1: \documentclass[12pt]{article}

2: \usepackage{amsfonts,latexsym,fullpage,amsmath,latexsym,amssymb}

3: 

4: 

5: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

6: % 

7: % Definitions

8: %

9: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

10: 

11: \newtheorem{Definition}{Definition}

12: \newtheorem{Theorem}{Theorem}

13: \newtheorem{Lemma}{Lemma}

14: \newtheorem{Corollary}{Corollary}

15: \newtheorem{Remark}{Remark}

16: \newtheorem{example}[Theorem]{Example}

17: \newtheorem{definition}[Theorem]{Definition}

18: 

19: \newcommand{\onemat}[0]{{\mathbf 1}}

20: \newcommand{\zeromat}[0]{{\mathbf 0}}

21: \newcommand{\cT}[0]{{\mathcal T}}

22: \newcommand{\cI}[0]{{\mathcal I}}

23: \newcommand{\cH}[0]{{\mathcal H}}

24: \newcommand{\cS}[0]{{\mathcal S}}

25: \newcommand{\C}[0]{{\mathbb{C}}}

26: \newcommand{\R}[0]{{\mathbb{R}}}

27: \newcommand{\N}[0]{{\mathbb{N}}}

28: \newcommand{\Z}[0]{{\rm Z}}

29: \newcommand{\SL}[0]{{\rm SL}}

30: \newcommand{\GL}[0]{{\rm GL}}

31: \newcommand{\DFT}[0]{{\rm DFT}}

32: \newcommand{\qed}[0]{\hfill $\Box$}

33: \newcommand{\cE}[0]{{\mathcal E}}

34: \newcommand{\cA}[0]{{\mathcal A}}

35: %\newcommand{\qed}[0]{\hfill\rule{0.86ex}{0.86ex}}

36: 

37: 

38: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

39: %

40: % Title and Authors

41: %

42: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

43: 

44: \title{\Large \textbf{Lower Bound on the Chromatic Number by\\

45: Spectra of Weighted Adjacency Matrices}} \author{ Pawe{\l}

46: Wocjan\thanks{e-mail: {\protect\tt

47: \{wocjan,janzing,eiss\_office\}@ira.uka.de}}, Dominik Janzing, Thomas

48: Beth \\ \small Institut f{\"u}r Algorithmen und Kognitive Systeme,

49: Universit{\"a}t Karlsruhe,\\[-1ex] \small Am Fasanengarten 5,

50: D-76\,128 Karlsruhe, Germany\\[-1ex] \small Forschungsgruppe Quantum

51: Computing}

52: 

53: \date{December 21, 2001}

54: 

55: \begin{document}

56: 

57: \maketitle

58: 

59: \abstract{A lower bound on the chromatic number of a graph is derived

60: by majorization of spectra of weighted adjacency matrices. These

61: matrices are given by Hadamard products of the adjacency matrix and

62: arbitrary Hermitian matrices.}

63: 

64: \section{Introduction}

65: Let $G=(V,E)$ be a graph on $n$ vertices $V=\{0,\ldots,n-1\}$. An

66: admissible vertex coloring of $G$ is an assignment of colors to its

67: vertices such that no two adjacent vertices receive the same

68: color. The minimum number of colors needed for such a coloring is

69: called the {\it chromatic number} of $G$, and is denoted

70: $\chi$. Determining the chromatic number of a graph is known to be

71: NP-hard \cite{garey}. Besides its theoretical significance as a

72: canonical NP-hard problem, graph coloring arises naturally in a

73: variety of applications such as e.g. register allocation and

74: scheduling. Therefore it is important to have bounds on the chromatic

75: number.

76: 

77: Bounds are often computed with the help of the adjacency matrix. Let

78: $A=(a_{ij})$ denote the adjacency matrix of $G$, i.\,e.

79: $$

80: a_{kl}=\left\{

81: \begin{array}{ll}

82: 1 & \mbox{if vertices $k$ and $l$ are adjacent} \\

83: 0 & \mbox{otherwise}

84: \end{array}

85: \right.

86: $$

87: and $\lambda_1\ge\lambda_2\ge\cdots\ge\lambda_n$ denote the

88: eigenvalues of $A$. The upper bound $\lambda_1+1\ge\chi$ has been

89: shown in \cite{wilf}. The lower bound

90: \begin{equation}\label{hoffman}

91: \chi\ge\frac{\lambda_1}{|\lambda_n|}+1\,

92: \end{equation}

93: has been shown in \cite{Hoffman70}. The following Theorem proved in

94: \cite{barnes} gives a generalization of the bound (\ref{hoffman}).

95: 

96: \begin{Theorem}\label{barnestheorem}

97: Let $A$ denote the adjacency matrix for a connected graph $G$ on $n$

98: vertices, and let $D=diag(d_1,\ldots,d_n)$ be a diagonal matrix such

99: that $A+D$ is positive semidefinite. Then each $d_i$ is positive, and

100: the largest eigenvalue of the matrix $D^{-\frac{1}{2}}A

101: D^{-\frac{1}{2}}+I$ is a lower bound for $\chi$.

102: \end{Theorem}

103: 

104: Note that the bound (\ref{hoffman}) can be obtained from

105: Theorem~\ref{barnestheorem} by taking $D=|\lambda_n| I$.

106: 

107: 

108: \section{Lower bounds for $\chi$}

109: We derive a method to obtain lower bounds on the chromatic

110: number. This method contains the above results on the lower bound as

111: special cases.

112: 

113: First we define the problem to reverse the sign of a matrix and the

114: cost of doing this. We consider weighted adjacency matrices, i.e.\

115: special matrices constructed from the adjacency matrix of a graph $G$,

116: and show that $\chi-1$ is an upper bound on the cost. Then a lower

117: bound is derived on the cost in terms of the spectra of the weighted

118: adjacency matrices. This yields a lower bound on the chromatic number.

119: \begin{Definition}[Sign reversal]

120: Let $M\in\C^{n\times n}$ be any traceless matrix. A sign reversal map

121: $\cI:=(r_1,U_1;r_2,U_2;\ldots;r_N,U_N)$ for $M$ is characterized by

122: $N$ positive real numbers $r_j$ and unitary matrices

123: $U_j\in\C^{n\times n}$ such that

124: \begin{equation}

125: \sum_{j=1}^N r_j U_j^{-1} M U_j = -M\,.

126: \end{equation}

127: We call $c(\cI):=\sum_{j=1}^N r_j$ the cost of the sign reversal map

128: $\cI$. 

129: \end{Definition}

130: Using basic results of representation theory we can show that this is

131: always possible. Let $G$ be a group acting irreducibly on $\C^n$ via

132: the unitary representation $g\mapsto U_g$. Let $M\in\C^{n\times n}$

133: matrix. By Schur's Lemma \cite{group} we have

134: \begin{equation}

135: \frac{1}{|G|}\sum_{g\in G} U_g^{-1} M U_g = \frac{{\rm tr}(M)}{n}\onemat\,, 

136: \end{equation}

137: where ${\rm tr}(M)$ denotes the trace of $M$ and $\onemat$ is the

138: identity matrix of size $N$. 

139: 

140: In the following the matrix $M$ is always traceless. Then by conjugating

141: $M$ with all elements of $G$ but the identity we obtain 

142: \[

143: \sum_{G\setminus\{1\}} U_g M U_g = -M

144: \]

145: since $M$ is traceless.

146: 

147: We consider matrices constructed from the adjacency matrix of a graph

148: $G$ and derive a lower bound on the cost of sign reversal in terms of

149: the chromatic number of $G$. To do this we need the definition of the

150: Hadamard product. If $A=(a_{ij})\in\C^{n\times n}$ and

151: $B=(b_{ij})\in\C^{n\times n}$ are given, then the Hadamard product of

152: $A$ and $B$ is the matrix $A*B=(a_{ij} b_{ij})\in\C^{n\times n}$

153: \cite{HJ90}.

154: 

155: \begin{Lemma}\label{invertChromatic}

156: Let $A=(a_{kl})$ be the adjacency matrix of a graph $G=(V,E)$ with $n$

157: vertices and chromatic number $\chi$. For an arbitrary matrix

158: $W=(w_{kl})\in\C^{n\times n}$ the sign of the Hadamard product

159: $M:=W*A$ can be reversed with cost $\chi-1$.

160: \end{Lemma}

161: {\bf Proof:} Choose a partition (disjoint union) of the vertices

162: $V=V_0\cup V_1\cup\ldots\cup V_{\chi-1}$ corresponding to a (minimal)

163: coloring. The set $V_c$ contains all vertices of color $c$

164: ($c\in\{0,\ldots,\chi-1\})$. We assume that the vertices are ordered

165: according to their colors, i.e. first come the vertices of color $0$,

166: than of color $1$, etc.

167: 

168: Let $\omega\in\mathbb{C}$ be a primitive $\chi$-th root of unity,

169: i.e. $\omega^{\chi}=1$. Define a diagonal matrix $D:={\rm

170: diag}(d_0,d_1,\ldots,d_{n-1})$ where $d_i:=\omega^c$ if $i\in V_c$. We

171: show that

172: \[

173: \bar{M}:=\sum_{j=0}^{\chi-1} D^{-j} M D^j\,.

174: \]

175: is the zero matrix. Let $M=(m_{kl})_{k,l=0,\ldots n-1}$ and

176: $\bar{M}=(\bar{m}_{kl})_{k,l=0,\ldots n-1}$. The entries of $\bar{M}$

177: are given by

178: \[

179: \bar{m}_{kl}=\sum_{j=0}^{\chi-1} \omega^{-kj} m_{kl} \omega^{lj}\,.

180: \]

181: Let $k,l\in V_c$ with $k\neq l$. The vertices $k$ and $l$ have the the

182: same color and consequently they cannot be adjacent,

183: i.e. $a_{kl}=0$. Therefore we have $\bar{m}_{kl}$ since $m_{kl}=w_{kl}

184: a_{kl}=0$. Note that $\bar{m}_{kk}=0$ since $a_{kk}=0$ (the diagonal

185: entries of the adjacency matrix are all zero).

186: 

187: Now let $k\in V_c$ and $l\in V_{\tilde{c}}$ with $c\neq\tilde{c}$. We

188: have

189: \[

190: \bar{m}_{kl}=\sum_{j=0}^{\chi-1} \omega^{-kj} m_{kl} \omega^{lj}=0

191: \]

192: since the vectors $({\omega}^{c\, 0},{\omega}^{c\, 1},\ldots,

193: {\omega}^{c\,(\chi-1)})$ and $({\omega}^{\tilde{c}\,

194: 0},{\omega}^{\tilde{c}\, 1},\ldots, {\omega}^{\tilde{c}\, (\chi-1)})$

195: are orthogonal. Note that they are rows of the matrix of the discrete

196: Fourier transform of size $\chi$.  \qed

197: 

198: \vspace*{12pt} To derive a lower bound on the cost of a sign reversal

199: map we introduce the concept that is known as spectral majorization of

200: matrices \cite{Bhatia:96}. Let $x=(x_1,\ldots,x_n)$ and

201: $y=(y_1,\ldots,y_n)$ be two $n$-dimensional real vectors. We introduce

202: the notation $\downarrow$ to denote the components of a vector

203: rearranged into non-increasing order, so

204: $x^\downarrow=(x_1^\downarrow,\ldots,x_n^\downarrow)$, where

205: $x_1^\downarrow\ge x_2^\downarrow\ge\ldots\ge x_n^\downarrow$. We say

206: that $x$ is majorized by $y$ and write $x\prec y$, if

207: \[

208: \sum_{i=1}^m x_i^\downarrow \le \sum_{i=1}^m y_i^\downarrow\,,

209: \]

210: for $m=1,\ldots,n-1$ and $\sum_{i=1}^n x_i^\downarrow = \sum_{i=1}^n

211: y_i^\downarrow$.

212: 

213: Let $\mathrm{Spec}(X)$ denote the spectrum of the Hermitian matrix

214: $X$, i.e.\ the vector of eigenvalues. Recall that the eigenvalues of a

215: Hermitian matrix are real. Ky Fan's maximum principle \cite{Bhatia:96}

216: gives rise to a useful constraint on the eigenvalues of a sum of

217: Hermitian matrices $A_j$ for $j=1,\ldots,N$:

218: \begin{equation}\label{major}

219: \mathrm{Spec}\big(\sum_{j=1}^N A_j\big)\prec\sum_{j=1}^N

220: \mathrm{Spec}(A_j)\,.

221: \end{equation}

222: The following lemma gives a lower bound on the cost of sign reversal

223: in terms of the spectrum of $M$.

224: \begin{Lemma}\label{lowerBound}

225: A lower bound on the negation cost of a Hermitian matrix $M$ is given

226: by the smallest positive real number $\tau$ such that

227: \begin{equation}

228: \mathrm{Spec}(-M)\prec\tau\,\mathrm{Spec}(M)\,.

229: \end{equation}

230: \end{Lemma}

231: {\bf Proof:} Let $\cI:=(r_1,U_1;r_2,U_2;\ldots r_N,U_N)$ be an

232: arbitrary sign reversal map for $M$, i.e.

233: \begin{equation}\label{inverteq}

234: \sum_{j=1}^N r_j\, U_j^{-1} M U_j = -M\,.

235: \end{equation}

236: Applying the inequality~(\ref{major}) to eq.~(\ref{inverteq}) yields

237: \[

238: \mathrm{Spec}(-M)=\mathrm{Spec}\big(\sum_{j=1}^N r_j U_j^{-1} M

239: U_j\big) \prec \sum_{j=1}^N r_j\,\mathrm{Spec}(U_j^{-1} M U_j) =

240: c(\cI)\,\mathrm{Spec}(M)\,.

241: \]

242: Therefore for all sign reversal maps we have $c(\cI)\ge\tau$.  \qed

243: 

244: \vspace*{12pt} Let $\lambda_1,\lambda_2,\ldots,\lambda_n$ be the

245: eigenvalues of $M$ with

246: $\lambda_1\ge\lambda_2\ge\ldots\ge\lambda_n$. Then the lower bound $\tau$

247: is given by

248: \begin{equation}\label{max}

249: \tau:=\max_{m=1,\ldots n-1}\Big\{ \frac{\sum_{i=1}^m \lambda_i}{-\sum_{i=1}^m

250: \lambda_{n+1-i}}\Big\}

251: \end{equation}

252: since for the eigenvalues $\mu_i$ of $-M$ we have

253: $\mu_i=\lambda_{n+1-i}$ ($i=1\ldots,n$).

254: 

255: By using the results on the lower and upper bounds on the cost of sign

256: reversal we obtain:

257: \begin{Theorem}[Lower bound on the chromatic number]

258: Let $A$ be the adjacency matrix of a graph $G$ with $n$ vertices and

259: $\chi$ its chromatic number. For any Hermitian matrix $W\in\C^{n\times

260: n}$ let $\tau_W$ denote the minimal positive real number such that

261: \begin{equation}\label{lowerW}

262: \lambda(-W*A)\prec\tau_W\,\lambda(W*A)\,.

263: \end{equation}

264: Let $\tau$ be the maximal $\tau_W$, where the maximum is taken over

265: all Hermitian matrices $W\in\C^{n\times n}$. Then we have

266: \begin{equation}

267: \chi \ge \tau+1\,.

268: \end{equation}

269: \end{Theorem}

270: {\bf Proof:} Lemma~\ref{invertChromatic} shows that the sign of all

271: matrices of the form $W*A$ can be reversed with cost

272: $\chi-1$. Lemma~\ref{lowerBound} gives $\tau_W$ in eq.~(\ref{lowerW})

273: as a lower bound on the sign reversal cost of $W*A$. Consequently, we

274: have $\chi\ge \tau_W +1$ for all Hermitian matrices. We choose the

275: maximal $\tau_W$ to obtain the best bound on $\chi$. \qed

276: 

277: \vspace*{12pt} Note that we obtain as a special case the well-known

278: lower bound $\chi\ge\frac{\lambda_1}{|\lambda_n|}+1$. Set $W$ to be

279: the matrix all of whose entries are equal to $1$. Then we have

280: $A=W*A$. Instead of taking the maximum over $m=1,\ldots,n-1$ consider

281: only $m:=1$ in eq.~(\ref{max}). This corresponds to the maximal and

282: minimal eigenvalues of $A$.

283: 

284: We now show that Theorem~1 can also be understood as a special case of

285: Theorem~2. Let $D:=diag(d_0,\ldots,d_{n-1})$ be the diagonal matrix

286: defined as in Theorem~1. Take $W$ to be the matrix with entries

287: $w_{kl}:=\sqrt{d_k}\sqrt{d_l}$. Then we have $W*A=D^{-\frac{1}{2}} A

288: D^{-\frac{1}{2}}$. This shows that this modified adjacency matrix can

289: also be expressed with the Hadamard product.

290: 

291: As a conclusion Theorem~2 permits to consider a larger class of

292: modified adjacency matrices and to take into account all eigenvalues

293: (Theorem~1 considers only the maximal and minimal eigenvalues). The

294: advantage of the method presented in \cite{barnes} is that the matrix

295: $D$ is the solution of a semidefinite programming problem that can be

296: computed by an algorithm given there. It remains to be shown whether

297: the larger class of modified adjacency matrices permits to obtain

298: better lower bounds and whether there are efficient algorithms to

299: compute them. Nevertheless, the derivation in our approach is easy and

300: gives a generalization of Theorem~2.

301: 

302: We would like to point out that the sign reversal map described here

303: is an abstraction of a problem in quantum physics. A physical system

304: evolving due to the Schr\"odinger equation

305: \[

306: \frac{d}{dt}

307: |\psi (t)\rangle = -i H |\psi (t)\rangle

308: \]

309: (where $H$ is the so-called ``Hamilton operator'') can be made evolve

310: backwards in time by interspersing the natural time evolution with

311: external control operations \cite{graph,arrow}. The unitary maps $U_j$

312: in Definition~1 are abstractions of control operations. The vertices

313: of the graph represent particles and the edges indicate whether there

314: is an interaction between the two particles or not. The physical

315: meaning of the inversion cost is the time overhead for simulating the

316: inverse evolution, i.e., the factor by which the reverse dynamics is

317: slower than the original one. The cost is therefore the {\it

318: complexity} of reversing the dynamical evolution and the chromatic

319: number is an upper bound on this time complexity.

320: 

321: 

322: \section*{Acknowledgments}

323: This work has been supported by grands of the DFG Schwerpunktprogramm

324: {\it Komplexit\"at und Energie}.

325: 

326: 

327: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

328: %

329: % The Literature

330: %

331: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

332: 

333: \begin{thebibliography}{1}

334: 

335: \bibitem{barnes}

336: E.~R. Barnes.

337: \newblock A lower bound for the chromatic number of a graph.

338: \newblock {\em Contemporary Mathematics}, 275:3--12, 2001.

339: 

340: \bibitem{Bhatia:96}

341: R.~Bhatia.

342: \newblock {\em {Matrix analysis}}.

343: \newblock Springer, Graduates Texts in Mathematics 184, 1996.

344: 

345: \bibitem{garey}

346: M.~R. Garey and D.~S. Williamson.

347: \newblock {\em Computers and Intractability: A Guide to the Theory of

348:   NP-Completeness}.

349: \newblock Freeman, 1979.

350: 

351: \bibitem{Hoffman70}

352: A.~J. Hoffman.

353: \newblock On eigenvalues and colorings of graphs.

354: \newblock {\em Graph Theory and its Applications (B. Harris, ed.), Academic

355:   Press}, pages 78--91, 1970.

356: 

357: \bibitem{HJ90}

358: R.A. Horn and C.R. Johnson.

359: \newblock {\em {Matrix Analysis}}.

360: \newblock Cambridge University Press, 1990.

361: 

362: \bibitem{group}

363: G.~James and M.~Liebeck.

364: \newblock {\em Representations and characters of groups}.

365: \newblock Cambridge University Press, 1993.

366: 

367: \bibitem{arrow}

368: D.~Janzing, P.~Wocjan, and Th. Beth.

369: \newblock {Complexity of inverting $n$-spin interactions: Arrow of time in

370:   quantum control}.

371: \newblock Technical report, Los Alamos National Laboratory, 2001.

372: \newblock LANL preprint quant--ph/0106085.

373: 

374: \bibitem{wilf}

375: H.~S. Wilf.

376: \newblock The eigenvalues of a graph and its chromatic number.

377: \newblock {\em J. London Math. Society}, 42:300--332, 1967.

378: 

379: \bibitem{graph}

380: P.~Wocjan, D.~Janzing, and Th. Beth.

381: \newblock {Simulating Arbitrary Pair-Interactions by a Given Hamiltonian:

382:   Graph-Theoretical Bounds on the Time Complexity}.

383: \newblock Technical report, Los Alamos National Laboratory, 2001.

384: \newblock LANL preprint quant--ph/0106077.

385: 

386: \end{thebibliography}

387: 

388: 

389: \end{document}

390: 

391: 

392: 

393: 

394: 

395: 

396: 

397: 

398: