A finite-precision Lanczos-Golub-Welsch route to probability-table construction in resonance self-shielding

This paper proposes a finite-precision Lanczos-Golub-Welsch algorithm for constructing probability tables in resonance self-shielding that reformulates Chiba's method as a polynomial-moment problem, yielding lower effective-cross-section errors and avoiding complex responses compared to conventional moment-Pade approaches.

The Gist

Imagine you are trying to predict how a crowd of people (neutrons) will behave in a crowded room (a nuclear reactor). The room has specific "danger zones" (resonances) where people get stuck or absorbed very quickly. To simulate this on a computer, you can't track every single person in real-time; it would take too long. Instead, you need to create a simplified "probability map" that says, "In this zone, 30% of people act like this, and 70% act like that."

This paper is about building a better, more reliable way to create that map.

The Problem: The "Fragile Blueprint"

The old way of making these maps (called the Moment-Padé method) is like trying to rebuild a complex 3D sculpture just by looking at its shadow on the wall.

1. The Shadow (Moments): You measure the "shadows" (mathematical averages) of the data.
2. The Reconstruction: You try to reverse-engineer the sculpture from those shadows.

The Flaw: Computers aren't perfect; they have tiny rounding errors (like a carpenter who is off by a fraction of a millimeter). When you try to reverse-engineer the sculpture from the shadows using the old method, those tiny errors get blown up.

• The Result: Instead of a solid sculpture, you get a mess. The computer might say, "Hey, 5% of the people in this zone have negative weight," or "This person is made of imaginary numbers." In physics, negative or imaginary probabilities make no sense. It's like a map telling you to drive backward into a wall. As you try to make the map more detailed (higher order), the old method tends to collapse into this nonsense.

The Solution: The "Lanczos-Golub-Welsch" Route

The author, Beichen Zheng, proposes a new way to build the map. Instead of trying to reverse-engineer the sculpture from its shadow, they build it directly from the raw material, using a very sturdy, step-by-step process.

Think of it like this:

1. The Raw Material (Positive Measure): Instead of starting with abstract shadows, we start with the actual "stuff" (the positive distribution of data). We know this stuff is real and positive (you can't have negative mass).
2. The Compression (Lanczos Reduction): Imagine you have a giant, messy pile of sand (the detailed data). You want to compress it into a small, neat bag (the probability table) that still holds the same "shape" and weight. The new method uses a mathematical tool called Lanczos to gently squeeze the sand.
   • The Analogy: It's like using a high-quality vacuum sealer that preserves the shape of the food inside, rather than a cheap press that squishes it into a weird, broken shape.
3. The Extraction (Golub-Welsch): Once the sand is compressed, we use a special key (Golub-Welsch) to pull out the exact locations and weights of the "grains" we need for our map.

Why is this better?
Because this method respects the "rules of the game" (mathematical positivity) at every single step. It's built on a foundation that guarantees the numbers stay real and positive, just like the original data. Even if the computer makes tiny mistakes, the structure is so robust that the final map doesn't break.

The Results: A Sturdier Map

The author tested this new method against the old one using real nuclear data (like Uranium-238).

• The Old Method: As they tried to make the map more precise, it started producing "ghosts" (complex numbers) and "negative people" (negative probabilities). It was like a bridge that looked great in the blueprint but collapsed when you added more weight.
• The New Method: It produced maps that were not only more accurate but also physically sensible. No ghosts, no negative numbers. Even when they pushed the math to its limits, the new method stayed stable.

The Big Picture

This paper is essentially saying: "Stop trying to guess the shape of the sculpture from its shadow. Let's build the sculpture directly using a method that guarantees it stays solid."

By changing the mathematical "recipe" for creating these nuclear safety maps, the author has provided a tool that is less likely to crash and produce impossible results, making nuclear reactor simulations safer and more reliable. It's a shift from a fragile, high-maintenance approach to a sturdy, "set-it-and-forget-it" approach that works even when the computer isn't perfect.

Technical Summary

1. Problem Statement

In multigroup reactor physics, resonance self-shielding causes neutron cross-sections to vary sharply within energy groups. To handle this computationally, the subgroup method replaces continuous within-group variations with a discrete set of representative levels (a probability table). This involves constructing a low-cardinality discrete approximation of a positive measure (representing the cross-section distribution) that preserves:

1. Accuracy: The ability to reproduce effective cross-sections over a range of background cross-sections (dilution).
2. Admissibility: The preservation of nonnegative-real properties for subgroup levels, probabilities, and effective cross-sections.

The Core Challenge:
The conventional approach (Ribon/Chiba) uses a moment–Padé pipeline:

1. Compute moments of the cross-section distribution.
2. Solve an ill-conditioned Hankel system to find Padé denominator coefficients.
3. Extract roots (subgroup levels) and residues (probabilities).
4. Solve a Vandermonde system for reaction-channel levels.

In finite-precision arithmetic, this pipeline is numerically fragile. The Hankel systems are often ill-conditioned, and root extraction is sensitive to coefficient perturbations. This leads to order-induced breakdown, where increasing the number of subgroups ($N$) causes the emergence of complex or negative subgroup levels and nonphysical effective cross-sections, rendering the solution unusable.

2. Methodology

The author proposes an alternative construction route based on positive-measure compression using the Lanczos–Golub–Welsch algorithm.

A. Theoretical Reformulation

• Transformed Measure: The paper reformulates Chiba's physically motivated "affine-order" moments as standard polynomial moments of a transformed positive measure $\mu_g$.
• Transformation: By defining a variable $z = \sigma_t^b$ (where $\sigma_t$ is the total cross-section and $b$ is a scaling factor), the affine moments become polynomial moments of a new measure. This allows the problem to be treated as a standard Gauss quadrature problem for a positive measure.

B. The Construction Route

Instead of explicit moment inversion, the proposed method proceeds in three stages:

1. Discrete Measure Realization: The continuous induced measure is approximated by a high-resolution discrete measure ($\mu_M$) on a union grid using trapezoidal or piecewise-linear integration. This step ensures the input is strictly positive.
2. Symmetric Lanczos Reduction: The Lanczos algorithm is applied to the diagonal matrix of the discrete measure and a starting vector. This generates a symmetric tridiagonal Jacobi matrix ($J_N$) that preserves the moment information of the measure up to order $2N-1$.
3. Golub–Welsch Extraction:
   • Total Levels & Probabilities: The eigenvalues of $J_N$ provide the compressed nodes (transformed levels), and the squared first components of the eigenvectors provide the weights (probabilities). This step is structure-preserving: in exact arithmetic, it guarantees real, nonnegative nodes and weights.
   • Reaction-Channel Levels: Instead of solving a Vandermonde system, reaction-channel levels are recovered by matching coefficients in an orthogonal basis (Lanczos vectors). This avoids the ill-conditioning associated with Vandermonde matrices.

C. Finite-Precision Stabilization

The paper investigates reorthogonalization strategies for the Lanczos process to prevent loss of orthogonality (which causes spurious eigenvalues):

• Full Reorthogonalization: Most stable but computationally expensive.
• Selective Reorthogonalization: A hybrid approach that locks converged Ritz vectors and reorthogonalizes only when semi-orthogonality is lost.
• Strategy: The paper recommends full reorthogonalization for low orders ($N=9$) and selective reorthogonalization for higher orders ($N \ge 30$) with large data sets ($M \ge 5000$).

3. Key Contributions

1. Reformulation: Recasting the probability-table construction as a structured positive-measure compression problem via a transformed variable, bypassing the need for explicit affine-moment inversion.
2. Algorithmic Shift: Replacing the unstable moment–Padé pipeline with a stable Lanczos–Golub–Welsch route that utilizes symmetric tridiagonal reduction and orthogonal-basis matching.
3. Error Decomposition: Distinguishing between realization error (discretization of the input measure) and compression error (reduction to $N$ points). The study identifies a transition point where compression error becomes negligible compared to realization error.
4. Numerical Robustness: Demonstrating that the new route maintains nonnegative-real properties even at high orders where the conventional method fails completely.

4. Results

The method was tested on five resonance-channel cases (including $^{238}\text{U}$ capture, $^{235}\text{U}$ fission, etc.) using both fine and coarse group structures.

• Accuracy: The proposed method yields lower effective-cross-section errors than the conventional method across all tested orders ($N=9, 30, 50$).
• Nonnegative-Real Preservation:
  • Conventional Method: Exhibits rapid order-induced breakdown. At $N=12$, it produces complex or negative subgroup levels and nonphysical effective cross-sections in all tested groups.
  • Proposed Method: Maintains nonnegative-real subgroup levels and probabilities across the entire tested range ($N$ up to 50). The effective cross-sections remain physically admissible.
• Order Scaling:
  • At low orders, the error is dominated by compression.
  • At high orders ($N > 26$ for analytic realization), the error is dominated by the initial measure realization. The proposed method reaches this "realization floor" without the numerical instability that plagues the conventional method.
• Reorthogonalization: Selective reorthogonalization offers a good trade-off between computational cost and orthogonality preservation for large datasets, though full reorthogonalization provides the most robust internal stability.

5. Significance

This work addresses a critical bottleneck in nuclear data processing: the numerical instability of high-order probability-table generation.

• Practical Impact: It enables the use of higher-order subgroup tables (which offer better accuracy) without the risk of generating nonphysical results that would crash reactor simulations.
• Theoretical Insight: It demonstrates that the instability of the conventional method is not inherent to the approximation problem itself, but rather an artifact of the moment–Padé reconstruction pipeline in finite precision.
• Route-Level Advantage: The paper establishes that a route-level change (from moment inversion to spectral extraction) provides a distinct finite-precision advantage, ensuring that the structural properties of the underlying positive measure (nonnegativity) are preserved throughout the computation.

In summary, the proposed Lanczos–Golub–Welsch route offers a numerically robust, accurate, and physically admissible alternative to the classical moment–Padé method for resonance self-shielding calculations.