Low-Dimensional Matrix Approximation

  • Context: Graduate 
  • Thread starter Thread starter jfy4
  • Start date Start date
  • Tags Tags
    Approximation Matrix
Click For Summary
SUMMARY

The discussion focuses on the optimal projection of a 4x4 matrix into a 2x2 matrix while retaining as much information as possible. The primary method mentioned is Singular Value Decomposition (SVD), where the matrix is decomposed into U, λ, and V components, allowing for the selection of the largest singular values to form the reduced matrix. Additionally, the conversation touches on the potential use of eigenvalues and eigenvectors for interpreting the physical implications of the matrix in specific applications, particularly for Hermitian or real symmetric matrices.

PREREQUISITES
  • Understanding of Singular Value Decomposition (SVD)
  • Knowledge of eigenvalues and eigenvectors
  • Familiarity with matrix operations and properties
  • Basic concepts of linear algebra and dimensionality reduction
NEXT STEPS
  • Research advanced techniques in matrix approximation, such as Principal Component Analysis (PCA)
  • Explore nonlinear dimensionality reduction methods like t-Distributed Stochastic Neighbor Embedding (t-SNE)
  • Study the physical interpretations of eigenvalues in Hermitian matrices
  • Investigate the applications of SVD in data compression and noise reduction
USEFUL FOR

Data scientists, mathematicians, and engineers working with matrix computations, dimensionality reduction, or those interested in optimizing data representation in machine learning and computational physics.

jfy4
Messages
645
Reaction score
3
Hi,

Lets say that I have a 4x4 matrix, and am interested in projecting out the most important information in that matrix into a 2x2 matrix. Is there an optimal projection to a lower dimensional matrix where one keeps most of the matrix intact as best as possible? Thanks.
 
Physics news on Phys.org
What information are you talking about, exactly? 4x4 and 2x2 matrices are rather different (they're functions on completely different spaces). If there's a particular (say, 2-dimensional) subspace of interest, then the restriction of a 4x4 matrix to the subspace may give you a 2x2 matrix, but we really can't say anything specific without more information.
 
I think you should be more specifc. What do you want it for? Do you have an example? Etc.

I can easily think of a (nonlinear) way to transform a ##4\times 4## into a ##2\times 2## such that all information is preserved, but I doubt you're looking for this.
 
good questions. True be told I'm not entirely sure... The matrices house local configurations on a lattice, and I can't keep them all due to computer memory cost, so I want to truncate the matrices and keep as much as I can. The only way I have been doing it is with SVD. Once I preform the SVD I rotate the matrix using the V ( from U λ V^{\dagger}) and then I project out the largest rows corresponding to the largest singular values.

Im interested in a nonlinear way to keep all the information too though lol :)
 
jfy4 said:
The only way I have been doing it is with SVD. Once I preform the SVD I rotate the matrix using the V ( from U λ V^{\dagger}) and then I project out the largest rows corresponding to the largest singular values.

You might like to think about that idea using the eigenvalues and vectors of the matrix rather than the singular values, to see what it means "physically" for your application.

For Hermitian (or real symmetric) matrices, you can interpret it in terms of partitioning the "energy" of the system and then throwing way the "least important" components.

Of course the SVs and EVs are closely related, and for arbitrary non-hermitian matrices the SVs might be easier to work with.
 

Similar threads

  • · Replies 2 ·
Replies
2
Views
3K
  • · Replies 1 ·
Replies
1
Views
2K
  • · Replies 12 ·
Replies
12
Views
3K
  • · Replies 1 ·
Replies
1
Views
2K
  • · Replies 3 ·
Replies
3
Views
4K
  • · Replies 2 ·
Replies
2
Views
2K
  • · Replies 2 ·
Replies
2
Views
5K
  • · Replies 3 ·
Replies
3
Views
3K
  • · Replies 10 ·
Replies
10
Views
3K
  • · Replies 3 ·
Replies
3
Views
1K