When people fit a linear model of a physical system from recorded trajectories, they usually report two things: whether the learned model is stable (its predictions do not blow up when run forward), and how many effective dimensions it needed. This proposal argues that, with today's methods, both numbers depend on undisclosed data-handling choices. The choices in question are how each sensor channel was rescaled, how finely time was sampled, how trajectories were cut into windows, and how train and test data were split. Rank-reduction methods shrink singular values, and singular values change when channels are rescaled. Stability verdicts for unconstrained fits change with the sampling interval, because finer sampling pushes every discrete-time eigenvalue toward the unit circle. Two groups can therefore start from the same raw data and report different ranks and different stability outcomes. Nobody would know why, because the preprocessing went unreported. The digital-pathology replication by Fell et al. showed this kind of collapse in vision: slide accuracy fell from 92% to 54.7% when only the published text was available. The cause was unreported preprocessing, not the architecture.
We propose a model whose stability and whose rank are properties of the underlying system rather than of the pipeline. The model is a stable-by-construction linear system in which only the energy-dissipation part is low-rank. We pair it with a preprocessing-sweep protocol that measures how much each reported quantity moves under realistic, undocumented data choices. In this note, dataset transparency is the evaluation lens. Stability parameterization is the core mechanism. Low-rank regularization is the specific prior that makes the reported rank meaningful.
Core Claim
We use the dissipative Hamiltonian form from the distance-to-stability literature. Gillis and Sharma show that every stable continuous-time matrix can be written as $(J-R)Q$, with $J$ skew-symmetric (energy-preserving rotation), $R$ positive semidefinite (energy loss), and $Q$ positive definite (stored energy). We restrict the dissipation to rank $k$:
$$A_c = (J - LL^\top)\,Q, \qquad L \in \mathbb{R}^{n\times k},\ k \ll n,$$
and we produce discrete-time predictions through the exact discretization $A_d = \exp(A_c\,\Delta t)$, where $\Delta t$ is the sampling interval. Because $A_c$ is stable for every value of the parameters, $A_d$ is stable for every $\Delta t$. The sampling interval therefore becomes an explicit, reported input instead of something silently baked into the fitted matrix.
The key observation is that this model class is closed under any invertible change of coordinates $T$, which includes per-channel standardization:
$$T A_c T^{-1} = \big(TJT^\top - (TL)(TL)^\top\big)\,\big(T^{-\top} Q T^{-1}\big).$$
The first factor is again skew-symmetric minus a rank-$k$ positive semidefinite term, and the second factor is again positive definite. The damping rank $k$ is therefore a coordinate-free property of the system: it is the number of independent directions through which energy leaves. Singular-value shrinkage of the transition matrix, as in reduced-rank regression, has no such invariance. The claim to test is that, under realistic preprocessing sweeps, the selected damping rank and the long-horizon behavior of this model stay nearly constant, while the rank and stability outcomes of reduced-rank-plus-stabilization pipelines vary substantially.
The prior is also physically different from the usual one. Low-rank transition matrices assume most modes vanish. Low-rank dissipation assumes most modes are lightly damped oscillations and a few channels carry the loss. Mass-spring chains with localized dampers fit this description, and so do fluid systems with a few outflows.
Relation to Prior Work
Several papers supply the method. Gillis and Sharma fit $(J-R)Q$ with dense $R$ using projected and accelerated gradient methods. They also note a scaling redundancy between $J-R$ and $Q$, which we remove by fixing the trace of $Q$. Our change is to replace the dense $R$ with $LL^\top$, which drops the semidefinite projection entirely and lowers the cost of handling $R$ from $O(n^3)$ to $O(nk)$ per step. Only the matrix exponential stays cubic in $n$. Finding the Nearest Positive-Real System extends the same structure to input-output passive systems and supplies our mass-spring-damper test systems. Stable Port-Hamiltonian Neural Networks and Dissipative Hamiltonian Neural Networks carry the energy decomposition into nonlinear networks, but they use full-capacity dissipation and do not study preprocessing sensitivity. We stay linear so that the invariance argument is exact.
Other papers motivate the proposal. Characteristic Root Analysis and Regularization for Linear Time Series Forecasting shows that least-squares linear forecasters learn spurious roots, because their noise error decays only as $O(1/\sqrt{T})$. It removes those roots with reduced-rank regression and a null-space penalty it calls Root Purge. Reduced-Rank Linear Dynamical Systems places nuclear-norm priors on the latent transition matrix and reports a consistent intrinsic rank of 5–6 on hippocampal data. Both papers treat rank as a singular-value property and neither guarantees stability. Our proposal moves the rank onto dissipation and makes stability unconditional. We also test whether the ranks those methods report survive rescaling.
Two discrete-time methods constrain the design and set the bar. Mamakoukas et al. parameterize stable matrices as $S^{-1}OCS$ with $O(n^2)$ memory. Efficient Learning of a Linear Dynamical System with Stability Guarantees stabilizes a least-squares estimate through a Riccati-equation projection. Both are stable but full-rank within their chosen state order, and both bake $\Delta t$ into the matrix. The stability/rank link in Low-Rank Matrix Approximation with Stability concerns algorithmic sensitivity to training samples, not eigenvalues. We borrow its question in a different form and ask whether our reported quantities are insensitive to pipeline perturbations.
Research Program
The work has three parts.
The first part is building and fitting the model. We fit $(J, L, Q)$ by multi-step rollout loss through $\exp(A_c\Delta t)$. We parameterize $Q$ as a Cholesky factor with fixed trace and choose $k$ by validation rollout error over a small grid. A nuclear-norm variant on a wider $L$ serves as a soft alternative to the fixed rank.
The second part is the preprocessing-sweep protocol, which is the transparency contribution. For every dataset we define a grid of defensible choices: per-channel standardization versus global scaling versus raw values; two to four sampling intervals obtained by subsampling; three window lengths; random versus chronological versus trajectory-held-out splits; and added sensor noise at several levels. Every method is refit on every cell of the grid. For each reported quantity we record its spread across the grid: selected rank, fraction of fits that are stable, one-step error, and long-horizon error. The result is a ranked list of which undisclosed choices actually move results in dynamical-systems learning. This list becomes the empirical basis for a short reporting card. Its fields are kept only if the sweep shows they matter, which addresses the open point in the file that control benchmarks have no transparency standard.
The third part is a light analysis. We will characterize which sweep axes the model is provably invariant to (coordinate rescaling and $\Delta t$, at the level of the model class) and which it is not (window length, split, and noise level, which change the estimator rather than the model class). The invariance is exact for the model class but not for the estimate, since the fitting loss is computed in the chosen coordinates. The empirical question is whether the estimate inherits most of the invariance.
Baselines
The naive reference is ordinary least squares on the transition matrix, refit on every sweep cell, with unstable fits reported as failures rather than repaired. It shows how often the pipeline alone flips a stability verdict. Mamakoukas et al. report that unconstrained least squares reached an average figure-8 tracking error of 7556 on the Franka Panda arm.
The strongest existing stable method is the SOC parameterization of Mamakoukas et al.. The same paper reports an average figure-8 tracking error of roughly 0.08 on the Panda arm. We run it with their fast gradient method, using the same state order, the same rollout loss, and the same number of optimizer steps and hyperparameter trials as our model. Its state order is tuned over the same grid we use for $k$. Beating it means matching its error on the Panda arm and dynamic textures in the default pipeline and showing clearly smaller spread across the sweep, or lower error on the systems where damping rank is genuinely low.
The strongest existing low-rank method is reduced-rank regression and Root Purge from Characteristic Root Analysis. We follow it with the Riccati-equation stabilization of Jongeneel et al. whenever the fit is unstable, which yields a two-stage "low-rank then stabilize" pipeline. It gets the same rank grid and tuning budget. For ETT and the other forecasting sets, no reported result was found on Lacuna for reduced-rank regression or Root Purge MSE. We will therefore reproduce it under the default protocol and report our reproduction alongside.
For the real damped pendulum, Sosanya and Greydanus report a test MSE of $9.5\times10^{-4}$ for their dissipative Hamiltonian network. We treat this as a nonlinear upper reference rather than a direct competitor.
Two ablations isolate the key ingredient. The first is the same $(J-R)Q$ model with dense $R$, which keeps stability but drops the rank restriction. The second is the same pipeline with low rank imposed on $A_c$ itself instead of on $R$, which keeps a rank restriction but puts it in the non-invariant place. If the dense-$R$ ablation shows the same sweep robustness, then the low-rank part adds nothing beyond stability.
Why not just…?
Why not just fit a reduced-rank model and then project it onto the stable set? This is the most serious alternative, and it is included as a baseline rather than argued away. The concern is structural. Reduced-rank truncation selects singular directions, and per-channel rescaling changes those directions, so the selected rank and the subsequent projection can both shift with preprocessing. The two-stage pipeline also fixes $\Delta t$ inside the matrix. The coordinate-rescaling and sampling-interval axes of the sweep test exactly this. If the two-stage pipeline shows comparable rank stability and error, the invariance argument has no practical value, and we will say so.
Why not just use SOC with a small state order? Shrinking the state order removes whole modes, including lightly damped oscillations that matter for long-horizon prediction, whereas low damping rank keeps every mode and restricts only where energy is lost. On the mass-spring-damper chains with a known number of localized dampers, this predicts a clean separation. SOC should need a large state order to keep the oscillations, while our model should recover the true number of dampers. If SOC matches us there, the distinction is moot.
Why not just require code release and a preprocessing description? Fell et al. found that fully specified architectures still failed to replicate when preprocessing details were missing. A description alone also does not tell readers which choices matter. The sweep produces that ranking, and a model that is robust to the dominant axes reduces how much must be disclosed for results to transfer.
Why not just put a nuclear-norm penalty on dense $R$? It is a plausible soft version and is included as a variant. It reintroduces a semidefinite projection at cubic cost and a penalty weight that, as She et al. note for their priors, needs empirical tuning. The penalty is also not congruence-invariant, so it may inherit some of the scaling sensitivity.
Proposed Experiments
The first test is decisive. It uses synthetic mass-spring-damper chains with 20 to 40 states, following Gillis and Sharma (2018), with $k^\star \in \{1,2,4\}$ dampers placed at chosen nodes and observation noise swept. We measure how often the selected $k$ equals $k^\star$, the rollout error to 10× the training horizon, and the spread of selected rank under random diagonal rescaling of channels. The prediction is exact recovery of $k^\star$ at moderate noise and near-zero rank spread under rescaling. For the reduced-rank pipeline, we predict visible rank spread and stability flips as $\Delta t$ shrinks.
On real systems, we run the full sweep on the Franka Panda arm and the UCLA, UCSD and DynTex dynamic textures from Mamakoukas et al., and on the cascaded tanks from Roth et al., where the correct long-run behavior (draining to zero) gives a qualitative check. No numeric cascaded-tanks error was found on Lacuna. On the dynamic textures, no per-video error numbers were found on Lacuna either, so all methods are reproduced under the same protocol.
On forecasting, ETT, Weather and Electricity test whether the method transfers to settings where the companion-matrix spurious-root story of Characteristic Root Analysis applies. Here we expect parity at best, and we frame this as a check that stability by construction does not cost accuracy.
The proposal fails if, across the sweep, the dense-$R$ ablation or the two-stage reduced-rank pipeline shows rank and error spreads within noise of ours. In that case the transparency output (the sensitivity ranking and reporting card) still stands, but the modeling claim does not.
References
- Reproducibility of deep learning in digital pathology whole slide image analysis
- On computing the distance to stability for matrices using linear dissipative Hamiltonian systems
- Finding the Nearest Positive-Real System
- Stable Port-Hamiltonian Neural Networks
- Dissipative Hamiltonian Neural Networks: Learning Dissipative and Conservative Dynamics Separately
- Characteristic Root Analysis and Regularization for Linear Time Series Forecasting
- Reduced-Rank Linear Dynamical Systems
- Memory-Efficient Learning of Stable Linear Dynamical Systems for Prediction and Control
- Efficient Learning of a Linear Dynamical System With Stability Guarantees
- Low-Rank Matrix Approximation with Stability