Kurdyka–Łojasiewicz Inequality Overview

Updated 9 February 2026

Kurdyka–Łojasiewicz inequality is a key analytic property that characterizes the quantified regularity of nonsmooth and nonconvex functions near critical points.
It establishes a precise relationship between the function's value and the size of its subdifferential, which in turn dictates convergence rates in optimization algorithms.
The framework extends to settings such as semi-algebraic, subanalytic, and infinite-dimensional variational problems, offering concrete error bounds and complexity estimates.

The Kurdyka–Łojasiewicz (KŁ) inequality is a fundamental analytic property describing a local geometric regularity at critical points of nonsmooth or nonconvex functions, extending the classical Łojasiewicz gradient inequality from real-analytic and subanalytic functions to wide classes including nonsmooth, semi-algebraic, and variational-analytic functions. At its core, the KŁ property asserts that the objective function exhibits a quantified relationship between proximity to criticality and subdifferential size, controlled by a desingularizing function. This relationship has powerful consequences for convergence rates and complexity bounds of optimization algorithms, error bounds in variational settings, and structural regularity conditions in real algebraic geometry and optimization theory.

1. Formal Definition and General Structure

Let $f:\mathbb{R}^n\to\mathbb{R}\cup\{+\infty\}$ be a proper, lower semicontinuous function and $\bar{x}$ a critical point (typically, $0\in\partial f(\bar{x})$ where $\partial f$ denotes a limiting or Mordukhovich subdifferential). The KŁ property at $\bar{x}$ consists of the existence of a neighborhood $U$ of $\bar{x}$ , a concave, strictly increasing desingularizing function $\varphi:[0,\eta)\to\mathbb{R}_+$ with $\varphi(0)=0$ , $\varphi'>0$ , and a constant $c>0$ , such that for all $x\in U$ , $0<f(x)-f(\bar{x})<\eta$ ,

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

For smooth $f$ , this reduces to

$\varphi'\bigl(f(x)-f(\bar{x})\bigr)\cdot\|\nabla f(x)\|\ge1.$

A canonical and highly relevant class of desingularizers is

$\varphi(s) = c\,s^{1-\theta},\qquad \theta\in[0,1),$

leading to the explicit form

$\|\nabla f(x)\|\ge\frac{1}{c(1-\theta)}(f(x)-f(\bar{x}))^\theta.$

Here, the exponent $\theta$ (alternatively $q$ in some literature with $\theta = q$ ) quantifies the “sharpness” of the regularity at $\bar{x}$ . The case $\theta=0$ corresponds to analytic or strongly convex-type regularity (finite-termination behavior), $\theta\in(0,\frac12]$ to linear convergence regimes, and $\theta\in(\frac12,1)$ to sublinear or polynomial convergence rates.

The property extends naturally to nonsmooth, nonconvex, and infinite-dimensional variational settings, with appropriate variational subdifferentials and possibly more general desingularizers, including nonsmooth or even discontinuous ones (Li et al., 2023, Wang et al., 2020, Wang et al., 2021).

2. Analytical and Algebraic Foundations

2.1. Semi- and Subanalytic Contexts

The KŁ property holds generically for real-analytic [classic Lojasiewicz], subanalytic, and definable (in o-minimal structures) functions on $\mathbb{R}^n$ (Gerth et al., 2019, Baldi et al., 2022, Dinh et al., 2015). For semi-algebraic sets

$S = \left\{ x\in\mathbb{R}^n : g_1(x)\geq0,\,\ldots,\,g_r(x)\geq0 \right\},$

and continuous functions $f$ semi-algebraic on $S$ , the existence of $(c,L)$ in inequalities of the form

$F(x)^L\leq c\,G(x)$

(“Łojasiewicz-type inequalities”) is supported by Bochnak–Coste–Roy’s theorem (Baldi et al., 2022). In particular, this yields

$\|\nabla f(x)\| \geq C |f(x)-f^*|^{\theta}$

for $x$ near a minimizer under suitable regularity hypotheses.

2.2. Explicit Exponents in Algebraic and Nash Geometries

For polynomial and Nash functions, the KŁ/Łojasiewicz exponent can be explicitly bounded in terms of the degree and the dimension, yielding effective inequalities even in non-isolated or singular configurations. For instance, for a Nash function $f$ on a compact semialgebraic set, the exponent $\varrho = 1-\frac{1}{S(n,d)}$ with $S(n,d)=2(2d-1)^{3n+1}$ , and even sharper when $f$ is algebraic with a degree- $d_1$ graph, $\varrho = 1-\frac{1}{R(n,d_1)}$ with $R(n,d_1)$ as in (Osińska-Ulrych et al., 2018): $|\nabla f(x)| \geq C |f(x)|^\varrho.$ This delivers fully explicit metrics for gradient domination tied to algebraic parameters (Dinh et al., 2015, Osińska-Ulrych et al., 2018).

3. Exponent, Modulus, and Desingularizing Function: Sharpness and Structure

Considerable attention is given in recent variational analysis to the optimality of the desingularizing function—the so-called “exact modulus” $\psi^*$ —defined as (for $U$ near $\bar{x}$ and $s\in(0,\eta)$ )

$h(s) = \sup\left\{ \frac{1}{\mathrm{dist}(0,\partial f(x))} : x\in U,\, s \leq f(x) - f(\bar{x}) < \eta \right\},\quad \psi^*(t) = \int_0^t h(s)\,ds.$

$\psi^*$ is always concave, and is the minimal possible modulus among all concave desingularizers: $\psi^* = \inf\{ \varphi : \varphi \text{ concave and satisfies the KŁ property on } U \}$

(Wang et al., 2020). This exact modulus provides the sharpest possible bound for the convergence/length of proximal- or first-order schemes based on the KŁ property.

For broad function classes (e.g., prox-regular, twice epi-differentiable, subdifferentially continuous functions) appearing in structured and nonsmooth optimization, the KŁ property with exponent $\theta=1/2$ —and an explicit, often computable modulus—is universally assured via second-order variational data (e.g., Moreau envelope, quadratic growth) (Li et al., 2023).

4. Calculus Rules and Structural Stability

The KŁ property is stable under a variety of function operations, subject to suitable regularity and interplay conditions. The calculus includes:

Sum rule: For $f = \sum_{i=1}^m f_i$ , the modulus is controlled by the maxima of the individual $f_i$ moduli, under a linear-regularity condition for the subdifferentials (Wang et al., 2021).
Minimum rule: For $f = \min_i f_i$ , the modulus at a minimizer is given in terms of the active indices’ moduli.
Separable sum rule: For block-separable $f = \sum_{i=1}^m f_i(x_i)$ , the modulus is again governed by the maximal individual modulus.
Composition rule: For $f = g\circ F$ , with $g$ KŁ and $F$ $C^1$ of full rank, the KŁ property at $x$ is inherited with a multiplicative scaling of the desingularizer (Wang et al., 2021).

These rules facilitate the propagation of the KŁ property through structured objective constructions, essential for composite and block-coordinate optimization.

5. Error Bounds, Polyak–Łojasiewicz, and Algorithmic Implications

The KŁ inequality entails explicit error bounds: for a function $f$ with KŁ exponent $\theta$ , the distance to the set of minimizers $S$ can be quantified as

$\mathrm{dist}(x, S) \leq C [f(x) - f^*]^{1-\theta}$

(e.g., for trust-region subproblems, the optimal bound is $p=1/2$ or $1/4$ in the so-called ill case, with KŁ exponent $\theta=1-p$ (Jiang et al., 2019)). This direct relationship underpins local and global metric regularity and is instrumental in establishing convergence and complexity of first-order algorithms.

The Polyak–Łojasiewicz (PL) and generalized $\alpha$ -PL (gradient domination) inequalities are special cases of the KŁ property, particularly prevalent in optimization. Under $\|\nabla f(x)\|\geq \mu(f(x)-f^*)^\theta$ with $\theta=1/2$ (PL), one obtains linear convergence rates. For $\theta\in(1/2,1)$ , only sublinear rates are possible (Bento et al., 2024, Fatkhullin et al., 2022, Ahookhosh et al., 13 Nov 2025). These rate regimes persist across stochastic, blockwise, and inexact frameworks, including SGD, variance-reduced methods (e.g., PAGER), and block-coordinate/reshuffling schemes (Fatkhullin et al., 2022, Li et al., 2021).

6. Infinite-Dimensional and Variational Setting

The KŁ property generalizes naturally to infinite-dimensional Hilbert and Banach spaces, replacing the norm-gradient by the minimal-norm subgradient or "slope" $|\partial\varphi|(u)$ . The KŁ–Simon inequality in this context states that for a proper, semiconvex, lower semicontinuous functional $\varphi:H\to(-\infty,+\infty]$ and a critical point $u_*$ ,

$\theta'\bigl( \varphi(u)-\varphi(u_*) \bigr) |\partial\varphi|(u) \geq 1,$

for some $C^1$ strictly increasing $\theta$ (Chill et al., 2016). This leads to stabilization results and convergence for abstract gradient flows and PDEs, provided only relative compactness in the relevant topology—a significant relaxation over previous prerequisites.

7. Applications, Examples, and Algorithmic Complexity

The presence of the KŁ property provides a rigorous analytic bridge from variational geometry to algorithmic performance:

Sums of squares and Positivstellensatz: The effective Putinar’s Positivstellensatz leverages KŁ–Łojasiewicz estimates on semi-algebraic distance functions to give degree bounds that are polynomial in the KŁ constant and in the inverse of function slack (Baldi et al., 2022).
Trust-region subproblems: The explicit KŁ exponent allows deriving tight convergence rates for projected gradient on nonconvex quadratics, distinguishing between genuinely hard and generic cases (Jiang et al., 2019).
Structured regularization: Composite functionals (e.g., $\ell_1$ -regularized objectives) and pointwise maxima over smooth components admit sharp exponents ( $\theta=1/2$ ) via subdifferential calculus and Moreau envelopes (Li et al., 2023).
Nonmonotone variational inequalities: KL-exponent $1/2$ and associated error bounds for D-gap functions guarantee global linear convergence of non-smooth, derivative-free descent schemes under mild regularity (Li et al., 2022).
Empirical and stochastic optimization: The explicit exponent determines the sample and iteration complexity across a wide array of first-order, stochastic, or splitting algorithms (Fatkhullin et al., 2022, Li et al., 2021).

The KŁ property acts as a unifying analytical regularity criterion, subsuming classical source and variational conditions for convergence rates in inverse problems and regularization methods (Gerth et al., 2019).

Key References:

Paper	Key Contribution
(Li et al., 2023)	Variational characterizations of KŁ exponent and modulus
(Wang et al., 2020, Wang et al., 2021)	Exact modulus, generalized concave KŁ property, calculus rules
(Baldi et al., 2022, Osińska-Ulrych et al., 2018, Dinh et al., 2015)	Effective exponents for algebraic, Nash, and matrix eigenvalue
(Fatkhullin et al., 2022, Ahookhosh et al., 13 Nov 2025, Bento et al., 2024)	Algorithmic complexity and stochastic optimization under KŁ
(Jiang et al., 2019, Li et al., 2022)	Error bounds, variational inequalities, and algorithm convergence
(Chill et al., 2016, Gerth et al., 2019)	Infinite-dimensional, variational, and functional-analytic forms

The ongoing refinement of the KŁ framework—including sharp modulus characterization, stability under function calculus, and tight algorithmic complexity—establishes it as a central theoretical pillar in contemporary analysis and optimization.