Constrained LiGME Model: Theory & Applications

• The constrained LiGME model extends regularized estimations by merging nonconvex penalties with hard convex constraints to preserve overall convexity and ensure global minimizers.
• It employs a Generalized Moreau Enhancement to refine convex penalties while enforcing operator inequalities that balance data fidelity with regularization.
• Recent developments include discrete-valued extensions and nonquadratic data-fidelity functions, achieving superior signal recovery in diverse applications.

The constrained LiGME model denotes a class of regularized estimation models built from the Linearly involved Generalized Moreau Enhanced (LiGME) construction, in which nonconvex regularizers are introduced without losing the overall convexity of the optimization problem, while convex prior information is imposed as hard constraints. In its canonical form, LiGME combines a data-fidelity term with a Generalized Moreau Enhancement (GME) of a convex penalty, and the constrained extension, usually abbreviated cLiGME, augments this structure with one or more convex feasibility conditions on the signal or on linear transforms of the signal (Abe et al., 2019, Yata et al., 2021). Subsequent developments enlarged the framework to simultaneous early and asymptotic constraints, discrete-valued estimation with designated isolated minimizers, and smooth nonquadratic data-fidelity functions (Yata et al., 2023, Shoji et al., 13 Mar 2025, Yata et al., 3 Sep 2025).

1. Conceptual lineage and scope

LiGME was introduced as a unified extension of regularized least-squares models that exploit nonconvex penalties while preserving overall convexity, with emphasis on sparsity- and rank-aware signal processing (Abe et al., 2019). The convexly constrained LiGME model then formalized the incorporation of multiple a priori knowledge sources as hard convex constraints, rather than as additional penalty terms (Yata et al., 2021). Later work refined the constraint semantics by distinguishing “early” constraints, enforced at every iteration, from “asymptotic” constraints, required only at the limit point of the iterative sequence (Yata et al., 2023). Other extensions specialized the framework to discrete-valued signal estimation and to nonquadratic smooth convex data fidelities such as Poisson and clipped-Gaussian models (Shoji et al., 13 Mar 2025, Yata et al., 3 Sep 2025).

Development	Main addition	Representative paper
LiGME	Nonconvex regularizers under overall convexity	(Abe et al., 2019)
cLiGME	Multiple hard convex constraints	(Yata et al., 2021)
EA-type constrained LiGME	Early and asymptotic constraints simultaneously	(Yata et al., 2023)
LiGME for discrete alphabets	Designated isolated minimizers, cLiGME variant	(Shoji et al., 13 Mar 2025)
Nonquadratic cLiGME extension	Smooth convex $f(Ax)$ beyond least squares	(Yata et al., 3 Sep 2025)

The unifying theme across these formulations is that the nonconvexity resides in the regularizer, whereas tractability is retained through algebraic conditions on the model operators. This design differs from unconstrained nonconvex penalization schemes that typically provide only critical-point guarantees; the LiGME line instead targets global minimization under explicit convexity-preserving conditions. This suggests a distinctive methodological position: LiGME is neither a purely convex regularizer nor an unconstrained nonconvex program, but a structured compromise between the two (Abe et al., 2019, Yata et al., 2021).

2. Canonical mathematical formulation

The foundational LiGME objective is

$$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$

where $x$ is the unknown signal, $y$ is the observation, $A$ is a known linear operator, $\mu>0$ is a regularization parameter, $\mathfrak{L}$ is a linear feature operator, and $\Psi$ is a proper, lower semicontinuous, convex penalty. The LiGME penalty is the Generalized Moreau Enhancement

$$\Psi_B(z) := \Psi(z) - \min_{v} \left[ \Psi(v)+\frac{1}{2}\|B(z-v)\|^2 \right],$$

with $B$ a tunable linear operator controlling the enhancement (Abe et al., 2019).

The convexly constrained LiGME model extends this by imposing multiple convex feasibility conditions: $$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$0 Here $$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$1 are linear operators and $$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$2 are closed convex sets encoding a priori information such as value ranges or structural relations (Yata et al., 2021).

A compact product-space form is obtained by defining

$$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$3

so that $$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$4 is equivalent to $$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$5 for all $$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$6. Introducing the indicator $$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$7, one may set

$$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$8

which yields the reformulation

$$J_{\Psi_B \circ \mathfrak{L}}(x) = \frac{1}{2}\|y-Ax\|^2 + \mu \Psi_B(\mathfrak{L}x),$$9

with

$x$0

This product-space representation is central because it keeps the constraint part explicitly convex while leaving the LiGME enhancement confined to the regularization channel (Yata et al., 2021).

The framework also admits sums of multiple LiGME penalties. In the foundational product-space treatment, the objective can be written as

$x$1

and later constrained variants use

$x$2

thereby combining multiple nonconvex enhancements with multiple convex constraints in a single model (Abe et al., 2019, Yata et al., 2023).

3. Overall convexity and the role of the GME operator

The central structural requirement in LiGME theory is the overall convexity condition. In the basic least-squares setting, the objective remains convex if

$x$3

For multi-penalty formulations, the corresponding condition becomes

$x$4

These inequalities express that the curvature contributed by the data-fidelity term must dominate the nonconvexity introduced through the GME matrices $x$5 or $x$6 (Abe et al., 2019, Yata et al., 2023).

This condition is the formal mechanism by which LiGME permits nonconvex regularization without sacrificing global convex solvability. A common misconception is that the LiGME penalty makes the whole objective nonconvex in the usual sense. The actual construction is more specific: $x$7 can be nonconvex, but the composite objective is designed to stay convex under the stated operator inequality. This is not a rhetorical distinction; it is the premise behind the global convergence claims of the associated splitting algorithms (Abe et al., 2019, Yata et al., 2021).

The discrete-valued extension makes this geometry particularly explicit. For a finite alphabet $x$8, the LiGME regularizer is defined as a GME of the sum-of-absolute-values (SOAV) convex regularizer,

$x$9

and if $y$0, the model reduces to the convex SOAV regularizer (Shoji et al., 13 Mar 2025). In the special case $y$1, the regularizer separates coordinatewise,

$y$2

where each $y$3 is a sum of MCP terms. Under the condition

$y$4

every vector in $y$5 is an isolated local minimizer of $y$6, while the overall convexity of the regularized least-squares model is preserved provided

$y$7

An explicit admissible construction is

$y$8

This suggests a characteristic LiGME design principle: local attraction toward desirable structures can be strengthened, but only within a curvature budget dictated by the observation operator (Shoji et al., 13 Mar 2025).

4. Algorithms and convergence guarantees

The original LiGME paper proposed a proximal splitting type algorithm under the overall-convexity condition, using primal and dual variables $y$9 and proximity operators associated with $A$0 and $A$1 (Abe et al., 2019). The constrained extension adapted this operator-splitting viewpoint to hard convex constraints. In the cLiGME formulation, the proximity operator of $A$2 decomposes as

$A$3

so the algorithm alternates proximal steps for the penalty with projection steps for the constraints (Yata et al., 2021).

The 2021 cLiGME algorithm is formulated as a proximal splitting method in product space. Its convergence analysis relies on averaged nonexpansive operators and explicit parameter conditions on $A$4, with $A$5. Under these choices and the convexity condition, the iterates converge to a global minimizer of the cLiGME problem (Yata et al., 2021).

The 2023 constrained LiGME model introduced a fixed-point method of Krasnosel’skiĭ-Mann type that can enforce two kinds of constraints simultaneously. Let

$A$6

The “early” constraint requires $A$7 for all iterations, while the “asymptotic” constraint requires only that the limit $A$8 belong to $A$9. Using auxiliary variables $\mu>0$0, the algorithm constructs iterates $\mu>0$1 through projections $\mu>0$2, $\mu>0$3 and proximities $\mu>0$4. Its convergence theorem states:

1. $\mu>0$5 for all $\mu>0$6;
2. $\mu>0$7;
3. the limit point is a global minimizer of $\mu>0$8 over $\mu>0$9 (Yata et al., 2023).

The nonquadratic extension preserved the same global-minimization orientation. For

$\mathfrak{L}$0

with $\mathfrak{L}$1, differentiable, convex, and with Lipschitz continuous gradient, an inner-loop free algorithm was derived on a product space involving primal, dual, and auxiliary variables. Under the corresponding overall convexity and existence conditions, the whole sequence converges to a global minimizer, rather than merely producing cluster-point or critical-point guarantees (Yata et al., 3 Sep 2025).

5. Constraint semantics and specialized variants

The distinction between early and asymptotic constraints is one of the most explicit conceptual additions to the constrained LiGME literature. In the EA-type constrained problem,

$\mathfrak{L}$2

the early constraint is attached to the base feasible set $\mathfrak{L}$3, while the asymptotic constraint is attached to the split-feasibility-type condition $\mathfrak{L}$4. The practical significance given in the paper is that every intermediate estimate remains feasible with respect to the simple constraint, which is useful if the algorithm must be stopped early, whereas the more complicated structural condition is guaranteed only at convergence (Yata et al., 2023).

This EA-type framework was instantiated in a bivariate nonconvex enhancement of fused lasso models for sparse piecewise constant signal estimation. With variables $\mathfrak{L}$5 and $\mathfrak{L}$6, and $\mathfrak{L}$7, the model constrains

$\mathfrak{L}$8

while applying GME-based nonconvex penalties to both the signal and its difference or latent components. In the formulation summarized in the data, the early constraint acts on $\mathfrak{L}$9 and the asymptotic constraint on $\Psi$0. Numerical evidence reported lower reconstruction error, better recovery of step edges, and robustness to baseline shifts relative to classical fused lasso and latent fused lasso (Yata et al., 2023).

A different specialization targeted discrete-valued estimation. There the cLiGME machinery is used as a tractable relaxation over a convex superset $\Psi$1 of $\Psi$2, followed by projection to the alphabet. Two modifications were introduced to improve practical accuracy. The first is iterative reweighting,

$\Psi$3

which updates the SOAV-type weights adaptively. The second is generalized superiorization,

$\Psi$4

which perturbs the current iterate toward the nearest discrete-valued point. The paper states that cLiGME outperforms convex SOAV in bit error rate, and that iterative reweighting and generalized superiorization further improve performance in MIMO signal detection (Shoji et al., 13 Mar 2025).

6. Nonquadratic data fidelity, empirical behavior, and terminological clarification

The most general formulation in the supplied material replaces the least-squares term by a smooth convex function: $\Psi$5 where $\Psi$6 is differentiable, convex, and has Lipschitz continuous gradient (Yata et al., 3 Sep 2025). The associated overall convexity condition is expressed as

$\Psi$7

or equivalently by requiring $\Psi$8 to be convex. The existence theorem reported for this model gives four sufficient conditions, including coercivity together with asymptotically multipolyhedral constraints, the condition $\Psi$9, the condition $$\Psi_B(z) := \Psi(z) - \min_{v} \left[ \Psi(v)+\frac{1}{2}\|B(z-v)\|^2 \right],$$0 with $$\Psi_B(z) := \Psi(z) - \min_{v} \left[ \Psi(v)+\frac{1}{2}\|B(z-v)\|^2 \right],$$1 bounded below, and boundedness of $$\Psi_B(z) := \Psi(z) - \min_{v} \left[ \Psi(v)+\frac{1}{2}\|B(z-v)\|^2 \right],$$2 (Yata et al., 3 Sep 2025).

Empirically, the papers consistently describe cLiGME-type models as improving estimation quality relative to convex baselines when prior structure is well matched to the application. In the foundational LiGME experiments, the framework was validated on piecewise constant 1D recovery, image deblurring, matrix completion, and joint piecewise-constancy/low-rank recovery, with lower mean squared error, sharper edge recovery, and more accurate rank recovery reported for the LiGME models (Abe et al., 2019). In the convexly constrained image-restoration study, cLiGME with box constraints $$\Psi_B(z) := \Psi(z) - \min_{v} \left[ \Psi(v)+\frac{1}{2}\|B(z-v)\|^2 \right],$$3 and a background equality constraint yielded lower mean squared error than plain total variation, and the paper states that cLiGME reached a target error level $$\Psi_B(z) := \Psi(z) - \min_{v} \left[ \Psi(v)+\frac{1}{2}\|B(z-v)\|^2 \right],$$4 that TV could not always achieve when constraints were imposed (Yata et al., 2021). In the nonquadratic extension, Poisson denoising and simultaneous declipping and denoising experiments were reported to show lower mean absolute and squared errors and a superior bias/sparsity tradeoff for the proposed nonconvex model relative to convex approaches (Yata et al., 3 Sep 2025).

A recurrent source of confusion is terminological rather than mathematical. The acronym “LiGME” refers to “Linearly involved Generalized Moreau Enhanced.” It should not be conflated with “LME” in the separate literature on constrained linear mixed effects models. For example, “Estimating Linear Mixed Effects Models with Truncated Normally Distributed Random Effects” studies sign-constrained mixed-effects regression through symmetric doubly truncated normal random effects, not the LiGME regularization framework (Chen et al., 2020). The two topics share the word “constrained,” but they concern different model classes, different objective functions, and different inferential settings.

Taken together, the constrained LiGME literature describes a coherent program in which nonconvex regularizers, hard convex constraints, and operator-splitting algorithms are engineered jointly. The core invariant across its variants is the preservation of overall convexity, which is the condition enabling global-minimizer-oriented algorithms while still admitting enhancement mechanisms that are absent from standard convex regularization (Abe et al., 2019, Yata et al., 2021).