Variable-Growth TV Regularization

• VGTV regularization is a convex variational method that adaptively modulates spatial growth to balance edge preservation with smooth transitions.
• It incorporates variable-exponent models and piecewise-Lipschitz constraints, reducing artifacts and enhancing reconstruction quality.
• Numerical approaches like PDHG and Bregman iterations efficiently solve VGTV minimization, yielding competitive PSNR and SSIM improvements in imaging tasks.

Variable-Growth Total Variation (VGTV) regularization is a class of convex variational methods that generalize classical total variation (TV) regularization by allowing the growth behavior of the regularizing term to vary spatially. This versatility enables adaptive image and signal reconstructions that can enforce spatially-varying smoothness, piecewise-Lipschitz constraints, or variable-exponent regularization, which helps to combine the edge-preserving characteristics of TV with the smoothness of higher-order functionals. VGTV can encompass $p(x)$-type variable-exponent models, double-phase integrands, and pointwise Lipschitz constrained approaches, and has been analyzed with rigorous variational, convex analytical, and partial differential equation tools (Górny et al., 18 Apr 2025, Burger et al., 2019, Kohr, 2017, Burger et al., 2019, Jia et al., 2016).

1. Mathematical Formulations of VGTV

Let $\Omega \subset \mathbb{R}^n$ be a bounded Lipschitz domain, and $u : \Omega \to \mathbb{R}$ a function (typically in $L^1$ or $L^2$). The broad class of VGTV functionals is built from Musielak–Orlicz integrands $\varphi : \Omega \times [0,\infty) \to [0,\infty)$, leading to the energy functional

$E(u) = \int_\Omega \varphi(x, |\nabla u(x)|)\,dx.$

For non-smooth $u$, $E$ is extended to $L^2(\Omega)$ via relaxation, exploiting convexity and lower semicontinuity.

Canonical model cases:

• Variable-exponent ($p(x)$) power growth:

$\varphi(x,t) = \frac{t^{p(x)}}{p(x)},$

where $p : \Omega \to [1, p^+]$ is measurable and log–Hölder continuous, and $p(x)\geq 1$ almost everywhere.

• Double-phase growth:

$\varphi(x, t) = t + a(x) t^q,$

with $q>1$ and $a : \Omega \to [0, A]$ upper semi-continuous, admitting almost-Hölder regularity.

Alternatively, widely used is the "piecewise-Lipschitz TV" (pwL) representation: for $\gamma \in M^+(\Omega)$ a nonnegative Radon measure (e.g., $\gamma(x)=\alpha(x)dx$ for density $\alpha(x)\geq0$),

$$\operatorname{pwL}^\gamma(u) = \inf_{g \in M(\Omega; \mathbb{R}^n): |g| \leq \gamma} \|Du - g\|_M,$$

where $Du$ is the distributional gradient.

A prominent special case is the variable-exponent Lebesgue modular (Kohr, 2017): $J(u) = \int_\Omega |\nabla u(x)|^{p(x)} dx,$ where $p(x): \Omega \rightarrow [1,2]$ can vary with location.

2. Analytical Properties and Dual Representations

VGTV regularizers are convex and lower semicontinuous due to their Musielak–Orlicz structure and infimal representation. The dual representation (via Fenchel–Rockafellar duality) for pwL-type models is

$$\operatorname{pwL}^\gamma(u) = \sup_{\varphi \in C^\infty_c(\Omega;\mathbb{R}^n),\;|\varphi(x)|_2 \leq 1} \left[ \int_\Omega u\,\mathrm{div} \varphi\, dx - \int_\Omega |\varphi(x)|_2\,d\gamma(x) \right].$$

For variable-exponent modulars: $J^*(\sigma) = \int_\Omega r^*(\sigma(x); p(x)) dx,$ where the pointwise convex conjugate $r^*$ depends on $p(x)$. For $p=1$, this yields an indicator function; for $1<p\leq2$, it has explicit form involving fractional powers.

VGTV preserves topological equivalence with TV: for all $u$ and finite $\gamma$,

$$\operatorname{TV}(u) - \gamma(\Omega) \leq \operatorname{pwL}^\gamma(u) \leq \operatorname{TV}(u).$$

The kernel of $\operatorname{pwL}^\gamma$ consists of all (piecewise) Lipschitz functions with local gradient norm bounded by $\gamma(x)$ almost everywhere (Burger et al., 2019).

3. Subdifferential and Euler–Lagrange Theory

VGTV functionals admit an $L^2$-subdifferential characterization (Górny et al., 18 Apr 2025). For the energy $E: L^2(\Omega) \to [0,\infty]$ as above: $$w \in \partial E(u) \iff (E(u) + E^*(w)) = (u, w)_{L^2},$$ where

$$E^*(w) = \inf \left\{\int_\Omega \varphi^*(x, |\xi|)\,dx : \xi \in L^1(\Omega)^n,\; \operatorname{div} \xi = w,\, \xi \cdot n = 0 \text{ on } \partial\Omega \right\},$$

and $\varphi^*$ is the convex conjugate.

The subdifferential admits a calibration: $w \in \partial E(u)$ iff there exists $z \in X_{\varphi^*,2}$ (vector fields with $\int \varphi^*(x,|z|)< \infty$ and $\mathrm{div}\,z \in L^2$) such that

$$\int_\Omega \varphi(x, |\nabla u|)dx + \int_\Omega \varphi^*(x, |z|)dx = \int_\Omega z \cdot \nabla u\, dx,$$

and $\operatorname{div}z = -w$, $z \cdot n = 0$.

In VGTV-regularized variational models (e.g., image denoising with data $f \in L^2$),

$J(u) = \frac{1}{2}\|u-f\|_{L^2}^2 + \lambda E(u),$

there exists a unique minimizer, and the optimality condition is

$$u-f - \lambda\,\operatorname{div}z = 0 \ \text{ in } \Omega, \qquad z \cdot n = 0 \text{ on } \partial\Omega,$$

with explicit $z(x)=\varphi'(x, |\nabla u|)\,\nabla u/|\nabla u|$ where $\nabla u \neq 0$ and where $\varphi$ is differentiable (Górny et al., 18 Apr 2025).

4. Numerical Algorithms

Efficient algorithms for VGTV minimization leverage the convex structure and separability of the functionals. Standard approaches include:

• Primal–dual hybrid gradient methods (PDHG/Chambolle–Pock): These are applicable to both inf-convolutions and modulars, enabling splitting between data fidelity and regularization terms (Burger et al., 2019, Burger et al., 2019, Kohr, 2017).
• Bregman iterative algorithms: In cases where the regularizer changes at each step, modified Bregman iterations efficiently decompose the problem, as in time–space fractional diffusion reconstructions (Jia et al., 2016).

For piecewise-Lipschitz models, the Chambolle–Pock scheme is used, with per-iteration cost essentially linear and explicit closed-form or pixelwise updates for all prox computations. The variable-exponent modular requires solving a scalar nonlinear equation per pixel (Newton iteration) for the proximal mapping, with moderate computational overhead. Algorithmic complexity is $O(NK)$ per iteration ($N=$ pixel number, $K=$ iterations per prox).

Parameter estimation for spatially-adaptive parameters (such as $\gamma(x)$ or $p(x)$) typically involves a pipeline: denoised pilot estimation, residual computation, smoothing, and differentiation. This enables automatic adaptation to image content and edge structure (Burger et al., 2019, Kohr, 2017).

5. Practical Performance and Comparative Results

VGTV methods substantially reduce staircasing artifacts compared to TV, while retaining sharp edge preservation. In denoising and inverse problems:

• Synthetic images: VGTV obtains PSNR/SSIM values close to or exceeding those of TGV², with performance near-ideal if the spatial parameter ($\gamma(x)$ or $p(x)$) is accurately estimated from the ground truth (Burger et al., 2019, Burger et al., 2019, Kohr, 2017).
• Natural images: VGTV has been shown to outperform TV and approach the performance of TGV², with SSIM and PSNR improvements and better preservation of fine textures and smooth shading.
• Time-space fractional inverse problems: VGTV regularization achieves lower reconstruction error than TV on smooth data and comparable performance on piecewise-constant data, with improved edge preservation over Tikhonov regularization (Jia et al., 2016).

VGTV-regularized problems exhibit robust convergence under PDHG/Bregman algorithms, and the reconstruction quality is resilient to moderate inaccuracies in spatial parameter estimation.

6. Connections to Related Regularization Methods

VGTV functionals interpolate between standard TV and higher-order regularization by spatially modulating the growth or penalty order:

• Relation to TV: As the spatial parameter approaches zero or $p(x)\to 1$, VGTV reduces to classical TV.
• Relation to TGV²: By allowing $\gamma(x)$ to be high in regions of smooth variation, VGTV can induce piecewise-affine behavior analogous to TGV², yet at first-order complexity.
• Infimal convolution and duality: VGTV admits infimal convolution representations and explicit dual forms paralleling classical TV, with richer structure for adaptive penalization (Burger et al., 2019, Burger et al., 2019).

Kernel characterizations demonstrate that VGTV can strictly enforce local Lipschitz‐type constraints or allow free variation within prescribed bounds, enabling flexible modeling not possible with TV or TGV alone.

7. Theoretical Guarantees and Regularity

VGTV energies are convex, lower semicontinuous, and coercive on $L^2$ or $L^1$ spaces under mild regularity on the variable-growth integrand. Their penalization of oscillations and jumps can be tuned pointwise, and they satisfy maximum principles analogous to TV. For ROF-type denoising, uniqueness and stability of minimizers hold; for flows, existence and uniqueness of strong solutions are established in Hilbert spaces (Górny et al., 18 Apr 2025). Convergence rates for first-order algorithms are $O(1/N)$ ergodic under standard convexity assumptions.

In summary, VGTV regularization offers a mathematically rigorous, computationally efficient, and highly adaptive framework for inverse problems and imaging, enabling local control of regularization behavior with provable existence, stability, and optimality properties, while bridging the practical and conceptual gap between TV and higher-order models (Górny et al., 18 Apr 2025, Burger et al., 2019, Kohr, 2017, Burger et al., 2019, Jia et al., 2016).