Information Geometry and Quantum Fields
Kevin T. Grosvenor
Julius-Maximilians Universität Würzburg, Germany
SPIGL ’20, Les Houches, France July 30, 2020 Collaborators
SciPost Phys. 8 (2020) 073 [arXiv:2001.02683]
Johanna Erdmenger Ro Jefferson
1 32 Sources and References
2 32 Outline
1. Introduction to AdS/CFT 2. Fisher metric, Symmetries and Uniqueness 3. Classical metric on States: Instantons 4. Classical metric on Theories: Ising Model 5. Quantum metric on States: Coherent states 6. Connections and Curvature 7. Conclusions and Outlook
3 32 AdS/CFT Correspondence
Holography: G. ’t Hooft (THU-93/26, ’93), L. Susskind (JMP 36, ’95) AdS/CFT: J. Maldacena (ATMP 2, ’97)
Weak-strong duality.
D R d E e d x φ0(x) O(x) = e−SSUGRA CFT φ(0,x)=φ0(x)
4 32 d,2 Embedding into R : −(X0)2 + (X1)2 + ··· + (Xd)2 − (Xd+1)2 = −L2
d(d−1) 1,...,d Isometries: 2 rotations among X , one rotation between X0 and Xd+1, and 2d boosts mixing X0 and Xd+1 with X1,...,d. Algebra of isometries is so(d, 2) .
Anti-de Sitter Spacetime
2 L2 2 r2 µ ν Poincaré patch: ds = r2 dr + L2 ηµνdx dx . Constant negative curvature, maximally symmetric spacetime. 1 R d+1 √ Vacuum of Einstein-Hilbert S = 16πG d x −g(R − 2Λ) with constant negative Λ (cosmological constant).
5 32 Anti-de Sitter Spacetime
2 L2 2 r2 µ ν Poincaré patch: ds = r2 dr + L2 ηµνdx dx . Constant negative curvature, maximally symmetric spacetime. 1 R d+1 √ Vacuum of Einstein-Hilbert S = 16πG d x −g(R − 2Λ) with constant negative Λ (cosmological constant). d,2 Embedding into R : −(X0)2 + (X1)2 + ··· + (Xd)2 − (Xd+1)2 = −L2
d(d−1) 1,...,d Isometries: 2 rotations among X , one rotation between X0 and Xd+1, and 2d boosts mixing X0 and Xd+1 with X1,...,d. Algebra of isometries is so(d, 2) .
5 32 CFT fields form representations of the conformal algebra.
Conformal Field Theory
Translations Pµ = i∂µ
Rotations/boosts Jµν = 2ix[µ∂ν]
µ Dilatations D = ix ∂µ
2 ν SCT Kµ = i(x ∂µ − 2xµx ∂ν)
Conformal algebra is so(d, 2) .
In 1 + 1d, extended to infinite Virasoro algebra.
6 32 Conformal Field Theory
Translations Pµ = i∂µ
Rotations/boosts Jµν = 2ix[µ∂ν]
µ Dilatations D = ix ∂µ
2 ν SCT Kµ = i(x ∂µ − 2xµx ∂ν)
Conformal algebra is so(d, 2) .
In 1 + 1d, extended to infinite Virasoro algebra.
CFT fields form representations of the conformal algebra.
6 32 AdS/CFT Textbook
7 32 Quantum Information Theory in Holography
Entanglement probes of the bulk.
[0603011 Ryu, Takayanagi; 0705.0016 Hubeny, Rangamani, Takayanagi; 1408.3203 Engelhardt, Wall] Quantum error correction =⇒ bulk reconstruction.
[1411.7041 Almheiri, Dong, Harlow] Tensor networks =⇒ bulk-boundary maps.
[1601.01694 Hayden et. al.] Holographic distance measures.
[pure: 1507.0755 Lashkari, van Raamsdonk; mixed: 1701.02319 Banerjee, Erdmenger, Sarkar]
8 32 Previous Info Geom + Physics work
Ruppeiner (’95,’96) - stat phys and thermo (e.g., ideal gas) Brody & Hook (2008) - stat phys and thermo (e.g., vdW gas) Janke et. al. (2002, 2003) - 1d Ising & Spherical models Dolan, Johnston, Kenna (2002) - Potts model Amari (’97) - Neural networks Ke & Nielsen (2016) - machine learning Heckman (2013) - string theory Blau, Narain, Thompson (2001) - YM instantons in 3+1-d
9 32 R Fisher metric: gij(ξ) ≡ dx pξ(x) ∂i ln pξ(x) ∂j ln pξ(x)
R R q q Fact 1: gij (ξ) = − dx pξ(x) ∂i∂j ln pξ(x) and gij (ξ) = 4 dx ∂i pξ(x) ∂j pξ(x)
Fact 2: gij is positive semi-definite.
Fisher Metric
x x x pξ(x) ≥ 0
p R R dx pξ(x) = 1
ξ injective ∂ ξ −−−−→ pξ smooth, ∂i ≡ ∂ξi
10 32 R R q q Fact 1: gij (ξ) = − dx pξ(x) ∂i∂j ln pξ(x) and gij (ξ) = 4 dx ∂i pξ(x) ∂j pξ(x)
Fact 2: gij is positive semi-definite.
Fisher Metric
x x x pξ(x) ≥ 0
p R R dx pξ(x) = 1
ξ injective ∂ ξ −−−−→ pξ smooth, ∂i ≡ ∂ξi
R Fisher metric: gij(ξ) ≡ dx pξ(x) ∂i ln pξ(x) ∂j ln pξ(x)
10 32 Fisher Metric
x x x pξ(x) ≥ 0
p R R dx pξ(x) = 1
ξ injective ∂ ξ −−−−→ pξ smooth, ∂i ≡ ∂ξi
R Fisher metric: gij(ξ) ≡ dx pξ(x) ∂i ln pξ(x) ∂j ln pξ(x)
R R q q Fact 1: gij (ξ) = − dx pξ(x) ∂i∂j ln pξ(x) and gij (ξ) = 4 dx ∂i pξ(x) ∂j pξ(x)
Fact 2: gij is positive semi-definite.
10 32 dµ2 + 2dσ2 ds2 = σ2
2 2 1 σ2 2 dµ +dσ Alternative example: pµ,σ (x) = gives ds = . πσ (x−µ)2+σ2 2σ2
Comments: 1. It is not true that all free theories have flat Fisher metrics. 2. This hyperbolic space is a priori unrelated to AdS/CFT.
Gaussian Distribution
A simple example (1d Gaussians): Random variable space is X = R.
Parameter space is Ξ = upper half-plane = (µ, σ) ∈ R × R>0
(x−µ)2 1 − p (x) = √ e 2σ2 µ,σ 2π σ
11 32 2 2 1 σ2 2 dµ +dσ Alternative example: pµ,σ (x) = gives ds = . πσ (x−µ)2+σ2 2σ2
Comments: 1. It is not true that all free theories have flat Fisher metrics. 2. This hyperbolic space is a priori unrelated to AdS/CFT.
Gaussian Distribution
A simple example (1d Gaussians): Random variable space is X = R.
Parameter space is Ξ = upper half-plane = (µ, σ) ∈ R × R>0
(x−µ)2 1 − p (x) = √ e 2σ2 µ,σ 2π σ
dµ2 + 2dσ2 ds2 = σ2
11 32 Comments: 1. It is not true that all free theories have flat Fisher metrics. 2. This hyperbolic space is a priori unrelated to AdS/CFT.
Gaussian Distribution
A simple example (1d Gaussians): Random variable space is X = R.
Parameter space is Ξ = upper half-plane = (µ, σ) ∈ R × R>0
(x−µ)2 1 − p (x) = √ e 2σ2 µ,σ 2π σ
dµ2 + 2dσ2 ds2 = σ2
2 2 1 σ2 2 dµ +dσ Alternative example: pµ,σ (x) = gives ds = . πσ (x−µ)2+σ2 2σ2
11 32 Gaussian Distribution
A simple example (1d Gaussians): Random variable space is X = R.
Parameter space is Ξ = upper half-plane = (µ, σ) ∈ R × R>0
(x−µ)2 1 − p (x) = √ e 2σ2 µ,σ 2π σ
dµ2 + 2dσ2 ds2 = σ2
2 2 1 σ2 2 dµ +dσ Alternative example: pµ,σ (x) = gives ds = . πσ (x−µ)2+σ2 2σ2
Comments: 1. It is not true that all free theories have flat Fisher metrics. 2. This hyperbolic space is a priori unrelated to AdS/CFT.
11 32 gij(ξ) = ∂i∂jψ(ξ)
R Fact: gij (ξ) = h(Fi − hFiiξ)(Fj − hFj iξiξ where hfiξ ≡ dx f(x) pξ(x).
Exponential Family
Generalization of the Gaussian (Exponential Family): i pξ(x) = exp C(x) + ξ Fi(x) − ψ(ξ) . R i ψ(ξ) = ln dx exp C(x) + ξ Fi(x) is a normalization factor.
Assume that {C,Fi} are lin. indep. so ξ → pξ is bijective.
12 32 Exponential Family
Generalization of the Gaussian (Exponential Family): i pξ(x) = exp C(x) + ξ Fi(x) − ψ(ξ) . R i ψ(ξ) = ln dx exp C(x) + ξ Fi(x) is a normalization factor.
Assume that {C,Fi} are lin. indep. so ξ → pξ is bijective.
gij(ξ) = ∂i∂jψ(ξ)
R Fact: gij (ξ) = h(Fi − hFiiξ)(Fj − hFj iξiξ where hfiξ ≡ dx f(x) pξ(x).
12 32 Gaussian example:
µ→µ+c −−−−−−−−→ −−−−→x→x+c
(µ,σ)→λ(µ,σ) −−−−−−−−→ −−−−−→x→λx
Symmetries
ξ → ξ˜(ξ) is a symmetry of p if it can be undone by x → x˜(x)
pξ˜(x) dx = pξ(˜x) dx˜
Fact: a symmetry of p is also a symmetry of gij .
13 32 Symmetries
ξ → ξ˜(ξ) is a symmetry of p if it can be undone by x → x˜(x)
pξ˜(x) dx = pξ(˜x) dx˜
Fact: a symmetry of p is also a symmetry of gij .
Gaussian example:
µ→µ+c −−−−−−−−→ −−−−→x→x+c
(µ,σ)→λ(µ,σ) −−−−−−−−→ −−−−−→x→λx
13 32 Non-Uniqueness
Infinitely many distributions give the same information geometry [Clingman, Murugan and Shock 2015] E.g., 1d Gaussian and Cauchy-Lorentz give 2d hyperbolic space Cauchy-Lorentz is not in the exponential family Weirder example in Clingman, Murugan, Shock (2015) of a 3d distribution that gives 2d hyperbolic space, but has none of its symmetries.
14 32 Lessons
Many statistical models S can lead to the same Fisher metric (Gaussian and Cauchy-Lorentz).
FM inherits the symmetries of S but can have more
15 32 R 4 µ 2 4 Simpler model: d x ∂µφ ∂ φ − g φ (Euclidean signature) 2 σ2 Exact solution to e.o.m.: φ~µ,σ(~x) = gσ |~x−~µ|2+σ2
16σ2 |~x−~µ|2−σ2 Lagrangian density: L~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4
L~µ,σ(~x) < 0 for |~x − ~µ| < σ, so not a good prob. distr.
2 2 4 16 σ4 If instead we use −φ ∂ φ − g φ , we get p~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4 and the information geometry is 5d hyperbolic space.
Classical metric on States: Instantons
Blau, Narain, Thompson (2001) - YM instantons in 3+1-d Uses prescription of Hitchin (1990): probability distribution is taken to be the Lagrangian density evaluated on the state.
16 32 16σ2 |~x−~µ|2−σ2 Lagrangian density: L~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4
L~µ,σ(~x) < 0 for |~x − ~µ| < σ, so not a good prob. distr.
2 2 4 16 σ4 If instead we use −φ ∂ φ − g φ , we get p~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4 and the information geometry is 5d hyperbolic space.
Classical metric on States: Instantons
Blau, Narain, Thompson (2001) - YM instantons in 3+1-d Uses prescription of Hitchin (1990): probability distribution is taken to be the Lagrangian density evaluated on the state. R 4 µ 2 4 Simpler model: d x ∂µφ ∂ φ − g φ (Euclidean signature) 2 σ2 Exact solution to e.o.m.: φ~µ,σ(~x) = gσ |~x−~µ|2+σ2
16 32 2 2 4 16 σ4 If instead we use −φ ∂ φ − g φ , we get p~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4 and the information geometry is 5d hyperbolic space.
Classical metric on States: Instantons
Blau, Narain, Thompson (2001) - YM instantons in 3+1-d Uses prescription of Hitchin (1990): probability distribution is taken to be the Lagrangian density evaluated on the state. R 4 µ 2 4 Simpler model: d x ∂µφ ∂ φ − g φ (Euclidean signature) 2 σ2 Exact solution to e.o.m.: φ~µ,σ(~x) = gσ |~x−~µ|2+σ2
16σ2 |~x−~µ|2−σ2 Lagrangian density: L~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4
L~µ,σ(~x) < 0 for |~x − ~µ| < σ, so not a good prob. distr.
16 32 Classical metric on States: Instantons
Blau, Narain, Thompson (2001) - YM instantons in 3+1-d Uses prescription of Hitchin (1990): probability distribution is taken to be the Lagrangian density evaluated on the state. R 4 µ 2 4 Simpler model: d x ∂µφ ∂ φ − g φ (Euclidean signature) 2 σ2 Exact solution to e.o.m.: φ~µ,σ(~x) = gσ |~x−~µ|2+σ2
16σ2 |~x−~µ|2−σ2 Lagrangian density: L~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4
L~µ,σ(~x) < 0 for |~x − ~µ| < σ, so not a good prob. distr.
2 2 4 16 σ4 If instead we use −φ ∂ φ − g φ , we get p~µ,σ(~x) = g2 (|~x−~µ|2+σ2)4 and the information geometry is 5d hyperbolic space.
16 32 Lessons
Many statistical models S can lead to the same Fisher metric (Gaussian and Cauchy-Lorentz).
FM inherits the symmetries of S but can have more
FM on the states of a field theory can be sensitive to stability (scalar instanton).
17 32 Classical metric on Theories: 2d Ising Model
Geometry of space of couplings of a theory. 2d classical Ising model on a square lattice:
N N X X H = −J σi,jσi+1,j − K σi,jσi,j+1 i,j=1 i,j=1
Partition function: Z = Q P e−βH(σ) i σi=±1 1 −βH(σ) Probability distribution: pβJ,βK (σ) = Z e . This is in the exponential family!
gij = ∂i∂jf where f is the reduced free energy per site
18 32 2d Ising Model Fisher Metric Ricci Scalar
βK
1.2
1.0
0.8 ferromagnetic phase
0.6
0.4
0.2 disordered phase
0.0 βJ
0.0 0.2 0.4 0.6 0.8 1.0 1.2 (a) Ricci curvature as function of (b) 2d Ising Model phase couplings J, K (β = 1) diagram for βJ, βK ≥ 0
The scalar curvature diverges at the critical curve given by sinh(2βJ) sinh(2βK) = 1
19 32 Lessons
Many statistical models S can lead to the same Fisher metric (Gaussian and Cauchy-Lorentz).
FM inherits the symmetries of S but can have more
FM on the states of a field theory can be sensitive to stability (scalar instanton).
FM on the space of theories is sensitive to phase transitions (2d Ising model)
20 32 R d2z Well-known effective theory: S = 2π ψ∂ψ + ψ∂ψ + imψψ . 1 Can we reproduce the ln |m| divergence of the Fisher metric in the effective field theory?
2d Isotropic Ising Model and 1d Majorana Fermions
√ 1 Set J = K = 1 and the critical temp. is βc = 2 ln 2 + 1 ≈ 0.44. gββ is the specific heat:
d2f 1 1 gββ = 2 ' ln ' ln , dβ |β − βc| |m|