Computational Complexity (cs.CC)

  • PDF
    We consider relative error low rank approximation of \it tensors with respect to the Frobenius norm: given an order-$q$ tensor $A \in \mathbb{R}^{\prod_{i=1}^q n_i}$, output a rank-$k$ tensor $B$ for which $\|A-B\|_F^2 \leq (1+\epsilon)$OPT, where OPT $= \inf_{\textrm{rank-}k~A'} \|A-A'\|_F^2$. Despite the success on obtaining relative error low rank approximations for matrices, no such results were known for tensors. One structural issue is that there may be no rank-$k$ tensor $A_k$ achieving the above infinum. Another, computational issue, is that an efficient relative error low rank approximation algorithm for tensors would allow one to compute the rank of a tensor, which is NP-hard. We bypass these issues via (1) bicriteria and (2) parameterized complexity solutions: (1) We give an algorithm which outputs a rank $k' = O((k/\epsilon)^{q-1})$ tensor $B$ for which $\|A-B\|_F^2 \leq (1+\epsilon)$OPT in $nnz(A) + n \cdot \textrm{poly}(k/\epsilon)$ time in the real RAM model. Here $nnz(A)$ is the number of non-zero entries in $A$. (2) We give an algorithm for any $\delta >0$ which outputs a rank $k$ tensor $B$ for which $\|A-B\|_F^2 \leq (1+\epsilon)$OPT and runs in $ ( nnz(A) + n \cdot \textrm{poly}(k/\epsilon) + \exp(k^2/\epsilon) ) \cdot n^\delta$ time in the unit cost RAM model. For outputting a rank-$k$ tensor, or even a bicriteria solution with rank-$Ck$ for a certain constant $C > 1$, we show a $2^{\Omega(k^{1-o(1)})}$ time lower bound under the Exponential Time Hypothesis. Our results give the first relative error low rank approximations for tensors for a large number of robust error measures for which nothing was known, as well as column row and tube subset selection. We also obtain new results for matrices, such as $nnz(A)$-time CUR decompositions, improving previous $nnz(A)\log n$-time algorithms, which may be of independent interest.

Recent comments

Robin Blume-Kohout Apr 07 2017 20:30 UTC

Zak, David: thanks! So (I think) this is a relation problem, not a decision problem (or even a partial function). Which is fine -- I'm happier with relation problems than with sampling problems, and the quantum part of Shor's algorithm is solving a relation problem, which is a pretty good pedigre

...(continued)
David Gosset Apr 06 2017 20:11 UTC

Thanks Zak, that's exactly right-- for each instance there is a set of possible solutions. Like in the Bernstein-Vazirani problem, a solution is a bit string. It can't just be a single bit since then we would have the problem you describe, Robin.

Zak Webb Apr 06 2017 17:15 UTC

You are completely correct that in order to check whether a give output is "correct" for the input, we would require an additional log-depth classical circuit, but this is not how the problem is defined. In particular, for each input there is a set of "accepting" outputs, and we only need to guaran

...(continued)
Robin Blume-Kohout Apr 06 2017 15:05 UTC

Is it okay to be a quantum supremacist? I thought I was, but maybe if it's "tainted" I should reconsider.

On a more serious note... a question for somebody who has read (or written) the paper. If the computation is performed on Poly(n) qubits, and all of them are relevant, and you are only allo

...(continued)
Steve Flammia Apr 04 2017 13:13 UTC

I would like to publicly thank the authors for using the term "advantage" instead of the tainted word "supremacy" that makes me cringe every time I hear it.

Also, great result!

Ashley Apr 04 2017 08:35 UTC

A provable separation between analogous quantum and classical circuit classes!

Māris Ozols Feb 21 2017 15:35 UTC

I'm wondering if this result could have any interesting consequences for Hamiltonian complexity. The LCL problem sounds very much like a local Hamiltonian problem, with the run-time of an LCL algorithm corresponding to the range of local interactions in the Hamiltonian.

Maybe one caveat is that thi

...(continued)
Jānis Iraids Jan 25 2017 11:35 UTC

You are correct, that is a mistake -- it should be $\\{0,1\\}^n\rightarrow\\{0,1\\}$. Thank you for spotting it!

Christopher Thomas Chubb Jan 25 2017 02:27 UTC

In the abstract, should the domain of $f$ be $\lbrace0,1\rbrace^n$ instead of just $\lbrace0,1\rbrace$?

Zoltán Zimborás Jan 12 2017 20:38 UTC

Here is a nice description, with additional links, about the importance of this work if it turns out to be flawless (thanks a lot to Martin Schwarz for this link): [dichotomy conjecture][1].

[1]: http://processalgebra.blogspot.com/2017/01/has-feder-vardi-dichotomy-conjecture.html