Contents lists available atScienceDirect
Operations Research Letters
www.elsevier.com/locate/orl
Parameterized complexity of configuration integer programs
Dušan Knop
a,
∗,
1, Martin Koutecký
b,
2, Asaf Levin
c,
3, Matthias Mnich
d,
4, Shmuel Onn
c,
5aDepartmentofTheoreticalComputerScience,FacultyofInformationTechnology,CzechTechnicalUniversityinPrague,Prague,CzechRepublic bCharlesUniversity,Prague,CzechRepublic
cTechnion–IsraelInstituteofTechnology,Haifa,Israel
dHamburgUniversityofTechnology,InstituteforAlgorithmsandComplexity,Hamburg,Germany
a rt i c l e i n f o a b s t r a c t
Articlehistory:
Received6August2021
Receivedinrevisedform31October2021 Accepted2November2021
Availableonline15November2021
Keywords:
Parameterizedalgorithms ConfigurationIP Surfing
Configuration integer programs (IP) have been key in the design of algorithms for NP-hard high- multiplicityproblems.First,wedevelopfastexact(exponential-time)algorithmsforConfigurationIPand matchinghardnessresults.Second, weshowcase the implications oftheseresults tobin-packing and facility-location-likeproblems.
©2021TheAuthor(s).PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-ND license(http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction
In 1961, Gilmore and Gomory [11] introduced the fundamen- tal and widely influential notion of Configuration IP (ConfIP for short), and applied it to the BinPacking problem. In BinPack- ing, one is given a set J ofn items with sizes p1
, . . . ,
pn∈
N, which need to be assigned (or packed) to a set of m identical bins of a common capacity T∈
N such that the total size of itemspackedper binisnotlarger thanitscapacity.Thisclassical problem has long been investigated from the perspective of ap- proximationalgorithms [13,20],andinrecentyearsalsofromthe perspective ofexactalgorithms [3,12,16,18].Gilmore andGomory usedtheConfIPtodescribeaBinPackingsolutionbyalistoftu- ples“(packingofonebins,multiplicityμ
ofbinswithpackings)”.In this paper,we consider a more generalform ofConfiguration IP.Formotivation,observethatthenaturalinputencoding forBin
*
Correspondingauthor.E-mailaddresses:dusan.knop@fit.cvut.cz(D. Knop),koutecky@iuuk.mff.cuni.cz (M. Koutecký),levinas@ie.technion.ac.il(A. Levin),matthias.mnich@tuhh.de (M. Mnich),onn@ie.technion.ac.il(S. Onn).
1 Partially supported by the OP VVV MEYS funded project CZ.02.1.01/0.0/0.0/16_019/0000765“ResearchCenterforInformatics”.
2 Partially supported by Charles University project UNCE/SCI/004, and bythe project19-27871XofGAˇCR.
3 PartiallysupportedbyIsraelScienceFoundationgrant308/18.
4 SupportedbyDFGgrantMN59/4-1.
5 PartiallysupportedbytheDresnerchairandIsraelScienceFoundationgrant 308/18.
Packing does not list the item sizes one by one; rather, the n itemsare classified intod
n itemtypes andthe input specifies the size pj∈
N andthe number nj of items oftype j. The so- calledhigh-multiplicityencoding(atermfirstcoinedbyHochbaum andShamir [14]) thus givesa size vector p= (
p1, . . . ,
pd)
anda multiplicityvector n= (
n1, . . . ,
nd)
with n1=
n1+ · · · +
nd=
n.Observenowthat, foranypacking
σ
,each bini definesa vector x= (
x1, . . . ,
xd)
,withxj beingthenumberofitemsoftype j as- signed to i byσ
; such a vector x is called a configuration. This high-multiplicitybinpackingproblem(andmanyother optimiza- tionproblems)canberephrasedinthefollowingway,aswasfirst observedbyGilmoreandGomory [11].LetNbetheset{
0,
1, . . . }
. LetCT=
x
∈
Nd|
px≤
Tdenotethesetofconfigurationsofsize at most T. Then, to decide if the given instance admits a feasi- ble solution amounts to deciding if n can be written as a sum ofm configurationsfromCT,i.e.,n
=
mi=1xi,withxi
∈
CT forall i=
1, . . . ,
m.Thistaskleads toa relativelysimpleformof config- urationIP,which hasan integralvariableλ
x∈
N foreach config- uration x∈
CT, and asks for a solution withx∈CT
λ
x=
m andx∈CT
λ
x·
x=
n.However, other optimization problems lead to more complex formsofConfigurationIP.Maybetherearebinsofdifferenttypes, inwhich case
τ
denotes the number ofbintypes withμ
i being thenumberofbinsoftype i,wherebinsofacommontypehave commoncharacteristics.Itemsmayalsohaveseveralothercharac- teristics,andthenan item typeisthe setofitemswithcommon characteristics. We may also incur a cost for each configuration.Thus,ourConfigurationIPisdefinedasfollows.
https://doi.org/10.1016/j.orl.2021.11.005
0167-6377/©2021TheAuthor(s).PublishedbyElsevierB.V.ThisisanopenaccessarticleundertheCCBY-NC-NDlicense (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Configuration IP(ConfIP)
Input: Dimensiond
∈
N,finitesetsX1, . . . ,
Xτ⊆
Zd, objectivefunctions f1, . . . ,
fτ:
Zd→
Z,numbersμ
1, . . . , μ
τ∈
N,andatargetvectorn∈
Zd. Find: min τi=1
x∈Xi fi
(
x) · λ
ix|
τi=1
x∈Xix
· λ
ix=
n,
x∈Xi
λ
xi= μ
i, λ
ix∈
N∀
x∈
Xi∀
i∈ {
1, . . . , τ }
.WerefertovandenAkkeretal. [26] forpracticalinvestigations ofConfIP,andtoHochbaumandShmoys [15],Alonetal. [1],Fer- nandez de laVega andLueker [9] andKarmarker andKarp [20]
forstudiesofapproximationalgorithmsbasedonConfIP. Previous studiesofthebinpackingprobleminthehigh-multiplicitysettings weredonebyJansenandSolis-Oba [17] andJansenandKlein [16].
Ourcontributionsandpaperoutline.Ourcontributionistwo-fold.
First, inSection 3we provideseveralfixed-parameter algorithms, andinSection4weprovehardnessresultsforConfIP,delineating thecomplexity landscapewithregardto themostnaturalparam- eters showing that some trade-offs evident in our analysis are inevitable. Second, to showcase the usefulness and versatility of ourapproach, weapply inSection 5ouralgorithms tohighmul- tiplicity problemsinbin packingandsurfing, a generalmodel of facilitylocationandmulticommodityflows.
2. Preliminaries
Forpositiveintegersm
,
nweset[
m,
n] = {
m,
m+
1, . . . ,
n}
and[
n] = [
1,
n]
. Wewrite vectorsinboldface(e.g.,x,
y) andtheiren- tries in normal font (e.g., the i-th entry of x is xi or x(
i)
). Forα ∈
R,α
isthefloorofα
,andα
istheceiling ofα
.Fordata object O,wedenotebyOitsbinaryencodinglength.ThesetN isthesetofnon-negativeintegers,thatis,N= {
0,
1,
2, . . . }
.TheinputofConfIPcanbegivenexplicitlyonlyinfairlylimited scenarios.Thus,weassumethateach(possiblyverylarge)setXiis definedsuccinctly.Thefollowingdefinitioncapturesthecasewhen each Xi isdefined asa projection of integerpoints ofa rational polytope.
Definition1(P -representation).Fori
=
1, . . . , τ
,let Pi⊆
Rd+di be a polytope andletπ
i((
x,
x)) =
x∈
Rd bea projection discarding the last di coordinates. We call the collection P1, . . . ,
Pτ a P - representation of X1, . . . ,
Xτ if Xi= π
i(
Pi) ∩
Zd,foreach i∈ [ τ ]
. Let each Pi by defined as Pi=
(
x,
x) |
Ai(
x,
x) ≤
bifor some Ai
∈
Zmi×(d+di) andbi∈
Zmi.The parametersofa P -representation arethefollowingquantities:M=
maxi∈[τ]mi,D=
maxi∈[τ]di,=
maxi∈[τ]Ai∞,L= ,
b1, . . . ,
bτ.We consistentlyusesuperscripts torefer toobjectsandquan- titiesrelatedtothe types(e.g., Xi
,
di,
fi, . . .
).Toavoidconfusion, we always use parentheseswhen intending to expressexponen- tiation (e.g.,(
di)
2). With each Xi given implicitly, the objective functions fi alsomust haveimplicitrepresentations, orbe given by oracles. We consider the following classes of objective func- tions:•
linear:givenvectorsw1, . . . ,
wτ∈
Zd,let fi(
x) =
wix.•
convex:each fi(
x)
isaconvexfunction.•
extension-separableconvex:each fi(
x) =
minx:(x,x)∈Pi∩Zd+dgi
(
x,
x)
for gi a separableconvex function. (In some ofour applica- tions theobjective is only expressible asa separableconvex functionintermsoftheauxiliaryvariablesx.)
•
concave:each fi(
x)
isaconcavefunction.•
fixed-charge:each fi(
x) =
ci∈
N ifx=
0and fi(
x) =
0 oth- erwise;wecallci apenalty.ForaConfIPinstancegiveninits P-representation,set fmax
=
maxi∈[τ] max
(x,x)∈Zd+di:Ai(x,x)≤bi
fi
(
x) .
CONFIP asN-foldIP.Below, weconnect ConfIP withaspecial classofintegerprograms(IPs).ThebaselineIPis:minf
(
x) :
Ax=
b,
l≤
x≤
u,
x∈
Zn where f:
Rn→
R, A∈
Zm×n, b∈
Zm, and l,
u∈ (Z ∪ {±∞} )
n.Wedenote fmax=
maxx∈Zn:l≤x≤u
|
f(
x) |
.Ageneral- izedN-foldIPmatrixisdefinedasE(N)
=
⎛
⎜ ⎜
⎜ ⎜
⎜ ⎝
E11 E21
· · ·
E1N E12 0· · ·
00 E22
· · ·
0.. . .. . . . . .. .
0 0· · ·
E2N⎞
⎟ ⎟
⎟ ⎟
⎟ ⎠ .
Here,r
,
s,
t,
N∈
N,E(N)isan(
r+
Ns) ×
Nt-matrix,Ei1∈
Zr×t and Ei2∈
Zs×t,i∈ [
N]
,are integermatrices.ProblemIPwith A=
E(N) is known as generalized N-foldinteger programming (generalized N-fold IP)[7]. The structure of E(N) allows us todivide any Nt- dimensionalobject,such asthevariablesofx,boundsl,
u,orthe objective f, into N bricks of sizet,e.g. x= (
x1, . . . ,
xN)
. Weuse subscriptstoindexwithin a brickandsuperscriptsto denotethe indexofthebrick,i.e., xij is the jth variableofthe ith brickwith j∈ [
t]
andi∈ [
N]
.Wecalla brickintegralifallofitscoordinates areintegral,andfractionalotherwise.ThehugeN-foldIP problemisthe(high-multiplicity)extension ofgeneralized N-fold IP wherethere arepotentially exponentially manybricks.The inputtoa hugeN-foldIP problemwith
τ
types ofbricksis definedby matrices Ei1∈
Zr×t and E2i∈
Zs×t,i∈ [ τ ]
, vectorsl1, . . . ,
lτ,u1, . . . ,
uτ∈
Zt,b0∈
Zr, b1, . . . ,
bτ∈
Zs,func- tions f1, . . . ,
fτ:
Rt→
R satisfying∀
i∈ [ τ ], ∀
x∈
Zt:
fi(
x) ∈
Z andgivenbyevaluationoracles,andintegersμ
1, . . . , μ
τ∈
Nsuch that τi=1
μ
i=
N. We say that a brick is of type i if its lower and upper bounds are li and ui, its right hand side is bi, its objectiveis fi, andthe matricesappearing at the corresponding coordinates are Ei1 and Ei2. We let E be the 2× τ
block ma- trix E=
E11 E21
· · ·
Eτ1 E12 222· · ·
Eτ2. We refer toOnn [25] and Knop et al. [22,23] forstudiesofhuge N-foldIP.
3. Algorithmsfor CONFIP
Ourgoalistoprovethefollowingtheorem.
Theorem1.LetSbeaConfIPinstancegiveninitsP -representation (if the objective f is convex or concave,we assume itis presented byanevaluationoracle),andletN
= μ
1=
τi=1
μ
iandLˆ =
L+
n,
fmax,
N.1 ConfIPwithalinear,convex,orfixed-chargeobjectivecanbesolved in time
(
N(
d+
D))
O(N(d+D))Lˆ
O(1), and is thus fixed-parameter tractableparameterizedbyN andd+
D.2 ConfIPwithaconcaveobjectivecanbesolvedintime
(
M N(
d+
D) ·
log)
O(N(d+D))Lˆ
O(1),andisthusfixed-parametertractableparam- eterizedbyN,M,d+
D,andwithgiveninunary.
3 ConfIP with a linear or an extension-separable convex objective canbe solvedintime
(
Md)
O(M2d+d2M)Lˆ
O(1),and isthusfixed- parametertractableparameterizedbyM,d,and.
4 ConfIPwithalinearorfixed-chargeobjectivecanbesolvedintime
( τ
dD Mlog)
τ(d+D)O(1)Lˆ
O(1),andisthusfixed-parametertractableparameterizedby
τ
,M,d,andD ifisgiveninunary.
Part 3 andPart 4thus meanthat we can solve ConfIP either in doubly-exponential time parameterized by d, D,
τ
, m, and M andall numbershave tobe giveninunary, orsingle-exponential timeparameterizedbyd,m, M,andthelargestcoefficient(but for polynomial
τ
). In Part 2 and Part 4 we use the fact that(
logα )
β≤
2β2/2+ α
o(1) [5,Hint 3.18] to saythat havinglogin thebaseamountstofixed-parameteralgorithms when
isgiven in unary. It is worth noting that the parameter-dependence for N-fold IP w.r.t. the “usual” parameters is
(
Md)
O(d2M+M2d)
. As thegeneralruntimedependson,e.g., ellipsoidmethod,we givea boundof(
L+
N+
D)
O(
1)
thatsufficestocomparewiththeabove Theorem1.ThenextlemmashowshowtomodelConfIPashuge N-foldIP.Lemma2.LetaConfIP instanceS begiven inits P -representation.
Thenintime
( τ +
D+
M+
L+ μ ,
n)
,onecanconstructahugeN-fold IPwhichmodelsSandhasparametersr=
d,s=
2M,t=
d+
D+
M, E∞=
,N= μ
1,andwith fi beingtheobjectiveforbricksof type i.Proof. LetE1
= (
I0) ∈
Zd×(d+D+M)whereIisthe(
d×
d)
-identity matrix and 0 is a(
d× (
D+
M))
-all-zero matrix, let E1i=
E1 for each i∈ [ τ ]
, and let b0=
n. The last M coordinates ofeach brick will play the role of slack variables in order to model in- equalities in the system Aix≤
bi. For each i∈ [ τ ]
, obtain Ei2 from Ai by adding M−
mi zero rows and D−
di zero columns, and then appending from the right the M×
M identity matrix, ensuring E2i has M rows and d+
D+
M columns, and append M−
mizeroes tobi.Formallyextendtheobjectivefunction fi to d+
D+
MdimensionsbymakingitignorethelastM+
D−
didi- mensions.Foreachi∈ [ τ ]
,defineli= {−∞}
d+di× {
0}
M+D−di and ui= {+∞}
d+di×{
0}
D−di×{+∞}
M.Letμ
ibethenumberofbricks oftype i,foreachi∈ [ τ ]
.It is easy to check that, foreach brick j
∈ [
N]
oftype i∈ [ τ ]
oftheresultinghuge N-fold formulation,xj restrictedtothefirst d
+
di coordinates can take on exactly the values of Pi∩
Zd+di. Moreover, theobjectivevalueofthe brickis exactlytheobjective value of corresponding point of Pi∩
Zd+di in the ConfIP prob- lem.Finally,bythedefinitionofE1,thesumoftherestrictionsof all bricks to the firstd coordinates is exactly n.This showsthat wehavereducedtheConfIPinstancetoahugeN-foldIPinstance in a way which allows us to recover the ConfIP optimum from thehuge N-foldIP optimum.Moreover,theboundsare clearlyas statedinthelemma.ThefirstthreepartsofTheorem1areestablishednextvia ap- plication ofprevious results,as we describe next: Use Lemma 2 to obtainan N-fold IP instance.Part 1forconvexor linearfunc- tionsfollowsbyapplyinganalgorithmbyDadushandVempala [6]
for solving convex IPs, which runs in time pO(p)
ˆ
LO(1), where p=
N(
d+
D)
is the dimension aswe can deletethe slack vari- ables and use the system of inequalities instead of the N-fold IP. For afixed-charge objective,we guess, for each i∈ [ τ ]
where 0∈
Xi, a numberμ ¯
i≤ μ
i such that an optimal solutionλ
hasλ
0i= μ
i− ¯ μ
i.With thisguess athand, the objectiveis fullyde- terminedto be τi=1
μ ¯
ici andit remainsto verifywhetherthere existsacorrespondingdecompositionofnbysolvingConfIPwith thevectorμ ¯
insteadofμ
andwithoutanyobjective.Finally,pick the bestamongall guesses whosecorresponding ConfIP is feasi- ble.There areatmostNτ≤
NN guesses. ToprovePart2,we useanalgorithmbyCooketal. [4] toenumerateallverticesofthecor- respondingpolyhedronintime
(
log·
M N(
D+
d))
O(N(D+d))ˆ
LO(1). Sinceaminimumofaconcavefunctionisalwaysattainedataver- tex,itsufficestoevaluatetheobjectiveoneach vertexandreturn thebest asthe output.Part 3is byapplying thefixed-parameter algorithmforhuge N-foldIPbyKnopetal. [23].Thus,itremainstoprovePart4ofTheorem1thatisournext goal. Our proof builds on a Structure Theorem of Goemans and Rothvoß and the idea of the proof of their main theorem [12, Theorem 2.2]. The Structure Theorem applies to the single-type settingandimpliesthatforanysolution
λ
correspondingtoade- compositionofn,thereexists asolutionλ ˆ
whosesupportmostly lies within a precomputable and not-too-large set Y of “impor- tant” configurations. We first extend the Structure Theorem into themultitypesetting,andthenuseit asfollows.Foreachtype i, we compute the set of “important” configurations Yi, and then guessfromit a smallsubset ofconfigurations whichwill appear inthesolution.Usingthis,weconstructanILPinsmalldimension, solveit,andderivefromitanoptimalsolutionλ
.Wetakespecial caretoenforcethemultiplicityconstraint(i.e.,λ
i1= μ
i,foreach i∈ [ τ ]
)andarguehow toencode alinearanda fixed-chargeob- jective.LetusbeginwiththeStructureTheoremofGoemans and Rothvoß [12]:Proposition3(StructureTheorem [12]).Let P
= {
x|
Ax≤
b} ⊆
Rd be a polytope with A∈
Zm×d and b∈
Zm such that all coeffi- cientsare bounded byin absolute value. Thenthere exists a set Y
⊆
P∩
Zd ofsize|
Y| ≤
S:=
mddO(d)(
log)
d thatcan becom- puted in time SO(1) with the following property. For every vector n=
x∈P∩Zd
λ
xxwithλ ∈
NP∩Zd,thereexistsavectorλ ˆ ∈
NP∩Zd suchthatn=
x∈P∩Zd
λ ˆ
xx,satisfyingthefollowingthreeconditionsλ ˆ
x∈ {
0,
1} ∀
x∈ /
Y ,|
supp(ˆ λ) ∩
Y| ≤
22d,|
supp(ˆ λ) \
Y| ≤
22d.Next,weextendProposition3tothemultitypesetting.
Lemma4.LetP1
, . . . ,
Pτ bea P -representationofX1, . . . ,
Xτ.Then, foreach i∈ [ τ ]
,there exists a set Yi⊆
Pi∩
Zd+di of size|
Yi| ≤
Si:= (
mi)
(d+di)(
d+
di)
O(d+di)(
log)
d+dithatcanbecomputedintime(
Si)
O(1)withthefollowingproperty.Foreveryvectorn
=
τ i=1(x,x)∈Pi∩Zd+di
λ
i(x,x)xwithnon-negativeintegral
λ
,thereexistsanon-negativeintegralvectorλ ˆ
suchthatn=
τi=1
(x,x)∈Pi∩Zd+di
λ ˆ
i(x,x)x,and,foreachi∈ [ τ ]
, a)λ ˆ
i(x,x)∈ {
0,
1} ∀ (
x,
x) / ∈
Yi,
b)|
supp( λ ˆ
i) ∩
Yi| ≤
22(d+di),
c)|
supp(ˆ λ
i) \
Yi| ≤
22(d+di),
d)ˆ λ
i1= λ
i1.
Proof. First,extendeach Pibyacoordinatewhichisalways1,i.e., replace Pi by
{(
1,
x,
x) | (
x,
x) ∈
Pi}
.This only increases the di- mensionby1 andrequiresanadditionalequalityconstraint.Then, applyProposition3toeachPiindividually.Letμ
i= λ
i1andN= λ
1.Observethatif(
N,
n) =
τi=1
(1,x,x)∈Pi∩Zd+di
λ
i(1,x,x)(
1,
x)
, thenthereisadecompositionof(
N,
n)
intoτ
summands( μ
i,
ni) =
(1,x,x)∈Pi∩Zd+di
λ
i(1,x,x)(
1,
x)
to which Proposition 3 applies di- rectly, and we obtain all points except for point d). To argue thislastpoint,notethatbothdecompositionsλ
i, λ ˆ
idecompose ni intoμ
i points of Pi∩
Zd+di andtheclaim holds. Thus, foreach i∈ [ τ ]
, Yi isobtained byapplying Proposition 3to theextended polytope Pi and then projecting out the first coordinate of each elementofthecomputedset.WearenowreadytofinishtheproofofPart4ofthetheorem.
First,computethesetsY1
, . . . ,
Yτ fromLemma4.Ourgoalnowis to setup an ILP insmall dimension,whose solutioncorresponds toanoptimalsolutionλ
withthepropertiesofLemma4.Fixsuch an optimalλ
. Foreachi∈ [ τ ]
,guessa subset Zi⊆
Yi whichsat- isfies|
Zi| ≤
22(d+di) and|
supp(λ
i) \
Zi| ≤
22(d+di).Alsoguess the numberμ ¯
i= |
supp(λ
i\
Zi) |
.There areτi=1
(
Si)
O(22(d+D))choices.Foreach guess, we apply the followingprocedure, andthen pick thebestobtainedvalue acrossallguesses andtransformitintoa solutionofConfIP.Introduceavariable
λ
i(z,z)foreachi∈ [ τ ]
and each(
z,
z) ∈
Zi.Additionally,foreachi∈ [ τ ]
,introduceμ ¯
ivectors ofvariables(
x,
x)
ij.Then,considerthefollowingconstraints:Ai
(
x,
x)
ij≤
bi∀
i∈ [ τ ] ,
j∈ [ ¯ μ
i]
(1) τi=1
⎡
⎣
(z,z)∈Zi
λ
(iz,z)z+
¯ μi
j=1 xij
⎤
⎦ =
n (2)(z,z)∈Zi
λ
i(z,z)= μ
i− ¯ μ
i∀
i∈ [ τ ]
with0∈ /
Xi (3)(z,z)∈Zi
λ
i(z,z)≤ μ
i− ¯ μ
i∀
i∈ [ τ ]
with0∈
Xi (4)λ
i(z,z)∈ N ∀
i∈ [ τ ], ∀(
z,
z) ∈
Zi (5)(
x,
x)
ij∈ Z
d+di∀
i∈ [ τ ], ∀
j∈ [ ¯ μ
i],
(6)and,dependingontheobjectiveoftheConfIPinstance,solvewith oneoftheobjectives
linear
(λ,
x,
x) =
τ i=1⎡
⎣
(z,z)∈Zi
λ
i(z,z)(
wiz) +
¯ μi
j=1
wixij
⎤
⎦ ,
or,
fixed-charge
(λ,
x,
x) =
τ i=1⎡
⎣
(z,z)∈Zi
λ
i(z,z)ci+ ¯ μ
ici⎤
⎦ .
Constraints (1) and (6) ensurethatthevariablevectors
(
x,
x)
ijas- sumevaluesfrom Pi∩
Zd+di,foreachi∈ [ τ ]
,enforcingthemean- ing that these variables represent the part of solutionλ ˆ
whose support does not lie in Yi.Constraint (2) ensures that the solu- tionindeedcorrespondstoadecompositionofnintopointsfrom(
Pi∩
Zd+di)
.Theconstraints (3)–(4) ensurethatthenumberof non-zeroconfigurationsoftype iisatmostμ
i.Let
(λ,
x,
x)
be an optimum of the ILP above, computed using an algorithm for ILP in small dimension (cf. [10,19]).We construct a solution
λ
∗ of ConfIP as follows. For each i∈ [ τ ]
and each z∈
Zd such that z∈ π
i(
Pi)
, letλ
∗(
i,
z) =
z∈Zdi:(z,z)∈Pi
λ(
i,
z,
z) +
j∈[ ¯μi],(x,x)ij=(z,z)1
,andlet
λ
∗(
i,
0)
= μ
i−
z∈(πi(Pi)\0)∩Zdλ
∗(
i,
z)
. We argue thatλ
∗ is an optimal solution.First, consider a linearobjective.Observe that anysolution of ConfIPinducesadecompositionn
=
τi=1ni,andthatthisdecom- position fully determines the objective function, which becomes τ
i=1wini.Furthermore, Part d) ofLemma 4guarantees that we can (almost) restrict our attentionto thespecial sets Yi without ruling out anydecompositionof ninto ni.Second, considering a fixed-chargeobjective,observethattheconstraint (4) andoursep- aratehandlingof
λ
∗(
i,
0)
encodesthisobjectiveappropriately.Regardingruntime,thenumberoftimeswesolvetheILPcon- structed above is equalto the number of guesses ofthe sets Zi andthenumbers
μ ¯
i,whichisboundedby τ i=1(
Si)
O(22(d+D))≤
(
M)
(d+dmax)(
d+
D)
O(d+D)(
log)
(d+D))
τ2O(d+D)≤
(
M+
d+
D+
log)
d+dmax τ2O(d+D)≤ (
M+
d+
D+
log)
τ(d+D)O(1)
.
TheILPwehaveconstructedhasdimensionatmostp
=
τi=1
( |
Zi| + ¯ μ
i(
d+
di)) ≤ τ
22(d+D)+
22(d+D)(
d+
D) ≤ ( τ +
d+
D)
22(d+D), andcanbe solvedintime pO(p)Lˆ
O(1) byKannan’salgorithm [19]forsolvingILPs(recallthat
ˆ
L=
N,
n,
b1, . . . ,
bτ, ,
fmax).Hence, the total runtime is bounded by(
dD Mlog)
τ(d+D)O(1)
ˆ
LO(1). This concludestheproofofthelastpartofthetheorem;thus,wehave establishedTheorem1.Remark.Goemans and Rothvoß prove a similar statement [12, Corollary5.1] to Part4ofTheorem 1,wheretheinputn andthe coefficients w have to be given in unary if one desires an FPT algorithm,whereas inour case they can be given inbinary. The difference is that they invoke the Structure Theorem on a poly- topeP whichisadisjunctiveformulationoftheunionofpolyhe- draP1
∪· · ·∪
Pτ.Thisdisjunctiveconstruction,however,introduces alargecoefficient,increasing.Similarly,alinearobjectivecould be handledintheir setting byintroducing an extra variablexd+1 andsettingxd+1
=
wx,butthisconstraintwouldagainincrease. We circumvent both of these limitations by using the Structure Theoremdirectly.
4. Hardnessof CONFIP
Proposition5.SolvingConfIP
1 isW
[
1]
-hardparameterizedbyd only,evenifisgiveninunaryand noobjectivefunction;
2 withafixed-chargeobjectiveisNP-hard evenwithd
=
1and=
1 (butwithlargepenalties);3 withaseparableconcavequadraticobjectiveisa)NP-hard evenwith d
=
2and=
1,andb)W[
1]
-hardparameterizedbyd evenwhen thelargestcoefficientoftheobjectiveisgiveninunaryand=
1.Proof. Part1. Consider the UnaryBinPacking problem, which takesas inputn items ofsizes o1
, . . . ,
on∈
N withmaxi∈[n]oi=
O≤
poly(
n)
,acapacity B∈
N,andanintegerk∈
N,andweask whethertheitemscanbepackedintokbinsofcapacityB.Jansen etal. [18] have shown that UnaryBinPacking is W[
1]
-hard pa- rameterizedbyk,evenfortightinstanceswhereni=1oi
=
kB. Weshallconstructa ConfIP instancewithntypes.We let Pi, foreachi∈ [
n]
,bedefinedbythesystemofinequalitieskj=1xj
=
oi,
kj=1yj
=
1,
xj≤
O yj, ∀
j∈ [
k] ,
x≥
0.Weletd=
k, di=
k, andμ
i=
1, for each i∈ [
n]
, and define the y variables to be the one which are discarded by the projectionπ
i. Finally, welet nbeak-dimensionalvectorofall B.Itiseasy toseethatπ
i(
Pi) ∩
Zk= {(
oi,
0, . . . ,
0), (
0,
oi,
0, . . . ,
0), . . . , (
0, . . . ,
0,
oi)}
and eachelementencodesthebinintowhichitemi isassigned.Thus, theConfIPinstanceisfeasibleifandonlyifthereexistsanassign- mentofitemstobinssuchthatthesumofitemsizesofeach bin isB.Part3b). We will continue working withthe ConfIP instance constructed above. Recall that minimizing a concave function is equivalenttomaximizingaconvexfunction,whichistheperspec- tive we shalltake here. Ourgoal nowis to modeltheconstraint