L16: SVD and Relatives
Je↵ M. Phillips
March 23, 2020
Dimensional its Reduction
-
→ PCA
Data Matrix Evers column-
↳ has the San , units
A c- IR" 'd
↳ representing-
d- Sama object⇐ii -
n# a >
, n weather station
d points in timetD÷÷÷i÷!÷÷÷:.±÷.
• tEach C stuck price )
Using norms Haik N dags closing value
n=N - d ai shinglesEuclidean distance of consecutive days
what to do if d is lars .
Reduce d s k te ad-
Projection unit vector WEIRD
date point pEIR d. p
Hull = 1
a O
÷t¥÷÷÷÷÷÷÷÷.
I ✓ subspace thrust
In° u C and a )
Basis B represented as 43=54 ,h
,. . Ya)
u ; ⇐ Rd
, Hillel ,Lu; ,
→
It!!,
←either :" " "
d ¥;• stfu.GS/t#.CplfDEIThk
Es fg. sarisA u ,
." "
.
¥¥.ae .us
Kp , vz ) origin
Sum of Soared Errors SSE C A,B)
SSECA,
B) =
a
.fi#llai-TbCailll2
Goal
tk3-dinensic.netsubspace B
BE a min SSE C An,
B )
t'sB
* contains0
Singular Value Decomposition
For n ⇥ d matrix A, define [U, S ,V ] = svd(A) so that USV T = A.
=A U SVT
d s, ,
=DiAh
- frightooo! o I sins .
y fat:÷÷÷o, . .
" "
°
O
n xnn xd
Singular Value Decomposition
=A U
S VT
one data point
left singular vector
right singular vector
singular valueimportance of singular vectorsdecreasing rank order: �j � �j+1
important directions (vj by �j)orthogonal: creates basis
maps contribution of data pointsto singular values
v1
v2
x
�Ax�
L u ; ,Uj ,)=o
•rthornormn
)
Hu; # =L,
11411=1
g./ ↳u ;,ujD=o
⑨ Iii u ;a
Tj I Fti .. -
Fe Zo
Tracing a Point through SVD
Consider a matrix
A =
0
BB@
4 32 2
�1 �3�5 �2
1
CCA ,
and its SVD [U, S ,V ] = svd(A):
U =
0
BB@
�0.6122 0.0523 0.0642 0.7864�0.3415 0.2026 0.8489 �0.34870.3130 �0.8070 0.4264 0.26250.6408 0.5522 0.3057 0.4371
1
CCA S =
0
BB@
8.1655 00 2.30740 00 0
1
CCA V =
✓�0.8142 �.5805�0.5805 0.8142
◆
n = 4
D= L
I Asa ,
•
$ '
Ga00n in nxd did
Tracing a Point through SVD
x = (0.243, 0.97), then what is ⇠ = V T x?
(V =)VT =
✓�0.8142 �.5805�0.5805 0.8142
◆
x1
x2
v1
v2
x
v1
v2
⇠
11×11=1Ax =
USVI
u :-
svtx
← .
.
. . .
T
* 0.9
Tracing a Point through SVDx = (0.243, 0.97), then what is SV T x = S⇠?
(V =)VT =
✓�0.8142 �.5805�0.5805 0.8142
◆
x1
x2
v1
v2
x
v1
v2
⇠
v1
v2�
÷'
Is
o.
..
4thBEEvis -
- Eu,
. . .ad
9
a' is'
-c
'
⇐ &, 11,9¥;Lai
, up - us lai. 4)1/2 top k right
pgflago.MYsing .
vectors
Him unit'D:
= em :÷.' ish . iii. " "
"
-
- ¥9 :.
aims.
=¥: " us
¥i÷¥ai:
Best Rank k-Approximation
=A U
S VTkk
kk
Find An sorank C Are )±k
minimize HA - Antle orHA - Anka
'
EAn --
'
II.,
virusc- Ilznxd
v,
chosen so Hu,
11=1
maximizes HA a 112
thanUr chosen so A.be/-t--1,LU,,u7-o
maximizes Htfvzll'