# Yang-Mills mass gap hunt, round 2: third beta point at 2.3, plus a 10^4 volume check

Date: 2026-11-20. Lattice: 8^4, periodic in all directions, with one extra run at 10^4 at the same beta. Gauge group: SU(2). Updates: Metropolis with checkerboard even/odd ordering, random SU(2) rotation near the identity, eps 0.35, cold start. Seed: 20261120 plus int(beta times 100), the same convention as round 1, so beta 2.3 uses seed 20261350. Thermalization: 400 sweeps. Measurement: 1500 sweeps. Errors are jackknife errors over 30 blocks of 50 sweeps, except Creutz ratios which use 20 blocks of the every-5-sweeps Wilson loop measurements. Runtime: 667.3 seconds for the 8^4 run and 1012.3 seconds for the 10^4 run, 1683.6 seconds total. The wall time is much longer than round 1 (338.6 seconds for two 8^4 runs) because the machine was heavily loaded during this run, with load average near 8.8 at the start; the code and protocol are unchanged. Full numbers live in data/yang_mills_r2.json, and the run log is data/wilson_r2.log. The runner is compute/run_r2.py, which calls the same run function in compute/wilson_lattice.py. Round 1 numbers are quoted from data/yang_mills.json for comparison.

Beta 2.3 sits between the round 1 points, beta 2.2 and beta 2.4, so the gap estimate can now be tracked at three lattice spacings. Nothing in this file proves the continuum mass gap. A lattice gap estimate at finite spacing is a measurement, not a proof. What follows is what the data says at these three spacings.

## Plaquette and string tension

Beta 2.3, 8^4: average plaquette 0.602788 plus or minus 0.000484, acceptance rate 0.445. Round 1 for comparison: 0.568932 plus or minus 0.000566 at beta 2.2, and 0.631483 plus or minus 0.000365 at beta 2.4. The plaquette at 2.3 sits between the two round 1 values, as expected when the coupling weakens with beta.

Creutz ratios chi(R, T), from Wilson loops measured every 5 sweeps:

Beta 2.3, 8^4: chi(2,2) 0.3166 plus or minus 0.0028, chi(2,3) 0.2770 plus or minus 0.0032, chi(3,3) 0.2168 plus or minus 0.0067. Round 1: chi(2,2) 0.4054 plus or minus 0.0027 at beta 2.2, and 0.2444 plus or minus 0.0018 at beta 2.4.

String tension estimate sigma a^2 from chi(2,2): 0.3166 plus or minus 0.0028 at beta 2.3, between 0.4054 at beta 2.2 and 0.2444 at beta 2.4. The Creutz ratios fall with loop size, which is the expected approach toward the asymptotic string tension from above, so chi(2,2) overestimates sigma somewhat. The tension in lattice units falls monotonically with beta, as expected when the lattice spacing shrinks.

## Gap channel 1: 0++ glueball from the plaquette correlator

Operator: O(t), the spatial plaquette average in time slice t. Connected correlator C(tau) equal to the average of O(0) O(tau) minus the square of the average of O. Effective mass m_eff(tau) equal to minus the log of C(tau+1) over C(tau).

Beta 2.3, 8^4: C(0) 0.00010051, C(1) 0.00000594, C(2) 0.00000091, C(3) 0.00000373, C(4) 0.00000583. Effective masses: m_eff(0) 2.8292 plus or minus 0.3597, m_eff(1) 1.8733 plus or minus 5.0683, m_eff(2) minus 1.4081 plus or minus 4.8275.

Reading: the correlator decays by a factor of about 17 from tau 0 to tau 1, which is the gap signal, but then it stops decaying: C(2) is smaller than C(3) and C(4), so the correlator has hit its noise floor after tau 1 and no later effective mass is meaningful. The m_eff(1) error is 270 percent of the value, so the pick rule falls back to m_eff(0), 2.8292 plus or minus 0.3597. That value sits between the round 1 primary estimates, 1.9691 plus or minus 0.1592 at beta 2.2 and 2.7052 plus or minus 0.3061 at beta 2.4, but closer to the 2.4 value than to the 2.2 value. All three primary 0++ estimates are m_eff(0), so all three carry excited-state contamination, and the comparison is at least apples to apples in that sense. In units of the string tension scale, m over sqrt(sigma) for the 0++ channel is 3.09 at beta 2.2, 5.03 at beta 2.3, and 5.47 at beta 2.4. The big move happens between 2.2 and 2.3, then the ratio flattens. The 0++ channel still cannot fix a stable continuum value.

## Gap channel 2: 2++ tensor glueball from the orientation-difference correlator

Operator: D(t) equal to P_xy(t) minus P_xz(t), the difference of two spatial plaquette orientations in time slice t. Cubic symmetry forces its vacuum expectation value to zero, so its correlator needs no disconnected subtraction: C_D(tau) equal to the average of D(0) D(tau).

Beta 2.3, 8^4: C_D(0) 0.00032839, C_D(1) 0.00001121, C_D(2) 0.00000077, C_D(3) minus 0.00000543, C_D(4) minus 0.00000271. Effective masses: m_eff(0) 3.3770 plus or minus 0.4166, m_eff(1) 2.6799 plus or minus 5.2697. No positive correlator at tau 2, so no m_eff(2).

Reading: the tensor correlator survives to tau 2 at beta 2.3, one step further than at beta 2.2, but m_eff(1) carries a 197 percent error, so the pick rule falls back to m_eff(0), 3.3770 plus or minus 0.4166. That sits between the round 1 values, 4.4826 plus or minus 0.8670 at beta 2.2 and 3.2237 plus or minus 0.3401 at beta 2.4, and falls monotonically with beta, which is the expected direction. The tensor mass sits above the scalar mass at all three betas, which matches the known ordering of glueball states. This channel corroborates rather than leads, as in round 1.

## Gap channel 3: torelon from the spatial Polyakov loop correlator

Operator: P_s, the Polyakov loop winding around a spatial direction, averaged over the three spatial directions and over transverse sites. Correlator C_T(tau) equal to the average of P_s(0) times the conjugate of P_s(tau), minus the square of the mean loop. It decays as exp(-m_tor tau) with m_tor equal to sigma times L_s in lattice units for a confining theory.

Beta 2.3, 8^4: C_T(0) 0.99948, C_T(1) 0.03334, C_T(2) 0.00519, C_T(3) 0.00147, C_T(4) 0.00058. Effective masses: m_eff(0) 3.4004 plus or minus 0.0362, m_eff(1) 1.8602 plus or minus 0.2302, m_eff(2) 1.2637 plus or minus 0.6599.

Reading: this is still the cleanest channel. The correlator stays positive through tau 4, and the effective mass falls with tau, as a spectral decomposition requires. The m_eff(1) value, 1.8602 plus or minus 0.2302, has a 12 percent error, so it is the picked estimate, and it sits between the round 1 values, 2.6731 plus or minus 0.7993 at beta 2.2 and 1.7393 plus or minus 0.1361 at beta 2.4. The fall from 2.2 to 2.3 is large, 0.81, and the fall from 2.3 to 2.4 is small, 0.12, so the torelon mass drops fast and then flattens. Cross-check against the string tension: sigma times L_s predicts 2.53 at beta 2.3, and the measured m_eff(1) sits below that prediction by 0.67, the same direction and similar size as round 1 (below by 0.57 at beta 2.2 and by 0.21 at beta 2.4), consistent with the flux tube Luescher correction and with residual excited-state contamination in m_eff(0). The torelon mass is nonzero at all three spacings with clear statistical significance.

## Confinement diagnostic: time-like Polyakov loop correlator

The correlator of time-like Polyakov loops at spatial separation r decays as exp(-N_t V(r)) with V(r) the static quark potential. Beta 2.3, 8^4: C_P(0) 1.00171, C_P(1) 0.03169, C_P(2) 0.00485, C_P(3) 0.00352, C_P(4) minus 0.00073. Static potential at separation 1: V(1) 0.4317 plus or minus 0.0088, between 0.4980 plus or minus 0.0117 at beta 2.2 and 0.3940 plus or minus 0.0067 at beta 2.4. The potential is large and positive at all three spacings and falls monotonically with beta, consistent with a confining potential that weakens in lattice units as the spacing shrinks.

## Gap estimates and the three-point continuum extrapolation

Primary gap estimate per channel at beta 2.3, 8^4, taking m_eff(1) when its relative jackknife error is under 50 percent and m_eff(0) otherwise, the same rule as round 1:

0++ glueball: 2.8292 plus or minus 0.3597 (m_eff(0); m_eff(1) error too large). 2++ tensor glueball: 3.3770 plus or minus 0.4166 (m_eff(0); m_eff(1) error too large). Torelon: 1.8602 plus or minus 0.2302 (m_eff(1)).

Three-point extrapolation, torelon channel. The torelon mass in units of sqrt(sigma), with sigma a^2 as the a^2 proxy at fixed physical sigma:

Beta 2.2: sigma a^2 0.4054, m_tor 2.673 plus or minus 0.799, m_tor over sqrt(sigma) 4.198. Beta 2.3: sigma a^2 0.3166, m_tor 1.860 plus or minus 0.230, m_tor over sqrt(sigma) 3.306. Beta 2.4: sigma a^2 0.2444, m_tor 1.739 plus or minus 0.136, m_tor over sqrt(sigma) 3.519.

Weighted linear fit of m_tor over sqrt(sigma) against sigma a^2, with errors on the ratio propagated from the m_tor and sigma errors: continuum intercept 3.473 plus or minus 1.506, slope 0.011 plus or minus 5.486, chi2 over degrees of freedom 0.53 over 1.

What this fit does support: a linear a^2 fit can now be performed at all, which two points could not do, and the intercept is nonzero at 2.3 sigma (3.473 divided by 1.506). The three points are mutually consistent with a flat line in a^2, since chi2 over 1 degree of freedom is 0.53.

What it does not support: almost everything else. The slope is 0.011 with an error of 5.486, so the fit cannot tell whether the ratio rises, falls, or is flat with a^2. The three points are not monotonic: the ratio falls from 4.198 to 3.306, then rises to 3.519. The intercept error is 43 percent of the value, dominated by the large beta 2.2 torelon error (0.799 on 2.673). Sigma a^2 is itself an estimate from chi(2,2) that overestimates the asymptotic tension, and the x errors are ignored in the fit. The 8^4 volume leaves finite-volume effects uncontrolled, and the linear a^2 form is assumed, not established: three points with two fit parameters leave one degree of freedom, so the fit cannot test the functional form. This is an extrapolation attempt, not a continuum result. It points in the direction of a nonzero continuum torelon mass, and the torelon channel stays nonzero at all three spacings, but it does not prove the continuum mass gap. A real extrapolation needs more beta values, larger volumes, better glueball operators, and a controlled continuum limit.

## Volume check: beta 2.3 on 10^4

Same beta, sweeps, and seed family on a 10^4 lattice. Plaquette 0.602805 plus or minus 0.000273, identical to the 8^4 value 0.602788 plus or minus 0.000484. String tension sigma a^2 0.3148 plus or minus 0.0017, identical to the 8^4 value 0.3166 plus or minus 0.0028. Creutz ratios agree: chi(2,2) 0.3148, chi(2,3) 0.2751, chi(3,3) 0.2214 on 10^4, against 0.3166, 0.2770, 0.2168 on 8^4. Static potential V(1) 0.4415 plus or minus 0.0091, against 0.4317 plus or minus 0.0088 on 8^4. The bulk observables show no finite-volume shift between 8^4 and 10^4 at this beta.

The gap channels move more. 0++ glueball m_eff(0) is 2.2472 plus or minus 0.2401 on 10^4, against 2.8292 plus or minus 0.3597 on 8^4, a shift of 0.58, about 1.35 combined sigma, downward on the larger volume. The glueball correlator amplitude C(0) halves on 10^4, 0.00005297 against 0.00010051, a factor of 1.90, close to the spatial volume ratio of 1.95, which is the expected scaling for the variance of a spatially averaged operator. The 10^4 glueball correlator goes non-positive at tau 2, so no m_eff(1) is available there. 2++ tensor m_eff(0) is 3.3040 plus or minus 0.4013 on 10^4, against 3.3770 plus or minus 0.4166 on 8^4, consistent within errors.

The torelon behaves as a flux tube state should. Its m_eff(0) is 4.2692 plus or minus 0.0421 on 10^4, against 3.4004 plus or minus 0.0362 on 8^4, a large difference, but the torelon mass scales with the spatial extent: sigma times L_s predicts 3.148 on 10^4 and 2.533 on 8^4. The 10^4 m_eff(1) is 3.7823 plus or minus 2.1275, a 56 percent error, so the pick rule falls back to m_eff(0) for the 10^4 torelon estimate, and that m_eff(0) must not be compared directly with the 8^4 m_eff(1): m_eff(0) is an upper estimate carrying excited-state contamination, and the two volumes have different L_s. The 10^4 m_eff(1), noisy as it is, is consistent with the predicted 3.148. The volume check therefore confirms the torelon interpretation, a state whose mass grows with L_s, rather than a fixed glueball mass, and it shows the bulk observables are volume stable while the glueball m_eff(0) shifts downward by about 1.35 sigma on the larger volume. One volume pair at one beta does not control finite-volume effects: it is a single check, not a finite-volume extrapolation.

## Notes on the run itself

No code changes were made for round 2. The runner compute/run_r2.py calls the same run function in compute/wilson_lattice.py with beta 2.3 and size 8, then size 10, using the same seed convention, seed 20261120 plus int(beta times 100). The receipt ledger was not touched. The long wall time, 1683.6 seconds against 338.6 seconds for round 1, is machine load during this run, not a change in the protocol: the 8^4 measurement phase alone took 639.1 seconds of wall time here.
