Opened 10 years ago
Last modified 3 years ago
#10950 closed defect
The hash function for matrices suffers from many collisions with permutation matrices — at Version 20
Reported by:  nthiery  Owned by:  nthiery 

Priority:  major  Milestone:  sage8.1 
Component:  linear algebra  Keywords:  Weyl groups, permutation matrices, hash 
Cc:  sagecombinat  Merged in:  
Authors:  Jeroen Demeyer  Reviewers:  
Report Upstream:  N/A  Work issues:  
Branch:  u/jdemeyer/the_hash_function_for_matrices_suffers_from_many_collisions_with_permutation_matrices (Commits)  Commit:  f67a8839d7dfb843a8ed8652a2734b644be579f1 
Dependencies:  #24090  Stopgaps: 
Description (last modified by )
The current hash function for generic dense matrices suffers from hard collisions with permutation matrices. For example: all permutation matrices of size 4 have hash 0!
sage: def mat(p): m = p.to_matrix(); m.set_immutable(); return m sage: [ hash(mat(p)) for p in Permutations(4) ] [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
After applying this branch:
sage: def hashmat(p): m = p.to_matrix(); m.set_immutable(); return hash(m) sage: len(set(hashmat(p) for p in Permutations(1))) 1 sage: len(set(hashmat(p) for p in Permutations(2))) 2 sage: len(set(hashmat(p) for p in Permutations(3))) 6 sage: len(set(hashmat(p) for p in Permutations(4))) 24 sage: len(set(hashmat(p) for p in Permutations(5))) 120 sage: len(set(hashmat(p) for p in Permutations(6))) 720 sage: len(set(hashmat(p) for p in Permutations(7))) 5040 sage: len(set(hashmat(p) for p in Permutations(8))) 40320
I stumbled on this when profiling some code using Weyl groups that heavily used caching (the hash of a weyl group element is the hash of the underlying matrix). I gained a speed factor of 10x by just tweaking the hash of matrices as in the attached patch. Now, I have no idea if in general that would be a good hash for matrices, so please some expert write an appropriate patch.
Cheers,
Nicolas
Change History (21)
Changed 10 years ago by
comment:1 Changed 10 years ago by
 Description modified (diff)
 Summary changed from The Hash function for matrices has many collisions with permutation matrices: to The Hash function for matrices suffers from many collisions with permutation matrices
comment:2 Changed 10 years ago by
 Summary changed from The Hash function for matrices suffers from many collisions with permutation matrices to The hash function for matrices suffers from many collisions with permutation matrices
comment:3 followup: ↓ 4 Changed 10 years ago by
comment:4 in reply to: ↑ 3 Changed 10 years ago by
 Keywords permutation matrices hash added
Replying to jhpalmieri:
Note that in
m = p.to_matrix()
, the matrix m is sparse, not dense (as far as I can tell), so you'll need to patch both matrix_dense.pyx and matrix_sparse.pyx.
Oops. I changed my examples at the last minute, thinking that permutation matrices would feel less exotic than Weyl groups, but I forgot to actually check that my patch fixed the problem for those. Thanks for pointing this out!
comment:5 Changed 10 years ago by
 Owner changed from jason, was to nthiery
Oops again ... Please ignore the second attachment; it's a duplicate of the first. Anyone with power: feel free to delete
comment:6 Changed 5 years ago by
This is still true. I'd say it's just true of "dense" collections of matrices:
def imm(A): A.set_immutable() return A
sage: V={imm(matrix([a,b,c,d])) for a in [1..1] for b in [1..1] for c in [1..1] for d in [1..1]} sage: H={hash(v) for v in V} sage: Ht={hash(tuple(v.list())) for v in V} sage: len(V); len (H); len(Ht) 81 14 69
so I think this problem might need a little bump in priority.
comment:7 Changed 3 years ago by
I opened #19050 while not knowing about this one! Funny numbers.
It would be easy to design a robust hash for dense matrices. But we do want to have the same hash for sparse and dense matrices. As a consequence the hash should just be computed from the set {(i,j,v)
} of nonzero values v
at position (i,j)
. We could do a reasonable hash by asking this list to be sorted lexicographically in (i,j)
(that would cost a log(num of entries)
).
For polynomials this is exactly the same setting with the list (exponent, coefficient)
. See #21284.
The question: for sparse matrices (respectively sparse or multivariate polynomials), would it be reasonable to sort the indices in the hash function?
comment:8 followup: ↓ 13 Changed 3 years ago by
This may be a silly question, but why does this happen?
sage: hash((0, 1, 1, 1)) == hash((0, 1, 0, 0)) True
comment:9 Changed 3 years ago by
How about just doing a frozenset
instead of sorting the tuples?
comment:10 followup: ↓ 11 Changed 3 years ago by
@tscrim: using frozenset
, the hash function would be less robust and, more importantly, would be very hard to optimize in specialized classes. You don't want the construction of a frozenset each time __hash__
is called (eg for dense matrices or univariate polynomial the list of exponents naturally come in order).
comment:11 in reply to: ↑ 10 ; followup: ↓ 14 Changed 3 years ago by
Replying to vdelecroix:
@tscrim: using
frozenset
, the hash function would be less robust
I don't see how, but maybe I don't understand what you think it will be less robust with respect to.
and, more importantly, would be very hard to optimize in specialized classes.
I don't see why you would need to optimize the hash function. The bottleneck would be iterating over the nonzero values, which is (would be?) a separate method.
You don't want the construction of a frozenset each time
__hash__
is called (eg for dense matrices or univariate polynomial the list of exponents naturally come in order).
Do you really want to sort a list of triples every time for sparse matrices? Inserting n objects into a hash table is much faster than that: O(n) vs O(n log n) for sorting them. We might want to consider (continuing) to cache the __hash__
as well.
I also don't see why polynomials should come into an argument here.
However, here is another thought for a hash function:
temp = [0,0,0] for i,j in self.nonzero_positions(): temp[0] += i temp[1] += j temp[2] += self[i,j] return hash(tuple(temp))
comment:12 followup: ↓ 17 Changed 3 years ago by
First of all, the hash function is time critical and caching a hash is useless. You are almost never trying to test A = A
(are you?). At least, upstream Python decided not to cache it for tuples and frozensets based on concrete benchmarking.
Let me consider a first concrete situation. I have 2 invertible matrices 3x3 matrices and I want to check whether there is a multiplicative relation among them. It is likely that I need to consider products of size 20, that is put in a set around 3^{20} matrices. Building a frozenset for each of them would be a considerable waste and caching would not help.
A situation for which caching might be useful is when you have a "combinatorial" free module with a basis made of matrices and you encode your vectors using your matrices as keys. Though, that might be a problem of datastructure here (the keys are completely static).
Do you have more concrete situations where hashing is involved?
For sorting the indices, I believe that in most situations hashing will be used on dense matrices. For them, you can go through the (i,j)
indices in any order you like.
Concerning collisions, I don't believe your hash function was a serious proposal. Let us consider n x n
matrices with entries {0,1}
. Let N = n^{2}. You know that you have 2^{N} such matrices and it is easy to see that you have a polynomial number of values produced by your hash function (around N^{2} N^{2} N = N^{5} if I am optimistic). Even on permutation matrices, your hash function would perform terribly.
The relevance of polynomials is that you have the same kind of troubles passing from univariate (coefficients in a list like structure) to multivariate (coefficients in a dictionary like structure). See #21284.
comment:13 in reply to: ↑ 8 Changed 3 years ago by
Replying to jhpalmieri:
This may be a silly question, but why does this happen?
sage: hash((0, 1, 1, 1)) == hash((0, 1, 0, 0)) True
Interesting. This is how Python implements hashing for tuples:
sage: def tuplehash(t): ....: n = len(t) ....: x = 3430008 ....: for i in range(n): ....: h = hash(t[i]) ....: x = (x ^^ h) * (1000003 + (82520 + 2*n)*i) ....: return x + 97531
The resulting value should be reduced mod 2^{32} or 2^{64} depending on the system.
The reason for the collision is that N ^^ 2 == N
for any odd N
. The choice of a XOR operation there is strange, since most simple hash functions would use addition instead. See https://en.wikipedia.org/wiki/Universal_hashing#Constructions for example.
comment:14 in reply to: ↑ 11 Changed 3 years ago by
Replying to tscrim:
However, here is another thought for a hash function:
temp = [0,0,0] for i,j in self.nonzero_positions(): temp[0] += i temp[1] += j temp[2] += self[i,j] return hash(tuple(temp))
I like the idea but not the concrete implementation. It does avoid sorting but there needs to be more mixing. Something like
h = 0 for i,j in self.nonzero_positions(): h += H(i, j, self[i,j]) return h
where H
is a simple hash function.
comment:16 Changed 3 years ago by
 Dependencies set to #24090
comment:17 in reply to: ↑ 12 Changed 3 years ago by
Replying to vdelecroix:
First of all, the hash function is time critical and caching a hash is useless. You are almost never trying to test
A = A
(are you?). At least, upstream Python decided not to cache it for tuples and frozensets based on concrete benchmarking.Let me consider a first concrete situation. I have 2 invertible matrices 3x3 matrices and I want to check whether there is a multiplicative relation among them. It is likely that I need to consider products of size 20, that is put in a set around 3^{20} matrices. Building a frozenset for each of them would be a considerable waste and caching would not help.
It depends on what you mean by "considerable waste." In terms of computational complexity, it just adds to the constant (or probably would not even be contributing asymptotically for dense matrices). However, I do agree that it effectively doubles the memory footprint for dense matrices.
A situation for which caching might be useful is when you have a "combinatorial" free module with a basis made of matrices and you encode your vectors using your matrices as keys. Though, that might be a problem of datastructure here (the keys are completely static).
Do you have more concrete situations where hashing is involved?
As you mentioned, when they are used for keys of a module implemented as a dict
, such for CombinatorialFreeModule
elements, or for defining objects such as FiniteDimensionalAlgebra
. Granted, in these situations the matrices are relatively small.
For sorting the indices, I believe that in most situations hashing will be used on dense matrices. For them, you can go through the
(i,j)
indices in any order you like.
I have been doing a bunch of stuff with sparse matrices recently (computing representations of Lie algebras) and have encountered things that are optimized for dense matrices that are used by sparse matrices as well. So I'd appreciate it if we didn't make the hash worse for sparse matrices just because it is better of dense matrices.
Concerning collisions, I don't believe your hash function was a serious proposal.
It was just a thought on how to remove the dependence on the order and not fully fleshed out. I know I'm less experienced in this area, but I am legitimately trying to help.
Let us consider
n x n
matrices with entries{0,1}
. Let N = n^{2}. You know that you have 2^{N} such matrices and it is easy to see that you have a polynomial number of values produced by your hash function (around N^{2} N^{2} N = N^{5} if I am optimistic). Even on permutation matrices, your hash function would perform terribly.
I agree that it would not perform well in this case. What do you think about Jeoren's modification. I think it works well for both dense and sparse matrices and should work better than my proposal.
The relevance of polynomials is that you have the same kind of troubles passing from univariate (coefficients in a list like structure) to multivariate (coefficients in a dictionary like structure). See #21284.
For the multivariate polynomials, it looks like what is used is Jeroen's proposal. Although we could have a completely different solution for polynomials because the data structure for dense univariate is 1D and sparse kvariate are kD, whereas matrices are strictly 2D.
comment:18 Changed 3 years ago by
 Branch set to u/jdemeyer/the_hash_function_for_matrices_suffers_from_many_collisions_with_permutation_matrices
comment:19 Changed 3 years ago by
 Commit set to f67a8839d7dfb843a8ed8652a2734b644be579f1
Branch pushed to git repo; I updated commit sha1. New commits:
f67a883  Fix doctests

comment:20 Changed 3 years ago by
 Description modified (diff)
Note that in
m = p.to_matrix()
, the matrix m is sparse, not dense (as far as I can tell), so you'll need to patch both matrix_dense.pyx and matrix_sparse.pyx.