Constrained Boltzmann-Gibbs measures and effective potential for glasses in hypernetted chain approximation and numerical simulations 8 9 9 Miguel Cardenas (*), Silvio Franz(**) and Giorgio Parisi(***) 1 (*) Scuola Normale di Pisa n Piazza de Cavalieri 7, 56126 Pisa, Italy a J (**) Abdus Salam International Center for Theoretical Physics 5 Strada Costiera 11, P.O. Box 563, 34100 Trieste (Italy) 1 (***) Universit`a di Roma “La Sapienza” ] Piazzale A. Moro 2, 00185 Rome (Italy) h e-mail: [email protected], [email protected], [email protected] c e m January 1998 - t a t s Abstract . t a By means of an effective potential associated to a constrained equilibrium measure and apt to study m frozen systems, we investigate glassy freezing in simple liquids in the hypernetted chain (HNC) approxi- - mation. Differently from other classical approximations of liquid theory, freezing is naturally embedded d in the HNC approximation. We get a detailed description of the freezing transition that is analogous to n o the one got in a large class of mean-field long range spin glass. We compare our findings with Monte c Carlosimulationsofthesamesystemandconcludethatmanyofthequalitativefeaturesofthetransition [ are captured by the approximated theory. 1 v 5 1 Introduction 5 1 1 Cooled at fixed rate, supercooled liquids stop flowing on observable time scales when the glassy transition 0 temperatureT ismet[1]. Atthatcoolingratedependenttemperaturethesesystemsget outofequilibrium. 8 g 9 Thelargescalemotionofthemoleculesisfrozen,andconsequentlytheentropy(inthesenseofthelogarithm / t of the phase space accessible to the system) is discontinuously reduced. This discontinuity is equal to a the configurational entropy at temperature T and is supposed to vanish if the system could be kept in m g equilibrium down to the point of the ideal glass transition T . This would be observable only for infinitely - 0 d slow cooling, and would be a real thermodynamic transition. n At temperatures smaller then T these systems finds themselves in a region of the configuration space o g c havingvanishingweightintheBoltzmanndistribution1,thevaluesoftheextensivequantities beingfarfrom : v these at equilibrium. One can also expect that they will remain conformationally close to the configuration i y reached at T where flow stopped. On the other hand small scale vibrations of the atoms are free to X g thermalize at the actual temperature T of the thermal bath. This situation, with extreme separation of r a time scales would be naturally described by a statistical ensemble where the slow degrees of freedom are quenched, and the fast degrees of freedom thermalize at temperature T. One can then define a conditional statistical ensemble, as a Boltzmann-Gibbs measure for fixed distance from the point y. The free-energy associated tothis distributionis afunctionoftheconstraineddistance, andisanatural“effective potential” for glassy systems. Thiseffective potential can becomputedwiththereplicamethodeven in nondisordered models, the role of the quenched variables being played by the reference configuration y. This reduces the 1Strictly speaking the in the metastable region the liquid has zero weight, the Boltzmann distribution being concentrated on crystal configurations. In this paperwe simply neglect theexistence of the crystal and we imagine that in the supercooled region the measure is concentrated on liquid configurations. This situation can also be realized by not considering in the partition sum the crystal like configurations or by modifying the potential in such a way that the crystal get an high free energy. 1 problem to the study of the free-energy of a multicomponent mixture, in which an analytic continuation on the number of components has to be performed at the end of the computation. In previous work [2, 3, 4], it was studied the shape of the potential, and the implications for the glass transition in long-range spin glasses, and supported the generality of the picture in numerical simulations of a binary mixture model [4]. In this paper we extend the analysis to models of simple liquids in the HNC approximation, and show how this approximation, devised to study the liquid phase naturally describe glassy freezing [5]. The same would not be true for other classical approximations of liquid theory as the Percus-Yevick or the Mean Spherical approximations. The implementation of the replica formalism for the effective potential in the HNC approximation is similar to the one used by Given and Stell [6] to study liquidsinrandomquenchedmatrices. ItalsobearresemblancewiththeoneusedbyZippeliusandcoworkers to implement random crosslinking in models of vulcanization. The main difference is that, while in these cases the replica method is used to deal with external quenched disorder, in our case we use quenched degrees of freedom to probe the configuration space of systems that freeze even in absence of quenched disorder. The approach of this paper is complementary to the one put forward recently in [5]. There it was shown how, combining HNC and replicas, one could reveal the glassy transition and find the properties of the system below the glass temperature. In this paper we will discuss the nature of the freezing in the HNC approximation finding a scenario very similar to that of mean-field spin glasses. The approximation is constructed in such a way that the critical density automatically coincides with the one obtained in [5]. The advantage of the “effective potential” framework with respect to that of [5] is to make conceptually clear the introduction of the replicas in the theory and to make testable predictions on the behaviour of the system at density less the critical one (as we shall see below) when we introduce a potential among two copies of the system. On the other hand we will see that for high values of the density the simpler approach of this paper, where we neglect replica symmetry breaking, leads to inconsistency, and there one need to resort to the approach of [5] for a coherent theory. The explicit computations are performed for the hard sphere potential. We support our findings with Monte Carlo simulations of the same system. A short account of our result has appeared in [7]. We organize the paper as follows: in section 2 we discuss the construction of the effective potential and we briefly review the results obtained for long-range spin glasses. In section 3 we discuss the potential for simple liquids in the HNC approximation. Section 4 is devoted to the presentation of the theoretical results on the hard sphere system, that in section 5 we compare with the numerical simulations. Finally in section 6 we present some conclusions and perspectives. 2 The effective potential In this section we review the construction of the effective potential [2, 3, 4]. For definiteness we discuss the case of a simple liquid composed by N identical point-like particles in a volume V, described by their coordinates x = (x ,...,x ), and interacting via a pair potential φ(x −x ). Suppose that, undergoing a 1 N i j cooling process from the liquid phase, the system falls out of equilibrium at a temperature T and remains g stuck in a region of the configuration space having vanishing weight in the Boltzmann-Gibbs measure. This commonly happens at the glassy transition of supercooled liquids, where the liquid stops flowing: large scale motion is frozen, while small scale motion of the atoms (vibration) can still equilibrate even below T . g In these conditions the observed values of extensive quantities can be far from their canonical equilibrium values, while keeping the external parameter constant, they do not vary over the laboratory time scale. It is appropriate then to restrict the measure in configuration space to the vicinity of the configuration y reached when crossing T . g Inordertodothatweneed todefinea notion of similarity (or codistance) amongconfigurations, q(x,y), that, with reference to spin glass terminology, we call overlap. The appropriate definition of the overlap dependson theproblem at hand,and ithas to besuch that to similar configurations correspondhigh values of q (with normalization q(x,x) = 1) and to very different configurations values close to zero. In our particle system an appropriate definition can be 1 q(x,y) = w(|x −y |) (1) i j N i,j X 2 where w(r) is a function close to one for r ≤ σr and close to zero for r ≥ σr , with r being the radius 0 0 0 of the particles and σ a number e.g. of the order of 0.3, such that couples of particles at small distances in the two configurations contribute positively to q. Specifically in the unit radius hard sphere problem of section 4 we will use w(r) = θ(r−0.3). As q(x,x) = 1 and q(x,y) ≤ 1 we can define a sort of distance as d(x,y) = 1−q(x,y). In the following we will speak indifferently about the two quantities, remembering that high overlap means small distance and vice-versa. Having now the definition of q(x,y) we can define a restricted Boltzmann-Gibbs distribution as 1 P(x|y) = exp(−βH(x))δ(q(x,y)−q) (2) Z(β,y) where Z(β,y) is the integral over x of the numerator of (2). Three comments are in order. • The value of q that appears in (2) is at this stage arbitrary. However, the system at temperature T will tend to adjust itself and select a given natural distance from the configuration y, according to local free-energy minimization. The selection of q can be well understood in a mean-field picture, and has been discussed in the case of long range spin glasses in [2]. At low temperature one expects metastable states in configuration space. This corresponds to a two-minima structure of the effective potential, withoneminimumatlow q representingthetypicaloverlap amongconfigurationsbelonging to different metastable states, and one minimum at high q representing the typical overlap among configurations in the same metastable state. This last is the q that would be naturally chosen by the system. • Thesecond comment concerns thedependenceof themeasure(2) on thereferenceconfiguration y. At a first sight, as different cooling experiment would produce different configuration y, it would appear that the measure (2) could be hardly of any use. However, y is supposed to be a configuration typical with respect to the Boltzmann-Gibbs probability at temperature T , µ(y) = exp(−β H(y))/Z(β ), g g g andwecanexpecttheextensivequantitiescomputedfrom(2)tobeself-averaging (i.e. y independent) inthethermodynamiclimit. Theroleoftheconfigurationy isanalogousinthisconstructiontotheone of the quenched variables in disordered systems. In this sense, the measure (2) is an implementation of the idea of “self-generated disorder” often advocated for structural glasses [9, 12]. • Thethirdcommentconcerns theselection of thetemperatureT . For thatwedonothave any a-priori g criterion, as it is a quantity that in experiments depends on the cooling rate. We have thought to ′ the temperature of the configuration y, that we will call T in the following, as the glass transition temperature in the purpose of illustrating the physical situation that we have in mind. As the matter ′ of fact our construction is well defined for arbitrary T , and interesting results are obtained even ′ for T = T. In this paper we will limit out analysis to this case using the measure (2) as powerful probe of configuration space. We insist however on the conceptual importance of considering two temperatures, and we will often refer to results obtained in spin glasses for this more complicated case. In the previous discussion and part of the following we have considered the temperature as the onlyexternal parameter. Itis clear thatmutatis mutandis analogous considerations holdforany other control parameter, as the density of section 5. The object on which we will concentrate our attention is the free-energy associated to the distribution (2) that, invoking the self-averaging property we can write as: T 1 ′ ′ V(q,β,β )= − dyexp(−β H(y))log dxexp(−βH(x))δ(q(x,y)−q) (3) N Z(β′) Z (cid:26)Z (cid:27) As the constraint implied by the delta function is global, we can enforce it through a Lagrange multiplier; and considering the quantity T 1 ′ ′ F(ǫ,β,β )= − dyexp(−β H(y))log dxexp(−β[H(x)−ǫq(x,y)]) (4) N Z(β′) Z (cid:26)Z (cid:27) 3 we find that V and F are related by the Legendre transform ′ ′ V(q,β,β )= min F(ǫ,β,β )+ǫq . (5) ǫ (cid:0) (cid:1) For practical purposes it is more convenient to work with F than with V, while the data are more easily interpreted in terms of V. We will pass freely from one representation to the other in the following. Inordertodealwiththeaverage ofthelogarithm in(4)weresorttothereplicamethod. Thisconsists in evaluatingthemomentsZr forintegerr andcomputingaverage thelogarithmfromananalyticcontinuation to non integer r from the formula logZ = limr→0 Zrr−1. Explicitly we can write: r r Zr = −T dx exp(−β′H(x ))/Z(β′) dx ...dx exp −β[ H(x )−ǫ q(x ,x )] . (6) 0 0 1 r a a 0 ! Z Z a=1 a=1 X X wherewehavewritteny = x . Theproblemisreducedtothecomputationofthermodynamicsforamixture 0 of r+1 components in the limit r → 0. Notice the non-symmetric role played by the replica x and the 0 replicas x for a ≥ 1. This implies that while there is symmetry underpermutation of replicas with positive a index there is not symmetry under interchange of the replica x with the others. Although we have r+1 0 replicas, the symmetry of the problem is only S , becoming S only for ǫ → 0. Technically, our approach r r+1 is similar to the one of Goldbart and Zippelius [13] to study vulcanization of rubber and the one of Stell et al. for liquids in random quenched matrices, where also one finds a number of replicas that tends to one. However, in their case real quenched disorder is present, whilein our case we usetheauxiliary configuration y to restrict the Boltzmann-Gibbs measure to small regions of configuration space. Before discussing in the next section the application of the present formalism to simple liquids in the hypernetted chain approximation, let us describe briefly what one can expect for the effective potential ′ when a glassy transition occurs, considering for simplicity the case T = T. In the supercooled phase the diffusion constant becomes lower and lower as the temperature is lowered. The “cage effect” takes place: the molecules get trapped for long times before they can diffuse. When the glass transition is met diffusion is completely stopped, at least on human time scale. The entropy associated to diffusion is lost at the transition. Ergodicity is broken and the configuration space is effectively split into an exponentially large number of practically mutually inaccessible regions. Making the approximation that the time to jump out of these regions is infinite, it is natural to expect the effective potential (2) to have two minima. One corresponding to the typical (low) overlap among configurationsbelongingtodifferentregionsinconfigurationspace,andanothercorrespondingtothetypical overlap (high) of different configurations belonging to the same region. The number of these regions, or metastable states N is related to the configurational entropy Σ by the relation: N = exp(NΣ). The probability of x to be in the same metastable state of y will be in such conditions 1/N = exp(−NΣ). Consequently the relative height of the high q minimum with respect to the low q one has to be equal to TΣ. This picture is realized and it has been discussed in [2, 3] in a large class of long range spin glass model. Theanalysis of these models tells us that the picturehas to berefineda little. To each disconnected region one can associate a free-energy f, with an energetic part and an entropic part. Defining Σ(T,f) the logarithm of the number of these regions as a function of f, one finds that, at low enough temperature, Σ is different from zero in a finite temperature dependent interval I(T) = [f (T),f (T)]. The states that m M dominate the partition function at temperature T are such if the quantity F = f −TΣ(T,f) (7) ′ is minimum [14]. The study of the effective potential for T 6= T shows that individual states do not disappear when the temperature is changed, but they remain stable for large ranges of temperatures [2, 15, 16]. As thetemperatureis lowered, states with lower andlower Σareselected in(7) until, foratemperature T with Σ = 0 are reached and the partition starts to be dominated by the lowest states. s 3 The HNC approach Let us now start the discussion of the implementation of the HNC approximation in the approach outlined in the previous section. As we stressed the use of the replica method reduces the problem of the evaluation 4 of the effective potential to the one of an r +1 component liquid mixture, in which there is a privileged component with which all the other replicas interact via the potential n n 1,N −Nǫ q(x ,x )= −ǫ w(x0−xa) (8) a 0 i j a=1 a=1 i,j X XX The basic quantities of the theory are the pair correlation functions among replicas ρ ρ g (x,y)+ρ δ δ(x−y) = δ(xa −x)δ(xb −y) (9) a b ab a ab i j i,j X ApartialresummationoftheMayer expansionallowstowriteaself-consistentexpressionforthefree-energy [17, 18, 5]: r −2βF = ddx ρ ρ g (x)[logg (x)−1+β φ(x)δ ] HNC a b ab ab a ab Z a,b=0 X r +2βǫ ρ ρ g (x)w(x)+Tr L(ρh) (10) 0 a 0a a=1 X where h = g −1 is the connected correlation function and L is an operator in physical and replica space, ab ab defined by L(u) = u−u2/2−log(1+u). (11) ′ ′ We have also put β = β , ρ = ρ and β = β, ρ = ρ for a> 1. The free-energy (10) has to be extremized 0 0 a a with respect to the g ’s and the terms of order r have to be extracted. The extremum conditions can be ab cast in the form: g (x,y) = exp(−β φ(x,y)δ +{δ (1−δ )β +δ (1−δ )β }ǫw(x)+h (x)−c (x)) (12) ab a ab 0a 0b b 0b 0a a ab ab with g and c related by the Ornstein-Zernike relation r h (x)= c (x)+ dy h (x−y)ρ c (y) (13) ab ab ac c cb c=0Z X As usual in the replica method one needs a parameterization of the matrix g that allows the analytic ab continuation tor → 0. Onthebasisof symmetryconsiderations analogous totheoneof[2], onecanpropose the structure: g a= b = 0 00 g = g a= 0, b 6= 0 or b = 0, a6= 0 (14) ab  10 ∗  gab a,b 6= 0 ∗ The r × r sub-matrix gab can be either replica symmetric or have an ultrametric structure [19], in the following we will limit ourselves to the replica symmetric structure g∗ = g11 a= b (15) ab g a 6= b. ( 12 which coincides with the choice done by Given and Stell [6]. In the case of the p-spin model, the replica symmetric choice gives the correct result for the effective potential in the high and in the low q regions and in particular around the minima. However it was found an intermediate q region where “one step replica ∗ symmetry breaking” in the matrix g was necessary to compute correctly the effective potential. Although ab we do not explore here the possibility of solutions with a structure more complicated then (15) and we limit ourself to the study of the high and low q parts of the effective potential, we warn the reader that replica symmetry breaking has also to be expected in this case for intermediate q. The interpretation of the different elements of the g matrix with the replica symmetric ansatz is ab straightforward. The element g represents the pair correlation function of the free system; as such the 00 equation determining itdecouples fromthe other components in thelimit r → 0. In turn,g represents the 11 5 paircorrelation functionofthecoupledsystem. g isthepaircorrelationamongthequenchedconfiguration 10 and the annealed one, while g represents the correlation between two systems coupled with the same 12 quenched system. This is the analogous of the Edwards-Anderson order parameter in disordered systems, and represents the long time limit of the time dependent autocorrelation function at equilibrium [2]. Straightforward algebra shows that the equations (13) reduce to the ones proposed in [6]. In particular one finds that, as it should, the equation for g , describing the correlation function of the quenched replica 00 decouples from the other and coincides with the usual HNC equation in absence of replicas. h (x) = c (x)+ ddy [ρ h (x−y)c (y)+rρ h (x−y)c (y)] 00 00 0 00 00 1 11 11 Z h (x) = c (x)+ ddy [ρ h (x−y)c (y)+ρ h (x−y)(c (y)−c (y)(1−r))] 10 10 0 00 10 1 10 11 12 Z h (x) = c (x)+ ddy [ρ h (x−y)c (y)+ρ (h (x−y)c (y)−h (x−y)c (y)(1−r))] 11 11 0 10 10 1 11 11 12 12 Z h (x) = c (x)+ ddy [ρ h (x−y)c (y) 12 12 0 10 10 Z + ρ (h (x−y)c (y)+h (x−y)c (y)−h (x−y)c (y)(2−r))] (16) 1 11 12 12 11 12 12 The overlap can be expressed in terms of the correlation function g , and reads 10 ∞ q = ρ dxw(x)g (x) = 4πρ r2w(r)g (r). (17) 10 10 Z Z0 According to the discussion of the previous section, we will associate glassy behavior to non convexity of the function V(q), and in particular to the existence of multiple solutions q(ǫ) for ǫ → 0. In the liquid phase we can expect instead q(ǫ) to be a single value function and a convex effective potential with a minimum at q = q = ρ dxw(x), corresponding to the absence of any structure in g , i.e. g (x) = 1 for all x. A 0 10 10 strong coupling ǫ, attracting the system towards the configuration y will force a structure in the g , which R 10 will have a higher peak in x = 0 the higher is the coupling. Similarly g will acquire a structure: if two 12 different systems are similar to the configuration y they will also besimilar to each other. When the system freezes, it will exist a solution to the HNC equations in which g and g will have a structure even for 10 12 small and vanishing ǫ. The value of ǫ at which we find a solution with non-zero g coincides with dynamical critical density 12 of [5] and it correspond to the phase transition point in a mode-coupling approach [1]. Indeed in this mean field approach ergodicity starts to be broken exactly at this point. 4 Results for HNC hard spheres Inthissectionwediscussthepicturecomingfromtheintegrationoftheequations(12,13)inthreedimension. We present systematic data in the case of the hard sphere potential 1 r < 1 1 r < 0.3 φ(r) = and with w(r) = (18) 0 r > 1 0 r > 0.3 ( ( The hard sphere potential has been chosen for practical convenience, the glassy transition picture that will emerge can be strongly expected to be very general. We have verified in non systematic investigations that the same picture indeed holds for soft sphere systems with φ(r) = r−12. The value 0.3 that appears in the definition of w has obviously nothing fundamental, and we have checked that the picture is insensitive to its precise value. The hard sphere model has no temperature, and the control parameter is the density. Numerical work report a glassy phase for values of the density higher then 1.15. We have solved the saddle point equation (12,13) by iteration for various values of the density and the coupling. For fixed density we start the integration of the equation at low (respectively high) coupling ǫ whereweknowthesolution andweincrease(respectively decrease)itatsmallsteps. Inthisway wecanfind the curves of q as a function of ǫ, and reconstruct from (10) the effective potential V(q). We checked that q ′ ′ up to a constant V(q) = dq ǫ(q ). In the low density region we were able in this way to fully reconstruct the shape of the potential. For higher densities, we could just reconstruct in this way the high and the low R 6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 Figure 1: The behavior of q as a function of ǫ for HNC hard spheres for ρ = 1.14,1.17,1.19,1.20. For high enough density q is a multivalued function of ǫ. We have shown only a portion of the curve in the region where it is multivalued. For graphical transparency in this and the next figure we have joined with a line the branches corresponding to the same density. parts of the effective potential. This is however enough to get a fully detailed picture of the freezing in the system. In figure 1 we present the curves of q as a function of ǫ for various values of the density, while in figure 2 we plot the corresponding curves V(q). At low density ǫ is a monotonic function of q, testifying ergodic behavior of the system. The potential V(q) is convex and has a single minimum for ǫ = 0, where the value of the overlap is q =ρ4π(0.3)3/3 = ρ×0.113, corresponding to g (x) = 1 for all x. 0 10 Interesting behavior appears for densities higher or equal to ρ ≈ 1.14. At ρ the function q(ǫ) begins cr cr to be multivalued, the potential looses the convexity property and a phase transition among a low q and a high q phase can be induced by a coupling. The point (ρ ,ǫ ), with ǫ = 0.305 is a critical point of cr cr cr second order phase transition, from which it departs a first order phase transition line ǫ (T) (fig. 3). The tr term −Nǫq(x,y) in the Hamiltonian implies an energetic advantage for the configurations x close to y and induces a transition between a high q “confined” phase with high energy and a low q “deconfined” phase with high entropy. The transition line tells that generic equilibrium configurations lie in metastable states for densities higher that ρ . For ρ < ρ < ρ = 1.17 the metastable states have a finite life, and a coupling c cr c ǫ ≥ ǫ (T) is needed to stabilize them. At ρ a minimum develops in the potential, and the metastable tr c states have an infinite time life. The equation q = q(ǫ = 0) has more then one solution. It has been shown by explicitcalculation in[2]thatasecond minimumin theeffective potential implies thatintheequilibrium dynamics the system remains confined in a region with a large overlap with the initial state. In figure 4 we plot the function g (r) for ρ = 1.20 and various values of ǫ in the high q and the low q solutions. We see 10 how the low q solution has little structure (g ≈ 1) while the high q solution has a very pronounced peaks 10 for integer values of r. In ordinary cases multiple minima in the effective potential as a function the order parameter signal the presence of different (stable or metastable) phases with different qualitative characteristics (e.g. liquid and gas). Here the implications of the two minima structure are different. The appearance of the secondary minimumsignalsthebreakingoftheergodicity, i.e. thesplitofthesupportoftheBoltzmann-Gibbsmeasure into many, mutually inaccessible, regions. It is easy to realize the link among the two minima structureand such non-ergodic situation. If we suppose that the different region have typically all the same distance, one 7 0.1 0.08 0.06 0.04 0.02 0 -0.02 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 Figure 2: The effective potential for HNC hard spheres. From to to bottom ρ = 1.0,1.14,1.17,1.19,1.20. For low density, high up in the liquid phase the potential is convex. In the glass phase two minima are present. 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 1.13 1.14 1.15 1.16 1.17 1.18 1.19 1.2 1.21 Figure 3: Phase diagram in the plane ǫ−ρ. A first order transition line terminating in a critical point separates a low q from a high q phase. 8 1000 100 10 10 g 1 0.1 0.01 0.1 1 10 100 r Figure 4: The function g (r) for ρ= 1.20 and ǫ = 0.10,0.15,0.20. The three curves on the top correspond 10 to the high q solution, the three curves on the bottom to the low q one. shall have a minimum corresponding to that distance, and another minimum corresponding to the typical distanceamongconfigurationsinthesameregion. Inthisperspectivetheminimaaredifferentmanifestation of the same phase. Coherently with this picture, the internal energy in the two minima should bethe same. Obviously this last sentence has no meaning for hard spheres where the internal energy is not defined, but it can be easily checked in soft sphere systems. The difference of height between the two minima, ∆V can also be understood in this perspective as being due to the fact that the number of regions in which the configuration space has split (N)is exponentially large in the number of particles N = exp(NΣ), and each region carry a vanishing weight in the measure. In that conditions, chosen y in a region at random, the probability that x falls in the same region, which should be exp(−β∆V) is equal to N−1 = exp(−NΣ). We find then that Σ, to be identified with the configurational entropy, is related to ∆V by2 ∆V = TΣ. (19) We have then a method to compute the configurational entropy, that we plot in figure 5. An equivalent method has been proposed in [20]. We see that Σ is a decreasing function of the density and vanishes at a density ρ ≈ 1.203,3 according the scenario of Gibbs and Di Marzio of the glass transition, and analogously s to long range spin glasses. At each value of the density one choose these states such that the total balance between f the internal free-energy of the region and the configurational entropy is such to minimize the total free-energy. Above ρ the equations we are considering give the clearly unacceptable result of a negative configura- s tional entropy, and the approach must be modified. Previous experience in spin-glasses tells us that the paradoxical behavior has to be ascribed to an incorrect description of the quenched replica y. For ρ > ρ s the lowest f states are chosen. These carry a finite Boltzmann weight and a correct description has to take this into account. In the correct HNC approach to the high density regime the quenched replica should be described by the replica formulation of M´ezard and Parisi with replica symmetry breaking [5]. The 2the previous considerations should be modified for T′ = T or ρ′ = ρ. In that case the secondary minimum reflects the 6 6 properties of thestates of equilibrium at theprimed values of theparameters when “followed” at the non primed values. 3The values of ρc and ρs are compatible with those found in [5], indeed the potential method reproduces the results of replica symmetry breaking for the static and the dynamiccritical density. 9 0.05 0.045 0.04 0.035 0.03 0.025 0.02 0.015 0.01 0.005 0 1.165 1.17 1.175 1.18 1.185 1.19 1.195 1.2 1.205 Figure 5: The configurational entropy Σ as a function of ρ. HNC approximation, originally devised to study liquids at equilibrium, naturally embeds glassy behavior in a glassy transition scenario completely analogous to the one of disordered models with “one step replica symmetry breaking”. In many senses the picture is genuinely mean-field like. In fact, it has been stressed many times that in realsystemsmetastablestates withinfinitelifetimedonotexist, andamechanismshouldrestoreergodicity between ρ andρ . Moreover itis easy to realize that theeffective potential we have definedmustbeconvex c s beyond mean field. This can be seen constructing configurations with overlap inhomogeneous in space and with a free energy lower or equal to the convex envelope of the potential. We expect however a reflex of the mean-field structure in real systems. What it is seen as a sharp transition in mean field can still be observed as a crossover in real life and some of the prediction of mean field can be expected to hold in finite dimensional systems. In particular the existence of a first order transition line in the plane ρ − ǫ, which depends on the existence of metastable states, regardless if their life is finite or infinite we expect to hold in real (or realistic) systems. In next section we will submit that to test. Beforeleavingthissectionletusremarkthat,howweshowinsomedetailintheappendix,thepossibility of a glass transition, associated to nontrivial g (x) for ǫ → 0 is excluded by other classical approximations 10 of liquid theory, the Percus-Yevick approximation and the Mean Spherical Approximation. 5 Numerical Simulations In order to test the predictions of the transition scenario of the previous section we have performed Monte Carlo simulations of a system of hard spheres in three dimension coupled with a quenched configuration. We have done simulations with a number of particles N ranging from 256 to 1024 and we have not observed any significant dependence on the volume. To generate the quenched equilibrium configurations at fixed density we start of N particles of zero radius in a box with periodicboundaryconditions, and we letthe radii grow until two particle do not get in contact. At that point we make a Monte Carlo sweep, i.e. we move the particles of random amount and we accept the change if two spheres do not overlap; the size of the proposed move is fixed is such a way to have 0.4 average acceptance. We iterate the procedure until the desired density is reached. The volume and the radius (r) are at the point rescaled in order to have r = 1. We thermalize then the system for 4000 Monte 10