---
title: Lanczos Coefficients in Quantum Dynamics
url: https://www.emergentmind.com/topics/lanczos-coefficients
type: topic
---

# Lanczos Coefficients in Quantum Dynamics

Lanczos coefficients are the recurrence parameters appearing in the three-term tridiagonalization of an operator—typically a matrix, Liouvillian, or moment-permuted Hamiltonian—in a Krylov (orthogonalized) basis. They underlie the continued-fraction expansion of Green’s functions, form the bridge to orthogonal polynomial theory, and encode key features of spectral densities, correlation functions, and operator complexity in quantum systems.

## 1. Definition and Origin of Lanczos Coefficients

The classical Lanczos algorithm constructs an orthonormal basis in a Krylov subspace by iterated application of an operator. For a Hermitian matrix \(A\), starting from a normalized vector \(v_0\), the recursion is
\[
A v_j = \beta_{j-1} v_{j-1} + \alpha_j v_j + \beta_j v_{j+1},
\]
with \(\alpha_j = \langle v_j, A v_j \rangle\) and \(\beta_j = \|A v_j - \alpha_j v_j - \beta_{j-1} v_{j-1}\|\).
Analogously, for a Liouvillian \(\mathcal{L}\) acting on operators \(\mathcal{O}_n\), the recursion is
\[
b_{n+1} \mathcal{O}_{n+1} = \mathcal{L} \mathcal{O}_n - b_n \mathcal{O}_{n-1}.
\]
Here, the \(\alpha_j\) and \(\beta_j\) (or, more commonly, the single off-diagonal sequence \(b_n\)) are the Lanczos (or Jacobi) coefficients. For orthogonal polynomials \(p_n(x)\) with respect to measure \(\mu\), similar recursions hold:
\[
x p_n(x) = \beta_{n-1} p_{n-1}(x) + \alpha_n p_n(x) + \beta_n p_{n+1}(x),
\]
and explicit formulas are available for classical weights, but generic measures require numerical schemes [2101.11963].

## 2. Structural Properties and Computational Algorithms

The Lanczos coefficients reflect the measure or operator’s moments and can be computed either explicitly (for classical cases) or via iterative/numerical techniques. Stabilized versions of the Lanczos algorithm with reorthogonalization are standard for discrete measures or finite matrices, while continuous or hybrid measures use predictor-corrector or Stieltjes-type approaches. For general measures,
- **Purely discrete:** Stabilized Lanczos iteration [Step 4 and 5 in 2101.11963].
- **Purely continuous:** Predictor–corrector algorithm.
- **Mixed measure:** Predictor–corrector–Lanczos hybrid (PCL).

Orthogonality defects and successive changes in \(\beta_n\) are used as convergence diagnostics. In quantum computing, block Lanczos algorithms extend the method to resolve degeneracies and treat non-Hermitian evolutions by bi-orthogonal bases [2109.14114, 2303.04175].

## 3. Connection to Continued Fractions and Functional Representation

Lanczos coefficients parameterize continued-fraction representations of Green’s functions and autocorrelation functions. The retarded Green’s function for operator \(O\) and Liouvillian \(\mathcal{L}\) is expressed as [2505.00089]:
\[
G(z) = \left\langle O, (z - \mathcal{L})^{-1} O \right\rangle = \cfrac{1}{z - \cfrac{b_1^2}{z - \cfrac{b_2^2}{z - \ddots}}},
\]
which recasts the entire spectral problem in terms of a one-dimensional tridiagonal system governed by \(\{b_n\}\). This enables highly efficient computation and asymptotic analysis of global spectral features, spectral densities, and moments [2505.00089, 2512.15857].

Such continued fractions also underlie recursion methods for quantum autocorrelation functions, where the Laplace or Fourier transforms are analytic functions determined by the sequence of \(b_n\). Corrections, truncations, and “stitching” approximations quantify finite-size errors or incomplete knowledge of the underlying coefficients [2505.00089, 2503.17555].

## 4. Physical Interpretation and Asymptotic Growth

The growth of Lanczos coefficients encodes deep physical and statistical properties:
- **Quantum Many-body Systems:** In chaotic, non-integrable models, \(b_n\) typically exhibits asymptotically linear growth, \(b_n \sim \alpha n\), with possible sublinear corrections (\(b_n \sim n/\ln n\) in one dimension) [2412.15932, 2505.00089].
- **Operator Growth Hypothesis (OGH):** Linear or nearly linear growth of \(b_n\) is conjectured as a hallmark of quantum chaos in the thermodynamic limit.
- **Free (quadratic) Models:** The Anderson impurity model demonstrates that diverse asymptotic forms of \(b_n\) (constant, \(\sqrt{n}\), linear) can all arise in an exactly solvable, integrable context [2601.13255].
- **Finite-size Scaling:** For finite systems, the late-time scaling and fluctuations of \(b_n\) (or the ratios \(b_{n+1}/b_n\)) control hydrodynamical relaxation, zero-mode behavior, and saturation plateaux of autocorrelation functions. Specific scaling conjectures relate these ratios to system size and the presence or absence of conserved quantities [2507.17424].
- **Open Systems:** In Lindbladian evolution, bi-Lanczos coefficients (off-diagonals and pure imaginary diagonals) reflect both operator scrambling and dissipation. Early-time behavior may distinguish chaos and integrability, but late-time fluctuations induced by dissipation universalize the saturation of K-complexity [2303.04175].

## 5. Error Analysis, Approximations, and Numerical Significance

Various truncation and approximation schemes rely on the rapid convergence properties of continued-fraction expansions with smoothly-growing \(b_n\):
- **Stitching Approximation:** Completing the continued fraction beyond a finite set of known \(b_n\) with an asymptotically-matching tail yields robust approximations. The rate of convergence depends on the decay of staggered subleading terms in the sequence; error terms scale as \(1/\mathrm{poly}(N)\) (best case, e.g., purely smooth spectral densities) to \(1/\mathrm{poly}(\log N)\) when subleading staggering decays only logarithmically [2505.00089].
- **Lanczos-Pascal Truncation:** For chaotic systems with smooth, near-linear \(b_n\), as few as 10–30 coefficients suffice to recover temporal autocorrelators and spectral areas to high precision; the method reduces the problem to a small set of damped oscillations [2503.17555].

In finite-dimensional settings, strong concentration theorems show that \(k = O(\log n)\) Lanczos steps yield coefficients overwhelmingly close to deterministic medians for “bulk” spectral features, with \(k = O(\sqrt{\log n})\) controlling convergence to limiting Jacobi parameters [1904.06012].

## 6. Mapping to Orthogonal Polynomials and Random Matrix Theory

In random matrix theory and the theory of orthogonal polynomials, the Lanczos coefficients are precisely the recurrence coefficients \(R_n\) and \(S_n\) (for off-diagonals and diagonals, respectively). In the large-\(N\) continuum limit, one finds the mapping
\[
\sqrt{R(x)} = b(1-x),\quad S(x) = a(1-x)
\]
where \(x = n/N\). The density of states, moments, and Krylov dynamics computed via the Jacobi (tridiagonal) matrix directly coincide whether generated from the Lanczos approach or from orthogonal polynomial theory [2512.15857].

Explicit models, for example the Gaussian Unitary Ensemble, yield algebraic forms for \(\{b_n\}\), the Hermite polynomial recurrence coefficients, and the Wigner semicircle law. Krylov state amplitudes, operator spread, and survival probabilities can all be expressed and analyzed in this unified framework.

## 7. Applications and Limitations

Lanczos coefficients underlie efficient algorithms for:
- Spectral density estimation and continued-fraction representations of Green’s functions
- Quantum many-body correlation functions, including autocorrelators at infinite temperature
- Extraction of hydrodynamical transport coefficients (e.g., diffusion constants via infinite product formulas over \(\{b_n\}\)) [2505.00089]
- Efficient time-evolution strategies for quantum systems, both in classical and quantum computing contexts (block Lanczos methods for excited states and non-Hermitian evolution) [2109.14114]
- Assessment of chaos, integrability, and hydrodynamic scales in both closed and open (dissipative) quantum systems [2601.13255, 2303.04175, 2507.17424]

However, the mere growth rate of \(\{b_n\}\) cannot distinguish chaos from integrability in general: quadratic models (non-interacting, integrable) can be tuned to display constant, \(\sqrt{n}\), or even linear growth in \(\{b_n\}\), and the Markovian limit in such cases always yields simple exponential decay of correlators regardless of \(\{b_n\}\) structure [2601.13255]. In open systems, dissipation-induced fluctuations in the Krylov chain wash out initial chaos/integrability signatures at late times [2303.04175].

Lanczos coefficients thus serve as a unifying language and computational toolset across numerical linear algebra, quantum dynamics, random matrix theory, orthogonal polynomials, and statistical physics, but with important model-specific caveats for physical interpretation.

Source: https://www.emergentmind.com/topics/lanczos-coefficients