1: \documentclass[12pt]{amsart}

2: 

3: %\addtolength{\hoffset}{-1.1cm}

4: %\addtolength{\textwidth}{2.2cm}

5: 

6: \usepackage[english]{babel}

7: \usepackage{amsmath}

8: \usepackage{amsfonts}

9: \usepackage{amssymb}

10: \usepackage{amsthm}

11: \usepackage{latexsym}

12: %\usepackage{amstext}

13: 

14: %\usepackage{subfigure}

15: 

16: \usepackage[dvips]{graphicx}

17: \usepackage{psfrag}

18: %\documentclass[12pt]{amsart}

19: \title{Holomorphic Matrix Integrals}

20: \thanks{This paper is based on the ETH diploma thesis

21: of the second author}

22: \author{Giovanni Felder}

23: \author{Roman Riser}

24: 

25: \begin{document}

26: 

27: 

28: 

29: \maketitle

30: \centerline{ \it Department of Mathematics, ETH Zurich}

31: \centerline{ \it 8092 Zurich, Switzerland}

32: 

33: \begin{abstract}

34: We study a class of holomorphic matrix models.

35: The integrals are taken over middle dimensional 

36: cycles in the space of complex square matrices. As

37: the size of the matrices tends to infinity, the

38: distribution of eigenvalues is given by a measure

39: with support on a collection of arcs in the complex

40: planes. We show that the arcs are level sets of

41: the imaginary part of a hyperelliptic integral

42: connecting branch points.

43: \end{abstract}

44: 

45: 

46: 

47: \section{Introduction}

48: 

49: 

50: According to Dijkgraaf and Vafa \cite{Dij1,Dij3}, the effective glueball superpotential

51: of $\mathcal N=1$ supersymmetric $U(n)$ gauge theory has an asymptotic expansion given by the 

52: planar part of the topological 

53: expansion of a matrix model. To give the effective potential

54: for various vacua of the gauge theory, Dijkgraaf and

55: Vafa propose a formula obtained by saddle point expansion

56: of the matrix integral around different critical points. The

57: filling fractions, i.e., the fraction of the eigenvalues sitting

58: close to each of the critical points of the potential

59: are the parameters selecting the vacua.

60: 

61: In this paper we give a way to define non-perturbatively

62: (i.e., beyond the saddle point expansion) the matrix

63: integrals. Different integrals are integrals over

64: different cycles in the space of normal matrices. In

65: the case of generic polynomial potentials it is possible

66: to construct a cycle for each of the critical points, so

67: that all effective potentials considered by Dijkgraaf

68: and Vafa arise in the asymptotic expansion of our integrals. 

69: We note that the idea of integrating over eigenvalues along 

70: suitable contours in the complex plane appears in special

71: situations in \cite{D}. David's contours were also more recently considered

72: in \cite{La},  where a (non-holomorphic) modification of the

73: integral was proposed to obtain arbitrary filling factors.  

74: 

75: 

76: A second result of this paper is the description of the

77: asymptotic distribution of eigenvalues for each critical 

78: point. The eigenvalues, as predicted by Dijkgraaf and Vafa,

79: lie asymptotically along arcs connecting 

80: branch points of a two-fold cover of 

81: the complex plane. We give

82: a reality condition which specifies the shape of the

83: arcs and the density of eigenvalues.

84:  

85: Before considering the case of general 

86: $N\times N$ matrices, it

87: is instructive to consider the case $N=1$ of ordinary integrals. 

88: Consider the Airy integral 

89: \[

90: u(x)=\int e^{\frac i\mu(xt+t^3/3)}dt.

91: \]

92: This integral gives a formal solution of the Airy differential equation

93: $-\mu^2u''(x)+xu(x)=0$. Indeed the saddle point expansion around

94: the two critical points of the integrand give the two linear

95: independent formal power series solutions. The non-perturbative

96: solutions can be obtained by integrating along paths in the

97: complex plane connecting any two of 

98: the three directions at infinity where the integrand goes to zero.

99: More generally,

100:  the integrals of the form $\int \exp(\frac i\mu(xt+P(t))dt$ 

101: for a generic polynomial $P(t)$ of degree $n$ are solutions of

102: a differential equation of order $n-1$. We show that a system of $n-1$ 

103: linearly independent solutions are obtained by choosing $n$

104: integration contours in the complex plane

105: going to infinity in directions where the integrand decays 

106: exponentially. Moreover the contours can be chosen

107: so that each of them passes through exactly one of the

108: $n$ critical points. It follows that the $n-1$ 

109: saddle point expansions at each of the critical points are 

110: the asymptotic expansions of true solutions given by convergent

111: integrals.

112: 

113: In the case of matrices we consider integrals of

114: the form 

115: \begin{equation}\label{e-hmi}

116: \langle F\rangle_N=

117: Z_N^{-1}\int_\Gamma F(\Phi)\exp\left(-\frac N\mu

118: \,\mathrm{tr}\,p(\Phi)\right) d\Phi

119: \end{equation}

120: over real $N^2$-dimensional cycles $\Gamma$ in

121: the space of complex $N\times N$ matrices.

122: The potential $p$ is a polynomial with complex

123: coefficients,

124:  $d\Phi=\wedge_{j,k} d\Phi_{jk}$ and

125: the normalization factor $Z_N$ is such that

126: $\langle 1\rangle_N=1$. The observables

127: $F(\Phi)$ are holomorphic functions

128: invariant under conjugation.

129:  In the well-studied

130: case of integrals over hermitian matrices, with a polynomial $p$ with real coefficients  bounded below, the

131: saddle point approximation becomes exact in the large

132: $N$ limit with fixed $\mu$ and the relevant critical

133: point is described by an asymptotic

134:  {\em density of eigenvalues} which has

135: support  on a union of intervals on the real axis. If

136: $\mu$ is small, these intervals are small neighborhoods

137: of the minima of the potential $p$. 

138: To obtain the densities of eigenvalues needed to make

139: contact to gauge theory one considers the variational

140: problem for critical points subject to the side condition

141: that the fraction of eigenvalues in the vicinity of

142: each of the critical points (not just minima) are given

143: numbers.

144: 

145: We consider here the case of a generic polynomial

146: $p$ of degree $n$

147: with complex coefficients, with distinct critical

148: points $z_1,\dots,z_{n-1}$, and propose to consider

149: integrals over cycles in the space of normal matrices

150: (a matrix is normal if it commutes with its adjoint or,

151: equivalently, if it is conjugated to a diagonal matrix

152: by a unitary matrix). The cycles we consider are 

153: parametrized by integers $N_1,\dots,N_{n-1}$ whose

154: sum is $N$ and are characterized by the condition

155: that $N_k$ eigenvalues belong to a path $\Gamma_k$

156: in the complex plane

157: going through $z_k$ and going to infinity in a direction

158: where $\mathrm{Re}(p(z))\to\infty$. In the

159: limit $N\to \infty$ with $n_k=N_k/N$ fixed, the

160: eigenvalues in the saddle point approximation (supposed

161: to be exact in the limit) are distributed along

162: $n-1$ arcs in the complex plane. For small $\mu$ the

163: arcs are close to the critical points and the

164: $k$th arcs contains the fraction $n_k$ of the

165: eigenvalues.

166: 

167: 

168: \section{The one-dimensional case}

169: Let $p(z)=a_nz^n+\cdots+a_1z+a_0$ be a 

170: polynomial of degree $n\geq 2$ with

171: complex coefficients. 

172: We want to consider integrals of the form

173: \[

174: \int_\Gamma q(z)e^{\textstyle{-\frac1\mu p(z)}}dz,\qquad\mu>0.

175: \]

176: for polynomials $q(z)$.

177: Before considering the question of integration cycles

178: we may evaluate such integrals as asymptotic

179: series as $\mu\to 0$ by formal application of the

180: saddle point method at each of the critical points

181: $z_1,\dots,z_{n-1}$, which we assume to be

182: distinct.

183: In this way we get $n-1$ asymptotic series

184: of the form $\exp(-p(z_k)/\mu)(c_1+c_2\mu+\cdots)$

185: and the question is whether these are asymptotic

186: expansions of our integral for suitable cycles 

187: $\Gamma_k$.

188: 

189: The cycles which we should consider here are

190: (linear combinations of) paths for which the

191: integral converges. As the integrand is holomorphic,

192: homotopic paths will give the same answer and what

193: matters is the behavior at infinity. As $z\to\infty$,

194: $p(z)\sim a_n z^n$, so there are $n$ directions

195: in the complex plane for which 

196: $\mathrm{Re}(p(z))\to +\infty$ as $z$

197: tends to infinity in these directions. Let us

198: call these asymptotic directions valleys as in these

199: directions the integrand decays exponentially.

200: Neighboring valleys are separated by hills, which

201: are directions of exponential increase of the

202: integrand.

203: So the cycles one needs to consider are linear

204: combinations of infinite

205: paths connecting  pairs of distinct valleys. As paths

206: connecting two valleys can be deformed into sums

207: of paths connecting the two valleys with any third

208: one, there are only $n-1$ linearly independent

209: cycles.

210: 

211: Let us assume for simplicity that the critical

212: values $p(z_k)$ have distinct imaginary parts.

213: Then there is a canonical way to associate to

214: each critical point $z_k$ a path $\Gamma_k$

215: in such a way that the asymptotic expansion

216: of the integral over $\Gamma_k$ as $\mu\to 0$

217: is obtained by the saddle point expansion

218: at $z_k$. Namely, we take the {\em steepest descent} paths(see, e.g., \cite{bru} and \cite{erd})

219: emerging from $z_k$, defined by the

220: condition that the tangent vector at each

221: point points in the direction of the

222: gradient of $\mathrm{Re}(p(z))$. As $p(z)$

223: is holomorphic, the gradient of $\mathrm{Re}\,p(z)$

224: is orthogonal to the gradient of $\mathrm{Im}\,p(z)$

225: by the Cauchy--Riemann equations. Thus steepest

226: descent paths are level lines for the imaginary

227: part of $p(z)$. Each non-degenerate

228: critical point $z_k$ is at the intersection of 

229: two such level lines. One of these two lines, the

230: one along which $\mathrm{Re}(p(z))$ takes its minimum

231: at $z_k$, is the steepest descent path $\Gamma_k$.

232: Along the other line $\Gamma_k'$, the real part

233: takes its maximum at $z_k$.

234: We claim that $\Gamma_k$ is a smooth path 

235: going to infinity in both directions and connecting

236: two valleys. Indeed, $\Gamma_k$ is (in suitable

237: parametrization) given by a solution of the

238: differential equation $\dot z(t)=\overline{p'(z(t))}$.

239: It follows that $\frac d{dt}\mathrm{Re}(p(z(t)))=

240: |p'(z(t))|^2$. By our assumption, critical

241: values have distinct imaginary parts, so the

242: steepest descent path passing through $z_k$

243: may not come close to any other critical point.

244: Thus $|p'(z(t))|$ is bounded below so that, 

245: as we go away from $z_k$,

246: $\mathrm{Re}\,p(z(t)$ must go to infinity and the

247: path $\Gamma_k$ connects two valleys. 

248: Similarly $\Gamma'_k$ connects two hills. As

249: $\Gamma_k$ and $\Gamma'_k$ cross at $z_k$, the

250: two valleys connected by $\Gamma_k$ are separated

251: by hills and are thus different.

252: 

253: We have thus shown that for each critical

254: point $z_k$ there is a  steepest

255: descent path $\Gamma_k$ going through $z_k$

256: and connecting pairs of different valleys.

257: On $\Gamma_k$ the real part of $p(z)$ is minimal

258: at $z_k$ so that the saddle point expansion

259: at $z_k$ indeed gives the asymptotic expansion

260: of the integral over $\Gamma_k$.

261: 

262: 

263: \section{Matrix integrals}

264: \subsection{Integration cycles}

265: We consider matrix integrals of the form \eqref{e-hmi}

266: for $p(z)$ a polynomial of degree $n$.

267: They are integrals of holomorphic differential forms 

268: over $N^2$ dimensional cycles. For each set of

269: natural numbers $N_1,\dots, N_{n-1}$ summing up to $N$,

270: we have a cycle in the normal matrices, characterized

271: by the condition that $N_k$ eigenvalues run over

272: the path $\Gamma_k$ of the previous section.

273: More precisely, the cycle $\Gamma$ is parametrized

274: by $\mathbb R^N\otimes U(N)/U(1)^N$:

275: \[

276: (t,U)\mapsto U\,

277: \left(\begin{array}{ccc}

278: \lambda_1(t_1)&    & \\

279:         &\ddots & \\

280:         &      & \lambda_N(t_{N})

281:       \end{array}

282: \right)

283: \,U^{-1}.

284: \]

285: The first $N_1$ diagonal elements

286:  $\lambda_1(t),\dots, \lambda_{N_1}(t)$ are parametrizations of

287: $\Gamma_1$, the next $N_2$ are parametrizations of $\Gamma_2$

288: and so on.

289: 

290: The usual argument to reduce the integral to an integral

291: over the eigenvalues (see \cite{BIZ}) gives

292: \[

293: \langle F\rangle_N=\frac 1{Z'_N}

294: \int F(\lambda_1,\dots,\lambda_N)

295: e^{\textstyle{-\frac 

296: N\mu\sum_{j=1}^n p(\lambda_j)}}\prod_{j<k}(\lambda_j-\lambda_k)^2

297: d\lambda_1\cdots d\lambda_N,

298: \]

299: The integral is over $\Gamma_1^{N_1}\times\cdots\times\Gamma_{n-1}

300: ^{N_{n-1}}$ and the function $F$, a function on matrices

301: invariant under conjugation, is regarded as a

302: symmetric function of the eigenvalues.

303: 

304: \subsection{The loop equation}

305: To study the large $N$ limit, 

306: it is useful to introduce the trace of the

307: resolvent: 

308: \[

309: \omega(z)=\frac1N\,\mathrm{tr}\left(\frac1{z-\Phi}

310: \right)

311:          =\frac1N\sum_{j=1}^N\frac1{z-\lambda_j},

312: \]

313: as products of such traces are generating functions of 

314: polynomial functions invariant under conjugation.

315: The ``loop equation'' \cite{W,M} for this quantity is

316: \[

317: \mu\langle\omega(z)^2\rangle_N-p'(z)\langle

318: \omega(z)\rangle_N+f_N(z)=0,

319: \]

320: where 

321: \[

322: f_N(z)=\frac1N

323: \left\langle \sum_i\frac{p'(z)-p'(\lambda_i)}{z-\lambda_i}

324: \right\rangle_N

325: \]

326: is a polynomial of degree $n-2$ with leading 

327: coefficient $na_n$.

328: This equation can be derived from the identity

329: \[

330: 0=\sum_{i=1}^N\int\frac\partial{\partial\lambda_i}

331: \left(\frac1{z-\lambda_i}e^{-\frac N\mu\sum_j

332: p(\lambda_j)}\prod_{j>k}(\lambda_j-\lambda_k)^2

333: \right)\prod d\lambda_j.

334: \]

335: In the limit $N\to\infty$, the matrix integral

336: is supposed to be dominated by an integral

337: over a region where the eigenvalues are close

338: (for small $\mu$)

339: to critical points of $p$. With our choice of

340: integration cycles and keeping $n_j=N_j/N$

341: fixed as $N\to\infty$, there will be a fraction

342: $n_j$  of the eigenvalues close to $z_j$.

343: 

344: In the limit $N\to\infty$ one expects that 

345: the saddle point approximation becomes exact

346: and thus  $\langle\omega(z)\rangle_N$ converges to 

347: \[

348: \langle\omega(z)\rangle=\int \frac1{z-z'}d\nu(z'),

349: \]

350: for some probability measure $\nu$ with support

351: in regions around the critical points of $p$ and

352: so that $n_k$ is 

353: the measure of the region around $z_k$.

354: Technically one assumes that the limit $N\to\infty$

355: exists and that $\lim_{N\to\infty}

356: (\langle\omega(z)^2\rangle_N-\langle\omega(z)\rangle_N^2)

357: =0$. 

358: The function $\langle\omega(z)\rangle$ is defined and

359: holomorphic for $z$ outside the support of the measure.

360: Setting 

361: \begin{equation}\label{e-y}

362: y(z)=2\mu\langle\omega(z)\rangle-p'(z),

363: \end{equation}

364: we finally obtain

365: \begin{equation}\label{e-hy}

366: y(z)^2=p'(z)^2-4\mu f(z),

367: \end{equation}

368: for some polynomial $f(z)$ of degree $n-2$ with

369: leading coefficient $na_n$. Thus the function $y(z)$,

370: which is a priori defined on the complement of the

371: support of the measure, has an analytic continuation

372: to a two-fold covering of the complex plane. The

373: original function $y$ is the branch of the function

374: defined by \eqref{e-hy} which behaves at infinity

375: as $-p'(z)$.

376: 

377: \subsection{The density of eigenvalues}

378: With the analogy with the case of hermitian matrices in

379: mind, it is reasonable to assume that the measure

380: $\nu$ has support on a collection of

381: arcs $\gamma_1,\dots,\gamma_{n-1}$ which for small

382: $\mu$ are close to the critical points of $p$, and

383: that $d\nu(t)=\rho(t)dt$ for some {\em density 

384: of eigenvalues} $\rho$ defined on

385: the arcs: namely, the measure $\nu(U)$ of a

386: set $U$ intersecting one of the arcs, say $\gamma_j$,

387: in a piece $\gamma_j\cap U$ is

388: \[

389: \nu(U)=\int_{\gamma_j\cap U}\rho(z)dz.

390: \]

391: Note that for the right-hand side to be defined we

392: need to fix an orientation on $\gamma_j$.

393:   Then $y(z)$ is a holomorphic function

394: outside on the complement of the arcs and the density

395: on $\gamma_j$ is related to the discontinuity of $y$:

396: \[

397: \rho(x)=\frac1{4\pi i\mu}(y(x^-)-y(x^+)),\qquad x\in\gamma_j.

398: \]

399: Here $y(x^+)$ ($y(x^-)$)

400:  denotes the limit of $y(z)$ as $z$

401: tends to $x$ from the left (from the right)

402:  of the oriented curve $\gamma_j$.

403: 

404: {}From this information we deduce that the arcs 

405: $\gamma_j$ connect pairs of zeros of $y$,

406: the branch points of the hyperelliptic

407: curve $y^2=p'(z)^2-4\mu f(z)$ of genus $n-2$. 

408: The measure

409: of the $j$th arc is then the period

410: \begin{equation}\label{e-pe}

411: n_j=\frac1{4\pi i\mu}\int_{A_j}y(z) dz ,\qquad j=1,\dots, n-1,

412: \end{equation}

413: over a cycle $A_j$ enclosing the pair of branch

414: points in counterclockwise direction. It follows from

415: the condition on the leading coefficient of $f$ that

416: $\sum n_j=1$.

417: We also note that

418:  since $y(x^-)=-y(x^+)$, we have the formula

419: $\rho(x)=y(x^-)/2\pi i\mu$.

420: 

421: \subsection{The shape of the arcs}

422: There remains to determine the precise form

423: of the arcs and the coefficients in $f$ as functions

424: of the filling fractions $n_j$.

425: First of all, the relation between the $n-2$ 

426: free complex coefficients $b_0,\dots, b_{n-3}$ of $f(z)$

427: (recall that

428: the leading coefficient is fixed to be $b_{n-2}=

429: na_n$) and the periods $n_j$ ($1\leq j\leq n-1$)

430: subject to $\sum n_j=1$ is, locally around any

431: point where the branch points are distinct,

432:  a holomorphic diffeomorphism,

433: since the Jacobian matrix

434: \[

435: \frac{\partial n_j}{\partial b_{k-1}}=-\frac 1{4\pi i\mu}

436: \int_{A_j}\frac {z^{k-1}}ydz, \qquad j,k=1,\dots,n-2.

437: \]

438: is non-degenerate, being the matrix of $a$-periods

439: of a basis of Abelian differentials on a smooth

440: curve. It follows that

441: locally there exists a real $(n-1)$-dimensional submanifold

442: in the complex space $\mathbb C^n$ of coefficients $b_j$

443: which maps to real positive $n_j$.

444: 

445: The condition that fixes the shape of the arcs is

446: the reality and positivity condition for the

447: density: if $t\mapsto z(t)$ is a parametrization

448: of the arc $\gamma_j$ respecting its orientation,

449: the condition is

450: \begin{equation}\label{e-re}

451: \rho(z(t))\frac {dz(t)}{dt}\geq 0.

452: \end{equation}

453: Using $\rho(z)=y(z^-)/2\pi i\mu$, we may parametrize

454: the arcs (away from the endpoints)

455: to be solutions of the differential equation

456: \[\dot z(t)=i\,\overline{ y(z(t)^-)}

457: ,\]

458: connecting branch points. Alternatively, arcs

459: may be described as level lines of a function: 

460: introduce the hyperelliptic integral

461: \[

462: F(z)=\frac 1{4\pi i\mu}\int_{z_0}^z y(z) dz.

463: \]

464: It is a holomorphic many-valued function on the complement

465: of the support of the measure. As we go around 

466: an arc $\gamma_j$, $F(z)$ increases by $n_j$ so

467: $\mathrm{Im}\,F(z)$ is single valued.

468: The measure of a piece between

469: two points $x_1$, $x_2$ on a curve $\gamma_j$ is  $F(x_2^-)-

470: F(x_1^-)$ which is real.

471: Thus the arcs $\gamma_j$ are level lines of the imaginary

472: part of $F(z)$. 

473: Around a branch point $z_0$ which is a simple

474: zero of $y^2$,

475: $F(z)\sim \mathrm{const}+ (z-z_0)^{3/2}$. Therefore

476: there are three smooth 

477: level lines of $F$ emerging from every

478: simple branch points. In the most general

479: situation the support of the measure may then be

480: a graph consisting of level lines of $\mathrm{Im}\,F$

481: joining branch points. 

482: A more precise description is possible in the case

483: of small $\mu$ to which we turn.

484: 

485: \subsection{Small coupling}

486: For small $\mu>0$ we claim that the arcs and the

487: density of eigenvalues are determined completely

488: by the filling fractions $n_j$ through \eqref{e-hy}

489: and \eqref{e-re}. 

490: To show this, notice first that as $\mu\to 0$,

491: pairs of branch points $z_j',z_j''$ converge

492: to the critical points $z_j$ and the periods $n_j$

493: (eq.~\eqref{e-pe}),

494: regarded as functions of $\mu$ and the coefficients

495: $b_j$ of $f$ are holomorphic at $\mu=0$. We have

496: \[

497: \lim_{\mu\to 0} n_j(\mu,b_0,\dots,b_{n-3})=

498: \frac{f(z_j)}{4\,p''(z_j)}.

499: \]

500: Since there is a bijective holomorphic

501: correspondence between

502: values $f(z_j)$ at the distinct points $z_j$ and

503: coefficients $b_j$ of $f$, we have at $\mu=0$,

504: and by analyticity also for small $\mu$, a

505: biholomorphic map $(b_0,\dots,b_{n-3})\to(n_1,\dots,n_{n-1})$,

506:  $\sum n_j=1$. In particular, we can invert this

507: map and find a unique $f$ for each set of $n_j\geq0$,

508: such that $\sum n_j=1$.

509: It remains to show that for all small $\mu>0$ there

510: is a level line $\gamma_j$ of $\mathrm{Im}\,F(z)$

511: connecting $z_j'$ to $z_j''$. As $\mu\to 0$,

512: $\mu F(z)\to 

513: -\frac1{4\pi i}p(z)+\mathrm{const}$. The level lines

514: of $\mu\,\mathrm{Im}\,F(z)$ at $\mu=0$ 

515: are thus the level lines of $\mathrm{Re}\,p(z)$.

516: In the neighborhood of a non-degenerate

517: critical point $z_j$ they

518: look like the left picture in Fig.~\ref{f-1}: on any 

519: small circle around $z_j$ each value is taken on at

520: most four times. 

521: \begin{figure}[h]

522: \begin{picture}(150,130)(60,0)\scalebox{0.55}

523: {\includegraphics{HMI.eps}}

524: \end{picture}

525: \caption{Level lines of $\mathrm{Im}\,F$ for $\mu=0$

526: (left) and for $\mu>0$ (right)}\label{f-1}

527: \end{figure}

528: For positive small $\mu$ the critical point $z_j$ splits 

529: into two branch points $z_j'$, $z_j''$ from

530: each of which three level lines emerge. The

531: condition that the period $n_j$ is real implies that 

532: $\mathrm{Im}\,F(z_j')=\mathrm{Im}\,F(z_j'')$. The

533: function $g(z)=\mathrm{Im}\,(F(z)-F(z_j'))$ is defined

534: up to a sign in a neighborhood

535: of $z_j',z_j''$, so that its zero level line

536:  is uniquely defined. It

537: follows that there is a level line $\gamma_j$ of $\mathrm{Im}\,F$, namely the zero set of $g$, joining $z_j'$

538: to $z_j''$ as on the right picture in Fig.~\ref{f-1}:

539: if none of the level lines emerging from $z_j'$ and

540: $z_j''$ were to join, $g(z)$ would take

541: the value zero six times on any small circle 

542: encircling $z_j',z_j''$, which cannot be, as this

543: does not happen at $\mu=0$. Also, a level line cannot go

544: from a point $z_j'$ or $z_j''$ to itself as the 

545: real part of $F$ is monotonic along level lines.

546: 

547: We conclude that for any small $\mu>0$, and any

548: given filling fractions $n_1,\dots,n_{n-1}\geq 0$

549: summing up to 1, there is a unique polynomial

550: $f(z)=na_nz^{n-2}+b_{n-3}z^{n-3}+\dots+b_0$, so that

551: the curve $y^2=p'(z)-4\mu f(z)$ has $a$-periods $n_j$.

552: The zeros of $y$ are connected in pairs by arcs

553: $\gamma_j$ obeying the reality condition \eqref{e-re}.

554: According to the discussion above, these arcs are

555: are the support of the measure and the density of

556: eigenvalue is $(2\pi i\mu)^{-1}y(x^-)$, $x\in\gamma_j$.

557: 

558: \section{Concluding remarks}

559: We have given a non-perturbative definition of 

560: the matrix integrals that in the large $N$ limit

561: give the superpotentials considered in \cite{Dij1,Dij3}.

562: We considered the case of a generic polynomial

563: potential with complex coefficients. For

564: small 't Hooft coupling $\mu$, the density

565: of eigenvalues was shown to be given by arcs

566: connecting pairs of branch points of a

567: hyperelliptic curve. The shape of the arcs is

568: uniquely determined by a reality condition.

569: For larger $\mu$ or for potentials with degenerate

570: critical points, one expect the arcs to combine into

571: graphs in the complex plane. It would be interesting

572: to understand what kind of graphs can arise in this

573: way.

574: 

575: 

576: 

577: 

578: \begin{thebibliography}{99} 

579: 

580: \bibitem{Dij1}R.~Dijkgraaf, C.~Vafa, ``Matrix Models, Topological Strings, and Supersymmetric Gauge Theories'' (arXiv:hep-th//0206255),

581: Nucl.Phys. B644 (2002) 3--20

582: %\bibitem{Dij2}R.~Dijkgraaf, C.~Vafa, ``On Geometry and Matrix Mo%dels'' (arXiv:hep-th/0207106), Nucl.Phys. B644 (2002) 21-39

583: 

584: \bibitem{Dij3}R.~Dijkgraaf, C.~Vafa, ``A Perturbative Window into Non-Perturbative Physics'' (arXiv:hep-th/0208048), 2002

585: \bibitem{D} F. David,

586: ``Non-Perturbative Effects in Matrix Models and Vacua of Two Dimensional Gravity'' Phys.Lett. B302 (1993) 403--410

587: \bibitem{La}

588: C. I. Lazaroiu

589: ``Holomorphic matrix models,''

590: hep-th/0303008, 

591: JHEP 0305 (2003) 044 

592: \bibitem{bru}N.G.~de~Bruijn, ``Asymptotic Methods in Analysis,'' North-Holland Publishing, Amsterdam, 1958

593: \bibitem{erd}A.~Erd\' elyi, ``Asymptotic Expansions,'' Dover Publications, New York, 1956

594: \bibitem{W} S. Wadia,

595: ``Dyson-Schwinger equation approach to the large-N limit:

596: model systems and string representation of Yang--Mills theory''

597: Phys.~Rev.~D 24 (1981) 970--978

598: \bibitem{M} A. A. Migdal, ``Loop equations and 1/N expansion''

599: Phys.~Rep.~102 (1983) 199--290

600: \bibitem{BIZ}D.~Bessis, C.~Itzykson, J.~B.~Zuber, ``Quantum Field Theory Techniques in Graphical Enumeration,'' Advances in Applied Mathematics, 1:109--157, 1980.

601: 

602: 

603: \end{thebibliography}

604: 

605: 

606: 

607: \end{document}

608: 

609: 

610: 

611: 

612: 

613: