SCM

SCM Repository

[matrix] Diff of /pkg/TODO
ViewVC logotype

Diff of /pkg/TODO

Parent Directory Parent Directory | Revision Log Revision Log | View Patch Patch

revision 530, Mon Feb 7 10:58:27 2005 UTC revision 2207, Mon Jul 7 22:34:52 2008 UTC
# Line 1  Line 1 
1    - Check for DimNames propagation in coercion and other operations.
2    
3  - Report the problem in the Linux ldexp manual page.  The second and  - Report the problem in the Linux ldexp manual page.  The second and
4    third calls in the Synopsis should be to ldexpf and ldexpl.    third calls in the Synopsis should be to ldexpf and ldexpl.
5    
6  - [,] indexing  - provide methods for "dspMatrix" and "dppMatrix"!
   
 - group generics: "Arith", but also "Ops" and "Math";  see  
   ?Math.data.frame  and the examples in  
   ?SetGeneric  
   
 - bCrosstab(): do we really want the diagonal "V:V" crosstabs?  
                (if so or in any case: add to  man/bCrosstab.Rd )  
   
 - src/Metis/ : one of the two Makefiles needs fixing, as changing  
              src/Metis/*.c  does not lead to recompilation.  
   --DB - it seems both Makefiles need fixing.  I think I have the  
              src/Metis/Makefile fixed but not src/Makefile  
   
 - man/Matrix.Rd :  has example with dimnames, but we just drop them!  
                 MM thinks dimnames should be supported (but then ...)  
   
 - data/ : names 'mm' and even more 'y' are  ``too short''.  
         If we really want to keep them, don't use "LazyData"  
         (such that one needs  data(*) explicitly);  
         But MM would rather want something like  ex.mm and ex.y  
   
 - "factors" maybe should move up to "Matrix"  
   
 -------  
   
 We have a (at least one) basic problem :  
    Currently the show() method fail sometime after coercion:  
    e.g. 'sy' show()s wrongly, even though it "str()" fine :  
   
       (po <- crossprod(Matrix(0:3, 2))) # ok  
       (ge <- as(po, "dgeMatrix")) # ok  
       (sy <- as(po, "dsyMatrix")) # BAD  
       str(sy) # looks fine  
   
   or  
       example(expand) # -> ex$L and ex$U look bad, however  
       as(ex$L, "dgeMatrix") # `works'  
7    
8    {Of course, we don't need a workaround but must understand  - implement (more) methods for supporting "packed" (symmetric / triangular)
9     and solve the problem}    matrices; particularly something like pack() and unpack()  [to/from our
10      classes from/to "numeric"] --- have already man/unpack.Rd but no method yet!
11    
12      (have some dtr* <-> dtp*)
13    
14    - combine the C functions for multiplication by special forms and
15      solution wrt special forms by using a 'right' argument and a
16      'classed' argument.
17       [done with dgeMatrix_matrix_mm();  not yet for other classes;
18        and for _crossprod()]
19    
20    -----
21    
22    - "Math2" , "Math", "Arith":
23       keep triangular and symmetric Matrices when appropriate:
24       particularly desirable for  "Math2": round(), signif()
25    
26      For triangular matrices, more specifically make sure the four rules of
27      "triangular matrix algebra" (Golub+Van Loan 1996, 3.1.8, p.93) are
28      fulfilled; now(2008-03-06) ok for Csparse; not yet for <dtr> %*% <dtr>
29    
30    - "d" <-> "l" coercion for all "[TCR]" sparse matrices is really trivial:
31      "d" -> "l" : drops the 'x' slot
32      "l" -> "d" : construct an 'x' slot of all '1'
33      We currently have many of these conversions explicitly, e.g.
34       setAs("dsTMatrix", "lsTMatrix",
35          function(from) new("lsTMatrix", i = from@i, j = from@j, uplo = from@uplo,
36                             Dim = from@Dim, Dimnames = from@Dimnames))
37      but I would rather want to automatically construct all these coercion
38      methods at once by a ``method constructor'', i.e.,
39      for all  "dsparse*" -> "lsparse*" and vice versa.
40      How can one do this {in a documented way} ?
41    
42    - Think of constructing  setAs(...) calls automatically in order to
43      basically enable all ``sensible'' as(fromMatrix, toMatrix)  calls,
44      possibly using canCoerce(.)
45    
46    - setAs(<Mcl>,  "[dln]Matrix") for <Mcl> in {Matrix or denseMatrix + sparseMatrix}
47    
48    - When we have a packed matrix, it's a waste to go through "full" to "sparse":
49      ==> implement
50            setAs("dspMatrix", "sparseMatrix")
51            setAs("dppMatrix", "sparseMatrix")
52            setAs("dtpMatrix", "sparseMatrix")
53      and the same for "lsp" , "ltp"  and  "nsp" , "ntp" !
54    
55    - tcrossprod(x, y) : do provide methods for y != NULL
56      calling Lapack's DGEMM for "dense"
57      [2005-12-xx: done for dgeMatrix at least]
58    
59    - BUGlet:  Shouldn't lose factorization here:
60      h6 <- Hilbert(6); chol(h6) ; str(h6) # has factor
61      str(H6 <- as(h6, "dspMatrix"))       # has lost factor
62      ## and the same in a similar situation involving  "dpo", "dpp"
63    
64    - Factorizations: LU done; also Schur()  for  *sparse*  Matrices.
65    
66    - is.na() method for all our matrices [ ==> which(*, arr.ind=TRUE) might work ]
67    
68    - use  .Call(Csparse_drop, M, tol) in more places,
69      both with 'tol = 0.' to drop "values that happen to be 0" and for
70      zapsmall() methods for Csparse*
71    
72    - implement .Call(Csparse_scale, ....) interfacing to cholmod_scale()
73      in src/CHOLMOD/Include/cholmod_matrixops.h : for another function
74      specifically for multiplying a cholmod_sparse object by a diagonal matrix.
75      Use it in %*% and [t]crossprod methods.
76    
77    - chol() should ``work'': proper result or "good" error message.
78    
79    - make sure *all* group methods have (maybe "bail-out") setMethod for "Matrix".
80      e.g. zapsmall(<pMatrix>) fails "badly"
81    
82    - Implement  expand(.) for the Cholesky() results
83      "dCHMsimpl" and  "dCHMsuper"  -- currently have no *decent* way to get at
84      the matrix factors of the corresponding matrix factorization !!
85    
86    - rbind2(<sparse>, <dense>) does not work  (e.g. <dgC>, <dge>)
87    
88    - <sparse> %*% <dense>  {also in crossprod/tcrossprod}  currently always
89      returns <dense>, since --> Csparse_dense_prod --> cholmod_sdmult
90      and that does only return dense.
91      When the sparse matrix is very sparse, i.e. has many rows with only zero
92      entries, it would make much sense to return sparse.
93    
94    - sparse-symmetric + diagonal should stay sparse-symmetric
95      (only stays sparse): Matrix(0, 4, 4) + Diagonal(4, 1:4)
96      --> R/diagMatrix.R ('FIXME')
97      but also R/Ops.R  to ensure  sp-sym. + sp-sym. |-> sp-sym.  etc
98    
99    - ! <symmetricMatrix>  loses symmetry, both for dense and sparse matrices.
100      !M  where M is "sparseMatrix", currently always gives dense. This only
101      makes sense when M is ``really sparse''.
102    
103    - example(Cholesky, echo=FALSE) ; cm <- chol(mtm); str(cm); str(mtm)
104    
105      shows that chol() does not seems to make use of an already
106      present factorization and rather uses one with more '0' in x slot.
107    
108    - diag(m) <- val    currently automatically works via  m[cbind(i,i)] <- val
109      This (`[<-` method) is now "smart" for diagonalMatrix, but needs also to
110      be for triangularMatrix, and probably also "dense*general*Matrix" since the
111      above currently goes via "matrix" and back instead of using the 'x' slot
112      directly; in particular, the triangular* "class property" is lost!
113    
114      Note that 'diag(M[,-1]) <- val' is deadly slow for large sparse M,
115      but that's because of the "outer" assignment in the equivalent
116      M[,-1] <- `diag<-`(M[,-1], val).
117    
118    - examples for solve( Cholesky(.), b, system = c("A", "LDLt"....))
119      probably rather in man/CHMfactor-class.Rd than man/Cholesky.Rd
120    
121    - LDL(<CHMsimpl>) looks relatively easy; via  "tCsparse_diag()"
122       {diagonal entries of *triangular* Csparse}
123      --> see comment in determinant(<dsC>) in R/dsCMatrix.R, will give
124      faster determinant
125    
126    - tr(A %*% B) {and even  tr(A %*% B %*% C) ...} are also needed
127      frequently in some computations {conditional normal distr. ...}.
128      Since this can be done faster than by
129        sum(diag(A %*% B))  even for traditional matrices, e.g.
130                   sum(A * t(B)) or {even faster for "full" mat}
131                   crossprod(as.vector(A), as.vector(B))
132      and even more so for, e.g.  <sparse> %*% <dense>
133      {used in Soeren's 'gR' computations},
134      we should also provide a generic and methods.
135    
136    - qr.R(qr(x)) may differ for the "same" matrix, depending on it being
137      sparse or dense:
138        "qr.R(<sparse>) may differ from qr.R(<dense>) because of permutations"
139    
140      This is not really acceptable and currently influences  rcond() as well.
141    
142    - eigen() should become generic, and get a method at least for diagonal,
143      but also for symmetric -> dsyMatrix  [LAPACK dsyev() uses UPLO !],
144      but also simply for dgeMatrix (without going via tradition matrices).
145      What about Sparse?  There's fill-in, but it may still be sensible, e.g.
146      mlist <- list(1, 2:3, diag(x=5:3), 27, cbind(1,3:6), 100:101)
147      ee <- eigen(tcrossprod(bdiag(lapply(mlist, as.matrix))))
148      Matrix( signif(ee$vectors, 3) )
149    
150    - facmul() has no single method defined;  it looks like a good idea though
151      (instead of the infamous qr.qy, qr.qty,.... functions)
152    
153    - symmpart() and skewpart()  for *sparse* matrices still use (x +/- t(x))/2
154      and could be made more efficient.
155      Consider going via  asTuniq() or something very close to
156      .Arith.Csparse() in R/Ops.R
157    
158    - many setAs(*, "[dl]..Matrix") are still needed, as long as e.g.
159      replCmat() uses as_CspClass() and drop0(.) which itself call
160      as_CspClass() quite a bit.  --> try to replace these by
161      as(*, "CsparseMatrix"); forceSymmetric, etc.
162    
163    - implement fast diag(<triangularCsparse>) via calling new
164      src/Csparse.c's diag_tC_ptr()
165    
166    - add examples (and tests!) for update(<CHMfactor>, ..) and
167      Cholesky(......, Imult), also tests for hidden {hence no examples}
168      ldetL2up() { R/CHMfactor.R }
169    
170    - chol(<nsCMatrix>)  gives "temporarily disabled"
171      but should give the *symbolic* factorization;
172      similarly Cholesky(.) is not enabled
173    
174    - writeMM(obj, file=stdout()) creates file "1" since file is silently
175      assumed to be a string, i.e. cannot be a connection.
176      An R (instead of C) version should be pretty simple, and would work with
177      connections automatically ["lsparse" become either "real" or
178      "pattern", "depending if they have NAs or not].
179    
180    - <diagMatrix> o <ddenseMatrix> currently works via sparse, but
181      should return <diagMatrix> in the same cases where
182      <diagMatrix> o <numeric> does.
183    
184    - look at solve.QP.compact() in \pkg{quadprog} and how to do that using
185      our sparse matrices.  Maybe this needs to be re-implemented using CHOLMOD
186      routines.
187    
188    - "sparseVector" : indices, i.e. @i and @length  should be changed
189                     from class "integer" to "numeric" (i.e. double prec),
190      since we want to be able to coerce large sparse matrices to sparse
191      vectors, where length maybe considerably larger than 2^32.

Legend:
Removed from v.530  
changed lines
  Added in v.2207

root@r-forge.r-project.org
ViewVC Help
Powered by ViewVC 1.0.0  
Thanks to:
Vienna University of Economics and Business Powered By FusionForge