4
$\begingroup$

Let $f: \mathbb R \rightarrow (-\infty,+\infty]$ be a lower-semicontinuous convex function.

Question

  • Under what futher conditions does there exists a convex decreasing function $\phi: \mathbb R \rightarrow \mathbb R$ such that $f(x) = \sup_y x\phi(y)-\phi(-y)$ for all $x \in \mathbb R$ ?
  • Construct such a $\phi$ explicitly.

Examples

$f(x) = -2\sqrt{x}$ if $ x \ge 0$ and $f(x)=+\infty$ else, then one may take $\phi(x) \equiv e^{-x}$.

Observations

Given a decreasing function $\phi$, define its (pseudo-)inverse by $\phi^{-1}(x):=\inf\{t \in \mathbb R \mid \phi(t) \le x\}$, with the usual convention that $\inf\emptyset = +\infty$. Let $T_\phi(x) := \phi(-\phi^{-1}(x))$, for all $x \in \mathbb R$. Then, with the change of variable $z := \phi(y)$ one has

$$ \sup_y x\phi(y)-\phi(-y)=\sup_zxz-\phi(-\phi^{-1}(z)) := T_\phi^*(x), $$ where $T_\phi^*$ is the convex conjugate of $T_\phi$. This suggests that one should consider the following problem rather:

Reformulated problem. Given a convex l.s.c function $g$, find decreasing convex function $\phi$ such that $g = T_\phi$.

$\endgroup$
2
+50
$\begingroup$

$\newcommand{\R}{\mathbb{R}} \newcommand{\tto}{\underset{\text{onto}}\to}$ Let us answer the reformulated question: given a convex function $g\colon C\to\R$, when is it possible to find a decreasing convex function $\phi\colon\R\to\R$ such that \begin{equation} \phi\circ(-\phi^{-1})=g? \tag{1} \end{equation} Here $C$ is a convex subset of $\R$, that is, an interval with some endpoints $c$ and $d$ such that $-\infty\le c<d\le\infty$ (excluding the trivial case when $c=d$).

First here, note that any convex function from $\R$ to $\R$ is continuous. So, the generalized inverse $\psi:=\phi^{-1}$ of a decreasing convex function $\phi\colon\R\to\R$ is just a regular inverse, and the function $\psi=\phi^{-1}$ is also decreasing and convex.

Also, (1) implies that $\psi=\phi^{-1}$ is defined only on $C$, so that we may write $\phi\colon\R\tto C$ (rather than $\phi\colon\R\to\R$) and $\psi=\phi^{-1}\colon C\tto\R$, where $\tto$ means that the map is onto. So, again in view of (1), we may write
\begin{equation} g=\phi\circ(-\phi^{-1})\colon C\tto C. \tag{1.00} \end{equation} Moreover, because the function $\phi\colon\R\tto C$ is decreasing and convex, we see that necessarily \begin{equation*} C=(c,d)\text{ for }d=\infty\text{ and some}\ c\in[-\infty,\infty). \tag{1.0} \end{equation*}

Next, (1) can be rewritten as \begin{equation} \psi\circ g=-\psi, \tag{1a} \end{equation} which implies $\psi\circ g\circ g=-\psi\circ g=\psi$ and hence \begin{equation} g\circ g=\text{id}_C, \tag{2} \end{equation} where $\text{id}_C$ is the identity map of $C$; this simple observation is crucial.

In particular, $g$ is one-to-one. Since $g$ is convex and l.s.c., it follows that $g$ is continuous. Therefore and because $g$ is one-to-one, we see that either $g$ is increasing on $C$ or decreasing on $C$. The convexity of $g$ also implies the existence of the left and right derivatives $g'_-(x)$ and $g'_+(x)$ at any point $x\in(c,d)$.

Consider first the case when $g$ is increasing on $C$. Then the identity $g(g(x))=x$ for $x\in C$ implies $g'_-(g(x))g'_-(x)=1$ and $g'_+(g(x))g'_+(x)=1$ for all $x\in(c,d)$. So, if $g'_-(a)<g'_+(b)$ for some $a$ and $b$ such that $c<a\le b<d$, then $g'_-(g(a))>g'_+(g(b))$, which latter contradicts the convexity of $g$ (which implies that $g'_-(x)\le g'_+(y)$ for all $x$ and $y$ such that $c<x\le y<d$). So, here we have $g'_-(a)=g'_+(b)$ for all $a$ and $b$ such that $c<a\le b<d$. Thus, if the function $g$ is increasing on $C$, it must be affine. It is then easy to see that the condition $g\circ g=\text{id}_C$ implies $g=\text{id}_C$. So, in this case the only solution of equation (1a) is $\psi=0$, which is not a decreasing function.

It remains to consider the case when $g\colon C\tto C$ is decreasing and convex. Then $g(c+)=d$ and $g(d+)=c$, and $g$ is a bijection of $C$. Since $g$ is continuous and $c<d$, the equation \begin{equation} g(z)=z \tag{3} \end{equation} has a unique root $z\in(c,d)$.

Since $g$ is decreasing and convex, we also have $g'_+(g(x))g'_-(x)=1$ for all $x\in(c,d)$. Hence, $g'_+(z)g'_-(z)=1$. Also, $g'_-(z)\le g'_+(z)\le0$. Therefore, \begin{equation} g'_-(z)\le-1. \tag{4} \end{equation}

For $x\in C$, let now \begin{equation*} \psi(x):=\left\{ \begin{aligned} z-x&\text{ if }x\ge z,\\ g(x)-z&\text{ if }x\le z. \end{aligned} \right. \tag{5} \end{equation*} This definition is valid, because, in view of (3), $z-x=0=g(x)-z$ if $x=z$. Next, $\psi$ is decreasing, since $g$ is decreasing.

Further, $\psi$ is obviously convex on $[z,\infty)\cap C$, and $\psi$ is convex on $(-\infty,z]\cap C$ -- because $g$ is convex. Also, in view of (4), $\psi'_+(z)=-1\ge g'_-(z)=\psi'_-(z)$. So, $\psi$ is convex on $C$.

Moreover, (i) in view of (1.0), $\psi(d-)=\psi(\infty-)=-\infty$ and (ii) in view of (1.00), $\psi(c+)=g(c+)-z=d-z=\infty$, so that $\psi\colon C\tto\R$ and hence $\phi=\psi^{-1}\colon\R\tto C$.

Finally, recalling that $g$ is decreasing and (2) holds, it is straightforward to check that (1a) holds for the so-defined $\psi$.

Thus, for any l.s.c. convex function $g\colon C\to\R$, the following two conditions are equivalent to each other:

(I) there exists a decreasing convex function $\phi\colon\R\to\R$ such that (1) holds;

(II) $C$ is as in (1.0) and $g$ is a decreasing involution of $C$ -- that is, $g$ is a decreasing bijection of $C$ satisfying condition (2).

One may also note that, if condition (II) holds and if $\psi$ is given by (5), then for $\phi=\psi^{-1}$ and all real $y$ we have \begin{equation*} \phi(y)=\left\{ \begin{aligned} z-y&\text{ if }y\le0,\\ g^{-1}(z+y)&\text{ if }y\ge0. \end{aligned} \right. \end{equation*}

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks. Your answer only deals with convex functions with full domans (in the convex analytical sense), i.e for which $g(x) < \infty$ for all $x$. Would you mind extending your answer to extended convex functions as in the question ? Also, what do you mean by $-\infty\pm$ ? $\endgroup$ – dohmatob Sep 22 '19 at 12:42
  • $\begingroup$ @dohmatob : I have now provided the requested modification, allowing $g$ to be defined on any convex subset of $\mathbb R$. As for $g(c+)$, it means $\lim_{x\downarrow c} g(x)$, as usual. Similarly defined is the left limit $g(d-)$. $\endgroup$ – Iosif Pinelis Sep 22 '19 at 15:27
  • $\begingroup$ Thanks very much for your very detailed and instructive answer! $\endgroup$ – dohmatob Sep 22 '19 at 19:26
  • $\begingroup$ As a supplement to this answer, let me add that if we define $U_C:=\{c-z \mid z \le c \in C\}$, then a $\phi$ which does the job is $\phi(x) = \psi^{-1}(x)= \begin{cases}z-x,&\mbox{ if }-x \in U_C,\\g^{-1}(z+x),&\mbox{ if }x \in U_C,\\+\infty,&\mbox{ else.}\end{cases}$ $\endgroup$ – dohmatob Sep 22 '19 at 22:34
  • 1
    $\begingroup$ Oops! Apparently I forgot to save the added explicit expression for that $\phi$. Have done this now. $\endgroup$ – Iosif Pinelis Sep 23 '19 at 13:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.