When physicists or machine-learning models learn a reduced description of a complex system (a few slow coordinates, plus a map of how strongly the system dissipates energy and how strongly it circulates at each point), the result is almost always a point estimate. Nobody attaches honest error bars. This matters most where data are scarce: tens to hundreds of expensive trajectories, statistics built from squared displacements, which are strongly skewed, and one-sided questions such as "is dissipation bounded away from zero in this region?" In that regime, bell-curve error bars are systematically wrong. This note proposes a way to build finite-sample-accurate confidence bands for these learned fields. The key is to exploit a symmetry that physics hands us for free.
The symmetry comes from a result by Carbone et al.. They show that Onsager's reciprocal relations (the rule that cross-couplings between fluxes are symmetric or antisymmetric) do not need the microscopic motion to be reversible. It is enough for the statistical ensemble, the distribution over system states, to be unchanged under a group of "generalized time reversals": time reversal combined with reflections, rotations, or spin flips. For charged particles in a fixed magnetic field, infinitely many such operations exist. Our idea is to use this invariance twice. First, it splits the estimated field into two pieces. One piece is symmetry-breaking and has an exactly symmetric sampling distribution, so it can be tested exactly with random sign flips and no approximation. The other piece is symmetry-preserving; it carries all the skewness and is where a higher-order correction to the bell curve (an Edgeworth expansion) is needed. Second, the symmetry ties the skewness at one state to the skewness at its mirror images. We can therefore pool those estimates, which makes the correction reliable at sample sizes where it would otherwise be swamped by noise.
Core Claim
Consider a nonparametric estimator $\hat A(z)$ of the state-dependent transport matrix on reduced coordinates $z$. It is obtained by regressing short-lag trajectory increments on the current state. Its symmetric part plays the role of the dissipation matrix $M(z)$ and its antisymmetric part the conservative matrix $W(z)$ of OnsagerNet. Suppose the stationary ensemble is invariant under a finite group $G$ of generalized involutions acting on $z$ through matrices $\rho(g)$. Then three things hold.
The first concerns the components of $\hat A$ lying in non-trivial symmetry sectors. These are the parts that would encode a violation of reciprocity or of the spatial symmetry. Their sampling distribution is symmetric about zero at every sample size, not just asymptotically. So all their odd cumulants vanish, and a max-statistic sign-flip test over independent trajectories is exact.
The second concerns the studentized symmetry-preserving component $T_n(z)$. It obeys a one-term Edgeworth expansion, which is valid for weakly dependent trajectory data by Edgeworth expansions for weakly dependent random variables:
$$P\big(T_n(z)\le x\big)=\Phi(x)+n^{-1/2}\,p_1\big(x;\kappa_3(z)\big)\,\phi(x)+O(n^{-1}).$$
Here $n$ is the number of independent trajectories, $\Phi$ and $\phi$ are the standard normal distribution and density, and $p_1$ is the usual quadratic polynomial in $x$ whose coefficients depend on the third cumulant (skewness) tensor $\kappa_3(z)$ of the increment statistic.
The third is that ensemble invariance forces
$$\kappa_3(gz)=\rho(g)^{\otimes 3}\,\kappa_3(z)\quad\text{for all } g\in G,$$
up to a sign for time-reversing elements. As a result, cumulant components that are odd under the subgroup fixing $z$ vanish outright, and the remaining ones can be averaged over the orbit of up to $|G|$ mirror-image points.
Together these give the method. We decompose the estimated field by symmetry sector, handle the symmetry-breaking sectors with exact sign-flip inference, and correct the symmetry-preserving sector with an Edgeworth expansion whose skewness estimates are pooled over orbits. One-sided and uniform coverage error should then fall from order $n^{-1/2}$ to order $n^{-1}$. The rate follows from removing the $n^{-1/2}$ skewness term in the expansion above. Pooling is what makes the correction worth its cost at small $n$.
Relation to Prior Work
Three lines of work motivate the gap. OnsagerNet and its stochastic extension S-OnsagerNet learn potentials, dissipation matrices, and noise tensors from trajectories, with positivity and skew-symmetry built in. They report point estimates only, with no confidence statements. Invertible coarse-graining guarantees Boltzmann consistency of generated ensembles but likewise has no finite-sample error characterization. Symmetry-driven variable discovery uses approximate Lie symmetries to find slow coordinates, not to do inference. We take coordinates from these pipelines as given and add inference on the fields defined over them.
Two works supply the higher-order method and also constrain it. Whang et al. build higher-order uniform bands for a nonparametric function and show the classical extreme-value (Gumbel) band is too conservative. Their method needs artificial undersmoothing (a deliberately small smoothing bandwidth) and works only for fewer than four evaluation dimensions. We inherit both limits, and we mitigate the variance cost of undersmoothing through orbit averaging. Gao et al. use a two-term Edgeworth correction for a scalar statistic comparing generative models. Our object is a matrix-valued field, and we use symmetry to decide where the correction is needed at all. Brown et al. and Haeusler et al. establish the coverage-error rates we target, for simpler statistics. Rearranging Edgeworth–Cornish–Fisher expansions supplies the fix for non-monotone corrected distribution functions, which we reuse directly.
The file's open problem on cumulant noise is a real negative constraint: estimated third cumulants can be noisy enough to cancel the benefit of the correction. Orbit pooling is aimed squarely at it. If pooling does not reduce that noise enough, the proposal fails.
To our knowledge, no prior work uses ensemble invariance to route different parts of an estimator to different inference procedures. Existing physics-informed learners use symmetry only as a constraint on the point estimate.
Research Program
The theoretical part has three pieces. First, we derive the symmetry-sector decomposition for the groups in Carbone et al.'s Theorem 3.10 (rotation-reflection involutions in a constant field) and for simple reflection groups. We then prove exact distributional symmetry of the symmetry-breaking sectors when trajectories are independent and start from the invariant ensemble. Second, we prove the pointwise and uniform Edgeworth expansion for the symmetry-preserving sector of a series or kernel regression estimator on a reduced space of dimension two or three. We follow Cohen et al.'s stability condition for the basis size and use the mixing-based validity results for within-trajectory dependence. Third, we quantify the variance reduction in the pooled cumulant estimate as a function of the orbit size and of how close $z$ sits to points fixed by the group. Near those fixed points pooling gains less, but more cumulants vanish exactly.
The practical output is a routine that takes three inputs: trajectories, learned coordinates, and a candidate group. It returns exact sign-flip bands for the symmetry-breaking sectors and Edgeworth-corrected, rearranged uniform bands for the symmetry-preserving sector. It also reports a sign-flip test of the symmetry assumption itself, so that pooling is never applied blindly.
We also state a known weak point. Where the true dissipation matrix is nearly singular, the positivity constraint creates a boundary where expansions break down. We will flag those regions rather than claim coverage there.
Baselines
The naive reference point is the first-order Gaussian band. It uses a pointwise studentized estimate, a Gaussian multiplier approximation for the supremum, and no symmetry use. This is what one would attach to an S-OnsagerNet-style estimate today. A classical Gumbel extreme-value band is included as the conservative reference that Whang et al. found too wide.
The strongest general competitor is a block multiplier bootstrap of the studentized sup-statistic. Resampling whole trajectories respects within-trajectory dependence, and bootstrap methods are known to capture skewness to second order in simpler settings. The strongest symmetry-aware competitor is the same bootstrap applied after augmenting the data with every group image of each trajectory. That tests whether "symmetrize, then bootstrap" already delivers everything we claim.
We also run deep ensembles of S-OnsagerNet as the practice-based machine-learning uncertainty reference. Ablations isolate the key ingredients: Edgeworth correction without orbit pooling, orbit pooling with a plain Gaussian band, and the full method without the sector split (Edgeworth applied to all components).
All methods get the same trajectories, the same learned coordinates, the same basis and bandwidth grid chosen by one shared rule, and 2000 bootstrap replicates where resampling is used. The file contains no reported coverage result for confidence bands on learned dissipation or transport fields. The closest reported numbers on these benchmarks are point-prediction errors, such as normalized error 0.021 for FNO-3D on Rayleigh-Bénard convection in Straat et al., which do not measure inference.
Beating a baseline therefore means two things across the small-sample range. The method must hold uniform and one-sided coverage within Monte Carlo error of nominal where the competitor visibly does not. It must also produce bands no wider than the symmetrized bootstrap, and the log-log slope of coverage error against $n$ must be near $-1$ rather than $-1/2$.
Why not just…?
Why not just use the bootstrap? This is the serious threat. For studentized statistics from independent units, the bootstrap is already second-order accurate, and the symmetrized bootstrap uses the same group information we do. Two things could still separate the methods. With few trajectories, resampling a handful of units gives a lumpy, discrete bootstrap distribution in the tails, while pooled analytic cumulants are smooth. Within-trajectory dependence also degrades block bootstrap accuracy. Neither advantage is guaranteed. We will state the outcome plainly. If the symmetrized bootstrap matches us at every $n$, the contribution narrows to two parts: the sector theory, which says which quantities need no approximation at all, and the exact sign-flip procedure.
Why not just use sign-flip randomization everywhere, since it is exact? It is exact only for statistics that the symmetry maps to their own negatives. Under ensemble invariance, only the symmetry-breaking sectors qualify. The dissipation field itself, which is positive and nonzero, has no symmetric null, so randomization cannot give its confidence band. This is precisely why the method splits the estimator rather than using one tool.
Why not just enforce the symmetry in the network and report ensemble spread? Building symmetry into OnsagerNet-type models improves the point estimate. The spread of a deep ensemble, however, reflects optimization randomness, not sampling variability, and it has no coverage guarantee. The experiments measure its coverage directly against ground truth, so this becomes an empirical question rather than an argument.
Proposed Experiments
We run four systems. All but the negative control allow ground truth from long reference runs, and all allow thousands of independent replicate datasets for measuring coverage.
The first is a controlled system: charged Brownian particles in a two-dimensional trap under a constant magnetic field, with state-dependent friction. Here the generalized reversal group and the relation $D_{xy}=-D_{yx}$ from Carbone et al. are known exactly. We vary the number of trajectories from 20 to 2000.
The second is the 900-degree-of-freedom polymer-stretching system with three learned coordinates from S-OnsagerNet, whose reflection symmetry gives $|G|=2$.
The third is two-dimensional Rayleigh-Bénard convection simulated with the Shenfun solver at $Ra$ from $10^5$ to $5\times10^6$, as in Straat et al.. We use the left-right reflection symmetry. Bands are evaluated on two-dimensional slices of the 11-dimensional OnsagerNet reduced space, because the method is limited to fewer than four evaluation dimensions.
The fourth is alanine dipeptide, where the only group is plain time reversal. This is a negative control where we predict little gain from pooling.
Metrics are uniform and one-sided coverage at 90% and 95%, the slope of coverage error against $n$, band width, the variance of cumulant estimates with and without pooling, and the size and power of the sign-flip test.
Three checks decide the claim. The coverage-error slope must steepen from about $-1/2$ to about $-1$ in the first three systems. The full method must beat the unpooled Edgeworth ablation at the smallest $n$, which shows pooling is what makes the correction pay off. And on a deliberately symmetry-broken variant of the first system, created by adding a weak perturbation that violates the invariance, the sign-flip test must detect the violation before pooling-induced bias corrupts the bands. If the symmetrized bootstrap ties us on all three systems, we report that as the finding.
References
- Time reversal symmetry for classical, non-relativistic quantum and spin systems in presence of magnetic fields
- OnsagerNet: Learning Stable and Interpretable Dynamics using a Generalized Onsager Principle
- Edgeworth expansions for weakly dependent random variables
- Constructing custom thermodynamics using deep learning
- Ensuring thermodynamic consistency with invertible coarse-graining
- Symmetry-Driven Discovery of Dynamical Variables in Molecular Simulations
- Doubly robust uniform confidence band for the conditional average treatment effect function
- Statistical Inference for Generative Model Comparison
- Confidence Intervals for a binomial proportion and asymptotic expansions
- Assessing confidence intervals for the tail index by Edgeworth expansions for the Hill estimator
- Rearranging Edgeworth-Cornish-Fisher Expansions
- On the stability and accuracy of least squares approximations
- Solving Turbulent Rayleigh-Bénard Convection using Fourier Neural Operators