Matrix Phase: Theory, Transitions & Applications

• Matrix Phase is a family of approaches that encode phase information via matrices, transforming scalar ambiguities into tractable recovery and analysis problems.
• It employs lifting techniques and statistical methods to reveal sharp phase transitions in matrix estimation and low-rank recovery tasks.
• Applications extend to entrywise phase constraints, optical phase control, and matrix phase-space formulations, impacting signal processing and quantum physics.

Searching arXiv for the primary paper and related matrix/phase-retrieval work. In the cited literature, matrix phase is not a single construction but a family of closely related ones in which phase information, phase transitions, or phase-space structure are encoded through matrices. The phrase appears in generalized phase retrieval, where quadratic measurements are lifted to rank-1 Hermitian matrix recovery (Wang et al., 2016); in low-rank recovery, where one studies sharp phase transitions of nuclear-norm or Bayes-optimal estimators (Donoho et al., 2013, Schülke et al., 2016); in entrywise phase-constrained matrix geometry through phase rank (Goucha et al., 2021); and in random-matrix, matrix-model, and matrix phase-space formulations of physical phases and symmetries (Vanderheyden et al., 2011, Rosso et al., 2021, Drummond et al., 11 Jun 2026). This suggests a common structural pattern: phase is made analyzable by replacing scalar descriptions with matrix-valued variables, matrix ensembles, or matrix kernels.

1. Scope of the term

Across the cited works, the term is used for several distinct technical objects rather than for a single invariant.

Usage	Core object	Representative paper
Generalized phase retrieval	Rank-1 Hermitian PSD lift $X=xx^*$	(Wang et al., 2016)
Low-rank recovery threshold	Phase transition curve $\delta^*(\rho;\beta)$	(Donoho et al., 2013)
Matrix compressed sensing	Easy, hard, impossible phases	(Schülke et al., 2016)
Entrywise phase geometry	$\operatorname{rank}_{\mathrm{phase}}(\Theta)$	(Goucha et al., 2021)
Random-matrix phase diagrams	Symmetry-constrained matrix ensembles	(Vanderheyden et al., 2011)
Matrix phase-space	Symmetry-projected matrix kernels	(Drummond et al., 11 Jun 2026)

The first major meaning is algebraic and inverse-problem oriented: a signal phase ambiguity is absorbed into an equivalence class, and the recovery task is reformulated as matrix recovery. A second meaning is asymptotic and statistical: a matrix problem exhibits sharp transitions between success and failure, or between easy, hard, and impossible regimes. A third meaning is geometric: the phase pattern of a matrix itself becomes the object of study. A fourth meaning is physical: phases of matter, spectral phases, or quantum phase-space dynamics are described by matrix ensembles or matrix kernels.

2. Generalized phase retrieval and the lifted matrix formulation

In generalized phase retrieval, one seeks to reconstruct $x \in \mathbb{F}^d$, with $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$, from quadratic samples

$y_i=x^*A_i x,\qquad i=1,\dots,N,$

with $A_i\in H_d(\mathbb{F})$, symmetric in the real case and Hermitian in the complex case. Recovery is modulo the natural phase ambiguity: $x\sim y$ iff $x=\pm y$ over $\mathbb{R}$, and $\delta^*(\rho;\beta)$0 iff $\delta^*(\rho;\beta)$1 over $\delta^*(\rho;\beta)$2. The measurement ensemble $\delta^*(\rho;\beta)$3 has the phase retrieval property when the map

$\delta^*(\rho;\beta)$4

is injective on the quotient space by this equivalence relation (Wang et al., 2016).

The decisive step is the lift

$\delta^*(\rho;\beta)$5

which turns the quadratic measurements into linear ones: $\delta^*(\rho;\beta)$6 Generalized phase retrieval therefore becomes recovery of a rank-1 Hermitian positive semidefinite matrix from linear measurements of the form $\delta^*(\rho;\beta)$7. In this form it contains standard phase retrieval as the special case $\delta^*(\rho;\beta)$8, where the data are $\delta^*(\rho;\beta)$9, and it also contains phase retrieval by orthogonal projections. Wang and Xu use this lift to connect phase retrieval to low-rank matrix recovery, to algebraic-geometric admissibility arguments, and to nonsingular bilinear forms (Wang et al., 2016).

The same paper gives several equivalent injectivity criteria. In the real case, phase retrieval is equivalent to nonsingularity of the bilinear form $\operatorname{rank}_{\mathrm{phase}}(\Theta)$0, to the condition that no nonzero $\operatorname{rank}_{\mathrm{phase}}(\Theta)$1 satisfy $\operatorname{rank}_{\mathrm{phase}}(\Theta)$2 for all $\operatorname{rank}_{\mathrm{phase}}(\Theta)$3, and to full-rank Jacobian $\operatorname{rank}_{\mathrm{phase}}(\Theta)$4 at every $\operatorname{rank}_{\mathrm{phase}}(\Theta)$5. In the complex case, the corresponding real Jacobian rank condition is $\operatorname{rank}_{\mathrm{phase}}(\Theta)$6 for all $\operatorname{rank}_{\mathrm{phase}}(\Theta)$7. The matrix-kernel formulation also yields spectral criteria in terms of rank-$\operatorname{rank}_{\mathrm{phase}}(\Theta)$8 Hermitian matrices $\operatorname{rank}_{\mathrm{phase}}(\Theta)$9 annihilated by all $x \in \mathbb{F}^d$0 (Wang et al., 2016).

For measurement numbers, the generic sufficiency bounds are $x \in \mathbb{F}^d$1 in the real case and $x \in \mathbb{F}^d$2 in the complex case for very general classes of $x \in \mathbb{F}^d$3, including prescribed-rank symmetric or Hermitian matrices, positive semidefinite matrices, and orthogonal projections. The same work gives lower bounds and exact minima in several dimensions: $x \in \mathbb{F}^d$4 for odd $x \in \mathbb{F}^d$5, $x \in \mathbb{F}^d$6 for even $x \in \mathbb{F}^d$7; $x \in \mathbb{F}^d$8; $x \in \mathbb{F}^d$9 for many odd $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$0; $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$1 for even $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$2, $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$3; and $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$4 or $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$5 for odd $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$6 according to the number of ones in the binary expansion of $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$7 (Wang et al., 2016).

Algorithmically, this lifted viewpoint is the same structural move used in PhaseLift, where one solves a convex semidefinite program over $\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$8 rather than working directly with nonconvex quadratic equations (Candes et al., 2011).

3. Recovery thresholds and phase transitions in matrix estimation

A second major meaning of matrix phase concerns phase transitions in matrix recovery. For low-rank matrix recovery from Gaussian linear measurements,

$\mathbb{F}\in\{\mathbb{R},\mathbb{C}\}$9

a widely studied convex estimator is nuclear norm minimization,

$y_i=x^*A_i x,\qquad i=1,\dots,N,$0

The relevant asymptotic parameters are the undersampling fraction $y_i=x^*A_i x,\qquad i=1,\dots,N,$1, the rank fraction $y_i=x^*A_i x,\qquad i=1,\dots,N,$2, and the aspect ratio $y_i=x^*A_i x,\qquad i=1,\dots,N,$3. Donoho, Gavish, and Montanari report a sharp empirical transition curve $y_i=x^*A_i x,\qquad i=1,\dots,N,$4 such that nuclear norm minimization typically succeeds above the curve and fails below it. Their central empirical statement is that this curve coincides with the asymptotic minimax denoising risk $y_i=x^*A_i x,\qquad i=1,\dots,N,$5 of singular-value soft-thresholding, so that

$y_i=x^*A_i x,\qquad i=1,\dots,N,$6

for Gaussian measurements and fixed asymptotic $y_i=x^*A_i x,\qquad i=1,\dots,N,$7 (Donoho et al., 2013).

That correspondence is quantitative. In the square case $y_i=x^*A_i x,\qquad i=1,\dots,N,$8, $y_i=x^*A_i x,\qquad i=1,\dots,N,$9 as $A_i\in H_d(\mathbb{F})$0, matching the Candès–Recht sufficient scaling in fractional form. For general $A_i\in H_d(\mathbb{F})$1,

$A_i\in H_d(\mathbb{F})$2

which explicitly shows easier recovery for tall-skinny matrices. The empirical transition is sharp: logistic slopes grow roughly linearly with $A_i\in H_d(\mathbb{F})$3, the transition width shrinks as $A_i\in H_d(\mathbb{F})$4, and reported finite-$A_i\in H_d(\mathbb{F})$5 fits typically agree with the predicted threshold within $A_i\in H_d(\mathbb{F})$6 in square and symmetric PSD settings (Donoho et al., 2013).

Matrix compressed sensing adds another layer: the unknown matrix is factorized as $A_i\in H_d(\mathbb{F})$7, measurements are $A_i\in H_d(\mathbb{F})$8, and Bayes-optimal inference is analyzed through replica equations and the state evolution of P-BiG-AMP. In this setting the phase diagram has three regimes. The easy phase has a unique low-error fixed point; the hard-but-possible phase has coexisting good and bad fixed points, so recovery is information-theoretically possible but algorithmically difficult; and the impossible phase has only a high-error optimum. In the noiseless symmetric rank-1 Gauss–Bernoulli case with global measurement rate $A_i\in H_d(\mathbb{F})$9, the information-theoretic threshold is $x\sim y$0, a stable bad fixed point at $x\sim y$1 persists for all $x\sim y$2, there is a second-order transition at $x\sim y$3, and a higher-$x\sim y$4 first-order jump to the easy phase (Schülke et al., 2016).

These results place the lifted phase-retrieval viewpoint inside a broader statistical picture. Recovering a rank-1 lifted matrix is one extremal point of a much larger landscape in which low-rank structure generates threshold phenomena, metastability, and algorithmic barriers.

4. Entrywise phase constraints and phase rank

A third meaning of matrix phase is intrinsic to the matrix itself. For a phase matrix $x\sim y$5, where every entry has modulus $x\sim y$6, the phase rank is

$x\sim y$7

Equivalently, one minimizes the rank of $x\sim y$8 over strictly positive real matrices $x\sim y$9. This notion generalizes sign rank, since for a sign matrix $x=\pm y$0 one has $x=\pm y$1, and it is complementary to phaseless rank, which fixes magnitudes and optimizes over phases (Goucha et al., 2021).

The paper proves several structural facts. Phase rank $x=\pm y$2 is equivalent to ordinary complex rank $x=\pm y$3. For $x=\pm y$4 with $x=\pm y$5, $x=\pm y$6 iff there exists a row scaling by scalars in $x=\pm y$7, not all zero, such that no column is colopsided. A volumetric bound then gives: if $x=\pm y$8, every $x=\pm y$9 phase matrix is phase-rank deficient, and in particular for $\mathbb{R}$0 no $\mathbb{R}$1 phase matrix has phase rank $\mathbb{R}$2 (Goucha et al., 2021).

The most complete result is the $\mathbb{R}$3 classification. Let $\mathbb{R}$4 be the six signed monomial evaluations from the Leibniz expansion of $\mathbb{R}$5. Then

$\mathbb{R}$6

Equivalently, determinant colopsidedness completely characterizes the $\mathbb{R}$7 phase-rank-$\mathbb{R}$8 locus. The same theorem implies that the coamoeba of the $\mathbb{R}$9 determinantal variety is determined solely by the colopsidedness criterion (Goucha et al., 2021).

This line of work changes the role of phase. Instead of recovering an unknown phase from measurements, it studies the rank geometry of matrices whose entrywise phases are prescribed. That shift connects matrix phase to coamoebas, ray nonsingularity, convex hull tests, and the complexity theory inherited from sign rank.

5. Measurement-matrix engineering and optical phase control

In inverse problems and photonic implementations, matrix phase also refers to the design or control of measurement matrices and phase shifts. For phase retrieval with measurements

$\delta^*(\rho;\beta)$00

Abdelhadi, Wornell, and Chan propose choosing a deterministic measurement matrix $\delta^*(\rho;\beta)$01 by maximizing the mutual information $\delta^*(\rho;\beta)$02. They recast the problem exactly as a linear MIMO channel on the lifted variable $\delta^*(\rho;\beta)$03, derive necessary optimality conditions for $\delta^*(\rho;\beta)$04, and show that in the low-SNR regime, if $\delta^*(\rho;\beta)$05 is Kronecker symmetric with covariance $\delta^*(\rho;\beta)$06, an optimal matrix is rank $\delta^*(\rho;\beta)$07: $\delta^*(\rho;\beta)$08 where $\delta^*(\rho;\beta)$09 is the principal eigenvector of $\delta^*(\rho;\beta)$10 and $\delta^*(\rho;\beta)$11. They also give practical design procedures for unconstrained and masked Fourier structures, with simulation gains of about $\delta^*(\rho;\beta)$12–$\delta^*(\rho;\beta)$13 dB over Gaussian measurements in some settings and lower sample complexity in both structured and Gaussian priors (Shlezinger et al., 2017).

A later data-driven variant is DLMMPR, which replaces fixed coded-diffraction operators by a learned measurement operator

$\delta^*(\rho;\beta)$14

inside a seven-stage unfolded network. The training objective uses the amplitude model $\delta^*(\rho;\beta)$15, with a subgradient-descent fidelity step and a learned proximal projection module. On $\delta^*(\rho;\beta)$16 images with four masks and Poisson intensity levels $\delta^*(\rho;\beta)$17, the paper reports average PSNR/SSIM improvements over DeepMMSE and PrComplex, including $\delta^*(\rho;\beta)$18 on UNT-6 at $\delta^*(\rho;\beta)$19, $\delta^*(\rho;\beta)$20 at $\delta^*(\rho;\beta)$21, and $\delta^*(\rho;\beta)$22 at $\delta^*(\rho;\beta)$23 (Liu et al., 16 Nov 2025).

In programmable photonic matrix-vector multiplication, the relevant phase variable is the programmable phase excursion of interferometer meshes. Perturbative programming operates the circuit near a fixed reference configuration and implements the target matrix through interferometric subtraction. For dense Ginibre targets normalized to $\delta^*(\rho;\beta)$24, the paper reports $\delta^*(\rho;\beta)$25 contraction of perturbative phase distributions, including near-extreme tails, and gives an entropic lower bound

$\delta^*(\rho;\beta)$26

The trade-off is a fixed subtraction overhead

$\delta^*(\rho;\beta)$27

with minimum $\delta^*(\rho;\beta)$28 at $\delta^*(\rho;\beta)$29 (Fldzhyan et al., 1 Jun 2026).

In this engineering literature, matrix phase is operational rather than purely algebraic: the central questions are how the measurement matrix should be designed, how the phase degrees of freedom of optical hardware should be distributed, and how noise, conditioning, and structural constraints affect recoverability.

6. Random matrices, matrix models, and physical phase structure

In physics, matrix phase often means a phase diagram generated by a matrix ensemble. A random-matrix approach to phase diagrams replaces a specific microscopic Hamiltonian by an ensemble constrained only by the relevant global symmetries. The resulting saddle-point thermodynamic potentials determine allowed and forbidden phase topologies, mean-field transitions, and mixed-phase criteria. In the competing-order expansion

$\delta^*(\rho;\beta)$30

the sign of

$\delta^*(\rho;\beta)$31

determines whether homogeneous mixed phases are allowed ($\delta^*(\rho;\beta)$32) or forbidden ($\delta^*(\rho;\beta)$33) (Vanderheyden et al., 2011).

For fuzzy scalar field theories, the associated multitrace matrix models exhibit three phases: a disordered symmetric one-cut phase, a uniform ordered asymmetric one-cut phase, and a non-uniform/striped two-cut phase. In the single-trace quartic model the one-cut to two-cut transition is at

$\delta^*(\rho;\beta)$34

After nonperturbative second-moment packing on the fuzzy sphere,

$\delta^*(\rho;\beta)$35

the symmetric transition line becomes

$\delta^*(\rho;\beta)$36

showing an essential singularity near the origin (Tekel, 2015).

Deformations of $\delta^*(\rho;\beta)$37 JT supergravity provide another matrix-model phase structure. The matrix completion of the theory exhibits a universal transition at $\delta^*(\rho;\beta)$38: in Type $\delta^*(\rho;\beta)$39, a single-cut phase turns into a double-cut phase with a gap $\delta^*(\rho;\beta)$40; in Type $\delta^*(\rho;\beta)$41, a hard-edge phase turns into a soft-edge phase with $\delta^*(\rho;\beta)$42. In both cases the transition is second order in the gravitational free energy. A further nonperturbative phase appears for $\delta^*(\rho;\beta)$43, where the $\delta^*(\rho;\beta)$44 theory develops a true spectral gap $\delta^*(\rho;\beta)$45 (Rosso et al., 2021).

The Penner matrix model with negative coupling shows that even the existence of a planar limit may depend on how the sequence of couplings approaches the large-$\delta^*(\rho;\beta)$46 regime. Ordinary ’t Hooft sequences fail because oscillatory Barnes-$\delta^*(\rho;\beta)$47 contributions prevent convergence. Kuijlaars–McLaughlin sequences restore a planar limit and produce a two-dimensional phase space $\delta^*(\rho;\beta)$48, with a first-order line at $\delta^*(\rho;\beta)$49 for $\delta^*(\rho;\beta)$50, a continuous transition at $\delta^*(\rho;\beta)$51, and a singular line $\delta^*(\rho;\beta)$52 interpreted as gap closing, eigenvalue tunneling, and Bose condensation (Álvarez et al., 2017). A related lesson from other ensembles is that exact equality of finite-$\delta^*(\rho;\beta)$53 partition functions does not imply identical large-$\delta^*(\rho;\beta)$54 phase diagrams: second- and third-order discrepancies can occur between equivalent matrix representations (Santilli et al., 2020).

7. Matrix phase-space and symmetry-resolved formulations

A final family of uses places phase directly in phase-space formulations. In matrix phase-space representations for quantum symmetries, a density operator is expanded as

$\delta^*(\rho;\beta)$55

where $\delta^*(\rho;\beta)$56, $\delta^*(\rho;\beta)$57 is a stochastic $\delta^*(\rho;\beta)$58 density matrix over symmetry sectors, and the kernel is projected by symmetry projectors $\delta^*(\rho;\beta)$59. This construction is complete for bosonic systems, unifies scalar phase-space methods with symmetry projection, and, in low-loss Gaussian boson sampling with photon-number-resolving detectors, parity symmetry reduces sampling errors by very large factors relative to earlier methods (Drummond et al., 11 Jun 2026). A closely related quantum-optics formulation reports improvements in sampling error by a factor of $\delta^*(\rho;\beta)$60 or more compared to unprojected methods and applies this to verification of low-loss GBS experiments with up to $\delta^*(\rho;\beta)$61 modes (Drummond et al., 17 Mar 2025).

A different operator-kernel viewpoint writes the Wigner function as

$\delta^*(\rho;\beta)$62

where $\delta^*(\rho;\beta)$63 is the state-specific density matrix and $\delta^*(\rho;\beta)$64 is a universal density matrix in phase space. In the harmonic-oscillator basis, the diagonal entries of $\delta^*(\rho;\beta)$65 are the usual oscillator Wigner functions, while the off-diagonal entries carry oscillatory factors $\delta^*(\rho;\beta)$66 along energy shells and encode coherence between levels (Perepelkin et al., 2019).

At a more structural level, the phase-space formulation of the $\delta^*(\rho;\beta)$67-matrix identifies the adjoint action of $\delta^*(\rho;\beta)$68 with a fuzzy diffeomorphism $\delta^*(\rho;\beta)$69 of the noncommutative $\delta^*(\rho;\beta)$70-algebra of symbols, with

$\delta^*(\rho;\beta)$71

and classical limit $\delta^*(\rho;\beta)$72. The quantum eikonal $\delta^*(\rho;\beta)$73 is the $\delta^*(\rho;\beta)$74-deformation of the classical Magnus generator obtained by replacing Poisson brackets with $\delta^*(\rho;\beta)$75-brackets (Kim, 28 Dec 2025).

The same theme appears in matrix models of dynamical phase space. An $\delta^*(\rho;\beta)$76-invariant matrix model derived from generalized Yang–Mills theory on the standard Courant algebroid organizes its degrees of freedom as

$\delta^*(\rho;\beta)$77

admits a Born-reciprocity symmetry $\delta^*(\rho;\beta)$78, and has classical solutions corresponding to noncommutative curved phase spaces. In this interpretation, emergent gravity is tied to the mixed position–momentum commutators and the associated frame fields rather than to position matrices alone (Chatzistavrakidis, 2014).

Taken together, these formulations extend matrix phase from inverse problems and random ensembles into operator geometry. Phase is no longer only a missing scalar degree of freedom; it becomes a matrix-resolved coordinate of symmetry sectors, a kernel on phase space, or an intrinsic part of a dynamical position–momentum algebra.