Main Page | Report this Page
 
   
Science Forum Index  »  Math - Numerical Analysis Forum  »  ? estimate F-norm
Page 1 of 1    
Author Message
Cheng Cosine
Posted: Fri Mar 16, 2007 6:20 am
Guest
Hi:

Given a linear system y = A*x, and one does not know the

details of A. One can only understand how A acts by feeding

different x's and measures y's. In this case, how does one

estimate the Frobenius norm of A?

Thanks,
by Cheng Cosine
Mar/16/2k7 NC
paya
Posted: Fri Mar 16, 2007 6:35 am
Guest
Since Frobenius norm is orthogonally invariant, you can try to compute
the actions on any orthogonal basis of the space which A acts on. Then
compute 2-norms of the images, sum their squares and take a square
root. If you have an orthogonal basis only of some subspace you can
get the lower bound in the same way.
P.J.
Gordon Sande
Posted: Fri Mar 16, 2007 8:31 am
Guest
On 2007-03-16 08:20:40 -0300, "Cheng Cosine" <acosine@nospam.com> said:

Quote:
Hi:

Given a linear system y = A*x, and one does not know the

details of A. One can only understand how A acts by feeding

different x's and measures y's. In this case, how does one

estimate the Frobenius norm of A?

Thanks,
by Cheng Cosine
Mar/16/2k7 NC

Try x of

e_1 = ( 1, 0, 0, ... )
e_2 = ( 0, 1, 0, ... )
.....
e_n = ( ..., 0, 0, 1 )

in turn and you will have all the details of A.
Duncan Muirhead
Posted: Fri Mar 16, 2007 9:30 am
Guest
On Fri, 16 Mar 2007 13:31:55 +0000, Gordon Sande wrote:

Quote:
On 2007-03-16 08:20:40 -0300, "Cheng Cosine" <acosine@nospam.com> said:

Hi:

Given a linear system y = A*x, and one does not know the

details of A. One can only understand how A acts by feeding

different x's and measures y's. In this case, how does one

estimate the Frobenius norm of A?

Thanks,
by Cheng Cosine
Mar/16/2k7 NC

Try x of

e_1 = ( 1, 0, 0, ... )
e_2 = ( 0, 1, 0, ... )
....
e_n = ( ..., 0, 0, 1 )

in turn and you will have all the details of A.

and indeed if n_j is the (euclidean) norm of A*e_i,
then the Frobenius norm of A is sqrt( sum n_i*n_i),
so there's no need to store all the A*e_i
Cheng Cosine
Posted: Fri Mar 16, 2007 4:47 pm
Guest
"Duncan Muirhead" <noone@this.address> wrote in message
news:pan.2007.03.16.14.30.33.570988@this.address...
Quote:
On Fri, 16 Mar 2007 13:31:55 +0000, Gordon Sande wrote:

On 2007-03-16 08:20:40 -0300, "Cheng Cosine" <acosine@nospam.com> said:

Hi:

Given a linear system y = A*x, and one does not know the

details of A. One can only understand how A acts by feeding

different x's and measures y's. In this case, how does one

estimate the Frobenius norm of A?
...
Try x of

e_1 = ( 1, 0, 0, ... )
e_2 = ( 0, 1, 0, ... )
....
e_n = ( ..., 0, 0, 1 )

in turn and you will have all the details of A.

and indeed if n_j is the (euclidean) norm of A*e_i,
then the Frobenius norm of A is sqrt( sum n_i*n_i),
so there's no need to store all the A*e_i


I should have mentioned another condition: A is of high dimension. So it is
highly

desired to use only "a small amount" of testing input vectors to get the
estimation.

Otherwise, just as those suggested, put any mutually orthogonal N vectors
works.

Here N = dim(A).


Thanks,
by Cheng Cosine
Mar/16/2k7 NC
Guest
Posted: Sun Mar 18, 2007 9:46 am
On 16 maalis, 23:47, "Cheng Cosine" <acos...@nospam.com> wrote:
Quote:
On 2007-03-16 08:20:40 -0300, "Cheng Cosine" <acos...@nospam.com> said:

Given a linear system y = A*x, and one does not know the
details of A. One can only understand how A acts by feeding
different x's and measures y's. In this case, how does one
estimate the Frobenius norm of A?

I should have mentioned another condition: A is of high dimension. So it is
highly
desired to use only "a small amount" of testing input vectors to get the
estimation.

Otherwise, just as those suggested, put any mutually orthogonal N vectors
works.

The problem I see is, the Frobenius norm is by nature difficult to
estimate unless you know each and every element. Consider a matrix
with only one nonzero element. If you don't know the location of this
element, you have to try on average O(N) trial vectors to find it.
This matrix could have arbitrarily large Frobenius norm, but you would
have no other estimate than 0 for it until the point that you've found
the element.

Maybe you have an application in mind where you do know something
about the structure of the matrix?
Cheng Cosine
Posted: Sun Mar 18, 2007 12:12 pm
Guest
<toni.lassila@gmail.com> wrote in message
news:1174229193.766392.82230@e1g2000hsg.googlegroups.com...
Quote:
On 16 maalis, 23:47, "Cheng Cosine" <acos...@nospam.com> wrote:
On 2007-03-16 08:20:40 -0300, "Cheng Cosine" <acos...@nospam.com
said:

Given a linear system y = A*x, and one does not know the
details of A. One can only understand how A acts by feeding
different x's and measures y's. In this case, how does one
estimate the Frobenius norm of A?

I should have mentioned another condition: A is of high dimension. So it
is
highly
desired to use only "a small amount" of testing input vectors to get the
estimation.

Otherwise, just as those suggested, put any mutually orthogonal N vectors
works.

The problem I see is, the Frobenius norm is by nature difficult to
estimate unless you know each and every element. Consider a matrix
with only one nonzero element. If you don't know the location of this
element, you have to try on average O(N) trial vectors to find it.
This matrix could have arbitrarily large Frobenius norm, but you would
have no other estimate than 0 for it until the point that you've found
the element.

Maybe you have an application in mind where you do know something
about the structure of the matrix?


I'm interseted in d/u/dt = div( a*grad(u) )+q, after discretization, we
should

get dv/dt = A*v+p where A is a symmetric matrix and all eigenvalues are
negative.

But different discretization can lead to different A, e.g., finite
difference, finite element,

and so on. Is that enough to estimate A's F-norm w/o O(N) trials? Or what
are other

constraints I need to impose?

Thanks,
by Cheng Cosine
Mar/18/2k7 NC
Guest
Posted: Mon Mar 19, 2007 7:57 am
On Mar 18, 7:12 pm, "Cheng Cosine" <acos...@nospam.com> wrote:
Quote:
toni.lass...@gmail.com> wrote in message

news:1174229193.766392.82230@e1g2000hsg.googlegroups.com...



On 16 maalis, 23:47, "Cheng Cosine" <acos...@nospam.com> wrote:
On 2007-03-16 08:20:40 -0300, "Cheng Cosine" <acos...@nospam.com
said:

Given a linear system y = A*x, and one does not know the
details of A. One can only understand how A acts by feeding
different x's and measures y's. In this case, how does one
estimate the Frobenius norm of A?

I should have mentioned another condition: A is of high dimension. So it
is
highly
desired to use only "a small amount" of testing input vectors to get the
estimation.

Otherwise, just as those suggested, put any mutually orthogonal N vectors
works.

The problem I see is, the Frobenius norm is by nature difficult to
estimate unless you know each and every element. Consider a matrix
with only one nonzero element. If you don't know the location of this
element, you have to try on average O(N) trial vectors to find it.
This matrix could have arbitrarily large Frobenius norm, but you would
have no other estimate than 0 for it until the point that you've found
the element.

Maybe you have an application in mind where you do know something
about the structure of the matrix?

I'm interseted in d/u/dt = div( a*grad(u) )+q, after discretization, we
should
get dv/dt = A*v+p where A is a symmetric matrix and all eigenvalues are
negative.

But different discretization can lead to different A, e.g., finite
difference, finite element,
and so on. Is that enough to estimate A's F-norm w/o O(N) trials? Or what
are other
constraints I need to impose?

For any chosen discretization method you will know that A has a
certain sparsity pattern and that the elements will be of certain
magnitude. This will certainly allow you to derive a decent lower
bound. But this calculation has to be done separately for each method,
so if you want a generic toolbox routine I'm afraid you'll have to
generate the matrices element-by-element.
Cheng Cosine
Posted: Mon Mar 19, 2007 8:06 pm
Guest
<toni.lassila@gmail.com> wrote in message
news:1174309028.553841.194600@l77g2000hsb.googlegroups.com...
Quote:
On Mar 18, 7:12 pm, "Cheng Cosine" <acos...@nospam.com> wrote:
...
For any chosen discretization method you will know that A has a
certain sparsity pattern and that the elements will be of certain
magnitude. This will certainly allow you to derive a decent lower
bound. But this calculation has to be done separately for each method,
so if you want a generic toolbox routine I'm afraid you'll have to
generate the matrices element-by-element.


Hmm, lower bound. In the case one cannot drive the system with those
standard

basis, e.g., e_1 = <1, 0, ..., 0>, but can only use a set of linearly
independent vectors,

how to estimate this lower bound for du/dt = A*u+q? Can this estimate be
determined

by a set of q's or must be determined by a set of initial u's?

Thanks,
by Cheng Cosine
Mar/19/2k7 NC
Peter Spellucci
Posted: Tue Mar 20, 2007 1:23 am
Guest
In article <45ff33a7$0$28118$4c368faf@roadrunner.com>,
"Cheng Cosine" <acosine@nospam.com> writes:
Quote:

toni.lassila@gmail.com> wrote in message
news:1174309028.553841.194600@l77g2000hsb.googlegroups.com...
On Mar 18, 7:12 pm, "Cheng Cosine" <acos...@nospam.com> wrote:
...
For any chosen discretization method you will know that A has a
certain sparsity pattern and that the elements will be of certain
magnitude. This will certainly allow you to derive a decent lower
bound. But this calculation has to be done separately for each method,
so if you want a generic toolbox routine I'm afraid you'll have to
generate the matrices element-by-element.


Hmm, lower bound. In the case one cannot drive the system with those
standard

basis, e.g., e_1 = <1, 0, ..., 0>, but can only use a set of linearly
independent vectors,

how to estimate this lower bound for du/dt = A*u+q? Can this estimate be
determined

by a set of q's or must be determined by a set of initial u's?

Thanks,
by Cheng Cosine
Mar/19/2k7 NC


up to this point anyone who answered thought that you can apply the

matrix to a vector. if this is the case and your matrix comes from some
discretization scheme (method of lines?) then take any unit vector
with some "middle" index (since usually the boundary rows will be influenced by
biundary conditions), apply it , take the norm and multiply by sqrt(n).
but your present question sounds as if you could get u(t) only, given
u(0) and q? then take a "typical" u(0), set q=0 , integrate and take
the norm of the difference quotient (u(t+delta_t)-u(t))/delat_t
hth
peter
 
Page 1 of 1       All times are GMT - 5 Hours
The time now is Thu Jan 08, 2009 3:19 am