elhnldthd d zfso - stanford universitymontanar/other/talks/oops4.pdf · 2020. 8. 30. · ft e 0,9 1...

5
Sherrington Kirkpatrick model Hn Ily R HN 161 21 6 Wo j Wn GOEIN W La G 1Gt Gigi N o N 10,1 Hw 16 oey yw centered Gaussian process ELHnldthd D zfso.TT General p spin model Hulot the yN centered Gaussian with ELHnlolH.de f Ngc EE l I 347 Zzz CE Xk Sk 3 G Hz Halo Ez LW oak 112 7 Wk indep gaussian tensors how convex many local minima maximize Hnl 6 subj to o c LI lb Q can we solve this approx in poly time input NH zz Output 5 S EL 11 14N st N 2 N p Ancoats Ci e max n HN lo n g

Upload: others

Post on 02-Feb-2021

0 views

Category:

Documents


0 download

TRANSCRIPT

  • Sherrington Kirkpatrick modelHn Ily R

    HN161 21 6 Wo j Wn GOEIN

    W La G1Gt Gigi N o N10,1Hw16 oey yw centered Gaussian process

    ELHnldthd D zfso.TTGeneral p spin model

    Hulot the yN centered Gaussian with

    ELHnlolH.de f Ngc EEl I

    347 ZzzCEXk Sk 3G Hz

    Halo Ez LW oak112 7 Wk indep gaussian tensors

    how convex

    many local minima

    maximize Hnl6subj to o c LI lb

    Q can we solve this approx in poly timeinput NH zz Output 5 SEL 11 14N st

    N 2 Np Ancoats Ci e max nHNlo n g

  • H IN feet yNIs this possible V E oIf not V E C 70

    Exact optimizationPCH.vlohst mgxth.to 21

    Refutation Upper bdALG w c IR st4 ALG W mixHalo2 P ALG W E HE m Halo 7

    Worst case Unless P NP no algorithm can doab Ao b 3 cmf 07

    Typical value Parisi's formula

    V Lt 0,1 2K o nondecr So'HHdt soozollt.xl l 54H 20lt.xl HHG.rolt xlTT oIo Dx IR 016 7 1 1 loRr Igloo I fo'tg lHHtIdt i EHIdHH N3tvtThe OPT v mix Hnk

    firm OPTN info Pcr

    The ft Pcr is strictly convex igf parEUachieved at unique 8 EU

  • Interpretation of IConsider Gibbs measure

    ZypEP NbMwp o L

    Se o LILY Halo 4 e mix Halo bMwp Unit Se E E Cfo 02 Iw Mwp Mwp

    Hid

    VPp N law of ko

    Nsoo N

    Ppw Up on o Dt line PipCEO t

    structure of rI Nooverlapgap r strictly incr on 6,1Il Overlap gap Fct ta EEo at

    rct e srctntl rlt.is Htt

    is tif Ceo t strictly increasing forte lo g constantabove

    1197 of'Ese ko left stts a

  • ft E 0,9 1AlgorithmsL rico D Rao 113 rll w.co sottEl9D Jo3 rfttodtJ3 Htt 3 htt HttL Z U r nondeer fittlott soo fU L n f non dear

    Thin AssumefifthPCH is achieved ThenHE of alg

    with linear complexity St

    PifHulot infzRH t 1Rmt in Pan signaler

    L Pcr

    intfift higfsinuf an.t factIhearoptLem If no overlap gap then

    in.fi crl iyfuPC8g OPT

    Greoltry If no gap TE te of linear time 1gN 70sit ACH Isa's 4 e mixHNK 21 a

    Conj For SK no overlapgapConjecture If overlap gap no poly time 21g V E

  • Unless P NP

    Proof Construct AMP algo Skxt W f try it ftp.t.sfs.icxi x Izt ftG

    InCzt zt zt o7

    t E o 8 7 8 I

    I 8 20stevolution SDE driftchoose coefficients of SDEoptsolving a stock opt contrinfzPL'T B

    E Exitx z Nco 8 Id