Approximation theory for Green's functions via the Lanczos algorithm

This paper develops a theoretical framework for the error convergence of the stitching approximation in Green's function calculations via the Lanczos algorithm, demonstrating that the convergence rate depends on the decay of subleading Lanczos coefficients and the smoothness of the spectral function, while also deriving a formula linking the spectral function at the origin to continued fraction coefficients to estimate the diffusion constant in the mixed-field Ising model.

The Gist

Imagine you are trying to predict the weather for a city that is infinitely large. You have a super-complex set of rules (the laws of physics) that tell you how the wind and rain interact. If you try to calculate the weather for every single molecule in the city, your computer would explode because there are simply too many variables.

This paper is about a clever shortcut scientists use to solve these "infinitely complex" problems without needing a supercomputer that doesn't exist yet. Here is the breakdown using everyday analogies.

1. The Problem: The Infinite Recipe

In quantum physics, scientists want to know how energy or information moves through a system (like heat spreading through a metal). To do this, they use a mathematical tool called a Green's function. Think of this function as a "recipe" that tells you exactly how the system behaves.

However, writing down this recipe perfectly requires an infinite list of numbers (called Lanczos coefficients). It's like trying to write down the exact value of $\pi$ (3.14159...) by listing every single digit. You can't do it because the list never ends.

2. The Shortcut: The "Stitching" Method

Since we can't calculate the infinite list, we calculate the first $N$ numbers (the first few digits of $\pi$) and then stop. But if we just stop there, our prediction is terrible. It's like guessing the rest of a story just by cutting off the book; the ending won't make sense.

The authors focus on a method called "Stitching" (also known as the recursion method).

• The Analogy: Imagine you are building a long bridge. You build the first 100 meters perfectly using precise measurements. For the rest of the bridge (the infinite part), instead of guessing randomly, you attach a pre-fabricated section that you know works perfectly.
• The Science: They take the exact numbers they calculated, and then they "stitch" them to a known, perfect mathematical pattern (called Meixner-Pollaczek polynomials) that mimics how the numbers should behave in the long run.

3. The Big Question: How Good is the Stitch?

The paper asks: How close is our "stitched" bridge to the real, perfect bridge?

If you only stitch a few meters, the error is huge. If you stitch a million meters, the error is tiny. But the authors wanted to know: How fast does the error disappear as we add more perfect numbers?

They found that the speed of this improvement depends on a hidden "glitch" in the numbers, which they call staggered terms.

• The Analogy: Imagine the bridge has a slight, rhythmic wobble (a zig-zag pattern) in its design.
  • If the wobble is strong and slow (it doesn't fade away quickly), the bridge remains shaky no matter how much you extend it. The error drops very slowly.
  • If the wobble is weak and fades away fast, the bridge becomes smooth very quickly. The error drops fast.

4. The Connection to "Smoothness"

The paper makes a fascinating connection between this "wobble" in the numbers and the "smoothness" of the physical system.

• The Analogy: Think of a smooth road versus a bumpy one.
  • If the road is very smooth (the physics is very regular), the "wobble" in the numbers fades away quickly, and our shortcut works great.
  • If the road is bumpy or has a sudden sharp turn (a "singularity" in the math), the "wobble" in the numbers is stubborn. It takes a lot of extra work to get a good answer.

The authors prove that if the physical system has a "kink" (is not perfectly smooth) at a specific point, the error in our calculation will decrease very slowly—so slowly that to get a precise answer, you might need to calculate an exponentially huge number of steps.

5. Real-World Application: The Diffusion Constant

The authors tested this theory on a specific problem: calculating the diffusion constant (how fast heat or particles spread out) in a chaotic quantum system (the Ising model).

• They used their "stitching" method to estimate this value.
• They compared their result to previous, more complicated calculations.
• The Result: Their simple "stitching" method gave the same answer as the complex methods, confirming their theory works.

Summary

• The Goal: Predict how quantum systems behave without doing impossible amounts of math.
• The Method: Calculate a few steps perfectly, then "stitch" them to a known perfect pattern.
• The Discovery: The accuracy of this method depends on a hidden "wobble" in the numbers.
• The Catch: If the physical system is "bumpy" (mathematically rough), that wobble is stubborn, and you need a massive amount of computing power to get a precise answer. If the system is "smooth," the method is very efficient.

Essentially, the paper provides a rulebook for scientists to know: "If your system looks like this, you need to calculate X steps. If it looks like that, you need to calculate a billion steps." This helps them decide if a calculation is even worth trying.

Technical Summary

Technical Summary: Approximation Theory for Green's Functions via the Lanczos Algorithm

Problem Statement
The paper addresses the computational challenge of approximating Green's functions in many-body quantum systems, specifically within the context of thermalization and hydrodynamic transport. While the Lanczos algorithm (or recurrence method) allows for the representation of Green's functions as continued fractions via Lanczos coefficients ($b_n$), practical applications are limited by the exponential growth of memory requirements with the number of steps. Consequently, only a finite number of coefficients ($N$) can be computed numerically. The central problem is determining how to best approximate the infinite continued fraction using only these first $N$ coefficients and, crucially, quantifying the rate at which the error of this approximation decays as $N$ increases. This is particularly critical for extracting hydrodynamic quantities, such as diffusion constants, which depend on the zero-frequency limit of the Green's function, a regime where standard error bounds often fail.

Methodology
The authors develop a theoretical framework for the "stitching" approximation (also known as the recursion method). In this approach, the exact Lanczos coefficients $\{b_n\}_{n=1}^N$ are computed, and the remaining infinite tail is replaced by a "stitched" sequence $\{b_{n,s}\}_{n>N}$ for which the Green's function is known analytically.

Key methodological steps include:

1. Formalism: The Green's function $G(z)$ is expressed as a composition of Möbius transformations acting on a level-$N$ Green's function. The error is formulated as the difference between the exact and stitched resolvents.
2. Assumptions: The analysis relies on two primary assumptions:
   • The Operator Growth Hypothesis (OGH), which posits that in chaotic systems, $b_n$ grows linearly (for $d>1$) or as $n/\log n$ (for $d=1$) asymptotically.
   • Assumption 1: The stitching approximation converges to the correct answer in the zero-frequency limit ($z \to -i0^+$), allowing the interchange of limits $N \to \infty$ and $z \to -i0^+$.
3. Error Analysis: The authors derive upper bounds for the error $e_{N,s}(z)$. While a standard bound exists for $\text{Im}(z) \neq 0$, the authors derive a new, improved bound valid along the imaginary axis (including $z=0$). This bound depends on the sum of the differences between exact and stitched coefficients, weighted by the inverse of the coefficients.
4. Spectral Smoothness Connection: The paper investigates the relationship between the differentiability of the spectral function $\rho(x)$ at the origin and the subleading "staggered" terms (oscillatory corrections) in the Lanczos coefficients. Using a theorem by Magnus and asymptotic analysis, they link the decay rate of these staggered terms to the smoothness of the Green's function.

Key Contributions and Results

• Convergence Rates and Staggering: The paper demonstrates that the convergence rate of the stitching approximation is not solely determined by the leading asymptotic growth of $b_n$, but is strongly governed by the decay of staggered subleading terms (terms of the form $(-1)^n s_n$).
  • Relevant Staggering: If the staggered term decays slowly (e.g., power-law slower than $1/n$), the convergence is slow, and advanced stitching methods (like Meixner-Pollaczek) offer no advantage over simple "constant stitching" (where $b_{n,s} = b_N$ for $n>N$).
  • Irrelevant Staggering: If the staggered term decays rapidly (faster than $1/n$), Meixner-Pollaczek stitching yields a significantly faster convergence rate than constant stitching.
• Zero-Frequency Limit and Diffusion Constants: The authors derive a formula linking the spectral function at the origin, $\rho(0)$, to an infinite product of Lanczos coefficients (Eq. 23). This allows for the direct estimation of diffusion constants ($D \propto \rho(0)$) without requiring the full spectral function.
• Dimensional Dependence:
  • For $d > 1$, where $b_n \sim \alpha n$, Meixner-Pollaczek polynomials provide an exact solution for the tail. The convergence rate depends on whether the staggering is relevant or irrelevant.
  • For $d = 1$, where $b_n \sim \alpha n / \log n$, no exact solution with the correct asymptotic growth is known. The authors propose using "constant stitching" and show that in the best case, convergence scales as $N^{-1}$, but can be arbitrarily slow if staggered terms decay slowly.
• Complexity Estimates: The paper establishes a link between the smoothness of the spectral function and computational cost. If the spectral function has a finite number of derivatives at the origin (a common scenario in hydrodynamics due to long-time tails), the staggered terms decay as a power of $\log n$. This implies that the error scales as $1/\text{poly}(\log N)$. Consequently, to achieve a precision $\epsilon$, the number of Lanczos steps required scales as $N \sim \exp(\text{poly}(1/\epsilon))$.
• Numerical Validation: The theory is validated using:
  • Toy models with known solutions for both irrelevant and relevant staggering cases, confirming the predicted error scaling ($N^{-2}$ vs. $N^{-2/3}$).
  • The mixed-field Ising model in 1D ($d=1$), where the diffusion constant is estimated using the infinite product formula. The results agree with previous, more complex extrapolation methods, despite the current method being simpler.

Significance
The paper provides a rigorous error analysis for the Lanczos-based approximation of Green's functions, moving beyond simple truncation to a systematic "stitching" approach. Its primary significance lies in identifying that the smoothness of the spectral function at zero frequency dictates the convergence speed of the algorithm. By linking the differentiability of the Green's function to the decay of staggered Lanczos coefficients, the authors explain why extracting transport coefficients in chaotic many-body systems is computationally demanding. The work suggests that for systems with hydrodynamic tails (non-analytic spectral functions), the Lanczos method may require exponentially many steps in the inverse of the desired error to achieve high precision, highlighting inherent limitations in simulating long-time hydrodynamics via this method. The derived infinite product formula offers a practical tool for estimating diffusion constants, validated against known results in the Ising model.