Fluctuation–Dissipation Diagnostics for Multi-Detector Joint Fits

When several particle-physics experiments are combined into one global fit, two things can go wrong, and today they are hard to tell apart. The sampler may not have explored the long, curved valleys of equally good parameter values that each detector leaves behind. Or one detector's model may be subtly wrong, so that the joint fit quietly drags the shared physics parameters toward a compromise instead of flagging a conflict. We propose a single-run toolkit that separates these failures. It is built on one structural fact. The Langevin sampler, a random walk pushed downhill on the negative log-posterior plus Gaussian noise, has a hidden supersymmetry: a fermionic symmetry of its path-integral description. That symmetry yields exact identities linking how the chain fluctuates to how it would respond if one detector's likelihood were turned up or down.

The idea in two sentences is this. Run one joint Langevin chain, and use the exact fluctuation–response identity to predict, with no refit, how each shared parameter would move if detector $k$ were reweighted or removed. Separately, check whether that same identity holds at every time lag: a violation means the chain has not converged, while a large predicted shift under a satisfied identity means the detectors genuinely disagree. The theoretical ingredient is the non-equilibrium extension of Parisi–Sourlas supersymmetry by Marguet et al.. The application regime is CEvNS (coherent elastic neutrino–nucleus scattering), dark-matter, and oscillation fits, where target and channel complementarity is the whole point of fusion.

Core Claim

The central mechanism of supersymmetric localization is that adding a "$Q$-exact" term, a term that is a total derivative under the fermionic symmetry $Q$, cannot change the integral. We take this lesson literally in the one place it applies exactly to data analysis: the path integral of Langevin sampling. Consider the sampler

$$ d\theta = -(I + \gamma J)\,\nabla U(\theta)\,dt + \sqrt{2}\,dW_t, \qquad U(\theta) = -\log \pi(\theta) - \sum_k \log \mathcal{L}_k(\mathcal{D}_k \mid \theta). $$

Here $\theta$ collects the shared physics parameters and the nuisance parameters. $\pi$ is the prior, $\mathcal{L}_k$ is detector $k$'s likelihood, $J$ is any antisymmetric matrix, $\gamma \ge 0$ sets its strength, and $W_t$ is Brownian motion. For every $J$ and $\gamma$ the stationary distribution is exactly the joint posterior, so these drifts are the sampler's analogue of $Q$-exact deformations: they change the dynamics but not the answer. For $\gamma > 0$ the dynamics is irreversible. The classic fluctuation–dissipation relation then fails, but the modified relation derived by Marguet et al. holds.

That paper's main stated limitation is that its construction requires the steady state to be known analytically. In Bayesian sampling the steady state is known, as the unnormalized posterior, so the limitation disappears in exactly this setting. That observation is what makes the proposal possible. It also means the theory is immediately usable for detector fusion, a use that paper never considers.

We then use two kinds of perturbation. For a non-exact perturbation, $U \to U - h\,\log \mathcal{L}_k$, the relation gives the response of $\langle\theta(t+\tau)\rangle$ to a small pulse of $h$ at time $t$. It does so as a correlation function between $\theta(t+\tau)$ and an explicitly known observable at time $t$, built from $\log\mathcal{L}_k$, the drift, and the stationary density. Integrating over lags gives the static influence $\partial\langle\theta\rangle/\partial h_k$, which reduces to $\mathrm{Cov}(\theta, \log\mathcal{L}_k)$. That quantity is the linearized "what if detector $k$ counted more or less" shift. The time-resolved identity, by contrast, holds only if the chain is stationary. The gap between the measured response and the response predicted from correlations is therefore a lag-by-lag convergence test. It is most sensitive to the slow modes along degeneracy valleys that summary statistics such as $\hat R$ average away.

The claim is that these two outputs together make a joint fit self-auditing from one run. We estimate the measured response directly with paired perturbed chains that share random numbers, which costs little. The first output is a detector-influence matrix; the second is a stationarity residual. A nonzero residual means "keep sampling". A satisfied identity combined with a large, predictive-inconsistent influence from one detector means "model tension". The file identifies this last case as the unsolved failure mode of joint likelihood minimization.

Research Program

The first component is theoretical and small. We will specialize the Marguet et al. identity to drifts of the form $-(I+\gamma J)\nabla U$ with $U$ a sum of detector terms, and write the conjugate observable for each detector explicitly. We will also derive its discrete-time counterpart for the Euler–Maruyama integrator, including the step-size bias that must be subtracted. The paper covers overdamped dynamics with additive white noise only, so we restrict to that class. Inertial samplers such as Hamiltonian Monte Carlo are out of scope for the identity, though not for comparison.

The second component chooses $J$. Each single detector's Fisher information picks out its nearly flat directions, its degeneracy valley. We will choose $J$ to rotate probability flow within the span of those valleys. The aim is not to beat every sampler; the aim is to make the irreversible chain mix along exactly the directions where the diagnostics matter.

The third component is software. It will be a JAX implementation on top of the stochastic-gradient Langevin code in srshtiputcha/sgmcmc_preferential_subsampling, exposing per-detector log-likelihoods, the influence matrix, and residual curves.

The role of the other directions should be stated honestly. Localization supplies the organizing principle, that exact deformations leave answers invariant while non-exact ones are measured by response. Its literal use, sending the deformation strength to infinity, would here reduce to a Laplace approximation, which fails on degenerate valleys; we do not pursue it. The pure spinor work of dos Santos, Guillen and Viana is a methodological precedent rather than a component. It shows that adding auxiliary variables which leave the physical content unchanged (their non-minimal pairs) makes constructible an operator that is provably impossible without them. Our analogue is the auxiliary response and ghost fields of the Langevin path integral: they leave the posterior unchanged but make response estimators available from a single equilibrium run.

The complexity-of-QFT direction frames the payoff. Baggio, Niarchos and Papadodimas reduce correlators to derivatives of one free energy, avoiding a sum over diagrams. In the same way, we reduce $K$ leave-one-detector-out refits to correlations within one chain.

Baselines

The naive reference point is separate per-detector fits combined by inverse-variance averaging. The file records why this fails on non-Gaussian valleys (T2K and NOvA joint analysis). We include it to show the size of the fusion effect on each benchmark. It is run on the identical binned data and priors.

The strongest method for the influence question is exact leave-one-detector-out refitting. The joint fit is simply rerun with detector $k$ removed, as the COHERENT collaboration effectively did when it reported argon alone against argon plus CsI (COHERENT argon+CsI). That paper gives $\sin^2\theta_W = 0.31 \pm 0.06$ from argon alone and $0.26 \pm 0.03$ combined. Removing CsI is a finite perturbation ($h=-1$), whereas our estimate is first-order. We will therefore report both our linear prediction and a second-order correction from third cumulants. We count it a win only if the prediction lands within the refit's statistical error while costing one chain instead of $K+1$. We give both methods the same total gradient evaluations.

The strongest method for sampling is a tuned No-U-Turn Hamiltonian Monte Carlo sampler. We add a flow-matching sampler using the continuous-normalizing-flow code in emil-ho/bg-cnf-fm-condensed-matter, and a random-walk Metropolis sampler representative of MCMC frameworks like ARIA and MaCh3. All are compared at matched gradient or likelihood evaluations, with equal tuning budgets, on effective sample size (ESS) per evaluation of the shared parameters. No reported sampler-efficiency result was found on Lacuna for any of these fits; the file reports physics results only.

For the convergence question, the baseline is split-$\hat R$ plus effective sample size. For the tension question, the baseline is per-detector posterior predictive p-values.

Two ablations isolate the key ingredient. The first sets $\gamma = 0$: reversible dynamics, where the standard fluctuation–dissipation relation and ordinary Stein-type zero-mean identities apply. This tests whether the non-equilibrium identity adds anything beyond equilibrium diagnostics. The second uses a random $J$ rather than a valley-aligned one, to test whether alignment matters.

Why not just…?

Why not just run NUTS and stop? On raw efficiency this is a real threat, and we expect Hamiltonian Monte Carlo to match or beat irreversible Langevin on smooth posteriors. The proposal does not rest on speed. Its contribution is the pair of single-run diagnostics, and those diagnostics require overdamped dynamics, for which the identity is proven. If the ESS comparison goes against us, we will recommend NUTS for production chains and a short irreversible-Langevin audit chain for diagnostics. We will report whether that audit costs under a stated fraction, such as 20%, of the production budget.

Why not just refit leaving each detector out? Refitting is exact and is the gold standard. It costs $K$ extra fits, which matters when $K$ counts detectors and systematic blocks such as the Germanium quenching-factor model in the Dresden-II + COHERENT analysis. There, extracted parameters depend strongly on which model is chosen. Its weakness for our method is nonlinearity: removing a dominant detector is not a small perturbation. The COHERENT argon/CsI check below settles this directly. If the linear prediction misses by more than the refit error, we restrict the claim to reweighting and systematic-block perturbations.

Why not just use $\hat R$, ESS, and Stein control variates? These are equilibrium, one-time statistics; the Stein identities are exactly the $\gamma=0$ Ward identities. They cannot, in principle, see a chain that is stationary in its fast directions but still drifting along a valley at long lags. The $\gamma=0$ ablation, run against injected slow-mixing failures, will show whether the two-time residual detects failures earlier at matched compute. If it does not, the non-equilibrium ingredient is unnecessary, and we will say so.

Why not apply supersymmetric localization to the posterior itself? A posterior has no fermionic symmetry beyond the one in its sampling dynamics. Sending a localizing term to infinity collapses onto the maximum-likelihood point with a Gaussian correction, which is the Laplace approximation. That is precisely what fails on the open degeneracy contours described by Pato et al. and Strege et al..

Proposed Experiments

The first testbed is a synthetic multi-target dark-matter fit (xenon, germanium, argon) in the $(m_\chi, \sigma)$ plane, following Pato et al. and Strege et al. The truth is known. We will inject two kinds of failure: deliberately under-run chains, and a mis-modeled germanium energy scale. The decisive checks are these. The stationarity residual must flag the under-run chains at fewer evaluations than $\hat R$. It must stay flat for correctly converged chains with the injected germanium bias. At the same time, the influence matrix must single out germanium as the detector pulling the mass away from truth.

The second testbed is COHERENT argon plus CsI, with the binned spectra used in the cited analysis. From the combined chain we predict the argon-only $\sin^2\theta_W$. The decisive target is the reported shift from $0.26 \pm 0.03$ to $0.31 \pm 0.06$. The linear prediction, plus a second-order correction, must reproduce the $\approx 0.05$ shift within the refit's uncertainty.

The third testbed adds Dresden-II germanium and treats the quenching-factor model as a continuous interpolation parameter. The influence matrix should reproduce the model dependence of $\sin^2\theta_W$ that the reanalysis found by swapping models. The residual should confirm that differences between models are physical rather than sampling artifacts.

The fourth testbed is T2K+NOvA. Full collaboration likelihoods are not public, so we will build a surrogate from published-scale oscillation probabilities with correlated flux and cross-section nuisances. The aim is to test valley-aligned $J$ on the $\delta_{CP}$–mass-ordering degeneracy. The reported normal-ordering result $\Delta m^2_{32} = 2.43^{+0.04}_{-0.03}\times 10^{-3}\,\mathrm{eV}^2$ (T2K+NOvA) serves only as a sanity anchor for the surrogate, not as a comparison target.

Across all four testbeds we report ESS per gradient against NUTS, flow matching, and random-walk Metropolis. We also report influence-prediction error against exact refits, and the lag at which the residual first exceeds its null band compared with when $\hat R$ first exceeds 1.01.