What Makes a Measure "Genuinely Natural"? Bits, Nats, and the Audit of Naturalness Claims in Measurement, Information, and Language

In plain language

The authors wanted to test if calling a measure “natural” means anything real. They made three simple rules for a truly natural measure. They checked the claim that nats are more natural than bits. They found a constant 1.4426950 shows up in every bit calculation, not in nat ones. Their eight-paper sample was 62.5% language, 25% imaging, and 12.5% theory. This matters because the 1.4426950 factor slows calculations, so nats are easier to use.

Restated for a non-expert reader; the original wording follows.

In the paper's words

The word "natural" carries persuasive weight across the quantitative sciences: it appears in "natural logarithm," "natural language," and "natural images," and most pointedly in the recent information-theoretic proposal that the nat — the information unit based on the natural logarithm — is a "genuinely natural" measure, in contrast to Shannon's bit. This paper asks whether such naturalness claims are substantive or rhetorical. We formalize three candidate criteria for a measurement convention being genuinely natural — ratio invariance, dimensionlessness of derived constants, and reference-class independence — and audit the bit-versus-nat question against them.

The abstract, verbatim, and its opening sentence is the claim. Every line below is quoted from the text.

Numbers it reports

1. For base 2, $$\frac{d}{dx} \log_2 x = \frac{d}{dx} \frac{\ln x}{\ln 2} = \frac{1}{x \ln 2} = \frac{1.4426950}{x}.$$ The factor $1/\ln 2 = 1.4426950$ (Derivation 1) is a pure convention artefact: it appears in every differentiated expression regardless of the function being studied.

   from 4.4 Criterion C4: does the bit–nat choice discriminate anything?

2. We then contrast this mathematical sense of "natural" with the sociolinguistic and dataset-geometric senses in which the term appears across the surveyed literature, and quantify the disciplinary composition of the eight-work corpus (62.5% language-focused, 25% imaging/measurement-focused, 12.5% information-theory-focused) as a descriptive, not causal, observation.

   from Abstract

3. Then $$h_e(X) = -\int_0^1 1 \cdot \ln 1\, dx = 0\ \text{nats}.$$ Now let $Y = 2X$, uniform on $[0, 2]$, density $g(y) = \frac{1}{2}$.

   from 4.3 Continuous entropy: the Gaussian and the scaling failure

4. Third, the corpus of eight entries is a small, non-random sample; the 62.5%/25%/12.5% composition is a fact about the sample, not about the wider literature, and no causal claim follows from it.

   from 6. Discussion

5. What differs is calculus convenience: $\frac{d}{dx}\ln x = 1/x$ against $\frac{d}{dx}\log_2 x = 1.4426950/x$.

   from 7. Conclusion

6. All quantitative content is derived explicitly: we compute $\ln 2 = 0.6931471808$ from a rapidly convergent series, the conversion constant $\log_2 e = 1.4426950$, the distribution-independence of entropy ratios, the differential entropy of a unit-variance Gaussian ($1.4189386$ nats $= 2.0470956$ bits), and the scaling failure of differential entropy under coordinate change.

   from Abstract

How it is done

1. All quantitative content is derived explicitly: we compute $\ln 2 = 0.6931471808$ from a rapidly convergent series, the conversion constant $\log_2 e = 1.4426950$, the distribution-independence of entropy ratios, the differential entropy of a unit-variance Gaussian ($1.4189386$ nats $= 2.0470956$ bits), and the scaling failure of differential entropy under coordinate change.

   from Abstract

2. We use this entry only as a further instance of sense-2 "natural language" terminology and of task difficulty being motivated by human labor costs; the truncated text supports no methodological detail. [4] Time, Tense and Aspect in Natural Language Database Interfaces (arXiv:cmp-lg/9803002v1).

   from 2. Background and Related Work

3. For senses 2 and 3 of "natural," we perform a qualitative audit: for each supplied entry using the term, we extract what property of the system the entry itself identifies as load-bearing (language-dependent rules in [1]; hard syntax versus natural intent in [2]; human labor cost in [3]; temporal mechanisms versus limited formal facilities in [4]; transparency and interpretability in [5]; intrinsic dimension $d_{data}$ in [8]) and check whether that property is invariant in any mathematical sense.

   from 3.3 Method for the cross-sense and corpus audit

4. Every input number below is either a definition, a standard mathematical constant computed here from first principles, or a value derived in a prior step.

   from 4. Analysis

What would falsify it

1. We close with falsification conditions and open questions.

   from Abstract

2. Section 6 discusses limitations, failure modes, and falsification conditions.

   from 1. Introduction

3. Our conclusion that C4 fails is therefore an audit of the mathematical substance available behind the bit–nat framing, not a refutation of the full argument of [6]; if the complete work identifies an invariance or qualitative property that discriminates the two units, our C4 conclusion would be falsified.

   from 6. Discussion

Where the claim stops

1. The audit of [6] rests on a truncated abstract: the entry supports only that Shannon proposed bits for discrete entropy and nats for continuous systems, and that the authors argue something further that the supplied text does not disclose.

   from 6. Discussion

2. Our conclusion that C4 fails is therefore an audit of the mathematical substance available behind the bit–nat framing, not a refutation of the full argument of [6]; if the complete work identifies an invariance or qualitative property that discriminates the two units, our C4 conclusion would be falsified.

   from 6. Discussion

3. The analysis could mislead if a reader takes "C4 fails" to mean the bit–nat choice is practically irrelevant: in numerical work the factor $c = 1.4426950$ enters every differentiated expression, and calculus convenience is a real, if conventional, advantage for nats.

   from 6. Discussion

Where to read

1. Abstract229 words · 1 min
2. 1. Introduction425 words · 2 min
3. 2. Background and Related Work1,054 words · 5 min
4. 3. Methods449 words · 2 min
5. 4. Analysis1,547 words · 7 min
6. 5. Results188 words · 1 min
7. 6. Discussion375 words · 2 min
8. 7. Conclusion149 words · 1 min
9. Appendix A. Divergence report125 words · 1 min
10. Appendix B. Claim attribution181 words · 1 min