Processing math: 100%

Singular Value Decomposition

序言:對於任意矩陣 A,可以運用 SVD 將其分解成 USVT
其中U是 AAT 之eigenvectors;V是 ATA 之eigenvectors


4. For any matrix A, AAT and ATA are both symmetric.

sol. (AAT)T=AAT; (ATA)T=ATA

5. For arbitrary matrices A, B, AB and BA have same nonzero eigenvalues. (if AB, BA both exist)

sol. ABu=λusinceλ,u0,Bu0 and
BA(Bu)(set=y)=B(ABu)=λBu=λy
{ABu=λuBAy=λy

From 4. and 5., we can write ATA and AAT as:
 AAT=UDUT; ATA=V˜DV and D, ˜D have the same nonzero eigenvalues.

6. Eigenvalues of AAT are all 0

sol. Let v be an arbitrary vector with shape (A.shape[0], 1).
vT1×mAm×nATn×mvm×1=yTy0 (set y=vTA)

Now we set v= eigenvector of  AAT (u)
uTAATu=uTλu=λuTu
Since uTu0 and uTAATu0, λ always 0 

And from 6. we can rewrite A as USVT such that 
AAT=USVTVSTUT=UDUT and ATA=VSTUTUSVT=V˜DVT
Si,i的值 = 
{λi (for nonzero eigenvalue λi)0

沒有留言:

張貼留言