Computational Many-Body QM

Many-body computation begins with a physical question and a representation that makes it tractable. Exact diagonalization, sampling, and tensor networks exploit different structures, so the appropriate method depends on the state, observable, size, and required accuracy.

This chapter develops many-body applications and comparisons. The Computational Many-Body Overview introduces the method families, and the Computational QM volume develops reusable algorithms and implementation details.

Required background. Scaling of Hilbert Space supplies the basis-growth problem. Math Needed for Computational QM supplies numerical linear algebra, floating-point, integration, and probability preparation.

Helpful background. Many-Body Hilbert Spaces and Operators prepares sectors, bases, and operator construction. Finite-Size Effects separates physical size mechanisms from numerical error.

Identify what limits the calculation

For $L$ spin-$1/2$ sites, an unrestricted state vector has $D=2^L$ complex amplitudes. At 16 bytes per complex double, one vector for $L=20$ uses 16 MiB, whereas a dense $D\times D$ matrix uses 16 TiB. Additional solver vectors need more memory. A local Hamiltonian can often be applied without storing the dense matrix, which is why sparse or matrix-free methods reach much larger problems.

State-space dimension is only one limitation. Monte Carlo may be limited by autocorrelation or cancellation of positive and negative weights. Tensor networks may be limited by entanglement or contraction cost. A real-frequency spectrum may be limited by time-window resolution or an ill-conditioned continuation even after the state is accurately computed.

Define the finite Hamiltonian, state or ensemble, symmetry sector, boundary conditions, and target observable first. Then estimate the relevant resources and choose an exact limit or small-system calculation to test the implementation. Increase numerical accuracy and physical size separately: a converged finite-chain result is not yet a thermodynamic result.

Keep error sources separate. Examples include a basis cutoff, eigenvalue residual, bond truncation, sampling uncertainty, finite-size correction, or artificial spectral broadening. Record the parameters, units, basis order, software version, and random seeds needed to repeat the calculation.

Choose a numerical method

Each branch uses a different structure in the problem.

1. Take the common first pass. Read Computational Many-Body Overview for representation choice, method families, error classes, validation, and finite-to-bulk inference.
2. Build a transparent finite baseline. Exact Diagonalization Preview constructs the basis and finite operator. Symmetry Sectors in Many-Body Numerics then reduces compatible blocks and explains what must still be reconstructed across sectors.
3. Target part of a sparse spectrum. After the finite operator is controlled, Lanczos Method Preview develops Krylov projection, Ritz values, residuals, reconstruction, and continued fractions. Symmetry reduction is useful but not universal.
4. Infer a large-system law only through a controlled family. Finite-Size Scaling in Numerics is a cross-cutting inference branch after any solver that supplies comparable sizes; it is a separate step from solver convergence.
5. Sample an equilibrium representation. Quantum Monte Carlo Preview develops configurations, estimators, equilibration, autocorrelation, and systematic limits. Pair it with Sign Problem Preview when weights cancel.
6. Compress low-entanglement structure. Enter the entanglement chapter’s conceptual tensor-network pages, then use Tensor Networks: Computational Guide for family selection and convergence design and DMRG Preview for finite-system MPS optimization.
7. Compute frequency-resolved observables. First declare the canonical correlator in the Correlations chapter. Then Dynamical Correlation Functions Numerically compares Lehmann sums, continued fractions, and time evolution with matched broadening and resolution.
8. Test and repeat the calculation. Benchmark Problems supplies small models with known answers. Reproducible Notebooks explains how to record the inputs, environment, outputs, and checks needed to repeat a result.

Example: the lowest gap of a spin chain

Suppose the target is the lowest gap of a periodic spin chain and its large-size behavior. First declare the Hamiltonian sign convention, chain lengths, periodic boundary, spin sector, momentum and parity labels, and whether the target is the lowest gap overall or within one sector. Exact diagonalization on the smallest chains tests basis construction, symmetry labels, degeneracies, and operator conventions. Sector-resolved Lanczos can extend the targeted eigenvalues, but a small Ritz residual supports a nearby eigenvalue—not every eigenvector-derived observable. Near degeneracy requires comparing the relevant subspace rather than one arbitrarily chosen eigenvector. The thermodynamic claim then requires a consistent sequence of sizes and sectors, checks of boundaries and commensurability, competing scaling forms, and overlap with exact limits or another method. Solver convergence and large-size inference remain separate arguments.

Continue from a computation to a physical conclusion

Finite-Size Effects explains physical size dependence. Phases, Order, and Criticality supplies phase diagnostics, and Nonequilibrium Dynamics explains relaxation and thermalization tests. Correlations and Response defines the exact observables before numerical approximation.

Use Computational QM for implementation, performance, and numerical analysis. For material-specific electronic-structure and response calculations, see Computational Quantum Matter.

Common pitfalls

“Exact diagonalization gives the exact physical answer.” It is exact only for the declared finite model, basis or cutoff, sector coverage, arithmetic, and tolerance. The thermodynamic or continuum inference is separate.

“A tiny residual certifies the eigenvector and every observable.” A residual strongly constrains a nearby eigenvalue; individual eigenvectors require spectral isolation, while near degeneracy calls for an invariant-subspace or projector statement.

“A converged variational energy certifies the phase.” Energy convergence does not by itself certify fidelity, long-distance correlations, topology, excitation structure, or freedom from metastability.

“The Monte Carlo error bar is the total uncertainty.” It normally measures sampling uncertainty after an autocorrelation analysis. Equilibration, discretization, estimator, finite-size, continuation, and sign-related biases remain separate.

“The sign problem is an intrinsic label on a Hamiltonian.” Its severity depends on basis, representation, decomposition, parameters, and algorithm. A sign-free formulation in one setting is not a universal cure.

“A discarded weight or bond dimension is an error bar.” It is a convergence diagnostic. Observable errors, finite-size effects, initialization, symmetry enforcement, and entanglement structure still require direct tests.

“One successful fit proves the thermodynamic law.” Size range, boundaries, corrections, covariance, competing forms, and stability under omitted sizes must be checked.

“Numerical broadening is a physical linewidth.” Finite time, recurrence, kernel choice, level spacing, truncation, and inverse reconstruction can create width without a physical decay rate.

“Having the notebook means the result can be reproduced.” The calculation also needs its data, environment, parameters, execution steps, and numerical checks. Verify that it runs successfully and reproduces the stated observables.

Exercises

Exercise 1: Route three targets for one Hamiltonian

For a spin-1/2 chain, route (a) the complete spectrum at ten sites, (b) the ground-state energy density at much larger size in a favorable one-dimensional regime, and (c) an equal-time finite-temperature structure factor when the sampling weights are nonnegative. State one mandatory validation test for each route.

Solution

For (a), use Exact Diagonalization with Symmetry Sectors and validate dimension, Hermiticity, trace or moment identities, and reconstruction across sectors. For (b), use the Tensor-Networks guide and DMRG, then a size and bond-dimension convergence study; compare the smallest systems with ED. For (c), use Quantum Monte Carlo, verify equilibration and integrated autocorrelation, test the estimator on a benchmark, and add finite-size scaling for a bulk claim. The same Hamiltonian does not imply the same method because the state, observable, size, and error structure differ.

Exercise 2: Repair an overclaim

Repair this statement: “The Lanczos residual is below $10^{-12}$, the DMRG discarded weight is small, and the Monte Carlo standard error is tiny, so all three methods prove the same thermodynamic phase.”

Solution

The three quantities diagnose different numerical layers. The Lanczos residual supports a finite eigenvalue or invariant-subspace claim, subject to sector and spectral-isolation checks. Discarded weight supports a bond-truncation convergence study but does not alone bound every observable or exclude metastability. A Monte Carlo standard error addresses correlated sampling after equilibration, not all biases. A phase claim additionally needs a common observable definition, cross-method agreement on overlapping finite systems, controlled size and boundary sequences, a phase diagnostic, uncertainty propagation, competing-explanation tests, and the declared thermodynamic limit.

References

• H. Fehske, R. Schneider, and A. Weiße, eds., Computational Many-Particle Physics, Springer (2008).
• M. E. J. Newman and G. T. Barkema, Monte Carlo Methods in Statistical Physics, Oxford University Press (1999).
• Y. Saad, Numerical Methods for Large Eigenvalue Problems, 2nd ed., SIAM (2011).
• A. W. Sandvik, “Computational Studies of Quantum Spin Systems,” in AIP Conference Proceedings 1297, 135–338 (2010).
• U. Schollwöck, “The Density-Matrix Renormalization Group in the Age of Matrix Product States,” Annals of Physics 326, 96–192 (2011).