[ncl-talk] Canonical Correlation Analysis via Barnett and Preisendorfer (BPCCA)

[Xiaoning Wu]

Thanks for clarifying! But it is surprising that I couldn't find anything
so far on the  Barnett and Preisendorfer CCA. I hope it is not because of
some technical intricacies that I'm currently not aware of ⊙﹏⊙
Nevertheless, extra gain if we can get cancor to work after all! Again,
thank you for looking into this issue.

Cheers,
Xiaoning

On Mon, Nov 13, 2017 at 10:28 AM, Dennis Shea <shea at ucar.edu> wrote:

> *I have no comment on the Barnett & Priesendorfer canonical correlation.*
> *Somebody else will have to comment.*
>
> ===
> re: "... the built-in “cancor” (which fails often) ..."
>
> This is true. There is an issue with NCL's 'cancor function.
> In fact, there is a 'trouble ticket' on this from **2010**:
>       JIRA 589: Possible bug in CANCOR
>
> My recollection is that It works correctly for "smaller" array sizes but
> fails when some critical size is reached.
> Obviously,  investigating this issue has been put on the back-burner.
> -----
>
> FYI: NCL's canonical correlation function uses a 'double precision'
> version of the f77 'SUBROUTINE CANOR' from the IBM Scientific Software
> Package [SPP].scientific library.
>
>   http://pdp-10.trailing-edge.com/decuslib20-02/01/decus/20-00
> 26/canor.ssp.html
>
> ----
> Actually, NCL has two hidden 'interfaces' between the NCL code and the f77
> subroutine. :
>
>   (a) a f77subroutine (written by me)
>   (b) a C-language interface between the NCL code and the fortran
> interface code (a)
> ----
> **Maybe**, I will take a look
>
> Cheers
>
>
>
>
>
>
>
> On Sun, Nov 12, 2017 at 12:53 PM, Xiaoning Wu <
> xiaoning.wu.1 at stonybrook.edu> wrote:
>
>> Hello,
>>
>> On a statistical application of NCL -  has anyone tried coding up a
>> script to perform the Barnett and Preisendorfer approach of Canonical
>> Correlation Analysis (BPCCA) between two datasets?
>>
>> Compared to the built-in “cancor” (which fails often), the trick with
>> BPCCA is to  decompose the data with PCA/EOF first, and the subsequent
>> linear algebra would involve the dominant modes only, thus bypassing direct
>> inversion of the original (large) matrices which is presumably where the
>> classical approach usually gets stuck. The method looks pretty manageable,
>> with existing matlab scripts to refer to, but I'm trying to avoid
>> reinventing the wheel here assuming it must have been done before in NCL.
>> Any thoughts would be very much appreciated!
>>
>> Thank you for your time. Have a great week!
>>
>> Best regards,
>> Xiaoning
>>
>> _______________________________________________
>> ncl-talk mailing list
>> ncl-talk at ucar.edu
>> List instructions, subscriber options, unsubscribe:
>> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.ucar.edu/pipermail/ncl-talk/attachments/20171114/1c2f0e40/attachment.html>