Tensor Network States: Algorithms and Applications December 1-5, 2014 Beijing, China The multi-scale Entanglement Renormalization Ansatz MERA -- a pedagogical introduction-- Guifre Vidal outline MERA • Definition • Efficiency • Structural properties: correlations and entropy The Renormalization Group • Goals • RG by isometries (TTN): why is it “wrong”? TRG • RG by isometries and disentanglers (MERA) (Zhiyuan’s lecture) TNR (Glen’s talk) MERA: definition š ⊗š Ψ ∈ (ā ) šš complex numbers Multi-scale entanglement renormalization ansatz (MERA) Matrix product state (MPS) MERA also MERA ! Efficiency š ⊗š Ψ ∈ (ā ) š + š š 1 1 + + … =š 1+ + +āÆ 2 4 2 4 šš ≤ 2š complex numbers Multi-scale entanglement renormalization ansatz (MERA) Matrix product state (MPS) š spins ⇒ š tensors ⇒ š(š) parameters log(š) š š spins ⇒ š log(š) tensors ? 2š tensors ⇒ š(š) parameters Q1 efficiency Matrix product state (MPS) Q2 〈Ψ Ψ cost š(š) 〈Ψ|š Ψ cost š(š) š” efficiency |Ψ š¤ š¢ isometric tensors! š¢ š¤ š” š”† 〈Ψ|Ψ = š¤ š¤† š¢ š¢† =1 = = Q3 cost = 0 !!! efficiency 〈Ψ|Ψ = isometric tensors! š” š”† š¤ š¤† š¢ š¢ † =1 = = efficiency 〈Ψ|Ψ = isometric tensors! š” š”† š¤ š¤† š¢ š¢ † =1 = = efficiency 〈Ψ|Ψ = isometric tensors! š” š”† š¤ š¤† š¢ š¢ † =1 = = efficiency 〈Ψ|Ψ = isometric tensors! š” š”† š¤ š¤† š¢ š¢ † =1 = = efficiency 〈Ψ|Ψ = isometric tensors! š” š”† š¤ š¤† š¢ š¢ † =1 = = efficiency 〈Ψ|Ψ = 1 isometric tensors! š” š”† š¤ š¤† š¢ š¢ † =1 = = š 〈Ψ|š Ψ = isometric tensors! š” š”† š¤ š¤† š¢ š¢† =1 = = = (1) = 〈Ψ|š Ψ = (2) = (3) cost O(log š ) = ΨšΨ Structural properties š ⊗š Ψ ∈ (ā ) šš • Decay of correlations • Scaling of entanglement complex numbers āØΨ|š 0 š(šæ)|Ψ MPS = = šæ−1 = ≈ ššæ−1 = š“š −šæ/š š≡− ⇒ Exponential decay of correlations 1 log š Ψš 0 š šæ Ψ āÆ āÆ āÆ = MERA āÆ āÆ āÆ Ψš 0 š šæ Ψ āÆ MERA āÆ = ≈ š = log3 šæ š log3 (šæ) = š2 log3(šæ) = šæ2 log3(š) = šæ−š š„ log3(š¦) = š¦ log3(š„) š ≡ −2 log 3 (š) ⇒ Polynomial decay of correlations āÆ āÆ Correlations: summary and interpretation matrix product state (MPS) multi-scale entanglement renormalization ansatz (MERA) šæ log(šæ) structure of geodesics: āØš 0 š šæ šæ ≈ š −šæ/š exponential structure of geodesics: āØš 0 š šæ ≈ šæ−š power-law Entanglement entropy matrix product state multi-scale entanglement renormalization ansatz (MPS) (MERA) š“ šæ log(šæ) connectivity: š“ š(š“) ≈ šššš š” boundary law! Q4 šæ connectivity: š(š“) ≈ log šæ logarithmic correction! Q5 ā® āÆ āÆ š(š“) ≈ log šæ Example: operator content of quantum Ising model š„ ššš„ ⊗ šš+1 +ā š»= š ššš§ for ā = āš = 1 š scaling dimension (exact ) scaling operators/dimensions: identity spin energy disorder fermions š 0 scaling dimension (MERA) 0 error ---- š 0.125 0.124997 0.003% š 0.99993 0.007% 0.1250002 0.0002% 0.5 <10−8 % 0.5 <10−8 % 1 ļ 0.125 ļ¹ 0.5 ļ¹ 0.5 OPE for local & non-local primary fields C ļ„ļ³ļ³ ļ½ 1 / 2 C ļ„ļ¹ ļ¹ ļ½ i C ļ„ļļ ļ½ ļ 1 / 2 C ļ„ļ¹ ļ¹ ļ½ ļ i Cļ¹ ļļ³ ļ½ e ļ iļ° / 4 Cļ¹ļļ³ ļ½ e iļ° / 4 / fusion rules ļ 2 ļ4 ( ļ± 6 ļ“ 10 ) / 2 ļ„ ļ“ļ„ ļ½ I ļ³ ļ“ ļ³ ļ½ I+ ļ„ ļ³ ļ“ļ„ ļ½ ļ³ ļ ļ“ļ ļ½ Iļ«ļ„ ļ ļ“ļ„ ļ½ ļ ļ¹ ļ“ļ¹ ļ½ I { I, ļ„ , ļ³ , ļ ,ļ¹ ,ļ¹ } local and semi-local subalgebras { I, ļ„ } { I, ļ„ , ļ³ } ļ¹ ļ“ļ¹ ļ½ I ļ¹ ļ“ļ¹ ļ½ ļ„ ļ¹ ļ“ļ„ ļ½ļ¹ ļ¹ ļ“ļ„ ļ½ļ¹ { I, ļ„ , ļ } { I, ļ„ ,ļ¹ ,ļ¹ } ... MERA and HOLOGRAPHY t t s x CFT1+1 x x AdS2+1 outline MERA • Definition • Efficiency • Structural properties: correlations and entropy The Renormalization Group • Goals • RG by isometries: why is it wrong? • RG by isometries and disentanglers The Renormalization Group: goals Given a local Hamiltonian š»= āš,š+1 on š sites (Hilbert space dimension š š ) š Two type of questions: 1) Low energy, large distance, UNIVERSAL behavior: e.g. disordered/symmetry-breaking phase, topological order (S,T modular matrices), quantum criticality (scaling operators, CFT data) fixed point š» → š» ′ → š» ′′ → āÆ → š» (ā) Ψ š š„ š š¦ |Ψ〉 for š„ − š¦ → ∞ 2) Low energy, short distance, detailed MICROSCOPIC properties e.g. 〈Ψ|š š„ |Ψ〉, Ψ š š„ š š¦ |Ψ〉, for all š„, š¦ The Renormalization Group: goals Example: given the (transverse field) Ising Hamiltonian š„ ššš„ ⊗ šš+1 +ā š»= š ššš§ š š(ā) spontaneous magnetization 0 ā āš magnetic field 1) Low energy, large distance, UNIVERSAL behavior? Is the spontaneous magnetization m(h) ≡ Ψ š š„ Ψ ≠ 0, or = 0? ordered phase disordered phase 2) Low energy, short distance, detailed MICROSCOPIC properties? How much is the spontaneous magnetization m(ā) ≡ 〈Ψ|š š„ |Ψ〉 as a function on ā? The Renormalization Group on Hamiltonians: on ground state wave-functions: on classical partition functions: š» → š»′ → š»′′ → āÆ → š» (ā) |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 šµ → šµ ′ → šµ ′′ → āÆ → šµ (∗) fixed point Hamiltonian fixed point ground state fixed point partition function Two types of Renormalization Group transformations: • Type 1 is only required to preserve UNIVERSAL properties • Type 2 is also required to preserve MICROSCOPIC properties š → š ′ → š ′′ → āÆ → š (ā) such that (for instance, š = š š„ in quantum Ising model) Ψ š Ψ = Ψ ′ |o′|Ψ ′ = š¹ ′′ š ′′ š¹ ′′ = āÆ = š¹ (∗) š(∗) š¹ (∗) For instance, TRG, TTN, MERA, are of type 2 The Renormalization Group • Type 1 is only required to preserve UNIVERSAL properties • Type 2 is also required to preserve MICROSCOPIC properties? š → š ′ → š ′′ → āÆ → š (ā) (for instance, š = š š„ in quantum Ising model) such that Ψ š Ψ = Ψ ′ |o′|Ψ ′ = š¹ ′′ š ′′ š¹ ′′ = āÆ = š¹ (∗) š(∗) š¹ (∗) Types 2A and 2B!! Type 2A: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 e.g. TTN, Zhiyuan’s lecture on TRG fixed point: mixture of UNIVERSAL and MICROSCOPIC properties Type 2B: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 e.g. MERA, Glen’s talk on TNR fixed point: only UNIVERSAL properties The Renormalization Group Type 2A: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 fixed point: mixture of UNIVERSAL and MICROSCOPIC properties example: TTN š ′ = š/3 sites š¤ š¤ ā′ š š¤† ā = š sites coarse-graining transformation š = š¤ ⊗ š¤ ⊗ āÆ⊗ š¤ Q6 = A’ B’ C’ D’ E’ F’ A’ B’ C’ D’ E’ F’ The Renormalization Group Type 2A: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 example: TTN fixed point: mixture of UNIVERSAL and MICROSCOPIC properties coarse-graining transformation š = š¤ ⊗ š¤ ⊗ āÆ⊗ š¤ Ψ → Ψ ′ = š † |Ψ〉 = Ψ š† A’ B’ C’ D’ F’ E’ Ψ′ A’ B’ C’ D’ E’ F’ š → š′ = š †š š š = A’ B’ C’ š D’ E’ F’ š† = A’ B’ C’ D’ E’ F’ A’ B’ C’ D’ E’ F’ š′ The Renormalization Group Type 2A: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 example: TTN fixed point: mixture of UNIVERSAL and MICROSCOPIC properties coarse-graining transformation š = š¤ ⊗ š¤ ⊗ āÆ⊗ š¤ Ψ → Ψ ′ = š † |Ψ〉 Ψ′ Ψ = š† Already a fixed point wave-function! It contains short-range entanglement = MICROSCOPIC details The Renormalization Group Type 2A: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 example: TTN fixed point: mixture of UNIVERSAL and MICROSCOPIC properties coarse-graining transformation š = š¤ ⊗ š¤ ⊗ āÆ⊗ š¤ |Ψ ′ 〉 retains short-range entanglement = MICROSCOPIC details Ψ š Ψ = Ψ ′ |o′|Ψ ′ ⇒ results are still accurate provided that we use a sufficiently large bond dimension Different fixed-point wave-function for the same phase! š(ā) (ā) |Ψh1 〉 → |Ψh1 ′〉 → |Ψā1 ′′〉 → āÆ → |Ψā1 〉 spontaneous magnetization (ā) (ā) |Ψā1 〉 |Ψā2 〉 0 (ā) |Ψh2 〉 → |Ψh2 ′〉 → |Ψā2 ′′〉 → āÆ → |Ψā2 〉 (ā) (ā) |Ψā1 〉 and |Ψā2 〉 have the same UNIVERSAL information, mixed with different MICROSCOPIC details ā ā1 ā2 magnetic field The Renormalization Group Type 2A: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 example: TTN fixed point: mixture of UNIVERSAL and MICROSCOPIC properties coarse-graining transformation š = š¤ ⊗ š¤ ⊗ āÆ⊗ š¤ |Ψ ′ 〉 retains short-range entanglement = MICROSCOPIC details Ψ š Ψ = Ψ ′ |o′|Ψ ′ ⇒ results are still accurate provided that we use a sufficiently large bond dimension (ā) (ā) |Ψā1 〉 and |Ψā2 〉 have the same UNIVERSAL information, mixed with different MICROSCOPIC details Two problems: • Computational: RG scheme is more expensive (ā) • Conceptual: How do we separate UNIVERSAL from MICROSCOPIC in |Ψā 〉 ? The Renormalization Group Type 2B: Proper RG transformation: fixed point: only UNIVERSAL properties |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 example: MERA š ′ = š/3 sites š¤ š¢ š¤ ā′ š¤† ā š¢ š š sites coarse-graining transformation š = (āÆ š¢ ⊗ š¢ ⊗ š¢ ⊗ āÆ )(āÆ š¤ ⊗ š¤ ⊗ š¤ ⊗ āÆ) Q7 = A’ B’ C’ D’ E’ F’ A’ B’ C’ D’ E’ F’ š¢† = = The Renormalization Group Type 2B: example: MERA Proper RG transformation: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 fixed point: only UNIVERSAL properties coarse-graining transformation š = (āÆ š¤ ⊗ š¤ ⊗ š¤ ⊗ āÆ )(āÆ š¢ ⊗ š¢ ⊗ š¢ ⊗ āÆ) Ψ → Ψ ′ = š † |Ψ〉 = Ψ š† Ψ′ A’ B’ C’ D’ E’ F’ š → š′ = š †š š = A’ B’ C’ š D’ E’ F’ = A’ B’ C’ D’ E’ F’ A’ B’ C’ D’ E’ F’ š′ The Renormalization Group Type 2B: example: MERA Proper RG transformation: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 fixed point: only UNIVERSAL properties coarse-graining transformation š = (āÆ š¤ ⊗ š¤ ⊗ š¤ ⊗ āÆ )(āÆ š¢ ⊗ š¢ ⊗ š¢ ⊗ āÆ) Ψ → Ψ ′ = š † |Ψ〉 Ψ š† Product state wave-function! It contains no MICROSCOPIC details = = Ψ′ A’ B’ C’ D’ E’ F’ = = The Renormalization Group Type 2B: example: MERA Proper RG transformation: fixed point: only UNIVERSAL properties |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 coarse-graining transformation š = (āÆ š¤ ⊗ š¤ ⊗ š¤ ⊗ āÆ )(āÆ š¢ ⊗ š¢ ⊗ š¢ ⊗ āÆ) Same fixed-point wave-function for the same phase! |Ψ 〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) |Ψ ā=0 (ā) |Ψ āš for ā < ā_š 〉 for ā = ā_š 〉 (ā) ā=∞ for ā > ā_š 〉 š(ā) magnetic field h spontaneous magnetization ā=0 ā′ |Ψ (ā) ā=0 〉 ā ā |Ψ (ā ) āš 〉 ā′ ā=∞ |Ψ (ā) ā=∞ 〉 The Renormalization Group Type 2B: example: MERA Proper RG transformation: |Ψ〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) 〉 fixed point: only UNIVERSAL properties coarse-graining transformation š = (āÆ š¤ ⊗ š¤ ⊗ š¤ ⊗ āÆ )(āÆ š¢ ⊗ š¢ ⊗ š¢ ⊗ āÆ) Same fixed-point wave-function for the same phase! |Ψ 〉 → |Ψ′〉 → |Ψ′′〉 → āÆ → |Ψ (ā) |Ψ ā=0 (ā) |Ψ āš for ā < ā_š 〉 for ā = ā_š 〉 (ā) ā=∞ 〉 for ā > ā_š With MERA, we have solved the two problems of TTN: • Computational: RG scheme is now scalable (ā) • Conceptual: |Ψ 〉 only contains UNIVERSAL information (it can be more easily extracted) summary MERA • Definition • Efficiency • Structural properties: correlations and entropy āØš 0 š šæ ≈ šæ−š š(š“) ≈ log šæ The Renormalization Group • Goals ground state wave-function |Ψ〉 classical partition function šµ • RG by isometries: why is it “wrong”? TTN TRG • RG by isometries and disentanglers MERA TNR (Zhiyuan’s lecture) (Glen’s talk)