# MQG test 21: which way does a collective-neutrino system leave an unstable point? A choice law and slaving, fixed in advance (public plan)

Ren Maxwell Chen. Version 1.0, 19 November 2026. Files, SHA-256 values and the smoke-test record are in Section 10.

## 1. Question

At the bipolar flavor instability of the two-beam model the system must leave the unstable point in some direction. A small tilt of the vacuum field, set by the mixing angle θ, favours one side, called "+" below. The quantum spread of the initial state favours no side. Which side does the system choose, and with what probability?

Tests 19 and 20 found that the seed variable x = sin 2θ √(N/2) decides how the sensitivity to θ scales with N. This test asks whether the same variable, with a factor fixed in advance by linear theory, decides the probability of the choice. It also asks whether the second beam is slaved to the first, with an amplitude ratio and phase fixed by the same theory.

## 2. Linear theory and the frozen predictions

**Model.** As in tests 19 and 20:
- H = B·(J_A − J_B) + (μ/N)(J² − 3N/4), with B = (sin 2θ, 0, cos 2θ), unstable ordering (b = +1), and beams of spin j = N/4 starting along +z and −z.
- Mean-field equations: ȧ = [B + (μ/2)(a + b)] × a and ḃ = [−B + (μ/2)(a + b)] × b, with J_A = j a and J_B = j b.

**Linearization.** Write ε = sin 2θ, α = 1 − μ/2, β = μ/2 and λ = √(μ − 1). Linearize about the fixed point a* = B and b* = −B, with u = a_x + i a_y and v = b_x + i b_y:
- δu̇ = i(α δu − β δv) and δv̇ = i(β δu − α δv).
- The growth rates are ±λ.
- The growing mode has right eigenvector r = (1, (α + iλ)/β) and left eigenvector l = (1, −(α + iλ)/β), with |l₂| = 1.

**Initial displacement from the fixed point.** δu(0) = −ε + ξ_A and δv(0) = +ε + ξ_B. The noises ξ_A and ξ_B are independent circular Gaussians with standard deviation σ = √(2/N) per real component (the transverse spread of a coherent spin state).

**Order-parameter amplitude.** A ∝ l·δw(0):
- deterministic part −ε(1 + iλ)/β, of modulus 2ε/√μ;
- noise variance 2σ² per real component.

**Prediction 1, the choice law.** The "+" side is the side of the deterministic part. Its direction in the u plane is

  φ₊ = arg[−(1 + iλ)/(λ − iα)] = −125.264°, −135.000° and −140.768° at μ = 1.5, 2.0 and 2.5.

The probability of leaving on the "+" side is

  **P₊ = Φ(κx),  κ = √(2/μ)**  (κ = 1.1547, 1.0000 and 0.8944),

where Φ is the standard normal distribution function.

**Prediction 2, slaving.** The partner mode decays as e^{−λt}, so once the system has left the unstable point the second beam follows the first: δv/δu → (α + iλ)/β, with modulus 1 and phase

  ψ = atan2(λ, α) = 70.529°, 90.000° and 101.537° at μ = 1.5, 2.0 and 2.5.

These values are printed by mqg_test21_predict.py, deposited with this plan (predictor_output.txt).

## 3. Points

- θ = 0.005, 0.015 and 0.03; N = 640, 1280, 2560 and 5120; μ = 1.5, 2.0 and 2.5; unstable ordering.
- 36 rows, MQG21-01 to 36, in the order θ, then N, then μ. Row k uses TWA seed 20272100 + k.
- No quantity has been computed at any of these points for any test.

| θ | N = 640 | 1280 | 2560 | 5120 |
|---|---|---|---|---|
| 0.005 | 0.179 | 0.253 | 0.358 | 0.506 |
| 0.015 | 0.537 | 0.759 | 1.073 | 1.518 |
| 0.03 | 1.073 | 1.517 | 2.145 | 3.034 |

**Collapse pairs** (same x, different θ and N): (0.015, 2560) with (0.03, 640), and (0.015, 5120) with (0.03, 1280), at each μ.

## 4. Observables (exact, for each row)

- **Displacement.** δu = (J_{A,x} + i J_{A,y})/j − ε, where the components are taken in the flavor frame.
- **Measurement time.** t* is the first time on a grid of step 0.02 at which the exact ⟨|δu|²⟩ = ⟨(J_{A,x}/j − ε)² + (J_{A,y}/j)²⟩ reaches 0.1. If it is not reached by t = 8, the row is "censored".
- **Choice probability.** P₊ = Prob(n·J_A > c) in the exact state at t*, with n = (cos φ₊, sin φ₊, 0) and c = j ε cos φ₊. It is computed from the eigen-decomposition of n·J_A (eigenvalues m = −j, …, j) by rotating beam A. If an eigenvalue lies within 10⁻⁹ of c, it counts with weight ½.
- **Slaving ratio.** r = ⟨δJ_{B,+} δJ_{A,−}⟩ / (j² ⟨|δu|²⟩) at t*, with δJ_{A,−} = J_{A,x} − jε − i J_{A,y} and δJ_{B,+} = J_{B,x} + jε + i J_{B,y}. The denominator is the same symmetric quantity that defines t*, so no operator-ordering term enters it. (The product δJ_{A,+} δJ_{A,−} would add ⟨J_{A,z}⟩ ≈ j, a shift of order 1/(0.1 j) in |r|.) The numerator needs no ordering, since operators of different beams commute.
- **TWA, for information.** The same P₊ and r from M = 8000 mean-field trajectories with the same initial sampling as tests 19 and 20 (RK4, step 0.01), at the TWA's own t*. Reported with the binomial standard error.

## 5. Rules (fixed before any data)

The tolerances come from a sandbox at non-test points (sandbox_bias.py, deposited as background): there, the TWA values of P₊ followed Φ(κx) to within 0.009 (mean +0.003); collapse pairs agreed to within 0.002; |r| = 0.998; and arg r lay 1°–3.5° below ψ. The exact runner at six non-test points (θ = 0.012 and 0.025, N = 1000; sandbox_exact.py) gave |P₊ − Φ(κx)| ≤ 0.002, |r| = 0.998 and arg r 0.2°–2.5° below ψ. The tolerances were not changed after it.

**Rule K (choice law).**
- Met if |P₊ − Φ(κx)| ≤ 0.02 at all 36 rows.
- Not met otherwise.
- Not evaluable if any row is censored.

**Censored rows (all rules).** A censored row has no P₊ and no r. Every rule that uses a censored row is "not evaluable": K and S if any row is censored; E if a deciding row is censored; D if a θ = 0.005 row is censored; C if a member of a collapse pair is censored. Censored rows are listed in the results.

**Rule E (rival: bias over noise only, κ = 1).** Decided only at the 7 rows where |Φ(κx) − Φ(x)| ≥ 0.025 (rows 10, 13, 16, 19, 21, 25 and 27, all at μ = 1.5 or 2.5, as listed by the predictor).
- A deciding row is a tie if the two distances |P₊ − Φ(κx)| and |P₊ − Φ(x)| differ by less than 0.002. Ties count for neither law.
- K favoured over E if at least 5 of the 7 deciding rows are not ties and Φ(κx) is the closer law at every one of them.
- E favoured if, under the same condition, Φ(x) is the closer law at every non-tie row.
- Otherwise neither.

**Rule D (rival: deterministic choice).**
- D is favoured if P₊ ≥ 0.99 at all 12 rows with θ = 0.005.
- D is rejected if P₊ ≤ 0.90 at all of them.
- Otherwise neither.

**Rule C (collapse).**
- Met if the two collapse pairs agree to within 0.01 at every μ (6 comparisons).
- Not met otherwise.

**Rule S (slaving).**
- Met if ||r| − 1| ≤ 0.03 and |arg r − ψ| ≤ 6° at all 36 rows.
- Not met otherwise.

**Information only (Rule Q).** P₊(exact) − P₊(TWA) at every row, with the TWA standard error. A difference beyond three standard errors at small x (x < 0.6: the 12 rows at θ = 0.005 and the 3 rows at θ = 0.015, N = 640) would be reported as a candidate quantum signature, to be tested in a later plan. It carries no verdict here.

**Void.** If a guard fails, the run is void; the fault is repaired in a new Zenodo version and the run repeated.

## 6. Guards

- **V1.** The propagation reproduces the Stage-4 reference solver (mqg_step2_robustness.py) at θ = 0.005, N = 640, μ = 2.0, t ≤ 5, to within 10⁻⁸ (relative). This runs only inside the RUN, after deposit; the smoke test runs it at θ = 0.01, N = 200, μ = 1.8, which is not a test point.
- **V2 (rotation).** At θ = 0.30 (not a test angle), for beam A at j = 160 and j = 1280: the distribution of n·J_A sums to 1 within 10⁻¹⁰; for n along the block z axis it reproduces the J_{A,z} distribution of the block solution within 10⁻¹⁰; for n along the flavor z axis at t = 0 it gives P(m = j) = 1 within 10⁻¹⁰ (frame check); and the eigenvalues equal m = −j, …, j within 10⁻⁸. The sum and eigenvalue checks are also applied in every row.
- **V3.** Norms are conserved to within 10⁻¹⁰.
- **V4.** With the same 8000 initial samples (same seed) for both steps, halving the TWA step from 0.01 to 0.005 changes the TWA P₊ at θ = 0.005, N = 640, μ = 2.0 by less than 0.002 (16 trajectories). Common samples remove the sampling noise from the comparison, so V4 measures only the step error.

## 7. What each outcome would mean

- **K, S and C met.** The choice and the slaving are fixed by linear theory with one variable: x, with the factor √(2/μ) from the noise of both beams. The choice is an imperfect bifurcation decided by the competition between the tilt and the quantum spread.
- **K met and E favoured.** Both laws lie within the tolerance of K, but Φ(x) is the closer one at the rows that can tell them apart. The verdict follows E; the pass of K is reported as a pass within tolerance only.
- **K not met, E favoured.** The bias-over-noise ratio alone sets the choice, and the second beam's noise does not enter as predicted.
- **D favoured.** The tilt alone decides, and the quantum spread plays no role in the choice.
- **C not met.** x is not the right variable for the choice.
- **S not met.** The second beam is not slaved as predicted.

## 8. Prior art

- H. Haken, *Synergetics: An Introduction* (Springer, Berlin, 1977), doi:10.1007/978-3-642-96363-6; and *Advanced Synergetics* (Springer Series in Synergetics, Vol. 20, Springer, Berlin, 1983): order parameters and the slaving principle.
- D. K. Kondepudi and G. W. Nelson, Physica A 125, 465 (1984), and Nature 314, 438 (1985), doi:10.1038/314438a0: at a noisy imperfect bifurcation, a small systematic bias sets the selection probability through its ratio to the noise. Rule K applies the same idea to a quantum spread with two noisy beams.
- Collective-neutrino flavor instabilities and their many-body treatment: as cited in the plans of tests 19 and 20 (Martin et al. 2022; Lacroix et al. 2022 and 2024).

## 9. Disclosures

- **Computed before deposit:** no quantity at a test point. The exact block code was checked against an independent dense calculation at N = 8 and 12 (dense_check.py; agreement 2 × 10⁻¹⁴), and run at six non-test points (θ = 0.012 and 0.025, N = 1000; sandbox_exact.py). The predictor uses analytic expressions only. The sandbox runs (TWA) used θ = 0.004, 0.007, 0.012, 0.025 and 0.04 with N = 1000 and 3000, none of which is a test point.
- **Design history.** The middle angle was changed from 0.01 to 0.015 before any computation, to put more rows where the two choice laws differ. At μ = 2.0 the two laws coincide (κ = 1), so those rows test Rule K but cannot decide between K and E.
- **Review changes (rev. 2), all before any computation at a test point.** After review by a second advisor: the denominator of r was matched to the t* quantity; V4 now uses common samples; rules for censored rows, a tie zone for Rule E and the meaning of "small x" were added; the references were checked.
- **The test is numerical.** It concerns the model, not neutrinos in nature.

## 10. Files, hashes and smoke test

All files are in this record (v1.0). SHA-256 values are in SHA256SUMS.txt; the main ones are:

- mqg_test21_runner.py: 2e5d6507edd14a13fdec553078224d3b0c5ac914ce1fc08aa1f4510411bd2eeb
- mqg_test21_score.py: a2928ae7914a9711ef1b277055fb87f87dceb8e1f9b6a8f215b1d5158473834e
- mqg_test21_predict.py: 2666e6c2c9e05047af1cbe1bd6c3179cffc48fe5ab4d26833e8a2d362b92eb3b (output: predictor_output.txt)
- mqg_step2_robustness.py (Stage-4 reference for V1): cb13dd79a8fc74be6a1a89799587f1362c4c45a96f27e67be821b636a0dc96d0

**Other files.** MQG_TEST_21_COLAB_CELLS.md (the RUN and VERIFY cells); background checks at non-test points or on made-up rows: sandbox_kappa.py, sandbox_bias.py, sandbox_exact.py with sandbox_exact_output.txt, dense_check.py with dense_check_output.txt, synth21.py with synth21_output.txt (15 made-up outcome cases, all scored as intended); smoke_log.txt.

**Smoke test.** Run on Colab on 19 November 2026 (Python 3.13.18, NumPy 2.4.6) at θ = 0.01, N = 200, μ = 1.8: V1 agreement 2.6 × 10⁻¹⁴, V2 ok, norm deviation 1.2 × 10⁻¹⁵. No P₊ or r was printed. After it, the runner was changed in one line only: the required NumPy version, from 2.1.3 to 2.4.6.

**Software.** The RUN uses NumPy 2.4.6 (the runner refuses any other version) and the SciPy of the Colab session, which the RUN cell prints.
