NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Continued Fraction Bag Of Tricks Bálint Joó School of Physics, University of Edinburgh UKQCD Collaboration In collaboration with Robert Edwards Tony Kennedy Kostas Orginos Urs Wenger NIVER IT Y R G H O F E Jlab LHPC Collaboration Edinburgh UKQCD Collaboration MIT LHPC Collaboration DESY Zeuthen UKQCD Collaboration S TH E U D I U N B -1- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B The Goal Chiral symmetry on the lattice Massless Dirac Operator, relation: , must satisfy Ginsparg-Wilson Explicit solution by Neuberger is an auxiliary Dirac Operator (the kernel) ) is the Matrix Sign Function ( NIVER S IT TH E U Y R G H O F E D I U N B -2- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Traditional Implementation Approximate Sign Function with rational function and . polynomials in H of degree with Write approximation as a sum over poles: : Use Multi-shift solver to apply poles : Use nested CG Solver IT TH E S Y R G H O F E NIVER U Inversion of Application of D I U N B -3- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Motivation for 5D approach: Nested solver is inefficient – Each outer solve throws away Krylov Subspace built up by inner solve. Can solve an enlarged 5D system: equivalent to solving by nested solve NIVER S IT TH E U Y R G H O F E D I U N B -4- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Basic Idea Consider the matrix equation: on Usual (4D): Invert Higher dimensional (5D): Solve enlarged system IT TH E S Y R G H O F E NIVER U From second row: From first row: D I U N B -5- R G H O F E March 9, 2005 H G R Y TH E Enlarged system for sum over poles: R G H TH E Y O F IT March 9, 2005 -6O F D I U N B E S NIVER U D I U N B E IT Testing 5D Chiral Operators S NIVER U NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B More than just an inversion trick: Domain Wall Fermions It has been shown (Boriçi, Edwards & Heller, Brower Nef and Orginos) that the 5D domain wall fermion matrix can be reduced to a 4D effective operator which approximates the Overlap LDU decomposition (see Appendix) ... ... ... IT TH E S Y R G H O F E NIVER U With .. D I U N B -7- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B is rational approximation to overlap action: is the 5D transfer matrix Hamiltonian traditional DWF - Higham representation of from Zolotarev approximation “Optimal DWF” (Chiu) is contribution from 5D Bulk (Pauli-Villars) modes IT TH E S Y R G H O F E NIVER U D I U N B -8- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Choice of Kernel Shamir-like DWF: From 5D transfer Matrix: is the Wilson Fermion Operator, where , is the Domain Wall Height Wilson/Boriçi (Overlap like) DWF: Recent Unification of kernels: Möbius Kernel (Brower, Nef, Orginos) is arbitrary Shamir), Borici, ( chooses scale factor NIVER S IT TH E U Y R G H O F E D I U N B -9- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B to shift For Möbius, with Higham (tanh) approx, use scale factor spectrum of to where is a good approximation. to where approximation be- But can push large eigenmodes of comes worse again −α|λ| −|λ| max max α NIVER min α min α |λ| α|λ| min min |λ| max α|λ| max IT Y R G H O F E −α|λ| −|λ| S TH E U Optimization problem for D I U N B -10- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B DWF Determinant Cancel Bulk Contribution using : 5D DWF action, generates 4D effective action, eg for 2 flavours: IT TH E S Y R G H O F E NIVER U D I U N B -11- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B DWF Pros and Cons Advantages – Admits red-black preconditioning (in 4D and 5D) can be chosen for optimal rational approximation – – Well known, both from 5D and 4D perspective Disadvantages from Zolotarev – Optimal numerically expensive – Apart from placement of Domain Wall (permutation freedom), operators cannot be tuned NIVER S IT TH E U Y R G H O F E D I U N B -12- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B An Alternative: Continued Fractions Different functional form for approximating the sign function. Write: specify approximation Coefficients Equivalence transformation between continued fractions TH E Y H G R D I U N B -13- March 9, 2005 IT O F E S NIVER U NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B Inversion Trick approach ( Wenger et al. hep-lat/0110070, hep-lat/0403003): Consider: ... with ) with even , For type 0 rational approx ( NIVER S IT TH E U Y R G H O F E D I U N B -14- March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Then solving .. . .. . is equivalent to solving: ie NIVER S IT TH E U Y R G H O F E D I U N B -15- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B -s. Large (continuous) freedom of choosing Naive choice: , ensures diagonal elements are (except for corner) Matrix is block tridiagonal – useful: easy decompositions NIVER S IT TH E U Y R G H O F E D I U N B -16- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B More than just an inversion trick: LDU Decomposition of Operator Set for convenience and consider operator (2n = 2): are hermitian and invertible. where NIVER S IT TH E U Y R G H O F E D I U N B -17- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B as: Can Decompose with with and . IT TH E S Y R G H O F E NIVER U Decomposition works for any positive , and D I U N B -18- R G H O F E March 9, 2005 R D I U N B E H Y G IT R TH E G H TH E Y O F March 9, 2005 -19O F D I U N B E S NIVER U IT Testing 5D Chiral Operators Now identify: and S NIVER U NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Comments are successive continued fractions growing from , , ..., “the tail inwards” Only is coupled to the mass term is the continued fraction we need: NIVER S IT TH E U Y R G H O F E D I U N B -20- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Determinant of Continued Fraction Operator we have that Since and since we have that IT TH E S Y R G H O F E NIVER Determinants of play same role as the bulk (Pauli– Villars) determinant in DWF (ie: they need to be cancelled) U D I U N B -21- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Cancelling The Unwanted Determinants Consider the operator ... ... ... .. which by construction has NIVER S IT TH E U Y R G H O F E D I U N B -22- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B We see immediately that .. ... ... ... Is an effective 4D overlap operator, with we have and since : S IT TH E NIVER Y R G H O F E U D I U N B -23- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Further, we can consider the action: and change variables: we have Since no extra Jacobean arises from changing variables. IT TH E S Y R G H O F E NIVER U D I U N B -24- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B We have formally reduced a 5D continued fraction theory to an effective 4D overlap theory If the are chosen from the Higham Representation for and is a Domain Wall kernel ( or Möbius) this theory is exactly equivalent to DWF. The difference is a change in representation (Cayley Transform form of Rational Function, v.s. Continued Fraction form of Rational Function) NIVER S IT TH E U Y R G H O F E D I U N B -25- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Practicalities: PV Matrix is problematic: ( Evaluating case) The element contains a continued fraction. S IT TH E NIVER Y R G H O F E U . Don’t want to evaluate this. In general it would be D I U N B -26- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U G H O F D I U N B and : However, defining dimension reduced forms of R E ... ... ... ... ... Can form ... ... ... ... is straightforward to evaluate (no Cont. Frac. terms) NIVER S IT TH E U Y R G H O F E D I U N B -27- March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B still have -s along diagonal so and so And (finally): S IT TH E NIVER Y R G H O F E U D I U N B -28- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B Choice of Kernel Recall definition of : ... , operator is straighforward: need (when ) If of applications has a demominator ( or Möbius) we can play additional trick: If ): Consider Möbius case (includes NIVER S IT TH E U Y R G H O F E D I U N B -29- March 9, 2005 ... Y H IT G R TH E R G H TH E Y O F March 9, 2005 -30O F D I U N B E S NIVER U . (One application of Q) , so 2. 1. Solve D I U N B E IT Testing 5D Chiral Operators S NIVER U Now solve Two step process: NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B Relative Costs of Applying Operators DWF (Shamir/Möbius) Cont. Frac. ( ) Cont. Frac. ( ) Relative cost per CG Iteration compared to DWF 1 1 ), with even order Cont. Frac.: type 0 approximation ( Continued Fraction can be competitive if matrix is sufficiently better conditioned than DWF matrix for same . NIVER S IT TH E U Y R G H O F E D I U N B -31- March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Chiral Symmetry Breaking & Measuring For approximate Overlap Operator, breaking of GW relation is: to get: Substitute in or All information about breaking GW relation lives in NIVER S IT TH E U Y R G H O F E D I U N B -32- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B (Kikukawa, Brower, Neff, Orginos): and : Usual DWF with Going through all the algebra IT TH E S Y R G H O F E NIVER U D I U N B -33- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B . is a matrix element of R G H O F E Can measure fraction version of . cancels in are renormalised by and for just effective 4D action, by using 4D partial . Relation between formulations: -s IT TH E S Y R G H O F E NIVER U with say for Shamir and for Boriçi, with the tuned to give the same pion mass in both cases D I U N B -34- March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U . Eg: Can have negative D I U N B Further Random Musings on always positive approx never crosses for For low order Zolotarev Approx, big wiggles may cancel to give small . Can I really trust to tell me the amount of chiral symmetry breaking? not sufficient measure of chiral symmetry breaking (may need to measure higher order operators in . Eg is always positive) NIVER S IT TH E U Y R G H O F E D I U N B -35- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Numerical Studies Performed test to study efficiency of various formulations. , 15 – Dynamical DWF dataset Used RBRC Configurations – Kostas also worked on sets Performed inversions with various formulations (details later) Computed Counted CG iterations NIVER S IT TH E U Y R G H O F E D I U N B -36- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U Matched pion mass between operators with kernels D I U N B and Möbius DWF ALL operators are even-odd preconditioned Did not project eigenvectors of . Computers used: Myrinet and Gig-E PC Clusters at JLab, UK National Grid Service Myrinet PC Clusters in Leeds & Oxford, UK HPCx IBM facility, Edinburgh BlueGene/L, Edinburgh QCDOC Operators are coded in Chroma, using BAGEL Dslash from Peter Boyle where appropriate (HPCx, QCDOC, BlueGene) NIVER S IT TH E U Y R G H O F E D I U N B -37- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Comparing Operators DWF Kernel, Higham Approximation Cπ(t) 0.0001 Moebius DWF (b5=1, c5=0) Continued Fraction (HT) Wilson Kernel, Higham Approxmation Moebius DWF (b5=1, c5=1) Continued Fraction (HW) 0.0001 Borici DWF Normal Shamir DWF 1e-05 1e-05 Relative difference (%) 1e-06 0.0025 0.015 0.002 0.01 0.0015 0.001 0.005 0.0005 0 0 10 20 30 0 0 10 t . Single Precision IT TH E S Y R G H O F E NIVER U 30 t Single Configuration (#806). Max Difference 20 D I U N B -38- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U between Operators with DWF and D I U N B kernel Matching Tuning quark mass to get same Pion Mass, Higham Approximation 0.01 Moebius (Hw mode) Ls=12, mf=0.0115 Cont. Frac (Hw mode) Ls=12, mf=0.0115 Moebius (Shamir mode), Ls=12, mf=0.02 Cπ(t) 0.001 0.0001 1e-05 1e-06 Effective Mass of π 0 2 4 6 8 0.5 10 12 14 16 t 18 20 22 24 26 28 30 Folded Correlation Function about Mid Point 0.4 0.3 0.2 0.1 0 0 2 4 8 6 10 12 14 t NIVER S IT TH E U Y R G H O F E D I U N B -39- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B using Computing 3 V=16 32, Ls=12,M=1.8, Shamir msea=0.02, 15 Configs, matched Borici mval=0.0115 0.0004 Moebius DWF (Midpoint/Pseudo) (H_w, Higham Approx) Continued Fraction (∆L/Pseudo), (H_w, Higham Approx) Cont. Frac. (∆L/Pseudo): bare mres = 0.000110(-12,13) mres ratio 0.0003 Moebius (Midpoint/Pseudo): bare mres = 0.000114(-13,+14) 2 χ ./ d.o.f = 1.27 2 χ ./ d.o.f = 1.33 0.0002 0.0001 0 NIVER 4 6 8 10 12 14 16 t 18 20 22 24 26 28 30 IT Y R G H O F E 2 S TH E U 0 D I U N B -40- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B measurements per configuration Results with Hw kernel, Ls=12 | mres | per conf 0.01 0.001 0.0001 1e-05 1e-06 1e-07 | λmin | / | λmax | 1 0.01 0.0001 1e-06 NIVER 2 3 4 5 7 8 9 10 6 (Configuration - 605)/50 11 12 13 14 15 IT Y R G H O F E 1 HT kernel Hw kernel S TH E U Moebius DWF (Higham Approx) Cont Frac (Higham Approx) Cont Frac (Zolotarev Approx) Moebius DWF (Zolotarev Approx) D I U N B -41- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U Cost Normalised by Unscaled Shamir DWF Ls=14 off graph Scaled Shamir (α=1.7, HT, Higham) Scaled Shamir (α=2, HT, Higham) Scaled Shamir (α=2.5, HT, Higham) Scaled Shamir (α=3, HT, Higham) Continued Fraction (α=1, Hw, Higham) 6 Continued Fraction (Hw, Zolotarev) Continued Fraction (α=1, HT, Higham) Continued Fraction (HT, Zolotarev, 3 Cfg) 4 Ls=6 2 1e-06 0.0001 | mres | (mShamir/m) 0.01 1 IT Y R G H O F E Chiu DWF (b5=1, c5=1, Zolotarev) S TH E NIVER Shamir (α=1, HT, Higham) Borici (α=1, Hw, Higham) 0 1e-08 U 8 Ls=16 D I U N B Cost measurement: Cost vs. D I U N B -42- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B Conclusions 1: Theory and Background Nice unified framework for 5D chiral actions. Both DWF and Continued Fraction – reduce to Approximate 4D Overlap Effective Action or full Möbius (including Shamir) kernels Separate issues of representation, approximation and kernel Should choose one or other on basis of numerical cost NIVER S IT TH E U Y R G H O F E D I U N B -43- March 9, 2005 – can accomodate – can accomodate Higham and Zolotarev approximations for NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B Conclusions 2: Numerical tests (so far) Continued Fraction with Zolotarev coefficients. – Kernel becomes competitive with DWF at reasonable Cost per application amortised by small (Preliminary) – Kernel seems least costly . – But need to check for other approximation ranges NIVER S IT TH E U Y R G H O F E D I U N B -44- March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Möbius appears best for kernels with denominator and Higham apkernel and Zolotarev approx. proximation but is costly with – scaling Shamir kernel helps reduce for Higham approx. Continued Fraction appears to be costly for Higham approx. equivalence transform coeffs – But can still try and tune NIVER S IT TH E U Y R G H O F E D I U N B -45- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Future/Current Work Complete results for Cont. Frac. with Shamir kernel (in production on Edinburgh BlueGene/L and QCDOC) Still to examine 3rd formulation, based on Partial Fraction approximation (Neuberger and Narayanan’s “Alternative to Domain Wall Fermions”) NIVER S IT TH E U Y R G H O F E D I U N B -46- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Appendix 1: Cayley form of Rational Functions . Write Define “Euclidean Cayley Transform”: Oddness of : IT TH E S Y R G H O F E NIVER U D I U N B -47- R G H O F E March 9, 2005 H G R Y TH E R G H TH E Y O F IT March 9, 2005 -48O F D I U N B E S NIVER U implies: and in factored form: and Writing write: To find form of D I U N B E IT Testing 5D Chiral Operators S NIVER U H G R Y TH E Finally And so R G H TH E Y O F IT March 9, 2005 -49O F D I U N B E S NIVER U D I U N B E IT Testing 5D Chiral Operators S NIVER U NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B Appendix 2: Domain Wall Fermions : R. Brower, H. Neff, K. Orginos hep-lat/0409118 Möbius kernel for Transfer Matrix Hamiltonian: Shamir) Borici, chooses operator ( eg: is arbitrary scale factor IT TH E S Y R G H O F E NIVER U D I U N B -50- March 9, 2005 H G R Y TH E R G H TH E Y O F IT March 9, 2005 -51O F D I U N B E S NIVER U is the Domain Wall height is the fermion mass, and D I U N B E IT Testing 5D Chiral Operators S NIVER U Möbius 5D Operator (R. Brower, H. Neff, K. Orginos hep-lat/0409118) NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B ¨ Reducing Mobius DWF to 4D effective theory First written down by Boriçi Approach presented here (a la Brower et al): – Define change of variables from new variables . using – “LDU” decompose – Define – Diagonal piece will yield effective 4D operator . – Ls=4 examples for simplicity NIVER S IT TH E U Y R G H O F E D I U N B -52- R G H O F E March 9, 2005 H G R Y TH E Definitions New variables: R G H TH E Y O F IT March 9, 2005 -53O F D I U N B E S NIVER U Define D I U N B E IT Testing 5D Chiral Operators S NIVER U H G Y R IT TH E R G H TH E Y O F March 9, 2005 -54O F D I U N B E S NIVER U Write D I U N B E IT Testing 5D Chiral Operators S NIVER U LDU Decomposition with H G R Y TH E and with i.e: R G H TH E Y O F IT March 9, 2005 -55O F D I U N B E S NIVER U D I U N B E IT Testing 5D Chiral Operators S NIVER U NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Putting it together: and and switch to Step 1: Apply : IT TH E S Y R G H O F E NIVER U D I U N B -56- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B -s appear and Step 2: Multiply in the NIVER S IT TH E U Y R G H O F E D I U N B -57- R G H O F E March 9, 2005 H G R Y TH E R G H TH E Y O F IT March 9, 2005 -58O F D I U N B E S NIVER U ) overlap with Cayley representation: ( extra bit to get: and the Step 3: Multiply in the D I U N B E IT Testing 5D Chiral Operators S NIVER U with NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B ? What about this arises from Jacobian from change of variables , so called high corresponds to modes with frequency, or bulk modes. In calculations, “cancelling high-frequency bulk modes with PauliVillars fields” corresponds to removing this extra piece. , eg 2 flavour Usually done by working with HMC Hamiltonian: IT TH E S Y R G H O F E NIVER U D I U N B -59- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B Summary of Domain Wall Fermions DW fermions are 5D realisations of Overlap fermions with Cayley representation for the sign function and extra high frequency (PV) modes that need to be cancelled Approximation specified by approximation (traditional DWF) Physics ( behaviour) dictated by choice of Möbius, by . , or in the case of IT TH E S Y R G H O F E NIVER U the extra scale factor When using Zolotarev D I U N B -60- March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Appendix 3: Details of Approximations NIVER S IT TH E U Y R G H O F E D I U N B -61- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Approximation: Higham (tanh), Neuberger’s Polar Approx. Approximation is with for all , Cayley form: – Conventional DWF. NIVER S IT TH E U Y R G H O F E D I U N B -62- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Approximation: Zolotarev’s approximation : Approximation is over range with come from Jacobian elliptic functions: where the coefficients , is the complete elliptic integral, and end of the approximation range. is the lower IT TH E S Y R G H O F E NIVER U D I U N B -63- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U R G H O F E D I U N B can be conveniently split into partial or continued fraction form. Cayley form is entirely trivial (according to Tony) (T.W. Chiu hep-lat/0209153) can be scaled to cover spectrum of operator: eg: Approximation is rationally BAD outside approx. interval Approximation is (minimax) optimal over the approx. interval. Maximum error decreases exponentially with . NIVER S IT TH E U Y R G H O F E D I U N B -64- March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Zolotarev or tanh ? Comparison of Tanh and Zolotarev Approximations 1 0.6 2 1 - ε (x) 0.8 0.4 0.2 0 -1.4 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 x 0.2 0.4 0.6 0.8 1 1.2 1.4 ε(x) 1 0 tanh approx - Ls=12 zolotarev approx - Ls=12 tanh approx - Ls=16 zolotarev approx - Ls =16 tanh approx - Ls=20 tanh approx - Ls=24 -1 -1.4 -1.2 NIVER IT Y R G H O F E -0.8 -0.6 -0.4 -0.2 S TH E U -1 D I U N B -65- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Zoom in on Graph showing approximation errors unscaled tanh - Ls=16 zolotarev - Ls=16, eps=0.01 unscaled tanh - Ls=20 zolotarev-Ls=20, eps=0.01 zolotarev - Ls = 24, eps=0.01 Zolotarev - Ls=12, eps=0.4 2e-05 2 1 - ε (x) 1e-05 0 -1e-05 -2e-05 0.2 NIVER 0.6 0.8 1 IT Y R G H O F E x S TH E U 0.4 D I U N B -66- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Maximum error as we increase the range or approximation maximal error δ rational approximations over +/- [ε,1] 1 δ = | 1 - R(ε) | 0.01 0.0001 1e-06 1e-08 1e-10 1e-06 NIVER IT Y R G H O F E 0.01 S TH E U Ls=12 - zolotarev Ls=16 - zolotarev Ls=20 - zolotarev Ls=24 - zolotarev Ls=12 - unscaled tanh Ls=16 - unscaled tanh Ls=20 - unscaled tanh Ls=24 - unscaled tanh 0.0001 ε D I U N B -67- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U D I U N B Maximum error as we change Ls maximal error δ of type 0 Zolotarev approximation over +/- [ε,1] for various Ls 0 10 -1 ε=0.005 ε=0.001 ε=0.0001 ε=0.00001 10 -2 10 -3 10 -4 δ = | 1 - R(ε) | 10 -5 10 -6 10 -7 10 -8 10 -9 10 -10 10 -11 10 -12 10 NIVER 20 Ls 30 40 IT Y R G H O F E 10 S TH E U 0 D I U N B -68- R G H O F E March 9, 2005 NIVER S Y Testing 5D Chiral Operators IT TH E U the error are important to get right? is unimportant, I can reapproximate Zolotarev with , and the approximation will be better than . if BUT which so that for Given an , there is some is smaller than Zolotarev. D I U N B Zolotarev approximation wiggles around Zolotarev is uniformly good for a given interval. is good for larger . (since it is not “uniform”) IT TH E S Y R G H O F E NIVER U Maximum error is always BAD for D I U N B -69- R G H O F E March 9, 2005