🔢 mathematics

Fast Evaluation of Truncated Neumann Series by Low-Product Radix Kernels

This paper introduces a novel framework for evaluating truncated Neumann series using higher-radix kernels, including the first exact rational construction for radix 9 and a residual-based approach for approximate kernels like radix 15, to achieve significant reductions in matrix-matrix product counts compared to traditional repeated squaring methods.

Original authors: Piyush Sao

Published 2026-02-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Piyush Sao

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 solve a massive puzzle, but instead of fitting pieces together one by one, you need to calculate a giant mathematical "sum" involving a matrix (a grid of numbers) raised to a very high power. This sum is called a Truncated Neumann Series.

In the world of supercomputers, calculating this sum is like trying to climb a mountain. The standard way to do it is to take one small step at a time (multiplying the matrix by itself over and over). If the mountain is 1,000 steps high, you have to take 1,000 steps. This is slow and exhausting.

Scientists have known for a while that you can climb faster by taking big leaps. Instead of one step, you can take two steps at a time (squaring the number), four steps, or even more. This is called "splitting." However, there's a catch: every time you take a big leap, you have to do a specific, complex calculation (a "kernel") to make sure you land exactly where you need to be.

This paper is about inventing better, faster leaping techniques to climb that mountain in record time.

The Problem: The "Step" Limit

For a long time, the best known technique was the "Quinary" (Radix-5) leap. It allowed you to jump 5 steps at a time using 2 complex calculations. It was great, but the researchers asked: Can we jump even further?

They wanted to jump 9 steps or even 15 steps at a time. The math suggested this would be possible, but nobody knew how to build the "engine" (the kernel) to make those jumps work without breaking the math.

The Breakthrough 1: The Perfect 9-Step Jump

The authors first built a Radix-9 Kernel.

The Analogy: Imagine you are a chef trying to bake a cake that requires 9 layers. The old way was to bake them one by one. The new way is to bake a "super-layer" that instantly creates 9 layers at once.
The Magic: They found a way to do this using only 3 multiplications (the "ingredients" needed for the calculation).
The Result: This is a "perfect" jump. It lands exactly where it's supposed to, with no errors. It saves about 21% more time than the previous best method. It's like finding a secret path up the mountain that cuts the travel time significantly.

The Breakthrough 2: The "Messy" 15-Step Jump

Next, they tried to jump 15 steps at once.

The Problem: When they tried to build the engine for a 15-step jump, they couldn't make it perfect. The math got too messy. The engine worked great for the first 14 steps, but the 15th step had a little bit of "spillover"—like a bucket of water that overflows slightly when you pour it.
The Old Fear: In the past, scientists thought if your calculation had even a tiny bit of "overflow" or error, the whole method was useless. You'd have to throw it away and go back to the slow, one-step-at-a-time method.
The New Solution: The authors invented a new framework (a new way of thinking). They realized: "What if we don't try to fix the overflow immediately? What if we just keep going, and the overflow fixes itself as we go?"
The Analogy: Imagine you are driving a car with a slightly leaky tire. The old rule was: "Stop immediately and fix the tire." The new rule is: "Keep driving, but check the tire pressure every few miles and add a little air." As long as the leak is small, you can still drive much faster than if you stopped every mile.
The Result: This "Radix-15" method is the fastest known way to climb the mountain. It saves even more time than the 9-step jump, reaching a new record speed.

Why Does This Matter?

You might ask, "Who cares about matrix math?"
This math is the engine behind many modern technologies:

AI and Machine Learning: Training neural networks often involves these calculations.
Medical Imaging: Reconstructing 3D images from MRI scans.
Weather Forecasting: Solving complex equations to predict storms.
Wireless Communication: Helping your phone connect to cell towers in crowded cities.

By making these calculations 25% faster, we can:

Save Energy: Supercomputers use massive amounts of electricity. Faster math means less energy wasted.
Solve Bigger Problems: We can tackle puzzles that were previously too big to solve in a reasonable time.
Get Answers Faster: Whether it's a weather forecast or a medical diagnosis, we get the results sooner.

The Bottom Line

This paper is a masterclass in efficiency.

They found a perfect shortcut for jumping 9 steps at a time.
They realized that imperfect shortcuts (with tiny errors) are actually okay if you have a smart system to manage them.
They used this to create the fastest possible shortcut for jumping 15 steps at a time.

They didn't just find a faster way to walk; they found a way to fly up the mountain, and they showed us how to handle the wind (the errors) so we don't crash.

1. Problem Statement

The Truncated Neumann Series is defined as $S_k(A) = \sum_{j=0}^{k-1} A^j = I + A + \dots + A^{k-1}$ , where $A$ is a matrix with spectral radius $\rho(A) < 1$ . This series is fundamental in numerical linear algebra for:

Approximate matrix inversion ( $(I-A)^{-1}$ ).
Polynomial preconditioning.
Resolvent approximation and log-determinant estimation.
Massive MIMO signal detection.

The Bottleneck: In dense matrix settings, the dominant computational cost is matrix-matrix multiplication (GEMM).

Naive evaluation requires $k-1$ products.
Binary splitting (repeated squaring) reduces this to $2 \log_2 k$ products.
Hybrid radix-2/3 schemes lower the constant slightly to $\approx 1.9 \log_2 k$ .

The asymptotic order $O(\log k)$ is already achieved; the current research goal is to minimize the leading coefficient of the product count. Existing methods rely on "radix kernels" (efficient ways to compute partial sums $T_m(B) = I + B + \dots + B^{m-1}$ ). While a radix-5 kernel exists (requiring only 2 products), explicit constructions for higher radices (e.g., radix-9, radix-15) were missing, and it was unclear if exact rational kernels existed for them.

2. Methodology

The paper proposes a framework to evaluate $S_k(A)$ using Radix- $m$ splitting, where the series is updated in blocks of size $m$ . The efficiency depends on the kernel cost $\mu_m$ , defined as the minimum number of matrix products required to compute $T_m(B)$ .

The total cost per update step is $C(m) = \mu_m + 2$ (where 2 accounts for concatenation and power updates). The asymptotic coefficient is $\frac{\mu_m + 2}{\log_2 m}$ .

The authors tackle two main challenges:

Exact Constructions: Finding explicit algebraic formulas for higher radices (specifically radix-9) that achieve the theoretical lower bound of products.
Approximate Constructions: Handling higher radices (specifically radix-15) where exact rational solutions may not exist, leading to "spillover" (unwanted higher-degree terms).

A. The Radix-9 Kernel (Exact)

The authors derive an exact kernel for $m=9$ using 3 matrix products ( $\mu_9 = 3$ ).

Strategy: They factorize the polynomial $T_9(B)$ into a product $P \cdot Q$ , where $P$ and $Q$ are linear combinations of basis terms $\{B, B^2, B^3+2B^4\}$ .
Result: The construction uses rational coefficients (denominators dividing 800), allowing for exact arithmetic.
Cost: $C(9) = 3 + 2 = 5$ products.
Coefficient: $\frac{5}{\log_2 9} \approx 1.58 \log_2 k$ .

B. The Radix-15 Kernel (Approximate)

For $m=15$ , the theoretical lower bound is $\mu_{15} \ge 4$ .

Challenge: No exact symbolic solution was found. Numerical optimization (L-BFGS-B) yielded a 4-product kernel that matches the target polynomial up to degree 14 but produces spillover (non-zero coefficients) for degrees $\ge 15$ .
Implication: Standard telescoping identities fail because the spillover prevents the exact extraction of $A^{mn}$ .

C. General Radix-Kernel Framework (Residual-Based)

To utilize the approximate radix-15 kernel, the authors developed a new framework that applies the kernel to the residual rather than the matrix power directly.

Mechanism: Instead of computing $S_{mn} = S_n \cdot T_m(A^n)$ , the method maintains an approximate inverse $Y_n$ and a residual $R_n = I - (I-A)Y_n$ .
Iteration: $Y_{n+1} = Y_n \cdot f(R_n)$ , where $f$ is the approximate kernel.
Error Analysis: The framework proves that even with spillover, the "prefix" (the correct low-degree terms) grows as $m^n$ , while the error (spillover) is pushed to higher degrees. As long as the spillover coefficient $|c| < 1$ , the method converges.
Cost: The step cost remains $\mu_m + 2$ .
Coefficient: For radix-15 with $\mu_{15}=4$ , the coefficient is $\frac{6}{\log_2 15} \approx 1.54 \log_2 k$ .

3. Key Contributions

First Exact Radix-9 Kernel: The paper presents the first explicit construction for a radix-9 kernel using exactly 3 products with rational coefficients, improving the asymptotic coefficient to 1.58.
Residual-Based Framework: A novel mathematical framework that allows the use of approximate kernels with spillover. This resolves the issue of "broken" telescoping identities, enabling the use of higher radices where exact solutions are elusive.
Approximate Radix-15 Kernel: Through numerical optimization, the authors constructed a 4-product kernel for radix-15. Within the new framework, this achieves the best-known asymptotic coefficient of 1.54.
Unification: The framework subsumes previous methods (binary splitting, Hensel lifting, exact telescoping) as special cases, providing a unified view of series evaluation.

4. Results

Theoretical Savings:
- Radix-9: Reduces product count by 21% compared to binary splitting ( $2.00 \to 1.58$ ).
- Radix-15: Reduces product count by 23% compared to binary splitting ( $2.00 \to 1.54$ ).
- For $k = 9^{10} \approx 3.5 \times 10^9$ , Radix-9 saves 13 matrix multiplications compared to binary splitting.
Numerical Experiments:
- Tested on random dense matrices ( $d=500$ ) and non-normal matrices.
- Exact Methods (Binary, Quinary, Radix-9): Achieved residuals near machine precision ( $\approx 10^{-14}$ ).
- Approximate Method (Radix-15): Converged to a residual floor of $\approx 10^{-6}$ , determined by the precision of the kernel coefficients (prefix error $< 9 \times 10^{-7}$ ). This is expected behavior for approximate kernels.
- Runtime: The reduction in matrix multiplications translated directly to wall-clock time savings.

5. Significance

Breaking the Barrier: This work challenges the long-held assumption that higher-radix splitting offers no benefit beyond radix-3 or hybrid schemes. It demonstrates that the leading coefficient can be pushed below 1.6.
Algebraic vs. Numerical: It highlights a transition in algorithm design: moving from purely exact algebraic constructions (radix-5, radix-9) to numerically optimized approximate constructions (radix-15) to gain efficiency.
Practical Impact: For large-scale dense linear algebra problems (e.g., in machine learning, physics simulations, or signal processing), reducing the number of GEMMs by 20-25% yields significant performance gains, especially on hardware where matrix multiplication is the bottleneck.
Future Directions: The paper opens questions regarding the existence of exact radix-15 kernels, the potential for even higher radices, and extensions to structured matrices (sparse/Toeplitz).

In summary, Piyush Sao presents a breakthrough in the computational complexity of Neumann series evaluation, introducing a new class of radix kernels and a robust framework for handling approximate solutions, achieving the fastest known asymptotic rate for this problem.