# Some not-so-recent ArXiV skims

I tend to flag papers on ArXiV that I want to take a look at in (soon to be defunct, *sniff*) Google Reader. Here are some papers from the last month that I found interesting. I’ll post a few more of these as I work through my backlog…

Local Privacy and Statistical Minimax Rates (John C. Duchi, Michael I. Jordan, Martin J. Wainwright) — this is a paper proving minimax lower bounds for differential privacy. The approach is based on the Fano/Le Cam style of getting minimax bounds by constructing a packing of instances of the problem.

Bernstein – von Mises Theorem for growing parameter dimension (Vladimir Spokoiny) — I’m generally interested in the consistency properties of Bayesian procedures, and this looks at the effect of asymptotically growing the problem size to see how fast the problem can grow while still getting the same consistency from the BvM theorem.

On the problem of reversibility of the entropy power inequality (Sergey G. Bobkov, Mokshay M. Madiman) — More results on the EPI. Reversing it is the same as reversing the Brunn-Minkowski inequality (consider uniform distributions), but there is an interesting impossibility result here (Theorem 1.3): “For any constant $C$, there is a convex probability distribution $\mu$ on the real line with a finite entropy, such that $\min \{ H(X+Y), H(X-Y) \} \ge C H(X)$, where $X$ and $Y$ are independent random variables, distributed according to $\mu$.” The distribution they use is a truncated Pareto distribution but the calculations seem hairy.

A universal, operational theory of unicast multi-user communication with fidelity criteria (Mukul Agarwal, Sanjoy Mitter, Anant Sahai) — This is the culmination of Mukul’s work starting from a very nice paper I cite all the time from Allerton. There are several results and commentary in here — there’s a fair bit of philosophy, so it’s worth a more patient read than I could give it so far (only so many hours in the day, after all!)

The Convergence Rate of Majority Vote under Exchangeability (Miles E. Lopes) — The title says it all, really. The bounds are actually in terms of the mixture distribution of the exchangeable sequence of Bernoulli votes.

/