A measure of intelligence of an approximation of a real number in a given model

Imagine you are trying to describe a complex, mysterious object (like the number $\pi$ ) to a friend using only a limited set of building blocks (like simple fractions or square roots).

Some descriptions are lazy. You might say, " $\pi$ is roughly 3.14159." That's accurate, but it's a mouthful. It's like describing a masterpiece painting by listing every single pixel's color code. It's technically correct, but it's boring and inefficient.

Other descriptions are brilliant. You might say, " $\pi$ is roughly $22/7$." This is slightly less accurate than the long decimal, but it's incredibly simple. It's like describing that same painting as "a sunset over a mountain." It captures the essence with very few words.

This paper by Bakir Farhi asks a simple but profound question: How do we mathematically measure the "intelligence" or "cleverness" of an approximation?

Here is the breakdown of his idea, translated into everyday language:

1. The Two Ingredients of a "Smart" Approximation

Farhi argues that a good approximation needs a balance of two things:

Simplicity (Size): How many "ingredients" does the formula use? Is it a tiny fraction like $22/7 $, or a massive, clunky number like$ 314159/100000$?
Accuracy (Error): How close is the guess to the real number?

The Analogy: Think of it like packing for a trip.

Naive Packing: You bring your entire house. You have everything you need (perfect accuracy), but it's impossible to carry (too complex).
Intelligent Packing: You bring a small backpack with just the essentials. It's not everything you own, but it's enough to survive, and it's easy to carry.
The Paper's Goal: To create a scorecard that tells you if your "backpack" (approximation) is packed intelligently.

2. The "Intelligence Score" ( $\mu$ )

Farhi invents a formula to give every approximation a score, which he calls $\mu$ (mu).

The Logic: The score compares the "cost" of the approximation (how many digits are in the numbers you used) against the "reward" (how many correct digits of the real number you got).
The Rule:
- If your score is less than 1, the approximation is Naive. You worked too hard for too little gain. (Like using a sledgehammer to crack a nut).
- If your score is 1 or higher, the approximation is Intelligent. You got a great return on your effort. (Like using a laser pointer to guide a robot).
- The Higher the Score, The Smarter: A score of 3.0 is "very intelligent," meaning you got a massive amount of accuracy for a tiny amount of complexity.

3. Real-World Examples from the Paper

Farhi tests his scorecard on famous numbers like $\pi$ and $e$ .

The Winner ( $\pi \approx 22/7$ ): This is a classic. It's a small fraction, but it gets you the first two decimal places of $\pi$ . It gets a high "Intelligence Score."
The Loser ( $\pi \approx 314159/100000$ ): This is technically more accurate, but the numbers are huge. You had to write down 10 digits to get just a tiny bit more accuracy. The score is low. It's "naive."
The "Magic" Approximations: Farhi finds some weird-looking formulas (like $\pi \approx \sqrt{2} + \sqrt{3}$ ) that look complicated but are actually very "intelligent" because they use small numbers to get surprisingly good results.

4. The "Continued Fraction" Connection

The paper dives deep into a mathematical tool called Continued Fractions. Imagine these as a special way of writing numbers that naturally breaks them down into their simplest, most efficient forms.

Farhi proves a cool rule: Any "convergent" (a step in the continued fraction process) is automatically an "intelligent" approximation.

Metaphor: If you are climbing a mountain and you take the most direct path (the continued fraction), every step you take is an "intelligent" move. You aren't wasting energy.

However, he also discovers that sometimes, you can take a "shortcut" that isn't on the standard path but is still a brilliant move. He proves there are "intelligent" approximations that don't follow the standard rules, but they are rare and hard to find.

5. The Big Mystery (The Open Problem)

The paper ends with a question that mathematicians are still scratching their heads over:

"Can we always find a 'smart' way to approximate any number, no matter how weird it is?"

For simple numbers (like $\pi$ or $e$ ), the answer seems to be "Yes, but there's a limit to how smart we can get."
For "super weird" numbers (called Liouville numbers), the answer is "Yes, and we can get infinitely smart!" You can find approximations that are so efficient they break the rules of normal math.

Summary

Bakir Farhi's paper is like a critic reviewing recipes.

He doesn't just care if the cake tastes good (accuracy).
He cares if the recipe was efficient (simplicity).
He gives a score to every recipe. If the score is high, the chef is a genius. If the score is low, the chef is just throwing ingredients at the wall and hoping for the best.

His work gives us a rigorous way to say, "Wow, that approximation is clever," rather than just guessing.

Here is a detailed technical summary of the paper "A measure of intelligence of an approximation of a real number in a given model" by Bakir Farhi.

1. Problem Statement

The paper addresses a fundamental gap in approximation theory: while mathematicians have long recognized that certain approximations of real numbers (e.g., $\pi \approx 22/7$ ) are "intelligent," "elegant," or "interesting," there has been no rigorous mathematical definition for these qualitative terms.

Intuitively, an approximation is considered "intelligent" if it achieves high accuracy with low complexity (simplicity). Conversely, an approximation that is highly accurate but requires massive parameters (e.g., $\pi \approx 314159/100000$ ) is considered "naive." The author seeks to formalize this intuition by defining a quantitative measure of intelligence ( $\mu$ ) that balances the size (complexity) of the approximation against its accuracy (error).

2. Methodology and Definitions

2.1 Models of Approximation

The author defines a model of approximation ( $M$ ) as a map from a set of integer parameters to the real numbers:
$M: A \subset (\mathbb{Z}^*)^n \to \mathbb{R}$
where $n$ is the number of parameters.

Examples:
- $M_1$ (Rational model): $M(a, b) = a/b$ .
- $M_2$ : $M(a, b) = a + b\sqrt{2}$ .
- $M_3$ : $M(a, b, c) = a + b\sqrt{c}$ .
Refinement: A model $M'$ is "finer" than $M$ if the image of $M$ is contained in the image of $M'$ .

2.2 Size and Logarithmic Size

For an approximation $x \approx M(a_1, a_2, \dots, a_n)$ , the size is defined as the product of the absolute values of the parameters:
$s(x \approx M(\dots)) = |a_1| \times |a_2| \times \dots \times |a_n|$
The logarithmic size is:
$s_{\log} = \log |a_1| + \dots + \log |a_n|$
This serves as a proxy for the "complexity" or the number of digits required to write the parameters.

2.3 The Measure of Intelligence ( $\mu$ )

The core contribution is the definition of the measure of intelligence $\mu$ . An approximation is deemed intelligent if the total number of digits in the parameters does not exceed the number of correct digits in the result.

Formally, for a real number $x$ and an admissible approximation $x \approx M(a_1, \dots, a_n)$ :
$\mu(x \approx M) = \frac{\log |x| - \log |x - M(a_1, \dots, a_n)|}{\log |a_1 a_2 \dots a_n|}$

Interpretation:
- The numerator represents the number of correct digits (integer part + decimal places).
- The denominator represents the complexity (number of digits) of the parameters.
Classification:
- Intelligent: $\mu \geq 1$ .
- Naive: $\mu < 1$ .
- Higher Intelligence: A larger $\mu$ indicates a "more intelligent" approximation.

3. Key Contributions and Results

3.1 Validation via Numerical Examples

The author applies the measure to historical and new approximations of $\pi$ and $e$ :

$\pi \approx 22/7$ : $\mu \approx 1.55$ (Intelligent).
$\pi \approx 355/113$ : $\mu \approx 1.53$ (Intelligent, though slightly less than $22/7$ despite higher accuracy).
$\pi \approx 314159/100000$ : Implied to be naive due to high denominator size relative to accuracy gain.
Ramanujan's $\pi$ approximations: Several are shown to be highly intelligent ( $\mu > 2$ ).
New Approximations: The author proposes several new intelligent approximations for $\pi$ , $e$ , and other constants (e.g., $\sqrt{e}$ , $\sqrt{\pi}$ ) with $\mu$ values ranging from 1.3 to 3.6.

3.2 Theoretical Analysis in the Rational Model

The paper focuses heavily on the rational model ( $x \approx p/q$ ) and establishes deep connections with Diophantine approximation and Continued Fractions.

Proposition 6.1: An approximation $p/q$ is intelligent iff $|\alpha - p/q| \leq |\alpha|/|pq|$ .
Corollary 6.3: Any regular continued fraction convergent of an irrational number $\alpha$ is an intelligent approximation in the rational model. This proves that the classical theory of continued fractions naturally yields intelligent approximations.
Existence of Non-Convergent Intelligent Approximations:
- Theorem 6.7 & 6.9: The author proves that there exist intelligent rational approximations that are not regular continued fraction convergents.
- Specifically, if the continued fraction expansion $[a_0; a_1, \dots]$ satisfies certain conditions (e.g., $a_{n+1} \geq a_n - 1$ ), then truncating or modifying the last term yields an intelligent approximation that is not a standard convergent.
- Application to $\sqrt{2}$ and $\sqrt{5}$ : The paper demonstrates that these numbers have infinitely many intelligent approximations outside their standard convergents.

3.3 Boundedness and Liouville Numbers

The paper investigates whether the measure of intelligence can be arbitrarily large.

Theorem 6.11: The set of intelligence measures for a real number $x$ $x$ is unbounded if and only if $x$ $x$ is a Liouville number.
- Liouville numbers can be approximated "too well" by rationals, allowing $\mu \to \infty$ .
Corollary 6.12: For any irrational algebraic number (and by extension, many transcendental numbers like $\pi$ $π$ , $e$ $e$ , $\log 2$ $lo g 2$ , $\zeta(3)$ $ζ (3)$ ), the set of intelligence measures is bounded.
- Significance: This implies that for standard mathematical constants, there is a theoretical limit to how "intelligent" an approximation can be; one cannot find an approximation with arbitrarily high intelligence for $\pi$ or $e$ .

4. Significance and Open Problems

Significance

Formalization of Intuition: The paper provides the first rigorous mathematical framework to quantify the "elegance" or "intelligence" of an approximation, moving beyond subjective preference.
Unification: It bridges the gap between classical Diophantine approximation (continued fractions) and the concept of approximation quality, showing that continued fraction convergents are inherently optimal in terms of this intelligence metric.
Classification of Constants: It categorizes real numbers based on the potential "intelligence" of their approximations, distinguishing between Liouville numbers (unbounded intelligence) and algebraic/transcendental numbers (bounded intelligence).

Open Problems

The author concludes with a significant open problem:

Generalization: While the existence of intelligent approximations is proven for the rational model (using Dirichlet's Pigeonhole Principle), it remains unproven whether any real number $x$ (that is a limit point of a given model $M$ ) admits an intelligent approximation in that model $M$ .
The author suggests that proving this for non-rational models (e.g., involving logarithms or roots) may require new methods involving density arguments or distribution modulo 1, as the pigeonhole principle does not directly apply.

Conclusion

Bakir Farhi's paper successfully defines a metric ( $\mu$ ) that quantifies the trade-off between the complexity of an approximation's parameters and its accuracy. The theory validates historical approximations, generates new ones, and reveals that while continued fractions provide a rich source of intelligent approximations, the "intelligence" of approximations for standard constants is fundamentally bounded, unlike for Liouville numbers.