1: \begin{abstract}

2: In this paper we consider the problem of estimating the \emph{$f$-moment} ($\sum_{v\in [n]} (f(\mathbf{x}(v))-f(0))$) of a dynamic vector $\mathbf{x}\in \G^n$ over some abelian group $(\G,+)$, where the $\|f\|_\infty$ norm is bounded.  We propose a simple sketch and new estimation framework based on the \emph{Fourier transform} of $f$. I.e., we decompose $f$ into a linear combination of homomorphisms $f_1,f_2,\ldots$ from $(\G,+)$ to $(\C,\times)$,

3: estimate the $f_k$-moment for each $f_k$, and synthesize them to obtain an estimate of the 

4: $f$-moment.

5: Our estimators are asymptotically unbiased and have variance asymptotic to $\|\mathbf{x}\|_0^2 (\|f\|_\infty^2 m^{-1} + \|\hat{f}\|_1^2 m^{-2})$, where the size of the sketch is $O(m\log n\log|\G|)$ bits.

6: 

7: 

8: When $\G=\Z$ this problem can also be solved using off-the-shelf $\ell_0$-samplers with space $O(m\log^2 n)$ bits, which does not obviously generalize to finite groups.  As a concrete benchmark, we extend \cite{ganguly2003processing}'s singleton-detector-based sampler to work over $\G$ using $O(m\log n\log|\G|\log(m\log n))$ bits.

9: 

10: We give some experimental evidence that the Fourier-based estimation framework is significantly more accurate than sampling-based approaches at the same memory footprint.

11: \end{abstract}

12: