1: \documentclass[12pt]{article}

2: \usepackage{epsf}

3: %\usepackage{graphicx}

4: \usepackage{a4}

5: \usepackage{amsfonts}		% open math symbols

6: \usepackage{cite}		% collapse citations

7: 

8: \def\hybrid{\topmargin 0pt

9:         \oddsidemargin 0pt

10:         \headheight 0pt \headsep 0pt

11: %       \textwidth 6.5in        % US paper

12: %       \textheight 9in         % US paper

13:         \textwidth 6.25in       % A4 paper

14:         \textheight 9.5in       % A4 paper

15:         \marginparwidth .875in

16:         \parskip 5pt plus 1pt   \jot = 1.5ex}

17: 

18: %\def\baselinestretch{1.5}

19: \catcode`\@=11

20: \def\marginnote#1{}

21: 

22: \newcount\hour

23: \newcount\minute

24: \newtoks\amorpm

25: \hour=\time\divide\hour by60

26: \minute=\time{\multiply\hour by60 \global\advance\minute by-\hour}

27: \edef\standardtime{{\ifnum\hour<12 \global\amorpm={am}%

28:         \else\global\amorpm={pm}\advance\hour by-12 \fi

29:         \ifnum\hour=0 \hour=12 \fi

30:         \number\hour:\ifnum\minute<10 0\fi\number\minute\the\amorpm}}

31: \edef\militarytime{\number\hour:\ifnum\minute<10 0\fi\number\minute}

32: 

33: \def\draftlabel#1{{\@bsphack\if@filesw {\let\thepage\relax

34:    \xdef\@gtempa{\write\@auxout{\string

35:       \newlabel{#1}{{\@currentlabel}{\thepage}}}}}\@gtempa

36:    \if@nobreak \ifvmode\nobreak\fi\fi\fi\@esphack}

37:         \gdef\@eqnlabel{#1}}

38: \def\@eqnlabel{}

39: \def\@vacuum{}

40: \def\draftmarginnote#1{\marginpar{\raggedright\scriptsize\tt#1}}

41: 

42: \def\draft{\oddsidemargin -.5truein

43:         \def\@oddfoot{\sl preliminary draft \hfil

44:         \rm\thepage\hfil\sl\today\quad\militarytime}

45:         \let\@evenfoot\@oddfoot \overfullrule 3pt

46:         \let\label=\draftlabel

47:         \let\marginnote=\draftmarginnote

48:    \def\@eqnnum{(\theequation)\rlap{\kern\marginparsep\tt\@eqnlabel}%

49: \global\let\@eqnlabel\@vacuum}  }

50: 

51: %       This causes equations to be numbered by section

52: 

53: \def\numberbysection{\@addtoreset{equation}{section}

54:         \def\theequation{\thesection.\arabic{equation}}}

55: 

56: \def\titlepage{\@restonecolfalse\if@twocolumn\@restonecoltrue\onecolumn

57:      \else \newpage \fi \thispagestyle{empty}\c@page\z@

58:         \def\thefootnote{\fnsymbol{footnote}}

59: 	\setcounter{page}{0} }

60: \def\endtitlepage{\if@restonecol\twocolumn \else  \fi

61:         \def\thefootnote{\arabic{footnote}}

62:         \setcounter{footnote}{0}}  %\c@footnote\z@ }

63: 

64: 

65: 

66: \catcode`@=12

67: \relax

68: \def\etc{\hbox{\it etc.}}

69: \def\ie{\hbox{\it i.e.}}

70: \def\nn{\nonumber}

71: \def\beq{\begin{equation}}

72: \def\eeq{\end{equation}}

73: \def\bea{\begin{eqnarray}}

74: \def\eea{\end{eqnarray}}

75: \relax

76: \numberbysection

77: \hybrid

78: %\draft

79: 

80: 

81: \begin{document}

82: \begin{titlepage}

83: \begin{center}

84: {\large\bf

85: % On the 2-D random-bond Ising model.

86: Strong disorder fixed points in the two-dimensional random-bond Ising model

87: }\\[.3in] 

88: 

89: {\bf M.\ Picco$^{1}$, A.\ Honecker$^{2}$ and P.\ Pujol$^{3}$}\\

90:         % {\bf (1)}

91: 	$^1$ {\it LPTHE\/}\footnote[1]{Unit\'e mixte de recherche du CNRS 

92: UMR 7589.}, % \\

93:         {\it  Universit\'e Pierre et Marie Curie-Paris6 and \\

94:               Universit\'e Denis Diderot-Paris7\\

95:               Bo\^{\i}te 126, Tour 24-25, 5 \`eme \'etage, \\

96:               4 place Jussieu,

97:               F-75252 Paris CEDEX 05, France, \\

98:     e-mail: {\tt picco@lpthe.jussieu.fr}. }\\

99:         % {\bf (2)}

100: 	$^2$ {\it Institut f\"ur Theoretische Physik, TU Braunschweig,\\

101:     Mendelssohnstr.\ 3, 38106 Braunschweig, Germany, and \\

102:     Institut f\"ur Theoretische Physik, Universit\"at G\"ottingen, \\

103:     Friedrich-Hund-Platz 1, 37077 G\"ottingen, Germany, \\

104:     e-mail: {\tt a.honecker@tu-bs.de}. } \\

105:         %{\bf (3)}

106: 	$^3$ {\it Laboratoire de Physique\footnote[3]{Unit\'e mixte de 

107:              recherche du CNRS UMR 5672 

108:             associ\'ee \`a l'Ecole Normale Sup\'erieure de Lyon.}, % \\

109:              ENS Lyon, \\

110:              46 All\'ee d'Italie, 69364 Lyon C\'edex 07, France, \\

111:     e-mail: {\tt Pierre.pujol@ens-lyon.fr}. }\\

112: \end{center}

113: %\vskip .04in

114: \centerline{(Dated: June 10, 2006)}

115: \vskip .2in

116: \centerline{\bf ABSTRACT}

117: \begin{quotation}

118: %{\bf Abstract.}

119: The random-bond Ising model on the square lattice has several

120: disordered critical points, depending on the probability distribution

121: of the bonds. There are a finite-temperature multicritical point,

122: called Nishimori point, and a zero-temperature fixed point, for both a

123: binary distribution where the coupling constants take the values $\pm

124: J$ and a Gaussian disorder distribution. Inclusion of dilution in the

125: $\pm J$ distribution ($J=0$ for some bonds) gives rise to another

126: zero-temperature fixed point which can be identified with percolation

127: in the non-frustrated case ($J \ge 0$).  We study these fixed points

128: using numerical (transfer matrix) methods. We determine the location,

129: critical exponents, and central charge of the different fixed points

130: and study the spin-spin correlation functions. Our main findings are

131: the following:

132: (1) We confirm that the Nishimori point is universal with respect to the type of

133: disorder, \ie\ we obtain the same central charge and critical exponents

134: for the $\pm J$ and Gaussian distributions of disorder.

135: (2) The Nishimori point, the zero-temperature fixed point for the $\pm

136: J$ and Gaussian distributions of disorder, and the percolation point

137: in the diluted case all belong to mutually distinct universality

138: classes.

139: (3) The paramagnetic phase is re-entrant below the Nishimori point, \ie\

140: the zero-temperature fixed points are not located exactly below the

141: Nishimori point, neither for the $\pm J$ distribution, nor for the

142: Gaussian distribution.

143: \vskip 0.5cm 

144: \noindent

145: %\pacs

146: {PACS numbers: 75.50.Lk, 05.50.+q, 64.60.Fr}

147: %PACS numbers: 05.70.Jk, 64.60.Ak, 64.60.Fr

148: 

149: % PACS 1999 classification: http://www.aip.org/pacs/pacs99/pacscheme.html

150: 

151: \end{quotation}

152: \end{titlepage}

153: \section{Introduction}

154: 

155: The problem of disordered magnetic systems has attracted great

156: interest in the past years and many questions still remain

157: unanswered. An interesting problem is the universality class of second

158: order phase transitions in two-dimensional systems. The random-bond

159: Ising model (RBIM) is one of the simplest and best known of these

160: systems \cite{DDSL,N,ON,LG}, but exhibits rich enough behaviour to

161: give a general understanding of the problem. Further interest in the

162: RBIM stems from analogies with the quantum Hall transition

163: \cite{CF,GRL,CRKHAL,MENMD06} and applications in coding theory

164: \cite{Sourlas,NishCod,Iba,Preskill}.

165: 

166: For a small amount of randomness, the universality class of the RBIM

167: remains unchanged, presenting only logarithmic corrections in some

168: correlation functions \cite{DDSL}. For some particular kind of

169: randomness, namely dilution, one can show that another non-trivial

170: fixed point corresponds to a percolation universality class at zero

171: temperature, since it becomes a purely geometric problem of having a

172: thermodynamic number of spins within the same cluster.  It is

173: important to notice that in this case, all the randomly distributed

174: bonds are non-negative.

175: 

176: The situation is quite different if some negative bonds are allowed in

177: the probability distribution. For certain distributions with negative

178: bonds, Nishimori \cite{N, ON} has shown that some exact statements can

179: be made about physical quantities. There is in particular a line in

180: which the internal energy can be calculated exactly, known as the

181: Nishimori line. The interest of this line goes further, since it has

182: been shown that this line is invariant under renormalization group

183: (RG) transformations \cite{LG}. Since this lines crosses the

184: ferromagnetic to paramagnetic transition line, the intersection point,

185: known as the Nishimori point is a fixed point. Examples of probability

186: distributions satisfying the Nishimori condition are the Gaussian,

187: $\pm J$ binary and $0, \pm J$ with appropriately chosen weights (see

188: below). Despite recent analytical approaches \cite{GRL}, the exact

189: characterization of the universality class of this non-perturbative

190: fixed point is still unknown. The underlying field theory describing

191: this point is certainly a good representative of the disordered

192: fixed-point behaviour with very interesting and rich phenomenology

193: \cite{Z,GRLC}. In a previous letter \cite{HPP}, we have studied

194: numerically the critical exponents and central charge of the Nishimori

195: point in the $\pm J$ RBIM.  We found in particular that its

196: universality class does not correspond to the one of percolation, as

197: one could have imagined in view of earlier numerical investigations of

198: this model \cite{MM,SA,AQdS}.  This is derived, first, from numerical

199: estimates for the critical exponents and central charge which differ

200: significantly from those of percolation, and confirmed by an analysis

201: of higher moments of the correlation functions (which are all equal in

202: percolation and only in pairs for the Nishimori point, see \cite{HPP}

203: for details and later in this paper).

204: 

205: The conclusion of a different universality class was confirmed in

206: another numerical analysis by Merz and Chalker \cite{MC}. Thanks to a

207: mapping to a network model, these authors reached big lattice sizes

208: and high accuracy in the measurement of critical exponents. An

209: interesting remark made by these authors concerns the dual theory of

210: the RBIM\footnote{Because of the randomness introduced in the bonds,

211: the model is not self-dual as the pure Ising model.}, in which the

212: different moments of the disorder field acquire negative dimensions

213: \cite{MC2}. More recently, Nishimori and Nemoto \cite{NN} used a

214: generalized model with self-duality, and conjectured that the

215: projection onto the RBIM gives the phase boundary for this model. This

216: result permits in particular to locate the Nishimori point analytically.

217: The conjectured location is however outside the accuracy range

218: of the most recent numerical works on the $\pm J$ binary disorder case

219: \cite{HPP,MC} and the validity of this conjecture is certainly a very

220: interesting open issue.

221: 

222: %\begin{table}[hpt]

223: \begin{table}

224: \centerline{\begin{tabular}{|c|c|c|}

225: \hline

226: $p_c$               & Method            & Reference \\ \hline

227: \multicolumn{3}{|c|}{Nishimori point in the $\pm J$ model} \\ \hline

228: $0.111 \pm 0.002$   & Transfer matrix   & \cite{OzekiNishimori} \\

229: $0.114 \pm 0.003$   & Series expansion  & \cite{SA} \\

230: $0.1128 \pm 0.0008$ & Non-equilibrium   & \cite{OzekiIto} \\

231: $0.1095 \pm 0.0005$ & Transfer matrix   & \cite{AQdS} \\

232: $0.1094 \pm 0.0002$ & Transfer matrix   & % This work, see also

233:  \cite{HPP} \\

234: $0.1093 \pm 0.0002$ & Fermionic transfer matrix & \cite{MC} \\

235: $0.110028$          & Duality           & \cite{NN} \\

236: $\le 0.178203$      & Rigorous upper bound & \cite{MNN03} \\ \hline 

237: \multicolumn{3}{|c|}{$T=0$ critical point in the $\pm J$ model} \\ \hline

238: $ \sim 0.099$ & Series expansion & \cite{Grinstein} \\

239: $ 0.105 \pm 0.01$ & Matching algorithm & \cite{Freund} \\

240: $ 0.095<p_c < 0.108$ &

241: % Efficient

242: Matching algorithm & \cite{Bendish} \\

243: ${\textstyle 0.104 \pm 0.001 \atop \textstyle

244: 0.106 \pm 0.002}$ & Exact ground states & \cite{KaRi} \\

245: $0.115$             & Ground state enumeration & \cite{BGP} \\

246: $0.1031 \pm 0.0001$   & Exact ground states & \cite{Preskill} \\

247: $0.103 \pm 0.001$   & Exact ground states & \cite{Hartmann} \\ \hline

248: \end{tabular}

249: }

250: \caption{

251: Overview of estimates for $p_c$ at fixed points in the two-dimensional

252: $\pm J$ random-bond

253: Ising model. The first part of this table is for the Nishimori point,

254: the second part for the zero-temperature critical point.

255: \label{sumPc}

256: }

257: \end{table}

258: 

259: The pure and the Nishimori point, when present in the phase diagram,

260: are not the only non-trivial fixed points of the model: there is in

261: any case a zero-temperature fixed point. While Nishimori's results

262: state rigorously that this point cannot be located at a higher

263: concentration of ``impure'' bonds, analytical and numerical works on

264: the $\pm J$ model

265: \cite{Grinstein,Freund,Bendish,Preskill,Hartmann,KaRi,MB,BGP} tend to

266: conclude that it is located at a smaller density of impurities,

267: indicating a re-entrance of the ferromagnetic phase. The properties of

268: the zero-temperature point vary considerably with the kind of disorder

269: introduced. For a symmetric Gaussian distribution, it has been shown

270: \cite{NS} that the lowest energy configuration is unique (modulo the

271: ${\mathbb Z}_2$ symmetry) with probability one. One is tempted to

272: check the extension of this result to our zero-temperature point, and

273: check for example that, for a given configuration of the disorder, any

274: spin-spin correlation function is $1$ or $-1$. Then, all the odd

275: moments of the spin-spin correlation functions are equal, and all the

276: even moments are just equal to $1$, a result that is similar, but not

277: exactly identical to the percolation case.  The situation is much more

278: subtle for distributions like the $\pm J$ one, since frustration plays

279: a crucial role. For a generic configuration of disorder, the lowest

280: energy states are expected to be highly degenerate. The results

281: obtained by different techniques for the location of both, the

282: Nishimori point and the zero-temperature critical point are summarized

283: in Table~\ref{sumPc} for the $\pm J$ model. As can be seen from the

284: most recent results, the zero-temperature critical point seems to be

285: located at a concentration of ``impure'' bonds strictly smaller than

286: the one of the Nishimori point.  A schematic picture of the phase

287: diagram of the $\pm J$ case in the $p-T$ plane ($p$ being the number

288: of antiferromagnetic bonds) is shown in Fig.~\ref{pd}, where the

289: separation of the paramagnetic and ferromagnetic phases, as well as

290: the Nishimori line are drawn.  The location of the Nishimori point N

291: and the zero-temperature critical point are representative of the

292: results shown in Table~\ref{sumPc}. For the case of the Gaussian

293: disorder, a similar diagram may be obtained by replacing the parameter

294: $p$ with the variance of the distribution of disorder, $\sigma$,

295: although the shape of the Nishimori line is different.

296: 

297: \begin{figure}

298: \centerline{\epsfxsize=300pt{\epsffile{pd.eps}}}

299: \protect\caption[2]{\label{pd} Phase diagram of the two-dimensional

300: $\pm J$ random-bond Ising model.

301: The arrows represent the flow under the renormalization group.} 

302: \end{figure}

303: 

304: %Moreover, this point seems to share

305: %the same critical exponents obtained for the finite temperature

306: %Nishimori point.

307: 

308: Given a simple phase diagram as shown in Fig.~\ref{pd}, it is tempting

309: to trace the massless flow backwards from the pure Ising model at $p=0$

310: to the Nishimori point N \cite{MS95,CHMP}. However, we will argue in

311: this paper that in a generalized parameter space there are in fact several

312: fixed points on the massless surface with different universality classes

313: such that a backtracking procedure is in general ambiguous. Indeed,

314: the problems encountered in \cite{CHMP} were one of the original

315: motivations for our earlier numerical study of the Nishimori point \cite{HPP}

316: which we continue and expand here.

317: 

318: In this paper we provide an extensive numerical analysis of the

319: random-bond Ising model for three different kinds of probability distributions. 

320: We first consider in parallel the case of Gaussian and $\pm J$ distributions

321: for which we estimate the location of the

322: fixed point along the Nishimori line using domain-wall techniques.

323: We compute the free energy with a transfer matrix technique \cite{Night}

324: and obtain the central charge by analyzing the finite-size corrections.

325: We also compute the spin-spin

326: correlation functions on long strips to obtain the magnetic exponent. This and other

327: measurement to check consistency clearly show that the models with $\pm J$ and Gaussian 

328: distributions share the same universality class at their respective Nishimori points.

329: The compatibility of the results between the $\pm J$ and Gaussian cases

330: further elucidates the nature of the universality class of this point which

331: was recently argued to be different from the one of percolation \cite{HPP,MC}.

332: We next consider dilution, which can be modeled by a probability

333: distribution allowing the values of $\pm J$ and $0$ for the coupling

334: constants. In the

335: case of pure dilution (allowing only values $+J$ and $0$ for the

336: coupling constants), the other fixed point apart the one of the pure model is located 

337: at zero temperature and dilution $q_c = 1/2$ and simply corresponds to

338: bond percolation. 

339: The $\pm J$, $0$ distribution allows us to study the behaviour in the critical

340: line between the Nishimori point and the percolation point. 

341: In particular, by studying the value of the effective central charge for different strip widths 

342: and its evolution towards larger sizes we show that the percolation fixed 

343: point is repulsive along this critical line in favour of the

344: Nishimori point. 

345: The same technique is used to confirm that the Nishimori point is unstable with respect

346: to the pure Ising fixed point when moving on the critical line connecting these two points.

347: We finally address the problem of the zero-temperature fixed point.

348: Although for the case of the $\pm J$ distribution there is an extensive list of numerical

349: works indicating re-entrance of the ferromagnetic phase

350: (see Table~\ref{sumPc}), to our knowledge there was

351: no conclusive evidence of the same fact for the Gaussian distribution.

352: We show here clear 

353: evidence for the re-entrance of the phase also for the Gaussian distribution.  

354: We compute also

355: different moments of spin-spin correlation functions and the magnetic exponent. Our results

356: in the magnetic sector clearly show that the universality class of this

357: zero-temperature critical point is once again

358: different from the one of the finite-temperature Nishimori point as well

359: as percolation, in contrast to what one could have thought considering previous numerical

360: results in two dimensions \cite{SA,KaRi}.   

361: 

362: \section{Some definitions}

363: 

364: In this section, we present some definitions which will be used

365: throughout this paper. We will consider two kinds of probability

366: distributions $P(J)$ for the bonds: a discrete distribution where the

367: coupling constants can take values $\pm 1$ and $0$, and a continuous Gaussian

368: distribution. In the case of a discrete distribution, two subclasses

369: can be considered, the first one allows only the

370: values $\pm 1$ while the second and more general one, allows also

371: the value $0$ for the coupling constants. This last case corresponds to

372: dilution. In any case we can imagine a phase diagram in which the

373: vertical axis is given by temperature and the horizontal axis by a parameter

374: representing the strength of the disorder (see Figure \ref{pd}). 

375: 

376: Let us define the Hamiltonian of the two-dimensional Ising model to be:

377: \beq

378: \label{HIsing}

379: H = - \sum_{\langle i,j \rangle} J_{i,j} \, \delta_{S(i),S(j)} \, ,

380: \eeq

381: were $S(i) = \pm 1$ and $\langle i,j \rangle$ means nearest neighbours on

382: a square lattice. The variables $J_{i,j}$ are random and independently

383: chosen with a probability distribution function $P(J)$. As usual, the Kronecker 

384: delta function is zero if the spins are different and one if they are equal. 

385: To establish the relation to other conventions for

386: the energy note that it can be expressed in terms of products of Ising

387: spins via $\delta_{S(i),S(j)} = \left(S(i)\,S(j)+1\right)/2$.

388: 

389: The Nishimori line is defined % in this phase diagram

390: by the condition \cite{N}:

391: %

392: \beq

393: P(-J)= \exp(- \beta J) \, P(J) \, ,

394: \label{nishcond}

395: \eeq

396: %

397: where $\beta = {1 \over k_{B} T}$ is the inverse temperature

398: (from now on we choose the convention $k_{B} =1$). 

399: Let us analyze the Nishimori condition for the different distributions

400: considered here

401: %

402: \begin{itemize}

403: \item $J=\pm 1$: The distribution is characterized by the concentration of

404: antiferromagnetic bonds $p$:

405: %

406: \beq

407: P(J)=(1-p)\,\delta(J-1)+p\,\delta(J+1) \; .

408: \eeq

409: %

410: Eq.~(\ref{nishcond}) gives the following condition for the Nishimori line

411: %

412: \beq

413: \beta=\ln{\left( 1-p \over p \right)} \; ,

414: \label{pmJnis}

415: \eeq

416: which is schematically depicted in Fig.~\ref{pd}. The line extends

417: from the (attractive) fixed

418: points given by $T=0$, $p=0$ and $T = \infty$, $p=1/2$ and crosses

419: the Para-Ferro transition

420: line at the critical concentration $p_c$. In the first part

421: of Table~\ref{sumPc} we summarize estimates for the location

422: of the critical point $p_c$.  With the exception of a conjectured duality

423: property \cite{NN} and a rigorous upper bound \cite{MNN03},

424: all the other results come from numerical simulations. 

425: %

426: \item Gaussian distribution:

427: %

428: \beq

429: \label{gaussdis}

430: P(J) = \sqrt{ 1 \over 2 \pi \sigma^2} \,

431: \exp{\left(-{(J-J_0)^2\over 2 \sigma^2}\right)} \; ,

432: \eeq

433: %

434: with the following condition

435: %

436: \beq

437: \beta= {J_0\over \sigma^2}

438: \eeq

439: %

440: for the Nishimori line. In the following, we will choose $J_0=1$

441: without any loss of generality. Thus, the distribution is

442: characterized by the parameter $\sigma$ along the Nishimori line which

443: extends from the (attractive) fixed points given by $T=0$, $\sigma=0$

444: and $T = \infty$, $\sigma=\infty$ and crosses the Para-Ferro transition

445: line at the value $\sigma_c$. The first numerical characterization of

446: the critical point was given in \cite{MM}, with a value of $\sigma_c

447: \sim 0.97$.

448: %

449: \begin{figure}

450: \centerline{\epsfxsize=300pt{\epsffile{pd2.eps}}}

451: \protect\caption[2]{\label{pdbin} Phase diagram of the diluted $\pm

452: J,~0$ random-bond Ising model. $p$ and $q$ are the concentration of

453: negative and zero bonds, respectively. The arrows represent the flow

454: under the renormalization group.

455: The dashed line represents the intersection between the Nishimori

456: surface and the Ferro-Para transition surface.} 

457: \end{figure}

458: %

459: \item Finally, the binary distribution can be generalized to include

460: dilution. In this case, we write:

461: %

462: \beq

463: P(J)=q \, \delta(J) + (1-q-p) \, \delta (J-1) + p \, \delta(J+1) \; .

464: \eeq

465: %

466: The case $q=0$ corresponds to the binary distribution discussed above,

467: while $p=0$ gives the ferromagnetic diluted model. In the diluted

468: model there is only a fixed point located at zero temperature and $q_c

469: = 1/2$, apart from the pure model fixed point \cite{YeSt79}.

470: This $T=0$ fixed point corresponds to percolation.

471: The reason for this is that in the absence of frustration

472: all the spins belonging to the same cluster must point to the same direction

473: at zero temperature. Then, whenever there is percolation, or a cluster containing a 

474: macroscopic number of spins, there is a macroscopic magnetization,

475: and this happens until the critical concentration of bonds $1/2$

476: (see for example \cite{YeSt79,StAh}).

477: 

478: The Nishimori surface in the $T$-$p$-$q$ space is now given by:

479: %

480: \beq

481: \label{nissurface}

482: \beta= \ln{\left( 1-p-q \over p \right)} \; .

483: \eeq

484: %

485: The intersection of this surface with the ferromagnetic-paramagnetic

486: transition surface gives a renormalization group invariant line

487: \cite{LG} whose end points are the percolation $q=1/2,$ $T=0$ fixed

488: point and the finite-temperature Nishimori point N of the purely

489: binary case as depicted in Fig.~\ref{pd}.

490: In the $T-p-q$ phase diagram, the location of the Nishimori point,

491: here called N', must be within the intersection line mentioned above

492: (the dashed line in Fig.~\ref{pdbin}).

493: In this sense, the point N depicted in Fig.~\ref{pd} is the representative

494: in the $T-p$ space of the more general location of the Nishimori point

495: denoted by N' in Fig.~\ref{pdbin}. 

496: \end{itemize}

497: 

498: \section{Domain-wall free energy}

499: 

500: \label{secDW}

501: 

502: In the following section we will discuss the domain-wall free energy

503: in a manner very similar to \cite{MM}.

504: For a strip of with $L$ the domain-wall free energy $d_L$ is defined as

505: %

506: \beq

507: d_L = L^2 \left(f_L^{(p)} - f_L^{(a)}\right) \, ,

508: \label{defDW}

509: \eeq

510: %

511: where $f_L^{(p)}$ is the free energy {\it per site} $f_L^{(p)} = {\ln

512: Z^{(p)} \over L M}$ of a strip of width $L$ and length $M$ with {\it

513: periodic} boundary conditions\footnote{Note that our sign conventions

514: differ from the standard ones.} and $f_L^{(a)} = {\ln Z^{(a)} \over L

515: M}$ the corresponding one with {\it antiperiodic} boundary conditions.

516: 

517: $d_L$ measures the free energy associated to a domain wall in the

518: system. We will first consider the $\pm J$ distribution of

519: disorder. For fixed parameters $p$ ($\beta$) in the disordered

520: (paramagnetic) phase, one should have $\lim_{L \to \infty} d_L \to 0$

521: while $d_L$ should diverge in the ordered phase ($d_L \to \infty$ as

522: $L \to \infty$). At the fixed point $p_c$ ($\beta_c$), $d_L$ should

523: converge quickly with $L$.  We can therefore use crossing points

524: between $d_{L_1}$ and $d_{L_2}$ to obtain finite-size estimates for

525: $p_c$.

526: 

527: Let us consider first the $\pm J$ distribution of disorder.

528: $f_L^{(p)}$ and $f_L^{(a)}$ are computed with the transfer matrix

529: technique (see, e.g., \cite{Night})

530: on strips of length $10^6$. Averages over up to $N \approx

531: 4000$ samples of such $L \times 10^6$ strips are taken in order to

532: average over randomness and to determine the statistical error.  It is

533: useful to fix the number of bonds on each $L \times 10^6$ strip to

534: approximate the chosen value of $p$ as accurately as possible: In our

535: implementation such a strip has $L \cdot (2 \cdot 10^6 -1)$ bonds out

536: of which we select the integer closest to $p \, L \cdot (2 \cdot 10^6

537: -1)$. We found that this method leads to 5 times smaller error bars for

538: $f_L^{(p)}$ than if one selects each bond {\it separately} at random with

539: probability $p$, {\it i.e.}\ for the same precision one needs 25 times

540: less samples with this method compared to generating each bond at

541: random with probability $p$ without constraining their total number.

542: %

543: For $d_L$ the improvement is even bigger and the error bars become 10 times

544: smaller -- or one needs 100 times less samples for the the same accuracy. 

545: %

546: \begin{figure}[ht]

547: \begin{center}

548: \epsfxsize=400pt{\epsffile{cross.ps}}

549: \end{center}

550: \caption{

551: Position $p$ of crossings between $d_{L_1}(p)$ and $d_{L_2}(p)$

552: on the Nishimori line of

553: the $\pm J$ random-bond Ising model with $L_2 \le 12$.

554: \label{figCross}

555: }

556: \end{figure}

557: 

558: Next we will present some details of the analysis leading to the results

559: presented in \cite{HPP} -- a very similar computation was also performed

560: in \cite{MC}. Fig.\ \ref{figCross} shows a doubly logarithmic plot

561: of the crossing point $p_c(L_1,L_2)$ between $d_{L_1}(p)$ and

562: $d_{L_2}(p)$, % determined from the new data.  anticipating already

563: using $p_c = 0.1094$ in the shift of the vertical axis.  It is known for the

564: pure Ising model that the finite-size corrections to $p_c$ scale

565: roughly $\sim L^{-2}$ with $L=(L_1+L_2)/2$ \cite{Sor}. In order to

566: extrapolate $p_c$ to the thermodynamic limit, one can therefore first

567: use a plot of finite-size estimates for $p_c$ as a function of

568: $L^{-2}$ and extrapolate to the vertical axis. A reasonable

569: extrapolation with a generous error bar is

570: %

571: \beq

572: p_c = 0.1093 \pm 0.0004 \, .

573: \label{pcrit}

574: \eeq           

575: %

576: Even if the correction is probably not of the form $1/L^2$ in the present

577: case, one can see in Fig.\ \ref{figCross} that the finite-size correction

578: to $p_c$ is well described by a power in $L$. For an improved extrapolation

579: we therefore use the following form for a fit:

580: %

581: \beq

582: p_c(L_1,L_2) = p_c + \alpha \left({L_1 + L_2 \over 2}\right)^{-\xi} \, .

583: \label{pcritFit}

584: \eeq

585: %

586: We then find \cite{HPP}

587: %

588: \beq

589: p_c = 0.1094 \pm 0.0002

590: \label{pcrit1}

591: \eeq

592: %

593: and an exponent

594: %

595: \beq

596: \xi = 1.5 \pm 0.3 \, .

597: \label{xiEst}

598: \eeq

599: %

600: The two estimates (\ref{pcrit}) and (\ref{pcrit1}) agree well with

601: each other -- the error bar of the second one is just a little smaller.

602: The exponent (\ref{xiEst}) cannot be determined very accurately.

603: 

604: \begin{figure}[ht]

605: \begin{center}

606: \epsfxsize=400pt{\epsffile{slopes.ps}}

607: \end{center}

608: \caption{

609: Slope $A(L)$ of the domain-wall free energy

610: for the $\pm J$ random-bond Ising model as a function of $1/L$.

611: \label{figSlopes}

612: }

613: \end{figure}

614: 

615: The estimates (\ref{pcrit}) and (\ref{pcrit1}) should be compared with

616: other estimates which are summarized in Table~\ref{sumPc}. At this

617: point comparison with the zero-temperature transition is meaningful

618: only if one assumes that the $T=0$ transition point is located

619: exactly below the Nishimori point. Our estimate agrees within error

620: bars with other transfer matrix computations

621: \cite{OzekiNishimori,AQdS,MC}.

622: %-- (\ref{pcrit1}) is actually very close to that of \cite{AQdS},

623: %just with a reduced error bar.

624: However, our estimate falls outside error bars of estimates obtained

625: by other methods: It is smaller than the estimates of

626: \cite{SA,OzekiIto} (in the latter case significantly).

627: % and only

628: %the estimates of \cite{KaRi} are smaller than ours (though it remains

629: %to be clarified to which extent the latter are relevant to the

630: %Nishimori point).

631: This discrepancy % with \cite{SA,OzekiIto}

632: becomes clearest if one notices that several of our finite-size

633: crossings have already reached the region $p \approx 0.1097$

634: (see also Fig.\ 2 of \cite{HPP}). If one

635: now assumes that these crossings decrease monotonically with

636: increasing $L$ (as the data in Fig.\ \ref{figCross} indeed does), one

637: does not need to actually carry out the extrapolation and estimate its

638: error in order to conclude that our extrapolated value for $p_c$ must

639: end up below the error windows of \cite{SA,OzekiIto}.

640: At this point it is particularly reassuring to observe that

641: our result is fully consistent with the one of \cite{MC} which

642: has been obtained using also substantially wider strips (up to $L=64$).

643: Furthermore, (\ref{pcrit1}) is in complete agreement with 

644: other quantities to be discussed in later sections. Thus, we

645: observe a significant difference between the analytic estimate

646: of \cite{NN} obtained from a duality argument and our value for $p_c$

647: as well as the one of Merz and Chalker \cite{MC}.

648: 

649: Now let us look at a first critical exponent, namely the

650: correlation length exponent $\nu$ along the Nishimori line. If one assumes

651: the scaling form

652: \beq

653: d_L(p-p_c) = d\left((p-p_c) L^{1/\nu} \right) \, ,

654: \label{dScal}

655: \eeq

656: one can expand around $p_c$ and finds

657: \beq

658: d_L(p - p_c) \sim \hbox{const.} - A(L)\, p

659: \label{dAsym}

660: \eeq

661: with

662: \beq

663: A(L) \sim L^{1/\nu} \, .

664: \label{nuFit}

665: \eeq

666: Then one can fit $d_L$ close to $p_c$ by a linear function and extract $A(L)$.

667: 

668: Fig.\ \ref{figSlopes} shows the values for $A(L)$ determined in this manner

669: on a doubly logarithmic scale\footnote{Our data for $d_L$ is in

670: perfect agreement with that of \cite{MC} where we overlap. However,

671: the windows in $p$ used for the present estimates of $A(L)$ differ from

672: those used in \cite{HPP} leading to slightly different results.

673: }. One can see that they follow indeed a power law.

674: Using (\ref{nuFit}) we extract

675: \beq

676: \nu = 1.48 \pm 0.03 \, ,

677: \label{nuVal}

678: \eeq

679: which amounts to a slight correction of the value $\nu = 1.33 \pm 0.03$

680: obtained in \cite{HPP}. The result (\ref{nuVal}) is now in excellent

681: agreement with the value $\nu = 1.50 \pm 0.03$

682: obtained by a fermionic transfer matrix \cite{MC}.

683: %Note that using a transfer matrix method McMillan \cite{MM} obtained

684: %$\nu = \lambda_1^{-1} \approx 1.58$ for the Gaussian random-bond Ising model.

685: The most recent results $\nu \approx 1.5$ do no longer agree

686: well with $\nu = 1.32 \pm 0.08$ obtained by high-temperature series \cite{SA}.

687: However, we already observed above that the series expansion method

688: does not yield a very accurate estimate for $p_c$ either. Furthermore,

689: the value $\nu = 4/3$ characteristic for percolation (see e.g.\ \cite{StAh})

690: now falls outside numerical errors and thus it

691: is possible to conclude that the Nishimori point is not in the universality

692: class of percolation already on the basis of the exponent $\nu$.

693: Note that the result (\ref{nuVal}) does not involve locating $p_c$ precisely

694: and should therefore be independent of errors which may have been made in the

695: location of $p_c$.

696: 

697: \begin{figure}

698: \begin{center}

699: \epsfxsize=400pt{\epsffile{crossg.ps}}

700: \end{center}

701: \protect\caption[2]{\label{crossg}Position $\sigma$ of crossings

702: between $d_{L_1}(\sigma)$ and $d_{L_2}(\sigma)$

703: on the Nishimori line of the Gaussian random-bond Ising model.

704: The figure also contains a linear fit to the large size data

705: and the value $\sigma=1/1.02177$ predicted by duality arguments

706: \cite{NN}.}

707: \end{figure}

708: 

709: Figures with scaling collapses of $d_L$ were presented in

710: \cite{HPP,HJPP} with $\nu = 1.33$ and in \cite{MC} with $\nu =

711: 1.50$. While they verify that $d_L$ obeys indeed the scaling form

712: (\ref{dScal}), the fact that reasonable collapses can be obtained for

713: different values shows that such a collapse is not a good criterion

714: for determining $\nu$.

715: 

716: \begin{figure}

717: \begin{center}

718: \epsfxsize=400pt{\epsffile{slopesg.ps}}

719: \caption{

720: Slope of the domain-wall free energy

721: of the Gaussian random-bond Ising model as a function of $1/L$.

722: \label{figSlopesg}

723: }

724: \end{center}

725: \end{figure}

726: 

727: \begin{figure}[t]

728: \begin{center}

729: \epsfxsize=400pt{\epsffile{margin.ps}}

730: \caption{Domain-wall free energy

731: of the Gaussian random-bond Ising model at $\sigma_c=0.97945$ as a function of

732: $L$. We also show a plot of the fit to the form (\ref{eqmargin}) with

733: $L \geq 7$.

734: \label{figmargin}

735: }

736: \end{center}

737: \end{figure}

738: 

739: We now present the same domain-wall free energy analysis for the

740: Gaussian distribution of disorder. Again, $f_L^{(p)}$ and $f_L^{(a)}$

741: are computed with the transfer matrix technique for small

742: sizes. For larger sizes ($L=15,18,21,24$), we employed a different

743: algorithm, developed by Merz and Chalker, which uses a mapping to a

744: network model, see \cite{MC} for details. For each point we

745: averaged over $10~000$ samples of size $L \times 10^6$.

746: In Fig.~\ref{crossg}, we show a plot of the crossing point $\sigma(L_1,L_2)$

747: between $d_{L_1}(\sigma)$ and $d_{L_2}(\sigma)$.  Contrary to the $\pm

748: J$ case where the crossing was converging quickly, here we see that

749: the situation is much more complex. For the smaller sizes with

750: $L_2=L_1+2$ and $2/(L_1+L_2) \geq 0.15$, it apparently converges to

751: $\sigma \simeq 0.9815$. For larger sizes, we see a different

752: behaviour. We explain this change of behaviour by the existence of an

753: irrelevant operator. If such an operator exists, we expect

754: that close to the critical point, the domain-wall free energy can be

755: expanded as \cite{MM,MC,Sor}:

756: %

757: \beq

758: \label{eqmargin}

759: d_L(\sigma) = a + b \, (\sigma-\sigma_c) \, L^{1/\nu} + c \, L^{-x} \; .  

760: \eeq 

761: %

762: Here we have neglected a further contribution coming from the thermal

763: exponent $\nu_T$ (see \cite{MC}) which is sub-leading compared to the

764: $(\sigma-\sigma_c)$ term. The last contribution corresponds to an

765: irrelevant operator, of dimension $2 +x$, associated to the Nishimori

766: point. In principle it should also be present in the scaling of the

767: Nishimori point with the $\pm J$ distribution although much bigger

768: sizes may be necessary for its observation. If we take only the points

769: with $2/(L_1+L_2) < 0.1$ then the data in Fig.\ \ref{crossg} seems to

770: converge linearly to a point that we determine by a fit to be

771: \begin{equation}

772: \sigma_c = 0.97945 \pm 0.00004 \, .

773: \label{sigmaCritVal}

774: \end{equation}

775: In Fig.\ \ref{crossg} we also show the plot corresponding to this

776: linear fit (for the larger sizes) as well as the value

777: $\sigma=1/1.02177=0.978694$ predicted by duality arguments \cite{NN}.

778: While the two values are close,

779: they are still not compatible from our analysis. If one changes the

780: parameters of the linear fit, allowing for smaller size data to be

781: taken in account, the smallest possible value that we get is

782: $\sigma=0.97939 \pm 0.00002$ which is still not compatible with the

783: duality argument value. The only other numerical estimate of

784: $\sigma_c$ that we are aware of is the one of McMillan \cite{MM},

785: $\sigma_c \simeq 0.93-0.97$. This result is not expected to be highly

786: accurate since it is obtained on small systems of linear sizes up to

787: $L=8$. 

788: 

789: 

790: 

791: Next we turn to the determination of $\nu$. Again, we compute it

792: by looking at the slope of the domain-wall free energy close to the

793: critical point. In Fig.\ \ref{figSlopesg}, we show the slope as a

794: function of the size. The slope is determined in the following way:

795: for each size, we computed $d_L(\sigma)$ for $\sigma=0.980$ and

796: $\sigma=0.982$. (For the sizes that we considered, the effective

797: critical value of $\sigma_c(L)$, determined as the crossing of the

798: domain-wall free energies, is always in this range, see Fig.\

799: \ref{crossg}). Thus the slope is determined directly from the

800: difference between these two values for each size. We also show in

801: this figure the best fit to the form $\simeq L^{1/\nu}$.

802: This fits perfectly the data for $L > 4$. We

803: obtain $\nu=1.52(3)$ which is very close to the result for the $\pm J$

804: disorder case.  Note that using a transfer matrix method McMillan

805: \cite{MM} obtained $\nu  \approx 1.58$ for the

806: Gaussian random-bond Ising model.

807: 

808: 

809: In Fig.\ \ref{figmargin}, we plot $d_L(\sigma)$ for the critical value

810: $\sigma_c=0.97945$ determined previously. One expects then that the

811: non-constant part comes entirely from the leading irrelevant

812: operator. A fit to the form (\ref{eqmargin}), also shown in

813: Fig.~\ref{figmargin}, gives excellent results if we remove the data

814: with $L\leq 6$. We obtain an exponent $x = 0.43(3)$.  Thus we predict

815: the existence of an irrelevant operator of dimension $2.43(3)$. We

816: observed previously that $\sigma_c(L)$ converges linearly (in $1/L$)

817: towards $\sigma_c(L\rightarrow +\infty)$, see Fig.\ \ref{crossg} . A

818: simple calculation starting from eq.(\ref{eqmargin}) shows that the

819: correction is of order $L^{-(1/\nu +x)}$ and the linear correction

820: corresponds to $x \simeq 0.33$. This is rather close to the

821: numerical result, the difference being easily explained by taking

822: into account further irrelevant operators. 

823: 

824: \section{Free energy and central charge}

825: 

826: \label{secC}

827: 

828: In this section, we use the free energy to

829: deduce the central charge. We obtain results for both the $\pm J$ and the

830: Gaussian distribution of disorder and compare these results to

831: check universality. 

832: 

833: As a byproduct of the measurements of the domain-wall free energy

834: presented in the previous section, one also obtains the free energy

835: $f_L^{(p)}$.  We will first discuss the $\pm J$ disorder case.

836: Averaging has been performed over sufficiently many samples to obtain

837: statistical errors of typically $\delta f_L^{(p)} < 4 \cdot 10^{-6}$.

838: Note that due to the aforementioned round-off error of one bond, $p$

839: has an error of about $10^{-7}$ on the strip sizes considered which is

840: not substantially below $\delta f_L^{(p)}$.

841: 

842: The (effective) central charge $c$ can be estimated\footnote{For a

843: non-unitary theory what we call `$c$' is in fact

844: $c_{\rm eff} = c - 12 \, \delta_{\min}$

845: where $c$ is the central term in the operator

846: product expansion of $T$ with itself, and $\delta_{\min}$ the smallest

847: dimension contributing to the free energy. See section \ref{secCexp}

848: for further details.} from this data since it appears as the universal

849: coefficient of the first finite-size correction \cite{central}

850: %

851: \beq

852: f_L^{(p)} = f_{\infty}^{(p)} + {c \, \pi \over 6 L^2} + \ldots

853: \label{cEst2}

854: \eeq

855: %

856: Note that the leading term $f_{\infty}^{(p)}$ is not universal.

857: The universal $1/L^2$ term originates from the

858: energy-momentum tensor $T$

859: which has scaling dimension 2. The next correction is expected to arise

860: from $T^2$

861: and should therefore give rise to a term of the form $L^{-4}$

862: %

863: \beq

864: f_L^{(p)} = f_{\infty}^{(p)} + {c \, \pi \over 6 L^2} + {d \over L^4} + \ldots

865: \label{cEst4}

866: \eeq

867: %

868: Estimates for $c$ are obtained by fitting the free energy data in

869: intervals $L_0 \le L \le L_{\rm max}$ to (\ref{cEst2}) or

870: (\ref{cEst4}). We have considered only intervals with $L_0 \le L_{\rm

871: max}-3$ (at least four-point fits).  For a given form of the fit and

872: interval of lattice sizes, we have chosen $c$ to lie in the center of

873: the resulting fits and adjusted the error to include all fits.

874: The estimates for $c$ obtained in this manner are shown in Fig.\

875: \ref{figC}. Fits to (\ref{cEst2}) for lattice sizes $5 \le L \le 9$

876: are denoted by diagonal crosses and those for $6 \le L \le 12$

877: by plusses. Fits to (\ref{cEst4}) were performed for $4 \le L_0

878: \le 6$ with $L_{\rm max}$ kept fixed at the largest available system

879: size. The corresponding estimates are denoted by boxes.

880: 

881: \begin{figure}[ht]

882: \begin{center}

883: %\leavevmode

884: \epsfxsize=400pt{\epsffile{c2a4.ps}}

885: \end{center}

886: \caption{

887: Estimates for the central charge $c$ along the $\pm J$ Nishimori line

888: (\ref{pmJnis}) in the

889: vicinity of the Nishimori point obtained from (\ref{cEst2}) for

890: $5 \le L \le 9$ (`$\times$'), $6 \le L \le 12$ (`$+$')

891: or (\ref{cEst4}) with $4 \le L_0 \le 6$ (boxes). The

892: grey shaded areas denote our final confidence intervals for $p_c$

893: (vertical) and the central charge (horizontal).

894: \label{figC}

895: }

896: \end{figure}

897: 

898: One notices that the estimates in Fig.\ \ref{figC} are almost

899: independent of $p$ in the range considered.  On the one hand, this

900: means that the estimates for $c$ are stable. On the other hand, this

901: implies that estimates for the central charge are not a good tool for

902: locating the critical point precisely for the $\pm J$ distribution of

903: disorder (the situation will be different for the Gaussian

904: distribution, see below).  Furthermore, one observes a slight trend of

905: the fits (\ref{cEst2}) obtained with only a $L^{-2}$ term to shift to

906: smaller values as the range of system sizes considered is increased

907: while the fits (\ref{cEst4}) with a $L^{-4}$ term included converge

908: rapidly with system size. Therefore, we quote as a final estimate with

909: a generous error bar \cite{HPP}

910: %

911: \beq

912: c = 0.464 \pm 0.004 \, ,

913: \label{cVal}

914: \eeq

915: %

916: which is shown by the grey shaded horizontal bar (the estimate

917: (\ref{pcrit1}) for the location of the critical point $p_c$ is shown

918: by the grey shaded vertical bar).  In any case, all estimates obtained

919: from (\ref{cEst2}) with $L_{\rm max} = 12$ and $L_0 \ge 6$ satisfy $c

920: < 0.469$.  Furthermore, also consideration of other possible

921: finite-size corrections to (\ref{cEst2}) yields results which are

922: consistent with (\ref{cVal}). Thus, we can safely exclude the value

923: $c= {5 \sqrt{3} \ln{2} \over 4 \pi} \approx 0.47769$ for percolation

924: in the Ising model \cite{JC} even if the absolute difference is not

925: big.

926: 

927: \begin{figure}

928: \begin{center}

929: %\leavevmode

930: \epsfxsize=400pt{\epsffile{Cgauss.ps}}

931: \end{center}

932: \protect\caption[2]{\label{cgaus} Central charge vs.\ $\sigma$ along

933: the Nishimori line for the Gaussian distribution of disorder, with $L_{\rm max}=8$.}

934: \end{figure}

935: 

936: We now turn to the Gaussian distribution of disorder. In this case, we

937: will argue that the central charge, along the Nishimori

938: line, gives a practical way of locating the multicritical point as the

939: point where the estimates of the central charge have a maximum.

940: We explained above that,

941: for the $\pm J$ distribution, the estimate of the central charge was

942: nearly independent of $p$ making it difficult to locate the

943: maximum. For the Gaussian distribution, the parameter $p$ is

944: replaced with the variance of the distribution $\sigma$, see

945: eq.~(\ref{gaussdis}). Thus, the situation is different since it is

946: possible to perform the simulations with the same disorder

947: configuration for each variance $\sigma$. More precisely, one defines

948: a bond $J$ in the following way: $J = J_0 + \sigma j$, $j$ being the

949: random part obtained from a random number generator which we choose to be

950: the same for different $\sigma$. As a consequence, most of the

951: measurements will be correlated and in particular this will be true

952: for the free energy and the deduced central charge (as was also the

953: case for the domain-wall free energy measured in the previous section).

954: 

955: \begin{figure}

956: \begin{center}

957: \epsfxsize=400pt{\epsffile{Cgauss3.ps}}

958: \end{center}

959: \protect\caption[2]{\label{cgausc} Central charge vs.\ $\sigma$ along

960: the Nishimori line for the Gaussian distribution of disorder close to the

961: maximum $\sigma_c \simeq 0.98$ and with $L_{\rm max}=10$.}

962: \end{figure}

963: 

964: \begin{figure}[t]

965: \begin{center}

966: \epsfxsize=400pt{\epsffile{Pccg.ps}}

967: \end{center}

968: \protect\caption[2]{\label{ccc} Central charge at the multicritical

969: fixed point on the Nishimori line for the $\pm J$ and Gaussian

970: distributions of disorder. We also show for comparison finite-size

971: estimates of the central charge for percolation. The horizontal line

972: shows the infinite system limit $c= {5 \sqrt{3} \ln{2} \over 4 \pi}$

973: for percolation on Ising clusters \cite{JC}.

974: }

975: \end{figure}

976: 

977: The central charge is shown in Fig.\ \ref{cgaus} where we present the

978: effective central charge obtained by a fit to eq.\ (\ref{cEst2}) using

979: free energies in the range $[L_0,L_{\rm max}=8]$ for increasing

980: $L_0$. For the Gaussian distribution, we cannot suppress disorder

981: fluctuations by fixing the number of bonds with a certain value

982: globally.  As a consequence, the data shown in Fig.~\ref{cgaus} which

983: has been extracted from runs of the same size as for the $\pm J$

984: disorder (10~000 samples of strips of sizes $L\times 10^6$) has much

985: bigger error bars. Still these larger error bars are partially

986: compensated by the fact that the measurements are strongly

987: sample-to-sample correlated, and thus the maximum is easy to identify:

988: we have a stable maximum at $\sigma \simeq 0.98$. In Fig.\

989: \ref{cgausc}, we show the measured central charge close to the maximum

990: with $L_{\rm max}=10$. An estimate of the maximum gives for all the

991: ranges of $L$ employed a value of $\sigma_c \simeq 0.9805 \pm 0.0005$,

992: with no measurable change of the central charge compared to $\sigma =

993: 0.98$. For the size that we consider here, the agreement is very good

994: with the $\sigma_c(L)$ obtained with the domain-wall measurements, see

995: Fig.\ \ref{crossg}. Additional runs were performed for the value

996: $\sigma = 0.98$, in particular for larger sizes, up to $L=12$. For

997: $\sigma=0.98$ the average is performed over at least 20~000 samples,

998: each sample being a strip of size $L \times 10^7$. We will employ in

999: the following this data to compute the central charge that we will

1000: compare to the $\pm J$ disorder case. In Fig.~\ref{ccc}, we compare

1001: the effective central charges obtained for the Gaussian and the $\pm

1002: J$ distributions of disorder. For the $\pm J$ distribution, we

1003: employed the data for $p=0.1095$. For comparison, we also show the

1004: central charge obtained for percolation on an Ising model. In all

1005: these cases, we show the central charge obtained by a fit to

1006: (\ref{cEst2}) for increasing $L_0$ while keeping the maximum size

1007: fixed to $L_{\rm max}=12$.  The agreement between the two types of

1008: disorder is excellent. We see that the central charges converge to

1009: $\simeq 0.465$ at the multicritical fixed point on the Nishimori line

1010: for both cases of disorder. We thus have strong evidence of the

1011: universality of the multicritical fixed point. Moreover, the

1012: asymptotic value is clearly different from the value of percolation,

1013: either the numerical one on a finite system or the infinite system

1014: limit also shown in Fig.~\ref{ccc}.

1015: 

1016: \section{Spin-spin correlation functions}

1017: 

1018: \label{secCor}

1019: 

1020: In the previous sections, we have already seen that the universality class

1021: of the multicritical point on the Nishimori line, for the two types of

1022: disorder considered, is different from the one of

1023: percolation. Still the central charge is very close and it is then

1024: useful to find further measurements to confirm this

1025: result. In this section, we present numerical results for the

1026: spin-spin correlation functions on the Nishimori line close to

1027: the Nishimori point.

1028: 

1029: A general correlation function $C_n$ between

1030: two points $\vec{r}_1$ and $\vec{r}_2$

1031: has the following power-law form with exponent $2\,x_n$

1032: in a two-dimensional conformal field theory on the infinite plane:

1033: %

1034: \beq

1035: C_n\left(\vec{r}_1, \vec{r}_2 \right) \propto

1036: \frac{1}{\left| \vec{r}_1 - \vec{r}_2\right|^{2\,x_n}} \, .

1037: \label{corInfPlane}

1038: \eeq

1039: In the transfer matrix computations we use long strips

1040: with periodic boundary conditions along the short directions which

1041: one can also interpret as a cylinder. Therefore, we

1042: consider the correlation function $C_n$ on the infinite

1043: cylinder of circumference $L$ with coordinates % $u, v$ with

1044: $u\in [1,L]$ and $v\in ]-\infty,+\infty[$. Applying a conformal mapping

1045: to (\ref{corInfPlane}), one infers the following behaviour on the 

1046: cylinder (see, e.g., chapter 11.2 of \cite{Cardy})

1047: \beq

1048: C_n\left((u_1,v_1) , (u_2,v_2) \right) \propto 

1049: {({2 \pi \over L})^{2 x_n}  \over 

1050: \left\{2 \cosh(2\pi (v_1-v_2)/L) -2 \cos(2 \pi (u_1-u_2)/ L)\right\}^{x_n} } \, .

1051: \label{fitsscf}

1052: \eeq

1053: %%

1054: %\beq

1055: %[\langle S(u_1,v_1) S(u_2,v_2)\rangle] \propto 

1056: %{({2 \pi \over L})^{2 x_n}  \over 

1057: %\left[2 \cosh(2\pi (v_1-v_2)/L) -2 \cos(2 \pi (u_1-u_2)/ L)\right]^{x_n} } \, .

1058: %\label{fitsscf}

1059: %\eeq

1060: %

1061: %Here $[\cdots]$ stands for the average over the disorder. 

1062: There are two particular cases that we will consider in the following: 

1063: 

1064: \begin{itemize}

1065: \item $v_1 = v_2$, $u=u_1-u_2$. The correlation function is measured

1066: for two points across the strip, separated by a distance $u$. The

1067: correlation function takes the form: 

1068: %

1069: \beq

1070: %[\langle S(0) \, S(u)\rangle]

1071: C_n(0,u) \propto

1072:     \left(\sin({\pi u\over L}) L\right)^{- 2 \, x_n} \; .

1073: \label{fitcf1}

1074: \eeq

1075: %

1076: One can then extract the exponent $2\,x_n$ by a direct fit of the

1077: measured correlation function, after averaging over the disorder, to

1078: this form. One disadvantage of this method is that we can use only

1079: points separated by distances $u \leq L/2$.

1080: 

1081: %\item If we consider the special case $u=L/2$ with again $v_1 = v_2$,

1082: %the correlation function simplifies further and takes the form 

1083: %%

1084: %\beq

1085: %%[\langle S(0) S(L/2)\rangle]

1086: %_n(0,L/2) \propto  L^{- 2\,x_n} \; ,

1087: %\label{fitcf2}

1088: %\eeq

1089: %which allows a direct determination of the exponent $2\,x_n$. 

1090: 

1091: \item % A third case is

1092: When one chooses $u_1=u_2$, $v=v_1 - v_2$ the

1093: correlation function takes the following form:

1094: %

1095: \beq

1096: %[\langle S(0) S(y)\rangle]

1097: C_n(0,v) \propto

1098:     \left(\sinh({\pi v \over L}) L\right)^{- 2 x_n} \; .

1099: \label{fitcf3}

1100: \eeq

1101: %

1102: Thus we consider two points separated along the strip with the same

1103: position across the strip, and it is possible to

1104: access long distances.

1105: % make fits

1106: %of numerical data with only points far apart, thus reducing the effect

1107: %of finite sizes.

1108: 

1109: \end{itemize}

1110: 

1111: We will apply these two forms to the $n$th moments of the

1112: spin-spin correlation function

1113: %

1114: %\beq

1115: $[\langle S(u_1,v_1) S(u_2,v_2)\rangle^n]$, % \, ,

1116: %\label{defsscf}

1117: %\eeq

1118: where $[\cdots]$ stands for the average over disorder. 

1119: The general result is then, for the two cases discussed above:

1120: %

1121: \begin{eqnarray}

1122: [\langle S(0) S(x)\rangle^n] &\propto &

1123:     \left(\sin({\pi x\over L}) L\right)^{- \eta_n} \; , \cr

1124: %[\langle S(0) S(L/2)\rangle^n]  &\propto&   L^{- \eta_n} \; , \cr

1125: [\langle S(0) S(y)\rangle^n]  &\propto& 

1126:     \left(\sinh({\pi y \over L}) L\right)^{- \eta_n} \; .

1127: \end{eqnarray}

1128: %

1129: Here, we have identified $\eta_n = 2 \,x_n$. All along the Nishimori line, the

1130: moments of these correlation functions are equal two by two \cite{N}, thus we

1131: have the general result that $\eta_{2k-1} = \eta_{2k}$. For a pure

1132: system, one has $\eta_n = n \times \eta$.  On the other hand, in the

1133: case of percolation over Ising clusters, it is easy to see that all

1134: moments of the spin-spin correlation functions are equal (and not only two

1135: by two). Then, if the Nishimori point would be in the percolation

1136: universality class, the exponents for the correlation functions

1137: should collapse to a unique value $\eta_n = \eta$ at

1138: the critical point.

1139: 

1140: As a first measurement for $\pm J$ disorder,

1141: we show in Fig.\ \ref{ss1} the effective

1142: magnetic exponent $\eta_1$ that we obtain from a fit to the form

1143: eq.\ (\ref{fitcf1}). The effective magnetic exponent $\eta_1$

1144: was computed for three points on the Nishimori line:

1145: at $p=0.105$, which is well inside the ferromagnetic phase and indeed,

1146: we see that the effective magnetic exponent tends to a small value;

1147: at $p=0.115$ which is well in the paramagnetic phase, which is

1148: confirmed by the fact that the effective magnetic exponent increases;

1149: at $p=0.1095 \simeq p_c$ where the effective magnetic exponent seems

1150: to converge to a value $\eta_1 \simeq 0.185$.

1151: %

1152: \begin{figure}

1153: \begin{center}

1154: %\leavevmode

1155: \epsfxsize=400pt{\epsffile{SS1.ps}}

1156: \end{center}

1157: \protect\caption[2]{\label{ss1}Effective magnetic exponent $\eta_1$ for

1158: the $\pm J$ RBIM with

1159: $p=0.105$, $p=0.1095$, $p=0.115$, the percolation model and the pure Ising

1160: model.} 

1161: \end{figure}

1162: %

1163: On this figure, we also show the corresponding magnetic exponent

1164: obtained from simulations of the pure Ising model and also of the

1165: percolation model, as well as the expected values for these two models

1166: in the infinite limit ($\eta = {1\over 4}$ and ${5 \over 24}$,

1167: respectively -- see, e.g., \cite{WK74} and \cite{StAh}, respectively).

1168: % \footnote{The latter exponents can be obtained from the identification of

1169: %critical percolation with the $q \to 1$ limit of the $q$ states Potts model, see for example

1170: %\cite{StAh}.}.

1171: These two measurements are presented in order to show

1172: what type of correction we can expect in such a measurement of the

1173: magnetic exponent. Indeed, we can see that this method gives rather

1174: accurate measurements for large $L$ (for $L=12$, the deviation is

1175: around $1\%$). From this, we can conclude that the value of $\eta_1$

1176: for $p=0.1095$ is significantly distinct from the one of percolation.

1177: This is an additional proof that the multicritical point on the Nishimori

1178: line is not in the universality class of percolation.

1179: 

1180: 

1181: \begin{figure}

1182: \begin{center}

1183: %\leavevmode

1184: \epsfxsize=400pt{\epsffile{SS1b.ps}}

1185: \end{center}

1186: \protect\caption[2]{\label{ss1b}Effective magnetic exponent $\eta_1$ for

1187: $p=0.108, \cdots, 0.111$, for the $\pm J$ disorder on the Nishimori line.} 

1188: \end{figure}

1189: 

1190: 

1191: Next, in Fig.\ \ref{ss1b}, we show the same quantity but only close to

1192: the critical point. In this figure, one can clearly distinguish a change

1193: of behaviour close to $p \simeq 0.1095$. For lower $p$, the exponent

1194: decreases, which is expected since we are in the ferromagnetic phase.

1195: On the contrary, for $p$ larger than $0.1095$, we observe that the

1196: magnetic exponent increases, as expected since we are in the

1197: paramagnetic phase.

1198: 

1199: Fig.\ \ref{ssFig} shows the moments of the spin-spin correlation

1200: function for $L=20$ and $p=0.1095$. There is a similar plot in

1201: \cite{HPP}, but the present data set is completely independent. The

1202: present data was obtained for a geometry which differs from

1203: \cite{HPP}, namely on $4001 \times 20$ strips with a globally fixed

1204: number of positive (negative) bonds while for that of \cite{HPP}, each

1205: bond was assigned a value independently. Here we have discarded the

1206: 1000 initial (and final) iterations before taking 101 measurements of

1207: the correlation functions every 20 iterations. The data in Fig.\

1208: \ref{ssFig} was obtained by averaging over 7623 such strips, resulting

1209: in statistical error bars which are much smaller than the size of the

1210: symbols. Despite these differences, the present results for $[\langle

1211: S(0) S(x)\rangle^n]$ agree with those of \cite{HPP} within error bars.

1212: 

1213: \begin{figure}[ht]

1214: \begin{center}

1215: %\leavevmode

1216: \epsfxsize=400pt{\epsffile{cor.ps}}

1217: \end{center}

1218: \smallskip

1219: \caption{Moments of the spin-spin correlation function

1220: for the $\pm J$ RBIM with

1221: $p=0.1095$ and $L=20$. We only show the odd moments.

1222: %: $n=1$ ($+$, red), $n=3$ ($\times$, green), $n=5$ ($*$, blue) and $n=7$

1223: %(open boxes, pink).

1224: Error bars are much smaller than the size of the symbols.

1225: \label{ssFig}

1226: }

1227: \end{figure}

1228: 

1229: A direct fit on a doubly logarithmic scale of the correlation functions for

1230: $p_c = 0.1095$ and $L=20$ to the form (\ref{fitcf1}) yields:

1231: %

1232: \bea

1233: \eta_{1}&=&\eta_{2}= 0.1848 \pm 0.0003 \, , \nonumber \\

1234: \eta_{3}&=&\eta_{4}= 0.2552 \pm 0.0009 \, , \nonumber \\

1235: \eta_{5}&=&\eta_{6}= 0.3004 \pm 0.0013 \, , \nonumber \\

1236: \eta_{7}&=&\eta_{8}= 0.3341 \pm 0.0016 \, .

1237: \label{valExp20}

1238: \eea

1239: %

1240: These estimates

1241: are consistent with those of \cite{HPP,QSnew} within error bars.

1242: 

1243: We now turn to the Gaussian distribution of disorder. In

1244: Fig.\ \ref{mgaus2}, we present the exponent $\eta_1$ obtained from a

1245: direct fit with eq.\ (\ref{fitcf1}) in function of $\sigma$. For that

1246: case, we have data for size up to $L=14$ and for each value of $L$ and

1247: $\sigma$ we average over 10~000 samples of geometry $L \times 200\,L$.

1248: We clearly see a crossing of the curves close to $\sigma = 0.98$,

1249: in agreement with the previous results. At the

1250: crossing point, we have $\eta_1 \simeq 0.18$, which is very close to

1251: the corresponding value for the $\pm J$ distribution of disorder.

1252: 

1253: 

1254: 

1255: \begin{figure}

1256: \begin{center}

1257: \epsfxsize=400pt{\epsffile{etag1.ps}}

1258: \end{center}

1259: \protect\caption[2]{\label{mgaus2} Effective magnetic exponent $\eta_1$

1260: vs.\ $L$ for the Gaussian distribution of disorder.}

1261: \end{figure}

1262: 

1263: Next, we present the exponents obtained, again for $L=14$, from a

1264: direct fit on a doubly logarithmic scale to the correlation functions for 

1265: $\sigma = 0.98$:

1266: %

1267: \bea

1268: \eta_{1}&=&\eta_{2}= 0.1818 \pm 0.0002 \, , \nonumber \\

1269: \eta_{3}&=&\eta_{4}= 0.2559 \pm 0.0002 \, , \nonumber \\

1270: \eta_{5}&=&\eta_{6}= 0.3041 \pm 0.0002 \, , \nonumber \\

1271: \eta_{7}&=&\eta_{8}= 0.3402 \pm 0.0002 \, .

1272: \label{valExp14}

1273: \eea

1274: %

1275: The value of these exponents is close to the ones for the $\pm J$

1276: distribution of disorder. Note that there is still a small difference,

1277: $\eta_1^{\pm J} \simeq 0.1848$ compared to $\eta_1^{\rm Gaussian} \simeq

1278: 0.1818$. This difference is due to the fact that the exponents are

1279: obtained only close to the critical points. For the $\pm J$ disorder

1280: case, the measurement is done at $p=0.1095$. From Fig.\ \ref{ss1b}, one

1281: can read off that this will imply a change of order $0.002$ on $\eta_1$

1282: if we take $p_c=0.1093$. Taking into account this correction, the

1283: correspondence of $\eta_1$ is nearly perfect between the two types of

1284: disorder, thus giving more support for the universality.

1285: 

1286: \begin{figure}

1287: \begin{center}

1288: \epsfxsize=400pt{\epsffile{PlogRMG.ps}}

1289: \end{center}

1290: \protect\caption[2]{\label{mgaus4} Effective magnetic exponent

1291: $\eta_1$ vs.\ $L$ for the Gaussian distribution of disorder with

1292: $\sigma=0.976,0.978,0.979,0.980,0.981,0.982,0.984$ (from bottom to

1293: top).}

1294: \end{figure}

1295:  

1296: In Fig.\ \ref{mgaus4}, we show the exponent $\eta_1$ obtained by an

1297: independent set of simulations on a geometry $L \times 100\,L$ and with

1298: the measurement of the correlation function {\it along} the strip. We

1299: simulated 10~000 samples for each size up to $L=12$ with this

1300: geometry. The exponent $\eta_1$ is obtained from a fit with

1301: eq.\ (\ref{fitcf3}). To perform the fit, we keep only the data for

1302: correlation functions with two points at a distance $y$ such that $10

1303: \leq y \leq 10\times L$. Thus one does not use the data for two

1304: operators very close, contrary to what is done while fitting with eq.\

1305: (\ref{fitcf1}) and we expect to reduce the finite-size corrections. In

1306: Fig.~\ref{mgaus4}, one sees that for the largest $L$, one obtains a

1307: constant exponent for $\sigma=0.98-0.981$ with $\eta_1 =

1308: 0.180-0.182$. Thus these results are in perfect agreement with the

1309: previous measurements on a different geometry.

1310: 

1311: \begin{figure}

1312: \begin{center}

1313: %\leavevmode

1314: \epsfxsize=400pt{\epsffile{Pnu.ps}}

1315: \end{center}

1316: \protect\caption[2]{\label{mgaus5} Thermal exponent $\nu_T$

1317: vs.\ $L$ for the $\pm J$ distribution of disorder.}

1318: \end{figure}

1319:  

1320: Finally, in Fig.\ \ref{mgaus5}, we present results for the correlation

1321: function of the energy operator $\epsilon_{ij}=J_{ij}S_i S_j$.

1322: A thermal exponent $\nu_T$ is associated to

1323: this operator, which corresponds to a perturbation in temperature,

1324: via the relation

1325: %

1326: \beq 

1327: \nu_T = {1\over (2 - x_T)} \; .  

1328: \eeq 

1329: %

1330: This exponent replaces the exponent $x_n$ in eq.\ (\ref{fitcf1}) if we

1331: replace the spin operator with the energy operator. In Fig.\

1332: \ref{mgaus5}, $\nu_T$ is shown for $p \simeq p_c$ and for increasing

1333: strip widths $L$. Finite-size corrections are very strong

1334: for $\nu_T$. We obtain a value of $\nu_T \simeq 3$ for $L=14,\ldots,16$,

1335: but it is not clear if we have reached a large enough size. With a

1336: different method, Merz and Chalker \cite{MC} obtained a value of

1337: $\nu_T = 4 \pm 0.5$ by measurements on system sizes up to $L=32$. We

1338: believe that the extrapolation of our results for larger sizes is

1339: compatible with this result.

1340: 

1341: \subsection{Central charge and magnetic exponents}

1342: 

1343: \label{secCexp}

1344: 

1345: We now return to the central charge. We mentioned at the beginning

1346: of section \ref{secC} that the measured central charge 

1347: is an effective quantity  which can be affected by operators with 

1348: negative dimensions. Indeed, if one considers a torus, \ie\ a strip of

1349: geometry $M \times L$ with periodic boundary conditions in both

1350: directions, one has the relation \cite{cardyr}

1351: %

1352: \beq 

1353: {Z \over Z_{\rm bulk}} = Q^{-c/12} \sum_{a,b} N_{a,b} Q^{\delta_a + n_b}

1354: \eeq

1355: %

1356: with 

1357: %

1358: \beq

1359: Q= e^{ -2 \pi {M \over L}} \; .

1360: \eeq

1361: Here, the index $a$ is associated to the primary operators $\phi_a$

1362: which appear in the transfer matrix, while the index $n_b$ is non zero

1363: (and positive) for the conformal descendants only. $N_{a,b}$ counts the 

1364: multiplicity of the descendants of the

1365: operator $\phi_a$ at the level $b$. 

1366: Next, we consider the free energy per spin:

1367: %

1368: \beq

1369: f(L,M) = {\ln{Z(M,L)} \over M L} = f_{\rm bulk} + {c \, \pi \over 6 L^2} +

1370: {1\over M L } \ln{\left[ \sum_{a,b} N_{a,b} Q^{\delta_a + n_b} \right]} \; .

1371: \label{eqLM}

1372: \eeq

1373: If all the operators which appear in the transfer matrix have a

1374: non-negative dimension, then in the limit $M \gg L$, the last term in 

1375: eq.~(\ref{eqLM}) can be dropped and one recovers eq.~(\ref{cEst2}).

1376: On the contrary, if an operator with 

1377: a negative dimension is present in the transfer matrix,  then for large $M$,

1378: the last term in eq.~(\ref{eqLM}) will be

1379: dominated by this operator. Let us call $\delta < 0$ the dimension

1380: associated to the operator with the lowest dimension. Then

1381: eq.~(\ref{eqLM}) will become, in the limit $M \gg L$:

1382: %

1383: \beq

1384: f(L,M) = f_{\rm bulk} + {\pi \over 6 L^2}

1385:  (c -12 \,\delta)  + \ldots

1386: \eeq

1387: %

1388: Thus, if a negative dimension operator appears in the transfer matrix,

1389: the measured central charge will be only an effective quantity

1390: $c_{\rm eff}=c-12 \, \delta > c$. If no negative dimension operators appear

1391: in the transfer matrix, then one has $\delta=0$ since the identity

1392: operator with zero dimension is always present. 

1393: 

1394: \begin{figure}

1395: \begin{center}

1396: \epsfxsize=400pt{\epsffile{Prccc.ps}}

1397: \end{center}

1398: \protect\caption[2]{\label{reccc} Fit of the free energy for the

1399: $\pm J$ RBIM with $p=0.1094$ to the form (\ref{estDelta1})

1400: in the range $1.5 \leq M/L \leq 5$ for $L=4, 6$ and $8$.}

1401: \end{figure}

1402: 

1403: It is known that negative exponents arise at the Nishimori point

1404: in certain moments of the correlation functions of the disorder operator

1405: \cite{MC2}. Still, these operators will not cause

1406: any problems in the present context since the

1407: disorder operator is not a local operator of our theory. 

1408: 

1409: We know of no direct way to determine if an operator with a negative dimension 

1410: is present in the transfer matrix. Still, one can make the following simple test. 

1411: In the pure Ising model, the lowest dimension corresponds to the magnetic operator. 

1412: Since one expects that the magnetic exponent is present and we know that its

1413: dimension is small, one can compare 

1414: it with the first dimension which appears in eq.~(\ref{eqLM}). Assuming 

1415: that there is no negative dimension operator in the transfer matrix, the last 

1416: part of eq.~(\ref{eqLM}) takes the form

1417: %

1418: \beq

1419: {1\over M L } \ln{\left( 1 + e^{(-2 \pi {M\over L} \delta_1)} + \ldots  \right)}

1420: \eeq

1421: with $\delta_1$ the lowest dimension. Thus, by computing 

1422: \beq

1423: \ln{\left( e^{ M L (f(L,M) - f(L,M\rightarrow +\infty))} -1 \right)} 

1424: \simeq -2 \pi {M\over L} \delta_1 + \ldots  \; ,

1425: \label{estDelta1}

1426: \eeq

1427: %

1428: one can estimate directly $\delta_1$. In Fig.~\ref{reccc}, we show

1429: this quantity for $L=4,6$ and $8$, as well as a fit to the data in the

1430: range $1.5 \leq M/L \leq 5$. These bounds are selected by imposing a

1431: good quality of the fit. The values extracted for $\delta_1$ are

1432: $0.108(1),0.104(1),0.102(1)$ for $L=4,6,8$, respectively. These values

1433: seem to converge to a value close to the one of the magnetic exponent

1434: $\eta_1/2 \simeq 0.0925$ such that it is reasonable to identify it with

1435: the magnetic operator. This suggests that the measured central

1436: charge is indeed the real central charge. 

1437: 

1438: \section{Other measurements}

1439: 

1440: In this section, we present measurements of other quantities, namely

1441: the Binder cumulant and the magnetic susceptibility. For each of

1442: these quantities, we do not expect to improve precision, the

1443: measurements are rather done in order to check the consistency

1444: of the previous measurements.

1445: 

1446: \subsection{Binder cumulant}

1447: We first present measurements of the magnetic Binder cumulant on the

1448: Nishimori line, in order to perform an independent measurement of the

1449: critical $p$ for the $\pm J$ RBIM.

1450: The magnetic Binder cumulant is defined as follows

1451: in terms of the moments of the magnetization $m$ \cite{Binder81,LB00}:

1452: %

1453: \beq

1454: B(L)  =  {1\over 2} \left(3 - { \left[\langle m^4\rangle \right]\over

1455: \left[\langle m^2\rangle\right]^2}\right) \; .

1456: \eeq

1457: We note that on the Nishimori line the magnetic Binder

1458: cumulant is identical with the overlap Binder cumulant which is

1459: usually employed in measurements for spin-glass models,

1460: since we have an equality of the first and second moment of the

1461: spin-spin correlation functions \cite{N}.

1462: 

1463: Simulations are performed on square lattices with periodic boundary

1464: conditions in both directions and with linear sizes in the range $L=3$

1465: up to $L=8$. We employ the transfer matrix to compute the

1466: partition function without and with a small magnetic field $h$ as

1467: well as with $2h$. Then we extract the second and fourth moment of

1468: the magnetization $m$ from the expansion

1469: %

1470: \beq 

1471: Z(h) = Z(h=0) \left(1 + {h^2 \over 2} \, \langle m^2 \rangle

1472:  + {h^4 \over 4!} \, \langle m^4\rangle + \cdots \right) \; , 

1473: \eeq 

1474: %

1475: and a similar expansion of $Z(2h)$. The terms $\langle m \rangle$ and

1476: $\langle m^3 \rangle$ do not appear in the expansion since they vanish

1477: at $h=0$ on a finite system. We used a value of $h \simeq 0.01 / L^2

1478: $ in our simulations. In order to reach a good precision, a large

1479: number of samples had to be simulated, typically one million samples

1480: for each size and value of disorder $p$. Fig.~\ref{b1} shows a

1481: plot of the Binder cumulant versus $p$. In this figure, we observe

1482: a crossing in the expected region, \ie\ $p \simeq 0.11$. Since

1483: the number of samples that we have to simulate is huge, it is

1484: difficult to reach sizes large enough to improve the previous estimate

1485: of the critical point on the Nishimori line. Thus the measurements

1486: that we show here should be regarded as a consistency test only.

1487: %

1488: \begin{figure}

1489: \begin{center}

1490: \epsfxsize=400pt{\epsffile{Binder1.ps}}

1491: \end{center}

1492: \protect\caption[2]{\label{b1} $B(L)$ vs.\ $p$ on the Nishimori line

1493: for the $\pm J$ disorder case. The inset shows $B(L)$

1494: vs.\ $(p-p_c) L^{1/\nu}$, with $p_c=0.1094$ and $\nu=3/2$.}

1495: \end{figure}

1496: 

1497: The inset of Fig.~\ref{b1} shows a plot of the Binder cumulant versus the

1498: rescaled variable, $(p-p_c)\,L^{1/\nu}$. Assuming the values that we

1499: obtained in section \ref{secDW}, $p_c=0.1094$ and $\nu=3/2$, we see that

1500: we have a good scaling behaviour of the Binder cumulant already for

1501: small lattices.

1502: 

1503: \begin{figure}

1504: \begin{center}

1505: \epsfxsize=400pt{\epsffile{Susnis5.ps}}

1506: \end{center}

1507: \protect\caption[2]{\label{sn5} Susceptibility exponent

1508: $\gamma/\nu$ vs.\ $p$ on

1509: the Nishimori line for the $\pm J$ disorder case. The inset

1510: shows $\chi (L) L^{-\gamma/\nu}$ vs.\ $(p-p_c)L^{1/\nu}$, with

1511: $p_c=0.1094$ and $\nu=3/2$.}

1512: \end{figure}

1513: 

1514: \subsection{Susceptibility}

1515: 

1516: \label{secSusc}

1517: 

1518: A second quantity of interest is the susceptibility $\chi$, which we measure

1519: using the transfer matrix. We compute the free energy, first without

1520: any applied magnetic, and next with a small magnetic field.

1521: $f_L(0)$ and $f_L(h)$ are determined with

1522: the same realization of disorder in order to decrease the

1523: fluctuations. The susceptibility is obtained by the following relation

1524: %

1525: \beq f_L(h)-f_L(0) 

1526: \simeq (\beta h)^2 \chi (L) \; .

1527: \eeq 

1528: %

1529: We need to choose a very small magnetic field $h$, typically $h \simeq

1530: 0.000001$, in order to ensure that we do not have a magnetic term.

1531: 

1532: We expect that the susceptibility is of the form $\chi (L) \simeq

1533: L^{\gamma/\nu}$ at a critical point. In Fig.~\ref{sn5}, we

1534: present the effective susceptibility exponent obtained by two-point

1535: fits for data with odd sizes. For each of these points, we made

1536: $10~000$ measurements on strips of sizes $L \times 10^5$. We see that the

1537: crossing converges towards a value of $p$ close to $0.109$. Moreover,

1538: we can see that the value of the susceptibility exponent seems to

1539: converge to a value close to $\gamma / \nu \simeq 1.82$. Thus one can

1540: also check that the hyperscaling relation ${\gamma/ \nu} + 2{\beta

1541: /\nu} = d$ is satisfied with $2 {\beta / \nu} = \eta_1 \simeq 0.18$.

1542: 

1543: Finally, we can also use the measured values of the susceptibility to

1544: perform a fit in the following form

1545: %

1546: \beq

1547: \label{fchi}

1548: \chi(L) \simeq L^{\gamma/\nu} a((p-p_c) L ^{1/\nu}) \; .

1549: \eeq

1550: %

1551: Since this fit involves three parameters, we obtain a

1552: large error bar on each of them. Keeping only the

1553: reasonable fits, one obtains good collapses of the data in

1554: the following range: $\gamma/\nu = 1.8 - 1.82, p_c=0.109 - 0.110, 1/\nu =

1555: 0.65 - 0.7$. In the inset of Fig.\ \ref{sn5} we present a plot of $\chi(L)

1556: L^{-\gamma/\nu}$ vs.\ the rescaled variable $(p-p_c) L ^{1/\nu}$ with the

1557: values $p_c = 0.1094$, $1/\nu = 2/3$ and $\gamma /\nu = 1.82$ with an

1558: excellent collapse of the data.

1559: 

1560: \section{Dilution}

1561: 

1562: In this section, we will consider the more general case of a binary

1563: distribution with dilution. We denote by $q$ the amount of dilution

1564: and by $p$ the amount of disorder (see section 2 for the definitions).

1565: The Nishimori line is now replaced by a surface in the $T-p-q$ space.

1566: For $p=0$, we expect only two critical points \cite{YeSt79}: an

1567: attractive point for $q=0$ (no dilution) and a repulsive point for

1568: $q=0.5$ which is a percolation fixed point (see figure \ref{pdbin})

1569: and is on the Nishimori surface.  Another fixed point, also on the

1570: Nishimori surface, is the fixed point on the Nishimori line determined

1571: previously for $q=0$. In this section we want to study the flow

1572: between these two fixed points on the Nishimori surface. In

1573: particular, obtaining a clear flow between these two fixed points will

1574: give further evidence that the multicritical point on the Nishimori

1575: line ({\it i.e.}\ without dilution) is not in the same universality

1576: class as percolation. Our measurements are carried out as follows: for

1577: a fixed dilution $q$, we perform simulations for varying $p$ and $T$

1578: in the Nishimori surface defined in eq.~(\ref{nissurface}) and look

1579: for a maximum in the effective central charge. Next we check the value

1580: of this maximum $c_{\rm eff}(q)\ {\rm vs.}\ q$ which is shown in Fig.\

1581: \ref{cdil0}, for central charges obtained numerically with a fit to

1582: the form eq.~(\ref{cEst2}) for the sizes $[4,5,6,7,8]$, $[5,6,7,8]$

1583: and $[6,7,8]$. As expected, $c_{\rm eff}(q)$ varies monotonically

1584: between $c_{\rm eff}(q=0)$ and $c_{\rm eff}(q=0.5)$ showing that there

1585: is no additional fixed point.

1586: 

1587: \begin{figure}

1588: \begin{center}

1589: %\leavevmode

1590: \epsfxsize=400pt{\epsffile{Pdil0.ps}}

1591: \end{center}

1592: \protect\caption[2]{\label{cdil0} Central charge vs.\

1593: dilution $q$ along the intersection of the $p-q-T$ Nishimori surface

1594: with the paramagnetic-ferromagnetic critical surface for

1595: the $\pm J$ disorder case.}

1596: \end{figure}

1597: 

1598: \begin{figure}

1599: \begin{center}

1600: %\leavevmode

1601: \epsfxsize=400pt{\epsffile{Pdil1.ps}}

1602: %\epsfxsize=200pt{\epsffile{Pdil2.ps}}

1603: \end{center}

1604: \protect\caption[2]{\label{cdil} $X$ vs.\

1605: dilution $q$ along the intersection of the $p-q-T$ Nishimori surface

1606: with the paramagnetic-ferromagnetic critical surface for

1607: the $\pm J$ disorder case.}

1608: \end{figure}

1609: 

1610: \begin{figure}[t]

1611: \begin{center}

1612: %\leavevmode

1613: \epsfxsize=400pt{\epsffile{PdilM.ps}}

1614: \end{center}

1615: \protect\caption[2]{\label{cdilM} Magnetic exponent $\eta$ vs.\

1616: dilution $q$ along the intersection of the $p-q-T$ Nishimori surface

1617: with the paramagnetic-ferromagnetic critical surface for

1618: the $\pm J$ disorder case.}

1619: \end{figure}

1620: 

1621: In practice, since $c_{\rm eff}$ changes on a very small range

1622: (between $0.47769$ for $q=0.5$ and $\simeq 0.464$ for $q=0$) with

1623: strong finite-size corrections, it is more convenient to consider the

1624: following quantity:

1625: %

1626: \beq

1627: \label{Xdef}

1628: X = {c_{\rm eff} - c_{\rm eff}(0) \over c_{\rm eff}(0.5)-c_{\rm

1629: eff}(0)} \; ,

1630: \eeq 

1631: %

1632: where for each range of sizes $L$ we consider the values obtained

1633: with the same corrections for $c_{\rm eff}(0)$ and $c_{\rm

1634: eff}(0.5)$. By construction, this quantity is equal to zero at $q=0$

1635: (no dilution) and is equal to 1 at $q=0.5$ (the percolation case).

1636: Fig.~\ref{cdil} shows the quantity $X$ as a function of dilution

1637: $q$, again for the sizes $[4,5,6,7,8]$, $[5,6,7,8]$ and $[6,7,8]$.

1638: Apart from the size of the errors, one does not observe significant

1639: differences. In all cases, one immediately sees that the central charge

1640: is dominated by the behaviour of the fixed point without dilution. If

1641: one starts at the percolation point, a small decrease in $q$ will

1642: appear as a jump towards the fixed point without dilution

1643: ($q=0$)\footnote{More precisely, critical properties at finite temperature

1644: are expected to be controlled by the fixed point N' in Fig.~\ref{pdbin}

1645: \cite{LG}. Fortunately, the crossover scale for the flow between N

1646: and N' appears to be so small that this difference can be neglected

1647: in practice.}. At $q\simeq 0.4$, $X$ is already indistinguishable

1648: from zero, indicating that the central charge is the

1649: same as the one for $q=0$. One expects that by increasing the sizes of

1650: the data in the fit for determining $c_{\rm eff}$, one should observe

1651: a crossover between the two fixed points $q=0.5$ and $q=0$ (such an

1652: effect will be shown in the next section for the flow between the

1653: Nishimori point and the fixed point of the pure Ising model). Here we

1654: cannot really observe this effect since the crossover is too fast.

1655: 

1656: Finally, Fig.~\ref{cdilM} shows the magnetic exponent vs.\ the

1657: dilution measured on long strips of width $L=8$. Here again, we

1658: observe the same effect. For $q=0.5$, the magnetic exponent is known

1659: exactly, it is $\eta = 5/24 \simeq 0.208333$ \cite{StAh}. As we decrease $q$,

1660: the magnetic exponent jumps to a value compatible with the one of the

1661: multicritical point on the Nishimori line $\eta \simeq 0.18$.

1662: 

1663: \section{Out of the Nishimori line}

1664: 

1665: In this section, we present some results off the Nishimori line. We

1666: will consider two cases separately. First we investigate the line which

1667: corresponds to the flow from the Nishimori point to the pure

1668: Ising model. This line corresponds to the ferromagnetic-paramagnetic

1669: transition line. We show in the next subsection that this line

1670: can be determined from the maximum of the central charge when varying

1671: temperature at fixed $p$. We show further that the flow on this line is

1672: from the Nishimori point towards the pure Ising model fixed

1673: point. Next we turn to the line which connects the

1674: Nishimori point with a fixed point at zero temperature. An

1675: important issue concerns the verticality or

1676: re-entrance of this line. By arguing that the nature of the Ferro-Para

1677: transition at the Nishimori point is of geometric origin, it was

1678: suggested \cite{NK,ON} that the transition should take place at the

1679: same concentration of impurities for any temperature below

1680: the Nishimori point, implying verticality of the line. However, recent

1681: numerical results for the $\pm J$ RBIM seem to advocate instead a

1682: re-entrance of the paramagnetic phase (see Table~\ref{sumPc}). The other

1683: important issue is the nature of the fixed point at zero temperature. It

1684: was argued in light of previous numerical results that the

1685: universality class of this point could be percolation \cite{SA,KaRi},

1686: although there is no obvious reasoning supporting this conclusion in

1687: contrast to the case of dilution.

1688: 

1689: \subsection{Flow from the Nishimori point to the pure Ising model}

1690: 

1691: First we study the line which separates the ferromagnetic

1692: phase from the paramagnetic phase between the

1693: Nishimori point and the pure Ising model. The phase boundary itself

1694: has already been determined with good accuracy \cite{MC}. Here

1695: we will use a computation of the central charge to follow this

1696: line in a similar way to what was done in the case of dilution.

1697: The motivation for studying this line is first to clearly

1698: see in which direction we flow (we expect towards the pure Ising model

1699: fixed point since this is a marginally attractive fixed point \cite{DDSL}).

1700: Furthermore, we want to check that there is no

1701: additional fixed point. In a similar study for the $3$-state Potts

1702: model \cite{JP}, an additional fixed point, predicted by perturbation

1703: theory \cite{Ludwig,DPP} was observed.

1704: 

1705: We start from the pure Ising model, with a small perturbation, say

1706: $p=0.01$ and $T \simeq T_c$, $T_c$ being the critical temperature of the

1707: pure Ising model. Next we vary $T$ and measure the central charge. For

1708: small $p$, these measurements are very simple to perform since we have

1709: only a weak disorder and moreover, measurements are performed at

1710: constant $p$. It is then easy to determine a maximum since the same

1711: configurations of disorder are employed for different $T$. Then we

1712: iterate the process for larger $p$ and follow the ferro-para line

1713: which is identified with the maximum of the central charge. The

1714: measurement is more complicated close to the Nishimori line because

1715: the transition line has a strong curvature. One needs to change simultaneously

1716: $p$ and $T$, thus we do not have any more correlated samples and

1717: it is much more time consuming to locate the maximum of the central charge

1718: (as is the case on the Nishimori line).

1719: Fig.~\ref{phd} shows the phase diagram obtained from

1720: the maximum of the central charge. Within error bars, our results agree

1721: with the phase diagram Fig.~7 determined in Ref.~\cite{MC} by a different

1722: method. In Fig.~\ref{ccfp} we plot the

1723: corresponding central charge obtained from three-point fits with

1724: $L=3,4,5$ and $L=4,5,6$. Since the difference of the central charge

1725: between the pure Ising model and the Nishimori point is very small,

1726: we employed a parameter $X$ defined as follows:

1727: %

1728: \beq

1729: X={c_{\rm eff}(p)-c_{\rm eff}(p=0.109) \over c_{\rm eff}(p=0)-c_{\rm eff}(p=0.109)} \; ,

1730: \label{label8.1}

1731: \eeq

1732: %

1733: similar to the one defined in the study of dilution, see

1734: eq.~(\ref{Xdef}). From this plot, we can see that by increasing the

1735: lattice size, the attractive fixed point is the pure Ising model

1736: (since the variation of the central charge increases, starting from the

1737: Nishimori point). This is consistent with a flow from the

1738: Nishimori point to the pure Ising model and moreover the absence of additional

1739: fixed points along this line.

1740: 

1741: \begin{figure}

1742: \begin{center}

1743: \epsfxsize=400pt{\epsffile{PD.ps}}

1744: \end{center}

1745: \protect\caption[2]{\label{phd} $T-p$ phase diagram for the

1746: $J=\pm 1$ disorder case.}

1747: \end{figure}

1748:  

1749: \begin{figure}

1750: \begin{center}

1751: \epsfxsize=400pt{\epsffile{PIN.ps}}

1752: \end{center}

1753: \protect\caption[2]{\label{ccfp} Variation of the effective central charge

1754: as characterized by the quantity $X$ (see (\ref{label8.1}))

1755: on the Ferro-Para line, for the $J=\pm 1$ disorder case.} 

1756: \end{figure}

1757: 

1758: \subsection{The zero-temperature fixed point} 

1759: 

1760: \label{secZeroT}

1761: 

1762: In this subsection, we study the zero-temperature fixed

1763: point. First we present results for the location of the fixed point

1764: at $T=0$. This was already determined with high precision for the $\pm J$

1765: disorder case by Wang et al.\ \cite{Preskill} and Amoruso and Hartmann

1766: \cite{Hartmann}. The result obtained by these groups is

1767: $p^0_c \simeq 0.103$, indicating a re-entrance of the paramagnetic

1768: phase. For comparison, we repeated these measurements, obtaining

1769: compatible results. We further considered the Gaussian case for which the

1770: same re-entrance effect is obtained.

1771: 

1772: The simulations were performed on square lattices with free boundary

1773: conditions along one direction

1774: and periodic/antiperiodic conditions in the other

1775: direction. We measure the difference of energy between the periodic

1776: case and the antiperiodic case

1777: %

1778: \beq

1779: \Delta E = E_p -E_a \; .

1780: \label{label8.2}

1781: \eeq

1782: %

1783: This difference of energy corresponds to the energy associated to a

1784: domain wall in the system. 

1785: We expect that after averaging over disorder, the domain-wall energy is

1786: characterized by some size exponent $\rho$

1787: %

1788: \beq

1789: [ \Delta E ] \propto L^\rho \; .

1790: \label{label8.3}

1791: \eeq

1792: %

1793: This quantity (as well as the one associated to the width of the

1794: distribution of domain-wall energies) was computed with high precision

1795: by Amoruso and Hartmann \cite{Hartmann} on very large lattices, up to

1796: $L=700$, by using a minimum-weight perfect matching algorithm. In

1797: Fig.~\ref{ztdwpm}, we present our results on smaller sizes (up to $L=100$)

1798: but with much larger statistics, while employing the same type of

1799: algorithm \cite{blossom4}. We simulated 1 million samples for each $p$

1800: and $L < 100$ and 0.5 million samples for $L=100$, compared to

1801: $30~000$ samples in \cite{Hartmann}. The reason for desiring even

1802: better precision

1803: is to be able to observe a crossover in the location of the fixed

1804: point, see the discussion of this point in the following

1805: subsection. In Fig.~\ref{ztdwpm}, we clearly distinguish a scaling in

1806: function of the size $L$ close to $p \simeq 0.103$ for large $L$, in

1807: perfect agreement with the previous results

1808: \cite{Preskill,Hartmann}. We also note that this power-law behaviour

1809: is apparent only for sizes $L > 20$. In Fig.~\ref{ztdwg}, we present

1810: the same quantity for the Gaussian distribution of disorder. Here

1811: again we observe a power law of $[\Delta E]$ with size $L$ close to $\sigma

1812: \simeq 0.97$ to be compared to the result (\ref{sigmaCritVal})

1813: $\sigma_c \simeq 0.97945$ on

1814: the Nishimori line. For the Gaussian disorder, we have data up to size

1815: $L=140$ and the number of samples is again 1 million for each value of

1816: $\sigma$ and $L$ except for $L=140$ where we have 0.5 million samples.

1817: 

1818: Thus in both cases, we observe a re-entrance of the paramagnetic phase. 

1819: 

1820: \begin{figure}

1821: \begin{center}

1822: \epsfxsize=400pt{\epsffile{PlotdwPMT0.ps}}

1823: \end{center}

1824: \protect\caption[2]{\label{ztdwpm} Domain-wall energy $[\Delta E]$ as a

1825: function of the system size $L$ for the $\pm J$ disorder at $T=0$ and

1826: for $p=0.100, \ldots, 0.105$.}

1827: \end{figure}

1828: 

1829: \begin{figure}

1830: \begin{center}

1831: \epsfxsize=400pt{\epsffile{PlotdwgaussianT0.ps}}

1832: %\epsfxsize=200pt{\epsffile{Plot2dwGaussianT0.ps}}

1833: \end{center}

1834: \protect\caption[2]{\label{ztdwg} Domain-wall energy $[\Delta E]$ as a

1835: function of the system size $L$ for the Gaussian disorder at $T=0$ and

1836: for $\sigma =0.96, \ldots, 0.98$.}

1837: \end{figure}

1838: 

1839: 

1840: \subsection{Magnetic exponent at zero temperature}

1841: 

1842: In order to characterize the fixed point at zero temperature

1843: we measure the magnetic exponent in the same manner as on the Nishimori

1844: line. Specifically, we use measurements of the spin-spin correlation

1845: functions which provide a direct estimate of the magnetic

1846: exponents and allow us to study higher moments.

1847: 

1848: Fig.\ \ref{mgaus01} shows the values of $\eta_1$, obtained from

1849: the measured spin-spin correlation functions with a fit to the form

1850: eq.\ (\ref{fitcf1}). We see in this figure that $\eta_1$ is constant

1851: close to $p=0.107-0.108$, at least for the largest sizes that we can

1852: reach, $L=20$. For larger sizes, we expect that this value can

1853: still decrease, see the discussion below.

1854: %

1855: \begin{figure}

1856: \begin{center}

1857: %\leavevmode

1858: \epsfxsize=400pt{\epsffile{PlotMPM01.ps}}

1859: \end{center}

1860: \protect\caption[2]{\label{mgaus01} Effective magnetic exponent $\eta_1$

1861: vs.\ $L$ for the $\pm J$ disorder case at $T=0$.}

1862: \end{figure}

1863: %

1864: Next, in Fig.\ \ref{mgaus02}, we compare the magnetic exponent (or

1865: more precisely $\eta_1 = 2 x_h$) obtained for the $\pm J$ distribution

1866: of disorder, both on the Nishimori line and at $T=0$. In this figure,

1867: one can see that the critical point at $T=0$, denoted by $p_c^0$ is

1868: very close to $p_c$, the critical point on the Nishimori line. The

1869: best measurements yields a value of $p_c^0$ slightly smaller than $p_c$,

1870: close to $0.108$. This value is far from the one obtained in the

1871: previous section, {\it i.e.} $p_c^0 = 0.103$. We explain this

1872: difference by the existence of strong finite-size corrections, a

1873: situation which is frequent for two-dimensional spin glasses, see

1874: \cite{HM}. These finite-size corrections can also be observed

1875: directly in Fig.\ \ref{mgaus01}. For small sizes, $6\le L \le10$, the

1876: effective exponent $\eta_1$ is constant for $p \simeq 0.109$ and only

1877: for this value of $p$. For $10\le L \le14$, it is almost constant for $p

1878: \simeq 0.108$. By increasing $L$, the estimate for $p_c^0$ will continue

1879: to decrease. For the largest size that we simulated, $p_c^0$ is estimated

1880: close to $0.107$. Presumably, this value will still decrease with increasing

1881: size. In the discussion in the previous subsection, we had already

1882: observed that a power-law behaviour is apparent only for $L > 20$.

1883: 

1884: \begin{figure}

1885: \begin{center}

1886: %\leavevmode

1887: \epsfxsize=400pt{\epsffile{PlotMPM03.ps}}

1888: \end{center}

1889: \protect\caption[2]{\label{mgaus02} Effective magnetic exponent $\eta_1$

1890: vs.\ $p$ for $\pm J$ disorder case on the Nishimori line and at $T=0$.}

1891: \end{figure}

1892: 

1893: 

1894: Fig.\ \ref{mgaus03} shows $\eta_1$ for the Gaussian distribution

1895: of disorder on the Nishimori line and at $T=0$. Here again, we clearly

1896: observe a small re-entrance, with $\sigma_c^0 \simeq

1897: 0.97$ compared to $\sigma_c \simeq 0.97945$ (see (\ref{sigmaCritVal}))

1898: on the Nishimori line. 

1899: Contrary to the $\pm J$ disorder case, there are very small finite-size

1900: effects. For $L \simeq 14$, one already obtains the same result for $\sigma_c^0$

1901: as by domain-wall measurements on much bigger systems. 

1902: 

1903: \begin{figure}

1904: \begin{center}

1905: \epsfxsize=400pt{\epsffile{PlotMGA.ps}}

1906: \end{center}

1907: \protect\caption[2]{\label{mgaus03} Effective magnetic exponent $\eta_1$

1908: vs.\ $\sigma$ for Gaussian disorder on the Nishimori line and at $T=0$.}

1909: \end{figure}

1910:  

1911: Finally, let us compare the value of the magnetic

1912: exponents for the two types of disorder at $T=0$. For the Gaussian

1913: disorder, one obtains $\eta_1 = 0.11-0.12$, see Fig.\ \ref{mgaus03}.

1914: The $\pm J$ disorder case is less clear. In Fig.~\ref{mgaus02},

1915: as discussed above, one still has strong finite-size

1916: corrections. If one considers the asymptotic point

1917: $p_c^0 \simeq 0.103$ (compare section \ref{secZeroT}),

1918: then the corresponding $\eta_1$ will be close to $0.11$, thus

1919: rather close to the one of the $\pm J$ disorder case. Still, this  is not

1920: sufficient to conclude that we have universality at $T=0$. Indeed, at

1921: $T=0$, the type of disorder influences the degeneracy of the ground

1922: states. For the Gaussian distribution of disorder, each correlation

1923: function (before averaging over disorder) will be equal to $\pm 1$.

1924: Thus the even moments will always be equal to one, while the odd

1925: moments will be equal to the first moment. Then, one obtains the

1926: general result, valid as long the ground state is unique (up to a

1927: global symmetry), $\eta_{2n+1} = \eta_1$ for any $n$ and $\eta_{2n}=

1928: 0$.  On the contrary, for the $\pm J$ disorder case, we have a huge

1929: degeneracy of the ground states. A direct determination of the

1930: exponents $\eta_i$ for $L=20$ and $p=0.107$ yields:

1931: %

1932: \bea

1933: \eta_1 = 0.1277 \pm  0.0004 \; &;& \; \eta_2 = 0.0593 \pm  0.0002 \nonumber \\

1934: \eta_3 = 0.1369 \pm  0.0004 \; &;& \; \eta_4 = 0.0719 \pm  0.0002 \nonumber \\

1935: \eta_5 = 0.1412 \pm  0.0004 \; &;& \; \eta_6 = 0.0776 \pm  0.0002 \nonumber \\

1936: \eta_7 = 0.1438 \pm  0.0004 \; &;& \; \eta_8 = 0.0810 \pm  0.0002 \, .

1937: \eea

1938: %

1939: The value $p=0.107$ is chosen even if it is not exactly at the

1940: asymptotic critical point $p_c^0 \simeq 0.103$ since for this value

1941: we have measurements for large sizes. We also computed the same

1942: exponents for smaller $p$ and smaller $L$. The change in the exponents

1943: is very small. We thus conclude that for the $\pm J$ disorder case,

1944: all moments are different.

1945: 

1946: \section{Summary and conclusions}

1947: 

1948: In this paper we have performed an extensive study of the random-bond

1949: Ising model with a particular emphasis on non-trivial fixed points.

1950: Concerning the Nishimori point, which is  located at the intersection 

1951: of the Ferro-Para critical line and the Nishimori line, we have used

1952: domain-wall free energy computations to provide an accurate estimate 

1953: for the location of this multicritical point. This has been done for both, the binary

1954: $\pm J$ distribution as well as for the Gaussian distribution.

1955: Our numerical results show that a conjecture for the location of

1956: the Nishimori point based on a duality argument \cite{NN} yields

1957: only a very good approximation, but is not exact.

1958: 

1959: Next, we estimated $\nu = 1.48(3)$ for $\pm J$ disorder and $\nu =

1960: 1.52(3)$ for Gaussian disorder. This agrees with other recent

1961: estimates for $\pm J$ disorder, namely $\nu = 1.50(3)$ on the square

1962: lattice \cite{MC}, and $\nu = 1.49(2)$ on the triangular and honeycomb

1963: lattices \cite{deQTH}. All these results are consistent with a

1964: universal value $\nu \approx 1.50$ for the Nishimori point. We have

1965: obtained further accurate results for the exponents $\eta_1, \ldots,

1966: \eta_8$ for the moments of the spin-spin correlation functions. The

1967: estimates for $\pm J$ disorder (\ref{valExp20}) and Gaussian disorder

1968: (\ref{valExp14}) are not only very close to each other, but also to

1969: recent estimates on the triangular and honeycomb lattices

1970: \cite{deQTH}.  Our analysis of the central charge $c$ at the Nishimori

1971: point for both types of disorder is also consistent with a universal

1972: value $c = 0.464 \pm 0.004$ \cite{HPP}.

1973: All these results\footnote{Essentially the same results are found \cite{LQ06}

1974: if the approximate value $p_c \approx 0.110028$ \cite{NN} is used

1975: instead of the numerically exact value (\ref{pcrit1}).}

1976: suggest a single universality class of the Nishimori point in the

1977: two-dimensional random-bond Ising model, and definitely exclude

1978: percolation as the possible universality class for this point.

1979: 

1980: We have also considered a probability distribution for the bonds with

1981: dilution, in which some of the coupling constants are zero.

1982: In the purely diluted case (a distribution containing

1983: only positive or zero bonds) there is only one non-trivial fixed point,

1984: namely the zero-temperature percolation point, apart from

1985: the critical point of the pure system \cite{YeSt79}.

1986: We have confirmed that this percolation point is unstable against

1987: the Nishimori point

1988: if one moves within the intersection of the critical transition

1989: surface and the Nishimori manifold by considering both, dilution and

1990: $\pm J$ couplings, and that these two points (percolation and Nishimori)

1991: are the only fixed points within this intersection line. 

1992: On the other hand, going off the Nishimori line but staying within

1993: the critical transition line we confirm that the Nishimori

1994: point is unstable in favour the pure Ising model fixed point. All these

1995: results are obtained by studying the crossover of the effective

1996: central charge, or first Lyapunov exponent, in strips of increasing

1997: width.

1998: 

1999: Finally, we have analyzed the critical point at zero temperature

2000: corresponding to the ferro-para transition in the model without dilution

2001: for both the binary and Gaussian distributions. Our numerical analysis

2002: confirms the strict re-entrance of the ferromagnetic

2003: phase \cite{Preskill,Hartmann}. We have also investigated the

2004: criticality of the zero-temperature fixed point and argued that it is,

2005: as for the finite-temperature Nishimori point, different from

2006: percolation. The results obtained in this paper raise the question

2007: of the apparent hierarchy of fixed points for strongly disordered systems,

2008: namely in our case the pure Ising transition, percolation, the Nishimori

2009: point and the zero-temperature point, all corresponding to different (and

2010: certainly non-unitary) conformal field theories with values for the

2011: central charge and critical exponents which are extremely close to

2012: each other but nevertheless different. Future analytical efforts will

2013: be needed to understand such effects

2014: of strong disorder in two-dimensional classical statistical systems.

2015: 

2016: %\noindent{\large\bf Acknowledgments} 

2017: 

2018: %\vspace*{0.7 true cm}

2019: %

2020: %\newpage

2021: %\small

2022: 

2023: \begin{thebibliography}{99}

2024: 

2025: 

2026: \bibitem{DDSL}

2027: Vik. S. Dotsenko and Vl. S. Dotsenko,

2028: \newblock {\it Sov.\ Phys.\ JETP Lett.}\ {\bf 33} (1981) 37; 

2029: \newblock {\it Adv.\ Phys.}\ {\bf 32} (1983) 129;

2030: B. N. Shalaev,

2031: \newblock {\it Sov. Phys. Solid State}~{\bf 26} (1984) 1811;

2032: A. W. W. Ludwig,

2033: \newblock {\it Nucl. Phys. }~B~{\bf 330} (1990) 639

2034: 

2035: \bibitem{N}

2036: H. Nishimori,

2037: \newblock {\it J.\ Phys.\ C: Solid State Phys.}\ {\bf 13} (1980) 4071;

2038: \newblock {\it Prog.\ Theor.\ Phys.}~{\bf 66} (1981) 1169

2039: 

2040: \bibitem{ON}

2041: Y. Ozeki and H. Nishimori,

2042: \newblock {\it J. Phys. A: Math. Gen.}~{\bf 26} (1993) 3399

2043: 

2044: \bibitem{LG}

2045: A. Georges and P. Le Doussal,

2046: \newblock {\it unpublished Preprint} (1988); 

2047: P.~Le Doussal and A.~B.\ Harris, {\it Phys.\ Rev.\ Lett.}~{\bf 61} (1988) 625;

2048: {\it Phys. Rev.}~B~{\bf 40} (1989) 9249

2049: 

2050: \bibitem{CF}

2051: S. Cho and M. P. A. Fisher,

2052: \newblock {\it Phys. Rev.}~B {\bf 55} (1997) 1025

2053: 

2054: \bibitem{GRL}

2055: I. A. Gruzberg, N. Read and A.~W.~W.~Ludwig,

2056: \newblock {\it Phys. Rev.}~B {\bf 63} (2001) 104422

2057: 

2058: \bibitem{CRKHAL} J.\ T.\ Chalker, N.\ Read, V.\ Kagalovsky, B.\ Horovitz, Y.\ Avishai and A.\ W.\ Ludwig

2059: \newblock {\it Phys. Rev.}~B {\bf 65} (2001) 012506

2060: 

2061: \bibitem{MENMD06} A.\ Mildenberger, F.\ Evers, R.\ Narayanan, A.~D.\ Mirlin

2062:               and K.\ Damle, {\it Phys.\ Rev.}\ B {\bf 73} (2006) 121301(R)

2063: 	      

2064: \bibitem{Sourlas} N.\ Sourlas, {\it Europhys.\ Lett.}\ {\bf 25} (1994) 159;

2065:                {\it Preprint} cond-mat/9811406

2066: 

2067: \bibitem{NishCod} H.\ Nishimori, {\it Physica} A {\bf 205} (1994) 1;

2068:                   {\it Physica} A {\bf 315} (2002) 243

2069: 

2070: \bibitem{Iba} Y.\ Iba, {\it J.\ Phys.\ A: Math.\ Gen.}\ {\bf 32} (1999) 3875

2071: 

2072: \bibitem{Preskill}

2073: C. Wang, J. Harrington and J. Preskill,

2074: \newblock {\it Annals Phys.}~{\bf 303} (2003) 31

2075: 

2076: \bibitem{Z}

2077: M. R. Zirnbauer,

2078: \newblock {\it J. Math. Phys.}~{\bf 37} (1996) 4986; 

2079: A. Altland and M. R. Zirnbauer,

2080: \newblock {\it Phys. Rev.}~B {\bf 55} (1997) 1142

2081: 

2082: \bibitem{GRLC}

2083: I. A. Gruzberg, N. Read and A.~W.~W.~Ludwig,

2084: \newblock {\it Phys. Rev. Lett.}~{\bf 82} (1999) 4524;

2085: J. Cardy,

2086: \newblock {\it Phys. Rev. Lett.}~{\bf 84} (2000) 3507

2087: 

2088: 

2089: \bibitem{HPP} A.\ Honecker, M.\ Picco and P.\ Pujol,

2090: \newblock {\it Phys. Rev. Lett.}~{\bf 87} (2001) 047201

2091: 

2092: \bibitem{MM}

2093: W. L. McMillan,

2094: \newblock {\it Phys. Rev.}~B {\bf 29} (1984) 4026

2095: 

2096: \bibitem{SA}

2097: R. R. P. Singh and J. Adler,

2098: \newblock {\it Phys. Rev.}~B {\bf 54} (1996) 364

2099: 

2100: \bibitem{AQdS}

2101: F. D. A. Ar\~ao Reis, S. L. A. de Queiroz and R. R. dos Santos,

2102: \newblock {\it Phys. Rev.}~B {\bf 60} (1999) 6740

2103: 

2104: \bibitem{MC}

2105: F. Merz and J.~T. Chalker,

2106: \newblock {\it Phys. Rev.}~B {\bf 65} (2002) 054425

2107: 

2108: \bibitem{MC2} F.\ Merz and J.~T.\ Chalker, 

2109: %{\it Negative Scaling Dimensions and

2110: %              Conformal Invariance at the Nishimori Point in the $\pm J$

2111: %              Random-Bond Ising Model}, % {\it Preprint} cond-mat/0201137

2112: \newblock {\it Phys.\ Rev.}~B~{\bf 66} (2002) 054413

2113: 

2114: \bibitem{NN}

2115: H. Nishimori and K. Nemoto,

2116: \newblock  {\it J. Phys. Soc. Jpn.}~{\bf 71} (2002) 1198;

2117: H.\ Nishimori, {\it Preprint} cond-mat/0602453

2118: 

2119: \bibitem{Grinstein}

2120: G. Grinstein, C. Jayaprakash and M. Wortis,

2121: \newblock {\it Phys. Rev.}~B {\bf 19} (1979) 260

2122: 

2123: \bibitem{Freund}

2124: H. Freund and P. Grassberger,

2125: \newblock {\it J. Phys. A: Math. Gen.}~{\bf 22} (1989) 4045

2126: 

2127: \bibitem{Bendish}

2128: J. Bendish, U. Derigs and A. Metz,

2129: \newblock {\it Discrete Applied Mathematics}~ {\bf 52} (1994) 139

2130: 

2131: \bibitem{KaRi} N.\ Kawashima and H.\ Rieger, 

2132: %{\it Finite Size Scaling Analysis of

2133: % Exact Ground States for $\pm J$ Spin Glass Models in Two

2134: %             Dimensions}, 

2135: \newblock {\it Europhys.\ Lett.}\ {\bf 39} (1997) 85

2136: 

2137: \bibitem{MB} G.\ Migliorini and A.\ N.\ Berker, 

2138: \newblock {\it Phys.\ Rev.}~B~{\bf 58} (1998) 426

2139: 

2140: \bibitem{BGP} J.~A.\ Blackman, J.~R.\ Gon\c{c}alves and J.\ Poulter, 

2141: %{\it Properties

2142: %              of the Two-Dimensional Random-Bond $\pm J$ Ising Spin Glass},

2143: \newblock {\it Phys.\ Rev.}~E~{\bf 58} (1998) 1502

2144: 

2145: \bibitem{Hartmann}

2146: C. Amoruso and A. K. Hartmann,

2147: \newblock {\it Phys. Rev.}~B {\bf 70} (2004) 134425

2148: 

2149: \bibitem{NS}

2150: C.~M. Newman and D.~L. Stein,

2151: \newblock {\it Phys. Rev. Lett.}~{\bf 84} (2000) 3966;

2152: \newblock {\it Commun.\ Math.\ Phys.}\ {\bf 224} (2001) 205 % cond-mat/0103395.

2153: 

2154: \bibitem{OzekiNishimori} 

2155: Y.\ Ozeki and H.\ Nishimori, 

2156: %{\it Phase Diagram of the

2157: %$\pm J$ Ising Model in Two Dimensions}, 

2158: \newblock {\it J.\ Phys.\ Soc.\ Jpn.}\ {\bf 56} (1987) 3265

2159: 

2160: \bibitem{OzekiIto} Y.\ Ozeki and N.\ Ito, 

2161: %{\it Multicritical Dynamics for the

2162: %              $\pm J$ Ising Model}, 

2163: \newblock {\it J.\ Phys.\ A: Math.\ Gen.}\ {\bf 31} (1998) 5451

2164: 

2165: \bibitem{MNN03} J.~M.\ Maillard, K.\ Nemoto and H.\ Nishimori, 

2166: %{\it Symmetry,

2167: %              Complexity and Multicritical Point of the Two-Dimensional Spin

2168: %              Glass}, 

2169: \newblock {\it J.\ Phys.\ A: Math.\ Gen.}\ {\bf 36} (2003) 9799

2170: 

2171: \bibitem{MS95} G.\ Mussardo and P.\ Simonetti, {\it Phys.\ Lett.}\ B {\bf 351}

2172:               (1995) 515

2173: 

2174: \bibitem{CHMP} D.~C.\ Cabra, A.\ Honecker, G.\ Mussardo and P.\ Pujol,

2175:           {\it J.\ Phys.\ A: Math.\ Gen.}\ {\bf 30} (1997) 8415

2176: 

2177: \bibitem{Night} M.~P.\ Nightingale, pp.\ 287-351 in: V.\ Privman (ed.),

2178:               {\it Finite Size Scaling and Numerical Simulations of

2179:               Statistical Physics}, World Scientific, Singapore (1990)

2180: 

2181: \bibitem{YeSt79} J.~M.\ Yeomans and R.~B.\ Stinchcombe,

2182:               {\it J.\ Phys.\ C: Solid State Phys}.\ {\bf 12} (1979) 347

2183: 

2184: \bibitem{StAh} D.\ Stauffer and A.\ Aharony, {\it Introduction to Percolation

2185:               Theory}, 2nd edition, Taylor \& Francis, London (1994)

2186:               

2187: 

2188: \bibitem{Sor} E.~S.\ S{\o}rensen,

2189: % {\it Logarithmic Corrections to the RG Flow

2190: %              for the Two-Dimensional Bond Disordered Ising Model},

2191:               {\it Preprint} cond-mat/0006233

2192: 

2193: 

2194: \bibitem{HJPP} A.\ Honecker, J.~L.\ Jacobsen, M.\ Picco and P.\ Pujol, 

2195: %{\it

2196: %              Nishimori Point in Random-Bond Ising and Potts Models in 2D},

2197: %              Proceedings of the NATO Advanced Research Workshop on

2198: \newblock {\it Statistical Field Theories}, Como, 18-23 June 2001,

2199:              eds.\ A.\ Cappelli, G.\ Mussardo, Kluwer Academic Publishers,

2200:              Dordrecht (2002) 251-261 [{\it Preprint} cond-mat/0112069].

2201: 

2202: \bibitem{central} H. W. J. Bl\"ote, J. L. Cardy and M. P. Nightingale,

2203: \newblock {\it Phys.\ Rev.\ Lett.}\ {\bf 56} (1986) 742;

2204: I.\ Affleck,

2205: \newblock {\it Phys.\ Rev.\ Lett.}\ {\bf 56} (1986) 746

2206: 

2207: \bibitem{JC}

2208: J.~L. Jacobsen and J.~L. Cardy, 

2209: {\it Nucl. Phys.}~B~{\bf 515} (1998) 701

2210: 

2211: \bibitem{Cardy} J.\ Cardy, {\it Scaling and Renormalization in Statistical

2212:               Physics}, Cambridge Lecture Notes in Physics 5, Cambridge

2213:               University Press (1996)

2214: 

2215: \bibitem{WK74} K.~G.\ Wilson and J.\ Kogut,

2216:               {\it Phys.\ Rep.}\ {\bf 12} (1974) 75

2217: 

2218: \bibitem{QSnew} S.~L.~A.\ de Queiroz and R.~B.\ Stinchcombe, 

2219: %{\it

2220: %              Correlation-Function Distributions at the Nishimori Point of

2221: %              Two-Dimensional Ising Spin Glasses}, 

2222: \newblock {\it Phys.\ Rev.}~B~{\bf 68} (2003) 144414

2223: 

2224: \bibitem{cardyr} J.~L.~Cardy, in {\it Phase Transitions and Critical Phenomena},

2225:                  Vol. 11,  edited by C. Domb and J. Lebowitz,

2226: 		 Academic Press, London (1987)

2227: 

2228: \bibitem{Binder81} K.\ Binder, {\it Z.\ Phys.}\ B {\bf 43} (1981) 119

2229: 

2230: \bibitem{LB00} D.~P.\ Landau and K.\ Binder, {\it A Guide to Monte Carlo

2231:      Simulations in Statistical Physics}, Cambridge University Press (2000)

2232: 

2233: \bibitem{NK}

2234: H. Nishimori,

2235: \newblock {\it J. Phys. Soc. Jpn.}~{\bf 55} (1986) 3305;

2236: H. Kitatani,

2237: \newblock {\it J. Phys. Soc. Jpn.}~{\bf 61} (1992) 4049

2238: 

2239: \bibitem{JP} J.~L.\ Jacobsen and M.\ Picco, 

2240: \newblock {\it Phys.\ Rev.}~E~{\bf 65} (2002) 026113

2241: 

2242: \bibitem{Ludwig} A.~W.~W.\ Ludwig, 

2243: \newblock {\it Nucl.\ Phys.}~B~{\bf 285} (1987) 97

2244: 

2245: \bibitem{DPP} V.~Dotsenko, M.\ Picco and P.\ Pujol,  

2246: \newblock {\it Nucl.\ Phys.}\ B {\bf 455} (1995) 701

2247: 

2248: \bibitem{blossom4} W.~Cook and A.~Rohe, {\it INFORMS Journal on

2249: Computing}~{\bf 11} (1999) 138

2250: 

2251: \bibitem{HM} A.~K.\ Hartmann and M.~A.\ Moore, 

2252: \newblock {\it Phys.\ Rev.\ Lett.}\ {\bf 90} (2003) 127201

2253: 

2254: \bibitem{deQTH} S.~L.~A.\ de Queiroz, {\it Phys.\ Rev.}~B~{\bf 73} (2006) 064410

2255: 

2256: \bibitem{LQ06} J.~C.\ Lessa and S.~L.~A.\ de Queiroz,

2257:          {\it Preprint} cond-mat/0605659

2258: 

2259: \end{thebibliography}

2260: 

2261: \end{document}

2262: