Orthogonal Invariant Sensing: State Evolution

• State evolution is a deterministic recursion framework that tracks AMP-type iterates in high-dimensional inference with orthogonally invariant sensing matrices.
• It rigorously predicts algorithmic performance, phase transitions, and MSE dynamics in applications like compressed sensing, phase retrieval, and spin glass models.
• The framework provides design insights for optimal denoisers and tuning, ensuring universality across diverse random matrix ensembles and structured noise settings.

State evolution for orthogonally invariant sensing matrices refers to a rigorous asymptotic framework describing the performance and dynamics of iterative signal recovery algorithms, most notably approximate message passing (AMP), when the measurement or noise matrices possess orthogonal invariance, rather than i.i.d. Gaussian structure. This theory is central to modern high-dimensional inference in settings ranging from compressed sensing, phase retrieval, and spiked matrix models to spin glasses, where classical probabilistic decoupling can be extended to much broader random matrix ensembles.

1. Orthogonally Invariant Ensembles and Signal Models

An orthogonally invariant matrix ensemble is defined via its singular value decomposition $A = U \Sigma V^\top$ or eigendecomposition $J = O^\top D O$, where $U$, $V$, and $O$ are Haar-distributed orthogonal matrices, and $\Sigma$, $D$ carry the spectrum. This includes sub-sampled Haar matrices, deterministic delocalized orthogonals (e.g., Hadamard–Walsh or partial DFT matrices), and random symmetric matrices with invariant measure. The orthogonally invariant property ensures that the law of $A$ (or $J$) depends only on its spectrum and not on the specific basis; formally, $A \sim U A V^\top$ in law for independent orthogonals $U$, $V$.

Signal models typically considered include:

• Linear models: $y = Ax + w$, $x$ sparse or structured, $w$ Gaussian or sub-Gaussian noise.
• Generalized linear models: $y = g(Ax, w)$, with $g$ potentially nonlinear or quantized.
• Spin glass models: $$P(\sigma) \propto \exp\left[\frac{\beta}{2} \sigma^\top J \sigma + h^\top \sigma\right]$$, $J$ orthogonally invariant.
• Phase retrieval: $y_i = |(A x)_i|$ for $i=1,\ldots,m$.

2. Message Passing Algorithms and Onsager Corrections

State evolution rigorously describes the iterates of AMP-type algorithms, which exploit the concentration properties of large random matrices. For orthogonally invariant ensembles, standard AMP with a classical Onsager term (calculated under Gaussian i.i.d. assumptions) must be generalized to account for arbitrary spectra and nontrivial eigenvector statistics.

A unifying framework employs two alternating modules, often dubbed Module A (linear estimation/MMSE filtering) and Module B (nonlinear denoising). The precise form of the Onsager correction depends on the law of $\Sigma$ or $D$, and, for generic orthogonally invariant ensembles, Onsager terms require careful construction using spectral transforms (e.g., the $R$-transform for spectral densities) or even convolutional memory kernels in iterative schemes such as convolutional AMP (CAMP) (Takeuchi, 2020).

3. State Evolution Recursion: General Theory

The central construct is a deterministic recursion for empirical means, variances, and covariances of iterates, capturing their large-system limits. For a broad class of models and algorithms:

• Linear AMP/OAMP/VAMP: SE tracks either scalar MSEs (for Bayes-optimal scalar denoisers) or vector/matrix covariances (for multivariate or long-memory AMP).
• For AMP with orthogonally invariant $A$, provided sufficient moment-matching to the Marchenko–Pastur law, the empirical distribution of errors at each iteration converges to a Gaussian whose variance is determined by a closed recursion (Takeuchi, 2019).
• OAMP/VAMP dispenses with moment-matching by exploiting explicit knowledge of the singular-value law, yielding SE valid for arbitrary spectra (Takeuchi, 2022).

In phase retrieval with orthogonally invariant $A$, a concise scalar SE controls the evolution of correlations between signal and iterates. For phase retrieval with sub-sampled Haar or Hadamard–Walsh $A$, the SE is shown to be universal: $$\alpha_{t+1} =\Bigl(\tfrac1κ−1\Bigr)\,\alpha_t E[Z^2\,\bar\eta_t(|Z|)],$$

$$\sigma_{t+1}^2 =\Bigl(\tfrac1κ−1\Bigr) \Big\{\alpha_t^2(E[Z^2\bar\eta_t(|Z|)^2] - E[Z^2\bar\eta_t(|Z|)]^2) + \sigma_t^2 E[\bar\eta_t(|Z|)^2]\Big\}$$

where $Z\sim N(0,1)$, and $\bar\eta_t(|z|)$ is the centered nonlinearity (Dudeja et al., 2020).

In multivariate or spatially coupled models, SE is a high-dimensional covariance recursion coupled across blocks, tracking both mean-square errors and cross-iteration covariances. For spatial coupling, the recursion propagates across sections and iterations, and convergence properties are linked to potential minimization and information dimension thresholds (Takeuchi, 2022).

4. Universality and Rigorous Guarantees

A key result is the universality of SE: for a given algorithm and prior, the deterministic recursion derived for classical random (i.i.d. Gaussian or Haar-distributed) matrices continues to hold for wide classes of deterministic or structured orthogonally invariant matrices, such as sub-sampled Hadamard, DFT, or coded diffraction patterns, provided the signal prior is sufficiently regular (e.g., i.i.d. Gaussian or sub-Gaussian entries) (Dudeja et al., 2020, Takeuchi, 2019). This universality is established by controlling the moments of traces and quadratic forms of alternating products of the matrix and Onsager terms, using tools such as free probability, combinatorial expansions, and central limit theorems for bilinear forms.

The convergence of AMP and validity of SE for $T$ iterations require that the first $2T$ moments of the spectrum match the Marchenko–Pastur law for classical AMP, while OAMP and its variants work for general spectra without such restrictions (Takeuchi, 2019).

5. Performance Predictions, Phase Transitions, and Applications

State evolution predicts key algorithmic performance metrics:

• Phase transitions, such as the weak- to exact-recovery thresholds, which depend only on the spectrum and prior but not matrix-specific randomness.
• Mean-square-error curves over iterations, which match the observed empirical error in simulations under the large-system limit (Liu et al., 2015, Takeuchi, 3 Dec 2025).
• The phase retrieval case shows that the SE equations and thus phase transitions are identical for both Haar and Hadamard–Walsh matrices under Gaussian priors (Dudeja et al., 2020).

In settings of sublinear sparsity, e.g., $k = N^\gamma, \gamma < 1$, the GOAMP framework exploits SE to determine the "all-or-nothing" threshold for reconstructability, which matches the Gaussian AMP threshold, and remains robust under ill-conditioned or non-Gaussian $A$ where classical AMP fails (Takeuchi, 3 Dec 2025).

In spatially coupled models, SE accurately tracks the sectionwise MSE evolution and confirms that Bayes-optimal OAMP achieves information-theoretic limits (approaching the Rényi information dimension), with threshold saturation under proper coupling (Takeuchi, 2022).

6. State Evolution in Structured and Nonlinear Regimes

Beyond standard compressed sensing, SE has been applied to:

• Multivariate/memory-augmented AMP (including Bayes-OAMP and PCA with orthogonally invariant noise): SE involves matrix-valued covariances and Onsager coefficients constructed using free cumulant expansions of the spectral law (Zhong et al., 2021).
• Spin glass models: SE tracks the evolution of magnetization estimates via a memory-free or divergence-free AMP. The fixed points of SE coincide with replica-symmetric TAP equations for models with orthogonally invariant couplings. The high-temperature regime guarantees $L^2$ convergence and validates the use of SE and Onsager terms in such disordered systems (Fan et al., 2021, Fan et al., 2022).
• Non-i.i.d. and quantized measurement models: For partial DFT/row-orthogonal measurement operators, SE recursions account for deterministic AWGN-like effective channels at each stage, with state variables reflecting only scalar effective variance and precision (Liu et al., 2015).

7. Algorithm Design, Practical Considerations, and Open Questions

State evolution provides a principled route to algorithm design:

• Tunability of denoisers, initialization, and damping parameters to optimize MSE and convergence, as the SE recursion quantifies performance for any nonlinearity.
• Validation of Bayes-optimality: If SE admits a unique fixed point under the replica/Bayesian equations, the underlying message-passing scheme achieves the optimal performance predicted by statistical mechanics (Takeuchi, 2020).
• Robustness and scalability: OAMP and related variants circumvent failure modes inherent to standard AMP under non-i.i.d. or ill-conditioned $A$, as detailed in recent empirical studies (Takeuchi, 3 Dec 2025).
• Remaining challenges include rigorous uniform-in-time SE for nonlinear or spatially coupled models (especially away from Bayes-optimal priors), characterizing finite-sample corrections, and extending universality to broader deterministic transforms and non-standard noise settings.

State evolution for orthogonally invariant sensing matrices thus constitutes a foundational bridge between random matrix theory, high-dimensional inference, and nonlinear algorithmic dynamics, providing both exact asymptotics and a flexible engineering framework for structured statistical estimation.