1Introduction
Around 1637, according to the tradition that rests on his son’s edition of 1670, Pierre de Fermat wrote in the margin of his copy of Bachet’s Diophantus that it is impossible to divide a cube into two cubes, or a fourth power into two fourth powers, or in general any power above the second into two of the same kind, and that he had found a truly marvellous proof of this, which the narrowness of the margin would not contain (hanc marginis exiguitas non caperet) [1, 3]. The copy itself has not survived. The sentence survives because Samuel de Fermat printed it as Observation II among the forty-eight marginal notes he took from his father’s volume.
The historical literature has treated the remark as a statement about mathematics. The question has been whether Fermat had a proof, and if so what it looked like [5, 6, 10]. Most authors who commit themselves doubt that he had a correct one for general exponent. The usual conjecture is that he had an argument of the kind Lamé presented to the Paris Academy in 1847, which fails where unique factorisation fails in cyclotomic integers [5]. The general case was settled by Wiles in 1995, in a paper of 109 pages with a further 20 pages by Taylor and Wiles [9, 11].
The remark is also a statement about a physical object. Whether a margin is too small depends on its capacity and on the length of what is to be written in it, and we found no published estimate of the capacity for this volume. A second observation follows from the first. Fermat made many other assertions of general results in the margins of the same book, and most of them have since been proved, at lengths that are on record. If the margin was the limiting factor, those assertions should carry the same remark. They do not. Observation II is the only one that mentions the margin.
This paper has three parts. We estimate the capacity of the outer margin of the 1621 edition at the opening containing Book II, Problem 8, from fourteen surviving copies and a sample of Fermat’s autograph hand. We then compare that capacity with the lengths of the proofs of eight of Fermat’s claims, using both the earliest published proof and the shortest proof we could find. Finally we test whether the frequency of Fermat’s remark on the narrowness of the margin agrees with the frequency with which the margin would have been too narrow. The measurements and the model are described in Section 3, the estimates in Section 4, and their consequences for the historiography in Section 5.
2Sources
2.1The edition and the missing copy
Bachet’s edition of 1621 prints the Greek text of the Arithmetica with a Latin translation in parallel and his own commentary [1]. It is a folio. In Samuel de Fermat’s edition of 1670 the observations of his father are printed as a separate series after Bachet’s commentary, each keyed to a problem [3], and they are reprinted with the rest of Fermat’s surviving writings in the standard edition [8]. The annotated volume is not known to exist. Its margins therefore cannot be measured, and we infer them from other copies.
2.2Which copies are relevant
Binders trimmed the fore-edge of a book when they sewed and plough-cut it, and rebinders trimmed it again. A margin of 47 mm in 1621 can be 30 mm in 1850. Two things follow for the choice of reference class. First, the note must have fitted inside the margin when it was written, so a later trim could not have limited it, though an earlier trim could. Second, a Toulouse magistrate who annotated a Paris folio in the 1630s was almost certainly annotating a volume in its first binding. We therefore take copies still in contemporary binding as the reference class and use rebound copies only as a check on how much a trimmed fore-edge changes the answer.
2.3The annotation hand
Fermat’s letters and memoranda survive in autograph, and the hand he used for marginalia is smaller and more contracted than the hand of his correspondence. We restricted the handwriting measurements to passages in the small hand, which we identified by an x-height under 1.5 mm in the facsimile plates of the standard edition [8]. Contractions follow the usual practice of seventeenth-century Latin and were expanded according to the standard repertory [2].
3Materials and methods
3.1Copies
We obtained measurements for 14 copies of the 1621 edition. Five were measured in person and nine from images with a scale in the frame or from dimensions supplied by the holding library. At the opening containing Book II, Problem 8, we recorded the leaf height and the width of the outer margin, taken from the edge of the leaf to the nearest printed text. The outer margin carries no printed matter at this opening in any of the copies. Each quantity was measured three times and averaged. The standard deviation of repeat measurements was 0.8 mm for copies measured in person and 1.1 mm for copies measured from images. Eight copies are in contemporary binding and six were rebound after 1700. We take the usable height of the margin to be mm, where is the leaf height, which leaves 20 mm at head and foot.
3.2Hand
We sampled 31 autograph documents. In each, we took five lines of the small hand and measured the mean horizontal distance per character over ten consecutive characters (, mm) and the mean distance between consecutive baselines (, mm). Document means were used as the unit of analysis. Standard errors in Table 2 are for the mean over documents.
3.3Contractions and word length
For twelve passages of ten lines each, we transcribed the autograph as written, expanded the contractions, and took the ratio of characters as written to characters in the expanded text. The mean length of a Latin word, including the following space, was taken from a sample of 1,900 words of the 1670 text and is characters.
3.4Capacity
If text is written horizontally in lines across the margin, the number of lines is , each line holds characters, and a word occupies written characters. Words are not split, so on average half a word is lost at the end of each line. The capacity in words is
Text rotated by a quarter turn and written along the length of the margin gives
Confidence intervals for the mean capacity of a contemporary copy come from 20,000 replicates in which the eight contemporary copies were resampled with replacement and , and were drawn from normal distributions with the means and standard errors in Table 2. Fermat’s own copy is one draw from the population of copies, so we also report a prediction interval, in which and are drawn from normal distributions with the across-copy means and standard deviations. The wording of the note in Observation II, Cubum autem in duos cubos through non caperet, is 36 words in the 1670 text. The space left for a proof is therefore .
3.5Claims and proof lengths
We took eight claims attributed to Fermat in the Observations and the correspondence for which a complete proof has since been published [4, 10]. For each we recorded the earliest published proof that is accepted today and the shortest proof we could find in the literature. The length of a proof is the number of words in a standard English presentation, with each displayed formula counted as eight words and each inline formula as one (Appendix A). For the proof of Wiles and of Taylor and Wiles we used the page count, 129 pages, multiplied by 640 words per page, the median of ten sampled pages counted by the same rule. A proof fits in the margin if .
3.6Reporting of margin insufficiency
Two readers coded the 48 observations of the 1670 edition independently. Set consists of the observations that assert a general result and give no proof. Set consists of those in which Fermat also states that he possesses a demonstration. Cohen’s for membership of was 0.81, and disagreements were settled by discussion. Let be the proportion of the eight claims whose shortest known proof does not fit in the margin. If Fermat remarked on the margin whenever the margin was too small, and only then, the number of observations in a set of size carrying the remark would follow a binomial distribution with parameters and . We report for the observed count , with at its point estimate and at the lower end of its exact 95% interval.
4Results
4.1Margins
Table 1 lists the copies. The eight copies in contemporary binding have a mean outer margin of 46.6 mm (SD 2.7) and a mean leaf height of 337.5 mm (SD 2.4). The six later-bound copies have a mean outer margin of 32.3 mm (SD 4.2) and a mean leaf height of 322.3 mm (SD 5.2). The difference in margin width is 14.3 mm (Welch , ). Trimming reduces both dimensions, and it reduces the margin proportionally more (Figure 1). The median page count of the eight contemporary copies is 482, with a range of 478 to 490.
| Copy | Binding | Access | Leaf height (mm) | Outer margin (mm) |
|---|---|---|---|---|
| A | C | Image | 338 | 49 |
| B | C | In person | 341 | 44 |
| C | C | Image | 335 | 51 |
| D | C | Image | 336 | 46 |
| E | C | In person | 340 | 43 |
| F | C | Image | 334 | 48 |
| G | C | Image | 337 | 45 |
| H | C | In person | 339 | 47 |
| I | L | Image | 322 | 33 |
| J | L | Image | 318 | 29 |
| K | L | In person | 326 | 36 |
| L | L | Image | 315 | 27 |
| M | L | Image | 324 | 31 |
| N | L | In person | 329 | 38 |
4.2Capacity of the margin
Table 2 gives the inputs to Eq. (1) and Eq. (2). The usable margin is 46.6 mm wide and 297.5 mm high. In the small hand, Fermat wrote 1.92 mm per character with a baseline distance of 3.4 mm, and contractions saved 11% of the characters of the expanded text.
| Quantity | Symbol | Value | SD | Unit | |
|---|---|---|---|---|---|
| Outer margin width (contemporary copies) | 46.6 | 2.7 | mm | 8 | |
| Usable margin height | 297.5 | 2.4 | mm | 8 | |
| Character pitch, small hand | 1.92 | 0.21 | mm | 31 | |
| Baseline distance, small hand | 3.4 | 0.45 | mm | 31 | |
| Contraction ratio | 0.89 | 0.04 | 12 | ||
| Expanded word length including space | 7.0 | characters | 1,900 words |
With these inputs, Eq. (1) gives words (95% interval 274.3 to 322.4). Rotated text gives words (310.4 to 360.2). The 95% prediction interval for the horizontal capacity of a single unmeasured contemporary copy is 254.6 to 342.2 words. If the six later-bound copies are used instead, the capacities are 183 and 220 words. The margin holds 0.999 words per millimetre of its length at the mean width. Figure 2 shows how the capacity responds to the two pitch parameters. Across the grid, which runs from 1.6 to 2.3 mm in character pitch and from 2.8 to 4.0 mm in baseline distance, capacity ranges from 205 to 444 words.
The note itself occupies 36 words, 12.1% of the capacity, which leaves words in the horizontal layout and 298.4 words in the rotated layout. We use the horizontal value, which is the smaller, in what follows.
4.3Lengths of the proofs of Fermat’s claims
Table 3 gives the lengths. None of the earliest published proofs fits in . The nearest is Euler’s proof of Fermat’s little theorem, at 380 words, which is 1.45 times the space available. Two of the shortest known proofs do fit: the textbook proof of Fermat’s little theorem, at 70 words, and Zagier’s proof that a prime congruent to 1 modulo 4 is a sum of two squares, at 90 words [12]. The other six exceed the margin at every length we found for them, by factors between 1.3 and 316 (Figure 3). The proof of the case by descent, which Fermat did write down, is 640 words in its original form and 450 in the shortest form we found. It exceeds in both.
| Claim | Earliest proof | Shortest proof | ||||
|---|---|---|---|---|---|---|
| Euler, 1736 | 380 | 1.45 | Textbook | 70 | 0.27 | |
| Prime is | Euler, 1749 | 2,900 | 11.09 | Zagier, 1990 | 90 | 0.34 |
| Every integer is a sum of four squares | Lagrange, 1770 | 2,200 | 8.41 | Hurwitz, 1896 | 550 | 2.10 |
| is solvable | Lagrange, 1768 | 3,100 | 11.85 | Pigeonhole argument | 340 | 1.30 |
| Polygonal number theorem | Cauchy, 1813 | 4,000 | 15.30 | Nathanson, 1987 | 1,600 | 6.12 |
| has no solution | Fermat, c. 1640 | 640 | 2.45 | Textbook | 450 | 1.72 |
| has no solution | Euler, 1770 | 1,900 | 7.27 | Textbook | 900 | 3.44 |
| , | Wiles, Taylor–Wiles, 1995 | 82,560 | 315.72 | Same | 82,560 | 315.72 |
4.4Reporting of margin insufficiency
Of the 48 observations, 19 are in set and 4 are in set . Observation II is in both sets, and it is the only observation in either that mentions the margin. Six of the eight shortest known proofs exceed , so (exact 95% interval 0.35 to 0.97). If the margin had been mentioned whenever it was too narrow, we would expect 14.3 of the 19 observations in to carry the remark, against one observed. The probability of one or none is at and at . In set the expected count is 3.0 of 4, and the probability of one or none is 0.051 at and 0.56 at the lower limit. With the earliest published proofs in place of the shortest, and the probability for set is zero.
4.5Scale
The proof of Wiles and Taylor–Wiles has an estimated length of 82,560 words. At the measured density it needs 82.6 m of margin 46.6 mm wide, or equivalently a single margin 12.9 m wide on a leaf 297.5 mm high. The outer margins of the 482 pages of the volume have a total length of 143.4 m, so the proof would take 57.6% of them.
5Discussion
For the nineteen unproved assertions the data do not agree with the hypothesis that Fermat remarked on the margin whenever it was too narrow. The probability of the observed count is small even at the least favourable in the interval. We read this as under-reporting. On the evidence of later proof lengths the margin would have been insufficient for at least six of his claims, and he recorded the insufficiency once.
There are two ways to avoid the conclusion. The first is to restrict the test to set , the four observations in which Fermat says he has a demonstration. A remark about the narrowness of a margin makes sense only for a proof one has, and in set the count of one is not unusual (0.051 at , 0.56 at the lower limit). We find this the more natural reading, though it leaves three observations in in which Fermat claimed a proof that, on our estimates, would probably not have fitted, and said nothing. The second is to suppose that for the other assertions Fermat did not have a proof, and had no occasion to complain about the space. The data cannot separate these accounts from under-reporting. They place a bound on how many observations in he could have mentioned.
Two of the shortest known proofs fit in the margin. One of them, the proof of the two-squares theorem, was published in 1990, and is shorter than Euler’s by a factor of 32. The other is the textbook proof of the little theorem. A margin of this size could have held a proof of either result, so a remark about the margin would not have been expected for them, and none is recorded. These two claims are therefore consistent with the test whichever way it is read.
The proof for sits awkwardly with the capacity estimate. At 640 words it is longer than two margins, yet Samuel de Fermat printed it as a single observation [3]. Either the descent was written across several consecutive margins or onto a blank leaf, or it was written at a place with a wider margin than the opening we measured, or it was composed more densely than the small hand we sampled. We did not identify the opening in any of the 14 copies. If the third explanation holds, the character pitch used here is too large, and the first column of Figure 2 (1.6 mm, a capacity of 366 words) is the more appropriate one. At 1.6 mm the residual margin is 330 words, and six of the eight shortest proofs still exceed it. A capacity above 376 words, which is within the grid, would bring the shortest proof for Pell’s equation (340 words) inside the margin and lower to 0.625.
A rhetorical reading of the phrase is possible, in which hanc marginis exiguitas non caperet is a conventional way of declining to continue and the margin plays no role. This reading predicts no relation between the margin and the length of the proof, and nothing in our data can test it. We have assumed throughout that the statement is literal. A related case is the note in the manuscript of Galois, Je n’ai pas le temps [7], which gives a temporal rather than a spatial limit. The analogue of Eq. (1) would need a writing speed, and we have no data on it.
The limitations are these. The fourteen copies are those we could reach, and they are not a random sample of the edition. We measured one opening, and the outer margin at Book II, Problem 8 may not be representative. Fermat’s copy might have been bound with blank leaves that we have not modelled. Eight claims is a small set, so is imprecise. The word counts in Table 3 are our own and depend on the presentation chosen. The count of 640 words per page for the Annals papers is a median over ten pages, and the formulae there are denser than in the classical proofs, so the length for the proof of 1995 is probably an underestimate in words.
6Concluding remarks
The uncertainty in for a single copy has two parts. Copy-to-copy variation in margin width accounts for about two thirds of the variance of the prediction, and the character pitch, baseline distance and contraction ratio account for the rest. The fore-edge width of Fermat’s own copy would be the single most useful measurement, and it is the only one that cannot be obtained. With it, the 95% prediction interval for would narrow from 88 words to about 51. The classification of the observations is the other open item. A third reader coding set would settle whether it contains four observations or five, and with five the probability in the last section changes from 0.051 to 0.016 at .
Data availability
Measurements for each of the 14 copies, the handwriting samples, the coding of the 48 observations and the code for Eq. (1) and the test in Section 3 are available from the corresponding author.
ACounting words in proofs
The word counts in Table 3 were made on the following rule. Prose was counted in words of English. A displayed formula was counted as eight words, which is the median number of tokens in the displayed formulae of ten sampled pages of the Annals paper, rounded. Inline formulae were counted as one word each. Diagrams were not counted. For the proofs of Euler, Lagrange, Cauchy and Fermat the counts are of the presentation in the book of Edwards or the history of Weil where one is given there [5, 10], and otherwise of a translation of the original. Rounding is to the nearest ten for lengths below 1,000 words and to the nearest hundred above, except for the final row, which is the product of 129 pages and 640 words.
References
- [1]Claude-Gaspar Bachet de Méziriac. Diophanti Alexandrini Arithmeticorum libri sex, et de numeris multangulis liber unus. Sebastien Cramoisy, Paris, 1621.
- [2]Adriano Cappelli. Lexicon Abbreviaturarum: Dizionario di abbreviature latine ed italiane. Ulrico Hoepli, Milan, 1899.
- [3]Samuel de Fermat, editor. Diophanti Alexandrini Arithmeticorum libri sex, et de numeris multangulis liber unus. Cum commentariis C. G. Bacheti V. C. et observationibus D. P. de Fermat Senatoris Tolosani. Bernard Bosc, Toulouse, 1670.
- [4]Leonard Eugene Dickson. History of the Theory of Numbers, Volume II: Diophantine Analysis. Carnegie Institution of Washington, Washington, DC, 1920.
- [5]Harold M. Edwards. Fermat's Last Theorem: A Genetic Introduction to Algebraic Number Theory. Springer, New York, 1977.
- [6]Michael S. Mahoney. The Mathematical Career of Pierre de Fermat, 1601–1665. Princeton University Press, Princeton, 1994.
- [7]Peter M. Neumann. The Mathematical Writings of Évariste Galois. European Mathematical Society, Zürich, 2011.
- [8]Paul Tannery and Charles Henry, editors. Œuvres de Fermat. Gauthier-Villars, Paris, 1891.
- [9]Richard Taylor and Andrew Wiles. Ring-Theoretic Properties of Certain Hecke Algebras. Annals of Mathematics, 141(3):553–572, 1995.
- [10]André Weil. Number Theory: An Approach through History from Hammurapi to Legendre. Birkhäuser, Boston, 1984.
- [11]Andrew Wiles. Modular Elliptic Curves and Fermat's Last Theorem. Annals of Mathematics, 141(3):443–551, 1995.
- [12]Don Zagier. A One-Sentence Proof That Every Prime p = 1 (mod 4) Is a Sum of Two Squares. American Mathematical Monthly, 97(2):144, 1990.