+T
1
N preN post
(cid:1)+(−)1 C post=e(i)
p
(a) C post=i(e)
p
(iv) Correct target specificity of physiological map
(cid:1)∆K (inhibitory connections analogously)
(v) Cm = 1
2(C a + C p)
Gray-matter long-range inputs
ext/K gm
In order to estimate the layer-specific number of external gray-matter inputs K gm
ext,
we first define the relative number of external gray-matter connections as γext =
K gm
total, where the total number of gray-matter connections can be divided
into local and external connections K gm
ext. Here, we assume
that the relative number of external gray-matter connections γext corresponds to
the weighted and normalized number of bouton clusters located outside of the
local cortical network as measured by Binzegger et al. (2007). The weighted and
normalized number of non-local bouton clusters is defined as follows: we count the
number of bouton clusters ξρ with a horizontal displacement greater than 0.56 mm
(i.e. outside of the local cortical volume below 1 mm2 of cortical surface) in
Binzegger et al. (2007), their figure 11, according to the rank ρ ∈ {1, 2, 3} (where
total = Klocal + K gm
47
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 14: Relative connectivity of the four connection types. The relative con-
vergent (left column) connectivity is the ratio of the relative averaged convergence
(weighted according to neuronal densities of postsynaptic cell types) and the rela-
tive neuronal densities of the presynaptic cell types. The relative divergent (right
column) connectivity is defined analogously for divergence. The relative connec-
tivity can be used to determine, which layer plays the dominant role in the local
network: If the overall effect of a neuron would be independent of its layer, any
neuron in the network, on average, sampled randomly from all other neurons and
the layer specificity of the average convergence and divergence corresponded to the
distribution of the number of neurons (relative connectivity of 1). We find that
neurons do not sample randomly from all other neurons in the network, for differ-
ent connection types, different layers play the dominant role: the highest number
of excitatory inputs relative to neuron numbers (to excitatory and inhibitory tar-
gets) originates from L2/3, inhibitory inputs (again to both target types) from
L4. Fewest inputs relative to neuron numbers are provided by L6 (e to e) and L5
(all other connection types). For outputs, the situation is different: the highest
number of outputs relative to neuron numbers are projected to L5 (e to e), to L4
(e to i) and to L2/3 (i to e/i). Fewest outputs relative to neuron numbers are
projected to L4 (i to e) and to L6 (all other types).
48
0.00.51.01.5e to erelative connectivityconvergence0.00.51.01.5e to i0.00.51.01.5i to e0.00.51.01.5i to iL2/3L4L5L60.00.51.01.5e to edivergence0.00.51.01.5i to e0.00.51.01.5e to i0.00.51.01.5i to iL2/3L4L5L6 Potjans & Diesmann, Connectivity explains cell-type specific activity
number of external clusters ξ1
number of external clusters ξ2
number of external clusters ξ3
number of clusters Ξtotal
cluster weight ω1
cluster weight ω2
cluster weight ω3
number of local connections klocal
number of gray-matter inputs kgm
ext
number of thalamic inputs
number of white-matter inputs
total number of external inputs ktotal
ext
L2/3e
1
2
9
24
0.65
0.18
0.17
4512
534
0
1072
1606
L4e
2
1
1
18
0.75
0.2
0.05
3277
353
93
1665
2111
L5e
1
0
1
6
0.55
0.25
0.2
2721
389
0
1609
1997
L6e
0
1
2
14
0.75
0.2
0.05
2702
79
47
2790
2915
Table 6: Layer-specific external input properties. From top to bottom: the ex-
tracted parameters required to calculate the number of gray-matter external con-
nections, the estimates of the three types of external inputs (gray-matter, thalamic
and white-matter) to excitatory neurons, and the total number of external inputs.
The total number of external inputs is rounded for simulations (compare Table 4).
rank 3 collapses rank > 2). The relative number of external bouton clusters
is calculated by dividing ξρ by the total number of clusters Ξtotal that can be
extracted from Binzegger et al. (2007), their figure 3B. In order to assess the
relative proportion of synapses formed by neurons outside of the local cortical
network, we multiply the relative number of external bouton clusters with the
cluster weights ωρ that can be extracted from Binzegger et al. (2007), their figures
4C and 4D for rank ρ = 1 and 2. As an estimate for rank 3 we use the remaining
weight ω3 = 1 − (ω1 + ω2). Finally, we sum over all ranks and obtain with
3(cid:88)
γext =
ξρωρ/Ξtotal
for the number of external gray-matter inputs:
ρ=1
K gm
ext = Klocal
= Klocal
γext
1 − γext
(cid:80)3
Ξtotal −(cid:80)3
ρ=1 ξρωρ
ρ=1 ξρωρ
.
The total number of layer-specific external inputs Kext then is the sum of the
external gray-matter inputs K gm
ext, the extrinsic (white-matter, wm) inputs K wm
ext
and specific thalamic inputs K th
ext. Table 6 contains detailed information on the
extracted values for all parameters of external inputs.
Cell-type specific input structure
To visualize the strongest pathways in the cell-type specific input structure we
calculate a input hierarchy. Therefore, we consider the graph G = (V, E) with the
neuronal populations as vertices V = {L2/3e, L2/3i, L4e, L4i, L5e, L5i, L6e, L6i}
and the edges E = {(v, w)C((v, w)) > 0.04} being those connections with
In the input hierarchy, the first
connection probabilities C greater than 0.04.
49
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 15: Differences in input structure for inhibitory cell types L2/3i (A), L4i
(B), L5i (C) and L6i (D) (large circles). The illustrations show the strongest path-
ways of direct (first-order, thick arrows) and indirect (second-order, thin arrows)
excitation (black) and inhibition (gray) of a given population. Triangles represent
excitatory and circles inhibitory populations.
order inputs to a vertex w are defined as the direct inputs
L1(w) = {v(v, w) ∈ E}.
The second order inputs to w are the strongest of the direct inputs to the elements
in L1(w):
L2(w) = {v∃v ∈ V \ L1(w)
∩∃w(cid:48) ∈ L1(w) : (v, w(cid:48)) ∈ E
∩∀w(cid:48)(cid:48) ∈ L1(w) \ {w(cid:48)} : C((v, w(cid:48)))C((w(cid:48), w)) > C((v, w(cid:48)(cid:48)))C((w(cid:48)(cid:48), w))}.
The definition can easily be extended to higher order inputs. Second and higher
order inputs are excitatory (inhibitory) if the product of mean synaptic strengths
in the pathway is positive (negative).
Wiring algorithm
The time consumption of the serial wiring algorithm described in Fig. 16A is O(K).
In a distributed simulation setup (Fig. 16B), the neurons are distributed over M
processes and the synapses are only created on the process where the postsynaptic
neuron is located.
In this case, the high-level connection routine for randomly
connecting neuronal populations is as follows: Let M be the number of processes,
K the number of synapses to create, Npre(post) the arrays containing all global
identifiers (GIDs) of the presynaptic (postsynaptic) neurons and Npost
M the dis-
tribution of postsynaptic neurons on processes, i.e. Npost
M [m] contains all GIDs
of postsynaptic neurons located on the mth process. The numbers of synapses
created on a given process KM is binomially distributed (because of the uniform
probability that a synapse is established on a given machine) with the boundary
m KM [m], i.e. KM is
multinomially distributed. In order to optimize the parallelization of the high-level
connection routine, we draw a priori a random sample KM from the multinomial
distribution depending on K, M and Npost
M (Davis, 1993). This sample has to be
identical on every process, which can be achieved by the usage of an identically
condition that the total number of connections is K = (cid:80)
50
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 16: Wiring algorithm. The cartoons depicts the presynaptic (bottom
triangles) and the postsynaptic (top triangles) populations, here both containing
N pre(post) = 21 neurons. The connection probability between the two populations
is C = 0.357, corresponding to K = 16 synapses (arrows, black indicating the
currently established synapses). (A) Serial version. For every synapse, the pre-
and postsynaptic neuron (highlighted triangles) are drawn from the corresponding
(B) Distributed version. The network is distributed over M = 5
populations.
processes (gray boxes) such that the first process hosts Npost
[0] = 5 postsynaptic
neurons and all other processes 4 postsynaptic neurons. The number of synapses on
each process (here KM = [3, 2, 4, 4, 3]) is multinomially distributed. The algorithm
can establish 5 connections in parallel (black arrows and highlighted triangles).
5
seeded random number generator. To create the synapses, we draw (on process
m and using independent random number generators on every process) KM [m]
times a presynaptic neuron j from Npre and a postsynaptic neuron i from Npost
M [m]
as well as the synaptic weight wij and delay dij from the gaussian distributions
and call the low-level function Connect(i,j,wij,dij). Fig. 16B shows that the
presynaptic neuron is drawn from all neurons in the presynaptic population, but
that the postsynaptic neuron on a given process is drawn only from postsynaptic
neurons located on this process.
The algorithm requires a small serial overhead in order to create the distribu-
tion of the postsynaptic neurons over processes and the random sample from the
multinomial distribution. Else, it guarantees that on every process, only calcula-
tions are carried out that are required to create synapses local to the process, i. e.
the time consumption is O(K/M ).
51
Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 17: Replacing the Poissonian background noise with constant input cur-
rents to all neurons yields basically the same activity, demonstrating that the ex-
ternal noise is not requirement for the asynchronous irregular firing. The raster
plot (A) exhibits only slightly more synchrony than the simulation with Poissonian
background (Fig. 7A). The input current to the different populations is chosen
identical to the mean input current evoked by the Poissonian background input
in the reference parametrization (Table 4). The resulting firing rates (B) are the
same as in Fig. 7B.
inputs kext with khom
Figure 18: The network simulations exhibit comparable spontanous activity fea-
tures when the external inputs are independent of the layer. (A) shows the raster
plot and (B) the cell-type specific firing rates after replacing the estimated num-
bers of external
ext = 2000 for excitatory populations and
khom
ext = 1850 for inhibitory populations. The order of firing rates is basically the
same as with data based external inputs: L2/3e and L6e exhibit lowest, L4e in-
termediate and L5e highest rates; inhibitory rates are higher than excitatory ones.
L6e is basically silent. This is due to the small excitatory input to the excitatory
population: the data based estimate assigns 50% more inputs to this population.
52
100200300400500600time [ms]L6L5L4L2/3AB02468firing rates [Hz]L6L5L4L2/3100200300400500600time [ms]L6L5L4L2/3AB02468firing rates [Hz]L6L5L4L2/3 Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 19: Dependence of stationary activity on the synaptic strength of the
L4e to L2/3e connection. The available experimental data do not constrain the
connection probabilities between L4 and L2/3 unambigously (see also Discussion).
Feldmeyer et al. (2006) find that the convergence onto L2/3 pyramids from other
L2/3 pyramids and from L4 cells is the same. In our model, however, the conver-
gence from L2/3 cells is twice as large as from L4. To test for the influence of
this connection, we increase the synaptic strength of the L4e to L2/3e connection.
The graph shows the firing rates of the excitatory populations in L2/3 (squares),
L4 (diamonds), L5 (circles) and L6 (triangles), lightness increases with depth as
a function of the relative increase of this synaptic strength. L2/3e fires a signifi-
cant number of spikes for an increase of > 1.6. The other layers are only slightly
affected by this modification. An increase of 2 is consistent with the relative con-
vergence described by Feldmeyer et al. (2006) and corresonds to our reference
parametrization.
53
1.01.21.41.61.82.0relative increase synaptic strength (L4e to L2/3e)246810firing rate [Hz] Potjans & Diesmann, Connectivity explains cell-type specific activity
Figure 20: Membrane potential synchrony according to (Golomb, 2007). The
graphs shows the membrane potential synchrony of (A) excitatory populations and
(B) inhibitory populations in L2/3 (squares), L4 (diamonds), L5 (circles) and L6
(triangles), lightness increases with depth, as a function of the number of recorded
membrane potential traces. The membrane potential synchrony χ depends on
+ O(N ) where a > 0 is a
the number of neurons N as χ(N ) = χ(∞) + a√
constant (Golomb, 2007). As known from previous studies on mono-layer models
(Brunel, 2000b), the asynchronous activity is not perfectly asynchronous (which
would correspond to χ(∞) = 0), but weakly synchronized. In the layered network
model, the subthreshold activity is similar, with χ(∞) ≈ 0.1. The membrane
potential synchrony of excitatory neurons is ordered as follows (from largest to
smallest values): L2/3e, L6e, L4e, L5e. For inhibitory neurons it is highest in L5i
and almost identical in the other layers.
N
Table 7: Numerical connectivity values. From left to right:
connection specifier, physiological connection probability es-
timates for raw and modified data, anatomical estimates for
the number of synapses (in billion) and the corresponding
connection probabilities for raw and modified data sets, re-
spectively. A dash ( -- ) indicates connections that have not
been measured. Non-zero but undetermined values of the
raw physiological data set are given with the estimates used
by Haeusler & Maass (2007) in brackets.
connection
L2/3e→L2/3e
L2/3e→L2/3i
L2/3i→L2/3e
L2/3i→L2/3i
L4e→L4e
L4e→L4i
L4i→L4e
L4i→L4i
L5e→L5e
L5e→L5i
L5i→L5e
L5i→L5i
L6e→L6e
p
C raw
0.223
0.212
0.182
0.4
0.174
0.19
0.095
0.5
0.092
0.096
0.123
0.6
--
p
C mod
0.122
0.416
0.413
0.4
0.122
0.282
0.304
0.5
0.122
0.109
0.194
0.6
--
a
K raw
73.6
10.3
16.7
2.9
23.5
3.5
10.8
1.9
3.1
0.42
0.42
0.07
15.7
54
a
K mod
73.6
10.3
20.3
3.4
23.5
3.5
18.0
2.9
3.1
0.42
4.0
0.47
15.7
a
C raw
0.172
0.086
0.139
0.086
0.05
0.029
0.09
0.065
0.13
0.081
0.082
0.062
0.075
a
C mod
0.172
0.086
0.169
0.101
0.05
0.029
0.15
0.098
0.13
0.081
0.781
0.414
0.075
101000200030004000number of recorded neurons0.00.10.20.3membrane potential synchronyAB101000200030004000number of recorded neurons0.00.10.20.3membrane potential synchrony Potjans & Diesmann, Connectivity explains cell-type specific activity
connection
L6e→L6i
L6i→L6e
L6i→L6i
L2/3e→L4e
L2/3e→L4i
L2/3i→L4e
L2/3i→L4i
L2/3e→L5e
L2/3e→L5i
L2/3i→L5e
L2/3i→L5i
L2/3e→L6e
L2/3e→L6i
L2/3i→L6e
L2/3i→L6i
L4e→L2/3e
L4e→L2/3i
L4i→L2/3e
L4i→L2/3i
L4e→L5e
L4e→L5i
L4i→L5e
L4i→L5i
L4e→L6e
L4e→L6i
L4i→L6e
L4i→L6i
L5e→L2/3e
L5e→L2/3i
L5i→L2/3e
L5i→L2/3i
L5e→L4e
L5e→L4i
L5i→L4e
L5i→L4i
L5e→L6e
L5e→L6i
L5i→L6e
L5i→L6i
L6e→L2/3e
L6e→L2/3i
L6i→L2/3e
L6i→L2/3i
L6e→L4e
L6e→L4i
L6i→L4e
L6i→L4i
L6e→L5e
L6e→L5i
p
p
C mod
--
--
--
0.0
0.163
0.0
0.0
0.180
0.0
--
--
--
--
0.112
0.098
0.327
C raw
--
--
--
0.0
0.163
0.0
0.0
0.58
0.0
>0.0(0.2) >0.0(0.2)
>0.0(0.0) >0.0(0.0)
--
--
--
--
0.147
0.098
0.327
>0.0(0.2) >0.0(0.2)
0.0
0.0
0.0
0.0
--
--
--
--
0.034
0.0
0.0
0.0
0.0
0.0
0.0
0.0
--
--
--
--
--
--
--
--
--
--
--
--
--
--
0.14
0.0
0.0
0.0
--
--
--
--
0.014
0.0
0.0
0.0
0.0
0.0
0.0
0.0
--
--
--
--
--
--
--
--
--
--
--
--
--
--
55
K raw
a
1.4
--
--
9.2
0.95
1.8
0.22
12.3
1.5
0.84
0.08
9.2
0.26
0.66
0.02
18.4
2.6
0.96
0.17
3.7
0.42
0.36
0.03
7.8
0.26
1.6
0.01
8.9
1.3
0.0
0.0
1.7
0.21
0.18
0.0
4.8
0.47
0.36
0.03
2.7
0.42
--
--
37.2
5.7
--
--
1.7
0.16
K mod
a
1.4
11.4
1.5
9.2
0.95
1.8
0.22
12.3
1.5
0.84
0.08
9.2
0.26
0.66
0.02
18.4
2.6
0.96
0.17
3.7
0.42
0.36
0.03
7.8
0.26
1.6
0.01
8.9
1.3
0.0
0.0
1.7
0.21
0.18
0.0
4.8
0.47
0.36
0.03
2.7
0.42
0.0
0.0
37.2
5.7
0.0
0.0
1.7
0.16
a
C raw
0.032
--
--
0.02
0.008
0.014
0.007
0.123
0.067
0.03
0.013
0.031
0.004
0.008
0.001
0.041
0.02
0.008
0.005
0.035
0.018
0.014
0.005
0.025
0.004
0.02
0.001
0.089
0.045
0.0
0.0
0.016
0.008
0.001
0.0
0.068
0.033
0.024
0.01
0.009
0.005
--
--
0.12
0.072
--
--
0.024
0.01
a
C mod
0.032
0.269
0.173
0.02
0.008
0.014
0.007
0.123
0.067
0.03
0.013
0.031
0.004
0.008
0.001
0.041
0.02
0.008
0.005
0.035
0.018
0.014
0.005
0.025
0.004
0.02
0.001
0.089
0.045
0.0
0.0
0.016
0.008
0.001
0.0
0.068
0.033
0.024
0.01
0.009
0.005
0.0
0.0
0.12
0.072
0.0
0.0
0.024
0.01
Potjans & Diesmann, Connectivity explains cell-type specific activity
connection
L6i→L5e
L6i→L5i
mean values
p
C raw
--
--
0.145
p
C mod
--
--
0.139
a
K raw
--
--
--
K mod
a
0.0
0.0
--
a
C raw
--
--
0.046
C mod
a
0.0
0.0
0.079
56
|
1811.07423 | 2 | 1811 | 2018-11-23T10:10:16 | The Frequent Network Neighborhood Mapping of the Human Hippocampus Shows Much More Frequent Neighbor Sets in Males Than in Females | [
"q-bio.NC"
] | In the study of the human connectome, the vertices and the edges of the network of the human brain are analyzed: the vertices of the graphs are the anatomically identified gray matter areas of the subjects; this set is exactly the same for all the subjects. The edges of the graphs correspond to the axonal fibers, connecting these areas. In the biological applications of graph theory, it happens very rarely that scientists examine numerous large graphs on the very same, labeled vertex set. Exactly this is the case in the study of the connectomes. Because of the particularity of these sets of graphs, novel, robust methods need to be developed for their analysis. Here we introduce the new method of the Frequent Network Neighborhood Mapping for the connectome, which serves as a robust identification of the neighborhoods of given vertices of special interest in the graph. We apply the novel method for mapping the neighborhoods of the human hippocampus and discover strong statistical asymmetries between the connectomes of the sexes, computed from the Human Connectome Project. We analyze 413 braingraphs, each with 463 nodes. We show that the hippocampi of men have much more significantly frequent neighbor sets than women; therefore, in a sense, the connections of the hippocampi are more regularly distributed in men and more varied in women. Our results are in contrast to the volumetric studies of the human hippocampus, where it was shown that the relative volume of the hippocampus is the same in men and women. | q-bio.NC | q-bio |
The Frequent Network Neighborhood Mapping of the
Human Hippocampus Shows Much More Frequent
Neighbor Sets in Males Than in Females
M´at´e Fellnera, B´alint Vargaa, Vince Grolmusza,b,∗
aPIT Bioinformatics Group, Eotvos University, H-1117 Budapest, Hungary
bUratim Ltd., H-1118 Budapest, Hungary
Abstract
In the study of the human connectome, the vertices and the edges of the network
of the human brain are analyzed: the vertices of the graphs are the anatomi-
cally identified gray matter areas of the subjects; this set is exactly the same for
all the subjects. The edges of the graphs correspond to the axonal fibers, con-
necting these areas. In the biological applications of graph theory, it happens
very rarely that scientists examine numerous large graphs on the very same,
labeled vertex set. Exactly this is the case in the study of the connectomes.
Because of the particularity of these sets of graphs, novel, robust methods need
to be developed for their analysis. Here we introduce the new method of the
Frequent Network Neighborhood Mapping for the connectome, which serves as
a robust identification of the neighborhoods of given vertices of special interest
in the graph. We apply the novel method for mapping the neighborhoods of the
human hippocampus and discover strong statistical asymmetries between the
connectomes of the sexes, computed from the Human Connectome Project. We
analyze 413 braingraphs, each with 463 nodes. We show that the hippocampi of
men have much more significantly frequent neighbor sets than women; therefore,
in a sense, the connections of the hippocampi are more regularly distributed in
men and more varied in women. Our results are in contrast to the volumetric
studies of the human hippocampus, where it was shown that the relative volume
of the hippocampus is the same in men and women.
Introduction
While it seems to be clear for all brain scientists that the complex connec-
tion patterns of the neurons play a fundamental role in brain function [1, 2, 3],
when the large-scale, macroscopic description of these connections has become
∗Corresponding author
Email addresses: [email protected] (M´at´e Fellner), [email protected]
(B´alint Varga), [email protected] (Vince Grolmusz)
1
available by the development of diffusion MRI techniques, it turned out that
novel methods are needed to handle these large graphs [1, 2]. MRI-mapped
human connectomes have only several hundred or at most one thousand ver-
tices today [4], and, therefore, more complex, more refined graph theoretical
algorithms [5, 6, 7] can be applied for their analysis than the widely followed
network science approach, originally developed for tens of millions of vertices of
the web graph [8].
The need for robust analytical methods
In the mathematical analysis of the large biological graphs, it is necessary
to apply methods that are capable of dealing with the frequently erroneous
networks, computed from high-deviation experimental data [9, 10, 11]. Our re-
search group has developed several fault-tolerant analytical methods for biologi-
cal graphs, based on a well-known robust network measure: Google's PageRank
[12, 13, 14, 15].
Today, one of the main challenges in brain science is the detailed mapping
of the brain circuitry with reliable methods. While diffusion MRI yields a rela-
tively reliable set of tools for brain imaging, the workflow of identifying the com-
mon, human brain areas across different subjects and the tractography phase of
the processing have numerous difficulties [16, 17, 18]. Human braingraphs, or
connectomes are very focused structures for the description of the cerebral con-
nections: the nodes of these graphs are the anatomically identified small (1-1.5
cm2) areas of the gray matter (called ROIs, Regions of Interests), and two nodes,
corresponding to two ROIs, are connected by an edge if in the tractography
phase an axonal fiber is found, connecting these two ROIs. These connectomes
describe the macroscopical scale set of connections between hundreds of brain
areas. Before the era of diffusion MR imaging, only very fractional knowledge
was available on these connections [1]; today tens of thousands of connections
can be identified and examined.
Previous Work
Perhaps the most straightforward robust approach to be considered is the
study of the frequently appearing cerebral connections. In work [19] we have
mapped the differences in the individual variability of the connections within
the lobes and some smaller brain areas.
We have constructed the Budapest Reference Connectome Server [20, 21]
at the address https://pitgroup.org/connectome/, which is capable of gen-
erating consensus connectomes from the data of 477 subjects, consisting of k-
frequent edges (i.e., edges that are present in at least k braingraphs), with
user-selected k and other parameters.
The Budapest Reference Connectome Server is an excellent tool for gener-
ating a robust human connectome, and, additionally, its construction has led
to the discovery of the phenomenon of the Consensus Connectome Dynamics
(CCD) [22, 23, 6, 24], mirroring the development of the axonal connections
within the human brain.
2
The global, or general approaches for describing the frequent (and, there-
fore, robust) cerebral connections are not always detailed enough for the study
of specific brain regions. Additionally, the frequently appearing connections
(e.g., in the Budapest Reference Connectome Server [20, 21]) may describe the
frequent neighbors of the individual vertices of these graphs, but not the fre-
quently appearing neighbor-sets of important vertices. The description of these
neighbor-sets is the goal of the present work.
Our contribution: the Frequent Network Neighborhood Mapping
Let us consider an important small area, corresponding to a vertex in the con-
nectome (called ROI, region of interest) of the brain, say the left hippocampus.
It is an important question to describe those ROIs, which are directly connected
to the left hippocampus since all the connections to and from the other cerebral
areas go through these edges and these neighbors of the important ROI (in our
case the hippocampus). It is a very interesting question whether almost all the
subjects have the same neighbors of the hippocampus, or there is a considerable
variability among the subjects.
If there were no any variability among the connectomes of the individual
subjects, then in each connectome, the left hippocampus would be connected
to the very same set of other nodes or ROIs. However, there is a considerable
variability of these connections between distinct subjects [19, 20, 21]. Therefore,
no such common neighbor-set exists for any vertex in the braingraphs.
Instead of the non-existent single, universal neighbor-set, which would have
been present in the connectome of all the subjects, we can still identify at least
the frequently appearing neighbor-sets of the left hippocampus (or any other
given vertex of interest in the graph).
Clearly, this is a completely different task than identifying the frequently ap-
pearing edges of the connectome, which are mapped by the Budapest Reference
Connectome Server at https://pitgroup.org/connectome. As an example,
let us consider a vertex u, and two other vertices v and w. Suppose, that both
edges {u, v} and {u, w} are present in 90% of all connectomes, but it may hap-
pen that the vertex-set {v, w} appears only in the 80% of the graphs as the
neighbors of vertex v:
if the connectomes are indexed from 1 through 100, it
may happen that in connectomes 1 through 90 {u, v} is an edge; in connectomes
11 through 100 {u, w} is an edge; so both of them are present in 90% of the
connectomes, but the set {v, w} is neighboring with u from connectomes 11
through 90, i.e., only 80 connectomes, i.e., 80%.
In the present contribution, we map the frequently appearing neighbor-sets
of the left and the right hippocampi of the human connectome, and we make
comparisons between the lateral and sex-differences in the frequent neighbor-
sets of the left- and right hippocampi. The neighbor set sizes are bounded by
4 in our present study, since if we considered larger sets than 4, the numbers of
the sets would be increased considerably. Our braingraphs have 463 vertices for
all subjects.
Our results show strong differences in the neighbor-sets of the hippocampi
between the sexes: we mapped the neighbor-sets, which have significant dif-
3
ferences in frequency in men and women (we call these vertex sets "significant
neighbor-sets"), and we have found that
• the number of the significant neighbor-sets of the left hippocampus is 65
times higher in males than in females;
• the number of the significant neighbor-sets of the right hippocampus is 16
times higher in males than in females;
In a sense, these results show that the neighbor sets of the hippocampus
of the women are more varied between individuals, while these sets are more
regular, that is, less varied in the case of men's connectomes. These results
complement our general studies of the deep graph-theoretical parameters of the
connectomes of men and women [5, 25, 7], where it was proven that women's
connectomes are better "connected", in precisely defined mathematical and com-
puter engineering terms.
Discussion and Results
Mapping the frequent graph-theoretical structures in the human braingraph
makes possible the robust analysis of the possibly noisy data. If the frequency
count is set to a high enough value, (say, the structure in question needs to
be present in 80% or 90% of the connectomes of the subjects in the group
analyzed), then image acquisition artifacts, data processing errors and small,
random individual variabilities in the connectomes could be counter-measured.
Here we consider the hippocampus and its neighbors of the human brain.
The hippocampus plays an important role in numerous brain functions, like the
processing of short-time memory and turning it into long-time memory, in spa-
tial memory and orientation [26, 27, 28, 29]. Today, the hippocampus is perhaps
the most widely studied functional and structural entity of the human brain,
and, consequently, the detailed study of its neighbor sets is a relevant area. Ad-
ditionally, the detailed study of the cerebral circuitry is an emphasized research
topic today: describing the direct neighbors of one of the most important brain
areas in this sense is also a crucial question.
In our present study, we have discovered and analyzed frequent 1,2,3 and 4
element neighbor-sets of the left and the right human hippocampi.
Frequent Network Neighborhood Mapping of the hippocampus
From now on, we refer to the brain areas, or ROIs, by their resolution-
[30]
listed at
250 parcellation labels, based on the Lausanne 2008 brain atlas
and computed by using FreeSurfer [31] and CMTK [32, 33],
https://github.com/LTS5/cmp_nipype/blob/master/cmtklib/data/parcellation/lausanne2008/ParcellationLausanne2008.xls.
The "lh" and the "rh" prefixes abbreviate the "left-hemisphere" and "right-
hemisphere" localizations.
We have mapped separately the direct neighbor sets of the left- and the right
hippocampi.
4
We need to recall an important property of the frequencies of the neighbor
sets of a given vertex, in our case the left- or the right hippocampus. We say
that a set U is a subset of set V , if every element of U is, at the same time, the
element of V . V is called the superset of U . We denote this relation as follows:
U ⊆ V or, equivalently, V ⊇ U .
Let us consider the left hippocampus. For every neighbor-set U of the left
hippocampus, we assign a frequency value as follows: we count the graphs of the
subjects, where every element of set U is connected to the left hippocampus,
and divide this number by the number of all the graphs. Let us denote this
frequency value by φ(U ). Clearly, if V is a superset of U , then the frequency of
V cannot be larger than the frequency of U : φ(U ) ≥ φ(V ).
This observation concerning the frequencies, of course, holds for the
neighbor-sets of any vertex of our graphs.
Table 1 lists the neighbor-sets of the left hippocampus with a minimum fre-
quency of 0.995. Three ROIs (the Left-Putamen, Left-Thalamus-Proper and the
lh.isthmuscingulate 3) are connected to the left hippocampus in all braingraphs,
while several other ROI (such as the lh.superiortemporal 2 and the lh.insula 2),
together with some subsets of the first three ROIs form the remaining neighbor
sets with a minimum frequency of 0.995.
The supporting Tables S1 and S2 contain the one, two three and four-element
subsets of the neighbor-sets of the left- and the right hippocampi, respectively,
with a minimum frequency of 0.9.
Sex differences
In what follows we compare the neighbor sets of the left- and right hip-
pocampi in braingraphs, computed from the male and female subjects of the
dataset of the Human Connectome Project [34]. We identify those neighbor
sets of the hippocampus that are significantly more frequent in male- and in fe-
male connectomes. We will see that male connectomes have much more frequent
hippocampal neighbor sets than female connectomes.
By our knowledge, this is the first observed significant sex difference in the
connections of the hippocampus in the literature.
Anatomical sex differences in the volume of the hippocampus were studied
in [35]: it was found that males have larger absolute hippocampus volumes, but
the relative hippocampus volume, compared to both the total brain volume or
the intracranial volume, are the same in the two sexes.
Here we show that there are 65 times more frequent neighbor sets of size
at most 4 of the left hippocampus in males than in females; and there are 16
times more frequent neighbor sets of size at most 4 of the right hippocampus in
males than in females. These results show that the variability of the neighbor
sets of the hippocampus is smaller in the case of males than in females: in males
there are much more frequent neighbor sets than in females. In other words,
the variability of the neighbor sets of the hippocampus in the connectomes of
women is greater than in the case of men.
Table 2 summarizes the results for the sex differences in the frequent neigh-
bor sets of the hippocampus. In size 1, 2, 3 and 4 neighbor-sets, males have
5
Frequency
1
1
1
1
1
1
1
.
0.99758
0.99758
0.99758
0.99758
0.99758
0.99758
0.99758
0.99758
.......
0.99517
0.99517
0.99517
0.99517
0.99517
0.99517
0.99517
0.99517
Subset
(Left-Putamen)
(Left-Thalamus-Proper)
(lh.isthmuscingulate 3)
(Left-Putamen)(Left-Thalamus-Proper)
(Left-Putamen)(lh.isthmuscingulate 3)
(Left-Thalamus-Proper)(lh.isthmuscingulate 3)
(Left-Putamen)(Left-Thalamus-Proper)(lh.isthmuscingulate 3)
...........................................................
(lh.superiortemporal 2)
(Left-Putamen)(lh.superiortemporal 2)
(Left-Thalamus-Proper)(lh.superiortemporal 2)
(lh.isthmuscingulate 3)(lh.superiortemporal 2)
(Left-Putamen)(Left-Thalamus-Proper)(lh.superiortemporal 2)
(Left-Putamen)(lh.isthmuscingulate 3)(lh.superiortemporal 2)
(Left-Thalamus-Proper)(lh.isthmuscingulate 3)(lh.superiortemporal 2)
(Left-Putamen)(Left-Thalamus-Proper)(lh.isthmuscingulate 3)(lh.superiortemporal 2)
...................................................................................
(lh.insula 2)
(Left-Putamen)(lh.insula 2)
(Left-Thalamus-Proper)(lh.insula 2)
(lh.insula 2)(lh.isthmuscingulate 3)
(Left-Putamen)(Left-Thalamus-Proper)(lh.insula 2)
(Left-Putamen)(lh.insula 2)(lh.isthmuscingulate 3)
(Left-Thalamus-Proper)(lh.insula 2)(lh.isthmuscingulate 3)
(Left-Putamen)(Left-Thalamus-Proper)(lh.insula 2)(lh.isthmuscingulate 3)
Table 1: The most frequent subsets of the neighbors of the left hippocampus (cut fre-
quency value: 0.995). 413 subjects from the Human Connectome Project public release
were examined. In all the subjects, the left hippocampus is connected to the following three
ROIs: the Left-Putamen, Left-Thalamus-Proper and the lh.isthmuscingulate 3, and, there-
fore, to all 7 (= 23 − 1) non-empty subsets of those (the first 7 lines of the table). The
lh.superiortemporal 2 ROI is connected to the left hippocampus in all, but one subjects, so
its frequency is 412/413=0.99758. Note that no subset, containing lh.superiortemporal 2,
may have a higher frequency than the frequency of
i.e.,
0.99758.
Indeed, all the 8 (empty and non-empty) subsets of the Left-Putamen, Left-
Thalamus-Proper and the lh.isthmuscingulate 3, together with the lh.superiortemporal 2
i.e., with 0.99517 frequency. The ROI lh.insula 2
is present in 412 out of 413 subjects,
is connected to the left hippocampus in 411 subjects out of the 413 subjects, there-
fore its frequency is 411/413=0.99517.
Similarly, all the 8 subsets of the ROIs Left-
Putamen, Left-Thalamus-Proper, and the lh.isthmuscingulate 3, together with lh.insula 2
lh.insula 2, together with the
have the same 411/413=0.99517 frequency. Theoretically,
lh.superiortemporal 2 may have the same frequency as lh.insula 2 alone (if the subject, where
the left hippocampus - lh.superiortemporal 2 edge is missing is one of the two subjects where
the left hippocampus - lh.insula 2 edge is missing), but this is not the case: the frequency
of the neighbor set {lh.superiortemporal 2,lh.insula 2} is below the cut frequency value for
Table 1, i.e., 0.995.
lh.superiortemporal 2 alone,
6
Size # significant
differences
Support Node
0.8
0.8
0.8
0.8
0.9
0.9
0.8
0.8
0.9
HPC left
HPC left
HPC right
HPC right
HPC
HPC
HPC left
HPC right
HPC
sex
male
female
male
female
male
female
all
all
all
Size
1
45
38
50
50
64
62
43
52
68
Size
2
844
679
1023
997
1781
1406
759
1059
1728
Size
3
9102
7128
11989
10725
28823
17976
7863
11963
24741
4
65150
48717
91498
73424
309659
150167
54861
85624
233548
15732
1762
19828
Sign. diff.
for whom
15497
235
1659
103
17688
2140
Table 2: The summary of the results for sex differences. The first column list the minimum
support, or, in other words, the frequency cut-off values: there are two values: 0.8 or 0.9, i.e.,
80% 90%. The second column denotes the righ-, left- or both hippocampi; the abbreviation
HPC stands for the word "hippocampus". In the third column the sex is given; the next four
columns contain the number size 1, 2, 3 and 4 frequent neighbor-sets of the hippocampus
considered. The next column gives the number of the neighbor-sets, which have significantly
different frequencies (p=0.001) in male and female connectomes. The last, ninth column
gives the number of neighbor-sets, which are significantly more frequent in male or in female
connectomes: the sum of the two numbers in the ninth column is equal to the number in the
eighth column. For example, in the first row, we can see that in males, the left hippocampus
has 45 frequent 1-element neighbor sets; 844 frequent 2-element neighbor sets, 9102 3-element
neighbor sets and 65150 frequent 4-element neighbor sets, where the frequency cut-off is 0.8.
Moreover, one can see that there are 15732 sets, differing significantly in frequency in males
and in females; and the last column says that from these 15732 sets, 15497 are present in the
braingraph of males and only 235 in the braingraphs of females.
more frequent sets than females. The only exception is the 1-element frequent
neighbors of the right hippocampus, where both males and females have 50 fre-
quent singleton sets. Since all the elements of the frequent 2,3 and 4 element
subsets need to be present also as a frequent 1-element set, it is surprising that
there is such a big difference in the 4 element frequent neighbors of the right
hippocampus in males (91498 sets) and females (73424 sets).
The following observation is much more surprising: There is a definitive,
but not two large difference between the numbers of the frequent size-1,2,3 and
4 neighbor sets of the left and the right hippocampi between the sexes. If we
consider, however, the number of neighbor sets with frequencies statistically
significantly differing (with p=0.001) between the two sexes, we have got that
15732 sets differ, and from these, 15497 is significantly more frequent in male-,
and 235 is significantly more frequent in female connectomes, in the case of
the neighbors of the left hippocampus. Neighbor sets of the right hippocampus
have 1762 significant differences, from which 1659 is significantly more frequent
in males, and 103 in female connectomes.
If we take the union of the neighbor sets of the left- and right hippocampi,
then the number of the neighbor-sets with significant differences is 19828, from
which 17688 are more frequent in male connectomes, and 2140 in female con-
nectomes (p=0.001).
Table 3 lists 10 neighbor-sets of the left hippocampus with the most signif-
icantly different frequencies in the sexes, where the higher frequency appears
at the male )5 sets) and also at the female subjects (5 sets). The supporting
7
frequency
male
0.823
0.823
0.823
0.823
0.823
frequency
female
0.556
0.561
0.561
0.561
0.561
name
(lh.fusiform 4)(lh.precuneus 10)(lh.precuneus 11)(lh.superiortemporal 6)
(lh.fusiform 4)(lh.isthmuscingulate 2)(lh.precuneus 6)(lh.superiortemporal 6)
(lh.fusiform 4)(lh.lingual 6)(lh.precuneus 10)(lh.superiortemporal 6)
(lh.fusiform 4)(lh.lingual 8)(lh.precuneus 10)(lh.superiortemporal 6)
(lh.fusiform 4)(lh.isthmuscingulate 2)(lh.precuneus 10)(lh.superiortemporal 6)
0.880
0.880
0.880
0.880
0.880
0.967
0.967
0.967
0.967
0.967
(Brain-Stem)(lh.bankssts 2)
(Brain-Stem)(Left-Thalamus-Proper)(lh.bankssts 2)(lh.isthmuscingulate 3)
(Brain-Stem)(Left-Putamen)(lh.bankssts 2)(lh.isthmuscingulate 3)
(Brain-Stem)(Left-Thalamus-Proper)(lh.bankssts 2)
(Brain-Stem)(lh.bankssts 2)(lh.isthmuscingulate 3)
Table 3: Several neighbor-sets of the left hippocampus with the most significant differences
in the frequencies between the sexes. The first five rows list five subsets, which are more
frequent in the braingraphs of men than of women. It is interesting that the lh.fusiform 4
ROI is present in the five sets. The next five are the most significant sets with frequencies
higher in females than in males. The lh.bankssts 2 and the Brain-Stem ROIs are present in
all five sets. The supporting Table S3 lists those neighbor-sets of the left hippocampus, which
are significantly more frequent in female connectomes; while supporting Table S4 lists those,
which are significantly more frequent in male connectomes, p ≤ 0.01. Similarly, Tables S5 and
S6 contain the analogous data for the right hippocampus.
Table S3 lists those neighbor-sets of the left hippocampus, which are signifi-
cantly more frequent in female connectomes; while Table S4 lists those, which
are significantly more frequent in male connectomes.
Table 4 lists 10 neighbor-sets of the right hippocampus with the most sig-
nificantly different frequencies in the sexes, where the higher frequency appears
at the male (5 sets) and also at the female subjects (5 sets). The supporting
Table S5 lists those neighbor-sets of the right hippocampus, which are signifi-
cantly more frequent in female connectomes; while Table S6 lists those, which
are significantly more frequent in male connectomes. The significance threshold
is p=0.01 (corrected by the Holm-Bonferroni method).
As we can see in Table 2, there is not a large difference between the frequent
1-element neighbors of the hippocampus in men and women. However, there
is a very significant difference in neighbor-set frequencies of higher cardinality,
as it is described in the last two columns of Table 2. One possible reason for
this could be that the neighbor sets of the hippocampus in general and the left
hippocampus, in particular, have less variability in the case of men than in the
case of women: men have the more regularly appearing neighbor-sets, while
women have more varied neighbor-sets.
Materials and Methods
The primary data source of the present study is Human Connectome
Project's website at http://www.humanconnectome.org/documentation/S500
[34], containing the HARDI MRI data of healthy human subjects between 22
and 35 years of age.
8
frequency
male
0.840
0.909
0.914
0.914
0.914
frequency
female
0.682
0.774
0.782
0.787
0.787
name
(rh.entorhinal 1)(rh.parahippocampal 2)(rh.precuneus 4)(rh.transversetemporal 1)
(rh.entorhinal 1)(rh.parahippocampal 2)(rh.precuneus 2)(rh.precuneus 4)
(Brain-Stem)(rh.entorhinal 1)(rh.parahippocampal 2)(rh.precuneus 4)
(Brain-Stem)(rh.entorhinal 1)(rh.precuneus 2)(rh.precuneus 4)
(rh.bankssts 2)(rh.entorhinal 1)(rh.parahippocampal 2)(rh.precuneus 4)
0.736
0.769
0.769
0.791
0.791
0.862
0.887
0.887
0.900
0.900
(rh.fusiform 8)(rh.inferiortemporal 1)(rh.middletemporal 9)
(rh.fusiform 8)(rh.inferiorparietal 4)(rh.lingual 7)(rh.middletemporal 9)
(rh.fusiform 8)(rh.middletemporal 9)(rh.supramarginal 9)
(rh.fusiform 8)(rh.lingual 7)(rh.middletemporal 9)
(Right-Amygdala)(rh.fusiform 8)(rh.middletemporal 9)
Table 4: Several neighbor-sets of the right hippocampus with the most significant differences in
the frequencies between the sexes. The first five rows list five subsets, which are more frequent
in the braingraphs of men than of women. It is interesting that the rh.precuneus 4 ROI is
present in all five sets. The next five sets are the most significant with frequencies higher in
females than in males. The rh.fusiform 8 and the rh.middletemporal 9 ROIs are present in all
five sets. The supporting Table S5 lists those neighbor-sets of the right hippocampus, which
are significantly more frequent in female connectomes; while Table S6 lists those, which are
significantly more frequent in male connectomes, p ≤ 0.01.
Construction of the Graphs
The workflow for generating the connectomes has applied the CMTK toolkit
[33], including the FreeSurfer tool and the MRtrix tractography processing tool
[36] with randomized seeding and with the deterministic streamline method,
with 1 million streamlines. For the present study we have applied the 463-
vertex graph resolution.
The parcellation data (containing the ROI labels) is listed in the CMTK
repository
nypipe
https://github.com/LTS5/cmp_nipype/blob/master/cmtklib/data/parcellation/lausanne2008/ParcellationLausanne2008.xls.
GitHub
The further details of braingraph-constructions are given in [37].
The
downloaded
from the https://braingraph.org/cms/download-pit-group-connectomes/
site, by choosing the "Full set, 413 brains, 1 million streamlines" option. In the
present study we have used the 463-node resolution graphs.
braingraphs
can
be
The Apriori Algorithm
The apriori algorithm [38] is a well-known tool in data mining for selecting
the frequent item sets from a large collection of subsets of a big item set. In
constructing association rules [39, 40] the selection of frequent subsets is a basic
In general, an n element set has 2n subsets,
step of the rule construction.
therefore for not-too-small n's it is not feasible to review all the 2n subsets.
However, if we want to identify only the subsets with high enough frequency (or
"support", in data mining terms [41]), then we can make use of the following
observation: Suppose that the set A has frequency α ≥ 0. Then all subsets of A
has a frequency at least α. Therefore, first, we identify those 1-element subsets,
which has a frequency at least α, it is an easy task. Then, for identifying all the
2-element subsets of frequency α we need to screen only the pairs of the frequent
9
1-element subsets. Next, for identifying the frequent 3-element subsets, we take
all the frequent 2-element subset appears with exactly one common element,
and verify if their union is frequent or not. The algorithm is continued in a
similar way, and usually, it finishes quickly.
Here we have applied an adaptation of the apriori code from the website
http://adataanalyst.com/machine-learning/apriori-algorithm-python-3-0/
with small modifications.
Statistical Analysis
For a chosen frequent neighbor set F , we have counted its occurrences in
the male dataset by count1(F ) and in the female dataset by count2(F ). The
support was calculated as suppi(F ) = counti(F )
, where Si, for i = 1, 2, wh is the
number of male and female braingraphs, respectively. For each set F we need
to determine whether supp1(F ) and supp2(F ) significantly differ. For this goal
we used the chi-squared test for categorical data:
Si
1st sample
2nd sample
total
contains F
count1(F )
count2(F )
does not contain F
S1 − count1(F )
S2 − count2(F )
total
S1
S2
count1(F ) + count2(F )
S1 + S2 − count1(F ) − count2(F )
S1 + S2
Then the test is calculated as
χ2 =
(count1(F ) · (S2 − count2(F )) − count2(F ) · (S1 − count1(F )))2 · (S1 + S2)
S1 · S2 · (count1(F ) + count2(F )) · (S1 + S2 − count1(F ) − count2(F ))
The degree of freedom for this test is one because it is the number of samples
minus one times the number of categories minus one.
Holm-Bonferroni correction [42]: After computing the p value for every
frequent set, we ordered these p-values p1 ≤ p2 ≤ p3 ≤ . . . ≤ pm. With a
significance level α = 0.01 let the Holm-Bonferroni value for kth frequent set be
k =
t: we have to
p
reject the null hypotheses for i indices i ≤ t. If t = 1, then we do not reject any
null hypotheses thus the difference in supports is significant.
m+1−k . Then let t be the minimal index such that pt > p
′
′
α
Conclusions
First in the literature, we have mapped the frequent neighbor sets of the hu-
man hippocampus, by applying the Frequent Network Neighborhood Mapping
method. We have identified the frequent neighbor sets of the human hippocam-
pus, and we have also compared the data of healthy young men and women in
respect to the neighborhood of the hippocampus. We have found that men have
much more significantly more frequent neighbor sets of the hippocampus than
women. Our results are in contrast with the generally much better connection
properties of the braingraphs of women than of men, as reported in [5, 25, 7].
Our results also need to be compared to the volumetric studies [35], where it
10
was shown that hippocampal volumes, relative to the intracranial volume and
also to the total brain volume are the same in the two sexes. Therefore, we
have shown that the neighbors of the human hippocampus significantly differ
in men and women, while there are no relative volumetric differences in the
hippocampus.
Data availability
The data source of this study is Human Connectome Project's website at
http://www.humanconnectome.org/documentation/S500 [34].
The
data,
containing the ROI labels, is listed in the CMTK nypipe GitHub repository
https://github.com/LTS5/cmp_nipype/blob/master/cmtklib/data/parcellation/lausanne2008/ParcellationLausanne2008.xls.
parcellation
The
braingraphs
downloaded
from the https://braingraph.org/cms/download-pit-group-connectomes/
site, by choosing the "Full set, 413 brains, 1 million streamlines" option. In the
present study we have used exclusively the 463-node resolution graphs.
can
be
Supplementary Tables S1 and S2 contain the one, two three and four-
element subsets of the neighbor-sets of the left- and the right hippocampi,
respectively, with a minimum frequency of 0.9. The Supplementary Ta-
ble S3 lists those neighbor-sets of the left hippocampus, which are signifi-
cantly more frequent in female connectomes; while Table S4 lists those, which
are significantly more frequent in male connectomes. Similarly, Tables S5
and S6 contain the analogous data for the right hippocampus. The supple-
mentary tables can be downloaded in a compressed MS Excel format from
http://uratim.com/hpc/supplementary_HCP_neighbors.zip.
Acknowledgments
Data were provided in part by the Human Connectome Project, WU-Minn
Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil;
1U54MH091657) funded by the 16 NIH Institutes and Centers that support
the NIH Blueprint for Neuroscience Research; and by the McDonnell Center
for Systems Neuroscience at Washington University. VG and BV were partially
supported by the VEKOP-2.3.2-16-2017-00014 program, supported by the Eu-
ropean Union and the State of Hungary, co-financed by the European Regional
Development Fund, VG and MF by the NKFI-126472 grant of the National
Research, Development and Innovation Office of Hungary. BV and MF was
supported in part by the EFOP-3.6.3-VEKOP-16-2017-00002 grant, supported
by the European Union, co-financed by the European Social Fund.
Conflict of Interest: The authors declare no conflicts of interest.
11
Author Contribution
MF analyzed braingraphs and developed statistical and graph theoretical
tools, BV constructed the image processing system and computed the brain-
graphs, VG has initiated the study, analyzed data and wrote the paper.
References
[1] H Sebastian Seung.
Reading
sampling versus dense mapping
(1):17 -- 29, Apr 2009.
http://dx.doi.org/10.1016/j.neuron.2009.03.020.
the book of memory:
of
10.1016/j.neuron.2009.03.020.
connectomes.
doi:
Neuron,
sparse
62
URL
[2] Olaf Sporns, Giulio Tononi, and Rolf Kotter. The human connec-
tome: A structural description of the human brain. PLoS Comput
Biol, 1(4):e42, Sep 2005.
10.1371/journal.pcbi.0010042. URL
http://dx.doi.org/10.1371/journal.pcbi.0010042.
doi:
[3] Jeff W. Lichtman, Jean Livet, and Joshua R. Sanes. A technicolour ap-
proach to the connectome. Nat Rev Neurosci, 9(6):417 -- 422, Jun 2008. doi:
10.1038/nrn2391. URL http://dx.doi.org/10.1038/nrn2391.
[4] Patric Hagmann, Patricia E. Grant, and Damien A. Fair. MR con-
nectomics: a conceptual framework for studying the developing brain.
Front Syst Neurosci, 6:43, 2012. doi: 10.3389/fnsys.2012.00043. URL
http://dx.doi.org/10.3389/fnsys.2012.00043.
[5] Bal´azs Szalkai, B´alint Varga, and Vince Grolmusz. Graph theoretical anal-
ysis reveals: Women's brains are better connected than men's. PLoS
One, 10(7):e0130045, 2015. doi: 10.1371/journal.pone.0130045. URL
http://dx.doi.org/10.1371/journal.pone.0130045.
[6] Csaba Kerepesi, Balazs Szalkai, Balint Varga, and Vince Grolmusz.
How to direct
the
consensus connectomes and the development of the connections in
the human brain.
URL
http://dx.doi.org/10.1371/journal.pone.0158680.
PLOS One, 11(6):e0158680, June 2016.
the connectomes: Dynamics of
the edges of
[7] Bal´azs Szalkai, B´alint Varga,
and Vince Grolmusz.
in women's
graph-theoretical
bias-compensated
size
better
havior,
http://dx.doi.org/10.1007/s11682-017-9720-0.
preprint arXiv:1512.01156.
connectomes.
doi:
Imaging
10.1007/s11682-017-9720-0.
parameters
Brain
2017.
Brain
are
also
and Be-
URL
Also in arXiv
[8] Michalis Faloutsos, Petros Faloutsos, and Christos Faloutsos. On power-
law relationships of the internet topology. In ACM SIGCOMM Computer
Communication Review, volume 29, pages 251 -- 262. ACM, 1999.
12
[9] Johannes Goll and Peter Uetz. The elusive yeast interactome. Genome
Biol, 7(6):223, 2006.
[10] Anne-Claude Gavin, Patrick Aloy, Paola Grandi, Roland Krause, Markus
Boesche, Martina Marzioch, Christina Rau, Lars Juhl Jensen, Sonja
Bastuck, Birgit Dumpelfeld, Angela Edelmann, Marie-Anne Heurtier,
Verena Hoffman, Christian Hoefert, Karin Klein, Manuela Hudak,
Anne-Marie Michon, Malgorzata Schelder, Markus Schirle, Marita Re-
mor, Tatjana Rudi, Sean Hooper, Andreas Bauer, Tewis Bouwmeester,
Georg Casari, Gerard Drewes, Gitte Neubauer, Jens M Rick, Bern-
hard Kuster, Peer Bork, Robert B Russell, and Giulio Superti-Furga.
Proteome survey reveals modularity of the yeast cell machinery. Na-
ture, 440(7084):631 -- 636, Mar 2006. doi: 10.1038/nature04532. URL
http://dx.doi.org/10.1038/nature04532.
[11] Nevan J Krogan, Gerard Cagney, Haiyuan Yu, Gouqing Zhong, Xinghua
Guo, Alexandr Ignatchenko, Joyce Li, Shuye Pu, Nira Datta, Aaron P
Tikuisis, Thanuja Punna, Jos´e M Peregr´ın-Alvarez, Michael Shales, Xin
Zhang, Michael Davey, Mark D Robinson, Alberto Paccanaro, James E
Bray, Anthony Sheung, Bryan Beattie, Dawn P Richards, Veronica
Canadien, Atanas Lalev, Frank Mena, Peter Wong, Andrei Starostine,
Myra M Canete, James Vlasblom, Samuel Wu, Chris Orsi, Sean R Collins,
Shamanta Chandran, Robin Haw, Jennifer J Rilstone, Kiran Gandi, Na-
talie J Thompson, Gabe Musso, Peter St Onge, Shaun Ghanny, Mandy
H Y Lam, Gareth Butland, Amin M Altaf-Ul, Shigehiko Kanaya, Ali
Shilatifard, Erin O'Shea, Jonathan S Weissman, C. James Ingles, Timo-
thy R Hughes, John Parkinson, Mark Gerstein, Shoshana J Wodak, Andrew
Emili, and Jack F Greenblatt. Global landscape of protein complexes in the
yeast saccharomyces cerevisiae. Nature, 440(7084):637 -- 643, Mar 2006. doi:
10.1038/nature04670. URL http://dx.doi.org/10.1038/nature04670.
[12] Vince I Grolmusz. Identifying diabetes-related important protein targets
with few interacting partners with the pagerank algorithm. Royal Society
Open Science, 2(4):140252, 2015.
[13] Vince Grolmusz. A note on the pagerank of undirected graphs. Information
Processing Letters, 115(6):633 -- 634, 2015.
[14] D´aniel B´anky, G´abor Iv´an, and Vince Grolmusz. Equal opportunity for
low-degree network nodes: a pagerank-based method for protein target
identification in metabolic graphs. PLoS One, 8(1):e54204, 2013.
[15] G´abor Iv´an and Vince Grolmusz. When the web meets the cell: using
personalized pagerank for analyzing protein interaction networks. Bioin-
formatics, 27(3):405 -- 407, 2011.
[16] Liang Zhan, Jiayu Zhou, Yalin Wang, Yan Jin, Neda Jahanshad, Gautam
Prasad, Talia M Nir, Cassandra D Leonardo, Jieping Ye, Paul M Thomp-
son, and For The Alzheimer's Disease Neuroimaging Initiative. Comparison
13
of nine tractography algorithms for detecting abnormal structural brain net-
works in Alzheimer's disease. Frontiers in Aging Neuroscience, 7:48, 2015.
ISSN 1663-4365. doi: 10.3389/fnagi.2015.00048.
[17] Saad Jbabdi and Heidi Johansen-Berg. Tractography: where do we go from
here? Brain Connect, 1(3):169 -- 183, 2011. doi: 10.1089/brain.2011.0033.
URL http://dx.doi.org/10.1089/brain.2011.0033.
[18] Matteo Bastiani, Nadim Jon Shah, Rainer Goebel, and Alard Roe-
Human cortical connectome reconstruction from diffusion
broeck.
weighted mri:
the effect of tractography algorithm. Neuroimage, 62
(3):1732 -- 1749, Sep 2012. doi: 10.1016/j.neuroimage.2012.06.002. URL
http://dx.doi.org/10.1016/j.neuroimage.2012.06.002.
[19] Csaba Kerepesi, Bal´azs Szalkai, B´alint Varga, and Vince Grolmusz. Com-
parative connectomics: Mapping the inter-individual variability of connec-
tions within the regions of the human brain. Neuroscience Letters, 662(1):
17 -- 21, 2018. doi: 10.1016/j.neulet.2017.10.003.
[20] Bal´azs Szalkai, Csaba Kerepesi, B´alint Varga, and Vince Grolmusz. The
Budapest Reference Connectome Server v2. 0. Neuroscience Letters, 595:
60 -- 62, 2015.
[21] Balazs Szalkai, Csaba Kerepesi, Balint Varga, and Vince Grolmusz. Param-
eterizable consensus connectomes from the Human Connectome Project:
The Budapest Reference Connectome Server v3.0. Cognitive Neurodynam-
ics, feb 2016. doi: http://dx.doi.org/10.1007/s11571-016-9407-z.
[22] Csaba Kerepesi, Balint Varga, Balazs Szalkai, and Vince Grolmusz. The
dorsal striatum and the dynamics of the consensus connectomes in the
frontal lobe of the human brain. Neuroscience Letters, 673:51 -- 55, 2018.
doi: 10.1016/j.neulet.2018.02.052.
[23] Bal´azs Szalkai, B´alint Varga, and Vince Grolmusz. The robustness and
the doubly-preferential attachment simulation of the consensus connectome
dynamics of the human brain. Scientific Reports, 7(16118), 2017. doi:
10.1038/s41598-017-16326-0.
[24] Balazs Szalkai, Csaba Kerepesi, Balint Varga, and Vince Grolmusz. High-
resolution directed human connectomes and the consensus connectome dy-
namics. arXiv, 1609.09036, September 2016.
[25] Bal´azs Szalkai, B´alint Varga, and Vince Grolmusz. The graph of our mind.
arXiv preprint arXiv:1603.00904, 2016.
[26] L J Santin, S Rubio, A Begega, R Miranda, and J L Arias. Spatial learning
ISSN
and the hippocampus. Revista de Neurologia, 31:455 -- 462, 2000.
0210-0010.
14
[27] Aristotle N Voineskos, Julie L Winterburn, Daniel Felsky, Jon Pipitone,
Tarek K Rajji, Benoit H Mulsant, and M Mallar Chakravarty. Hippocampal
(subfield) volume and shape in relation to cognitive performance across the
adult lifespan. Human brain mapping, 36:3020 -- 3037, August 2015. ISSN
1097-0193. doi: 10.1002/hbm.22825.
[28] Frauke Nees and Sebastian T. Pohlack. Functional MRIstudies of the hip-
pocampus. Frontiers of Neurology and Neuroscience, 34:85 -- 94, 2014. ISSN
1662-2804. doi: 10.1159/000356427.
[29] Juan Eugenio Iglesias, Jean C Augustinack, Khoa Nguyen, Christopher M
Player, Allison Player, Michelle Wright, Nicole Roy, Matthew P Frosch,
Ann C McKee, Lawrence L Wald, Bruce Fischl, Koen Van Leemput, and
Alzheimer's Disease Neuroimaging Initiative. A computational atlas of the
hippocampal formation using ex vivo, ultra-high resolution mri: Applica-
tion to adaptive segmentation of in vivo mri. NeuroImage, 115:117 -- 137,
July 2015. ISSN 1095-9572. doi: 10.1016/j.neuroimage.2015.04.042.
[30] Patric Hagmann, Leila Cammoun, Xavier Gigandet, Reto Meuli,
Christopher J. Honey, Van J. Wedeen, and Olaf Sporns. Map-
PLoS Biol,
ping the structural core of human cerebral cortex.
6(7):e159, Jul 2008.
URL
http://dx.doi.org/10.1371/journal.pbio.0060159.
10.1371/journal.pbio.0060159.
doi:
[31] Bruce Fischl. Freesurfer. Neuroimage, 62(2):774 -- 781, 2012.
[32] Stephan Gerhard, Alessandro Daducci, Alia Lemkaddem, Reto Meuli, Jean-
Philippe Thiran, and Patric Hagmann. The connectome viewer toolkit:
an open source framework to manage, analyze, and visualize connec-
tomes. Front Neuroinform, 5:3, 2011. doi: 10.3389/fninf.2011.00003. URL
http://dx.doi.org/10.3389/fninf.2011.00003.
[33] Alessandro Daducci, Stephan Gerhard, Alessandra Griffa, Alia Lemkad-
dem, Leila Cammoun, Xavier Gigandet, Reto Meuli, Patric Hag-
mann, and Jean-Philippe Thiran. The connectome mapper: an open-
source processing pipeline to map connectomes with MRI.
PLoS
10.1371/journal.pone.0048121. URL
One, 7(12):e48121, 2012.
http://dx.doi.org/10.1371/journal.pone.0048121.
doi:
[34] Jennifer A. McNab, Brian L. Edlow, Thomas Witzel, Susie Y. Huang, Hi-
manshu Bhat, Keith Heberlein, Thorsten Feiweier, Kecheng Liu, Boris
Keil, Julien Cohen-Adad, M Dylan Tisdall, Rebecca D. Folkerth, Han-
nah C. Kinney, and Lawrence L. Wald. The Human Connectome Project
initial applications of 300 mT/m gradients. Neuroimage,
and beyond:
80:234 -- 245, Oct 2013.
doi: 10.1016/j.neuroimage.2013.05.074. URL
http://dx.doi.org/10.1016/j.neuroimage.2013.05.074.
[35] Anh Tan, Wenli Ma, Amit Vira, Dhruv Marwha, and Lise Eliot. The human
hippocampus is not sexually-dimorphic: Meta-analysis of structural MRI
15
volumes. NeuroImage, 124:350 -- 366, January 2016. ISSN 1095-9572. doi:
10.1016/j.neuroimage.2015.08.050.
[36] J Tournier, Fernando Calamante, Alan Connelly, et al. Mrtrix: diffusion
tractography in crossing fiber regions. International Journal of Imaging
Systems and Technology, 22(1):53 -- 66, 2012.
[37] Csaba Kerepesi, Balazs Szalkai, Balint Varga, and Vince Grolmusz. The
braingraph. org database of high resolution structural connectomes and the
brain graph tools. Cognitive Neurodynamics, 11(5):483 -- 486, 2017.
[38] Rakesh Agrawal, Ramakrishnan Srikant, et al. Fast algorithms for mining
In Proc. 20th int. conf. very large data bases, VLDB,
association rules.
volume 1215, pages 487 -- 499, 1994.
[39] Rakesh Agrawal, Tomasz Imielinski, and Arun N. Swami. Mining associa-
tion rules between sets of items in large databases. In Peter Buneman and
Sushil Jajodia, editors, Proceedings of the 1993 ACM SIGMOD Interna-
tional Conference on Management of Data, Washington, D.C., May 26-28,
1993, pages 207 -- 216. ACM Press, 1993.
[40] Rakesh Agrawal and Ramakrishnan Srikant. Fast algorithms for mining
association rules in large databases. In Jorge B. Bocca, Matthias Jarke, and
Carlo Zaniolo, editors, Proc. of the 20th International Conference on Very
Large Data Bases (VLDB '94),, volume 1215, pages 487 -- 499. Kaufmann
Publishers Inc.
1994.
"
[41] Jiawei Han and Micheline Kamber. Data Mining: Concepts and Techniques.
Morgan Kaufmann Publishers, 2000.
[42] Sture Holm. A simple sequentially rejective multiple test procedure. Scan-
dinavian Journal of Statistics, pages 65 -- 70, 1979.
16
|
1210.4695 | 1 | 1210 | 2012-10-17T11:12:02 | Regulating the information in spikes: a useful bias | [
"q-bio.NC",
"cs.IT",
"cs.LG",
"cs.IT"
] | The bias/variance tradeoff is fundamental to learning: increasing a model's complexity can improve its fit on training data, but potentially worsens performance on future samples. Remarkably, however, the human brain effortlessly handles a wide-range of complex pattern recognition tasks. On the basis of these conflicting observations, it has been argued that useful biases in the form of "generic mechanisms for representation" must be hardwired into cortex (Geman et al).
This note describes a useful bias that encourages cooperative learning which is both biologically plausible and rigorously justified. | q-bio.NC | q-bio |
Regulating the information in spikes:
a useful bias
David Balduzzi
Department of Computer Science
ETH Zurich, Switzerland
[email protected]
The bias/variance tradeoff is fundamental to learning: increasing a model's complexity can improve
its fit on training data, but potentially worsens performance on future samples [1]. Remarkably,
however, the human brain effortlessly handles a wide-range of complex pattern recognition tasks.
On the basis of these conflicting observations, it has been argued that useful biases in the form of
"generic mechanisms for representation" must be hardwired into cortex [2].
This note describes a useful bias that encourages cooperative learning which is both biologically
plausible and rigorously justified [3 -- 9].
Let us outline the problem. Neurons learn inductively. They generalize from finite samples and
encode estimates of future outcomes (for example, rewards) into their spiketrains [10]. Results from
learning theory imply that generalizing successfully requires strong biases [1] or, in other words,
specialization. Thus, at any given time some neurons' specialties are more relevant than others.
Since most of the data neurons receive are other neurons' outputs, it is essential that neurons indicate
which of their outputs encode high quality estimates. Downstream neurons should then be biased to
specialize on these outputs.
The proposed biasing mechanism is based on a constraint on the effective information, ei, gener-
ated by spikes, see Eq. (*) below. The motivation for using effective information comes from a
connection to learning theory explained in §2. There, we show the ei generated by empirical risk
minimization quantifies capacity: higher ei yields tighter generalization bounds.
Sections §3 and §4 consider implications of the constraint in two cases: abstractly and for a concrete
model. In both cases we find that imposing constraint (*) implies: (i) essentially all information is
carried by spikes; (ii) spikes encode reward estimates and (iii) the higher the effective information,
the better the guarantees on estimates.
Although the proposal is inspired by cortical learning, the main ideas are information-theoretic,
suggesting they may also apply to other examples of interacting populations of adaptive agents.
1 Information
We model physical systems as input/output devices. For simplicity, we require that inputs X and
outputs Y form finite sets. Systems are not necessarily deterministic. We encode the probability that
system m outputs y ∈ Y given input x ∈ X in Markov matrix Pm(yx).
Consider two perspectives on a physical system. The first is computational: the system receives
an input, and we ask what output it will produce according to the probabilistic or deterministic
rule encoded in Pm. The second is inferential: the system produces an output, and we ask what
information it generates about its input.
The inferential perspective is based on the notion of Bayesian information gain. Suppose we have
model PM(dh) that specifies the probability of observing data given a hypothesis and also prior
distribution P (h) on hypotheses. If we observe data d, how much have we learned about the hy-
1
potheses? The Bayesian information gain is
Pi log2
D(cid:2)PM(Hd)(cid:13)(cid:13) P (H)(cid:3), where D[P kQ] :=Xi
Pi
Qi
is the Kullback-Leibler divergence and PM(hd) is computed via Bayes' rule.
Effective information is simply1 the Bayesian information gain of physical system m, where inputs
and outputs correspond to hypotheses and data respectively [3]. In the absence of additional con-
straints, we place the uniform (maximum entropy) prior on inputs. The effective information about
X generated by m when it outputs y is
(2)
Effective information quantifies the distinctions in the input that are "visible" to m, insofar as they
make a different to its output.
ei(m, y) := D(cid:2)Pm(Xy)(cid:13)(cid:13) Punif (X)(cid:3).
(1)
Finally, observe that effective information quantifies selectivity:
ei(m, y) = H(cid:2)Punif (X)(cid:3)
}
total bits available
{z
− H[Pm(Xy)]
bits indistinguishable to m
{z
}
=nselectivity of response y by mo,
(3)
where H[•] denotes entropy. Outputs generating higher effective information therefore trace back
to more specific causes (i.e. more concentrated posteriors).
2 Learning
The information one system generates about another should have implications for their future in-
teractions. We show this holds for the well-studied special case of empirical risk minimization
(ERM). Results are taken from [4], which should be consulted for details.
Given function class F ⊂ ΣX = {σ : X → ±1} and unlabelled data D = (x1, . . . , xℓ) ∈ X ℓ,
ERM takes labelings of D to the empirical error of the best-performing classifier in F
EF ,D : ΣX −→ R : σ 7→ min
f ∈F
1
ℓ
ℓXi=1
f (xi)6=σ(xi).
I
(4)
It is easy to show that ei(EF ,D, 0) = ℓ − V CF (D), where V CF (D) is the empirical VC-entropy
[4, 12]. It follows with high probability that
EhI
bf (x)6=σ(x)i
{z
}
expected error
≤
1
ℓ
I
ℓXi=1
bf (xi)6=σ(xi)
training error
{z
+c1vuuut
}
ei(EF ,D, 0)
1 −
effective information
ℓ
{z
}
+nconfidence termo.
(5)
Effective information answers the question: To what extent does the error take the form it does
because of the supervisor? Note that the optimal classifier does not depend on the supervisor ("en-
vironment") directly, but rather on the supervisor factored through EF ,D, which outputs the error:
LF ,D : ΣD −→ F : σ 7→ arg min
f ∈F
1
ℓ
f (xi)6=σ(xi)
I
ℓXi=1
{z
EF ,D(σ)
}
=: bf
(6)
A learning algorithm that shatters the data achieves zero empirical error for all supervisors, which
implies the error is independent of the supervisor. Equivalently, it implies the error achieved by
ERM generates no information about the supervisor. Increasing the effective information generated
by EF ,D progressively concentrates the distribution of likely future errors. Guarantees tighten as ei
increases.
An empirical risk minimizer's training error is a meaningful indicator of future performance to
the extent that it generates high effective information. This motivates investigating whether the
effective information generated when performing more biologically plausible optimizations also has
implications for future performance.
1In more general settings, effective information makes essential use of Pearl's interventional calculus [11].
2
3 Cooperative learning in abstract
Organisms aim to choose beneficial actions in the situations they encounter. Responsibility for mak-
ing choices falls largely on the cortex. However, cortical neurons interact extraordinarily indirectly
with the external environment -- their inputs and outputs are mediated through millions of other
neurons that constantly rewire themselves. It follows that, at best, neurons can encode provisional
estimates of expected outcomes into their spiketrains.
Guaranteeing and highlighting high quality estimates of future outcomes is therefore essential. In
particular, since interneuronal communication is dominated by spikes, it is necessary that spikes
provide meaningful indicators of future outcomes.
We propose an information-theoretic constraint that simultaneously guarantees and highlights esti-
mates in populations of interacting learners [5 -- 8]. The next section considers a particular model of
cortical neurons in more detail.
Let us model abstract learners as adaptable channels X
spikes y1 and silences y0. Impose the following
m−→ Y with two outputs, suggestively called
constraint:
ei(m, y1) = λ bits, where λ ≫ ei(m, y0).
(*)
Parameter λ controls the fraction of inputs causing the learner to produce y1 (the higher λ, the
smaller the fraction). It makes sense that each learner in a large population should specialize on a
small fraction of possible inputs. Indeed, increasing λ decreases the frequency of y1, so Pm(y1) ≪
Pm(y0).
Consequence 1: Spikes carry (essentially all) information. Under constraint (*), the informa-
tion transferred by learners is carried by spikes alone
I(X; Y ) = Pm(y1) · ei(m, y1) + O(cid:0)Pm(y1)2(cid:1) ≈ Pm(y1) · ei(m, y1)
(7)
to a first-order approximation [7]. It follows that silent learners can be ignored, despite the fact that
they typically constitute the bulk of the population's responses (recall: Pm(y1) ≪ Pm(y0)).
This suggests a principled way to distribute credit [6]. When a positive/negative global signal is
released, the few learners with informative responses -- that trace back to specific stimuli, recall (3)
-- should reinforce/weaken their behaviors. Conversely, the many learners producing uninformative
responses that do not trace back to specific stimuli should not modify their synapses. It turns out
this is exactly how neurons modify their synapses, see §4.
Consequence 2: Spikes encode high quality reward estimates. Under constraint (*), the effec-
tive information generated by reward maximizing2 learners essentially equals their empirical relative
reward [7]. Thus, outputs with high ei indicate high empirical reward relative to alternatives.
Furthermore, constraint (*) ensures that the information encoded in spikes is reliable. Applying a
PAC-Bayes inspired variant of Ockham's razor [15], it can be shown that the higher the effective
information generated by spikes, the smaller the difference between the empirical reward estimate
.
(8)
bRm and expected reward Rm:
(cid:12)(cid:12)Rm − bRm(cid:12)(cid:12)
{z
}
accuracy of reward estimate
≤r c1
T1
exp[ei(m, y1)]
·s ei(m, y1) + 1
+nconfidence termo
{z
}
term that decreases as ei increases
Summary. The information-theoretic constraint (*) introduces an asymmetry into outputs, ensures
that spikes carry reliable information about future rewards.
4 Cooperative learning in cortex
Spike-timing dependent plasticity (STDP) is a standard and frequently extended model of plasticity
[16]. Unfortunately, it operates in continuous time and is difficult to analyze using standard learning-
theoretic techniques. Recently [5], we investigated the fast-time constant limit of STDP. Taking the
2formalized as a constrained optimization following the free-energy approaches developed in [13, 14].
3
α+ · e(cid:16) tj −tk
τ+ (cid:17)
τ− (cid:17)
−α− · e(cid:16) tk −tj
∆wjk =
if tj ≤ tk
else.
(9)
STDP modifies synapses when input and output spikes co-occur in a short time window, so that
spikes gate learning. This makes sense if spikes are selective, and so by (7) carry most of the
information in cortex.
Synaptic weights control expected error.
If STDP incorporates neuromodulatory signals then it
can be shown to encode reward estimates into spikes in the fast-time constant limit τ• → 0, see [5].
The accuracy of these estimates is controlled by synaptic weights.
More precisely, let ν : X → R be a random variable representing neuromodulatory signals; l denote
the error, i.e. whether a spike is followed by a negative neuromodulatory signal; and R the empirical
reward after spiking [5]. The expected error is controlled by the sum ω of neuron m's synaptic
weights and m's empirical reward:
limit strips out the exponential discount factors, essentially reducing STDP to a simpler discrete time
algorithm whilst preserving its essential structure.
We recall STDP. Given a presynaptic spike at time tj and postsynaptic at time tk, the strength of
synapse j → k is updated according to
E(cid:2)l(x, m, ν)(cid:3)
}
{z
expected error
≤ c1 · ω2 −Xi
R(cid:0)x(i), m, ν(cid:1)
}
{z
empirical reward
+ ω ·ncapacity termo +nconfidence termo.
(10)
Thus, the lower synaptic weights ω, the better the quality of spikes as indicators of future outcomes.
Synaptic weights are homeostatically regulated. Spikes are both metabolically expensive [17]
and selective: they typically occur in response to specific stimuli -- e.g. an edge or a familiar face
[18]. Using spikes selectively reduces metabolic expenditures, which is important since the brain
accounts for a disproportionate fraction of the body's total energy budget [19].
There is evidence that synaptic strengths increase on average during wakefulness and are downscaled
during sleep [9, 20, 21]. This suggests that homeostatic regulation during sleep may both reduce
metabolic costs and simultaneously improve guarantees on reward estimates. Finally, it is easy to
show that decreasing synaptic weights increases the effective information generated by spikes.
5 Discussion
There is an interesting analogy between spikes and paper currency. Money plays many overlapping
roles in an economy, including: (i) focusing attention; (ii) stimulating activity; and (iii) providing a
quantitative lingua franca for tracking revenues and expenditures.
Spikes may play similar roles in cortex. Spikes focus attention: STDP and other proposed learning
rules are particularly sensitive to spikes and spike timing. Spikes stimulate activity: input spikes
cause output spikes. Finally, spikes leave trails of (Calcium) traces that are used to reinforce and
discourage neuronal behaviors in response to neuromodulatory signals.
Neither money nor spikes are intrinsically valuable. Currency can be devalued by inflation. Simi-
larly, the information content and guarantees associated with spikes are eroded by overpotentiating
synapses which reduces their selectivity (potentially leading to epileptic seizures in extreme cases).
Regulating the information content of spikes is therefore essential. Eq. (*) provides a simple con-
straint that can be approximately imposed by regulating synaptic weights. Indeed, there is evidence
that one of the functions of sleep is precisely this.
Spikes with high information content are valuable because they come with strong guarantees on
their estimates. They are therefore worth paying attention to, worth responding to, worth keeping
track of, and worth learning from.
Acknowledgements.
sions.
I thank Michel Besserve, Samory Kpotufe, Pedro Ortega for useful discus-
4
References
[1] Vapnik V: Estimation of Dependencies Based on Empirical Data. Springer 1982.
[2] Geman S, Bienenstock E, Doursat R: Neural Networks and the Bias/Variance Dilemma. Neural Comp
1992, 4:1 -- 58.
[3] Balduzzi D, Tononi G: Integrated Information in Discrete Dynamical Systems: Motivation and The-
oretical Framework. PLoS Comput Biol 2008, 4(6):e1000091.
[4] Balduzzi D: Falsification and Future Performance. In Proceedings of Solomonoff 85th Memorial Con-
ference, LNAI, Springer in press.
[5] Balduzzi D, Besserve M: Towards a learning-theoretic analysis of spike-timing dependent plasticity.
To appear in Adv Neural Information Processing Systems (NIPS) 2012.
[6] Balduzzi D, Tononi G: What can neurons do for their brain? Communicate selectivity with spikes.
To appear in Theory in Biosciences 2012.
[7] Balduzzi D, Ortega PA, Besserve M: Metabolic cost as an organizing principle for cooperative learn-
ing. Under review, 2012.
[8] Nere A, Olcese U, Balduzzi D, Tononi G: A neuromorphic architecture for object recognition and
motion anticipation using burst-STDP. PLoS One 2012, 7(5):e36958.
[9] Tononi G, Cirelli C: Sleep function and synaptic homeostasis. Sleep Med Rev 2006, 10:49 -- 62.
[10] Gottfried JA, O'Doherty J, Dolan RJ: Encoding predictive reward value in human amygdala and
orbitofrontal cortex. Science 2003, 301(5636):1104 -- 7.
[11] Pearl J: Causality: models, reasoning and inference. Cambridge University Press 2000.
[12] Boucheron S, Lugosi G, Massart P: A Sharp Concentration Inequality with Applications. Random
Structures and Algorithms 2000, 16(3):277 -- 292.
[13] Ortega PA, Braun DA: Information, utility and bounded rationality. The fourth conference on artificial
general intelligence 2011, :269 -- 274.
[14] Tishby N, Polani D: Information theory of decisions and actions. In Perception-Reason-Action. Edited
by Vassilis, Hussain, Taylor, Springer 2011.
[15] Seldin Y, Tishby N: Multi-Classification by Categorical Features via Clustering. In Proceedings of the
25th International Conference on Machine Learning 2008.
[16] Song S, Miller KD, Abbott LF: Competitive Hebbian learning through spike-timing-dependent
synaptic plasticity. Nature Neuroscience 2000, 3(9).
[17] Hasenstaub A, Otte S, Callaway E, Sejnowski TJ: Metabolic cost as a unifying principle governing
neuronal biophysics. Proc Natl Acad Sci U S A 2010, 107(27):12329 -- 34.
[18] Quiroga RQ, Reddy L, Koch C, Fried I: Invariant visual representation by single neurons in the
human brain. Nature 2005, 435(1102-1107).
[19] Attwell D, Laughlin SB: An energy budget for signaling in the grey matter of the brain. J Cereb Blood
Flow Metab 2001, 21(10):1133 -- 45.
[20] Vyazovskiy VV, Cirelli C, Pfister-Genskow M, Faraguna U, Tononi G: Molecular and electrophysio-
logical evidence for net synaptic potentiation in wake and depression in sleep. Nat Neurosci 2008,
11(2):200 -- 8.
[21] Maret S, Faraguna U, Nelson AB, Cirelli C, Tononi G: Sleep and waking modulate spine turnover in
the adolescent mouse cortex. Nat Neurosci 2011, 14(11):1418 -- 1420.
5
|
1705.00550 | 1 | 1705 | 2017-05-01T14:50:52 | Topography of the nuclei and distribution of Acetylcholinesterase activity in the septum of the telencephalon in man | [
"q-bio.NC"
] | Distribution of acetylcholinesterase (AChE) activity in the septum of the telencephalon in man was studied in 15 human brains using the acetylthiocholine method. Highest activity of AChE was found in the nucleus of the diagonal band and nucleus accumbens, and lowest in the Lateral nucleus. Comparison of histochemical results with cellular structure and wi.th the course of fibers showed absence in man of some of the nuclei described in animals such as the anterior medial nucleus, triangular nucleus, and marked reduction of the septo-hippocampal nucleus and fimbriate nucleus. Areas of the septum showing AChE activity were divided into an anterior and posterior system. The applicability to man of some neurophysiologic findings in animals is discussed. | q-bio.NC | q-bio | FOLIA MORPHOL. (WARSZ.), 1976, XXXV, 4, 405-412
JACEK BIAŁOW ĄS
TOPOGRAPHY OF THE NUCLEI AND DISTRIBUTION
OF ACETYLCHOLINESTERASE ACTIVITY IN THE SEPTUM
OF THE TELENCEPHALON IN MAN
Departrnent of Normal Anatomy, Institute of Medical Biology, Medical Academy,Gdańsk
Director of the Department: Prof. Dr. O. Nar'kiewicz
Distribution of acetylcholinesterase (AChE) activity in the septum
of the telencephalon in man was studied in 15 human brains using the
acetylthiocholine method. Highest activity of AChE was found in the
nucleus of the diagonal band and nucleus accumbens, and lowest in the
Lateral nucleus. Comparison of histochemical results with cellular
structure and wi.th the course of fibers showed absence in man of some
of the nuclei described in animals such as the anterior medial nucleus,
triangular nucleus, and marked reduction of the septo-hippocampal
nucleus and fimbriate nucleus. Areas of the septum showing AChE
activity were divided into an anterior and posterior system. The
applicability to man of some neurophysiologic findings in animals is
discussed.
The region of the septum of the telencephalon is one of the centers of the so-
called limbic system, which embraces the cortex of the cingulated gyrus and
some subcortical centers such as the amygdaloid body and hippocampus. The
name of this septum should not be confused with that of the septum
pellucidum, which applies only to the upper, thinned, part of the septum in
man and some monkeys. On the other hand, it includes the much wider basal
part and the paraterminal gyrus on the medial surface of the hemisphere,
corresponding to the vertical part of the diagonal band of Broca. Studies on
some animals show that this area contains, among others, inhibitory systems
which are active in emotional states (6, 9, 10, 14, 15, 17, 28). Lesion of the
septum is followed by increased activity in rats and cats, which become less
timorous. Zeeman and King (30) observed similar symptoms in human beings
with tumors in this area. Nowicki (22), in a pneumoencephalographic study,
noted a relation between symptoms of characteropathy and thickening
Fol. Morph. – 3
406 J. Białowąs Nr 4
of the septum in man. Various vegetative disorders have been reported after
lesions of the septum in rats (for literature, see Cytawa- 9).
Hamilton et al. and Igic et al. (15, 17) noted an effect of atropine and amiton
on functions of the septum, suggesting existence of specific systems of
cholinergic conduction in this area. Animals treated locally with these drugs
exhibit symptom s suggesting damage of the septum. Studies on rats, cats and
monkeys (4, 5, 12, 13, 18, 20, 21, 27) showed characteristic distribution of
acetylcholinesterase (AChE) in the septal nuclei. We considered that a similar
study in man would be interesting, especially since, according to the literature,
cholinergic transmission seems to play an important role in this region (4, 5,
15, 17, 19, 20, 23).
MATERIAL AND METHODS
Sections from 15 human brains obtained from the Department of Pathological
Anatomy of the Medical Academy, Gdańsk, taken 19-30 hours after death
from subjects aged 2 weeks to 74 years, were studied. A separate subgroup
consisted of brains from children (1 from a 2-weekold infant, and 2 from 3-
month-old children); all other brains were from adults. The patients from
whom the brain sections were taken had died from causes unrelated to the
central nervous system, with the exception of two who had hemorrhagic foci
in the brain, but distant from the septum, which was grossly unaffected. After
fixation in 10% formalin neutralized with calcium carbonate and cooled to
4°C, serial coryostat sections were stained by Koelle's acetylthiooholine
method in the modification of Gerebtzoff (11). Etopropazine was added to the
incubation mixture in the concentration of 10-4M to inhibit nonspecific
cholinesterases. The reaction was carried out at pH 5.6 at room temperature.
Sections next to those stained with acetylthiocholine were stained with cresyl
violet according to Nissl and with hematoxylin by the method of Weil.
RESULTS
Basing on a preliminary analysis of preparations stained with cresyl
violet and on published data from similar studies on animals (1, 2, 4, 5,
12, 13, 16, 21, 29), the septum was divided as proposed by Andy and
Stephan (3) with some modifications. The general scheme of the different
parts of the septum in man in serial frontal sections is illustrated in Fig. 1.
Nr 4 Nuclei of the septum of the telencephalon 407
The septum contains the fallowing nuclei:
1. Septo-hippocampal nucleus (HA),
2. Nucleus of the vertical part of the diagonal tract of Broca (NB),
3. Nuc1eus accumbens of the septum (NA),
4. Lateral nuc1eus (L),
5. The dorsal nuc1eus, consisting of an inner (Dl), intermediate (DM)
and outer part (DE),
6. The fimbriate nucleus (NF).
First, typical sections of the septum, from before backward, stained for AChE,
will be discussed.
Anterior part of the septum
Fig. 2A represents the most anterior segment of the septum, with distinctly
standing out rostrurn of the corpus callosum. AChE activity is increased in the
septum pellucidum 'in the form of two symmetric bands. Just below the
rostrum of the corpus callosum, a band of increased reactivity can be seen,
which corresponds to the rudimentary form of the septo-hippocampal nucleus
in man. In preparations stained by the method of Nissl, dark cells similar to
those encountered in the hippocampus can be seen in moderate numbers. Thin
active bands running to the sides and downward connect this area with the
deepest layer of cortex.
The next section in' the backward direction (Fig. 2B) shows high AChE
activity in the most anterior upper part of the diagonal band of Broca. An
anteromedial nuc1eus with increased enzyme activity, as seen in animals,
cannot be recognized here.
Below, the dark area of the nucleus accumbens forming the lower limit of
the ventric1e and indenting the septum is seen. In this region, high AChE
activity has characteristic diffuse dark-brown coloration. The area of increased
AChE activity coincides strictly with the boundaries of the .nucleus
accumbens, as determined by cellular structure. The rather small cells
intensely stained by cresyl violet with characteristic fine granules in their
cytoplasm farm irregular groups, especially in the medial part of the nuc1eus
accumbens. It is hard to say whether the increased reaction pertains mainly to
neuropil, as also to the cells as suggested by different intensity of staining
corresponding to the cell groups described above.
Between the nuc1eus accumbens and nuc1eus of the diagonal band,
bands giving an intense 'reaction for AChE connect the two. They run
diagonally upward and medial ward showing stratified activity; their external
part merges with the lateral part of the septum, while only the internal
408 J. Białowąs Nr 4
part reaches the upper part of the diagonal tract. In preparations stained with
cresyl violet, small groups of dark cells of a type similar to those in the
nucleus accumbens can be seen. These areas are usually distinguished
by darker brown coloration, indicating high AChE content. In horizontal
sections (Fig. 4 C, D) these bands extend fairly far in the forward direction,
running on the boundary between the deepest layer of the cortex and the white
substance of the cingulum.
Middle part of the septum
The next sections (Fig. 2C, D) represent the middle part of the septum. The
distinctly dark region in the medial part of the septum in the neighborhood of
the longitudinal fissure of the brain is the nucleus of the vertical part of the
diagonal band of Broca. The boundaries of the area showing increased AChE
activity in the form of diffuse brown coloration of varying intensity are quite
sharp. They correspond to the area containing the large irregularly shaped cells
with cresyl-violet stained granules in their cytoplasm, which are characteristic
of the nucleus of the diagonal band. The fairly numerous darker spots on the
brown background probably represent activity of the enzyme in the perikaryon
of these cells; the nuclei, on the other hand, give a negative reaction. In
sections further backward (Fig. 3A), the active area of the nucleus of the
diagonal band is reduced to a thin layers surrounding bundles of fibers of the
fornices from the side and from below, as can be seen also in the horizontal
sections (Fig. 4A, B).
Laterally to the diagonal tract and above the nucleus accumbens, the basal
area of the septum, occupied by the lateral nucleus, is characterized by lack of
AChE activity (Fig. 2, B, C, D). In preparations stained with cresyl violet, the
cells of this nucleus have various shape and size and are not very intensely
stained so that the boundaries of the nucleus are not sharply defined. In
preparations stained by the thiocholine method, the boundaries, particularly
with the dorsal nucleus, are somewhat more distinct (Fig. 3A),
The next frontal section (Fig: 3A) visualizes the part of the septum where the
crura of the anterior commissure approach the median plane and tend to join.
Besides the previously mentioned thin band surrounding the columns of the
fornix from the sides, which is the furthest backward reaching part of the
diagonal band, two other thin bands of strong AChE activity are visible in the
lateral and upper area of the septum, situated in the so-called dorsal nucleus of
the septum, in its lateral part upward and backward of the lateral nucleus. In
preparations stained with cresyl violet, three layers can be distinguished,
although not as distinctly as in some animals:
J. Białowąs
Fig. 1. Medial surface of the cerebral hemisphere in man. Planes of the frontal
sections in Figs. 2 and 3 are shown.
Designations : AS - area subcallosa, CA - commissura anterior,
CC - corpus callosum, CHO - chiasma opticum, CM - corpus mamillare, F -
jornix, PCH - plexus chorioideus, SP - septum pellucidum, SV - septum verum,
TH - thalamus.
Fig. 2. Consecutive frontal sections through the Septum in man. 'Ihiochołine
technique. The nuclei are designated, A, B, D - magn, X6; C - magn. X4. The
arrow points to active bands connecting the nucleus accumbens with the
nucleus of the diagonal band.
Designations : CA - commissura anterior, CC - corpus callosum,
HA - nucl. septo-hippocampalis, L - nucl. lateralis, NA - nucl. accumbens, NB -
nucl. tractus diagonalis Broca, NC - nucl. caudatus, R - rostrum corporis
callosi, SP - septum pellucidum.
J. Białowąs
Fig. 3. Further frontal sections through the septum in man. Thiocholine technique.
A - magn, X4; B - magn, X6.
Designations : CA - commissura anterior, DE - nucl. dorsalis - pars
externa, Dl - nucl. dorsalis - pars interna, DM - nucl. dorsalis - pars
intermedia, F - fornix, NC - nucl. caudatus, NF - nucl. fimbratus septi .
Fig. 4. Horizontal sections through the septum in man, from above downward.
Thiocholine technique. A, B, C, D - magn. X3. Arrows point to AOhE activity of
the band s connecting the septum with the cortex. Designations as in the preceding
figures (Dorsal nucleus – D).
Nr 4 Nuclei of the septum of the telencephalon 409
: an internal layer, neighboring with the posterior part of the diagonal band
and fornix, and containing fairly intensely stained cells; an intermediate part,
and an external part, neighboring with the ventricle, with less densely
distributed cells. The course of the bands of intense AChE activity agrees with
the course of the layers, but the bands are narrower than the layers. The first
band, reaching furthest downward, is situated in the internal part; and the
second band, which is very poorly visible, is situated in the intermediate part.
A stratification of the reaction in the postero-superior part of the septum can
be discerned also in the horizontal sections (Fig. 4A).
Posterior part of the septum
The last of the frontal sections (Fig. 3B) represents the septum just before
the interventricular foramen. The upper part of the septum is thick and hardly
fits here the term septum pellucidum any more. The columns of the fornix are
inactive, and higher up the somewhat darker area of the dorsal nucleus can be
seen. Of the bands of increased activity, only the internal one, separating the
fimbriate from the dorsal nucleus, remains. Compared to other animals, the
area occupied by the fimbriate nucleus is small. Only a few dark bands,
presumably representing acetylcholinesterase-rich nerve fibers, can be seen
above the columns of the fornix.
The triangular nucleus below the columns is indistinguishable in man,
probably owing to marked reduction of the anterior commissure of the fornix
to a thin band of fibers connecting both fornices on the ventral side, posteriorly
to the interventricular foramen. No significant differences were noted between
the preparations incubated with etopropazine and those incubated without it.
Preparations from brains received 26-30 hours after death gave less strongly
positive reactions compared with fresher brains.
DISCUSSION
The available literature contains precise descriptions of the distribution
of acetylcholinesterase activity in the septum in the cat, monkey
and rat (4, 5, 12, 13, 16, 18, 21), but only fragmentary information about
its distribution in man (16). The reports agree that highest AChE activity
is found in the nuclei of the diagonal band and nucleus accumbens of
the septum. According to the results of the present study, this applies
410 J. Białowąs Nr 4
also to man. In preceding studies (4, 5, 16, 21), stratification of the reaction
in the dorsal nucleus in the cat and rat was described. Girgiss (12, 13) has
described similar differentiation of the activity in Galago and monkeys. In the
present study, layers of increased activity were observed in the area of the
dorsal nucleus in man, in confirmation of the division of this nucleus into
layers proposed by Andy and Stephan (1-3, 29). Similar results were found by
Harkmark et al. (16).
In the light of the presented data, the division of Andy and Stephan reflects
the structural organization better than the former division by Shimazono and
Brockhaus (7, 8, 26). Nevertheless, AChE activity allows further
simplification. For instance, an anterior medial nucleus, which in animals
shows moderately high AChE activity, is not distinguishable in man. The area
adjacent to the ventricle given this name by Andy and Stephan should be
considered a part of the lateral nucleus because of the lack of AChE activity.
The triangular nucleus is absent in man, and the septo-hippocampal and
fimbriate nuclei are markedly reduced.
In previous studies on rats the acetylcholinesterase-rich centers in the septum
were divided into connected systems. The same can be done with the human
brain. For instance, the nucleus of the diagonal band together with the
fimbriate nucleus form a single system which, according to experimental
studies on animals, has massive cholinergic connections with the hippocampus
(5, 19-21, 23-25), and which may be called the posterior system.
A second group of AChE-rich centers in the septum consists of the
septo -hippocampal and medial anterior nuclei together with active bands
which connect them with the cingulate cortex. In man, in spite of absence of
the medial anterior nucleus, this system is better developed than in animals
owing to the active bands reaching far into the cingulated cortex. The name
anterior system is proposed.
The remaining centers showi.ng AChE activity are difficult to systematize.
The nucleus accumbens is probably connected with the striatum, and the
dorsal nucleus, which receives stratified projection from the hippocampus (24,
25), may also be the site of terminals of collaterals of fibers from the diagonal
tract (16).
The studies cited in the introduction indicate that the septum plays a
considerable role in the regulation of emotional states and vegetative functions
in animals and in man. Establishment of homology of the various areas of the
septum in animals and man by further comparative anatomical studies should
permit comparison of the results of some experimental studies on animals with
clinical observations.
Nr 4 Nuclei of the septum of the telencephalon 411
REFERENCES
l. Andy O. J., Stephan H.: J. Comp. Neurol., 1959, 111, 503.
2. Andy O. J., Stephan H.: J. Comp. Neurol., 1961, 1117,251.
3. Andy O. J., Stephan H.: J. Comp. Neurol., 1·968,133, 383.
4. Bialowąs J., Narkiewicz O.: Acta Neurobiol. Exp. (Warsz.), 1974, 34, 573.
5. Bialowąs J., Nitecka L., Narkiewicz O.: Folia Morphol. (Warsz.), 1972, 31,
145.
6. Brady J. V., Nauta W. J. H.: J. Comp. Physiol. Psychol., 1953, 46, 339.
7. Brockhaus H.: J. Psychol. Neurol., 1942a, 51, l.
8. Brockhaus H.: J. Psychol. Neurol., 1942b, 51, 57.
9. Cytawa J.: Acta Neurobiol. Exp. (Warsz.), 1974, 34, 161.
10. Donovik P. J.: J. Comp. Physioi. Psychol., 1968, 66, 569.
11. Gerebtzoij M. A.: Acta Anat. (Basel), 1953, 19, 366.
12. Girgis M.: Acta Anat. (Basel), 1969, 72, 94.
13. Girgis M.: Acta Anat, (Basel), 1973, 84, 202.
14. Hamilton L. W., Kelsey J. E., Grossman S. P.: J. Comp. Physiol. Psychol.,
1970, 70, 79.
15. Hamilton L. W., McCleary R. A., Grossman S. P.: J. Comp. Physiol.
Psychol., 1968, 66, 563.
16. Harkmark W., Mellgren S. I., Srebro B.: Brain Res., 1975, 95, 281.
17. Iqic R., Stern P., Basagic E.: Neuropharmacology, 1970, 9, 73.
18. Koelle G. B.: J. Comp. Neurol., 1954, 100, 21.1.
19. Lewis P. R., Shute C. C. D.: Brain, 1967, 90, 5211.
20. Lewis P. R., Shute C. C. D.: J. Physiol., 1957, UH, 2115.
21. Mellgren S. I., Srebro B.: Brain Res., 1973, 52, 19.
22. Nowicki Z.: Znaczenie rozszerzonych jam przegrody przeźroczyste] w
psychopatologii. Thesis for degree of medicine doctor, Medical Academy,
Gdańsk 1973.
23. Oderfeld-Nowak B., Narkiewicz O., Białowąs J., Dąbrowska J.,
Wieraszko A., Grądkowska M.: Acta Neurobiol Exp. (Warsz.), 1974, 34,
583...
24. Raisman G.: Brain, 1966, 89, 317.
25. Raisman G.: Exp. Brain Res., 1969, 7, 317.
26. Shimazono J.: Arch. Anat. Physiol., Anat. Abt., 1912, 55.
27. Shute C. C. D., Lewis P. R.: Nature (Lond.), 1963, 199, 1160.
412 J. Białowąs Nr 4
28. Smythies J. R.: Brain mechanisms and behavior. Blackwell Scientific
Publications, Oxford and Edinburgh 1970.
29. Stephan H., Andy O. J.: J. Hirnforsch., 1964, 7, 1.
30. Zeeman W., King F. A.: J. Nerv. Ment. Dis., 1958, 127, 490.
The author's present contact address: [email protected] (Dr. med.
Jacek Białowąs in Polish , Jacek Bialowas MD in English),
|
1207.0033 | 2 | 1207 | 2012-10-08T16:32:07 | Quantum-like model of behavioral response computation using neural oscillators | [
"q-bio.NC"
] | In this paper we propose the use of neural interference as the origin of quantum-like effects in the brain. We do so by using a neural oscillator model consistent with neurophysiological data. The model used was shown to reproduce well the predictions of behavioral stimulus-response theory. The quantum-like effects are obtained by the spreading activation of incompatible oscillators, leading to an interference-like effect mediated by inhibitory and excitatory synapses. | q-bio.NC | q-bio |
Quantum-like model of behavioral response computation
using neural oscillators
J. Acacio de Barros
Liberal Studies Program, San Francisco State University, San Francisco, CA 94132
Abstract
In this paper we propose the use of neural interference as the origin of quantum-
like effects in the brain. We do so by using a neural oscillator model consistent
with neurophysiological data. The model used was shown elsewhere to reproduce
well the predictions of behavioral stimulus-response theory. The quantum-like
effects are brought about by the spreading activation of incompatible oscilla-
tors, leading to an interference-like effect mediated by inhibitory and excitatory
synapses.
Keywords: disjunction effect, quantum cognition, quantum-like model, neural
oscillators, stimulus-response theory
1. Introduction
Quantum mechanics is one of the most successful scientific theories in his-
tory. From it, detailed predictions and extremely precise descriptions of a wide
range of physical systems are given. Thus, it is not surprising that researchers
from social and behavioral sciences disciplines started asking whether the ap-
paratus of quantum mechanics could be useful to them (Khrennikov, 2010). In
applying the quantum formalism, most researchers do not claim to have found
social or behavioral phenomena that are determined by the physical laws of
quantum mechanics (Bruza et al., 2009).
Instead, they claim that by using
the representation of states in a Hilbert space, together with the correspond-
ing non-Kolmogorovian interference of probabilities, a better description of ob-
served phenomena could be achieved. To distinguish the idea that the formalism
describing a system is quantum without requiring the system itself to be, re-
searchers often use the term quantum-like.
One of the first applications of quantum-like dynamics outside of physics was
in finances (Baaquie, 1997; Haven, 2002) and in decision making (Busemeyer et al.,
2006; Aerts, 2009).
In economy, Baaquie (1997) showed that a more general
form of the Black-Scholes option pricing equation was given by a Schrödinger-
type equation, and Haven (2002) linked the equivalent of the Plank constant to
the existence of arbitrage in the model. In psychology, most of the initial ef-
fort was focused on using quantum-like dynamics and Hilbert-space formalisms
Preprint submitted to BioSystems
June 25, 2018
to describe discrepancies between empirical data and "classical" models using
standard probability theory (see, for example, Aerts (2009); Asano et al. (2010);
Busemeyer et al. (2006); Busemeyer and Wang (2007); Busemeyer et al. (2009);
Khrennikov (2009); Khrennikov and Haven (2009); Khrennikov (2011)).
To try to describe quantum-like effects in cognitive sciences, in a recent paper
Khrennikov (2011) used a classical electromagnetic-field model for the mental
processing of information. In it, Khrennikov argues that the collective activity
of neurons involved in mental processing produces electromagnetic signals. Such
signals then propagate in the brain, leading to interference and the appearance
of quantum-like correlations. He used this model to show how quantum-like
effects help understand the binding problem.
In this paper we set forth a neural oscillator model exhibiting many of the
characteristics of Khrennikov's. The neural model we employ was initially pro-
posed as an attempt to obtain behavioral stimulus-response theory from the
neuronal activities in the brain (Suppes et al., 2012). There are many different
ways to use neural oscillators to model learning (see Vassilieva et al., 2011,
and references therein), but to our knowledge, this is the only one that makes
a clear connection between neural oscillator computations and the behavioral
responses. Furthermore, the model is simple enough to help understand the
basic principles behind the computation of a response from a stimulus without
compromising the neurophysiological interpretation of it. This must be empha-
sized, as most neural models constructed from a neurophysiological basis are so
computationally intensive and complex that it is very hard to understand why
or how one response was select instead of another. We shall see later that this
is not the case with Suppes et al. (2012).
As with Khrennikov's electromagnetic field model, our model has many
quantum-like characteristics, such as signaling, superposition, and interference.
However, it does not rely on the propagation of electromagnetic fields in the
brain. Instead, we show that these quantum-like effects could be obtained by
directly considering the interference of the activities of collections of firing neu-
rons coupled via inhibitory and excitatory synapses. In particular, we show a
violation of Savage's sure-thing principle, often cited as an example of quantum-
like behavior.
This paper is organized as follows.
In section 2 we introduce Suppes, de
Barros, and Oas's (2012) oscillator model, with special emphasis to its neu-
rophysiological interpretation and connection to SR theory. Then, in section
3.1, we present a contextual theory of quantum-like phenomena coming from
classical interference of waves (de Barros and Suppes, 2009). Finally, in section
2 we show computer simulations of neural oscillators and how they fit some
quantum-like experimental data in the literature. We end with some remarks.
2. Neural oscillator model of SR-theory
In this section, we delineate the modeling of SR theory by neural oscillators.
This section is divided into three parts. First, in subsection 2.1 we sketch the
mathematical version of behavioral SR theory we model. Then, in subsection
2
2.2 we present the model of Suppes et al. (2012) in a qualitative way. Finally,
in subsection 2.3 we give a mathematical description of the oscillator model.
Our goal is not to be exhaustive, but only to focus on the main features of
the model that are relevant for the emergence of quantum-like effects derived
later. Readers interested in more detail are referred to (Suppes et al., 2012;
de Barros et al., 2012).
2.1. Review of SR theory
Stimulus-response theory (or SR theory) is one of the most successful be-
havioral learning theories in psychology (Suppes and Atkinson, 1960). Among
the many reasons for its success, it has been shown to fit well empirical data
in a variety of experiments. This fitting requires few parameters (the learning
probability c and the number of stimuli), giving it strong prediction powers from
few assumptions. Also, because it requires a rigid trial structure, its concepts
can be formally axiomatized, resulting in many important non-trivial experi-
ments. Finally, despite its simplicity, SR theory is rich enough such that even
language can be represented within its framework (Suppes, 2002). Thus, SR
theory presents an ideal mathematical structure to be modeled at a neuronal
level1.
Here we present the mathematical version of SR theory for a continuum of re-
sponses, formalized in terms of a stochastic process (here we follow Suppes et al.,
2012). Let (Ω,F , P ) be a probability space, and let Z, S, R, and E be random
variables, with Z : Ω → ES S : Ω → S, R : Ω → R, and E : Ω → E, where
S is the set of stimuli, R the set of responses, and E the set of reinforcements.
Then a trial in SR theory has the following structure:
Zn → Sn → Rn → En → Zn+1.
(1)
Intuitively, the trial structure works the following way. Trial n starts with a
certain state of conditioning and a sampled stimulus. Once a stimulus is sam-
pled, a response is computed according to the state of conditioning. Then,
reinforcement follows, which can lead (with probability c) to a new state of con-
ditioning for trial n + 1 . In more detail, at the beginning of a trial, the state
, . . . , z(n)
1
m (cid:17).
, . . . , z(n)
m (cid:17) associates to each stimuli si ∈ S, i = 1, . . . , m,
of conditioning is represented by the random variable Zn = (cid:16)z(n)
The vector (cid:16)z(n)
where m = S is the cardinality of S, a value z(n)
Sn = si is sampled with probability P (Sn = sisiǫS) =
z(n)
i
i (cid:17),
K(cid:16)rz(n)
i.e. P (cid:16)a1 ≤ Rn ≤ a2Sn = si, Zn,i = z(n)
determines the probability of responses in R by the probability distribution
on trial n. Once a stimulus
i (cid:17) = ´ a2
a1
k(cid:16)xz(n)
i (cid:17) dx,
1
i
1
m
, its corresponding
1SR theory, in the form presented here, has lost most of its support among learning theo-
rists, mainly because of of inability to satisfactorly explain certain experiments (Mackintosh,
1983). However, many of its theorems are relevant to current neurophysiological experiments.
3
where k(cid:16)xz(n)
where Zn,i is the i-th component of the vector (cid:16)z(n)
distribution K(cid:16)rz(n)
variance and mode z(n)
fective with probability c, i.e. P (cid:16)Zn+1,i = ySn = si, En = y, Zn,i = z(n)
and P (cid:16)Zn+1,i = z(n)
i (cid:17) is the probability density associate to the distribution, and
m (cid:17). The probability
i (cid:17) is the smearing distribution, and it is determined by its
i (cid:17) = c
i (cid:17) = 1− c. The trial ends with
Sn = si, En = y, Zn,i = z(n)
, . . . , z(n)
1
. The next step is the reinforcement En, which is ef-
i
a new (with probability c) state of conditioning Zn+1.
i
2.2. Qualitative description of oscillator model
Because of its focus on behavioral outcomes, mathematical SR theory does
not provide a clear connection to the internal processing of information by the
brain. To bridge this gap, Suppes et al. (2012) proposed a neural-oscillator
based response computation model able to reproduce the main stochastic fea-
tures of SR theory, including the conditional probabilities for a continuum of
responses. It is this model that we describe qualitatively in this subsection.
In SR theory, a trial is given by the sequence in (1). In the oscillator model,
with the exception of Zn, each random variable is represented by a corresponding
oscillator. To better understand the model, let us examine each step of an SR
trial in terms of oscillators.
Let us start with the sampling of a stimulus, Sn. At the beginning of a trial,
a (distal) stimulus is presented. Through the perceptual system processing, it
reaches an area of the brain where it produces a somewhat synchronized firing
of a set of neurons. This set of firing neurons is what we think as the activation
of the sampled stimulus oscillator2, Os (t). The oscillator Os (t) is a simplified
description of the semi-periodic activity of the collection of (perhaps thousands
of) neurons.
We now turn to the computation of a response, Rn. For simplicity, we first
consider two possible responses, R1 and R2 (i.e. R = {R1, R2}); later we show
how we generalize to any number of to a continuum of responses. Once Os (t) is
activated, its couplings to other neurons may lead to a spreading of its activation
to other sets of neurons, including the response neurons. Say we can split the
neurons activated by Os (t)'s couplings into two sets, one representing R1 and
the other R2. Since those neurons were activated, as the stimulus neurons, they
fire in relative synchrony, and can be represented by the oscillators Or1 (t) and
Or2 (t). Of course, the activation of both Or1 (t) and Or2 (t) is not enough for the
computation of a response. To see how a response is computed in the oscillator
model from Or1 (t) and Or2 (t), we need think of what happens if two oscillators
are in-phase or out-of-phase. When two neural oscillators are in phase, the
firing of one adds to the firing of the other, perhaps reaching a certain response
2We limit our discussion to a single stimulus oscillator, but its generalization to N stimuli
is straightforward.
4
threshold; if they are out of phase, such threshold may not be reached, as the
firings do not add above their activated intensities. So, a response is chosen
depending on the relative phases of the oscillators. If Os (t) is in phase with
Or1 (t) and out of phase with Or2 (t), response R1 is selected. Which response
is selected depends on the actual couplings between Os (t), Or1 (t), and Or2 (t)
(see next subsection for details).
We can now talk about En, Zn, and Zn+1. Since the selected response is
determined by the couplings between neural oscillators, clearly the state of con-
ditioning Zn depends on those couplings. Therefore, on the oscillator model,
saying that a trial starts with a state of conditioning Zn is the same as say-
ing that a trial starts with a set of synaptic connections between the neurons
belonging to the ensembles representing the activated stimulus and responses.
But, as the reinforcement En may change Zn into a different state of condi-
tioning, Zn+1, it follows that during reinforcement the neural couplings need to
change. So, in the oscillator model, such couplings change when a neural rein-
forcement oscillator (corresponding to the activation of an ensemble of neurons
representing reinforcement) interacts with the stimulus and response oscillators,
thus driving a Hebb-like change in synaptic connection, resulting in learning.
The previous paragraphs summarize qualitatively the model, but there are
three points that need to be further discussed before we proceed to the quan-
titative description. First, since neural oscillators correspond to a tremendous
dimensional reduction in the neuronal dynamical system, we need to justify their
use. Second, as it is crucial for the model, an argument needs to be made as to
how oscillators synchronize. Finally, we should the two-response case, but how
can a finite number of oscillators encode a large number of responses, possibly
a continuum of responses?
Why neural oscillators? First, we note that Suppes et al. (2012) does not
propose the use of neural oscillators for any type of modeling. Instead, they
argue that if we are interested in higher cognitive functions, such as language,
it can be showed that they can be represented on electroencephalogram (EEG)
oscillation patterns (see Suppes, 2002, Chapter 8, and corresponding refer-
ences). But EEG signals are the result of a large assembly of neurons firing
synchronously or quasi-synchronously (Nunez and Srinivasan, 2006). Thus, we
can think of a word as represented in the brain by a set of coupled neurons.
When this word is activated by an external stimulus, this set of neurons fire
coherently. Such an ensemble of synchronizing neurons could, in first approxi-
mation, be represented by a periodic function FT (t), where T is the period of
the function3. So, for Suppes et al. (2012) a neural oscillator is simply a set of
firing neurons represented by such periodic functions, and a good candidate as
a unit of processing of information in the brain.
It must be emphasized that representing a collection of neurons by a periodic
function, as we did above, does not mean that all neurons fire simultaneously.
3We remark that neurons in this ensemble do not need to fire at the same time, but only
that their frequencies synchronize. They may still be in synchrony but out of phase.
5
It just means that the firing frequencies of individual neurons become very close
to each other. But neurons with the same frequency do not need to fire at
the same times. In other words, they may synchronize out of phase. But, more
importantly, once they synchronize, the initial dynamical system with thousands
of degrees of freedom become describable by a single dynamical variable, the
periodic function FT (t). It is this dramatic dimensional reduction that allows
us to produce a model whose behavior we can understand in certain specific
situations.
We now move to the question of how neural oscillators synchronize. Let us
first look at the qualitative behavior of individual neurons in two ensembles,
A and B, each with a large number of neurons firing coherently. Let nA be a
single neuron in ensemble A firing with period TA, and let nB be another neu-
ron in B firing with period TB. Let t′ be a particular time such that nB would
fire if A were not activated, i.e. the time it would fire because of its natural
periodicity TB. If nA and nB are synaptically coupled, a firing of nA at a time
t / t′ would anticipate nB's firing to a time t′′ closer to t, thus approaching
the timing of firing for both neurons4. If more neurons from ensemble A are
coupled to nB, the stronger the effect of anticipating its firing time. Further-
more, since nB is also coupled to neurons in A, its firing has the same effect
on A (though this effect does not need to be symmetric). So, the couplings
between A and B lead both periods TA and TB to approach each other, i.e., A
and B synchronize. In fact, it is possible to prove mathematically that, under
the right conditions, if the number of neurons is large enough, the sum of the
several weak synaptic interactions can cause a strong effect, making all neurons
fire close together (Izhikevich, 2007). In other words, even when weakly coupled,
two neural ensembles represented by oscillators may synchronize.
We end this subsection with the representation of a continuum of responses in
terms of neural oscillators. Continuum of responses show up in some behavioral
experiments. For example, a possible task could be to predict the position of an
airplane on a radar screen (represented by a flashing dot appearing periodically)
after observing its position for several trials. Since the radar position can be
anywhere in the screen, this cannot be represented by a discrete set of responses.
So, how could we represent a continuum of responses with neural oscillators?
With a discrete number of responses, what we did was to create a representation
for each response in terms of a distinct oscillator (in our example above, Or1 (t)
and Or2 (t)). But this approach would lead to an uncountable number of neural
oscillators for the continuum case, clearly a physically unreasonable assumption.
But there is another alternative. When we used oscillators for the two responses,
we thought about Or1 (t) and Or2 (t) as being either in phase or out of phase with
Os (t). However, they do not need to be in perfect phase. In the next section,
we show that by using specific differences of phase between Or1 (t) and Or2 (t)
and Os (t), we can control the amount of "intensity" of a neural activation on
4This is true for excitatory synapses. As we shall see later on this paper, inhibitory synapses
will play a different yet important role.
6
each oscillator. This relative intensity between Or1 (t) and Or2 (t) can be used
as a measure of how close a response is to either R1 or R2; in other words, we
can think of the response as being between R1 and R2, which would correspond
to a continuum. So, by thinking of R1 and R2 as extremes in a possible range of
responses in a continuum interval, the relative intensities of Or1 (t) and Or2 (t),
given by their phase relation with Os (t), could code any response in between.
2.3. Quantitative description of the oscillator model
We now turn to a more rigorous mathematical description. Our underlying
assumption is that we have two neural ensembles A and B described by periodic
functions FTA (t) and FTB (t) with periods TA and TB, respectively. Without
loss of generality in our argument, we assume that FTA (t) and FTB (t) have
the same shape, i.e., there exists a constant c such that FTA (t) = FTB (tc).
A simple way to describe the synchronization of A and B is to rewrite the
time argument in the functions, rewriting them as FTA (tαA) and FTB (tαB).
Clearly, if no couplings exist between A and B, their dynamics is not modified
and αA = αB = 1. However, when the ensembles are couples, αA and αB
are a function of time, their synchronization means that they evolve such that
FTA (tαA (t)) = FTB (tαB (t)). Thus, when we are studying the dynamics of
coupled neural oscillators, it suffices to study the dynamics of the arguments of
FTA and FTB above, which we call their phases.
To illustrate this point, let us imagine a simple case where when there is no
interaction A follows a harmonic oscillator with angular frequency ωA = 2π/TA.
Then FTA (t) = A cos (ωAt + δA), where and δA is a constant, and expressions
hold for FTB (t). Following the above paragraph, we rewrite
FTA (t) = A cos (ϕA (t)) ,
where
ϕA (t) = ωAt + δA
(2)
is the phase. Let us now focus on the phase dynamics.
For a set of coupled phase oscillators, Kuramoto (1984) proposed a simple set
of dynamical equations for ϕA (t). To better understand them, let us assume
that in the absence of interactions ϕA evolves according to (2). Thus, the
differential equation describing ϕA would be
and for ϕB,
dϕA (t)
dt
= ωA,
dϕB (t)
dt
= ωB.
(3)
(4)
In the presence of a synchronizing interaction, Kuramoto proposed that equa-
tions (3) and (4) should be replaced by
dϕA (t)
dt
= ωA − kAB sin (ϕA (t) − ϕB (t)) ,
(5)
7
and for ϕB,
dϕB (t)
dt
= ωB − kBA sin (ϕB (t) − ϕA (t)) .
(6)
It is easy to see how equations (5) and (6) work. When the phases ϕA and ϕB
are close to each other, we can linearize the sine term, and rewrite equations
(5) and (6) as
and for ϕB,
dϕA (t)
dt ∼= ωAt − kAB (ϕA (t) − ϕB (t)) ,
dϕB (t)
dt ∼= ωBt − kBA (ϕB (t) − ϕA (t)) .
(7)
(8)
The effect of the added term in equation (7) is to make the instantaneous fre-
quency increase above its natural frequency ωA if ϕB is greater than ϕA, and
decrease otherwise. In other words, it pushes the phases ϕA and ϕB towards
synchronization. We should emphasize that Kuramoto's equations were pro-
posed for a large number of oscillators, and his goal was to find exact solutions
for such large systems. However, as we described above, his equations also
make sense for systems with small number of oscillators (for applications of
Kuramoto's equations to systems with small number of oscillators, in addition
to Suppes et al. (2012), see also Billock and Tsou (2005, 2011); Seliger et al.
(2002); Trevisan et al. (2005) and references therein). For a set of N coupled
phase oscillators, Kuramoto's equations are
dϕi (t)
dt
= ωi −
N
Xj=1
j6=i
kij sin (ϕi (t) − ϕj (t)) .
(9)
But independent of whether we have large or small number of oscillators or not,
it is possible to prove that for oscillating dynamical systems near a bifurcation
(which is the case for many neuronal models), Kuramoto's equations present a
first order approximation for the synchronizing coupled dynamics (Izhikevich,
2007).
In the model proposed in (Suppes et al., 2012), each stimulus in the set S of
stimuli corresponds to an oscillator, i.e. there are m stimulus neural oscillators,
{si (t)}, i = 1, . . . , N . Here we use the notation si (t) to distinguish between the
oscillator and the actual stimulus si ∈ S. Once a distal stimulus is presented,
the processing of the proximal stimulus leads, after sensory processing, to the
activation of neurons in the brain corresponding to a neural oscillator repre-
sentation of such stimulus. Once a stimulus oscillator is activated by starting
to fire synchronously, a set of response oscillators, r1 (t) and r2 (t), connected
to the active stimulus oscillator via the couplings ksi,rj and kr1,r2 lead to the
synchronization of the stimulus and response oscillators. The couplings ksi,rj
and kr1,r2 correspond to the state of conditioning in SR-theory. Depending on
the details of the synchronization dynamics, as explained below, a response is
selected (this being the equivalent of sampling Xn). Finally, during reinforce-
ment, a reinforcement oscillator ey (t) is activated, and its couplings with si (t),
8
r1 (t), and r2 (t), together with ksi,rj and kr1,r2, leads to a dynamics that allow
for changes in ksi,rj and kr1,r2, which accounts for the last step in the SR trial
(1), with a new state of conditioning.
Before we detail the dynamics of the oscillator model, it is important to
discuss how si (t), r1 (t), and r2 (t) can encode a response through their cou-
plings, and how can we interpret such couplings. Let us write the case where
si is sampled, and let us assume that once the dynamics is acting, all oscilla-
tors synchronize, acquiring the same frequency (though perhaps with a phase
difference). Then,
si(t) = A cos (ϕsi (t)) = A cos (ω0t) ,
r1(t) = A cos (ϕr1 (t)) = A cos (ω0t + δφ1) ,
r2(t) = A cos (ϕr2 (t)) = A cos (ω0t + δφ2) ,
(10)
(11)
(12)
where s(t), r1(t), and r2(t) represent harmonic oscillations, ϕs(t), ϕr1 (t), and
ϕr2(t) their phases, δφ1 and δφ2 are constants, and A their amplitude, assumed
to be the same. As argued in (Suppes et al., 2012), since neural oscillators have
a wave-like behavior (Nunez and Srinivasan, 2006), their dynamics satisfy the
principle of superposition, thus making oscillators prone to interference effects.
As such, the mean intensity in time gives us a measure of the excitation of the
neurons in the neural oscillators. For the superposition of s(t) and r1(t),
where
I1 = D(si(t) + r1(t))2Et
= (cid:10)s(t)2(cid:11)t +(cid:10)r1(t)2(cid:11)t + h2s(t)r1(t)it ,
hf (t)it = lim
T →∞
1
T T
0
f (t) dt
is the time average. It is easy to compute that
and
I1 = A2 (1 + cos (δφ1)) ,
I2 = A2 (1 + cos (δφ2)) .
From the above equations, the maximum intensity for each superposition is 2A2,
and the minimum is zero. Thus, the maximum difference between I1 and I2
happens when their relative phases are π. We also expect a maximum contrast
between I1 and I2 when a response is not in between the responses represented
by the oscillators r1 (t) and r2 (t). For in between responses, we should expect
less contrast, with the minimum contrast happening when the response lies on
the mid-point of the continuum between the responses associated to r1(t) and
r2(t). The ideal balance of responses happen if we impose
δφ1 = δφ2 + π ≡ δφ,
9
(13)
which results in
and
I1 = A2 (1 + cos (δφ)) ,
(14)
(15)
From equations (14) and (15), let x ∈ [−1, 1] be the normalized difference in
intensities between r1 and r2, i.e.
I2 = A2 (1 − cos (δφ)) .
x ≡
I1 − I2
I1 + I2
= cos (δϕ) ,
(16)
0 ≤ δϕ ≤ π. The parameter x codifies the dispute between response oscillators
r1 (t) and r2 (t). So, we can use arbitrary phase differences between oscillators
to code for a continuum of responses between −1 and 1.
As we mentioned above, the dynamics of the phase oscillators leading to
(10) -- (12) are modeled here by Kuramoto equations (9). However, as we also
mentioned, (9) leads to the synchronization of all oscillators with the same
phase, which is not what we need to code responses according to equation (16).
Naturally, Kuramoto's equations need to be modified to encode the phase dif-
ferences δϕ in (10) -- (12). This can be accomplished by adding a term inside the
sine, resulting in
dϕi (t)
dt
= ωi −
N
Xj=1
j6=i
kij sin (ϕi (t) − ϕj (t) + δij ) .
(17)
The main problem with such change is how to interpret it. Equation (9) had a
clear interpretation: the firing of oscillator neurons coupled through excitatory
synapses displaced the phase of the other oscillator into synchrony. But this
interpretation does not make sense for (17). Why would oscillators be brought
close to each other, but be kept at a phase distance of δij?
To understand the origin of δij , let us rewrite (17) as
dϕi
dt
= ωi −
N
Xj=1
j6=i
kij cos (δij ) sin (ϕi − ϕj)
−
N
Xj=1
j6=i
kij sin (δij) cos (ϕi − ϕj) .
(18)
Since the terms involving the phase differences δij are constant, we can write
(18) as
dϕi
dt
= ωi −
N
Xj=1
j6=i
(cid:2)kE
ij sin (ϕi − ϕj) + kI
ij cos (ϕi − ϕj)(cid:3) ,
(19)
10
ij ≡ kij cos (δij) and kI
where kE
ij ≡ kij sin (δij). Equation (17) now has a clear
ij makes oscillators ϕi (t) and ϕj (t) approach each other, kE
interpretation: kE
ij
makes them move further apart. Thus, kE
ij corresponds to excitatory couplings
between neurons and kI
ij to inhibitory couplings. In other words, we give mean-
ing to equations (17) by rethinking their couplings in terms of excitatory and
inhibitory neuronal connections.
To summarize the response process, once a stimulus oscillator si (t) is acti-
vated at time ts,n on trial n, the response oscillators are also activated. Because
of the excitatory and inhibitory couplings between stimulus and response os-
cillators, after a certain time their dynamics lead to their synchronization, but
with phase differences given by (13). Here we point out that, due to stochas-
tic variations of biological origin, the initial conditions si (ts,n), r1 (ts,n), and
r2 (ts,n) vary according to the distribution
f (ϕi) =
1
σϕ√2π
exp(cid:18)−
ϕi
2σ2
ϕ(cid:19) .
(20)
Therefore, the phase differences at the time of response, tx,n, may not be exactly
the ones given in (13), which gives an underlying explanation for a component
of the smearing distribution5. Interference between oscillators, determined by
their phase differences, leads to a relative intensity that codes a continuum
of responses according to x in equation (16). The variable x is the response
computed by the oscillators.
We now turn to learning. Here we will only present the main features of
learning relevant to this paper. The actual details are numerous, and we refer
ij and kI
the to (Suppes et al., 2012). For learning to happen, the couplings kE
ij
need to change during reinforcement. Since reinforcement must have a brain
representation in terms of synchronously firing neurons, during reinforcement
a neural oscillator ϕy (t) gets activated. We assume that ϕy (t) is a fixed rep-
resentation in the brain, with frequencies that are independent of the stimulus
and response oscillators, and therefore can be represented by
ϕy (t) = ωet.
Furthermore, since ϕy is coding a response y being reinforced, from (16) the
phase of the oscillators must be related to y via
δϕ = arccos y.
5We note that we here only model the brain computation of a stimulus and response, and
not the representation of the distal stimulus into brain oscillators and the representation of
responses into actual muscle movement or sound production. Of course, those extra steps
between the stimulus and the response will lead to further smearing of the actual observed
distribution, and a detailed brain computation of K (cid:16)rz
i (cid:17) should include such factors. The
interested reader should refer to Suppes et al. (2012) for details about the used parameters
and the fitting of the model to experimental data.
(n)
11
During the reinforcement period, the (active) stimulus and response oscillators
evolve according to the same equations as before, but now with an extra term
due to the reinforcement oscillator, according to
dϕi
dt
= ωi − Xsi,r1,r2
kE
ij sin (ϕi − ϕj)
− Xsi,r1,r2
kE sin (ϕi − ϕj )
−K0 sin(cid:0)ϕsj − ωet + ∆i(cid:1) ,
where ∆ij reflects the reinforced response y,
(21)
∆i = − (δi,r1 + δi,r2) arccos y + δi,r2 π,
and where δi,j is Kronecker's delta. During reinforcement, couplings also change
according to a Hebbian rule represented by the equations (Seliger et al., 2002)
and
where
dkE
ij
dt
dkI
ij
dt
= ǫ (K0)(cid:2)α cos (ϕi − ϕj) − kE
ij(cid:3) ,
= ǫ (K0)(cid:2)α sin (ϕi − ϕj) − kI
ij(cid:3) ,
ǫ (K0) = (cid:26) 0 if K0 < K ′
ǫ0 otherwise,
(22)
(23)
(24)
where ǫ0 ≪ ω0, α and K0 are constant during ∆te, and K ′ is a threshold
constant throughout trials (Hoppensteadt and Izhikevich, 1996a,b). We also
assume that that there is a normal probability distribution governing the cou-
pling strength K0 between the reinforcement and the other active oscillators. It
has mean K 0 and standard deviation σK0 . Its density function is:
f (K0) =
1
σK0√2π
exp(cid:26)−
1
2σ2
K0 (cid:0)K0 − K 0(cid:1)2(cid:27) .
(25)
So, when a reinforcement oscillator is activated, the system evolves under the
coupled set of differential equations (21), (22), and (23).
To summarize, we described a model of SR theory in terms of neural oscilla-
tors. In such a model, stimulus and responses are represented by collections of
neurons firing in synchrony. Such collections of neurons, when coupled, synchro-
nize to each other. The relative phases of synchronization lead to interference
of sets of neurons, and such interference results in the coding of a continuum of
responses. We left our many of the details of the stochastic processes involved
in this model, as they are not necessary for this paper, but details can be found
in reference (Suppes et al., 2012).
12
3. Quantum-like neural oscillator effects
In this section we discuss possible quantum-like effects in the brain and
how they can be understood in terms of oscillators. We start by discussing
what we believe are quantum-like effects in the brain. We then describe how
quantum-like effects show up in decision-making experiments. Finally, we show
that neural oscillators can reproduce such effects.
3.1. What are quantum-like effects?
As we mentioned earlier, quantum mechanics is one of the most successful
theories in history. However, such success did not come without a price, as
quantum mechanics presents a view of the world that most people would find
disturbing, to say the least. This was certainly the view of many of the founders
of quantum mechanics, such as Plank, Einstein, and Pauli. At the core of their
concerns were three characteristics of quantum mechanics: nondeterminism,
contextuality, and nonlocality. Let us examine each one of those characteristics,
and see how they can be relevant to brain processes.
We start with nondeterminism. Intuitively, a dynamical system is determin-
istic if the current state of the system completely determines the future state of
such system, and it is nondeterministic otherwise. Quantum mechanics is non-
deterministic because a quantum state only gives the probabilities for the values
of certain observables during a measurement process. Of course, nondetermin-
ism is not unique to quantum mechanics. For practical purposes, there are many
nondeterministic processes, such as the famous Brownian motion, or the tossing
of a coin. But when we toss a coin or throw a die, physicists believe that the
newtonian dynamics would allow for the computation of the final state of the
die or coin, if we were to know exactly the initial conditions, i.e., the state of the
system (Ford, 1983; Vulović and Prange, 1986). In other words, before quantum
mechanics physicists used to believe that the underlying dynamics for nonde-
terministic processes was actually deterministic; we just could not distinguish
the deterministic from the nondeterministic because we could not know enough
details about the system. But whether we can distinguish between a determin-
istic and stochastic dynamics is not the important issue (in fact sometimes we
can't; see Suppes and de Barros (1996) for detail). The main point is that for
quantum theory the underlying dynamics is essentially nondeterministic.
Nondeterministic stochastic processes are often used in cognitive modeling
(Busemeyer and Diederich, 2010). For example, the stochastic SR theory men-
tioned above is nondeterministic, as conditioning happens probabilistically. Yet,
as in the coin tossing example, there is no reason to believe that the underlying
dynamics must be nondeterministic, as its stochasticity may come from our lack
of detailed knowledge about the brain or its biological noises (Josić et al., 2009).
So, it should be clear that when we talk about quantum-like brain processes,
nondeterminism should not play a prominent role. Insofar as nondeterminism is
concerned, there is no need to add further quantum-mechanical mathematical
or conceptual machinery to account for it in the brain.
13
Now let us examine nonlocality. To discuss nonlocality, we should first dis-
cuss causality, albeit briefly (interested readers are referred to Suppes (1970)).
Let A and B be two events, and let A and B be ±1-random variables corre-
sponding to whether A and B occurred (+1 if occurred, −1 otherwise). We
say that C is a cause of A if P (A = 1C = 1) > P (A = 1) and there not
exist a spurious cause (represented by a hidden variable D) that could account
for the correlations. Now, it is possible to prove that for some quantum sys-
tems, there are space-like events A and A′ such that A is a the cause of A′.
What is key in the previous statement is that A and A′ are space-like separated
events; in other words, whatever mechanism is making A′ influence A must be
superluminal6. This, it may be argued, is the main puzzling characteristic of
quantum processes: to understand them we need to accept superluminal cau-
sation7. However, we believe that quantum-like effects in the brain are not the
outcome of superluminal causation. The brain is small, of the order of tens
of centimeters, and an electromagnetic field would only need of the order of
10−10 seconds to travel the whole brain. Given that brain processes are orders
of magnitude slower than 10−10 seconds, no phenomena in the brain could not
be accounted for non-superluminal causality. In other words, we doubt it would
be possible to create an experiment where nonlocal effects in the brain would
be detected in a superluminal way.
Here we need to differentiate our use of nonlocality from what is some-
times found in the literature on quantum cognition (Nelson et al., 2003). Some-
times, authors consider a violation of some form of Bell's inequalities (Bell, 1964,
1966) as evidence of nonlocality. However, from our discussion, the correlated
quantities must be measured such that no superluminal signal explaining the
correlations is possible, otherwise a violation of Bell's inequalities would only
imply a contextuality (Suppes et al., 1996a). Classical fields, such as the electro-
magnetic fields used by Khrennikov, also exhibit contextuality, violating Bell's
inequalities and not allowing for certain configurations a joint probability distri-
bution (Suppes et al., 1996b). Here we are using nonlocality in the strict sense
of "spooky" action-at-a-distance correlations that cannot be explained without
using superluminal interactions.
This brings us to the last quantum-like issue we wish to discuss: contex-
tuality. Early on, the wave description of a particle led physicists to realize,
through Fourier's theorem, that it was impossible to simultaneously assign val-
ues of momentum and position to it. This developed into the concept of comple-
mentarity. Without trying to be too technical, we can say that two observables
are complementary when they cannot be simultaneously observed, i.e., when
the corresponding hermitian operators O1 and O2 for such properties does not
6We are summarizing a very complex and subtle discussion in a few sentences. The reader
interested in more details and in a mathematically rigorous treatment is directed to Suppes
(2002) and references therein.
7Or abandon realism, in the sense that the act of measurement of an observer (a free choice
of the experimenter) determines the existence of the quantity being measured (though realist
theories may also have other problems; see de Barros et al. (2007)).
14
commute ([ O1, O2] 6= 0). But the question is not whether we can measure O1
and O2 simultaneously, but whether we can assign those properties values even
when we cannot measure them. In other words, can we generate a data table
that fills out the values of O1 and O2 for every trial? If the number of observ-
ables is large enough, the answer to that question was given in the negative
first by Bell (1966), who showed that any attempts to fill out such tables would
result in values inconsistent with those predicted by quantum mechanics, and
later by Kochen and Specker (1975). In other words, there are quantum systems
whose observables cannot be assigned values when we do not measure them, and
therefore we cannot provide a consistent joint probability distribution for such
observables (de Barros and Suppes, 2000, 2001, 2010; de Barros, 2011)8.
This impossibility of defining a joint probability distribution is related to the
change of values of the random variables when contexts change. For example, a
random variable O1 representing O1 the outcomes of an experiment may not be
the same random variable when O2 is being measured. Such random variables
change with the context. Contextuality is not new in physics, nor in psychology.
It happens in classical systems, such as with the classical electromagnetic field
(de Barros and Suppes, 2009). However, we claim it is contextuality, with its
associated impossibility of defining joint probability distributions, that leads to
quantum-like events in brain processing.
At this point, it is important to raise the issue of how much of the quantum
mechanical apparatus is needed to describe brain processes. Quantum mechan-
ics brings to the table more than just contextuality, nonlocality, and nondeter-
minism. It also brings an additional mathematical structure, formalized by the
algebra of observables on a Hilbert space, initially advocated by von Neumann
(1996). To see this, let us examine the following case, analyzed in details by
Suppes and Zanotti (1981); Suppes et al. (1996a). Let X, Y, and Z be three
±1-valued random variables with observed pairwise joint expectations given by
E (XY) = E (XZ) = E (YZ) = −ǫ,
where ǫ > 1/3. Suppes and Zanotti (1981) showed that such expectations are
not consistent with the existence of a joint probability distribution for X, Y,
and Z if
−1 ≤ E (XY) + E (XZ) + E (YZ) ≤ 1 + 2 min{E (XY) , E (XZ) , E (YZ)} .
Thus, if they represent the outcomes of local measurements, they are contextual.
If, on the other hand, they each represent space-like separated measurement
events, they are non-local. And each random variable is indistinguishable from a
tossed coin, and are therefore nondeterministic9. However, X, Y, and Z cannot
8Though with the redefinition of what constitutes a particle, local models reproducing
many of the characteristics of quantum mechanics non-locality are possible. See, for example,
Suppes and de Barros (1994b,a); Suppes et al. (1996c,d) for a particular approach.
9The distinction between nondeterministic and deterministic is a complicated and subtle
one, c.f. Suppes and de Barros (1996), and we refrain from it in this paper. The interested
reader should also refer to Werndl (2009).
15
represent quantum mechanical observables, as, since they all commute, we could
in principle simultaneously observe X, Y, and Z, which is a sufficient condition
for the existence of a joint probability distribution10. But we could, on the
other hand, imagine a social-science or psychology example (albeit probably a
contrived one) where correlations such as the ones exhibited above could appear.
Furthermore, as shown by de Barros (2012), such random variable correlations
can be obtained from the same neural oscillator model from section 2.
In this section we discussed the different ways in which quantum-like pro-
cesses may show up in the brain. We argued that among them, the most prob-
able is contextuality. Contextuality may appear in a system for several rea-
sons. For example, certain systems evolve according to a dynamic dependent
on boundary conditions, and those may change with the context. Additionally,
a complex system may have different parts that can interfere, changing their
behavior due to alterations in the conditions of other parts. The latter pos-
sibility is mainly what Khrennikov (2011) explored in his field model. In the
next section, we will show that the complex interaction of neural oscillators are
another plausible explanation for quantum-like behavior in the brain.
3.2. Quantum-like effects in decision-making
The experimental violation of Savage's sure-thing principle (STP) is con-
sidered an example of a quantum-like decision making, so, let us start with a
description of the STP. Savage presents the following example.
"A businessman contemplates buying a certain piece of property. He
considers the outcome of the next presidential election relevant to the
attractiveness of the purchase. So, to clarify the matter for himself,
he asks whether he should buy if he knew that the Republican can-
didate were going to win, and decides that he would do so. Similarly,
he considers whether he would buy if he new that the Democratic
candidate were going to win, and again finds that he would do so.
Seeing that he would buy in either event, he decides that he should
buy, even though he does not know which event obtains, or will ob-
tain, as we would ordinarily say. It is all too seldom that a decision
can be arrived at on the basis of the principle used by this business-
man, but, except possibly for the assumption of simple ordering, I
know of no other extralogical principle governing decisions that finds
such ready acceptance." (Savage, 1972, pg. 21)
The idea illustrated is that if X (buying) is preferred to Y (not buying) under
condition A (Republican wins) and also under condition ¬A (Democrat wins),
then X is preferred to Y . Savage calls this the sure-thing principle.
10This comes from the fact that, since they commute with each other, there exists a set of
orthogonal base vectors in the Hilbert space representation such that the observables corre-
sponding to X, Y, and Z, namely X, Y , and Z, are all diagonal in this base.
16
Putting it in a more formal way, consider the following three propositions,
A, X, and Y , and let P (XA), the conditional probability of X given A, rep-
resent a measure of a rational belief on whether X is true given that A is true.
Since the measure P requires rationality of belief, it follows that it satisfies Kol-
mogorov's axioms of probability (Jaynes, 2003; Galavotti, 2005). We start with
the assumption that
This expression can be interpreted as stating that if A is true, then X is preferred
over Y. If we also assume
P (XA) > P (Y A).
P (X¬A) > P (Y ¬A),
then we obtain, multiplying each inequality by P (A) and P (¬A), respectively,
P (XA)P (A) + P (X¬A)P (¬A) > P (Y A)P (A) + P (Y ¬A)P (¬A) .
From P (A&¬A) = 1 and the definition of conditional probabilities we have
P (X) > P (Y ).
In other words, if we prefer X from Y when A is true, and we also prefer X
from Y when ¬A is true, then we should prefer X over Y regardless of whether
A is true. This, of course, is equivalent to the Savage's example.
Though STP should hold if agents are making rational decisions, Tversky and Shafir
(1992) and Shafir and Tversky (1992) showed that it was violated by decision
makers. In Tversky and Shafir (1992), they presented several simple decision-
making problems to Stanford students. For example, in one question students
were told about a game of chance, to be played in two steps. In the first step,
not voluntary, players had a 50% probability of winning $200 and 50% of loosing
$100. In the second step, a choice needed to be made: whether to gamble once
again or not. If a player accepted a second gamble, the same odds and payoffs
would be at stake. When told that that they won the first bet, 69% of subjects
said they would gamble again on the second step. When told they lost, 59% of
the subjects also said they would gamble again. Following the discussion above,
we translate the experiment into the following propositions.
and we have that
and
A = "Won first bet,"
¬A = "Lost first bet,"
X = "Accept second gamble,"
Y = "Reject second gamble,"
P (XA) = 0.69 > P (Y A) = 0.31,
P (X¬A) = 0.59 > P (Y ¬A) = 0.41,
17
since P (X) = 1 − P (Y ), as X = ¬Y . Clearly, X is preferred over Y regardless
of A. Later on the semester, the same problem was presented, but this time
students were not told whether the bet was won or not in the first round. In
other words, they had to make a decision between X and Y without knowing
whether A or ¬A. This time, 64% of students chose to reject the second gamble,
and 36% chose to accept it. Since in this case
P (X) = 0.36 < P (Y ) = 0.64,
there is a clear violation of the STP.
3.3. Violation of the sure-thing-principle with neural oscillators
We now turn to the representation of the above situation in terms of neural
oscillators. Following Tversky and Shafir (1992), we call "Won/Lost Version"
the case when students were told about A, and the "Disjunctive Version" when
students were not told about A. Let us start with the Won/Lost Version. In
the simplest case, we have two stimulus oscillators corresponding to the stim-
uli associated to the brain representation of A and ¬A. Let us call sa (t) the
oscillator corresponding to A, and sa (t) to ¬A. As before, response oscillators
are r1 (t) and r2 (t). For simplicity, we assume that all oscillators have the same
natural frequency ωi = ω0. We emphasize that this is not a necessary assump-
tion, but since the dynamics will lead to synchronization, this will make the
overall computations simpler. Then, when sa is activated, so are the response
oscillators, and the dynamics is given by
ϕsa = ω0 − kE
sa,r1 sin (ϕsa − ϕr1 )
−kE
−kI
−kI
sa,r2 sin (ϕsa − ϕr2 )
sa,r1 cos (ϕsa − ϕr1)
sa,r2 cos (ϕsa − ϕr2) ,
ϕr1 = ω0 − kE
r1,sa sin (ϕr1 − ϕsa )
−kE
−kI
−kI
r1,r2 sin (ϕr1 − ϕr2 )
r1,sa cos (ϕr1 − ϕsa )
r1,r2 cos (ϕr1 − ϕr2) ,
ϕr2 = ω0 − kE
r2,r1 sin (ϕr2 − ϕr1)
−kE
−kI
−kI
r2,sa sin (ϕr2 − ϕsa )
r2,r1 cos (ϕr2 − ϕr1 )
r2,sa cos (ϕr2 − ϕsa ) .
18
(26)
(27)
(28)
For such a system, it is possible to show (Suppes et al., 2012, Appendix) that a
response x is selected when
and
kE
sa,r1 = αx,
kE
sa,r2 = −αx,
kE
r1,r2 = −α,
kE
r1,sa = αx,
kE
r2,sa = −αx,
kE
r2,r1 = −α,
kI
sa,r1 = αp1 − x2,
sa,r2 = −αp1 − x2,
kI
kI
r1,r2 = 0,
r1,sa = −αp1 − x2,
kI
r2,sa = αp1 − x2,
kI
kI
r2,r1 = 0,
(29)
(30)
(31)
(32)
(33)
(34)
(35)
(36)
(37)
(38)
(39)
(40)
where α is a coupling strength parameter that determines how fast the solution
converges to the phase differences given in (13). We have similar equations for
sa (t), namely
ϕsa = ω0 − kE
sa,r1 sin (ϕsa − ϕr1 )
−kE
−kI
−kI
sa,r2 sin (ϕsa − ϕr2 )
sa,r1 cos (ϕsa − ϕr1)
sa,r2 cos (ϕsa − ϕr2) ,
ϕr1 = ω0 − kE
r1,sa sin (ϕr1 − ϕsa )
−kE
−kI
−kI
r1,r2 sin (ϕr1 − ϕr2 )
r1,sa cos (ϕr1 − ϕsa )
r1,r2 cos (ϕr1 − ϕr2) ,
ϕr2 = ω0 − kE
r2,r1 sin (ϕr2 − ϕr1)
−kE
−kI
−kI
r2,sa sin (ϕr2 − ϕsa )
r2,r1 cos (ϕr2 − ϕr1 )
r2,sa cos (ϕr2 − ϕsa ) ,
19
(41)
(42)
(43)
Figure 1: Histogram density of 300 oscillator models reinforced using the dynamics given in
equations (21) -- (23) for y = 0.5. The histogram was obtained using 6300 points corresponding
to 300 oscillators and 21 reinforcement trials per set of oscillators (trials 40 to 60). All
parameters used were the same as those described in Suppes et al. (2012), and K
0 was set to
4, 500. The fitted Gaussian has µ = 0.51 and σ = 0.05, with p < 10−3.
′
and the equivalent equations for (29) -- (40).
Equations (26) -- (40) determine a given phase relation for a response x in a
continuous interval, but they do not necessarily model the discrete case of a se-
lecting X over Y . To do so, let us recall that because of the stochasticity of the
initial conditions, if a particular value x is conditioned through couplings (29) --
(40), such response is selected according to a smearing distribution K(xksa ).
In fact, if f (y) is a simple reinforcement given by a Dirac delta function cen-
tered in z, i.e. f (y) = δ(y − z), then the response distribution r(x) becomes
asymptotically (Suppes, 1959)
r(x) = 1
−1
ksa (xy)δ(y − z)dy = ks(xz),
i.e. r(x) coincides with the smearing distribution at the point of reinforcement
z. Thus, by using a reinforcement density given by a Dirac delta function
centered at z, we are able to obtain the shape of the smearing distribution from
the oscillator dynamics. Figure 1 shows the density histogram for a MATLAB
simulation of the three oscillator model. Not surprisingly, the model's smearing
distribution ks(xz) fits well a Gaussian
1
ks (xz) =
σ√2π
e− (x−z)2
2σ2
,
where σ = 0.05. We should mention that the estimation of the smearing distri-
bution from a reinforcement schedule given by the Dirac delta function is not
realistic for real psychological experiments. However, because we are running
simulations without taking into consideration other behavioral and environmen-
tal aspects, we are able to extract ks(xy) from it.
20
With the smearing distribution, we may now describe how a stochastic de-
cision making processes may happen in the brain. Let X and Y be two of the
possible responses for a certain behavioral experiment. In our model, response
X would be associated with an oscillator, r1 (t), and Y with the other, r2 (t). If
I1 = 2A2 and I2 = 0, then the response would clearly be I1. Since −1 ≤ x ≤ 1,
x given by 16, we may state that x > 0 corresponds to a response X and x ≤ 0
to response Y . If we do this, when the reinforcement is a value different from 1
or −1, there is a nonzero chance, given by the smearing distribution, to select
each response. In other words, say a stimulus sa is conditioned to −1 ≤ z ≤ 1.
Then,
P (X) = 1
ksa (xz) dx
erf(cid:18) z
√2σ(cid:19)
1
2
and
≈
0
1
2
+
P (Y ) = 0
−1
1
2 −
≈
ksa (xz) dx
erf(cid:18) z
1
2
√2σ(cid:19) ,
for z , σ ≪ 1, where erf (x) is the error function. So, we can use the expression
z = √2σerf−1 (2P (R1) − 1)
to estimate the reinforcement value z leading to the desired probabilities of
responses. Let us summarize the oscillator implementation of the Won/Lost
version. We started with two stimulus, either Won or Lost, corresponding to
the oscillators sa (t) and sa (t). Once sa (t) or sa (t) are activated, the dynamics
of the system leads to a selection of either X or Y , corresponding to the decisions
of betting or not betting on the next step of the game.
We now model the "Disjunctive Version." Because the response Win or Lost
is not known, we assume that both oscillators sa (t) and sa (t) are activated
simultaneously. Furthermore, since the subject also knows that we either Win
or we Lose, oscillators sa (t) and sa (t) are incompatible, in the sense that they
must represent off-phase oscillations in an SR-oscillator model. Therefore, the
response system is composed not only of three oscillators, but four: the two
stimulus and the two response oscillators. The equivalent dynamical expressions
for (26) -- (43) when both oscillators are activated are the following.
ϕsa = ω0 − kE
sa,r1 sin (ϕsa − ϕr1 )
−kE
−kI
−kI
−kI
sa,r2 sin (ϕsa − ϕr2 )
sa,sa cos (ϕsa − ϕsa )
sa,r1 cos (ϕsa − ϕr1)
sa,r2 cos (ϕsa − ϕr2) ,
21
(44)
ϕsa = ω0 − kE
sa,r1 sin (ϕsa − ϕr1 )
−kE
−kI
−kI
−kI
sa,r2 sin (ϕsa − ϕr2 )
sa,sa cos (ϕsa − ϕsa )
sa,r1 cos (ϕsa − ϕr1)
sa,r2 cos (ϕsa − ϕr2) ,
ϕr1 = ω0 − kE
r1,sa sin (ϕr1 − ϕsa )
−kE
−kE
−kI
−kI
−kI
r1,sa sin (ϕr1 − ϕsa )
r1,r2 sin (ϕr1 − ϕr2 )
r1,sa cos (ϕr1 − ϕsa )
r1,sa cos (ϕr1 − ϕsa )
r1,r2 cos (ϕr1 − ϕr2) ,
ϕr2 = ω0 − kE
r2,r1 sin (ϕr2 − ϕr1)
−kE
−kE
−kI
−kI
−kI
r2,sa sin (ϕr2 − ϕsa )
r2,sa sin (ϕr2 − ϕsa )
r2,r1 cos (ϕr2 − ϕr1 )
r2,sa cos (ϕr2 − ϕsa )
r2,sa cos (ϕr2 − ϕsa ) ,
(45)
(46)
(47)
with the couplings between sa and r1 and r2 as well as sa and r1 and r2 being
the same, but
sa,sa = α′.
kI
(48)
As mentioned above, the parameter α′ in (48) represents a measure of the the
degree of incompatibility of stimulus oscillators sa (t) and sa (t).
1 be the mean intensity at r1 (t) given by
Now, as above, let I ′
I ′
1 = D(sa(t) + sa (t) + r1(t))2E ,
where this time we have both sa (t) and sa (t) contributing to the intensity.
Defining δϕr1,a ≡ ϕr1 − ϕsa and δϕr1,a ≡ ϕr1 − ϕsa , it is straightforward to
show that
I ′
1 = A2(cid:18) 3
2
+ cos (δϕr1,a) + cos (δϕr1,a) + cos (δϕr1,a − δϕr1,a)(cid:19) .
Similarly, for r2 (t) we have
I ′
2 = A2(cid:18) 3
2
+ cos (δϕr2,a) + cos (δϕr2,a) + cos (δϕr2,a − δϕr2,a)(cid:19) .
22
We notice that, because we now have more oscillators, extra interference terms
shows up in the intensity. We now use the new intensities, I ′
2 to compute
the response, using
1 and I ′
b′ =
I ′
1 − I ′
2
I ′
1 + I ′
2
.
At this point, we should remark on the appearance of quantum-like effects.
The model we are using, of interfering oscillators, carry some similarities with
the famous two-slit experiment (Suppes and de Barros, 1994a). In the two-slit
experiment, a quantum particle can go through two slits, and then hit a phos-
phorous screen, thus being detected. Since a particle is a localized object, going
through one slit should mean no interaction with the other slit, and therefore
we should expect to make no difference whether we keep the other slit closed
or open. However, if we open both slits, we observe different probability pat-
terns than if we close one of the slits. In other words, if we close one of the
slits, and therefore know through which slit the particle went through, we get
a probability of observing the particle in a certain region of the screen. If, on
the other hand, we do not close one of the slits, we do not know through which
slit it went, and when we do not know, the probability of detecting the particle
changes. Formally, let L be the proposition "went through the slit on the left
because the slit on the right was closed," R ="went through the slit on the right
because the slit on the left was closed," and A ="was observed at a certain
region on the screen." Then, it is possible to choose a region on the screen such
that
and
P (AR) > P (¬AR) ,
P (AL) > P (¬AL) ,
i.e., such that there are more particles being detected on it than outside of it.
However, because of quantum interference, if we do not close the slits, and do
not know where they went, it is again possible to have an A such that the above
holds and
This is the two slit formal equivalent of the violation of STP.
P (A) < P (¬A) .
Now, let us go back to the oscillator model. For the the Win/Lost Version,
a single stimulus oscillator is active at a time. Such oscillator produces a dy-
namics that leads to a certain balance between the responses X and Y . This
corresponds to one slit open, when there is no interference from the other slit.
However, for the Disjunctive Version, both oscillators are active, and their ac-
tivity interferes with each other, in a way similar to the interference from the
two-slit experiment. The strength of interference, i.e. the degree of coherence
between the two oscillators, is determined by the inhibitory coupling parame-
ter kI
sa,sa. Therefore, when both oscillators are active, we obtain a probability
distribution violating the standard axioms of probability, and consequently STP.
To end this section, let us look at a simulation of the violation of the STP
using neural oscillators. We simulated in Matlab R2012a the oscillators' dynam-
ics determined by equations (26) -- (43) and (44) -- (48) using the Dormand-Prince
23
sa,r1 = −kE
sa,r2 = kE
r1,sa = kE
r1,r2 = kE
sa,r2 = kE
r1,r2 = kI
r2,r1 = −11 Hz, kI
r1,sa = kE
r2,r1 = 0, and kI
method. For simplicity, and without loss of generality, all oscillators were as-
sumed to have the same frequency of 11 Hz, as this would allow the use of the
couplings given by (35) -- (40) and (48) without having to invoke learning equa-
tions, as in (Suppes et al., 2012). For our system, the parameters utilized were
the following: ∆tr = 0.2 s, σϕ = pπ/4, kE
r2,sa =
sa,r2 = −kI
−0.011 Hz, kE
r2,sa = 1,
sa,r1 = −kI
sa,r1 = −kE
kE
r1,sa =
kI
r2,sa = 1, kI
sa,sa = −0.011, where ∆tr is
the time of response after the onset of stimulus, and σϕ is the variance of the
initial conditions for the phase oscillators. We ran 1,000 trials with the above
parameter values, for the Win/Lost Version we obtained X as response 63%
when the Lost oscillator was used and 58% with the Win oscillator. However,
for the Disjunctive Version, when the two oscillators were active, we obtained
X as a response 36% of the runs, clearly showing a violation of Savage's STP
and an interference effect, as expected.
sa,r1 = −kI
r2,sa = −0.017 Hz, kI
sa,sa = kI
r1,sa = kI
sa,r2 = −kI
4. Conclusions
In this paper we presented a neural oscillator model based on reasonable
assumptions about the behavior of coupled neurons. We then showed that
the predictions of such model presents interference effects between incompat-
ible stimulus oscillators. Such interference leads to computation of responses
which are inconsistent with standard probability laws, but are consistent with
quantum-like effects in decision making processes Busemeyer et al. (2009). Of
course, the fact that we get interference from s1 and s2 does not preclude us from
necessarily making rational decisions. We could, for example, modify the above
model, and include an extra oscillator, such that after reinforcement we could
eliminate interference, therefore satisfying the STP. But our model suggests that
a plausible mechanism for quantum effects in the brain may not need to rely on
actual quantum processes, as advocated by Penrose (1989), but instead can be
a consequence of neuronal "interference."
One interesting feature of the oscillator model is that the amount of inter-
ference between the oscillators is determined by the couplings between s1 and
s2. The stronger the incompatibility between the two stimulus, the more we
should see interference between the responses. Furthermore, because interfer-
ence is a consequence of inhibitory couplings, we could in principle design exper-
iments where pharmacological interventions could suppress inhibitory synapses,
destroying such interference effects.
An attentive reader may question the possibility of modeling decision-making
processes from neurons.
In particular, how can we account for what Pattee
claims are non-holonomic constraints not internalized by the dynamics (Pattee,
1972, 2001, 2012). To address this issue, let us examine this claim in more
detail. First, we should notice that there are many examples of non-holonomic
constraints in physical systems that are determined by an underlying and well-
understood dynamics. For example, a ball rolling without slipping on a flat
24
surface satisfies such type of constraint. But most physicists would not think
that such an example presents fundamental difficulties to the description of the
system ball-surface. In this case, what is a constraint and what is a physical
system is just a matter of computational convenience. As a trained physicists
himself, we believe this is not what Pattee meant.
For Pattee, decision making processes require the type of discontinuous jump
that cannot be associated to classical dynamical systems, but can be thought
of as determined by non-holonomic constraints in a classical dynamics. Since
the fundamental laws of physics do not have such characteristics, he reasons,
it remains a mystery how they emerge from complex systems. He goes on to
propose that the jumps come from the emergence of quantum decoherence (using
the modern parlance).
Going back to the model presented here, the question remains on how we can
derive such emergent characteristics. First, we emphasize that we are not mod-
eling decision making processes from basic physics, but instead from emerging
properties of collections of neurons. In other words, we are assuming an emerg-
ing dynamics with the necessary features described by Pattee (1972): nonlin-
earity and time-dependent dynamics (Guckenheimer and Holmes, 1983). For
instance, overall dynamics of a highly complex system of interacting neurons
is modeled mathematically with nonlinear equations describing the behavior
of phases (which adds to the nonlinearity). Furthermore, the decision-making
dynamics changes in time by reinforcements which are outside of the dynami-
cal system itself (Suppes et al., 2012). Such reinforcements are not only time-
dependent, but can also be viewed as non-holonomic constraints to the dynam-
ics. Thus, in a certain sense, we are not addressing the fundamental problem
posed by Pattee, but we do push it to another level in the description of neural
interactions.
It is also important to clarify the relationship between our model and quan-
tum mechanics. We are not claiming that we derive quantum mechanics from
neuronal dynamics. We are showing that certain quantum-like features (but not
all; see de Barros and Suppes, 2009, for details) are present in our model. In
other words, we start with a classical stochastic description and obtain some
quantum-like results. This is to be contrasted with Pattee, who argues that
classical-like dynamics emerges from a quantum world, and such emergence
happens with the appearance of life itself11.
We end this paper with a comparison between our model and the one pro-
posed by Khrennikov (2011). Whereas in Khrennikov's model the interference
of electromagnetic fields account for quantum-like effects in the brain, in our
model we use Kuramoto oscillators to describe the evolution of coupled sets of
dynamical oscillators close to Andronov-Hopf bifurcations. Because such os-
cillators obey a wave equation, they are subject to interference. So, in some
11In an earlier paper (Suppes and de Barros, 2007), we actually discussed the limits of quan-
tum effects in decision-making processes, and proposed an experiment where the entanglement
of photons could be used to condition an insect.
25
sense, we may say that the oscillator and the field models are equivalent repre-
sentations of quantum-like interference in the brain. However, though we can
think of the weak couplings between oscillators originating from the synapses
between neurons, it is also conceivable that the weak oscillating electric fields
generated by sets of oscillators provide sufficient additional coupling to account
for their synchronization and interference. Thus, we believe that there are
not only equivalencies between the models, but also that our oscillator model
might provide an underlying neurophysiological justification for Khrennikov's
quantum-like field processing. It would be interesting to prove some representa-
tion theorem between neural oscillators and Khrennikov's field model. However,
this seems implausible, since it was shown in de Barros (2012) that oscillators
may generate processes that are not representable in terms of observables in a
Hilbert space, and Khrennikov's model relies on a quantum representation of
states in terms of Hilbert spaces.
Acknowledgments. The author wishes to thank Dr. Gary Oas for critical reading
and discussion on the contents of this manuscript, as well as the anonymous
referee for useful comments and suggestions.
References
Aerts, D., 2009. Quantum structure in cognition. Journal of Mathematical
Psychology 53, 314 -- 348.
Asano, M., Ohya, M., Khrennikov, A., 2010. Quantum-like model for decision
making process in two players game. Foundations of Physics 41, 538 -- 548.
Baaquie, B.E., 1997. A path integral approach to option pricing with stochastic
volatility: Some exact results. Journal de Physique I 7, 1733 -- 1753.
de Barros, J.A., 2011. Comments on "There is no axiomatic system for the quan-
tum theory". International Journal of Theoretical Physics 50, 1828 -- 1830.
de Barros, J.A., 2012. Joint probabilities and quantum cognition, in: Khren-
nikov, A. (Ed.), Proceedings of Quantum Theory: Reconsiderations of Foun-
dations - 6, Institute of Physics, Växjö, Sweden. To appear.
de Barros, J.A., de Mendonça, J.P.R.F., Pinto-Neto, N., 2007. Realism in
energy transition processes: an example from Bohmian quantum mechanics.
Synthese 154, 349 -- 370.
de Barros, J.A., Oas, G., Suppes, P., 2012. Response selection using neural
phase oscillators. arXiv:1208.6041 Submitted to the Festschrift in Honor of
Patrick Suppes's 90th Birthday.
de Barros, J.A., Suppes, P., 2000. Inequalities for dealing with detector inef-
ficiencies in Greenberger-Horne-Zeilinger type experiments. Physical Review
Letters 84, 793 -- 797.
26
de Barros, J.A., Suppes, P., 2001. Probabilistic results for six detectors in a
three-particle GHZ experiment, in: Bricmont, J., Dürr, D., Galavotti, M.C.,
Ghirardi, G., Petruccione, F., Zanghi, N. (Eds.), Chance in Physics, p. 213.
de Barros, J.A., Suppes, P., 2009. Quantum mechanics, interference, and the
brain. Journal of Mathematical Psychology 53, 306 -- 313.
de Barros, J.A., Suppes, P., 2010. Probabilistic inequalities and upper proba-
bilities in quantum mechanical entanglement. Manuscrito 33, 55 -- 71.
Bell, J.S., 1964. On the Einstein-Podolsky-Rosen paradox. Physics 1, 195 -- 200.
Bell, J.S., 1966. On the problem of hidden variables in quantum mechanics.
Rev. Mod. Phys. 38, 447 -- 452.
Billock, V.A., Tsou, B.H., 2005. Sensory recoding via neural synchronization:
integrating hue and luminance into chromatic brightness and saturation. J.
Opt. Soc. Am. A 22, 2289 -- 2298.
Billock, V.A., Tsou, B.H., 2011. To honor Fechner and obey Stevens: rela-
tionships between psychophysical and neural nonlinearities. Psychological
Bulletin 137, 1 -- 18.
Bruza, P., Busemeyer, J.R., Gabora, L., 2009. Introduction to the special issue
on quantum cognition. Journal of Mathematical Psychology 53, 303 -- 305.
Busemeyer, J., Wang, Z., 2007. Quantum information processing explanation for
interactions between inferences and decisions, in: Proceedings of the Quantum
Interaction Symposium AAAI Press, AAAI Press, Menlo Park, CA. pp. 91 --
97.
Busemeyer, J., Wang, Z., Lambert-Mogiliansky, A., 2009. Empirical comparison
of markov and quantum models of decision making. Journal of Mathematical
Psychology 53, 423 -- 433.
Busemeyer, J.R., Diederich, A., 2010. Cognitive modeling. SAGE.
Busemeyer, J.R., Wang, Z., Townsend, J.T., 2006. Quantum dynamics of human
decision-making. Journal of Mathematical Psychology 50, 220 -- 241.
Ford, J., 1983. How random is a coin toss? Physics Today 36, 40.
Galavotti, M.C., 2005. Philosophical introduction to probability. volume 167 of
CSLI Lecture Notes. CSLI Publications, Stanford, CA.
Guckenheimer, J., Holmes, P., 1983. Nonlinear oscillations, dynamical systems,
and bifurcations of vector fields. volume 42. Springer-Verlag New York.
Haven, E., 2002. A discussion on embedding the Black -- Scholes option pricing
model in a quantum physics setting. Physica A: Statistical Mechanics and its
Applications 304, 507 -- 524.
27
Hoppensteadt, F.C., Izhikevich, E.M., 1996a. Synaptic organizations and dy-
namical properties of weakly connected neural oscillators i. analysis of a
canonical model. Biological Cybernetics 75, 117 -- 127.
Hoppensteadt, F.C., Izhikevich, E.M., 1996b. Synaptic organizations and dy-
namical properties of weakly connected neural oscillators II. learning phase
information. Biological Cybernetics 75, 129 -- 135.
Izhikevich, E.M., 2007. Dynamical Systems in Neuroscience: The Geometry of
Excitability and Bursting. The MIT Press, Cambridge, Massachusetts.
Jaynes, E.T., 2003. Probability theory: the logic of science. Cambridge Univ
Press, Cambridge, Great Britain.
Josić, K., Rubin, J., Matias, M., 2009. Coherent Behavior in Neuronal Networks.
Springer Verlag, Dordrecht, Holland.
Khrennikov, A., 2009. Quantum-like model of cognitive decision making and
information processing. Biosystems 95, 179 -- 187.
Khrennikov, A., 2010. Ubiquitous Quantum Structure. Springer Verlag, Hei-
delberg.
Khrennikov, A., 2011. Quantum-like model of processing of information in the
brain based on classical electromagnetic field. Biosystems 105, 250 -- 262.
Khrennikov, A., Haven, E., 2009. Quantum mechanics and violations of the
sure-thing principle: The use of probability interference and other concepts.
Journal of Mathematical Psychology 53, 378 -- 388.
Kochen, S., Specker, E.P., 1975. The problem of hidden variables in quantum
mechanics, in: Hooker, C.A. (Ed.), The Logico-Algebraic Approach to Quan-
tum Mechanics. D. Reidel Publishing Co., Dordrecht, Holland, p. 293 -- 328.
Kuramoto, Y., 1984. Chemical Oscillations, Waves, and Turbulence. Dover
Publications, Inc., Mineola, New York.
Mackintosh, N.J., 1983. Conditioning and associative learning. Clarendon Press
Oxford.
Nelson, D.L., McEvoy, C.L., Pointer, L., 2003. Spreading activation or spooky
action at a distance? Journal of Experimental Psychology: Learning, Mem-
ory, and Cognition 29, 42 -- 51.
von Neumann, J., 1996. Mathematical Foundations of Quantum Mechanics.
Princeton University Press, Princeton, NJ.
Nunez, P., Srinivasan, R., 2006. Electric Fields of the Brain: The Neurophysics
of EEG, 2nd Ed. Oxford University Press, New York, NY. 2nd edition.
28
Pattee, H., 2012. Epistemic, evolutionary, and physical conditions for biological
information. Biosemiotics , 1 -- 23.
Pattee, H.H., 1972. Physical problems of decision-making constraints. Interna-
tional Journal of Neuroscience 3, 99 -- 105.
Pattee, H.H., 2001. The physics of symbols: bridging the epistemic cut. Biosys-
tems 60, 5 -- 21.
Penrose, R., 1989. Emperor's New Mind. Oxford University Press, New York,
NY.
Savage, L.J., 1972. The foundations of statistics. Dover Publications Inc.,
Mineola, New York. 2nd edition.
Seliger, P., Young, S.C., Tsimring, L.S., 2002. Plasticity and learning in a
network of coupled phase oscillators. Physical Review E 65, 041906 -- 1 -- 7.
Shafir, E., Tversky, A., 1992. Thinking through uncertainty: Nonconsequential
reasoning and choice. Cognitive Psychology 24, 449 -- 474.
Suppes, P., 1959. A linear learning model for a continuum of responses, in:
Bush, R.R., Estes, W.K. (Eds.), Studies in Mathematical Leaning Theory.
Stanford University Press, Stanford, CA, pp. 400 -- 414.
Suppes, P., 1970. A probabilistic theory of causality. volume 24 of Acta Philo-
sophica Fennica. North-Holland Publishing Company, Amsterdam, Holland.
Suppes, P., 2002. Representation and Invariance of Scientific Structures. CSLI
Publications, Stanford, California.
Suppes, P., Atkinson, R.C., 1960. Markov learning models for multiperson
interactions. Stanford University Press, Stanford, California.
Suppes, P., de Barros, J.A., 1994a. Diffraction with well-defined photon trajec-
tories: A foundational analysis. Foundations of Physics Letters 7, 501 -- 514.
Suppes, P., de Barros, J.A., 1994b. A random-walk approach to interference.
International Journal of Theoretical Physics 33, 179 -- 189.
Suppes, P., de Barros, J.A., 1996. Photons, billiards and chaos, in: Weingartner,
P., Schurz, G. (Eds.), Law and Prediction in the Light of Chaos Research.
Springer Berlin Heidelberg. volume 473, pp. 189 -- 201.
Suppes, P., de Barros, J.A., 2007. Quantum mechanics and the brain, in: Quan-
tum Interaction: Papers from the AAAI Spring Symposium, AAAI Press,
Menlo Park, CA. p. 75 -- 82.
29
Suppes, P., de Barros, J.A., Oas, G., 1996a. A collection of probabilistic
hidden-variable theorems and counterexamples, in: Pratesi, R., Ronchi, L.
(Eds.), Waves, Information, and Foundations of Physics: a tribute to Giuliano
Toraldo di Francia on his 80th birthday, Italian Physical Society, Florence,
Italy.
Suppes, P., de Barros, J.A., Oas, G., 2012. Phase-oscillator computations as
neural models of stimulus -- response conditioning and response selection. Jour-
nal of Mathematical Psychology 56, 95 -- 117.
Suppes, P., de Barros, J.A., Sant'Anna, A.S., 1996b. A proposed experi-
ment showing that classical fields can violate bell's inequalities. arXiv:quant-
ph/9606019 .
Suppes, P., de Barros, J.A., Sant'Anna, A.S., 1996c. Violation of bell's in-
equalities with a local theory of photons. Foundations of Physics Letters 9,
551 -- 560.
Suppes, P., Sant'Anna, A., Barros, J., 1996d. A particle theory of the casimir
effect. Foundations of Physics Letters 9, 213 -- 223.
Suppes, P., Zanotti, M., 1981. When are probabilistic explanations possible?
Synthese 48, 191 -- 199.
Trevisan, M.A., Bouzat, S., Samengo, I., Mindlin, G.B., 2005. Dynamics of
learning in coupled oscillators tutored with delayed reinforcements. Physical
Review E 72, 011907.
Tversky, A., Shafir, E., 1992. The disjunction effect in choice under uncertainty.
Psychological Science 3, 305 -- 309.
Vassilieva, E., Pinto, G., de Barros, J.A., Suppes, P., 2011. Learning pattern
recognition through quasi-synchronization of phase oscillators. IEEE Trans-
actions on Neural Networks 22, 84 -- 95.
Vulović, V.Z., Prange, R.E., 1986. Randomness of a true coin toss. Physical
Review A 33, 576 -- 582.
Werndl, C., 2009. Are deterministic descriptions and indeterministic descrip-
tions observationally equivalent? Studies in history and philosophy of science
part B: studies in history and philosophy of modern physics 40, 232 -- 242.
30
|
1705.09205 | 3 | 1705 | 2018-01-05T14:10:07 | Firing rate equations require a spike synchrony mechanism to correctly describe fast oscillations in inhibitory networks | [
"q-bio.NC",
"nlin.AO"
] | Recurrently coupled networks of inhibitory neurons robustly generate oscillations in the gamma band. Nonetheless, the corresponding Wilson-Cowan type firing rate equation for such an inhibitory population does not generate such oscillations without an explicit time delay. We show that this discrepancy is due to a voltage-dependent spike-synchronization mechanism inherent in networks of spiking neurons which is not captured by standard firing rate equations. Here we investigate an exact low-dimensional description for a network of heterogeneous canonical type-I inhibitory neurons which includes the sub-threshold dynamics crucial for generating synchronous states. In the limit of slow synaptic kinetics the spike-synchrony mechanism is suppressed and the standard Wilson-Cowan equations are formally recovered as long as external inputs are also slow. However, even in this limit synchronous spiking can be elicited by inputs which fluctuate on a time-scale of the membrane time-constant of the neurons. Our meanfield equations therefore represent an extension of the standard Wilson-Cowan equations in which spike synchrony is also correctly described. | q-bio.NC | q-bio | Firing rate equations require a spike synchrony mechanism to correctly
describe fast oscillations in inhibitory networks
Federico Devalle,1, 2 Alex Roxin,3 and Ernest Montbri´o1
1Center for Brain and Cognition. Department of Information and Communication Technologies,
Universitat Pompeu Fabra, 08018 Barcelona, Spain
2Department of Physics, Lancaster University,
Lancaster LA1 4YB, United Kingdom
3Centre de Recerca Matem`atica, Campus de Bellaterra,
Edifici C, 08193 Bellaterra, Barcelona, Spain.
Abstract
Recurrently coupled networks of inhibitory neurons robustly generate oscillations in the gamma band.
Nonetheless, the corresponding Wilson-Cowan type firing rate equation for such an inhibitory population
does not generate such oscillations without an explicit time delay. We show that this discrepancy is due
to a voltage-dependent spike-synchronization mechanism inherent in networks of spiking neurons which is
not captured by standard firing rate equations. Here we investigate an exact low-dimensional description
for a network of heterogeneous canonical Class 1 inhibitory neurons which includes the sub-threshold dy-
namics crucial for generating synchronous states. In the limit of slow synaptic kinetics the spike-synchrony
mechanism is suppressed and the standard Wilson-Cowan equations are formally recovered as long as ex-
ternal inputs are also slow. However, even in this limit synchronous spiking can be elicited by inputs which
fluctuate on a time-scale of the membrane time-constant of the neurons. Our meanfield equations therefore
represent an extension of the standard Wilson-Cowan equations in which spike synchrony is also correctly
described.
8
1
0
2
n
a
J
5
]
.
C
N
o
i
b
-
q
[
3
v
5
0
2
9
0
.
5
0
7
1
:
v
i
X
r
a
1
I.
INTRODUCTION
Since the seminal work of Wilson and Cowan [1], population models of neuronal activity have
become a standard tool of analysis in computational neuroscience. Rather than focus on the micro-
scopic dynamics of neurons, these models describe the collective properties of large numbers of
neurons, typically in terms of the mean firing rate of a neuronal ensemble. In general, such popu-
lation models, often called firing rate equations, cannot be exactly derived from the equations of a
network of spiking neurons, but are obtained using heuristic mean-field arguments, see e.g. [2–6].
Despite their heuristic nature, heuristic firing rate equations (which we call H-FRE) often show
remarkable qualitative agreement with the dynamics in equivalent networks of spiking neurons
[7–10], and constitute an extremely useful modeling tool, see e.g.
[11–28]. Nonetheless, this
agreement can break down once a significant fraction of the neurons in the population fires spikes
synchronously, see e.g. [29]. Such synchronous firing may come about due to external drive, but
also occurs to some degree during spontaneously generated network states.
As a case in point, here we focus on partially synchronized states in networks of heteroge-
neous inhibitory neurons. Inhibitory networks are able to generate robust macroscopic oscillations
due to the interplay of external excitatory inputs with the inhibitory mean field produced by the
population itself. Fast synaptic processing coupled with subthreshold integration of inputs intro-
duces an effective delay in the negative feedback facilitating the emergence of what is often called
Inter-Neuronal Gamma (ING) oscillations [30–38]. Modeling studies with networks of spiking
neurons demonstrate that, in heterogeneous inhibitory networks, large fractions of neurons be-
come frequency-entrained during these oscillatory episodes, and that the oscillations persist for
weak levels of heterogeneity [30, 32, 34]. Traditional H-FRE (also referred to as Wilson-Cowan
equations) fail to describe such fast oscillations. To overcome this limitation, explicit fixed time
delays have been considered in H-FRE as a heuristic proxy for the combined effects of synaptic
and subthreshold integration [9, 10, 36, 39].
Here we show that fast oscillations in inhibitory networks are correctly described by a recently
derived set of exact macroscopic equations for quadratic integrate-and-fire neurons (that we call
QIF-FRE) which explicitly take into account subthreshold integration [40]. Specifically, the QIF-
FRE reveal how oscillations arise via a voltage-dependent spike synchronization mechanism, miss-
ing in H-FRE, as long as the recurrent synaptic kinetics are sufficiently fast. In the limit of slow
recurrent synaptic kinetics intrinsically generated oscillations are suppressed, and the QIF-FRE re-
2
duce to an equation formally identical to the Wilson-Cowan equation for an inhibitory population.
However, even in this limit, fast fluctuations in external inputs can drive transient spike synchrony
in the network, and the slow synaptic approximation of the QIF-FRE breaks down. This suggests
that, in general, a correct macroscopic description of spiking networks requires keeping track of
the mean subthreshold voltage along with the mean firing rate.
Additionally, the QIF-FRE describe the disappearance of oscillations for sufficiently strong
heterogeneity which is robustly observed in simulations of spiking networks. Finally, we also
show that the phase diagrams of oscillatory states found in the QIF-FRE qualitatively match those
observed in simulations of populations of more biophysically inspired Wang-Buzs´aki neurons
[30]. This shows that not only are the QIF-FRE an exact mean-field description of networks
of heterogeneous QIF neurons, but that they also provide a qualitatively accurate description of
dynamical states in networks of spiking neurons more generally, including states with significant
spike synchrony.
II. RESULTS
Recurrent networks of spiking neurons with inhibitory interactions readily generate fast os-
cillations. Figure 1 shows an illustration of such oscillations in a network of globally coupled
Wang-Buzs´aki (WB) neurons [30]. Panels (a,c) show the results of a numerical simulation of the
network for fast synapses (time constant τd = 5 ms), compared to the membrane time constant of
the neuron model (τm = 10 ms). Although the neurons have different intrinsic frequencies due
to a distribution in external input currents, the raster plot reveals that fast inhibitory coupling pro-
duces the frequency entrainment of a large fraction of the neurons in the ensemble. This collective
synchronization is reflected at the macroscopic scale as an oscillation with the frequency of the
synchronous cluster of neurons [41, 42]. Indeed, panel (a) shows the time series of both the mean
synaptic activation variable S, and the mean firing rate R, which display ING oscillations. Panels
(b,d) of Fig. 1 show the disappearance of the synchronous state when the synaptic kinetics is slow
(τd = 50 ms).
3
FIG. 1. Networks of heterogeneous inhibitory neurons with fast synaptic kinetics (τd = 5 ms) display
macroscopic oscillations in the gamma range (ING oscillations) due to collective synchronization. Panels
(a) and (c) show the time series of the synaptic variable S (red) and mean firing rate R (blue), and the
raster plot of a population of N = 1000 inhibitory Wang-Buzs´aki neurons [30] with first order fast synaptic
kinetics. The oscillations are suppressed considering slow inhibitory synapses (τd = 50 ms), as shown in
Panels (b) and (d). See Materials and Methods for details on the numerical simulations.
A. A heuristic firing rate equation
A heuristic firing rate description of the spiking network simulated in Fig. 1 takes the form [1, 5]
R = −R + Φ(−JτmS + Θ),
S = −S + R.
τm
τd
(1a)
(1b)
where R represents the mean firing rate in the population, S is the synaptic activation, and the
time constants τm and τd are the neuronal and synaptic time constants respectively [39, 43]. The
input-output function Φ, also known as the f-I curve, is a nonlinear function, the form of which
depends on the details of the neuronal model and on network parameters. Finally, J ≥ 0 is the
synaptic strength and Θ is the mean external input current compared to threshold. In contrast with
the network model, the H-FRE Eqs. (1) cannot generate sustained oscillations. In fact, a linear
stability analysis of steady state solutions in Eqs. (1) shows that the resulting eigenvalues are
λ = −α(1 ±(cid:112)1 − β),
(2)
where the parameter α = (τm + τd)/(2τmτd) is always positive. Additionally, the parameter
β = [4τmτd(1 + JτmΦ(cid:48))]/(τm + τd)2 is also positive, since the f-I curve Φ(x) is an increasing
4
function, and its derivative evaluated at the steady state is then Φ(cid:48) > 0. Therefore the real part of
the eigenvalue λ is always negative and hence steady states are always stable, although damped
oscillations are possible, e.g. for strong enough coupling J. Introducing an explicit fixed time
delay in Eqs. (1) can lead to the generation of oscillations with a period on the order of about
twice the delay [9, 10, 36]. Nonetheless, inhibitory networks of spiking neurons robustly show
oscillations even in the absence of explicit delays, as seen in Fig. 1. This suggests that there is
an additional mechanism in the network dynamics, key for driving oscillatory behavior, which
H-FRE do not capture.
B. An exact firing rate equation which captures spike synchrony
Here we show that the mechanism responsible for the generation of the oscillations shown
in Fig. 1 is the interplay between the mean firing rate and membrane potential, the dynamics of
which reflect the degree of spike synchrony in the network. To do this, we use a set of exact macro-
scopic equations which have been recently derived from a population of heterogeneous quadratic
integrate-and-fire (QIF) neurons [40]. We refer to these equations as the QIF-FRE. The QIF-FRE
with exponential synapses have the form
∆
πτm
+ 2RV,
R =
V = V 2 − (πτmR)2 − JτmS + Θ,
S = −S + R.
τm
τm
τd
(3a)
(3b)
(3c)
where ∆ is a parameter measuring the degree of heterogeneity in the network and the other param-
eters are as in the H-FRE Eqs. (1). Eqs. (3) are an exact macroscopic description of the dynamics
in a large network of heterogeneous QIF neurons with inhibitory coupling. In contrast with the
traditional firing rate equations Eqs. (1), they contain an explicit dependence on the subthreshold
state of the network, and hence capture not only macroscopic variations in firing rate, but also in
spike synchrony. Specifically, a transient depolarizing input which drives the voltage to positive
values (the voltage has been normalized such that positive values are suprathrehsold, see Materials
and Methods) will lead to a sharp growth in the firing rate through the bilinear term in Eq. (3a).
Simulations in the corresponding network model reveal that this increase is due to the synchronous
spiking of a subset of neurons [40]. This increase in firing rate leads to a hyperpolarization of the
mean voltage through the quadratic term in R in Eq. (3b). This term describes the effect of the
5
neuronal reset. This decrease in voltage in turn drives down the mean firing rate, and the process
can repeat. Therefore the interplay between mean firing rate and mean voltage in Eqs. (3) can
generate oscillatory behavior; this behavior corresponds to transient bouts of spike synchrony in
the spiking network model.
It is precisely the latency inherent in this interplay which provides the effective delay, which
when coupled with synaptic kinetics, generates self-sustained fast oscillations. In fact, in the limit
of instantaneous synapses (τd → 0), Eqs. (3) robustly display damped oscillations due to the spike
generation and reset mechanism described in the preceding paragraph [40]. Contrast this with the
dynamics in Eqs.(1) in the same limit: the resulting H-FRE is one dimensional and hence cannot
oscillate.
On the face of things the Eqs. (3) appear to have an utterly distinct functional form from the
traditional Wilson-Cowan Eqs.(1). In particular, the f-I curve in H-FRE traditionally exhibits an
expansive nonlinearity at low rates and a linearization or saturation at high rates, e.g. a sigmoid.
There is no such function visible in the QIF-FRE which have only quadratic nonlinearities. How-
ever, this seeming inconsistency is simply due to the explicit dependence of the steady-state f-I
curve on the subthreshold voltage in Eqs. (3). In fact, the steady-state f-I curve in the QIF-FRE is
"typical" in the qualitative sense described above. Specifically, solving for the steady state value
of the firing rate in Eqs. (3) yields
where
R∗ = Φ(−JτmR∗ + Θ),
(cid:113)
I +
√
I 2 + ∆2.
Φ(I) =
1√
2πτm
(4)
(5)
The f-I curve Eq. (5) is shown in Fig. 2 for several values of the parameter ∆, which measures the
degree of heterogeneity in the network. Therefore, the steady-state firing rate in Eqs. (3), which
corresponds exactly to that in a network of heterogeneous QIF neurons, could easily be captured
in a heuristic model such as Eqs. (1) by taking the function Φ to have the form as in Eq. (5). On
the other hand, the non-steady behavior in Eqs. (3), and hence in spiking networks as well, can be
quite different from that in the heuristic Eqs. (1).
6
FIG. 2. The f-I curve Φ(I), Eq. (5), for several values of the heterogeneity parameter ∆. The membrane
time constant is τm = 10ms.
1. Fast oscillations in the QIF-FRE
We have seen that decreasing the time constant of synaptic decay τd in a network of inhibitory
spiking neurons lead to sustained fast oscillations, while such a transition to oscillations is not
found in the heuristic rate equations, in which the synaptic dynamics are taken into account
Eqs. (1). The exact QIF-FRE, on the other hand, do generate oscillations in this regime. Fig-
ure 3 shows a comparison of the firing rate R and synaptic variable S from simulations of the
QIF-FRE Eqs.(3), with that of the H-FRE Eqs. (1), for two different values of the synaptic time
constants. Additionally, we also performed simulations of a network of N = 5× 104 QIF neurons.
The mean firing rate of the network is shown in red, and perfectly agrees with the firing rate of the
low dimensional QIF-FRE (solid black lines). The function Φ in Eqs. (1) is chosen to be that from
Eq. (5), which is why the firing rate from both models converges to the same steady state value
when oscillations are not present (panels (b,d) for τd = 50 ms). We will study the transition to fast
oscillations in Eqs.(3) in greater details in the following sections.
C. Linear stability analysis of the QIF-FRE
We can investigate the emergence of sustained oscillations in Eqs. (3) by considering small
amplitude perturbations of the steady-state solution. If we take R = R∗ + δReλt, V = V∗ + δV eλt
and S = S∗ + δSeλt, where δR, δV , δS (cid:28) 1, then the sign of the real part of the eigenvalue
λ determines whether the perturbation grows or not. Plugging this ansatz into Eqs. (3) yields
7
FIG. 3. Heuristic FRE Eqs. (1) do not display inhibition-based fast oscillations. In contrast, networks of
QIF neurons (red) and their corresponding QIF-FRE Eqs. (3) (solid black) do show ING oscillations for fast
synaptic kinetics (τd = 5 ms). These oscillations are suppressed for slow synaptic kinetics (τd = 50 ms), as
in the Wang-Buzs´aki model shown in Fig. 1. Panels (a,b) show the times series of the Firing Rate variable
R of the FRE models, as well as the mean firing rate of a population of N = 5 × 104 QIF neurons (red).
Panels (c,d) show the time series of the synaptic S variables for the H-FRE (dashed line) and QIF-FRE
(solid line). Parameters: τm = 10 ms, J = 21, Θ = 4, ∆ = 0.3. Initial values: R(0) = S(0) = 5 Hz,
V (0) = 0.
three coupled linear equations which are solvable if the following characteristic equation also has
a solution
− 2JτmR∗ = (1 + τdλ)
(2πτmR∗)2 +
τmλ +
.
(6)
(cid:34)
(cid:18)
(cid:19)2(cid:35)
∆
πτmR∗
The left hand side of Eq. (6) is always negative and, for τd = 0, this implies that the solutions λ are
necessarily complex. Hence, for instantaneous synapses, the fixed point of the QIF-FRE is always
of focus type, reflecting transient episodes of spike synchrony in the neuronal ensemble [40]. In
contrast, setting τd = 0 in the H-FRE, the system becomes first order and oscillations are not
possible. This is the critical difference between the two firing rate models. In fact, and in contrast
with the eigenvalues Eq. (2) corresponding to the growth rate of small perturbations in the H-FRE,
here oscillatory instabilities may occur for nonvanishing values of τd. Figure 4 shows the Hopf
√
boundaries obtained from Eq. (6), as a function of the normalized synaptic strength j = J/
Θ
8
FIG. 4. The ratio of the width to the center of the distribution of currents Eq. (14), δ = ∆/Θ, determines
the presence of fast oscillations in the QIF-FRE. Oscillations disappear above the critical value given by
Eq. (8). The panels show the Hopf boundaries of QIF-FRE with first-order synapses, for different values of
δ, obtained solving the characteristic Eq. (6) with Re(λ) = 0, see Materials and Methods. Shaded regions
are regions of oscillations. Symbols in the right panel correspond to the parameters used in Fig 3.
Θτd/τm, and for different values
and the ratio of the synaptic and neuronal time constants τ =
of the ratio δ = ∆/Θ -see Materials and Methods, Eqs.(19-21). The shaded regions correspond
to parameter values where the QIF-FRE display oscillatory solutions.
√
1.
Identical neurons
In the simplest case of identical neurons we find the boundaries of oscillatory instabilities
explicitly. Indeed, substituting λ = ν + iω in Eq. (6) we find that, near criticality (ν (cid:28) 1), the
real part of the eigenvalue is
ν ≈ Jτ
R∗
1 + (2πτdR∗)2 .
(7)
Thus, the fixed point (4) is unstable for Jτ > 0, and changes its stability for either J = 0, or τ = 0.
In particular, given a non-zero synaptic time constant there is an oscillatory instability as the sign
of the synaptic coupling J changes from positive to negative. Therefore oscillations occur only
for inhibitory coupling [44–46]. The frequency along this Hopf bifurcation line is determined
entirely by the intrinsic firing rate of individual cells ωc = 2πR∗.
On the other hand, in the limit of fast synaptic kinetics, i.e. τd = 0 in Eq. (6), we find another
(cid:112)2τmR∗(J + 2π2τmR∗). This reflects the fact that oscillations
Hopf bifurcation with ωc = 1
τm
cannot be induced if the synaptic interactions are instantaneous. The left panel of Figure 4 shows
the phase diagram with the Hopf boundaries depicted in red, reflecting that oscillations are found
for all values of inhibitory coupling and for non-instantaneous synaptic kinetics.
9
2. Heterogeneous neurons
Once heterogeneity is added to the network the region of sustained oscillatory behavior shrinks,
see Fig.4, center and right. The red closed curves correspond to the Hopf bifurcations, which have
been obtained in parametric form from the characteristic equation (6), see Materials and Methods.
Note that for small levels of δ, oscillations are present in a closed region of the phase diagram,
and disappear for large enough values of τ (the synaptic time constant relative to the neuronal time
constant). Further increases in δ gradually reduce the region of oscillations until it fully disappears
at the critical value
(cid:18)∆
(cid:19)
(cid:113)
(8)
δc =
=
1
5
Θ
c
√
5 − 2
5 = 0.1453 . . . ,
which has been obtained analytically from the characteristic Eq. (6), see Materials and Meth-
ods. This result is consistent with numerical studies investigating oscillations in heterogeneous
inhibitory networks which indicate that gamma oscillations are fragile against the presence of
quenched heterogeneity [30, 32, 34].
In the following, we compare the phase diagrams of Fig. 4 with numerical results using het-
erogeneous ensembles of Wang-Buzs´aki neurons with first order synapses. Instead of using the
population mean firing rate or mean synaptic activation, in Fig. 5 we computed the amplitude of
the population mean membrane potential. This variable is less affected by finite-size fluctuations
and hence the regions of oscillations are more easily distinguishable. The results are summarized
in Fig. 5 for different values of δ and have been obtained by systematically increasing the cou-
pling strength k for fixed values of τd. The resulting phase diagrams qualitatively agree with those
shown in Fig. 4 . As predicted by the QIF-FRE, oscillations are found in a closed region in the
(τd, k) parameter space, and disappear for large enough values of δ. Here, the critical value of
δ = σ/ ¯I is about 6%, smaller than the critical value given by Eq. (8). This is due to the steep f-I
curve of the WB model, which implies a larger dispersion in the firing rates of the neurons even
for small heterogeneities in the input currents.
Additionally, for small τd (fast synaptic kinetics) and strong coupling k, we observed small
regions where the oscillations coexist with the asynchronous state -not shown. Numerical sim-
ulations indicate that this bistability is not present in the QIF-FRE. For strong coupling, and co-
existing with the asynchronous state, we also observed various clustering states, already reported
in the original paper of Wang & Buzs´aki [30]. Clustering in inhibitory networks has also been
observed in populations of conductance-based neurons with spike adaptation [47] or time delays
10
FIG. 5. Amplitude of the oscillations of the mean membrane potential for a population of N = 1000 WB
neurons. From left to right: δ = σ/ ¯I = 0, 0.05 and 0.06. Central and Right panels have σ = 0.01 µA/cm2.
See Materials and Methods for details.
[48]. The fact that such states do not emerge in the model Eqs. (12) may be due to the purely
sinusoidal shape of the phase resetting curve of the QIF model [49–54].
D. Firing Rate Equations in the limit of slow synapses
We have seen that the oscillations which emerge in inhibitory networks for sufficiently fast
synaptic kinetics are correctly described by the firing rate equations Eqs. (3), but not by the heuris-
tic Eqs. (1). The reason for this is that the oscillations crucially depend on the interaction between
the population firing rate and the subthreshold membrane potential during spike initiation and re-
set; this interaction manifests itself at the network level through spike synchrony. Therefore, if
one could suppress the spike synchrony mechanism, and hence the dependence on the subthresh-
old membrane potential, in Eqs. (3), the resulting equations ought to bear resemblance to Eqs. (1).
In fact, as we saw in Fig. 3, the two firing rate models become more similar when the synaptic
kinetics become slower.
Next we show that the two models become identical, formally, in the limit of slow synaptic
kinetics. To show this, we assume the synaptic time constant is slow, namely τd = ¯τd/ where
0 < (cid:28) 1, and rescale time as ¯t = t. In this limit we are tracking the slow synaptic dynamics in
while the neuronal dynamics are stationary to leading order, i.e.
R∗ = Φ(−JτmS + Θ).
Therefore, the dynamics reduce to the first order equation
S = −S + Φ(−JτmS + Θ).
τd
11
(9)
(10)
FIG. 6. The reduction of the QIF-FRE to Eq.(10) breaks down when neurons receive time-varying inputs.
Panels (a-c): S-variable time series for QIF-FRE (solid Black), H-FRE (dashed Black) and Eq. (10) (Blue),
for decreasing values of the period TΘ of the external periodic forcing Θ(t) = 4 + [1 + sin(2πt/Tθ)]3
-shown in panels (g-i). In all cases, the synaptic time constant is slow τd = 100 ms, compared to the
membrane time constant τm = 10 ms. Panels (d-f): R-variable time series. In the case of model Eq. (10),
the firing rate has been evaluated using Eq.(9). Other parameters are J = 21, ∆ = 0.3.
Notably, this shows that the QIF-FRE Eqs. (3), and the H-FRE (1), do actually have the same
dynamics in the limit of slow synapses. In other words, Eq. (10) is formally equivalent to the
Wilson-Cowan equations for a single inhibitory population, and this establishes a mathematical
link between the QIF-FRE and Heuristic firing rate descriptions. Additionally, considering slow
second order synaptic kinetics (not shown), allows one to reduce the QIF-FRE with either alpha or
double exponential synapses to a second-order system that formally corresponds to the so-called
neural mass models largely used for modeling EEG data, see e.g. [6, 55–58].
1. External inputs and breakdown of the slow-synaptic limit Eq. (10)
It is important to note that, in the derivation of Eq. (10) we considered external inputs Θ to be
constant. Then, if synapses are slow, the neuronal variables (R in the case of Eqs. (1) and R and
12
V in the case of Eqs. (3)) decay rapidly to their fixed point values. However even in the limit of
slow synapses, such reduction can break down if external inputs are time-varying Θ = Θ(t), with
a time-scale which itself is not sufficiently slow.
To demonstrate this, in Fig. 6, we compared the dynamics of the QIF-FRE and H-FRE with
the approximation Eq. (10), for periodic stimuli of various periods -panels (g-i)-, and always
considering slow synapses, τd = 100 ms. As expected, the models show good agreement for
very slow external inputs -see panels (a,d)-, but this discrepancy is strongly magnified for fast
external inputs Specifically, for fast inputs -see panels (c,f)-, the dynamics of the S and R
variables of the QIF-FRE are clearly different form those of the other models. This illustrates
that even in the limit of slow synapses, the response of spiking networks to arbitrary time-varying
inputs will always generate some degree of spike synchrony.
III. DISCUSSION
Firing rate models, describing the average activity of large neuronal ensembles are broadly used
in computational neuroscience. However, these models fail to describe inhibition-based rhythms,
typically observed in networks of inhibitory neurons with synaptic kinetics [30–38]. To overcome
this limitation, some authors heuristically included explicit delays in traditional FRE, and found
qualitative agreement with the oscillatory dynamics observed in simulations of spiking neurons
with both synaptic kinetics and fixed time delays [9, 10, 36, 39]. Nonetheless it remains unclear
why traditional H-FRE with first order synaptic kinetics do not generate inhibition-based oscilla-
tions.
Here we have investigated a novel class of FRE which can be rigorously derived from popu-
lations of spiking (QIF) neurons [40]. Networks of globally coupled QIF neurons with fast in-
hibitory synapses readily generate fast self-sustained oscillations. The corresponding exact FRE,
which we call the QIF-FRE, therefore also generates oscillations. The benefit of having a simple
macroscopic description for the network dynamics is its amenability to analysis. In particular,
the nonlinearities in Eqs.(3), which arise due to the spike initiation and reset mechanism in the
QIF model, conspire to generate damped oscillations which reflect transient spike synchrony in
the network. This oscillatory mode can be driven by sufficiently fast recurrent inhibitory synaptic
activation, leading to sustained oscillations. This suggests that any mean-field description of net-
work activity which neglects subthreshold integration will not properly capture spike-synchrony-
13
dependent dynamical behaviors, including fast inhibitory oscillations.
The QIF-FRE have also allowed us to generate a phase diagram for oscillatory behavior in a
network of QIF neurons with ease via a standard linear stability analysis, see Fig.4. This phase dia-
gram agrees qualitatively with that of an equivalent network of Wang-Buzs´aki neurons, suggesting
that the QIF-FRE not only provide an exact description of QIF networks, but also a qualitatively
accurate description of macroscopic behaviors in networks of Class I neurons in general. In par-
ticular, the QIF-FRE capture the fragility of oscillations to quenched variability in the network, a
feature that seems to be particularly pronounced for Class 1 neuronal models compared to neural
models with so-called Class 2 excitability [59].
Finally we have shown that the QIF-FRE and the heuristic H-FRE are formally equivalent in
the limit of slow synapses. In this limit the neuronal dynamics is slaved to the synaptic activation
and well-described by Eq. (10), as long as external inputs are stationary. In fact, in the absence
of quenched heterogeneity (∆ = 0), the approximation for slow synapses Eq. (10) is also fully
equivalent to the reduction for slow synapses in networks of Class 1 neurons derived by Ermentrout
in [60]. This further indicates that the QIF-FRE are likely valid for networks of Class 1 neurons in
general. However, we also show that in the more biologically plausible scenario of time-varying
external drive some degree of neuronal synchronization is generically observed, as in Fig. (6), and
the slow-synapse reduction Eq. (10) is not valid.
The results presented here are obtained under two important assumptions that need to be taken
into account when comparing our work to the existing literature on fast oscillations in inhibitory
networks. (i) A derivation of an exact firing rate model for a spiking neuron network is only possi-
ble for ensembles of QIF neurons, which are the canonical model for Class 1 excitability [45, 61].
Although many relevant computational studies on fast inhibitory oscillations also consider Class
1 neurons [30, 32, 34, 39, 62–64], experimental evidence indicates that fast spiking interneurons
are highly heterogeneous in their minimal firing rates in response to steady currents, and that a
significant fraction of them are Class 2 [65–68] -but see also [69]. (ii) Our derivation of the QIF-
FRE is valid for networks of globally coupled QIF neurons, with Lorentzian-distributed currents.
In this system inhibition-based oscillations are only possible when the majority of the neurons are
self-sustained oscillators, i.e. for Θ > 0 in Eq. (14), and are due to the frequency locking of a frac-
tion of the oscillators in the population [41, 42] -as it can be seen in the raster plot of Fig. 1(c).
In this state, the frequency of the cluster of synchronized oscillators coincides with the frequency
of the mean field. The value of the frequency itself is determined through an interplay between
14
single-cell resonance and network effects. Specifically, the synchronized neurons have intrinsic
spiking frequencies near that of the mean-field oscillation and hence are driven resonantly. This
collective synchronization differs from the so-called sparse synchronization observed in inhibitory
networks of identical Class 1 neurons under the influence of noise [34, 36, 62, 63]. In the sparsely
synchronized state neurons fire stochastically at very low rates, while the population firing rate
displays the fast oscillations as the ones reported here.
Oscillatory phenomena arising from single-cell resonances, and which reflect spike synchrony
at the population level, are ubiquitous in networks of spiking neurons. Mean-field theory for noise-
driven networks leading to a Fokker-Planck formalism, allows for a linear analysis of the response
of the network to weak stimuli when the network is in an asynchronous state [43, 70]. Reso-
nances can appear in the linear response when firing rates are sufficiently high or noise strength
sufficiently low. Recent work has sought to extend H-FRE in this regime by extracting the com-
plex eigenvalue corresponding to the resonance and using it to construct the linear operator of a
complex-valued differential equation, the real part of which is the firing rate [29]. Other work
has developed an expression for the response of spiking networks to external drive, which of-
ten generates resonance-related damped oscillations, through an eigenfunction expansion of the
Fokker-Planck equation [71]. Our approach is similar in spirit to such studies in that we also
work with a low dimensional description of the network response. In contrast to previous work
our equations are an exact description of the macroscopic behavior, although they are only valid
for networks of heterogeneous QIF neurons. Nonetheless, the QIF-FRE are simple enough to al-
low for an intuitive understanding of the origin of fast oscillations in inhibitory networks, and in
particular, of why these oscillations are not properly captured by H-FRE.
IV. ACKNOWLEDGMENTS
F.D. and E.M acknowledge support by the European Union's Horizon 2020 research and in-
novation programme under the Marie Skłodowska-Curie grant agreement No. 642563. A.R. ac-
knowledges a project grant from the Spanish ministry of Economics and Competitiveness, Grant
No. BFU2012-33413. A.R. has been partially funded by the CERCA progam of the Generalitat de
Catalunya. E.M. acknowledges the projects grants from the Spanish ministry of Economics and
Competitiveness, Grants No. PSI2016-75688-P and No. PCIN-2015-127.
15
V. MATERIALS AND METHODS
A. Populations of inhibitory Quadratic Integrate and Fire neurons
We model fast-spiking interneurons, the dynamics of which are well-described by the Hodgkin-
Huxley equations with only standard spiking currents. Many models of inhibitory neurons are
Class 1 excitable [72], including for example the Wang-Busz´aki (WB) [30], and the Morris-Lecar
models [73]. Class 1 models are characterized by the presence of a saddle-node bifurcation on an
invariant circle at the transition from quiescence to spiking. One consequence of this bifurcation
structure is the fact the spiking frequency can be arbitrarily low near threshold. Additionally, near
threshold the spiking dynamics are dominated by the time spent in the vicinity of the saddle-node
itself, allowing for a formal reduction in dimensionality from the full neuron model to a reduced
normal form equation for a saddle-node bifurcation [2, 45, 61]. This normal form, which is valid
for any Class 1 model near threshold, is known as the quadratic integrate-and-fire model (QIF).
Using a change of variables, the QIF model can be transformed to a phase model, called Theta-
Neuron model [74], which has an strictly positive Phase Resetting Curve (PRC). Neuron models
with strictly positive PRC are called Type 1 neurons, indicating that perturbations always produce
an advance (and not a delay) of their phase. In general, Class 1 neurons have a Type 1 PRC [45],
but see [75, 76].
In a network of QIF neurons, the neuronal membrane potentials are { Vi}i=1,...,N, which obey
the following ordinary differential equations [7, 64, 74]:
C
d Vi
dt
= gL
( Vi − Vt)( Vi − Vr)
(Vt − Vr)
+ I0,i
(11)
where C is the cell capacitance, gL is the leak conductance and I0,i are external currents. Addi-
tionally, Vr and Vt represent the resting potential and threshold of the neuron, respectively. Us-
i = Vi − (Vt + Vr)/2, and then rescaling the shifted voltages as
ing the change of variables V (cid:48)
Vi = V (cid:48)
i /(Vt − Vr), the QIF model (11) reduces to
τm
Vi = V 2
i + Ii
(12)
where τm = C/gL is the membrane time constant, Ii = I0,i/(gL(Vt − Vr)) − 1/4 and the overdot
represents derivation with respect to time t. Note that in the model (12) the voltage variables Vi
and the inputs Ii do not have dimensions. Thereafter we work with QIF model its simplest form
16
Eq. (12). We assume that the inputs are
Ii = ηi − JτmS,
(13)
where J is the inhibitory synaptic strength, and S is the synaptic gating variable. Finally, the
currents ηi are constants taken from some prescribed distribution that here we consider it is a
Lorentzian of half-width ∆, centered at Θ
g(η) =
1
π
∆
(η − Θ)2 + ∆2 .
(14)
In numerical simulations the currents were selected deterministically to represent the Lorentzian
distribution as: ηi = Θ + ∆ tan(π/2(2i − N − 1)/(N + 1)), for i = 1, . . . , N. In the absence
of synaptic input, the QIF model Eqs.(12,13) exhibits two possible dynamical regimes, depend-
√−ηi,
ing on the sign of ηi. If ηi < 0, the neuron is excitable, and an initial condition Vi(0) <
asymptotically approaches the resting potential −√−ηi. For initial conditions above the excitabil-
√−ηi, the membrane potential grows without bound. In this case, once the
ity threshold, Vi(0) >
neuron reaches a certain threshold value Vθ (cid:29) 1, it is reset to a new value −Vθ after a refractory
period 2τm/Vθ (in numerical simulations, we choose Vθ = 100). On the other hand, if ηj > 0, the
√
neuron behaves as an oscillator and, if Vθ → ∞, it fires regularly with a period T = πτm/
ηi.
The instantaneous population mean firing rate is
N(cid:88)
(cid:88)
(cid:90) t
R = lim
τs→0
1
N
1
τs
j=1
k
t−τs
dt(cid:48)δ(t(cid:48) − tk
j ),
(15)
j is the time of the kth spike of jth neuron, and δ(t) is the Dirac delta function. Finally,
where tk
the dynamics of the synaptic variable obeys the first order ordinary differential equation
S = −S + R.
τd
(16)
For the numerical implementation of Eqs. (15,16), we set τs = 10−2τm. To obtain a smoother time
series, the firing rate plotted in Fig. 3 was computed according to Eq. (15) with τs = 3 · 10−2τm.
B. Firing Rate Equations for populations of Quadratic Integrate and Fire neurons
Recently Luke et al. derived the exact macroscopic equations for a pulse-coupled ensemble
of Theta-Neurons [77], and this has motivated a considerable number of recent papers [78–86,
88]. This work applies the so-called Ott-Antonsen theory [89–91] to obtain a low-dimensional
17
description of the network in terms of the complex Kuramoto order parameter. Nevertheless, it is
is not obvious how these macroscopic descriptions relate to traditional H-FRE.
As we already mentioned, the Theta-neuron model exactly transforms to the Quadratic Inte-
grate and Fire (QIF) model by a nonlinear change of variables [45, 61, 74]. This transformation
establishes a map between the phase variable θi ∈ (−π, π] of a Theta neuron i, and the mem-
brane potential variable Vi ∈ (−∞, +∞) of the QIF model Eq. (12). Recently it was shown that,
under some circumstances, a change of variables also exists at the population level [40]. In this
case, the complex Kuramoto order parameter transforms into a novel order parameter, composed
of two macroscopic variables: The population-mean membrane potential V , and the population-
mean firing rate R. As a consequence of that, the Ott-Antonsen theory becomes a unique method
for deriving exact firing rate equations for ensembles of heterogeneous spiking neurons -see
also [92–94] for recent alternative approaches.
Thus far, the FRE for QIF neurons (QIF-FRE) have been successfully applied to investigate
the collective dynamics of populations of QIF neurons with instantaneous [40, 86, 87], time de-
layed [95] and excitatory synapses with fast synaptic kinetics [96]. However, to date the QIF-FRE
have not been used to explore the dynamics of populations of inhibitory neurons with synaptic
kinetics -but see [83] for a numerical investigation using the low-dimensional Kuramoto order
parameter description. The method for deriving the QIF-FRE corresponding to a population of
QIF neurons Eq. (12) is exact in the thermodynamic limit N → ∞, and, under the assumption
that neurons are all-to-all coupled. Additionally, if the parameters ηi in Eq. (13) (which in the
thermodynamic limit become a continuous variable) are assumed to be distributed according to
the Lorentzian distribution Eq. (14), the resulting QIF-FRE become two dimensional. For instan-
taneous synapses, the macroscopic dynamics of the population of QIF neurons (12) is exactly
described by the system of QIF-FRE [40]
∆
πτm
+ 2RV,
R =
V = V 2 − (πτmR)2 − JτmR + Θ,
τm
τm
(17a)
(17b)
where R is the mean firing rate and V the mean membrane potential in the network. With ex-
ponentially decaying synaptic kinetics the QIF-FRE Eqs. (17) become Eqs.(3). In our study, we
consider Θ > 0, so that the majority of the neurons are oscillatory -see Eq.(14).
18
1. Fixed points
The fixed points of the QIF-FRE (3) are obtained imposing R = V = S = 0. Substituting
this into Eqs. (3), we obtain the fixed point equation V ∗ = −∆/(2πτmR∗), the firing rate given
by Eq. (4) and S∗ = R∗. Note that for homogeneous populations, ∆ = 0, the f-I curve Eq. (5)
reduces to
(cid:112)I+,
1
π
Φ(I) =
which displays a clear threshold at I = 0 (Here, I+ = I if I ≥ 0, and vanishes for I < 0.) This
function coincides with the squashing function found by Ermentrout for homogeneous networks
of Class 1 neurons [60]. As expected, for heterogeneous networks, the well-defined threshold of
Φ(I) for ∆ = 0 is lost and the transfer function becomes increasingly smoother.
2. Nondimensionalized QIF-FRE
The QIF-FRE (3) have five parameters. It is possible to non-dimensionalize the equations so
that the system can be written solely in terms of 3 parameters. Generally, we adopt the following
notation: we use capital letters to refer to the original variables and parameters of the QIF-FRE,
and lower case letters for non-dimensional quantities. A possible non-dimensionalization, valid
for Θ > 0, is
r = δ/π + 2rv,
v = v2 − π2r2 − js + 1,
τ s = −s + r,
where the overdot here means differentiation with respect to the non-dimensional time
√
Θ
τm
t.
t =
The other variables are defined as
r =
τm√
Θ
R,
v =
V√
Θ
,
s =
τm√
Θ
S.
On the other hand, the new coupling parameter is defined as
j =
J√
Θ
.
19
(18a)
(18b)
(18c)
(19)
and the parameter
δ =
∆
Θ
,
(20)
describes the effect of the Lorentzian heterogeneity (14) into the collective dynamics of the
FRE (17). Though the Lorentzian distribution does not have finite moments, for the sake of com-
parison of our results with those of studies investigating the dynamics of heterogeneous networks
of inhibitory neurons, e.g.[30, 32], the quantity δ can be compared to the coefficient of variation,
which measures the ratio of the standard deviation to the mean of a probability density function.
Finally, the non-dimensional time
τ =
τd,
(21)
√
Θ
τm
measures the ratio of the synaptic time constant to the most-likely period of the neurons (times π),
¯T = π
τm√
Θ
.
In numerical simulations we will use the original QIF-FRE (3), with Θ = 4, and τm = 10ms.
Thus ¯T = 10π/3 ≈ 15.71ms, so that the most likely value of the neurons' intrinsic frequency
is ¯f ≈ 63.66 Hz. However, our results are expressed in a more compact form in terms of the
quantities j, δ, τ, and we will use them in some of our calculations and figures.
C. Parametric formula for the Hopf boundaries
To investigate the existence of oscillatory instabilities we use Eq. (6) written in terms of the
non-dimensional variables and parameters defined previously, which is
(cid:34)
(cid:18)
(cid:19)2(cid:35)
− 2jr∗ = (1 + λτ )
(2πr∗)2 +
λ +
δ
πr∗
.
(22)
Imposing the condition of marginal stability λ = iω in Eq. (22) gives the system of equations
0 = 2jr∗ + 4π2r2∗ + 4v2∗ − (1 − 4v∗τ )ω2
0 = ω(4v∗ − 4π2r2∗τ − 4v2∗τ + τ ω2)
where the fixed points are obtained from Eq. (4) solving
0 = v2∗ − π2r2∗ − jr∗ + 1,
(23a)
(23b)
(24)
with
v∗ = − δ
2πr∗
20
Eq. (23b) gives the critical frequency
ω =
2
τ
(cid:112)(πτ r∗)2 + τ v∗(τ v∗ − 1).
The Hopf boundaries can be plotted in parametric form solving Eq. (24) for j, and substituting j
and ω into Eq. (23a). Then solving Eq. (23a) for τ gives the Hopf bifurcation boundaries
τ±(r∗) =
Using the parametric formula
π2r2∗ − 1 + 7v2∗ ±(cid:112)(π2r2∗ − 1)2 − (14 + 50π2r2∗)v2∗ − 15v4∗
(cid:0)j(r∗), τ±(r∗)(cid:1)±
=(cid:0)v2∗/r∗ + 1/r∗ − π2r∗, τ±(r∗)(cid:1) .
16v∗(π2r2∗ + v2∗)
.
(25)
we can be plot the Hopf boundaries for particular values of the parameter δ, as r∗ is changed.
Figure 4 shows these curves in red, for δ = 0.05 and δ = 0.075. They define a closed region in
parameter space (shaded region) where oscillations are observed.
1. Calculation of the critical value δc, Eq (8)
The functions τ± meet at two points, when the argument of the square root in Eq. (25) is zero.
This gives four different roots for δ, and only one of them is positive and real
(cid:113)
8(cid:112)1 + 5π2r2∗ + 10π4r4∗ − 7 − 25π2r2∗.
δ∗(r∗) =
2πr∗√
15
This function has two positive zeros, one at r∗min = 0, and one at r∗max = 1/π, corresponding,
respectively, to the minimal (j → ∞) and maximal (j = 0) values of the firing rate for identical
neurons (δ = 0). Between these two points the function attains a maximum, where r∗min =
r∗max = r∗c, with
1(cid:112)
√
2
5π
r∗c =
= 0.1505 . . .
The function δ∗(r∗) evaluated at its local maximum r∗ = r∗c gives Eq. (8).
D. Populations of Wang-Buzski neurons
We perform numerical simulations using the the Wang-Buzs´aki (WB) neuron [30], and com-
pare them with our results using networks of QIF neurons. The onset of oscillatory behavior in
the WB model is via a saddle node on the invariant circle (SNIC) bifurcation. Therefore, the
21
populations of WB neurons near this bifurcation are expected to be well described by the theta-
neuron/QIF model, the canonical model for Class 1 neural excitability [45, 74].
We numerically simulated a network of N all-to-all coupled WB neurons, where the dynamics
of each neuron is described by the time evolution of its membrane potential [30]
Vi = −INa,i − IK,i − IL,i − Isyn + Iapp,i + I0.
Cm
The cell capacitance is Cm = 1 µF/cm2. The inputs Iapp (in µA/cm2) are distributed accord-
ing to a Lorentzian distribution with half width σ and center ¯I. In numerical simulations these
currents were selected deterministically to represent the Lorentzian distribution as Iapp,i = ¯I +
σ tan(π/2(2i− N − 1)/(N + 1)), for i = 1, . . . , N. The constant input I0 = 0.1601 µA/cm2 sets
the neuron at the SNIC bifurcation when Iapp = 0. The leak current is
IL,i = gL (Vi − EL) ,
with gL = 0.1 mS/cm2, so that the passive time constant τm = Cm/gL = 10 ms. The sodium
current is
INa,i = gNam3∞h (Vi − ENa) ,
where gNa = 35 mS/cm2, ENa = 55 mV, m∞ = αm/ (αm + βm) with αm (Vi) = −0.1 (Vi + 35) / (exp (−0.1 (Vi + 35) − 1)),
βm (Vi) =4 exp (− (Vi+60) /18). The inactivation variable h obeys the differential equation
h = φ (αh (1 − h) − βhh) ,
with φ = 5, αh (Vi) = 0.07 exp (− (Vi + 58) /20) and βh (Vi) = 1/ (exp (−0.1 (Vi + 28)) + 1).
The potassium current follows
IK,i = gKn4 (Vi − EK) ,
with gK = 9 mS/cm2, EK = −90 mV. The activation variable n obeys
n = φ (αn (1 − n) − βnn) ,
where αn (Vi) = −0.01 (Vi + 34) / (exp (−0.1 (Vi + 34)) − 1) and βn (Vi) = 0.125 exp (− (Vi + 44) /80).
The synaptic current is Isyn = kCmS, where the synaptic activation variable S obeys the first
order kinetics Eq. (16) and k is the coupling strength (expressed in mV). The factor Cm ensures
that the effect of an incoming spike to the neuron is independent from its passive time constant.
The neuron is defined to emit a spike when its membrane potential crosses 0 mV. The population
22
firing rate is then computed according to Eq. (15), with τs = 10−2 ms. In numerical simulations
we considered N = 1000 all-to-all coupled WB neurons, using the Euler method with time step
dt = 0.001 ms. In Fig. 1, the membrane potentials were initially randomly distributed according
to a Lorentzian function with half width 5 mV and center −62 mV. Close to the bifurcation
point, this is equivalent to uniformly distribute the phases of the corresponding Theta-Neurons in
[−π, π][2, 7, 61, 74]. The parameters were chosen as ¯I = 0.5 µA/cm2, σ = 0.01 µA/cm2 and
k = 6 mV. The population firing rate was smoothed setting τs = 2 ms in Eq.(15).
In Fig. 5, we systematically varied the coupling strength and the synaptic time decay constant
to determine the range of parameters displaying oscillatory behavior. For each fixed value of τd
we varied the coupling strength k; we performed two series of simulations, for increasing and
decreasing coupling strength. In Fig. 5 we only show results for increasing k.
All quantities were measured after a transient of 1000 ms. To obtain the amplitude of the
oscillations of the mean membrane potential, we computed the maximal amplitude ¯Vmax − ¯Vmin
over time windows of 200 ms for 1000 ms, and then averaged over the five windows.
23
[1] Wilson HR, Cowan JD. Excitatory and inhibitory interactions in localized populations of model neu-
rons. Biophys J. 1972;12(1):1–24.
[2] Ermentrout GB, Terman DH. Mathematical foundations of neuroscience. vol. 64. Springer; 2010.
[3] Gerstner W, Kistler WM, Naud R, Paninski L. Neuronal dynamics: From single neurons to networks
and models of cognition. Cambridge University Press; 2014.
[4] Dayan P, Abbott LF. Theoretical neuroscience. Cambridge, MA: MIT Press; 2001.
[5] Cowan J. A personal account of the development of the field theory of large-scale brain activity from
1945 onward. In: Neural fields. Springer; 2014. p. 47–96.
[6] Coombes S, beim Graben P, Potthast R. Tutorial on neural field theory. In: Neural fields. Springer;
2014. p. 1–43.
[7] Latham P, Richmond B, Nelson P, Nirenberg S. Intrinsic dynamics in neuronal networks. I. Theory.
Journal of Neurophysiology. 2000;83(2):808–827.
[8] Shriki O, Hansel D, Sompolinsky H. Rate models for conductance-based cortical neuronal networks.
Neural Comput. 2003;15(8):1809–1841.
[9] Roxin A, Brunel N, Hansel D. Role of delays in shaping spatiotemporal dynamics of neuronal activity
in large networks. Phys Rev Lett. 2005;94(23):238103.
[10] Roxin A, Montbri´o E. How effective delays shape oscillatory dynamics in neuronal networks. Physica
D. 2011;240(3):323–345.
[11] Wilson HR, Cowan JD. A mathematical theory of the functional dynamics of cortical and thalamic
nervous tissue. Kybernetik. 1973;13(2):55–80. doi:10.1007/BF00288786.
[12] Amari Si.
A method of
statistical neurodynamics.
Kybernetik. 1974;14(4):201–215.
doi:10.1007/BF00274806.
[13] Nunez PL. The brain wave equation: a model for the EEG. Mathematical Biosciences. 1974;21(3):279
– 297. doi:http://dx.doi.org/10.1016/0025-5564(74)90020-0.
[14] Ermentrout GB, Cowan JD. A mathematical theory of visual hallucination patterns. Biological Cy-
bernetics. 1979;34(3):137–150. doi:10.1007/BF00336965.
[15] Ben-Yishai R, Bar-Or RL, Sompolinsky H. Theory of orientation tuning in visual cortex. Proc Nat
Acad Sci. 1995;92(9):3844–3848.
[16] Pinto DJ, Brumberg JC, Simons DJ, Ermentrout GB, Traub R. A quantitative population model of
24
whisker barrels: Re-examining the Wilson-Cowan equations. Journal of Computational Neuroscience.
1996;3(3):247–264. doi:10.1007/BF00161134.
[17] Hansel D, Sompolinsky H. Modeling Feature Selectivity in Local Cortical Circuits. In: Koch C, Segev
I, editors. Methods in Neuronal Modelling: From Ions to Networks. Cambridge: MIT Press; 1998. p.
499–567.
[18] Tsodyks M MH Pawelzik K. Neural networks with dynamic synapses. Neural Comput. 1998;10:821.
[19] Wilson HR. Spikes, decisions, and actions: the dynamical foundations of neurosciences. 1999;.
[20] Tabak J, Senn W, ODonovan MJ, Rinzel J. Modeling of spontaneous activity in developing spinal
cord using activity-dependent depression in an excitatory network. J Neurosci. 2000;20:3041–3056.
[21] Bressloff PC, Cowan JD, Golubitsky M, Thomas PJ, Wiener MC. Geometric visual halluci-
nations, Euclidean symmetry and the functional architecture of striate cortex.
Philosophical
Transactions of the Royal Society of London B: Biological Sciences. 2001;356(1407):299–330.
doi:10.1098/rstb.2000.0769.
[22] Laing CR, Troy WC, Gutkin B, Ermentrout GB. Multiple bumps in a neuronal model of working
memory. SIAM Journal on Applied Mathematics. 2002;63(1):62–97.
[23] Holcman D, Tsodyks M. The emergence of up and down states in cortical networks. PLoS Comput
Biol. 2006;2(3):e23.
[24] Moreno-Bote R, Rinzel J, Rubin N. Noise-induced alternations in an attractor network model of
perceptual bistability. J Neurophysiol. 2007;98(3):1125–1139.
[25] Mongillo G, Barak O, Tsodyks M.
Synaptic theory of working memory.
Science.
2008;319(5869):1543–1546.
[26] Touboul J, Wendling F, Chauvel P, Faugeras O. Neural mass activity, bifurcations, and epilepsy.
Neural computation. 2011;23(12):3232–3286.
[27] Mart´ı D, Rinzel J. Dynamics of feature categorization. Neural computation. 2013;25(1):1–45.
[28] Ton R, Deco G, Daffertshofer A. Structure-function discrepancy: inhomogeneity and delays in syn-
chronized neural networks. PLOS Comput Biol. 2014;10(7):e1003736.
[29] Schaffer ES, Ostojic S, Abbott L. A Complex-Valued Firing-Rate Model That Approximates the
Dynamics of Spiking Networks. PLoS Comput Biol. 2013;9(10):e1003301.
[30] Wang XJ, Buzs´aki G. Gamma oscillation by synaptic inhibition in a hippocampal interneuronal net-
work model. The journal of Neuroscience. 1996;16(20):6402–6413.
[31] Whittington MA, Traub RD, Jefferys JG. Synchronized oscillations in interneuron networks driven
25
by metabotropic glutamate receptor activation. Nature. 1995;373:612–615.
[32] White JA, Chow CC, Rit J, Soto-Trevino C, Kopell N. Synchronization and oscillatory dynamics in
heterogeneous, mutually inhibited neurons. Journal of computational neuroscience. 1998;5(1):5–16.
[33] Whittington MA, Traub RD, Kopell N, Ermentrout B, Buhl EH.
Inhibition-based rhythms: ex-
perimental and mathematical observations on network dynamics.
Int Journal of Psychophysiol.
2000;38(3):315 – 336. doi:http://dx.doi.org/10.1016/S0167-8760(00)00173-2.
[34] Tiesinga P, Jos´e JV. Robust gamma oscillations in networks of inhibitory hippocampal interneurons.
Network: Computation in Neural Systems. 2000;11(1):1–23.
[35] Brunel N, Hansel D. How noise affects the synchronization properties of recurrent networks of in-
hibitory neurons. Neural Comput. 2006;18(5):1066–1110.
[36] Brunel N, Hakim V. Sparsely synchronized neuronal oscillations. Chaos: An Interdisciplinary Journal
of Nonlinear Science. 2008;18(1):015113.
[37] Bartos M, Vida I, Jonas P. Synaptic mechanisms of synchronized gamma oscillations in inhibitory
interneuron networks. Nature reviews neuroscience. 2007;8(1):45–56.
[38] Wang XJ. Neurophysiological and computational principles of cortical rhythms in cognition. Physio-
logical reviews. 2010;90(3):1195–1268.
[39] Keeley S, Fenton AA, Rinzel J. Modeling fast and slow gamma oscillations with interneurons of
different subtype. Journal of Neurophysiology. 2017;117(3):950–965. doi:10.1152/jn.00490.2016.
[40] Montbri´o E, Paz´o D, Roxin A. Macroscopic Description for Networks of Spiking Neurons. Phys Rev
X. 2015;5:021028. doi:10.1103/PhysRevX.5.021028.
[41] Winfree AT. Biological rhythms and the behavior of populations of coupled oscillators. J Theor Biol.
1967;16:15–42.
[42] Kuramoto Y. Chemical Oscillations, Waves, and Turbulence. Berlin: Springer-Verlag; 1984.
[43] Ledoux E, Brunel N. Dynamics of networks of excitatory and inhibitory neurons in response to time-
dependent inputs. Frontiers Comp Neurosci. 2011;5:25.
[44] Van Vreeswijk C, Abbott LF, Bard Ermentrout G. When inhibition not excitation synchronizes neural
firing. Journal of Computational Neuroscience. 1994;1(4):313–321. doi:10.1007/BF00961879.
[45] Ermentrout B. Type I membranes, phase resetting curves, and synchrony. Neural Comp. 1996;8:979–
1001.
[46] Hansel D, Mato G, Meunier C. Synchrony in excitatory neural networks. Neural Comput. 1995;7:307–
337.
26
[47] Kilpatrick ZP, Ermentrout B. Sparse Gamma Rhythms Arising through Clustering in Adapting Neu-
ronal Networks. PLoS Comput Biol. 2011;7(11):e1002281. doi:10.1371/journal.pcbi.1002281.
[48] Ernst U, Pawelzik K, Geisel T. Delay-induced multistable synchronization of biological oscillators.
Physical review E. 1998;57(2):2150.
[49] Okuda K. Variety and generality of clustering in globally coupled oscillators. Physica D: Nonlinear
Phenomena. 1993;63(3-4):424–436.
[50] Hansel D, Mato G, Meunier C. Clustering and slow switching in globally coupled phase oscillators.
Phys Rev E. 1993;48:3470–3477. doi:10.1103/PhysRevE.48.3470.
[51] Kori H, Kuramoto Y. Slow switching in globally coupled oscillators: robustness and occurrence
through delayed coupling. Phys Rev E. 2001;63:046214. doi:10.1103/PhysRevE.63.046214.
[52] Kori H. Slow switching and broken cluster state in a population of neuronal oscillators. Int J Mod
Phys B. 2003;17:4238–4241. doi:10.1142/S0217979203022246.
[53] Politi A, Rosenblum M. Equivalence of phase-oscillator and integrate-and-fire models. Phys Rev E.
2015;91:042916. doi:10.1103/PhysRevE.91.042916.
[54] Clusella P, Politi A, Rosenblum M. A minimal model of self-consistent partial synchrony. New J
Phys. 2016;18(9):093037.
[55] Freeman WJ. Mass action in the nervous system. Academic Press, New York; 1975.
[56] Jansen BH, Rit VG. Electroencephalogram and visual evoked potential generation in a math-
ematical model of coupled cortical columns.
Biological Cybernetics. 1995;73(4):357–366.
doi:10.1007/BF00199471.
[57] Robinson PA, Rennie CJ, Wright JJ. Propagation and stability of waves of electrical activity in the
cerebral cortex. Phys Rev E. 1997;56:826–840. doi:10.1103/PhysRevE.56.826.
[58] Ashwin P, Coombes S, Nicks R. Mathematical Frameworks for Oscillatory Network Dynamics in
Neuroscience. The Journal of Mathematical Neuroscience. 2016;6(1):1–92. doi:10.1186/s13408-015-
0033-6.
[59] Tikidji-Hamburyan RA, Mart´ınez JJ, White JA, Canavier CC. Resonant Interneurons Can In-
crease Robustness of Gamma Oscillations.
Journal of Neuroscience. 2015;35(47):15682–15695.
doi:10.1523/JNEUROSCI.2601-15.2015.
[60] Ermentrout B. Reduction of conductance-based models with slow synapses to neural nets. Neural
Comput. 1994;6(4):679–695.
[61] Izhikevich EM. Dynamical Systems in Neuroscience. Cambridge, Massachusetts: The MIT Press;
27
2007.
[62] Brunel N, Hakim V. Fast global oscillations in networks of integrate-and-fire neurons with low firing
rates. Neural Comput. 1999;11(7):1621–1671.
[63] Brunel N, Wang XJ. What determines the frequency of fast network oscillations with irregular neural
discharges? I. Synaptic dynamics and excitation-inhibition balance.
Journal of neurophysiology.
2003;90(1):415–430.
[64] Hansel D, Mato G. Asynchronous states and the emergence of synchrony in large networks of inter-
acting excitatory and inhibitory neurons. Neural Computation. 2003;15(1):1–56.
[65] Golomb D, Donner K, Shacham L, Shlosberg D, Amitai Y, Hansel D. Mechanisms of firing patterns
in fast-spiking cortical interneurons. PLoS Computational Biology. 2007;3(8):e156.
[66] Tateno T, Harsch A, Robinson HPC. Threshold Firing Frequency–Current Relationships of Neu-
rons in Rat Somatosensory Cortex: Type 1 and Type 2 Dynamics.
Journal of Neurophysiology.
2004;92(4):2283–2294. doi:10.1152/jn.00109.2004.
[67] Tateno T, Robinson HPC. Phase Resetting Curves and Oscillatory Stability in Interneurons of Rat
Somatosensory Cortex. Biophys J. 2007;92(2):683–695. doi:10.1529/biophysj.106.088021.
[68] Mancilla JG, Lewis TJ, Pinto DJ, Rinzel J, Connors BW. Synchronization of Electrically Cou-
pled Pairs of Inhibitory Interneurons in Neocortex. Journal of Neuroscience. 2007;27(8):2058–2073.
doi:10.1523/JNEUROSCI.2715-06.2007.
[69] La Camera G, Rauch A, Thurbon D, Luscher HR, Senn W, Fusi S. Multiple Time Scales of
Temporal Response in Pyramidal and Fast Spiking Cortical Neurons. Journal of Neurophysiology.
2006;96(6):3448–3464. doi:10.1152/jn.00453.2006.
[70] Ostojic S, Brunel N. From spiking neuron models to linear-nonlinear models. PLoS Comput Biol.
2011;7(1):e1001056.
[71] Mattia M, Del Giudice P. Population dynamics of interacting spiking neurons. Phys Rev E.
2002;66:051917. doi:10.1103/PhysRevE.66.051917.
[72] Rinzel J, Ermentrout B. Analysis of neural excitability and oscillations. In: Koch C, Segev I, editors.
Methods in Neuronal Modelling: From Ions to Networks. Cambridge: MIT Press; 1989. p. 135–171.
[73] Morris C, Lecar H. Voltage oscillations in the barnacle giant muscle fiber. Biophysical journal.
1981;35(1):193–213.
[74] Ermentrout B, Kopell N. Parabolic bursting in an excitable system coupled with a slow oscillation.
SIAM J Appl Math. 1986;46:233–253.
28
[75] Achuthan S, Butera RJ, Canavier CC.
Synaptic and intrinsic determinants of the phase re-
setting curve for weak coupling.
Journal of Computational Neuroscience. 2011;30(2):373–390.
doi:10.1007/s10827-010-0264-1.
[76] Ermentrout GB, Glass L, Oldeman BE. The Shape of Phase-Resetting Curves in Oscillators with
a Saddle Node on an Invariant Circle Bifurcation. Neural Computation. 2012;24(12):3111–3125.
doi:10.1162/NECO a 00370.
[77] Luke TB, Barreto E, So P. Complete classification of the macroscopic behavior of a heterogeneous
network of theta neurons. Neural Comput. 2013;25(12):3207–3234.
[78] So P, Luke TB, Barreto E.
Networks of
theta neurons with time-varying excitability:
Macroscopic chaos, multistability, and final-state uncertainty.
Physica D. 2014;267(0):16–26.
doi:http://dx.doi.org/10.1016/j.physd.2013.04.009.
[79] Laing CR. Derivation of a neural field model from a network of theta neurons. Phys Rev E.
2014;90:010901. doi:10.1103/PhysRevE.90.010901.
[80] Laing CR. Exact Neural Fields Incorporating Gap Junctions. SIAM Journal on Applied Dynamical
Systems. 2015;14(4):1899–1929.
[81] Laing CR. Travelling waves in arrays of delay-coupled phase oscillators. Chaos. 2016;26(9).
doi:http://dx.doi.org/10.1063/1.4953663.
[82] Laing CR. Bumps in Small-World Networks. Frontiers in Computational Neuroscience. 2016;10:53.
doi:10.3389/fncom.2016.00053.
[83] Coombes S, Byrne ´A. Next generation neural mass models. in Lecture Notes in Nonlinear Dynamics
in Computational Neuroscience: from Physics and Biology to ICT Springer (In Press).
[84] Roulet J, Mindlin GB. Average activity of excitatory and inhibitory neural populations. Chaos: An
Interdisciplinary Journal of Nonlinear Science. 2016;26(9):093104. doi:10.1063/1.4962326.
[85] O'Keeffe KP, Strogatz SH. Dynamics of a population of oscillatory and excitable elements. Phys Rev
E. 2016;93:062203. doi:10.1103/PhysRevE.93.062203.
[86] Pietras B, Daffertshofer A.
Ott-Antonsen attractiveness for parameter-dependent oscillatory
systems.
Chaos: An Interdisciplinary Journal of Nonlinear Science. 2016;26(10):103101.
doi:10.1063/1.4963371.
[87] Esnaola-Acebes JM, Roxin A, Avitabile D, Montbri´o E. Synchrony-induced modes of oscillation of a
neural field model. Phys Rev E. 2017;96:052407. doi:10.1103/PhysRevE.96.052407.
[88] Chandra S, Hathcock D, Crain K, Antonsen TM, Girvan M, Ott E. Modeling the network dy-
29
namics of pulse-coupled neurons. Chaos: An Interdisciplinary Journal of Nonlinear Science.
2017;27(3):033102. doi:10.1063/1.4977514.
[89] Ott E, Antonsen TM. Low dimensional behavior of large systems of globally coupled oscillators.
Chaos. 2008;18(3):037113. doi:10.1063/1.2930766.
[90] Ott E, Antonsen TM. Long time evolution of phase oscillator systems. Chaos. 2009;19(2):023117.
doi:10.1063/1.3136851.
[91] Ott E, Hunt BR, Antonsen TM. Comment on "Long time evolution of phase oscillators systems".
Chaos. 2011;21:025112.
[92] Mattia M. Low-dimensional firing rate dynamics of spiking neuron networks.
arXiv preprint
arXiv:160908855. 2016;.
[93] Augustin M, Ladenbauer J, Baumann F, Obermayer K. Low-dimensional spike rate models derived
from networks of adaptive integrate-and-fire neurons: comparison and implementation. PLOS Com-
putational Biology. 2017;13(6). doi:10.1371/journal.pcbi.1005545.
[94] Schwalger T, Deger M, Gerstner W. Towards a theory of cortical columns: From spiking neurons
to interacting neural populations of finite size. PLOS Computational Biology. 2017;13(4):1–63.
doi:10.1371/journal.pcbi.1005507.
[95] Paz´o D, Montbri´o E.
From Quasiperiodic Partial Synchronization to Collective Chaos
in Populations of
Inhibitory Neurons with Delay.
Phys Rev Lett. 2016;116:238101.
doi:10.1103/PhysRevLett.116.238101.
[96] Ratas I, Pyragas K. Macroscopic self-oscillations and aging transition in a network of
synaptically coupled quadratic integrate-and-fire neurons.
Phys Rev E. 2016;94:032215.
doi:10.1103/PhysRevE.94.032215.
30
|
1907.13046 | 1 | 1907 | 2019-07-30T16:08:29 | Visual Information flow in Wilson-Cowan networks | [
"q-bio.NC"
] | In this work we study the communication efficiency of a psychophysically-tuned cascade of Wilson-Cowan and Divisive Normalization layers that simulate the retina-V1 pathway. This is the first analysis of Wilson-Cowan networks in terms of multivariate total correlation. The parameters of the cortical model have been derived through the relation between the steady state of the Wilson-Cowan model and the Divisive Normalization model.
The communication efficiency has been analyzed in two ways: First, we provide an analytical expression for the reduction of the total correlation among the responses of a V1-like population after the application of the Wilson-Cowan interaction. Second, we empirically study the efficiency with visual stimuli and statistical tools that were not available before: (1) we use a recent, radiometrically calibrated, set of natural scenes, and (2) we use a recent technique to estimate the multivariate total correlation in bits from sets of visual responses which only involves univariate operations, thus giving better estimates of the redundancy.
The theoretical and the empirical results show that although this cascade of layers was not optimized for statistical independence in any way, the redundancy between the responses gets substantially reduced along the neural pathway. Specifically, we show that (1)~the efficiency of a Wilson-Cowan network is similar to its equivalent Divisive Normalization model, (2) while initial layers (Von-Kries adaptation and Weber-like brightness) contribute to univariate equalization, the bigger contributions to the reduction in total correlation come from the computation of nonlinear local contrast and the application of local oriented filters, and (3)~psychophysically-tuned models are more efficient (reduce more total correlation) in the more populated regions of the luminance-contrast plane. | q-bio.NC | q-bio | Visual Information flow in Wilson-Cowan networks
A. Gomez-Villa,a M. Bertalmío,a J. Malob,1
aDept. Inf. Comm. Tech. Universitat Pompeu Fabra, Barcelona, Spain
bImage Processing Lab, Universitat de València, Valencia, Spain
E-mail: [email protected], [email protected],
[email protected]
Abstract:
In this work we study the communication efficiency of a psychophysically-tuned cascade of
Wilson-Cowan and Divisive Normalization layers that simulate the retina-V1 pathway. This
is the first analysis of Wilson-Cowan networks in terms of multivariate total correlation.
The parameters of the cortical model have been derived through the relation between the
steady state of the Wilson-Cowan model and the Divisive Normalization model.
The communication efficiency has been analyzed in two ways: First, we provide an ana-
lytical expression for the reduction of the total correlation among the responses of a V1-like
population after the application of the Wilson-Cowan interaction. Second, we empirically
study the efficiency with visual stimuli and statistical tools that were not available before:
(1) we use a recent, radiometrically calibrated, set of natural scenes, and (2) we use a recent
technique to estimate the multivariate total correlation in bits from sets of visual responses
which only involves univariate operations, thus giving better estimates of the redundancy.
The theoretical and the empirical results show that although this cascade of layers was
not optimized for statistical independence in any way, the redundancy between the responses
gets substantially reduced along the neural pathway. Specifically, we show that (1) the
efficiency of a Wilson-Cowan network is similar to its equivalent Divisive Normalization
model, (2) while initial layers (Von-Kries adaptation and Weber-like brightness) contribute
to univariate equalization, the bigger contributions to the reduction in total correlation come
from the computation of nonlinear local contrast and the application of local oriented filters,
and (3) psychophysically-tuned models are more efficient (reduce more total correlation) in
the more populated regions of the luminance-contrast plane. These results are an alternative
confirmation of the Efficient Coding Hypothesis for the Wilson-Cowan systems. And from
an applied perspective, they suggest that neural field models could be an option in image
coding to perform image compression.
9
1
0
2
l
u
J
0
3
]
.
C
N
o
i
b
-
q
[
1
v
6
4
0
3
1
.
7
0
9
1
:
v
i
X
r
a
1Corresponding author
Contents
1 Introduction
2 Models, materials and methods
2.1 Model: A psychophysically-tuned Wilson-Cowan network
2.2 Material: radiometrically calibrated stimuli and color adaptation
2.3 Method: measuring Total Correlation with Gaussianization transforms
3 Results: Redundancy reduction via the Wilson-Cowan interaction
3.1 Theoretical analysis
3.1.1 Expression for Total Correlation in Wilson-Cowan systems.
3.1.2 Experiment in a reduced-scale system.
3.2 Empirical analysis of the full-scale model
3.2.1 Marginal equalization
3.2.2 Mutual Information along the network
3.2.3 Total Correlation along the network
4 Discussion
A Appendix: Psychophysically inspired reduced-scale model
1
2
2
8
10
11
11
11
12
16
16
18
21
24
26
1
Introduction
The Wilson-Cowan equations [1] and Divisive Normalization [2] are alternative influential
models of interaction between cortical neurons. While the Divisive Normalization compu-
tation has been extensively studied from an information theoretic perspective [3 -- 7], the
efficiency of Wilson-Cowan networks has not been analyzed in such detail through accurate
redundancy measures.
Such analysis is interesting not only because of the classical Efficient Coding Hypothe-
sis [8 -- 10], which is the information-theoretic version of the free-energy principle [11], but
also because of practical applications. Note that equivalent analysis in the Divisive Nor-
malization case led to substantial improvements in image compression using brain inspired
architectures [12, 13]. In a low noise context [14 -- 16] optimal systems are those that mini-
mize the redundancy among the components of the signal representation [5, 17, 18].
In this paper we quantify how effective is each layer of the considered network in
achieving the density factorization goal. This information theoretic analysis is applied to a
recent psychophysically-tuned Wilson-Cowan model [19, 20], using a recent radiometrically
calibrated database of visual scenes [21, 22], and using a recent accurate statistical tool to
measure redundancy [23, 24].
-- 1 --
The appropriate (multivariate) concept to describe the redundancy in vectors (or neu-
ral populations) is the total correlation [25, 26]. Direct computation of total correlation
from its definition is not straightforward because it involves multivariate probability den-
sity estimation. That is why, in the past, different computationally convenient (bivariate)
surrogates of the total correlation have been used. For instance, the conditional proba-
bility between pairs of responses [3, 4, 27 -- 29], the correlation between the energies of the
responses [30 -- 34], or the mutual information between pairs of responses [5, 12, 35].
Now, by using a recent multivariate Gaussianization technique [23, 24], total correlation
can be estimated from (easy to compute) marginal equalizations, so we can use the proper
measure of communication efficiency in psychophysically-tuned networks. This study, which
measures total correlation on Wilson-Cowan responses for the first time, generalizes previous
analysis that used total correlation but were restricted to linear models to avoid estimation
problems [17], or addressed simpler (just-one-layer) linear+nonlinear models with bivariate
mutual information measures [5].
2 Models, materials and methods
In this section we first describe the considered model for the retina-V1 pathway that consists
of a cascade of equivalent Divisive Normalization and Wilson-Cowan modules. Then, we
present the visual stimuli over which the considered model will be applied. And finally, we
will review the statistical methods to assess the communication efficiency of the model.
2.1 Model: A psychophysically-tuned Wilson-Cowan network
In this work the theory is illustrated in the context of models of the retina-cortex pathway.
The considered framework follows the program suggested in [36] and implemented in [7]: a
cascade of four isomorphic linear+nonlinear modules. These modules address brightness,
contrast, frequency filtered contrast masked in the spatial domain, and orientation/scale
masking. An example of the transforms of the input in such models is shown in Fig. 1.
In this illustration the input is the spatial distribution of the spectral irradiance at the
retina. This input undergoes the following transforms: (1) The linear part of the first layer
consist of three positive LMS spectral sensitivities and a linear recombination of the LMS
values with positive/negative weights. This leads to three tristimulus values in each spatial
location: one of them is proportional to the luminance, and the other two have opponent
chromatic meaning (red-green and yellow-blue). The linear tristimulus values are normal-
ized by the corresponding values of the white in the scene (Von-Kries adaptation), and then,
the normalized opponent responses undergo parallel saturation transforms. Perception of
brightness is mediated by an adaptive Weber-like nonlinearity applied to the luminance at
each location. This nonlinearity enhances the response in the regions with small linear in-
put (low luminance). (2) The linear part of the second layer computes the deviation of the
brightness at each location from the local brightness. Then, this deviation is nonlinearly
normalized by the local brightness to give the local nonlinear contrast. (3) The responses to
local contrast are convolved by center surround receptive fields (or filtered by the Contrast
Sensitivity Function). Then the linearly filtered contrast is nonlinearly normalized by the
-- 2 --
Figure 1: Cascade of Linear+Nonlinear layers.
The network addresses in turn
(1) tristimulus-from-irradiance integrals and color opponency + Von-Kries adaptation together with
Weber-like brightness and nonlinear opponent channels, (2) local contrast (subtraction of local mean
+ division by local mean), (3) CSF filter + masking in the spatial domain, and (4) Wavelet filters
+ masking in the wavelet domain. See details on the formulation of each Divisive Normalization
layer in [7]. In this work we substitute the last nonlinearity by the equivalent Wilson-Cowan model.
-- 3 --
local contrast. Again normalization increases the response in the regions with small input
(low contrast). (4) After a linear wavelet transform modelling the response of simple cells
in V1, each response is normalized by the activity of the neurons in the surround. Again,
the activity relatively increases in the regions with low input. As shown in the marginal
and joint PDFs below, the common effect of the nonlinear modules throughout the network
is response equalization.
Divisive Normalization is the conventional model used to describe the nonlinearities
in contrast perception psychophysics [37, 38], but here we will also explore the equivalent
Wilson-Cowan model in the last layer.
Below we introduce the notation of both interaction models (the Divisive Normaliza-
tion and the Wilson-Cowan models), and the relation between them so that we can infer
psychophysically plausible parameters for the Wilson-Cowan model from the parameters
already tuned for Divisive Normalization.
In the case of the V1 cortex, we refer to the set of
Modelling cortical interactions.
responses of a population of simple cells as the vector r. The considered models (Divisive
Normalization and Wilson-Cowan) define a nonlinear mapping, N , that transforms the
input vector r (before the interaction among neurons) into the output vector x (after the
interaction),
N '
r
x
(2.1)
In this setting, responses are called excitatory or inhibitory, depending on the corresponding
sign of the signal: r = sign(r)r, and x = sign(x)x. The map N is an adaptive saturating
transform, but it preserves the sign of the responses (i.e. sign(x) = sign(r)). Therefore,
the models care about cell activation (the modulus · ) but not about the excitatory or
inhibitory nature of the sensors (the sign(·) = ±).
We will refer to as the energy of the input responses to the vector e = rγ, where this
is an element-wise exponentiation of the amplitudes ri. Given the sign-preserving nature
of the nonlinear mapping, for the sake of simplicity in notation, in the rest of the paper the
variables r and x refer to the activations r and x.
The Divisive Normalization model. Forward transform: The input-output transform
in the Divisive Normalization is (in matrix notation [7]),
x = Dk · D−1
(b+H·e) · e
(2.2)
where the output vector of nonlinear activations in V1, x, depends on the energy of the
input linear wavelet responses, e, which are dimension-wise normalized by a sum of neighbor
energies. Note that in this matrix notation, Dv, stands for a diagonal matrix with the
vector, v, in the diagonal. The non-diagonal nature of the interaction kernel H in the
denominator, b + H · e, implies that the i-th element of the response may be attenuated
if the activity of the neighbor sensors, ej with j (cid:54)= i, is high. Each row of the kernel H
describes how the energies of the neighbor simple cells attenuate the activity of each simple
-- 4 --
'
cell after the interaction. The each element of the vectors b and k respectively determine
the semisaturation and the dynamic range of the nonlinear response of each sensor.
Inverse transform: The relation between the two models is easier to obtain by iden-
tifying the corresponding decoding transforms in both models. In the case of the Divisive
Normalization, the analytical inverse is [7, 12],
(2.3)
e =(cid:0)I − D−1
k · Dx · H(cid:1)−1 · Db · D−1
k · x
The Wilson-Cowan model. Dynamical system: In the Wilson-Cowan model the vari-
ation of the activation vector, x, increases with the energy of the input, e, but, for each
sensor, this variation is also moderated by its own activity and by a linear combination of
the activities of the neighbor sensors,
x = e − Dα · x − W · f (x)
(2.4)
where W is the matrix that describes the damping factor between sensors, and f (x) is a
dimension-wise saturating nonlinearity (see Fig. 2). Note that in Eq. 2.4 both the inhibitory
and the excitatory responses are considered just as negative and positive components of the
same vector. Therefore, following [39], the two equations in the traditional Wilson-Cowan
formulation are represented here by a single expression.
Steady state and inverse. The stationary solution of the above differential equation,
x = 0 in Eq. 2.4, leads to the following decoding (input-from-output) relation:
e = Dα · x + W · f (x)
(2.5)
The identification of the decoding equations in both models, Eq. 2.3 and Eq. 2.5, is the
key to obtain simple analytical relations between their parameters.
Equivalence of models. The analytical relation between the steady state of the Wilson-
Cowan model and the Divisive Normalization was originally proposed at the Conference
Figure 2: Saturation in the Wilson-Cowan model. Left: Illustrative saturating function
in blue and linear approximation around the origin in green. Right: Derivative of the saturating
function decreases with amplitude.
-- 5 --
celebrating the 50th anniversary of Prof. Cowan at the University of Chicago [19], but
detailed demonstration and discussion of the properties are given in [20].
In summary, just to get a simpler analytical relation between the parameters, in [20],
an assumption was done in each model: (1) a first order approximation of the nonlinear
saturation of the Wilson-Cowan system (green line in Fig. 2), and (2) a first order expansion
of the inverse in decoding the Divisive Normalization (in Eq. 2.3). In this way, it is easy to
see that the parameters of both models introduced above are related as:
b
α =
k
W = D
k ) · H · D−1
( x
b (cid:12) df
( k
dx )
(2.6)
where the (cid:12) stands for the dimension-wise Hadamard product, and the divisions are also
Hadamard quotients. This expression allows us to obtain the interaction kernel and the
attenuation of a Wilson-Cowan model which leads to a steady state compatible with the
Divisive Normalization response.
In this
The considered Wilson-Cowan model: parameters and performance.
work we take an architecture as the one considered in Fig. 1, in which different Divisive
Normalization layers were obtained through different experimental methods. For instance,
Maximum Differentiation psychophysics was used to get the 2nd and 3rd layers [40], the
1st layer was obtained by fitting human opinion in subjective image distortion [7], and the
last layer was tuned to reproduce contrast response curves [41].
In this psychophysically-tuned network, the nonlinearity in the cortical layer (the 4th
layer consisting of a linear wavelet transform followed by Divisive Normalization) was substi-
tuted by the equivalent Wilson-Cowan interaction. Following [1, 42] we assumed a Gaussian
interaction between the sensors tuned to different locations, scales and orientations, and
it makes sense to have different attenuation coefficients per subband. However, instead
of performing additional psychophysics to determine these parameters, here we took the
Divisive Normalization layer optimized in [7, 41], and we applied the relation proposed in
[19, 20], i.e. Eq. 2.6. The resulting kernel and attenuation for the Wilson-Cowan interaction
are shown in Fig. 3.
Using these parameters, simple Euler integration from an initial output given by the
input energy converges to the Divisive Normalization solution, see Fig. 4. A more detailed
discussion on the Divisive Normalization as a stable node of this Wilson-Cowan system
(theoretical stability, phase diagrams) is given in [20], together with evidences of their
perceptual equivalence beyond the mathematical equivalence.
Perceptual evidences in [20] were only focused on the visual consistence between the
original psychophysically-tuned kernel, H, and the signal-dependent kernel obtained from
Eq. 2.6, but always within the Divisive Normalization context. Specifically, it was shown
that Divisive Normalization with these two different kernels has similar contrast response
curves, and achieves similar correlation with human opinion in subjective image quality.
In this introductory section we present a new (more direct) evidence of the perceptual
plausibility of the Wilson-Cowan model derived from the psychophysically-tuned Divisive
-- 6 --
Figure 3: Parameters in the Wilson-Cowan cortical layer derived from the Divisive
Normalization cortical layer. Left: The wavelet diagram represents the auto-attenuation val-
ues, αi, for each sensor of the V1-like population. Gray values are linearly scaled to represent
attenuation values from the minimum (black, for low frequencies) to the maximum (white, for high
frequencies). Right: Illustrative interaction kernel for a specific coefficient (the one highlighted in
red). Here lighter gray represents stronger interaction between the considered sensor and the rest
of the sensors in the wavelet-like structure. Darker gray represents zero interaction. This diagram
represents the corresponding row of W arranged as a wavelet vector. In both cases, α and W , were
computed from the average response, x, in a large image database using the equivalence relation
in Eq. 2.6. The width of the Gaussian kernel in space is 0.08 deg, in scale is 1.1 octave, and in
orientation is π/6 rad, and the values of the attenuation vector are in the range [3 × 103, 1 × 105].
Figure 4: Convergence of the Wilson-Cowan model to the Divisive Normalization
solution. Left: evolution of the relative energy of the update of the solution along the integration.
Right: evolution of the relative energy of the difference between the Wilson-Cowan solution and the
Divisive Normalization response along the integration. The curves in blue are the average of the
update and difference over 35 natural images of the Van Hateren database [43], and the intervals
in red represent 3 standard deviations below and above the mean.
-- 7 --
Figure 5: Wilson-Cowan improves the description of image distortion. Left: Alignment
between subjective opinion and distance in the input representation. Right: Alignment between
subjective opinion and distance in after the proposed Wilson-Cowan interaction. Stimuli and ground
truth were taken from the subjectively rated image quality database TID [44].
Normalization. In the image quality context, here we explicitly compute the response of the
dynamic Wilson-Cowan model (integrating the equation until the steady state is achieved)
for the original and the distorted images. And then we check if the proposed Wilson-Cowan
representation is more perceptually Euclidean than the input representation (the response
after the three initial layers plus the linear wavelet transform).
In Fig. 5 we compare the correlation between the experimental subjective visibility
of distortions and the Euclidean distance between responses computed in the wavelet do-
main, r, and after the convergence of the Wilson-Cowan network, i.e. in the representation
x. In our implementation (restricted to patches subtending 0.63 degrees of visual angle),
this illustration including 68 distortions applied to one full-size image of the database, im-
plied the integration of the Wilson-Cowan equation in 8970 image regions. The result shows
that the considered Wilson-Cowan model does improve the description of the perceived dis-
tortion in naturalistic environments with regard to the previous layer of the model.
In summary, the proposed Wilson-Cowan model, with parameters obtained according
to Eq. 2.6, has the proper mathematical and perceptual behavior: (a) it converges to the
Divisive Normalization solution (Fig. 4), and (b) it describes the visibility of distortions
better than the image representation given by the linear simple cells (Fig. 5).
2.2 Material: radiometrically calibrated stimuli and color adaptation
In this work we use stimuli from the databases of Foster and Nascimento [21,
Scenes.
22] because (a) the scenes consist of radiometrically calibrated spectra in each pixel, and
(b) they also include information about the illumination through gray spheres of controlled
reflectance. While the spectral nature of the scenes is useful to simulate the perception
process from the initial integration over wavelengths [45, 46], the information about the
-- 8 --
Figure 6: Examples of Von-Kries adaptation. Top: Representative scenes from [21, 22]
under the original illumination. Spectral integration and tristimulus to digital counts rendering
was done with Colorlab [52]. Note that (in the common illumination setting of the figure) the
appearance of the reference gray is quite variable in the different scenes (different brightness, hue,
and even saturation). Right: Same scenes after Von-Kries adaptation. All the gray references are
normalized so their appearance is more stable and also the gamut of colors in the scenes.
illumination allows a straightforward implementation of Von-Kries adaptation with no extra
gray-world assumptions [47].
Alternative possibilities to perform Von-Kries adaptation in sensible LMS spaces in-
clude the Barcelona database [48], in which CIE XYZ images also include gray spheres;
and the IPL database [49, 50], in which CIE XYZ scenes were illuminated using standard
CIE D65 and CIE A spectra.
Standard chromatic adaptation transforms. We applied the Stockman and Sharpe
LMS fundamentals [51] to 5700 spectral image patches of size 40 × 40 × 33 (including 33
wavelegths in the [380,700] nm range). In this way we got 5700 tristimulus image patches
of size 40 × 40 × 3. We assumed that these patches subtend 0.625 degrees of visual angle.
This implies assuming certain observation distance and spatial sampling frequency (in our
case 64 cycles/deg), which is important to apply spatially calibrated models. Each of these
patches was associated to a specific scene, and the white point of each scene was computed
from the average LMS tristimulus values in the gray spheres of the scene (segmented by
hand). Each patch in the original LMS representation was Von-Kries normalized [47] by
the tristimulus values of the corresponding white point leading to 5700 Von-Kries-adapted
LMS images of size 40× 40× 3. Fig. 6 shows the transform of different scenes to a common
(Von-Kries adapted) canonical chromatic representation.
Information-theoretic considerations on the color manifold. The above chromatic
adaptation stage reduces the variability and simplifies the structure of the color manifold
in the LMS space because tristimulus measurements corresponding to objects seen under
different illuminations are aligned and represented in a common scale [18, 49, 53, 54]. As
a result, Von-Kries necessarily reduces the entropy of the color manifold. Additionally, the
-- 9 --
considered linear opponent transform [55] is similar to PCA [56], so it strongly reduces the
redundancy between the color components. Even though the actual nonlinearities applied
to the linear opponent spaces involve the three chromatic channels [47, 49], here we took
a simplified (dimension-wise) approximation [46]: we applied saturating nonlinearities only
depending on the average tristimulus value in the scene in each separated channel to get
the brightness and the corresponding nonlinear RG and YB values. These additional non-
linearities should not modify the redundancy shared among the linear opponent tristimulus
values because the total correlation is invariant under dimension-wise transforms [26, 57].
Despite the deep impact of these transforms on the redundancy among the compo-
nents of the spectra and the components of the colors, in this work we restrict ourselves
to the effects of the transforms on the spatial information since the Wilson-Cowan inter-
action considered above acts on spatial features. Further work may use the data available
in the supplementary material and the statistical methods reviewed below to compare the
efficiency of the (relatively simple) chromatic transforms used here with regard to more so-
phisticated color appearance models [47], or other color representations specifically designed
for information maximization or error minimization [18, 49, 58].
2.3 Method: measuring Total Correlation with Gaussianization transforms
Information transference from stimuli to response is maximized if the components of the
inner representation are statistically independent [14, 15, 59]. Therefore, the appropriate
description of the communication efficiency of a perception system consist of tracking the
amount of information shared by the different components of the signal along the neural
pathway. This redundancy (or shared information) is the total correlation [25] or multi-
information [26]. The total correlation, T , is the Kullback-Leibler divergence between the
joint PDF and the product of its marginal PDFs. Unfortunately, direct computation of T
from its definition is not straightforward because it involves multivariate PDF estimation
in spaces with a huge number of dimensions.
The problems for direct computation of T imply that the quantification of the efficiency
of image representations, x = N (r), is done using the variations ∆T from the input, r, to
the output, x [26]:
d(cid:88)
∆T = T (r) − T (x) =
(h(ri) − h(xi)) + E (log∇r N (r))
(2.7)
i=1
where the term with the sum over the dimensions is easy to compute because the marginal
entropies, h(·), only depend on univariate PDF estimations, but the expected value of the
Jacobian of the transform is required. This second term complicates the estimation and,
as a result, sometimes the analysis is restricted to linear transforms [17], where this second
term is just a constant; or surrogates of total correlation have to be used, as for instance
multiple measures of mutual information between pairs of responses [4, 5, 12], which only
involve bivariate PDF estimations.
In this work we solve the above problems by using a novel estimator of T which only
relies on univariate estimations: the Rotation-Based Iterative Gaussianization (RBIG) [24].
-- 10 --
The RBIG is a cascade of nonlinear+linear layers, each one made of (easy) marginal Gaus-
sianizations followed by an (easy) rotation. This invertible architecture is able to transform
any input PDF into a zero-mean unit-covariance multivariate Gaussian even if the chosen
rotations are random [23]. This ability to completely remove the structure of any PDF is
useful to estimate T of arbitrary vectors x: as the redundancy of a Gaussianized signal is
zero, T (x) corresponds to the sum of the individual variations, ∆Tl, that take place along
the layers of RBIG while Gaussianizing x. Interestingly, the individual variation in each
RBIG layer only depends on (easy to compute) marginal negentropies [23]:
L(cid:88)
L(cid:88)
T RBIG(x) =
∆Tl =
Jm(x(l))
(2.8)
l=1
l=1
because the marginal negentropies, Jm(x(l)), are just the sum of divergences between the
marginal PDFs of the signal that is being Gaussianized at each layer (the vectors x(l)) and
a univariate Gaussian.
In the Results section the theoretical predictions on efficiency (obtained from the ana-
lytical Jacobian of a Wilson-Cowan system plugged into Eq. 2.7) are empirically confirmed
by the computationally-convenient RBIG total correlation estimate, Eq. 2.8.
3 Results: Redundancy reduction via the Wilson-Cowan interaction
Here we study the evolution of the statistical dependence between the responses along the
psychophysically-tuned neural pathway described in Models, Materials and Methods.
First, we present an analytical result for the reduction of total correlation due to a
Wilson-Cowan interaction. We check the validity of this theoretical result with an illus-
trative reduced-scale system. This reduced-scale example is useful not only to confirm the
theory but also to illustrate the accuracy of the RBIG estimates of total correlation.
Then, we empirically analyze the behavior of the full-scale model in different ways:
(1) by analyzing the shape of the marginal PDFs for different kinds of sensors at different
layers of the network, (2) by computing the mutual information, I, between the responses
of multiple pairs of sensors at different layers, and finally, (3) by computing the (more
appropriate) total correlation, T , among the responses at the different layers.
3.1 Theoretical analysis
Here, we first present an analytical expression for the most interesting (multivariate) term
in the reduction of total correlation for the specific case of the Wilson-Cowan interaction.
Afterwards, we exhaustively check the validity of this expression in a reduced-scale exam-
ple, and we compare the efficiency of the Wilson-Cowan system with the efficiency of the
equivalent Divisive Normalization system.
3.1.1 Expression for Total Correlation in Wilson-Cowan systems.
The problem in estimating the variation of total correlation under arbitrary transforms
(Eq. 2.7) is the term depending on the Jacobian. The determinant ∇rN (r) represents
-- 11 --
the compression or expansion of the response space at the point r. Therefore, the last
term is the average variation of the volume of the response space over the PDF of natural
signals. As shown in the experiments below, this term determines the basic trends of the
redundancy reduction.
In this section we give an analytical expression for this Jacobian-dependent term for
a Wilson-Cowan interaction. We use this expression to predict how the efficiency of the
system is going to be for natural images of different luminance/contrast.
Note that in the steady state of the Wilson-Cowan system there is a straightforward
expression for the inverse (eq. 2.5). Taking into account that ∇rN (r0) = (∇xN −1(x0))−1,
∀r0, where x0 = N (r0), the critical term of Eq. 2.7 is:
E (log∇rN) = log(γ)− E(cid:16)
log(Dα + W · D df
dx
)
E (log(αixi + vi)) (3.1)
(cid:17)− (γ−1 − 1)
d(cid:88)
i=1
where D df
of the diagonal, and v = W · f (x).
dx
is a diagonal matrix with the derivative of the sigmoidal functions in each element
Eq. 3.1 is interesting because one can infer the redundancy reduction as a function
of relevant visual features such as luminance and contrast. First, note that the activity
of cortical sensors tuned to DC (flat patterns) increase with luminance and the activity of
those tuned to AC (textured patterns) increase with contrast. Then, note that the first
term in Eq. 3.1 is a constant, but the derivative in the second term decreases with contrast
(see Fig. 2), so this negative term subtracts less for bigger contrasts. On the contrary, the
last term (also negative because typically γ−1 > 1) quickly increases with luminance and
contrast, note that both components of this term increase with x.
The above considerations suggest that, for constant contribution of the marginal en-
tropies (which is the case, as shown below), the efficiency of a Wilson-Cowan network should
be bigger in the low-luminance / low-contrast region of the image space. This trend sug-
gested by the analysis of the terms in Eq. 3.1 is interesting because natural images typically
have low luminance and low contrast [16, 28, 60, 61].
3.1.2 Experiment in a reduced-scale system.
Here we consider a simplified example with the basic elements of the considered network
but in a reduced-scale scenario: 3-pixel images. In this situation the Jacobian matrices are
small so the theory can be visualized and systematically checked. The structure of this
reduced-scale perception system is as follows: first, 3-pixel luminance images are trans-
formed into brightness through a pixel-wise (dimension-wise) Weber-like saturation [47].
Then, 3 Fourier-like analyzers extract the DC, the low-frequency, and the high-frequency
components of the brightness simulating frequency-selective cortical filters [62]. This fre-
quency representation is weighted by a low-pass transfer function that simulates contrast
sensitivity [63]. Finally, this set of responses interact either according to a static Divisive
Normalization transform, or dynamically through the equivalent Wilson-Cowan equation.
The specific parameters of this reduced-scale system are given in Appendix A.
Figure 7 shows how this system transforms the manifold of 3-pixel luminance images of
the considered database. First, note how the frequency sensors (second scatter plot) look
-- 12 --
Figure 7: Natural images through the reduced-scale system. The scatter plots (2 × 103
samples randomly chosen from a set of 5 × 106 samples used in the experiments) display the
changes on the PDF of the responses of the system to 3-pixel natural images at different layers
of the network. From left to right: (1) linear responses to luminance at the retina, (2) responses
of frequency analyzers applied to brightness (saturated luminance), and (3) two versions of the
nonlinear interaction at the cortex: the Divisive Normalization and the Wilson-Cowan responses.
Absolute luminance (in cd/m2) has been normalized by the 95th percentile luminance value.
for the axes of symmetry of the manifold of natural images in the input domain, similarly
to PCA [64]. However, the luminance-to-brightness saturating transform (scatter plot not
shown) expands the low-luminance region and compresses the high-luminance region so the
tail of brighter stimuli (high values in r2
1 in the second scatter plot) is shorter. This will be
discussed in further detail in analyzing the behavior of the full-scale model. Finally, note
how the two interaction schemes considered here lead to a sort of factorization of the PDF
in four separate blobs in the AC frequency components. This effect is also discussed in the
next section with the marginal PDFs of the full-scale model in the wavelet domain.
The deformation of the manifolds shown in Fig. 7 is an interesting illustration of the
multivariate equalization/factorization effect of the model. This suggests the model(s)
are actually operating under an information maximization goal. However, this qualitative
intuition has to be quantified. To do so, we computed the average luminance and the RMSE
contrast of 5× 106 three-dimensional samples extracted from the considered natural scenes,
and we estimated their distribution in the luminance/contrast plane. Then, these images
were injected through the model: on the one hand the static Divisive Normalization response
was computed, and, on the other hand, the equivalent Wilson-Cowan system was stimulated
with the linear frequency representation of the brightness of each sample until it converged
to a steady state (after 500 Euler integration steps). Then, we computed the reduction in
total correlation for the stimuli at different locations of the luminance/contrast plane. We
did this in two ways: (1) theoretically, through the proposed analytical expression, Eq. 3.1,
and (2) empirically, by using the RBIG method reviewed above, Eq. 2.8. In each case, we
considered up to 2.5 × 104 samples per location in the luminance/contrast plane to do the
estimations.
Fig. 8 shows the elements of the theoretical computation: the final result comes from
(i) a Jacobian-dependent term (Eq. 3.1 in the Wilson-Cowan case, and Eqs. 42 and 43 in
-- 13 --
Figure 8: Theoretical efficiency of Wilson-Cowan and Divisive Normalization. The
theoretical reduction of total correlation for natural images of different luminance/contrast (surfaces
at the left) is computed from a term that depends on the Jacobian of the transform (surfaces at
the center), and a term that depends on the sum of marginal entropies of the input and output
(surfaces at the right). The dark surfaces at the left are the (small) standard deviations of the
estimates over 10 different realizations.
[7] in the case of Divisive Normalization), and (ii) and a term that depends on the marginal
entropies. As discussed above, the Jacobian-dependent term decreases with luminance and
contrast, while the variation of the entropy-dependent term is smaller across the image
space. Therefore, the Jacobian determines the general behavior of the system.
Results of the theoretical and the empirical descriptions of the efficiency are compared
in Fig. 9 together with the estimated PDF in the luminance/contrast plane. The general
trend of the efficiency surfaces for Divisive Normalization and for the Wilson-Cowan model
are the same (with both estimation approaches): first, the redundancy is reduced all over
the image space (the reduction of total correlation is positive almost everywhere), but more
importantly, the efficiency is clearly bigger for the low-luminance / low-contrast region,
-- 14 --
Figure 9: Efficiency is consistent with natural image statistics. The general trend of the
efficiency surfaces in the luminance/contrast plane for Divisive Normalization and Wilson-Cowan
are the same. Moreover, both are consistent with natural image statistics since efficiency is bigger in
the most populated region of the image space. The dark surfaces are the (small) standard deviations
of the estimates over 10 different realizations.
which also is the most populated region (see the PDF of the same calibrated training set
at the left).
These results confirm the accuracy of the empirical estimation of T and the correctness
of the theoretical expression for the Wilson-Cowan nonlinearity (or more precisely, they
show the consistency between the theoretical and the empirical estimation). Moreover, re-
sults show the equivalent behavior of Divisive Normalization and Wilson-Cowan in coding
efficiency: both models focus on the same region of the image space. And, more interest-
ingly, they show that psychophysically-inspired nonlinearities have their peak performance
in the proper region of the image space even though they were not optimized in any way
to that end.
-- 15 --
3.2 Empirical analysis of the full-scale model
In this section we study the efficiency of the full-scale model using a purely empirical ap-
proach. This is because the size of the Jacobian for each sample is huge and it is difficult
to collect a big enough set for a reliable estimation of Eq. 3.1. For instance, in the cur-
rent implementation of the model (using a 3-scale, 4-orientation, steerable transform [65]),
discrete luminance patches of size 40 × 40 turn into response vectors x ∈ R10025, so the
Jacobian matrices are 10025 × 10025. One could restrict the interest to a selected set of
sensors (as done in some empirical analysis below). However, note that this restriction does
not imply a reduced-size Jacobian in the theoretical expression. No rows/columns in the
Jacobian can be neglected since the evolution of a subset of responses in the Wilson-Cowan
integration depends on all the responses because, in principle, the matrix W is dense. And
this dense nature makes perceptual sense because (stronger or weaker) there is interaction
between all scales/orientations/positions [37, 38].
Here we analyze the behavior of the full-scale model in three different empirical ways:
(1) by discussing the shape of the marginal PDFs for the responses of sensors at different
layers of the network, (2) by computing the mutual information, I, between the responses
of multiple pairs of sensors at different layers (as in [5]), and finally, and more interestingly,
(3) by computing the total correlation, T , among the responses at the different layers. In the
latter, different spatial sampling schemes are explored to capture how visual information
depends on the field of view.
In all the experiments in this section we start from the responses of the considered
model to the 5700 spectral image patches described in the Materials section. In each case,
the responses at the different layers are further subsampled according to the goal of the
specific experiment and to get a representative set for the considered estimation.
3.2.1 Marginal equalization
Descriptions of natural image statistics usually start from the marginal PDFs because they
show the basic complexity of the signal (e.g. naive coding would only consider zero-order
entropy [57]). Marginal PDFs are relevant because once 2nd order correlation has been
removed, marginal non-Gaussianity (or sparsity) is an appropriate description of the total
correlation in certain cases [66]. It is also interesting to see how a perception system modifies
the marginal PDFs (e.g. eventual equalization, increased sparsity, Gaussianization [16, 28,
60, 67, 68]) because these changes may reveal an information maximization goal [67, 68] and
the general multivariate factorization goal [16] may be achieved through more complicated
marginal PDFs [5].
The initial layers of the system we are considering (1st to 3rd layer) consist of sensors
tuned to specific spatial locations, while the 4th layer is formed by sensors with wavelet-
like receptive fields.
In the marginal approach considered here we assume the signal is
stationary across space and orientation. Therefore we pool together samples from different
spatial locations and (in the case of wavelets) also corresponding to different orientations.
In all the results in this section we collected 1 × 107 randomly chosen samples to estimate
each marginal PDF.
-- 16 --
Figure 10: Marginal PDFs of responses of sensors tuned to spatial locations. Left: PDFs
for the original luminance, Von-Kries luminance, and Brightness. Center: Adaptive luminance to
brightness transform. Right: PDF of nonlinear contrast. See [7] for details on specific Divisive
Normalization expressions of the transforms.
Figure 10 shows the effect of the layers of sensors tuned to spatial locations over the
original PDF of luminance at the retina. We see that the linear luminance representation
(blue line in Fig. 10-left) is strongly biased and has several peaks in the high-luminance tail.
These different peaks correspond to high reflectance objects seen in different illumination
conditions. Note how these peaks disappear after the Von-Kries normalization that puts
every scene in the same relative-luminance range. Then, the brightness transform that
depends on the background luminance (Fig. 10-center) tends to expand the low-luminance
range so that the resulting marginal PDF is relatively more flat.
All the above transforms make no use of spatial or contextual information (appart from
the global scaling factors in Von-Kries obtained from average illumination). The variation
of the PDFs in these point-wise layers is consistent with equalization goals [67]. This com-
pletely changes in the 2nd and 3rd layers that compute contrast from local normalizations of
brightness and apply linear center-surround receptive fields whose response is subsequently
normalized by the local activity.
After contrast computation the mean is removed (see that the peak in Fig. 10-right is
in zero). The distribution is not symmetric around zero because darker regions (now below
zero) are more frequent in natural images.
Interestingly, the contrast masking through
Divisive Normalization in the third layer generates a bimodal distribution above and below
zero. This effect in Divisive Normalization has been interpreted as predictive coding, where
the numerator is predicted from the neighbors in the denominator leading to peaks above
and below zero where this prediction is successful. The predictive coding interpretation has
been suggested many times [12, 27, 33, 69], and peaks of this kind have been consistently
found in [5, 7] using different models and scenes.
Figure 11 shows the marginal PDFs of the responses of sensors tuned to oriented features
-- 17 --
Figure 11: Marginal PDFs of responses of sensors tuned to local oriented fea-
tures.Left: PDFs of the responses to distributions of nonlinear contrast of V1-like linear sensors.
Center: PDFs after Divisive Normalization. Right: PDFs after Wilson-Cowan interaction.
of different scales. These sensors are the linear wavelet-like filters at the 4th layer that
simulate V1 simple cells, and the corresponding mechanisms after the nonlinear interactions
(either Divisive Normalization or Wilson-Cowan). The linear wavelet-like sensors at the
4th layer display heavy tailed PDFs with decreasing variance for finer scales. However the
sparsity of these responses is lower than the sparsity of the same filterbank applied on
luminance images (results not shown). This lower sparsity maybe because wavelet filters
certainly lead to highly sparse (non-Gaussian) responses when applied to luminance images,
but the nonlinear-contrast images are substantially different from luminance images (see
for example the output of 3rd layer at Fig. 1).
More interestingly, these (moderately) sparse response distributions turn into bimodal
distributions after the nonlinear interactions, both for Divisive Normalization and Wilson-
Cowan. Again (as in Fig. 10-right) the masking interaction generates bimodal PDFs. How-
ever, this is not a exclusive feature of Divisive Normalization:
it happens in equivalent
Wilson-Cowan systems as well. Note that this marginal behavior (two modes around a
depression in zero) can be understood in multivariate terms using the visualization of the
reduced-scale example in Fig. 7: the projection of the blobs on the different axes leads to
the peaks in the marginals.
3.2.2 Mutual Information along the network
Criticisms to linear Independent Component Analysis (ICA) pointed out that sparsity max-
imization in the marginal PDFs does not guarantee complete statistical independence [16,
70]. This is obvious from the mathematical point of view [66], however, in practice, a
substantial amount of work was devoted to point out the existence of residual statistical
relations after ICA-like filters had been applied to images. The problems for direct estima-
tion of total correlation mentioned in the Methods section, impled a variety of surrogates to
measure this remaining redundancy, as for instance the analysis of conditional probabilities
-- 18 --
Figure 12: Mutual information between a spatial sensor and its neighbors. Different
layers along the network are considered from left to right. The value of auto-mutual information
has been set to zero for better visualization. We did that because this value is arbitrarily large
(every response contains arbitrarily large amount of information about itself), and it should not be
considered in describing the interaction with the neighbors.
of neighbor responses, the so-called bow-ties [3 -- 5, 27 -- 29], the analysis of the correlation
between the energies of pairs of responses as a way to identify subspaces with residual rela-
tions [30 -- 34], or the measure of mutual information between pairs of responses [5, 12, 35].
Here we quantify the relation between pairs of responses using mutual information, I,
as a function of the separation between the sensors in the corresponding feature space
(spatial departure for the first three layers and departure in space, scale and orientation
for the wavelet-like layer). We applied the straightforward definition of mutual information
because each computation reduces to the estimation of a joint (bivariate) PDF and two
marginal PDFs. We did that by gathering neighbors from sliding windows of size smaller
than the available visual field. In that way we got 8 × 104 samples for each pair of sensors.
We computed 10 estimations of I using a randomly chosen subset with 80% of those samples
in each estimation. We chose the bin size according to the Silverman rule of thumb [71].
The results in the plots are the average of those estimations.
Figure 12 shows the evolution of the interactions as the signal goes through the three
initial layers. In every case, the relation is higher with closer neighbors and decays with
distance. However, in the explored spatial range (which is limited, about 0.5 degrees) these
pair-wise relations do not drop to zero in layers before contrast computation. The I would
eventually arrive to zero for sufficiently large separation, but we didn't have access to big
enough distances. Von-Kries adaptation and brightness computation reduce the amount of
paired relations about 0.5 bits, but the substantial change again comes with the introduction
of spatial processing in the third layer: the relations with the closer neighbors are smaller,
they drop faster, and actually arrive to zero for small departures. The division by the local
activity removes the relation with the spatial neighborhood:
it seems that the predictive
coding interpretation of normalization mentioned above [12, 27, 33, 69] also applies here.
Figure 13 shows the mutual information values for the responses of two sensors (those
highlighted in red) with all their neighbors in the response vector at the V1 layer before
-- 19 --
Figure 13: Mutual information between two wavelet-like sensors and all their neigh-
bors. Left: interactions of a sensor tuned to vertical patterns of middle frequency (highlighted in
red). Right: interactions of a sensor tuned to diagonal patterns of middle frequency (highlighted in
red). Top: interactions between the linear sensors. Top: interactions between the sensors after the
Wilson-Cowan recurrence. Lighter gray indicate higher mutual information. All figures are scaled
in the same way. Comparing top and bottom one can see that the Wilson-Cowan recurrence reduces
the statistical relations a little bit (darker diagrams at the bottom).
and after the application of the Wilson-Cowan interaction. Finally, Fig. 14 shows a subset
of the results shown in Fig. 13 represented as a surface as in Fig. 12. As previously reported
for linear wavelet domains [5, 35, 72], here mutual information also decays with distance
in space scale and orientation. After the nonlinear interactions (we only show the Wilson-
Cowan result) the relations between the coefficients seem to be reduced.
However, the reduction obtained here is substantially smaller than the one obtained
with previous Divisive Normalization models where I was computed in the same way [5].
It is not a specific matter of Wilson-Cowan, but it is related to the multi-layer architecture:
results with the current architecture with Divisive Normalization (not shown) are similar
-- 20 --
Figure 14: Mutual information between a wavelet-like sensor and its neighbors in
space and orientation. The considered sensor is at the center of the domains represented at the
left column. Left: Decay of the interaction in space (for the same orientation -or subband-). Right:
Decay of the relation in space for a different orientation. Top: result before the Wilson-Cowan
recurrence. Bottom: result after the Wilson-Cowan recurrence.
to the Wilson-Cowan results in Figs. 13 and 14, i.e. only moderate gain in I. It is not
the database either: we made preliminary experiments with the current architecture and
Divisive Normalization using the Van Hateren stimuli [43] (the data used in [5]), and the
results (not shown) are similar: also a very moderate gain.
The difference in the architecture explored here with regard to the one in [5] is that
the current one is deeper: the current wavelet-like layer comes after previous layers that
already are doing spatially significant Divisive Normalization (in our case the 3rd layer).
The presence of previous layers which already reduce the redundancy, may limit the ability
of the considered nonlinearities at this later stage. Nevertheless, note that the set of mutual
information measures is only a surrogate of the conceptually appropriate measure, the total
correlation, which is addressed in the next section.
3.2.3 Total Correlation along the network
In this section we measure the shared information (total correlation) in the responses of
sensors covering progressively bigger portions of the visual field, and we see how this shared
information modifies along the neural pathway. We take two different spatial sampling
strategies: (1) no-subsampling, i.e. considering all the sensors tuned to the spatial region,
so bigger regions imply a bigger number of sensors, and (2) subsampling, i.e. taking a fixed
number of sensors progressively more separated, so the bigger the visual field, the poorer
the sampling of the visual patterns.
An illustration of the dimensions of the response vector considered at the wavelet-like
layer in both experiements is shown in Fig. 15. The corresponding selection for the layers
-- 21 --
where sensors have spatial meaning is straightforward.
Taking such subsets of responses (or sensors) keeps the dimension small and makes
the empirical estimation easier. This specific sampling schemes were proposed to allow
gathering response vectors by sliding windows like those in Fig. 15. Collecting big datasets
is particularly important since the dimensions of the considered vectors range from 5 (in
the subsampling experiment in the layers with spatial meaning) up to 321 (in the no-
subsampling experiment with bigger field of view in the wavelet-like layers). In every case,
the total correlation is reported by coefficient. All the total correlation measures in this
section were computed using 2.5 × 104 samples and RBIG with 600 layers to ensure the
convergence to the Gaussian. In each case, 10 realizations of the estimates were computed
using 80% of the samples.
The sampling strategies were selected because they lead to a clear pattern in the total
correlation in the input representation. This pattern is a safety check in this full-scale
situation where the theoretical computation is not feasible due to the size of the Jacobians.
Note that given the smoothness of objects in natural scenes, the consideration of bigger
visual fields in the no-subsampling experiment should include more coherent structures
(e.g. bigger portions of objects in the images). As a result, one would expect increased
redundancy given the bigger coherence of the visual structures. On the other hand, if a
fixed number of sensors is progressively separated covering bigger visual fields, as in the
subsampling experiment, the statistical relation between the sensors in the extremes of the
visual field should decrease with separation.
This is exactly the pattern that is found in the results, see Fig. 16: (1) in luminance
Figure 15: Wavelet sensors in total correlation experiments. Visual angle and spatial
sampling strategy. From left to right:
increasing visual angle covered by the sensors. Top and
bottom: no sub-sampling, versus sparse sampling.
-- 22 --
images in the no-subsampling experiment, total correlation increases with the size of the
visual field, and (2) in luminance images in the subsampling experiment, total correlation
decreases with the size of the visual field. Note also that the stimuli that can be properly
called images are those of the no-subsampling experiment. In the subsampling experiment
we just take samples at the corners and at the center of the visual field.
Regarding the efficiency of the signal representations, Fig. 16-left shows how the total
correlation is reduced for images (note that the signals in the subsampling experiment are
not strictly images). Results show that the initial layers (Von-Kries chromatic adaptation,
and Weber-like adaptive brightness) do not contribute to redundancy reduction in images.
This is because these operations introduce no fundamental spatial information, so the re-
dundancy between neighbor spatial locations is basically preserved. Redundancy starts to
be removed in the third layer (Nonlinear Contrast) where the Contrast Sensitivity Function
is applied to the local contrast and then the result is nonlinearly transformed using Divisive
Normalization. In this layer, bigger reductions are obtained in smaller visual fields due to
the (relatively small) size of the masking kernel of Divisive Normalization in the spatial
domain (about 0.02 degrees). On the contrary, the bigger the visual field (which involves
more complicated visual structures), the more effective is the linear filterbank of wavelet-
like receptive fields (Linear V1 layer). Finally, the nonlinear interaction at the cortical level
(after all the previous stages) does not imply substantial reductions of redundancy in terms
of total correlation. The values at the Nonlinear V1 layer in Fig. 16 display the specific
result for the Wilson-Cowan interaction, but Table 1 shows that the removed redundancy
for the equivalent Divisive Normalization model is virtually the same: in the biggest visual
Figure 16: Reduction of Total Correlation along the layers of the network. Visual
signals of different spatial size are considered. Left: no spatial subsampling (actual images). Right:
with spatial subsampling (samples at the corners and the center of the visual field).
-- 23 --
Table 1: Removed Redundancy (in %). Percentage of Total Correlation removed at the inner
representation with regard to the values at the input representation. Visual fields of different size
(in degrees) with no subsampling (left) and with subsampling (right).
NO-SUBSAMPL.
Divisive Norm.
Wilson-Cowan
∆x = 0.24
84 ± 3
85 ± 4
∆x = 0.12
67 ± 3
67 ± 3
∆x = 0.06 ∆x = 0.44
54 ± 2
54 ± 1
87 ± 2
86 ± 3
SUBSAMPL.
∆x = 0.32
84 ± 4
84 ± 3
∆x = 0.18
79 ± 3
79 ± 3
field explored for images the inner representation removes about 85% of the total correlation
in the input representation.
Fig. 16-right shows the equivalent results for visual fields in which spatial samples
were taken from the corners and the center of the considered regions. In this configuration
the inner representation also removes about 85% of the total correlation at the input.
However, in this case, most of the redundancy reduction comes from the nonlinear contrast
computation.
Similarly to what was found in the mutual information analysis, in this cascade (for
images) the biggest reduction in redundancy comes from the wavelet-like transform and the
contribution of the nonlinear interaction between simple cells is relatively small.
4 Discussion
In
Alternative confirmation of Barlow's hypothesis for Wilson-Cowan systems.
this work we performed the first analysis of the communication efficiency of Wilson-Cowan
networks in accurate information theoretic terms using the appropriate (multivariate) de-
scription of redundancy: the total correlation. Pointing out the efficiency of a psychophys-
ically tuned system is an alternative confirmation of Barlow's Efficient Coding Hypothesis.
This hypothesis states that natural perception systems evolved to efficiently encode
natural signals [8 -- 10]. The conventional way to check this hypothesis is training artificial
systems with natural scenes assuming some communication efficiency goal (as for instance
redundancy reduction), and see the emergence of natural properties from the statistical
learning. This logic (from-statistics-to-perception) has been applied to explain the linear
image representation in V1 [32, 50, 73], the nonlinear interactions [3, 4, 49, 58], and even
the origin of visual illusions [18, 74].
However, note that here we reasoned in the opposite direction: we took a psychophys-
ically plausible network and analyzed its behavior in information-theoretic terms after no
statistical learning whatsoever. We found that the considered systems have bigger efficiency
(bigger reductions in T ) in the most populated regions of the image space, see Fig. 9, and
they remove about 85% of the total correlation of natural scenes in the input representa-
tion, see Fig. 16 and Table 1. The approach used here (from-perception-to-statistics) is an
alternative way to confirm the Efficient Coding Hypothesis, as suggested in [5].
Efficiency of Wilson-Cowan is similar to Divisive Normalization. Several evi-
dences suggest that the Wilson-Cowan interaction can be as effective in information theo-
-- 24 --
retic terms as the Divisive Normalization model. First, consider that the interaction kernel
W derived here from the relation proposed in [19, 20], roughly decays with space, orienta-
tion, and scale. The similarity between the interaction kernel and the mutual information
in the linear wavelet-like representation (compare Figs. 3-right and 13-left) has been related
to the effectiveness of this kind of interactions to capture the structure of the signal [5, 72].
Second, recurrent subtraction of the activity of the neighbors weighted in this way in the
integration of the Wilson-Cowan equation lead to similar bimodal marginal PDFs, and to
similar clusters in the joint PDF as the Divisive normalization (see Figs. 11 and 7). Third,
the Jacobian of the Wilson-Cowan nonlinearity induces similar trends in the reduction of
total correlation as the Divisive Normalization (see Figs. 8 and 9): both models have better
efficiency in the low-luminance / low-contrast region. And finally, note that the redundancy
reduction in the full-scale model is roughly the same using the Wilson-Cowan interaction
and the Divisive Normalization (see Table 1).
The above statistical similarities between the Wilson-Cowan and the Divisive Normal-
ization models are consistent with the ability of Wilson-Cowan interaction to improve the
explanation of pattern visibility, which is similar to Divisive Normalization. Note that the
results in Fig. 5 with Wilson-Cowan are consistent with better explanations of subjective
image quality reported for Divisive Normalization [7, 75, 76]. These information theoretic
and perceptual similarities suggest the Divisive Normalization actually is the steady state
of the Wilson-Cowan dynamics, as assumed in [20].
Accuracy of total correlation estimates. A side technical benefit of the theory pre-
sented here (the expression of the reduction of total correlation for Wilson-Cowan systems,
Eqs. 2.7 and 3.1) is that this kind of theoretical results can be used to check the accuracy
of empirical estimates of total correlation in neuroscience applications. This is also true for
the equivalent result for Divisive Normalization presented in [7]. Here we directly compare
these theoretical results with empirical estimates (in Fig. 9), which was not done in [7].
These ground truth results are relevant because information measures using Gaussian-
ization [23, 24] had been used to identify interesting patterns in other spatio-spectral and
spatio-temporal data [77, 78], but in those cases there was no theoretical bound to compare
with.
The success of the empirical measure of total correlation [23, 24] shown in this work
suggests that it can be used as a goal function or as an analysis tool to optimize artificial
systems, or to check natural systems according to the information bottleneck principle [79].
Differences with shallower networks. Despite here we quantify redundancy through a
more appropriate concept, the analysis presented in this work is similar to the one done in [5]
for a shallower network based on another phychophysically tuned Divisive Normalization.
The model consisted of a single linear+nonlinear layer: a wavelet transform plus a linear
weighting to simulate the contrast sensitivity followed by a Divisive Normalization of the
energies of the weighted wavelet.
However, as noted in section 3.2, the results obtained here differ from those reported
in [5]: the reduction of the mutual information with regard to the linear wavelet represen-
tation obtained here is substantially smaller. This is consistent with the small reduction in
-- 25 --
total correlation with regard to the linear wavelet representation shown in Fig. 16, which
was not measured in [5] because the statistical tool was not available. This could mean
that the efficiency of this interaction (either implemented through Wilson-Cowan or Divisive
Normalization) is not as big as suggested in [5].
On the contrary, we argue that the small gain with regard to the linear wavelet is not
a fundamental limitation of the Divisive Normalization or the Wilson-Cowan models, but
a simple consequence of the fact that we are looking at a deeper layer: the presence of
previous layers which already reduce the redundancy limits the ability of the considered
nonlinearities a later stage. In fact, the reduced sparsity displayed by our linear wavelet
representation in Fig. 11-left has the same origin: the modified effect of wavelets on contrast
images (instead of the conventional luminance images).
Reproducible results. Code to reproduce all the empirical results on the behavior of
the full-scale model as well as the experiment of the reduced-scale model that includes the
theoretical result are available at https://github.com/alviur/information_wc.git
In summary, the considered psychophysically tuned 4-layer network re-
Final remarks.
moves 85% of the total correlation from achromatic (luminance) images subtending 0.24 de-
grees. For this field of view the Wilson-Cowan model has the same communication efficiency
as the equivalent Divisive Normalization model. In a reduced-scale scenario, the theoretical
and the empirical results on redundancy reduction are consistent, and both show that the
psychophysically inspired Wilson-Cowan and Divisive Normalization networks are more ef-
ficient in the regions of the image space where natural scenes are more frequent. The above
results represent a confirmation of Barlow's Efficient Coding Hypothesis for Wilson-Cowan
models in the perception-to-statistics direction (alternative to the conventional statistics-
to-perception approach).
The similarities in the statistical effect of the Wilson-Cowan and the Divisive Normal-
ization suggest that recurrent neural field models could be an alternative in image coding
applications where Divisive Normalization has beaten both JPEG and JPEG2000 [12, 13].
From a more fundamental point of view, future work should use the proposed analysis
before the stationary state is reached. This will show how the information efficiency may
evolve while adaptation takes place.
A Appendix: Psychophysically inspired reduced-scale model
Overview. The reduced-scale model consist of two linear+nonlinear layers: (1) a linear
radiance-to-luminance transform using a standard Spectral Sensitivity Function, Vλ, in the
spectral integration [45], followed by a simple exponential for the luminance-to-brightness
nonliniearity applied pixel-wise in the spatial domain, that simulates the Weber-Fechner
response to luminance [47], and (2) a linear+nonlinear layer in which the linear transform
is a discrete cosine transform (a orthonormal rotation) followed by a low-pass weighting
function that simulate frequency-tuned sensors and the Contrast Sensitivity Function (CSF)
[63]. Then, the outputs of the frequency sensors undergo a nonlinear interaction that may be
-- 26 --
a Divisive Normalization [2, 7, 41], or its equivalent Wilson-Cowan network, with parameters
computed according to Eq. 2.6 [20].
L(1)
N (1)
L(2)
N (2)
x0
r1
x1
r2
x2
(A.1)
Transform. The actual inputs of our code are the responses of the linear photoreceptors:
3-pixel image vectors with normalized luminance values, i.e. r1 ∈ R3. The normalized lumi-
nance was computing dividing the absolute luminance in cd/m2 by the value corresponding
to the 95% percentile of the luminance, in our case 260 cd/m2.
• The luminance-to-brightness transform, N (1), is just:
x1 = (r1)γ
where γ = 0.6
• The linear transform of frequency-tuned sensors with CSF gain, L(2), is:
r2 = GCSF · F · x1
where
3
(cid:113) 1
(cid:113) 1
−(cid:113) 1
1
2
6
0
0
3
(cid:113) 1
(cid:113) 1
0 −(cid:113) 1
(cid:113) 2
3 −(cid:113) 1
6
3
2
0
0.5
0
0
0
0.3
F =
GCSF =
(A.2)
(A.3)
• The Divisive Normalization of the frequency-tuned sensors, N (2)
(b+H·r2γ ) · r2γ where γ = 0.7,
x2 = sign(x2) (cid:12) Dk · D−1
DN, is:
and,
(A.4)
0.18
0
0
Dk =
H = Dl · W · Dr =
0.03
0
0
0
0
0.01
0.06
0
0
0
0.35
0
0
0
0.27
and the vector of semisaturations, b, is:
b =
·
0.95
0
0
0.06
0.93
0.02
0.01
0.05
0.98
0
0.27
0
0
0
0.13
0.04
·
0.93
0.08
0
0.03
0.01
-- 27 --
$
$
%
%
$
$
%
%
• The equivalent Wilson-Cowan interaction, N (2)
WC, is defined by the differential equation
2.4, where the auto-attenuation, α, and the interaction matrix, W , are:
α =
W =
and the saturation function is:
0.06
0.93
0.02
0.01
0.05
0.98
(A.5)
(A.6)
0.41
0.93
1.10
1.30
0.04
0
1.12
0.02
0.01
x =
(A.7)
the scaling constant is, c = x1−γ, and x is the average response over natural images
(for the Divisive Normalization transform):
f (x) = c xγ where γ = 0.4,
and,
This exponent is also used for the definition of energy in Wilson-Cowan, e = rγ.
Note that the interaction neighborhoods have unit volume, (cid:80)
j Wij = 1 ∀j, as sug-
gested in [38], and then, the Divisive Normalization kernel is given by the product of this
unit-volume neighborhood and two left and right filters in the diagonal matrices, Dl and
Dr [41]. The values for the semisaturation, b, and the diagonal matrices Dl and Dr were
inspired by the contrast response results in [41]: we set the semisaturation according to
the average response of natural images (low-pass in nature), and we initialized the left and
right filters to high-pass. However, afterwards, in order to make NDN and NWC consistent,
we applied the Divisive Normalization over natural images and we iteratively updated the
values of the right and left filters according to Eq. 2.6. In the end, we arrived to the values
in the above expressions (where the filter at the left is high-pass, but the filter at the right
is not). Note that the attenuation in Wilson-Cowan is computed using Eq. 2.6.
Jacobian. The information theoretic computations strongly depend on how the system
(locally) deforms the signal representation (e.g. Eq. 3.1). This is described by the Jacobian
of the transform with regard to the signal, ∇r1S = ∇r2N (2) · ∇x1L(2) · ∇r1N (1). In this
reduced-scale model, this Jacobian (for the Wilson-Cowan case) is:
(cid:16)Dα + W · D df
(cid:17)−1 ·(cid:0)D(Dα·x2+W·f (x2))
(cid:1)1− 1
dx
∇r1S = γ2
γ2 · GCSF · F · D
γ1(r1)γ1−1
(A.8)
Acknowledgments
This work was partially funded by the Spanish Government and EU FEDER fund through
the MINECO grants TIN2015-71537-P and DPI2017-89867-C2-2-R; and by the European
Union's Horizon 2020 research and innovation programme under grant agreement number
761544 (project HDR4EU) and under grant agreement number 780470 (project SAUCE).
-- 28 --
References
[1] H. R. Wilson and J. D. Cowan, A mathematical theory of the functional dynamics of cortical
and thalamic nervous tissue, Kybernetik 13 (1973), no. 2 55 -- 80.
[2] M. Carandini and D. Heeger, Summation and division by neurons in visual cortex, Science
264 (1994), no. 5163 1333 -- 6.
[3] O. Schwartz and E. Simoncelli, Natural signal statistics and sensory gain control, Nature
Neurosci. 4 (2001), no. 8 819 -- 825.
[4] J. Malo and J. Gutiérrez, V1 non-linear properties emerge from local-to-global non-linear
ICA, Network: Computation in Neural Systems 17 (2006), no. 1 85 -- 102.
[5] J. Malo and V. Laparra, Psychophysically tuned divisive normalization approximately
factorizes the pdf of natural images, Neural computation 22 (2010), no. 12 3179 -- 3206.
[6] R. Coen-Cagli, P. Dayan, and O. Schwartz, Cortical surround interactions and perceptual
salience via natural scene statistics, PLOS Computational Biology 8 (03, 2012) 1 -- 18.
[7] M. Martinez-Garcia, P. Cyriac, T. Batard, M. Bertalmío, and J. Malo, Derivatives and
inverse of cascaded linear+nonlinear neural models, PLOS ONE 13 (10, 2018) 1 -- 49.
[8] H. Barlow, Sensory mechanisms, the reduction of redundancy, and intelligence, Proc. of the
Nat. Phys. Lab. Symposium on the Mechanization of Thought Process (1959), no. 10 535 -- 539.
[9] H. Barlow, Possible principles underlying the transformation of sensory messages, in Sensory
Communication (W. Rosenblith, ed.), pp. 217 -- 234. MIT Press, Cambridge, MA, 1961.
[10] H. Barlow, Redundancy reduction revisited, Network: Comp. Neur. Syst. 12 (2001), no. 3
241 -- 253.
[11] K. Friston, The free-energy principle: a rough guide to the brain?, Trends in cognitive
sciences 13 (2009), no. 7 293 -- 301.
[12] J. Malo, I. Epifanio, R. Navarro, and E. P. Simoncelli, Nonlinear image representation for
efficient perceptual coding, IEEE Transactions on Image Processing 15 (2006), no. 1 68 -- 80.
[13] J. Ballé, V. Laparra, and E. P. Simoncelli, End-to-end optimized image compression, in 5th
Int. Conf. Learn. Repres., ICLR 2017, 2017.
[14] A. J. Bell and T. J. Sejnowski, An information maximization approach to blind separation
and blind deconvolution., Neural Computation 7 (1995) 1129 -- 1159.
[15] A. J. Bell and T. J. Sejnowski, The independent components of natural scenes are edge
filters, Vision Research 37 (1997), no. 23 3327 -- 3338.
[16] E. P. Simoncelli and B. A. Olshausen, Natural image statistics and neural representation,
Annual Review of Neuroscience 24 (2001), no. 1 1193 -- 1216.
[17] M. Bethge, Factorial coding of natural images: how effective are linear models in removing
higher-order dependencies?, JOSA A 23 (2006), no. 6 1253 -- 1268.
[18] V. Laparra and J. Malo, Visual aftereffects and sensory nonlinearities from a single
statistical framework, Frontiers in Human Neuroscience 9 (2015) 557.
[19] J. Malo and M. Bertalmio, Appropriate kernels for divisive normalization explained by
Wilson-Cowan equations, Celebration of Cowan's 50th anniv. at Univ. Chicago. ArXiv:
Quant. Biol. https://arxiv.org/abs/1804.05964 (2018).
-- 29 --
[20] J. Malo, J. Esteve-Taboada, and M. Bertalmío, Divisive normalization from Wilson-Cowan
dynamics, ArXiv: Quant. Biol. https://arxiv.org/abs/1906.08246 (2019).
[21] D. H. Foster, K. Amano, and S. M. Nascimento, Time-lapse ratios of cone excitations in
natural scenes, Vision research 120 (2016) 45 -- 60.
[22] S. M. Nascimento, K. Amano, and D. H. Foster, Spatial distributions of local illumination
color in natural scenes, Vision Research 120 (2016) 39 -- 44.
[23] V. Laparra, G. Camps-Valls, and J. Malo, Iterative gaussianization: from ICA to random
rotations, IEEE Trans. Neural Networks 22 (2011), no. 4 537 -- 549.
[24] J. Johnson, V. Laparra, R. Santos, G. Camps, and J. Malo, Information theory in density
destructors, in 7th Int. Conf. Learn. Repres., ICLR 2019, Workshop on Invertible
Normalization Flows, 2019.
[25] S. Watanabe, Information theoretical analysis of multivariate correlation, IBM Journal of
research and development 4 (1960), no. 1 66 -- 82.
[26] M. Studeny and J. Vejnarova, The Multi-information function as a tool for measuring
stochastic dependence, pp. 261 -- 298. Kluwer, January, 1998.
[27] R. W. Buccigrossi and E. P. Simoncelli, Image compression via joint statistical
characterization in the wavelet domain, IEEE Transactions on Image Processing 8 (Dec,
1999) 1688 -- 1701.
[28] Jinggang Huang and D. Mumford, Statistics of natural images and models, in Proceedings.
1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat.
No PR00149), vol. 1, pp. 541 -- 547 Vol. 1, June, 1999.
[29] A. Hyvärinen, J. Hurri, and J. Väyrynen, Bubbles: a unifying framework for low-level
statistical properties of natural image sequences, JOSA A 20 (2003), no. 7 1237 -- 1252.
[30] A. Hyvärinen and P. O. Hoyer, A two-layer sparse coding model learns simple and complex
cell receptive fields and topography from natural images, Vision Research 41 (2001), no. 18
2413 -- 2423.
[31] A. Hyvärinen and P. Hoyer, Emergence of phase- and shift-invariant features by
decomposition of natural images into independent feature subspaces, Neural Computation 12
(2000), no. 7 1705 -- 1720.
[32] A. Hyvärinen, J. Hurri, and P. Hoyer, Natural Image Statistics: A probabilistic approach to
early computational vision. Springer, 2009.
[33] J. Malo, R. Navarro, I. Epifanio, F. Ferri, and J. Artigas, Non-linear invertible representation
for joint statistical and perceptual feature decorrelation, Lect. Not. Comp. Sci. 1876 (2000)
658 -- 667.
[34] I. Epifanio, J. Gutierrez, and J. Malo, Linear transform for simultaneous diagonalization of
covariance and perceptual metric matrix in image coding, Pattern Recognition 36 (2003),
no. 8 1799 -- 1811.
[35] J. Liu and P. Moulin, Information-theoretic analysis of interscale and intrascale dependencies
between image wavelet coefficients, IEEE Trans. Im. Proc. 10 (2001), no. 11 1647 -- 1658.
[36] M. Carandini and D. J. Heeger, Normalization as a canonical neural computation, Nature
Rev. Neurosci. 13 (2012), no. 1 51 -- 62.
-- 30 --
[37] J. Foley, Human luminance pattern mechanisms: Masking experiments require a new model,
Journal of the Optical Society of America A 11 (1994), no. 6 1710 -- 1719.
[38] A. B. Watson and J. A. Solomon, Model of visual contrast gain control and pattern masking,
JOSA A 14 (1997), no. 9 2379 -- 2391.
[39] P. C. Bressloff and J. D. Cowan, The functional geometry of local and horizontal connections
in a model of v1, Journal of Physiology-Paris 97 (2003), no. 2 221 -- 236. Neurogeometry and
visual perception.
[40] J. Malo and E. Simoncelli, Geometrical and statistical properties of vision models obtained
via maximum differentiation, in SPIE Electronic Imaging, pp. 93940L -- 93940L, International
Society for Optics and Photonics, 2015.
[41] M. Martinez, M. Bertalmío, and J. Malo, In praise of artifice reloaded: Caution with natural
image databases in modeling vision, Front. Neurosci. doi: 10.3389/fnins.2019.00008 (2019).
[42] P. Chossat and O. Faugeras, Hyperbolic planforms in relation to visual edges and textures
perception, PLoS Computational Biology 5 (12, 2009) e1000625.
[43] J. H. v. Hateren and A. v. d. Schaaf, Independent component filters of natural images
compared with simple cells in primary visual cortex, Proceedings: Biological Sciences 265
(Mar, 1998) 359 -- 366.
[44] N. Ponomarenko, M. Carli, V. Lukin, K. Egiazarian, J. Astola, and F. Battisti, Color image
database for evaluation of image quality metrics, Proc. Int. Workshop on Multimedia Signal
Processing (Oct., 2008) 403 -- 408.
[45] G. Wyszecki and W. Stiles, Color Science: Concepts and Methods, Quantitative Data and
Formulae. John Wiley & Sons, New York, 1982.
[46] A. Stockman and D. Brainard, OSA Handbook of Optics (3rd. Ed.), ch. Color vision
mechanisms, pp. 147 -- 152. McGraw-Hill, NY, 2010.
[47] M. Fairchild, Color Appearance Models. The Wiley-IS&T Series in Imaging Science and
Technology. Wiley, 2013.
[48] J. Vazquez-Corral, C. Párraga, R. Baldrich, and M. Vanrell, Color constancy algorithms:
Psychophysical evaluation on a new dataset, Journal of Imaging Science and Technology 53
(2009), no. 3 31105 -- 1 -- 31105 -- 9.
[49] V. Laparra, S. Jiménez, G. Camps-Valls, and J. Malo, Nonlinearities and adaptation of color
vision from sequential principal curves analysis, Neural Computation 24 (2012), no. 10
2751 -- 2788.
[50] M. U. Gutmann, V. Laparra, A. Hyvärinen, and J. Malo, Spatio-chromatic adaptation via
higher-order canonical correlation analysis of natural images, PloS ONE 9 (2014), no. 2
e86481.
[51] A. Stockman and L. T. Sharpe, The spectral sensitivities of the middle- and
long-wavelength-sensitive cones derived from measurements in observers of known genotype,
Vision Research 40 (2000), no. 13 1711 -- 1737.
[52] J. Malo and M. Luque, "ColorLab: A Matlab Toolbox for Color Science and Calibrated
Color Image Processing." http://isp.uv.es/code/visioncolor/colorlab.html, 2002.
[53] M. A. Webster and J. D. Mollon, Colour constancy influenced by contrast adaptation, Nature
373 (1995) 694 -- 698.
-- 31 --
[54] M. A. Webster and J. Mollon, Adaptation and the color statistics of natural images, Vision
Research 37 (1997), no. 23 3283 -- 3298.
[55] C. R. Ingling Jr and B. H.-P. Tsou, Orthogonal combination of the three visual channels,
Vision research 17 (1977), no. 9 1075 -- 1082.
[56] G. Buchsbaum and A. Gottschalk, Trichromacy, opponent colours coding and optimum
colour information transmission in the retina, Proc. Roy. Soc. Lond. B Biol. Sci. 220 (1983),
no. 1218 89 -- 113.
[57] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd Edition.
Wiley-Interscience, 2 ed., July, 2006.
[58] D. MacLeod and T. von der Twer, The pleistochrome: optimal opponent codes for natural
colors, in Color Perception: From Light to Object (D. Heyer and R. Mausfeld, eds.). Oxford
Univ. Press, Oxford, UK, 2003.
[59] A. Hyvärinen and E. Oja, Independent component analysis: algorithms and applications,
Neural networks 13 (2000), no. 4-5 411 -- 430.
[60] E. Simoncelli, Statistical models for images: Compression, restoration and synthesis, in
Conference Record of the Asilomar Conference on Signals, Systems and Computers
(M. Farques and R. Hippenstiel, eds.), vol. 1, pp. 673 -- 678, IEEE Comp Soc, 1998.
[61] J. Malo, F. Ferri, J. Albert, J. Soret, and J. Artigas, The role of perceptual contrast
non-linearities in image transform quantization, Image and Vision Computing 18 (2000),
no. 3 233 -- 246.
[62] A. Watson, The cortex transform: Rapid computation of simulated neural images, Computer
Vision, Graphics and Image Processing 39 (1987) 311 -- 327.
[63] F. Campbell and J. Robson, Application of Fourier analysis to the visibility of gratings,
Journal of Physiology 197 (1968) 551 -- 566.
[64] P. Hancock, R. Baddeley, and L. Smith, The principal components of natural images,
Network 3 (1991) 61 -- 70.
[65] E. P. Simoncelli, W. T. Freeman, E. H. Adelson, and D. J. Heeger, Shiftable multi-scale
transforms, IEEE Trans Information Theory 38 (Mar, 1992) 587 -- 607. Special Issue on
Wavelets.
[66] J. Cardoso, Dependence, correlation and gaussianity in independent component analysis, J.
Mach. Learn. Res. 4 (2003) 1177 -- 1203.
[67] S. B. Laughlin, Matching coding to scenes to enhance efficiency, in In Braddick, O.J. &
Sleigh, A.C. (Eds) Physical and Biological Processing of Images, pp. 42 -- 52, Springer, 1983.
[68] D. L. Ruderman, The statistics of natural images, Network: computation in neural systems 5
(1994), no. 4 517 -- 548.
[69] J. Ballé, V. Laparra, and E. P. Simoncelli, Density modeling of images using a generalized
normalization transformation, in Int. Conf. Learn. Rep. (ICLR 16'), 2016.
[70] S. Lyu and E. P. Simoncelli, Nonlinear extraction of independent components of natural
images using radial gaussianization, Neural Comp. 21 (2009), no. 6 1485 -- 1519.
[71] B. W. Silverman, Density Estimation for Statistics and Data Analysis. Chapman & Hall,
London, 1986.
-- 32 --
[72] V. Laparra, J. Gutiérrez, G. Camps-Valls, and J. Malo, Image denoising with kernels based
on natural image relations, J. Mach. Learn. Res. 11 (2010) 873 -- 903.
[73] B. Olshausen and D. Field, Emergence of simple-cell receptive field properties by learning a
sparse code for natural images, Nature 281 (1996) 607 -- 609.
[74] H. Barlow, Vision: Coding and Efficiency, ch. A theory about the functional role and
synaptic mechanism of visual aftereffects. Cambridge Univ. Press, Cambridge, UK, 1990.
[75] V. Laparra, J. Muñoz-Marí, and J. Malo, Divisive normalization image quality metric
revisited, JOSA A 27 (2010), no. 4 852 -- 864.
[76] A. Berardino, J. Balle, V. Laparra, and E. Simoncelli, Eigen-distortion of hierarchical
representations, Adv. Neur. Inf. Proc. NIPS-17 (2017).
[77] J. E. Johnson, E. Diaz, V. Laparra, M. Mahecha, D. Miralles, and G. Camps-Valls,
Estimating information in Earth data cubes, Geophysical Research Abstracts Vol. 20 (2018)
12702.
[78] V. Laparra and R. Santos-Rodríguez, Spatial/spectral information trade-off in hyperspectral
images, in 2015 IEEE Int. Geosci. Rem. Sens. Symp. (IGARSS), pp. 1124 -- 1127, July, 2015.
[79] A. Painsky and N. Tishby, Gaussian lower bound for the information bottleneck limit,
Journal of Machine Learning Research 18 (2018), no. 213 1 -- 29.
-- 33 --
|
1710.02563 | 2 | 1710 | 2018-11-02T17:53:21 | Expectation-induced modulation of metastable activity underlies faster coding of sensory stimuli | [
"q-bio.NC"
] | Sensory stimuli can be recognized more rapidly when they are expected. This phenomenon depends on expectation affecting the cortical processing of sensory information. However, virtually nothing is known on the mechanisms responsible for the effects of expectation on sensory networks. Here, we report a novel computational mechanism underlying the expectation-dependent acceleration of coding observed in the gustatory cortex (GC) of alert rats. We use a recurrent spiking network model with a clustered architecture capturing essential features of cortical activity, including the metastable activity observed in GC before and after gustatory stimulation. Relying both on network theory and computer simulations, we propose that expectation exerts its function by modulating the intrinsically generated dynamics preceding taste delivery. Our model, whose predictions are confirmed in the experimental data, demonstrates how the modulation of intrinsic metastable activity can shape sensory coding and mediate cognitive processes such as the expectation of relevant events. Altogether, these results provide a biologically plausible theory of expectation and ascribe a new functional role to intrinsically generated, metastable activity. | q-bio.NC | q-bio | Expectation-induced modulation of metastable activity underlies
faster coding of sensory stimuli
L. Mazzucato1,2, G. La Camera1,3,*, A. Fontanini1,3,*
1 Department of Neurobiology and Behavior, State University of New York at Stony Brook,
Stony Brook, NY 11794
2 Departments of Biology and Mathematics and Institute of Neuroscience, University of Oregon,
Eugene, OR 97403
3 Graduate Program in Neuroscience, State University of New York at Stony Brook, Stony
Brook, NY 11794
* Co-senior authors
Running Title: Metastable activity mediates expectation
Contact:
Alfredo Fontaninia and Giancarlo La Camerab
a Department of Neurobiology & Behavior, Life Sciences Building 516
State University of New York at Stony Brook, Stony Brook, NY 11794
[email protected]
b Department of Neurobiology & Behavior, Life Sciences Building 513
State University of New York at Stony Brook, Stony Brook, NY 11794
[email protected]
Abstract
Sensory stimuli can be recognized more rapidly when they are expected. This
phenomenon depends on expectation affecting the cortical processing of sensory
information. However, virtually nothing is known on the mechanisms responsible for
the effects of expectation on sensory networks. Here, we report a novel computational
mechanism underlying the expectation-dependent acceleration of coding observed in
the gustatory cortex (GC) of alert rats. We use a recurrent spiking network model with a
clustered architecture capturing essential features of cortical activity, including the
metastable activity observed in GC before and after gustatory stimulation. Relying both
on network theory and computer simulations, we propose that expectation exerts its
function by modulating the intrinsically generated dynamics preceding taste delivery.
Our model, whose predictions are confirmed in the experimental data, demonstrates
how the modulation of intrinsic metastable activity can shape sensory coding and
mediate cognitive processes such as the expectation of relevant events. Altogether, these
results provide a biologically plausible theory of expectation and ascribe a new
functional role to intrinsically generated, metastable activity.
Introduction
Expectation exerts a strong influence on sensory processing. It improves stimulus
detection, enhances discrimination between multiple stimuli and biases perception
towards an anticipated stimulus1-3. These effects, demonstrated experimentally for
various sensory modalities and in different species2,4-7, can be attributed to changes in
sensory processing occurring in primary sensory cortices. However, despite decades of
investigations, little is known regarding how expectation shapes the cortical processing
of sensory information.
While different forms of expectation likely rely on a variety of neural mechanisms,
modulation of pre-stimulus activity is believed to be a common underlying feature8-10.
Here, we investigate the link between pre-stimulus activity and the phenomenon of
general expectation in a recent set of experiments performed in gustatory cortex (GC) of
alert rats6. In those experiments, rats were trained to expect the intraoral delivery of one
of four possible tastants following an anticipatory cue. The use of a single cue allowed
the animal to predict the availability of gustatory stimuli, without forming expectations
on which specific taste was being delivered. Cues predicting the general availability of
taste modulated the firing rates of GC neurons. Tastants delivered after the cue were
encoded more
improvement was
phenomenologically attributed to the activity evoked by the preparatory cue. However,
tastants, and
this
rapidly
than uncued
2
the precise computational mechanism linking faster coding of taste and cue responses
remains unknown.
Here we propose a mechanism whereby an anticipatory cue modulates the timescale of
temporal dynamics in a recurrent population model of spiking neurons. In the model
proposed here, neurons are organized in strongly connected clusters and produce
sequences of metastable states similar to those observed during both pre-stimulus and
evoked activity periods11-17. A metastable state is a vector of firing rates across
simultaneously recorded neurons that can last for several hundred milliseconds before
giving way to the next state in a sequence. The ubiquitous presence of state sequences
in many cortical areas and behavioral contexts18-24 has raised the issue of their role in
sensory and cognitive processing. Here, we elucidate the central role played by pre-
stimulus metastable states in processing forthcoming stimuli, and show how cue-
induced modulations of state sequences drive anticipatory coding. Specifically, we
show that an anticipatory cue affects sensory coding by decreasing the duration of
metastable states and accelerating the pace of state sequences. This phenomenon, which
results from a reduction in the effective energy barriers separating the metastable states,
accelerates the onset of specific states coding for the presented stimulus, thus mediating
the effects of general expectation. The predictions of our model were confirmed in a
new analysis of the experimental data, also reported here.
Altogether, our results provide a model for general expectation, based on the
modulation of pre-stimulus ongoing cortical dynamics by anticipatory cues, leading to
acceleration of sensory coding.
Results
Anticipatory cue accelerates stimulus coding in a clustered population of neurons
To uncover the computational mechanism linking cue-evoked activity with coding
speed, we modeled the gustatory cortex (GC) as a population of recurrently connected
excitatory and inhibitory spiking neurons. In this model, excitatory neurons are
arranged in clusters12,25,26 (Fig. 1a), reflecting the existence of assemblies of functionally
correlated neurons in GC and other cortical areas27,28. Recurrent synaptic weights
between neurons in the same cluster are potentiated compared to neurons in different
clusters, to account for metastability in GC11,16 and in keeping with evidence from
electrophysiological and imaging experiments27,28 29,30. This spiking network also has
bidirectional random and homogeneous (i.e., non-clustered) connections among
inhibitory neurons and between inhibitory and excitatory neurons. Such connections
stabilize network activity by preventing runaway excitation and play a role in inducing
the observed metastability11,12,15.
3
The model was probed by sensory inputs modeled as depolarizing currents injected
into randomly selected neurons. We used four sets of simulated stimuli, wired to
produce gustatory responses reminiscent of those observed in the experiments in the
presence of sucrose, sodium chloride, citric acid, and quinine (see Supplementary
Methods for details). The specific connectivity pattern used was inferred by the
presence of both broadly and narrowly tuned responses in GC31,32, and the temporal
dynamics of the inputs were varied to determine the robustness of the model
(Supplementary Results, Sec. 1.3).
In addition to input gustatory stimuli, we included anticipatory inputs designed to
produce cue-responses analogous to those seen experimentally in the case of general
expectation. To simulate general expectation, we connected anticipatory inputs with
random neuronal targets in the network. The peak value of the cue-induced current for
each neuron was sampled from a normal distribution with zero mean and fixed
variance (see Fig. S1-S2 and Supplementary Results for details), thus introducing a
spatial variance in the afferent currents. This choice reflected the large heterogeneity of
cue responses observed in the empirical data, where excited and inhibited neural
responses occurred in similar proportions10 and overlapped partially with taste
responses6,10. Fig 1b shows two representative cue-responsive neurons in the model: one
inhibited by the cue and one excited by the cue (more details and examples are reported
in the Supplementary Results).
Given these conditions, we simulated the experimental paradigm adopted in awake-
behaving rats to demonstrate the effects of general expectation6,10. In the original
experiment, rats were trained to self-administer into an intra-oral cannula one of four
possible tastants following an anticipatory cue. At random trials and time during the
inter-trial interval, tastants were unexpectedly delivered in the absence of a cue. To
match this experiment, the simulated paradigm interleaves two conditions: in expected
trials, a stimulus (out of 4) is delivered at t=0 after an anticipatory cue (the same for all
stimuli) delivered at t=-0.5s (Fig. 1b); in unexpected trials the same stimuli are presented
in the absence of the cue. Importantly, in the general expectation paradigm adopted
here, the anticipatory cue is identical for all stimuli in the expected condition. Therefore,
it does not convey any information regarding the identity of the stimulus being
delivered.
We tested whether cue presentation affected stimulus coding. A multi-class classifier
(see Methods and Fig. S3) was used to assess the information about the stimuli encoded
in the neural activity, where the four class labels correspond to the four tastants.
Stimulus identity was encoded well in both conditions, reaching perfect average
accuracy across the four tastants after a few hundred milliseconds (Fig. 1c, across-taste
average decoding accuracy). However, comparing the time course of the decoding
accuracy between conditions, we found that the increase in decoding accuracy was
4
significantly faster in expected than in unexpected trials (Fig. 1c, pink and blue curves
represent expected and unexpected conditions, respectively). Indeed, the onset time of a
significant decoding occurred earlier in the expected vs. the unexpected condition
(decoding latency was 0.13±0.01 s [mean±s.e.m.] for expected compared to 0.21±
0.02 s for unexpected, across 20 independent sessions; p=0.002, signed-rank=14,
d.o.f.=39; inset in Fig. 1c). Similar decoding accuracies were obtained for each individual
tastant separately (see Supplementary Results and Fig. S3). Thus, in the model network,
the interaction of cue response and activity evoked by the stimuli results in faster
encoding of the stimuli themselves, mediating the expectation effect.
To clarify the role of neural clusters in mediating expectation, we simulated the same
experiments in a homogeneous network (i.e., without clusters) operating in the
balanced asynchronous regime25,26 (Fig. 1d, intra- and inter-cluster weights were set
equal, all other network parameters and inputs were the same as for the clustered
network). Even though single neurons' responses to the anticipatory cue were
comparable to the ones observed in the clustered network (Fig. 1e, Fig. S2 and
Supplementary Results), stimulus encoding was not affected by cue presentation (Fig.
1f). In particular, the onset of a significant decoding was similar in the two conditions
(latency of significant decoding was 0.17±0.01 s for expected and 0.16±0.01 s for
unexpected tastes averaged across 20 sessions; p=0.31, signed-rank=131, d.o.f.=39; inset
in Fig. 1f).
The anticipatory activity observed in the clustered network was robust to variations in
key parameters related to the sensory and anticipatory inputs, as well as network
connectivity, size and architecture (Fig. S4-6 and Supplementary Results). Furthermore,
acceleration of coding depended on the patterns of connectivity of anticipatory inputs,
specifically on the fact that it increased the spatial variance in the cue afferent currents
(Fig. S6). In a model where the cue recruited the recurrent inhibition (by increasing the
input currents to the inhibitory population), stimulus coding was decelerated (Fig. S7),
suggesting a potential mechanism mediating the effect of distractors.
Overall, these results demonstrate that a clustered network of spiking neurons can
successfully reproduce the acceleration of sensory coding induced by expectation and
that removing clustering impairs this function.
Anticipatory cue speeds up the network's dynamics
Having established that a clustered architecture mediates the effects of expectation on
coding, we investigated the underlying mechanism.
Clustered networks spontaneously generate highly structured activity characterized by
coordinated patterns of ensemble firing. This activity results from the network hopping
5
between metastable states
in which different combinations of clusters are
simultaneously activated11,14,15. To understand how anticipatory inputs affected network
dynamics, we analyzed the effects of cue presentation for a prolonged period of 5
seconds in the absence of stimuli. Activating anticipatory inputs led to changes in
network dynamics, with clusters turning on and off more frequently in the presence of
the cue (Fig. 2a). We quantified this effect by showing that a cue-induced increase in
input spatial variance (𝜎*) led to a shortened cluster activation lifetime (top panel in
Fig. 2b; Kruskal-Wallis one-way ANOVA: p<10-17, 𝜒*(5)=91.2), and a shorter cluster
𝜒*(5)=98.6).
inter-activation interval (i.e., quiescent intervals between consecutive activations of the
same cluster, bottom panel in Fig. 2b, kruskal-wallis one-way ANOVA: p<10-18,
Previous work has demonstrated that metastable states of co-activated clusters result
from attractor dynamics11,14,15. Hence, the shortening of cluster activations and inter-
activation intervals observed in the model could be due to modifications in the
network's attractor dynamics. To test this hypothesis, we performed a mean field
theory analysis33-36 of a simplified network with only two clusters, therefore producing a
reduced repertoire of configurations. Those include two configurations in which either
cluster is active and the other inactive ('A' and 'B' in Fig. 2c), and a configuration where
both clusters are moderately active ('C'). The dynamics of this network can be analyzed
using a reduced, self-consistent theory of a single excitatory cluster, said to be in focus33
(see Methods for details), based on the effective transfer function relating the input and
output firing rates of the cluster (r and rout, Fig. 2c). The latter are equal in the A, B and C
network configurations described above -- also called 'fixed points' since these are the
points where the transfer function intersects the identity line, 𝑟012=Φ(𝑟56).
Configurations A and B would be stable in an infinitely large network, but they are only
metastable in networks of finite size, due to intrinsically generated variability15.
Transitions between metastable states can be modeled as a diffusion process and
analyzed with Kramers' theory37, according to which the transition rates depend on the
height D of an effective energy barrier separating them15,37. In our theory, the effective
energy barriers (Fig. 2c, bottom row) are obtained as the area of the region between the
identity line and the transfer function (shaded areas in top row of Fig. 2c; see Methods
for details). The effective energy is constructed so that its local minima correspond to
stable fixed points (here, A and B) while local maxima correspond to unstable fixed
points (C). Larger barriers correspond to less frequent transitions between stable
configurations, whereas lower barriers increase the transition rates and therefore
accelerate the network's metastable dynamics.
This picture provides the substrate for understanding the role of the anticipatory cue in
the expectation effect. Basically, the presentation of the cue modulates the shape of the
effective transfer function, which results in the reduction of the effective energy
6
barriers. More specifically, the cue-induced increase in the spatial variance, s2, of the
afferent current flattens the transfer function along the identity line, reducing the area
between the two (shaded regions in Fig. 2c). In turn, this reduces the effective energy
barrier separating the two configurations (Fig. 2c, bottom row), resulting in faster
dynamics. The larger the cue-induced spatial variance s2 in the afferent currents, the
faster the dynamics (Fig. 2d; lighter shades represent larger ss).
In summary, this analysis shows that the anticipatory cue increases the spontaneous
transition rates between the network's metastable configurations by reducing the
effective energy barrier necessary to hop among configurations. In the following we
uncover an important consequence of this phenomenon for sensory processing.
Anticipatory cue induces faster onset of taste-coding states
The cue-induced modulation of attractor dynamics led us to formulate a hypothesis for
the mechanism underlying the acceleration of coding: The activation of anticipatory
inputs prior to sensory stimulation may allow the network to enter more easily
configurations encoding stimuli while exiting more easily non-coding configurations.
Fig. 3a shows simulated population rasters in response to the same stimulus presented
in the absence of a cue or after a cue. Spikes in red hue represent activity in taste-
selective clusters and show a faster activation latency in response to the stimulus
preceded by the cue compared to the uncued condition. A systematic analysis revealed
that in the cued condition, the clusters activated by the subsequent stimulus had a
significantly faster activation latency than in the uncued condition (Fig. 3b, 0.22±0.01 s
(mean±s.e.m.) during cued compared to 0.32±0.01 s for uncued stimuli; p<10-5, rank
sum test R(39)=232).
We elucidated this effect using mean field theory. In the simplified two-cluster network
of Fig. 3c (the same network as in Fig. 2d), the configuration where the taste-selective
cluster is active ("coding state") and the nonselective cluster is active ("non-coding
state") have initially the same effective potential energy, in the absence of stimulation
(local minima of the black line in Fig. 3c), separated by an effective energy barrier
whose height is reduced by the anticipatory cue (dashed vs. full line). When the taste
stimulus is presented, it activates the stimulus-selective cluster, so that the coding state
will now sit in a deeper well (lighter lines) compared to non-coding state. Stronger
stimuli (lighter shades in Fig. 3c) increase the difference between the wells' depths
breaking their initial symmetry, so that now a transition from the non-coding to the
coding state is more likely than a transition from the coding to the non-coding state37,
(also in the absence of the cue; full lines). The anticipatory cue reduces further the
existing barrier and thereby increases the transition rate into coding configurations.
This results into faster coding, on average, of the stimuli encoded by those states.
7
We tested this model prediction on the data from Samuelsen et al. (Fig 4)6. To compare
the data to the model simulations, we randomly sampled ensembles of model neurons
so as to match the sizes of the empirical datasets. Since we only have access to a subset
of neurons in the experiments, rather than the full network configuration, we
segmented the ensemble activity in sequences of metastable states via a Hidden Markov
Model (HMM) analysis (see Methods). Previous work has demonstrated that HMM
states, i.e. patterns of coordinated ensemble firing activity, can be treated as proxies of
metastable network configurations11.
taste-coding
configurations for a particular stimulus results in HMM states containing information
about that stimulus (i.e., taste-coding HMM states). If the hypothesis originating from
the model is correct, transitions from non-coding HMM states to taste-coding HMM
states should be faster in the presence of the cue compared to uncued trials. We indeed
found faster transitions to HMM coding states in cued trials for both model and data
(Fig. 4a and 4c, respectively; color-coded horizontal bars overlay coding states). The
latency of coding states was significantly faster during cued compared to uncued trials
in both the model (Fig. 4b, mean latency of the first coding state was 0.32±0.02 s for
expected vs 0.38±0.01 s for unexpected trials; rank sum test R(39)=319, p=0.014) and
the empirical data (Fig. 4d: 0.46±0.02 s for expected vs 0.56±0.03 s for unexpected
In particular, activation of
trials; rank sum test R(37)=385, p=0.026).
Altogether, these results demonstrate that anticipatory inputs speed up sensory coding
by reducing the effective energy barriers from non-coding to coding metastable states
(i.e., the transitions facilitated by the stimulus).
Discussion
Expectations modulate perception and sensory processing. Typically, expected stimuli
are recognized more accurately and rapidly than unexpected ones 1-3. In the gustatory
cortex, acceleration of taste coding has been related to changes in firing activity evoked
by cues predicting the general availability of tastants6. However, the computational
mechanisms linking pre-stimulus activity with changes in the latency of sensory coding
are still unknown. Here we propose a novel mechanism that explains the effects of
expectation through the modulation of the dynamics intrinsically generated by the
cortex. Our results provide a new functional interpretation for the intrinsically
generated activity that is ubiquitously observed in cortical circuits11,18-21,24,38-42.
The proposed mechanism requires a recurrent spiking network where excitatory
neurons are arranged in clusters, which has been demonstrated to capture essential
features of the dynamics of neural activity in sensory circuits11,16. In such a model,
network activity during both spontaneous and stimulus-evoked periods unfolds
through state sequences, each state representing a metastable network attractor. In
8
response to an anticipatory cue, the pace of state sequences speeds up, accelerated by a
higher transition probability among states. The latter is caused by lowering the
potential barrier separating metastable states in the attractor landscape. This anticipates
the offset of states not conveying taste information and the onset of states containing the
most information about the delivered stimulus ('coding states'), causing the faster
decoding observed by Samuelsen et al6 (see Fig. 1c).
Notably, this novel mechanism for anticipation is unrelated to increases in network
excitability which would lead to unidirectional changes in taste-evoked firing rates. It
relies instead on an increase in the spatial variance of the cue afferent currents to the
sensory network brought about by the anticipatory cue. This increase in the input's
variance is observed experimentally after training10, and is therefore the consequence of
having learned the anticipatory meaning of the cue. The acceleration of the dynamics of
state sequences predicted by the model was also confirmed in the data from ensembles
of simultaneously recorded neurons in awake-behaving rats.
These results provide a precise explanatory link between the intrinsic dynamics of
neural activity in a sensory circuit and a specific cognitive process, that of general
expectation6 (see also 43,44).
Clustered connectivity and metastable states
A key feature of our model is the clustered architecture of the excitatory population.
Removing excitatory clusters eliminates the cue-induced anticipatory effect (Fig. 1d-f).
Theoretical work in recurrent networks had previously shown that a clustered
architecture can produce stable patterns of population activity called attractors12. Noise
(either externally14,45 or internally generated11,15) may destabilize those states, driving the
emergence of temporal dynamics based on the progression through metastable states.
Network models with clustered architecture provide a parsimonious explanation for the
state sequences that have been observed ubiquitously in alert mammalian cortex,
during both task engagement17,18,46,47 and inter-trial periods.11,39,40 In addition, this type of
models accounts for various physiological observations such as stimulus-induced
reduction of trial-to-trial variability11,14,15,48, neural dimensionality16, and firing rate
multistability11 (see also49,50). In particular, models with metastable attractors have been
used to explain the state sequences observed in rodent gustatory cortex during taste
processing and decision making11,45,51.
In this work, we propose that clustered networks have the ability to modulate coding
latency, and demonstrate one specific mechanism for modulation that can underlie the
phenomenon of general expectation. Changes in the depth of attractor wells, induced by
a non-stimulus specific anticipatory cue (which in turn may depend on the activation of
top-down and neuromodulatory afferents6,52), can accelerate or slow down network
9
dynamics. The acceleration resulting from shallower wells leads to a reshaping of
ongoing activity and to a quicker recruitment of states coding for sensory information.
To our knowledge, the link between generic anticipatory cues, network metastability,
and coding speed as presented here is novel and represents the main innovation of our
work.
Functional role of heterogeneity in cue responses
As stated in the previous section, the presence of clusters is a necessary ingredient to
obtain a faster latency of coding. Here we discuss the second necessary ingredient, i.e.,
the presence of heterogeneous neural responses to the anticipatory cue (Fig. 1b).
Responses to anticipatory cues have been extensively studied in cortical and subcortical
areas in alert rodents6,10,53,54. Cues evoke heterogeneous patterns of activity, either
exciting or inhibiting single neurons. The proportion of cue responses and their
heterogeneity develops with associative learning,10,54 suggesting a fundamental function
of these patterns. In the generic expectation paradigm considered here, the anticipatory
cue does not convey any information about the identity of the forthcoming tastant,
rather it just signals the availability of a stimulus. Experimental evidence suggests that
the cue may induce a state of arousal, which was previously described as "priming" the
sensory cortex6,55,56. Here, we propose an underlying mechanism in which the cue is
responsible for acceleration of coding by increasing the spatial variance of pre-stimulus
activity. In turn, this modulates the shape of the neuronal current-to-rate transfer
function and
the effective energy barriers between metastable
configurations.
We note that the presence of both excited and inhibited cue responses poses a challenge
to simple models of neuromodulation. The presence of cue-evoked suppression of
firing10 suggests that cues do not improve coding by simply increasing the excitability of
cortical areas. Additional mechanisms and complex patterns of connectivity may be
required to explain the suppression effects induced by the cue. However, here we
provide a parsimonious explanation of how heterogeneous responses can improve
coding without postulating any specific pattern of connectivity other than i) random
projections from thalamic and anticipatory cue afferents and ii) the clustered
organization of the intra-cortical circuitry. Notice that the latter contains wide
distributions of synaptic weights and can be understood as the consequence of Hebb-
like re-organization of the circuitry during training57,58.
It is also worth noting that our model incorporates excited and inhibited cue responses
in such a manner to affect only the spatial variance of the activity across neurons, while
leaving the mean input to the network unaffected. As a result, the anticipatory cue
leaves average firing rates unchanged in the clustered network (Fig. S10), and only
lowers
thus
10
modulates the network temporal dynamics. Our model thus provides a mechanism
whereby increasing the spatial variance of top-down inputs has, paradoxically, a
beneficial effect on sensory coding.
Specificity of the anticipatory mechanism
Our model of anticipation relies on gain reduction in clustered excitatory neurons due
to a larger spatial variance of the afferent currents. This model is robust to variations in
parameters and architecture (Fig. S1, S4-S6). A priori, this effect might be achieved
through different means, such as: increasing the strength of feedforward couplings;
decreasing the strength of recurrent couplings; or modulating background synaptic
inputs59. However, when scoring those models on the criteria of coding anticipation and
heterogeneous cue responses, we found that they failed to simultaneously match both
criteria, although for some range of parameters they could reproduce either one (see
Fig. S8-9 and Supplementary Results for a detailed analysis). Thus, we concluded that
only the main mechanism proposed here (Fig. 1a) captures the plurality of experimental
observations pertaining anticipatory activity in a robust and biologically plausible way.
Cortical timescales, state transitions, and cognitive function
In populations of spiking neurons, a clustered architecture can generate reverberating
activity and sequences of metastable states. Transitions from state to state can be
typically caused by external inputs11,15. For instance, in frontal cortices, sequences of
states are related to specific epochs within a task, with transitions evoked by behavioral
events18,19,22. In sensory cortex, progressions through state sequences can be triggered by
sensory stimuli and reflect the dynamics of sensory processing23,46. Importantly, state
sequences have been observed also in the absence of any external stimulation,
promoted by intrinsic fluctuations in neural activity11,41. However, the potential
functional role, if any, of this type of ongoing activity has remained unexplored.
Recent work has started to uncover the link between ensemble dynamics and sensory
and cognitive processes. State transitions in various cortical areas have been linked to
decision making45,60, choice representation22, rule-switching behavior24, and the level of
task difficulty23. However, no theoretical or mechanistic explanations have been given
for these phenomena.
Here we provide a mechanistic link between state sequences and expectation, by
showing that intrinsically generated sequences can be accelerated, or slowed down,
thus affecting sensory coding. Moreover, we show that the interaction between external
stimuli and intrinsic dynamics does not result in the simple triggering of state
transitions, but rather in the modulation of the intrinsic transition probabilities. The
11
modulation of intrinsic activity can dial the duration of states, producing either shorter
or longer timescales. A shorter timescale leads to faster state sequences and coding
anticipation after stimulus presentation (Fig. 1 and 4). Other external perturbations may
induce different effects: for example, recruiting the network's inhibitory population
slows down the timescale, leading to a slower coding (Fig. S7).
The interplay between intrinsic dynamics and anticipatory influences presented here is
a novel mechanism for generating diverse timescales, and may have rich computational
consequences. We demonstrated its function in increasing coding speed, but its role in
mediating cognition is likely to be broader and calls for further explorations.
Author Contributions
LM, GLC, and AF designed the project, discussed the models and the data analyses,
and wrote the manuscript; LM performed the data analysis, model simulations and
theoretical analyses.
Acknowledgements
This work was supported by a National Institute of Deafness and Other
Communication Disorders Grant K25-DC013557 (LM), by the Swartz Foundation
Award 66438 (LM), by National Institute of Deafness and Other Communication
Disorders Grants NIDCD R01DC012543 and R01DC015234 (AF), and partly by a
National Science Foundation Grant IIS-1161852 (GLC). The authors would like to thank
Drs. S. Fusi, A. Maffei, G. Mongillo, and C. van Vreeswijk for useful discussions.
Competing Financial Interests
The authors declare no competing financial interests.
References
•
Gilbert, C. D. & Sigman, M. Brain states: top-down influences in sensory
1
processing. Neuron 54, 677-696 (2007).
Jaramillo, S. & Zador, A. M. The auditory cortex mediates the perceptual effects
2
of acoustic temporal expectation. Nature neuroscience 14, 246-251, doi:10.1038/nn.2688
(2011).
3
synchrony in top -- down processing. Nature Reviews Neuroscience 2, 704-716 (2001).
4
neurobiology 17, 465-470 (2007).
Nobre, A. C., Correa, A. & Coull, J. T. The hazards of time. Current opinion in
Engel, A. K., Fries, P. & Singer, W. Dynamic predictions: oscillations and
•
•
•
12
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
on
cortical
to discriminate
learning changes cross-modal
Vincis, R. & Fontanini, A. Associative
Samuelsen, C. L., Gardner, M. P. & Fontanini, A. Effects of cue-triggered
410-422,
taste. Neuron
Niwa, M., Johnson, J. S., O'Connor, K. N. & Sutter, M. L. Active engagement
temporal
Mazzucato, L., Fontanini, A. & La Camera, G. Dynamics of Multistable States
J Neurosci 35, 8214-8231,
5
Doherty, J. R., Rao, A., Mesulam, M. M. & Nobre, A. C. Synergistic effect of
combined temporal and spatial expectations on visual attention. Journal of Neuroscience
25, 8259-8266 (2005).
6
expectation
doi:10.1016/j.neuron.2012.02.031 (2012).
7
improves primary auditory cortical neurons' ability
modulation. Journal of Neuroscience 32, 9323-9334 (2012).
8
Yoshida, T. & Katz, D. B. Control of prestimulus activity related to improved
sensory coding within a discrimination task. The Journal of neuroscience : the official journal
of the Society for Neuroscience 31, 4101-4112, doi:10.1523/jneurosci.4380-10.2011 (2011).
9
Gardner, M. P. & Fontanini, A. Encoding and tracking of outcome-specific
expectancy in the gustatory cortex of alert rats. The Journal of neuroscience : the official
journal of the Society for Neuroscience 34, 13000-13017, doi:10.1523/jneurosci.1820-14.2014
(2014).
10
representations in the gustatory cortex. eLife 5, doi:10.7554/eLife.16420 (2016).
11
during Ongoing and Evoked Cortical Activity.
doi:10.1523/JNEUROSCI.4819-14.2015 (2015).
12
Amit, D. J. & Brunel, N. Model of global spontaneous activity and local
structured activity during delay periods in the cerebral cortex. Cerebral cortex (New York,
N.Y. : 1991) 7, 237-252 (1997).
13
Renart, A., Moreno-Bote, R., Wang, X. J. & Parga, N. Mean-driven and
fluctuation-driven persistent activity in recurrent networks. Neural computation 19, 1-46,
doi:10.1162/neco.2007.19.1.1 (2007).
14
variability
doi:10.1371/journal.pcbi.1002395 (2012).
Litwin-Kumar, A. & Doiron, B. Slow dynamics and high variability in balanced
15
cortical networks with clustered connections. Nature neuroscience 15, 1498-1505,
doi:10.1038/nn.3220 (2012).
16
of cortical activity. Front Sys Neurosci 10, 11 (2016).
17
cortex
doi:10.1038/nature10918 (2012).
18
Acad Sci USA 92, 8616-8620 (1995).
I., Abeles, M., Bergman, H. & Vaadia, E.
19
Simultaneously recorded single units in the frontal cortex go through sequences of
Harvey, C. D., Coen, P. & Tank, D. W. Choice-specific sequences in parietal
62-68,
virtual-navigation
Abeles, M. et al. Cortical activity flips among quasi-stationary states. Proc Natl
Mazzucato, L., Fontanini, A. & La Camera, G. Stimuli reduce the dimensionality
Deco, G. & Hugues, E. Neural network mechanisms underlying stimulus driven
e1002395,
computational
reduction.
biology
PLoS
8,
Seidemann, E., Meilijson,
decision
task. Nature
484,
during
a
processing
of
74,
13
Rich, E. L. & Wallis, J. D. Decoding subjective decisions from orbitofrontal cortex.
Renart, A. et al. The asynchronous state in cortical circuits. Science 327, 587-590,
discrete and stable states in monkeys performing a delayed localization task. The Journal
of neuroscience : the official journal of the Society for Neuroscience 16, 752-768 (1996).
20
Arieli, A., Shoham, D., Hildesheim, R. & Grinvald, A. Coherent spatiotemporal
patterns of ongoing activity revealed by real-time optical imaging coupled with single-
unit recording in the cat visual cortex. Journal of neurophysiology 73, 2072-2093 (1995).
21
Arieli, A., Sterkin, A., Grinvald, A. & Aertsen, A. Dynamics of ongoing activity:
explanation of the large variability in evoked cortical responses. Science 273, 1868-1871
(1996).
22
Nature neuroscience 19, 973-980, doi:10.1038/nn.4320 (2016).
23
Ponce-Alvarez, A., Nacher, V., Luna, R., Riehle, A. & Romo, R. Dynamics of
cortical neuronal ensembles transit from decision making to storage for later report. The
Journal of neuroscience : the official journal of the Society for Neuroscience 32, 11956-11969,
doi:10.1523/jneurosci.6176-11.2012 (2012).
24
Durstewitz, D., Vittoz, N. M., Floresco, S. B. & Seamans, J. K. Abrupt transitions
between prefrontal neural ensemble states accompany behavioral transitions during rule
learning. Neuron 66, 438-448, doi:10.1016/j.neuron.2010.03.029 (2010).
25
excitatory and inhibitory activity. Science 274, 1724-1726 (1996).
26
doi:10.1126/science.1179850 (2010).
27
Chen, X., Gabitto, M., Peng, Y., Ryba, N. J. & Zuker, C. S. A gustotopic map of
taste qualities in the mammalian brain. Science (New York, N.Y.) 333, 1262-1266,
doi:10.1126/science.1204076 (2011).
28
Fletcher, M. L., Ogg, M. C., Lu, L., Ogg, R. J. & Boughter, J. D., Jr. Overlapping
Representation of Primary Tastes in a Defined Region of the Gustatory Cortex. The
Journal of neuroscience : the official journal of the Society for Neuroscience 37, 7595-7605,
doi:10.1523/jneurosci.0649-17.2017 (2017).
Kiani, R. et al. Natural Grouping of Neural Responses Reveals Spatially
29
Segregated Clusters in Prearcuate Cortex. Neuron, doi:10.1016/j.neuron.2015.02.014
(2015).
30
Miller, J. E., Ayzenshtat, I., Carrillo-Reid, L. & Yuste, R. Visual stimuli recruit
intrinsically generated cortical ensembles. Proceedings of the National Academy of Sciences
of the United States of America 111, E4053-4061, doi:10.1073/pnas.1406077111 (2014).
31
Katz, D. B., Simon, S. A. & Nicolelis, M. A. Dynamic and multimodal responses
of gustatory cortical neurons in awake rats. The Journal of neuroscience : the official journal
of the Society for Neuroscience 21, 4478-4489 (2001).
32
Jezzini, A., Mazzucato, L., La Camera, G. & Fontanini, A. Processing of hedonic
and chemosensory features of taste in medial prefrontal and insular networks. The
Journal of neuroscience : the official journal of the Society for Neuroscience 33, 18966-18978,
doi:10.1523/JNEUROSCI.2974-13.2013 (2013).
33
population states. Network (Bristol, England) 10, 351-373 (1999).
van Vreeswijk, C. & Sompolinsky, H. Chaos in neuronal networks with balanced
Mascaro, M. & Amit, D. J. Effective neural response function for collective
14
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Hänggi, P., Talkner, P. & Borkovec, M. Reaction-rate theory: fifty years after
Kenet, T., Bibitchkov, D., Tsodyks, M., Grinvald, A. & Arieli, A. Spontaneously
representations of visual attributes. Nature 425, 954-956,
34
premotor cortex. Journal of Neuroscience 33, 11155-11168 (2013).
35
La Camera, G., Giugliano, M., Senn, W. & Fusi, S. The response of cortical
neurons to in vivo-like input current: theory and experiment : I. Noisy inputs with
stationary statistics. Biological cybernetics 99, 279-301, doi:10.1007/s00422-008-0272-7
(2008).
36
Curti, E., Mongillo, G., La Camera, G. & Amit, D. J. Mean field and capacity in
realistic networks of spiking neurons storing sparsely coded random memories. Neural
computation 16, 2597-2637, doi:10.1162/0899766042321805 (2004).
37
Kramers. Reviews of modern physics 62, 251 (1990).
38
emerging cortical
doi:10.1038/nature02078 (2003).
39
Pastalkova, E., Itskov, V., Amarasingham, A. & Buzsaki, G. Internally generated
cell assembly sequences in the rat hippocampus. Science (New York, N.Y.) 321, 1322-1327,
doi:10.1126/science.1159775 (2008).
40
Euston, D. R., Tatsuno, M. & McNaughton, B. L. Fast-forward playback of recent
memory sequences in prefrontal cortex during sleep. Science (New York, N.Y.) 318, 1147-
1150, doi:10.1126/science.1148979 (2007).
41
possible
doi:10.1016/j.neuron.2009.03.014 (2009).
42
Based Probabilistic Representations
doi:10.1016/j.neuron.2016.09.038 (2016).
43
Lee, J. H., Whittington, M. A. & Kopell, N. J. Top-down beta rhythms support
selective attention via interlaminar interaction: a model. PLoS computational biology 9,
e1003164, doi:10.1371/journal.pcbi.1003164 (2013).
44
Puccini, G. D., Sanchez-Vives, M. V. & Compte, A. Integrated mechanisms of
anticipation and rate-of-change computations in cortical circuits. PLoS computational
biology 3, e82, doi:10.1371/journal.pcbi.0030082 (2007).
45
Miller, P. & Katz, D. B. Stochastic transitions between neural states in taste
processing and decision-making. J Neurosci 30, 2559-2570, doi:10.1523/jneurosci.3047-
09.2010 (2010).
46
Jones, L. M., Fontanini, A., Sadacca, B. F., Miller, P. & Katz, D. B. Natural stimuli
evoke dynamic sequences of states in sensory cortical ensembles. Proceedings of the
National Academy of Sciences of the United States of America 104, 18772-18777,
doi:0705546104 [pii]
10.1073/pnas.0705546104 (2007).
47
population coding across cortex. Nature, doi:10.1038/nature23020 (2017).
Orban, G., Berkes, P., Fiser, J. & Lengyel, M. Neural Variability and Sampling-
in the Visual Cortex. Neuron 92, 530-543,
Luczak, A., Bartho, P. & Harris, K. D. Spontaneous events outline the realm of
in neocortical populations. Neuron 62, 413-425,
responses
sensory
Mattia, M. et al. Heterogeneous attractor cell assemblies for motor planning in
Runyan, C. A., Piasini, E., Panzeri, S. & Harvey, C. D. Distinct timescales of
15
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Churchland, M. M. et al. Stimulus onset quenches neural variability: a
Murray, J. M. & Escola, G. S. Learning multiple variable-speed sequences in
Grewe, B. F. et al. Neural ensemble dynamics underlying a long-term associative
48
widespread cortical phenomenon. Nature neuroscience 13, 369-378, doi:nn.2501 [pii]
10.1038/nn.2501 (2010).
49
striatum via cortical tutoring. eLife 6, doi:10.7554/eLife.26084 (2017).
50
Stern, M., Bolding, K., Abbott, L. & Franks, K. A Transformation From Latency
To Ensemble Coding In A Model Of Piriform Cortex. bioRxiv, doi:10.1101/118364 (2017).
Sadacca, B. F. et al. The Behavioral Relevance of Cortical Neural Ensemble
51
Responses Emerges Suddenly. The Journal of neuroscience : the official journal of the Society
for Neuroscience 36, 655-669, doi:10.1523/jneurosci.2265-15.2016 (2016).
52
Linster, C. & Fontanini, A. Functional neuromodulation of chemosensation in
vertebrates. Current opinion in neurobiology 29, 82-87, doi:10.1016/j.conb.2014.05.010
(2014).
Liu, H. & Fontanini, A. State Dependency of Chemosensory Coding in the
53
Gustatory Thalamus (VPMpc) of Alert Rats. The Journal of neuroscience : the official journal
of the Society for Neuroscience 35, 15479-15491, doi:10.1523/jneurosci.0839-15.2015 (2015).
54
memory. Nature 543, 670-675, doi:10.1038/nature21682 (2017).
55
47, doi:10.1186/1471-2202-5-47 (2004).
Chow, S. S., Romo, R. & Brody, C. D. Context-dependent modulation of
56
functional connectivity: secondary somatosensory cortex
to prefrontal cortex
connections in two-stimulus-interval discrimination tasks. The Journal of neuroscience : the
official journal of the Society for Neuroscience 29, 7238-7245, doi:10.1523/jneurosci.4856-
08.2009 (2009).
57
Zenke, F., Agnes, E. J. & Gerstner, W. Diverse synaptic plasticity mechanisms
orchestrated to form and retrieve memories in spiking neural networks. Nat Commun 6,
6922, doi:10.1038/ncomms7922 (2015).
58
assemblies
doi:10.1038/ncomms6319 (2014).
59
synaptic input. Neuron 35, 773-782 (2002).
60
Science 354, 1140-1144 (2016).
Litwin-Kumar, A. & Doiron, B. Formation and maintenance of neuronal
5319,
Salinas, E. Context-dependent selection of visuomotor maps. BMC neuroscience 5,
Engel, T. A. et al. Selective modulation of cortical state during spatial attention.
through
synaptic
plasticity. Nature
communications
5,
Chance, F. S., Abbott, L. F. & Reyes, A. D. Gain modulation from background
16
Figures
Fig. 1
Fig. 1: Anticipatory activity requires a clustered network architecture. Effects of
anticipatory cue on stimulus coding in the clustered (a-c) and homogeneous (d-f)
network. a: Schematics of the clustered network architecture and stimulation paradigm.
A recurrent network of inhibitory (red circles) and excitatory neurons (triangles)
arranged in clusters (ellipsoids) with stronger intra-cluster recurrent connectivity. The
network receives bottom-up sensory stimuli targeting random, overlapping subsets of
clusters (selectivity to 4 stimuli is color-coded), and one top-down anticipatory cue
inducing a spatial variance in the cue afferent currents to excitatory neurons. b:
Representative single neuron responses to cue and one stimulus in expected trials in the
clustered network of a). Black tick marks represent spike times (rasters), with PSTH
(mean±s.e.m.) overlaid in pink. Activity marked by horizontal bars was significantly
areas represent mean±s.e.m. across four tastes in 20 simulated sessions; color-dotted
different from baseline (pre-cue activity) and could either be excited (top panel) or
inhibited (bottom) by the cue. c: Time course of cross-validated stimulus-decoding
accuracy in the clustered network. Decoding accuracy increases faster during expected
(pink) than unexpected (blue) trials in clustered networks (curves and color-shaded
17
lines around gray shaded areas represent 95% C.I. from shuffled datasets). A separate
classifier was used for each time bin, and decoding accuracy was assessed via a cross-
validation procedure, yielding a confusion matrix whose diagonal represents the
accuracy of classification for each of four tastes in that time bin (see text and Fig. S3 for
details). Inset: aggregate analysis across simulated sessions of the onset times of
significant decoding in expected (pink) vs. unexpected trials (blue). d: Schematics of the
homogenous network architecture. Sensory stimuli modeled as in a). e: Representative
single neuron responses to cue and one stimulus in expected trials in the homogeneous
network of d), same conventions as in b). f: Cross-validated decoding accuracy in the
homogeneous network (same analysis as in c). The latency of significant decoding in
expected vs. unexpected trials is not significantly different. Inset: aggregate analysis of
onset times of significant decoding (same as inset of c). Panels b, c, e, f: pink and black
horizontal bars, p < 0.05, t-test with multiple-bin Bonferroni correction. Panel c: ** = p <
0.01, t-test. Panel f: n.s.: non-significant.
18
Fig. 2
Fig. 2: Anticipatory cue speeds up network dynamics. a: Raster plots of the clustered
network activity in the absence (left) and in the presence of the anticipatory cue (right),
with no stimulus presentation in either case. The dynamics of cluster activation and
deactivation accelerated proportionally to the increase in afferent currents' variance s2
induced by the cue. Top panels: distribution of cue peak values across excitatory
neurons: left, no cue; right, distribution with S.D. 𝜎=10% in units of baseline current.
decrease when increasing 𝜎. c: Schematics of the effect of the anticipatory cue on
19
Bottom panels: raster plots of representative trials in each condition (black: excitatory
neurons, arranged according to cluster membership; red: inhibitory neurons). b: The
average cluster activation lifetime (left) and inter-activation interval (right) significantly
network dynamics. Top row: the increase in the spatial variance of cue afferent currents
(insets: left: no cue; stronger cues towards the right) flattens the "effective f-I curve"
(sigmoidal curve) around the diagonal representing the identity line (straight line). The
case for a simplified two-cluster network is depicted (see text). States A and B
correspond to stable configurations with only one cluster active; state C corresponds to
an unstable configuration with 2 clusters active. Bottom row: shape of the effective
potential energy corresponding to the f-I curves shown in the top row. The effective
potential energy is defined as the area between the identity line and the effective f-I
curve (shaded areas in top row; see formula). The f-I curve flattening due to the
anticipatory cue shrinks the height D of the effective energy barrier, making cluster
transitions more likely and hence more frequent. d: Effect of the anticipatory cue (in
units of the baseline current) on the height of the effective energy barrier D, calculated
via mean field theory in a reduced two-cluster network of LIF neurons (see Methods).
20
Fig. 3
Fig. 3: Anticipatory cue induces faster onset of stimulus-coding states. a: Raster plots
of representative trials in the expected (left) and unexpected (right) conditions in
response to the same stimulus at t=0. Stimulus-selective clusters (red tick marks, spikes)
activate earlier than non-selective clusters (black tick marks, spikes) in response to the
stimulus when the cue precedes the stimulus. b: Comparison of activation latency of
selective clusters after stimulus presentation during expected (pink) and unexpected
(blue) trials (mean±s.e.m. across 20 simulated sessions). Latency in expected trials is
significantly reduced. c: The effective energy landscape and the modulation induced by
stimulus and anticipatory cue on two-clustered networks, computed via mean field
theory (see Methods). Left panel: after stimulus presentation the stimulus-coding state
(left well in left panel) is more likely to occur than the non-coding state (right well).
Right panel: barrier heights as a function of stimulus strength in expected (`cue ON') and
unexpected trials (`cue OFF'). Stronger stimuli (lighter shades of cyan) decrease the
barrier height D separating the non-coding and the coding state. In expected trials
21
trials (for stimulus ≥4% the barrier vanishes leaving just the coding state). Panel b: ***
(dashed lines), the barrier D is smaller than in unexpected ones (full lines), leading to a
faster transition probability from non-coding to coding states compared to unexpected
= p < 0.001, t-test.
22
Fig. 4
Fig. 4: Anticipation of coding states: model vs. data. a: Representative trials from one
ensemble of 9 simultaneously recorded neurons from clustered network simulations
during expected (left) and unexpected (right) conditions. Top panels: spike rasters with
latent states extracted via a HMM analysis (colored curves represent time course of state
probabilities; colored areas indicate intervals where the probability of a state exceeds
80%; thick horizontal bars atop the rasters mark the presence of a stimulus-coding
state). Bottom panels: Firing rate vectors for each latent state shown in the corresponding
top panel. b: Latency of stimulus-coding states in expected (pink) vs. unexpected (blue)
trials (mean±s.e.m. across 20 simulated datasets). Faster coding latency during expected
trials is observed compared to unexpected trials. c-d: Same as a)-b) for the empirical
datasets. * = p < 0.05, ***= p < 0.001, t-test.
23
Methods
Experimental dataset. The experimental data come from a previously published dataset
Ref.1 (for details, see Supplementary Methods). Experimental procedures were
approved by the Institutional Animal Care and Use Committee of Stony Brook
University and complied with university, state, and federal regulations on the care and
use of laboratory animals.
Ensemble states detection. A Hidden Markov Model (HMM) analysis was used to detect
ensemble states in both the empirical data and model simulations. Here, we give a brief
description of the method used and we refer the reader to Refs. 2-5 for more detailed
information.
The HMM assumes that an ensemble of N simultaneously recorded neurons is in one of
M hidden states at each given time bin. States are firing rate vectors 𝑟5(𝑚), where 𝑖=
1,...,𝑁 is the neuron index and 𝑚=1,...,𝑀 identifies the state. In each state, neurons
were assumed to discharge as stationary Poisson processes (Poisson-HMM) conditional
on the state's firing rates. Trials were segmented in 2 ms bins, and the value of either 1
(spike) or 0 (no spike) was assigned to each bin for each given neuron (Bernoulli
approximation for short time bins); if more than one neuron fired in a given bin (a rare
event), a single spike was randomly assigned to one of the firing neurons. A single
HMM was used to fit all trials in each recording session, resulting in the emission
probabilities 𝑟5(𝑚) and in a set of transition probabilities between the states. Emission
and transition probabilities were calculated with the Baum-Welch algorithm6 with a
fixed number of hidden states M, yielding a maximum likelihood estimate of the
parameters given the observed spike trains. Since the model log-likelihood LL increases
with M, we repeated the HMM fits for increasing values of M until we hit a minimum
of the Bayesian Information Criterion (BIC, see below and Ref.6). For each M, the LL
used in the BIC was the sum over 10 independent HMM fits with random initial
guesses for emission and transition probabilities. This step was needed since the Baum-
Welch algorithm only guarantees reaching a local rather than global maximum of the
likelihood. The model with the lowest BIC (having M* states) was selected as the
winning model, where 𝐵𝐼𝐶=−2𝐿𝐿+[𝑀(𝑀−1)+𝑀𝑁]∙ln 𝑇, T being the number of
observations in each session (= number of trials × number of bins per trials). Finally, the
winning HMM model was used to "decode" the states from the data according to their
posterior probability given the data. During decoding, only those states with
probability exceeding 80% in at least 25 consecutive 2ms-bins were retained (henceforth
denoted simply as "states")3,5. This procedure eliminates states that appear only very
transiently and with low probability, also reducing the chance of overfitting. A median
of 6 states per ensemble was found, varying from 3 to 9 across ensembles.
24
Coding states. In each condition (i.e., expected vs. unexpected), the frequencies of
occurrence of a given state across taste stimuli were compared with a test of proportions
(chi-square, p<0.001 with Bonferroni correction to account for multiple states). When a
significant difference was found across stimuli, a post-hoc Marascuilo test was
performed7. A state whose frequency of occurrence was significantly higher in the
presence of one taste stimulus compared to all other tastes was deemed a 'coding state'
for that stimulus (Fig. 4).
Spiking network model. We modeled the local neural circuit as a recurrent network of N
Networks of different architectures were considered: i) networks with segregated
clusters (referred to as "clustered network," parameters as in Tables 1 and 2); ii)
networks with overlapping clusters (see Suppl. Table S2 and Suppl. Methods for
details), iii) homogeneous networks (parameters as in Table 1). In the clustered network,
leaky-integrate-and-fire (LIF) neurons, with a fraction 𝑛O=80% of excitatory (E) and
𝑛P=20% of inhibitory (I) neurons.8 Connectivity was random with probability 𝑝OO=
0.2 for E to E connections and 𝑝OP=𝑝PO=𝑝PP=0.5 otherwise. Synaptic weights 𝐽5S from
pre-synaptic neuron 𝑗∈𝐸,𝐼 to post-synaptic neuron 𝑖∈𝐸,𝐼 scaled as 𝐽5S=𝑗5S/ 𝑁, with
𝑗5S drawn from normal distributions with mean 𝑗XY (for 𝛼,𝛽=𝐸,𝐼) and 1% SD.
E neurons were arranged in Q clusters with 𝑁\=100 neurons per clusters on average
(1% SD), the remaining fraction 𝑛]^ of E neurons belonging to an unstructured
cluster had intra-cluster synaptic weights potentiated by a factor 𝐽_; synaptic weights
between neurons belonging to different clusters were depressed by a factor 𝐽'=1−
𝛾𝑓(𝐽_−1) <1 with 𝛾=0.5; 𝑓=(1−𝑛]^)/𝑄 is the average number of neurons in each
cluster.8 When changing the network size N, all synaptic weights 𝐽5S were scaled by 𝑁,
the intra-cluster potentiation values were 𝐽_=5, 10, 20, 30, 40 for 𝑁=1,2,4,6,8 ×10e
parameters were kept fixed. In the homogeneous network, there were no clusters (𝐽_=
𝐽'=1).
"background" population. In the clustered network, neurons belonging to the same
neurons, respectively, and cluster size remained unchanged (see also Table 1); all other
𝑑𝑑𝑡𝑉=−𝑉𝜏j +𝐼kl\+𝐼lm2 ,
Model neuron dynamics. Below threshold the LIF neuron membrane potential evolved in
time as
where 𝜏j is the membrane time constant and the input currents I are a sum of a
recurrent contribution 𝐼kl\ coming from the other network neurons and an external
current 𝐼lm2= 𝐼n+𝐼o25j+𝐼\1l (units of Volt/s). Here, 𝐼n is a constant term representing
25
𝐽5S
sStu
𝛿(𝑡−𝑡wS)
,
w
𝜏or6𝑑𝑑𝑡𝐼kl\=−𝐼kl\+
th neuron was a low-pass filter of the incoming spike trains
input from other brain areas; 𝐼o25j and 𝐼\1l represent the incoming stimuli and cue,
respectively (see Stimulation protocols below). When 𝑉 hits threshold 𝑉2pk , a spike is
emitted and 𝑉 is then clamped to the rest value 𝑉klol2 for a refractory period 𝜏klq .
Thresholds were chosen so that the homogeneous network neurons fired at rates 𝑟O=5
spks/s and 𝑟P=7 spks/s. The recurrent contribution to the postsynaptic current to the i-
where 𝜏or6 is the synaptic time constant; 𝐽5S is the recurrent synaptic weights from
presynaptic neuron j to postsynaptic neuron i, and 𝑡wS is the k-th spike time from the j-th
presynaptic neuron. The constant external current was 𝐼n=𝑁lm2𝑝5n𝐽5n𝜈lm2, with 𝑁lm2=
𝑛O𝑁, 𝑝5n=0.2, 𝐽5n=𝑗5n/ 𝑁 with 𝑗On for excitatory and 𝑗Pn for inhibitory neurons (see
Table 1), and 𝑟lm2=7 spks/s. For a detailed mean field theory analysis of the clustered
targeting 50% of neurons in stimulus-selective clusters 𝐼o25j(𝑡)=𝐼n∙𝑟o25j(𝑡), where
𝑟o25j(𝑡) was expressed as a fraction of the baseline external current 𝐼n. Each cluster had a
(expressed as fraction of the baseline current 𝐼n; s=20% unless otherwise specified). The
In both the unexpected and the unexpected conditions, stimulus onset at 𝑡=0 was
followed by a linear increase 𝑟o25j(𝑡) in the stimulus afferent current to stimulus-
selective neurons reaching a value 𝑟jym at 𝑡=1 s (𝑟jym=20% , unless otherwise
𝑟\1l(𝑡) with onset at 𝑡=−0.5s before stimulus presentation.
50% probability of being selective to a given stimulus, thus different stimuli targeted
overlapping sets of clusters. The anticipatory cue, targeting a random 50% subset of E
neurons, was modeled as a double exponential with rise and decay times of 0.2 s and 1
s, respectively, unless otherwise specified; its peak value for each selective neuron was
sampled from a normal distribution with zero mean and standard deviation s
network and a comparison between simulations and mean field theory during ongoing
and stimulus-evoked periods we refer the reader to the Suppl. Methods and Refs.5,9.
Stimulation protocols. Stimuli were modeled as time-varying stimulus afferent currents
cue did not change the mean afferent current but only its spatial (quenched) variance
across neurons.
specified). In the expected condition, stimuli were preceded by the anticipatory cue
Network simulations. All data analyses, model simulations and mean field theory
calculations were performed using custom software written in MATLAB (MathWorks),
and C. Simulations comprised 20 realizations of each network (each one representing a
different experimental session), with 20 trials per stimulus in each of the 2 conditions
26
equations8
'u,
the population and the condition. In the absence of the anticipatory cue, the LIF current-
(1)
is the population firing rate vector (boldface represents
𝑟X=𝐹X𝜇X(𝒓),𝜎X*(𝒓) ,
(unexpected and expected); or 40 trials per session in the condition with "cue-on" and
no stimuli (Fig. 2). Dynamical equations for the LIF neurons were integrated with the
Euler method with 0.1 ms step.
Mean field theory. Mean field theory was used in a simplified network with 2 excitatory
clusters (parameters as in Table 2) using the population density approach10-12: the input
to each neuron was completely characterized by the infinitesimal mean 𝜇X and variance
𝜎X* of the post-synaptic current (see Sec. 2.3 of Supplementary Methods for their
expressions). The network fixed points satisfied the 𝑄+2 self-consistent mean field
where 𝒓= 𝑟u,…,𝑟~,𝑟O]^,𝑟P
vectors). 𝐹X is the current-to-rate function for population 𝛼, which varied depending on
to-rate function 𝐹Xn was used
𝐹Xn𝜇X,𝜎X = 𝜏klq+𝜏j,X 𝜋
𝑒1(cid:129) 1+erf𝑢
(cid:134)(cid:135)(cid:136)(cid:136),(cid:137)
(cid:138)(cid:135)(cid:136)(cid:136),(cid:137)
+𝑎𝑘X , 𝐻lqq,X=(cid:140)(cid:143)(cid:135)(cid:149)(cid:135)(cid:141),(cid:137)'(cid:144)(cid:137)
where 𝛩lqq,X=(cid:140)(cid:141)(cid:142)(cid:143),(cid:137)'(cid:144)(cid:137)
+𝑎𝑘X . Here, 𝑘X= 𝜏or6,X 𝜏j,X , 𝑎=
(cid:150)u ** ~1.03.13,14 In the presence of the anticipatory cue, a modified current-to-rate
(cid:145)(cid:137)
(cid:145)(cid:137)
function 𝐹X\1l was used to capture the cue-induced Gaussian noise in the cue afferent
currents to the cue-selective populations (𝛼=1,…,𝑄):
𝐹X\1l𝜇X,𝜎X = 𝐷𝑧 𝐹Xn𝜇X+ 𝑧 𝜎 𝜇lm2,𝜎X,
where 𝐷𝑧=𝑑𝑧 exp (−(cid:156)(cid:129)*)/ 2𝜋 is the Gaussian measure with zero mean and unit
variance, 𝜇lm2=𝐼n is the baseline afferent current and 𝜎 is the anticipatory cue's SD as
fraction of 𝜇lm2 (Fig. 3d; in Fig. 3c, for illustration purposes we used 𝐹On=0.5(1+tanh).
Fixed points 𝒓∗ of equation (1) were found with Newton's method; the fixed points
evaluated at 𝒓∗ was negative definite. Stability was defined with respect to an
approximate linearized dynamics of the mean 𝑚X and SD 𝑠X of the input currents15
−¥ƒ(cid:137)(cid:144)(cid:137)(𝒓),(cid:145)(cid:137)(cid:129)(𝒓)
¥(cid:145)(cid:137)(cid:129)
27
𝑆XY= u¢(cid:149)£⁄,(cid:137) ¥ƒ(cid:137)(cid:144)(cid:137)(𝒓),(cid:145)(cid:137)(cid:129)(𝒓)
¥k§
were stable (attractors) when the stability matrix
¥(cid:145)(cid:137)(cid:129)¥k§−𝛿XY ,
(2)
𝜏or6,X'j(cid:137)'2 =− 𝑚X+𝜇X(𝒓) ,
'o(cid:137)(cid:129)'2=− 𝑠X*+𝜎X*(𝒓) ,
¢(cid:149)£⁄,(cid:137)*
𝑟X𝑡 =𝐹X𝑚X(𝒓),𝑠X*(𝒓) ,
where 𝜇X and 𝜎X* are the stationary values given in the Suppl. Methods.
(3)
out-of-focus populations
𝑟Y𝑟u,…,𝑟" =𝐹Y[𝜇Y(𝑟u,…,𝑟",𝑟"_u,…,𝑟«),𝜎Y*(𝑟u,…,𝑟",𝑟"_u,…,𝑟«)] ,
Effective mean field theory for the reduced network. The mean field equations (1) for the
P=Q+2 populations may be reduced to a set of effective equations governing the
dynamics of a smaller subset of q<P of populations, henceforth referred to as
populations in focus16. The reduction is achieved by integrating out the remaining P-q
out-of-focus populations. This procedure was used to estimate the energy barrier
separating the two network attractors in Fig. 2d and Fig. 3c. Given a fixed set of values
𝒓=[𝑟u,…,𝑟"] for the in-focus populations, one solves the mean field equations for P-q
for 𝛽=𝑞+1,…,𝑃 to obtain the stable fixed point 𝒓fi𝒓 ≐[𝑟"_u𝒓,…,𝑟«𝒓] of the out-
of-focus populations as functions of the in-focus firing rates 𝒓. Stability of the solution
𝒓fi𝒓 is computed with respect to the stability matrix (2) of the reduced system of P-q
out-of-focus populations. Substituting the values 𝒓fi𝒓 into the fixed-point equations for
the q populations in focus yields a new set of equations relating input rates 𝒓 to
"output" rates 𝒓012:
for 𝛼=1,…,𝑞. The input 𝒓 and output 𝒓012 firing rates of the in-focus populations will
for 𝛼=1,…,𝑞 in-focus populations at the point 𝒓. The fixed points 𝑟X012𝒓∗ =𝑟X∗ of the
the out-of-focus populations attain multiple attractors for a given value of 𝒓, in which
case the set of effective transfer functions 𝐹Xlqq is labeled by the chosen attractor; in our
of 𝒓.
𝑟X012𝒓 =𝐹X[𝜇X(𝒓,𝒓fi𝒓),𝜎X*(𝒓,𝒓fi𝒓)] ,
𝑟X012𝒓 =𝐹Xlqq𝜇X𝒓,𝜎X*𝒓 ,
be different, except at a fixed point of the full system where they coincide. The
correspondence between input and output rates of in-focus populations defines the
effective current-to-rate transfer functions
analysis of the two-clustered network, only one attractor was present for a given value
in-focus equations (4) are fixed points of the entire system. It may occur, in general, that
28
(4)
so that the dynamics can be understood as a motion in an effective potential energy
𝛥=
†
‡
(𝑟−𝑟012𝑟)𝑑𝑟
,
dynamics, while its maxima represent the effective energy barriers between two
minima, as illustrated in Fig. 2c. The one-cluster network has 3 fixed points, two stable
Energy potential. In a network with Q=2 clusters, one can integrate out all populations
(out-of-focus) except one (in-focus) to obtain the effective transfer functions for the in-
Network dynamics can be visualized on a one-dimensional curve, where it is well
approximated by the first-order dynamics (see ref. 16 for details):
𝜏or6,X𝑑 𝑟𝑑𝑡 =−𝑟+𝑟012𝑟.
𝜏or6,X𝑑 𝑟𝑑𝑡 =−𝜕𝐸𝑟𝜕𝑟 ,
focus population representing a single cluster, with firing rate 𝑟 (equation (4) for q=1).
These dynamics can be expressed in terms of an effective energy function 𝐸𝑟 as
landscape, as if driven by an effective force −¥Ok¥k =−(𝑟−𝑟012𝑟). The minima of the
energy with respect to 𝑟 are the stable fixed points of the effective 1-dimensional
attractors ('A' and 'B' in Fig. 2c) and a saddle point ('C'). We estimated the height 𝛥 of
which represents the area between the identity line (𝑦=𝑟) and the effective transfer
function (𝑦=𝑟012(𝑟)) (see Fig. 2c). In the finite network, where the dynamics comprise
frequency that depends on the effective energy barrier Δ, as explained in the main text.
homogeneous networks). Spike counts from all trials of 𝑛o25j taste stimuli in each
Population decoding. The amount of stimulus-related information carried by spike trains
was assessed through a decoding analysis17 (see Fig. S3a for illustration). A multiclass
classifier was constructed from Q neurons sampled from the population (one neuron
from each of the Q clusters for clustered networks, or Q random excitatory neurons for
the potential energy barrier on the trajectory from A to B through C as minus the
integral of the force from the first attractor A to C:
stochastic transitions among the states, switching between A and B would occur with a
condition (expected vs. unexpected) were split into training and test sets for cross-
validation. A "template" was created for the population PSTH for each stimulus,
condition and time bin (200 ms, sliding over in 50 ms steps) in the training set. The
PSTH contained the trial-averaged spike counts of each neuron in each bin (the same
number of trials across stimuli and conditions were used). Population spike counts for
each test trial were classified according to the smallest Euclidean distance from the
templates across 10 training sets ('bagging' or bootstrap aggregating procedure18).
Specifically, from each training set L, we created bootstrapped training sets 𝐿], for 𝑏=
1,..,𝐵=10, by sampling with replacement from L. In each bin, each test trial was then
classified B times using the B classifiers, obtaining B different "votes", and the most
29
frequent vote was chosen as the bagged classification of the test trial. Cross-validated
decoding accuracy in a given bin was defined as the fraction of correctly classified test
trials in that bin.
Significance of decoding accuracy was established via a permutation test: 1000 shuffled
datasets were created by randomly permuting stimulus labels among trials, and a
'shuffled distribution' of 1000 decoding accuracies was obtained. In each bin, decoding
accuracy of the original dataset was deemed significant if it exceeded the upper bound,
𝛼n.n•, of the 95% confidence interval of the shuffled accuracy distribution in that bin
(this included a Bonferroni correction for multiple bins, so that 𝛼n.n•=1−0.05/𝑁],
with 𝑁] the number of bins). Decoding latency (insets in Figs. 1c and 1f) was estimated
interval, and latency; see Figs. 2 and 3), cluster spike count vectors 𝑟5 (for 𝑖=1,...,𝑄) in
as the earliest bin with significant decoding accuracy.
Cluster dynamics. To analyze the dynamics of neural clusters (lifetime, inter-activation
5 ms bins were obtained by averaging spike counts of neurons belonging to a given
cluster. A cluster was deemed active if its firing rate exceeded 10 spks/s. This threshold
was chosen so as to lie between the inactive and active clusters' firing rates, which were
obtained from a mean field solution of the network5.
Code and data availability statement. Experimental datasets are available from the
corresponding authors upon request. All data analysis and network simulation scripts
are available from the authors upon request.
Symbol
𝑗OO
𝑗OP
𝑗PO
𝑗PP
𝑗On
𝑗Pn
𝐽_
𝑟lm2O
𝑟lm2P
𝑉2pkO
𝑉2pkP
𝑉klol2
tm
tref
tsyn
nbg
Nc
Description
Mean EàE synaptic weight × 𝑁
Mean IàE synaptic weight × 𝑁
Mean EàI synaptic weight × 𝑁
Mean IàI synaptic weight × 𝑁
Mean afferent synaptic weights to E neurons × 𝑁
Mean afferent synaptic weights to I neurons × 𝑁
Potentiated intra-cluster EàE weights factor
Average afferent rate to E neurons (baseline)
Average afferent rate to I neurons (baseline)
E neuron threshold potential
I neuron threshold potential
E and I neurons reset potential
E and I membrane time constant
Absolute refractory period
E and I synaptic time constant
Background neurons fraction
Average cluster size
30
Value
1.1 mV
5.0 mV
1.4 mV
6.7 mV
5.8 mV
5.2 mV
7 spks/s
7 spks/s
3.9 mV
4.0 mV
0 mV
see caption
20 ms
5 ms
4 ms
10%
100 neurons
Symbol
Description
Table 1: Parameters for the clustered and homogeneous networks with N LIF neurons.
In the clustered network, the intra-cluster potentiation parameter values were 𝐽_= 5, 10,
20, 30, 40 for networks with 𝑁=1,2,4,6,8 ×10e neurons, respectively. In the
homogeneous network, 𝐽_=1.
𝑗OO
𝑗OP
𝑗PO
𝑗PP
𝑗On
𝑗Pn
𝐽_
𝑉2pkO
𝑉2pkP
Mean EàE synaptic weight × 𝑁
Mean IàE synaptic weight × 𝑁
Mean EàI synaptic weight × 𝑁
Mean IàI synaptic weight × 𝑁
Mean afferent synaptic weights to E neurons × 𝑁
Mean afferent synaptic weights to I neurons × 𝑁
0.8 mV
10.6 mV
2.5 mV
9.7 mV
14.5 mV
12.9 mV
4.6 mV
8.7 mV
Potentiated intra-cluster EàE weights factor
E neuron threshold potential
I neuron threshold potential
Background neurons fraction
Value
65%
nbg
9
Table 2: Parameters for the simplified 2-cluster network with N=800 LIF neurons (the
remaining parameters were as in Table 1).
of
on
74,
cortical
processing
taste. Neuron
Abeles, M. et al. Cortical activity flips among quasi-stationary states. Proc Natl
Samuelsen, C. L., Gardner, M. P. & Fontanini, A. Effects of cue-triggered
410-422,
1
expectation
doi:10.1016/j.neuron.2012.02.031 (2012).
2
Acad Sci USA 92, 8616-8620 (1995).
3
Jones, L. M., Fontanini, A., Sadacca, B. F., Miller, P. & Katz, D. B. Natural stimuli
evoke dynamic sequences of states in sensory cortical ensembles. Proceedings of the
National Academy of Sciences of the United States of America 104, 18772-18777 (2007).
4
Ponce-Alvarez, A., Nacher, V., Luna, R., Riehle, A. & Romo, R. Dynamics of
cortical neuronal ensembles transit from decision making to storage for later report. The
Journal of neuroscience : the official journal of the Society for Neuroscience 32, 11956-11969,
doi:10.1523/jneurosci.6176-11.2012 (2012).
5
during Ongoing and Evoked Cortical Activity.
doi:10.1523/JNEUROSCI.4819-14.2015 (2015).
Mazzucato, L., Fontanini, A. & La Camera, G. Dynamics of Multistable States
J Neurosci 35, 8214-8231,
31
References
•
•
•
•
•
Lansky, P. & Sato, S. The stochastic diffusion models of nerve membrane
Zucchini, W. & MacDonald, I. L. Hidden Markov models for time series: an
Tuckwell, H. C. Introduction to theoretical neurobiology. Vol. 2 (Cambridge
6
introduction using R. (CRC Press, 2009).
7
La Camera, G. & Richmond, B. J. Modeling the violation of reward maximization
and invariance in reinforcement schedules. PLoS computational biology 4, e1000131,
doi:10.1371/journal.pcbi.1000131 (2008).
8
Amit, D. J. & Brunel, N. Model of global spontaneous activity and local
structured activity during delay periods in the cerebral cortex. Cerebral cortex (New York,
N.Y. : 1991) 7, 237-252 (1997).
9
Mazzucato, L., Fontanini, A. & La Camera, G. Stimuli Reduce the Dimensionality
of Cortical Activity. Frontiers in systems neuroscience 10, 11, doi:10.3389/fnsys.2016.00011
(2016).
10
University Press, 1988).
11
depolarization and interspike interval generation. J Peripher Nerv Syst 4, 27-42 (1999).
12
Richardson, M. J. The effects of synaptic conductance on the voltage distribution
and firing rate of spiking neurons. Physical review. E, Statistical, nonlinear, and soft matter
physics 69, 051,918 (2004).
13
synaptic
dynamics.
doi:10.1006/jtbi.1998.0782 (1998).
14
Fourcaud, N. & Brunel, N. Dynamics of the firing probability of noisy integrate-
and-fire neurons. Neural computation 14, 2057-2110, doi:10.1162/089976602320264015
(2002).
15
La Camera, G., Rauch, A., Luscher, H. R., Senn, W. & Fusi, S. Minimal models of
adapted neuronal response to in vivo-like input currents. Neural Comput 16, 2101-2124
(2004).
16
population states. Network (Bristol, England) 10, 351-373 (1999).
17
Nature 497, 585-590, doi:10.1038/nature12160 (2013).
18
Brunel, N. & Sergi, S. Firing frequency of leaky intergrate-and-fire neurons with
87-95,
Rigotti, M. et al. The importance of mixed selectivity in complex cognitive tasks.
Mascaro, M. & Amit, D. J. Effective neural response function for collective
Breiman, L. Bagging predictors. Machine learning 24, 123-140 (1996).
current
Journal
of
theoretical
biology
195,
•
•
•
•
•
•
•
•
•
•
•
•
•
32
Expectation-induced modulation of metastable
activity underlies faster coding of sensory stimuli
Supplementary Material
Luca Mazzucato1,2 Giancarlo La Camera*1,3 and Alfredo Fontanini*1,3
1Department of Neurobiology and Behavior, State University of New York at Stony Brook, Stony Brook, NY 11794
2Departments of Biology and Mathematics and Institute of Neuroscience, University of Oregon, Eugene, OR 97403
3Graduate Program in Neuroscience, State University of New York at Stony Brook, Stony Brook, NY 11794
1 Supplementary Results
1.1 Cue responses in data vs. model
1.2 Anticipation of stimulus decoding
1.3 Robustness of anticipatory activity
1.4 Distracting cue model
1.5 Specificity of anticipatory mechanism (alternative models)
1.6 Firing rate coding of expectation
2 Supplementary Methods
2.1 Behavioral training and electrophysiology
2.2 Cue responsiveness
2.3
2.4 Network with overlapping clusters
Infinitesimal moments for mean field theory
1
1
2
3
6
6
10
11
11
11
12
13
Supplementary Results
1.1 Cue responses in data vs. model
We compared the cue responses of the clustered network to those in the dataset published in Ref. [1]. The
entity of the response was estimated by the ∆PSTH with the procedure described in the Suppl. Methods,
Sec. 2.2. In the data [1], the PSTH modulation in cue-responsive neurons was consistent across modalities
(excited responses peaked at 12±1.0 spks/s, inhibited responses peaked at −4.8 ± 0.3 spks/s; see Fig. S1a).
In the model, modulation of cue responses depended on the cue-induced modulation of the spatial variance
σ2 of the afferent currents (expressed in percent of baseline in Fig. S1b). As shown in the figure, the PSTH
modulation of cue-responsive neurons slightly increased with σ, however, over a wide range of parameters, was
in quantitative agreement with the data shown in Fig. S1a.
In Fig. 1a -- c of the main text we chose σ = 20%, giving peak firing rate modulations of 11.1 ± 0.8 spks/s
(excited) and −6.2 ± 0.3 spks/s (inhibited). Single neurons' responses thus reflected responses observed in the
empirical data (Fig. S1a and refs. [1, 2]). In the homogeneous network of Fig. 1d -- f, single neurons' responses
were comparable to the ones observed in the clustered network. With σ = 20%, peak firing rate modulations
were 11.2±1.2 spks/s (excited) and −3.9±0.2 spks/s (inhibited). Both in the clustered and in the homogeneous
networks, neurons could be nonresponsive to the cue (Fig. S2).
-- 1 --
Fig. S1. Cue responses in data
and model.
Peak firing rate re-
sponses of cue-excited (green) and
cue-inhibited (red) neurons. a: Data
from different sensory modalities to-
gether with across-modalities aver-
age (rightmost bar), adapted from
[1]. b: Clustered network model
peak firing rate responses for dif-
ferent values of cue-induced spatial
variance. Peak responses (∆PSTH,
mean±s.e.m.) were computed in
both panels as the difference between
the activity post-cue and the activity
pre-cue. All panels: t-test, n.s.=non-
significant. "Data" in panel b is the
same as "Average" in panel a.
Fig.
S2. Model neurons
with no cue response. Raster
plots of representative single
neuron responses to cue and
one stimulus in expected tri-
als in the clustered (a) and ho-
mogeneous (b) networks (pink
curve: PSTH (mean±s.e.m.).
Top row: non-responsive neu-
rons. Bottom row: neurons
that are stimulus-responsive
but not cue-responsive accord-
ing to whether or not
the
PSTH was significantly differ-
ent from baseline (pink hori-
zontal bars: p < 0.05, t-test
with multiple-bin correction).
1.2 Anticipation of stimulus decoding
In the main text we have shown that stimulus decoding is accelerated in the presence of the anticipatory cue (for
example, in Fig. 1c). As a measure of performance we used the decoding accuracy averaged across all tastes.
Here we show that similar results are obtained when considering the decoding performance for single tastes
separately. In either case, the decoding procedure is described in Methods (Sec. Population decoding) and is
summarized in Fig. S3a.
As shown in figure, decoding was performed using a multi-class classifier where each class corresponded
to one of the four taste stimuli. A separate classifier was used for each time bin, and decoding performance
in each bin was assessed via a cross-validation procedure (see Methods for details). This procedure yields a
confusion matrix whose diagonal elements represent the accuracy of classification for each taste in that time
bin (Fig. S3a). The performance for each time bin can be averaged across instances of the same taste (as in the
-- 2 --
Fig. S3. Multi-class classifier for stimulus decoding. a: Schematics of the classification algorithm used for population
decoding. The colored curves represent the temporal evolution of neural activity across N simultaneously recorded neurons.
The four colors label trajectories obtained with four different stimuli (tastants). Hues help visualize the time course within
each trajectory. Each time bin of activity (black dots) was decoded using an independent classifier. Decoding performance
was assessed via a cross-validation procedure, yielding a confusion matrix whose diagonal represents the accuracy of
classification for each taste in that time bin. b: Time course of decoding accuracy in expected vs. unexpected conditions
for each individual tastant (left) and for the across-taste average (right). The left panel demonstrates coding anticipation for
each tastant separately. Color-coded horizontal bar represents significant difference between decoding accuracy in expected
vs. unexpected trials, p < 0.05, t-test with multiple bin Bonferroni correction (notations as in Fig. 1c of the main text).
left panel of Fig. S3b) or across all tastes simultaneously (right panel of Fig. S3b). In either case, significant
decoding accuracy is anticipated in expected trials compared to unexpected trials.
1.3 Robustness of anticipatory activity
To test the robustness of anticipatory activity in the clustered network, we systematically varied key parameters
related to the sensory and anticipatory inputs, as well as network connectivity and architecture (Fig. S4-5).
Increasing stimulus intensity led to a faster encoding of the stimulus in both conditions (Fig. S4a). Stimuli were
always decoded faster when preceded by the cue (Fig. S4a; two -- way ANOVA with factors 'stimulus slope',
p< 10−18, F (4) = 30.4, and 'condition' (expected vs. unexpected), p< 10−15, F (1) = 79.8). The amount
of anticipatory activity induced by the cue depended on stimulus intensity only weakly (p(interaction)= 0.05,
F (4) = 2.4), and was present even in the case of step -- like stimulus (Fig. S4b). Moreover, we found that
anticipatory activity was maintained when we increased the number of presented stimuli, while keeping a fixed
50% probability that each cluster be selective to a given stimulus (Fig. S4c; two-way ANOVA with factors
'number of stimuli' (p = 0.57, F (3) = 0.67) and 'condition' (p < 10−7, F (1) = 35.3; p(interaction)= 0.66,
F (3) = 0.54). Anticipatory activity was present also when the stimulus selectivity targeted both E and I
neurons, rather than E neurons only (Fig. S4d; 50% of both E and I neurons were targeted by the cue).
We then tested robustness of anticipatory activity to variations in network size and architecture. We esti-
mated the decoding accuracy in ensembles of constant size (20 neurons) sampled from networks of increasing
size (Fig. S5a). When scaling the network size we kept fixed both the clusters' size and the probability that
each cluster was selective to a given stimulus (50%). Network synaptic weights scaled as reported in Table 1
and 2 of Methods. Cue-induced anticipation was even more pronounced in larger networks (Fig. S5a, two --
way ANOVA with factors 'network size' (p < 10−20, F (3) = 49.5), 'condition' (p < 10−16, F (1) = 90;
p(interaction)= 10−10, F (3) = 20).
-- 3 --
Fig. S4. Robustness of anticipation
effects to variations in stimulus
model. a: Stimulus intensity (slope).
Latency of significant decoding in-
creased with stimulus intensity in
both conditions (top, darker shades
represent stronger stimuli; bottom,
decoding latency, mean±s.e.m.), and
it is faster in expected (pink) than
unexpected trials (blue). b: Step-
like stimuli. Anticipatory activity is
present in the case of step-like stim-
uli. Bottom panel: time course of
decoding accuracy (notations as in
Fig. 1c of the main text). c: Number
of stimuli. Anticipatory activity dis
not depend on the number of stimuli
presented to the network. d: Stim-
uli targeting inhibitory neurons. An-
ticipatory activity was present when
stimuli targeted both E and I neu-
rons. Notations as in bottom row of
panel a. Panels a-d: ∗ = p < 0.05,
∗∗ = p < 0.01, ∗∗∗ = p < 0.001,
post-hoc multiple -- comparison t -- test
with Bonferroni correction. Hori-
zontal black bar, p < 0.05, t -- test
with multiple -- bin Bonferroni correc-
Insets: ∗∗∗ = p < 0.001,
tion.
signed -- rank test.
In our main model the neural clusters were segregated in disjoint groups. In Fig. S5b, we investigated an
alternative scenario where neurons may belong to multiple clusters, resulting in an architecture with overlapping
Fig. S5. Robustness of anticipatory activity to variations in network architecture. a: Variations in network size.
Anticipatory activity was present for a large range of network sizes (J+ = 5, 10, 20, 30, 40 for N = 1, 2, 4, 6, 8 × 103
neurons, respectively). b: Network with overlapping clusters. Anticipatory activity was present also when neurons were
shared among multiple clusters (see the text for detail, notations as in Fig. 1c of the main text). Panels a-b: ∗∗ = p < 0.01,
∗∗∗ = p < 0.001, post-hoc multiple -- comparison t -- test with Bonferroni correction. Horizontal black bar, p < 0.05, t -- test
with multiple -- bin Bonferroni correction. Inset: ∗∗ = p < 0.01, signed -- rank test.
-- 4 --
Fig. S6. Robustness of anticipatory activity to variations in cue model. a: Robustness to variations in spatial variance.
Increasing the cue-induced spatial variance in the afferent currents σ (top: distribution of peak afferent currents across
neurons) leads to more pronounced anticipatory activity (bottom, latency in unexpected (red) and expected (pink) trials).
b: Robustness to variations in kinetics (time course of cue stimulation). Anticipatory activity is present for a large range
of cue time courses (top, double exponential cue profile with rise and decay times [τ1, τ2] = g × [0.1, 0.5] s, for g in
the range from 1 to 3; bottom, decoding latency during unexpected, red, and unexpected, pink, trials). c: Robustness to
step -- like kinetics. This example shows the case of step -- like time course for the cue with spatial variance σ = 20%. d:
Robustness to inclusion of inhibitory neurons as targets. Anticipatory activity was also present when the cue targets 50%
of E and I neurons (σ = 20% in baseline units). Panels a-d: ∗ = p < 0.05, ∗∗ = p < 0.01, ∗∗∗ = p < 0.001, post-hoc
multiple -- comparison t -- test with Bonferroni correction. Horizontal black bar, p < 0.05, t -- test with multiple -- bin Bonferroni
correction; Insets: ∗∗ = p < 0.01, ∗∗∗ = p < 0.001, signed -- rank test. Panels c-d: notations as in Fig. 1c of the main text.
cluster membership [3]. We found that anticipatory activity was present also in networks with overlapping
clusters (Fig. S5b; see Sec. 2.4 of Supplementary Methods below for details of this model).
-- 5 --
The results so far show that anticipatory activity was robust to changes in network architecture. Next, we
show that anticipatory activity was robust to variations in cue parameters -- specifically, (i) variations in the
spatial variance σ2 of the cue-induced afferent currents, (ii) variations in the kinetics of cue stimulation, and
(iii) variations in the type of neurons targeted by the cue.
(i) We found that coding anticipation was present for all values of σ above 10% (Fig. S6a, two -- way
ANOVA with factors 'cue variance' (p = 0.09, F (7) = 1.8) and 'condition' (p < 10−7, F (1) = 31);
p(interaction)= 0.07, F (7) = 1.9).
(ii) Anticipatory activity was also robust to variations in the time course of the cue-evoked currents (Fig.
S6b, two -- way ANOVA with factors 'time course' (p = 0.03, F (4) = 2.8) and 'condition' (p < 10−15,
F (1) = 72.7; p(interaction)= 0.12, F (4) = 1.8). We also considered a model with constant cue-evoked
currents (step-like model), to further investigate a potential role (if any) of the cue's variable time course on
anticipatory activity (Fig. S6c). Again, we found anticipatory coding also in this case (Fig. S6c, bottom panel).
(iii) Finally, we tested whether or not anticipation was present if I neurons, in addition to E neurons, were
targeted by the cue (while maintaining stimulus selectivity for E neurons only). We assumed that 50% of both
E and I neurons were targeted by the cue, and found robust anticipatory activity also in that case (Fig. S6d).
We did not find any anticipatory coding in the homogeneous network model as a result of the same manip-
ulations of Fig. S4-S5 (not shown).
1.4 Distracting cue model
To show that the anticipatory effect of our model cue is specific, we give here an example of a manipulation
leading to the opposite effect. Specifically, if the cue is modeled as recruiting recurrent inhibition rather than
excitation (we refer to such a cue as a 'distractor'), stimulus decoding is slowed down rather than accelerated,
as shown in Fig. S7a-b. The coding delay following such 'distracting cue' is due to a phenomenon of delayed
dynamics which mirrors the anticipatory dynamics of the main model, as shown next.
First, from the analysis of model simulations we found that the activation latency of stimulus-selective
clusters after stimulus presentation was longer during distracted trials (Fig. S7c), which resulted in the delayed
onset of the coding states (see the main text for the definition of coding state in section Anticipatory cue induces
faster onset of taste-coding states). This, in turn, was the result of higher energy barriers between coding and
non-coding states, as confirmed in a mean field analysis of a simplified 2-cluster network (Fig. S7d). Except
for the different cue targets, the 2-cluster network is the same as used to study anticipatory coding (main text,
Anticipatory cue speeds up the networks dynamics). The distracting cue increased the energy barriers for a
range of stimulus intensities (Fig. S7d, right panel).
1.5 Specificity of anticipatory mechanism (alternative models)
How specific is our model of anticipatory activity? To answer this question, we compared our model to several
alternative models where the anticipatory cue modulates, respectively: the feedforward couplings Jext; the
recurrent couplings JEE; and the background synaptic input.
All models had the same architecture of the main clustered network model and the same cue temporal
profile, which was modeled as a double exponential with rise and decay times of 0.2s and 1s, respectively. In
the alternative models, the cue targeted all clustered E neurons, rather than just 50% as in the main model of
Fig. 1a, to avoid the introduction of spatial variance across neurons (this could confound the results with those
-- 6 --
Fig. S7. A "distracting" cue slows down stimulus coding (model). a: Schematics of clustered network architecture and
stimulation (notations as in Fig. 1a of the main text). Unlike the main model, here the cue targets the inhibitory neurons.
b: Time course of cross -- validated population decoding during distracted (brown) trials was slower than during unexpected
(red) trials (notations as in Fig. 1c). Inset: aggregate analysis across 20 simulated sessions (mean±s.e.m.). c: Activation
latency of stimulus-selective clusters after stimulus presentation was delayed during distracted trials (latency, mean±s.e.m.).
d: Mean field theory of simplified 2-cluster network (notations as in Fig. 3c, lighter brown denotes stronger stimuli). Left
panel: the transition probability from the non-coding (right well) to the coding state (left well) increased with larger stimuli.
In distracted trials (dashed curves) the barrier height ∆ from the non-coding to the coding state is larger than in unexpected
trials (full curves), leading to slower coding in the distracted condition. Right panel: effective energy barriers as a function
of stimulus intensity, with (full lines) and without the cue (dashed). Panels b and c: ∗∗ = p < 0.01, ∗∗∗ = p < 0.001,
t-test.
of the main model). All models were scored on the ability to match the experimental data on the ∆PSTH of
cue responses and coding anticipation (see Fig. S8 and Table 1). The performance of the model of the main
text is reported in Fig. S8a for comparison.
Feedforward coupling model (Fig. S8b). In this model, The cue was a time dependent modulation of the
external synaptic coupling Jext = JE0, identical for all clustered excitatory neurons. Coding anticipation was
absent for either moderate or strong positive modulations varying from 10% to 20% above baseline. Cue re-
sponses were heterogeneous but to a different degree than the experimental data. For negative cue modulations,
only inhibited cue responses were observed and no coding anticipation was present (not shown).
Recurrent coupling model (Fig. S8c).
In this model, the cue was a time-dependent modulation of the
recurrent coupling strength JEE between E neurons. For negative modulation of the recurrent couplings (the
more likely to produce a faster dynamics), coding anticipation was present, however, peak cue responses were
-- 7 --
Fig. S8. Alternative models. a: Performance results of the main model ('spatial variance model' of the main text, see Fig.
1a-c), reported here for comparison. Top row: Temporal profile modeling the presentation of the cue (targeting excitatory
neurons). Bottom row, left: ∆PSTH, defined as the peak cue response minus baseline rate, in data and model (excitatory
responses in green, inhibited responses in red). Bottom row, right: Decoding performance in expected (pink) and unexpected
(blue) conditions averaged across 20 simulated sessions. b: Feedforward coupling model. In this model, the cue elicited
a modulation of the feedforward couplings Jext (see text). Same notations as in a. c: Recurrent coupling model. In this
model, the cue elicited a modulation of all excitatory recurrent couplings JEE (see text). Same notations as in a. All panels:
Network parameters as in Table 1. t-test, ∗ = p < 0.05, ∗∗ = p < 0.01, ∗∗∗ = p < 0.001.
strongly inhibited over a wide range of parameters, and thus incompatible with the empirical data. For positive
modulation (not shown), coding anticipation was absent and cue responses were mostly excited.
ext → Xσ2
Background synaptic input model (Fig. S9). In this model, the cue induced an increase in the background
synaptic input, which in turn drove a simultaneous increase in background noise and shunting inhibition as
in ref.
[4]. The baseline noise level was modeled after an Ornstein-Uhlenbeck process with zero mean and
variance σext = 0.5rext, where rext was the mean afferent current. The cue increased the background noise
ext while shunting the membrane time constant by a factor 1/X: τm → τm/X
by a factor X: σ2
(we refer to [4] for details). Both modulations followed the same double exponential time course of the spatial
variance model of the main text. With even a moderate factor of X = 1.5, the cue induced mostly inhibited cue
responses and no anticipation (Fig. S9a), due to the strong shunting effect on the network dynamics, leading to
a strong reduction of excitability in the clusters. At X = 2, excitatory neurons were transiently silenced (not
shown). In an effort to obtain a fair comparison with the spatial variance model, we reduced the shunting effect
(using peak value τm/X , with = 1/2, 1/4) while keeping the same amount of background noise (Fig. S9b).
In this case, the cue induced heterogeneous cue responses more similar to the data, but still no anticipation
of coding. We explored the parameter space by independently scaling background noise, shunting, and mean
afferent current rext (transiently increased with the same time course as the other quantities -- to counteract the
shunting effect), but found that coding anticipation was never present (Fig. S9c, bottom row). We concluded
that a model cue inducing an increase in background synaptic activity did not lead to anticipatory activity.
In summary, while our spatial variance model produces robust coding anticipation and quantitatively
matches cue responses in a wide range of parameters and architectures (Fig. S1, S4, S5, S6), all the poten-
tially alternative models above failed to capture the empirical results and reproduce key features of the data (see
Table 1 for a summary).
-- 8 --
Fig. S9. Background synaptic
input model. a: Main findings for
the alternative model wherein the
cue elicits a change in background
synaptic input to E neurons (see
the text for details). Same nota-
tions as in Fig. 8a, with X: amount
of cue-induced modulation ('scal-
ing parameter'); rext: mean affer-
ent current to the network; σext:
background noise input to the net-
work; τm: E neurons membrane
time constant. b: Model with re-
duced shunting inhibition. Same
as a, with the membrane time con-
stant scaled by a lower power of
c: Effect of different scal-
X.
ing regimes on the performance
of the background synaptic in-
put model of panel a. Top row:
ext →
∆PSTH. Key: Black, σ2
ext, rext → rext, τm →
Xσ2
τm/X (same as panel a); Dark
ext, rext →
grey, σ2
X 1/2rext, τm → τm/X; Light
ext, rext →
grey, σ2
Xrext, τm → τm/X; Dark
ext, rext →
brown: σ2
rext, τm → τm/X 1/2; Light
ext, rext →
brown: σ2
rext, τm → τm/X 1/4. Right-
most bar: ∆PSTH of data. Bottom
row: decoding latency as a func-
tion of the scaling parameter X for
the 5 scaling regimes shown in top
row (same color code). Dashed
line: coding latency in the absence
of the cue. All panels: t-test, ∗ =
p < 0.05, ∗∗ = p < 0.01, ∗∗∗ =
p < 0.001.
ext → Xσ2
ext → Xσ2
ext → Xσ2
ext → Xσ2
Model
Fwd. Jext Rec. JEE Backgr. Syn. Input
Yes (fine tuned)
No
S9
Modulation
Anticipation
Cue responses
Figure
Spatial Var.
Yes (robust)
Yes (robust)
S1, S4, S5, S6
No
No
S8b
Yes
No
S8c
Table. S1. Comparison of the performance of alternative models of the anticipatory cue (see Figs. S8, S9). First row ('Mod-
ulation') reports type of modulation due to the cue (feature defining the model; see the text). Second row ('Anticipation')
reports the presence (Yes) or absence (No) of coding anticipation. Third row ('Cue responses') reports whether the ∆PSTH
is compatible (Yes) or not (No) with the empirical data. Last row: figure where the corresponding result is shown.
-- 9 --
1.6 Firing rate coding of expectation
Stimulus coding could accelerate if the cue increased the firing rate of the stimulus-selective neurons compared
to the non-selective neurons. Here we show that coding anticipation in the clustered network is not driven by
such cue-induced changes in firing rate selectivity. To prove this point, we estimated the firing rate difference
∆r between stimulus-selective and nonselective neurons in the expected and unexpected conditions in the
clustered network, and found no difference in ∆r (Fig. 10a).
Similarly, the finding that no coding acceleration is found in the homogeneous network (Fig. 1f of main
text), suggests either no difference in ∆r, or a difference in the opposite direction (i.e., larger ∆r in non-
expected trials). An analysis of ∆r in the homogeneous network (Fig. 10b) confirmed that ∆r was larger in the
unexpected condition, in agreement with the reversed trend in coding speed found in Fig. 1f.
Fig. S10. Time course of the firing rate difference ∆r between stimulus-selective and nonselective clusters in the expected
(pink) and unexpected (blue) conditions. a: Clustered network of Fig. 1a of main text. ∆r is not significantly different be-
tween expected and unexpected trials. b: Homogeneous network (Fig. 1d of main text). In this network, ∆r is significantly
larger in the unexpected condition compared to the expected one (black horizontal bar: p< 0.05, t-test with multiple bin
Bonferroni correction). Both panels: main curves represent means of ∆r over 20 sessions; shaded area represents s.e.m..
-- 10 --
Supplementary Methods
2.1 Behavioral training and electrophysiology
Adult female Long -- Evans rats were used in the experiment. Movable bundles of 16 microwires attached to a
"mini -- microdrive" were implanted bilaterally in the gustatory cortex and intraoral cannulae (IOC) were inserted
bilaterally and cemented. All experimental procedures were approved by the Institutional Animal Care and Use
Committee of Stony Brook University and complied with university, state, and federal regulations on the care
and use of laboratory animals (for more details, see [2]). After postsurgical recovery, rats were trained to self --
administer fluids through IOCs by pressing a lever under head-restraint within 3s presentation of an auditory
cue ("expected trials"; a 75 dB pure tone at a frequency of 5 kHz). The inter-trial interval was progressively
increased to 40 ± 3s. Early presses were discouraged by the addition of a 2s delay to the inter-trial interval.
During training and electrophysiology sessions, additional tastants were automatically delivered through the
IOC at random times near the middle of the inter -- trial interval and in the absence of the anticipatory cue
("unexpected trials"). The following tastants were delivered: 100 mM NaCl, 100 mM sucrose, 100 mM citric
acid, and 1 mM quinine HCl. Water (50 µl) was delivered to rinse the mouth clean through a second IOC, 5s
after the delivery of each tastant.
Multiple single -- unit action potentials were amplified, bandpass filtered, and digitally recorded. Single
units were isolated using a template algorithm, clustering techniques, and examination of inter -- spike interval
plots (Offline Sorter, Plexon). Starting from a pool of 299 single neurons in 37 sessions, neurons with peak
firing rate lower than 1 spks/s (defined as silent) were excluded from further analysis, as well as neurons with
a large peak around the 6 -- 10 Hz in the spike power spectrum, which were classified as somatosensory [5 -- 7].
Only ensembles with 5 or more simultaneously recorded neurons were included in the rest of the analyses.
Ongoing and evoked activity were defined as occurring in the 5 seconds-long interval before or after taste
delivery, respectively.
2.2 Cue responsiveness
A neuron was deemed responsive to the cue if a sudden change in its firing rate was observed during the post-cue
interval. To detect changes in firing rates we used the 'change-point' (CP) procedure described in ref. [6] (an
adaptation of a method presented in ref. [8]). Briefly, we built the cumulative distribution function of the spike
count (CumSC) across all trials in each session, in the interval starting 0.5s before and ending 1 s after the cue
delivery. We then ran the CP detection algorithm on CumSC record with a given tolerance level corresponding
to a desired confidence level of p = 0.05 for a two-sample problem in spike counts. This algorithm looks for
putative CPs as those points xC such that CumSC(xC) is located at the maximal distance from the straight
line joining the first and any subsequent point xn along the record. If the spike counts before xC and between
xC and xn are significantly different (according to the chosen tolerance level), xC is accepted as a legitimate
change point. Normally, one would at this point repeat the procedure considering xC as the new start of the
record, to look for additional CPs; however, we stopped the procedure after finding the first CP (if any), as this
was sufficient to claim responsiveness. If, with this method, a CP was detected before cue presentation, the
whole algorithm was repeated with lower tolerance (lower p-value) to enforce a more stringent criterion. If a
legitimate CP was found anywhere within 1s after cue presentation, the neuron was deemed responsive. If no
CP was found, the neuron was deemed not responsive. For neurons with excited (inhibited) cue responses, the
positive (negative) peak response was estimated as the difference between the post-cue activity and the mean
-- 11 --
baseline activity in the 0.5s preceding cue presentation. This measure is reported as the ∆PSTH in e.g. Sec. 1.1
and Fig. S1.
2.3
Infinitesimal moments for mean field theory
Here we report the formulae for the infinitesimal mean and variance of the input current needed for mean field
theory (see Methods, section Mean field theory).
We need the infinitesimal moments for Q + 2 neural populations: the first Q populations representing the
Q excitatory clusters; the (Q + 1) -- th population representing the "background" unstructured excitatory popu-
lation; and the (Q + 2) -- th population representing the inhibitory population. The theory can be derived from a
population density approach where the input to each neuron in population α is completely characterized by the
infinitesimal mean µα and variance σ2
α of the post-synaptic current (see Table 1 of Methods for parameter val-
ues). The mean and variance of the infinitesimal moments of the input current to the αth excitatory population
is
+ nE(1 − f )pEEJ−jEEr(bg)
E
pEEJ−jEErβ
√
N
µα = τm,E
Q
pEEJ+jEErα +
(cid:16) nEf
Q−1(cid:88)
(cid:17)
pEE(J+jEE)2(1 + δ2)rα +
β=1
,
(cid:16) nEf
Q
−nI pEI jEI rI + nEpE0jE0rext
σ2
α = τm,E
Q−1(cid:88)
β=1
pEE(J−jEE)2(1 + δ2)rβ
(cid:17)
+nE(1 − f )pEE(J−jEE)2(1 + δ2)r(bg)
E + nI pEI j2
EI (1 + δ2)rI
,
where rα, rβ, with α, β = 1, ..., Q, are the firing rates of the excitatory clusters; r(bg)
is the firing rate of the
background E population; rI is the firing rate of the I population; nE = 4/5 and nI = 1/5 are the fractions
of excitatory and inhibitory neurons in the network, respectively, and δ = 1% is the SD of the distribution of
synaptic weights. The infinitesimal moments of the afferent current to a neuron belonging to the background E
population read
E
µ(bg)
E = τm,E
√
N
(cid:16) nEf
Q(cid:88)
Q
β=1
pEEJ−jEErβ + nE(1 − f )pEEjEEr(bg)
E
(cid:17)
(σ(bg)
E )2 = τm,E
pEE(J−jEE)2(1 + δ2)rβ + nE(1 − f )pEEj2
EE(1 + δ2)r(bg)
E
−nI pEI jEI rI + nEpE0jE0rext
,
(cid:16) nEf
Q(cid:88)
Q
β=1
(cid:17)
+nI pEI j2
EI (1 + δ2)rI
,
-- 12 --
and, similarly, for the current to a neuron of the inhibitory population we have
pIEjIErβ + nE(1 − f )pIEjIEr(bg)
E
(cid:17)
−nI pII jII rI + nEpI0jI0rext
,
√
N
µI = τm,I
(cid:16) nEf
Q(cid:88)
Q
β=1
(cid:16) nEf
Q(cid:88)
Q
β=1
σ2
I = τm,I
+nI pII j2
II (1 + δ2)rI
.
pIEj2
IE(1 + δ2)rβ + nE(1 − f )pIEj2
(cid:17)
IE(1 + δ2)r(bg)
E
(2.1)
These expressions enter the mean field equations as described in Methods (section Mean field theory).
2.4 Network with overlapping clusters
In the spiking network with overlapping clusters of Fig. S5b, each neuron (network size: N = 2000) had a
probability f = 0.06 of belonging to one of the Q = 14 clusters [3]. Hence, the fraction of neurons in a
population shared by a set of k specific clusters (a population of 'rank' k) was f k(1 − f )Q−k (there are(cid:0)Q
(cid:1)
such populations); and the fraction of neurons in the unstructured background population (rank 0) was (1−f )Q.
Synaptic weights Jij between pre- and post-synaptic excitatory neurons j and i respectively, were given by the
stochastic variable
k
Jij = pEE (ijξ+ + (1 − ij)ξ−) ,
where pEE = 0.2 is the E → E connection probability, ξ± are the two possible values of the synaptic weights
(see below), and ij and 1−ij are the potentiation and depression probabilities, P (Jij = ξ+) and P (Jij = ξ−),
respectively. Weights were potentiated or depressed according to the rule [3, 9]
where Pij =(cid:80)Q
ij =
,
(2.2)
Pij
Pij + ρf Dij
i = 0 otherwise), while Dij =(cid:80)Q
j is the number of clusters in common between neurons i and j (ηk
i = 1 if cluster k
i ηk
k=1 ηk
contains neuron i and ηk
j ), and ρ = 2.75. To motivate Eq. 2.2,
we note that this is the asymptotic value of the potentiation probability obtained from the stochastic Hebbian
learning rule of ref. [9] in the limit of slow learning. Finally, ξ+ and ξ− were random variables sampled from
+ and δ2J 2−, respectively (δ = 0.01). The E
normal distributions with means J+ and J− and variances δ2J 2
→ I, I → E, and I → I connection probability was the same as for the homogeneous and clustered networks
described (Table 1 in the main text); the stimuli and the anticipatory cue targeted the excitatory neurons as in
the main model with segregated clusters. The remaining parameters of the network are reported in Table S2.
i (1 − ηk
k=1 ηk
References
[1] R. Vincis and A. Fontanini, Associative learning changes cross-modal representations in the gustatory cortex, Elife 5
(2016) e16420.
[2] C. L. Samuelsen, M. P. Gardner, and A. Fontanini, Effects of cue-triggered expectation on cortical processing of taste,
Neuron 74 (2012), no. 2 410 -- 422.
-- 13 --
Symbol Description
N.
jEE
jEI
jIE
jII
jE0
jI0
J+
rE
ext
rI
ext
V thr
E
V thr
I
V reset
τm
τref
τsyn
N.
N.
N.
Mean E→E weights ×√
Mean I→E weight ×√
Mean E→I weight ×√
Mean I→I weight ×√
Mean afferent synaptic weights to E neurons ×√
Mean afferent synaptic weights to I neurons ×√
Potentiated intra-cluster E→E weights factor.
Average afferent rate to E neurons (baseline).
Average afferent rate to I neurons (baseline).
E neuron threshold potential.
I neuron threshold potential.
E and I neurons reset potential.
E and I membrane time constant.
Absolute refractory period.
E and I synaptic time constant.
N.
N.
Value
1.4 mV
5.0 mV
2.5 mV
6.1 mV
7.3 mV
6.5 mV
10.5
5 spks/s
7 spks/s
3.6 mV
5.7 mV
0 mV
20 ms
5 ms
4 ms
Table. S2. Parameters for the network with overlapping clusters with N = 2000 LIF neurons. See Sec. 2.4 for details.
[3] E. Curti, G. Mongillo, G. La Camera, and D. J. Amit, Mean field and capacity in realistic networks of spiking neurons
storing sparsely coded random memories, Neural computation 16 (2004), no. 12 2597 -- 2637.
[4] F. S. Chance, L. F. Abbott, and A. D. Reyes, Gain modulation from background synaptic input, Neuron 35 (2002),
no. 4 773 -- 782.
[5] D. B. Katz, S. A. Simon, and M. A. Nicolelis, Dynamic and multimodal responses of gustatory cortical neurons in
awake rats, J Neurosci 21 (2001), no. 12 4478 -- 89.
[6] A. Jezzini, L. Mazzucato, G. La Camera, and A. Fontanini, Processing of hedonic and chemosensory features of taste
in medial prefrontal and insular networks, The Journal of Neuroscience 33 (2013), no. 48 18966 -- 18978.
[7] N. K. Horst and M. Laubach, Reward-related activity in the medial prefrontal cortex is driven by consumption, Front
Neurosci 7 (2013) 56.
[8] C. R. Gallistel, S. Fairhurst, and P. Balsam, The learning curve: implications of a quantitative analysis, Proc Natl
Acad Sci U S A 101 (2004), no. 36 13124 -- 31.
[9] N. Brunel, F. Carusi, and S. Fusi, Slow stochastic hebbian learning of classes of stimuli in a recurrent neural network,
Network: Computation in Neural Systems 9 (1998), no. 1 123 -- 152.
-- 14 --
|
1306.3285 | 1 | 1306 | 2013-06-14T03:36:08 | Third-Order Edge Statistics: Contour Continuation, Curvature, and Cortical Connections | [
"q-bio.NC"
] | Association field models have attempted to explain human contour grouping performance, and to explain the mean frequency of long-range horizontal connections across cortical columns in V1. However, association fields only depend on the pairwise statistics of edges in natural scenes. We develop a spectral test of the sufficiency of pairwise statistics and show there is significant higher order structure. An analysis using a probabilistic spectral embedding reveals curvature-dependent components. | q-bio.NC | q-bio |
Third-Order Edge Statistics: Contour Continuation,
Curvature, and Cortical Connections
Matthew Lawlor
Applied Mathematics
Yale University
New Haven, CT 06520
[email protected]
Steven W. Zucker
Computer Science
Yale University
New Haven, CT 06520
[email protected]
Abstract
Association field models have attempted to explain human contour grouping per-
formance, and to explain the mean frequency of long-range horizontal connections
across cortical columns in V1. However, association fields only depend on the
pairwise statistics of edges in natural scenes. We develop a spectral test of the suf-
ficiency of pairwise statistics and show there is significant higher order structure.
An analysis using a probabilistic spectral embedding reveals curvature-dependent
components.
1
Introduction
Natural scene statistics have been used to explain a variety of neural structures. Driven by the
hypothesis that early layers of visual processing seek an efficient representation of natural scene
structure, decorrelating or reducing statistical dependencies between subunits provides insight into
retinal ganglion cells [16], cortical simple cells [13, 2], and the firing patterns of larger ensembles
[17]. In contrast to these statistical models, the role of neural circuits can be characterized function-
ally [3, 14] by positing roles such as denoising, structure enhancement, and geometric computations.
Such models are based on evidence of excitatory connections among co-linear and co-circular neu-
rons [5], as well as the presence of co-linearity and co-circularity of edges in natural images [8],
[7]. The fact that statistical relationships have a geometric structure is not surprising: To the extent
that the natural world consists largely of piecewise smooth objects, the boundaries of those objects
should consist of piecewise smooth curves.
Common patterns between excitatory neural connections, co-occurrence statistics, and the geometry
of smooth surfaces suggests that the functional and statistical approaches can be linked. Statistical
questions about edge distributions in natural images have differential geometric analogues, such as
the distribution of intrinsic derivatives in natural objects. From this perspective, previous studies
of natural image statistics have primarily examined "second-order" differential properties of curves;
i.e., the average change in orientation along curve segments in natural scenes. The pairwise statistics
suggest that curves tend toward co-linearity, in that the (average) change in orientation is small.
Similarly, for long-range horizontal connections, cells with similar orientation preference tend to be
connected to each other.
Is this all there is? From a geometric perspective, do curves in natural scenes exhibit continuity in
curvatures, or just in orientation? Are edge statistics well characterized at second-order? Does the
same hold for textures?
To answer these questions one needs to examine higher-order statistics of natural scenes, but this
is extremely difficult computationally. One possibility is to design specialized patterns, such as in-
tensity textures [15], but it is difficult to generalize such results into visual cortex. We make use
of natural invariances in image statistics to develop a novel spectral technique based on preserving
1
a probabilistic distance. This distance characterizes what is beyond association field models (dis-
cussed next) to reveal the "third-oder" structure in edge distributions. It has different implications
for contours and textures and, more generally, for learning.
2 Edge Co-occurrence Statistics
Edge co-occurrence probabilities are well studied [1, 8, 6, 11]. Following them, we use random
variables indicating edges at given locations and orientations. More precisely, an edge at position,
orientation (xi, yi, θi), denoted Ei, is a {0, 1} valued random variable. Co-occurrence statistics
examine various aspects of pairwise marginal distributions, which we denote by P (Ei, Ej).
The image formation process endows scene statistics with a natural translation invariance. If the
camera were allowed to rotate randomly about the focal axis, natural scene statistics would also
have a rotational invariance. For computational convenience, we enforce this rotational invariance
by randomly rotating our images. Thus,
P (E(xi,yi,θi), ..., E(xj ,yj ,θj )) = P (E(xi+δx,yi+δy,θi+δθ), ..., E(xj +δx,yj +δy,θj +δθ))
We can then estimate joint distributions of nearby edges by looking at patches of edges centered at
a (position, orientation) location (xi, yi, θi) and rotating the patch into a canonical orientation and
position that we denote E0 = E(x0,y0,θ0).
P (E(xi,yi,θi), ..., E(xj ,yj ,θj )) = P (E(xi−xj ,yi−yj ,θi−θj ), ..., E(x0,y0,θ0))
August and Zucker, 2000
Geisler et al, 2001
Elder & Goldberg, 2002
Figure 1: Association fields derive from image co-occurrence statistics. We interpret them as illus-
trating the probability (likelihood) of an edge near a horizontal edge at the center position.
Several examples of statistics derived from the distribution of P (Ei, E0) are shown in Fig. 1. These
are pairwise statistics of oriented edges in natural images. The most important visible feature of
these pairwise statistics is that of good continuation: Conditioned on the presence of an edge at the
center, edges of similar orientation and horizontally aligned with the edge at the center have high
probability. Note that all of the above implicitly or explicit enforced rotation invariance, either by
only examining relative orientation with respect to a reference orientation or by explicit rotation of
the images.
It is critical to estimate the degree to which these pairwise statistics characterize the full joint dis-
tribution of edges (Fig. 2). Many models for neural firing patterns imply relatively low order joint
statistics. For example, spin-glass models imply pairwise statistics are sufficient, while Markov
random fields have an order determined by the size of neighborhood cliques.
3 Contingency Table Analysis
To test whether the joint distribution of edges can be well described by pairwise statistics, we per-
formed a contingency table analysis of edge triples at two different threshold levels. We computed
estimated joint distributions for each triple of edges in an 11× 11× 8 patch, not constructed to have
an edge at the center. Using a χ2 test, we computed the probability that each edge triple distribution
could occur under hypothesis H0 : {No three way interaction}. This is a test of the hypothesis that
log P (Ei, Ej, Ek) = f (Ei, Ej) + g(Ej, Ek) + h(Ei, Ek)
2
Figure 2: Two approximately equally likely triples of edges under the pairwise independence as-
sumption of Elder et. al. Conditional independence is one of several possible pairwise distributional
assumptions. Intuitively, however, the second triple is much more likely. We examine third-order
statistics to demonstrate that this is in fact the case.
for each triple (Ei, Ej, Ek), and includes the cases of independent edges, conditionally independent
edges, and other pairwise interactions. For almost all triples, this probability was extremely small.
(The few edge triples for which the null hypothesis cannot be rejected consisted of edges that were
spaced very far apart, which are far more likely to be nearly statistically independent of one another.)
n = 150705016
percentage of triples where pH0 > .05
threshold = .05
0.0082%
threshold = .1
0.0067%
4 Counting Triple Probabilities
We chose a random sampling of black and white images from the van Hataren image dataset[10].
They were randomly rotated and then filtered using oriented Gabor filters covering 8 angles from
[0, π). Each Gabor has a carrier period of 1.5 pixels per radian and an envelope standard deviation of
5 pixels. The filters were convolved in near quadrature pairs, squared and summed. To restrict anal-
ysis to the statistics of curves, we applied local non-maxima suppression across orientation columns
in a direction normal to the given orientation. This removes effects of filter blurring and oriented tex-
ture patterns (considered shortly). The resulting edge maps were subsampled to eliminate statistical
dependence due to overlapping filters.
Thresholding the edge map yields E : U → {0, 1}, where U ⊂ R2 × S is a discretization of R2 × S.
We treat E as a function or a binary vector as convenient. We randomly select 21 × 21 × 8 image
patches with an oriented edge at the center, and denote these characteristic patches by Vi
(a)
(b)
Figure 3: Example image (a) and edges (b) for statistical analysis. Note: color corresponds to
orientation
3
Since edges are significantly less frequent than their absence, we focus on (positive) edge co-
occurance statistics. For simplicity, we denote P (Ei = 1, Ej = 1, Ek = 1) by P (i, j, k) (A
small orientation anisotropy has been reported in natural scenes (e.g., [9]), but does not appear in
our data because we effectively averaged over orientations by randomly rotating the images.)
We compute the matrix P (i, j0) where
P (i, j0) = P (Ei = 1, Ej = 1E0 = 1)
n(cid:88)
i=1
∼ 1
n
ViV T
i
where Vi is a (vectorized) random patch of edges centered around an edge with orientation θi = 0.
5 Visualizing Triples of Edges
By analogy with the pairwise analysis above, we seek to find those edge triples that frequently co-
occur. But this is significantly more challenging. For pairwise statistics, one simply fixes an edge to
lie in the center and "colors" the other edge by the joint probability of the co-occurring pair (Fig. 1).
No such plot exists for triples of edges. Even after conditioning, there are over 12 million edge
triples to consider.
Our trick: Embed edges in a low dimensional space such that the distance between the edges rep-
resents the relative likelihood of co-occurrence. We shall do this in a manner such that distance in
Embedded Space ∼ Relative Probability.
As before, let Ei be a binary random variable, where Ei = 1 means there is an edge at location
vi = (xi, yi, θi). We define a distance between edges
+(vi, vj) = P (i, i0) − 2P (i, j0) + P (j, j0)
D2
The first and the last terms represent pairwise co-occurrence probabilities; i.e., these are the asso-
ciation field. The middle term represents the interaction between Ei and Ej conditioned on the
presence of E0. Thus this distance is zero if the edges always co-occur in images, given the hori-
zontal edge at the origin, and is large if the pair of edges frequently occur with the horizontal edge
but rarely together. (The relevance to learning is discussed below.)
We will now show how, for natural images, edges can be placed in a low dimensional space where
the distance in that space will be proportional to this probabilistic distance.
6 Dimensionality Reduction via Spectral Theorem
Exploiting the fact that P (i, j0) is a positive semi-definite matrix, we introduce the spectral expan-
sion
P (i, j0) =
λlφl(i)φl(j)
l=1
n(cid:88)
(cid:32) xi
(cid:112)
yi
θi
(cid:33)
→ Rn
where φl is an eigenvector of the K matrix.
Define the SPECTRAL EMBEDDING Φ :
Φ(vi) = {(cid:112)
λ1φ1(i),
λ2φ2(i), ...,
(cid:112)
λnφn(i)}
(1)
The Euclidean distance between embedded points is then
(cid:107)Φ(vi) − Φ(vj)(cid:107)2 = (cid:104)Φ(vi), Φ(vi)(cid:105) − 2(cid:104)Φ(vi), Φ(vj)(cid:105) + (cid:104)Φ(vj), Φ(vj)(cid:105)
= P+(i, i0) − 2P+(i, j0) + P+(j, j0)
= D2
+(vi, vj)
4
Figure 4: Spectrum of P (i, j0). Other spectra are similar. Note rapid decay of the spectrum indi-
cating the diffusion distance is well captured by embedding using only the first few eigenfunctions.
Φ maps edges to points in an embedded space where squared distance is equal to relative probability;
we refer to this as diffusion distance.
The usefulness of this embedding comes from the fact that the spectrum of P (i, j, 0) decays rapidly
(Fig. 6). Therefore we truncate Φ, including only dimensions with high eigenvalues. This gives a
dramatic reduction in dimensionality, and allows us to visualize the relationship between triples of
edges (Fig. 6).
φ2
φ3
φ4
Figure 5: Display of third-order edge structure showing how oriented edges are related to their
spectral embeddings. (top) Spectral embeddings. Note clusters of co-occuring edges. (bottom) Edge
distributions. The eigenvectors of P (i, j, 0) are used to color both the edges and the embedding.
Clusters of edges in this embedding have a high probability of co-occurring along with the edge in
the center. Compare with Fig. 1 where red edges all have high probability of occurring with the
center, but no information is known about their co-occurrence probability.
To highlight information not contained in the association field, we normalized our probability matrix
by its row sums, and removed all low-probability edges. Embedding the mapping from R2 × S →
Rm reveals the cocircular structure of edge triples in the image data (Fig. 5. The colors along each
column correspond, so similar colors map to nearby points along the dimension corresponding to
the row. Under this dimensionality reduction, each small cluster in diffusion space corresponds to
half of a cocircular field. In effect, the coloring by φ2 shows good continuation in orientation (with
our crude quantization) while the coloring by φ4 shows co-circular connections. In effect, then, the
association field is the union of co-circular connections, which also follows from marginalizing the
third-order structure away. We used 40,000 (21 × 21 × 8) patches.
5
051015202530354000.20.40.60.811.2lambdaSpectrum of co−occurance kernel−0.2−0.15−0.1−0.0500.050.10.15−0.2−0.15−0.1−0.0500.050.10.15phi2(x)phi4(x)Diffusion map colored by phi2−0.2−0.15−0.1−0.0500.050.10.15−0.2−0.15−0.1−0.0500.050.10.15phi2(x)phi4(x)Diffusion Map colored by phi3−0.2−0.15−0.1−0.0500.050.10.15−0.2−0.15−0.1−0.0500.050.10.15phi2(x)phi4(x)Diffusion Map colored by phi4 7
Implications: "Cells that fire together wire together"
Shown above are low dimensional projections of the diffusion map and their corresponding color-
ings in R2 × S. To provide a neural interpretation of these results, let each point in R2 × S represent
a neuron with a receptive field centered at the point (x, y) with preferred orientation θ. Each cluster
then signifies those neurons that have a high probability of co-firing given that the central neuron
fires, so clusters in diffusion coordinates should be "wired" together by the Hebbian postulate. Such
curvature-based facilitation can explain the non-monotonic variance in excitatory long-range hori-
zontal connections in V1 [3, 4]. It may also have implications for the receptive fields of V2 neurons.
As clusters of co-circular V1 complex cells are correlated in their firing, it may be efficient to rep-
resent them with a single cell with excitatory feedforward connections. This predicts that efficient
coding models that take high order interactions into account should exhibit cells tuned to curved
boundaries.
Our approach also has implications beyond excitatory connections for boundary facilitation. We
repeated our conditional spectral embedding, but now conditioned on the absence of an edge at the
center (Fig. 6). This could provide a model for inhibition, as clusters of edges in this embedding are
likely to co-occur conditioned on the absence of an edge at the center. We find that the embedding is
only one dimensional, with the significant eigenvector a monotonic function of orientation difference
from the center orientation. Compared to excitatory connections, this suggests that inhibition is
relatively unstructured, and agrees with many neurobiological studies.
Figure 6: Embeddings conditioned on the absence of an edge at the center location. Note how
less structured it is, compared to the positive embeddings. As such it could serve as a model for
inhibitory connections, which span many orientations.
Finally, we repeated this third-order analysis (but without local non-maxima supression) on a struc-
tured model for isotropic textures on 3D surfaces and again found a curvature dependency (Fig. 7).
Every 3-D surface has a pair of associated dense texture flows in the image plane that correspond to
the slant and tilt directions of the surface. For isotropic textures, the slant direction corresponds to
the most likely orientation signalled by oriented filters. Hence, the distribution of these directions in
3-D shape are an excellent proxy for measuring the response of orientation filters to dense textured
objects.
As this is a representation of a dense vector field, it is more difficult to interpret than the edge map.
We therefore applied k-means clustering in the embedded space and segmented the resulting vector
field. The resulting clusters show two-sided continuation of the texture flow with a fixed tangential
curvature (Fig. 8).
In summary, then, we have developed a method for revealing third-order orientation structure by
spectral methods. It is based on a diffusion metric that makes third-order terms explicit, and yields
a Euclidean distance measure by which edges can be clustered. Given that long-range horizontal
connections are consistent with these clusters, how biological learning algorithms converge to them
remains an open question. Given that research in computational neuroscience is turning to third-
order [12] and specialized interactions, this question now becomes more pressing.
6
(cid:239)10(cid:239)50510(cid:239)1001012345678−0.2−0.15−0.1−0.0500.050.10.15−0.2−0.15−0.1−0.0500.050.10.15 (a)
(b)
φ2
φ3
φ4
Figure 7: (top) Oriented textures provide information about surface shape.
(bottom) As before,
we looked at the conditional co-occurrence matrices of edge orientations over a series of randomly
generated shapes. Slant orientations and embedding colored by each eigenvector. The edge map is
thresholded to contain only orientations of high probability. The resulting embedding φ(vi) of those
orientations is shown below. The eigenvectors of P (i, j, 0) are used to color both the orientations and
the embedding. Clusters of orientations in this embedding have a high probability of co-occurring
along with the edge in the center.
Figure 8: Clustering of dense texture flows. Color corresponds to the cluster index. Clusters were
separated into different figures so as to minimize the x, y overlap of the orientations. Embedding on
the right is identical to the embeddings above, but viewed along the φ3, φ4 axes.
References
[1] Jonas August and Steven W Zucker. The curve indicator random field: Curve organization
via edge correlation. In Perceptual organization for artificial vision systems, pages 265 -- 288.
Springer, 2000.
7
−0.06−0.04−0.0200.02−0.1−0.0500.050.1−0.1−0.0500.050.1−0.06−0.04−0.0200.02−0.1−0.0500.050.1−0.1−0.0500.050.1−0.06−0.04−0.0200.02−0.1−0.0500.050.1−0.1−0.0500.050.1−0.0500.05−0.06−0.04−0.0200.020.040.060.08−0.08−0.06−0.04−0.0200.020.040.060.080.1 [2] A.J. Bell and T.J. Sejnowski. The independent components of natural scenes are edge filters.
Vision research, 37(23):3327 -- 3338, 1997.
[3] O. Ben-Shahar and S. Zucker. Geometrical computations explain projection patterns of long-
range horizontal connections in visual cortex. Neural Computation, 16(3):445 -- 476, 2004.
[4] William H Bosking, Ying Zhang, Brett Schofield, and David Fitzpatrick. Orientation selectivity
and the arrangement of horizontal connections in tree shrew striate cortex. The Journal of
Neuroscience, 17(6):2112 -- 2127, 1997.
[5] Heather J. Chisum, Franois Mooser, and David Fitzpatrick. Emergent properties of layer 2/3
neurons reflect the collinear arrangement of horizontal connections in tree shrew visual cortex.
The Journal of Neuroscience, 23(7):2947 -- 2960, 2003.
[6] James H Elder and Richard M Goldberg. Ecological statistics of gestalt laws for the perceptual
organization of contours. Journal of Vision, 2(4), 2002.
[7] J.H. Elder and RM Goldberg. The statistics of natural image contours. In Proceedings of the
IEEE Workshop on Perceptual Organisation in Computer Vision. Citeseer, 1998.
[8] WS Geisler, JS Perry, BJ Super, and DP Gallogly. Edge co-occurrence in natural images
predicts contour grouping performance. Vision research, 41(6):711 -- 724, 2001.
[9] Bruce C Hansen and Edward A Essock. A horizontal bias in human visual processing of
orientation and its correspondence to the structural components of natural scenes. Journal of
Vision, 4(12), 2004.
[10] J. H. van Hateren and A. van der Schaaf.
Independent component filters of natural images
compared with simple cells in primary visual cortex. Proceedings: Biological Sciences,
265(1394):359 -- 366, Mar 1998.
[11] Norbert Kruger. Collinearity and parallelism are statistically significant second-order relations
of complex cell responses. Neural Processing Letters, 8(2):117 -- 129, 1998.
[12] Ifije E Ohiorhenuan and Jonathan D Victor. Information-geometric measure of 3-neuron fir-
ing patterns characterizes scale-dependence in cortical networks. Journal of computational
neuroscience, 30(1):125 -- 141, 2011.
[13] Bruno A Olshausen et al. Emergence of simple-cell receptive field properties by learning a
sparse code for natural images. Nature, 381(6583):607 -- 609, 1996.
[14] T.K. Sato, I. Nauhaus, and M. Carandini. Traveling waves in visual cortex. Neuron, 75(2):218 --
229, 2012.
[15] Gasper Tkacik, Jason S Prentice, Jonathan D Victor, and Vijay Balasubramanian. Local statis-
tics in natural scenes predict the saliency of synthetic textures. Proceedings of the National
Academy of Sciences, 107(42):18149 -- 18154, 2010.
[16] JH Van Hateren. A theory of maximizing sensory information. Biological cybernetics,
68(1):23 -- 29, 1992.
[17] William E Vinje and Jack L Gallant. Sparse coding and decorrelation in primary visual cortex
during natural vision. Science, 287(5456):1273 -- 1276, 2000.
8
|
1210.6989 | 1 | 1210 | 2012-10-25T20:01:51 | The impact of short term synaptic depression and stochastic vesicle dynamics on neural variability | [
"q-bio.NC"
] | Neural variability plays a central role in neural coding and neuronal network dynamics. Unreliability of synaptic transmission is a major source of neural variability: synaptic neurotransmitter vesicles are released probabilistically in response to presynaptic spikes and are recovered stochastically in time. The stochastic dynamics of this process interacts with variability in the arrival times of presynaptic spikes to shape the variability of the postsynaptic response. We use continuous time Markov chain methods to analyze a model of short term synaptic depression with stochastic vesicle dynamics coupled with three different models of presynaptic spiking: one model in which the timing of presynaptic spikes are modeled as a Poisson process, one in which spikes occur more regularly than a Poisson process and one in which spikes occur more irregularly. We use this analysis to investigate how variability in a presynaptic spike train is transformed by short term synaptic depression and stochastic vesicle dynamics to determine the variability of the postsynaptic response to a presynaptic spike train. We find that regular presynaptic spiking increases the average rate at which vesicles are released, that the number of vesicles released over a time window is more variable for smaller time windows than for larger time windows, and that fast presynaptic spiking gives rise to Poisson-like variability of the postsynaptic response even when presynaptic spike times are highly non-Poisson. Our results complement and extend previously reported theoretical results and provide possible explanations for some trends observed in recorded data. | q-bio.NC | q-bio | The impact of short term synaptic depression and
stochastic vesicle dynamics on neuronal variability
Steven Reich · Robert Rosenbaum
2
1
0
2
t
c
O
5
2
]
.
C
N
o
i
b
-
q
[
1
v
9
8
9
6
.
0
1
2
1
:
v
i
X
r
a
Abstract Neuronal variability plays a central role in
neural coding and impacts the dynamics of neuronal
networks. Unreliability of synaptic transmission is a
major source of neural variability: synaptic neurotrans-
mitter vesicles are released probabilistically in response
to presynaptic action potentials and are recovered stochas-
tically in time. The dynamics of this process of vesicle
release and recovery interacts with variability in the
arrival times of presynaptic spikes to shape the vari-
ability of the postsynaptic response. We use continuous
time Markov chain methods to analyze a model of short
term synaptic depression with stochastic vesicle dynam-
ics coupled with three different models of presynaptic
spiking: one model in which the timing of presynaptic
action potentials are modeled as a Poisson process, one
in which action potentials occur more regularly than a
Poisson process and one in which action potentials oc-
cur more irregularly. We use this analysis to investigate
how variability in a presynaptic spike train is trans-
formed by short term depression and stochastic vesicle
dynamics to determine the variability of the postsynap-
tic response. We find that regular presynaptic spiking
increases the average rate at which vesicles are released,
that the number of vesicles released over a time win-
This work supported by National Foundation of Science grant
NSF-EMSW21-RTG0739261 and National Institute of Health
grant NIH-1R01NS070865-01A1.
Steven Reich
Department of Mathematics, University of Pittsburgh, Pitts-
burgh PA, USA
Robert Rosenbaum (corresponding author)
Department of Mathematics and Center for the Neural Basis
of Cognition, University of Pittsburgh, Pittsburgh PA, USA
Tel.: +1412-624-8375
Fax: +1412-624-83970
E-mail: [email protected]
dow is more variable for smaller time windows than
larger time windows and that fast presynaptic spiking
gives rise to Poisson-like variability of the postsynaptic
response even when presynaptic spike times are non-
Poisson. Our results complement and extend previously
reported theoretical results and provide possible expla-
nations for some trends observed in recorded data.
Keywords Short term depression · Synaptic variabil-
ity · Fano factor
1 Introduction
Variability of neural activity plays an important role in
population coding and network dynamics [21]. Random
fluctuations in the number of action potentials emitted
by a population of neurons affects the firing rate of
downstream cells [50,49]. In addition, spike count vari-
ability over both short and long timescales can impact
the reliability of a rate-coded signal [16]. It is therefore
important to understand how this variability is shaped
by synaptic and neuronal dynamics.
Several studies examine the question of how intrin-
sic neuronal dynamics interact with variability in presy-
naptic spike timing to determine the statistics of a post-
synaptic neuron's spiking response, but many of these
studies do not account for dynamics and variability in-
troduced at the synaptic level by short term synaptic
depression and stochastic vesicle dynamics. Synapses
release neurotransmitter vesicles probabilistically in re-
sponse to presynaptic spikes and recover released vesi-
cles stochastically over a timescale of several hundred
milliseconds [65,22]. The dynamics and variability in-
troduced by short term depression and stochastic vesi-
cle dynamics alter the response properties of a postsy-
naptic neuron [60,2,10,34,26,48,25,28,17,19,44,6,43]
and therefore play an important role in information
2
Steven Reich, Robert Rosenbaum
transfer [63,18,24,36,45], neural coding [57,12,1,27,20,
32,40] and network dynamics [55,23,7,62,56,5]. Under-
standing how variability in presynaptic spike times in-
teract with short term depression and stochastic vesicle
dynamics to determine the statistics of the postsynaptic
response is therefore an important goal.
In this study, we use a model of short term synap-
tic depression with stochastic vesicle dynamics to ex-
amine how variability in a presynaptic input is trans-
ferred to variability in the synaptic response it pro-
duces. We use the theory of continuous-time Markov
chains to construct exact analytical methods for calcu-
lating the the statistics of the postsynaptic response to
three different presynaptic spiking models: one model
with Poisson spike arrival times, one with more regular
spike arrival times, and one with more irregular spike
arrival times. We find that depressing synapses shape
the timescale over which neuronal variability occurs:
the number of neurotransmitter vesicles released over
a time interval is highly variable for shorter time win-
dows, but less variable for longer time windows when
variability is quantified using Fano factors. Addition-
ally, we find that when presynaptic inputs are highly ir-
regular (Fano factor greater than 1), synaptic dynamics
cause a reduction in Fano factor, consistent with pre-
vious studies [26,25,18,19]. On the other hand, when
presynaptic input is more regular (Fano factor less than
1), synaptic dynamics often cause an increase in Fano
factor. This observation suggests a mechanism through
which irregular and Poisson-like variability can be sus-
tained in spontaneously spiking neuronal networks [54,
51,8,9,35,11], which complements previously proposed
mechanisms [58,61,52,29,33].
2 Methods
We begin by introducing the synapse model used through-
out this study. We then proceed by analyzing the statis-
tics of the synaptic response to three different input
models.
2.1 Synapse model
A widely used model of depressing synapses [57,2,55,
34,48] does not capture stochasticity in vesicle recovery
and release. As a result, this model underestimates the
variability of the synaptic response [19,43]. For this rea-
son, we use a more detailed synapse model that takes
stochastic recovery times and probabilistic release into
account [60,62,22,43].
We consider a presynaptic neuron with spike train
I(t) = Pj δ(t − tj) that makes M functional contacts
onto a postsynaptic cell. Here, tj is the time of the
jth presynaptic action potential. Define m(t) to be the
number of contacts with a readily releasable neuro-
transmitter vesicle at time t (so that 0 ≤ m(t) ≤ M ).
For simplicity, we assume that each contact can release
at most one neurotransmitter vesicle in response to a
presynaptic spike. When a presynaptic spike arrives,
each contact with a releasable vesicle releases its vesi-
cle independently with probability pr. After releasing
a vesicle, a synaptic contact enters a refractory period
during which it is unavailable to release a vesicle again
until it recovers by replacing the released vesicle. The
recovery time at a single contact is modeled as a Pois-
son process with rate 1/τu. Equivalently, the duration of
the refractory period is exponentially distributed with
mean τu.
Define wj to be the number of contacts that release
a vesicle in response to the presynaptic spike at time
tj (so that 0 ≤ wj ≤ m(t−
j ) =
m(t)). The synaptic response is quantified by
limt→t−
the marked point process
j ) ≤ M where m(t−
j
x(t) =Xj
wjδ(t − tj).
Since the signal observed by the postsynaptic cell is de-
termined by x(t), we quantify synaptic response statis-
tics in terms of the statistics of x(t) in our analysis. The
process x(t) can be convolved with a post-synaptic re-
sponse kernel to obtain the conductance induced on the
postsynaptic cell [43]. The effects of this convolution on
response statistics is well understood [53], so we do not
consider it here.
This model can be described more precisely using
the equation [43]
dm(t) = −dNx(t) + dNu(t)
where dNu(t) = u(t)dt is the increment of an inhomo-
geneous Poisson process with instantaneous rate that
depends on m(t) through hdNu(t)i m(t)i = dt(M −
m(t))/τu (here, h· ·i denotes conditional expectation),
0 x(s)ds is the number of vesicles released up
to time t, and each wj is a binomial random variable
with mean prm(tj) and variance m(tj)pr(1 − pr).
Nx(t) =R t
2.2 Statistical measures of the presynaptic spike train
and the synaptic response
We focus on steady state statistics in this article, and
therefore assume that the presynaptic spike trains are
stationary and that the synapses have reached statis-
tical equilibrium. The intensity of a presynaptic spike
train is quantified by the mean presynaptic firing rate,
rin = hI(t)i = hNI (T )/T i
Short term depression and neuronal variability
3
where h·i denotes the expected value and
2.3 Model analysis with Poisson presynaptic inputs
NI (T ) =Z T
0
I(s)ds
represents the number of spikes in the time interval
[0, T ]. Temporal correlations in the presynaptic spike
times are quantified by the auto-covariance,
Rin(τ ) = cov(I(t), I(t + τ )),
and the variability in the presynaptic spike train is
quantified by its Fano factor,
Fin(T ) =
var(NI (T ))
T rin
.
For much of this work, we will focus on Fano factors
over large time windows which, through a slight abuse
of notation, we denote by Fin = limT →∞ Fin(T ). To
compute Fano factors, we will often exploit their rela-
tionship to auto-covariance functions,
Fin(T ) =
1
rin Z T
−T
Rin(τ )(1 − τ /T )dτ
and
Fin =
1
rin Z ∞
−∞
Rin(τ )dτ.
(1)
(2)
The statistics of the synaptic response, x(t), are de-
fined analogously to the statistics of I(t). The steady
state rate of vesicle release is defined as
rx = hx(t)i = hNx(T )/T i
where Nx(T ) = R T
0 x(s)ds represents the number of
vesicles released in the time interval [0, T ]. Temporal
correlations in the synaptic response are quantified by
the auto-covariance,
We first consider a homogeneous Poisson input, I(t),
with rate rin. The input auto-covariance for this model
is given by Rin(τ ) = rinδ(τ ) and the Fano factor is given
by Fin(T ) = 1 for any T > 0. The mean rate of vesicle
release for this model is given by
rx =
M prrin
prrinτu + 1
which saturates to M/τu for large presynaptic rates, rin.
A closed form approximations to the auto-covariance
function of the response for this Poisson input model
are derived in [43,36] (see also [17]) and consist of a
sum of a delta function and an exponential,
Rx(τ ) = Drxδ(τ ) − Erxe−τ /τ0,
(4)
where the mass of the delta function is given by
D =
2pr (rinτu + M − 1) + 2 − pr
2rinτu
(2 − pr)prrinτu + 2
> 0,
(5)
the timescale of the exponential decay is given by
τ0 =
τu
1 + prrinτu
,
and the peak of the exponential is given by
E =
prrin((M − 2)pr + 2)τu + 2(M − 1)pr + 2
M (2 − pr)prrinτu + 2M
rx. (6)
It can easily be checked that E > 0 whenever M ≥ 1,
0 ≤ pr ≤ 1, rin > 0, and τu > 0 so that the peak of the
exponential in (4) is negative. For finite T , the Fano
factor, Fx(T ), is given by
Fx(T ) = D − 2Eτ0 −(cid:16)e− T
and, in the limit of large T ,
τ0 − 1(cid:17) 2Eτ 2
T
0
(7)
(8)
Rx(τ ) = cov(x(t), x(t + τ ))
Fx = D − 2Eτ0.
and response variability is quantified by the Fano factor
of the number of vesicles released,
Fx(T ) =
var(Nx(T ))
T rx
.
As above, we define Fx = limT →∞ Fx(T ) and note that
Fx(T ) =
1
rin Z T
−T
Rx(τ )(1)dτ and Fx =Z ∞
−∞
Rx(τ )dτ.
(3)
To test the accuracy of these approximations, ex-
act solutions can be found numerically using standard
methods for the analysis of continuous-time Markov
chains, as described for alternate input models below.
This analysis is a special case of the analysis for the
"regular" input model described below that is achieved
by taking θ = 1. Alternatively, exact numerical results
can be achieved by taking rs = rf for the "irregular"
input model. In figures showing results for the Poisson
input model, we plot the closed form approximations
described above along with exact numerical results ob-
tained using the regular input model with θ = 1.
4
Steven Reich, Robert Rosenbaum
2.4 Model analysis with irregular presynaptic inputs
then calculated explicitly to obtain
Spike trains measured in vivo often exhibit irregular
spiking statistics indicated by Fano factors larger than
1 [4,15,3,11]. To describe the synaptic response to ir-
regular inputs, we use a model of presynaptic spiking in
which the instantaneous rate of the presynaptic spike
train, I(t), randomly switches between two values, rs
and rf > rs, representing a slow spiking state and a
fast spiking state. The time spent in the slow state
before transitioning to the fast state is exponentially
distributed with mean τs. Likewise, the amount of time
spent in the fast state before switching to the slow state
is exponentially distributed with mean τf . Transition
times are independent from one another and from the
spiking activity. Between transitions, spikes occur as a
Poisson process.
To find rin, Rin(τ ), and Fin, we represent this model
as a doubly stochastic Poisson process. Define r(t) ∈
{rs, rf } to be the instantaneous firing rate at time t.
Then r(t) is a continuous time Markov chain [31] on
the state space Γ = (rs, rf ) with infinitesimal generator
matrix
A =(cid:20) −1/τs 1/τs
1/τf −1/τf(cid:21) .
Clearly, r(t) spends a proportion τs/(τs + τf ) of its
time in the slow state (defined by r(t) = rs) and a pro-
portion τf /(τs + τf ) of its time in the fast state (defined
by r(t) = rf ). This gives a steady-state mean firing rate
of
rin =
rsτs + rf τf
τs + τf
.
At non-zero lags (τ 6= 0), the auto-covariance of a
doubly stochastic Poisson process is the same as the
auto-covariance of r(t) [43], which we can compute us-
ing techniques for analyzing continuous time Markov
chains. For τ > 0, we have
hr(t)r(t + τ )i = rs Pr(r(t) = rs)hr(t + τ ) r(t) = rsi
+ rf Pr(r(t) = rf )hr(t + τ ) r(t) = rf i
=
+
rsτs
τs + τf
rf τf
τs + τf
hr(t + τ ) r(t) = rsi
(9)
hr(t + τ ) r(t) = rf i
where h··i denotes conditional expectation and
hr(t + τ ) r(t) = rsi = rs Pr(r(t + τ ) = rs r(t) = rs)
+ rf (1 − Pr(r(t + τ ) = rs r(t) = rs)).
The probability in this expression can be written in
terms of an exponential of the generator matrix, A, and
Pr(r(t + τ ) = rs r(t) = rs) =(cid:20)eAT τ(cid:18) 1
0(cid:19)(cid:21)1
τf e
=
− τ
τf
− τ
τs + τs
τf + τs
where [v]k denotes the kth component of a vector, v.
An identical calculation can be performed to obtain an
analogous expression for hr(t + τ ) r(t) = rf i. Combin-
ing these with Eq. (9) gives
hr(t)r(t + τ )i =
τf τs (rf − rs)2
(τf + τs)2
− τ
τs
e
− τ
τf + r2
in.
For positive τ , we have Rin(τ ) = hr(t)r(t + τ )i − r2
in . As
with all stationary point processes Rin(τ ) = Rin(−τ )
and Rin(τ ) has a Dirac delta function with mass rin
at the origin [14]. Thus, the auto-covariance of I(t) is
given by
Rin(τ ) = rinδ(τ ) +
τf τs (rf − rs)2
(τf + τs)2
− τ
τs
− τ
τf .
e
(10)
For finite T , the Fano factor, Fin(T ), can be computed
using Eqs. (1) and (10). In the limit of large T , we can
use Eqs. (2) and (10) to obtain a closed form expression,
Fin = 1 +
2τ 2
f τ 2
s (rf − rs)2
(τf + τs)2 (rf τf + rsτs)
.
(11)
Poisson spiking is recovered by setting rf = rs,
τf = 0, or τs = 0. For any other parameter values (i.e.,
when rf 6= rs and τf , τs > 0), it follows from Eq. (11)
that Fin(T ) > 1 for any T . Therefore this input model,
hereafter referred to as the "irregular spiking" model,
represents spiking that is more irregular than a Poisson
process.
The analysis in [43] used to derive closed form ex-
pressions for the response statistics with Poisson in-
puts cannot easily be generalized to derive expressions
with non-Poisson inputs like those considered here. In-
stead, we analyze the synaptic response for the irregu-
lar input model using techniques for analyzing contin-
uous time Markov chains. First note that the process
b(t) = (m(t), r(t)) is a continuous-time Markov chain on
the discrete state space {0, 1, . . . , M } × {rs, rf }. Here,
m(t) denotes the size of readily releasable pool and r(t)
represents the instantaneous presynaptic rate (which
switches between rs and rf ). We enumerate all 2(M +1)
elements of this state space and denote the jth ele-
ment of this enumeration as Γj = (mj, rj) for j =
1, . . . , 2(M + 1).
Short term depression and neuronal variability
5
The infinitesimal generator, B, of b(t) is a 2(M +
1) × 2(M + 1) matrix with off-diagonal terms defined
by the instantaneous transition rates,
Bj,k = lim
h→0
1
h
Pr(b(t + h) = Γk b(t) = Γj), j 6= k (12)
and with diagonal terms chosen so that the rows sum
to zero: Bj,j = −Pk6=j Bj,k [31].
To fill the matrix B, we consider each type of tran-
sition that the process b(t) undergoes. Vesicle recovery
events occur at the instantaneous rate (M − m(t))/τu
and increment the value of m(t) by one vesicle. There-
fore
lim
h→0
1
h
Pr(b(t+h) = (m+1, r) b(t) = (m, r)) =
M − m
τu
for m ∈ {0, . . . , M − 1} and r ∈ {rs, rf }. Vesicle release
events occur at the instantaneous rate r(t) and decre-
ment the value of m(t) by a random amount k with a
binomial distribution so that
lim
h→0
1
h
r
Pr(b(t + h) = (m − k, r) b(t) = (m, r)) =
m!
(m − k)!
pr
k(1 − pr)m−k
for m ∈ {1, . . . , M }, k ∈ {0, . . . , m}, and r ∈ {rs, rf }.
The value of r(t) switches from rs to rf with instanta-
neous rate 1/τs so that
lim
h→0
1
h
Pr(b(t + h) = (m, rf ) b(t) = (m, rs)) =
and, similarly,
lim
h→0
1
h
Pr(b(t + h) = (m, rs) b(t) = (m, rf )) =
1
τs
1
τf
.
These four transition types account for all of the tran-
sitions that b(t) undergoes. They can be used to fill
the off-diagonal terms of the matrix B. The diagonal
terms are then filled to make the rows sum to zero, as
discussed above.
Once a the infinitesimal generator matrix, B, is ob-
tained, the probability distribution of b(t) given an ini-
tial distribution p(0) is given by
p(t) = etBT
p(0).
The stationary distribution, p0, of b(t) is given by the
vector in the one-dimensional null space of B with ele-
ments that sum to one [31].
The instantaneous rate of vesicle release, conditioned
on the current state of r(t) and m(t), is given by
hx(t) r(t) = r, m(t) = mi = rprm.
Averaging over r and m in the steady state gives
rx =
2(M+1)
Xj=1
[p0]jrj prmj
where [·]j denotes the jth element. The auto-covariance,
Rx(τ ), has a Dirac delta function at τ = 0. We separate
this delta function from the continuous part by writing
Rx(τ ) = Axδ(τ ) + R+
x (τ ) is a continuous
function. The area of the delta function can be found by
conditioning on the current state of r(t) in the steady
state to get
x (τ ) where R+
Ax = lim
t→∞
= lim
t→∞
hx(t)2dti
hdN 2
x (t)/dti
2(M+1)
=
Xj=1
rj[p0]j lim
k→∞
hw2
k m(t−
k ) = mji
(13)
where wk is the number of vesicles released by the
kth presynaptic spike. Conditioned on the size, m(t−
k ),
of the readily releasable pool immediately before the
presynaptic spike arrives, wk has a binomial distribu-
tion with second moment,
lim
k→∞
hw2
k m(t−
k ) = mji = mjpr(1 − pr) + m2
j pr
2
which can be substituted into Eq. (13) to calculate Ax.
All that remains is to calculate the continuous part,
R+
x (τ ), of Rx(τ ). First note that, for τ > 0,
lim
t→∞
hx(t)x(t + τ )i
2(M+1)
=
[p0]i Pr(b(t + τ ) = Γj b(t) = Γi)
(14)
Xi,j=1
× hdNx(t)dNx(t + τ ) b(t) = Γi, b(t + τ ) = Γji/dt2.
The second term in Eq. (14) can be computed as
Pr(b(t + τ ) = Γj b(t) = Γi) = [eτ BT
ei]j = [eτ BT
]j,i
where ei is the 2(M + 1) × 1 vector whose ith element is
1 and all other elements are zero, which represents an
initial distribution concentrated at Γi. The last term in
Eq. (14) is given by
hdNx(t)dNx(t + τ ) b(t) = Γi, b(t + τ ) = Γji/dt2 =
rirjpr
2mimj.
Finally, R+
so that
x (τ ) = limt→∞hx(t)x(t + τ )i − r2
x for τ 6= 0
Rx(τ ) =
Axδ(τ ) − r2
x +
2(M+1)
Xi,j=1
[p0]iheτ BTij,i
rirjpr
2mimj
6
Steven Reich, Robert Rosenbaum
which can be computed efficiently using matrix multi-
plication. The response Fano factor, Fx, can then be
found by integrating Rx(τ ) according to Eqs. (1) and
(2).
2.5 Model analysis with regular presynaptic inputs
We now consider a spiking model that gives Fano fac-
tors smaller than 1 and therefore spike trains that are
more regular than Poisson processes. We achieve this by
defining a renewal process with gamma-distributed in-
terspike intervals (ISIs). Such a process can be obtained
by first generating a Poisson process,Pk δ(t − sk) with
rate r = θ rin for some positive integer θ, then keep-
ing only every θth spike to build the spike train I(t).
More precisely, the first spike of the gamma process is
obtained by choosing an integer, k, uniformly from the
set {1, . . . , θ} and defining defining t1 = sk. The re-
maining spikes are defined by tj+1 = sjθ+k to obtain
the stationary renewal process, I(t) =Pj δ(t − tj) [13].
Clearly, this process has rate rin since the original
Poisson process has rate θrin and a proportion 1/θ of
these spikes appear in I(t). The auto-covariance is given
by [42]
Rin(τ ) = rinδ(τ ) + rin ∞
Xk=1
fk(τ ) − rin! .
(15)
where
fk(t) =
tkθ−1(θrin)kθe−θrint
(kθ − 1)!
is the density of the waiting time between the first spike
and the (k + 1)st spike (i.e., the duration of k consecu-
tive ISIs).
For finite T , the Fano factor, Fin(T ), can be com-
puted using Eqs. (1) and (15). In the limit of large T ,
we can use Eq. (2) or use the fact that, for renewal pro-
cesses, Fin = var(ISI)/hISIi2 where var(ISI) = 1/(r2
inθ)
is the variance and hISIi = 1/rin is the mean of the
gamma distributed ISIs [13]. This gives
Fin =
1
θ
.
Poisson spiking is recovered by setting θ = 1. When
θ > 1, we have that Fin(T ) < 1 for any T . Therefore
this model, hereafter referred to as the "regular" input
model, represents spiking that is more regular than a
Poisson process.
The synaptic response with the regular input model
can be analyzed using methods similar to those used for
the irregular model. We introduce an auxiliary process,
q(t), that transitions sequentially through the state space
{1, . . . , θ}. Once reaching θ, q(t) transitions back to
state 1. Transitions occur as a Poisson process with
rate θrin. The waiting times between transitions from
q = θ to q = 1 are gamma distributed. Thus, to recover
the regular input model, we specify that each transi-
tion from 1 = θ to q = 1 represents a single presynaptic
spike. The process g(t) = (m(t), q(t)) is then a con-
tinuous time Markov chain on the discrete state space
{1, . . . , θ} × {0, . . . , M }. We enumerate all θ(M + 1)
elements of this space and denote the jth element as
Γj = (mj , qj) for j = 1, . . . , θ(M + 1).
The infinitesimal generator, G, which is a θ(M +
1) × θ(M + 1) matrix is defined analogously to the ma-
trix B in Eq. (12) above. The elements of G can be
filled using the following transition probabilities. As for
the irregular input model, vesicle recovery occurs as a
Poisson process with rate (M − m(t))/τu so that
lim
h→0
1
h
Pr(g(t+h) = (m+1, q) g(t) = (m, q)) =
M − m
τu
for m = 0, . . . , M and q = 1, . . . , θ. Transitions that
increment q(t) occur with instantaneous rate, θrin so
that
lim
h→0
1
h
Pr(g(t + h) = (m, q + 1) g(t) = (m, q)) = θrin
for q = 1, . . . , θ − 1 and m = 0, . . . , M . The only other
transitions are those from q(t) = θ to q(t) = 1, which
represent a presynaptic spike and are therefore accom-
panied by a release of vesicles. The transitions con-
tribute the following,
lim
h→0
1
h
Pr(g(t + h) = (1, m − k) g(t) = (θ, m)) =
θrin
m!
(m − k)!
pr
k(1 − pr)m−k
for m ∈ {0, 1, . . . , M } and k ∈ {0, . . . , m}. These tran-
sition rates can be used to fill the off-diagonal terms
of the matrix G. The diagonal terms are then filled so
that the rows sum to zero. The stationary distribution,
p0, of g(t) = (m(t), q(t)) is given by the vector in the
one-dimensional null space of G with elements that sum
to one.
A proportion [p0]γ(k) of time is spent in state (m(t), q(t)) =
(k, θ) where γ(k) represents the index of the element
(k, θ) in the enumeration chosen for Γ (i.e., the index,
j, at which Γj = (k, θ)). In that state, the transition
to q(t) = 1 occurs with instantaneous rate θrin and re-
leases average of prm(t) vesicles. Thus, the mean rate
of vesicle release is given by
rx =
M
Xk=1
θrinprk [p0]γ(k).
As above, we separate the auto-covariance into a
delta function and a continuous part by writing Rx(τ ) =
Short term depression and neuronal variability
Axδ(τ ) + R+
x (τ ) is a continuous function.
The area of the delta function at the origin is given by
x (τ ) where R+
Ax =
M
Xk=1
θrin[p0]γ(k)(cid:0)kpr(1 − pr) + k2pr
2(cid:1)
by an argument identical to that used for the irregular
input model above. Also by a similar argument used for
the irregular input model, we have that
Rx(τ ) =
Axδ(τ ) − r2
x + θ2r2
in
M
Xk,l=1
[p0]γ(k)heτ GTiγ(l),γ(k)
pr
2kl.
2.6 Parameters used in figures
Theoretical results are obtained for arbitrary parameter
values, but for all figures we use a set of parameter val-
ues that are consistent with experimental studies. For
synaptic parameters, we use τu = 700 ms and pr = 0.5
consistent with measurements of short term depression
in pyramidal-to-pyramidal synapses in the rat neocor-
tex [57,22]. We also choose M = 5 which is within the
range observed in several cortical areas [6].
The Poisson presynaptic input model is determined
completely by its firing rate and the regular input model
is determined completely by its firing rate and Fano
factor. Presynaptic firing rates and Fano factors are
reported on the axes or captions of each figure. The
irregular input model has four parameters that deter-
mine the firing rate and Fano factor. In all figures, we
set τb = τs = 1.315/c, rb = 37c, and rs = 3c which
gives a Fano factor of Fin = 20.0017 ≈ 20 for any value
of c (from Eq. (11)). Changing c effectively scales the
timescale of presynaptic spiking, hence scaling rin, with-
out changing Fin.
3 Results
We analyze the synaptic response to different patterns
of presynaptic inputs using a stochastic model of short
term synaptic depression in which a presynaptic neuron
makes M functional contacts onto a postsynaptic neu-
ron [60,22,24,19]. The input to the presynaptic neuron
is a spike train denoted by I(t). Neurotransmitter vesi-
cles are released probabilistically in response to each
presynaptic spike. Specifically, a contact with a readily
available vesicle releases this vesicle with probability pr
in response to a single presynaptic spike. After a synap-
tic contact has released its neurotransmitter vesicle, it
enters a refractory state where it is unable to release
again until the vesicle is replaced. The duration of this
7
in
F = 20
F = 1
F = 0.1
in
in
)
c
e
s
/
s
e
c
s
e
v
(
i
l
x
r
7
6
5
4
3
2
1
0
−1
10
0
10
1
10
r
in
(spikes/sec)
2
10
Fig. 1 Rate of vesicle release as a function of presynaptic
firing rate for various presynaptic Fano factors. The rate
of vesicle release, rx, is an increasing function of presynap-
tic firing rate, rin. Vesicle release is slower for the irregular
spiking model than for the Poisson and regular spiking model.
refractory period is an exponentially distributed ran-
dom variable with mean τu, so that vesicle recovery is
Poisson in nature.
We are interested in how the statistics of the presy-
naptic spike train determine the statistics of the synap-
tic response. The presynaptic statistics are quantified
using the presynaptic firing rate, rin, the presynaptic
auto-covariance function, Rin(τ ), and the Fano factor,
Fin(T ), of the number of presynaptic spikes during a
window of length T . Similarly, we quantify the statistics
of the synaptic response using the mean rate of vesicle
release, rx, the auto-covariance of vesicle release, Rx(τ ),
and the Fano factor, Fx(T ), of the number of vesicles
released during a window of length T . We will espe-
cially focus on Fano factors over large time windows
and define Fin = limT →∞ Fin(T ), Fx = limT →∞ Fx(T )
accordingly. See Methods for more details.
We begin by considering the effect of Fin on the
mean rate of vesicle release, rx. We then examine the
dependence of Fx(T ) on the length, T , of the time win-
dow over which vesicle release events are counted. Fi-
nally, we show that short term synaptic depression pro-
motes Poisson-like responses to non-Poisson presynap-
tic inputs.
3.1 Irregular presynaptic spiking reduces the rate at
which neurotransmitter vesicles are released
We first briefly investigate the dependence of the rate
of vesicle release, rx, on the rate and variability of the
presynaptic spike train, as measured by rin and Fin re-
Steven Reich, Robert Rosenbaum
8
A
)
t
(
I
)
t
(
m
)
t
(
x
3
2
1
0
2
1
0
B
3
2
1
0
2
1
0
0
0.5
1
0
0.5
t (sec)
1
Fig. 2 Synaptic response to an irregular and a regular presynaptic spike train. A) An irregular spike train, I(t), drives
a depressing synapse. Each vesicle release event is indicated by a vertical bar with height indicating the number of vesicles
released (here, all events release just one vesicle). Each time a vesicle is released, the number of available vesicles, m(t), is
decremented accordingly. Vesicle recovery increments m(t) and occurs randomly in time (vesicle recovery events indicated by
filled triangles). B) Same as (A) except for a more regular presynaptic spike train. Note that, even though the same number of
spikes occur in (A) and (B), the irregular spike train is less effective in releasing vesicles. This occurs because all vesicles are
depleted by the first few spikes in a burst and subsequent spikes in that burst are unable to release vesicles. For illustrative
purposes, we set M = 3 in this figure.
spectively. Vesicle release rate generally increases with
rin, but saturates to rx = M/τu whenever prrin ≫ 1/τu
since synapses are depleted in this regime (Fig. 1).
When presynaptic spike times occur as a Poisson
process (so that Fin = 1), the mean rate of vesicle
release is given by rx = M prrin/(prrinτu + 1) [22,19,
43]. Interestingly, vesicle release is slower for more ir-
regular presynaptic spiking and faster for more regular
presynaptic spiking even when presynaptic spikes ar-
rive at the same mean rate (Fig. 1, also see [18]). This
can be understood by noting that, for the irregular in-
put model, spikes arrive in bursts of higher firing rate
followed by durations of lower firing rate. Vesicles are
depleted by the first few spikes in a burst and subse-
quent spikes in that burst are ineffective and therefore
essentially "wasted" spikes (Fig. 2A). When presynap-
tic spikes arrive more regularly, more vesicles are re-
leased on average (Fig. 2B).
3.2 Variability in the number of vesicles released in a
time window decreases with window size
We now consider how the the variability of the synap-
tic response to a presynaptic input depends on the
timescale over which this variability is measured. We
quantify the variability of the synaptic response us-
ing the Fano factor, Fx(T ), which is defined to be the
variance-to-mean ratio of the number of vesicles re-
leased in a time window of length T (see Methods)
and can be calculated from an integral of the auto-
covariance function, Rx(τ ), using Eq. (3).
The auto-covariance of a Poisson presynaptic spike
train is simply a delta function at the origin, Rin(τ ) =
rinδ(τ ), and the Fano factor over any window size is
therefore equal to one, Fin(T ) = 1 (Figs. 3C and 4C).
The auto-covariance of the synaptic response when presy-
naptic inputs are Poisson consists of a delta function
at the origin surrounded by a double-sided exponential
with a negative peak (see Eq. (4) and Fig. 3D) that de-
cays with timescale τ0 = τu/(1 + prrinτu). The fact that
the auto-covariance is negative away from τ = 0 implies
that the Fano factor, Fx(T ), is monotonically decreas-
ing in the window size, T (see Eq. (7) and Fig. 4D).
For small T , the mass of the delta function at the ori-
gin dominates the integral in Eq. (3) so that the Fano
factor is approximately equal to the ratio of this mass
to the mean rate, rx, at which vesicles are released. As
T increases, the negative mass of the exponential peak
subtracts from the positive contribution of the delta
function and decreases the Fano factor. In particular,
Fx(T ) ≈ D − ET + O(T 2) where Drx is the mass of
the delta function in Rx(τ ) and −Erx is the peak of
the exponential in Rx(τ ) (see Eqs. (5) and (6)). As T
continues to increase, Fx(T ) monotonically decreases
towards its limit, Fx := limT →∞ Fx(T ) = D − 2Eτ0.
Thus, short term synaptic depression converts a Fano
factor that is constant with respect to window size into
one that decreases with window size (Fig. 4C,D).
Short term depression and neuronal variability
A
300
)
τ
(
n
i
R
200
100
= 20
F
in
C
= 1
F
in
1
0
−1
9
= 0. 1
F
in
E
0
−200
−400
−2
−1
0
1
2
−0.5
0
0.5
−0.2
0
0.2
B
)
τ
(
x
R
0
−2
−4
−2
−1
0
1
2
τ (sec)
D
0
−10
−20
−0.5
F
0
−20
−40
0
0.5
−0.2
0
0.2
τ
(sec)
τ
(sec)
Fig. 3 Presynaptic auto-covariance functions and auto-covariance of vesicle release for three input models. Auto-
covariance functions of presynaptic spike trains (top row) and the synaptic response they evoke (bottom row) for three
different presynaptic input models: A,B)irregular (Fin = 20), C,D) Poisson (Fin = 1) and E,F) regular (Fin = 0.1). Each
auto-covariance function has a Dirac delta function at the origin that is not depicted here. Dotted line in (D) is from the
approximation in Eq. (4) and solid line is from exact calculation obtained using numerics for the regular input model with
θ = 1 (see Methods), but the two are nearly indistinguishable. Rin(τ ) has units (spikes/sec)2 and Rx(τ ) has units (vesicles/sec)2.
Short term depression introduces negative temporal correlations even when presynaptic spike trains are temporally uncorrelated
(C,D) or positively correlated (A,B).
When presynaptic spike times are not Poisson, the
statistics of the postsynaptic response cannot be de-
rived analytically using the methods utilized for the
Poisson input model. Instead, we use the fact that the
synapse model can be represented using a continuous
time Markov chain, which can be analyzed to derive
expressions for the response statistics in terms of an
infinitesimal generator matrix (see Methods).
Irregular presynaptic spiking (i.e., inputs with Fin >
1) is achieved by varying the rate of presynaptic spik-
ing randomly in time to produce a doubly stochas-
tic Poisson process (see Methods). For this model, the
input auto-covariance is a delta function at the ori-
gin surrounded by an exponential peak (see Eq. 10
and Fig. 3A). The input Fano factor therefore increases
with window size (see Eq. 11 and Fig. 4A). The posi-
tive temporal correlations exhibited in the input auto-
covariance function are canceled by the temporal de-
correlating effects of short term synaptic depression [26,
25,24]. For the parameters chosen in this study, this
de-correlation outweighs the positive presynaptic cor-
relations so that the auto-covariance function of the
response is negative away from τ = 0 (Fig. 3B), al-
though parameters can also be chosen so that temporal
correlations in the response are small and positive [25].
As with the Poisson input model, negative temporal
correlations cause the response Fano factor to decrease
with window size (Fig. 4B). Thus short term synaptic
depression and stochastic vesicle dynamics can convert
a presynaptic Fano factor that increases with window
size into one that decreases.
Regular presynaptic spiking is achieved by generat-
ing a renewal process with gamma-distributed interspike-
intervals. The input auto-covariance function for this
model exhibits temporal oscillations (Eq. (15) and Fig. 3E)
and the Fano factor generally decreases with window
size (Fig. 4E). Perhaps unsurprisingly, the auto-covariance
function of the synaptic response exhibits oscillations
and the response Fano factor decreases with window
size (Figs. 3F and 4F).
For all three input models, the variability of the
synaptic response is larger over shorter time windows
and smaller over larger time windows. A postsynap-
tic neuron that is in an excitable regime will generally
respond most effectively to inputs that exhibit more
variability over short time windows [46,47,39,38]. In
addition, rate coding is often more efficient when spike
counts over larger time windows are less variable [64].
Thus, the dependence of Fx(T ) on window size is espe-
cially efficient for the neural transmission of rate-coded
information [24].
In addition to the temporal dependence of Fx(T )
introduced by short term depression, note that the re-
sponse Fano factor for the irregular input model is sub-
10
A
)
T
(
n
Fi
20
15
10
5
B
1.5
1.4
)
T
(
Fx
1.3
0
=20
F
in
C
=1
F
in
2
1
0
Steven Reich, Robert Rosenbaum
= 0. 1
F
in
E
1
0.8
0.6
0.4
0.2
0
5
10
15
0
1
2
0
0.5
1
D
1.2
1
0.8
F
1.2
1.1
1
0.9
10
5
T (sec)
15
0
1
T (sec)
2
0
0.5
T (sec)
1
Fig. 4 Presynaptic and response Fano factor as a function of window size for three input models. Presynaptic and
response Fano factors, Fin(T ) and Fx(T ), as a function of the window size over which inputs or vesicles are counted (see
Methods), obtained by applying Eq. (1) to the auto-covariance functions in Fig. 3. Short term depression causes response Fano
factor to decrease with window size even when presynaptic Fano factor increases with window size (A,B) or is independent of
window size (C,D). Also, response Fano factors are near 1 even when presynaptic Fano factors are not (B and F).
stantially smaller than the input Fano factor (Fig. 4).
Conversely, the response Fano factor for the regular in-
put model is larger than the input Fano factor (Fig. 4).
For both models, the response Fano factor is substan-
tially nearer to 1 than the input Fano factor. We explain
this phenomenon next.
3.3 Depleted synapses exhibit Poisson-like variability
even when presynaptic inputs are highly non-Poisson
We now investigate the dependence of the variability
in synaptic response on the rate and variability of the
presynaptic input. Since we have already discussed the
dependence of Fx(T ) on T above, we will focus here
on the Fano factor calculated over long time windows,
Fx = limT →∞ Fx(T ).
We first consider parameter regimes where the ef-
fective rate of presynaptic inputs is much slower than
the rate of vesicle recovery (prrin ≪ 1/τu). In such
a regime, each contact is likely to recover between two
consecutive presynaptic spikes and therefore all M con-
tacts are likely to have a vesicle ready to release when
each spike arrives (Fig. 6A). In this limit, the number
of vesicles released by each spike is an independent bi-
nomial variable with mean hwj i = prM and variance
var(wj ) = M pr(1 − pr). The number, Nx(T ), of vesi-
cles released in a time window of length T can then be
represented as a sum of Nin(T ) independent binomial
random variables (i.e., a random sum). The mean of
this sum is given by hNx(T )i = hNin(T )ihwji, which
implies that rx = M prrin in this limit. Similarly, the
variance of this sum is given by [31] var(Nx(T )) =
hNin(T )ivar(wj ) + hwj i2var(Nin(T )), which implies
lim
rin→0
Fx(T ) = hwj iFin(T ) + var(wj )/hwj i
(16)
= 1 + pr(M Fin(T ) − 1).
Eq. (16) is verified for the Poisson input model by tak-
ing rin → 0 in Eqs. (7). For the irregular and regular
input models, Eq. (16) should be interpreted heuristi-
cally, as it was derived heuristically. A counterexample
to Eq. (16) for the irregular input model can be con-
structed by fixing rf and τf , then letting τs → ∞ and
rs → 0 to achieve the rin → 0 limit. In this case, our
assumption that each contact is increasingly likely to
recover between two consecutive spikes is violated and
Eq. (16) is not valid (not pictured). Regardless, we ver-
ify numerically that Eq. (16) is accurate when rin is
decreased toward zero while keeping Fin fixed (Fig. 5).
We now discuss the statistics of the postsynaptic re-
sponse when the effective presynaptic spiking is much
faster than vesicle recovery (prrin ≫ 1/τu). In such a
regime, incoming spikes occur much more frequently
than recovery events and synapses becomes depleted.
As a result, the number of vesicles released over a long
time window is determined predominantly by the num-
ber of recovery events in that time window and largely
independent from the number of presynaptic spikes (Fig. 6B)
Short term depression and neuronal variability
11
A
)
t
(
I
)
t
(
m
)
t
(
x
3
2
1
0
3
2
1
0
B
3
2
1
0
2
1
0
0
1
2
3
4
5
0
0.5
t (sec)
1
Fig. 6 Vesicle release dynamics at low and high presynaptic firing rates. A) At low presynaptic rates, vesicles are
recovered (i.e., m(t) returns to M = 3) between presynaptic spikes. Thus, the number of vesicles released by each presynaptic
spike is approximately an independent binomial random variable with mean prM and variance prM (1 − M ). B) At high
presynaptic rates, vesicles are released almost immediately after they are recovered. Thus, the number of vesicles released over
a time window of length T is approximately a Poisson random variable with mean and variance T M/τu.
[19,43]. The synaptic response therefore inherits the
Poisson statistics of the recovery events so that
lim
rin→∞
Fx(T ) = 1.
For the Poisson input model, this limit can be made
more precise in the T → ∞ limit by expanding Eq. (8)
in terms of the parameter α = 1/(prrinτu) to obtain
Fx = 1 − 2α + 4α2 + O(α3)
(17)
which converges to 1 as rinτu → ∞. For the irregular
and regular input models, we verify in Fig. 5 that Fx →
1 when rin is increased while keeping Fin fixed.
The time constant, τu, at which a synapse recov-
ers from short term depression has been measured in a
number of experimental studies and is often found to
be several hundred milliseconds [57,59,34,23,22,28,41].
Therefore, for even moderate presynaptic firing rates,
synapses are often in a highly depleted state. As dis-
cussed above, this promotes Poisson-like variability in
the synaptic response. This provides one possible mech-
anism through which irregular Poisson-like firing can be
sustained in neuronal populations [11].
4 Discussion
We used continuous time Markov chain methods to de-
rive the response statistics of a stochastic model of short
term synaptic depression with three different presynap-
tic input models. We then used this analysis to un-
derstand how the mean presynaptic firing rate and the
variability of presynaptic spiking interact with synaptic
dynamics to determine the mean rate of vesicle release
and variability in the number of vesicles released. This
analysis revealed a number of fundamental, qualitative
dependencies of response statistics on presynaptic spik-
ing statistics. Some of the dependencies have been pre-
viously noted in the literature and some have not.
The number of vesicles released over a time win-
dow is smaller for irregular inputs than for more reg-
ular inputs (Figs. 1 and 2) given the same number of
presynaptic spikes. Thus, regular presynaptic spiking
is more efficient at driving synapses. This mechanism
competes with a well-known property of excitable cells:
that they are driven more effectively by irregular, posi-
tively correlated synaptic input currents [46,47,39,38].
In addition, a population of presynaptic spike trains
drives a postsynaptic neuron more efficiently when the
population-level activity is more irregular, for example
due to pairwise correlations [19]. Together, these results
suggest that a postsynaptic neuron is most efficiently
driven by presynaptic populations that exhibit small or
negative auto-correlations, but positive pairwise cross-
correlations.
Our model predicts that the de-correlating effects
of short term depression and stochastic vesicle dynam-
ics can produce negative temporal auto-correlations in
the synaptic response even when presynaptic spiking
is temporally uncorrelated or positively correlated, in
agreement with previous studies [25,19]. This yields a
response Fano factor that decreases with window size,
as observed in some recorded data [30]. We note, though,
F
= 20
in
−1
10
0
10
1
10
2
10
F
= 1
in
−1
10
0
10
1
10
2
10
F
= 0.1
in
12
A
Fx
B
64
16
4
1
3
1.75
Fx
1
0.6
1
C
Fx
0.75
0.56
−1
10
0
10
1
10
2
10
r
in
(spikes/sec)
Fig. 5 Response Fano factor as a function of presynap-
tic firing rate for three input models. Response Fano fac-
tors calculated over large windows for A) the irregular input
model B) Poisson input model and C) regular input model.
Fano factors approach 1 at high presynaptic firing rates re-
gardless of the presynaptic Fano factor (triangle on right is
placed at Fx = 1). At low presynaptic firing rates, response
Fano factors approach the value given in Eq. (16) (indicated
by triangle on left). Dotted line in (B) is from closed form ap-
proximation in Eq. (8) and dashed line is from the expansion
given in Eq. (17).
Steven Reich, Robert Rosenbaum
that some parameter choices can yield positively a cor-
related synaptic response when presynaptic inputs are
positively correlated [25] and neuronal membrane dy-
namics can introduce positive correlations to a post-
synaptic spiking response even when synaptic currents
are not positively correlated in time [37]. This is con-
sistent with several studies showing positive temporal
correlations in recorded spike trains [4,15,3,11].
We predict that moderate or high firing rates can in-
duce a Poisson-like synaptic response even when presy-
naptic inputs are non-Poisson (Fig. 5 and [18]). This is
because even moderate firing rates can deplete synapses
and depleted synapses inherit the Poisson-like variabil-
ity of synaptic vesicle recovery (Fig. 6B and [19,43]).
At lower firing rates, short term depression and synap-
tic variability can increase or decrease Fano factor. For
example, in Fig. 5B, the response Fano factor is larger
than the presynaptic Fano factor (Fin = 1) at low firing
rates, decreases at higher firing rates, then approaches
Fx = 1 at higher firing rates. This complex dependence
of firing rate on Fano factor might be related to the
stimulus dependence of Fano factors observed in sev-
eral cortical brain regions [11].
Our conclusion that fast presynaptic spiking causes
Poisson-like variability in the synaptic response relied
on the assumption that vesicle recovery times are ex-
ponentially distributed. The exponential distribution
is a justifiable choice for recovery times only if recov-
ery times obey a memoryless property: having already
waited t units of time for a recovery event, the proba-
bility of waiting an additional s units of time does not
depend on t. The precise mechanics of vesicle re-uptake
and docking determine whether this is an appropriate
assumption. If recovery times have a different probabil-
ity distribution, then the synaptic response will inherit
the properties of this distribution at high presynaptic
firing rates instead of inheriting the Poisson-like nature
of exponentially distributed recovery times.
Previous methods have been developed to analyze
the synaptic response of the model used here. In [18,
24], the model restricted to the M = 1 case is ana-
lyzed for presynaptic spike trains that are renewal pro-
cesses. This includes the Poisson and the regular input
model discussed here, but excludes the irregular input
model in which the spike train is a non-renewal inho-
mogeneous Poisson process. In [43] approximations are
obtained for the case where the presynaptic spike train
is an inhomogeneous Poisson process, but the approx-
imation is only valid when the rate-modulation of the
Poisson process is small compared to the average fir-
ing rate. Thus, these approximations are only valid for
the irregular input model when rf − rs ≪ rs. Other
studies [32,36] use a deterministic synapse model that
Short term depression and neuronal variability
13
implicitly treats the number of available vesicles as a
continuous rather than a discrete quantity. This deter-
ministic model represents the trial average of the model
considered here and can vastly underestimate the vari-
ability of a synaptic response [43].
A more detailed synapse model allows for multiple
docking sites at a single contact [62,19]. This model can
yield different response properties than the model used
here in certain parameter regimes [19]. Even though this
more detailed model can be represented as a continuous-
time Markov chain, the analysis of this model would be
significantly more complex than the analysis considered
here since it would be necessary to keep track of the
number of readily releasable vesicles at each contact
separately. This would result in a Markov chain with
K × N M states where M is the number of contacts, N
is the number of docking sites per contact and K is the
number of states used for the presynaptic input model
(K = 1 for the Poisson input model, K = 2 for the
irregular input model, and K = θ for the regular input
model).
To quantify the synaptic response to a presynaptic
spike train, we focused on the statistics of the num-
ber of vesicles released in a time window. Postsynaptic
neurons observe changes in synaptic conductance in re-
sponse to presynaptic spikes. The synaptic conductance
are often modeled in such a way that they can be easily
derived from our process x(t) through a convolution:
0 x(t − s)α(s)ds where g(t) is the synaptic con-
ductance elicited by a presynaptic spike train and α(s)
is a kernel representing the characteristic postsynaptic
conductance elicited by the release of a single neuro-
transmitter vesicle. Since this mapping is linear, the
statistics of g(t) can easily be derived in terms of the
statistics of x(t) [53,43].
g(t) =R t
References
1. L F Abbott and Wade G Regehr. Synaptic computation.
Nature, 431(7010):796–803, October 2004.
2. L F Abbott, J A Varela, K Sen, and S B Nelson.
Synaptic depression and cortical gain control. Science,
275(5297):220–224, January 1997.
3. R. Baddeley, L.F. Abbott, M.C.A. Booth, F. Sengpiel,
T. Freeman, E.A. Wakeman, and E.T. Rolls. Responses
of neurons in primary and inferior temporal visual cor-
tices to natural scenes. Proceedings of the Royal Society
of London. Series B: Biological Sciences, 264(1389):1775–
1783, 1997.
4. W. Bair, C. Koch, W. Newsome, and K. Britten. Power
spectrum analysis of bursting cells in area mt in the be-
having monkey. The Journal of neuroscience, 14(5):2870–
2892, 1994.
5. F. Barbieri and N. Brunel. Can attractor network mod-
els account for the statistics of firing during persistent
activity in prefrontal cortex? Front in Neurosci, 2(1):114,
2008.
6. Tiago Branco and Kevin Staras. The probability of neu-
rotransmitter release: variability and feedback control at
single synapses. Nature Rev Neurosci, 10(5):373–383, May
2009.
7. P. C. Bressloff. Mean-field theory of globally cou-
pled integrate-and-fire neural oscillators with dynamic
synapses. Phys. Rev. E, 60:2160–2170, Aug 1999.
8. K.H. Britten, M.N. Shadlen, W.T. Newsome, J.A.
Movshon, et al. Responses of neurons in macaque mt to
stochastic motion signals. Visual neuroscience, 10:1157–
1157, 1993.
9. G.T. Buracas, A.M. Zador, M.R. DeWeese, and T.D. Al-
bright. Efficient discrimination of temporal patterns by
motion-sensitive neurons in primate visual cortex. Neu-
ron, 20(5):959–969, 1998.
10. F.S. Chance, S.B. Nelson, and L.F. Abbott. Synaptic
depression and the temporal response characteristics of
v1 cells. J Neurosci, 18(12):4785, 1998.
11. M.M. Churchland, M.Y. Byron, J.P. Cunningham, L.P.
Sugrue, M.R. Cohen, G.S. Corrado, W.T. Newsome,
A.M. Clark, P. Hosseini, B.B. Scott, et al. Stimulus onset
quenches neural variability: a widespread cortical phe-
nomenon. Nature neuroscience, 13(3):369–378, 2010.
12. Daniel L Cook, Peter C Schwindt, Lucinda A Grande,
and William J Spain. Synaptic depression in the local-
ization of sound. Nature, 421(6918):66–70, January 2003.
13. D Cox. Renewal Theory. London: Methuen and Co, 1962.
14. D Cox and V Isham. Point processes. London: Chapman
and Hall, 1980.
15. Y. Dan, J.J. Atick, and R.C. Reid. Efficient coding of
natural scenes in the lateral geniculate nucleus: experi-
mental test of a computational theory. The Journal of
Neuroscience, 16(10):3351–3362, 1996.
16. P. Dayan and L.F. Abbott. Theoretical Neurosci: Compu-
tational and mathematical modeling of neural systems. Tay-
lor & Francis, 2001.
17. J de la Rocha and R Moreno. Correlations modulate
the non-monotonic response of a neuron with short-term
plasticity. Neurocomputing, 2004.
18. J de la Rocha and A Nevado. Information transmission
by stochastic synapses with short-term depression: neural
coding and optimization. Neurocomputing, 2002.
19. Jaime de la Rocha and N´estor Parga. Short-term synaptic
depression causes a non-monotonic response to correlated
stimuli. J Neurosci, 25(37):8416–8431, September 2005.
20. Jaime de la Rocha and N´estor Parga. Thalamocortical
transformations of periodic stimuli: the effect of stimu-
lus velocity and synaptic short-term depression in the
vibrissa-barrel system. J Comput Neurosci, 25(1):122–
140, August 2008.
21. A.A. Faisal, L.P.J. Selen, and D.M. Wolpert. Noise in
the nervous system. Nature Rev Neurosci, 9(4):292–303,
2008.
22. G. Fuhrmann, I. Segev, H. Markram, and M. Tsodyks.
Coding of temporal information by activity-dependent
synapses. J Neurophys, 87(1):140, 2002.
23. M. Galarreta and S. Hestrin.
Frequency-dependent
synaptic depression and the balance of excitation and
inhibition in the neocortex. Nature Neurosci, 1:587–594,
1998.
24. M.S. Goldman. Enhancement of information transmis-
sion efficiency by synaptic failures. Neural Comput,
16(6):1137–1162, 2004.
25. M.S. Goldman, P. Maldonado, and LF Abbott. Re-
dundancy reduction and sustained firing with stochas-
tic depressing synapses. The Journal of Neuroscience,
22(2):584–591, 2002.
14
Steven Reich, Robert Rosenbaum
26. M.S. Goldman, S.B. Nelson, and LF Abbott. Decorrela-
tion of spike trains by synaptic depression. Neurocomput-
ing, 26:147–153, 1999.
27. Lucinda A Grande and William J Spain. Synaptic de-
pression as a timing device. J Physiol, 20:201–210, June
2005.
28. Jesse E Hanson and Dieter Jaeger. Short-term plasticity
shapes the response to simulated normal and parkinso-
nian input patterns in the globus pallidus. J Neurosci,
22(12):5164–5172, June 2002.
29. A. Harsch and H.P.C. Robinson. Postsynaptic variability
of firing in rat cortical neurons: the roles of input synchro-
nization and synaptic nmda receptor conductance. The
Journal of Neuroscience, 20(16):6181–6192, 2000.
30. P. Kara, P. Reinagel, and R.C. Reid. Low response vari-
ability in simultaneously recorded retinal, thalamic, and
cortical neurons. Neuron, 27(3):635–646, 2000.
31. S. Karlin and H.M. Taylor. A first course in stocastic
processes. Academic press, 1975.
32. B Lindner, D Gangloff, A Longtin, and J E Lewis.
Broadband Coding with Dynamic Synapses. J Neurosci,
29(7):2076–2087, February 2009.
33. A Litwin-Kumar and B Doiron. Slow dynamics and high
variability in balanced cortical networks with clustered
connections. Nat Neurosci, 2012.
34. H. Markram, Y. Wang, and M. Tsodyks. Differential sig-
naling via the same axon of neocortical pyramidal neu-
rons. Proc Natl Acad Sci USA, 95(9):5323, 1998.
35. C.J. McAdams and J.H.R. Maunsell. Effects of attention
on the reliability of individual neurons in monkey visual
cortex. Neuron, 23(4):765–773, 1999.
36. Matthias Merkel and Benjamin Lindner. Synaptic filter-
ing of rate-coded information. Phys Rev E, 81(4), April
2010.
37. R. Moreno-Bote and N. Parga. Auto-and crosscorrel-
ograms for the spike response of leaky integrate-and-
fire neurons with slow synapses. Physical review letters,
96(2):28101, 2006.
38. R. Moreno-Bote and N. Parga. Response of integrate-
and-fire neurons to noisy inputs filtered by synapses with
arbitrary timescales: Firing rate and correlations. Neural
Computation, 22(6):1528–1572, 2010.
39. R. Moreno-Bote, A. Renart, and N. Parga. Theory of
input spike auto- and cross-correlations and their effect
on the response of spiking neurons. Neural Computation,
20(7):1651–1705, 2008.
41. M Rav-Acha, N Sagiv,
40. AM Oswald and NN Urban.
Interactions between be-
haviorally relevant rhythms and synaptic plasticity alter
coding in the piriform cortex. J Neurosci (In Press), 2012.
I Segev, H Bergman, and
Y Yarom. Dynamic and spatial features of the inhibitory
pallidal GABAergic synapses. J Neurosci, 135(3):791–
802, 2005.
42. R Rosenbaum. The transfer and propagation of correlated
neuronal activity. PhD thesis, University of Houston,
2011.
43. R Rosenbaum, J Rubin, and B Doiron. Short term synap-
tic depression imposes a frequency dependent filter on
synaptic information transfer. Public Library of Sci Com-
put Biol, 8(6), 2012.
44. Jason S Rothman, Laurence Cathala, Volker Steuber,
and R Angus Silver. Synaptic depression enables neu-
ronal gain control. Nature, 457(7232):1015–1018, Febru-
ary 2009.
45. Z Rotman, P Y Deng, and V A Klyachko. Short-Term
Plasticity Optimizes Synaptic Information Transmission.
J Neurosci, 31(41):14800–14809, October 2011.
46. E. Salinas and T.J. Sejnowski.
Impact of correlated
synaptic input on output firing rate and variability
in simple neuronal models.
Journal of Neuroscience,
20(16):6193, 2000.
47. E. Salinas and T.J. Sejnowski. Integrate-and-fire neurons
driven by correlated stochastic input. Neural computation,
14(9):2111–2155, 2002.
48. W Senn, H Markram, and M Tsodyks. An algorithm
for modifying neurotransmitter release probability based
on pre- and postsynaptic spike timing. Neural Comput,
13(1):35–67, January 2001.
49. M.N. Shadlen and W.T. Newsome. Noise, neural codes
and cortical organization. Findings and Current Opinion
in Cognitive Neuroscience, 4:569–579, 1998.
50. M.N. Shadlen and W.T. Newsome. The variable dis-
charge of cortical neurons: implications for connectivity,
computation, and information coding. Journal of Neuro-
science, 18(10):3870–3896, 1998.
51. W.R. Softky and C. Koch. The highly irregular firing of
cortical cells is inconsistent with temporal integration of
random EPSPs. Journal of Neuroscience, 13(1):334–350,
1993.
52. C.F. Stevens, A.M. Zador, et al. Input synchrony and the
irregular firing of cortical neurons. Nature neuroscience,
1(3):210–217, 1998.
53. T. Tetzlaff, S. Rotter, E. Stark, M. Abeles, A. Aertsen,
and M. Diesmann. Dependence of neuronal correlations
on filter characteristics and marginal spike train statis-
tics. Neural Comput, 20(9):2133–2184, 2008.
54. DJ Tolhurst, JA Movshon, and AF Dean. The statistical
reliability of signals in single neurons in cat and monkey
visual cortex. Vision research, 23(8):775–785, 1983.
55. M Tsodyks, K Pawelzik, and H Markram. Neural net-
works with dynamic synapses. Neural Comput, 10(4):821–
835, May 1998.
56. M. Tsodyks, A. Uziel, H. Markram, et al. Synchrony gen-
eration in recurrent networks with frequency-dependent
synapses. J Neurosci, 20(1):825–835, 2000.
57. M V Tsodyks and H Markram. The neural code be-
tween neocortical pyramidal neurons depends on neuro-
transmitter release probability. Proc Natl Acad Sci USA,
94(2):719–723, January 1997.
58. C. Van Vreeswijk and H. Sompolinsky. Chaos in neu-
ronal networks with balanced excitatory and inhibitory
activity. Science, 274(5293):1724–1726, 1996.
59. J A Varela, K Sen, J Gibson, J Fost, L F Abbott, and S B
Nelson. A quantitative description of short-term plastic-
ity at excitatory synapses in layer 2/3 of rat primary vi-
sual cortex. J Neurosci, 17(20):7926–7940, October 1997.
60. D. Vere-Jones. Simple stochastic models for the release of
quanta of transmitter from a nerve terminal. Australian
& New Zealand Journal of Statistics, 8(2):53–63, 1966.
61. C. Vreeswijk and H. Sompolinsky. Chaotic balanced
state in a model of cortical circuits. Neural Computation,
10(6):1321–1371, 1998.
62. X J Wang. Fast burst firing and short-term synaptic
plasticity: a model of neocortical chattering neurons. J
Neurosci, 89(2):347–362, 1999.
63. A. Zador.
Impact of synaptic unreliability on the in-
formation transmitted by spiking neurons. J Neurophys,
79(3):1219, 1998.
64. E Zohary, MN Shadlen, and WT Newsome. Correlated
neuronal discharge rate and its implications for psy-
chophysical performance. Nature, 370(6485):140–3, Jul
1994.
65. R.S. Zucker and W.G. Regehr. Short-term synaptic plas-
ticity. Annual Rev of Phys, 64(1):355–405, 2002.
|
1610.05872 | 1 | 1610 | 2016-10-19T05:32:32 | Making brain-machine interfaces robust to future neural variability | [
"q-bio.NC",
"stat.ML"
] | A major hurdle to clinical translation of brain-machine interfaces (BMIs) is that current decoders, which are trained from a small quantity of recent data, become ineffective when neural recording conditions subsequently change. We tested whether a decoder could be made more robust to future neural variability by training it to handle a variety of recording conditions sampled from months of previously collected data as well as synthetic training data perturbations. We developed a new multiplicative recurrent neural network BMI decoder that successfully learned a large variety of neural-to- kinematic mappings and became more robust with larger training datasets. When tested with a non-human primate preclinical BMI model, this decoder was robust under conditions that disabled a state-of-the-art Kalman filter based decoder. These results validate a new BMI strategy in which accumulated data history is effectively harnessed, and may facilitate reliable daily BMI use by reducing decoder retraining downtime. | q-bio.NC | q-bio | Making brain-machine interfaces robust to future neural variability
David Sussillo1,5,*, Sergey D. Stavisky2,*, Jonathan C. Kao1,*, Stephen I Ryu1,7,
Krishna V. Shenoy1,2,3,4,5,6
*Equal Contribution
1Electrical Engineering Department
2Neurosciences Graduate Program
3Neurobiology and Bioengineering Departments
4Bio-X Program
5Stanford Neurosciences Institute
6Howard Hughes Medical Institute at Stanford University
Stanford University, Stanford, 94305 CA, USA.
7Palo Alto Medical Foundation
Palo Alto, 94301 CA, USA
Contact: [email protected]
arXiv pre-print, 18 October 2016
ABSTRACT
A major hurdle to clinical translation of brain-machine interfaces (BMIs) is that current
decoders, which are trained from a small quantity of recent data, become ineffective
when neural recording conditions subsequently change. We tested whether a decoder
could be made more robust to future neural variability by training it to handle a variety of
recording conditions sampled from months of previously collected data as well as
synthetic training data perturbations. We developed a new multiplicative recurrent
neural network BMI decoder that successfully learned a large variety of neural-to-
kinematic mappings and became more robust with larger training datasets. When tested
with a non-human primate preclinical BMI model, this decoder was robust under
conditions that disabled a state-of-the-art Kalman filter based decoder. These results
validate a new BMI strategy in which accumulated data history is effectively harnessed,
and may facilitate reliable daily BMI use by reducing decoder retraining downtime.
Page 1
Brain-machine interfaces (BMIs) can restore motor function and communication to
people with paralysis1,2. Progress has been particularly strong towards enabling two-
dimensional computer cursor control, which may allow versatile communications
prostheses3–5. Cursor control performance has approached that of the native hand in
recent macaque studies6,7, but this was done under favorable laboratory conditions
where neural recordings are often stable both during and across BMI sessions8–11. In
contrast to these pre-clinical studies, one of the major challenges impeding BMI use by
human clinical trial participants is the high degree of within- and across-day variability in
neural recording conditions (Fig. 1a)12–16. We use the term "recording condition" to
broadly encompass the combination of factors which together determine the relationship
between observed neural activity and intended kinematics. These factors include the
relative position of the electrodes and surrounding neurons (diagrammed in Fig 1b),
variability in sensor properties such as impedance or wiring quality, noise sources, and
biological factors such as cognitive state or medications. Existing neural decoding
algorithms are poorly suited to handle variability in recording condition, resulting in
intermittent performance and a need for frequent decoder retraining4,5,13,17.
The clinical viability of BMIs would be much improved by making decoders robust to
recording condition changes18,19, and several recent studies have begun to focus on this
problem. We can broadly divide the conditions that a BMI will encounter into one of two
types: 1) conditions that are completely different from what has been previously
encountered, and 2) conditions that share some commonality with ones previously
encountered. For existing BMI methods, both of these situations necessitate some
Page 2
interruption of function while the decoder is updated to handle the new condition. One
strategy for minimizing this interruption is to use adaptive decoders, which update their
parameters based on new data collected during the BMI's use (rather than collecting
new training data for a de novo decoder) to try to better match the new recording
condition 4,10,20–29. In the first case, this is likely the best that can be done. But in the
second case, BMI interruption could in principle be avoided altogether by a decoder
capable of exploiting the similarities between the current and previously encountered
conditions (Fig. 1c).
We were motivated to try this complimentary strategy because chronic BMI systems do
typically encounter recording conditions in which there is some commonality with past
recording conditions8,10,13,14,27,28,30–32. Furthermore, these systems generate and store
months, or even years, of neural and kinematic data as part of their routine use. Almost
all of this past data is left unused in existing BMI systems: decoders are trained using
the most recently available data, typically from a block of calibration trials at the start of
that day's experiment, or from a recent previous experiment33. Using this historical data
would be difficult for most BMI decoders, as they are linear (e.g.,2,6). Linear decoders
are prone to underfitting heterogeneous training sets, such as those that might be
sampled from months of data. To overcome this limitation, an essential aspect of our
approach is to use a nonlinear and computationally 'powerful' decoder (i.e. one capable
of approximating any complex, nonlinear dynamical systems), which should be capable
of learning a diverse set of neural-to-kinematic mappings.
Page 3
Specifically, we tested whether one could gain traction on the decoder robustness
problem by exploiting this idle wealth of stored data using an artificial recurrent neural
network (RNN). We did this with a three-pronged approach. The first was the use of the
nonlinear RNN. The second was to train the decoder from many months of previously
recorded data. Third, to 'harden' the decoder against being too reliant on any given
pattern of inputs, we artificially injected additional variability into the data during decoder
training.
The fact that conventional state-of-the-art decoding methods, which tend to be linear or
at least of limited computational complexity34, work well for closed-loop BMI control of
2D cursors demonstrates that the model mismatch of assuming linear neural-to-
kinematic mappings is well tolerated for a given recording condition. Nevertheless, when
neural-to-kinematic mappings change over time, a conventional decoder trained on
many days' data is almost certainly not going to fully benefit from this abundance of
data. This is because it requires a powerful nonlinear algorithm to learn a set of different
context-dependent mappings, even if these individual mappings from neural firing rates
to kinematics were entirely linear (which they are not). Methods such as linear Kalman
filters can at best only learn an average mapping, 'splitting the difference' to reduce
error across the training set. This approach is not well-suited for most of the recording
conditions. We therefore developed a new BMI decoder using a nonlinear RNN variant
called the multiplicative recurrent neural network (MRNN) developed by Sutskever and
colleagues35 using their Hessian-free technique for training RNNs36. Several properties
of the MRNN architecture, which was originally used for character-level language
Page 4
modeling, make it attractive for this neural prosthetic application. First, it is recurrent,
and can therefore 'remember' state across time (e.g. during the course of a movement),
potentially better matching the time-varying, complex relationships between neural firing
rates and kinematics37,38. Second, its 'multiplicative' architecture increases
computational power by allowing the neural inputs to influence the internal dynamics of
the RNN by changing the recurrent weights (Fig. 2a). Loosely speaking, this allows the
MRNN to learn a 'library' of different neural-to-kinematic mappings that are appropriate
to different recording conditions. The MRNN was our specific choice of non-linear
method for learning a variety of neural-to-kinematic mappings, but this general approach
is likely to work well with many out-of-the-box RNN variants, such as a standard RNN
(e.g. 38) or LSTM39. It is also completely complementary to adaptive decoding.
We evaluated decoders using two non-human primates implanted with chronic
multielectrode arrays similar to those used in ongoing clinical trials. We first show that
training the MRNN with more data from previous recording sessions improves accuracy
when decoding new neural data, and that a single MRNN can be trained to accurately
decode hand reach velocities across hundreds of days. We next present closed-loop
results showing that an MRNN trained with many days' worth of data is much more
robust than a state-of-the-art Kalman filter based decoder (the Feedback Intention
Trained Kalman filter "FIT-KF"40) to two types of recording condition changes likely to be
encountered in clinical BMI use: the unexpected loss of signals from highly-informative
electrodes, and day-to-day changes. Finally, we show that this robustness does not
come at the cost of reduced performance under more ideal (unperturbed) conditions: in
Page 5
the absence of artificial challenges, the MRNN provides excellent closed-loop BMI
performance and slightly outperforms the FIT-KF. To our knowledge, this is the first
attempt to improve robustness by using a large and heterogeneous training dataset: we
used roughly two orders of magnitude more data than in previous closed-loop studies.
RESULTS
An MRNN can leverage large amounts of training data to improve decoder
performance
We first tested whether training the MRNN with many days' worth of data can improve
offline decoder performance across a range of recording conditions. This strategy was
motivated by our observation that the neural correlates of reaching – as recorded with
chronic arrays – showed day-to-day similarities (Supplementary Fig. 1). For a typical
recording session, the most similar recording came from a chronologically close day, but
occasionally the most similar recording condition was found in chronologically distant
data. MRNN decoders were able to exploit these similarities: Fig. 2b shows that as more
days' data (each consisting of ~500 point to point reaches) were used to train the
decoder, the accuracy of reconstructing reach velocities of test datasets increased
(positive correlation between number of training days and decoded velocity, r2 = 0.24 for
monkey R, r2 = 0.20 for monkey L, p < 0.001 for both). Two results from this analysis
were particularly encouraging. First, using more training data substantially increased the
decode accuracy for the 'hard' days that challenged decoders trained with only a few
days' data (e.g., test day 51 for monkey R). Second, this improvement did not come at
the cost of worse performance on the initially 'easy' test days. These results
demonstrate that larger training datasets better prepare the MRNN for a variety of
recording conditions, and that learning to decode additional recording conditions did not
Page 6
diminish the MRNN's capability to reconstruct kinematics under recording conditions
that it had already 'mastered'. There was not a performance versus robustness trade-
off.
We then tested whether the MRNN's computational capacity could be pushed even
further by training it using data from 154 (250) different days' recording sessions from
monkey R (L), which spanned 22 (34) months (Fig. 2c). Across all these recording
sessions' held-out test trials, the MRNN's offline decode accuracy was r2 = 0.81 ± 0.04
(mean ± s.d., monkey R) and r2 = 0.84 ± 0.03 (monkey L). For comparison, we tested
the decode accuracy of the FIT-KF trained in two ways: either specifically using
reaching data from that particular day ("FIT Sameday"), or trained on the same large
multiday training dataset ("FIT Long"). Despite the multitude of recording conditions that
the MRNN had to learn, on every test day each monkey's single MRNN outperformed
that day's FIT Sameday filter (monkey R: FIT-Sameday r2 = 0.57 ± 0.05, p < 0.001
signed-rank test comparing all days' FIT-Sameday and MRNN r2; monkey L: r2 = 0.52 ±
0.05, p < 0.001). Unsurprisingly, a linear FIT-KF did not benefit from being trained with
the same large multiday training set and also performed worse than the MRNN (monkey
R: FIT Long r2 = 0.56, p< 0.001 comparing all days' FIT Long to MRNN r2; monkey L: r2
= 0.46 ± 0.05, p < 0.001).
We note that, while these offline results demonstrate that the MRNN can learn a variety
of recording conditions, experiments are required to evaluate whether this type of
training leads to increased decoder robustness under closed-loop BMI cursor control. In
Page 7
closed-loop BMI cursor control, the BMI user updates his or her motor commands as a
result of visual feedback, resulting in distributions of neural activity that are different than
that of the training set. Thus, results from offline simulation and closed-loop BMI control
may differ 32,41–43. To this end, we next report closed-loop experiments that demonstrate
the benefit of this training approach.
Robustness to unexpected loss of the most informative electrodes
We next performed closed-loop BMI cursor control experiments to test the MRNN's
robustness to recording condition changes. The first set of experiments challenged the
decoder with an unexpected loss of inputs from multiple electrodes. The MRNN was
trained with a large corpus of hand reaching training data up through the previous day's
session (119 - 129 training days for monkey R, 212 - 230 days for monkey L). Then, its
closed-loop performance was evaluated on a Radial 8 Task while the selected
electrodes' input firing rates were artificially set to zero. By changing how many of the
most informative electrodes were dropped ('informative' as determined by their mutual
information with reach direction; see Methods), we could systematically vary the severity
of the challenge. Since this experiment was meant to simulate sudden failure of
electrodes during BMI use (after the decoder had already been trained), we did not
retrain or otherwise modify the decoder based on knowledge of which electrodes were
dropped. There were no prior instances of these dropped electrode sets having zero
firing rates in the repository of previously collected training data (Supplementary Fig. 2).
Thus, this scenario is an example of an unfamiliar recording condition (zero firing rates
on the dropped electrodes) having commonality with a previously encountered condition
(the patterns of activity on the remaining electrodes).
Page 8
We found that the MRNN was robust to severe electrode dropping challenges. It
suffered only a modest loss of performance after losing up to the best 3 (monkey R) or 5
(monkey L) electrodes (Fig. 3). We compared this to the electrode-dropped performance
of a FIT-KF decoder trained with hand reaching calibration data from the beginning of
that day's experiment6,40 ("FIT Sameday") by alternating blocks of MRNN and FIT
Sameday control in an "ABAB" interleaved experiment design. FIT Sameday decoder's
performance worsened dramatically when faced with this challenge. Across all
electrodes-dropped conditions, Monkey R acquired 52% more targets per minute using
the MRNN, while Monkey L acquired 92% more targets. Supplementary Video 2 shows
a side-by-side comparison of the MRNN and FIT-KF decoders with the 3 most
informative electrodes dropped.
Although the past datasets used to train the MRNN never had these specific sets of
highly important electrodes disabled, our technique of artificially perturbing the true
neural activity during MRNN training did generate training examples with reduced firing
rates on various electrodes (as well as examples with increased firing rates). The
MRNN had therefore been broadly trained to be robust to firing rate reduction on
subsets of its inputs. Subsequent closed-loop comparisons of MRNN electrode dropping
performance with and without this training data augmentation confirmed its importance
(Supplementary Fig. 3a). An additional offline decoding simulation, in which MRNN
decoders were trained with varying dataset sizes with and without training data
augmentation, further shows that both the MRNN architecture and its training data
Page 9
augmentation are important for robustness to electrode dropping (Supplementary Fig.
4). These analyses also suggest that large training dataset size does not impart
robustness to these particular recording condition changes, which is not surprising given
that the previous datasets did not include examples of these electrodes being dropped.
Robustness to naturally occurring recording condition changes
The second set of closed-loop robustness experiments challenged the MRNN with
naturally occurring day-to-day recording condition changes. In contrast to the highly
variable recording conditions encountered in human BMI clinical trials, neural recordings
in our laboratory setup are stable within a day and typically quite stable on the timescale
of days (Supplementary Fig. 1 and 10). Therefore, in order to challenge the MRNN and
FIT-KF decoders with greater recording condition variability, we evaluated them after
withholding the most recent several months of recordings from the training data. We
refer to this many month interval between the most recent training data day and the first
test day as the training data "gap" in these "stale training data" experiments. The gaps
were chosen arbitrarily within the available data, but to reduce the chance of outlier
results, we repeated the experiment with two different gaps for each monkey.
For each gap, we trained the MRNN with a large dataset consisting of many months of
recordings preceding the gap and compared it to two different types of FIT-KF
decoders. The "FIT Old" decoder was trained from the most recent available training
day (i.e. the day immediately preceding the gap); this approach was motivated under
the assumption that the most recent data were most likely to be similar to the current
day's recording condition. The "FIT Long" decoder was trained from the same multiday
Page 10
dataset used to train the MRNN and served as a comparison in which a conventional
decoder is provided with the same quantity of data as the MRNN. The logic underlying
this FIT Long approach was that despite the Kalman filter being ill-suited for fitting
multiple heterogeneous datasets, this 'averaged' decoder might still perform better than
the FIT-KF Old trained using a single distant day.
We found that the MRNN was the only decoder that was reliably usable when trained
with stale data (Fig. 4). FIT Old performed very poorly in both monkeys, failing
completely (defined as the monkey being unable to complete a block using the decoder,
see Methods) in 4/6 monkey R experimental sessions and 6/6 monkey L sessions. FIT
Long performed better than FIT Old, but its performance was highly variable – it was
usable on some test days but failed on others. In Monkey R, the across-days average
acquisition rate was 105% higher for the MRNN than FIT Long (p < 0.01, paired t-test).
Monkey L's MRNN did not perform as consistently well as Monkey R's, but nevertheless
demonstrated a trend of outperforming FIT Long (32% improvement, p = 0.45), in
addition to decidedly outperforming FIT Old, which failed every session. Although
monkey L's FIT Long outperformed the MRNN on one test day, on all other test days
FIT Long was either similar to, or substantially worse than, MRNN. Moreover, whereas
the MRNN could be used to control the cursor every day, FIT Long was not even
capable of acquiring targets on some days. Further tests of additional FIT Old decoders
confirmed that they generally perform poorly (Supplementary Fig. 5). The lack of
consistent usability by any of the FIT-KF decoders (Old or Long) demonstrates that
having access to a large repository of stale training data does not enable training a
Page 11
single Kalman filter that is robust to day-to-day variability in recording conditions. In
contrast, an MRNN trained with this large dataset was consistently usable.
To further demonstrate the consistency of these results, we performed offline
simulations in which we tested MRNN decoders on additional sets of training and test
datasets separated by a gap. Each set was non-overlapping with the others, and
together they spanned a wide range of each animal's research career. We observed the
same trends in these offline simulations: MRNNs trained with many previous days of
training data outperformed FIT Old and FIT Long decoders (Supplementary Fig. 6). In
these analyses we also dissected which components of our decoding strategy
contributed to the MRNN's robustness. We did this by comparing MRNNs trained with
varying numbers of days preceding the gap, with or without training data spike rate
perturbations. The results show that training using more data, and to a lesser extent
incorporating data augmentation (see also closed-loop comparisons in Supplementary
Fig. 3b), contributed to the MRNN's robustness to naturally occurring recording
condition changes.
High performance BMI using the MRNN decoder
Finally, we note that the MRNN's robustness to challenging recording conditions did not
come at the cost of reduced performance under more 'ideal' conditions, i.e. without
electrode dropping or stale training data. During the electrode dropping experiments, we
also evaluated the MRNN's closed-loop performance after being trained using several
months' data up through the previous day. In this scenario, the MRNN enabled both
monkeys to accurately and quickly control the cursor. Supplementary Movie 1 shows
Page 12
example cursor control using the MRNN. This data also allowed us to compare the
MRNN's performance to that of a FIT Sameday decoder in back-to-back "ABAB" tests.
Figure 5a shows representative cursor trajectories using each decoder, as well as under
hand control. Figure 5b shows that across nine experimental sessions and 4,000+ trials
with each decoder, Monkey R acquired targets 7.3% faster with the MRNN (619 ± 324
ms mean ± s.d. vs. 668 ± 469 ms, p < 0.01, rank-sum test). Monkey L acquired targets
10.8% faster with the MRNN (743 ± 390 ms vs. 833 ± 532 ms, p < 0.01, rank-sum test)
across 8 sessions and 2,500+ trials using each decoder. These online results
corroborate the offline results presented in Fig. 2c; both show that an MRNN trained
from many days' recording conditions outperforms the FIT Kalman filter trained from
training data collected at the start of the experimental session.
A potential risk inherent to a computationally powerful decoder such as the MRNN is
that it will overtrain to the task structure of the training data and fail to generalize to
other tasks. Most of our MRNN training data were from arm reaches on a Radial 8 Task
similar to the task used for evaluation (albeit with 50% further target distance). We
therefore also tested whether the MRNN enabled good cursor control on the Random
Target Task, in which the target could appear in any location in a 20 x 20 cm workspace
(Supplementary Fig. 7). Monkey R performed the Random Target Task on two
experimental sessions and averaged a 99.4% success rate, with mean distance-
normalized time to target of 0.068 s/cm. Monkey L performed one session of this task at
a 100% success rate with mean normalized time to target of 0.075 s/cm. To provide
context for these metrics, we also measured Random Target Task performance using
Page 13
arm control. Monkey R's arm control success rate was 100%, with 0.055 s/cm mean
normalized time to target, during the same experimental sessions as his MRNN
Random Target Task data. Monkey L's arm control success rate was 97.7%, with 0.055
s/cm mean normalized time to target, during one session several days following his
MRNN test.
DISCUSSION
We developed the MRNN decoder to help address a major problem hindering the
clinical translation of BMIs: once trained, decoders can be quickly rendered ineffective
due to recording condition changes. A number of complementary lines of research are
aimed at making BMIs more robust, including improving sensors to record from more
neurons more reliably (e.g.,44); decoding multiunit spikes10,30,45 or local field
potentials31,32,46 which appear to be more stable control signals than single unit activity;
and using adaptive decoders that update their parameters to follow changing neural-to-
kinematic mappings 4,10,20–29,47. Here we present the MRNN as a proof-of-principle of a
new and different approach: build a fixed decoder whose architecture allows it to be
trained to be inherently robust to recording condition changes it has previously
encountered as well as to new conditions that have some similarity to a previously
encountered condition.
We stress that all of these approaches are complementary in several respects. For
example, a decoder that is inherently more robust to neural signal changes, such as the
MRNN, would still benefit from improved sensors, could operate on a mix of input signal
types including single- and multiunit spikes and field potentials, and is especially well
Page 14
positioned to benefit from decoder adaptation. When performance degrades due to
recording condition changes, both supervised10,21–23,25,27,29 and unsupervised4,20,24,26
adaptive decoders need a period of time in which control is at least good enough that
the algorithm can eventually infer the user's intentions and use these to update its
neural-to-kinematic model. Improved robustness may "buy enough time" to allow the
decoder's adaptive component to rescue performance without interrupting prosthesis
use. Here we've demonstrated the MRNN's advantages over a state-of-the-art static
decoder, but comparing this strategy both against and together with adaptive decoding
remains a future direction.
We demonstrated the MRNN's robustness to two types of recording condition changes.
These changes were chosen because they capture key aspects of the changes that
commonly challenge BMI decoders during clinical use. The stale training data
experiments showed that the MRNN was usable under conditions where the passage of
time would typically require recalibration of conventional decoders such as the FIT-KF.
We do not mean to suggest that in a clinical setting one would want to – or would often
have to – use a BMI without any training data from the past several months. Rather, we
used this experimental design to model recording condition changes that can happen on
the timescale of hours in human BMI clinical trials13. Possible reasons for the greater
recording condition variability observed in human participants compared to non-human
primates include: more movement of the array relative to the human brain due to larger
cardiovascular pulsations and epidural space; greater variability in the state of the BMI
user (health, medications, fatigue, cognitive state); and more electromagnetic
Page 15
interference from the environment. The MRNN can take advantage of having seen the
effects of these sources of variability in previously accumulated data; it can therefore be
expected to become more robust over time as it builds up a "library" of neural-to-
kinematic mappings under different recording conditions.
The electrode dropping experiments, which demonstrated the MRNN's robustness to an
unexpected loss of high-importance electrodes, are important for two reasons. Firstly,
sudden loss of input signals (e.g., due to a electrode connection failure48,49), is a
common BMI failure mode that can be particularly deleterious to conventional BMI
decoders50. The MRNN demonstrates considerable progress in addressing this so-
called "errant unit" problem. Secondly, these results demonstrate that the MRNN trained
with artificially perturbed neural data can be relatively robust even to a recording
condition change that has not been encountered in past recordings.
The MRNN's robustness did not come at the cost of diminished performance under
more ideal conditions. This result is nontrivial given the robustness-focused decisions
that went into its design (e.g. perturbing the input spike trains in the training set).
Instead, we found that the MRNN was excellent under favorable conditions, slightly
outperforming a state-of-the-art same day trained FIT-KF decoder. Taken together,
these results demonstrate that the MRNN exhibits robustness to a variety of clinically
relevant recording condition changes, without sacrificing peak performance. These
advances may help to reduce the onerous need for clinical BMI users to collect frequent
retraining data.
Page 16
One disadvantage of this class of nonlinear decoders trained from large datasets, when
compared to traditional linear decoders trained on smaller datasets, is the longer
training time required to fit their more numerous parameters. In the present study, which
we did not optimize for fast training, this took multiple hours. This could be substantially
sped up by iteratively updating the decoder with new data instead of retraining de novo
and by leveraging faster computation available with graphics processing units, parallel
computing, or custom hardware. A second disadvantage of the MRNN is that it appears
to require more training data to saturate its performance (Fig. 2b) compared to
conventional methods, such as FIT-KF, that are trained from sameday calibration data.
We do not view this as a major limitation because the motivation for using the MRNN is
to take advantage of accumulated previous recordings. Nonetheless, it will be valuable
to compare the present approach with other decoder architectures and training
strategies, which may yield similar performance and robustness while requiring less
training data.
The MRNN decoder's robustness was due to the combination of a large training data
corpus, deliberate perturbation of the training data, and a computationally powerful
architecture that was able to effectively learn this diverse training data. While it may
seem obvious that successfully learning more training data is better, this is not
necessarily true. Older data only helps a decoder if some of these past recordings
capture neural-to-kinematic relationships that are similar to that of the current recording
condition. Our offline and closed-loop MRNN robustness results suggest that this was
Page 17
indeed the case for the two monkeys used in this study. While there are indications that
this will also be true in human BMI studies14, validating this remains an important future
question. The relevance of old data to present recording conditions also enables a
different robustness-enhancing approach: store a library of different past decoders and
evaluate each to find a decoder well-suited for the current conditions (e.g.,10). However,
since offline analyses are poor predictors of closed-loop performance32,42,45,51, this
approach necessitates a potentially lengthy decoder selection process. Using a single
decoder (such as the MRNN) that works across many recording conditions avoids
switching-related downtime.
In addition to training with months of previous data, we improved the MRNN's
robustness by intentionally perturbing the training neural data. In the present study we
applied random Gaussian firing rate scaling based on a general assumption that the
decoder should be broadly robust to both global and private shifts in observed firing
rates. This perturbation type proved effective, but we believe that this approach (called
data augmentation in the machine learning community) can potentially be much more
powerful when combined with specific modeling of recording condition changes that the
experimenter wants to train robustness against. For example, data augmentation could
incorporate synthetic examples of losing a particularly error-prone set of electrodes;
recording changes predicted by models of array micro-movement or degradation; and
perhaps even the predicted interaction between kinematics and changes in cognitive
state or task context. We believe this is an important avenue for future research.
Page 18
We view the success of our specific MRNN decoder implementation as a validation of
the more general BMI decoder strategy of training a computationally powerful nonlinear
decoder to a large quantity of data representing many different recording conditions.
This past data need not have been collected explicitly for the purpose of training as was
done in this study; neural data and corresponding kinematics from past closed-loop BMI
use can also serve as training data4,10. It is likely that other nonlinear decoding
algorithms will also benefit from this strategy, and that there are further opportunities to
advance the reliability and performance of BMIs by starting to take advantage of these
devices' ability to generate large quantities of data as part of their regular use.
Page 19
METHODS
Animal model and neural recordings
All procedures and experiments were approved by the Stanford University Institutional
Animal Care and Use Committee. Experiments were conducted with adult male rhesus
macaques (R and L, ages 8 and 18, respectively), implanted with 96-electrode Utah
arrays (Blackrock Microsystems Inc., Salt Lake City, UT) using standard neurosurgical
techniques. Monkeys R and L were implanted 30 months and 74 months prior to the
primary experiments, respectively. Monkey R had two electrode arrays implanted, one
in caudal dorsal premotor cortex (PMd) and the other in primary motor cortex (M1), as
estimated visually from anatomical landmarks. Monkey L had one array implanted on
the border of PMd and M1. Within the context of the simple point-to-point arm and BMI
reach behavior of this study, we observed qualitatively similar response properties
between these motor cortical areas; this is consistent with previous reports of a gradient
of increasing preparatory activity, rather than stark qualitative differences, as one moves
more rostral from M1 52–56. Therefore, and in keeping with standard BMI decoding
practices 6,8,10,24,38,40,46, we did not distinguish between M1 and PMd electrodes.
Behavioral control and neural decode were run on separate PCs using the xPC Target
platform (Mathworks, Natick, MA), enabling millisecond-timing precision for all
computations. Neural data were initially processed by Cerebus recording system(s)
(Blackrock Microsystems Inc., Salt Lake City, UT) and were available to the behavioral
control system within 5 ±
1 ms. Spike counts were collected by applying a single
negative threshold, set to -4.5 times the root mean square of the spike band of each
electrode. We decoded "threshold crossings", which contain spikes from one or more
Page 20
neurons in the electrode's vicinity, as per standard practice for intracortical BMIs
1,4,6,7,10,15,16,31,38,40 because threshold crossings provide roughly comparable population-
level velocity decode performance to sorted single-unit activity, without time-consuming
sorting30,45,57–59, and may be more stable over time30,45. To orient the reader to the
quality of the neural signals available during this study, Supplementary Data 1 provides
statistics of several measures of electrodes' 'tuning' and cross-talk.
Behavioral tasks
We trained the monkeys to acquire targets with a virtual cursor controlled by either the
position of the hand contralateral to the arrays or directly from neural activity. Reaches
to virtual targets were made in a 2D fronto-parallel plane presented within a 3D
environment (MSMS, MDDF, USC, Los Angeles, CA) generated using a Wheatstone
stereograph fused from two LCD monitors with refresh rates at 120 Hz, yielding frame
updates within 7 ±
4 ms 43. Hand position was measured with an infrared reflective
bead tracking system at 60 Hz (Polaris, Northern Digital, Ontario, Canada). During BMI
control, we allowed the monkey's reaching arm to be unrestrained47,60 so as to not
impose a constraint upon the monkey that during BMI control he must generate neural
activity that does not produce overt movement61.
In the Radial 8 Task the monkey was required to acquire targets alternating between a
center target and one of eight peripheral targets equidistantly spaced on the
circumference of a circle. For our closed-loop BMI experiments, the peripheral targets
were positioned 8 cm from the center target. In hand reaching datasets used for
decoder training and offline decode, the targets were either 8 cm or 12 cm (the majority
of datasets) from the center. In much of Monkey L's training data, the three targets
Page 21
forming the upper quadrant were placed slightly further (13 and 14 cm) based on
previous experience that this led to decoders with improved ability to acquire targets in
that quadrant. To acquire a target, the monkey had to hold the cursor within a 4 cm × 4
cm acceptance window centered on the target for 500 ms. If the target was acquired
successfully, the monkey received a liquid reward. If the target was not acquired within
5 s (BMI control) or 2 s (hand control) of target presentation, the trial was a failure and
no reward was given.
Although the data included in this study's decoder training datasets and offline analyses
span many months of each animal's research career, these data start after each animal
was well-trained in performing point-to-point planar reaches; day-to-day variability when
making the same reaching movements was modest. To quantify behavioral similarity
across the study, we took advantage of having collected the same "Baseline Block" task
data at the start of most experimental sessions: 171/185 monkey R days, 398/452
monkey L days. This consisted of ~200 trials of arm-controlled Radial 8 Task reaches,
with targets 8 cm from the center. For each of these recording sessions, we calculated
the mean hand x- and y-velocities (averaged over trials to/from a given radial target)
throughout a 700 ms epoch following radial target onset for outward reaches and 600
ms following center target onset for inward reaches (inward reaches were slightly
faster). We concatenated these velocity time series across the 8 different targets,
producing 10,400 ms x-velocity and y-velocity vectors from each recording session.
Behavioral similarity between any two recording sessions was then measured by the
Pearson correlation between the datasets' respective x- and y- velocity vectors. Then,
Page 22
the two dimensions' correlations were averaged to produce a single correlation value
between each pair of sessions. These hand velocity correlations were 0.90 ± 0.04
(mean ± s.d. across days) for monkey R, and 0.91 ± 04 for monkey L.
We measured closed-loop BMI performance on the Radial 8 Task using two metrics.
Target acquisition rate is the number of peripheral targets acquired divided by the
duration of the task. This metric holistically reflects cursor control ability because, unlike
time to target, it is negatively affected by failed trials and directly relates to the animal's
rate of liquid reward. Targets per minute is calculated over all trials of an experimental
condition (i.e., which decoder was used) and therefore yields a single measurement
datum per day/experimental condition. Across-days distributions of a given decoder's
targets per minute performance were consistent with a normal distribution (Kolmogorov-
Smirnov test), justifying our use of paired t-tests statistics when comparing this metric.
This is consistent with the measure reflecting the accumulated outcome of many
hundreds of random processes (individual trials). As a second measure of performance
that is more sensitive when success rates are high and similar between decoders (such
as the "ideal" conditions where we presented no challenges to the decoders), we
compared times to target. This measure consists of the time between when the target
appeared and when the cursor entered the target acceptance window prior to
successfully acquiring the target, but does not include the 500 ms hold time (which is
constant across all trials). Times to target are only measured for successful trials to
peripheral targets, and were only compared when success rates were not significantly
different (otherwise, a poor decoder with a low success rate that occasionally acquired a
Page 23
target quickly by chance could nonsensically "outperform" a good decoder with 100%
success rate but slower times to target). Because these distributions were not normal,
we used the Mann–Whitney–Wilcoxon rank-sum tests when comparing two decoders'
times to target.
In the Random Target Task each trial's target appeared at a random location within a 20
cm × 20 cm region centered within a larger workspace that was 40 x 30 cm. A new
random target appeared after each trial regardless of whether this trial was a success or
a failure due to exceeding the 5 s time limit. The target location randomization enforced
a rule that the new target's acceptance area could not overlap with that of the previous
target. Performance on the Random Target Task was measured by success rate (the
number of successfully acquired targets divided by the total number of presented
targets) and the normalized time to target. Normalized time to target is calculated for
successful trials following another successful trial, and is the duration between target
presentation and target acquisition (not including the 500 ms hold time), divided by the
straight-line distance between this target's center and the previously acquired target's
center 62.
Decoder comparison experiment design
All offline decoding comparisons between MRNN and FIT-KF were performed using test
data that was held out from the data used to train the decoders. Thus, although the
MRNN has many more parameters than FIT-KF, both of these fundamentally different
Page 24
algorithm types were trained according to best practices with matched training data.
This allows their performance to be fairly compared.
When comparing online decoder performance using BMI-controlled Radial 8 Target or
Random Target Tasks, the decoders were tested in using an interleaved "block-set"
design in which contiguous ~200 trial blocks of each decoder were run followed by
blocks of the next decoder, until the block-set comprising all tested decoders was
complete and the next block-set began. For example, in the electrode dropping
experiments (Figure 3), this meant an "AB AB" design where A could be a block of
MRNN trials and B could be a block of FIT Sameday trials. For the stale training data
experiments (Figure 4), an "ABCD ABCD ABCD… " design was used to test the four
different decoders. When switching decoders, we gave the monkey ~20 trials to
transition to the new decoder before starting 'counting' performance in the block; we
found this to be more than sufficient for both animals to adjust. For electrode dropping
experiments, the order of decoders within each block-set was randomized across days.
For stale training data experiments, where several decoders often performed very
poorly, we manually adjusted the order of decoders within block-sets so as to keep the
monkeys motivated by alternating what appeared to be more and less frustrating
decoders. All completed blocks were included in the analysis. Throughout the study, the
experimenters knew which decoder was in use, but all comparisons were quantitative
and performed by the same automated computer program using all trials from
completed blocks. The monkeys were not given an overt cue to the decoder being used.
Page 25
During online experiments, we observed that when a decoder performed extremely
poorly, such that the monkey could not reliably acquire targets within the 5 second time
limit, the animal stopped performing the task before the end of the decoder evaluation
block. To avoid frustrating the monkeys, we stopped a block if the success rate fell
below 50% after at least 10 trials. This criterion was chosen based on pilot studies in
which we found that below this success rate, the monkey would soon thereafter stop
performing the task and would frequently refuse to re-engage for a prolonged period of
time. Our interleaved block design meant that each decoder was tested multiple times
on a given experimental session, which in principle provides the monkey multiple
attempts to finish a block with each decoder. In practice, we found that monkeys could
either complete every block or no blocks with a given decoder, and we refer to decoders
that could not be used to complete a block as having "failed". These decoders'
performance was recorded as 0 targets per minute for that experimental session. The
exception to the above was that during an electrode dropping experiment session, we
declared both FIT-KF Sameday and MRNN as having failed for a certain number of
electrodes dropped if the monkey could not complete a block with either decoder. That
is, we did not continue with a second test of both (unusable) decoders as per the
interleaved block design, because this would have unduly frustrated the animal.
We performed this study with two monkeys, which is the conventional standard for
systems neuroscience and BMI experiments using a non-human primate model. No
monkeys were excluded from the study. We determined how many experimental
sessions to perform as follows. For all offline analyses, we examined the dates of
Page 26
previous experimental sessions with suitable arm reaching data and selected sets of
sessions with spacing most appropriate for each analysis (e.g., closely spaced sessions
for Figure 2b, all of the available data for Figure 2c, two clusters with a gap for stale
training analyses). All these pre-determined sessions were then included in the
analysis. For the stale training data experiments (Figure 4), the choice of two gaps with
three test days each was pre-established. For the electrode dropping experiments
(Figure 3), we did not know a priori how electrode dropping would affect performance
and when each decoder would fail. We therefore determined the maximum number of
electrodes to drop during the experiment and adjusted the number of sessions testing
each drop condition during the course of experiments to comprehensively explore the
"dynamic range" across which decoder robustness appeared to differ. For both of these
experiments, during an experimental session additional block-sets were run until the
animal became satiated and disengaged from the task. We did not use formal effect
size calculations to make data sample size decisions, but did perform a variety of
experiments with large numbers of decoder comparison trials (many tens of thousands)
so as to be able to detect substantial decoder performance differences. For secondary
online experiments (Supplementary Figures 3 and 7), which served to support offline
analyses (Supplementary Figure 3) or demonstrate that the MRNN could acquire other
target locations (Supplementary Figure 7), we chose to perform only 1-3 sessions per
animal in the interest of conserving experimental time.
At a high level, the MRNN decoder transforms inputs 𝐮𝑡 , the observed spike counts on
Neural decoding using a Multiplicative Recurrent Neural Network (MRNN)
each electrode at a particular time, into a cursor position and velocity output. This is
Page 27
accomplished by having first trained the weights of an artificial recurrent neural network
such that when the network is provided a time series of training neural data inputs, the
training data kinematic outputs can be accurately 'read out' from this neural network's
state. The rest of this section will describe the architecture, training, and use of the
MRNN for the purpose of driving a BMI.
The generic recurrent network model is defined by an 𝑁-dimensional vector of activation
variables, 𝐱, and a vector of corresponding "firing rates'', 𝐫=tanh𝐱. Both 𝐱 and 𝐫 are
continuous in time and take continuous values. In the standard RNN model, the input
affects the dynamics as an additive time-dependent bias in each dimension. In the
MRNN model, the input instead directly parameterizes the artificial neural network's
recurrent weight matrix, allowing for a multiplicative interaction between the input and
the hidden state. One view of this multiplicative interaction is that the hidden state of the
recurrent network is selecting an appropriate decoder for the statistics of the current
dataset. The equation governing the dynamics of the activation vector is of the form
suggested in 35, but adapted in this study to continuous time in order to control the
smoothness to MRNN outputs, 𝜏 𝐱𝑡 =−𝐱𝑡 +𝐉𝐮!𝐫𝑡 +𝐛!.
The 𝑁×𝑁×𝐮 tensor 𝐉𝐮! describes the weights of the recurrent connections of the
network, which are dependent on the 𝐸-dimensional input, 𝐮𝑡 . The symbol 𝐮 denotes
the number of unique values 𝐮𝑡 can take. Such a tensor is unusable for continuous
valued 𝐮(𝑡) or even discrete valued 𝐮(𝑡) with prohibitively many values. To make these
computations tractable, the input is linearly combined into 𝐹 factors and 𝐉𝐮! is
factorized 35 according to the following formula:
Page 28
∙𝐉!",
𝐉𝐮! =𝐉!"∙diag 𝐉!"𝐮𝑡
where 𝐉!" has dimension 𝑁×𝐹, 𝐉!" has dimension 𝐹×𝐸, 𝐉!" has dimension 𝐹×𝑁, and
diag(𝐯) takes a vector, 𝐯, and returns a diagonal matrix with 𝐯 along the diagonal. One
can directly control the complexity of interactions by choosing 𝐹. Additionally, the
network units receive a bias 𝐛! . The constant 𝜏 sets the time scale of the network, so
we set 𝜏 on the order of hundreds of milliseconds to allow meaningful interactions. The
output of the network is read out from a weighted sum of the network firing rates plus a
σtrial
σelectrode
𝐳𝑡 =𝐖𝐎𝐫𝑡 +𝐛!,
bias, defined by the equation
Monkey R Monkey L
20-30 ms
20 ms
100 ms
100-150 ms
50
100
50
100
0.045
0.045
0.3
0.3
1.0
1.0
1.0
1.0
1.0
1.0
96
192
where 𝐖𝐎 is an 𝑀×𝑁 matrix, and 𝐛! is an M-dimensional bias.
𝛥𝑡
𝜏
𝑁
𝐹
𝑔!"
𝑔!"
𝑔!"
𝐸
#
params
in
each
MRNN
𝛽
We began decoder training by instantiating MRNNs of network size 𝑁=100 (monkey
R) and 𝑁=50 (monkey L) with 𝐹=𝑁 in both cases (see Table 1 for all key
parameters). For monkey R, who was implanted with two multielectrode arrays, 𝐸=
Table 1. Network and training parameters used for the closed-loop MRNN BMI decoder
MRNN training
days of
training data
years
spanned
82-129
189-230
1.59
39502
0.99
2.77
9952
0.99
Page 29
192, while for monkey L with one array, 𝐸=96. The non-zero elements of the non-
sparse matrices 𝐉!",𝐉!",𝐉!" are drawn independently from a Gaussian distribution with
zero mean and variance 𝑔!"/𝐹,𝑔!"/𝐸, and 𝑔!"/𝑁, with 𝑔!",𝑔!", and 𝑔!" set to 1.0 in
this study. The elements of 𝐖𝐎 are initialized to zero, and the bias vectors 𝐛! and 𝐛! are
The input 𝐮(𝑡) to the MRNN (through the matrix 𝐉𝐮(!)) is the vector of binned spikes at
each time step. Concatenating across time in a trial yields training data matrix, 𝐔!, of
binned spikes of size 𝐸×𝑇!, where 𝑇! is the number of times steps for the 𝑗!" trial. Data
also initialized to 0.
from five consecutive actual monkey-reaching trials are then concatenated together to
make one "MRNN training" trial. The first two actual trials in an MRNN training trial were
used for seeding the hidden state of the MRNN (i.e., not used for learning), whereas the
next three actual trials were used for learning. With the exception of the first two actual
trials from a given recording day, the entire set of actual trials are used for MRNN
learning by incrementing the actual trial index that begins each training trial by one.
The parameters of the network were trained offline to reduce the averaged squared
error between the measured kinematic training data and the output of the network, 𝐳(𝑡).
Specifically, we used the Hessian-Free (HF) optimization method36,63 for RNNs (but
adapted to the continuous-time MRNN architecture). HF is an exact 2nd order method
that uses back-propagation through time to compute the gradient of the error with
respect to the network parameters. The set of trained parameters is
{𝐉!",𝐉!",𝐉!",𝐛!,𝐖𝐎,𝐛!}. The HF algorithm has three critical parameters: the minibatch
size, the initial lambda setting, and the max number of conjugate gradient iterations. We
Page 30
set these parameters to one-fifth the total number of trials, 0.1, and 50, respectively.
The optimizations were run for 200 steps and a snapshot of the network was saved
every 10 steps. Among these snapshots, the network with the lowest cross-validation
error on held-out data was used in the experiment.
We independently trained two separate MRNN networks to each output a 2-dimensional
(𝑀=2) signal, 𝐳(𝑡). The first network learned to output the normalized hand position
through time in both the horizontal (𝑥) and vertical (𝑦) spatial dimensions. The second
MRNN learned to output the hand velocity through time, also in the 𝑥 and 𝑦 dimensions.
We calculated hand velocities from the positions numerically using central differences.
In this study we trained a new MRNN whenever adding new training data; this allowed
us to verify that the training optimization consistently converged to a high-quality
decoder. However, it is easy to iteratively update an MRNN decoder with new data
without training from scratch. By adding the new data to the training corpus and using
the existing decoder weights as the training optimization's initial conditions, the MRNN
will more rapidly converge to a new high-quality decoder.
Improving the MRNN by training with many datasets and intentionally perturbed
inputs
A critical element of achieving both high performance and robustness in the MRNN
decoder was training the decoder using data from many previous recording days
spanning many months. When training datasets included data from more than one day,
we randomly selected a small number of trials from each day for a given minibatch. In
this way, every minibatch of training data sampled the input distributions from all training
days.
Page 31
𝑛!!=𝜂!𝜂!𝑛!!,
A second key element of training robustness to recording condition changes was a form
of data augmentation in which we intentionally introduced perturbations to the neural
spike trains that were used to train the MRNN. The concatenated input,
𝐔=[𝐔!,…,𝐔!!!] was perturbed by adding and removing spikes from each electrode.
We focus on electrode c of the 𝑗!" training trial, i.e., a row vector of data 𝐔!,:! . Let the
number of actual observed spikes in 𝐔!,:! be 𝑛!!. This number was perturbed according to
where both 𝜂! and 𝜂! are Gaussian variables with a mean of one and standard
deviations of 𝜎!"#$% and 𝜎!"!#$%&'!, respectively. Conceptually, 𝜂! models a global firing
𝜂! models electrode by electrode perturbations such as electrode dropping or moving
baselines in individual neurons. If 𝑛!! was less than zero or greater than 2𝑛!!, it was
same electrode and training trial. Otherwise, if 𝑛!! was greater than 𝑛!!, then 𝑛!!−𝑛!!
spikes were added to random time bins of the training trial. If 𝑛!! was less than 𝑛!!, then
𝑛!!−𝑛!! spikes were randomly removed from time bins of the training trial that already
had spikes. Finally, if 𝑛!!=𝑛!!, nothing was changed.
the optimization algorithm, i.e. in the HF algorithm, the perturbation 𝑛!!=𝜂!𝜂!𝑛!! occurs
The process of perturbing the binned spiking data occurred anew on every iteration of
rate modulation across all electrodes of the array (e.g., array movement, arousal), while
resampled, which kept the average number of perturbed spikes in a given electrode and
training trial roughly equal to the average number of true (unperturbed) spikes in the
after each update of the network parameters. Note that these input data perturbations
were only applied during MRNN training; when the MRNN was used for closed-loop BMI
Page 32
control, true neural spike counts were provided as inputs. Supplementary Figure 3
shows the closed-loop control quality difference between the MRNN trained with and
without this data augmentation.
Controlling a BMI cursor with MRNN network output
Once trained, the MRNNs were compiled into the embedded real-time operating system
and run in closed-loop to provide online BMI cursor control. The decoded velocity and
position were initialized to 0, as was the MRNN hidden state. Thereafter, at each
decode time step the parallel pair of MRNNs received binned spike counts as input and
had their position and velocity outputs blended to yield a position estimate. This was
used to update the drawn cursor position. The on-screen position that the cursor moves
to during BMI control, 𝑑!𝑡,𝑑!𝑡, is defined by
where 𝑣!, 𝑣!, 𝑝!, 𝑝! are the normalized velocity and positions in the x and y dimensions
and 𝛾!,𝛾! are factors that convert from the normalized velocity and position,
respectively, to the coordinates of the virtual-reality workspace. The parameter 𝛽 sets
𝑑!𝑡 =𝛽𝑑!𝑡−𝛥𝑡 +𝛾!𝑣!𝑡−𝛥𝑡𝛥𝑡 + 1−𝛽𝛾!𝑝!𝑡
𝑑!𝑡 =𝛽𝑑!𝑡−𝛥𝑡 +𝛾!𝑣!𝑡−𝛥𝑡𝛥𝑡 +(1−𝛽)𝛾!𝑝!(𝑡)
the amount of position versus velocity decoding and was set to 0.99. In effect, the
decode was almost entirely dominated by velocity, with a slight position contribution to
stabilize the cursor in the workplace (i.e., offset accumulated drift). Note that when
calculating offline decode accuracy (Fig. 2), we set 𝛽 to 1 to more fairly compare the
MRNN to the FIT-KF decoder, which decodes velocity only.
We note that although 1) the MRNN's recurrent connections mean that previous inputs
affect how subsequent near-term inputs are processed, and 2) our standard procedure
Page 33
was to retrain the MRNN with additional data after each experimental session, the
MRNN is not an "adaptive" decoder in the traditional meaning of the term. Its
parameters are fixed during closed-loop use, and therefore when encountering
recording condition changes, the MRNN cannot "learn" from this new data to update its
neural-to-kinematic mappings in the way that adaptive decoders do (e.g., 4,24,27). Insofar
as its architecture and training regime make the MRNN robust to input changes, this
robustness is "inherent" rather than "adaptive."
Neural decoding using a Feedback Intention Trained Kalman Filter (FIT-KF)
We compared the performance of the MRNN to FIT-KF 40. The FIT-KF is a Kalman filter
where the underlying kinematic state, 𝐳(𝑡), comprises the position and velocity of the
cursor as well as a bias term. Observations of the neural binned spike counts, 𝐲(𝑡), are
used to update the kinematic state estimate. With 𝛥𝑡 denoting bin width (25 ms in this
study), the FIT-KF assumes the kinematic state gives rise to the neural observations
according to the following linear dynamical system:
𝐳𝑡+𝛥𝑡 = 𝐀𝐳𝑡 +𝐰𝑡
𝐲𝑡 =𝐂𝐳𝑡 +𝐪(𝑡)
where 𝐰(𝑡) and 𝐪𝑡 are zero-mean Gaussian noise with covariance matrices 𝐖 and 𝐐
respectively. The Kalman filter is a recursive algorithm that estimates the state 𝐳(𝑡)
using the current observation 𝐲(𝑡) and the previous state estimate 𝐳(𝑡−𝛥𝑡). Previous
The parameters of this linear dynamical system, 𝐀,𝐖,𝐂,𝐐, are learned in a supervised
studies have used such decoders to drive neural cursors (e.g. 5,38,64).
fashion from hand reach training data as previously reported6,65. The FIT-KF then
incorporates two additional innovations. First, it performs a rotation of the training
kinematics using the assumption that at every moment in time, the monkey intends to
Page 34
move the cursor directly towards the target. Second, it assumes that at every time step,
the monkey has perfect knowledge of the decoded position via visual feedback. This
affects Kalman filter inference in two ways: first, the covariance of the position estimate
in Kalman filtering is set to 0, and secondly, the neural activity that is explainable by the
cursor position is subtracted from the observed binned spike counts. These innovations
are further described in 6,40.
Mutual information for determining electrode dropping order
When testing the decoders' robustness to unexpected electrode loss, we determined
which electrodes to drop by calculating the mutual information between each electrode's
binned spike counts and the reach direction. This metric produced a ranking of
electrodes in terms of how statistically informative they were of the reach direction;
importantly, this metric is independent of the decoder being used. Let 𝑝 denote the
distribution of an electrode's binned firing rates, 𝑦 denote the binned spike counts lying
in a finite set 𝛶 of possible binned spike counts, 𝑀 denote the number of reach
directions, and 𝑥! denote reach direction j. The set 𝛶 comprised {0,1,2,3,4,5+} spike
counts, where any spike counts greater than or equal to 5 were counted towards the
same bin ("5+", corresponding to an instantaneous firing rate of 250 Hz in a 20 ms bin).
We calculated the entropy of each electrode,
as well as its entropy conditioned on the reach direction
!∈!
𝐻𝑌 =−
𝑝𝑥!
𝐻𝑌𝑋 =−
!
!!!
𝑝𝑦 log𝑝(𝑦),
𝑝𝑦𝑥!
!∈!
log𝑝𝑦𝑥!
.
Page 35
From these quantities, we calculated the mutual information between the neural activity
and the reach direction as 𝐼!"#$𝑋;𝑌 =𝐻𝑌 −𝐻(𝑌𝑋). We dropped electrodes in
order from highest to lowest mutual information.
Principal angles of neural subspaces analysis
For a parsimonious scalar metric of how similar patterns of neural activity during
reaching were between a given pair of recording days (used in Supplementary Fig. 1),
we calculated the minimum principal angle between the neural subspaces of each
recording day. We defined the neural subspace on a recording day as the top 𝐾
similar day 𝑖 and day 𝑗's motifs of covariance between electrodes' activity were during
arm reaching. Specifically, we started with a matrix 𝒀! from each day 𝑖 consisting of
principal components of the neural coactivations. Put more simply, we asked how
the number of non-overlapping 20 ms bins comprising the duration of this task. We next
neural activity collected while the monkey performed ~200 trials of a Radial 8 Task (8
cm distance to targets) using arm control; this task has been run at the start of almost
every experimental session conducted using both monkeys R and L since array
implantation. 𝒀! is of dimensionality 𝐸×𝑇, where E is the number of electrodes and T is
subtracted from each row of 𝒀! that electrode's across-days mean firing rate (we also
eigenvalue decomposition on the covariance matrix 𝒀!𝒀!! (note, 𝒀! is zero mean), and
defined the matrix 𝑽! as the first 𝐾 eigenvectors. 𝑽! had dimensions 𝐸×𝐾, where each
column 𝑘 is the vector of principal component coefficients (eigenvector) corresponding
to the 𝑘!! largest eigenvalue of the decomposition. Supplementary Fig. 1 was generated
repeated this analysis without across-days mean subtraction and observed qualitatively
similar results, not shown). To obtain the principal components, we performed
Page 36
using 𝐾=10, i.e., keeping the first 10 PCs, but the qualitative appearance of the data
was similar when 𝐾 was varied from 2 to 30 (not shown). Finally, the difference metric
between days 𝑖 and 𝑗 was computed as the minimum of the 𝐾 subspace angles
between matrices 𝑽! and 𝑽!. Subspace angles were computed using the subspacea
MATLAB function66.
Data and code availability
All relevant data and analysis code can be made available from the authors upon
request.
Acknowledgments
We thank Mackenzie Risch, Michelle Wechsler, Liana Yates, Shannon Smith, and Rosie
Steinbach for surgical assistance and veterinary care; Evelyn Castaneda and Beverly
Davis for administrative support; Boris Oskotsky for information technology support.
This work was supported by the National Science Foundation Graduate Research
Fellowship (J.C.K., S.D.S.); NSF IGERT 0734683 (S.D.S.); Christopher and Dana
Reeve Paralysis Foundation (S.I.R. and K.V.S.); and the following to K.V.S.: Burroughs
Welcome Fund Career Awards in the Biomedical Sciences, Defense Advanced
Research Projects Agency Reorganization and Plasticity to Accelerate Injury Recovery
N66001-10-C-2010, US National Institutes of Health Institute of Neurological Disorders
and Stroke Transformative Research Award R01NS076460, US National Institutes of
Health Director's Pioneer Award 8DP1HD075623-04, US National Institutes of Health
Director's Transformative Research Award (TR01) from the NIMH #5R01MH09964703,
Page 37
and Defense Advanced Research Projects Agency NeuroFAST award from BTO
#W911NF-14-2-0013.
Author contributions
D.S., S.D.S., and J.C.K. designed the experiments. D.S. trained the MRNN decoders.
S.D.S. and J.C.K. conducted the experiments and data analysis. S.D.S. wrote the
manuscript. J.C.K., D.S., and S.I.R. assisted in manuscript preparation. S.I.R. was
responsible for surgical implantation. K.V.S. was involved in all aspects of
experimentation, data review, and manuscript preparation.
Competing financial interests
The authors declare no competing financial interests.
Page 38
REFERENCES
1.
Hochberg,
L.
R.
et
al.
Reach
and
grasp
by
people
with
tetraplegia
using
a
neurally
controlled
robotic
arm.
Nature
485,
372–5
(2012).
2.
Collinger
et
al.
High--‐performance
neuroprosthetic
control
by
an
individual
with
tetraplegia.
Lancet
381,
557–64
(2013).
3.
Gilja,
V.
et
al.
Clinical
translation
of
a
high--‐performance
neural
prosthesis.
Nat.
Med.
21,
1142–1145
(2015).
4.
Jarosiewicz,
B.
et
al.
Virtual
typing
by
people
with
tetraplegia
using
a
self--‐calibrating
intracortical
brain--‐computer
interface.
Sci.
Transl.
Med.
7,
313ra179--‐313ra179
(2015).
5.
Bacher,
D.
et
al.
Neural
Point--‐and--‐Click
Communication
by
a
Person
With
Incomplete
Locked--‐In
Syndrome.
Neurorehabil.
Neural
Repair
29,
462–471
(2015).
6.
Gilja,
V.
et
al.
A
high--‐performance
neural
prosthesis
enabled
by
control
algorithm
design.
Nat.
Neurosci.
15,
1752–7
(2012).
7.
Nuyujukian,
P.,
Fan,
J.
M.,
Kao,
J.
C.,
Ryu,
S.
I.
&
Shenoy,
K.
V.
A
High--‐Performance
Keyboard
Neural
Prosthesis
Enabled
by
Task
Optimization.
IEEE
Trans.
Biomed.
Eng.
62,
21–29
(2015).
8.
Ganguly,
K.
&
Carmena,
J.
M.
Emergence
of
a
stable
cortical
map
for
neuroprosthetic
control.
PLoS
Biol.
7,
e1000153
(2009).
Flint,
R.
D.,
Ethier,
C.,
Oby,
E.
R.,
Miller,
L.
E.
&
Slutzky,
M.
W.
Local
field
potentials
9.
allow
accurate
decoding
of
muscle
activity.
Journal
of
Neurophysiology
108,
18–24
(2012).
10.
Nuyujukian,
P.
et
al.
Performance
sustaining
intracortical
neural
prostheses.
J.
Neural
Eng.
11,
66003
(2014).
11.
Chestek,
C.
a
et
al.
Single--‐neuron
stability
during
repeated
reaching
in
macaque
premotor
cortex.
J.
Neurosci.
27,
10742–50
(2007).
Simeral,
J.
D.,
Kim,
S.--‐P.,
Black,
M.
J.,
Donoghue,
J.
P.
&
Hochberg,
L.
R.
Neural
control
of
12.
cursor
trajectory
and
click
by
a
human
with
tetraplegia
1000
days
after
implant
of
an
intracortical
microelectrode
array.
J.
Neural
Eng.
8,
25027
(2011).
13.
Perge,
J.
A.
et
al.
Intra--‐day
signal
instabilities
affect
decoding
performance
in
an
intracortical
neural
interface
system.
J.
Neural
Eng.
10,
36004
(2013).
14.
Perge,
J.
A.
et
al.
Reliability
of
directional
information
in
unsorted
spikes
and
local
field
potentials
recorded
in
human
motor
cortex.
J.
Neural
Eng.
11,
46007
(2014).
15.
Wodlinger,
B.
et
al.
Ten--‐dimensional
anthropomorphic
arm
control
in
a
human
brain−machine
interface:
difficulties,
solutions,
and
limitations.
J.
Neural
Eng.
12,
16011
(2015).
16.
Aflalo,
T.
et
al.
Decoding
motor
imagery
from
the
posterior
parietal
cortex
of
a
tetraplegic
human.
Science
(80--‐.
).
348,
906–910
(2015).
17.
Jarosiewicz,
B.
et
al.
Advantages
of
closed--‐loop
calibration
in
intracortical
brain--‐
computer
interfaces
for
people
with
tetraplegia.
J.
Neural
Eng.
10,
46012
(2013).
18.
Huggins,
J.
E.,
Wren,
P.
A.
&
Gruis,
K.
L.
What
would
brain--‐computer
interface
users
want?
Opinions
and
priorities
of
potential
users
with
amyotrophic
lateral
sclerosis.
Amyotroph.
Lateral
Scler.
12,
318–324
(2011).
19.
Collinger
et
al.
Functional
priorities,
assistive
technology,
and
brain--‐computer
interfaces
after
spinal
cord
injury.
J.
Rehabil.
Res.
Dev.
50,
145
(2013).
20.
Srinivasan,
L.,
Eden,
U.
T.,
Mitter,
S.
K.
&
Brown,
E.
N.
General--‐Purpose
Filter
Design
for
Neural
Prosthetic
Devices.
J.
Neurophysiol.
98,
2456–2475
(2007).
Page 39
21.
Wu,
W.
&
Hatsopoulos,
N.
G.
Real--‐time
decoding
of
nonstationary
neural
activity
in
motor
cortex.
IEEE
Trans.
Neural
Syst.
Rehabil.
Eng.
16,
213–222
(2008).
Shpigelman,
L.,
Lalazar,
H.
&
Vaadia,
E.
Kernel--‐ARMA
for
Hand
Tracking
and
Brain--‐
22.
Machine
Interfacing
During
3D
Motor
Control.
Neural
Inf.
Process.
Syst.
1489–1496
(2008).
doi:citeulike--‐article--‐id:6002127
23.
DiGiovanna,
J.,
Mahmoudi,
B.,
Fortes,
J.,
Principe,
J.
C.
&
Sanchez,
J.
C.
Coadaptive
Brain
--‐
Machine
Interface
via
Reinforcement
Learning.
IEEE
Trans.
Biomed.
Eng.
56,
54–64
(2009).
24.
Li,
Z.,
O'Doherty,
J.
E.,
Lebedev,
M.
A.
&
Nicolelis,
M.
A.
L.
Adaptive
Decoding
for
Brain--‐
Machine
Interfaces
Through
Bayesian
Parameter
Updates.
Neural
Comput.
23,
3162–
3204
(2011).
25.
Dangi,
S.,
Gowda,
S.,
Heliot,
R.
&
Carmena,
J.
M.
Adaptive
Kalman
filtering
for
closed--‐
loop
Brain--‐Machine
Interface
systems.
2011
5th
Int.
IEEE/EMBS
Conf.
Neural
Eng.
NER
2011
609–612
(2011).
doi:10.1109/NER.2011.5910622
26.
Mahmoudi,
B.
&
Sanchez,
J.
C.
A
symbiotic
brain--‐machine
interface
through
value--‐
based
decision
making.
PLoS
One
6,
(2011).
27.
Orsborn,
A.
L.
et
al.
Closed--‐Loop
Decoder
Adaptation
Shapes
Neural
Plasticity
for
Skillful
Neuroprosthetic
Control.
Neuron
82,
1380–1393
(2014).
28.
Bishop,
W.
et
al.
Self--‐recalibrating
classifiers
for
intracortical
brain--‐computer
interfaces.
J.
Neural
Eng.
11,
26001
(2014).
29.
Pohlmeyer,
E.
A.,
Mahmoudi,
B.,
Geng,
S.,
Prins,
N.
W.
&
Sanchez,
J.
C.
Using
reinforcement
learning
to
provide
stable
brain--‐machine
interface
control
despite
neural
input
reorganization.
PLoS
One
9,
e87253
(2014).
Chestek,
C.
A.
et
al.
Long--‐term
stability
of
neural
prosthetic
control
signals
from
30.
silicon
cortical
arrays
in
rhesus
macaque
motor
cortex.
J.
Neural
Eng.
8,
45005
(2011).
Flint,
R.
D.,
Wright,
Z.
A.,
Scheid,
M.
R.
&
Slutzky,
M.
W.
Long
term,
stable
brain
31.
machine
interface
performance
using
local
field
potentials
and
multiunit
spikes.
J.
Neural
Eng.
10,
56005
(2013).
32.
Stavisky,
S.
D.,
Kao,
J.
C.,
Nuyujukian,
P.,
Ryu,
S.
I.
&
Shenoy,
K.
V.
A
high
performing
brain–machine
interface
driven
by
low--‐frequency
local
field
potentials
alone
and
together
with
spikes.
J.
Neural
Eng.
12,
36009
(2015).
33.
Homer,
M.
L.,
Nurmikko,
A.
V,
Donoghue,
J.
P.
&
Hochberg,
L.
R.
Sensors
and
decoding
for
intracortical
brain
computer
interfaces.
Annu.
Rev.
Biomed.
Eng.
15,
383–405
(2013).
34.
Kao,
J.
C.,
Stavisky,
S.
D.,
Sussillo,
D.,
Nuyujukian,
P.
&
Shenoy,
K.
V.
Information
Systems
Opportunities
in
Brain--‐Machine
Interface
Decoders.
Proc.
IEEE
102,
666–
682
(2014).
35.
Sutskever,
I.,
Martens,
J.
&
Hinton,
G.
Generating
text
with
recurrent
neural
networks.
in
Proceedings
of
the
28th
International
Conference
on
Machine
Learning
(2011).
36.
Martens,
J.
&
Sutskever,
I.
Learning
recurrent
neural
networks
with
Hessian--‐free
optimization.
Proc.
28th
Int.
Conf.
Mach.
Learn.
ICML
2011
1033–1040
(2011).
37.
Churchland,
M.
M.
&
Shenoy,
K.
V.
Temporal
complexity
and
heterogeneity
of
single--‐
neuron
activity
in
premotor
and
motor
cortex.
J.
Neurophysiol.
97,
4235–57
(2007).
Sussillo,
D.
et
al.
A
recurrent
neural
network
for
closed--‐loop
intracortical
brain–
38.
machine
interface
decoders.
J.
Neural
Eng.
9,
26027
(2012).
39.
Hochreiter,
S.
&
Schmidhuber,
J.
Long
Short--‐Term
Memory.
Neural
Comput.
9,
1735–
1780
(1997).
Page 40
40.
Fan,
J.
M.
et
al.
Intention
estimation
in
brain–machine
interfaces.
J.
Neural
Eng.
11,
16004
(2014).
Chase,
S.
M.,
Schwartz,
A.
B.
&
Kass,
R.
E.
Bias,
optimal
linear
estimation,
and
the
41.
differences
between
open--‐loop
simulation
and
closed--‐loop
performance
of
spiking--‐
based
brain--‐computer
interface
algorithms.
Neural
networks
22,
1203–13
(2009).
42.
Koyama,
S.
et
al.
Comparison
of
brain--‐computer
interface
decoding
algorithms
in
open--‐loop
and
closed--‐loop
control.
J.
Comput.
Neurosci.
29,
73–87
(2010).
43.
Cunningham,
J.
P.
et
al.
A
closed--‐loop
human
simulator
for
investigating
the
role
of
feedback
control
in
brain--‐machine
interfaces.
J.
Neurophysiol.
105,
1932–1949
(2011).
44.
Seo,
D.
et
al.
Wireless
Recording
in
the
Peripheral
Nervous
System
with
Ultrasonic
Neural
Dust
Neuron
NeuroResource
Wireless
Recording
in
the
Peripheral
Nervous
System
with
Ultrasonic
Neural
Dust.
Neuron
91,
1–11
(2016).
Fraser,
G.
W.,
Chase,
S.
M.,
Whitford,
A.
&
Schwartz,
A.
B.
Control
of
a
brain--‐computer
45.
interface
without
spike
sorting.
J.
Neural
Eng.
6,
55004
(2009).
So,
K.,
Dangi,
S.,
Orsborn,
A.
L.,
Gastpar,
M.
C.
&
Carmena,
J.
M.
Subject--‐specific
46.
modulation
of
local
field
potential
spectral
power
during
brain–machine
interface
control
in
primates.
J.
Neural
Eng.
11,
26002
(2014).
47.
Shenoy,
K.
V.
&
Carmena,
J.
M.
Combining
Decoder
Design
and
Neural
Adaptation
in
Brain--‐Machine
Interfaces.
Neuron
84,
665–680
(2014).
48.
Barrese,
J.
C.
et
al.
Failure
mode
analysis
of
silicon--‐based
intracortical
microelectrode
arrays
in
non--‐human
primates.
J.
Neural
Eng.
10,
66014
(2013).
49.
Gold,
C.,
Henze,
D.
a,
Koch,
C.
&
Buzsáki,
G.
On
the
origin
of
the
extracellular
action
potential
waveform:
A
modeling
study.
J.
Neurophysiol.
95,
3113–3128
(2006).
50.
Kao,
J.
C.
et
al.
Investigating
the
role
of
firing--‐rate
normalization
and
dimensionality
reduction
in
brain--‐machine
interface
robustness.
in
2013
35th
Annual
International
Conference
of
the
IEEE
Engineering
in
Medicine
and
Biology
Society
(EMBC)
2013,
293–298
(IEEE,
2013).
51.
Cunningham,
J.
P.
et
al.
A
Closed--‐Loop
Human
Simulator
for
Investigating
the
Role
of
Feedback--‐Control
in
Brain--‐Machine
Interfaces.
J.
Neurophysiol.
(2010).
doi:10.1152/jn.00503.2010
52.
Schwartz,
A.,
Ketter,
R.
&
Georgopoulos,
A.
Primate
motor
cortex
and
free
arm
movements
to
visual
targets
in
three--‐dimensional
space.
I.
Relations
between
single
cell
discharge
and
direction
of
movement.
J.
Neurosci.
8,
2913–2927
(1988).
53.
Riehle,
A.
&
Requin,
J.
Monkey
primary
motor
and
premotor
cortex:
single--‐cell
activity
related
to
prior
information
about
direction
and
extent
of
an
intended
movement.
J.
Neurophysiol.
61,
534–549
(1989).
54.
Wise,
S.
P.,
Moody,
S.
L.,
Blomstrom,
K.
J.
&
Mitz,
A.
R.
Changes
in
motor
cortical
activity
during
visuomotor
adaptation.
Exp.
brain
Res.
121,
285–99
(1998).
Crammond,
D.
&
Kalaska,
J.
Differential
relation
of
discharge
in
primary
motor
cortex
55.
and
premotor
cortex
to
movements
versus
actively
maintained
postures
during
a
reaching
task.
Exp.
Brain
Res.
108,
45–61
(1996).
56.
Messier,
J.
&
Kalaska,
J.
F.
Covariation
of
primate
dorsal
premotor
cell
activity
with
direction
and
amplitude
during
a
memorized--‐delay
reaching
task.
J.
Neurophysiol.
84,
152–165
(2000).
57.
Stark,
E.
&
Abeles,
M.
Predicting
movement
from
multiunit
activity.
J.
Neurosci.
27,
8387–94
(2007).
58.
Perel,
S.
et
al.
Single--‐unit
activity,
threshold
crossings,
and
local
field
potentials
in
Page 41
motor
cortex
differentially
encode
reach
kinematics.
J.
Neurophysiol.
114,
1500–1512
(2015).
Christie,
B.
P.
et
al.
Comparison
of
spike
sorting
and
thresholding
of
voltage
59.
waveforms
for
intracortical
brain–machine
interface
performance.
J.
Neural
Eng.
12,
16009
(2015).
60.
Nuyujukian,
P.
et
al.
Monkey
Models
for
Brain--‐Machine
Interfaces:
The
Need
for
Maintaining
Diversity.
in
33rd
Annual
International
Conference
of
the
IEEE
Engineering
in
Medicine
and
Biology
Society
(2011).
61.
Kaufman,
M.
T.,
Churchland,
M.
M.,
Ryu,
S.
I.
&
Shenoy,
K.
V.
Cortical
activity
in
the
null
space:
permitting
preparation
without
movement.
Nat.
Neurosci.
17,
440–8
(2014).
62.
Suminski,
A.
J.,
Tkach,
D.
C.,
Fagg,
A.
H.
&
Hatsopoulos,
N.
G.
Incorporating
Feedback
from
Multiple
Sensory
Modalities
Enhances
Brain--‐Machine
Interface
Control.
J.
Neurosci.
30,
16777–16787
(2010).
63.
Martens,
J.
Deep
learning
via
Hessian--‐free
optimization.
27th
Int.
Conf.
Mach.
Learn.
951,
735–742
(2010).
64.
Kim,
S.--‐P.,
Simeral,
J.
D.,
Hochberg,
L.
R.,
Donoghue,
J.
P.
&
Black,
M.
J.
Neural
control
of
computer
cursor
velocity
by
decoding
motor
cortical
spiking
activity
in
humans
with
tetraplegia.
J.
Neural
Eng.
5,
455–476
(2008).
65.
Wu,
W.
et
al.
Neural
Decoding
of
Cursor
Motion
Using
a
Kalman
Filter.
Adv.
Neural
Inf.
Process.
Syst.
15
Proc.
2002
Conf.
133–140
(2003).
doi:10.1.1.6.8776
66.
Knyazev,
A.
V
&
Argentati,
M.
E.
Principal
Angles
Between
Subspaces
in
An
A--‐Based
Scalar
Product:
Algorithms
and
Perturbation
Estimates.
SIAM
J.
Sci.
Comput.
23,
2008–2040
(2002).
Page 42
FIGURES (with legends below)
a
100
20
)
z
h
(
e
t
a
R
g
n
i
r
i
F
0
100
l
t
c
e
r
r
o
C
s
a
i
r
T
%
0
20
40
60 80
Time (min)
b
Recording Condition 1
Electrodes
c
Decoder Trained
From Chronic
Recordings
Traditional
Decoding
Dall
D1 D2 D1 D3 D4 D1 D5 D5 D1 D1 D2 D2 D3 D1 D3 D6
D5 D1
Recording Condition 2
Recording Condition 1
day 2
day 1
Time
…
?
previous day now
Recording Condition 3
(broken)
(lost)
(added)
Figure 1. Our strategy for training a decoder robust to recording condition
changes
(a) Example data from a BMI clinical trial showing sudden decoder failure caused by a
recording condition change. The black trace shows the participant's closed-loop
performance over the course of an experiment using a fixed Kalman filter. An abrupt
drop in performance coincides with a reduction in the observed firing rate (red trace) of
a neuron with a high decoder weight. Both the neuron's firing rate and decoder
performance spontaneously recover ~40 minutes later. Adapted from figure 7 of 13. (b) A
cartoon depicting one hypothetical cause of the aforementioned change: micro-motion
of the electrodes leads to Recording Condition 2, in which spikes from the red-shaded
neuron are lost. BMI recovery corresponds to a shift back to Condition 1. Over time,
further changes will result in additional recording conditions; for example, Condition 3 is
shown caused by a disconnected electrode and an additional neuron entering recording
range. (c) Recording conditions (schematized by the colored rectangles) will vary over
the course of chronic intracortical BMI use. We hypothesize that oftentimes new
conditions are similar to ones previously encountered (repeated colors). Typically,
decoders are fit from short blocks of training data and are only effective under that
recording condition (decoders D1, D2, …). Consider the goal of training a decoder for
use at time "now" (black rectangle on right). Standard practice is to use decoder D1
trained from the most recently available data (e.g. from the previous day or the start of
the current experiment). D1 will perform poorly if the recording condition encountered
differs from its training data. To increase the likelihood of having a decoder that will
perform well given the current recording condition, we tested a new class of decoder,
Dall, trained using a large collection of previous recording conditions.
Page 43
a
b
)
2
r
(
y
c
a
r
u
c
c
A
e
d
o
c
e
D
c
)
2
r
(
y
c
a
r
u
c
c
A
e
d
o
c
e
D
vy(t)
vx(t)
px(t)
py(t)
1
Training Days
Monkey R
1
Training Days
Monkey L
37 Training Days
25 Training Days
17 Training Days
11 Training Days
7 Training Days
5 Training Days
3 Training Days
2 Training Days
1 Training Day
7
)
2
r
(
y
c
a
r
u
c
c
A
e
d
o
c
e
D
1
51
Test Day
50 Training Days
38 Training Days
25 Training Days
17 Training Days
11 Training Days
7 Training Days
5 Training Days
3 Training Days
2 Training Days
1 Training Day
1
14
34
Test Day
1
0
r2 = 0.20
47
55
1
0
r2 = 0.24
59
0
1
MRNN
FIT Sameday
FIT Long
1
Recording Day
***
1
0
663
0
1
***
1
0
Recording Day
1024
0
1
0
Figure 2. An MRNN decoder can harness large training datasets
(a) Schematic of the experimental setup and MRNN decoder. A monkey performed a
target acquisition task using his hand while multiunit spikes were recorded from
multielectrode arrays in motor cortex. Data from many days of this task were used to
train two MRNNs such that velocity and position could be read out from the state of their
respective internal dynamics. These internal dynamics are a function of the binned spike
counts; thus, the MRNN can conceptually be thought of as selecting which internal
dynamics rules are active at any given time based on the input signals. During closed-
loop use, the decoded velocity and position outputs were blended together to
continuously update the on-screen cursor. (b) The MRNN was better prepared for future
days' recording conditions after being trained with more previously collected data. We
evaluated its ability to reconstruct offline hand velocity on 12 (16) monkey R (L) test
days after training with increasing numbers of previous days' datasets. Training data
Page 44
were added by looking further back in time so as to not conflate training data recency
with data corpus size. In monkey R the early test days also contributed training data
(with the test trials held out). In monkey L, from whom more suitable data was available,
the training datasets start with the day prior to the first test day. Thus, except for the few
monkey R overlap days, all training data comes chronologically before the test days.
More training data (darker colored traces) improved decode accuracy, especially when
decoding more chronologically distant recording conditions (later test days). For
comparison, we also plotted performance of a traditional decoder (FIT Kalman filter)
trained from each individual day's training data ("FIT Sameday", light blue). (Insets)
show the same MRNN data in a scatter plot of decode accuracy versus number of
training days (99 data points for monkey R, 160 for L), with linear fit trend lines revealing
a significant positive correlation. (c) An MRNN could successfully learn even larger
datasets spanning many more recording days. An MRNN (red trace) was trained with
data from 154 (250) monkey R (L) recording days spanning many months. Its offline
decoding accuracy on held-out trials from each of these same days was compared to
that of the FIT Sameday (light blue). We also tested a single FIT-KF trained using the
same large dataset as the MRNN ("FIT Long", dark blue). Gaps in the connecting lines
denote recording gaps of more than ten days. (Insets) mean ± s.d. decode accuracy
across all recording days. Stars denote p < 0.001 differences (signed-rank test). The
MRNN outperformed both types of FIT-KF decoders on every day's dataset.
Page 45
30
e
t
u
n
M
i
r
e
P
s
t
e
g
r
a
T
0
Monkey R
*
*
*
30
*
Monkey L
*
*
5/9 6/7 6/7
0/9
1/7
0
1/7
1/7
0/7
2 3
10
Top N Electrodes Dropped
1/7
1/7
7
2/7
2/7
5
0
7/11
3/11
3
1/4
2/8
0/4
0/8
0
10
Top N Electrodes Dropped
5/7
0/7
5
5/5
0/5
7
Figure 3. Robustness to unexpected loss of the most important electrodes
Closed-loop BMI performance using the MRNN (red) and FIT Sameday (blue) decoders
while simulating an unexpected loss of up to 10 electrodes by setting the firing rates of
these electrodes to zero. The mean and s.e.m. across experimental sessions' targets
per minute performance is shown for each decoder as a function of how many
electrodes were removed. Stars denote conditions for which the MRNN significantly
outperformed FIT Sameday across sessions (p < 0.05, paired t-test). The text above
each condition's horizontal axis tick specifies for how many of the individual evaluation
days MRNN (red fraction) or FIT Sameday (blue fraction) performed significantly better
according to single-session metrics of success rate and time to target. Electrode-
dropping order was determined by the mutual information between that electrode's
spike count and target direction during arm-controlled reaches.
Page 46
Monkey R
Decoder Set 2 Training Data
MRNN
FIT Long
FIT Old
Arm Control
FIT Sameday
Decoder Set 1 Training Data
MRNN
FIT Long
FIT Old
384
Day 1
30
e
t
u
n
M
i
r
e
P
s
t
e
g
r
a
T
0
Monkey L
Decoder Set 2 Training Data
MRNN
FIT Long
FIT Old
Decoder Set 1 Training Data
MRNN
FIT Long
FIT Old
706
Day 1
810
30
e
t
u
n
M
i
r
e
P
s
t
e
g
r
a
T
0
964
447
558
Set 1 Test{{Set 2 Test
Set 1 Test{{Set 2 Test
Figure 4. Robustness to naturally occurring recording condition changes
Robustness to a natural sampling of neural input variability. We created decoder
evaluation conditions in which the neural inputs were likely to be different from much of
the training data by withholding access to the most recent several months of data. Each
circle corresponds to the mean closed-loop BMI performance using these "stale" MRNN
(red), FIT Long (dark blue), and FIT Old (teal) decoders when evaluated on six different
experiment days spanning 7 (13) days in monkey R (L). Each test day, these three
decoders, as well as a FIT Sameday decoder trained from that day's arm reaches, were
evaluated in an interleaved block design. The legend bars also denote the time periods
from which training data for each stale decoder came from. We repeated the
experiments for a second set of decoders to reduce the chance that the results were
particular to the specific training data gap chosen. The training data periods contained
82 and 92 datasets (monkey R), and 189 and 200 training datasets (monkey L). The
only decoder that was consistently usable, i.e. did not fail on any test days, was the
MRNN. To aid the interpretation of these stale decoder performances, we show the
average performance across the six experiment days using arm control (gray dashed
line) or a FIT Sameday decoder (blue dashed line).
Page 47
3
6
2
6
4 cm
7
Monkey R
5
1
b
)
8
m
c
(
t
e
g
r
a
T
o
4
t
e
c
n
a
t
s
0
D
i
0
250
500
Trial Time (ms)
a
7
Arm Control
8
4
MRNN
1
3
FIT Sameday
2
6
8
2
4
8
4
0
0
750
3
1
4
8
5
7
5
Monkey L
250
500
Trial Time (ms)
750
Figure 5. MRNN has high performance under "ideal" conditions
(a) We compared cursor control using the MRNN (red) trained from many datasets up
through the previous day to the FIT Sameday (blue) trained from data collected earlier
the same day, without any artificial challenges (i.e. no electrodes dropped or stale
training data). Cursor trajectories are shown for eight representative and consecutive
center-out-and-back trials of the Radial 8 Task. Gray boxes show the target acquisition
area boundaries, and the order of target presentation is denoted with green numbers.
For comparison, cursor trajectories under arm control are shown in gray. From dataset
R.2014.04.03.
(b) Mean distance to target, across all Radial 8 Task trials under these favorable
conditions, as a function of trial time using each cursor control mode. Thickened
portions of each trace correspond to "dial-in time", i.e. the mean time between the first
target acquisition and the final target acquisition. These MRNN and FIT Sameday data
correspond to the drop 0 electrodes condition data in Figure 3, and include 4,094
(3,278) MRNN trials and 4119 (3,305) FIT Sameday trials over 9 (8) experimental days
in Monkey R (L).
Page 48
Supplementary Information
a
Monkey R
Session # 1
8
16
24
32
40
48
56
64
71
Monkey L
1
14
28
42
56
70
84
98
112
125
b
1 1
Recording Session #
71
20
1
1
Recording Session #
125
30
)
°
(
e
g
n
A
l
i
l
a
p
c
n
i
r
P
.
n
M
i
i
#
n
o
s
s
e
S
g
n
d
r
o
c
e
R
i
2
#
i
n
o
s
s
e
s
4
2
#
i
n
o
s
s
e
s
2
8
4
#
i
n
o
s
s
e
s
0
7
#
i
n
o
s
s
e
s
125
32
2
#
i
n
o
s
s
e
s
2
4
#
i
n
o
s
s
e
s
4
8
#
i
n
o
s
s
e
s
5
4
2
1
#
i
n
o
s
s
e
s
Recording Day
658
3
1
34%
Recording Day
1003
i
#
n
o
s
s
e
S
g
n
d
r
o
c
e
R
i
70
c
20
e
c
a
p
s
b
u
S
.
i
n
M
)
°
(
l
e
g
n
A
2
d
y
c
n
e
u
q
e
r
F
e
c
n
e
r
u
c
c
O
1
35%
1
10
20
Number of Recording Sessions Apart
30
1
10
20
30
40
Number of Recording Sessions Apart
Supplementary Figure 1. For a given day, similar neural recording conditions can be
found on some other day(s). Chronologically close days tend to have more similar neural
recordings, but for a given day there are occasional similar recordings from more distant days.
(a) To minimize the potential effect of behavioral variability on neural variability, we restricted
this analysis to recording sessions with very consistent Radial 8 Target task behavior. Hand
velocity correlations between all pairs of sessions within the included set were at least 0.9.
Representative hand position traces (mean over trials towards each target) are shown for ten
sessions spanning the months analyzed. (b) Between-day variability of the structure of neural
activity recorded during reaches over the course of many months (71 recording sessions over a
658 day period in monkey R, and 125 sessions spanning 1003 days in monkey L; these
correspond to a subset of the days included in Fig. 2c). The color at the intersection of row i
Page 49
and column j corresponds to how differently the observed neural activity covaried during
recording sessions i and j. Specifically, we have plotted the minimum principal angle between
subspaces spanned by the top 10 eigenvectors of each day's mean-activity-subtracted
covariance matrix (see Methods). These 10 eigenvectors captured on average 51 (46)% of
single-trial variance for monkeys R (L). Sharp "block" structure transitions typically correspond
to a long (many weeks') interval between consecutive recording sessions. (c) Four slices
through each monkey's principal angle matrix show that for these example days, there were
similar recording conditions on other days both near and far in time. Each color series shows
the minimum principal angle between every recording day's data and the reference day
marked with this color's arrow in panel b. Note that the horizontal axis, which spans the same
time range as in panel b, is now in units of (calendar) days rather than session number. Each
series' reference day is marked with a vertical line. (d) Histograms showing the distribution,
across each monkey's recordings, of how many recording sessions apart (either forward or
back in time) we observed the most similar neural correlates of reaching as measured by
minimum principal angle.
Monkey R
Dropped
Electrode
192
160
140
120
100
80
60
40
20
e
d
o
r
t
c
e
E
l
)
z
H
(
e
t
a
R
g
n
i
r
i
F
e
g
a
r
e
v
A
e
d
o
r
t
c
e
E
l
96
80
60
40
20
40
30
20
10
0
Monkey L
)
z
H
(
e
t
a
R
g
n
i
r
i
F
e
g
a
r
e
v
A
20 40 60 80 100 120 140 160 180
1
Previous Training Data
212 231
Experiments
60
50
40
30
20
10
0
20
1
Previous Training Data
40
60
80 100
119
129
Experiments
Recording Session #
Recording Session #
Supplementary Figure 2. Artificially dropped electrodes were active in the training data.
These plots show each electrode's average firing rate during each dataset used to train the
MRNN; electrodes are ordered by descending average firing rate across all recording sessions.
Recording sessions numbered in gray were only used for training data. The electrode dropping
experiments (Fig. 3) were conducted during the sessions numbered in black. Zero firing rates
(i.e. non-functional electrodes) are shown in purple for emphasis, while electrodes selected for
dropping on a particular day are shown in red (note that although on a given test session we
evaluated different numbers of electrodes dropped, this plot shows each day's broadest
dropped set). These dropped electrodes rarely recorded zero firing rates in the training data
sessions, and the specific sets of dropped electrodes used to challenge the decoders never all
had zero firing rates in the training data.
Page 50
a
30
e
t
Monkey R
30
Monkey L
u
n
M
i
r
e
P
s
t
e
g
r
a
T
0
b
800
600
)
s
m
(
t
e
g
r
a
T
o
400
t
e
m
T
i
200
0
2/2
0/2
7
1/1
0/1
5
2/2
0/1
0/1
0/2
2 3
2/2
0/2
0/2
0/2
10
0
Top N Electrodes Dropped
Monkey R
Training Set 1
Training Set 2
0
not
1/3
tested
0/3
0
10
Top N Electrodes Dropped
1/1
0/1
7
2/2
0/2
5
1/1
0/1
3
Monkey L
Training Set 1
800
600
400
200
0
l
s
a
i
r
t
5
1
2
=
n
4
1
2
=
n
4
2
2
=
n
6
2
2
=
n
8
6
2
=
n
9
2
2
=
n
Block
1600
1200
800
400
0
6
0
2
=
n
5
1
2
=
n
0
1
2
=
n
9
0
2
=
n
8
1
2
=
n
5
3
2
=
n
Block
0
8
1
=
n
3
9
1
=
n
6
8
1
=
n
7
9
1
=
n
3
7
1
=
n
6
9
1
=
n
Block
Supplementary Figure 3. Training data spike rate perturbations improve closed-loop
MRNN robustness.
(a) Robustness to electrode dropping. We evaluated the closed-loop BMI performance of the
MRNN decoder trained with (red) and without (brown) the spike rate perturbations training data
augmentation. Both decoders were evaluated on the same day with firing rates on varying
numbers of the most informative electrodes set to zero (similar to Fig. 3). Each circle
corresponds to a decoder's targets per minute performance on a given evaluation day. In total
there were 3 sessions per monkey. Filled circles denote conditions where there was a
significant within-session performance difference between the two decoders tested according
to: p < 0.05 binomial test on success rate, followed, if success rate was not significantly
different, by a more sensitive comparison of times to target (p < 0.05, rank-sum test). Fractions
above the horizontal axis specify for how many of the individual evaluation days each decoder
performed significantly better. Trend lines show the across-sessions mean targets per minute
performance for each decoder. The MRNN trained with perturbed firing rates outperformed the
MRNN trained without data augmentation when encountering electrode-dropped neural input.
(b) Robustness to naturally occurring neural recording condition changes. MRNNs were trained
without access to recent training data, as in the Fig. 4 stale training data experiments, either
with (red) or without (brown) training data spike rate perturbations. We trained decoders from
both of monkey R's stale training data periods and from monkey L's longer stale training data
period. Closed-loop BMI performance using these decoders was then compared on the same
evaluation day in alternating blocks. Bars show mean ± s.e.m. time to target for each block of
trials (success rates using both training paradigms were close to 100%). The MRNN with spike
rate perturbations had significantly faster times to target in monkey R (p < 0.05, rank-sum test
aggregating trials across blocks) and showed a trend of faster times to target in monkey L (p =
0.066). Datasets R.2014.03.21 & L.2014.04.04.
Page 51
Monkey R
)
2
r
(
y
c
a
r
u
c
c
A
e
d
o
c
e
D
0.8
0.6
0.4
0.2
0
MRNN Big w/ SP
MRNN Big no SP
MRNN Small w/SP
MRNN Small no SP
MRNN 1 day w/SP
MRNN 1 day no SP
FIT Sameday
FIT Long
0
1
2
3
7
Top N Electrodes Dropped
4
5
6
0.8
0.6
0.4
0.2
0
8
9
0
1
2
Monkey L
3
4
Top N Electrodes Dropped
5
6
7
8
9
Supplementary Figure 4. Offline decode experiments testing various training paradigms'
robustness to unexpected loss of electrodes
We performed offline decoding analyses to test how each of the three main components of our
method – the use of an MRNN architecture instead of a linear Kalman filter, the use of large
training datasets, and the spike count perturbation data augmentation – contributed to
improved robustness to unexpected loss of the most important electrodes, similar to the online
tests shown in Figure 3. The results suggest that both the use of the MRNN architecture and
training data augmentation contributed to the complete system's improved robustness to a
novel recording condition consisting of electrode dropping.
We trained MRNNs with different quantities of data: 1 day of (held out) data from the
test day, a 'Small' dataset consisting 10-13 days up to and including the test day, or a "Big"
dataset of 40 – 100 days up to and including the test day, with ("w/SP") or without ("no SP")
additional spike count perturbations during training. We also trained a FIT-KF Sameday
decoder and a FIT-KF Long which used the same datasets as the MRNN Big datasets. We
compared the offline decoding accuracy of each decoder as a function of the number of
electrodes dropped, using the same electrode dropping order determination method as in the
Figure 3 online experiments. Three decoders were trained for each training paradigm using
data from different periods of each monkey's research career; these decoders' training dates
correspond to exactly the same as those in Supplementary Figure 6, and each decoder was
tested on held out data from its last day of training. We averaged offline hand velocity
reconstruction accuracy across each monkey's three testing days. We found that applying the
spike count perturbation always increased MRNN robustness to electrode dropping (compiling
r2 across all SP vs all no SP decoders across all three test days, SP decoders performed better
than no SP decoders, p < 0.001, signed-rank test). Note that, as expected, training with larger
dataset sizes did not strongly affect performance or robustness to electrode dropping, since all
MRNN's have 'seen' data collected on the same day as the withheld testing data.
Page 52
35
e
t
u
n
M
i
r
e
P
s
t
e
g
r
a
T
0
Monkey R
35
Monkey L
FIT Old from 2013-09-19
FIT Old from 2014-02-03
FIT Old from 2013-04-16
FIT Old from 2013-11-21
FIT Old from 2012-09-21
0
FIT Old from 2013-10-22
FIT Old from 2013-10-09
FIT Old from 2013-07-10
FIT Old from 2011-08-04
Days Between Training Data and Decoder Test
272
965 978
31
Days Between Training Data and Decoder Test
559
155
Supplementary Figure 5. Additional tests showing that FIT Old typically performs poorly.
We investigated whether our result that three of the four different FIT Old decoders tested in
the main stale training data experiments (Fig. 4) failed was due to a particularly unlucky choice
of FIT Olds. To better sample the closed-loop performance of FIT-KF decoders trained using
old training data, we trained FIT Old decoders from 3 (monkey R) and 2 (Monkey L) additional
arbitrarily chosen arm reaching datasets from the monkey's prior experiments. We evaluated all
5 (4) FIT Old decoders on a number of additional days over the course of the current study (8
total test days for monkey R, 13 total test days for monkey L). Each point shows the
performance of a particular FIT Old decoder on one test day. Different days' evaluations of the
same FIT Old decoder are shown in the same color. Black borders denote data points and
black underlines denote decoders that are shared with Fig. 4.
Page 53
)
2
r
(
y
c
a
r
u
c
c
A
e
d
o
c
e
D
1
0.8
0.6
0.4
0.2
0
Monkey R
w/SP
no SP
w/SP
MRNN 1 day MRNN Small MRNN Big
no SP
no SP
w/SP
1
0.8
0.6
0.4
0.2
0
FIT
Old
FIT
Long
Monkey L
w/SP
no SP
w/SP
MRNN 1 day MRNN Small MRNN Big
no SP
no SP
w/SP
FIT
Old
FIT
Long
Supplementary Figure 6. Offline decode experiments testing various training paradigms'
robustness to naturally occurring recording condition changes
We performed offline decoding analyses to test how each of the three main components of our
method – the use of an MRNN architecture instead of a linear Kalman filter, the use of large
training datasets, and the spike count perturbation data augmentation – contributed to
improved robustness to naturally occurring recording condition changes similar to the online
'stale' training data tests shown in Figure 4. We performed offline decoding evaluations across
three different training data gaps and found that using more previously collected data and
incorporating data augmentation both improved the MRNN's performance on the future test
data. MRNNs trained with many datasets outperformed FIT-KFs trained using the same data
("FIT Long") or just the most recent data ("Fit Old"). These results suggest that all three
components of the method contributed to the complete system's improved robustness.
(Left) Offline decode results for Monkey R. We performed an offline decode for 8 different types
of decoders and aggregated each decoder's performance over the three gaps. The MRNN
decoders were either trained with 1 day of data "1 day," a "Small" dataset (gap 1: 13 datasets,
gaps 2 and 3: 10 datasets) or a "Big" dataset (gap 1: 40 datasets, gap 2: 44 datasets, gap 3:
37 datasets). The MRNN decoders were also either trained with no spike rate perturbations
("no SP") or with spike rate perturbations ("w/SP"). We also trained a FIT Old using the most
recent dataset and a FIT Long which used the same datasets as the MRNN Big datasets. Gap
1 comprised training data from 2012-10-22 (YYYY-MM-DD) to 2013-04-19 and testing data
from 2013-07-29 to 2013-11-21 (44 testing days). Gap 2 comprised training data from 2013-
07-29 to 2013-11-21 and testing data from 2014-02-03 to 2014-04-07 (37 testing days). Gap 3
comprised training data from 2014-02-03 to 2014-04-07 and testing data from 2014-06-16 to
2014-08-19 (33 testing days). The bars show the mean ± s.d. performance of each training
approach across all 3 gaps. We observed the same trends across individual gaps, with the
MRNN Big w/SP decoder always achieving the best performance (p < 0.01, signed-rank test
with every other decoder, all gaps).
(Right) Same for Monkey L. The Small datasets comprised 10 datasets (gap 1 and 2) or 11
datasets (gap 3), while the Big datasets comprised 103 datasets (gap 1), 105 datasets (gap 2),
and 77 datasets (gap 3). Gap 1 comprised training data from 2010-03-04 to 2010-10-26 and
testing data from 2011-01-17 to 2011-04-28 (51 testing days). Gap 2 comprised training data
from 2011-01-18 to 2011-10-04 and testing data from 2012-04-02 to 2012-07-19 (51 testing
days). Gap 3 comprised training data from 2012-04-02 to 2012-10-12 and testing data from
2013-01-26 to 2013-07-10 (37 testing days). Across individual gaps, the same trends showed
were displayed, with the MRNN Big w/SP decoder always achieving the best performance (p <
0.01, signed-rank test with every other decoder, all gaps, except in Gap 2 when comparing to
MRNN Small w/SP, p = 0.1, and Gap 3 where it on average achieved a lower r2 than MRNN
Small w/SP and MRNN Big no SP).
Page 54
a
0.4
l
s
a
i
r
T
Monkey R
474 ms median Arm Control (n = 1,018)
562 ms median MRNN (n = 649)
0.4
Monkey L
557 ms (n = 519)
692 ms (n = 207)
f
o
n
o
i
t
c
a
r
F
b
0
0
1000 2000 3000 4000
Time to Target (ms)
fail
0
0
1000 2000 3000 4000
Time to Target (ms)
fail
4 cm
fail
2001-5000
2000
)
s
m
(
t
1100
e
g
r
a
t
o
t
e
m
T
i
200
Page 55
Supplementary Figure 7. Closed-loop MRNN decoder performance on the Random
Target Task.
Both monkeys were able to use the MRNN decoder to acquire targets across a broad
workspace in which targets often appeared at locations that differed from the target locations
dominating the training datasets.
(a) Histograms of Random Target Task times to target (time of final target entry minus target
onset time, not including the 500 ms target hold period) using the MRNN decoder are shown in
red. For comparison, histograms of performance on the same task using arm control are
shown in gray.
(b) Task workspace plots showing the location of each Random Target Task trial's target
during MRNN decoder evaluation. Each point corresponds to the center of one trial's target,
and its color represents the time it took the monkey to acquire this target. The location of the
one failed trial (for monkey R) is shown with a black 'x'. The acquisition area boundaries of the
nine Radial 8 Task targets used for the majority of the training data are shown as black
squares. Monkey R's data are aggregated across the two experimental sessions in which he
performed this task. Monkey L's data are from one session.
Supplementary Video 1. Example closed-loop performance of the MRNN
The MRNN was trained using reaching data from 125 recording sessions up through
the previous day. The video shows a continuous 90 seconds of monkey R using this
decoder to perform the Radial 8 Task. He controls the white cursor and acquires the
green target (which turns blue during the 500 ms target hold period). Dataset
R.2014.04.03. This is a portion of the data used to generate the drop 0 electrodes
condition of Fig. 3.
Supplementary Video 2. Side-by-side comparison of the MRNN and FIT Sameday
decoders after electrode dropping
During the experiment, the two decoders were evaluated in alternating blocks after the
same 3 most important electrodes were dropped. Here we show a continuous 60
seconds of each decoder's closed-loop performance from consecutive blocks. The
MRNN (right side) was trained using reaching data from 125 recording sessions up
through the previous day, while the FIT Kalman filter (left side) was trained using
reaching data from earlier that same day. Dataset monkey R.2014.04.03. This is a
portion of the data used to generate the drop 3 electrodes condition of Fig. 3.
Page 56
Supplementary Data 1. Array recording quality measurements across this study's
recording days
For task consistency, these analyses were restricted to those recording days when
Baseline Block data was collected. Reach direction tuning was calculated as in 32: for
each recording day, we calculated each electrode's average firing rate over the course
of each trial (analysis epoch: 200 to 600 ms after target onset) to yield a single data
point per trial, and then grouped trials by target location. 69.0 ± 7.7 of monkey R's
electrodes (mean ± s.d. across recording days) and 66.3 ± 12.3 of monkey L's
electrodes exhibited significantly different firing rates when reaching to at least one of
the eight different targets (p < 0.01, one-way ANOVA). These tuned electrodes'
modulation range, defined here as the trial-averaged rates firing rate difference
between reaches to the two targets evoking the highest and lowest rates, was 26.4 ±
5.9 Hz in monkey R monkey (mean ± s.d., averaged first across all electrode pairs in a
given recording day, and then over days) and 23.8 ± 5.6 Hz in monkey L. We did not
observe cross-talk between electrodes' threshold crossings, consistent with recording
spiking activity from electrodes at least 400 μm apart: pairwise electrode cross-
correlations, computed using the time-series of firing rates in consecutive non-
overlapping 20 ms bins spanning a given day's Baseline Block, was 0.0089 ± 0.0021 in
monkey R (mean ± s.d., averaged first across all electrode pairs in a given recording
day, and then over days) and 0.0150 ± 0.0058 in monkey L.
Page 57
|
1309.3535 | 1 | 1309 | 2013-09-13T18:33:55 | Identification of criticality in neuronal avalanches: II. A theoretical and empirical investigation of the driven case | [
"q-bio.NC",
"math-ph",
"math-ph"
] | The observation of apparent power-laws in neuronal systems has led to the suggestion that the brain is at, or close to, a critical state and may be a self-organised critical system. Within the framework of self-organised criticality a separation of timescales is thought to be crucial for the observation of power-law dynamics and computational models are often constructed with this property. However, this is not necessarily a characteristic of physiological neural networks - external input does not only occur when the network is at rest/a steady state. In this paper we study a simple neuronal network model driven by a continuous external input (i.e.\ the model does not have a separation of timescales) and analytically tuned to operate in the region of a critical state (it reaches the critical regime exactly in the absence of input - the case studied in the companion paper to this article). The system displays avalanche dynamics in the form of cascades of neuronal firing separated by periods of silence. We observe partial scale-free behaviour in the distribution of avalanche size for low levels of external input. We analytically derive the distributions of waiting times and investigate their temporal behaviour in relation to different levels of external input, showing that the system's dynamics can exhibit partial long-range temporal correlations. We further show that as the system approaches the critical state by two alternative `routes', different markers of criticality (partial scale-free behaviour and long-range temporal correlations) are displayed. This suggests that signatures of criticality exhibited by a particular system in close proximity to a critical state are dependent on the region in parameter space at which the system (currently) resides. | q-bio.NC | q-bio |
Identification of criticality in neuronal avalanches: II. A the-
oretical and empirical investigation of the driven case
Caroline Hartley1,2, Timothy J Taylor3 , Istvan Z Kiss4 , Simon F Farmer5,6 and Luc Berthouze∗2,3
1 Centre for Mathematics and Physics in the Life Sciences and Experimental Biology, University College London, London, UK.
2 Institute of Child Health, University College London, London, UK.
3 Centre for Computational Neuroscience and Robotics, University of Sussex, Falmer, UK.
4 School of Mathematical and Physical Sciences, Department of Mathematics, University of Sussex, Falmer, UK.
5 National Hospital of Neurology and Neurosurgery, London, UK.
6 Institute of Neurology, University College London, London, UK.
Email: Caroline Hartley - [email protected]; Timothy J Taylor - [email protected]; Istvan Z Kiss - [email protected];
Simon F Farmer - [email protected]; Luc Berthouze - [email protected];
∗Corresponding author
Abstract
The observation of apparent power-laws in neuronal systems has led to the suggestion that the brain is at,
or close to, a critical state and may be a self-organised critical system. Within the framework of self-organised
criticality a separation of timescales is thought to be crucial for the observation of power-law dynamics and
computational models are often constructed with this property. However, this is not necessarily a characteristic of
physiological neural networks - external input does not only occur when the network is at rest/a steady state. In
this paper we study a simple neuronal network model driven by a continuous external input (i.e. the model does
not have a separation of timescales) and analytically tuned to operate in the region of a critical state (it reaches
the critical regime exactly in the absence of input - the case studied in the companion paper to this article). The
system displays avalanche dynamics in the form of cascades of neuronal firing separated by periods of silence.
We observe partial scale-free behaviour in the distribution of avalanche size for low levels of external input. We
analytically derive the distributions of waiting times and investigate their temporal behaviour in relation to different
levels of external input, showing that the system's dynamics can exhibit partial long-range temporal correlations.
We further show that as the system approaches the critical state by two alternative 'routes', different markers
of criticality (partial scale-free behaviour and long-range temporal correlations) are displayed. This suggests that
signatures of criticality exhibited by a particular system in close proximity to a critical state are dependent on the
1
region in parameter space at which the system (currently) resides.
Introduction
In recent years, apparent power-laws (i.e. where a power-law is the best model for the data using a model
selection approach [1,2]) have been observed experimentally in neurophysiological data leading to the sugges-
tion that the brain is a critical system [3, 4]. These observations have included that of neuronal avalanches -
cascades of neuronal firing recorded in vivo and in vitro whose size and duration appear to follow power-law
distributions [5 -- 9]. Recently it has been claimed that equivalent neuronal avalanche behaviour with the same
power-law relationship can be identified in human MEG (magnetoencephalography) recordings [10]. On a
wider scale, fluctuations in oscillation amplitude in human (adult and child) EEG (electroencephalography)
and MEG exhibit a power-law decay of the autocorrelation function of the signal - a property known as
long-range temporal correlations (LRTCs) [4, 11 -- 15]. These observations and the idea that the brain is a
critical system have drawn much attention as critical systems have been shown to exhibit optimal dynamic
range and optimal information processing [16,17]. Moreover, it has led to the hypothesis that brain dynamics
may fit within the framework of self-organised criticality (SOC), i.e. a system that does not require external
tuning of parameters to reach the critical state [4, 18, 19].
While the observation of power-laws within neuronal activity may be attractive we must address the
issue of whether (specifically) a neuronal system in the region of a critical state can produce this type of
dynamics. Propagation of the spiking of neurons within a network has been interpreted within the context
of percolation dynamics and the theory of branching processes [20, 21]. A critical branching process is a
process such that one active node will activate on average one other node at the next time-step and so one
can discern how this would relate to neuronal systems whereby the system is critical if one active neuron
on average activates one other neuron at the next time-step. A critical branching process will display
power-law dynamics, however, a number of assumptions underlying branching processes do not hold true in
neurophysiological systems. Firstly, the theoretical analysis of branching processes relies on full-sampling of
the system. Full-sampling is unlikely to occur in the experimental setting and this can have a profound effect
on the observed distributions [22]. Additionally, re-entrant connections invalidate the standard theory of
branching processes [21] which brings into question the idea that neuronal systems can be modelled as critical
branching processes. Moreover, the strict definition of a critical system is one that operates at a second order
phase transition which applies only to systems with infinite degrees of freedom. Therefore, we may expect
a critical system to exhibit an exact power-law distribution in the case of infinite size but what should we
2
expect if the system is finite? As neuronal systems are necessarily finite this is an important question in the
field but one that has yet to be fully addressed. Within experimental results this fact has been accounted
for by the concept of finite-size effects - where a power-law is observed up to a cut-off value [2, 5, 6, 19]. This
cut-off value has been suggested to coincide with the size of the system and distributions from networks of
different sizes have been shown to exhibit an exact scaling relationship - a phenomenon known as finite-size
scaling [2, 23]. However, the finite-size effect with a cut-off value at system size has been assumed without
analytical derivation (though, see the companion paper to this article [24], as described below) and the
question of how a finite critical system behaves and the types of dynamics possible from such a system
remains open in the field. Whether a finite-size system should display the same signatures of criticality as
the system in the limit of system size is not known.
In the companion paper to this article [24] we examined a computational model of a finite neuronal
system analytically tuned to its critical state, defined as a transcritical bifurcation. There we showed that
the dynamics of the system, which by analogy with experimental neuronal avalanches could be termed
avalanches (discrete cascades of neuronal firing), exhibited scaling which does not follow an exact power-law
but does exhibit partial scale-free behaviour. We were able to show that the cut-off value is approximately
the system size, as suggested experimentally by the finite-size effect, but is analytically related to the lead
eigenvalue of the transition matrix (the matrix of all possible transitions at each simulation step). This is an
important observation given that avalanches in systems with re-entrant connections could in principle be of
infinite size and yet experimental observations have suggested that neuronal avalanches exhibit a finite-size
cut-off [2, 5]. Overall, the results suggested that finite systems at criticality exhibit signatures of critical
systems dynamics but do not (at least in this instance) exhibit exact power-laws as had previously been
suggested.
While the system studied in the companion paper leads us to a greater understanding of the dynamics
displayed by a finite neuronal system, there is still an important difference between the system studied there
and physiological neuronal systems.
In the companion paper the system was seeded by setting a single
neuron in the network into the active state and an avalanche was defined as the firing that occurred until the
network returned to a stable state (the fully quiescent state). After this point no more firing could occur until
the system was reseeded. This imposed a strict separation of timescales, with all avalanches and neuronal
firing occurring on a much faster timescale than the timescale of the 'external input' reseeding the system.
Many other computational models have also taken this approach [18, 25, 26], with a separation of timescales
thought to be necessary for the observation of self-organised critical dynamics [23]. While a separation of
3
timescales is likely to occur in some natural systems such as earthquakes, where friction in the Earth's plates
build up over the course of years but energy is released in a matter of minutes, this is not a physiologically
realistic assumption for a neuronal system. External input (be it from the environment or other areas of
the nervous system) will not arrive only once the neuronal population has returned to a set state. Before
physiological recordings can be interpreted within the field of critical systems we must address the question
of the types of dynamics that should be expected by not only a finite-size system but also a system that is
driven by a physiologically realistic external input. Can a finite-size system without an explicit separation
of timescales in the region of a critical regime exhibit markers of criticality? How might the external input
to the system affect these markers?
Previous authors examining computational neuronal networks with continuous driving (i.e. no explicit
separation of timescales) have observed power-law dynamics [16, 27 -- 29].
In particular, Kinouchi and
Copelli [16] and Larremore et al. [29] analytically determined the parameters required such that the model
they studied was at criticality and displayed peak dynamic range, in fully connected networks and networks
with a range of topologies, respectively. However, these authors did not explicitly examine the firing dy-
namics of the system in the region of the critical regime, concentrating on average activity levels.
In a
SOC system such as the sandpile model [18] the waiting times (periods of inactivity between avalanches)
have been shown to follow an exponential distribution [30]. However, these waiting times are related to the
reseeding of the system - sand is added to cells chosen at random and the next avalanche begins when a cell
exceeds the threshold. In contrast, recent experimental work has shown that waiting times between neuronal
avalanches in cultures have a distribution with two trends - a (short) initial power-law region thought to
relate to neuronal up-states and a bump in the distribution at longer waiting times thought to relate to neu-
ronal down-states [31]. Could this difference in these waiting time distributions (between the SOC sandpile
model and the neuronal avalanches in culture) be explained by the fact that physiological neuronal systems
do not have a separation of timescales?
As mentioned previously, another signature of criticality that has been reported in neural systems is
the presence of LRTCs. In the majority of cases they have been observed in large scale neuronal signals
such as human brain oscillations. Recent endeavours have been made to link these observations of scale-
free behaviour on large scales with neuronal avalanches [32, 33]. Poil and colleagues demonstrated in a
computational neuronal network that power-law distributed avalanches and LRTCs in oscillations emerge
concurrently.
In addition, LRTCs have also been detected in the waiting times of bursts of activity in
cultures [34] and the discontinuous burst activity recorded from extremely preterm human neonates [35].
4
Thus, LRTCs have been demonstrated in discrete neuronal activity yet they have not been examined in the
waiting times of neuronal avalanches themselves. While LRTCs in avalanche activity would not be possible
in a seeded computational system (where the activity is initiated 'by hand' and there is no memory within
the system's dynamics) it is conceivable that a driven system, which is more akin to physiological networks
which can display LRTCs, might display this type of dynamics in the waiting times of neuronal avalanches.
In summary, in this paper we aim to address the following questions:
1. Assuming that the brain, or population of neurons under study, operates in the region of a critical
regime can it be expected to display power-law statistics given that it is a finite-size system? If not
what distribution should we expect? As discussed, this question was also addressed in the companion
paper [24], where we studied a system without an external input. However, here we specifically
consider this question in the context of a driven system, i.e. with a non-zero external input and no
explicit separation of timescales.
2. Can we expect a finite-size neuronal system in the region of a critical regime to exhibit other markers
of criticality, and specifically the presence of LRTCs? Does the presence of LRTCs relate to that of
power-law distributions? As described above, LRTCs have been observed in neurophysiological data
sets. However, a systematic examination of how LRTCs may relate to other markers of criticality in
neuronal systems is lacking.
3. How are signatures of criticality (power-law distributions and LRTCs) affected by proximity to the
critical regime? One might assume that a system which is closer to a critical regime may exhibit
signatures of criticality, whereas a system that is further from the critical regime will not. Importantly,
our analysis shows that this assumption is in fact not (always) true.
In this paper, as in the companion paper, we examine a purely excitatory stochastic neuronal model. As
in the companion paper, a number of assumptions are made to simplify the model with the outcome that it is
analytically tractable and therefore can be tuned to operate in the region of a critical regime. This approach
is taken as it allows direct exploration of the above questions, which would not be possible with a more
complex system. We begin by examining the distributions of avalanche size and duration, investigating the
presence of scale-free behaviour. We also show that as the system approaches the theoretical critical regime
by decreasing the external input, there is a change in the distributions of avalanche characteristics with
the appearance of partial scale-free behaviour in avalanche size. It is important to note that the definition
5
of avalanches strongly depends on the choice of binning method. In the literature different definitions of
avalanches are used in models with seeded systems and with systems where the dynamics are continuous
(including physiological recordings). We will return to this in the discussion.
Unlike in the companion paper where the system was seeded after each avalanche, the system studied
here does not have an explicit separation of timescales. This allows us to additionally assess the waiting
times, which are intrinsic to the system, and we are able to analytically derive the distribution of waiting
times. We then investigate the presence of partial LRTCs in the empirically derived waiting times. Finally,
we show that as the system size increases (and the system approaches the theoretical critical regime from a
different route) the range over which the correlations extend also increases. Overall we find that the system
displays different signatures of criticality depending on the region of the parameter space around the critical
regime.
The model
As in the companion paper, we study a stochastic model based on that of Benayoun et al. [36]. Though
greatly simplified from a physiological neural network, the model is chosen as it is analytically tractable and
thus enables direct derivation of the parameters such that there is a critical regime. With this approach it
is therefore possible to assess the dynamics of a neuronal system in the region of (or at) a critical regime.
While Benayoun et al. considered a network with both excitatory and inhibitory connections, we simplify
the system further (as in the companion paper), considering a network with purely excitatory synaptic
connections. As will be discussed later, this type of network can be set within the context of early brain
development.
We consider a system of N fully connected neurons, with each neuron in one of two states - active (A)
or quiescent (Q). For a small time step dt → 0 the probability of transition for a neuron between the two
states is given by:
where si(t) =(cid:80)
j
P (Q → A, in time dt) = f (si(t))dt,
P (A → Q, in time dt) = αdt,
wij
N aj(t) + hi(t) is the input to neuron i, f is an activation function, hi(t) is the external
input to neuron i, wij is the connection strength from neuron i to neuron j and aj(t) = 1 if neuron j is
active at time t and zero otherwise. Finally, α is a constant rate at which neurons change from the active to
inactive (quiescent) state.
6
For analytical tractability and characterisation of the critical state we make the following additional
simplifications:
1. The synaptic connection strengths are the same for all connections with wij = w > 0.
2. The external input is constant to all neurons and at all simulation steps so that hi(t) = h > 0.
3. The activation function is linear with f (x) = x.
While the first and third assumptions are the same as in the companion paper, we make the additional
assumption of constant positive external input here as opposed to the companion paper where we examined
the system with no external input (h = 0). As the network is fully connected, and the system is closed so
that A + Q = N (where A is the number of active neurons and Q is the number of quiescent neurons), the
system can be described by the mean field equation:
(cid:18) wA
(cid:19)
dA
dt
=
+ h
(N − A) − αA.
N
As stated in the companion paper, we can use this equation to analyse the stability of the system about the
fixed point and determine the parameters for which the system is at the threshold of stability, i.e. when the
fixed point is critical. This threshold occurs when the eigenvalue (λ) of the fixed point is zero, which can
alternatively be stated, borrowing terms from the epidemiology literature, as R0 = 1 (the basic reproductive
ratio). Moreover, this is also equivalent to a branching parameter of one. In the companion paper it was
shown that with h = 0, R0 = w
α and so for R0 = w
α = 1 ⇒ α = w the system is critical.
Here we study the system in the presence of a positive external input, h > 0. In this case the fixed point
of the system is given by:
w
N
−
A2 + wA + h(N − A) − αA = 0,
and the eigenvalue at the fixed point is:
λ = −2
w
N
A + w − h − α.
For a fixed point to be critical we require that both these equations be satisfied. However, solving them
simultaneously we find that there are no real roots when w, h, N > 0. This implies that there is no
parameter region such that the system (with this activation function and positive external input) has a
critical fixed point. However, considering again the case with no external input (h = 0) for which the critical
state occurred with parameters α = w, if this system is driven by a 'sufficiently low' level of external input
7
it should still be within the region of the critical state. There has been some suggestion that the brain is
not directly at a critical point but is in fact just very close to the critical regime and it has been speculated
that the brain may actually be slightly supercritical [32]. Additionally, it been shown that a computational
model of neuronal avalanches which follows a SOC approach [25] is actually a system that 'hovers' close to
the critical state [23]. Therefore, the question of how a finite driven system behaves in the region of a critical
regime is pertinent to the neuroscience field.
An additional motivation for considering a non-zero external input is the dynamic range (∆) of the
system. Larremore and colleagues [29] describe the dynamic range as "the range of stimuli over which
there is significant variation in the collective response of the network". Kinouchi and Copelli [16] examined
dynamic range in models of networks with uniform connectivity operating with discrete time dynamics where
multiple firings can occur within each time step. They found that the dynamic range was maximised when
the local branching ratio was equal to one. Larremore and colleagues [29] considered a version of this model
but with the introduction of heterogeneity in connections, showing that it is the lead eigenvalue, λ, of the
connectivity matrix that governs the dynamic range and that the dynamic range is maximised when λ = 1.
In Appendix 1 we provide an analytic calculation for the dynamic range of our continuous model. Analogous
to the results described above [16, 29] the dynamic range is maximised when R0 = 1 (w/α = 1 when h = 0).
This is illustrated in Fig. 1 where results from simulations are compared with the analytic solution. It is
important to emphasise that the parameterisation of the dynamic range is in terms of the value R0 calculated
for networks when there is no external input. When this parameterisation is such that R0 = 1 ⇒ α = w
(and therefore when the system is tuned to the critical state) external input to the system will give rise to
dynamics for which the dynamic range is maximised. This point will be considered further in the discussion.
Throughout this study we will examine the system in the presence of an external input of h = 1/N
or less. This level of the external input is equivalent to setting a single neuron to the active state and so
corresponds to seeding the system in the zero input case. We therefore deem this level of external input to
be sufficiently low such that we could expect the system to remain within the region of the critical regime.
As in the companion paper we set w = α = 1. With these parameters and with positive external input we
find that the fixed point of the system is given by
A = −
hN
2 ±
N 2h2
4
+ N 2h,
(cid:114)
8
Figure 1: R0 versus ∆. Plot of R0 versus ∆ (the dynamic range - see Appendix 1) for both the analytic result
(black line) and simulations (red dots). For the comparable simulated result we average 10, 000 realisations
that have run until time t = 200. To obtain a reasonable spread of h we used the conjugation of the intervals
[0 : 0.002 : 0.2] and [0.4 : 0.2 : 18].
and the eigenvalue of this fixed point is given by
λ = −
(cid:112)
h2 + 4h.
With lower levels of external input the system approaches the critical regime (see Fig. 2). Note that this
approach is in fact from a slightly subcritical state given these values of α and w and with positive external
input. Under these conditions it is not possible to consider an approach from a supercritical regime with a
positive eigenvalue.
As described above, we (initially) set h = 1/N . With this level of external input:
and the eigenvalue of the fixed point is given by
A = −
+ N ,
1
2 ±
(cid:114) 1
(cid:114) 1
4
λ = −
N 2 +
4
N
.
As N → ∞, λ → 0 (see Fig. 2). Thus, for this level of the external input (h = 1/N ), as the system size
(N ) increases the system approaches the critical state (as the system reaches the critical state exactly when
the eigenvalue λ = 0). We will examine the effect on the dynamics of decreasing the external input, thereby
9
00.20.40.60.811.21.41.61.820100200300400500600700800R0∆ Figure 2: The eigenvalue of the system compared with the level of external input and system
size. (a) With w = α the eigenvalue decreases with lower levels of external input h, with the system reaching
the critical regime in the absence of external input (λ = 0 i.e. R0 = 1, the case studied in the companion
paper). (b) With h = 1/N and w = α the eigenvalue of the fixed point λ → 0 as N → ∞. Thus, the system
approaches the critical state as the system size increases.
allowing the system to approach the critical regime. We will also investigate an alternative route to the
critical regime by increasing system size at constant (overall) level of external input.
Model simulations and burst analysis
As in the companion paper and in Benayoun et al. [36], simulations of the network dynamics were carried
out using the Gillespie algorithm for stochastic simulations [37]. Briefly, at each step in the simulation
• The total transition rate r for all the neurons within the network is calculated, with r = raq + rqa
where raq is the total rate of active → quiescent transitions and is given by raq = αA and rqa is the
total rate of all quiescent → active transitions which is given by rqa = f (si)(N − A).
• The time to the next transition dt is selected at random from an exponential distribution of rate r.
• The type of transition is selected by generating a random number n ∈ [0, 1]. If n < raq
then a randomly
r
chosen active neuron becomes quiescent, otherwise a (randomly chosen) quiescent neuron switches to
the active state.
At each step in the simulation a single neuron makes a transition, though the rate at which transitions occur
changes and so the simulation step changes. If the network is in a fully quiescent state (Q = N ) then, with
positive external input, raq = 0 but rqa = hN and consequently there will necessarily be a transition of a
10
-0.07-0.04-0.01hλ(a)00.58001800050000100000150000-0.07-0.05-0.03-0.01Nλ(b) neuron from the quiescent to the active state. Similarly, when the network is in the fully active state (A = N )
rqa = 0 but raq = αN and so there will necessarily be a transition of a randomly chosen neuron from the
active to the quiescent state. From all other starting points transitions from the active to the quiescent or
from the quiescent to the active state are possible. Thus, from all network states one neuron will change
state. This is unlike the companion paper where with no external input the network must be seeded when in
the fully quiescent state. Instead in this case network dynamics are continuous (i.e. no re-seeding is required)
and are of finite length only in-so-far as they are restricted by simulation lengths.
We define a neuron as firing at the first time step at which the neuron switches from the quiescent to
the active state. Fig. 3 shows raster plots of network firings for the first 1 second of simulations with three
different levels of the external input. As was described above, unlike in the companion paper where there was
no external input, the dynamics continue even if the system reaches the fully quiescent state. Interestingly,
we can also notice that the network dynamics appear to exhibit burst-like behaviour, with periods of high
neuronal firing interspersed with periods without network firing. It is important to realise that these bursts
are intrinsic to the system and are not directly related to the dynamics of the external input (the input
is constant to all neurons in each of the simulations) nor due to a saturation of the network - the bursts
themselves consist of different numbers of neuronal firing. In all three cases the parameters are set to the
critical state (with no external input). With lower levels of the external input the system approaches the
critical regime and we see that the bursts become further apart and more distinct. We will characterise
these dynamics below. (See also Appendix 2 where we examine driving the system from subcritical and
supercritical states.)
These burst dynamics are analogous to the neuronal avalanches observed experimentally in that they are
discrete cascades of firing. Neuronal avalanches observed experimentally in physiological networks are so
called because they have sizes which are distributed according to a power-law and while the size distribution
of the burst activity in this network has yet to be presented we will refer to the activity throughout the rest of
this paper as avalanches due to their discrete burst behaviour. To determine the distribution of the avalanches
we divided the activity into individual avalanches using the approach of Benayoun et al. [36]. This method
divides consecutive neuronal spiking between any two neurons within the network into separate avalanches if
the time difference between the spikes is greater than the average difference (δt) between consecutive spikes
within the simulation. This approach (referred to later in the text as the binning method) is similar to the
method used to define neuronal avalanches within physiological data [5, 6] - though the choice of binning
method will be discussed later in the paper. It is important to note that no binning approach was needed in
11
Figure 3: Raster plots of the network dynamics for different levels of the external input. Neuronal
firing across 1 second of a simulation with an external input of (a) h = 1/N , (b) h = 0.1/N and (c)
h = 0.01/N . For all three simulations N = 800, α = w = 1 and the red line indicates the rate of firing in
1 ms bins. As can be observed, as the level of the external input decreases the firing rate decreases and the
time between avalanches increases.
the companion paper since an avalanche was naturally defined as all firing that occurred before the network
reached the fully quiescent state and was reseeded. This has been used as a standard classification for
discontinuous data, stemming from the sandpile model of criticality [18]. However, as the firing dynamics
here continue for the entire simulation it was instead appropriate to use an approach that had been used
previously for continuous dynamics.
Throughout the remainder of this paper we examine characteristics of these avalanches: namely the size
and duration of avalanches as well as the inter-avalanche intervals (IAIs). The size of an avalanche is defined
(in the standard way) as the number of firings within the avalanche. If a single neuron fires more than once
within a single avalanche it is also counted more than once. The duration of an avalanche is defined as the
time between the start of the avalanche (the first neuron firing) and the end of the avalanche. Note that if
the avalanche consists of a single neuron firing then the duration of the avalanche is 0 (and the size of the
avalanche is 1). Similarly, an IAI is defined as the time between the end of one avalanche and the start of
the next avalanche, i.e. the waiting time between avalanches. Note that the minimum IAI is bounded below
by δt as a separation between two consecutive spikes of greater than δt defines separate avalanches.
12
020040060080010000400800Time (ms)Neuron Index(a)020040060080010000400800Time (ms)Neuron Index(b)20040060080010000400800Time (ms)Neuron Index(c) Figure 4: Distributions of avalanche size and duration for varying levels of the external input.
The distributions of avalanche size (a,b,c) and duration (d,e,f) from simulations with h = 1/N (a,d), h =
0.1/N (b,e) and h = 0.01/N (c,f). As the level of the external input is decreased the system approaches
the critical regime. For all simulations N = 800, α = w = 1 and the distributions are pooled from 10
simulations each of length 104 seconds. The red lines indicates linear fits on the double logarithmic scale
(where appropriate), i.e. fitted power-laws with exponents of (b) 1.68, (c) 1.48, (e) 1.88 and (f) 1.64.
Distributions of avalanche size and duration
Fig. 4 shows the distributions of avalanche size and duration from example simulations for the three different
levels of external input investigated -- this can be compared with Fig. 3 of the companion paper [24] which
shows the avalanche size distribution in the absence of external input. With lower levels of external input
the system approaches the critical regime and the distributions of avalanche size appear scale-free across a
range of scales. The distribution approaches the distribution found in the companion paper for the system
exactly tuned to the critical state. With h = 0.01/N (i.e. the lowest level of eternal input) the exponent
of the fitted power-law is approximately 1.5, see Fig. 4(c), which is consistent with experimentally observed
neuronal avalanche sizes [5, 6]. However, for higher levels of the external input this scale-freeness of the
distribution is lost which coincides with moving away from the critical regime.
In the case of avalanche
duration a similar relationship with the critical regime is seen with a scale-free portion in the middle ranges
of the distribution (between approximately 2 and 50 ms) with lower levels of external input. Thus, for lower
levels of external input, when the system approaches the critical regime, the distributions, in particular that
of avalanche size, exhibit partial scale-free behaviour.
13
Avalanche sizeProbability10−710−410−1110100500(a)11010010000Avalanche sizeProbability10−510−310−111010010000(b)Avalanche sizeProbability10−510−310−111010010000(c)0.10.52.05.0Avalanche durationProbability10−710−410−1(d)Avalanche durationProbability10−710−410−10.1110100(e)Avalanche durationProbability10−510−310−10.1110100(f) It is worth considering what leads to the changes seen in the distributions as the level of external input
is varied. As stated, as the level of external input decreases, the system approaches the critical regime
and so it is perhaps not surprising that signatures of criticality (i.e. scale-free behaviour) emerge in the
distribution of avalanche size. Examining the raster plots of firing for the different levels of external input,
see Fig. 3, we see that at lower levels the avalanches are further apart and more distinct. While the external
input itself is continuous, at lower levels of external input there is a separation of timescales, where one
avalanche always finishes well before the next avalanche begins. The distribution therefore appears to
follow similar characteristics to a system with a built-in separation of timescales and we confirm that the
distribution is similar to that found in the companion paper (in which the model had an explicit separation
of timescales) where an exponent close to 1.5 was also observed for the distribution of avalanche size. As
the external input is increased there are no longer such distinct periods between avalanches. This leads to
a superposition effect, with the next (actual) avalanche starting before the previous avalanche has finished
(i.e. a new network cascade is initiated before the previous one has finished). This leads to these 'avalanches'
being defined using the binning approach as a single avalanche (see discussion). The scale-free behaviour in
the distributions of avalanche size and duration is therefore lost.
Theoretical derivation of the distribution of the IAIs and comparison with simulated data
The temporal patterning of activity within networks of neurons has long been investigated as a property of
key importance, with rate and temporal coding suggested as potential substrates for information propagation.
As it remains to be fully determined how different neuronal firing properties may lead to information transfer
we suggest that in addition to the distribution of avalanches sizes the intervals between avalanches need to
be considered as a functional entity in their own right. In this section we derive the theoretical distribution
of IAIs and compare it with results from simulations.
We begin by noting that a single IAI is a period during which there is no neuronal firing, i.e. neurons can
only be switching from the active to the quiescent states or an IAI may be a period with a single quiescent to
active transition which is preceded by another quiescent to active transition. Let us initially ignore the fact
that there is a minimum duration (δt) of an IAI and first consider the distribution of all consecutive active
to quiescent transitions (we will return to the distribution of single quiescent to active transitions later).
14
Distribution of consecutive active to quiescent transitions
Let N0 be the number of active neurons at a time point in the simulation. After a single simulation step
the number of active neurons will be N0 + 1 or N0 − 1, as at every simulation step only one neuron makes a
transition. Let qi be the probability that an active neuron goes back to the quiescent state given that there
are i active neurons. Note that from the transition rates:
qi =
αi
( w
N i + h)(N − i) + αi
=
αiN
(wi + hN )(N − i) + αiN
.
Starting with N0 active neurons the probability that there are N0− 1 active neurons after a single simulation
step is qN0 and the probability that there are N0 + 1 active neurons is 1 − qN0. Given these probabilities
we can construct a probability tree, shown in Fig. 5, particularly concentrating on the portion of the tree
corresponding to active to quiescent transitions, i.e. those transitions that form a period of consecutive active
to quiescent transitions. (Note that this probability tree focuses on different aspects of the model to that of
the probability tree in the companion paper.)
Figure 5: Probability tree of consecutive active to quiescent transitions. Starting from a state with
N0 active neurons the probability tree diagram indicates the possible transitions from each state specifically
concentrating on the active to quiescent transitions. The probability qi of each transition is as indicated in
the main text and is dependent on the number of active neurons, i.
From this tree approach we can calculate that the probability of exactly k consecutive active to quiescent
transitions (note that to consist of exactly k active to quiescent transitions the transition sequence must be
ended by a quiescent to active transition):
P (IAIk) = p(N0, k) = (1 − qN0−k)
15
k−1(cid:89)
j=0
qN0−j.
(1)
N0 @@@@@R1 qN0qN0N0+1N0 1 @@@@@R1 qN0 1qN0 1N0N0 21 @@@@@R1 q1q120 111 The duration of this period of consecutive active to quiescent transitions is given by the sum of the times
for each of these k transitions (plus the time for the quiescent to active transition). As the Gillespie algorithm
is used for simulations, at each simulation step the time to the next transition is drawn randomly from an
exponential distribution with rate r (see above), where r is dependent on the number of active neurons and so
changes at each simulation step. The duration of consecutive active to quiescent transitions is therefore the
sum of exponentially distributed variables drawn from distributions of different rates, i.e. the distribution of
consecutive active to quiescent transitions is a hypoexponential. Thus, the duration distribution, f (x, N0, k),
of consecutive active to quiescent transitions of length x, consisting of k transitions, ending with an additional
quiescent to active transition and starting from N0 active neurons is [38]:
f (x, N0, k) =
rN0−je−rN0−j x
k(cid:88)
j=0
k(cid:89)
i=0,i(cid:54)=j
rN0−i
rN0−i − rN0−j
,
(2)
when rN0−i (cid:54)= rN0−j and where rm is the total transition rate for all neurons within a network with m
active neurons and is the rate of the exponential distribution from which the time to the next transition is
randomly drawn. This equation holds provided that rN0−i (cid:54)= rN0−j ∀i, j. If this is not the case and there
exists A, B ∈ {1, . . . N} such that (α + w − h)/w = A + B ⇒ rA = rB then we use the more general form --
assuming there are a distinct rates, which we label β1, β2, . . . , βa that occur c1, c2, . . . , ca times respectively
(i.e. c1 + c2 + . . . ca = k + 1) -- given by [39]:
f (x, N0, k) = B
a(cid:88)
ck(cid:88)
k=1
l=1
where
a(cid:89)
j=1
B =
βcj
j and φk,l(t) =
φk,l(−βk)xck−le−βkx
(ck − l)!(l − 1)!
,
a(cid:89)
j=1,j(cid:54)=k
dt−1
dtt−1
(βj + t)−cj .
(3)
Whilst this involves higher order derivatives a closed-form solution is provided by Amari and Misra [40].
From equation 1 we know the probability of k consecutive active to quiescent transitions. This equation
holds true for any k up to k = N0, which is the maximum number of consecutive active to quiescent
transitions as the fully quiescent state is then reached. Therefore, the distribution, F (x, N0), of consecutive
active to quiescent transitions of duration x starting with N0 active neurons but consisting of any number
of transitions is a weighted sum of hypoexponentials:
F (x, N0) =
f (x, N0, k)p(N0, k).
(4)
N0(cid:88)
k=1
16
Probability distribution of the initial number of active neurons
Finally, to calculate the full probability distribution of consecutive active to quiescent transitions for a
network of set system size, N , we must combine equation 4 with the probability of the initial number of
active neurons being equal to N0 ∈ {1, 2, ...N} (note that N0 = 0 is not considered as the next transition
will necessarily be an activation). To determine this probability, let us first consider the simple case of
N = 3. We assume that the simulation starts from a state with no active neurons. Fig. 6 shows all possible
transitions between the number of active neurons in a network of this size.
Figure 6: Probability tree of all possible transitions in a network of size N = 3. Simulations start
from a state with no active neurons: N0 = 0. The diagram shows the possible transitions at each step,
along with the probability of making that transition. The probabilities are as defined in the main text with
qi being the probability of a neuron switching from the active to the quiescent state given i initially active
neurons. Dotted lines indicate transitions that are already shown elsewhere in the tree and so the tree shown
here completely describes all possible transitions in a network of this size.
From this probability tree the probabilities, P (i), of the number of active neurons being equal to i, where
i ∈ {0, 1, 2, 3} are given by:
P (0) = q1P (1),
P (1) = P (0) + q2P (2),
P (2) = (1 − q1)P (1) + P (3),
P (3) = (1 − q2)P (2),
17
(5)
0❄11\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}✠❅❅❅❅❅❘q11−q102\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}\x{FFFF}✠❅❅❅❅❅❘q21−q213❄121 (assuming a steady state has been reached such that the probabilities are time independent). Rearranging
and substituting to write the equations in terms of P (1):
P (0) = q1P (1),
(1 − q1)
P (2) =
q2
P (1),
P (3) =
(1 − q1)(1 − q2)
q2
P (1).
Furthermore, the sum of all the probabilities must equal 1 and so
(cid:18)
q1 + 1 +
(1 − q1)
q2
+
(1 − q1)(1 − q2)
q2
(cid:19)
P (1) = 1.
Therefore,
P (1) =
q2
q1q2 + q2 + (1 − q1) + (1 − q1)(1 − q2)
.
(6)
(7)
(8)
By substituting this value back into the set of equations 6 the probabilities for the full system can be
calculated.
Generalisation to a system of any size N
From considering this simple example we can generalise to derive the probabilities of the number of active
neurons for a system of any size N . Firstly, as in equation 4 the probabilities can be written as (again
assuming a steady state):
P (0) = q1P (1),
P (1) = P (0) + q2P (2),
P (2) = (1 − q1)P (1) + q3P (3),
...
P (k) = (1 − qk−1)P (k − 1) + qk+1P (k + 1),
...
P (N − 1) = (1 − qN−2)P (N − 2) + qN P (N ),
P (N ) = (1 − qN−1)P (N − 1),
(9)
where qN = 1 but it will remain in the equations so as to aid notation. Rearranging gives
P (2) =
(1 − q1)
q2
P (1),
18
and by induction:
P (k + 1) =
=
1
(P (k) − (1 − qk−1)P (k − 1))
qk+1
(1 − q1)(1 − q2) . . . (1 − qk)
q2q3 . . . qk+1
P (1).
(10)
Summing all the probabilities and setting this equal to 1:
P (1) =
q1q2 . . . qN + q2q3 . . . qN + (1 − q1)q3 . . . qN + . . . + (1 − q1)(1 − q2) . . . (1 − qN )
q2q3 . . . qN
.
(11)
Having determined these probabilities we then need to take into account the fact that the consecutive
active to quiescent transitions must be preceded by a quiescent to active transition, i.e. they must be preceded
by a neuron firing (otherwise they would be a chain of k + 1 consecutive active to quiescent transitions and
so included elsewhere in the distribution). We are therefore only interested in the probability PA(N0) of the
number of active neurons being equal to N0 given that a quiescent to active transition has just occurred.
Considering again the probability tree, Fig 6, we find that these probabilities, are given by:
PA(0) = 0,
PA(1) = P (0),
...
PA(k) = (1 − qk−1)P (k − 1),
...
PA(N ) = (1 − qN−1)P (N − 1),
(12)
where we make use of the previously defined probabilities P (k). From these probabilities the full probability
distribution of the duration of consecutive active to quiescent transitions can be calculated. As was shown
above, for a set initial number of active neurons N0, the probability distribution of consecutive active to
quiescent transitions is given by a weighted sum of hypoexponentials, see equation 4. This can then be
further weighted by the probability PA(N0) that the initial (at the start of the sequence of transitions)
number of active neurons is equal to N0 and the previous transition was quiescent to active. Thus, the
℘(x) =
PA(i)
f (x, i, m)p(i, m)
.
(13)
To confirm that this theoretically derived distribution compares with results from simulations, we de-
termined the distribution of the lengths of periods of any consecutive active to quiescent transitions from
i=0
m=1
19
overall probability distribution of consecutive active to quiescent transitions is given by:
(cid:32)
N(cid:88)
i(cid:88)
(cid:33)
Figure 7: Theoretical and simulated distributions of periods of consecutive active to quiescent
transitions. The simulated distribution (black) is compared with the theoretically derived probability
density function (shown in red, see equation 13). For both distributions α = 1, w = 1, N = 50 and h = 1/N .
The mean difference between consecutive spikes (δt) within the simulation (green) is used to define avalanches
through the binning approach described in the main text. Thus, the portion of the distribution for which
the length of active to quiescent transitions are greater than this average time between consecutive spikes
form the distribution of IAIs when combined with the distribution of single quiescent to active transitions.
simulations. Fig. 7 shows the good agreement between the distribution of consecutive active to quiescent
transitions from a simulation with N = 50 and the theoretical distribution.
Distribution of single quiescent to active transitions
As was described above, a single period in between two neurons firing can also be an IAI, provided that the
duration of this IAI is longer than the average time between spikes as accounted for below. Note, however,
that only single periods are considered, as consecutive periods necessarily include neurons switching to the
active state and therefore cannot form part of an IAI. Thus, the distribution of IAIs should also take into
account the distribution of single quiescent to active transitions. As we make use of the Gillespie algorithm,
the duration distribution of these single transitions is an exponential with rate given by the total transition
rate, which is dependent on the number of active neurons, N0. This is then weighted by the probability of
a quiescent to active transition given N0 active neurons (i.e. by 1 − qN0 ) and additionally weighted by the
probability of starting with N0 active neurons following a quiescent to active transition as calculated above.
Thus, the probability distribution of single quiescent to active transitions of length x is given by:
(14)
(cid:0)PA(i)(1 − qi)rie−rix(cid:1) .
ρ =
N(cid:88)
i=0
Fig. 8(a) shows the good agreement between simulated distribution of single quiescent to active transitions
and theoretical distribution.
20
00.511.5200.020.040.06PeriodProbability SimulationTheoretical PDFδ t Figure 8: Theoretical distributions of single quiescent to active transitions and the combined
distributions. Theoretical (red) and simulated (black) probability distributions for (a) single quiescent to
active transitions and (b) this distribution combined with the distribution of consecutive active to quiescent
transitions (see Fig. 7). In both cases α = 1, w = 1, N = 50 and h = 1/N . The green line indicates the
average time between consecutive spikes (δt) within the simulations. Thresholding the combined distribution
(b) at this level determines the IAI distribution.
The IAI distribution
As discussed above, the IAI distribution combines these two distributions - the distribution of consecutive
active to quiescent transitions and the distribution of single quiescent to active transitions. This combined
distribution, along with simulated values, is shown in Fig. 8(b). As was described above, avalanches are
defined from the network firing pattern as consecutive spikes where the time difference between them is
no greater than the average time difference between consecutive spikes, δt, within the network. Thus, the
minimum IAI is bounded below by δt and all consecutive active to quiescent transitions or single quiescent
to active transitions whose total duration is greater than δt will be an IAI. Thresholding the combined dis-
tribution at δt determines the IAI distribution. Fig. 9(a) shows theoretical and simulated IAI distributions
displayed on a double logarithmic scale. Despite the fact that the distribution is not a power-law (theoret-
ically we know that it is a weighted sum of hypoexponentials), it appears scale-free over a range of scales
on this double logarithmic scaling. As we will show below, the distribution can also pass statistical tests for
power-law distributions, suggesting partial scale-free behaviour of the system close to the critical regime.
Fig. 9 also shows the theoretical and simulated distributions for lower levels of external input. With
lower levels of external input (as the system approaches the critical regime) the average IAI increases and
21
00.10.20.30.40.50.60.70.800.10.2PeriodProbability(a) SimulationTheoretical PDFδ t00.20.40.60.811.21.41.61.8200.050.10.15PeriodProbability(b) SimulationTheoretical PDFδ t Figure 9: Distribution of IAIs for varying levels of the external input. The theoretical (red) and
simulated (black) IAI distributions with (a) h = 1/N , (b) h = 0.1/N and (c) h = 0.01/N . These distributions
were for N = 800 with α = w = 1, with a simulation length of 104 seconds. The distributions from the
simulated data are pooled from 10 simulations. The theoretical distributions were calculated up to the level
of active neurons which occur with a cumulative probability of 0.9 (see main text). The blue line in plot (a)
indicates a linear fit, i.e. a fitted power-law with an exponent of 2.71.
the distribution changes, no longer exhibiting scale-free behaviour. Even when considering the same scale for
all levels of the external input (IAIs in the region of 0.05-5 ms) it is only for h = 1/N that the distribution
is scale-free.
Indeed, at the lowest level of input h = 0.01/N the distribution is in fact best fit by an
exponential, in this case y = 0.028e−0.01x as seen in Fig. 10, indicating the loss of the scale-free behaviour
in the distribution. Thus, scale-free behaviour in the case of the IAI distribution does not increase with
proximity to the critical regime.
As an aside, note that due to the product in the hypoexponential (see equation 2) determination of the
probabilities for large N can become computationally intractable. For simulations larger than with N = 50
we therefore only determined the theoretical distribution up to a set level of the number of active neurons.
We set the threshold level of the number of active neurons according to the probability distribution of
starting from a particular number of active neurons (calculating the cumulative probability from zero active
neurons), and sufficiently low so that the calculations were computationally viable. However, the theoretical
distributions calculated using this threshold are still a good fit to the simulated data - Fig. 9.
Distributions of avalanche size and duration
As we have shown, the theoretical distribution of IAIs can be calculated by assessing periods of consecutive
active to quiescent transitions and single quiescent to active transitions. It is also possible to derive the
distribution of consecutive quiescent to active transitions. However, if a period of active to quiescent transi-
22
0.050.505.00IAIProbability10−710−410−1(a)0.55.050.0IAIProbability10−910−510−1(b)5201001000IAIProbability10−810−510−2(c) Figure 10: The IAI distribution with h = 0.01/N is well fitted by an exponential distribution. The
theoretical IAI distribution at the lowest level of the external input (shown in red, see also Fig. 9c) compared
with the fitted exponential distribution (black dashed). The exponential is given by y = 0.028e−0.01x. This
indicates that as the external input is decreased and the system approaches the critical regime the IAI
distribution loses the scale-free behaviour seen at higher levels of the external input and is dominated by an
exponential.
tions (a period without firing) has a duration less than the average time difference between two spikes then
this interval does not separate an avalanche into two. Therefore, the distributions of number and length of
consecutive quiescent to active transitions does not describe the distributions of avalanche size and duration
- these distributions can also contain periods of active to quiescent transitions within two or more periods of
quiescent to active transitions. Note that a period of active to quiescent transitions having a duration less
than the average difference between consecutive spikes is not dependent on the number of active to quiescent
transitions within the interval, as the length of each transition is drawn at random from an exponential
distribution. It was therefore not possible for us to determine a theoretical distribution of avalanche size
and duration using this approach.
Statistical comparison with a power-law distribution
The influential paper by Clauset et al. [1] developed a model selection based methodology to determine
whether empirical data are likely to be power-law distributed. This method has been used to assess physio-
logical neuronal avalanches and the results have shown that the power-law hypothesis is not rejected for this
data [2]. It is therefore of interest to determine whether this is also the case for the data from the model
studied here. Briefly, this method finds the best fit to a power-law of the distribution under study. The
empirical data are then compared to distributions of the same size that are generated by randomly drawing
values to follow the best-fit power-law distribution. A p-value is calculated as the proportion of times that
the empirical data are a better fit to the power-law than the generated data (using the Kolmogorov-Smirnov
test). As per Clauset et al. [1] the hypothesis (that the data come from a power-law) is rejected if the p-
23
IAIProbability10−810−510−25201001000 value is less than 0.1. As we have observed (Figs. 4, 9), the distribution of avalanche sizes appears to exhibit
partial scale-free behaviour for low levels of external input (h = 0.1/N, 0.01/N ) and the IAI distribution
appears scale-free over a range of scales for h = 1/N . As in the companion paper [24], we fit a truncated
power-law distribution up to an avalanche size of xmax = 9
10 N in the case of avalanche size distributions.
We fit a power-law distribution without truncation to the IAI distribution. Testing the entire avalanche
size distributions (consisting of over 900,000 avalanches) yielded p = 0 indicating that the hypothesis that
the distribution follows a power-law should be rejected. Similarly, taking the IAI distribution for h = 1/N ,
testing the whole distribution of over 6,000,000 IAIs (note that there are more avalanches and therefore IAIs
with larger h due to the higher firing rate) yielded p = 0. Testing instead the first 100,000 avalanches (a
similar order of magnitude to the number of neuronal avalanches tested experimentally) with h = 0.1/N
yielded p = 0.46 indicating instead that the power-law hypothesis should not be rejected. Similarly, for
h = 0.01/N testing the first 10,000 avalanches yielded p = 0.13. These results are similar to those of the
companion paper, where the power-law hypothesis was not rejected when the number of avalanches included
in the distribution was of the same order as those tested experimentally, and are indicative of the partial
scale-free behaviour of the system in proximity of the critical regime.
In the case of the IAI distribution testing the first 100,000 IAIs yielded p = 0.44 indicating that a power-
law is a good fit to the data. Given that in this case we know that the IAI distribution is not a power-law
(but is in fact a weighted sum of hypoexponentials), it is interesting to note that the hypothesis that the data
follow a power-law is not rejected when the number of data points is of the same order as that which have
been tested experimentally, an observation that will be explained in the discussion. When the power-law
hypothesis is not rejected, Clauset et al. [1] employ a model selection process to determine the best model
for the data. We did not carry out this testing here (as, at least in the case of the IAI distribution, we
already know what the distribution is) and it may be that such a process would suggest that a power-law is
not the best fit to the data. However, the results here (and those of the companion paper) are indicative of
the partial scale-free behaviour exhibited by the system in the region of the critical regime.
Long-range temporal correlations
As mentioned in the introduction, long-range temporal correlations are another possible signature of a system
at (or near) a critical state and have also been observed in neurophysiological data [4, 11 -- 15]. It is therefore
of interest to determine whether this finite-size neuronal system with external input displays LRTCs - given
that it is in the region of a critical regime - and whether LRTCs relate to other signatures of criticality, i.e.
24
the presence of partial scale-free behaviour in the data distributions themselves. The latter is of particular
interest given that we have seen a change in distributions as the system approaches the critical regime. As
with all simulations the level of external input is constant, i.e., it does not itself display LRTCs, and therefore
we stress at the outset that any LRTCs present in the dynamics of the system would have to be intrinsic to
the system. Furthermore, it is useful to remember that a power-law distribution of any data set does not
imply that the data will exhibit LRTCs and vice-versa -- consider points drawn at random from a power-law
distribution - such a data set would not exhibit LRTCs.
In neurophysiological data, LRTCs have been observed in fluctuations of oscillation amplitude (i.e. within
continuous data) [4,11 -- 15] and also in discrete burst activity in our recent analysis of the inter-event intervals
of bursts of nested oscillations in EEG recordings of extremely preterm human neonates [35]. Moreover,
LRTCs in discrete data have previously been investigated by Peng et al. [41] and a number of other authors,
for example [42 -- 44], in their analysis of inter-heartbeat intervals. As the data from the model analysed
here result in discrete avalanche activity, we adopt the approach used in these previous studies of LRTCs in
discrete data and examine LRTCs in waiting times, i.e.
in IAIs.
We assessed the presence of LRTCs in IAIs through estimating the Hurst exponent, H, which describes
the degree of self-similarity within the data. A Hurst exponent of H ≈ 0.5 indicates that there are no
correlations in the data or short-range correlations only, for example a white noise process, whereas a Hurst
exponent of 0.5 < H < 1.0 indicates LRTCs in the data. Additionally, an exponent of 1 corresponds to 1/f
noise [41]. We estimated the Hurst exponent using detrended fluctuation analysis (DFA) -- an approach that
has been shown to produce more accurate estimates of the Hurst exponent than some other approaches [45]
and has been used previously to assess the presence of LRTCs in neurophysiological data sets [4, 11, 15, 35].
DFA is a graphical method whereby the average root mean square fluctuations at a given box size are
compared across different box sizes and the gradient of the line of best-fit is the estimate of the Hurst
exponent (for more detail see Peng et al. [41, 46]). We used a minimum box size of 5, with 50 box sizes
linearly spaced on a logarithmic scale up to a maximum box size of 1/10 of the length of the IAI sequence [47].
Calculations were carried out using the MATLAB code of McSharry [48].
Fig. 11 shows example DFA plots for IAIs from three simulations with α = w = 1, h = 1/N .
It is
important to notice from these plots that there is not a single linear trend across all box sizes. Hu et al. [47]
discussed the importance of identifying crossover points - box sizes at which there is a change in the linear
fit of the data - within DFA plots. Failure to examine these trends leads to erroneous estimates of the Hurst
exponents. A single linear fit across all the points would give an estimate of the Hurst exponent for that
25
Figure 11: DFA plot examining the presence of temporal correlations in IAIs. Plot of the average
fluctuations F (n) against box size n for IAIs from simulations with α = w = 1, h = 1/N and (a) N = 800,
(b) N = 3200 and (c) N = 172800. The data is best fit not by a single linear trend but by three lines (red,
green, blue) between two crossover points (dashed black lines). For smaller box sizes the Hurst exponents
(slope of the line - as annotated next to the individual lines: the DFA exponent α) indicates that correlations
extend across these regions. However, for larger box sizes the exponents are closer to 0.5 suggesting that the
correlations do not extend across these larger box sizes. With a larger system size (c) the upper crossover
point increases to larger box sizes. With increasing system size the system approaches the critical regime -
(a) λ = −0.07, (b) λ = −0.04 and (c) λ = −0.005.
sequence. However, crossover points indicate that the same correlations (i.e. temporal behaviour) do not
extend across the whole sequence. In the DFA plots here there are in fact three regions, each with a different
linear trend, between two crossover points. The best-fit to the data by three linear regions was found using
the nonlinear regression function 'nlinfit' in the MATLAB environment, therefore determining the crossover
points. In Fig. 11(a), the Hurst exponent (slope of the line) of the first two regions (at smaller box sizes)
are 0.83 and 0.62 respectively - exponents which indicate the presence of LRTCs within the data. However,
the third region across the largest box sizes has an exponent of 0.51 indicating that there are no correlations
in the data. This change in the exponents therefore suggests that the correlations observed in the data at
small box sizes do not extend across the entire sequence length.
When examining the presence of LRTCs it is standard practice to compare the exponent of the actual data
to the exponent of the data randomly shuffled [4]. Shuffling the data should destroy any correlations present
and therefore the exponent of the shuffled data is expected to be approximately 0.5. We compared the original
sequence (whose DFA plot is shown in Fig. 11) with 500 shuffled sequences. As expected, the DFA plots for
the shuffled sequences (data not shown) did not exhibit crossover points and showed a mean exponent for
the shuffled sequences of 0.50 with a range of 0.48-0.52. The fact that the exponents of the original sequence
(at smaller box sizes) do not fall within the distribution of exponents for the shuffled sequences therefore
26
12345−2−10123log10(n)log10(F(n))(cid:95)=0.83(cid:95)=0.62(cid:95)=0.51(a)12345−2−10123log10(n)log10(F(n))(cid:95)=0.75(cid:95)=0.89(cid:95)=0.56(b)12345−2−10123log10(n)log10(F(n))(cid:95)=0.72(cid:95)=0.95(cid:95)=0.56(c) demonstrates that the original sequence exhibits complex temporal ordering with correlations that extend
across a range of box sizes (up to the upper crossover).
Increasing the system size
As noted previously (see Fig. 2), as N → ∞ the eigenvalue of the system λ → 0, i.e., the system approaches
the critical regime with increasing system size. We might expect that as the system approaches the critical
regime it is more likely to exhibit signatures of criticality and therefore that LRTCs would extend to larger
box sizes as the system size is increased. We therefore investigated whether there was a change in the temporal
correlations of the IAIs with system size, while maintaining all other parameters including h = 1/N . For all
system sizes investigated the DFA plots displayed three regions with different linear trends, as was discussed
above. Fig. 11 shows example DFA plots for the IAIs of three simulations with the smallest and largest
system sizes examined. It was observed that the pattern of the exponents in each of the cases remained
the same - with the two lower regions having exponents indicative of LRTCs, while the exponent across the
largest box sizes is closer to 0.5. Additionally, we found that the location of the upper crossover increased
with system size. Fig. 12 shows the location of the upper crossover (the crossover at the higher box regions)
for different system sizes (from N=3200 to N=172800) normalised with respect to the largest box size. We
did not find a change, other than small fluctuations, in the exponents themselves for any of the three regions
for all system sizes. Namely, across all system sizes considered, the mean exponent across the smallest box
sizes (up to the first crossover point) was 0.73 with a range of 0.70 - 0.76.
It was 0.95 with a range of
0.89 - 0.99 (close to an exponent of 1 which would indicate 1/f noise) between the first and second crossover
points. The largest variation in exponents was for the region above the upper crossover point with an average
exponent of 0.59 and a range of 0.46 - 0.73; on average this indicates that temporal correlations do not extend
beyond the upper crossover. Thus, overall, as the system size is increased the temporal correlations extend
across larger box sizes. This is consistent with the idea that, when the system reaches the critical regime,
LRTCs could extend to infinite length (i.e., all possible box sizes) in the limit of system size.
Next we considered whether the distributions of IAIs and avalanche size themselves changed with system
size and whether the change in the correlation length observed above was reflected in a change in the
distributions. Fig. 12 also shows the IAI and avalanche size distributions for different network sizes.
In
both cases, the distributions for different system sizes show only small changes which can be accounted for
by noise. Thus, as the system approaches the critical regime through increasing the system size there does
not appear to be a change in the distributions despite the change in the temporal correlations. Moreover,
27
Figure 12: Changes with increasing system size.
(a) The normalised (with respect to the largest
box size) upper crossover box size increases with system size. The plot shows the average value across 10
simulations at the same system size and error bars indicate the standard deviation. (b) The IAI distributions
and (c) the distributions of avalanche size for 6 different system sizes, scaled with respect to the mean IAI
or avalanche size of each distribution respectively. For all simulations α = w = 1, h = 1/N .
LRTCs are present in the data but the distribution of avalanche sizes does not exhibit scale-free behaviour,
i.e., these markers of criticality do not occur simultaneously in this case. By contrast, through approaching
the critical regime by lowering the external input we have shown that the avalanches are more distinct and
the distribution of avalanche sizes exhibits partial scale-free behaviour.
The effect on LRTCs of decreasing the level of external input
We also examined the DFA exponents at lower levels of external input (h = 0.1/N and h = 0.01/N ). In both
cases there were no crossover points with a single linear trend across all box sizes (data not shown). The
exponents were 0.50 (range 0.49-0.51, across 10 simulations with N = 800) and 0.56 (range 0.55-0.57) for
h = 0.01/N and h = 0.1/N respectively. Thus, at the lowest level of external input the IAIs do not exhibit
LRTCs and there is a slight increase in the exponent as the external input increases. This suggests that as
the system approaches the critical regime through a decrease in the external input the temporal correlations
are lost. Thus, the existence of LRTCs as the system approaches the critical regime is dependent on how
the critical regime is approached, namely, approaching the critical regime through increasing the system
size extends the temporal correlations whereas decreasing the external input leads to a loss of long-range
correlations. In addition, this signature of criticality is independent from the other marker we investigated
-- the presence of scale-free behaviour in the avalanche size distribution. Considering avalanche size and
duration, scale-free behaviour is present at the lowest level of external input, when LRTCs are lost. Thus,
28
00.511.52x 1050.020.040.060.080.1NNormalised Upper Crossover(a) 10010110210310−810−610−410−2100X=Scaled IAIProbability(b) N=3200N=16000N=32000N=48000N=64000N=17280010010110210−810−610−410−2100X=Scaled avalanche sizeProbability(c) N=3200N=16000N=32000N=48000N=64000N=172800 we find that markers of criticality are not only dependent on the region around the critical regime but also
may not be present for the same parameter set.
Discussion
This paper specifically examined a finite-size neuronal system without a separation of timescales between
external input and the avalanches themselves. By analytically tuning the system to be in the region of a
critical regime we were able to examine the type of dynamics displayed by such a system and to investigate
whether the dynamics display signatures of criticality. In summary, we have shown that:
1. As the system approaches the critical regime through a reduction in the external input the avalanches
become more distinct and the distribution of avalanche sizes displays scale-free behaviour.
2. With h = 1/N the IAIs exhibit temporal correlations which extend across a range of bin sizes to an
upper crossover. As the system approaches the critical regime through increasing the system size the
length of the temporal correlations is extended across a wider range of bin widths. These correlations
(one noted signature of a critical system) are observed despite the fact that the distribution of avalanche
sizes does not exhibit scale-free behaviour and does not change with the increase in system size. These
temporal correlations are lost if the critical regime is instead approached through reducing the external
input.
3. The distribution of IAIs was theoretically derived and was shown to be a weighted sum of hypoex-
ponentials. However, for h = 1/N (when the number of avalanches considered was of the same order
as those tested experimentally) the hypothesis that the IAI distribution follows a power-law was not
rejected by statistical testing indicating the scale-free nature of the distribution at this level of the
external input.
Validity of the model
The model considered in this paper was a highly simplified neuronal system with a number of assumptions,
such as equally weighted synapses and continuous constant external input. These assumptions were necessary
in order to analytically tune the system to be in the region of a critical regime. Therefore, while this should
not be taken as an accurate model of a real neuronal system it is important that we first consider models
such as this, examining markers of criticality, which will then aid our understanding when building on this
work with more complex models. This paper opens the way for future work examining the role of external
29
input on signatures of criticality and the importance of the region of parameter space on network dynamics.
Future work should also investigate the effect of topology on the dynamics [29, 49] and the effect of external
input with different temporal and spatial characteristics.
A purely excitatory network
The synaptic connections investigated in this model were purely excitatory. This not only simplifies the
model for analytical investigations but is also of interest from a neurological perspective in terms of early
brain development. Before birth, GABA is thought to have a depolarising effect on postsynaptic neurons
and it is not until the nervous system reaches a more mature state that this neurotransmitter becomes
inhibitory [50,51]. While presynaptic inhibition is thought to be present at all developmental stages [52] this
effect can be considered to be taken into account in the model by the fact that neurons cannot re-fire until
they have returned to the quiescent state. We have recently shown that EEG recordings from very preterm
infants (when GABA is still thought to be purely excitatory) exhibit LRTCs in the temporal occurrence
of bursts of activity [35]. The model studied here may be a candidate mechanism for the generation of
this temporal patterning in the discontinuous activity of the developing brain. Moreover, it is interesting
to note that despite the fact that the system has purely excitatory postsynaptic connections and input, for
these parameter regions, the model does not exhibit runaway excitation (saturation) but is able to maintain
stable dynamics through the 'balance' of individual neuronal dynamics resulting from a trade-off between
the rates at which neurons become active and quiescent. Indeed, while a number of authors have suggested
that a balance of excitation and inhibition in neuronal networks leads to critical behaviour [53], the work
here and in the companion paper shows that excitatory networks can display the same behaviour. It can be
speculated that this type of balanced activity in the region of a critical regime might be a way in which the
brain avoids (for the most part) epileptic behaviour during early development.
The activation function
Here we used a linear activation function for the transition of neurons from the quiescent to the active states.
However, physiologically neurons behave more like a saturating function. The linear activation function used
here was chosen so as to be analytically tractable and is also equivalent to a saturating function at low input
levels. However, considering instead a saturating function (see Appendix 3) we found the dynamics in the
region of the critical regime to show similar behaviour to the system with the linear activation function.
With both the linear and saturating activation functions, the critical regime can only be reached exactly
30
in the absence of external input. A positive external input therefore drives the system away from the critical
regime. However, with a quadratic activation function (see Appendix 3) the system does have a critical
fixed point even with a positive external input and it can be tuned directly to this regime. With such an
activation function the dynamics do not appear to exhibit burst like behaviour, however, analysis shows
that the activity fluctuates about the critical regime in an 'avalanche-like' manner. Thus, while a quadratic
function does not best describe activation in a neuronal network, we can emphasise our conclusion that
signatures of criticality are not universal and can be examined only in relation to the specific critical regime
of the system (see Appendix 3).
The binning approach
As described previously, the binning method separated avalanches where the time difference between con-
secutive spikes was greater than the average time difference between consecutive spikes across the entire
simulation. This was the approach taken by Benayoun et al. [36]. However, it is worth noting that this is
a slightly different approach to the method first proposed by Beggs and Plenz [5, 6] to separate neuronal
avalanches. In their analysis neuronal firing is distributed into bins of width of the average time difference
between consecutive spikes (δt) and firing is separated into avalanches by bins in which no firing occurs.
Thus, two spikes may be separated by more than the average time difference δt but still be considered part
of the same avalanche if they fell within consecutive bins. Our theoretical derivation of the IAI distribution
relied on the fact that all consecutive active to quiescent transitions or single quiescent to active transitions
with a length greater than the average time between two spikes is an IAI. This would not be the case if
the alternative (Beggs and Plenz) binning approach was used to determine avalanches. If this alternative
approach was used the distributions of consecutive active to quiescent transitions and single quiescent to
active transitions would be the same, but transitions of length slightly greater than or equal to the average
time between consecutive spikes (in fact up to twice this average) may or may not form part of the IAI
distribution depending on the exact binning. It is also important to note that with the binning method used
here, even with dense neuronal firing (which occurs if the external input is increased from the levels studied
here), it is always possible to separate the dynamics into 'avalanches' since there is always an average time
between consecutive spikes.
Additionally, both these binning approaches differ from that used in non-driven systems such as the
classical sand-pile model [18] and the system investigated in the companion paper to this article [24]. In
those models an avalanche consists of all firings until the system returns to the fully quiescent state. This
31
means that although the system may have a long period without firing -- during which neurons switch to the
inactive state -- activity before and after this period will not be considered as separate avalanches even if the
period in-between exceeds the average difference between consecutive spikes. Future work is needed to fully
investigate how differences in the definitions of avalanche affect the distributions of size, duration and IAIs
and care needs to be taken when interpreting the results from these different approaches.
Validity of DFA and the investigation of LRTCs
DFA is one method by which to estimate the Hurst exponent and was chosen here as it has been shown to
be an accurate estimate [45]. Moreover, it is a graphical approach and so can be used to check for crossover
points [47] -- see also the recently proposed ML-DFA [54] whereby the existence of crossover behaviour can
be rigorously tested. As the Hurst exponent can only be estimated it is is considered best practice to check
the consistency of the exponents using two methods [55]. However, as non-graphical methods only give
single numerical values they cannot be interpreted when crossover behaviour exists. Given that there were
crossover points we only considered DFA with this analysis.
Crossover points within a DFA plot have been shown to exist when the same correlations do not extend
across the whole data sequence in analytically constructed data [47]. It is important to correctly interpret
these crossover points. Being box sizes, they define sequence lengths, e.g., a box size of 10 indicates detrending
across a sequence of 10 consecutive IAIs. Note that as the IAIs are of variable length the box size does not
specify a particular simulation time but merely a number of events. Future investigation is therefore needed
to determine the relationship between the model and its crossover points.
Correlations extended only across a range of box sizes, with this range extending as the system size
increased and the system approached the critical regime. It appears that correlations could potentially extend
across an arbitrarily large box size in the limit of system size. Thus, as the critical regime is approached in
this way, this signature of a critical system emerges. LRTCs have been demonstrated previously in discrete
neurophysiological data, in the waiting times of burst activity in cultures [34] and in the bursts of activity
recorded using EEG in very preterm human neonates [35]. To our knowledge, waiting times of neuronal
avalanches have yet to be examined in this manner. However, such a study would provide an additional
link between studies on the neuronal scale and studies on a wider network scale for which LRTCs have been
observed in the fluctuations of oscillation amplitude. Palva and colleagues demonstrated strong correlations
between power-law exponents of avalanche size distributions and exponents of LRTCs in fluctuations of
oscillation amplitude in human MEG recordings [33]. Recent computational work also demonstrated a link
32
between neuronal avalanches on the one scale and LRTCs on a wider temporal scale and the authors called
for future work in this area [32]. However, these authors did not investigate LRTCs in the waiting times of
the avalanches themselves. Interestingly, in our model, LRTCs were observed when h = 1/N but not for
lower levels of external input. Thus, they were not observed when the avalanche size distribution exhibited
scale-free behaviour -- the type of distribution observed for avalanches recorded in vivo and in vitro [5, 6, 9].
It would therefore also be interesting to assess whether altering the driving force experimentally in vitro
would lead to the types of dynamics (LRTCs) observed here.
Partial scale-free behaviour in avalanche size
Statistical testing of the avalanche size distribution (with h = 0.1/N, 0.01/N ) did not reject the hypothesis
that the distribution followed a power-law when the number of points within the distribution was of the order
of the number of avalanches recorded in the experimental setting. Only with larger numbers of avalanches
was the hypothesis that the distribution is a power-law rejected. This is to be expected, as has been discussed
by Klaus and Plenz [2]. When a distribution deviates from the expected distribution by more than noise from
sampling then given a large enough number of samples the power-law hypothesis will eventually be rejected.
The fact that the power-law hypothesis was not rejected for lower numbers of avalanches demonstrates the
partial scale-free behaviour of the system in the region of the critical regime. Further, it highlights the
fact that stringent statistical testing, such as this, with high sampling may lead to rejecting the power-law
hypothesis and so rejecting the criticality hypothesis even when the system is critical.
Waiting times
In addition to increasing the physiological realism of the model, investigating the system with continuous
external input also has the advantage of producing waiting times (termed IAIs throughout). In the companion
paper the simple reseeding of the network with a neuron set to the active state implied that there was no
waiting times between avalanches. Other authors have reseeded by increasing the membrane potential
but stipulated that neurons must reach a threshold for them to become active (and a new avalanche to
start) [18, 25]. This does lead to waiting times, however, these are not the same as the waiting times
investigated in this model which are intrinsic to the network dynamics rather than as a result of an explicit
separation of timescales.
Recent work by Lombardi and colleagues [31] showed that the waiting times between neuronal avalanches
recorded in vitro have a distribution with an initial power-law regime. The authors suggest that the shape
33
of the distribution relates to up and down states within the network (which exhibit critical and subcritical
dynamics respectively) and are able to reproduce the non-monotonic waiting time distribution in a compu-
tational model in which neurons switch between up and down states depending on short-term firing history.
Interestingly, the distribution they observe is similar to the IAI distribution for the system with h = 0.1/N ,
see Fig. 9(b), which also has a scale-free initial regime albeit over a shorter range to that presented by
Lombardi et al. It is therefore possible that the waiting time distribution observed experimentally fits with
the model constructed here. It would be interesting to investigate whether a change in input to the network
in vitro alters the distribution in a similar way to those distributions seen in Fig. 9.
Additionally, for different parameter ranges different distributions were observed, in the IAI distribution
as well as the distributions of avalanche size and duration. This leads us to the important conclusion
that power-law distributions will not necessarily be displayed by systems in the region of a critical regime.
Therefore, this work suggests that the absence of a power-law in experimental data should not necessarily
be taken to conclude that the system does not lie in the region of a critical regime. This was also seen in the
companion paper where it was shown that despite being analytically tuned to the critical state (in absence of
external input) the avalanche size distribution was not a power-law although it did exhibit partial scale-free
behaviour. The fact that the system may not exhibit power-laws when close to (or at) the critical regime is
an important finding given that the system is finite-size as will be the case in an experimental setting. This
highlights the necessity of examining other markers of criticality before conclusions about the critical nature
of a system can be drawn.
Dynamic range and power-laws
Coinciding with results from previous authors [16, 29] we showed that the system exhibits optimal dynamic
range when the branching parameter is equal to one. When calculating the dynamic range of a system, we
emphasised that this value was dependent on the critical state of the system calculated when there was no
external input. We have shown that tuning a system to this critical point but then driving it with different
levels of external input has considerable effect on the distribution of avalanche sizes. For non-zero h the
corresponding ODE would, in the strictest sense, not be considered critical. Importantly, however, tuning
to the critical point of the system with zero external input, maximises the dynamic range.
Dehghani and colleagues [56] showed that in vivo (against the results of Petermann et al. [9] and Hahn et
al. [8]) avalanches were not well approximated by power-laws, but were more likely to approach exponential
distributions. They contrast this with the evidence that the brain is operating at criticality from in vitro
34
studies [5,57] where avalanches are well approximated by power-laws. Here we argue that external input and
functional benefits [17] such as dynamic range, information transmission and information capacity, provide
an interesting possibility as to the reason why in vivo and in vitro studies could potentially give different
results. The critical brain hypothesis demands that in isolation from its natural surroundings (in vitro)
and whilst having no external influences acting upon it (akin to the model with h = 0 we studied in the
companion paper [24]), a culture should exhibit signs that it is tuned to criticality (i.e. avalanches that are
well approximated by power-laws). However, when observed in vivo, and thus with external inputs acting
upon it, a critical brain may no longer exhibit avalanches approximated by power-laws but still optimise
functional benefits such as the dynamic range and information transmission [17]. In our model we have shown
that tuning the parameters to the critical regime does indeed maximise the dynamic range, but it is the level
of external input that dictates whether the avalanche distributions exhibit partial scale-free behaviour. For
this reason, avalanches recorded in vivo that lacked a power-law distribution would not be in contradiction
with the criticality hypothesis but rather an expected result. This further supports our suggestion in the
companion paper [24] that future work should shift its focus away from characterising avalanche distributions
and towards more appropriate metrics.
Two routes to criticality
In this paper we examined two different parameter changes such that the system approaches the critical
state: increasing the system size and lowering the overall level of the external input. Despite the fact that
in both cases the critical regime is approached, the dynamics and the signatures of criticality observed are
different. With increasing system size the temporal correlations extend across a wider range. However, the
distributions of the avalanche characteristics remain the same and the distribution of avalanche size does
not exhibit scale-free behaviour. By contrast, for lower overall levels of the external input the distributions
of avalanche size and duration do exhibit partial scale-free behaviour. However, in this case as the critical
regime is approached the temporal correlations in the avalanches are lost. At these lower levels of the
external input we also observe a greater separation of the avalanches suggesting that the avalanches have
less of an influence on each other which would explain this loss of LRTCs. Thus, as the system approaches
the critical state in two different regions of the parameter space the dynamical properties of the system are
very different. Significantly, this implies that not just the critical state alone but the region around the
critical regime is an important factor in the system's dynamics.
In conclusion, we have shown here and in the companion paper that in a finite-size neuronal system
35
in the region of a critical regime the distributions of avalanche attributes need not be a power-law. The
current assumption in the literature is that power-law dynamics imply criticality and vice versa that systems
without power-law dynamics are not in the region of a critical regime, however, the results here suggest that
this assumption need not be true. Moreover, we found that long-range temporal correlations and scale-free
distributions are not dependent on proximity to the critical regime alone but on the region of the parameter
space. The results further highlight the need for future work examining the type of dynamics we might
expect from such systems.
Appendix 1: Dynamic range
Whilst [16] and [29] consider a discrete model where multiple events can happen per time step, here we
show analytically that our continuous model will exhibit the same maximisation of the dynamic range when
R0 = 1. We use the calculation of R0 for a system where there is no external input (h = 0) and thus
R0 = w/α.
We begin by defining (as in Kinouchi and Copelli [16]) Fmax(R0) as the saturation level of neurons in a
network assuming a large external input h. For our model Fmax(R0) = N for all R0. Similarly we define
F0(R0) as the steady state solution of the mean field ODE for the system when there is zero external input,
i.e.
dA
dt
=
=
Therefore, solving this we have that
(cid:19)
+ h
(N − A) − αA
(N − A) − αA.
(cid:18) wA
N
wA
N
(cid:40)
0
Additionally let Fx(R0) = F0(R0) + x [Fmax(R0) − F0(R0)] giving
F0(R0) =
if R0 ≤ 1
if R0 > 1.
w
N(cid:0)1 − α
(cid:1)
N(cid:2)1 − α
w (1 − x)(cid:3)
N x
(cid:40)
Fx(R0) =
if R0 ≤ 1
if R0 > 1.
Finally, let A(σ, y) be the number of active neurons at the steady state in a regime where R0 = σ and h = y
(where σ and y are dummy variables and h is the external input), then the dynamic range ∆(R0) is defined
(similarly to [16]) as:
∆(R0) =
h0.9
h0.1
,
36
where
h0.1 is the level of external input such that A(R0, h0.1) = F0.1(R0) = F0.1
and h0.9 is the level of the external input such that A(R0, h0.9) = F0.9(R0) = F0.9.
We note that in [16, 29], the logarithm of this is taken but as the logarithm is an increasing function it is
unnecessary to scale in this way for the result we obtain. Whilst using F0.1 and F0.9 is the standard for
calculating the dynamic range these values are somewhat arbitrary [16] and can be generalised to k1 and k2
respectively. To calculate the dynamic range analytically we consider the two regimes of R0, firstly R0 ≤ 1
and secondly R0 > 1.
R0 ≤ 1
Here the steady state is given by
(cid:18) wFk
(cid:19)
N
+ hk
(N − Fk) − αFk = 0
⇒ (wk + hk) (N − N k) − αN k = 0
⇒hk =
1 − k − wk,
αk
∆ =
hk2
hk1
=
=
[αk2 − wk2(1 − k2)] (1 − k1)
[αk1 − wk1(1 − k1)] (1 − k2)
k2(1 − k1) [1 − R0(1 − k2)]
k1(1 − k2) [1 − R0(1 − k1)]
(w − α + αk)
∆ =
hk2
hk1
=
=
k2(1 − k1)(w − α + αk2)
k1(1 − k2)(w − α + αk1)
k2(1 − k1)(R0 − 1 + k2)
k1(1 − k2)(R0 − 1 + k1)
37
thus
R0 > 1
(cid:19)
(cid:18) wFk
(cid:16)
(cid:104)
N
w
α
w
1 −
k
1 − k
⇒hk =
thus
Here the steady state is given by
+ hk
(N − Fk) − αFk = 0
(cid:17)
(cid:105)(cid:20) N α
(cid:21)
⇒
(1 − k)
+ hk
(1 − k)
− αN
w
(cid:104)
1 −
(cid:105)
= 0
α
w
(1 − k)
Maximum of ∆(R0)
Calculating the derivative of ∆(R0) we find that if 0 < k1 < k2 < 1, then for R0 ≤ 1, d∆
R0 > 1, d∆
dR0
< 0. Thus the maximum of ∆(R0) is achieved for R0 = 1. It is worth noting that ∆(R0) is
> 0, whilst for
dR0
independent of N and only depends on the choice of k1 and k2.
Appendix 2: Driving the system from a subcritical and supercritical state
Throughout the paper we have examined parameters such that the system is critical when there is no external
input. In the presence of a small external input we therefore investigate driving the system in the region of
this critical state. In the companion paper [24], with no external input, we also investigated the system with
subcritical and supercritical parameters. In this appendix we briefly examine the dynamics of the system as
it is driven from these states by an external input.
Fig. 13 shows raster plots of network firings when the system is driven from a subcritical and supercritical
state with h = 1/N . Compared with the critical case, see Fig. 3(a), with the subcritical parameter set the
bursts appear to be shorter and consist of fewer neurons firing. Conversely, in the supercritical case the bursts
appear longer and consist of denser network firing. Fig. 14 shows the IAI distributions for the subcritical
and supercritical parameters. As expected from the raster plots, the IAIs are longer in the subcritical case
compared with the critical (Fig. 9(a)) and the supercritical. While the subcritical distribution appears
to exhibit partial scale-free behaviour similar to the critical case, the supercritical distribution loses this
appearance. The distributions from simulations are shown with the theoretical distribution calculated as
previously described as a weighted sum of hypoexpontials.
Fig. 15 shows the distributions of avalanche size and duration in the subcritical and supercritical cases.
Contradicting what we would expect from the raster plots we find that the avalanche sizes are smaller (on
average) in the supercritical system. In the companion paper we showed that the supercritical distribution
(without the presence of external input) had an increased number of large avalanches compared with the
distribution for the system at criticality. However, we do not find this here. As the firing with the supercritical
parameters is relatively dense we believe that this highlights a limitation with the binning method in this
case. We suggest that future research should focus on how binning can influence avalanche distributions.
Appendix 3: Altering the activation function
Throughout this paper we considered a linear activation function. What happens if a different activation
function is chosen? Do we observe the same type of dynamics? In this appendix we briefly investigate two
38
Figure 13: Raster plots of neuronal firing for the network driven from subcritical and super-
critical states. The network firing for (a) subcritical, α = 1.1, w = 1 ⇒ λ < 0 and (b) supercritical,
α = 0.9, w = 1 ⇒ λ > 0 parameter sets. Here we investigate the system with a small external input
(h = 1/N ) which drives the system slightly away from these fixed points. The red line indicates the level
of firing in 1 ms bins. The subcritical case appears to give rise to smaller bursts and the supercritical case
leads to a greater level of firing and longer burst activity compared with the critical system (see Fig. 3).
Figure 14: Distribution of IAIs for the system driven from subcritical and supercritical states.
The theoretical (red) and simulated (black) IAI distributions with (a) α = 1.1 (subcritical) and (b) α = 0.9
(supercritical) parameters. The blue line in (a) indicates a linear fit, i.e. a fitted power-law with an exponent
of 2.37. These distributions are with N = 800, w = 1, h = 1/N and the theoretical distributions were
calculated up to a level of initial active neurons which occur with a cumulative probability of 0.9 and 0.13
respectively (see main text).
39
0.10.52.05.0IAIProbability10−810−510−1(a)0.020.100.502.00IAIProbability10−810−510−1(b) Figure 15: Distributions of avalanche size and duration for the system driven from subcritical
and supercritical regimes. Avalanche size and duration distributions for the system driven from (a,c)
subcritical (α = 1.1) and (b,d) supercritical (α = 0.9) states. Simulations were run with N = 800, w =
1, h = 1/N . The red line in (a) shows the linear fit with a slope of -2.66.
40
1550500Avalanche sizeProbability(a)10−610−410−1Avalanche sizeProbability(b)10−710−410−11520500.10.52.010.0Avalanche durationProbability(c)10−610−410−10.010.050.200.50Avalanche durationProbability(d)10−710−410−1 other activation functions: an exponential and a quadratic.
First let us consider the system with an exponential activation function such that:
(cid:18)
(cid:19)
dA
dt
=
1
N A+h) −
1 + e−( w
1
2
(N − A) − αA.
This function saturates and so is somewhat more realistic than the linear function considered previously.
Also, note that the function is set such that when A and h are both zero we also have f (x) = 0, i.e. without
any external input and with no active neurons the network will remain in this state. With this activation
function the eigenvalues of the fixed points are given by:
We have that
f(cid:48)(x) =
w
N e−( w
(1 + e−( w
N A+h)
N A+h))2
=
w
N
f (x) +
1
2
2 − f (x)
λ = f(cid:48)(x)(N − A) − f (x) − α.
(cid:18)
(cid:19)(cid:18) 1
(cid:19)
(cid:18) 1
4 − f 2(x)
(cid:19)
,
=
w
N
and so this could be used to find a critical fixed point along with the fact that at the fixed point of the
system we have that:
f (x) =
αA
,
(N − A)
which defines the level of the external input at the critical fixed point.
As before, consider initially the case where there is no external input (h = 0). In this case A = 0 is a
fixed point, which is critical (with λ = 0) if and only if α = w/4 by the above equations. What happens
to this system in the presence of small external input? Fig. 16 shows the raster plot for the three different
levels of the external input considered previously: h = 0.01/N, 0.1/N, 1/N . Comparing with Fig. 3, the
firing rate is lower with the saturating function studied here, however, the overall pattern of firing is the
same. For all three levels of the external input we continue to observe avalanche dynamics and for lower
levels of the external input (as the system approaches the critical regime) these avalanches become more
distinct. Fig. 17 shows the avalanche size, duration and IAI distributions for each of these three levels of the
external input. Comparing with Figures 4 and 9 we find that a similar relationship with the critical regime
emerges. With h = 1/N the IAI distribution shows scale-free behaviour (note that by the same derivation
as previously, theoretically the distribution is a weighted sum of hypoexponentials). For lower levels of the
external input the scale-free behaviour in the IAI distribution is lost but the distribution of avalanche sizes
appears scale-free. As was shown previously for the system with a linear activation function, we also found
that when h = 1/N the IAIs exhibited LRTCs up to a crossover point (data not shown). For lower levels of
the external input these correlations were lost.
41
Figure 16: Raster plots for different levels of the external input with the saturating activation
function. Neuronal firing during the first 2 seconds of example simulations with (a) h = 1/N , (b) h = 0.1/N
and (c) h = 0.01/N . For all three simulations w = 1, α = 0.25 and N = 800. Comparing with Figure 3 we
find that while in this case the firing rate is lower (note the longer time scale over which the raster plot is
displayed) the overall pattern is the same, with the avalanches becoming more distinct with lower levels of
the external input.
Figure 17: IAI, avalanche size and duration distributions for the system with the saturating
activation function. The distributions are for simulations with h = 1/N (a,d,g), h = 0.1/N (b,e,h) and
h = 0.01/N (c,f,i). For all simulations N = 800, α = 0.25, w = 1 and the distributions are pooled from 10
simulations each of length 104 seconds. The red lines indicates linear fits on the double logarithmic scale,
i.e. fitted power-laws with exponents of (a) 2.65, (e) 1.81, (f) 1.49 and (h) 1.84.
42
0500100020000400800Time (ms)Neuron Index(a)0500100020000400800Time (ms)Neuron Index(b)0500100020000400800Time (ms)Neuron Index(c)0.21.05.050.0IAIProbability10−710−410−1(a)IAIProbability10−510−310−11550500(b)10505005000IAIProbability10−510−3(c)1520100Avalanche sizeProbability10−710−410−1(d)Avalanche sizeProbability10−510−310−111010010000(e)Avalanche sizeProbability10−510−310−111010010000(f)Avalanche durationProbability10−810−410−10.11520(g)Avalanche durationProbability10−710−410−10.51050500(h)Avalanche durationProbability(i)10−510−310−10.1110500 With both the linear and saturating activation function we considered the system in the region of the
critical regime, with the system driven from the critical regime by the positive external input. Consider the
system instead with a quadratic activation function:
(cid:16) w
N
dA
dt
=
(cid:17)
A2 + h
(N − A) − αA,
With this activation function the fixed points are given by:
with eigenvalues:
g(A) =
dA
dt
w
N
= −
+ wA2 − (h + α)A + hN = 0,
λ = g(cid:48)(A) = −3
w
N
A2 + 2wA − h − α.
Solving these simultaneously we find:
w
N
A2 + 2wA − h,
α = −3
2wA3 − wN A2 + hN 2 = 0,
which define the parameter space and the value of the fixed point for which a critical fixed point can be
obtained. Thus, we find that unlike the model with the linear (and saturating) activation function, here
with a non-zero external input it is possible to tune the system so that it is directly at the critical regime.
Upon examining this parameter space one can note that in many cases there also exists a stable (positive)
fixed point as well as the critical fixed point. From simulating such a system we found (data not shown)
that the dynamics of the system are quickly attracted to the stable fixed point and so the critical fixed
point has little affect on the dynamics. Therefore, to have a system which is affected by a critical fixed
point in the presence of a non-zero external input (in the case of this activation function and where positive
parameters are required) the critical regime must be the only fixed point of the system. Given that g(A) is
a cubic equation, to achieve a single fixed point which is critical this point must be an inflection point with
g(cid:48)(A) = 0 and g(cid:48)(cid:48)(A) = 0. From these equalities we find that the critical fixed point is A = N/3 and we
must also have h = wN
27 and α = 8wN
27 .
Fig. 18 shows a raster firing plot and the number of active neurons throughout a simulation for the system
with a single critical fixed point. As would be expected, the number of active neurons fluctuates about the
critical point. Previously when considering avalanche dynamics we have binned the firing. However, as
noted in the discussion the binning method will always separate firing into avalanches and as there are no
clear periods of inactivity this does not seem appropriate here. Recall that in the zero input case (see the
43
companion paper) we seeded the system so perturbing it away from the fully quiescent state (which was
the critical fixed point) and defined an avalanche as the firing that occurred before the system returned to
the fully quiescent state. In a similar approach here it is possible to define an avalanche as the number of
neurons that fire in a single excursion from the critical fixed point. We therefore counted the number of
neurons that fired from when the system was deflected (either in a positive or negative direction) from the
fixed point (A = N/3) until the next time at which the system had exactly N/3 active neurons. Fig. 18
shows the probability distribution of the size of the avalanches defined in this way. The distribution appears
to be scale-free over a range of scales.
Thus, while critical dynamics may not be apparent initially when examining data (for example if we were
to look at the overall dynamics from the simulations with quadratic activation function), we can observe
signatures of criticality when the dynamics are examined in relation to the known critical regime. Here we
can note that the network firing fluctuates about the critical regime - that is the number of active neurons
fluctuates about this regime and so the average number of active neurons across the course of a simulation is
approximately equal to the critical state of N/3. It might therefore be interesting to examine the fluctuations
about the mean activity level in experimental settings where activity is continuous (i.e. cannot be described
as intermittent avalanche-like activity) to determine whether signatures of criticality are present. Indeed,
such an approach has been taken previously to examine MEG data, thresholding at the median level [58].
Acknowledgements
Caroline Hartley is funded through CoMPLEX (Centre for Mathematics and Physics in the Life Sciences
and Experimental Biology), University College London. Timothy Taylor is funded by a PGR studentship
from MRC, and the Departments of Informatics and Mathematics at University of Sussex. Istvan Z. Kiss ac-
knowledges support from EPSRC (EP/H001085/1). Simon Farmer acknowledges support from the National
Institute for Health Research University College London Hospitals Biomedical Research Centre.
44
Figure 18: Dynamics of the network with a quadratic activation function. (a) Raster plot of network
firing for simulation with N = 800, h = wN
27 , w = 0.01. (b) For the same simulation this plot
shows the number of active neurons (A) at each simulation step. The distribution of avalanches (c) and
positive avalanches only (d) - as described in the main text - pooled from 20 simulations of 1000 seconds in
length. The red lines indicate linear fits, i.e. fitted power-laws, both with exponents of 1.48.
27 , α = 8wN
45
References
1. Clauset A, Shalizi CR, Newman MEJ: Power-Law Distributions in Empirical Data. Siam Review 2009,
51(4):661 -- 703.
2. Klaus A, Yu S, Plenz D: Statistical analyses support power law distributions found in neuronal
avalanches. PLoS One 2011, 6(5):e19779.
3. Chialvo DR: Emergent complex neural dynamics. Nature Physics 2010, 6(10):744 -- 750.
4. Linkenkaer-Hansen K, Nikouline VV, Palva JM, Ilmoniemi RJ: Long-range temporal correlations and scal-
ing behavior in human brain oscillations. J Neurosci 2001, 21(4):1370 -- 7.
5. Beggs JM, Plenz D: Neuronal avalanches in neocortical circuits. J Neurosci 2003, 23(35):11167 -- 77.
6. Beggs JM, Plenz D: Neuronal avalanches are diverse and precise activity patterns that are stable for
many hours in cortical slice cultures. J Neurosci 2004, 24(22):5216 -- 29.
7. Gireesh ED, Plenz D: Neuronal avalanches organize as nested theta- and beta/gamma-oscillations
during development of cortical layer 2/3. Proc Natl Acad Sci U S A 2008, 105(21):7576 -- 81.
8. Hahn G, Petermann T, Havenith MN, Yu S, Singer W, Plenz D, Nikolic D: Neuronal avalanches in sponta-
neous activity in vivo. J Neurophysiol 2010, 104(6):3312 -- 22.
9. Petermann T, Thiagarajan TC, Lebedev MA, Nicolelis MAL, Chialvo DR, Plenz D: Spontaneous corti-
cal activity in awake monkeys composed of neuronal avalanches. Proc Natl Acad Sci U S A 2009,
106(37):15921 -- 6.
10. Shriki O, Alstott J, Carver F, Holroyd T, Henson RNA, Smith ML, Coppola R, Bullmore E, Plenz D: Neuronal
avalanches in the resting MEG of the human brain. J Neurosci 2013, 33(16):7079 -- 90.
11. Linkenkaer-Hansen K, Nikulin VV, Palva JM, Kaila K, Ilmoniemi RJ: Stimulus-induced change in long-
range temporal correlations and scaling behaviour of sensorimotor oscillations. Eur J Neurosci 2004,
19:203 -- 11.
12. Nikulin VV, Brismar T: Long-range temporal correlations in alpha and beta oscillations: effect of
arousal level and test-retest reliability. Clin Neurophysiol 2004, 115(8):1896 -- 908.
13. Nikulin VV, Brismar T: Long-range temporal correlations in electroencephalographic oscillations:
Relation to topography, frequency band, age and gender. Neuroscience 2005, 130(2):549 -- 58.
14. Smit DJA, de Geus EJC, van de Nieuwenhuijzen ME, van Beijsterveldt CEM, van Baal GCM, Mansvelder HD,
Boomsma DI, Linkenkaer-Hansen K: Scale-free modulation of resting-state neuronal oscillations reflects
prolonged brain maturation in humans. J Neurosci 2011, 31(37):13128 -- 36.
15. Berthouze L, James LM, Farmer SF: Human EEG shows long-range temporal correlations of oscilla-
tion amplitude in Theta, Alpha and Beta bands across a wide age range. Clin Neurophysiol 2010,
121(8):1187 -- 97.
16. Kinouchi O, Copelli M: Optimal dynamical range of excitable networks at criticality. Nature Physics
2006, 2(5):348 -- 352.
17. Shew WL, Plenz D: The Functional Benefits of Criticality in the Cortex. Neuroscientist 2013, 19:88 -- 100.
18. Bak P, Tang C, Wiesenfeld K: Self-organized criticality: An explanation of the 1/f noise. Phys Rev Lett
1987, 59(4):381 -- 384.
19. Jensen HJ: Self-organized criticality: emergent complex behaviour in physical and biological systems. Cambridge
University Press 1998.
20. Essam J: Percolation Theory. Rep Prog Phys 1980, 43:833 -- 912.
21. Harris TE: The theory of branching processes. Die Grundlehren der mathematischen Wissenschaften in Einzel-
darstellungen, Berlin: Springer 1963.
22. Priesemann V, Munk MHJ, Wibral M: Subsampling effects in neuronal avalanche distributions recorded
in vivo. BMC Neurosci 2009, 10:40.
23. Bonachela JA, Munoz MA: Self-organization without conservation:
true or just apparent scale-
invariance? J. Stat. Mech. 2009, :P09009.
46
24. Taylor TJ, Hartley C, Simon PL, Kiss IZ, Berthouze L: Identification of criticality in neuronal avalanches:
I. A theoretical investigation of the non-driven case. J Math Neurosci 2013, 3:5.
25. Levina A, Herrmann JM, Geisel T: Dynamical synapses causing self-organized criticality in neural
networks. Nature Physics 2007, 3:857 -- 860.
26. Olami Z, Feder H, Christensen K: Self-organized criticality in a continuous, nonconservative cellular
automaton modeling earthquakes. Phys Rev Lett 1992, 68(8):1244 -- 1247.
27. Ribeiro TL, Copelli M: Deterministic excitable media under Poisson drive: power law responses,
spiral waves, and dynamic range. Phys Rev E Stat Nonlin Soft Matter Phys 2008, 77(5 Pt 1):051911.
28. Rubinov M, Sporns O, Thivierge JP, Breakspear M: Neurobiologically realistic determinants of self-
organized criticality in networks of spiking neurons. PLoS Comput Biol 2011, 7(6):e1002038.
29. Larremore DB, Shew WL, Restrepo JG: Predicting criticality and dynamic range in complex networks:
effects of topology. Phys Rev Lett 2011, 106(5):058101.
30. Boffetta G, Carbone V, Giuliani P, Veltri P, Vulpiani A: Power laws in solar flares: Self-organized criti-
cality or turbulence? Physical Review Letters 1999, 83(22):4662 -- 4665.
31. Lombardi F, Herrmann HJ, Perrone-Capano C, Plenz D, de Arcangelis L: Balance between Excitation
and Inhibition Controls the Temporal Organization of Neuronal Avalanches. Phys Rev Lett 2012,
108(22):228703.
32. Poil SS, Hardstone R, Mansvelder HD, Linkenkaer-Hansen K: Critical-state dynamics of avalanches and
oscillations jointly emerge from balanced excitation/inhibition in neuronal networks. J Neurosci
2012, 32(29):9817 -- 23.
33. Palva JM, Zhigalov A, Hirvonen J, Korhonen O, Linkenkaer-Hansen K, Palva S: Neuronal long-range tem-
poral correlations and avalanche dynamics are correlated with behavioral scaling laws. Proc Natl
Acad Sci U S A 2013, 110(9):3585 -- 90.
34. Segev R, Benveniste M, Hulata E, Cohen N, Palevski A, Kapon E, Shapira Y, Ben-Jacob E: Long term behavior
of lithographically prepared in vitro neuronal networks. Phys Rev Lett 2002, 88(11):118102.
35. Hartley C, Berthouze L, Mathieson SR, Boylan GB, Rennie JM, Marlow N, Farmer SF: Long-range temporal
correlations in the EEG bursts of human preterm babies. PLoS One 2012, 7(2):e31543.
36. Benayoun M, Cowan JD, van Drongelen W, Wallace E: Avalanches in a stochastic model of spiking
neurons. PLoS Comput Biol 2010, 6(7):e1000846.
37. Gillespie D: Exact stochastic simulation of coupled chemical reactions. Journal of Physical Chemistry
1977, 81(25):2340 -- 2361.
38. Ross SM: Introduction to probability models. Amsterdam: Academic Press, 10th ed edition 2010.
39. Scheuer E: Reliability of an m-out of-n system when component failure induces higher failure rates
in survivors. IEEE Transactions On Reliability 1988, 37:73 -- 74.
40. Amari S, Misra R: Closed-form expressions for distribution of sum of exponential random variables.
IEEE Transactions On Reliability 1997, 46:519 -- 522.
41. Peng CK, Havlin S, Stanley HE, Goldberger AL: Quantification of scaling exponents and crossover phe-
nomena in nonstationary heartbeat time series. Chaos 1995, 5:82 -- 7.
42. Toweill DL, Kovarik WD, Carr R, Kaplan D, Lai S, Bratton S, Goldstein B: Linear and nonlinear analysis of
heart rate variability during propofol anesthesia for short-duration procedures in children. Pediatr
Crit Care Med 2003, 4(3):308 -- 14.
43. Castiglioni P, Parati G, Di Rienzo M, Carabalona R, Cividjian A, Quintin L: Scale exponents of blood
pressure and heart rate during autonomic blockade as assessed by detrended fluctuation analysis.
J Physiol 2011, 589(Pt 2):355 -- 69.
44. Ho KK, Moody GB, Peng CK, Mietus JE, Larson MG, Levy D, Goldberger AL: Predicting survival in heart
failure case and control subjects by use of fully automated methods for deriving nonlinear and
conventional indices of heart rate dynamics. Circulation 1997, 96(3):842 -- 8.
45. Taqqu M, Teverovsky V, Willinger W: Estimators for long-range dependence: An empirical study.
Fractals 1995, 3(4):785 -- 798.
47
46. Peng CK, Buldyrev SV, Havlin S, Simons M, Stanley HE, Goldberger AL: Mosaic organization of DNA
nucleotides. Phys Rev E Stat Phys Plasmas Fluids Relat Interdiscip Topics 1994, 49(2):1685 -- 9.
47. Hu K, Ivanov PC, Chen Z, Carpena P, Stanley HE: Effect of trends on detrended fluctuation analysis.
Phys Rev E Stat Nonlin Soft Matter Phys 2001, 64(1 Pt 1):011114.
48. McSharry P: DFA Matlab code found on the website of the Systems analysis, modelling
and prediction group, Department of Engineering Sciences, University of Oxford. Available:
http://www.eng.ox.ac.uk/samp/software/ cardiodynamics/dfa.m. DFA Matlab code found on the web-
site of the Systems analysis, modelling and prediction group, Department of Engineering Sciences, University
of Oxford. Available: http://www.eng.ox.ac.uk/samp/software/ cardiodynamics/dfa.m 2005, Last checked: 11
January 2011.
49. Sporns O: The non-random brain: efficiency, economy, and complex dynamics. Front Comput Neurosci
2011, 5:5.
50. Cherubini E, Gaiarsa JL, Ben-Ari Y: GABA: an excitatory transmitter in early postnatal life. Trends
Neurosci 1991, 14(12):515 -- 9.
51. Ben-Ari Y: Excitatory actions of gaba during development: the nature of the nurture. Nat Rev
Neurosci 2002, 3(9):728 -- 39.
52. Holmes GL, Khazipov R, Ben-Ari Y: New concepts in neonatal seizures. Neuroreport 2002, 13:A3 -- 8.
53. Shew WL, Yang H, Petermann T, Roy R, Plenz D: Neuronal avalanches imply maximum dynamic range
in cortical networks at criticality. J Neurosci 2009, 29(49):15595 -- 600.
54. Botcharova M, Farmer SF, Berthouze L: A maximum likelihood based technique for validating de-
trended fluctuation analysis (ML-DFA). arXiv:1306.5075 2013.
55. Gao J, Hu J, Tung WW, Cao Y, Sarshar N, Roychowdhury VP: Assessment of long-range correlation in
time series: how to avoid pitfalls. Phys Rev E Stat Nonlin Soft Matter Phys 2006, 73(1 Pt 2):016117.
56. Dehghani N, Hatsopoulos NG, Haga ZD, Parker R, Greger B, Halgren E, Cash SS, Destexhe A: Avalanche
analysis from multi-electrode ensemble recordings in cat, monkey and human cerebral cortex
during wakefulness and sleep. Frontiers in Physiology 2012, 3(302).
57. Friedman N, Ito S, Brinkman BAW, Shimono M, DeVille REL, Dahmen KA, Beggs JM, Butler TC: Universal
Critical Dynamics in High Resolution Neuronal Avalanche Data. Phys. Rev. Lett. 2012, 108:208102.
58. Poil SS, van Ooyen A, Linkenkaer-Hansen K: Avalanche dynamics of human brain oscillations: relation
to critical branching processes and temporal correlations. Hum Brain Mapp 2008, 29(7):770 -- 7.
48
|
1706.03836 | 1 | 1706 | 2017-06-12T20:16:42 | The interplay between long- and short-range temporal correlations shapes cortex dynamics across vigilance states | [
"q-bio.NC",
"nlin.AO"
] | Increasing evidence suggests that cortical dynamics during wake exhibits long-range temporal correlations suitable to integrate inputs over extended periods of time to increase the signal-to-noise ratio in decision-making and working memory tasks. Accordingly, sleep has been suggested as a state characterized by a breakdown of long-range correlations; detailed measurements of neuronal timescales that support this view, however, have so far been lacking. Here we show that the long timescales measured at the individual neuron level in freely-behaving rats during the awake state are abrogated during non-REM (NREM) sleep. We provide evidence for the existence of two distinct states in terms of timescale dynamics in cortex: one which is characterized by long timescales which dominate during wake and REM sleep, and a second one characterized by the absence of long-range temporal correlations which characterizes NREM sleep. We observe that both timescale regimes can co-exist and, in combination, lead to an apparent gradual decline of long timescales during extended wake which is restored after sleep. Our results provide a missing link between the observed long timescales in individual neuron fluctuations during wake and the reported absence of long-term correlations during deep sleep in EEG and fMRI studies. They furthermore suggest a network-level function of sleep, to reorganize cortical networks towards states governed by slow cortex dynamics to ensure optimal function for the time awake. | q-bio.NC | q-bio | The interplay between long- and short-range temporal
correlations shapes cortex dynamics across vigilance
states
Christian Meisel1∗, Andreas Klaus1, Vladyslav V. Vyazovskiy2 and Dietmar
Plenz1
1 Section on Critical Brain Dynamics, National Institute of Mental Health, Bethesda, Maryland
20892, USA
2 Department of Physiology, Anatomy and Genetics, University of Oxford, Parks Road, Oxford
OX1 3PT, UK
∗ corresponding author email address: [email protected]
Short title: Cortical timescales in different vigilance states
Number of pages: 31
Number of figures: 4
Number of words for Abstract: 248
Number of words Introduction: 390
Number of words Discussion: 600
Conflict of Interest: N/A
Acknowledgements: This study was supported by the Intramural Research Program of the
NIMH. This study utilized the high-performance computational capabilities of the Biowulf
Linux cluster at the National Institutes of Health, Bethesda, Md.
7
1
0
2
n
u
J
2
1
]
.
C
N
o
i
b
-
q
[
1
v
6
3
8
3
0
.
6
0
7
1
:
v
i
X
r
a
1
Increasing evidence suggests that cortical dynamics during wake exhibits long-
range temporal correlations suitable to integrate inputs over extended periods
of time to increase the signal-to-noise ratio in decision-making and working
memory tasks. Accordingly, sleep has been suggested as a state character-
ized by a breakdown of long-range correlations; detailed measurements of
neuronal timescales that support this view, however, have so far been lack-
ing. Here we show that the long timescales measured at the individual neuron
level in freely-behaving rats during the awake state are abrogated during non-
REM (NREM) sleep. We provide evidence for the existence of two distinct
states in terms of timescale dynamics in cortex: one which is characterized by
long timescales which dominate during wake and REM sleep, and a second one
characterized by the absence of long-range temporal correlations which char-
acterizes NREM sleep. We observe that both timescale regimes can co-exist
and, in combination, lead to an apparent gradual decline of long timescales
during extended wake which is restored after sleep. Our results provide a
missing link between the observed long timescales in individual neuron fluctu-
ations during wake and the reported absence of long-term correlations during
deep sleep in EEG and fMRI studies. They furthermore suggest a network-
level function of sleep, to reorganize cortical networks towards states governed
by slow cortex dynamics to ensure optimal function for the time awake.
2
Significance Statement
The lack of sleep deteriorates several key cognitive functions, yet the neuronal underpinnings
of these deficits have remained elusive. Cognitive capabilities are generally believed to bene-
fit from a neural circuit's ability to reliably integrate information. Persistent network activity
characterized by long timescales may provide the basis for this integration in cortex. Here we
show that long-range temporal correlations indicated by slowly decaying autocorrelation func-
tions in neuronal activity are dependent on vigilance states. While wake and REM sleep exhibit
long timescales, these long-range correlations break down during NREM sleep. Our findings
thus suggest two distinct states in terms of timescale dynamics. During extended wake, the
intermittency of short timescales can lead to an overall decline in cortical timescales.
3
Introduction
The ability to integrate information over time is a pre-requisite for brain function during wake.
Accumulating evidence suggests that the brain integrates information in a distinct progression
of increasingly longer temporal timescales that closely follows the hierarchy of cortical organi-
zation (Chaudhuri et al., 2015). Evidence for such a hierarchical ordering and the existence of
long timescales in higher order cortical areas has come from various experimental modalities in-
cluding studies on non-human primates (Ogawa and Komatsu, 2010; Murray et al., 2014), hu-
man electrocorticography (Honey et al., 2012) as well as fMRI (Stephens et al., 2013). Func-
tionally, these differentially expressed timescales provide an intriguing perspective on tem-
poral information integration in the brain: by setting the duration over which a neural cir-
cuit integrates its inputs, longer times allow to increase the signal-to-noise ratio in decision-
making and working memory (Kiebel et al., 2008; Friston et al., 2012; Chaudhuri et al., 2015;
Kringelbach et al., 2015).
While these long timescales associated with long-range temporal correlations in the history
of neural activity may provide a plausible mechanism for temporal integration during wake,
their dependence on vigilance levels is currently less well understood. It has been suggested that
during sleep the brain loses its ability to effectively integrate information across different corti-
cal areas and across time, a functional loss that has also been proposed a potential mechanism
for the concomitant loss of consciousness (Tononi, 2008). Indeed, a decrease in information
integration during nonrapid eye movement (NREM) sleep has been observed in space, quan-
tified by the effective connectivity between different cortical regions (Massimini et al., 2005),
and time (Tagliazucchi et al., 2013). However, these studies have focused on large-scale brain
signals measured by EEG and fMRI; their link to the long timescales governing individual
neuron activity is missing. Apart from the effect of sleep on cortical timescales, it is conceiv-
4
able that also more subtle vigilance changes may impact the timescales of cortical dynamics.
For example, cognitive capabilities and the ability to process information are known to be af-
fected by extended waking (Van Dongen et al., 2003; Banks and Dinges, 2007; Mignot, 2008;
Killgore, 2010). The impact of extended wakefulness on cortical timescales, however, is cur-
rently unknown.
Here, we sought to systematically characterize the timescales governing cortical dynam-
ics during wake, extended waking, and its dependency on sleep. We hypothesized that the
breakdown in long-range temporal correlations during sleep will be visible at the level of indi-
vidual neuron spiking activity. For this purpose, we analyzed neuron activity at the individual
and multi-unit level along with local field potentials (LFPs) across different cortical regions in
freely-behaving rats.
Materials and Methods
We investigated two datasets where cortical activity was monitored across different vigilance
states in rats. In the first dataset, local field potentials (LFP) and multi-unit activity (MUA)
were recorded from frontal cortex during 6 hours of sleep deprivation and after a consecutive
recovery period (n=7 animals; dataset 1). In the second dataset, LFP and single-unit activity
were recorded from frontal and parietal cortex areas during 4 hours of sleep deprivation and
during consecutive recovery sleep (n=13 animals; dataset 2).
Dataset 1
Adult male rats (Spraque Dawley, total n=7) were used in this dataset. All rats were housed
individually in transparent Plexiglas cages. Lighting and temperature (21 ± 0.5C) were kept
constant (LD 12:12, light on at 6 am, food and water available ad libitum).
5
Implantation of microelectrode arrays
Procedures were in accordance with National Institutes of Health guidelines. Animal proce-
dures were approved by the National Institute of Mental Health Animal Care and Use Commit-
tee. At age 4-6 weeks, multichannel microelectrode arrays were implanted in the right frontal
cortex (B: +1-2 mm, L: 2.5-3.5 mm) under deep isoflurane anesthesia (1.5 − 2% volume) and
presence of the analgesic ketoprofen (5 mg/kg, subcutaneous). Ketoprofen was given for up
to two days post-surgery and animals were allowed to recover for 5 days before recordings.
The arrays had 32 channels (8x4, 200µm inter-electrode spacing; 23µm electrode diameter;
Neuronexus, Ann Arbor, MI, USA). The surgical procedure was performed in sterile conditions
where, first, a ∼ 2x2 mm craniotomy was made using a high-speed surgical drill. After an inci-
sion in the dura the electrode array was then lowered into the brain tissue by penetrating the pia
mater. Electrodes were slowly advanced further into brain tissue until all rows of the array were
at the level of superficial cortical layers (layers 2/3) and most channels showed robust single-
or multiunit activity. After application of silicon gel to seal the craniotomy and to protect the
surface of the brain, dental acrylic was placed around the electrode array to fix it to the skull.
The ground wire was connected to a scull screw located 1 mm posterior to bregma.
Experimental design
At least 5 days were allowed for recovery after surgery. Sleep deprivation began at light onset
(6 am) and involved continuous observation of the animal and its LFP recording. Animals
were sleep deprived for a total of 6 hours after which an undisturbed recovery period of 5-6
hours duration followed. Prolonged wakefulness was achieved by providing rats with novel
objects and/or delivering low-level auditory or visual stimuli. Prior to the experiment, rats
were habituated to the experimenter and to the exposure to novel objects (including nesting and
bedding material from other rat cages, tubes of different shape and color, little colored blocks,
6
rubber balls and boxes of different shapes and colors). Care was taken not to stress animals;
they never appeared to engage in freezing or aggressive behavior, and exhibited no signs of
discomfort.
LFP and MUA recording during prolonged wakefulness
Neuronal activity including LFP and extracellular unit activity was recorded at 30 kHz and
stored for offline analysis (Cerebus, Blackrock Microsystems). At least 20 minutes of continu-
ous LFP and MUA were recorded for each hour and experiment during prolonged wakefulness
(from 0 to 6 hours, corresponding to SD0 to SD6, respectively) as well as after the recovery
period (REC). It has previously been pointed out that wake is not a homogenous state, and that
it is therefore important to perform analyses within substates (quiet and active wake) as similar
as possible (Fisher et al., 2016). We here, similar to previous work (Vyazovskiy et al., 2011),
focused on "quiet wake" (Q) episodes, i.e. periods when the animal is alert and has its eyes
open, readily responds to stimuli while maintaining vigilance and posture, but is immobile. To
find these episodes, we first marked periods when the animal behaviorally appeared to be in a
Q state, i.e. had its eyes open, readily responded to stimuli, maintained vigilance and posture,
but was immobile. This included the removal of episodes with rearing, grooming, eating or
drinking. Channels containing artifacts or noise were removed from further analysis of LFP
and MUA. We then carefully selected time periods of at least 4 s duration that showed LFP
signals characterized by typical wake activity, i.e. LFP signals characterized by mainly low
voltage activity with theta waves and no extended periods of slow waves typical for sleep (≥∼
1 s).
Similar to previous reports (Vyazovskiy et al., 2011) we observed occasional large-amplitude
activity in these select periods of Q, which typically presented as single positive half-waves in
the frequency range from 2-6 Hz, but could occasionally also appear in succession of 2 or more
7
waves (Fig. 1). Events like these are often not detectable in EEG (Vyazovskiy et al., 2011) and
are therefore part of what is traditionally considered the wake state. To test if or to what extent
our results depended on the occurrence of these events, all analyses were repeated on data where
segments with this kind of slow wave activity had been eliminated (see below). As a further
control, such large-amplitude segments were carefully removed from analyses in dataset 2 (see
below).
Multi-unit spikes were extracted from the high-pass filtered signal (≥ 250Hz) by applying
a threshold at negative 5 times the root mean square of the signal using the Cerebus Central
software (Blackrock Microsystems). Thresholded waveforms were subsequently offline-sorted
using the Offline Spike Sorter (Plexon Inc., Dallas, TX, USA) to eliminate artifactual wave-
forms caused by electrical or mechanical noise. Principal components (PCs, set to 3) were
extracted and automatic cluster selection was performed using the T-Distribution Expectation-
Maximization (E-M) method (Shoham et al., 2003). Only electrodes and units were used for
analysis for which the typical biphasic waveform was evident. LFPs were first downsampled
to 500 Hz and band-pass filtered (1-40 Hz) using a phase neutral filter by applying a second
order Butterworth filter in both directions. A notch filter to eliminate line noise was applied
subsequently.
Dataset 2
For dataset 2, we used recordings from a previous study described in detail in (Vyazovskiy et al., 2011).
In short, data consisted of male WKY rats from which LFP and unit activity were recorded in
frontal (n=11) and/or parietal (n=9) cortex with 16-channel micro-arrays. Similar to dataset
1, rats were housed individually in transparent Plexiglas cages (light:dark 12:12 h, light on at
10:00; food and water ad libitum and replaced daily at 10:00). Data recording and online spike
sorting were done with the Multichannel Neurophysiology Recording and Stimulation System
8
(Tucker-Davis Technologies Inc). Unit activity was collected continuously (25 kHz, 300 -- 5,000
Hz), concomitantly with the LFPs from the same electrodes and epidural EEGs (both 256Hz,
0.1 -- 100 Hz). Data consisted of 4 hours of sleep deprivation and 6 hours of consecutive recov-
ery. In each animal, two to four experiments with 4 h of sleep deprivation were performed (at
least 5 days apart). Sleep stages were scored offline by visual inspection of 4-s epochs. Spike
sorting was performed by principal component analysis (PCA) followed by split and merge ex-
pectation maximization (SMEM) clustering algorithm. Individual unit activity was binned in
2-ms bins and stored for further offline analysis. All analyses in this study were performed on
artifact-free 4s-intervals classified as vigilance stages pertaining to either quiet wake (Q), REM
sleep (R) and NREM sleep (N). Quiet wake epochs were defined as in dataset 1 (see above)
while, additionally, carefully excluding also short epochs with large-amplitude LFP activity.
For further details regarding these data, see (Vyazovskiy et al., 2011).
Spike-count autocorrelation
The temporal autocorrelation of spike counts was computed for MUA (dataset 1) and putative
single units (dataset 2) similar to previous work (Ogawa and Komatsu, 2010; Murray et al., 2014).
The 4s-intervals were divided into separate, successive time bins of duration ∆ which we set to
∆ = 100ms. Across different 4s-intervals within the same animal and hour during sleep depri-
vation (datasets 1 and 2), during recovery sleep (dataset 2) and after recovery sleep (dataset 1),
we calculated the correlation between spike counts N in two time bins, indexed by their onset
times i∆ and j∆, using the Pearsons correlation coefficient R:
R(i∆, j∆) =
Cov(N(i∆), N(j∆))
qVar(N(i∆)) × Var(N(j∆))
(1)
where covariance (Cov) and variance (Var) are computed across 4s-intervals for those time
bins. To allow reliable computation of spike-count autocorrelation, we required that each (multi-
9
)unit fired at least 8 times during the 4s-interval and that at least 10 4s-intervals within each ani-
mal and hour were available (Ogawa and Komatsu, 2010); similar to previous work this thresh-
old level was not essential to observe our main findings. One rat of dataset 1 did not exhibit
MUA at all time points during the sleep deprivation period and was thus excluded from the
analysis, so that spike-autocorrelation was derived from a total of n=6 rats in dataset 1. The
decay of autocorrelation was fit by an exponential decay to the population of neurons, i.e. the
mean across autocorrelation functions from all units, for each time point and rat:
R(k∆) = A(e−λk∆ + B),
(2)
where λ is the decay rate. Fitting was done using the PYTHON (Python Software Foun-
dation, version 2.7) function SCIPY.OPTIMIZE.CURVE FIT. As a minimal goodness of fit we
required R2 ≥ 0.5 which excluded some rare outliers due to poor fits.
Recovery from large-activity events
MUA activity increased with negative LFP (nLFP) excursions (Fig. 2). To analyze how MUA
and LFP activity recovered from such large-activity events in dataset 1, we identified these
events of increased MUA activity in each channel and for each 4s-interval as nLFP excursions
greater or equal -2 standard deviations (SD) after z-transformation (Fig. 2 a). For each such
nLFP event, both the LFP as well as the MUA in this channel for the time ±100ms around
the most negative LFP excursion was then saved for further analysis. We required the next
consecutive nLFP event to be at least 200 ms away. Next, the recovery from these nLFP was
analyzed. nLFP segments were averaged for each time point (hour during sleep deprivation or
after recovery period) and animal (Fig. 2 b) and fit by an exponential decay to derive recovery
rates for each time point and each animal. The recovery rates λ were derived by fitting an
exponential function to the nLFP:
10
S(t) = Ae−λt + B
(3)
using the PYTHON (Python Software Foundation, version 2.7) function SCIPY.OPTIMIZE.CURVE
FIT. The recovery was fit from 30 ms after the most-negative LFP deflection to the local max-
imum within 100 ms after the most-negative LFP deflection since LFP sometimes exhibited a
bleed-through of the spike trace in the vicinity of the most negative excursion (Ray et al., 2008;
Ray and Maunsell, 2011; Nauhaus et al., 2009; Nauhaus et al., 2012). To avoid distortion by
this distinct wave, we chose to fit the LFP recovery from large-activity events at 30 ms after the
most-negative deflection.
Detection of bimodality in 2-6 Hz half-waves
For this analysis, LFP signals were further band-pass filtered (2-6 Hz) using a phase neutral
filter by applying a second order Butterworth filter in both directions in each 4s-interval. The
histogram of positive 2-6 Hz half-waves exhibited a clear bimodality in double-logarithmic
plot when the occasional brief large-amplitude activity events were not removed (dataset1; see
above; Fig. 1 c).
In dataset 1, to test the dependence of our results on the occurrence of these events, all
analyses were repeated for segments with or without this kind of large-amplitude slow wave
activity. Specifically, we determined whether the maximum of the 2-6 Hz filtered LFP signal in
each channel and 4-s segment was larger than a given threshold, and then separately analyzed
spike autocorrelation and nLFP recovery for units and LFP data based on whether there was an
event larger than this threshold or not. We set the threshold individually for each channel and
each 4-s segment as the median plus one standard deviation of the positive 2-6 Hz half-waves
(Fig. 1 c, vertical magenta line). This threshold reliably removed all large-amplitude events (i.e.
the bimodality in the LFP amplitude distribution) while maintaining a large-enough number of
11
data segments for analysis both above and below threshold.
In dataset 2, segments during quiet-wake (Q) were carefully chosen not to contain any of
these high-amplitude LFP events. This is evidenced by the 2-6 Hz LFP amplitude distribution
which did not exhibit a prominent bimodality (Fig. 3 d).
Computational model
To develop better insights into the determinants of timescale changes in cortical network ac-
itivty, we studied a neuron network model that could exhibit long timescales associated with
long-range temporal correlations. The neuron network model consisted of N=100 all-to-all
coupled, binary-state neurons with the following dynamical rules: If neuron j spiked at time t
(i.e., sj(t) = 1) as part of an active set of neurons J(t), a postsynaptic neuron i will spike at
time t + 1 according to
si(t + 1) = Θ[1 − Y
j∈J(t)
(1 − pij) − ζ(t)]
(4)
where Θ[x] is the unit step function and ζ(t) is a random number from a uniform distribution
on [0,1] to account for the probabilistic nature and variability of unitary synaptic efficacy. pij
are the asymmetric synaptic coupling strengths between each pair of neurons which are first
drawn from a uniform distribution on [0,1] and then multiplied by N · K/ P pij. The parameter
K is consequently related to the average connectivity and is therefore a control parameter for
the dynamics of the network: at K = 1 each spiking neuron excites, on average, exactly one
postsynaptic neuron meaning the network is critical. Conversely, at K < 1 activity dies out
prematurely and the system is subcritical; at K > 1 each neuron excites on average more than
on postsynaptic neuron and the system is supercritical (Shew et al., 2009). We instantiated the
model at different average connectivies K = 1 to study the spike autocorrelation and recovery
from large-activity events in similar ways as in the experimental data. Network activity was
12
introduced by randomly setting one neuron to active at each time step.
The temporal autocorrelation of neurons was studied by simulating network activity for
5050 timesteps. After the initial 50 timesteps during which network activity typically reached a
stable level, neuron activity was binned in 50-timestep bins, resulting in spike-count timeseries
of length 100 for each neuron. This process was repeated for 100 network simulation runs which
allowed the calculation of Pearsons correlation coefficients R (eq. 1) for different time bins and
across simulation runs analogously to the experimental data. The population autocorrelation
function was, also analogously to experimental data, derived by averaging across all neurons
and consequently fit by an exponential decay (eq. 2) to derive a recovery rate λ.
To study the recovery from large-activity network events relevant to our experimental anal-
ysis, the recovery from large, intrinsic fluctuations in network activity was monitored. We
simulated network activity for 5050 timesteps. Network activity typically reached a stable level
after about 50 timesteps. From the following 5000 timesteps, we identified large intrinsic fluc-
tuations in the spike histogram greater than a threshold T which we set to T = 2 standard
deviations. Once a large fluctuation of network activity, i.e. in the spike histogram, had been
identified, we required the following fluctuation event to be at least 50 timesteps apart. We av-
eraged over all fluctuation events from a total of 100 network simulation. We fit an exponential
decay (eq. 3) to these averages to derive a recovery rate λ.
To implement offline periods of spiking similar to those observed during sleep deprivation,
the activity of an individual neurons was set to zero with a probability pOFFLINE at each time
step during the simulation.
Results
We first assessed the timescales governing cortical dynamics at the individual neuron level in
three different vigilance states: REM sleep, NREM sleep and quiet wake. For each vigilance
13
state, the spike-count autocorrelation was calculated as the correlation coefficient between the
number of putative single-unit spikes in pairs of time bins across all artifact-free 4s-segments.
Following previous work (Ogawa and Komatsu, 2010; Murray et al., 2014), we derived an au-
tocorrelation function reflective of the population- or network-level activity in each animal by
averaging functions across units (Fig. 1 a). We observed slow autocorrelation decays indicative
of long-range temporal correlations during quiet wake in frontal cortex. While REM sleep ex-
hibited a similar, slowly decaying autocorrelation function, NREM sleep differed profoundly,
exhibiting a fast decaying autocorrelation function at significantly higher rate (ANOVA and post
hoc unpaired t-test; Fig. 1 a, b). To assure that these timescale differences were independent of
firing rates, we repeated the analyses using only neurons with average firing rates between 3 and
5 Hz in all vigilance states. While this selection abolished all firing rate differences between
vigilance states, the autocorrelation decay differences were maintained (Fig. 1 c; ANOVA,
F=29, p=4e-13).
Next, we extended our analysis to recordings from parietal cortex to investigate how univer-
sal the vigilance-dependent timescale differences were across cortex. No statistically significant
difference between frontal and parietal areas was observed when all quiet wake segments were
compared (frontal: 0.0017 ± 0.0002; parietal: 0.0015 ± 0.0002; p=0.39; unpaired t-test). In
absolute terms, these values correspond to timescales of about 400 ms, which is in the or-
der of magnitude also observed in areas of higher cortical hierarchy in non-human primates
(Murray et al., 2014). Similar to frontal cortex, we observed a much more rapid spike-count au-
tocorrelation decay during NREM sleep in comparison to quiet wake and REM sleep indicative
of a loss of long-range temporal correlations (Fig. 1 d -- f).
These results indicate distinct timescales in different vigilance states: long timescales asso-
ciated with a slow autocorrelation decay during quiet wake and REM sleep, and short timescales
during NREM sleep. We next investigated the persistence of timescales within vigilance states
14
and across time. Previous work has shown that wake is not a homogeneous state. Sponta-
neous waking is associated with more pronounced slow rhythms in the EEG (Finelli et al., 2000;
Strijkstra et al., 2003), changes in firing rates (Fisher et al., 2016) and the occurrence of local
OFF periods where neurons stop firing and go "offline" (Poulet and Petersen, 2008; Vyazovskiy et al., 2011;
McGinley et al., 2015). Some of these changes, like local offline periods, may not show in scalp
EEG (Vyazovskiy et al., 2011); the underlying state may thus still be considered wake in the
traditional sense. We therefore studied neuron timescale dynamics in wake by systematically
taking the effect of offline periods into consideration.
We first analyzed multi-unit activity of frontal cortex in freely-behaving rats during 6 hours
of sleep deprivation and after a consecutive recovery period (dataset 1). In these data, the spike
autocorrelation analysis revealed a progressive increase of decay rates during sleep deprivation
reflective of the faster autocorrelation decay (ANOVA, F=2.91, p=0.015; Fig. 2 a, b). Consecu-
tive sleep, conversely, recovered the slower rates. Post hoc statistical comparison of beginning
(SD0) to end of sleep deprivation (SD6), and end of sleep deprivation (SD6) to post recov-
ery sleep (REC) revealed both changes to be significant (p=0.002 and p=0.01, respectively;
two-tailed paired t-test). Similar to previous work (Vyazovskiy et al., 2011), we observed oc-
casional large positive LFP half-waves in the 2-6 Hz frequency band that were accompanied by
pauses in the MUA (Fig. 2 c). When plotted in double logarithmic coordinates, the amplitude of
positive half-waves in this frequency band exhibited a distinct bimodality in its density distribu-
tion (Fig. 2 c). Such a bimodality has also been observed in other frequency bands, where it was
taken as indication for the existence of two co-existing different states between which dynam-
ics can switch (Freyer et al., 2011). In our case, the observed bimodality may similarly suggest
two distinct states characterized by different low-frequency power levels. We hypothesized that
large-amplitude LFP events might correspond to a more sleep-like state and thus potentially im-
print shorter timescales. We thus repeated our analysis in select data with or without any such
15
events. Specifically, we defined a threshold for each animal and channel individually, that sep-
arated the two peaks in the LFP amplitude distribution (Fig. 2 c right, magenta line). We then
separately analyzed spike-count autocorrelations in all channels where amplitudes in the 2-6 Hz
frequency band either exceeded this threshold at least once or stayed below it throughout a 4-s
segment. The separation of sleep deprivation data along these two peaks in the LFP distribution
revealed that timescale changes were only apparent in segments that contained high-amplitude
LFP events. Timescales did not change during sleep deprivation in the low-amplitude regime
(ANOVA; Fig. 2 d, e). These observations indicate distinct dynamics with respect to the two
peaks in the LFP distribution. The intermittency of both dynamical states leads to the appar-
ent shortening of intrinsic timescales characterizing spontaneous fluctuations of neuron activity
during sustained wakefulness.
To confirm the causal role of intermittent large-amplitude LFP events for the apparent
timescale decline during extended wake, we repeated our analysis in dataset 2 where such
large-amplitude LFP events had been excluded altogether in the selection of quiet wake data
(as confirmed by the absence of a distinct bimodality in the LFP amplitude distribution; Fig.
2 f, blue). Consequently, autocorrelation decays obtained from quiet wake exhibited no sta-
tistically significant change during sleep deprivation in this dataset (frontal and parietal; SD0
compared to SD4; unpaired t-test; Fig. 2 g). During extended wake, the occurrence of shorter
timescales is thus linked to the presence of intermittent large-amplitude LFP events. Over the
course of NREM sleep, autocorrelation decays became significantly less steep in frontal cortex
(ANOVA, F=3.84, p=0.003; Fig. 2 i). A similar trend was visible in parietal cortex although
it did not reach significance level there. REM sleep, conversely, did not exhibit a significant
change in any region (Fig. 2 h). The differential changes observed over the course of NREM
sleep but not REM sleep thus provide further evidence for the relevance of low frequency large-
amplitude events as the essential mechanism for the apparent timescale change. Low frequency
16
large-amplitude events were present during NREM but not REM sleep (Fig. 2 f, brown vs.
green line), and are known to decrease over the course of NREM sleep (Borbely et al., 1981;
Finelli et al., 2000; Vyazovskiy et al., 2011) which can account for the less steep spike autocor-
relation decays observed.
There exists an intimate relationship between the autocorrelation function governing sponta-
neous fluctuations and the time it takes a signal to recover from small perturbations (Wissel, 1984;
Ives, 1995; Scheffer et al., 2009). From this dynamical systems perspective, a slow autocorre-
lation decay also implies a slow recovery from perturbations or intrinsic fluctuations. We hy-
pothesized that equivalent wake-time dependent changes quantified by autocorrelation decays
should also manifest in recovery rates from intrinsic fluctuations. We first identified sizeable
negative excursions from the mean in each LFP channel (Fig. 3 a). Next, we averaged across
all these LFP segments for each rat and at each time point (dataset 1) to quantify the return to
baseline activity during sleep deprivation and after consecutive recovery sleep. Negative LFP
averages typically coincided with increased spiking activity (Fig. 3 b, c). To quantify the re-
covery rate by which LFPs returned to baseline after a negative excursion, we fit the return
with an exponential function. Recovery rates obtained from these fits increased during sleep
deprivation and reversed to lower values after consecutive sleep (ANOVA, F=3.5, p=0.004; Fig.
3 d, e). Post hoc statistical comparison between the beginning (SD0) and the end of sleep
deprivation (SD6), and between the end of sleep deprivation (SD6) and after recovery sleep
(REC) revealed both changes to be significant (p=0.002 and p=0.01, respectively; two-tailed
paired t-test). We hypothesized that large-amplitude LFP events would again drive the observed
timescale decline. As with spike-count autocorrelation, we thus repeated our analysis for both
the low-amplitude and the high-amplitude regimes separately. In agreement with our obser-
vations from spike-count autocorrelations, timescales changes were only apparent in segments
also containing high-amplitude LFP events, while timescales did not change during sleep de-
17
privation in the low-amplitude regime (ANOVA; Fig. 3 f, g).
Together, results indicated that intermittent periods of large-amplitude LFP can lead to an
apparent decline of long cortical timescales during wake. Our observations also suggested
that these LFP events are characterized by pauses in neuron spiking. To test whether these
intermittent offline periods were sufficient to account for timescale changes, we studied a neu-
ron network model exhibiting long timescales associated with long-range temporal correlations
analogous to experiment. Specifically, the network's average connectivity was set so that net-
work dynamics was located right at the transition between an inactive and an active network
phase (Shew et al., 2009). At this transition state, the network exhibited long timescales associ-
ated with both the autocorrelation decay and the recovery from intrinsic fluctuations. We then
implemented the possibility that neurons go "offline" for a short amount of time with a certain
probability. We observed that this ability to go offline was sufficient to reproduce our exper-
imental findings during wake. In particular, with higher probability for neurons to go offline,
faster declines in spike-count autocorrelation and recovery from intrinsic fluctuations were ob-
served (Fig. 4). These modeling results thus suggest that an increasing tendency for neurons
to go offline by themselves (e.g. during sleep deprivation) or in synchrony (e.g. during NREM
sleep) is sufficient to account for the shortened timescales.
Discussion
The ability to integrate information over time is a pre-requisite for brain function during wake.
Information integration is believed to be supported by long timescales associated with long-
range temporal correlations in neuron spiking activity. We here provide first-time evidence at the
neuronal level that these timescales are abrogated during NREM sleep. In contrast, quiet wake
and REM sleep are characterized by similar, long timescales. Our results provide indication for
the existence of distinct states in terms of timescale dynamics: one, which is characterized by
18
long timescales that dominate during wake and REM sleep, and a second one characterized by
the absence of long timescales predominantly observed in NREM sleep. Our sleep deprivation
results suggest that transitions between these two states can occur rapidly and intermittently
which, on average, can lead to an apparent progressive decline of timescales.
The present work provides a missing link between long cortical timescales (Ogawa and Komatsu, 2010;
Honey et al., 2012; Stephens et al., 2013; Murray et al., 2014; Chaudhuri et al., 2015) to obser-
vations indicating disrupted long-range correlations during NREM sleep (Massimini et al., 2005;
Tagliazucchi et al., 2013). The observed breakdown of long timescales in ongoing spiking ac-
tivity here provides insights that complements these reports based on large-scale signals mea-
sured by EEG and fMRI. From a theory point of view, the timescales quantified by the autocor-
relation of intrinsic fluctuations are intimately linked to the timescales that control the response
to external perturbations (Wissel, 1984; Ives, 1995; Scheffer et al., 2009; Meisel et al., 2015).
The fast autocorrelation decay observed in our data during NREM sleep thus links to observa-
tions of reduced stimulation responses using TMS in the same vigilance state (Massimini et al., 2005).
In detailed analyses of the apparent gradual timescale decline during extended wake, we
identified intermittent transitions between two different states. We found that the state charac-
terized by a fast decaying autocorrelation increasingly dominated the overall dynamics, while
autocorrelation decay rates did not change for the low-amplitude state during sleep depriva-
tion. The two peaks visible in the LFP amplitude distribution thus point to distinct dynamical
states, governed by long timescales in the low amplitude regime, and shorter timescales in
the high amplitude regime. The identification of two different states being at play relied on
the ability to record data at high temporal and spatial resolution in our data. It is therefore
possible that the gradual decline in timescales observed during the transition into NREM sleep
(Tagliazucchi et al., 2013) is similarly the result of two different co-existing states which a more
coarse method such as fMRI may not be able to resolve in the temporal domain. Several sub-
19
cortical structures in the brainstem, hypothalamus and basal forebrain are known to regulate
the maintenance of waking and sleep states through neuro-modulatory action (Jones, 2005). It
is conceivable that these subcortical structures control the transient occurrences of the large-
amplitude LFP epochs observed in some of our data during wake.
Functionally, the shorter timescales during NREM sleep indicate a severely reduced ability
to integrate information. During sleep deprivation, shorter timescales induced by intermittent,
sleep-like, high-amplitude LFP events may be similarly detrimental for cortical networks to pro-
cess information. As such, this decline may provide a perspective on several cognitive capabili-
ties known to decrease during extended wake (Van Dongen et al., 2003; Banks and Dinges, 2007;
Mignot, 2008; Killgore, 2010; Lo et al., 2012). In fact, a two-state dynamics, as observed here,
may help to explain some at times controversial findings with respect to the specifics perfor-
mance impairments. Although it is well-established that performance, for example quantified
in the psychomotor vigilance test (PVT), decreases, on average, as a function of time awake,
a subject may still perform almost normally even after extended periods of sleep deprivation
during individual trials in the PVT (Banks and Dinges, 2007).
It is conceivable that normal
performance could still be achieved at times when dynamics is not compromised by intermit-
tent high-amplitude events, while performance will decrease on average due to the growing
dominance of intermittent epochs with disrupted timescales during sleep deprivation.
Long-range temporal correlations are generic features that arise when a system is in the
vicinity of a critical state (Bak et al., 1987). Mounting evidence from computational and ex-
perimental studies indicates that cortical neural networks operate at some sort of critical state
(Bornholdt and Rohlf, 2000; Beggs and Plenz, 2003; Haldeman and Beggs, 2005; Shew et al., 2009;
Meisel et al., 2012; Markram et al., 2015). The observation of long timescales during wake is
in agreement with this criticality hypothesis (Linkenkaer-Hansen et al., 2001). Conversely, fast
decaying autocorrelation functions such as during NREM sleep and also during sleep depriva-
20
tion may be seen as indication for a dynamical state where critical dynamics is disrupted. In a
previous study based on EEG, we reported fading signatures of critical dynamics during sleep
deprivation (Meisel et al., 2013). Our results here demonstrate an increasing disruption of long
timescales during sleep deprivation at the individual neuron level which complements these
earlier findings. From this perspective, our observations support an intriguing hypothesis for
a function of sleep, to re-organize cortical networks towards critical dynamics with long-range
temporal correlations for optimal functioning during wake.
References
Bak P, Tang C, Wiesenfeld K (1987) Self-organized criticality: An explanation of the 1/f
noise. Phys Rev Lett 59:381 -- 384.
Banks S, Dinges DF (2007) Behavioral and physiological consequences of sleep restriction. J
Clin Sleep Med 3:519 -- 528.
Beggs JM, Plenz D (2003)
Neuronal avalanches in neocortical circuits.
J Neu-
rosci 23:11167 -- 11177.
Borbely AA, Baumann F, Brandeis D, Strauch I, Lehmann D (1981) Sleep deprivation: ef-
fect on sleep stages and EEG power density in man. Electroencephalogr Clin Neurophys-
iol 51:483 -- 495.
Bornholdt S, Rohlf T (2000) Topological evolution of dynamical networks: Global criticality
from local dynamics. Phys Rev Lett 84:6114 -- 6117.
Chaudhuri R, Knoblauch K, Gariel MA, Kennedy H, Wang XJ (2015) A Large-Scale
Circuit Mechanism for Hierarchical Dynamical Processing in the Primate Cortex. Neu-
ron 88:419 -- 431.
21
Finelli LA, Bauman H, Borb´ely AA, Achermann P (2000) Dual electroencephalogram mark-
ers of human sleep homeostasis: correlation between theta activity in waking and slow-wave
activity in sleep. Neuroscience 101:523 -- 529.
Fisher SP, Cui N, McKillop LE, Gemignani J, Bannerman DM, Oliver PL, Peirson SN, Vya-
zovskiy VV (2016) Stereotypic wheel running decreases cortical activity in mice. Nat Com-
mun 7:13138.
Freyer F, Roberts JA, Becker R, Robinson PA, Ritter P, Breakspear M (2011) Biophysical
mechanisms of multistability in resting-state cortical rhythms. J. Neurosci. 31:6353 -- 6361.
Friston K, Breakspear M, Deco G (2012) Perception and self-organized instability. Front
Comput Neurosci 6:44.
Haldeman C, Beggs J (2005) Critical branching captures activity in living neural networks and
maximizes the number of metastable states. Phys Rev Lett 94:058101.
Honey CJ, Thesen T, Donner TH, Silbert LJ, Carlson CE, Devinsky O, Doyle WK, Rubin N,
Heeger DJ, Hasson U (2012) Slow cortical dynamics and the accumulation of information
over long timescales. Neuron 76:423 -- 434.
Ives AR (1995) Measuring resilience in stochastic systems. Ecol Monogr 65:217233.
Jones BE (2005) From waking to sleeping: neuronal and chemical substrates. Trends Phar-
macol. Sci. 26:578 -- 586.
Kiebel SJ, Daunizeau J, Friston KJ (2008) A hierarchy of time-scales and the brain. PLoS
Comput. Biol. 4:e1000209.
Killgore WD (2010) Effects of sleep deprivation on cognition. Prog. Brain Res. 185:105 -- 129.
22
Kringelbach ML, McIntosh AR, Ritter P, Jirsa VK, Deco G (2015) The Rediscovery of Slow-
ness: Exploring the Timing of Cognition. Trends Cogn. Sci. (Regul. Ed.) 19:616 -- 628.
Linkenkaer-Hansen K, Nikouline VV, Palva JM, Ilmoniemi RJ (2001) Long-range temporal
correlations and scaling behavior in human brain oscillations. J Neurosci 21:13701377.
Lo JC, Groeger JA, Santhi N, Arbon EL, Lazar AS, Hasan S, von Schantz M, Archer SN, Dijk
DJ (2012) Effects of partial and acute total sleep deprivation on performance across cognitive
domains, individuals and circadian phase. PLoS ONE 7:e45987.
Markram H, Muller E, Ramaswamy S, Reimann MW, Abdellah M, Sanchez CA, Ailamaki
A, Alonso-Nanclares L, Antille N, Arsever S, Kahou GA, Berger TK, Bilgili A, Buncic N,
Chalimourda A, Chindemi G, Courcol JD, Delalondre F, Delattre V, Druckmann S, Dumusc R,
Dynes J, Eilemann S, Gal E, Gevaert ME, Ghobril JP, Gidon A, Graham JW, Gupta A, Haenel
V, Hay E, Heinis T, Hernando JB, Hines M, Kanari L, Keller D, Kenyon J, Khazen G, Kim Y,
King JG, Kisvarday Z, Kumbhar P, Lasserre S, Le Be JV, Magalhaes BR, Merchan-Perez A,
Meystre J, Morrice BR, Muller J, Munoz-Cespedes A, Muralidhar S, Muthurasa K, Nachbaur
D, Newton TH, Nolte M, Ovcharenko A, Palacios J, Pastor L, Perin R, Ranjan R, Riachi I,
Rodriguez JR, Riquelme JL, Rossert C, Sfyrakis K, Shi Y, Shillcock JC, Silberberg G, Silva
R, Tauheed F, Telefont M, Toledo-Rodriguez M, Trankler T, Van Geit W, Diaz JV, Walker R,
Wang Y, Zaninetta SM, DeFelipe J, Hill SL, Segev I, Schurmann F (2015) Reconstruction and
Simulation of Neocortical Microcircuitry. Cell 163:456 -- 492.
Massimini M, Ferrarelli F, Huber R, Esser SK, Singh H, Tononi G (2005) Breakdown of
cortical effective connectivity during sleep. Science 309:2228 -- 2232.
23
McGinley MJ, Vinck M, Reimer J, Batista-Brito R, Zagha E, Cadwell CR, Tolias AS, Cardin
JA, McCormick DA (2015) Waking State: Rapid Variations Modulate Neural and Behavioral
Responses. Neuron 87:1143 -- 1161.
Meisel C, Klaus A, Kuehn C, Plenz D (2015) Critical slowing down governs the transition to
neuron spiking. PLoS Comput. Biol. 11:e1004097.
Meisel C, Olbrich E, Shriki O, Achermann P (2013) Fading signatures of critical brain dy-
namics during sustained wakefulness in humans. J. Neurosci. 33:17363 -- 17372.
Meisel C, Storch A, Hallmeyer-Elgner S, Bullmore E, Gross T (2012) Failure of adaptive
self-organized criticality during epileptic seizure attacks. PLoS Comput Biol 8:e1002312.
Mignot E (2008) Why we sleep: the temporal organization of recovery. PLoS Biol. 6:e106.
Murray JD, Bernacchia A, Freedman DJ, Romo R, Wallis JD, Cai X, Padoa-Schioppa C,
Pasternak T, Seo H, Lee D, Wang XJ (2014) A hierarchy of intrinsic timescales across primate
cortex. Nat. Neurosci. 17:1661 -- 1663.
Nauhaus I, Busse L, Carandini M, Ringach DL (2009) Stimulus contrast modulates functional
connectivity in visual cortex. Nat. Neurosci. 12:70 -- 76.
Nauhaus I, Busse L, Ringach DL, Carandini M (2012) Robustness of traveling waves in
ongoing activity of visual cortex. J. Neurosci. 32:3088 -- 3094.
Ogawa T, Komatsu H (2010) Differential temporal storage capacity in the baseline activity of
neurons in macaque frontal eye field and area V4. J. Neurophysiol. 103:2433 -- 2445.
Poulet JF, Petersen CC (2008) Internal brain state regulates membrane potential synchrony in
barrel cortex of behaving mice. Nature 454:881 -- 885.
24
Ray S, Hsiao SS, Crone NE, Franaszczuk PJ, Niebur E (2008) Effect of stimulus intensity
on the spike-local field potential relationship in the secondary somatosensory cortex. J. Neu-
rosci. 28:7334 -- 7343.
Ray S, Maunsell JH (2011) Different origins of gamma rhythm and high-gamma activity in
macaque visual cortex. PLoS Biol. 9:e1000610.
Scheffer M, Bascompte J, Brock WA, Brovkin V, Carpenter SR, Dakos V, Held H, van Nes EH,
Rietkerk M, Sugihara G (2009) Early-warning signals for critical transitions. Nature 461:5359.
Shew WL, Yang H, Petermann T, Roy R, Plenz D (2009) Neuronal avalanches imply maximum
dynamic range in cortical networks at criticality. J Neurosci 9:15595 -- 15600.
Shoham S, Fellows MR, Normann RA (2003) Robust, automatic spike sorting using mixtures
of multivariate t-distributions. J. Neurosci. Methods 127:111 -- 122.
Stephens GJ, Honey CJ, Hasson U (2013) A place for time: the spatiotemporal structure of
neural dynamics during natural audition. J. Neurophysiol. 110:2019 -- 2026.
Strijkstra AM, Beersma DGM, Drayer B, Halbesma N, Daan S (2003) Subjective sleepiness
correlates negatively with global alpha (8-12 Hz) and positively with frontal theta (4-8 Hz)
frequencies in the human resting awake electroencenphalogram. Neurosci Lett 340:17 -- 20.
Tagliazucchi E, von Wegner F, Morzelewski A, Brodbeck V, Jahnke K, Laufs H (2013) Break-
down of long-range temporal dependence in default mode and attention networks during deep
sleep. Proc. Natl. Acad. Sci. U.S.A. 110:15419 -- 15424.
Tononi G (2008) Consciousness as integrated information: a provisional manifesto. Biol.
Bull. 215:216 -- 242.
25
Van Dongen HP, Maislin G, Mullington JM, Dinges DF, Van Dongen HP, Mullington JM,
Dinges DF (2003) The cumulative cost of additional wakefulness: dose-response effects on
neurobehavioral functions and sleep physiology from chronic sleep restriction and total sleep
deprivation. Sleep 26:117 -- 126.
Vyazovskiy VV, Olcese U, Hanlon EC, Nir Y, Cirelli C, Tononi G (2011) Local sleep in awake
rats. Nature 472:443 -- 447.
Wissel C (1984) A universal law of the characteristic return time near thresholds. Oecolo-
gia 65:101 -- 107.
26
Acknowledgements
This study was supported by the Intramural Research Program of the NIMH. This study utilized
the high-performance computational capabilities of the Biowulf Linux cluster at the National
Institutes of Health, Bethesda, Md.
27
Figure 1: Breakdown of long-range temporal spiking correlations during NREM sleep. a,
Spike-count autocorrelation (AC) decays rapidly in NREM sleep while other vigilance states
have similar slow decays. Mean AC over animals (n=13; s.e.m.; single-unit activity; frontal
cortex; quiet wake, Q, REM sleep, R, NREM sleep N). b, Summary of decay rates. Decay
rates obtained from exponential fits over the whole autocorrelation function. c, Changes in
autocorrelation decays are independent of differences in spiking rate between vigilance states.
d -- f, Breakdown of long-range temporal spiking correlation in NREM sleep in parietal cortex.
Equivalent analysis as in a -- c.
Figure 2: Two-state dynamics underlies the decline in long-range temporal correlations during
sleep deprivation. a, Decay in spike-count autocorrelation (AC) accelerates as a function of time
awake and recovers after sleep (multi-unit activity; quiet wake; 6 hours of sleep deprivation
SD0 to SD6 followed by 5 -- 6 hr recovery sleep, REC). Mean AC over all multi-units (n=26;
1 rat; s.e.m.). Solid lines: exponential fit. Left: linear plot. Right: double-logarithmic plot.
SD3 line slightly shifted upwards for visibility. b, Summary of decline in decay rates and
recovery for all n=6 animals (s.e.m). c, Identification of two-state dynamics underlying decay
in long-range temporal correlations with sleep deprivation. Large-amplitude LFP events are
intermittently observed and typically coincide with pauses in multi-unit activity (top; offline
periods marked as gray areas in zoomed inset). These LFP events typically present as one
or more positive half-waves predominantly in the frequency range 2-6 Hz (bottom, wavelet
spectrogram and positive LFP trace bandpass filtered between 2-6 Hz). Right: Bimodality
in the histogram of positive 2-6 Hz waves. The bimodality was used to separate the large-
amplitude LFP containing putative offline periods from other activity for each rat individually
(magenta line). d, e, Decline in long-range temporal correlations is restricted to data segments
containing large-amplitude LFP events. f, Histogram of positive LFP 2-6 Hz waves sorted by
vigilance states (dataset 2; top row frontal, bottom row parietal). g, Autocorrelation decays
remains unchanged (right) in data without large-amplitude LFP events during sleep deprivation
(quiet wake; 4 hours of sleep deprivation SD0, SD4). h, Autocorrelation decay rates during
REM sleep remain unchanged (dataset 2; 6 hour recovery sleep; R1-R6).
i, Autocorrelation
decay rates in NREM sleep decrease over the 6 hour recovery sleep (dataset 2; N1-N6).
28
Figure 3: Two-state dynamics underlies the accelerated recovery from large-activity events
during sleep deprivation. a, Identification of large-activity events from LFP. Negative LFP
(nLFP; ¡-2 standard deviations) deflections were averaged over each time point during the sleep
deprivation experiment (SD0 to SD6, REC) and for each rat. b, Mean nLFP over all events and
channels (1 rat; s.e.m.). c, Corresponding nLFP-triggered spike histogram. The histogram is
normalized to mean activity from -150 to -100 ms. d, Recovery from intrinsic nLFP excursions
accelerates as a function of time awake and recovers after sleep. Mean nLFP traces from the
beginning and end of sleep deprivation from one rat. e, Summary of nLFP recovery rates for all
n=7 animals (s.e.m.). f, g, More rapid recovery from large-activity events is restricted to data
segments containing large-amplitude LFP events.
Figure 4: Neuron network model with the ability of neurons to go offline captures the differ-
ential decline of long-range temporal correlations. The network was instantiated at a connec-
tivity K=1 and with different probablities pOFFLINE for neurons to go offline for a brief amount
of time. a, Raster plot without the presence of offline periods. b, Raster plot with offline
periods(pOFFLINE = 0.1, grey shaded areas). c, Autocorrelation functions for spiking activity
in a network without offline periods (pOFFLINE = 0, blue) and in a network with offline pe-
riods (pOFFLINE = 0.3, red) in neuron activity. d, Autocorrelation decay rates increase with
higher probability for offline periods. e, Recovery from intrinsic fluctuations (pOFFLINE = 0,
blue; pOFFLINE = 0.3, red). f, Recovery rates from intrinsic fluctuations increase with higher
probability for offline periods.
29
|
1411.7635 | 3 | 1411 | 2015-10-23T13:13:20 | A statistical model for in vivo neuronal dynamics | [
"q-bio.NC"
] | Single neuron models have a long tradition in computational neuroscience. Detailed biophysical models such as the Hodgkin-Huxley model as well as simplified neuron models such as the class of integrate-and-fire models relate the input current to the membrane potential of the neuron. Those types of models have been extensively fitted to in vitro data where the input current is controlled. Those models are however of little use when it comes to characterize intracellular in vivo recordings since the input to the neuron is not known. Here we propose a novel single neuron model that characterizes the statistical properties of in vivo recordings. More specifically, we propose a stochastic process where the subthreshold membrane potential follows a Gaussian process and the spike emission intensity depends nonlinearly on the membrane potential as well as the spiking history. We first show that the model has a rich dynamical repertoire since it can capture arbitrary subthreshold autocovariance functions, firing-rate adaptations as well as arbitrary shapes of the action potential. We then show that this model can be efficiently fitted to data without overfitting. Finally, we show that this model can be used to characterize and therefore precisely compare various intracellular in vivo recordings from different animals and experimental conditions. | q-bio.NC | q-bio |
1
A statistical model for in vivo neuronal dynamics
Simone Carlo Surace1,2,∗, Jean-Pascal Pfister2
1 Department of Physiology, University of Bern, Bern, Switzerland
2 Institute of Neuroinformatics, University of Zurich and ETH Zurich, Zurich, Switzerland
∗ E-mail: [email protected]
Abstract
Single neuron models have a long tradition in computational neuroscience. Detailed biophysical models
such as the Hodgkin-Huxley model as well as simplified neuron models such as the class of integrate-and-
fire models relate the input current to the membrane potential of the neuron. Those types of models have
been extensively fitted to in vitro data where the input current is controlled. Those models are however
of little use when it comes to characterize intracellular in vivo recordings since the input to the neuron is
not known. Here we propose a novel single neuron model that characterizes the statistical properties of in
vivo recordings. More specifically, we propose a stochastic process where the subthreshold membrane po-
tential follows a Gaussian process and the spike emission intensity depends nonlinearly on the membrane
potential as well as the spiking history. We first show that the model has a rich dynamical repertoire
since it can capture arbitrary subthreshold autocovariance functions, firing-rate adaptations as well as
arbitrary shapes of the action potential. We then show that this model can be efficiently fitted to data
without overfitting. Finally, we show that this model can be used to characterize and therefore precisely
compare various intracellular in vivo recordings from different animals and experimental conditions.
1
Introduction
During the last decade, there has been an increasing number of studies providing intracellular in vivo
recordings. From the first intracellular recordings performed in awake cats [1, 2] to more recent recording
in cats [3], monkeys [4], mice [5], and even in freely behaving rats [6], it has been shown that the
membrane potential displays large fluctuations and is very rarely at the resting potential. Some recent
findings in the cat visual cortex have also suggested that the statistical properties of spontaneous activity
is comparable to the neuronal dynamics when the animal is exposed to natural images [7]. Similar results
have been found in extracellular recordings in the ferret [8]. Those data are typically characterized by
simple quantifications such as the firing rate or the mean subthreshold membrane potential [5], but a more
comprehensive quantification is often missing. So the increasing amount of intracellular data of awake
animals as well as the need to compare in a rigorous way the data under various recording conditions call
for a model of spontaneous activity in single neurons.
Single neuron models have been studied for more than a century. Simple models such as the integrate-
and-fire model [9, 10] and its more recent nonlinear versions [11, 12, 13] describe the relationship between
the input current and the membrane potential in terms of a small number of parameters and are therefore
convenient for analytical treatment, but do not provide much insight about the underlying biophysical
processes. On the other end of the spectrum, biophysical models such as the Hodgkin-Huxley model
[14, 15] relate the input current to the membrane potential through a detailed description of the various
transmembrane ion channels, but estimating the model parameters remains challenging [16, 17]. Despite
the success of those types of models, none of them can be directly applied to intracellular in vivo recordings
for the simple reason that the input current is not known.
Another reason why a precise model of spontaneous activity is needed is that there are several theories
that have been proposed that critically depend on statistical properties of spontaneous activity. For ex-
2
ample Berkes et al. validate their Bayesian treatment of the visual system by comparing the spontaneous
activity and the averaged evoked activity [8]. Another Bayesian theory proposed the idea that short-term
plasticity acts as a Bayesian estimator of the presynaptic membrane potential [18]. To validate this the-
ory, it is also necessary to characterize spontaneous activity with a statistical model that describes the
subthreshold as well as the spiking dynamics.
The last motivation for a model that describes both the subthreshold and the suprathreshold dynamics
is the possibility to separate those two dynamics in a principled way. Indeed, it is interesting to know
from the recordings what reflects the input dynamics and what aspect comes from the neuron itself (or
rather what is associated with the spiking dynamics). Of course a simple voltage threshold can separate
the sub- and a suprathreshold dynamics, but the value of the threshold is somewhat arbitrary and could
lead to undesirable artifacts. Therefore a computationally sound model that decides itself what belongs
to the subthreshold and what belongs to the suprathreshold dynamics is highly desirable.
Here, we propose a single neuron model that describes intracellular in vivo recordings as a sum of a
sub- and suprathreshold dynamics. This model is flexible enough in order to capture the large diversity
of neuronal dynamics while remaining tractable, i.e. the model can be fitted to data in a reasonable time.
More precisely, we propose a stochastic process where the subthreshold membrane potential follows a
Gaussian process and the firing intensity is expressed as a non-linear function of the membrane potential.
Since we further include refractoriness and adaptation mechanisms, our model, which we call the Adaptive
Gaussian Point Emission process (AGAPE), can be seen as an extension of both the log Gaussian Cox
process [19] and the generalized linear model [20, 21, 22].
2 Results
Here we present a statistical model of the subthreshold membrane potential and firing pattern of a single
neuron in vivo. See Fig. 1A for such an in vivo membrane potential recording. We first provide a formal
definition of the model and then show a range of different results. 1) The model is flexible and supports
arbitrary autocorrelation structures and adaptation kernels. Therefore, the range of possible statistical
features is very large. 2) The model is efficiently fittable and the learning procedure is validated on
synthetic data. 3) The model can be fitted to in vivo datasets. 4) All the features included in the model
are required to provide a good description of in vivo data.
2.1 Definition of the AGAPE model
The AGAPE model is a single neuron model where the input to the neuron is not known, which is
typically the case under in vivo conditions. The acronym AGAPE stands for Adaptive GAussian Point
Emission process since the subthreshold membrane potential follows a Gaussian process and since the
spike emission process is adaptive.
membrane voltage trace usom(t) and the spike train s(t) =Pns
More formally, the AGAPE process defines a probability distribution p(usom, s) over the somatic
i=1 δ(t − ti) where ti, i = 1, ..., ns are the
nominal spike times (decision times), which occur a certain fixed time period δ > 0 before the peak
of the action potential. From this probability distribution (or generative model) it is possible to draw
samples that look like intracellular in vivo activity (for practical purposes, the samples will be compared
to the preprocessed recordings, see explanations below). The AGAPE model assumes that the somatic
membrane voltage as a function of time usom(t) is given by (see Fig. 1)
usom(t) = ur + u(t) + uspike(t),
(1)
where ur is a constant (the reference potential), u(t) describes the subthreshold membrane potential as
a stochastic function drawn from a stationary Gaussian process (GP) [23]
u ∼ GP [0, k(t − t′)]
(2)
A
V
m
0
1
100 ms
B
3
u
usom
s
Figure 1. (A) A sample in vivo membrane potential trace from an intracellular recording of a neuron
in HVC of a Zebra Finch. (B) The generative AGAPE model can generate a trace of subthreshold
membrane potential u (top trace). Based on this potential, a spike train s is generated (middle, dashed
vertical lines). Finally, a stereotypic spike-related kernel is convolved with the spike train and added to
u, giving rise to usom (bottom, thick line). This quantity is the synthetic analog of the recorded,
preprocessed in vivo membrane potential.
with covariance function k(t − t′) (which can be parametrized by a weighted sum of exponential decays
with weights σ2
i and inverse time constants θi, see Materials and Methods). For small values of δ (e.g.
1-3 ms), u(t) can be seen as the net contribution from the unobserved synaptic inputs and uspike(t) is
the spike-related contribution (see Fig. 1B) which consists of the causal convolution of a stereotypical
spike-related kernel α with the spike train s(t), i.e.
uspike(t) =Z ∞
0
α(t′)s(t − t′)dt′.
(3)
where α can be parametrized by a weighted sum of basis functions with weights ai, see Materials and
Methods. Here, we have made a separation of subthreshold and suprathreshold layers, in that whatever
is stereotypic and triggered by the point-like spikes s(t) is attributed to uspike(t), and the rest belongs to
the fluctuating signal u(t). This separation need not correspond to the biophysical distinction between
synaptic inputs and active processes of the recorded cell (i.e. the positive feedback loop of the spiking
Indeed, especially for a choice of large δ (e.g. ∼ 20 ms), uspike(t) also contains large
mechanism).
depolarizations due to strong synaptic input which cannot be explained by the GP signal u(t).
Note that this model could easily be extended by including an additional term in Eq. (1) which
depends on an external input, e.g. a linear filter of the input (see also Discussion). However, since this
input current was not accessible in our recordings, its contribution was assumed to be part of u(t) or
uspike(t).
Now we proceed to the coupling between the subthreshold potential u(t) and the spiking output, as well
as adaptive effects associated with spike generation. These effects are summarized by an instantaneous
firing rate r(t) -- as in the generalized linear model (GLM) [20, 21, 22] or escape-rate models [24] -- which
is computed from the value of the subthreshold membrane potential at time t, u(t), and the spike history
as
where β ≥ 0 is the coupling strength between u and the spikes, and A(t) is the adaptation variable which is
the convolution of an adaptation kernel η (which can be parametrized by a weighted sum of basis functions
with weights wi, see Materials and Methods) with the past spike train. Also note that we choose not
r(t) = g [A(t) + βu(t)] , A(t) =Z ∞
0
η(t′)s(t − t′)dt′,
(4)
4
to model adaptation currents explicitly, since they would simultaneously impact the membrane potential
and the firing probability (see Discussion). The function g is called gain function, and here we use an
exponential one, i.e. g [A(t) + βu(t)] = elog r0+A(t)+βu(t). Other functional forms such as rectified linear
or sigmoidal could be used depending on the structure of the data. However, this choice has important
implications on the efficiency of learning of the model parameters [25]. We define the probability density
for s on an interval [0, T ] conditioned on u as
p(su) = exp −Z T
0
r(t) dt! nsYi=1
r(ti),
s(t) =
nsXi=1
δ(t − ti),
0 ≤ t1 < ... < tns ≤ T, ns ∈ N0.
(5)
The parameter β connects the subthreshold membrane potential u to the rate fluctuations. The magnitude
of the rate fluctuations depend on the variance σ2 of u, and therefore we use βσ as a measure of the
effective coupling strength. When β > 0 the quantity θ(t) = −A(t)/β can be regarded as a soft threshold
variable which is modulated after a spike, and u(t) − θ(t) is the effective membrane potential relevant
for the spike generation. This spiking process is a point process which generalizes the log Gaussian Cox
process. Indeed, when A = 0, Eq. (5) describes an inhomogeneous Poisson process with rate g [βu(t)].
Practically, if we want to draw a sample from the AGAPE process, we first draw a sample u from
the Gaussian Process (see S1 Text for how to do this efficiently), then for each time t we draw spikes s(t)
with probability density r(t) and update the adaptation variable A(t). Finally, the somatic membrane
potential is calculated using Eq. 1.
It is important to emphasize at this point that while the model may be directly fitted to the raw
membrane potential uraw as recorded by an intracellular electrode, we median filter the data in order to
avoid artifacts and downsample for computational efficiency (see 'Materials and Methods'). In this study
the model is always fitted to the preprocessed recordings u∗som and this is reflected e.g. in the shape of
α which is most strongly affected by the preprocessing. It is important to keep in mind this point while
interpreting the results of model fitting. The details of the preprocessing steps which were used are given
in the 'Materials and Methods' section.
2.2 The model has a rich dynamical repertoire
The AGAPE provides a flexible framework which can be adjusted in complexity to model a wide range
of dynamics. While for the datasets presented here a covariance function was used which consists of
a sum of Ornstein-Uhlenbeck (OU) kernels, the Gaussian Process (GP) allows for arbitrary covariance
functions to be used. This includes simple exponential decay (as produced by a leaky integrate-and-fire
neuron driven by white noise current), but it can produce also more interesting covariance functions such
as power-law covariances, which are reported in [26, 7], or subthreshold oscillations, as reported in [27].
The model is also able to reproduce a wide range of firing statistics. A common measure of firing
irregularity is the coefficient of variation (CV , i.e. the ratio of standard deviation and mean) of the inter-
spike interval distribution. In the absence of adaptation, the AGAPE is a Cox process and therefore has
a coefficient of variation CV ≥ 1 [28]. The precise value of the CV is a function of the coupling strength
(βσ) as well as the autocorrelation of the GP. To illustrate this, we sampled synthetic data from a simple
version of the AGAPE where the subthreshold potential u is an OU process with time-constant τ . As
shown in Fig. 2A, the CV is an increasing function of the membrane time-constant τ , baseline firing rate
r0, and dimensionless coupling parameter between membrane potential and firing rate βσ. Moreover, the
range of the CV extends from 1 to ≈ 8 within a range of βσ ∈ [0, 2] and r0τ ∈ [2−2, 28]. In the presence
of adaptation, firing statistics are markedly different and can produce values of CV < 1 [24, 29]. To
illustrate this point, we considered an exponential adaptation kernel, i.e. η(t) = −η0e−t/τr . While the
CV increases as a function of βσ and r0τ as before, the range of values of the CV now also covers the
interval (0, 1) which is not accessible by the Cox process but which is observed in many neurons across
the brain [30]. In order to study the influence of the parameters of the adaptation mechanism, we fix
5
βσ = r0τ = 1 and plot the CV as a function of r0τr and η0 (see Fig. 2B). Within the parameter region
explored in Fig. 2B, the CV spans values from 0.1 up to 1.6.
2.3 The model can be learned efficiently
The parameters of the AGAPE model are learned through a maximum likelihood approach. More pre-
cisely, we fit the model to an in vivo sample (highlighted by a '∗') of preprocessed somatic membrane
potential u∗som and spike train s∗,δ by maximizing the log likelihood applied to the joint data set (u∗som, s∗,δ)
over the parameter space of the model (i.e. ur, log r0, β, the coefficients of the kernels k, η, and α, and
the delay parameter δ). The empirical spike train s∗,δ depends on the parameter δ because the formal
spike times ti are assigned to be a time period δ before the recorded peak of the action potential. The
joint probability of the data can be expressed as a product
p(u∗som, s∗,δ) =Z p(u)p(s∗,δu)p(u∗somu, s∗,δ)Du
=Z p(u)p(s∗,δu)δ(u∗som − u − ur − uspike)Du
= p(u = u∗som − ur − uspike)p(s∗,δu = u∗som − ur − uspike)
≡ ps∗,δ (u∗som)p(s∗,δu∗som).
(6)
The subscript s∗,δ of the first factor denotes the explicit dependence on the spike train. The individual
terms on the r.h.s. will be given below. The function we are optimizing is the logarithm of the above
joint probability which we can write as
L(ur, k, α, log r0, β, η, δ) = log ps∗,δ (u∗som; ur, k, α) + log p(s∗,δu∗som; ur, α, log r0, β, η).
(7)
It should be noted that the presence of the spike-related kernel α in both terms produces a trade-off
situation: removing the spike-related trajectory improves the Gaussianity of the membrane potential
u (and therefore boosts the first term) at the cost of the of the second term by removing the short
upward fluctuation that leads to the spike. This trade-off situation makes maximum likelihood parameter
estimation a non-concave optimization problem. Moreover, the evaluation of the GP likelihood of n
samples, where n = O(105), comes at a high computational cost. Two important techniques make the
parameter learning both tractable and fast: the first is the use of the circulant approximation of the GP
covariance matrix which makes the evaluation of the likelihood function fast. The second is the use of an
alternating fitting algorithm which (under an appropriate parametrization, see 'Materials and Methods')
replaces the non-concave optimization in the full parameter space with two concave optimizations and a
non-concave one in suitable parameter subspaces. Those two techniques are further described in the next
section.
2.3.1 Efficient likelihood computation
The log-likelihood function is evaluated in its discrete-time form with n time points separated by a time-
step ∆t. The GP variable u (which leads to usom through Eq. (1)) is multivariate Gaussian distributed
with a covariance matrix Kij = k(ti − tj), where ti = i∆t. The matrix K is symmetric and, by virtue of
stationarity, Toeplitz. Evaluation of the GP likelihood requires inversion of K, which is computationally
expensive (the time required to invert a matrix typically scales with n3). For this reason we approximate
this Toeplitz matrix by the circulant matrix C which minimizes the Kullback-Leibler divergence (see
[31, 32, 33] and S1 Text5)
C = argmin
D circulant
DKL [N (m, K)N (m, D)]
(8)
6
E
α(t)
V
m
3
0
4
ms
10
t [ms]
0
5
10
15
20
δ [ms]
F
ground truth
synth. data
η(t)
ML fit
0
t [ms]
1
20
ms
C
-0.63
n
b
i
r
e
p
d
o
o
h
i
l
e
k
i
l
g
o
l
-0.65
-0.67
-0.69
D
20
10
2
1
]
2
V
m
[
)
t
(
k
1
10
100
1000
t [ms]
2
A
σ
β
1
0
1
100
10
r0τ
B
10
0
η
−
8
6
4
2
0
1
10
r0τr
100
8
7
6
5
4
3
2
1
V
C
1.6
1.4
1.2
1
.8
.6
.4
.2
0
V
C
Figure 2. The model has a rich dynamical repertoire (A,B) and can be correctly fitted to synthetic
data (C-F). (A,B) The coefficient of variation (CV ) of the inter-spike interval distribution is computed
for parameter values shown as black dots and then linearly interpolated. (A) The CV of a simple
version of the AGAPE (k(t) = σ2e−t/τ , α = η = 0) as a function of the model parameters (membrane
time-constant τ , baseline firing rate r0 and coupling strength βσ). (B) CV of the AGAPE model with
an exponentially adaptive process with fixed membrane time-constant, firing rate and coupling
(βσ = r0τ = 1) as a function of the parameters describing adaptation, namely adaptation strength η0
and time-constant τr. (C,D,E,F) Synthetic data is sampled from the AGAPE model with GP (D),
spike-related (E), and adaptation (F) kernels as depicted in black, and δ = 4 ms, r0 = 4.15 Hz,
β = 0.374 mV−1. Then the AGAPE is fitted to the synthetic data by maximum likelihood (ML). (C)
The maximum log likelihood per bin as a function of the parameter δ has its maximum at the ground
truth value δ = 4 ms. (D,E,F) The ML estimates (red) of the GP, spike-related and adaptation kernels
lie within two standard deviations (red shaded regions, estimated by means of the observed Fisher
information) from the ground truth.
7
between the two multivariate Gaussian distributions with the same mean but different covariance matrices.
This optimization problem can be solved by calculating the derivative of DKL [N (m, K)N (m, D)] with
respect to D and using the diagonalization of D by a Fourier transform matrix [33]. After a bit of
algebra (see S1 Text5), denoting ki = K1i and kn+1 ≡ 0, the optimal circulant matrix can be written as
Cij = c(i−j mod n)+1, where i, j = 1, ..., n and
ci =
1
n
[(n − i + 1)ki + (i − 1)kn−i+2] .
(9)
The replacement of K by C is equivalent to having periodic boundary conditions on u, which has a small
effect under the assumption that the time interval spanned by the data is much longer than the largest
temporal autocorrelation length of k. So the first term on the r.h.s. of Eq. (6) is a multivariate Gaussian
density N (0, C). The determinant of the covariance matrix C is the product of eigenvalues, which for a
circulant matrix are conveniently given by the entries of c, the discrete Fourier transform of c (see the S1
Text5 for our conventions regarding discrete Fourier transforms). Also the scalar product uT C−1u can
be written in terms of c. Together, the first term on the r.h.s. of Eq. (6) takes the simple form
log ps∗,δ (u∗som) = −
1
2
nXi=1 log(2πci) +
ui2
ci ! ,
1
n
(10)
where ui are the components of the discrete Fourier transform of u∗. The Gaussian component of the
membrane potential u is implicitly given by the discretized somatic voltage modified by a discrete-time
version of the spike-related kernel convolution,
u∗i = u∗som,i − ur −
i−1Xj=1
αjs∗,δ
i−j,
(11)
i
where s∗,δ
is the binned spike train (see below), αi is a discretized version of the spike-related kernel. The
time required to compute log ps(u∗som) is determined by the complexity of the Fourier transform, which is
of the order of n log n. This dramatic reduction in complexity (compared to n3) allows a fast evaluation
of the log-likelihood.
The spiking distribution p(s∗,δu∗som) is approximated by a Poisson distribution with constant rate
counts the number of spikes that occur in that bin, and the
within one time bin. For each bin, s∗,δ
conditional likelihood of the spikes therefore reads
i
log p(s∗,δu∗som) =
nXi=1ns∗,δ
i
log [ri∆t] − ri∆t − loghs∗,δ
i
!io ,
(12)
where ri = g[βu∗i +Pi−1
i−j] and u∗i as defined in Eq. (11). If s∗,δ
can be accomplished given small enough bins), the last term log s∗,δ
j=1 ηjs∗,δ
contains only zeros and ones (which
! vanishes.
i
i
2.3.2 Efficient parameter estimation
Except for the parameter δ, which takes discrete values of multiples of the discretization step ∆t, it is
possible to analytically calculate the first and second partial derivatives of the likelihood function defined
in Eq. (6) with respect to the model parameters (ur, k, α, log r0, β, η) (see S1 Text5) to facilitate the
use of gradient ascent, Newton, and pseudo-Newton optimization algorithms. A desirable feature of an
optimization problem is concavity of the objective function (in our case, the log-likelihood function).
Even though the problem of finding optimal parameters for the AGAPE process is not concave, the
optimization can be done in three alternating subspaces (see Fig. 3). The full set of parameters Θ is
8
divided into three parts: θGP for the GP parameters (ur, parameters of k), θspike kernel for the spike-related
kernel parameters, and θspiking for the parameters controlling spike emission (log r0, β and parameters of
η). The optimization then proceeds according to the following cycle: (1) the GP parameters are learned,
(2) the spike-related kernel parameters are learned, and lastly (3) the spiking parameters are learned. In
each step the remaining parameters are held fixed. The cycle is repeated until the parameters reach a
region where the log likelihood is locally concave in the full parameter space, after which the optimization
can be run in the full parameter space until it converges. Joint concavity of the log likelihood holds if
all the eigenvalues of the Hessian matrix are strictly negative. As shown in [25], step (3) is concave for a
certain class of gain functions g, including the exponential function, and linear parametrizations of the
adaptation kernel. The same holds for the spiking term of the log-likelihood in step (2). The voltage
term of the log likelihood of step (2) is concave by numerical inspection in the cases we considered. To
summarize, steps (2) and (3) are concave and Newton's method can be used in these steps as well as for
the final concave optimization in the full space. Step (1) is non-concave and therefore a simple gradient
ascent algorithm is used.
The optimization over (ur, k, α, log r0, β, η) is repeated for every δ = 0, ∆t, 2∆t, ..., δmax in order to
select the one δ that maximizes the log-likelihood L(ur, k, α, log r0, β, η, δ). The value of δmax is chosen
such that it is less than the least upper bound of the support of the basis of the spike-related kernel α.
Since the parameters ur, k, α, log r0, β, η are expected to change only a little when going from one δ to
the next, δ + ∆t, learned parameters for δ can be used as initial guesses for nearby δ + ∆t or δ − ∆t. We
thus get two different initializations, which we can exploit by starting e.g. with δ = 0, ascending through
the sequence of candidate δ's up to the maximum δ, and descending back to zero.
non-concave
concave
Optimize over
θGP = (ur, k)
Optimize over
θspike shape = (α)
concave
Optimize over
θspiking =
(log r0, β, η)
neg.
Check concavity
pos.
Optimize over
Θ =(ur, k, α,
log r0, β, η)
Figure 3. This schematic shows the optimization scheme that is used to learn the parameters of the
AGAPE model when it is fitted to the data (for a given δ). As long as the current parameter estimate
sits in a non-concave region of the likelihood function, the top cycle optimizes over different subspaces
of the parameter space. If and when a concave point is reached, the optimization proceeds in the full
parameter space. This whole scheme is repeated for each value of δ in order to find the optimal one.
2.4 Validation with synthetic data
Despite this improvement in speed and tractability, the optimization is still riddled with multiple local
minima which require the use of multiple random initializations. In order to demonstrate the validity of
9
the fitting method, synthetic data of length 270.112 seconds (n = 270112, the same as in vivo dataset
D1, see below) was generated with known parameters (δ = 4 ms, r0 = 4.15 Hz, β = 0.374 mV−1 and
GP, spike-related kernel and adaptation kernels as depicted in Fig. 2D-F). The learning algorithm was
initialized with least-squares estimates of the covariance function parameters σ2
i based on the empirical
autocorrelation function of usom and spike-related kernel and adaptation kernels set to zero. The true
underlying δ can be recovered from the synthetic data (Fig. 2C). Moreover, the algorithm converges after
a few dozen iterations (taking only three minutes on an ordinary portable computer) and -- with δ set
to 4 ms -- recovers the correct GP, spike-related, and adaptation kernels (Fig. 2D-F). All ML estimates
lie within a region of two standard deviations around the ground truth, where standard deviations are
estimated from the observed Fisher information [34].
2.5 The model can fit in vivo data
We fitted the model to a number of in vivo traces from different animals and conditions (see 'Materials
and Methods' for a detailed description of the data sets). We would like to remind the reader at this point
that the model is never fitted to the raw membrane potential, but to a preprocessed, i.e. median-filtered
and downsampled dataset (see Materials and Methods). Because of this preprocessing stage, the model
only sees the truncated action potentials which emerge from the median filter. This is reflected in the
extracted spike-related kernel α, which is characterized by a smaller amplitude than the original action
potential in the raw membrane potential data.
We show the detailed results of the model fitting for the example songbird HVC dataset D1. The
optimal value of δ for this dataset was δ = 18 ms (see S3 Fig), with which the model captures the
subthreshold and suprathreshold statistics (smaller values of δ compromise both the subthreshold and
suprathreshold description because the large upward fluctuations which preceed spikes in this dataset are
unlikely to arise from a GP). In particular, the stationary distribution of the membrane potential u is
well approximated by a Gaussian (Fig. 4B) and pronounced after-hyperpolarization is seen in the spike-
related kernel (Fig. 4D). The subthreshold autocorrelation structure is well reproduced by the parametric
autocorrelation function k (Fig. 4C). The adaptation kernel reveals an interesting structure in the way
the spiking statistics deviates from a Poisson process (Fig. 4E). This feature of the spiking statistics is
also reflected in the inter-spike interval (ISI) distribution (Fig. 4F). Both the data and the fitted model
first show an increased, and then a significantly decreased probability density when compared to a pure
Poisson process. The remaining parameters are listed in Tab. 1 (errors denote two standard deviations,
estimated from Fisher information, see Materials and Methods). The model can be used to generate
synthetic data, which is shown in Fig. 4H.
In order to show the generality of the model, we fitted the model on two more datasets, D3 from another
HVC neuron and D4 from mouse visual cortex. The parameter δ was found to take the optimal value of 12
ms for D3 and 32 ms for D4 (to see how fitted parameters change as a function of δ, see S4 Fig and S5 Fig).
The comparison of the GP, spike-related and adaptation kernels is shown in Fig. 5, and the remaining
parameters are listed in Tab. 1. The three cells show pronounced differences in autocorrelation structure,
In particular the two datasets D1 and D4 show rather
spike-related kernel and spike-history effects.
long autocorrelation lengths of the membrane potential and asymmetric spike-related kernels, whereas
the cell in D3 has comparatively short autocorrelation length and very pronounced hyperpolarization.
Adaptation is much stronger in D3 than in D1, balancing the much higher baseline firing rate r0, see
Tab. 1. The error bars on the adaptation kernel are small for datasets D1 and D3 due to the abundance
of spikes. On the other hand, the adaptation kernel of dataset D4 is poorly constrained by the available
data. This is due to the fact that dataset D4 consists of very short trials with very few spikes. Despite
this fact, good agreement is achieved between the distribution of inter-spike intervals of the in vivo data
and ISI statistics sampled from the AGAPE (see Fig. 5, bottom row) for all datasets.
10
data
A = 0
A 6= 0
A
V
m
5
50
ms
uraw
u + ur
C
data
ML fit
10
5
1
]
2
V
m
[
)
t
(
k
B
)
u
(
p
D
10
5
0
]
V
m
[
)
t
(
α
STA
ML fit
F
data
ML fit
E
]
1
[
)
t
(
η
-
30
0
18
50
t [ms]
)
τ
(
ρ
ML fit
1
0
-
1
-
2
-
10
mV
0
10
u [mV]
1
10
100
1000
0
50
100
0
100
t [ms]
t [ms]
200
τ [ms]
300
400
G
data
simulation
H
V
m
0
1
150 ms
Figure 4. The results of maximum likelihood (ML) parameter fitting to dataset D1. After fitting, we
see (A) the removal of the spike-related kernel through the difference between the recorded trace u∗som
and the subthreshold membrane potential u + ur; (B) the match of the stationary distribution of the
subthreshold potential u and a Gaussian. We also observe that (C) the autocorrelation function of the
data, Eq. (14), is well reproduced by k(t) in Eq. (15); (D) the spike-related kernel α(t) starts at
−δ = −18 ms relative to the peak of the action potential. The difference between the spike-triggered
average (STA) and the spike-related kernel is attributed to the GP; and (E) that the adaptation kernel
η(t) shows distinct modulation of firing rate which produces firing statistics significantly different from
a Poisson process. This is also reflected in the inter-spike interval density ρ(τ ) (F) of the data, which
shows good qualitative agreement with a simulated AGAPE with adaptive kernel as in (E) (thick red
line), but not by a non-adaptive (i.e. Poisson) process (thin red line). After fitting, a two second sample
of synthetic data (H) looks similar as the in vivo data (G). In (G,H) vertical lines are drawn at the
spiking times. All red shaded regions denote ± 2 standard deviations, estimated from the observed
Fisher information.
11
D4
10
102
103
D3
10
102
103
D1
15
10
5
]
2
V
m
[
)
t
(
k
0
1
10
102
103
]
V
m
[
)
t
(
α
10
5
0
-
5
15
10
5
0
1
10
5
0
-
5
15
10
5
0
1
10
5
0
-
5
-
50
-
18
0
50
-
50
-
12
0
50
-
50
-
32
0
50
200
100
0
-
100
-
200
-
300
50
100
0
50
100
0
-
2
-
4
-
6
-
8
-
10
0
50
100
0.5
0.0
-
0.5
-
1.0
]
1
[
)
t
(
η
-
1.5
-
2.0
0
)
t
(
ρ
0
100
200
300
400
0
t [ms]
100
t [ms]
200
0
1000
2000
3000
t [ms]
Figure 5. Fitting results for three different datasets. Dataset D1 is the same as in Fig. 4, i.e. an HVC
neuron from anesthetized Zebra Finch. D3 is from HVC in awake Zebra Finch, and D4 is from mouse
visual cortex in awake mouse. The different panels show the results after fitting; in the first line the GP
covariance function k(t) (red) and the empirical autocorrelation (black), Eq. (14), in the second line the
spike-related kernel α(t), in the third line the adaptation kernel η(t), and in the fourth line the
inter-spike interval density ρ(t) (data ISI histogram in gray, simulated ISI distribution from AGAPE in
red). There are pronounced differences between datasets in all three kernels, showing the flexibility of
the AGAPE model in describing a wide range of statistics. All red shaded regions denote ± 2 standard
deviations, estimated from the observed Fisher information.
12
Dataset
D1
D3
D4
δ [ms]
ur [mV]
r0 [Hz]
β [mV−1]
18
-52.9±0.2
11.7±0.6
0.12±0.01
12
-66.6±0.1
71±5
0.24±0.02
32
-51.5±0.5
0.15±0.05
0.46±0.05
βσ [1]
0.45±0.03
0.67±0.04
1.3±0.1
Table 1. The values (p.m. two standard deviations, estimated from the observed Fisher information)
of the fitted parameters not shown in Fig. 5 for the in vivo datasets described in the main text. The
last row shows the effective coupling strength between the membrane potential and the firing rate, given
by β times the standard deviation σ of the membrane potential.
2.6 The model does not overfit in vivo data
The AGAPE process has a fairly large number of parameters. Therefore it is important to check whether
the model overfits the data, compromising its generalization performance. In short, when a model has
too many parameters, it tends to be poorly constrained by the data and therefore when the model is
first trained on one part of the data and then tested on another part on which it is not trained, the test
performance will be significantly worse than the training performance.
Here, we use cross-validation to perform a factorial model comparison on an exemplary dataset in
order to validate the different structural parts of the model. The procedure is described in detail in the
Materials and Methods.
Model comparison is performed on the dataset D2 and the results are shown in Fig. 6, where the
mean difference of per-bin log-likelihood (see 'Materials and Methods')
∆pvalid
i
= h∆pvalid
ij
ij , ∆ptest
i = h∆ptest
ij ij ,
∆pvalid,test
ij
= pvalid,test
i,j
− pvalid,test
Gαβη,j
(13)
is shown for all models i ∈ {0, ..., Gαβη} (here, h·ij denote averages over chunks j of the cross-validation).
The results are very similar for both validation data (which was left-out during training, but appeared in
other training runs) and the test data which was never seen during training. The most complex model
(MGαβη) performs significantly better than any one of the simpler models on validation data except
MGαη where the difference is too small and lies inside a region of two standard errors of the mean.
This confirms that most of the model features are required to provide an accurate description of the
experimental data.
3 Discussion
In this study, we introduced the AGAPE generative model for single-neuron statistics in order to describe
the spontaneous dynamics of the somatic potential without reference to an input current. We showed
that this model has a rich dynamical repertoire and can be fitted to data efficiently. By fitting a het-
erogeneous set of data, we finally demonstrated that the AGAPE model can be used for the systematic
characterization and comparison of in vivo intracellular recordings.
3.1 Flexibility and tractability of the model
The AGAPE model provides a unified description of intracellular dynamics, offering a large degree of
flexibility in accounting for the distinct statistical features of a neuron. As the example datasets demon-
strate, the model readily teases apart the differences in the statistics which exist between different cells
13
models
®
r
e
t
t
e
b
e
c
n
a
m
r
o
f
r
e
p
l
e
d
o
m
.
l
e
r
e
s
r
o
w
¬
0.
-0.1
-0.2
-1.2
-1.3
-1.4
-1.5
Η
Β
Α
G
MMMMMMMM
Β
Α
G
Η
Α
G
Α
G
Η
Β
Α
Η
Α
Α
Β
Α
Dpvalid
Dptest
G: multiple OUs
Α: spike shape
Β: coupling
Η: adaptation
Α
h
t
i
w
s
l
e
d
o
m
Α
t
u
o
h
t
i
w
Β
G
Η
Β
G
Η
Β
Β
MMMMM
models
Η
G
Η G 0
MMM
Figure 6. Comparison of the different models on dataset D2. The relative measure of model
performance, i.e. the per-bin log-likelihood ∆p (see Eq. (13)) between any model and the most complex
model (MGαβη) are significantly negative (with exception of MGαη, and trivially MGαβη) , implying
that the added complexity improves the model fit without overfitting. This holds for both validation
scores ∆pvalid (black) and scores from unseen test data ∆ptest (red). Error bars denote one standard
error of the mean (S.E.M.). The biggest improvement of fit quality is achieved by including the
spike-related kernel (upper vs. lower part of the figure).
14
in different animals (see Fig. 5). This shows that the model is sensitive enough to distinguish between
datasets which are in fact very similar.
We used a set of approximations and techniques to make the model fitting tractable, despite the non-
concavity of the log likelihood function. It is still the case that multiple local maxima of the likelihood
function can make the fitting somewhat hard, especially if the quantity of data available for fitting is quite
low. However, since one run of the fitting itself takes only a few minutes even on a portable computer,
multiple initializations can be tried out in a relatively short amount of time.
3.2 Comparison with existing models
From an operational perspective, existing spiking neuron models can be divided into three main categories:
stimulus-driven, current-driven and input-free spiking neurons. The first category contains phenomenolog-
ical models that relate sensory stimuli to the spiking output of the neuron. The linear-nonlinear-Poisson
model (LNP) [35], the generalized linear model (GLM) [20, 21, 22] or the GLM with additional latent
variables [36] are typical examples in this category. Even though the spike generation of the AGAPE
shares some similarities with those models, there is an important distinction to make. In those models
the convolved input (i.e. the output of the 'L' step of the LNP or the input filter of the GLM) is an
internal variable that does not need to be mapped to the somatic membrane potential whereas in our
case, the detailed modeling of the membrane potential dynamics is an important part of the AGAPE.
Consequently, those phenomenological models are descriptions of extracellular spiking recordings whereas
the AGAPE models the dynamics of the full membrane potential accessible with intracellular methods.
The second class of spiking models aims at bridging the gap between the input current and the spiking
output. The rather simple integrate-and-fire types of models such as the exponential integrate-and-fire
[13] or the spike-response model [24, 37] as well as the more biophysical models such as the Hodgkin-
Huxley model [15] fall within this category.
In contrast to those models where the action potentials
are caused by the input current, the AGAPE produces a fluctuating membrane potential and stochastic
spikes without a reference to an input current.
The last category of models aims at producing spontaneous spiking activity without an explicit de-
pendence to a given input [38, 18, 39]. For example, Cunningham et al. propose a doubly stochastic
process where the spiking generation is given by a gamma interval process and the firing intensity by
a rectified Gaussian process, which provides a flexible description of the firing statistics [38]. However,
the membrane potential dynamics is not modeled.
In opposition, the neuronal dynamics assumed by
Pfister et al.
[18] models explicitly the membrane potential (as a simple Ornstein-Uhlenbeck process)
but is not flexible enough to capture the dynamics of in vivo recordings. Also any of the current-driven
spiking neuron models mentioned above can be turned into an input-independent model by assuming
some additional input noise. So why is there a need to go beyond stochastic versions of those models?
An integrate-and-fire model with additive Gaussian white noise is certainly fittable, but does not have
the flexibility to model arbitrary autocorrelation for the membrane potential. At the other end of the
spectrum, a Hodgkin-Huxley model with some colored noise would certainly be able to model a richer
dynamical repertoire, but the fitting of it remains challenging [16] (but see [17]). The main advantage of
the AGAPE is that it is at the same time very flexible and easily fittable. The flexibility mostly comes
from the fact that any covariance function can be assumed for the GP process. The relative ease of
fitting comes from the circulant approximation as well as from the presence of concave subspaces in the
full parameter space.
Another distinct feature of our model with respect to other existing models is the explicit modeling
of the spike-related trajectory instead of the spike-triggered average (as e.g. in [40]). Even though both
concepts share similarities - both would capture a sudden and strong input that lead to a spike - there
is an important distinction. The spike-triggered average also captures the (possibly smaller) upward
fluctuations of the membrane potential which causes the spike while the spike-related kernel α precisely
avoids capturing those fluctuations, letting the GP kernel explain them.
15
So if we removed the spike-triggered average e.g. in synthetic data where the true coupling parameter
β is large, we would also remove the characteristic upward fluctuation of the membrane potential which
causes the spike. By doing so, the fitting procedure would not find the correct relation between the
values of the membrane potential and the observed spike patterns and therefore choose a β close to
zero. Thus, if something has to be removed around an action potential (and our model comparison,
Fig. 6, demonstrates convincingly that this is necessary), the formulation of the model demands that it
is parametrically adjustable. This is the main reason why in our model framework the spike-triggered
average has to be rejected as a viable extraction method. Note that if the true coupling parameter β is
close to zero, the spike-triggered average is close to the extracted spike-related kernel α. For data where
the action potential shape shows considerable variability, the model could be generalized to include a
stochastic or a history-dependent spike-related kernel.
3.3 Extensions and future directions
Despite the focus of the present work on single-neuron spontaneous dynamics, the AGAPE model admits
a straightforward inclusion of both stimulus-driven input and recurrent input. The inclusion of stimulus-
driven input is similar as for the GLM model and allows the model to capture the neuronal correlate
of stimulus-specific computation. The recurrent input makes the framework adaptable to multi-neuron
recordings in vivo. While intracellular recordings from many neurons in vivo are very hard to perform,
the rapid development of new recording techniques (e.g. voltage-sensitive dyes) makes the future avail-
ability of subthreshold data with sufficient time-resolution at least conceivable. The full-fledged model
would allow questions regarding the relative importance of background activity, recurrent activity due to
computation in the circuit, and activity directly evoked by sensory stimuli to be answered in a systematic
way. In this setup, the contribution of the GP-distributed membrane potential to the overall fluctuations
would be reduced (since it has to capture less unrecorded neurons) while the contribution of the recorded
neurons would increase. This modified model can be seen as a generalization of the stochastic spike-
response model [24] or a generalization of the GLM (if the internal variable of the GLM is interpreted as
the membrane potential).
So far, we assumed that weak synaptic inputs are captured by the Gaussian process while the strong
inputs that lead to the spikes are captured by the spike-related kernel α. A straightforward extension
of the model would be to consider additional intermediate inputs that cannot be captured by the GP
nor by the spike-related kernel α but that can drive the neuron to emit (with a given probability) an
action potential. Those intermediate input could be modeled as filtered Poisson events. The inclusion
of those latent events would increase the complexity of the model and at the same time change some
of the fitted parameters.
In particular, we expect that it would increase the coupling β between the
membrane potential and the firing rate and reduce the optimal delay δ between the decision time and the
peak of the action potential. This could also provide a better way to separate the subthreshold dynamics
(which depends on the input activity) from the suprathreshold dynamics (which would depend only on
the neuron dynamics, and not on the strong inputs that it receives, as it is the case now).
A central assumption of our model is that of a Gaussian marginal distribution of the subthreshold
potential. Although it is remarkably valid for the dataset considered here (i.e. the HVC dataset D1 see
also Fig. 4B), datasets characterized by a distinctly non-Gaussian voltage distribution even after spike-
related kernel removal are beyond the scope of the current model. In order to address this limitation, the
Gaussian process could be extended to a different stochastic process, e.g. a nonlinear diffusion process,
permitting non-Gaussian and in fact arbitrary marginal distributions. Moreover, a reset behavior similar
to the one exhibited by an integrate-and-fire model [13] could be achieved with a non-stationary GP
which features a mean which is reset after a spike. Both modifications would have a severe impacton
the technical difficulty of model fitting. Therefore, the Gaussian assumption can be regarded as a useful
compromise which is preferable over a perfect account for the skewness of the marginal distribution.
The spike-related kernel method to separate subthreshold and suprathreshold dynamics is an impor-
16
tant feature of the model which is used to rid the membrane potential recording of stereotypic waveforms
associated with a spike. The spike related kernel as modeled in the AGAPE has no bearing on the
probability of the spikes, whereas the adaptation kernel η which modulates the firing rate after a spike is
not visible in the somatic membrane potential dynamics. A simple extension of the model could include
spike-triggered adaptation currents which affect both the somatic membrane potential as well as the firing
intensity. Another possible extension is to allow the firing probability to depend on a filtered version of
the subthreshold potential u instead of the instantaneous value of u at a time δ before the peak of the
action potential. Both of the mentioned extensions would improve the biophysical interpretability of the
AGAPE, but they would also vastly increase the number of parameters. Therefore, a model comparison
would be required to determine what level of model complexity is required in order to characterize the
statistics of the recording.
In the present study, the AGAPE was fit to different datasets of two different animals and brain
regions. A systematic fitting to in vivo intracellular data from a wide range of animals and brain regions
would constitute a classification scheme which does not only complement existing classifications of neurons
which are based on electrophysiological, morphological, histological, and biochemical data; such as the
one in [41], but which is in direct relationship with the computational tasks the brain is facing in vivo.
Another application of the AGAPE could be in the context of a normative theory of short-term plas-
ticity. Indeed, it has been recently hypothesized that short-term plasticity performs Bayesian inference
of the presynaptic membrane potential based on the observed spike-timing [18, 42]. According to this
theory, short-term plasticity properties have to match the in vivo statistics of the presynaptic neuron.
Since the AGAPE provides a realistic generative model of presynaptic activity under which inference is
supposedly performed, our model can be used to make testable predictions on the dynamical properties
of downstream synapses.
4 Materials and Methods
4.1 Description of the datasets used
1. Dataset D1 is a recording from a HVC neuron of an anesthetized Zebra Finch (Ondracek and
Hahnloser, unpublished recordings). The recording has a total length of 270 seconds at 32 kHz (see
Fig. 1A for a snippet of this recording) and contains 2281 action potentials.
2. Dataset D2 is another recording from a projection cell in HVC of Zebra Finch, but this time the
animal is awake (Vallentin and Long, unpublished recordings). It consists of 6 individual recordings
which together have a length of 152.5 seconds at 40 kHz. This dataset is used for model comparison
(see below).
3. Dataset D3 is from similar conditions as D2 (Vallentin and Long, unpublished recordings, see
[43, 44, 45] for similar recordings) and has a length of 60 seconds.
4. Dataset D4 consists of 19 individual trials of 4.95s duration at 20 kHz. The recording was obtained
from a pyramidal neuron in layer 2/3 of awake mouse visual cortex [46].
4.2 Preprocessing
Intracellular voltage traces are often recorded at a rate between 20 and 40 kHz. This allows the action
potentials to be resolved very clearly and precise spike timings to be extracted. However, for the study
of the subthreshold regime, this high sampling rate is not required, and therefore the data may be down-
sampled to roughly 1 kHz after obtaining the precise spike timings. Prior to down-sampling, we smooth
with a median filter of 1ms width in order to truncate the sharp action potential peaks and avoid artifacts
(see details below).
17
We define the spike peak times ti,peak operationally as the time where the local maximum of the
action potential is reached. This means that ti,peak occurs after action potential onset, and hence the
spike-related kernel has to extend to the past of ti,peak. The spike-related kernel starts at the nominal
spike time ti which is shifted from the peak time by a fixed amount δ, i.e. ti,peak = ti + δ. The nominal
spike times ti are then binned to 1 ms, yielding a binary spike train si = 0, 1.
For usom(t) we use a preprocessed version of the recorded trace which has been median-filtered with
a width of the filter of 1 ms and then down-sampled to 1 kHz, making it the same length as the binary
spike train. This procedure preserves the relevant correlation structure of the membrane potential while
reducing the computational demands of fitting as much as possible. In the data we examined, the median-
filtered membrane potential has a dip after ti,peak, but unless down-sampling is done carefully, this dip
sometimes occurs one timestep after ti,peak and sometimes right at ti,peak in the downsampled usom.
Since this dip will have to be captured by the spike-related kernel which has a fixed shape for all action
potentials, the down-sampling procedure has to ensure that the dip occurs always in the first time-step.
We solved this problem by setting the down-sampled value of usom at ti,peak (rounded to 1 ms) to the
value of usom at ti,peak before down-sampling.
While applying the model to the raw recording uraw directly (without first filtering and downsampling
it) is possible in principle, it comes at a massively increased computational cost. In the interest of time
required to fit the model and amount of data having to be handled, it is therefore sensible to include that
pre-processing stage.
4.3 Parametrizations and initializations
We already introduced the parameters ur, r0 and β. Additional parameters are needed to describe the
autocorrelation k(t), the spike-related kernel α(t) and the adaptation kernel η(t).
The covariance function of the GP has to be parametrized such that it can explain the autocorrelation
structure of the data. Therefore, an initial examination of the empirical autocovariance of usom, i.e.
kemp(j∆t) =
1
n − j − 1
n−jXi=1 usom,i −
1
n − j
n−jXk=1
usom,k! usom,i+j −
1
n − j
usom,k+j! ,
n−jXk=1
(14)
for j = 0, ..., jmax, is done in order to determine a suitable basis. Here, we used a sum of Ornstein-
Uhlenbeck (OU) kernels, i.e.
k(t) =
σ2
i e−θit,
(15)
nkXi=1
nkXj=1
where nk = 10 and θi = 2−i ms−1. The autocovariance has to remain positive definite. This induces the
following linear constraints:
ci =
j c(j)
σ2
i > 0,
∀i = 1, ..., n,
(16)
i
i , where c(j)
on σ2
are the discrete Fourier transforms of the circulant basis vectors. The optimization
problem is non-concave in the subspace of σ2
i and multiple local maxima and saddle points can occur.
Therefore, multiple initializations have to be made in order to find a potential global optimum. In general,
the least-squares fit of k(t) to the empirical autocovariance function (14) yields a good starting point for
the optimization.
The spike-rate adaptation kernel is chosen to be a linear combination of ten different alpha shapes
η(t) =(Pnη
0,
i=1 wi [exp(−νit) − exp(−ωit)] ,
t > 0,
t ≤ 0,
(17)
18
where we chose nη = 10, νi = 2ωi and νi = 2−i ms−1.
Since the median filter time constant is short, the voltage change around the spike can be fast, requiring
flexible spike-related kernel basis. Most of this flexibility is required around t = δ. Because δ is adapted,
we choose a discrete parametrization which has equal flexibility from t = 0 up to a maximum t. In our
case, this maximum is at t = 60 ms, and therefore our parametrization of the spike-related kernel reads
α(t) =(ai,
0,
if
else
t ∈ [i∆t, (i + 1)∆t)
(18)
where ai ∈ R, i = 1, ..., 60 are the free parameters. Since the spike-related kernel fitting is concave, the
large number of parameters does not lead to a dramatic increase of computational time. It also does not
lead to overfitting, as is evidenced by the smoothness of the fitted kernels (see updated Figs.4,5 in the
main text and the new S3-S5) and by the new model comparison results (see updated Fig.6 in the main
text).
4.4 Model validation
We performed a factorial model comparison (see Fig. 6) where the four factors were the presence/absence
of each of the following: multiple OU components in the GP autocorrelation function (see Eq. (15),
as opposed to only one OU kernel with variable time-constant), the spike-related kernel α, coupling
between u and s (through β) and adaptation η, which gives a total of 16 different models. We use the
nomenclature that M0 is the simplest model, e.g. α = β = η = 0 and only one OU component, having
only four parameters (ur, θ, σ and r0). A subscript G (for GP) indicates that we use the multiple OU
basis and any other subscript indicates that the corresponding parameter is adjustable in addition to the
parameters already present in M0 and the parameters that are associated with the subscribed ones. E.g.
MGα indicates that we use the multiple OU basis and allow a non-zero spike-related kernel and that
there are now 73 parameters (δ, ur, θi, ai for i = 1, ..., 60, and log r0). The parameter δ is optimized only
for the 12 out of 16 models which depend on this parameter, i.e. that have at least β 6= 0 or α 6= 0.
For each of the models M ∈ {M0, ..., MGαβη}, we performed eight-fold cross-validation [47] on dataset
D2 in order to assess the models' generalization performance. The entire dataset was cut into eight equally-
sized chunks dj , where j = 1, ..., 8, each of length 15s (n = 15000), and six chunks of 3s d′j , j = 1, ..., 6
set aside as a test set (n′ = 3000). Each model was then trained on seven out of eight chunks (treating
them as independent samples) giving an optimal set of parameters Θi
j = argmaxΘ p({dk, k 6= j} Mi, Θ)
and training per-bin log-likelihood ptrain
7n log p({dk, k 6= j} Mi, Θi
j). Then the validation likelihood
ij = 1
j) of the left-out chunk #j was evaluated. The unseen data d′j is used for a
pvalid
final benchmark of model performance, where the best set of parameters is selected for each model, i.e.
ij = 1
ptest
n′ maxk=1,...,8 log p(d′jMi, Θi
n log p(djMi, Θi
ij = 1
k).
5 Acknowledgments
We would like to thank Janie Ondracek and Richard Hahnloser in Zurich, Switzerland, Daniela Vallentin
and Michael Long at NYU, Bilal Haider and Matteo Carandini at UCL for kindly providing the data for
this study. We also thank M´at´e Lengyel, Christian Pozzorini and Johanni Brea for helpful discussions.
19
References
1. Woody CD, Gruen E. Characterization of electrophysiological properties of intracellularly recorded
neurons in the neocortex of awake cats: a comparison of the response to injected current in spike
overshoot and undershoot neurons. Brain Research; 1978;158(2):343 -- 357.
2. Baranyi A, Szente MB, Woody CD. Electrophysiological characterization of different types of
neurons recorded in vivo in the motor cortex of the cat. ii. membrane parameters, action poten-
tials, current-induced voltage responses and electrotonic structures. Journal of Neurophysiology;
1993;69(6):1865 -- 1879.
3. Steriade M, Timofeev I, Grenier F. Natural waking and sleep states: a view from inside neocortical
neurons. Journal of Neurophysiology; 2001;85(5):1969 -- 1985.
4. Matsumura M, Cope T, Fetz EE. Sustained excitatory synaptic input to motor cortex neurons
in awake animals revealed by intracellular recording of membrane potentials. Experimental Brain
Research; 1988;70(3):463 -- 469.
5. Poulet JFA, Petersen CCH. Internal brain state regulates membrane potential synchrony in barrel
cortex of behaving mice. Nature; 2008;454(7206):881 -- 885.
6. Lee AK, Manns ID, Sakmann B, Brecht M. Whole-cell recordings in freely moving rats. Neuron;
2006;51(4):399 -- 407.
7. El Boustani S, Marre O, B´ehuret S, Baudot P, Yger P, Bal T, et al. Network-state modu-
lation of power-law frequency-scaling in visual cortical neurons. PLoS Computational Biology;
2009;5(9):e1000519.
8. Berkes P, Orb´an G, Lengyel M, Fiser J. Spontaneous cortical activity reveals hallmarks of an
optimal internal model of the environment. Science; 2011;331(6013):83 -- 87.
9. Lapicque L. Recherches quantitatives sur l'excitation ´electrique des nerfs trait´ee comme une polar-
isation. Journal de Physiologie et de Pathologie G´en´erale; 1907;9(1):620 -- 635.
10. Stein RB. The information capacity of nerve cells using a frequency code. Biophysical journal;
1967;7(6):797 -- 826.
11. Latham PE, Richmond BJ, Nelson PG, Nirenberg S. Intrinsic dynamics in neuronal networks. i.
theory. Journal of Neurophysiology; 2000;83(2):808 -- 827.
12. Fourcaud-Trocm´e N, Hansel D, van Vreeswijk C, Brunel N. How spike generation mechanisms deter-
mine the neuronal response to fluctuating inputs. The Journal of Neuroscience; 2003;23(37):11628 --
11640.
13. Brette R, Gerstner W. Adaptive exponential integrate-and-fire model as an effective description
of neuronal activity. Journal of Neurophysiology; 2005;94(5):3637 -- 3642.
14. Hille B. Ion channels of excitable membranes. Sinauer Associates Incorporated; 2001.
15. Hodgkin AL, Huxley AF. A quantitative description of membrane current and its application to
conduction and excitation in nerve. The Journal of Physiology; 1952;117(4):500.
16. Gerstner W, Naud R. How good are neuron models? Science; 2009;326(5951):379 -- 380.
20
17. Druckmann S, Banitt Y, Gidon A, Schurmann F, Markram H, Segev I. A novel multiple objective
optimization framework for constraining conductance-based neuron models by experimental data.
Frontiers in Neuroscience; 2007;1(1):7 -- 18.
18. Pfister JP, Dayan P, Lengyel M. Synapses with short-term plasticity are optimal estimators of
presynaptic membrane potentials. Nature Neuroscience; 2010;13(10):1271 -- 1275.
19. Møller J, Syversveen AR, Waagepetersen RP. Log gaussian cox processes. Scandinavian Journal
of Statistics; 1998;25(3):451 -- 482.
20. Truccolo W, Eden UT, Fellows MR, Donoghue JP, Brown EN. A point process framework for
relating neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects.
Journal of Neurophysiology; 2005;93(2):1074 -- 1089.
21. Pillow J, Shlens J, Paninski L, Sher A, Litke AM, Chichilnisky EJ, et al. Spatio-temporal correla-
tions and visual signalling in a complete neuronal population. Nature; 2008;454(7207):995 -- 999.
22. Paninski L, Ahmadian Y, Ferreira DG, Koyama S, Rahnama Rad K, Vidne M, et al. A new look at
state-space models for neural data. Journal of Computational Neuroscience; 2009;29(1-2):107 -- 126.
23. Rasmussen CE, Williams CKI. Gaussian processes for machine learning. The MIT Press; 2006.
24. Gerstner W, Kistler WM. Spiking neuron models. Single Neurons, Populations, Plasticity. Cam-
bridge University Press; 2002.
25. Paninski L. Maximum likelihood estimation of cascade point-process neural encoding models.
Network: Computation in Neural Systems; 2004;15(4):243 -- 262.
26. Pozzorini C, Naud R, Mensi S, Gerstner W. Temporal whitening by power-law adaptation in
neocortical neurons. Nature Neuroscience; 2013;16(7):942 -- 948.
27. Buzs´aki G. Theta oscillations in the hippocampus. Neuron; 2002;33(3):325 -- 340.
28. Shinomoto S, Tsubo Y. Modeling spiking behavior of neurons with time-dependent poisson pro-
cesses. Physical Review E; 2001;64(4):041910.
29. Lindner B, Schimansky-Geier L, Longtin A. Maximizing spike train coherence or incoherence in
the leaky integrate-and-fire model. Physical Review E; 2002;66(3):031916.
30. Softky WR, Koch C. The highly irregular firing of cortical cells is inconsistent with temporal
integration of random epsps. The Journal of Neuroscience; 1993;13(1):334 -- 350.
31. Katsaggelos AK, Lay KT. Maximum likelihood blur identification and image restoration using the
em algorithm. IEEE Transactions on Signal Processing; 1991;39(3):729 -- 733.
32. Bach FR, Jordan MI. Learning graphical models for stationary time series. IEEE Transactions on
Signal Processing; 2004;52(8):2189 -- 2199.
33. Gray RM. Toeplitz and circulant matrices: A review. Foundations and Trends in Communications
and Information Theory; 2006;2(3):155 -- 239.
34. Efron B, Hinkley DV. Assessing the accuracy of the maximum likelihood estimator: observed
versus expected fisher information. Biometrika; 1978;65(3):457 -- 487.
35. Chichilnisky EJ. A simple white noise analysis of neuronal light responses. Network: Computation
in Neural Systems; 2001;12(2):199 -- 213.
21
36. Vidne M, Ahmadian Y, Shlens J, Pillow J, Kulkarni JE, Litke AM, et al. Modeling the impact of
common noise inputs on the network activity of retinal ganglion cells. Journal of Computational
Neuroscience; 2012;33(1):97 -- 121.
37. Jolivet R, Rauch A, Luscher HR, Gerstner W. Predicting spike timing of neocortical pyramidal
neurons by simple threshold models. Journal of Computational Neuroscience; 2006;21(1):35 -- 49.
38. Cunningham JP, Byron MY, Shenoy KV, Sahani M. Inferring neural firing rates from spike trains
using gaussian processes. In: Advances in neural information processing systems; 2007. p. 329 -- 336.
39. Macke JH, Buesing L, Cunningham JP. Empirical models of spiking in neural populations. Ad-
vances in Neural Information Processing Systems; 2011;24:1350 -- 1358.
40. Mensi S, Naud R, Pozzorini C, Avermann M, Petersen CCH, Gerstner W. Parameter extraction and
classification of three cortical neuron types reveals two distinct adaptation mechanisms. Journal
of Neurophysiology; 2012;107(6):1756 -- 1775.
41. Markram H, Toledo-Rodriguez M, Wang Y, Gupta A, Silberberg G, Wu C. Interneurons of the
neocortical inhibitory system. Nature Reviews Neuroscience; 2004;5(10):793 -- 807.
42. Pfister JP, Dayan P, Lengyel M. Know thy neighbour: A normative theory of synaptic depression.
Advances in Neural Information Processing Systems; 2009;22:1464 -- 1472.
43. Long MA, Jin DZ, Fee MS. Support for a synaptic chain model of neuronal sequence generation.
Nature; 2010;468(7322):394 -- 399.
44. Hamaguchi K, Tschida KA, Yoon I, Donald BR, Mooney R. Auditory synapses to song premotor
neurons are gated off during vocalization in zebra finches. eLife; 2014;3:e01833.
45. Vallentin D, Long MA. Motor origin of precise synaptic inputs onto forebrain neurons driving a
skilled behavior. The Journal of Neuroscience; 2015;35(1):299 -- 307.
46. Haider B, Hausser M, Carandini M. Inhibition dominates sensory responses in the awake cortex.
Nature; 2013;493(7430):97 -- 100.
47. Arlot S, Celisse A. A survey of cross-validation procedures for model selection. Statistics surveys;
2010;4:40 -- 79.
22
Supplementary Text
A statistical model for in vivo neuronal dynamics
Simone Carlo Surace1,2,∗, Jean-Pascal Pfister2
1 Department of Physiology, University of Bern, Bern, Switzerland
2 Institute of Neuroinformatics, University of Zurich and ETH Zurich, Zurich, Switzerland
∗ E-mail: [email protected]
Discrete Fourier Transform
In the following and in the main text, we denote discrete Fourier transforms of vectors of length n by a
hat. The Fourier transformed vector is again of length n and can be formally expressed as
v = Fv,
(Fn)ij = e
2πI(i−1)(j−1)
n
,
i, j = 1, ..., n,
(19)
where I denotes the imaginary unit. In practice, discrete Fourier transforms are not actually computed
by matrix multiplication, but by means of a Fast Fourier Transform (FFT) algorithm.
Circulant matrices
In order to reduce the computational complexity of the likelihood estimation, we approximate the auto-
covariance matrix K (which is a Toeplitz matrix K) with a circulant matrix C. By definition a circulant
matrix can be expressed as
we write C = Cn(c). All circulant matrices of dimension n can be diagonalized by the unitary discrete
Fourier transform matrix U = 1√n
Fn:
Cij = c(i−j mod n)+1
(20)
Cn(c) = U†ndiag(Fnc)Un = U†ndiag(c)Un,
(21)
where † is the conjugate transpose. This implies that c is the vector of eigenvalues of C, and it is a
vector with real entries. This makes calculation of inverse and determinant of C extremely cheap, as is
multiplication of C−1 by a vector x ∈ Rn, which simplifies to
C−1
n (c)x =
1
n
c(cid:19)
F†n(cid:18) x
(22)
where the vector in brackets is the component-wise quotient of the vectors Unx and Fnc.
Circulant approximation
The task is now to find a circulant matrix which is as close as possible to the covariance matrix K. This
can be formalized as the following minimization problem:
C = argmin
D circulant
DKL (N (m, K)N (m, D)) ,
∀m
(23)
where N denotes a multivariate Gaussian with specified mean vector and covariance matrix. This problem
has the unique solution
ci =
1
n
{(n − i + 1)ki + (i − 1)kn−i+2} ,
1 ≤ i ≤ n,
kn+1 ≡ 0
(24)
Proof: The Kullback-Leibler divergence between two Gaussians is given by
1
1
2
n
2
1
ci
det C−1 =
2(cid:2)tr(C−1K) − log det(C−1K)(cid:3) −
c(cid:19) Un,
nXi=1(cid:20) (UnKU†n)ii
2ci(cid:20)1 −
nYi=1
+ log ci(cid:21) + const.
(cid:21) = 0
(UnKU†n)ii
ci
ci
We have
and hence
DKL (N (m, K)N (m, C)) =
C−1 = U†ndiag(cid:18) 1
DKL (N (m, K)N (m, C)) =
where the constant does not depend on c. We obtain the derivative
DKL (N (m, K)N (m, C)) =
1
∂
∂ci
and therefore, at the stationary point we have
ci = (UnKU†n)ii
(F†n)ij (Fn)jlKlm(F†n)mj
23
(25)
(26)
(27)
(28)
(29)
(30)
ci =
=
=
1
n2
1
n2
1
n2
nXj,l,m=1
nXj,l,m=1
nXj,l,m=1
n
Klm exp(cid:20) 2πI
kl−m+1 exp(cid:20) 2πI
(−(i − 1)(j − 1) + (j − 1)(l − 1) − (m − 1)(j − 1))(cid:21)
(j − 1)(l − m + 1 − i)(cid:21)
n
The sum of roots of unity over j only gives a non-zero value if the integer q = l − m + 1 − i is a multiple
of n. Since q has a maximum of q = n − 1 when l = n, m = i = 1 and a minimum of q = 2 − 2n when
l = 1, m = i = n, only q = 0 and q = −n are eligible. Hence,
nXl,m=1
n−1Xr=−n+1
ci =
1
n2
=
=
1
n
1
n
nkl−m+1 (δ0,l−m+1−i + δ−n,l−m+1−i)
(n − r)kr+1 (δi−1,r + δi−1,n+r)
{(n − i + 1)ki + (i − 1)kn−i+2} ,
kn+1 ≡ 0
The second equality is obtained by reparametrizing r = l − m. (cid:3)
Sampling using FFTs
In order generate a sample u of length n from a multivariate Gaussian with mean vector m and circulant
covariance matrix C = Cn(c), one generates a zero-mean white noise vector x with unit variance and
then uses FFTs to compute u, i.e.
u = m +
1
n
F†nh(c)1/2 xi = m +
1
n
Cn(cid:16)F†n (c)1/2(cid:17) x
(31)
as the following calculation shows, the covariance comes out correctly
D(u − m)(u − m)TE =
=
Optimization method
1
n2 Cn(cid:16)F†n (c)1/2(cid:17)DxxTECn(cid:16)F†n (c)1/2(cid:17)T
Cn(cid:0)F†nc(cid:1) = Cn(c) = C
1
n
24
(32)
The optimization scheme used is a quasi-Newton method, where the vector of parameters Θ is updated
according to
where f is the function to be minimized (e.g. − log p(usom, s)), ∇f is the gradient, and the matrix B is
chosen to be
Θ(k+1) = Θ(k) − B(k)∇f(cid:16)Θ(k)(cid:17)
(cid:0)Θ(k)(cid:1) ,
if Hf(cid:0)Θ(k)(cid:1) positive definite
else
γ(k)G−1,
B(k) =(H−1
f
Where Hf is the Hessian of f , γ(k) denotes a learning rate, and G is a metric tensor on the parameter space
which is used to rescale the parameters to lie in similar ranges. The learning rate γ(k) < 0 is increased
when the previous step was successful (typically, by 10 percent), and reduced when the optimizer either
runs into boundaries of the admissible parameter region or increases the value of the function (we used
a reduction by a factor of 2).
Below, we derive the formulae for the gradient and Hessian required for the optimization. The
derivations hold for the case where the GP covariance function k is parametrized arbitrarily by θi and
the spike-shape kernel and adaptation kernel are given by linear combinations of basis functions α(k),
k = 1, ..., na and η(k), k = 1, ..., nw respectively.
Gradient
The likelihood function has the form
log p(usom, s) =
where
nXi=1"−
1
2
log(2πci) −
1
2n
ui2
ci
+ si log qi + (1 − si) log [1 − qi]# ,
qi = ∆teβui+Ai+log r0
is the probability of a spike in bin i and depends on all parameters except the ones that parametrize the
covariance function k. The membrane potential is given implicitly by
It is worthwhile to write the derivatives of log p in the following form:
u = usom − ur − α ∗ s.
d log p(usom, s) =
nXi=1"−
1
2 1
ci
i ! dci −
ui2
c2
1
n
1
nci
−
ℜ {u∗i dui} +
si − qi
qi(1 − qi)
dqi# ,
Now, let us evaluate all the terms one by one. The Fourier transform of the circulant covariance c only
depends on the GP kernel parameters θ, i.e.
(33)
(34)
(35)
(36)
(37)
(38)
dci =
∂ci
∂θk
dθk =
∂θk!i
nkXk=1 d∂c
dθk.
(39)
Let us turn to the q terms next. Their differential is
dqi =
∂qi
∂ui
dui +
∂qi
∂Ai
dAi +
∂qi
∂r0
dr0 +
∂qi
∂β
dβ = qi(βdui + dAi + d log r0 + uidβ),
(40)
25
where by (37) and using the fact that α is a linear combination of basis kernelsPnα
k=1 akα(k)
dui = −dur −
i dak, S(k)
S(k)
i = (α(k) ∗ s)i.
nαXk=1
Moreover, Ai is also a linear combination, so
i dwk, A(k)
A(k)
i = (η(k) ∗ s)i.
S(k)
i dak +
i dwk + d log r0 + uidβ! ,
A(k)
dui = −nδ1idur −
S(k)
i dak.
nηXk=1
nαXk=1
vi =
si − qi
1 − qi
,
Therefore (40) can be written as
dAi =
nηXk=1
dqi = qi −βdur − β
nαXk=1
Lastly, by (41) we also have
Using
all the previous results can be regrouped to yield
d log p(usom, s) =
ci
1
nXi=1"−
2 1
nαXk=1(cid:18) 1
+
nci
(41)
(42)
(43)
(44)
(45)
(46)
−
1
n
ui2
c2
ℜnu∗i
S(k)
dθk
∂θk!i
i ! nkXk=1 d∂ci
i vi(cid:19) dak
i o − βS(k)
nηXk=1
i vidwk
A(k)
+
+vid log r0
+uividβ
ℜ {ui} − βvi(cid:19) dur#,
+(cid:18) δ1i
ci
Hessian
For the Hessian, we mainly need the following
d 1
ci
1
2
i ! = −
ui2
c2
1
n
−
−
1
2 1
c2
i
i ! dci +
ui2
c3
2
n
1
nc2
i
−
ℜ {u∗i dui} ,
(47)
d(cid:18) 1
nci
ℜnu∗i
S(k)
1
nc2
i
i o(cid:19) = −
dvi = d(cid:18) si − qi
ℜnu∗i
1 − qi(cid:19) =
S(k)
i o dci +
1
nci
ℜn S∗(k)
i
duio ,
si − 1
(1 − qi)2 dqi =
ξi
qi
dqi,
where
ξi
qi
=
si − 1
(1 − qi)2 .
The components of the Hessian matrix are computed as follows
∂ci
∂θl" 1
c2
i
i #) ,
ui2
c3
2
n
−
ui
2# −
ci(cid:12)(cid:12)(cid:12)(cid:12)
∂ci
∂θk
1
2
1
n
∂2 log p(usom, s)
∂θk∂θl
∂2 log p(usom, s)
∂θk∂al
∂2 log p(usom, s)
∂θk∂ur
= −
= −
= −
∂c1
∂θk
1
ci
−
n(cid:12)(cid:12)(cid:12)(cid:12)
i ) ,
S(l)
i
c2
nXi=1( ∂2ci
∂θk∂θl" 1
ℜ( u∗i
nXi=1
1(cid:27) ,
ℜ(cid:26) u1
∂ci
∂θk
c2
∂2 log p(usom, s)
∂θk∂wk
=
∂2 log p(usom, s)
∂2 log p(usom, s)
=
= 0,
∂θk∂r0
∂θk∂β
ℜ( S∗(k)
ci
i
S(l)
i
) + β2S(k)
i S(l)
i ξi# ,
∂2 log p(usom, s)
∂ak∂al
∂2 log p(usom, s)
∂ak∂wl
∂2 log p(usom, s)
∂ak∂ur
∂2 log p(usom, s)
∂ak∂ log r0
∂2 log p(usom, s)
∂ak∂β
=
1
n
= −β
nXi=1"−
nXi=1
= −ℜ( S(k)
nXi=1
nXi=1
= −β
= −β
1
i A(l)
S(k)
i ξi,
c1 ) + β2
S(k)
i
S(k)
i
ξi,
ξiui,
S(k)
i
ξi,
nXi=1
∂2 log p(usom, s)
∂wk∂wl
∂2 log p(usom, s)
∂wk∂ur
∂2 log p(usom, s)
∂wk∂ log r0
∂2 log p(usom, s)
∂wk∂β
=
nXi=1
= −β
=
=
nXi=1
nXi=1
i A(l)
A(k)
i ξi,
A(k)
i ξi,
nXi=1
A(k)
i ξi,
A(k)
i ξiui,
26
(48)
(49)
(50)
(51)
(52)
(53)
(54)
(55)
(56)
(57)
(58)
(59)
(60)
(61)
(62)
(63)
∂2 log p(usom, s)
∂u2
r
= −
n
c1
+ β2
ξi,
nXi=1
∂2 log p(usom, s)
∂ur∂ log r0
∂2 log p(usom, s)
∂ur∂β
= −β
ξi,
= −
[vi + βuiξi] ,
=
nXi=1
nXi=1
nXi=1
nXi=1
nXi=1
=
=
ξi,
uiξi,
u2
i ξi.
∂2 log p(usom, s)
∂(log r0)2
∂2 log p(usom, s)
∂ log r0∂β
∂2 log p(usom, s)
∂β2
27
(64)
(65)
(66)
(67)
(68)
(69)
The components as given by equations (33-51) are then combined to form the Hessian matrix Hf , which
is used in Eq. (34).
28
Figure 7. S2 Fig. Supplementary Figure. Comparison of in vivo and artificial data snippets for
datasets D3 and D4, analogous to Fig. 3G,H. The scale (shown on panel D) is the same for all four
panels. Vertical lines are drawn at the spiking times. (A) A 2-second sample of in vivo activity from
dataset D3 (Zebra Finch HVC). (B) Artificial data sampled from AGAPE with parameters learned from
dataset D3. (C) A 2-second sample of in vivo activity from dataset D4 (mouse visual cortex). (D)
Artificial data sampled from AGAPE with parameters learned from dataset D4.
29
Figure 8. S3 Fig. Supplementary Figure. The fitting result as a function of the parameter δ for
dataset D1, see color code next to the plot of the marginal distribution of u in the second row of the left
column. The top left panel shows that the log likelihood peaks at δ = 18 ms, and the bottom right
panel shows the decrease of the effective coupling stength as δ increases.
30
Figure 9. S4 Fig. Supplementary Figure. The fitting result as a function of the parameter δ for
dataset D3, see color code next to the plot of the marginal distribution of u in the second row of the left
column. The top left panel shows that the log likelihood peaks at δ = 12 ms, and the bottom right
panel shows the decrease of the effective coupling stength as δ increases.
31
Figure 10. S5 Fig. Supplementary Figure. The fitting result as a function of the parameter δ for
dataset D4, see color code next to the plot of the marginal distribution of u in the second row of the left
column. The top left panel shows that the log likelihood peaks at δ = 32 ms, and the bottom right
panel shows the decrease of the effective coupling stength as δ increases.
|
1505.04432 | 1 | 1505 | 2015-05-17T18:34:12 | Coupled Hebbian learning and evolutionary dynamics in a formal model for structural synaptic plasticity | [
"q-bio.NC"
] | Theoretical models of neuronal function consider different mechanisms through which networks learn, classify and discern inputs. A central focus of these models is to understand how associations are established amongst neurons, in order to predict spiking patterns that are compatible with empirical observations. Although these models have led to major insights and advances, they still do not account for the astonishing velocity with which the brain solves certain problems and what lies behind its creativity, amongst others features. We examine two important components that may crucially aid comprehensive understanding of said neurodynamical processes. First, we argue that once presented with a problem, different putative solutions are generated in parallel by different groups or local neuronal complexes, with the subsequent stabilization and spread of the best solutions. Using mathematical models we show that this mechanism accelerates finding the right solutions. This formalism is analogous to standard replicator-mutator models of evolution where mutation is analogous to the probability of neuron state switching (on/off). The second factor that we incorporate is structural synaptic plasticity, i.e. the making of new and disbanding of old synapses, which we apply as a dynamical reorganization of synaptic connections. We show that Hebbian learning alone does not suffice to reach optimal solutions. However, combining it with parallel evaluation and structural plasticity opens up possibilities for efficient problem solving. In the resulting networks, topologies converge to subsets of fully connected components. Imposing costs on synapses reduces the connectivity, although the number of connected components remains robust. The average lifetime of synapses is longer for connections that are established early, and diminishes with synaptic cost. | q-bio.NC | q-bio |
Coupled Hebbian learning and evolutionary
dynamics in a formal model for structural
synaptic plasticity
H.P. Vladar and E. Szathm´ary
Parmenides Foundation
Pullach/Munich, Germany
November 18, 2018
Abstract
Theoretical models of neuronal function consider different mech-
anisms through which networks learn, classify and discern inputs. A
central focus of these models is to understand how associations are
established amongst neurons, in order to predict spiking patterns that
are compatible with empirical observations. Although these models
have led to major insights and advances, they still do not account for
the astonishing velocity with which the brain solves certain problems
and what lies behind its creativity, amongst others features. We ex-
amine two important components that may crucially aid comprehen-
sive understanding of said neurodynamical processes. First, we argue
that once presented with a problem, different putative solutions are
generated in parallel by different groups or local neuronal complexes,
with the subsequent stabilization and spread of the best solutions.
Using mathematical models we show that this mechanism accelerates
finding the right solutions. This formalism is analogous to standard
1
replicator-mutator models of evolution where mutation is analogous
to the probability of neuron state switching (on/off). Although in evo-
lution mutation rates are constant, we show that neuronal switching
probability is determined by neuronal activity and their associative
weights, described by the network of synaptic connections. The sec-
ond factor that we incorporate is structural synaptic plasticity, i.e.
the making of new and disbanding of old synapses, which we apply
as a dynamical reorganization of synaptic connections. We show that
Hebbian learning alone does not suffice to reach optimal solutions.
However, combining it with parallel evaluation and structural plastic-
ity opens up possibilities for efficient problem solving. In the resulting
networks, topologies converge to subsets of fully connected compo-
nents. Imposing costs on synapses reduces the connectivity, although
the number of connected components remains robust. The average
lifetime of synapses is longer for connections that are established early,
and diminishes with synaptic cost.
1
Introduction
Many mechanisms of cognition, memory and other aspects of brain function
remain unclear.
It is acknowledged that associations build up by updat-
ing synapses between neurons that spike (nearly) synchronously to a given
stimulus.
In this way some neuronal circuits can predispose or anticipate
a response to similar stimuli by retrieving information stored in synaptic
weights. Synaptic weights may in turn be systematically altered by success-
ful anticipation or recognition activity. At the same time, given the mul-
tidimensional space of alternative neuronal circuits and spiking sequences,
undirected random variation in circuitry and spiking are extremely unlikely
to produce better solutions for each new problem.
The connectivity overall of the human brain is sparse where, roughly, 1011
neurons are estimated to connect through some 1015 synapses. Learning and
2
cognition have been understood in terms of changes in associative weights
on networks of fixed topology. However, the discovery that rewiring this
network is not uncommon even in adult brains challenges the former views
regarding the mechanisms of learning. This rewiring, known as structural
synaptic plasticity (SSP), has been well documented experimentally [1, 2].
However, neither the full consequences nor the central role of SSP have been
fully clarified. Yet, it is not only reasonable, but also supporting evidence
exists, that SSP can encode information [3]. Thus, associative weights and
SSP are two mechanisms that have an effect on learning. These need not
be mutually exclusive; rather, as we show in this article, they both seem to
be necessary for different stages of learning, such as short and long term,
respectively.
Our knowledge about what determines the establishment of new synapses
is still limited, especially taking into account the sparseness and dimensions
of the brain. Neither synaptic weights nor SSP explain on their own vari-
ability in circuitry associated with a particular stimulus. As such, they only
show variability in time.
If trial solutions to a problem (such as learning
or recognizing a pattern) rely on serial evaluations, SSP is a poor candidate
mechanism, even for long-term learning. Under serial evaluations the time
for establishing new synapses would be prohibitively large to account for
randomly testing connections amongst pairs of neurons.
Changeux [4, 5] and Edelman [6] proposed a selectionist [7] framework
for brain function. They noted that selection acts, through preferentially
reinforcing and stabilizing some synaptic patterns over others, and through
the elimination of dysfunctional neurons and neuronal connections. Although
these ideas are correct, they are incomplete because they only consider the
fate of initial topological variability in circuitry, thought to occur only during
development. In their framework, selection acts on this standing variation,
stabilizing functional circuits that remain unchanged throughout life, with
later learning and problem solving resulting only from changing synaptic
3
weights. In this sense, the role of selection is limited to establishing functional
neuronal network at early stages. The ideas that we investigate in this article
go beyond this view: we consider that selection of novel variation plays an
active role in learning through life.
Kilgard proposed a verbal model that accounts for circuitry variation
during learning periods [8]. In his 'expansion-renormalization model' he en-
visions that SSP accounts for such variation. The mechanism is as follows.
When a cortical subnetwork is challenged by a novel task, new synapses are
being generated in response, out of which only the functionally important
ones are kept, while the obsolete ones are eliminated. This is like an iterated
Changeux-type overproduction-selective stabilisation mechanism, and is be-
ing explicitly regarded as a Darwinian mechanism by the author. However,
he fails to discuss particulars such as: what are the true units of variation,
and how this mechanism quantitatively acts. Our ideas are conceptually
similar, but we pin them down to specific 'learning' units and develop quan-
titative models to understand how this variability is generated and how it
affects learning.
We note that there are at least two other sources of neuronal variabil-
ity. The first one is the variance in spiking patterns and is due only to the
stochastic behaviour of neurons (cf. [9, 10]). The second one, which is more
fundamental, is due to SSP, which acts by rewiring the set of neurons in a
complex. Selection is then able to act on the variation that is generated by
the three mechanisms. We point out that the crucial one is SSP, but as we
will explain throughout this article, the three mechanisms play different roles
in learning.
We assume that circuits that result in a sub-optimal solution relative to
the rest of the circuits not only receive less reward, but also are more likely
to be 'overwritten' by transmitting the information in the form of synaptic
weights and structure from other local complexes. During this transmission
process, small variations are introduced to the new circuit through SSP.
4
Iterating this mechanism results in the increase in the representation of the
circuit that gives the best solution, gradually replacing other circuits until
no better variants are further produced, and finally (and ideally) a solution
is found. Our central aim is to understand how different neuronal complexes
might evaluate possible solutions in parallel and thus compete to converge
to an optimal result during learning (Fig. 1). For this, we put together all
these verbal ideas into a quantitative framework.
We study the properties that need to be sought in order to understand
more accurately how learning occurs. For this reason we build up from lo-
cal mechanisms of neural learning. That is, we set our problem at a time
scale that allows us follow whether neurons are found to be on or off. Each
neuron is assumed to fire stochastically, but with a probability given by the
input activity of other neurons in the complex. We will assume reinforcement
learning, and as other works, employ simple measures such as distance be-
tween the output and the target. We emphasize that this is analogous to the
gradient of a fitness landscape in evolution [11]. This analogy will allow us
to tackle the problem with full force, partly by employing the mathematical
models developed in evolutionary biology.
Despite the high level of abstraction of our approach, we acknowledge
that an ultimate verification of our hypothesis needs to come from experimen-
tal neuroscience. However at the moment we intentionally avoid discussing
molecular or physiological aspects, which although essential to understand
the problem experimentally, at this point would simply obscure understand-
ing what we propose are the strategic means through which the brain works
at the level we aim to describe.
1.1 Analogy with Darwinian Evolution
As stated above, so-called Neural Darwinism does not account for the gener-
ation of post-developmental variation repeatedly in circuitry, on which selec-
tive mechanisms could act. However, this is still not enough for a complete
5
Figure 1: Replicative neurodynamics.
(A) The input is fed into several local neuronal groups. (B) Each of these
groups evaluate the input independently, thus trying in parallel distinct spik-
ing patterns (represented by neurons in white and grey states), and (C) pro-
ducing distinct outputs with corresponding reward/fitness values W .
(D)
Groups that result in higher fitness transmit their synaptic weights to other
groups that performed poorly (connections amongst groups are assumed to
exist but are not displayed on the figure, and not explicitly modelled). This
parallel evaluation is repeated until an optimal solution spreads across all
groups.
6
implementation of adaptive evolution in the functioning of the brain. What
is missing is an interpretation of heredity in terms of neurophysiology, so
that the selected variants can be expanded and, from them further variants
be generated. In this way the interaction between selection, variability and
heredity can find the right spiking patterns to solve a problem [7].
The mechanisms for generating variability of neural spiking patterns are
relatively simple to rationalize, and there are many works in the literature
that take this aspect as modelling objective [12]. But it is less obvious,
of deeper implications and of far-reaching consequences, to realise that a
mechanism of 'neuronal heredity' between local complexes might exist.
As explained above, for neuronal heredity to occur it is necessary that
several local networks can act in parallel; stochastic variation in spiking is
not enough. Heredity occurs when circuits that have reached satisfactory so-
lutions transmit their contents to some other circuits that did not perform as
well (Fig. 1). Although there is no replication of the population of neurons
per se (as in a biological population), these repeated rounds of evaluation
and replacement implement a mechanism of heredity [7,13] that is analogous
to genetic inheritance. Admittedly, models of 'neuronal replication' thus far
have relied too much on accurate topographic mapping between local net-
works.
In general this assumption is unjustified. We hasten to note that
a neurobiologically realistic solution to this problem is already in our hands
and will be subject of forthcoming publications. We regard our present paper
a 'formal' one partly because we treat the component process of replication
as a black box of which the content will be revealed later. Thus we perform
an abstract analysis of evolutionary neurodynamics by linking basic theory
in neuroscience and evolutionary biology under the assumption that neuronal
heredity is solved. Note that the discussions on two mechanisms of accumu-
lating knowledge (by evolution and by learning) have been largely isolated
from each other. These two sides of the discussion are not exclusive. We of
course recognize that spiking neurons, Hebbian learning and SSP exist, and
7
are central components of cognition, but we argue that on their own, they
do not suffice for explaining how complex tasks are solved.
In this paper we merge these concepts formally and investigate how learn-
ing and Darwinian selection can work together to drive the system to optimal
solutions. Our approach is akin to models coupling learning and selection
by directly rewarding over stimuli (cf. the direct actor; [14], pp. 344-346),
which, as in our evolutionary approach, results in maximising the average
reward (i.e. fitness). However, our work has a wider scope since it also
merges other ideas such as neuronal copying and structural synaptic plas-
ticity (though see Discussion). We take a mathematical and computational
approach to understand the principal factors that are relevant and isolate
aspects that are in principle falsifiable. We show that for eyes educated in
evolutionary biology, the equations that describe the whole process are as-
tonishingly similar to the mutation-selection equations, albeit with a twist.
The quantity that is analogous to mutation rates is not constant since it
depends on the state of the whole system. The relevance of this difference
is that such 'mutation rates' derive from the associative (Hebbian) learning
mechanism, where the synaptic weights represent the strength of the associ-
ations, effectively represented by a graph with weighted edges. (This would
be equivalent to mutation rates that depend on the extant population diver-
sity.) When coupled to selection, the learning system is more efficient than
simple hill-climbing. This is our first central result. Second, we will consider
asymmetric landscapes and show that the learning weights correlate with the
fitness gradient. That is, the neuronal complexes learn the local properties
of the fitness landscape, resulting in the generation of variability directed to-
wards the direction of fitness increase, as if mutations in a genetic pool were
drawn such that they would increase reproductive success. Third, we study
how this mechanism is efficient in reaching optimal solutions in rugged land-
scapes. We show that the system often reaches sub-optimal solutions when
we assume random networks of synaptic connections. We identify these as
8
impasse states. That is, the network reaches a sub-optimal state where each
possible modification of the synaptic and spiking activity only decreases the
quality of the solution. However, when synaptic plasticity is considered, these
states are easily overcome. For this dynamics, we characterize the distribu-
tion of lifetimes of synaptic connections. This is a central aspect because it
is a quantity that can be measured experimentally. Although at this stage
we are not concerned with parameter estimation or inference, we do point
out that our results are in principle falsifiable.
2 Models
We note that on short time scales (milliseconds) spikes take place and the
selective dynamics can act by rewarding different sub-networks of the neu-
ronal circuit. Yet, variation in spiking can be produced due to changes in
synaptic weights. On a larger time scale, SSP generates novel circuitry. For
simplicity we separate these two time scales. We first describe the joint ac-
tion of selection in several groups and learning. For now we assume that all
groups have the same topology of connections, but each one has a different
spiking pattern; we describe SSP below.
2.1 Learning in parallel groups
In the spiking models, learning occurs by updating the weights that determine
the probability that a neuron fires. This update follows Hebb's rule, verbally
stated as 'neurons that fire together, wire together'. Hebb's rule has been
modelled with fixed connection topology where the weights are allowed to
change according to the covariance amongst neurons, as for example [15]:
∆φij = λXiYj
(1)
9
Table 1: Analogy between the concepts in evolutionary genetics
and neurodynamics.
Evolutionary genetics Neurodynamics
Loci/genes
Nr. of loci
Alleles
Allele frequency
Population
Adaptive landscape
Mutation rate
−
−
−
Neurons
Nr. of Neurons in a group
Neuron state (on/off)
Firing probability
Groups of neurons
Rewarding mechanism, score W
Switching probability
Hebbian weights
Learning rate
Synaptic cost
Symbol
i, j, k
n
X
ρ
N (= ∞)
A, M
φ
λ
k
does not (off), Yk =(cid:80)
where λ is the learning rate, X = 1 if the neuron fires (on) and X = −1 if it
i φkiXi is the output of neuron k, and φijis the weight
between neuron i and j. (Note: in the neuroscience literature, weights are
denoted by w, however this notation is potentially confusing in the context
of evolutionary analyses because a similar symbol is employed for fitness; see
Table 1).
Hebb's rule is problematic because it allows weights to increase unbound-
edly. Thus, for computational convenience, we employ Oja's rule, which is a
version of Hebb's rule with normalised weights:
∆φij = λYj(Xi − φijYj) .
(2)
Oja's rule ensures that the weights are normalised, in this case with Eucledian
norm, so that(cid:80)
i φ2
ik = 1.
Whether any one neuron spikes or not is assumed to be a random event.
The probability with which neurons change state (switch on or off) is given
by an update rule A that depends on the state of the input neurons and their
weights. Thus, the probability that a neuron i is on, P [Xi = 1] ≡ ρi is given
10
by the master equation
dρi
dt
= Aon
i = Pr[Xi : 0 → 1] and Aoff
i (1 − ρi) − Aoff
i ρi
i = Pr[Xi : 1 → 0] are the probabilities
where Aon
that inactive neurons spike and spiking neurons shut dow, respectively. We
assume that the update rule takes into consideration the state of both the fo-
cal neuron and the rest of the neurons in the group at the previous evaluation
round. We also assume a time scale that is larger than the refractory period,
so that that spiking is only affected by the previous state of the network.
(3)
Some models, such as Boltzmann machines (see [14]), assume that the
neuron itself has no memory, and its spiking probability is independent from
i = Pr[Xi = 1X] = 1 − Pr[X =
previous neuronal states. In this case Aon
0X]; the master equation gives
= Aon
i − ρi
dρi
dt
and the 'on' probability is given by
i = [1 + exp(−Yi)]
Aon
−1 .
(4)
(5)
For firing neurons the weights φ increase, so Aon
i
tends to grow to 1.
Hence, co-spiking neurons become more likely to fire in each evaluation
round.
However, we take a different approach by assuming that learning can
be modulated more efficiently by allowing Aon
to have an effect
i
on the network. Note that this description of learning is coarse-grained: it
only tracks how often a neuron tends to be on as learning proceeds. This
is a different view than that of machine learning, where neural networks are
trained by a set of examples from which the weights are inferred. Then, from
this inference the model can be used to predict or classify data that were not
and Aoff
i
11
included in the training set. Our goal in this paper is different: we consider
parallel networks that try to solve a specific problem.
Solving a particular problem requires some comparison between the tar-
get, T and the output or solution Ω. For this purpose, we can employ the
square deviation, ∆2 = (Ω − T )2 which we want to minimize. We assume
that T is a given parameter and Ω is the output evaluation of the network.
Each network presents an alternative solution, thus having a different devia-
tion ∆ from the target. Hence, we minimize the mean value of the deviation,
¯∆2 = E [(Ω − T )2]. Under a proper scheme of neuronal network replication,
this minimization amounts to Darwinian selection. This selective scheme is
the following: First, each local network is weighted according to its fitness,
given by W = exp(−β∆2]. Second, groups that have larger fitness are kept.
Third, networks with lower fitness are overwritten with the content (spiking
and/or weight states) of the groups with large fitness. (There are several
ways in which this copying can be implemented: this is the black box part
as explained above.). Since in the present model we assume that there are
infinitely many groups, replacement need not be done explicitly: we simply
consider the proportions of groups (this in order to have a direct link to clas-
sical population genetics models that assume infinite population size). Since
we assume that copying is random across different neuronal loci, then the
proportion of a group with a specific configuration is simply the product of
the probability of the state of each neuronal locus (this is analogous to the
Hardy-Weinberg assumption of population genetics; [11] pp. 34-39). Mathe-
matically, we track the proportions, ρ, of active neurons and the distribution
of weights across groups. For the former:
= ρi(1 − ρi)
where Wi is the fitness, and ¯W =(cid:80)
dρi
dt
i ρiWi the mean fitness. The first term
is well known to evolutionary biologists and computer scientists: it describes
∂ log( ¯W )
∂ρi
+ Mi(2ρi − 1)
(6)
12
the process of hill-climbing in the direction of fitness increase [11]. Note that
we can approximate log ¯W = ¯∆2 + var(∆). The second term represents the
variability that is generated throughout learning. For simplicity, we assumed
that the switching probability is symmetric (Mi = Aoff
i ), and which
is given by the activity rule
i = 1−Aon
where Yi =(cid:80)
Mi =
1
1 + eYi
(7)
j φijXj is the activity or current of the the input neurons, and
φijare the weights determining the associations amongst neurons, and which
evolve according to Oja's rule. As more spiking neurons are connected the
activity of the focal neuron increases and its switching probability decreases
asymptotically to zero. Whether a neuron stays on or off however depends
non-trivially on the collective success of reaching the target T .
each network has, on average, input activity ¯Yi =(cid:80)
Appendix A shows that to first-order approximation we can track only
the mean weight at every synapse and apply a general learning rule to all the
average activities of the ensemble of groups. That is, we approximate that
¯φij(2ρj − 1). This still
allows each network to have a different spiking pattern from those of other
groups. We will see below that even under these simplifying assumptions,
evolution has a dramatic effect by accelerating convergence to maximum
fitness (or minimum ¯∆2). We will assume small initial values of the synaptic
weights. Moreover, the variance of these becomes increasingly small as the
neuronal complexes converge to a solution. Thus, we will make no further
distinction between ¯φ and φ.
j
Summarizing, for n neurons and k synapses amongst them we work with
n + k ordinary differential equations. Initial conditions are assumed to be
random and the topology of the network is assumed to be fixed during each
learning round.
13
2.2
Information content of a synapse
How to measure the information content of a synapse is not obvious. First,
we point out that it is important to discern between local measures, such
information in a particular synapse, or global information in the whole cir-
cuitry of a complex, or even across complexes [16]. Different choices depend
on the specific purpose, and can be made according to distinct epistemologi-
cal views and/or experimental purposes. For the aim of this article we choose
one of the simplest measures, mutual information, H, because it describes
the interdependency amongst two specific neurons in the context of a specific
complex. Mutual information is defined as
(cid:88)
Hij =
r,s∈{0,1}
Pr[Xi = rXj = s]ρj log
(cid:20)Pr[Xi = rXj = s]
(cid:21)
ρi
A.
(8)
To calculate the conditional probabilities we first evaluate the input ac-
tivity Eij of neuron i by fixing the value of neuron j to ρj = 0 or 1. This gives
a conditioned value of the switching probability, Mij. Then, the solution to
Eq. 6, using the conditioned switching probability Mij, gives the desired
conditioned probabilities. In this case we keep the weights constant because
we are only assessing the information capacity of the specific synapse and not
the information capacity of the whole network. The exact expression of H is
derived in the Appendix B, where we also show that for Gaussian selective
landscapes information is approximately:
Hij = 6φ2
ijMiMj .
(9)
Mutual information quantifies how likely it is that, if one neuron spikes,
the other one will also do so.
If the neurons are not connected, φij = 0
implying Hij = 0. If the focal neuron i spikes randomly (large M ) then the
information content is low (in that case φij is expected to be close to zero).
Since 0 < φij < 1, the quadratic term dominates over M , which makes Hij
14
proportional to φ2
ij. Hence, the information of a synapse increases as the
weight increases. Note that for a given switching probability, the learning
weights are higher in sparse networks than in fully connected ones. Thus, to
an extent, the former encode more information than the later.
2.3 Structural synaptic plasticity
We implement SSP on a time scale different from that of associative learning.
The system described above is in terms of statistical averages, and can be
regarded as conditioned on a given network of connections. We assume that
synaptic connectivity changes occur in one arbitrary group (below we explain
how changes are introduced). If the new topology improves fitness, it spreads
across all groups. However, we allow for a random component to avoid
trapping in local optima, where the network is stranded in a state of impasse.
For simplicity this is implemented through a Metropolis-Hastings algorithm.
That is, if fitness increases with the new topology, this spreads to all groups.
If fitness remains unchanged or decreases, then the layer might spread with
probability exp(Wnew − Wold). Allowing for this fitness decrease facilitates
the escape from states of impasse.
We assume that changes in synaptic structure follow two heuristic rules
inspired from neuroscience. First, if two neurons are unconnected but they
are highly likely to spike, then a new synapse amongst them can be intro-
duced. There is evidence that synaptic rearrangements result from circuit
rewiring upon (e.g.
in neocortical pyramidal neurons) stimulation [17 -- 19].
Algorithmically, we randomly choose pairs of neurons with probability ρiρj
amongst the set of unconnected pairs of neurons. Second, we allow existing
synapses to be disconnected randomly with probability
R = exp[−αH] ,
(10)
where H is the synaptic information (Eq. 9. That is, if a synapse is in-
15
formative, then it is unlikely to be disconnected, whereas if it contains no
information, it is likely to be disconnected [3, 20, 21].
In addition, we perform a Metropolis acceptance/rejection system: if the
change in the network structure results in fitness increase, it is accepted. If
on the contrary, the change results in fitness decrease, it is accepted with a
probability proportional to the fitness ratio between that of the new network
to that of the old one.
3 Results
3.1 Selection and learning together speed up finding
solutions
W = exp[S(cid:80)
To understand how learning and selection jointly act we first assume an el-
ementary scenario, i.e. a simple hill climbing process where we target for
all neurons to be on. This situation results from a fitness function given by
i ρi], which has a constant gradient, ∂ρ log ¯W = S for every
neuron. (This can be seen as a limit where T is far from the current state,
thus S = 2βT ). We assume that Hebbian learning is slower than selection;
i.e. λ < S. This regime describes the coupling of rounds of learning with
copying across groups. Otherwise, learning would be completed indepen-
dently in each group, associating random spikes and would not be able to
learn the relevant features of the landscape. The associations that the net-
work makes would relate back to previous states, effectively acting against
hill climbing. However, in the regime where learning is slower than selection,
fitness increases the representation of solutions, and once these are stabilised,
learning can create meaningful associations. Figure 2 presents a typical out-
come where the process is characterized by three stages. We first observe an
initial exponential increase in the proportion of active neurons. Compared
with systems that do not learn the dynamics are similar on short time scales.
16
Figure 2: Example of selection-learning dynamics.
(A) Selection-learning dynamics
(black lines) compared to standard
mutation-selection with naıve transition rates (M ∼ 1/2; blue) and to the
run with the transition rates already learnt (red). Inset: evolution of fitness.
(B) Evolution of the transition rates. Inset: evolution of Hebbian weights.
n = 20, S = 5, λ = 0.001.
Initial conditions for allele frequencies and for
initial weights are randomly sampled from a uniform distribution U [0, 0.01].
The learning network is fully connected. Note the log-time scale.
This is because selection increases the representation of groups that provide
better solutions, but these are initially in very low proportions. These fitter
solutions are simply products of lucky stochastic events.
Initially there is
hardly any learning, indicated by light weights, and selective expansion sim-
ply amplifies those groups that have higher activity. This amplification takes
on the order of 1/S rounds of evaluation. As groups become selected learn-
ing takes over, entering an incubation period where associations are built up
because a good proportion of neurons fire correctly. After a learning period,
associations are fully strengthened and the solution is finally reached whereby
switching probabilities reach a minimum (Fig. 2B). The width of the plateau
has a duration of roughly 1
S . This regime is noticeable on a log-scale.
Although in absolute time the selective process is so quick that it might pass
unnoticed, this selection stage is crucial to explore configurations that can
be fixed through learning. We emphasize that this early stage corresponds
to the selective stabilization in the Neural Darwinism theory.
λ − 1
17
(A)�/��/λ0.010.1011010010001040.00.20.40.60.81.0TimeSpikingfrequencies0.010.1011010010001040.00.20.40.60.81.0Fitness0.010.1011010010001040.00.20.40.60.8TimeSwitchingprobability(B)0.010.101101001000104-0.20.00.20.4Associativeweights,ϕ making W = exp[(cid:80)
Instead of favouring equally all neurons to spike, the landscape can be
set to favour distinct neurons to fire preferentially over others. For instance,
i Siρi] and allowing Si to take any arbitrary value intro-
duces asymmetries to the landscape. Crucially, if the dynamics are re-run
with the learnt weights, the equilibrium is reached order of magnitudes faster.
We stress that this is true even if initial spiking probabilities are randomised
(Appendix C).
Importantly, the switching probabilities become strongly correlated with
the fitness gradient, meaning that the associative weights learn the local
properties of the landscape. Figure 3 shows that scaling the relative switching
probabilities as Mi = Mi/Mm, where Mm is the largest M of all the neuronal
n − 1S/Mi of the form:
loci, then there is a universal behaviour with S =
√
M = e
− 1
S .
(11)
This is an approximate form which breaks down at very low values of S (see
Appendix D). Intuitively, we would expect that larger values of S would result
in small switching probabilities. What happens is that the coupling of the
system is sensitive to the distribution of S. For example, if all the selective
values are similar (i.e. the variance is low) then the corresponding M 's are
all low (e.g. red curves in the inset of Fig. 3). If the S's are distributed in a
wider range, then the M 's are also distributed more widely, but have, overall
larger values (e.g. black curves in the inset of Fig. 3). In other words, for a
given mean value of S, the average M increases with var(S). In fact, there
is a positive correlation between the variance of S and the mean switching
probability (data not shown). In turn, for a given variance of S, M decrease
with the mean of S. This means that most efficient complexes are composed
by neurons that are required to fire more specifically.
For a given system, the associative weights increase (asymptotically) with
the strenght of selection (data not shown). We performed performed Spear-
man's ranked correlation test to measure the strength of the dependency.
18
Figure 3: Relationship between selection intensity and associative
weights.
By scaling the switching probabilities with the maximum M of a network,
n − 1S/M , we obtain a universal
and scaling selection intensity as S =
behaviour (black line). The distinct symbols represent runs with (squares)
n = 5, (circles) n = 10 and (diamonds) n = 15 neuronal loci. For each n we
report 300 independent runs. The selective values at each locus are chosen
independently from a U [0, 10]. λ = 0.01, ρ0 U [0, 0.1], φ0 U [0, 0.01].
√
(Because of the non-linearity, 'standard' Pearson's correlation is not a good
measure for the dependency between S and φ.) In absolutely all cases the p-
values were numerically zero, indicating strong dependence amongst S and φ
(data not shown). This strong statistical support indicates that the synapses
encode the fitness gradient, directing variant spiking patterns accordingly:
strong selection results in strong weights, which in turn decrease the switch-
ing probability. This leads in minimal variability of spiking, which maximises
speed of fitness increase. Conversely, weak selection leads to poor associa-
tions resulting in large spiking variability, which allows exploration of the
landscape.
We note the learnt equilibrium point is independent of the learning rate
19
□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□□○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○○◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇◇02004006008000.750.800.850.900.951.00Scaledselectionstrength,n-1S/MScaledswitchingprob.M/M*●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●2468100.02400.02450.0250Selectivevalue,SSwitchingprob. λ. This turns out to be generally true, regardless of the fitness landscape.
We also note that under these 'directional landscapes' the initial conditions
(of both weights and allele frequencies) do not affect the equilibrium state
of the system. (However, later we will see that under more complex fitness
landscapes this is not true.)
3.2 Formal analogy between evolutionary dynamics and
neurodynamics
At this point we formalise further the analogy with evolutionary biology, and
more specifically with population genetics. First we realise that the bimodal
neuron model is analogous to a biallelic genetic system. We start by clarifying
a small but crucial difference in the notation. While in the models considered
in this paper neurons take states {−1, +1}, in population genetics alleles are
typically denoted as {0, 1}. The +/- notation is convenient mathematically
in order to describe Hebb's rules, thus in our evolutionary analogy we also
require this property. Hence if G is the value of a gene or allele, then we
define X = 2G − 1. In this way we can readily apply the machinery from
evolution to neuronal networks.
Second, we consider the spiking probability of a neuronal locus (node)
across all groups (Fig. 1). This average, which is the probability Pr(Xi) that
a neuronal locus i fires in some of the groups, is thus analogous to the average
E[2Gi − 1] = 2ρi − 1, where ρi are allele frequencies at locus i. Note that
allele frequencies are interpreted as the probability of sampling a particular
allele in the population. Thus, for the analogy to be consistent, population
size needs to be analogous to the number of groups involved in the learning.
Although in both populations of individuals and of neuronal groups numbers
are in fact finite, in this work we consider, as a first approximation, an
infinite number. In this way we do not need to worry about stochastic effects
that complicate the analyses. However, we recognise that randomness due
to finite population size (a.k.a. genetic drift) can play a crucial role in both
20
evolution and in learning. This is because randomness facilitates escaping
local peaks and exploring the landscape in a less constrained manner. But
before taking stochastic factors into consideration we want to focus on the
interaction between selection and learning in the infinite population model.
Third, upon reproduction a population generates a new set of individuals,
which sooner or later replaces the parental population. However, in neuro-
dynamics, reproduction has to be interpreted in a particular way, because
there is no generation of a new set of neuronal groups. However, the selec-
tive copying into groups with inferior performance effectively corresponds to
a new population of groups (Fig. 1D).
Given the analogies above, we can ask the converse question: what is the
interpretation of the learning process in evolutionary dynamics?
Equation 3 describes the activity changes of neural networks across it-
erations, leading to an update rule of the spiking frequency of each neuron.
In population genetics, this transition probability corresponds to a mutation
rate. In molecular evolution mutation rates are normally state-independent,
dictated by, for example, copying errors of the polymerases that replicate
DNA, repair mechanisms, or other molecular processes that do not depend
on the genetic states of the individual or population. (Although there are
genetic models that consider evolvable mutation rates; see Discussion.) The
switching probability Mi = 1/(1 + exp[Yi]) is dependent on the state of the
system, and follows directly from the update rule. Apart from this depen-
dency, the equations (Eq. 6) are analogous to a selection-mutation equation.
The resemblance is a natural outcome from the analogy laid out above.
But beyond the cosmetic similarity between the replicator-mutator equa-
tion and neural dynamics, the crucial difference is that the update rule is
able to learn the local properties of the fitness landscape. By doing so, hill
climbing towards a fitness peak is facilitated by generating variation directed
towards the fitness increase.
21
Figure 4: Neurodynamics in a stabilizing rugged landscape.
Model parameters: T = 15, otherwise as in Fig. 2.
3.3 Learning in rugged landscapes
We now consider the more complex adaptive landscape, given by W =
exp [−β∆2]. In evolution this kind of landscapes are known as 'stabilizing se-
lection'. The complexity of this landscape results from the non-linear effects
(known as epistasis in genetics and evolution). These are hard landscapes
to explore because there are many local peaks or solutions, some equally
optimal, some sub-optimal, and simple hill-climbing algorithms often fail to
converge to an absolute maximum of fitness.
Figure 4 shows the neurodynamics. We find that exactly 15 neurons fire
(with probability ρ = 0.995) and the remaining five remain off. In this case
the uninformative neurons are shut down. Which neurons spike and which do
not is contingent on the initial conditions, but in this landscape the identity
of each neuronal locus is meaningless. Different initial conditions can lead to
different but equivalent solutions (data not shown).
3.4 Random and sparse topologies of the neuronal con-
nections impair learning
So far we assumed that there are synapses amongst all pairs of neurons. Re-
laxing that assumption corresponds mathematically to fixing certain weights
22
0.010.1011010010001040.00.20.40.60.81.0TimeSpikingprobability(A)0.010.1011010010001040.00.20.40.60.81.0Fitness0.010.1011010010001040.00.20.40.60.8TimeSwitchingprobability(B)0.010.101101001000104-0.20.00.20.4Associativeweights,ϕ φij to zero, indicating that no synapse exists among neurons i and j. Un-
der these circumstances the equilibrium switching and spiking probabilities
are more variable, with the spread determined by the connectivity of the
underlying learning network. Appendix E presents some neurodynamic out-
comes using different random topologies under directional and stabilising
landscapes. These topologies are drawn from different random graph models
with various degrees (see Methods). We tried Erdos-R´enyi, Barab´asi-Albert
(scale free) and Watts-Strogatz small world topologies. Each of these mod-
els has different statistical properties.
Irrespective of these, there are two
central conclusions. First, random networks lead to unfit solutions, where
the systems cannot reach the target. This is true regardless the target value,
number of neurons and type of topology. The systems typically converge to a
suboptimal solution where no further learning can happen and cannot escape
local optima. We regard this as a situation where a network that was previ-
ously functional for another task is repurposed for a new task, and the initial
topology is, regarding to the new task, arbitrary. Thus, the initial circuit is
not expected to be adapted to the new task. Consequently, what the system
can learn is only limited, and in the vast majority of cases, suboptimal. We
identify these solutions as states of impasse, which means there is no further
progress possible, because any small modification to the synaptic weights or
spiking patterns leads to a state with lower fitness score.
The second central conclusion is that poorly connected neurons have very
low input activity, leading to high switching probabilities. Highly connected
nodes have small switching probabilities with spiking frequencies close to
unity. Hence, only highly connected nodes (the less frequent) can learn ef-
ficiently. Since random topologies give suboptimal results, we consider that
details regarding specific network distributions are secondary and discuss
them only in the Appendix E.
Our choice of network distributions is arbitrary, motivated by mathemat-
ical convenience and scientific hype. Hence, the results above do not nec-
23
essarily imply that brains are suboptimal unless fully connected. However,
they reveal that in sparse networks, as in real neuronal complexes, Hebbian
learning does not suffice to solve complex problems because it too often leads
to impasses.
3.5 Structural Synaptic Plasticity
Structural synaptic plasticity is a mechanism that goes beyond the update of
existing synaptic weights (i.e. Hebbian learning) by allowing new synapses to
be established and old ones eliminated. This dynamical restructuring of the
topology of the neuronal networks as the system learns has been shown to be
important for the transfer of short-term to long-term memory [8]. However,
we test the role of SSP in the more general scenario of problem and impasse
solving.
Above we found that network topology impairs problem solving on com-
plex learning landscapes. This is paradoxical because there is clear evidence
that the brain is not fully connected, even though the type of connectivity
is disputed and tissue-dependent. But our negative results do not rule out
that there might be specific topologies that facilitate or optimize learning.
We now show that under synaptic plasticity, the neuronal complexes form
particular structures, which are unlikely to be recovered randomly, and thus
accounts for the negative results above.
A mechanistic description of SSP rooted in neurophysiological processes
is beyond the scope of this article, in part because much is unknown. Instead,
we propose a simple phenomenological model to show that SSP can not only
affect the outcome of the learning process fundamentally, but also that this
mechanism can resolve impasses. Further specific aspects keep the essential
features unchanged, even though some of these might be crucial for the actual
implementation of the cellular mechanisms that we explain using a simple
model.
A central feature of SSP in the neurodynamics is that, by modifying the
24
distribution of synapses, it provides new pathways to explore the space of
solutions. This is why SPP can be an efficient mechanism to escape impasses.
3.6 Dependency on the number of neurons
This random mechanism allows the neuronal networks to explore the space
of configurations, leading, on average, to an increase of fitness. Figure 5
shows that systems with few neurons evolve good solutions more easily than
larger systems. This is clear: finding an optimal configuration with a few
neurons requires fewer evaluations than larger networks simply because the
search space of the former is much smaller than that of the latter. The
number of possible configurations increases with n2, thus on the basis of
trying one modification at a time the convergence time increases non-linearly
with the number of neurons. However, although this holds true for our
model, there is no reason to think that there cannot be parallel evaluations
of different topologies in different complexes, dramatically alleviating this
inefficiency. However, in this paper we restrict ourselves to evaluating one
modification per iteration. Note that one step in the iteration does not
correspond to a physiological time unit because the Metropolis algorithm only
ensures convergence to the equilibrium distribution as dictated by detailed
balance and considers no information regarding the diffusion leading to said
equilibrium.
In each round of learning (i.e. after the system converged to equilibrium
with a newly tested network) the current weights are being kept, and new
connections are assigned to new random initial values. Alternatively, we can
simply reset all weights to random values. The second strategy proves to be
more efficient than the first. Whether spiking probabilities are reset or not,
proved irrelevant (data not shown).
25
Figure 5: Dynamics of structural synaptic plasticity.
Example of a random realization with n = 30 neuronal loci (fitness is scaled
to the maximum value). Inset: Absolute fitness as a function of neuronal
loci. Parameters: T = 7; otherwise as in Fig. 2. Synapses are assumed to
have no cost.
3.7 Structural plasticity leads to maximal connected-
ness
The resulting synaptic networks are straightforward. Recall that our test
problem chooses for a target number T of spiking neurons. Thus, the optimal
state has exactly T neurons on and the rest are off. Ideally, these T neurons
are fully connected amongst them. We find that the complexes correctly
converge to solve the problems, and, thanks to SSP, the networks that evolve
fully connect these active neurons (Fig. 6).
In other words, the systems
converge to networks that fully connect the required components to solve
the problem.
The convergence to fully connected networks is due to two factors. The
26
0501001502002500.650.700.750.800.850.900.951.00IterationRelativefitness●●●●●●101520253010-710-50.0010.100Numberofneurons,nAbsolutefitness Figure 6: Evolved learning networks.
Example of evolved networks with n = 15 (left) and n = 20 (right) neuronal
loci. The colors indicate the frequency of active neurons. In both cases the
particular node labels are irrelevant, and the proportion of active neurons
depend on the initial conditions and the history of the process. Note that
irrespective of the number of neuronal loci, the number of active components
is correct. Parameters as in Fig. 5
first is the need to switch on the right number of neurons, which requires
strong synapses amongst them. The second is to switch off the unneeded com-
ponents; this also requires connected components because negative weights
between active and inactive decrease the probability of firing.
If negative
weights are not allowed, the system can only maximise fitness by ensuring
the right neurons are on, and the networks converge to fully connect these
components (Fig. 6).
3.8 Neuronal networks are robust to small costs of
synaptic connections
Now we penalize for the amount of connections that the networks have (Fig.
7). There are various reasons to assume this constraint. First, there are costs
associated to synaptogenesis. Second, there are higher metabolic costs due to
the transmission of action potentials, which increases at least proportionally
27
123456789101112131415123456789101112131415161718192010 (if not allometrically) with wiring. Third, there are major spatial constraints
in the brain, limiting the amount of white matter that can be packed. In
order to take into account these and other reasons for limiting the amount
of neurons, we include a fitness cost to the system, exp[−kd], where k is the
cost per synapse, and d is the number of synapses (number of edges in the
learning network).
In Fig. 7 we observe that finding the solution to a problem is impaired
as the cost of establishing synapses is increased.
(In these examples we
target T = 7, but the particular choice of the target value is unimportant; in
the Appendix F we present results for other target values.) Clearly, this is
because the number of connections decreases with increasing cost, which in
turn compromises spiking specificity.
It might be unsurprising that the number of synaptic connections de-
creases with their cost, and naturally, the networks become less discrimi-
native as they lose connections. However, it is also true that they show a
notable level of resilience (graceful degradation), in the sense that even if
performance is somewhat impaired as synapses are eliminated the required
number of connected neurons is robust to the cost. In other words as the
cost increases the networks still converge to structures that connect (even
if sparsely) the necessary neurons (see Figs. 7B). The networks lose perfor-
mance as they lose synapses because neurons receive less input and therefore
their switching probability becomes higher. Nevertheless, they tend to re-
main connected with as many components as possible.
In the stationary state the distribution of networks is broad. As indicated
in Fig. 7B, the average number of synapses decreases with the cost; this is also
true for its variance (in Appendix G we present the degree distributions). The
meaning is that as the cost increases, each neuron establishes fewer synapses
with other ones. Also note that some of the components tend to have only
few connections. This is indicated by the notion of connectivity (Fig. 7B),
which is the number of synapses that we need to remove to separate the
28
network into two unconnected subsets. Typically in the evolved networks
this is due to a single poorly connected network, rather than to connected
sub-complexes interconnected by a few synapses. As costs are very high
(k ∼ 1), the networks are sparse and have several unconnected components.
We point out two important differences between random networks and
the evolved distribution of networks. First, taking the random network as
a null model (the Erdos-R´enyi, is the one that best matches the evolved
distributions, Appendix G) we expect a binomial distribution B[n−1, p]. The
observed distributions are reminiscent of the binomial using the empirical p's.
However in all cases we rejected the null hypothesis (χ2 tests, all p-values
numerically zero); the expected variances are too low.
Second, despite the variability in the distribution of the evolved networks,
these solutions are not in states of impasse. With fewer synapses, the input
activities of the neurons are lower, translating into larger switching proba-
bilities (Fig. 7C). This does not reduce specificity of firing: still the correct
neurons are more likely to fire in an idiosyncratic manner. However, there
is more 'background' noise due to fluctuations. We could say that for larger
costs, neurons are still accurate, albeit less precise.
3.9 Distribution of synaptic lifetimes
Part of the reason why our model shows a large variance in synaptic network
topologies is that we allow for synapses to continuously be established and
disbanded. We still ignore the quantitative and dynamical extent of these
processes, reason why we can only make arbitrary assumptions. Precisely
this, however, is an important point of falsification of our ideas. Irrespective
on the details regarding the assumed synaptic distribution, the observable
distribution can lead us to refine our model. Yet, the relevant aspect here is
not the details of the predicted distribution (since, as argued above we present
a simplified model), but the fact that we predict that there is a distribution
of synaptic lifetimes at all. For instance, Fig. 8 presents the timelines of
29
synapses. We note that many are short lived and a few live long. This result is
expected to hold regardless of the specific problem being solved, in as long as
we do not trace the identity of the neurons that the synapses are connecting.
Clearly, the number of long-lived synapses must increase with the number
of neurons in a complex. However, the lifetime distribution should hold
relatively robust. This is expected because what determines the lifetime of a
synapse is essentially its local information, and this is independent of other
properties of the network. For example, increasing the parameter α, which
controls how weaker connections are penalised, modifies the probability R
of synaptic disbanding so that the average synaptic lifetime become shorter.
If however the target or intensity of selection is changed, the effect on the
distribution of synaptic lifetimes is minimal (Appendix H).
Our results show that synapses that are established early tend to last
longer than synapses established in later developmental stages (Fig. 9A).
This is a natural outcome of the model; we have not imposed any mecha-
nism that presupposes this behaviour purposely. This happens due to his-
torical contingencies. Some of the synapses that are initially established lead
the network close to a local peak. Afterwards other synapses fine-tune the
network, increasing fitness in smaller amounts. Only after the network has
been populated, new connections can substitute the original ones. Figure
9B reveals how synaptic lifetime increases with the cost. For newly formed
synapses, low costs do not show a central trend, implying a certain degree
of resilience or robustness. In the long run, in the small cost range, synapses
established later live shorter as the cost increases. For larger cost range,
synaptic lifetimes tend to increase with cost. However, recall that for costs
k (cid:38) 0.5 the networks have low or no connectivity. As the cost increases, the
network starts to dispense with inhibitory connections. Recall that shutting
unwanted neurons down only decreases the background noise, but does not
affect the result of the 'good' neurons. Thus, synapses amongst these 'good'
neurons tend to live longer, as they are essential for the system's performance.
30
Due to the redundancy of optimal solutions occasional major restruc-
turing of the networks are expected to occur during lifetime for two reasons.
First, once a peak has been found, disbanding the principal connections leads
to major function impairment (fitness decrease). Second, establishing new
synapses that are potentially as good as the existing central ones is unlikely
due to the costs of synaptic connections. However, we do find occasional ma-
jor network restructuration. Because of the strong coupling amongst several
neurons and synapses, once a major connection is destroyed this can cause
further impairments by subsequent changes that result in even worse fitness.
At some point there is a restitution of the system when a new fitness peak is
approached. This has the qualities of self-organised criticality. However, it
might also be that the stochastic behaviour allows a few groups to shift from
sub-optimal solutions to better ones, effectively jumping across fitness peaks.
The subsequent replication of these successful solutions to other groups can
result on a full escape from impasse states.
(In evolution this process is
known as shifting balance; [22, 23].
4 Discussion
4.1 Relationship to previous models
Using a Bayesian framework, [12] proposed a model that explains aspects
of cognitive learning in children.
In their model, the brain implements a
Bayesian update algorithm to form theories based on observed data. They
stress that in contrast to a more reductionist 'connectionist' view, Bayesian
learners explain neuropsychological aspects of the dynamics of learning. In
their framework, theories map to a multidimensional landscape where well-
formed theories lie at peaks. The dynamics include learning, but only at as
a local process. They argue that learning cannot account for the invention of
new theories, but rather, only modify the degree to which we believe in any
given theory. That is, learning acts to fine tune the theory around a peak.
31
The proposal of new theories does not happen through learning, but through
stochastic modification of the existing theories (in a data-independent man-
ner, that is, there is random variation). If a new theory scores better than
the previous ones, it is adopted with certain probability [24].
This model is similar to ours, particularly in the implementation. How-
ever, there is nothing mystical about this coincidence. What Ullman et al..
and we describe belongs to the Markov Chain Monte Carlo class of mod-
els, which is a popular methodological toolbox in stochastic processes. One
important common aspect is that learning alone does not produce any new
configurations (networks in our case, theories in theirs), but only improves lo-
cal adaptation given the current configuration. Whilst their model describes
processes occurring at high level of cognition, we describe simpler processes
at the neurophysiological scale. However, we reach similar conclusions re-
garding the need and limitations of learning in relation to other processes
that can generate variant configurations that lead to a better performance.
This coincidence and its consequences (see discussion in Ullman's paper) are
preliminary evidence supporting our proposed physiological mechanisms.
Although Ullman et al.. do not discuss the states of impasse these are
implicit in their models. Also, the explanations they give in terms of 'theory
formation' are to a large extent equivalent to those of our model's. That
is, learning a local peak restricted to a given configuration results in weight
values that always lead to lower scores if any modification is introduced.
They resort to stochasticity (see below) as a mean to jump across peaks. In
our case this stochasticity is introduced via synaptic plasticity. There can
be other sources of stochasticity, which we address below. However, synaptic
plasticity is a component accounting not only for the necessary randomness to
escape peaks, but is also known to be an important component of learning.
The question of what generates circuit diversity remains so far open: we
propose that SSP can account for this diversity.
Note also the crucial difference in the search mechanisms in the two mod-
32
els. Kemp and Tenenbum [24] use a greedy search algorithm on stochastically
generated variation, whereas we adhere to the view of the Darwinian neu-
rodynamics, also known as the neuronal replicator hypothesis [25, 26]. The
point is that on vast combinatorial landscapes, evolutionary search is known
to produce impressive results. The greedy algorithm works for relatively
small spaces but for larger spaces more efficient search is needed, as Kemp
and Tenenbaum [24] acknowledge. It is remarkable that although Ullmann
et al. [12] explicitly draw a rugged conceptual landscape, the possibility of
evolutionary search is not mentioned. Interestingly, this can lead to predic-
tions. That is, we expect an inverse correlation between the rate of synaptic
turnover and the capacity to overcome impasses. For instance, the rate of
synaptic turnover tends to be slower in adults, who are often less able (or do
it at a slower rate) to overcome impasses in certain neuropsychological test
problems than children. It has not escaped our attention that this can also
be addressed through experimentation in animal models from several points
of view, from behavioural to neurophysiological.
Besides stochasticity in circuitry, as addressed in this paper, we iden-
tify at least two more stochastic sources. One has been addressed already
in the first part of or work. That is, the stochastic nature of neuron firing.
Particularly in our infinite layer model, the solution to any posed problem ex-
ists a priori, because possible spiking patterns combinations are represented.
Clearly, any arbitrary configuration will occur in an infinitesimally low pro-
portion. However, any random spiking pattern that is not causally supported
by a synaptic network is extremely unlikely to be repeated in the following
rounds of evaluation. Therefore even though it might be selected for one
round, this layer will be overwritten at a later stage. Thus the stochastic-
ity from neural spiking is not a source of heritable variation across groups.
Since neurons spike with probabilities given by their input activity the only
meaningful variation that can be produced is that which follows from mod-
ification of synaptic weights. However, as we have extensively argued, only
33
this provides the substrate for selection to act on a local basis.
We call attention to a partly related model by Seung [9] invoking 'he-
donistic synapses' that would release neurotransmitters stochastically, and
an immediate reward would either strengthen or weaken them according to
whether vesicle release of failure preceded reward, respectively. It was noted
that such randomness in synaptic transmission would play the role of muta-
tions in a Darwinian analogy. Seung also notes that stochasticity in action
potentials could play a similar role, and that mechanism would be faster [27].
But since ultimately these mechanisms operate on a fixed topology, the lim-
itations without structural plasticity remain. Note that 'copying' in our
mechanism is a fast component, intermediate between spikes and Hebbian
synaptic plasticity. This is a valid assumption if we assume that copying is
aided by dedicated adaptations (cf. Adams, 1998).
Another source of stochasticity can result from the competition amongst
not an infinite but a relatively small number of groups. In this case the de-
tails of how neuronal copying occurs can be important. Some of the N groups
that have a solution with superior fitness score have to overwrite some groups
that have lower score. This cannot be done in a proportional way, as in the
infinite layer model, partly due to the topographic, rather than merely topo-
logical structure of neuronal networks. Therefore, there is some randomness
of which groups are overwritten and which ones not. Naturally, the ones
with higher fitness are more likely to be transcribed to other groups, but
ultimately the process introduces fluctuations due to finite size. The nature
of these fluctuations is simply binomial on each neuronal locus (assuming
that each one is copied independently), introducing a variance of the order
of ρ(1 − ρ)/N . Thus if there are relatively few groups, the stochasticity is
strong, allowing for a shift in configurations, avoiding impasses and facilitat-
ing the convergence to the optimum. This might lead us to think that fewer
groups would be advantageous for this purpose. However, with very few
groups, this sampling effect overrides selection, no hill climbing can occur,
34
and thus no learning can happen. Hence, it becomes clear that there must be
a compromise between these factors, or even an optimal layer number that
ensures learning but facilitate impasses. (However we are of the opinion that
what limits number of groups are physiological costs, not this trade-off.)
Summarising, there are three central sources of stochasticity: in spiking,
in circuitry and random sampling. The first one is important for short-term
learning, whilst the second and third are important for long-term learning.
4.2 Gating can act as a selective mechanism
It could be argued that we do not require invoking layer copying in order
to solve problems or even impasses. For instance, gating mechanisms could
account for the selective amplification simply by overweighting the outcome
of solutions with larger score. In other words, gating can implement selection
as efficiently as a more complex network to network transmission mechanism.
This is because instead of increasing the number of groups that produce good
solutions, a single group with such a solution is rewarded preferentially.
To visualise the difference we make an analogy. Suppose we are mixing
several paint colours to reach a given tonality. For this purpose we have a
stock of basic colours and a set of pipes to a recipient where the colours are
mixed. Thus we need to pour a given amount a% of base paint A, b% of base
paint B, etc. Suppose that a > b. We can increase the flow of A over B (i)
through a single pipe, by adjusting the tap or (ii) we can use identical pipes
all of which have the same flow, just that for A we use more pipes than for
B. The former case is analogous to a gating mechanism, whereas the second
case is analogous to increasing the representation of the groups.
The direct actor models [14] describe a hill-climbing situation that is most
consistent with this gating mechanism. This is because the rewards, being di-
rected to neuronal activity, do not require a mechanism of neuronal copying.
Rather, it is enough for two (or more) competing agents to adjust output
voltage. This effectively implements hill climbing.
In the case of the di-
35
rect actor, gating mechanisms are conspicuously clear: learning acts directly
on the strength of selection, and the closer a complex is to a solution, the
stronger selection for it becomes, and the weaker the competing complexes
become. Although invoking gating seemingly renders the assumption of neu-
ronal replication obsolete, it also poses some issues. First of all, selection
through gating remains a local learning mechanism that can lead to local fit-
ness peaks resulting on impasses. In fact, if we assume that during copying
the whole content (not just one or few neuronal loci) is transferred to a new
layer, the equations that describe gating or copying are the same. Second, if
we want to consider SSP for long-term learning, then we require invoking a
mechanism that creates and tests novel circuitry, rather than acting on the
standing variation of the available strategies. In essence this mechanism has
to include copying of the group in order to evaluate which circuit is better,
and then discard the worse. Thus in any case, processes of neuronal copying
must exist.
However this criticism does not mean we are rejecting gating as a selec-
tive mean; we just point out that it is not enough to implement the complete
procedure of evolution-enhanced learning. In fact we embrace the possibility
that both mechanisms, namely gating and copying, can act together. This
is not only plausible for finite number of groups, but also an efficient way to
modulate the signal between selection and random sampling [13, 28], poten-
tially facilitating overcoming impasses.
4.3 Size of the neuronal complexes
The complexity of the brain is reflected by the dimension of its constituent
cells (billions of neurons) and by the intricate number of synapses (on the
order of trillions). In some way this accounts for the cognitive capacities of
humans, although how, it is not fully clear. We have presented a hypothesis
that serves as an organising principle for this complexity. However, we have
considered systems that employ as few as 10 neuronal loci on each layer.
36
Although this small number is partly motivated by computational easiness,
there are reasons to think that each layer might not require excessive number
of neurons. First of all, it is well known that the brain is highly modular,
with different neuronal complexes allocated to specific functions. We believe
that some of these modules might be specialized for processing information
in the way we propose, and thus, are expected to be sub-structured into
smaller functional complexes, each of which is constituted by a system of
interconnected groups acting in parallel and competing to solve tasks. Thus,
the actual number of neurons dedicated to any given task depends on the
number of groups that are recruited for processing a given input, not just
on the number of neuronal loci. Second, modular networks also facilitate
rewiring because finding the right network configuration becomes increasingly
harder for larger numbers of neurons. Hence, for an efficient implementation
of SSP, brains might work on a modular way to facilitate rewiring of small
complexes. Third, most complex tasks are likely to be split into subtasks,
each of lower complexity employing relatively small circuits. In this divide
& conquer strategy, these smaller circuits can in turn be included on larger
complexes to accomplish more elaborate tasks.
We may additionally argue that neuronal systems can show invariant
properties, so that larger networks behave in a way similar to smaller net-
works. For instance, distribution of synaptic lifetimes does not depend on
the size of the system, and is therefore invariant. Although we expect the
time required to reach a given solution increases with the size of the system,
we expect this time to scale in a particular way. This scaling may also imply
that more complex problems need to recruit more groups, although this is
not a straightforward requirement.
4.4 The expansion-renormalisation model
Kilgard proposed a verbal theory based on Darwinian dynamics that ac-
counts also for circuitry variation and stabilization, which he termed the
37
expansion-renormalisation model [8]. As mentioned in the introduction, the
ERM assumes the generation of variant circuitry, which results in accelerated
learning. He correctly points out that previous Darwinian frameworks do not
take into consideration mechanisms of variation. Kilgard accounts for such
variation in a verbal model; we have a similar stance regarding these ideas,
but our approach is a formal one. More specifically, unlike Kilgard's work,
we assume specific neuronal rules that account for circuit variability. SSP
constitutes the basic process that allows circuits to be modified. However,
resorting to SSP also demands understanding or assuming factors that drive
SSP with the purpose of circuitry modification. We have assumed two prin-
cipal means for variation of the network structures, namely establishment of
new synapses and disbanding of old synapses. The former follows a higher
instance of Hebb's rule, which simply means that unconnected neurons that
co-spike can become connected. The specific neurophysiological processes
that facilitate rewiring amongst two arbitrary neurons are unknown. Re-
garding the second factor of SSP, the disbanding of existing synapses, we
have introduced a novel idea. That is, we assume that the disbanding prob-
ability decreases with the amount of local information. Moreover, we have
shown that synaptic information is proportional to the square of the synap-
tic weights. This is an interesting result because it relates the mechanistic
aspects to the intuitive notion of neuronal function. Together, these two
mechanisms determine the dynamics of variation after layer replication.
On this line, Fauth et al. [20] propose and analyse a model for the distri-
bution of synapses between two neurons, by studying the interplay between
Hebbian learning and SSP in a similar way we have done. They assume a
constant rate of synaptogenesis for unconnected neurons, unlike the Hebbian-
like mechanism we employ. Synaptic disbanding occurs with probability
P r = po exp(−αφa
ij), where po, a and α are positive constants, which is of
similar form to our's, R (Eqns. 9 and 10). Although in their case this form
is not motivated by information content, they do point out that the topology
38
of the network might constitute the basis of information storage, and that
the role of this storage in memory. The topology, in turn, is determined by
the balance between synaptogenesis and disbanding. Thus, although they do
not explicitly assume that disbanding decreases with information capacity of
a synapse, in the context of our model their model comes down to precisely
that.
Kilgard's ERM hypothesizes that there must be a transient increase of
circuitry variability (expansion) with a subsequent pruning of sub-optimal
synapses, reducing the variation (renormalization). We have not seen evi-
dence for this. Rather, we find an increase of circuitry variability, with an
eventual stabilization. The Monte Carlo implementation allows for persis-
tent fluctuations and occasional 'avalanches', which afterwards recover and
re-establish the network functionality. Although the behaviour of Kilgard's
model and ours is different, we think that the disparity is superficial. The
expansion that creates a standing variation that is later reduced might occur
under specific fitness landscapes and might thus be problem-dependent. In
our case, since the fitness landscape has many equivalent maxima allowing
for equally good solutions, there is no force that generates excess variabil-
ity. However, certain types of non-linarites in the fitness landscape can cer-
tainly lead to that behaviour. It maybe that the 'expansion' phase of the
ERM is not a pervasive feature of neurodynamics, but rather, a context and
stimulus-dependent attribute. However, we must also point out that in our
implementation we only allow for circuit modifications in one layer at a time
(coincidentally, as Fauth et al. [20] do). This circuit might be either copied
to all groups, or discarded altogether. By assuming that many groups can
develop different circuits in the same evaluation round we might neverthe-
less find the expansion phase. (In fact in evolutionary models that allow for
high mutation rates there can be a transient increase of genetic variability,
which is equivalent to the expansion phase of the ERM; e.g. [29]). Finally,
it may well be true that new synapse formation is adaptive in that its rate
39
is increased by the appearance of novel tasks, for which there is some evi-
dence [1,30 -- 32]. This provocation-based mechanism would easily lead to the
expansion phase.
4.5 Towards replicative neurodynamics
In this article we theoretically test a novel mechanism for neural function. We
have found a crucial synergy between learning and fitness climbing, strength-
ening previous, related findings [13]. We have used simple models to show
that the combination of selection and learning is an extremely efficient one.
Moreover, we also showed the relevance of SSP in this context: modifications
of the learning topology of networks. However, there is another open possibil-
ity that can result from a combination of learning and plasticity. We showed
that certain networks are in general terms more efficient learners. Thus, the
recruitment of an existing, efficient network could in principle lead through
synaptic plasticity to the copying of its structure in the current network. In
an analogous way in which DNA is copied, an existing network could replicate
and such a structure would spread, allowing for problem solving. This has
been previously proposed as the neuronal replicator hypothesis [13,25,26,28].
The idea is very recent, reason why there has still been no experimental ver-
ification. But in this article we advanced the mechanisms that justify the
neural replicators. It remains open to study how the copying of the network
topologies can occur.
Related to the issue of exponential strengthening versus exponential repli-
cation, the path evolution algorithm by Fernando et al.. [33] is a remarkable
suggestion.
In that model, neurons along a path are assumed to code for
some behaviour. Whilst neuronal activity is fixed, paths grow collaterals
and thus recruit new nodes. Neuronal activity can spread along different
paths probabilistically that can be evaluated and compared according to
some performance (fitness) measure. Good paths become strengthened by
reward, whereas bad ones are weakened. Various paths can have few or many
40
common neurons. This algorithm explicitly incorporates structural plasticity
and selection and, despite the differences, is thus the closest precedent to our
model. We improve on that model in two respects: we present a mathemat-
ical framework (as opposed to mere simulations) that takes the first steps
to unite theory of learning with that of natural selection, and we consider
recurrent networks that posed a special problem for path evolution.
4.6 Mutations and recombination as creative sources
We should call attention to two possible usage of the term 'mutation' in the
neuronal context. One we have seen before: stochasticity in firing or trans-
mitter release. Another one structural plasticity itself: the term 'synaptic
mutation' was coined by Adams [34] in this latter sense, who by the way fore-
saw the potential importance of the phenomenon for the performance of the
nervous system. Note that [33] in their path evolution model use structural
plasticity to implement 'crossover' between different paths.
Although visually reminiscent to genetic recombination, the synaptic mu-
tations and the path crossovers are not formally equivalent to DNA crossover.
In genetics, recombination does not create new allelic variability (on/off prob-
abilities). Instead, it reshuffles the existing variants at any given locus. This
certainly results in a 'macromutation' at the phenotypic level, but the genetic
variability of the population remains intact. Thus, recombination does not
increase allelic variation, but it does increase variation across groups. The
distinction is important in the context of our work: we assume the equivalent
to high recombination. Namely, at any given neuronal locus, the copying can
occur from any other group, irrespective of the state of the other neuronal
loci. This provides the highest rate of reshuffling, and is thus 'creative'. The
contrary limit is when only the complete content of a selected group can
overwrite an out-selected group. This is an 'asexual' limit in that there is no
recombination. The latter provides the fastest selective response, but is less
creative because it has no combinatorial power.
41
4.7 Relationship to evolvability
Adam's synaptic mutations and Fernando's path-crossover are analogous to
modifications in the architecture of traits. This is a more powerful type of
macro-mutation because it truly modifies the decoding of the information
stored in neuronal states, in an equivalent way as development decodes the
genetic information into a trait, which is one of the fundamental aspects of
evolvability.
Evolvability is understood as the potential of a population to respond to
selection. How fast the response to selection is depends on the amount of
genetic (or heritable) variation that can be produced. This can be given by
standing variation, cryptic variation (due to epistasis, for example), or due
to mutational variance. Although high mutation rates will provide source
'material' to respond to selection, these will also create load that keeps the
population maladapted. However, the optimal scenario is achieved if muta-
tion rates can be increased as selection is started, and tuned down once the
population approaches adaptation.
As we saw above, this is precisely what happens with the neurodynamics
we have described. Of course, genetic systems do not have a learning mecha-
nism as the brain does. Nevertheless, these are analogous. We want to bring
the analogy further and interpret that the input current E as a quantitative
trait, with the weights φ taking the role of additive effects.
A previous model in quantitative genetics has taken an approach reminis-
cent of ours [35]. They did not apply a learning mechanism, but considered
modifier alleles for the mutational effects. These are selected indirectly, in-
creasing the transition rate in the direction of the highest increase in fitness,
in a way that is analogous to our switching probability, which generates
variability in order to improve fitness increase.
42
4.8
Information storage in neuronal circuits
Understanding the relationship between information capacity and synaptic
changes is central in order to understand learning, memory and other aspects
of cognition [3,17]. Under Hebbian learning, information storage relies solely
on the modification of synaptic weights and is contingent on the existing
connections. In the SSP scenario, the information capacity of neuronal com-
plexes is adjusted through modification of the connections [21]. Despite the
advantages of this mechanism in learning, it poses a challenge because of all
possible synapses that can exist between every pair of neurons, only a small
fraction is realised, even within a complex.
Even granting that SPP is a mechanism for the exploration of alternative
circuits, discovering the right synaptic configurations in this vast combina-
torial space is a major problem. Merely grasping the efficiency of the brain
in managing multiple tasks of combinatorial complexity demands an under-
standing of the mechanisms behind such capabilities. In other words, how is
the brain implementing algorithms for searching the combinatorial space of
solutions?
Our model introduces means to explore the combinatorial space, imple-
menting mechanisms that are analogous to biological evolution. From this
perspective, understanding the algorithmic means used by the brain comes
down to understanding and identifying what constitutes the units of selec-
tion and, crucially, what are the units of variability. This has been one of
the central questions of our research, which we have clarified by employing
our analogy with genetic systems.
Understanding the physiological basis of learning and cognition requires
the identification of the modifications that occur during learning, and which
result on specific circuitries directed by different stimuli and experiences.
However, we point out that there is a fundamental relationship between
units of learning and units of variability, which we must consider in order
to understand and identify which are the units of learning.
43
In our model information content is stored not in the activity of neurons,
but in the synaptic weights and switching probabilities. This has two impor-
tant implications. First, this suggests that the loci of memory are circuits
(not neurons), even though the mapping between memory loci and cognitive
functioning might be mediated through the coordinated spiking if individual
neurons (cf. [10]).
However, most importantly, we have shown that SSP is the process that
directly mediates establishment and disbanding of synapses resulting in cir-
cuits that represent solutions to specific problems. Previous findings also
support the notion that information is stored in circuits, not neurons [21,36].
Furthermore, at a higher cognitive level, it has been proposed that conscious-
ness can be gauged through information integration measures between neu-
ronal complexes [16]. At a phenomenological level it is well-known that the
SSP at the level of both spine growth and modifications of the synaptic net-
works is directly inflected by sensory experience [17, 18], or by manipulating
neuromodulators [19]. Despite these lines of evidence, which are compelling
for our theory, we still require and lack direct experimental verification re-
garding the minimal complexity in the circuit distribution that results from
solving particular tasks.
The second important implication is the procedural relationship between
learning and variability. Even when these two are not the same and they
constitute fundamentally two different processes, we have shown how learn-
ing fine-tunes the generation of variability. At the synaptic level, Hebbian
learning modifies switching probabilities, which are the mechanism for gener-
ating variability in spiking. At the level of circuitry, SSP dictates longer-term
changes, where informative synapses persist and uninformative synapses are
disbanded. Altogether these two processes mediate the exploration of the
complex combinatorial space by generating the required variability, guided by
learning. These are the 'fuel' for the motor that results in effective changes,
which are, ultimately, selection mechanisms.
44
Although the question of neuronal and circuit information storage has
been extensively discussed, to our knowledge, we are the first to consider
informational aspects as mechanisms that direct synaptic 'survival' and life-
times (see below). By showing that synaptic information between two neu-
rons is proportional to the learning weight and to the switching probabilities,
we identified a plausible and verifiable mechanism through which SPP can be
directed. This is crucial because we still ignore the causes and consequences
of structural plasticity at a cognitive level. Still our hypothesis is supported
by known facts, chiefly the experiences-driven SSP, where dendritic spine
growth is observed in vivo in the neocortex of adult rats [18, 21, 36].
4.9 The lifetime of synapses
Despite the known potential of neuronal complexes to undergo experience-
driven synaptic network restructuration [37,38], we still ignore how frequently
synaptic connections are modified in adult brains [17]. On the basis that the
number of synapses remains practically constant during adulthood, it has
been long argued that the rates of synapse establishment and disbanding
balance each other. Our result indicates that the length of synaptic lifetimes
follows a statistical distribution, and the question of synaptic stability is a
quantitative one rather than a yes/no statement. Once a stationary state has
been reached, the number of connections remains more or less constant with
an unchanging distribution of synaptic lifetimes. Even then, circuits are by
no mean fixed but, rather, show quite a degree of dynamism.
The SPP model of Fauth et al. [20] studies the equilibrium distribution
of synaptic connections between two neurons. In their work, as in ours, the
general shape of the distribution of synapses is robust to model parame-
ters. However, they analyse synaptic stability between two neurons, not the
general distribution of synapses of a network (whereas they allow multiple
synapses between neuron pairs, we do not take into account this possibility).
Through three-dimensional reconstruction of cultured neocortical cells it
45
was determined that 75% of the connections can be explained by assuming
that pairs of neurons connect randomly. However, the remaining 25% of the
connections requires invoking function, anatomic differences and positioning,
amongst others factors that facilitate chemical mechanisms for attraction or
repulsion in order to complete or avoid the establishment of new synapses
[39].
Direct experimental information regarding the rates of structural plastic-
ity comes from murine models. Although dendritic spines in the cortex of
adult brains tend to be rather stable, a proportion of them show considerable
variation, and experience can induce not only their growth or shrinkage [40]
but also de novo formation or elimination of spines, which, importantly, occur
at balanced rates. Importantly, these structural spine dynamics are known
to be similar in distinct parts of the brain [40 -- 42]. Consequently, it is safe
to assume that the physio-anatomical principles, if not the rates, that main-
tain the distribution of spine size and numbers are similar in distinct cortical
regions .
Although direct observation of synaptogenesis is more elusive (mostly due
to technical limitations), it is known that dendritic spines sometimes result
in new synapses [43]. Changes in axon terminals occur over several weeks in
young brains [44]. A compelling example is that of the visual cortex at early
ages, where synapses established during he developmental critical period are
stable for more than 13 months [45].
5 Numerical and Simulation Methods
5.1 Neuronal dynamics
To simulate the time evolution of a complex of multiple groups (effectively
infinite in number), where each has a fixed number of n of neurons, we solve
a system of coupled ordinary differential equations where n of them represent
the change in spiking probability, and, along with it, a set of differential equa-
46
tions that describe the change in learning weight (i.e. Oja's rule for Hebbian
learning). The number of learning equations depends on the connectivity of
the learning network, which we assume undirected. The initial conditions
for the spiking equations are random deviates from a uniform distribution
between zero and one (unless otherwise stated). The initial conditions for the
learning equations are random deviates from a uniform distribution between
(0, 0.01], unless otherwise stated. The system of equations are solved numer-
ically for t = 10000 time units. (This number ensures convergence of the
system for all parameters used, and overall is considered to be on the order
of ∼ 10 ms. Although we could choose S and other parameters such that the
system is measured in the relevant units, numerically it is more convenient to
employ the mentioned scale). All simulations were implemented and solved
in Mathematica 9.0 and/or 10.0.
5.2 Random networks
To explore learning network topologies, we generated random graphs from
three classes of distributions. The first one is the Erdos-R´enyi model (ER),
where nodes are connected randomly. This model assumes a fixed number of
nodes n and certain probability r that each node is connected to any other
node. The second model is the Barab´asi-Albert (BA), famous for its scale-
free properties. The BA model also employs two parameters that control
the network topology: the fixed number of nodes n and number k of vertices
that are preferentially attached to each node. The third model is the Watts-
Strogratz (WS), or so called small-world networks, which takes as parameters
n nodes and a probability r of rewiring a vertex amongst two nodes in such
a way as to avoid loops.
In all cases we forbid multiple edges and self-
connections. These network models are built-in Mathematica, and employed
as indicated in the software's Documentation Centre.
47
5.3 Structural synaptic plasticity
We assume that Hebbian learning happens faster than SSP, leading to a
separation of time scales. Consequently, we allow the learning equations to
reach equilibrium, and only then modify the learning network. We consider
the three modifications of a network indicated in the Model section. First,
we allow synapses to be formed by adding an edge amongst two unconnected
nodes i and j with a probability qij ∝ ρiρj. In this way, neurons that co-
fire tend to be wired together. Second, we allow synapses to be eliminated
by randomly removing edges between two connected neurons i and j with
probability Rij = exp[−αHij], so that neurons that do not co-fire tend to
be disconnected. Third, we also allow random rewiring (irrespective of firing
probabilities) with a small probability u = 0.01: we randomly and uniformly
choose a connected pair i, j and eliminate the edge, and at the same time
choose an unconnected pair l,m and establish an edge.
In each time step
any (including all) of the above events are allowed to happen. Once the
networks have been rewired, a new round of learning is performed. Initial
conditions may or may not be modified (see Results section). After a new
equilibrium is reached, the new fitness is compared to the fitness before the
rewiring. This is implemented through a Metropolis algorithm: if the fitness
is increased, the change is accepted, but if the fitness decreases, the change
is accepted with probability proportional to the ratio of new to old fitness.
We additionally impose a multiplicative fitness cost per synapse of exp[−kd],
where k is the penalty of each edge in the network, and d is the number of
edges of a given network. We typically run the simulations for at least 250
steps, to ensure convergence. However, in different experiments larger step
numbers are used, as indicated in the figures.
48
6 Acknowledgments
This work was financially supported by the EU FET-open project 'INSIGHT',
agreement number no 308943, by the European Research Council project
'EvoEvo', grant agreement no 294332 and by the Templeton Foundation
project 'Darwin and Hebb', FQEB Grant #RFP-12-19.
7 References
References
[1] Butz M, Worgotter F, Van Ooyen A. Activity-dependent structural
plasticity. Brain Res Rev. 2009;60:287 -- 305.
[2] Holtmaat A, Sovoboda K. Experience-dependent structural plasticity in
the mammalian brain. Nature Reviews Neurosicence. 2009;10:647 -- 658.
[3] Chklovskii DB, Mel BW, Svoboda K. Cortical rewiring and information
storage. Nature. 2004 Oct;431(7010):782 -- 788.
[4] Changeux JP, Courrege P, Danchin A. A Theory of the Epigenesis of
Neuronal Networks by Selective Stabilization of Synapses. Proc Natl
Acad Sci USA. 1973;20:2974 -- 2978.
[5] Changeux JP. Neuronal Man: The Biology of Mind. Princeton, NJ:
Princeton University Press; 1973.
[6] Edelman GM. Neural Darwinism. the theory of neuronal group selection.
Arizona: Basic Books; 1987.
[7] Fernando C, Szathm´ary E, Husbands P. Selectionist and evolutionary
approaches to brain function: a critical appraisal. Front Comput Neu-
rosci. 2012;6(24).
49
[8] Kilgard MP. Harnessing plasticity to understand learning and treat
disease. Trends Neurosci. 2012 Dec;35(12):715 -- 722.
[9] Seung HS. Learning in spiking neural networks by reinforcement of
stochastic synaptic transmission. Neuron. 2003 Dec;40(6):1063 -- 1073.
[10] Rolls ET. Advantages of dilution in the connectivity of attractor net-
works in the brain. Biologically Inspired Cognitive Architectures. 2012
Jul;1(C):44 -- 54.
[11] Crow JF, Kimura M. An Introduction to Population Genetics Theory.
Harper & Row; 1970.
[12] Ullman TD, Goodman ND, Tenenbaum JB. Theory learning as
stochastic search in the language of thought. Cognitive Development.
2012;27(4):455 -- 480.
[13] Fernando C, Goldstein R, Szathm´ary E. The neuronal replicator hy-
pothesis. Neural Comput. 2010 Nov;22(11):2809 -- 2857.
[14] Dayan P, Abbott LF. Theoretical Neuroscience. Cambridge, MA: MIT
Press; 2005.
[15] MacKay DJC. Information theory, inference, and learning algorithms.
Cambridge Univ. Press; 2003.
[16] Tononi G. An information integration theory of consciousness. BMC
Neurosci. 2004 Nov;5(42).
[17] Caroni P, Chowdhury A, Lahr M.
Synapse rearrangements upon
learning: from divergent -- sparse connectivity to dedicated sub-circuits.
Trends Neurosci. 2014 Oct;37(10):604 -- 614.
[18] Trachtenberg JT, Chen BE, Knott GW, Feng GP, Sanes JR, Welker
E, et al. Long-term in vivo imaging of experience-dependent synaptic
plasticity in adult cortex. Nature. 2002;420(6917):788 -- 794.
50
[19] Le B´e JV, Markram H. Spontaneous and evoked synaptic rewiring in
the neonatal neocortex. P Natl Acad Sci USA. 2006 Aug;103(35):13214 --
13219.
[20] Fauth M, Worgotter F, Tetzlaff C. The Formation of Multi-synaptic
Connections by the Interaction of Synaptic and Structural Plastic-
ity and Their Functional Consequences.
PLoS Comp Biol. 2015
Jan;11(1):e1004031.
[21] Poirazi P, Mel BW. Impact of active dendrites and structural plasticity
on the memory capacity of neural tissue. Neuron. 2001 Mar;29(3):779 --
796.
[22] Wright S. Evolution in Mendelian populations. Genetics. 1931;16:97 --
159.
[23] Wright S. The shifting balance theory and macroevolution. Annu Rev
Genet. 1982;16:1 -- 19.
[24] Kemp C, Tenenbaum JB. The discovery of structural form. Proc Natl
Acad Sci USA. 2008 Aug;105(31):10687 -- 10692.
[25] Fernando C, Szathm´ary E. Natural selection in the brain. In: Glatzeder
B, Goel V, von Muller A, editors. Toward a Theory of Thinking. Berlin:
Springer; 2010. p. 291 -- 340.
[26] Fernando C, Szathm´ary E. Chemical, neuronal and linguistic replicators.
In: Pigliucci M, Muller G, editors. Towards an Extended Evolutionary
Synthesis. Cambridge, MA: MIT Press; 2010. p. 209 -- 249.
[27] Werfel JK, Xie X, Seung HS. Learning curves for stochastic gradient
descent in linear feedforward networks. Adv Neural Info Proc Syst.
2004;in press.
51
[28] Fernando C, Karishma KK, Szathm´ary E. Copying and Evolution of
Neuronal Topology. PLoS ONE. 2008;3(11):e3775.
[29] Burger R. The mathematical theory of selection, recombination, and
mutation. Chichester: Wiley; 2000.
[30] Levy WB. Contrasting rules for synaptogenesis, modification of existing
synapses, and synaptic removal as a function of neuronal computation.
Neurocomputing. 2004;5860:343 -- 350.
[31] Butz M, Steenbuck ID, Van Ooyen A. Homeostatic structural plasticity
increases the efficiency of small-world networks. Front Synaptic Neu-
rosci. 2014;6:7.
[32] Butz M, Steenbuck ID, Van Ooyen A. Homeostatic structural plasticity
can account for topology changes following deafferentation and focal
stroke. Front Neuroanat. 2014;8:115.
[33] Fernando C, Vasas V, Szathm´ary E, Husbands P. Evolvable Neuronal
Paths: A Novel Basis for Information and Search in the Brain. PLoS
ONE. 2011 Aug;6(8):e23534.
[34] Adams P. Hebb and Darwin. J Theor Biol. 1998;195(4):419 -- 38.
[35] Jones AG, Arnold S, Burger R. The mutation matrix and the evolution
of evolvability. Evolution. 2007;61(4):727 -- 745.
[36] Stepanyants A, Hof PR, Chklovskii DB. Geometry and structural plas-
ticity of synaptic connectivity. Neuron. 2002 Apr;34(2):275 -- 288.
[37] Buonomano DV, Merzenich MM. ortical plasticity:
from synapses to
maps. Annu Rev Neurosci. 1998;21:149 -- 186.
[38] Knott GW, Quairiaux C, Genoud C, Welker E. Formation of den-
dritic spines with GABAergic synapses induced by whisker stimulation
in adult mice. Neuron. 2002;34:265 -- 273.
52
[39] Hill SL, Wang Y, Riachi I, Schurmann F, Markram H. Statistical con-
nectivity provides a sufficient foundation for specific functional connec-
tivity in neocortical neural microcircuits. Proc Natl Acad Sci USA. 2012
Oct;109(42):E2885 -- 94.
[40] Boothe RG, Greenough WT, Lund JS, Wrege K. A quantitative inves-
tigation of spine and dendrite development of neurons in visual cortex
(area 17) of Macaca nemestrina monkeys. J Comp Neurol. 1979;186:473 --
489.
[41] Bourgeois JP, Goldman-Rakic PS, Rakic P. Synaptogenesis in the pre-
frontal cortex of rhesus monkeys. Cereb Cortex. 1994;4:78 -- 96.
[42] Rakic P, Bourgeois JP, Eckenhoff MF, Zecevic N, Goldman-Rakic PS.
Concurrent overproduction of synapses in diverse regions of the primate
cerebral cortex. Science. 1986;11:232 -- 235.
[43] Toni N, Buchs PA, Nikonenko I, Bron CR, Muller D. LTP promotes
formation of multiple spine synapses between a single axon terminal
and a dendrite. Nature. 1999;402:421 -- 425.
[44] Purves D, Hadley RD. Changes in the dendritic branching of adult
mammalian neurons revealed by repeated imaging in situ. Nature.
1985;315:404 -- 406.
[45] Grutzendler J, Kasthuri N, Gan WB. Long-term dendritic spine stability
in the adult cortex. Nature. 2002 Dec;420(6917):812 -- 816.
53
Figure 7: Neuronal systems under costly synapses
(A) Dynamics of the fitness (relative to the maximum) of neuronal systems
under different cost per connection. Black curves on top:
low costs (k <
0.05). Colour curves:
intermediate costs increasing from k = 0.05 (red) to
k = 0.5 (blue). Grey curves at the bottom: high costs (k > 0.5). Each
curve is a replica of 77 independent simulations. (B) Equilibrium fitness as
a function of the synaptic costs. Inset: Average number of synapses of each
neuron (network degree) against the cost per synapse. (C) Mean switching
probabilities of the connected nodes against synaptic cost. Each point is
an average at the stationary values (last 50 time points) and over the 77
simulations. Parameters: n = 10, T = 7, S = 10, λ = 0.01.
54
0501001502000.000.050.100.150.20IterationMeanfitnessA●●●●●●●●●●●●●●●●0.010.050.100.5010.020.050.10SynapticcostEq.meanfitness●●●●●●●●●●●●●●●●0.010.050.100.5010.00.51.01.52.02.53.03.5MeandegreeB●●●●●●●●●●●●●●●●0.010.050.100.5010.300.320.340.360.380.40CostpersynapseMeanswitchingprobability,〈M〉A Figure 8: Timeline of synapses.
Each row corresponds to a particular synapse between a given pair of neurons.
Colors reflect the synapse length (only for visualization): blue: short lasting
synapses; red: short lasting synapses. In this case the cost of synapses is
k = 0; otherwise as in Fig. 7.
55
02004006008001000TimeSynapse Figure 9: Average lifetime of synapses, measured over one history.
(A) Relationship between average and standard deviation of the synaptic
lifetimes for 14 independent realisations. Colors indicate the synaptic costs
k as in the inset legend. (B) Average lifetime as a function of the cost per
synapse k. Different colors represent 14 different (independent) realisations.
In both panels: squares indicate synapses established at early stages (before
200 iterations) and the bullets: synapses established after reaching stationary
states (later than 200 generations). Parameters as in Fig. 7.
56
40608010012050100150200MeansynapticlifetimeStd.Dev.lifetimeAα=0.01α=0.02α=0.03α=0.06α=0.12α=0.22α=0.40α=0.740.010.050.100.501406080100120140Synapticcost,kSynapticlifetimeBLearningperiodPost-learningstationaryperiod A Hebb's and Oja's rule on the group ensem-
ble
In this appendix we want to show that we can approximate the ensemble of
weights by their mean. We will show that under certain assumptions, Hebb's
and Oja's rules apply on the average, plus some variance terms. We argue
that the variance terms are small and thus the average learning is enough.
This approximation allows a tractable analysis of the neurodynamics without
needing to follow the full distribution of associative weights in the ensemble.
We first work out the simpler case of Hebb's rule (Eq. 1 in the main text)
and then Oja's rule (Eq. 2 in the main text). We interpret the change in the
average weight as the average weight change, i.e.
(cid:28) dφij
(cid:29)
d(cid:104)φij(cid:105)
dt
(cid:39)
dt
,
(12)
where (cid:104). . .(cid:105) denotes average on the ensemble population (for simplicity we will
also use the 'bar' notation, e.g. φ to denote the same quantity. The average
is in principle taken on a joint distribution P[ψ, X], where the bolds indicate
the vector of values (weights and neuronal states). However, we will make
the simplifying approximation that all these quantities are independent. In
population genetics this corresponds to the 'Hardy-Weinberg' equilibrium,
which we take throughout this article.
Mean-field Hebb's rule From Eq. 1 in the main text we have that
dφji
dt
= λ(cid:104)XiYj(cid:105) .
(13)
57
Using the definition of the activity, Yj =(cid:80)n
k(cid:54)=j φjkXk we have that
(cid:104)XiφjkXk(cid:105) +(cid:10)φjiX 2
i
(cid:11)(cid:33)
(cid:32) n(cid:88)
k(cid:54)=j
dφji
dt
= λ
.
(14)
We now use the independency assumption to further develop the two
averages. The first term is
(cid:104)XiφjkXk(cid:105) = (cid:104)Xi(cid:105)(cid:104)φjk(cid:105)(cid:104)Xk(cid:105) = φjk(2ρk − 1)(2ρi − 1)
(15)
(in as long as i (cid:54)= j). The second term is
(cid:10)φjiX 2
i
(cid:11) = (cid:104)φji(cid:105)(cid:10)X 2
i
(cid:11) = φij.
(16)
Now we put together the two expressions, complete the sum to include
terms i and some algebra gives that
(cid:32)
dφji
dt
= λ
(2ρi − 1)
(cid:33)
φjk(2ρk − 1) + φji4ρi(1 − ρi)
n(cid:88)
k(cid:54)=j
(17)
Note that (2ρ − 1) is the average input activity and that 4ρi(1 − ρi) is its
variance. Then
dφji
dt
= λ ¯Xi
¯Yi + λφjivar(Xi),
(18)
where the notation ¯Y is a shorthand for the output of the average activity
(and not the average output, which is a correlation between X and φ).
The last expression shows that, to a first order approximation, Hebb's
rule can be applied to the average. In most cases var(Xi) is small. Moreover,
once the system has substantially or fully learned, then ρ = 0, 1 making the
second term vanish.
(cid:4)
58
Mean-field Oja's rule We proceed in a similar way as above, averaging
over Oja's rule (Eq. 2 in the main text). First note that there are two terms:
= λ(cid:0)(cid:104)XiYj(cid:105) −(cid:10)φjiY 2
(cid:11)(cid:1) ;
j
dφji
dt
(19)
(21)
(22)
(23)
the first terms is given by the right-hand side of Eq. 18. Thus we concentrate
in working out the second term. We begin by expanding the Y 2:
(cid:10)φjiY 2
j
(cid:11) = 2
(cid:88)
k(cid:54)=l(cid:54)=j
(cid:104)φjiφjkφjlXkXl(cid:105) +
jkX 2
k
(20)
(cid:10)φjiφ2
(cid:88)
k(cid:54)=j
(cid:11) .
Now we use the assumption of independency. However, we must note that
not all the terms on the first sum are independent, bacause some symbols
can repeat. We first write
(cid:104)φjiφjkφjl(cid:105) (2ρk − 1)(2ρi − 1) +
(cid:10)φjiφ2
jk
(cid:11)
(cid:88)
k(cid:54)=j
(cid:10)φjiY 2
j
(cid:11) = 2
(cid:88)
k(cid:54)=l(cid:54)=j
Now consider that
(cid:104)φjiφjkφjl(cid:105) =
ji
if k (cid:54)= l (cid:54)= i
if l (cid:54)= k = i
if k (cid:54)= l = i
φjiφjkφjl
(cid:10)φ2
(cid:11) φjl
(cid:11) φjk
(cid:10)φ2
(cid:40)
(cid:10)φ2
(cid:11) if k (cid:54)= i
(cid:11) =
(cid:10)φ3
(cid:11)
if k = i
φji
jk
ji
ji
and in the second sum(cid:10)φjiφ2
jk
This separates the two sums into five sums. The rest of the derivation
It continues by completing the
is somewhat lengthy but straightforward.
59
missing terms of sums and then simply rearranging, and it leads to
j
(cid:10)φjiY 2
where(cid:10)∆φ3
(cid:11) = φjiY
(cid:11) =(cid:10)φ3
ji
ji
2
j + var(φji)X iY j + φji
(cid:11) − φji
(cid:10)φ2
ji
(cid:11).
(cid:88)
(cid:10)φ2
(cid:11) (var(Xk) − 1)
+ φjivar(φji)(var(Xi) − 1) +(cid:10)∆φ3
k(cid:54)=j
jk
(cid:11) (24)
ji
Finally, we note that altogether we can write the mean field on Oja's rule
as the rule of the average activity:
dφji
dt
= λY j(X i − φjiY j) + λ × variance terms.
(25)
(cid:4)
B Mutual information contained in a synapse
Mutual Information is defined as
(cid:88)
r,s∈{0,1}
(cid:20)Pr [Xi = r Xj = s]
(cid:21)
Pr [Xi = r]
H[I, J] =
Pr [Xi = r Xj = s] Pr [Xj = s] log
(26)
The unconditioned probabilities are simply the 'allele' frequencies. Namely,
Pr [Xj = 1] = ρj
Pr [Xj = 0] = 1 − ρj
(27)
(28)
Conditional spiking and switching probabilities. To compute the
conditional probabilities we evaluate the activity of the neuron i with the
60
conditioned value for the neuron j, namely
Yij =
φki (2ρk − 1) + φji (2Xj − 1) .
(cid:88)
k(cid:54)=i,j
This leads to a conditioned switching probability Mij
Mij =(cid:0)1 + exp(cid:2)Yij
(cid:3)(cid:1)−1
(29)
(30)
This conditional switching probability leads to a dynamical equation with the
new activity, whose equilibrium frequencies are Pr [Xi = 1Xj = s] = ρij and
Pr [Xi = 0Xj = s] = 1−ρij. The conditional frequency is by the equilibrium
of
(cid:0)1 − ρij
(cid:1) ∂ρi log(cid:2) ¯W(cid:3) + Mij
(cid:0)1 − 2ρij
(cid:1)
(31)
d ρij
d t
= ρij
For the directional selection case the solution is
ρij =
and for stabilising selection
(cid:113)
S − 2Mij +
2S
4M 2
ij + S2
(cid:32)
1 +
(cid:114)
1 − 4Mij
S
(cid:33)
.
ρij =
1
2
,
(32)
(33)
We evaluate on the evolved weights because we want to measure how one
neuron is dependend on one other. Clearly another possible mutual infor-
mation measure is to allow the whole network to relax with the constraint
(including weights). However, this global measure asseses the capacity of the
whole neuronal complex. By evaluating only the local mutual information
we measure directly the information content contained in it.
61
Note that Yij can be written as
Yij =
φki (2ρk − 1) + φji (2Xj − 1) =
(cid:88)
k(cid:54)=i,j
(cid:88)
k(cid:54)=i
φki (2ρk − 1) + φji (2Xj − 1) − φji (2ρj − 1)
(cid:88)
k(cid:54)=i
=
φki (2ρk − 1) + 2φjiδρ(Xk)
(34)
with δρ(Xk) = Xk − ρk. Consequently, we can express the conditional switch-
ing probability as
Mij = (1 + exp [Yi] exp [2φjiδρj]) −1.
(35)
We assume that both φji and δρ(Xk) are small and expand Mij in series of
δρ, to get
Mij = Mi − 2Mi (1 − Mi) φijδρj + 2Mi (1 − Mi) (1 − 2Mi) φ2
ijδρ2
j
(36)
Since this last expression does not depend explicitly on ρi, the equilibrium
value of ρij is algebraically the same as that of ρi but with Mi → Mij (Eqns.
32 and 33).
62
Approximating mutual information. The sum in Eq. 26 resumes four
terms that we can write explicitly:
H[I, J] = (1 − ρi0) (1 − ρj) log
(cid:20)(1 − ρi0)
(1 − ρi)
+ ρi0 (1 − ρj) log
(cid:21)
(cid:21)
(cid:20)ρi0
(cid:20)(1 − ρi1)
ρi
(1 − ρi)
+ (1 − ρi1)ρj log
(cid:21)
(cid:20)ρi1
(cid:21)
ρi
(37)
+ ρi1ρj log
Substituting Eqns. 32/33 and 36 into the last formula gives the exact
expression for H[I, J], which is lengthy and complicated. Using automated
algebra software (Mathematica 10.0) we perform a series expansion on φij to
second order. Using the fact that S >> M we get that, in both selective
scenarios:
+ O(cid:0)M 2(cid:1)
2MiMjφ2
ij
S2
(38)
H[I, J] =
(cid:4)
C Dependency of the dynamics on the initial
conditions of neuronal spiking
In this appendix we present results that show that once a system learns the
landscape, the retrieval of the information is much quicker than learning, and
it is independent of the initial conditions of the spiking probabilities.
The way we assess this quick retrieval effect is by first solving the full
dynamics. Once in equilibrium, we fix the Hebbian weights and compute the
corresponding switching probabilities Mi at each neuronal locus.
After the learning phase has been completed, we run the dynamics for
63
the neuronal probabilities ρi with this fixed Mi's but by choosing random
initial spiking frequencies (uniformly in the interval (0, 1)). We measure the
state of the system relative to the fixed point of the learning neurodynamics,
ρ∗, and calculate the euclidian distance from this state:
(cid:118)(cid:117)(cid:117)(cid:116) n(cid:88)
Df p =
i − ρi(τ )]2.
[ρ∗
i
Because by time 1/S the ρi have increased in representation, by time τ = S
(1/S < S << 1/λ) they are very close to the fixed point, and we expect Df p
to be very close to zero.
Figure 10A presents the results form simulations where we randomise
initial conditions after learning. The figure shows how the average and the
variance of the distance (over multiple realisations) to the fixed point quickly
decays to zero, and by time τ = S it is already negligible. Figure 10B
shows the mean and variance of a set of ensembles under number of different
neuronal loci n and different fitness landscapes.
D Scaling of selection intensity and associa-
tive weights
We show here the scaling of Fig. 3 in the main text.
First, consider that of all neuronal loci there is one that has the maximum
switching probability, which we denote with the subscript m. We denote the
relative switching probability as Mi, which by the definition of M is
Mi =
Mi
Mm
=
1 + eEm
1 + eEi
.
(39)
64
Figure 10: Effect of initial conditions on the dynamics of learning.
(A) Eucledian distance to the fixed point Df p of the learning dynamics
(red) and the average on an ensemble (100 samples) of dynamics with learnt
weights (black) with random initial conditions for the spiking probabilities.
The shadowed region in grey covers the standard deviation.
In this ex-
ample n = 10 ; the landscape is asymmetric with gradients Si given by
{3.1, 3.2, 4.6, 4.8, 4.8, 4.9, 6.2, 7.8, 8.1, 9.9}. (B) Mean and variance of the dis-
tance to the fixed point Df p of an ensemble of simulations under different
numbers of neuronal loci n (see inset legend) at time τ = S = 10. Each
point is an average over 30 runs with learnt weights and randomised initial
conditions for the spiking probabilities with distribution U (0, 1). For each
of these ensembles the landscape is a random vector with each component
distributed as U (0, S]. Other parameters as in Fig. 3 in the main text.
65
0.010.101101000.00.51.01.52.02.53.0TimeDfpA01.×10-102.×10-103.×10-104.×10-105.×10-106.×10-1001.×10-202.×10-203.×10-204.×10-205.×10-206.×10-20\.08〈Dfp〉\.08Var(Dfp)n=5n=6n=7n=8n=9n=10n=11n=12n=13n=14n=15B As a first approximation we note that for large activity values
Mi (cid:39) exp(Em − Ei) .
(40)
The activity difference ∆i = Em − Ei is:
(cid:88)
j(cid:54)=m
φmj(2ρj − 1) −(cid:88)
k(cid:54)=i
∆i = Em − Ei =
φik(2ρk − 1) .
(41)
The two sums overlap at all but two indices: j = i and k = m, respectively.
Hence
∆i = φmi(2ρi − 1) − φim(2ρm − 1) +
(φmj − φij)(2ρj − 1) .
(42)
(cid:88)
j(cid:54)=m,i
As a second approximation we assume that weights are of the order φ ∼
(n − 1)− 1
2 . Consequently the sum is negligible and
∆i (cid:39) 2
ρi − ρm√
n − 1
.
The equilibrium condition of the neurodynamical equation gives
(cid:16)(cid:112)
ρ =
1
2
4µ2 + 1 − 2µ + 1
(cid:17)
(43)
(44)
where µ = M/S. Assuming that S >> M we have that to first order in µ,
ρ (cid:39) 1 − µ. Thus,
∆i (cid:39) 2
µm − µi
√
n − 1
.
(45)
A third approximation follows: since µm is the term with the largest S,
√
then, compared to most µi, it is negligible. Therefore ∆i (cid:39) −2µi/
n − 1.
Summarising, the relative switching probability is:
Mi (cid:39) exp
− 2µi√
n − 1
.
(46)
(cid:21)
(cid:20)
66
Now, simply call S =
√
(cid:4)
n − 1S/M , so we get the scaled relationship
(cid:104)−1/ S
(cid:105)
Mi (cid:39) exp
.
(47)
E Effect of the topology of the network
The number of neurons in the brain is vast; yet it is very sparsely connected,
with a modular nature. Empirical measurements suggest small world topolo-
gies. However, what lies behind this distribution, and how truly 'random'
it is, remains unclear. From the point of view of our framework, we can
evaluate what effect different topologies have on the neurodynamics.
We can interpret the random topologies in two ways. First, under the
Changeux selective (but not evolvable) scenario where neuronal circuits are
fixed, we can think that from the point of view of an arbitrary cognitive task,
any pre-established circuitry is essentially random. Second, under the 'neu-
ronal replicator hypothesis' , we can think that once some neuronal groups
allocated for a new task, has a current state of the circuitry optimised for a
previous task. However from the point of view of the new task, the current
configuration is random and potentially uninformative.
In this supporting text, we study how random topologies affect the neu-
rodynamics. We consider learning networks that have topologies drawn from
random network distributions, but which remain constant during the learning
process. For any given random topology we run the selection-learning system
on a stabilising landscape with a randomly placed optimum value (which also
remains constant during a run). We consider systems with between 5 to 30
neuronal loci.
We first evaluate the effect of Erdos-Renyi (ER) topologies. These are
random networks built simply by placing random connections amongst two
nodes with a probability p. Although we do not have any reason to believe
67
Figure 11: Distribution of switching probabilities of neuronal circuits
with Erdos-Renyi topologies.
(A) Empirical probability function on the outcome of the selection-learning
process. (B) Relationship between switching probability and the degree of
each neuronal locus. In each simulation we randomise the initial conditions
(as in the figures in the main text), the position of the optimum (uniformly
in the real interval [1, n − 1]) and the learning network (using a random p).
For each n we perform 1000 simulations. λ = 0.01.
that this ER network distribution is realised in neuronal circuits, it is the
most basic pattern of random networks, providing a baseline of uncertainty.
Figure 11A shows the distribution of the equilibrium switching probabilities
M under ER distributions for different number of neuronal loci. The pattern
is clear: most neurons spike in a random, unspecified way. This is reflecting
the fact that the systems are unable to find solutions. Figure 11B shows the
dependency of the switching probabilities against the degree of each neutron
(numbers of synaptic connections with other neurons). Only as the degree
increases some neurons can be more specific. However, generally speaking
the spread the spread of the values is high.
We can compare these patterns with that of fully connected neuronal
circuits (Fig. 12). In this case the switching probabilities are bimodal with
peaks close to M = 0 and M= 1, indicating highly specific firing, as argued
in the main text.
We also perform a similar analysis with Watts-Strogats (WS) 'small
world' topologies. These kinds of networks have become popularised through
68
0.00.20.40.60.81.00.00.20.40.60.81.0Switchingprobability,MCumulativeprobabilityn=5n=10n=15n=20n=25n=30A Figure 12: Distribution of switching probabilities for fully connected
neuronal circuits.
As in figure 11, but with using fully connected networks. n as in the inset of
Fig. 11A.
the "six degrees of separation" motto. The key property of small world net-
works is that any node can be reached from any other by connecting through
a small number of intermediary nodes. Unlike the ER networks which often
have disconnected nodes, the WS networks are very well connected. Some
works have reported that neuronal networks have this kind of properties [].
Figure 13A shows that under this random network models, there is an
almost uniform distribution of switching probabilities M , although there are
almost no circuits with neurons that fire specifically. This occurs because
although the circuits as a whole are well connected, each neuron has only
very few connections (Fig. 13B), which in turn, results in low activities at
most neurons.
Figures 14-15 and shows the neurodynamics assuming a random networks
for the circuit topology. These examples show that in both cases the systems
are unable to learn do to the constraints imposed by the network topology.
We see that most neurons reach a state where they do not fire specifically.
We see this learning process that is unable to learn on an existing circuitry
as a state of impasse as defined in cognitive psychology: any alternative that
can be reached under the current configuration or representation leads to an
even worse 'solution'. Algorithmically, this corresponds to a local optimum
69
0.00.20.40.60.81.00.00.10.20.30.40.50.6Switchingprobability,MCumulativeprobability Figure 13: Distribution of switching probabilities of neuronal circuits
with Watts-Strogatz small world topologies.
As in figure 11, but with Watts-Strogatz deviates with a random rewiring
probability. n as in the inset of Fig. 11A.
on the fitness landscape.1
F Dynamics of structural plasticity under synap-
tic costs
Figure 16 shows the evolution of fitness for neuronal systems targeting dif-
ferent values T . Although the mean fitness decreases with T , this has no
consequences on the outcome of the system. This is because the success of
neuronal group only depends on the relative advantage to other neuronal
groups with different configuration. The relevant pattern is the decrease of
performance as the synaptic costs increase.
G Distribution of the evolved networks
As reported in the main text, the neuronal circuits become less connected
as the cost per synapse k increases. Figure 17 shows that the distribution of
degrees (number of synapses per neuron) becomes more skewed as k becomes
larger.
70
0.00.20.40.60.81.00.00.20.40.60.81.0SwitchingprobabilityCumulativeprobabilityA Figure 14: Neurodynamics under random (Erdos-Renyi) circuit
topologies.
(A) Neurodynamics (inset: fitness) (B) Switching probability (inset: associa-
tive weights). (C) Topology of the neuronal circuit. Parameters as in Figs.
2,4 of the main text.
Figure 15: Neurodynamics under random (Watts-Strogatz) circuit
topologies.
(A) Neurodynamics (inset: fitness) (B) Switching probability (inset: associa-
tive weights). (C) Topology of the neuronal circuit. Parameters as in Figs.
2,4 of the main text.
71
0.010.1011010010001040.00.20.40.60.81.0TimeSpikingprobabilityA0.010.10110100100010400.000010.000020.000030.00004Fitness0.010.1011010010001040.00.20.40.60.81.0TimeSwitchingprobabilityB0.010.101101001000104-0.50.00.51.0Associativeweights,ϕC0.010.1011010010001040.00.20.40.60.81.0TimeSpikingprobabilityA0.010.10110100100010405.×10-60.000010.0000150.000020.000025Fitness0.010.1011010010001040.00.20.40.60.81.0TimeSwitchingprobabilityB0.010.101101001000104-0.50.00.51.0Associativeweights,ϕC Figure 16: Neuronal systems under costly synapses
(A-D) Dynamics of the fitness (relative to the maximum) of neuronal systems
under different cost per connection (see legend in A) and different target val-
ues (A: 4; B: 5; C:10). Each curve is a replica of 16 independent simulations.
(D) Equilibrium fitness as a function of the synaptic costs. Each point is an
average at the stationary values (last 50 time points) and over 16 simula-
tions (except for T = 7 which is the data in Fig 7 int he main text). Filled
diamonds T = 4; open diamonds T = 5; filled bullets T = 7; open bullets
T = 10. Parameters as in Fig. 7 in the main text.
72
0501001502000.010.050.100.501IterationMeanfitnessAα=0.01α=0.02α=0.03α=0.06α=0.12α=0.22α=0.40α=0.740501001502000.010.050.100.501IterationMeanfitnessB0501001502000.010.050.100.501IterationMeanfitnessC◆◆◆◆◆◆◆◆◇◇◇◇◇◇◇◇○○○○○○○○●●●●●●●●●●●●●●●●0.010.050.100.5010.020.050.100.200.50SynapticcostEq.meanfitnessD Figure 17: Degree distribution of the evolved neuronal circuits.
The data corresponds to the simulations of Fig.
in the main text. The
histograms are constructed by taking the last 50 points of 112 simulations
and pooling the degrees of all networks. Since each circuit consists of n = 10
neuronal loci, there are in total 55000 data points in each histogram.
Note that there are always zero classes because the target value T deter-
mines that the optimal circuit consist of T neurons that are fully connected
(particularly with high in-degree) and n − T that are either disconnected or
have zero in-degree. Consequently, because this sets a lower bound for the
number of unconnected neurons. The distribution of degrees is dependent
on both the synaptic cost and the target values (Fig. 18) .
We tried fitting several distributions: Erdos-Renyi, Watts-Strogatz, Pois-
son, Binomial, and scale free networks, and none fits well the empirical dis-
tributions.
73
DegreeProbability02468100.00.20.40.60.81.0k=0.0102468100.00.20.40.60.81.0k=0.01402468100.00.20.40.60.81.0k=0.01802468100.00.20.40.60.81.0k=0.02502468100.00.20.40.60.81.0k=0.03402468100.00.20.40.60.81.0k=0.04602468100.00.20.40.60.81.0k=0.06302468100.00.20.40.60.81.0k=0.08602468100.00.20.40.60.81.0k=0.1202468100.00.20.40.60.81.0k=0.1602468100.00.20.40.60.81.0k=0.2202468100.00.20.40.60.81.0k=0.2902468100.00.20.40.60.81.0k=0.402468100.00.20.40.60.81.0k=0.5402468100.00.20.40.60.81.0k=0.7402468100.00.20.40.60.81.0k=1. Figure 18: Mean degree of the evolved neuronal circuits for different
targets and synaptic costs.
Each point is an average at the stationary values (last 50 time points) and
over 16 simulations (except for T = 7 which is the data in Fig. 7 int he main
text). Filled diamonds T = 4; open diamonds T = 5; filled bullets T = 7;
open bullets T = 10. Parameters as in Fig. 7 in the main text.
74
◆◆◆◆◆◆◆◆◇◇◇◇◇◇◇◇○○○○○○○○●●●●●●●●●●●●●●●●0.010.050.100.50101234Synapticcost,kMeandegree H Distribution of synaptic lifetimes
In this Appendix we present further results on the distribution of synaptic
lifetimes. We ran extensive simulations with an alternative target values
from that on the text, T = 4. (Due to prohibitively long computation times
we chose to sacrifice analysing more target values in order to focus on one
value, but with sufficient statistical power.)
Figure 19 shows results for pooled samples of synaptic lifetimes of 14
simulations each lasting 1000 steps. We partition the synaptic lifetimes into
'early (learning) period', and 'late (post-learning) period'. At the learning
period, there seems to be no relationship between the average lifetime and
the standard deviation of the synaptic lifetimes, whereas for the pos-learning
stationary period there is a clear trend (Fig. 19). This trend seems to be
independent of the target values. The average synaptic lifetimes seem to be
largely independent of the synaptic costs (Fig. 19B).
Despite the fact that there is a clear difference between the synapses
established early and those established late, Mann-Whitney tests on the each
of the different simulations give p−values that are largely variable, and very
often fail to reject the null hypothesis that the distributions of early and late
synapses are distinct (Fig. 20). However, pooling the replicate simulations
leads, in all cases to reject the null hypothesis (Table 2).
75
Figure 19: Synaptic lifetimes for different costs and targets
(A) Relationship between the average and standard deviation of the synaptic
lifetimes. (B) In both panels: blue T = 4, red T = 7; rectangles: synapses
established during early, learning periods (before 200 iterations); bullets:
synapses established in post-learning periods (at or after 200 iterations).
The data is pooled from 14 independent runs each with of 1000 iterations.
Other parameters as in Fig. 9 in the main text.
Cost \Target T = 4
k = 0.010
k = 0.018
k = 0.034
k = 0.063
k = 0.117
k = 0.215
k = 0.398
k = 0.736
Table 2: p−values of pooled simulations
1. × 10−2
4. × 10−5
4. × 10−4
3. × 10−5
3. × 10−9
2. × 10−11
2. × 10−15
5. × 10−12
T = 7
4. × 10−4
2. × 10−2
3. × 10−5
8. × 10−5
5. × 10−8
5. × 10−6
1. × 10−8
1. × 10−6
Test as in Fig. 20 but with pooled samples.
76
40608010012050100150200250MeansynapticlifetimeSTDevsynapticlifetimeA0.010.050.100.501406080100120Synapticcost,kSynapticlifetimeB Figure 20: Significance of the difference of synaptic lifetimes during
and after learning
The plots show the p−values for Mann-Whitney tests (note the log-scale);
all with confidence 0.05. Null hypothesis H0: the mean of the distributions
of synaptic lifetimes during learning and post-learning periods are different.
Alternative hypothesis: Ha the mean of the distributions are different. Marks
at p = 0.01 actually indicate that p ≤ 0.01. The data corresponds to the
simulations of the previous figure (and also Fig. 9 of the main text).
77
|
1805.09999 | 1 | 1805 | 2018-05-25T06:30:50 | Subdivisions of the posteromedial cortex in disorders of consciousness | [
"q-bio.NC"
] | Evidence suggests that disruptions of the posteromedial cortex (PMC) and posteromedial corticothalamic connectivity contribute to disorders of consciousness (DOCs). While most previous studies treated the PMC as a whole, this structure is functionally heterogeneous. The present study investigated whether particular subdivisions of the PMC are specifically associated with DOCs. Participants were DOC patients, 21 vegetative state/unresponsive wakefulness syndrome (VS/UWS), 12 minimally conscious state (MCS), and 29 healthy controls. Individual PMC and thalamus were divided into distinct subdivisions by their fiber tractograpy to each other and default mode regions, and white matter integrity and brain activity between/within subdivisions were assessed. The thalamus was represented mainly in the dorsal and posterior portions of the PMC, and the white matter tracts connecting these subdivisions to the thalamus had less integrity in VS/UWS patients than in MCS patients and healthy controls, as well as in patients who did not recover after 12 months than in patients who did. The structural substrates were validated by finding impaired functional fluctuations within this PMC subdivision. This study is the first to show that tracts from dorsal and posterior subdivisions of the PMC to the thalamus contribute to DOCs. | q-bio.NC | q-bio | Subdivisions of the posteromedial cortex in disorders of consciousness
Running head: Posteromedial cortex subdivision and consciousness
Yue Cui, PhD1,2,3,*, Ming Song, PhD1,2,3,*, Darren M. Lipnicki, PhD4, Yi Yang, MD5,
Chuyang Ye, PhD1,2,3, Lingzhong Fan, MD1,2,3, Jing Sui, PhD1,2,3,6,9, Tianzi Jiang,
PhD1,2,3,7,8,9,#, Jianghong He, MD5,#
1 Brainnetome Center, Institute of Automation, Chinese Academy of Sciences, Beijing
100190, China
2 National Laboratory of Pattern Recognition, Institute of Automation, Chinese
Academy of Sciences, Beijing 100190, China
3 University of Chinese Academy of Sciences, Beijing, China
4 CHeBA (Centre for Healthy Brain Ageing), School of Psychiatry, University of New
South Wales, Randwick, NSW 2031, Australia
5 Department of Neurosurgery, PLA Army General Hospital, Beijing, China
6 The Mind Research Network and Lovelace Biomedical and Environmental Research
Institute, Albuquerque, NM 87106, USA
7 Key Laboratory for NeuroInformation of Ministry of Education, School of Life
Science and Technology, University of Electronic Science and Technology of China,
Chengdu 610054, China
8 Queensland Brain Institute, University of Queensland, Brisbane, QLD 4072,
Australia
1
9 CAS Center for Excellence in Brain Science and Intelligence Technology, Institute
of Automation, Chinese Academy of Sciences, Beijing 100190, China
Corresponding authors:
# Jianghong He, MD, Department of Neurosurgery, PLA Army General Hospital,
Beijing, China (Phone: +86-1-66721961; Fax: +86-1-64057752; E-mail:
[email protected])
# Tianzi Jiang, PhD, Brainnetome Center, Institute of Automation, Chinese Academy
of Sciences, Beijing 100190, China (Phone: +86-1-82544778; Fax: +86-1-82544777;
E-mail: [email protected])
* These authors contributed equally to the manuscript.
2
Abstract
Evidence suggests that disruptions of the posteromedial cortex (PMC) and
posteromedial corticothalamic connectivity contribute to disorders of consciousness
(DOCs). While most previous studies treated the PMC as a whole, this structure is
functionally heterogeneous. The present study investigated whether particular
subdivisions of the PMC are specifically associated with DOCs. Participants were
DOC patients, 21 vegetative state/unresponsive wakefulness syndrome (VS/UWS), 12
minimally conscious state (MCS), and 29 healthy controls. Individual PMC and
thalamus were divided into distinct subdivisions by their fiber tractograpy to each
other and default mode regions, and white matter integrity and brain activity
between/within subdivisions were assessed. The thalamus was represented mainly in
the dorsal and posterior portions of the PMC, and the white matter tracts connecting
these subdivisions to the thalamus had less integrity in VS/UWS patients than in MCS
patients and healthy controls, as well as in patients who did not recover after 12
months than in patients who did. The structural substrates were validated by finding
impaired functional fluctuations within this PMC subdivision. This study is the first to
show that tracts from dorsal and posterior subdivisions of the PMC to the thalamus
contribute to DOCs.
3
Key words: default mode network; diffusion tensor imaging; disorders of
consciousness; parcellation; parietal cortex; persistent vegetative state
4
1. Introduction
Patients surviving severe brain damage may develop a long-term disorder of
consciousness (DOC), such as vegetative state/unresponsive wakefulness syndrome
(VS/UWS) or minimally conscious state (MCS). Evidence from brain imaging studies
suggests that disconnections within thalamocortical areas of the default mode network
(DMN) are implicated in DOCs (Boly et al., 2009; Fernández-Espejo et al., 2011;
Fernandez-Espejo et al., 2012; He et al., 2015; Rosazza et al., 2016; Soddu et al.,
2012; Vanhaudenhuyse et al., 2010). The posteromedial cortex (PMC, i.e.,
precuneus/posterior cingulate cortex) is the structural (Hagmann et al., 2008) and
functional (Utevsky et al., 2014) core of the DMN, and diffusion magnetic resonance
imaging (dMRI) of DOC patients has revealed white matter damage in connections
between the PMC and thalamus (Fernandez-Espejo et al., 2012). A resting state
functional magnetic resonance imaging (fMRI) study found reduced PMC-thalamic
fluctuations in DOC patients compared with healthy controls (Boly et al., 2009),
which accords well with our previous findings of disrupted functional connectivity
between the PMC and thalamus in DOC patients (He et al., 2015). These findings
suggest that the PMC and thalamus play important roles in determining levels of
consciousness. Other studies have shown fMRI or PET can help in predicting
recovery from a DOC. For example, functional connectivity strength discriminated
between DOC patients who regained consciousness within 3 months and those who
did not, with the most discriminative region being the PMC (Wu et al., 2015), and
5
PET predicted long-term recovery (i.e., 12 months after baseline assessment) of
patients with VS/UWS (Stender et al., 2014).
Previous studies of the PMC in DOC patients have investigated this large region as a
whole (Boly et al., 2009; Fernandez-Espejo et al., 2012; He et al., 2015; Laureys et
al., 1999a; Vanhaudenhuyse et al., 2010). However, recent work provides strong
evidence of anatomical diversity and functional heterogeneity in the PMC (Margulies
et al., 2009; Zhang et al., 2014). We previously used whole brain white matter
tractography in healthy participants to reveal particularly strong connections between
dorsal parts of the PMC and the thalamus (Zhang et al., 2014). Dorsal parts of the
PMC are reported to have fundamental roles in self-awareness (Den Ouden et al.,
2005; Kjaer et al., 2002) and in altered state of consciousness (Maquet et al., 1999),
which suggests that, rather than the structure as a whole, particular parts of the PMC
may be specifically involved in determining levels of consciousness. If so, this could
have implications for using neuroimaging to facilitate more accurate diagnoses of
DOCs and more reliably predict patient outcomes. The aim of the present study was
to identify subdivisions of the PMC that may be specifically implicated in DOCs. We
first used fiber tractography within regions of the DMN to identify subdivisions of the
PMC and thalamus. White matter integrity between PMC and thalamus subdivisions,
and brain activity within PMC subdivisions, was then compared across VS/UWS
patients, MCS patients, and healthy controls, and between DOC patients who
recovered after 12 months and those who did not.
6
2. Materials and methods
2.1 Participants
We recruited 79 patients with a DOC (VS/UWS, n = 62; MCS, n = 17) from the PLA
Army General Hospital in Beijing between January 2014 and May 2016. All patients
had received a severe brain injury caused by trauma, anoxia or stroke/cerebrovascular
accident more than one month prior to recruitment. Patients were evaluated at least
twice weekly within the two weeks before baseline, and their diagnosis of either
VS/UWS or MCS was based on the Coma Recovery Scale-Revised (CRS-R)
assessment with the highest diagnostic level during this time, as per Giacino et al.
(2004). Exclusion criteria were having brain damage exceeding 30% of total brain
volume, any contraindications to MRI scanning, being sedated or anesthetized during
MRI acquisition, or having died during follow-up. The duration of follow-up was at
least 12 months after baseline, with clinical examinations using the CRS-R and
Glasgow Outcome Scale (GOS) (Jennett and Bond, 1975). Patients with a GOS level
of 3-5 were considered to have recovered from their DOC (levels 1 and 2 represent
dead and VS/UWS or MCS, respectively) The GOS has good inter-rater reliability of
92% (Wilson et al., 1998) and is commonly used to assess recovery from DOCs (e.g.,
Norton et al., 2012; Perlbarg et al., 2009; Qin et al., 2015; Van Der Eerden et al.,
2014; Wu et al., 2015). The GOS has the advantage of being easy to administer and
has a good inter-rater reliability of 92% (Wilson et al., 1998) that suggests the
7
potential for bias across multiple raters is minimal. Most patients were evaluated at
follow-up as outpatients, with those outside Beijing evaluated by local health care
physicians. After excluding patients with incomplete follow-up assessments (n = 6),
severe motion artifacts and poor quality MRI (n = 27), or failed registration or
preprocessing because of severe structural deformations (n = 13), our study sample
comprised 21 VS/UWS and 12 MCS patients. A group of 29 gender-matched healthy
controls (HCs) were recruited. None of the HCs had a history of psychiatric or
neurological illness, head injury, or drug or alcohol abuse. The study was approved by
the Ethics Committee of the PLA Army General Hospital. All HCs and surrogates for
the DOC patients provided written informed consent.
2.2 MRI data acquisition
All participants underwent dMRI, resting state fMRI and T1-weighted MRI scanning
on a 3 T GE Discovery MR750 scanner (GE Medical Systems). The major acquisition
parameters for dMRI data included: repetition time (TR)/echo time (TE) = 9000/81.9
ms, thickness/gap = 2.0/0 mm, flip angle = 90°, matrix size = 128 × 128 × 75, voxel
size = 2 × 2 × 2 mm3. For each participant, a total of 67 volumes were acquired,
including 3 non-diffusion-weighted volumes (b = 0 s mm-2) and 64 non-collinear
gradient directions (b = 1000 s mm-2). Resting state fMRI data were acquired axially
using an echo-planar imaging sequence sensitive to blood oxygen level-dependent
contrast. The acquisition parameters were TR/TE = 2000/30 ms, thickness/gap =
4.0/0.6 mm, flip angle = 90°, matrix size = 64 × 64 × 39, voxel size = 3.75 × 3.75 × 4
8
mm3, 210 volumes were obtained for each participant. Sagittal T1-weighted structural
MRI scans were obtained using the following optimized acquisition parameters:
TR/TE = 8.16/3.18 ms, thickness/gap = 1.0/0 mm, flip angle = 7°, matrix size = 256 ×
256 × 188; and voxel size = 1 × 1 × 1 mm3.
2.3 MRI data preprocessing
The diffusion MR data were preprocessed using the FSL Diffusion Toolbox (FMRIB
Software Library, http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSL). First, corrections for eddy
current distortions and head motion were performed by aligning all diffusion-
weighted images to the non-diffusion-weighted image (the b0 volume). To ensure
accurate brain masks and anatomical registration, T1-weighted images were processed
with the Computational Anatomy Toolbox (CAT12, Structural Brain Mapping Group,
http://dbm.neuro.uni-jena.de/cat/). The brain mask in diffusion space was then
separated from the skull using the binarized skull-stripped T1-weighted images. Next,
the skull-stripped T1-weighted image was aligned with b0 volume and diffusion
images. In addition, each coregistered T1-weighted image was first linearly then
nonlinearly warped to the Montreal Neurological Institute (MNI) template. The
derived transformation parameters were then inverted and used to warp the seed and
target masks from MNI space to the native dMRI space using a nearest neighbor
interpolation. The registration procedures were performed using Aladin and F3D tools
from the NiftyReg package (https://sourceforge.net/projects/niftyreg/). The diffusion-
9
weighted data for each participant were visually inspected to ensure there were no
apparent artifacts arising from acquisition or data processing procedures.
Resting state fMRI data were preprocessed in the native space of each participant
using Data Processing & Analysis for Brain Imaging toolbox (DPABI) (Yan et al.,
2016). Of the 210 volumes for each participant, the first 10 were discarded and the
remainders underwent slice timing and motion correction, and were resampled to a
stereoscopic 3 mm3. In order to retain the temporal structures of patients with large
head displacement, images with maximum displacement in the cardinal direction > 3
mm, or maximum spin > 3° from the previous frame were treated as outliers and
included as nuisance regressors (Demertzi et al., 2015). In subsequent functional
statistical analysis, we included head motion indexed by mean framewise
displacement (Power et al., 2012) as a confounding factor in the statistical models.
Linear regression was performed to remove the influence of head motion, linear
trends, white matter and CSF signals. Finally, data were temporally filtered to keep
frequencies of 0.01-0.08 Hz, to reduce low-frequency drift and high-frequency noise.
Images were smoothed with a Gaussian filter of full width at half maximum value of
6 mm.
2.4 Structural connectivity-based parcellation of the PMC and thalamus to the
DMN
We investigated 6 ROIs: the PMC, thalamus, superior frontal gyrus (SFG), angular
gyrus, hippocampus, and medial prefrontal cortex (MPFC). Our connectivity-based
10
parcellation scheme is similar to that previously reported by Cunningham et al
(Cunningham et al., 2017). The mask for the PMC was manually drawn on a template
of standard MNI152 space, with a detailed definition of the boundaries described
elsewhere (Zhang et al., 2014). Other masks were obtained either from FSL's
Harvard-Oxford Cortical and Subcortical Structural probabilistic atlases, including
those for the thalamus, hippocampus, SFG and angular gyrus, or from the Stanford
FIND Lab functional ROI database, for the MPFC (Shirer et al., 2012).
In our structural connectivity-based parcellation, either the PMC or thalamus was the
seed, and the remaining 5 ROIs were targets. Probabilistic tractography in FSL was
used to determine the number of streamlines (out of 5000) from each voxel of the
seed region that had a 50% or greater chance of reaching each target ROI. Each voxel
in the seed was assigned to the target ROI that received the highest number of
streamlines from that voxel.
2.5 Structural and functional analysis of consciousness levels at baseline
Probabilistic tractography was performed in each participant's diffusion space to
define pathways between distinct PMC subdivisions to corresponding thalamic
subdivision and DMN regions. Voxels comprising PMC subdivisions were set as
probabilistic tractography seed points (i.e., head region), with the corresponding
thalamic subdivision and other DMN ROIs set as the waypoint, termination, and
classification target (i.e., tail region, and vice versa). To minimize the effect of
spurious connections, mean weighted fractional anisotropy (FA) values for the
11
obtained path connecting head and tail were used to quantify path integrity. A voxel's
weighted FA was calculated as the voxel's FA multiplied by the ratio of streamlines
through the voxel to all streamlines in the path. Tractography is dependent on the
seeding location, and the probability from the head to tail region is not necessarily
equivalent to the probability from the tail to head region (Cao et al., 2013). For this
reason, fiber tracking was performed in both directions between the ROIs, and the
mean weighted FA is the average across these. For the resting state fMRI, we
extracted the voxel-wise fractional amplitude of low-frequency fluctuations (fALFF)
to generate a map for each subject as per Zou et al. (2008) using DPARSFA software.
Individual fALFF maps were then registered into MNI standard space.
We analyzed main effects of group (i.e., VS/UWS, MCS, and HC) for mean weighted
FA, with age and gender as statistical covariates. Functional validation was performed
by comparing fALFF signals from the whole PMC between the VS/UWS and MCS
groups. An AlphaSim approach (Ward, 2000) was used to correct for multiple
comparisons. Potentially confounding factors controlled for in the functional analysis
included age, gender, etiology, duration of illness, and mean framewise displacement
movement factor.
2.6 Structural and functional analysis in the prediction of recovery
To investigate the prognostic capacity of our findings, we compared the non-
recovered (GOS < 3) and recovered (GOS >= 3) groups on the structural integrity of
tracts from the identified PMC subdivisions to the thalamus and other ROIs, and on
12
PMC activity. For this we used general linear models that included age, gender,
etiology and duration of illness as covariates. Statistical significance was set as a two-
tailed p < 0.05.
3. Results
3.1 Demographic and clinical characteristics
Demographic and clinical characteristics of the participants are presented in Table 1
(with individual-level details for the patients in the Supplementary Table). There were
no statistically significant differences in age, gender, or etiology between VS/UWS
and MCS patients, though the VS/UWS patients had a shorter mean illness duration.
At follow-up, 2 of 21 VS/UWS patients and 10 of 12 MCS patients were considered
to have recovered (GOS >= 3).
3.2 Individual parcellation of the PMC and thalamus
The fine-grained architecture of the PMC and thalamus are shown by maximum
probability maps (MPMs) and spaces for representative HC, MCS and VS/UWS
individuals in Figure 1. We observed the thalamus to be represented mainly in the
dorsal and posterior portions of the PMC (dpPMC), including the mesial extent of
Brodmann areas 7 and 31, as well as area v23. The MPFC was represented mainly in
the ventral and anterior parts of the PMC, and the PMC was represented mainly in
regions of the thalamus that included the ventral posterior lateral nucleus (VPL),
pulvinar (Pu) and parts of the centromedian nucleus (CM). The average streamlines
13
from each subregion to corresponding target ROIs showed that structural connections
decreased across the groups from HC to MCS to VS/UWS.
3.3 Structural and functional analysis of consciousness levels at baseline
As shown in Figure 2A, the tract between the dpPMC and VPL/Pu/CM mainly
includes the splenium (posterior) and body of the corpus callosum, the posterior
corona radiata, and retrolenticular part of the internal capsule. The mean weighted FA
between the dpPMC and VPL/Pu/CM was significantly different among VS/UWS,
MCS, and HC groups after controlling for age and gender (F = 56.64, p < 0.001,
Figure 3B), with post hoc analyses showing that both the VS/UWS and MCS patient
groups had less FA than the HC group (p < 0.001), and that VS/UWS patients had less
FA than MCS patients (p = 0.017, Bonferroni corrected value). In addition, we found
VS/UWS patients continued to have less FA than MCS patients (p = 0.036) when
etiology and duration of illness were included as covariates in addition to age and
gender. We also investigated whether the tracts between the ventral and anterior
subdivisions of the PMC and the MPFC differed between groups. These tracts showed
a significant group effect, with post-hoc analyses revealing differences between both
patient groups and the HC group (p < 0.05), but not between the VS/UWS and MCS
groups. Results of fMRI analyses revealed that VS/UWS patients had significantly
lower fALFF values than MCS patients in the dorsal and ventral posterior PMC
(AlphaSim corrected p < 0.05), with the peak location comprising 68.5% of all voxels
being in the dpPMC (Figure 3). The remaining voxels were located in PMC
subdivisions connecting to the MPFC (22.3%) and hippocampus (9.2%).
14
3.4 Structural and functional analysis in the prediction of recovery
Compared to the recovered group (GOS >= 3), the non-recovered group (GOS < 3)
had less mean FA between the dpPMC and VPL/Pu/CM (p = 0.031). No significant
effects were found for tracts between the ventral and anterior PMC and MPFC. The
fALFF values were significantly decreased in the non-recovered group (AlphaSim
corrected p < 0.05, Figure 4B); the peak voxel and 81.4% of the voxels showed
reduced fALFF in the dpPMC. The remaining voxels were located in PMC
subdivisions connecting to the MPFC (11.8%) and hippocampus (6.8%). Specifically,
the fALFF values were significantly greater in recovered than in non-recovered
VS/UWS patients (Figure 4C). To further validate our individual-level parcellation
scheme, we used the MPM derived from our HCs, and a publicly available
Brainnetome Atlas (Fan et al., 2016) with corresponding subregions of the dpPMC,
and projected this into individual space. The results were unchanged from those found
using our individualized parcellation scheme.
4. Discussion
To the best of our knowledge, this is the first study to show that tracts connecting
specific PMC and thalamic subdivisions are implicated in DOCs. We found that
dorsal and posterior portions of the PMC had pathways to the thalamus (VPL, Pu, and
CM) with less white matter integrity, and had less resting state brain activity, in
VS/UWS patients than in MCS patients, as well as in patients who did not recover
after 12 months than in patients who did.
15
There being less PMC activity in VS/UWS patients than in MCS patients is consistent
with previous work (Vanhaudenhuyse et al., 2010), and with evidence suggesting that
the PMC plays a central role in mediating activity within the DMN (Fransson and
Marrelec, 2008). The PMC has been previously reported as the region where activity
is most predictive of whether coma and VS/UWS patients wake within 3 months (Wu
et al., 2015), and seems to have a crucial role in the recovery of consciousness
(Laureys et al., 1999b; Voss et al., 2006) that includes being the first cerebral region
to increase activity in conscious waking from a vegetative state (Andreasen et al.,
1995; Laureys et al., 1999a). A mesocircuit model of consciousness posits that
cortico-thalamic-cortical loop systems involving the precuneus/posterior medial
parietal complex are particularly vulnerable to injury (Giacino et al., 2014; Schiff,
2010) and that a progressive increase in neural activity within these networks leads to
regaining consciousness (Laureys and Schiff, 2012). Our findings support the notion
that disruptions within the corticothalamic system, particularly involving distinct
subdivisions of the PMC and thalamus, underpin DOCs and the emerging from these.
The PMC subdivisions we identified include the dorsal and posterior precuneus and
ventral posterior cingulate cortex (PCC) posterior to the splenium of the corpus
callosum. Previous research has shown important roles for these regions in various
states of consciousness and conscious experience (Kjaer et al., 2002; Northoff and
Bermpohl, 2004), and functional connectivity between the ventral PMC and MPFC is
reported to predict recovery from coma (Silva et al., 2015). Our study extends the
16
known roles of these regions in consciousness, by showing their involvement in
DOCs.
The presence of dense connections between dorsal and posterior parts of the PMC and
the thalamus in HCs has been previously reported (Behrens et al., 2003; Cunningham
et al., 2017), and suggests the capacity of these tracts to convey rich information flow
consistent with a role in the neural network subserving consciousness (Boly et al.,
2009; Fernandez-Espejo et al., 2012). It is thus not surprising that we found disrupted
connectivity between the PMC and thalamus in DOC patients and that this was
greater for those with VS/UWS than with MCS, supporting previous reports of the
structural integrity of these tracts being correlated with CRS-R scores
(Vanhaudenhuyse et al., 2010; Wu et al., 2015), DOC categories (Fernandez-Espejo et
al., 2012), and the recovery of consciousness from VS/UWS (Laureys et al., 2004).
Others have also implicated tracts between the PMC and thalamus in the
pathophysiologic basis of DOCs (Laureys and Schiff, 2012), and in the recovery from
these (Koenig et al., 2014; Norton et al., 2012; Wu et al., 2015). Recovery from a
vegetative state is also more likely with less damage to the corpus callosum (Kampfl
et al., 1998), through which many of the PMC to thalamus tracts pass. These previous
reports on recovery from DOCs fit with our finding of less PMC-thalamic disruption
in DOC patients who recovered within 12 months than in those who did not. Some of
our patients had traumatic head injuries, to which tracts between the PMC and
thalamus are vulnerable (Leclercq et al., 2001) because of mechanical shearing forces
17
(Gennarelli et al., 1982). However, while other of our DOC patients did not have a
traumatic etiology, such cases have been previously reported to nevertheless have
damage in white matter bundles including the corpus callosum (Molteni et al., 2017;
Newcombe et al., 2010; Van Der Eerden et al., 2014) and thalamus (Newcombe et al.,
2010).
We found that the tracts from specific subdivisions of the PMC targeted specific
subdivisions of the thalamus, including the VPL/Pu/CM. Previous evidence also
implicates the thalamus in DOCs. Thalamic lesions are common in VS/UWS patients
(Adams et al., 1999), and more prevalent than in MCS patients (Jennett et al., 2001) in
whom the relative preservation of corticothalamic connections might support
consciousness (Fernández-Espejo et al., 2011) and residual cognitive functions
(Giacino et al., 2014). It is also reported that deep brain stimulation of the bilateral
centromedian-parafascicular complex can improve behavioral responsiveness and
awareness in DOC patients (Chudy et al., 2017), and that a circuit-level mechanism
involving the central thalamus and parietal regions helps regulate awareness (Giacino
et al., 2014). Additionally, the thalamic subdivision we identified includes higher
order nuclei rich in widely projecting matrix cells (Jones, 1998), the activity of which
is thought to represent content-specific neural correlates of consciousness (Koch et
al., 2016).
There are some limitations to the present study. The number of patients included in
18
the analysis is relatively small, and the validity of our findings remains uncertain
without replication in larger samples. A greater ratio of MCS patients to VS/UWS
patients is also needed. It should also be noted that the generalizability of our study
can apply only to patients able to be scanned without head movement artifacts, which
can be especially present in DOC patients (Stender et al., 2014) because of limited or
absent body control. We also did not study patients with severe structural deformation
or damage in regions selected as ROIs. With only one follow-up we were not able to
investigate relationships between neuroimaging markers and signs of transition from
VS/UWS to MCS, and from MCS to normal consciousness. Having more follow-ups
would also facilitate the development of an index or model for predicting the course
and time to recovery.
The present study has shown the fine-grained architecture of the PMC and thalamus
using diffusion tensor tractography, and revealed an involvement of dorsal and
posterior sections of the posteromedial cortex, and the tracts between dpPMC and the
VPL/Pu/CM region of the thalamus, in DOCs. This furthers our understanding of the
neural basis of DOC, and offer potential avenues for treating these. Our findings may
also casts light on the neurological mechanisms underpinning consciousness.
19
References
Adams, J.H., Jennett, B., McLellan, D., Murray, L., Graham, D., 1999. The neuropathology of the
vegetative state after head injury. Journal of clinical pathology 52, 804.
Andreasen, N.C., Leary, D.S., Cizadlo, T., Arndt, S., 1995. Remembering the past: two facets of episodic
memory explored with positron emission tomography. The American Journal of Psychiatry 152, 1576.
Behrens, T.E., Johansen-Berg, H., Woolrich, M.W., Smith, S.M., Wheeler-Kingshott, C.A., Boulby, P.A.,
Barker, G.J., Sillery, E.L., Sheehan, K., Ciccarelli, O., Thompson, A.J., Brady, J.M., Matthews, P.M., 2003.
Non-invasive mapping of connections between human thalamus and cortex using diffusion imaging.
Nature neuroscience 6, 750-757.
Boly, M., Tshibanda, L., Vanhaudenhuyse, A., Noirhomme, Q., Schnakers, C., Ledoux, D., Boveroux, P.,
Garweg, C., Lambermont, B., Phillips, C., Luxen, A., Moonen, G., Bassetti, C., Maquet, P., Laureys, S.,
2009. Functional connectivity in the default network during resting state is preserved in a vegetative
but not in a brain dead patient. Human Brain Mapping 30, 2393-2400.
Cao, Q., Shu, N., An, L., Wang, P., Sun, L., Xia, M.-R., Wang, J.-H., Gong, G.-L., Zang, Y.-F., Wang, Y.-F.,
2013. Probabilistic diffusion tractography and graph theory analysis reveal abnormal white matter
structural connectivity networks in drug-naive boys with attention deficit/hyperactivity disorder. Journal
of Neuroscience 33, 10676-10687.
Chudy, D., Deletis, V., Almahariq, F., Marčinković, P., Škrlin, J., Paradžik, V., 2017. Deep brain stimulation
for the early treatment of the minimally conscious state and vegetative state: experience in 14 patients.
Journal of Neurosurgery, 1-10.
Cunningham, S.I., Tomasi, D., Volkow, N.D., 2017. Structural and functional connectivity of the
precuneus and thalamus to the default mode network. Human Brain Mapping 38, 938-956.
Demertzi, A., Antonopoulos, G., Heine, L., Voss, H.U., Crone, J.S., de Los Angeles, C., Bahri, M.A., Di Perri,
C., Vanhaudenhuyse, A., Charland-Verville, V., 2015. Intrinsic functional connectivity differentiates
minimally conscious from unresponsive patients. Brain 138, 2619-2631.
Den Ouden, H.E., Frith, U., Frith, C., Blakemore, S.-J., 2005. Thinking about intentions. NeuroImage 28,
787-796.
Fan, L., Li, H., Zhuo, J., Zhang, Y., Wang, J., Chen, L., Yang, Z., Chu, C., Xie, S., Laird, A.R., 2016. The human
brainnetome atlas: a new brain atlas based on connectional architecture. Cerebral Cortex 26, 3508-3526.
Fernández-Espejo, D., Bekinschtein, T., Monti, M.M., Pickard, J.D., Junque, C., Coleman, M.R., Owen,
A.M., 2011. Diffusion weighted imaging distinguishes the vegetative state from the minimally conscious
state. NeuroImage 54, 103-112.
Fernandez-Espejo, D., Soddu, A., Cruse, D., Palacios, E.M., Junque, C., Vanhaudenhuyse, A., Rivas, E.,
Newcombe, V., Menon, D.K., Pickard, J.D., Laureys, S., Owen, A.M., 2012. A role for the default mode
network in the bases of disorders of consciousness. Annals of Neurology 72, 335-343.
Fransson, P., Marrelec, G., 2008. The precuneus/posterior cingulate cortex plays a pivotal role in the
default mode network: Evidence from a partial correlation network analysis. NeuroImage 42, 1178-1184.
Gennarelli, T.A., Thibault, L.E., Adams, J.H., Graham, D.I., Thompson, C.J., Marcincin, R.P., 1982. Diffuse
axonal injury and traumatic coma in the primate. Annals of Neurology 12, 564-574.
20
Giacino, J.T., Fins, J.J., Laureys, S., Schiff, N.D., 2014. Disorders of consciousness after acquired brain
injury: the state of the science. Nature Reviews Neurology 10, 99-114.
Giacino, J.T., Kalmar, K., Whyte, J., 2004. The JFK Coma Recovery Scale-Revised: measurement
characteristics and diagnostic utility. Archives of physical medicine and rehabilitation 85, 2020-2029.
Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C.J., Wedeen, V.J., Sporns, O., 2008. Mapping
the structural core of human cerebral cortex. PLoS biology 6, e159.
He, J., Cui, Y., Song, M., Yang, Y., Dang, Y., Jiang, T., Xu, R., 2015. Decreased functional connectivity
between the mediodorsal thalamus and default mode network in patients with disorders of
consciousness. Acta Neurologica Scandinavica 131, 145-151.
Jennett, B., Adams, J.H., Murray, L.S., Graham, D.I., 2001. Neuropathology in vegetative and severely
disabled patients after head injury. Neurology 56, 486-490.
Jennett, B., Bond, M., 1975. Assessment of outcome after severe brain damage: a practical scale. The
Lancet 305, 480-484.
Jones, E.G., 1998. A new view of specific and nonspecific thalamocortical connections. Advances in
neurology 77, 49-71; discussion 72-43.
Kampfl, A., Schmutzhard, E., Franz, G., Pfausler, B., Haring, H.-P., Ulmer, H., Felber, S., Golaszewski, S.,
Aichner, F., 1998. Prediction of recovery from post-traumatic vegetative state with cerebral magnetic-
resonance imaging. The Lancet 351, 1763-1767.
Kjaer, T.W., Nowak, M., Lou, H.C., 2002. Reflective self-awareness and conscious states: PET evidence for
a common midline parietofrontal core. NeuroImage 17, 1080-1086.
Koch, C., Massimini, M., Boly, M., Tononi, G., 2016. Neural correlates of consciousness: progress and
problems. Nature Reviews Neuroscience 17, 307-321.
Koenig, M.A., Holt, J.L., Ernst, T., Buchthal, S.D., Nakagawa, K., Stenger, V.A., Chang, L., 2014. MRI default
mode network connectivity is associated with functional outcome after cardiopulmonary arrest.
Neurocritical care 20, 348-357.
Laureys, S., Goldman, S., Phillips, C., Van Bogaert, P., Aerts, J., Luxen, A., Franck, G., Maquet, P., 1999a.
Impaired effective cortical connectivity in vegetative state: preliminary investigation using PET.
NeuroImage 9, 377-382.
Laureys, S., Lemaire, C., Maquet, P., Phillips, C., Franck, G., 1999b. Cerebral metabolism during
vegetative state and after recovery to consciousness. Journal of Neurology, Neurosurgery & Psychiatry
67, 121-122.
Laureys, S., Owen, A.M., Schiff, N.D., 2004. Brain function in coma, vegetative state, and related
disorders. The Lancet Neurology 3, 537-546.
Laureys, S., Schiff, N.D., 2012. Coma and consciousness: paradigms (re) framed by neuroimaging.
NeuroImage 61, 478-491.
Leclercq, P.D., McKenzie, J.E., Graham, D.I., Gentleman, S.M., 2001. Axonal injury is accentuated in the
caudal corpus callosum of head-injured patients. Journal of Neurotrauma 18, 1-9.
Maquet, P., Faymonville, M.E., Degueldre, C., Delfiore, G., Franck, G., Luxen, A., Lamy, M., 1999.
Functional neuroanatomy of hypnotic state. Biological Psychiatry 45, 327-333.
Margulies, D.S., Vincent, J.L., Kelly, C., Lohmann, G., Uddin, L.Q., Biswal, B.B., Villringer, A., Castellanos,
F.X., Milham, M.P., Petrides, M., 2009. Precuneus shares intrinsic functional architecture in humans and
monkeys. Proceedings of the National Academy of Sciences 106, 20069-20074.
Molteni, E., Rocca, M.A., Strazzer, S., Pagani, E., Colombo, K., Arrigoni, F., Boffa, G., Copetti, M., Pastore,
V., Filippi, M., 2017. A diffusion tensor magnetic resonance imaging study of paediatric patients with
21
severe non‐traumatic brain injury. Developmental Medicine & Child Neurology 59, 199-206.
Newcombe, V.F., Williams, G.B., Scoffings, D., Cross, J., Carpenter, T.A., Pickard, J.D., Menon, D.K., 2010.
Aetiological differences in neuroanatomy of the vegetative state: insights from diffusion tensor imaging
and functional implications. Journal of Neurology, Neurosurgery & Psychiatry 81, 552-561.
Northoff, G., Bermpohl, F., 2004. Cortical midline structures and the self. Trends in cognitive sciences 8,
102-107.
Norton, L., Hutchison, R., Young, G., Lee, D., Sharpe, M., Mirsattari, S., 2012. Disruptions of functional
connectivity in the default mode network of comatose patients. Neurology 78, 175-181.
Perlbarg, V., Puybasset, L., Tollard, E., Lehericy, S., Benali, H., Galanaud, D., 2009. Relation between brain
lesion location and clinical outcome in patients with severe traumatic brain injury: a diffusion tensor
imaging study using voxel‐based approaches. Human Brain Mapping 30, 3924-3933.
Power, J.D., Barnes, K.A., Snyder, A.Z., Schlaggar, B.L., Petersen, S.E., 2012. Spurious but systematic
correlations in functional connectivity MRI networks arise from subject motion. NeuroImage 59, 2142-
2154.
Qin, P., Wu, X., Huang, Z., Duncan, N.W., Tang, W., Wolff, A., Hu, J., Gao, L., Jin, Y., Wu, X., 2015. How are
different neural networks related to consciousness? Annals of Neurology 78, 594-605.
Rosazza, C., Andronache, A., Sattin, D., Bruzzone, M.G., Marotta, G., Nigri, A., Ferraro, S., Rossi
Sebastiano, D., Porcu, L., Bersano, A., 2016. Multimodal study of default‐mode network integrity in
disorders of consciousness. Annals of Neurology 79, 841-853.
Schiff, N.D., 2010. Recovery of consciousness after brain injury: a mesocircuit hypothesis. Trends in
neurosciences 33, 1-9.
Shirer, W., Ryali, S., Rykhlevskaia, E., Menon, V., Greicius, M., 2012. Decoding subject-driven cognitive
states with whole-brain connectivity patterns. Cerebral Cortex 22, 158-165.
Silva, S., De Pasquale, F., Vuillaume, C., Riu, B., Loubinoux, I., Geeraerts, T., Seguin, T., Bounes, V.,
Fourcade, O., Demonet, J.-F., 2015. Disruption of posteromedial large-scale neural communication
predicts recovery from coma. Neurology 85, 2036-2044.
Soddu, A., Vanhaudenhuyse, A., Bahri, M.A., Bruno, M.A., Boly, M., Demertzi, A., Tshibanda, J.F., Phillips,
C., Stanziano, M., Ovadia-Caro, S., Nir, Y., Maquet, P., Papa, M., Malach, R., Laureys, S., Noirhomme, Q.,
2012. Identifying the default-mode component in spatial IC analyses of patients with disorders of
consciousness. Human Brain Mapping 33, 778-796.
Stender, J., Gosseries, O., Bruno, M.-A., Charland-Verville, V., Vanhaudenhuyse, A., Demertzi, A.,
Chatelle, C., Thonnard, M., Thibaut, A., Heine, L., 2014. Diagnostic precision of PET imaging and
functional MRI in disorders of consciousness: a clinical validation study. The Lancet 384, 514-522.
Utevsky, A.V., Smith, D.V., Huettel, S.A., 2014. Precuneus is a functional core of the default-mode
network. Journal of Neuroscience 34, 932-940.
Van Der Eerden, A.W., Khalilzadeh, O., Perlbarg, V., Dinkel, J., Sanchez, P., Vos, P.E., Luyt, C.-E., Stevens,
R.D., Menjot de Champfleur, N., Delmaire, C., 2014. White matter changes in comatose survivors of
anoxic ischemic encephalopathy and traumatic brain injury: comparative diffusion-tensor imaging study.
Radiology 270, 506-516.
Vanhaudenhuyse, A., Noirhomme, Q., Tshibanda, L.J., Bruno, M.A., Boveroux, P., Schnakers, C., Soddu,
A., Perlbarg, V., Ledoux, D., Brichant, J.F., Moonen, G., Maquet, P., Greicius, M.D., Laureys, S., Boly, M.,
2010. Default network connectivity reflects the level of consciousness in non-communicative brain-
damaged patients. Brain : a journal of neurology 133, 161-171.
Voss, H.U., Uluç, A.M., Dyke, J.P., Watts, R., Kobylarz, E.J., McCandliss, B.D., Heier, L.A., Beattie, B.J.,
22
Hamacher, K.A., Vallabhajosula, S., 2006. Possible axonal regrowth in late recovery from the minimally
conscious state. Journal of Clinical Investigation 116, 2005.
Ward, B.D., 2000. Simultaneous inference for fMRI data.
Wilson, J.L., Pettigrew, L.E., Teasdale, G.M., 1998. Structured interviews for the Glasgow Outcome Scale
and the extended Glasgow Outcome Scale: guidelines for their use. Journal of Neurotrauma 15, 573-
585.
Wu, X., Zou, Q., Hu, J., Tang, W., Mao, Y., Gao, L., Zhu, J., Jin, Y., Wu, X., Lu, L., 2015. Intrinsic functional
connectivity patterns predict consciousness level and recovery outcome in acquired brain injury. Journal
of Neuroscience 35, 12932-12946.
Yan, C.-G., Wang, X.-D., Zuo, X.-N., Zang, Y.-F., 2016. DPABI: data processing & analysis for (resting-state)
brain imaging. Neuroinformatics 14, 339-351.
Zhang, Y., Fan, L., Wang, J., Zhu, M., Yu, C., Jiang, T., 2014. Connectivity-based parcellation of the human
posteromedial cortex. Cerebral Cortex 24, 719-727.
Zou, Q.-H., Zhu, C.-Z., Yang, Y., Zuo, X.-N., Long, X.-Y., Cao, Q.-J., Wang, Y.-F., Zang, Y.-F., 2008. An
improved approach to detection of amplitude of low-frequency fluctuation (ALFF) for resting-state fMRI:
fractional ALFF. Journal of neuroscience methods 172, 137-141.
23
Funding
This work was partially supported by the Natural Science Foundation of China (Grant
Nos. 31771076, 81771128, 91432302, 31620103905), the Science Frontier Program
of the Chinese Academy of Sciences (Grant No. QYZDJ-SSW-SMC019), National
Key R&D Program of China (Grant No. 2017YFA0105203), Beijing Municipal
Science & Technology Commission (Grant Nos. Z141107002514111,
Z161100000216152, Z161100000216139), and the Guangdong Pearl River Talents
Plan (2016ZT06S220).
Declaration of interest:
All of the authors declare nothing to report.
24
Figure legends
Figure 1. The seeds and targets for the parcellation of the posteromedial cortex
(PMC, A), and parcellation maps of the PMC (B) and the thalamus (C). First row
shows the maximum probabilistic map for PMC and thalamic subdivisions and ROIs
from a group of healthy controls (HCs) registered to Montreal Neurological Institute
space. Second to fourth rows show parcellation maps and ROIs in individual space for
a representative HC, minimally conscious state (MCS) patient, and a vegetative state /
unresponsive wakefulness syndrome (VS/UWS) patient. Each voxel of the PMC (or
thalamus) was assigned to 1 of 5 target DMN ROIs, which include the thalamus (or
PMC), MPFC, hippocampus, superior frontal gyrus, and angular gyrus, based on the
number of white matter connections to that region. Right column in B and C show
structural connectivity fingerprints of all participants from HC, MCS, and VS/UWS
groups showing the average number of streamlines from subregions connecting to
corresponding targets. Red lines are the average number of streamlines for each
group. DMN, default mode network; dpPMC, dorsal and posterior portions of the
posteromedial cortex; MPFC, medial prefrontal cortex; SFG, superior frontal gyrus;
VPL/Pu/CM, ventral posterior lateral nucleus, pulvinar and centromedian nucleus.
25
26
Figure 2. Structural connectivity analysis of consciousness levels at baseline. A,
group probability maps of reconstructed tracts in healthy controls (HCs) between
dpPMC (dorsal and posterior parts of the posteromedial cortex) and VPL/Pu/CM
(ventral posterior lateral nucleus, pulvinar and center median nucleus). Maps are
thresholded at presence in at least 25% of the participants. B, The mean weighted
fractional anisotropy (FA) between dpPMC and VPL/Pu/CM was significantly
different among VS/UWS, MCS and HC groups, controlling for the effects of age and
gender, and with Bonferroni correction for multiple comparisons. ***, p < 0.001; *, p
< 0.05.
27
Figure 3. Functional activity analysis between vegetative state / unresponsive
wakefulness syndrome (VS/UWS) and minimally conscious state (MCS) patients at
baseline. A, Patients with VS/UWS showed significantly lower fractional amplitude
of low-frequency fluctuations (fALFF) signal values than MCS patients (corrected p <
0.05), with the peak abnormal voxel and 68.5% of abnormal voxels in the subregion
connecting to the thalamus (i.e., dorsal and posterior portions of the PMC, B and C).
28
Figure 4. Structural and functional analysis in the prediction of recovery. A, The
mean weighted fractional anisotropy (FA) was significantly lower in the non-
recovered group (GOS < 3) than the recovered group (GOS >= 3). B, The fractional
amplitude of low-frequency fluctuations (fALFF) signal values were significantly
lower in the non-recovered group than the recovered group, with the peak abnormal
voxel and 81.4% of abnormal voxels in the subregion connecting to the thalamus. C,
The fALFF values were significantly greater in recovered than in non-recovered
VS/UWS groups. Error bars represent SD. *, p < 0.05.
29
Table 1: Demographic and clinical characteristics of the participants
VS/UWS
MCS
HC
Statistics
N = 21
N = 12
N = 29
F/𝜒2/t
p
Age, years
45.43 (14.33) 37.25 (12.76) 37.24 (11.27) F = 2.9
n.s.
Gender (male : female)
8 : 13
8 : 4
15 : 14
Etiology (TBI : CVA : ABI)
7 : 5 : 9
6 : 5 : 1
𝜒2= 1.27 n.s.
𝜒2= 1.73 n.s.
Duration of illness, months
3.38 (1.72)
14.42 (23.36)
t = -2.18
0.04
Baseline CRS-R total
6.05 (1.12)
10.92 (3.85)
t = -5.47 < 0.001
Time interval between baseline
and follow-up, months
21.14 (8.37)
18.42 (12.83)
t = 0.74
n.s.
Follow-up CRS-R total
8.76 (4.40)
17.42 (4.21)
Recovered
Yes (GOS >= 3), N
No (GOS < 3), N
2
19
10
2
t = -5.52 < 0.001
ABI, Anoxic brain injury; CRS-R, Coma Recovery Scale-Revised; CVA,
stroke/cerebrovascular accident; GOS, Glasgow Outcome Scale; HC, healthy
controls; MCS, minimally conscious state; MRI, magnetic resonance imaging; n.s.,
not significant; TBI, traumatic brain injury; VS/UWS, vegetative state/unresponsive
wakefulness syndrome.
Mean and SD are reported unless otherwise specified.
30
Supplementary Table. Detailed demographic and clinical characteristics of patients
Structural lesions
(caudate)
Patient Gender
Age,
y
Group Etiology
time prior to
Duration,
MRI, m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
M
M
39 MCS
ABI
23 MCS
TBI
M
23 MCS
CVA
M
23 MCS
TBI
M
M
M
M
F
F
F
F
F
F
F
M
M
F
F
F
F
61 MCS
ABI
29 MCS
42 MCS
TBI
ABI
53 MCS
ABI
46 MCS
47 MCS
28 MCS
TBI
ABI
TBI
33 MCS
TBI
64 VS/UWS ABI
26 VS/UWS ABI
60 VS/UWS CVA
61 VS/UWS
TBI
51 VS/UWS CVA
38 VS/UWS CVA
30 VS/UWS CVA
58 VS/UWS
TBI
25 VS/UWS CVA
22
M
39 VS/UWS ABI
23
M
41 VS/UWS CVA
F
F
53 VS/UWS ABI
35 VS/UWS CVA
24
25
1
5
3
6
2
9
3
7
77
47
1
12
1
4
4
9
3
2
2
2
3
3
2
3
3
31
Follow-up,
m
12
12
13
13
12
39
12
14
51
12
19
12
17
12
13
16
28
19
38
14
36
12
13
28
19
temporo-parietal
lobe
thalamus, basal ganglia
Brainstem damage
Right basal ganglia
damage
Diffuse cortical & subcortical
atrophy
Diffuse cortical & subcortical
atrophy
Bilateral
damage
Bilateral temporo-parietal damage
Left basal ganglia damage
Diffuse cortical & basal ganglia
(caudate) damage
Diffuse axonal injury
Right basal ganglia damage
Left frontotemporal lobe damage
Bilateral frontal lobe damage and
atrophy
Left
lesions
Left basal ganglia damage
Diffuse
cortical lesions
Right frontotemporal lobe damage
Diffuse cortical & subcortical
atrophy
Diffuse cortical & subcortical
atrophy
Bilateral basal ganglia damage
Diffuse cortical & subcortical
atrophy
Diffuse cortical & subcortical
atrophy
Right basal ganglia damage and
atrophy
Diffuse cortical & subcortical
atrophy
Bilateral brainstem and midbrain
damage
Diffuse cortical & subcortical
hyperintense
anoxic
26
27
28
29
30
31
32
33
M
M
M
M
F
F
F
F
46 VS/UWS
TBI
39 VS/UWS
TBI
22 VS/UWS
TBI
51 VS/UWS
TBI
69 VS/UWS ABI
68 VS/UWS
TBI
35 VS/UWS CVA
43 VS/UWS CVA
4
5
3
3
4
6
3
2
(caudate)
atrophy
Right temporo-parietal damage
Right basal ganglia
damage
Left fronto-temporo-parietal lobe
damage
Right frontotemporal lobe damage
Diffuse cortical & subcortical
atrophy
Diffuse axonal injury
Bilateral basal ganglia damage
Diffuse cortical & subcortical
atrophy
13
16
27
17
33
17
27
29
32
Supplementary Table. Continued
Baseline
Patient
CRS-R
Auditory
Function
Visual
Motor
Function
Function
Scale
Scale
Scale
Oromotor/Verbal
Communication
Arousal
Function Scale
Scale
Scale
4
0
3
3
3
5
0
3
2
1
3
0
1
0
0
0
1
1
0
0
0
1
0
1
0
0
1
1
1
0
1
1
1
3
1
2
1
1
3
1
3
2
1
3
1
1
1
1
1
1
1
0
0
1
1
0
1
1
1
1
1
1
1
1
2
1
17
7
10
9
11
18
7
11
11
8
15
7
7
6
6
6
7
6
4
3
5
7
5
5
6
6
7
7
7
6
7
7
7
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
1
1
1
1
1
1
1
1
2
1
1
2
1
1
1
1
1
1
0
1
0
1
1
1
1
2
1
1
1
1
1
1
1
5
3
2
2
4
5
3
2
2
3
5
2
2
2
2
2
2
1
2
2
2
2
2
2
2
1
2
2
2
2
2
1
2
33
1
0
0
0
0
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
3
2
2
2
2
3
2
2
2
2
3
2
2
2
2
2
2
2
2
0
2
2
2
0
2
2
2
2
2
2
2
2
2
Supplementary Table. Continued
Follow-up
Oromoto
Patient
CRS-R
Auditory
Function
Visual
Motor
r/
Function
Function
Verbal
Scale
Scale
Scale
Function
Communic
ation Scale
Arousal
Scale
GOS
Recov
ered
Scale
4
2
4
4
2
4
4
3
3
2
4
4
2
1
1
1
1
4
0
0
1
1
0
1
1
1
3
3
3
1
1
2
2
4
2
5
4
2
5
1
3
3
2
5
5
3
0
0
2
1
5
2
0
1
1
0
1
1
0
2
3
3
0
3
3
0
19
11
21
19
16
22
19
12
13
13
22
22
11
6
6
8
7
23
7
3
6
7
5
7
8
6
11
15
16
7
9
9
8
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
5
3
5
4
3
6
6
2
2
2
6
5
3
2
2
2
2
6
2
2
2
2
2
2
2
1
2
4
5
2
2
1
2
1
2
2
2
1
2
3
1
2
2
2
3
1
1
1
1
1
3
1
1
0
1
1
1
2
2
1
1
1
2
1
1
1
34
2
0
2
2
1
2
2
2
1
1
2
2
0
0
0
0
0
2
0
0
0
0
0
0
0
0
1
2
1
0
0
0
1
3
2
3
3
2
3
3
3
2
2
3
3
2
2
2
2
2
3
2
1
2
2
2
2
2
2
2
2
3
2
2
2
2
3
2
4
3
3
4
3
3
2
3
4
4
2
2
2
2
2
5
2
2
2
2
2
2
2
2
2
2
3
2
2
2
2
Yes
No
Yes
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
No
No
No
No
No
Yes
No
No
No
No
No
No
No
No
No
No
Yes
No
No
No
No
ABI, Anoxic brain injury; CRS-R, Coma Recovery Scale-Revised; CVA, stroke/cerebrovascular accident;
GOS, Glasgow Outcome Scale; MCS, minimally conscious state; TBI, traumatic brain injury; VS/UWS,
vegetative state/unresponsive wakefulness syndrome.
35
|
1804.08513 | 2 | 1804 | 2018-04-24T09:16:36 | An accumulator model for primes and targets with independent response activation rates: Basic equations for average response times | [
"q-bio.NC"
] | In response priming tasks, speeded responses are performed toward target stimuli preceded by prime stimuli. Responses are slower and error rates are higher when prime and target are assigned to different responses, compared to assignment to the same response, and those priming effects increase with prime-target SOA. Here, we generalize Vorberg et al.'s (2003) accumulator model of response priming, where response activation is first controlled exclusively by the prime and then taken over by the actual target. Priming thus occurs by motor conflict because a response-inconsistent prime can temporarily drive the process towards the incorrect response. While the original model assumed prime and target signals to be identical in strength, we allow different rates of response activation (cf. Mattler & Palmer, 2012; Schubert et al., 2012). Our model predicts that stronger primes mainly increase priming effects in response times and error rates, whereas stronger targets mainly diminish response times and priming effects. | q-bio.NC | q-bio | Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
1
An accumulator model for primes and targets with independent response
activation rates: Basic equations for average response times
arXiv.com/1804.08513 [q-bio.NC]
Thomas Schmidt
Experimental Psychology Unit
University of Kaiserslautern
Kaiserslautern, Germany
Department of General and Experimental Psychology
Filipp Schmidt
University of Giessen
Giessen, Germany
Abstract
In response priming tasks, speeded responses are performed toward target stimuli
preceded by prime stimuli. Responses are slower and error rates are higher when
prime and target are assigned to different responses, compared to assignment to
the same response, and those priming effects increase with prime-target SOA.
Here, we generalize Vorberg et al.'s (2003) accumulator model of response priming,
where response activation is first controlled exclusively by the prime and then
taken over by the actual target. Priming thus occurs by motor conflict because a
response-inconsistent prime can temporarily drive the process towards the
incorrect response. While the original model assumed prime and target signals to
be identical in strength, we allow different rates of response activation (cf. Mattler
& Palmer, 2012; Schubert et al., 2012). Our model predicts that stronger primes
mainly increase priming effects in response times and error rates, whereas stronger
targets mainly diminish response times and priming effects.
1. Introduction
Response priming (Klotz & Neumann, 1999; Klotz & Wolff, 1995; Vorberg,
Mattler, Heinecke, Schmidt, & Schwarzbach, 2003) is a response-conflict paradigm
that serves to study the time-course of sequential response activation (Schmidt et al.,
2011). On each trial, observers respond to a target stimulus by pressing a left or right
key (e.g., to indicate whether the target is red or green; Schmidt, 2000). The target is
preceded by a prime stimulus which is either mapped to the same response as the
target (consistent prime; e.g., red target preceded by red prime) or to the alternative
response (inconsistent prime; e.g., red target preceded by green prime). Typically,
consistent primes will speed responses to the target and produce fewer errors while
inconsistent primes will slow responses and produce more errors. This difference
between response times or error rates in consistent versus inconsistent trials is
labeled the priming effect.
The magnitude of this priming effect critically depends on the stimulus-onset
asynchrony (SOA), that is, the time between the onsets of prime and target.
Generally, response priming occurs for SOAs up to about 100 ms and increases
approximately linearly with SOA (Vorberg et al., 2003). Accumulator models of
response priming (Vorberg et al., 2003; cf. Mattler & Palmer, 2012; Schubert,
Palazova, & Hutt, 2013) assume that the prime starts activating the response assigned
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
2
to it, followed by response activation by the target after the SOA. Depending on
whether the prime is consistent or inconsistent with the target, it will drive the
response process into either the correct or incorrect direction, and may even cause a
response error. The longer the SOA, the more time is available for the prime to
activate the response, and the larger the priming effect in both response times and
error rates (e.g., Vorberg et al., 2003; Klotz, Heumann, Ansorge, & Neumann, 2007;
Leuthold & Kopp, 1998; Schmidt, 2002; Vath & Schmidt, 2007).
Furthermore, the magnitude of response priming depends on the strength of
the prime and the target in activating the response. Prime and target strength can be
controlled by luminance or color contrast (Vath & Schmidt, 2007), familiarity (Kiefer,
Sim, & Wentura, 2015; Pfister, Pohl, Kiesel, & Kunde, 2012), stimulus eccentricity
(Lingnau & Vorberg, 2005), similarity of stimulus alternatives (Schmidt & Schmidt,
2013, 2014), by spatial attention to the prime (Kiefer & Brendel, 2006; Schmidt &
Seydell, 2008), or by feature-based attention to the prime (Schmidt & Schmidt, 2010;
Tapia, Breitmeyer, & Shooner, 2010). Reversely, the prime can be weakened by
degradation of its response-defining features (Schmidt, Weber, & Schmidt, 2014).
Generally, stronger primes lead to larger priming effects, and stronger targets lead to
faster overall responses but smaller priming effects (Schmidt, Haberkamp, &
Schmidt, 2011).
In the following, we will describe our generalization of Vorberg et al.'s (2003)
accumulator model of response priming, and describe the basic behavior of the
model.
2. The free-rate accumulator model
We define two counters, n1(t) and n2(t), gathering evidence for leftward or
rightward responses to a prime followed by a target (Fig. 1). Counters are defined
such that n1(t) gathers information for the response afforded by the prime, and n2(t)
for the opposite response. Therefore, when the prime affords a leftward response,
n1(t) is the leftward counter and n2(t) is the rightward counter. To follow the
accumulation process through time, we distinguish two time periods. Let t = 0 mark
the time when the accumulation process for prime-related information starts, and let
s be the prime-target SOA. Then, for 0 ≤ t < s the process is entirely controlled by the
prime. When the target arrives at time s, it takes over the accumulation process from
the state that it was in at time t = s.
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
3
Fig. 1A: In consistent trials, counter n1 (blue) starts gathering evidence for the response afforded by the
prime. When the target arrives at time s, it proceeds in the same direction at a new rate. Counter n2
(pink) remains inactive because there is no input to which it is tuned. Three different SOAs are shown.
B: In inconsistent trials, the initial activation of n1 decays after time s because there is no further input,
while n2 starts getting active after delay s. C: Responses are registered when the difference between the
counters crosses the threshold ±c. D: Response time effects as a continuous function of SOA, now
relative to target onset. We have to subtract a constant containing all the slack due to conduction
times, nonmotor processing time, mechanical inertia etc. E: Priming function, P(s). F: Response-speed
function, R(s), also defined up to the slack constant.
A counter is incremented by a stimulus to which it is tuned; it is decremented
by random decay. Accumulation and decay occur in discrete counts, and each
increment or decrement is a Poisson event. The resulting behavior is an immigration-
death process, where the count at time t is distributed as a Poisson variable with mean
and variance ni(t). Therefore, when a counter is activated by relevant information, the
expectancy of the accumulated evidence for its respective response is
ni(t)= 1
ν λ(1−e−ν t)
,
where λ (read: lambda) is the rate at which evidence accumulates and ν (read: nu) is
the rate at which it decays. If the counter is no longer activated by any relevant
information after time s, its current activation level ni(s) decays passively to zero
according to
ni(t−s)=ni(s)e−ν(t−s)
.
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
4
(Read t – s as "time since target started to accumulate".) In fact, this is all the
mathematical arsenal we need, and everything else follows from these basic
equations. The logic of the model is the following (Vorberg et al., 2003). Each counter
gathers evidence for its respective response over time. The time when the difference
n1(t) - n2(t) between the counters crosses a threshold c, the response afforded by the
prime is given; when it crosses the threshold –c, the opposite response is given.
In a consistent trial, the prime will first activate n1 at rate λP until time t = s, the
point in time when the prime is replaced by the target in the accumulation process.
Then, starting from time t = s, n1 is further activated, but now at rate λT. Because n2 is
not activated by any stimulus, it remains at resting level, n2(t) = 0. Therefore, for t < s:
ν λP(1−e−ν t)
n1(t)= 1
n2(t)=0 ;
;
and for
t⩾s:
ν λT (1−e−ν(t−s))
n1(t)=n1(s)+ 1
ν λP(1−e−νs)+ 1
= 1
ν λT (1−e−ν(t−s))
= 1
ν [λP(1−e−νs)+λT (1−e−ν(t−s))]
;
n2(t)=0 .
The response is given as soon as the difference n1(t) - n2(t) crosses the threshold
c. To find this crossing time, we just have to set set n1(t) equal to c and solve for time.
Assuming the response time is longer than the prime-target SOA, this time is:
t⩾s:
dcon(t ')=n1(t ')−n2(t ')=c
1
ν [λP(1−e−ν s)+λT (1−e−ν(t'−s))]−0=c
⇒t '=t'con= 1
ν log
λT eνs
λP(1−e−ν s)+λT−νc
,
(Eq. 1a)
where the log symbol denotes the natural logarithm, and the prime in t'
indicates that we are now looking at a particular value of the variable t.
The inconsistent case is more involved. In an inconsistent trial, the prime will
again activate n1 at rate λP until time t = s. However, because the prime is inconsistent
with the target, it activates the wrong counter. When the target arrives at time t = s, it
starts activating the correct counter, n2, at rate λT, while activation in n1 starts to decay
at rate ν. Therefore, for t < s:
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
5
ν λP(1−e−ν t)
n1(t)= 1
n2(t)=0 ;
and for
t⩾s:
n1(t)=n1(s)e−ν(t−s)= 1
= 1
ν λP(eν s−1)e−ν t
;
ν λP(1−e−ν s)e−ν(t−s)
.
ν λT (1−e−ν(t−s))
n2(t)= 1
Because the difference n1(t) - n2(t) starts into the wrong direction, when the
inconsistent target takes over the process, it has to start from state n1(s), which is way
positive, and then move the process all the way down towards the opposite threshold,
–c. It is this time loss that generates the priming effect. Again assuming that the
response time is longer than the prime-target SOA:
t⩾s:
dincon(t')=n1(t ')−n2(t')=−c
1
ν λP(eν s−1)e−νt '− 1
1
ν {[(λP+λT)eνs−λP]e−ν t−λT}=−c
ν λT (1−e−ν(t '−s))=−c
⇒t '=t'incon= 1
ν log
(λP+λT)eνs−λP
λT−νc
.
(Eq. 1b)
Note that t'con and t'incon are defined relative to the onset of the counting process
triggered by the prime. To define them relative to target onset, as is common in
response time research, we have to subtract the SOA, s. Further note that response
times as measured in the experiment involve some additional processing time due to
conduction delays, mechanical slack, and other factors, which we assume to be
constant. Therefore, empirical response times tcon and tincon are related to the model
values by ti = t'i – s + const.
If we define response time as a function of SOA in consistent and inconsistent
trials as C(s) and I(s), respectively, we can decompose the entire response time
pattern into two effects. First, the priming function is defined as
P(s)=I(s)−C (s)=t'incon(s)−t 'con(s)
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
6
= 1
ν log
[(λP+λT)eνs−λP][λP(1−e−νs)+λT−νc]
λT(λT−ν c)eν s
.
(Eq. 2a)
Second, the response speed function is defined as
R(s)= 1
2
[I (s)+C(s)]= 1
2
[t'incon(s)+tcon(s)]
= 1
2 ν
log
λT (λP+λT )e2 ν s−λP λT eνs
(λT−νc)λP(1−e−ν s)+(λT−νc)2
.
(Eq. 2b)
The decomposition can be helpful because λT and λP have separable effects on
response speed and priming. Note that both functions depend on all four
parameters. The functions are related by:
I(s)=R(s)+ 1
2
P(s)
;
C (s)=R(s)− 1
2
P (s)
.
Note that priming effects are symmetric about R(s) simply because R(s) is
defined as the average of I(s) and C(s). This does not imply that inconsistent and
consistent primes induce symmetric costs and benefits in absolute time, which is
generally not the case. Also remember that, just as t'con and t'incon, R(s) is only defined
up to a constant. This assumption of constant residual time may be violated by
several factors, like foreperiod or attentional effects.
3. Basic behavior of the model
•
•
•
•
The four free parameters of the model all have separable effects on the data.
λP (evidence accumulation by the prime) mainly controls the magnitude of
priming effects in both response times and error rates. The larger λP, the
deeper the motor conflict induced by inconsistent primes, and the larger the
priming effect in response times and error rates.
λT (evidence accumulation by the target) both controls the magnitude of
priming effects and the overall speed of responses. The larger λT, the faster the
overall response and the smaller the priming effect (because stronger targets
are quicker in counteracting the effect of an inconsistent prime).
ν (decay rate) controls the curvature of all functions. If response times for
consistent and inconsistent trials depend linearly on SOA, so that the
curvature is low, then ν can be assumed to be small relative to λP and λT.
Larger ν leads to negatively accelerating functions for response times and
priming, and it can also increases overall response times.
c (response threshold) affects the overall response time and the error rate. It is
intuitively clear that a lower response threshold allows inconsistent primes to
induce more errors, and at shorter SOAs.
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
7
Unfortunately, it is quite difficult to model error rates in closed form, and we
did not attempt it here. Instead, we are currently running simulations of the
underlying Poisson process to simulate not only error rates but also response time
distributions. It turns out that if c is set to a high value (e.g., 100 counts, as is
customary in the literature), the stochastic process is so fine-grained that errors at
short SOAs become very unlikely. To have a smooth, linear increase in error rate
with SOA, a very low response threshold is needed (10 counts or even less).
It is important to consider the special case of simultaneous presentation of
primes and targets, because this is the standard setting in some response conflict
paradigms closely related to response priming (like the Eriksen and the Stroop
effect). In our model, setting s to 0 yields identical predictions for consistent and
inconsistent response times that do no longer depend on prime strength:
t'con(s=0)=t'incon(s=0)= 1
ν log (
λT
(λT−νc)
)
,
so that P(s = 0) = 0. (Setting s = 0 in Eq. 3a gives the same result, P(s = 0) = [log
(λT / λT)]/ν = 0.) Simultaneous presentation of primes and targets (e.g., arranged as
flankers) is thus predicted to generate no response priming, contrary to empirical
findings from flanker tasks. In general, temporally overlapping primes and targets
may require explicit modeling of joint response activation, such as a prime and target
that together activate a response at rate λT + λP (consistent trials) or λT – λP
(inconsistent trials) for the time period in question.
4. Recommendations for model fitting
The model has four free parameters: λP, λT, ν, and c. The full power of the
model can be harnessed in a parametric research design where priming functions are
traced through variations in prime or target strength. If only a single priming
function has to be predicted, the full four-parameter model will probably overfit the
data. For that purpose, we recommend restricting the parameter space. First, the
response threshold c can be set to a low, constant value (e.g., 10 counts). Second, if
the response-time and priming functions have little curvature, ν can be set to a small
constant value as well (e.g., 0.001). For this range of values, plausible initial values
for λP and λT are between 0 and 1.
Keep in mind that the parameters have to obey the condition that λT>νc .
Otherwise, the priming function is undefined either because of division by zero or
because of taking the log of a negative value. This side condition makes sense
because νc is an upper limit of how many counts can potentially decay (namely, if the
process is very close to the response threshold). If λT is smaller than that value, then
the rate of decay exceeds the rate of accumulation, and the threshold cannot be
reached.
Consistent and inconsistent response times should always be estimated jointly,
not separately. To do this, observe that t'incon = t'con + P, where P is the priming effect.
Let D be a dummy variable that equals 1 for inconsistent trials and 0 otherwise. Then,
functions can be fitted to t'con + DP(s) in a single step. A similar trick can be used to
test the difference of accumulation between two conditions, for example, two
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
8
different primes or targets. Let D be a dummy variable that equals 1 for condition 1
and 0 otherwise. Then, replacing λ with λ + D(λ + δ) enables us to estimate δ, the
difference in processing rates.
Remember that model fitting and model testing should be separated. One
straightforward technique to cross-validate the fit is to estimate parameters with one
half of the data set (e.g., odd-numbered trials) and then fit it to the other half to do
the statistical test. An even more elegant technique is bootstrapping. Generally,
model parameters should be estimated separately for each participant.
5. Relation to previous models
Our model is based on the general principle of direct parameter specification
(Neumann, 1990), which states that single stimulus parameters can trigger
preprogrammed choice responses without the need for awareness. More specifically,
it is an instance of our chase theory of response priming (Schmidt et al., 2011), which
states that prime and target activate their response as a sequence of neuronal
feedforward sweeps (Lamme & Roelfsema, 2000) independent of visual awareness of
the prime (Schmidt & Vorberg, 2006; Schmidt, 2007). Please keep in mind that our
model only describes the response conflict that occurs for prime-target SOAs up to
about 100 ms. For longer SOAs, additional processes come into play, especially
processes of response inhibition (Panis & Schmidt, 2016; Schmidt, Hauch, & Schmidt,
2015). Response inhibition can lead to reversal of the priming effect, with faster
responses in inconsistent than in consistent trials (negative compatibility effect; Eimer &
Schlaghecken, 1998; Lingnau & Vorberg, 2005).
The response-time model we describe here is a straightforward generalization
of Vorberg et al.'s (2003) accumulator model to arbitrary, independent processing
rates of prime and target. Vorberg et al. (2003) assume that prime and target are
processed by the same rate, λ = λP = λT, and that the system does not distinguish
between primes and targets: When a consistent target appears, it does not refresh the
accumulation process, but the process simply continues at the same rate of
deceleration as if the prime had never changed. This simplification leads to an
extremely economical model: The priming function depends only on ν but is
independent of c and λ. However, since the decay parameter ν is a system constant
that does not depend on the stimulus, the model cannot explain how priming
functions depend on stimulus parameters.
Mattler and Palmer (2012) published a simulation model that makes the same
generalization as ours but assumes that response thresholds become narrower
during the trial. Our explicit model of the response times should basically
correspond to their simulations, but there may be details in their simulation that do
not exactly fit our modeling here. Schubert, Palazova, and Hutt (2013) have
presented a model that makes the exact same generalization we attempt here, but
additionally proposes that the prime processing rate depends on the prime-target
interval. To facilitate comparisons between those models, we use the same
mathematical notation as previous papers.
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
9
References
Eimer, M., & Schlaghecken, F. (1998). Effects of masked stimuli on motor
activation: Behavioral and electrophysiological evidence. Journal of
Experimental Psychology: Human Perception & Performance, 24, 1737-
1747.
Kiefer, M., & Brendel, D. (2006). Attentional modulation of unconscious "automatic"
processes: Evidence from event-related potentials in a masked priming
paradigm. Journal of Cognitive Neuroscience, 18(2), 184-198.
Kiefer, M., Sim, E. J., & Wentura, D. (2015). Boundary conditions for the influence of
unfamiliar non-target primes in unconscious evaluative priming: The
moderating role of attentional task sets. Consciousness and Cognition, 35, 342-
356.
Klotz, W., Heumann, M., Ansorge, U., & Neumann, O. (2007). Electrophysiological
activation by masked primes: Independence of prime-related and target-
related activities. Advances in Cognitive Psychology, 3, 449-465.
Klotz, W., & Neumann, O. (1999). Motor activation without conscious discrimination
in metacontrast masking. Journal of Experimental Psychology: Human Perception
and Performance, 25 (4), 976-992.
Klotz, W., & Wolff, P. (1995). The effect of a masked stimulus on the response to the
masking stimulus. Psychological Research, 58, 92-101.
Lamme, V. A. F., & Roelfsema, P. R. (2000). The distinct modes of vision offered by
feedforward and recurrent processing. Trends in Neurosciences, 23, 571-579.
Leuthold, H., & Kopp, B. (1998). Mechanisms of priming by masked stimuli:
Inferences from event-related brain potentials. Psychological Science, 9, 263-
269.
Lingnau, A., & Vorberg, D. (2005). The time course of response inhibition in
masked priming. Perception & Psychophysics, 67, 545-557.
Mattler, U., & Palmer, S. (2012). Time course of free-choice priming effects explained
by a simple accumulator model. Cognition, 123, 347-360.
Neumann, O. (1990). Direct parameter specification and the concept of perception.
Psychological Research, 52, 207-215.
Panis, S., & Schmidt, T. (2016). What is shaping reaction-time and accuracy
distributions? Active and selective response inhibition causes the Negative
Compatibility Effect. Journal of Cognitive Neuroscience, 28, 1651-1671.
Pfister, R., Pohl, C., Kiesel, A., & Kunde, W. (2012). Your unconscious knows your
name. PLoS One, 7(3), e32402.
Schmidt, T. (2000). Visual perception without awareness: Priming responses by color.
In T. Metzinger (Ed.), Neural Correlates of Consciousness: Empirical and
Conceptual Questions (pp. 157-169). Cambridge, MA: MIT Press.
Schmidt, T. (2007). Measuring unconscious cognition: Beyond the zero-awareness
criterion. Advances in Cognitive Psychology, 3, 275-287.
Schmidt, F., Haberkamp, A., & Schmidt, T. (2011). Dos and don'ts in response
priming research. Advances in Cognitive Psychology, 7, 120-131.
Schmidt, T., Haberkamp, A., Veltkamp, G. M., Weber, A., Seydell-Greenwald, A., &
Schmidt, F. (2011). Visual processing in rapid-chase systems: Image
processing, attention, and awareness. Frontiers in Psychology, 2, Article 169.
Cite as: Schmidt, T., & Schmidt, F. (2018). An accumulator model for primes and targets with
independent response activation rates: Basic equations for average response times.
arXiv:1804.08513 [q-bio.NC]
10
Schmidt, T., Hauch, V., & Schmidt, F. (2015). Mask-triggered thrust reversal in the
Negative Compatibility Effect. Attention, Perception, & Psychophysics, 77, 2377-
2398.
Schmidt, F., & Schmidt, T. (2010). Feature-based attention to unconscious shapes and
colors. Attention, Perception & Psychophysics, 72, 1480-1494.
Schmidt, F., & Schmidt, T. (2013). Grouping principles in direct competition. Vision
Research, 88, 9-21.
Schmidt, F., & Schmidt, T. (2014). Rapid processing of closure and viewpoint-
invariant symmetry: Behavioral criteria for feedforward processing.
Psychological Research, 78, 37-54.
Schmidt, T., & Seydell, A. (2008). Visual attention amplifies response priming of
pointing movements to color targets. Perception & Psychophysics, 70, 443-455.
Schmidt, F., Weber, A., & Schmidt, T. (2014). Activation of response force by self-
splitting objects: Where are the limits of feedforward gestalt processing?
Journal of Vision, 14(9):20, 1-14.
Schubert, T., Palazova, M., & Hutt, A. (2013). The time course of temporal attention
Attention, Perception, &
effects on nonconscious prime processing.
Psychophysics, 75, 1667-1686.
Tapia, E., Breitmeyer, B. G., & Shooner, C. R. (2010). Role of task-directed attention in
nonconscious and conscious response priming by form and color. Journal of
Experimental Psychology: Human Perception and Performance, 36(1), 74.
Vath, N., & Schmidt, T. (2007). Tracing sequential waves of rapid visuomotor
activation in lateralized readiness potentials. Neuroscience, 145, 197-208.
Vorberg, D., Mattler, U., Heinecke, A., Schmidt, T., & Schwarzbach, J. (2003).
Different time courses for visual perception and action priming. Proceedings
of the National Academy of Sciences USA, 100, 6275-6280.
|
1812.06594 | 1 | 1812 | 2018-12-02T15:15:44 | Computational EEG in Personalized Medicine: A study in Parkinson's Disease | [
"q-bio.NC",
"cs.LG",
"eess.SP",
"q-bio.QM",
"stat.ML"
] | Recordings of electrical brain activity carry information about a person's cognitive health. For recording EEG signals, a very common setting is for a subject to be at rest with its eyes closed. Analysis of these recordings often involve a dimensionality reduction step in which electrodes are grouped into 10 or more regions (depending on the number of electrodes available). Then an average over each group is taken which serves as a feature in subsequent evaluation. Currently, the most prominent features used in clinical practice are based on spectral power densities. In our work we consider a simplified grouping of electrodes into two regions only. In addition to spectral features we introduce a secondary, non-redundant view on brain activity through the lens of Tsallis Entropy $S_{q=2}$. We further take EEG measurements not only in an eyes closed (ec) but also in an eyes open (eo) state. For our cohort of healthy controls (HC) and individuals suffering from Parkinson's disease (PD), the question we are asking is the following: How well can one discriminate between HC and PD within this simplified, binary grouping? This question is motivated by the commercial availability of inexpensive and easy to use portable EEG devices. If enough information is retained in this binary grouping, then such simple devices could potentially be used as personal monitoring tools, as standard screening tools by general practitioners or as digital biomarkers for easy long term monitoring during neurological studies. | q-bio.NC | q-bio |
Computational EEG in Personalized Medicine:
A study in Parkinson's Disease
Sebastian Mathias Keller1, Maxim Samarin1, Antonia Meyer2, Vitalii Kosak (Cozak)1,2, Ute
Gschwandtner2, Peter Fuhr2, and Volker Roth1
1University of Basel
2Hospital of the University of Basel
Abstract
Recordings of electrical brain activity carry information about a person's cognitive
health. For recording EEG signals, a very common setting is for a subject to be at
rest with its eyes closed. Analysis of these recordings often involve a dimensionality
reduction step in which electrodes are grouped into 10 or more regions (depending
on the number of electrodes available). Then an average over each group is taken
which serves as a feature in subsequent evaluation. Currently, the most prominent
features used in clinical practice are based on spectral power densities. In our work
we consider a simplified grouping of electrodes into two regions only. In addition
to spectral features we introduce a secondary, non-redundant view on brain activity
through the lens of Tsallis Entropy Sq=2. We further take EEG measurements not
only in an eyes closed (ec) but also in an eyes open (eo) state. For our cohort of
healthy controls (HC) and individuals suffering from Parkinson's disease (PD), the
question we are asking is the following: How well can one discriminate between
HC and PD within this simplified, binary grouping? This question is motivated by
the commercial availability of inexpensive and easy to use portable EEG devices.
If enough information is retained in this binary grouping, then such simple devices
could potentially be used as personal monitoring tools, as standard screening tools
by general practitioners or as digital biomarkers for easy long term monitoring
during neurological studies.
1
Introduction
The recording of the first human EEG was performed in 1924 by German physician Hans Berger.
Since then relative band power has become an established measure to quantify deviations from
normal oscillatory brain activity: Several minutes of EEG signal, usually recorded under resting
state condition and in an 'eyes closed' (ec) setting, are filtered into four to seven non-overlapping
frequency bands covering a range from 0.5Hz up to 70Hz. Then, for each of those bands, the relative
signal power is calculated. Based on this signal power, multiple studies have established a statistical
connection between the neurological condition of their subjects (as measured e.g. by psychological
testing) and the distribution of spectral power within these bands [1], [2], [3].
High density EEG machines available today provide up to 256 individual electrodes recording
brain activity at a frequency of 1000Hz or higher. If one is not interested in performing source
reconstruction or spatial filtering such a high number of electrodes is usually not necessary, especially
as electrodes in proximity to each other are very highly correlated. It is thus expected that features
based on a grouping of electrodes into 10 or more regions as shown in fig. 4b are only slightly
more informative as a grouping into only 2 such regions as shown in fig. 4a. In fact a color coded
representation of the mean Tsallis entropy (cid:104) Sq=2(cid:105) per electrode, as shown in fig. 1, indicates that
the simplest viable partitioning is obtained by grouping electrodes into a frontal and an occipital
Machine Learning for Health (ML4H) Workshop at NeurIPS 2018.
Figure 1: Estimated Tsallis entropy during eo state for the 4-8Hz band (averaged over several minutes)
is shown for every electrode. Most but not all subjects of our cohort fall into this "front versus back"
pattern where entropy is higher in the frontal brain regions. As subjects transition into ec state, frontal
entropy is significantly reduced (not shown). CT068 and CT086 are healthy controls whereas CT016
and CT040 have been diagnosed with Parkinson's disease.
Figure 2: A significant change in brain oscillation can be observed when transitioning from an
eyes open to an eyes closed state and vice versa. This change is known as the "Berger effect" or
"α-blockade" as the amplitude of EEG oscillations in the α-band (8-13 Hz) decreases significantly
when subjects open their eyes. Visible are also the three eye blinks during eyes open state.
region. If these two regions still carry enough information about a persons cognitive health it is
potentially possible that two electrodes located at the front and back of the head would capture enough
information to present a simplified alternative to a much denser electrode setting. Consequentially,
as portable, longterm EEG devices with up to five electrodes are already available, these devices
might prove useful as future long term monitoring tools analog to how ECG devices are used today
to monitor heart rate variability.
1.1 Parkinson's Disease and Dementia
According to the American Psychiatric Association [4], Dementia is a syndrome that consists of a
decline in cognitive and intellectual abilities occurring in an awake and alert patient. As a symptom
complex it is known to occur in over 70 disorders [5], among which Parkinson's Disease (PD) is
the second most common age-related neurodegenerative disorder after Alzheimer's disease. It is
estimated to affect nearly 2 percent of those over age 65. According to the National Parkinson
Foundation (NPF), over the entire course of their illness, 50 to 80 percent of those suffering from
Parkinson's disease eventually develop dementia. The average time from onset of movement problems
to the development of dementia is about 10 years. Accurately diagnosing Parkinson's Disease --
especially in its early stages -- requires experienced practitioners. Thus providing tools for detecting
early changes in brain activity that are as easy to use as taking ones own blood pressure is an important
first step.
2 Materials and Methods
In this study, we focus on how to complement and improve existing signal power based classification
methods by (i) introducing additional entropy based features, (ii) including eo and ec states into the
measurement process and (iii) using the Berger effect (see fig. 2) as a potential feature indicative
of abnormal brain activity. At the same time we simplify the grouping of electrodes and investigate
2
whether a measurement of frontal and occipital brain activity might potentially suffice to indicate a
deviation from healthy brain oscillations.
2.1 Sample Description
We recruited participants for this study from the outpatient clinic for movement disorders of the
Department of Neurology and Neurophysiology of the University Hospital of Basel (Universitätsspital
Basel), Switzerland in the period from 2011 to 2012. We selected 46 persons with Parkinson's disease
and labelled them as "PD". Parkinson's disease was diagnosed according to the United Kingdom
Parkinson's Disease Society Brain Bank criteria. The patients who had dementia (Diagnostic and
Statistical Manual of Mental Disorders, 4th Edition), history of stroke, epilepsy, multiple sclerosis
and surgical interventions to the brain, or/and insufficient knowledge of German language, were
excluded. We selected 16 persons without Parkinson's disease, matched to patients according to sex,
age, and education, and labelled them as "HC". For controls, inclusion criteria were a subjective
report of good health and a neuropsychological examination within normal limits. Exclusion criteria
were identical to those for patients. The local ethics committee approved the study protocol. All
participants provided written informed consent.
2.2 Tsallis Entropy Sq=2
1 −(cid:80)W
i=1 pq
i
The q-entropy, also known as the Tsallis entropy, was proposed in [6] and takes the following form
Sq(pi) =
(1)
with probabilities pi and q ∈ R. In the limit q → 1, eq. 1 recovers the standard Shannon entropy.
In analogy to the Normal distribution, which is the maximum entropy distribution of the Shannon
entropy within the class of distributions with E(x) = µ and E((x − µ)2) = σ2, Tsallis entropy Sq is
maximized by a q-Gaussian distribution. The pdf pqg of a q-Gaussian is given by
q − 1
(cid:20)
2(cid:21)1/(1−q)
pqg(x) = p0
1 − (1 − q)
x
x0
,
(2)
for details see section 3 of [7]. Fig. 5b shows the shape of such q-Gaussians for four different values
of q where q = 1 is the Normal distribution. Higher values of q lead to an increased probability in the
tails of the distribution. Random samplings from a heavy tailed Gaussian with q = 2 are shown in fig.
5a and compared to samples obtained for q = 1. When band filtering a Gaussian white noise signal
and a q-Gaussian (q = 2) white noise signal we find that the signal obtained from the q-Gaussians
has more similarities to a band filtered EEG signal for the same band in terms of amplitude variance
than its Gaussian counterpart. This leads to the conclusion that Tsallis entropy might be a more
appropriate entropy measure for EEG signals. This view is supported by several studies achieving
improved classification results in similar settings based on Tsallis entropy, e.g. [8], [9], [10].
For estimating q-Tsallis entropy for q = 2 an estimator Sq=2 has been proposed in [11]. It is
parameter-free, easy to calculate, designed to estimate the entropy of time series and given by
(cid:80)n
(3)
i=1(xi − µ)2. To estimate the entropy of a time series it is first divided into N bins
with s2
where each extreme point of the signal marks the beginning, resp. the end, of a bin. The variance
within each bin is given by s2
i whereas σ2 is the variance of the entire time series.
Sq=2 = 1 − 1
i = 1
n
σ2
N
(cid:80) s2
i
3 Results and Discussion
EEG measurements in ec and eo states were filtered into four frequency bands (1-4Hz, 4-8Hz, 8-13Hz
and 13-45Hz). For each band Tsallis entropy (ts) for q = 2 and relative band power (bp) was
calculated for the frontal (f) and occipital (b) region. PD vs. HC classification was performed with
repeated cross validation on a train/test split of 80/20 using gradient boosting machines (GBM). Due
to class imbalance the loss function was weighted accordingly and classification was measured using
3
(a) HC: ec → eo: for HC the occip-
ital region has a higher increase in
entropy in the eo state than for PD
(b) PD: ec → eo: for PD the increase
in occipital entropy content is less
pronounced
(c) HC and PD group: the front to
back ratio of entropy in the 4Hz-8Hz
band is higher for healthy controls
Figure 3: 4Hz-8Hz band: HC vs. PD in ec and eo state. In each panel the stars mark the means of the
resp. groups. In panel (c) the crosses mark the extremes for HC and PD (during eo state) calculated
using Archetype analysis [12]. They turn out to be the two most important features in table 1, row 4.
Proximity to either extreme is an indicator of a person's neurological status, i.e. PD or HC.
Table 1: Binary classification with repeated cross validation (80/20) using gradient boosting machines
(gbm) on a 2-region frontal vs. occipital setting with target PD vs. HC. BP is band power, TS is
Tsallis Entropy. Measure is area under the ROC curve. Due to the imbalance between HC and
PD group weighted classification and SMOTE were used but no significant difference in ROC was
obtained.
Classification (gbm)
Feature Set
BP
BP
TS
TS
BP + TS
Ratios(BP, TS)
Ratios Modality
no
no
no
no
no
yes
five most important Features (descending order)
bp.b.ec.4-8, bp.b.ec.13-45, bp.b.ec.1-4, bp.f.ec.1-4, bp.f.ec.4-8
EC
bp.b.eo.1-4, bp.b.eo.4-8, bp.b.eo.13-45, bp.f.eo.13-45, bp.f.eo.4-8
EO
ts.f.ec.1-4, ts.f.ec.4-8, ts.b.ec.8-13, ts.b.ec.13-45, ts.b.ec.4-8
EC
EO
ts.b.eo.4-8, ts.f.eo.4-8, ts.b.eo.13-45, ts.f.eo.8-13, ts.b.eo.8-13
EC + EO bp.b.eo.1-4, bp.f.ec.8-13, ts.f.eo.4-8, ts.b.ec.8-13, ts.b.eo.1-4
EC + EO bp.feo.vs.beo.1-4, ts.fec.vs.bec.8-13, bp.feo.vs.beo.4-8,
BP + TS
yes
EC + EO bp.b.eo.1-4, bp.feo.vs.beo.1-4, bp.feo.vs.beo.4-8,
ts.bec.vs.beo.13-45, bp.fec.bs.bec.1-4
ts.fec.vs.bec.8-13, ts.f.eo.4-8
ROC
64.0%
74.8%
65.5%
71.8%
75.7%
67.0%
71.5%
ROC. Our best result in the binary grouping setting is achieved with a combination of ec and eo
measurements using entropy and band power features. Using only the ratios of band power and
entropy in front and back gives only a moderate classification of ROC=67%. Thus the absolute value
of measured quantities in back and front are meaningful -- not only individual ratios. Combining
ratios and absolute values of power and entropy does not improve overall classification: the possible
benefit of additional features is outweighed by the larger search space. The largest increase in ROC
was obtained when switching from "eyes closed" to "eyes open" measurements. We thus recommend
to make it a standard procedure to record EEG measurements during both ec and especially eo state.
Our best classification result compares favourably -- given the reduced dimensionality -- with results
based on 10 region groupings presented in [13], where ROC=80% is achieved on a similar patient
cohort also from the University Hospital Basel.
Acknowledgment
This project is supported by the the Swiss National Science Foundation project CR32I2-159682
References
[1] Justin Dauwels, François-Benoît Vialatte, and Andrzej Cichocki. On the early diagnosis of alzheimer's
disease from eeg signals: A mini-review, 2011.
4
[2] Laura Bonanni, Astrid Thomas, Pietro Tiraboschi, Bernardo Perfetti, Sara Varanese, and Marco Onofrj.
Eeg comparisons in early alzheimer's disease, dementia with lewy bodies and parkinson's disease with
dementia patients with a 2-year follow-up. Brain : a journal of neurology, 131(Pt 3):690 -- 705, Mar 2008.
doi: 10.1093/brain/awm322.
[3] Martin Hardmeier, Florian Hatz, Habib Bousleiman, Christian Schindler, Cornelis Jan Stam, and Peter
Fuhr. Reproducibility of functional connectivity and graph measures based on the phase lag index (pli) and
weighted phase lag index (wpli) derived from high resolution eeg. PloS one, 9(10):e108648, 2014.
[4] Diagnostic and Statistical Manual of Mental Disorders: DSM-IV. American Psychiatric Association,
Washington, DC, 4 edition, 1994. Committee on Nomenclature and Statistics.
[5] Geldmacher D.S. and Whitehouse P.J. Differential diagnosis of alzheimer's disease. Neurology 48(suppl
6):S2-S9, pages 1125 -- 1138, 1997.
[6] Constantino Tsallis. Possible generalization of boltzmann-gibbs statistics. Journal of Statistical Physics,
52(1):479 -- 487, Jul 1988. ISSN 1572-9613. doi: 10.1007/BF01016429.
[7] Jr. S. Picoli, R. S. Mendes, L. C. Malacarne, and R. P. B. Santos. q-distributions in complex systems: a brief
review. Brazilian Journal of Physics, 39(2a):468 -- 474, 2009. doi: 10.1590/S0103-97332009000400023.
[8] J. McBride, X. Zhao, T. Nichols, V. Vagnini, N. Munro, D. Berry, and Y. Jiang. Scalp eeg-based
discrimination of cognitive deficits after traumatic brain injury using event-related tsallis entropy analysis.
IEEE Transactions on Biomedical Engineering, 60(1):90 -- 96, Jan 2013. ISSN 0018-9294. doi: 10.1109/
TBME.2012.2223698.
[9] T. J. De Bock, S. Das, M. Mohsin, N. B. Munro, L. M. Hively, Y. Jiang, C. D. Smith, D. R. Wekstein,
G. A. Jicha, A. Lawson, J. Lianekhammy, E. Walsh, S. Kiser, and C. Black. Early detection of alzheimer's
disease using nonlinear analysis of eeg via tsallis entropy. In 2010 Biomedical Sciences and Engineering
Conference, pages 1 -- 4, May 2010. doi: 10.1109/BSEC.2010.5510813.
[10] Alberto Capurro, Luis Diambra, Daisy Lorenzo, Macadar O, M.T. Martin, Mostaccio C, A Plastino,
Perez J, Rozman E, Maria Eugenia Torres, and Velluti J. Human brain dynamics: The analysis of eeg
signals with tsallis information measure. Physica A Statistical and Theoretical Physics, 265, 1999. doi:
10.1016/S0378-4371.
[11] Robert Sneddon. The tsallis entropy of natural information. Physica A: Statistical Mechanics and its
Applications, 386(1):101 -- 118, 2007. URL https://EconPapers.repec.org/RePEc:eee:phsmap:v:
386:y:2007:i:1:p:101-118.
[12] Adele Cutler and Leo Breiman. Archetypal analysis. Technometrics, 36(4):338 -- 347, 1994.
[13] M. Chaturvedi, F. Hatz, and U. Gschwandtner. Quantitative eeg (qeeg) measures differentiate parkinson's
disease (pd) patients from healthy controls (hc). Frontiers in aging neuroscience, 9(3), 2017. doi:
10.3389/fnagi.2017.00003.
5
A Appendices
A.1 Simplified Partitioning: 10 regions vs. 2 regions
(a) Electrodes partitioned into frontal and occipital
region
(b) A standard partitioning into 10 brain regions as
used in a clinical setting
Figure 4: The observed division into frontal and occipital region shown in fig. 1 motivates the
simplification of the standard partitioning with its 10 regions to a binary "front vs. back" partitioning
of electrodes.
A.2 Family of q-Gaussians
(a) Random samples from a Normal distri-
bution and from a heavy tailed q-Gaussian.
(b) Within the family of q-Gaussian, the Normal distribu-
tion is recovered for q = 1. Any q-Gaussian with q > 1
has tails with higher probabilities.
Figure 5: In the family of q-Gaussians a distribution with lighter or heavier tails than the Normal
distribution (q = 1) is obtained for q < 1 or q > 1.
6
|
1505.02142 | 2 | 1505 | 2016-02-09T13:23:40 | Porting HTM Models to the Heidelberg Neuromorphic Computing Platform | [
"q-bio.NC",
"cs.NE"
] | Hierarchical Temporal Memory (HTM) is a computational theory of machine intelligence based on a detailed study of the neocortex. The Heidelberg Neuromorphic Computing Platform, developed as part of the Human Brain Project (HBP), is a mixed-signal (analog and digital) large-scale platform for modeling networks of spiking neurons. In this paper we present the first effort in porting HTM networks to this platform. We describe a framework for simulating key HTM operations using spiking network models. We then describe specific spatial pooling and temporal memory implementations, as well as simulations demonstrating that the fundamental properties are maintained. We discuss issues in implementing the full set of plasticity rules using Spike-Timing Dependent Plasticity (STDP), and rough place and route calculations. Although further work is required, our initial studies indicate that it should be possible to run large-scale HTM networks (including plasticity rules) efficiently on the Heidelberg platform. More generally the exercise of porting high level HTM algorithms to biophysical neuron models promises to be a fruitful area of investigation for future studies. | q-bio.NC | q-bio | Porting HTM Models to the Heidelberg Neuromorphic
Computing Platform
Sebastian Billaudelle 1,∗, Subutai Ahmad 2,†
1KirchhoU-Institute for Physics, Heidelberg, Germany
2Numenta, Inc., Redwood City, CA
6
1
0
2
b
e
F
9
]
.
C
N
o
i
b
-
q
[
2
v
2
4
1
2
0
.
5
0
5
1
:
v
i
X
r
a
ABSTRACT
Hierarchical Temporal Memory (HTM) is a computational theory
of machine intelligence based on a detailed study of the neocortex.
The Heidelberg Neuromorphic Computing Platform, developed
as part of the Human Brain Project (HBP), is a mixed-signal
(analog and digital) large-scale platform for modeling networks
of spiking neurons. In this paper we present the Vrst eUort in
porting HTM networks to this platform. We describe a framework
for simulating key HTM operations using spiking network models.
We then describe speciVc spatial pooling and temporal memory
implementations, as well as simulations demonstrating that
the fundamental properties are maintained. We discuss issues
in implementing the full set of plasticity rules using Spike-
Timing Dependent Plasticity (STDP), and rough place and route
calculations. Although further work is required, our initial studies
indicate that it should be possible to run large-scale HTM networks
(including plasticity rules) eXciently on the Heidelberg platform.
More generally the exercise of porting high level HTM algorithms
to biophysical neuron models promises to be a fruitful area of
investigation for future studies.
1 INTRODUCTION
The mammalian brain, particularly that of humans, is able to
process diverse sensory input, learn and recognize complex
spatial and temporal patterns, and generate behaviour based
on context and previous experiences. While computers
are eXcient in carrying out numerical calculations, they
fall short in solving cognitive tasks. Studying the brain
and the neocortex in particular is an important step to
develop new algorithms closing the gap between intelligent
organisms and artiVcial systems. Numenta is a company
dedicated to developing such algorithms and at the same
time investigating the principles of the neocortex. Their
Hierarchical Temporal Memory (HTM) models are designed
to solve real world problems based on neuroscience results
and theories.
EXciently simulating large-scale neural networks in
software is still a challenge. The more biophysical details
a model features,
the more computational ressources
it requires. DiUerent techniques for speeding up the
execution of such implementations exist, e.g. by parallelizing
∗email: [email protected]
†email: [email protected]
calculations. Dedicated hardware platforms are also
being developed. Digital neuromorphic hardware like
the SpiNNaker platform often features highly parallelized
processing architectures and optimized signal routing
[Furber et al., 2014]. On the other hand, analog systems
directly emulate the neuron's behavior in electronic
microcircuits. The Hybrid Multi-Scale Facility (HMF) is
a mixed-signal platform developed in the scopes of the
BrainScaleS Project (BSS) and Human Brain Project (HBP).
In this paper we present eUorts in porting HTM networks
to the HMF. A framework for simulating HTMs based on
spiking neural networks is introduced, as well as concrete
network models for the HTM concepts spatial pooling and
the temporal memory. We compare the behavior to software
implementations in order to verify basic properties of the
HTM networks. We discuss the overall applicability of
these models on the target platform, the impact of synaptic
plasticity, and connection routing considerations.
1.1 Hierarchical Temporal Memory
HTM represents a set of concepts and algorithms for machine
intelligence based on neocortical principles [Hawkins et al.,
2011]. It is designed to learn spatial as well as temporal
patterns and generate predictions from previously seen
sequences. It features continuous learning and operates on
streaming data. An HTM network consists of one or multiple
hierarchically arranged regions. The latter contain neurons
organized in columns. The functional principle is captured
in two algorithms which are laid out in detail in the original
whitepaper [Hawkins et al., 2011]. The following paragraphs
are intended as an introductory overview and introduce the
properties relevant to this work.
The spatial pooler is designed to map a binary input
vector to a set of columns. By recognizing previously seen
input data, it increases stability and reduces the system's
susceptibility for noise. Its behaviour can be characterized
by the following properties:
1. The columnar activity is sparse. Typically, 40 out of 2,048
colums are active, which is approximately a sparsity of
2 %. The number of active columns is constant in each
time step and does not depend on the input sparsity.
2. The spatial pooler activates the k columns which receive
the most input. In case of a tie between two columns,
the active column is selected randomly, e.g. through
1
Billaudelle et al.
structural advantages of certain cells compared to its
neighbors.
3. Stimuli with low pairwise overlap counts are mapped
to sparse columnar representations with low pairwise
overlap counts, while high overlaps are projected onto
representations with high overlap. Thus, similar input
vectors lead to a similar columnar activation, while
disjunct stimuli activate distinct columns.
4. A column must receive a minimum input (e.g. 15 bits) to
become active.
The temporal memory operates on single cells within
columns and further processes the spatial pooler's output.
Temporal sequences are learned by the network and can be
used for generating predictions and highlighting anomalies.
Individual cells receive stimuli from other neurons on their
distal dendrites. This lateral input provides a temporal
context. By modifying a cell's distal connectivity, temporal
sequences can be learned and predicted. The temporal
memory's behavior can be summarized by the following:
1. Individual cells receive lateral input on their distal
dendrites. In case a certain threshold is crossed, the
cells enter a predictive (depolarized) state.
2. When a column becomes active due to proximal input,
it activates only those cells that are in predictive state.
3. When a column with no predicted cells becomes active
due to proximal input, all cells in the column become
active. This phenomenon is referred to as columnar
bursting.
1.2 Heidelberg Neuromorphic Computing Platform
The HMF is a hybrid platform consisting of a traditional
high-performance cluster and a neuromorphic system. It is
developed primarily at the KirchhoU-Institute for Physics in
Heidelberg and the TU Dresden while receiving funding from
the BSS and HBP [HBP SP9 partners, 2014]. The platform's
core is the wafer-scale integrated High Input Count Analog
Neural Network (HICANN) chip as shown in Figure 1. Part
of the chip's unique design is its mixed-signal architecture
featuring analog neuron circuits and a digital communication
infrastructure. Due to the short intrinsic time constants of
the hardware neurons, the system operates on an accelerated
timescale with a speed-up factor of 10 × 104 compared to
biological real-time.
HICANN features 512 neurons or dendritic membrane
circuits. Each circuit can be stimulated via 226 synapses on
two synaptic inputs. As a default, the latter are conVgured
for excitatory and inhibitory stimuli, respectively. However,
they can be set up to represent e.g. two excitatory inputs
with diUerent synaptic time constants or reversal potentials.
2
Fig. 1. A wafer containing 384 HICANN chips. The undiced wafer
undergoes a custom post-processing step where additional metal
layers are applied to establish inter-reticle connectivity and power
distribution. (Photo courtesy of the Electronic Vision(s) group,
Heidelberg.)
By connecting multiple dendritic membranes larger neurons
with up to 14 × 103 synapses can be formed.
A single wafer contains 384 chips with 200 × 103 neurons
and 45 × 106 synapses. Multiple wafers can be connected to
form even larger networks. The BSS's infrastructure consists
of six wafers and is being extended to 20 wafers for the Vrst
HBP milestone.
1.3 Spiking Neuron Model
There exist diUerent techniques of varying complexity for
simulating networks of spiking neurons. The reference
implementation we use for HTM networks is based
on Vrst generation, binary neurons with discrete time
steps [Numenta, Inc]. Third generation models, however,
incorporate the concept of dynamic time and implement
inter-neuron communication based individual spikes.
Starting from the original Hodgkin-Huxley equations
[Hodgkin and Huxley, 1952], multiple spiking neuron
models were developed that feature diUerent levels of detail
and abstraction. The HICANN chip implements Adaptive
Exponential Integrate-and-Fire model (AdEx) neurons [Brette
and Gerstner, 2005]. At its core,
it represents a simple
Leaky Integrate-and-Fire (LIF) model but features a detailed
spiking behavior as well as spike-triggered and sub-threshold
adaption. It was found to correctly predict approximately
96 % of the spike times of a Hodgkin-Huxley-type model
neuron and about 90 % of the spikes recorded from a cortical
neuron [Jolivet et al., 2008]. On the HMF and thus also
in the following simulations, the neurons are paired with
conductance-based synapses allowing for a Vne-grained
control of the synaptic currents and the implementation
of e.g. shunting inhibition.
Porting HTM Models to the Heidelberg Neuromorphic Computing Platform
2 SPIKING NETWORK MODELS
Implementing neural network models for a neuromorphic
hardware platform or dynamic software simulations requires
an abstract network description deVning the individual cell
populations as well as the model's connectivity. For this
work, our primary focus was on developing mechanistic
and functional implementations of the software reference
models while staying within the topological and parameter
restrictions imposed by the hardware platform. A more
detailed biophysical approach should begin with simulations
of single HTM neurons and their dendritic properties before
advancing to more complex systems, e.g. full networks.
In the following sections we describe spatial pooler
and temporal memory models that
incorporate basic
HTM properties. These models are able to reproduce the
fundamental behaviour of existing software implementations.
The simulations were set up in Python using the PyNN
library [Davison et al., 2008]. Besides supporting a wide
range of software simulators, this high-level interface is also
supported by the HMF platform [Billaudelle, 2014a]. NEST
was used as a simulation backend [Gewaltig and Diesmann,
2007]. To enable multiple synaptic time constants per neuron,
a custom implementation of the AdEx model was written.
2.1 Spatial Pooler
At its core the spatial pooler resembles a k-Winners-Take-
All (kWTA) network: k out of m columns are chosen to
be active in each time step. In fact, kWTA networks have
often been mentioned as an approximation for circuits
naturally occurring in the neocortex [Felch and Granger,
2008]. Continuous-time and VLSI implementations of such
systems have been discussed in the literature [Erlanson and
Abu-Mostafa, 1991, Tymoshchuk, 2012, Maass, 2000]. In
the implementation below we describe a novel approach to
maintaining stable sparsity levels even with a large number
of inputs.
The network developed for this purpose is presented in
Figure 2. It follows a purely time-based approach and is
designed for LIF neurons. It allows for very fast decision
processes based on a single input event per source. Each
column is represented by a single cell which accumulates
feed-forward input from the spike sources. Here, the rise
time of the membrane voltage decreases with the number
of presynaptic events seen by the cell: cells receiving the
most input will Vre before the others. An inhibitory pool
consisting of a single cell collects the network's activity.
Low membrane and high synaptic time constants lead to
a reliable summation of events. When a certain number of
spikes have been collected – and thus the cell's threshold
has been crossed – the pool strongly inhibits all cells of the
network suppressing subsequent spike events.
The model is extended by adding subtle feed-forward
shunting inhibition. The inhibitory conductance increases
Fig. 2. Timing based implementation of the spatial pooler. Each
column is represented by a single cell C and receives sparse input
from the input vector 11 . The columns become active when the
number of connected active inputs crosses a threshold. The rise
time of the membrane voltage highly depends on the number of
coincident inputs: cells with more presynaptic activity will Vre
before those with less stimuli do. Inhibitory pool I accumulates
the columnar spikes 22 and in doing so acts as a counter. After
a certain number of columns have become active, the pool will
inhibit and shut down all columns preventing any further activity
33 . To stabilize this kWTA model, all columns receive a subsampled
feed-forward inhibition 44 . This eUectively prolongs the decision
period for high input activity.
with the overall
input activity νin. With the reversal
potential set to match the leakage potential, the conductance
contributes to the leakage term
g(cid:48)
l = gl + ginh(νin) .
This eUectively slows down the neurons' responses and
thus prolongs the decision period of the network. With
this inhibition, the resulting system is able to achieve stable
sparsity levels with a large number of inputs, at the cost of
slightly slower response times.
Tie situations between columns receiving the same
number of presynaptic events are resolved by adding slight
gaussian jitter to the weights of the excitatory feed-forward
connections. This gives some columns structural advantages
over other columns resulting in a slightly faster response to
the same stimulus. By increasing the standard deviation σj
of the jitter, the selection criterion can be blurred.
2.2 Temporal Memory
Similar to the spatial pooler,
the temporal memory
implementation was designed for fast reaction times
and spike-timing based response patterns. A complete
network consists of m identical columns with n HTM
cells each. Modelling these cells is a challenge in itself. A
multicompartmental neuron model would represent the best
Vt. While a neuromorphic hardware chip implementing such
3
13IC24 Billaudelle et al.
Fig. 3. Implementation of the temporal memory not including
plasticity. Every HTM cell within a column is modeled with
three individual LIF cells modeling diUerent compartments (distal
dendrites D, soma S and a lateral inhibition segment I – which is not
biologically inspired). Per column, there exist multiple cell triples
as well as one "head" cell P which participates in the columnar
competition and collects proximal input for the whole column.
Activity of this cell is forwarded to the individual soma cells of the
column 11 . Without a previous prediction, this results in all soma
cells Vring. However, the distal compartment sums over the input of
the previous time step. When a threshold is reached, the inhibitory
compartment as well as the soma are depolarized 33 44 . Together
with proximal input 22 , the inhibitory partition Vres and inhibits
all other cells in the column 55 .
a model is planned and Vrst steps in that direction have
already been taken [Millner, 2012], the current system does
not provide this feature. Since HTM cells primarily depend
on the active properties of a compartment, it can be modelled
by a triple of individual LIF cells as shown in Figure 3.
A column collects proximal input using a single cell. In fact,
this cell can be part of a spatial pooler network as presented
in section 2.1. When the column becomes active, this cell
emits a spike and excites both the neurons representing the
HTM cells' somae as well as inhibitory cells. The inhibitory
projection, however, is not strong enough to activate the
target compartment alone. Instead, it only leads to a partial
depolarization. The soma neuron, however, reaches the
4
Vring threshold for a single presynaptic event. This suXces
as a columnar bursting mechanism (i.e. temporal memory
property 3): without predictive input, all soma compartments
will Vre as a response to the proximal stimulus.
Distal input is processed for each cell individually by
their dendritic segment compartments. A cell's dendritic
segment receives input from other cells' somae. When its
Vring threshold is crossed, it partly depolarizes the inhibitory
helper cell of the same triplet. This synaptic projection is
set up with a relatively long synaptic time constant and
a reversal potential matching the threshold voltage. This
ensures that the predictive state is carried to the next time
step and prohibits the cell from becoming active due to distal
input alone. On proximal input, the already depolarized
helper cell Vres before the somatic compartments. The latter
are then inhibited instantly, with the exception of the own
triplet's soma. As described, this basic predictive mechanism
fails when multiple cells are predicted, since the inhibitory
compartments laterally inhibit every cell. The solution is
to also depolarize the somatic compartments of predicted
cells. In summary this mechanism satisVes temporal memory
properties 1 and 2.
3 RESULTS
The network models presented in the previous section
were simulated in software to investigate their behavior.
In the following, respective experiments and their results
are shown. Additionally, plasticity rules and topological
requirements are discussed in respect of the HMF.
3.1 Network Simulations
The spatial pooler was analyzed for a network spanning
1,000 columns and an input vector of size 10,000. To speed
up the simulation, the input connectivity was preprocessed
in software by multiplying the stimulus vector to the
connectivity matrix.
A Vrst experiment was designed to verify the basic
kWTA functionality. A random pattern was presented to
the network. The number of active inputs per column – the
input overlap score – can be visualized in a histogram as
shown in Figure 5. By highlighting the columns activated by
that speciVc stimulus, one can investigate the network's
selection criteria. Complying with the requirements for
a spatial pooler, only the rightmost bars – representing
columns with the highest input counts – are highlighted.
Furthermore, the model's capability to resolve ties between
columns receiving the same input counts is demonstrated:
the bar at the decision boundary was not selected as a whole
but only a few columns were picked. This veriVes spatial
pooler property 2.
In a second scenario, input vectors with varying sparsity
were fed into the network, as shown in Figure 6. The number
DIPSsingle predictive element21354 Porting HTM Models to the Heidelberg Neuromorphic Computing Platform
80
70
60
50
#
40
30
20
10
0
All Columns
Active Columns
]
%
[
y
t
i
s
r
a
p
S
t
u
p
t
u
O
6
5
4
3
2
1
0
20
30
40
50
60
0
1
2
3
4
5
6
Input Events
Input Sparsity [%]
Fig. 5. Histogram showing the distribution of overlap scores
individual columns receive. Columns activated by the spatial pooler
network are highlighted. This conVrms that only competitors
with the highest input enter an active state. Furthermore, tie
situations between columns with the same overlap score are
resolved correctly.
Fig. 6. The relative number of active columns is plotted against
the input vector's sparsity. After a certain level of input sparsity
is reached, columns start to enter active states. With higher
presynaptic activity, columnar competition increases and the
output sparsity reaches a plateu. The curve's exact course can
be manipulated through the neurons' parameters as can the size of
the plateau. Error bars indicate the standard deviation across Vve
trials.
of active columns stays approximately constant across a
wide range of input sparsity. Additionally the plot shows
that columns must receive a minimum amount of input to
become active at all. This veriVes the underlaying kWTA
approach as well as spatial pooler properties 1 and 4.
To verify the general functionality of a spatial pooler,
expressed in property 3, a third experiment was set up. Input
data sets with a variable overlap were generated starting
from an initial random binary vector. For each stimulus, the
overlap of the columnar activity with the initial dataset was
calculated while sweeping the input's overlap. The resulting
relation of input and output overlap scores is shown in
Figure 7. Also included are the results of a similar experiment
performed with a custom Python implementation of the
spatial pooler directly following the original speciVcation
[Hawkins et al., 2011]. Multiple simulation runs all yielded
results perfectly matching the reference data, thus verifying
property 3.
The experiments have shown that the model presented in
this section does fulVll the requirements for a spatial pooler
and can be considered a solid kWTA implementation. The
speciVc results of course depend on the individual network
size and conVguration. In this case, the network – most
importantly the columnar neurons' time constants – was
conVgured for a relatively short time step of T = 50 ms.
By choosing diUerent parameter sets, the network can be
tuned towards diUerent operational scenarios, e.g. further
increasing the model's stability.
The temporal memory was veriVed in a Vrst sequence
prediction experiment. A reference software implementation
reference data
simulation 1
simulation 2
simulation 3
simulation 4
simulation 5
1.0
0.8
0.6
0.4
0.2
p
a
l
r
e
v
O
t
u
p
t
u
O
0.0
0.0
0.2
0.4
0.6
Input Overlap
0.8
1.0
Fig. 7. Output overlap as a dependency of the input vector's
overlap score. In each of the Vve simulation runs, the stimulus'
was gradually changed starting from a random vector. As required
for a spatial pooler, two similar input stimuli get mapped to similar
output patterns, while disjunct input vectors result in low overlap
scores. The simulations fully reproduce data from an existing
software implementation which is also shown in this Vgure.
was trained with three disjunct sequences of length three.
Consecutive sequences were separated by a random input
pattern. The trained network's lateral connectivity was
dumped and loaded in a simulation. When presented with
the same stimulus, the LIF-based implementation was able
to correctly predict sequences, as shown in Figure 8.
5
Billaudelle et al.
1.0
0.8
0.6
0.4
0.2
p
a
l
r
e
v
O
t
u
p
t
u
O
reference data
simulation 1
simulation 2
simulation 3
simulation 4
simulation 5
0.0
0.0
0.2
0.4
0.6
Input Overlap
0.8
1.0
Fig. 9. Dependency of output and input overlap for a trained spatial
pooler. Results of Vve independent simulation runs are shown as
well as reference data from a custom software implementation.
3.2 Learning Algorithms
Implementing online learning mechanisms in neuromorphic
hardware is a challenge, especially for accelerated systems.
Although the HMF features implementations of nearest-
neighbor Spike-Timing Dependent Plasticity (STDP) and
Short Term Plasticity (STP) [Friedmann, 2013a, Billaudelle,
2014b], more complex update algorithms are hard to
implement. Numenta's networks rely on structural plasticity
rules which go beyond these mechanisms.
The spatial pooler's stimulus changes signiVcantly for
learned input patterns. VeriVcation of its functionality under
these conditions is important. In order to follow the HTM
speciVcation as closely possible, a supervised update rule was
implemented in an outer loop: for each time step, a matrix
containing the connections' permanence values is updated
according to the activity patterns of the previous time step.
This allows us to implement the concepts of structural
plasticity presented in the original whitepaper. For the target
platform, the learning algorithms could be implemented on
the Plasticity Processing Unit (PPU) which is planned for
the next version of the HICANN chip [Friedmann, 2013b].
Simulation results of the implementation described above
are shown in Figure 9.
Experiments to replace the HTM structural plasticity
rules by a classic nearest-neighbor STDP model did not
yield the desired results. The HTM learning rules require
negative modiVcations to inactive synapses in segments that
contribute to cell activity. In contrast, STDP does not aUect
inactive synapses.
3.3 Map and Route
Applying abstract network models to the hardware platform
requires algorithms for placing the neuron populations and
6
routing the synaptic connections. In a best-case scenario,
this processing step results in an isomorphic projection of
the network graph to the hardware topology. For networks
with extreme connectivity requirements, however, synaptic
losses must be expected.
Mapping the simulated networks does not represent a
challenge for the routing algorithms. The temporal memory
can be projected to a single wafer without synaptic loss. The
same still applies with assumed lateral all-to-all connectivity
resulting in approximately 2 million synapses. The latter
network corresponds to a network with a potential pool
of 100 % which would allow the exploration of learning
algorithms even without creating and pruning hardware
synapses.
On the hardware platform, a tradeoU between the number
of aUerent connections per cell and the number of neurons
must be taken into consideration: while it is possible to
connect the dendritic membrane circuits such that a single
neuron can listen on roughly 14 × 103 synases, such a
network could only consist of approximately 3 × 103 neurons
per wafer. With just 226 synapses, just under 200 × 103
neurons can be allocated per wafer.
Scaling up the proof-of-concept models to a size useful
for production purposes, however, challenges the hardware
topology as well as the projection algorithms.
A minimal useful HTM network spans 1024 columns
with 8 cells each. In such a scenario the neurons would
receive lateral input on 32 dendritic segments. Allowing
approximately 1 × 103 aUerent connections per dendritic
segment, this network could be realized on approximately
1 × 106 dendritic membrane circuits, or six wafers. The
existing system set up for the BSS would suXce for this
scenario. Even larger networks could be brought to the HBP's
platform.
4 CONCLUSION AND OUTLOOK
Implementing machine intelligence algorithms as spiking
neural networks and porting them to a neuromorphic
hardware platform presents high demands in terms of
precision and scalability.
We have shown in this paper that HTMs can be
successfully modeled in dynamic simulations. The basic
functionality of spatial pooler and temporal memory
networks could be reproduced based on AdEx neurons.
In theory, the proof of concept networks can be easily
transferred to the HMF, since the high-level software
interfaces are designed to be interchangable. Of course,
emulating the models on the actual hardware platform will
bring up a new set of challenges.
Adapting the HTM's learning rules to the native plasticity
features available on the HMF has turned out to be nontrivial.
The learning rules could not be replicated with the current
Porting HTM Models to the Heidelberg Neuromorphic Computing Platform
implementation of classic STDP. As a freely programmable
microprocessor directly embedded into the neuromorphic
core, the PPU provides the ability to extend the system's
plasticity mechanisms in order to implement the HTM rules.
Further investigation is required to map out a complete
implementation of the HTM update rules on the PPU.
Analog neuromorphic hardware is susceptible to transistor
mismatches due to e.g. dopand Wuctuations in the production
process [Mihai A. Petrovici, 2014]. A careful calibration
of the individual neurons is required to compensate for
these variations. Due to the complexity of the problem
and the high number of interdependent variables, a perfect
calibration is hard to accomplish. Therefore, network models
are required to be tolerant regarding certain spatial, and trial-
to-trial variations on the computing substrate. Carrying out
additional Monte Carlo simulations with slightly randomized
parameters is important to investigate the robustness of the
presented networks.
Finally, a multicompartmental neuron model is planned
for a later version of the neurmorphic platform. Making use
of this extended feature set will signiVcantly increase the
level of biophysical detail. This will account for the detailed
dendritic model used in HTMs and help to stay closer to the
whitepaper as well as the reference implementation.
Besides paving the road towards a highly accelerated
the HMF oUers a high
execution of HTM models,
level of detail in its neuron implementation. With the
multicompartmental extension and a Wexible plasticity
framework, we anticipate the platform will prove valuable
as a tool for further low-level research on HTM theories.
ACKNOWLEDGEMENTS
Special thanks to JeU Hawkins, Prof. Dr. Karlheinz Meier,
Paxon Frady, and the Numenta Team.
REFERENCES
Sebastian Billaudelle. PyHMF – eine PyNN-kompatible
Schnittstelle für das HMF-System, 2014a.
Sebastian Billaudelle. Characterisation and calibration of
short term plasticity on a neuromorphic hardware chip.
Bachelor's thesis, Universität Heidelberg, 2014b.
Romain Brette and Wulfram Gerstner. Adaptive exponential
integrate-and-Vre model as an eUective description of
Journal of neurophysiology, 94(5):
neuronal activity.
3637–3642, 2005.
Andrew P Davison, Daniel Brüderle, Jochen Eppler, Jens
Kremkow, Eilif Muller, Dejan Pecevski, Laurent Perrinet,
and Pierre Yger. Pynn: a common interface for neuronal
network simulators. Frontiers in neuroinformatics, 2, 2008.
Ruth Erlanson and Yaser S Abu-Mostafa. Analog neural
networks as decoders. In Advances in neural information
processing systems, pages 585–588, 1991.
Andrew C Felch and Richard H Granger. The hypergeometric
connectivity hypothesis: Divergent performance of brain
circuits with diUerent synaptic connectivity distributions.
Brain research, 1202:3–13, 2008.
Simon Friedmann.
A new approach to learning in
neuromorphic hardware. PhD thesis, Heidelberg, Univ.,
Diss., 2013, 2013a.
Simon Friedmann.
A new approach to learning in
neuromorphic hardware. PhD thesis, Heidelberg, Univ.,
Diss., 2013, 2013b.
SB Furber, F Galluppi, S Temple, and LA Plana. The spinnaker
project. 0018-9219, (99):1–14, 2014.
Marc-Oliver Gewaltig and Markus Diesmann. Nest (neural
simulation tool). Scholarpedia, 2(4):1430, 2007.
JeU Hawkins, Subutai Ahmad, and Donna Dubinsky.
Cortical Learning Algorithm and Hierarchical Temporal
Memory, 2011. URL http://numenta.org/resources/HTM_
CorticalLearningAlgorithms.pdf.
HBP SP9 partners. Neuromorphic Platform SpeciVcation.
Human Brain Project, March 2014.
A.L. Hodgkin and A.F. Huxley. A quantitative description of
membrane current and its application to conduction and
excitation in nerve. Journal of Physiology, 117:500–544,
1952.
Renaud Jolivet, Felix Schürmann, Thomas K Berger,
Richard Naud, Wulfram Gerstner, and Arnd Roth.
The quantitative single-neuron modeling competition.
Biological cybernetics, 99(4-5):417–426, 2008.
Wolfgang Maass. Neural computation with winner-take-all
as the only nonlinear operation. Citeseer, 2000.
Paul Müller Oliver Breitwieser Mikael Lundqvist Lyle Muller
Matthias Ehrlich Alain Destexhe Anders Lansner RenÃľ
SchüUny Johannes Schemmel Karlheinz Meier Mihai
A. Petrovici, Bernhard Vogginger.
Characterization
and compensation of network-level anomalies in mixed-
signal neuromorphic modeling platforms. PLOS ONE,
October 2014.
doi: dx.doi.org/10.1371/journal.pone.
URL http://www.plosone.org/article/info%
0108590.
3Adoi%2F10.1371%2Fjournal.pone.0108590.
Sebastian Millner. Development of a Multi-Compartment
Neuron Model Emulation. PhD thesis, Heidelberg, Univ.,
Diss., 2013, 2012.
Numenta, Inc. Numenta Platform for Intelligent Computing
(NuPIC). URL http://numenta.org/.
PavloV. Tymoshchuk.
Shigang Yue,
A continuous-time model of
analogue k-winners-take-all neural circuit. In Chrisina
Jayne,
editors,
Engineering Applications of Neural Networks, volume 311
of Communications in Computer and Information Science,
pages 94–103. Springer Berlin Heidelberg, 2012.
ISBN
978-3-642-32908-1.
and Lazaros Iliadis,
7
Billaudelle et al.
Cell 1
]
V
m
[
l
a
t
s
i
d
V
]
V
m
[
h
n
V
i
]
V
m
[
a
m
o
s
V
−50
−55
−60
−65
−70
−50
−55
−60
−65
−70
−50
−55
−60
−65
−70
]
V
m
[
l
a
t
s
i
d
V
]
V
m
[
h
n
V
i
]
V
m
[
a
m
o
s
V
−50
−55
−60
−65
−70
−50
−55
−60
−65
−70
−50
−55
−60
−65
−70
]
V
m
[
l
a
t
s
i
d
V
]
V
m
[
h
n
V
i
]
V
m
[
a
m
o
s
V
−50
−55
−60
−65
−70
−50
−55
−60
−65
−70
−50
−55
−60
−65
−70
Cell 2
Cell 3
0
50
100
150
200
t [ms]
250
300
350
400
Fig. 4. Neuron traces for a temporal memory column containing three HTM cells. Each of these cells is represented by a somatic
compartment, an inhibitory helper cells and two dendritic segments. The column is activated by proximal input in every time step and
receives random distal stimulus predicting none, one or more cells per step. As indicated by the automatic classiVcation algorithm, the
column exhibits a correct response pattern to these predictions.
8
Porting HTM Models to the Heidelberg Neuromorphic Computing Platform
x
e
d
n
I
l
l
e
C
x
e
d
n
I
l
l
e
C
x
e
d
n
I
l
l
e
C
x
e
d
n
I
l
l
e
C
7
6
5
4
3
2
1
0
7
6
5
4
3
2
1
0
7
6
5
4
3
2
1
0
7
6
5
4
3
2
1
0
a1
a2
a3
b1
b2
b3
c1
c2
c3
random
random
random
0 20 40 60 80 100 120
0 20 40 60 80 100 120
0 20 40 60 80 100 120
Column Index
Column Index
Column Index
Fig. 8. A LIF neuron based temporal memory implementation correctly predicting diUerent patterns. Predicted cells are marked blue, active
cells in purple. The network spans 128 columns with each of their eight HTM cells collecting distal stimuli via two dendritic segments.
Connectivity for the distal inputs was conVgured externally. The model was presented three disjunct sequences of size three. The individual
patterns were separated by a random input Sparse Distributed Representation (SDR).
9
|
1304.1565 | 1 | 1304 | 2013-04-04T21:27:40 | Bayesian Odds-Ratio Filters: A Template-Based Method for Online Detection of P300 Evoked Responses | [
"q-bio.NC",
"physics.med-ph",
"stat.AP"
] | Template-based signal detection most often relies on computing a correlation, or a dot product, between an incoming data stream and a signal template. While such a correlation results in an ongoing estimate of the magnitude of the signal in the data stream, it does not directly indicate the presence or absence of a signal. Instead, the problem of signal detection is one of model-selection. Here we explore the use of the Bayesian odds-ratio (OR), which is the ratio of posterior probabilities of a signal-plus-noise model over a noise-only model. We demonstrate this method by applying it to simulated electroencephalographic (EEG) signals based on the P300 response, which is widely used in both Brain Computer Interface (BCI) and Brain Machine Interface (BMI) systems. The efficacy of this algorithm is demonstrated by comparing the receiver operating characteristic (ROC) curves of the OR-based (logOR) filter to the usual correlation method where we find a significant improvement in P300 detection. The logOR filter promises to improve the accuracy and speed of the detection of evoked brain responses in BCI/BMI applications as well the detection of template signals in general. | q-bio.NC | q-bio | 1
Bayesian Odds-Ratio Filters:
A Template-Based Method for Online Detection of
P300 Evoked Responses
Asim M. Mubeena, Kevin H. Knutha,b,c
aKnuth Cyberphysics Lab, Department of Physics, University at Albany, Albany NY, USA
bDepartment of Informatics, University at Albany, Albany NY, USA
cAutonomous Exploration Inc., Andover MA , USA
Abstract: Template-based signal detection most often relies on computing a correlation, or a dot product, between an incoming
data stream and a signal template. While such a correlation results in an ongoing estimate of the magnitude of the signal in the
data stream, it does not directly indicate the presence or absence of a signal. Instead, the problem of signal detection is one of
model-selection. Here we explore the use of the Bayesian odds-ratio (OR), which is the ratio of posterior probabilities of a
signal-plus-noise model over a noise-only model.
We demonstrate
this method by applying
it
to simulated
electroencephalographic (EEG) signals based on the P300 response, which is widely used in both Brain Computer Interface
(BCI) and Brain Machine Interface (BMI) systems. The efficacy of this algorithm is demonstrated by comparing the receiver
operating characteristic (ROC) curves of the OR-based (logOR) filter to the usual correlation method where we find a significant
improvement in P300 detection. The logOR filter promises to improve the accuracy and speed of the detection of evoked brain
responses in BCI/BMI applications as well the detection of template signals in general.
Index Terms— biomedical signal processing, digital signal processing, medical signal detection, Bayesian, Bayesian model testing
I. INTRODUCTION
Neurophysiology in general, and brain machine interface (BMI) in particular, relies on the detection and
characterization of the electric and magnetic field potentials produced by the brain in response to sensory
stimulation or in association with its cognitive and/or motor operations and planning. These potentials
originate from the transmembrane current flow produced by multiple ensembles of hundreds of thousands
of synchronously firing neurons [1], [2]. Human scalp electroencephalographic (EEG) recording has the
advantage of being noninvasive, inexpensive and portable, which make it a very popular technique among
the BCI/BMI community.
Event-related potentials (ERPs) are evoked brain responses synchronized to sensory, cognitive and motor
events [3]. As such, they consist of relatively reproducible waveshapes embedded in background EEG
activity [4], [5]. In particular, the P300 evoked potential [6], [7] is a positive peak that is evoked 300 ms
after stimulus onset. The P300 is widely used in both BCI and BMI applications, but recent studies show
that it is also useful in the diagnosis of neurological disorders and lie detection. Despite its wide use,
detection of the P300 component in single-trial recordings remains challenging. More often, multiple trials
are needed to detect the P300 component in ongoing EEG activity, which decreases the overall speed of
BCI/BMI systems.
In general, the detection of an ERP is made difficult by its low signal-to-noise (SNR) ratio compared to
the ongoing background EEG. The most commonly-used method to estimate the ERP is coherent
averaging, which averages a large number of time-locked epochs of the identical stimuli presented to the
subject. This technique is used in online applications of BCI [8], [9]; however, it suffers from several
drawbacks [10]. One drawback is that coherent averaging implicitly assumes that the ERP waveshape is
identical from trial-to-trial. This is known to be a poor assumption [11], [12] and, in fact, such trial-to-trial
variability is utilized in some advanced source separation methods [4], [13], [14].
2
Several detection methods work by correlating a template signal with the ongoing EEG. For example,
the Woody filter [15], [16], performs a correlation by taking the dot product of the ongoing EEG signal
with the signal template. Other correlation-based classification techniques like Pearson’s correlation
method (PCM), Fisher’s linear discriminant (FLD), stepwise linear discriminant analysis (SWLDA), linear
support vector machine (LSVM), and the Gaussian kernel support vector machine (GSVM) have been
presented and compared [17], [18]. These methods, which convert the problem of detecting P300 into a
binary search problem (P300 present or not-present), have been widely used in BCI applications. In this
paper we present a Bayesian odds ratio-based (OR-based) technique, which relies on a signal template to
detect the P300 response in ongoing EEG. The OR-based detection method is not specific to EEG signals
and can be applied to any kind of template-based signal detection. The technique is demonstrated by
applying it to synthetic P300 responses. We compare our results to the template correlation (dot product /
Woody filter [16]) method and demonstrate efficacy by comparing the resulting receiver operating
characteristic (ROC) curves [19], [20], [21].
II. THE LOG ODDS-RATIO FILTER
A. Bayes’ Theorem
Bayes’ theorem transforms the problem of signal detection into one of model selection where the
probabilities of the considered models can be computed and compared. The posterior probability
gives us the probability of the parameter values, m, of the model M, given the recorded data D
and relevant prior information I
.
(1)
The posterior probability
depends on both the prior probability of the model parameter
and the data-dependent ratio of the likelihood of the data given the model and its specific
values
model parameter values
, which represents the probability that the
to the evidence
data could have resulted from the model irrespective of the specific model parameter values. Since
summing the posterior probability over all possible model parameter values results in unity, we can write
the evidence as
,
(2)
which demonstrates the reason that this quantity is also referred to as the marginal likelihood.
We consider the problem of signal detection as a model selection problem where we compare the
evidence provided by the data given one model to the evidence of the data provided by the data given
another model. As we are trying to detect the P300 signal from ongoing EEG activity, we will refer to the
model as noise-only (background EEG activity)
,
(3)
where is the time series data recorded from the mth electrode and is the time series of ongoing
EEG activity recorded at each channel. We refer to such activity as noise and assume it to be independent
in each channel.
),,(IMDmPIMDPIMmDPIMmPIMDmP,,,,,,),,(IMDmP),(IMmP),,(IMmDP),(IMDPIMmDPIMmPdmIMDP,,,,)()(:ttxMmmN
3
Similarly, we refer to the model as signal (P300) plus noise (background EEG activity) and write the
model as
,
(4)
where is the time series data recorded from the mth electrode, is the coefficient that couples the
signal source to the mth electrode, α is the amplitude of the signal in the given trial. We assume that the
coefficients and the source signal waveshape are both known. The value of the single-trial
amplitude parameter α in the signal-plus-noise model is assumed to be between 0 and some maximum
value a, so that
.
B. Odds Ratio
To compare the two models, we compute the odds ratio (OR), which is the ratio of the marginal
likelihoods, or evidences, of each model. However, in this problem, the noise-only model has no
model parameters. So we compare the marginal likelihood of to the likelihood :
(5)
where
And
(6)
(7)
We assign a Gaussian likelihood to both models, and note that this does not necessarily mean that the noise
is Gaussian, but rather that the expected squared deviation
from the mean is a relevant quantity. We
then have that (6) becomes
and that the likelihood term in (7) becomes
(8)
(9)
We also assume that the distribution of single-trial amplitudes α is Gaussian with mean
so that
and variance
(10)
Applying the probability assignments in (9) and (10) to the integral in (7), we find that
(11)
)()()(:ttsCtxMmmmNSa0NNSNNSZZIMDPIMDPOR),(),()),(,)((),(ItCtxPIMDPZNNaNSNSItCtxPIPdIMDPZ0),),(,)(()(,2MmTtmMTNtxZ11222/2))((21exp)2(MmTtmmmsMTsttsCtxItCtxP11222/2)()()(21exp)2(),),(,)((2222/12)(21exp)2()(IP22222222/122/22)(221exp)2()2(DEerfDEaDerfDFDEZssssMTsNS
4
where the function erf(x) is the error function defined as the integral of a Gaussian with zero mean and unit
variance from zero to x [22].
Also by defining
(12)
we have that
and
,
(13)
(14)
.
(15)
It is often easier to work with the logarithm of the probabilities. Taking the logarithm of the odds ratio (5)
we have that
, (16)
which after some algebraic simplification, including the assumption that
, results in
(17)
Since the log odds ratio compares the evidence of the two models, it can be used as an index for signal
detection. We will refer to this filter as the logOR filter.
The remainder of this paper focuses on evaluating the performance of this proposed method by
comparing and contrasting its receiver operator characteristic (ROC) curves with those produced by the
usual correlation method (dot product). These ROC curves are constructed by applying these techniques to
times-series data consisting of synthetic EEG background signals in which are embedded P300 targets.
The next section discusses the construction of the synthetic EEG data.
III. SYNTHETIC EEG DATA
To analyze the performance of the logOR filter, we generated synthetic EEG data representing both EEG
background and the P300 evoked response. Three channels of synthetic EEG data (Figure 1a) were
generated to simulate recordings from Cz, Pz and Fz, which are commonly used for P300 -based BMI
applications [18]. The dipole model is used to scale the data among different electrodes. The data from
each of these channels consisted of 300 epochs each being 800 ms in length and comprised of 200 samples,
which is consistent with a sampling rate of 250 Hz. Thirty epochs were selected to each host a single P300
response.
222sSTtMmmtsCSD11222)(TtMmmmtstxCSE112)()(TtMmmtxSF11222))(()log()log(loglogNNSNNSZZZZORs22222222222)(2loglog2log2121logDEerfDEaDerfDDEORssss
5
(a)
(c)
(b)
FIGURE 1. (a) An illustration of the synthetic EEG data generated for the filter evaluation. The three panels represent data
recorded from different channels: Cz, Pz, Fz (top to bottom). Two P300s, indicated by arrows are embedded in these traces.
(b) This figure illustrates the mean power spectrum of the synthetic EEG data. (c) The target signal (P300) template is shown
in this figure. Note that we are using a noisy template to simulate a template generated by taking averages of P300 recordin g
epochs.
We randomly generated a sequence of stimulus event times and used these times as latencies at which to
superimpose the P300 template onto the synthetic ongoing background EEG signal, which we refer to as
noise [23]. The stimulus event times were stored and used to identify as true/false-positives and true/false-
negatives during the filter evaluation. The remaining 270 epochs exhibited only ongoing background EEG
(noise).
The synthetic P300 waveform was represented using the classical theory [23], [24] where an ERP
waveform reflects a phasic-burst of activity. The P300 template (Figure 1b) was created with by producing
a peak of width of 100ms starting around 300 ms after onset. To produce a P300 template more similar to
what one would generate in the lab by averaging recorded P300 responses, we added low-level background
noise to the P300 template.
To generate the ongoing EEG background, we used MATLAB code provided online by Yeung and
Bogacz [25]. Fifty different sine waves were superimposed to create the background signal. This was done
by randomly selecting a frequency for each sine wave between 0 and 50 Hz. The phase was selected at
random, and the amplitude was chosen to be consistent with the Human EEG power law spectrum on
which was superimposed two peaks at 10 and 60 Hz [26] (Figure 1c).
To study the effect of the SNR on the filter performance, we created 17 data sets where the SNR,
calculated by the formula
,
(19)
was varied in integral steps from -6 dB to 7 dB as well as 10, 15 and 20 dB. This covers the typical SNR
range seen in BCI EEG applications, which is approximately from -6 dB to 6 dB [27].
210log10NoiseSignaldBAASNR
(a)
(c)
(e)
(g)
6
(b)
(d)
(f)
FIGURE 2. Receiver Operative Characteristic (ROC)
curves for both the Correlation Method (dashed) and logOR
filter (solid). The area under the curve quantifies the
efficacy of the detection method. Each panel shows results
for a different SNR value varying from -5dB to 15 dB. As
the SNR value increases, the area under the curves increases
for both methods. The logOR filter out -performs the
Correlation Method at all SNR levels.
7
IV. RECEIVER OPERATIVE CHARACTERISTICS (ROC) CURVES
ROC curves serve as a performance measure of signal detection [20], [21] by plotting the true positive rate
(also known as sensitivity) versus the false positive rate (1–specificity). Sensitivity is computed by
identifying the fraction of correctly detected targets (P300) and specificity refers to the fraction of non-
targets (non-P300) identified as non-targets. The fractional area under the ROC curve serves to quantify
the performance of the detection methods so that a perfect detection method will result in unity; whereas a
completely failed method will result in zero. The result is such that the more accurate the detection
method, the greater the fractional area under the ROC curve.
V. DETECTION THRESHOLD
The selection of a detection threshold value is a difficult task. As the detection threshold increases, the
sensitivity decreases while the specificity increases, which means that the false positive fraction (1–
specificity) decreases. To produce ROC curves, we calculate sensitivity and (1-specificity) for each
distinct value of the detection measure (i.e logOR / Correlation) to consider it as a candidate for detection
cutoff. By plotting (1–specificity) versus sensitivity, the efficacy of the detection method can be quantified
by the area under the ROC curve. The ROC curve starts from the most strict decision cutoff point where
sensitivity and (1–specificity) values are zero where there will be no positive detection, and ends at most
lenient cutoff point where sensitivity and (1–specificity) values are one and all detection values are taken to
represent a positive detection. The value of the area under the curve for the best detection method should
be one [28], [29]. However, when a method results in an ROC curve where the area under the curve is less
than one; there is always a tradeoff between sensitivity and (1–specificity). So, the detection threshold can
be found by minimizing the distance between sensitivity and specificity. One can also find the threshold
value by plotting the sensitivity and specificity versus the detection criteria, and selecting the decision
threshold value based on the intersection point of sensitivity and specificity curves.
FIGURE 3. SNR vs. Area under ROC curves for the Correlation Method
and the logOR filter. The value of Area under the curve increases as SNR
increases for both methods. For higher values of SNR, the logOR filter
performs much better than the Correlation Method.
VI. RESULTS
8
In this paper we developed and evaluated a template-based detection method that relies on the Bayesian log
odds ratio (logOR filter) to detect the presence or absence of a P300 signal in a synthetic ongoing EEG
dataset representing three commonly-used EEG channels: Cz, Pz and Fz.
We applied the logOR filter to the synthetic EEG data and compared the results to those obtained using
the traditional correlation filter. Using the detections and the sequence of events used to generate the data,
we computed the ROC curves for different SNR ranging from -6 dB to 20 dB to evaluate the performance
of both methods. Figure 2 shows the resulting ROC curves for several SNR values ranging from -5 dB to
15 dB. The area under the ROC curve is consistently greater for the logOR filter indicating that the logOR
filter out-performs the traditional correlation filter. A plot of the area under the ROC curves as a function
of P300 SNR can be seen in Figure 3. Both methods exhibit similar performance at the lowest SNR value,
but as the SNR increases, one can see that the logOR filter out-performs the correlation-based filter.
VII. DISCUSSION
Most template-based methods designed to detect a target signal in a continuous EEG stream rely on the
cross correlation of the incoming data with the template or target signal [15], [16]. Here we present a new
approach in which we recast the problem of signal detection in terms of evidence-based model selection.
By computing the Bayesian log odds-ratio between two models (the signal plus noise model, , and the
noise only model, ), we produce an index, which we refer to as the logOR filter. The value of the
logOR filter index is high when the target signal, for example a P300, is present and low otherwise. We
studied the performance of these two methods by applying them to synthetic EEG signals with P300 targets
exhibiting SNRs ranging from -6 to 20 dB. Performance was quantified by constructing the ROC curves
for each method and computing the area under the ROC curve. We found that the logOR filter out-
performs the correlation method. This suggests that systems aiming to minimize the number of target
trials, such as P300-based BCI/BMI systems, may benefit from employing the logOR filter. However it is
observed that logOR filter performance is very sensitive to differences between the assigned and true
standard deviation values of the background noise and the trial-to-trial amplitude variability of the P300
response. Therefore in a practical application one must choose the values of these parameters very
carefully to calculate the logOR filter value.
ACKNOWLEDGMENTS
The authors would like to thank Dr. Dennis J. McFarland for valuable discussions, access to EEG data and
instruction regarding the BCI2000 system
REFERENCES
[1] M. A. B. Brazier. “Evoked responses recorded from the depths of the human brain,” Annals of the New York Academy of Sciences, 112(1):
pp. 33–59, 1964.
[2] C. C. Duncan, R. J. Barry, J. F. Connolly, C. Fischer, P. T. Michie, R. Näätänen, J. Polich, I. Reinvang, and C. V. Petten. “Event-related
potentials in clinical research: guidelines for eliciting, recording, and quantifying mismatch negativity, P300, and N 400,” Clinical
Neurophysiology, 120(11): pp. 1883–1908, November 2009. PMID: 19796989.
[3] D. Friedman and R. Johnson. “Event-related potential (ERP) studies of memory encoding and retrieval: a selective review,” Microsc Res
Tech, 51(1): pp. 6–28, 2000.
9
[4] K. H. Knuth, A. S. Shah, W. A. Truccolo, M. Ding, S. L. Bressler, and C. E. Schroeder, “Differentially variable component analysis:
Identifying multiple evoked components using trial-to-trial variability,” Journal of Neurophysiology, 95(5): pp. 3257–3276, May 2006.
PMID: 16467431.
technique,”
[5] A. Mouraux and G. D. Iannetti. “Across-trial averaging of event-related EEG responses and beyond,” Magnetic Resonance Imaging,
26(7): pp. 1041–1054, September 2008. PMID: 18479877.
[6] U. Hoffmann. Bayesian machine learning applied in a brain-computer interface for disabled users, PhD thesis, STI, Lausanne, 2007.
[7] Sutton, M. Braren, J. Zubin, and E. R. John. “Evoked-potential correlates of stimulus uncertainty,” Science, 150(700): pp. 1187–1188,
1965.
[8] K. H. Chiappa. Evoked Potentials in Clinical Medicine, 3e. Lippincott Williams & Wilkins, 1997.
[9] E. Donchin, K. M. Spencer, and R. Wijesinghe, “The mental prosthesis: assessing the speed of a P300-based brain-computer interface,”
IEEE Transactions on Rehabilitation Engineering, 8(2): pp. 174 –179, June 2000.
[10] G. Sparacino, S. Milani, E. Arslan, and C. Cobelli, “A Bayesian approach to estimate evoked potentials,” Computer Methods and
Programs in Biomedicine, 68(3): pp. 233–248, June 2002.
[11] G Fein and B Turetsky. “P300
latency variability
in normal elderly: effects of paradigm and measurement
Electroencephalography and clinical neurophysiology, 72(5): pp. 384–394, May 1989. PMID: 2469563.
[12] J. M. Ford, P. White, K. O. Lim, and A. Pfefferbaum, “Schizophrenics have fewer and smaller P300s: a single-trial analysis,” Biological
Psychiatry, 35(2): pp. 96–103, January 1994. PMID: 8167215.
[13] C. D’Avanzo, S. Schiff, P. Amodio, and G. Sparacino, “A Bayesian method to estimate single-trial event-related potentials with
application to the study of the p300 variability,” Journal of Neuroscience Methods, 198(1): pp. 114–124, 2011.
[14] W. Truccolo, K. H. Knuth, A. Shah, S. L. Bressler, C. E. Schroeder, and M. Ding. “Estimation of single-trial multicomponent ERPs:
differentially variable component analysis (dVCA),” Biological Cybernetics, 89(6): pp. 426–438, December 2003. PMID: 14673654.
[15] H. Serby, E. Yom-Tov, and G.F. Inbar. “An improved P300-based brain-computer interface,” IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 13(1): pp. 89–98, 2005.
[16] C. D. Woody. “Characterization of an adaptive filter for the analysis of variable latency neuroelectric signals ,” Medical and Biological
Engineering, 5(6): pp. 539–554, November 1967.
[17] D. J. Krusienski, E. W. Sellers, F. Cabestaing, S. Bayoudh, D. J. McFarland, T. M. Vaughan, and J. R. Wolpaw. “A comparison of
classification techniques for the P300 speller,” Journal of Neural Engineering, 3(4): pp. 299–305, December 2006.
[18] E. W. Sellers, D. J. Krusienski, D. J McFarland, T. M. Vaughan, and J. R. Wolpaw. “A P300 event-related potential brain-computer
interface (BCI): the effects of matrix size and inter stimulus interval on performance ,” Biological Psychology, 73(3): pp. 242–252,
October 2006. PMID: 16860920.
[19] J. T. Bushberg, J. A. Seibert, E. M. Leidholdt, and J. M. Boone. The Essential Physics of Medical Imaging. Lippincott Williams &
Wilkins, December 2011.
[20] S. Schinkel, O. Dimigen, and N. Marwan. “Selection of recurrence threshold for signal detection ,” The European Physical Journal Special
Topics, 164(1): pp. 45–53, November 2008.
[21] M. H. Zweig and G. Campbell. “Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine ,”
Clinical Chemistry, 39(4): pp. 561–577, April 1993.
[22] M. Abramowitz, I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, New York:
Dover, 1965, Chapter 7, pp. 297.
[23] N. Yeung, R. Bogacz, C. B. Holroyd, and J. D. Cohen, “Detection of synchronized oscillations in the electroencephalogram: an evaluation
of methods,” Psychophysiology, 41(6): pp. 822–832, Nov. 2004. PMID: 15563335.
[24] A. S. Shah, S. L. Bressler, K. H. Knuth, M. Ding, A. D. Mehta, I. Ulbert, and C. E. Schroeder, “Neural dynamics and the fundamental
mechanisms of event-related brain potentials,” Cerebral Cortex, 14(5): pp. 476–483, May 2004.
[25] N. Yeung, R. Bogacz, C. B. Holroyd, S. Nieuwenhuis and J. D. Cohen Generation of simulated EEG data.
http://www.cs.bris.ac.uk/~rafal/phasereset/
[26] T. C. Ferree and R. C. Hwa. “Power-law scaling in human EEG: relation to Fourier power spectrum,” Neurocomputing, 52-54: pp. 755–
761, June 2003.
[27] J. Kronegg, S. Voloshynovskiy, and T. Pun. “Analysis of bit-rate definitions for brain-computer interfaces,” In Conference on Human-
Computer Interaction (HCI-2005), Las Vegas, 2005.
[28] N. A. Obuchowski. “Receiver operating characteristic curves and their use in radiology,” Radiology, 229(1): pp. 3–8, October 2003.
[29] N. A. Obuchowski. “ROC analysis,” American Journal of Roentgenology, 184(2): pp. 364–372, February 2005.
|
1703.04182 | 7 | 1703 | 2018-08-28T20:42:53 | Optimal control of eye-movements during visual search | [
"q-bio.NC"
] | We study the problem of optimal oculomotor control during the execution of visual search tasks. We introduce a computational model of human eye movements, which takes into account various constraints of the human visual and oculomotor systems. In the model, the choice of the subsequent fixation location is posed as a problem of stochastic optimal control, which relies on reinforcement learning methods. We show that if biological constraints are taken into account, the trajectories simulated under learned policy share both basic statistical properties and scaling behaviour with human eye movements. We validated our model simulations with human psychophysical eye-tracking experiments | q-bio.NC | q-bio |
Optimal control of eye-movements during visual search
A.Y.Vasilyev
Queen Mary University of London, Mile End Road, E1 4NS
Abstract
We study the problem of an optimal oculomotor control during the execution of visual search tasks. We
introduce a computational model of human eye movements, which takes into account various constraints
of the human visual and oculomotor systems. In the model, the choice of the subsequent fixation location
is posed as a problem of a stochastic optimal control, which relies on reinforcement learning methods. We
show that if biological constraints are taken into account, the trajectories simulated under a learned policy
share both basic statistical properties and a scaling behaviour with human eye movements. We validated
our model simulations with human psychophysical eye-tracking experiments.
Keywords:
scaling in biology, visual search, reinforcement learning, multifractal analysis
1. Introduction
The human oculomotor system performs hun-
dreds of thousands of eye-movements per day during
the execution of different behavioral tasks. In order
to find the details of a visual scene related to the
tasks, humans direct foveal vision to the most in-
formative locations via saccades - high-velocity con-
jugate gaze shifts. Saccades are followed by a vi-
sual fixation, during which the human oculomotor
system generates fixational eye movements involun-
tarily. Despite the remarkable achievements in the
modelling of fixational eye movements and the inter-
pretation of their fundamental properties [1, 2, 3],
there is no comprehensive generic model of fixation
selection [4, 5, 6], which takes into account the un-
derlying mechanisms of visual attention [5, 7, 8] and
qualitatively describes the statistical properties of
saccadic eye-movements during the execution of vi-
sual tasks [9, 10, 11, 12].
Previously, the problem of fixation selection was
studied in the framework of control models of eye
movements [13, 11, 9].
In control models the ob-
server gathers information about the world during
each fixation, integrates information over all fixa-
tions into a belief state and makes a choice of the
next location on which to fixate. This choice is gov-
erned by the policy of gaze allocation - a function
that specifies the action of decision-maker in a cer-
tain belief state. It was shown that the policy based
on information maximization criteria [9] generates
trajectories that share basic statistical properties
with human eye movements.
In this research, we
set the goal of developing a control model of fixation
selection that is capable of interpreting the scaling
behaviour of human eye-movements [10, 14, 12, 15]
and provides a human level of performance to a com-
putational agent.
In contrast to the previous research on control
models, we take into account the inherent uncer-
tainty of human oculomotor system and the dura-
tion of saccadic eye movements.
It's well known
that any motor action of humans is executed with
random error, which increases with movement mag-
nitude [16, 17]. Despite the oculomotor system hav-
ing developed a correction mechanism for saccade
Preprint submitted to Example: Nuclear Physics B
August 30, 2018
2.1 World state
2
errors [18], these result in inevitable temporal costs.
Furthermore, the duration of saccades is empirically
correlated with their magnitude as well [19]. These
factors result in situations where the observer has
to choose between more informative remote (and
riskier) locations and those nearby (but less infor-
mative ones). We show that if these constraints
are taken into consideration, the trajectories sim-
ulated under a learned policy share both basic sta-
tistical properties and scaling behavior with human
eye movements, which is not achievable with the
conventional infomax model [9].
On the basis of our results, we argue that we have
made the following contribution:
• The formulation of the biologically plausible
model of gaze allocation in the human observer
from the point of view of stochastic optimal
control. The representation of the model in
the form of partially observable Markov deci-
sion process (PO-MDP) and the proposal of a
heuristic policy.
• The development of robust and high perfor-
mance algorithms of simulation of PO-MDP.
The implementation of reinforcement learning
algorithms of policy optimization and numeri-
cal estimation of the optimal policy of gaze al-
location.
• The comprehensive statistical analysis of sim-
ulated trajectories and data from our psycho-
physical experiments. The policy, which is
learned with the policy gradient REINFORCE
algorithm, shows the highest level of statisti-
cal similarity with human eye-movements.
In
our experiments we discovered the dependency
of the mean saccade length and q-order Hurst
exponent on visibility of the target, which was
explained by our model.
2. Model of the ideal observer
In this section we formulate the model of the ideal
observer, which aims to localize the single target
object on the stationary 2D image. We represent
the model in the form of partially observable Markov
Figure 1:
Flow chart of the model of the ideal observer.
The visual search is a recurrent process that starts after ini-
tialization of the world state. On each consequent step the
observer receives the observation vector Wn, which is then
used for estimation of belief state pn using Bayesian inference
[20]. After the update of the belief state the observer makes
the decision Dn where to fixate next according to the policy
of gaze allocation: Dn = µ(pn). The next fixation location
is defined by execution function: An+1 = α(Dn). If the ob-
server fixates on the location of target An+1 = m the process
of visual search is terminated, otherwise the next step starts
with updated values of variables.
decision process (PO-MDP), which is summarized
by flow chart on Figure 1.
2.1. World state
At the beginning of each episode the target ob-
ject appears randomly at one of L possible locations.
We assume that the target is placed on background
noise or surrounded by distractors, which are placed
on vacant locations. The world state Sn is repre-
sented as a tuple:
Sn = (m, An, tn)
(1)
Initialize world state Extract observation Update belief state Execute saccade YES NO End Visual search Make decision Initialization 2.3 Execution of saccades
where m is a location of the target on the image and
An is gaze fixation location that changes with the
number of step n, and tn is time passed from the
start of a trial and the step n.
If the observer fixates the gaze on the location of
target:
An = m
(2)
the visual task is considered to be accomplished.
This formulation of the terminal state reflects the
necessity to foveate the target in order to extract as
much information about its identity and details as
possible. The location of target m doesn't change
during a trial.
2.2. Update of belief state
The decision-making of the observer is modeled as
PO-MDP with a belief state pn - a discreet probabil-
ity distribution function of target location given all
observations received up to the step n. Because the
observer is instructed that the target appears ran-
domly, the initial belief state p0 is a discrete uniform
distribution.
On each step n observer receives the observation
vector Wn = (W1,n, . . . , WL,n), whose elements rep-
resent the perceptual evidence that the target is at
corresponding locations. The probability distribu-
tion function is updated using Bayesian inference
[20]:
(cid:80)
pn−1 (l )p(Wnl , An )
pn−1 (k )p(Wnk , An )
pn (l ) =
(3)
k
where l is the index of the location and p(Wl, A) is
an observation model. In order to take into account
the uncertainty of the processing of perceptual in-
formation within the neural circuits of the observer,
we follow the "noisy observation" paradigm [9]. In
this paradigm the observation model p(Wl, A) re-
flects the presence of the observer's internal sources
of inefficiency, such as physical neural noise on all
stages of information processing. According to the
perceptual model [11] the observation W may be
represented as a random variable with Gaussian dis-
tribution with mean depending on the location m of
the center of target on the lattice:
(cid:89)
p(Wl, A) =
p(WlA) =
l
Wl; δl,m,
1
F ((cid:107)l − A(cid:107))
(cid:17)
(cid:16)
(cid:89)
l
N
3
(4)
where δi,j is Kronecker delta, N (x, µ, ν) is a value
of Gaussian function with mean µ and variance ν
for argument x; (cid:107)l − A(cid:107) is Euclidean distance be-
tween the locations l and the current fixation A,
and F is a Fovea-Peripheral Operating Character-
istic (FPOC) [13]. FPOC is a function that repre-
sents the dependence of a signal-to-noise ratio on
the eccentricity. Figure 2 demonstrates FPOC cal-
culated for several values of RMS contrast of the
background 1/f noise: en ∈ (0.1, 0.15, 0.2, 0.25) and
a single value of RMS contrast of target et = 0.2.
The calculation are based on the analytical expres-
sions from [11]. The signal-to-noise ratio has a peak
at fovea and decreases rapidly with eccentricity.
In our simulations we consider only the case of the
rotationally symmetric FPOC. This assumption is
not correct for human observers, and better generic
model of FPOC can be found in [21]. The broken
circular symmetry of FPOC inevitably results in the
asymmetry of the visual search process [22]. We sim-
plify the model of FPOC, because in this research
we focus our attention more on the temporal struc-
ture of eye-movements rather than on their spatial
distribution.
2.3. Execution of saccades
The decision of which location to fixate next, Dn,
is made on each step of PO-MDP according to the
policy of gaze allocation µ :
Dn = µ(pn)
(5)
After making the decision, the coordinates of the
next fixation location An+1 are defined by the exe-
cution function:
An+1 = α(Dn) = Dn + Jn
(6)
where Jn is a Gaussian-distributed random error
with zero mean and standard deviation ν defined
in [17]:
2.5 Value function
4
of these time intervals are empirically correlated
with a magnitude of the saccade preceding the fix-
ation [19, 23, 24]. The duration of saccadic eye-
movements Θsac(n) in range of magnitudes from
1.5◦ to 30◦ is possible to approximate as [25]:
Θsac(n) = τsac (cid:107)An − An−1(cid:107)0.4
(9)
where τsac = 21ms · deg
−0.4. Besides the magni-
tude of saccade, the fixation duration Θf ix(n) is in-
fluenced by various factors as a discriminability of
the target [26], its complexity and the visual task
of the observer [24, 27]. However, if the observer
is correctly informed about the targets' properties
before the task execution and performs the visual
task without any interruptions, the contribution of
these factors to the fixation duration (with excep-
tion of magnitude) is constant during each trial. The
eye-tracking experiments with the fixations tasks
[23, 24, 28] found that the dependence of fixation
duration on saccade amplitude is linear:
Θf ix(n) = (cid:107)An − An−1(cid:107) τf ix + Θ0,f ix
(10)
with a slope τf ix = 6ms/ deg. The constant
Θ0,f ix = 250ms is an intercept, averaged from val-
ues from eye-tracking data [29, 30]. Finally, the du-
ration of step n is:
Θ(n) = Θsac(n) + Θf ix(n)
(11)
sac = 20± 3ms· deg
The values of parameters used in simulations are
consistent with our estimates from the eye-tracking
f ix = 5.8±
experiments: τ∗
1.8ms/ deg, Θ0,f ix = 241±42ms. Within this range
of the parameters' values we didn't find a substantial
difference in the estimates of the learned policy of
gaze allocation.
−0.4, τ∗
2.5. Value function
Given the initial world state S0, we define the cost
function for policy µ as an expectation of a random
variable V :
Vµ(S0) = E [V µ, S0]
(12)
Figure 2: Fovea-Peripheral Operating Characteristic was cal-
culated for several values of the RMS contrast of background
noise: en ∈ (0.1, 0.15, 0.2, 0.25) and the single value of RMS
contrast of target et = 0.2. The signal-to-noise ratio has a
peak at fovea and decreases rapidly with eccentricity.
ν = ζ0 + ζ1 (cid:107)Dn − An(cid:107)
(7)
The error of the saccade execution is proportional
to intended saccade amplitude (cid:107)Dn − An(cid:107) given in
degrees, the value of parameters: ζ0 = 0.87 deg,
ζ1 = 0.084 (from[17]).
The next step of PO-MDP starts after the tran-
sition to the location An+1. This decision-making
model may be easily extended in order to take into
account the extraction of visual
information be-
tween the moment of making the decision where to
fixate next and completion of the saccade.
2.4. Duration of the steps
After each consequent step the time variable t of
world state (1) is updated in a deterministic way:
tn = tn−1 + Θ(n)
(8)
where Θ(n) is a duration of step n. The duration of
time step Θ(n) is considered as a total time, which
is required for the relocation of the gaze from a pre-
vious location A(n − 1) to the current one A(n)
and the extraction of visual information from the
location A(n). Therefore, we consider Θ(n) as a
sum of durations of the fixation Θf ix(n) and the
saccade Θsac(n). According to the literature, both
12345eccentricity, deg02468valueFPOC0.10.150.20.25 3.2 Optimal policy estimation
5
The random variable V denotes the cost and is de-
fined by:
N(cid:88)
V ≡ c
Θ(n) = ctN
(13)
n=0
where N is a total number of steps in the episode,
and c is a time cost constant.
Formulation of the cost function in a real time
sets this study separately from the previous works
[31, 11, 13]. We show below, that the policy µ opti-
mized for the cost function with the reward defined
in (13) generates the sequences of actions with sta-
tistical characteristics close to the human saccadic
eye-movements.
3. Policy of gaze allocation
3.1. Infomax approach
In this section we describe two heuristic policies
related to the model of Entropy Limit Minimization
searcher [9]. We define the information gain on the
step n + 1 as: (cid:52)I(n + 1) = H(pn)− H(pn+1), where
H (·) is Shannon entropy. The heuristic policy π0
is defined as a policy which chooses such decision
Dn that maximizes the expected information gain
(cid:52)I(n + 1):
π0(pn) = Dn = arg max
D
[E [(cid:52)I(n + 1)]]
(14)
The term E [(cid:52)I(n + 1)] is calculated analytically in
[9] for the case of the saccadic eye-movements with-
out uncertainty (An+1 ≡ Dn):
1
2
E [(cid:52)I(n + 1)] =
(15)
where sign ∗ denotes a convolution operator, and
F is FPOC represented as a radially symmetric 2D
function: F (A) ≡ F ((cid:107)A(cid:107)). The expression (15)
gives an approximate value of the expected infor-
mation gain in the case of the stochastic saccadic
placement (6).
The figure 3 illustrates the decision-making pro-
cess, which corresponds to the policy π0. The colour
(cid:0)pn ∗ F 2(cid:1) (Dn)
map (left) represents the function of the expected in-
formation gain (equation (15)). The blue cross cor-
responds to the location of the current fixation on
the step n. The observer makes a decision to fixate
at the location defined by the policy: Dn = π0(pn).
This decision results in a saccadic eye-movements to
location An+1 = α(Dn) marked by the green cross.
After receiving the observation at the step n + 1,
the observer updates the belief state and evaluate
the information gain for the next decision. In this
particular situation, the target was absent at the
vicinity of An+1, and the observation resulted in the
decline of probability pn+1 in the area around the
green cross (figure 3 right). This area is effectively
inhibited from subsequent fixations due to low prob-
ability. The size of this area is defined by values of
FPOC (et = 0.2, en = 0.1 in this case). We call the
policy π0 "infomax greedy" in the text below.
The trajectories generated with infomax greedy
policy match the basic properties of human eye
movements [9]. However, the policy (15) doesn't
consider the correlation between the magnitude of
saccades and the durations of steps of MDP. We
show later that the policy π0 is inferior to the pol-
icy that optimizes the expected rate of information
gain E [(cid:52)I(n + 1)/Θ(n + 1)]:
π1(pn) = arg max
D
[E [(cid:52)I(n + 1)/Θ(n + 1)]]
(16)
Using the expression for E[(cid:52)I(n + 1)] (15), for
the deterministic saccadic placement (An+1 ≡ Dn):
(cid:34)(cid:0)pn ∗ F 2(cid:1) (Dn)
(cid:35)
Θ(n + 1)
(17)
π1(pn) = Dn = arg max
D
The policy π1 is called "infomax rate" in the text
below. The performance of these two heuristic poli-
cies will be compared with a performance of the pol-
icy learned with reinforcement learning algorithms
in the section Appendix A.4.1.
3.2. Optimal policy estimation
In this section we describe the evaluation of the
policy of gaze allocation that optimizes the cost
3.2 Optimal policy estimation
6
Figure 3: The decision-making process under the infomax greedy policy π0 [9]. The colour map (left) represents the function
of the expected information gain (equation (15)). The blue cross corresponds to location of the current fixation on the step
n. The observer makes a decision to fixate at the location defined by the policy: Dn = π0(pn). This decision results in the
saccadic eye-movement to location An+1 = α(Dn) marked by the green cross. After receiving the observation at the step n + 1,
observer updates the belief state and evaluates the information gain for the next decision. In this particular situation, the
target is absent in the vicinity of An+1, and the observation resulted in the decline of probability pn+1 in the area around the
fixation (the green cross). This area is effectively inhibited from the subsequent fixations due to low probability pn+1. The size
of this area is defined by the values of FPOC (in this case et = 0.2, en = 0.1).
function (12) for any starting world state S0. We
start with the representation of the stochastic pol-
icy µ in the on [20]:
µ(D, p) =
exp(f (D, p))
exp(f (l, p))
(18)
(cid:80)
l
(cid:88)
where f (D, p) is a function of expected reward gain
after making the decision D with the belief state
p. In this study we limit the search of f (D, p) to a
convolution [20] of belief state p:
f (D, p) =
K(D − l)p(l)
(19)
l
In supplementary materials Appendix A.1 we jus-
tify this choice of the policy and evaluate the form
of a kernel function K that allows us to effectively
solve the optimization problem with the policy gra-
dient algorithms. Our task is the search of the kernel
function K (A.4), which corresponds to the policy
that optimizes the cost function Vµ:
K∗ = arg min
K
(20)
for any starting world state ∀S0. The policy µ(K∗)
is called the optimal policy of gaze allocation.
Vµ(K)(S0)
We approach the optimization problem (20) with
an algorithm named "REINFORCE with optimal
baseline" [32] according to the procedure described
in Supplementary materialAppendix A. The perfor-
mance of REINFORCE was compared with one of
the optimization algorithms named "policy gradi-
ent parameter exploration" (PGPE) adopted from
[33]. The algorithm of REINFORCE with an op-
timal baseline belongs to the class of the likelihood
ratio methods, whereas PGPE is related to the finite
difference methods. Despite the distinction between
these two approaches, both algorithms give a close
estimation of the optimal policy Appendix A.4.1.
We simulated trajectories for the data analysis in
section 4 using the solution provided by REIN-
FORCE due its better performance comparing to
PGPE. Figure 4 demonstrates the decision-making
process under the policy µ learned for FPOC corre-
Expected information gain on step N, bits03691215X, deg03691215Y, deg00.10.20.30.40.5Expected information gain on step N+1, bits03691215X, deg03691215Y, deg00.10.20.30.40.50.6 4.1 Performance
7
Figure 4: The decision-making process under the policy learned for FPOC corresponding to conditions et = 0.2, en = 0.1.
At the step n observer fixates the location marked by a blue cross. The policy µ defines a probability density function of a
decision D where to fixate next (A.5). Observer chooses the decision Dn according to the policy, which results in a saccadic
eye-movement to location An+1 = α(Dn) (the green cross). As well as in the case of dynamics under the heuristic policy π0
previously visited locations are inhibited from the subsequent fixations. Note that movements to remote locations are inhibited
by the radial function. This results in co-directed short movements, which are also characteristic of human observer.
sponding to the conditions et = 0.2, en = 0.1 (see
figure A.13 for its kernel function). At the step n the
observer fixates the location marked by a blue cross.
The policy µ defines a probability density function
of the decision D where to fixate next (A.5). The
observer chooses the decision Dn = µ (pn), which
results in a saccadic eye-movement to the location
An+1 = α(Dn) (green cross). As well as in the case
of dynamics under the heuristic policy π0, previously
visited locations are inhibited from the subsequent
fixations.
4. Basic properties of trajectories
In this section we discuss the statistical proper-
ties of trajectories generated with the learned pol-
icy µ and the heuristic policies π0 and π1. The
simulations were performed on the grid with size
128 × 128 that corresponds to the visual field with
size of 15×15 deg in the psychophysical experiment.
In order to justify our computational model, we re-
produced the psychophysical experiments from [9].
The detailed description of the experiments can be
found in AppendixAppendix C.
4.1. Performance
Although this computational model was not de-
signed for an exact prediction of a response time
of human observers, it demonstrates a high level
of consistency in a performance of the visual task
execution with human observers. The performance
was measured as an average time to reach the tar-
get (the mean response time) and as a percentage
of the correct fixations on target's location on an
N-Alternative Forced Choice task (N-AFC). The
unsuccessful trials from the psychophysical exper-
iments were excluded from the consideration. We
found that the number of the unsuccessful trials
grows with the contrast of noise: 2.3%, 5.7%, 9.8%,
16.4% for the corresponding numbers of the contrast
n = (0.1, 0.15, 0.2, 0.25).
Figure 5 (left) demonstrates the percentage of cor-
rect fixations on the target location for the exper-
imental conditions: et = 0.2, en = 0.15. Means
and standard errors of the response time of the hu-
man observers is presented on Figure 5 (right) to-
gether with means of the response time for three
policies estimated from 104 episodes of PO-MDP.
The learned policy outperforms two heuristics and
PDF of decision at step N 03691215X, deg03691215Y, deg012#10-4PDF of decision at step N+1 03691215X, deg03691215Y, deg01234#10-4 4.2 Amplitude distribution
8
the human observers both in the mean response time
and the percentage of the correct fixations for all
experimental conditions. Human observers signif-
icantly outperformed the infomax rate for the ex-
perimental conditions: en = (0.2, 0.25) (Student's
t-test p < 0.05) and the infomax greedy for the con-
ditions en = (0.15, 0.2, 0.25) (p < 0.05) on the mean
response time, which was previously found in [22, 9].
In the same time the learned policy outperformed
the human observers significantly for the condition
en = 0.25, while for other conditions t-test didn't re-
ject hypothesis that distributions have equal means
at 5% significance level.
4.2. Amplitude distribution
The Figure 6 (left) shows the length distributions
of saccades of the human observers and the simu-
lated agents performing the visual search task corre-
sponding to the experimental conditions: en = 0.2,
et = 0.2.
The distributions for all policies and the human
observers exhibit an ascent between 0 deg and max-
imum around 2 deg. The difference in the behaviour
of the distributions starts from 4 deg. In this exper-
imental conditions the share of the saccades of the
human observer with the length larger than 4 deg
is 18%, whereas this value for π0 is 38%. The
length distribution for π0 stabilizes on the interval
[4.0, 14.0] deg that was observed in the earlier work
[11], and we found that length of this "stability"
interval increases linearly with the grid size. The
reason behind this is an uniform radial ranking of
policy π0 for all locations due to the constant radial
function (15). The decline of probability starts only
at a distance compared to the size of visual field.
On the other hand the length distributions of
trajectories under µ, π1 are concave on an interval
[4.0, 14.0] deg, which is also a characteristic for hu-
man eye-movement [34, 35]. The behaviour of the
radial function of µ reflects the non-uniform radial
ranking (a preference in decision-making, see figure
(4)) of the locations. As a result, the remote lo-
cations have significantly lower probabilities to be
chosen as the next destinations.
We performed Kolmogorov-Smirnov (K-S) test to
check equality of distributions of experimental and
Cases
noise
contrast
0.1
0.15
0.2
0.25
K-S test statistics
infomax
infomax
greedy
0.312
0.256
0.212
0.201
rate
0.121
0.082
0.053
0.095
learned
policy
0.124
0.081
0.04
0.074
critical
value
0.067
0.052
0.047
0.035
Table 1: Statistics for K-S test between the experimental
distribution of saccade length and simulated distributions for
different policies. The first and the second columns show the
values of RMS contrast in the psycho-physical experiment
and corresponding critical values of test statistics for signifi-
cance level α = 0.01. The next three columns show K-S test
statistics for distribution of saccades simulated under differ-
ent policies.
simulated saccades. The results of the test are
summarized in table 1 . The first and the second
columns show the values of RMS contrast in the
psycho-physical experiment and corresponding crit-
ical values of the test statistics at a significance level
α = 0.01. The critical values are different due to the
difference in number of saccades for each experimen-
tal condition. The next three columns show K-S test
statistics for the distributions of the saccades simu-
lated under the different policies. K-S test indicated
a higher statistical similarity between the distribu-
tions of the experimental saccades and the saccades
simulated under the learned policy µ for the cases
en = (0.15, 0.2, 0.25). In the case en = 0.1 the in-
fomax rate and the learned policy explained the ex-
perimental distribution equally well. From this, we
can make a conclusion that simulations under the
learned policy explained the best the length distri-
bution of the human eye-movements.
The mean length of the saccades was estimated
from 104 episodes of PO-MDP for all three policies
and compared with the mean length of the saccades
of the human observers (see figure 6 right). Ac-
cording to our results, the mean length of the sac-
cades decreases with en, which is consistent with
our simulations. It's an immediate consequence of
the decrease of values of FPOC with the increase
of the RMS contrast of noise, which is illustrated
on figure 2. The amplitude of the signal exceeds
4.2 Amplitude distribution
9
Figure 5: The performance of the human observers and the simulated agents. The learned policy outperforms two heuristics
both in the mean completion time and the percentage of correct responses in N-AFC task (left) for the experimental conditions:
et = 0.2, en = 0.15. The dependence of mean completion time (right) for the learned policy resembles one for the human
observer.
Figure 6: The histograms of the length distribution of the saccadic events (left) for trajectories generated under the policies
π0, π1 and µconv and the human eye-movements corresponding to the experimental conditions: en = 0.2, et = 0.2. The data
was binned with the resolution of 0.1 deg. The distribution function for all policies and human observer exhibits an ascent
between 0 deg and maximum around 2 deg. The distribution of length corresponding tothe infomax greedy π0 stabilizes after
4 deg and declines only after 10 deg. It is not consistent with length distribution of human saccadic eye-movements, which is
concave on an interval [4.0◦, 14.0◦]. The mean length of saccades decreases with en (right). It's immediate consequence of the
decrease of the width of FPOC with en, which defines the area of inhibition from the subsequent fixations.
time, sec0123456AFC % Correct020406080100infomax greedyinfomax ratehumanlearned kernel0.10.150.20.25RMS contrast of noise0246810response time, seclength, deg02468101214frequency012345learned kernelinfomax greedyinfomax ratehuman0.10.150.20.25RMS contrast of noise123456Saccade length, deg 4.3 Geometrical persistence
10
the amplitude of noise within the circle area with
radius r that satisfies the condition F (r) = 1 (we
call this radius the "width of FPOC"). This circle
area is effectively inhibited from the subsequent fix-
ations (see figures (3) and (4)), because information
is already gathered with a sufficient level of confi-
dence. However, we found that our model provides
close estimates of the mean length only for high val-
ues of the RMS contrast of noise. Our experimental
findings are consistent with previously reported re-
sults [36], where the visual search experiments were
set for several levels of the RMS contrast of back-
ground noise. In future works we plan to incorporate
more complex saccade execution model that takes
into account the bias toward the optimal saccade
length [17] in order to explain a lower variability of
the saccade length in the experiments.
4.3. Geometrical persistence
In this section we analyze the distribution of the
directional angle θd (this notation was introduced in
[10]) of the human saccadic eye-movements and the
simulated trajectories. The directional angle is the
angle between two consequent saccades, and, there-
fore, can be defined as θd = tan−1 (yn+1/xn+1) −
tan−1 (yn/xn), where (yn, xn) are the coordinates of
nth fixation. According to this definition, the move-
ment is related to a persistent one if the directional
angle is close to 0 or 2π. The angles with the values
close to π correspond to anti-persistent movements.
The distributions of the directional angle were cal-
culated for the trajectories generated by Markov de-
cision process with the policies π0, π1 and µ. Figure
(7)(left) demonstrates the distribution of the direc-
tional angle of the saccadic events for the human ob-
servers and the simulated trajectories for en = 0.2
and et = 0.2. The infomax greedy policy π0 gen-
erates the trajectories with stable anti-persistent
movements, because the policy π0 chooses the next
fixation location without taking the current location
into consideration. Due to the inhibitory behavior
of infomax, it's much less likely to choose the nearby
location instead of remote and relatively unexplored
ones. Only geometrical borders limit the choice of
the next fixation, which results in fixations on the
opposite side of the visual field (as the most remote
point, look at the figure (3)).
In contrast, the decision process under the learned
policy µ tends to preserve the direction of the move-
ment. The dynamic of the system under the policy µ
is quite similar to self-avoiding random walk model
described in [1]. Due to the asymptotic behavior of
the kernel function K(x, y), the reward gain from
the remote locations is suppressed, meanwhile, the
locations, which are already visited, are also inhib-
ited (look at the figure (4)). This results in short-
range self-avoiding movements, which demonstrate
the persistent behavior [37, 1], and, therefore, the
probability distribution of the directional angle θd
is biased towards values 0 or 2π. According to the
Figure 7 (left), the dynamics under the heuristics π1
is also characterized as a persistent random walk.
The learned policy µ has, in general, a stronger ra-
dial ranking of locations than π1, which results in
a shorter range of saccades, and a repulsion, caused
by inhibition, becomes more relevant. The distri-
bution of average length of saccades depending on
θd is shown on Figure 7 (right). On average the
co-directed movements are shorter than the reversal
ones for all policies.
In our experiments we discovered that the geo-
metrical persistence depends on the visibility of tar-
get (on FPOC in the simulations). We measured
the share of the saccades, which retain the direction
of the previous movement: cos(θd) > 0. This quan-
tity is called "persistence coefficient". The figure 8
demonstrates the dependency of the persistence co-
efficient on the RMS contrast of background noise
for the human observers and the simulated trajec-
tories. As it was mentioned previously, the aver-
age saccade length is decreasing with the growth of
RMS contrast (6). Therefore, the linear term (10)
in the duration of steps becomes less relevant, and
the decision-making becomes more agnostic about
the temporal costs (closer to the information greedy
π0). The decline of the persistent coefficient is also a
characteristic of human eye movements, which was
not covered in the previous research.
11
Figure 7: The histograms of the directional angle (left) and the distributions of the mean length of the saccades to the
directional angle (right). Data for both histograms was binned with the resolution of 20 deg. The infomax greedy policy π0
generates the trajectories with stable anti-persistent movements (left), with a high degree of separation between large and small
movements (right). In contrast, the decision-making process under the infomax rate policy π1 tends to preserve the direction
of movement. The dynamics under the learned policy µ is also characterized as a persistent random walk.
5. Statistical persistence
In the previous section we have analyzed the ge-
ometrical persistence of the human eye-movements
and the trajectories simulated under three different
policies. However, this statistical property doesn't
give any insight into a long-range correlation in
time-series.
In this section we show that dynam-
ics under the learned policy µ have a multifractal
behavior, which is similar to that of the human
eye-movements during execution of the visual search
task .
In contrast to the previous research [10] in our
analysis we distinguish between two different types
of the multifractality by a calculation of a gener-
alized Hurst exponent for shuffled time series. We
separate the time series on fixational and saccadic
eye-movements, which allows us to demonstrate the
fundamental difference in the temporal structure of
these types of eye-movements. It was shown that the
behaviour of the generalized Hurst exponent is con-
sistent with the basic statistical properties of eye-
movements. After this we demonstrate that the dy-
namics under the optimal policy of gaze allocation
explains the changes in scaling behaviour of eye-
movements with difficulty of the visual task both
on qualitative and quantitative levels.
Figure 8: The share of saccades, which retain the direction
of the previous movement: cos(θd) > 0, is called the "per-
sistence coefficient". This quantity demonstrates the depen-
dence of the persistence on the visibility of the target. As
it was mentioned previously, the average saccade length is
decreasing with the growth of the RMS contrast (6). There-
fore, the linear term (10) in the duration of the steps becomes
less relevant, and the decision-making becomes more agnostic
about temporal costs (closer to the information greedy π0).
The decline of the persistent coefficient is also a characteristic
of the human eye movements, which was not covered in the
previous research.
020406080100120140160180directional angle, deg0.511.522.533.5frequencyinfomax greedylearned kernelhumaninfomax rate020406080100120140160180directional angle, deg1.522.533.54average length, degrms contrast of noise0.10.150.20.25persistence coefficient0.20.30.40.50.60.70.8humaninfomax ratelearned kernelinfomax greedy 5.1 Multifractality of human eye movements
12
For statistical analysis of simulated trajectories
we use a multifractal detrended fluctuation analy-
sis (MF-DFA)[38], which is a widely-used method
for detection of long-range correlations in stochastic
time-series. It has found successful applications in
the field of bioinformatics [39, 40], nano and geo-
physics [41]. This method is based on the approxi-
mation of trends in time-series and the subtraction
of detected trends (detrending) from original data
on different scales. The detrending allows deducting
the undesired contribution to long-range correlation,
which is a result of non-stationarities of physical pro-
cesses. We use the package provided by Espen Ihlen
[42] for all our estimations of the generalized Hurst
exponent in this section.
In the appendix Appendix B we thoroughly ex-
plain the details of the multifractal analysis. The
subsection Appendix B.1 presents the details of MF-
DFA algorithm.
In the subsections Appendix B.3
and Appendix B.3 we explain how MF-DFA is per-
formed over the simulated trajectories. The re-
sults of the multifractal analysis of the human eye-
movements are presented in 5.1. The subsection 5.2
summarizes our findings and compares the general-
ized Hurst exponent of the simulated trajectories to
one of human eye-movements for different experi-
mental conditions.
5.1. Multifractality of human eye movements
We perform MF-DFA over the difference of time
series of the human gaze positions and in order to
compare the estimated generalized Hurst exponent
with the simulations. The differentiated time series
was estimated from raw data of coordinates of the
gaze fixations A = {(x1, y1) , ... (xN , yN )}with the
resolution of 7 ms:
∆X = {(x2 − x1) , . . . (xN − xN−1)}
∆Y = {(y2 − y1) , . . . (yN − yN−1)}
(21)
(22)
The time series ∆X and ∆Y were estimated for
each trial with certain experimental conditions and
concatenated over all participants. After this, we
represent the differentiated time series in the fol-
lowing way: ∆X = {F1, S1, ...., Fm−1, Sm−1, Fm},
where Fi and Si correspond to the sequences of the
movements during time interval of i-th fixation and
saccade respectively [10]. We separate the differen-
tiated time series on the fixational and the saccadic
time series:
∆XF = {F1, 0s1, . . . Fm−1, 0sm, Fm}
(cid:9)
∆XS =(cid:8)0f1, S1, . . . 0fm−1 , Sm, 0fm
(23)
(24)
where 0n corresponds to zero array with the length
n, and fm and sm are the lengths of corresponding
sequences Fm and Sm.
The figure 9 demonstrates the scaling of the q-
order fluctuation function Fq(s) (B.4). This graph is
a result of the application of MF-DFA over the hor-
izontal concatenated differentiated time series ∆X
of the human scan-paths for the experimental condi-
tions: et = 0.2, en = 0.25. The red, blue and green
lines correspond to the linear approximation of func-
tion log2 (Fq(s)) for the orders q = {−10; 0; 10}.
The scaling of Fq(s) exhibits the crossover on a time
scale of 256 ms. The crossover separates the "lower"
and "upper" regimes mentioned in [10]. According
to Amor et. al. the crossover is caused by the pres-
ence of two different generative mechanisms of eye-
movements. The lower regime is related to fixational
eye-movements (which is supported by the value of
crossover scale scros being close to the average fixa-
tion duration), and upper regime - to the saccadic
ones. The crossover in the scaling of Fq(s) was ob-
served for all experimental conditions. The value
of generalized Hurst exponent H(q) (Figure 9 right)
is obtained through linear regression of log2 (Fq(s)).
Our estimates of H(q) are consistent with the ones
of Amor et. al. for both directions and all regimes.
In order to distinguish between two different types
of multifractality [38] we calculated the generalized
Hurst exponent Hshuf (q) for the shuffled differen-
tiated time series. The first type of multifractal-
ity is a consequence of a broad probability density
5.1 Multifractality of human eye movements
13
Figure 9: The scaling of the q-order fluctuation function Fq(s) (left), and the generalized Hurst exponent H(q) (right) computed
through a linear regression of log2 (Fq(s)). This graph is a result of application of MF-DFA over the horizontal differentiated
time series ∆X of the concatenated human scan-paths for the experimental conditions: et = 0.2, en = 0.25. The red, blue and
green lines correspond to the linear approximation of function log2 (Fq(s)) for the orders q = {−10; 0; 10}. The scaling of Fq(s)
exhibits the crossover the crossover on a time scale of 256 ms. The crossover separates the "lower" and the "upper" regimes
mentioned in [10]. The lower regime is related to fixational eye-movements (which is supported by the value of crossover scale
scros being close to average fixation duration), and the upper regime - to saccadic ones. The crossover in the scaling of Fq(s)
was observed for all experimental conditions.
function for the values of time series. If only mul-
tifractality of the first type presents in time series:
H(q) = Hshuf (q). The second type of multifractal-
ity is caused by the difference in correlation between
large and small fluctuations, which is a scenario de-
scribed in [10].
In this case Hshuf (q) = 0.5 and
H(q) = 0.5 + Hcorr(q), where Hcorr(q) is (nega-
tive) positive for the long-range (anti-)correlation.
If both types of multifractality present in time se-
ries: H(q) = Hshuf (q) + Hcorr(q).
The figure 10 demonstrates our estimates of the
Hurst exponent of the shuffled time series Hshuf (q)
(top) and the correlational Hurst exponent Hcorr(q)
(bottom) for the horizontal (left) and the vertical
components (right). We estimated both exponents
for the saccades (green dashed line) and FEM (pur-
ple dashed line) in the upper and the lower regimes
of scales respectively. As well as a previous graph
9, this one is a result of an application of MF-DFA
over the concatenated differentiated time series of
the human eye-movements for the experimental con-
ditions: et = 0.2, en = 0.25. The behaviour of
(cid:40)
Hshuf (q) for the full time series and the saccadic
time series in the upper regime corresponds to the
one mentioned in [38] (eq. 27):
H(q) ∼
1/q (q > α)
1/α (q ≤ α)
(25)
with α ∼ 1. The equation 25 was derived for time
series of uncorrelated random values with the power
law distribution:
(cid:26) αx−(α+1) x ≥ 1
0 x < 1
P =
(26)
One can see a similarity of the function (26) with
the distribution of the amplitude of the saccadic
events for humans (see figure 6). The amplitude
distribution of the saccades demonstrates the power
law behavior on the interval [4.0◦, 14.0◦] with α ≈ 1.
The probability distribution function (26) also re-
flects an absence of saccades with the length lower
than minimal one. Therefore, the first type of mul-
tifractality of the saccadic time series is caused by
16326412825651210242048Scale,ms-10123log2(Fq)Scaling function Fq (q-order RMS)q =-10q =0q =10q =-10q =0q =10-10-8-6-4-20246810q00.511.52Hqq-order Hurst exponentdata1Hq(-10) =1.5083Hq(0) =0.83336Hq(10) =0.248 5.1 Multifractality of human eye movements
14
Figure 10: The Hurst exponent of the shuffled time series Hshuf (q) (top) and the correlational Hurst exponent Hcorr(q)
(bottom) for the horizontal (left) and the vertical components (right) of the human eye-movements. As well as a previous graph
9, this one is a result of an application of MF-DFA over the concatenated human scan-paths for the experimental conditions:
et = 0.2, en = 0.25. The behaviour of Hshuf (q) for both horizontal and vertical shifts for full scales corresponds to the one
mentioned in [38] (eq. 27). We assume that multifractality of the first type is caused by an asymptotic behaviour of the
amplitude distribution of saccades (see figure 6). The difference in the long-range correlation of large and small fluctuations is
reflected by Hcorr(q) (figure 10 bottom). Due to the properties of the fluctuation function B.4 for positive (negative) q-orders
the main contribution are coming from segments containing small (large) fluctuations [38]. The positive (negative) long range
correlation (Hcorr(q) > 0) is, therefore, a characteristic of small (large) fluctuations in the upper and the full scales regimes for
both directions. In general, these results are consistent with the distribution of the average length of saccades to the directional
angle (see figure 7 right), which also indicates the difference in persistence of large and small saccades.
-10-8-6-4-20246810q00.511.5H(q)H(q), shuffled horizontal shiftssaccadesFEMall scalesupper regimelower regime-10-8-6-4-20246810q00.511.5H(q)H(q), shuffled vertical shifts-10-8-6-4-20246810q-0.4-0.200.20.40.6H(q)correlational H(q), horizontal shifts-10-8-6-4-20246810q-0.200.20.4H(q)correlational H(q), vertical shifts the broad probability distribution of saccade mag-
nitude.
The difference in the long-range correlation of
large and small fluctuations is reflected by Hcorr(q)
(figure 10 bottom). Due to the properties of fluc-
tuation function (B.4) for the positive (negative)
q-orders the main contribution are coming from
segments containing the large (small) fluctuations
[38]. The positive (negative) long-range correlation
(Hcorr(q) > 0) is, therefore, a characteristic of the
small (large) fluctuations in the upper regime for
the saccadic and the full time series. These results
are consistent with the distribution of the average
length of saccade to the directional angle (see fig-
ure 7 right), which also indicates the difference in
the persistence of large and small saccades. There-
fore, we confirm here that the small saccadic eye-
movements demonstrate the long-range correlations
as well as fixational eye-movements.
The time series of FEM demonstrates
the
monofractal behaviour and the positive correlations
with H ≈ 0.8 in the lower regime of scales [10]. How-
ever, the behaviour of both Hcorr(q) and Hshuf (q)
for the full time series in the lower regime indicates
the presence of multifractalities of both types. At
the present moment we have no explanation of the
multifractality in the lower regime and leave this
problem for a future work.
5.2. Dependence on visibility
In this section we present a comparison of the
generalized Hurst exponent for the human eye-
movements in the upper regime and the simulated
trajectories under the learned policy. As well as in
the case of the geometrical persistence, we claim the
quantitative properties of the statistical persistence
depend on the visibility of the target.
We estimated the correlational Hurst exponent
Hcorr(q) and the Hurst exponent of the shuffled
time series Hshuf (q) for the differentiated trajec-
tories of the human eye-movements for all
lev-
els of the RMS contrast of background noise:
en ∈ (0.1, 0.15, 0.2, 0.25). Figure 11 (left) shows
Hcorr(q) (left) of simulated trajectories (blue) un-
der the learned policy µ and the correlational
Hurst exponent for the human eye-movements
15
corr(q) + H y
(pink) averaged over two directions: Hcorr(q) =
corr(q)) /2 in the upper regime. The
(H x
correlational Hurst exponents for the negative q-
orders declines with the growth of the RMS con-
trast of background noise both for the human eye-
movements and the simulated trajectories. This in-
dicates the weakening of the correlation between
small fluctuations. For the positive q-orders the
correlational Hurst exponent is less affected by the
change of the visibility of target. The Hcorr(q) for
q = 10 stabilized on values 0.04 and −0.12 for hu-
man eye-movements and the simulated trajectories
correspondingly. In general, the correlations weaken
with the growth of the RMS contrast, which is con-
sistent with the decline of the geometrical persis-
tence 8. The decline of the Hurst exponent with the
increase of difficulty of visual search task was also
observed in the previous work [12].
The Hurst exponent of the shuffled time series
(Figure 11 right), as well as the correlational Hurst
exponent, demonstrates the decline with the growth
of the RMS contrast for the negative q-orders both
for the human eye-movements and the simulated
trajectories.
In the subsection 5.1 we mentioned
that the behaviour of Hshuf (q) resembles the one
related to time series of random values with the
power law distribution 26. The average value of
this time series equals 1/ (α − 1) for α > 1. The
increase of α results both in the decrease of the av-
erage value in time series and the decrease of the
value of Hshuf (q) ∼ 1/α for q < 0. Therefore,
the average value in time series and the values of
Hshuf (q) for the negative q-orders are correlated in
the assumption of the power-law distribution. Pre-
viously we found the decrease of the average saccade
length with the growth of RMS of background noise
6, which is consistent with the decrease of values
of Hshuf (q) for negative q-orders. We assume that
this correlation is caused by the power-law asymp-
totic behaviour of the length distribution of human
eye-movements (26).
6. Conclusion
We have presented a computational model of the
ideal observer that both qualitatively and quantita-
16
tional Hurst exponent for the human eye-movements (pink) averaged over two directions: Hcorr(q) =(cid:0)H x
Figure 11: This figure demonstrates Hcorr(q) (left) of the simulated trajectories (blue) under the learned policy µ and the correla-
in the upper regime. The correlational Hurst exponents for negative q-orders declines with the growth of RMS contrast of noise
both for human eye-movements and the simulated trajectories. For the positive q-orders the correlational Hurst exponent is less
affected by the change of the visibility of the target. In general, the correlations weaken with the growth of the RMS contrast,
which is consistent with the decline of the geometrical persistence 8. The Hurst exponent of the shuffled time series (right), as
well as the correlational Hurst exponent, demonstrates the decline with the growth of the RMS contrast for negative q-orders
both for the human eye-movements and the simulated trajectories. In the subsection (5.1) we mentioned that the behaviour of
Hshuf (q) resembles the one related to time series of random values with power law distribution 26.
corr(q)(cid:1) /2
corr(q) + H y
tively describes the human visual behaviour during
the execution of the visual search task. The basis
of this model is the observer's representation of the
constraints of its own visual and oculomotor sys-
tems. We demonstrated that a consideration of the
temporal costs and uncertainty of the execution of
saccades results in the dramatic change of the basic
statistical properties and the scaling behavior of the
simulated time series.
We performed the multifractal analysis of our
data and discovered the presence of two types of
multifractality both in time series of the human eye-
movements and the model simulations. The multi-
fractality caused by the broad amplitude distribu-
tion of the saccades (the first type of multifractal-
ity) makes a significant contribution to the multi-
fractal behaviour of time series, which was not cov-
ered in the previous work [10]. After the estimation
of the correlational part of the Hurst exponent [38]
we confirmed the presence of the long-range posi-
tive correlations of the small saccades in the upper
regime. On the contrary, the large saccades exhibit
the weak long-range anti-correlations for the model
simulations and the human eye-movements in the
upper regime. As well as in the case of the geomet-
rical persistence, we found that the long-range cor-
relations between eye-movements weaken with the
decline of the target's visibility, which is consistent
with the previous work on this topic [12].
In this research we focused our attention more
on the persistence of eye-movements rather than on
their spatial distribution. That's why we didn't con-
sider the factors that are not directly related to the
trade-off between the temporal costs and the ex-
pected information gain. We estimate the optimal
policy under the assumption that the visual search
process is characterized by shift-rotational symme-
try [20], which was not observed in the previous
work with similar experimental settings [22]. The
symmetry of the visual search can be broken by an-
gular dependency of FPOC in both cases of normal
controls and patients with vision disabilities [43].
We plan to include the angular dependency to radial
and smoothing functions of policy (see eq. (A.4)) in
order to consider the asymmetry of the visual field
in our future works.
-10.25-10-8-6-400.2-2RMS of noiseq0Correlational H(q)H(q)20.1514680.1102humanlearned kernel0-100.25-8-60.5-40.2-2H(q)H(q)of shuffled dataqRMS of noise0120.15461.58100.1humanlearned kernel Appendix A.1 Kernel function
17
To sum up, this framework provides an elegant
explanation of scaling and persistent dynamic of the
voluntary saccades from an optimality point of view.
It clearly demonstrates that control models are able
to describe human eye-movements far beyond their
basic statistical properties.
Appendix A. Implementation of reinforce-
ment learning algorithms
Appendix A.1. Kernel function
We assume that the process of the visual search is
characterized by shift-rotational invariance [13]. In
this research we focus our attention on the persis-
tence of eye-movements rather than on their spatial
distribution. That's why we use the approximation
of the shift-rotational invariance in which we don't
need to consider the factors that are not directly
related to the trade off between the temporal costs
and the expected information gain, such as an asym-
metry of FPOC.
The coefficients in the set of dynamic equations
(4,6,11) are unaltered under any distance preserv-
ing transformations. The last dynamic equation,
which is the policy of gaze allocation (18), should
be shift-rotational invariant as well. The policy (18)
is determined by function of an expected reward
f (D, p). Due to the property of shift invariance we
can represent the function of an expected reward
with Volterra series [44]:
N(cid:88)
L(cid:88)
··· L(cid:88)
f (D, p) = f0+
Kn(D−l1, .., D−ln)
p(lj)
n(cid:89)
n=1
l1=1
ln=1
j=1
(A.1)
Where Kn(l1, .., ln) are called Volterra kernels. The
constant f0 is eliminated in the equation 18, and,
therefore, will not be considered. The dimensional-
ity of Volterra kernels Kn scales with the number
of the potential locations as Ln. The estimation of
Volterra kernel for n ≥ 2 is computationally unfeasi-
ble for the grid size in our simulations: L = 27 × 27.
For this reason we consider only the linear term:
f (D, p) =
K(D − l)p(l)
(A.2)
(cid:88)
l
We do not expect that the estimation of higher order
terms will result in improvement of the performance
of the policy. The current observation model 4 is
based on independent inputs Wl on each individual
location, which results in independence of the values
of the probability distribution pl for a sufficiently
large grid size. Therefore, higher order terms don't
provide additional information on the location of the
target.
f (D, p) = (cid:80)
The function of the expected reward should be
computed taking into account the current location
of the gaze A. Considering its rotational invari-
ance the most general
form of this function is:
K((cid:107)D − l(cid:107) ,(cid:107)D − A(cid:107))p(l). The soft-
max policy (18) for the function of the expected re-
ward is:
l
µ(Dn, pn) ∝
(cid:33)
(cid:32)(cid:88)
exp
pn(l)K ((cid:107)Dn − l(cid:107) ,(cid:107)Dn − An(cid:107))
(A.3)
l
Together with the set of the equations (4,6,11), this
form of the policy keeps the evolution of the system
invariant under any distance-preserving transforma-
tion. The convolution of the probability distribution
with the kernel function K(x, y) in general form A.3
is difficult to optimize, and the problem can be ef-
fectively solved only in a separable approximation:
K((cid:107)Dn − An(cid:107) ,(cid:107)Dn − x(cid:107))
≈ R ((cid:107)Dn − An(cid:107)) S ((cid:107)Dn − x(cid:107))
(A.4)
We call R and S the radial and the smoothing
functions correspondingly. The first one character-
izes the dependence of the expected reward on the
intended saccade length. The motivation behind
the introduction of the radial function R are both
growing uncertainty of the fixation placement (6)
and the duration of the step (11) with the length
of the saccade. We assume that the radial func-
tion R equals zero outside an interval [amin, amax],
where amin and amax are minimal and maximal sac-
cade length correspondingly. The minimal saccade
Appendix A.3 REINFORCE
18
√
length amin = 1 deg [45] is chosen as a magnitude
of the shortest possible voluntary movement. The
2 · 15 deg is equal
maximal saccade length amax =
to the length of the diagonal of the stimulus image
in our experiments. The smoothing function S de-
scribes the relative contribution of the surrounding
locations to the reward. The smoothing function
has the same role as a term F (see eq.
(15)) in
the definition of the information maximization pol-
icy π0, and it basically defines how meaningful the
certain location is without consideration of the time
costs of a relocation.
The form of policy (A.3) in the separable approx-
imation is:
µ(Dn, pn) ∝
exp (R ((cid:107)Dn − An(cid:107)) (pn ∗ S) (Dn))
(A.5)
which is used in the simulation of the trajectories
and the training phase. Two heuristic policies pre-
sented in section 3.1 are both special cases of the
general form of a policy in the separable approxi-
mation(A.5).
Appendix A.2. Parametrization of policy
The radial R(x) and the smoothing S(x) functions
are represented with Fourier-Bessel series:
(cid:40) (cid:80)Ξ
(cid:16) u1:ξ(x−amin)
amax−amin
(cid:17)
0,
else
R(x) =
ξ=1 rξJ1
, amin < x < amax
(cid:40) (cid:80)Ξ
ξ=1 sξJ0
0,
S(x) =
(cid:0) u0:ξx
(cid:1) , x < b
b
else
(A.6)
(A.7)
where ui:ξ are zeros of Bessel function of order i
and b is the radii of the visual field. This repre-
sentation allows us to control the dimensionality
of the kernel and to effectively store the policy in
memory. The choice of orders (i = 0, 1) of Bessel
functions in (A.6,A.7) is caused by boundary con-
ditions for the radial and the smoothing functions:
R(amin) = R(amax) = 0; S(b) = 0. The boundary
conditions on the radial function forbid the model
observer to fixate the same location again R(0) = 0
and to make unlikely large saccades R(amax) = 0.
The condition on the smoothing function S(b) = 0
corresponds to the absence of any information gain
from remote locations, and, therefore, their irrele-
vance to the process of fixation selection. So, the
policy µ(ϑ) is represented by set of parameters:
ϑ ≡ (r0:Ξ, s0:Ξ ).
Appendix A.3. REINFORCE
We solve the optimization problem for the value
function (12) with a policy gradient algorithm
adopted from [32]. This optimization procedure is
represented as an iterative process of a gradient es-
timation and an update of the policy parameters at
the end of each training epoch - the sequence of M
episodes.
Repeat
optimal
each
m
t
element
ξ:
baseline
gradient
2. Estimate
1. Perform a training epoch with M episodes
and get the sequence of observations, ac-
tions and costs for each time step t and
episode m: (pt,m, at,m, Vm).
(cid:18)(cid:80)
(cid:80)
(cid:18)(cid:80)
(cid:80)
(cid:18)(cid:80)
(cid:80)
for
ηξ
=
(bξ − Vm)
4. Update policy parameters: ϑ ← ϑ + αη
∇ξ log µϑ(at,m,pt,m)
∇ξ log µϑ(at,m,pt,m)
∇ξ log µϑ (at,m, pt,m)
the
element:
(cid:19)2
(cid:19)2
3. Estimate
gradient
(cid:19)2
m
t
each
for
=
bξ
m
t
Vm
until gradient η converges.
Appendix A.4. PGPE
The second approach to the optimization prob-
lem (12) is a parameter exploring policy gradient
presented in [33]. As well as in the previous section,
we estimate the gradient and update the policy pa-
rameter at the end of each training epoch. We use
a symmetric sampling of the policy parameters for
gradient estimation. At the beginning of each step
i
j
M(cid:88)
(cid:1)2 − σ2
j=1
i
σi
j
(cid:1)
(cid:0)V −
j − V +
(cid:33)(cid:32)
(cid:104)V (cid:105) − V +
(A.8)
(cid:33)
j + V −
j
2
µi = µi + α
(cid:32)(cid:0)i
j
M(cid:88)
σi = σi + α
j=1
we generate the perturbation from normal distri-
bution N (0, Iσ2) and create the symmetric param-
eter samples ϑ+ = µ + and ϑ+ = µ − , where µ is
the current values of the policy parameters for the
training epoch. Then we simulate one episode for
each parameter sample and denote the cost V + for
the episode generated with ϑ+ , and V − for ϑ− cor-
respondingly. At the end of each training epoch the
policy parameters and the standard deviation of the
distribution of perturbation are updated according
to the equations:
(A.9)
j is the perturbation for the
where for jth episode i
parameter i and V ±
are sampled costs. The cost
baseline is chosen as a mean cost for the training
epoch.
j
Appendix A.4.1. Convergence of policy gradient
The Markov decision process defined by set of
dynamic equations (4,6,11,A.5) was simulated on
N × N grid, which comprises the N 2 possible tar-
get locations, where N=128. At the beginning of
the optimization procedure we pick the policy pa-
rameters ϑ randomly from the uniform distribution
U (−0.5, 0.5) and fix parameter λ = 0.001. For both
algorithms we use the same parametrization of pol-
icy. The training epoch for both PGPE and RE-
INFORCE consists of 400 episodes. Learning rate
α = 0.2 was the same for both algorithms.
Figure A.12 illustrates the performance of two
policy gradient methods we used for search of the
optimal policy for the case of FPOC corresponding
to en = 0.25 and et = 0.2. Both algorithms used
Fourier-Bessel parametrization of policy with a di-
mensionality Ξ = 45 for the radial and the smooth-
ing functions. REINFORCE performed better for
all parameter settings. On average, it takes around
19
Figure A.12: The performance of parameter exploration pol-
icy gradient (PGPE) and episodic REINFORCE with an op-
timal baseline.
50 and 40 learning epochs to converge for REIN-
FORCE and PGPE correspondingly. The choice of
the dimensionality higher than 45 doesn't improve
the performance of both algorithms.
Figure A.13 shows the results of optimization: the
radial R(x) and the smoothing S(x) functions. Both
REINFORCE and PGPE provide close estimates of
the smoothing and the radial functions for eccen-
tricity smaller than < 3◦.
In order to compare
the solution with the heuristic policies (15,17), we
presented FPOC on the same plot with the smooth-
ing function. The smoothing function provided by
REINFORCE is monotonously decreasing as well as
FPOC, whereas for PGPE we have a fluctuating so-
lution with a decreasing amplitude of oscillations.
The behavior of the radial function is similar for
both solutions, with higher amplitude of oscillations
for PGPE solution.
Appendix B. Implementation of MF-DFA
Appendix B.1. Multifractal analysis
In this chapter we present the details of MF-DFA
algorithm used here for calculation of the general-
ized Hurst exponent. All of this section is based on
Kantelhardt et al.[38].
Learning epochs01020304050Response time, sec1.522.533.5PerformancePGPEREINFORCE Appendix B.2
Interpolation of simulated trajectories
20
Figure A.13: The results of the optimization: the smoothing S(x) funtion (left) and the radial R(x) function (right).
The procedure of MF-DFA starts with definition
of a profile for time series X = {x1, ..., xN} with a
compact support:
i(cid:88)
Y (i) =
(xk − (cid:104)x(cid:105))
(B.1)
k=1
The profile Y (i) is divided on Ns ≡ int(N/s) seg-
ments, where s is chosen among some linear space
s ∈ S = {smin,smin + (cid:52)s, ..., smax}. The segmenta-
tion starts from the beginning of time series, there-
fore, there are residual N ÷ s number of the ele-
ments at the end of time-series. In order to process
the residual elements, the segmentation is also per-
formed from the end of time series. So, at the end of
segmentation procedure we have 2Ns segments for
each value of s.
The calculation of the variance is based on an
approximation of local trend for each segment ν =
1, ..., Ns with a polynomial function yv. Then, the
variance on each segment is calculated as:
s(cid:88)
i=1
1
s
F 2(ν, s) =
{Y [(ν − 1) s + i] − yv(i)} (B.2)
for each segment ν = 1, ..., Ns and
s(cid:88)
i=1
1
s
F 2(ν, s) =
{Y [N − (ν − Ns) s + i] − yv(i)}2
(B.3)
for ν = Ns+1, ..., 2Ns. The order m of polynomial
function must satisfy the condition m ≤ s − 2. The
variance over all segments are averaged to obtain
the qth order fluctuation function:
(cid:40)
2Ns(cid:88)
ν=1
1
2Ns
(cid:2)F 2(ν, s)(cid:3)q/2
(cid:41)1/q
Fq(s) =
(B.4)
According to the properties of qth order fluctua-
tion function [46], the scaling behavior of Fq(s) is
governed by the generalized Hurst exponent:
Fq(s) ∼ sH(q)
(B.5)
The value of H(q) is usually obtained through a
linear regression of log2 (Fq(s)).
Appendix B.2. Interpolation of simulated trajecto-
ries
We perform MF-DFA analysis on the magnitude
of the saccadic events simulated by MDP defined
above. Each episode of MDP provides the sequence
of vectors of gaze positions: A1, ..., AN. In order to
01245678910eccentricity, deg-0.500.511.522.5weightSmoothing functionREINFORCEFPOCPGPE012345678910eccentricity, deg01234weightRadial functionREINFORCEPGPE Appendix B.3 Multifractality of simulated trajectories
21
get the time series of the gaze allocation in real time
- ¯A, we follow the simple procedure of an interpola-
tion:
(cid:80)N
• The calculation of the duration of each time
step n with 11, the total time of episode T =
i=1 Θi and start time of each discrete step:
(cid:26)
(cid:80)n−1
Tn =
0, if n = 1
i=1 Θi, if n > 1
(B.6)
• The choice of the length of the real-time se-
quence M = 25 ∗ T , which corresponds to 40
millisecond resolution.
• For each element t of ¯A we define, which dis-
crete time step it belongs: Tn ≤ t
25 < Tn+1.
t
• If time step t of ¯At corresponds to the fixation
25 − Tn <
during the discrete time step n :
Θf ix(n), than ¯At = An. In the other case, if
time t corresponds to the saccadic movement
within the discrete time interval n, we have:
¯At = An +τsac
Therefore, we have defined the function that
maps the discrete sequence A to real-time se-
quence ¯A.
(cid:0) t
25 − Tn − Θf ix(n)(cid:1).
An+1−An
An+1−An
The real-time sequences ¯A from 1000 episode corre-
sponding to each policy are merged, and the result-
ing sequences ¯Aµ, ¯Aπ1, ¯Aπ0 are analyzed with MF-
DFA.
Appendix B.3. Multifractality of simulated trajecto-
ries
We perform MF-DFA over the differentiated tra-
jectories generated with PO-MDP under the heuris-
tic policies π0, π1 and the learned policy µ. Before
differentiation trajectories were represented as real
time sequences with the procedure of the interpola-
tion Appendix B.2.
The model presented here is not devoted to FEM
and can't describe the combined movement of both
FEM and the saccades. The results of our analy-
sis should be compared with the scaling behavior
of Fq(s) for human eye-movements on the scales
s ≥ 256 ms, which corresponds to the upper regime.
Therefore, we set the minimal time scale smin =
256ms. The choice of smax = 2∗ 103ms corresponds
to the average length of episode. We assume that
there is no correlation between the episodes due to a
random location of the first fixation and the location
of the target.
The figure B.14 demonstrates the scaling of the
q-order fluctuation function Fq(s) (B.4) for simu-
lated trajectory under the infomax greedy policy π0
for the conditions: et = 0.2, en = 0.15. The red,
blue and green lines correspond to the linear ap-
proximation of the function log2 (Fq(s)) for the or-
ders q = {−10; 0; 10}. The scaling of Fq(s) doesn't
exhibit the crossover for positive q-orders on an
interval of scales [smin,smax], however the behav-
ior of log2 (Fq(s)) deviates from linear at the large
scales s ∼ smax. The simulations on different grid
sizes, which correspond to different average time of
task execution, have shown that the interval of lin-
ear behavior of log2 (Fq(s)) always coincides with
[smin,smax]. The scaling of Fq(s) on [smin,smax] is
different for different orders q and, therefore, the
trajectories ¯Aπ1 are multifractal time series.
The figure B.15 demonstrates our estimates of the
correlational Hurst exponent Hcorr(q) (left) and the
Hurst exponent of shuffled time series Hshuf (q) of
time series simulated under the different policies.
As well as in the case of the human eye-movements,
two types of multifractality present in the simulated
time series. The behavior of Hshuf (q) resembles the
power-law distribution scenario 25 for all policies,
except the infomax greedy π0. The distribution of
the saccade length doesn't correspond to the power-
law for π0 , which was demonstrated on figure 6 .
On the contrary, the infomax rate π1 and the learned
policy µ generate the movement with the distribu-
tional multifractality that presents in human eye-
movements as well.
For all policies the correlational Hurst exponent
is positive for the negative q-orders. This indicates
the presence of long-range correlations for small fluc-
tuations. The large fluctuations are anticorrelated
for π1 and exhibit weak anti-correlation for µ. We
observe the last scenario for upper regime of hu-
man eye-movements 10, where the large fluctuations
demonstrate a weak anticorrelation in a contrary to
Appendix B.3 Multifractality of simulated trajectories
22
Figure B.14: The scaling of the q-order fluctuation function Fq(s) (left) estimated for the simulated trajectories under the
infomax rate policy π1, and the generalized Hurst exponent H(q) (right) computed through a linear regression of log2 (Fq(s)).
The red, blue and green lines correspond to the linear approximation of function log2 (Fq(s)) for orders q = {−10; 0; 10}. The
scaling of Fq(s) doesn't exhibit the crossover for the positive q-orders on an interval of scales [smin,smax].
Figure B.15: The Hurst exponent of the shuffled time series Hshuf (q) (right) and the correlational Hurst exponent Hcorr(q)
(left) of the trajectories simulated under the different policies. As well as in the case of human eye-movements, two types of
multifractality present in the simulated time series. The behavior of Hshuf (q) resembles the power-law distribution scenario
25 for all policies, except the infomax greedy π0. The distribution of saccade length doesn't correspond to the power-law for
π0 , which was demonstrated on the figure 6 . On the contrary, the infomax rate π1 and the learned policy µ generate the
movement with the distributional multifractality that presents in human eye-movements as well.
25651210282048Scale, ms01234log2(Fq)Scaling function Fq (q-order RMS)q =-10q =0q =10q =-10q =0q =10-10-8-6-4-20246810q0.20.40.60.811.2Hqq-order Hurst exponent-10-8-6-4-20246810q-0.500.511.5H(q)Correlational H(q)infomax ratelearned kernelinfomax greedy-10-8-6-4-20246810q0.450.50.550.60.650.70.750.8H(q)H(q) of shuffled time series positively correlated small fluctuations.
Appendix C. Implementation
of
chophysical experiments
psy-
We set a goal to reproduce the eye tracking exper-
iment described in [11]. In this section we provide
the description of the psychophysical experiments.
Appendix C.1. Participants
The group of nine patients with normal to
corrected-to-normal vision participated in the ex-
periment. The group included four postgraduate
students (age 23 ± 7, 4 males) from Queen Mary
University of London. This group was aware of
the experimental settings and passed 10 minutes
of training sessions with four different experimental
conditions, which correspond to the certain value
of the RMS contrast of background noise. The ex-
periments were approved by the ethics committee
of Queen Mary University of London and informed
consent was obtained.
Appendix C.2. Equipment
We used DELL P2210 22" LCD monitor (resolu-
tion 1680 × 1050, refresh rate 60 Hz) driven by a
Dell Precision laptop for all experiments. The eye
movements of the right eye were registered using Eye
Tracker device SMI-500 with a sampling frequency
of 120 Hz. The Eye tracker device was mounted on
the monitor. Matlab Psychtoolbox was used to run
the experiments and generate the stimulus images.
Appendix C.3. Stimulus and procedure
The participants set in front of the monitor with
their heads fixed with a chin rest at a distance of
110 cm from the monitor. The monitor subtended
a visual angle of 21 × 15 deg. Each participant was
shown the examples of the stimulus image before the
experiments and was instructed to fixate the target
object as fast as possible and to press the certain
button on a keyboard to indicate that they found the
target. All four participants completed one practice
session with 40 trials before the experiment.
23
The stimuli were the static images generated be-
fore each session according to the description from
the original experiment [11]. The 1/f noise was gen-
erated on a square region on the screen, which spans
the visual angle of 15 × 15 deg. The target was sine
−1 framed by symmetric raised cosine.
grating 6 deg
The target appeared randomly at any possible loca-
tion on the stimuli image within the square region.
The experiments were provided for one level of RMS
contrast of target et = 0.2 and several levels of 1/f
noise RMS contrast en ∈ (0.1, 0.15, 0.2, 0.25).
The participants completed four experimental ses-
sions with 120 trials. The experimental session
started after inbuilt nine-point grid calibration of
the eye-tracking device. The participants were given
3 minutes of rest between sessions. One of 120 stim-
uli images were shown at the beginning of each trial.
The participants are assumed to perform the vi-
sual search task, which is finished by pressing the
"END" button.
In our experimental settings, the
signal from the participants was blocked for 300 ms
from the start of each trial.
If the gaze position
measured by the eye tracking device is in the vicin-
ity of 2 deg around the location of the target at the
moment participant presses the "END" button, the
task is considered successful. Due to the presence
of a temporal delay between the moments of local-
ization of target and pushing of "END" button we
block the signal from END button for 400 ms. After
a completion of each trial the central fixation cross
was shown for 500 ms, then the next trial started
and new stimulus image was shown to participants.
Appendix D. Influence of saccade latency
According to the literature, the saccade program-
ming is assumed to be the two-stage process that
consists of labile and non-labile stages [17]. The la-
bile stage is the first stage of the saccade program-
ming, during which the initial saccade command can
be cancelled in a favour of saccade to another loca-
tion. The saccade to the next location is executed
after the non-labile stage. The visual input is active
during both labile and non-labile stages and sup-
pressed during the execution of a saccade. There-
REFERENCES
24
Figure C.16: The completion time (left) and the saccade length (right) for the human observers, the simulated agents with the
initial observation model and the extended one.
fore, the decision Dn is made at the end of the labile
stage, and the visual input received at the location
An during non-labile stage of saccade programming
can be used for decision-making only at the next
step n + 1. As a result, the observer receives two
separate observation vectors Wn+1
nlb from
the previous and the current fixation locations. The
observation model 4 doesn't take into account the
duration of the observation. We assume that ob-
servation vector is integrated continuous-time Gaus-
w(t, l, A)dt, which
sian white noise: W (l, A) =(cid:82) t+Θ
and Wn+1
lb
satisfies following:
t
1. E[w(t, l, A)] = δl,m/θ0
2. E[w(t1, l1, A), w(t2, l2, A)] = δt1 ,t2 δl1,l2
θ0F 2((cid:107)l−A(cid:107))
where θ0 = 250 ms is a time interval is the detec-
tion experiment [11], for which the visibility maps
were measured. This model of noise generalizes the
"noisy observation" paradigm for variable fixation
duration. The result of integration of continuous
time noise w is Gaussian white noise W with mean
Θ·E[w] and variance Θ·σ2[w]. Next, we assume that
the duration of the non-labile stage is Θnlb = 41, 6
ms [47] and the rest of fixation duration is allocated
for the labile stage Θlb ≈ 200 ms (the average fix-
ation duration according to our data: Θf ix = 240
ms). Using 1 and 2 we compute mean and variance
of observation inputs Wn+1
nlb , and suc-
cessively apply the equation 3 to evaluate the belief
state pn+1.
and Wn+1
lb
We learned the policy of gaze allocation for the
extended observation model using REINFORCE.
We compared the basic characteristic of trajecto-
ries simulated under this policy with the simulations
for the initial model and data from the human ob-
servers (Look at C.16. The initial model outper-
formed the extended one, but no significant differ-
ence was found. We didn't expect any significant
difference in performance, because the observer re-
ceives the same amount of information on average
in both models.
References
References
[1] R. Engbert, K. Mergenthaler, P. Sinn,
A. Pikovsky, An integrated model of fixational
eye movements and microsaccades, Proceedings
of the National Academy of Sciences 108 (39)
(2011) E765 -- E770.
[2] R. Engbert, K. Mergenthaler, Microsaccades
are triggered by low retinal image slip, Pro-
ceedings of the National Academy of Sciences
103 (18) (2006) 7192 -- 7197.
[3] R. Engbert, R. Kliegl, Microsaccades keep the
eyes' balance during fixation, Psychological Sci-
ence 15 (6) (2004) 431 -- 431.
0;.10.150.20.25RMS contrast of noise0246810complection time, secinitial modelhumanextended model0.10.150.20.25RMS contrast of noise123456Saccade length, deg REFERENCES
25
[4] B. W. Tatler, M. M. Hayhoe, M. F. Land,
D. H. Ballard, Eye guidance in natural vision:
Reinterpreting salience, Journal of vision 11 (5)
(2011) 5.
[5] A. Borji, L. Itti, State-of-the-art in visual at-
tention modeling, Pattern Analysis and Ma-
chine Intelligence, IEEE Transactions on 35 (1)
(2013) 185 -- 207.
[6] D. H. Ballard, M. M. Hayhoe, Modelling the
role of task in the control of gaze, Visual cog-
nition 17 (6-7) (2009) 1185 -- 1204.
[7] H.-k. Ko, M. Poletti, M. Rucci, Microsaccades
precisely relocate gaze in a high visual acuity
task, Nature neuroscience 13 (12) (2010) 1549 --
1553.
[15] P. Grigolini, G. Aquino, M. Bologna,
M. Luković, B. J. West, A theory of 1/f
noise in human cognition, Physica A: Statis-
tical Mechanics and its Applications 388 (19)
(2009) 4192 -- 4204.
[16] R. J. van Beers, The sources of variability in
saccadic eye movements, The Journal of Neu-
roscience 27 (33) (2007) 8757 -- 8770.
[17] R. Engbert, A. Nuthmann, E. M. Richter,
R. Kliegl, SWIFT: a dynamical model of sac-
cade generation during reading., Psychological
review 112 (4) (2005) 777.
[18] J. Tian, H. S. Ying, D. S. Zee, Revisiting cor-
rective saccades: role of visual feedback, Vision
research 89 (2013) 54 -- 64.
[8] P. Sinn, R. Engbert, Small saccades versus
microsaccades: Experimental distinction and
model-based unification, Vision research .
[19] S. P. Lee, J. B. Badler, N. I. Badler, Eyes alive,
in: ACM Transactions on Graphics (TOG),
vol. 21, ACM, 637 -- 644, 2002.
[9] J. Najemnik, W. S. Geisler, Simple summa-
tion rule for optimal fixation selection in visual
search, Vision Research 49 (10) (2009) 1286 --
1294.
[20] N. J. Butko, J. R. Movellan, I-POMDP: An
infomax model of eye movement,
2008
7th IEEE International Conference on Devel-
opment and Learning, IEEE, 139 -- 144, 2008.
in:
[10] T. A. Amor, S. D. Reis, D. Campos, H. J. Her-
rmann, J. S. Andrade Jr, Persistence in eye
movement during visual search, Scientific re-
ports 6.
[11] J. Najemnik, W. S. Geisler, Optimal eye
movement strategies in visual search, Nature
434 (7031) (2005) 387 -- 391.
[12] D. G. Stephen, J. Anastas, Fractal fluctuations
in gaze speed visual search, Attention, Percep-
tion, & Psychophysics 73 (3) (2011) 666 -- 677.
[13] N. J. Butko, J. R. Movellan, Infomax control
of eye movements, IEEE Transactions on Au-
tonomous Mental Development 2 (2) (2010) 91 --
107.
[21] C. Bradley, J. Abrams, W. S. Geisler, Retina-
V1 model of detectability across the visual field,
Journal of vision 14 (12) (2014) 22 -- 22.
[22] J. Najemnik, W. S. Geisler, Eye movement
statistics in humans are consistent with an op-
timal search strategy, Journal of Vision 8 (3)
(2008) 4 -- 4.
[23] A. E. Bartz, Eyemovement latency, duration,
and response time as a function of angular dis-
placement., Journal of Experimental Psychol-
ogy 64 (3) (1962) 318.
[24] T. A. Salthouse, C. L. Ellis, Determinants of
eye-fixation duration, The American journal of
psychology (1980) 207 -- 234.
[14] S. Wallot, C. A. Coey, M. J. Richardson, Cue
predictability changes scaling in eye-movement
fluctuations, Attention, Perception, & Psy-
chophysics 77 (7) (2015) 2169 -- 2180.
[25] S. Lebedev, P. Van Gelder, W. H. Tsui, Square-
root relations between main saccadic param-
eters., Investigative Ophthalmology & Visual
Science 37 (13) (1996) 2750 -- 2758.
REFERENCES
26
[26] I. T. C. Hooge, C. J. Erkelens, Control of fix-
ation duration in a simple search task, Percep-
tion & Psychophysics 58 (7) (1996) 969 -- 976.
[27] R. Kliegl, A. Nuthmann, R. Engbert, Track-
ing the mind during reading: the influence of
past, present, and future words on fixation du-
rations., Journal of experimental psychology:
General 135 (1) (2006) 12.
[28] K. Rayner, G. W. McConkie, S. Ehrlich, Eye
movements and integrating information across
fixations., Journal of Experimental Psychol-
ogy: Human Perception and Performance 4 (4)
(1978) 529.
[29] H. H. Greene, The control of fixation duration
in visual search, Perception 35 (3) (2006) 303 --
315.
[30] P. J. Unema, S. Pannasch, M. Joos, B. M.
Velichkovsky, Time course of information pro-
cessing during scene perception: The relation-
ship between saccade amplitude and fixation
duration, Visual cognition 12 (3) (2005) 473 --
494.
[31] V. Navalpakkam, C. Koch, A. Rangel, P. Per-
ona, Optimal reward harvesting in complex
perceptual environments, Proceedings of the
National Academy of Sciences 107 (11) (2010)
5232 -- 5237.
[32] J. Peters, S. Schaal, Policy gradient methods
for robotics, in: 2006 IEEE/RSJ International
Conference on Intelligent Robots and Systems,
IEEE, 2219 -- 2225, 2006.
[33] F. Sehnke, C. Osendorfer, T. Rückstiess,
A. Graves,
J. Schmidhuber,
Parameter-exploring policy gradients, Neural
Networks 23 (4) (2010) 551 -- 559.
J. Peters,
[34] B. W. Tatler, R. J. Baddeley, B. T. Vincent,
The long and the short of it: Spatial statis-
tics at fixation vary with saccade amplitude and
task, Vision research 46 (12) (2006) 1857 -- 1862.
[35] M. S. Castelhano, M. L. Mack, J. M. Hender-
son, Viewing task influences eye movement con-
trol during active scene perception, Journal of
Vision 9 (3) (2009) 6 -- 6.
[36] W. S. Geisler, J. S. Perry, J. Najemnik, Visual
search: The role of peripheral information mea-
sured using gaze-contingent displays, Journal of
Vision 6 (9) (2006) 1 -- 1.
[37] S. Isogami, M. Matsushita, Structural and
statistical properties of self-avoiding fractional
Brownian motion, Journal of the Physical So-
ciety of Japan 61 (5) (1992) 1445 -- 1448.
[38] J. W. Kantelhardt,
S. A. Zschiegner,
E. Koscielny-Bunde, S. Havlin, A. Bunde,
H. E. Stanley, Multifractal detrended fluctu-
ation analysis of nonstationary time series,
Physica A: Statistical Mechanics and its
Applications 316 (1) (2002) 87 -- 114.
[39] A. Rosas, E. Nogueira Jr, J. F. Fontanari,
Multifractal analysis of DNA walks and trails,
Physical Review E 66 (6) (2002) 061906.
[40] S. Dutta, D. Ghosh, S. Chatterjee, Multifractal
detrended fluctuation analysis of human gait
diseases, Frontiers in physiology 4 (2013) 274.
[41] N. Vandewalle, M. Ausloos, M. Houssa,
P. Mertens, M. Heyns, Non-Gaussian behavior
and anticorrelations in ultrathin gate oxides af-
ter soft breakdown, Applied physics letters 74
(1999) 1579.
[42] E. A. F. E. Ihlen, Introduction to multifractal
detrended fluctuation analysis in Matlab, Fron-
tiers in physiology 3 (2012) 141.
[43] S. Van der Stigchel, R. A. Bethlehem, B. P.
Klein, T. T. Berendschot, T. Nijboer, S. O. Du-
moulin, Macular degeneration affects eye move-
ment behavior during visual search, Frontiers in
psychology 4 (2013) 579.
[44] V. Volterra, M. Long, Theory of Function-
als and of
Integral and Integro-differential
Equations, Blackie and Son Limited, URL
REFERENCES
27
https://books.google.co.uk/books?id=
iKUVswEACAAJ, 1944.
[45] J. Otero-Millan, X. G. Troncoso, S. L. Mack-
nik, I. Serrano-Pedraza, S. Martinez-Conde,
Saccades and microsaccades during visual fixa-
tion, exploration, and search: foundations for a
common saccadic generator, Journal of Vision
8 (14) (2008) 21 -- 21.
[46] H.-O. Peitgen, H. Jürgens, D. Saupe, Chaos
and fractals: new frontiers of science, Springer
Science & Business Media, 2006.
[47] R. Engbert, A. Longtin, R. Kliegl, A dynamical
model of saccade generation in reading based
on spatially distributed lexical processing, Vi-
sion research 42 (5) (2002) 621 -- 636.
|
0905.0701 | 3 | 0905 | 2011-05-17T14:56:45 | How effective delays shape oscillatory dynamics in neuronal networks | [
"q-bio.NC",
"nlin.PS",
"q-bio.QM"
] | Synaptic, dendritic and single-cell kinetics generate significant time delays that shape the dynamics of large networks of spiking neurons. Previous work has shown that such effective delays can be taken into account with a rate model through the addition of an explicit, fixed delay [Roxin et al. PRL 238103 (2005)]. Here we extend this work to account for arbitrary symmetric patterns of synaptic connectivity and generic nonlinear transfer functions. Specifically, we conduct a weakly nonlinear analysis of the dynamical states arising via primary instabilities of the asynchronous state. In this way we determine analytically how the nature and stability of these states depend on the choice of transfer function and connectivity. We arrive at two general observations of physiological relevance that could not be explained in previous works. These are: 1 - Fast oscillations are always supercritical for realistic transfer functions. 2 - Traveling waves are preferred over standing waves given plausible patterns of local connectivity. We finally demonstrate that these results show a good agreement with those obtained performing numerical simulations of a network of Hodgkin-Huxley neurons. | q-bio.NC | q-bio |
How effective delays shape oscillatory dynamics in
neuronal networks
Alex Roxina,b, Ernest Montbri´oa,c,1
aComputational Neuroscience, Department of Information and Communication
Technologies, Pompeu Fabra University, Barcelona 08018, Spain.
bCenter for Theoretical Neuroscience, Columbia University, New York, New York, USA.
cCenter for Neural Science, New York University, New York, 10012 New York, USA.
Abstract
Synaptic, dendritic and single-cell kinetics generate significant time delays that
shape the dynamics of large networks of spiking neurons. Previous work has
shown that such effective delays can be taken into account with a rate model
through the addition of an explicit, fixed delay [1, 2]. Here we extend this work
to account for arbitrary symmetric patterns of synaptic connectivity and generic
nonlinear transfer functions. Specifically, we conduct a weakly nonlinear analy-
sis of the dynamical states arising via primary instabilities of the asynchronous
state. In this way we determine analytically how the nature and stability of
these states depend on the choice of transfer function and connectivity. We
arrive at two general observations of physiological relevance that could not be
explained in previous works. These are: 1 - Fast oscillations are always super-
critical for realistic transfer functions. 2 - Traveling waves are preferred over
standing waves given plausible patterns of local connectivity. We finally demon-
strate that these results show a good agreement with those obtained performing
numerical simulations of a network of Hodgkin-Huxley neurons.
Keywords:
delay, neuronal networks, neural field, amplitude equations, Wilson-Cowan
networks, rate models, oscillations
PACS: 87.19.lj, 87.19.lp, 05.45.-a, 84.35.+i, 89.75.-k
1. Introduction
When studying the collective dynamics of cortical neurons computationally,
networks of large numbers of spiking neurons have naturally been the bench-
mark model. Network models incorporate the most fundamental physiological
properties of neurons: sub-threshold voltage dynamics, spiking (via spike gen-
eration dynamics or a fixed threshold), and discontinuous synaptic interactions.
For this reason, networks of spiking neurons are considered to be biologically
realistic. However, with few exceptions, e.g.
[3 -- 5], network models of spiking
neurons are not amenable to analytical work and thus constitute above all a
Preprint submitted to Elsevier
November 5, 2018
computational tool. Rather, researchers use reduced or simplified models which
describe some measure of the mean activity in a population of cells, oftentimes
taken as the firing rate (for reviews, see [6, 7]). Firing-rate models are simple,
phenomenological models of neuronal activity, generally in the form of contin-
uous, first-order ordinary differential equations [8, 9]. Such firing-rate models
can be analyzed using standard techniques for differential equations, allowing
one to understand the qualitative dependence of the dynamics on parameters.
Nonetheless, firing-rate models do not represent, in general, proper mathemati-
cal reductions of the original network dynamics but rather are heuristic, but see
[10]. As such, there is in general no clear relationship between the parameters
in the rate model and those in the full network of spiking neurons, although for
at least some specific cases quasi-analytical approaches may be of value [11]. It
therefore behooves the researcher to study rate models in conjunction with net-
work simulations in order to ensure there is good qualitative agreement between
the two.
Luckily, rate models have proven remarkably accurate in capturing the main
types of qualitative dynamical states seen in networks of large numbers of asyn-
chronously spiking neurons. For example, it is well known that in such networks
the different temporal dynamics of excitatory and inhibitory neurons can lead
to oscillations. These oscillations can be well captured using rate models [8].
When the pattern of synaptic connectivity depends on the distance between
neurons, these differences in the temporal dynamics can also lead to the emer-
gence of waves [9, 12 -- 14]. This is certainly a relevant case for local circuits in
cortical tissue, where the likelihood of finding a connection between any two
neurons decreases as a function of the distance between them, e.g. [15].
When considering the spatial dependence of the patterns of synaptic connec-
tivity between neurons, one must take into account the presence of time delays
due to the finite velocity propagation of action potentials along axons. Such
delays depend depend linearly on the distance between any two neurons. This
has been the topic of much theoretical study using rate models with a space-
dependent time delay e.g. [13, 14, 16 -- 24]. The presence of propagation delays
can cause an oscillatory instability of the unpatterned state leading to homoge-
neous oscillations and waves [17, 20]. The weakly nonlinear dynamics of waves
in spatially extended rate models, i.e. describing large-scale (on the order of
centimeters) activity, is described by the coupled mean-field Ginzburg-Landau
equations [22], and thus exhibits the phenomenology of small amplitude waves
familiar from other pattern forming systems [25]. Also, it is important to note
that discrete fixed delays have been used to model the time delayed interaction
between discrete neuronal regions, as well as to model neuronal feedback, e.g.
[26 -- 30].
Localized solutions of integro-differential equations describing neuronal ac-
tivity, including fronts and pulses, are also affected by distance-dependent ax-
onal delays [13, 14, 17 -- 19]. Specifically, the velocity of propagation of the lo-
calized solution is proportional to the conduction velocity along the axon for
small conduction velocities, while for large conduction velocities it is essentially
constant. This reflects the fact that the propagation of activity in neuronal
2
tissue is driven by local integration in which synaptic and membrane time con-
stants provide the bottleneck. Also, allowing for different conduction velocities
for separate excitatory and inhibitory populations can lead to bifurcations of
localized bump states to breathers and traveling pulses [23].
Although the presence of time delays in the nervous system are most often
associated with axonal propagation, significant time delays are also produced by
the synaptic kinetics and single-cell dynamics (see the next section for a detailed
discussion about the origin of such effective time delays in networks of spiking
neurons). As a relevant example for the present work, it was shown in [1, 2] that
the addition of an explicit, non-space-dependent delay in a rate equation was
sufficient to explain the emergence of fast oscillations prevalent in networks of
spiking neurons with strong inhibition and in the absence of any explicit delays.
Specifically, in [1, 2] the authors studied a rate model with a fixed delay on
a ring geometry with two simplifying assumptions. First they assumed that the
strength of connection between neurons could be expressed as a constant plus
the cosine of the distance between the neurons. Secondly, they assumed a linear
rectified form for the transfer function which relates inputs to outputs. These
assumptions allowed them to construct a detailed phase diagram of dynamical
states, to a large degree analytically. In addition to the stationary bump state
(SB) which had been studied previously [31, 32], the presence of a delay led
to two new states arising from primary instabilities of the stationary uniform
state (SU): an oscillatory uniform state (OU) and a traveling wave state (TW).
Secondary bifurcations of these three states (SB,OU,TW) led to yet more com-
plex states including standing waves (SW) and oscillatory bump states (OB).
Several regions of bistability between primary and secondary states were found,
including OU-TW, OU-SB and OU-OB. They subsequently confirmed these re-
sults through simulations of networks of Hodgkin-Huxley neurons. Despite the
good agreement between the rate equation and network simulations, several
important issues remain unresolved:
• The rate equation predicted that the primary instability of the SU state
to waves should be to traveling waves, while in the network simulations
standing waves were robustly observed.
• The linear-threshold transfer function, albeit amenable to analysis, nonethe-
less leads to degenerate behavior at a bifurcation point. Specifically, any
perturbations with a positive linear growth rate will continue to grow until
the lower threshold of the transfer function is reached. This means that
the amplitude of new solution branches at a bifurcation is always finite,
although the solution itself may not be subcritical. In a practical sense
then, this means that it is not possible to assess whether a particular solu-
tion, for example oscillations or bumps, will emerge continuously from the
SU state as a parameter is changed, or if it will appear at finite amplitude
and therefore be bistable with the SU state over some range.
• The previous work only considered a simplified cosine connectivity. More
realistic patterns of synaptic connectivity such as a Gaussian dependence
3
of connection strength as a function of distance might lead to different
dynamical regimes. It remains to be explored the effect of a general con-
nectivity kernel in the dynamics of both the rate equation and the spiking
neurons network with fixed time delays.
In order to address these issues, and provide a more complete analysis of the
role of fixed delays in neuronal tissue, we here study a rate equation with delay
without imposing any restrictions on the form of the transfer function beyond
smoothness or on the shape of the connectivity kernel beyond being symmetric.
Our approach is similar to that of Curtu and Ermentrout in [33], who extended
a simplified rate model with adaptation for orientation selectivity [32] to include
a nonlinear transfer function and general connectivity kernel. Here we do the
same for a rate model with a fixed time delay.
Thus in what follows we will study a rate equation with fixed delay and spa-
tially modulated connectivity. In conjunction with this analysis we will conduct
numerical simulations of a network of large numbers of spiking neurons in order
to assess the qualitative agreement between the rate model and the network for
the delay-driven instabilities, which are the primary focus of this work.
This article is organized as follows: In section 2 we provide an overview
of the origin of the effective delay. We do this by looking at the dynamics of
synaptically coupled conductance-based neurons. This will motivate the pres-
ence of an explicit fixed delay in a rate-model description of the dynamics in
recurrently coupled networks of neurons.
In section 3 we formulate the rate
model and conduct a linear stability analysis of the SU state. In section 4 we
conduct a weakly nonlinear analysis for the four possible primary instabilities
of the SU state (asynchronous unpatterned state in a network model), thereby
deriving amplitude equations for a steady, Turing (bumps), Hopf (global oscilla-
tions), and Turing-Hopf (waves) bifurcations. We will focus on the delay-driven
instabilities, i.e. Hopf and Turing-Hopf. Finally, in section 5 we will study
the interactions of pairs of solutions: bumps and global oscillations, and global
oscillations and waves, respectively.
2. The origin of effective time delays
This section is intended to provide an intuitive illustration of the origin of
an effective fixed delay in networks of spiking neurons. A detailed, analytical
study of this phenomenon can be found in [3, 34, 35].
Fig.1 illustrates the origin of the effective delay in networks of model neu-
rons.
In this case we look at a single neuron pair: one presynaptic and one
postsynaptic. The single-neuron dynamics are described in detail in Appendix
A. The top panel of Fig.1 shows the membrane potential of an excitatory neu-
ron subjected to a current injection of Iapp = 0.2µA/cm2 which causes it to
fire action potentials. Numerically, an action potential is detected whenever the
membrane voltage exceeds 0mV from below. When this occurs, an excitatory
postsynaptic current (EPSC) is generated in the post-synaptic neuron, as seen
in the second panel. This current is generated by the activation of an excitatory
4
conductance which has the functional form of a difference of exponentials with
rise time τ1 = 1ms and decay time τ2 = 3ms. The different colored curves
correspond to different conductance strengths: black gE = 0.05, red gE = 0.1,
blue gE = 0.2 and orange gE = 0.4 mS·ms/cm2. The resulting excitatory post-
synaptic potential (EPSP) in millivolts is shown in the third panel. At this
point it is already clear that the postsynaptic response, although initiated here
simultaneously with the presynaptic action potential, will nonetheless take a
finite amount of time to bring the postsynaptic cell to threshold, thereby alter-
ing its firing rate. This is shown in the bottom panel. In this case the weakest
input (black) was insufficient to cause the neuron to spike, while the other three
inputs were all strong enough to cause an action potential. The latency until
action potential firing is a function of the synaptic strength, with the latency
going to zero as the synaptic strength goes to infinity. The very long latency for
gE = 0.1 mS·ms/cm2 (red curve) is due in part to the intrinsic action potential
generating mechanism. In fact, an input which brings the neuron sufficiently
close to the bifurcation to spiking can generate arbitrarily long latencies.
Fig.2 illustrates how this effective delay is proportional to both the rise and
decay time of the EPSC. In the top panel, the decay time is fixed at 3ms while
the rise time is varied, while in the bottom panel, the rise time is fixed at 0.1ms
and the decay time is varied. From these figures it is clear that the effective
delay is proportional to both the rise and decay times. Simulations with an
EPSC modeled by a jump followed by a simple exponential decay reveal that
the effective delay is proportional to the decay time in this case (not shown).
It is instructive to note that the effective delay, due to the time course of the
synaptic kinetics in the model neuron, can be captured by modeling the EPSC
as a Dirac delta function with a fixed delay. This is shown in Fig.3. In Fig.3,
the curves shown in black are the same as in Fig. 1 for gE = 0.4 mS·ms/cm2,
while the red EPSC is a Dirac delta function which arrives with a fixed delay of
3.2ms. Note that the decay of the EPSP and the postsynaptic spike time are well
captured here. The fact that a jump in voltage with a fixed delay can capture
the effect of having continuous synaptic kinetics was described already in [3]. In
that work, the authors studied a network of recurrently coupled integrate-and-
fire neurons with inhibitory synapses, the time course of which was modeled
as jump in the voltage at a fixed delay. They showed that the fixed delay led
to the emergence of fast oscillations, the period of which was proportional to
approximately several times the delay. The advantage of using EPSCs modeled
as Dirac delta functions is that the input current is delta-correlated in time,
allowing one to solve the associated Fokker-Planck equation for the distribution
of the membrane voltages in a straightforward way.
Subsequent work studied the emergence of fast oscillations in networks of
integrate-and-fire neurons with continuous synaptic kinetics [34]. There the au-
thors determined the frequency of oscillations analytically and found that it
is proportional to both the rise and decay times of the synaptic response. An
extension of that work showed that for networks of Hogkin-Huxley conductance-
based neurons, the frequency of oscillations also depends on the single cell dy-
namics and specifically the membrane time constant and action potential gener-
5
Figure 1: (Color online) The synaptic time course introduces an effective fixed delay in the
interaction between model neurons. Top panel: The membrane voltage of the presynaptic
neuron which is driven by a steady input current of Iapp = 0.2µA/cm2 causing it to spike
periodically. Here only one action potential is shown. The time of the action potential,
defined as the time at which the membrane potential crosses zero mV from below, is indicated
by the dashed line labeled tpre. Second panel from top: The EPSC in the postsynaptic cell
is initiated at time tpre with no delay. It consists of a difference-of-exponential conductance
change with rise and decay times of 1 and 3 milliseconds respectively, times the voltage
of the postsynaptic cell (the reversal potential for excitatory synapses is 0mV). The colors
indicate different values of the maximum conductance: black gE = 0.05, red gE = 0.1, blue
gE = 0.2 and orange gE = 0.4 mS·ms/cm2. See Appendix A for details of the model and
explanation of units. Third panel from top: The EPSP in the postsynaptic cell. These curves
were generated by eliminating the action potential generating currents from the model, i.e.
Na and K, and subtracting off the rest potential ∼ −64mV. Bottom panel: The membrane
potential of the postsynaptic cell with N a and K currents intact. Note that the time of the
postsynaptic action potential, indicated by the dashed lines, approaches tpre with increasing
synaptic strength. The very long latency at time t3 is due to an input which puts the cell
membrane potential very close to threshold and is therefore due in part to the action potential
generating mechanism of the model. The postsynaptic cell is driven by a steady input current
of Iapp = 0.1µ A/cm2 which is insufficient to cause it to spike.
6
-60-40-20020Vpre (mV)tpre00.020.040.060.08EPSC (µA/cm2)051015EPSP (mV)859095100105110115120125time (ms)-60-40-2002040Vpost (mV)t1t2t3 Figure 2: (Color online) The effective delay is affected by both the rise and decay time of the
EPSC. Top: The decay is kept fixed at 3ms and the rise time is varied. Bottom: The rise
time is kept fixed at 0.1ms and the decay time is varied. The parameter values are the same
as in Fig.1 with gE = 0.4 mS· ms/cm2.
7
-60-40-2002040Vpre (mV)00.050.10.15EPSC (µA/cm2)τ1 = 0.1τ1 = 1τ1 = 3859095100105time (ms)-60-40-2002040Vpost (mV)τ2 = 3-60-40-2002040Vpre (mV)00.511.5EPSC (µA/cm2)τ2 = 0.1τ2 = 1τ2 = 10859095100105time (ms)-80-60-40-2002040Vpost (mV)τ1 = 0.1 Figure 3: (Color online) Continuous synaptic kinetics can be replaced by a jump in the voltage
which occurs at a fixed delay. The black curves in the four panels are the same as in Fig.1
with gE = 0.4 mS·ms/cm2. The red curves show the effect of an EPSC modeled as a Dirac
delta function which occurs with a delay of 3.2ms after the presynaptic action potential.
ation mechanism [35]. This is consistent with the effect of the synaptic response
and single-cell dynamics on the response latency that we have illustrated above.
Thus the same mechanism which generates an effective delay in the response
of a postsynaptic neuron to a single excitatory presynaptic input, can also gen-
erate coherent oscillations in a network of neurons coupled through inhibitory
synapses. This can be seen in Fig. 4, which shows the results of simulations
of a network of 2000 recurrently coupled inhibitory cells. The single-cell model
is the same conductance based model used in Figs.1 -3, see Appendix A for
details. Synaptic connections are made between neurons with a probability of
p0 = 0.2, leading to a sparse, random connectivity with each cell receiving an
average of 400 connections. Synapses are modeled as the difference of exponen-
tials with a rise time and a decay time of 1ms and 3ms respectively and gI = 0.01
mS·ms/cm2. All cells receive uncorrelated Poisson inputs with a rate of 12000Hz
and gext = 0.0019mS·ms/cm2, and there is no delay in the interactions. Fig.
4A shows a raster plot of the network activity in the top panel. The activity
is noisy although periods of network synchrony are visible. The middle panel
shows the firing rate averaged over all neurons in time bins of 0.1ms (black)
and smoothed by averaging with a sliding window of 10ms (red). The large
fluctuations in the firing rate indicate network synchrony, while the averaged
trace shows clear periodic oscillations. This is even more evident in the bottom
panel which shows the subthreshold input current averaged over all cells. One
can clearly see the ongoing oscillation, the amplitude of which undergoes slow
8
-60-40-20020Vpre (mV)00.020.040.060.08EPSC (µA/cm2)05101520EPSP (mV)859095100105110115120125time (ms)-80-60-40-2002040Vpost (mV) fluctuations due to the noisy dynamics. Fig. 4B shows the smoothed firing rate
(top) and average input current (bottom) from the same simulation, but on a
shorter time scale. Note that the sign of the input current has been inverted
so that downward deflections mean increasing positive currents. Here it is clear
that the input current is a delayed copy of the firing rate, with a delay on the
order of ∼ 2 − 3ms which matches with the time scale of the synaptic kinetics
(τ1 = 1ms, τ2 = 3ms).
Therefore, Fig. 4 provides a clear prescription for developing a rate model
description of fast oscillations in networks in the asynchronous regime. The
input a neuron receives is not simply a nonlinear function of the instantaneous
firing rate, rather it is a function of the delayed firing rate. Thus one should
introduce a fixed delay in the rate model description. This was the underlying
assumption behind the work in [1].
Before presenting the model we would like to emphasize that fixed delays,
which are primarily due to synaptic and dendritic integration, and conduction
delays due to the propagation of action potentials along the axon, are both
present in real neuronal systems.
Importantly, this means that the delay in
neuronal interactions at zero distance is not zero.
In fact, fixed delays are
always observed in paired intracellular recordings in cortical slices. The latency
from the start of the fast rising phase of the action potential to the start of the
post-synaptic current (or potential) has been measured for pairs of pyramidal
cells in rat layers 3 to 6 and is on the order of milliseconds, see [37] for a
recent review. Recordings from cat cortex and between pyramidal cells and
other cells including spiny cells and interneurons in the rat cortex also reveal
fixed delays which are rarely less than a millisecond. These delays are seen
when neurons are spatially adjacent, indicating that axonal propagation is not
an important contributing factor. On the other hand the speed of propagation
of action potentials along unmyelinated axons in mammals is on the order of
10−1 − 101 m/s, which means a delay of 0.1-10 ms for neurons separated by
1 millimeter [38, 39]. Thus fixed delays and conduction delays are of similar
magnitude within a local patch of cortex and both would be expected to shape
the dynamics of non-steady activity, i.e. neither is negligible. Here we have
decided to focus on fixed delays, as in previous work [1, 2], due both to their
physiological relevance and prevalence in networks of spiking neurons.
3. The rate model with fixed time delay
An effective delay roughly proportional to the time scale of the post-synaptic
currents is always present in networks of spiking neurons as we have illustrated
in the previous section and has been shown extensively elsewhere, e.g. [34, 35].
In particular, this is true for networks in the asynchronous regime, for which
a rate-equation description is, in general, appropriate. Given this, we consider
here a rate model with fixed delay. Specifically, we study a heuristic equation
describing the activity of a small patch of neural tissue consisting of two pop-
ulations of recurrently coupled excitatory and inhibitory neurons respectively.
9
Figure 4: (Color Online) The effective delay can lead to oscillations in networks of recurrently
coupled inhibitory neurons. A. Raster plot of the spiking activity in a network of 2000 in-
hibitory cells (top). The average firing rate shows large fluctuations (center, black), which
when smoothed are clearly identifiable as noisy, periodic oscillations (red). Network oscilla-
tions are also clearly visible in the average subthreshold input (bottom). B. A blowup of the
average firing rate (top) and average input current (bottom) from the simulation in A. The
input current is clearly a time-delayed copy of the firing rate. The delay is between 2 and
3ms.
10
8101214161820firing rate (Hz)380385390395400405410time (ms)-0.6-0.4-0.200.20.4-<Isyn> (µA/cm2)effective delayB (cid:21)
(cid:21)
(1a)
. (1b)
Our formulation is equivalent to the Wilson-Cowan equations without refrac-
tory period [8], and with spatially dependent synaptic connectivity which was
studied originally in [40]. Additionally, we consider a fixed delay in the neuronal
interactions. Given these assumptions, the full Wilson-Cowan equations are
τe re = −re + Φe
dy[Jee(x − y)re(y, t − de) − Jei(x − y)ri(y, t − di)] + Ie
,
τi ri = −ri+Φi
dy[Jie(x−y)re(y, t−de)−Jii(x−y)ri(y, t−di)]+Ii
Ω
In the original formulation [8], re(x, t) and ri(x, t) represent the average number
of active cells in the excitatory and inhibitory populations respectively, in this
case at a position x and at a time t. The time constant τe (τi) is roughly the
time it takes for a an excitatory (inhibitory) cell receiving "at least threshold
excitation" [8] to generate a spike. This can reasonably be taken as the mem-
brane time constant which is generally on the order of 10-20 ms. The functions
Φa(x)(a = e, i) are usually taken to be sigmoidal. Specifically, if all neurons in
the population receive equal excitatory drive, and there is heterogeneity in some
parameter across neurons, e.g. the threshold to spiking, which obeys a unimodal
distribution, then the fraction of active neurons is just the integral over the
distribution, up to the given level of excitation. The integral of a unimodal dis-
tribution is sigmoidal. In Eqs.(1a-1b), the functions Jab(x)(a = e, i)(b = e, i)
represent the strength of synaptic connection from a neuron in population b to
a neuron in population a separated by a distance x. Here the neurons are ar-
ranged in one dimension on a domain Ω. Input from excitatory (inhibitory) cells
is furthermore delayed by a fixed amount de (di), which, as we have discussed
in the introduction, is on the order of one millisecond. Finally, the excitatory
and inhibitory populations are subject to an external drive of strength Ie and
Ii respectively.
A general analysis of Eqs.(1a-1b) would be technically arduous although it is
a natural extension of the work presented here. Rather, we choose to study the
dynamics of this system under the simplifying assumption that the excitatory
and inhibitory neurons follow the same dynamics, i.e. τe = τi = τ , de = di = d,
Jee = Jie = Je, Jei = Jii = Ji, Φe = Φi = Φ, Ie = Ii = I. If this the case, then
re = ri = r and the variable r follows the dynamics given by
r(x, t) = −r(x, t) + Φ
dyJ(x − y)r(y, t − D) + I
,
(2)
(cid:20)(cid:90)
(cid:20)(cid:90)
Ω
(cid:20) 1
(cid:90) π
2π
−π
(cid:21)
where we have chosen the domain Ω to be a ring of normalized length L = 2π.
Furthermore, we have re-scaled time by the time constant τ . The normalized
delay is therefore D = d/τ , which is the ratio of the effective delay in neuronal
interactions to the integration time constant and should be much less than one
in general. The synaptic connectivity expressed in terms of the excitatory and
inhibitory contributions is J(x) = Je(x) − Ji(x), and thus represents an
effective mixed coupling which may have both positive and negative regions.
11
Eq.(2) with the choice of Φ(I) = I for x > 0 and 0 otherwise and with
J(x) = J0 + J1 cos (x) is precisely the model studied in [1, 2]. We now wish to
study Eq.(2) for arbitrary choices of Φ(I) and J(x).
In presenting Eq.(2) we have relied on the heuristic physiological motivation
first put forth in [8]. Nonetheless, as a phenomenological model, the terms and
parameters in Eq.(2) may have alternative and equally plausible interpretations.
Indeed, the variable r is often thought of as the firing rate as opposed to the
fraction of active cells, in which case the function Φ(I) can be thought of as the
transfer function or fI curve of a cell.
Experimentally the function Φ(I) has been found to be well approximated
by a power-law nonlinearity with a power greater than one [36, 41]. Modeling
studies show that the same nonlinearity applies to integrate-and-fire neurons
and conductance based neurons driven by noisy inputs [36]. Therefore it may
be that such a choice of Φ leads to better agreement of Eq.(2) with networks of
spiking neurons and hence with actual neuronal activity. More fundamentally,
we may ask if choosing Φ as a sigmoid or a power law qualitatively alters the
dynamical states arising in Eq.(2). This is precisely why we choose here not to
impose restrictions on Φ but rather conduct an analysis valid for any Φ. How
the choice of Φ affects the generation of oscillations and waves is an issue we
will return to in the corresponding sections of this paper.
3.1. Linear stability analysis
Stationary uniform solutions (SU) of Eq.(2) are given by
R = Φ
J0R + I
,
(3)
where R is a constant non-zero rate, J0 is the zeroth order spatial Fourier
coefficient of the symmetric connectivity which can be expressed as
J(x) = J0 +
Jkeikx + c.c.
(4)
(cid:105)
(cid:33)
(cid:104)
(cid:32) ∞(cid:88)
k=1
∞(cid:88)
and k is an integer. Depending on the form of Φ, Eq.(3) may admit one or
several solutions.
We study the linear stability of the SU state with the ansatz
r(x, t) = R +
δrkeikx+α(k)t,
(5)
where δrk (cid:28) 1 and the spatial wavenumber k is an integer due to the periodic
boundary conditions. Plugging Eq.(5) into Eq.(2) leads to an equation for the
complex eigenvalue α(k)
k=0
α(k) = −1 + Φ
(cid:48)
Jke−α(k)D,
(6)
(cid:48)
where the slope Φ
is evaluated at the fixed point given by Eq.(3). The real and
imaginary parts of the eigenvalue α(k) = λ(k)+iω(k) represent the linear growth
12
rate and frequency of perturbations with spatial wavenumber k respectively. At
the bifurcation of a single mode, the growth rate will reach zero at exactly one
point and be negative elsewhere. That is, λ(kcr) = 0 for the critical mode kcr.
Given this, Eq.(6) yields the dispersion relation for the frequency of oscillation
of the critical mode
iω(kcr) = −1 + Φ
(cid:48)
Jkcr e−iω(kcr)D.
(7)
From Eq.(7) it is clear that the wavelength of the critical mode depends crucially
on the synaptic connectivity. In particular, the spatial Fourier coefficients of the
connectivity kernel J(x) depend on the wavenumber k, i.e. Jk = J(k). Thus, the
critical wavenumber is, in effect, selected by the choice of connectivity kernel.
It is in this way that the nature of the instability depends on the synaptic
connectivity at the linear level.
Depending on the values of ω and kcr in Eq.(7) at the bifurcation from the
SU state, four types of instabilities are possible:
• Steady (ω = 0, kcr = 0): The instability leads to a global increase in
activity.
• Turing (ω = 0, kcr (cid:54)= 0): The instability leads to a stationary bump state
(SU).
• Hopf (ω (cid:54)= 0, kcr = 0): The instability leads to an oscillatory uniform
state (OU).
• Turing-Hopf (ω (cid:54)= 0,kcr (cid:54)= 0): The instability leads to waves (SW, TW).
For the non-oscillatory instabilities (i.e. ω = 0), Eq.(7) gives the critical value
¯Jk = 1/Φ(cid:48)
(8)
while for the oscillatory ones Eq.(7) is equivalent to the system of two transcen-
dental equations
Note that we have defined the critical values as Jkcr ≡ ¯Jk and ωcr ≡ ¯ω.
3.1.1. The small delay limit (D → 0)
It is possible to gain some intuition regarding the effect of fixed delays on
the dynamics, by deriving asymptotic results in the limit of small delay. This
limit is a relevant one physiologically, since fixed delays are on the order of a few
milliseconds and the integration time constant is about an order of magnitude
larger. Therefore throughout this work we will present asymptotic results, and
compare them to the full analytical formulas as well as to numerical simulations.
In the limit D → 0, the asymptotic solutions of Eq.(9a) can be easily ob-
tained graphically. Fig.5 shows two curves (black and grey) representing the
13
¯ω = − tan ¯ωD,
¯ω = −Φ
(cid:48) ¯Jk sin ¯ωD.
(9a)
(9b)
Figure 5: The critical frequency at the instability to oscillations is given by the intersection
of the grey (D = 0.1) and black curves, the left and right hand sides of Eq.(9a) respectively.
As D → 0, solutions clearly approach φ = (2n + 1)π/2 (n integer). Eq.(9b) shows that the
first potentially unstable mode corresponds to the solutions φ = ±π/2 (see text).
right and left hand sides of Eq.(9a) respectively, where we defined φ ≡ ¯ωD.
The intersections of these curves correspond to the roots of Eq.(9a). The
plot shows three solutions, the trivial one φ = 0 (corresponding to the non-
oscillatory instabilities), and an infinite number of solutions that clearly ap-
proach φ = (2n + 1)π/2 (n integer) in the small delay limit, since the slope of
the straight line goes to infinity as D → 0. Substituting these solutions into
Eq. (9b), we find that the first potentially unstable solution of the kth spatial
Fourier mode is φ = ±π/2, that occurs at the critical value of the coupling
with a frequency
¯Jk = − π
2DΦ
(cid:48) ,
¯ω =
π
2D
.
(10)
(11)
Fig. (6) shows the critical frequency and coupling as a function of the delay, up
to a delay D = 1. The solution obtained from the dispersion relation Eqs.(9a)
and (9b) are given by solid lines, while the expressions obtained in the small
delay limit are given by dotted lines. Thus the expressions in the small delay
limit agree quite well with the full expressions even for D = 1.
3.2. An illustrative Phase Diagram
Throughout the analysis which follows we will illustrate our results with
a phase diagram of dynamical states. Specifically, we will follow the analysis
in [1, 2] in constructing a phase diagram of dynamical states as a function of
14
-π-π/20π/2πφ=ω D-15015-tan φφ/D Figure 6: Top: The critical frequency of oscillatory instabilities as a function of the delay D
from the dispersion equation Eq.(9a) (solid line) and in the small delay limit (dotted line).
Bottom: The critical coupling as a function of the delay D from Eq.(9b) (solid) and in the
small delay limit (dotted).
J0 and J1, the first two Fourier coefficients of the synaptic connectivity. We
will set the higher order coefficients to zero for this particular phase diagram,
although we will discuss the effect of additional modes in the text. Furthermore,
unless otherwise noted, for simulations we choose a sigmoidal transfer function
Φ(I) = α
1+e−βI with α = 1.5 and β = 3. As we vary the connectivity in the phase
diagram, we also vary the constant input I in order to maintain the same level
of mean activity, i.e. we keep R = 0.1 fixed. For the values of the parameters
we have chosen here this results in I ∼ −0.1J0 − 0.88. We also take D = 0.1
unless noted otherwise
The primary instability lines for the SU state can be seen in the phase
diagram, Fig. (7). The region in (J0, J1) space where the SU state is stable is
shown in gray, while the primary instabilities, listed above, are shown as red
lines.
In Section 4 we will provide a detailed analysis of the bumps, global
oscillations and waves (SB, OU and SW/TW) which arise due to the Turing,
Hopf and Turing-Hopf instabilities respectively. The derivation of the amplitude
equations is given in Appendix B, as well as a brief discussion of the steady,
transcritical bifurcation which occurs for strong excitatory coupling and is not
of primary interest for this study. Finally, in Section 5 we will analyze the
codimension 2 bifurcations: Hopf and Turing-Hopf (OU and waves), and Turing
and Hopf (SU and OU). This analysis will allow us to understand the dynamical
states which appear near the upper and lower left hand corners of the grey
shaded region in Fig. (7), i.e. the SW/OU and OB states.
4. Bifurcations of codimension 1
As we are interested in creating a phase diagram as a function of the con-
nectivity, we will take changes in the connectivity as the bifurcation parameter.
15
110100ωcr0.11D1101001000-Jcr Figure 7: (Color online) Phase diagram of the rate model Eq.(2). In each region, the type of
solution seen in numerical simulations is indicated by a letter code: SU - stationary uniform
(grey region), HA - high activity, SB - stationary bump, OB - oscillatory bump, SW - standing
waves, TW - traveling waves. Solid lines indicate analytical expressions. In particular, the four
possible instabilities of the SU state are depicted in red (thick lines correspond to subcritical
bifurcations) and are given by the linear stability criteria Eqs.(8-9b). The four lines emanating
from the upper and lower left corners of the SU region were determined from a weakly nonlinear
analysis at the two corners (codimension two points) [see section IV]. The region marked OB
corresponds to a mixed mode solution of SB-OU, while in the lower left-hand region the OU
and SW solutions are bistable. Parameters: Φ(x) =
1+e−βx where α = 1.5 and β = 3. We
consider the coupling function J(x) = J0 + 2J1 cos x. The time delay is D = 0.1 and the input
current I is varied so as to keep the uniform stationary solution fixed at R = 0.1.
α
16
-100-500J0-100-500J1SUSBHAOUTWOBSB/HATW/HASWSU/HASU/SWSWTW/SWSW/OUTuring-HopfTuringTranscriticalHopf The small parameter is therefore defined by the expansion
Jk = ¯Jk + 2∆Jk,
(12)
The perturbative method we apply, which makes use of this small parameter, is
called the multiple-scales method and is a standard approach for determining
the weakly nonlinear behavior of pattern-forming instabilities [25]. We choose
the particular scaling of 2 in the foreknowledge that if the amplitudes of the
patterns of interest are scaled as , a solvability condition will arise at order 3.
This solvability condition yields a dynamical equation governing the temporal
evolution of the pattern (see Appendix A for details). Without loss of generality
we will assume that an instability of a nonzero spatial wavenumber is for k = 1.
We will furthermore co-expand the constant input I so as to maintain a fixed
value for the spatially homogeneous steady state solution R
I = ¯I + 2∆I,
r = R + r1 + 2r2 + . . . ,
(13)
(14)
where the small parameter is defined by Eq.(12). Additionally we define the
slow time
T = 2t.
(15)
4.1. Turing Bifurcation
The emergence and nature of stationary bumps in rate equations have been
extensively studied elsewhere, e.g. [40]. We briefly describe this state here for
completeness. The kth spatial Fourier mode of the connectivity is given by
the critical value Eq. (8), while we assume that all other Fourier modes are
sufficiently below their critical values to avoid additional instabilities. Without
loss of generality we assume k = 1 here.
We expand the parameters J1, I and r as in Eqs. (12,13,14), and define
the slow time Eq. (15). The solution of Eq.(2) linearized about the SU state
R is a spatially periodic amplitude which we allow to vary slowly in time, i.e.
r1 = A(T )eix + c.c.. Carrying out a weakly nonlinear analysis to third order in
leads to the amplitude equation
∂T A = η∆J1A + ΓA2A,
with the coefficients
Γ =
¯J 3
1
1 + D
(cid:32)
(cid:48)
Φ
η =
1 + D
,
J2(Φ(cid:48)(cid:48))2
2(1 − J2Φ
(cid:33)
.
(cid:48)(cid:48)(cid:48)
Φ
2
+
(cid:48)
)
(cid:48)(cid:48)
)2
J0(Φ
1 − J0Φ
(cid:48) +
(16)
(17a)
(17b)
The nature of the bifurcation (sub- or supercritical) clearly depends strongly
on the sign and magnitude of mean connectivity J0 and the second spatial
Fourier mode J2. Figure 8 shows a phase diagram of the bump state at the crit-
ical value of ¯J1 = 3.54. The red lines indicate oscillatory and steady instability
17
Figure 8: (Color online) The phase diagram for stationary bumps as a function of the zeroth
and second spatial Fourier modes of the connectivity kernel. The region of bistability between
the unpatterned and the bump state is shaded. Here the critical spatial Fourier coefficient
¯J1 = 3.54. Red lines indicate the boundaries of the SU state (obtained via Eqs.(8-9b). The
functions Φ and J(x) as well as the input current I and the delay D are taken as in Fig. (7).
Insets: example connectivity patterns corresponding to the values of J0 and J2 marked by
the square and triangle respectively. Note that standard Mexican Hat connectivity tends to
favor bistability.
boundaries for the modes J0 and J2. Clearly J0 < 0 and J2 < 0 over most of
the region of allowable values, and the bump is therefore supercritical. There is
only a narrow region of predominantly positive values (shaded region in Fig. 8
for which the cubic coefficient is positive. This indicates that the bifurcating
solution branch is unstable. However, neuronal activity is bounded, which is
captured in Eq.(2) by a saturating transfer function Φ. Thus the instability will
not grow without bound but rather will saturate, producing a finite amplitude
bump solution. This stable, large amplitude branch and the unstable branch
annihilate in a saddle-node bifurcation for values of J1 below the critical value
for the Turing instability. Such finite-amplitude bumps are therefore bistable
with the SU state. In Fig. 8, the two insets show the connectivity kernel J(x)
for parameter values given by the placement of the open triangle (subcritical
bump) and the open square (supercritical bump).
In the phase diagram (7), the Turing instability line (upper horizontal red
line) is shown thin for supercritical, and thick for subcritical bumps (here J2 =
0).
4.2. Hopf Bifurcation
4.2.1. Network simulations
As shown elsewhere previously [1, 3, 4], a network of recurrently coupled
inhibitory neurons can generate fast oscillations due to the effective delay in
the synaptic interactions. Fig.9A shows raster plots of the spiking activity in
18
-60-40-200J2-60-40-200J0SubcriticalSupercritical+++----- Figure 9: Fast oscillations in a sparse, randomly connected network of 2000 randomly con-
nected inhibitory neurons without delay (pI
0 = 0.2). A: Raster plots of spiking activity in four
networks with increasing values of the synaptic weight. From top to bottom: gI = 0, 0.005,
0.008, 0.01mS·ms/cm2. The mean firing rate was kept fixed at approximately 14Hz by varying
the input current. From top to bottom: νext = 2150, 7700, 10000 and 12000Hz. See Appendix
A for additional model details. B: The autocorrelation function of the network-averaged firing
rate for the four networks.
19
020406080100τ (ms)-0.200.20.4AC(τ)B a network randomly connected inhibitory neurons as the synaptic weight gI is
increased (see Appendix A for a description of the network). The raster plots
clearly show oscillations emerging as the parameter gI increases in strength.
Fig.9B shows the autocorrelation function of the network-averaged firing rate
from four seconds of simulation time. The curve is completely flat for the un-
coupled case as expected, while periodic peaks appear and grow as the synaptic
weights are increased, indicating the presence of coherent network oscillations.
Note that the oscillations appear to emerge continuously, which would indicate
a supercritical bifurcation. This is consistent with the finding in [3], where
fast oscillations in a network of integrate-and-fire neurons were shown to be
supercritical analytically.
We may ask if fast oscillations are in general expected to bifurcate super-
critically, or if a subcritical bifurcation is also possible. To this end we study
the rate equation which allows us to detemine the nature of the bifurcation
analytically as a function of the transfer function and connectivity.
4.2.2. Rate model
In the rate model there is a spatially homogeneous oscillatory instability
with frequency ω given by Eq.(9a). This occurs for a value of the 0th spatial
Fourier mode of the connectivity given by Eq.(9b), while we assume that all
other Fourier modes are sufficiently below their critical values to avoid additional
instabilities. We expand the parameters J0, I and r as in Eqs. (12,13,14), and
define the slow time Eq. (15). The linear solution has an amplitude which we
allow to vary slowly in time, i.e. r1 = H(T )eiωt + c.c.. Carrying out a weakly
nonlinear analysis to third order in leads to the amplitude equation
∂T H = (µ + iΩ)∆J0H + (α + iβ)H2H,
(18)
where the coefficients (µ + iΩ) and (α + iβ) are specified by the Eqs. (B.4) and
(B.5) in the Appendix B.
Figure 10 shows a typical bifurcation diagram (in this case J1 = 0) for the
Hopf bifurcation. Plotted is the amplitude of the limit cycle as a function of J0
where symbols are from numerical simulation of Eq.(2) and the lines are from
the amplitude equation, Eq.(18).
In the small delay limit (D → 0) we can use the asymptotic values (11) to
obtain, to leading order,
(cid:48)
,
2 + i)Φ
1 + π2
4
µ + iΩ = − ( π
(cid:32)
11π − 4
(cid:48)(cid:48)
(cid:20) 11 + π
10
(cid:21)(cid:33)
(19a)
,(19b)
(cid:48)(cid:48)(cid:48)
− Φ
2
α + iβ = − χ
(DΦ(cid:48))3
(Φ
Φ
)2
(cid:48) − πΦ
4
(cid:48)(cid:48)(cid:48)
− i
20
where we have defined the quantity χ ≡ π3/(8 + 2π2). Figure 11 shows a
comparison of the full expressions for the coefficients of the amplitude equation,
20
Figure 10: The bifurcation diagram for a supercritical Hopf bifurcation. Shown is the ampli-
tude of the limit cycle as a function of the 0th order spatial Fourier coefficient of the coupling
J(x). Open circles are from numerical simulation of Eq.(2) and solid lines show the solution
from the amplitude equation, Eq.(16). The functions Φ and J(x) as well as the input current
I and the delay D are taken as in Fig. 7.
Eqs. (B.4-B.5) with the expressions obtained in the limit D → 0, Eqs. (19a-19b).
Again, the agreement is quite good, even up to D = 1, especially for the real
part of the cubic coefficient α, which is of primary interest here.
The asymptotic expression for the cubic coefficient α, Eq.(19b), indicates
)2 > (11π − 4)/(5π).
that a subcritical limit cycle should occur for Φ
This provides a simple criterion for determining whether or not a particular
choice of the transfer function can generate oscillations which are bistable with
the SU state.
In fact, it is a difficult condition to fulfill given a sigmoidal-
like input-output function. For example, given a sigmoid of the form Φ(x) =
α/(1 + e−βx), one finds that
/(Φ
Φ
(cid:48)(cid:48)(cid:48)
(cid:48)(cid:48)
(cid:48)
(cid:48)(cid:48)(cid:48)
(cid:48)
Φ
Φ
(cid:48)(cid:48)
(Φ
)2 = 1 − 2
e−3βx
(e−4βx − 2e−3βx + e−2βx)
.
(20)
It is straightforward to show that the expression of the right hand side of
)2 < 1 < (11 −
Eq.(20) is bounded above by 1.
4π)/(5π) ∼ 1.95. Such a nonlinear transfer function will therefore always gen-
erate supercritical oscillations.
In fact, −∞ ≤ Φ
/(Φ
Φ
(cid:48)(cid:48)(cid:48)
(cid:48)(cid:48)
(cid:48)
If the nonlinear transfer function is interpreted as the single-cell fI curve,
which is common in the literature, then we can use the fact that cortical cells
operate in the fluctuation-driven regime. In particular, the mean input current
to cortical cells is too low to cause spiking. Rather, this occurs at very low rates
due to fluctuations in the membrane voltage. Although the fI curve for spiking
neurons in the supra-threshold regime is concave down and saturates, in the
fluctuation-driven, sub-threshold regime the fI curve exhibits a smoothed out
tail which is concave up. It has been shown that the sub-threshold portion of
the fI curve of actual cells can be well fit by a function of the form Φ(x) = Axγ,
21
-64-62-60-58-56J000.01LC amplitudenumericsamplitude eqs. Figure 11: Top: The real part of the linear coefficient µ. Bottom: Minus the real part of the
cubic coefficient −α. Solid lines are from the full expressions Eq.(B.4-B.5) and dotted lines
are the leading order terms in the small delay limit, Eqs. (19a-19b). The functions Φ and
J(x) as well as the input current I are taken as in Fig. (7).
where γ > 1 (see e.g. [36, 41]). In this case
(cid:48)(cid:48)(cid:48)
(cid:48)
Φ
(Φ
Φ
(cid:48)(cid:48)
)2 = 1 − 1
γ − 1
,
which again is bounded between −∞ and 1. This again rules out subcritical
oscillations in the small delay limit.
Nonetheless, suitable functions Φ for generating subcritical oscillations can
be contrived, as shown in Fig. 12 A. Numerical simulation of Eq.(2) indeed
reveals a subcritical bifurcation in this case (see Fig. 12 B). However, this type
of transfer function does not seem consistent with the interpretation of Φ as a
single-cell fI curve, nor with that of Φ as a cumulative distribution of activation,
i.e. a sigmoid. This strongly suggests that delay-driven oscillations in networks
of spiking neurons will be generically supercritical.
4.3. Turing-Hopf Bifurcation
4.3.1. Network simulations
As shown previously in [1], given an inverted Mexican-hat connectivity for
which inhibition dominates locally, fast waves may emerge in networks of spiking
neurons. This is illustrated in Fig.13A, where raster plots of three networks are
shown with the degree of spatial modulation increasing from top to bottom.
Additionally, Fig. 13B (top) shows a spatial profile of the network activity
averaged over 2500 ms for the three networks, while the bottom panel shows
the autocorrelation function (AC) of the network firing rate for the three cases
(same color scheme). Note that for small inhibitory spatial modulation pI
1 (black
curve) the profile is essentially flat while the AC exhibits an initial peak and
dip, but an absence of multiple peaks which would indicate fast oscillations. As
pI
1 is increased (red curve), the profile remains flat but the AC clearly exhibits
22
-0.16-0.12-0.08η00.20.40.60.81D110001e+06-α Figure 12: A. An example of a function Φ(x) for which subcritical oscillations are possible.
The dotted curve indicates the range of the function Φover which oscillations are subcritical.
B. A bifurcation diagram for subcritical oscillations when the function Φ(x) is the same as in
panel A. Open circles: the limit cycle amplitude computed numerically as a function of J0.
Here D = 0.1 and the critical coupling is ¯J0 = −15.89. The fixed point is held at R = 0.1 and
thus the value of x in panel A is close to 0.1 (x + x3 = 0.1).
periodic peaks indicating that fast oscillations are present in the firing rate.
The corresponding raster plot in the middle panel of Fig. 13 shows intermit-
tent standing wave patterns which emerge and later disappear giving rise to a
new pattern with a different spatial orientation (not shown). This explains why
the time average of this spatial profile becomes flat. Finally, for strong enough
spatial modulation, a stationary standing wave pattern is seen in Fig. 13A (bot-
tom).
In this case the time-averaged spatial profile shown in Fig. 13B (top,
blue) shows two distinct maxima, whereas the AC indicates fast oscillations in
the firing rate.
Extensive simulations with such a cosine connectivity always yielded stand-
ing wave patterns for various choices of synpatic weights and input rates (not
shown). We seek to understand why this is so, and if delay-driven traveling
wave patterns can also be seen in network numerical simulations. To this end
we study the emergence of fast oscillations in the rate equation.
4.3.2. Rate equation
There is a spatially inhomogeneous oscillatory instability with frequency ω
given by Eq.(9a). This occurs for a value of the kth spatial Fourier mode of the
connectivity given by Eq.(9b), while we assume that all other Fourier modes are
sufficiently below their critical values to avoid additional instabilities. Without
loss of generality we assume k = 1.
We expand the parameters J1, I and r as in Eqs. (12,13,14) and define the
slow time Eq. (15). The linear solution consists of leftwards and rightwards
traveling waves with an amplitude which we allow to vary slowly in time, i.e.
r1 = A(T )eiωt+ix+B(T )e−iωt+ix+c.c.. Carrying out a weakly nonlinear analysis
to third order in leads to the coupled amplitude equations
∂T A = (µ + iΩ)∆J1A + (a + ib)A2A + (c + id)B2A,
(21a)
23
00.20.40.60.81x00.511.52Φ(x) = x+x3(A)-17-16.5-16-15.5-15-14.5-14J000.0020.0040.0060.008LC amplitude(B) 0 = pI
0 = 0.2, pE
1 = 0 and pE,I
Figure 13: (Color online) Standing waves in a spiking network with inverted Mexican-hat
connectivity with pE
2 = 0 (see Eq.(A.2) in the Appendix A)
and gE = 0.01, gI = 0.028. The rate of the external Poisson inputs is νext = 5000Hz and
gext = 0.001. A: Raster plots of spiking activity for three simulations with increasing spatial
modulation of the connection probability between neurons. From top to bottom: pI
1 = 0.15,
0.17 and 0.20. B: Top: The profile of activity in the three simulations averaged over 2500ms.
The color code is black pI
1 = 0.15, red 0.17 and blue 0.2. Bottom: The autocorrelation function
of the firing rate averaged over all neurons in each simulation. Note that the instantaneous
firing rate itself is not shown here.
24
0500100015002000neuron index010203040506070firing rate (Hz)020406080100τ-0.500.51AC(τ)B Figure 14: Top: The real part of the cubic coefficient a. Bottom: The real part of the cross-
coupling coefficient c. Solid lines are from the full expressions Eq.(B.6-B.7) and dotted lines
are the leading order terms in the small delay limit, Eqs. (22a-22b). The functions Φ and
J(x) as well as the input current I are taken as in Fig. (7).
∂T B = (µ − iΩ)∆J1B + (a − ib)B2B + (c − id)A2B,
(21b)
where the coefficients (a+ib), (c+id) and (µ+iΩ) are given by the Eqs. (B.6,B.7,B.4),
respectively.
In the small delay limit (D → 0) we can use the asymptotic values (11) to
obtain, to leading order,
a + ib =
(cid:33)
,
(cid:48)(cid:48)
J0(Φ
1 − Φ
(cid:48)
)2
J0
(cid:48)(cid:48)(cid:48)
Φ
2
+
(cid:32)
χ( π
(DΦ
2 + i)
)3
(cid:48)
(cid:32)
(cid:33)
,
(22a)
(22b)
c + id =
χ( π
(DΦ
2 + i)
)3
(cid:48)
(cid:48)(cid:48)
J0(Φ
1 − Φ
(cid:48)
)2
J0
(cid:48)(cid:48)(cid:48)
+
+ Φ
(cid:48)(cid:48)
J2(Φ
1 − Φ
(cid:48)
)2
J2
where χ ≡ π3/(8 + 2π2). Figure 14 shows a comparison of the full expressions
(solid lines) for the real parts of the cubic and cross-coupling coefficients a and
c with the asymptotic expressions above (dotted lines).
The equations (21a) and (21b) admit solutions of the form (A, B) = (AeiθA ,BeiθB ),
4.3.3. Wave solutions and their stability.
where the amplitudes A and B obey
A = µ∆J1A + aA3 + cB2A,
B = µ∆J1B + aB3 + cA2B.
(23a)
(23b)
Traveling waves: Leftward and rightward traveling waves in Eqs. (23a)
and (23b) are given by (AT W , 0) and (0,AT W ) respectively, where AT W =
−µ∆J1/a. The stability of traveling waves can be determined with the ansatz
25
1100100001e+06-a00.20.40.60.81D1100100001e+06c Figure 15: The existence and stability of traveling and standing waves as a function of the
cubic and cross-coupling coefficients a and c given by Eqs. (22a,22a) and, in the small delay
limit, by Eq. (24). In each sector of parameter space a representative bifurcation diagram is
shown. Supercritical (subcritical) solutions are shown growing from left to right (right to left).
Stable (unstable) solutions are given by solid (dashed) lines. Also indicated in each sector is
the type of solution which will be seen numerically. A question mark is placed wherever the
type of stable solution cannot be determined through a weakly nonlinear analysis.
(A,B) = (AT W , 0) + (δA, δB)eλt. The resulting eigenvalues are λ1 = −2µ∆J1
and λ2 = −µ∆J1(c/a − 1).
Standing waves: Standing waves in Eqs. (23a) and (23b) are given by (ASW ,ASW ),
where AT W = −µ∆J1/(a + c). The stability of standing waves can be de-
termined with the ansatz (A,B) = (ASW ,ASW ) + (δA, δB)eλt. The resulting
eigenvalues are λ1 = −2µ∆J1 and λ2 = −2µ∆J1(a − c)/(a + c).
The existence and stability of small-amplitude waves as described above is
completely determined by the values of the cubic and cross-coupling coefficients
a and c. This is illustrated in Fig. 15, where the parameter space is divided
into five sectors.
In each sector the type of solution which will be observed
numerically is indicated where known, and otherwise a question mark is placed.
Illustrative bifurcation diagrams are also given. Specifically, in the region la-
beled 1 (red online), the SW solution is supercritical and unstable while the
TW solution is supercritical and stable. TW will therefore be observed. In the
region labeled 2, the SW solution is supercritical and unstable while the TW
solution is subcritical. Finite-amplitude TW are therefore expected to occur
past the bifurcation point. In the region labeled 3, both solution branches are
subcritical, indicating that the analysis up to cubic order is not sufficient to
identify the type of wave which will be observed. In the region labeled 4, TW
are supercritical and unstable while SW are subcritical. Finite amplitude SW
are therefore expected past the bifurcation point. In the region labeled 5, the
TW solution is supercritical and unstable while the SW solution is supercritical
and stable. SW will therefore be observed.
26
-1-0.500.51c-1-0.500.51aTWTWSWSW?12345 A
C
time
time
B
D
time
time
Figure 16: Examples of wave solutions from numerical simulation of Eq.(2). The functions Φ
and J(x) as well as the input current I and the delay D are taken as in Fig. 7, with J1 = −120.
A. Supercritical standing waves: J0 = −40 and 5 units of time are shown. B. Supercritical
standing waves: J0 = −9 and 5 units of time are shown. C. Subcritical standing waves:
J0 = −5 and 40 units of time are shown. D. Subcritical traveling waves: J0 = 0 and 5 units
of time are shown.
Performing the small delay limit we find, from Eqs. (22a,22b), that
a = c − π
2
χ
(DΦ)3
(cid:48)(cid:48)(cid:48)
Φ
2
(cid:48)(cid:48)
J2(Φ
1 − Φ
(cid:48)
)2
J2
+
.
(24)
(cid:32)
(cid:33)
From Fig. 15 we can see that the nature of the solution seen will depend crucially
on the sign of the second term of the right-hand side of Eq.(24). In particular,
the diagonal a = c divides the the parameter space into two qualitatively differ-
ent regions. Above this line TWs are favored while below it SWs are favored.
In the small delay limit, Eq.(24) indicates that the balance between the third
derivative of the transfer function Φ
and the second spatial Fourier mode of
the connectivity kernel will determine whether TW or SW are favored.
(cid:48)(cid:48)(cid:48)
For sigmoidal transfer functions, the third derivative changes sign from posi-
tive to negative already below the inflection point, while for expansive power-law
nonlinearities, which fit cortical neuronal responses quite well in the fluctuation-
driven regime, the third derivative is positive if the power is greater than 2
and is negative otherwise. The contribution of this term therefore will depend
on the details of the neuronal response.
In simulations of large networks of
conductance-based neurons in the fluctuation-driven regime in which J2 was
> 0 [1, 2].
zero, the standing wave state was always observed, indicating a Φ
The phase diagram for J2 = 0, Fig. 7, clearly shows the dominance of the
> 0 for the parameter values chosen. Specifically,
SW solution, indicating Φ
(cid:48)(cid:48)(cid:48)
(cid:48)(cid:48)(cid:48)
27
for values of J0 < −6.3, supercritical standing waves are stable (see region 5 in
Fig. 15). Figures 16 A and 16 B show supercritical SW patterns for J0 = −40
and J0 = −9, respectively. For −6.3 < J0 < −2.6, TW are supercritical and
unstable while SW are subcritical [see region 4 in Fig. 15]. An example of
subcritical SW is shown in Fig. 16 C. For −2.6 < J0 < 3.58, both SW and
TW are subcritical (see region 3 in Fig. 15). Numerical simulations reveal TW
patterns in this region (see an example in Fig. 16 D). In the region where SW
are subcritical there is a small sliver in (J0, J1) where the SW state is bistable
with a TW state (TW/SW in the phase diagram). This TW branch most likely
arises in a secondary bifurcation slightly below the Turing-Hopf bifurcation line.
There is also a small region of bistability between large amplitude TW and the
spatially uniform high activity state (TW/HA in the phase diagram Fig. 7).
(cid:48)(cid:48)(cid:48)
Thus for Φ
> 0 and with a simple cosine connectivity, SW arise for most
values of J0. However, adding a non-zero J2 can lead to the TW solution
winning out. The phase diagram of wave states as a function of J0 and J2 is
shown in Fig. 17 A. In Fig. 17 A, the light shaded region indicates values of
J0 and J2 for which TW are expected, whereas SW are expected in the dark
shaded region. In the unshaded region, both TW and SW are subcritical and
the solution type is therefore not determined by the analysis up to cubic order.
These regions, delimited by the solid lines, were determined by numerically
evaluating the real parts of the full expressions for the cubic and cross-coupling
coefficients, Eqs. (B.6-B.7). Each region is furthermore numbered according to
the existence and stability of the TW and SW solution branches as shown in
Fig. 15. The dashed lines show the approximation to the solid lines given by
the asymptotic formulas (22a-22b). The set of allowable values for J0 and J2 is
bounded by the conditions (8-9b) corresponding to steady or oscillatory linear
instabilities. These stability conditions are shown by the horizontal and vertical
bounding lines (red online). All parameter values are as in Fig. 7.
From Fig. 17 we can now understand the discrepancy between the analytical
results in [1] using a rate equation with a linear threshold transfer function,
which predicted TW, and network simulations, which showed SW. Specifically,
given a nonlinear transfer function with Φ
> 0, then with a simple cosine
coupling SW are predicted over almost the entire range of allowable J0 (dark
shaded region for J2 = 0). The nonlinear transformation of inputs into outputs
is thus crucial in determining the type of wave solution. The choice of a threshold
linear transfer function results in the second and all higher order derivatives
being zero. In this sense it produces degenerate behavior at a bifurcation point,
and by continuation of the solution branches, in a finite region of the phase
diagram.
(cid:48)(cid:48)(cid:48)
4.3.4. 'Difference-of-Gaussian' connectivities
We have shown that varying J0 can change the nature of the bifurcation, e.g.
supercritical to subcritical, while varying J2 can switch the solution type, e.g.
from SW to TW. As an example of a functional form of connectivity motivated
by anatomical findings, e.g. [42], we consider a difference of Gaussians, written
28
Figure 17: (Color online) A. Phase diagram for waves as a function of the zeroth and second
spatial Fourier coefficients of the connectivity kernel. The dark-shaded region indicates SW,
whereas the light shaded region indicates TW. Red lines indicate boundaries for primary
instabilities of J0 and J2 given by Eqs. (8-9b). Solid stability lines for waves are from Eqs. (B.6-
B.7) while the dashed line are from the asymptotic expressions (22a-22b). Here ¯J1 = −58.4.
The function Φ as well as the input current I and the delay D are taken as in Fig. 7. Insets:
example connectivity patterns corresponding to the values of J0 and J2 marked by the square
and triangle respectively. B. The same phase diagram as in A, now showing where various
types of 'difference-of-Gaussian' connectivities, Eq. (25), would lie. Each dotted line indicates
the border of a region in which the standard deviations of the excitatory and inhibitory
connectivities are below a certain threshold (0.7, 1.0 and 1.5, respectively). Relatively narrow
connectivities compared to the system size will always generate TW solutions. See text for
details.
29
-60-40-200J2-60-40-200J0TW supercr (1)TW subcr (2)SW subcr (4)SW supercr (5)--++++(A)(3)-60-40-200J2-60-40-200J0σe,i<0.7σe,i<1σe,i<1.5(B) as
J(x) =
Je√
2πσe
− x2
2σ2
e
e − Ji√
2πσi
− x2
2σ2
i .
e
(25)
In this case, one finds that the Fourier coefficients are
Jk = Jee−k2σ2
e /2f (k, σe) − Jie−k2σ2
i /2f (k, σi).
(26)
√
where f (k, σe,i) = Re[Erf((π/σe,i + ik2)/
2)]/π. Once J1 has been fixed at
the critical value for the onset of waves, from Eq.(26) it is straightforward to
show that J0 = −pJ2 + q where both p and q are constants which depend
on σe and σi, the widths of the excitatory and inhibitory axonal projections
respectively. Thus a difference-of-Gaussian connectivity, constrains the possible
values of J0 and J2 to lie along a straight line for fixed connectivity widths.
This is illustrated in Fig. 17B where three dashed lines are superimposed on the
phase diagram, corresponding to the values σe,i = (1.5, 1.49); σe,i = (1, 0.99);
and σe,i = (0.7, 0.69). Each of these lines is bounding a region to the left
where σe and σi are less than 0.7, 1.0 and 1.5 respectively. Given periodic
boundary conditions with a system size of 2π, a Gaussian with σ = 1.5 is already
significantly larger than zero for x = π or −π. Thus, restricting ourselves to
Gaussians which essentially decay to zero at the boundaries means that TW
will always be observed. The same holds true for qualitatively similar types of
connectivity.
4.3.5. Classes of waves in Network Simulations
Our analytical results concerning waves from the rate equation Eq.(2) predict
that a connectivity with a sufficiently strong second Fourier component with
a negative amplitude will lead to traveling waves (see the phase diagram in
Fig. 17(A)).
Here we have confirmed this prediction performing numerical simulations of
the network of spiking neurons described in the Appendix A. Indeed, Fig. 18
shows that the addition of the second spatial Fourier component to the in-
hibitory connections converts standing waves (SW) into travelling waves (TW).
5. Bifurcations of codimension 2
For certain connectivity kernels we may be in the vicinity of two distinct in-
stabilities. This is the case for certain Mexican hat connectivities (OU and SB)
and certain inverted Mexican hat connectivities (OU and SW/TW). Although
two instabilities will co-occur only at a single point in the phase diagram Fig. 7,
i.e. J0 and J1 are both at their critical values, the competition between these
instabilities may lead to solutions which persist over a broad range of connectiv-
ities. This is the case here. We can investigate this competition once again using
a weakly nonlinear approach. The main results of this section are summarized
in table 1.
30
Figure 18: The transition from (A) standing to (B) travelling waves in a network of
conductance-based neurons takes place by increasing the second Fourier mode of the synaptic
connectivity pI
1,2 = 0,
gE = 0.01, gI = 0.028, gext = 0.001mS ms/cm2, and νext = 5000Hz.
2 = 0.1. Remaining parameters: pE,I
1 = 0.2, pE
2. (A) pI
2 = 0, (B) pI
0 = pI
5.1. Hopf and Turing-Hopf bifurcations
Here we consider the co-occurrence a spatially homogeneous oscillation and
a spatially inhomogeneous oscillation (OU and SW/TW), both with frequency
ω given by Eq.(9a). This instability occurs when the zeroth and kth spatial
Fourier mode of the connectivity both satisfy the relation, Eq.(9b), while we
assume that all other Fourier modes are sufficiently below their critical values
to avoid additional instabilities. Without loss of generality we take k = 1 for
the SW/TW state.
We expand the parameters J0, J1, I and r as in Eqs. (12,13,14), and define
the slow time (15). The linear solution consists of homogeneous, global oscilla-
tions, leftwards and rightwards traveling waves with amplitudes which we allow
to vary slowly in time, i.e. r1 = H(T )eiωt + A(T )eiωt+ix + B(T )e−iωt+ix + c.c..
Carrying out a weakly nonlinear analysis to third order in leads to the coupled
amplitude equations
∂T H = (µ + iΩ)∆J0H
+2(α + iβ)[(
+ A2 + B2)H + H∗AB∗],
H2
2
∂T A = (µ + iΩ)∆J1A + (a + ib)A2A + (c + id)B2A
+(α + iβ)[2H2A + H 2B],
∂T B = (µ − iΩ)∆J1B + (a − ib)B2B + (c − id)A2B
+(α − iβ)[2H2B + H∗2A],
(27a)
(27b)
(27c)
where α + iβ, a + ib and c + id are given by Eqs. (B.5,B.6,B.7), respectively.
The overbar in H∗ represents the complex conjugate.
31
9009501000time (ms)010002000neuron index(A)9009501000time (ms)010002000neuron index(B) 5.1.1. Solution types and their stability
Eqs. (27a-27c) admit several types of steady state solutions including os-
cillatory uniform solutions (OU), traveling waves (TW), standing waves (SW)
and mixed mode oscillations/standing waves (OU-SW). The stability of these
solutions depends on the values of the coefficients in Eqs. (27a-27c). In addi-
tion, non-stationary solutions are also possible. Here we describe briefly some
stationary solutions and their stability. For details see Appendix B.
Oscillatory Uniform (OU): The oscillatory uniform solution has the form (H, A, B) =
(Heiωt, 0, 0) where
(cid:114)−µ∆J0
(cid:17)
(cid:16)
α
Ω − βµ
α
,
∆J0.
H =
ω =
The OU state undergoes a steady instability along the line
∆J1 = ∆J0.
(28)
This stability line agrees very well with the results of numerical simulations of
Eq.(2) [see the phase diagram Fig. 7].
Traveling Waves (TW): The traveling wave solution has the form (H, A, B) =
(0,AT W eiωt, 0) or (0, 0,AT W e−iωt), where
(cid:114)−µ∆J1
(cid:17)
(cid:16)
a
Ω − bµ
a
,
∆J1.
AT W =
ω =
The TW state undergoes an oscillatory instability along the line
with a frequency
(cid:16)
¯ω =
∆J1 =
a
2α
∆J0,
Ω(1 − a
2α
) + (b − 2β)
µ
2α
(cid:17)
∆J0.
(29)
Standing Waves (SW): The standing wave solution has the form (H, A, B) =
(0,ASW eiωt,ASW e−iωt), where
ASW =
ω =
(cid:17)
µ
∆J1.
(30)
(31)
(cid:115)−µ∆J1
(cid:16)
(a + c)
Ω − (b + d)
(a + c)
,
32
An oscillatory instability occurs along the line
with a frequency
(cid:114)
¯ω =
[Ω(
∆J1 =
(a + c)
4α
∆J0,
a + c
4α
− 1) − µ
b + d − 4β
4α
]2 − µ2 α2 + β2
4α2 ∆J0.
A stationary instability occurs along the line
where
k1 =
(cid:104)
Ψ =
(b + d − 4β)
Ω − µ
∆J1 = Ψ∆J0,
−k2 +(cid:112)k2
2 − 4k1k3
(cid:105)2
2k1
+ µ2 (12α2 − 4β2)
(a + c)2
(b + d − 4β)
(a + c)
− 2Ω2 + 2Ωµ
,
,
,
(a + c)
k2 = −8µ2
α
(a + c)
(32)
(33)
k3 = µ2 + Ω2.
For Eq. (2) with the parameters used to generate the phase diagram Fig. 7,
we find that the stationary instability precedes the oscillatory one and that
Ψ ∼ 0.6. This agrees well with the numerically determined stability line near
the co-dimension 2 point in the diagram 7.
Mixed Mode: We can study the mixed mode solutions in Eqs. (27a-27c) by
assuming an ansatz
(H, A, B) = (Heiθ,AeiψA ,Be−iψB ),
(35)
which leads to four coupled equations [see (Appendix B)]. We do not study the
stability of mixed mode solutions in this work.
5.1.2. A simple example
We now turn to a simple example in order to illustrate the two main types
bifurcation scenarios that can arise when small amplitude waves and oscillations
interact in harmonic resonance.
i. Bistability: Here we take the parameters 1. Given these parameter values
one finds, from the analysis above, that the oscillatory uniform state has an
amplitude H = 1 and destabilizes along the line ∆J1 = −1. The standing wave
1µ = −1, ∆J0 = −1, α = −1, a = −1, b = c = d = β = ω = Ω = 0
33
solution (traveling waves are unstable [see Fig. 15] has an amplitude ASW =
√−∆J1 which undergoes a steady bifurcation to the oscillatory uniform state
at ∆J1 = −1/2. Both solutions are therefore stable in the region −1 < ∆J1 <
−1/2. This analysis is borne out by numerical simulation of Eqs. (27a-27c) [see
Fig. 19a]. Solid and dotted lines are the analytical expressions for the stable and
unstable solution branches respectively (red is OU and black is SW). Circles are
from numerical simulation of the amplitude equations (27a-27c).
Note that this scenario is the relevant one for the phase diagram shown in
Fig. 7. That is, we find there is a region of bistability between the OU and SW
solutions, bounded between two lines with slope ∼ 0.6 and 1 respectively.
an amplitude ASW = (cid:112)−∆J1/8 (traveling waves are again unstable) which
ii. Mixed Mode: Here we consider the parameters 2. Given these parameter
values one finds that the oscillatory uniform state has an amplitude H = 1
and destabilizes along the line ∆J1 = −1. The standing waves solution has
undergoes an oscillatory instability at ∆J1 = −2. The mixed-mode solution is
given by
H =
ASW =
4 + ∆J1(2 − cos φ − sin φ)
4 − (2 − cos φ − sin φ)2
∆J1 + (2 − cos φ − sin φ)
8 − 2(2 − cos φ − sin φ)2 ,
,
(36a)
(36b)
1 = ∆J1(1 − 4 cos φ − 2 sin φ + 2 sin φ cos φ) − 4 cos φ + 8 cos φ2. (36c)
It is easy to show that for ∆J1 → −1 the mixed mode amplitudes approach
(H,ASW ) = (1, 0) and the phase φ → 0. The mixed-mode solution thus bi-
furcates continuously from the oscillatory pure mode. Figure 19b shows the
corresponding bifurcation diagram where solid and dotted lines are the ana-
lytical expressions for the solution branches and symbols are from numerical
simulation of Eqs. (27a-27c). As ∆J1 increase from the left we see that the SW
solution indeed undergoes an oscillatory instability at ∆J1 = −2 leading to an
oscillatory mixed-mode solution indicated by small circles (the maximum and
minimum amplitude achieved on each cycle is shown). This oscillatory solution
disappears in a saddle-node bifurcation, giving rise to a steady mixed-mode
solution whose amplitude is given by Eq.(36a- 36c). This steady mixed-mode
solution bifurcates from the pure oscillatory mode at ∆J1 = −1 as predicted.
5.1.3. Summary
The interaction between the oscillatory uniform state and waves may lead to
mixed mode solutions or bistability. The OU state always destabilizes along the
line J1 = J0, irrespective of parameter values or the choice of Φ or J(x). This
result from the weakly nonlinear analysis, agrees with numerical simulations of
2 µ = −1, ∆J0 − 1, α = −1, a = −8, β = 1, b = c = d = ω = Ω = 0
34
Figure 19: Two typical bifurcation diagrams for the case of harmonic resonance between
small-amplitude oscillations and small-amplitude standing waves. A: Here oscillations and
standing waves are bistable for −1 < ∆J1 < −1/2. ∆J0 = −1, α = −1, a = −1, b = c =
d = β = Ω = 0, µ = −1. B: Here the standing wave solution loses stability to an oscillatory
mixed-mode solution at ∆J1 = −2. At ∆J1 ∼ −1.75 a steady mixed-mode solution arises in
a saddle-node bifurcation and continuously approaches the oscillatory pure-mode solution at
∆J1 = −1. Parameters are a = −8, β = 1, α = −1, µ = −1, b = c = d = Ω = 0. The phase φ
of the mixed-mode solution is not shown.
Eq.(2) over the entire range of values of J0 and J1 used in the phase diagram,
Fig. 7 and appears to be exact. Depending on the value of J2, supercritical TW
or supercritical SW will be stable near the codimension 2 point. In the case of
TW, the slope of the stability line is one half the ratio of the cubic coefficient
of waves to that of oscillations. In the small delay limit this expression can be
simplified to
a
2α
∼ π
4
(37)
(cid:16) (Φ
(cid:16) (11π−4)
Φ
20
(cid:48)(cid:48)
(cid:48)(cid:48)(cid:48)
(cid:48) − Φ
)2
2
)2
(cid:48)(cid:48)
(cid:48) − πΦ
(cid:48)(cid:48)(cid:48)
4
(Φ
Φ
(cid:17)
(cid:17) ,
which depends only on shape of the transfer function Φ. For the parameter
values used in the phase diagram Fig. 7 this yields a line with slope close to
one half. Thus TW and OU are expected to be bistable in the wedge between
∆J1 = ∆J0/2 and ∆J1 = ∆J0. In the case of SW, the slope of the stability line
is a complicated function of the shape of Φ and the second Fourier coefficient
J2. For the parameter values used in the phase diagram Fig. 7 the slope is close
to 0.6. Therefore the OU and SW states are bistable in the wedge between
∆J1 = 0.6∆J0 and ∆J1 = ∆J0.
5.1.4. Network simulations
Given that network simulations robustly reveal standing wave patterns, we
would expect to find either mixed-mode SW-OU or bistability between SW and
OU. As we have shown previously, e.g.
[1], there is a region of bistability be-
tween SW and OU in network simulations for strongly modulated inhibitory
connectivity. Here we show additional network simulations that suggest this
bistable region is in the vicinity of the codimension two point, i.e. it is a bista-
bility between the OU and SW states arising via primary bifurcations of the
unpatterned state.
35
-2-1.5-1-0.50∆J100.511.5amplitudeHASW(A)-3-2-10∆J100.20.40.60.81amplitudeHASWHMMAMMMM osc(B) Figure 20: Bistability between SW and OU states in an inhibitory network with pI (r) = 0.4 +
0.2 cos r, where r is the distance between neurons, νext = 4500Hz, and gI = 0.1mS·ms/cm2.
This is, the only network simulation for which an explicit delay has been added of δ = 0.5ms.
Removing the explicit delay for these parameter values eliminates the bistability. Parameter
values are identical for both simulations.
In the simulation shown in the bottom raster a
hyperpolarizing current of Iapp = −5.0µA/cm2 was injected into cells 1-1000 for 30ms at time
t = 400ms, switching the state from OU to SW.
Figure 20 shows two rasters from simulations of a purely inhibitory network
with strongly spatially modulated connectivity. The top raster shows 300 mil-
liseconds of activity in which homogeneous oscillations are clearly observable.
In the bottom raster, the network is started from the precisely the same initial
condition, but a hyperpolarizing input current is applied to neurons 1 to 1000
from t=400 to t=430ms. The network activity clearly switches to a SW state
in response to this input. The SW state persists for as long as simulations were
carried out (10sec). The network thus exhibits bistability between the OU and
SW states.
In order to determine if this region of bistability is related to the codimension
2 point, we adiabatically increased the recurrent excitatory connectivity in the
network, thereby mimicking an increase in J0 in the rate model. This is done
by generating a network of both inhibitory and excitatory neurons with pI (r) =
0.4 + 0.2 cos r and pE = 0.2. The weights of the inhibitory synapses are taken
as gI = 0.1 while excitatory weights are allowed to vary. Specifically, gE = 0
for the first 500ms of the simulation and are then slowly increased according
to gE(t) = 1.0526 · 10−5(t − 500). At the same time we slowly decrease the
external drive in order to maintain mean firing rates. Thus νext = 4500 for
the first 500ms and is then varied according to νext = 4500 − 0.158(t − 500)
thereafter. This particular functional form was determined empirically to keep
the mean firing rates steady. Therefore, for the first 500ms of simulation time
36
Figure 21: Increasing the strength of recurrent excitatory connections drives the system out
of the bistable regime and into a SW state. A: A raster plot of 3 seconds of simulation.
For the first 500ms the simulation is identical to that shown in the top panel of Fig.20.
Thereafter the weights of the recurrent excitatory synaptic are slowly increased while the
external drive is slowly decreased so as to maintain stable mean firing rates. There is a
spontaneous destabilization of the OU state to a SW state around t = 1500ms. B: The
network firing rate for the simulation shown in A (top), the time course of the excitatory
synaptic weights (middle) and external drive (bottom). C: For the first 500ms the simulation
is identical to that shown in the bottom panel of Fig.20, i.e. there is a hyperpolarizing current
injected at time t = 400ms which switches the state from OU to SW. Thereafter the weights of
recurrent excitatory synapses are slowly increased while the external drive is slowly decreased
so as to maintain stable mean firing rates. D: The network firing rate and time courses of the
excitatory synaptic weights and external drive, as in B.
37
the network is equivalent to that shown in Fig. 20 while at later times the
network is no longer purely inhibitory.
Carrying out such an adiabatic increase in the recurrent excitation should
cause the network to cross the line of instability of the OU state, leading to stable
SW. Thus, if we begin simulations in the OU state, they should destabilize at
some point to SW while if we begin in the SW state they should persist. This is
precisely what occurs. Fig.21A shows a raster plot of three seconds of simulation
time begining in the OU state. A transition to the SW occurs around 1500ms.
Fig.21B shows the network firing rate during these three seconds (top) as well
as the time course of the excitatory synaptic weights and external drive (middle
and bottom respectively). Note that the mean firing rate is relatively steady
in the OU state indicating that the co-variation of the synaptic weights and
external drive balance one another. Fig.21C and D analogously show the raster
and firing rate of a simulation in which the network is switched into a SW state
at t = 400ms. Note that the SW state persists over the whole 3 sec. simulation.
5.2. Hopf and Turing bifurcations
(cid:48)
We consider the co-occurrence of two instabilities: a spatially homogeneous
oscillation and a spatially inhomogeneous steady solution. This occurs when the
zeroth spatial Fourier mode of the connectivity satisfies the relation, Eq.(9b) and
the kth spatial Fourier mode satisfies J1 = 1/Φ
, while we assume that all other
Fourier modes are sufficiently below their critical values to avoid additional
instabilities. Without loss of generality we take k = 1 for the Turing instability.
We expand the parameters J0, J1, I and r as in Eqs. (12,13,14), and define
the slow time (15). The linear solution consists of homogeneous, global oscilla-
tions and stationary, spatially periodic bumps with amplitudes which we allow
to vary slowly in time, i.e. r1 = H(T )eiωt + A(T )eix + c.c.. Carrying out a
weakly nonlinear analysis to third order in leads to the coupled amplitude
equations
∂T H = (µ + iΩ)∆J0H + (α + iβ)H2H + (κ + iΛ)A2H,
∂T A = ¯η∆J1A + ΓA2A + σH2A,
(38a)
(38b)
where µ + iΩ, α + iβ, ¯η, Γ, κ + iΛ and σ are given by Eqs. (B.4,B.5,17a,17b,
B.15,B.16), respectively.
5.2.1. Solution types and their stability
Steady state solutions to Eqs. (38a,38b) include pure mode OU, pure mode
SB and mixed mode solutions (OU-SB). We look at the stability of the OU and
SB solutions in turn for the general case and then look specifically at the case
of small delay in Eq.(2). Since the coupling in Eqs. (38a-38b) is only through
the amplitudes we can simplify the equations by taking (H, A) = (Heiθ,Aeiφ)
which yields
H = µ∆J0H + αH3 + κA2H,
(39a)
38
(39b)
Oscillatory Uniform (OU): The uniform oscillations have the form (H∗, 0) where
A = η∆J1A + ΓA3 + σH2A.
(cid:114)
H =
− µ∆J0
α
.
The linear stability of this solution can be calculated with the ansatz
(H,A) = (H∗ + δHeλt, δAeλt),
which yields the two eigenvalues
λH = −2µ∆J0,
µσ
α
λA = η∆J1 +
∆J0.
If we assume a supercritical uniform oscillatory state then the first eigenvalue
is always negative, while the second becomes positive along the line
∆J1 =
µσ
ηα
∆J0,
(41)
indicating the growth of a bump solution.
For Eq.(2) with the parameters used to generate the phase diagram Fig. 7,
we find from Eq.(41) that the OU state destabilizes along the line ∆J1 ∼
−0.026∆J0.
Stationary Bump (SB): The stationary bump solution has the form (0,A∗)
where
(cid:114)
A =
− η∆J1
Γ
.
The linear stability of this solution can be calculated with the ansatz
(H,A) = (δHeλt,A∗ + δAeλt),
which yields the two eigenvalues
λH = µ∆J0 − ηκ
Γ
λA = −2η∆J1.
∆J1,
If we assume a supercritical stationary bump state then the second eigenvalue
is always negative, while the first becomes positive along the line
∆J1 =
µΓ
ηκ
∆J0,
(43)
indicating the growth of uniform oscillations.
39
Figure 22: Typical bifurcation diagrams for the competition between bumps and global oscil-
lations. µ = −1, ∆J0 = −1, α = −1, η = 1, Γ = −1. A: κ = −2, σ = −2. The limit cycle
and bump solutions are bistable in the range 1/2 < ∆J1 < 2. B: A mixed-mode solution is
stable in the range 1/2 < ∆J1 < 2. κ = −0.5, σ = −0.5. Symbols are from simulation of the
amplitude equations Eqs. (39a-39b) while lines are the analytical expressions.
For Eq.(2) with the parameters used to generate the phase diagram 7, we find
from Eq.(43) that the SB state destabilizes along the line ∆J1 ∼ −0.144∆J0.
Mixed Mode (OU-SB): The mixed-mode solution satisfies the following ma-
trix equation
which yields
(cid:18) α κ
(cid:19)
(cid:19)(cid:18) H2
A2
(cid:18) µ∆J0
η∆J1
(cid:19)
,
= −
σ Γ
H2 =
A2 =
−µΓ∆J0 + ηκ∆J1
αΓ − σκ
µσ∆J0 − ηα∆J1
αΓ − σκ
,
.
We do not study the stability of the mixed-mode solution here.
5.2.2. A simple example
We once again illustrate the scenarios of bistability and mixed-mode solu-
tions with a simple example.
√
i. Bistability 3: In this case, the limit cycle has an amplitude H = 1 and
undergoes an instability at ∆J1 = 2. The bump solution has an amplitude
A =
∆J1 and becomes unstable at ∆J1 = 1/2. The oscillatory and bump
solutions are therefore bistable in the range 1/2 < ∆J1 < 2. This is borne out
in numerical simulations of Eqs. (39a-39b) [see Fig. 22A]. Symbols are from nu-
merical simulation (circles:limit cycle, squares:bump), while lines are analytical
solutions.
3 µ = α = Γ = ∆J0 = −1, σ = κ = −2
40
00.511.522.53∆Jk00.511.52amplitude(A)AH00.511.522.53∆Jk00.511.52amplitude(B)AHHMMAMM Codim.-2 bifurcations
Hopf and Turing-Hopf Oscillatory Uniform
Solution types calculated
Hopf and Turing
Travelling Waves
Standing Waves (osc.)
Standing Waves (stat.)
Mixed-Mode
Oscillatory Uniform
Stationary Bump
Mixed-Mode (OU-SB)
Instability boundaries
∆J1 = ∆J0
∆J1 = a/(2α)∆J0
∆J1 = (a + c)/(4α)∆J0
∆J1 = Ψ∆J0
Not calculated
∆J1 = (µσ)/(ηα)∆J0
∆J1 = (µΓ)/(ηκ)∆J0
Not calculated
Table 1: Some existing dynamical states that are present close to the codimension-2 bifurca-
tions, and their corresponding instability boundaries Eqs. (28,29,32,33,41,43), except for the
Mixed Mode solutions.
ii.Mixed-mode 4: In this case, the limit cycle has an amplitude H = 1 and
√
undergoes an instability at ∆J1 = 1/2. The bump solution has an amplitude
A =
∆J1 and becomes unstable at ∆J1 = 2. The mixed-mode solution is sta-
ble in the range 1/2 < ∆J1 < 2 and has amplitudes HMM = 2(cid:112)(1 − ∆J1/2)/3
and A = 2(cid:112)(∆J1 − 1/2)3. The corresponding bifurcation diagram is shown in
Fig. 22B where symbols are from simulation of Eqs. (39a-39b) and lines are the
analytical results.
5.2.3. Summary
The interaction between the SB and OU states can lead to one of two sce-
narios. Either there is a region of bistability between bumps and oscillations,
or there is a mixed-mode solution which, near the codimension-2 point at least,
will consist of bumps whose amplitude oscillates in time, i.e. oscillating bumps
(OB).
In the limit of small D the instability lines for the OU and SB states in the
vicinity of the codimension 2 point are given by the equations
(cid:48)(cid:48)(cid:48)(cid:17)
Φ
(cid:16) (Φ
(cid:16) (11π−4)
(cid:16) (Φ
(cid:16) (Φ
20
(cid:48)(cid:48)
(cid:48) − Φ
(Φ
Φ
)2
Φ
D
Φ
(cid:48)(cid:48)
(cid:48) − Φ
)2
)2
(cid:17) ∆J0,
(cid:17)
(cid:48)(cid:48)
(cid:48) − π Φ
(cid:48)(cid:48)(cid:48)
4
(cid:48)(cid:48)(cid:48)(cid:17)
2 − J2(Φ
(cid:48) − Φ
(cid:48)(cid:48)
)2
(cid:48)
2(1−J2Φ
(cid:48)(cid:48)(cid:48)
)2
(cid:48)(cid:48)
)
∆J1 = −D
∆J1 = − 2
π
(45)
∆J0.
(46)
respectively. The slope of both of the stability lines is proportional to D, indi-
cating that in the small D limit any region of bistability or mixed mode solution
will be limited to a narrow wedge close to the J0 axis. Which scenario will be
(cid:48)
observed (bistability or mixed-mode) depends on the particular choice of Φ
and the value of the second spatial Fourier mode of the connectivity J2. For
4µ = α = Γ = ∆J0 = −1, σ = κ = −1/2
41
Figure 23: A mixed mode (OU/SB) solution in a network of spiking neurons with pE =
0.2 + 0.1 cos r, pI = 0.2, gE = 0.1, gI = 0.28, νext = 2000 and gext = 0.01. See text for
details.
the parameters values used to generate the phase diagram Fig. 7 the slopes
are ∼ −0.026 and ∼ −0.144 for the OU and SB stability lines respectively,
indicating a mixed mode solution.
5.2.4. Network simulations
As shown in [1], oscillating bump solutions can be found in networks of
spiking neurons with Mexican-hat connectivity and strong inhibition. Here we
have identified such oscillating bumps as arising via a bifurcation to mixed-
mode OU and SB in the vicinity of the codimension-two point for homogeneous
oscillations and Turing patterns. Fig.23 shows a sample raster from a numerical
simulatiom simulation. The top panel shows a mixed-mode solution which drifts
in time. The bottom panel is a blow-up of the raster for 500ms< t < 700ms
where the fast oscillations are clearly visible.
6. Conclusions
Our main objective in this paper was to understand in greater detail the
dynamical states which arise in networks of spiking neurons. Specifically, we
are interested in large networks of irregularly spiking neurons for which a re-
duced, phenomenological description in terms of mean firing rates is a reasonable
approximation. This is the case, for example, if the connectivity is sparse and
cross-correlations of the input currents to different cells are therefore weak. The
42
particular form of the rate equation description is motivated by the observation
of emergent fast oscillations in simulations of networks with dominant inhibi-
tion, e.g.
[3]. The origin of these oscillations has been well studied in model
networks and is due to an effective delay in neuronal interactions generated by
the synaptic kinetics and single-cell dynamics [34, 35]. The dynamics of the
mean firing rate of the resulting oscillatory states can be captured by a rate
equation with an explicit time delay [1].
However, the rate equation considered previously did not agree entirely with
network simulations. Specifically, it predicted a large region of traveling wave
solutions while in the network simulations only standing waves were found to be
stable. Our analysis here has shown that given a more realistic transfer function,
i.e. sigmoidal or expansive power-law nonlinearity, the primary inhomogeneous
oscillatory instability given a cosine connectivity will be to standing waves. This
suggests that the standing wave states robustly observed in network simulations
were due to the nonlinearity in the single cell fI curve, and that the rate model
with threshold linear transfer function studied in [1] was not able to capture
this effect.
We predicted further that altering the connectivity could stabilize the trav-
eling wave state. Specifically, more realistic patterns of connectivity, such as
Gaussian, affect the competition between traveling and standing waves through
the second spatial Fourier mode. In the case of delay-driven waves, where the
primary instability occurs only for strong inhibition, we show that the sign of the
second mode will be negative and that in the limit of small delay this will always
lead to traveling waves. This prediction is borne out in network simulations,
see Fig.18.
We furthermore show that the bifuration to homogeneous oscillations is su-
percritical for standard choices of transfer functions, again sigmoidal and ex-
pansive power-law. This agrees with network simulations of conductance-based
neurons we have conducted, and with the amplitude equation derived for a
network of integrate-and-fire neurons [3].
It is more difficult to draw clear-cut conclusions regarding the dynamical
states seen in the vicinity of the two codimension-two points we have studied:
Hopf/Turing-Hopf and Hopf/Turing. In general there will either be a region of
bistability between the two states which bifurcate via primary instabilities in the
vicinity of the codimension-two point, or there will be one or several mixed-mode
solutions.
In network simulations we have shown examples of both: bistable
OU/SW and mixed mode OU/SB. Many more complex dynamical states can
be observed both in the rate model and the network simulations further from
the primary bifurcations. Here we have chosen an analytical approach which is
powerful enough to allow for arbitrary transfer functions and connectivity, but
which is limited to the parameter space in the vicinity of the primary bifurca-
tions. This is a complementary approach to that taken in [1], where a specific
choice of transfer function and connectivity allowed for an analysis of several
nonlinear states, even far from the primary bifurcations.
The approach we have taken in this paper is similar in spirit to that of Curtu
and Ermentrout [33]. In that work, they study an extension of a rate model
43
with adaptation proposed by Hansel and Sompolinsky [32]. As in our work here,
they allow for a generic transfer function and generic connectivity and derive
amplitude equations for the primary Turing and Turing-Hopf bifurcations which
occur. Thus both adaptation and the effective delay in neuronal interactions
can lead to waves. However, the waves in these two cases arise via distinct
physiological mechanisms and can exhibit very different propagation velocities.
In the case of adaptation, waves arise given patterns of synaptic connectivity
which, in the absence of adaptation would lead to the emergence of bump states,
i.e. a Turing instability. These tend to be Mexican-hat like connectivities.
In the presence of adaptation, stationary bump states may destabilize since
the peak of activity is preferentially suppressed compared to activity at the
edge of the bump. As a result the bump begins to move via a symmetry-
breaking bifurcation, generally leading to traveling waves. This only occurs if
the adaptation is significantly strong, i.e. above a critical threshold. In a two-
population model with adaptation studied in [33], the frequency of the waves
was shown to be proportional to the square root of the difference between the
strength of adaptation and a critical value below which no bifurcation is possible.
Thus the resulting waves can be arbitrarily slow depending on the strength of
adaptation.
In the case of rate equations with delay, a Turing-Hopf instability occurs only
for connectivities which are strongly inhibitory locally, i.e. inverted Mexican-hat
connectivities. A transient perturbation which increases the firing rate locally
will, after a delay, strongly self-inhibit while increasing the firing rate of more
distant neurons. The process then repeats, leading to a propagation of activity.
The frequency of such waves is clearly related to the delay, which itself depends
on the synaptic time constants as well as the spike generation currents [34, 35].
Finally, we have tried to emphasize the importance of fixed delays in shap-
ing the dynamics described by Eq.(2) and by extension in networks of spiking
neurons. Nonetheless both fixed and conduction delays are present in neuronal
systems and are roughly of the same order of magnitude in a small patch of
cortex of ∼ 1mm in extent. It remains to be studied how these delays interact
to shape patterns of spontaneous activity.
Appendix A. The network of spiking neurons
In this appendix we describe the network of conductance-based neurons used
in simulations. The single cell model is taken from [43]. The network consists of
two populations of neurons: one excitatory and one inhibitory. The number of
neurons in each population is NE and NI respectively. The membrane voltage of
the ith neuron in the excitatory population evolves according to the differential
equation
Cm Vi = −IN a,i − IK,i − IL,i − Isyn,i + Iapp,i,
where the membrane capacitance Cm = 1µF/cm2 and the applied current Iapp,i
has the units µA/cm2.
44
(cid:16)
(cid:16)
(cid:17)
,
(cid:17)
The leak current is
IL,i = gL(Vi − EL),
where gL = 0.1mS/cm2. Action potential generation is dependent on a sodium
and a potassium current.
The sodium current is
IN a,i = gN am3∞,ihi(Vi − EN a),
where gN a = 35mS/cm2, EN a = 55mV and the activation variable m is as-
sumed fast and therefore taken at its equilibrium value m∞,i = αm,i/(αm,i +
βm,i), where αm,i = −0.1(Vi+35)/(exp(−0.1(Vi+35))−1) and βm,i = 4exp(−(Vi+
60)/18). The inactivation variable hi follows the first order kinetics
hi = φ
αh,i(1 − hi) − βh,ihi
where φ = 5, αh,i = 0.07exp(−(Vi + 58)/20) and βh,i = (exp(−0.1(Vi + 28)) +
1)−1.
The potassium current is
IK,i = gKn4
i (Vi − Ek),
where gK = 9mS/cm2 and EK = −90mV . The activation variable ni follows
the first order kinetics
ni = φ
αn,i(1 − ni) − βn,ini
where φ = 5, αn,i = −0.01(Vi + 34)/(exp(−0.1(Vi + 34)) − 1) and βn,i =
0.125exp(−(Vi + 44)/80).
The synaptic current is
Isyn,i = gEEsE,i(t)(Vi−Esyn,E)+gEI sI,i(t)(Vi−Esyn,I )+gextsext,i(t)(Vi−Esyn,E),
where the reversal potentials for excitatory and inhibitory synapses are Esyn,E =
0 and Esyn,I = −80 respectively. The conductance change from the activation
of recurrent excitatory connections is given by gEEsE,i(t) where
τE,2 sE,i = −sE,i + xE,i,
τE,1 xE,i = −xE,i +
NE(cid:88)
wij
(cid:88)
δ(t − tk
j ),
(A.1)
j=1
k
where the wijs ∈ {0, 1} indicate the presence or absence of a synaptic contact
from cell j to cell i. The double sum in Eq.(A.1) is over all neurons in the
excitatory population and over all spikes, i.e. tk
j is the time of the kth spike
of neuron j which will cause a jump of amplitude 1/τE,1 in the variable xi of
neuron i if wij = 1. The resulting post-synaptic conductance change in cell i
from a single presynaptic spike at time t∗ is given by
(cid:16)
gEEsE,i(t) =
gEE
τE,2 − τE,1
(cid:17)
,
e−(t−t∗)/τE,2 − e−(t−t∗)/τE,1
45
which has units of mS/cm2. The time course is therefore a difference of expo-
nentials with a rise time given by τE,1 and a decay time τE,2. Note that the
time integral of the response sE,i from t = t∗ to t = ∞ has been normalized to
1 and so gEE has units of mS · ms/cm2. The synaptic current from inhibitory
connections is analogous, with time constants τI,1 and τI,2.
Finally, external inputs have the same functional form as the recurrent
excitatory inputs. External presynaptic spikes to excitatory cells are mod-
eled as a Poisson process with rate νE,ext. The Poisson process is indepen-
dent from cell to cell. Unless otherwise noted, all external inputs have weight
gext = 0.0019mS·ms/cm2, and synpatic time constants are taken to be τ1 = 1ms
and τ2 = 3ms.
Inhibitory neurons are modeled analogously to excitatory ones. For this
work, we take all single cell parameters to be identical to the excitatory cells.
Synaptic time constants are taken to be the same, i.e. τE,1 = τI,1 = τ1 and
τE,2 = τI,2 = τ2. In addition we take gEE = gIE = gE and gII = gEI = gI .
Thus excitatory and inhibitory synapses have identical time courses but may
have different strengths.
Appendix A.1. Connectivity
We choose a prescription for choosing wijs which leads to sparse, random
connectivity which is spatially modulated. We do this by defining a probability
for a connection to be made from cell j in population β ∈ {E, I} to a cell i in
population α ∈ {E, I} of the form
pαβ(i, j) = pαβ
0 + pαβ
1 cos r + pαβ
2 cos 2r,
(A.2)
where r is the distance between cells i and j which are situated on a ring ,
normalized such that r ∈ {−π, π}. In order to compare with the rate model we
choose pEE
l where l ∈ {0, 1, 2}.
l = pIE = pE and pII
l = pEI
l = pI
Appendix B. Amplitude Equations
In this Appendix we outline the calculation of the amplitude equations which
describe the slow temporal evolution of the various instabilities near their re-
spective bifurcations.
Appendix B.1. General framework for the weakly nonlinear calculation: Codi-
mension 1 bifurcations
Here we briefly describe the general framework for the weakly nonlinear
calculation for the Turing, Hopf and Turing-Hopf bifurcations. We use the
standard multiple-scales approach which takes advantage of the fact there is a
near-zero eigenvalue in the vicinity of a bifurcation which is responsible for the
slow temporal evolution of the critical eigenmode (see e.g. [44]).
For simplicity we first rewrite Eq.(2) as
,
(B.1)
(cid:16)(cid:104)Jr(cid:105) + I
(cid:17)
r = −r + Φ
46
(cid:82) π
−π dyf (y− x)g(y, t− D). We study the stability of the steady
n=1 Jn cos nx. We
(cid:17)
, where J(x) = J0 + 2(cid:80)∞
(cid:16)
where (cid:104)f g(cid:105) ≡ 1
state solution R = Φ
expand the rates, the connectivity and the input current as
J0R + I
2π
r(x, t) = R + r1(x, t, T ) + r2(x, t, T ) + . . . ,
J(x) = ¯J(x) + 2∆J(x),
I = ¯I + 2∆I,
where the small parameter is defined by the distance from the critical value of
the connectivity, given by Eqs. (9a,9b). Plugging these expansions into Eq.(B.1)
yields
(L + 2L2)(r1 + 2r2 + ...) = 2N2(r1) + 3N3(r1, r2),
where
Lr = ∂tr + r − (cid:104) ¯Jr(cid:105),
L2r = ∂T(cid:104) ¯Jr(cid:105) − (cid:104)∆Jr(cid:105),
N2 =
N3 = Φ
(cid:48)(cid:48)
(cid:104) ¯Jr1(cid:105)2,
Φ
2
(cid:48)(cid:48)(cid:104) ¯Jr1(cid:105)(cid:104) ¯Jr2(cid:105) +
(cid:48)(cid:48)(cid:48)
Φ
6
(cid:104) ¯Jr1(cid:105)3.
We now collect terms by order in . At first order we have
ϑ() : Lr1 = 0.
This equation gives the linear dispersion relation Eq.(7). The values of the
connectivity and input current for which it is satisfied are J(x) = ¯J(x) and
I = ¯I. At second order we obtain
ϑ(2) : Lr2 = N2(r1).
The second order solution r2 is the particular solution of this linear differential
equation. And finally, at third order
ϑ(3) : Lr3 = N3(r1, r2) − L2r1.
At this order secular terms arise which have the same temporal and/or spatial
frequency as the linear solution. In order for the above equation to have a solu-
tion, these terms must therefore be eliminated, yielding the desired amplitude
equation for the instability.
Appendix B.1.1. Steady Bifurcation: ω = 0, k = 0
For completeness we include here the derivation of the amplitude equation
for the transcritical bifurcation.
47
Figure B.24: (Color online) Bifurcation diagram for the steady instability. Open circles:
numerical simulation of Eq.(2). Red Lines: amplitude equation solution from Eq.(B.3). Black
lines: steady-state solution of Eq.(2) using a Newton-Raphson solver. Solid lines indicate
stable solutions and dotted lines unstable ones. Φ(x) =
1+e−βx where α = 1.5 and β = 3.
J(x) = J0 + J1 cos x where J1 = 0. The input current I is varied so as to keep the uniform
stationary solution fixed at R = 0.1.
α
The 0th spatial Fourier mode of the connectivity is given by the critical value
(cid:48) , while we assume that all other Fourier modes are sufficiently below
¯J0 = 1
Φ
their critical values to avoid additional instabilities. We expand
J0 = ¯J0 + ∆J0,
I = ¯I + ∆I,
r = R + r1 + 2r2 + . . . ,
(B.2)
where the small parameter is defined by Eq.(B.2). We define the slow time
T = t. The linear solution is a spatially homogeneous amplitude which we allow
to vary slowly in time, i.e. r1 = A(T ). Carrying out a weakly nonlinear analysis
to second order in leads to the normal form for a transcritical bifurcation given
by
∂T A = η∆J0A + γA2,
(cid:48)
Φ
η =
γ =
¯J 2
0 .
(B.3)
,
1 + D
(cid:48)(cid:48)
Φ
2(1 + D)
48
012345J001RNumericsNewton-RaphsonAmplitude eqs. Appendix B.1.2. Turing bifurcation
ϑ(): The solution to the linear equation is spatially periodic with slowly varying
amplitude A,
r1 = A(T )eix + c.c.
ϑ(2): The nonlinear forcing and resulting second order solution are
(cid:48)(cid:48)
1 (A2e2ix + c.c. + 2A2)/2,
J 2
N2 = Φ
r2 = r22e2ix + c.c. + r20,
r22 =
r20 =
(cid:48)(cid:48)
A2,
(cid:48)
Φ
J 2
1
2(1 − J2Φ
J 2
Φ
1
1 − J0Φ
(cid:48)(cid:48)
)
(cid:48) A2.
ϑ(3): The nonlinear forcing at cubic order is
J1J2A∗r22 + Φ
(cid:48)(cid:48)
N3 = (Φ
J1J0Ar20 + Φ
(cid:48)(cid:48)
(cid:48)(cid:48)(cid:48)
1A2A/2)eix + c.c. + . . .
J 3
Eliminating all terms of periodicity eix at this order yields the amplitude equa-
tion, Eq.(16).
∂T A = η∆J1A + ΓA2A,
with the coefficients
η =
Γ =
,
(cid:32)
(cid:48)
Φ
1 + D
¯J 3
k
1 + D
(cid:48)(cid:48)
)2
J0(Φ
1 − J0Φ
(cid:48) +
J2(Φ(cid:48)(cid:48))2
2(1 − J2Φ
(cid:48)
)
(cid:48)(cid:48)(cid:48)
Φ
2
+
(cid:33)
.
Appendix B.1.3. Hopf bifurcation
ϑ(): The solution to the linear equation is a time periodic function with slowly
varying amplitude H
r1 = H(T )eiωt + c.c.
ϑ(2): The nonlinear forcing and resulting second order solution are
(cid:48)(cid:48)
N2 = Φ
r2 = r22e2iωt + c.c. + r20,
0 (H 2e2iω(t−D) + c.c. + 2H2)/2,
J 2
e−2iωDH 2,
J0e−2iωD)
r22 =
r20 =
(cid:48)(cid:48)
Φ
J 2
0
2(2iω + 1 − Φ
(cid:48)
J 2
(cid:48) H2.
Φ
0
1 − J0Φ
(cid:48)(cid:48)
ϑ(3): The nonlinear forcing at cubic order is
49
N3 = (Φ
(cid:48)(cid:48)(cid:48)
+ Φ
(cid:48)(cid:48)
(cid:48)(cid:48)
0 H∗r22
J 2
J 2
0 Hr20 + Φ
0A2A/2)eiω(t−D) + c.c. + . . .
J 3
Eliminating all terms of periodicity eiωt at this order yields the amplitude equa-
tion, Eq.(18).
∂T H = (µ + iΩ)∆J0H + (α + iβ)H2H,
with the coefficients
µ + iΩ =
α + iβ =
(cid:48)
e−iωD
Φ
1 + D(1 + iω)
e−iωD
1 + D(1 + iω)
(cid:32) ¯J 4
,
×
(cid:48)(cid:48)
)2
0 (Φ
1 − Φ
(cid:48) ¯J0
(cid:48)(cid:48)
¯J 4
0 (Φ
)2e−2iωD
+
2(2iω + 1 − Φ
(cid:48) ¯J0e−2iωD)
(B.4)
(B.5)
(cid:33)
.
(cid:48)(cid:48)(cid:48)
¯J 3
0 Φ
2
+
Appendix B.1.4. Turing-Hopf bifurcation
ϑ(): The solution to the linear equation are two sets of periodic waves, one
left-traveling with slowly varying amplitude A and the other right-traveling
with slowly varying amplitude B
r1 = A(T )eiωt+ix + B(T )e−iωt+ix + c.c.
= A(T )eψ + B(T )eφ + c.c.
ϑ(2): The nonlinear forcing and resulting second order solution are
r2ψ =
rψφ =
rψφ∗ =
r2φ =
r20 =
(cid:48)(cid:48)
Φ
J 2
1
2(2iω + 1 − Φ
(cid:48)
J 2
Φ
1
1 − Φ
(cid:48)
J2
AB,
(cid:48)(cid:48)
(cid:48)(cid:48)
Φ
J 2
1
2iω + 1 − J0Φ
J 2
Φ
1
(cid:48)(cid:48)
(cid:48)
(cid:48)(cid:48)
2(−2iωD + 1 − J2Φ
1 (A2 + B2)
J 2
Φ
1 − J0Φ
.
(cid:48)
e−2iωDA2,
J2e−2iωD)
e−2iωD e−2iωDAB∗,
e2iωDB2,
(cid:48)
e2iωD)
50
(cid:104)
N2 = Φ
(cid:48)(cid:48)
J 2
1
A2e2ψ−2iωD + 2ABeψ+φ + 2AB∗eψ+φ∗−2iωD +
B2e2φ+2iωD + c.c. + 2(A2 + B2)
/2,
r2 = r2ψe2ψ + rψφeψ+φ + rψφ∗ eψ+φ∗
+ r2φe2φ + c.c. + r20,
(cid:105)
ϑ(3): The nonlinear forcing at cubic order is
(cid:48)(cid:48)
N3 = [Φ
(cid:48)(cid:48)(cid:48)
Φ
(cid:48)(cid:48)
[Φ
(cid:48)(cid:48)(cid:48)
Φ
J1(J0Ar20 + J2A∗r2ψ + J0Brψφ∗ + J2B∗rψφ) +
(A2A/2 + B2A)]eψ−iωD +
J1(J0Ar∗
(B2B/2 + A2B)]eφ+iωD + . . .
ψφ∗ + J2A∗rψφ + J0Br20 + J2B∗r2φ) +
Eliminating all terms with dependencies eiψ and eiφ yields the two coupled
amplitude equations (21a,21b):
∂T A = (µ + iΩ)∆J1A + (a + ib)A2A + (c + id)B2A,
∂T B = (µ − iΩ)∆J1B + (a − ib)B2B + (c − id)A2B,
with the coefficients
a + ib =
c + id =
×
×
(cid:48)
(cid:32)
k e−iωD
¯J 3
(cid:48) ¯Jke−iωD
1 + DΦ
(cid:48)(cid:48)
)2
J0(Φ
1 − Φ
+
J0
k e−iωD
¯J 3
(cid:48) ¯Jke−iωD
1 + DΦ
(cid:48)(cid:48)
)2
J0(Φ
1 − Φ
J0
(cid:32)
+
(cid:48)
(cid:48)(cid:48)
)2e−2iωD
J2(Φ
2(2iω + 1 − Φ
(cid:48)
J2e−2iωD)
(cid:48)(cid:48)
)2e−2iωD
J0(Φ
2iω + 1 − Φ
(cid:48)
J0e−2iωD +
(cid:33)
,
(cid:48)(cid:48)(cid:48)
Φ
2
+
(cid:48)(cid:48)
J2(Φ
1 − Φ
(cid:48)
)2
J2
(cid:48)(cid:48)(cid:48)
+ Φ
(B.6)
(cid:33)
.(B.7)
Appendix B.2. Codimension 2 bifurcations
Appendix B.2.1. Double zero eigenvalue: Hopf, Turing-Hopf
ϑ(): The solution to the linear equation are periodic oscillations and traveling
waves
r1 = H(T )eiωt + A(T )eψ + B(T )eφ + c.c.
where we have defined ψ ≡ i(kx + ωt) and φ ≡ i(kx − ωt).
ϑ(2): The nonlinear forcing and resulting second order solution are
51
N2 = Φ
0 (H 2e2iω(t−D) + c.c. + 2H2)/2 + Φ
J 2
(cid:48)(cid:48)
+2AB∗eψ+φ∗−2iωD + B2e2φ+2iωD + c.c. + 2(A2 + B2)]/2
+Φ
J0J1(HAe2iω(t−D)+ix + H∗Be−2iω(t−D)+ix + AH∗eix + HBeix + c.c.),
1 [A2e2ψ−2iωD + 2ABeψ+φ
J 2
(cid:48)(cid:48)
(cid:48)(cid:48)
r2 = r22e2iωt + r2ψe2ψ + rψφeψ+φ + rψφ∗ eψ+φ∗
rBH∗ e−2iωt+ix + rAH∗ eix + rHBeix + c.c. + r20,
+ r2φe2φ + rHAe2iωt+ix +
J1e−2iωD e−2iωDHA,
J1e2iωD e2iωDBH∗,
(cid:48)(cid:48)
Φ
J0J1
2iω + 1 − Φ
(cid:48)(cid:48)
(cid:48)
(cid:48)
(cid:48)(cid:48)
Φ
J0J1
−2iω + 1 − Φ
(cid:48)
AH∗,
Φ
J0J1
1 − Φ
J1
Φ
J0J1
1 − Φ
J1
0H2 + J 2
(cid:48)(cid:48)
(J 2
Φ
HB,
(cid:48)(cid:48)
(cid:48)
rHA =
rBH∗ =
rAH∗ =
rHB =
r20 =
1A2 + J 2
1B2)
.
1 − J0Φ
(cid:48)
(cid:48)(cid:48)
0 (Hr20 + H∗r22) + Φ
J 2
BH∗ + Ar∗
J1(J0Ar20 + J2A∗r2ψ + J0Brψφ∗ + J2B∗rψφ +
ϑ(3): The nonlinear forcing at cubic order is
1 (Ar∗
J 2
HB + B∗rHB) + Φ
(cid:48)(cid:48)
N3 = [Φ
Br∗
(cid:48)(cid:48)
[Φ
J0(HrAH∗ + HrHB + H∗rHA)) + Φ
[Φ
J0(HrBH∗ + H∗rAH∗ + H∗rHB)) + Φ
ψφ∗ + J2A∗rψφ + J0Br20 + J2B∗r2φ] +
J1(J0Ar∗
(cid:48)(cid:48)(cid:48)
(cid:48)(cid:48)(cid:48)
(cid:48)(cid:48)(cid:48)
(cid:48)(cid:48)
AH∗ + A∗rHA + B∗rAH∗ +
0H2H/2]eiω(t−D) +
J 3
(A2A/2 + B2A)]eψ−iωD +
(B2B/2 + A2B)]eφ+iωD + . . .
Eliminating terms with dependencies eiωt, eiψ and eiφ yields the three coupled
amplitude equations (27a-27c):
(cid:104)
H2
2
+ A2 + B2)H + AH∗B∗(cid:105)
,
∂T H = (µ + iΩ)∆J0H + 2(α + iβ)
∂T A = (µ + iΩ)∆J1A + (a + ib)A2A + (c + id)B2A +
(
∂T B = (µ − iΩ)∆J1B + (a − ib)B2B + (c − id)A2B +
(α + iβ)(2H2A + H 2B),
(α − iβ)(2H2B + H∗2A),
where α + iβ, a + ib and c + id are given by Eqs. (B.5, B.6,B.7), respectively.
In the small delay limit (D → 0) we can use the asymptotic values given by
52
Figure B.25: Top: The real part of the cubic coefficient at the codimension-2 point. The solid
line is the full expression, Eq.(B.6) and the dotted line is the asymptotic result in the D → 0
limit, Eq.(B.8). Bottom: The real part of the cross-coupling coefficient at the codimension-2
point. The solid line in the full expression, Eq.(B.7) and the dotted line is the asymptotic
results in the D → 0 limit, Eq.(B.9).
Eqs. (11) to obtain, to leading order,
(cid:32)
(cid:33)
,
− π
2
2 + i)
(DΦ)3
(cid:32)
a + ib = − χ( π
(cid:34)
c + id = − χ
(cid:48)(cid:48)
(DΦ)3
(cid:48)(cid:48)
(Φ
Φ
(cid:48)(cid:48)(cid:48)
)2
(cid:48) − Φ
2
(3π − 2)
(cid:48)(cid:48)
)2
(cid:48)
(Φ
Φ
5
(B.8)
(cid:48)(cid:48)(cid:48)
− πΦ
2
+
(cid:48)(cid:48)
J2(Φ
1 − Φ
(cid:35)(cid:33)
)2
J2
(cid:48)
i
)2
(cid:48)
(Φ
Φ
(6 + π)
5
(cid:48)(cid:48)
− J2(Φ
1 − Φ
(cid:48)
)2
J2
(cid:48)(cid:48)(cid:48)
− Φ
.
(B.9)
where χ ≡ π3/(8 + 2π2) and α + iβ is given by Eq.(19b).
Solutions and their stability
Oscillatory uniform OU:
This solution can be expressed as (H, A, B) = (Heiωt, 0, 0), where
(cid:114)−µ∆J0
(cid:17)
(cid:16)
α
Ω − βµ
α
,
∆J0.
H =
ω =
The stability of this solution can be studied with the ansatz
(H, A, B) = (MHeiωt(1 + δH+eλt + δH∗−eλ∗t), eiωt(δA+eλt + δA∗−eλ∗t),
e−iωt(δB+eλt + δB∗−eλ∗t)),
53
1100100001e+06-a00.20.40.60.81D1100100001e+06-c which leads to three pairs of coupled linear equations which determine the six
eigenvalues λ. The first pair is restricted to the linear subspace of the small
amplitude limit cycle and results in the standard stability problem which yields
one stable eigenvalue λ = −µ∆J0 and one zero eigenvalue corresponding to
a shift in the phase of the oscillation. The other two pairs, which span the
subspaces of (δA+, δB+) and (δA−, δB−) respectively, give
(cid:18)
M
−(α − iβ)H2
−(α + iβ)H2
M∗
(cid:19)(cid:18) δA+
(cid:19)
δB+
= 0,
where M = λ − µ∆J1 − 2αH2 + i(Ω(∆J0 − ∆J1) − βH2), and the complex
conjugate matrix spanning (δA−, δB−). Setting the determinant equal to zero
yields the characteristic equation
0 = λ2 − 2λ(∆J1 − 2∆J0)µ + µ2(∆J1 − 4∆J1∆J0 + 3∆J0) +
Ω2(∆J1 − ∆J0)2 − 2
β
α
µΩ∆J0(∆J1 − ∆J0).
We find an oscillatory instability for ∆J1 = 2∆J2 while a steady instability
occurs for ∆J1 = ∆J0. The steady instability therefore always precedes the
oscillatory one.
Traveling waves (TW):
This solution can be expressed as (H, A, B) = (0,AT W eiωt, 0), where
(cid:114)−µ∆J1
(cid:17)
(cid:16)
a
Ω − bµ
a
,
∆J1.
AT W =
ω =
The stability of this solution can be studied with the ansatz
(H, A, B) = (eiωt(δH+eλt + δH∗−eλ∗t),AT W eiωt(1 + δA+eλt + δA∗−eλ∗t),
e−iωt(δB+eλt + δB∗−eλ∗t)),
which results in four coupled linear equations corresponding to the stability
problem for TW in the competition between SW and TW (see section D, Turing-
Hopf Bifurcation). Here we assume that the TW solution is supercritical and
stable. We then turn our attention to the remaining two linear equations which
describe the growth of the oscillatory uniform mode. These equations are un-
coupled and yield the complex conjugate eigenvalues
(cid:16)
(cid:17) ± i
(cid:16)
λ = −µ
∆J1 − ∆J0
2
α
a
Ω(1 − a
2α
) + (b − 2β)
µ
2α
∆J0,
(cid:17)
from which it is easy to see that an instability occurs for ∆J1 = a
instability will generically occur with non-zero frequency.
2α ∆J0. This
54
Standing waves (SW):
This solution can be expressed as (H, A, B) = (0,ASW eiωt,ASW e−iωt),
where ASW and ω are given by Eqs. (30, 31)
(cid:115)−µ∆J1
(cid:16)
(a + c)
Ω − (b + d)
(a + c)
,
(cid:17)
µ
∆J1.
ASW =
ω =
The stability of this solution can be studied with the ansatz
(H, A, B) = (eiωt(δH+eλt + δH∗−eλ∗t),ASW eiωt(1 + δA+eλt + δA∗−eλ∗t),
ASW e−iωt(1 + δB+eλt + δB∗−eλ∗t)).
This ansatz results in four coupled equations for the stability of SW in the
competition between SW and TW. Here we assume that the SW solution is
supercritical and stable. The remaining two equations describe the growth of
the oscillatory uniform mode.
N
−2ASW (α − iβ)
−2ASW (α + iβ)
N∗
(cid:19)
(cid:19)(cid:18) δH+
δH−
where N = λ + iω − (µ + iΩ)∆J0 − 4ASW (α + iβ). Setting the determinant to
zero yields the characteristic equation for the eigenvalues
0 = µ2
∆J0 − 4α∆J1
(a + c)
1 (α2 + β2)
4µ2∆J 2
(a + c)2
+
∆J1
Ω − (b + d)µ
(a + c)
+
4βµ
(a + c)
(cid:105)
− 2µλ
∆J0 − 4α∆J1
(a + c)
+ λ2.
(B.11)
The conditions for oscillatory and steady instabilities
setting λ equal to i¯ω and 0 respectively.
(32,33), are found by
Mixed Mode:
Mixed mode solutions are found by applying the ansatz Eq.(35) to Eqs. (27a-
27c). This gives
H = µ∆J0H + α(H2 + 2A2 + 2B2)H + 2HAB(α cos φ − β sin φ),
A = µ∆J1A + aA3 + cB2A + 2αH2A + H2B(α cos φ + β sin φ),
B = µ∆J1B + aB3 + cA2B + 2αH2B + H2A(α cos φ + β sin φ),
φ = 2Ω(∆J1 − ∆J0) + 2βH2 + (b + d − 4β)(A2 + B2) −
(cid:16)H2B
B + 4AB(cid:17)
H2A
α sin φ
A +
(cid:16)H2B
B − 4AB(cid:17)
H2A
,(B.12)
+ β cos φ
A +
55
(cid:18)
(cid:18)
(cid:104)
(cid:19)2
(cid:16)
(cid:104)
= 0.
(B.10)
(cid:105) − Ω∆J0
(cid:17)2 −
where φ = ψA − ψB − 2θ. One steady state solution of these equations takes
the form (H,A,B, φ) = ( H, A,− A, φ), where
(cid:0) − 2α + α cos φ − β sin φ(cid:1)
−µ∆J0(a + c) − 2µ∆J1
α(a + c) −(cid:2)4α − 2(α cos φ − β sin φ)(cid:3)(cid:2)2α − α cos φ + β sin φ(cid:3) ,(B.13)
α(a + c) −(cid:2)4α − 2(α cos φ − β sin φ)(cid:3)(cid:2)2α − α cos φ + β sin φ(cid:3) ,(B.14)
−µ∆J0(−2α + α cos φ − β sin φ) − µ∆J1α
H2 =
A2 =
and φ is found by plugging Eqs. (B.13,B.14) into Eq.(B.12) and setting the left
hand side equal to zero. We do not study the stability of the mixed-mode state
here.
Appendix B.2.2. Double zero eigenvalue: Turing, Hopf
ϑ(): The solutions to the linear equation are time periodic oscillations and
spatially periodic functions
r1 = H(T )eiωt + A(T )eix + c.c.
ϑ(2): The nonlinear forcing and resulting second order solution are
(cid:48)(cid:48)
N2 =
Φ
2
(cid:48)(cid:48)
Φ
0 (H 2e2iω(t−D) + c.c. + 2H2) +
J 2
J0J1HAeix+iωt + Φ
(cid:48)(cid:48)
J0J1AH∗e−iωt+ix + c.c.,
Φ
2
(cid:48)(cid:48)
1 (A2e2ix + c.c. + 2A2)
J 2
r2 = r2H e2iωt + r2Ae2ix + rAH eiωt+ix + rAH∗ e−iωt+ix + c.c. + r20,
rAH =
rAH∗ =
r20 =
Φ
(cid:48)(cid:48)
iω + 1 − Φ
(cid:48)(cid:48)
J0J1
(cid:48)
Φ
J0J1
−iω + 1 − Φ
(cid:48)
(cid:48)(cid:48)
Φ
0H2 + J 2
(J 2
1 − J0Φ
(cid:48)
e−iωDAH,
J1e−iωD)
J1eiωD eiωDAH∗,
1A2)
.
(cid:104)
(cid:104)
(cid:48)(cid:48)
Φ
N3 =
ϑ(3): The nonlinear forcing at cubic order is
(cid:105)
0 H∗r2H + J 2
1 Ar∗
1A2A/2)
eiωt +
(cid:105)
0 Hr20 + J 2
(J 2
0 J1H2A + J 3
(J 2
(J0J1HrAH∗ + J0J1H∗rAH + J0J1Ar20 + J1J2A∗r2A) +
(cid:48)(cid:48)
0 J1H2A + J 3
(J 2
1 A∗rAH ) +
1A2A/2)
A H + J 2
eix.
Φ
Φ
Φ
(cid:48)(cid:48)(cid:48)
(cid:48)(cid:48)(cid:48)
Eliminating terms with dependencies eiωt, eix yields the two coupled amplitude
equations, Eqs. (38a-38b),
∂T H = (µ + iΩ)∆J0H + (α + iβ)H2H + (κ + iΛ)A2H,
∂T A = ¯η∆J1A + ΓA2A + σH2A,
56
where µ + iΩ, α + iβ, ¯η and Γ are given by Eqs. (B.4, B.5,17a, 17b) respectively,
and
+ 2
(cid:48)(cid:48)
¯J 3
¯J0(Φ
1
iω + 1 − Φ
)2e−iωD
(cid:48) ¯J1e−iωD
+
(B.15)
κ + iΛ =
σ =
(cid:32) ¯J 2
0
¯J 2
k
1 − Φ
(cid:48) ¯J0
1 + DΦ
(cid:48) ¯J0e−iωD
e−iωD
(cid:33)
(cid:32) ¯J 3
0
(cid:48)(cid:48)
¯J1(Φ
)2
1 − Φ
(cid:48) ¯J0
(cid:48)(cid:48)(cid:48) ¯J0 ¯J 2
1
Φ
,
1
1 + DΦ
(cid:48) ¯J1
(cid:16)
+ Φ
(cid:48)(cid:48)(cid:48) ¯J 2
0
¯J1 + ¯J 2
0
¯J 2
1 (Φ
)2 ×
(cid:48)(cid:48)
(cid:17)(cid:33)
eiωD
−iω + 1 − Φ
(cid:48) ¯J1eiωD
+
e−iωD
iω + 1 − Φ
(cid:48) ¯J1e−iωD
.
(B.16)
In the small delay limit (D → 0) we can use the asymptotic values (11) to
obtain, to leading order,
κ + iΛ = −
(π/2 + i)
(1 + π2/4)
(cid:48)
)3
(cid:48)(cid:48)
(Φ
Φ
)2
(cid:48) − Φ
(cid:48)(cid:48)(cid:48)
(cid:32)
(cid:33)
π
2D(Φ
(cid:48)
(cid:32)
η = Φ
+ ϑ(D),
Γ =
1
(cid:48)
(Φ
)3
(cid:48)(cid:48)
)2
− (Φ
(cid:48) +
Φ
(cid:48)(cid:48)(cid:48)
Φ
2
+
σ = −
π2
4D2(Φ
(cid:48)
)3
(cid:48)(cid:48)
(Φ
Φ
)2
(cid:48) − Φ
(cid:48)(cid:48)(cid:48)
(cid:32)
(cid:48)(cid:48)
J2(Φ
)
2(1 − J2Φ
(cid:33)
+ ϑ(1/D),
+ ϑ(1),
(cid:33)
+ ϑ(D),
(cid:48)
)
and µ + iΩ and α + iβ are given by Eqs. (19a,19b) respectively.
References
[1] A. Roxin, N. Brunel and D. Hansel, Phys. Rev. Lett 94, 238103 (2005).
[2] A. Roxin, N. Brunel, D. Hansel, Prog. Theor. Phys. 161, 68 (2006).
[3] N. Brunel, and V. Hakim, Neural Comp. 11, 1621 (1999).
[4] N. Brunel, J. Comput. Neurosci. 8, 183 (2000).
[5] D. J. Amit, and N. Brunel, Cerebral Cortex 7, 237 (1997).
[6] G. B. Ermentrout, Rep. Prog. Phys. 61, 353 (1998).
[7] S. Coombes, Biol. Cybern. 93, 91 (2005).
[8] H. R. Wilson, and J. D. Cowan, Biophys. J. 12, 1 (1972).
57
[9] S. Amari, Biol. Cybern. 27, 77 (1977).
[10] G. B. Ermentrout, Neural Comp. 6, 679 (1994).
[11] O. Shriki, D. Hansel, and H. Sompolinsky, Neural Comput. 15 1809-1841
(2003)
[12] G. B. Ermentrout, and J. B. McLeod, Proc. R. Soc. Edin. 123A, 461
(1993).
[13] D. J. Pinto, and G. B. Ermentrout, SIAM J. Appl. Math. 62, 206 (2001).
[14] D. J. Pinto, and G. B. Ermentrout, SIAM J. Appl. Math. 62, 226 (2001).
[15] C. Holmgren, T. Harkany, B. Svennenfors, and Y. Zilberter, J. Physiol.
551, 139 (2003).
[16] A. Hutt, M. Bestehorn, and T. Wennekers, Network 14, 351 (2003).
[17] S. Coombes, G. J. Lord and M. R. Owen, Physica D 178, 219 (2003).
[18] A. Hutt, Phys Rev. E 70, 052902 (2004).
[19] A. Hutt, and F. M. Atay, Phys. Rev. E 73, 021906 (2006).
[20] F. M. Atay, and A. Hutt, SIAM J. Appl. Math. 65, 644 (2005).
[21] V. K. Jirsa, and H. Haken, Phys. Rev. Lett. 77, 960 (1996)
[22] N. A. Venkov, S. Coombes, and P. C. Matthews, Physica D 232, 1 (2007).
[23] C. R. Laing, and S. Coombes, Network 17, 151 (2005).
[24] S. Coombes, N. A. Venkov, L. Shiau, I. Bojak, D. T. J. Liley, and
C. R. Laing, Phys. Rev. E 76, 051901 (2007).
[25] M. C. Cross, and P. C. Hohenberg, Rev. Mod. Phys. 65, 851 (1993).
[26] F. M. Atay, and A. Hutt, SIAM J. Appl. Dyn. Syst. 5, 670 (2006).
[27] A. Hutt, C. Sutherland, and A. Longtin, Phys. Rev. E 78, 021911 (2008).
[28] J. A. Roberts, and P. A. Robinson, Phys. Rev. E 78, 051901 (2008).
[29] D. Battaglia, N. Brunel, and D. Hansel, Phys. Rev. Lett. 99, 238106 (2007)
[30] S. Coombes, and C. Laing, Philos. Trans. R. Soc. A 367, 1117 (2009).
[31] R. Ben-Yishai, R. Lev Bar-Or, and H. Sompolinsky, Proc. Natl. Acad. Sci.
USA 92, 3844 (1995).
[32] D. Hansel, and H. Sompolinsky, inMethods in Neuronal Modeling, edited
by C. Koch and I. Segev (MIT press, Cambridge, MA, 1998), 2nd ed.
58
[33] R. Curtu and B. Ermentrout, SIAM J. Applied Dynamical Systems. 3, 191
(2004).
[34] N. Brunel and X. J. Wang, J. Neurophysiol. 90, 415 (2003).
[35] C. Geisler, N. Brunel and X. J. Wang, J. Neurophysiol. 94, 4344 (2005).
[36] D. Hansel, and C. van Vreeswijk, J. Neurosci. 22, 5118 (2002).
[37] A. M. Thomson, and C. Lamy, Fontiers in Neuroscience 1, 19 (2007).
[38] E. Kandel, J. H. Schwartz, and T. M. Jessell, Principles of Neural Science
(Mc.Graw Hill, 1991).
[39] P. Girard, J. M. Hup´e, and J. Bullier, J. Neurophysiol. 85,1328 (2001).
[40] G. B. Ermentrout, and J. D. Cowan, SIAM J. on Appl. Math. 38, 1 (1980).
[41] E. K. Miller, and T. W. Troyer, J. Neurophysiol. 87, 653 (2002).
[42] B. Hellwig, Biol. Cyber. 82, 111 (2000).
[43] X.-J. Wang, and G. Buzs´aki, J. Neurosci. 16, 6402 (1996).
[44] M. H. Holmes, Introduction to Perturbation Methods (Springer, New York,
1995).
59
|
1506.06581 | 1 | 1506 | 2015-06-22T12:58:47 | Macroscopic description for networks of spiking neurons | [
"q-bio.NC",
"cond-mat.dis-nn",
"nlin.AO",
"nlin.CD"
] | A major goal of neuroscience, statistical physics and nonlinear dynamics is to understand how brain function arises from the collective dynamics of networks of spiking neurons. This challenge has been chiefly addressed through large-scale numerical simulations. Alternatively, researchers have formulated mean-field theories to gain insight into macroscopic states of large neuronal networks in terms of the collective firing activity of the neurons, or the firing rate. However, these theories have not succeeded in establishing an exact correspondence between the firing rate of the network and the underlying microscopic state of the spiking neurons. This has largely constrained the range of applicability of such macroscopic descriptions, particularly when trying to describe neuronal synchronization. Here we provide the derivation of a set of exact macroscopic equations for a network of spiking neurons. Our results reveal that the spike generation mechanism of individual neurons introduces an effective coupling between two biophysically relevant macroscopic quantities, the firing rate and the mean membrane potential, which together govern the evolution of the neuronal network. The resulting equations exactly describe all possible macroscopic dynamical states of the network, including states of synchronous spiking activity. Finally we show that the firing rate description is related, via a conformal map, with a low-dimensional description in terms of the Kuramoto order parameter, called Ott-Antonsen theory. We anticipate our results will be an important tool in investigating how large networks of spiking neurons self-organize in time to process and encode information in the brain. | q-bio.NC | q-bio | Macroscopic description for networks of spiking neurons
Ernest Montbri´o,1 Diego Paz´o,2 and Alex Roxin3
1Center for Brain and Cognition. Department of Information and
Communication Technologies, Universitat Pompeu Fabra, 08018 Barcelona, Spain
2Instituto de F´ısica de Cantabria (IFCA), CSIC-Universidad de Cantabria, 39005 Santander, Spain
3Centre de Recerca Matem`atica, Campus de Bellaterra, 08193 Bellaterra (Barcelona), Spain.
5
1
0
2
n
u
J
2
2
]
.
C
N
o
i
b
-
q
[
1
v
1
8
5
6
0
.
6
0
5
1
:
v
i
X
r
a
(Dated: September 10, 2018)
A major goal of neuroscience, statistical physics and nonlinear dynamics is to understand how brain function
arises from the collective dynamics of networks of spiking neurons. This challenge has been chiefly addressed
through large-scale numerical simulations. Alternatively, researchers have formulated mean-field theories to
gain insight into macroscopic states of large neuronal networks in terms of the collective firing activity of the
neurons, or the firing rate. However, these theories have not succeeded in establishing an exact correspondence
between the firing rate of the network and the underlying microscopic state of the spiking neurons. This has
largely constrained the range of applicability of such macroscopic descriptions, particularly when trying to de-
scribe neuronal synchronization. Here we provide the derivation of a set of exact macroscopic equations for
a network of spiking neurons. Our results reveal that the spike generation mechanism of individual neurons
introduces an effective coupling between two biophysically relevant macroscopic quantities, the firing rate and
the mean membrane potential, which together govern the evolution of the neuronal network. The resulting equa-
tions exactly describe all possible macroscopic dynamical states of the network, including states of synchronous
spiking activity. Finally we show that the firing rate description is related, via a conformal map, with a low-
dimensional description in terms of the Kuramoto order parameter, called Ott-Antonsen theory. We anticipate
our results will be an important tool in investigating how large networks of spiking neurons self-organize in time
to process and encode information in the brain.
PACS numbers: 87.19.lj 05.45.Xt 87.10.-Ed 05.65.+b
Processing and coding of information in the brain necessar-
ily imply the coordinated activity of large ensembles of neu-
rons. Within sensory regions of the cortex, many cells show
similar responses to a given stimulus, indicating a high de-
gree of neuronal redundancy at the local level. This suggests
that information is encoded in the population response and
hence can be captured via macroscopic measures of the net-
work activity [1]. Moreover, the collective behavior of large
neuronal networks is particularly relevant given that current
brain measurement techniques, such as electroencephalogra-
phy (EEG) or functional magnetic resonance imaging (fMRI),
provide data which is necessarily averaged over the activity of
a large number of neurons.
The macroscopic dynamics of neuronal ensembles has been
extensively studied through computational models of large
networks of recurrently coupled spiking neurons, including
Hodgkin-Huxley-type conductance-based neurons [2] as well
as simplified neuron models, see e.g. [3 -- 5]. In parallel, re-
searchers have sought to develop statistical descriptions of
neuronal networks, mainly in terms of a macroscopic observ-
able that measures the mean rate at which neurons emit spikes,
the firing rate [6 -- 22]. These descriptions, called firing-rate
equations (FREs), have been proven to be extremely useful
in understanding general computational principles underlying
functions such as memory [23, 24], visual processing [25 -- 27],
motor control [28] or decision making [29].
Despite these efforts, to date there is no exact theory link-
ing the dynamics of a large network of spiking neurons with
that of the firing rate. Specifically, current macroscopic de-
scriptions do not offer a precise correspondence between the
microscopic dynamics of the individual neurons, e.g. individ-
ual membrane potentials, and the firing rate dynamics of the
neuronal network.
Indeed, FREs are generally derived through heuristic ar-
guments which rely on the underlying spiking activity of the
neurons being asynchronous and hence uncorrelated. As such,
firing rate descriptions are not sufficient to describe network
states involving some degree of spike synchronization. Syn-
chronization is, however, an ubiquitous phenomenon in the
brain, and its potential role in neuronal computation is the
subject of intense research [14, 30 -- 36]. Hence, the lack of
firing rate descriptions for synchronous states limits the range
of applicability of mean-field theories to investigate neuronal
dynamics.
Here we propose a method to derive the FREs for networks
of heterogeneous, all-to-all coupled quadratic integrate-and-
fire (QIF) neurons, which is exact in the thermodynamic limit,
i.e. for large numbers of neurons. We consider an ansatz for
the distribution of the neurons' membrane potentials that we
denominate the Lorentzian ansatz (LA). The LA solves the
corresponding continuity equation exactly, making the system
amenable to theoretical analysis. Specifically, for particular
distributions of the heterogeneity, the LA yields a system of
two ordinary differential equations for the firing rate and mean
membrane potential of the neuronal population. These equa-
tions fully describe the macroscopic states of the network --
including synchronized states -- , and represent the first exam-
ple of an exact firing-rate description for a network of recur-
rently connected spiking neurons. We finally show how the
LA transforms, via a conformal mapping, into the so-called
Ott-Antonsen ansatz (OA) that is extensively used to inves-
tigate the low-dimensional dynamics of large populations of
phase oscillators in terms of the Kuramoto order parameter
[37].
I. MODEL DESCRIPTION
Hodgkin-Huxley-type neuronal models can be broadly
classified into two classes, according to the nature of their
transition to spiking in response to an injected current [38, 39].
Neuronal models with so-called Class I excitability, generate
action potentials with arbitrarily low frequency, depending on
the strength of the applied current. This occurs when a resting
state disappears through a saddle-node bifurcation. In con-
trast, in neurons with Class II excitability the action potentials
are generated with a finite frequency. This occurs when the
resting state loses stability via a Hopf bifurcation. The QIF
neuron is the canonical model for Class I neurons, and thus
generically describes their dynamics near the spiking thresh-
old [5, 40, 41]. Our aim here is to derive the FREs correspond-
ing to a heterogeneous all-to-all coupled population of N QIF
neurons. The correspondence is exact in the thermodynamic
limit, i.e. when N → ∞ (this convergence has been recently
studied in [42]).
The microscopic state of the population of QIF neurons is
given by the membrane potentials {Vj}j=1,...,N , which obey
the following ordinary differential equations [5]:
Vj = V 2
j + Ij ,
if Vj ≥ Vp, then Vj ← Vr.
(1)
Here the overdot denotes the time derivative, and Ij represents
an input current. Each time a neuron's membrane potential Vj
reaches the peak value Vp, the neuron emits a spike and its
voltage is reset to the value Vr. In our analysis we consider
the limit Vp = −Vr → ∞. This resetting rule captures the
spike reset as well as the refractoriness of the neuron. Without
loss of generality we have rescaled the time and the voltage in
(1) to absorb any coefficients which would have appeared in
the first two terms. The form for the input currents is:
Ij = ηj + Js(t) + I(t),
(2)
where the external input has a heterogeneous, quenched com-
ponent ηj as well as a common time-varying component I(t),
and the recurrent input is the synaptic weight J times the mean
synaptic activation s(t), which is written:
s(t) =
1
N
N
j <tZ t
Xj=1 Xk\tk
dt′aτ (t − t′)δ(t′ − tk
j ).
(3)
−∞
Here, tk
j is the time of the kth spike of jth neuron, δ(t) is the
Dirac delta function, and aτ (t) is the normalized synaptic ac-
tivation caused by a single pre-synaptic spike with time scale
τ, e.g. aτ (t) = e−t/τ /τ.
A. Continuous formulation
In the thermodynamic limit N → ∞, we drop the indices
in Eqs. (1), (2) and denote ρ(V η, t)dV as the fraction of neu-
rons with membrane potentials between V and V + dV , and
parameter η at time t. Accordingly, parameter η becomes now
2
a continuous random variable distributed according to a prob-
ability distribution function g(η). The total voltage density at
time t is then given byR ∞
lowing continuity equation:
The conservation of the number of neurons leads to the fol-
−∞ ρ(V η, t) g(η) dη.
∂tρ + ∂V (cid:2)(V 2 + η + Js + I)ρ(cid:3) = 0,
where we have explicitly included the velocity given by (1)
and (2).
(4)
II. RESULTS
The continuity equation (4) without
temporal forcing
I(t) = 0 has a trivial stationary solution. For each value
of η, this solution has the form of a Lorentzian function:
ρ0(V η) ∝ (V 2 + η + Js)−1. Physically, the Lorentzian den-
sity means that firing neurons with the same η value will be
scattered with a density inversely proportional to their speed
(1), i.e. they will accumulate at slow places and thin out at
fast places on the V axis. In addition, for those η values cor-
responding to quiescent neurons, the density ρ0 collapses at
the rest state in the form of a Dirac delta function.
Next we assume that, independently of the initial condition,
solutions of (4) generically converge to a Lorentzian-shaped
function, so that all relevant dynamics occur inside that lower
dimensional space. This fact is mathematically expressed by
the following Lorentzian Ansatz (LA) for the conditional den-
sity functions:
ρ(V η, t) =
1
π
x(η, t)
[V − y(η, t)]2 + x(η, t)2
,
(5)
which is a Lorentzian function with time-dependent half-
width x(η, t) and center at y(η, t). In the following we as-
sume that the LA (5) completely describes the macroscopic
dynamics of the network of QIF neurons and postpone the
mathematical justification of its validity to Sec. II E.
A. Macroscopic observables: Firing rate and mean membrane
potential
The half-width x(η, t) of the LA has a particularly simple
relation with the firing rate of the neuronal population (i.e. the
number of spikes per unit time). Indeed, the firing rate for
each η value at time t, r(η, t), can be computed by noting that
neurons fire at a rate given by the probability flux at infinity:
r(η, t) = ρ(V → ∞η, t) V (V → ∞η, t). The limit V → ∞
on the right hand side of this equation can be evaluated within
the LA, and gives the simple identity
x(η, t) = πr(η, t).
The (total) firing rate r(t) is then
r(t) =
1
π Z ∞
−∞
x(η, t)g(η)dη.
(6)
(7)
3
(a)
20
2
/
1
∆
/
J
10
Bistability
Stable
focus
Stable node
(b)
(c)
1
2
/
1
∆
/
r
0
-10
0
2
/
1
∆
/
v
0
-10
-5
η/∆
0
-3
-10
-5
-5
η/∆
(d)
(▲
)
1
0
-1
2
/
1
∆
/
v
-2
0
0
0
0,5
1
r/∆1/2
1,5
2
FIG. 1. (color online). Analysis of the steady states of FREs (12). (a) Phase diagram: In the wedge-shaped cyan-shaded region there is
bistability between a high and a low activity state. The boundary of the bistability region is the locus of a saddle-node bifurcation which is
exactly obtained in parametric form: (¯η, J)SN = [−π2r2 − 3∆2/(2πr)2, 2π2r + ∆2/(2π2r3)]. To the right of the dashed line, defined by
¯ηf = −[J/(2π)]2 − (π∆/J)2, there is a stable focus (shaded regions). (b) r-¯η and (c) v-¯η bifurcation diagrams for J/∆1/2 = 15. Square
symbols: Results obtained from numerical simulations of QIF neurons (see Appendix A for details). (d) Phase portrait of the system in the
bistable region (¯η/∆ = −5, J/∆1/2 = 15, triangle in panel (a)) with three fixed points: a stable focus (with its basin of attraction shaded), a
stable node, and a saddle point.
Additionally, the quantity y(η, t) is, for each value of η, the
mean of the membrane potential:
ρ(V η, t)V dV.
−∞
y(η, t) = P.V.Z ∞
−∞ f (x)dx = limR→∞R R
Here we take the Cauchy principal value, defined as
−R f (x)dx, to avoid the oth-
erwise ill-defined integral. The mean membrane potential is
then
P.V.R ∞
(8)
v(t) =Z ∞
−∞
y(η, t)g(η)dη.
(9)
B. Firing-rate equations
Substituting the LA (5) into the continuity equation (4), we
find that, for each value of η, variables x and y must obey two
coupled equations which can be written in complex form as
(10)
∂tw(η, t) = i(cid:2)η + Js(t) − w(η, t)2 + I(t)(cid:3) .
where w(η, t) ≡ x(η, t) + iy(η, t). Closing this equation re-
quires expressing the mean synaptic activation s(t) as a func-
tion of w(η, t). The simplest choice is to take the limit of in-
finitely fast synapses, τ → 0 in (3), so that we get an equality
with the firing rate: s(t) = r(t). This allows for the system of
QIF neurons (1)-(3) to be exactly described by Eqs. (10) and
(7); an infinite set of integro-differential equations if g(η) is a
continuous distribution.
Equation (10) is useful for general distributions g(η) (see
Appendix B), but a particularly sharp reduction in dimension-
ality is achieved if η is distributed according to a Lorentzian
distribution of half-width ∆ centered at ¯η:
g(η) =
1
π
∆
(η − ¯η)2 + ∆2 .
(11)
Note that this distribution accounts for the quenched variabil-
ity in the external inputs. The fact that it is Lorentzian is
unrelated to the LA for the density of membrane potentials.
Assuming (11) the integrals in (7) and (9) can be evaluated
closing the integral contour in the complex η-plane and using
the residue theorem[43]. Notably, the firing rate and the mean
membrane potential depend only on the value of w at the pole
of g(η) in the lower half η-plane:
πr(t) + iv(t) = w(¯η − i∆, t).
As a result, we only need to evaluate (10) at η = ¯η − i∆, and
thereby obtain a system of FREs composed of two ordinary
differential equations
r = ∆/π + 2rv,
v = v2 + ¯η + Jr + I(t) − π2r2.
(12a)
(12b)
This nonlinear system describes the macroscopic dynamics of
the population of QIF neurons in terms of the population firing
rate r and mean membrane potential v.
It is enlightening to compare mean-field model (12) with
the equations of the spiking neurons. Note that Eq. (12b) re-
sembles equations (1) and (2) for the individual QIF neuron,
but without spike resetting. Indeed, the macroscopic firing-
rate variable r enters as a negative feedback term in Eq. (12b),
and impedes the explosive growth of the mean membrane po-
tential v.
This negative feedback, combined with the coupling term
on the right hand side of (12a), describe the effective interac-
tion between the firing rate and mean membrane potential at
the network level. Therefore, the FREs (12) describe the ef-
fect of the single-cell spike generation and reset mechanism at
the network level.
In the following we examine the dynamics described by
Eq. (12), and show that they fully reproduce the macro-
scopic dynamics of the network of QIF neurons, even during
episodes of strong spike synchrony.
4
(a)
3
r
0
2
0
(c)
v
(e)
-2
0
0
3
x
e
d
n
i
n
o
r
u
e
N
0
(g)
)
t
(
I
3
0
-10
(b)
3
0
2
0
-2
0
0
3
(d)
(f)
(h)
0
3
0
-3
-10
0
10
time
20
30
40
0
10
20
30
time
40
50
60
70
80
FIG. 2. (color online). The transient dynamics of an ensemble of QIF model neurons (1)-(2) are exactly described by the FREs (12). The
instantaneous firing rate (Panels a,b) and mean membrane potential (Panels c,d) of the QIF neurons and the FREs are depicted in black and
orange, respectively. Panels (e,f) show the raster plots of 300 randomly selected QIF neurons of the N = 104 in the ensemble. Left Panels
(a,c,e,g): At time t = 0, a current I = 3 is applied to all neurons, and set to zero again at t = 30; stimulus I(t) shown in panel (g). Right
panels (b,d,f,h): At time t = 0 a sinusoidal current is applied to all neurons I(t) = I0 sin(ωt), with I0 = 3, ω = π/20; stimulus I(t) shown
in panel (h). Parameters: J = 15, ¯η = −5, ∆ = 1.
C. Analysis of the firing-rate equations
To begin with the analysis of Eq. (12), we first note that in
the absence of forcing, I(t) = 0, the only attractors of (12) are
fixed points. Figure 1(a) shows a phase diagram of the system
as a function of the mean external drive ¯η and synaptic weight
J, both normalized by the width of the input distribution [44].
There are three qualitatively distinct regions of the phase dia-
gram: 1 - A single stable node corresponding to a low-activity
state (white), 2 - A single stable focus (spiral) generally cor-
responding to a high-activity state (gray), and 3 - A region of
bistability between low and high firing rate (cyan; see a phase
portrait of this region in Fig. 1(d)). Comparison of a sample
bifurcation diagram of the fixed points from numerical simu-
lation of networks of QIF neurons with that obtained from the
FREs (12) shows an excellent correspondence, see Fig. 1(b,c).
A similar phase diagram can be readily reproduced by tra-
ditional heuristic firing-rate models, with one significant qual-
itative difference: the presence of a stable focus -- and hence
damped oscillations. Specifically, in the gray region of the
phase diagram in Fig. 1(a), the system undergoes oscillatory
decay to the stable fixed point. This oscillatory decay occurs
as well for the high-activity state over a large extent of the
region of bistability (cyan), see e.g. Fig. 1(d).
The presence of damped oscillations at the macroscopic
level reflects the transitory synchronous firing of a fraction
of the neurons in the ensemble. This behavior is common in
spiking neuron models with weak noise, and is not captured
by traditional firing rate models (see e.g. [22]).
D. Analysis of the firing-rate equations: Non-stationary inputs
stimuli (up to finite-size effects), we consider two types of
stimulus I(t): 1 - a step function and 2 - a sinusoidal forcing.
In both cases we simulate the full system of QIF neurons and
the FREs (12).
Figure 2 shows the system's response to the two different
In both cases the system is initially (t < 0) in a
inputs.
bistable regime and set in the low activity state, with parame-
ters corresponding to those of Fig. 1(d). Left panels of Fig. 2
show the response of the system to a step current, applied at
t = 0. The applied current is such that the system abandons
the bistable region -- see Fig. 1(a) -- and approaches the high
activity state, which is a stable focus. This is clearly reflected
in the time series r(t), v(t), where the rate equations (12) ex-
actly predict the damped oscillations exhibited by the mean-
field of the QIF neurons. The raster plot in panel (e) shows
the presence of the oscillations, which is due to the transitory
synchronous firing of a large fraction of neurons in the popula-
tion. Finally, at t = 30, the current is removed and the system
converges -- again, showing damped oscillations -- to the new
location of the (focus) fixed point, which clearly coexists with
the stable node where it was originally placed (t < 0).
The right hand panels of Fig. 2 show the response of the
model to a periodic current, which drives the system from one
side of the bistable region to the other. As a result, we observe
periodic bursting behavior when the system visits the stable
focus region of the phase diagram Fig. 1(a).
To further illustrate the potential of the FREs (12) to pre-
dict and investigate complex dynamics in ensembles of spik-
ing neurons, we present here a simple situation where the sys-
tem of QIF neurons exhibits macroscopic chaos. This is ob-
served by increasing the frequency ω of the sinusoidal driving,
so that the system cannot trivially follow the stable fixed point
at each cycle of the applied current.
To show that the FREs (12) fully describe the macroscopic
response of the population of QIF neurons to time-varying
Figure 3(a) shows a phase diagram obtained using Eq. (12).
The shaded regions indicate parameter values where the rate
(b)
2
v
0
-2
2
(c)
v
0
-2
-2
0
η
r
1
r
2
(a)
20
16
12
J
8
-4
2
(d)
r
1
(e)
0
0
0
3
x
e
d
n
i
n
o
r
u
e
N
0
(f)
0
0
3
x
e
d
n
i
n
o
r
u
e
n
d
e
r
e
d
r
O
0
0
10
20
30
time
40
50
60
FIG. 3. (color online). Firing rate model predicts the existence of
macroscopic chaos in a ensemble QIF neurons (1) with an injected
periodic current I(t) = I0 sin(ωt). (a) Phase diagram with the re-
gions where chaos is found, obtained using the rate model (12). In the
black-shaded region, there is only one chaotic attractor, whereas in
the cyan region the chaotic attractor coexists with a periodic attractor.
Dotted lines correspond to the bistability boundary of Fig. 1(a), de-
picted to facilitate comparison. (b) Chaotic trajectory obtained sim-
ulating the FREs (12); the Lyapunov exponent is λ = 0.183 . . .. (c)
Chaotic trajectory obtained from the QIF neurons -- parameters cor-
responding to the (yellow) triangle symbol in panel (a).
(d) Time
series for the rate model (orange) and QIF model (black). (e) Raster
plot corresponding to 300 randomly selected neurons. (f) Same raster
plot as in (e), ordering neurons according to their intrinsic current ηk.
Parameters: I0 = 3, ω = π; and ¯η = −2.5, J = 10.5 (triangle sym-
bol in (a)).
model has either a single chaotic attractor (in black) or a
chaotic attractor coexisting with a periodic orbit (in cyan). A
trajectory on the chaotic attractor is depicted in Fig. 3(b), and
the clearly irregular time series of the firing rate is shown in
orange in Fig. 3(d). Using the same parameters we performed
numerical simulations of the QIF neurons (1)-(2), finding a
similar attractor and irregular dynamics as in the rate model,
see Figs. 3(c) and (d). To obtain the time series shown in
Figs. 3(d), we ran the QIF neurons numerically and, after a
long transient, at time t = 0, the rate model (12) was initiated
with the values of r and v obtained from the population of
QIF neurons. The time series of the two systems, which are
initially close, rapidly diverge reflecting the chaotic nature of
5
the system.
Finally, to illustrate this chaotic state at the microscopic
level, Fig. 3(e) shows the raster plot for 300 randomly chosen
neurons, corresponding to the time series in Fig. 3(d). The
irregular firing of neurons in Fig. 3(e) has some underlying
structure that may be visualized ordering the same set of neu-
rons according to their intrinsic currents as ηk < ηk+1, with
k ∈ [1, 300], see Fig. 3(f). Clearly, the maxima of the firing
rate coincide with the synchronous firings of clusters of neu-
rons with similar η values. The size of these clusters is highly
irregular in time, in concomitance with the chaotic behavior.
E. Validity of the Lorentzian ansatz
Thus far, we have shown that the LA (5) solves the conti-
nuity equation (4), and confirmed that these solutions agree
with the numerical simulations of the original system of QIF
neurons (1)-(2). Here we further clarify why the LA holds for
ensembles of QIF neurons.
Transforming the voltage of the QIF neuron into a phase via
Vj = tan(θj/2), the system (1)-(2) becomes an ensemble of
'theta neurons' [40]:
θj = (1 − cos θj) + (1 + cos θj)[ηj + Js(t) + I(t)].
In the new phase variable θ ∈ [0, 2π), the LA (5) becomes:
(13)
ρ(θη, t) =
1
2π
Re(cid:20) 1 + α(η, t)eiθ
1 − α(η, t)eiθ(cid:21) ,
(14)
(15)
where the function α(η, t) is related with w(η, t) as
α(η, t) =
1 − w(η, t)
1 + w(η, t)
.
Equations (5) and (14) are two representations of the so-
called Poisson kernel on the half-plane and on the unit disk,
respectively. These representations are related via equation
(15), that establishes a conformal mapping from the half-plane
Re(w) ≥ 0 onto the unit disk α ≤ 1. In the next section we
show that variables r and v can be related, via the same con-
formal map, with the Kuramoto order parameter, which is a
macroscopic measure of phase coherence [45, 46].
The key observation supporting the applicability of the LA
is the fact that equation (14) turns out to be the ansatz dis-
covered in 2008 by Ott and Antonsen [37]. According to the
Ott-Antonsen (OA) theory, in the thermodynamic limit the dy-
namics of the class of systems
(16)
∂tθ(η, t) = Ω(η, t) + Im(cid:2)H(η, t)e−iθ(cid:3) ,
generally converges to the OA manifold (14). In our case, for
equation (13), we have Ω(η, t) = 1+η+Js+I and H(η, t) =
i(−1 + η + Js + I). Thus far, the convergence of (16) to
the OA manifold has been proven only for H(η, t) = H(t).
This includes the well-known Kuramoto and Winfree models
[47, 48], but not system (13). However, there are theoretical
arguments [49] that strongly suggest that the OA manifold is
also attracting for η-dependent H, as numerically confirmed
by a number of recent papers using theta neurons [50 -- 52] and
other phase-oscillator models [53 -- 57].
6
We would like to emphasize that the derivation of the FREs
does not rely on assumptions of weak coupling, separation of
time-scales, averaging, or any other approximation. Rather,
the assumptions underlying the validity of equation (12) are
that i) the QIF neurons be all-to-all connected, ii) heterogene-
ity be quenched, and iii) inputs be distributed according to a
Lorentzian distribution.
The last two assumptions may seem particularly restric-
tive. Concerning (iii), it must be stressed that for arbitrary
distributions of quenched heterogeneity the LA (5) remains
generally valid, and therefore equation (10) can be used to
discern the stability of the network states (see Appendix C).
Furthermore, in Appendix C we also show numerical simu-
lations using uniform- and Gaussian-distributed inputs, that
reveal very similar macroscopic dynamics in response to time
varying inputs. Even relaxing assumption (ii), numerical sim-
ulations with identical QIF neurons driven by external inde-
pendent Gaussian noise sources show qualitative agreement
with the FREs (12) (see Appendix D). In sum, the choice of
a quenched Lorentzian distribution g(η) is thus a mere math-
ematical convenience, whereas the insights gained from the
resulting firing-rate description are valid more generally.
Our derivation represents a sharp departure from, and a ma-
jor advance compared to previous studies in several regards.
Firstly, in the past it has only been possible to calculate the
approximate firing rate of networks of spiking neurons for
stationary states, or for weak deviations of such states [9 --
12, 15, 20, 22]. The equations which result from these deriva-
tions are difficult to solve, and typically require special nu-
merical methods. In contrast, FREs (12) can be easily ana-
lyzed and simulated, and exactly reproduce the behavior of the
spiking network far from any fixed point and for arbitrary ex-
ternal currents. Secondly, firing-rate descriptions traditionally
assume that the activity of a population of neurons is equiva-
lent to a set of uncorrelated stochastic processes with a given
rate, see e.g. [13, 14]. However, in simulations of spiking net-
works it is well known that the response of the network to
non-stationary inputs generally involves some degree of spike
synchrony. Recent theoretical work has sought to improve
on classical rate models, which act as a low-pass filter of in-
puts, by fitting them with linear filters extracted from the cor-
responding Fokker-Planck equation for the network [20, 22].
These filters tend to generate damped oscillations when the
external noise level is not too high, reflecting the presence of
spike synchrony [15, 22]. The FREs (12) also capture this
phenomenon, and reveal that the underlying mechanism is an
interplay between firing rate and subthreshold voltage. Fur-
thermore, because these equations are exact, we can explore
the full nonlinear response of the network, such as the gen-
eration of chaotic states of synchronous bursting as shown in
Fig. 3.
Recently it has been shown that, not only Kuramoto-like
models, but a much wider class of phase-models, evolve in
the OA manifold defined by (14). Specifically, networks of
pulse-coupled oscillators [48] and theta-neurons [50 -- 52] al-
low for an exact, low-dimensional description in terms of the
Kuramoto order parameter (17). Although these later works
use a finite-width phase coupling function that differs from the
FIG. 4. (color online). The conformal map (18) transforms the right
half-plane onto the unit disk. This transformation and its inverse
define a one-to-one mapping between the Kuramoto order parameter
Z and the macroscopic quantities of the population of QIF neurons:
r: firing rate, and v: mean membrane potential. Note that in the limit
Z = eiΨ (full coherence, ρ(θ) = δ(θ − Ψ)) we recover v = i(1 −
eiΨ)/(1+eiΨ) = tan(Ψ/2), as the original mapping V = tan(θ/2)
between QIF- and theta-neurons dictates.
F. Firing rate and Kuramoto order parameter
There exists a mapping between the macroscopic variables
r and v and the so-called Kuramoto order parameter Z. Equa-
tion (15) relates, for each value of η, the firing rate and the
mean membrane potential, both contained in w, with the uni-
formity of the phase density, measured by α -- note that α = 0
in (14) yields a perfectly uniform density. The Kuramoto or-
der parameter is obtained by integrating α over the whole pop-
ulation as
−∞
Z(t) =Z ∞
=Z ∞
−∞
g(η)Z 2π
0
ρ(θη, t)eiθdθ dη
g(η)α∗(η, t) dη,
(17)
where we assumed that the density ρ is the OA ansatz Eq. (14).
The quantity Z can be seen as the center of mass of a popula-
tion of phases distributed across the unit circle eiθ.
For a Lorentzian distribution of currents, g(η), we may de-
rive the exact formula (see Appendix B)
Z =
1 − W ∗
1 + W ∗ .
(18)
relating the Kuramoto order parameter Z with the firing-rate
quantity W ≡ πr + iv. Figure 4 illustrates this conformal
mapping of the right half-plane (r ≥ 0) onto the unit disk
(Z ≤ 1).
III. CONCLUSIONS
We have presented a method for deriving firing rate equa-
tions for a network of heterogeneous QIF neurons, which is
exact in the thermodynamic limit. To our knowledge, the re-
sulting system of ordinary differential equations (12) repre-
sents the first exact FREs of a network of spiking neurons.
synaptic coupling (3), the obtained low-dimensional descrip-
tion in terms of the Kuramoto order parameter is analogous to
ours, but in a different space [58]. Indeed, we showed that the
OA ansatz (14) is related, via the nonlinear transformation of
variables (15), to the LA ansatz (5). Remarkably, this transfor-
mation establishes an exact correspondence between the Ku-
ramoto order parameter and a novel, biophysically maningful
macroscopic observable which describes the firing rate and
mean membrane potential of the neuronal network. Interest-
ingly, the low-dimensional description in terms of firing rates
seems to be a more natural description for networks of spik-
ing neurons, compared to that in terms of the Kuramoto order
parameter. The firing rate equations (12) take a surprisingly
simple form in the LA manifold, what makes them a valuable
tool to explore and and understand the mechanisms governing
the macroscopic dynamics of neuronal networks.
Finally, since the OA ansatz is the asymptotic solution for
systems of the form in equation (16), applying the change of
variables V = tan(θ/2) automatically implies that the LA
should hold for populations governed by:
Vj = A(ηj, t) + B(ηj , t)Vj + C(ηj, t)V 2
j ,
(19)
7
where A, B, and C are related with Ω and H as A =
[Ω + Im(H)]/2, B = −Re(H), C = [Ω − Im(H)]/2. No-
tably, equation (19) defines a wide family of ensembles of QIF
neurons.
Therefore, the LA is actually valid for more general net-
works beyond the one investigated here -- ηj in Eq. (19) may
also be a vector containing several forms of disorder. As par-
ticularly relevant cases, in Appendix E we provide the FREs
governing the dynamics of an excitatory network with both,
heterogeneous inputs η and synaptic weights J, as well as pair
of interacting excitatory and inhibitory populations of QIF
neurons. In addition, according to Eq. (19), the LA is also
valid if synapses are modeled as conductances, in which case
the reversal potentials may be distributed as well. Moreover,
the role of gap-junctions or synaptic kinetics may be consid-
ered within the same framework.
[1] B. B. Averbeck, P. E. Latham, and A. Pouget, "Neural corre-
lations, population coding and computation," Nature Reviews
Neuroscience 7, 358 -- 366 (2006).
[2] A. L. Hodgkin and A. F. Huxley, "A quantitative description of
membrane current and its application to conduction and excita-
tion in nerve," J Physiol. 117, 500 -- 544 (1952).
[3] Romain Brette, Michelle Rudolph, Ted Carnevale, Michael
Hines, David Beeman, James M Bower, Markus Diesmann,
Abigail Morrison, Philip H Goodman, Frederick C Harris Jr,
et al., "Simulation of networks of spiking neurons: a review of
tools and strategies," Journal of computational neuroscience 23,
349 -- 398 (2007).
[4] E. M. Izhikevich and G. M. Edelman, "Large-scale model of
mammalian thalamocortical systems," Proc. Nat. Acad. Sci.
105, 3593 -- 3598 (2008).
[5] E. M. Izhikevich, Dynamical Systems in Neuroscience. (The
MIT Press, Cambridge, Massachusetts, 2007).
[6] Hugh R Wilson and Jack D Cowan, "Excitatory and inhibitory
interactions in localized populations of model neurons," Bio-
phys. J. 12, 1 -- 24 (1972).
[7] Shun-ichi Amari, "A method of statistical neurodynamics," Ky-
bernetik 14, 201 -- 215 (1974).
[8] Walter J Freeman, Mass action in the nervous system (Aca-
demic Press, New York, 1975).
[9] W. Gerstner, "Time structure of the activity in neural network
models," Phys. Rev. E 51, 738 -- 758 (1995).
[10] S. Fusi and M. Mattia, "Collective behavior of networks with
linear (vlsi) integrate-and-fire neurons," Neural Comput. 11,
633 -- 652 (1998).
[11] Daniel J Amit and Nicolas Brunel, "Model of global sponta-
neous activity and local structured activity during delay periods
in the cerebral cortex." Cerebral Cortex 7, 237 -- 252 (1997).
[12] N. Brunel, "Dynamics of sparsely connected networks of ex-
citatory and inhibitory spiking neurons," Journal of computa-
tional neuroscience 8, 183 -- 208 (2000).
[13] P. Dayan and L. F. Abbott, Theoretical neuroscience (Cam-
bridge, MA: MIT Press, 2001).
[14] W. Gerstner and W. M. Kistler, Spiking neuron models: Single
neurons, populations, plasticity (Cambridge university press,
2002).
[15] Oren Shriki, David Hansel,
and Haim Sompolinsky, "Rate
models for conductance-based cortical neuronal networks,"
Neural Comput. 15, 1809 -- 1841 (2003).
[16] Stephen Coombes, "Waves, bumps, and patterns in neural field
theories," Biol. Cybern. 93, 91 -- 108 (2005).
[17] A. Roxin, N. Brunel, and D. Hansel, "Role of delays in shap-
ing spatiotemporal dynamics of neuronal activity in large net-
works," Phys. Rev. Lett. 94, 238103 (2005).
[18] G. Deco, V. K. Jirsa, P. A. Robinson, M. Breakspear,
and
K. Friston, "The dynamic brain: from spiking neurons to neu-
ral masses and cortical fields," PLoS Computational biology 4,
e1000092 (2008).
[19] G. B. Ermentrout and D. H. Terman, Mathematical foundations
of neuroscience, Vol. 64 (Springer, 2010).
[20] S. Ostojic and N. Brunel, "From spiking neuron models to
linear-nonlinear models," PLoS Comput. Biol. 7, e1001056
(2011).
[21] A. Roxin and E. Montbri´o, "How effective delays shape oscilla-
tory dynamics in neuronal networks," Physica D 240, 323 -- 345
(2011).
[22] Evan S Schaffer, Srdjan Ostojic, and LF Abbott, "A complex-
valued firing-rate model that approximates the dynamics of
spiking networks," PLoS Comput. Biol. 9, e1003301 (2013).
[23] J. J. Hopfield, "Neurons with graded response have collective
computational properties like those of two-state neurons," Pro-
ceedings of the national academy of sciences 81, 3088 -- 3092
(1984).
[24] G. Mongillo, O. Barak, and M. Tsodyks, "Synaptic theory of
working memory," Science 319, 1543 -- 1546 (2008).
[25] R Ben-Yishai, R Lev Bar-Or, and H Sompolinsky, "Theory of
orientation tuning in visual cortex." Proc. Nat. Acad. Sci. 92,
3844 -- 3848 (1995).
[26] David Hansel and Haim Sompolinsky, "Modeling feature selec-
tivity in local cortical circuits," in Methods in Neuronal Mod-
elling: From Ions to Networks, edited by C. Koch and I. Segev
(MIT Press, Cambridge, 1998) pp. 499 -- 567.
[27] Rub´en Moreno-Bote, John Rinzel, and Nava Rubin, "Noise-
induced alternations in an attractor network model of perceptual
bistability," J. Neurophysiol. 98, 1125 -- 1139 (2007).
[28] K. Zhang, "Representation of spatial orientation by the intrin-
sic dynamics of the head-direction cell ensemble: a theory," J.
Neurosci. 16, 2112 -- 2126 (1996).
[29] K.-F. Wong and X.-J. Wang, "A recurrent network mechanism
of time integration in perceptual decisions," J. Neurosci. 26,
1314 -- 1328 (2006).
[30] Moshe Abeles, Corticonics: Neural circuits of the cerebral cor-
tex (Cambridge University Press, 1991).
[31] F. Varela, J.-P. Lachaux, E. Rodriguez, and J. Martinerie, "The
brainweb: phase synchronization and large-scale integration,"
Nature reviews neuroscience 2, 229 -- 239 (2001).
[32] Andreas K Engel and Wolf Singer, "Temporal binding and the
neural correlates of sensory awareness," Trends in cognitive sci-
ences 5, 16 -- 25 (2001).
[33] L. M. Ward, "Synchronous neural oscillations and cognitive
processes," Trends in cognitive sciences 7, 553 -- 559 (2003).
[34] P. Fries, "A mechanism for cognitive dynamics: neuronal com-
munication through neuronal coherence," Trends in cognitive
sciences 9, 474 -- 480 (2005).
[35] Gyorgy Buzsaki, Rhythms of the Brain (Oxford University
Press, 2006).
[36] J. Fell and N. Axmacher, "The role of phase synchronization
in memory processes," Nature Reviews Neuroscience 12, 105 --
118 (2011).
[37] E. Ott and T. M. Antonsen, "Low dimensional behavior of large
systems of globally coupled oscillators," Chaos 18, 037113
(2008).
[38] AL Hodgkin, "The local electric changes associated with repet-
itive action in a non-medullated axon," The Journal of physiol-
ogy 107, 165 -- 181 (1948).
[39] J. Rinzel and B. Ermentrout, "Analysis of neural excitability
and oscillations," in Methods in Neuronal Modelling: From
Ions to Networks, edited by C. Koch and I. Segev (MIT Press,
Cambridge, 1998) pp. 251 -- 291.
[40] B. Ermentrout and N. Kopell, "Parabolic bursting in an ex-
citable system coupled with a slow oscillation," SIAM J. Appl.
Math. 46, 233 -- 253 (1986).
[41] PE Latham, BJ Richmond, PG Nelson, and S Nirenberg, "In-
trinsic dynamics in neuronal networks. i. theory," Journal of
Neurophysiology 83, 808 -- 827 (2000).
[42] M. A. Buice and C. C. Chow, "Dynamic finite size effects in
spiking neural networks," PLoS Comput. Biol. 9, e1002872
(2013).
[43] We make an analytic continuation of w(η, t) from real η into
complex-valued η = ηr + iηi. This is possible into the lower
half-plane ηi < 0, since this guarantees the half-width x(η, t)
remains positive zero: ∂tx(η, t) = −ηi > 0 at x = 0. There-
fore we closed the integrals in (7) and (9) with an arc ηeiϑ with
η → ∞ and ϑ ∈ (−π, 0). This contour encloses one pole of
the Lorentzian distribution (11), which written in partial frac-
tions reads: g(η) = (2πi)−1[(η− ¯η−i∆)−1−(η− ¯η+i∆)−1].
[44] The number of effective parameters can be reduced by nondi-
mensionalizing the system as: η = ¯η/∆, J = J/√∆, (r, v) =
(r, v)/√∆, t = √∆t, and I(t) = I(t/√∆)/∆.
8
[45] Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence
(Springer-Verlag, Berlin, 1984).
[46] A. S. Pikovsky, M. G. Rosenblum, and J. Kurths, Synchroniza-
tion, a Universal Concept in Nonlinear Sciences (Cambridge
University Press, Cambridge, 2001).
[47] E. Ott and T. M. Antonsen, "Long time evolution of phase os-
cillator systems," Chaos 19, 023117 (2009); E. Ott, B. R. Hunt,
and T. M. Antonsen, "Comment on "long time evolution of
phase oscillators systems"," ibid. 21, 025112 (2011).
[48] D. Paz´o and E. Montbri´o, "Low-dimensional dynamics of pop-
ulations of pulse-coupled oscillators," Phys. Rev. X 4, 011009
(2014).
[49] A. Pikovsky and M. Rosenblum, "Dynamics of heterogeneous
oscillator ensembles in terms of collective variables," Physica
D 240, 872 -- 881 (2011).
[50] T. B. Luke, E. Barreto, and P. So, "Complete classification of
the macroscopic behavior of a heterogeneous network of theta
neurons," Neural Comput. 25, 3207 -- 3234 (2013).
[51] P. So, T. B. Luke, and E. Barreto, "Networks of theta neurons
with time-varying excitability: Macroscopic chaos, multistabil-
ity, and final-state uncertainty," Physica D 267, 16 -- 26 (2014).
[52] C. R. Laing, "Derivation of a neural field model from a network
of theta neurons," Phys. Rev. E 90, 010901 (2014).
[53] E. Montbri´o and D. Paz´o, "Shear diversity prevents collective
synchronization," Phys. Rev. Lett. 106, 254101 (2011).
[54] D. Paz´o and E. Montbri´o, "The Kuramoto model with dis-
tributed shear," EPL (Europhys. Lett.) 95, 60007 (2011).
[55] E. Montbri´o and D. Paz´o, "Collective synchronization in the
presence of reactive coupling and shear diversity," Phys. Rev. E
84, 046206 (2011).
[56] D. Iatsenko, S. Petkoski, P. V. E. McClintock, and A. Ste-
fanovska, "Stationary and traveling wave states of the Ku-
ramoto model with an arbitrary distribution of frequencies and
coupling strengths," Phys. Rev. Lett. 110, 064101 (2013).
[57] D. Iatsenko, P. V. E. McClintock, and A. Stefanovska, "Oscilla-
tor glass in the generalized Kuramoto model: synchronous dis-
order and two-step relaxation," Nat. Commun. 5, 4188 (2014).
[58] However, note that the phase diagrams in Luke et al. [50] do not
fully agree with ours, and show additional instabilities. These
bifurcations are associated with the finite width of the pulses,
as it accurs in the Winfree model [48].
[59] J.A.G. Roberts and G.R.W. Quispel, "Chaos and time-reversal
symmetry. Order and chaos in reversible dynamical systems,"
Phys. Rep. 216, 63 -- 177 (1992).
[60] S. H. Strogatz and R. E. Mirollo, "Stability of incoherence in
a population of coupled oscillators," J. Stat. Phys. 63, 613 -- 635
(1991).
[61] S. H. Strogatz, R. E. Mirollo, and P. C. Matthews, "Coupled
nonlinear oscillators below the synchronization threshold: Re-
laxation by generalized Landau damping," Phys. Rev. Lett. 68,
2730 -- 2733 (1992).
[62] H. Chiba and I. Nishikawa, "Center manifold reduction for
large populations of globally coupled phase oscillators," Chaos
21, 043103 (2011).
[63] H. Daido, "Onset of cooperative entrainment in limit-cycle os-
cillators with uniform all-to-all interactions: bifurcation of the
order function," Physica D 91, 24 -- 66 (1996).
[64] Equation (10) becomes then a reversible system [59] of ordi-
nary differential equations with the invariance under time re-
versal and w → w∗, e.g. Eq. (12) for ∆ = 0 is reversible,
with invariance under t → −t, v → −v. It is advisable to ap-
proximate Eq. (10) by a set of ODEs (one for each η value)
with additional small terms breaking time-reverisibility. This
is achieved including a small term in each ODE, like ∆ in
Eq. (12a), so that g(η) is actually approximated by a sum of
very sharp Lorentzian functions.
[65] S. Watanabe and S. H. Strogatz, "Constant of motion for super-
conducting Josephson arrays," Physica D 74, 197 -- 253 (1994).
[66] A. Pikovsky and M. Rosenblum, "Partially integrable dynamics
of hierarchical populations of coupled oscillators," Phys. Rev.
Lett. 101, 264103 (2008).
[67] A subtle question is that, as it occurs in the Kuramoto model,
there exists a continuous spectrum of eigenvalues at λ =
−2w0(η)i and the complex conjugate. Indeed, for η values
such that neurons are actively firing (y0(η) = 0), the eigen-
values lie exactly on the imaginary axis. Apparently, this seems
to forbid the exponential stability obtained from Eq. (12) for
the Lorentzian distribution. The same mathematical conundrum
was encountered and solved in the Kuramoto model [60 -- 62],
what suggests similar mechanisms are operating here.
[68] Note that considering stronger forms of heterogeneity for the
synaptic weights may generally not lead to Eq. (19).
9
APPENDIX A: NUMERICAL SIMULATIONS
To numerically simulate the population of QIF neurons (1)
we used the Euler method, with time step dt = 10−4. The
population had N = 104 neurons, and the Lorentzian dis-
tribution (11) was deterministically generated using: ηj =
¯η+∆ arctan[π/2(2j−N−1)/(N +1)], where j = 1, . . . , N ,
and ∆ = 1.
The time it takes for the membrane potential of a QIF neu-
ron (with Ij > 0) to reach infinity from a given positive
value of the membrane potential is arctan(pIj /Vp)/pIj.
ForpIj ≪ Vp, this expression can be approximated as:
1
Vp
.
≈
arctan(cid:0)pIj/Vp(cid:1)
pIj
In simulations we considered Vp = −Vr = 100, and used
the previous approximation. Thus, the time for the neurons to
reach infinity from Vp is 1/Vp ≈ 10−2. Additionally, the time
taken from minus infinity to Vr is 1/Vr ≈ 10−2. Numerically,
once a neuron's membrane potential Vj satisfies Vj ≥ Vp the
neuron is reset to −Vj and held there for a refractory time
2/Vj. The neurons produce a spike when Vj reaches infinity,
i.e. a time interval 1/Vj after crossing Vp. The exact time of
the spike can not be evaluated exactly in numerical simula-
tions, since dt is finite. However, simulations agree very well
with the theory provided that 1/Vp ≫ dt.
To evaluate the mean membrane potential v, the popula-
tion average was computed discarding those neurons in the
refractory period. The choice Vp = −Vr is not needed, but
in this way the observed mean membrane potential v agrees
with the theory, consistent with the definition in (8). The mean
synaptic activation (3) was evaluated using the Heaviside step
function aτ (t) = Θ(τ − t)/τ with τ = 10−3. The instan-
taneous firing rates were obtained binning time and counting
spikes within a sliding time window of size δt = 2 × 10−2
(Figs. 2, C2, C3, D1) and δt = 4 × 10−2 (Fig. 3).
APPENDIX B: PROOF OF EQUATION (18)
The inverse of Eq. (18) reads:
W =
1 − Z ∗
1 + Z ∗
(B1)
Recalling Eqs. (7) and (9), we write the macroscopic field W
as
W (t) =Z ∞
−∞
w(η, t)g(η)dη.
where w ≡ x + iy. Inserting the conformal mapping w =
(1 − α)/(1 + α) -- the inverse of Eq. (15) -- we get
1 + α(η, t)(cid:21) g(η)dη.
−∞(cid:20) 1 − α(η, t)
W (t) =Z ∞
Using the geometric series formula, and grouping the powers
of α, we may evaluate the integrals obtaining
and integration over all the firing neurons, gives the self-
consistent condition for r0:
10
W (t) = 1 − 2Z ∗(t) + 2Z ∗
2 (t) − 2Z ∗
3 (t) + ··· ,
where the Zm are the generalized order parameters [63]:
Zm(t) =Z ∞
=Z ∞
−∞
−∞
g(η)Z 2π
0
ρ(θη, t)eimθdθ dη
g(η)[α∗(η, t)]m dη.
(B2)
r0 =Z ∞
−Jr0
c(η)g(η) dη
(C1)
which is valid for any distribution g(η) (an analogous expres-
sion was obtained in Eq. (45) of [42]). For Lorentzian g(η),
the integral in (C1) can be evaluated, and solving the result-
ing equation for r0 one obtains the steady states in agreement
with the result of setting r = v = 0 in the FREs (12).
Since, for Lorentzian g(η), Zm(t) = [α∗(η = ¯η − i∆, t)]m =
Z(t)m, we can revert the power series in (B2) and obtain the
result in (B1).
APPENDIX C: RESULTS FOR ARBITRARY
DISTRIBUTIONS OF CURRENTS
In this Appendix we compare the results in the main text,
obtained using a Lorentzian distribution of currents g(η), with
other distributions. The results are qualitatively similar as ev-
idenced by Figs. C1, C2 and C3 (cf. Figs. 1, 2, and 3 in the
main text).
Note that if a particular distribution g(η) has 2n poles (all of
them off the real axis) one can readily obtain the FREs con-
sisting of n complex-valued ordinary differential equations.
Even if g(η) does not fulfill this condition, as in the case of
Gaussian or uniform distributions, the Lorentzian Ansatz still
holds (see below).
Moreover, it is possible to efficiently simulate the dynamics
of a population by integrating Eq. (10) for a sample of η val-
ues that approximate a particular g(η) [64]. However, if g(η)
is a discrete distribution (or has a discrete part), -- as in the
case of the Dirac delta function g(η) = δ(η − ¯η) -- , the so-
called Watanabe-Strogatz theory applies [49, 65, 66], and the
'Lorentzian manifold' is no longer attracting since the dynam-
ics is extremely degenerate, akin to a Hamiltonian system.
Steady states
Considering I(t) = 0 in Eqs. (1) and (2), it is possible to
find the solutions with steady firing rate r(t) = r0 for arbitrary
distributions of currents.
As we pointed out in the main text, the stationary solution
of the continuity equation (4) for those neurons that are intrin-
sically active (η > −Jr0) must be is inversely proportional to
their speed, that is
Linear stability analysis of steady states
Assuming that the LA captures the actual dynamics of the
population of QIF neurons, one can also investigate the stabil-
ity of steady states for arbitrary distributions g(η). Indeed, the
evolution of infinitesimal perturbations from a steady state
w0(η) ≡ x0(η) + iy0(η) =(√η + Jr0
−i√−η − Jr0
if η > −Jr0
if η ≤ −Jr0
is determined by the linearization of Eq. (10):
∂tδw(η, t) = i(Jδr − 2w0δw)
(C2)
−∞[δw(η, t) + δw∗(η, t)]g(η)dη.
where δr(t) = (2π)−1R ∞
To find the discrete spectrum of eigenvalues [67], we use
the ansatz δw(η, t) = b(η)eλt in (C2) and obtain
[λ+2w0(η)i]b(η)eλt =
iJ
2π Z ∞
−∞
[b(η)eλt+b∗(η)eλ∗t]g(η)dη.
Solving the equation for beλt, we find
b(η)eλt =
1
2w0(η) − iλ
J
2π Z ∞
−∞
[b(η)eλt+b∗(η)eλ∗t]g(η)dη.
After summing the complex conjugate at each side of the
equation, we multiply by g(η) and integrate over η. This
allows one to cancel out the integrals R (beλt + b∗eλ∗t)gdη.
Then, imposing self-consistency, we find
1 =
J
2π Z ∞
−∞(cid:20)
1
2w0(η) − iλ
+ c.c.(cid:21) g(η)dη.
(C3)
Local bifurcations are associated with marginal stability of the
fixed points: Re(λ) → 0. In correspondence with the results
obtained for a Lorentzian distribution in the main text, we ex-
pect bifurcations of saddle-node type associated with λ = 0.
Therefore, from Eq. (C3) we find that the loci of the saddle-
node bifurcations are located at
ρ0(V η) =
c(η)
V 2 + η + Jr0
,
JSN =
where the normalizing constant is c(η) = √η + Jr0/π. The
firing rate for each value of η is then
r0(η) = ρ0(V → ∞η) V (V → ∞η, t) = c(η),
Since x0 = √η + Jr0 for η > −Jr0, and zero otherwise,
we can restrict the integration over η to the range (−Jr0,∞).
Now multiplying Eq. (C4) by Eq. (C1) we get, after defining
2π
x0(η)
w0(η)2 g(η)dη
.
(C4)
−∞
R ∞
(a)
2
/
1
γ
/
J
20
10
0
-2
(b)
20
2
/
1
σ
/
J
10
-1
η/γ
0
0
-3
-2
η/σ
-1
0
FIG. C1.
(color online). Phase diagrams for (a) uniform and (b)
Gaussian distributions of currents g(η). The cyan-shaded region
represents the bistable region, with the solid lines corresponding to
saddle-node bifurcations analytically obtained from Eqs. (C5) and
(C6). Square symbols are estimations of the bifurcations loci ob-
tained from by direct numerical simulations of N = 104 QIF neu-
rons. Triangle symbols indicate parameter values used in numerical
simulations of Figs. C2 and C3.
ξ = JSN r0, two equations that permit to find the locus of the
saddle-node bifurcations systematically for arbitrary distribu-
tions of currents:
with ξ obtained solving
JSN =
ξ =
2π
R ∞
0 η−1/2g(η − ξ)dη
2R ∞
0 η1/2g(η − ξ)dη
R ∞
0 η−1/2g(η − ξ)dη
.
,
(C5)
(C6)
Bistability region for uniform and Gaussian distributions
Equations (C5) and (C6) are particularly easy to solve for
uniform distributions of currents:
g(η) =( 1
2γ
0
forη − ¯η < γ
otherwise
After defining the rescaled parameters η = ¯η/γ and J =
J/√γ, we find two branches of saddle-node bifurcations em-
anating from the cusp point at η = −1/3:
These functions are plotted in Fig. S1(a). Numerical simula-
tions using the network of QIF neurons confirm the correct-
ness of these boundaries (see the square symbols in Fig. C1).
For Gaussian distributions,
g(η) =
1
√2πσ
e−(η−¯η)2/(2σ2),
the solutions of (C5) and (C6) can be numerically found. The
results are shown in Fig. C1(b). Again, there is a perfect
agreement between Eqs. (C5) and (C6) -- derived assuming
the Lorentzian Ansatz -- and the numerical estimations ob-
tained simulating the network of QIF neurons.
J (1)
SN =
J (2)
SN =
2π
√3η + 3
rη + 1 + 2q 1
2π
3 + η2 −rη − 1 + 2q 1
3 + η2
11
Excitatory networks with external periodic currents
Firing rate equations (12) predict the existence of a stable
focus in the shaded region of Fig. 1. Trajectories attracted to
this fixed point display oscillations in the firing rate and mean
membrane potential due to the transient synchronous firing of
the QIF neurons.
When an external periodic current is injected to all neurons
in the network, the spiral dynamics around the fixed point is
responsible for the bursting behavior observed in Fig. 2, as
well as for the emergence of the macroscopic chaos shown in
Fig. 3. Here we investigate whether similar phenomena oc-
cur when an external periodic current of the same frequency
is injected in an excitatory network with either uniform or
Gaussian-distributed currents.
We introduced a sinusoidal forcing I(t) = I0 sin(ωt) ob-
serving a behavior qualitatively identical to the one reported
in the main text with Lorentzian g(η). Under a low-frequency
forcing the system displays periodic bursting, see Fig. C2,
provided the parameters are set inside the bistable region (see
the black triangles in Fig. C1). Furthermore, note that the
range of firing rates and mean membrane potentials in Fig-
ure C2 are similar to those of Fig. 2 -- for Lorentzian distribu-
tions of currents.
As for the simulations shown in Figure (3), we next increase
the frequency of the injected current up to ω = π to investi-
gate the emergence of macroscopic chaos in a network with
Gaussian-distributed currents. Fig. C3 shows the emergence
of an apparently chaotic state. The observed dynamics is sim-
ilar to that of Fig. 3, which suggests it is chaotic. This type
of chaos persists in the thermodynamic limit N → ∞. On
top of this, highly-dimensional but weakly chaotic dynamics
is probably also present due to 'residual' finite-size effects.
APPENDIX D: IDENTICAL QIF NEURONS DRIVEN BY
INDEPENDENT NOISE TERMS
We compare now the results obtained above and in the main
text for quenched heterogeneity g(η) with the results for iden-
tical neurons g(η) = δ(η − ¯η) driven by noise. Specifically,
the inputs currents are now taken as
Ij = ¯η + Jr(t) + ξj (t),
(D1)
where ξj are independent white noise terms with expected val-
ues hξj (t)i = 0, and hξj (t)ξk(t′)i = 2D δjk δ(t − t′).
Fokker-Planck equation:
In the thermodynamic limit, the density ρ(V, t) obeys the
(D2)
∂tρ + ∂V (cid:2)(V 2 + ¯η + Jr)ρ(cid:3) = D ∂2
V ρ.
The Lorentzian ansatz is not a solution of this equation,
but we demonstrate here that the phenomena observed with
quenched heterogeneity arise also with independent noise
sources. This qualitative similarity at the macroscopic level
between quenched Lorentzian heterogeneity and Gaussian
noise has been noted in previous work [60].
(a)
3
r
(b)
v
0
2
0
-2
-4
-6
(c)
0
0
3
x
e
d
n
i
n
o
r
u
e
N
0
12
(e)
3
(f)
(g)
0
2
0
-2
0
0
3
0
(d)
)
t
(
I
1
0
-1
-10
0
10
20
30
time
40
50
60
70
80
(h)
1.5
0
-1.5
-10
0
10
20
30
time
40
50
60
70
80
FIG. C2. Numerical simulations of the excitatory network of QIF neurons, Eqs. (1), and (2). with (a-d) uniform and (e-h) Gaussian distributions
of currents. As in Fig. 2, at time t = 0, an external sinusoidal current I(t) = I0 sin(ωt) -- shown in panels (d,h) -- is applied to all neurons.
Panels (a,e) show the time series of the firing rate, and (b,f) the mean membrane potential of all cells. In panels (c,g) we depicted the raster
plots of 300 randomly chosen neurons. Parameters correspond to the black triangle symbols in Fig. C1: (a-d) ¯η = −1, J = 8, γ = 1, I0 = 1;
(e-h) ¯η = −2, J = 10, σ = 1, I0 = 1.5.
(a)
r
1
0
(b)
0
0
3
x
e
d
n
i
n
o
r
u
e
N
0
(c)
0
0
3
x
e
d
n
i
n
o
r
u
e
n
d
e
r
e
d
r
O
0
0
to each fixed point (see figure caption for details).
Additionally, in order to investigate and compare the dy-
namical behavior of the identical noise-driven neurons with
that of the FREs (12), an external current of intensity I0 = 3
is injected to all neurons at time t = 0, like in Fig. 2.
In
Fig. D1(b,c) the time series of the firing rate and the mean
membrane potential clearly display damped oscillations after
the injection of the current, confirming the existence of a sta-
ble focus, exactly as observed in the FREs (12). The existence
of a stable focus reflects the presence of transient spike syn-
chrony in the network, as seen by the raster plot in Fig. D1(d).
Remarkably, the raster plot and the time series closely resem-
ble those of Fig. 2. The resemblance is not just qualitative,
but rather there is a near quantitative fit between the network
of QIF neurons driven by Gaussian noise and the FREs (12),
which were derived assuming quenched Lorentzian hetero-
geneity (orange curves).
20
time
40
60
APPENDIX E: MODEL GENERALIZATIONS
FIG. C3. Chaotic state in a network of N = 104 QIF neurons with
Gaussian-distributed currents. Network parameters correspond to the
yellow-triangle symbol in the phase diagram of Fig. S1(b). Like in
Fig. 3, neurons receive a common periodic current I(t) = I0 sin(ωt)
of frequency ω = π. Parameters: ¯η = −1, J = 8, σ = 1, I0 = 1.5.
To illustrate the potential of the LA for investigating more
sophisticated networks, here we provide the low-dimensional
FREs corresponding to an excitatory network of QIF neurons
with independently distributed currents and synaptic weights,
and to two interacting populations of excitatory and inhibitory
QIF neurons with distributed currents.
A subtle point in Eq. (D2) is that its nondimensionalization
entails a different (compared to ∆) rescaling with D: V =
V /D1/3, η = η/D2/3, J = J/D1/3, t = tD1/3 (implying
r = r/D1/3).
Numerical simulations reveal the existence of a region of
bistability, see Fig. D1(a), analogous to the one observed for
quenched heterogeneity, cf. Figs. 1(a) and C1. Obviously, true
bistability only holds in the thermodynamic limit, while what
we observe are rather exceedingly long residence times close
Excitatory population with heterogeneous currents and synaptic
weights
As a first example, let us assume that both the currents η
and the synaptic weights J are distributed -- this type of het-
erogeneity was also considered in [42]. The input currents
read then [68]
Ij = ηj + Jjr(t) + I(t).
(a)
20
15
3
/
1
D
/
J
10
5
0
13
(b)
3
r
(c)
v
0
2
0
-2
0
0
3
0
x
e
d
n
i
n
o
r
u
e
N
(d)
(e)
-10
-5
η/D2/3
0
)
t
(
I
3
0
-10
0
10
time
20
30
40
FIG. D1. (color online). Numerically obtained phase diagram (a) and time series (b-d) for a network of identical QIF neurons driven by
independent white noise terms. The square symbols in panel (a) represent the boundaries that enclose the region of bistability between a
high and a low activity state. The boundaries have been obtained using a network of 104 neurons, and by numerical continuation of a low
activity solution (filled squares) and a high activity solution (empty squares). Specifically, using a noise intensity D = 1 and a particular
value of J, the system was initialized either at ¯η = −5 or at ¯η = 0 and, after t = 10 time units, parameter ¯η was decreased/increased an
amount 0.025/0.1, respectively. This continuation was made until the relative change of two successive values of the averaged firing rate
(time-averaged over the last time unit for each parameter value) was larger than 50%. Other parameters for the numerical simulations are
the same as in all other figures, and are described in the Material and Methods section. The triangle symbol indicates the parameter value,
(¯η/D2/3, J/D1/3) = (−5, 15), corresponding to the numerical simulations shown in black in the right panels -- for the simulations we used
D = 1. To facilitate the comparison with the FREs (12), the orange curves show the time series of the FREs, using the same parameters
(¯η, J) = (−5, 15), but with ∆ = 1
For simplicity, let us additionally assume that η and J are
distributed independently, with a joint distribution p(η, J) =
g(η)h(J).
In the simplest situation of Lorentzian g(η) and
h(J)
g(η) =
∆/π
(η − ¯η)2 + ∆2 ;
h(J) =
Γ/π
(J − ¯J)2 + Γ2 .
the problem is extremely simplified using the Lorentzian
ansatz, which now trivially reads:
ρ(V η, J, t) =
1
π
x(η, J, t)
[V − y(η, J, t)]2 + x(η, J, t)2 .
The firing rate and mean membrane potential are determined
only by the value of w ≡ x + iy at the poles in the lower half
planes:
r(t) + iv(t) =ZZ w(η, J, t)g(η)h(J)dηdJ
= w(¯η − i∆, ¯J − iΓ, t).
Finally, evaluating Eq. (10) in the main text at the poles (η =
¯η − i∆, J = ¯J − iΓ) we get the exact FREs:
r = ∆/π + Γr/π + 2rv,
v = v2 + ¯η + ¯Jr + I(t) − π2r2.
(E1a)
(E1b)
These equations contain simply an extra term '+Γr/π' com-
pared to equations (12). Figure E1 shows the bistability
boundaries obtained from (E1) for different ratios of Γ and
∆, and I(t) = 0. Note that the region of bistability shifts
to lower values of ¯η/∆ and to higher values of ¯J/√∆ as the
level of heterogeneity in the synaptic coupling Γ is increased.
20
15
2
/
1
∆
/
J
10
5
0
-10
Γ/ ∆1/2= 0
Γ/∆1/2= 5
Γ/ ∆1/2= 1
-5
η/∆
0
FIG. E1. (color online). Phase diagram for an excitatory population
with heterogeneous currents and synaptic weights obtained using the
FREs (E1).
Firing-rate equations for a pair of Excitatory-Inhibitory
populations
The microscopic state of each population of QIF neurons is
characterized by the membrane potentials {V (e,i)
}j=1,...,N ,
which obey the following ordinary differential equations:
j
V (e,i)
j
= (V (e,i)
j
)2 + I (e,i)
j
, if V (e,i)
j
≥ Vp, then V (e,i)
j ← Vr.
j
Here, V (e,i)
represents the membrane potential of neuron j in
either the excitatory (e), or in the inhibitory population (i).
The external currents for the excitatory and inhibitory popula-
tions are, respectively:
j = η(e)
I (e)
j = η(i)
I (i)
j + Jees(e)(t) − Jies(i)(t) + I (e)(t),
j + Jeis(e)(t) − Jiis(i)(t) + I (i)(t),
where the synaptic weights are Jee, Jii, Jie, Jei. Finally, the
mean synaptic activation for each population is
s(e,i)(t) =
1
N
N
j )(e,i)<tZ t
Xj=1 Xk\(tk
dt′aτ (t−t′)δ(t′−(tk
j )(e,i)).
−∞
Here, (tk
j )(e,i) is the time of the kth spike of jth neuron in
either the excitatory (e), or in the inhibitory population (i).
Additionally, δ(t) is the Dirac delta function, and aτ (t) is the
normalized synaptic activation caused by a single pre-synaptic
spike with time scale τ, e.g. aτ (t) = e−t/τ /τ.
14
It is straightforward to apply the LA and the method de-
scribed in the main text, to obtain the firing rate equations
corresponding to the two-population model. Considering the
limit of infinitely fast synapses, τ → 0, we get s(e,i)(t) =
r(e,i)(t). Finally, assuming the Lorentzian distributions of
currents for both populations:
ge,i(η) =
1
π
∆e,i
(η − ¯ηe,i)2 + ∆2
e,i
,
we obtain the firing-rate equations:
r(e) = ∆e/π + 2r(e)v(e),
v(e) = (v(e))2 + ¯ηe + Jeer(e) − Jier(i) + I (e)(t) − (πr(e))2,
r(i) = ∆i/π + 2r(i)v(i),
v(i) = (v(i))2 + ¯ηi + Jeir(e) − Jiir(i) + I (i)(t) − (πr(i))2.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.