3
$\begingroup$

I want to compute a Phase Transition Diagram as shown here ($A \in \mathbb{R}^{n \times N}$ and $k$ is the sparsity: $\vert \vert x \vert \vert_0 = k $ )

My question is: For $n=1$ I can only compute $k=1$, for $n=2$ I can only compute $k=1,2$ and so on. Therefore the computable number of $\rho$ values is different for each $\delta$ value, but in the plots it is even for $\delta=0$ a continuous, changing plot into the y-direction. My plots are currently looking like this (yellow means reconstruction with probability 1):

PTD

$\endgroup$
  • $\begingroup$ The parameters in your question seem inconsistent: your $A \in \mathbb R^{n \times p}$, but after this you talk about $m$. That is, you effectively have four parameters $n$, $p$, $k$, and $m$. There should only be three parameters; in Donoho & Tanner they call them $N$, $n$, and $k$. From your specification of $A$, your $n$ and $p$ correspond to $n$ and $N$ respectively. $\delta = n/N$ and $\rho = k/n$. Once we get this clarified, I think I can explain how to achieve what you want. $\endgroup$ – Thomas Arildsen Jul 27 '17 at 7:31
  • $\begingroup$ I hope it is clear now. $\endgroup$ – N8_Coder Jul 28 '17 at 8:16
  • $\begingroup$ If not please let me know what is confusing you $\endgroup$ – N8_Coder Jul 31 '17 at 9:55
  • $\begingroup$ I will try to make time to answer later today $\endgroup$ – Thomas Arildsen Jul 31 '17 at 9:56
  • $\begingroup$ I feel really bad to bother you, but would be glad to get an answer. Thanks in advance! $\endgroup$ – N8_Coder Aug 2 '17 at 14:39
1
$\begingroup$

I recommend first fixing $N$. By the way, remember here that compressed sensing works asymptotically well; for $N \rightarrow \infty$ your phase transition is going to move up/left (better) in the $\delta, \rho$ diagram, see for example Donoho & Tannner 2010.

Your size of $N$ essentially determines the resolution you can achieve in your phase transition diagram. For example we can look at Maleki & Donoho 2010, Section IV: they describe how they select $N = 800$ and then calculate $n$ and $k$ from a grid of values $(\delta, \rho)$ between 0.05 and 1 in 30 steps. For each value $\delta$ or $\rho$, the corresponding $n$ or $k$ are calculated as $$n = \lceil \delta N\rceil,$$ respectively $$k = \lceil \rho n \rceil.$$

As you can see there is generally some imprecision involved during to the rounding. If you select a too fine grid of values $(\delta, \rho)$ for a given $N$, several of the resulting $n$ or $k$ values are going to round to the same number. You will have to either select a larger $N$ or choose a coarser grid for your $(\delta, \rho)$ to avoid this.

Finally, you should (if you want the same type of plot as Donoho et al.) plot your success ratio in each of your points in a $(\delta, \rho)$ coordinate system. Your figure above looks as if it was perhaps plotted with $k$ on the first axis?

Now, if you wish to estimate the actual phase transition boundary, you can also see in Maleki & Donoho 2010, Section IV, how they fit a logistic curve to the success ratios and can then read a specific percentage reconstruction boundary $\rho$ for each value $\delta$ from the fitted logistic function.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.