Under the auspices of the Computational Complexity Foundation (CCF)

REPORTS > DETAIL:

### Paper:

TR10-072 | 19th April 2010 21:30

#### Constructive Proofs of Concentration Bounds

TR10-072
Authors: Russell Impagliazzo, Valentine Kabanets
Publication: 19th April 2010 21:30
Keywords:

Abstract:

We give a simple combinatorial proof of the Chernoff-Hoeffding concentration bound~\cite{Chernoff, Hof63}, which says that the sum of independent \$\{0,1\}\$-valued random variables is highly concentrated around the expected value. Unlike the standard proofs,
our proof does not use the method of higher moments, but rather uses a simple and intuitive counting argument. In addition, our proof is constructive in the following sense: if the sum of the given random variables is not concentrated around the expectation, then we can efficiently find (with high probability) a subset of the random variables that are statistically dependent. As simple corollaries, we also get the concentration bounds for
\$[0,1]\$-valued random variables and Azuma's inequality for martingales~\cite{Azuma}.

We interpret the Chernoff-Hoeffding bound as a statement about Direct Product Theorems. Informally, a Direct Product Theorem says that the complexity of solving all \$k\$ instances of a hard problem
increases exponentially with \$k\$; a Threshold Direct Product Theorem says that it is exponentially hard in \$k\$ to solve even a significant fraction of the given \$k\$ instances of a hard problem. We show the equivalence between optimal Direct Product Theorems and optimal Threshold Direct Product Theorems. As an application of this connection, we get the Chernoff bound for expander walks~\cite{Gil98} from the (simpler to prove) hitting
property~\cite{AKS87}, as well as an optimal (in a certain range of parameters) Threshold Direct Product Theorem for weakly verifiable puzzles from the optimal Direct Product Theorem~\cite{CHS05}. We
also get a simple constructive proof of Unger's result~\cite{Ung09} saying that XOR Lemmas imply Threshold Direct Product Theorems.

ISSN 1433-8092 | Imprint