Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Recent Advances in Algorithmic
Heavy -Tailed Statistics
Sam Hopkins , UC Berkeley (→ MIT)Based on joint work with Yeshwanth Cherapanamjeri , Tarun Kathvria ,Prasad Raghavendra , and Nilesh Tripuraneni .
Measuring success probability in estimation ( confidence intervals)" rate
"
--
Given X . . . . Xu- Po , find ⑤ sit . 110--041 Mdg's failurewith prob . x t - s no '
ambient ratesamples
dimension
For PoEPTclass of distributions
Our game : large class P leg . all distributions with
0111 bold . moments) , but similar guarantees as if
P contains Gaussian s.
Example : Estimating the mean in kElltnexi- Ex H
e-EnXi
, . . . ,Xn -X , Cov IX ) t I +myGaussian : Ht -2Xi - EXILE flattens up . I - s
Drop the Gaussian assumption ?
Is there an estimator right . . . . . Xn) s .t .
this-nut Offal- FEI ) up . I- s ?
Example : Estimating the mean in k
D= I : median of means , truncated mean,. .-
Lugosi -Mendelson ' 18 : estimator based on high -dimensional
quantile 's / median of means (exp . time,"""dits )
H.
'
18 : algorithmic approach to high - dimensional quartilesTheorem : for every 8 > 2-
n
exists As 1h . . .Xn) sit .
Huis-MH e Offa +FLI ) w . p .I -8
, poly In , d , log' 1st
time
AgendaI. Survey of recent ( algorithmic) developments
2. Heavy - tailed covariance estimation
• Algorithm of fcherapanamjeri - H .
- Kathrin - Raghavendra- Tripuraneni)
• 2 key techniques for concentration t convex prog -
- bounded differences Kini Hop -- Yg:ix. Emil
try
- Sos Bernstein { Hemings -- %%s" 'Emil
[Lugosi-Mendelson ‘19, ‘19, ‘19, ‘20, ‘20, Mendelson-Zhivotovskiy ‘20, Minsker ‘15, ‘18, ‘20, Hsu-Sabato ‘16, Lerasale-Oliveira ‘12, Joly-Lugosi-Oliveira ‘17, ... ]
Matching Gaussian Confidence w/ only 0111 moments
• Mean estimation , lz - median of means , trimmed mean
• Mean estimation,Holl
• Covariance estimation
• regression• regularized regression /sparse recovery
B) r i.
Matching Gaussian Confidence w/ only 0111 moments
in polynomial time
Mean estimation,la covariance , H' Hop
authors time authors suboptimal by
H. Ofnd)
's Minster FE
(herapanamjeri - -
s,CIIKRT ( log Ys)
" "
Flammarion - BartlettOln th'd)
similar for linear regressionDeepersin - Leave ,Lei -Luh -Venkat -Zhang Oln'd) ( Hsv -Sabato )
[Depersin-Lecue ‘19, Diakonikolas-Kane-Pensia ‘20, H.-Li-Zhang ‘20,...]
Connections to Robust Statistics
*
I;¥ → Y . . . . Yn→ - n"" "" "
# An• related techniques - new notions of high -dimensional quantile
4 algorithms to compute them• simultaneously robust and high - confidence alqos
AgendaI. Survey of recent ( algorithmic) developments
2. Heavy - tailed covariance estimation
• Algorithm of fcherapanamjeri - H .
- Kathrin - Raghavendra- Tripuraneni)
• 2 key techniques for concentration t convex prog -
- bounded differences- Sos Bernstein
Setup : Xi.. . .
,Xn - X on Rd
,covariance EHR
""'
for simplicity , TRI = Old) , HI Hop -- od )
Theorem : can find Is set.HIS-E Hop a-Offal . Hog Ysl"" t FEI)
Wp . I- S in time poly In ,d , log' Is) if X is certifiably 12,81-hypercontractive.
• sub -optimal by ( log' Is )""- can remove in
:
• exp . time (Mendelson -Zhivatovsky)• poly In ,dits ) time (C # KRT)
⑧ best previous :ois, """ "er, / "''m: " -
-
Theorem : can find Is set.HIS-Ello ,. a-Offal . Hog Ysl"" t FEI)
Wp . I- S in time poly In ,d , log' Is) if X is certifiably 12,81-hypercontractive.
( 2,8) - hypercontractility : Echr)'E OCEAN't
"
Certifiable :O ( ElXiii)"- Efx ,vY= E picul
'
• products of heavy -tailed dis tins
• mixtures thereof
• affine transformations thereof
⑨. . .
The High -Dimensional Median -of -Means ParadigmB,- Z , = FIB,XiXit① Buckets ×n} :
tBlog lls - Flog 'ls= ftp.?gYig② Quantile
Z ."
spectral r- center"
: •
Zz
Hr,Kziivvt) - (Minister •
..
•
for 60% of Zi 's . Zs
II Zi- Mllop Er for 60% ofzi's
The High -Dimensional Median -of -Means Paradigm"
spectral r- center"
: Z ,
•
Hv,Kziivvt) - (Mini) Itr Zz
•
'
for 60% of Zi 's . •
Zs
• -2 is a spectral fat Fin center of Z , . . -Ziogyg w-p- l- S
(Lugosi -Mendelson , Mendelson - Zhiva?
•M.M'
spectral r - centers → HM -M '
Hope LrJust need to find a spectral r-center !
Our Strategy : Man , → . . .⇒M
• Attempt to certify that Mi is a spectralr center
• Success → output Mi
•• Failure → witness V, update Min = Mi INT
→•
Mitt•
• Zi 'sMic. •
[CFB '
19 , CHKRT'24
If GIVEN E and Zi, .. .
,Zioglls ,
can you check that
I (Zi,wt) - ( I
,WTI l E r
for 60% of Zi's and all unit v ?
If GIVEN E and Zi, .. .
,Zioglls ,
Optimization Approach can you check that
Variables bi . . - biogys , Vi . . - Nd Kzinti - Himmler
may"b ;
for 60% of Zi's and all unit v ?
b.V
s 't . bi c- {oil's , Hulk l ,
bikzisvvtl-SM.mil/xbir
If GIVEN E and Zi, .. . ,Ziog4s ,
Optimization Approach can you check that
Variables bi . . - biogys , H . . - Nd -- v Kzinti) - Himmler
May"b ;
for 60% of Zi's and all unit v ?
b.V
s 't . bi C- {oil} , Hull =/ ,• Z .
Et•
Zs
bikzisvvtl-SM.mil/xbir • •
M Zz
If GIVEN E and Zi, .. .
,Zioglls ,
Optimization Approach can you check that
Variables bi . . - biogys , Vi . . - Nd Kzinti) - Himmler
Max"b ;
for 60% of Zi's and all unit v ?
b.v•• Know : E 0.6 - log
' Is forSt . bi c- {oil's , Hulk l , F- I'Intf,bikini) - GM
,wt) ) > bi . r WP . I -8
• Convex relaxation : t 0.610948 , for r -- dog 'lsl""ft t Ft,
W - p . I - S
AgendaI. Survey of recent ( algorithmic) developments
2. Heavy - tailed covariance estimation
• Algorithm of fcherapanamjeri - H .
- Kathrin - Raghavendra- Tripuraneni)
• 2 key techniques for concentration t convex prog -
- bounded differences- Sos Bernstein
Optimization Approach If GIVEN E and Z. .. .
.Zions ,
can you check thatVariables bi . . -biogys , h . . - Nd
Kzinti) - SM ,will er
Max"
bi
b,v for 60% of Zi's and all unit v ?
St . bi c- {oil's , Hulk l,
bi#innit - Gezi,vvt) ) " bi ' r
Any" reasonable
"
convex relaxation will satisfy bounded dias .
~ ~•
E'
( OPT ( Z . . -
y Zqogyg ) - O PT ( Zi . .. .
.E.gig ) ( E l
t '.
- * 9 :• &GR .g
Bi n bi O E Bi E Iµ
. -3
New Goal : convex relaxation of zi-ienpi.MEmax
"
bi
b,V
s 't . bi c- {oil} , Hulk I,
bikini) - Ke,vvt) ) x bir
St . z.IE#.,O.,gPTEO.6log'Is , assuming onlycertifiable 12,8) -hypercontractility
We use Sum of Squares Semidefinite Program-
hierarchy of convex relaxations for anypolynomial optimization problem
max"
bi
b.vg.iii.biko =3"
'
proxy variables for bi,bib ; , vivjbk
2 ✓
St . bit {oil} , Hull =/ , Bi. . .
I
bKZi x bi ' r
① max"
bi
b,V
s 't . bi c- {on} , Huh =/ ,E Mb.at/trEbiKtiivvTI-sqwtDbiKZisvvTI-52.vvTHxbir± Max t's's Kei -e.until
"
WH -- I
Suffices to bound
a.EE?,so4Hn9IiEiIzi-e.uutiy
Suffices to bound
# Himmat zie.mil
( svmotfiidpolynomiab in v
) Max [( (Zi -E) ① ( Zi -E ) , Y )YeyzdlxdlYE Sos
Try ''
,Yao
Suffices to bound
HE?.gs/Min9IiiEizi-e.uui-iy( svmotfiidpolynomiab in v
¥ Max [(Ei -2) ① ( Zi -e ) , Y )YeyzdlxdlHE Sos
= EH Elzi -Elahi -E) H- Sos
② EH -2 Hi -Elahi -sell!Sos Bernstein : control via HE-2 Hi -250 Hi -E5 Hs
.se- -
degree 8 polynomial in X
⇒ control via 12,81 certifiable
hypercontractivity :
HEX 8/1so,I 041
Yields covariance estimation algo . with r -- 041091"" ft + FIH
Open : remove (log'
1st" " ?
Open : optimal heavy- tailed linear regression ?
Applications of• bold . Gifts for convex programs ?
• Sos Bernstein ?