Generalized Concave KL Property

• Generalized Concave KL property is a framework that extends classical KL analysis using concave, potentially nondifferentiable desingularizers.
• It introduces the exact modulus as a minimal concave desingularizer to capture intrinsic geometric regularity near critical points.
• Calibrated calculus rules and algorithmic applications yield optimal finite-length bounds and sharper convergence rate estimates.

The generalized concave Kurdyka–Łojasiewicz (KL) property is an extension of the classical KL property, central to the modern analysis of convergence rates in nonconvex and nonsmooth optimization. This property employs general (possibly nondifferentiable) concave desingularizing functions, and is equipped with a canonical "exact modulus"—the minimal such function—which captures intrinsic geometric regularity near critical points. The theory enables optimal finite-length and convergence rate bounds for a wide range of first-order algorithms.

1. Formal Definition of the Generalized Concave KL Property

Let $f:\mathbb{R}^n \to (-\infty, +\infty]$ be a proper, lower semicontinuous function, and let $\partial f(x)$ denote its limiting (Mordukhovich) subdifferential. Fix $\bar{x} \in \operatorname{dom} \partial f$. For $\eta > 0$, define the class

$$\Phi_\eta = \{ \varphi:[0,\eta) \to \mathbb{R}_+ \mid \varphi(0)=0,\ \varphi \text{ strictly increasing, right-continuous at } 0,\ \varphi \text{ concave} \}.$$

Then $f$ is said to have the pointwise generalized concave KL property at $\bar{x}$ if there exist a neighborhood $U \ni \bar{x}$, $\eta > 0$, and $\varphi \in \Phi_\eta$ such that, for all $x \in U$ with $0 < f(x) - f(\bar{x}) < \eta$, it holds that

$$\varphi_-'\bigl(f(x) - f(\bar{x})\bigr) \cdot \mathrm{dist}(0, \partial f(x)) \ge 1,$$

where $\varphi_-'$ denotes the left derivative. A setwise version extends this from a singleton critical point to a compact level set $V \subset \operatorname{dom}\,\partial f$ on which $f \equiv \mu$, with uniformity over $V$ (Wang et al., 2020, Wang et al., 2021, Wang et al., 2021).

This generalization departs from earlier formulations by accepting nonsmooth, nondifferentiable $\varphi$, accommodating cases where the minimal admissible desingularizer is not differentiable.

2. The Exact Modulus: Minimal Concave Desingularizer

Given the generalized concave KL property at $\bar{x}$, one can define the exact modulus $\varphi_{\text{exact}}$ by

$$h(s) := \sup \left\{ \frac{1}{\mathrm{dist}(0, \partial f(x))} : x \in U,\, 0 < f(x) - f(\bar{x}) < \eta,\, f(x) - f(\bar{x}) \ge s \right\},$$

and set

$\varphi_{\text{exact}}(t) := \int_0^t h(s)\,ds.$

$\varphi_{\text{exact}} \in \Phi_\eta$ is strictly increasing, concave, and right-continuous at $0$ with $\varphi_{\text{exact}}(0)=0$. Moreover,

$$\varphi_{\text{exact}\,-}'(s) = h(s) \ge \sup_{x: f(x)-f(\bar{x}) = s} \frac{1}{\mathrm{dist}(0, \partial f(x))}.$$

$\varphi_{\text{exact}}$ is shown to be the infimum—pointwise and minimal—among all admissible concave desingularizers. That is, for any other $\psi \in \Phi_\eta$ satisfying the KL-inequality, $\varphi_{\text{exact}}(t) \le \psi(t)$ for all $t \in [0, \eta)$ (Wang et al., 2020).

3. Calculus Rules and Broader Scope

Unlike classical calculus for KL exponents, the generalized concave KL framework enables calculus rules for sum, minimum, separable sum, and composition operations without imposing differentiability or power-law forms on desingularizers. Key results include:

• Sum Rule: Suppose $f = \sum_{i=1}^m f_i$, each with desingularizer $\varphi_i$. Under a linear regularity condition, $f$ has as desingularizer

$$\varphi(t) = \frac{1}{\alpha} \int_0^t \max_{1 \le i \le m} (\varphi_i)_-'(s/m) ds,$$

where $\alpha > 0$ is the regularity constant.

• Minimum Rule: For $f(x) = \min_{1 \le i \le m} f_i(x)$, $f$ inherits the maximum of the one-sided derivatives among active indices.
• Separable Sum and Composition: Structured constructions for $\varphi$ arise from the properties and domains of building-block functions and their desingularizers.

These rules go beyond the scope of previous frameworks (e.g., Li–Pong’s KL exponent calculus), applying to composite and piecewise functions, even when no power-law desingularizer exists (Wang et al., 2021).

4. Representative Examples and Nondifferentiability

Classic and pathological examples expose the necessity and impact of nondifferentiable exact modulus:

• For the piecewise function

$$f(x) = \begin{cases} \frac{1}{2}x^2 & |x| \le \rho \ 2\rho|x| - \frac{3}{4}\rho^2 & |x|>\rho \end{cases}$$

around $x=0$, the exact modulus is

$$\varphi_{\text{exact}}(t) = \begin{cases} \sqrt{2t}, & 0 \le t \le \rho^2/2\ t/(2\rho) + (3/4)\rho, & t > \rho^2/2 \end{cases}$$

which is nondifferentiable at $t = \rho^2/2$ (Wang et al., 2020).

• For $h(x) = \min\{e^{-1/x^2}, |x|\}$, the exact modulus is $\varphi_h(t) = \sqrt{-1/\ln t}$ on $(0, e^{-3/2})$, which cannot be expressed as any $ct^{1-\theta}$ power law (Wang et al., 2021).

Such examples demonstrate that the exact modulus is often non-power-law, and may be strictly smaller than those produced by previous theories (e.g., BDLM integrals, classical exponent rules).

5. Consequences for Algorithmic Convergence and Complexity Bounds

The sharpness and minimality of the exact modulus yield optimal finite-length and convergence rate bounds for descent-type algorithms. For instance, under the standard PALM (Proximal Alternating Linearized Minimization) assumptions,

$$\sum_{k=1}^\infty \|z_{k+1} - z_k\| \leq A + C\cdot \varphi_{\text{exact}}(\Psi(z_{l+1}) - \Psi(z^*)),$$

with $A$ and $l$ determined by the early iterates, yielding the strongest possible bound of this type. No smaller (i.e., sharper) estimate of this algorithmic path length is possible without further assumptions on $f$ (Wang et al., 2020).

Similarly, in the Malitsky–Tam forward-reflected-backward (FRB) splitting for nonconvex composite minimization, the exact modulus desingularizer is central to deriving finite-length convergence and the best-possible upper bounds on cumulative step lengths. If the merit function $H$ has the setwise generalized concave KL property with exact modulus $\tilde\varphi$, one obtains

$\sum_{k} \|z_{k+1}-z_k\| < \infty,$

with explicit rates determined by $\tilde\varphi$ (Wang et al., 2021). In the semialgebraic case, the classical finite/linear/sublinear rate trichotomy is recovered as a special case; the generalized property allows non-power-law regimes.

6. Implications and Applications

The introduction and algebraic manipulation of the exact modulus relax classical rigidities—differentiability and power form—enabling convergence theory for broader function classes such as piecewise, composite, and nonsmooth (including non-semi-algebraic) objectives. This provides a more nuanced understanding of descent geometry, allowing for the discovery of new complexity bounds and sharper iteration estimates, particularly for composite optimization and structured nonconvex models.

Applications include analysis of structured feasibility problems, nonconvex quadratically constrained programs, and zero-norm regularization, with calculation or tight estimation of $h(s)$ directly from problem structure or through variational analysis (Wang et al., 2021). Numerical experiments in nonconvex sparse-feasibility further confirm that algorithms leveraging the exact modulus can outperform those using coarser desingularizing functions (Wang et al., 2021).

7. Relation to Classical Theory and Current Research Directions

In the classical KL framework, desingularizing functions are typically restricted to power forms, and differentiability is assumed, limiting calculus rules and applicability. The generalized property with its exact modulus overcomes these limitations, bridging to more general settings and finer complexity results (Wang et al., 2020, Wang et al., 2021).

Contemporary research focuses on exploiting this added flexibility for:

• Extension to stochastic or variational frameworks,
• Development of new complexity bounds for composite optimization,
• Exploration of convergence behaviors outside the classical rate trichotomy,
• Applications to models where subdifferential geometry is highly irregular.

A plausible implication is that future algorithmic convergence theory will increasingly rely on the computation and estimation of the exact modulus, approaching the theoretical limit of KL-based regularity (Wang et al., 2020, Wang et al., 2021, Wang et al., 2021).