betaFeedback

WorksMathematics

A non-zero gap for the planar Mumford-Shah convex lifting and the nonexistence of crack-tip calibrations

Abstract

We prove that the convex lifting of the Mumford-Shah functional constructed through the notion of calibration introduced in [2] is not sharp in dimension two. For Dirichlet boundary data prescribed by the crack-tip function, we construct a lifted competitor whose energy is strictly lower than that of the crack-tip subgraph. Since the crack-tip minimizes the original Mumford-Shah functional, this establishes a strict gap between the infima of the original and lifted problems. As a consequence we establish that the crack-tip admits no calibration.

Introduction

In this work we consider the homogeneous Mumford-Shah functional defined for functions of special bounded variation v∈SBV(Ω)v \in\mathrm{SBV}(\Omega) as

MS(v)=∫Ω∣∇v∣2 dx+Hn−1(Sv),\mathrm{MS}(v) = \int_{\Omega} |\nabla v|^2 \, dx + \mathcal{H}^{n-1}(S_v),

where Ω⊂Rn\Omega\subset\mathbb{R}^n is an open, bounded set, SvS_v is the singular set of v∈SBV(Ω)v \in\mathrm{SBV}(\Omega) and Hn−1\mathcal{H}^{n-1} is the n−1n - 1 dimensional Hausdorff measure. Originally introduced by Mumford and Shah [33, 32] as a model for image segmentation, the full functional includes an additional fidelity term penalizing deviations from the observed image gg:

F(v)=MS(v)+∥v−g∥L2(Ω)2F(v) = \mathrm{MS}(v) + \lVert v - g \rVert_{L^2(\Omega)}^2

again considered on v∈SBV(Ω)v \in\mathrm{SBV}(\Omega). The existence and regularity of minimizers of FF, together with the characterization of local and global minimizers of MS\mathrm{MS}, have been the subject of extensive mathematical research over the past four decades [16, 18, 11, 17, 3, 19]. This interest is motivated, in particular, by the central role that the characterization of minimizers of MS\mathrm{MS} plays in understanding the regularity of minimizers of FF [5], and by its close connection to the longstanding Mumford–Shah conjecture [29].1 Notably, a ground-breaking result by Bonnet and David [6] has shown that in R2\mathbb{R}^{2} the so-called crack-tip, defined in polar coordinates as

u^(ρ,θ)=2ρπsin⁡(θ2),−π<θ<π(1)\hat{u}(\rho,\theta)=\sqrt{\frac{2\rho}{\pi}}\sin\left(\frac{\theta}{2}\right),\quad-\pi<\theta<\pi \tag*{(1)}

is a global minimizer of MS.

With the goal of proving the minimality of candidate minimizers, in [2], Alberti, Bouchitté and Dal Maso introduced the theory of calibrations for the Mumford–Shah functional. Their approach builds on foundational work by Federer [23] in geometric measure theory, by Harvey and Lawson [26] in calibrated geometry and generalizes previous approaches related to null-lagrangians of integral functionals [25]. Calibrations are a tool to certify the minimality of a given candidate and, in case of the Mumford–Shah functional, they are divergence free vector fields ϕ\phi satisfying suitable properties tailored to the candidate minimizer. In particular, their construction can provide an elegant proof of minimality with respect to either MS (with Dirichlet boundary conditions) or FF. Following [2] the calibration theory for the Mumford–Shah functional and its variants has also been extensively developed in subsequent years [31, 15, 30]. Moreover, similar approaches have been used for various integral and free discontinuity functionals [1, 36], free boundary problems [28, 20, 12, 27, 13, 34] and mean curvature flows [24, 35]. It was therefore natural to seek a proof of the minimality of the crack-tip through the explicit construction of a suitable calibration. However, despite considerable efforts by several authors, including the first author of this paper, such a construction has remained elusive.

A closely related, more fundamental question is whether the crack-tip admits a calibration at all. A negative answer would explain why attempts to construct one have been unsuccessful. The main goal of this work is to show that the crack-tip does not admit a calibration of the form proposed in [2].

Questions of this kind have been investigated for various classes of functionals [23, 9, 14, 10], and their answers are often closely connected to the sharpness of a suitable convex lifting of the original functional and, more generally, to the presence of a duality gap [7, 8]. For the homogeneous Mumford–Shah functional, the convex lifting has been introduced in [2] as a byproduct of the notion of calibration. It takes the form

Flift(w)=sup⁡ϕ∈K∫ϕ⋅DwF_{\mathrm{lift}}(w)=\sup_{\phi\in K}\int\phi\cdot Dw

where w∈BVloc(Ω×R)w\in BV_{\mathrm{loc}}(\Omega\times\mathbb{R}) and KK is the set of bounded Borel vector fields ϕ:Ω×R→Rn+1\phi:\Omega\times\mathbb{R}\to\mathbb{R}^{n+1} satisfying the properties:

{ϕt(x,t)≥∣ϕx(x,t)∣24for a.e. x∈Ω and every t∈R,∣∫t1t2ϕt(x,s) ds∣≤1for every x∈Ω and every t1,t2∈R.\begin{cases} \phi^{t}(x,t) \ge\frac{|\phi^{x}(x,t)|^{2}}{4} & \text{for a.e. } x \in\Omega\text{ and every } t \in\mathbb{R}, \\ \left|\int_{t_{1}}^{t_{2}} \phi^{t}(x,s)\,\mathrm{d}s\right| \le1 & \text{for every } x \in\Omega\text{ and every } t_{1},t_{2} \in\mathbb{R}. \end{cases}

To establish that this lifting is not sharp for prescribed boundary data, it suffices to find a Dirichlet minimizer uu of MS and a competitor ww, agreeing with 1{u>t}\mathbf{1}_{\{u>t\}} (the indicator function of the subgraph of uu) outside a given set, such that Flift(w)<MS(u)F_{\mathrm{lift}}(w)<\mathrm{MS}(u), that is a non-zero gap exists. Crucially, such a non-zero gap also rules out the existence of a calibration for uu.

In this work, we show that the convex lifting of the Mumford–Shah functional is not sharp in dimension two. We take u^\hat{u} to be the crack-tip, whose minimality follows from Bonnet and David [6]. We show that for this choice a non-zero gap exists effectively ruling out the existence of calibration for the crack-tip. Our result also identifies a limitation of numerical approaches based on this convex lifting. For boundary data exhibiting a positive gap, a minimizer of the lifted problem cannot be the subgraph lift of an admissible SBVSBV function. Therefore, solving the lifted problem does not, in general, recover a minimizer of the original functional.

We conclude this introduction by outlining the main idea of the proof. We construct a competitor wL∈BVloc(Ω×R)w_{L}\in BV_{\mathrm{loc}}(\Omega\times\mathbb{R}) where Ω\Omega is chosen to be the annulus Ω={x∈R2:e−4L<∣x∣<2e−L}\Omega=\{x\in\mathbb{R}^{2}:e^{-4L}<|x|<2e^{-L}\}. Here LL is a positive parameter that will be chosen sufficiently big. This annular choice allows us to establish a strictly positive lifting gap in Ω\Omega and it is made to avoid the fact that calibrations for the crack-tip (if exists) need to be chosen singular close to the origin, due to the crack-tip singularity (see Remark 3.1). We construct wLw_{L} by perturbing the lifted subgraph of the crack-tip while preserving its boundary data at the annulus. Specifically, we use the layer-cake construction

wL(x,t)=12∫−111{vL,a>t}(x,t) dawithvL,a(x)=u^(x)+ahL(x)w_{L}(x,t)=\frac{1}{2}\int_{-1}^{1}\mathbf{1}_{\{v_{L,a}>t\}}(x,t)\,\mathrm{d}a \quad\text{with} \quad v_{L,a}(x)=\hat{u}(x)+ah_{L}(x)

where hLh_{L} is a radial cut-off function vanishing for ∣x∣≤e−3L|x|\le e^{-3L} and ∣x∣≥e−L|x|\ge e^{-L} and having a plateau equal to e−Le^{-L} on a smaller annulus inside Ω\Omega. Such perturbation is not modifying the jump set of the crack-tip. Moreover, on the plateau, we have ∇vL,a=∇u^\nabla v_{L,a}=\nabla\hat{u}. There, the key observation is that for any admissible field ϕ\phi the lifting ∫ϕ⋅DwL\int\phi\cdot Dw_{L} cannot recover the Dirichlet energy of the crack-tip. Indeed, due to the first constraint in the definition of KK, no-gap implies that

ϕx(x,t)=2∇u^(x),ϕt(x,t)=∣∇u^(x)∣2.\phi^{x}(x,t)=2\nabla\hat{u}(x), \qquad\phi^{t}(x,t)=|\nabla\hat{u}(x)|^{2}.

The second condition in KK rules this out wherever 4e−L∣∇u^∣>14e^{-L}|\nabla\hat{u}|>1, which holds throughout a smaller annulus inside the plateau because of the crack-tip singularity: ∣∇u^(x)∣=(2π∣x∣)−1/2|\nabla\hat{u}(x)|=(2\pi|x|)^{-1/2}. Quantitatively, this construction produces an averaged deficit of order e−2Le^{-2L}, uniformly over all admissible fields. The cutoff, that is crucially chosen logarithmically, restores the boundary data at a cost of order e−2L/Le^{-2L}/L, so choosing LL sufficiently large provides a competitor with strictly smaller lifted energy than the crack-tip.

AI statement. The initial idea for constructing a competitor exhibiting a positive gap emerged from discussions between the authors and GPT-6 Astra in September 2026. Following several unsuccessful attempts to construct a calibration for the crack-tip, GPT-6 Astra suggested investigating the existence of a positive gap in the convex lifting and subsequently proposed the initial competitor. The paper is built on the initial idea and it has been integrally written without the use of AI, apart from a grammar and a notational check. The authors take full responsibility for the mathematical content, references and presentation of the manuscript.

Reflections and Acknowledgments 2. Having spent years of the PhD trying to calibrate the crack-tip, this problem has obviously a deeply personal significance for me. Perhaps this explains why completing the paper and finally having an answer to this question feels bittersweet almost ten years after completing my PhD. On one hand, I am enthusiastic about the result and glad to understand why my attempts to construct a calibration could not succeed. On the other hand, knowing that the key idea came from an AI leaves me with this sense of bitterness.

Anyway I believe that some acknowledgements are in order. First and foremost, I would like to thank my PhD advisor, Bernd Kirchheim, for introducing me to this problem and for the many hours we spent exploring strategies, working through computations, and discussing unsuccessful attempts. I am also grateful to those who worked on this problem before me and generously shared their approaches and insights. In particular, I have fond memories of discussions with Giovanni Alberti, Antonin Chambolle, Guy David, and Massimo Gobbino that also helped me get through my PhD.

Preliminaries

We first recall some basic notations and definitions we will use throughout the paper. We always denote by Ω\Omega an open, bounded set of Rn\mathbb{R}^n. Moreover, we denote by Hn−1(A)\mathcal{H}^{n-1}(A) the n−1n-1 dimensional Hausdorff measure of a given Borel set AA. We will work with functions with bounded variation defined on an open, bounded set UU in Rd\mathbb{R}^d (in this paper d=nd=n or d=n+1d=n+1) and denote them by BV⁡(U)\operatorname{BV}(U). They are defined as functions in L1(U)L^1(U) such that their distributional derivative is a bounded measure on UU. Given v∈BV⁡(U)v\in\operatorname{BV}(U) we denote by SvS_v its singular set. Since SvS_v is (d−1)(d-1)-rectifiable, its measure-theoretical normal νSv\nu_{S_v} is well-defined Hd−1\mathcal{H}^{d-1} almost everywhere [4]. Moreover, for Hd−1\mathcal{H}^{d-1}-a.e. x∈Svx\in S_v the traces of vv, denoted by v+(x)v^+(x) and v−(x)v^-(x), are well-defined as the approximate limits of vv at x∈Svx\in S_v [4]. We use the convention that v+(x)≥v−(x)v^+(x)\geq v^-(x).

It is well-known, that the derivative of a function of bounded variation v∈BV⁡(U)v\in\operatorname{BV}(U) can be decomposed as

Dv=∇v dx+(v+−v−)νSvHd−1⌞Sv+DcvDv = \nabla v\,dx + (v^+ - v^-)\nu_{S_v}\mathcal{H}^{d-1}\llcorner S_v + D^c v

where DvcD^c_v is the Cantor part of the derivative [4]. The set of special functions of bounded variation, denoted by SBV(Ω\Omega), are the functions v∈BV(Ω)v \in BV(\Omega) such that Dvc=0D^c_v = 0.

Given v∈SBV⁡(Ω)v \in\operatorname{SBV}(\Omega) we will denote by 1{v>t}\mathbf{1}_{\{v>t\}} the indicator function of its subgraph, that is

1{v>t}(x,t)={1if v(x)>t,0otherwise.\mathbf{1}_{\{v>t\}}(x,t) = \begin{cases} 1 & \text{if } v(x)>t,\\ 0 & \text{otherwise.} \end{cases}

Since v∈SBV⁡(Ω)v \in\operatorname{SBV}(\Omega), it holds that 1{v>t}∈SBV⁡(Ω×I)\mathbf{1}_{\{v>t\}} \in\operatorname{SBV}(\Omega\times I) for every open, bounded interval II.

The homogeneous Mumford-Shah functional is defined for v∈SBV⁡(Ω)v \in\operatorname{SBV}(\Omega) as

MS⁡(v)=∫Ω∣∇v∣2 dx+Hn−1(Sv).\operatorname{MS}(v) = \int_{\Omega} |\nabla v|^2 \, dx + \mathcal{H}^{n-1}(S_v).

The study of local and global minimizers of the Mumford-Shah functional has been an area of active research since the model was defined by Mumford and Shah. Here is the precise definition that is also relevant for this paper.

Definition 2.1 (Local and global minimizers). We say that u∈SBV⁡(Ω)u \in\operatorname{SBV}(\Omega) is a local minimizer of the Mumford-Shah functional if MS⁡(u)<∞\operatorname{MS}(u) < \infty and

MS⁡(u)≤MS⁡(v)\operatorname{MS}(u) \leq\operatorname{MS}(v)

for every v∈SBV⁡(Ω)v \in\operatorname{SBV}(\Omega) such that {u≠v}\{u \neq v\} is compactly contained in Ω\Omega. We also say that u∈SBV⁡loc(Rn)u \in\operatorname{SBV}_{\mathrm{loc}}(\mathbb{R}^n) is a global minimizer if its restriction to any open bounded set is a local minimizer.

In particular one of the most striking result concern proving the for n=2n = 2 the crack-tip defined in (1) is a global minimizer as stated in the next theorem due to Bonnet and David.

Theorem 2.2 (Bonnet, David). The crack-tip is a global minimizer of the Mumford-Shah functional

Due to the length of their proof, around 25 years ago, researchers have wondered if the crack-tip minimality can be proven by a calibration argument, considerably shortening it. Indeed, a calibration method was introduced by Alberti, Bouchitté and Dal Maso in 2001 [2] and thanks to that examples show that the construction of a calibration allowed to easily prove the local minimality of the triple junction and of harmonic functions under suitable assumptions. However, despite the effort in constructing such calibration all the attempts failed. We would like to summarize briefly the calibration method for the Mumford-Shah functional.

Definition 2.3. Given u∈SBV⁡(Ω)u \in\operatorname{SBV}(\Omega), a calibration for uu is a C1C^1 vector field in ϕ=(ϕx,ϕt):Rn+1→Rn+1\phi= (\phi^x,\phi^t): \mathbb{R}^{n+1} \to\mathbb{R}^{n+1} such that div⁡ϕ=0\operatorname{div}\phi= 0 and with the following properties:

a) ϕt(x,t)≥∣ϕx(x,t)∣24\phi^t(x,t) \geq\frac{|\phi^x(x,t)|^2}{4} for a.e. x∈Ωx \in\Omega and every t∈Rt \in\mathbb{R},

b) ∣∫t1t2ϕx(x,s) ds∣≤1\left|\int_{t_1}^{t_2}\phi^x(x,s)\,\mathrm{d}s\right| \le1 for Hn−1\mathcal{H}^{n-1}-a.e. x∈Ωx \in\Omega and every t1,t2∈Rt_1,t_2 \in\mathbb{R},

c) ϕx(x,u(x))=2∇u(x)\phi^x(x,u(x)) = 2\nabla u(x) and ϕt(x,u(x))=∣∇u(x)∣2\phi^t(x,u(x)) = |\nabla u(x)|^2 for a.e. x∈Ωx \in\Omega,

d) ∫u−(x)u+(x)ϕx(x,s) ds=νSu(x)\int_{u_-(x)}^{u_+(x)}\phi^x(x,s)\,\mathrm{d}s = \nu_{S_u}(x) for Hn−1\mathcal{H}^{n-1}-a.e. x∈Sux \in S_u.

As shown in [2] the existence of a calibration of a given u∈SBV⁡(Ω)u \in\operatorname{SBV}(\Omega) automatically implies its local minimality in Ω\Omega. Moreover, it is worth to remark that the C1C^1 regularity of ϕ\phi can be relaxed by essentially asking enough regularity to ensure the validity of the divergence theorem. For example, in [2], the class of approximately regular vector fields has been considered allowing for discontinuous calibrations.

The convex lifting for the Mumford-Shah functional and its gap

Due to the presence of Hn−1(Sv)\mathcal{H}^{n-1}(S_v) the Mumford-Shah functional is naturally non-convex. However, the set of calibrations can be used to define a convex lifting. More precisely, for w∈BV⁡loc(Ω×R)w \in\operatorname{BV}_{\mathrm{loc}}(\Omega\times\mathbb{R}) one can define the convex functional

Flift(w)=sup⁡ϕ∈K∫ϕ⋅DwF_{\mathrm{lift}}(w)=\sup_{\phi\in K}\int\phi\cdot Dw

where KK is the set of bounded Borel vector fields ϕ:Ω×R→Rn+1\phi:\Omega\times\mathbb{R}\to\mathbb{R}^{n+1} satisfying the properties:

{ϕt(x,t)≥∣ϕx(x,t)∣24for a.e. x∈Ω and every t∈R∣∫t1t2ϕx(x,s) ds∣≤1for every x∈Ω and every t1,t2∈R.\left\{ \begin{aligned} \phi^t(x,t)&\ge\frac{|\phi^x(x,t)|^2}{4} &&\text{for a.e. }x\in\Omega\text{ and every }t\in\mathbb{R}\\ \left|\int_{t_1}^{t_2}\phi^x(x,s)\,\mathrm{d}s\right|&\le1 &&\text{for every }x\in\Omega\text{ and every }t_1,t_2\in\mathbb{R}. \end{aligned} \right.

It is easy to see that for every v∈SBV⁡(Ω)v \in\operatorname{SBV}(\Omega) it holds that Flift(1{v>t})=MS(v)F_{\mathrm{lift}}(\mathbf{1}_{\{v>t\}})=\mathrm{MS}(v) (Remark 3.3 [2]). However, a much deeper question is establishing if the convex lifting is sharp, meaning if it has the same minimum value as the Mumford-Shah functional under suitable boundary conditions. In more precise words, given u∈SBV⁡(Ω)u\in\operatorname{SBV}(\Omega) a local minimizer of MS one asks if

inf⁡w∈GuFlift(w)=MS(u).\inf_{w\in\mathcal{G}_u} F_{\mathrm{lift}}(w)=\mathrm{MS}(u).

where Gu={v∈BV⁡loc(Ω×R):{v≠1{u>t}}⊂Ω×R}\mathcal{G}_u=\{v\in\operatorname{BV}_{\mathrm{loc}}(\Omega\times\mathbb{R}):\{v\ne\mathbf{1}_{\{u>t\}}\}\subset\Omega\times\mathbb{R}\}. In dimension one (n=1n=1), this was proven to be true by Chambolle in [14]. The goal of this paper is to show that this is false in dimension two.

The lifting FliftF_{\mathrm{lift}} is connected to the notion of calibration by the following simple observation. Given u∈SBV⁡(Ω)u\in\operatorname{SBV}(\Omega) a local minimizer of the Mumford-Shah functional, if there exists a calibration for uu according to Definition 2.3, then

inf⁡w∈GuFlift(w)=MS(u).\inf_{w\in\mathcal{G}_u}F_{\mathrm{lift}}(w)=\mathrm{MS}(u).

This can be easily seen by observing that if ϕˉ\bar{\phi} is a calibration for uu, then for all w∈BVloc(Ω×R)w \in\mathrm{BV}_{\mathrm{loc}}(\Omega\times\mathbb{R}) agreeing with 1{u>t}1_{\{u>t\}} outside a compact set contained in Ω×R\Omega\times\mathbb{R}, it holds

Flift(w)≥∫Ω×Rϕˉ⋅Dw=∫Ω×Rϕˉ⋅D1{u>t}=MS(u).F_{\mathrm{lift}}(w) \ge\int_{\Omega\times\mathbb{R}} \bar{\phi} \cdot Dw = \int_{\Omega\times\mathbb{R}} \bar{\phi} \cdot D1_{\{u>t\}} = \mathrm{MS}(u).

In this paper we show that this is not the case when u^\hat{u} is the crack-tip.

Non-zero gap for the crack-tip

In this section, let L>0L > 0 be a parameter that will be chosen big enough and we consider the annulus

Ω={x∈R2:e−4L<∣x∣<2e−L}.\Omega= \{x \in\mathbb{R}^{2}: e^{-4L} < |x| < 2e^{-L}\}.

Note that for notational convenience we do not make the dependence on LL in Ω\Omega explicit. Moreover, we set I=(−2,2)I = (-2,2) and we let u^∈SBV(Ω)\hat{u} \in\mathrm{SBV}(\Omega) be the crack-tip. Our goal is to show that for big enough LL it holds that

inf⁡w∈Gu^Flift(w)<MS(u^).\inf_{w \in\mathcal{G}_{\hat{u}}} F_{\mathrm{lift}}(w) < \mathrm{MS}(\hat{u}).

where

Gu^={w∈BV(Ω×I):{w≠1{u^>t}}⊆Ω×I}.\mathcal{G}_{\hat{u}} = \{w \in\mathrm{BV}(\Omega\times I): \{w \ne1_{\{\hat{u}>t\}}\} \subseteq\Omega\times I\}.

In particular, we aim to construct a competitor w∈Gu^w \in\mathcal{G}_{\hat{u}} such that Flift(w)<MS(u^)F_{\mathrm{lift}}(w) < \mathrm{MS}(\hat{u}). By proving this statement we will show the non-existence of calibration for the crack-tip in Ω\Omega and also in any set Ω′⊃Ω\Omega' \supset\Omega by a simple restriction argument. Also note that the set II is useful to work with functions defined on compact cylinders. It is enough to choose any open interval containing the image of the crack-tip in Ω\Omega. From now on we will work with functions defined in Ω×I\Omega\times I, replacing also BVloc\mathrm{BV}_{\mathrm{loc}} with BV\mathrm{BV}.

Remark 3.1. Note that in this paper we always work with the annulus Ω\Omega as a domain. This choice is done to have a non-trivial result. Indeed, due to the condition ϕx=2∇u^\phi^{x} = 2\nabla\hat{u} in Definition 2.3 any calibrating vector field would have the singularity of the crack-tip at the origin. Therefore there is no hope to find C1C^{1} calibrations according to Definition 2.3, or even approximately regular calibrations according to [2].

Our goal is to construct a family of functions wL:Ω×I→Rw_{L} : \Omega\times I \to\mathbb{R} and to show that for suitable choice of L>0L > 0 such wLw_{L} belongs to Gu^\mathcal{G}_{\hat{u}} and Flift(wL)<MS(u^)F_{\mathrm{lift}}(w_{L}) < \mathrm{MS}(\hat{u}). We first define

q(τ)={1∣τ∣≤122(1−∣τ∣)12<∣τ∣<10∣τ∣≥1q(\tau) = \begin{cases} 1 & |\tau| \le\frac{1}{2} \\ 2(1-|\tau|) & \frac{1}{2} < |\tau| < 1 \\ 0 & |\tau| \ge1 \end{cases}

and the cut-off function hL:Ω→Rh_L:\Omega\to\mathbb{R}

hL(x)=e−Lq(log⁡(∣x∣)+2LL).h_L(x)=e^{-L}q\left(\frac{\log(|x|)+2L}{L}\right).

The family of competitors wL:Ω×I→Rw_L:\Omega\times I \to\mathbb{R} is defined as:

vL,a(x)=u^(x)+ahL(x)a∈(−1,1),wL(x,t)=12∫−111{vL,a>t}(x,t) da.v_{L,a}(x)=\hat{u}(x)+ah_L(x) \qquad a\in(-1,1), \qquad w_L(x,t)=\frac{1}{2}\int_{-1}^{1}\mathbf{1}_{\{v_{L,a}>t\}}(x,t)\,da.

It is easy to see that for L≥2L\geq2 the function wLw_L is indeed an admissible competitor as next lemma shows.

Lemma 3.2. The function wLw_L belongs to BV(Ω×I)\mathrm{BV}(\Omega\times I) and, for L≥2L\geq2, it holds that

i)i) wL(x,t)=1{u^>t}(x,t)w_L(x,t)=\mathbf{1}_{\{\hat{u}>t\}}(x,t) if x∈Ωx\in\Omega, ∣x∣≤e−3L|x|\leq e^{-3L} or ∣x∣≥e−L|x|\geq e^{-L}, and t∈It\in I,

ii)ii) wL(x,t)=1w_L(x,t)=1 for all (x,t)∈Ω×(−2,−1)(x,t)\in\Omega\times(-2,-1),

iii)iii) wL(x,t)=0w_L(x,t)=0 for all (x,t)∈Ω×(1,2)(x,t)\in\Omega\times(1,2).

In particular wL∈Gu^w_L\in\mathcal{G}_{\hat{u}}.

Proof. First, let us prove that wL∈BV(Ω×I)w_L\in\mathrm{BV}(\Omega\times I). Note that wLw_L is bounded in Ω×I\Omega\times I and thus L1(Ω×I)L^1(\Omega\times I). Moreover, since the total variation of its distributional derivative can be computed as

∣DwL∣(Ω×I)=12∫−11∣D1vL,a∣(Ω×I) da==12∫−11(∫Ω1+∣∇u^(x)+a∇hL(x)∣2 dx+∫Su^∣u^+(x)−u^−(x)∣dH1(x))da<∞,\begin{aligned} |Dw_L|(\Omega\times I)=\frac{1}{2}\int_{-1}^{1}|D\mathbf{1}_{v_{L,a}}|(\Omega\times I)\,da = \\ &=\frac{1}{2}\int_{-1}^{1}\left(\int_{\Omega}\sqrt{1+|\nabla\hat{u}(x)+a\nabla h_L(x)|^2}\,dx+\int_{S_{\hat{u}}}|\hat{u}^{+}(x)-\hat{u}^{-}(x)|d\mathcal{H}^{1}(x)\right)da \\ &<\infty, \end{aligned}

since hLh_L is Lipschitz. Since q(τ)=0q(\tau)=0 for ∣τ∣≥1|\tau|\geq1, we have hL(x)=0h_L(x)=0 whenever ∣x∣≤e−3L|x|\leq e^{-3L} or ∣x∣≥e−L|x|\geq e^{-L}. This proves (i)(i). Note also that for x∈Ωx\in\Omega, a∈(−1,1)a\in(-1,1) and L≥2L\geq2 it holds

vL,a(x)=u^(x)+ahL(x)≤2π+e−L≤2π+e−2<1v_{L,a}(x)=\hat{u}(x)+ah_L(x)\leq\sqrt{\frac{2}{\pi}}+e^{-L}\leq\sqrt{\frac{2}{\pi}}+e^{-2}<1

proving iii)iii). ii)ii) can be proven similarly, implying that wL∈Gu^w_L\in\mathcal{G}_{\hat{u}}.

We also compute the Dirichlet energy of hLh_L in Ω\Omega.

Lemma 3.3. It holds that

∫Ω∣∇hL(x)∣2 dx=e−2L8πL.\int_{\Omega}|\nabla h_L(x)|^2\,dx=e^{-2L}\frac{8\pi}{L}.

Proof. We compute

∫Ω∣∇uL(x)∣2 dx=e−2LL2∫Ωq′(log⁡(∣x∣)+2LL)21∣x∣2 dx=2πe−2LL2∫e−4L2e−Lq′(log⁡(r)+2LL)21r dr.\int_{\Omega} |\nabla u_L(x)|^2\,dx = \frac{e^{-2L}}{L^2}\int_{\Omega} q'\left(\frac{\log(|x|)+2L}{L}\right)^2\frac{1}{|x|^2}\,dx = \frac{2\pi e^{-2L}}{L^2}\int_{e^{-4L}}^{2e^{-L}}q'\left(\frac{\log(r)+2L}{L}\right)^2\frac{1}{r}\,dr.

Set now τ=log⁡(r)+2LL\tau=\frac{\log(r)+2L}{L}. This gives

=2πe−2LL∫−2log⁡2+LLq′(τ)2 dτ=2πe−2LL(∫−1−1/24 dτ+∫1/214 dτ)=e−2L8πL= \frac{2\pi e^{-2L}}{L}\int_{-2}^{\frac{\log2+L}{L}}q'(\tau)^2\,d\tau= \frac{2\pi e^{-2L}}{L}\left(\int_{-1}^{-1/2}4\,d\tau+\int_{1/2}^{1}4\,d\tau\right)=e^{-2L}\frac{8\pi}{L}

as we wanted to show. □

Theorem 3.4. There exists a sufficiently large L>2log⁡4L>2\log4 such that

Flift(uL)<MS⁡(u^).F_{\mathrm{lift}}(u_L)<\operatorname{MS}(\hat{u}).

Proof. Let ϕ=(ϕx,ϕt)∈K\phi=(\phi^x,\phi^t)\in K and v∈SBV(Ω)v\in SBV(\Omega) with ∣v(x)∣≤1|v(x)|\leq1 almost everywhere. Then denoting by Q=Ω×IQ=\Omega\times I, we can compute

MS⁡(v)−∫Qϕ⋅D1{v(x)>t}=∫Ω∣∇v(x)∣2−ϕx(x,v(x))⋅∇v(x)+ϕt(x,v(x)) dx+(∫Sv1−νSv(x)⋅∫v−(x)v+(x)ϕx(x,t) dt)dH1(x)=∫Ω∣∇v(x)−12ϕx(x,v(x))∣2+ϕt(x,v(x))−14∣ϕx(x,v(x))∣2 dx++(∫Sv1−νSv(x)⋅∫v−(x)v+(x)ϕx(x,t) dt)dH1(x).\begin{aligned} \operatorname{MS}(v)-\int_Q\phi\cdot D\mathbf{1}_{\{v(x)>t\}}=\int_{\Omega}|\nabla v(x)|^2-\phi^x(x,v(x))\cdot\nabla v(x)+\phi^t(x,v(x))\,dx \\ &+\left(\int_{S_v}1-\nu_{S_v}(x)\cdot\int_{v^-(x)}^{v^+(x)}\phi^x(x,t)\,dt\right)d\mathcal{H}^1(x) \\ &=\int_{\Omega}\left|\nabla v(x)-\frac{1}{2}\phi^x(x,v(x))\right|^2+\phi^t(x,v(x))-\frac{1}{4}|\phi^x(x,v(x))|^2\,dx+ \\ &+\left(\int_{S_v}1-\nu_{S_v}(x)\cdot\int_{v^-(x)}^{v^+(x)}\phi^x(x,t)\,dt\right)d\mathcal{H}^1(x). \end{aligned}

Applying the previous computation with v=vL,av=v_{L,a} it holds that

∫Qϕ⋅DwL=12∫Q∫−11ϕ⋅D1{vL,a(x)>t} da=12∫−11∫Qϕ⋅D1{vL,a(x)>t} da−12∫−11MS⁡(vL,a) da+12∫−11MS⁡(vL,a) da=12∫−11MS⁡(vL,a) da−12∫−11∫Ω∣∇vL,a(x)−12ϕx(x,vL,a(x))∣2+ϕt(x,vL,a(x))−14∣ϕx(x,vL,a(x))∣2 dx da\begin{aligned} \int_Q\phi\cdot D w_L=\frac{1}{2}\int_Q\int_{-1}^{1}\phi\cdot D\mathbf{1}_{\{v_{L,a}(x)>t\}}\,da \\ &=\frac{1}{2}\int_{-1}^{1}\int_Q\phi\cdot D\mathbf{1}_{\{v_{L,a}(x)>t\}}\,da-\frac{1}{2}\int_{-1}^{1}\operatorname{MS}(v_{L,a})\,da+\frac{1}{2}\int_{-1}^{1}\operatorname{MS}(v_{L,a})\,da \\ &=\frac{1}{2}\int_{-1}^{1}\operatorname{MS}(v_{L,a})\,da-\frac{1}{2}\int_{-1}^{1}\int_{\Omega}\left|\nabla v_{L,a}(x)-\frac{1}{2}\phi^x(x,v_{L,a}(x))\right|^2+\phi^t(x,v_{L,a}(x))-\frac{1}{4}|\phi^x(x,v_{L,a}(x))|^2\,dx\,da \end{aligned}
−12∫−11(∫JvL,a1−νSL,a(x)⋅∫vL,a−(x)vL,a+(x)ϕx(x,t) dt)dH1(x) da≤12∫−11MS⁡(vL,a) da−12∫−11∫Ω∣∇vL,a(x)−12ϕx(x,vL,a(x))∣2 dx da(2)\begin{aligned} -\frac{1}{2}\int_{-1}^{1}\left(\int_{J_{v_{L,a}}}1-\nu_{S_{L,a}}(x)\cdot\int_{v_{L,a}^-(x)}^{v_{L,a}^+(x)}\phi^x(x,t)\,dt\right)d\mathcal{H}^1(x)\,da \\ &\le\frac{1}{2}\int_{-1}^{1}\operatorname{MS}(v_{L,a})\,da-\frac{1}{2}\int_{-1}^{1}\int_{\Omega}\left|\nabla v_{L,a}(x)-\frac{1}{2}\phi^{x}(x,v_{L,a}(x))\right|^{2}\,dx\,da \tag*{(2)} \end{aligned}

where in the last inequality we used that ϕ∈K\phi\in K. Now note that

12∫−11MS⁡(vL,a) da=12∫−11∫Ω∣∇u^(x)+a∇hL(x)∣2 dx da+12∫−11H1(Su)=MS⁡(u)+12∫−11∫Ωa2∣∇hL(x)∣2 dx da+∫−11∫Ωa∇u^(x)⋅∇hL(x) dx da=MS⁡(u)+12∫−11∫Ωa2∣∇hL(x)∣2 dx da=MS⁡(u)+13∫Ω∣∇hL(x)∣2 dx.(3)\begin{aligned} \frac{1}{2}\int_{-1}^{1}\operatorname{MS}(v_{L,a})\,da &=\frac{1}{2}\int_{-1}^{1}\int_{\Omega}\left|\nabla\widehat{u}(x)+a\nabla h_{L}(x)\right|^{2}\,dx\,da+\frac{1}{2}\int_{-1}^{1}\mathcal{H}^{1}(S_{u})\\ &=\operatorname{MS}(u)+\frac{1}{2}\int_{-1}^{1}\int_{\Omega}a^{2}\left|\nabla h_{L}(x)\right|^{2}\,dx\,da+\int_{-1}^{1}\int_{\Omega}a\nabla\widehat{u}(x)\cdot\nabla h_{L}(x)\,dx\,da\\ &=\operatorname{MS}(u)+\frac{1}{2}\int_{-1}^{1}\int_{\Omega}a^{2}\left|\nabla h_{L}(x)\right|^{2}\,dx\,da\\ &=\operatorname{MS}(u)+\frac{1}{3}\int_{\Omega}\left|\nabla h_{L}(x)\right|^{2}\,dx. \tag*{(3)} \end{aligned}

Moreover, using Jensen’s inequality and the fact that ϕ∈K\phi\in K one has that

12∫−11∫Ω∣∇vL,a(x)−12ϕx(x,vL,a(x))∣2 dx da=12∫Ω∫−11∣∇u^(x)+a∇hL(x)−12ϕx(x,u(x)+ahL(x))∣2 da dx≥∫Ω∣12∫−11∇u^(x)+a∇hL(x)−12ϕx(x,u^(x)+ahL(x)) da∣2 dx=∫Ω∣12∫−11∇u^(x)−12ϕx(x,u^(x)+ahL(x)) da∣2 dx=∫Ω∣∇u^(x)−∫−1114ϕx(x,u^(x)+ahL(x)) da∣2 dx≥∫Ω∩{hL>0}∣∇u^(x)−14hL(x)∫u^(x)−hL(x)u^(x)+hL(x)ϕx(x,z) dz∣2 dx≥∫Ω∩{hL>0}(∣∇u^(x)∣−14hL(x))2 dx.(4)\begin{aligned} \frac{1}{2}\int_{-1}^{1}\int_{\Omega}\left|\nabla v_{L,a}(x)-\frac{1}{2}\phi^{x}(x,v_{L,a}(x))\right|^{2}\,dx\,da &=\frac{1}{2}\int_{\Omega}\int_{-1}^{1}\left|\nabla\widehat{u}(x)+a\nabla h_{L}(x)-\frac{1}{2}\phi^{x}(x,u(x)+ah_{L}(x))\right|^{2}\,da\,dx\\ &\ge\int_{\Omega}\left|\frac{1}{2}\int_{-1}^{1}\nabla\widehat{u}(x)+a\nabla h_{L}(x)-\frac{1}{2}\phi^{x}(x,\widehat{u}(x)+ah_{L}(x))\,da\right|^{2}\,dx\\ &=\int_{\Omega}\left|\frac{1}{2}\int_{-1}^{1}\nabla\widehat{u}(x)-\frac{1}{2}\phi^{x}(x,\widehat{u}(x)+ah_{L}(x))\,da\right|^{2}\,dx\\ &=\int_{\Omega}\left|\nabla\widehat{u}(x)-\int_{-1}^{1}\frac{1}{4}\phi^{x}(x,\widehat{u}(x)+ah_{L}(x))\,da\right|^{2}\,dx\\ &\ge\int_{\Omega\cap\{h_{L}>0\}}\left|\nabla\widehat{u}(x)-\frac{1}{4h_{L}(x)}\int_{\widehat{u}(x)-h_{L}(x)}^{\widehat{u}(x)+h_{L}(x)}\phi^{x}(x,z)\,dz\right|^{2}\,dx\\ &\ge\int_{\Omega\cap\{h_{L}>0\}}\left(\left|\nabla\widehat{u}(x)\right|-\frac{1}{4h_{L}(x)}\right)^{2}\,dx. \tag*{(4)} \end{aligned}

for all ϕ∈K\phi\in K. Therefore defining the annulus

AL={x:e−2L4≤∣x∣≤e−2L}\mathcal{A}_{L}=\left\{x:\frac{e^{-2L}}{4}\leq|x|\leq e^{-2L}\right\}

we note that hL(x)=e−Lh_{L}(x)=e^{-L} in AL\mathcal{A}_{L}. Indeed, since L>2log⁡4L>2\log4, we get that 0≥log⁡(∣x∣)+2LL≥−log⁡4L>−120\geq\frac{\log(|x|)+2L}{L}\geq-\frac{\log4}{L}>-\frac{1}{2}. Therefore, using Lemma 3.3

∫Qϕ⋅DwL≤MS⁡(u^)+13∫Ω∣∇hL(x)∣2 dx−∫Ω∩{hL>0}(∣∇u^(x)∣−14hL(x))2 dx+\int_{Q}\phi\cdot D w_{L}\leq\operatorname{MS}(\widehat{u})+\frac{1}{3}\int_{\Omega}\left|\nabla h_{L}(x)\right|^{2}\,dx-\int_{\Omega\cap\{h_{L}>0\}}\left(\left|\nabla\widehat{u}(x)\right|-\frac{1}{4h_{L}(x)}\right)^{2}\,dx+
≤MS⁡(u^)+e−2L8π3L−∫AL(12π∣x∣−14e−L)+2 dx=MS⁡(u^)+e−2L8π3L−∫AL(12π∣x∣−14e−L)2 dx=MS⁡(u^)+e−2L8π3L−2πe−2L∫1/41r(12πr−14)2 dr=MS⁡(u^)+e−2L(8π3L−2π∫1/41r(12πr−14)2 dr).\begin{aligned} &\leq\operatorname{MS}(\hat{u}) + e^{-2L}\frac{8\pi}{3L} - \int_{A_L}\left(\frac{1}{\sqrt{2\pi|x|}}-\frac{1}{4e^{-L}}\right)_{+}^{2}\,dx \\ &= \operatorname{MS}(\hat{u}) + e^{-2L}\frac{8\pi}{3L} - \int_{A_L}\left(\frac{1}{\sqrt{2\pi|x|}}-\frac{1}{4e^{-L}}\right)^{2}\,dx \\ &= \operatorname{MS}(\hat{u}) + e^{-2L}\frac{8\pi}{3L} - 2\pi e^{-2L}\int_{1/4}^{1}r\left(\frac{1}{\sqrt{2\pi r}}-\frac{1}{4}\right)^{2}\,dr \\ &= \operatorname{MS}(\hat{u}) + e^{-2L}\left(\frac{8\pi}{3L}-2\pi\int_{1/4}^{1}r\left(\frac{1}{\sqrt{2\pi r}}-\frac{1}{4}\right)^{2}\,dr\right). \end{aligned}

Therefore, choosing LL big enough and since the right hand side does not depend on ϕ\phi we immediately obtain that sup⁡ϕ∈K∫Qϕ⋅DuL<MS⁡(u^)\sup_{\phi\in K}\int_Q\phi\cdot D u_L<\operatorname{MS}(\hat{u}) as we wanted to prove.

Remark 3.5. Our construction provides an explicit lower bound on the relaxation gap, although it does not determine its exact value. By considering MS⁡\operatorname{MS} and FlitF_{\mathrm{lit}} defined on ΩL={x∈R2:e−2L<∣x∣<2e−L}\Omega_L=\{x\in\mathbb{R}^2:e^{-2L}<|x|<2e^{-L}\} we define

ΔL:=MS⁡(u^)−inf⁡w∈Gu^Flit(w).\Delta_L:=\operatorname{MS}(\hat{u})-\inf_{w\in\mathcal{G}_{\hat{u}}}F_{\mathrm{lit}}(w).

Set

η:=2π∫1/41r(12πr−14)2 dr=34−72π24+15π256≈0.203\eta:=2\pi\int_{1/4}^{1}r\left(\frac{1}{\sqrt{2\pi r}}-\frac{1}{4}\right)^2\,dr=\frac{3}{4}-\frac{7\sqrt{2\pi}}{24}+\frac{15\pi}{256}\approx0.203

The estimate in the proof of Theorem 3.4 thus gives

ΔL≥e−2L(η−8π3L).\Delta_L\geq e^{-2L}\left(\eta-\frac{8\pi}{3L}\right).

This lower bound is strictly positive whenever L>8π3η≈41.273L>\frac{8\pi}{3\eta}\approx41.273. So choosing L=42L=42 we get Δ42>e−90\Delta_{42}>e^{-90}. This is obviously a crude estimate; however, we believe that the real gap is also small. This could explain why a positive gap has never been convincingly detected using numerical methods.

References

  1. [1]G. Alberti, L. Ambrosio, and X. Cabré. On a long-standing conjecture of E. De Giorgi: symmetry in 3D for general nonlinearities and a local minimality property. Acta Applicandae Mathematica, 65(1-3):9–33, 2001.DOI
  2. [2]G. Alberti, G. Bouchitté, and G. Dal Maso. The calibration method for the Mumford-Shah functional and free-discontinuity problems. Calculus of Variations and Partial Differential Equations, 16(3):299–333, 2003.DOI
  3. [3]L. Ambrosio, N. Fusco, and J. E. Hutchinson. Higher integrability of the gradient and dimension of the singular set for minimisers of the mumford–shah functional. Calculus of Variations and Partial Differential Equations, 16(2):187–215, 2003.DOI
  4. [4]L. Ambrosio, N. Fusco, and D. Pallara. Functions of bounded variation and free discontinuity problems. Oxford university press, 2000.
  5. [5]A. Bonnet. On the regularity of edges in image segmentation. In Annales de l’Institut Henri Poincaré C, Analyse non linéaire, volume 13, pages 485–528. Elsevier, 1996.DOI
  6. [6]A. Bonnet and G. David. Cracktip is a global Mumford-Shah minimizer. Société mathématique de France, 2001.
  7. [7]G. Bouchitté. Convex analysis and duality. arXiv preprint arXiv:2004.09330, 2020.arxiv.org/abs/2004.09330
  8. [8]G. Bouchitté and I. Fragalà. Duality for non-convex variational problems. Comptes Rendus. Mathématique, 353(4):375–379, 2015.
  9. [9]G. Bouchitté and M. Phan. Calibration for minimal surfaces with free boundary and cheeger-type problems. arXiv preprint arXiv:2511.03322, 2025.arxiv.org/abs/2511.03322
  10. [10]M. Carioni. A discrete corea-type formula for the Mumford-Shah functional in dimension one. arXiv preprint arXiv:1610.01846, 2016.arxiv.org/abs/1610.01846
  11. [11]M. Carioni. Existence of calibrations for the mumford-shah functional and the reinitialization of the distance function in the framework of chan and vese algorithms. 2017.
  12. [12]M. Carioni and A. Pluda. Calibrations for minimal networks in a covering space setting. ESAIM: Control, Optimisation and Calculus of Variations, 26:40, 2020.arxiv.org/abs/1707.01448
  13. [13]M. Carioni and A. Pluda. On different notions of calibrations for minimal partitions and minimal networks in R2. Advances in Calculus of Variations, 14(3):401–417, 2021.
  14. [14]A. Chambolle. Convex representation for lower semicontinuous envelopes of functionals in L1. Journal of Convex Analysis, 8(1):149–170, 2001.
  15. [15]G. Dal Maso, M. G. Mora, and M. Morini. Local calibrations for minimizers of the mumford–shah functional with rectilinear discontinuity sets. Journal de mathématiques pures et appliquées, 79(2):141–162, 2000.
  16. [16]G. David. Singular sets of minimizers for the Mumford-Shah functional. Springer, 2005.
  17. [17]C. De Lellis and M. Focardi. Higher integrability of the gradient for minimizers of the 2d mumford–shah energy. Journal de Mathématiques Pures et Appliquées, 100(3):391–409, 2013.
  18. [18]C. De Lellis and M. Focardi. The regularity theory for the Mumford-Shah functional on the plane. EMS Press, 2025.arxiv.org/abs/2308.14660
  19. [19]G. De Philippis and A. Figalli. Higher integrability for minimizers of the mumford–shah functional. Archive for Rational Mechanics and Analysis, 213(2):491–502, 2014.
  20. [20]G. De Philippis and E. Paolini. A short proof of the minimality of simons cone. Rendiconti del Seminario Matematico della Università di Padova, 121:233–241, 2009.
  21. [21]F. Deangelis. Classification of boundary blow-ups for the Mumford-Shah functional with Dirichlet boundary conditions. arXiv preprint arXiv:2610.23513, 2026.arxiv.org/abs/2610.23513
  22. [22]F. Deangelis. Solution of the Mumford-Shah conjecture. arXiv preprint arXiv:2609.26732, 2026.arxiv.org/abs/2609.26732
  23. [23]H. Federer. Real flat chains, cochains and variational problems. Indiana University Mathematics Journal, 24(4):351–407, 1974.DOI
  24. [24]J. Fischer, S. Hensel, T. Laux, and T. M. Simon. Local minimizers of the interface length functional based on a concept of local paired calibrations. arXiv preprint arXiv:2212.11840, 2022.arxiv.org/abs/2212.11840
  25. [25]M. Giaquinta and S. Hildebrandt. Calculus of variations II. Springer Science & Business Media, 2013.
  26. [26]R. Harvey and H. B. Lawson. Calibrated geometries. 1982.DOI
  27. [27]F. Hélein. Isoperimetric inequalities and calibrations. PITMAN RESEARCH NOTES IN MATHEMATICS SERIES, pages 92–105, 1996.
  28. [28]G. Lawlor and F. Morgan. Paired calibrations applied to soap films, immiscible fluids, and surfaces or networks minimizing other norms. Pacific journal of Mathematics, 166(1):55–83, 1994.DOI
  29. [29]A. Lemenant. A selective review on mumford–shah minimizers. Bollettino dell’Unione Matematica Italiana, 9(1):69–113, 2016.DOI
  30. [30]M. G. Mora. Local calibrations for minimizers of the mumford–shah functional with a triple junction. Communications in Contemporary Mathematics, 4(02):297–326, 2002.DOI
  31. [31]M. G. Mora and M. Morini. Local calibrations for minimizers of the Mumford–Shah functional with a regular discontinuity set. In Annales de l’Institut Henri Poincaré C, Analyse non linéaire, volume 18, pages 403–436. Elsevier, 2001.DOI
  32. [32]D. Mumford and J. Shah. Boundary detection by minimizing functionals. In IEEE Conference on computer vision and pattern recognition, volume 17, pages 137–154. San Francisco, 1985.
  33. [33]D. Mumford and J. Shah. Optimal approximations by piecewise smooth functions and associated variational problems. 1989.DOI
  34. [34]V. Pagliari. Halfspaces minimise nonlocal perimeter: a proof via calibrations: V. pagliari. Annali di Matematica Pura ed Applicata (1923-), 199(4):1685–1696, 2020.DOI
  35. [35]A. Pluda and M. Pozzetta. Minimizing properties of networks via global and local calibrations. Bulletin of the London Mathematical Society, 55(6):3029–3052, 2023.DOI
  36. [36]T. Pock, A. Chambolle, D. Cremers, and H. Bischof. A convex relaxation approach for computing minimal partitions. In 2009 IEEE conference on computer vision and pattern recognition, pages 810–817. IEEE, 2009.

Paper details

vlnc:W_de28a3st
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents