0 ratings 0% found this document useful (0 votes) 12 views 8 pages ConvexSpring25 Week6
The document discusses various mathematical concepts related to optimization problems, including properties of eigenvalues and conditions for optimality. It also addresses issues of unbounded solutions and the necessity of regularization in regression problems to achieve sparsity. Additionally, the duality of quadratic programming is explored, highlighting the relationship between primal and dual formulations.
AI-enhanced title and description
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content,
claim it here .
Available Formats
Download as PDF or read online on Scribd
Go to previous items Go to next items
Save ConvexSpring25_Week6 For Later UB TURE I BA Fe
how Hee Eymno arc Nath (AF) = that CHT
Henee Nui CH) 1 Range CH).
New let us examine
rit Later x 447 His sypmetnc
OO uneque sol”
el) His positive Akaike y vvboonded 25 0 ARCH)
KUr: is posite Semidetinile wile one af its eigenvalues =O
coug : H ws 0 me Uynvalue > unvounded
be a ve thak is -ve u be he
eosin Biganverfor .
Zz Au, KER
the fimohion value Ltiliu ~ duutda
Lek Aa) Fea te
Gis Wu + Tau ta
"
os, X20, tne fitet form will qo fo be
ZB flaw AQ.
Dine problem is unbounded and Aveg Mob wave <
finite opal value -
Cosel. WK) = Haye, the eptinal solution P Satis res
Phos =H Havre 20 .
: ° ZF fe ve + ad sol”.Coe 2: EO wlth of lent one etqenvalve. A=0. O.
vfta= He0-
necessary bondétten fer optimality is Hat + C=O
Hote -c
Case 29 = je Z Range (t)| 9 XA satisfying Hee =<
gine A=O
lek ote oly where ut the clgpnwentoe °® 420.
$lau)= L OuThutbatutd = desert
=O
td
Since Hu fo, we can Cngoge Aap A X3~v St
F (AU) > -%. Optioal value is not finite &
the poolem fe unbounded -
cate 2b: ¢ © Rane Cl) 35 satistyng (WR =-C)
then any X+aXu whene Ue Nall (4)
als sabstes H(Ktoxu)= HR+e Hus -
Let us
$ (x4 au) ad (Rretuy ACR ta) + Ot (CR tae) th
= $ a
Yous Cueto Thue oj-
eLeHY +044, Iwvespeotie 4 Chette “tx andy -
—yT
ve Haw = TBVIR, 30Now te \ bl mia, id -F ib =: fle)
Keon)! that tue linger Regressi'ow problem min Ide NE =: fle)
es
feod= ( PlRu-F ) (giQ0-5) = ST PAT Pie - 2 GRID
Pi 2) + GS
PATPR ER = ha xe Tard
whene x20, He 2P (RSX)
d= (598
Cs -2 prays
TE HY O, we have a Unigve aptimaf volute w= e -
F GTR: PRIS
2 y's Paar pei) pals -
TE H has an eigenvalue =O, we can shold tak
PRTG €RangeCH) -
ond we ave infinitely many solutions, volta he same
fiotte optimal value »
when tre number 4 feabres K > number 4 data potnts,
dA) eR ® nae Fewer Toms nat columns.
T
tank Pa) 4K» tete arte fohinitely many
Solutions ,
‘ ty, ven).
‘When a new data pois ts give, then,
ND
In fack, the entnes 4 iD ond DtHv ane going jo
he very Different from gach fray , an
th Is Aifiodt fo pedi Y¥ on ned datesTn onder to aljouiate tne above
fssue, the wegression problem
ts moditred to thelude regulars zat ferme , ettres tn the
cost fonction oy @ —_ tonstrasats.
Ex?
min PR o-~ Sih + A Yeo, A: hyper paxameber
OO eos Foretion ys NeW
Strongly Convex &
we nave onrqve
Sparsity promoting solutong -
Cone = O) m=
(fe |, £04
[mint PRyo- Sys
wd
4: Cenadtor) < mn)
en.
> NP-Hand problem 3
Card tw) ts nob a, Conver
problem
For a funttion f , de (eonvex)
ae sup $B: g is convex }
Under ceniatn agevmp tia, we cam shoo jolly < camden.
wt
K @xdle)Neehate 208 TB Fe
Qo neguiadzed
Motivated by tne Above obsesvation, we deting
Begesiow problem aS
ia, Were - SIE + Oily,
a
iver
an Oltenative [min \ $(&yw-5 >
formulation wees ( ———
Sb qos € E).
Larger value foc smaller value of C megult 1) goluttong
hot hove most Of frets enftes arual bb or Clee fo O,
Despising ond Precewise wonstant aspx) matron
Given signed oie eR me
AAO: vave tt Signal +t g
tee #, oeb eT te |g :
+ 4
84. YU ts a Smoot appr winahin :
ott, oO Tt
ys yl) gtk) og tk-!)
yl) 2 90) UU
The signal 4 should have two _propesties :
DD yet lly -x741L_ shoub be small
Q) [yw ybayyt Should be smal).
(ec
min py tT (yGo-yd >
yer 22Tn orden to achiwe sparsity tn [y(k ~yCk-1)'| vector
io = Rega OW.
% pyar $
fo ~ U4 i)-4li-
newt a 7 t S gli) - yG-1)
Dual of a Quatrahe Program (ap)
Consider. the &P wn Sal tot ole $4, H espace,
4 Ansb 1 4 peste
_ bl4,4)= Lo he +o nth + x CAx-b)
| lt CMe) edb
bt
ALA Inf Atwa) =F -we Aare g RH)
eR
} }
U
5 Uk = Bar(Aate) =0
toe -(crAa).
BECCRAAJERLH), thoy Jz ct Hes —(CFAA)
z= -H'G+AA)
deat (—Hitertay) & (at@eaan) — (CRA) i (@raA) +4
= aN
a ok (ee Ra) Hee RA) th TDdual_optinrzation pblem : mn Aa)
q
ot equivalently min t(or Ra wterda) +b -2-
nD
quedvatic tn A -
Hence, dual of a BP is onaney RP.
— Quadratically tonstrained Quadeahic _progeams (CAP)
Dif FATHOM + CIA +d
et. LolWn telaed: 20, tel2---m 2 A
LR
7, = bE1)2--- Dk
bol Ha fee sOr # \2---P 2 i
The above proviem is Comex whey H-%@, %20,/|,2--M
G:=O , b=L2-- P
Let us derive tts dual whew tye problem is comrex & thro
“
44 (2 Pix tarde)
t=
+ 5 bs (frre; )
aap)
fae eon! log > HMO
ca e Tati!
— clay) 70
equivalentiy < mia a
B+ avy) fre)” 9 yO:
= clam Hla) .
& es
Fot 2 aet +7 4545 | Linear makix
Snequal'ty ,
will be discusted yn deta]
Subrequetty -