Can Gamma-convergence be made quantitative through variational extrapolation?

Gamma-convergence guarantees stability but not approximation rates

De Giorgi's Gamma-convergence theory is the standard tool for showing that a family of energy functionals converges to a limit functional, and that minimizers of the approximating problems converge to minimizers of the limit. What it does not intrinsically supply is a rate: it tells you that a sequence of minimizers approaches the limit minimizer, but not how fast, and not which regularity of the data controls that speed.

The paper frames this as the central limitation it addresses. Classical interpolation theory, built around the K-functional between two fixed normed spaces, is not directly suited to sequences of functionals in which the penalization itself evolves with the approximation parameter. The authors' response is to merge De Giorgi's variational framework with Gagliardo's theory of relative completions, so that the evolving penalization becomes part of the machinery rather than an obstacle.

The variational K-functional replaces two fixed norms with a family of functionals

The central new device is the variational K-functional. For two functionals F and G mapping a normed space X to [0, +infinity], it is defined for each f and t > 0 as the infimum of G(g) + t F(h) over decompositions f = g + h. When F and G are norms, this coincides with the classical K-functional; when they are general functionals, it becomes a nonlinear extension that can track a sequence of penalizations rather than a single fixed pair.

This construction is what makes the framework dynamic. Instead of measuring how far an element sits from a fixed space, the variational K-functional measures how the penalization behaves as the approximation parameter changes, which is exactly the regime in which Gamma-convergence operates. The paper also extends Gamma-convergence itself: the generalized upper and lower Gamma-limits are defined relative to a second sequence of functionals G_n that controls the sense in which test sequences approach f, and the classical theory is recovered when G_n is just the norm on X.

Gagliardo completions appear as the endpoint limit of interpolation norms

The bridge to Gagliardo's work comes from a limiting behavior of real interpolation norms. For a compatible pair of normed spaces (A0, A1), the normalized interpolation norm of an element f converges, as the interpolation parameter theta approaches an endpoint, to the norm of f in the corresponding Gagliardo relative completion. The paper states this as Proposition 2.1 and proves it by combining earlier endpoint results with the fact that the Gagliardo pair preserves the underlying K-functional structure.

This is the technical hinge of the whole framework. It means that the limit of a family of interpolation-type functionals as the parameter approaches 1 is not an abstract object but a Gagliardo completion, and that completion can be identified in concrete cases. The authors use exactly this identification in image processing, where the Gagliardo completion of (L1, W-dot-1,1) is the space BV of functions of bounded variation, which is the natural home of the total variation.

ROF convergence rates are controlled by the decay of a variational K-functional

The headline application is the inverse regularity problem for discrete approximations of the Rudin-Osher-Fatemi model. For a noisy image f in L2 on the torus, the ROF functional is one half of the squared L2 distance to f plus the total variation of u. The paper considers E_n, the approximation of this functional on piecewise constant functions over an n-by-n grid, with minimizers u-bar and u-bar_n and minimal values E-bar and E-bar_n.

Theorem 1.1 states that for p in (1, 2] and gamma = 2 - 2/p, there is a constant C independent of f such that the square root of the energy gap plus the L2 distance between the two minimizers is bounded by C times K(n to the minus gamma, f; L2, W-dot-1,p). Under the additional assumption that the continuous minimizer lies in BV intersect L-infinity, the squared L2 error is bounded by C n to the minus one half times that same K-functional times a factor involving the L-infinity norm and the total variation of the continuous minimizer.

The important structural point is that the rate is governed by a variational K-functional of the observed data f, which can in turn be estimated by a classical K-functional. The regularity of the image, not the discretization alone, sets the achievable speed.

Which image regularity guarantees a prescribed rate depends on p

The paper turns the estimate into an answer to a concrete question: given a target rate alpha in (0,1), what regularity of f guarantees that the L2 distance between minimizers is of order n to the minus alpha? Because the K-functional decay of order t to the alpha is equivalent to membership in the real interpolation space (L2, W-dot-1,p) with parameters alpha and infinity, the question reduces to characterizing that interpolation space.

In the Hilbert case p = 2, the answer is classical: the K-functional is comparable to the modulus of smoothness, and the interpolation space is the Besov space B-dot-alpha-2,infinity. The resulting rates split at alpha = 1/2: order n to the minus alpha for alpha in [1/2, 1), and order n to the minus (2 alpha + 1)/4 for alpha in (0, 1/2). The paper notes this decay is optimal for every alpha in (0,1).

Outside the Hilbert setting, where p is not 2, there is no suitable analogue of the modulus-of-smoothness comparison, and the authors describe this as a genuine obstruction. Their route around it uses a recent characterization of the relevant interpolation spaces in terms of the Brezis-Van Schaftingen-Yung spaces, defined through a weak-type supremum over level sets of difference quotients. With that identification, the L2 error is bounded by n to the minus (2 - 2/p) alpha times the B-dot-alpha-q(gamma,infinity) norm of f, provided alpha is at least p divided by 4(p-1).

Extrapolation of compactness unifies compactness criteria and filter limits

A second technical pillar is extrapolation of compactness. The paper proves compactness extrapolation theorems that also involve Gagliardo completions, and then derives several known compactness criteria as corollaries. One corollary recovers a compactness criterion stated in terms of BMO-type functionals; another recovers a criterion stated in terms of heat kernels, where a bound on a rescaled heat-kernel energy forces strong local Lp convergence, with the limit lying in W-dot-1,p locally when p > 1 and in BV locally when p = 1. A third recovers a criterion on compact metric measure spaces involving the KS-1,p space.

The same machinery explains convergence of filters in image processing. The normalized interpolation functionals play the role of non-local filters, and the Gagliardo completion plays the role of the local filter. Applying the main continuity theorem with B = L2, A0 = L1 and A1 = W-dot-1,1 on a fixed cube, together with the identification of the Gagliardo completion as BV, gives L2-stability of minimizers of the interpolated functionals as the parameter approaches 1, recovering the classical ROF model. The choice B = L1 gives the corresponding L1-stability for the Chan-Esedoglu denoising model. The authors present these as instances of one general phenomenon rather than as separate arguments.

The framework is quantitative but not a universal rate formula

Several boundaries matter for reading the results correctly. The ROF rate estimates are stated for p in (1, 2]; the paper explains that p greater than 2 would impose continuity on the observed image through the classical embedding, which the authors regard as uninteresting. The sharper squared-error bound requires the additional assumption that the continuous minimizer lies in BV intersect L-infinity, and the paper notes that while this L-infinity property is automatically satisfied for general data in L2, no estimate of the L-infinity norm in terms of the L2 norm of f is available.

The characterization outside the Hilbert case depends on a recent interpolation formula for (L2, W-dot-1,p) with p not equal to 2, and the resulting rate statement carries the restriction alpha at least p divided by 4(p-1). The compactness corollaries are derived for specific settings, including rearrangement-invariant spaces with absolutely continuous norm on a cube, heat-kernel energies, and compact metric measure spaces satisfying the assumptions of the relevant section. These are the conditions under which the extrapolation machinery applies, and they should travel with any use of the corresponding statements.

The decisive payoff is reading image regularity through K-functional decay

The most decision-relevant content of the paper is not the general framework statement but the specific translation it enables for the ROF model: the regularity of the observed image that guarantees a prescribed optimal rate of convergence is measured by the decay of variational K-functionals, and that decay is captured by the Brezis-Van Schaftingen-Yung spaces. This is what converts a qualitative stability statement into a checkable condition on the data.

For a reader deciding how to use the result, the practical split is between the Hilbert case p = 2, where the classical Besov characterization applies and the rate threshold at alpha = 1/2 governs the exponent, and the non-Hilbert case p not equal to 2, where the Brezis-Van Schaftingen-Yung characterization is required and the additional restriction on alpha applies. The paper's own emphasis is that these two regimes are not interchangeable, and that the K-functional decay, not the discretization scheme alone, is the quantity that determines which convergence rate is actually available.

Sources used in this answer