density of states approach with ffa for field theories...
TRANSCRIPT
Density of states approach with FFA for fieldtheories with a complex action problem
Mario GiulianiChristof Gattringer, Pascal Törek
Karl Franzens Universität Graz
Mario Giuliani (Universität Graz) Graz, 11th October 2015 1 / 37
Table of contents
1 Sign Problem: a brief review
2 Density of States: general introduction
3 Density of States for SU(3)-spin system
4 Density of States for U(1) with a topological term
5 Conclusions
Mario Giuliani (Universität Graz) Graz, 11th October 2015 2 / 37
Sign Problem: a brief review
Quantum field theory is based on two quantities:
Z =
∫D[C ]eS[C ] 〈O〉 =
1Z
∫D[C ]O(C )eS[C ]
Probabilistic interpretation: p(C) = eS[C]
* Monte-Carlo: to produce a set of M configurations {ci} with probability p(C)
* 〈O〉 ≈ O = 1M
M∑i=1O(ci )
Sign problem: Chemical potential µ 6= 0 or a topological term:
* S acquires an imaginary part
* Factor eS[C ] cannot be used as a probability weight
Mario Giuliani (Universität Graz) Graz, 11th October 2015 3 / 37
Sign Problem: a brief review
Different approaches to solve this problem:
– Reweighting
– Expansion methods
– Stochastic differential equations
– Mapping to dual variables =⇒ Complete Solution
– Et cetera...
Density of states approach:
– Density of states (DoS). An approach developed by K. Langfeld, B. Lucini, A.Rago. LLR Local Linear Relaxation [1–4]
– Different method FFA Functional Fit Approach [5, 6]
Mario Giuliani (Universität Graz) Graz, 11th October 2015 4 / 37
References:
Kurt Langfeld et al. “An efficient algorithm for numerical computations of continuous densities ofstates”. In: (2015). arXiv: 1509.08391 [hep-lat].
Kurt Langfeld e Biagio Lucini. “Density of states approach to dense quantum systems”. In: Phys.Rev. D90.9 (2014), p. 094502. DOI: 10.1103/PhysRevD.90.094502. arXiv: 1404.7187 [hep-lat].
Kurt Langfeld e Jan M. Pawlowski. “Two-color QCD with heavy quarks at finite densities”. In:Phys. Rev. D88.7 (2013), p. 071502. DOI: 10.1103/PhysRevD.88.071502. arXiv: 1307.0455[hep-lat].
Kurt Langfeld, Biagio Lucini e Antonio Rago. “The density of states in gauge theories”. In: Phys.Rev. Lett. 109 (2012), p. 111601. DOI: 10.1103/PhysRevLett.109.111601. arXiv: 1204.3243[hep-lat].
Christof Gattringer e Pascal Törek. “Density of states method for the Z3 spin model”. In: Phys.Lett. B747 (2015), pp. 545–550. DOI: 10.1016/j.physletb.2015.06.017. arXiv: 1503.04947[hep-lat].
Ydalia Delgado Mercado, Pascal Törek e Christof Gattringer. “The Z3 model with the density ofstates method”. In: PoS LATTICE2014 (2015), p. 203. arXiv: 1410.1645 [hep-lat].
Mario Giuliani (Universität Graz) Graz, 11th October 2015 5 / 37
Density of States Method
Density of states
Z =
∫D[ψ]e−S[ψ] 〈O〉 =
1Z
∫D[ψ]O[ψ]e−S[ψ]
In the density of states approach we divide the action into two parts:
S [ψ] = Sρ[ψ] + cX [ψ]
* Sρ[ψ] and X [ψ] are real functionals of the fields ψ
* Sρ[ψ] is the part of the action that we include in the weighted density ρ
* Here c is purely imaginary: c = iξ
Mario Giuliani (Universität Graz) Graz, 11th October 2015 6 / 37
DoS: toy models
Examples discussed here:
Effective Polyakov loop model:
Sρ[P] = Re(S [P])
X [P] = Im(S [P])/ξ
ξ = 2k sinh(µ)
2D U(1) lattice gauge theory with a topological term
Sρ[U] = SG [U] (gauge action)
X [U] = Q[U] (discretization of the topological charge)
ξ = θ
Mario Giuliani (Universität Graz) Graz, 11th October 2015 7 / 37
Density of States Method
The weighted density is defined as:
ρ(x) =
∫D[ψ]e−Sρ[ψ]δ(X [ψ]− x)
Using ρ(x) we can write Z and 〈O〉s:
Z =
∫ xmax
xmin
dx ρ(x) e−iξx 〈O〉 =1Z
∫ xmax
xmin
dx ρ(x) e−iξxO[x ]
Usually there is a symmetry ψ −→ ψ′ such that:
Sρ[ψ′] = Sρ[ψ], X [ψ′] = −X [ψ],
∫D[ψ′] =
∫D[ψ]
The symmetry implies that Z is real
Mario Giuliani (Universität Graz) Graz, 11th October 2015 8 / 37
Density of States Method
The symmetry also implies that ρ(x) is an even function ρ(x) = ρ(−x):
ρ(−x) =
∫D[ψ]e−Sρ[ψ]δ(X [ψ] + x) =
∫D[ψ′]e−Sρ[ψ
′]δ(−X [ψ′] + x) = ρ(x)
The partition function simplifies to:
Z = 2∫ xmax
0dx ρ(x) cos(ξx)
... and for the observables O:
〈O〉 = 2∫ xmax
0dx ρ(x) {cos(ξx)Oeven[x ]− i sin(ξx)Oodd [x ]}
Where Oeven = O[x]+O[−x]2 and Oodd = O[x]−O[−x]
2
Mario Giuliani (Universität Graz) Graz, 11th October 2015 9 / 37
First example
SU(3) spin system with chemical potential
Natural evolution of the Z3 model: from discrete d.o.f to continuous
See Pascal Törek Master thesis for the discussion of the Z3 model
Mario Giuliani (Universität Graz) Graz, 11th October 2015 10 / 37
SU(3)-spin model
SU(3) spin model is a 3D effective theory for heavy dense QCD
Relevant d.o.f. is the Polyakov loop P(n) ∈ SU(3) (static quark source at n)
The model has a real and positive dual representation ⇒ reference data
We have an action:
S [P] = −τ∑n
3∑ν=1
[TrP(n) TrP(n+ν)†+c .c .
]−k
∑n
[eµ TrP(n) +e−µ TrP(n)†]
The action depends only on the trace ⇒ simple parametrization
P(n) =
e iθ1(n) 0 00 e iθ2(n) 00 0 e−i(θ1(n)+θ2(n))
Mario Giuliani (Universität Graz) Graz, 11th October 2015 11 / 37
Imaginary part of the action
Term with chemical potential can be decomposed into real and imaginary parts:
k∑n
[eµ TrP(n) + e−µ TrP(n)
]=
2k cosh(µ)∑n
Re[TrP(n)
]+ i2k sinh(µ)
∑n
Im[TrP(n)
]Definition:
X [P] =∑n
Im[TrP(n)
]=∑n
[sin(θ1(n)) + sin(θ2(n))− sin(θ1(n) + θ2(n))
]
X [P] ∈ [− 3√3
2 V ; 3√3
2 V ]≡
[−xmax , xmax ]
θ1
θ2
Mario Giuliani (Universität Graz) Graz, 11th October 2015 12 / 37
Reduced Haar Measure
The partition function is:
Z =
∫D[P]e−S[P] , D[P] =
∏n
dP(n)
dP(n) = sin2(θ1(n)−θ2(n)
2
)sin2(θ1(n)+2θ2(n)
2
)sin2(2θ1(n)+θ2(n)
2
)dθ1(n)dθ2(n)
θ1
θ2
Mario Giuliani (Universität Graz) Graz, 11th October 2015 13 / 37
Simplification of the partition function
Charge conjugation symmetry:
θi (n)∗−→ −θi (n) Sρ[P]
∗−→ Sρ[P]
dP(n)∗−→ dP(n) X [P]
∗−→ −X [P]
Z =∫D[P]e−Sρ[P]−iξX [P] ∗
=⇒∫D[P]e−Sρ[P]+iξX [P]
=⇒ Z =∫D[P]e−Sρ[P] cos(ξX [P])
Mario Giuliani (Universität Graz) Graz, 11th October 2015 14 / 37
Definition of density of states
We define the weighted density of states:
ρ(x) =
∫D[P] e−Sρ[P] δ(x − X [P]) x ∈ [−xmax , xmax ]
Symmetry P(n)→ P(n)∗ implies ρ(−x) = ρ(x)
This simplifies the partition function:
Z =
xmax∫−xmax
dx ρ(x) cos(2k sinh(µ)x) = 2
xmax∫0
dx ρ(x) cos(2k sinh(µ)x)
〈O[X]〉 =
2Z
xmax∫0
dx ρ(x)[OE [x ] cos(2k sinh(µ)x) + i OO [x ] sin(2k sinh(µ)x)
]
Mario Giuliani (Universität Graz) Graz, 11th October 2015 15 / 37
Parametrization of the density ρ(x)
Ansatz for the density: ρ(x) = e−L(x), normalization ρ(0) = 1⇒ L(0) = 0
We divide the interval [0, xmax ] into N intervals n = 0, 1, . . . ,N − 1.
L(x) is continuous and linear on each of the intervals, with a slope kn:
0
1
2
3
4
5
∆0 ∆1∆2 ∆3 ∆4∆5 ∆6 ∆N−2 ∆N−1
k0
k1k2
k3
k4 k5k6
kN−2
kN−1
xmax
L(x)
Mario Giuliani (Universität Graz) Graz, 11th October 2015 16 / 37
Determination of the slopes kn
How do we find the slopes kn?Restricted expectation values which depend on a parameter λ ∈ R:
〈〈O〉〉n(λ) =1
Zn(λ)
∫D[P] e−Sρ[P]+λX [P]O
[X [P]
]θn[X [P]
]Zn(λ) =
∫D[P] e−Sρ[P]+λX [P] θn
[X [P]
]θn[x]
=
{1 for x ∈ [xn, xn + 1]
0 otherwise
Update with a restricted Monte Carlo
Vary the parameter λ to fully explore the density
Mario Giuliani (Universität Graz) Graz, 11th October 2015 17 / 37
Functional Fit Approach FFA
Expressed in terms of the density:
Zn(λ) =
xmax∫−xmax
dx ρ(x) eλx θn[x]
=
xn+1∫xn
dx ρ(x) eλx = c
xn+1∫xn
dx e(−kn+λ)x
= ce(λx−kn)xn+1 − e(λx−kn)xn
λ− kn
For computing the slopes we use as observable X [P]:
〈〈X [P]〉〉n(λ) =1
Zn(λ)
xn+1∫xn
dx ρ(x) eλx x =∂
∂λln[Zn(λ)
]
Mario Giuliani (Universität Graz) Graz, 11th October 2015 18 / 37
Functional Fit Approach FFA
Explicit expression for restricted expectation values:
1∆
[〈〈X [P]〉〉n(λ)− n∆
]− 1
2= h((λ− kn)∆n
)h(r) =
11− e−r
− 1r− 1
2
Strategy to find kn:
1 Evaluate 〈〈X [P]〉〉n(λ) for different values of λ
2 Fit these Monte Carlo data h((λ− kn)∆n)
3 kn are obtained from simple one parameter fits
All Monte Carlo data are used in the process
Mario Giuliani (Universität Graz) Graz, 11th October 2015 19 / 37
Properties of the h(r) function
The universal function h(r) behaves like that:
-0.5-0.4-0.3-0.2-0.1
00.10.20.30.40.5
-15 -12 -9 -6 -3 0 3 6 9 12 15
h((λ−
k n)∆
n)
λ
kn = 3
∆n = 1∆n = 2∆n = 4
Mario Giuliani (Universität Graz) Graz, 11th October 2015 20 / 37
Overview on the Monte Carlo
λ allows to explore the density:Example: 83, N = 120, n = 70
λ = −1, 30000 steps λ = 10, 30000 steps
X [θ1, θ2]
Haar(θ1, θ2)eX [θ1,θ2]
Mario Giuliani (Universität Graz) Graz, 11th October 2015 21 / 37
Fit of slopes ⇒ density ρ(l)
Example: 83, k = 0.005, µ = 0.0
λ4− 2− 0 2 4
21)
n
λ(n
>>
F[P
]<
<∆1
0.5−
0.4−
0.3−
0.2−
0.1−
0
0.1
0.2
0.3
0.4
0.5
0
10
20
30
40
50
60
70
80
90
100
110
120
130
140
150
160
170
180
190
200
210
220
230
240
-12000
-10000
-8000
-6000
-4000
-2000
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400ln(ρ(x))
x
τ=0.025τ=0.050τ=0.075τ=0.100τ=0.125τ=0.150
kn L(x) ρ(x) = e−L(x)
τ = 0.075
Mario Giuliani (Universität Graz) Graz, 11th October 2015 22 / 37
Observables
1 Particle number density n:
n =1V
12k
∂
∂ sinh(µ)lnZ =
1V
2Z
xmax∫0
dx ρ(x) sin(2k sinh(µ)x) x
2 ... and the corresponding susceptibility χn:
χn =12k
∂
∂ sinh(µ)n
=1V
{2Z
xmax∫0
dx ρ(x) cos(2k sinh(µ)x) x2 +( 2Z
xmax∫0
dx ρ(x) sin(2k sinh(µ)x)2}
Mario Giuliani (Universität Graz) Graz, 11th October 2015 23 / 37
Comparison with dual approach
Particle number density n
-0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0 0.5 1 1.5 2 2.5 3 3.5
n
µβ
Lattice:83, τ=0.066, k=0.005Density of StatesDual Formulation
For dual results see Y.D.Mercado, C. Gattringer Nuclear Physics B862 (2012) 737-750We find good agreement for chemical potential up to µβ ≈ 3
Mario Giuliani (Universität Graz) Graz, 11th October 2015 24 / 37
Comparison with dual approach
Susceptibility χn
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0 0.5 1 1.5 2 2.5 3 3.5
χn
µβ
Lattice:83,τ=0.066,k=0.005
Density of StatesDual formulation
We find a good agreement for chemical potential up to µβ ≈ 2.5
We need to reduce the size of the intervals ∆n to explore larger µβ
Mario Giuliani (Universität Graz) Graz, 11th October 2015 25 / 37
Second example
2D U(1) LGT with a topological term
One of the simplest system with a topological term
We have analytical results for this system
Mario Giuliani (Universität Graz) Graz, 11th October 2015 26 / 37
U(1) LGT in 2D with a topological term
The partition function is:
Z =
∫D[U]e−SG [U]−iθQ[U]
Where:
SG = −β2
∑n
[Up(n) + U∗p (n)], Q[U] =1i4π
∑n
[Up(n)− U∗p (n)]
The DoS is defined as:
ρ(q) =
∫D[U]e−SG ρ(q − Q[U])
The topological charge is bounded:Q[U] = 1
i4π
∑n[Up(n)− U∗p (n)] = 1
i4π
∑n 2 i ImUp(n) = 1
2π
∑n ImUp(n)
Q[U] ∈ [−Qmax ,Qmax ], Qmax = L1L22π
Mario Giuliani (Universität Graz) Graz, 11th October 2015 27 / 37
U(1): Partition function and observables
Symmetry: charge conjugation Uµ(n) −→ U ′µ(n) = U∗µ(n)
ρ(−q) = ρ(q)
Z = 2∫ Qmax
0dr ρ(q) cos(θq)
〈O〉 =2Z
Qmax∫0
dq ρ(q)[Oeven[q] cos(θq) + i Oodd[q] sin(θq)
]
Simpler problem:
* small range of vacuum angle θ ∈ [0; 2π]
* oscillation frequency linear in θ (no exponential growth)...
* ...but more challenging structure of the density ρ
Mario Giuliani (Universität Graz) Graz, 11th October 2015 28 / 37
Parametrization of the density ρ(q)
Ansatz for the density: ρ(q) = e−L(q), normalization ρ(0) = 1⇒ L(0) = 0
We divide the interval [0,Qmax ] into N intervals n = 0, 1, . . . ,N − 1
L(q) is continuous and linear on each of the intervals, with a slope kn:
0
1
2
3
4
5
∆0 ∆1∆2 ∆3 ∆4∆5 ∆6 ∆N−2 ∆N−1
k0
k1k2
k3
k4 k5k6
kN−2
kN−1
Qmax
L(q)
Mario Giuliani (Universität Graz) Graz, 11th October 2015 29 / 37
Semi-analytical result
We have a theoretical result for the partition function:
Z (θ) =
q=∞∑q=−∞
I|q|
2
√(β
2
)2
−(θ
4π
)2
(β2 −
θ4π
)2(β2
)2−(θ4π
)2
q2NsNt
We can obtain the density ρ(q) from the partition function:
ρ(q) = 2∫ ∞0
Z (θ) cos(θq)
We can study the continuum limit sending β →∞ and L1 = L2 →∞ with βL1L2
constant
Mario Giuliani (Universität Graz) Graz, 11th October 2015 30 / 37
Theoretical DoS
The DoS that we want to find looks like:
0
0.2
0.4
0.6
0.8
1
0 0.5 1 1.5 2 2.5 3 3.5 4
ρ(q
)
q
20x20, β = 4.032x32, β = 10.24
The peaks can have a big non linear component, so it can be difficult to obtainthe right shapeMario Giuliani (Universität Graz) Graz, 11th October 2015 31 / 37
A different parametrization for ρ(q)
We tried different parametrizations for this system, for example:
We divide the interval [0,Qmax ] into N intervals n = 0, 1, . . . ,N − 1
ρ(q) = e−L(q), L(q) is constant on each of the intervals, with a value hn
The simulation is done on two neighbouring intervals
0
1
2
3
4
5
∆0 ∆1 ∆2 ∆3 ∆N−2 ∆N−1
h1 h2
Qmax
h3 h4 hN−2 hN−1
h2
h3h4
hN−1
L(q)
We don’t find better results using it
Mario Giuliani (Universität Graz) Graz, 11th October 2015 32 / 37
DoS from our simulation
We obtain a reasonable approximation for 202:
0
0.2
0.4
0.6
0.8
1
0 0.5 1 1.5 2 2.5 3 3.5 4
ρ(q
)
q
Lattice:202 β : 4.0 βNsNt
= 0.01
Analytical resultDensity of States
Mario Giuliani (Universität Graz) Graz, 11th October 2015 33 / 37
DoS from our simulation
It’s not so simple to get the same for 322:
0
0.2
0.4
0.6
0.8
1
0 0.5 1 1.5 2 2.5 3 3.5 4
ρ(q
)
q
Lattice:322 β : 10.24 βNsNt
= 0.01
Analytical resultDensity of States
Mario Giuliani (Universität Graz) Graz, 11th October 2015 34 / 37
DoS from our simulation
We can try to improve the statistic, for example of a factor 3:
0
0.2
0.4
0.6
0.8
1
0 0.5 1 1.5 2 2.5 3 3.5 4
ρ(q
)
q
Lattice:322 β : 10.24 βNsNt
= 0.01
Analytical resultDensity of States
We need to use considerable smaller intervalsMario Giuliani (Universität Graz) Graz, 11th October 2015 35 / 37
Observables
Observable: expectation value of the topological charge:
〈Q〉 = − 1L1L2
∂
∂θlnZ =
1L1L2
2Z
Qmax∫0
dq ρ(q) q sin(θq)
Analytical result for the topological charge:
-0.02
-0.01
0
0.01
0.02
0 2 4 6 8 10 12
〈Q〉
θ
Lattice:202 β : 4.0 βNsNt
= 0.01
Analytical result
It’s really difficult to get this sawtooth shapeMario Giuliani (Universität Graz) Graz, 11th October 2015 36 / 37
Conclusions
DoS is a general approach
Crucial: accuracy of ρ
FFA uses restricted Monte Carlo and probes the density with an additionalBoltzmann weight
Tested in SU(3) Spin model: with a modest simulation agreement until µ ≈ 2.5
Interesting applications: Theories with a topological term (Work in progress....)
Mario Giuliani (Universität Graz) Graz, 11th October 2015 37 / 37