ResearchStatistical physicsGSI-R-0016 · Note · Measurement

Our symmetry-reduced transfer-matrix program took 1.6 times as long as our engine to compute the periodic 4 × 4 × 4 Ising density of states exactly.

Ω(b), the zero-field Ising density of states, counts ±1 assignments to 64 spins of the periodic 4 × 4 × 4 simple-cubic lattice with b bonds joining unequal spins. Our engine and our symmetry-reduced transfer-matrix program [1], [2], [3] agree on Ω; in one single-thread run each, the program took 72.8 s wall-clock, 1.6 times the engine’s user CPU time.

Before this work
Pearson [4] published the partition function of this lattice in 1982, and OEIS A008551 [5] lists Ω(b) for even b ≤ 50 (read 15 November 2026). We compare with our own implementation of the method of [1] and [2], the baseline below.
Evidence
Internal. The engine and a plain transfer matrix, written in the same project and sharing no code, each computed the 4 × 4 × 4 table; the engine alone computed the 5 × 5 × 4 table. No one outside the lab has reviewed any of it.
What you can check
One half: the table. Run the script in How to check it on the two files below (Python 3, no packages, under a second). The timings cannot be rerun, because neither the engine nor the baseline is released.
Not established
That the engine is faster than the published method. The baseline, our own, reuses part of the engine’s source; each ran once, on one lattice, with different timers, under load.
Method
Not published.

What was known

The microcanonical transfer matrix [1], [2] counts configurations by energy one layer at a time; Kim [6] gives it for periodic layers with free ends and Valani [3] reduces layer states to symmetry classes. Valani’s Table 3.1 of maximal figure sizes lists the periodic 4 × 4 × 4 and 4 × 4 × 10, the 4 × 5 × 5 of Bhanot and Sastry [7] with one open direction, and no fully periodic lattice with a side of 5. Bhanot et al. [8] approximated the partition function of the 5 × 5 × 5 lattice in 1987 (abstract read 16 November 2026).

The result

The baseline, our symmetry-reduced program, implements [1], [2] in the form of [6], closed periodically by summing over the starting layer, cache-blocked, with one starting layer per class of layer configurations under translations, square symmetries and global spin flip, weighted by class size [3].

ProgramTimerTime (s)
Plain transfer matrix [1], [2], [6]: all 32,768 starting layers (the 216 layer states with the global flip fixed), 12 threadsuser CPU18,120
Plain, 4 starting layers, 1 thread, scaled to 32,768wall-clock9,511
Cache-blocked, the same 4 layers, scaledwall-clock5,622
Cache-blocked, one starting layer per class, 1 threadwall-clock72.8
Single runs; rows 2 and 3 are projections from 4 of 32,768 starting layers. The engine’s time is not published.

Conditions: one x86-64 virtual machine with AVX-512 (model and core count not recorded). The 1-minute load average before and after each block of runs was 29.6 and 25.2 for the baseline, 31.4 and 40.0 for the engine; the 12-thread run has no record. The baseline’s timer covers its transfer loop, excluding set-up; the engine’s covers its process, excluding an untimed post-processing step.

  b  Ω(b)
  0  2
  6  128
 10  384
 12  3648
 14  1920
 16  20448
...
 96  2158800489098126488
...
186  128
192  2
Levels of omega_4x4x4_periodic.txt: 91 levels with Ω > 0, b even from 0 to 192, maximum at b = 96. Levels with Ω = 0 are not listed.

The engine also computed the density of states of the fully periodic 5 × 5 × 4 lattice (100 spins); it passes the five checks below that apply.

How to check it

Put both files and check.py in one folder and run python3 check.py.

OEIS = [2, 0, 0, 128, 0, 384, 3648, 1920, 20448, 71680, 125952, 537600, 1445440, 3756288, 11173440, 29955072,
        79741908, 214759680, 564284736, 1466972160, 3794966880, 9719376896, 24675685440, 62292331392,
        156254486400, 389861188608]  # OEIS A008551: entry n is the number of configurations with b = 2n

def check(path, N, B, F4, bipartite):
    om = {int(a): int(c) for a, c in (line.split() for line in open(path))}
    m = lambda k: sum(w * (B - 2 * b) ** k for b, w in om.items())  # sum of E^k over all configurations, E = B - 2b
    ok = {"sum": sum(om.values()) == 2**N, "m1": m(1) == 0, "m2": m(2) == 2**N * B,
          "m4": m(4) == 2**N * (3 * B * B - 2 * B + 24 * F4),
          "levels": [om[b] for b in (0, 6, 10, 12, 14)] == [2, 2 * N, 6 * N, N * (N - 7), 30 * N]}
    if bipartite:
        ok["symmetry"] = all(om.get(B - b) == w for b, w in om.items())
        ok["A008551"] = all(om.get(2 * n, 0) == v for n, v in enumerate(OEIS))
    print(f"{path}: {len(om)} levels, {sum(ok.values())} of {len(ok)} checks pass", *[k for k in ok if not ok[k]])

check("omega_4x4x4_periodic.txt", N=64, B=192, F4=240, bipartite=True)
check("omega_5x5x4_periodic.txt", N=100, B=300, F4=325, bipartite=False)
check.py. N spins, B bonds, F4, the number of 4-cycles. Levels 0, 6, 10, 12, 14: ground state, one flipped spin, two adjacent, two non-adjacent, three in a path, each with global flip. Symmetry: the 4 × 4 × 4 lattice is bipartite.
omega_4x4x4_periodic.txt: 91 levels, 7 of 7 checks pass
omega_5x5x4_periodic.txt: 128 levels, 5 of 5 checks pass
Output with Python 3.15.0 on an Apple M4, 0.02 s.

The checks cover the sum, three moments and the lowest levels, not individual bulk entries. Levels 0 to 50 are also in [5] and 142 to 192 follow by symmetry. A failed check, or a computation that contradicts an entry, would refute the table.

Limits

  • The comparison. The baseline, our own program, reuses part of the engine’s source; each program ran once, on one lattice, under load. Repeated runs on a named CPU, a second lattice and an implementation by others would close it. Valani [3] offers code, which we did not run.
  • The timers differ. In the 12-thread plain run, 18,120 s of user time corresponded to 24,088 thread-seconds of wall-clock; the same factor on the baseline would give a ratio near 1.2, but for one thread it is unknown. User CPU time for both programs would close it.
  • Levels 52 to 140 of the 4 × 4 × 4 table rest on the moment checks and on our plain transfer matrix, which gave a table identical to the engine’s; we have not compared them with [4]. A comparison with [4] or a computation by others would close it.
  • The 5 × 5 × 4 table rests on the engine and the checks alone. On a different lattice, 5 × 5 × 10 with open ends, the engine’s polynomial equals the 621 coefficients of Table E.1 in [3]. A computation by others would close it.

Files

References

  1. [1]G. Bhanot, A numerical method to compute exactly the partition function with application to Z(n) theories in two dimensions, J. Stat. Phys. 60, 55–75 (1990). Link
  2. [2]R. J. Creswick, Transfer matrix for the restricted canonical and microcanonical ensembles, Phys. Rev. E 52, R5735–R5738 (1995). Link
  3. [3]Y. P. Valani, On the partition function for the three-dimensional Ising model, PhD thesis, City University London, August 2011 (unpublished; City Research Online). Link
  4. [4]R. B. Pearson, Partition function of the Ising model on the periodic 4×4×4 lattice, Phys. Rev. B 26, 6285–6290 (1982). Link
  5. [5]OEIS A008551 (M. Karliner, 1996): number of times the n-th energy state appears in the statistical sum of the Ising model on a 4 × 4 × 4 cube; read 15 November 2026. Link
  6. [6]S.-Y. Kim, Partition function zeros of the Q-state Potts model on the simple-cubic lattice, Nucl. Phys. B 637, 409–426 (2002), arXiv:cond-mat/0205451. Link
  7. [7]G. Bhanot, S. Sastry, Solving the Ising model exactly on a 5×5×4 lattice using the Connection Machine, J. Stat. Phys. 60, 333–346 (1990). Link
  8. [8]G. Bhanot, S. Black, P. Carter, R. Salvador, A new method for the partition function of discrete systems with application to the 3D Ising model, Phys. Lett. B 183, 331–336 (1987). Link

Cite asKishan Kumar, The 4 × 4 × 4 Ising density of states: a timing record, GSI-R-0016, https://generalised.si/research/ising-3d-density-of-states, 2026.

Published16 November 2026

History
  • This account published.

All results