Fastest way to solve non-negative linear diophantine equations

215 views Asked by At

Let A be a list of n lists of m non-negative integers, such that for all j there is i with A[i][j] nonzero. Let V be a list of m positive integers.

Question: What is the fastest way to find all the lists X of n non-negative integers such that for all i then sum_j A[i][j] X[j] = V[i]?

The assumptions implies that the number of solutions is finite. See below an example of A and V, with 5499 solutions X.

Let me reformulate the problem using matrix and vector. Let A be a n-by-m matrix with non-negative integral entries and without zero column. Let V be a vector with positive integral entries. What is the fastest way to find all the vectors X, with non-negative integral entries, such that AX=V?

The usual functions for solving such a system may underuse the non-negativity. To prove so, I wrote a brute-force code finding all the solutions of such a system and applied it to an example (see below, and these crossposts on mathoverflow and on ask.sagemath), but I'm still looking for something significantly faster than this; in fact I'm looking for the fastest way.


Example

Here is the kind of system I need to solve (where x_i is non-negative integral), but with possibly more equations and variables.

[
5*x0 + 5*x1 + 5*x2 + 6*x3 + 7*x4 + 7*x5 == 24,
5*x1 + 7*x10 + 5*x6 + 5*x7 + 6*x8 + 7*x9 == 25,
5*x11 + 6*x12 + 7*x13 + 7*x14 + 5*x2 + 5*x7 == 25,
5*x12 + 6*x15 + 7*x16 + 7*x17 + 5*x3 + 5*x8 == 30,
5*x13 + 6*x16 + 7*x18 + 7*x19 + 5*x4 + 5*x9 == 35,
5*x10 + 5*x14 + 6*x17 + 7*x19 + 7*x20 + 5*x5 == 35,
5*x1 + 7*x10 + 5*x6 + 5*x7 + 6*x8 + 7*x9 == 25,
5*x21 + 5*x22 + 6*x23 + 7*x24 + 7*x25 + 5*x6 == 24,
5*x22 + 5*x26 + 6*x27 + 7*x28 + 7*x29 + 5*x7 == 25,
5*x23 + 5*x27 + 6*x30 + 7*x31 + 7*x32 + 5*x8 == 30,
5*x24 + 5*x28 + 6*x31 + 7*x33 + 7*x34 + 5*x9 == 35,
5*x10 + 5*x25 + 5*x29 + 6*x32 + 7*x34 + 7*x35 == 35,
5*x11 + 6*x12 + 7*x13 + 7*x14 + 5*x2 + 5*x7 == 25,
5*x22 + 5*x26 + 6*x27 + 7*x28 + 7*x29 + 5*x7 == 25,
5*x11 + 5*x26 + 5*x36 + 6*x37 + 7*x38 + 7*x39 == 24,
5*x12 + 5*x27 + 5*x37 + 6*x40 + 7*x41 + 7*x42 == 30,
5*x13 + 5*x28 + 5*x38 + 6*x41 + 7*x43 + 7*x44 == 35,
5*x14 + 5*x29 + 5*x39 + 6*x42 + 7*x44 + 7*x45 == 35,
5*x12 + 6*x15 + 7*x16 + 7*x17 + 5*x3 + 5*x8 == 30,
5*x23 + 5*x27 + 6*x30 + 7*x31 + 7*x32 + 5*x8 == 30,
5*x12 + 5*x27 + 5*x37 + 6*x40 + 7*x41 + 7*x42 == 30,
5*x15 + 5*x30 + 5*x40 + 6*x46 + 7*x47 + 7*x48 == 35,
5*x16 + 5*x31 + 5*x41 + 6*x47 + 7*x49 + 7*x50 == 42,
5*x17 + 5*x32 + 5*x42 + 6*x48 + 7*x50 + 7*x51 == 42,
5*x13 + 6*x16 + 7*x18 + 7*x19 + 5*x4 + 5*x9 == 35,
5*x24 + 5*x28 + 6*x31 + 7*x33 + 7*x34 + 5*x9 == 35,
5*x13 + 5*x28 + 5*x38 + 6*x41 + 7*x43 + 7*x44 == 35,
5*x16 + 5*x31 + 5*x41 + 6*x47 + 7*x49 + 7*x50 == 42,
5*x18 + 5*x33 + 5*x43 + 6*x49 + 7*x52 + 7*x53 == 48,
5*x19 + 5*x34 + 5*x44 + 6*x50 + 7*x53 + 7*x54 == 49,
5*x10 + 5*x14 + 6*x17 + 7*x19 + 7*x20 + 5*x5 == 35,
5*x10 + 5*x25 + 5*x29 + 6*x32 + 7*x34 + 7*x35 == 35,
5*x14 + 5*x29 + 5*x39 + 6*x42 + 7*x44 + 7*x45 == 35,
5*x17 + 5*x32 + 5*x42 + 6*x48 + 7*x50 + 7*x51 == 42,
5*x19 + 5*x34 + 5*x44 + 6*x50 + 7*x53 + 7*x54 == 49,
5*x20 + 5*x35 + 5*x45 + 6*x51 + 7*x54 + 7*x55 == 48
]

Here are explicit A and V from above system (in list form):

A=[  
[5,5,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,5,0,0,0,0,5,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,5,0,0,0,0,5,0,0,0,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,5,0,0,0,0,5,0,0,0,5,0,0,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,5,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,5,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,5,0,0,0,0,5,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,0,0,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,5,0,0,0,0,5,0,0,0,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0],
[0,0,0,5,0,0,0,0,5,0,0,0,5,0,0,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,0,0,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,6,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,7,7,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,7,7,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,7,7,0,0,0,0],
[0,0,0,0,5,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,7,7,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,0,7,7,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,0,7,7,0],
[0,0,0,0,0,5,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,6,0,7,7,0,0,0,0,0,0,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,7,7,0,0,0,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,0,7,7,0],
[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,0,0,0,0,5,0,0,0,0,0,6,0,0,7,7]  
]

V=[24,25,25,30,35,35,25,24,25,30,35,35,25,25,24,30,35,35,30,30,30,35,42,42,35,35,35,42,48,49,35,35,35,42,49,48]

Computation

I wrote a brute-force code finding all the solutions of such a system, then applied it to A, V above. It took 12 seconds to find all 5499 solutions. I'm looking for something significantly faster than this.

sage: %time LX=NonnegativeSolverPartition(A,V)
CPU times: user 11.8 s, sys: 0 ns, total: 11.8 s
Wall time: 11.8 s
sage: len(LX)
5499

Note that the time reduces to 3 seconds with PyPy3, but it is still too slow for other (bigger) such systems I need to solve.


Code

Here is my (python) code, improved by Peter Taylor (see his comment):

def NonnegativeSolverPartition(A,V):
    WB = WeakBasis(A)
    VB = VarBound(A,V)
    PP = []

    for i, ll in WB:
        L = tuple(A[i][j] for j in ll)
        B = tuple(VB[j] for j in ll)
        PP.append(WeightedPartitionSolver(L, B, V[i]))

    return list(NonnegativeSolverPartitionInter(A, V, PP, WB, [-1] * len(A[0])))


def NonnegativeSolverPartitionInter(A, V, PP, WB, X):
    if any(len(P) > 1 for P in PP):
        _, ii = min((len(P), i) for i, P in enumerate(PP) if len(P) > 1)
        for p in PP[ii]:
            PPP = PP[:ii] + [[p]] + PP[ii+1:]
            Fi = Filter(PPP, list(X), WB) 
            if Fi:
                PPPP, XX = Fi
                yield from NonnegativeSolverPartitionInter(A, V, PPPP, WB, XX)
    else:
        assert -1 not in X
        yield X


def WeakBasis(A):
    return tuple(enumerate([j for j, tmp in enumerate(row) if tmp] for row in A))


def WeightedPartitions(ws, n):
    def inner(ws, n):
        if n == 0:
            yield (0,) * len(ws)
        elif ws:
            w = ws[0]
            lim = n // w
            ws = ws[1:]
            for i in range(lim + 1):
                for tl in inner(ws, n - w * i):
                    yield (i,) + tl

    return list(inner(ws, n))


def VarBound(A,V):
    nvars = len(A[0])

    # Solve the individual constraints and then intersect the solutions.
    possible_values = [None] * nvars
    row_solns = []
    for row, v in zip(A, V):
        lut = []
        ws = []
        var_assignments = []
        for j, val in enumerate(row):
            if val:
                lut.append(j)
                ws.append(val)
                var_assignments.append(set())

        for soln in WeightedPartitions(ws, v):
            for i, w in enumerate(soln):
                var_assignments[i].add(w)

        for j, assignments in zip(lut, var_assignments):
            if possible_values[j] is None:
                possible_values[j] = assignments
            else:
                possible_values[j] &= assignments

    return tuple(frozenset(x) for x in possible_values)


def WeightedPartitionSolver(L, B, n):
    # the entries of L must be non-negative
    # B gives valid values coming from other equations (see VarBound)
    def inner(L, B, n):
        if n == 0:
            yield (0,) * len(L)
        elif L:
            w, allowed = L[0], B[0]
            L, B = L[1:], B[1:]
            for i in range(n // w + 1):
                if i in allowed:
                    for tl in inner(L, B, n - w * i):
                        yield (i,) + tl

    return list(inner(L, B, n))


def Filter(PP, X, W):
    if [] in PP:
        return None

    while True:
        for Wi, P in zip(W, PP):
            F = FixedPoints(P)
            for j in F:
                P0j = P[0][j]
                Wij = Wi[1][j]
                if X[Wij] == -1:
                    X[Wij] = P0j
                elif X[Wij] != P0j:
                    return None

        LL=[]
        for Wi, P in zip(W, PP):
            LL.append([p for p in P if not any(X[idx] not in (-1, pval) for idx, pval in zip(Wi[1], p))])
            if not LL[-1]:
                return None

        if PP == LL:
            return LL, X

        PP = LL


def FixedPoints(P):
    # This would prefer P to be transposed
    m=len(P)
    n=len(P[0])
    return tuple(i for i in range(n) if all(P[j][i] == P[0][i] for j in range(m)))

A simpler brute-force code by Max Alekseyev is available in this answer.

0

There are 0 answers