Exact Pseudofermion Action for Hybrid Monte Carlo Simulation of One-Flavor Domain-Wall Fermion Yu-Chih Chen (for the TWQCD Collaboration) Physics Department National Taiwan University Collaborators: Ting-Wai Chiu Content • Lattice Dirac Operator for Domain-Wall Fermion (DWF) • Two-Flavor Algorithm (TFA) • TWQCD’s One-Flavor Algorithm (TWOFA) • Rational Hybrid Monte Carlo (RHMC) Algorithm • TWOFA vs. RHMC with Domain-Wall Fermion • Concluding Remarks Lattice Dirac Operator for Domain-Wall Fermion (DWF) For domain-wall fermion, in general, the lattice Dirac operator reads π·ππ€π π = ππ π·π€ + πΌ + ππ π·π€ − πΌ πΏ π where ππ = π + π ππ , ππ = π − π ππ and π = diag π1 , π2 , β― , πππ , and π·π€ is the standard Wilson-Dirac operator with −π0 0 < π0 < 2 , πΏ π = π+ πΏ+ π + π− πΏ− π = πΏ+ π π ,π ′ πΏ+ π 0 0 πΏ− π π·ππππ πΏπ ′ ,π −1 , 1 < π ≤ ππ = −ππΏπ ′ ,ππ , π = 1, with π = πππ , π = 1/ 2π0 1 − ππ0 πΏ− π = πΏ+ π π πΏ± (π) are the matrices in the fifth dimension and depend on the quark mass. Lattice Dirac Operator for Domain-Wall Fermion (DWF) For ππ = 4, the form of πΏ± are ο© 0 οͺ ο1 οͺ Lο« ο½ οͺ 0 οͺ ο« 0 0 0 0 0 ο1 0 0 ο1 ο mοΉ οΊ 0 οΊ 0 οΊ οΊ 0 ο» ο© 0 οͺ 0 οͺ Lο ο½ οͺ 0 οͺ ο«ο m ο1 0 0 ο1 0 0 0 0 0 οΉ οΊ 0 οΊ ο 1οΊ οΊ 0 ο» Using ππ = π + πππ and ππ = π − πππ , the π·ππ€π (π) can be written as π·ππ€π (π) = π«π ππ π° + π³(π) + π π° − π³(π) Matrix in 4D Matrices in 5-th dimension πΈ+ π π π πΈ− π + π° − π³(π) π«ππππ Lattice Dirac Operator for Domain-Wall Fermion (DWF) π«π ππ (π) = π«π ππ π° + π³(π) + π π° − π³(π) + π° − π³(π) 1. If ππ are the optimal weights given in Ref. [1], it gives Optimal Domain-Wall Fermion π» π»2 ≈ π(π») = π» π ππ βΆ Zolotarev Optimal Rational Approximation π» 2 + ππ π(π») = 1− 1+ ππ π =1 ππ ππ π =1 ππ → π» π»2 , as ππ → ∞ where ππ = (1 − π»π ) (1 + π»π ), π»π = ππ π» and π» = ππ»π€ 1 + ππΎ5 π»π€ [1] T. W. Chiu, Phys. Rev. Lett. 90, 071601 (2003) −1 π»π€ = πΎ5 π·π€ Lattice Dirac Operator for Domain-Wall Fermion (DWF) π«π ππ (π) = π«π ππ π° + π³(π) + π π° − π³(π) + π° − π³(π) 2. If ππ = π, π = π. π, π = π. π, it gives Domain-Wall Fermion with Shamir Kernel 3. If ππ = π, π = π. π, π = π. π, it gives Domain-Wall Fermion with Scaled (πΆ = π) Shamir Kernel π» π»2 ≈ π(π») = π» π π(π») = π′π βΆ Polar Approximation 2 π» + π′π 1− 1+ ππ π =1 ππ ππ π =1 ππ → π» π»2 , as ππ → ∞ where ππ = (1 − π») (1 + π»), and π» = ππ»π€ 1 + ππΎ5 π»π€ −1 Lattice Dirac Operator for Domain-Wall Fermion (DWF) For a physical observable πͺ(π) 1 πͺ(π) = π 1 = π π·ππ·ππ·ππͺ(π)exp −ππ·π π π − ππ π π·ππͺ(π)det[π·π (π)]exp −ππ π If π·π (π) = πΎ ∗ π·(π), where the matrix K is independent of the gauge field π·ππͺ(π)det[π·π (π)]exp −ππ π π·πdet[π·π (π)]exp −ππ π = π·ππͺ(π)det[π·(π)]exp −ππ π π·πdet[π·(π)]exp −ππ π Lattice Dirac Operator for Domain-Wall Fermion (DWF) Using the redefined operator π·(π) 1 πͺ(π) = π 1 = π π·ππͺ(π)det[π·(π)]exp −ππ π π·ππ·π † π·ππͺ(π) exp −π † π»−1 π π − ππ π where π» satisfies: 1) det π» = det π· 2) H is Hermitian 3) H is positive-definite Lattice Dirac Operator for Domain-Wall Fermion (DWF) For DWF, since π and πΏ are independent of the gauge field, + ππΌ πΌ+−πΏ(π) πΏ(π) + π + πΌ πΌ−−πΏ(π) πΏ(π) π·ππ€π π = π·π€ ππ πΌ + πΏ(π) ππ π«π ππ → π« π = π«π + π΄ π = π«π + π΄+ π π π± π = π −1/2 ππ± π + π−1 π −1 π −1/2 π± π = 1 + πΏ± (π) 1 − πΏ± (π) −1 π π΄− π −1 π«ππππ Two-Flavor Algorithm (TFA) [2] For the DWF Dirac operator π· π = π·π€ + π(π) = π·π€ + π+ π+ π + π− π− (π) we can apply the Schur decomposition with the even-odd preconditioning 4 − π0 + π(π) π· π = ππ π·π€ (π) where πΌ ππ = π·π€ π5 (π) 0 πΌ π·π€ππ(π) π5 (π)−1 ≡ 4 − π0 + π(π) π·π€ππ π5 (π)−1 0 0 π5 (π)−1 πΆ(π) ππ π·π€ π5 (π)−1 πΌ 0 ππ π5 (π)π·π€ πΌ ππ πΆ π = I − π5 (π)π·π€ π5 (π)π·π€ππ We then have det π· π = det[π5 π ]−2 × det[πΆ(π)] [2] T. W. Chiu, et al. [TWQCD Collaboration],PoS LAT 2009, 034 (2009); Phys. Lett. B 717, 420 (2012) . Two-Flavor Algorithm (TFA) The pseudofermion action for HMC simulation of 2-flavor QCD with DWF is πππ 1 † † = π πΆ (1) πΆ(1)π † πΆ π πΆ π πΆ(π) det πΆ(1) 2 Two flavor The field π can be generated by the Gaussian noise field π πππ = π†πΆ † 1 1 1 † πΆ 1 π = π†π πΆ π πΆ π 1 1 π= πΆ 1 π βΊ π= πΆ(π)πΌ πΆ π πΆ(1) generated from Gaussian noise TWQCD’s One Flavor Algorithm (TWOFA) For one-flavor of domain-wall fermion in QCD, we have devised an exact pseudofermion action for the HMC simulation, without taking square root. det[π·(π)] det[π·(π)] det[π· (π, 1)] = × det[π·(1)] det[π·(1)] det[π· (π, 1)] where π· π = π·π€ + π(π) = π·π€ + π+ π+ π + π− π− (π) In Dirac space π − π0 + π+ (π) π· π = − π⋅π‘ † π⋅π‘ π − π0 + π+ (π) π − π0 + π+ (π) π· π, 1 = − π⋅π‘ † π⋅π‘ π − π0 + π+ (1) TWQCD’s One Flavor Algorithm (TWOFA) Use type I Schur decomposition to π· π, 1 , and π·(π) π΄ πΆ π΅ πΌ = π· πΆπ΄−1 0 πΌ π΄ 0 0 π· − πΆπ΄−1 π΅ πΌ π΄−1 π΅ 0 πΌ we then have det π· π, 1 det π· π det[π» π + Δ− (π)] 1 = = det πΌ + Δ− (π) det[π» π ] π» π where π» π = π 5 π − π0 + π− π + π ⋅ π‘ † 1 π⋅π‘ π − π0 + π+ (π) Δ− π = π 5 π− 1 − π− (π) , π 5 = πΏπ +π ′,ππ TWQCD’s One Flavor Algorithm (TWOFA) Use type II Schur decomposition to π· 1 , and π· (π, 1) π΄ πΆ −1 π΅ πΌ π΅π· = π· 0 πΌ π΄ − π΅π·−1 πΆ 0 0 π· πΌ π·−1 πΆ 0 πΌ we then have det π· 1 det π· π, 1 det[π»(1)] 1 = = det πΌ + Δ+ (π) det[π» 1 − Δ+ (π)] π» 1 − Δ+ (π) where 1 π» 1 = π 5 π − π0 + π+ 1 + π ⋅ π‘ + π⋅π‘ π − π0 + π− (1) Δ+ π = π 5 π+ 1 − π+ (π) , π 5 = πΏπ +π ′,ππ † TWQCD’s One Flavor Algorithm (TWOFA) By using the Sherman-Morrison formula, we have found that the fifth dimensional matrices π± can be rewritten as π΄± π = π−π/π 𨱠−π π−π/π πππ + πΉπ π−π/π π± π± π» π−π/π π + π − ππππ where π is a scalar function of π, π and π, π£± are the vector functions of π, π and π, and here we have defined π΄± = ππ± 0 + π−1 π With these form of π± (π), Δ± (π) can be rewritten as Δ± π = π 5 π± 1 − π± (π) = ππ −1/2 π£± π£± π π −1/2 π 1−π π= 1 − ππ 1 + π(1 − 2ππ) TWQCD’s One Flavor Algorithm (TWOFA) Next we use det πΌ + π΄π΅ = det[πΌ + π΅π΄], one then has det π· π, 1 det π· π 1 = det πΌ + Δ− (π) π» π =det πΌ + ππ −1/2 π£− π£− π π −1/2 =det πΌ + ππ£− π π −1/2 1 π» π 1 π −1/2 π£− π» π Positive definite and Hermitian TWQCD’s One Flavor Algorithm (TWOFA) Again, with det πΌ + π΄π΅ = det[πΌ + π΅π΄], one also has det π· 1 det π· π, 1 1 = det πΌ + Δ+ (π) π» 1 − Δ+ (π) =det πΌ + ππ −1/2 π£+ π£+ π π −1/2 =det πΌ + ππ£+ π π −1/2 1 π» 1 − Δ+ (π) 1 π −1/2 π£+ π» 1 − Δ+ (π) Positive definite and Hermitian TWQCD’s One Flavor Algorithm (TWOFA) π1 = π1 † πΌ + ππ£− π π −1/2 π2 = π2 † πΌ + ππ£+ π π −1/2 × 1 π −1/2 π£− π1 π» π 1 π −1/2 π£+ π2 π» 1 − Δ+ (π) det π· π, 1 det π· 1 det π· π det π· π, 1 det[π·(π)] = det[π·(1)] One flavor determinant TWQCD’s One Flavor Algorithm(TWOFA) Use π1 , π2 and some algebra, the pseudofermion action of one-flavor domain-wall fermion can be written as 1 0 † π −1/2 −1/2 πππ = 0 π1 πΌ − ππ£− π π π£− π1 π»(π) + π2 † 0 πΌ + ππ£+ π π −1/2 1 π −1/2 π£+ π» 1 − Δ+ (π)π+ where π» π = πΎ5 π 5 π·(π) , π 5 = πΏπ +π ′,ππ Δ± π = ππ −1/2 π£± π£± π π −1/2 π 1−π π= 1 − ππ 1 + π(1 − 2ππ) π2 0 TWQCD’s One Flavor Algorithm(TWOFA) The initial pseudofermion fields of each HMC trajectory are generated by Gaussian noises as follows. π1 = π1 † π πΌ + ππ£− π −1/2 1 π −1/2 π£− π1 = π1 † π1 π» π 1 π −1/2 ⇒ π1 = πΌ + ππ£− π π −1/2 π£− π» π π2 = π2 † πΌ + ππ£+ π π −1/2 −1/2 π1 1 π −1/2 π£+ π2 π» 1 − Δ+ (π) 1 π −1/2 ⇒ π2 = πΌ + ππ£+ π π −1/2 π£+ π» 1 − Δ+ (π) Gaussian noise −1/2 π2 TWQCD’s One Flavor Algorithm (TWOFA) The invesre square root can be approximated by the Zolotarev optimal rational approximation π1 π1 ππ = π=1 ππ πΌ 1 + ππ π2 π2 ππ = π=1 Gaussian noise ππ πΌ 1 + ππ Rational Hybrid Monte Carlo(RHMC) Algorithm A widely used algorithm to do the one-flavor HMC simulation is the rational hybrid Monte Carlo (RHMC)[3], which can be used for any lattice fermion. 1/4π 1/4π 1 † † † πππ = ππ πΆ(1)πΆ 1 πΆ(1)πΆ 1 ππ 1/2π πΆ(π)πΆ † π π 1. det πΆ(1) πΆ π = π det πΆ(1)πΆ † 1 1/4π 1 πΆ(π)πΆ † π 1/2π πΆ(1)πΆ † 1 1/4π 2. Positive definite and Hermitian The fields ππ are generated by the Gaussian noise fields ππ ππ = 1 πΆ(1)πΆ † 1 1/4π πΆ(π)πΆ † π 1/4π Gaussian noise ππ [3] M. A. Clark and A. D. Kennedy, Phys. Rev. Lett. 98, 051601 (2007) Rational Hybrid Monte Carlo(RHMC) Algorithm πππ = ππ † † πΆ(1)πΆ 1 π ππ π΄1/πΌ = π0 + π=1 ππ π΄ + ππ 1/4π 1 πΆ(π)πΆ † π 1/2π † πΆ(1)πΆ 1 1/4π ππ rational approximation where ππ is the number of poles for rational approximation. The ππ numbers of inverse matrices 1 (π΄ + ππ ) can be obtained from the multiple mass shift conjugate gradient. TWOFA vs. RHMC with DWF I. Memory Usage : We list memory requirement (in unit of bytes) for links, momentum and 5D vectors as follows, 1) ππ ≡ 8 ∗ ππ₯ 3 ∗ ππ‘ 2) ππ = 48 ∗ ππ , link variables 3) ππ = 32 ∗ ππ , momentum 4) ππ = 24 ∗ ππ ∗ ππ , 5D vector Then the ratio of the memory usage for RHMC and TWOFA is 20 + 3 3 + 2ππ ππ ππ π»ππΆ = πππππΉπ΄ 32 + 10.5ππ where ππ is the number of poles for MMCG in RHMC algorithm TWOFA vs. RHMC with DWF 20 + 3 3 + 2ππ ππ ππ π»ππΆ = πππππΉπ΄ 32 + 10.5ππ TWOFA vs. RHMC with DWF II. Efficiency: The lattice setup is π = π. ππ, ππ = π. ππ, π³π = ππ , π» = ππ, π΅π = ππ, HMC Steps: (Gauge, Heavy, Light) = (1, 1, 10), π΅π = ππ for RHMC We compare RHMC and TWOFA for the following cases: 1) DWF with π = π. π, π = π. π and ππππ ππππ = π. ππ π. π (Optimal DWF) 2) DWF with π = π. π, π = π. π and ππ = π (Shamir) 3) DWF with π = π. π, π = π. π (πΆ = π) and ππ = π (Scaled Shamir) TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice Maximum Forces 1. Optimal Domain-Wall Fermion : Maximum Forces ππ = π. π ππ = π. ππ ππ = π. π ππ = π. ππ TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice 1. Optimal Domain-Wall Fermion: βπ― βπ» π−βπ― = π. ππππ(ππ) erfc π«π― = π. πππ(ππ) π Accept = π. πππ(ππ) π−βπ― = π. ππππ(ππ) erfc π«π― = π. πππ(ππ) π Accept = π. πππ(ππ) TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice 1. Optimal Domain-Wall Fermion : π = π. ππ, ππ = π. ππ, π³ = π, π» = ππ, HMC Steps: (Gauge, Heavy, Light) = (1, 1, 10), π΅π = ππ, π΅π = ππ for RHMC ODWF (kernel π―π ) with π = π. π, π = π. π and ππππ ππππ = π. ππ π. π Algorithm ππ ππ Plaquette Force (Gauge) Force (heavy) Force (light) TWOFA 1.3 0.4 0.58051(09) 5.15555(34) 0.18971(13) 0.01401(29) RHMC 1.3 0.4 0.58100(10) 5.15762(36) 0.35334(11) 0.06946(44) Algorithm Accept erfc( βπ― π) π−βπ― Memory π»ππππ. π»ππππ. (sec.) TWOFA 0.980(8) 0.981(11) 0.9992(16) 1.00 1.00 0ππππ(πππ) RHMC 0.987(7) 0.994(18) 1.0003(16) 6.58 1.21 πππππ(πππ) TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice Maximum Forces 2. Shamir Kernel (ππ = π) : Maximum Forces ππ = π. π ππ = π. π ππ = π. ππ ππ = π. ππ TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice 2. Shamir Kernel (ππ = π) : βπ― βπ» π−βπ― = π. ππππ ππ erfc π«π― = π. πππ(ππ) π Accept = π. πππ(ππ) π−βπ― = π. ππππ ππ erfc − π«π― = π. πππ(ππ) π Accept = π. πππ(ππ) TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice 2. Shamir Kernel (ππ = π) : π = π. ππ, ππ = π. ππ, π³ = π, π» = ππ, π΅π = ππ, HMC Steps: (Gauge, Heavy, Light) = (1, 1, 10), π΅π = ππ for RHMC DWF (Shamir kernel) with π = π. π, π = π. π and ππ = π Algorithm ππ ππ Plaquette Force (Gauge) Force (heavy) Force (light) TWOFA 1.8 0.1 0.59061(09) 5.17686(34) 0.14663(45) 0.03578(18) RHMC 1.8 0.1 0.59094(14) 5.17866(35) 0.28522(68) 0.10757(06) Algorithm Accept erfc( βπ― π) π−βπ― Memory π»ππππ. π»ππππ. (sec.) TWOFA 0.987(7) 0.987(13) 0.9999(16) 1.00 1.00 ππππ(ππ) RHMC 0.997(3) 0.953(06) 1.0074(18) 6.58 1.00 ππππ(ππ) TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice Maximum Forces 3. Scaled Shamir Kernel (ππ = π and πΆ = π) : Maximum Forces ππ = π. π ππ = π. π ππ = π. ππ ππ = π. ππ TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice 3. Scaled Shamir Kernel (ππ = π and πΆ = π) : βπ― βπ» π−βπ― = π. ππππ ππ erfc π«π― = π. πππ(ππ) π Accept = π. πππ(ππ) π−βπ― = π. ππππ ππ erfc − π«π― = π. πππ(ππ) π Accept = π. πππ(ππ) TWOFA vs. RHMC with DWF on the ππ × ππ × ππ Lattice 3. Scaled Shamir Kernel (ππ = π and πΆ = π) : π = π. ππ, ππ = π. ππ, π³ = π, π» = ππ, π΅π = ππ, HMC Steps: (Gauge, Heavy, Light) = (1, 1, 10), π΅π = ππ for RHMC DWF (scaled Shamir kernel) with π = π. π, π = π. π (πΆ = π)and ππ = π Algorithm ππ ππ Plaquette Force (Gauge) Force (heavy) Force (light) TWOFA 1.8 0.1 0.59061(09) 5.17854(34) 0.14646(13) 0.03359(13) RHMC 1.8 0.1 0.59032(09) 5.17670(32) 0.28559(39) 0.10775(06) Algorithm Accept erfc( βπ― π) π−βπ― Memory π»ππππ. π»ππππ. (sec.) TWOFA 0.983(7) 0.990(14) 1.0000(15) 1.00 1.00 ππππ(ππ) RHMC 0.997(3) 0.967(07) 1.00038(18) 6.58 1.17 ππππ(ππ) Concluding Remarks 1. We have derived a novel pseudofermion action for HMC simulation of one-flavor DWF, which is exact, without taking square root. 2. It can be used for any kinds of DWF with any kernels, and for any approximations (polar or Zolotarev) of the sign function. 3. The memory consumption of TWOFA is much smaller than that of RHMC. This feature is crucial for using GPUs to simulate QCD. 4. The efficiency of TWOFA of is compatible with that of RHMC. For the cases we have studied, TWOFA outperforms RHMC. 5. TWQCD is now using TWOFA to simulate (2+1)-flavors QCD, and (2+1+1)-flavors QCD, on 323 × 64 × 16, and 243 × 48 × 16 lattices. Thank You