Closed-form conjecture for the rectangular optimal shrinker

Let Υγ,β(y)\Upsilon_{\gamma,\beta}(y) be the shrinker defined in equation (Upsilon), and let zγ,β+\mathsf{z}^{+}_{\gamma,\beta} and zγ,β\mathsf{z}^{-}_{\gamma,\beta} be the associated spectral-edge quantities. Closed-form conjecture for the rectangular optimal shrinker. For all γ(0,)\gamma\in (0,\infty) and β(0,1)\beta\in (0,1) with γβ<1\gamma\beta<1, the shrinker Υγ,β()\Upsilon_{\gamma,\beta}(\cdot) has the closed-form expression

Υγ,β(y)={1y(y2zγ,β+)(y2zγ,β)if y>zγ,β+,0if yzγ,β+.\Upsilon_{\gamma,\beta}(y) = \begin{cases} \frac{1}{y}\sqrt{(y^2-\mathsf{z}^+_{\gamma,\beta})(y^2-\mathsf{z}^-_{\gamma,\beta})}&\quad\textrm{if } y>\sqrt{\mathsf{z}^+_{\gamma,\beta}},\\ 0 &\quad\textrm{if } y\le \sqrt{\mathsf{z}^+_{\gamma,\beta}} \end{cases}.

This would extend the known explicit optimal shrinker for the full SVD to the rectangular setting; the source reports exhaustive numerical verification, but does not provide a proof.

Sources & referencesView supporting material

Primary source

Elad Romanov, “On the Noise Sensitivity of the Randomized SVD”, arXiv:2305.17435 (2023).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.