Doubling-of-Variables Method in PDE Analysis

• The doubling-of-variables method is a technique used to derive comparison principles and prove uniqueness in nonlinear and linear PDEs by aligning candidate solutions via penalization.
• It utilizes specific penalization functions, such as quadratic or geometric constructs, to force convergence of subsolutions and supersolutions, ensuring rigorous analytic bounds.
• The method extends to stochastic analysis and probability, providing quantitative estimates for central limit theorems and proving uniqueness even in degenerate Kolmogorov equations.

The doubling-of-variables method is a powerful technique for deriving comparison principles, proving uniqueness of solutions, and obtaining quantitative estimates in a range of nonlinear and linear partial differential equations (PDEs), stochastic equations, and probability limit theorems. Its core idea is to compare two candidate solutions—typically a subsolution and a supersolution—by introducing a penalization that forces their arguments to coincide, then exploiting optimality or stationarity conditions to extract rigorous analytic bounds or identities.

1. Foundational Principles and Variants

The classical doubling-of-variables technique originated in the theory of viscosity solutions for nonlinear PDEs, notably Hamilton–Jacobi equations of the form

$u(x) + H(x, d_x u) = F(x),\quad x \in M,$

where $(M, g)$ is a complete boundaryless Riemannian manifold, $H : T^* M \to \mathbb{R}$ is typically continuous (or $C^1$ in the convex case), and $F: M \to \mathbb{R}$ is given. The central objective is to prove a comparison principle: if $u$ (upper-semicontinuous) is a subsolution and $v$ (lower-semicontinuous) a supersolution, then $\sup_M (u - v) \leq \sup_M (F - F)$. This ensures uniqueness and stability of solutions, a cornerstone in the well-posedness of PDEs (Bertucci et al., 3 Dec 2025).

A critical feature is the choice of penalization function $\theta(x, y)$ used to enforce $x \approx y$ in the sup–inf arguments or in weak formulations over couplings. The architecture of $\theta$—quadratic, geometric, or otherwise—determines to what extent analytical or geometric regularity is required of the underlying data.

In stochastic analysis and kinetic equations, the method is also foundational for uniqueness proofs of invariant measures and for deriving contraction or continuity estimates, notably in the context of degenerate Kolmogorov (Fokker–Planck) equations (Bogachev et al., 11 Nov 2025). In probability theory, it is recently shown to yield new proofs and bounds for the central limit theorem by quantifying the gap between discrete and continuous semigroups (Addario-Berry et al., 2022).

2. The Classical Quadratic Construction

The original methodology, as formulated by Crandall and Ishii, employs a symmetric nonnegative penalization $\theta(x, y)$—most commonly $\theta(x, y) = |x - y|^2/(2\varepsilon)$. Defining

$\Phi(x, y) = u(x) - v(y) - \theta(x, y),$

one locates maximizers $(x_\varepsilon, y_\varepsilon)$ and uses viscosity sub- and supersolution properties, combined with the structure of the penalization, to extract test differentials $p_1 \in \partial^+ u(x_\varepsilon)$, $p_2 \in \partial^- v(y_\varepsilon)$. The penalization ensures that for small $\varepsilon$, the difference $|x_\varepsilon - y_\varepsilon|$ and the associated penalized momentum both become negligible.

The comparison then hinges on pointwise inequalities:

$$u(x_\varepsilon)+H(x_\varepsilon,p_1) \leq F(x_\varepsilon),\quad v(y_\varepsilon)+H(y_\varepsilon,p_2)\geq F(y_\varepsilon),$$

and on controlling $H(y_\varepsilon,p_2) - H(x_\varepsilon,p_1)$ via the Lipschitz/Hölder continuity of $H$ and the scaling imposed by the choice of $\theta$. In the limit $\varepsilon \to 0$, one recovers the comparison principle (Bertucci et al., 3 Dec 2025).

In degenerate stationary Kolmogorov equations, a similar penalization is used in weak formulations: choosing a penalty $\Phi(x, y)$ (e.g., $|x-y|^2/2$ or a logarithmic variant), one integrates against a coupling $\pi$ of two candidate invariant measures. The key quantity is then the sign or growth of a function $q(x, y)$ derived from the generator, allowing one to demonstrate that the coupling is supported on the diagonal $x = y$ and thus show uniqueness (Bogachev et al., 11 Nov 2025).

3. Geometric and Action-Based Penalization

A major advance is the geometric adaptation of the penalization, replacing purely Euclidean constructions with ones that encode the intrinsic metric or control structure of the problem. For Hamilton–Jacobi equations arising from convex optimal control, the penalization is chosen as the minimal action:

$$D(\varepsilon; x, y) := \inf_{\gamma(0)=x,\, \gamma(\varepsilon)=y} \int_0^\varepsilon L(\gamma(t), \dot{\gamma}(t))\,dt,$$

where $L$ is the Lagrangian related to $H$ via Fenchel duality. This construction shifts regularity demands from the solutions $u, v$ and the Hamiltonian $H$ onto the geometric properties (e.g., monotonicity, superdifferentiability) of $D(\varepsilon; x, y)$. Conservation of Hamiltonian energy along action-minimizing curves ensures exact cancellation of Hamiltonian terms, reducing analysis to the source term $F$ and yielding the desired comparison inequalities (Bertucci et al., 3 Dec 2025).

The superdifferentiability of $D$ with respect to $x$ or $y$ is crucial: for a minimizer $\gamma_{x \to y}$,

$$- \frac{\partial L}{\partial v}(x, \dot{\gamma}(0)) \in \partial^+_x D(\varepsilon;\cdot, y)(x).$$

This geometric penalization principle generalizes to infinite-dimensional spaces, such as the $p$-Wasserstein space $\mathcal{P}_p(M)$, via optimal transport costs and action functionals, preserving the core comparison mechanism while adapting to the geometry of measures (Bertucci et al., 3 Dec 2025).

4. Applications Across Deterministic and Stochastic PDEs

The doubling-of-variables method is central in several research areas:

• Nonlinear first-order Hamilton–Jacobi equations: Enables robust comparison principles on manifolds, with classical or geometric penalizations depending on convexity and the regularity of $H$ and $F$ (Bertucci et al., 3 Dec 2025).
• Second-order Kolmogorov equations: Proves uniqueness of stationary measures even for degenerate diffusion matrices, by coupling two solutions via an extended Fokker–Planck operator and penalizing off-diagonal points (Bogachev et al., 11 Nov 2025).
• Central Limit Theorem: Provides a new, real-analytic proof for the multidimensional CLT, comparing a discrete semigroup to the solution of the continuous heat equation using a “doubled” test function in time (and, in principle, space), and deriving quantitative Berry–Esseen-type bounds as a consequence (Addario-Berry et al., 2022).
• Optimal control and mean-field systems: Extends to control problems in mean-field (probability measure) spaces, constructing penalizations directly from the underlying Lagrangian or transport cost (Bertucci et al., 3 Dec 2025).

The underlying structure of the method is summarized in three general steps (Bogachev et al., 11 Nov 2025):

Step	Action	Outcome
(i)	Couple two candidate solutions via a product-space or measure coupling	Reduces to a joint stationary problem
(ii)	Test against a penalty that vanishes on the diagonal	Forces convergence to coinciding points
(iii)	Use sign, regularity, or growth to deduce support on the diagonal	Concludes equality or uniqueness

5. Handling Degeneracy and Regularity Transfer

A salient advantage of the method is its flexibility with respect to degeneracy and regularity. In the Kolmogorov setting, the only structural requirement is that $A(x)$ is symmetric and nonnegative-definite ($A = \Sigma \Sigma^T$ for some $\Sigma$), with no need for uniform ellipticity or a Hörmander condition. The method relies on estimating quantities such as

$$q(x, y) = \langle x - y, b(x) - b(y) \rangle + \text{tr}[(\Sigma(x) - \Sigma(y))(\Sigma(x) - \Sigma(y))^T],$$

and constructing a penalty $\Phi(x, y)$ for which sign conditions on $q(x, y)$ can be effectively used. Under strict or controlled sign assumptions, uniqueness follows even in highly degenerate scenarios (Bogachev et al., 11 Nov 2025).

In geometric penalization, regularity transitions from analytic conditions on $u, v$ and $H$ to geometric properties of the minimal action $D$, such as monotonicity and $\Gamma$-convergence. This transfer is advantageous in optimal control and non-Euclidean settings (Bertucci et al., 3 Dec 2025).

6. Quantitative Results and Probability Limit Theorems

Employing the doubling-of-variables approach in probability leads to new quantitative results. By comparing the discrete evolution $u_n(x, t)$ of rescaled random walks to the heat equation solution $u(x, t)$, and analyzing the maximizer of a doubled test function, one derives explicit Berry–Esseen-type rates:

$$\left| \mathbb{E}[f(S_n/\sqrt n)] - \mathbb{E}[f(\xi)] \right| \leq C\, \mathbb{E}[|X|^{2 + \gamma}]\, n^{-\gamma/2},$$

for smooth test functions $f$, with $\gamma \in (0, 1]$ and $S_n$ the sum of i.i.d. steps with covariance $\Sigma$. For $\gamma = 1$, this recovers the classical $O(n^{-1/2})$ Berry–Esseen bound in $\mathbb{R}^d$ (Addario-Berry et al., 2022).

This analytic approach requires no reliance on characteristic functions or Fourier methods and demonstrates the adaptability of the doubling-of-variables principle to quantitative probabilistic questions.

7. Extensions, Limitations, and Research Directions

The generality of the doubling-of-variables technique allows adaptation to:

• Full optimal transport and mean-field control settings, via lifting penalizations to measure or phase spaces (Bertucci et al., 3 Dec 2025).
• Degenerate and infinite-dimensional PDEs, where explicit analytic test functions may be unavailable or unwieldy, but geometric or measure-theoretic penalizations can be constructed (Bertucci et al., 3 Dec 2025, Bogachev et al., 11 Nov 2025).
• Nonlinear parabolic equations, via monotone finite-difference schemes and stability analysis (Addario-Berry et al., 2022).

A plausible implication is that further research will refine the geometric penalization paradigm—transferring more regularity from the solutions to structural properties of the penalization function—enabling comparison principles in even broader classes of degenerate or nonclassical PDEs.

Current limitations include sensitivity to the form of the penalization and the necessity of identifying conditions under which the diagonalization (forcing $x = y$ or their measure-theoretic analog) is enforced. The method's effectiveness depends on explicit or implicit control of cross-variation—such as $q(x, y)$—and the precise behavior of the penalization function near the diagonal.

---

Relevant research:

• (Bertucci et al., 3 Dec 2025) On the doubling of variables technique in first order Hamilton–Jacobi equations
• (Bogachev et al., 11 Nov 2025) Doubling variables and uniqueness of probability solutions to degenerate stationary Kolmogorov equations
• (Addario-Berry et al., 2022) The central limit theorem via doubling of variables