---
title: Low-Rank State Approximation
url: https://www.emergentmind.com/topics/low-rank-state-approximation
type: topic
---

# Low-Rank State Approximation

Low-rank state approximation refers to the task of representing a high-dimensional state, operator, or function by an object of much lower rank, typically for purposes of compression, denoising, complexity reduction, or model reduction. In both classical and quantum settings, the core objective is to approximate the dynamics or statistical structure of a system while restricting the approximant to lie in a low-rank manifold or subset. This notion arises in numerous domains of computational mathematics, physics, control, and machine learning.

## 1. Formal Problem Statement

Given a high-dimensional or operator-valued state (e.g., the state vector or density matrix of a dynamical system, or a trajectory ensemble), the low-rank approximation problem seeks the closest state (according to a specified norm or metric) of rank at most $k$, where $k$ is much smaller than the ambient dimension. In the classical linear-dynamical setting, this translates to

\[
A_k^* = \arg\min_{A \in \mathbb{R}^{n \times n}, \ \mathrm{rank}(A) \leq k} \|Y - A X \|_F^2
\]

where $X$ and $Y$ collect snapshot data (states before and after evolution), and the goal is to find a best-fitting linear dynamics $A$ of rank at most $k$ [1610.02962]. In quantum state approximation, the analogous task is

\[
\sigma^*(R) = \arg\min_{\substack{\sigma \geq 0, \\ \mathrm{Tr} \, \sigma = 1, \\ \mathrm{rank}(\sigma) \leq R}} D(\rho, \sigma)
\]

where $D( \cdot, \cdot )$ is a chosen metric (e.g., Hilbert-Schmidt or trace distance) and $\rho$ is a density matrix [2203.00811]. The principal aim across contexts is dimensionality and computational reduction while retaining essential dynamic, statistical, or physical fidelity.

## 2. Exact Closed-form and Algorithmic Solutions

The central theoretical advance for linear systems is the discovery of an exact, closed-form minimizer of the empirical $\ell_2$-error under a rank constraint [1610.02962]. The solution proceeds by projecting the unconstrained minimizer $A_m^* = Y X^\dagger$ onto its best rank-$k$ approximation:

\[
A_k^* = U_{Z,k} U_{Z,k}^{\top} Y X^\dagger
\]

where $Z = Y X^\dagger$ and $U_{Z,k}$ consists of the top $k$ left singular vectors of $Z$. This result enables an explicit, polynomial-time algorithm:

1. Compute SVD of $X$ ($O(n m^2)$ if $X \in \mathbb{R}^{n \times m}$).
2. Compute $Z = Y X^\dagger$.
3. Compute leading $k$-SVD of $Z$.
4. Return $A_k^* = U_{Z,k} (U_{Z,k}^\top Y X^\dagger)$.

Alternative representations yield SVD-based (primal) and EVD-based (modal) reduced models, useful for analyzing or simulating the dynamics in a reduced subspace [1610.02962].

In the quantum setting, the optimal low-rank $\sigma$ minimizing the Hilbert-Schmidt distance to a density matrix $\rho = \sum_{i=1}^r \lambda_i |e_i\rangle \langle e_i|$ is given in closed form:

\[
\sigma^*_{HS} = \sum_{i=1}^R \left( \lambda_i + \frac{1 - \sum_{j=1}^R \lambda_j}{R} \right) |e_i\rangle \langle e_i|
\]

The optimal solution for trace distance admits a degenerate set of minimizers, always supported on the leading $R$ eigenvectors but with a simplex of admissible eigenvalue vectors [2203.00811].

## 3. Error Characterization and Robustness

The exact approximation error for the optimal low-rank solution in the linear setting is

\[
\| Y - A_k^* X \|_F^2 = \sum_{i=k+1}^m \sigma_{Z,i}^2 + \| (I - \mathcal{P}_{X^\top}) Y \|_F^2
\]

When $X^\top$ is full row rank, the second term vanishes, and the error reduces to the sum of squared singular values beyond $k$ in $Z$ [1610.02962]. Hence, the method is near-optimal if the singular spectrum of $Z$ decays rapidly.

Numerical benchmarks show that this closed-form method not only dominates traditional two-step truncation/projection or DMD variants in accuracy, but is also robust under moderate noise. Competing methods (truncated DMD, projected DMD, or convex relaxations) can be significantly suboptimal or unstable, especially under noise or when the true dynamics are only approximately low-rank.

Quantum analogues similarly quantify the minimum distance; for Hilbert-Schmidt, the minimum error is

\[
D^*_{HS} = \sum_{i>R} \lambda_i^2 + \frac{(1-\sum_{j=1}^R \lambda_j)^2}{R}
\]

for eigenvalues $\lambda_i$ of $\rho$ [2203.00811].

## 4. Reduced-order Model Construction

Low-rank models can be constructed directly in factorized form to facilitate simulation and further analysis. For the SVD-based primal form, the optimal low-rank $A_k^*$ is factorized as $A_k^* = P Q^\top$ with $P = U_{Z,k}$ and $Q = (X^\top U_{Z,k})^\dagger{}^\top$. The reduced system is then

\[
z_{t} = S z_{t-1}, \qquad \tilde{x}_t = R z_t
\]

where $R = P$, $L = Q$, $S = Q^\top P$ [1610.02962].

Alternatively, for diagonalizable $A_k^*$, the EVD-based construction analytically recovers the right and left eigenvectors and their associated invariants within the $k$-dimensional optimal subspace, further enabling modal analysis and control.

## 5. Polynomial Complexity and Implementation

All critical steps—economy-size SVDs, linear solves via pseudoinverse, and construction of projectors—can be implemented in $O(m^2(m+n))$ time off-line for $X, Y \in \mathbb{R}^{n \times m}$ and $k \ll \min\{m,n\}$. On-line simulation costs for the reduced model are $O(r n + T r^2)$, where $r$ is the reduced rank and $T$ the time horizon [1610.02962].

There is no numerical evidence of instability or pathological slow-down in any tested regime (linear, weakly nonlinear, high noise). The algorithm is effective even at large scale (e.g., $n=1024$ in benchmarked cases).

## 6. Applications and Empirical Demonstrations

Low-rank state approximation delivers substantial benefits in high-dimensional dynamical model reduction, system identification, and data-efficient simulation. Experimental evaluation in both synthetic and physical systems demonstrates that:

- For linear systems, the closed-form low-rank solution is optimal whenever the true system lies in the data span and achieves minimal error for $k$ exceeding the intrinsic dimension.
- For weakly nonlinear or heavily noisy systems, only the closed-form low-rank solution matches the minimum approximation error; standard DMD truncation and convex relaxations can lag by orders of magnitude.
- Modal reconstructions and spatial eigenvectors retain greater interpretability and physically meaningful structures when computed from the optimal low-rank model; other truncation-based approaches degrade substantially under noise.

The framework is robust to noise, achieves minimal error, supports construction of interpretable reduced models, and offers computational efficiency on par with the best classical approaches [1610.02962].

## 7. Comparative Analysis and Limitations

Relative to DMD-based two-stage truncations, projected DMD, and convex relaxations, the closed-form low-rank approach is universally superior in empirical $\ell_2$-error, stability, and physical fidelity, while not requiring significantly more computational effort. Its only essential assumption is that the input data has sufficient rank and coverage to support the relevant SVD computations.

Generalization to nonlinear or time-varying systems requires either re-linearization or extension to kernelized or manifold-based settings. For quantum systems, Hilbert-Schmidt low-rank approximation is unique and well-posed; for the trace norm, the solution is non-unique and care must be taken in variational or learning-based applications [2203.00811]. Open questions include systematic handling of more complex (nonlinear, time-inhomogeneous) dynamical processes and extension to online or streaming data regimes.

---

**References**

- "Low-Rank Dynamic Mode Decomposition: An Exact and Tractable Solution" [1610.02962]
- "The quantum low-rank approximation problem" [2203.00811]

Source: https://www.emergentmind.com/topics/low-rank-state-approximation