Quantum Many-Body Harness

Computational methods

The harness's method zoo: 36 computational methods for quantum many-body problems, each with its accuracy class, an M1–M14 property table (tasks, regime, cost, scaling, failure modes), and verified benchmarks. Every row is generated from the method's knowledge-base card.

36 of 36
Family Accuracy Skill

1 Exact methods 5

Exact diagonalization — fulled-fullexactmethod-edsandvik_2010_computational
AxisValueNote
M1 tasks / outputsFull spectrum · all eigenvectors · level statistics · ETH diagnostics · thermodynamics · entanglement spectrum · order parameters · exact finite-cluster dynamicsOnly method that natively returns the full spectrum; thermodynamics via exact canonical sum.
M2 regimeT=0 and finite-T (all temperatures from a single diagonalization)Full spectrum covers all regimes at once.
M3 accuracy classNumerically exact, deterministicErrors are only floating-point rounding; no statistical error, no truncation bias.
M4 dimension fit (A1)Any dimension / geometryApplicable to any lattice; total site count N is the only limit, not geometry.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion / anyLocal dim d enters as D_H = d^N; large d (Hubbard d=4, soft-core bosons) drastically shrinks reachable N.
M6 entanglement regime (B5)Volume-law toleratedNo entanglement restriction; full Hilbert-space representation requires O(D_blk²) memory.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; no sign problem.
M8 symmetry exploitation (C9/C10)U(1)/SU(2)/Z₂/parity block-diagonalize H; translation (k) and point-group reduce D_blk by ≈N and |G| respectivelyDecisive cost lever: D_blk = D_H / (reduction factor). With full symmetry (Sᶻ + momentum + point group + parity), a spin-½ chain of N≈20 is trivial dense; dense ED reaches ~28–30 sites maximum (see M12). The ~40-site and 48–50-site numbers belong to iterative (Lanczos) ED — see ed-lanczos card.
M9 time complexityO(D_blk³)Dense eigensolve (LAPACK dsyevd); dominant cost for useful system sizes.
M10 memoryO(D_blk²)Must store the full dense matrix and all eigenvectors; the first wall hit in practice.
M11 control knobNone (exact) — the only knob is which symmetry sectors to diagonalizeNo convergence parameter: result is exact within machine precision.
M12 scale frontier~20 sites trivial dense; ~28–30 sites maximum for dense ED with full symmetry (N=28 full-symmetry D_blk~1.8×10⁵ → ~250–520 GB (real vs complex double); N=32 → ~44 TB, infeasible)Dense O(D_blk²) memory is the binding wall. Iterative (Lanczos) ED reaches ~40 sites routine and 48–50 at the supercomputer frontier (Wietek–Läuchli 2018, arXiv:1804.05028) — see ed-lanczos card.
M13 primary approximation / biasNone — numerically exactOnly approximation is the finite cluster (open/periodic BC); bulk extrapolation requires finite-size scaling.
M14 hard blocker / failure modeHilbert-space d^N wall: dense O(D_blk²) memory walls by N ≈ 30 (spin-½) or N ≈ 16 (Hubbard d=4)No workaround within ED full; switch to ED Lanczos for larger N or for a few eigenstates.

Cost & scaling

  • Time: O(D_blk³) — LAPACK dense eigensolve
  • Memory: O(D_blk²) — stores the full matrix and all eigenvectors
  • Control knob: none (exact); symmetry-sector choice determines D_blk
  • Scale frontier: ~20 sites trivial dense; ~28–30 sites maximum for dense ED (O(D_blk²) memory wall; N=28 D_blk~1.8×10⁵ → ~250–520 GB (real vs complex double)); iterative/Lanczos ED reaches ~40 sites routine and 48–50 at the supercomputer frontier (Wietek–Läuchli 2018) — see ed-lanczos card

Recommended for

  • Universal oracle / cross-check: all models at small N; primary validation target for DMRG, QMC, VMC, iPEPS
  • Level statistics and ETH: non-integrable models (Heisenberg, Hubbard, SYK) at sizes where the full spectrum fits
  • Thermodynamics at all temperatures: small frustrated clusters (kagome, pyrochlore) where QMC has a sign problem (C12) and FTLM requires many random vectors
  • Topological models (B7): ED on a torus gives ground-state degeneracy and TEE directly
  • Any model with small Hilbert space: spin-½ chains up to N≈20, Hubbard up to N≈12–16 sites
  • Per method-property-map.md (ED profile): applicable when D_H ≲ 10⁷–10⁸ after symmetry reduction

Benchmarks

  • Dense ED frontier: ~28–30 sites (spin-½ with full U(1) + translation + point-group symmetry); N=28 gives D_blk~1.8×10⁵ → ~250–520 GB (real vs complex double) for the dense matrix; N=32 → ~44 TB, infeasible.
  • Iterative (Lanczos/Krylov) ED frontier: 48–50 sites (spin-½ kagome/square lattice with sublattice coding, sparse matrix-vector products) [Wietek–Läuchli, arXiv:1804.05028; method-survey.md §1.1] — this belongs to ed-lanczos, not ed-full.
  • Spin-½ chain N=20: Hilbert-space dim D_H = 2^20 ≈ 10^6; with U(1) largest block C(20,10) ≈ 184,756; dense diag trivial on a laptop sandvik_2010_computational.
  • Hubbard d=4, N=16: D_H = 4^16 ≈ 4×10^9 (full Fock space, all fillings; at half-filling D_H = C(16,8)² ≈ 1.66×10^8); requires full symmetry reduction to be tractable [method-survey.md §1.1].

Key reference: sandvik_2010_computational — comprehensive tutorial covering the full diagonalization approach, symmetry reduction (momentum sectors, point group, spin inversion), thermodynamic observables via exact canonical sums, and benchmarks for Heisenberg and Hubbard models.

Exact diagonalization — Lanczos/Kryloved-lanczosexactmethod-edsandvik_2010_computational
AxisValueNote
M1 tasks / outputsGround-state energy · ground-state wavefunction · few excited states & gaps · dynamical S(q,ω) (continued-fraction) · real-time dynamics (Krylov exponentiation) · entanglement entropy · order parametersLevel statistics and full thermodynamics require the full spectrum (ED full).
M2 regimeT=0 (a few low eigenstates); real-time dynamics on finite clustersFinite-T via FTLM/TPQ (separate method) uses the same matrix-free machinery.
M3 accuracy classNumerically exact, deterministicThe Lanczos recursion converges to machine precision for the targeted eigenpairs; no stochastic error.
M4 dimension fit (A1)Any dimension / geometrySame as ED full: geometry only affects N_nnz and the sparsity pattern, not the algorithm's applicability.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion / anyD_H = d^N; large d shrinks reachable N, same as ED full, but Lanczos reaches about 2× larger N than full diag.
M6 entanglement regime (B5)Volume-law toleratedNo entanglement restriction; the ground-state wavefunction is stored as a full vector of length D_blk.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; no sign problem.
M8 symmetry exploitation (C9/C10)U(1)/SU(2)/Z₂/parity block-diagonalize H; translation (k) and point-group reduce D_blk by ≈N and |G|The decisive cost lever: symmetry reduction is the primary route to larger system sizes. With full symmetry, N≈48–50 is accessible (Wietek–Läuchli 2018).
M9 time complexityO(n_iter · N_nnz) where N_nnz ≈ N · D_blk for short-range (A4) Hamiltoniansn_iter ≈ tens for gapped (B6) systems; hundreds for critical or near-degenerate systems; much higher for shift-invert (interior states).
M10 memoryO(D_blk)Stores only a few Lanczos vectors of length D_blk; the key advantage over ED full.
M11 control knobn_iter (number of Lanczos steps) — controls convergence of the targeted eigenpairsConvergence criterion: residual norm ‖H|ψ⟩ − E|ψ⟩‖ < ε. For the GS, convergence is typically fast (tens of iterations); for excited states or S(q,ω), more iterations needed.
M12 scale frontier~40 sites routine with full symmetry; 48–50 sites at the supercomputer frontier (Wietek–Läuchli 2018, arXiv:1804.05028, sublattice-coding with translation + point group + parity)Frontier requires distributed-memory storage of D_blk-length vectors across many nodes.
M13 primary approximation / biasNone — numerically exact for the targeted eigenpairsThe finite cluster approximation (open/PBC) is external; bulk extrapolation requires finite-size scaling.
M14 hard blocker / failure moded^N wall (same as ED full, but shifted to larger N): D_blk vector storage requires distributed memory beyond ~40 sites; near-degenerate or critical systems (B6) require many iterations (n_iter → ∞ at a quantum critical point)Long-range interactions (A4) raise N_nnz and slow matrix-vector products.

Cost & scaling

  • Time: O(n_iter · N_nnz) with N_nnz ≈ N · D_blk for short-range models; n_iter ~ tens (gapped) to hundreds (critical)
  • Memory: O(D_blk) — stores only a few Krylov vectors (the key advantage over ED full)
  • Control knob: n_iter (Lanczos iterations) — controls eigenpair convergence to machine precision
  • Scale frontier: ~40 sites routine; 48–50 sites at the supercomputer frontier with full symmetry reduction

Recommended for

  • Primary GS solver for small-to-medium clusters: all lattice models up to N≈40; the benchmark oracle for DMRG, QMC, VMC at these sizes
  • Dynamical properties S(q,ω) on small clusters: frustrated magnets (B8, C12 sign-ful) where QMC cannot reach real-frequency data without analytic continuation
  • Finite-cluster Krylov dynamics: quench dynamics, spectral weight computation on clusters where TEBD would be overkill
  • Interior excited states (shift-invert): eigenstate thermalization hypothesis (ETH) studies in specific energy windows
  • Per method-property-map.md (ED profile): applicable when D_H ≲ 10⁷–10⁸ after symmetry reduction; preferred over ED full whenever the full spectrum is not required

Benchmarks

  • ED Lanczos frontier: 48–50 sites (spin-½ kagome/triangular with full symmetry, sublattice coding) [Wietek–Läuchli, arXiv:1804.05028; verified in method-survey.md §1.2].
  • GS energy convergence: residual < 10⁻¹² in < 200 Lanczos steps for gapped 1D Heisenberg (N=36) sandvik_2010_computational.
  • S(q,ω) via continued-fraction: n_CF ≈ 200 Lanczos steps per q point for Heisenberg chain N=24 gagliano_1987_dynamical.
  • n_iter ~ 50 (gapped) vs ~ 500 (critical, near-degenerate) — B6 dependence verified in method-survey.md §1.2.

Key reference: sandvik_2010_computational — covers the Lanczos algorithm, symmetry reduction, continued-fraction computation of S(q,ω), and benchmarks for Heisenberg chains and ladders.

Finite-T Lanczos / thermal pure statesftlm-tpqcontrolledmethod-edjaklic_1994_lanczos
AxisValueNote
M1 tasks / outputsFinite-T thermodynamics (free energy, internal energy, specific heat C_v, susceptibility χ) · finite-T dynamical S(q,ω) (FTLM) · canonical and grand-canonical ensembles (canonical TPQ)Does NOT natively give the full spectrum or level statistics (need ED full for those).
M2 regimeFinite-T (all temperatures including T→0 limit via large β)T=0 ground-state properties accessible via large-β limit; in practice ED Lanczos is cheaper for T=0.
M3 accuracy classStochastic (statistical error ∝ 1/√(R·D_H)); controlled-bias (improvable by increasing R)Error shrinks with system size (typicality) — a key practical advantage at N ≳ 20. Deterministic per random vector; stochastic only across the average.
M4 dimension fit (A1)Any dimension / geometrySame as ED full/Lanczos: geometry enters only through N_nnz and the symmetry block structure.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion / anyD_H = d^N; memory O(D_blk) means the accessible N matches ED Lanczos (not ED full).
M6 entanglement regime (B5)Volume-law toleratedOperates in the full Hilbert space; no entanglement restriction.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; random vectors in Hilbert space, not field-space sampling.
M8 symmetry exploitation (C9/C10)U(1)/SU(2)/Z₂/parity block-diagonalize; translation and point-group reduce D_blkSame symmetry reduction as ED Lanczos; random vectors drawn within the symmetry sector.
M9 time complexityO(R · n_iter · N_nnz) where R = number of random vectors, n_iter = Lanczos steps per vector, N_nnz ≈ N · D_blkAt high T, R can be small (1–10); at low T, thermal fluctuations are larger and R must grow.
M10 memoryO(D_blk)Stores only the current Lanczos basis vectors; same footprint as ED Lanczos.
M11 control knobR (number of random vectors) — controls statistical error 1/√(R·D_H)At large N, even R=1 can give < 1% error (typicality); at small N, need R ≈ 10–100. Standard canonical TPQ (Sugiura–Shimizu 2012/2013) applies e^{-βH/2}|r⟩ via Krylov/Chebyshev — exact to machine precision with no Trotter error; Δτ is a knob only in Trotter-based TPQ implementations.
M12 scale frontierSame as ED Lanczos: ~40 sites routine; frontier ~48–50 sites with full symmetryTypicality means error improves with system size; FTLM/TPQ is most powerful for N ≳ 20 where it beats ED full on memory while giving accurate finite-T results.
M13 primary approximation / biasStochastic trace approximation with R random vectors; error ∝ 1/√(R·D_H) (controlled)Finite cluster approximation for bulk thermodynamics; bulk limit requires finite-size scaling.
M14 hard blocker / failure modeVery low T (large β): thermal fluctuations grow → R must increase to maintain accuracy; long imaginary-time Lanczos evolution becomes numerically delicateD_H wall same as ED; sign-immune so no QMC sign problem. At very low T, ED Lanczos (T=0) or DMRG (finite-T purification) may be preferred.

Cost & scaling

  • Time: O(R · n_iter · N_nnz) with N_nnz ≈ N · D_blk for short-range models
  • Memory: O(D_blk) — same as ED Lanczos; no full matrix or all-eigenvector storage
  • Control knob: R (random vectors) controlling statistical error 1/√(R · D_H); standard Krylov-TPQ has no Δτ (exact imaginary-time evolution); Δτ (error O(Δτ²)) applies only in Trotter-based TPQ implementations
  • Scale frontier: ~40 sites routine; ~48–50 sites at the supercomputer frontier (same as ED Lanczos)

Recommended for

  • Finite-T thermodynamics on frustrated or sign-ful clusters: kagome Heisenberg, pyrochlore, J₁-J₂ square lattice, Kitaev honeycomb — where QMC has a sign problem (C12) and LTRG/XTRG is not yet implemented
  • Medium-sized clusters N ≈ 20–40: where FTLM/TPQ beats ED full (memory O(D_blk) vs O(D_blk²)) while giving accurate finite-T results
  • Specific heat and susceptibility peaks: identification of crossover temperatures, spin-gap detection, Schottky anomalies
  • Finite-T S(q,ω): dynamical neutron-scattering structure factor at finite T on finite clusters
  • Per method-property-map.md: falls under the ED harness skill profile; preferred for D13 finite-T when QMC sign is present and LTRG/XTRG is unavailable

Benchmarks

  • Statistical error ∝ 1/√(R · D_H): for N=20 spin-½ (D_H ≈ 10^6), R=10 gives error ≈ 3×10^{-4} per observable sugiura_2012_thermal.
  • FTLM specific heat for Heisenberg chain N=28: agrees with Bethe ansatz TBA to < 0.5% for T ≥ J/10 with R=20 jaklic_1994_lanczos.
  • Cost per random vector: n_iter ≈ 100–200 Lanczos steps for convergence at intermediate T [method-survey.md §1.3].

Key reference: jaklic_1994_lanczos — original FTLM paper by Jaklič and Prelovšek; establishes the random-vector Lanczos approach for finite-T thermodynamics and dynamical functions in correlated systems.

Kernel polynomial methodkpmcontrolledmethod-edweisse_2006_kernel
AxisValueNote
M1 tasks / outputsDensity of states (DOS) · spectral function A(ω) · dynamical structure factor S(q,ω) · optical / dynamical conductivity · finite-T traces via Chebyshev expansion of e^{-βH}Does NOT provide individual eigenstates, the GS wavefunction, or entanglement data (these require Lanczos).
M2 regimeT=0 spectral densities · finite-T thermodynamic traces (stochastic Chebyshev)Ground-state GS energy not native (no variational principle); spectral densities at all ω.
M3 accuracy classStochastic-trace + controlled-bias; deterministic per random vector, stochastic across R vectorsBias: Gibbs ripples from truncating the Chebyshev series at order M — controlled by Jackson/Lorentz kernel; statistical error ∝ 1/√(R · D_H) from the stochastic trace. Resolution ≈ bandwidth/M improves with M.
M4 dimension fit (A1)Any dimension / geometry — wins for large sparse Hamiltonians in any dSparsity (A4) is the key gate; short-range interactions → small N_nnz per site → cheap matrix-vector products.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion / anyD_H = d^N; memory O(D_H) (stores one or a few state vectors of length D_H).
M6 entanglement regime (B5)Volume-law toleratedNo entanglement restriction; KPM never constructs a wavefunction ansatz — it works with the full D_H-dimensional Hilbert space.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; random vectors sampled in Hilbert space, not field space.
M8 symmetry exploitation (C9/C10)Block structure can reduce D_H (same as ED Lanczos); random vectors drawn within a symmetry sector for sector-resolved DOS/S(q,ω)Symmetry is less critical than for ED (no need to diagonalize): KPM already reaches full D_H scale cheaply.
M9 time complexityO(M · R · N_nnz) where M = #Chebyshev moments, R = #random vectors, N_nnz = #nonzeros in HLinear in M and R; no D_H³ or D_H^2 factor. Scales to very large sparse Hamiltonians inaccessible to Lanczos.
M10 memoryO(D_H) — stores one or a few state vectors of length D_HNo matrix storage (matrix-free); no eigenvalue array; just the state vectors and the recursion coefficients.
M11 control knobM (#Chebyshev moments) — controls spectral resolution ≈ bandwidth/M and Gibbs-ripple suppressionLarger M → finer spectral resolution → higher cost (linearly). R controls statistical error 1/√(R · D_H). Both knobs are independent and can be tuned separately.
M12 scale frontierLargest system sizes accessible among the exact methods: D_H ≈ 10^{9}–10^{10} feasible on a single workstation since only O(D_H) memory is needed for the vectorsWithout the O(D_H²) or O(D_H³) wall, KPM reaches much larger D_H than ED full/Lanczos for the tasks it computes (spectral densities).
M13 primary approximation / biasTruncated Chebyshev expansion at order M — spectral resolution limited to ≈ bandwidth/M; Gibbs oscillations damped by kernel (controlled)The kernel choice (Jackson, Lorentz) determines the trade-off between resolution and Gibbs suppression; this is a controlled, improvable approximation.
M14 hard blocker / failure modeO(D_H) memory wall — must store state vectors of length D_H; not a route to individual eigenstates or the GS wavefunction; poor for sharp spectral features if M is limitedLong-range interactions (A4) raise N_nnz and slow the matrix-vector products; dense Hamiltonians break the sparse advantage.

Cost & scaling

  • Time: O(M · R · N_nnz) — linear in M and R; N_nnz ≈ N · D_blk for short-range models
  • Memory: O(D_H) — only a few state vectors; no eigenvalue storage
  • Control knob: M (#moments) → spectral resolution ≈ bandwidth/M; R (#random vectors) → statistical error 1/√(R · D_H)
  • Scale frontier: D_H ≈ 10^9–10^{10} feasible (memory-limited only by the state vector size)

Recommended for

  • Large sparse Hamiltonians where only spectral densities are needed: Anderson localization, random Hamiltonians, large Heisenberg clusters — situations where D_H is too large for Lanczos eigenstates but O(D_H) memory is feasible
  • DOS and optical conductivity surveys: tight-binding / Hubbard models on large clusters where Lanczos continued-fraction would require many eigenstates
  • Disorder-averaged spectral functions: quench-disorder studies where R random vectors serve dual purpose as disorder-average samples
  • Transport coefficients: electrical/thermal conductivity via the Kubo formula on large clusters
  • Per method-property-map.md (ED profile, extended): a matrix-free spectral-density method that extends the ED harness to larger D_H for spectral tasks

Benchmarks

  • Time cost: O(M · R · N_nnz) — verified scaling in weisse_2006_kernel §III; resolution ≈ bandwidth/M confirmed for the 1D Heisenberg chain DOS.
  • Memory: O(D_H) — storing two state vectors of length D_H; verified in [method-survey.md §1.4].
  • Stochastic-trace error: ∝ 1/√(R · D_H) — for D_H = 10^8 and R = 10, error ≈ 3×10^{-5} per spectral bin weisse_2006_kernel §II.D.
  • KPM vs Lanczos for DOS: KPM at M = 2000 gives resolution ≈ bandwidth/2000; Lanczos continued-fraction achieves similar resolution at comparable cost but requires a starting vector near the target state weisse_2006_kernel §IV.

Key reference: weisse_2006_kernel — authoritative Reviews of Modern Physics paper by Weisse, Wellein, Alvermann, and Fehske; defines the Chebyshev expansion, kernel choices (Jackson, Lorentz, Dirichlet), stochastic-trace evaluation, and applications to DOS, dynamical conductivity, and spectral functions. The primary reference for all KPM implementations.

Full configuration interactionfciexactmethod-edbooth_2009_fermion
AxisValueNote
M1 tasks / outputsExact ground-state energy and wavefunction · few excited states (EOM-FCI / excited-state CI) · one- and two-particle reduced density matrices (1-RDM, 2-RDM) · benchmark for approximate quantum-chemistry methods (CCSD, CASPT2, DMRG-CI)FCI is the benchmark target; other methods are benchmarked against it.
M2 regimeT=0 ground state (primary); few excited states via sector targetingFinite-T FCI possible (canonical-ensemble ED) but rarely used for chemistry; more relevant in lattice-model context.
M3 accuracy classNumerically exact, deterministic (exact FCI); controlled-bias stochastic (FCIQMC, selected-CI)Exact FCI = numerically exact within the chosen single-particle basis. FCIQMC = stochastic, error ∝ 1/√N_w·N_steps; selected-CI = variational upper bound with improvable selection threshold.
M4 dimension fit (A1)Molecules and lattice clusters in any geometry (A1 irrelevant — FCI is basis-set, not lattice, driven)Applicable to any molecular or cluster geometry; cost set by #orbitals, not spatial dimension.
M5 statistics & local dim (A3)Fermions only; local dimension d = 4 per spatial orbital (empty/up/down/doubly-occupied)d=4 per orbital is a factor of 4 harder than spin-½ per site; D_H = C(2·N_orb, N_e) grows exponentially in N_orb.
M6 entanglement regime (B5)Volume-law toleratedNo entanglement restriction; exact FCI stores the full coefficient vector of length D_H.
M7 sign-problem dependence (C12)Sign-immune (exact FCI, selected-CI) · stochastic with sign (FCIQMC, controlled by annihilation)Exact FCI: sign-immune (direct matrix diagonalization). FCIQMC: fermionic sign problem controlled by the annihilation step, giving unbiased results in principle booth_2009_fermion.
M8 symmetry exploitation (C9/C10)Sᶻ (N_↑, N_↓) + S² (spin symmetry) + spatial point-group (irreducible representations) reduce D_HPoint-group symmetry (C₂v, D₂h, etc.) of the molecule reduces the Hamiltonian block size; key cost lever analogous to translation symmetry in lattice ED.
M9 time complexityExponential in N_orb: O(D_H³) for direct dense diagonalization (LAPACK); O(N_orb² · D_H) per Davidson/Lanczos iteration (dominant two-electron sigma-vector step)In practice, the Davidson iterative solver is used (analogous to Lanczos); memory O(D_H) per vector. Exponential wall prevents scaling beyond ~20 orbitals for exact FCI.
M10 memoryO(D_H) for iterative (Davidson/Lanczos FCI); O(D_H²) for dense diagonalizationDavidson FCI: stores a few vectors of length D_H; dense FCI (LAPACK): stores the full matrix D_H × D_H.
M11 control knobN_orb (number of active orbitals) — the primary convergence parameter; for selected-CI: selection threshold ε₁ (controls which determinants are included)Increasing N_orb → exponential cost increase. Selected-CI threshold ε₁ → 0 → exact FCI. FCIQMC: N_w (number of walkers) controls statistical error.
M12 scale frontier~18–20 spatial orbitals in routine codes (Davidson FCI); ~10¹² determinants at the distributed-memory frontier (trillion-determinant FCI, JCTC 2024); selected-CI extends effective reach via sparsityThe 10¹² frontier uses specialized bit-string algorithms and distributed memory; routine use is ~N_orb ≈ 16 (CASSCF active space).
M13 primary approximation / biasBasis-set incompleteness (choice of single-particle orbitals); none within the chosen orbital space — FCI is exact in the given basisExtrapolation to the complete-basis-set (CBS) limit is the main source of remaining error. For selected-CI: selection threshold bias (variational, controlled).
M14 hard blocker / failure modeExponential scaling in N_orb: D_H = C(2·N_orb, N_e) grows as ~4^{N_orb} (at half-filling) — the d^N wall translated to orbital language; standard FCI is limited to active spaces of ~20 orbitalsFCIQMC and selected-CI push the frontier but cannot escape the exponential wall for strongly correlated systems with many active orbitals.

Cost & scaling

  • Time: exponential in N_orb; Davidson iterative FCI ≈ O(N_orb² · D_H) per iteration (dominant two-electron sigma-vector); dense diag O(D_H³)
  • Memory: O(D_H) for iterative (Davidson); O(D_H^2) for dense; D_H = C(2·N_orb, N_e)
  • Control knob: N_orb (active orbitals, exponential cost); selected-CI threshold ε₁ (improvable bias); FCIQMC walker count N_w (statistical error 1/√N_w)
  • Scale frontier: ~18–20 orbitals routine; ~10¹² determinants at the distributed-memory frontier (JCTC 2024)

Recommended for

  • Quantum chemistry benchmarks: small molecules (H₂, N₂, F₂, H₂O) in moderately large basis sets; the exact reference for all approximate methods
  • Strongly correlated active spaces: any system with static correlation (B8 frustration in the molecular sense: degeneracies, open shells) up to N_orb ≈ 20
  • Lattice Hubbard / Anderson models in orbital language: FCI is identical to lattice ED full when the single-particle basis is the site basis (d=4 Hubbard); selected-CI (FCIQMC, CIPSI) pushes to larger N_orb
  • FCIQMC / selected-CI for larger active spaces: N_orb ≈ 30–100 with near-FCI accuracy via stochastic sampling of the dominant determinants
  • Per method-property-map.md (ED profile): FCI is the quantum-chemistry arm of the ED harness; /method-ed skill; fermionic, any geometry

Benchmarks

  • Routine FCI scale: ~18–20 active orbitals in codes like PySCF, OpenFCI, Dice (CASSCF active spaces) [method-survey.md §1.5].
  • Distributed-memory frontier: ~10¹² determinants (trillion-determinant FCI) [method-survey.md §1.5, citing JCTC 2024].
  • FCIQMC on C₂ (cc-pVDZ, 28 electrons, 28 orbitals): energy converged to within 0.1 mEh of exact FCI using N_w ≈ 10^8 walkers booth_2009_fermion.
  • Selected-CI (CIPSI/heat-bath CI): near-FCI accuracy for N_orb ≈ 30–50 with selected-CI threshold ε₁ ≈ 10^{-5} [method-survey.md §1.5].

Key reference: booth_2009_fermion — original FCIQMC paper by Booth, Thom, and Alavi; introduces the Monte Carlo imaginary-time projection in determinant space, demonstrating exact FCI-quality results for systems beyond the reach of conventional FCI (C₂ molecule in cc-pVDZ, 10⁸–10¹⁰ determinants); the foundational reference for stochastic approaches to FCI.

2 Classical Monte Carlo & critical RG 3

Metropolis Monte Carlometropolis-mcexactclassical-mcsandvik_2010_computational
AxisValueNote
M1 tasks / outputsFinite-T thermodynamics: energy, magnetization, magnetic susceptibility, specific heat, Binder cumulants, structure factors, correlation functionsNOT ground-state energy or quantum amplitudes; NOT critical exponents (use MCRG).
M2 regimeFinite temperature (thermal equilibrium)Classical models only, or quantum models mapped sign-free to a classical representation (e.g. transfer-matrix / SSE at finite β).
M3 accuracy classNumerically exact (within Boltzmann statistics), stochasticStatistical error ∝ 1/√(N_s/τ_auto); systematic bias only from finite τ_auto (equilibration) and finite lattice.
M4 dimension fit (A1)Any (1D / 2D / 3D / ∞-D)Per-sweep cost O(N) is independent of geometry; suited to classical Ising, Heisenberg, XY, Potts models in any dimension.
M5 statistics & local dim (A3)Classical spin models (Ising d=2, Potts d=q, continuous O(n) spins); sign-free quantum mappingsNot applicable to genuine quantum models with a sign problem. Local dim d only affects the update cost per site, not the algorithmic structure.
M6 entanglement regime (B5)N/A — classical model; no entanglement restrictionFor quantum models accessed via sign-free mappings, entanglement is implicit in the weight.
M7 sign-problem dependence (C12)Requires sign-free (classical models or sign-free quantum mappings)Not a quantum MC method per se. A genuine fermionic/frustrated quantum model is NOT accessible unless a sign-free classical representation exists.
M8 symmetry exploitation (C9/C10)Global symmetry used to measure order parameters (e.g. Z₂ for Ising, O(n) for Heisenberg); translation used for structure-factor measurementSymmetry does not reduce per-sweep cost (unlike ED block-diagonalization), but allows optimized estimators.
M9 time complexityO(N) per sweep; total O(N · τ_auto · N_s); near T_c: τ_auto ~ L^z with z = 2.1665(12) (2D Ising, Glauber/Metropolis, Nightingale–Blöte 1996)Critical slowing down τ_auto ~ L^z dominates at a continuous transition (B6). Larger lattice L is far more expensive at T_c.
M10 memoryO(N) — stores spin configuration + neighbor listMinimal memory footprint; scales to thermodynamic-limit lattices (millions of sites).
M11 control knobN_s (number of sweeps after equilibration) — statistical error ∝ 1/√(N_s/τ_auto)Increasing N_s systematically reduces statistical error; τ_auto must be measured to set N_s reliably.
M12 scale frontierMillions of sites in practice (off-critical); thousands of sites near T_c (critical slowing down limits effective sample count)The method scales to very large L away from criticality; near T_c the effective sample count is N_s/τ_auto.
M13 primary approximation / biasNone — sampling from the exact Boltzmann distribution (within statistical error and given exact weights)Finite-size effects require finite-size scaling to extract thermodynamic-limit behavior.
M14 hard blocker / failure modeCritical slowing down (τ_auto ~ L^z, z = 2.1665(12)) near a continuous transition; frustration/glassiness (B8) inflates τ_auto further, causing metastability; NOT applicable to sign-ful quantum modelsCritical slowing down is partially defeated by cluster algorithms (§2.2 of method-survey.md).

Cost & scaling

  • Time: O(N) per sweep; O(N · τ_auto · N_s) total; at criticality τ_auto ~ L^z with z = 2.1665(12) for 2D Ising
  • Memory: O(N) — spin configuration and neighbor list only
  • Control knob: N_s (sweeps) — controls statistical error as 1/√(N_s/τ_auto)
  • Scale frontier: millions of sites off-critical; thousands near T_c before critical slowing down dominates

Recommended for

  • Classical Ising / Heisenberg / XY / Potts in any dimension at finite temperature, away from the critical point
  • Finite-size scaling studies at or near T_c (combined with Binder cumulants), though cluster MC is preferred when z matters
  • Sign-free quantum models mapped to classical weights (e.g. transverse-field Ising via Suzuki–Trotter at finite β)
  • Per method-property-map.md: classical models (B8 unfrustrated) at finite T; not for critical-exponent extraction (use MCRG) or quantum sign-ful models (use MPS/PEPS/VMC)

Benchmarks

  • 2D Ising T_c/J = 2/ln(1+√2) ≈ 2.2692 (exact Onsager): located to < 0.01% precision by Binder-cumulant crossing at L ≈ 32–64 sandvik_2010_computational.
  • Dynamical critical exponent (local Metropolis/Glauber): z = 2.1665(12) for 2D Ising nightingale_1996_dynamic [verified in method-survey.md §2.1].
  • 2D Ising critical specific heat: C ~ L^{α/ν} with α = 0 (log divergence) captured correctly after N_s ≫ τ_auto ~ L^{2.1665} sweeps sandvik_2010_computational.

Key reference: sandvik_2010_computational — comprehensive treatment of classical and quantum Monte Carlo for spin systems, including Metropolis/heat-bath algorithms, finite-size scaling with Binder cumulants, and structure-factor estimators.

Cluster MC / Wang–Landau / parallel temperingcluster-mcexactclassical-mcwolff_1989_collective
AxisValueNote
M1 tasks / outputsSame thermodynamic observables as Metropolis MC: energy, magnetization, susceptibility, Binder cumulants, structure factors, correlation functions; Wang–Landau additionally returns the full density of states g(E)Cluster updates are most valuable for making large-L critical-region measurements affordable.
M2 regimeFinite temperature (thermal equilibrium)Classical models; or quantum models with a sign-free classical representation.
M3 accuracy classNumerically exact (Boltzmann statistics), stochasticStatistical error ∝ 1/√(N_s/τ_auto); cluster MC's win is that τ_auto is orders of magnitude smaller than Metropolis near T_c.
M4 dimension fit (A1)Any (1D / 2D / 3D / ∞-D)SW and Wolff are most impactful in 2D and 3D where critical slowing down is severe; applicable to any dimension where the FK mapping holds.
M5 statistics & local dim (A3)Classical spin models (Ising, Heisenberg O(n), Potts) with ferromagnetic / unfrustrated couplingsFK cluster construction requires non-negative bond weights; does not apply to general antiferromagnetic or frustrated models.
M6 entanglement regime (B5)N/A — classical modelSame as Metropolis MC.
M7 sign-problem dependence (C12)Requires sign-free / unfrustrated (Fortuin–Kasteleyn construction requires ferromagnetic couplings)Frustration breaks the FK bond construction — this is the hard blocker (see M14).
M8 symmetry exploitation (C9/C10)Global spin-flip symmetry exploited by Wolff (projects a random direction, builds the cluster in that projected space); SW flips each cluster independentlyTranslation / point-group symmetry used for structure-factor estimators.
M9 time complexityO(N) per sweep (amortized over cluster size); near T_c: τ_auto ~ L^z with Swendsen–Wang z ≈ 0.22 and Wolff: ~ ln L equilibrium-autocorrelation; nonequilibrium-relaxation fits give a larger effective z ≈ 1.2 for WolffNear-O(N) effective cost at criticality — the decisive advantage over Metropolis (z = 2.1665(12)). Parallel tempering multiplies total cost by N_replicas.
M10 memoryO(N) — spin configuration; cluster labels O(N) (union-find)Cluster label array adds a small constant factor over plain Metropolis.
M11 control knobN_s (sweeps/cluster updates) — controls statistical error; for Wang–Landau: the modification factor f → 1 controls flatness of g(E); for parallel tempering: N_replicas and temperature gridStatistical error ∝ 1/√(N_s/τ_auto_eff) where τ_auto_eff ≪ τ_auto(Metropolis) near T_c.
M12 scale frontierMillions of sites even near T_c (SW/Wolff make large-L critical-region sampling affordable); parallel tempering / Wang–Landau: hundreds of sites (limited by N_replicas × L^d)Cluster algorithms enable finite-size scaling studies at L ~ 10³–10⁴ near T_c that are prohibitive with Metropolis.
M13 primary approximation / biasNone — samples from exact Boltzmann distribution (for cluster MC); Wang–Landau has a systematic ln(f) bias that vanishes as f → 1Finite-size effects remain; Wang–Landau convergence must be checked by flatness criterion.
M14 hard blocker / failure modeFrustration (B8) breaks the Fortuin–Kasteleyn cluster construction — SW/Wolff do not apply; frustrated/spin-glass systems require parallel tempering or simulated annealing (at higher cost and with no z suppression); sign-ful quantum models inaccessibleParallel tempering is a workaround for glassy landscapes but does not reduce z; it incurs an N_replicas overhead.

Cost & scaling

  • Time: O(N) per cluster sweep (union-find construction); effective critical cost O(N · L^z) with z ≈ 0.22 (SW) vs z ≈ 2.17 (Metropolis); parallel tempering: × N_replicas
  • Memory: O(N) — spin configuration + cluster label array
  • Control knob: N_s (sweeps) + N_replicas (parallel tempering) — controls statistical error and barrier crossing
  • Scale frontier: millions of sites near T_c for SW/Wolff; hundreds of sites for parallel tempering with many replicas

Recommended for

  • Ferromagnetic or unfrustrated classical models (Ising, Potts, O(n)) near their continuous phase transitions — SW and Wolff are the standard choice
  • Large-L finite-size scaling at criticality (B6): SW/Wolff make L ~ 10³–10⁴ routine where Metropolis is stuck
  • Glassy or rough free-energy landscapes (B8, D15): parallel tempering and Wang–Landau for landscape exploration
  • Per method-property-map.md: classical models, sign-free, B8 unfrustrated (for SW/Wolff); B8 frustrated → parallel tempering only (no z suppression)

Benchmarks

  • Swendsen–Wang dynamical critical exponent: z ≈ 0.22 for 2D Ising model (equilibrium-autocorrelation convention) [verified in method-survey.md §2.2].
  • Wolff: ~ ln L equilibrium-autocorrelation for 2D Ising; nonequilibrium-relaxation fits give a larger effective z ≈ 1.2 for Wolff [verified in method-survey.md §2.2].
  • Comparison at L = 512, 2D Ising at T_c: SW requires ~L^{0.22} effective sweeps vs Metropolis ~L^{2.1665} sweeps per independent sample — roughly 10⁵ speedup (ratio L^{z_M − z_SW} = L^{1.95} ≈ 1.3×10⁵ at L=512).
  • 2D Ising T_c/J = 2.2692 (Onsager): cluster MC locates to sub-0.01% at L = 1024 with O(10⁴) cluster updates swendsen_wang_1987_nonuniversal.

Key reference: wolff_1989_collective — original Wolff single-cluster algorithm with the demonstration that z is effectively ~ ln L for 2D Ising, making large-L critical simulations affordable.

Monte Carlo renormalization groupmcrgcontrolledmethod-mcrgwu_2017_variational
AxisValueNote
M1 tasks / outputsCritical exponents y_t, y_h (→ ν, β, γ, δ, η); critical coupling K_c; critical manifold and its tangent space; renormalized coupling constants as a function of RG scaleNOT ground-state energies, NOT thermodynamic-limit observables (e.g., magnetization or susceptibility away from criticality), NOT the full spectrum.
M2 regimeFinite temperature — specifically at or near the critical fixed pointMCRG is designed for the critical fixed point; away from T_c it gives flow information but not exponents.
M3 accuracy classControlled-bias, stochasticStatistical error from MC sampling ∝ 1/√N_s; systematic bias from finite block-spin operator basis (controlled by expanding the basis) and finite lattice (controlled by L → ∞).
M4 dimension fit (A1)Any dimension (1D / 2D / 3D / ∞-D) — applicable to classical spin models in any dimensionBest applied where a critical fixed point exists (2D Ising / Potts, 3D Ising / Heisenberg, etc.).
M5 statistics & local dim (A3)Classical spin models: Ising (d=2), Potts, O(n) continuous spinsApplies to any classical statistical model with a critical point; local dim only affects MC cost.
M6 entanglement regime (B5)N/A — classical modelNo entanglement restriction; thermodynamic configurations sampled.
M7 sign-problem dependence (C12)Requires sign-free / classical modelsMCRG operates on classical Boltzmann weights; no sign problem. For quantum models, the sign-free auxiliary-field / world-line mapping must exist first.
M8 symmetry exploitation (C9/C10)Symmetry constrains the operator basis for the RG matching (fewer independent coupling constants); translation used in momentum-space matching variantsSymmetry reduces the number of couplings that must be matched, improving accuracy.
M9 time complexityMC sampling cost O(N · τ_auto · N_s) × #RG iterations; cluster MC (§2.2) used to reduce τ_auto; variational MCRG (Wu–Car 2017) significantly alleviates critical slowing downLarger L near T_c improves exponent accuracy (finite-size corrections ∝ L^{-ω}); #RG steps typically 3–5 per lattice.
M10 memoryO(N) — spin configurations + blocked configurations at each RG levelMemory is dominated by storing configurations for the blocked lattice hierarchy.
M11 control knobLattice size L (controls finite-size corrections ~ L^{-ω}); number of RG iterations; operator-basis size (controls truncation bias)Accuracy of exponents improves with L and operator-basis completeness; convergence checked by stability of eigenvalues across RG steps.
M12 scale frontierLarge lattices near T_c (L ~ 10²–10³ typical; L ~ 10⁴ with cluster MC or variational MCRG); exponent accuracy ~ 10⁻³–10⁻⁴ at the frontierFinite-size corrections to exponents scale as L^{-ω} (correction-to-scaling exponent); 3–4 RG levels require L ≥ 2^4 = 16 at minimum, L ≥ 256 for percent-level precision.
M13 primary approximation / biasFinite block-spin operator basis (truncation of RG operator space) — controlled by expanding the basis; finite-size effects — controlled by increasing LBoth biases are systematically reducible; variational MCRG optimizes the basis to minimize truncation error.
M14 hard blocker / failure modeNot applicable to non-critical systems or for computing thermodynamic-limit observables / ground-state energies — MCRG extracts only fixed-point data; requires a valid critical point (no topological phases without Landau order); frustrated / glassy systems (B8) complicate locating T_cAway from criticality, MCRG returns RG-flow information but not universal exponents.

Cost & scaling

  • Time: MC sampling O(N · τ_auto · N_s) + O(N) per RG blocking iteration; cluster MC cuts τ_auto ~ L^{0.22} (SW) vs L^{2.17} (Metropolis)
  • Memory: O(N) — configurations at each RG level (shrink by factor 2–4 per level)
  • Control knob: L (lattice size, controls finite-size corrections ~ L^{-ω}); operator-basis size; #RG steps
  • Scale frontier: L ~ 10²–10³ routine; L ~ 10⁴ with variational MCRG / cluster sampling

Recommended for

  • Critical-exponent extraction for classical spin models at their continuous phase transition (B6 = gapless/critical)
  • Any dimension where a critical fixed point exists: 2D Ising, 2D Potts, 3D Ising, 3D Heisenberg, O(n) universality classes
  • Cross-validation of exponents from finite-size scaling (Binder cumulants) or field-theoretic calculations
  • Per method-property-map.md (MCRG profile): applicable when B6 = critical; blocked when the goal is thermodynamic-limit observables or GS energies

Benchmarks

  • 2D Ising y_t = 1.000, y_h = 1.875 (exact from Onsager/Kaufman/Yang): recovered to < 0.1% by MCRG with L = 256 and 4 RG levels swendsen_1979_montecarlo.
  • 3D Ising exponents from variational MCRG (Wu–Car 2017): ν = 0.6299(4) (cf. conformal-bootstrap 0.6299(2)), η = 0.0363(4) wu_2017_variational.
  • Triangular-lattice ferromagnetic Ising model: MCRG recovers the 2D Ising universality class y_t = 1.00(1), y_h = 1.875(5) — a non-square-lattice test of the method's geometry-independence guo_2005_montecarlo.
  • Finite-size corrections to exponents scale as L^{-ω} with ω ≈ 0.83 (3D Ising) — L ≥ 64 required for sub-percent precision [verified in method-survey.md §2.3].

Key reference: wu_2017_variational — variational MCRG that optimizes the block-spin rule to minimize a variational distance, removes critical slowing down, and extracts the critical manifold tangent space; provides benchmark exponents for 2D Ising and 3D Ising.

3 Ground-state tensor networks 4

Density-matrix renormalization groupdmrgcontrolledmethod-mpsschollwoeck_2010_density
AxisValueNote
M1 tasks / outputsGS energy & observables · gap & few excited states (targeted DMRG) · correlation length (transfer-matrix spectrum) · entanglement entropy & spectrum (free from MPS) · order parameters · S(q,ω) (post-process via DMRG + correction-vector or TDVP)Finite-T and real-time are separate algorithms (§5 of survey); dynamical spectra require additional passes.
M2 regimeT=0 ground state (primary); thermodynamic limit via iDMRG/VUMPSFinite-T via purification/METTS (separate algorithm using same MPS machinery).
M3 accuracy classControlled-bias (truncation error ε_trunc is the systematic knob), deterministicVariational upper bound on energy at each fixed χ; bias → 0 as χ → ∞.
M4 dimension fit (A1)1D / quasi-1D near-exact; walls in 2D (cylinder width W → cut entropy ∝W → χ ~ e^{cW})1D gapped: small constant χ; 1D critical: χ grows polynomially; 2D: feasible only for narrow cylinders (W ≲ 6–8 for spin-½).
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion (Jordan-Wigner or MPO fermion)Local dimension d appears in cost as O(d·χ³) per site; large d (Hubbard d=4, soft-core boson) raises prefactor but not the exponent. Fermions handled natively via Jordan-Wigner string in MPS.
M6 entanglement regime (B5)Area-law (cheap, constant χ) / area+log 1D-critical (χ grows polynomially)Volume-law states (excited, thermal, real-time beyond t*) are inaccessible at reasonable χ.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; no sign problem regardless of frustration or doping.
M8 symmetry exploitation (C9/C10)U(1) / SU(2) block-sparse tensors; Z₂ / parity; translation (iDMRG unit cell); point group (quantum numbers per site)SU(2) multiplet tensors give ≳10× speedup at fixed accuracy; U(1) particle-number conservation shrinks block sizes roughly by √χ.
M9 time complexityO(N · w · d · χ³) per sweepN = sites, w = MPO bond (long-range → w large), d = local dim, χ = bond dim. Leading cost is from the two-site update contraction: O(χ³) at each of N sites (w, d prefactors suppressed; full cost O(N·w·d·χ³) — see table value).
M10 memoryO(d · χ²)Stores one MPS tensor per site (d × χ × χ) plus the environment and Hamiltonian MPO; dominant term is O(d·χ²) per tensor.
M11 control knobBond dimension χ (truncation error ε_trunc = discarded weight)Increasing χ reduces truncation error; convergence checked by extrapolating energy vs ε_trunc → 0.
M12 scale frontierLarge 1D chains (N ~ 10³–10⁴ routine); narrow 2D cylinders up to W ~ 6–8; thermodynamic limit directly via iDMRG/VUMPSPractical χ ceiling set by χ³ time cost; supercomputer runs reach χ ~ 10⁴ for critical 1D chains.
M13 primary approximation / biasFinite-χ truncation (controlled): the MPS ansatz restricts entanglement; truncation error ε_trunc → 0 as χ → ∞Bias is systematic and controlled; energy is a variational upper bound at fixed χ.
M14 hard blocker / failure mode2D systems: χ ~ e^{cW} per cylinder width W (exponential wall); volume-law entanglement (real-time t > t*, excited states, thermal states); long-range interactions raise MPO bond w, inflating costFrustrated 2D magnets, doped Hubbard at moderate χ, real-time beyond the entanglement barrier all push beyond practical reach.

Cost & scaling

  • Time: O(N · w · d · χ³) per sweep (N sites, MPO bond w, local dim d, bond dim χ)
  • Memory: O(d · χ²) — one MPS tensor per site plus environment tensors
  • Control knob: bond dimension χ — controls truncation error ε_trunc (discarded weight)
  • Scale frontier: N ~ 10³–10⁴ (1D); narrow cylinders W ≲ 6–8 (2D); thermodynamic limit via iDMRG/VUMPS

Recommended for

  • Primary GS solver for 1D gapped systems (spin chains, Hubbard chain, t-J chain): essentially exact at moderate χ ~ 100–500
  • 1D critical systems (XXX chain, transverse-field Ising at Jc): larger χ (polynomial growth with L), finite-size scaling needed
  • Quasi-1D / ladder geometries: ladders, zig-zag chains, narrow cylinders — gold standard, χ scales exponentially with ladder width
  • 2D as benchmarking tool on cylinders: torus / cylindrical Hubbard, Heisenberg on width-W cylinders (W ≤ 6 practical)
  • Per method-property-map.md: optimal when A1=1D, B5=area-law or area+log, C12=sign-ful (frustrated), C9/C10 present (big speedup); disfavored for A1=2D bulk, D13=real-time, or B5=volume-law

Benchmarks

  • 1D Heisenberg chain: GS energy per site e₀ = −0.44315 J (N→∞, iDMRG χ~500) — reproduced to 6+ digits by χ extrapolation schollwoeck_2010_density.
  • 2D Heisenberg square lattice (cylinder W=8, χ=4096): e₀ ≈ −0.3347 J/site, consistent with QMC benchmark −0.33699 within 0.5% [survey §3.1].
  • Entanglement entropy scaling: S ~ (c/3) ln(ξ) for 1D CFT (c=1), verified for XXX chain χ~200 [survey §3.1].

Key reference: schollwoeck_2010_density — the authoritative modern reference: formulates DMRG as MPS variational optimization, covers finite/infinite systems, symmetries, excited states, dynamical spectra, and cost scaling.

Projected entangled pair statespeps-ipepscontrolledmethod-pepsnaumann_2023_introduction
AxisValueNote
M1 tasks / outputsGS energy & order parameters · correlation functions · correlation length (boundary transfer matrix) · thermodynamic limit directly (iPEPS) · entanglement entropy (approximate; from the finite-χ_env CTMRG environment, not a limitation of the PEPS ansatz) · topological order (representable at fixed D)Dynamics and finite-T require specialized extensions (thermal-state PEPS, D-doubling); excitations via PEPS tangent space.
M2 regimeT=0 ground state (primary); thermodynamic limit via iPEPSFinite-T harder: thermal-state PEPS or D-doubling approach; real-time dynamics via time-evolved PEPS (much harder).
M3 accuracy classControlled-bias (finite-D truncation + approximate contraction), deterministic; variational upper bound in the limit D→∞, χ_env→∞; at finite χ_env the approximate (CTMRG/boundary-MPS) contraction can break the strict variational boundBoth D (ansatz) and χ_env (contraction accuracy) must be converged; two independent knobs.
M4 dimension fit (A1)2D (its home); 3D explored but expensive; 1D → use MPS insteadNatural for square, triangular, kagome, honeycomb lattices. Quasi-1D (ladders) → MPS preferred.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion (fPEPS with parity tensors)Fermions need parity tensors (fermionic PEPS / fPEPS) to handle the fermionic sign consistently; modest additional overhead.
M6 entanglement regime (B5)2D area-law at fixed D; B6 criticality and B8 frustration inflate DArea-law GS at manageable D ~ 4–16; critical / frustrated states need larger D for quantitative accuracy.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; applicable to frustrated magnets and doped fermions without sign problem.
M8 symmetry exploitation (C9/C10)U(1) / SU(2) / Z₂ block-sparse virtual tensors; lattice point-group symmetry via symmetric iPEPSSymmetric tensors reduce cost and improve convergence; AD-based optimization fully compatible with U(1)/SU(2) symmetry.
M9 time complexityObservable eval O(χ_env³ D⁶) via CTMRG/boundary-MPS; variational full-update / AD optimization O(D¹⁰)–O(D¹²); simple update ~O(D⁵) per stepContraction bottleneck dominates. χ_env ∝ D² is the standard choice. Simple update is fast but uses a mean-field environment.
M10 memoryO(χ_env² D⁴) for the environment tensors; O(D⁵) per site tensorEnvironment tensors dominate for large χ_env.
M11 control knobVirtual bond D (ansatz bias) + environment bond χ_env (contraction bias)Both must be extrapolated: energy vs 1/D and energy vs 1/χ_env. Converged result requires D and χ_env both large.
M12 scale frontierThermodynamic limit directly (iPEPS, single unit cell); finite PEPS on patches up to ~20×20; practical D ≲ 16 for spin-½iPEPS is genuinely a thermodynamic-limit method; finite-size effects absent in principle for translationally invariant systems.
M13 primary approximation / bias(1) Finite-D MPS-ansatz truncation (controlled, → 0 as D → ∞); (2) approximate contraction bias via χ_env (controlled, → 0 as χ_env → ∞)Two independent biases must both be converged; simple-update environment is a less controlled approximation of the full environment.
M14 hard blocker / failure modeContraction cost scales as O(D¹⁰⁻¹²) for optimization → accessible D is small (typically D ≤ 16); 3D extremely expensive; real-time dynamics entanglement barrier; non-Hermitian / open system extensions immatureFor strongly frustrated 2D magnets, very large D may be needed to resolve competing orders — computationally prohibitive at present.

Cost & scaling

  • Time: O(χ_env³ D⁶) for observable evaluation; O(D¹⁰)–O(D¹²) for variational full-update/AD optimization; ~O(D⁵) for simple update per step
  • Memory: O(χ_env² D⁴) environment + O(D⁵) per site tensor
  • Control knob: virtual bond D (ansatz accuracy) and environment bond χ_env (contraction accuracy) — both must be converged
  • Scale frontier: thermodynamic limit (iPEPS); practical D ≲ 16 for spin-½ on standard hardware

Recommended for

  • Primary GS solver for 2D lattice models: 2D Heisenberg on square/triangular/kagome, 2D Hubbard model at half-filling and doped, frustrated J₁-J₂ models
  • Sign-ful 2D regimes (frustrated magnets, doped fermions): sign-immune alternative to QMC
  • Thermodynamic-limit observables without finite-size extrapolation (iPEPS)
  • Per method-property-map.md: optimal when A1=2D, B5=2D-area-law, C12=sign-ful; disfavored for A1=3D, D13=real-time, or D=very large (cost prohibitive)

Benchmarks

  • 2D Heisenberg square lattice (iPEPS, D=10, χ_env=100): e₀ = −0.66934(4) J/site, consistent with QMC benchmark −0.66934(4) [Corboz 2016; survey §3.2].
  • 2D Hubbard (iPEPS, D=8, half-filling): double occupancy and energy vs doping in good agreement with diagrammatic QMC at moderate U/t [survey §3.2].
  • Simple-update D=8 vs full-update D=8: full-update lowers energy by ~0.5–1% and is more accurate for frustrated models [survey §3.2].

Key reference: naumann_2023_introduction — pedagogical introduction to iPEPS with AD-based optimization; covers CTMRG environment, variational algorithms, and modern AD workflow; most accessible entry point for practitioners.

Tensor RG (TRG / HOTRG)trg-hotrgcontrolledmethod-pepsxie_2012_coarse
AxisValueNote
M1 tasks / outputsFree energy & partition function · critical temperature T_c · critical exponents & central charge (from transfer-matrix / RG-spectrum eigenvalues) · correlation length · thermodynamic observables; also contracts 2D quantum-thermal & PEPS networksCFT data (scaling dimensions, central charge c) extracted from the fixed-point spectrum of the transfer matrix or RG superoperator.
M2 regimeFinite-T (classical equilibrium); T=0 quantum via path-integral / PEPS contractionImaginary-time direction mapped onto a spatial leg; real-time dynamics not natural in this framework.
M3 accuracy classControlled-bias (finite-χ truncation at each coarse-graining step), deterministicBias → 0 as χ → ∞; no stochastic error. TNR removes short-range correlations that contaminate the TRG fixed point.
M4 dimension fit (A1)2D (TRG / CTMRG); 2D & 3D (HOTRG)TRG cost grows as O(χ⁶) per step (feasible in 2D); HOTRG cost O(χ^{4d−1}) (d = spatial dimension here, not local Hilbert dim) makes 3D practical at moderate χ but expensive.
M5 statistics & local dim (A3)Classical spins / hard-core variables; quantum via path integral (bosons, spins); fermions need Grassmann tensor extensionsLocal dimension d absorbed into the initial tensor; classical Ising/Potts are natural applications. Fermionic path integrals require careful sign handling via Grassmann tensors.
M6 entanglement regime (B5)Area-law classical correlations; handles critical (log-growing) correlations at finite χ via RG fixed-pointUnlike MERA, TRG/HOTRG do not explicitly install disentanglers; TNR variant removes the short-range entanglement that spoils the TRG fixed point.
M7 sign-problem dependence (C12)Sign-free (real Boltzmann weights for classical models)Classical statistical models have non-negative Boltzmann weights; no sign problem. Quantum path integrals can carry a sign, but this is treated by explicit tensor structure rather than Monte Carlo sampling.
M8 symmetry exploitation (C9/C10)Global symmetry (Z₂, U(1), etc.) enforced via symmetric tensors; translation exploited by the coarse-graining geometrySymmetric tensor implementations reduce the effective bond dimension and improve convergence near criticality.
M9 time complexityTRG O(χ⁶) per step (2D); HOTRG O(χ^{4d−1}) per step = O(χ⁷) in 2D, O(χ¹¹) in 3DDO NOT confuse with ATRG O(χ^{2d+1}) — that is a different (cheaper) algorithm. TNR scales higher than HOTRG but yields clean scale invariance.
M10 memoryTRG O(χ⁴); HOTRG O(χ^{2d}) (d = spatial dimension here, not local Hilbert dim) = O(χ⁴) in 2D, O(χ⁶) in 3DStores one or two local tensors per coarse-graining step; environment tensors in CTMRG are O(χ³).
M11 control knobBond dimension χ (number of singular values retained after each truncation)Increasing χ reduces the truncation bias; convergence monitored via free-energy density as a function of χ.
M12 scale frontierThermodynamic limit directly (coarse-graining runs to infinite volume); classical 2D models routine at χ ~ 20–60; 3D at χ ~ 10–20Iterative coarse-graining naturally accesses the infinite-volume fixed point without finite-size extrapolation.
M13 primary approximation / biasFinite-χ truncation at each coarse-graining step (controlled); TRG additionally accumulates short-range correlation errors near criticality (partially fixed by TNR)Truncation bias is the only source of error; systematically controlled by increasing χ.
M14 hard blocker / failure mode3D at large χ (O(χ¹¹) wall); sign-ful quantum path integrals; real-time dynamics; volume-law entanglement phasesFrustrated quantum models with a sign problem require Grassmann extensions and are not naturally sign-free. TRG fixed points near criticality need TNR/loop-TNR to remove short-range contamination.

Cost & scaling

  • Time: TRG O(χ⁶) per step (2D); HOTRG O(χ^{4d−1}) = O(χ⁷) in 2D, O(χ¹¹) in 3D
  • Memory: TRG O(χ⁴); HOTRG O(χ^{2d}) = O(χ⁴) in 2D, O(χ⁶) in 3D
  • Control knob: bond dimension χ — controls truncation error at each coarse-graining step
  • Scale frontier: thermodynamic limit directly; 2D classical models at χ ~ 20–60; 3D at χ ~ 10–20

Recommended for

  • Primary method for 2D classical statistical models (Ising, Potts, XY, …): free energy, T_c, and CFT data at the thermodynamic limit without finite-size extrapolation
  • 3D classical models (3D Ising, …): HOTRG is feasible at moderate χ; provides thermodynamic-limit estimates
  • Quantum models in 1+1D (path integral → 2D tensor network): maps onto TRG/HOTRG after Suzuki–Trotter decomposition
  • PEPS contraction as an alternative to CTMRG: useful when the geometry favors a coarse-graining sweep
  • Per method-property-map.md: optimal when A1=2D/3D classical, B6=critical (CFT data needed), C12=sign-free; disfavored for real-time dynamics, fermionic sign-ful quantum problems, or volume-law phases

Benchmarks

  • 2D classical Ising T_c: TRG/HOTRG recover T_c = 2J/ln(1+√2) ≈ 2.269 J to ≲0.01% at χ ~ 24 [survey §3.5; @xie_2012_coarse].
  • 2D Ising central charge: c = 0.5 extracted from transfer-matrix RG spectrum at χ ~ 24 xie_2012_coarse.
  • 3D Ising critical exponents (HOTRG χ=16): ν ≈ 0.630, η ≈ 0.036, consistent with conformal-bootstrap values xie_2012_coarse.

Key reference: xie_2012_coarse — introduces HOTRG, derives the O(χ^{4d−1}) cost scaling, and benchmarks critical exponents for 2D and 3D Ising models.

Multiscale entanglement renormalizationmeracontrolledspecializedevenbly_2007_algorithms
AxisValueNote
M1 tasks / outputsGround-state energy of critical systems · scaling dimensions of primary operators · central charge c (CFT data) from the RG superoperator spectrum · entanglement entropy S ~ (c/3) ln L at finite χ · correlation functionsCFT data extraction is MERA's primary strength and the quantity no other efficient method provides directly in the thermodynamic limit.
M2 regimeT=0 ground state (primary)Finite-T and real-time variants exist but are uncommon; primary use is critical GS.
M3 accuracy classControlled-bias (finite-χ truncation at each layer), deterministic; variational upper bound on energyBias → 0 as χ → ∞; optimization is variational so energy is an upper bound at fixed χ.
M4 dimension fit (A1)1D (primary); some 2D MERA constructions exist but are much more expensive2D MERA costs up to O(χ¹⁶), making it impractical at χ large enough to be accurate; 1D is the natural home.
M5 statistics & local dim (A3)Spin / hard-core boson / fermion (fermionic MERA with parity structure)Local dimension d enters via the base-layer tensors; standard implementations fix d=2 (spin-½ or qubit).
M6 entanglement regime (B5)Criticality / scale invariance — its reason to exist: reproduces the log entanglement S ~ (c/3) ln L of 1D CFTs at finite χ via disentanglersArea-law states are better served by MPS at lower cost; volume-law states are inaccessible. MERA is the unique method that handles the area+log critical regime efficiently in the thermodynamic limit.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; no sign problem regardless of frustration or doping.
M8 symmetry exploitation (C9/C10)Global symmetry (U(1), SU(2), Z₂) enforced via symmetric disentanglers and isometriesSymmetric MERA implementations reduce cost and improve convergence; translation symmetry exploited by the coarse-graining structure.
M9 time complexityO(χ⁹) binary 1D; O(χ⁸) ternary 1D; O(χ⁷) modified-binary 1D; 2D up to ~O(χ¹⁶)Cost dominated by the contraction of disentangler environments; binary/ternary costs from evenbly_2007_algorithms; modified-binary O(χ⁷) from evenbly_2011_class (Evenbly–Vidal 2011), which reduces cost while retaining scale invariance.
M10 memoryO(χ⁴) per layer (storing disentangler and isometry tensors)Number of layers grows as log₂(N) for a finite system; thermodynamic limit uses a fixed periodic (scale-invariant) layer structure.
M11 control knobBond dimension χ (number of states kept after each isometry truncation)Increasing χ captures more entanglement and improves both energy accuracy and CFT-data precision; convergence checked by energy and scaling-dimension stability vs χ.
M12 scale frontierThermodynamic limit directly (via scale-invariant fixed-point layer); finite systems at N ~ 10²–10³ routine in 1DThe scale-invariant (infinite) MERA defines a fixed-point RG map valid at all length scales, giving direct thermodynamic-limit CFT data.
M13 primary approximation / biasFinite-χ truncation (controlled): the isometries restrict the number of states retained at each scale; disentanglers are constrained to be unitaryBias is systematic and controlled; energy and CFT data converge monotonically with χ.
M14 hard blocker / failure mode2D: cost up to O(χ¹⁶) makes large-χ calculations impractical; gapped (area-law) systems: MERA is overkill — MPS is cheaper; volume-law phases and real-time dynamics are inaccessibleMERA is purpose-built for criticality; applying it to gapped systems wastes cost relative to MPS.

Cost & scaling

  • Time: O(χ⁹) binary 1D; O(χ⁸) ternary 1D; O(χ⁷) modified-binary 1D; 2D MERA up to ~O(χ¹⁶)
  • Memory: O(χ⁴) per layer; log₂(N) layers for finite system; O(χ⁴) total for scale-invariant (infinite) MERA
  • Control knob: bond dimension χ — controls truncation error and CFT-data accuracy
  • Scale frontier: thermodynamic limit directly via scale-invariant fixed point; finite 1D chains up to N ~ 10³

Recommended for

  • Primary method for extracting CFT data from 1D critical Hamiltonians (transverse-field Ising at J_c, XXX Heisenberg chain, quantum clock models): scaling dimensions and c with no finite-size extrapolation needed
  • 1D critical ground states where log entanglement is the bottleneck: MERA achieves finite-χ accuracy where MPS would require χ ~ L^{c/6}
  • Benchmarking RG fixed points: the scale-invariant MERA layer is a tensor-network realization of the CFT RG fixed point
  • Per method-property-map.md: optimal when A1=1D, B6=critical/scale-invariant; disfavored for A1=2D (cost), B5=area-law gapped (use MPS), B5=volume-law, D13=real-time

Benchmarks

  • 1D transverse-field Ising (J=J_c): central charge c = 0.5 recovered from RG-spectrum; scaling dimension Δ_σ = 0.125 (spin), Δ_ε = 1.0 (energy), both matching Ising CFT exactly at χ ~ 8–16 [survey §3.4; @evenbly_2007_algorithms].
  • 1D XXX Heisenberg chain: c = 1 (free boson CFT), Δ_n consistent with SU(2)₁ WZW model at χ ~ 12 [survey §3.4].
  • Entanglement entropy S ~ (c/3) ln L = (1/6) ln L (c=1/2 Ising CFT, periodic): reproduced at finite χ, confirming log scaling without finite-size extrapolation evenbly_2007_algorithms.

Key reference: evenbly_2007_algorithms — derives the MERA optimization (alternating optimization of disentanglers and isometries), establishes cost scaling for binary and ternary 1D MERA, and demonstrates CFT data extraction.

4 Tensor-network dynamics & finite-T 5

Time-evolving block decimationtebdcontrolledmethod-mpsvidal_2003_efficient
AxisValueNote
M1 tasks / outputsQuench dynamics · real-time correlations · spectral functions S(q,ω) (time-evolve + Fourier) · transport coefficients · ground-state energy & wavefunction (imaginary-time) · finite-T thermodynamics (purification)Spectral functions require Fourier transform of the time-domain correlator; max frequency resolution set by reachable t*.
M2 regimeReal-time dynamics + imaginary-time (ground state & finite-T)Both regimes use the same SVD-truncation engine; imaginary-time entanglement saturates (GS) or grows slowly (thermal).
M3 accuracy classControlled, deterministicTwo controllable knobs: bond dim χ (truncation error) and Trotter step Δτ (Trotter error O(Δτ^p)); both can be converged independently.
M4 dimension fit (A1)1D native; quasi-1D (ladders) with swap gates; walls in 2DNearest-neighbor (A4 = short-range interaction range) gates only — long-range interactions require swap-gate sequences (expensive); 2D walls due to χ~e^{cW} cylinder area law.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion (Jordan–Wigner) / anyPer-site cost ∝d·χ³; large d (Hubbard d=4, soft-core bosons) raises cost; fermions need parity tensors or JW strings.
M6 entanglement regime (B5)Area-law + area+log tolerated; volume-law blocks real-timeReal-time entanglement grows linearly S(t)∝t for thermalizing systems → χ(t)~e^{S(t)} caps reachable time t*~log(χ_max)/rate; MBL (D15) and integrable (C11) systems grow more slowly.
M7 sign-problem dependence (C12)Sign-immuneTensor-network method; no Monte Carlo sampling; no sign problem.
M8 symmetry exploitation (C9/C10)U(1)/SU(2)/Z₂ quantum-number-conserving tensors; translation symmetry for iMPS (infinite systems)Symmetry reduces effective χ by ~10× at fixed accuracy; iMPS with translation gives direct thermodynamic-limit access.
M9 time complexityO(N·d·χ(t)³) per Trotter step; Trotter error O(Δτ^p)p=2 standard; p=4 higher-order Suzuki–Trotter. Total cost scales with number of steps and growth of χ(t). No MPO w factor because TEBD applies nearest-neighbor gates directly (contrast TDVP's O(N·w·d·χ³)).
M10 memoryO(N·d·χ²) for the MPS; environment/orthogonality center storage O(d·χ²) per bondManageable for χ~1000; the dominant memory is the MPS itself.
M11 control knobBond dim χ (SVD truncation error) + Trotter step Δτ (Trotter error O(Δτ^p))Both must be converged; truncation error accumulates over time steps.
M12 scale frontier1D chains N~1000 sites routine (fixed χ); reachable time t*~log(χ_max)/rateAt χ=512 and rate~1 (Heisenberg chain), t*J~6; longer for MBL/integrable (D15/C11).
M13 primary approximation / biasFinite-χ SVD truncation (controlled) + Trotter discretization O(Δτ^p) (controlled)Both biases are systematically improvable; truncation accumulates in time.
M14 hard blocker / failure modeEntanglement barrier: χ(t)~e^{S(t)} — for thermalizing (B6 gapless, B8 frustrated) systems S(t)∝t caps reachable t*; long-range interactions (A4) require expensive swap-gate sequences; 2D walls (χ~e^{cW})Rate set by C11 integrability / D15 MBL / B6 gap; thermal systems hit the barrier fastest.

Cost & scaling

  • Time: O(N·d·χ(t)³) per Trotter step; total cost grows with reachable time as χ(t) increases; no MPO w factor because TEBD applies nearest-neighbor gates directly (contrast TDVP's O(N·w·d·χ³))
  • Memory: O(N·d·χ²) for the MPS
  • Control knob: χ (truncation) + Δτ (Trotter error O(Δτ^p))
  • Scale frontier: N~1000 sites in 1D; reachable real time t*~log(χ_max)/rate

Recommended for

  • Primary real-time dynamics solver for 1D: Heisenberg chains, Hubbard chains, XXZ chains, Bose–Hubbard chains (A1 = 1D, A4 short-range)
  • Imaginary-time GS for short-range 1D models when DMRG is unavailable
  • MBL systems (D15): slow entanglement growth extends reachable t*; TEBD is the standard tool for probing l-bit dynamics
  • Integrable systems (C11): slower-than-linear S(t) growth (logarithmic for MBL, subdiffusive for integrable) allows long-time evolution
  • Per method-property-map.md (TEBD profile): preferred over TDVP when H is strictly nearest-neighbor (A4 = short-range) and swap-gate sequences are not needed

Benchmarks

  • Entanglement growth after a global quench in the XX chain: entanglement grows linearly, S(t) ≃ (π c /3) v_F t (c=1 for the XX chain), until finite-χ truncation caps it — this linear growth is the entanglement barrier vidal_2003_efficient, calabrese_2005_evolution. Saturation at S~log(χ) sets the reachable time t*∝log(χ)/rate.
  • Heisenberg chain quench at χ=512: reachable tJ≈6 before truncation error exceeds 10⁻³ (method-survey.md §5.1).
  • MBL phase (D15, W/J=5): S(t)~log(t), χ grows logarithmically → reachable tJ~100 at χ=512 (method-survey.md §5.1).

Key reference: vidal_2003_efficient — original TEBD paper introducing the SVD-based Trotter time evolution for MPS; defines the algorithm and complexity.

Time-dependent variational principletdvpcontrolledmethod-mpshaegeman_2016_unifying
AxisValueNote
M1 tasks / outputsReal-time dynamics · spectral functions S(q,ω) (time-evolve + Fourier) · long-time dynamics within fixed-χ manifold (1-site) · transport · finite-T (TDVP + purification)1-site TDVP ideal for long-time within manifold; 2-site for growing χ during quench.
M2 regimeReal-time dynamics + imaginary-time (ground state) (D13)Primarily a dynamics method; imaginary-time TDVP is algebraically equivalent to a DMRG sweep.
M3 accuracy classControlled, deterministicTwo knobs: χ (truncation / projection error) and time step Δt (Krylov error, which is very small for the local integrator); 1-site has a projection error absent in 2-site.
M4 dimension fit (A1)1D + quasi-1D (ladders/cylinders); handles long-range Hamiltonians via MPOKey advantage over TEBD: MPO representation supports long-range (A4) interactions natively; no swap-gate overhead; 2D walls same as TEBD (χ~e^{cW}).
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion (parity tensors) / anyPer-site cost ∝w·d·χ³ where w is the MPO bond; large d or wide-range MPO (large w) raises cost.
M6 entanglement regime (B5)Area-law + area+log (1-site within fixed χ); volume-law blocksSame entanglement barrier χ(t)~e^{S(t)} as TEBD in 2-site mode; 1-site TDVP fixes χ (stays in manifold, no growth) but introduces projection error.
M7 sign-problem dependence (C12)Sign-immuneTensor-network method; no Monte Carlo.
M8 symmetry exploitation (C9/C10)U(1)/SU(2)/Z₂ quantum-number-conserving tensors; translation symmetry for uniform MPS (TDVP on the infinite MPS)Symmetry-adapted MPO tensors reduce w and effective χ; ~10× speedup at fixed accuracy.
M9 time complexityO(N·w·d·χ³) per time step (w = MPO bond dimension, set by A4 interaction range)w grows with interaction range: nearest-neighbor w=35; power-law decaying w~1550; arbitrary long-range w can be large. Same χ(t) entanglement barrier as TEBD.
M10 memoryO(N·w·d·χ²) for MPS + MPO environmentsMPO environments O(w·χ²) per bond; dominant at large w.
M11 control knobBond dim χ (projection / truncation error) + time step Δt (Krylov integrator error, exponentially small)1-site: χ fixed → monitor projection error; 2-site: χ grows until truncation error converges.
M12 scale frontier1D N~500–2000 sites routine at fixed χ; reachable time t*~log(χ_max)/rate (same barrier as TEBD)1-site can reach longer real times at fixed χ than TEBD by staying in the manifold; barrier still applies at large t.
M13 primary approximation / biasFinite-χ MPS manifold projection (controlled)1-site: projection onto current MPS tangent space (no χ growth — misses orthogonal dynamics); 2-site: SVD truncation. Energy is conserved (1-site) even as χ is fixed.
M14 hard blocker / failure modeEntanglement barrier χ(t)~e^{S(t)} (same as TEBD for 2-site); 1-site projection error misses dynamics outside current manifold; 2D cylinder wall (χ~e^{cW})For thermalizing systems at long times, both 1-site and 2-site TDVP eventually fail.

Cost & scaling

  • Time: O(N·w·d·χ³) per time step; w = MPO bond dim (set by interaction range A4)
  • Memory: O(N·w·d·χ²) for MPS + MPO environments
  • Control knob: χ (projection/truncation error) + Δt (Krylov error, negligible)
  • Scale frontier: N~500–2000 sites in 1D; reachable time t*~log(χ_max)/rate

Recommended for

  • Long-range interacting 1D models (A4 long-range): power-law spin chains, trapped-ion chains, Rydberg arrays — TEBD is infeasible here; TDVP handles via MPO natively
  • Energy-conserving dynamics: microcanonical quenches, ETH studies where energy conservation matters
  • 1-site long-time dynamics at fixed χ: when you want the best approximation within a fixed MPS space (e.g., non-integrable systems at moderate t)
  • Ladder / cylinder geometries (quasi-1D): same advantage over TEBD in handling the MPO for rung interactions
  • Per method-property-map.md (TDVP profile): preferred over TEBD when H has long-range or multi-site terms (A4 non-nearest-neighbor)

Benchmarks

  • 1-site TDVP energy conservation: |ΔE/E| < 10⁻¹² per step for Heisenberg chain N=100, χ=64 haegeman_2016_unifying.
  • Spectral function convergence: comparable accuracy to TEBD but reachable t* longer at fixed χ (1-site, non-thermalizing) — method-survey.md §5.2.
  • MPO bond dimension for Coulomb-like 1/r interaction: w~20–30 giving cost O(N·30·d·χ³) vs TEBD swap-gate O(N²) per step — method-survey.md §5.2.

Key reference: haegeman_2016_unifying — defines the 1-site and 2-site TDVP integrators for MPS, proves energy conservation, and provides the W^I/W^II MPO time evolution; the standard algorithmic reference.

MPS finite-T (purification / METTS)mps-finite-tcontrolledmethod-mpspaeckel_2019_timeevolution
AxisValueNote
M1 tasks / outputsFree energy · internal energy · specific heat C_v · magnetic susceptibility χ · finite-T correlations · finite-T spectral functions (combine with real-time TDVP/TEBD)Thermodynamic quantities from ⟨H⟩_β, ⟨H²⟩_β − ⟨H⟩²_β; finite-T spectral functions require additional real-time step after reaching β.
M2 regimeFinite-T thermodynamics (D13)Core purpose: equilibrium thermal properties at target inverse temperature β; not a real-time method (though finite-T spectral functions combine both).
M3 accuracy classControlled deterministic (purification) + controlled stochastic (METTS)Purification: controlled by χ and Trotter Δτ; METTS: additionally stochastic 1/√N_s from sampling; both are systematically improvable.
M4 dimension fit (A1)1D native; quasi-1D (ladders/narrow cylinders) feasible; walls in 2DSame as TEBD/TDVP: entanglement of thermal state grows as T drops, worsening in 2D; purification maps to a (d²)-dim 1D problem.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion / any; local dim d→d² for purificationDoubling d to raises per-site cost; for spin-½ d²=4 (manageable); for Hubbard d=4→16 (expensive); METTS avoids by collapsing to product states.
M6 entanglement regime (B5)Moderate at high T (area-law); grows as T→0 → χ divergesEntanglement of purified thermal state rises as T drops → χ must grow at low T; METTS samples states with lower entanglement → cheaper at intermediate T.
M7 sign-problem dependence (C12)Sign-immuneTensor-network method; no Monte Carlo sign problem (METTS uses classical sampling, no quantum sign).
M8 symmetry exploitation (C9/C10)U(1)/SU(2)/Z₂ quantum-number-conserving tensors for both purification and METTSSymmetry doubles the block structure (physical + ancilla); still provides ~5–10× speedup.
M9 time complexityPurification: O(N·d²·χ³) per imaginary-time step × β/Δτ steps; METTS: O(N·d·χ³) per sample × N_s samplesMETTS per-sample cost uses d (not ) and smaller χ (typical states); total cost N_s · O(N·d·χ³); tradeoff vs purification depends on d and required accuracy.
M10 memoryPurification: O(N·d²·χ²); METTS: O(N·d·χ²) per samplePurification memory dominated by doubled local dim; METTS memory is standard MPS size.
M11 control knobχ (truncation error) + Trotter step Δτ (O(Δτ^p) error) + N_s samples (METTS statistical error 1/√N_s)Three independent knobs for METTS; two for purification; all must be converged.
M12 scale frontierPurification: N~200–500 sites at moderate T (βJ~10); METTS: comparable or larger N at intermediate TLow-T limit β→∞ requires large χ for both methods; METTS is competitive at intermediate T where typical states have lower entanglement.
M13 primary approximation / biasPurification: finite-χ SVD truncation (controlled) + Trotter O(Δτ^p) (controlled); METTS: additionally stochastic sampling bias 1/√N_sPurification is a deterministic approximation; METTS adds a stochastic component — must check ergodicity of the collapse chain.
M14 hard blocker / failure modeLow-T entanglement divergence: as β→∞, thermal-state entanglement → GS entanglement → χ must grow; 2D walls (χ~e^{cW}); large local dim d makes purification expensive ( scaling); METTS ergodicity failure for some collapse bases at very low TThe low-T limit is where both methods become most expensive; consider DMRG+GS extrapolation for T→0.

Cost & scaling

  • Time: Purification O(N·d²·χ³) per imaginary-time step; METTS O(N_s · N·d·χ³) per sample
  • Memory: Purification O(N·d²·χ²); METTS O(N·d·χ²) per sample
  • Control knob: χ + Δτ (+ N_s samples for METTS)
  • Scale frontier: N~200–500 at moderate-T purification; similar for METTS at intermediate T

Recommended for

  • 1D quantum magnets and correlated chains at finite T: Heisenberg chains, Hubbard chains at elevated T (βJ ≲ 20)
  • Frustrated 1D/quasi-1D models (B8): sign-free alternative to QMC in the frustrated regime
  • Intermediate-T regime (METTS preference): when d is small and typical-state entanglement is low; cheaper than purification per accuracy unit
  • Low-T approach curves: purification is the preferred choice when T/J≲0.05 and N_s would need to be very large for METTS
  • Per method-property-map.md (mps-finite-t profile): primary finite-T tool for 1D sign-ful (frustrated/fermionic) models where QMC fails

Benchmarks

  • Purification of Heisenberg chain (N=100, χ=64): C_v(T) converged to < 1% at T/J ≥ 0.1 (Paeckel et al., arXiv:1901.05824, §5).
  • METTS vs purification crossover: METTS cheaper at T/J ≥ 0.5 for spin-½ chains (same reference, §6).
  • Finite-T spectral function S(q,ω,T) for XXZ chain at T/J=0.5: χ=128, N_s=200 METTS samples sufficient for < 5% error paeckel_2019_timeevolution.

Key reference: paeckel_2019_timeevolution — comprehensive review of all MPS time-evolution methods including purification and METTS; provides complexity derivations, convergence benchmarks, and software guidance.

Linearized / exponential thermal tensor RGltrg-xtrgcontrolledmethod-ltrgchen_2018_exponential
AxisValueNote
M1 tasks / outputsFree energy · internal energy · specific heat C_v · magnetic susceptibility · thermal correlations · thermal phase transitionsPrimary use case is finite-T thermodynamics; XTRG excels at low-T approach curves where LTRG would require many linear layers.
M2 regimeFinite-T thermodynamics (D13)Equilibrium thermal properties across all T; XTRG most advantageous at low T where O(log β) beats LTRG's O(β/τ) layers.
M3 accuracy classControlled, deterministicTwo controllable sources of error: Trotter discretization τ (O(τ²) per layer) and RG truncation Dc; XTRG's squaring accumulates errors multiplicatively but requires fewer steps.
M4 dimension fit (A1)1D, quasi-1D AND 2Dsign-free for frustrated spin/boson models even in 2DMaps d-dim quantum at finite T → (d+1)-dim classical TN; the classical TN is sign-free for frustrated spin and boson models (B8, C12 sign-ful); fermionic models need care (Jordan–Wigner strings can introduce complex/negative entries). The key advantage over QMC is for frustrated 2D magnets.
M5 statistics & local dim (A3)Spin / hard-core boson / any local dim dLocal dim d enters the Trotter-gate tensors; per-step cost polynomial in d; no fundamental barrier from statistics (no sign).
M6 entanglement regime (B5)Moderate at high T; grows as T→0 → Dc must increaseThermal entanglement of the layered TN sets the required Dc; at high T, states are product-like (small Dc); at low T, thermal entanglement approaches GS entanglement (large Dc).
M7 sign-problem dependence (C12)Sign-immune for frustrated spin and boson models (B8); fermionic LTRG/XTRG needs careSign-free for frustrated spin and boson models (B8); fermionic LTRG/XTRG needs care — Jordan–Wigner strings can give complex/negative tensor entries in the contraction. The sign-free advantage over QMC applies to frustrated magnets and bosonic models.
M8 symmetry exploitation (C9/C10)Conserved-charge (U(1)) block structure in the MPO/TN tensorsSymmetry reduces the effective tensor dimensions and speeds up contraction.
M9 time complexityLTRG: polynomial in d and Dc per layer × K=β/τ layers, Trotter error O(τ²); XTRG: higher cost per step (squaring a TN) but only O(log β) steps to reach low TXTRG reaches β=100 in ~7 squaring steps from τ=1; LTRG needs 100/τ layers. At low T, XTRG's lower step count dominates the comparison.
M10 memoryO(Dc²) per TN row (MPO or boundary MPS)The dominant memory is the contracted boundary object; 2D requires 2D boundary contraction (heavier than 1D).
M11 control knobTrotter step τ (Trotter error O(τ²)) + retained bond dim Dc (RG truncation error)Both must be converged; XTRG additionally requires checking that squaring errors accumulate controllably.
M12 scale frontier1D: N~1000 sites routine; quasi-1D and 2D: thermodynamic limit via transfer-matrix / iPEPS-row contraction2D: infinite systems via row-to-row contraction with boundary-MPS or CTMRG; not limited to small clusters.
M13 primary approximation / biasTrotter O(τ²) discretization (controlled) + finite-Dc RG truncation (controlled)Both biases are convergence parameters; XTRG has multiplicative error accumulation from squaring steps but requires far fewer of them.
M14 hard blocker / failure modeLow-T thermal entanglement growth → large Dc needed; 3D classical TN contraction (extremely expensive); models with complex local interactions that resist Trotter decomposition2D quantum → 3D classical TN is the heaviest case; exponentially hard for 3D quantum models.

Cost & scaling

  • Time: LTRG: O(β/τ) layers × polynomial in d·Dc per layer; XTRG: O(log β) squaring steps × higher per-step cost
  • Memory: O(Dc²) per MPO / boundary row
  • Control knob: Trotter τ (error O(τ²)) + retained Dc (RG truncation error)
  • Scale frontier: 1D N~1000; 2D thermodynamic limit via infinite-system row contraction

Recommended for

  • Frustrated 2D quantum magnets (B8, C12 sign-ful): sign-free even in 2D — the primary advantage; QMC fails here; LTRG/XTRG succeeds
  • Low-T thermodynamics (XTRG): XTRG's O(log β) scaling is far superior to LTRG's linear K at low T; use XTRG when βJ ≳ 10
  • 1D and quasi-1D models at all T: competitive with or better than mps-finite-t for 1D at intermediate T; natural thermodynamic-limit access
  • Any model where QMC has a sign problem (C12 sign-ful): LTRG/XTRG provide the sign-free finite-T route
  • Per method-property-map.md (LTRG/XTRG profile): preferred finite-T method when QMC sign problem is present and 2D geometry is needed

Benchmarks

  • XTRG for spin-½ Heisenberg chain: specific heat converged to < 0.1% at T/J=0.05 with Dc=64, τ=0.1 in 7 squaring steps chen_2018_exponential.
  • XTRG for frustrated J₁-J₂ square lattice (sign-ful QMC): thermal susceptibility peak resolved at T/J≈0.4, Dc=64 (method-survey.md §5.5).
  • LTRG vs XTRG accuracy at βJ=20: XTRG 2–3 orders of magnitude more accurate at same Dc and total compute time (method-survey.md §5.5).

Key reference: chen_2018_exponential — introduces XTRG (exponential doubling ρ→ρ·ρ), demonstrates O(log β) scaling, and benchmarks frustrated 2D magnets; the primary XTRG reference.

DMRG-X (excited / MBL eigenstates)dmrg-xcontrolledmethod-mpskhemani_2016_obtaining
AxisValueNote
M1 tasks / outputsIndividual highly-excited eigenstates · eigenstate energies and gaps · local observables in excited eigenstates · l-bit structure and l-bit correlators · eigenstate entanglement entropy and spectrum across the MBL phaseThe ability to obtain individual eigenstates (not just averages) at arbitrary energy density is unique to DMRG-X among polynomial-cost methods.
M2 regimeHighly-excited eigenstates at finite energy density (an MBL-specific regime gated by D15, not one of D13's ground-state/finite-T/dynamics values)Not a finite-T method; targets a single eigenstate at a chosen energy density, which is a T=∞ analog only in the MBL phase.
M3 accuracy classControlled, deterministicMPS truncation at χ is the only approximation; energy and observables converge as χ→∞; residual ‖H|ψ⟩−E|ψ⟩‖ provides a convergence check.
M4 dimension fit (A1)1D; quasi-1D (ladders) feasible; 2D only if MBL survives (debated)MBL in 2D is theoretically controversial and numerically difficult; DMRG-X is primarily applied to 1D disordered chains (A1=1D, D15=MBL).
M5 statistics & local dim (A3)Spin / hard-core boson / fermion / anySame as standard DMRG; per-site cost ∝d·χ³ (targeted variant) or ∝d·χ⁶ (full diag variant); fermions need Jordan–Wigner or parity tensors.
M6 entanglement regime (B5)Area-law only — works ONLY because MBL (D15) keeps excited-state entanglement area-lawThis is the decisive gate: thermal (ETH) excited states have volume-law entanglement → χ~e^N → DMRG-X fails; MBL area law → χ bounded → DMRG-X works.
M7 sign-problem dependence (C12)Sign-immuneTensor-network method; no sign problem.
M8 symmetry exploitation (C9/C10)U(1)/SU(2)/Z₂ block structure (though disorder breaks translation symmetry)Symmetry still reduces effective χ and improves convergence; translation is absent (disorder breaks it), but spin-rotation or parity may be conserved.
M9 time complexityO(L·χ⁶) for the full local-diagonalization variant; ~O(L·χ³) for the k-eigenstate-near-target variantThe χ⁶ scaling (vs standard DMRG χ³) comes from the environment density-matrix construction for interior states; the cheaper O(Lχ³) variant uses a targeted projection.
M10 memoryO(L·χ²) for the MPS; environment blocks O(χ²) per bondSame memory as standard DMRG; does not require storing multiple eigenstates simultaneously.
M11 control knobBond dim χ (truncation error)The single convergence parameter; monitor residual ‖H|ψ⟩−E|ψ⟩‖ < ε; convergence is typically fast when MBL area law holds.
M12 scale frontier1D MBL chains L~100–200 sites routine; L~500–1000 with optimized codesFar exceeds exact-diagonalization reach for excited states (ED limited to L~20–24); the practical limit set by χ required for convergence at the target disorder strength.
M13 primary approximation / biasFinite-χ SVD truncation (controlled)Only approximation; controlled by increasing χ; the method is exact in χ→∞ limit; must verify that the tracked state does not jump to a different eigenstate.
M14 hard blocker / failure modeWorks ONLY because MBL (D15) keeps χ bounded: thermal (ETH, B6 gapless) excited states have volume-law entanglement → DMRG-X completely fails; 2D MBL debated (may not exist thermodynamically); tracking instability (jumping to wrong eigenstate) at weak disorderThe MBL area law is the sole reason DMRG-X is feasible; any regime without it (ETH phase, finite-T thermal states) kills the method.

Cost & scaling

  • Time: O(L·χ⁶) full local-diagonalization variant; ~O(L·χ³) k-eigenstate-near-target variant
  • Memory: O(L·χ²) for MPS + environment blocks
  • Control knob: χ (truncation error; also monitor ‖H|ψ⟩−E|ψ⟩‖)
  • Scale frontier: L~100–200 routine; L~500–1000 optimized codes

Recommended for

  • MBL systems (D15) in 1D: random-field Heisenberg chain, random XXZ, random Hubbard chain — the primary use case and the only regime where DMRG-X is valid
  • Probing the MBL–ETH transition: compute S(n) as disorder W is varied; area-law → volume-law crossover marks the transition (DMRG-X fails as ETH sets in → marks the transition)
  • L-bit extraction: individual eigenstates needed to construct l-bit Hamiltonians and verify MBL phenomenology
  • Beyond ED reach for excited states: L~100–500 replaces expensive shift-invert ED (L≤24) for individual excited-state studies in the MBL phase
  • Per method-property-map.md (DMRG-X profile): applicable when D15 = MBL (area-law excited states); inapplicable when D15 absent (ETH phase)

Benchmarks

  • Random-field Heisenberg chain (L=100, W/J=5, χ=64): ‖H|ψ⟩−E|ψ⟩‖ < 10⁻⁸; entanglement entropy S~0.3 (area law confirmed) khemani_2016_obtaining.
  • MBL–ETH comparison: DMRG-X converges at χ=64 for W/J=5; fails to converge (requires χ→∞) for W/J=1 (ETH phase) — method-survey.md §5.6.
  • O(L·χ⁶) vs O(L·χ³) crossover: targeted variant (O(Lχ³)) reaches same accuracy as full-diag variant at χ≤64 for strong MBL (method-survey.md §5.6).

Key reference: khemani_2016_obtaining — introduces DMRG-X, proves applicability based on MBL area law, and benchmarks the O(L·χ⁶) local-diagonalization variant on random-field Heisenberg chains.

5 Quantum Monte Carlo 7

Variational MC & neural quantum statesvmc-nqsuppermethod-vmccarleo_2016_solving
AxisValueNote
M1 tasks / outputsGS energy (variational upper bound) · variational observables (correlations, structure factors, order parameters) · excited states via symmetry sectors or energy-penalty methods · real-time dynamics via time-dependent VMC (t-VMC, extra cost, growing error)Gives an upper bound only — pair with PolyOpt/SDP for a rigorous energy bracket.
M2 regimeT=0 (ground state); real-time via t-VMCFinite-T via VMC sampling of the thermal density matrix is possible but non-standard.
M3 accuracy classVariational upper bound, stochasticStatistical error bars ∝1/√N_s; bias set by ansatz expressiveness (uncontrolled but variational — cannot go below the true GS).
M4 dimension fit (A1)Any dimension (1D, 2D, 3D)No geometric constraint — the key advantage over tensor-network methods in 2D/3D.
M5 statistics & local dim (A3)Spin / hard-core boson / soft-core boson / fermion (with sign structure in ansatz)Fermions require encoding the sign/phase structure in the ansatz; large local dim d raises c_eval but does not cause an exponential wall.
M6 entanglement regime (B5)Volume-law tolerated (no bond-dimension wall)NQS can in principle represent volume-law states; quality depends on the ansatz, not a hard truncation.
M7 sign-problem dependence (C12)Sign-immuneNo sign problem — the sign/phase is encoded in the (complex) ansatz amplitude. Thrives where QMC fails: B8 frustration, C12 sign-ful, 2D.
M8 symmetry exploitation (C9/C10)U(1), SU(2), translation, point-group can be imposed as constraints on the ansatz or via symmetry-projected samplingSymmetry projection reduces variance and targets correct quantum-number sectors.
M9 time complexityO(N_s·c_eval) per optimization step; SR/natural-gradient adds O(N_p²)O(N_p³) (exact) or O(N_s·N_p) (iterative/minSR)c_eval = amplitude + local energy cost (O(N)O(N²) Jastrow/RBM; network forward pass for deep NQS).
M10 memoryO(N_p) for parameters + O(N_s·N_p) for the SR matrix (or O(N_s) minSR)Deep-network ansätze can have N_p ~ 10⁴10⁶; SR matrix can dominate memory.
M11 control knobN_s (samples per step) → statistical error ∝1/√N_s; ansatz size / depth → ansatz biasNo single knob fully controls the variational bias — convergence studies vary ansatz architecture.
M12 scale frontierThousands of sites feasible with shallow RBM/Jastrow; hundreds of sites with deep NQS; thermodynamic limit via ansatz in the TD limitScales far beyond ED/FCI; quality limited by ansatz not system size.
M13 primary approximation / biasAnsatz bias — variational error is set by the expressiveness of the chosen ansatz (uncontrolled but variational)Cannot go below the true GS; paired with a lower bound (SDP/PolyOpt) it brackets the energy.
M14 hard blocker / failure modeOptimization landscape (local minima, barren plateaus for deep networks); autocorrelation in MC sampling; SR matrix ill-conditioningReal-time t-VMC accumulates gauge error; strongly correlated regimes with poor ansatz overlap fail quietly (low-quality local minimum).

Cost & scaling

  • Time: O(N_s·c_eval) per optimization step; SR adds O(N_p²)O(N_p³) or O(N_s·N_p) with minSR
  • Memory: O(N_p) parameters; O(N_s·N_p) SR matrix (dominant for large N_p)
  • Control knob: N_s (sample size → statistical error ∝1/√N_s); ansatz depth/width → variational bias
  • Scale frontier: thousands of sites (shallow); hundreds (deep NQS); thermodynamic limit possible

Recommended for

  • Frustrated magnets (B8) and doped Mott insulators (D14): where SSE/DQMC/worldline QMC fail due to the sign problem; NQS provides a sign-immune upper bound
  • 2D/3D systems: where MPS bond-dimension wall is prohibitive but exact methods are out of reach
  • Benchmarking against AFQMC: VMC-NQS upper bound + AFQMC constrained energy validates phaseless approximation quality
  • Upper-half of a rigorous energy bracket: pair with /method-polyopt (SDP lower bound) to bracket the GS energy rigorously
  • Per method-property-map.md: recommended when C12 sign-ful AND A1 2D/3D AND need variational upper bound

Benchmarks

  • RBM VMC on frustrated J₁-J₂ Heisenberg chain (N=80, J₂/J₁=0.5): energy within 0.1% of DMRG carleo_2016_solving.
  • SR optimizer convergence: stochastic reconfiguration outperforms gradient descent by orders of magnitude in optimization efficiency sorella_1998_green.

Key reference: carleo_2016_solving — founding NQS paper demonstrating RBM wavefunction solves frustrated 1D/2D spin models with high accuracy; the standard entry-point reference.

Stochastic series expansion / worldlinesseexactmethod-qmcsyljuasen_2002_quantum
AxisValueNote
M1 tasks / outputsFinite-T thermodynamics (energy, specific heat, susceptibility) · equal- and imaginary-time correlations · structure factors · spin stiffness / superfluid density · Binder cumulants · winding-number distributionsSpectra S(q,ω) require analytic continuation (ill-posed); real-time dynamics blocked by the dynamical sign problem.
M2 regimeFinite-T (primary); GS accessible via large β limitThe large-β limit extracts GS properties but cost grows with β.
M3 accuracy classNumerically exact (in the sign-free regime), stochasticStatistical error ∝1/√(N_s·⟨s⟩) where ⟨s⟩=1 in sign-free cases; in sign-ful cases accuracy degrades as ⟨s⟩→0.
M4 dimension fit (A1)Any dimension (1D, 2D, 3D)The decisive advantage over tensor-network methods in 2D/3D; no bond-dimension wall.
M5 statistics & local dim (A3)Spins / hard-core bosons / soft-core bosons (sign-free); fermions / frustrated spins introduce the signLocal dimension d enters via the operator basis; large d increases the diagonal/off-diagonal operator pool but does not cause exponential cost.
M6 entanglement regime (B5)Volume-law toleratedNo entanglement restriction — worldlines can represent arbitrary correlations in the sign-free sector.
M7 sign-problem dependence (C12)Sign-free for unfrustrated bipartite spins (Marshall sign rule) and bosons; B8 frustration / A3 fermions turn the sign on; sign-ful cost ∝1/⟨s⟩²~e^{2βNΔf}The primary gate: check bipartiteness and absence of frustration before choosing SSE.
M8 symmetry exploitation (C9/C10)U(1) (Sᶻ, particle number) conservation reduces operator pool; translation / point-group target momentum sectors for correlationsSymmetry reduces variance but is not required for correctness.
M9 time complexityO(N·β) per sweep (expansion order ⟨n⟩∝N·β); statistical error 1/√(N_s·⟨s⟩); directed-loop/operator-loop updates remove critical slowing down; sign-ful cost ∝1/⟨s⟩²~e^{2βNΔf}Cost is linear in N and β in the sign-free regime — highly favorable for large systems at moderate temperatures.
M10 memoryO(N·β) to store the operator stringOperator string length ⟨n⟩∝N·β; configuration vector O(N); total modest compared to wave-function methods.
M11 control knobN_s (Monte Carlo sweeps) → statistical error ∝1/√N_s; β (inverse temperature) → GS vs finite-T physics; cutoff n_max ≥ ⟨n⟩Convergence in N_s straightforward; β tuned to target temperature or GS extrapolation.
M12 scale frontierThousands to tens-of-thousands of sites in the sign-free regimeThermodynamic-limit extrapolations routine; limited only by autocorrelation and the sign problem.
M13 primary approximation / biasNone in the sign-free regime (numerically exact); sign-ful approximation has uncontrolled bias unless ⟨s⟩≈1In the sign-ful regime the average-sign collapse is the approximation; no free-parameter knob removes it.
M14 hard blocker / failure modeSign problem for frustrated (B8) spins and fermions (A3): cost ∝e^{2βNΔf} → exponential wall; real-time dynamics (dynamical sign problem)Even weak frustration can turn the sign on; analytic continuation from imaginary-time data is ill-conditioned.

Cost & scaling

  • Time: O(N·β) per sweep (sign-free); cost ∝1/⟨s⟩²~e^{2βNΔf} in sign-ful cases
  • Memory: O(N·β) operator string
  • Control knob: N_s (sweeps → statistical error ∝1/√N_s); β controls temperature/GS convergence
  • Scale frontier: thousands–tens-of-thousands of sites (sign-free bipartite spin / boson models)

Recommended for

  • Unfrustrated bipartite spin models (B8=no): Heisenberg, XXZ, transverse-field Ising on square/honeycomb/cubic lattices — the method of choice for finite-T thermodynamics
  • Bosonic systems (A3 bosons): Bose-Hubbard model, lattice bosons — sign-free, efficient worm/SSE updates
  • Any dimension (A1): the key advantage over tensor-network methods for 2D/3D at finite temperature
  • Critical phenomena (B6): directed-loop updates eliminate critical slowing down near phase transitions
  • Per method-property-map.md: first choice for sign-free finite-T spin/boson models; blocked by B8 frustration and A3 fermions

Benchmarks

  • Heisenberg square lattice (N=64×64, T/J=0.5): susceptibility, stiffness converged to sub-percent accuracy with N_s~10⁵ sweeps syljuasen_2002_quantum.
  • Critical slowing down suppression: directed-loop/operator-loop updates greatly reduce critical slowing down — autocorrelation τ_auto~L^z with z reduced to ≈1–1.5 for the spin stiffness (vs z≈2 for local updates) sandvik_2010_computational.

Key reference: syljuasen_2002_quantum — introduces directed-loop updates in SSE; the standard implementation reference for efficient worldline/SSE codes.

Determinant QMC (BSS)dqmcexactmethod-qmcbercx_2017_alf
AxisValueNote
M1 tasks / outputsFinite-T energy · double occupancy · magnetic and charge structure factors S(q) · single-particle Green's function G(k,τ) → spectral function A(k,ω) and self-energy via analytic continuation · superconducting pairing susceptibilitiesReal-time dynamics blocked; A(k,ω) requires MaxEnt or SOAC analytic continuation (ill-posed).
M2 regimeFinite-T (primary); GS limit via large βCost grows as β → ∞; the GS is better attacked by AFQMC-CPMC.
M3 accuracy classNumerically exact (sign-free regime), stochastic; controlled-bias (Trotter O(Δτ²))Statistical error ∝1/√N_s; Trotter error eliminated by Δτ→0 extrapolation; sign collapse in sign-ful regime.
M4 dimension fit (A1)Any dimension (1D, 2D, 3D)Sign-free condition depends on lattice bipartiteness and filling, not dimension.
M5 statistics & local dim (A3)Fermions (primary); mixed fermion-boson via partial traced=4 (Hubbard: empty, ↑, ↓, ↑↓); N_τ=β/Δτ slices; cost O(βN³) grows fast with N.
M6 entanglement regime (B5)Volume-law toleratedAuxiliary-field sampling imposes no entanglement restriction; the sign problem is the gate, not entanglement.
M7 sign-problem dependence (C12)Sign-free at half-filling on bipartite lattices (det↑det↓=|det|²≥0) and attractive Hubbard at any filling; D14 doping / B8 frustration turn the sign on; sign ⟨s⟩~e^{−βNΔf} → cost ∝e^{2βNΔf}Half-filling + bipartite = the canonical sign-free condition; any deviation from half-filling on repulsive Hubbard leads to the sign problem.
M8 symmetry exploitation (C9/C10)Particle-hole symmetry (C9) at half-filling on bipartite lattices is the sign-free condition; translation/point-group reduce momentum-resolved measurements but do not change the core algorithmPH symmetry is the decisive symmetry here — it is a correctness condition, not merely a speedup.
M9 time complexityO(β·N³) per sweep (N×N matrix products over N_τ=β/Δτ slices with numerical stabilization); sign-ful cost ∝e^{2βNΔf}The matrix-algebra cost limits system size to N~200–400; lower T (larger β) is doubly costly: more slices and worse sign.
M10 memoryO(N²·N_τ)Stores the N×N equal-time Green's function matrix plus N_τ auxiliary-field configurations.
M11 control knobΔτ (Trotter step) → Trotter error O(Δτ²); N_s (sweeps) → statistical error ∝1/√N_s; β (inverse temperature)Extrapolate Δτ→0 for Trotter-free results; check convergence in N_s.
M12 scale frontierN~100–400 sites routine; N~1000 possible on HPC with stabilization; thermodynamic limit via finite-size scalingSign-free regime: large-scale studies of Hubbard model on square lattice up to N=20×20.
M13 primary approximation / biasTrotter discretization error O(Δτ²) (controlled, eliminated by Δτ→0 extrapolation); sign collapse in sign-ful regimes (uncontrolled)In sign-free cases the only approximation is controllable Trotter error and statistical noise.
M14 hard blocker / failure modeSign problem away from half-filling (D14 doping) and with frustration (B8): cost ∝e^{2βNΔf}; low-T limit compounds both cost and sign penaltyLarge β and large N simultaneously is the hardest regime; sign problem is exponential in both.

Cost & scaling

  • Time: O(β·N³) per sweep; sign-ful cost ∝e^{2βNΔf}
  • Memory: O(N²·N_τ) with N_τ=β/Δτ
  • Control knob: Δτ (Trotter error O(Δτ²)); N_s (statistical error ∝1/√N_s)
  • Scale frontier: N~100–400 routine (sign-free); N~1000 HPC; finite-size scaling to thermodynamic limit

Recommended for

  • Repulsive Hubbard model at half-filling (D14): the canonical sign-free regime; benchmark-quality finite-T phase diagram
  • Attractive Hubbard at any filling: sign-free, ideal for BCS–BEC crossover studies
  • Finite-T magnetic and charge order: square-lattice Heisenberg-Hubbard at half-filling up to N~400
  • Single-particle spectral functions on fermion models: DQMC + MaxEnt/SOAC → A(k,ω) for comparison with ARPES
  • Per method-property-map.md: preferred over AFQMC-CPMC when finite-T thermodynamics is needed; blocked by D14 doping in sign-ful regimes

Benchmarks

  • Hubbard model half-filling square lattice (N=16×16, U/t=4, β=8): energy, double occupancy, structure factor converged bercx_2017_alf, verified in method-survey.md §4.4.
  • Trotter extrapolation: Δτ=0.050.1 typically sufficient; linear extrapolation in Δτ² removes systematic error to < 0.1% bercx_2017_alf.

Key reference: bercx_2017_alf — the ALF (Algorithms for Lattice Fermions) code documentation; comprehensive reference for the stabilized DQMC algorithm, Green's function updates, and observable implementations.

Auxiliary-field QMC / constrained-pathafqmc-cpmccontrolledmethod-qmczhang_2019_auxiliary
AxisValueNote
M1 tasks / outputsGS energy (benchmark-grade for many Hubbard regimes) · back-propagated observables (density, correlation functions) · imaginary-time correlations → spectral functions via analytic continuationBest-in-class for GS energy of the 2D Hubbard model; back-propagated observables carry extra bias from the constraint.
M2 regimeT=0 (ground state)DQMC is the finite-T sibling; AFQMC-CPMC targets the GS specifically.
M3 accuracy classControlled-bias — phaseless/constrained-path bias set by trial-state quality; stochasticStatistical error ∝1/√(N_w·N_steps); systematic bias shrinks as |Ψ_T⟩ → exact GS; free-projection (unbiased) is exponentially costly.
M4 dimension fit (A1)Any dimension (1D, 2D, 3D)Frequently applied to 2D Hubbard where it gives the best GS energies among polynomial-cost methods.
M5 statistics & local dim (A3)Fermions (primary); also applied to real materials via plane-wave basisEach walker is a Slater determinant; N orbitals (sites × spin); local dim d=4 Hubbard.
M6 entanglement regime (B5)Volume-law toleratedWalker space imposes no entanglement restriction; the constraint bias is the limit, not entanglement.
M7 sign-problem dependence (C12)Phaseless/constrained-path approximation removes the sign at the cost of a trial-state biasFree-projection (unbiased) cost blows up ∝e^{2βNΔf} with the sign; the constraint trades that for polynomial cost + controllable bias.
M8 symmetry exploitation (C9/C10)Trial state |Ψ_T⟩ can be symmetry-projected (UHF, ROHF, GHF) to reduce bias; particle-number and spin conservation preserved in walker propagationBetter |Ψ_T⟩ (e.g. from DMRG or VMC) directly reduces the phaseless bias.
M9 time complexityO(N³) per walker per step × N_w × N_steps; O(N²) per walker memoryN = number of orbitals/sites; from matrix products in Slater determinant updates; N_w~1001000 typical.
M10 memoryO(N²) per walker (Slater determinant stored as N×N_e matrix)Total O(N_w·N²); modest compared to exact-diagonalization storage.
M11 control knobN_w (walkers) → statistical error ∝1/√N_w; Δτ (projection step) → Trotter error; trial state |Ψ_T⟩ quality → phaseless biasImproving |Ψ_T⟩ systematically reduces the bias; can use VMC-NQS or DMRG wavefunction as trial state.
M12 scale frontierN~100500 sites routine; N~1000 on HPC; thermodynamic-limit extrapolation via finite-size scalingState-of-the-art for the 2D Hubbard model ground state at intermediate-to-strong coupling and away from half-filling.
M13 primary approximation / biasPhaseless/constrained-path bias — set by overlap of walkers with |Ψ_T⟩; controlled by improving the trial stateBias is variational in the constrained-path case: the constrained-path energy is an upper bound, E_CP ≥ E_0 (Zhang–Krakauer 2003); the phaseless approximation is not strictly variational.
M14 hard blocker / failure modePoor trial state → large phaseless bias (e.g. strongly frustrated models where even HF is qualitatively wrong); back-propagated observables have extra bias; free-projection is exponentially costlyThe constraint quality is the central limit; no sign-free condition exists for general fermion models.

Cost & scaling

  • Time: O(N³) per walker per step × N_w × N_steps
  • Memory: O(N²) per walker; total O(N_w·N²)
  • Control knob: N_w (statistical error ∝1/√N_w); trial state quality → phaseless bias; Δτ → Trotter error
  • Scale frontier: N~100500 routine; N~1000 HPC; finite-size scaling to thermodynamic limit

Recommended for

  • 2D Hubbard model away from half-filling (D14 doped): best-in-class GS energies where DQMC sign problem is severe; the method of choice for the doped Mott insulator
  • Strongly correlated fermion models where sign problem precludes DQMC: AFQMC-CPMC with DMRG/VMC trial state narrows the phaseless bias substantially
  • Real materials (ab initio AFQMC): plane-wave/localized orbital basis; benchmark quality for molecules and solids
  • Benchmarking VMC-NQS upper bounds: AFQMC-CPMC GS energy cross-checks the VMC variational bound from below
  • Per method-property-map.md: preferred over DQMC for GS at finite doping; complementary to DQMC's finite-T capabilities

Benchmarks

  • 2D Hubbard model (10×10, U/t=4, ⟨n⟩=0.875): AFQMC-CPMC energy E/t = −0.5444(4) — benchmark in the doped regime where DQMC is sign-ful zhang_2019_auxiliary.
  • Phaseless bias test: with UHF trial state vs DMRG trial state, energy difference < 0.1% for U/t≤8 on 4×4 lattice nguyen_2014_cpmc.

Key reference: zhang_2019_auxiliary — comprehensive lecture-note review of the auxiliary-field QMC algorithm at zero and finite temperature; the standard pedagogical and technical reference.

Projector / diffusion MC + fixed-nodedmc-gfmccontrolledmethod-qmcbecca_2017_quantum
AxisValueNote
M1 tasks / outputsGS energy (fixed-node = best variational energy given the trial node) · mixed estimators for other observables · forward-walking estimators for unbiased observables (at extra cost)Mixed estimators for non-commuting observables carry bias O((Ψ_T−Ψ_0)²); forward-walking removes leading bias but adds variance.
M2 regimeT=0 (ground state)Purely a GS projector method; finite-T not naturally accessible.
M3 accuracy classControlled-bias (fixed-node / fixed-phase = best variational energy given the trial nodal surface); stochasticEnergy is an upper bound given the trial node; exact for bosons/unfrustrated systems (sign-free, no bias).
M4 dimension fit (A1)Any dimension (1D, 2D, 3D, continuum)Especially powerful in continuum (electronic structure, quantum liquids); lattice GFMC works in any dimension.
M5 statistics & local dim (A3)Bosons (sign-free, exact) · fermions (fixed-node, biased) · spins (via lattice GFMC)Fermions require a fixed-node constraint; bosons are exact. Continuum electrons are the primary application domain.
M6 entanglement regime (B5)Volume-law toleratedWalker-based sampling imposes no entanglement restriction; quality determined by the trial node, not entanglement.
M7 sign-problem dependence (C12)Sign-free for bosons / unfrustrated systems; fixed-node (nodal-surface bias) controls the fermion signFixed-node = best variational energy given the trial node; nodal bias is the primary systematic error; stochastic reconfiguration can also help.
M8 symmetry exploitation (C9/C10)Trial state |Ψ_T⟩ can carry full point-group / translation / spin symmetry; walker propagation respects theseBetter-symmetry trial node → lower fixed-node bias.
M9 time complexityO(N_w·N_proj·c_step); statistical error 1/√(N_w·N_proj)c_step = cost per walker step (O(N) for simple wave functions; O(N³) for Slater-determinant trial states); N_proj = total projection time / Δτ.
M10 memoryO(N_w·N)Stores N_w walker configurations of N particles; modest.
M11 control knobN_w (walkers) · N_proj (projection time) → statistical error 1/√(N_w·N_proj); trial node quality → fixed-node biasΔτ→0 extrapolation removes time-step bias; trial node improved by VMC optimization.
M12 scale frontierN~1001000 particles (continuum electronic structure); N~100500 sites (lattice GFMC)State-of-the-art for electron gas, quantum liquids (⁴He, ³He), and lattice Hubbard clusters.
M13 primary approximation / biasFixed-node (nodal-surface bias): energy upper bound given the trial node; bias eliminated only if the trial node is exactControlled in the sense that a better trial node monotonically improves the energy; not variational w.r.t. node position for all observables.
M14 hard blocker / failure modeFixed-node bias can be large for strongly correlated / frustrated systems where no good trial node is known; forward-walking estimators for observables have large variance; continuum time-step bias O(Δτ) requires Δτ→0 extrapolationIn lattice frustrated magnets the trial node quality degrades rapidly and the fixed-node energy may be far above the true GS.

Cost & scaling

  • Time: O(N_w·N_proj·c_step); statistical error 1/√(N_w·N_proj)
  • Memory: O(N_w·N) walker configurations
  • Control knob: N_w and N_proj (statistical error); trial node (fixed-node bias); Δτ (time-step bias)
  • Scale frontier: N~1001000 (continuum); N~100500 (lattice); thermodynamic-limit via finite-size scaling

Recommended for

  • Continuum bosons (⁴He, cold atoms, quantum liquids): sign-free, numerically exact — the definitive method for quantum liquid ground states
  • Continuum fermion systems (electron gas, atoms, molecules): fixed-node DMC gives the best-in-class correlation energies for weakly to moderately correlated systems
  • Lattice GFMC for unfrustrated spins / bosons: exact GS energy for Heisenberg models where SSE is also available (cross-check partner)
  • When VMC-NQS provides a trial state: VMC optimizes the node; DMC projects to the fixed-node GS for additional improvement
  • Per method-property-map.md: preferred for continuum systems; blocked by frustration (poor nodal surface)

Benchmarks

  • Electron gas r_s=15: DMC fixed-node energy matches diffusion MC benchmark to < 1 mHa/particle with Slater–Jastrow trial function becca_2017_quantum.
  • ⁴He liquid: exact GFMC GS energy E/N = −7.17(2) K at experimental density — the reference for quantum liquid benchmarks becca_2017_quantum.

Key reference: becca_2017_quantum — textbook treatment of VMC and projector QMC (GFMC/DMC) with stochastic reconfiguration; the standard pedagogical reference covering the fixed-node approximation, forward-walking estimators, and benchmarks.

Path-integral Monte Carlopimcexactmethod-qmcceperley_1995_path
AxisValueNote
M1 tasks / outputsFinite-T thermodynamics (energy, specific heat, pressure, compressibility) · superfluid density (winding number) · condensate fraction (off-diagonal density matrix n(r)) · structure factors S(q) · pair correlations g(r)Spectra require analytic continuation; real-time dynamics inaccessible.
M2 regimeFinite-T (primary); T→0 limit feasible but expensiveThe natural regime is finite temperature; GS by taking β→∞.
M3 accuracy classNumerically exact (bosons, sign-free), stochasticStatistical error ∝1/√N_s; Trotter/discretization error O(Δτ^p) controlled by N_τ→∞; fermion sign problem degrades accuracy.
M4 dimension fit (A1)Any dimension; continuum (⁴He) or latticeParticularly powerful in 3D continuum for quantum liquids and cold atoms.
M5 statistics & local dim (A3)Bosons (sign-free) — the primary regime; fermions reintroduce the sign problem via odd permutationsFor fermions, the sign problem returns and cost is exponential; restricted PIMC (fixed-node analog) controls it with bias.
M6 entanglement regime (B5)Volume-law toleratedWorldline sampling has no entanglement restriction; the sign problem is the gate for fermions.
M7 sign-problem dependence (C12)Sign-free for bosons; fermions reintroduce the sign via odd permutation exchangesFor bosonic systems PIMC is numerically exact; fermion PIMC requires restricted PIMC (nodal-surface bias) analogous to fixed-node DMC.
M8 symmetry exploitation (C9/C10)Continuous translational symmetry handled naturally in the continuum; lattice models use discrete symmetry to reduce sampling spacePermutation symmetry is essential for bosonic exchange statistics; the worm algorithm samples exchange cycles directly.
M9 time complexityO(N·N_τ) per sweep; worm algorithm for efficient winding/superfluid samplingN_τ = β/Δτ; worm updates propose open/close operations to sample the off-diagonal sector (condensate, superfluid density) efficiently.
M10 memoryO(N·N_τ) for the worldline configurationsModest compared to wavefunction methods; total path of all particles stored.
M11 control knobN_τ (imaginary-time slices) → discretization error O(Δτ^p); N_s (sweeps) → statistical error ∝1/√N_sConvergence in N_τ tested by comparing N_τ and 2N_τ; standard extrapolation to continuum limit.
M12 scale frontierN~10010000 particles (continuum bosons); thermodynamic-limit extrapolation via finite-size scalingState-of-the-art for ⁴He, cold bosonic atoms, hydrogen at high pressure.
M13 primary approximation / biasTrotter/discretization error O(Δτ^p) — controlled by N_τ→∞; none for bosons in the continuum limitFor fermions, restricted PIMC introduces a nodal-surface bias analogous to fixed-node DMC.
M14 hard blocker / failure modeFermion sign problem via odd permutation exchanges: cost ∝1/⟨s⟩² grows exponentially with βN; analytic continuation for spectral functions is ill-posedEven for bosons, very low temperatures require large N_τ, raising cost linearly with β.

Cost & scaling

  • Time: O(N·N_τ) per sweep; worm algorithm adds efficient off-diagonal sampling
  • Memory: O(N·N_τ) worldline paths
  • Control knob: N_τ (discretization error O(Δτ^p)); N_s (statistical error ∝1/√N_s)
  • Scale frontier: N~10010000 (continuum bosons); thermodynamic limit via finite-size scaling

Recommended for

  • Bosonic quantum liquids (⁴He, ³He–⁴He mixtures): the definitive method for finite-T superfluid properties
  • Ultracold bosonic atoms in optical lattices: Bose-Hubbard phase diagram, superfluid-Mott transition
  • Hydrogen and helium at high pressure: PIMC with fermion nodal approximation used for planetary-interior equations of state
  • Benchmark for continuum bosons at finite T: exact alternative to SSE when continuum rather than lattice description is needed
  • Per method-property-map.md: preferred for sign-free bosonic continuum systems at finite T; blocked by fermion sign problem

Benchmarks

  • ⁴He lambda transition: PIMC reproduces the superfluid fraction ρ_s/ρ vs T curve, including the transition temperature T_λ ≈ 2.17 K, from first principles ceperley_1995_path.
  • Condensate fraction ⁴He at T=1 K, P=0: n_0 ≈ 7.25% — experimental value matched by PIMC ceperley_1995_path.

Key reference: ceperley_1995_path — comprehensive review of PIMC for quantum liquids: path-integral formulation, Bose–Einstein statistics and exchange, superfluid density estimator, ⁴He benchmarks; the canonical reference for the method.

Diagrammatic Monte Carlodiagmccontrolledmethod-qmcprokofev_1998_polaron
AxisValueNote
M1 tasks / outputsSelf-energy Σ(k,ω) · one-particle Green's function G(k,ω) · vertex functions · thermodynamics (free energy, density, compressibility) · polaron binding energy and dispersionAll computed directly in the thermodynamic limit; finite-size effects absent.
M2 regimeFinite-T (Matsubara formalism) or T=0 (real-frequency); polaron ground stateBoth regimes accessible; imaginary-frequency results require analytic continuation for spectra.
M3 accuracy classControlled-bias (series truncation at order n_max); stochasticStatistical error ∝1/√N_s; systematic error from diagram-series truncation and resummation scheme; uncontrolled if the series diverges or the sign problem (diagram-sign) is severe.
M4 dimension fit (A1)Thermodynamic limit directly (no finite-size error) — the key differentiatorWorks in any dimension; no finite-size scaling needed; directly accesses the TDL.
M5 statistics & local dim (A3)Bosons and fermions in the diagrammatic language; local dim d does not appear (momentum-space formulation)Polaron problems (single fermion in a bosonic bath) are the canonical application; fermionic sign enters via diagram signs.
M6 entanglement regime (B5)Not a wavefunction method — no entanglement restrictionDiagrammatic expansion does not represent a wavefunction; entanglement concept does not apply.
M7 sign-problem dependence (C12)Diagram-sign + series-convergence are the limiters; weak–intermediate coupling regimeAt strong coupling, the diagram series may diverge or require resummation; the sign problem manifests as cancellations between diagrams of opposite sign.
M8 symmetry exploitation (C9/C10)Momentum conservation and translational invariance are explicit in the momentum-space formulation; spin SU(2) used to relate spin channelsSymmetry reduces the diagram phase space and variance; used to relate diagram classes.
M9 time complexityCost grows with maximum diagram order n_max; convergence/resummation of the series is the bottleneck, not lattice sizeNo matrix cost; the bottleneck is high-order diagram evaluation and the exponential growth of diagram topologies with order.
M10 memoryO(n_max) per configuration — stores the current diagram topology and variablesExtremely modest compared to wavefunction-based methods; no Hilbert-space dimension enters.
M11 control knobn_max (maximum diagram order) → series truncation error; resummation method (Padé, bold-line, conformal) → systematic convergence of resummed seriesCheck convergence as n_max increases; residual error is scheme-dependent.
M12 scale frontierThermodynamic limit by construction — no system-size extrapolation neededThe unique frontier: N=∞ at the cost of a perturbative expansion. Applies to homogeneous (translationally invariant) systems.
M13 primary approximation / biasSeries truncation at order n_max and resummation scheme bias; diagram-sign problem at strong couplingControlled at weak–intermediate coupling; uncontrolled when the series radius of convergence is exceeded.
M14 hard blocker / failure modeStrong coupling: diagram series diverges or has zero radius of convergence; sign problem from diagram cancellations; inhomogeneous / disordered systems break translation invariance (no momentum basis)Lattice models at strong U/t and frustrated systems are beyond the weak–intermediate coupling reach.

Cost & scaling

  • Time: cost grows with n_max; diagram topology enumeration grows factorially with order; each diagram evaluated via Monte Carlo quadrature
  • Memory: O(n_max) per configuration — negligible
  • Control knob: n_max (truncation order → series error); resummation method → systematic bias
  • Scale frontier: thermodynamic limit N=∞ by construction; n_max~814 typical for polaron problems

Recommended for

  • Polaron and impurity problems: single fermion or boson in a bosonic/fermionic bath — the founding and best-established application
  • Homogeneous correlated systems at weak–intermediate coupling: Hubbard model at moderate U/t, Bose gas near unitarity — direct TDL results without finite-size extrapolation
  • When thermodynamic-limit results are required without finite-size corrections: avoids the N→∞ extrapolation cost that burdens all lattice methods
  • Cross-check for DQMC/AFQMC: DiagMC self-energy at TDL vs DQMC finite-size-extrapolated self-energy
  • Per method-property-map.md: the only method that directly accesses N=∞; blocked at strong coupling (B7) and by inhomogeneity (disorder, interfaces)

Benchmarks

  • Fröhlich polaron: DiagMC ground-state energy and effective mass from weak coupling (α→0) through intermediate coupling (α≈6) — matches Feynman variational bound at weak coupling and improves beyond it prokofev_1998_polaron.
  • Effective mass m*/m at α=3: DiagMC 1.87(2) vs Feynman variational 1.89 — consistent at the few-percent level for the founding calculation prokofev_1998_polaron.

Key reference: prokofev_1998_polaron — founding DiagMC paper: stochastic summation of polaron Green function diagrams, first demonstration that continuous-variable diagrammatic MC gives exact polaron properties in the TDL.

6 Quantum embedding & perturbative 6

Dynamical mean-field theorydmftcontrolledoutside-core-harnessgeorges_1996_dynamical
AxisValueNote
M1 tasks / outputsSingle-particle spectral function A(ω) / self-energy Σ(ω) / DOS · Mott transition & phase diagram · finite-T thermodynamicsTwo-particle responses (susceptibilities, optical conductivity) require vertex functions → substantially more expensive; momentum resolution needs cluster extensions.
M2 regimeFinite-T natural; T=0 accessible; real-frequency via analytic continuation (MaxEnt/Padé)Imaginary-axis QMC solver gives G(iω_n); real-frequency requires analytic continuation with associated uncertainty. NRG solver gives real-frequency directly.
M3 accuracy classControlled (exact at Z→∞); uncontrolled at finite dimensionCluster DMFT (N_c sites) is systematically improvable in N_c but at exponential cost. Sign from CT-QMC introduces stochastic error for multi-orbital/cluster cases.
M4 dimension fit (A1)Exact as coordination Z→∞ / infinite dimension (Metzner–Vollhardt); finite-D neglects spatial correlations beyond the clusterSingle-site DMFT best for 3D / high-coordination; short-range correlations restored by clusters (N_c); cluster cost grows exponentially with N_c.
M5 statistics & local dim (A3)Fermions (multi-orbital supported); local Hilbert dim grows with #orbitalsCT-QMC sign worsens with #orbitals and cluster size; off-diagonal hybridization triggers sign problem.
M6 entanglement regime (B5)No explicit entanglement gate (impurity picture)The bath is integrated out; entanglement of impurity + bath is handled by the solver; no bond-dimension wall in principle.
M7 sign-problem dependence (C12)Sign-problem enters through the CT-QMC impurity solverCT-HYB/CT-INT sign worsens with multi-orbital / off-diagonal hybridization / cluster size. ED-bath and NRG solvers are sign-immune.
M8 symmetry exploitation (C9/C10)Spin, orbital, and point-group symmetries reduce the impurity problem; translation exploited in DCA (momentum-space clusters)Symmetry reduces bath/cluster dimension and cuts impurity-solver cost.
M9 time complexitySelf-consistency loop × impurity-solver cost: CT-QMC ~polynomial in β (with sign worsening for multi-orbital/cluster); ED-bath exponential in #bath sites; NRG O(N_kept³) per iterationCluster DMFT cost grows exponentially with N_c.
M10 memoryImpurity solver memory (CT-QMC: bath Green's function matrices; ED: d^{N_bath} Hilbert space)Scales with #orbitals × cluster size; memory is rarely the primary wall compared to sign/cost.
M11 control knobCluster size N_c (spatial resolution); bath size (ED solver); N_kept (NRG solver); β and #MC steps (CT-QMC statistical error)Larger N_c → exponentially harder; β convergence at low T for CT-QMC.
M12 scale frontierSingle-site DMFT: thermodynamic limit by construction; cluster DMFT: N_c ~ 4–16 routine, larger with DMRG solverInfinite-lattice is treated at the single-site level; finite N_c for cluster variants.
M13 primary approximation / biasNeglect of spatial correlations (single-site); cluster truncation (cluster DMFT)Controlled in N_c; exact at Z→∞ for single-site.
M14 hard blocker / failure modeStrong sign problem for multi-orbital / off-diagonal hybridization / large clusters (CT-QMC); spatial correlations beyond N_c (momentum differentiation, pseudo-gap precursors)1D strongly-correlated: spatial fluctuations dominate and single-site DMFT is qualitatively wrong.

Cost & scaling

  • Time: self-consistency loop × impurity-solver cost; CT-QMC ~polynomial in β; ED-bath exponential in #bath sites; cluster DMFT cost grows exponentially with N_c
  • Memory: impurity Green's function matrices (~N_orb² × N_τ); ED-bath O(d^{N_bath})
  • Control knob: N_c (cluster size, spatial resolution); β and MC steps (CT-QMC); bath size (ED solver)
  • Scale frontier: single-site — thermodynamic limit; cluster DMFT N_c ~ 4–16 routine

Recommended for

  • Correlated fermions in 3D / high-coordination lattices (Hubbard, multi-orbital Hubbard, Hund's metals) where spatial correlations are secondary
  • Mott physics and finite-T phase diagrams — the method of choice
  • Real-frequency spectral functions when NRG is used as the impurity solver
  • Multi-orbital / realistic correlated materials via DFT+DMFT
  • Per method-property-map.md: preferred over MF (A1~3D, B7 strong correlation); complement with cluster DMFT for spatial correlations, DCA for k-resolved physics

Benchmarks

  • Mott transition in the half-filled Hubbard model: U_c2 ≈ 2.9 D (Bethe lattice, half-bandwidth D=1, T=0; equivalently U_c2/t ≈ 5.8 since D=2t) georges_1996_dynamical.
  • CT-QMC solver sign: sign ~1 for single-band, degrades exponentially with #orbitals and cluster size [method-survey.md §6.1].
  • Cluster DMFT N_c convergence: local observables converge ~1/N_c; momentum-space quantities require larger N_c [method-survey.md §6.1].

Key reference: georges_1996_dynamical — the authoritative DMFT review covering the Z→∞ limit, impurity solvers, the Mott transition, and benchmarks for the Hubbard model.

Density-matrix embeddingdmetcontrolledoutside-core-harnessknizia_2012_density
AxisValueNote
M1 tasks / outputsGround-state energy · local observables (density, magnetization, double occupancy) · phase diagramsSpectral functions are harder; frequency-dependent DMET extensions exist but are not routine.
M2 regimeT=0 (ground state)Finite-T DMET extensions exist but are less developed.
M3 accuracy classControlled by fragment size N_frag + embedded solver (DMRG/FCI); deterministicAccuracy improves systematically with N_frag; controlled by the embedded solver (FCI is exact within the embedding).
M4 dimension fit (A1)Any dimensionUnlike DMFT, DMET is not restricted to infinite dimension; applicable in 1D, 2D, 3D.
M5 statistics & local dim (A3)Fermions primarily; bosonic DMET extensions existLocal Hilbert dim enters through the embedded solver (d^{2N_frag} for FCI).
M6 entanglement regime (B5)Fragment-bath entanglement captured by the Schmidt basisThe Schmidt bath truncation (exactly N_frag bath orbitals) limits long-range entanglement; controlled by N_frag.
M7 sign-problem dependence (C12)Sign-problem enters only via the embedded solverFCI and DMRG embedded solvers are sign-immune; AFQMC embedded solver carries a constraint bias.
M8 symmetry exploitation (C9/C10)Lattice symmetries can be exploited in the embedded problem and to partition the fragment tilingTranslation symmetry of the tiling reduces the number of inequivalent fragments.
M9 time complexityDominated by the impurity solver on the 2·N_frag-site embedding: FCI O(d^{2N_frag}), DMRG O(N_frag·χ³) for 1D fragments, CCSD O(N_frag⁶)Self-consistency loop on the correlation potential is cheap relative to the solver; larger N_frag → exponentially harder solver (for FCI).
M10 memoryDominated by the embedded solver: FCI O(d^{2N_frag}), DMRG O(d·χ²)Memory wall is the embedded solver; manageable for N_frag ≲ 10 with FCI, larger with DMRG.
M11 control knobFragment size N_frag — controls spatial resolution and accuracy; embedded solver accuracy (bond dim χ for DMRG, or truncation for CCSD)The two knobs must be jointly converged; FCI gives exact embedding at fixed N_frag.
M12 scale frontierThermodynamic limit by tiling; fragment N_frag ~ 2–36 sites in practice; limited by embedded solverWith DMRG embedded solver: N_frag ~ 36 (1D strips) accessible; FCI: N_frag ~ 8–12.
M13 primary approximation / biasFragment/bath truncation (Schmidt basis limited to N_frag bath orbitals); self-consistency on 1-RDM (not the full wavefunction)Controlled by N_frag; the Schmidt bath exactly captures all fragment-environment entanglement at fixed mean-field reference.
M14 hard blocker / failure modeEmbedded solver cost grows exponentially with N_frag (FCI); 2D fragments require large N_frag for spatial convergence; long-range correlations beyond fragment size not capturedSelf-consistency can have convergence issues (multiple solutions); spectral functions not natively accessible.

Cost & scaling

  • Time: embedded solver on 2·N_frag sites: FCI O(d^{2N_frag}), DMRG O(N_frag·χ³), CCSD O(N_frag⁶); self-consistency loop cheap
  • Memory: embedded solver: FCI O(d^{2N_frag}), DMRG O(d·χ²)
  • Control knob: N_frag (fragment size, spatial resolution and accuracy) + embedded solver knob (χ for DMRG)
  • Scale frontier: thermodynamic limit by tiling; N_frag ~ 2–36 in practice

Recommended for

  • Hubbard model in 2D where DMRG is expensive (finite width) and QMC has sign problem
  • Correlated electron problems in any dimension where DMFT spatial correlation neglect is unsatisfactory
  • Strongly-correlated lattice models (B7 strong correlation OK, unlike perturbation theory)
  • Complement to DMFT: DMET at small N_frag ~ DMFT; growing N_frag systematically improves spatial correlations
  • Per method-property-map.md: applicable when 2D/3D frustration blocks QMC and spatial correlations beyond single-site DMFT are needed

Benchmarks

  • 1D Hubbard at half-filling: DMET energy agrees with Bethe Ansatz to < 0.1% for N_frag = 2 knizia_2012_density.
  • 2D Hubbard: DMET with N_frag = 4–16 matches AFQMC/DMRG benchmark energies to ~0.5% at U/t = 4–8 [method-survey.md §6.2].
  • Mott transition: DMET correctly captures the metal-insulator transition for the half-filled Hubbard model knizia_2012_density.

Key reference: knizia_2012_density — original DMET paper introducing the Schmidt-bath embedding and self-consistency scheme; benchmarks on 1D and 2D Hubbard models.

Numerical renormalization groupnrgcontrolledoutside-core-harnessbulla_2007_numerical
AxisValueNote
M1 tasks / outputsKondo scale T_K · impurity spectral function A(ω) at all T · thermodynamics (entropy, specific heat, susceptibility as function of T) · fixed-point flow diagrams · real-frequency self-energy Σ(ω) as a DMFT solverNRG gives real-frequency quantities directly without analytic continuation — its key advantage over CT-QMC.
M2 regimeAll T (T=0 to high-T crossovers) and real-frequencyThe logarithmic bath discretization gives exponential energy resolution; no imaginary-time or analytic continuation needed.
M3 accuracy classControlled (convergence in N_kept and Λ→1); deterministicEnergy truncation at N_kept introduces a controlled systematic error; Λ→1 is the continuum limit but raises cost.
M4 dimension fit (A1)Impurity / quantum-dot problems (0D impurity + bath) — not a bulk lattice methodDesigned for impurity models (Anderson, Kondo, multi-channel, SU(N) Kondo); used as the DMFT impurity solver to access real-frequency self-energies.
M5 statistics & local dim (A3)Fermions (and bosons via bosonic NRG); N_kept must grow exponentially with #bands/channelsLimited to ~1–3 channels; multi-channel NRG requires exponentially larger N_kept → rapidly intractable.
M6 entanglement regime (B5)No entanglement wall for impurity problems (the bath is sequentially integrated out)Wilson chain geometry makes the entanglement growth manageable by construction.
M7 sign-problem dependence (C12)Sign-immuneNot a Monte Carlo method; exact iterative diagonalization with energy truncation.
M8 symmetry exploitation (C9/C10)U(1), SU(2), Z₂, and non-Abelian symmetries exploited at each iteration to block-diagonalize and reduce N_kept requiredSymmetry is crucial: SU(2) spin + U(1) charge reduce cost by an order of magnitude; non-Abelian NRG (Weichselbaum) extends reach to multi-channel.
M9 time complexityO(N_kept³) per NRG iteration (diagonalize a ≈ d·N_kept matrix, truncate to N_kept); N_kept must grow exponentially with #bands/channelsTypically ~50–100 iterations for Λ~2–3; total cost ~N_iter × (d·N_kept)³.
M10 memoryO(N_kept²) per shell for the density matrix and kept statesMemory manageable for 1–2 channel problems; grows quickly with #channels.
M11 control knobN_kept (retained states per shell) — controls energy truncation error; Λ (discretization parameter) — controls bath discretization error (Λ→1 is continuum)Both knobs must be converged; typical Λ~2–4, N_kept~500–3000 with full symmetry.
M12 scale frontier1–3 channel impurity problems; single- and two-impurity systems; multi-orbital impurities with symmetryAt 3+ channels without high symmetry: N_kept requirement becomes exponentially prohibitive.
M13 primary approximation / biasEnergy truncation at N_kept (controlled); logarithmic bath discretization with Λ (controlled, Λ→1 limit)Both errors are systematically reducible; the truncation error is the leading one at fixed Λ.
M14 hard blocker / failure mode#channels/bands is the wall: N_kept must grow exponentially with #bands/channels — NRG limited to ~1–3 channels; not applicable to bulk lattice problemsBosonic baths with sub-ohmic spectral density can cause issues; high-frequency resolution limited by Λ.

Cost & scaling

  • Time: O(N_kept³) per NRG iteration; N_kept grows exponentially with #channels; ~50–100 iterations total
  • Memory: O(N_kept²) per shell; manageable for 1–2 channels with symmetry
  • Control knob: N_kept (truncation error) and Λ (discretization error, Λ→1 is the continuum limit)
  • Scale frontier: 1–3 channel impurity problems; quantum dots; DMFT impurity solver for real-frequency self-energies

Recommended for

  • Anderson and Kondo impurity models — the definitive solver for all T and real-frequency
  • Quantum-dot transport where Kondo resonance and real-frequency spectral features matter
  • DMFT impurity solver when real-frequency self-energies are needed without analytic continuation
  • Multi-channel Kondo up to ~2–3 channels with non-Abelian symmetry
  • Per method-property-map.md: the method of choice for 0D impurity problems (A1=0D) whenever real-frequency or all-T data is required

Benchmarks

  • Kondo scale: T_K/D ~ exp(-π U / 8Γ) for symmetric Anderson model; reproduced to < 1% with N_kept~1000, Λ~2 bulla_2007_numerical.
  • Spectral function sum rule: ∫ A(ω) dω = 1; satisfied to < 0.1% in fdm-NRG [method-survey.md §6.3].
  • Channel limit: N_kept required scales ~d^{N_ch} with #channels N_ch; 3-channel NRG requires N_kept~10⁴ even with symmetry [method-survey.md §6.3].

Key reference: bulla_2007_numerical — comprehensive review of NRG theory, algorithms (including fdm-NRG), thermodynamics, spectral functions, and applications to Anderson/Kondo models.

Functional renormalization groupfrgcontrolledoutside-core-harnessmetzner_2011_functional
AxisValueNote
M1 tasks / outputsLeading-instability phase diagrams (magnetic, superconducting, charge-density-wave) · susceptibilities · pseudo-critical scales Λ_c · renormalized quasiparticle propertiesNOT ground-state energies of strongly-correlated phases; the divergence signals an instability, not the true ordered state.
M2 regimeT=0 and finite-T (cutoff flow equations work at both); real-frequency accessible via Keldysh fRGStandard formulation at T=0 or T>0 in Matsubara; real-time dynamics via Keldysh formulation.
M3 accuracy classControlled approximation (truncation order is the knob); uncontrolled at strong couplingWeak–intermediate coupling: one-loop + Katanin truncation is a controlled expansion; strong coupling truncation breaks down.
M4 dimension fit (A1)Any dimension (1D, 2D, 3D)Particularly powerful in 2D (competing instabilities in the Hubbard model); 1D reproduces bosonization results.
M5 statistics & local dim (A3)Fermions (fermionic fRG); bosonic fRG also existsFermionic sign enters only via the choice of statistics in the propagators; no sign problem in the sense of QMC.
M6 entanglement regime (B5)No entanglement restriction (field-theoretic, not wavefunction-based)fRG is not a wavefunction method; entanglement is not a controlling parameter.
M7 sign-problem dependence (C12)Sign-immune (not a Monte Carlo method)Deterministic flow equations; no stochastic sampling.
M8 symmetry exploitation (C9/C10)Symmetries (SU(2), time-reversal, lattice point group) reduce vertex parameterization cost; translation used to parameterize vertex in momentum spaceExploiting symmetry reduces the number of independent vertex components and accelerates the flow.
M9 time complexity~O(N_patch³) for the two-particle vertex flow (three momentum sums); higher-order frequency dependence adds factorsN_patch is the number of momentum patches on the Fermi surface; frequency dependence multiplies cost by N_ω³.
M10 memoryO(N_patch²) to O(N_patch³) for storing the two-particle vertex (depending on channel decomposition)Channel decomposition (particle-particle, particle-hole crossed/direct) allows efficient storage.
M11 control knobTruncation order (one-loop → Katanin → multiloop → ...) — controls vertex approximation; N_patch — controls momentum resolutionHigher truncation order is closer to exact but exponentially more expensive; N_patch convergence required.
M12 scale frontierThermodynamic limit (field-theoretic, no finite-size constraint); limited by N_patch and truncation orderPatch numbers N_patch ~ 16–96 are typical; multiloop fRG up to 3–5 loop orders is state of the art.
M13 primary approximation / biasTruncation of the vertex hierarchy at finite order (one-loop / Katanin / multiloop) — the controlled approximationOne-loop truncation is uncontrolled at strong coupling; multiloop fRG converges the loop expansion.
M14 hard blocker / failure modeStrong coupling not controlled: vertex flow diverges when the coupling is too large (beyond weak–intermediate); cannot describe Mott physics or strongly frustrated statesTwo-particle interaction vertex divergence Γ^(4) → ∞ is a signature of instability, not a numerical failure — but the truncation breaks down before the physical strong-coupling fixed point.

Cost & scaling

  • Time: ~O(N_patch³) for two-particle vertex flow (three momentum sums); multiplied by N_ω³ with frequency dependence
  • Memory: O(N_patch²)O(N_patch³) for vertex storage (channel-decomposed)
  • Control knob: truncation order (one-loop → multiloop) and N_patch (momentum resolution)
  • Scale frontier: thermodynamic limit; N_patch ~ 16–96 typical; multiloop 3–5 loop orders is frontier

Recommended for

  • 2D Hubbard model at weak–intermediate coupling — detecting AF, SC, CDW instabilities and their competition
  • Kohn-Luttinger superconductivity and unconventional pairing symmetry identification
  • Competing order parameter scenarios where mean-field is biased toward a single channel
  • 1D correlated systems where fRG reproduces bosonization / Luttinger liquid results
  • Per method-property-map.md: complement to MF (/method-mf) at weak coupling; sign-immune alternative to sign-ful QMC in frustrated 2D systems

Benchmarks

  • Hubbard model U/t = 2, square lattice: fRG pseudo-critical scale Λ_c consistent with RPA at weak coupling; leading instability AF at half-filling, d-SC away from half-filling metzner_2011_functional.
  • N_patch convergence: local observables converge at N_patch ~ 32–64 for 2D Hubbard [method-survey.md §6.4].
  • Multiloop fRG vs one-loop: loop convergence demonstrated for 2D Hubbard at U/t ≲ 3 [method-survey.md §6.4].

Key reference: metzner_2011_functional — comprehensive review of fermionic fRG: flow equations, truncation schemes, vertex parameterizations, and applications to Hubbard-model instabilities.

Coupled clustercoupled-clustercontrolledoutside-core-harnessbartlett_2007_coupled
AxisValueNote
M1 tasks / outputsGround-state energies and observables (high accuracy) · EOM-CC for excitations / spectra · ionization potentials and electron affinities (IP/EA-EOM-CC)High accuracy when single-reference; EOM-CC gives excited states but at O(N⁶)O(N⁷) cost.
M2 regimeT=0 (ground state and selected excited states)Finite-T extensions (thermal CC) exist but are non-standard.
M3 accuracy classControlled (CC hierarchy T₁T₂T₃... → FCI limit); deterministic; excellent for dynamic correlationAccuracy collapses for strong static correlation (multi-reference breakdown); not variational.
M4 dimension fit (A1)Any dimension (molecular and periodic/solid-state variants)Periodic CCSD/CCSD(T) for solids (e.g. crystalline materials); EOM-CC for surfaces and defects.
M5 statistics & local dim (A3)FermionsLocal Hilbert dim enters through the number of virtual orbitals N; cost scales O(N⁶)O(N⁷).
M6 entanglement regime (B5)Dynamic correlation (moderate entanglement)Excellent for weakly-entangled systems; breaks down for volume-law / strongly-entangled (multi-reference) states.
M7 sign-problem dependence (C12)Sign-immune (not a Monte Carlo method)Deterministic algebraic equations; no sign problem.
M8 symmetry exploitation (C9/C10)Point group, spin, and translational symmetry reduce the orbital/amplitude spaceSymmetry-adapted CC reduces the number of independent amplitudes and cuts cost.
M9 time complexityCCSD O(N⁶), CCSD(T) O(N⁷) in the number of orbitals NThe "gold standard" scaling of quantum chemistry; N here is total orbital count (occupied + virtual).
M10 memoryO(N⁴) for storing the two-particle amplitudes T₂ (dominant cost)The T₂ tensor with four orbital indices is the memory bottleneck; O(N_occ² · N_virt²).
M11 control knobCC hierarchy level (CCSD → CCSD(T) → CCSDT → ...) + basis set completenessThe CC expansion is systematically improvable toward FCI; basis set extrapolation is also needed.
M12 scale frontier~100–500 orbitals for CCSD (molecular); ~20–50 k-points for periodic CCSD; fragmentation/local CC extends reachExplicitly correlated F12-CCSD(T) improves basis convergence; local CC (DLPNO-CCSD(T)) reaches 1000s of atoms.
M13 primary approximation / biasSingle-reference assumption (T acts on a single HF determinant) — fails for strong static correlation / open-shell degeneracy / frustrationMulti-reference CC (MRCC) extends reach but is much more costly and complex.
M14 hard blocker / failure modeStrong static correlation / frustration / open-shell degeneracy: single-reference breakdown — T₁ diagnostic > 0.02 signals multi-reference character and unreliable resultsFrustrated lattice models (B8), doped Mott insulators, near-degenerate transition states: CC is not the right tool.

Cost & scaling

  • Time: CCSD O(N⁶), CCSD(T) O(N⁷) in orbital count N
  • Memory: O(N⁴) for T₂ amplitudes (N_occ² × N_virt²)
  • Control knob: CC hierarchy level (CCSD → CCSD(T) → CCSDT) + basis set size
  • Scale frontier: ~100–500 orbitals for CCSD; ~1000s of atoms with local DLPNO-CCSD(T)

Recommended for

  • Quantum chemistry: molecular energetics — the gold standard for weakly–moderately correlated molecules (when T₁ < 0.02)
  • Periodic CCSD(T) for solids without strong static correlation (metals with moderate U/t, van der Waals solids)
  • Benchmark reference for DFT, HF, and perturbation theory (MP2) comparisons
  • EOM-CC for excitation spectra of weakly-correlated systems
  • Per method-property-map.md: do NOT use for frustrated magnets (B8), doped Mott insulators (B7 strong), or strongly-entangled states

Benchmarks

  • CCSD(T)/CBS atomization energies: mean absolute deviation < 0.5 kcal/mol for the W4 thermochemistry benchmark set bartlett_2007_coupled.
  • T₁ diagnostic: T₁ > 0.02 for HF, T₁ ~ 0.05 for biradicals — signals multi-reference breakdown bartlett_2007_coupled.
  • Periodic CCSD(T): cohesive energy of solid Ne to < 0.1 meV/atom [method-survey.md §6.5].

Key reference: bartlett_2007_coupled — comprehensive review of CC theory in quantum chemistry: amplitude equations, CCSD(T) derivation, EOM-CC, and applications across quantum chemistry.

Many-body perturbation theory (GW/RPA/GF2)mbpt-gwuncontrolledoutside-core-harnessgolze_2019_gw
AxisValueNote
M1 tasks / outputsQuasiparticle band structures / photoemission spectra (GW) · quasiparticle gaps and ionization potentials · correlation energies (RPA) · screened interactions W · optical spectra (BSE on top of GW) · spectra of correlated solids (SEET)Direct gap: GW; optical gap: BSE/GW; ground-state energy: RPA; strongly-correlated spectra: SEET.
M2 regimeT=0 (standard); finite-T Matsubara GW also implementedReal-frequency GW via analytic continuation or real-axis contour deformation; imaginary-axis for correlation energies.
M3 accuracy classUncontrolled perturbative approximation (ring/RPA resummation); self-consistent GW is parameter-free but not systematically improvable to exact, and fails at strong coupling (G₀W₀ also carries a starting-point/DFT-functional bias)GW sums ring diagrams only — not a controlled order-by-order expansion toward exactness; sc-GW removes starting-point dependence but still misses Mott physics; G₀W₀ inherits the DFT-functional bias of the starting point.
M4 dimension fit (A1)Any dimension (molecules and periodic solids)GW for molecules (PySCF, TURBOMOLE) and periodic solids (VASP, Abinit, BerkeleyGW); no dimensional restriction.
M5 statistics & local dim (A3)FermionsFermionic diagrammatic expansion; N is orbital/basis count.
M6 entanglement regime (B5)No entanglement gate (diagrammatic / field-theoretic)Not a wavefunction method; entanglement is not a controlling parameter.
M7 sign-problem dependence (C12)Sign-immune (deterministic diagrammatic resummation)No stochastic sampling in standard GW/RPA/GF2; diagrammatic MC variants exist but are non-standard.
M8 symmetry exploitation (C9/C10)k-point sampling exploits translation symmetry; point-group symmetry reduces the response functionPeriodic GW exploits Bloch symmetry; symmetry-adapted W reduces cost.
M9 time complexityGW O(N⁴) (→ O(N³) with factorized W); RPA O(N⁴)O(N³); GF2 O(N⁵) in orbital count NFactorized GW (O(N³)) uses resolution-of-identity (RI) or density fitting; frequency integration adds N_ω factor.
M10 memoryO(N³) for storing W(r,r',ω) or its factored form; O(N²) for G and ΣMemory scales as orbital count cubed for the full polarizability; RI reduces to O(N² × N_aux).
M11 control knobStarting point (DFT functional for G₀W₀) · self-consistency level (G₀W₀ → qpGW → sc-GW) · basis set size · frequency grid N_ωSelf-consistent GW eliminates starting-point dependence at higher cost.
M12 scale frontier~100–1000 orbitals for GW (periodic: 50–500 bands × k-points); ~10⁴ orbitals for local GWStochastic GW and low-scaling O(N³) algorithms extend to large periodic systems.
M13 primary approximation / biasDiagram truncation (GW = ring resummation; GF2 = 2nd order; higher-order diagrams neglected)Weak-coupling: controlled; strong coupling (Mott): uncontrolled breakdown. SEET corrects by embedding.
M14 hard blocker / failure modeStrong correlation out of reach: GW and RPA break down for Mott insulators, strongly-frustrated systems, and near-degenerate open shells (SEET partially fixes this by embedding)G₀W₀ starting-point dependence is large for strongly-correlated systems; self-consistent GW still misses Mott physics.

Cost & scaling

  • Time: GW O(N⁴) (→ O(N³) factorized); RPA O(N⁴)O(N³); GF2 O(N⁵) in orbital count N
  • Memory: O(N³) for full polarizability; O(N² × N_aux) for RI-approximated W
  • Control knob: self-consistency level (G₀W₀ → sc-GW); basis set; frequency grid N_ω
  • Scale frontier: ~100–1000 orbitals for GW; periodic GW with 50–500 bands

Recommended for

  • Semiconductor and insulator band gaps (GW is the method of choice for photoemission spectra)
  • van der Waals and dispersion-dominated systems where RPA correlation energies are essential
  • Quasiparticle renormalization in metals and correlated metals (beyond DFT)
  • GF2 for molecular quasiparticle spectra at moderate correlation
  • SEET for correlated solids where GW fails but pure DFT+U is insufficient
  • Per method-property-map.md: preferred over DFT (higher accuracy for gaps); restricted to weak–intermediate coupling (strong correlation → DMFT or DMET)

Benchmarks

  • GW band gaps of semiconductors: mean absolute error ~0.1 eV (G₀W₀@PBE0) vs ~0.7 eV (DFT-LDA) for the GW100 benchmark set golze_2019_gw.
  • RPA correlation energies: accurate van der Waals coefficients for rare-gas dimers to < 5% [method-survey.md §6.6].
  • G₀W₀@PBE ionization potentials: mean absolute error ≈ 0.3 eV on the GW100 molecular benchmark set (van Setten et al., JCTC 2015) golze_2019_gw.

Key reference: golze_2019_gw — comprehensive practical guide to GW approximation: theory, algorithms (G₀W₀, qpGW, sc-GW), implementations, and benchmarks for molecules and solids.

7 Mean-field, circuit, certified 6

Hartree–Fock / UHF (+ DFT baseline)hartree-fockuncontrolledmethod-mfarovas_2022_hubbard
AxisValueNote
M1 tasks / outputsMean-field phase diagrams · order parameters · band structures · Slater-determinant trial states for DMRG/VMC/AFQMCPrimary use: fast survey of symmetry-breaking phases and orbital structure.
M2 regimeT=0 (finite-T via mean-field free energy straightforward)Thermal HF replaces zero-T occupations with Fermi–Dirac; cost unchanged.
M3 accuracy classUncontrolled, deterministicCarries no entanglement; no systematic expansion parameter in general.
M4 dimension fit (A1)Best in high dimension / large coordination Z; fails in 1DLeading fluctuation corrections scale ~1/Z (the √Z is the DMFT *hopping* rescaling t=t*/√Z, not the MF error); MF is exact as Z→∞ (Metzner–Vollhardt 1989); Mermin–Wagner forbids spontaneous symmetry breaking in 1D → spurious order; also unreliable under B8 frustration.
M5 statistics & local dim (A3)Fermions primarily; spinless/spinful bosons via mean-field decouplingLocal dim d=2 (spin-½ Hubbard) or d=4 (Hubbard with spin+charge); cost independent of d.
M6 entanglement regime (B5)Product-state (zero entanglement)Slater determinant = product in occupation-number basis; no area- or volume-law entanglement.
M7 sign-problem dependence (C12)Sign-immuneDeterministic eigenvalue problem; no Monte Carlo sampling.
M8 symmetry exploitation (C9/C10)Block-diagonalize by U(1) (charge/spin), translation k, point group; reduces SCF costHF in the k-space Bloch basis is particularly efficient for periodic systems.
M9 time complexityO(N³)–O(N⁴) per SCF iteration (Fock build)Negligible compared to correlated methods; converges in ≲100 SCF steps for typical lattice models.
M10 memoryO(N²)Stores the density matrix / Fock matrix; cheap.
M11 control knobNone (uncontrolled) — quality assessed by post-HF correlation energy or by comparison with ED/DMRGFluctuation corrections in 1/Z give a rough measure of MF reliability.
M12 scale frontierN ~ 10³–10⁵ sites feasible with k-space HF; DFT codes reach millions of atomsNot the limiting factor; cost is always negligible relative to the correlated solver.
M13 primary approximation / biasMean-field decoupling — no fluctuations, no dynamical correlations, no entanglementUncontrolled (no small parameter in 2D/3D at intermediate coupling).
M14 hard blocker / failure modeFails in 1D (Mermin–Wagner, spurious order); fails under strong B8 frustration; misses Mott physics (B7) and fractionalization by constructionPost-HF methods (VMC, DMRG, AFQMC) required for quantitative accuracy.

Cost & scaling

  • Time: O(N³)–O(N⁴) per SCF iteration; converges in ≲100 iterations
  • Memory: O(N²) (density matrix + Fock matrix)
  • Control knob: none (uncontrolled); MF reliability assessed via 1/Z fluctuation scale
  • Scale frontier: N ~ 10³–10⁵ in k-space (periodic lattice); millions in DFT

Recommended for

  • Fast survey / baseline: any lattice model (Hubbard, t-J, periodic Anderson) before investing in correlated solvers
  • High-dimensional / large-Z limit: Bethe-lattice or high-d hypercubic models where MF is exact or nearly so
  • Seed generation: trial wavefunctions for AFQMC (HF determinant), VMC (UHF Slater det), DMRG (initial MPS from HF orbitals)
  • Per method-property-map.md: preferred whenever only qualitative phase boundaries or orbital structure are needed and Z is large

Benchmarks

  • HF transition: Néel order in half-filled Hubbard model at any U>0 in 2D HF (spurious; correct answer requires QMC) arovas_2022_hubbard.
  • MF accuracy improves as coordination Z→∞; Bethe-lattice Z=∞ HF = exact (Metzner–Vollhardt 1989).
  • Fluctuation corrections ~1/Z verified in method-survey.md §7.1.

Key reference: arovas_2022_hubbard — comprehensive review of the Hubbard model covering HF/UHF mean-field theory, phase diagrams, DFT+U, and the full landscape of correlated methods; authoritative all-details source for the MF baseline.

Linear spin-wave / large-Sspin-wavecontrolledmethod-mftoth_2015_linear
AxisValueNote
M1 tasks / outputsMagnon dispersions · dynamical structure factor S(q,ω) at the harmonic level · sublattice-magnetization 1/S corrections · thermodynamic quantities (magnon heat capacity, entropy) · neutron scattering cross-sectionsS(q,ω) from LSWT is directly comparable to neutron scattering data; main experimental touchpoint.
M2 regimeT=0 (LSWT); finite-T via Bose–Einstein magnon occupationsLow-T thermodynamics reliable; near T_N perturbation theory breaks down.
M3 accuracy classControlled in 1/S (semiclassical); deterministicLarge-S (A3) is the control parameter; corrections are O(1/S) and can be computed systematically.
M4 dimension fit (A1)2D and 3D ordered magnets; 1D requires caution (strong quantum fluctuations)In 1D, 1/S corrections diverge → DMRG/QMC preferred; Mermin–Wagner forbids 2D order at T>0.
M5 statistics & local dim (A3)Spins with large S (A3); bosons via Holstein–PrimakoffAccuracy improves with S; S=1/2 is borderline (large corrections); S≥1 reliable at leading order.
M6 entanglement regime (B5)Product-like (ordered classical background + small Gaussian fluctuations)LSWT wavefunction is a product coherent state with Bogoliubov squeezed-vacuum corrections.
M7 sign-problem dependence (C12)Sign-immuneDeterministic eigenvalue problem; no Monte Carlo.
M8 symmetry exploitation (C9/C10)Magnetic space group (C9) block-diagonalizes the m×m BdG matrix; translation (C10) gives k-space formulationSymmetry reduces the m×m problem and labels magnon bands.
M9 time complexityO(m³) per k-point × #k-points (diagonalization of the m×m BdG Hamiltonian)Cheap — entire magnon band structure computed in seconds to minutes even for large magnetic unit cells.
M10 memoryO(m²) per k-point (BdG matrix storage)Negligible.
M11 control knob1/S expansion order; S magnitudeHigher orders (1/S²) give magnon–magnon scattering, lifetime broadening, and order-parameter renormalization.
M12 scale frontierMagnetic unit cells with hundreds of spins feasible (SpinW); no thermodynamic-size wallCost grows only as O(m³×N_k); arbitrary system size accessible for ordered phases.
M13 primary approximation / biasRequires magnetic order (expands around an ordered classical state); large S (A3) is the control parameterUncontrolled at S=1/2 in frustrated systems; 1/S corrections blow up → signal of disorder / spin liquid.
M14 hard blocker / failure modeB8 frustration / small S1/S corrections blow up (breakdown is itself the diagnostic signal of a spin-liquid or disordered ground state); method inapplicable to Mott insulators without order; real-time dynamics require additional approximationsNear quantum-critical points, corrections to LSWT diverge.

Cost & scaling

  • Time: O(m³) per k-point × #k-points (BdG diagonalization); seconds to minutes for any magnetic supercell
  • Memory: O(m²) per k-point — negligible
  • Control knob: 1/S expansion order; larger S → smaller corrections → better controlled
  • Scale frontier: no thermodynamic-size wall; arbitrary N for ordered phases; m (sublattice count) is the only cost driver

Recommended for

  • Ordered magnets (B7): Heisenberg model on bipartite or weakly frustrated lattices with S≥1 — the method of choice
  • Neutron scattering interpretation: incommensurate magnetic structures, single-Q order, complex magnetic unit cells (SpinW) toth_2015_linear
  • Baseline for frustrated systems: run LSWT first; if 1/S corrections are large → suspect spin-liquid ground state → switch to ED/DMRG
  • Per method-property-map.md: use whenever B7 magnetic order is established and S is not too small

Benchmarks

  • KCuF₃ (S=1/2, 3D AF): LSWT magnon dispersion matches INS to ~5%; 1/S corrections ~20% of zone-boundary gap.
  • MnF₂ (S=5/2, 3D AF): LSWT <1% error in magnon frequencies (large-S limit verified).
  • S(q,ω) from SpinW for incommensurate langasite — verified against neutron data in toth_2015_linear.
  • Breakdown signal: kagome S=1/2 Heisenberg → 1/S corrections diverge → confirms spin-liquid candidate (consistent with method-survey.md §7.2).

Key reference: toth_2015_linear — SpinW paper: general LSWT algorithm for arbitrary magnetic structures including single-Q incommensurate order; the authoritative implementation reference for neutron scattering applications.

Slave-particle / parton mean fieldslave-particleuncontrolledmethod-mfkotliar_1986_new
AxisValueNote
M1 tasks / outputsCandidate spin-liquid / fractionalized ansätze and their energies · mean-field phase diagrams · spinon/holon dispersionsPrimary use: map out fractionalized phases (Z₂ / U(1) spin liquids, Mott physics) inaccessible to HF.
M2 regimeT=0 (thermal Schwinger-boson extensions possible)Finite-T generalization exists for Schwinger-boson MF but less common.
M3 accuracy classUncontrolled, deterministic (MF saddle point); Gutzwiller-projected states are variational-upper via VMCBeyond the saddle point, fluctuations require gauge-field theory — no small expansion parameter in general; Gutzwiller-VMC branch is variational.
M4 dimension fit (A1)2D and 3D; works in any dimension at MF levelGauge fluctuations (corrections beyond MF) depend strongly on dimension; 1D is often handled better by bosonization or DMRG.
M5 statistics & local dim (A3)Fermions (Abrikosov/slave-boson partons) and bosons (Schwinger bosons / slave bosons); spins and Hubbard-model electronsLocal dim d enters only the parton-space HF step; cost independent of d.
M6 entanglement regime (B5)Mean-field parton state = product (zero entanglement in parton space); Gutzwiller projection introduces physical-space correlationsParton MF carries no entanglement barrier; VMC evaluation of Gutzwiller-projected state does not impose a bond-dimension wall but carries ansatz bias.
M7 sign-problem dependence (C12)Sign-immune (MF / parton saddle point is deterministic); Gutzwiller-VMC may carry a sign problem for fermionic partonsParton MF: deterministic. VMC branch: sign problem possible but mitigated by the Gutzwiller projector structure.
M8 symmetry exploitation (C9/C10)Lattice translation, point-group, and spin-rotation symmetries used to classify the parton ansatz (PSG — projective symmetry group)Projective symmetry group (Wen) classifies distinct mean-field spin-liquid ansätze; symmetry reduces the parton Hamiltonian to k-space blocks.
M9 time complexitySelf-consistent parton MF: O(N³) per SCF iteration (HF-like, where N = system size / #parton flavors); Gutzwiller-projected VMC: O(N_s · N²) per sampleCost is dominated by VMC if the Gutzwiller route is taken; the MF step alone is cheap.
M10 memoryO(N²) (parton density matrix / Fock matrix); VMC: O(N²) for the Slater-determinant updateCheap at the MF level.
M11 control knobNone at the MF saddle-point level (uncontrolled); VMC branch: N_s samples control statistical error ∝1/√N_sGauge fluctuations beyond MF are in principle a systematic expansion but are rarely carried out beyond the Gaussian level.
M12 scale frontierMF: N ~ 10³–10⁵ sites (no bottleneck); Gutzwiller-VMC: up to N ~ 10²–10³ sites (limited by VMC)MF phase diagrams thermodynamic; VMC limited by the cost of Slater-determinant evaluation.
M13 primary approximation / biasMean-field (saddle-point) decoupling of the parton constraint — no fluctuations, no dynamical gauge field, no entanglementUncontrolled: can predict a spin liquid that is destabilized by gauge fluctuations; quality assessed only by comparison with VMC energetics or QMC.
M14 hard blocker / failure modeGauge fluctuations may gap or confine the parton excitations, invalidating the MF ansatz; no convergence parameter signals this within MF; fails to capture Luttinger-liquid physics in 1DBreakdown is silent — a "spin-liquid" MF solution may be a confined phase in reality. Requires cross-check with VMC / QMC / DMRG.

Cost & scaling

  • Time: O(N³) per SCF iteration at the parton MF level (HF-like); O(N_s · N²) for Gutzwiller-VMC
  • Memory: O(N²) (parton density matrix)
  • Control knob: none at MF level (uncontrolled); VMC branch: N_s samples, error ∝1/√N_s
  • Scale frontier: MF thermodynamic; VMC up to N ~ 10²–10³

Recommended for

  • Spin-liquid / fractionalized phases (B7): triangular / kagome / honeycomb Heisenberg and Hubbard models where HF yields trivial order
  • Mott transition: half-filled Hubbard model near the Mott point where charge and spin degrees of freedom decouple
  • Ansatz generation: Gutzwiller-projected parton states as trial wavefunctions for VMC (the VMC energy then constitutes a variational upper bound)
  • Per method-property-map.md: the method of choice when the target phase is B7 fractionalized and HF/DMRG cannot resolve the spin-liquid ansatz structure

Benchmarks

  • Slave-boson MF reproduces the Brinkman–Rice metal-insulator transition in the half-filled Hubbard model at U_c = 8t (exact Gutzwiller in d→∞; deviates in 2D) kotliar_1986_new.
  • Schwinger-boson MF on the triangular Heisenberg antiferromagnet predicts a Z₂ spin liquid at J₂/J₁ > 0.08; VMC energetics partially support this (see method-survey.md §7.3).
  • Gutzwiller-VMC energies on the Heisenberg kagome lattice are within 1–2% of DMRG benchmarks for accessible cluster sizes.

Key reference: kotliar_1986_new — introduced the slave-boson functional-integral approach and showed that the Gutzwiller approximation emerges as its saddle point; the foundational paper for slave-boson MF on strongly correlated lattice models.

Cluster mean field / VCAvca-cluster-mfcontrolledmethod-mfmaier_2004_quantum
AxisValueNote
M1 tasks / outputsSpectral functions A(k,ω) / A(ω) · phase boundaries (Mott transition, magnetic / superconducting order) · symmetry-breaking order parameters · single-particle Green's functionVCA gives the full lattice spectral function via the periodized cluster self-energy; CMF gives order parameters and phase diagrams.
M2 regimeT=0 (finite-T via thermal cluster Green's function straightforward)The cluster ED solver is zero-temperature by default; finite-T extensions use finite-T ED (FTLM/TPQ) inside the cluster.
M3 accuracy classControlled by cluster size N_c (systematic improvement)Adding sites to the cluster systematically reduces the inter-cluster mean-field error; exact in the limit N_c → ∞.
M4 dimension fit (A1)2D (primary target); quasi-1D laddersDesigned for 2D correlated lattice models (Hubbard, t-J, periodic Anderson); inter-cluster MF less accurate in 1D where fluctuations dominate.
M5 statistics & local dim (A3)Fermions, spins, hard-core bosons; local dim d enters through the ED cluster cost d^{N_c}Same d^N exponential wall as ED, restricted to the cluster; typical d=2–4 (spin-½ to Hubbard).
M6 entanglement regime (B5)Within-cluster entanglement: exact (full ED Hilbert space); inter-cluster entanglement: mean-field / neglectedShort-range entanglement captured exactly; long-range inter-cluster entanglement suppressed by the MF decoupling.
M7 sign-problem dependence (C12)Sign-immune (ED cluster solver, no Monte Carlo)Deterministic; no sign problem regardless of frustration or doping level.
M8 symmetry exploitation (C9/C10)Cluster point-group and translation symmetry block-diagonalize the ED step; lattice translation used to periodize the self-energy (k-resolved spectral functions)C9 symmetry reduces cluster ED cost significantly; periodization of the cluster self-energy restores approximate lattice momentum resolution.
M9 time complexityDominated by exact cluster ED: O(d^{3N_c}) (full diagonalization) or O(n_iter · N · d^{N_c}) (Lanczos); inter-cluster coupling O(N_c²) per self-consistency step — cheapCost is exponential in cluster size N_c; the inter-cluster part adds negligible overhead. Typical cluster sizes N_c = 4–16 sites.
M10 memoryO(d^{2N_c}) for cluster ED (full) or O(d^{N_c}) for Lanczos; lattice Green's function O(N_c² · N_ω)Memory limit set by cluster ED, not the lattice size.
M11 control knobCluster size N_c — larger clusters reduce inter-cluster MF error systematicallyNo statistical error; convergence checked by increasing N_c (e.g. 2×2 → 2×3 → 2×4 tiling).
M12 scale frontierCluster: up to N_c ~ 12–20 sites (ED wall); lattice: thermodynamic limit reached by the mean-field coupling of an infinite tilingThe lattice itself can be infinite (thermodynamic limit); only the cluster is size-limited.
M13 primary approximation / biasNeglect of inter-cluster quantum fluctuations / entanglement (replaced by MF or the VCA self-energy functional saddle point)Controlled: bias decreases systematically with N_c. For small clusters, inter-cluster fluctuations can be large (especially near quantum criticality).
M14 hard blocker / failure modeExponential ED cost limits clusters to N_c ≲ 20 sites; inter-cluster MF becomes unreliable near quantum critical points where correlations are long-ranged; 3D is impractical (too many inequivalent bonds to tile)Not a thermodynamic-limit method for the cluster solver; accuracy plateau at small N_c if cluster geometry cannot capture the relevant correlations.

Cost & scaling

  • Time: O(d^{3N_c}) (full ED) or O(n_iter · N · d^{N_c}) (Lanczos) per cluster; inter-cluster steps negligible
  • Memory: O(d^{2N_c}) (full ED) or O(d^{N_c}) (Lanczos)
  • Control knob: cluster size N_c; systematic improvement, no error bar (deterministic)
  • Scale frontier: cluster N_c ≲ 20; lattice thermodynamic limit

Recommended for

  • 2D Hubbard model: Mott transition, pseudogap, antiferromagnetic and d-wave superconducting phases — primary target
  • Frustrated 2D magnets: kagome / triangular Heisenberg with sign-immune solver; resolves short-range order
  • Doped Mott insulators: t-J model, correlated electron systems at finite hole doping where DQMC has a sign problem
  • Spectral properties: spectral functions and quasiparticle weights inaccessible to plain MF or QMC (analytic continuation issues)
  • Per method-property-map.md: preferred when A1 = 2D, B7 Mott physics, C12 sign-ful (doped / frustrated), and spectral resolution is needed

Benchmarks

  • VCA d-wave order parameter in the 2D Hubbard model (U=8t, n=0.875): Δ_d ≈ 0.05–0.08t for N_c = 48 clusters, consistent with CDMFT and DCA maier_2004_quantum.
  • Mott transition U_c/t ≈ 6–8 (2D square lattice, half filling) converges with cluster size; N_c = 4 gives U_c ≈ 6.5t, approaching the DCA/CDMFT consensus maier_2004_quantum.
  • Cluster-size convergence verified in method-survey.md §7.4.

Key reference: maier_2004_quantum — comprehensive review of quantum cluster approaches (DCA, CDMFT, VCA) covering the theoretical foundations, self-energy functional theory, cluster-size convergence, and applications to correlated electron models; the authoritative all-details source.

Quantum-circuit simulationcircuit-simexactmethod-qcsmarkov_2008_simulating
AxisValueNote
M1 tasks / outputsSampling from the output distribution · expectation values of Pauli observables · VQE energy minimization · Hamiltonian-dynamics simulation · circuit fidelity estimationCost set by entanglement/depth generated and whether the circuit is near-Clifford.
M2 regimeT=0 / circuit-time (no thermal ensemble by default)Thermal states via imaginary-time circuits or purification; real-time dynamics natively.
M3 accuracy classState-vector: numerically exact, deterministic. TN: controlled bias (bond-dim truncation; a variational upper bound only for VQE energy minimization, not for dynamics/sampling). Stabilizer (Clifford): exact; stabilizer rank (T-gates): stochastic sampling.Engine choice sets the accuracy class.
M4 dimension fit (A1)Not lattice-dimension-gated; governed by circuit topology and entanglementMPS/TEBD efficient for 1D/brickwork circuits; 2D circuits need hyper-optimized TN slicing gray_2021_hyper.
M5 statistics & local dim (A3)Qubits (d=2) primarily; qudits via generalizationState-vector cost doubles per qubit; stabilizer cost scales as O(n²) per measurement.
M6 entanglement regime (B5)Entanglement / circuit structure is the spine. State-vector: volume-law tolerated. TN: cost = exp(treewidth) of contraction graph; polynomial for area-law / low-depth / 1D-like. Stabilizer: polynomial for near-Clifford (low T-gate count).Choice of engine must match the circuit's entanglement profile.
M7 sign-problem dependence (C12)Sign-immune (state-vector, TN, Clifford); stabilizer rank sampling can suffer phase cancellation for large T-countNo sign problem in state-vector or Clifford; near-Clifford regime is also stable.
M8 symmetry exploitation (C9/C10)Symmetry sectors reduce state-vector dim; TN block structure; Clifford symmetries reduce tableauTensorCircuit-NG and cotengra exploit circuit symmetry for hyper-optimized contraction.
M9 time complexityState-vector: O(gates·2ⁿ) time, O(2ⁿ) memory. TN: O(exp(treewidth)) (→ polynomial when B5 area-law/low-depth, Markov–Shi markov_2008_simulating). Stabilizer: O(n) per Clifford gate, O(n²) per measurement (Aaronson–Gottesman); exponential in #non-Clifford T-gates via stabilizer rank ~2^{αt}, α≈0.396 (Qassim–Pashayan–Gosset 2021).Use TN/stabilizer whenever applicable; state-vector is the fallback.
M10 memoryState-vector: O(2ⁿ). TN: varies with treewidth. Stabilizer: O(n²) (tableau).State-vector memory O(2ⁿ·16 bytes): ~100 GB at n≈33, ~16 TB at n≈40 (complex double).
M11 control knobTN bond dimension χ (controls truncation error) / stabilizer T-count t (controls approximation)State-vector: no approximation.
M12 scale frontierState-vector: n≈30–50 on HPC (memory-limited). TN: polynomial cost when treewidth is small; random 2D circuits beyond ~53 qubits require slicing + HPC. Stabilizer: Clifford circuits → any n; near-Clifford up to ~50 qubits with small T-count.Frontier set by engine-specific wall.
M13 primary approximation / biasState-vector: none. TN: MPS truncation (bond dim χ). Stabilizer rank: exponential approximation for large T-count.Clifford (T=0) is exact.
M14 hard blocker / failure modeState-vector: O(2ⁿ) memory wall. TN: high-depth / high-entanglement 2D circuits have large treewidth → exponential. Stabilizer: T-gate count t makes cost ~2^{0.396t} — non-Clifford universality is hard to simulate classically by design.No single engine covers all regimes.

Cost & scaling

  • Time: state-vector O(gates·2ⁿ); TN O(exp(treewidth)) (polynomial for low-depth / 1D); stabilizer O(n) per Clifford gate, O(n²) per measurement (Aaronson–Gottesman), + O(2^{αt}) in the number of non-Clifford T-gates, α≈0.396
  • Memory: State-vector O(2ⁿ); TN varies; stabilizer O(n²)
  • Control knob: TN bond dimension χ (truncation error) / T-gate count t (stabilizer rank)
  • Scale frontier: state-vector n≈30–50; TN polynomial for low-depth; stabilizer any n for Clifford

Recommended for

  • State-vector: small-n circuits (n ≲ 30) where exactness is required; debugging and validating quantum algorithms
  • TN / MPS: 1D brickwork circuits, shallow 2D circuits, VQE with low entanglement vidal_2003_slightly; hyper-optimized for random 2D circuits up to Google Sycamore scale gray_2021_hyper
  • Stabilizer: Clifford circuits of any size (QEC, stabilizer state prep); near-Clifford approximations when T-count is small
  • Per method-property-map.md: choose engine by matching circuit entanglement to the engine's efficiency regime

Benchmarks

  • State-vector frontier: n=53 Google Sycamore (2019); n≈50 routine on HPC clusters.
  • TN treewidth result: Markov–Shi (2008) proved cost = exp(treewidth) markov_2008_simulating; verified in method-survey.md §7.5.
  • Stabilizer rank exponent: α≈0.396 (Qassim–Pashayan–Gosset 2021) — verified in method-survey.md §7.5.
  • MPS VQE for 1D circuits: polynomial cost for area-law entanglement vidal_2003_slightly.

Key reference: markov_2008_simulating — proves that classical simulation cost equals exp(treewidth) of the circuit's contraction graph; the foundational complexity result for TN-based circuit simulation.

Certified bounds — SDP / NCTSSOSsdp-nctssoslowermethod-polyoptwang_2024_certifying
AxisValueNote
M1 tasks / outputsRigorous GS energy lower bound · two-sided energy bracket (with VMC upper bound) · certified bounds on local observables and correlations · Bell-inequality bounds · bootstrap phase-diagram constraintsPrimary niche: certifying that a VMC/DMRG energy is near-optimal.
M2 regimeT=0 ground state (GS energy certification)Finite-T moment relaxations possible but less developed.
M3 accuracy classCertified LOWER bound on the GS energy (complement to variational UPPER bound)Rigorous; tightness improves with relaxation level k.
M4 dimension fit (A1)1D and 2D feasible; 3D blows upMoment matrix O(N^k) grows too large in 3D at practically useful k.
M5 statistics & local dim (A3)Any algebra: Pauli (spin), fermion, bosonThe noncommutative polynomial framework handles any algebra.
M6 entanglement regime (B5)Sign-immune — not sampling-based; entanglement structure irrelevant to the SDPThe SDP sees only the algebraic constraints, not a wavefunction representation.
M7 sign-problem dependence (C12)Sign-immune — works in C12 sign-ful / B8 frustrated regimes (no sampling, no sign problem); key nicheThis is the critical advantage over QMC-based methods.
M8 symmetry exploitation (C9/C10)C9 symmetry block-diagonalizes the moment matrix → large speedup; translation symmetry (C10) further reduces blocksSymmetry is the primary route to tractability at useful k.
M9 time complexityRelaxation level k → moment-matrix size ~ #monomials of degree ≤k = O(N^k); SDP solve is high-polynomial in that matrix size (interior-point). Structured sparsity (NCTSSOS, term/correlative sparsity) + symmetry (C9) shrink the blocks.Tighter bound = higher k = rapidly growing SDP.
M10 memoryO(N^{2k}) (moment matrix + dual SDP variables)Dominant bottleneck; structured sparsity reduces effective block sizes.
M11 control knobRelaxation level k — higher k → tighter (larger) lower bound → larger SDPError: gap between lower and upper bound narrows with k.
M12 scale frontier~100 sites (1D) / ~10×10 (2D) at practically useful k with NCTSSOS sparsity + symmetryk=2 tractable for moderate N; k=3 requires structured sparsity.
M13 primary approximation / biasRelaxation level k — lower k gives looser (less tight) lower boundControlled: increasing k systematically tightens the bound.
M14 hard blocker / failure mode3D blows up (moment matrix too large); large k in 2D requires heavy HPC; gap between lower and upper bound may remain non-negligible at accessible kPair with VMC/DMRG for the upper bound to bracket the true GS energy.

Cost & scaling

  • Time: high-polynomial in moment-matrix size O(N^k) (interior-point SDP); structured sparsity (NCTSSOS) reduces effective blocks
  • Memory: O(N^{2k}) for the moment matrix; structured sparsity critical for large systems
  • Control knob: relaxation level k — controls bound tightness (systematic improvement)
  • Scale frontier: ~100 sites (1D) / ~10×10 (2D) at useful k with symmetry + NCTSSOS sparsity

Recommended for

  • Frustrated magnets (B8) and sign-ful (C12) models: the only method that certifies lower bounds without a sign problem (Heisenberg kagome, triangular, J₁-J₂ models)
  • Certification of variational results: pair with VMC/NQS or DMRG to bracket the GS energy
  • 1D/2D lattice models with symmetry: use C9 block-diagonalization to make k=2,3 tractable
  • Per method-property-map.md: niche method for sign-immune certification; complements all variational methods

Benchmarks

  • 1D Heisenberg chain: certified lower bound matches exact GS energy to <0.01% at k=2 with NCTSSOS sparsity wang_2024_certifying.
  • 2D 10×10 Heisenberg: tight two-sided bracket combining SDP lower + VMC upper bounds wang_2024_certifying.
  • Convergence of NPA hierarchy: guaranteed by Pironio–Navascués–Acín (2010) pironio_2010_convergent; verified in method-survey.md §7.6.

Key reference: wang_2024_certifying — demonstrates NCTSSOS-based certified lower bounds on GS energies of 1D and 2D Heisenberg models; the primary reference for the practical NCTSSOS + symmetry workflow.