Methods that learn a linear dynamical system from data usually test it by shaking it with random noise and recording what comes out. A long-standing question is whether a cleverer, adaptive choice of inputs could learn the system faster. A recent answer for systems whose full internal state is measured was "only sometimes": active design helps a lot when the learner's uncertainty is concentrated in a few directions, and barely helps when it is spread evenly. That answer was proven only for fully observed state. Most real systems are seen through a few sensors, and their hidden internal modes show up in the outputs with very uneven strength. This note proposes to answer the same question for partially observed linear systems. We will build an adaptive input-design method that targets the weakest hidden modes, and we will test a sharp prediction about when it beats random noise and when it cannot.
The core idea fits in two sentences. When the state is hidden, how hard identification is depends on the system's weakest "Hankel mode", meaning the internal mode that contributes least to how past inputs shape future outputs. Concentrating input power at the frequencies where that mode resonates should cut the required data by a factor that we can predict from the separation between modes in frequency, and the factor should shrink to nothing when the modes overlap. That prediction is falsifiable. It says partial observability widens the regime in which active learning gives a large gain, but only along one specific geometric axis.
Core Claim
High Effort, Low Gain studies active experiment design for picking among a finite set of fully observed linear models. Its result is that the learner only needs to inject energy where the candidate models disagree, and that this beats white Gaussian noise substantially only when that disagreement is anisotropic, meaning it lies in a few directions. The authors explicitly leave partial observability open. A separate line of work, Lee (2020), Lee (2021) and Fattahi (2021), identifies partially observed systems from passive i.i.d. Gaussian inputs. These methods estimate the impulse response, the sequence of matrices $G_i$ mapping an input $i$ steps ago to today's output. Lee then recovers a state-space model through a rank-$d$ factorization of the Hankel matrix built from those $G_i$. Lee's error bound $\widetilde{\mathcal{O}}(\sqrt{d(d_u+d_y)/T})$, with $d$ the system order, $d_u, d_y$ the input and output dimensions and $T$ the trajectory length, matches a worst-case lower bound. A skeptic will therefore say there is nothing left to gain.
We claim the opposite holds instance by instance, and that the gain can be predicted. Worst-case-optimal rates hide constants that depend on the smallest nonzero Hankel singular value $\sigma_d$. Recovering the order-$d$ model from a noisy Hankel matrix requires the noise to be small relative to $\sigma_d$, so the burn-in length and the realization error both scale with $1/\sigma_d$. Partial observability is exactly what makes $\sigma_d$ small, because a mode seen weakly through the output map $C$ contributes little to the outputs. Under an input with power spectral density $\Phi_u(\omega)$ and a fixed power budget, the effective strength of mode $k$ with transfer function $G_k(e^{i\omega})$ is
$$ s_k(\Phi_u) = \frac{1}{2\pi}\int_{-\pi}^{\pi} \|G_k(e^{i\omega})\|^2 \, \Phi_u(\omega)\, d\omega, \qquad \frac{1}{2\pi}\int \mathrm{tr}\,\Phi_u(\omega)\, d\omega \le P, $$
where $P$ is the input power. White noise sets $\Phi_u$ flat. Our predicted sample-complexity gain is the ratio between the best achievable $\min_k s_k$ under a shaped spectrum and $\min_k s_k$ under a flat one.
This ratio is large when the weakest mode resonates in a narrow band that the other modes do not occupy. Lightly damped poles near the unit circle are the typical case. The ratio approaches one when modes overlap in frequency, or when the weak mode is weak because of $C$ over a broad band. This is the partial-observability counterpart of the "high gain versus low gain" split, and it is stated in terms of a quantity we can compute.
Research Program
The method is a sequential, certainty-equivalence design loop. It begins with a short burst of white noise and runs Lee's multi-scale Hankel estimator to obtain an initial order-$d$ model. It then repeats three steps in epochs of doubling length. First, it computes the estimated mode transfer functions in a balanced realization, a coordinate system in which each internal mode is equally easy to reach and to observe. Second, it solves the convex program of maximizing $\min_k s_k(\Phi_u)$ over discretized spectra subject to the power budget. Because $s_k$ is linear in $\Phi_u$, this is a linear program. Third, it draws the next epoch's input as Gaussian noise with that spectrum, mixed with a small white-noise floor so that every frequency stays excited. All data collected so far are re-estimated with the multi-scale estimator.
Keeping the excitation stochastic with a shaped spectrum is deliberate. It preserves the Gaussian-input structure that Lee's concentration arguments rely on, so the analysis can reuse those arguments. It also avoids the non-convex trajectory optimization that High Effort, Low Gain reports scales poorly. And it needs no particles, unlike BAD-PODS. BAD-PODS does online Bayesian design for partially observed state-space models with nested particle filters, and its authors note that the particle budget becomes prohibitive in high dimension. We give up that full Bayesian generality in exchange for linear structure and a convex design step.
On the analysis side, we aim for an instance-dependent upper bound in which the realization error scales as $1/\sqrt{T \cdot \min_k s_k(\Phi_u)}$. Its form follows from the standard argument that the Hankel estimation error is $\mathcal{O}(1/\sqrt{T})$ in the input-whitened coordinates, and that factorization amplifies this error by the inverse of the smallest effective singular value. We also aim for a matching local lower bound over spectra with the same power. Together these would turn the gain ratio from a heuristic into a characterization. We will separately record the regime where the bound says active design cannot help: weakness that comes from $C$ and is spread across frequency. That regime is the partially observed "low gain" case.
Baselines
The naive reference point is passive white Gaussian excitation of the same total power, fed to Lee's multi-scale rank-$d$ estimator (Lee, 2021). It is the strongest passive competitor because it is worst-case optimal. We run it on identical systems, noise realizations, power budgets and trajectory lengths, with the same estimator, so that the only difference is the input. The file gives Lee's theoretical rates but no empirical $\mathcal{H}_2$ error values. We found no reported result on Lacuna for numeric $\mathcal{H}_2$ or realization errors on Lee's synthetic configurations, so we will reproduce them ourselves. Beating this baseline means at least a twofold reduction in the trajectory length needed to reach a target realization error, on systems where the predicted gain ratio exceeds two, with no loss on systems where it is near one.
The second baseline is the $\ell_1$-regularized impulse-response estimator of Fattahi (2021). It is the strongest method for high dimension and short data, and the file reports that ordinary least squares "failed entirely" when $N < Tp$ in that paper's $n=200$, $p=50$ setting. We run it with both passive and actively designed inputs, tuning the regularization weight on the same held-out validation budget. This tests whether active design adds to sparsity or is made redundant by it.
The strongest existing active method is BAD-PODS, run on the low-dimensional systems where its particle filter is feasible, with the same compute budget in GPU-hours. On its own SIR benchmark, the file reports 2.5× lower runtime and 10GB versus 25GB of memory relative to static design at $T=150$. Those figures measure efficiency relative to static design, not identification error, so they do not set a target error for us. Beating BAD-PODS means matching its error at a fraction of its compute on small systems, and continuing to work at $d=10$ and in the $n=200$ setting, where we expect it to exhaust its particle budget.
Two ablations isolate the key ingredient. The first, which we call the oracle-spectrum design, uses the true system's optimal spectrum. It bounds what adaptivity could achieve. The second applies the full-state "tailored excitation" rule of High Effort, Low Gain naively to the estimated state from an observer instead of to Hankel modes. It tests whether reasoning in Hankel coordinates, rather than state coordinates, is what matters.
Why not just…?
Why not just use white noise and run longer? This is the real threat, and in part of the space it wins. The theory above predicts it is near-optimal whenever modes overlap in frequency or the observability weakness is broadband. The experiments therefore include a sweep that deliberately enters that regime. If white noise matches us even where the predicted ratio is large, the claim is wrong.
Why not just pick a fixed shaped input spectrum from engineering intuition? For known systems, classical frequency-domain design does exactly this, and a well-chosen prior spectrum could capture most of the gain. The obstacle is that the weak mode's resonance is what we do not know. We include a "prior spectrum" baseline drawn from the distribution that generates the test systems, plus a two-phase variant that designs a spectrum once after the initial burst. If the two-phase variant matches the fully sequential loop, we will report the simpler method as the contribution.
Why not just use particle-based Bayesian design, which handles nonlinearity too? BAD-PODS is more general, but its authors identify the particle budget as prohibitive for high-dimensional hidden states. The comparison at $d \in \{5,10\}$ and in the $n=200$ setting will show whether that limit binds in practice for linear systems.
Why not just train a recurrent network on input–output histories? That approach gives no identified model, and Shi et al. show that unstructured partially observed learning faces exponential lower bounds. Linear structure is what makes a convex design step possible. We do not run this baseline because it answers a different question.
Proposed Experiments
The primary testbed reproduces Lee's synthetic configurations: orders $d \in \{1,3,5,10\}$, $d_u=d_y \in \{1,3\}$, memory lengths $L \in \{27,81\}$ and spectral radii $\{0.9, 0.95\}$. We augment it with a controlled family where we set the weakest mode's damping, its frequency separation from the other modes, and whether its weakness comes from $B$ or from $C$. The secondary testbed is Fattahi's $n=200$, $p=50$ setting. A third testbed is the structured versus unstructured hypothesis classes of High Effort, Low Gain, recast with partial outputs, so that discrimination error can be compared directly.
Metrics are the $\mathcal{H}_2$ error of the impulse response, a similarity-invariant realization error (transfer-function error and pole error), and the trajectory length needed to reach a target error, all at equal input power and averaged over at least 20 seeds.
The decisive check is a scatter plot of the measured data-saving factor against the predicted gain ratio computed from the true system. The claim survives only if the two correlate strongly. It must also hold that systems with C-driven, broadband weakness show savings near one, as predicted, rather than the large gains a naive "partial observability always helps active learning" story would expect. A secondary check pushes the spectral radius toward one, where predicted gains grow and the existing passive guarantees stop applying. This shows whether the method degrades gracefully at the edge of stability, which the file identifies as an open problem.
References
- High Effort, Low Gain: Fundamental Limits of Active Learning for Linear Dynamical Systems
- Improved rates for identification of partially observed linear dynamical systems
- Improved rates for prediction and identification for partially observed linear dynamical systems
- Learning Partially Observed Linear Dynamical Systems from Logarithmic Number of Samples
- Online Bayesian Experimental Design for Partially Observed Dynamical Systems
- Near-Optimal Partially Observable Reinforcement Learning with Partial Online State Information