The Graph Isomorphism Problem and the Module Isomorphism Problem. Harm Derksen

Size: px

Start display at page:

Download "The Graph Isomorphism Problem and the Module Isomorphism Problem. Harm Derksen"

Mervyn Underwood
5 years ago
Views:

1 The Graph Isomorphism Problem and the Module Isomorphism Problem Harm Derksen Department of Mathematics Michigan Center for Integrative Research in Critical Care University of Michigan Partially supported by NSF DMS Symmetry vs Regularity Pilsen, July, 2018

2 6/30/2018 MR: Publications results for "Author=(derksen, *) AND Title=(graph)" Click here to activate Remote Access Univ of Michigan Publications results for "Author=(derksen, *) AND Title=(graph)" MR Reviewed Derksen, Harm(1-MI) Department of Mathematics, University of Michigan Ann Arbor, Michigan, Citations From References: 0 From Reviews: 0 The graph isomorphism problem and approximate categories. (English summary) J. Symbolic Comput. 59 (2013), C60 (20G05 68Q25) Summary: "It is unknown whether two graphs can be tested for isomorphism in polynomial time. A classical approach to the Graph Isomorphism Problem is the d-dimensional Weisfeiler- Lehman algorithm. The d-dimensional WL-algorithm can distinguish many pairs of graphs, but the pairs of non-isomorphic graphs constructed by Cai, Fürer and Immerman it cannot distinguish. If d is fixed, then the WL-algorithm runs in polynomial time. We will formulate the Graph Isomorphism Problem as an Orbit Problem: Given a representation V of an algebraic group G and two elements v1, v2 V, decide whether v1 and v2 lie in the same G-orbit. Then we attack the Orbit Problem by constructing certain approximate categories Harm Derksen

3 Orbit Problems Orbit Problem G group acting on k-vector space V given v, v V, does there exist g G with g v = v?

4 Orbit Problems Orbit Problem G group acting on k-vector space V given v, v V, does there exist g G with g v = v? Graph Isomorphism Problem (Hard) Γ, Γ graphs with n vertices, A, A Mat n,n adjacency matrices G = {permutation matrices} acts on Mat n,n by conjugation does there exists a permutation matrix P with PAP 1 = A?

5 Orbit Problems Orbit Problem G group acting on k-vector space V given v, v V, does there exist g G with g v = v? Graph Isomorphism Problem (Hard) Γ, Γ graphs with n vertices, A, A Mat n,n adjacency matrices G = {permutation matrices} acts on Mat n,n by conjugation does there exists a permutation matrix P with PAP 1 = A? Module Isomorphism Problem (Easy) G = GL n acts on Mat m n,n by simultaneous conjugation A = (A 1,..., A m ), A = (A 1,..., A m) Mat m n,n is there a P GL n with (PA 1 P 1,..., PA m P 1 ) = (A 1,..., A m)?

6 Module Isomorphism Problem (Easy) R = k x 1,..., x m free associative algebra (A 1,..., A m ) Mat m n,n corresponds to module M = k n, where x i v = A i v for v M

7 Module Isomorphism Problem (Easy) R = k x 1,..., x m free associative algebra (A 1,..., A m ) Mat m n,n corresponds to module M = k n, where x i v = A i v for v M (A 1,..., A m) Mat m n,n corresponds to module M = k n Hom R (M, M ) = {P Mat n,n i PA i = A i P}

8 Module Isomorphism Problem (Easy) R = k x 1,..., x m free associative algebra (A 1,..., A m ) Mat m n,n corresponds to module M = k n, where x i v = A i v for v M (A 1,..., A m) Mat m n,n corresponds to module M = k n Hom R (M, M ) = {P Mat n,n i PA i = A i P} Probabilistic Module Isomorphism Algorithm choose P Hom R (M, M ) Mat n,n at random if P invertible, then M = M if P not invertible, then M = M with high probability polynomial time de-randomized algorithms for module isom.: Chistov-Ivanyos-Karpinsky 1997, Brooksbank-Luks 2008 (for arbitrary finitely generated associative k-algebras)

9 Graph Isomorphism by Solving Polynomial Equations Γ, Γ graphs on n vertices with adjacency matrices A = (a i,j ), A = (a i,j ) does there exists a permutation n n matrix X = (x i,j ) with XAX 1 = A?

10 Graph Isomorphism by Solving Polynomial Equations Γ, Γ graphs on n vertices with adjacency matrices A = (a i,j ), A = (a i,j ) does there exists a permutation n n matrix X = (x i,j ) with XAX 1 = A? We need to solve a system of polynomial equations:

11 Graph Isomorphism by Solving Polynomial Equations Γ, Γ graphs on n vertices with adjacency matrices A = (a i,j ), A = (a i,j ) does there exists a permutation n n matrix X = (x i,j ) with XAX 1 = A? We need to solve a system of polynomial equations: X is a permutation matrix, means: (1) x i,j x i,l = 0 = x j,i x l,i for all i and all j l (2) n j=1 x i,j 1 = n j=1 x j,i 1 = 0 for all i

12 Graph Isomorphism by Solving Polynomial Equations Γ, Γ graphs on n vertices with adjacency matrices A = (a i,j ), A = (a i,j ) does there exists a permutation n n matrix X = (x i,j ) with XAX 1 = A? We need to solve a system of polynomial equations: X is a permutation matrix, means: (1) x i,j x i,l = 0 = x j,i x l,i for all i and all j l (2) n j=1 x i,j 1 = n j=1 x j,i 1 = 0 for all i XA = A X gives us the linear equations: (3) j=1 x i,ja j,l j=1 a i,j x j,l = 0 for all i, l system of linear and quadratic equations in n 2 variables

13 Gröbner Basis? R = k[x 1,1, x 1,2,..., x n,n ] polynomial ring in n 2 variables define Eq(Γ, Γ ) R as the set of poly s from our system of equations (1)-(3) let I = (Eq(Γ, Γ )) R be the ideal generated by Eq(Γ, Γ ) Hilbert s Nullstellensatz 1 I Γ = Γ

14 Gröbner Basis? R = k[x 1,1, x 1,2,..., x n,n ] polynomial ring in n 2 variables define Eq(Γ, Γ ) R as the set of poly s from our system of equations (1)-(3) let I = (Eq(Γ, Γ )) R be the ideal generated by Eq(Γ, Γ ) Hilbert s Nullstellensatz 1 I Γ = Γ Algorithm 1, Gröbner basis (GB) compute Gröbner basis G of I using Buchberger s algorithm then 1 G 1 I Γ = Γ

15 Gröbner Basis? R = k[x 1,1, x 1,2,..., x n,n ] polynomial ring in n 2 variables define Eq(Γ, Γ ) R as the set of poly s from our system of equations (1)-(3) let I = (Eq(Γ, Γ )) R be the ideal generated by Eq(Γ, Γ ) Hilbert s Nullstellensatz 1 I Γ = Γ Algorithm 1, Gröbner basis (GB) compute Gröbner basis G of I using Buchberger s algorithm then 1 G 1 I Γ = Γ Computing a Gröbner basis is known to be very slow there is no reason to believe Algorithm 1 could be polynomial time This is a stupid approach!

16 Truncated Ideals... or is it?

17 Truncated Ideals... or is it? we restrict ourselves to computations in low degree fix positive integer d 2

18 Truncated Ideals... or is it? we restrict ourselves to computations in low degree fix positive integer d 2 R d = k[x 1,1, x 1,2,..., x n,n ] d space of polynomials of degree d dim R d is polynomial in n (for fixed d)

19 Truncated Ideals... or is it? we restrict ourselves to computations in low degree fix positive integer d 2 R d = k[x 1,1, x 1,2,..., x n,n ] d space of polynomials of degree d dim R d is polynomial in n (for fixed d) we construct subspaces I [0] I [1] of R d as follows: I [0] R d is the k-span of Eq(Γ, Γ ) (Eq(Γ, Γ ) was the set of linear and quadratic equations)

20 Truncated Ideals... or is it? we restrict ourselves to computations in low degree fix positive integer d 2 R d = k[x 1,1, x 1,2,..., x n,n ] d space of polynomials of degree d dim R d is polynomial in n (for fixed d) we construct subspaces I [0] I [1] of R d as follows: I [0] R d is the k-span of Eq(Γ, Γ ) (Eq(Γ, Γ ) was the set of linear and quadratic equations) I [j+1] = d e=0 (I [j] R e )R d e for all j

21 Truncated Ideals... or is it? we restrict ourselves to computations in low degree fix positive integer d 2 R d = k[x 1,1, x 1,2,..., x n,n ] d space of polynomials of degree d dim R d is polynomial in n (for fixed d) we construct subspaces I [0] I [1] of R d as follows: I [0] R d is the k-span of Eq(Γ, Γ ) (Eq(Γ, Γ ) was the set of linear and quadratic equations) I [j+1] = d e=0 (I [j] R e )R d e for all j

22 Truncated Ideals... or is it? we restrict ourselves to computations in low degree fix positive integer d 2 R d = k[x 1,1, x 1,2,..., x n,n ] d space of polynomials of degree d dim R d is polynomial in n (for fixed d) we construct subspaces I [0] I [1] of R d as follows: I [0] R d is the k-span of Eq(Γ, Γ ) (Eq(Γ, Γ ) was the set of linear and quadratic equations) I [j+1] = d e=0 (I [j] R e )R d e for all j for some l dim R d, I [l] = I [l+1] = I [l+2] = Let (Eq(Γ, Γ )) d = I [l] be this limit this is the d-truncated ideal generated by Eq(Γ, Γ )

23 Comparison to the Weisfeiler-Leman Algorithm a basis of (Eq(Γ, Γ )) d (as a k-vector space) can be computed with a polynomial number of arithmetic operations in the field k Algorithm 2, Truncated Ideals (TI d ) compute (Eq(Γ, Γ )) d and test whether 1 (Eq(Γ, Γ )) d if 1 (Eq(Γ, Γ )) d then Γ = Γ

24 Comparison to the Weisfeiler-Leman Algorithm a basis of (Eq(Γ, Γ )) d (as a k-vector space) can be computed with a polynomial number of arithmetic operations in the field k Algorithm 2, Truncated Ideals (TI d ) compute (Eq(Γ, Γ )) d and test whether 1 (Eq(Γ, Γ )) d if 1 (Eq(Γ, Γ )) d then Γ = Γ this algorithm is polynomial time if we work over a finite field k = F q and q = q(n) = 2 O(poly(n)) Theorem if q is a prime > n, k = F q if WL d distinguishes Γ and Γ, then TI 2d+2 distinguishes Γ and Γ so TI is as powerful as WL (but perhaps not more powerful) but there is more structure...

25 An Associative Algebra recall R = k[x 1,1, x 1,2,..., x n,n ] and R d = k[x 1,1, x 1,2,..., x n,n ] d matrix multiplication gives a ring homomorphism ϕ : R = k[x 1,1,..., x n,n ] k[y 1,1,..., y n,n, z 1,1,..., z n,n ] = R R defined by ϕ(x i,j ) = n l=1 y i,lz l,j

26 An Associative Algebra recall R = k[x 1,1, x 1,2,..., x n,n ] and R d = k[x 1,1, x 1,2,..., x n,n ] d matrix multiplication gives a ring homomorphism ϕ : R = k[x 1,1,..., x n,n ] k[y 1,1,..., y n,n, z 1,1,..., z n,n ] = R R defined by ϕ(x i,j ) = n l=1 y i,lz l,j this ring homomorphism restricts to a linear map R d R d R d,

27 An Associative Algebra recall R = k[x 1,1, x 1,2,..., x n,n ] and R d = k[x 1,1, x 1,2,..., x n,n ] d matrix multiplication gives a ring homomorphism ϕ : R = k[x 1,1,..., x n,n ] k[y 1,1,..., y n,n, z 1,1,..., z n,n ] = R R defined by ϕ(x i,j ) = n l=1 y i,lz l,j this ring homomorphism restricts to a linear map R d R d R d, which dualizes to a linear map R d R d R d,

28 An Associative Algebra recall R = k[x 1,1, x 1,2,..., x n,n ] and R d = k[x 1,1, x 1,2,..., x n,n ] d matrix multiplication gives a ring homomorphism ϕ : R = k[x 1,1,..., x n,n ] k[y 1,1,..., y n,n, z 1,1,..., z n,n ] = R R defined by ϕ(x i,j ) = n l=1 y i,lz l,j this ring homomorphism restricts to a linear map R d R d R d, which dualizes to a linear map R d R d R d, which defines a bilinear multiplication R d R d R d,

29 An Associative Algebra recall R = k[x 1,1, x 1,2,..., x n,n ] and R d = k[x 1,1, x 1,2,..., x n,n ] d matrix multiplication gives a ring homomorphism ϕ : R = k[x 1,1,..., x n,n ] k[y 1,1,..., y n,n, z 1,1,..., z n,n ] = R R defined by ϕ(x i,j ) = n l=1 y i,lz l,j this ring homomorphism restricts to a linear map R d R d R d, which dualizes to a linear map Rd R d R d, which defines a bilinear multiplication Rd R d R d, which makes Rd into an associative algebra

30 The Category C n,d Definition (Approximate Category C n,d ) objects of C n,d are graphs on n vertices,

31 The Category C n,d Definition (Approximate Category C n,d ) objects of C n,d are graphs on n vertices, Hom Cn,d (Γ, Γ ) = (R d /(Eq(Γ, Γ )) d ) R d

32 The Category C n,d Definition (Approximate Category C n,d ) objects of C n,d are graphs on n vertices, Hom Cn,d (Γ, Γ ) = (R d /(Eq(Γ, Γ )) d ) R d multiplication R d R d R d restricts to a bilinear map Hom Cn,d (Γ, Γ ) Hom Cn,d (Γ, Γ ) Hom Cn,d (Γ, Γ )

33 The Category C n,d Definition (Approximate Category C n,d ) objects of C n,d are graphs on n vertices, Hom Cn,d (Γ, Γ ) = (R d /(Eq(Γ, Γ )) d ) R d multiplication R d R d R d restricts to a bilinear map Hom Cn,d (Γ, Γ ) Hom Cn,d (Γ, Γ ) Hom Cn,d (Γ, Γ ) if 1 (Eq(Γ, Γ )) d then (Eq(Γ, Γ )) d = R d and Hom Cn,d (Γ, Γ ) = 0

34 Properties of C n,d suppose Γ, Γ graphs on n vertices with adjacency matrices A, A if Γ = Γ then there is a permutation matrix P with PA = A P

35 Properties of C n,d suppose Γ, Γ graphs on n vertices with adjacency matrices A, A if Γ = Γ then there is a permutation matrix P with PA = A P if ev P : R d k is evaluation at P, then ev P Hom Cn,d (Γ, Γ ) R d and ev P is an isomorphism in C n,d (with inverse ev P 1)

36 Properties of C n,d suppose Γ, Γ graphs on n vertices with adjacency matrices A, A if Γ = Γ then there is a permutation matrix P with PA = A P if ev P : R d k is evaluation at P, then ev P Hom Cn,d (Γ, Γ ) R d and ev P is an isomorphism in C n,d (with inverse ev P 1) Theorem let T = Hom Cn,d (Γ, Γ) (an associative k-algebra) Γ, Γ are isomorphic in C n,d Hom Cn,d (Γ, Γ) and Hom Cn,d (Γ, Γ) are isomorphic T -modules

37 Properties of C n,d suppose Γ, Γ graphs on n vertices with adjacency matrices A, A if Γ = Γ then there is a permutation matrix P with PA = A P if ev P : R d k is evaluation at P, then ev P Hom Cn,d (Γ, Γ ) R d and ev P is an isomorphism in C n,d (with inverse ev P 1) Theorem let T = Hom Cn,d (Γ, Γ) (an associative k-algebra) Γ, Γ are isomorphic in C n,d Hom Cn,d (Γ, Γ) and Hom Cn,d (Γ, Γ) are isomorphic T -modules we can test whether Γ, Γ are isomorphic in C n,d in polynomial time

38 Properties of C n,d suppose Γ, Γ graphs on n vertices with adjacency matrices A, A if Γ = Γ then there is a permutation matrix P with PA = A P if ev P : R d k is evaluation at P, then ev P Hom Cn,d (Γ, Γ ) R d and ev P is an isomorphism in C n,d (with inverse ev P 1) Theorem let T = Hom Cn,d (Γ, Γ) (an associative k-algebra) Γ, Γ are isomorphic in C n,d Hom Cn,d (Γ, Γ) and Hom Cn,d (Γ, Γ) are isomorphic T -modules we can test whether Γ, Γ are isomorphic in C n,d in polynomial time Algorithm 3 (AC d ) test whether Γ, Γ are isomorphic in the category C n,d for all fields k = F q with q a prime 2n if not isomorphic for some k, then Γ and Γ are non-isomorphic graphs Harm Derksen

39 if V = {1, 2,..., n} is the set of vertices, then WL d 1 captures reasoning on subsets of V d it is as powerful as d-variable logic with counting (see Cai-Fürer-Immerman) if W = kv = k n is the vector space whose basis is the set of vertices, then AC 2d captures reasoning with subspaces of W d = W W with operations such as tensor products, sums, intersections, projections and dimension count.

40 Cai-Fürer-Immerman constructed families of pairs of nonisomorphic graphs that cannot be distinguished by WL d for any fixed d so WL d does not give a polynomial time algorithm for a pair of CFI graphs (Γ, Γ ), we can construct matrices B and B from the adjacency matrices A and A such that B and B do not have the same rank if k = F 2 AC 3 can distinguish each pair of CFI-graphs (Γ, Γ ) if we work over k = F 2

Tractable Approximations of Graph Isomorphism.

Tractable Approximations of Graph Isomorphism. Anuj Dawar University of Cambridge Computer Laboratory visiting RWTH Aachen joint work with Bjarki Holm Darmstadt, 5 February 2014 Graph Isomorphism Graph