Desingularizing Functions in Theory & Applications

• Desingularizing functions are concave, strictly increasing functions that re-parameterize errors to control singular behavior near critical points.
• They play a crucial role in dynamical systems, PDEs, and algebraic geometry by enabling precise convergence estimates and effective resolution of singularities.
• Their calculus, including power-law moduli and optimal constructions, underpins robust algorithmic guarantees and structural transformations in diverse mathematical settings.

A desingularizing function arises in multiple branches of mathematics, notably in the analysis of dynamical systems, PDE theory, algebraic geometry, and nonsmooth analysis, whenever it is necessary to control or compensate for singular behavior either of operators, functionals, or algebraic structures. In the analytic context, desingularizing functions are central to the Kurdyka–Łojasiewicz (KL) theory and its generalizations, where they facilitate quantitative estimates on convergence rates near critical points of non-smooth or nonconvex objective functions. In PDEs and stochastic processes, desingularizing weights play a fundamental role in compensating for degeneracy or singular drift terms, enabling sharp kernel bounds and functional inequalities. In desingularization of algebraic or topological objects, analogous concepts underlie the explicit transformation of singular spaces into regular ones.

1. Desingularizing Functions in the Kurdyka–Łojasiewicz Framework

The classical Kurdyka–Łojasiewicz (KL) property for a proper lower semicontinuous function $f:\mathbb{R}^n\to(-\infty,+\infty]$ at a critical point $\bar{x}$ asserts the existence of a continuous, concave, strictly increasing function $\varphi:[0,\eta)\to\mathbb{R}_+$, with $\varphi(0)=0$ and $\varphi'\!>\!0$ on $(0,\eta)$, for which the KL inequality holds: $$\varphi'\big(f(x)-f(\bar{x})\big) \cdot \operatorname{dist}(0, \partial f(x)) \geq 1, \qquad x \in U\cap\{0 < f(x)-f(\bar{x}) < \eta\}.$$ This desingularizing function $\varphi$ “re-parametrizes” the error so that the gradient (or subdifferential) does not degenerate excessively as the optimum is approached. The canonical example is the power-law class $\varphi(s) = cs^{1-\theta}$, $0\le\theta<1$, emphasizing the function’s flatness at criticality and determining the convergence speed of descent methods. This form directly controls whether the corresponding algorithm achieves finite, linear, or sublinear convergence (Johnstone et al., 2016, Jendoubi et al., 2014).

The generalized Kurdyka–Łojasiewicz property, as formalized in (Wang et al., 2020, Wang et al., 2021), allows for nonsmooth (even nondifferentiable) concave functions $\bar{x}$0 equipped with well-defined left derivatives. The “exact modulus” or “optimal desingularizer” $\bar{x}$1 is constructed by

$\bar{x}$2

This $\bar{x}$3 is minimal among all concave desingularizing functions for $\bar{x}$4 at $\bar{x}$5 and yields the sharpest possible rate bounds for iterative optimization sequences (Wang et al., 2020).

2. Quantitative Role in Dynamical Systems and PDEs

In the study of damped second order gradient systems $\bar{x}$6, desingularizing functions $\bar{x}$7 connect the geometry of the potential $\bar{x}$8 to explicit convergence rates of trajectories. When $\bar{x}$9 is $\varphi:[0,\eta)\to\mathbb{R}_+$0 and definable, any desingularizing function must satisfy $\varphi:[0,\eta)\to\mathbb{R}_+$1 for small $\varphi:[0,\eta)\to\mathbb{R}_+$2. This lower bound propagates to similar algebraic rates for the decay of the total energy and distance to equilibrium (Jendoubi et al., 2014).

In the context of non-local PDEs, specifically the fractional Kolmogorov operator with Hardy-type drifts, desingularizing weights are applied to neutralize the singularity of the drift $\varphi:[0,\eta)\to\mathbb{R}_+$3. The key construction is a weight $\varphi:[0,\eta)\to\mathbb{R}_+$4 near the singularity, with $\varphi:[0,\eta)\to\mathbb{R}_+$5 chosen via a transcendental equation balancing the singularity in weighted $\varphi:[0,\eta)\to\mathbb{R}_+$6-spaces, leading to a cancellation of the critical drift in the generator's adjoint and enabling sharp two-sided heat-kernel estimates. This method generalizes the ground-state transform for local operators to non-local, nonsymmetric cases (Kinzebulatov et al., 2020).

3. Structural and Algorithmic Desingularization

In algebraic geometry and singularity theory, desingularization refers to explicit procedures transforming singular algebraic objects into regular ones. The role of desingularizing functions manifests in algorithmic invariants (such as the Hilbert–Samuel function or the Bierstone–Milman invariant) and local bases (e.g., standard bases along Samuel strata), underpinning the construction of canonical Rees algebras and functorial resolution algorithms (Wlodarczyk, 2015, Bierstone et al., 2013). In this context, the desingularization process is guided by local invariants that reflect singular complexity, with stepwise reduction via blow-ups and invariants decreasing at each stage until a nonsingular or stable-snc structure emerges.

In simplicial topology, the desingularization functor $\varphi:[0,\eta)\to\mathbb{R}_+$7 systematically collapses non-embedded simplices through iterated “enforced collapses” until a non-singular simplicial set is produced. This is achieved by a transfinite sequence of quotienting steps indexed by ordinals, with effective termination due to the finiteness of non-degenerate simplices (Fjellbo, 2020).

4. Calculus and Composition of Desingularizing Functions

The calculus of desingularizing functions, particularly in nonsmooth optimization, encompasses systematic rules for determining the desingularizable structure of composite or aggregate functions. Notably, for sums, minima, separable sums, and compositions with smooth maps, explicit constructions for the smallest possible (optimal) concave desingularizer are established. For instance, the sum rule produces a desingularizer for $\varphi:[0,\eta)\to\mathbb{R}_+$8 by

$\varphi:[0,\eta)\to\mathbb{R}_+$9

with $\varphi(0)=0$0 the desingularizers for $\varphi(0)=0$1 and $\varphi(0)=0$2 a regularity constant, recovering and extending the exponent rules of Li–Pong for the power-law case (Wang et al., 2021). For the minimum of finitely many functions, the maximum of the respective left-derivatives under the integral yields the optimal modulus.

This calculus generalizes the exponent-only framework of semialgebraic functions, allowing the treatment of fully general nondifferentiable concave desingularizers. The approach is essential in establishing optimal rates for proximal-type algorithms and in analyzing complex functionals arising in signal processing and machine learning (Johnstone et al., 2016, Wang et al., 2021).

5. Applications and Case Studies

The desingularizing function formalism underpins several concrete results and algorithmic guarantees:

• In inertial and proximal splitting algorithms for nonconvex composite optimization, the KL exponent $\varphi(0)=0$3 in the power-law desingularizer $\varphi(0)=0$4 determines the rate of convergence: super-fast (finite) for $\varphi(0)=0$5, linear for $\varphi(0)=0$6, and sublinear $\varphi(0)=0$7 for $\varphi(0)=0$8 (Johnstone et al., 2016).
• For the PALM algorithm, the optimal modulus $\varphi(0)=0$9 controls the exact bound on the total length of the optimization path, with minimal $\varphi'\!>\!0$0 providing the sharpest guarantee (Wang et al., 2020).
• In the desingularization of algebraic varieties, the use of standard bases and Rees algebras along Samuel strata yields canonical stratifications and explicit local normal forms, crucial for functorial and characteristic-0 resolution algorithms (Wlodarczyk, 2015).
• The weighted Nash inequality and kernel estimates for drift-PDEs with critical singularities rely decisively on the choice and properties of desingularizing weights, as detailed in the analysis of the fractional Kolmogorov operator (Kinzebulatov et al., 2020).

6. Broader Implications and Open Problems

Desingularizing function techniques traverse analytic, algebraic, and topological frameworks, providing a unifying lens for singularity compensation. For KL-type inequalities, the search for minimal desingularizing functions (optimal moduli) has resolved several open problems in obtaining tight convergence and path-length estimates (Wang et al., 2020). However, extending these results to more general classes of functions, such as those with genuinely anisotropic singularities, multi-point singularities, and variable-coefficient or time-dependent operators, remains a significant open research direction (Kinzebulatov et al., 2020).

In singularity theory and resolution of varieties, open challenges include the extension of functorial and stratification-based desingularization algorithms to positive characteristic and the construction of canonical desingularizers in equivariant or stack-theoretic contexts (Wlodarczyk, 2015, Bierstone et al., 2013).

Recent advances in the calculus for composite and nonsmooth desingularizers strongly suggest a pathway toward a comprehensive theory encompassing broader classes of variational problems and dynamical systems (Wang et al., 2021). The explicit construction of optimal moduli and their application to algorithm design and analysis in high-dimensional and nonconvex landscapes is an active and rapidly developing area.