Long-Context Linear System Identification

2024-10-08

ICLR 2025International Conference on Learning Representations

Long-Context Linear System Identification

Oğuz Kaan Yüksel·Mathieu Even·Nicolas Flammarion
October 2024

Research context

Open in graph

Produces

Sample-complexity bound for high-order linear systems

For a broad class of high-order linear systems—equivalently, vector autoregressive processes—the constrained empirical-risk minimizer attains the i.i.d. parametric rate up to logarithmic factors. Its dependence on the system-response condition number \(\kappa\) is only logarithmic, and it incurs no multiplicative mixing-time penalty.

Statement
\[\left\|\widehat{\boldsymbol A}-\boldsymbol A^\star\right\|_F^2 \leq \widetilde{\mathcal O}\!\left(D^2\left(1+\ln\kappa\right)\frac{p d^2}{N(T-p)}\right)\]
\(p\)
autoregressive order
\(d\)
dimension of each state \(\boldsymbol x_t\in\mathbb R^d\)
\(N\)
number of independent trajectories
\(T\)
length of each trajectory
\(D\)
known bound on the prediction-operator norm and the estimator search space
\(\kappa\)
condition number of the system-response operator mapping innovations to the full trajectory
\(\widetilde{\mathcal O}\)
asymptotic upper-bound notation suppressing logarithmic factors in the other problem parameters
\(\boldsymbol A^\star\)
true block transition matrix \([A_1^\star,\ldots,A_p^\star]\in\mathbb R^{d\times pd}\)
\(\widehat{\boldsymbol A}\)
constrained empirical-risk minimizer, also in \(\mathbb R^{d\times pd}\)