Locally Polyak–Łojasiewicz Region

• LPLR is a local geometric property where a function satisfies the Polyak–Łojasiewicz inequality, enabling linear convergence of gradient-based methods.
• It describes a tubular neighborhood around a manifold of minimizers with quantifiable size and curvature, linking local geometry to convergence rates.
• LPLR underpins applications in deep learning, optimal transport, and stochastic dynamics by providing diagnostic tools for verifying rapid, linear convergence in nonconvex settings.

A Locally Polyak–Łojasiewicz Region (LPLR) is a subset of the parameter or state space in which a function (such as a loss, potential, or objective) satisfies the Polyak–Łojasiewicz (PL) inequality with uniform positive constant. This local geometric property guarantees that gradient-based optimization or stochastic dynamics exhibit linear convergence rates, even in nonconvex or degenerate settings. The LPLR concept is central to recent theoretical advances explaining efficient optimization in high-dimensional, nonconvex systems including deep neural networks, non-isolated minimizer structures, and regularized optimal transport (Aich et al., 29 Jul 2025, Gong et al., 8 Feb 2025, Feng et al., 23 Mar 2026, González-Sanz et al., 26 May 2026, Gong et al., 2024). The LPLR framework extends global PL theory by recognizing that local landscape geometry—not global convexity—is sufficient for fast convergence, and characterizes the size, structure, and analytical consequences of regions where the local PL inequality is active.

1. Formal Definition of the Locally Polyak–Łojasiewicz Region

A function $F$ defined on an open subset of $\mathbb{R}^d$ satisfies the PL inequality locally if, in a region $R \subset \mathbb{R}^d$, there exists $\mu>0$ such that

$$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$

where $F^*$ denotes the minimum value of $F$ on $R$ or an associated optimal set $S$.

• In deep learning, for losses $L(\theta)$ with parameters $\mathbb{R}^d$0, an LPLR is a region $\mathbb{R}^d$1 around initialization such that

$\mathbb{R}^d$2

• In quadratic optimal transport duals, the LPLR is the set of pairs $\mathbb{R}^d$3 in function space with $\mathbb{R}^d$4 for which a uniform PL constant $\mathbb{R}^d$5 exists.

The size of the region $\mathbb{R}^d$6 is determined by local curvature, the smoothness of the function ($\mathbb{R}^d$7 regularity), and, in certain applications, model-dependent quantities such as the smallest eigenvalue of the Neural Tangent Kernel or geometric properties of minimizer manifolds (Aich et al., 29 Jul 2025, Feng et al., 23 Mar 2026, González-Sanz et al., 26 May 2026, Gong et al., 2024).

2. Geometric and Manifold Structure of LPLR Minimizer Sets

In nonconvex optimization with non-isolated minimizers, the optimal set $\mathbb{R}^d$8 to which the PL region is tubularly centered is rigorously identified as a compact, $\mathbb{R}^d$9-embedded submanifold of $R \subset \mathbb{R}^d$0 without boundary. Under mild regularity and no-saddle conditions:

• $R \subset \mathbb{R}^d$1 is a connected component.
• The second fundamental form of $R \subset \mathbb{R}^d$2 is bounded, ensuring geometric control.
• Examples include spheres (Morse-Bott functions) and other noncontractible manifolds, as confirmed in deep models and low-temperature potentials (Gong et al., 8 Feb 2025, Gong et al., 2024, Feng et al., 23 Mar 2026).

The region $R \subset \mathbb{R}^d$3 is typically a “tubular” neighborhood around $R \subset \mathbb{R}^d$4:

$R \subset \mathbb{R}^d$5

where $R \subset \mathbb{R}^d$6, $R \subset \mathbb{R}^d$7 is a scale parameter (such as temperature), and $R \subset \mathbb{R}^d$8 the PL exponent ($R \subset \mathbb{R}^d$9) (Gong et al., 2024). This structure guarantees concentration properties of induced Gibbs or learning dynamics and facilitates spectral analysis (e.g., for Poincaré inequalities and mixing rates).

3. Analytical Consequences: Local Linear Convergence

Within an LPLR, first-order optimization methods and stochastic dynamics exhibit linear (exponential) convergence to a local minimizer or minimizer set:

• For $\mu>0$0 objective functions, Nesterov’s accelerated gradient descent achieves optimal local linear convergence, with the spectral rate

$\mu>0$1

where $\mu>0$2 and $\mu>0$3 are local Hessian eigenvalues at the minimizer (Feng et al., 23 Mar 2026).

• In deep networks, under local NTK stability, full-batch gradient descent and SGD converge at rate $\mu>0$4 as long as iterates remain in the LPLR (Aich et al., 29 Jul 2025).
• The overdamped Langevin SDE with a local PL potential $\mu>0$5 mixes in time $\mu>0$6 (logarithmic factors in $\mu>0$7 are suppressed), where $\mu>0$8 is the diffusion or “temperature” parameter (Gong et al., 2024, Gong et al., 8 Feb 2025).

For linearly convergent algorithms, the PL constant $\mu>0$9 translates geometric properties (local strong convexity, kernel conditioning, or spectral gap) into explicit rates and step size bounds.

4. Verification Criteria and Diagnostics in Applied Settings

The presence of an active LPLR during training or algorithmic execution can be confirmed by the following diagnostics (Aich et al., 29 Jul 2025):

• PL-ratio: $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$0 stabilizes at a positive constant as optimization progresses.
• NTK Conditioning: For DNNs, empirically monitoring the smallest eigenvalue $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$1 of the NTK over trajectories signals a positive local PL constant.
• Loss and Gradient Decay: Exponential (linear-linear scale) drop in training loss and gradient norm confirms operation within an LPLR.

Algorithmically, ensuring that iterations remain within the defined neighborhood $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$2 is essential for maintaining theoretical convergence guarantees (e.g., in regularized OT, initializations must satisfy $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$3 for coordinate ascent methods) (González-Sanz et al., 26 May 2026).

5. Poincaré Inequalities and Spectral Structure

In the low-temperature regime, for Gibbs measures

$$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$4

the local PL region ensures satisfaction of a Poincaré inequality with an equilibrium-independent lower bound:

$$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$5

where $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$6 is the first positive eigenvalue of the Laplace–Beltrami operator on the minimizer manifold $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$7 (Gong et al., 2024, Gong et al., 8 Feb 2025). This result holds even in non-log-concave, nonconvex landscapes with non-isolated minimizers, distinguishing the LPLR framework from global PL or (log-)convex settings.

This spectral gap underpins subexponential Langevin mixing and Langevin Monte Carlo iteration complexity, and forms a technical bridge to spectral theory via manifold geometry.

6. Applications and Empirical Observations

The LPLR framework explains observed linear convergence in diverse nonconvex optimization scenarios:

• Deep Learning: Empirically, robust LPLR structure occurs in both MLP and ResNet architectures across various widths and initialization schemes, with the local PL constant $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$8 sensitive to network architecture and initialization (Aich et al., 29 Jul 2025).
• Quadratic Optimal Transport: Local PL geometry arises in quadratically regularized duals, leading to explicit linear rates for standard first-order algorithms in function space, provided local regions defined by the $$\|\nabla F(x)\|^2 \geq 2\mu \big( F(x)-F^* \big), \qquad \forall x \in R,$$9-norm stay bounded (González-Sanz et al., 26 May 2026).
• Stochastic Dynamics: Analysis of nonconvex, overparameterized landscapes with non-isolated minima and bounded curvature shows that Gibbs measures concentrate on compact manifolds, and the stochastic processes converge quickly even when the global function is highly nonconvex (Gong et al., 2024, Gong et al., 8 Feb 2025).

A plausible implication is that rapid convergence of gradient-based methods in high-dimensional, real-world nonconvex objectives is largely explained by the emergence and prevalence of LPLRs under standard initialization and overparameterization.

7. Theoretical Extensions and Limitations

The LPLR paradigm generalizes PL-based analyses beyond globally strongly convex or nondegenerate settings:

• It accounts for manifolds of minimizers, degenerate Hessians, and non-contractible optimal sets.
• It provides a uniform framework for the spectral gap analysis of nonconvex Gibbs measures and for the Lyapunov stability of optimization methods (Gong et al., 8 Feb 2025, Feng et al., 23 Mar 2026).
• It rigorously justifies local-to-global reductions via subdomain truncation, Holley–Stroock perturbation, and spectral stability arguments.

However, the guarantees are local: iterates must remain in the region where the PL inequality, its associated curvature conditions, and, when relevant, the NTK stability or minimum eigenvalue bounds are valid. Global convergence, nonlocal saddle avoidance, and flow between disconnected components fall outside the LPLR scope.

---

References:

• "From Sublinear to Linear: Fast Convergence in Deep Networks via Locally Polyak-Lojasiewicz Regions" (Aich et al., 29 Jul 2025)
• "Poincaré Inequality for Local Log-Polyak-Lojasiewicz Measures : Non-asymptotic Analysis in Low-temperature Regime" (Gong et al., 8 Feb 2025)
• "Optimal local linear convergence of Nesterov's accelerated gradient method for $F^*$0 functions under the Polyak--Łojasiewicz inequality" (Feng et al., 23 Mar 2026)
• "Polyak-Lojasiewicz Inequality for Quadratically Regularized Optimal Transport" (González-Sanz et al., 26 May 2026)
• "Poincare Inequality for Local Log-Polyak-Łojasiewicz Measures: Non-asymptotic Analysis in Low-temperature Regime" (Gong et al., 2024)