The main file in this package, rsparse.m, creates a sparse array object that is a subclass of MATLAB's native sparse data type, but which is robust against certain bugs and errors (present as of R2009b).
The I/O syntax and options for rsparse() are identical to that of sparse().
Where sparse() would return a double sparse matrix, rsparse() will return a matrix of class RobustSparseDouble, a class defined in this package. Similarly, wherever sparse() would return a logical sparse matrix, rsparse() will return a matrix of user-defined class RobustSparseLogical.
EXAMPLE 1: Indexing robustness - this utilizes the sparse-sub-access package by Bruno Luong.
>> A=sparse(100000,100000,1), %native MATLAB sparse matrix
>> B=rsparse(A), %convert to robust type
>> whos A B
Name Size Bytes Class Attributes
A 100000x100000 400016 double sparse
B 100000x100000 400072 RobustSparseDouble
>> A(logical(A))=2, %Wrong result due to indexing bug
>> B(logical(B))=2, %RIGHT!!!
EXAMPLE 2: Robustness to mixed data type operations.
For some reason, MATLAB decides to issue an error when a mathematical operation involving sparse and non-double numeric data is attempted:
>> y=A*x, %ANNOYING....
??? Error using ==> mtimes
Sparse single arrays are not supported.
The RobustSparse types, however, will silently pre-convert the foreign data to doubles:
I'm hoping that TMW will fix these problems soon, rendering this package obsolete. Once they do, simply replace the code in rsparse.m with the following, and rsparse() will become equivalent to sparse()
The package also includes "robustified" versions of other common sparfun toolbox routines: rspeye, rsprand, rsprandn, rprandsym,... They too will revert to their native MATLAB versions when the above replacement is made.
(1) Because this package uses various M-coded wrappers for built-in sparse routines, there is obviously some compromise in speed.
(2) I've overloaded all math operators (+ , - , .* , >=, <=, etc...) and some common functions like sum(), inv(), spfun(), etc... to return RobustSparse data types, where they would otherwise return ordinary sparse types.
For obvious reasons, however, I cannot do this for all functions out there capable of returning a sparse matrix (e.g., triu, tril, sin, cos, etc ... ), nor can I keep up with TMW's releases of new functions.
Therefore, you should bear in mind that B=triu(rsparse(A)) and similar will return an ordinary native sparse data type, B, with all of its vulnerabilities. You should take care to post convert such results to robust type, via B=rsparse(B). Or, you can add overloaded methods as you see fit, mimicking the others in RobustSparseDouble.m and RobustSparseLogical.m