Seventy-two problems, nearly all solved in the dark, 1872 to 1891, every answer checked by a program
Having Eluded Us Both
In 1893 Lewis Carroll published seventy-two problems, nearly all solved in his head while lying awake, listed the mistakes he had found in them when he wrote them out, and left the rest for “some critical reader”. A program has now checked 577 claims in the book, and 537 of them hold. It finds fifteen mistakes in the printed text (one of them perhaps a joke): among them a theorem that the numbers 1, 4 and 13 refute, a chance of 7/12 that should be 1/2, a table of chances wrong in every line beneath an answer that is right, a coin he left out, and two angles each half a degree out.
The answer, first of all
In 1893 Charles Lutwidge Dodgson, "Student and late Mathematical Lecturer of Christ Church Oxford", whom the world knew as Lewis Carroll, published Curiosa Mathematica, Part II: Pillow-Problems. His Introduction says how the problems were made:
NEARLY all of the following seventy-two Problems are veritable "Pillow-Problems", having been solved, in the head, while lying awake at night. (I have put on record the exact dates of some.) No. 37 and one or two others belong to the daylight, having been solved while taking a solitary walk; but every one of them was worked out, to the very end, before drawing any diagram or writing down a single word of the solution. I generally wrote down the answer, first of all: and afterwards the question and its solution.
He wrote the answer down before the working that would test it, which is the order a careful experimenter keeps today when a prediction is registered before the data come in. Then he tested it. "It is not, perhaps, much to be wondered at that, when these Problems came to be re-written and arranged for publication, a good many mistakes were discovered." The worst he dropped; the rest he corrected in his solutions, "but, that I may not be credited with an amount of accuracy, as a computator, which I am well aware I do not possess, I here append a list of them." Ten problems are on the list. F. G. Brabant of Corpus Christi College read the proofs, "first working out each result independently". And the Introduction ends with an invitation:
Other mistakes may perchance, having eluded us both, await the penetrating glance of some critical reader, to whom the joy of discovery, and the intellectual superiority which he will thus discern, in himself, to the author of this little book, will, I hope, repay to some extent the time and trouble its perusal may have cost him!
C. L. D., Christ Church, Oxford, May 1893 (Introduction, pp. xii to xix of the fourth edition, 1895)
This page takes him up on it, with a program rather than a glance. Every answer he printed, the result of every solution and the figures in it where a slip could hide, and every entry in his list of mistakes is a claim; each claim is checked by a computation that starts from the question, not from his solution. The checks are short JavaScript files, and every one of them runs in your browser from the ledger below.
What the program found
The checks make 577 claims about the book: each figure and formula of the 44 printed answers, the results of the 72 solutions and the figures and steps in them where a slip could hide, what his questions and prefaces assert, the ten entries of his list of mistakes, and the other things his Introduction says about particular problems. 537 of them hold. The other 40 come down to 15 mistakes that got past him and Brabant, 10 questions whose words allow more than one answer, and one theorem that no check by examples can settle. Nearly everything he worked out in the dark is right. What is not right is worth seeing one at a time.
A theorem that is not true (No. 61)
No. 61 asks for a proof that if three whole numbers cannot be arranged in arithmetical progression, and their sum is a multiple of 3, the sum of their squares is also the sum of three other squares, with no square in common. His proof makes the new numbers by a rule (each new number is two-thirds of the sum, less one of the old), shows that the squares add up to the same total, and shows that a new number can equal an old one only if the three are in progression. But two numbers can have the same square without being equal: a new number can be minus an old one. Take 1, 4 and 13. They are not in progression, and their sum, 18, is a multiple of 3. Their squares make 186; his rule gives 11, 8 and −1, and 1 is one of the old squares. Nor is there any other way out: 186 is a sum of three squares in exactly three ways, 1 + 16 + 169, 1 + 64 + 121 and 16 + 49 + 121, and the other two each share a square with the first. With repeats allowed (his own examples include 1, 4, 4), the smallest failure is 1, 1, 7, or 1, 1, 4 if 0 does not count as a square; 1, 4 and 13 fail either way.
It is the problem whose proposed solution he took apart in the preface to his second edition, in September 1893. A correspondent of the Educational Times had offered one family of examples, 3m, 21m and 30m, whose squares make the same sum as four other sets of three, and all that proved, Carroll wrote, "is that certain selected Numbers, which have property 'α', have also property 'β'", not that every set with the first property has the second. He was right about the logic; it is the theorem itself that fails. The journal had printed his theorem, under his name, as its question 9636, and the solution printed for it in Mathematical Questions and Solutions, the volumes that collected the journal's problems, "by W. S. Foster, Professor Nash, and others", makes the same move as his proof and stops at the same place: "2m−a is not equal to a or b or c unless these quantities are in Arithmetical Progression" (vol. LXI, 1894, p. 86).
A remark that adds nothing, and changes the answer (No. 41)
A friend brings a bag of four counters, each black or white, and has you draw two: both white. "I meant to tell you, before you began, that there was at least one white counter in the bag. However, you know it now, without my telling you. Draw again." What is now the chance of white, and what would it have been if he had not spoken? Carroll answers 7/12 and 1/2. The friend has it right, though: two whites drawn already prove at least one white, so the remark removes no possibility and the two chances are the same, 1/2. The error is in the first line of his solution, where "at least one white" becomes the chances 1/8, 3/8, 3/8, 1/8 for four, three, two and one whites. Those are the chances for a bag in which one particular counter is known to be white. A bag with at least one white has the chances 1/15, 4/15, 6/15, 4/15: his own 1, 4, 6, 4 from the second part, with the all-black bag struck out. It is the trap of "at least one" against "this one".
Fill the bags
A table wrong in every line, under an answer that is right (No. 50)
Two bags, H and K, each hold two counters; a white is added to H, one counter goes from H to K, and one comes back. His solution ends with a table of the chances that H now holds three, two, one or no whites, each line made of products of chances he has already found, and the lines come to 25, 50, 29 and 4 in 108ths. They should be 24, 51, 30 and 3. Each of his products multiplies the chance of what H kept by the chance that the counter coming back is white (5/9) or black (4/9), as if the two were independent; but the counter that comes back is drawn from K, which holds the one H gave up, and what H gave up depends on what it kept. Given that H kept two, one or no whites, the returning counter is white with chance 8/15, 5/9 and 2/3, not 5/9 each time. His answer, 17/27, is right anyway, because it needs only the average number of whites, and an average adds up whether or not its parts are independent. His Introduction confesses a slip in this very table ("bringing out 41/108, instead of 50/108"): the line he corrected to 50/108 should have been 51/108.
A coin he left out (No. 10)
A triangular billiard-table has a pocket in each corner, one holding only one ball and the others two each. Three balls, each containing a coin, run into one corner, nobody knows which, and the expected contents of the pocket are 2s. 6d.; what are the coins? His working is sound: the three coins must make 4s. 6d. His answer is "Either 2 florins and a sixpence; or else a half-crown and 2 shillings", and his Introduction confesses that he first missed the half-crown and 2 shillings. There is a third set: a double florin and two threepenny pieces. The double florin, a four-shilling piece, was struck from 1887 until August 1890, the month in which he dated this problem, and stayed in circulation afterwards; silver threepences were struck for circulation in every year from 1838 to 1901 but three. The double florin was unpopular ("Up to this date it has failed to obtain a general currency", the London correspondent of the Liverpool Mercury wrote in May 1888, as Wikipedia's article on the coin quotes it), so a reader may forgive him for not reaching for it. But his answer says either, or else, and his solution "Hence the coins must be", which is a claim that there are no others.
Two angles, half a degree out (No. 70)
The longest problem in the book wraps triangles round a tetrahedron and asks, among other things, for two angles. His exact answers are right: the angles whose tangents are √3/5 and √3/7. Turning them into degrees by "mem. tech." (angles he had by heart for round values of the sine), he printed 18·65° and 14·53°. They are 19.11° and 13.90°. In the first, the sine is √84/28, and he knew √84 was "rather over 9", but he worked with 9, so he placed the angle a fifth of the way between his remembered angles where it lies 0.27 of the way. In the second, the sine .24 lies two-fifths of the way between .2 and .3, and he went "about half-way".
His Introduction quotes the first answers he wrote down, before he wrote any of the solution: about 18° 18′ and about 14°. "These answers are not quite correct", he says, and his list of mistakes explains each by a decimal he forgot to add. For the second angle that explanation is half the story. His first answer, 14°, was right to the nearest degree, because the decimal he forgot and the "half-way" he did not notice cancelled out. Putting back the decimal and keeping the "half-way" gave the 14·53° he printed, which is the answer more than half a degree out.
Where the sine lies
A subtraction in the seventh place (No. 45)
No. 45 asks for the chance that, of infinitely many rods broken, at least one breaks in the middle. His answer comes to 1 − 1/e, which he works out to seven places below a note: "[N. B. What follows here was not thought out.]" So this is paper, not pillow. His own lines give 3·0861612 − 2·7182818 = ·3678794, and he wrote ·3678793; the answer, ·6321207, should be ·6321206. (What the question means by infinitely many rods is another matter: see below.)
His column, and the subtraction
A decimal, and a confession under the wrong number (No. 37)
Two circles cross; the common chord subtends 30° at one centre and 60° at the other; what fraction of the smaller lies within the larger? (This is one he solved on a walk, by daylight.) His exact answer is right for the figure he draws, and he gives it as "about ·044". It is 0.0429, which is ·043 to three places. In the working, one term is rounded up (·478) and the other cut short (·434); treated alike, both ways give ·043. His Introduction confesses a slip in these very figures, that in his head he had made ·478 into ·476 and so brought out ·042 for ·044, but it files the confession under No. 38, a problem about bags of counters with no decimal in it. And his mental ·042 was, as it happens, nearer the truth than the ·044 he printed.
Five small slips, some perhaps the printer's
In No. 31 a line gives the watch a rate that belongs to the clock, and the next line uses it, rightly, for the clock. In No. 33 a bracket is out of place in one line of algebra, and the next line is right. In No. 47 a strict inequality should not be strict (the next sentence has it right). In No. 65 one of three sets of fractions reads 1/2, 1/5, 1/20 where it must be 1/4, 1/5, 1/20, and a list of values includes a 1/2 that his own condition excludes; the count he draws from it is right. None reaches an answer, and any of the first four could be the compositor's rather than his; only a printed page can say.
The paradox (No. 72)
The last problem asks for the colours of two counters in a bag "without taking them out of the bag", and answers: one black and one white. Add a black counter, he argues: the chance of then drawing black works out at 2/3, and a bag of two blacks and a white is the only state that gives 2/3; so the bag must have held one of each. The 2/3 is an average over the bags the question could be holding. That one of them, a black and a white with the black added, gives 2/3 by itself does not make it yours: the chance that yours holds one of each is still 1/2. Carroll calls it his "one Problem in 'Transcendental Probabilities'", which "may seem abnormal, and even paradoxical". Warren Weaver, in 1956, took it for "two dreadful mistakes"; Amirouche Moktefi, in 2008, calls it a "controversial joke-problem". The page counts it among the mistakes as mathematics and does not claim to know which he meant.
Fill the bags
Questions with more than one answer
Ten questions are worded so that more than one answer is defensible. In each, his answer is right under the reading his solution takes, and the check gives the answer under the others too. The best known is No. 58, three points "taken at random on an infinite Plane": there is no way to choose points evenly from an infinite plane, so the chance of an obtuse triangle depends on how they are chosen. His procedure gives 0.6394, three points spread evenly over a disc give 0.7197, over a square 0.7252, and independent bell-curve points exactly 3/4. Stephen Portnoy (1994), Ruma Falk and Ester Samuel-Cahn (2001) and Jason Cantarella and colleagues (2019) have written about this one at length.
No. 45 is the same kind of question in disguise. With N rods, each breaking at one of n equally likely points, the chance that at least one breaks in the middle tends to 1 − e−c when N/n tends to c. "An infinite number of rods" does not say what c is; his solution quietly takes as many rods as points, c = 1.
As many rods as points, or not
No. 37's circles can cross in a second way the words allow, with both centres on the same side of the chord; then 0.9852 of the smaller circle is inside the larger. No. 38's "one proves to be red" gives 49/72 if a particular counter is looked at (his reading, which the words favour) and 49/95 if we are told only that at least one is red. Nos. 14 and 29, two theorems about sums of squares, are true if 0 counts as a square and fail in some cases if it does not. No. 13's area formula gives minus the area when the tetragon is re-entrant; No. 69's ratio of the sides of two similar triangles, 2 cos θ, is negative when his variable θ passes 90°, which his own working allows; No. 25's least number of a hospital's patients who have lost an eye, an arm and a leg goes below zero when the three fractions who have lost each add up to less than 2; and three of No. 47's four sets of solutions depend on letting 0/0 stand for any number. Every one is in the ledger.
Two figures the words allow
Seventy-two nights
Carroll dated most of the problems, day and month and year or month and year. Here are the dated ones on a line from his earliest, January 1872, to his latest, December 1891, each coloured by what the checks found. Choose one to open it.
His list, checked
Each entry in the Introduction's list of mistakes is itself a claim: that he made the slip, and that the corrected figure is right. The table also checks what the rest of the Introduction says about particular problems: Brabant's neater answer to No. 59, a figure and the dates in his "biography" of No. 63, the first answers to No. 70, and his "one Problem in 'Transcendental Probabilities'".
What he did not say
Many of the problems ask for a construction in "a given Triangle", and a construction is only half an answer until it says when it can be done. His usually do exactly what he claims whenever an answer exists, and say nothing about when one does not. None of these is counted as a mistake; each is in the ledger with what the check found. In No. 18 and No. 20 there is no such point at all when a base angle is obtuse. No. 26 can be done only in an acute-angled triangle. No. 11's inner triangle exists only while the given angle is less than 180° less the largest angle of the triangle. No. 56's construction from three altitudes works exactly when the altitudes belong to some triangle, which is not always. No. 46's works in 691 of 1,500 random cases, exactly those where an answer exists with the two remaining angles placed as in his figure; in 234 more an answer exists only with those two angles the other way round, and in 575 there is none. In No. 57 the three squares lie inside the triangle in 232 of 400 random triangles, and not in 168 of the 242 obtuse ones. And in No. 51, when an obtuse angle makes the problem impossible, his steps still draw a line, sometimes one that ends on the right side and fails only when measured. In the other direction, No. 9 has two answers whenever the angle is acute; his construction can find either, and he does not say there are two.
How the checks were made
The text is Project Gutenberg's transcription of the fourth edition (Macmillan, 1895; eBook 79080, released in July 2026, made by Aaron Adrignola, Laura Natal and the Online Distributed Proofreading Team from HathiTrust's page images), cut into the 72 problems by the book's own numbering: 72 questions, 44 printed answers, 72 solutions. A claim is anything he asserts that a computation can test, and every claim quotes his words; a program refuses any quotation that is not in the book. Each problem has a check: a short JavaScript file that builds the question from its data alone and computes what he claims, in exact fractions where it can, in floating point with a stated tolerance where it cannot, on hundreds of random triangles where the claim is a formula or a construction, and, where it is about chances, by listing every case when the cases are finite and by integration or a limit when they are not. His solutions are read only to learn his reading and to test his steps.
The checks were written by AI agents: No. 37's by the agent that built this page, as the worked example for a written brief, and the other 71 by eight more, each with a batch of problems and that brief. Every verdict that says Carroll erred, or that a question can be read two ways, was then worked out again from the book by a second agent that had not seen the first one's code or reasoning, with problems known to be right mixed in among them unmarked. The two agreed on every mistake on this page. They differed on three labels. The second checker would have called No. 25's missing condition a mistake and No. 69's missing absolute value a slip, and would have said No. 38's answer holds, its words favouring his reading; the page counts all three as questions of reading. A verdict is one of five words: holds (right, exactly or to the figures printed), slip (a figure wrong, the method sound), wrong (more than a slip: a formula, a prior, a count, or a list that claims to be complete), depends on the reading (the words allow more than one answer, and his is right under one of them), and not checkable (a theorem about every figure, No. 15, which every figure tried obeys).
Three limits, stated rather than implied. A check on random instances is evidence, not a proof. A reading of a question is a choice, and each is printed beside its verdict for you to dispute. And the figures are the transcription's: we could not read a page image of the book ourselves, because the libraries holding scans (HathiTrust and Google Books) turn automated readers away. The transcribers say that they replaced his private symbols for sine and cosine with sin and cos, and that where the print does not distinguish a from α they corrected certain expressions. A second transcription of the fourth edition, at schnark.github.io, made separately (it has slips of its own that Gutenberg's does not), agrees with Gutenberg's on the disputed figures we compared in it, in Nos. 10, 31, 33, 37, 38, 41, 45, 47, 50, 65 and 70, except in two places: in No. 50's table it has 22/108 where Gutenberg's has 29/108, and in No. 70 "about 1½ degrees" where Gutenberg's has 1⅕; in each, Gutenberg's reading is the one his own arithmetic gives (29/108 is his product, and 1⅕ is the 1·20° printed beside it). No. 61's theorem we also read on a page image of another printing, the Educational Times volume above, where it is set as the same theorem in slightly different words. Even so, any single figure called a slip here could in principle be the printer's or a transcriber's rather than Carroll's, and the five small slips above are the likeliest to be. The first edition of 1893 was not read at all, and its figures may differ.
Who looked before
As far as we could find, nobody has published a check of all 72 answers before. Carroll's own list, and F. G. Brabant's working of every result, are the first. Eugene Seneta's centenary paper of 1993 examines a few of the thirteen problems on chances and says "Some are badly posed with imaginative but incorrect solutions"; we could read only its abstract, so we do not know which he meant, and it may already say some of what this page says about Nos. 38 and 41. No. 58 has a literature of its own (Portnoy 1994; Falk and Samuel-Cahn 2001; Cantarella, Needham, Shonkwiler and Stewart 2019), and so does No. 72 (Weaver 1956; Moktefi 2008). We found no earlier statement that No. 61 is false, or of the mistakes in Nos. 10, 37, 45, 50 and 70; a scout read what it could reach under the site's rules for automated reading, and the pages those rules refuse were not read (one was fetched by mistake, and deleted at once unused). The Gutenberg volunteers made the text this page is built on, and a separate transcription at schnark.github.io let us compare the figures we dispute.
Check it yourself
Every check runs in your browser from the ledger above. To run all of them, and a second set of arithmetic that recomputes every finding on this page without using them, download the verifier and run it with Node (version 22 or later, no other installation):
curl -sSO https://artwaste.land/checks/verify-having-eluded-us-both.mjs
node verify-having-eluded-us-both.mjs
It downloads the page's data and check files into a folder beside it, runs every check again, holds every quotation the checks make to the book's text, recomputes each finding by its own route, and fails if any of the 59 figures that this page's source marks for it differs from what it computes. --online also fetches the book from Project Gutenberg, confirms its hash is the one recorded, and finds in it every part of the book the checks read and every quotation they make; --mutate plants errors in a copy of the page and confirms each one is caught.