1: \documentclass[11pt,onecolumn]{IEEEtran}

2: \usepackage{subfigure}

3: \usepackage{epsfig,graphicx,psfrag}

4: \usepackage{amsfonts,amsmath,amssymb}

5: %amsthm}

6: %\usepackage{amsfonts,amsmath,amssymb,amsthm}

7: \newtheorem{lemma}{Lemma}

8: \newtheorem{theorem}{Theorem}

9: %\theoremstyle{definition}

10: \newtheorem{cor}{Corollary}

11: \newtheorem{assumption}{Assumption}

12: \newtheorem{conjecture}{Conjecture}

13: \newtheorem{definition}{Definition}

14: \newtheorem{example}{Example}

15: \newcommand{\tends}{{\rightarrow}}

16: \newcommand{\hGe}{\hat{G}^\epsilon}

17: \newcommand{\hEe}{\hat{E}^\epsilon}

18: \newcommand{\lf}{\left}

19: \newcommand{\rf}{\right}

20: \newtheorem{corollary}{Corollary}

21: \newtheorem{condition}{Condition}

22: \newcommand{\bone}{{\mathbf 1}}

23: 

24: 

25: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

26: %delimiters

27: 

28: \let\bl\bigl

29: \let\bbl\Bigl

30: \let\bbbl\biggl

31: \let\bbbbl\Biggl

32: \let\br\bigr

33: \let\bbr\Bigr

34: \let\bbbr\biggr

35: \let\bbbbr\Biggr

36: \let\eps\epsilon

37: \let\eps\epsilon

38: \newcommand{\I}{\mathcal{I}}

39: \newcommand{\E}{\mathbb{E}}

40: \newcommand{\R}{\mathbb{R}}

41: \newcommand{\C}{\mathbb{C}}

42: \newcommand{\cC}{\mathcal{C}}

43: \newcommand{\cA}{\mathcal{A}}

44: 

45: \newcommand{\V}{\mathbb{V}}

46: \renewcommand{\l}{\lambda}

47: \newcommand{\N}{\mathbb{N}}

48: \newcommand{\Z}{\mathbb{Z}}

49: \newcommand{\F}{\mathbb{F}}

50: \renewcommand{\P}{\mathbb{P}}

51: \newcommand{\cP}{\mathcal{P}}

52: \newcommand{\HH}{\mathbb{H}}

53: \renewcommand{\S}{\mathbb{S}}

54: \newcommand{\cS}{\mathcal{S}}

55: 

56: %\pagestyle{empty}

57: \newcommand{\ie}{\emph{i.e.}}

58: \newcommand{\eg}{\emph{e.g.}}

59: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

60: \begin{document}

61: \title{\huge Product Multicommodity Flow in Wireless Networks}

62: \author{Ritesh Madan, Devavrat Shah, and Olivier Leveque

63: \thanks{R. Madan is at QUALCOMM-Flarion Technologies, D. Shah is with Departments of EECS and ESD at MIT, and Olivier

64: Leveque is at EPFL. Emails:~{ \tt rkmadan@stanfordalumni.org, devavrat@mit.edu,

65: olivier.leveque@epfl.ch.}

66:  }}

67: %\date{}

68: \maketitle

69: 

70: %\markboth{Submitted to the IEEE Trans. on Information

71: %Theory}{Submitted to the IEEE Trans. on Information Theory }

72: 

73: \begin{abstract}

74: 

75: We provide a tight approximate characterization of the

76: $n$-dimensional product multicommodity flow (PMF)

77: region for a wireless network of $n$ nodes. Separate

78: characterizations in terms of the spectral properties of

79: appropriate network graphs are obtained in both an information theoretic

80: sense and for a combinatorial interference model (\eg, Protocol

81: model). These provide an inner approximation to the $n^2$ dimensional

82: capacity region. These results answer the following

83: questions which arise naturally from previous work: (a) What is the

84: significance of $1/\sqrt{n}$ in the scaling laws for the Protocol

85: interference model obtained by Gupta and Kumar (2000)? (b) Can we

86: obtain a tight approximation to the ``maximum supportable flow" for

87: node distributions more general than the geometric random

88: distribution, traffic models other than randomly chosen

89: source-destination pairs, and under very general assumptions on the

90: channel fading model?

91: 

92: We first establish that the random source-destination model is

93: essentially a one-dimensional approximation to the capacity region,

94: and a special case of product multi-commodity flow. For

95: a wireline network (graph), a series of results starting from the result

96: of Leighton and Rao (1988) relate the product multicommodity flow to

97: the spectral (or cut) property of the graph. Building on these results,

98: for a combinatorial interference model given by a network

99:     and a conflict

100: graph, we relate the product multicommodity flow to the spectral

101: properties of the underlying graphs resulting in computational upper

102: and lower bounds. These results show that the $1/\sqrt{n}$ scaling

103: law obtained by Gupta and Kumar for a geometric random network can

104: be explained in terms of the combinatorial properties of a

105: geometric random network and the scaling law of the \emph{conductance}

106: of a grid graph. For the more interesting random fading

107: model with additive white Gaussian noise (AWGN), we show that the

108: scaling laws for PMF can again be tightly characterized by the

109: spectral properties of appropriately defined graphs. As an

110: implication, we obtain computationally efficient upper and lower

111: bounds on the PMF for any wireless network with a guaranteed

112: approximation factor.

113: 

114: 

115: \end{abstract}

116: 

117: \begin{keywords}

118: Product multicommodity flow, wireless network, scaling law, capacity

119: region.

120: \end{keywords}

121: % Note that keywords are not normally used for peer review papers.

122: 

123: % For peer review papers, you can put extra information on the cover

124: % page as needed:

125: % \begin{center} \bfseries EDICS Category: 3-BBND \end{center}

126: %

127: % For peerreview papers, inserts a page break and creates the second title.

128: % Will be ignored for other modes.

129: %\IEEEpeerreviewmaketitle

130: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

131: \section{Introduction}

132: 

133: \subsection{Prior Work}\label{s:intro} \vspace{.1in}

134: 

135: An important open question in network information theory is that of

136: characterizing the capacity region of a wireless network of $n$

137: nodes, i.e., the set of all achievable rates between the $n^2$ pairs of

138: nodes in terms of the joint statistics of the channels between

139: these nodes. This has proved to be a very challenging question; even

140: the capacity of a relay network comprising of three nodes is

141: not known in complete generality.

142: 

143: Instead of trying to characterize the

144: capacity region for a general wireless network, the seminal paper by

145: Gupta and Kumar~\cite{GK} concentrated on obtaining the maximum achievable rate

146: for a particular communication model, geometric random distribution

147: of nodes, and randomly chosen source-destination pairs. They showed

148: that the maximum rate for the protocol interference model scales as

149: $\Theta(1/\sqrt{n})$ for $n$ nodes randomly placed on a sphere of

150: unit area. This precise characterization has been followed by many

151: interesting results for both combinatorial interference models and

152: the random fading information theoretic model for large random

153: networks; these include~\cite{GT, KV, EMPS_journal,F, GV, XXK} for

154: communication theoretic models, and~\cite{LT,XK} for

155: information theoretic results. These results are crucially based on

156: the assumption that a large number of nodes are randomly distributed

157: in a certain region, and the inherent symmetry in the random

158: source-destination pair traffic model.

159: 

160: 

161: Since the relative locations of wireless nodes play an

162: important role in the characterization of the capacity region, the

163: notion of transport capacity was defined in~\cite{GK}. A scaling law

164: for the transport capacity for the protocol interference model was

165: obtained in~\cite{GK}, while that for an information theoretic

166: setting was obtained in~\cite{XXK}. The transport capacity can be used

167: to obtain an upper bound on the achievable rate-region for

168: certain rate-tuples, but is not of much use in determining the feasibility of

169: a certain rate-tuple.

170: More

171: recently, information theoretic  outer bounds to the capacity

172: region of a wireless network with a finite number of nodes were

173: obtained in~\cite{AJV} for any wireless network, using the cut-set

174: bound~\cite[Ch. 14]{Cover_book}. We note that any achievable scheme

175: can be used to obtain a set of lower bounds. While the above is only

176: a discussion of a representative set of results in this area

177: (see~\cite{XK_mono} for a more detailed summary), we note that there

178: is no result which provides upper and lower bounds with a guaranteed

179: approximation factor for a general wireless network with a generic

180: random fading model. In this paper, we take the first steps towards

181: providing such a tight characterization under very general

182: assumptions. In doing so, we make  connections between

183: spectral graph theoretic results and network information theory.

184: This results in efficient methods to compute the above tight upper

185: and lower bounds.

186: 

187: %\vspace{.05in}

188: 

189: \subsection{Contribution and Organization}

190: \vspace{.1in}

191: 

192: In Section~\ref{sec1}, we consider the product multicommodity flow (PMF)

193: as an $n$ dimensional approximation of the $n^2$ dimensional

194: capacity region. We show that the random souce-destination pair

195: traffic model is a special case of PMF and it is essentially a

196: one-dimensional approximation of the capacity region.

197: 

198: In Section~\ref{sec2}, we study the PMF for an arbitrary topology and a general

199: combinatorial interference model, of which the protocol model is a

200: special case. We show that the normalized cut capacity

201: (equivalently conductance) of a capacitated network graph induced by

202: the node placement and the interference model characterizes the PMF

203: (within $\log n$ factor). For this model, we also obtain a precise

204: scaling law for average delay using very elementary and almost

205: structure independent arguments.

206: We provide, possibly simpler, re-derivation of the (weaker by $\log^{2.5}n$ factor)

207: lower bound on the maximum flow obtained by Gupta and Kumar for randomly chosen

208: permutation flow on a geometric random graph with a protocol interference model.

209: Our derivation illustrates the connections between the combinatorial properties of geometric random

210: graphs and the maximum PMF.

211: While we do not discuss in detail in this

212: paper, the spectral properties of appropriate induced capacitated

213: graphs characterize the scaling laws obtained for mobile networks

214: in~\cite{GT}, \cite{DGT}.

215: 

216: 

217: In Section~\ref{sec3}, we address the question of characterizing the

218: PMF for a wireless network with Gaussian channels and random fading.

219: This is substantially more challenging than for the combinatorial

220: interference model because there is no obvious underlying network

221: graph that specifies the links which should be used for data

222: transmission. We construct a capacitated graph whose cut capacity

223: characterizes (in terms of tight upper and lower bounds) the PMF in

224: the wireless network. This construction allows one to use classical

225: network flow arguments to characterize and compute the PMF. We

226: illustrate the generality of our results by obtaining scaling laws for a

227: geometric random network and for a network where the number of

228: commodities is constant.

229: 

230: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

231: \section{Traffic Flows}\label{sec1}

232: 

233: In this section, we describe a class of traffic flows, namely,

234: the class of product multicommodity flows, that we study in this paper, and its

235: relevance. Consider a wireless network of $n$ nodes and denote the

236: node set as $V=\{1,\hdots,n\}$. A traffic matrix $\lambda =

237: [\lambda_{ij}] \in \R_+^{n\times n}$ is said to be {\em feasible},

238: if for each pair of nodes $(i,j), ~1\leq i, j \leq n$, data can be

239: transmitted from node $i$ to node $j$ at rate $\lambda_{ij}$. Note

240: that whether a traffic matrix $\lambda$ is feasible or not depends

241: on the model for the underlying wireless network, and we shall

242: describe the precise models for wireless networks in the later

243: sections.

244: 

245: We denote the capacity region by  $\Lambda$, i.e., $\Lambda$ is the

246: set of all feasible traffic matrices. Ideally, we would like to

247: characterize $\Lambda$. However, this is a hard problem in most cases. Instead, we

248:  characterize an {\em approximation} of $\Lambda$ under

249: general assumptions on the wireless network. For this, we consider

250: product multicommodity flow (PMF), defined as follows.

251: 

252: \vspace{.1in}

253: 

254: \begin{definition}[Product Multicommodity Flow (PMF)]

255: Let node $i$ be assigned a weight $\pi(i)$, for $1\leq i\leq n$.

256: Then the PMF corresponding to the weights $\pi\in \R_{+}^n$ and a flow rate $f\in \R_{+}$ is given by the function~\cite{LR99}

257: $M: \R^{n+1} \mapsto \R^{n \times n} $:

258: \[ M(f,\pi) = f \left[ \begin{array}{llll}

259: 0 & \pi(1)\pi(2) &\hdots & \pi(1)\pi(n)\\

260: \pi(2)\pi(1) & 0 & \hdots & \pi(2)\pi(n)\\

261: \vdots & \vdots &\vdots &\vdots\\

262: \pi(n) \pi(1) & \pi(n)\pi(2)&\hdots& \ \ 0

263: \end{array}\right].

264: \]

265: \end{definition}

266: \vspace{.1in} The PMF is an $n$-dimensional approximation of the

267: $n^2$ dimensional capacity region $\Lambda$ with {\em product}

268: constraints. An important special case arises when

269: all the weights are 1, i.e., $\pi(i) = 1$ for $i=1,\hdots,n$. We call

270: such a flow \emph{uniform multicommodity flow (UMF)}. \vspace{.1in}

271: 

272: \begin{definition}[Uniform Multicommodity Flow (UMF)]

273: UMF with flow rate $f\in \R_{+}$ is given by $U(f) = f {\bf 1}$, where

274: ${\bf 1}\in \mathbb{R}_+^{n\times n}$ is a matrix with all

275: entries equal to 1.

276: \end{definition}

277: \vspace{0.1in}

278: We denote by $f_\pi^*$ as the supremum over the flow rates for which the PMF

279: corresponding to the weights $\pi$ is feasible, i.e.,

280: \[ f^*_\pi = \sup \{ f\in R_{+}~:\ M(f, \pi) \text{ is feasible}\}.\]

281: We abuse notation and denote the corresponding quantity for UMF as

282: simply $f^*$.

283: 

284: 

285: 

286: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

287: 

288: 

289: 

290: 

291: \subsection{Inner Approximation to $\Lambda$}

292: 

293: 

294: 

295: 

296: We first show that the maximum UMF $f^*$ is a one-parameter

297: approximation to the capacity region $\Lambda$. Consider the

298: following parameter defined in terms of the capacity region as

299: follows.

300: 

301: \vspace{.1in}

302: 

303: \begin{definition}[$\rho^*$]

304: For any $\lambda \in \R^{n\times n}_+$, let $ \rho(\lambda)

305: \stackrel{\triangle}{=} \max_{i} \left\{  \sum_{k=1}^n \lambda_{ik},

306: \sum_{k=1}^n \lambda_{ki} \right\}.$ Let \\\mbox{$L(x) =

307: \{\lambda\in\R_+^{n\times n}: \rho(\lambda)\leq x\}$}. Then,  define

308: $\rho^*$ as

309: $$\rho^* = \sup\{x\in \R_+: L(x)\subseteq \Lambda\}.$$

310: \end{definition}

311: \vspace{.1in} Thus the quantity $\rho^*$ is a parametrization of a

312: (regular) polyhedral inner approximation to the capacity region

313: $\Lambda$. It is tight in the sense for any $x>\rho^*$, there is an

314: infeasible traffic matrix in the set $L(x)$.

315: %Since the set

316: %$\{x\in \R_+: L(x)\subseteq \Lambda\}$ is closed and bounded, any $\lambda$

317: %with $\rho(\lambda)\leq \rho^*$ is feasible.

318: 

319: Roughly speaking, the following result shows that UMF $f^*$ and

320: $\rho^*$ are equally good approximations to the capacity region

321: $\Lambda$.

322: 

323: \vspace{.1in}

324: \begin{lemma}

325: \label{lem0} If $U(f)$ is feasible, then any $\lambda\in

326: \R_{+}^{n\times n}$ such that $\rho(\lambda)\leq nf/2$ is feasible.

327: \end{lemma}

328: \begin{proof}

329: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

330: Consider any $\lambda$ such that $\rho(\lambda) \leq nf/2$. Suppose

331: that $U(f)$ is feasible. Then there exists a transmission scheme such

332: supports $U(f)$.

333: We now consider the two stage routing scheme of Valiant and Brenber \cite{valiant}

334: which routes $U(\rho(\lambda)/n)$ in each stage. Since $U(f)$ is feasible,

335: any $\lambda$ with $\rho(\lambda) \leq nf/2$ is supportable by time sharing

336: between the two transmission schemes corresponding to the two stages.

337: To complete the

338: proof of the Lemma, we need to describe this two stage routing

339: scheme.

340: 

341: In the first stage, each node $i$ sends data to all the remaining

342: nodes uniformly (ignoring its actual destination). Thus, node $i$

343: sends data to any node $j$ at rate $\sum_{k} \lambda_{ik}/n \leq

344: \rho(\lambda)/n$.  In the second stage, a node, say $j$, on

345: receiving data (from the first stage) from any source $i$ sends it

346: to the appropriate destination. It is easy to see that due to the

347: uniform spreading of data in the first stage, each node say $j$

348: routes data at rate $\sum_k \lambda_{ki}/n \leq \rho(\lambda)/n$ to

349: node $i$ in the second stage. Thus, the traffic matrices routed in

350: both the stages are dominated by $U(\rho(\lambda)/n)$. That is, the

351: sum traffic matrix is dominated by $U(2\rho(\lambda)/n)$. Hence, if

352: $U(f)$ is feasible then $\rho(\lambda)\leq nf/2$ is feasible. This

353: completes the proof of Lemma \ref{lem1}.

354: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

355: \end{proof}

356: 

357: 

358: \vspace{.1in}

359: \begin{theorem}

360: \label{lem1}

361: $f^*$ and $\rho^*$ are related as

362: \[ \frac{nf^*}{2} \leq \rho^*  \leq nf^* \]

363: \end{theorem}

364: \begin{proof} Note that in general, the capacity region $\Lambda$ may not be

365: closed, and so we need a more careful argument\footnote{We present such a formal

366: argument only once; similar arguments are implicit in many results that follow}.

367: We first show that $\frac{nf^*}{2} \leq \rho^*$. By definition of

368: $\sup$ it follows that for any $\epsilon>0$, $U(f^* - n\epsilon/2)$

369: is feasible. Hence, from Lemma~\ref{lem0}, any $\lambda\in

370: \R_{+}^{n\times n}$ such that $\rho(\lambda)\leq nf^*/2 - \epsilon$

371: is feasible. Hence, again using the definition of $\sup$,

372: $\rho^*\geq nf^*/2$.

373: 

374: Now for the other bound, assume that $\rho^*  > nf^*$, and $\epsilon

375: = (\rho^* - nf^*)/2$. Then, by definition of $\sup$ and $\rho$,

376: $U(nf^* + \epsilon/2)$ is feasible, which is a contradiction. Hence,

377: it follows that $\rho^*  \leq nf^*$.

378: 

379: 

380: 

381: 

382: \end{proof}

383: 

384: 

385: 

386: \vspace{.1in}

387: Thus, bounds on $f^*$ give bounds on $\rho^*$ which differ by at most a factor of 2. Subsequently,

388: a  scaling law for $f^*$ as a function of $n$ is the same as a scaling law for $\rho^*$,

389: i.e. $f^* = \Theta(\rho^*)$ as a function of $n$.

390: 

391: 

392: The set of all feasible PMF clearly provides an $n$ dimensional

393: inner approximation to the capacity region, which is, in general,

394: $n^2$ dimensional. Thus the characterization of the set of feasible

395: PMFs provides a much better approximation to the capacity region

396: than that the one-dimensional approximation given by set of feasible

397: UMF. We next establish the equivalence of UMF with a traffic model

398: with a randomly chosen permutation flow.

399: 

400: \vspace{.1in}

401: 

402: \subsubsection{UMF and Random Permutation Flow}

403: 

404: In some previous work, (e.g.,~\cite{GK}), the capacity scaling laws

405: were derived for the case where $n$ distinct source-destination

406: pairs are chosen at random such that each node is a source

407: (destination) for exactly one destination (source) and such a

408: pairing is done uniformly at random over all possible such pairings.

409: Thus the traffic matrix corresponds to a randomly chosen permutation

410: flow which is defined as follows.

411: 

412: \vspace{.1in}

413: 

414: \begin{definition}[Permutation Flow]

415: Let $S_n$ denote the set of permutation matrices in $\mathbb{R}_+^{n\times n}$.

416: Then the permutation flow corresponding to a permutation $\Sigma\in S_n$ and flow rate

417: $f \in \R_+$ is given by  $S(f,\Sigma)= f \Sigma$.

418: \end{definition}

419: \vspace{.1in}

420: Many previous works  study the scaling of $\bar{f}$, where $\bar{f}$

421: is the supremum over the set of $f \in \R_+$  such when a permuation

422: $\Sigma$ is randomly chosen from $S_n$, the permutation flow $S(f,\Sigma)$

423: is feasible with probability at least $1-1/n^2$. We now show that

424: when a permutation flow with flow rate $nf$ and a randomly chosen permutation

425: is feasible with a high enough probability, then the uniform mulicommodity

426: flow $U(f)$ can be ``almost" supported when $n$ is large enough.

427: \vspace{.1in}

428: \begin{lemma}\label{lem2}

429: For $\Sigma \in S_n$ chosen uniformly at random, if $(nf)\Sigma$

430: is feasible with probability at least $1-n^{-1-\alpha}, \alpha > 0$, then

431: there exists a sequence of feasible rate matrices $\Gamma_n$ such that

432: \[  \| {U_n(f) - \Gamma_n \|} = O(f n^{-\alpha}) \to 0 \mbox{~as~$n\to\infty$,}\]

433: where $\| \cdot \|$ denotes the standard 2-norm for matrices\footnote{Given

434: a matrix $M \in \R^{n \times n}$, the 2-norm of $M$ is $\| M \| = \sup\{ \| Mx \| : x \in \R^n , \|x\| = 1 \}$,

435: where $\|x\|$ is the $\ell_2$ norm of vector $x \in \R^n$.}, and

436: $U_n(f)$ is the uniform multicommodity flow for $n$ nodes.

437: \end{lemma}

438: \begin{proof}

439: From the hypothesis of the Lemma, it is clear that for at least

440: $(1-n^{-1-\alpha})$ fraction of all $n!$ permutations in $S_n$, the

441: permutation flow $(nf)\Sigma$ is feasible. By definition and

442: symmetry of permutations, we can write

443: \[ U_n(f) = \frac{1}{n!}\sum_{i=1}^{n!} (nf)\Sigma_i.\]

444: Let us define the following indicator function

445: \[{\bf 1}_i = \left\{\begin{array}{ll}

446: 1&\quad\text{$(nf)\Sigma_i$ is supportable}\\

447: 0&\quad\text{otherwise}

448: \end{array}\right. \]

449: Consider a uniform time sharing scheme between all the $n!$

450: permutation flows. Then the following traffic matrix is supportable.

451: \[ \Gamma_n = \frac{1}{n!}\sum_{i=1}^{n!} {\bf 1}_i (nf)\Sigma_i\]

452: Thus

453: \begin{eqnarray*}

454: \|{U_n(f) - \Gamma_n\|} &=  &\left|\left|\frac{1}{n!} \sum_{i=1}^{n!}(1- {\bf 1}_i)(nf)\Sigma_i\right |\right|\\

455:                            &\stackrel{(a)}{\leq} & \frac{1}{n!}\sum_{i=1}^{n!} \left|\left| (1- {\bf 1}_i)(nf)\Sigma_i\right |\right|\\

456:                            &\stackrel{(b)}{=} & \frac{1}{n!}\sum_{i=1}^{n!}(1- {\bf 1}_i) nf ~

457:                             \leq ~ \frac{nf}{n!}\frac{n!}{n^{1+\alpha}}\\

458:                            &= & \frac{f}{n^\alpha}

459:                             ~\rightarrow 0, ~~\mbox{as $n\to\infty$.}

460: \end{eqnarray*}

461: Step (a) uses  triangle inequality for a norm and step (b) uses

462: $||\Sigma_i||=1$ for any permutation matrix $\Sigma_i$.

463: \end{proof}

464: 

465: 

466: 

467: \vspace{.1in} From Lemma~\ref{lem1}, if $U(f)$ is feasible, then

468: $S(nf/2, \Sigma)$ is feasible for all $\Sigma\in S_n$. Thus, using

469: an argument identical to that in the proof of Theorem~\ref{lem1}, a

470: scaling law for $\bar{f}$ is equivalent to a scaling law for $f^*$,

471: i.e.

472: $$ f^* = \Theta(n\bar{f}).$$

473: 

474: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

475: 

476: 

477: 

478: \subsection{Wireline Networks: PMF Over a Graph}

479: 

480: We briefly review the key results known for PMF on graphs with

481: fixed edge capacities. These results will be useful in our analysis

482: for PMF for wireless networks.

483: 

484: Consider a directed graph $G=(V,E)$, where an edge $(i,j)\in E$ has

485: a capacity $C(i,j)$. Also, for $(i,j)\notin E$, we take $C(i,j)=0$.

486:  Then for a given $\pi$,

487: $f_\pi^*$ for graph $G=(V,E)$ is given by the solution of the

488: following linear program (LP).

489: %which can be solved in polynomial time~\cite{karmarkar_1984}.

490: \[\begin{aligned}

491:  \text{maximize}  \qquad & \qquad \qquad f,\\

492: \text{subject to} \qquad &

493: \sum_{k: (i,k)\in E}\bbl( x_{ij}(i,k) - x_{ij}(k,i) \bbr)  = f\pi(i)\pi(j), \ 1\leq i,j \leq n,\\

494: & \sum_{m: (k,m)\in E}\bbl( x_{ij}(k,m) - x_{ij}(m,k) \bbr)  = 0, \ \forall k\neq i,j,\ 1\leq i,j\leq n,\\

495: & \sum_{i=1}^n\sum_{j=1}^n x_{ij}(k,m)  \leq C(k,m), \ \forall (k,m)\in E,

496: \end{aligned}\]

497: where the variables are $f$ and $\{x_{ij}(k,m): (k,m)\in E, i,j,k,m = 1,\hdots, n\}$. The first two

498: are flow conservation constraints and the third one is the capacity constraint. The total number of variables is

499: less than $2n^4$ and the total number of constraints is less than $(n^3 + 2n^2)$. Hence, the above LP can

500: be solved in polynomial time~\cite{karmarkar_1984}.

501: 

502: The well-known max-flow min-cut characterization for a single

503: commodity flow naturally gives rise to the following question.

504: Though the maximum PMF $f^*_\pi$ for a given weight

505: vector can be computed in polynomial time, is there a corresponding

506: result that relates $f^*_\pi$ and the properties of the graph.

507: In their seminal paper, Leighton

508: and Rao \cite{LR88} obtained a characterization of $f^*_\pi$ in

509: terms of the weighted min-cut of graph. We summarize their

510: main result below. Let $p_\pi = | \{ i \in V : \pi(i) > 0 \}| $,  denote

511: the number of nodes for which the corresponding element of $\pi$

512: is non-zero. Then, without loss of generality we assume that

513: $\sum_{i=1}^n \pi(i) = p_\pi$.

514: 

515: \vspace{.1in}

516: 

517: 

518: \begin{definition}

519: For the graph $G$ and weight vector $\pi$, define the min-cut by

520: \[ \Upsilon(G, \pi) = \min_{U\subseteq V} \frac{\sum_{(i,j):i\in U, j\in U^C} C(i,j)}{\pi(U) \pi(U^c)},\]

521: with notation that $\pi(S) = \sum_{i\in S} \pi(i)$ for any set $S$.

522: \end{definition}

523: \begin{theorem}[Theorem 17, \cite{LR99}]

524: \label{thm:PMF_LR}

525: In any directed graph $G$, the maximum PMF for weight $\pi$ is related to $\Upsilon(G, \pi)$ as follows:

526: \[ \Omega\left( \frac{ \Upsilon(G, \pi)} {\log p_\pi}\right ) \leq f^*_\pi \leq  \Upsilon(G, \pi), \]

527: where the constants for the lower bound do not depend on the graph.

528: %where $n_\pi = \{ i \in V : \pi(i) > 0 \}$.

529: \end{theorem}

530: \vspace{.1in}

531: Note that the upper bound follows easily because for a given PMF $f_\pi$, the total flow

532: from $U$ to $U^C$ is $\pi(U)\pi(U^C)f_\pi$, which has to be less than the total capacity of the

533: links from $U$ to $U^C$.

534: The above characterization was crucial to the design of subsequent

535: approximation algorithms for many NP-hard problems; a summary of these algorithms can be found in~\cite{LR99}. An important case

536: of the above result is when $\pi(i)=1$ for all $i=1,\hdots, n$, i.e., the special case of uniform mulitcommodity flow.

537: In this case, we have

538: \[ \Upsilon(G) = \Upsilon(G, {\bf 1}) = \min_{U\subseteq V} \frac{\sum_{(i,j):i\in U, j\in U^C} C(i,j)}{|U||U^C|},\]

539: and

540: \[ \Omega\left( \frac{ \Upsilon(G)} {\log n}\right ) \leq f^* \leq  \Upsilon(G).\]

541: 

542: \vspace{0.1in}

543: %We note that  identical results can also be derived for directed graphs; the details can again be found in~\cite{LR99}.

544: 

545: 

546: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

547: \section{Combinatorial Interference Model}\label{sec2}

548: 

549: 

550: A combinatorial interference model defines constraints such that simultaneous data transmissions

551: over only certain sets of links (or edges) can be successful. This is a simplified abstraction of

552: a wireless network because in reality whether or not multiple simultaneous data transmissions

553: are successful depends on the rate of data transmission and the interference power at the various receivers.

554: We next describe the combinatorial interference model formally and illustrate it with example scenarios where

555: this abstraction is a reasonable one.

556: 

557: \subsection{Model}

558: \label{subsec:model}

559: A combinatorial interference model for a given set of wireless nodes $V=\{1,\hdots,n\}$

560: defines the following two objects:

561: \begin{itemize}

562: 

563: \item[(a)] A directed graph $G = (V, E)$ where $E$ is the set of directed links (edges)

564: over which data can be transmitted.

565: %Let $\delta$ denote the maximum vertex degree of this graph.

566: 

567: \item[(b)] For each directed edge $e \in E$, let  $\I(e) = \{ \hat{e} \in E \}$ be the set

568: of edges (directed links) that interfere with a transmission on link

569: $e$. Data can be successfully transmitted on link $e$ at rate $W(e)$

570: if and only if no transmission on any link in $\I(e)$ takes place

571: simultaneously. In general, the rate $W(e)$ for a given power

572: constraint can be different for different edges.

573: % Further, it can be a function of the set of links that transmit simultaneously.

574: The

575: proof methods and results of the paper will not change

576: (qualitatively) in this scenario. However, for ease of exposition

577: we will assume $W(e)=1$\footnote{As long

578: as $W(e)$ is bounded below and above by a constant, scaling laws

579: do not change even though the bounds for a given number of nodes $n$ will change.} for all $e \in E$.

580: \end{itemize}

581: 

582: \vspace{0.1in}

583: We assume that for every edge $(i,j)\in E$, edge $(j,i)\in E$,  i.e., the graph $G$ is essentially

584: an undirected graph without the interference constraints given by the sets $\I(e)$'s. This is a reasonable

585: assumption in many time division and frequency division systems where the channels are reciprocal~\cite{molisch_book}.

586: The interference sets $\I((i,j))$ and $\I((j,i))$ may not be identical because the transmissions which interfere

587: with a signal received at node $i$ may not be the same as transmissions which interfere with a signal

588: received at node $j$.

589: 

590: 

591: The above definitions can be used to induce a \emph{dual conflict

592: graph} as follows.

593: 

594: \vspace{0.1in}

595: \begin{definition}

596: The dual conflict graph  is an undirected graph $G^D = (E,E^D)$ with

597: vertex set $E$ and edge set $E^D$, where an edge $e^D\in E^D$

598: exists between $e_1$ and $e_2$ if $e_1$ and $e_2$ cannot transmit

599: simultaneously due to interference constraints.

600: Thus, each link $e\in E$ is connected to all links in

601: $\I(e)$.

602: \end{definition}

603: 

604: \vspace{0.1in}

605: 

606: 

607: For the rest of the section, we will suppress the explicit

608: dependence of all quantities on the combinatorial interference model

609: parameterized by the graphs $G$ and $G^D$; this helps to simplify

610: notation. Let us denote the node degree  and the chromatic number\footnote{The chromatic

611: number of a graph is the minimum number of colors needed to color the vertices of the graph

612: such that no two nodes  of the graph which are connected by an edge share the same color.}

613: dual conflict graph $G^D$ by $ \Delta = \max_{e \in E} |\I(e)|$ and

614: $\kappa$ respectively. Note that $\kappa\leq

615: (1+\Delta)$. %~(see, for example,~\cite{color}).

616: Let $\{E_k\}, E_k

617: \subseteq E$, be the set of all possible link sets that can be active

618: simultaneously, i.e., simultaneous transmissions

619: on all the links in $E_k$ at rate $W(e)=1$ are feasible for

620: the given interference model. Each $E_k$

621: corresponds to a vector $C_k\in \R^{|E|}$, where $C_k(e) =

622: {\mathbf 1}_{\{e \in E_k\}} = {\mathbf 1}_{\{e \in E_k\}} $.

623: Let $\cC$

624: be the convex-hull of all such vectors $\{C_k\}$. Thus $\cC$ is the set of all

625: vectors $C$ such that link capacities $C(e)$ (for link $e$) can be obtained

626: by time-sharing between the $C_k$'s for the given interference model. We then define the capacity

627: region to be the set of traffic matrices which can be routed over the graph $G = (V,E)$, where each edge

628: $e$ has capacity $C(e)$ for $C \in \cC$. The formal definition is as follows.

629: 

630: \vspace{0.1in}

631: \begin{definition}[Capacity Region ($\Lambda$)]

632: \label{def:cap_region}

633: The capacity region is the set of traffic matrices $\lambda \in \R^{n\times n}$ such that the following set

634: of conditions are feasible for some $C\in \cC$:

635: \begin{equation}

636: \label{eqn:cap_reg}

637: \begin{aligned}

638: &\sum_{k: (i,k)\in E}\bbl( x_{ij}(i,k) - x_{ij}(k,i) \bbr) = \lambda_{ij}, \ 1\leq i,j \leq n,\\

639: &\sum_{m: (k,m)\in E}\bbl( x_{ij}(k,m) - x_{ij}(m,k) \bbr) = 0, \ \forall k\neq i,j,\ 1\leq i,j \leq n,\\

640: &\sum_{i=1}^n\sum_{j=1}^n x_{ij}(k,m) \leq C(k,m), \ \forall (k,m)\in E,

641: \end{aligned}\end{equation}

642: where $C(k,m) = C(e)$ with $e=(k,m)\in E$; variables are $\{x_{ij}(k,m): (k,m)\in E, 1\leq i,j,k,m \leq n\}$.

643: \end{definition}

644: \vspace{0.1in} Thus, the capacity region consists of all traffic

645: matrices which are feasible using routing and link scheduling

646: (time-sharing between the sets $\{E_k\}$).  We now illustrate this

647: capacity region by a couple of special cases corresponding to widely

648: used models for wireless networks.

649: 

650: \subsubsection{Protocol Model}

651: The protocol model parameterized by the maximum radius of

652: transmission, $r$, and the amount of acceptable interference,

653: $\eta$, is defined in~\cite{GK} as follows.

654: \begin{itemize}

655:  \item[(a)] A node $i$ can transmit to any

656: node $j$ if the distance between $i$ and $j$, $r_{ij}$ is less than

657: the transmission radius $r$.

658: \item[(b)] For transmission from node $i$ to $j$ to be successful, no other node

659: $k$ within distance $(1+\eta) r_{ij}$ ($\eta > 0$ a constant) of

660: node $j$ should transmit simultaneously.

661: \end{itemize}

662: The corresponding definitions of $E$ and $E_D$ follow. A directed

663: link from node $i$ to node $j$ is in $E$ if $r_{ij}\leq r$. For a

664: link, $e\in E$, let $e^+$ denote the transmitter and let $e^-$

665: denote the receiver. Then

666: %\[{\cal I}(e)=\left\{\hat{e}\in E: \min\bl(d(e^+,\hat{e}^-),d(\hat{e}^+,e^-)\br)\leq (1+\eta)d(e^+,e^-)\right\}.\]

667: \[{\cal I}(e)=\left\{\hat{e}\in E:  r_{\hat{e}^+e^-} \leq (1+\eta)r_{e^+e^-}\right\}.\]

668: Thus the protocol model is a special case of the combinatorial interference model.

669: \vspace{.1in}

670: 

671: \subsubsection{SINR Threshold Model}

672: Assume that all transmissions occur at power $P$, and

673: the channel gain from the transmitter

674: of node $j$ to the receiver of node $i$ is given by $h_{ij}$, i.e.,

675: if node $j$ transmits at power $P$, the received signal power at

676: node $i$ will be $h_{ij}P$. A signal to interference and noise ratio (SINR) threshold model is parametrized by

677: a threshold $\alpha$ such that a transmission from node $i$ to node $j$ is successful if and only if the SINR is above

678: $\alpha$, i.e.,

679: \[ \frac{P h_{ij}}{ \sum_{k\neq i}Ph_{kj}+ N_0B} \geq \alpha \]

680: For example, if we assume that each link transmits Gaussian signals and that the Shannon capacity on each link is achievable,

681: then the threshold is given by $N_0B(2^W-1)$ (assuming $W(e)=W$ for all $e\in E$ as before).

682: 

683: We can define a corresponding combinatorial interference model such that the feasible

684: simultaneous transmissions defined by the combinatorial interference model are a subset of that

685: described by the SINR threshold model.

686: Consider the set of directed links

687: $E_\gamma$ such that a link, $(i,j)$, from node $i$ to node $j$, is

688: in $E_\gamma$ if and only if $h_{ji}\geq \gamma$. Also, define

689: ${\cal I}(e) = \{\hat{e}\in E_\gamma: h_{e^-\hat{e}^+ }\geq \beta\}

690: $.

691: Then link $e$

692: can transmit at rate $W$ if no other links in ${\cal I}(e)$ transmit simultaneously,

693: if and only if $\gamma$ and $\beta$ are such that

694: \begin{equation}

695: \label{eqn:interference}   \frac{P h_{e^-e^+}}

696: {\sum_{\hat{e}\in E_\gamma ,\ \hat{e}\notin {\cal

697: I}(e)}Ph_{e^-\hat{e}^+} + N_0B} \geq \alpha, \qquad \forall e\in

698: E_\gamma

699: \end{equation}

700: It is easy to see that the above condition is satisfied

701: if the following condition holds.

702: \[ \beta \leq \frac{1}{nP}\left(\frac{P\gamma}{2^W-1}-N_0B\right)\]

703: 

704: 

705: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

706: 

707: \subsection{Results}

708: 

709: We now derive results for the combinatorial interference model

710: which relate the maximum PMF $f_\pi^*$ to spectral properties of the

711: underlying graphs induced by the interference model. Most of the results

712: in this section use ideas from known results. While important

713: in their own right, these results and their proofs motivate the

714: results for an information theoretic setting for wireless networks

715: with Gaussian channels. Also, they provide alternate derivations

716: for known capacity scaling laws in random networks. Towards

717: the end of this section, we obtain simple bounds on the delay.

718: 

719: \vspace{.2in}

720: 

721: \subsubsection{Bounds on PMF}

722: For any $C\in \cC$, we denote the maximum PMF

723: on graph $G$ where each edge $e$ has capacity $C(e)$, by $f_\pi(C)$ , and the corresponding

724: min-cut by

725: $$ \Psi_\pi(C) = \min_{S \subset V} \frac{\sum_{(i,j): i \in S, j\in S^c} C(i,j)}{\pi(S)\pi(S^c)}. $$

726: We denote the corresponding quantities for the special case of UMF by $f(C)$ and $\Psi(C)$, respectively.

727: Then we have the following lemmas.

728: 

729: \begin{lemma}

730: \label{lem:cut_continuity}

731: $\Psi_\pi: \cC \mapsto \R$ is a continuous function for any $\pi \geq 0$.

732: \end{lemma}

733: \begin{proof}

734: Consider a cut $S$ such that $\pi(S)\pi(S^C) >0$. Then, the following is a continuous function of $C$:

735: $$\frac{\sum_{(i,j): i \in S, j\in S^c} C(i,j)}{\pi(S)\pi(S^c)}.$$

736: The lemma then follows since the minimum of a finite number of continuous functions is continuous.

737: \end{proof}

738: 

739: 

740: \vspace{0.1in}

741: 

742: \begin{lemma}

743: \label{lem:pmf_continuity}

744: $f_\pi: \cC \mapsto \R$ is a continuous function for any $\pi \geq 0$.

745: \end{lemma}

746: \begin{proof}

747: \iffalse

748: We first show that for any $C \in \cC$, we have the following:

749: \[f_\pi(C) =  \frac{f_\pi(\alpha C)}{\alpha},\]

750: where $\alpha >0$ is chosen such that $\alpha C\in \cC$.

751: By scaling all the variables in~(\ref{eqn:cap_reg}), it follows that

752: the flow $U(\alpha f)$ is feasible for $\alpha C$, and so

753: \[f_\pi(\alpha C) \geq \alpha f_\pi(C).\]

754: For the other direction, we argue by contradiction. Assume that

755: \[f_\pi(C) <  \frac{f_\pi(\alpha C)}{\alpha}.\]

756: Then for the LP in~(\ref{eqn:cap_reg}), for the vector $\alpha C$, we can scale

757: down all variables by $\alpha$, which means that a flow of greater than

758: $f_\pi(C)$ is feasible, which is incorrect.

759: \fi

760: For $\hat{C}\in \cC$ and any $\epsilon>0$ define the set

761: \[ {\cal B}_\delta = \{C \in \cC : ||C - \hat{C}||< \delta\}.\]

762: To prove the lemma, we have to show that for any $\epsilon >0$, there exists a

763: $\delta >0$ such that for all $C\in {\cal B}_\delta$,

764: $|f_\pi(C) - f_\pi(\hat{C})|<\epsilon$.

765: 

766: For  $\hat{C}\in \cC$ consider

767: \[ \delta_1 = \min_{(k,m)\in E} \left\{ \alpha \hat{C}(k,m) : \hat{C}(k,m) >0\right \}, \qquad 0<\alpha < 1,\]

768: and

769: \[ \delta = \min \left\{ \delta_1, \min_{i,j: \pi(i)\pi(j)>0}\frac{\epsilon}{2\pi(i)\pi(j)}\right \}.\]

770: Then for any $C \in {\cal B}_\delta $, it follows from~\ref{eqn:cap_reg} that $f_\pi(C) \leq f_\pi(\hat{C}) + \epsilon$.

771: It only remains to show that $f_\pi(C) \geq  f_\pi(\hat{C}) - \epsilon$.

772: For this note that $C \succ \underline{C}$  for all $C\in {\cal B}_\delta $, where $\underline{C}$ is as follows:

773: \[ \underline{C}(k,m) = \left \{

774: \begin{array}{ll}

775: 0 & \text{C(k,m) = 0}\\

776: \hat{C}(k,m) - \delta_1 & \text{otherwise}

777: \end{array}

778:   \right..\]

779: Now by scaling all the variables by $(1-\alpha)$ in the LP~(\ref{eqn:cap_reg}) for $\hat{C}$ and

780: using the monotonicity of $f_\pi(C)$ in $C$, we can see that

781: $f_\pi(C) \geq (1-\alpha) f_\pi(\hat{C})$ for all $C\in {\cal B}_\delta $. If $f_\pi(\hat{C})=0$,

782: we are done. If not, choose $\alpha = \min(\epsilon/f_\pi(\hat{C}), 0.5)$, which gives

783: $f_\pi(C) \geq  f_\pi(\hat{C}) - \epsilon$, and so we are done.

784: 

785: \end{proof}

786: \vspace{0.1in}

787: 

788: 

789: We now define a quantity for the combinatorial interference model

790: corresponding to the min-cut of a graph.

791: 

792: \vspace{0.1in}

793: \begin{definition}

794: The min-cut for the combinatorial interference model is defined as

795: $$ \Psi_\pi^* = \max_{C \in \cC} \min_{S \subset V} \frac{\sum_{(i,j): i \in S, j\in S^c} C(i,j)}{\pi(S)\pi(S^c)}. $$

796: \end{definition}

797: 

798: \vspace{0.1in}

799: 

800: Note that $\Psi_\pi^*$ is well defined since $\Psi_\pi(C)$ is a continuous function of $C$, and $\cC$ is closed and

801: bounded because it is the convex hull of a finite number of points.

802: The above definition can be interpreted as the min-cut of the graph $G$, where each edge has capacity $C(e)$, and

803: the vector $C$ is chosen from the set $\cC$ such that it maximizes the min-cut of this graph $G$.

804: The following result is an extension of Theorem~\ref{thm:PMF_LR} to combinatorial

805: interference models.

806: \vspace{0.1in}

807: \begin{theorem}

808: \label{thm:UMF_mincut}

809: $f^*_\pi$ is bounded as

810: \begin{equation}

811: \Omega\left(\frac{\Psi_\pi^*}{\log p_\pi} \right) \leq f_\pi^* \leq \Psi_\pi^*.

812: \end{equation}

813: %where we recall that $n_\pi = \{ i \in V : \pi(i) > 0\}$.

814: \end{theorem}

815: \vspace{0.1in}

816: \begin{proof}

817: Since $\cC$ is closed and bounded, it follows from Lemma~\ref{lem:pmf_continuity} that

818: there exists $C^*\in \cC$ such that $f^*_\pi = f_\pi(C^*)$. Then, using

819: Theorem~\ref{thm:PMF_LR}, it follows that

820: \[

821: f^*_\pi \leq   \Psi(C^*) \leq \Psi_\pi^*.

822: \]

823: %where $n_\pi = \{ i \in V : \pi(i) > 0 \}$.

824: 

825: Now, from Lemma~\ref{lem:cut_continuity}, there is $\hat{C}\in \cC$ such that

826: $ \Psi_\pi^* = \Psi_\pi(\hat{C})$. Using Theorem~\ref{thm:PMF_LR}, it follows that

827: \[f_\pi^* \geq  f_\pi(\hat{C}) = \Omega\left(\frac{\Psi_\pi^*}{\log p_\pi } \right). \]

828: This completes the proof of Theorem~\ref{thm:UMF_mincut}.

829: 

830: 

831: \end{proof}

832: 

833: 

834: 

835: Note that unlike the case for wireline networks (or equivalently graphs), $f^*$ is a hard

836: quantity to compute. Also, note that $\Psi$ is a function of both $G$ and the dual graph $G^D$.

837: We next relate the maximum UMF $f^*$ to spectral properties of graphs $G$ and $G^D$.

838: 

839: \vspace{0.1in}

840: 

841: \begin{definition}

842: The conductance of a graph $G$ is defined as follows.

843: \[\Phi(G) = \min_{U\subseteq V, |U|\leq n/2}\frac{\sum_{i\in U, j\in U^C} {\bf 1}_{[(i,j)\in E] }}{|U|},\]

844: where ${\bf 1}_{[.]}$ is the indicator function.

845: \end{definition}

846: %Then we have the following result.

847: \vspace{0.1in}

848: 

849: \begin{corollary}

850: \label{thm:UMF_cond}

851: Recall that $\kappa$ is the chromatic number of the dual graph $G^D$.

852: Then, $f^*$ is related to $\Phi(G)$ as follows.

853: \[\Omega\left(\frac{\Phi(G)}{\kappa n\log n} \right) \leq f^* \leq \frac{\Phi(G)}{n}.\]

854: \end{corollary}

855: \begin{proof}

856: Consider vertex coloring for the dual graph $G^D = (E,E^D)$.

857: The chromatic number of $G^D$ is defined to be $\kappa$ and hence we need

858: $\kappa$ colors for vertex coloring of $G^D$.  Thus we have partitioned

859: the set $E$ into subsets, say, $E_1,\hdots, E_{\kappa}$ such that the links

860: in each subset can transmit simultaneously at rate 1.

861: Now let $C_k(e) = {\bf 1}_{\{e\in E_k\}}$. Then, $C$ corresponding to

862: uniform time-sharing between the $\kappa$ edge sets $E_1,\dots, E_k$

863: is given  by

864: \[ C = \frac{1}{\kappa}(C_1 + \hdots + C_{\kappa}), \]

865: which is a convex combination of $C_1,\hdots,C_{\kappa}\in \cC$.

866: Hence, $C(i,j) = 1/\kappa$ for all $i\neq j$, and $C\in \cC$. Then, using Theorem~\ref{thm:PMF_LR} and the definition

867: of conductance above,

868: \[ f^* \geq f^*(C) \geq \Omega\left( \frac{\Psi}{\log n} \right) = \Omega\left(\frac{\Phi(G)}{\kappa n\log n} \right).\]

869: \end{proof}

870: 

871: For the upper bound, note that for any $C\in \cC$, $C \preceq {\bf 1}$, i.e., $C$ is lexicographically

872: less than ${\bf 1}$, and $f(C_1) \geq f(C_2)$ if $C_1\geq C_2$. Hence, $f^* = \max_{C\in \cC} f(C) \leq f({\bf 1})$. Then, the upper bound follows again

873: by a straightforward use of Theorem~\ref{thm:PMF_LR} and the definition

874: of conductance.

875: \vspace{0.2in}

876: 

877: \subsubsection{Average Delay}

878: 

879: We now provide bounds on the average delay for a class of traffic matrices.

880: We measure delay in number of hops. We assume that the packet size is small

881: enough so that the packet delay is essentially equal to the number of hops

882: taken by the packet. This is similar to the assumptions

883: in, for example, \cite{GK},\cite{GT}, and \cite{EMPS_journal}.

884: 

885: We restrict ourselves to periodic link scheduling schemes (similar

886: arguments extend to any ergodic scheduling scheme as well). For

887: fixed networks, $\cC$ is the convex hull of the set, $\{C_k\}$,

888: which has a finite cardinality. Hence, any vector in $\cC$ can be

889: written as a linear combination of the $C_k$'s. Thus to maximize the

890: supportable uniform multicommodity flow it is sufficient to optimize

891: over transmission schemes with periodic scheduling of links where

892: the periodic schedule corresponds to time division between the

893: $C_k$'s.

894: 

895: We obtain the following general scaling of delay.

896: \vspace{.05in}

897: \begin{theorem}\label{thm4}

898: Let $S(n)$ be the total number of transmissions by the $n$ wireless nodes

899: on average per unit time\footnote{The quantity $S(n)$ is well defined

900: since we consider periodic scheduling of links.}. When data is transmitted according to rate matrix

901: $\lambda\in \Lambda$,  the average delay, $D(n)$, over all packets  scales as

902: $$ D(n) = \Theta\left(\frac{S(n)}{\bar{\lambda}}\right),

903: \mbox{~~where $\bar{\lambda} = \sum_{i,j} \lambda_{ij}$.}$$

904: \end{theorem}

905: \begin{proof}

906: Let $\Gamma$ denote the set of all  possible

907: paths (without cycles) in the network. The amount of flow generated at node $i$

908: to be transmitted to node $j$ is $\lambda_{ij}$. Let us consider an arbitrary but fixed\footnote{Here,

909: we consider a fixed deterministic scheme. However, it is easy to see that the result

910: extends for any randomized scheme as well.}

911: routing scheme where a fraction $\alpha_{ij}^\gamma$ of

912: the flow from node $i$ to node $j$ is routed over path $\gamma \in\Gamma$. We assume that

913: the traffic matrix $\lambda$ is feasible.  Hence, there exists a link scheduling

914: and routing scheme to support it. The total number of transmissions per unit time at node $l$

915: is $\sum_{\gamma \ni l}\sum_{i,j} \alpha_{ij}^\gamma\lambda_{ij}$. Hence, the average

916: number of transmissions per unit time in the entire network, denoted by $S(n)$, is

917: \[ \begin{aligned}

918: S(n) \ =\ \sum_{l=1}^n\sum_{\gamma \ni l}\sum_{i,j}

919: \alpha_{ij}^\gamma\lambda_{ij}

920:      \ =\  \sum_{\gamma} H^\gamma \sum_{i,j} \alpha_{ij}^\gamma \lambda_{ij}\\

921: \end{aligned}

922: \]

923: where $H^\gamma$ is the number of hops on path $\gamma$.

924: The total flow over a path $\gamma$ is $\sum_{i,j} \alpha_{ij}^\gamma\lambda_{ij}$, i.e.,

925: the fraction of total flow over path $\gamma$ is $\sum_{i,j} \alpha_{ij}^\gamma\lambda_{ij}/\bar{\lambda}$.

926: Hence, the average number of hops traversed by all

927: packets is given by

928: \[D(n) = \frac{1}{\bar{\lambda}}\sum_\gamma H^\gamma \sum_{i,j} \alpha_{ij}^\gamma\lambda_{ij} =\frac{S(n)}{\bar{\lambda}}. \]

929: This completes the proof of Theorem \ref{thm4}.

930: \end{proof}

931: 

932: 

933: \vspace{.2in}

934: 

935: We note that the above result uses very little information about the specific underlying

936: transmission scheme.

937: Next, we present an immediate corollary of the above result and Theorem~\ref{thm:UMF_cond} that

938: bounds the delay for a scheme that maximizes the UMF in the network.

939: 

940: \vspace{.05in}

941: 

942: \begin{cor}

943: \label{cor2} Since $f^* = \Omega\left(\frac{\Phi(G) }{\kappa n\log n} \right)$, the

944: corresponding delay scales as $D(n) = O\left( \frac{|E|\log n}{n\Phi(G)} \right)$.

945: \end{cor}

946: \begin{proof}

947: Consider the link scheduling scheme in the proof of

948: Corollary~\ref{thm:UMF_cond}, where we partition the set of links $E$ into

949: subsets $E_1,\hdots,E_\kappa$ such that all the links in each subset

950: $E_i$ can transmit simultaneously. Note that this scheme can support UMF

951: $f = \Omega\left(\frac{\Phi(G)}{\kappa n\log n}\right)$. For this transmission

952: scheme, every link transmits

953: at rate 1 for at most $1/\kappa$ fraction of the time. Hence, we have $ S(n) \leq

954: \frac{|E|}{\kappa}$. Thus, it follows from Theorem~\ref{thm4} that

955: $D(n) = O\left( \frac{|E|\log n}{n\Phi(G)} \right)$.

956: 

957: 

958: \end{proof}

959: 

960: \vspace{.1in}

961: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

962: 

963: \subsection{Computational Methods}

964: 

965: We now describe computational methods to obtain bounds on $f^*$ (the extensions

966: to PMF are straightforward). As

967: noted earlier, for wire-line networks, the computation of $f^*$ is

968: equivalent to solving an LP. However, in a wireless network, the

969: link capacity is a function of the link schedule. Since, the number

970: of link schedules is combinatorial, it is a hard problem.

971: Specifically, the question of checking feasibility of a rate vector

972: $\lambda$ was proved to be NP-hard by Arikan \cite{Ar84}, that is

973: there exists an interference model and graph under which checking

974: feasibility of $\lambda$ is NP-hard.

975: % citation: E. Arikan, Some Complexity Results about Packet Radio Networks,

976: % IEEE Transaction on Information Thoery, 30:4, pp: 681-685.

977: Motivated by this, here we address the question of providing a

978: simple computational method to bound $f^*$.

979: 

980: We use ideas of node coloring to induce a link schedule in a way

981: similar to, for example,~\cite{kodialam_2003}. In particular,

982: we can obtain an upper bound $f_1^*$ and a lower

983: bound $f_2^*$ for maximum UMF $f^*$ in polynomial time such that

984: \[ f_1^* \leq \kappa f_2^* \]

985: The upper bound can be computed by solving the LP in~(\ref{eqn:cap_reg})

986: with $C(e)=1$ for all $e\in E$. For the lower bound, since the dual

987: graph $G^D$ has chromatic number $\kappa$, we can color the nodes of

988: $G^D$ (which are given by the set $E$ of wireless links) such no

989: two nodes which share an edge share the same color. This in turn

990: induces a link scheduling scheme, where each link in $E$ is scheduled

991: for at least a fraction $1/\kappa$ of time, and the resutling $C$ is

992: such that $C(e)\geq 1/\kappa$ for all $e\in E$. Again, the lower bound

993: can be computed by solving the LP in~(\ref{eqn:cap_reg}) with

994: $C(e)=1/\kappa$ for all $e\in E$. It is easy to see that

995: $f_1^* \leq \kappa f_2^*$.

996: 

997: Now from Theorem~\ref{thm:UMF_mincut}, we know that

998: $\Omega\left(\frac{\Psi}{\log n} \right) \leq f^* \leq \Psi$.

999: Thus, we can now also bound $\Psi$ as

1000: \[ f_2^*     \leq   \Psi \leq O\bl( f_1^* \log n\br ) \]

1001: Thus, the upper and lower bounds differ by at most a factor

1002: of $\kappa \log n$. In addition, using the algorithm

1003: in~\cite{LR88}, we can find a vector $C(e)$ and the

1004: corresponding cut $(U,U^C)$ such that the capacity of

1005: this cut, $\min_{S \subset V} \frac{\sum_{(i,j): i \in S, j\in S^c} C(i,j)}{|S||S^c|}$,

1006: is within a factor $\kappa (\log n)^2$ of $\Psi$.

1007: 

1008: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1009: 

1010: \subsection{Application}

1011: We now illustrate our results for the combinatorial interference model through an application

1012: to \emph{geometric random graphs}.

1013: The geometric random graph has been widely used to model the topology of wireless networks after the work

1014: of Gupta and Kumar~\cite{GK}. However, it has been a combinatorial object of interest for more than 60 years.

1015: We derive scaling laws for a combinatorial interference model which is more restrictive than the protocol model. Note that

1016: the lower bound hence obtained is also a lower bound for the protocol model. Specifically, the lower bound is weaker by $\log^{2.5}n$

1017: compared to the lower bound obtained in~\cite{GK}. Specifically, we show that the scaling of the lower bound is closely tied

1018: to the known combinatorial properties of geometric random graphs.

1019: 

1020: 

1021: We first define the \emph{restricted protocol interference model}.

1022: It is also parameterized by the maximum radius of

1023: transmission, $r$, and the amount of acceptable interference,

1024: $\eta$.

1025: \begin{itemize}

1026:  \item[(a)] A node $i$ can transmit to any

1027: node $j$ if the distance between $i$ and $j$, $r_{ij}$ is less than

1028: the transmission radius $r$.

1029: \item[(b)] For transmission from node $i$ to $j$ to be successful, no other node

1030: $k$ within distance $(1+\eta) r$ (where $\eta > 0$  is a constant) of

1031: node $j$ should transmit simultaneously.

1032: \end{itemize}

1033: 

1034: \vspace{0.2in}

1035: 

1036: We now state a version of the well-known Chernoff's bound

1037: for binomial random variables that we use multiple times in this paper.

1038: 

1039: 

1040: 

1041: \begin{lemma}\label{lem:chernoff}

1042: Let $X_1,\dots, X_N$ be i.i.d. binary random variables with $\Pr(X_1 = 1) = p$. Let $S_n = \sum_{k=1}^n X_k$ for

1043: $n=1,\hdots, N$.

1044: Then, for

1045: any $\delta \in (0,1)$

1046: $$ \Pr\left( | S_n - np | \geq \delta n p \right) \leq 2 \exp\left(-\frac{\delta^2np}{2}\right).$$

1047: Specifically, for $\delta = \sqrt{\frac{2L \log n}{np}}$, we have

1048: $$ \Pr\left( |S_n - np| \geq \sqrt{2L np \log n} \right) \leq \frac{2}{n^L}.$$

1049: \end{lemma}

1050: 

1051: 

1052: \vspace{.2in}

1053: 

1054: 

1055: Consider $n$ wireless nodes  distributed uniformly at random in a unit

1056: square, and the interference model given by the protocol model with transmission

1057: radius $r$. We denote such a wireless network by $G(n,r)$.

1058: It is well-known that for $G(n,r)$ to be connected with high probability,

1059: it is necessary to have $r = \Omega(\sqrt{\log n/n})$.

1060: We take $r = \Theta(\log^{3/4} n/\sqrt{n})$ and prove the following bounds on the maximum UMF, $f^*$,

1061: for the restrictive protocol model;

1062: the lower bound is only $\log^{2.5} n$ weaker than the result of Gupta and Kumar for the protocol model

1063: with $r = \Theta(\sqrt{\log n/n})$.

1064: \begin{lemma}\label{lem:gnrcomb}

1065: For $G(n,r)$, with $r = \Theta(\log^{3/4} n/\sqrt{n})$, maximum UMF is bounded as

1066: $$ \Omega\left(\frac{1}{n^{3/2} \log^{5/2} n}\right) ~\leq~ f^* ~\leq~O\left(\frac{1}{n^{3/2} \log^{3/4} n}\right).$$

1067: \end{lemma}

1068: \begin{proof}

1069: To prove the above bounds, we obtain appropriate upper and lower bounds on the

1070: quantity $\Psi$. These bounds along with  Theorem \ref{thm:UMF_mincut}

1071: imply Lemma \ref{lem:gnrcomb}. To obtain an upper bound on $\Psi$,

1072: we evaluate the cut-capacity for a specific cut-set. For

1073: the lower bound, we first, establish that a grid

1074: graph on $n$ nodes is a sub-graph of $G(n,r)$ and then use the known conductance

1075: of the grid graph.

1076: 

1077: First, consider the upper bound on $\Psi$. Specifically, consider the square, say $\cS$, of area

1078: $1/9$ (of side $1/3$) that is in the center of the unit square. Let

1079: $S$ be the set of nodes that fall inside this square.

1080:  By definition, we have

1081:  $$ \Psi ~\leq~ \Psi(S) ~=~ \sup_{C\in \cC} \frac{\sum_{i\in S, j\in S^c} C(i,j)}{|S||S^c|}.$$

1082: Therefore, it is sufficient to required obtain bound on $\Psi(S)$.

1083: 

1084: 

1085: Corresponding to

1086: node $i$, define a random variable $X_i \in \{0, 1\}$  which is $1$ is $i$

1087: is in $S$, and 0 otherwise.

1088: Since nodes

1089: are placed uniformly and independently at random in the unit area square,

1090: $X_i$ are i.i.d. binary random variable with $\Pr(X_i = 1) = 1/9$. Now,

1091: $\sum_{i=1}^n X_i$ is the number of nodes in $S$. Using Lemma~\ref{lem:chernoff} with

1092: $\delta = 0.009$, it follows that for large enough $n$,

1093: $|S| \in (0.1n,0.2n)$ (and so $|S^c| \in [0.8n, 0.9n]$) with probability at least

1094: $1-n^{-4}$. Now, consider squares $\cS^0, \cS^1$ of sides $1/3 + 2r$

1095: and $1/3 -2r$ respectively with their centers being the same as

1096: that of $\cS$. That is, $\cS^1 \subset \cS \subset \cS^0$. Let

1097: $\cA^0 = \cS^0 - \cS$ and $\cA^1 = \cS - \cS^i$.

1098: Thus, $\cA^1$ is a strip of width $r$ surrounding $\cS$ of

1099: and $\cA^0$ is a strip of width $r$ on the boundary and inside $\cS$.

1100: Since, $r = \Theta(\log^{3/4} n/\sqrt{n})$, it can be easily shown that

1101: $\cA^0$ and $\cA^1$ is $\Theta(r)$.

1102: 

1103: \iffalse It is useful

1104: to visualize that $\cA^o$ is a "strip" surrounding $\cS$ of

1105: width $r$ while $\cA^i$ is a "strip" on the boundary of $\cS$

1106: of width $r$. By Protocol model, two nodes can transmit

1107: to each other only if they are within distance $r$ of each

1108: other. Based on this, it is not hard to show that area of

1109: $\cA^o$ and $\cA^i$ is $\Theta(r)$ for large enough $n$.

1110: \fi

1111: 

1112: 

1113: Now, nodes that are

1114: in $S$ (i.e. physically inside $\cS$) can only be connected

1115: to those nodes in $S^c$ that lie in $\cA^1$. Similarly, nodes

1116: in $S^c$ that are connected to nodes $S$ must lie in $\cA^1$.

1117: Thus, nodes that can communicate across the cut $(S, S^c)$ must

1118: lie within a region of area $\Theta(r)$. For the protocol model,

1119: if a node transmits, nodes within distance $r(1+\eta)$ of the receiver

1120: must not

1121: transmit. That is, each transmission effectively silences nodes

1122: within an area of $\Theta(r^2)$.

1123: Thus, at any given time, the maximum number

1124: of simultaneous transmissions between $S$ and $S^c$ is $\Theta(1/r)$.

1125: This along with $|S|, |S^c| =\Theta(n)$ implies that

1126: $$ \Psi ~\leq~ \Psi(S) ~= ~ \frac{O(1/r)}{\Theta(n^2)} = O\left(\frac{1}{n^2 r}\right)~=~O\left(\frac{1}{n^{3/2} \log^{3/4} n}\right).$$

1127: 

1128: 

1129: 

1130: For the lower bound, we identify

1131: a grid subgraph of $G(n,r)$

1132: with $r = \Theta(\log^{3/4} n/\sqrt{n})$.

1133: Consider a grid graph $G_n$ of $\sqrt{n}\times\sqrt{n}$ nodes

1134: with each node connected to one of its four neighbors (with

1135: suitable modifications at the boundaries). The nodes of $G_n$ are placed in

1136: a uniform manner in a unit square; each node is at a distance

1137: $1/\sqrt{n}$ from its neighbors.

1138: Now consider a minimax matching between

1139: nodes of $G_n$ and $n$ randomly placed nodes in the unit square,

1140: where a minimax matching is a perfect matching between the $n$

1141: nodes of $G_n$ and the nodes of $G(n,r)$ with maximum

1142: length minimized. Leighton and Shor \cite{LS86} established

1143: that the maximum edge length in a minimax matching, say $r^*$, is

1144: $\Theta(\log^{3/4} n/\sqrt{n})$ with probability at least $1-1/n^4$.

1145: Now we identify the subgraph $G^\prime_n$ (with grid graph

1146: structure) of $G(n,r)$ as follows. $G^\prime_n$ has all $n$ nodes.

1147: Consider the minimax matching between $G_n$ and $G(n,r)$. If a node

1148: of $G(n,r)$ is connected to node number $m$ of $G_n$, then renumber

1149: it as $m$ to obtain nodes of $G^\prime_n$. Now by setting

1150: $r \geq r^* + 2/\sqrt{n}$, clearly a node $m$ and $m'$ are connected

1151: in $G^\prime_n$ if they are connected in $G_n$. Thus, we have established

1152: that $G_n \subset G^\prime_n$. Now, we will focus only on the edges

1153: of $G(n,r)$ that belong to $G^\prime_n$ and provide them with positive capacity by an appropriate

1154: communication scheme that is feasible for the restricted protocol model.

1155: For this, note that in $G(n,r)$

1156: each node is connected to at most $O(\log^{3/2}n)$ nodes with probability

1157: at least $1-1/n^4$ (using Chernoff's bound and Union bound) for large enough

1158: $n$. Hence, using a simple TDMA scheme based on vertex coloring of $G(n,r)$,

1159: each node gets to transmit once in every $\Theta(1/\log^{3/2} n)$ time slots.

1160: This transmission can be along any outgoing edge.  Since, we are interested

1161: in providing positive capacity to only at most $4$ outgoing edges,

1162: we have established that there is a simple TDMA scheme which provides

1163: $\Theta(1/\log^{3/2} n)$ capacity to each edge of a grid subgraph of

1164: $G(n,r)$. To complete the proof, we recall that

1165: the conductance of a grid graph is $\Theta(1/\sqrt{n})$ \cite{info_gossip}.

1166: That is,

1167: $$ \Phi(G_n) = \min_{S} \frac{\sum_{i\in S, j\in S^c} {\mathbf 1}_{\{(i,j) \in E\}}}{|S||S^c|} = \Theta\left(\frac{1}{n^{3/2}}\right).$$

1168: Now, putting all the above discussion together we have the following.

1169: \begin{eqnarray}

1170: \Psi & = & \sup_{C\in\cC} \min_{S\subset V} \frac{\sum_{i \in S, j \in S^c} C(i,j)}{|S||S^c|} \nonumber \\

1171:      & \geq & \Phi(G_n) \Theta\left(\frac{1}{\log^{3/2} n}\right) \nonumber \\

1172:      & = & \Omega\left(\frac{1}{n^{3/2} \log^{3/2} n}\right).

1173: \end{eqnarray}

1174: In summary, upper and lower bound on $\Psi$ along with Theorem \ref{thm:UMF_mincut}

1175: implies the Lemma \ref{lem:gnrcomb}.

1176: \end{proof}

1177: \vspace{.1in} Now, we discuss briefly delay. In~\cite{EMPS_journal}, delay

1178: was defined as the average number of hops per packet, and the

1179: packet size was assumed to scale to an arbitrarily small value. For any

1180: communication scheme feasible for the protocol model with maximum transmission radius

1181: $r = \Theta(\log^{3/4} n/\sqrt{n})$, the maximum number of transmissions per

1182: unit time is upper bounded as $O(n/\log^{3/2} n)$. Using this and

1183: Theorem \ref{thm4} we obtain the following result immediately.

1184: 

1185: \begin{lemma}\label{lem:delay}

1186: The delay $D(n)$ for any scheme achieving

1187: $f^* = \Omega\left(\frac{1}{n^{3/2} \log^{5/2} n}\right)$ is bounded

1188: above as

1189: $$D(n)  =  O\left(\sqrt{n}\log n\right).$$

1190: \end{lemma}

1191: 

1192: \vspace{.2in}

1193: 

1194: 

1195: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1196: 

1197: \section{Gaussian Fading Channel Model}\label{sec3}

1198: 

1199: In the previous section, we assumed that the wireless

1200: network was defined by two graphs $G$ and $G^D$.

1201: We extended the results of Leighton and Rao

1202: to wireless networks modeled by a combinatorial interference model;

1203: this mainly exploited the fact that all possible transmission schemes could

1204: be described in terms of routing over a set of capacitated graphs, where

1205: the set of edge capacity vectors belonged to the convex hull of a finite

1206: number of vectors. Thus, in this sense, the inherent \emph{discrete}

1207: nature of the model worked to our advantage.

1208: 

1209: While the combinatorial interference model

1210: can allow for arbitrary scheduling and routing

1211: schemes, it does not model all the degrees of freedom

1212: in a wireless network. Specifically, the results are

1213: not {\em information theoretic}.

1214: In this section, we provide an information

1215: theoretic characterization of PMF in a wireless

1216: network with Gaussian fading channels.

1217: The techniques for the combinatorial model can be easily extended to obtain a feasible scheme and

1218: a lower bound on the maximum PMF $f^*$. However, for information theoretic upper

1219: bounds we have to work harder, especially to obtain a bound that relates to the lower bound

1220: and allows us to quantify the gap.

1221: 

1222: Our key contribution is in quantifying the suboptimality of the UMF/PMF for

1223: a \emph{simple feasible} scheme and an upper bound on the UMF/PMF

1224: for an arbitrary network topology, in terms of a simple

1225: graph property. The bound is general when channel side information (CSI) is assumed to be

1226: available only at the receiver.

1227: For AWGN channels, we quantify only for UMF, and when the SNR is low enough.   To the best

1228: of our knowledge, this is the first such result which guarantees that a feasible scheme

1229: achieves rates within a certain factor of an outer bound for an arbitrary graph. We also

1230: illustrate these results through applications.

1231: The results hence obtained are interesting in their own right.

1232: 

1233: 

1234: 

1235: 

1236: Our main approach is as follows. We construct

1237: two directed capacitated graphs $G^U$ and $G^L$ for

1238: the given wireless network. The graph $G^U$ is such that

1239: the capacity (defined appropriately later) of each

1240: cut in $G^U$ upper bounds the corresponding cut-capacity

1241: in the wireless network. The graph $G^L$ is such that

1242: there exists a communication scheme

1243: that simultaneously achieves the capacity of each edge in

1244: $G^L$, and the ratio of capacity of each cut in $G^U$

1245: and $G^L$ is bounded above by a quantifiable term.

1246: This leads to an approximate characterization of

1247: PMF in an arbitrary wireless network with Gaussian

1248: fading channels. Moreover, the feasible scheme that induces the capacities in $G^L$

1249: supports PMF which is within a quantifiable factor of the optimal.

1250: 

1251: 

1252: 

1253: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1254: 

1255: \subsection{Channel Model}

1256: 

1257:  This is similar to the model in, for example,~\cite{JVK}. We have

1258: $V=\{1,\dots, n\}$ wireless nodes with transceiver capabilities

1259: located arbitrarily in a plane. Node transmissions happen at

1260: discrete times, $t \in \Z_+$. Let $X_i(t)$ be the signal transmitted

1261: by node $i$ at time $t \in \Z_+$. We assume that each node has a

1262: power constraint\footnote{For notational simplicity we assume that

1263: each node has the same power constraint. The general case, where

1264: each node has different maximum average power can be handled using

1265: identical techniques.} such that

1266: \mbox{$\limsup_{N\rightarrow\infty}\frac{1}{N}\sum_{t=1}^N

1267: |X_i^2(t)| \leq P$.}  Then $Y_i(t)$, the signal received by node $i$

1268: at time $t$, is given by

1269: \begin{equation} \label{eqn:gauss_ch}

1270: \textstyle Y_i(t) = \sum_{k \neq i} H_{ik} X_k(t) + Z_i(t),

1271: \end{equation}

1272: where $Z_i(t)$ denotes a complex zero mean white Gaussian noise

1273: process with independent real and imaginary parts with variance 1/2

1274: such that $Z_i(t)$ are i.i.d. across all $i$. Let $r_{ij}$ denote the

1275: distance between nodes $i$ and $j$. Let $H_{ik}(t)$ be such that

1276: $$ H_{ik}(t) = \sqrt{g(r_{ik})} \hat{H}_{ik}(t), $$

1277: where $\hat{H}_{ik}(t)$ is a stationary and ergodic zero mean

1278: complex Gaussian process with independent real and imaginary parts

1279: (with variance 1/2). It models channel fluctuations due to frequency

1280: flat fading. Also, $g(\cdot)$ is a monotonically decreasing function

1281: that models path loss with $g(x) \leq 1$ for all $ x \geq 0$. We

1282: assume also that the $\hat{H}_{ik}(t)$'s are independent.

1283: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1284: 

1285: \subsection{Graph Definitions}

1286: 

1287: Consider the following two graphs induced by a wireless network of

1288: $n$ nodes:

1289: \begin{itemize}

1290: \item[(1)]~$K_n$ is the fully

1291: connected graph with node set $V$;

1292: \item[(2)]

1293: $G_r$ is the graph where

1294: each node $i \in V$ is connected to all nodes that are within a

1295: distance $r$ of $i$. Let $E_r$ denote the edge set of $G_r$.

1296: Let $\Delta(r)$ be the maximum vertex degree of $G_r$. Finally, define

1297: $$ r^* = \min\{r : ~G_r ~\text{~is connected}\}. $$

1298: \end{itemize}

1299: 

1300: \vspace{0.1in}

1301: 

1302: 

1303: 

1304: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1305: \subsection{Preliminaries}

1306: In the analysis in this section, we utilize the following two simple lemmas.

1307: 

1308: \begin{lemma}\label{lem:ineq1}

1309: Given $x_i \in (0,1), 1\leq i\leq N$,

1310: $$ \textstyle \sum_{i=1}^N \log (1+\sqrt{x_i}) \leq \sqrt{2N} \sqrt{\sum_{i=1}^N \log(1+x_i)}.$$

1311: \end{lemma}

1312: \begin{proof}

1313: For any $x \in (0,1)$, $x/2 \leq \log (1+x) \leq x$, so

1314: \begin{eqnarray}

1315: \textstyle \sum_{i=1}^N \log (1+\sqrt{x_i}) & \leq & \textstyle \sum_{i=1}^N \sqrt{x_i} \leq  \sqrt{N} \sqrt{\sum_{i=1}^N x_i}  \label{eq:leq2x}\\

1316: & \leq & \textstyle \sqrt{2N} \sqrt{\sum_{i=1}^N \log(1+x_i)},\nonumber

1317: \end{eqnarray}

1318: where (\ref{eq:leq2x}) follows from Cauchy-Schwarz inequality.

1319: \end{proof}

1320: 

1321: \vspace{0.2in}

1322: 

1323: \begin{lemma}

1324: \label{lem:ineq2}

1325: For any $x \geq 0$, $\alpha \in (0,1)$,

1326: $\frac{1}{\alpha} \log (1 + \alpha x) \geq \log (1+x)$.

1327: \end{lemma}

1328: \begin{proof}

1329:     Define $f(x) = \frac{1}{\alpha} \log (1+ \alpha x) - \log (1+x)$. Note that $f'(x) \geq 0$ for $x \geq 0$ and $f(0) = 0$.

1330: \end{proof}

1331: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1332: \subsection{Results}

1333: We obtain bounds on the maximum PMF for three different cases:

1334: \begin{itemize}

1335: \item[(1)] fading channel with AWGN, and channel side information (CSI) available only

1336: at the receiver,

1337: \item[(2)] deterministic (no fading) additive white Gaussian noise (AWGN) channel,

1338:  and

1339: \item[(3)] fading channel with AWGN, and CSI available at both the transmitter and the receiver.

1340: \end{itemize}

1341: The exact bounds for the above cases are different, but the analysis and bounding techniques are similar.

1342: 

1343: \vspace{0.3in}

1344: 

1345: \subsubsection{Random Fading with Rx-only CSI}

1346: We first obtain bounds on the PMF for Gaussian channels with random fading under the assumption

1347: that CSI is available at the receiver, but not the transmitter. We then relate the bounds for PMF, and show

1348: that the gap can be quantified well, and under very general assumptions. We note that this is the case for which we can obtain the

1349: strongest results.

1350: \vspace{0.1in}

1351: \begin{theorem}

1352: \label{thm:coh_bounds}

1353: With channel state information (CSI) only at receivers, $f^*_\pi$ is

1354: bounded as follows:

1355: \[\begin{aligned}

1356: &f^*_\pi \leq  \min_{S\subset V} \frac{\sum_{i\in S, j\in S^c}

1357: \E ( \log(1+ P |H_{ji}|^2 )) }{\pi(S)\pi(S^c)}, \\

1358: &f^*_\pi = \Omega\left(\sup_{r\geq r^*, \ \eta\geq

1359: 0}\left[\frac{1}{1+\Delta(r)\Delta(r(1+\eta))}\right]  \times

1360: \left[\min_{S\subset V}\frac{\sum_{i\in S, j\in S^c} {\bf 1}_{(i,j)\in

1361: E_r}\E\log\left( 1 + \frac{P|H_{ji}|^2}{1+nPg(r(1+\eta))}\right ) }{\log

1362: p_\pi \pi(S)\pi(S^c)} \right] \right).

1363: \end{aligned}

1364: \]

1365: \end{theorem}

1366: \vspace{.1in}

1367: 

1368: Theorem \ref{thm:coh_bounds} provides bounds on $f^*$ which relates to the ``cut capacity" of appropriate capacitated graphs. Specifically, \emph{we can compute the information theoretic upper bound (for any PMF) in polynomial time using flow arguments, and by solving an LP as detailed in Sec.~\ref{subsec:info_comp}}.

1369: However, it is not clear how {\em tight} these bounds are. We now quantify the \emph{gap} between

1370: the upper and lower bounds.

1371: 

1372: 

1373: \begin{corollary}

1374: \label{cor:fading}

1375: For any $r\geq r^*$, denote $\delta(r) = \max_{i}\sum_{j: r_{ij} \geq r} Pg(r_{ij})$. Then,

1376: $$

1377: \Omega\left(\frac{\Upsilon }{(1+\Delta^2(r)) (1+\delta(r)) \log p_\pi} \right) \leq f_\pi^* \leq (1+\gamma(r))\Upsilon , $$

1378: where

1379: $$\Upsilon = \min_{S\subset V}\frac{\sum_{i\in S, j\in S^c:  r_{ij} \leq r}

1380: \E \log(1+P |H_{ij}|^2)}{\pi(S)\pi(S^C)}, $$

1381: and

1382: $$ \gamma(r) = \max_{S\subset V: \pi(S), \pi(S^C)>0} \frac{\sum_{i\in S, j\in S^c:  r_{ij} > r} \E \log(1+P |H_{ij}|^2)}

1383: {\sum_{i\in S, j\in S^c:  r_{ij} \leq r} \E \log(1+P |H_{ij}|^2)} .$$

1384: \end{corollary}

1385: 

1386: 

1387: 

1388: \vspace{0.2in}

1389: 

1390: Note that both $\gamma(r)$ and $\delta(r)$ are decreasing functions of $r$, while $\Delta(r)$ is an increasing function

1391: of $r$. Also, since power typically decays as $1/r^a$ for $2\leq a\leq 6$, while  for uniformly distributed networks

1392: $\Delta(r)$ grows only linearly with $r$, the decay of $\gamma(r)$ and $\delta(r)$ is much faster than the growth of $\Delta(r)$.

1393: Hence, for $r$ large enough the \emph{gap} is dominated by the term $\log p_\pi(1+\Delta(r)^2)$.

1394: Specifically, assume that there exists an $\epsilon > 0$ such that the graph $\hGe =

1395: (V, \hEe)$ is connected, where $\hEe = \{(i,j) : \E \lf[\log \lf(1+

1396: P |H_{ij}|^2\rf)\rf] \geq n^{-\epsilon/2}\}$. Then the above bound for UMF reduces to~\cite{ISIT_paper}

1397: $$

1398: \Omega\left(\frac{\mbox{\sf min-cut}_R}{\Delta^2(r_\epsilon) \log n} \right) = f^* = O(\mbox{\sf min-cut}_R),

1399: $$

1400: where $\mbox{\sf min-cut}_R = \min_{S\subset V}\frac{\sum_{i\in S, j\in S^c:  r_{ij} \leq r_\epsilon}

1401: \E \log(1+P |H_{ij}|^2)}{|S||S^C|}$, and $r_\epsilon$ is such that $\delta(r_\epsilon)\leq \frac{1}{n^{1+\epsilon}}$.

1402: 

1403: 

1404: \vspace{0.2in}

1405: 

1406: \begin{proof}[Theorem \ref{thm:coh_bounds}]

1407: We first prove the upper bound.

1408: Following the steps in the proof of Theorem~2.1 in~\cite{JVK} and using $\left(1+\sum_{i=1}^n \alpha_i\right) \leq \prod_{i=1}^n (1+\alpha_i)$ for $\alpha_i>0$, we obtain that for $\lambda \in \Lambda$,

1409: \begin{eqnarray}

1410: \textstyle \sum_{i \in S, j \in S^c} \lambda_{ij} & \leq & \max_{Q_S\succeq 0, (Q_S)_{ii}\leq P} \E[\log \det (I + H_S Q_S

1411: H_S^*)] \nonumber \\

1412: & \leq & \textstyle \sum_{i\in S, j\in S^c} \E \left( \log(1+ P

1413: |H_{ji}|^2) \right). \label{ezx1}

1414: \end{eqnarray}

1415: Now, for any PMF $M = M(f_\pi, \pi)$,  it must be that $\sum_{i\in S, j\in S^C}M_{ij} = f_\pi \pi(S)\pi(S^C)$.

1416: Hence, for any such PMF $M(f,\pi)\in \Lambda$, the upper bound in the Theorem holds.

1417: \vspace{.03in}

1418: 

1419: To establish the lower bound, we construct a

1420: transmission scheme for which the PMF is greater

1421: than or equal to that in the lower bound. For $r\geq r^*$, consider

1422: the graph $G_r = (V,E_r)$ on the $n$ nodes defined above. We use

1423: $\Delta(r(1+\eta))$ to denote the maximum vertex degree of the graph

1424: $G_{r(1+\eta)}$. Now, consider the following transmission scheme. A

1425: node $i$ can transmit to a node $j$ only if $r_{ij}\leq r$. Also,

1426: when a node $i$ transmits, no node within a distance $r(1+\eta)$ of

1427: the receiver can transmit. Thus, when a link $(i,j)\in E_r$ is

1428: active, at most $\Delta(r(1+\eta))$ nodes are constrained to remain

1429: silent, i.e., at most $\alpha=\Delta(r(1+\eta))\Delta(r)$ links are

1430: constrained to remain inactive. Hence, the chromatic number of the

1431: dual graph is at most $(1+\Delta(r(1+\eta))\Delta(r)$. In addition,

1432: we assume that the signal transmitted by each node has a Gaussian

1433: distribution.

1434: %For the lower bound, we will use the

1435: %time-division scheme in the proof of Theorem~\ref{thm:awgn} which

1436: %schedules every link in the network for a fraction $\alpha

1437: %=\frac{1}{1+\Delta(r)\Delta(r(1+\eta))}$ of the time.

1438: For any given

1439: link that transmits data at a particular time, we treat all other

1440: simultaneous transmissions in the network as interference. Now

1441: focus on any one link, say link $(1,2)$ between node $1$ and $2$,

1442: without loss of generality. We claim the following.

1443: \begin{lemma}\label{lemma:cx1}

1444: For the above scheme, the following rate on link $(1,2)$ is achievable:

1445: \[ \textstyle \lambda_{12} =\  \alpha^{-1} \E \log \left( 1 + \frac{P |H_{21}|^2}{1 + n Pg(r(1+\eta))} \right).\]

1446: \end{lemma}

1447: We prove Lemma \ref{lemma:cx1} later. First we explain how it implies the proof

1448: of Theorem \ref{thm:coh_bounds}.  A similar analysis holds for

1449: other links that $(1,2)$ in $E_r$. Thus, for graph $G_r$

1450: the following rate are achievable on link $(i,j)\in E_r$:

1451: $$ \alpha^{-1} \E \log \left( 1 + \frac{P |H_{ji}|^2}{1 + n Pg(r(1+\eta))} \right),$$

1452: Now given the capacitated graph $G_r$, we can

1453: use classical wireline network based routing algorithms for obtaining

1454: a product multicommodity flow that is lower bounded by the following

1455: quantity:

1456: $$ f_{LB}(r, \eta) = \Omega\left(\left[\frac{1}{1+\Delta(r)\Delta(r(1+\eta))}\right]  \times

1457: \left[\min_{S\subset V}\frac{\sum_{i\in S, j\in S^c} {\bf 1}_{(i,j)\in

1458: E_r}\E\log\left( 1 + \frac{P|H_{ji}|^2}{1+nPg(r(1+\eta))}\right ) }{\log

1459: p_\pi \pi(S)\pi(S^c)} \right] \right).$$

1460: This implies the following lower bound on $f^*_\pi$:

1461: $$ f^*_\pi \geq \sup_{r \geq r^*, \eta \geq 0} f_{LB}(r, \eta).$$

1462: This is precisely the claimed lower bound in the statement of Theorem \ref{thm:coh_bounds}

1463: and thus completing the proof.

1464: \end{proof}

1465: 

1466: \begin{proof}[Lemma \ref{lemma:cx1}]

1467: We will use the following result, that follows directly from Theorem~1 in~\cite{kashyap}.

1468: \begin{theorem}

1469: \label{thm:jamming} Consider a complex scalar channel where the

1470: output $Y$ when $X$ is transmitted is given by

1471: \[ \textstyle Y = hX + Z + S,\]

1472: where $Z$ is a complex circularly symmetric Gaussian random variable

1473: with unit variance, and $S$ satisfies $\E[S^*S]\leq \hat{P} $. Also,

1474: $h$ is zero mean and i.i.d over channel uses. If $X$ is a complex

1475: zero mean circularly symmetric Gaussian random variable with

1476: $\E[X^*X]= P $, then $ \textstyle I(X;(Y,h)) \geq  \E \log\left ( 1

1477: + \frac{P|h|^2}{1 + \hat{P}} \right )$.

1478: \end{theorem}

1479: 

1480: \vspace{0.1in}

1481: 

1482: We consider a transmission scheme where the signal transmitted over

1483: each link, when active, is a complex zero mean white circularly

1484: symmetric Gaussian with variance $P$.

1485:  Moreover, we assume that the transmissions on

1486: all links are mutually independent. Let $t_1,t_2,\hdots$ denote

1487: times at which link $(1,2)$ is scheduled. Hence, at any such time

1488: $t\in\{t_1,t_2,\hdots\}$, the received signal at node 2 is given by

1489: \[ \textstyle Y_2(t) = H_{21}(t)X_1(t) + \sum_{k\neq 1,2} H_{2k}(t) X_k(t) + Z_2(t).\]

1490: Using the mutual independence of transmissions and zero mean

1491: property along with the construction of the scheduling scheme,

1492: \[ \textstyle \E \left| \sum_{k\neq 1,2} H_{2k}(t) X_k(t) + Z_2(t)  \right |^2 \leq 1 + nPg(r(1+\eta)).\]

1493: From Theorem~\ref{thm:jamming},

1494: \begin{equation} \label{zza}

1495: \textstyle I( X_1(t); (Y_2(t), H_{21}(t))) \geq \E \log \left ( 1 + \frac{P|H_{21}|^2}{(1 +n Pg(r(1+\eta) )} \right).

1496: \end{equation}

1497: Since the channel is assumed to be i.i.d. over channel uses, a random

1498: coding argument can be used to achieve this rate with a probability

1499: of error that goes to zero as the block length goes to infinity.

1500: 

1501: Combining this with the time-sharing between different sets of links

1502: described above, since each link gets to transmit at least once in $\alpha$

1503: times slots, or at least $1/\alpha$ fraction of the time, it follows that

1504: \[ \lambda_{12}  \geq \alpha^{-1} \E \log \left( 1 + \frac{P |H_{21}|^2}{1 + n

1505: Pg(r(1+\eta))} \right).\]

1506: \end{proof}

1507: 

1508: 

1509: 

1510: \begin{proof}[Corollary \ref{cor:fading}]

1511: Consider any $S$ such

1512: that $\pi(S), \pi(S^C)>0$. Then,

1513: \begin{eqnarray}

1514: & &  \textstyle \mbox{\sf Cut}(S,S^c) = \sum_{i\in S, j\in S^c} \E \left( \log(1+ P

1515: |H_{ji}|^2) \right). ~~~~~~ \nonumber \\

1516: & & \textstyle  ~\leq (1+\gamma(r)) \sum_{i\in S, j\in S^c: r_{ij} \leq r} \E \left( \log(1+ P

1517: |H_{ji}|^2) \right), ~~~~~\label{eq:cut_bound_rxcsi}

1518: \end{eqnarray}

1519: where the second line follows from the concavity of the $\log$ function, Jensen's inequality,  $\log(1+x)\leq x$

1520: for $x>0$ and definition of $\gamma(r)$.

1521: Thus,

1522: \begin{equation}

1523: \frac{\textstyle \mbox{\sf Cut}(S,S^c)}{\pi(S)\pi(S^C)}\leq \Upsilon (1+\gamma(r)).

1524: \end{equation}

1525: The upper

1526: bound then follows from the upper bound in Theorem~\ref{thm:coh_bounds}.

1527: 

1528: Next, we consider the transmission scheme that led to the lower bound

1529: in~(\ref{zza}) with $\eta = 0$.

1530: Note that in (\ref{zza}), we used the term $n Pg(r(1+\eta))$ as a bound on the interference power. However, here we consider

1531:  the actual interference $I_{ij} = \sum_{k \in V: r_{jk} \geq r}  Pg(r_{jk})$ for a transmission from $i$ to $j$. Note that

1532:  $I_{ij}\leq \delta(r)$.

1533: Now, by Lemma \ref{lem:ineq2}, we have

1534: \begin{equation} \label{x2aa}

1535: \textstyle \E \lf[\log\lf( 1+

1536: \frac{P|H_{ji}|^2}{1 + I }\rf)\rf] \geq

1537: \frac{1}{1+\delta(r)}\textstyle \E\lf(\log(1+ {P|H_{ji}|^2})

1538: \rf).

1539: \end{equation}

1540: Using (\ref{eq:cut_bound_rxcsi}) and (\ref{x2aa}) along with the

1541: lower bound obtained via time-division scheme that led to (\ref{zza}),

1542: the lower bound in Theorem~\ref{thm:coh_bounds} gives us

1543: \begin{eqnarray}

1544: f^* & = & \Omega\lf( \min_{S\subset V} \frac{\sum_{i\in S, j\in S^c: r_{ij} \leq r} \E \left( \log(1+ P

1545: |H_{ji}|^2) \right)}{\pi(S)\pi(S^C)(1+ \Delta(r)^2) (1+\delta(r))\log p_\pi} \rf) \nonumber \\

1546: & = & \Omega\lf( \frac{\Upsilon}{(1+\Delta(r)^2) (1+\delta(r))\log p_\pi}  \rf). \label{fx0}

1547: \end{eqnarray}

1548: \end{proof}

1549: 

1550: \vspace{.2in}

1551: \subsubsection{Deterministic AWGN Channels}

1552: We now consider an AWGN channel without fading, i.e., we have $\hat{H}_{kj}=1$ w.p.

1553: $1$, $\forall k,j=1,\hdots,n$.

1554: We first obtain the following set of bounds on maximum PMF using standard arguments.

1555: \begin{theorem}

1556: \label{thm:awgn} The

1557: maximum PMF $f^*_\pi$ is bounded as

1558: follows.

1559: %\footnote{For the lower bound, we use the $\Omega$ notation in

1560: %order not to have to write out constants explicitly; we note that the

1561: %constants are independent of the graph structure and $n$~\cite{LR88}.}.

1562: \[

1563: f^*_\pi \leq \min_{S\subset V} \frac{2\sum_{i\in S, j\in S^c} \log(1+

1564: \sqrt{P g(r_{ij}}))}{\pi(S)\pi(S^c)},

1565: \]

1566: \[\begin{aligned} &f^*_\pi = \Omega\left(\sup_{r\geq r^*, \ \eta\geq 0}\left[\frac{1}{1+\Delta(r)\Delta(r(1+\eta))}\right]

1567: \times \left[\min_{S\subset V}\frac{\sum_{i\in S, j\in S^c:  r_{ij} \leq r}

1568: \log\left( 1 + \frac{Pg(r_{ij})

1569: %\Delta(r(1+\eta))

1570: }{1+nPg(r(1+\eta))}\right) }{\log p_\pi \pi(S)\pi(S^c)} \right]\right).

1571: \end{aligned}\]

1572: \end{theorem}

1573: 

1574: \vspace{0.2in}

1575: 

1576: 

1577: 

1578: 

1579: 

1580: Next, we present a Corollary of Theorem \ref{thm:awgn} which characterizes the tightness of the above bound for UMF for

1581: low signal to noise ratio (SNR).

1582: 

1583: \begin{corollary}\label{cor:lbub}

1584: Define $I(r) = \min\{ I > 0: \sum_{j: r_{ij} \geq r} Pg(r_{ij}) \leq I, ~\mbox{for all $i$}\}$;

1585: $r(\delta) = \min\{ r > 0: I(r) \leq \delta\}$ for $\delta > 0$. Then,

1586: $$

1587: \Omega\left(\frac{n}{(1+\delta)\Delta(r(\delta))(1+\Delta(r(\delta))^2)\log n }  ~{\Upsilon}^2   \right) \leq f^* \leq  2\Upsilon + O\left(\frac{\delta}{n}\right), $$

1588: where

1589: $$\Upsilon = \min_{U\subset V} \frac{\sum_{i\in U, j\in U^c:  r_{ij} \leq r(\delta)} \log(1+ \sqrt{P g(r_{ij}}))}{|U||U^C|}.$$

1590: \end{corollary}

1591: 

1592: \vspace{.1in}

1593: 

1594: 

1595: We now present the proofs of Theorem~\ref{thm:awgn} and Corollary~\ref{cor:lbub}.

1596: The main idea in the proof of Theorem~\ref{thm:awgn}

1597: is to neglect interference to upper bound achievable rates on links, and

1598: to construct a transmission scheme to induce achievable rates on the links. In particular the scheme

1599: that we construct consists of time sharing between multiple transmission schemes, each of which

1600: enables direct transmissions between nodes that are separated by at most distance $r$. Then the lower

1601: bound on $f^*$ is obtained by routing over graph $G_r$, where each edge has a capacity given by this

1602: time division scheme.

1603: 

1604: \begin{proof}[Theorem \ref{thm:awgn}]

1605: We first prove the upper bound. In order to bound the sum-rate across

1606: each given cut, we refer to the proof of the max-flow min-cut lemma

1607: in \cite{XK}, which yields for any $S \subset V$ and $\lambda \in \Lambda$,

1608: $$

1609: \textstyle \sum_{i \in S, j \in S^c} \lambda_{ij} \leq \sum_{j \in S^c} \log (1 + \E(|\tilde{X}_j|^2)),

1610: $$

1611: where $\tilde{X}_j = \sum_{i \in S} \sqrt{g(r_{ji})} \, X_i$.

1612:  We therefore deduce that

1613: \begin{eqnarray*}

1614: \lefteqn{\textstyle \sum_{i \in S, j \in S^c} \lambda_{ij}}\\

1615: &\leq & \textstyle \sum_{j \in S^c} \log

1616: [1 + \sum_{i,k \in S} \sqrt{g(r_{ji}) \, g(r_{jk})} \,

1617: |\E(X_i \overline{X_k})| ]\\

1618: & \leq & \textstyle \sum_{j \in S^c} \log [ 1 + P (\sum_{i \in S}

1619: \sqrt{g(r_{ji})})^2],

1620: \end{eqnarray*}

1621: since $|\E(X_i \overline{X_k})| \leq \sqrt{P_i P_k} \leq P$. Finally,

1622: we obtain

1623: \[\begin{aligned}

1624: \textstyle \sum_{i \in S, j \in S^c} \lambda_{ij} &\textstyle \leq \sum_{j \in S^c} 2 \log (1 + \sqrt{P} \sum_{i \in S} \sqrt{g(r_{ji})})\\

1625: & \textstyle \leq \sum_{i \in S, j \in S^c} 2 \log (1 + \sqrt{P g(r_{ji})}).

1626: \end{aligned}\]

1627: Now, for any PMF $M = M(f_\pi, \pi)$,  it must be that $\sum_{i\in S, j\in S^C}M_{ij} = f_\pi \pi(S)\pi(S^C)$.

1628: Hence, for any such PMF $M(f,\pi)\in \Lambda$, the upper bound in the Theorem holds.

1629: 

1630: To establish the lower bound, we construct a

1631: transmission scheme for which the PMF is greater

1632: than or equal to that in the lower bound. For $r\geq r^*$, consider

1633: the graph $G_r = (V,E_r)$ on the $n$ nodes defined above. We use

1634: $\Delta(r(1+\eta))$ to denote the maximum vertex degree of the graph

1635: $G_{r(1+\eta)}$. Now, consider the following transmission scheme. A

1636: node $i$ can transmit to a node $j$ only if $r_{ij}\leq r$. Also,

1637: when a node $i$ transmits, no node within a distance $r(1+\eta)$ of

1638: the receiver can transmit. Thus, when a link $(i,j)\in E_r$ is

1639: active, at most $\Delta(r(1+\eta))$ nodes are constrained to remain

1640: silent, i.e., at most $\Delta(r(1+\eta))\Delta(r)$ links are

1641: constrained to remain inactive. Hence, the chromatic number of the

1642: dual graph is at most $(1+\Delta(r(1+\eta))\Delta(r)$. In addition,

1643: we assume that the signal transmitted by each node has a Gaussian

1644: distribution. Then, subject to the maximum average power constraint,

1645: for any node pair $i, j$, such that $r_{ij} \leq r$,

1646: the following rate is achievable from $i \to j$:

1647: \begin{equation}

1648: \label{cap_lb} \lambda_{ij} \geq  \frac{\log\left( 1 +

1649: \frac{Pg(r_{ij})}{ 1 + nPg(r(1+\eta))}\right )}{1

1650: + \Delta(r)\Delta(r(1+\eta))}.

1651: \end{equation}

1652: Note that the interference is due to at most $n$ nodes and all the interfering nodes are at least a distance $r(1+\eta)$ away from the receiver. We now consider routing over the graph $G_r$, where each edge $(i,j)$ has capacity $\lambda_{ij}$. The lower bound

1653: then follows from the lower bound in Theorem~\ref{thm:PMF_LR}.

1654: \end{proof}

1655: 

1656: 

1657: 

1658: 

1659: 

1660: \vspace{0.2in}

1661: 

1662: \begin{proof}[Corollary \ref{cor:lbub}]

1663: Consider any cut defined by $(S,S^c)$. Due to the symmetry of the upper bound in Theorem~\ref{thm:awgn}, without loss of generality, assume $|S| \leq n/2$.

1664: Consider any $\delta$ such that $r(\delta)\geq r^*$. Then,

1665: \begin{eqnarray}

1666: & &  \textstyle \mbox{\sf Cut}(S,S^c) =  \sum_{i\in S, j\in S^c} \log (1+\sqrt{Pg(r_{ij})})~~~~~~ \nonumber \\

1667: & & \textstyle  ~=  \sum_{i\in S, j\in S^c :  r_{ij} \leq r(\delta)} \log \lf(1+\sqrt{Pg(r_{ij})}\rf) +

1668: \sum_{i\in S, j\in S^c :  r_{ij} > r(\delta)} \log \lf(1+\sqrt{Pg(r_{ij})}\rf)\nonumber \\

1669: & & \textstyle ~\leq  \sum_{i\in S, j\in S^c :  r_{ij} \leq r(\delta)} \log \lf(1+\sqrt{Pg(r_{ij})}\rf) + |S|\delta,~~~~~\label{eq:leq3}

1670: \end{eqnarray}

1671: where the last step follows from the definition of $r(\delta)$, and $\log(1+\sqrt{x})\leq x$ for

1672: $x\leq 1$.

1673: Hence, the upper bound in the Corollary follows from the upper bound in Theorem~\ref{thm:awgn}.

1674: Since we assume $Pg(r_{ij})\leq 1$ for all $i$ and $j$, from Lemma~\ref{lem:ineq1}, we have

1675: {\small

1676: \begin{eqnarray}

1677: & & \sum_{i\in S, j\in S^c :  r_{ij} \leq r(\delta)} \log \left[ 1+\sqrt{Pg(r_{ij})} \right ]

1678:   ~~\leq  \sqrt{2 \Delta(r(\delta)) |S| \sum_{i\in S, j\in S^c :  r_{ij} \leq r(\delta)} \log \lf(1+Pg(r_{ij})\rf)}. \label{eq:leq4}

1679: \end{eqnarray}}

1680: 

1681: 

1682: For the lower bound, consider the choice of $r = r(\delta)$ and $\eta = 0$ for the

1683: scheme described in the proof of Theorem~\ref{thm:awgn}. Then, the interference during data transmission

1684: from $i$ to $j$, $I_{ij} = \sum_{k \in V: r_{jk} \geq r(\delta)}  Pg(r_{jk})\leq \delta$.

1685: Now, Lemma~\ref{lem:ineq2}, implies that

1686: \begin{equation}

1687: \label{eqn:xyz}

1688: \log \left(1 + \frac{Pg(r_{ij})}{1+I_{ij}}\right)

1689: \geq \frac{1}{1+\delta}\log( 1 +  Pg(r_{ij})).

1690: \end{equation}

1691: 

1692: Using an appropriately modified lower bound in

1693: Theorem \ref{thm:awgn} for the choice of $r = r(\delta)$, $\eta=0$, it follows that

1694: {\small \begin{eqnarray}

1695: f^* & = &  \Omega\left[ \min_{S\subset V} \frac{\sum_{i\in S,

1696: j\in S^c :  r_{ij} \leq r(\delta)} \log (1+Pg(r_{ij}))}

1697: {(1+\delta)(1+\Delta(r(\delta))^2) \log n|S||S^c|}\right] \nonumber \\

1698: & = & \Omega\left[

1699: \frac{\Upsilon^2 n}{(1+\delta)\Delta(r(\delta))(1+\Delta(r(\delta))^2)\log n }\right], \label{x2}

1700: \end{eqnarray}}

1701: where the second step follows from~(\ref{eq:leq4}). The lower bound

1702: in Theorem~\ref{thm:awgn} then implies the lower bound in the Corollary.

1703: This completes the proof.

1704: \end{proof}

1705: 

1706: %

1707: 

1708: 

1709: \vspace{0.3in}

1710: 

1711: \subsubsection{Random Fading with CSI at both Tx and Rx}

1712: We now obtain bounds on the PMF for a Gaussian channel with random fading when CSI is available at

1713: both the transmitter and the receiver.

1714: Qualitatively, these bounds are very similar to the case of deterministic

1715: AWGN channels. The main result is as follows.

1716: \vspace{.1in}

1717: \begin{theorem}

1718: \label{thm:coh_bounds1}

1719: With CSI at both transmitters and receivers, $f^*_\pi$ is bounded as follows.

1720: \[

1721: f^*_\pi \leq \min_{S\subset V}\frac{\sum_{i\in S, j\in S^c} 2

1722: \, \E ( \log(1+ \sqrt{P} \, |H_{ji}|)}{\pi(S)\pi(S^c)}.

1723: \]

1724: The lower bound for the receiver only CSI case is a (weak) lower

1725: bound for this case as well.

1726: \end{theorem}

1727: \vspace{.2in}

1728: \begin{proof}

1729: The upper bound follows again from the proof of Theorem 2.1 in \cite{JVK}, from which we deduce that for

1730: any $\lambda \in \Lambda$,

1731: \begin{eqnarray*}

1732: \lefteqn{\textstyle \hspace{-0.2in}\sum_{i \in S, j \in S^c} \lambda_{ij} \leq \E [\max_{Q\succeq 0, Q_{ii}\leq P} \log \det (I + H_S Q_S H_S^*)]}\\

1733: & \leq & \textstyle \sum_{j \in S^c} \E [\max_{Q\succeq 0, Q_{ii}\leq P} \log (1 + h_j Q_S h_j^*)],

1734: \end{eqnarray*}

1735: where $h_j$ is the $j^{th}$ row of $H$. Since $h_j Q_S h_j^*$

1736: is maximum when $(Q_S)_{ik} \equiv P$ for all $i,k \in S$,

1737: we obtain, following the steps of the proof of Theorem~\ref{thm:awgn},

1738: \begin{eqnarray*}

1739: \textstyle \sum_{i \in S, j \in S^c} \lambda_{ij}

1740: & \leq & \textstyle \sum_{j \in S^c} \E \log (1 + P (

1741: \sum_{i \in S} |H_{ij}|)^2 ]\\

1742: & \leq & \textstyle \sum_{j \in S^c} 2 \, \E \log (1 + \sqrt{P}

1743: \sum_{i \in S} |H_{ij}| )]\\

1744: & \leq & \textstyle \sum_{i  \in S, j \in S^c}

1745: 2 \, \E (\log (1 + \sqrt{P} \, |H_{ij}|),

1746: \end{eqnarray*}

1747: so the upper bound on $f^*_\pi$ follows from Theorem \ref{thm:PMF_LR}.

1748: \end{proof}

1749: 

1750: 

1751: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1752: \subsection{Computational Methods}

1753: \label{subsec:info_comp}

1754: We discuss the implications of the bounds

1755: for the case of CSI availability at the receiver only as stated in Corollary \ref{cor:fading}.

1756: Similar implications follow for the case where CSI is available to both transmitters and

1757: receivers as well.

1758: 

1759: Corollary~\ref{cor:fading} shows that an upper bound

1760: on $f^*_\pi$ can be obtained via the maximum PMF on graph $K_n$,

1761: where each edge $(i,j)$ has a capacity  $\log(1+P |H_{ij}|^2)$, and there is no interference;

1762: specifically, $\log n$ times the PMF thus computed for $K_n$ is an upper bound on $f^*$.

1763: The lower bound is obtained via routing on $G_{r}$ with edge $(i,j)$

1764: having capacity $\frac{\log(1+P |H_{ij}|^2)}{(1+\Delta^2(r))(1+\delta)}$. Hence, the PMF

1765: on $G_{r(\delta)}$ is a lower bound on $f_\pi^*$.

1766: Both the above computations can be done by solving an LP in polynomial time. Moreover, the ratio

1767: of the bounds is quantified in~\ref{cor:fading}. We note that

1768: such an approximation ratio could be obtained easily for the

1769: combinatorial interference model using node coloring arguments.

1770: The arguments  here are more complicated, as detailed

1771: in the proof of  Corollary~\ref{cor:fading}.

1772: 

1773: 

1774: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1775: 

1776: 

1777: 

1778: 

1779: \subsection{Application}

1780: 

1781: We now apply the information

1782: theoretic characterization of PMF in the previous subsection to

1783: obtain a scaling law for average UMF in a geometric random network

1784: with a fading channel, and when CSI is available at the receivers. The scaling law

1785: we obtain is along similar lines to those that exist in the literature.

1786: Similar bounds can be obtained when CSI is available both

1787: at the transmitter and the receiver or when the channels are AWGN channels.

1788: 

1789: We consider a geometric random graph model

1790: with a constant node density:~$n$ nodes are placed uniformly

1791: at random in a torus of area $n$ (and not unit area). Thus the

1792: distance between two nodes is a random variable taking values

1793: in $(0,\Theta(\sqrt{n}))$. We assume that all nodes have the

1794: same transmission power equal to $1$, i.e. $P_i = 1$ for all

1795: $1\leq i\leq n$. We state the following result characterizing

1796:  $f^*$.

1797: \begin{lemma}

1798: \label{lem:coh_expl}

1799: Consider the Gaussian channel model with random fading and CSI available only

1800: at the receivers. Let $g(r) = (1+r)^{-\alpha}$, $\alpha > 3$ and $P = 1$. Then

1801: for a geometric random graph with constant node density (described above), the average (over

1802: random position of nodes) $f^*$ is bounded as

1803: $$ \Omega\left(\frac{1}{n^{3/2} \log^{1+\alpha} n}\right) \leq \E[f^*] \leq O\left(\frac{1}{n^{3/2}}\right),$$

1804: if $\Pr(|\hat{H}_{ij}|^2 \geq \beta) \geq \gamma$ for some strictily positive constants

1805: $\beta, \gamma$ (independent of $n$) for all $1\leq i, j \leq n$. (Note that the condition is for

1806: the normalized channel gains $\hat{H}_{ij}$s and not the actual gains $H_{ij}$s.)

1807: \end{lemma}

1808: \begin{proof}

1809: We use Theorem \ref{thm:coh_bounds} to evaluate the bounds. First we

1810: obtain an upper bound by evaluating the bound of Theorem \ref{thm:coh_bounds}

1811: for a specific cut $(U, U^c)$. Then, we  evaluate lower bound by

1812: relating it to an appropriate grid-graph as in Lemma \ref{lem:gnrcomb}.

1813: 

1814: 

1815: Now, we consider the upper bound. Consider a horizontal line dividing

1816: the square of area $n$ into equal halves. Let $U$ be set of nodes

1817: that lie in bottom half, and so $U^c$ is the set of nodes that lie in the

1818: top half. From Theorem \ref{thm:coh_bounds}, we have

1819: \begin{eqnarray}

1820: f^*|U||U^c| & \leq & \sum_{i \in U, j \in U^c} \E\log (1+P|H_{ij}|^2) \nonumber \\

1821:   & \leq & \sum_{i \in U, j \in U^c} \log (1+P\E[|H_{ij}|^2]) \nonumber \\

1822:   & =  & \sum_{i \in U, j \in U^c} \log (1+Pg(r_{ij})) \nonumber \\

1823:   & \leq & \sum_{i \in U, j\in U^c} Pg(r_{ij}) \nonumber \\

1824:   & = & \sum_{i\neq j} (1+r_{ij})^{-\alpha} \bone_{\{i \in U\}}\bone_{\{j\in U^c\}}, \label{eq:xx1}

1825: \end{eqnarray}

1826: where we have used Jensen's inequality, $\log (1+x) \leq x$ for all $x \geq 0$,

1827: and the hypothesis of Lemma. Since, the nodes are thrown uniformly at random,

1828: the expectation of each term in (\ref{eq:xx1}) for a pair $(i,j)$ is the same.

1829: Using linearity of expectation, we obtain that

1830: \begin{eqnarray}

1831: \E[\sum_{i\neq j}(1+ r_{ij})^{-\alpha} \bone_{\{i \in U\}}\bone_{\{j\in U^c\}}]

1832: & = & n(n-1) \E[(1+r_{12})^{-\alpha}\bone_{\{1 \in U\}}\bone_{\{2\in U^c\}} ] \nonumber \\

1833: & \leq & O\left(n^2 \int_{1}^{\sqrt{n}}  \int_{r}^{\sqrt{n}} s^{-\alpha} \frac{s ds}{n}

1834: \frac{\sqrt{n} dr}{n}\right) \nonumber \\

1835: & = & O\left(\sqrt{n} \int_1^{\sqrt{n}} \int_{r}^{\sqrt{n}} s^{-\alpha+1} ds dr \right) \nonumber \\

1836: & = & O\left(\sqrt{n}\int_1^{\sqrt{n}} r^{-\alpha+2} dr\right) \nonumber \\

1837: & = & O(\sqrt{n}),\label{lx1}

1838: \end{eqnarray}

1839: where we used the fact that for $\alpha > 3$ the last integral is bounded

1840: above by a constant. The above evaluation can be justified as follows. First

1841: note that $\Pr(1 \in U, 2 \in U^c) = 1/4$. Given $\{1\in U, 2 \in U^c\}$,

1842: node $1$ in the bottom rectangle and node $2$ in the top rectangle

1843: are uniformly distributed. Now, consider a thin horizontal strip of

1844: width $dr$ and length $\sqrt{n}$ at distance $r$ below the

1845: horizontal line dividing the square (and inducing $U, U^c$). The node

1846: $1 \in U$ belongs to this strip with probability $2\sqrt{n} dr/n$. Now,

1847: the node $2$ is at distance at least $r$ from node $1$. Consider a ring

1848: of width $ds$, centered at node $1$'s position and of radius $s \geq r$.

1849: The area of this ring is $2\pi s ds$. The probability of node $2$ being

1850: in this ring is bounded above by $4\pi s ds/n$. When the above described

1851: condition is true, the nodes $1$ and $2$ are at distance $s$. Integrating

1852: over the appropriate ranges justifies the final outcome (\ref{lx1}).

1853: 

1854: 

1855: 

1856: 

1857: 

1858: Now, it is easy to see that under

1859: any configuration of nodes, $f^* = O(n^2)$ since $g(r) \leq 1$ for any $r \geq 0$,

1860: $P = 1$ and elementary arguments. Let

1861: event $A = \{ |U||U^c| = \Theta(n^2)\}$. Using Chernoff's bound, it is

1862: easy to see that (with appropriate selection of constants in definition of $A$)

1863: for large enough $n$, we have

1864: $$ \Pr(A) = 1-1/n^6. $$

1865: Using this estimate and bound $f^* = O(n^2)$ we obtain that

1866: \begin{eqnarray}

1867: \E[f^*] & = & \E[f^* \bone_A] + \E[f^* \bone_{A^c}] \nonumber \\

1868:         & \leq & \E[f^* \bone_A] + O\left(\frac{1}{n^4}\right) \nonumber \\

1869:         & = & \Theta\left(\frac{\E[f^* |U||U^c|\bone_{A}]}{n^2}\right) + O\left(\frac{1}{n^4}\right) \nonumber \\

1870:         & \leq & \Theta\left(\frac{\E[f^* |U||U^c|]}{n^2}\right) + O\left(\frac{1}{n^4}\right) \nonumber \\

1871:         & = & O\left(\frac{1}{n^{3/2}}\right).

1872: \end{eqnarray}

1873: 

1874: Next, we prove the lower bound. For this we construct a graph

1875: with achievable link capacities for which the average $f^*$ is

1876: lower bounded as claimed in the Lemma.

1877: Consider $r = \Theta(\log n)$. Then

1878: the corresponding $G_r$, which is the geometric random graph $G(n,r)$, is

1879: connected with high probability (at least $1-1/n^4$ by

1880: appropriate choice of constants in selection of $r$).

1881: For this choice of $r$, using the Chernoff and Union bounds it follows

1882: that with probability at least $1-1/n^4$,

1883: $$ \Delta(2r) = \Theta\left(\log^{2} n\right).$$

1884: Again, we can identify

1885: a grid graph structure as a subgraph structure of

1886: $G_r$ based on the argument used in Lemma \ref{lem:gnrcomb}.

1887: Denote the edges of this grid sub-graph structure as $\hat{E}$.

1888: We note that $\Theta(1)$ edges are incident on each

1889: of the $n$ nodes that belong to $\hat{E}$ (which is a property of

1890: the grid-graph structure). Next, we design a feasible transmission scheme

1891: for which each edge in $\hat{E}$ can support a transmission rate of

1892: $\Omega(\log^{-\alpha} n)$.

1893: 

1894: Specifically, we consider a TDMA schedule for the

1895: graph $G_r$ similar to that described in the proof of the lower bound

1896: for Theorem \ref{thm:coh_bounds}.

1897: It is easy to

1898: see that $G_{2r}$ can be vertex colored using

1899: $\Theta(\Delta(2r))$ colors.

1900: We use a randomized

1901: scheme to do  TDMA scheduling

1902: as follows: in each time-slot, each node becomes

1903: tentatively active with probability $1/\Delta(2r)$ and

1904: remains inactive otherwise. If a node becomes tentatively active

1905: and none of its neighbors in $G_{2r}$ is tentatively-active, then

1906: it will become active. Else, it becomes inactive. All

1907: active nodes transmit in the time-slot simultaneously.

1908: It is easy to see that each node transmits for

1909: $\Theta(1/\Delta(2r))$ fraction of the time on an average. The

1910: randomization here is used to facilitate the  computation

1911: of a simple bound on the average interference experienced

1912: by a node due to transmissions by  nodes that are

1913: not its neighbor.

1914: 

1915: Now under the above vertex coloring, each node

1916: gets to transmit once in $\Theta(\Delta(2r))$ time-slots on average

1917: at power $\Theta(P \Delta(2r)) = \Theta(\Delta(2r))$.

1918: We wish to concentrate on transmissions for

1919: edges that belong to $\hat{E}$, which are a subset of

1920: edges of $G_r$. For any such transmission,

1921: say from $u \to v$ with $(u,v) \in \hat{E}$, according

1922: to the above coloring of $G_{2r}$ no other node within distance $r$ of $v$ is transmits

1923: simultaneously. Also, any node

1924: that is at least distance $r$ away from $v$ can be active

1925: with probability at most $1/\Delta(2r)$. Hence, the average power

1926: corresponding to the interference received

1927: by node $v$, say $I_v$, can be bounded above as follows:

1928: \begin{eqnarray}

1929:  I_v & = & O\left(\sum_{j: r_{vj} > r} \frac{P\Delta(2r)\E[|H_{ij}|^2]}{\Delta(2r)}

1930:  \right) = O\left(\sum_{j: r_{vj} > r} g(r_{vj})\right). \label{eq:xx4}

1931: \end{eqnarray}

1932: where we used the fact that each node transmits at

1933: power $P\Delta(2r) = \Delta(2r)$ for $1/\Delta(2r)$ fraction

1934: of the time and $\E[|H_{ij}|^2] = g(r_{ij})$.

1935: By another application of Chernoff's bound and union bound, it

1936: can be shown that the number of nodes in an annulus around node $v$ with unit

1937: width and radius $R$ for $R \in \N, r \leq R \leq \sqrt{n}$, is

1938: $\Theta(R)$ with probability at least $1-1/n^5$. Then it follows that

1939: $$ \sum_{j: r_{vj} > r} g(r_{vj}) = O\left(\sum_{R=\lfloor r \rfloor}^{\sqrt{n}} g(R) R\right) = O\left(r^{-\alpha+2}\right),$$

1940: where we have used fact that $\alpha > 3$. Using this in

1941: (\ref{eq:xx4}), we obtain that with probability at least $1-1/n^{5/2}$ we

1942: have

1943: \begin{eqnarray}

1944:  I_v & = & O\left(r^{-\alpha+2}\right) = O\left(\log^{2-\alpha} n\right).

1945:  \end{eqnarray}

1946: That is, $I_v \to 0$ as $n\to\infty$ for $\alpha > 3$. Thus, by selection

1947: of large enough $n$, $I_v$ can be made as small as possible.

1948: That is, when transmission from $u\to v$ happens, the average noise

1949: received by node $v$ due to other simultaneous transmission is very

1950: small, say less than $\delta$ for some small enough $\delta >0$.

1951: 

1952: Given this, the arguments used in Lemma \ref{lemma:cx1} imply that

1953: when $u$ transmits to $v$ at power $P\Delta(2r)$ once

1954: in $\Theta(\Delta(2r))$ time-slot, considering other transmissions

1955: as noise, we obtain that the effective rate between $u \to v$ is

1956: lower bounded as

1957: $$ \lambda_{u\to v} = \Omega\left(\Delta(2r)^{-1} \E\left[\log \left( 1+ \frac{P \Delta(2r) g(r_{uv}) |\hat{H}_{uv}|^2}{1+I_v}\right ) \right]\right).$$

1958: Now, $\hat{H}_{uv}$ is independent of everything else and

1959: $$\Pr(|\hat{H}_{uv}|^2 \geq \beta) \geq \gamma,  $$

1960: for some positive constants $\beta, \gamma$ as per our hypothesis.

1961: Therefore, use of Lemma \ref{lem:ineq2} implies

1962: $$ \lambda_{u\to v} \geq \Omega\left(\Delta(2r)^{-1} \log \left (1+ \frac{P \Delta(2r) g(r_{uv})}{1+I_v}\right) \right).$$

1963: Further, $I_v \leq \delta$ for small enough $\delta$. Therefore,

1964: another use of Lemma \ref{lem:ineq2} implies that

1965:  \[\lambda_{u\to v} \geq \Omega\left(\Delta(2r)^{-1} \log \left( 1+ P \Delta(2r) g(r_{uv})\right) \right).\]

1966: Now

1967: $$ P g(r_{uv})\Delta(2r) = \Omega\left( \log^{2-\alpha} n\right),$$

1968: where we have used the fact that $r_{uv} \leq r$ and $g(\cdot)$ is

1969: monotonically decreasing.  For $x \in (0,0.5)$, $\log (1+x) \geq x/2$.

1970: Therefore, for $\alpha > 3$

1971: $$ \lambda_{u\to v} = \Omega\left(\Delta(2r)^{-1} \log^{2-\alpha} n\right).$$

1972: Since $\Delta(2r) = \Theta(\log^2 n)$,  we have established that

1973: the effective capacity of transmissions for each edge under

1974: the above described TDMA scheme is $\Omega(\log^{-\alpha} n)$.

1975: That is, each edge of $\hat{E}$ gets capacity at least

1976: $\Omega(\log^{-\alpha} n)$. Now recall that a grid graph

1977: with unit capacity has $f^*$ lower bounded

1978: as $ \Omega\left(\frac{1}{n^{3/2} \log n}\right)$.

1979: Hence, using this routing of UMF along edges of $\hat{E}$

1980: with capacity $\Omega(\log^{-\alpha} n)$ we obtain that

1981: $$ f^* = \Omega\left(\frac{1}{n^{3/2} \log^{1+\alpha} n}\right).$$

1982: By careful accounting of probability of relevant events above

1983: and Union bound of events will imply that the above stated lower

1984: bound on $f^*$ holds with probability at least $1-1/n^2$. Since

1985: $f^* \geq 0$ with probability $1$, it immediately implies the

1986: desired lower bound of Lemma:

1987: $$ \E[f^*] = \Omega\left(\frac{1}{n^{3/2} \log^{1+\alpha} n}\right).$$

1988: This completes the proof of Lemma \ref{lem:coh_expl}.

1989: 

1990: 

1991: 

1992: \end{proof}

1993: 

1994: \vspace{.1in}

1995: 

1996: 

1997: 

1998: 

1999: \vspace{0.3in}

2000: 

2001: %\subsubsection{Computational Example}

2002: 

2003: \section*{Acknowledgment}

2004: The authors would like to thank A. Jovicic and S. Tavildar for a helpful discussion on the information

2005: theoretic min-cut bound for networks with Gaussian channels.

2006: We would like to thank Ashish Goel for sharing his proof about evaluating

2007: mixing time of $G(n,r)$ -- it has influenced the proof of Lemma \ref{lem:gnrcomb}.

2008: 

2009: 

2010: \bibliographystyle{IEEEtran}

2011: \bibliography{cap}

2012: 

2013: 

2014: \end{document}

2015: 

2016: 

2017: