LECTURF-23.

io
last day : how to visualize a
high -
dim .
distribution ? leg .
human
genome)

✗ c- Bd r .
vector
,
F- ✗ =D
,
Cov ( X) = F- XXI:E .

of random person the world

g. ✗
a in
e.
genome
=

principal components oof ✗

"

of E
"

Eigenvectors Vi =

•
PCA : reduce dimension ÑʰÑ by projecting ✗
:

↑
"
onto
Spain { vi. v2 }
u
,

•
PCA assumes that we can compute the poptulan car .
matrix

I = Cov ( X ) = F- ✗ *
↑
average
over population
don't have We have
But we data of all population .
:

Finitesa-u.pk/y...,XnELRdiidcopiesofX approximate Ely

we
•
.

|J_ÉF
" "

Sample covariance matrix

and hope that PCA (sample ) = PCA ( population ] ,

i e
-
.

Ii ( En) ≈ ai (E) and Vi (En) =

Vi (E) .
(F)

• How
-
large is u
? n=0( lgd ) ? nz ◦ (d) ?
n=o4eᵈ)?_
Curse of V. D ?
COVARIANCE ESTIMATION PROBLEM

OURG-OALtn-ocdsuf.fi#-forC*D*PLAN-
•

1 .
Approximate Eu ≈ E in operator norm

2. Use perturbation theory to conclude (* )

-
I -
 for symmetric matrices )
•

By HW ( operator norm
,

Sd '
-

A
◦

§"= sphere Rᵈ
vt(En E) v1
unit
/
where in
HEN Ell max
-
-
=

Esd"
✓
=
VTIV
VTENV -

TIE ✗ ✗Tv EVTXXTV = F- (✗ v7

• VTIV =
= ,

TT
(✗it G. v7

•
FIX =
± § / Xi ,
v72

112-211 _= max & <✗ i. v5 -

EH , v51
Sd '

#
-

vs

-2 (
"

v ) random variable .

Sd *

(2-6))v€gᵈ Randemproc indexed

by
-

✓ c-
.
•
-1

Wiener process]
•

Compare to the Brownianmokon-ca.k.ae .

time
@ (4) to ↳ →
indexed
by .

Fi E- max / 2- a) I ≤ ?
F- max 1B€> I ≈
•
Sd-1
.

✓ c-
+≤ T

Sd " Discretion
of points
:

Difficulty continuum in
a
.
•
:

2-
 THF-E-NF-TMETHODPY-l j . i e #Ti ? n
.+speeresᵈ"hasane-netm %°i#
ti.e.IS#?-.-.Eei:V-sec-Sd-l7-i:Nx-ail z
I N≤⇐+Dᵈ

≤ E

II. "

g
centered at ni cover
Choose V34 E-balls
orithm :

{
•

%
-
^
.

"" * * &" " "

"
% '

'

at dist > c from { Ki Uz }

Choose the}
.
, .

•
.

at dist
2C from { ka Mie } } .

Choose thee
.
,

•
,

E- Ép wherever impossible

disjoint
%) centered at ai are
•

Uai The -

balls

FETT :{
Hai yE≤ 42
-

1- i≠j 7- Éi j
, , y
Haj YNz≤ 42
.
-

us

Hai Rj 112 { {
≤
=E
-

+ .

But all Ri are c-

separated by
construction
£
• All these balls lie in the
@ %) -
ball '

421
.

centered at 0 .

'

,1+{
vol(B(tekD ≥ N .
Voll BCEIZD

n≤%i:; +1
.

j
- .

Regale Covering ≈
packing _

z -
Prog ( Co¥Ét)

HAtl-1-max.mx#--
""

let A be an mxn matrix Nes an E- net .

Then
,

(By
""
7- ufs
def of operator norm
,
:

(*)
KAUAI KAU .

of
ac-N:/ x-u.kz
By def E- net
,
I

≤ E.

KAGE a) Nz HAK ka cell , ( def of operator )

norm

MAX Aullz
≤ - -

-
= -

≤ HAH -

E. (☒ * )
-

Nz (Dineq )
112 KAU Hz 11Am Au .

( Au Au )
-

11 An ≥
-

HAWK
- -

≥ 11AM
-
11AM -
e ( by ( *) and )

= d- e) 11AM .
]

4-