ECT* Marco Cristoforetti ECT* Abhishek Mukherjee ECT* Luigi Scorzato INFN/TIFPA Francesco Di Renzo University of Parma PRD86, 074506 (2012) PRD Rapid 88, 051501 (2013) PRD Rapid 88, 051502 (2013) Centre for Information Technology - CIT_irst Centre for Materials and Microsystems - CMM Centre for Italian-German Historical Studies Centre for Religious Sciences - ISR Lattice QCD at finite temperature and density KEK - January 21 - 2014 Quantum field theory on the Lefschetz thimble FBK Research Centres European Centre for Theoretical Studies in Nu Introduction: dense systems Nuclear physics High energy physics Condensed matter Introduction: dense systems ! Studied using Monte Carlo simulations Introduction: dense systems Condensed matter High Tc superconductivity Hubbard model h (HK +HV +Hµ ) Z = Tr e HK = Hµ = t X i (c†i cj + c†j ci ) hi,ji, µ HV = U X (ni" + ni# ) i X✓ i ni" ◆✓ 1 ni# 2 1 2 ◆ . Introduction: dense systems Nuclear physics stellar nucleosynthesis Shell model h (HM F +HVres ) Z = Tr e HM F = X ✏↵ a†↵ a↵ i ↵ HVres 1 X = h↵ |V | 2 ↵ ia†↵ a† a a Introduction: dense systems High energy physics Heavy ion collisions QCD on the lattice Z= Z DU det[M (U )]Nf e SG [U ] = 3 XX n2⇤ µ<⌫ Re tr[1 SG [U ] Uµ⌫ (n)] Introduction: sign problem Hubbard Model Shell Model QCD ! What do they have in common? SIGN PROBLEM Introduction: dense systems Z= DA det[D / +m The sign problem at finite chemical potential the fermionic determinant is complex: standard Monte Carlo methods fails det M(μ) = ¦det M(μ)¦ eiθ µ 0 /2]eSY M [det M(μ)]* = det M(-μ*) Unfortunately we cannot simply neglect the phase of the determinant. Phase quenched theory can be very different from the real world An example of that difference that we will treat later is the Silver Blaze phenomenon Introduction: Lefschetz thimble We want to overcome sign problem for Lattice QCD We must be extremely careful not destroying physics [ Silver Blaze phenomenon ] whatever machinery we use to solve the theory Integration on a Lefschetz thimble Before applying the idea to full QCD we choose to start from something more manageable: we consider here integration on Lefschetz thimbles for the case of a simple 0-dim field theory and the 4-dim scalar field with a quartic interaction Lefschetz thimble on a lattice Saddle point integration the Airy function 1 Ai(x) = 2⇡ Z 1 1 e i ⇣ t3 3 ⌘ +x t dt Re e 1.0 tI i ⇣ t3 3 +x t ⌘ 0.5 0.0 -0.5 tR -1.0 -10 -5 0 5 10 Lefschetz thimble on a lattice Saddle point integration the Airy function 1 Ai(x) = 2⇡ Z complexify 1 e i 1 1 2⇡ ⇣ t3 3 Z ⌘ +x t dt e i ⇣ z3 3 +xz ⌘ dz Complexify the variable t ! tR + i tI Stationary point Steepest descent for the real part of the exponent starting at the stationary point Imaginary part of the exponent is constant integrate on SD 1 i e 2⇡ Z e h ⇣ 3 ⌘i z R i 3 +xz dz tI SR steepest descent tR Lefschetz thimble on a lattice Saddle point integration the Airy function 1 Ai(x) = 2⇡ Z 1 1 e i ⇣ t3 3 Re e ⌘ +x t dt 1.0 i ⇣ t3 3 +x t ⌘ 0.5 Complexify the variable t ! tR + i tI Stationary point Steepest descent for the real part of the exponent starting at the stationary point Imaginary part of the exponent is constant tI steepest descent 0.0 -0.5 -1.0 -10 SR -5 0 5 10 0.5 0.4 0.3 0.2 0.1 tR 0.0 -4 -2 0 2 4 From saddle point to Lefschetz thimble Saddle point integration ! Works extremely well for low dimensional oscillating integrals. ! Usually combined with an asymptotic expansion around the stationary point (sort of perturbative expansion). ! The phase is stationary + important contributions localized = good for sign problem tI SR What about a Monte Carlo integral along the curves of steepest descent tR Lefschetz thimble on a lattice Path integral and Morse theory E. Witten arXiv:1009.6032 (2010) Complexify the degrees of freedom Deform appropriately the original integration path (Morse theory) L C= ( ) R z = x + iy ( ) ( ) ( ) C ( ) ( ) = C ( ) ( ) L for each stationary point pσ the Lσ (thimble) is the union of the paths of steepest descent that fall in pσ at L the thimbles provide a basis of the relevant homology group, with integer coefficients Generalization of the one dimensional SD to n-dim problems is called Lefschetz thimble Lefschetz thimble on a lattice Path integral and Morse theory E. Witten arXiv:1009.6032 (2010) Complexify the degrees of freedom ( ) R ( ) C ( ) = ( ) ( ) C Deform appropriately the original integration path (Morse theory) ( ) z = x + iy steepest descent tI ( ) ( ) L tR # of intersections between steepest ascent and original integration domain steepest ascent Lefschetz thimble on a lattice Lefschetz thimbles and QFT Can we use the thimble basis to compute the path integral for a QFT ? R Q d CR x Q hOi = C xe xd S[ ] xe O[ ] S[ ] hOi = P n P R n J R Q J In principle yes but we have to discuss the details xd Q xe xd S[ ] xe O[ ] S[ ] Lefschetz thimble on a lattice Lefschetz thimbles and QFT R Q d CR x Q hOi = C S[ ] xe x d xe O[ ] S[ ] hOi = P n P R n J R Q J xd Q xe xd S[ ] xe Computing the contribution from all the thimbles is probably not feasible On a Lefschetz thimble the imaginary part of the action is constant but the measure term does introduce a new residual phase, due to the curvature of the thimble O[ ] S[ ] Lefschetz thimble on a lattice Lefschetz thimbles and QFT the residual phase hOi = R J0 R Q J0 xd Q xe xd S[ ] xe O[ ] S[ ] Additional phase coming from the Jacobian of the transformation between the canonical complex basis and the tangent space to the thimble Does it lead to a sign problem? No formal proof but ... dΦ=1 at leading order and <dΦ> 1 are strongly suppressed by e-S there is strong correlation between phase and weight (precisely the lack of such correlation is the origin of the sign problem) In fact this residual phase is completely neglected in the saddle point method Lefschetz thimble on a lattice Lefschetz thimbles and QFT the residual phase hOi = R J0 R Q J0 xd Q xe xd S[ ] xe O[ ] S[ ] Additional phase coming from the Jacobian of the transformation between the canonical complex basis and the tangent space to the thimble Does it lead to a sign problem? Hopefully not but nevertheless the calculation of that phase cannot be avoid! highly demanding in terms of computation power H. Fujii et al JHEP 1310 (2013) 147 Next talk: Y Kikukawa Lefschetz thimble on a lattice Lefschetz thimbles and QFT R Q d CR x Q hOi = C S[ ] xe x d xe O[ ] S[ ] hOi = P n P R n J R Q J xd Q Computing the contribution from all the thimbles is probably not feasible Is it necessary in order to have the correct physics? I will try to convince you that in many cases we can consider only one thimble xe xd S[ ] xe O[ ] S[ ] Lefschetz thimble on a lattice Lefschetz thimbles and QFT choosing the stationary point The study of the stationary points of the complexified theory is mandatory and has to be done on a case-by-case The system has a single global minimum? There are degenerate global minima, that are however connected by symmetries? These cases are good! There is a large number of stationary points that accumulate near the global minimum giving a finite contribution? This case can be bad! There are degenerate global minima, with vanishing probability of tunneling? Lefschetz thimble on a lattice Lefschetz thimbles and QFT choosing the stationary point hOi = P n P R n J R Q J d Q x xe xd S[ ] xe O[ ] S[ ] This is an exact reformulation of the original path integral. For a QFT reproducing the original integral is both impractical and unnecessary Consider the stationary point with the lower value of the real part of the action and with nσ 0 Define a QFT on the thimble attached to this point. If The degrees of freedom are the same The symmetries are the same The same perturbative expansion The same continuum limit By universality this is a legitimate regularisation of the original QFT Lefschetz thimble on a lattice Lefschetz thimbles and QFT choosing the stationary point hOi = P n P R n J R Q J d Q x xe xd S[ ] xe O[ ] S[ ] This is an exact reformulation of the original path integral. For a QFT reproducing the original integral is both impractical and unnecessary Consider the stationary point with the lower value of the real part of the action and with nσ 0 Define a QFT on the thimble attached to this point. If The degrees of freedom are the same The symmetries are the same The same perturbative expansion Universality is not a theorem BUT it is an assumed property studying QFTs on a lattice The same continuum limit By universality this is a legitimate regularisation of the original QFT Lefschetz thimble on a lattice Lefschetz thimbles and QFT choosing the stationary point hOi = P n P R n J R Q J xd Q xe xd S[ ] xe O[ ] S[ ] hOi = Consider the stationary point with the lower value of the real part of the action and with nσ 0 Define a QFT on the thimble attached to this point. If The degrees of freedom are the same The symmetries are the same The same perturbative expansion R J0 R Q J0 xd Q xe xd S[ ] xe O[ ] S[ ] Universality is not a theorem BUT it is an assumed property studying QFTs on a lattice The same continuum limit By universality this is a legitimate regularisation of the original QFT Lefschetz thimble: algorithm Integration on a Lefschetz thimble M. C., F. Di Renzo and L. Scorzato PRD86, 074506 (2012) Is it numerically applicable to QFT s on a Lattice? Before applying the idea to full QCD we choose to start from something more manageable: we consider here integration on the Lefschetz thimbles for the case of a 0 dimensional field theory with U(1) symmetry the four dimensional scalar field with a quartic interaction Lefschetz thimble: algorithm Langevin PRD Rapid 88, 051501 (2013) d R i (⌧ ) d⌧ d I i (⌧ ) d⌧ = S R ( (⌧ )) R + ⌘ i (⌧ ) R (⌧ ) i = S R ( (⌧ )) I + ⌘ (⌧ ) i I (⌧ ) i d⌘i (⌧ ) X = ⌘(⌧ )k @k @j SR d⌧ k projection of the noise on the tangent space Metropolis PRD Rapid 88, 051502 (2013) d i (r) 1 = dr r S i (r) i (n + 1) = i (n) + r S i Other methods (example HMC) H. Fujii et al JHEP 1310 (2013) 147 Next talk: Y Kikukawa Lefschetz thimble: algorithm Langevin PRD Rapid 88, 051501 (2013) We want to compute this: 1 hOi = e Z0 iSI Z boundend from below on J0 Y d xe SR [ ] J0 x constant on J0 O[ ] preserve J0 by construction We can use a Langevin algorithm but how can we stay on the thimble? d R i (⌧ ) d⌧ d I i (⌧ ) d⌧ = S R ( (⌧ )) R + ⌘ (⌧ ) i R (⌧ ) i = S R ( (⌧ )) I + ⌘ i (⌧ ) I (⌧ ) i Need to be projected on the tangent space to J0 Lefschetz thimble: algorithm Langevin PRD Rapid 88, 051501 (2013) preserve J0 by construction We can use a Langevin algorithm but how can we stay on the thimble? d R i (⌧ ) d⌧ d I i (⌧ ) d⌧ = S R ( (⌧ )) R + ⌘ (⌧ ) i R (⌧ ) i = S R ( (⌧ )) I + ⌘ i (⌧ ) I (⌧ ) i Need to be projected on the tangent space to J0 The tangent space at the stationary point is easy to compute ! We can get tangent vectors at any point if we can transport the noise along the gradient flow so that it remains tangent to the thimble L@SR (⌘) = 0 [@SR , ⌘] = 0 d⌘i (⌧ ) X = ⌘(⌧ )k @k @j SR d⌧ k projection of the noise on the tangent space Lefschetz thimble: algorithm Langevin PRD Rapid 88, 051501 (2013) d R i (⌧ ) d⌧ d I i (⌧ ) d⌧ = S R ( (⌧ )) R + ⌘ i (⌧ ) R (⌧ ) i = S R ( (⌧ )) I + ⌘ (⌧ ) i I (⌧ ) i 1 hOi = e Z0 iSI d⌘i (⌧ ) X = ⌘(⌧ )k @k @j SR d⌧ Z Y J0 x d xe SR [ ] O[ ] k Start from the global minimum of the real part of the action, generate a noise vector projected on the thimble and follow the steepest ascent Perform a Langevin step using the noise evolved along the steepest descent and compute the observables Go back along the steepest descent until you are in a region where quadratic approx. is valid and then project the configuration on the thimble Generate a new noise and go back along the steepest ascent * in the plot you have -SR that is the exponent in the integrand of the partition function : where I wrote ascent (descent) you see the opposite in the figure Lefschetz thimble: algorithm Langevin on the Lefschetz thimble vs Complex Langevin Lefschetz Langevin d R i (⌧ ) d⌧ d I i (⌧ ) d⌧ Complex Langevin = S R ( (⌧ )) R + ⌘ i (⌧ ) R (⌧ ) i d = S R ( (⌧ )) I + ⌘ (⌧ ) i I (⌧ ) i d d⌧ I i (⌧ ) d⌧ 3 3 2 2 0 -1 -2 -2 -3 -3 0 1 2 fR 3 4 5 = S R ( (⌧ )) I + ⌘ i (⌧ ) I (⌧ ) i 0 -1 -1 = S R ( (⌧ )) R + ⌘ (⌧ ) i R (⌧ ) i 1 fI The relation between the two approaches should be studied carefully! 1 fI R i (⌧ ) -1 0 1 2 fR 3 4 5 Lefschetz thimble: algorithm Metropolis PRD Rapid 88, 051502 (2013) In the neighbourhood of a critical point 3 i S[ ] = S[ 0 ] + SG [⌘] + O(|⌘| ) 1X 2 SG = k ⌘k 2 Hwk = 0 i + wki ⌘k k G0 is the flat thimble associated to the gaussian action SG k The λ and w are solutions of = X where H is the Hessian k w̄k ηreal are the direction of steepest descent of SR and the equations of steepest descent of η for the Gaussian action can be explicitly solved in term of a new parameter r=e-τ ¯G d⌘k 1 @S 1 = = dr r @⌘k r but for r=ε infinitesimal the Lefschetz and Gaussian thimbles coincide k ⌘k i (✏) = 0 i + ¯ d i 1 @S = dr r@ i X ⌘k / r k ✏ k wki ⌘k k r 2 [✏, 1] Start with a random real η vector, compute Φ(ε) and evolve using steepest descent Lefschetz thimble: algorithm Metropolis PRD Rapid 88, 051502 (2013) In the neighbourhood of a critical point 3 i S[ ] = S[ 0 ] + SG [⌘] + O(|⌘| ) 1X 2 SG = k ⌘k 2 η Hwk = k w̄k ¦η¦/δr + wki ⌘k G0 is the flat thimble associated to the gaussian action SG where H is the Hessian n-dim random vector living on the manifold defined by the eigenvectors of the Hessian computed at the critical point with positive eigenvalues ¦η¦ 0 i k k The λ and w are solutions of = X distance along the thimble number of steps along the steepest descent d i (r) 1 = dr r S i (r) i (n + 1) = i (n) + r S i Lefschetz thimble: algorithm Gaussian thimble Gaussian manifold: flat manifold defined by the directions of steepest descent at the critical point r!0 Lefschetz thimble number of steps along the steepest descent Decreasing δr your manifold get closer and closer to the Lefschetz thimble ¦η¦/δr = N If the action decreases fast away from the stationary point integrating on the Gaussian thimble can be sufficient U(1) one plaquette model PRD Rapid 88, 051502 (2013) Can be seen as the limiting case of the more interesting three-dimensional XY model One dimensional problem: the integration on the Lefschetz thimble can be plotted ACTION S= OBSERVABLE i 2 U +U 1 = i cos J1 ( ) he i = i J0 ( ) i On the thimble P hO( )i = P m m R RJ J d O( )e d ( )e S( ) S( ) SR = sin SI = cos R sinh I R cosh I constant on the thimble U(1) one plaquette model PRD Rapid 88, 051502 (2013) The stationary points are in (0,0) and (π,0) and the thimble can be computed also analytically Exact thimbles: have to pass from the critical point and the imaginary part of the action has to be constant SI (⌧ ) = cos R (⌧ ) cosh CRITICAL POINTS THIMBLES I (⌧ ) = cp SI U(1) one plaquette model PRD Rapid 88, 051502 (2013) The stationary points are in (0,0) and (π,0) and the thimble can be computed also analytically In order to perform the integration on the thimble we use a Metropolis algorithm Gaussian manifold Increasing the accuracy in the integration of the steepest descent we move closer to the exact thimble U(1) one plaquette model PRD Rapid 88, 051502 (2013) OBSERVABLE J1 ( ) he i = i J0 ( ) i There are parameter regions where integration on the Gaussian manifold is sufficiently accurate U(1) one plaquette model PRD Rapid 88, 051502 (2013) 0.1 Residual phase is well under control and is not a source 0.5 1 of additional sign problem (at least in0 this case) hO( )i = P m m R RJ J d O( )e d ( )e This phase should be essentially constant over the portion of phase space which dominates the integral. Residual phase 2 β There is an additional phase coming from the Jacobian of the transformation between the i FIG. 2. Expectation value of e and as athe function ofspace . canonical complex basis tangent S( ) to the thimble S( ) 1 0.8 cos (arg {Jφη e-S }) P 1.5 0.6 0.4 0.2 0 0.0001 Probability measure Gaussian Nτ=6 Nτ=20 Nτ=200 0.001 Log scale 0.01 ⏐Jφη e-S⏐ 0.1 1 λΦ4 theory on the lattice PRD Rapid 88, 051501 (2013) [ , ]= (| | +( µ )| | + | | + µ( ) Continuum action T Silver Blaze problem when T=0 and μ<μc physics is independent from the chemical potential <n>=0 <n>≠0 μc We will study the system at zero temperature μ λΦ4 theory on the lattice PRD Rapid 88, 051501 (2013) [ , [ , ]= ( ]= [( µ + , | +( (| ) +ˆ + ( + +ˆ µ µ )| | + | | + µ( Continuum action Lattice action: chemical potential introduced as an imaginary constant vector potential in the temporal direction ) )) , ) = in term of real fields [ ]= ( = , ) ( = + ( ) , + + ) ( , ) , = cosh µ , , +ˆ + sinh µ , , +ˆ , +ˆ λΦ4 theory on the lattice PRD Rapid 88, 051501 (2013) PHASE QUENCHED [ , O ]= full = (| | +( µ )| | + | | + µ( D |e S |ei O ei O pq = S i ei pq D |e |e Let us try ignoring the phase O pq = D |e S |O D |e S | ) 0 0 λΦ4 theory on the lattice PRD Rapid 88, 051501 (2013) PHASE QUENCHED [ , 0.2 ]= | +( (| µ )| | + | | + µ( 1 L=4 L=6 L=10 L=4 L=6 L=10 0.8 Re < eiφ > 0.15 <n> ) 0.1 0.05 0.6 0.4 0.2 0 0 0 0.2 0.4 1 n = V 0.6 μ ln Z µ 0.8 1 1.2 0 0.2 0.4 0.6 μ 0.8 1 λΦ4 on a Lefschetz thimble PRD Rapid 88, 051501 (2013) On the Lefschetz thimble M. C., F. Di Renzo, A. Mukherjee and L. Scorzato arXiv:1303.7204 (2013) Fields are complexified a ! R a +i I a The integration on the thimble performed with a Langevin algorithm In this case calculations in Gaussian approximation are sufficient to obtain the exact result λΦ4 on a Lefschetz thimble PRD Rapid 88, 051501 (2013) On the Lefschetz thimble 0.2 0.8 pq L=6 pq L=10 AA 64 AA 84 0.6 φ i Re[<e >] 0.15 <n> 1 pq L=6 pq L=10 AA 64 AA 84 0.1 0.05 0.4 0.2 0 0 0 0.2 0.4 0.6 0.8 1 1.2 0.5 µ Silver Blaze ! solving sign problem we have the correct physics 0.6 0.7 0.8 0.9 µ 1 1.1 1.2 λΦ4 on a Lefschetz thimble On the Lefschetz thimble 0.7 0.6 AA 84 WA 84 Comparison with Worm Algorithm 0.5 (courtesy of C. Gattringer and T. Kloiber) <n> 0.4 ! 0.3 1 n = V 0.2 0.1 0 -0.1 0.95 1 1.05 1.1 µ 1.15 1.2 1.25 ln Z µ What about QCD? Consider the stationary point with the lower value of the real part of the action and with nσ 0 Define a QFT on the thimble attached to this point. In QCD this is the trivial vacuum Complexification: a,R Aa⌫ (x) ! Aa,R (x) + iA ⌫ ⌫ (x) a = 1...Nc 1 SU (3)4V ! SL(3, C)4V Covariant derivative: @ rx,⌫,a F [U ] := F [ei↵Ta U⌫ (x)]|↵=0 @↵ rx,⌫,a = rR x,⌫,a irIx,⌫,a I rx,⌫,a = rR + ir x,⌫,a x,⌫,a Equation of steepest descent: d U⌫ (x; ⌧ ) = ( iTa rx,⌫,a S[U ])U⌫ (x; ⌧ ) d⌧ Defining the thimble for gauge theories is possible: substitute the concept of nondegenerate critical point with that of non-degenerate critical manifold All the ingredients are there What about QCD? The symmetries are the same? Yes It can be proved starting from the invariance of the SD equation: d U⌫ (x; ⌧ ) = ( iTa rx,⌫,a S[U ])U⌫ (x; ⌧ ) d⌧ Under gauge transformations it changes as: (Ta rx,⌫,a S[U ]) ! (⇤(x) 1 † ) (Ta rx,⌫,a S[U ])⇤(x)† U⌫ (x) ! ⇤(x)U⌫ (x)⇤(x + ⌫ˆ) 1 The full SD equation is invariant only under the SU(3) subgroup of SL(3,C) this is very interesting: the gauge links are not in SU(3) but the gauge invariance is exactly the same! The perturbative expansion is the same? Yes It is an expansion around the trivial vacuum where the integrand in the partition function has the form of a gaussian times polynomials (let me skip the details) Something else on a Lefschetz thimble Next steps - Theoretical questions (single thimble, residual phase, reflection positivity ... ) ! - 0-dim Φ4 (finished) ! - Chiral random matrix theory ! - Thirring model ! - Hubbard model (repulsive almost done) ! - SU(3) pure gauge with theta term ! - QCD in 0+1 dimension ! -... ! - QCD thank you