In this note, we prove the seminal result that : All statements that have statistical zero knowledge proofs reduce to the statistical distance problem1. To do so, we begin by defining what the statistical distance (SD) problem even is. And in order to describe (SD), we first need to describe what we mean by distributions encoded by circuits. Then the major technical results can be found by clicking the pick tiles in each section.
This reduction only works for distributions encoded by circuits. If we did not have circuits but instead drew samples using some black box oracle, then it is an open problem to show that this reduction, still holds.
Let X be a boolean circuit (with possibly unbounded fan in) with m input gates and n output gates. The distribution encoded by X is a probability distribution on the set \{ 0, 1\}^n, induced by feeding X with inputs sampled uniformly randomly from \{ 0, 1\}^m. We will often abuse notation by referring to X directly as the distribution instead of the circuit for the remainder of this document.
When we sample from distributions this way – we will often say that we have white-box access to the distribution.
For constants 0 \leq \beta < \alpha \leq 1, Statistical Difference (\mathsf{SD}^{\alpha, \beta}) is a the promise problem \mathsf{SD}^{\alpha, \beta}_\mathsf{Yes}, \mathsf{SD}^{\alpha, \beta}_\mathsf{No}, where
\mathsf{SD}^{\alpha, \beta}_\mathsf{Yes}= \{(X,Y) : d_\mathtt{TV} (X, Y) > \alpha \} \mathsf{SD}^{\alpha, \beta}_\mathsf{No}= \{(X,Y) : d_\mathtt{TV} (X, Y) < \beta \}
where X and Y are distributions encoded by circuits.