What's the one-line summary?

Harold Jeffreys (1946) 'An Invariant Form for the Prior Probability in Estimation Problems' derived the SQUARE-ROOT-OF-FISHER-INFORMATION prior — the unique prior invariant under reparametrization of the parameter space (Section 2 eq. 12 verbatim). Applied to a binomial (Section 2 eqs 19-21): reparametrizing p = sin²α with a uniform prior on α in [0, π/2] recovers exactly the Beta(0.5, 0.5) density used in today's Stats #44 Jeffreys credible interval. Full primary source verified via WebFetch + pdfminer.six on 2026-09-13 from a Wayback capture of the Royal Society PDF; 9 pages, 23,168 characters extracted. 34th PT of 44 with full primary-source access.

Why did Jeffreys reject the uniform Bayes-Laplace prior?

The uniform-on-p prior (Bayes-Laplace's Beta(1, 1)) is NOT invariant under reparametrization: if you reparametrize p → q = f(p), the uniform prior on p becomes NON-uniform on q, and your interval statement changes depending on which parameter you happen to have chosen. Jeffreys argued the ONLY prior whose interval statement is invariant under any smooth reparametrization is the one proportional to sqrt(det(I(θ))), where I is the Fisher information matrix — see Section 2 equation 12 verbatim (I quote the abstract in the tool): 'if the prior probability density of the parameters is taken as proportional to ||g_ik||^(1/2), the prior probability over any region will be invariant for all ways of choosing the parameters.' For a Bernoulli likelihood I(p) = 1/(p(1-p)), so the Jeffreys prior density ∝ 1/sqrt(p(1-p)) — the Beta(0.5, 0.5) density. Section 2 explicitly compares to Bayes-Laplace's uniform and to Haldane's improper dp/[p(1-p)] and calls Jeffreys's the 'intermediate' rule.

Where does the arcsine reparametrization show up in the paper?

Section 2 eqs (19)-(21), page 457 verbatim: 'Next consider simple sampling. Denote the chance of a success by sin²α, that of a failure by cos²α. Then for variations of α I_1 = (sin α − sin α')² + (cos α − cos α')² = 4 sin²½(α'−α), I_2 = (sin²α' − sin²α) log[sin²α'/sin²α] = 4(α'−α)². Then the rule gives, since 0 ≤ α ≤ ½π, P(dα | H) = dα/(½π).' The transformation is p = sin²α — mathematically identical to α = arcsin(√p) — and the uniform prior on α on [0, π/2] recovers a prior on p whose density is proportional to |dα/dp| = 1/(2 * sqrt(p(1-p))). That's the Beta(0.5, 0.5) density up to a factor of 2. Jeffreys DOES NOT use the words 'arc-sine' or 'sin⁻¹√p' anywhere in the paper — the 'arcsine' naming is a later convention.

How does this pair with the Stats #44 credible interval?

Prior Beta(0.5, 0.5) × Binomial(n, p) likelihood → posterior Beta(k + 0.5, n - k + 0.5) by conjugacy. The 95% central credible interval is [Beta.ppf(0.025, k+0.5, n-k+0.5), Beta.ppf(0.975, k+0.5, n-k+0.5)]. On today's slot 1 GBPNZD 9/11 sample: [53.28%, 96.02%] — tighter than Wilson [52.30%, 94.86%], Agresti-Coull [51.15%, 96.01%], and Clopper-Pearson [48.22%, 97.72%] (fails to reject). The Bayesian interpretation: 95% posterior probability the true up-rate is in [53.28%, 96.02%] given the sample and the Jeffreys prior. The frequentist interpretation is more delicate — BCD 2001 (yesterday's PT #43) Section 4 shows the Jeffreys interval's frequentist coverage oscillates near 95% rather than sitting at or above it, but on average it's competitive with Wilson.

Where can I actually read the paper?

The live royalsocietypublishing.org PDF returned 403 to WebFetch on 2026-09-13 (the Royal Society requires either institutional access or JSTOR credentials for pre-2000 articles). The verified working source for this post was a Wayback Machine capture from 2026-09-06 at https://web.archive.org/web/20260906062915/https://royalsocietypublishing.org/doi/pdf/10.1098/rspa.1946.0056 — 1.59 MB, 9 pages, 23,168 characters extracted via pdfminer.six. If the Wayback capture goes dead, the paper is reprinted in H. Jeffreys, 'Theory of Probability' (3rd ed., 1961), Oxford University Press, Section 3.10 pp. 179-181. Google Scholar also indexes several university mirrors that were 403 as of today. The paper is short (9 pages) and rewards direct reading — Section 2 is the whole binomial argument.

← All insights

Research13 September 2026· 5 min read

Paper Trail #44: An Invariant Form for the Prior Probability in Estimation Problems (Jeffreys, 1946) — Proceedings of the Royal Society A 186:453-461. FULL PRIMARY SOURCE VERIFIED via WebFetch + pdfminer.six on 2026-09-13 from a Wayback Machine capture of the Royal Society Publishing PDF (9 pages, 1.59 MB, 23,168 characters). 34th PT of 44 with full primary-source access. Derived the SQUARE-ROOT-OF-FISHER-INFORMATION prior that today's Stats #44 uses on the binomial: Section 2 eqs (19)-(21) verbatim show that reparametrizing p = sin²α with uniform prior on α in [0, π/2] recovers exactly the Beta(0.5, 0.5) density. Directly closes the four-paper small-sample-inference arc PT #40/#41/#42/#43 with the Bayesian sibling BCD 2001 recommends.

Harold Jeffreys (1946), “An Invariant Form for the Prior Probability in Estimation Problems.” Proc. Royal Society A 186:453-461. Full primary source verified via WebFetch + pdfminer.six on 2026-09-13 from a Wayback Machine capture of the Royal Society Publishing PDF (9 pages, 1.59 MB, 23,168 characters extracted). 34th PT of 44 with full primary-source access. Derived the SQUARE-ROOT-OF-FISHER-INFORMATION prior that today’s Stats #44 uses on the binomial: Section 2 eqs (19)-(21) verbatim show that reparametrizing p = sin²(alpha) with a uniform prior on alpha in [0, pi/2] recovers exactly the Beta(0.5, 0.5) density. Directly closes the four-paper small-sample- inference arc PT #40/#41/#42/#43 with the Bayesian sibling BCD 2001 recommends alongside Wilson for n ≤ 40.

Three prior densities on the unit interval for a binomial proportion p: Jeffreys Beta(0.5, 0.5) shown in orange with a U-shape with poles at 0 and 1 (density 3.20 at p=0.01, 0.64 at p=0.5), Bayes-Laplace uniform Beta(1, 1) shown as a coral horizontal line at density 1.0, and Haldane improper dp/[p(1-p)] shown as a light-grey dashed line that goes to infinity at both endpoints (density 100 at p=0.01, 4 at p=0.5, 100 at p=0.99). Jeffreys is intermediate between uniform and Haldane. — The three priors Jeffreys 1946 Section 2 compares. Uniform (coral) is Bayes-Laplace’s Beta(1, 1); Haldane (grey dashed) is improper dp/[p(1-p)]; Jeffreys (orange) is Beta(0.5, 0.5) — the “intermediate” rule Jeffreys derived from invariance.

The invariance derivation — Section 2 eq. (12)

Jeffreys 1946 abstract (p. 453) verbatim: “It is shown that a certain differential form depending on the values of the parameters in a law of chance is invariant for all transformations of the parameters when the law is differentiable with regard to all parameters ... This form has the properties required to give a general rule for stating the prior probability in a large class of estimation problems.” Section 2 equation (12) verbatim: “if the prior probability density of the parameters is taken as proportional to ||g_ik||^(1/2), the prior probability over any region will be invariant for all ways of choosing the parameters.” g_ik is defined in eq. (7) as the Fisher information matrix: g_ik = lim(δx_r → 0) Σ (1/p_r)(∂p_r/∂α_i)(∂p_r/∂α_k). The prior is sqrt(det(I(theta)))— later canonized as “the Jeffreys prior.”

The binomial special case — eqs (19)-(21)

Section 2, p. 457, eqs (19)-(21) verbatim: “Next consider simple sampling. Denote the chance of a success by sin²(alpha), that of a failure by cos²(alpha). Then for variations of alpha I_1 = (sin alpha − sin alpha’)² + (cos alpha − cos alpha’)² = 4 sin²(½(alpha’−alpha)), I_2 = (sin²alpha’ − sin²alpha) log[sin²alpha’ / sin²alpha] = 4(alpha’−alpha)². Then the rule gives, since 0 ≤ alpha ≤ ½pi, P(dalpha | H) = dalpha/(½pi). According to the Bayes-Laplace rule the prior probability of p should be taken uniform. Haldane has suggested dp/[p(1−p)]... The present rule is intermediate.” Reparametrizing p = sin²(alpha) with uniform prior on alpha in [0, pi/2] recovers a prior on p whose density is proportional to |d(alpha)/dp| = 1/(2 * sqrt(p(1-p))) — the Beta(0.5, 0.5) density up to a factor of 2.

Direct pairing with today’s Stats #44

Prior Beta(0.5, 0.5) × Binomial(n, p) likelihood → posterior Beta(k + 0.5, n - k + 0.5) by conjugacy. Today’s Stats #44 uses this to compute equal-tailed 95% credible intervals for the past-week samples: on today slot 1 GBPNZD 9/11 up, the Jeffreys interval is [53.28%, 96.02%] — the tightest lower bound of the four methods. BCD 2001 (yesterday’s PT #43) Section 4 recommended Wilson OR Jeffreys as the two small-n methods of choice; today’s PT #44 closes that thread with Jeffreys’s original 1946 derivation.

Small-sample-inference arc — five papers mapped

Post	Paper	Year	Contribution
PT #40	Hanley & Lippman-Hand JAMA	1983	Named the 'rule of three' — 3/n upper CI for k=0
PT #41	Wilson JASA	1927	Score-test-inversion interval (60 years pre-Wilson-name)
PT #42	Agresti & Coull AmStat	1998	'Add 2 successes and 2 failures' pedagogic simplification
PT #43	Brown, Cai, DasGupta Stat.Sci	2001	Evaluation — recommend Wilson OR Jeffreys for n ≤ 40
PT #44	Jeffreys ProcRoySocA	1946	Invariance derivation of Beta(0.5, 0.5) prior

Arc closed.BCD 2001 (PT #43) is the evaluation paper — it read the small-sample landscape and recommended Wilson (PT #41) or Jeffreys (today PT #44) for n ≤ 40. Today’s post gives Jeffreys its due primary-source citation.

Provenance and access

Verified working source on 2026-09-13: https://web.archive.org/web/20260906062915/ https://royalsocietypublishing.org/doi/pdf/10.1098/rspa.1946.0056 — Wayback Machine capture from 2026-09-06 of the Royal Society Publishing PDF. 1,590,772 bytes, 9 pages (pp. 453-461), 23,168 characters extracted via pdfminer.six. The live royalsocietypublishing.org PDF returned 403 to WebFetch (pre-2000 articles gated behind JSTOR / institutional access). The paper is also reprinted in H. Jeffreys, “Theory of Probability” (3rd ed., 1961), Oxford University Press, Section 3.10 pp. 179-181.

Not verified from this fetch

Pages 458-461 of the paper (sections beyond the binomial case — the location-scale extension, the Cauchy example, and the discussion of “restricted invariance” for laws not everywhere differentiable) exist in the fetched PDF but are not cross-checked line-by-line in this post. The four-item verbatim scope above covers the direct linkage to today’s Stats #44 Jeffreys interval; anything else claimed about the paper should be re-verified against the primary source before quoting. Jeffreys does NOT use the words “arc-sine” or sin⁻¹√panywhere in the paper — the “arcsine” naming for the Beta(0.5, 0.5) distribution is a later convention.

Keep reading