## Uniqueness properties of functionals with Lipschitzian derivative

### November 27, 2006

It is, of course, of intrinsic mathematical interest to know whether a real-valued function has a unique global maximum (or minimum) in a given set. In particular, it is interesting to know whether for a point ${x}_{0}$  out of a subset $C$  of a metric space $\left(X,d\right)$  there is exactly one point $y\in C$  such that $d\left({x}_{0},y\right)=dist\left({x}_{0},C\right)$  .
The aim of this short note is to establish such uniqueness properties in connection with functionals on Hilbert spaces having Lipschitzian derivative.
With the usual convention $inf\varnothing =+\infty$  , our main result reads as follows:
THEOREM 1. - Let $X$  be a real Hilbert space and let $J:X\to \mathbf{R}$  be a sequentially weakly upper semicontinuous ${C}^{1}$  functional, with Lipschitzian derivative. Let $L$  be the Lipschitz constant of ${J}^{\prime }$  .
Then, for each ${x}_{0}\in X$  with ${J}^{\prime }\left({x}_{0}\right)\ne 0$  , if we set ${\alpha }_{0}={inf}_{x\in {M}_{\frac{1}{L}}}J\left(x\right)$  and ${\beta }_{0}=dist\left({x}_{0},{M}_{\frac{1}{L}}\right),$  where ${M}_{\frac{1}{L}}$  is the set of all global minima of the functional $x\to \frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}-\frac{1}{L}J\left(x\right)$  , we have ${\alpha }_{0}>J\left({x}_{0}\right)$  , ${\beta }_{0}>0$  , and the following properties hold:
$\left(i\right)$  for every $r\in \right]J\left({x}_{0}\right),{\alpha }_{0}\left[$  there exists a unique ${y}_{r}\in {J}^{-1}\left(r\right)$  such that $\parallel {x}_{0}-{y}_{r}\parallel =dist\left({x}_{0},{J}^{-1}\left(r\right)\right);$  $\left(ii\right)$  for every $r\in \right]0,{\beta }_{0}\left[$  the restriction of the functional $J$  to the set $\left\{x\in X:\parallel x-{x}_{0}\parallel =r\right\}$  has a unique global maximum. The main tool used to get Theorem 1 is the following particular case of Theorem 3 of [1].
THEOREM A. - Let $X$  be a reflexive real Banach space, $I\subseteq \mathbf{R}$  an interval and $\Psi :X×I\to \mathbf{R}$  a function such that $\Psi \left(x,\cdot \right)$  is concave for all $x\in X$  , while $\Psi \left(\cdot ,\lambda \right)$  is sequentially weakly lower semicontinuous for all $\lambda \in I$  and coercive, with a unique local minimum for all $\lambda \in int\left(I\right)$  .
Then, one has ${sup}_{\lambda \in I}{inf}_{x\in X}\Psi \left(x,\lambda \right)={inf}_{x\in X}{sup}_{\lambda \in I}\Psi \left(x,\lambda \right).$  We will also use the two propositions below.
PROPOSITION 1. - Let $Y$  be a nonempty set, $f,g:Y\to \mathbf{R}$  two functions, and $a,b$  two real numbers, with $a  . Let ${y}_{a}$  be a global minimum of the function $f-ag$  and ${y}_{b}$  a global minimum of the function $f-bg$  .
Then, one has $g\left({y}_{a}\right)\le g\left({y}_{b}\right)$  . If either ${y}_{a}$  or ${y}_{b}$  is strict and ${y}_{a}\ne {y}_{b}$  , then $g\left({y}_{a}\right)  .
PROOF. We have $f\left({y}_{a}\right)-ag\left({y}_{a}\right)\le f\left({y}_{b}\right)-ag\left({y}_{b}\right)$  as well as $f\left({y}_{b}\right)-bg\left({y}_{b}\right)\le f\left({y}_{a}\right)-bg\left({y}_{a}\right).$  Summing, we get $-ag\left({y}_{a}\right)-bg\left({y}_{b}\right)\le -ag\left({y}_{b}\right)-bg\left({y}_{a}\right)$  and so $\left(b-a\right)g\left({y}_{a}\right)\le \left(b-a\right)g\left({y}_{b}\right)$  from which the first conclusion follows. If either ${y}_{a}$  or ${y}_{b}$  is strict and ${y}_{a}\ne {y}_{b}$  , then one of the first two inequalities is strict and hence so is the third one. $△$  PROPOSITION 2. - Let $Y$  be a real Hibert space and let $\phi :Y\to \mathbf{R}$  be a sequentially weakly upper semicontinuous ${C}^{1}$  functional whose derivative is a contraction.
Then, for every ${y}_{0}\in Y$  , the functional $y\to \frac{1}{2}\parallel y-{y}_{0}{\parallel }^{2}-\phi \left(y\right)$  is coercive and has a unique local minimum.
PROOF. Let $\nu$  be the Lipschitz constant of ${\phi }^{\prime }$  . So, $\nu <1$  , by assumption.
For each $y\in Y$  , we have $\phi \left(y\right)=\phi \left(0\right)+{\int }_{0}^{1}〈{\phi }^{\prime }\left(ty\right),y〉dt$  and so $|\phi \left(y\right)|\le |\phi \left(0\right)|+{\int }_{0}^{1}|〈{\phi }^{\prime }\left(ty\right),y〉|dt\le |\phi \left(0\right)|+\parallel y\parallel {\int }_{0}^{1}\parallel {\phi }^{\prime }\left(ty\right)dt\le$  $\le |\phi \left(0\right)|+\parallel y\parallel \left({\int }_{0}^{1}\parallel {\phi }^{\prime }\left(ty\right)-{\phi }^{\prime }\left(0\right)\parallel dt+\parallel {\phi }^{\prime }\left(0\right)\parallel \right)\le |\phi \left(0\right)|+\frac{\nu }{2}\parallel y{\parallel }^{2}+\parallel {\phi }^{\prime }\left(0\right)\parallel \parallel y\parallel .$  From this, we then get $\frac{1}{2}\parallel y-{y}_{0}{\parallel }^{2}-\phi \left(y\right)\ge \frac{1-\nu }{2}\parallel y{\parallel }^{2}-\left(\parallel {\phi }^{\prime }\left(0\right)\parallel +\parallel {y}_{0}\parallel \right)\parallel y\parallel +\frac{1}{2}\parallel {y}_{0}{\parallel }^{2}-|\phi \left(0\right)|$  and hence ${lim}_{\parallel y\parallel \to +\infty }\frac{1}{2}\parallel y-{y}_{0}{\parallel }^{2}-\phi \left(y\right)=+\infty$  which yields our first claim. Then, the functional $y\to \frac{1}{2}\parallel y-{y}_{0}{\parallel }^{2}-\phi \left(y\right)$  has a global minimum, since it is sequentially weakly lower semicontinuous. But the critical points of this functional are exactly the fixed points of the operator ${\phi }^{\prime }+{y}_{0}$  that is a contraction. So, the functional has a unique local minimum (that is its global minimum). $△$  Proof of Theorem 1. Since ${J}^{\prime }\left({x}_{0}\right)\ne 0$  , we can choose $y\in X$  such that $〈{J}^{\prime }\left({x}_{0}\right),y〉>0$  . We then have
 $\begin{array}{c}{lim}_{\lambda \to {0}^{+}}\frac{J\left({x}_{0}+\lambda y\right)-J\left({x}_{0}\right)}{{\lambda }^{2}}=+\infty .\end{array}$ (1)
Fix any $\gamma >L$  . Therefore, $\frac{1}{\gamma }{J}^{\prime }$  is a contraction, and so, by Proposition 2, the functional $x\to \frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}-\frac{1}{\gamma }{J}^{\prime }\left(x\right)$  has a strict global minimum, say ${x}_{\frac{1}{\gamma }}$  .
Now, by $\left(1\right)$  , we can choose $\lambda >0$  so that $\frac{J\left({x}_{0}+\lambda y\right)-J\left({x}_{0}\right)}{{\lambda }^{2}}>\frac{\gamma \parallel y{\parallel }^{2}}{2}.$  Then, if we put ${z}_{0}={x}_{0}+\lambda y$  , we have $-\frac{1}{\gamma }J\left({x}_{0}\right)>\frac{1}{2}\parallel {z}_{0}-{x}_{0}{\parallel }^{2}-\frac{1}{\gamma }J\left({z}_{0}\right).$  So, a fortiori, we have $-\frac{1}{\gamma }J\left({x}_{0}\right)>\frac{1}{2}\parallel {x}_{\frac{1}{\gamma }}-{x}_{0}{\parallel }^{2}-\frac{1}{\gamma }J\left({x}_{\frac{1}{\gamma }}\right).$  From this, it follows that ${x}_{\frac{1}{\gamma }}\ne {x}_{0}$  . Then, by Proposition 1, we have $J\left({x}_{0}\right)  for all $x\in {M}_{\frac{1}{L}}$  , and so $J\left({x}_{0}\right)<{\alpha }_{0}$  . Clearly, ${x}_{\frac{1}{\gamma }}$  is the global minimum of the functional $x\to \frac{\gamma }{2}\parallel x-{x}_{0}{\parallel }^{2}-J\left(x\right)$  , while any $z\in {M}_{\frac{1}{L}}$  is a global minimum of the functional $x\to \frac{L}{2}\parallel x-{x}_{0}{\parallel }^{2}-J\left(x\right)$  . Consequently, if we apply Proposition 1 again (with $f\left(x\right)=-J\left(x\right)$  , $g\left(x\right)=-\parallel x-{x}_{0}{\parallel }^{2}$  , $a=\frac{L}{2}$  , $b=\frac{\gamma }{2}$  ), for any $z\in {M}_{\frac{1}{L}}$  , we get $-\parallel z-{x}_{0}{\parallel }^{2}\le -\parallel {x}_{\frac{1}{\gamma }}-{x}_{0}{\parallel }^{2},$  and so ${\beta }_{0}\ge \parallel {x}_{\frac{1}{\gamma }}-{x}_{0}\parallel >0.$  Now, to prove $\left(i\right)$  , fix $r\in \right]J\left({x}_{0}\right),{\alpha }_{0}\left[$  and consider the function $\Psi :X×\left[0,\frac{1}{L}\right]\to \mathbf{R}$  defined by $\Psi \left(x,\lambda \right)=\frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}+\lambda \left(r-J\left(x\right)\right)$  for all $\left(x,\lambda \right)\in X×\left[0,\frac{1}{L}\right]$  . Taken Proposition 2 into account, it is clear that the function $\Psi$  satisfies all the assumptions of Theorem A. Consequently, we have ${sup}_{\lambda \in \left[0,\frac{1}{L}\right]}{inf}_{x\in X}\Psi \left(x,\lambda \right)={inf}_{x\in X}{sup}_{\lambda \in \left[0,\frac{1}{L}\right]}\Psi \left(x,\lambda \right).$  The functional ${sup}_{\lambda \in \left[0,\frac{1}{L}\right]}\Psi \left(\cdot ,\lambda \right)$  is sequentially weakly lower semicontinuous and coercive, and so there exists ${x}^{*}\in X$  such that ${sup}_{\lambda \in \left[0,\frac{1}{L}\right]}\Psi \left({x}^{*},\lambda \right)={inf}_{x\in X}{sup}_{\lambda \in \left[0,\frac{1}{L}\right]}\Psi \left(x,\lambda \right).$  Also, the function ${inf}_{x\in X}\Psi \left(x,\cdot \right)$  is upper semicontinuous, and so there exists ${\lambda }^{*}\in \left[0,\frac{1}{L}\right]$  such that ${inf}_{x\in X}\Psi \left(x,{\lambda }^{*}\right)={sup}_{\lambda \in \left[0,\frac{1}{L}\right]}{inf}_{x\in X}\Psi \left(x,\lambda \right).$  Hence, from this it follows that $\frac{1}{2}\parallel {x}^{*}-{x}_{0}{\parallel }^{2}+{\lambda }^{*}\left(r-J\left({x}^{*}\right)\right)={inf}_{x\in X}\frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}+{\lambda }^{*}\left(r-J\left(x\right)\right)={sup}_{\lambda \in \left[0,\frac{1}{L}\right]}\frac{1}{2}\parallel {x}^{*}-{x}_{0}{\parallel }^{2}+\lambda \left(r-J\left({x}^{*}\right)\right).$  We claim that $J\left({x}^{*}\right)=r$  . Indeed, if it were $J\left({x}^{*}\right)  , then we would have ${\lambda }^{*}=\frac{1}{L}$  , and so ${x}^{*}\in {M}_{\frac{1}{L}}$  , against the fact that $r<{\alpha }_{0}$  . If it were $J\left({x}^{*}\right)>r$  , then we would have ${\lambda }^{*}=0$  , and so ${x}^{*}={x}_{0}$  , against the fact that $J\left({x}_{0}\right)  .
We then have $\frac{1}{2}\parallel {x}^{*}-{x}_{0}{\parallel }^{2}={inf}_{x\in X}\frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}+{\lambda }^{*}\left(r-J\left(x\right)\right).$  This implies, on one hand, that ${\lambda }^{*}<\frac{1}{L}$  (since $r<{\alpha }_{0}$  ) and, on the other hand, that each global minimum of the restriction to ${J}^{-1}\left(r\right)$  of the functional $x\to \frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}$  is a global minimum in $X$  of the functional $x\to \frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}-{\lambda }^{*}J\left(x\right)$  . But this functional (just because ${\lambda }^{*}<\frac{1}{L}$  ) has a unique global minimum, and so $\left(i\right)$  follows. Let us now prove $\left(ii\right)$  . To this end, fix $r\in \right]0,{\beta }_{0}\left[$  and consider the function $\Phi :X×\left[L,+\infty \left[\to \mathbf{R}$  definded by $\Phi \left(x,\lambda \right)=\frac{\lambda }{2}\left(\parallel x-{x}_{0}{\parallel }^{2}-{r}^{2}\right)-J\left(x\right)$  for all $\left(x,\lambda \right)\in X×\left[L,+\infty \left[$  . Applying Theorem A, we get ${sup}_{\lambda \in \left[L,+\infty \left[}{inf}_{x\in X}\Phi \left(x,\lambda \right)={inf}_{x\in X}{sup}_{\lambda \in \left[L,+\infty \left[}\Phi \left(x,\lambda \right).$  Arguing as before (note, in particular, that ${lim}_{\lambda \to +\infty }{inf}_{x\in X}\Phi \left(x,\lambda \right)=-\infty$  ), we get $\stackrel{^}{x}\in X$  and $\stackrel{^}{\lambda }\in \left[L,+\infty \left[$  such that ${sup}_{\lambda \in \left[L,+\infty \left[}\Phi \left(\stackrel{^}{x},\lambda \right)={inf}_{x\in X}{sup}_{\lambda \in \left[L,+\infty \left[}\Phi \left(x,\lambda \right)$  and ${inf}_{x\in X}\Phi \left(x,\stackrel{^}{\lambda }\right)={sup}_{\lambda \in \left[L,+\infty \left[}{inf}_{x\in X}\Phi \left(x,\lambda \right).$  So that $\frac{\stackrel{^}{\lambda }}{2}\left(\parallel \stackrel{^}{x}-{x}_{0}{\parallel }^{2}-{r}^{2}\right)-J\left(\stackrel{^}{x}\right)={inf}_{x\in X}\frac{\stackrel{^}{\lambda }}{2}\left(\parallel x-{x}_{0}{\parallel }^{2}-{r}^{2}\right)-J\left(x\right)={sup}_{\lambda \in \left[L,+\infty \left[}\frac{\lambda }{2}\left(\parallel \stackrel{^}{x}-{x}_{0}{\parallel }^{2}-{r}^{2}\right)-J\left(\stackrel{^}{x}\right).$  From this it follows at once that $\parallel \stackrel{^}{x}-{x}_{0}{\parallel }^{2}\le {r}^{2}$  . But, if it were $\parallel \stackrel{^}{x}-{x}_{0}{\parallel }^{2}<{r}^{2}$  we would have $\stackrel{^}{\lambda }=L$  . This, in turn, would imply that $\stackrel{^}{x}\in {M}_{\frac{1}{L}}$  , against the fact that $r<{\beta }_{0}$  . Hence, we have $\parallel \stackrel{^}{x}-{x}_{0}{\parallel }^{2}={r}^{2}$  . Consequently $-\frac{1}{\stackrel{^}{\lambda }}J\left(\stackrel{^}{x}\right)={inf}_{x\in X}\frac{1}{2}\left(\parallel x-{x}_{0}{\parallel }^{2}-{r}^{2}\right)-\frac{1}{\stackrel{^}{\lambda }}J\left(x\right).$  This implies, on one hand, that $\stackrel{^}{\lambda }>L$  (since $r<{\beta }_{0}$  ) and, on the other hand, that each global maximum of the restriction of the functional $J$  to the set $\left\{x\in X:\parallel x-{x}_{0}\parallel =r\right\}$  is a global mimimum in $X$  of the functional $x\to \frac{1}{2}\parallel x-{x}_{0}{\parallel }^{2}-\frac{1}{\stackrel{^}{\lambda }}J\left(x\right)$  . Since $\stackrel{^}{\lambda }>L$  , this functional has a unique global minimum, and so $\left(ii\right)$  follows. $△$  REMARK 1. It is clear from the proof that the assumption ${J}^{\prime }\left({x}_{0}\right)\ne 0$  has been used to prove ${\alpha }_{0}>J\left({x}_{0}\right)$  and ${\beta }_{0}>0$  , while it has no role in showing $\left(i\right)$  and $\left(ii\right)$  . From the proof again, it also follows that, instead of ${J}^{\prime }\left({x}_{0}\right)\ne 0$  , one could require the less restrictive condition ${sup}_{x\in X\\left\{{x}_{0}\right\}}\frac{J\left(x\right)-J\left({x}_{0}\right)}{\parallel x-{x}_{0}{\parallel }^{2}}>\frac{L}{2}.$  However, without this latter condition, it can happen that ${\alpha }_{0}=J\left({x}_{0}\right)$  , ${\beta }_{0}=0$  , with $\left(i\right)$  (resp. $\left(ii\right)$  ) holding for no $r>{\alpha }_{0}$  (resp. for no $r>0$  ). To see this, take, for instance, $X=\mathbf{R}$  , $J\left(x\right)=\frac{1}{2}{x}^{2}$  , ${x}_{0}=0$  . References [1] B. RICCERI, Minimax theorems for limits of parametrized functions having at most one local minimum lying in a certain set, preprint.
Departement of Mathematics University of Catania Viale A. Doria 6 95125 Catania Italy e-mail address: ricceri@dmi.unict.it