Perturb-Then-Diagonalize (PTD) Approach

• PTD is a mathematical framework that introduces controlled perturbations to regularize and stabilize the diagonalization of ill-conditioned matrices.
• The method involves a two-step process: first perturbing the operator to induce manageable spectral properties, then applying a suitable diagonalization technique.
• It has practical applications in quantum mechanics, signal processing, MIMO communications, and machine learning for robust system representations.

The Perturb-Then-Diagonalize (PTD) approach encompasses a collection of mathematical strategies for resolving diagonalization and block-diagonalization problems under nontrivial structural constraints or in the presence of small perturbations. PTD methods are widely utilized in fields ranging from quantum many-body theory and condensed matter to signal processing, wireless communications, numerical linear algebra, and modern machine learning. The PTD paradigm generally involves two key steps: first, introducing or leveraging a perturbation (either intrinsic or constructed) to regularize or simplify the spectral structure, and second, applying an appropriate diagonalization scheme to the resulting system so as to extract canonical, block-diagonal, or otherwise decoupled representations. Below, the central theoretical and algorithmic frameworks of PTD are surveyed, incorporating significant advances and variants across multiple domains.

1. Formal Structure of the PTD Paradigm

In its broadest sense, the PTD approach addresses problems where direct diagonalization is numerically unstable, ill-posed, or fails to respect auxiliary physical or engineering constraints. The archetype is a system described by an operator or matrix $A = A_0 + V$, with $A_0$ admitting a simple structure (e.g., diagonal or block-diagonal), and $V$ a perturbing interaction or constraint-induced term. The PTD methodology employs a perturbation—either intrinsic to the problem or artificially introduced—to regularize spectral pathologies, after which an exact or approximate diagonalization is performed.

The overall logic can be summarized as:

1. Perturbation: Modify the matrix/pencil/operator (possibly by a random or structured perturbation with norm $\|E\| \leq \epsilon$) to circumvent ill-conditioning or to induce spectral shattering/pseudospectral regularity.
2. Diagonalization: Apply a (possibly backward-stable or recursive) diagonalization algorithm—ranging from truncated SVD, Schur decomposition, fixed-point maps, unitary transformations, or divide-and-conquer—to the perturbed object.

This two-stage process enables block-diagonalization, exact or approximate spectral decomposition, or canonical system representations even when direct algebraic methods fail due to instability or non-normality (Yu et al., 2023, Demmel et al., 2023).

2. PTD in Linear Representation Problems

The PTD framework is fundamental to a variety of linear algebraic problems, such as generalized eigenvalue problems, block-diagonalization of Hermitian (or more general) matrices, and canonical forms of matrix pencils.

2.1. Block Diagonalization and Canonical Forms

For Hermitian $H = H_0 + \lambda V$, with $H_0$ explicitly block-diagonal (projections $\{P_i\}$), PTD constructs unitary transformations $T = e^{-iS}$ so that $T^\dagger H T$ is block-diagonal. The least-action (direct-rotation) approach, following Cederbaum, identifies the unique minimal-norm $T$ by minimizing $\|T - I\|$; the associated generator $S$ is obtained as a power series in $\lambda$, with explicit formulas through third order. This is contrasted with the Schrieffer–Wolff (block-off-diagonal generator) prescription, which diverges from the least-action path at $O(\lambda^3)$ for $k > 2$ blocks (Mankodi et al., 2024).

2.2. Generalized Matrix Pencil Diagonalization

For pencils $(A, B)$ (possibly singular), PTD regularizes the pencil by random Ginibre perturbations and scaling, creating a well-behaved $\epsilon$-pseudospectrum ("shattering" into $n$ components), amenable to randomized, inverse-free divide-and-conquer diagonalization. The result is a backward-stable, high-probability construction of invertible $S, T$ and diagonal $D$ such that

$$\|A - S D T^{-1}\|_2 \leq \epsilon, \quad \|B - S T^{-1}\|_2 \leq \epsilon$$

achieved in $O(\log^2(n/\epsilon) T_{\mathrm{MM}}(n))$ operations (Demmel et al., 2023).

Table 1: Core Steps of PTD for Pencil Diagonalization

Step	Operation	Purpose
Perturb	Add Ginibre noise	Regularize and shatter spectrum
Scale	Pencil scaling	Optimize conditioning
Diagonalize	Randomized D&C	Reach block/diagonal form

3. PTD Approaches in Physical and Engineering Systems

3.1. Quantum Hamiltonians and Renormalization

The PTD framework unifies stepwise/unitary renormalization (discrete projector-based and continuous flow-equation perspectives), block-diagonalization, and perturbative expansions in quantum many-body theory. Generalized projection operators $P_\lambda$, $Q_\lambda$ are defined with respect to the instantaneous eigenstructure; a succession of unitary rotations $e^{X_{\lambda,\Delta\lambda}}$ eliminates off-block-diagonal elements with $X^{(n)}$ constructed recursively to cancel high-energy transitions. The resulting flow yields explicit renormalization-group equations for running parameters (e.g., in the Holstein or Fano–Anderson models) (Sykora et al., 2020).

3.2. Time-dependent Diagonalization

In time-dependent quantum systems, the PTD scheme is closely related to the time-dependent Schrieffer–Wolff transformation. The generator $S(t)$ is constructed order-by-order in the small off-diagonal parameter ($\epsilon\sim\|V\|/\Delta$), using projectors onto diagonal/off-diagonal sectors at each order:

$[H_0, S_n] + i\partial_t S_n = -Q_0 V^{(n)}(t)$

This yields effective Hamiltonians $H_{\mathrm{eff}}(t)$ diagonal up to desired order, capturing nontrivial phenomena such as tunable Bloch–Siegert shifts and parametric dispersive features (Xiao et al., 2021).

4. PTD in MIMO Communications and Signal Processing

A notable engineering highlight of PTD is channel diagonalization under arbitrary transmit covariance in multiple-input multiple-output (MIMO) systems. Given a MIMO channel $H$ and constraint $E[xx^H] = Q \succeq 0$, the PTD solution absorbs $Q^{1/2}$ into the channel to form $\Phi = H Q^{1/2}$, performs its truncated SVD, and reconstructs linear transceivers:

$V = Q^{1/2} V_\Phi, \quad U^H = U_\Phi^H$

guaranteeing both $V V^H = Q$ and exact diagonalization, achieving the constrained capacity $C(Q) = \log_2\det(I + H Q H^H)$. The method subsumes and generalizes classical SVD-based diagonalization and has demonstrated complete recovery of capacity in interference-limited and covariance-constrained scenarios (Liu et al., 2016).

5. PTD in Machine Learning: Robust Approximate Diagonalization

In state-space models (SSMs) for sequence modeling, PTD addresses the instability of direct diagonalization for non-normal or ill-conditioned state matrices (e.g., HiPPO-LegS). Exact diagonalization is ill-posed due to exponential ill-conditioning. PTD introduces a carefully chosen perturbation $E$ so that $A_H + E$ becomes suitable for backward-stable eigendecomposition. Only the transfer function $G(s) = C(sI - A)^{-1}B + D$ needs to be preserved up to small backward error, allowing robust approximations. This strongly improves the spectral and dynamical robustness of deep models such as S4-PTD and S5-PTD, with empirical improvements documented in long-range sequence benchmarks (Yu et al., 2023).

Model	ListOps	Text	Retrieval	Image	Pathfinder	Path-X	Avg
S4D	60.47%	86.18%	89.46%	88.19%	93.06%	91.95%	84.89%
S4-PTD	60.65%	88.32%	91.07%	88.27%	94.79%	96.39%	86.58%
S5	62.15%	89.31%	91.40%	88.00%	95.33%	98.58%	87.46%
S5-PTD	62.75%	89.41%	91.51%	87.92%	95.54%	98.52%	87.61%

6. Iterative and Dynamical PTD Methods

Beyond classical power-series or closed-form approaches, PTD admits iterative formulations. For Hermitian problems $H = H_0 + \epsilon V$, one can define fixed-point maps in projective space such that the eigenvector solutions correspond to attractors. This “dynamical diagonalization” is computationally advantageous, extends the domain of convergence beyond standard power-series, and outperforms conventional perturbative methods (e.g., Rayleigh–Schrödinger) both in rate of convergence and per-iteration cost, especially in large-scale high-precision settings (Kenmoe et al., 2020).

7. PTD for Ill-conditioned Generalized Eigenproblems

Quantum subspace algorithms for eigenvalue estimation produce generalized eigenpairs $(H, S)$ contaminated by noise far above machine precision. PTD regularizes the problem by threshold truncation of $S$, thus discarding spurious subspaces and transforming the generalized eigenproblem into a well-conditioned dense instance. Rigorous $O(\eta)$ error bounds in terms of noise scale $\eta$ and threshold $\varepsilon$ can be established, and practical guidance involves adaptive threshold selection for optimal stability and bias-variance trade-off (Epperly et al., 2021).

• (Liu et al., 2016): Closed-form MIMO PTD diagonalization (engineering/communications)
• (Mankodi et al., 2024): Block-diagonalization, least-action PTD, and differentiating Schrieffer–Wolff approaches
• (Yu et al., 2023): PTD for robust SSMs and deep sequence models, backward stability, transfer function bounds
• (Demmel et al., 2023): PTD for matrix pencil diagonalization, pseudospectral shattering, randomized eigensolvers
• (Xiao et al., 2021): PTD in time-dependent quantum systems, Schrieffer–Wolff, effective Hamiltonians
• (Sykora et al., 2020): Projector-based unitary renormalization and PTD in quantum many-body systems
• (Kenmoe et al., 2020): Dynamical perturbation theory, fixed-point PTD, improved convergence
• (Epperly et al., 2021): PTD for noisy quantum subspace diagonalization, thresholding, error bounds