1: 

2: %%%%%%%%%%%%%%%%%%%%%%%%%\pagestyle{empty}

3: %

4: %%%%%%%%%%%%%%%%%%%%%

5: 

6: %\documentclass[a4,notitlepage,11pt]{article}

7: \documentstyle[12pt]{article}

8: %%\textheight=21cm \textwidth=15cm \oddsidemargin=0.2in

9: %%\topmargin=-0.1in

10: 

11: %%\pagenumbering{Roman}

12: %\documentstyle[12pt,fullpage]{article}

13: 

14: %\input{defs}

15: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

16: \newtheorem{theorem}{Theorem}

17: \newtheorem{proposition}{Proposition}

18: \newtheorem{corollary}{Corollary}

19: \newtheorem{lemma}{Lemma}

20: \newtheorem{exercise}{Exercise}

21: \newtheorem{claim}{Claim}

22: \newtheorem{consequence}{Consequence}

23: \newtheorem{definition}{Definition}

24: \newtheorem{example}{Example}

25: \newtheorem{conjecture}{Conjecture}

26: \newenvironment{conex}[1]{\vspace{1mm}\noindent Example #1 \em (continued):\ \ }{\vspace{1mm}}

27: \newcommand{\note}[1]{\\  *** {\bf #1} *** \\}  

28: \newenvironment{notation}{\noindent\bf Notation:\em\penalty1000}{}

29: \newenvironment{continue}{\noindent\bf Example {\em (continued)}: \em\penalty1000}{}

30: \newcommand{\short}{\setlength\topsep{0mm}\setlength\partopsep{0mm}\setlength\itemsep{0em}}

31: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

32: \newcommand{\blackslug}{\mbox{\hskip 1pt \vrule width 4pt height 8pt 

33: depth 1.5pt \hskip 1pt}}

34: \newcommand{\QED}{\quad\blackslug\lower 8.5pt\null\par}

35: \newcommand{\proof}{\par\penalty-1000\vskip .5 pt\noindent{\bf Proof\/: }}

36: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

37: \newcommand{\comment}[1]{}

38: \newcommand{\N}{{\em Normal\/\/}}

39: \newcommand{\normal}{{\cal N}_{\bar {\em x\/}}}

40: \newcommand{\normalal}{{\cal N}_{\bar {\em x\/}}^\alpha}

41: \newcommand{\norm}[2]{{\cal N}_{\bar {\em #1\/}}^{#2}}

42: \newcommand{\normm}[2]{{\cal N}_{\em #1\/}^{#2}}

43: \newcommand{\allx} {\forall {\bar {\em x\/}}}

44: \newcommand{\ally} {\forall {\bar {\em y\/}}}

45: \newcommand{\Ex} {\exists {\bar {\em x\/}}}

46: \newcommand{\Ey} {\exists {\bar {\em y\/}}}

47: \newcommand{\bx} {\bar {\em x\/}}

48: \newcommand{\by} {\bar {\em y\/}}

49: \newcommand{\bz} {\bar {\em z\/}}

50: \newcommand{\subs}[2]{[^{#1}_{#2}]}

51: \newcommand{\subsb}[2]{[^{\bar {\em #1\/}}_{\bar {\em #2\/}}]}

52: \newcommand{\na} {\neg \alpha}

53: \newcommand{\nb} {\neg \beta}

54: \newcommand{\nc} {\neg \gamma}

55: \newcommand{\Bm} {\begin{math}}

56: \newcommand{\Em} {\end{math}}

57: \newcommand{\bR} {{\bf R}}

58: \newcommand{\bF} {{\bf F}}

59: \newcommand{\bM} {{\bf M}}

60: \newcommand{\bL} {{\bf L}}

61: \newcommand{\bE} {{\bf E}}

62: \newcommand{\bS} {{\bf S}}

63: \newcommand{\bG} {{\bf \Gamma}}

64: \newcommand{\eqR}{{=}_{\bR}}

65: \newcommand{\smR}{{<}_{\bR}}

66: \newcommand{\smeR}{{\leq}_{\bR}}

67: \newcommand{\grR}{{>}_{\bR}}

68: \newcommand{\Mm} {{\cal M} \models}

69: \newcommand{\Mmf} {{\cal M} \models_{f}}

70: \newcommand{\Mmx}[1] {{\cal M} \models_{#1}}

71: \newcommand{\hal}{\hat {\bf \alpha}}

72: \newcommand{\hab}{\hat {\bf \beta}}

73: \newcommand{\dom}{\|\cM\|}

74: \newcommand{\nca}{{\bf NC}(\alpha )}

75: \newcommand{\cR}{$\cal R$}

76: \newcommand{\ncR}{$\notR$}

77: \newcommand{\bxp}{\bar_{{\em x}\prime}}

78: \newcommand{\xp}{{\em x}\prime}

79: \newcommand{\ba}{$\bar{\alpha}$}

80: \newcommand{\bb}{$\bar{\beta}$}

81: \newcommand{\hb}{$\hat{\beta}$}

82: \newcommand{\ha}{$\hat{\alpha}$}

83: \newcommand{\gaR}{$\alpha{\cal R}$}

84: \newcommand{\ganR}{$\alpha\notR$}

85: \newcommand{\gbR}{$\beta{\cal R}$}

86: \newcommand{\gbnR}{$\beta\notR$}

87: \newcommand{\CB}{C\hspace{-1.00mm}B}

88: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

89: %\newcommand{\SEP}{\makebox[0in]{\rule{.5mm}{4.5mm}}}

90: \newcommand{\ru}{\rule[-0.4mm]{.1mm}{3mm}}

91: \newcommand{\nni}{\ru\hspace{-3.5pt}}

92: \newcommand{\nnj}{\ru\hspace{-4.5pt}}

93: \newcommand{\sni}{\ru\hspace{-1pt}}

94: \newcommand{\pre}{\hspace{0.28em}}

95: \newcommand{\post}{\hspace{0.1em}}

96: \newcommand{\NIm}{\pre\nni\sim}

97: \newcommand{\NI}{\mbox{$\: \nni\sim$ }}

98: \newcommand{\notNIm}{\pre\nni\not\sim}

99: \newcommand{\notNI}{\mbox{ $\nni\not\sim$ }}

100: \newcommand{\NIVm}{\pre\nni\sim_V}

101: \newcommand{\NIV}{\mbox{ $\nni\sim_V$ }}

102: \newcommand{\notNIVm}{\pre\sni{\not\sim}_V\post}

103: \newcommand{\notNIV}{\mbox{ $\sni{\not\sim}_V$ }}

104: \newcommand{\NIWm}{\pre\nni\sim_W}

105: \newcommand{\NIW}{\mbox{ $\nni\sim_W$ }}

106: \newcommand{\NIWp}{\mbox{ $\nni\sim_{W'}$ }}

107: \newcommand{\notNIWm}{\pre\sni{\not\sim}_W\post}

108: \newcommand{\notNIW}{\mbox{ $\sni{\not\sim}_W$ }}

109: \newcommand{\eem}{\hspace{0.8mm}\rule[-1mm]{.1mm}{4mm}\hspace{-4pt}}

110: \newcommand{\EM}{\eem\equiv}

111: \newcommand{\notEM}{\eem\not\equiv}

112: \newcommand{\notR}{\not {\hspace{-1.5mm}{\cal R}}}

113: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

114: \newcommand{\bK}{{\bf K}}

115: \newcommand{\bKp}{${\bf K}^p$}

116: \newcommand{\oK}{$\overline {\bf K}$}

117: \newcommand{\bP}{{\bf P}}

118: \newcommand{\ga}{\mbox{$\alpha\ $}}

119: \newcommand{\gb}{\mbox{$\beta\ $}}

120: \newcommand{\gc}{\mbox{$\gamma\ $}}

121: \newcommand{\gd}{\mbox{$\delta\ $}}

122: \newcommand{\gS}{\mbox{$\Sigma\ $}}

123: \newcommand{\gep}{\mbox{$\varepsilon$}}

124: \newcommand{\gf}{\mbox{$\zeta$}}

125: %\newcommand{\ga}{\mbox{\''{a}}}

126: \newcommand{\cA}{\mbox{${\cal A}$}}

127: \newcommand{\cB}{\mbox{${\cal B}$}}

128: \newcommand{\cC}{\mbox{${\cal C}$}}

129: \newcommand{\cD}{\mbox{${\cal D}$}}

130: \newcommand{\cE}{\mbox{${\cal E}$}}

131: \newcommand{\cF}{\mbox{${\cal F}$}}

132: \newcommand{\cG}{\mbox{${\cal G}$}}

133: \newcommand{\cK}{\mbox{${\cal K}$}}

134: \newcommand{\cL}{\mbox{${\cal L}$}}

135: \newcommand{\cI}{\mbox{${\cal I}$}}

136: \newcommand{\cM}{\mbox{${\cal M}$}}

137: \newcommand{\cN}{\mbox{${\cal N}$}}

138: \newcommand{\cO}{\mbox{${\cal O}$}}

139: \newcommand{\cP}{\mbox{${\cal P}$}}

140: \newcommand{\cS}{\mbox{${\cal S}$}}

141: \newcommand{\cU}{\mbox{${\cal U}$}}

142: \newcommand{\cW}{\mbox{${\cal W}$}}

143: \newcommand{\ab}{\mbox{\ga \NI \gb}}

144: \newcommand{\cd}{\mbox{\gc \NI \gd}}

145: \newcommand{\ef}{\mbox{\gep \NI \gf}}

146: \newcommand{\xe}{\mbox{$\xi$ \NI $\eta$}}  

147: \newcommand{\pht}{\mbox{$\varphi$ \NI $\theta$}}

148: \newcommand{\rt}{\mbox{$\rho$ \NI $\tau$}}

149: \newcommand{\Cn}{\mbox{${\cal C}n$}}

150: \newcommand{\Pf}{\mbox{${\cal P}_{f}$}}

151: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

152: \newcommand{\ra}{\rightarrow}

153: \newcommand{\Ra}{\Rightarrow}

154: \newcommand{\eqdef}{\stackrel{\rm def}{=}}

155: \newcommand{\absv}[1]{\mid #1 \mid}

156: \newcommand{\vstar}{\mbox{$V\sstar_{\infty}$}}

157: \newcommand{\sumstar}{\mbox{$\sum\sstar$}}

158: \newcommand{\tilh}{\mbox{$\tilde{h}$}}

159: \newcommand{\tilep}{\mbox{$\tilde{\varepsilon}$}}

160: \newcommand{\tilf}{\mbox{$\tilde{f}$}}

161: \newcommand{\gahat}{\mbox{$\hat{\ga}$}}

162: \newcommand{\gafalse}{\mbox{$\ga\NI{\bf false}$}}

163: \newcommand{\sstar}{^{*}}

164: \newcommand{\calR}{\mbox{${\cal R}\sstar$}}

165: \newcommand{\eq}{=}

166: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

167: \newcommand{\st}[2]{(\star\ #1\ #2)}

168: \newcommand{\str}{$\star$}

169: \newcommand{\dx}{\$ x}

170: \newcommand{\dy}{\$ y}

171: \newcommand{\dz}{\$ z}

172: \newcommand{\du}{\$ u}

173: \newcommand{\dv}{\$ v}

174: \newcommand{\imm}{\leftrightarrow}

175: \newcommand{\Imm}{\Leftrightarrow}

176: \newcommand{\seq}{$x_1,\ldots ,x_n$}

177: \newcommand{\se}[1]{{#1}_1,\ldots ,{#1}_n}

178: \newcommand{\imp}{\Rightarrow}

179: \newcommand{\dip}{\diamond _{pma}}

180: \newcommand{\domn}[1]{$\,\succ _{_#1}$}

181: \newcommand{\lek}{\preceq _{\cal K}}

182: \newcommand{\lk}{\prec _{\cal K}}

183: \newcommand{\vspc}{\vspace{0.5cm}}

184: \newcommand{\binom}[2]   {{#1\choose#2}}

185: %%%%%robot motion%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

186: \newcommand{\allnec}{\forall _{\Box}}

187: \newcommand{\allpos}{\forall _{\Diamond}}

188: \newcommand{\somenec}{\exists _{\Box}}

189: \newcommand{\somepos}{\exists _{\Diamond}}

190: \newcommand{\pallnec}{\forall _{\pnec}}

191: \newcommand{\pallpos}{\forall _{\hspace{-2mm}\ppos}\hspace{-1.5mm}}

192: \newcommand{\psomenec}{\exists _{\pnec}}

193: \newcommand{\psomepos}{\exists _{\hspace{-2mm}\ppos}\hspace{-1.5mm}}

194: \newcommand{\ballnec}{\forall _{\bnec}}

195: \newcommand{\ballpos}{\forall _{\bpos}}

196: \newcommand{\bsomenec}{\exists _{\bnec}}

197: \newcommand{\bsomepos}{\exists _{\bpos}}

198: \newcommand{\bpallnec}{\forall _{\overline{\pnec}}}

199: \newcommand{\bpallpos}{\forall _{\bppos}}

200: \newcommand{\bpsomenec}{\exists _{\overline{\pnec}}}

201: \newcommand{\bpsomepos}{\exists _{\bppos}}

202: \newcommand{\bpnec}{\overline{\pnec}}

203: \newcommand{\pnec}

204: {\setlength{\unitlength}{0.0025in}%

205: \begin{picture}(20,20)(760,800)

206: \thinlines

207: \put(760,800){\framebox(20,20){}}

208: \put(765,810){\line( 1, 0){ 10}}

209: \end{picture}}

210: 

211: \newcommand{\ppos}

212: {\setlength{\unitlength}{0.00375in}%

213: \begin{picture}(60,20)(60,520)%40,20

214: \thinlines

215: \put( 60,540){\line( 0, 1){  0}}

216: \multiput( 80,530)(0.40000,0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

217: \multiput( 90,540)(0.40000,-0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

218: \put(100,530){\line( 0, 1){  0}}

219: \put( 60,540){\line( 0, 1){  0}}

220: \put( 85,535){\line( 0, 1){  0}}

221: \put( 85,535){\line( 0, 1){  0}}

222: \multiput( 80,530)(0.40000,-0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

223: \multiput( 90,520)(0.40000,0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

224: \put( 85,530){\line( 1, 0){ 10}}

225: \end{picture}}

226: 

227: \newcommand{\bpos}

228: {\setlength{\unitlength}{0.005in}%

229: \begin{picture}(20,25)(70,795)

230: \thinlines

231: \multiput( 80,815)(-0.40000,-0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

232: \multiput( 70,805)(0.40000,-0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

233: \multiput( 80,795)(0.40000,0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

234: \multiput( 90,805)(-0.40000,0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

235: \put( 75,820){\framebox(10,0){}}

236: \end{picture}

237: }

238: 

239: \newcommand{\bnec}

240: {\setlength{\unitlength}{0.0075in}%

241: \begin{picture}(10,15)(80,790)

242: \thinlines

243: \put( 80,790){\framebox(10,10){}}

244: \put( 80,805){\framebox(10,0){}}

245: \end{picture}

246: }

247: \newcommand{\bppos}

248: {\setlength{\unitlength}{0.005in}%

249: \begin{picture}(20,25)(70,795)

250: \thinlines

251: \multiput( 80,815)(-0.40000,-0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

252: \multiput( 70,805)(0.40000,-0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

253: \multiput( 80,795)(0.40000,0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

254: \multiput( 90,805)(-0.40000,0.40000){26}{\makebox(0.4444,0.6667){\sevrm .}}

255: \put( 80,815){\line( 0, 1){  0}}

256: \put( 75,805){\line( 1, 0){ 10}}

257: \put( 75,820){\line( 1, 0){ 10}}

258: \end{picture}

259: }

260: 

261: %%%%%%%%%%%%%%%%%%%%%%%%%%%%

262: %\newcommand{\cons}{\NIm}

263: \newcommand{\cons}{\joinrel{\mid}{\sim}\;}

264: \newcommand{\dcons}{\joinrel{\mid}{\sim}_{\scriptsize \Delta}\;}

265: %\newcommand{\cons2}{|\nnj\sim}

266: \newcommand{\nocons}{\pre\nni\not\sim}

267: 

268: \newcommand{\mat}[4]

269: {

270: \begin{tabular}{cc||c|c||}

271: \multicolumn{3}{r}{$s_1$}&\multicolumn{$s_2$}{c}{}\\ 

272: \	&\ 	& 1 	& 2 \\  \cline{2-4} \cline{2-4}

273: $act_1$	&1 	& #1	& #2 \\ \cline{2-4}

274: $act_2$	&2	& #3 	& #4 \\  \cline{2-4}

275: \end{tabular}

276: }

277: 

278: \newcommand{\prisd}[4]

279: {

280: \begin{center}

281: \begin{tabular}{c||c|c||}

282: \multicolumn{3}{r}{teacher} \\ \cline{2-3}

283: student	& Coop	& Defect\\ \hline\hline 

284: Coop    & #1	& #2 \\ \hline

285: Defect  & #3 	& #4 \\  \hline

286: \end{tabular}

287: \end{center}

288: }

289: 

290: \newcommand{\matd}[4]

291: {

292: \begin{center}

293: \begin{tabular}{c||c|c||}

294: \multicolumn{3}{r}{teacher} \\ \cline{2-3}

295: student	& I	& II\\ \hline\hline 

296: 1    	& #1	& #2 \\ \hline

297: 2	& #3 	& #4 \\  \hline

298: \end{tabular}

299: \end{center}

300: }

301: 

302: \newcommand{\denselist}{\itemsep 0pt} %still separated by \parsep

303: \newcommand{\closelist}{\setlength{\partopsep}{-1pt}

304:                         \setlength{\topsep}{\.5topsep}}

305: \hfuzz=2pt  %ignore small overflows

306: \newenvironment{closeitems}{\begin{itemize}\vspace{-.5\topsep}\denselist}{\vspace{-.4\topsep}\end{itemize}}

307: \newenvironment{closenumerals}{\begin{itemize}\vspace{-.5\topsep}\denselist}{\vspace{-.4\topsep}\end{itemize}}

308: 

309: 

310: 

311: 

312: 

313: 

314: 

315: \def\eqnum#1{\eqno (#1)}

316: \def   \ni     {\noindent}

317: \renewcommand{\baselinestretch}{1.2}

318: %%%%%%%%%%%%%%%%%%%%%%%

319: %%\pagestyle{empty}

320: 

321: 

322: \title{

323:  Bundling Equilibrium in Combinatorial Auctions\footnote{First version: June 2001.}}

324: 

325: %\author{

326: %Ron Holzman, Noa Kfir-Dahav*, Dov Monderer** and Moshe Tennenholtz*** \\

327: %Faculty of Industrial Engineering and Management \\

328: %Technion -- Israel Institute of Technology \\

329: % Haifa 32000, Israel\\

330: % RON EMAIL\\

331: % * kfir@ie.technion.ac.il\\

332: % ** dov@ie.technion.ac.il\\

333: % ***MOSHE EMAIL}

334: 

335: 

336: %\title{Bundling equilibrium: social surplus and communication complexity

337: %in combinatorial auctions}

338: 

339: \author{

340:      \begin{tabular}{ccc}

341:          \large Ron Holzman\thanks{Research supported

342:          by the Fund for the Promotion of Research at

343:          the Technion and by Technion V.P.R. Fund -

344:          M. and M. L. Bank Mathematics Research Fund}

345:          & \hspace{.2in} & \large Noa Kfir-Dahav\\

346:   \small Department of Mathematics  & &

347:  \small  Faculty of IE and Management\\

348:  \small Technion--Israel Institute of Technology  & &

349:  \small Technion--Israel Institute of Technology\\

350:  \small Haifa 32000, Israel  & &

351:  \small Haifa 32000, Israel\\

352:   \small holzman@tx.technion.ac.il & &

353:   \small kfir@ie.technion.ac.il\\

354:    \small  & & \\

355:    \small  & & \\

356:          \large Dov Monderer\thanks{Research supported

357:          by the Fund for the Promotion of Research at

358:          the Technion, and by the Israeli Academy of Science.} & \hspace{.2in} & \large Moshe Tennenholtz\\

359:  \small  Faculty of IE and Management & &

360:  \small  Computer Science Department\\

361:  \small Technion--Israel Institute of Technology  & &

362:  \small Stanford University\\

363:  \small Haifa 32000, Israel  & &

364:  \small Stanford, CA 94305, USA\\

365:   \small dov@ie.technion.ac.il & &

366:    \small moshe@robotics.stanford.edu\\

367:      \end{tabular}

368:   }

369: 

370: 

371: 

372: 

373: 

374: 

375: \date{September 18, 2001}

376: 

377: 

378: 

379: 

380: 

381: 

382: \begin{document}

383: 

384: 

385: 

386: 

387: 

388: \maketitle

389: 

390: \newpage

391: 

392: \noindent{\bf Abstract:}

393: 

394: This paper analyzes  individually-rational ex post equilibrium in

395: the VC (Vickrey-Clarke) combinatorial auctions. If $\Sigma$ is a

396: family of bundles of goods, the organizer may restrict the

397: participants by requiring them to submit their bids only for

398: bundles in $\Sigma$. The $\Sigma$-VC combinatorial auctions

399: (multi-good auctions) obtained in this way are known to be

400: individually-rational truth-telling mechanisms. In contrast, this

401: paper deals with non-restricted VC  auctions, in which the

402: buyers restrict themselves to bids on bundles in $\Sigma$, because

403: it is rational for them to do so. That is,  it may be that when

404: the buyers report their valuation of the bundles in $\Sigma$, they

405: are in an equilibrium. We fully characterize those $\Sigma$ that

406: induce individually rational equilibrium in every VC auction, and

407: we refer to the associated equilibrium as a bundling equilibrium.

408: The number of bundles in $\Sigma$ represents the communication

409: complexity of the equilibrium. A special case of bundling

410: equilibrium is partition-based equilibrium, in which $\Sigma$ is a

411: field, that is, it is generated by a partition. We analyze

412:  the tradeoff between communication complexity and economic

413:  efficiency of bundling equilibrium, focusing in particular on

414:  partition-based equilibrium.

415: 

416: 

417: 

418: 

419: 

420: 

421: \section{Introduction}

422: 

423: The Vickrey-Clarke-Groves (VCG) mechanisms

424: \cite{Vickrey,Clarke,Groves} are central to the design of

425: protocols with selfish participants (e.g.,

426: \cite{NisanRonenGames,Tijcai,Varian95}), and in particular for

427: combinatorial auctions (e.g.,

428: \cite{Weber,KrishnaPerry,deVriesVohra,Wellmangeb,MonTenAsymp,Nisan00,Lehmann99}), in which the

429: participants submit bids, through which they can express

430: preferences over bundles of goods. The organizer allocates the

431: goods and collects payments based on the participants'

432: bids.\footnote{Motivated by the FCC auctions (see e.g.,

433: \cite{Cramton,McMillan,Milgromfcc} ) there is an extensive recent

434: literature devoted to the design and analysis of multistage

435: combinatorial auctions, in which the bidders express partial

436: preferences over bundles at each stage. See

437: e.g.,\cite{Wellmangeb,perryrenyb,Ausubel2000,Parkes99,ParkesUngar,AusubelMilgrom}

438: .} These protocols allow to allocate a set of goods (or services,

439: or tasks) in a socially optimal (surplus maximizing) manner,

440: assuming there are no resource bounds on the agents' computational

441: capabilities.\footnote{There are  at least two sources of

442: computational issues, which arise when dealing with combinatorial

443: auctions; Winner determination --finding the optimal allocation

444: (see e.g.,\cite{RothPekHar,Nisan00,Tenn,Sandholm99,FujiBrownShoham,Anderson,Sandholm01,HoosBoutilier})  , and bid communication --

445: the transfer of information (see e.g.,

446: \cite{nisancommunication}).} The VCG protocols are designed in a

447: way that truth-revealing of the agents' private

448: information\footnote{This paper deals with the private-values

449: model, in which every buyer knows his own valuations of bundles

450: of goods. In contrast, in a correlated-values model, every buyer

451: receives a signal (possibly about all buyers' valuation

452: functions), and this signal does not completely reveal his own

453: valuation function (see e.g.

454: \cite{MilgromWeber,jehielmsignals,mcafeereny,dasmaskin,perryrenya,perryrenyb}

455: for

456:   discussions of

457:  models

458: in which valuations are correlated).} is a dominant strategy to

459: them. Moreover, VCG protocols can be applied in the context of

460: games in informational form, where no probabilistic assumptions

461: about agents' types are required.\footnote{ A game in

462: informational form is a pre-Bayesian game. That is, it has all

463: the ingredients of a Bayesian game except for the specification

464: of probabilities. Unlike Bayesian games, games in informational

465: form do not necessarily possess a solution: a recommendation for

466: rational players how to play. However, in many important models

467: such solutions do exist. See Section 2 for a precise definition.}

468: We shortly define domination and equilibrium in such games. These

469: solutions are called ex post solutions because they have the

470: property that if the players were told about the true state,

471: after they choose their actions, they would not regret their

472: actions.\footnote{Alternatively, ex post solutions may be called

473: probability-independent solutions because, up to some

474: technicalities concerning the concept of measurable sets, they

475: form Bayesian solutions for every specification of probabilities.}

476: 

477: 

478: 

479: In this paper we deal with a special type of VCG  mechanisms --

480: the VC mechanisms. Amongst the VCG mechanisms the VC mechanisms

481: are characterized by two additional important properties: Truth

482: telling satisfies the participation constraint, that is, it is

483: preferred to non participation,\footnote{ An equilibrium that

484: satisfies the participation constraint  is said to be

485: individually- rational.} and the seller's revenue is always non

486: negative.

487: 

488: 

489:  A famous observation of the theory of

490: mechanism design in economics, termed {\bf the revelation

491: principle} (see e.g. \cite{Myerson79}), implies that the

492: discussion of additional individually rational (IR) equilibria of

493: the VC mechanisms may seem unneeded, and indeed it has been

494: ignored by the literature.  It can be proved that every mechanism

495: with an ex post equilibrium is economically equivalent to another

496: mechanism -- a direct mechanism -- in which every agent is

497: required to submit his information. In this direct mechanism,

498: revealing the true type is an ex post dominating strategy for

499: every agent, and it yields the same economics parameters as the

500: original mechanism. However, the two mechanisms differ in the set

501: of inputs that the player submits in equilibrium. This difference

502: may be crucial when we deal with communication complexity. Thus,

503: two mechanisms that are equivalent from the economics point of

504: view, may be considered different mechanisms from the CS point of

505: view.

506: 

507: Thus,  tackling the VC mechanisms from a computational perspective

508: introduces a vastly different picture. While the revelation of the

509: agents' types  defines one IR equilibrium, there are other (in

510: fact, over-exponentially many) IR equilibria for the VC auctions.

511: Moreover, these equilibria have different communication

512: requirements.

513: 

514: 

515: In this paper we analyze ex post equilibria in the VC mechanisms.

516: 

517: 

518: 

519: 

520: 

521: 

522: Let $\Sigma$ be a family of bundles of goods. We characterize

523: those $\Sigma$, for which the strategy of reporting the true

524: valuation over the bundles in $\Sigma$ is a player-symmetric IR ex

525: post equilibrium.  An equilibrium that is defined by such $\Sigma$

526: is called a bundling equilibrium. We prove that $\Sigma$ induces a

527: bundling equilibrium if and only if  it is a quasi field of

528: bundles.\footnote{A quasi field is a nonempty set of sets that is

529: closed under complements and under disjoint unions.} The number of

530: bundles in $\Sigma$ represents the communication complexity of the

531: equilibrium, and the  economic efficiency of an equilibrium is

532: measured by the generated social surplus.

533:    A special type of bundling

534: equilibria are partition-based equilibria, in which $\Sigma$ is a

535: field (i.e. it is generated by a partition).  The partition-based

536: equilibria are ranked according to the usual partial order on

537: partitions: If one partition is finer than another one, then it

538: yields higher communication complexity as well as higher social

539: surplus.\footnote{  It is worth mentioning that the various

540: equilibria cannot be ranked according to the revenue of the

541: seller. That is, under some conditions, a partition-based

542: equilibrium may simultaneously yield more revenue and less

543: communication complexity than the truth revealing equilibrium (an

544: example is provided).}

545: 

546: We analyze the least upper bound  (over all possible profiles of

547: valuation functions, one for each buyer) of the ratio between the

548: optimal surplus and the surplus obtained in a partition-based

549: equilibrium. We express this least upper bound in terms of the

550: partition's structure. We provide an upper bound for this ratio,

551: which is proved to be tight in infinitely many cases.

552: 

553: 

554: 

555: 

556: 

557: In Section~2 we present the concept of ex post equilibrium in

558: games in informational form. In Section 3 we discuss combinatorial

559: auctions. Together, Sections 2 and 3 provide the reader with a

560: rigorous framework for general analysis of VC protocols for

561: combinatorial auctions.  In Section 4  we introduce bundling

562: equilibrium, and provide a full characterization of bundling

563: equilibria for VC protocols.

564:  Then we discuss bundling equilibrium that is generated by a partition,

565:  titled  partition-based equilibrium.

566:  In Section 5 we deal with the surplus of VC protocols

567: for combinatorial auctions

568: when following partition-based equilibrium, exploring the spectrum between

569: economic efficiency and communication efficiency.

570: 

571: \section{Ex post equilibrium in games in informational form}

572:  A game in  {\bf informational form} $G=G(N,\Omega,T,(\tilde t_i)_{i\in N},X,(u_i)_{i\in N})$

573:   is defined

574:  by the following parameters:

575: 

576:  $\bullet$ {\bf Agents}: Let $N=\{1,\ldots,n\}$ be the

577:  set of agents.

578: 

579: $\bullet$ {\bf States}: Let $\Omega$ be the set of  (relevant)

580: states.

581: 

582: 

583: $\bullet$ {\bf Types}: Let $T_i$ be the set of types of agent $i$,

584: $T=\times_{i\in N}T_i$.

585: 

586: $\bullet$ {\bf Signaling functions}: Let $\tilde t_i:\Omega\to

587: T_i$ be the signaling function of agent $i$. Without loss of

588: generality, it is assumed that every type $t_i\in T_i$ is

589: possible. That is $\tilde t_i(\Omega)=T_i$.

590: 

591: $\bullet$ {\bf Actions}: Let $X_i$ be the set of actions of $i$,

592: $X=\times_{i\in N}X_i.$

593: 

594: 

595: $\bullet$ {\bf Utility functions}: Let $u_i(\omega, x)$ be the

596: utility of $i$ at state $\omega$, when the agents choose the

597: action profile $x$.

598: 

599:  Every $\omega\in\Omega$ defines a game in strategic

600: (normal) form, $G(\omega)$. In this game agent $i$ receives

601: $u_i(\omega, x)$, when it chooses $x_i$, and all other agents

602: choose $x_{-i}$. However, the agents do not know which game they

603: play.

604: 

605: For $t_i\in T_i$ let $\Omega_i(t_i)$ be the set of states that

606: generate the signal $t_i$, that is

607:  $$\Omega_i(t_i)=\{\omega\in\Omega|\tilde t_i(\omega)=t_i\}.$$

608: 

609:   A {\bf strategy}\footnote{In this paper

610:   we do not deal with mixed strategies.}

611:    of $i$ is a function $b_i:T_i\to X_i$;

612:  The associated {\bf implied strategy} is the function $\hat

613:  b_i:\Omega\to X_i$ given by

614:  $$\hat b_i(\omega)=b_i(\tilde t_i(\omega)).$$

615: %We denote $b(v)=(b_i(v_i))_{i \in N}$ and

616: %$b_{-i}(v_{-i})=(b_j(v_j))_{j \in N\setminus \{i\}}$.

617: 

618:  A profile of strategies $b=(b_1,\cdots,b_n)$ is an {\bf

619:  ex post equilibrium}, if for every agent $i$, for every $t_i\in T_i$, for

620: every $\omega\in\Omega_i(t_i)$, and for every $x_i\in X_i$,

621: $$u_i(\omega,b_i(t_i),\hat b_{-i}(\omega))\ge u_i(\omega,x_i,\hat

622: b_{-i}(\omega)).$$

623: 

624: A strategy $b_i$ of $i$ is an {\bf ex post dominant strategy} for

625: $i$, if for every profile of strategies $b_{-i}$ of the other

626: players, for every  $t_i\in T_i$, for every

627: $\omega\in\Omega_i(t_i)$, and for every $x_i\in X_i$,

628: $$u_i(\omega,b_i(t_i),\hat b_{-i}(\omega))\ge u_i(\omega,x_i,\hat

629: b_{-i}(\omega)).$$ Obviously, if $b_i$ is an  ex post dominant

630: strategy for every $i$, $b$ is an ex post equilibrium, but not

631: necessarily  vice versa. An ex post equilibrium $b$, in which

632: every strategy $b_i$ is ex post dominant is called an ex post {\bf

633: domination equilibrium}.

634: 

635: 

636: 

637: 

638: 

639: \section{Combinatorial auctions}

640: In a combinatorial auction there is a seller, denoted by $0$, who

641: wishes to sell a set of  $m$ items $A=\{a_1,\ldots,a_m\}$, $m\ge

642: 1$, that are owned by her. There is a set of (potential) buyers

643: $N=\{1,2,\ldots,n\}$, $n\ge 1$. We take $N$ as the set of agents.

644: Let $\Gamma$ be the set of all allocations of the goods. That is,

645: every $\gamma\in \Gamma$ is an ordered partition of $A$,

646: $\gamma=(\gamma_i)_{i\in N\cup\{0\}}$.

647: % An allocation $\gamma$ is {\bf buyer complete} if $\gamma_0=\emptyset$.

648: A {\bf valuation function} of buyer $i$ is a function $v_i:2^A\to

649: \Re$, where $\Re$ denotes the set of real numbers, with the

650: normalization $v_i(\emptyset)=0$. In a more general setup, a buyer

651: may care about the distribution of goods that he does not own. In

652: such a setup the utility  of an agent may depend on the whole

653: allocation $\gamma$ rather than on $\gamma_i$. Hence, by dealing

654: with valuation functions we actually assume:

655: 

656: $\bullet$ {\bf No allocative externalities}.\footnote{For auctions

657: in which externalities are assumed see, e.g.,

658: \cite{jmols,jmolsAER}.}

659: 

660: We also assume:

661: 

662: $\bullet$ {\bf Free disposal}: If $B\subseteq C$, $B,C\in 2^A$,

663: then $v_i(B)\le v_i(C)$.

664: 

665: Let $V_i$ be the set of all possible valuation functions of $i$

666: (obviously $V_i=V_j$ for all $i,j\in N$), and let $V=\times_{i\in

667: N}V_i$. We refer to $V$ as the set of {\bf states} ($\Omega=V$).

668: In a general model, each buyer receives a signal $t_i$  through a

669: signaling function $\tilde t_i$ defined on $V$. We assume:

670: 

671: $\bullet$  {\bf Private value model}: $\tilde t_i(v)=v_i$. That

672: is, $T_i=V_i$ and each buyer knows his valuation function

673: only.\footnote{ See e.g.

674:  \cite{MilgromWeber,jehielmsignals,mcafeereny,dasmaskin,perryrenya,perryrenyb} for

675:   discussions of

676:  models

677: in which valuations are correlated and buyers do not know their

678: own valuation.}

679: 

680: 

681: A {\bf mechanism}  for allocating the goods is defined by sets of

682: messages $X_i$, one set for each buyer $i$, and by a pair $(d,c)$

683: with $d:X\to \Gamma$, and $c:X\to \Re^n$, where $X=\times X_i$.

684: $d$ is called the allocation function and $c$ the transfer

685: function; if the buyers send the profile of messages $x\in X$,

686: buyer $i$ receives the set of goods $d_i(x)$ and pays $c_i(x)$ to

687: the seller. We assume:

688: 

689: $\bullet$  {\bf Quasi linear utilities}: If agent $i$ with the

690: valuation function $v_i$ receives the set of goods $\gamma_i$ and

691: pays $c_i$, his utility equals $v_i(\gamma_i)-c_i$.

692: 

693: As the seller cannot force the buyer to participate, a full

694: description of a mechanism should describe the allocation of goods

695: and transfers for cases in which not all agents participate.

696: However, we adopt the way this issue is treated in economics: The

697: mechanism $(X,d,c)$ defines a game in informational form. An ex

698: post equilibrium $b$ in this game satisfies the {\bf participation

699: constraint} if for every buyer $i$, $$v_i(d_i(b(v)))-c_i(b(v))\ge

700: 0\quad\mbox{for every $v \in V$,}\eqno{(3.1)}$$ where

701: $b(v)=(b_1(v_1), \ldots ,b_n(v_n))$. If an ex post equilibrium

702: satisfies the participation constraint, we call it {\bf

703: individually rational}. If the buyers use an individually rational

704: ex post equilibrium profile $b$, then a deviation of a buyer to

705: non participation is not profitable for him.\footnote{ Thus, $b$

706: remains an ex post equilibrium profile if every set of messages is

707: extended by a null message, and an agent whose input is null

708: receives no good and pays nothing. Nevertheless, if the issue of

709: uniqueness of equilibrium is important, dealing with ex post

710: individually rational equilibrium instead of dealing with ex post

711: equilibrium in the extended model is not without loss of

712: generality. The extended model may have more equilibrium profiles,

713: that cannot be expressed in the reduced model, that is,

714: equilibrium profiles in which some of the agents do not

715: participate in some of the cases.} Similarly, a dominant strategy

716: of $i$, $b_i$, is individually rational if (3.1) is satisfied for

717: every profile $b_{-i}$ of the other buyers' strategies.

718: 

719:  For an allocation $\gamma$ and a profile of types

720: $v$ we denote by $S(v,\gamma)$ the {\bf total social surplus} of

721: the buyers, that is $$S(v,\gamma)=\sum_{i\in N} v_i(\gamma_i).$$

722: We also denote: $$S_{max}(v)=\max_{\gamma\in \Gamma}S(v,\gamma).$$

723: 

724: 

725: Consider a mechanism $M=(X,d,c)$ and an individually rational ex

726: post equilibrium $b$.

727:  For every profile $v$ we denote the surplus generated by $b$ by

728: $S^M_b(v)=S(v,d(b(v)))$, and the revenue collected by the seller

729: by $R^M_b(v)=\sum_{i\in N}c_i(b(v))$.

730: 

731: Because of the participation constraint, $$R^M_b(v)\le S^M_b(v)\le

732: S_{max}(v)\quad\mbox{for all $v\in V$.}$$ A mechanism and an

733: individually rational ex post equilibrium $(M,b)$ are called {\bf

734: socially optimal} if $$S^M_b(v)=S_{max}(v)\quad\mbox{ for all

735: $v\in V$.}$$

736: 

737: Note that the seller controls the mechanism, but she does not

738: control the strategies used by the buyers. However, it is assumed

739: that if the mechanism possesses an individually rational ex post

740: equilibrium, the agents use such an equilibrium.\footnote{ For

741: example,  the agents may reach the equilibrium by a process of

742: learning  (see e.g. \cite{HonSela}).}

743: %Moreover, if the game defined by the mechanism

744: %has multiple equilibria, the seller can recommend a particular

745: %equilibrium profile ( e.g., by supplying software that utilize

746: %this equilibrium).

747: 

748: A public seller may wish to generate a socially optimal mechanism,

749: whereas a selfish seller may be interested in the revenue function

750: only. Such a seller would rank mechanisms according to the revenue

751: they generate.

752: 

753: A mechanism $(X,d,c)$ is called a  {\bf direct} mechanism if

754: $X_i=V_i$ for every $i\in N$. That is, in a direct mechanism a

755: buyer's message contains a full description of some valuation

756: function. A direct mechanism is called {\bf truth revealing} if

757: for every buyer $i$, telling the truth ($b_i(v_i)=v_i$) is an

758:  individually rational ex post dominating strategy. (Of course the profile of

759:  strategies $b=(b_i)_{i\in N}$ is an individually rational

760: ex post equilibrium.) By the revelation principle,\footnote{See

761: e.g. \cite{ Myerson79}.} given a mechanism and an individually

762: rational ex post equilibrium $(M,b)$ one can find a direct truth

763: revealing mechanism that yields the same distribution of goods and

764: the same payments (and in particular, the same revenue and surplus

765: functions).\footnote{ This strong version of the revelation

766: principle is due to our private values assumption. Otherwise, the

767: revelation principle guarantees the existence of an equivalent

768: direct mechanism in which telling the truth is an individually

769: rational ex post equilibrium (but not necessarily a domination

770: equilibrium).} It may seem therefore that the concept of ex post

771: equilibrium is not interesting in our setup (private values), and

772: indeed most of the economics literature of mechanism design with

773: private values deals only with direct and truth revealing

774: mechanisms.

775:  However, when we deal with computational issues, two mechanism

776: that are equivalent in economics may differ in their complexity.

777: The time, space, and  communication required to compute and communicate

778: the message of

779: an agent  as well as the chosen allocation may depend on the

780: messages sent in equilibrium. Thus, the concept of ex post

781: equilibrium may be very important even if private values are

782: assumed.

783: 

784:  Well-known  truth revealing mechanisms are the VC

785: mechanisms.  These mechanisms are parameterized by an allocation

786: function $d$, that is socially optimal. That is,

787: $S_{max}(v)=S(v,d(v))$ for every $v\in V$.  The transfer functions

788: are defined as follows: $$c^d_i(v)=\max_{\gamma\in

789: \Gamma}\sum_{j\ne i}v_j(\gamma_j)-\sum_{j\ne

790: i}v_j(d_j(v)).\eqno{(3.2)}$$ Note that $c_i^d(v) \ge 0$ for every

791: $v \in V$.

792: 

793: The mechanisms differ in the allocation they pick in cases in

794: which there exist more than one socially optimal allocation, and

795: therefore in the second term in (3.2).\footnote{By VC mechanisms

796: we refer here to what is also known as Clarke mechanisms or the

797: Pivotal mechanism. More general mechanisms are the VCG mechanisms.

798: Every VCG mechanism is obtained from some VC mechanism by changing

799: the transfer functions: A VCG mechanism is defined by a socially

800: optimal allocation function $d$ and by a family of functions

801: $h=(h_i)_{i\in N}$. The transfer functions are defined by:

802: $$c^d_i(v)=\max_{\gamma\in \Gamma}\sum_{j\ne

803: i}v_j(\gamma_j)-\sum_{j\ne i}v_j(d_j(v))+h_i(v_{-i}).$$ Truth

804: telling is an ex post equilibrium in every VCG mechanism, but it

805: is not necessarily an individually rational ex post equilibrium.}

806: It is well-known that {\bf all VC mechanisms yield the same

807: utility to a truth telling buyer}: For a VC mechanism

808: $d$\footnote{Since in all VC mechanisms $M=(X,d,c)$, $X$ is $V$

809: and $c$ is defined as in (3.2), it is enough to specify $d$ in

810: order to specify the mechanism.} we denote by

811: $u_i^d(v_i,(v'_i,v_{-i}))$ the utility of buyer $i$ with the

812: valuation function $v_i$, when he declares $v'_i$ and the other

813: buyers declare $v_{-i}$. That is,

814: $$u_i^d(v_i,v')=v_i(d_i(v'))-c^{d}_i(v'),\eqno{(3.3)}$$

815:  where $v'=(v'_i,v_{-i})$. Therefore, by (3.2),

816: $$u_i^d(v_i,v')=S(v,d(v'))-g_i(v_{-i}),\eqno{(3.4)}$$

817: where $v=(v_i,v_{-i})$, and

818: $$g_i(v_{-i})=\max_{\gamma\in \Gamma}\sum_{j\ne

819: i}v_j(\gamma_j).\eqno{(3.5)}$$

820:  If $i$ declares $v_i$,

821: $$ u_i^d(v_i,v)=S(v,d(v))-g_i(v_{-i})=S_{max}(v)-g_i(v_{-i}).\eqno{(3.6)}$$

822: As the right-hand side of (3.6) does not depend on $d$, a truth

823: telling buyer receives the same utility at all VC mechanisms.

824: Note that truth revealing is indeed a dominant strategy in a VC

825: mechanism.

826: 

827: 

828: 

829: 

830: In the next section we discuss other (not truth revealing)

831: individually rational ex post equilibrium profiles in the VC

832: mechanisms. We will focus on {\bf player symmetric} equilibria

833: $b=(b_i)_{i \in N}$, where $b_i=b_j$ for all $i,j\in N$, which are

834: in equilibrium in {\bf every} VC mechanism.

835: 

836: 

837: \section{Bundling equilibrium}

838: Let $\Sigma\subseteq 2^A$ be a family of bundles of goods. We

839: deal only with such families $\Sigma$ for which

840: \begin{itemize}

841: \item $\emptyset\in\Sigma$.

842: \end{itemize}

843: A valuation function $v_i$ is a {\bf $\Sigma$-valuation function}

844: if $$v_i(B)=\max_{C\in\Sigma,C\subseteq B}v_i(C),\quad\mbox{for

845: every $B\in 2^A$.}$$

846: 

847: The set of all $\Sigma$-valuation functions in $V_i$ is denoted by

848: $V_i^\Sigma$. We further denote $V^\Sigma=\times_{i\in

849: N}V_i^\Sigma$. For every valuation function $v_i$ we denote by

850: $v_i^\Sigma$ its projection on $V_i^\Sigma$, that is:

851: $$v_i^\Sigma(B)=\max_{C\in\Sigma,C\subseteq

852: B}v_i(C),\quad\mbox{for every $B\in 2^A$.}$$ Obviously

853: $v_i^\Sigma\in V_i^\Sigma$, and for $v_i\in V_i^\Sigma$,

854: $v_i^\Sigma=v_i$. In particular $(v_i^\Sigma)^\Sigma=v_i^\Sigma$

855: for every $v_i\in V_i$. Let $f_\Sigma:V_i\to V_i^\Sigma$ be the

856: projection function defined by $$f^\Sigma(v_i)=v_i^\Sigma.$$

857: 

858:  An

859: allocation $\gamma$ is a {\bf $\Sigma$-allocation} if

860: $\gamma_i\in\Sigma$ for every buyer $i\in N$. The set of all

861: $\Sigma$-allocations is denoted by $\Gamma^\Sigma$.

862: 

863: 

864: 

865: We are interested in the following question: For which $\Sigma$,

866: do we have that $f^\Sigma$ is a player-symmetric individually

867: rational ex post equilibrium in every   VC mechanism (with any

868: number of buyers)? In such a case we call $f^\Sigma$ a {\bf

869: bundling equilibrium }for the VC mechanisms and say that $\Sigma$

870: induces a bundling equilibrium. The next example shows that not

871: every $\Sigma$ induces a bundling equilibrium.

872: 

873: Before we present the example we need the following notation: Let

874: $B\in 2^A$, we denote by $w_B$ the following valuation function:

875: $$\quad\mbox{If $B\neq \emptyset$, } w_B(C)=1\quad\mbox{if

876: $B\subseteq C$, and}\quad w_B(C)=0

877: \quad\mbox{otherwise}.\footnote{For $B\ne\emptyset$, a valuation

878: function of the form $w_B$ is called a unanimity TU game in

879: cooperative game theory. An agent with such a valuation function

880: is called by Lehmann, O'Callaghan, and Shoham \cite{Lehmann99} a

881: single-minded agent.}$$ $$\quad\mbox{If $B=\emptyset$, } w_B(C)=0

882: \quad\mbox{for all } C \in 2^A.$$

883: % If $\alpha\in \Re$, we

884: %denote by $\alpha w_B$ the valuation function that assigns the

885: %value $\alpha w_B(D)$ for every $D\subseteq A$.

886: 

887: \noindent{\bf Example 1}

888: 

889: Let $A$ contain four goods $a,b,c,d.$ Let $$\Sigma=\{a,d, bcd,

890: abc, A,\emptyset\}.\footnote{We omit braces and commas when

891: writing subsets of $A$.}$$ Let $v_2=w_{a}, v_3=w_{d}$. Consider

892: buyer 1 with $v_1=w_{bc}$. Note that $v_i\in V_i^\Sigma$ for

893: $i=2,3$. If buyer 1 uses $f^\Sigma$ he declares $

894: v_1'(bcd)=v_1'(abc)=v_1'(A)=1$ and $v_1'(C)=0$ for all other $C$,

895:  and there

896: exists a VC mechanism  that allocates $a$ to 2, $d$ to 3 and $bc$

897: to the seller. In this mechanism the utility of 1 from using

898: $f^\Sigma$ is zero. On the other hand, if agent 1 reports the

899: truth ($w_{bc}$) he receives  (in every VC mechanism) $bc$ and

900: pays nothing. Hence, his utility would be 1. Therefore $f^\Sigma$

901: is not in equilibrium in this VC mechanism, and hence $\Sigma$

902: does not induce a bundling equilibrium.

903: 

904: \subsection{A characterization of bundling equilibria}

905: $\Sigma\subseteq 2^A$ is called a {\bf quasi field} if it

906: satisfies the following properties:\footnote{Recall our assumption

907: that we deal only with $\Sigma$ such that $\emptyset\in\Sigma$.}

908: \begin{itemize}

909: \item $B\in \Sigma$ implies that $B^c\in \Sigma$, where

910: $B^c=A\setminus B$.

911: %\item If $B_1,...,B_k\in\Sigma$ and $B_j\cap B_s=\emptyset$ for

912: %all $j\ne s$, then $\cup_{j=1}^k B_j\in\Sigma$.

913: \item $B,C \in \Sigma$ and $B \cap C= \emptyset$ imply that $B

914: \cup C \in \Sigma$.\footnote{Equivalently, the union of any number

915: of pairwise disjoint sets in $\Sigma$ is also in $\Sigma$.}

916: \end{itemize}

917: 

918: 

919: 

920: \begin{theorem} $\Sigma$ induces a bundling equilibrium if and only if

921: it is a quasi field.

922: \end{theorem}

923: 

924: \begin{proof}

925: 

926: 

927: 

928: {\bf Suppose $\Sigma$ is a quasi field}:

929: 

930:  Consider a VC mechanism with

931: an allocation function $d$. We show that $f^\Sigma$ is an

932: individually rational ex post equilibrium in this VC mechanism.

933: 

934: 

935: Assume that every buyer $j$, $j\ne i$, uses the strategy

936: $b_j=f^\Sigma$. Let $v_{-i}\in V_{-i}$. We have to show that for

937: buyer $i$ with valuation $v_i$,  $v_i^\Sigma$ is a best reply to

938: $v_{-i}^\Sigma$. As truth revealing is a dominating strategy in

939: every VC mechanism, it suffices to show that buyer $i$'s utility

940: when submitting $v_i^\Sigma$ is the same as when submitting

941: $v_i$.\footnote{Note that this will imply not only that $f^\Sigma$

942: is in equilibrium but also that it is individually rational.} That

943: is, we need to show that

944: 

945: $$S_{max}(v_i,v^\Sigma_{-i})-\alpha=S((v_i,v^\Sigma_{-i}),\gamma)-\alpha,$$

946: where $\alpha=g_i( v_{-i}^\Sigma)$, and

947: $\gamma=d(v_i^\Sigma,v^\Sigma_{-i})$.

948: 

949: Hence, we have to show that

950: $$S_{max}(v_i,v^\Sigma_{-i})=S((v_i,v^\Sigma_{-i}),\gamma).\eqno{(4.1)}$$

951: Obviously,

952: $$S_{max}(v_i,v^\Sigma_{-i})\ge S((v_i,v^\Sigma_{-i}),\gamma).\eqno{(4.2)}$$

953: 

954:  As $v_i(B)\ge v_i^\Sigma(B)$ for every $B\in 2^A$,

955: $$S((v_i,v^\Sigma_{-i}),\gamma)\ge

956: S((v_i^\Sigma,v^\Sigma_{-i}),\gamma)=S_{max}(v_i^\Sigma,v^\Sigma_{-i}).\eqno{(4.3)}$$

957: 

958: Let $ \xi=d(v_i,v^\Sigma_{-i})$. For $j\ne i$ and $j\ne 0$, let

959: $\xi_j^\Sigma\in\Sigma$ be such that $\xi^\Sigma_j\subseteq \xi_j$

960: and $v^\Sigma_j(\xi^\Sigma_j)=v^\Sigma_j(\xi_j).$ Let

961: $\xi^\Sigma_i=(\cup_{j\ne 0,i} \xi^\Sigma_j)^c$, and let

962: $\xi_0^\Sigma=\emptyset$.

963: 

964:  Because $\Sigma$ is a quasi field,

965: $\xi^\Sigma_i\in\Sigma$, and hence $\xi^\Sigma\in\Gamma^\Sigma$.

966: As $\xi_i\subseteq \xi^\Sigma_i$, $\xi^\Sigma$ is also optimal for

967: $(v_i,v^\Sigma_{-i})$. However

968: $$S((v_i,v^\Sigma_{-i}),\xi^\Sigma)=S((v_i^\Sigma,v^\Sigma_{-i}),\xi^\Sigma)\le

969: S_{max}(v_i^\Sigma,v^\Sigma_{-i}).\eqno{(4.4)}$$ Combining (4.2),

970: (4.3), and (4.4) yields

971: $$S_{max}(v_i^\Sigma,v^\Sigma_{-i}) \ge S_{max}(v_i,v^\Sigma_{-i}) \ge

972: S((v_i,v^\Sigma_{-i}),\gamma) \ge

973: S_{max}(v_i^\Sigma,v^\Sigma_{-i}).$$ Therefore (4.1) holds.

974: 

975: {\bf Suppose $\Sigma$ induces a bundling equilibrium}:

976: 

977: We first show that if $B\in\Sigma$, then $B^c\in\Sigma$.  If $B=A$

978: then by definition $B^c=\emptyset \in \Sigma$. Let $B\subset A$.

979: Assume, for the sake of contradiction, that $B^c\not\in\Sigma$.

980: Let $v_2=w_B$ and $v_1=w_{B^c}$. Note that $v_2^\Sigma=v_2$. Thus,

981: if buyer 2 uses $f^\Sigma$, he declares $v_2$. If buyer 1 uses

982: $f^\Sigma$, he declares $v_1^\Sigma$, where $v_1^\Sigma(B^c)=0$.

983: Hence, there exists a VC mechanism $d$, that allocates $B$ to

984: agent 2 and $B^c$ to the seller. However, if buyer 1 deviates and

985: declares his true valuation, then this VC mechanism allocates to

986: him $B^c$, and he pays nothing. Hence, there is a profitable

987: deviation from $f^\Sigma$, a contradiction.

988: 

989: Next, we show that if $B,C \in \Sigma$ are disjoint then $B \cup C

990: \in \Sigma$. By the first part of the proof, it suffices to show

991: that $(B \cup C)^c \in \Sigma$. Clearly, we may assume that the

992: sets $B$, $C$, and $(B \cup C)^c$ are all non empty. Assume, for

993: the sake of contradiction, that $(B \cup C)^c \notin \Sigma$.

994: Consider three buyers with valuations $v_1=w_{(B \cup C)^c}$,

995: $v_2=w_B$, $v_3=w_C$. Proceeding as in the

996:  first part of the current part of the  proof yields a similar

997:  contradiction.\QED

998:  \end{proof}

999: 

1000:  It may be useful to note that if $f^\Sigma$ is a buyer-symmetric

1001:  equilibrium for a fixed set of buyers, then $\Sigma$ is not

1002:  necessarily a quasi field. For example, if there is only one

1003:  buyer, every $\Sigma$ such that $A\in\Sigma$ induces an

1004:  equilibrium. In the case of two buyers, being closed under

1005:  complements is necessary and sufficient for $\Sigma$ to induce

1006:  an equilibrium. However, it can be deduced from the proof of the

1007:  only if part of Theorem 1, that for a fixed set of buyers $N$,

1008:  if $n=|N| \ge 3$, then $\Sigma$ must be a quasi field if it induces

1009:  an equilibrium for the set of buyers $N$.

1010: 

1011: \subsection{Partition-based equilibrium}

1012: Let $\pi=\{A_1,...,A_k\}$ be a partition of $A$ into non empty

1013: parts. That is, $A_i\ne\emptyset$ for every $A_i\in\pi$,

1014: $\cup_{i=1}^k A_i=A$, and $A_i\cap A_j=\emptyset$ for every $i\ne

1015: j$. Let $\Sigma_\pi$ be the field generated by $\pi$. That is,

1016: $\Sigma_\pi$ contains all the sets of goods of the form

1017: $\cup_{i\in I} A_i$, where $I\subseteq\{1,...,k\}$. To avoid

1018: confusion: $\emptyset\in \Sigma_\pi$. For convenience, we will

1019:  use $f^\pi$ to denote

1020: $f^{\Sigma_\pi}$.  A

1021: corollary of Theorem 1 is:

1022: \begin{corollary}

1023: $f^\pi$ is a bundling equilibrium.

1024: \end{corollary}

1025: \begin{proof}

1026: As $\Sigma_\pi$ is a field it is in particular a quasi field.

1027: Hence, the proof follows from Theorem 1.\QED

1028: \end{proof}

1029: 

1030: A bundling equilibrium of the form $f^\pi$, where $\pi$ is a

1031: partition, will be called a {\bf partition-based equilibrium}.

1032: Thus, a partition-based equilibrium is a bundling equilibrium

1033: $f^\Sigma$ that is based on a field $\Sigma=\Sigma_\pi$.

1034:  It is important to note that there exist quasi fields, which are

1035:  not fields.

1036:  For example, let $A=\{a,b,c,d\}$. $\Sigma=\{ab, cd, ac, bd, A ,

1037: \emptyset\}$ is a quasi field, which is not a field. We note,

1038: however, that when $m=|A| \le 3$, the notions of quasi field and

1039: field coincide.

1040: 

1041: \section{Surplus and communication complexity}

1042:  Let $\Sigma\subseteq 2^A$. If every buyer uses

1043: $f^\Sigma$, then in every VC mechanism $d$, the total surplus

1044: generated when the types of the buyers are given by $v\in V$ is

1045: $S(v^\Sigma,d(v^\Sigma))=S_{max}(v^\Sigma)$.

1046: 

1047:  We denote

1048: $$S_{\Sigma-max}(v)=\max_{\gamma\in\Gamma^\Sigma}S(v,\gamma).$$

1049: Obviously,

1050: 

1051: % \noindent{\bf Lemma 1} For every $\Sigma$,

1052: $$S_{\Sigma-max}(v)=S_{\Sigma-max}(v^\Sigma)=S_{max}(v^\Sigma),\quad\mbox{for

1053: every $v\in V$.}$$

1054: 

1055: %The following useful Lemma follows easily from the method of

1056: %proof of Theorem 1.

1057: 

1058: %\noindent{\bf Lemma 2 HOPEFULLY IT IS NOT JUST EASY but also

1059: %CORRECT}

1060: 

1061: %$\Sigma$ is a quasi field if and only if

1062: %$$S_{\Sigma-max}(v)=\max_{\gamma\in\Gamma^\Sigma, \gamma\quad

1063: %is\quad buyer-complete}S(v,\gamma)\quad\mbox{for every $v\in

1064: %V$.}$$

1065: For convenience we denote $S_{\Sigma-max}$ by $S_\Sigma$, and we

1066: call $S_\Sigma$ the {\bf $\Sigma$-optimal surplus function} (note

1067: that $S_{2^A}=S_{max}$). When $\Sigma$ is a field generated by a

1068: partition $\pi$ we write $S_\pi$ for $S_{\Sigma_\pi}$.

1069: 

1070: If $\Sigma$ is a quasi field we say that the {\bf communication

1071: complexity} of the equilibrium $f^\Sigma$ is the number of bundles

1072: in $\Sigma$, that is $|\Sigma|$. Notice that this is a natural definition

1073: because a buyer who is using $f^\Sigma$ has to submit  a vector of

1074: $|\Sigma|$ numbers to the seller.%

1075: \footnote{A discussion of the way this can be extended to deal with

1076: the introduction of concise bidding languages \cite{Nisan00,BoutilierHoos}

1077:  is beyond the scope of this

1078: paper.} Thus, if $\pi$ is a partition, the communication

1079: complexity is $2^{|\pi|}$. If $\Sigma_1\subseteq\Sigma_2$, then

1080: $S_{\Sigma_1}(v)\le S_{\Sigma_2}(v)$ for every $v\in V$. So,

1081: $\Sigma_2$ induces more surplus (a proxy for economic efficiency)

1082: than $\Sigma_1$, but $\Sigma_2$ also induces higher  communication

1083: complexity.

1084: %\footnote{as

1085: % well as more winner determination problem complexity.}.

1086: Hence, there is a tradeoff between

1087:   economic efficiency and computational

1088:  complexity.

1089:  The next example shows that as far as the {\bf revenue} of the

1090:  seller is concerned,

1091:   there is no clear

1092:  comparison between the revenues obtained by quasi

1093:  fields ranked by inclusion.\footnote{ In spite of our example, it is commonly

1094:  believed that social optimality is a good proxy for revenue.

1095:  This was proved to be asymptotically correct when the number of

1096:  buyers is large, and the organizer has a Bayesian belief over the

1097:  distribution of valuation functions, which assumes independence

1098:  across buyers (see \cite{MonTenAsymp}).}

1099:  Before we present the example, note that for two partitions

1100:  $\pi_1,\pi_2$, $\Sigma_{\pi_1}\subseteq\Sigma_{\pi_2}$ if and

1101:  only if $\pi_2$ refines $\pi_1$.

1102: 

1103: 

1104: 

1105: 

1106: 

1107: 

1108: 

1109: 

1110: 

1111:  \noindent{\bf Example 2}

1112: 

1113: 

1114:  Assume there are two buyers, $N=\{1,2\}$, and two goods, $A=\{a,b\}$.

1115: Assume $v_1=w_a$ and $v_2=w_b$. In any VC mechanism, in the truth

1116: revealing equilibrium buyer 1 gets $a$, buyer 2 gets $b$, and they

1117: pay nothing. Hence the level of social surplus is $2$ and the

1118: revenue of the seller at $v=(v_1,v_2)$ is zero.

1119:  Let $\pi$ be the trivial partition $\{A\}$

1120:  ($\Sigma_\pi=\{\emptyset,A\}$).

1121: If  each  buyer uses  the equilibrium strategy $f^{\pi}$, they

1122: both report $w_A$. Hence, one  of the buyers gets $ab$ and pays 1.

1123: The seller collects a revenue of $1$, and the social surplus

1124: equals 1. Hence, $S_{max}(v)>S_\pi(v)$ and $R(v)<R_\pi(v)$. On the

1125: other hand, if $N=\{1,2,3,4\}$ where $v_1$ and $v_2$ are defined

1126: as before and $v_3=v_1$ and $v_4=v_2$, $S_{max}(v)=2$ and $R(v)=2$

1127: while $S_\pi(v)=1$ and $R_\pi(v)=1$.

1128: 

1129: For every family of bundles $\Sigma$ with $A\in\Sigma$ we define

1130: $$r_\Sigma^n=\sup_{v\in V,v\ne 0}\frac{

1131: S_{max}(v)}{S_{\Sigma}(v)},\eqno{(5.1)}$$ where $V=V_1 \times

1132: \cdots \times V_n$. Thus, $r_\Sigma^n$ is a worst-case measure of

1133: the economic inefficiency that may result from using the strategy

1134: $f^\Sigma$ when there are $n$ buyers. Obviously $r_\Sigma^n \ge

1135: 1$, and equality holds for $\Sigma = 2^A$. A standard argument

1136: using homogeneity and continuity of $S_{max}/S_\Sigma$ shows that

1137: the supremum in (5.1) is attained, i.e., it is a maximum.

1138: 

1139: 

1140: 

1141: The following remark gives a simple upper bound on the

1142: inefficiency associated with $\Sigma$.

1143: 

1144: 

1145: 

1146: \noindent {\bf Remark 1}

1147:  For every $\Sigma \subseteq 2^A$ with $A \in \Sigma$, and for every $v\in V$,

1148:  $$S_{max}(v)\le n S_\Sigma(v),$$

1149:  where $n$ is the number of potential buyers.

1150: Consequently,

1151:  $$r^n_\Sigma\le n.$$

1152: 

1153:  \begin{proof}

1154:  Let $\gamma=d(v)$, where $d$ is any VC mechanism.

1155:  $$S_{max}(v)=S(v,\gamma)=\sum_{i\in N}v_i(\gamma_i)\le \sum_{i\in N}v_i(A)=

1156:  \sum_{i\in N}v^\Sigma_i(A)

1157:  \le n S_\Sigma(v).$$\QED

1158:  \end{proof}

1159: 

1160: However, we are interested mainly in upper bounds on the economic

1161: inefficiency that are independent of the number of buyers. For

1162: every family of bundles $\Sigma$ with $A \in \Sigma$ we define $$

1163: r_\Sigma = \sup_{n \ge 1} r_\Sigma^n.\eqno{(5.2)}$$ It is easy to

1164: see that, since any allocation assigns non empty bundles to at

1165: most $m=|A|$ buyers, the supremum in (5.2) is attained for some $n

1166: \le m$. When $\Sigma = \Sigma_\pi$ for a partition $\pi$, we write

1167: $r_\pi$ instead of $r_{\Sigma_\pi}$.

1168: 

1169: In the following subsection we characterize and estimate $r_\pi$,

1170: thereby obtaining a quantitative form of the tradeoff between

1171: communication and economic efficiency in partition-based

1172: equilibria.

1173: 

1174: 

1175:  \subsection{ Communication efficiency vs.

1176:  economic efficiency in partition-based equilibria}

1177: 

1178: 

1179: 

1180:  We first express $r_\pi$ in

1181: terms of the partition $\pi = \{A_1,...,A_k\}$ only. A {\bf

1182: feasible} family for $\pi$ is a family $\Delta =(H_i)_{i=1}^s$ of

1183: (not necessarily distinct) subsets of $\{1,...,k\}$ satisfying the

1184: following two conditions:

1185: 

1186: \begin{itemize}

1187: \item

1188:  $H_i\cap

1189: H_j\ne\emptyset$ for every $1\le i,j\le s$.

1190: \item $|\{i:l \in H_i \}| \le |A_l|$ for every $1 \le l \le k$.

1191: \end{itemize}

1192: We write $s=s(\Delta)$ for the number of sets in the family

1193: $\Delta$ (counted with repetitions).

1194: 

1195: \begin{theorem}

1196: For every partition $\pi$, $$r_\pi= \max s(\Delta),$$ where the

1197: maximum is taken over all families $\Delta$ that are feasible for

1198: $\pi$.

1199: \end{theorem}

1200: \begin{proof}

1201: 

1202: We first prove that $r_\pi \le \max s(\Delta)$. It suffices to

1203: show that for every $v \in V$ there exists a feasible family

1204: $\Delta$ for $\pi$ such that $$S_{max}(v) \le s(\Delta) \cdot

1205: S_\pi(v).$$

1206: 

1207: Let $v\in V$. Let $\gamma$ be a socially optimal allocation. That

1208: is, $$S_{max}(v)=\sum_{i\in N}v_i(\gamma_i).$$ For every

1209: $\gamma_i$ let $\gamma_i^\pi$ be the minimal set in $\Sigma_\pi$

1210: that contains $\gamma_i$. That is $\gamma_i^\pi=\cup_{l\in

1211: J_i}A_l$ where $J_i=\{l \in \{1,\ldots,k\}:A_l \cap \gamma_i \neq

1212: \emptyset\}$.

1213: 

1214: Let $\xi$ be a partition of $N$ to $r$ subsets, such that for

1215: every $i,j\in I\in\xi$, $i\ne j$, $\gamma_i^\pi\cap

1216: \gamma_j^\pi=\emptyset$. Assume $r$ is the {\bf minimal}

1217: cardinality of such a partition. For every $I\in\xi$ let

1218: $H_I=\cup_{i \in I} J_i$. That is, each $H_I$ is a set of indices

1219: of parts $A_l$ in $\pi$ that should be allocated to the buyers in

1220: $I$ in order for each of them to get the goods they received in

1221: the optimal allocation $\gamma$. Note that if $I\ne J$, $H_I\cap

1222: H_J\ne\emptyset$, otherwise we can join $I$ and $J$ together in

1223: contradiction to the minimality of  the cardinality of $\xi$.

1224: Hence, $\Delta =(H_I)_{I\in\xi}$ is a family of subsets of

1225: $\{1,...,k\}$ that satisfies that any two subsets in $\Delta$

1226: intersect. Furthermore, the second condition for feasibility is

1227: also satisfied, because for any given $l \in \{1,...,k\}$ there

1228: are at most $|A_l|$ buyers $i$ with $\gamma_i \cap A_l \ne

1229: \emptyset$, and hence at most $|A_l|$ parts $I \in \xi$ such that

1230: $l \in H_I$. Thus, $\Delta$ is a feasible family for $\pi$ with

1231: $s(\Delta)=r$.

1232: 

1233:  Every $H_I$, $I\in\xi$ defines a

1234: $\Sigma_\pi$-allocation. In this allocation every $i\in I$

1235: receives $\gamma_i^\pi$, and the seller receives all other goods.

1236:  Therefore $\sum_{i\in I}v_i(\gamma_i^\pi)\le S_\pi(v)$ for every

1237:  $I\in\xi$. Hence,

1238:  $$S_{max}(v)\le \sum_{i\in

1239:  N}v_i(\gamma_i^\pi)=\sum_{I\in\xi}\sum_{i\in

1240:  I}v_i(\gamma_i^\pi)\le\sum_{I\in\xi}S_\pi(v)=rS_\pi(v).$$

1241: 

1242: Next, we prove that $r_\pi \ge \max s(\Delta)$. It suffices to

1243: show that for every feasible family $\Delta$ for $\pi$ there

1244: exists a profile of valuations $v=(v_1,...,v_n) \ne 0$ for some

1245: number $n$ of buyers satisfying $$S_{max}(v) \ge s(\Delta) \cdot

1246: S_\pi(v).$$

1247: 

1248: Let $\Delta =(H_i)_{i=1}^s$ be a feasible family for $\pi$. By the

1249: second condition of feasibility, we can associate with each $H_i$

1250: a set of goods $B_i$ containing one good from each $A_l$ such that

1251: $l \in H_i$, in such a way that the sets $B_i$ are pairwise

1252: disjoint. By the first condition of feasibility, for every $1 \le

1253: i,j \le s$ there can be no two disjoint sets $C_i, C_j \in

1254: \Sigma_\pi$ such that $B_i \subseteq C_i$, $B_j \subseteq C_j$.

1255: 

1256: Now, we take $n=s$ buyers, and let buyer $i$ have the valuation

1257: $v_i=w_{B_i}$. Then $S_{max}(v)=s$ whereas $S_\pi(v)=1$.\QED

1258: \end{proof}

1259: 

1260: Theorem~2 reduces the determination of the economic inefficiency

1261: measure $r_\pi$ to a purely combinatorial problem. However, this

1262: combinatorial problem does not admit an easy

1263: solution.\footnote{The special case of this problem, in which

1264: $|A_i|=|A_j|$ for all $A_i, A_j \in \pi$, has been treated in the

1265: combinatorial literature using a different but equivalent

1266: terminology (see e.g. \cite{Furedi}). But even in this case, a

1267: precise formula for $\max s(\Delta)$ seems out of reach.}

1268: Nevertheless, we will use Theorem~2 to calculate $r_\pi$ in some

1269: special cases, and to obtain a general upper bound for it which is

1270: tight in infinitely many cases.

1271: 

1272: The following proposition determines $r_\pi$ for partitions $\pi$

1273: with a small number of parts. We use the notations $\lfloor \cdot

1274: \rfloor$ and $\lceil \cdot \rceil$ for the lower and upper integer

1275: rounding functions, respectively.

1276: 

1277: \begin{proposition}

1278: Let $|A|=m$, and let $\pi = \{A_1,...,A_k\}$ be a partition of $A$

1279: into $k$ non empty sets.

1280: \begin{itemize}

1281: \item If $k=1$ then $r_\pi=m$.

1282: \item If $k=2$ then $r_\pi= \max \{|A_1|,|A_2|\}$. Consequently,

1283: the minimum of $r_\pi$ over all partitions of $A$ into 2 parts is

1284: $\lceil \frac{m}{2} \rceil$.

1285: \item If $k=3$ then $r_\pi= \max \{|A_1|,|A_2|,|A_3|, \lfloor

1286: \frac{m}{2} \rfloor \}$. Consequently, the minimum of $r_\pi$ over

1287: all partitions of $A$ into 3 parts is $\lfloor \frac{m}{2}

1288: \rfloor$.

1289: \end{itemize}

1290: \end{proposition}

1291: 

1292: \begin{proof}

1293: 

1294: In each case, we determine the maximum of $s(\Delta)$ over all

1295: families $\Delta$ that are feasible for $\pi$.

1296: 

1297: For $k=1$, a feasible family consists of at most $|A_1|=m$ copies

1298: of $\{1\}$, and therefore $\max s(\Delta)=m$.

1299: 

1300: A feasible family for $k=2$ cannot contain two sets, $H_i$ and

1301: $H_j$, such that $1 \notin H_i$ and $2 \notin H_j$, because such

1302: sets would be disjoint. Hence, for any feasible family $\Delta$,

1303: either all sets contain 1 or all of them contain 2. Therefore,

1304: $s(\Delta) \le \max \{|A_1|,|A_2|\}$. On the other hand, feasible

1305: families of size $|A_1|$, $|A_2|$ trivially exist.

1306: 

1307: Suppose $k=3$, and denote $$\beta_l = |A_l| \quad\mbox{for }

1308: l=1,2,3.$$ We first show that $s(\Delta) \le \max \{ \beta_1,

1309: \beta_2, \beta_3, \lfloor \frac{m}{2} \rfloor \}$ for every

1310: feasible family $\Delta$. If $\Delta$ contains some singleton

1311: $\{l\}$, then all sets in $\Delta$ must contain $l$, and hence

1312: $s(\Delta) \le \beta_l$. Otherwise, $\Delta$ consists of $s_{12}$

1313: copies of $\{1,2\}$, $s_{13}$ copies of $\{1,3\}$, $s_{23}$ copies

1314: of $\{2,3\}$, and $s_{123}$ copies of $\{1,2,3\}$, for some non

1315: negative integers $s_{12}, s_{13}, s_{23}, s_{123}$. We have the

1316: following inequalities: $$s_{12}+s_{13}+s_{123} \le \beta_1,$$

1317: $$s_{12}+s_{23}+s_{123} \le \beta_2,$$ $$s_{13}+s_{23}+s_{123} \le

1318: \beta_3.$$ Upon adding these inequalities we obtain

1319: $$2(s_{12}+s_{13}+s_{23})+3s_{123} \le m,$$ which implies

1320: $$s(\Delta)=s_{12}+s_{13}+s_{23}+s_{123} \le \lfloor \frac{m}{2}

1321: \rfloor.$$ We show next that there exists a feasible family

1322: $\Delta$ with $s(\Delta)= \max \{\beta_1, \beta_2, \beta_3,

1323: \lfloor \frac{m}{2} \rfloor \}$. If this maximum is one of the

1324: $\beta_l$'s, this is trivial. So assume that $\beta_l < \lfloor

1325: \frac{m}{2} \rfloor$ for $l=1,2,3$. If $m$ is even then the family

1326: $\Delta$ that consists of $$s_{12}=\frac{\beta_1 +\beta_2

1327: -\beta_3}{2} \quad\mbox{copies of } \{1,2\},$$

1328: $$s_{13}=\frac{\beta_1 + \beta_3 - \beta_2}{2} \quad\mbox{copies

1329: of } \{1,3\},$$ $$s_{23}= \frac{\beta_2 + \beta_3 - \beta_1}{2}

1330: \quad\mbox{copies of } \{2,3\},$$ is feasible (note that the

1331: prescribed numbers are non negative because $\beta_l < \lfloor

1332: \frac{m}{2} \rfloor$ for $l=1,2,3$, and they are integers because

1333: $\beta_1 + \beta_2 + \beta_3 =m$ is even). The size of this family

1334: is $s(\Delta)=s_{12}+s_{13}+s_{23}= \frac{m}{2}$. If $m$ is odd,

1335: we make slight changes in the values of $s_{12},s_{13},s_{23}$: we

1336: add $\frac{1}{2}$ to one of them and subtract $\frac{1}{2}$ from

1337: the other two. In this way we get a family $\Delta$ with

1338: $s(\Delta)= \lfloor \frac{m}{2} \rfloor$.\QED

1339: \end{proof}

1340: 

1341: We see from Proposition~1 that if we use partitions into two parts

1342: (entailing a communication complexity of 4), the best we can do in

1343: terms of economic efficiency is $r_\pi = \lceil \frac{m}{2}

1344: \rceil$, and this is achieved by partitioning $A$ into equal or

1345: nearly equal parts. Allowing for three parts (and therefore a

1346: communication complexity of 8) permits only a small gain in

1347: $r_\pi$ (in fact, no gain at all when $m$ is even).

1348: 

1349: 

1350: 

1351: We will now state the two parts of our main result.

1352:  \begin{theorem}

1353: Let $\pi = \{A_1,...,A_k\}$ be a partition of $A$ into $k$ non

1354: empty sets of maximum size $\beta(\pi)$. (That is, $\beta(\pi) =

1355: \max \{|A_1|,...,|A_k|\}$.) Then $$r_\pi \le \beta(\pi) \cdot

1356: \varphi(k),$$ where $$\varphi(k)=\max_{j=1,...,k} \min \{

1357: j,\frac{k}{j} \}.$$

1358: 

1359:  \end{theorem}

1360: 

1361:  The proof of Theorem~3 is given in the following subsection.

1362: 

1363: Note that $$\varphi(k)\le\sqrt{k}.$$ In particular, if all sets in

1364: $\pi$ have equal size $\frac{m}{k}$, we obtain the upper bound

1365: $$r_\pi \le \frac{m}{\sqrt{k}}.$$

1366: 

1367: Now, consider the case when, for some non negative integer $q$, we

1368: have $$k=q^2+q+1, \quad\mbox{and} \eqno{(5.3)}$$ $$|A_i|=q+1

1369: \quad\mbox{for } i=1,...,k. \eqno{(5.4)}$$ In this case

1370: $$\varphi(k)= \frac{q^2+q+1}{q+1},$$ and hence the upper bound of

1371: Theorem~3 takes the form $$r_\pi \le k.$$ The second part of our

1372: main result implies that in infinitely many of these cases this

1373: upper bound is tight.

1374: 

1375: 

1376: 

1377: 

1378:  \begin{theorem}

1379: Let $\pi = \{A_1,...,A_k\}$ be a partition that satisfies (5.3)

1380: and (5.4) for some $q$ which is either 0 or 1 or of the form $p^l$

1381: where $p$ is a prime number and $l$ is a positive integer. Then

1382: $$r_\pi = k.$$

1383: 

1384: 

1385:  \end{theorem}

1386: 

1387: We prove Theorems 3 and 4 in the following subsection.

1388: 

1389: \subsection{Proofs of Theorems~3 and 4}

1390: We begin with some preparations. Let $\Delta=(H_i)_{i=1}^s$ be a

1391: family of (not necessarily distinct) subsets of $\{1,...,k\}$. A

1392: vector of non negative  numbers $\delta=(\delta_i)_{i=1}^s$ is

1393: called a {\bf

1394:  semi balanced}\footnote{This concept is equivalent to what is

1395:  called a fractional matching in combinatorics. We chose the term

1396:  semi balanced, because balanced vectors, defined by requiring

1397:  equality instead of weak inequality, are a familiar concept in

1398:  game theory (see e.g. \cite{Shapley67}).}

1399:  vector for $\Delta$ if for every $l \in \{1,...,k\}$,

1400: $$\sum_{i: l\in H_i}\delta_i\le 1.$$%

1401: \begin{proposition}

1402: Let $\Delta=(H_i)_{i=1}^s$ be a family of (not necessarily

1403: distinct) subsets of $\{1,...,k\}$ such that $H_i\cap

1404: H_j\ne\emptyset$ for every $1\le i,j\le s$. Let

1405: $\delta=(\delta_i)_{i=1}^s$ be a semi balanced vector for

1406: $\Delta$. Then $$\sum_{i=1}^s\delta_i\le\varphi(k),$$ where

1407: 

1408: $$\varphi(k)=\max_{j=1,...,k}\min \{j,\frac{k}{j}\}.$$

1409: \end{proposition}

1410: \begin{proof}

1411: 

1412: Assume without loss of generality that $h=|H_1|$ is the minimal

1413: number of elements in a member of $\Delta$. The proposition will

1414: be proved if we prove the following two claims:

1415: 

1416:  \noindent{\bf Claim 1:}

1417: $\sum_{i=1}^s\delta_i\le h.$

1418: 

1419: \noindent{\bf Claim 2:} $\sum_{i=1}^s\delta_i\le\frac{k}{h}.$

1420: 

1421: \noindent{\bf Proof of Claim 1:}

1422: 

1423: Let $z=\sum_{l\in H_1}\sum_{i: l\in H_i}\delta_i$. As every $H_i$

1424: intersects $H_1$, every $\delta_i$ appears in $z$ at least once.

1425: Therefore, $z \ge \sum_{i=1}^s\delta_i$. Because $\delta$ is

1426:  semi balanced, $\sum_{i: l\in H_i} \delta_i \le 1$ for every $l$, and in particular

1427: for $l \in H_1$. Hence, $z\le \sum_{l\in H_1} 1=h.$ \QED

1428: 

1429: \noindent{\bf Proof of Claim 2:}

1430: 

1431: Let $w= \sum_{l=1}^k \sum_{i:l \in H_i} \delta_i$. Every

1432: $\delta_i$ appears in $w$ exactly $|H_i|$ times. Since $|H_i| \ge

1433: h$ for every $i$, we have $w \ge h \sum_{i=1}^s \delta_i$. On the

1434: other hand, as in the proof of Claim~1, we obtain $w \le

1435: \sum_{l=1}^k 1=k$. Combining the two inequalities, we get

1436: $\sum_{i=1}^s \delta_i \le \frac{k}{h}$. \QED

1437: 

1438:  Therefore,

1439:  $$\sum_{i=1}^s\delta_i\le

1440: \min \{h,\frac{k}{h}\}\le\varphi(k).$$

1441: 

1442:  \QED

1443: 

1444: \end{proof}

1445: 

1446: We are now ready for the proof of Theorem 3:

1447: 

1448: \noindent{\bf Proof of Theorem 3:}

1449: 

1450: Let $\pi = \{A_1,...,A_k\}$ be a partition of $A$ into $k$ non

1451: empty sets of maximum size $\beta(\pi)$. We have to prove that

1452: $r_\pi \le \beta(\pi) \cdot \varphi(k)$. By Theorem~2, it suffices

1453: to show that for every feasible family $\Delta$ for $\pi$, we have

1454: $$s(\Delta) \le \beta(\pi) \cdot \varphi(k).$$ Let

1455: $\Delta=(H_i)_{i=1}^s$ be such a family. Consider the vector

1456: $\delta = (\delta_i)_{i=1}^s$ with $$\delta_i =

1457: \frac{1}{\beta(\pi)}, \quad i=1,...,s.$$ By the second condition

1458: of feasibility, this vector is semi balanced. Hence we may apply

1459: Proposition~2 and conclude that $$\sum_{i=1}^s \delta_i \le

1460: \varphi(k),$$ or equivalently, $$\frac{s}{\beta(\pi)} \le

1461: \varphi(k),$$ as required. \QED

1462: 

1463: In order to prove Theorem~4 we invoke a result about finite

1464: geometries (see e.g. \cite{Demb}). A {\bf finite projective plane}

1465: of order $q$ is a system consisting of a set $\Pi$ of points and a

1466: set $\Lambda$ of lines (in this abstract setting, a line is just a

1467: set of points, i.e., $L \subseteq \Pi$ for every $L \in \Lambda$),

1468: satisfying the following conditions:

1469: \begin{itemize}

1470: \item $|\Pi|=|\Lambda|=q^2+q+1$.

1471: \item Every point is incident to $q+1$ lines and every line

1472: contains $q+1$ points.

1473: \item There is exactly one line containing any two points, and

1474: there is exactly one point common to any two lines.

1475: \end{itemize}

1476: Such a system does not exist for every $q$. However, it trivially

1477: exists for $q=0$ (a single point) and for $q=1$ (a triangle) and

1478: it is known to exist for every $q$ of the form $q=p^l$, where $p$

1479: is a prime number and $l$ is a positive integer. The first non

1480: trivial example, corresponding to $q=2$, is called the {\bf Fano

1481: plane}: $$\Pi= \{1,2,3,4,5,6,7\},$$ $$\Lambda=

1482: \{124,235,346,457,561,672,713\}.$$

1483: 

1484: \noindent {\bf Proof of Theorem~4:}

1485: 

1486: Let $\pi = \{A_1,...,A_k\}$ be a partition that satisfies (5.3)

1487: and (5.4) for some $q$ which is either 0 or 1 or of the form $p^l$

1488: where $p$ is a prime number and $l$ is a positive integer. As

1489: $r_\pi \le k$ follows from Theorem~3 (see the discussion preceding

1490: the statement of Theorem~4), we need to prove only that $r_\pi \ge

1491: k$. By Theorem~2, it suffices to show that there exists a family

1492: $\Delta$ with $s(\Delta)=k$ which is feasible for $\pi$. Such a

1493: family is given by the system of lines of a projective plane of

1494: order $q$, when the points are identified with $1,...,k$. \QED

1495: 

1496: \subsection{More on the ranking of equilibria}

1497: 

1498: The tradeoff between communication complexity and economic

1499: efficiency, as delineated above, may be made concrete by the

1500: following scenario. Suppose that a set $A$ of $m$ goods is given,

1501: and we are in a position to recommend to the potential buyers an

1502: equilibrium strategy. Assume further that a certain level $M$ of

1503: communication complexity is considered the maximum acceptable

1504: level. If we are going to recommend a partition-based equilibrium

1505: $f^\pi$, then the number of parts in $\pi$ should be at most $k=

1506: \lfloor \log_2 M \rfloor$. From the viewpoint of economic

1507: efficiency, we would like to choose such a partition $\pi$ with

1508: $r_\pi$ as low as possible. Which partition should it be?

1509: 

1510: According to Theorem~3, we obtain the lowest guarantee on $r_\pi$

1511: by making the maximum size of a part in $\pi$ as small as

1512: possible, which means splitting $A$ into $k$ equal (or nearly

1513: equal, depending on divisibility) parts. This leads to the

1514: question whether, for given $m$ and $k$, the lowest value of

1515: $r_\pi$ itself (not of our upper bound) over all partitions $\pi$

1516: of $A$ into $k$ parts is achieved at an equi-partition, i.e., a

1517: partition $\pi =\{A_1,...,A_k\}$ such that $\lfloor \frac{m}{k}

1518: \rfloor \le |A_i| \le \lceil \frac{m}{k} \rceil$, $i=1,...,k$.

1519: 

1520: While Proposition~1 gives an affirmative answer for $k=1,2,3$, it

1521: turns out, somewhat surprisingly, that this is not always the

1522: case. This is shown in the following example.

1523: 

1524: \noindent {\bf Example~3}

1525: 

1526: Let $m=21$ and $k=7$. If $\pi$ is an equi-partition of the 21

1527: goods into 7 triples then, by Theorem~4, $r_\pi =7$. Consider now

1528: a partition $\pi'=\{A_1,...,A_7\}$ in which $$|A_1|=2, |A_2|=4,

1529: |A_3|= \cdots = |A_7|=3.$$ We claim that $r_{\pi'} \le 6$.

1530: 

1531: In order to prove this, it suffices to show that there exists no

1532: feasible family of 7 sets for $\pi'$. Suppose, for the sake of

1533: contradiction, that $\Delta=(H_i)_{i=1}^7$ is such a family. Let

1534: $H_i$ be an arbitrary set in $\Delta$. It follows from the second

1535: condition of feasibility that if $H_i$ contains the element 1 then

1536: it shares it with at most one other set in $\Delta$. Similarly, if

1537: $H_i$ contains the element 2 then it shares it with at most three

1538: other sets in $\Delta$. For $l=3,...,7$, if $H_i$ contains the

1539: element $l$ then it shares it with at most two other sets in

1540: $\Delta$. This implies that $H_i$ must contain at least three

1541: elements (because it must share an element with every other set,

1542: and $3+2<6$). Moreover, if $H_i$ contains exactly three elements

1543: and one of them is 1, then it also contains 2 (since $1+2+2<6$).

1544: On the other hand, we have $$\sum_{i=1}^7 |H_i| = \sum_{i=1}^7

1545: \sum_{l \in H_i} 1= \sum_{l=1}^7 \sum_{i:l \in H_i} 1=

1546: \sum_{l=1}^7 |\{i:l \in H_i\}| \le \sum_{l=1}^7 |A_l|=21.$$ Since

1547: every $H_i$ has at least three elements, it follows that every

1548: $H_i$ has exactly three elements, and all the weak inequalities

1549: $|\{i:l \in H_i\}| \le |A_l|$ must in fact hold as equalities. In

1550: particular, there exist two sets in $\Delta$, say $H_i$ and $H_j$,

1551: that contain the element 1. By the above, they both contain 2 as

1552: well. Let $l$ be the third element of $H_i$. Then among the

1553: remaining five sets in $\Delta$, the set $H_i$ shares the element

1554: 1 with none of them, it shares the element 2 with two of them, and

1555: the element $l$ with at most two of them. This contradicts the

1556: fact that $H_i$ intersects every other set in $\Delta$.

1557: 

1558: It can be checked that in fact $r_{\pi'}=6$ and this is the lowest

1559: achievable value among all partitions of 21 goods into 7 sets. We

1560: omit the detailed verification of this.

1561: 

1562: The tradeoff between communication complexity and economic

1563: efficiency was quantitatively analyzed above only for

1564: partition-based equilibria. It is natural to ask whether it is

1565: possible to beat this tradeoff using the more general bundling

1566: equilibria. The answer is, in a sense made precise below:

1567: sometimes yes, but not by much.

1568: 

1569: \noindent {\bf Example~4}

1570: 

1571: Assume that the number of goods $m$ is even, and let the set of

1572: goods $A$ be partitioned into two equal parts $B$ and $C$.

1573: Consider $\Sigma \subseteq 2^A$ defined by $$\Sigma = \{ D

1574: \subseteq A: |D \cap B|=|D \cap C| \}.$$ It is easy to check that

1575: $\Sigma$ is a quasi field, and hence it induces a bundling

1576: equilibrium. The communication complexity is

1577: $$|\Sigma|=\sum_{j=0}^{m/2} {m/2 \choose j}^2=\sum_{j=0}^{m/2}

1578: {m/2 \choose j}{m/2 \choose m/2 -j}= {m \choose m/2}.$$

1579: 

1580: We claim that $r_\Sigma =2$. That $r_\Sigma \ge 2$ can be seen by

1581: taking two buyers with valuations $w_B$ and $w_C$, respectively.

1582: To see that $r_\Sigma \le 2$, suppose that $v$ is a profile of

1583: valuations for a set of buyers $N$, and let $\gamma$ be an optimal

1584: allocation. Split the set $N$ into the following two sets: $$N_B=

1585: \{i \in N:|\gamma_i \cap B| \ge |\gamma_i \cap C| \},$$ $$N_C= \{i

1586: \in N:|\gamma_i \cap B| < |\gamma_i \cap C| \}.$$ Note that the

1587: sets of goods $\gamma_i, i \in N_B$, can be expanded to pairwise

1588: disjoint sets of goods that belong to $\Sigma$. In other words,

1589: there exists a $\Sigma$-allocation $\xi$ such that $\gamma_i

1590: \subseteq \xi_i$ for every $i \in N_B$. Similarly, there exists a

1591: $\Sigma$-allocation $\eta$ such that $\gamma_i \subseteq \eta_i$

1592: for every $i \in N_C$. Hence $$S_{max}(v)= \sum_{i \in N}

1593: v_i(\gamma_i)= \sum_{i \in N_B} v_i(\gamma_i) + \sum_{i \in N_C}

1594: v_i(\gamma_i) \le \sum_{i \in N} v_i(\xi_i) + \sum_{i \in N}

1595: v_i(\eta_i) \le 2S_\Sigma(v).$$ Thus, $r_\Sigma \le 2$.

1596: 

1597: We claim further that if a partition $\pi$ of $A$ satisfies $r_\pi

1598: \le 2$ then $|\Sigma_\pi| \ge 2^{m-2}$. Indeed, suppose $\pi =

1599: \{A_1,...,A_k\}$. It is easy to find a feasible family of 3 sets

1600: for $\pi$ if one of the $A_l$'s has three or more elements, or if

1601: three of the $A_l$'s have two elements each. Therefore, $r_\pi \le

1602: 2$ implies that at most two of the sets $A_1,...,A_k$ have two

1603: elements and the rest are singletons. Thus $k \ge m-2$ and

1604: $|\Sigma_\pi| \ge 2^{m-2}$.

1605: 

1606: Since ${m \choose m/2} < 2^{m-2}$ for all even $m \ge 10$, we have

1607: the following conclusion: If $m \ge 10$ then every partition-based

1608: equilibrium that matches the economic efficiency of $f^\Sigma$ has

1609: a higher communication complexity than $f^\Sigma$. In other words,

1610: the quasi field $\Sigma$ offers an efficiency/complexity

1611: combination that cannot be achieved or improved upon (in the

1612: Pareto sense) by any field.

1613: 

1614: The above example notwithstanding, the efficiency/complexity

1615: combinations which arise from arbitrary quasi fields are still

1616: subject to a tradeoff that is not much better than for fields. This

1617: is the content of our final remark.

1618: 

1619: \noindent {\bf Remark 2}

1620:  Let $m=|A|$ and let $k$ be a positive integer. Any quasi field

1621: $\Sigma \subseteq 2^A$ with $r_\Sigma \le \frac{m}{k}$ must

1622: contain a partition of $A$ into $k$ non empty parts, and therefore

1623: must satisfy $|\Sigma| \ge 2^k$.

1624: 

1625:  \noindent {\bf Proof:}

1626: 

1627: Let there be $m$ buyers, each with valuation $w_a$ for a distinct

1628: $a \in A$. For this $v$ we have $S_{max}(v)=m$. If $r_\Sigma \le

1629: \frac{m}{k}$ then we must have $S_\Sigma(v) \ge k$. Hence an

1630: optimal $\Sigma$-allocation has to assign non empty bundles of

1631: goods to at least $k$ buyers. Thus $\Sigma$ contains $k$ pairwise

1632: disjoint non empty sets of goods, and therefore, being a quasi

1633: field, also a partition of $A$ into $k$ non empty parts. \QED

1634: 

1635: \begin{thebibliography}{10}

1636: 

1637: \bibitem{Anderson}

1638: A.~Anderson, M.~Tenhunen, and F.~Ygge.

1639: \newblock Integer programming for combinatorial auction winner determination.

1640: \newblock In {\em ICMAS}, pages 39--46, 2000.

1641: 

1642: \bibitem{Ausubel2000}

1643: L.~Ausubel.

1644: \newblock An efficient dynamic auction for heterogeneous commodities.

1645: \newblock Mimeo, University of Maryland, 2000.

1646: 

1647: \bibitem{AusubelMilgrom}

1648: L.~Ausubel and P.~Milgrom.

1649: \newblock Ascending auctions with package bidding.

1650: \newblock Mimeo, University of Maryland, 2001.

1651: 

1652: \bibitem{BoutilierHoos}

1653: C.~Boutilier and H.H. Hoos.

1654: \newblock {Bidding languages for combinatorial auctions}.

1655: \newblock In {\em The 17th international joint conference artificial

1656:   intelligence}, pages 1211--1216, 2001.

1657: 

1658: \bibitem{Clarke}

1659: E.~Clarke.

1660: \newblock Multipart pricing of public goods.

1661: \newblock {\em Public Choice}, 18:19--33, 1971.

1662: 

1663: \bibitem{Cramton}

1664: P.C. Cramton.

1665: \newblock Money out of thin air: The nationwide narrowband pcs auction.

1666: \newblock {\em Journal of Economics and Management Strategy}, 1995.

1667: 

1668: \bibitem{dasmaskin}

1669: P.~Dasgupta and E.~Maskin.

1670: \newblock Efficient auctions.

1671: \newblock Discussion Paper Harvard University, to appear in Quarterly Journal

1672:   of Economics, 1999.

1673: 

1674: \bibitem{deVriesVohra}

1675: S.~de~Vries and R.~Vohra.

1676: \newblock Combinatorial auctions: A brief survey.

1677: \newblock Unpublished manuscript, 2000.

1678: 

1679: \bibitem{Demb}

1680: P.~Dembowski.

1681: \newblock {\em Finite geometries}.

1682: \newblock Springer, 1968.

1683: 

1684: \bibitem{FujiBrownShoham}

1685: Y.~Fujishima, K.~Leyton-Brown, and Y.~Shoham.

1686: \newblock Taming the computational complexity of combinatorial auctions:

1687:   Optimal and approximate approaches.

1688: \newblock In {\em IJCAI-99}, 1999.

1689: 

1690: \bibitem{Furedi}

1691: Z.~F\"uredi.

1692: \newblock Covering pairs by $q^2+q+1$ sets.

1693: \newblock {\em Journal of Combinatorial Theory Ser. A}, 54(2):248--271, 1990.

1694: 

1695: \bibitem{Groves}

1696: T.~Groves.

1697: \newblock Incentives in teams.

1698: \newblock {\em Econometrica}, 41:617--631, 1973.

1699: 

1700: \bibitem{HonSela}

1701: S.~Hon-Snir, D.~Monderer, and A.~Sela.

1702: \newblock A learning approach to auctions.

1703: \newblock {\em Journal of Economic Theory}, 82:65--88, 1998.

1704: 

1705: \bibitem{HoosBoutilier}

1706: H.H. Hoos and C.~Boutilier.

1707: \newblock Solving combinatorial auctions using stochastic local search.

1708: \newblock In {\em The 17th national conference on artificial intelligence},

1709:   pages 22--29, 2000.

1710: 

1711: \bibitem{jehielmsignals}

1712: P.~Jehiel and B.~Moldovanu.

1713: \newblock Efficient design with interdependent valuations.

1714: \newblock Working Paper University of Mannheim, 2000.

1715: 

1716: \bibitem{jmolsAER}

1717: P.~Jehiel, B.~Moldovanu, and E.~Stacchetti.

1718: \newblock How (not) to sell nuclear weapons.

1719: \newblock {\em American Economic Review}, 86(4):814--829, 1996.

1720: 

1721: \bibitem{jmols}

1722: P.~Jehiel, B.~Moldovanu, and E.~Stacchetti.

1723: \newblock Multidimensional mechanism design for auctions with externalities.

1724: \newblock {\em Journal of Economic Theory}, 85(2):258--294, 1999.

1725: 

1726: \bibitem{KrishnaPerry}

1727: V.~Krishna and M.~Perry.

1728: \newblock Efficient mechanism design.

1729: \newblock Working Paper, 1998.

1730: 

1731: \bibitem{Lehmann99}

1732: D.~Lehmann, L.I. O'Callaghan, and Y.~Shoham.

1733: \newblock Truth revelation in rapid, approximately efficient combinatorial

1734:   auctions.

1735: \newblock In {\em ACM Conference on Electronic Commerce}, 1999.

1736: 

1737: \bibitem{mcafeereny}

1738: P.R. McAfee and P.~Reny.

1739: \newblock Correlated information and mechanism design.

1740: \newblock {\em Econometrica}, 60:395--421, 1992.

1741: 

1742: \bibitem{McMillan}

1743: J.~McMillan.

1744: \newblock Selling spectrum rights.

1745: \newblock {\em Journal of Economic Perspectives}, 8:145--162, 1994.

1746: 

1747: \bibitem{Milgromfcc}

1748: P.~Milgrom.

1749: \newblock Putting auction theory to work: The simultaneous ascending auction.

1750: \newblock Technical Report 98-0002, Department of Economics, Stanford

1751:   University, 1998.

1752: 

1753: \bibitem{MilgromWeber}

1754: P.~Milgrom and R.J. Weber.

1755: \newblock A theory of auctions and competitive bidding.

1756: \newblock {\em Econometrica}, 50:1089--1122, 1982.

1757: 

1758: \bibitem{MonTenAsymp}

1759: D.~Monderer and M.~Tennenholtz.

1760: \newblock {Asymptotically optimal multi-object auctions}.

1761: \newblock Working paper, Technion, 2000.

1762: 

1763: \bibitem{Myerson79}

1764: R.~Myerson.

1765: \newblock Incentive compatibility and the bargaining problem.

1766: \newblock {\em Econometrica}, 47:61--74, 1979.

1767: 

1768: \bibitem{Nisan00}

1769: N.~Nisan.

1770: \newblock Bidding and allocation in combinatorial auctions.

1771: \newblock In {\em ACM Conference on Electronic Commerce}, 2000.

1772: 

1773: \bibitem{nisancommunication}

1774: N.~Nisan.

1775: \newblock The communication complexity of combinatorial auctions.

1776: \newblock Working paper, the Hebrew University, 2001.

1777: 

1778: \bibitem{NisanRonenGames}

1779: N.~Nisan and A.~Ronen.

1780: \newblock Algorithmic mechanism design.

1781: \newblock {\em Games and Economic Behavior}, 35:166--196, 2001.

1782: 

1783: \bibitem{Parkes99}

1784: D.~C. Parkes.

1785: \newblock ibundle: An efficient ascending price bundle auction.

1786: \newblock In {\em ACM Conference on Electronic Commerce}, 1999.

1787: 

1788: \bibitem{ParkesUngar}

1789: D.~C. Parkes and L.~H. Ungar.

1790: \newblock Iterative combinatorial auctions: Theory and practice.

1791: \newblock In {\em AAAI,IAAI}, pages 74--81, 2000.

1792: 

1793: \bibitem{perryrenyb}

1794: M.~Perry and P.~Reny.

1795: \newblock An ex post efficient multi-unit ascending auction for agents with

1796:   interdependent valuations.

1797: \newblock Working paper, University of Chicago, 1999.

1798: 

1799: \bibitem{perryrenya}

1800: M.~Perry and P.~Reny.

1801: \newblock An ex post efficient multi-unit auction for agents with

1802:   interdependent valuations.

1803: \newblock Working paper, University of Pittsburgh, 1999.

1804: 

1805: \bibitem{RothPekHar}

1806: M.H. Rothkopf, A.~Pekec, and R.M. Harstad.

1807: \newblock Computationally manageable combinatorial auctions.

1808: \newblock {\em Management Science}, 44(8):1131--1147, 1998.

1809: 

1810: \bibitem{Sandholm99}

1811: T.~Sandholm.

1812: \newblock An algorithm for optimal winner determination in combinatorial

1813:   auctions.

1814: \newblock In {\em IJCAI-99}, 1999.

1815: 

1816: \bibitem{Sandholm01}

1817: T.~Sandholm, S.~Suri, A.~Gilpin, and D.~Levine.

1818: \newblock Cabob: A fast optimal algorithm for combinatorial auctions.

1819: \newblock In {\em 17th International Joint Conference on Artificial

1820:   Intelligence}, pages 1102--1108, 2001.

1821: 

1822: \bibitem{Shapley67}

1823: L.~Shapley.

1824: \newblock {On Balanced Sets and Cores}.

1825: \newblock {\em Naval Research Logistics Quarterly}, 14:453--560, 1967.

1826: 

1827: \bibitem{Tijcai}

1828: M.~Tennenholtz.

1829: \newblock Electronic commerce: From game-theoretic and economic models to

1830:   working protocols.

1831: \newblock In {\em IJCAI-99}, 1999.

1832: 

1833: \bibitem{Tenn}

1834: M.~Tennenholtz.

1835: \newblock Some tractable combinatorial auctions.

1836: \newblock Proceedings of AAAI-2000, 2000.

1837: 

1838: \bibitem{Varian95}

1839: H.R. Varian.

1840: \newblock Economic mechanism design for computerized agents.

1841: \newblock In {\em Proceedings of Usenix Conference on Electronic Commerce, New

1842:   York, July 11-12}, 1995.

1843: 

1844: \bibitem{Vickrey}

1845: W.~Vickrey.

1846: \newblock Counterspeculations, auctions, and competitive sealed tenders.

1847: \newblock {\em Journal of Finance}, 16:15--27, 1961.

1848: 

1849: \bibitem{Weber}

1850: R.J. Weber.

1851: \newblock Multiple-object auctions.

1852: \newblock In R.~Engelbreach, M.~Shubik, and J.~Stark, editors, {\em Auctions,

1853:   Bidding, and Contracting}, pages 165--191. 1983.

1854: 

1855: \bibitem{Wellmangeb}

1856: M.P. Wellman, P.R. Wurman, W.E. Walsh, and J.K. MacKie-Mason.

1857: \newblock Auction protocols for decentralized scheduling.

1858: \newblock {\em Games and Economic Behavior}, 35:271--303, 2001.

1859: 

1860: \end{thebibliography}

1861: 

1862: 

1863: 

1864: 

1865: %\bibliography{bibnew11}

1866: %\bibliographystyle{plain}

1867: 

1868: 

1869: 

1870: 

1871: 

1872: 

1873: 

1874: 

1875: 

1876: 

1877: 

1878: 

1879: 

1880: 

1881: 

1882: 

1883: 

1884: \end{document}

1885: 

1886: 

1887: 

1888: 

1889: 

1890: 

1891: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1892: \noindent{Bundling VCG mechanisms}

1893: 

1894: {\bf Every} $\Sigma$ defines a family of truth telling mechanism

1895: -- the $\Sigma$-VCG mechanisms, which are just  the VCG mechanisms

1896: applied to the set of allocation $\Gamma^\Sigma$. In these

1897: mechanisms, which are parameterized by allocation function

1898: $d^\Sigma: V\to \Gamma^\Sigma$, every buyer $i$ reports a

1899: valuation function, say $v_i$. The mechanism picks an allocation

1900: $d^\Sigma(v)$, that maximizes $S(v,\gamma)$ over

1901: $\gamma\in\Gamma^\Sigma$, and buyer $i$ pays according to (3.2),

1902: where $d$ is replaced with $d^\Sigma$ and $\Gamma$ is replaced

1903: with $\Gamma^\Sigma$.

1904: 

1905: \begin{theorem}

1906: \end{theorem}

1907: 

1908: 

1909: 

1910: 

1911: 

1912: Note that the same outcome is generated if the organizer implement

1913: the $\pi$-VCG mechanism, and the buyers use their (unique)

1914: dominating strategy in this mechanism -- $b^{\pi}$, or when she

1915: uses the full $VCG$ mechanism, and the buyer use the player

1916: symmetric equilibrium $b^{\pi}$.

1917:   When using the $\pi$ mechanism (when $\pi$ contains small number of sets), the

1918: organizer knows that the players use their (unique) dominating

1919: strategy -- $b^{\pi}$. While by using the full mechanism, the

1920: buyers may learn to play in equilibrium  the appropriate

1921: partition-based equilibrium.  In the later case, the organizer

1922: cannot control the particular $\pi$ chosen in equilibrium (or any

1923: other equilibrium that may be chosen ( see the coming

1924: conjecture)). We will further discuss this issue when we deal with

1925: the organizer's incentives.

1926: 

1927: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%5

1928: 

1929: section{Bayesian Games} A game in informational form becomes a

1930: Bayesian game if we assign to it a probabilistic information, and

1931: make additional "probabilistic assumptions" assumptions. A

1932: probabilistic information for the game $G$ requires the

1933: specification of a sigma field $\Sigma$ of subsets of $\Omega$,

1934: sigma fields $F_i$ of subsets of $T_i$ for all $i\in N$, a sigma

1935: fields $Q_i$ of subsets of $X_i$ for all $i\in N$, and a

1936: probability measure $\lambda$ on $(\Omega,\Sigma)$. We denote the

1937: Bayesian game by $G(\lambda)$, assuming that the measure structure

1938: is understood from the context. If $G$ is a finite game we will

1939: always assume that all sets under discussion are measurable and

1940: the probabilistic information will be given by a vector

1941: $\lambda=(\lambda(\omega))_{\omega\in\Omega}$. It is always

1942: implicitly assumed that every $u_i$ and every $\tilde t_i$ is a

1943: Borel measurable function. A Bayesian equilibrium of $i$ in

1944: $G(\lambda)$ is a measurable strategy of $i$ in $G$. A profile of

1945: Bayesian strategies $b=(b_i)_{i\in N}$ is a Bayesian-Nash

1946: equilibrium in $G(\lambda)$, if the following definition makes

1947: sense in probability theory: For every version of the conditional

1948: expectation,

1949: 

1950: For every agent $i$,   and for every $x_i\in X_i$,

1951: $$E_\lambda\left(u_i(.,b_i(t_i),\hat b_{-i}(.))|\tilde t_i=t_i\right)\ge

1952: E_\lambda\left(u_i(.,x_i,\hat b_{-i}(.))|\tilde t_i=t_i\right),$$

1953: for $\lambda_i-$almost every $t_i\in T_i$, where $\lambda_i$ is

1954: the probability distribution on $T_i$ induced by $\lambda$ and by

1955: $\tilde t_i$. The amount of technical and uninteresting work

1956: needed  to make the following result correct does not justify

1957: this. Therefore we state it as a meta theorem:

1958: 

1959: 

1960: $\bullet${\bf Principle}

1961: 

1962:  $b$ is a $PI$ equilibrium in $G$ if and only if $b$

1963: is a Bayesian-Nash equilibrium in $G(\lambda)$ for every $\lambda$

1964: for which $b$ is a  profile of Bayesian strategies(i.e.,

1965: measurable), and for which all integrals make sense.

1966: