ResearchCombinatorial designsGSI-R-0005 · Paper · Existence
Proper circulant weighing matrices exist for CW(120, 36), which Tan (2018) excludes; CW(180, 36), listed “Open”; and 209 pairs listed “No” in La Jolla version 1.3.
A circulant weighing matrix CW(n, k) is an n × n circulant W with entries 0, ±1, k nonzero per row (the weight) and WWT = kI; it is proper if its first row’s support lies in no coset of a proper subgroup of ℤn. We give 6 proper CW(120, 36) and 12 proper CW(180, 36), inequivalent under shifts, negation and multipliers (units of ℤn), and proper matrices for 209 pairs (n, k) listed “No” in La Jolla version 1.3.
- Before this work
- Tan [1], Theorem 4.24: “There is no CW(120, 36)”. La Jolla version 1.3 [2] lists it “No” (comment “Tan (2018)”), CW(180, 36) “Open” (also open in Table 10 of [3]) and 209 pairs “No”; version 2.0 lists the 209 as existing (read 15 November 2026).
- Evidence
- Independently checked. The label covers existence; each matrix is its own certificate. Programs written separately by our own system verified them from the definition, and a Lean 4 kernel check covers weight and autocorrelation of one matrix of each order. All share the first-row files, and no person outside the lab has reviewed our files.
- What you can check
- One half: the existence of the matrices. Run the short check below on one matrix, or check_objects.py (Python 3, under a second) on all 6 + 12 + 209; nothing published tests completeness, the 12 non-existence results or the count.
- Not established
- Completeness of the 6 and 12 classes, non-existence at the 12 parameter pairs, and the count of decided open entries. They rest on unpublished exhaustive computations, most by floating-point enumeration.
- Method
- Not published.
What was known
Theorem 4.24 of [1] (arXiv version 4) has no proof of its own: Tan sketches an argument from computer classifications that are not printed and defers the details to his 2014 thesis [4].
La Jolla keys its entries by n and s = √k, and its statuses refer to proper matrices (version 2.0 notebook): CW(120, 36) is its entry CW(120, 6), the weight-4 entries are CW(n, 2) and the weight-16 entries CW(n, 4). To our knowledge no earlier public matrix of weight 36 at order 120 or 180 exists (searched: web, 24 October and 15 November; arXiv, Zenodo and GitHub, 1–2 November 2026). The latest table of known proper CW(n, 36) that we read, Table 7 of Gordon (2026) [5], lists n = 26m, 48m, 91, 168 and 182.
Arasu, Gordon and Zhang (2021) [3] credit the classification of weight 4 to Eades and Hain (1976) [6] and give the weight-16 family as their Theorem 6.2. Version 1.3 cites “Arasu, Leung, Ma, Nabavi and Ray-Chaudhuri (2006)” [7] for the 10 weight-16 entries without naming the paper. The explicit matrices below show the 209 entries to be errors in version 1.3. Version 2.0 lists the 199 of weight 4 with the comment “Eades and Hain; unique equivalence class” and the 10 of weight 16 with comments citing constructions or verifications of September 2026, 2 credited to “ChatGPT Sep 2026”.
On 19 October 2026 D. M. Gordon, who maintains the La Jolla repository, wrote publicly [8] that no CW(112, 36), CW(117, 36), CW(195, 36) or CW(156, 81) exists, that he had one existence result and that a paper was in preparation; the comment names no case for that result. We cede priority on those 4 non-existence results.
The result
CW(120, 36), CW(180, 36) and the 209 pairs in the last two rows below have proper matrices, each given as a first row in the published files and verified from the definition.
from math import gcd
n = 120
P = [17, 26, 30, 37, 42, 54, 57, 58, 60, 62, 68, 70, 77, 84, 94, 97, 98, 106, 112, 116, 117]
N = [0, 2, 6, 8, 14, 18, 24, 50, 52, 56, 66, 74, 78, 102, 110]
a = [0] * n
for i in P: a[i] = 1
for i in N: a[i] = -1
assert sum(x * x for x in a) == 36
assert all(sum(a[i] * a[(i + s) % n] for i in range(n)) == 0 for s in range(1, n))
assert gcd(n, *[i - P[0] for i in P + N]) == 1
print('CW(120, 36): weight 36, autocorrelation 0 at all 119 shifts, proper')| Parameters | Our matrices | Version 1.3 | Version 2.0 |
|---|---|---|---|
| CW(120, 36) | 6 classes | “No” | unchanged |
| CW(180, 36) | 12 classes | “Open” | unchanged |
| weight 4, 199 even n from 404 to 996 | 199 | “No” | “All” |
| weight 16, n = 98j, j = 1 to 10 | 10 | “No” | “Yes” |
Weight 4: for n = 2m put y = xm, so y2 = 1 and a = −1 + x + xm + xm+1 = (x − 1) + y(x + 1). Then a(x)a(x-1) = [(x − 1)(x-1 − 1) + (x + 1)(x-1 + 1)] + y[(x − 1)(x-1 + 1) + (x + 1)(x-1 − 1)] = 4 + 0. Weight 16: the 10 rows equal A = (1 − X7m) C(X2m) + (1 + X7m) X C(Xm), C = −1 + X + X2 + X4, n = 14m, the construction in the proof of Theorem 6.2 of [3], rebuilt in check_objects.py.
An exhaustive computation of ours finds no further class of CW(120, 36) or CW(180, 36), and no circulant weighing matrix of any kind, proper or not, at the 12 parameter pairs below.
At order 180, 2 separately written programs return the same 12 classes; at order 120, 2 programs return the same 6, one in exact arithmetic.
| Parameters | Announced 19 October | Exact re-derivation | Version 2.0 |
|---|---|---|---|
| CW(112, 36) | yes | yes | “Open” |
| CW(117, 36) | yes | no | “No” |
| CW(195, 36) | yes | no | “No” |
| CW(156, 81) | yes | yes | “Open” |
| CW(195, 81) | no | no | “Open” |
| CW(198, 81) | no | yes + verifier | “Open” |
| CW(112, 100) | no | yes | “Open” |
| CW(120, 100) | no | yes | “Open” |
| CW(156, 100) | no | yes | “Open” |
| CW(165, 100) | no | no | “Open” |
| CW(182, 100) | no | no | “Open” |
| CW(195, 100) | no | yes + verifier | “Open” |
Two other programs, written separately, reproduce all 12 non-existence results. A third, reusing modules of one of them, reproduces 9 of the 12; for CW(117, 36), CW(195, 36) and CW(165, 100) the two others are the only second checks. All 12 were open in Table 10 of [3] and in version 1.3, and none is settled in Tan’s 2026 preprint [9].
Of the 2,763 pairs (n, k) that version 1.3 lists “Open” and that our computation covered, 2,397 were recorded as decided by 26 October 2026: 2,342 with no matrix of any kind, 25 with a proper matrix and 30 with improper matrices only. The 25 proper matrices are explicit and verified from the definition; they are not published here.
Of the 2,342 non-existence results, a second program covers 899 (the 12 above among them), 561 more have an agreeing certificate file of the first program (2 a check of another kind), and 882 rest on one program or carry a recorded caveat. The 30 improper-only results rest on the first program; 13 have an agreeing certificate file and CW(155, 25) a second program. The 2,342 include CW(182, 144), whose non-existence conflicts with Table 11 of [3] read at m = 1, and CW(155, 100), which we do not claim; we treat both as unresolved, so without them 2,395 are decided and 2,340 have no matrix.
In version 2.0, 115 of the 2,342 non-existence results are listed “No” (113 credited to ChatGPT, August or September 2026, 2 to “Gordon (2026)”) and 15 of the 25 proper matrices “Yes”, citing published constructions or others’ work; the other 10, CW(180, 36) among them, remain “Open”. We made no per-case literature search and claim no priority for these decisions.
How to check it
- Run the 11 lines above (Python 3.9 or newer); one line prints, in about 0.02 seconds. Moving one position of P (17 to 19) fails the second assertion.
- Download the files below into one folder and run the next block (Python 3, standard library, about 0.5 seconds; Lean 4.35.0, under 20 seconds). Lean accepts one CW(120, 36) and one CW(180, 36) (support size 36, zero autocorrelation) and proves that a copy of the first with one position moved is not a CW.
- In a clone of [2] at commit a0b0e52, apply its is_cwm (which does not test properness) to get_cwm(n, 2, 0) for the 199 values of n in lj_weight4.json and to get_cwm(n, 4, 0) for the 10 in lj_weight16.json; is_cwm accepts 199 of 199 and 10 of 10.
$ python3 -I check_objects.py .
PASS CW(120,36): 6 classes, weight 36 and zero autocorrelation at all 119 shifts: True; proper: True; pairwise inequivalent: True
PASS CW(180,36): 12 classes, weight 36 and zero autocorrelation at all 179 shifts: True; proper: True; pairwise inequivalent: True
PASS weight 4: 199 values of n (404 to 996), each -1 + x + x^m + x^(m+1) with n = 2m, valid and proper; failures: 0
ALL CHECKS PASSED
$ python3 -I check_orbits.py .
CW(120,36): orbit sizes (both signs) [1920, 1920, 3840, 3840, 3840, 3840] total first rows 19200 with row sum +6 9600
CW(180,36): orbit sizes (both signs) [2880, 2880, 2880, 2880, 5760, 5760, 5760, 5760, 5760, 5760, 5760, 5760] total first rows 57600 with row sum +6 28800
$ lean CW120_180.lean
(no output, exit status 0)Passing these checks shows that each listed row is a proper CW of its parameters, that the 6 and the 12 rows are pairwise inequivalent, and that the 209 entries are contradicted. The first two checks use our rows; the third uses the maintainer’s rows and checker. A matrix of any kind at one of the 12 parameter pairs would refute the non-existence results; a further class of CW(120, 36) or CW(180, 36) would refute the lists.
What it changes
Theorem 4.24 as printed in arXiv version 4 is false for proper matrices, and the La Jolla entry for CW(120, 36) cites Tan (2018) alone.
Limits
- Completeness and non-existence. The lists of 6 and 12 classes, the 12 non-existence results and the count of 2,397 rest on unpublished exhaustive computations, most by floating-point enumeration without an exact certificate of completeness; 5 of the 12 have no exact re-derivation. Publication or exact re-derivation would close it.
- Priority. Gordon’s existence result of 19 October 2026 is unnamed and its paper unpublished, so a matrix of order 120 or 180 may be his; the paper would close this.
- Weight 25. Gordon [8] says an unpublished paper of Leung and Ma settles all CW(v, 25) in theory; version 2.0 lists CW(155, 25), CW(165, 25) and CW(186, 25), 3 of our 30 improper-only decisions, as “No”, citing Leung and Ma; [3] attributes n ≡ 0 (mod 5) to their published paper [10] and the other n to a 2011 preprint.
- Tan’s argument. We do not know where the sketch for Theorem 4.24 fails; the thesis [4] would close it.
- The table audit. The 209 are the version 1.3 “No” entries of weight 4 at even n and weight 16 at n divisible by 14, with n < 1000. We audited part of the “No” entries by computation and none of the rest, so other errors may remain. A full audit would close it.
Files
- cw120_36_classes.jsonfirst rows of the 6 classes of CW(120, 36), as sets P (+1) and N (−1)
- cw180_36_classes.jsonfirst rows of the 12 listed classes of CW(180, 36); the list is not claimed complete
- lj_weight4.jsonthe 199 values of n and, for each, the first row −1 + x + x^m + x^(m+1)
- lj_weight16.jsonthe 10 values n = 98j and a first row of CW(n, 16) for each
- check_objects.pyweight, autocorrelation, properness and inequivalence of all of the above; Python 3, standard library
- check_objects.logfull output of check_objects.py
- check_orbits.pyorbit sizes and first-row counts of the 6 and 12 classes; Python 3, standard library
- CW120_180.leanLean 4 check of weight and autocorrelation of one CW(120, 36) and one CW(180, 36), and of a mutant that must fail
References
- [1]M. M. Tan, Group invariant weighing matrices, arXiv:1610.01914v4 (2 March 2018); Des. Codes Cryptogr. 86 (2018) 2677–2702 (journal version not read). Link
- [2]D. M. Gordon, La Jolla Circulant Weighing Matrices Repository, Zenodo: version 1.3, doi:10.5281/zenodo.19741187 (24 April 2026); version 2.0, doi:10.5281/zenodo.24170503 (12 November 2026; GitHub dmgordo/circulant-weighing-matrices, tag v2.0, commit a0b0e52). Link
- [3]K. T. Arasu, D. M. Gordon, Y. Zhang, New nonexistence results on circulant weighing matrices, arXiv:1908.08447v3 (11 May 2021): Tables 10 and 11, Theorem 6.2 (the arXiv text was read, not the journal version, Cryptogr. Commun. 13 (2021) 775–789). Link
- [4]M. M. Tan, Relative difference sets and circulant weighing matrices, PhD thesis, Nanyang Technological University, 2014 (cited in [1]; not obtained).
- [5]D. M. Gordon, Cyclic relative difference sets and circulant weighing matrices, arXiv:2501.14924v2 (8 April 2026): Table 7 (arXiv text read). Link
- [6]P. Eades, R. M. Hain, On circulant weighing matrices, Ars Combin. 2 (1976) 265–284 (not read).
- [7]K. T. Arasu, K. H. Leung, S. L. Ma, A. Nabavi, D. K. Ray-Chaudhuri, Circulant weighing matrices of weight 2^(2t), Des. Codes Cryptogr. 41 (2006) 111–123, cited in [3] for the weight-16 orders (not read); the same authors also published Determination of all possible orders of weight 16 circulant weighing matrices, Finite Fields Appl. 12 (2006) 498–538 (not read), which is also a candidate for the La Jolla citation.
- [8]D. M. Gordon, comment of 19 October 2026 on pull request 2 of dmgordo/circulant-weighing-matrices (read 15 November 2026). Link
- [9]M. M. Tan, Character and multiplier obstructions for circulant weighing matrices, arXiv:2608.18468v2 (1 September 2026; not refereed); we searched its full text for CW(120, 36), CW(180, 36), Theorem 4.24 and the 12 parameter pairs. Link
- [10]K. H. Leung, S. L. Ma, Proper circulant weighing matrices of weight p², Des. Codes Cryptogr. 72 (2014) 539–550 (cited in [3]; not read). Link