P-Stationary Points in ℓ0 Factor Analysis

• P-stationary points are defined for ℓ0-regularized factor analysis via a proximal fixed-point condition using hard-thresholding to enforce exact sparsity.
• They replace classical subdifferential inclusions with an exact proximal map that identifies zero entries, preserving the support structure of the sparse component.
• P-stationarity integrates traditional KKT optimality for smooth components with nonconvex sparsity constraints, ensuring that global minimizers satisfy local optimality conditions under strict regularity.

Searching arXiv for the cited paper and closely related work on proximal stationarity and $\ell_0$ optimization. A P-stationary point—short for proximal-stationary point—is a stationarity notion introduced for the nonconvex, nonsmooth $\ell_0$-regularized factor-analysis problem studied in " $\ell_0$ Factor Analysis: A P-Stationary Point Theory" (Wang et al., 2024). In that setting, the objective is to decompose a sample covariance matrix into a low-rank component and a sparse component by solving

$$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$

where $f$ is smooth and convex and $\|S\|_0$ counts the nonzero entries of $S$ (Wang et al., 2024). Because the $\ell_0$ term is discrete and nonconvex, classical first-order KKT conditions do not directly provide an adequate stationarity concept for the $S$-variable. P-stationarity addresses this by replacing a subdifferential inclusion with a proximal fixed-point condition built from the hard-thresholding proximal map (Wang et al., 2024).

1. Problem setting and motivation

The underlying model arises in factor analysis for stationary time series, where one seeks a parsimonious covariance representation through a hidden low-rank plus sparse structure (Wang et al., 2024). The same parsimonious modeling perspective is also described as important in systems and control (Wang et al., 2024). The optimization model combines a smooth convex term $f(L,S)$ with an $\ell_0$0 penalty on the sparse component $\ell_0$1, under semidefinite and strict-positivity constraints on $\ell_0$2.

Within this formulation, the variable $\ell_0$3 encodes the low-rank-related part through the feasible set constraint $\ell_0$4, while $\ell_0$5 is constrained by $\ell_0$6 and regularized by $\ell_0$7 (Wang et al., 2024). The condition $\ell_0$8 prevents degeneracy of the total covariance model. The nonconvexity does not primarily come from the smooth term $\ell_0$9, which is assumed smooth and convex, but from the combinatorial nature of $\ell_0$0.

This structure motivates a stationarity concept that preserves the familiar KKT logic for the cone constraints while handling sparsity by means of the exact proximal operator of $\ell_0$1. The paper identifies this as the natural analogue of first-order optimality for the problem class (Wang et al., 2024).

2. Formal definition

Definition 3 in the paper states that a feasible pair $\ell_0$2 is a P-stationary point of $\ell_0$3 if there exist symmetric multiplier matrices $\ell_0$4, $\ell_0$5 and a scalar $\ell_0$6 such that the following conditions hold (Wang et al., 2024):

1. Primal feasibility:

$\ell_0$7

2. Dual feasibility:

$\ell_0$8

3. Complementary slackness:

$\ell_0$9

4. Stationarity in $$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$0:

$$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$1

5. Proximal stationarity in $$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$2:

$$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$3

The proximal operator is applied element-wise through the hard-thresholding rule

$$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$4

This makes the sparsity mechanism explicit: the threshold $$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$5 determines whether an entry is forced to zero or retained (Wang et al., 2024).

The definition is asymmetric in $$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$6 and $$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$7. For $$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$8, ordinary gradient stationarity with semidefinite multipliers is sufficient. For $$(\mathcal P)\qquad \min \; F(L,S):=f(L,S)+C\|S\|_0 \quad \text{subject to } (L,S)\in D:=\{L\succeq 0,\; S\succeq 0,\; L+S\succ 0\},$$9, the $f$0 term necessitates a fixed-point condition under the proximal map, which encodes exact hard thresholding rather than convex relaxation.

3. Interpretation as a KKT analogue for $f$1

The paper interprets conditions (a)–(c) as the usual linear-constraint KKT conditions for $f$2 and $f$3 (Wang et al., 2024). Condition (d) is the vanishing of the Lagrangian gradient with respect to $f$4. The distinctive element is condition (e), which replaces the nonexistent subdifferential inclusion for $f$5 by a proximal-map fixed-point relation.

The paper also presents the equivalent inclusion

$f$6

with

$f$7

and equivalently

$f$8

Because $f$9 is nonconvex and discrete, this proximal fixed-point condition is presented as the natural analogue of a first-order condition of the form $\|S\|_0$0 in the nonsmooth setting (Wang et al., 2024).

The comparison with more classical stationarity notions is explicit. When $\|S\|_0$1 is replaced by a smooth or convex nonsmooth term, the $\|S\|_0$2-condition reduces to the standard subgradient inclusion

$\|S\|_0$3

The paper further notes that, in Clarke generalized-gradient language, one could require

$\|S\|_0$4

but P-stationarity is described as stronger and more explicit because it identifies which entries of $\|S\|_0$5 are zero through the threshold $\|S\|_0$6 and matches the exact form of the $\|S\|_0$7 proximal operator (Wang et al., 2024).

A plausible implication is that P-stationarity is not merely a formal substitute for KKT, but a structure-preserving optimality notion tailored to exact sparsity, since it retains the combinatorial support information that convex surrogates typically blur.

4. Existence of solutions and the link to optimality

The paper establishes that $\|S\|_0$8 has at least one global minimizer under its standing assumptions: $\|S\|_0$9 jointly strictly convex and smooth, $S$0 lower-semicontinuous, and $S$1 nonempty (Wang et al., 2024). It also states that the set of all global minimizers is bounded. The proof sketch proceeds by showing that $S$2 is lower semicontinuous, that sublevel sets are closed and bounded because the objective diverges on the boundary or at infinity, and then invoking the Weierstrass theorem (Wang et al., 2024).

The central optimality result is Theorem 2, which gives a two-way relation between minimizers and P-stationary points under additional regularity. Assuming $S$3 is Lipschitz on a level set with constant $S$4, and $S$5 is strictly convex, the theorem states (Wang et al., 2024):

• If $S$6 is a global minimizer and $S$7, then for any $S$8 there exist multipliers $S$9 such that $\ell_0$0 satisfies the P-stationarity conditions.
• Conversely, any P-stationary point is a local minimizer of $\ell_0$1.

The proof sketch given in the paper is informative about the role of the definition. For necessity, the optimization is split into a convex subproblem in $\ell_0$2 and a convex subproblem in $\ell_0$3, using classical KKT for the first and a KKT-plus-proximal argument for the second, with strong convexity and Lipschitz continuity controlling the proximal map (Wang et al., 2024). For sufficiency, the proof partitions the index set of entries of $\ell_0$4 into zero and nonzero entries at $\ell_0$5, builds a local neighborhood that preserves the zero pattern, and then uses convexity, complementary slackness, and proximal properties to show local minimality (Wang et al., 2024).

This yields a particularly sharp characterization: global minimizers satisfy P-stationarity under the stated assumptions, while P-stationary points are guaranteed to be local minimizers. The paper therefore places P-stationarity between exact optimality and generic first-order criticality.

5. Threshold structure and local support identification

The hard-thresholding form of the proximal map gives P-stationarity an entrywise interpretation. Section III.A recalls that $\ell_0$6 is the hard-thresholding operator, and the paper notes that P-stationarity explicitly “locks in” which entries of $\ell_0$7 are zero through inequalities of the form

$\ell_0$8

(Wang et al., 2024). The exact zero pattern is therefore encoded by the balance between the smooth gradient term, the multiplier term, and the threshold set by $\ell_0$9 and $S$0.

Remark 3.1 emphasizes a limiting case: when $S$1 is positive definite and hence has no zero entries, the proximal condition reduces to the usual gradient-equals-zero stationarity (Wang et al., 2024). This shows that P-stationarity is compatible with the classical smooth picture when sparsity is absent, while remaining meaningful when zero entries are essential to the model.

The paper also discusses the role of the parameter $S$2. The necessity of choosing $S$3 small enough, specifically $S$4 in Theorem 2(1), is illustrated numerically: if $S$5 is too large, the proximal step immediately zeros out all entries; if too small, the proximal step is nearly the identity and no sparsity is enforced (Wang et al., 2024). This suggests that P-stationarity formalizes a scale-dependent trade-off: the data-fit signal carried by $S$6 competes directly against the sparsity-inducing hard-threshold level $S$7.

6. Role in ADMM design and convergence analysis

The paper uses P-stationarity not only as an optimality concept but also as the target limit condition for its algorithmic framework. In Section IV, an ADMM algorithm is designed for $S$8 by introducing splitting variables $S$9 and $f(L,S)$0 and augmenting the Lagrangian (Wang et al., 2024). The $f(L,S)$1-update is given by

$f(L,S)$2

The convergence analysis is organized around the same proximal structure. By choosing $f(L,S)$3, Lemma 6 shows that the augmented Lagrangian $f(L,S)$4 decreases sufficiently at each iteration (Wang et al., 2024). Proposition 4 establishes boundedness of the iterates, and Proposition 5 shows that successive differences converge to zero (Wang et al., 2024). Proposition 6 then states that any cluster point $f(L,S)$5 of the full ADMM sequence satisfies the limit-point KKT conditions of each subproblem (Wang et al., 2024).

The paper’s key conclusion is that these limit-point KKT conditions are exactly the P-stationarity conditions of Definition 3 for $f(L,S)$6 (Wang et al., 2024). Hence every limit point of the ADMM is a P-stationary point, and by Theorem 2, a local minimizer. In this way, P-stationarity functions as the bridge between the nonconvex optimization model and the asymptotic interpretation of the algorithm’s output.

7. Conceptual status and related interpretations

P-stationarity is presented as the “right” analogue of KKT for problems involving $f(L,S)$7 (Wang et al., 2024). The reason is structural rather than rhetorical: the usual KKT template remains intact for the cone constraints and the smooth $f(L,S)$8-component, while the $f(L,S)$9-component is expressed through the exact proximal geometry of the $\ell_0$00 penalty. This avoids invoking a direct subdifferential set-inclusion for an object that is discrete and nonconvex.

A common misconception is to treat any criticality notion for $\ell_0$01 problems as interchangeable with convex-subgradient stationarity. The formulation in (Wang et al., 2024) shows that this is not the case. P-stationarity is stronger and more explicit than a generalized-gradient condition because it specifies support behavior through the hard-thresholding rule itself. The zero pattern is therefore part of the stationarity statement, not an incidental byproduct.

Another possible misconception is that proximal conditions are merely algorithmic artifacts. In the framework of (Wang et al., 2024), the proximal fixed-point is instead an intrinsic variational characterization: it appears in the definition of stationary points, in the optimality theory connecting stationary points to local minimizers, and in the ADMM convergence proof. A plausible implication is that the concept is valuable precisely because it unifies modeling, theory, and computation within the same nonsmooth nonconvex formalism.

Taken together, these results position P-stationary points as a rigorous first-order notion for exact-sparsity factor-analysis models, combining primal-dual feasibility, complementary slackness, classical smooth stationarity in $\ell_0$02, and hard-thresholding fixed-point stationarity in $\ell_0$03 (Wang et al., 2024).