Conjugated function – Wikipedia
In mathematics, and more precisely in convex analysis, the conjugated function is a function built from a real function
Defined on a vector space
, which is useful:
The combined function of
is most often noted
. It is a convex function, even if
is not, defined on slopes , that is to say on the elements of the dual vector space of
. The definition is motivated and specified below.
L’application
is called transformation de Fenchel or Legendre transformation or Transformation de Legendre-Fnechel , the APRIIIIr Adrien-Mary Legendrere it Werner Fenchel.
Supposed knowledge : linear algebra, differential calculation, the bases of convex analysis (in particular the main concepts attached to convex sets and functions); The subdifferential of a convex function is only used to motivate the definition of combined function.
The apparent complexity of the concept of combined function requires motivating its definition.
Function convexification [ modifier | Modifier and code ]
By definition, a function is convex if its epigraph is convex. Convex a function
consists in determining the largest convex function, let’s say closed, reduced
. In terms of epigraph, this amounts to finding the smallest closed convex containing the epigraph of
, that is to say to take the closed convex envelope of the epigraph
Like any closed convex envelope, that of
is the intersection of every closed half-space containing
, that is, sets of the form
Or
. It is easy to see that when
contains an epigraph, we must have
. It is more thin to show that, in the expression of the closed convex envelope of
as an intersection of
, we can only keep half-spaces with
. However, these are the epigraphs of the Minurants Affines de
of shape
Summons. As you would expect, the closed convenience of
is the upper envelope of all the Minurant Affines of
. Parmi les diminishing affines
having a slope
fixed, one can also keep only the one that is the highest, that is to say the one that has the greatest
. We must therefore determine the greatest
such as
We clearly see that the smallest value of
is given by
It is the value of the combination of
in
. We are interested in
rather than
to define
in order to obtain a conjugated conjugated function.
Access route to the subdifferential [ modifier | Modifier and code ]
In this section, we show the interest of the concept of combined function as a tool to calculate the subdifferential of a convex function.
The subdifferential
of a convex function
in
is the set of slopes of minorating the affines of
(i.e., des fonctions affines that less
), which are accurate in
(i.e., who have the same value as
in
). For
given, it is not always easy to specify all the minor affins exact in
. It is sometimes easier to give yourself a slope
of Minurant Affine
and seek the points it may be exact by modifying
. From this point of view, we are looking for the largest
such as
We clearly see that the smallest value of
is given by
It is the value of the combination of
in
. We are interested in
rather than
to define
in order to obtain a conjugated conjugated function. Let’s get back to the problem we were asking at the start of this paragraph: if
is solution of the maximization problem above, then
for everything
or
These inequalities show that
is a minor refine
, Exact and
.
It is assumed in this section that the functions are defined in a euclidean vector space
(of finished dimension therefore), whose scalar product is noted
and the associated standard
.
On note
The combined and its properties [ modifier | Modifier and code ]
We chose to designate by
l’Argment of
to remember that this is a slope ( slope in English), that is to say an element of the dual space of
, here identified with
via the scalar product.
We are interested below in the convexity and the clean and closed character of the conjugated
. The convexity of
is a remarkable property of the combined, since it is remembered that
is not necessarily convex.
It will be remembered that a convex and clean function necessarily has an affine minor; It therefore checks the properties of point 1 above, so that
The biconjugated and its properties [ modifier | Modifier and code ]
We can of course apply the transformation of Legendre-Fenchel to the combined function
; We thus obtain the biconjugated of
, noted
.
The convex, closed and clean character of the biconjugated is examined in the following result.
Closed convex biconjugate – Regardless
, Biconjugaée
is convex and closed. In addition, the following properties are equivalent:
- is clean and has a Minurant Affine,
- .
If the argument
of
is a slope (identified with an element of
), l’argument
of
is in the starting space
. We can then wonder if there is a link between
And
. The following proposal examines this question. We noted there
, the closure of
.
This result makes it possible to compare the values of
And
.
If we take the combined of the biconjugated, we find the conjugated. There is therefore no concept of triconjuguée .
No triciejugated – Either
A clean function with a minor refine. SO
.
Conjugated rules for calculating [ modifier | Modifier and code ]
Inf-image under a linear application [ modifier | Modifier and code ]
Let us recall the definition of Inf-image of a function under a linear application . We give ourselves two Euclidian spaces
And
(we will need a scalar product here on
And
, while this structure is not necessary in the definition of
), a function
and a linear application
. Then the inf-image of
below
is the noted application
and defined in
about
Pre-composition with a linear application [ modifier | Modifier and code ]
Examples [ modifier | Modifier and code ]
Rules [ modifier | Modifier and code ]
Either
A standard on a Euclidean space
, not necessarily derived from the scalar product
of
. We introduce the dual standard
and the closed duality ball
The combined function
of
is the indicator of the dual unit ball; So she takes in
the following value
Now consider the power
Its conjugated function
Take in
the following value
Or
:
Distance to a convex [ modifier | Modifier and code ]
Maximum proper value [ modifier | Modifier and code ]
Elements of history [ modifier | Modifier and code ]
The combined function was introduced by Mandelbrojt (1939) for a function of a single real variable; Then specified and improved by Fenchel (1949) with convex functions depending on a finished number of variables. The latter introduces the notation
For the combined of
. Conjugation generalizes a function transformation introduced much earlier by Legendre (1787). The extension to topological vector spaces is due to Brønsted (1964), Moreau (1967) and Rockafellar.
Related articles [ modifier | Modifier and code ]
Bibliography [ modifier | Modifier and code ]
- (in) J. M. Borwein et A. S. Lewis, Convex Analysis and Nonlinear Optimization , New York, Springer, .
- (in) A. Brøndsted, ‘ Conjugate convex functions in topological vector spaces. » , Mathematical Physical Messages Published by the Royal Danish Society of Sciences , n O 34, , p. 1–26 .
- (in) W. Fenchel, ‘ On conjugate functions » , Canadian Journal of Mathematics , vol. 1, , p. 73–77 .
- (in) J.-B. Hiriart-Urrey et C. Lemaréchal, ‘ Convex Analysis and Minimization Algorithms » , Basics of mathematical Sciences , Springer-Verlag, , p. 305-306 .
- (in) J.-B. Hiriart-Urrey et C. Lemaréchal, Fundamentals of convex analysis , Berlin, Springer-Verlag, .
- (in) A.M. Legendre, Memory on the integration of some equations with partial differences , coll. «My. ACAD. Sciences », , p. 309–351 .
- (in) S. Mandelbrojt, ‘ On convex functions » , C. R. Acad. Sci. Paris , vol. 209, , p. 977-978 .
- J.J. Moreau, Convex functional , France secondary school, coll. “Seminar with partial drift equations”, .
- (in) R.T. Rockafellar, ‘ Convex Analysis » , Princeton Mathematics , Princeton, New Jersey, Princeton University Press,, 28 It is series, .
Recent Comments