Computable Time-Uniform Error Certificates for Koopman Forecasts, and Where They Break

When a model forecasts a dynamical system by applying a learned one-step map over and over, the useful question is how far the forecast drifts from the truth over every horizon at once, not just at step 10. Theory now gives such time-uniform guarantees for some learned forecasters. These are upper bounds on forecast error that do not grow with the number of steps, because the system forgets its past fast enough to absorb the error injected at each step. Yet nobody has checked whether these bounds say anything at realistic sample sizes. The guarantees depend on constants that cannot be observed, such as how strongly the dynamics contract and how much temporary error amplification the learned operator allows. On systems with slow transitions between long-lived states, which is the regime practitioners care about, those constants may be so large that the bound is meaningless.

We propose to turn the time-uniform bound for Koopman forecasting into a number computed from data. A Koopman forecaster is a linear operator, learned in a kernel feature space, that pushes observables one step forward in time. The computed number is a certificate: a high-probability ceiling on worst-case rollout error over all horizons, built only from the learned operator and the training trajectory. We then audit it systematically. We measure how loose it is, where it becomes vacuous, and why. Our hypothesis is that the certificate is informative on quickly mixing systems and fails sharply on metastable molecular systems. We expect the failure to trace to a single computable quantity, the sum of the norms of the learned operator's powers. If that holds, the same quantity can choose the time lag at which the operator should be learned. The result is a forecasting method whose guarantee and empirical error can be compared directly on the same plot.

Core Claim

Consistent Long-Term Forecasting of Ergodic Dynamical Systems supplies the method we build on. It removes the known invariant component (the eigenvalue-1 mode) before regression, learns the residual operator, and adds the invariant mean back. This is their Deflate-Learn-Inflate (DLI) scheme. They prove a uniform-in-time bound using the Kreiss constant, a measure of how much a non-normal matrix can amplify before it decays. The paper shows large empirical gains. It never compares the bound itself to observed errors, and its theory assumes unique ergodicity. Alanine dipeptide, one of its own benchmarks, is metastable, so this assumption is strained there.

We replace the worst-case Kreiss argument with a quantity we can compute. Let $K$ be the true residual operator, $\hat K$ the learned one, and $\Delta = K - \hat K$. Unrolling the difference of powers gives

$$\sup_{t\ge 1}\|K^t - \hat K^t\| \;\le\; \|\Delta\|\,\sup_{t}\sum_{s=0}^{t-1}\|K^{s}\|\,\|\hat K^{t-1-s}\| .$$

The right side depends on the unknown $K$ only through its power norms. Define the transient sum $S = \sum_{s\ge 0}\|\hat K^s\|$, which we compute by powering the learned finite-rank matrix. We will prove a perturbation step showing that whenever $\|\Delta\| S < 1$, the true operator's power sum is at most $S/(1-\|\Delta\|S)$. This yields the certificate

$$C \;=\; \frac{\widehat{\varepsilon}\, S^2}{1-\widehat{\varepsilon}\, S},$$

where $\widehat\varepsilon$ is a high-probability bound on $\|\Delta\|$ from the data-dependent empirical Bernstein inequality for dependent data of An Empirical Bernstein Inequality for Dependent Data in Hilbert Spaces and Applications. That paper bounds operator estimation error on a training trajectory but not multi-step rollout error. Our certificate supplies the missing propagation step.

The prediction is specific. When the second-largest eigenvalue $\lambda_2$ of the residual dynamics approaches 1, as it does for slow barrier crossings, $S$ grows roughly like $1/(1-\lambda_2)$. The certificate then goes vacuous once $\widehat\varepsilon S \ge 1$. Learning the operator at a longer lag $L$ shrinks $S$ toward $1/(1-\lambda_2^L)$. The cost is fewer effective samples, so $\widehat\varepsilon$ grows, and coarser time resolution. The certificate therefore defines an optimal lag that can be computed before any rollout is run.

Research Program

The theoretical part is deliberately small. It consists of the perturbation lemma above, its combination with the empirical Bernstein bound (including the block-size choice that paper leaves sensitive), and a statement for the deflated operator with inflation restored. We will also give a simple corollary relating $S$ to $\lambda_2$ and to the operator's non-normality. This corollary explains when the certificate is vacuous. It covers the ergodic case only and makes no claim about non-ergodic systems.

The empirical part is the main contribution and addresses the validation gap directly. For each system we run long held-out rollouts across many seeds and training sizes $n$. We report the observed worst-case error over all horizons, the certificate $C$, the original Kreiss-based bound with its constants evaluated, and the coverage rate, meaning the fraction of runs in which the certificate actually holds. The diagnostic quantity throughout is slack, the ratio of certificate to observed worst-case error. We also report the horizon and sample size at which each bound stops being finite or below the trivial error level.

The method part uses the certificate to select the learning lag $L$, and optionally the rank in reduced-rank regression. We compare this choice against selection by validation rollout error.

Baselines

The naive reference point is standard uncentered reduced-rank regression of the Koopman operator, rolled out with no deflation. Consistent Long-Term Forecasting of Ergodic Dynamical Systems reports that it drifts on alanine dipeptide and has conditional-variance RMSE 0.0470 on the Cox–Ingersoll–Ross model. Its certificate, computed the same way, should be vacuous or diverge. That outcome checks that the certificate detects instability at all.

The strongest existing forecaster is DLI with reduced-rank regression from the same paper. It reports conditional-variance RMSE of 0.0124 on Cox–Ingersoll–Ross. For alanine dipeptide the file gives only a figure and no numbers, so we will regenerate them. We run DLI with identical kernels, ranks, training trajectories, and a matched hyperparameter search budget. Our lag-selected variant beats it if its worst-case rollout error over all horizons is lower on the metastable systems at equal training data. On Cox–Ingersoll–Ross and Ornstein–Uhlenbeck it must stay within noise of DLI, because the fix should do nothing where it is not needed.

The strongest existing bound is the Kreiss-based time-uniform bound of the same paper with its constants evaluated. Our certificate beats it if slack is smaller by at least an order of magnitude at the same coverage.

Three ablations isolate the key ingredients. The first replaces $S$ with $1/(1-\hat\rho)$, using the spectral radius only. This ignores transient growth and should under-cover on non-normal operators. The second replaces the empirical Bernstein $\widehat\varepsilon$ with a worst-case concentration bound, which isolates how much data-dependence matters. The third fixes the lag at 1, which isolates the lag-selection rule.

Why not just…?

Why not just measure held-out rollout error, or wrap the forecaster in conformal prediction? This is the real threat. With many independent long trajectories, empirical time-uniform intervals could be tighter than any bound. Our setting is a single long correlated trajectory, which is the norm for molecular dynamics. Exchangeability then fails, and every independent validation rollout consumes data that could be used for training. We will test this directly on the Ornstein–Uhlenbeck and multi-well systems. There we can generate many independent trajectories, so we can compare conformal coverage and width against the certificate as the number of available held-out trajectories shrinks. If conformal intervals dominate even with one trajectory, we will report that, and the certificate's remaining value is then only as a diagnostic.

Why not just clip learned eigenvalues inside the unit disk? Clipping prevents blow-up but does not bound error. Non-normal operators with spectral radius below 1 can still spike transiently, and this is exactly why the DLI paper needed the Kreiss argument. The spectral-radius ablation tests this, and we expect it to under-cover.

Why not just use a deep forecaster such as Informer or the Augmented Invertible Koopman Autoencoder? Informer's single-pass decoder reduced MSE by 26.8–60.1% relative to LSTMs on ETT. The autoencoder ranks well on the Informer benchmarks. Both answer a different question: they offer accuracy without any guarantee. We include the autoencoder on alanine dipeptide as an accuracy reference only, and we do not claim to beat it on raw error.

Proposed Experiments

The systems form a ladder of mixing speed. Ornstein–Uhlenbeck has a known Gaussian invariant law, so ground-truth error can be computed exactly. Cox–Ingersoll–Ross is the second rung. Next is a synthetic one-dimensional multi-well Langevin system whose barrier height we sweep, so that $1-\lambda_2$ moves through several decades. These three follow Consistent Long-Term Forecasting of Ergodic Dynamical Systems and the double-well setting of Yang and Wang. The top rungs are alanine dipeptide dihedral angles and Chignolin, using the aligned data distributed with jiajunhe98/FEAT. Chignolin's folding transition is far slower than alanine's, so it stresses the regime where we predict failure.

The metrics are worst-case and per-horizon MAE and relative MMD over rollouts of length well beyond the slowest relaxation time, certificate slack, coverage over at least 50 seeds, and the critical sample size at which $\widehat\varepsilon S$ crosses 1.

The barrier sweep is the decisive check. If certificate vacuity tracks $1/(1-\lambda_2)$ as predicted, and the certificate-selected lag matches the lag that minimizes observed worst-case error, the central claim holds. If slack stays below roughly 10 on Ornstein–Uhlenbeck and Cox–Ingersoll–Ross while coverage stays at the nominal level, the certificate is useful where theory says it should be. If the certificate remains vacuous on alanine dipeptide at every lag and every available $n$, the honest conclusion is that time-uniform guarantees for kernel Koopman forecasting do not reach metastable molecules at current data scales. That negative result is itself worth publishing, as concrete evidence of the validation gap.