partonmap
Study · 4.4

Closure tests and calibration

L0/L1 statistics derived — expected χ², coverage, the health identity — and why certificates are ratio-specific.

Chapters 4.1–4.3 built a beautiful machine. This chapter is about refusing to trust it. Before an astronomer points a new telescope at the sky, she points it at a source she already knows — a calibration star — and demands the instrument return the truth. Closure testing is exactly that discipline for a fitting pipeline: plant a known truth, run the full machinery blind, and check that it comes back — the right central values, the right $\chi^2$ statistics, bands with the advertised coverage. We derive the expected statistics at each rung (L0: noiseless data; L1: one noisy realization), including the subtle question of what "$\chi^2 \to 0$" can actually mean for a sampler, put error bars on coverage itself, and end with the two honest limits: why a certificate earned at one parameter count does not transfer to another, and what closure can never certify at all.

The logic of planted truth

Pick a truth vector $\theta^*$ inside the model family, generate pseudo-data from it, and fit with the identical pipeline — same priors, same sampler settings, same $\chi^2$, same termination — that will later run on real data. Because the truth is known, every output becomes checkable: the best-fit $\chi^2$ has a computable expected value, the posterior bands have a measurable coverage against $f(x;\theta^*)$, and any systematic deviation is a pipeline pathology by construction, not a physics surprise — there is no model mismatch to hide behind. The ladder ascends in realism: L0, noiseless data $D = T(\theta^*)$, isolates the optimizer and sampler; L1, one noisy draw $D = T(\theta^*) + \eta$ with $\eta \sim \mathcal N(0, C)$, adds the statistical fluctuations of a real experiment while keeping the truth in-family. Our current certificate at 52 parameters is PPDF-11: L0 $\chi^2/N = 0.0074$, L1 $\chi^2/N = 0.979$ — and this chapter derives what those two numbers had to be for the machinery to pass.

Derivation 1 — L0: zero in principle, floored in practice

With noiseless in-family data, $\chi^2(\theta) = \big(T(\theta^*) - T(\theta)\big)^{\mathsf T} C^{-1} \big(T(\theta^*) - T(\theta)\big) \ge 0$, with equality at $\theta = \theta^*$: the true minimum is exactly zero. So why did our nested-sampling L0 report a total $\chi^2_{\min} = 0.0074 \times 648 = 4.80$ rather than $10^{-8}$? Because a sampler's "best point" is its luckiest draw, not a polished optimum. Near the minimum the posterior is $\propto e^{-\chi^2/2}$ with $\chi^2 = \Delta^{\mathsf T} A \Delta$, so for a draw from the posterior, $\chi^2 \sim \chi^2_{d_{\text{eff}}}$ — a chi-square with one unit per constrained direction, mean $d_{\text{eff}}$ (§4.3's complexity, $\approx 13.7$ here). The best of $n$ effectively independent draws sits at the $\sim 1/n$ quantile, and the small-$c$ tail of a chi-square is

$$ P\big(\chi^2_{d} \le c\big) \;\approx\; \frac{(c/2)^{d/2}}{\Gamma(d/2 + 1)}\, e^{-c/2} $$

For $d = 13.7$, $c = 4.8$: $P \approx 10^{-2}$ — the observed floor is what the minimum over $\mathcal O(10^2)$ independent-equivalent posterior draws should be. The floor scales like the complexity, gently reduced by sample size: a sampler's L0 lands at $\mathcal O(C)$, orders of magnitude above machine zero, and that is correct behaviour. Gradient descent from the sampler's best point is a different animal — it polishes one point down the quadratic bowl to machine precision: our Hessian-GD refit reached $\chi^2 = 1.9\times10^{-8}$. Both numbers certify the same thing at their own resolution: the likelihood surface has its zero where the truth was planted. The gate is set accordingly: L0 passes on $\chi^2/N \ll 1$ (ours: 0.0074), not on $\chi^2 = 0$. $\blacksquare$

Derivation 2 — L1: $\mathbb E[\chi^2_{\min}] = N - d_{\text{eff}}$

Now $D = T(\theta^*) + \eta$, $\eta \sim \mathcal N(0, C)$. Whiten: $e = C^{-1/2}\eta \sim \mathcal N(0, \mathbb 1_N)$, and write the whitened residual at parameters $\theta$ as $e - \tilde J\,\delta\theta$ with $\tilde J = C^{-1/2} \partial T/\partial\theta$ (linearizing around the truth — legitimate because the fit lands near it, by L0). Minimizing over $\delta\theta$ projects $e$ onto the orthogonal complement of the model's tangent space, $\mathrm{col}(\tilde J)$, of dimension $d_{\text{eff}}$ (the constrained directions; flat directions contribute no tangent motion the data can see):

$$ \chi^2_{\min} = \big\| (\mathbb 1 - P)\, e \big\|^2, \qquad P = \text{projector onto } \mathrm{col}(\tilde J) $$

For a standard normal vector, $\|(\mathbb 1 - P)e\|^2 \sim \chi^2_{N - d_{\text{eff}}}$ exactly. Hence

$$ \mathbb E[\chi^2_{\min}] = N - d_{\text{eff}}, \qquad \operatorname{var}[\chi^2_{\min}] = 2\,(N - d_{\text{eff}}) $$

The fit "spends" one unit of $\chi^2$ per constrained direction fitting noise — the same $d_{\text{eff}}$ that appeared as complexity in §4.3, now counted from the other side. An L1 result far below $N - d_{\text{eff}}$ means overfitting or an inflated covariance; far above means the machinery failed to find the planted truth. $\blacksquare$

Derivation 3 — coverage, with error bars on the calibration itself

Coverage asks a sharper question than $\chi^2$: are the bands honest? Define pointwise band coverage: at each of $n$ probe points $x_j$, check whether the truth $f(x_j; \theta^*)$ lies inside the posterior 68% credible interval for $f(x_j)$; report the fraction $\hat p$. If the machinery is calibrated, each check is (marginally) a Bernoulli trial with $p = 0.68$, so the measured coverage carries its own binomial uncertainty:

$$ \hat p \;=\; p \;\pm\; \sqrt{\frac{p(1-p)}{n}} $$

— a calibration experiment with finite statistics. Vocabulary: calibrated, $\hat p$ consistent with 0.68; conservative, $\hat p$ significantly above (bands wider than advertised — inefficient but safe); anti-conservative, $\hat p$ significantly below (bands lie — the one unforgivable failure). Caveat that matters in practice: neighbouring $x_j$ share parameters, so the $n$ trials are strongly correlated and the effective $n$ is far smaller than the point count — the binomial bar is a lower bound on the real uncertainty. Worked Example 2 runs the numbers. $\blacksquare$

$$ \text{L0: } \chi^2_{\min} \to 0 \;(\text{floor} \sim \mathcal O(C)), \qquad \text{L1: } \mathbb E[\chi^2_{\min}] = N - d_{\text{eff}}, \;\; \sigma = \sqrt{2(N - d_{\text{eff}})} $$

In-family vs out-of-family truth

Everything above assumed the truth is representable: $\theta^*$ exists with $T(\theta^*)$ exact. Plant an out-of-family truth instead — pseudo-data from a different parametrization — and the two questions "does the sampler work?" and "can the model bend far enough?" collapse into one confounded test: an L1 excess could be sampler failure or model rigidity, and you cannot tell which. That is why our MSHT20 closure drills (PPDF-11) plant the truth with the model itself: it isolates the machinery, which is the thing a closure test can actually certify. Out-of-family studies are valuable but answer a different question (parametrization bias — the multi-model program of §4.3's averaging section), and mixing the two produces tests that neither certify the sampler nor measure the bias cleanly.

Certificates are ratio-specific: why 13-parameter closure did not cover 52

The same Colibri machinery that passed closure at 13 parameters was re-certified from scratch before the 52-parameter runs — and that was not bureaucracy. What closure certifies is the machinery in a posterior regime, and the regime is set by the information ratio $N/d$: at $648/13 \approx 50$ the posterior is a single compact quasi-Gaussian blob; at $648/52 \approx 12$, with only $\sim 14$ directions constrained (complexity 13.7, §4.3), the posterior is a 14-dimensional shell stretched along $\sim 38$ prior-dominated plateaus. Every hard part of §4.2 scales with that geometry: slice decorrelation lengths ($156 = 3\times52$ steps), live-point coverage of flat directions (the 300$\to$1200 population growth), termination on plateaus, weight concentration. A sampler can be flawless in the first regime and silently biased in the second — so the L0/L1 gates were re-run at 52 params, giving PPDF-11's 0.0074 and 0.979. Standing lab rule: new parameter count, new dataset, or new prior geometry $\Rightarrow$ new certificate. Ratios change; certificates do not transfer.

What closure cannot certify

A closure test plants truths the model can represent — so it is structurally incapable of telling you whether the model can represent the proton. That question — adequacy on real data — belongs to goodness-of-fit: compare the real-data $\chi^2_{\min}$ against the same yardstick $N - d_{\text{eff}}$ that L1 derived, now with nature supplying the truth. Our real-data run answers it: $\chi^2/N = 1.002$ (PPDF-12), against an expectation of $(N - d_{\text{eff}})/N \approx 0.98$ with spread $\pm 0.055$ — adequate. The division of labour is exact: closure certifies the instrument, goodness-of-fit judges the model, and the evidence (§4.3) ranks the models. Three questions, three statistics, never interchangeable: a pipeline that passes closure can still carry a wrong model (the toy's $\chi^2/N \approx 5.6$ says so at $+1453$'s expense), and a perfect real-data $\chi^2$ from an uncertified pipeline proves nothing, because uncalibrated machinery can reach good $\chi^2$ with dishonest bands.

Worked example 1 — scoring our own L1 certificate

Expected $\chi^2/N$ at $N = 648$, $d_{\text{eff}} = 11$ — vs the measured 0.979

Expectation: $\mathbb E[\chi^2_{\min}] = N - d_{\text{eff}} = 648 - 11 = 637$, so $\mathbb E[\chi^2/N] = 637/648 = 0.983$.

Spread: $\sigma(\chi^2_{\min}) = \sqrt{2 \times 637} = 35.7$, so $\sigma(\chi^2/N) = 35.7/648 = 0.055$.

Score the measurement: PPDF-11's L1 gave $\chi^2/N = 0.979$, i.e. $\chi^2_{\min} = 634.4$. Pull: $(634.4 - 637)/35.7 = -0.07\sigma$. As calibrated as a single realization can look.

Two lessons hiding in the arithmetic. First, the test is one draw from a distribution with $\sigma(\chi^2/N) = 0.055$: any single L1 between roughly $0.87$ and $1.09$ would also have passed, so one L1 can catch gross failures but cannot certify percent-level calibration — that is what coverage (many probe points) and repeated-realization L1 batteries add. Second, inverting the logic to measure $d_{\text{eff}}$ from one L1 gives $\hat d = N - 634.4 = 13.6 \pm 35.7$ — uselessly wide. The sharp estimate of constrained directions comes from the complexity $C = 13.7$ of the evidence run (§4.3), whose posterior average beats one noisy realization; the L1's job is only to be consistent with it, and (13.6 vs 11 vs 13.7, all within a fraction of $\sigma$) it is.

Worked example 2 — coverage arithmetic on 120 probe points

What "65–100% across flavours" does and does not mean

Binomial band: for $n = 120$ $x$-points and true coverage $p = 0.68$: $\sigma = \sqrt{0.68 \times 0.32 / 120} = \sqrt{0.001813} = 0.043$. So a calibrated pipeline should return measured coverage in $[63.7\%, 72.3\%]$ at $1\sigma$, $[59.5\%, 76.5\%]$ at $2\sigma$ — if the 120 checks were independent.

Score our spread: across flavours we measured 65% to 100%. The 65% flavour sits $0.7\sigma$ below 0.68 — textbook calibrated. The 100% flavour is formally $(1.00 - 0.68)/0.043 = 7.4\sigma$ high — but here the correlation caveat of Derivation 3 bites: adjacent $x$-points of one flavour's curve are driven by the same few parameters, so the effective number of independent checks is closer to the number of constrained directions feeding that flavour — order 10, not 120. With $n_{\text{eff}} = 10$, $\sigma = \sqrt{0.68 \times 0.32/10} = 0.148$, and 100% is a $2.2\sigma$ fluctuation on the conservative side.

Interpretation, stated the way the result card states it: no flavour is anti-conservative (nothing below the band) — the bands never lie downward; some flavours are conservative — bands wider than 68% where the data barely constrain them, which is the prior (honestly) doing the talking, §4.1's flat-direction regime. That pattern — calibrated where constrained, conservative where not — is precisely the signature of a healthy Bayesian pipeline with uninformative directions, and it is what PPDF-11's certificate records. The follow-up that would sharpen it from qualitative to quantitative: many-realization L1 batteries to measure $n_{\text{eff}}$ per flavour directly — logged as future work in the wiki.

Hold on to three things: (1) L0's floor is not zero and should not be — a sampler's best draw lands at $\mathcal O(C)$ above the true minimum (ours: 4.80 total, polished to $1.9\times10^{-8}$ by gradient descent; both pass), while L1 must land at $\mathbb E[\chi^2_{\min}] = N - d_{\text{eff}}$, and ours sits $0.07\sigma$ from it; (2) coverage is itself a measurement with (at least) binomial error bars, and correlated probe points make honest interpretation coarser than the raw point count suggests — calibrated-where-constrained, conservative-where-not is the healthy signature; (3) certificates are ratio-specific ($N/d$ sets the posterior regime — recertify at every new parameter count) and closure can never certify model adequacy: that is goodness-of-fit's job on real data ($\chi^2/N = 1.002$), and ranking models is the evidence's job (§4.3). The instrument is now calibrated end to end; Chapter 5 turns to the rival uncertainty philosophies — Hessian, replicas, Bayesian — and the theorem that reconciles them.