[Octave-bug-tracker] [bug #65032] [Octave Forge] (Symbolic) @sym/qr outputs q1 of q=[q1 q2]

[Liang Tang]

URL:
  <https://savannah.gnu.org/bugs/?65032>

                 Summary: [Octave Forge] (Symbolic) @sym/qr outputs q1 of
q=[q1 q2]
                   Group: GNU Octave
               Submitter: lt1234
               Submitted: Thu 14 Dec 2023 01:51:27 PM UTC
                Category: Octave Forge Package
                Severity: 3 - Normal
                Priority: 5 - Normal
              Item Group: Matlab Compatibility
                  Status: None
             Assigned to: None
         Originator Name: lt1234
        Originator Email: 
             Open/Closed: Open
                 Release: 8.2.0
         Discussion Lock: Any
        Operating System: Microsoft Windows
           Fixed Release: None
         Planned Release: None


    _______________________________________________________

Follow-up Comments:


-------------------------------------------------------
Date: Thu 14 Dec 2023 01:51:27 PM UTC By: Liang Tang <lt1234>
Hi, 

There is no Octave compatibility option for me to pick for reporting. 
Therefore I picked the matlab compatibility instead.  

@sym/qr clearly documents the differences of @sym/qr compared to Octave/qr. [q
r]=@sym/qr(A) behaves the same as Octave [q r] = qr(A, 'econ').  There is no
input argument for @sym/qr to force @sym/qr to output the full q=[q1 q2]. 
Effectively, for numeric variables, @sym/qr output is a subset of Octave qr. 


By reading the documentation, matlab qr and matlab symbolic qr appear to be
equivalent in syntax. It is also known that Octave/qr and Matlab/qr are
compatible. That leaves @sym/qr inconsistent in definition and output, i.e.,
same inputs but different (or subset of) results. 

@sym/qr returns the q equivalent to the q of qr with 'econ' argument. There is
no option to output the full q matrix.  
sympy 1.12 documentation page 1330 states the same limitation.  So the 'econ'
limitation is at sympy, not the symbolic package.  

sympy recommends post process the symbolic q1 of the q=[q1 q2] for the full q.
 There are three items to note: 

(1) sympy recommends Gram-Schmidt but I can not identify the availability of
symbolic Gram-Schmidt in the symbolic package.  I think Gram-Schmidt is a
method of sympy matrix object, 
similar to a method of an Octave class.  Since I have not found any mean to
access the GS method in @sym, with a m file, only numeric values from @sym/qr
can be processed.  
(2) The null space of the input matrix A to qr is not unique.  Therefore the
q2 of q=[q1 q2] from Octave/qr may not be reproduced.  In addition, I think
both Octave and Matlab qr are built-in.  It is unclear to me for what a user
can do.  
(3) Knowing the matrix null space is not unique, we also find null of @sym and
null of Octave produce different results.  I may be able to replace Octave q2
with Octave null(A).  But the same cannot be done for @sym.  Mathematically
both codes are correct but output values are different.   It is about
compatibility, not incorrect results.    
       
I would like you to consider the following.  Thanks.    

(A) @sym/syntax: delete this option  -- Method on @sym: [Q, R] = qr (M).  
(B) @sym/syntax: modify this option  -- Method on @sym: R = qr (M), replaced
with  @sym: R = qr (M, 0)
(C) Documentation:  If you agree with the deletion/modification, please update
the text accordingly.   Give workaround for the full q if possible. 
(D) Investigate the feasibility to implement the @sym/qr [Q, R] = qr (M)
according to the definition of Octave [Q, R] = qr (M).


Example:  

% sym qr 

ver('symbolic')
pkg load symbolic 

% @sym/qr outputs q1  

A=sym('A'); A=[sym(1/sqrt(2)) 1/sqrt(2) 0]';
[q_s, r_s]=qr(A);

double(q_s), double(q_s)*double(q_s)',  double(q_s)'*double(q_s)

% Octave qr (full q) 

[q r ]= qr([1/sqrt(2) 1/sqrt(2) 0]');

q, q*q'  

% Octave qr + econ = @sym/qr 

[q_e r_e ]= qr([1/sqrt(2) 1/sqrt(2) 0]', 'econ');

q_e, q_e*q_e',  q_e'*q_e

%  @sym q1 + modified GS, per sympy

[q_G, r_G] = mgson([q(:,1)'; 1 2 3; 4 5 6]') ; % mgson (modofied GS)
downloaded from github for demo purpose 

%{ 
q_G = ...
  [-0.7071  -0.1622   0.6882;
   -0.7071   0.1622  -0.6882;
         0   0.9733   0.2294];
%}

q_G, q_G*q_G'

% this shows the difference for null (same null space uniqueness discussion) 

[q r]=qr([1 2 3]'); q(:, 2:end), null([1 2 3]), null([sym('1') sym('2')
sym('3')])
 

    








    _______________________________________________________

Reply to this item at:

  <https://savannah.gnu.org/bugs/?65032>

_______________________________________________
Message sent via Savannah
https://savannah.gnu.org/