Skip to main content
Browse by:

Effective March 10, 2020, all Duke-sponsored events over 50 people have been cancelled, rescheduled, postponed or virtualized.
Please check with the event contact regarding event status. For more information, please see

Simple statistical tasks can be hard on average.

Event Image
Icon calendar
Monday, January 14, 2019
Icon time
12:00 pm - 1:00 pm
Icon speaker
Aukosh Jagannath (Harvard University)
Icon series
Mathematics Colloquium Seminar

Consider the problem of recovering a rank 1 tensor of order k that has been subject to Gaussian noise. We will begin by reviewing results surrounding the statistical limits of maximum likelihood estimation for this problem and discuss an geometric analogue of the well-known BBP phase transition from the matrix setting. We then discuss recent analyses of the behavior of this problem from an optimization perspective. While the threshold for estimation occurs at a finite signal-to-noise ratio, it is expected that one needs a polynomially diverging signal-to-noise ratio to be able to do so efficiently. We present a recent study of the thresholds for efficient recovery for a simple family of algorithms, Langevin dynamics and gradient descent, to better understand the mechanism for this diverging statistical-to-computational gap. I will report on recent works with Ben Arous-Gheissari on the algorithmic threshold and Lopatto-Miolane on the statistical threshold.