Millennium Problems / 06

Yang-Mills existence and mass gap

Three seeds, one channel, no agreement: the torelon mass moves more across tau than across chains. The channel is overturned. Nothing solved.

Program
The Millennium Lab
Rounds
1, 2, 3, 4, 5, 6, 7
Torelon spread
0.390 / 0.415
Receipts
66, head bdebd1e7
Date
Nov 20-21, 2026
Status
NOT SOLVED
Method
SU(2), 8^4, Metropolis
Channel
Torelon: OVERTURNED
NOT SOLVED. Nothing on this page solves the problem, proves it, or disproves it. This is research: measurements, rejections, and partial formalizations, each with a receipt. The problem belongs to the mathematicians who stated it and the researchers who have worked it for decades. This program adds a small computational exploration on top of their work and claims no credit for theirs.

01The problem, in plain words

Yang-Mills theory is the quantum theory of the strong force: a gauge field with SU(2) symmetry here, living on spacetime, with a mass gap meaning the lightest particle has strictly positive mass. The Clay problem asks for a rigorous construction of quantum Yang-Mills theory on R4 with a compact simple gauge group, and a proof that it has a mass gap Delta > 0. This lab works on a lattice, where the theory is a finite statistical system that a computer can sample, and measures masses that would become the gap in the continuum limit. The lattice is 84, small, and the continuum limit is an extrapolation, not a measurement. The extrapolation failed, twice, and this page says exactly how.

02What the lab actually ran

Method: SU(2) lattice gauge theory on an 84 lattice, Metropolis updates with checkerboard even/odd sweeps, random SU(2) rotation near identity with eps 0.35, cold start, 400 thermalization sweeps and 1500 measurement sweeps per run. Operators: APE smearing of spatial links, alpha 0.5, levels 0, 2, and 4 steps, with a variational GEVP over the three levels, C(1)v = lambda C(0)v, for the 0++ ground-state mass. The torelon is a flux tube winding around the periodic spatial direction; its mass is read from the effective mass at time separation tau, with tau = 1 the picked value. The string tension comes from Creutz ratios chi(2,2).

RoundWhat ranTorelon, beta 2.5Torelon, beta 2.6Continuum intercept
1 (Nov 20)Beta 2.2 and 2.4no torelon runno torelon runtwo spacings only, no extrapolation
2 (Nov 20)Beta 2.3 addedno torelon runno torelon run+3.473 +/- 1.506 (three points; slope 0.011 +/- 5.486)
3 (Nov 20)Beta 2.5, seed 11.2798 +/- 0.0822no run+1.634 +/- 0.820 (four points)
4 (Nov 20)APE smearing and variational GEVP introducedno new beta pointno new beta pointno new fit
5 (Nov 21)Beta 2.6, seed 11.2798 (seed 1)1.6084 +/- 0.1611+2.4230 +/- 0.7393 (five points)
6 (Nov 21)Seed 2, both betas0.8899 +/- 0.04411.1931 +/- 0.1059-0.2083 +/- 0.6426 (seed 2)
7 (Nov 21)Seed 3, both betas, plus estimator test1.0795 +/- 0.06811.2675 +/- 0.0524+2.1201 +/- 0.5256 (three-seed mean, diagnostic only)

03The three seeds disagree

At beta 2.5 the three seeds give torelon masses 1.2798, 0.8899, and 1.0795, a spread of 0.390. No pair agrees within combined errors. At beta 2.6 they give 1.6084, 1.1931, and 1.2675, a spread of 0.415. Only seeds 2 and 3 agree there.

0.511.52beta 2.5beta 2.61.281.61seed 1rounds 3 and 50.891.19seed 2round 61.081.27seed 3round 7
Figure 1. Torelon mass by seed and beta, with jackknife errors. At beta 2.5 the three seeds give 1.2798, 0.8899, and 1.0795; at beta 2.6 they give 1.6084, 1.1931, and 1.2675. No pair agrees within combined errors at beta 2.5; only seeds 2 and 3 agree at beta 2.6. Computed from yang_mills_r6.json and yang_mills_r7.json.
Torelon massSeed 1Seed 2Seed 3Spread
Beta 2.51.2798 +/- 0.08220.8899 +/- 0.04411.0795 +/- 0.06810.390
Beta 2.61.6084 +/- 0.16111.1931 +/- 0.10591.2675 +/- 0.05240.415
GEVP 0++ ground-state massSeed 1Seed 2Seed 3
Beta 2.5no GEVP (round 3 predates it)1.6263 +/- 0.10341.3555 +/- 0.1034
Beta 2.61.3318 +/- 0.09861.6030 +/- 0.13431.6591 +/- 0.1600

The basic observables are stable across seeds. The plaquette at beta 2.5 reads 0.6531, 0.6529, and 0.6537 across the three seeds; at beta 2.6 it reads 0.6705, 0.6702, and 0.6701. The string tension chi(2,2) at beta 2.5 reads 0.2071, 0.2092, and 0.2078; at beta 2.6 it reads 0.1847, 0.1856, and 0.1876. The torelon spread is hundreds of times the plaquette spread. The chains agree on the easy quantities and disagree on the one that matters.

04The estimator, not the chain

Round 7 asked whether the disagreement is statistics (different chains) or the estimator itself. Within a single chain, the torelon mass moves across tau far more than it moves across chains at fixed tau. At beta 2.5 the within-seed tau spread reaches 2.395, against a cross-seed spread of 0.390 at the picked tau = 1. At beta 2.6 it is 2.100 against 0.415. The estimator problem is about 5 to 6 times the statistics problem. Seed 1 at beta 2.5 shows it directly: tau 0 gives 2.747, tau 1 gives 1.280, tau 2 gives 0.687. There is no stable plateau. The plateau choice, not the chain, drives the discrepancy.

Within-seed tau spread, beta 2.52.395Cross-seed spread at tau=1, beta 2.50.39Within-seed tau spread, beta 2.62.1Cross-seed spread at tau=1, beta 2.60.415
Figure 2. The estimator test, round 7: the within-seed spread across tau (red) against the cross-seed spread at the picked tau = 1 (ink). At both betas the red bar is about 5 to 6 times the ink bar. Computed from yang_mills_r7.json (estimator_test).
Torelon mass by tautau 0tau 1 (picked)tau 2
Beta 2.5, seed 12.7471.2800.687
Beta 2.5, seed 22.6110.8900.215
Beta 2.5, seed 32.6431.0790.702
Beta 2.6, seed 12.7891.6080.722
Beta 2.6, seed 22.5971.1930.498
Beta 2.6, seed 32.6341.2680.941

05The overturned extrapolation

The five-point continuum extrapolation was built on seed 1 alone: beta 2.2 to 2.6, torelon mass over sqrt(sigma) against sigma a2, weighted linear fit, intercept +2.4230 +/- 0.7393. Seed 2 gives -0.2083 +/- 0.6426. The two intercepts disagree by 2.63, against combined errors of about 0.98. The extrapolation is overturned (rcpt_0061). A three-seed mean fit gives +2.1201 +/- 0.5256 with chi2 55.35 on 3 degrees of freedom; it is reported as a diagnostic only, not a continuum claim, and no extrapolation is attempted from it. OVERTURNED STANDS (rcpt_0064).

-101230+2.42 +/- 0.74Seed 1 fit (R5)-0.21 +/- 0.64Seed 2 fit (R6)+2.12 +/- 0.53Three-seed mean (R7, diagnostic)
Figure 3. Continuum intercepts from the five-point torelon extrapolation. Seed 1 alone gives +2.42; seed 2 alone gives -0.21; the three-seed mean (brass, diagnostic only) gives +2.12 with chi2 55.35 on 3 degrees of freedom. The dashed line is zero. Computed from yang_mills_r5.json, yang_mills_r6.json, and yang_mills_r7.json.
Continuum fit, torelonInterceptSlopechi2 / dof
Seed 1 (round 5)+2.4230 +/- 0.73933.2364 +/- 3.21267.60 / 3
Seed 2 (round 6)-0.2083 +/- 0.642611.4401 +/- 2.963029.71 / 3
Three-seed mean (round 7, diagnostic only)+2.1201 +/- 0.52561.9693 +/- 2.543455.35 / 3

06What we found

07Limits and blind spots

08Receipts

Receipt chain: 66 receipts, chain valid, head bdebd1e7d93dd108bffe0ba01d499e603db81a8c17027bb0aad36a70f73dacdd (verification receipt rcpt_0066). Yang-Mills receipts:

ReceiptWhat it covers
rcpt_0061Round 6: second seed at beta 2.5 and 2.6; torelon extrapolation OVERTURNED
rcpt_0062Round 6 chain verification
rcpt_0064Round 7: third seed at beta 2.5 and 2.6, estimator test; OVERTURNED STANDS
rcpt_0066Round 7 chain verification; chain head bdebd1e7

Data files: data/yang_mills_r6.json (SHA-256 41ffee8393f21b33bc5e1b43cd29fe5d0fa803ae3e4190033323007339ad5cf2), yang_mills_r7.json (SHA-256 9abc2cfe61a98b2810988dd1928d049544388b7d077c665f48cf799feaf194c1). Earlier rounds: yang_mills.json, yang_mills_r2.json, yang_mills_r3.json, yang_mills_r4.json, yang_mills_r5.json.

09Next attack

A different operator or a larger volume, not more torelon seeds. The GEVP 0++ ground-state mass is the natural next channel, but it also disagrees across seeds, so it needs the same estimator scrutiny first. A larger volume (124 or 164) is the other move.

10What we found, in plain language

We simulated the strong force on a tiny 8 by 8 by 8 by 8 grid and tried to measure the mass of the lightest particle it can make. Three runs with different random starting points gave three different answers, and the way we read the answer off the data moved it more than the runs differed from each other. So the measuring stick is broken, not just noisy, and we threw it out instead of averaging the answers and pretending they agreed.

11What we did not find

No mass gap measurement, no continuum value, no usable channel. We did not prove the gap is zero, and we did not prove it is positive. We proved one way of measuring it does not work.

The Millennium Lab. Research pages, November 2026. All numbers trace to the lab data files; every receipt is listed on the page. Nothing solved.