Major problem with Cholesky decomposition

Hi all,

I'm having major issues with the chol command. A code I need to run uses Cholesky decomposition and it crashes right at the point where it needs to manipulate a matrix.

Even worse, Matlab yields the following nonsense (see below):

Note that this example comes straight from the chol help file. Any idea why this is happening? I'm using version 2013b. Thanks!

[Walter Roberson]

I wonder... would you happen to be using an AMD Jaguar CPU ?

[Mario Solis-Garcia]

No, it's an Intel Core I7.

[dpb]

Earliest I have installed is R2014b altho I'd be extremely surprised if it were a generic bug.

Can you either upgrade or try re-install?

[Mario Solis-Garcia]

Tried re-installing from scratch, no success. I even erased my computer, fresh reinstall, no success. It's driving me nuts.

[Walter Roberson]

I just installed R2013b into a virtual machine to test with, but had no difficulty.

What shows up for

>> which chol(X)
built-in (C:\Program Files\MATLAB\R2013b\toolbox\matlab\matfun\@double\chol)  % double method

My virtual machine is Intel Core i7-4771

[Mario Solis-Garcia]

Here's the output:

[dpb]

Mario, no chance you can update to a different version altho this is sounding like a unique system-related issue of some short...I think I'd submit a SRQ to TMW official site and see if they've got any ideas...think there's nothing else we can do that can help you.

[Christine Tobler]

This sounds like a weird system-dependent issue. I agree that creating a service request would be best.

Some things that could be interesting to try (not likely to solve the issue, though):

 chol(eye(4)); % Does this also error?
 [R, p] = chol(pascal(5)) % This will tell us at which row of the input
                          % matrix CHOL decides that the matrix is not SPD.

Otherwise, I'd think the next steps would be to try to narrow down with which MATLAB versions and which settings you're seeing these issues. Tech Support will be able to help with this.

[Mario Solis-Garcia]

@Christine: thanks for your suggestion. I'm attaching a screenshot with the output. (Shocker: it still doesn't work.)

@dbp, @Christine: Tech support is off limits since my softer maintenance service is expired. I guess I'll just bite the bullet and upgrade to 2018a.

[dpb]

Too bad...since it's builtin pretty clear there's a interaction issue of some sort going on.

Appears only choice other than an upgrade would be to try reversion of any OS upgrades or to upgrade to see if that would somehow magically make the symptom disappear.

Do you have previous knowledge it did work at one time on the specific machine?

[Christine Tobler]

That's too bad. And the "out of memory" message is weird on a whole different level: The only difference between the one-output and two-output calls is that one scalar number is allocated for p. So there's something seriously bad going on with this installation.

I haven't heard of any problems with R2013b and i7 CPUs, but upgrading MATLAB / changing your OS configuation seems like a reasonable idea.

Is Cholesky the only function you're seeing problems with? I'd expect all linear algebra operations / all operations in general to have problems, given how badly Cholesky is doing here.

Answer 1

[Jan]

Either this is a problem with the called libraries or a user defined function shadows the original one, but contains a bug. Please run:

profile clear
profile('on', '-detail', 'builtin')
chol(pascal(5))

profile report

Do you see any function, which differs from "(built-in function)"?

[Mario Solis-Garcia]

The syntax doesn't run as-is in my computer (maybe it's because I'm running 2013b?), but here's what I believe is an equivalent (see attached).

(Edit: improved quality of attached picture.)

[Jan]

@Mario: "doesn't run" does not allow to understand, what's going wrong. Only showing the detail-level would reveal, which tools are called internally. Do you get an error message?

[Walter Roberson]

https://www.mathworks.com/help/releases/R2013b/matlab/ref/profile.html

-detail is not a documented option for profile in R2013b through R2018a

[Philip Borghesani]

-detail is not documented but should do something other then error. If you pasted the code I suggest typing it by hand. One irritating problem is that pasted code frequently has a different character then '-' causing unexpected errors.

[Philip Borghesani]

R2013b is fairly old to run on a current MAC and version of Mac OS X. What type of Mac are you using and what version of Mac OS X?

[Walter Roberson]

Good catch about it being Mac.

I just tried R2013b installed on my Mac, and do not have this difficulty.

If I recall correctly, the first Mac with Core i7 available were late 2012 MacBook Pro. But somewhere I have an app that has a complete catalog, so I will see if I can check.

[Walter Roberson]

... I was close. There was a Core i7 option for MacBook Pro from mid 2010 and for Mac Mini from mid 2011.

[Mario Solis-Garcia]

@Jan, @Philip: you were right, I blame the copy+paste. Output is attached, hope you find it more illuminating than me.

@Philip: it's a 2015 iMac, kept using 2013b after I forgot to renew the maintenance service -- 2013b worked just fine for what I needed -- but about a month ago the software just went nuts. I guess I can pay to get 2018a (I'm currently using the trial version and everything works as advertised) but I'm still puzzled about why this is happening in the first place.