The Calibration Method

arXiv:math/0006016v1 [math.OC] 2 Jun 2000

for Free Discontinuity Problems

Gianni Dal Maso ∗

Abstract
The calibration method is used to identify some minimizers of the
Mumford-Shah functional. The method is then extended to more general
free discontinuity problems.

1 Introduction
In [5] De Giorgi introduced the name free discontinuity problems to denote a
wide class of minimum problems for functionals of the form
Z Z
F (u) := f (x, u(x), ∇u(x))dx + ψ(x, u+(x), u−(x), νu (x))dHn−1, (1)
Ω\Su Su

where Ω is a given bounded domain in Rn with Lipschitz boundary, f : Ω × R ×

Rn → [0, +∞] and ψ: Ω × R × R × Sn−1 → [0, +∞] are given Borel functions,
Sn−1 = {v ∈ Rn : |v| = 1}, Hn−1 is the (n − 1)-dimensional Hausdorff measure,
and the unknown function u: Ω → R is assumed to be regular out of a (partially
regular) singular set Su of dimension n − 1, with unit normal νu , on which u
admits unilateral traces u+ and u− . The main feature of these problems is that
the shape and location of the discontinuity set Su are not prescribed. Thus
minimizing F means optimizing both the function u and the singular set Su ,
which is indeed often regarded as an independent unknown.
These problems have an increasing importance in many branches of applied
analysis, such as image processing (Mumford-Shah functional for image seg-
mentation) and fracture mechanics (Griffith’s criterion and Barenblatt cohesive
zone model).
The Mumford-Shah functional was introduced in [10] in the context of a
variational approach to image segmentation problems (for which we refer to
[9]). It can be written, in dimension n, as
Z Z
Fgα,β (u) := |∇u(x)|2 dx + αHn−1 (Su ) + β |u(x) − g(x)|2 dx , (2)
Ω\Su Ω\Su
∗ SISSA, Via Beirut 4, 34014 Trieste, Italy, e-mail address: [email protected]

1
where g is a given function in L∞ (Ω) (interpreted as the grey level of the image
to be analysed), and α > 0 and β ≥ 0 are constants. When n = 2 (the only case
considered in image processing), the singular set Su of a minimizer u of Fgα,β is
interpreted as the set of the most relevant segmentation lines of the image.
Using different classes of infinitesimal variations, one can show that every
minimizer must satisfy certain equilibrium conditions, which could be globally
called Euler-Lagrange equations for Fgα,β . For instance, u must satisfy the equa-
tion ∆u = β(u − g) on Ω \ Su , with Neumann boundary conditions on Su ∪ ∂Ω.
Moreover, there is a link between the mean curvature of Su (where defined)
and the traces of u and ∇u on the two sides of Su ; for instance, when β = 0,
the mean curvature of Su must be equal to the difference of the squares of the
norms of the traces of ∇u. Additional conditions have been derived for the two-
dimensional case. We refer the reader to [10] and [2] for a precise description of
these equilibrium conditions.
However, since Fgα,β is not convex, all conditions which can be derived by
infinitesimal variations are necessary for minimality, but never sufficient. The
purpose of this note is precisely to present a sufficient condition for minimality
(Theorem 3.1 for Fgα,β and Theorem 3.4 for F ), and give a few applications
(Examples 4.1–4.8). Detailed proofs and further results will be given in the
forthcoming paper [1].

2 Notation and preliminaries

For a complete mathematical treatment of the minimum problems for the func-
tional F considered in (1), we use the space SBV (Ω) of special functions of
bounded variation, introduced by De Giorgi and Ambrosio in [6]. A self-con-
tained presentation of this space can be found in the recent book [2], which
contains also the complete proof of the existence of a minimizer u of Fgα,β , and
of the partial regularity of the corresponding singular set Su (the regularity of
u on Ω \ Su follows from the standard theory of elliptic equations).
We recall that for every u ∈ SBV (Ω) the approximate upper and lower limits
u+(x) and u−(x) at a point x ∈ Ω are defined by

u±(x) := ± inf {t ∈ R : lim ρ−n Ln ({±u > t} ∩ Bρ (x)) = 0} ,

ρ→0+

where Bρ (x) is the open ball with centre x and radius ρ. The singular set (or
jump set ) of u is defined by Su := {x ∈ Ω : u−(x) < u+(x)}. It is known that Su
is countably (Hn−1 , n − 1)-rectifiable and that there exists a Borel measurable
function νu : Su → Sn−1 such that for Hn−1 -a.e. x ∈ Su we have
1
Z
lim n |u(y) − u±(x)| dy = 0 , (3)
ρ→0+ ρ ±
Bρ (x)

where Bρ± (x) := {y ∈ Bρ (x) : ±(y − x) · νu (x) > 0} and · denotes the scalar
product in Rn (see [7, Theorem 4.5.9]). Condition (3) says that νu (x) points
from the side of Su corresponding to u−(x) to the side corresponding to u+(x).

2
 The gradient Du of u is a measure that can be decomposed as the sum of
two measures Du = Da u + Ds u, where Da u is absolutely continuous and Ds u
is singular with respect to the Lebesgue measure Ln . The density of Da u with
respect to Ln is denoted by ∇u. Since u ∈ SBV (Ω), for every Borel set B in Ω
we have
Z Z
(Du)(B) = ∇u(x) dx + (u+(x) − u−(x)) νu (x) dHn−1 .
B B∩Su

The graph of u is defined as

Γu := {(x, t) ∈ Ω × R : u−(x) ≤ t ≤ u+(x)} .

The characteristic function of the subgraph {(x, t) ∈ Ω × R : t ≤ u(x)} is de-

noted by 1u . It is defined by 1u (x, t) := 1 if t ≤ u(x), and 1u (x, t) := 0 if
t > u(x). It belongs to SBV (Ω × R) and its gradient D1u is a measure concen-
trated on Γu .

3 The main results

We fix an open subset U of Ω × R of the form

U := {(x, t) ∈ Ω × R : τ1 (x) < t < τ2 (x)} , (4)

where τ1 and τ2 are two continuous functions on Ω such that −∞ ≤ τ1 (x) ≤

τ2 (x) ≤ +∞ for every x ∈ Ω.
Let F be the functional introduced in (1). We say that a function u ∈
SBV (Ω), with graph Hn -contained in U (i.e., Hn (Γu \ U ) = 0), is a Dirichlet
U -minimizer of F , if F (u) ≤ F (v) for every v ∈ SBV (Ω) with the same trace
as u on ∂Ω and with graph Hn -contained in U . If the inequality F (u) ≤ F (v)
holds for every v ∈ SBV (Ω) with graph Hn -contained in U , we say that u is a
U -minimizer of F . We omit U when U = Ω × R.
The symbol φ will always denote a bounded Borel measurable vectorfield
defined on U with values in Rn+1 = Rn × R, with components φx ∈ Rn and
φt ∈ R. The divergence of φ is then divφ(x, t) = divx φx (x, t) + ∂t φt (x, t).
We begin with a theorem concerning the functional Fgα,β introduced in (2).

Theorem 3.1 Let u ∈ SBV (Ω) with graph Hn -contained in U . Assume that
there exists a bounded vectorfield φ of class C 1 on U with the following proper-
ties:
1 x 2
(a1) 4 |φ (x, t)| ≤ φt (x, t) + β|t − g(x)|2
n
for L -a.e. x ∈ Ω and for every τ1 (x) < t < τ2 (x);

(a2) φx (x, u(x)) = 2 ∇u(x) and φt (x, u(x)) = |∇u(x)|2 − β|u(x) − g(x)|2
for Ln -a.e. x ∈ Ω;

3
 Z t2
(b1) φx (x, t) dt ≤ α
t1
for Hn−1 -a.e. x ∈ Ω and for every τ1 (x) < t1 < t2 < τ2 (x);
Z u+ (x)
(b2) φx (x, t) dt = α νu (x) for Hn−1 -a.e. x ∈ Su ;
u− (x)

(c1) divφ(x, t) = 0 for every (x, t) ∈ U .

Then u is a Dirichlet U -minimizer of Fgα,β . If, in addition, φx (x, t) satisfies the

boundary condition

(c2) lim φx (y, s) · ν(x) = 0

(y,s)→(x,t)

for Hn−1 -a.e. x ∈ ∂Ω and for L1 -a.e. t ∈ [τ1 (x), τ2 (x)],

where ν(x) is the outer unit normal to ∂Ω, then u is a U -minimizer of Fgα,β .

A vectorfield φ which satisfies conditions (a1)–(c1) of Theorem 3.1 is called

a calibration for the functional Fgα,β on U . If φ satisfies also (c2), it is called a
Neumann calibration. Theorem 3.1 is an immediate consequence of the following
lemmas.

Lemma 3.2 Let φ be a vectorfield which satisfies conditions (a1) and (b1) of
Theorem 3.1. Then for every u ∈ SBV (Ω) with graph Hn -contained in U we
have Z
α,β
Fg (u) ≥ φ · d(D1u ) . (5)
U
Moreover, equality holds in (5) for a given u if and only if conditions (a2) and
(b2) of Theorem 3.1 are satisfied.

The next lemma is a consequence of the divergence theorem.

Lemma 3.3 Suppose that φ is of class C 1 and that divφ = 0 on U . Then

Z Z
φ · d(D1u ) = φ · d(D1v ) (6)
U U

for every pair of functions u, v in BV (Ω) with the same trace on ∂Ω and with
graphs Hn -contained in U . If, in addition, φ satisfies condition (c2) of Theo-
rem 3.1, then (6) holds for every pair of functions u, v in BV (Ω) with graphs
Hn -contained in U .

As a matter of fact, the method of calibrations can be easily adapted to the

functional F defined in (1).

Theorem 3.4 Let u ∈ SBV (Ω) with graph Hn -contained in U . Assume that
there exists a bounded vectorfield φ of class C 1 on U with the following proper-
ties:

4
(a1) φx (x, t) · v ≤ φt (x, t) + f (x, t, v) for Ln -a.e. x ∈ Ω,
for every τ1 (x) < t < τ2 (x), and for every v ∈ Rn ;

(a2) φx (x, u(x)) · ∇u(x) = φt (x, u(x)) + f (x, u(x), ∇u(x))

for Ln -a.e. x ∈ Ω;
Z t2
(b1) ν· φx (x, t) dt ≤ ψ(x, t1 , t2 , ν) for Hn−1 -a.e. x ∈ Ω,
t1
for every τ1 (x) < t1 < t2 < τ2 (x), and for every ν ∈ Sn−1 ;
Z u+(x)
(b2) νu (x) · φx (x, t) dt = ψ(x, u− (x), u+ (x), νu (x))
u−(x)
n−1
for H -a.e. x ∈ Su ;

(c1) divφ(x, t) = 0 for every (x, t) ∈ U .

Then u is a Dirichlet U -minimizer of F . If φx (x, t) satisfies also the boundary

condition (c2) of Theorem 3.1, then u is a U -minimizer of F .

Remark 3.5 We note that in Theorem 3.4 there is no regularity or convexity

hypothesis on f or ψ. If f ∗ (x, t, v ∗ ) is the the convex conjugate of f (x, t, v) with
respect to v, condition (a1) is equivalent to

(a1′ ) f ∗ (x, t, φx (x, t)) ≤ φt (x, t)

for Ln -a.e. x ∈ Ω and for every τ1 (x) < t < τ2 (x).

If this condition is satisfied, and f (x, t, v) is convex and differentiable with

respect to v, then condition (a2) is equivalent to

φx (x, u(x)) = ∂v f (x, u(x), ∇u(x))


(a2′ ) for Ln -a.e. x ∈ Ω.
φt (x, u(x)) = f ∗ (x, u(x), φx (x, u(x)))

Remark 3.6 In Theorems 3.1 and 3.4 the hypothesis that φ is of class C 1 is
too strong for many applications. It is used only in Lemma 3.3 and it can be
relaxed in several ways (see [1] for details). For instance, one may consider
piecewise C 1 vectorfields, which may be discontinuous along sufficiently regular
interfaces. In this case the divergence-free condition (c1) must be understood in
the distributional sense, i.e., the pointwise divergence vanishes (where defined)
and the normal component of φ is continuous across the discontinuity surfaces.

5
4 Some examples
The following examples show that the calibration method is very flexible, and
can be used to prove the minimality of a given function u in many different situ-
ations. In the first examples we will consider only
R the “homogeneous” functional
F α := Fgα,0 , in which the lower order term β Ω |u − g|2 dx vanishes.
Example 4.1 (Affine function in one dimension) Let n := 1, Ω := ]0, a[ ,
and u(x) := λx, with λ > 0. It is easy to see that u is a Dirichlet minimizer of
F α if and only if aλ2 ≤ α. In this case a calibration is given by the piecewise
constant function
(
(2λ, λ2 ), if λ2 x ≤ t ≤ λ2 (x + a),
φ(x, t) := (7)
(0, 0), otherwise.
Another calibration is given by

2 t , ( t )2 ,


if 0 ≤ t ≤ λx,
 x x


λa−t λa−t 2


φ(x, t) := 2 a−x , ( a−x ) , if λx ≤ t ≤ λa, (8)


(0, 0), otherwise.


If aλ2 > α, then the function u(x) := λx is not a Dirichlet minimizer of F α ,

but it is still a Dirichlet U -minimizer with
α α
U := {(x, t) ∈ ]0, a[ × R : λx − 4λ < t < λx + 4λ } .

A calibration on U is given by φ(x, t) := (2λ, λ2 ).

Example 4.2 (Jump in one dimension) Let n := 1, Ω := ]0, a[ , u(x) := 0

for 0 < x < c, and u(x) := h for c < x < a, with 0 < c < a and h > 0. It is easy
2
to see that u is a Dirichlet minimizer of F α if and only if aα√≤ h√ . In this case
two different calibrations are given by (7) and (8) with λ = α/ a.√
Suppose now that aα > h2 . Let ε > 0 be a constant such that 2ε+ 2αε ≤ h,
let
−ε, if x ≤ c,



h
τ1 (x) = − ε + ε (x − c), if c ≤ x ≤ c + ε,

h − ε, if c + ε ≤ x,


let τ2 (x) = τ1 (x + ε) + 2ε, and let U be the open set defined by (4). Then u is
a Dirichlet U -minimizer of F α , and a calibration on U is given by the piecewise
constant function

2
 (2λ, λ ), if c − ε < x < c + ε and

φ(x, t) := ε + λ2 (x − c + ε) < t < ε + λ2 (x − c + ε) + 2λ
α
,

(0, 0), otherwise,


α
where λ > 0 is any constant such that ε + ελ + ≤ h − ε, for instance
√ √ 2λ
λ = α/ 2ε.

6
Example 4.3 (Harmonic function) Let Ω be a bounded domain in Rn , n
arbitrary, and let u be a harmonic function on Ω. As pointed out by Chambolle
[3], u is a Dirichlet minimizer of F α if

osc u sup |∇u| ≤ α , (9)

Ω Ω

where oscΩ u := supΩ u − inf Ω u. Note that for n = 1 this condition reduces to
the constraint aλ2 ≤ α of Example 4.1. Inspired by the one dimensional case
(see (7)), we construct the calibration
(
2∇u(x), |∇u(x)|2 , if 12 (u(x) + m) ≤ t ≤ 21 (u(x) + M ),


φ(x, t) := (10)
(0, 0), otherwise,

where m := inf Ω u and M := supΩ u. Another calibration (see (8)) is given by

t−m t−m
)2 |∇u(x)|2 , if m ≤ t ≤ u(x),



 2 u(x)−m ∇u(x), ( u(x)−m

M−t M−t
)2 |∇u(x)|2 , if u(x) ≤ t ≤ M ,


φ(x, t) := 2 M−u(x) ∇u(x), ( M−u(x)



(0, 0), otherwise.

If (9) is not satisfied, u is still is a Dirichlet U -minimizer of F α , for

U := {(x, t) ∈ Ω × R : u(x) − α4 |∇u(x)|−1 < t < u(x) + α4 |∇u(x)|−1 } , (11)

and a calibration in U is given by φ(x, t) := 2∇u(x), |∇u(x)|2 .

Example 4.4 (Pure jump) Let n ≥ 2 and let Ω := ]0, a[ × V , where V is a

bounded domain in Rn−1 with Lipschitz boundary. Denoting the first coordi-
nate of x by x1 , let u(x) := 0 for 0 < x1 < c, and u(x) := h for c < x1 < a, with
0 < c < a and h > 0. Using the results of Example 4.2 it is easy to see that u
is a Dirichlet minimizer of F α if aα ≤ h2 . In this case two different calibrations
can be constructed in the following way: the projection √ of √
these calibrations
onto the (x1 , t)-plane are given by (7) and (8), with λ = α/ a and x replaced
by x1 , while all other components of these calibrations vanish.
If aα > h2 , it may happen that u is still a Dirichlet minimizer of F α . For
instance, if n = 2 and V = ]0, b[ , with bαπ ≤ 2h2 , a different calibration has
been constucted in [1]. Therefore u is a Dirichlet minimizer of F α even if aα is
very large with respect to h2 , provided that bα is small enough.
Arguing as in the last part of Example 4.2 one can prove that for every a
and V there exists an open set U of the form (4), containing Γu , such that u is
a Dirichlet U -minimizer of F α .

Example 4.5 (Triple junction) Let n := 2, let Ω := B(0, r) be the open ball
with radius r > 0 centered at the origin, and let u be given, in polar coordinates,
by u(ρ, θ) := a for 0 ≤ θ < 32 π, u(ρ, θ) := b for 23 π ≤ θ < 34 π, and u(ρ, θ) := c

7
for 34 π ≤ θ < 2π, where a, b, and c are distinct constants. Thus Su is given by
three line segments meeting at the origin with equal angles. If

2αr ≤ min{|a − b|2 , |b − c|2 , |c − a|2 } , (12)

then u is a Dirichlet minimizer of F α . To construct a calibration, it is not

restrictive to assume a < b = 0 < c. Inspired
√ by the one dimensional case
described in Example 4.2, we take e± := (± 3/2, −1/2), and λ > 0 such that
λr α
2 + λ ≤ min{−a, c} (which is possible by (12)), and we define the calibration
by

(λe+ , λ2 /4), if λ4 (r + x · e+ ) ≤ t ≤ λ4 (r + x · e+ ) + αλ ,



φ(x, t) := (λe− , λ2 /4), if λ4 (−r + x · e− ) − αλ ≤ t ≤ λ4 (−r + x · e− ), (13)


(0, 0), otherwise.

If αr is much larger than min{|a− b|2 , |b − c|2 , |c− a|2 }, it is easy to construct
a comparison function v with the same boundary values as u and such that
F α (v) < F α (u). This shows that in this case u is not a Dirichlet minimizer.
However, for every value of the parameters α, r, a, b, c, one can construct a
suitable neighbourhood U of the graph Γu , of the form (4), such that a variant
of (13) is a calibration in U , and therefore u is a Dirichlet U -minimizer of F α .
We refer to [1] for the details.

We consider now the functional Fgα,β , with β > 0.

Example 4.6 (Solution of the Neumann problem) Let Ω be a bounded

open set in Rn with boundary of class C 1,ε for some ε > 0, and let u be the
solution of the Neumann problem
(
∆u = β(u − g) on Ω,
∂u
(14)
∂ν = 0 on ∂Ω,

with β > 0 and g ∈ L∞ (Ω). Assume that condition (9) of Example 4.3 is
satisfied. Then u is a minimizer of Fgα,β . If the strict inequality holds in (9),
then u is the unique minimizer. A Neumann calibration φ(x, t) is given by
 u(x) 2 u(x)
0, β| m m 2
if t − u(x) m


2 − 2 | − β| 2 + 2 − g(x)| , 2 < 2,



u(x)
2∇u(x), |∇u(x)|2 − β|t − g(x)|2 + β|t − u(x)|2 , if m M


 2 ≤t− 2 ≤ 2 ,
u(x) 2 u(x) u(x)

2
0, β| M M
if M


2 − 2 | − β| 2 + 2 − g(x)| , 2 < t− 2 ,


where m := inf Ω u and M := supΩ u.

If (9) is not satisfied, u is still is a U -minimizer of Fgα,β , where U is the open
set defined by (11). A Neumann calibration on U is given by

φ(x, t) := (2∇u(x), |∇u(x)|2 − β|t − g(x)|2 + β|t − u(x)|2 ) .

8
 The hypothesis that ∂Ω is of class C 1,ε is used only to obtain the boundary
condition (c2) of Theorem 3.1, which, in this case, becomes

lim ∇u(y) · ν(x) = 0 for Hn−1 -a.e. x ∈ ∂Ω . (15)

y→x

It is clear that (15) is still true if for Hn−1 -a.e. x ∈ ∂Ω there exists an open
neighbourhood Vx of x in Rn such that Vx ∩ ∂Ω is a manifold of class C 1,ε (see
[2, Theorem 7.5.2]). Therefore the result of this example is true also when Ω is
polyhedral.

In the next examples we construct a calibration for Fgα,β when the parameter
β is large enough.

Example 4.7 (Smooth g and large β) Let Ω be a bounded open set in Rn

with smooth boundary, and let g ∈ C 2 (Ω). There exists a constant β0 ≥ 0,
depending on g and α, such that for every β > β0 the solution u of the Neumann
problem (14) of Example 4.6 is the unique minimizer of Fgα,β . A Neumann
calibration is constructed in [1].
This shows that the minimizer of Fgα,β is smooth, provided that g is smooth
and β is large enough. Therefore the solution of the image segmentation problem
(n = 2) based on the minimization of Fgα,β has an empty set of segmentation
lines if the
R “grey level” function g is smooth and the parameter β in the fidelity
term β Ω |u − g|2 dx is large.

Example 4.8 (Function g with only two values) Let Ω be an open set in
Rn and let E be a compact set contained in Ω with boundary of class C 2 . Let
g(x) := a for x ∈ E and g(x) := b for x ∈ Ω \ E, with a 6= b. There exists
a constant β0 ≥ 0, depending on g and α, such that for every β > β0 the
function u := g is the unique minimizer of Fgα,β . To construct a calibration, it
is not restrictive to assume a < b. We take a C 1 vectorfield v: Ω → Rn with
compact support in Ω such that |v(x)| ≤ 1 for every x ∈ Ω and v(x) is the
outer unit normal to ∂E for every x ∈ ∂E. Then we set φx (x, t) = σ(t)v(x),
where σ is a fixed positive smooth function with integral equal to α and support
contained in ]a, b[ . We see that conditions (b1), (b2), and (c2) of Theorem 3.1
are satisfied by construction. It remains to choose φt so that (a1), (a2), and
(c1) hold. Condition (a2) forces us to set φt (x, t) = 0 for t = g(x), while (c1)
gives ∂t φt (x, t) = −σ(t)divx v(x). These two conditions determine φt (x, t) at
every point (x, t). It is then easy to see that (a1) holds if β is large enough. We
refer to [1] for the details.
This example shows that, if g ∈ SBV (Ω) has only two values, and Sg is
smooth enough, then the minimizer of the Mumford-Shah functional Fgα,β re-
constructs g exactly, when β is large enough.

Recently the following question has been studied by using the calibration
method: is it true that a function u is a (Dirichlet) minimizer of Fgα,β , if it
satisfies the Euler-Lagrange equations and the domain Ω is sufficiently small ?

9
For the moment we have only a partial answer. In [4] we have considered the
case where n := 2 and Su is a line segment joining two points of the boundary of
Ω. If u satisfies the Euler-Lagrange equations for the “homogeneous functional”
F α := Fgα,0 , then for every x0 ∈ Su there exists an open neighbourhood Ω0
of x0 , contained in Ω, such that u is a Dirichlet minimizer of F α in Ω0 . The
minimality is proved by constructing a complicated calibration on Ω0 × R.
This result has been extended in [8] to the case where Su is an analytic curve
joining two points of ∂Ω. The (more difficult) construction of the calibration
presented in this paper shows that one can take the same set Ω0 for every
x0 ∈ Su ; in other words, one can take as Ω0 a suitable tubular neighbourhood
of Su . Moreover, it is proved in [8] that an additional condition on u and Su
implies that u is a Dirichlet U -minimizer for a suitable open neighbourhood U
of the graph Γu . A counterexample (where Su is a line segment joining two
points of ∂Ω) shows that this is not always true when u is just a solution of
the Euler-Lagrange equations with Su 6= ∅, in contrast to the case Su = ∅ (see
Example 4.6).

References
[1] G. Alberti, G. Bouchitté, G. Dal Maso: The calibration method for the
Mumford-Shah functional, paper in preparation.
[2] L. Ambrosio, N. Fusco, D. Pallara: Functions of Bounded Variation and
Free Discontinuity Problems, Oxford Mathematical Monographs, Oxford
University Press, Oxford, 2000.
[3] A. Chambolle: Personal communication, Trieste, 1996.
[4] G. Dal Maso, M.G. Mora, M. Morini: Local calibration for minimizers of
the Mumford-Shah functional with rectilinear discontinuity sets, J. Math.
Pures Appl., 79 (2000), 141–162.
[5] E. De Giorgi: Free discontinuity problems in calculus of variations, in:
Frontiers in Pure and Applied Mathemathics, a collection of papers ded-
icated to Jacques-Louis Lions on the occasion of his sixtieth birthday, R.
Dautray ed., North Holland, Amsterdam (1991), 55–62.
[6] E. De Giorgi, L. Ambrosio: Un nuovo funzionale del calcolo delle variazioni,
Atti Accad. Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur., 82 (1988), 199–210.
[7] H. Federer: Geometric Measure Theory, Springer-Verlag, Berlin, 1969.
[8] M.G. Mora, M. Morini: Local calibration for minimizers of the Mumford-
Shah functional with a regular discontinuity sets, preprint SISSA, Trieste,
2000.
[9] J.-M. Morel, S. Solimini: Variational Methods in Image Segmentation,
Progr. Nonlinear Differential Equations Appl., 14 (1995), Birkhäuser,
Boston.

10
[10] D. Mumford, J. Shah: Optimal approximation by piecewise smooth func-
tions and associated variational problems, Comm. Pure Appl. Math., 42
(1989), 577–685.

11