I'm going through the last steps of Bourgain and Demeter's proof of the $l^2$ decoupling conjecture, but I'm unable to see how the first inequality in (43) goes through. I'll water down the question a little to make it more transparent.

Let $f_1, \dots, f_n$ be $n$ functions whose Fourier transform is supported on $B(1)$. Bourgain claimed that by the heuristics he and Guth used in their paper we can think that $f_i$ is "essentially constant" on $B_1$. This seems to allow the following "reverse Holder" inequality

$$\prod_{i=1}^n \|f_i\|_{L^p(B_1)}^{1/n}\ll_{n,p} \|\prod_{i=1}^n f_i^{1/n}\|_{L^p(B_1)}. $$

But I skimmed that paper and only found an estimate morally like the following:

$$ f(x)=f*\eta(x)\le \|f\|_{L^1}\|\eta\|_{L^\infty}, $$

where $\hat\eta$ is compactly supported and equals 1 on $B_1$. How do I deduce the reverse Holder inequality from that?

Added: Just for reference, the inequality (43) that I referred to is

$$ \prod_{i=1}^n \|\widehat{g_id\sigma}\|_{p,\delta,B_{\delta^{-1/2}}}^{1/n}\ll_{n,p} \|\prod_{i=1}^n (\sum_{\theta\subset\tau_i:\delta^{1/2}\text{ cap}} |\widehat{g_{i,\theta}d\sigma}|^2)^{1/2n}\|_{L^p(w_{B_{\delta^{-1/2}}})}. $$

In the above, the $f_i$ are supposed to be $\widehat{g_{i,\theta}d\sigma}$, and I'm assuming $\tau_i=\text{supp }g_i$ contains only one $\delta^{1/2}$ cap, which is a slab of size $\delta^{1/2}\times\cdots\times\delta^{1/2}\times\delta$. $\tau_i$ are assumed to be $\nu-$transverse, but I feel this is unimportant in this inequality. $w_{B_{\delta^{-1/2}}}$ is just a smooth cutoff weight that equals 1 on $B_{\delta^{-1/2}}$. I have normalized $\delta$ to 1 in the above.

  • 4
    $\begingroup$ Could you remind us what B(1) and O(1) denote? $\endgroup$
    – Yemon Choi
    Commented Mar 2, 2016 at 23:12
  • $\begingroup$ The idea in all of these "essentially constant" arguments (commonly referred to as "the uncertainty principle," as popularized by Tao) is as follows: Assume that $f$ is a function on $R^d$ such that $\hat{f}$ is supported on a box B. Then the essentially (but not formally) true claim is that $f$ is essentially constant on translations of the polar of $B$, say $B_{0}$. In the $1$-d case if $B$ is an interval of length $L$ then $B_{0}$ is an interval of length $1/L$. $\endgroup$
    – Mark Lewko
    Commented Mar 3, 2016 at 1:24
  • $\begingroup$ Now what is formally true is that if $\hat{\eta}$ is a a smooth bump function equal to $1$ on $B$ and rapidly decaying away form B, then $f(x) = f * \eta$. In addition, it isn't hard to see that $\eta$ is now a function that is nearly constant on the $b_{0}$ and decays rapidly away from it. By nearly constant we mean bounded above and below by universal constants. $\endgroup$
    – Mark Lewko
    Commented Mar 3, 2016 at 1:25
  • $\begingroup$ Thus, recalling the definition of convolution, $f(x)$ is upper-bounded by essentially the average of $f(x)$ over a translate of $B_{0}$ centered at $x$. In just about any harmonic analysis estimate this observation leads to the same conclusion that would be obtained if $f$ was in fact a step function constant on a latices of boxes of the shape $B_{0}$. $\endgroup$
    – Mark Lewko
    Commented Mar 3, 2016 at 1:26
  • $\begingroup$ @MarkLewko This is basically what I got from the paper they cite, but it seems nontrivial to make it work with geometric means of a bunch of functions. $\endgroup$
    – Fan Zheng
    Commented Mar 3, 2016 at 1:40

1 Answer 1


I checked with Jean and Ciprian about this, and there is indeed a small issue here; the bound indicated is "morally" correct, and may possibly even be true (using the weights $w_B$ rather than a sharp truncation $1_B$), but it does not quite follow from the standard device of representing a Fourier-localised function by a convolution with a Schwartz function. In later papers, most notably in the Vinogradov conjecture paper with Guth, the issue was addressed by inserting an additional $L^p$ norm at small scales, replacing expressions such as $\| \widehat{g d\sigma} \|_{p,\delta,\Delta}$ with somewhat more complicated looking expressions $A_p(u,B^n,u)$. From what I understand, it is possible to adapt the arguments in these later papers (which are more streamlined than the original arguments in some respects) to reprove the hypersurface decoupling estimate rigorously. The authors plan to update the arXiv version of the original decoupling paper with some additional remarks addressing these issues.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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