Calculate the Hafnian as quickly as possible

Question

The challenge is to write the fastest code possible for computing the Hafnian of a matrix.

The Hafnian of a symmetric 2n-by-2n matrix A is defined as:

Here S_2n represents the set of all permutations of the integers from 1 to 2n, that is [1, 2n].

The wikipedia link also gives a different looking formula which may be of interest (and even faster methods exist if you look further on the web). The same wiki page talks about adjacency matrices but your code should work for other matrices as well. You can assume the values will all be integers but not that that they are all positive.

There is also a faster algorithm but it seems hard to understand. and Christian Sievers was the first to implement it (in Haskell).

In this question matrices are all square and symmetric with even dimension.

Reference implementation (note this is using the slowest possible method).

Here is some example python code from Mr. Xcoder.

from itertools import permutations
from math import factorial

def hafnian(matrix):
    my_sum = 0
    n = len(matrix) // 2
    for sigma in permutations(range(n*2)):
        prod = 1
        for j in range(n):
            prod *= matrix[sigma[2*j]][sigma[2*j+1]]
        my_sum += prod
    return my_sum / (factorial(n) * 2 ** n)

print(hafnian([[-1, 1, 1, -1, 0, 0, 1, -1], [1, 0, 1, 0, -1, 0, -1, -1], [1, 1, -1, 1, -1, -1, 0, -1], [-1, 0, 1, -1, -1, 1, -1, 0], [0, -1, -1, -1, -1, 0, 0, -1], [0, 0, -1, 1, 0, 0, 1, 1], [1, -1, 0, -1, 0, 1, 1, 0], [-1, -1, -1, 0, -1, 1, 0, 1]]))
4

M = [[1, 1, 0, 0, 0, 0, 0, 1, 0, 0], [1, 1, -1, 0, -1, 1, 1, 1, 0, -1], [0, -1, -1, -1, 0, -1, -1, 0, -1, 1], [0, 0, -1, 1, -1, 1, -1, 0, 1, -1], [0, -1, 0, -1, -1, -1, -1, 1, -1, 1], [0, 1, -1, 1, -1, 1, -1, -1, 1, -1], [0, 1, -1, -1, -1, -1, 1, 0, 0, 0], [1, 1, 0, 0, 1, -1, 0, 1, 1, -1], [0, 0, -1, 1, -1, 1, 0, 1, 1, 1], [0, -1, 1, -1, 1, -1, 0, -1, 1, 1]]

print(hafnian(M))
-13

M = [[-1, 0, -1, -1, 0, -1, 0, 1, -1, 0, 0, 0], [0, 0, 0, 0, 0, -1, 0, 1, -1, -1, -1, -1], [-1, 0, 0, 1, 0, 0, 0, 1, -1, 1, -1, 0], [-1, 0, 1, -1, 1, -1, -1, -1, 0, -1, -1, -1], [0, 0, 0, 1, 0, 0, 0, 0, 0, 1, -1, 0], [-1, -1, 0, -1, 0, 0, 1, 1, 1, 1, 1, 0], [0, 0, 0, -1, 0, 1, 1, -1, -1, 0, 1, 0], [1, 1, 1, -1, 0, 1, -1, 1, -1, -1, -1, -1], [-1, -1, -1, 0, 0, 1, -1, -1, -1, 1, -1, 0], [0, -1, 1, -1, 1, 1, 0, -1, 1, -1, 1, 1], [0, -1, -1, -1, -1, 1, 1, -1, -1, 1, 0, -1], [0, -1, 0, -1, 0, 0, 0, -1, 0, 1, -1, 1]]

print(hafnian(M))
13

M = [[-1, 1, 0, 1, 0, -1, 0, 0, -1, 1, -1, 1, 0, -1], [1, -1, 1, -1, 1, 1, -1, 0, -1, 1, 1, 0, 0, -1], [0, 1, 1, 1, -1, 1, -1, -1, 0, 0, -1, 0, -1, -1], [1, -1, 1, -1, 1, 0, 1, 1, -1, -1, 0, 0, 1, 1], [0, 1, -1, 1, 0, 1, 0, 1, -1, -1, 1, 1, 0, -1], [-1, 1, 1, 0, 1, 1, -1, 0, 1, -1, -1, -1, 1, -1], [0, -1, -1, 1, 0, -1, -1, -1, 0, 1, -1, 0, 1, -1], [0, 0, -1, 1, 1, 0, -1, 0, 0, -1, 0, 0, 0, 1], [-1, -1, 0, -1, -1, 1, 0, 0, 1, 1, 0, 1, -1, 0], [1, 1, 0, -1, -1, -1, 1, -1, 1, 1, 1, 0, 1, 0], [-1, 1, -1, 0, 1, -1, -1, 0, 0, 1, -1, 0, -1, 0], [1, 0, 0, 0, 1, -1, 0, 0, 1, 0, 0, 1, 1, 1], [0, 0, -1, 1, 0, 1, 1, 0, -1, 1, -1, 1, 1, -1], [-1, -1, -1, 1, -1, -1, -1, 1, 0, 0, 0, 1, -1, -1]]

print(hafnian(M))
83

The task

You should write code that, given an 2n by 2n matrix, outputs its Hafnian.

As I will need to test your code it would be helpful if you could give a simple way for me to give a matrix as input to your code, for example by reading from standard in. I will test your code in randomly chosen matrices with elements selected from {-1, 0, 1}. The purpose of testing like this is to reduce the chance the Hafnian will be a very large value.

Ideally your code will be able to read in matrices exactly as I have them in the examples in this question straight from standard in. That is the input would look like [[1,-1],[-1,-1]] for example. If you want to use another input format, please ask and I will do my best to accommodate.

Scores and ties

I will test your code on random matrices of increasing size and stop the first time your code takes more than 1 minute on my computer. The scoring matrices will be consistent for all submissions in order to ensure fairness.

If two people get the same score then the winner is the one which is fastest for that value of n. If those are within 1 second of each other then it is the one posted first.

Languages and libraries

You can use any available language and libraries you like but no pre-existing function to compute the Hafnian. Where feasible, it would be good to be able to run your code so please include a full explanation for how to run/compile your code in Linux if at all possible.`

My Machine The timings will be run on my 64-bit machine. This is a standard ubuntu install with 8GB RAM, AMD FX-8350 Eight-Core Processor and Radeon HD 4250. This also means I need to be able to run your code.

Call for answers in more languages

It would be great to get answers in your favorite super fast programming language. To start things off, how about fortran, nim and rust?

Leaderboard

• 52 by miles using C++. 30 seconds.
• 50 by ngn using C. 50 seconds.
• 46 by Christian Sievers using Haskell. 40 seconds.
• 40 by miles using Python 2 + pypy. 41 seconds.
• 34 by ngn using Python 3 + pypy. 29 seconds.
• 28 by Dennis using Python 3. 35 seconds. (Pypy is slower)

Answer 1

Haskell

import Control.Parallel.Strategies
import qualified Data.Vector.Unboxed as V
import qualified Data.Vector as VB

type Poly = V.Vector Int

type Matrix = VB.Vector ( VB.Vector Poly )

constpoly :: Int -> Int -> Poly
constpoly n c = V.generate (n+1) (\i -> if i==0 then c else 0)

add :: Poly -> Poly -> Poly
add = V.zipWith (+)

shiftmult :: Poly -> Poly -> Poly
shiftmult a b = V.generate (V.length a) 
                           (\i -> sum [ a!j * b!(i-1-j) | j<-[0..i-1] ])
  where (!) = V.unsafeIndex

x :: Matrix -> Int -> Int -> Int -> Poly -> Int
x  _    0  _ m p = m * V.last p
x mat n c m p =
  let mat' = VB.generate (2*n-2) $ \i ->
             VB.generate i       $ \j ->
                 shiftmult (mat!(2*n-1)!i) (mat!(2*n-2)!j) `add`
                 shiftmult (mat!(2*n-1)!j) (mat!(2*n-2)!i) `add`
                 (mat!i!j)
      p' = p `add` shiftmult (mat!(2*n-1)!(2*n-2)) p
      (!) = VB.unsafeIndex
      r = if c>0 then parTuple2 rseq rseq else r0
      (a,b) = (x mat (n-1) (c-1) m p, x mat' (n-1) (c-1) (-m) p')
              `using` r
  in a+b

haf :: [[Int]] -> Int
haf m = let n=length m `div` 2
        in x (VB.fromList $ map (VB.fromList . map (constpoly n)) m) 
             n  5  ((-1)^n)  (constpoly n 1) 

main = getContents >>= print . haf . read

This implements a variation of Algorithm 2 of Andreas Björklund: Counting Perfect Matchings as Fast as Ryser.

Compile using ghc with compile time options -O3 -threaded and use run time options +RTS -N for parallelization. Takes input from stdin.

Answer 2

Python 3

from functools import lru_cache

@lru_cache(maxsize = None)
def haf(matrix):
	n = len(matrix)
	if n == 2: return matrix[0][1]
	h = 0
	for j in range(1, n):
		if matrix[0][j] == 0: continue
		copy = list(matrix)
		del copy[:j+1:j]
		copy = list(zip(*copy))
		del copy[:j+1:j]
		h += matrix[0][j] * haf(tuple(copy))
	return h

print(haf(tuple(map(tuple, eval(open(0).read())))))

Try it online!

Answer 3

C++ (gcc)

#define T(x) ((x)*((x)-1)/2)
#define S 1
#define J (1<<S)
#define TYPE int

#include <iostream>
#include <vector>
#include <string>
#include <pthread.h>

using namespace std;

struct H {
    int s, w, t;
    TYPE *b, *g;
};

void *solve(void *a);
void hafnian(TYPE *b, int s, TYPE *g, int w, int t);

int n, m, ti = 0;
TYPE r[J] = {0};
pthread_t pool[J];

int main(void) {
    vector<int> a;
    string s;
    getline(cin, s);

    for (int i = 0; i < s.size(); i++)
        if (s[i] == '0' || s[i] == '1')
            a.push_back((s[i-1] == '-' ? -1 : 1)*(s[i] - '0'));

    for (n = 1; 4*n*n < a.size(); n++);
    m = n+1;

    TYPE z[T(2*n)*m] = {0}, g[m] = {0};

    for (int j = 1; j < 2*n; j++)
        for (int k = 0; k < j; k++)
            z[(T(j)+k)*m] = a[j*2*n+k];
    g[0] = 1;

    hafnian(z, 2*n, g, 1, -1);

    TYPE h = 0;
    for (int t = 0; t < ti; t++) {
        pthread_join(pool[t], NULL);
        h += r[t];
    }

    cout << h << endl;

    return 0;
}

void *solve(void *a) {
    H *p = reinterpret_cast<H*>(a);
    hafnian(p->b, p->s, p->g, p->w, p->t);
    delete[] p->b;
    delete[] p->g;
    delete p;
    return NULL;
}

void hafnian(TYPE *b, int s, TYPE *g, int w, int t) {
    if (t == -1 && (n < S || s/2 == n-S)) {
        H *p = new H;
        TYPE *c = new TYPE[T(s)*m], *e = new TYPE[m];
        copy(b, b+T(s)*m, c);
        copy(g, g+m, e);
        p->b = c;
        p->s = s;
        p->g = e;
        p->w = w;
        p->t = ti;
        pthread_create(pool+ti, NULL, solve, p);
        ti++;
    }
    else if (s > 0) {
        TYPE c[T(s-2)*m], e[m];
        copy(b, b+T(s-2)*m, c);
        hafnian(c, s-2, g, -w, t);
        copy(g, g+m, e);

        for (int u = 0; u < n; u++) {
            TYPE *d = e+u+1,
                  p = g[u], *x = b+(T(s)-1)*m;
            for (int v = 0; v < n-u; v++)
                d[v] += p*x[v];
        }

        for (int j = 1; j < s-2; j++)
            for (int k = 0; k < j; k++)
                for (int u = 0; u < n; u++) {
                    TYPE *d = c+(T(j)+k)*m+u+1,
                          p = b[(T(s-2)+j)*m+u], *x = b+(T(s-1)+k)*m,
                          q = b[(T(s-2)+k)*m+u], *y = b+(T(s-1)+j)*m;
                    for (int v = 0; v < n-u; v++)
                        d[v] += p*x[v] + q*y[v];
                }

        hafnian(c, s-2, e, w, t);
    }
    else
        r[t] += w*g[n];
}

Try it online! (13s for n = 24)

Based on the faster Python implementation in my other post. Edit the second line to #define S 3 on your 8-core machine and compile with g++ -pthread -march=native -O2 -ftree-vectorize.

The splits the work in half, so the value of S should be log2(#threads). The types can easily be changed between int, long, float, and double by modifying the value of #define TYPE.

Answer 4

Python 3

This computes haf(A) as a memoised sum(A[i][j] * haf(A without rows and cols i and j)).

#!/usr/bin/env python3
import json,sys
a=json.loads(sys.stdin.read())
n=len(a)//2
b={0:1}
def haf(x):
 if x not in b:
  i=0
  while not x&(1<<i):i+=1
  x1=x&~(1<<i)
  b[x]=sum(a[i][j]*haf(x1&~(1<<j))for j in range(2*n)if x1&(1<<j)and a[i][j])
 return b[x]
print(haf((1<<2*n)-1))

Answer 5

C

Another impl of Andreas Björklund's paper, which is much easier to understand if you also look at Christian Sievers's Haskell code. For the first few levels of the recursion, it distributes round-robin threads over available CPUs. The last level of the recursion, which accounts for half of the invocations, is optimised by hand.

Compile with: gcc -O3 -pthread -march=native; thanks @Dennis for a 2x speed-up

n=24 in 24s on TIO

#define _GNU_SOURCE
#include<sched.h>
#include<stdio.h>
#include<stdlib.h>
#include<memory.h>
#include<unistd.h>
#include<pthread.h>
#define W while
#define R return
#define S static
#define U (1<<31)
#define T(i)((i)*((i)-1)/2)
typedef int I;typedef long L;typedef char C;typedef void V;
I n,ncpu,icpu;
S V f(I*x,I*y,I*z){I i=n,*z1=z+n;W(i){I s=0,*x2=x,*y2=y+--i;W(y2>=y)s+=*x2++**y2--;*z1--+=s;}}
typedef struct{I m;V*a;V*p;pthread_barrier_t*bar;I r;}A;S V*(h1)(V*);
I h(I m,I a[][n+1],I*p){
 m-=2;I i,j,k=0,u=T(m),v=u+m,b[u][n+1],q[n+1];
 if(!m){I*x=a[v+m],*y=p+n-1,s=0;W(y>=p)s-=*x++**y--;R s;}
 memcpy(b,a,sizeof(b));memcpy(q,p,sizeof(q));f(a[v+m],p,q);
 for(i=1;i<m;i++)for(j=0;j<i;j++){f(a[u+i],a[v+j],b[k]);f(a[u+j],a[v+i],b[k]);k++;}
 if(2*n-m>8)R h(m,a,p)-h(m,b,q);
 pthread_barrier_t bar;pthread_barrier_init(&bar,0,2);pthread_t th;
 cpu_set_t cpus;CPU_ZERO(&cpus);CPU_SET(icpu++%ncpu,&cpus);
 pthread_attr_t attr;pthread_attr_init(&attr);
 pthread_attr_setaffinity_np(&attr,sizeof(cpu_set_t),&cpus);
 A arg={m,a,p,&bar};pthread_create(&th,&attr,h1,&arg);
 I r=h(m,b,q);pthread_barrier_wait(&bar);pthread_join(th,0);pthread_barrier_destroy(&bar);
 R arg.r-r;
}
S V*h1(V*x0){A*x=(A*)x0;x->r=h(x->m,x->a,x->p);pthread_barrier_wait(x->bar);R 0;}
I main(){
 ncpu=sysconf(_SC_NPROCESSORS_ONLN);
 S C s[200000];I i=0,j=0,k,l=0;W((k=read(0,s+l,sizeof(s)-l))>0)l+=k;
 n=1;W(s[i]!=']')n+=s[i++]==',';n/=2;
 I a[T(2*n)][n+1];memset(a,0,sizeof(a));k=0;
 for(i=0;i<2*n;i++)for(j=0;j<2*n;j++){
  W(s[k]!='-'&&(s[k]<'0'||s[k]>'9'))k++;
  I v=0,m=s[k]=='-';k+=m;W(k<l&&('0'<=s[k]&&s[k]<='9'))v=10*v+s[k++]-'0';
  if(i>j)*a[T(i)+j]=v*(1-2*m);
 }
 I p[n+1];memset(p,0,sizeof(p));*p=1;
 printf("%d\n",(1-2*(n&1))*h(2*n,a,p));
 R 0;
}

Algorithm:

The matrix, which is symmetric, is stored in lower-left triangular form. Triangle indices i,j correspond to linear index T(max(i,j))+min(i,j) where T is a macro for i*(i-1)/2. Matrix elements are polynomials of degree n. A polynomial is represented as an array of coefficients ordered from the constant term (p[0]) to xⁿ's coefficient (p[n]). The initial -1,0,1 matrix values are first converted to const polynomials.

We perform a recursive step with two arguments: the half-matrix (i.e. triangle) a of polynomials and a separate polynomial p (referred to as beta in the paper). We reduce the size-m problem (initially m=2*n) recursively to two problems of size m-2 and return the difference of their hafnians. One of them is to use the same a without its last two rows, and the very same p. Another is to use the triangle b[i][j] = a[i][j] + shmul(a[m-1][i],a[m-2][j]) + shmul(a[m-1][j],a[m-2][i]) (where shmul is the shift-multiply operation on polynomials - it's like polynomial product as usual, additionally multiplied by the variable "x"; powers higher than x^n are ignored), and the separate polynomial q = p + shmul(p,a[m-1][m-2]). When recursion hits a size-0 a, we return the major coefficient of p: p[n].

The shift-and-multiply operation is implemented in function f(x,y,z). It modifies z in-place. Loosely speaking, it does z += shmul(x,y). This seems to be the most performance-critical part.

After the recursion has finished, we need to fix the sign of the result by multiplying by (-1)ⁿ.

Answer 6

Python

This is pretty much a straight-forward, reference implementation of Algorithm 2 from the mentioned paper. The only changes were to only keep the current value of B, dropping the values of β by only updating g when i ∈ X, and truncated polynomial multiplication by only calculating the values up to degree n.

from itertools import chain,combinations

def powerset(s):
    return chain.from_iterable(combinations(s, k) for k in range(len(s)+1))

def padd(a, b):
    return [a[i]+b[i] for i in range(len(a))]

def pmul(a, b):
    n = len(a)
    c = [0]*n
    for i in range(n):
        for j in range(n):
            if i+j < n:
                c[i+j] += a[i]*b[j]
    return c

def hafnian(m):
    n = len(m) / 2
    z = [[[c]+[0]*n for c in r] for r in m]
    h = 0
    for x in powerset(range(1, n+1)):
        b = z
        g = [1] + [0]*n
        for i in range(1, n+1):
            if i in x:
                g = pmul(g, [1] + b[0][1][:n])
                b = [[padd(b[j+2][k+2], [0] + padd(pmul(b[0][j+2], b[1][k+2]), pmul(b[0][k+2], b[1][j+2]))[:n]) if j != k else 0 for k in range(2*n-2*i)] for j in range(2*n-2*i)]
            else:
                b = [r[2:] for r in b[2:]]
        h += (-1)**(n - len(x)) * g[n]
    return h

Try it online!

Here is a faster version with some of the easy optimizations.

def hafnian(m):
  n = len(m)/2
  z = [[0]*(n+1) for _ in range(n*(2*n-1))]
  for j in range(1, 2*n):
    for k in range(j):
      z[j*(j-1)/2+k][0] = m[j][k]
  return solve(z, 2*n, 1, [1] + [0]*n, n)

def solve(b, s, w, g, n):
  if s == 0:
    return w*g[n]
  c = [b[(j+1)*(j+2)/2+k+2][:] for j in range(1, s-2) for k in range(j)]
  h = solve(c, s-2, -w, g, n)
  e = g[:]
  for u in range(n):
    for v in range(n-u):
      e[u+v+1] += g[u]*b[0][v]
  for j in range(1, s-2):
    for k in range(j):
      for u in range(n):
        for v in range(n-u):
          c[j*(j-1)/2+k][u+v+1] += b[(j+1)*(j+2)/2][u]*b[(k+1)*(k+2)/2+1][v] + b[(k+1)*(k+2)/2][u]*b[(j+1)*(j+2)/2+1][v]
  return h + solve(c, s-2, w, e, n)

Try it online!

For added fun, here is a reference implementation in J.

Answer 7

Octave

This is basically a copy of Dennis' entry, but optimized for Octave. The main optimization is done by using the full input matrix (and its transpose) and recursion using only matrix indices, rather than creating reduced matrices.

The main advantage is reduced copying of matrices. While Octave does not have a difference between pointers/references and values and functionally only does pass-by-value, it's a different story behind the scenes. There, copy-on-write (lazy copy) is used. That means, for the code a=1;b=a;b=b+1, the variable b is only copied to a new location at the last statement, when it is changed. Since matin and matransp are never changed, they will never by copied. The disadvantage is that the function spends more time calculating the correct indices. I may have to try different variations between numerical and logical indices to optimize this.

Important note: input matrix should be int32! Save the function in a file called haf.m

function h=haf(matin,indices,matransp,transp)

    if nargin-4
        indices=int32(1:length(matin));
        matransp=matin';
        transp=false;
    end
    if(transp)
        matrix=matransp;
    else
        matrix=matin;
    end
    ind1=indices(1);
    n=length(indices);
    if n==2
        h=matrix(ind1,indices(2));
        return
    end
    h=0*matrix(1); 
    for j=1:(n-1)
        indj=indices(j+1);
        k=matrix(ind1,indj);
        if logical(k)
            indicestemp=true(n,1);
            indicestemp(1:j:j+1)=false;
            h=h+k.*haf(matin,indices(indicestemp),matransp,~transp);
        end
    end
end

Example test script:

matrix = int32([0 0 1 -1 1 0 -1 -1 -1 0 -1 1 0 1 1 0 0 1 0 0 1 0 1 1;0 0 1 0 0 -1 -1 -1 -1 0 1 1 1 1 0 -1 -1 0 0 1 1 -1 0 0;-1 -1 0 1 0 1 -1 1 -1 1 0 0 1 -1 0 0 0 -1 0 -1 1 0 0 0;1 0 -1 0 1 1 0 1 1 0 0 0 1 0 0 0 1 -1 -1 -1 -1 1 0 -1;-1 0 0 -1 0 0 1 -1 0 1 -1 -1 -1 1 1 0 1 1 1 0 -1 1 -1 -1;0 1 -1 -1 0 0 1 -1 -1 -1 0 -1 1 0 0 0 -1 0 0 1 0 0 0 -1;1 1 1 0 -1 -1 0 -1 -1 0 1 1 -1 0 1 -1 0 0 1 -1 0 0 0 -1;1 1 -1 -1 1 1 1 0 0 1 0 1 0 0 0 0 1 0 1 0 -1 1 0 0;1 1 1 -1 0 1 1 0 0 -1 1 -1 1 1 1 0 -1 -1 -1 -1 0 1 1 -1;0 0 -1 0 -1 1 0 -1 1 0 1 0 0 0 0 0 1 -1 0 0 0 1 -1 -1;1 -1 0 0 1 0 -1 0 -1 -1 0 0 1 0 0 -1 0 -1 -1 -1 -1 -1 1 -1;-1 -1 0 0 1 1 -1 -1 1 0 0 0 -1 0 0 -1 0 -1 -1 0 1 -1 0 0;0 -1 -1 -1 1 -1 1 0 -1 0 -1 1 0 1 -1 -1 1 -1 1 0 1 -1 1 -1;-1 -1 1 0 -1 0 0 0 -1 0 0 0 -1 0 0 -1 1 -1 -1 0 1 0 -1 -1;-1 0 0 0 -1 0 -1 0 -1 0 0 0 1 0 0 1 1 1 1 -1 -1 0 -1 -1;0 1 0 0 0 0 1 0 0 0 1 1 1 1 -1 0 0 1 -1 -1 -1 0 -1 -1;0 1 0 -1 -1 1 0 -1 1 -1 0 0 -1 -1 -1 0 0 -1 1 0 0 -1 -1 1;-1 0 1 1 -1 0 0 0 1 1 1 1 1 1 -1 -1 1 0 1 1 -1 -1 -1 1;0 0 0 1 -1 0 -1 -1 1 0 1 1 -1 1 -1 1 -1 -1 0 1 1 0 0 -1;0 -1 1 1 0 -1 1 0 1 0 1 0 0 0 1 1 0 -1 -1 0 0 0 1 0;-1 -1 -1 1 1 0 0 1 0 0 1 -1 -1 -1 1 1 0 1 -1 0 0 0 0 0;0 1 0 -1 -1 0 0 -1 -1 -1 1 1 1 0 0 0 1 1 0 0 0 0 1 0;-1 0 0 0 1 0 0 0 -1 1 -1 0 -1 1 1 1 1 1 0 -1 0 -1 0 1;-1 0 0 1 1 1 1 0 1 1 1 0 1 1 1 1 -1 -1 1 0 0 0 -1 0])

tic
i=1;
while(toc<60)
    tic
    haf(matrix(1:i,1:i));
    i=i+1;
end

I have tried this out using TIO and MATLAB (I've actually never installed Octave). I imagine getting it to work is as simple as sudo apt-get install octave. The command octave will load the Octave GUI. If it's any more complicated than this, I will delete this answer until I've provided more detailed installation instructions.

Answer 8

Recently Andreas Bjorklund, Brajesh Gupt and myself published a new algorithm for Hafnians of complex matrices: https://arxiv.org/pdf/1805.12498.pdf . For an n \times n matrix it scales like n^3 2^{n/2}.

If I understand correctly Andreas' original algorithm from https://arxiv.org/pdf/1107.4466.pdf it scales like n^4 2^{n/2} or n^3 log(n) 2^{n/2} if you used Fourier transforms to do polynomial multiplications.
Our algorithm is specifically taylored for complex matrices so it will not be as fast as the ones developed here for {-1,0,1} matrices. I wonder however if one can use some of the tricks you used to improve our implementation? Also if people are interested I would like to see how their implementation do when given complex numbers instead of integers. Finally, any comments, criticism, improvements, bugs, improvements are welcomed in our repository https://github.com/XanaduAI/hafnian/

Cheers!