1: \documentclass[prl,aps,twocolumn,superscriptaddress,showpacs]{revtex4}
2: %\topmargin=0.5cm
3: \usepackage{amsbsy}
4: \usepackage{amssymb}
5: \usepackage[dvips]{graphicx}
6:
7: \begin{document}
8: \title{Magnetic exponents of two-dimensional Ising spin glasses}
9:
10: \author{F. Liers}
11: \affiliation{Institut f{\"u}r Informatik, Universit{\"a}t zu K{\"o}ln,
12: Pohligstra{\ss}e 1, D-50969 K{\"o}ln,Germany.}
13:
14: \author{O. C. Martin}
15: \affiliation{Univ Paris-Sud, UMR8626, LPTMS, Orsay, F-91405; CNRS, Orsay,
16: F-91405, France.}
17:
18: \date{\today}
19:
20: \begin{abstract}
21: The magnetic critical properties of two-dimensional
22: Ising spin glasses are controversial.
23: Using exact ground state determination,
24: we extract the properties of clusters flipped
25: when increasing continuously a uniform field. We show that these clusters
26: have many holes but otherwise have statistical properties
27: similar to those of zero-field droplets. A detailed analysis
28: gives for the magnetization exponent
29: $\delta \approx 1.30 \pm 0.02$ using
30: lattice sizes up to $80 \times 80$; this is
31: compatible with the droplet model prediction $\delta = 1.282$.
32: The reason for previous disagreements stems from the need to analyze both
33: singular and analytic contributions in the low-field regime.
34: \end{abstract}
35: \pacs{75.10.Nr, 75.40.-s, 75.40.Mg}
36:
37: \maketitle
38:
39: Spin glasses~\cite{MezardParisi87b,Young98} have been the focus of
40: much interest because of their many remarkable features:
41: they undergo a subtle freezing transition as temperature is lowered,
42: their relaxational dynamics is slow (non-Arrhenius), they
43: exhibit ageing, memory effects, etc.
44: Although there are
45: still some heated disputes concerning three-dimensional
46: spin glasses, the case of two dimensions is relatively consensual,
47: at least in the absence of a magnetic field. Indeed,
48: two recent
49: studies~\cite{KatzgraberLee04,HartmannHoudayer04}
50: found that the thermal properties of two-dimensional
51: Ising spin glasses with Gaussian couplings agreed very well
52: with the predictions of the scaling/droplet
53: pictures~\cite{BrayMoore86,FisherHuse86}. Interestingly,
54: the situation in the presence of a magnetic field remains
55: unclear; in particular, some Monte Carlo
56: simulations~\cite{KinzelBinder83}
57: and basically all ground state
58: studies~\cite{KawashimaSuzuki92,Barahona94,RiegerSanten96} seem
59: to go against the scaling/droplet pictures. Nevertheless, since
60: spin glasses often have large corrections to scaling,
61: the apparent disagreement with the
62: droplet picture resulting from these studies may be misleading and
63: tests in one dimension give credence to this claim~\cite{CarterBray03}.
64:
65: In this study we use state of the art algorithms for determining
66: exact ground states in the presence of a magnetic field and
67: treat significantly larger lattice sizes than in previous work.
68: By finding the precise points where the ground states change
69: as a function of the field, we extract the excitations
70: relevant in the presence of a field which can then be
71: compared to the zero-field droplets.
72: Although for small size lattices we agree
73: with previous studies, at our larger ones
74: a careful analysis, taking into account both the analytic
75: and the singular terms, gives excellent agreement with the droplet
76: picture.
77:
78: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
79: \paragraph*{The model and its properties ---} We work on an $L \times L$
80: square lattice having Ising spins on its sites and
81: couplings $J_{ij}$ on its bonds. The
82: Hamiltonian is
83: \begin{equation}
84: \label{eq:H}
85: H(\{\sigma_{i}\}) \equiv
86: -\sum_{\langle ij \rangle} J_{ij}\sigma_{i}\sigma_{j}
87: - B\sum_{i} \sigma_{i}
88: \end{equation}
89: The first sum runs over all nearest neighbor sites using
90: periodic boundary conditions to minimize
91: finite size effects.
92: The $J_{ij}$ are independent random variables of
93: either Gaussian or exponential distribution.
94: %We shall not discuss the case of the $\pm J$ model which even in zero field
95: %has a subtle behavior.
96:
97: It is generally agreed that two-dimensional spin glasses
98: have a unique critical point at $T=B=0$. There,
99: the free energy is non-analytic and in fact,
100: standard arguments~\cite{Cardy96} suggest that
101: as $T \to 0$ and $B \to 0$ the free energy goes as
102: $\beta F(L,\beta) \sim \beta E_0 + G_s(T L^{y_T},B L^{y_B})$
103: where $E_0$ is the ground-state energy, $\beta$ the inverse temperature, while
104: $y_T$ and $y_B$ are the thermal and magnetic exponents.
105: Previous work when $B=0$ is compatible with this
106: form and in fact also agrees with the scaling/droplet picture
107: of Ising spin glasses in which one has
108: $y_T=-\theta\approx 0.282$.
109: The stumbling block concerns the behavior when $B\ne 0$;
110: there, the droplet prediction in general dimension $d$ is
111: %
112: \begin{equation}
113: \label{eq:y_B}
114: y_B = y_T + d/2
115: \end{equation}
116: %
117: but the numerical evidence for this is muddled
118: at best. It is thus worth
119: reviewing the hypotheses assumed by the droplet model
120: so that they can be tested directly.
121:
122: We begin with the fact that in any dimension $d$,
123: a magnetic field destabilizes
124: the ground state beyond a characteristic length scale $\xi_B$.
125: To see this, consider an infinitesimal field and
126: zero-field droplets of scale $\ell$.
127: These are expected to be compact.
128: The interfacial energy of such droplets is $O(\ell^{\theta})$
129: while their total magnetization goes as
130: $\ell^{d/2}$. The magnetic and
131: interfacial energy are then balanced when
132: $B$ reaches a value $O(1/\ell^{d/2-\theta})$: at that value
133: of the field, some of the droplets will flip and the ground
134: state will be destabilized. We then see that for each
135: field strength there is an associated magnetic length scale
136: $\xi_B$
137: \begin{equation}
138: \xi_B \approx B^{-\frac{1}{{\frac{d}{2}-\theta}}}
139: \end{equation}
140: This leads to the identification $y_B = d/2 - \theta$
141: in agreement with Eq.~\ref{eq:y_B}, giving
142: $y_b \approx 1.282$ at $d=2$.
143:
144: The droplet model also predicts the scaling
145: of the magnetization in the $B \to 0$ limit via
146: the exponent $\delta$:
147: \begin{equation}
148: m(B) \sim B^{1/\delta}
149: \end{equation}
150: If this form also holds for infinitesimal
151: fields at finite $L$,
152: % (without first taking the $L\to \infty$ limit),
153: we can consider the field $B^*$ for which
154: system-size droplets flip; this happens when $B = O(1/L^{y_B})$
155: and then the magnetization is $O(L^{-d/2})$, the droplets
156: having random magnetizations. This leads to
157: $m(B^*) \sim L^{-d/2}$ and $m(B^*) \sim [1/L^{y_B} ]^{1/\delta}$ so that
158: \begin{equation}
159: d \delta = 2 y_B
160: \end{equation}
161:
162: Although the droplet model arguments are not proofs,
163: they seem quite convincing. Nevertheless, the
164: numerical studies measuring $\delta$ do not give
165: good agreement with the prediction $\delta =1.282$.
166: For instance, using Monte Carlo at ``low enough'' temperatures,
167: Kinzel and Binder~\cite{KinzelBinder83}
168: find $\delta \approx 1.39$. Since thermalization
169: is difficult at low temperatures, it is preferable to
170: work directly with ground states, at least when that
171: is possible. This was done by three independent
172: groups~\cite{KawashimaSuzuki92,Barahona94,RiegerSanten96}
173: with increasing power, leading
174: to $\delta \approx 1.48$, $\delta \approx 1.54$
175: and $\delta \approx 1.48$. Taken together, these
176: studies show a real discrepancy with the droplet
177: prediction.
178: To save the droplet model from this thorny situation, one can appeal
179: to large corrections to scaling.
180: Such potential effects have been considered~\cite{CarterBray03}
181: in dimension one where it was shown
182: that $\xi_B$ was poorly fitted by a pure power law
183: unless the fields were very small. Here we
184: revisit the two-dimensional case
185: to reveal either the size of the corrections
186: to scaling or a cause for the break down of the
187: droplet reasoning.
188:
189: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
190: \paragraph*{Computation of ground states ---}
191:
192: We determine the exact ground state of the Hamiltonian (\ref{eq:H}) by
193: computing a maximum cut in the graph of
194: interactions~\cite{Barahona82}, a prominent problem in
195: combinatorial optimization. Whereas it is polynomially solvable on
196: two-dimensional grids without a field
197: and couplings bounded by a polynomial in the size of the input,
198: it is NP-hard with an
199: external field. In practice, we rely on a branch-and-cut
200: algorithm~\cite{BarahonaGrotschel88,LiersJunger04}.
201:
202: Let the ground state at a field $B$ be denoted as
203: $\{\sigma^{(G)}(B)\}$. To study the magnetization,
204: we computed the ground states at increasing values of $B$,
205: in steps of size $0.02$. When focusing instead on
206: the flipping clusters, we had to determine
207: the intervals in which the ground state was constant
208: and in what manner it changed when going from one
209: interval to the next. In Fig.~\ref{fig:fig1}
210: we show the associated piecewise constant magnetization curves
211: for three samples of the disorder variables $J_{ij}$ at $L=10$.
212: %
213: \begin{figure}[h]
214: %\includegraphics[width=8.5cm, height=5cm,angle=0]{fig1.eps}
215: \includegraphics[width=8.5cm, height=5cm,angle=0]{fig1.eps}
216: \caption{Magnetization as a function of $B$ for three
217: typical $L=10$ samples.\protect\label{fig:fig1}}
218: \end{figure}
219: %
220: To get the sequence of intervals or break points
221: associated with such a function
222: exactly, we start by computing the ground state in zero field. By
223: applying postoptimality analysis from linear programming
224: theory, we determine~\cite{RiegerSanten96,LiersJunger04} a range
225: $\Delta B$ such that the ground state at a field $B$ remains the
226: optimum in the interval $[B,B+\Delta B]$. We reoptimize at
227: $B+\Delta B+\epsilon$, with $\epsilon$ being a sufficiently small number. By
228: repeatedly applying this procedure, we get a new ground state
229: configuration, and increase $B$ until all spins are aligned with the
230: field. This procedure works for system sizes in which the
231: branch-and-cut program can prove optimality without branching, i.e.,
232: without dividing the problem into smaller sub-problems.
233: If the algorithm branches (this occurs
234: only for the largest system sizes studied here), we apply a
235: divide-and-conquer strategy for determining
236: $\{\sigma^{(G)}(B)\}$ in an interval, say $[B_1,B_2]$. For a fixed
237: configuration the Hamiltonian (\ref{eq:H}) is linear in the field, the
238: slope being the system's magnetization. Let $f_1,f_2$ be the two
239: linear functions associated with $\{\sigma^{(G)}(B_1)\}$ and
240: $\{\sigma^{(G)}(B_2)\}$. If $f_1$ and $f_2$ are equal,
241: we are done. Otherwise, we determine the field $B_3$ at which the
242: functions intersect and recursively solve the problem in the intervals
243: $[B_1,B_3]$ and $[B_3,B_2]$.
244: %This turns out to be quite effective,
245: %relatively few ground states need to be calculated for determining
246: %$\{\sigma^{(G)}(B)\}$.
247:
248: A typical sample at $L=80$ requires about 2 hours of cpu on a work
249: station for determining the ground states when $B$ goes through
250: the multiples of $0.02$. The more time consuming computation
251: of the exact break points takes about 4 hours
252: on typical samples with $L=60$, but less than a minute if $L\le 30$
253: because the ground-state determinations are fast and branching almost
254: never arises. For our work, we considered mainly the
255: case of Gaussian $J_{ij}$, analyzing 2500 samples at $L=80$, 5000 at
256: $L=70$, and from 2000 to 11000 instances for sizes $L=60, 50, 40, 30,
257: 24, 20, 14$.
258: % detailed numbers of samples:
259: % instances L
260: % 11281 10
261: % 12504 14
262: % 3739 20
263: % 6979 24
264: % 5993 30
265: % 3612 40
266: % 2008 50
267: % 2567 60
268: We also analyzed a smaller number of samples for $J_{ij}$
269: taken from an exponential distribution; exponents showed
270: no significant differences when comparing to the
271: Gaussian case.
272:
273: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
274: \paragraph*{The exponent $\delta$ ---}
275: Given the Hamiltonian, it is easy to see that
276: for each sample the magnetization (density)
277: \begin{equation}
278: m_J(B) = \frac{\sum_i \sigma_i^{(G)}}{L^2}
279: \end{equation}
280: must be an increasing function of $B$. (The index $J$ on
281: the magnetization is to recall that it depends on the
282: disorder realisation, but in the large $L$ limit
283: $m_J$ is self averaging; also, without loss of generality,
284: we shall work with $B>0$.) At large fields $m_J$ saturates
285: to $1$, while at low fields, its growth law must be above
286: a linear function of $B$. Indeed, for continuous $J_{ij}$,
287: the distribution of local fields has a finite density at zero and so
288: small clusters of spins will flip and will lead to a linear
289: contribution to the magnetization. A more singular
290: behavior is in fact predicted by the droplet model
291: since $\delta > 1$, indicating that
292: the system is anomalously sensitive to the magnetic field perturbation.
293: % Note that in some rare cases,
294: % the zero-field ground state will have $m_J(B=0)=0$ exactly.
295: %From the disorder average of $m_J$, we now consider estimating
296: %the exponent $\delta$ using our data with $L\le 80$.
297:
298: If $B$ is not too small, the convergence to the thermodynamic
299: limit ($L \to \infty$) is rapid, and in fact one expects
300: exponential convergence in $L/\xi_B$. We should thus see
301: an envelope curve $m(B)$ appear as $L$ increases; to make
302: a power dependence on $B$ manifest, we show in Fig.~\ref{fig:fig2}
303: a log-log plot of the ratio $m(B)/B^{1/\delta}$ where
304: $\delta$ is set to its droplet scaling value of 1.282.
305: %
306: \begin{figure}[h]
307: %\includegraphics[width=8.5cm, height=6cm]{fig2.eps}
308: \includegraphics[width=8.5cm, height=6cm]{fig2.eps}
309: \caption{Magnetization divided by $B^{1/\delta}$ as a function of $B$;
310: the $B^{1/1.45}$ line is to guide the eye. From top to bottom,
311: $L=14$, 20, 24, 30, 40, 50, 60, 70, and 80.
312: Inset: $m-\chi_1 B$ divided by
313: $B^{1/\delta}$ as a function of $B$. (Same $L$ and
314: symbols as in main
315: part of the figure.) In both cases, $\delta$ is set to
316: its droplet model value, $\delta=1.282$.\protect\label{fig:fig2}}
317: \end{figure}
318: %
319: For that value of $\delta$ there is not much indication that a flat
320: region is developping when $L$ increases, while at $L=50$ a direct fit
321: to a power gives $\delta = 1.45$ (cf. line displayed
322: in the figure to guide the eye), as found in previous
323: work~\cite{KawashimaSuzuki92,Barahona94,RiegerSanten96}.
324: The problem with this simple
325: analysis is that $m$ has both analytic and non-analytic contributions;
326: to lowest order we have
327: \begin{equation}
328: \label{eq:m_expansion}
329: m = \chi_1 B + \chi_S B^{1/\delta} + \ldots
330: \end{equation}
331: Although $\chi_1 B$ is sub-dominant, it is far from negligible
332: in practice; for instance for it to contribute to less than
333: $10\%$ of $m$, one would need $B < (0.1 \chi_S/\chi_1)^{1/0.282}$.
334: This could easily mean $B < 10^{-3}$ for which there would be huge finite size
335: effects since $L$ would then be much smaller than the magnetic length
336: $\xi_B$. We thus must take into account the term $\chi_1 B$; we have
337: done this, adjusting $\chi_1$ so that $(m - \chi_1 B)/B^{1/\delta}$
338: has an envelope as flat as possible. The result is displayed
339: in the inset of Fig.~\ref{fig:fig2}, showing that the
340: droplet scaling fits very well the data as long as
341: the $\chi_1 B$ term is included. In fact, direct fits to
342: the form of Eq.~\ref{eq:m_expansion}
343: give $\delta$ in the range 1.28 to 1.32 depending on the sets of $L$'s
344: included in the fits.
345:
346: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
347: \paragraph*{The clusters that flip are like zero-field droplets ---}
348: %If the droplet model gave an erroneous prediction for
349: %$\delta$, we would need to find out why.
350: The fundamental hypothesis
351: in the droplet argument relating $\delta$ or $y_B$ to $\theta$ is the fact
352: that in an infinitesimal field one flips droplets defined
353: in zero field, droplets which are compact and
354: have random (except for the sign) magnetizations. We
355: therefore now focus on the
356: properties of the actual clusters that are flipped at low fields.
357:
358: At zero field, the droplet of lowest energy
359: almost always is a single spin (this
360: follows from the large number of such droplets,
361: in spite of their typically higher energy).
362: Thus as the field is turned on, the ground state changes
363: first mainly via single spin flips, and
364: when large clusters do flip (they finally
365: do so but at larger fields),
366: they necessarily have many ``holes''
367: and thus do not correspond exactly to
368: zero-field droplets. This is not a problem for the droplet
369: argument as long as these clusters are
370: compact and have random magnetizations.
371: %
372: \begin{figure}[h]
373: %\includegraphics[width=8.5cm, height=5cm]{fig3.eps}
374: \includegraphics[width=8.5cm, height=5cm]{fig3.eps}
375: \caption{The cluster magnetization divided
376: by the square root of cluster volume --- for the largest cluster flipped
377: in each sample --- is insensitive to $L$.
378: Inset: The clusters's mean surface scales as $L^{d_S}$
379: with $d_S\approx 1.32$.\protect\label{fig:fig3}}
380: \end{figure}
381: %
382: To test this, we consider for each realization of the $J_{ij}$ disorder
383: the \emph{largest} cluster that flips during
384: the whole passage from $B=0$ to $B=\infty$.
385: According to the droplet picture, this cluster
386: should contain a number of spins $V$ that scales
387: as $L^2$ (compactness) and have a total magnetization
388: $M$ that scales as $\sqrt{V}$ (randomness).
389: This is confirmed by our data where we find $M/V\sim 2/L$;
390: in Fig.~\ref{fig:fig3} we plot the disorder
391: mean of $M/\sqrt{V}$ for increasing $L$; manifestly,
392: this mean is remarkably insensitive to $L$.
393: Similar conclusions apply to $V/L^2$. For completeness,
394: we show in the inset of the figure that the surface of
395: these clusters, defined as the number of lattice
396: bonds connecting them to their complement,
397: grows as $L^{d_S}$ with $d_S\approx 1.32$;
398: this is to be compared to the value $d_S=1.27$ for
399: zero-field droplets~\cite{HartmannYoung02}, in spite
400: of the fact that our clusters have holes.
401: All in all, we find that the clusters considered have statistical
402: properties that are completely compatible with those assumed in
403: the droplet scaling argument, thereby directly validating
404: the associated hypotheses.
405:
406:
407: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
408: \paragraph*{The magnetic exponent $y_B$ and finite size scaling of the magnetization ---}
409: One can also measure the exponent $y_B$ directly
410: via the magnetic length which scales as $\xi_B \sim B^{-1/y_B}$.
411: For each sample, define $B_J^*$ as that field where the ground
412: state changes by the largest cluster of spins as described in the
413: previous paragraph. Since these clusters involve a number of spins
414: growing as $L^2$, we can identify $\xi_B(B_J^*)$
415: with $L$. Let $B^*$ be the disorder average of $B_J^*$; then
416: $B^* \sim L^{-y_B}$ from which we can estimate $y_B$. We find that a pure
417: power with $y_B$ set to its value in the droplet picture describes
418: the data quite well; in the inset of Fig.~\ref{fig:fig4}
419: %
420: \begin{figure}[h]
421: %\includegraphics[width=8.5cm, height=6cm]{fig4.eps}
422: \includegraphics[width=8.5cm, height=6cm]{fig4.eps}
423: \caption{Inset: Field $B^*$ times $L^{1/\delta}$
424: as a function of $1/L$ shows
425: a limit at large $L$
426: as expected in the droplet model ($\delta=1.282$).
427: Main figure:
428: Data collapse plot exhibiting finite size scaling of the
429: singular part of the magnetization
430: ($L=10$, 14, 20, 24, 30, 40, 50, and 60).
431: \protect\label{fig:fig4}}
432: \end{figure}
433: %
434: we display the product $L^{1.282} B^*$
435: as a function of $1/L$ and see that the behavior is compatible
436: with a large $L$ limit with $O(1/L)$ finite size
437: effects. Direct fits to the form $B^*(L) = u L^{-y_B} (1 + v/L)$
438: give $y_B$s in the range $1.28$ to $1.30$ depending on the
439: points included in the fit.
440:
441: Given the magnetic length, one can perform finite size scaling (FSS) on the
442: magnetization data $m(B,L)$. Since FSS applies to the singular
443: part of an observable, we should have a data collapse according to
444: \begin{equation}
445: \frac{m(B,L)-\chi_1 B}{m(B^*,L)-\chi_1 B^*} = W(B/B^*)
446: \end{equation}
447: $W$ being a universal function, $W(0)=O(1)$ and
448: $W(x)\sim x^{1/\delta}$ at large $x$.
449: Using the value of $\chi_1$ previously determined,
450: we display in Fig.~\ref{fig:fig4}
451: the associated data. The collapse is excellent and we have checked
452: that this also holds when the $J_{ij}$ are drawn
453: from an exponential distribution. Added to the figure is the
454: function $x^{1/\delta}$ to guide the eye ($\delta=1.282$ as
455: predicted by the droplet model).
456:
457:
458: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
459: \paragraph*{Conclusions ---}
460: We have investigated the $2d$ Ising spin glass with Gaussian and
461: exponential couplings
462: at zero temperature as a function of the magnetic field.
463: The magnetization exponent $\delta$ can be measured; previous
464: studies did not find good agreement with the droplet model
465: prediction $\delta=1.282$ because the analytic contributions
466: to the magnetization curve were mishandled, while in this work we
467: found instead $1.28 \le \delta \le 1.32$.
468: We also performed a direct measurement of the magnetic length,
469: obtaining for the associated exponent $1.28 \le y_B \le 1.30$, again in
470: excellent agreement with the droplet prediction.
471: With this length we showed that finite size scaling is
472: realized without going to infinitesimal fields or
473: huge lattices.
474: Finally, we validated the hypotheses underlying
475: the arguments of the droplet model inherent to the
476: in-field case; we find in particular that in the low-field
477: limit the spin clusters
478: that are relevant are compact and have random magnetizations.
479: In summary, by combining improved computational techniques
480: and greater care in the analysis, we have lifted
481: the discrepancy on the magnetic exponents
482: that has existed for over a decade
483: between numerics and droplet scaling.
484:
485: We thank T. Jorg for helpful comments.
486: The computations were performed on the cliot cluster of the
487: Regional Computing Center and on the scale cluster of E.
488: Speckenmeyer's group, both in Cologne.
489: FL has been supported by the German Science Foundation in the projects
490: Ju 204/9 and Li 1675/1
491: and by the Marie Curie RTN ADONET 504438 funded by the EU.
492: This work was supported also by the EEC's
493: %FP6 Information Society
494: %Technologies Programme under contract IST-001935 EVERGROW (www.evergrow.org)
495: %and by the EEC's
496: HPP under contract HPRN-CT-2002-00307 (DYGLAGEMEM).
497: %The LPTMS is an Unit\'e de Recherche de l'Universit\'e Paris-Sud
498: %associ\'ee au CNRS.
499:
500: \bibliographystyle{apsrev}
501:
502: \bibliography{/home/martino/Papers/Bib/references,/home/martino/Papers/Bib/co}
503:
504: \end{document}
505:
506:
507:
508: