Re: SVD when calculated for a corpus of similar category?


Peter Spellucci wrote:
In article <1163579376.066684.82740@xxxxxxxxxxxxxxxxxxxxxxxxxxxx>,
"paluri" <santoshpaluri@xxxxxxxxx> writes:
>Hello everybody,
>
>I have a doubt regarding SVD. Suppose i compute SVD for a huge corpus
>of similar category, and i have the decomposition as [USV^T] . Are the
>singular values in the diagonal matrix S arranged in descending order
>along the diagonal will be very near to each other? i mean , is the
>numerical difference between one singular value and the next in the
>diagonal will be negligible?. I would be thankful for the response...
>

???????? what please is a "huge corpus of similar category"?
you mean a set of nearby matrices ???
then the answer is yes:
let sigma(A,i) denote the singualr values of A in descending order
and sigma(B,i) those of B. Then
|sigma(A,i)-sigma(B,i)|<= ||A-B|| for all i
||A-B||=sigma(A-B,1)
hence if A is near B, then all the singular values of A and B can be paired
corresponding to this order with this universal error bound
(follows from the courant-fischer-minimax characterization of eigenvalues)
hth
peter


By "huge corpus of similar category", i mean web pages downloaded from
a similar category ,
Actually i am creating a term by document matrix(rows indicating the
terms, columns the documents and each element of the matrix indicating
the frequency of each term in the corresponding document) of certain
number of web pages and then i will aplly SVD to that term by document
matrix in order to calculate the similarity bwetween the documents or
web pages.

Now, what i am asking is , if i create the term by document matrix of
pages or documents taken from the same category, i.e if they are
already similar, then in the SVD of the Term by Document matrix which i
create using these similar pages, does the singular values in the
diagonal will be very near to each other, i.e. the numerical difference
between one and next singular value in the diagonal will be very
small..?

.

Relevant Pages

  • Re: SVD when calculated for a corpus of similar category?
    ... Peter Spellucci wrote: ... Suppose i compute SVD for a huge corpus ... >singular values in the diagonal matrix S arranged in descending order ...
    (sci.math.num-analysis)
  • Re: SVD when calculated for a corpus of similar category?
    ... I have a doubt regarding SVD. ... Suppose i compute SVD for a huge corpus ... singular values in the diagonal matrix S arranged in descending order ...
    (sci.math.num-analysis)
  • Re: SVD Algorithm
    ... > I will use SVD for Latent Semantic Indexing. ... In effect if you sort into descending order, ... could copy the singular values to a single-dimensioned array, ... This indicates that using LAPACK routines or C/C++ wrappers around them ...
    (sci.stat.math)
  • Re: SVD Algorithm
    ... I will use SVD for Latent Semantic Indexing. ... So I need singular value in descending order (to take the first "k" ... > eigenvectors of the untransposed matrix, ...
    (sci.stat.math)
  • A query about SVD
    ... I have a doubt regarding SVD.Suppose i compute SVD for a huge corpus ... element of the matrix indicating ... does the singular values in the ...
    (sci.math)