ResearchCoding theoryGSI-R-0009 · Note · Count
We computed the numbers of inequivalent binary self-dual codes of even length from 42 to 56 and of inequivalent stabilizer states on 13 to 20 qubits.
We computed the numbers bn of binary self-dual codes of length n (subspaces C = C⊥ of GF(2)n) up to coordinate permutation, for even n from 42 to 56, and tn of n-qubit stabilizer states (self-dual additive codes over GF(4)) up to single-qubit Clifford operations and qubit permutations, for n from 13 to 20. No code or state was listed.
- Before this work
- OEIS A003179 [1] lists bn up to length 36, and A094927 [2] and A090899 [3] list tn and in up to n = 12 (read 15 November 2026). Published: counts at lengths 38 [4] and 40 [5]; t13 ≥ 72,573,549 (Danielsen and Parker, 2006 [6]).
- Evidence
- One engine. Two programs of ours, written separately, agree on every intermediate value; they share a counting framework and small-size test data. A third, written without reading either’s code, recomputed the cases it could finish, 22.6% to 30.0% of intermediate values at lengths 52 to 56, with no mismatch. No review outside the lab has taken place.
- What you can check
- Counts only. A Python script (under a second) asserts that in follows from tn by the Euler transform of [6] and that the ratios of counts to mass-formula lower bounds exceed 1 and decrease with n. It recomputes no count.
- Not established
- That the 16 counts are correct: they rest on two programs of one lab that share a counting framework, and no review outside the lab has taken place.
- Method
- Not published.
What was known
Bouyuklieva and Bouyukliev [4] counted 38,682,183 inequivalent self-dual codes of length 38 (Theorem 1). At length 40, Bouyukliev et al. [5] list 94,343 doubly even codes, the count of Betsumiya, Harada and Munemasa [7] (Theorem 1), and 8,249,963,738 singly even codes (Table 1; they attribute the completed classification to [8]). At length 42 they report 28,376,939 codes as a partial result (Section 7), a lower bound for b42.
Danielsen and Parker [6] classified stabilizer states up to length 12. They wrote that length 13 was not feasible with their method and the computational resources available to them, and gave t13 ≥ 72,573,549, the quotient in their Eq. (7) rounded up.
The result
We find t13 = 99,941,357; the lower bound of [6] is 72,573,549.
| n | bn | rn |
|---|---|---|
| 42 | 3,934,676,259,154 | 1.4091 |
| 44 | 3,790,315,996,196,030 | 1.2246 |
| 46 | 7,070,315,085,481,125,613 | 1.1274 |
| 48 | 25,030,154,105,275,764,972,367 | 1.0734 |
| 50 | 166,477,921,531,760,133,020,482,546 | 1.0425 |
| 52 | 2,070,320,728,008,136,175,486,544,187,343 | 1.0247 |
| 54 | 48,053,939,802,408,011,950,341,021,178,783,945 | 1.0143 |
| 56 | 2,081,592,146,371,900,329,408,119,219,024,584,572,060 | 1.0083 |
| n | tn | in | rn |
|---|---|---|---|
| 13 | 99,941,357 | 98,572,066 | 1.3771 |
| 14 | 16,754,750,239 | 16,653,480,463 | 1.1836 |
| 15 | 5,654,176,880,374 | 5,637,322,136,143 | 1.0970 |
| 16 | 3,705,084,048,798,379 | 3,699,413,115,770,518 | 1.0530 |
| 17 | 4,654,021,830,321,003,627 | 4,650,311,091,993,729,907 | 1.0293 |
| 18 | 11,153,427,803,766,906,353,415 | 11,148,770,076,835,673,172,116 | 1.0162 |
| 19 | 50,929,537,738,536,576,682,934,734 | 50,918,379,656,705,307,906,016,480 | 1.0090 |
| 20 | 443,254,456,383,906,740,560,646,355,926 | 443,203,515,692,736,689,377,101,042,816 | 1.0050 |
Of the two programs, one was written with read access to the other’s tables and totals, not its code, at lengths 42 to 46 and for 13 ≤ n ≤ 17. Both reproduce every published value (bn, lengths 2 to 40; tn, n ≤ 12). Further programs of ours, using other formulations of the count, reproduce tn for 13 ≤ n ≤ 16 and, separately, t13.
How to check it
from fractions import Fraction
from math import factorial, prod
A = [ # t_1..t_20: A094927 for n <= 12, then the counts above
1, 2, 3, 6, 11, 26, 59, 182, 675, 3990, 45144, 1323363, 99941357, 16754750239, 5654176880374,
3705084048798379, 4654021830321003627, 11153427803766906353415, 50929537738536576682934734,
443254456383906740560646355926,
]
C = [ # i_1..i_20: A090899 for n <= 12, then the values above
1, 1, 1, 2, 4, 11, 26, 101, 440, 3132, 40457, 1274068, 98572066, 16653480463, 5637322136143,
3699413115770518, 4650311091993729907, 11148770076835673172116, 50918379656705307906016480,
443203515692736689377101042816,
]
B = [ # b_42, b_44, ..., b_56
3934676259154, 3790315996196030, 7070315085481125613, 25030154105275764972367,
166477921531760133020482546, 2070320728008136175486544187343, 48053939802408011950341021178783945,
2081592146371900329408119219024584572060,
]
t = [1] # Euler transform [6]: n t_n = sum_{k=1..n} c_k t_(n-k), where c_k = sum_{d|k} d i_d
for n in range(1, 21):
s = sum(sum(d * C[d - 1] for d in range(1, k + 1) if k % d == 0) * t[n - k] for k in range(1, n + 1))
assert s % n == 0 and s // n == A[n - 1], n
t.append(s // n)
print("Euler transform of i_n reproduces t_n for n = 1..20")
mass = lambda m: prod(2**i + 1 for i in range(1, m + 1))
print("bound on t_13 of [6]:", -(-mass(13) // (6**13 * factorial(13))))
rt = [Fraction(A[n - 1] * 6**n * factorial(n), mass(n)) for n in range(13, 21)]
rb = [Fraction(b * factorial(n), mass(n // 2 - 1)) for n, b in zip(range(42, 58, 2), B)]
assert all(x > 1 for x in rt + rb) and rt == sorted(rt, reverse=True) and rb == sorted(rb, reverse=True)
r = lambda x: format(float(x), ".4f")
print("t_n / bound, n = 13..20:", *map(r, rt))
print("b_n / bound, n = 42..56:", *map(r, rb))
print("all ratios exceed 1 and decrease with n")
python3 check.pyEuler transform of i_n reproduces t_n for n = 1..20
bound on t_13 of [6]: 72573549
t_n / bound, n = 13..20: 1.3771 1.1836 1.0970 1.0530 1.0293 1.0162 1.0090 1.0050
b_n / bound, n = 42..56: 1.4091 1.2246 1.1274 1.0734 1.0425 1.0247 1.0143 1.0083
all ratios exceed 1 and decrease with nFor n ≥ 13, in was obtained from tn, so the Euler transform test checks the inversion, not tn. Bouyuklieva and Bouyukliev [4] conjecture that rn decreases for binary codes of length 20 and above. rn > 1 is necessary for a correct count and says nothing about an overcount, or an undercount that stays above the mass-formula lower bound. A count by any correct method that differs from an entry refutes that entry.
What it changes
The values 1/|Aut| over the equivalence classes sum to the mass-formula lower bound [4], [6] (|Aut|: the stabiliser in the equivalence group, of order n! for codes and 6n n! for states), and a class with |Aut| ≥ 2 adds at least 1/2 to the excess of the count over it. If the counts are correct, at most 2(1 − 1/rn) of the classes therefore have a non-trivial automorphism group: 1.6% at length 56, 1.0% at 20 qubits.
Limits
- Correctness. The 16 counts rest on two programs of one lab that share a counting framework, with no review outside the lab. No further program has recomputed lengths 42 to 50, 17 ≤ n ≤ 20 or 70% to 77% of intermediate values at lengths 52 to 56. An independent count at one size would close it for that size.
- Priority. To our knowledge no total is published for length 42 or above or for n ≥ 13 (searched in OEIS, arXiv, the authors’ databases and the web on 26 October and, briefly, 15 November 2026). A dated earlier record would change this.
References
- [1]OEIS A003179, Number of self-dual binary codes of length 2n (up to column permutation equivalence), read 15 November 2026 (record last edited 30 May 2026). Link
- [2]OEIS A094927, Number of nonisomorphic (possibly decomposable) self-dual quantum codes on n qubits, read 15 November 2026 (record last edited 5 November 2025). Link
- [3]OEIS A090899, Number of nonisomorphic indecomposable self-dual quantum codes on n qubits, read 15 November 2026 (record last edited 29 June 2026). Link
- [4]S. Bouyuklieva, I. Bouyukliev, An algorithm for classification of binary self-dual codes, IEEE Trans. Inform. Theory 58(6), 3933–3940 (2012), doi:10.1109/TIT.2012.2190134. The version read is arXiv:1106.5930v1 (29 June 2011), titled On the Classification of Binary Self-Dual Codes: mass formula, Theorem 1, Conjecture 1. Link
- [5]I. Bouyukliev, S. Bouyuklieva, M. Dzhumalieva-Stoeva, V. Monev, What is GenSelfDual?, Serdica J. Comput. 10(3–4), 231–244, doi:10.55630/sjc.2016.10.231-244. The version read, the authors’ preprint of 30 November 2014, has Table 1 and Section 7; the Serdica text was not compared. Link
- [6]L. E. Danielsen, M. G. Parker, On the classification of all self-dual additive codes over GF(4) of length up to 12, J. Combin. Theory Ser. A 113(7), 1351–1367 (2006), doi:10.1016/j.jcta.2005.12.004; the version read is arXiv:math/0504522v4 (17 February 2006): Eqs. (5) and (7), the Euler transform, Tables 1 and 2, Section 6. Link
- [7]K. Betsumiya, M. Harada, A. Munemasa, A complete classification of doubly even self-dual codes of length 40, Electron. J. Combin. 19(3), #P18 (2012), arXiv:1104.3727; Theorem 1 (94,343 codes), read in the journal text. Link
- [8]I. Bouyukliev, M. Dzhumalieva-Stoeva, V. Monev, Classification of binary self-dual codes of length 40, IEEE Trans. Inform. Theory 61(8), 4253–4258 (2015), doi:10.1109/TIT.2015.2440266 (bibliographic record only; not read). Link