Rectangle Norm Does not Change Under Relaxation

28 Views Asked by At

At the beginning of Section 4.1 of the paper Limits of Dense Graph Sequences by Lovasz and Szegedy, the following is stated:

Let $U:[0, 1]^2\to \mathbf R$ be an integrable function. Define the rectangle norm of $U$ as $$\|U\|_{\blacksquare} = \sup_{A, B\subseteq [0, 1]}\left|\int_A\int_BU(x, y)\ dx dy\right|$$ (where $A$ and $B$ are implicitly assumed to be measurable).

Then the authors say that it is easy to see that

$$\|U\|_{\blacksquare}= \sup_{f, g:[0, 1]\to [0, 1]} \left| \int_0^1\int_0^1 U(x, y)f(x)g(y)\ dx dy \right|$$

(where $f$ and $g$ are implicitly assumed to be measurable.) (There is a misprint in the paper. The absolute value symbole above is missing.)

Question. I don't see how the second equation is true. Can somebody please provide a proof. Thanks.

1

There are 1 best solutions below

1
On BEST ANSWER

Characteristic functions of measurable sets are measurable $[0,1]$-valued functions, this implies that $$\sup_{A,B\subseteq[0,1]}\left|\int_A\int_B U(x,y)\,dxdy\right|≤\sup_{f,g:[0,1]\to[0,1]}\left|\int_{[0,1]^2}U(x,y)f(x)g(y)\,dxdy\right|.$$ Now lets look at some specific $f,g:[0,1]\to[0,1]$ and show that we can make the integral larger by having these functions be only $\{0,1\}$-valued, ie by having them be characteristic functions. This would imply "$≥$".

Suppose for convenience that $\int_0^1\int_0^1 U(x,y)f(x)g(y)\,dxdy$ is positive. Look at the function $y\mapsto \int_0^1dx\, U(x,y)f(x)$ and let $\tilde g$ be the characteristic function of the set on which that function is positive (this is measurable). Since $g$ maps into $[0,1]$ we have: $$\tilde g(y)\int_0^1U(x,y)f(x)\,dx≥g(y)\int_0^1U(x,y)f(x)\,dx$$ and as such the integral can be made bigger by replacing $g$ with $\{0,1\}$-valued function. You can do the same now with $f$.