📊 statistics

Classification of real hyperplane singularities by real log canonical thresholds

This paper establishes explicit combinatorial formulas and a general algebraic theory for the real log canonical threshold and its multiplicity of real hyperplane arrangements, supported by a SageMath implementation and applications to statistical model analysis and high-dimensional volume integrals.

Original authors: Dimitra Kosta, Daniel Windisch

Published 2026-01-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dimitra Kosta, Daniel Windisch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to understand how "messy" or "complex" a specific shape is. In mathematics, these shapes are called singularities. Think of a singularity like a sharp corner on a piece of paper, a point where a surface folds over itself, or a place where several lines cross perfectly. The messier the point, the harder it is to do math around it.

This paper is about measuring that "messiness" for a specific type of shape: hyperplane arrangements. If you imagine a 3D room, a hyperplane is just a flat wall. An arrangement is a collection of these walls (which might cross each other, overlap, or be stacked). The authors are interested in the points where these walls meet.

Here is the breakdown of their work using simple analogies:

1. The Two Rulers: Complex vs. Real

Mathematicians usually have two different rulers to measure this messiness:

The Complex Ruler (lct): This ruler looks at the shape as if it exists in a magical, multi-dimensional world where numbers can be imaginary (like $\sqrt{-1}$ ). It's very powerful but abstract.
The Real Ruler (rlct): This ruler looks at the shape only in our "real" world, where numbers are the ones we use for counting and measuring.

For a long time, mathematicians thought these two rulers might give different answers for the same shape. This paper proves that for flat walls (hyperplanes), the two rulers actually agree. If you measure the messiness of a crossing of flat walls using the complex ruler, you get the exact same number as if you used the real ruler. This is a big deal because the "Real Ruler" is the one that matters for real-world applications.

2. Why Do We Care? (The "Learning Coefficient")

The authors mention that this "Real Messiness Score" (called the Real Log Canonical Threshold or rlct) is crucial for Machine Learning and Statistics.

Imagine you are training a robot to recognize cats. You have to choose a model (a set of rules) for the robot to learn.

Simple models are like a straight line; they are easy to understand but might miss details.
Complex models are like a tangled knot of wires; they can learn very complex patterns but might get confused (overfit).

There is a famous formula (BIC) that helps humans decide which model is best. However, this formula breaks down when the model is a "tangled knot" (a singular model). The authors show that to fix the formula for these tangled knots, you need to plug in the Real Messiness Score.

The Score ( $\lambda$ ): Tells you how "heavy" the complexity is.
The Multiplicity ( $m$ ): Tells you how many different ways the model can be complex at that specific point.

If you get these numbers right, you can accurately predict how well your robot will learn and how much data it needs.

3. The "Building Set" Recipe

Before this paper, if you wanted to calculate this score for a specific arrangement of walls, you had to do it case-by-case, like solving a unique puzzle every time.

The authors created a universal recipe (a combinatorial formula).

The Ingredients: You just need to know the geometry of the walls (where they cross) and how many times each wall is counted (some walls might be "double" or "triple" layers).
The Method: They use a concept called a "building set." Imagine you are building a tower out of blocks. You look at every possible way the walls intersect (a single wall, two walls crossing, three walls meeting at a point).
The Calculation: For every intersection, you calculate a simple ratio: How many dimensions does this intersection lose? divided by How many wall-layers are there?
The Result: The lowest ratio you find is your "Messiness Score" ( $\lambda$ ). The longest chain of intersections that all share this lowest score gives you the "Multiplicity" ( $m$ ).

4. The Computer Tool

The authors didn't just write the math; they built a calculator (a SageMath program).

You can feed it a list of equations for your walls.
It instantly crunches the numbers to tell you the Messiness Score and the Multiplicity.
They tested it and found it is much faster than existing tools, capable of handling complex arrangements of up to 15 walls in a few seconds.

5. Real-World Example: The "Volume" of a Mist

The paper also explains how this score predicts the behavior of volume integrals.
Imagine you have a cloud of mist (a volume) defined by a complex shape. You want to know how much mist is inside a tiny bubble of size $\epsilon$ around a messy corner.

As the bubble gets smaller, the amount of mist shrinks.
The Messiness Score tells you exactly how fast it shrinks.
If the score is low, the mist disappears slowly. If the score is high, it vanishes quickly.
The Multiplicity adds a "logarithmic" twist, like a slight delay or acceleration in that shrinking process.

Summary

In short, this paper:

Proves that for flat, crossing walls, the "real world" complexity measure is the same as the "imaginary world" measure.
Provides a simple, step-by-step recipe to calculate this measure for any arrangement of walls.
Builds a fast computer program to do the math for anyone.
Shows how this measure helps statisticians and machine learning experts choose the best models and understand how their models behave as they get more data.

It turns a very abstract, difficult mathematical problem into a solvable, calculable recipe for understanding complexity in both math and machine learning.

Technical Summary: Classification of Real Hyperplane Singularities by Real Log Canonical Thresholds

Problem Statement
The log canonical threshold (lct) is a fundamental invariant in birational geometry used to measure the complexity of singularities in algebraic varieties. Its real counterpart, the real log canonical threshold (rlct), also known as the learning coefficient, has become critical in statistics and machine learning for model selection and error estimation in singular statistical models (e.g., Gaussian mixtures and neural networks). While the rlct and its multiplicity are essential parameters in Watanabe's Singular Bayesian Information Criterion (sBIC) and the asymptotic analysis of volume integrals, explicit computational methods have historically been limited to specific examples or coordinate hyperplane arrangements.

This paper addresses the gap in the systematic computation of the rlct and its multiplicity for the general class of real hyperplane arrangements, which may be non-reduced (i.e., involving powers of linear forms). The authors aim to derive explicit combinatorial formulas for these invariants and provide a general algebraic framework connecting real and complex singularity theory.

Methodology
The authors employ techniques from birational geometry, specifically log resolutions and blow-ups, adapted to the real setting. The methodology proceeds through the following steps:

Algebraic Foundation: The paper establishes a rigorous connection between complex log canonical thresholds and real log canonical thresholds. It utilizes the concept of "wonderful compactifications" (introduced by De Concini and Procesi) to construct log resolutions for hyperplane arrangements.
Real vs. Complex Alignment: A central theoretical component involves proving that for real hyperplane arrangements (defined by products of linear forms with real coefficients), the real log canonical threshold and its multiplicity coincide with their complex counterparts. This is achieved by demonstrating that the prime divisors in the log resolution of a real arrangement always contain real points, satisfying specific conditions (Lemmas 3.5 and 3.6) that allow the transfer of results from the complex to the real domain.
Combinatorial Formulas: Using the intersection lattice of the hyperplane arrangement and the theory of building sets, the authors derive formulas for the rlct and multiplicity. These formulas rely on the codimension of intersections of hyperplanes and the sum of the multiplicities of the linear forms defining those intersections.
Implementation: The theoretical results are implemented in a SageMath package, utilizing linear algebra to compute intersections and the necessary combinatorial data.

Key Contributions and Results

General Algebraic Theory: The paper provides a sufficient condition (Lemma 3.5) under which the real log canonical threshold $\lambda_R$ equals the complex log canonical threshold $\lambda_C$ , and a stronger condition (Lemma 3.6) for the equality of their multiplicities ( $m_R = m_C$ ). These results formalize and prove assertions previously suggested in the literature (e.g., by Saito).
Combinatorial Formulas for Hyperplane Arrangements:
- Theorem 4.2: Provides explicit formulas for the complex log canonical threshold $\lambda_C$ and multiplicity $m_C$ for central hyperplane arrangements (products of linear forms) using a building set $G$ . The threshold is given by $\min_{W \in G} \frac{\text{codim}(W)}{s(W)}$ , where $s(W)$ is the sum of multiplicities of hyperplanes containing the intersection $W$ .
- Theorem 4.3: Proves that for any real hyperplane arrangement $f$ , $\text{rlct}(f) = \text{lct}(f)$ . This allows the direct application of the complex combinatorial formulas to the real case.
- Corollary 4.4: Presents the final combinatorial formulas for the real log canonical threshold $\lambda_R$ and multiplicity $m_R$ based on the intersection lattice $L(A)$ .
Extension to Non-Central Arrangements: Remark 4.5 demonstrates that the restriction to central arrangements is not a loss of generality. The global rlct of a non-central arrangement can be determined by minimizing the local rlcts of the central arrangements formed by the hyperplanes passing through each point in the ambient space.
Applications to Asymptotics: The paper illustrates how these invariants determine the asymptotic behavior of high-dimensional volume integrals of the form $V(\epsilon) = \int_{|f(w)| \leq \epsilon} \phi(w) dw$ , which behaves as $C\epsilon^{\lambda}(-\ln \epsilon)^{m-1}$ for small $\epsilon$ .
Software Implementation: The authors provide a SageMath implementation (Listing 1) that efficiently computes the rlct and multiplicity for hyperplane arrangements. Benchmarks indicate the code performs significantly faster than existing Macaulay2 implementations for the log canonical threshold of similar inputs.

Significance
The paper claims significance primarily in providing the first systematic, explicit combinatorial formulas for the real log canonical threshold and its multiplicity for the broad class of real hyperplane arrangements. By generalizing previous results limited to coordinate arrangements or specific statistical models, the work offers a unified algebraic theory that bridges complex birational geometry and singular learning theory.

The results enable the direct computation of the learning coefficients required for the sBIC in singular statistical models, facilitating better model selection and error estimation. Furthermore, the provided SageMath implementation offers a practical tool for researchers to analyze the asymptotic behavior of volume integrals and singularities in high-dimensional spaces without resorting to case-by-case geometric analysis. The paper emphasizes that these contributions formalize previously scattered results and extend the scope of computable invariants in both algebraic geometry and statistical learning theory.