I’ve been working on an interesting high dimensional statistics problem 1 for which we needed to prove the usual Restricted Eigenvalue (RE) condition and stuck.
In a nutshell, RE condition helps to prove non-asymptotic consistency of an estimator. Meaning that using the RE condition one can show that for a finite number of samples \(n\) the estimation error of a parameter of interest is upper bounded by some number.2
There are a few methods for showing the RE condition among which Mendelson’s small ball method is very popular. Describing the method needs another post (and actually is not necessary because Joel Tropp has already done a great job explaining it here.).
I just wanted to mention that after stalling over the problem for a while I completely stopped thinking about it and worked on other things for a few weeks and when I re-examined it today (after some context switch warm up) I looked through the problem and quickly saw the answer!
The main point that I want to share with other knowledge-workers out there is that living with a problem seems desirable and one feels that he makes progress by continually thinking and pursuing a solution but unfortunately, this is not true in many situations for abstract problems. You need to relax or switch to another task for a while and let your diffuse mode^ take over and find a solution in the back of your mind. Then when you come back to the tough problem, you can see with the fresh eyes and hopefully come up with a novel line of attack.