💰 quantitative finance

A New Approximation to the Normal Distribution Quantile Function

This paper introduces a faster, albeit slightly less accurate, approximation to the normal distribution quantile function compared to the Beasley and Springer method, offering a maximum absolute error under $2.5 \cdot 10^{-5}$ that remains sufficient for many applications, particularly in financial markets.

Original authors: Paul M. Voutier

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 3.0

Original authors: Paul M. Voutier

Original paper licensed under CC BY 3.0 (http://creativecommons.org/licenses/by/3.0/). ✨ 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 a chef trying to bake a perfect cake (a "Normal Distribution"). To do this, you need a very specific, secret ingredient: the Quantile Function. Think of this function as a magical map that tells you exactly how much of an ingredient to add based on a probability number you give it.

For a long time, statisticians have used two main recipes for this map: one by Beasley and Springer (very precise, but slow to cook) and one by Abramowitz and Stegun (a bit faster, but less precise).

Paul Voutier, the author of this paper, wanted a recipe that was faster than the others, even if it meant sacrificing a tiny bit of precision. Here is the simple breakdown of what he did:

1. The Strategy: Splitting the Kitchen

The author realized that the "map" behaves differently in the middle of the room versus the corners.

The Middle (Central Region): This is where most of the action happens. Here, the map is smooth and predictable.
The Corners (The Tails): These are the extreme ends (very low or very high probabilities). Here, the map gets tricky and requires expensive, time-consuming math operations (like taking square roots and logarithms) to calculate.

The Old Way: The Beasley-Springer recipe used a very complex, high-precision tool for the middle section. It was accurate but took a long time to compute.

The New Way: Voutier decided to use a simpler, lighter tool for the middle section. He swapped a heavy, complex machine for a quick, efficient hand-tool.

He calls this a (2, 2) scheme. Imagine it as a simple 2-step ladder instead of a 5-step staircase.
For the tricky corners (the tails), he used a slightly different, slightly more complex tool (a (3, 2) scheme) to make sure the map didn't break at the edges.

2. The Trade-off: Speed vs. Perfection

Voutier admits his new map isn't as perfectly precise as the Beasley-Springer one.

Beasley-Springer: Like a master watchmaker; the error is microscopic (less than 1 in a billion).
Voutier's New Method: Like a very good pocket watch; the error is still tiny (less than 1 in 40,000), but it's not quite as perfect.

However, the paper argues that for many real-world jobs (like managing financial markets), being "good enough" is fine, as long as you get the answer much faster.

3. The Results: The Race

The author put his new recipes to the test against the old ones using a computer (a Dell laptop with an Intel processor). He ran the calculations 200,000 times to see who was fastest.

Here is how they stacked up (in milliseconds):

Abramowitz-Stegun (Old Standard): 25,210 ms (The slowest)
Beasley-Springer (The Accurate One): 10,212 ms (The middle ground)
Voutier's New Method (Rat22A): 8,052 ms (Faster!)
Voutier's Optimized Method (Rat22B): 6,649 ms (The Winner!)

The Big Win: Voutier's fastest version was nearly 4 times faster than the old standard and about 1.5 times faster than the highly accurate Beasley-Springer method.

4. Why It Matters

The paper highlights a key lesson: Sometimes, doing less math is better.
By making the "middle" part of the calculation simpler, the computer didn't have to do as many heavy lifting operations (like square roots and logs). Even though these heavy operations only happen at the extreme edges of the data, cutting them down by half saved nearly 20% of the total time.

Summary

Paul Voutier built a new, streamlined calculator for a specific statistical map. It's not the most precise map in the world, but it is the fastest one available. If you are in a hurry (like in high-speed financial trading) and can tolerate a tiny bit of rounding error, this new method is the best tool for the job.

Technical Summary: A New Approximation to the Normal Distribution Quantile Function

Problem Statement
The paper addresses the need for a computationally efficient approximation of the normal distribution quantile function, $N^{-1}(x)$ , which is the inverse of the cumulative distribution function (CDF). While existing methods, such as the Beasley-Springer approximation [3] and Acklam's approximation [2], offer high accuracy, the author identifies a specific requirement for a method that prioritizes speed while maintaining "reasonably accurate" results (specifically, a maximum absolute error less than $4.5 \cdot 10^{-4}$ , comparable to the Abramowitz-Stegun approximation [1]). The primary challenge is to reduce the computational cost of expensive operations, specifically logarithms and square roots, which are typically required for tail calculations.

Methodology
The proposed approach follows the established framework of separating the domain into a wide central region and narrow tails, leveraging the odd symmetry of $N^{-1}(x - 1/2)$ . However, the author introduces a modification to the rational function schemes used for approximation:

Central Region Approximation:
- Instead of the higher-order schemes used by Beasley-Springer (a (3,4) scheme) or Acklam (a (5,5) scheme), this paper proposes a (2,2) scheme for the central region.
- The function is defined as $f(p) = q \frac{a_2 r^2 + a_1 r + a_0}{r^2 + b_1 r + b_0}$ , where $q = p - 0.5$ and $r = q^2$ .
- Two variations are presented:
  - Rat22A: Uses boundaries at $p = 0.0465$ and $p = 0.9535$ .
  - Rat22B: Uses wider boundaries at $p = 0.025$ and $p = 0.975$ to further reduce the frequency of tail calculations.
- Coefficients were derived using the minimax function from Maple's numapprox package to minimize the maximum absolute error (Chebyshev approximation).
Tail Region Approximation:
- For the tails ( $p < 0.0465$ and $p > 0.9535$ ), a (3,2) scheme is employed.
- The approximation utilizes the variable $r = \sqrt{\log(1/p^2)}$ (or $\sqrt{\log(1/(1-p)^2)}$ for the upper tail).
- The function is defined as a rational function of $r$ : $f_{tail}(p) = \frac{c_3 r^3 + c_2 r^2 + c_1 r + c_0}{r^2 + d_1 r + d_0}$ .
- Coefficients were similarly optimized using Maple's minimax algorithm.
Optimization Techniques:
- The rational functions are structured to minimize the number of multiplications (e.g., normalizing the denominator so the leading coefficient is 1).
- The wider central region (Rat22B) is specifically designed to reduce the number of times the computationally expensive log and sqrt operations must be executed.

Key Contributions and Results

Accuracy: The new (2,2) central approximation achieves a maximum absolute error of approximately $2.5 \cdot 10^{-5}$ $2.5 \cdot 1 0^{- 5}$ in the central region and $2.458 \cdot 10^{-5}$ $2.458 \cdot 1 0^{- 5}$ in the tails. The total maximum absolute error is less than $2.5 \cdot 10^{-5}$ $2.5 \cdot 1 0^{- 5}$ .
- Note on Accuracy: The author explicitly acknowledges that this is less accurate than the Beasley-Springer approximation (which has an error under $1.85 \cdot 10^{-9}$ in the central region) and the Acklam approximation.
Performance: The primary contribution is speed. Benchmarks performed on a Dell Inspiron 1525 (Intel Core 2 Duo T5800, 2.00 GHz) using Java (JDK 1.6.0 17) over 200,000 iterations yielded the following results:
- Abramowitz-Stegun (AS): 25,210 ms
- Beasley-Springer (BS): 10,212 ms
- Rat22A (New, narrow central): 8,052 ms
- Rat22B (New, wide central): 6,649 ms
Efficiency Gains: The new approximations are significantly faster than existing methods. The comparison between Rat22A and Rat22B demonstrates that reducing the subset of values requiring log and sqrt calculations by roughly 50% (by expanding the central region) reduced CPU time by nearly 20%.

Significance and Claims
The paper posits that while the proposed approximation sacrifices some precision compared to state-of-the-art methods like Beasley-Springer, the trade-off is justified by the substantial gain in computational speed. The author states that an error margin of less than $2.5 \cdot 10^{-5}$ is "sufficient for many applications," specifically citing financial markets as a domain where speed is crucial.

Additionally, the paper revisits the Abramowitz-Stegun approximations (Equations 26.2.22 and 26.2.23 in [1]). The author presents improved coefficients for these equations that offer over five times the accuracy of the original within the standard IEEE-754 representable range, though the accuracy degrades as the range approaches the extremes of 0 and 1.

In summary, the work provides a "faster" alternative to established quantile function approximations, prioritizing execution time for applications where the slightly higher error bound remains acceptable.