I am reading a paper where the author derives the following Lagrangian dual problem :

$\min_v \int_R \frac{1}{4} \frac{\beta^2}{v-2\|x\|}dx+v\;\;\;\text{s.t.}\;\;\;v\geq 2\|x\|\;\;\;\forall x \in R$

from the primal problem :

$\max_{f(.)} \int_R (2\|x\| f(x) + \beta \sqrt{f(x)})dx\;\;\; \text{s.t.}\;\;\;\int_R f(x) dx= 1 \;\;\;\text{and}\;\;\;f(x) \geq 0\;\;\;\forall x \in R$

where $f(.)$ belongs to the Banach space $L^2$ over a compact set $R$ (a distribution function). Do you know how to construct a dual problem in case the objective and constraints of the primal include integrals. It was said that standard techniques of vector space optimization could be used to approach the function $f(.)$, but this is maybe not obvious. I could not pinpoint the starting point.


$\newcommand\R{\mathbb R}$ A convenient way to derive the dual problem from a primal one is by using the minimax duality for the Lagrangian, which is given here by the formula $$L(f,v):=\int_R\big[2|x|f(x)+b\sqrt{f(x)}\big]\,dx-v\Big(\int_Rf(x)\,dx-1\Big),$$ where $|x|:=\|x\|$ and $b:=\beta$. Clearly, $$\sup_{f\ge0}\inf_{v\in\R} L(f,v) =\sup\Big\{\int_R\big[2|x|f(x)+b\sqrt{f(x)}\big]\,dx\colon f\ge0,\int_R f(x)\,dx=1\Big\},$$ which is value of the primal problem.

The value of the dual problem is $$\inf_{v\in\R}\sup_{f\ge0} L(f,v) =\inf_{v\in\R}\Big(v+\sup\Big\{\int_R\big[2|x|f(x)+b\sqrt{f(x)}-vf(x)\big]\,dx\colon f\ge0\Big\}\Big) =\inf_{v\in\R}\Big(v+\int_R s(|x|,v)\,dx\Big), $$ where $$s(a,v):=\sup\{2at+b\sqrt t-vt\colon t\ge0\}.$$ For any real $a>0$, it is easy to see that $$s(a,v)=\frac{b^2}{4 (v-2 a)}$$ if $b>0$ and $v>2a$, $s(a,v)=0$ if $b\le0$ and $v\ge2a$, and $s(a,v)=\infty$ otherwise; in particular, $s(a,v)=\infty$ if $v<2a$. Thus, with $|R|:=\max\{|x|\colon x\in R\}$, the value of the dual problem is $$\inf_{v>2|R|}\Big(v+\int_R \frac{\max(0,b)^2}{4 (v-2|x|)}\,dx\Big). $$

  • $\begingroup$ The equality in the interchanging of $sup_{f}$ and $\int_R$ comes from the continuity and the measurability of the function $f(.)$, right. $\endgroup$ – Omar R. Sep 19 '20 at 19:52
  • $\begingroup$ @OmarR. : The inrterchange of $\sup_f$ and $\int_R$ is possible because, for $v>2a$, the maximizer of $2at+b\sqrt t-vt$ in $t\ge0$ is $\max(0,b)^2/(4(v-2a)^2)$, which is a measurable function of $a$ ($a$ here representing $|x|$). $\endgroup$ – Iosif Pinelis Sep 20 '20 at 1:06
  • $\begingroup$ Thank you very much. $\endgroup$ – Omar R. Sep 20 '20 at 10:17

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.