Widely accepted mathematical results that were later shown to be wrong?

It is a memorable fact that the first known proof of history is false: the first proposition of Euclid which announces that one can construct an equilateral triangle with a given segment for one of the sides. We cannot always construct this triangle for the reason that the circles which intersect in the Euclidean plane ($mathbb R^2$) do not necessarily intersect in the rational Euclidean plane ($mathbb Q^2$)

Aristotle stated as a fact that the regular tetrahedron tiles space. This was accepted and repeated in commentaries on Aristotle for 1800 years, until Regiomontanus showed it was wrong. A detailed history is given in Lagarias and Zong, Mysteries in packing regular tetrahedra, Notices of the AMS, Volume 59, Number 11, December 2012, pages 1540 to 1549.

Very recently, Dobbs A minimal ring extension of a large finite local prime ring is probably ramified, ZBL 07192436. identified an error in the proof that "a separable extension of finite rings is always Galois" (Corollary XV.3 of McDonald, Bernard R., Finite rings with identity, Pure and Applied Mathematics. Vol. 28. New York: Marcel Dekker, Inc. IX, 429 p. (1974). ZBL 0294.16012.).

In 1892, Michel Frolov Equalities of the second and third degree, Bull. Soc. Math. Fr. 20, 69-84 (1892). JFM 24.0176.01. claimed a proof that there are no 7th-order bimagic squares, since there is no series of 7 distinct odd numbers, from 1 to 49, with sum 175 and sum of squares 5775; this is, however, not true (e.g. 1,7,25,31,33,37,41). The non-existence of 7th-order bimagic squares has been only confirmed in 2004 via computer calculations, see http://www.multimagie.com/English/Smallestbi.htm.

A recent example: The main theorem of [Masa-Hiko Saito, On the infinitesimal Torelli problem of elliptic surfaces, J. Math. Kyoto Univ. 23, 441-460 (1983). ZBL 0532.14019] has been shown to be incorrect due to a counterexample in [Atsushi Ikeda, Bielliptic curves of genus three and the Torelli problem for certain elliptic surfaces, Adv. Math. 349, 125-161 (2019). ZBL 1414.14004].

Pontryagin made a famous mistake in A classification of continuous transformations of a complex into a sphere which led him to the false conclusion that the homotopy group $pi_{n+2}(S^n)$ is zero. Later Freudenthal in Über die Klassen den Sphärenabbilgunden I. Grosse Dimensionen and Whitehead in The $(n+2)^{text{nd}}$ homotopy group of the $n$-sphere showed with different methods that $pi_{n+2}(S^n) cong mathbb Z_2$ and Pontryagin corrected his mistake in Homotopy classification of the mappings of an $(n+2)$-dimensional sphere on an $n$-dimensional one.

Let me try to explain what Pontryagin got wrong: Let $f colon S^{n+2}to S^n$ be a smooth map which represents an element of $pi_{n+2}(S^n)$ (in every homotopy class of continouus maps between manifolds there is a smooth representative). Following Sard's Theorem there is an $x_0 in S^n$ which is a regular value of $f$, thus $Sigma:=f^{-1}(x_0)$ is a closed $2$-dimensional submanifold of $S^n$. The normal bundle of $Sigma$ in $S^{n+2}$ is trivial and has a natural framing induced by the derivative of $f$ and a choice of a basis in $T_{x_0}S^n$. Pontryagin defined a map $$ varphi colon H_1(Sigma;mathbb Z_2) to mathbb Z_2, $$ where he assigned to every closed curve $C$ representing an element of $H_1(Sigma;mathbb Z_2) $ if the normal bundle over $C$ is framed trivially or not (over a circle there are only two homotopy classes of trivializations of the trivial vector bundle since $pi_1(SO(n))=mathbb Z_2$ provided $ngeq 3$). Pontryagin assumed that $varphi$ is a homomorphism and concluded by a surgery argument that every surface $Sigma$ is framed bordant to the 2-sphere $S^2$ which would mean that the map $f$ is null homotopic (see here for more details and nice pictures!).

Later Pontryagin corrects his mistake here. He shows $$ varphi(C_1+C_2) = varphi(C_1)+varphi(C_2) + I(C_1,C_2), $$ where $I(C_1,C_2)$ is the intersection number of the two curves $C_1$ and $C_2$. Thus $varphi$ is a quadratic refinement of $I$ and one can associated the Arf invariant $A(varphi)$ to $varphi$ (see Wikipedia). This can be used to enumerated $pi_{n+2}(S^n)$.

Grunwald's Theorem (1933) says that an element of a number field is an $n$-th power if and only if it is locally almost everywhere. As anyone who studied number theory now should guess, there is a problem with even primes, as discovered by Wang in 1948. This resulted the the corrected Grunwald-Wang theorem.

In the Wikipedia link above, Tate is quoted as saying:

Some days later I was with Artin in his office when Wang appeared. He said he had a counterexample to a lemma which had been used in the proof. An hour or two later, he produced a counterexample to the theorem itself... Of course he [Artin] was astonished, as were all of us students, that a famous theorem with two published proofs, one of which we had all heard in the seminar without our noticing anything, could be wrong.

This thread doesn't seem to mention Gauss's first proof of the fundamental theorem of algebra, from 1799. He claimed it as the first really rigorous proof, but it had a topological gap discovered 120 years later by Ostrowski. I think this is pretty famous since I first heard of it in an introductory abstract algebra class. It is discussed a bit in Smale's article about the FTA and complexity theory, here (p. 4).

The following is very far from my areas of competence, but I think it would fit the bill if the authors are correct:

https://arxiv.org/abs/1801.06359 L. Rempe-Gillen, D. Sixsmith, On connected preimages of simply-connected domains under entire functions.

Quoting from the abstract:

Let $f$ be a transcendental entire function, and let $U,Vsubsetmathbb{C}$ be disjoint simply-connected domains. Must one of $f^{-1}(U)$ and $f^{-1}(V)$ be disconnected?

In 1970, Baker implicitly gave a positive answer to this question... It was recently observed by Julien Duval that there is a flaw in Baker's argument (which has also been used in later generalisations and extensions of Baker's result). We show that the answer to the above question is negative; so this flaw cannot be repaired. Indeed, there is a transcendental entire function $f$ for which there are infinitely many pairwise disjoint simply-connected domains $(U_i)_{i=1}^{infty}$, such that each $f^{-1}(U_i)$ is connected... On the other hand, if $S(f)$ is finite (or if certain additional hypotheses are imposed), many of the original results do hold.

For the convenience of the research community, we also include a description of the error in the proof of Baker's paper, and a summary of other papers that are affected.

No less a mathematician than Kurt Gödel was guilty of claiming to have proved a result that was accepted for decades, and even used by others, before being shown to be wrong. Stål Aanderaa showed that Gödel's argument was incorrect and Warren D. Goldfarb showed that the result itself was false. The claimed result was about the decidability of a class of formulas including equality; see here for more details.

I'm surprised that this one has not already been mentioned. Voevodsky wrote an article explaining that one of the main motivations for his interest in homotopy type theory and univalent foundations was his personal experience with incorrect results being widely accepted for many years. For example, a 1989 paper by Kapranov and Voevodsky on ∞-groupoids contained a false result that was accepted until Simpson published a counterexample in 1998 (and even then, it took many more years before the community fully accepted Simpson's counterexample).

I think that this is a particularly important example from a sociological or historical point of view, since it spurred Voevodsky, a "mainstream" mathematician, to take seriously computerized proof assistants, which had been (and perhaps still is!) regarded by most people as a specialized subject of little interest to most mathematicians.

The wronskian determinant of $n$ functions which are $(n-1)$ times differentiable is $$W(f_1,dotsc,f_n)=detbegin{pmatrix} f_1 & f_2 & dots & f_n\ f_1' & f_2' & dots & f_n'\ vdots & & & vdots\ f_1^{(n-1)} & f_2^{(n-1)} & dots & f_n^{(n-1)} end{pmatrix}.$$

It is clear that if $f_1$, ..., $f_n$ are linearly dependent, $W(f_1,dotsc,f_n)=0$. For some years, the converse was assumed to be true too, until Peano gave the counterexample: $f_1=x^2$ and $f_2=xcdot |x|$ are linearly independent, though $W(f_1,f_2)=0$ everywhere. Later, Bôcher even gave counter examples with infinitely differentiable functions.

Bôcher also proved that the converse holds as soon as the functions are analytic. Other conditions are also known for the converse to hold.

Engdahl and Parker describe the history of the wronskian [1]. For a nice proof of Bôcher's result, one can have a look at a paper of Bostan and Dumas [2].

[1] Susannah M. Engdahl and Adam E. Parker. Peano on Wronskians: A Translation, Loci (April 2011), DOI:10.4169/loci003642.
[2] Alin Bostan and Philippe Dumas. Wronskians and linear independence, American Mathematical Monthly, vol. 117, no. 8, pp. 722–727, 2010.

I guess one major example is that unique factorisation doesn't always hold in rings of integers of number fields.

Classical attempts at solving Fermat's Last Theorem resulted in moving to cyclotomic fields $mathbb{Q}(zeta_n)$ and noting that Fermat's equation factorises to give:

$$prod_{k=0}^{n-1}(x + zeta_n^k y) = z^n$$

an equality in the ring of integers $mathbb{Z}[zeta_n]$ of such fields.

Lame offered a full proof of FLT along these lines but a crucial assumption was unique factorisation. Kummer was able to provide a counter-example that shows non-unique factorisation for $n=23$ and this spurred off the process of inventing ideals as well as lots of other cool stuff in algebraic number theory.

Any rational function field over a finite field has genus $0$ and class number $1$, where the class number of a function field over a finite field is the number of degree-zero elements of the divisor class group. In 1975, Leitzel, Madan, and Queen proved there are exactly $7$ nonisomorphic function fields over finite fields with positive genus and class number $1$. Almost 40 years later, in 2014, Stirpe found an $8$th example (see http://arxiv.org/abs/1311.6318)! A precise gap was then found in the original proof, and once fixed the theorem is that there are $8$ examples (see http://arxiv.org/abs/1406.5365 and http://arxiv.org/abs/1412.3505).

Hilbert's sixteenth problem. In his speech, Hilbert presented the problems as:

The upper bound of closed and separate branches of an algebraic curve of degree n was decided by Harnack (Mathematische Annalen, 10); from this arises the further question as of the relative positions of the branches in the plane. As of the curves of degree 6, I have - admittedly in a rather elaborate way - convinced myself that the 11 branches, that they can have according to Harnack, never all can be separate, rather there must exist one branch, which have another branch running in its interior and nine branches running in its exterior, or opposite.

According to Arnold (see his book "What is mathematics?") Gudkov has found 3rd posiible configuration (exist one branch, which has 5 branches running in its interior and 5 branches running in its exterior).

Lebesgue famously "proved" that the projection of a Borel set in $mathbb R^2$ is a Borel set in $mathbb R$. Famously disproved by Souslin a decade later. See this answer by Gerald Edgar.

The Auslander Conjecture states: Every crystallographic subgroup $Gamma$ of $mathrm{Aff}(mathbb{R}^n)$ is virtually solvable, i.e. contains a solvable subgroup of finite index.

He published an incorrect proof in 1964 of this statement.

In 1983 Fried and Goldman proved Auslander’s conjecture for $n = 3$.

Abels, Margulis and Soifer proved the conjecture for $nleq 6$ in 2012.

Although it is not my area of expertise, I believe it is considered to be an important open conjecture and has led to active research.

The Steiner Ratio Gilbert–Pollak Conjecture was proved by Dingzhu Du and Frank Hwang in 1990, and published in Algorithmica in 1992. In 2001, a gap of the original proof was found by A.O. Ivanov and A.A. Tuzhilin. And the conjecture remains open now. See: http://link.springer.com/article/10.1007%2Fs00453-011-9508-3

This blog post, the previous one (linked inside) and the addendum caused by reader response, treat exactly this question, including Euler's polyhedra formula from Micah Miller's response.

Carmichael's totient function conjecture (stating that the equation $phi(x)=n$ never has a unique solution) was a theorem until an error was found in 1922 (apparently after the proof was left as an exercise in a textbook); since then, it is a conjecture. See: http://mathworld.wolfram.com/CarmichaelsTotientFunctionConjecture.html

In 1959 Kravetz published a proof that the Teichmuller metric on Teichmuller space is negatively curved in the sense of Buseman. This was widely quoted and used until Linch found a gap in 1971.

In 1974, Howard Masur showed that the Teichmulller metric is not negatively curved, by exhibiting two distinct geodesic rays which have a common starting point but stay a bounded distance apart. There is now a whole subfield studying Teichmuller geometry, which grew out of the failure of Kravetz's theorem.

Euler in his 1759 paper on knight's tours claimed that closed tours were not possible on any board with 4 or fewer ranks, though he gave no explicit proof. The claim was repeated by other influential writers such as E. Lucas and W. Ahrens. It was proved true for 4-rank boards by C. Flye Sainte-Marie in 1877. It was finally disproved by Ernest Bergholt in 1918 by constructing closed tours on 3x10 and 3x12 boards. Algorithms for enumerating tours on 3xn boards have now been devised by D. E. Knuth. This is a case of a famous mathematician's statements being taken as gospel and not really subjected to testing.

There are also numerous sources that state that Euler constructed a magic knight's tour on the 8x8 board. Where this mis-statement originated I'm not sure, but it has proved difficult to eradicate from the literature. In fact the first such tours were found by W. Beverley in 1848 and C. Wenzelides in 1849.

I heard that Bott's theorem on the periodicity of the stable homotopy of the unitary group was delayed for some time by an erroneous computation in dimension 10, possibly due to Pontryagin.

Another one from Grunbaum's paper:

According to Branko Grunbaum, An enduring error, Elemente der Mathematik 64 (2009) 89-101, reprinted in Mircea Pitici, ed., The Best Writing On Mathematics 2010, Bruckner enumerated 4-dimensional simple polytopes with eight facets, in 1909. But one of Bruckner's polytopes does not exist, according to Grunbaum and Sreedharan, An enumeration of simplicial 4-polytopes with 8 vertices, J Combin Theory 2 (1967) 437-465.

This thread on the Italian tradition in algebraic geometry contains some important examples.

According to Branko Grunbaum, An enduring error, Elemente der Mathematik 64 (2009) 89-101, reprinted in Mircea Pitici, ed., The Best Writing On Mathematics 2010, Andreini in 1905 claimed that there are precisely 25 types of uniform tilings of three-dimensional space. But in 1994 Grunbaum showed that the correct number is 28. The reference is B Grunbaum, Uniform tilings of 3-space, Geombinatorics 4 (1994) 49-56.

Grunbaum has yet more examples in this paper, which I may write up for this question.

According to Branko Grunbaum, An enduring error, Elemente der Mathematik 64 (2009) 89-101, reprinted in Mircea Pitici, ed., The Best Writing On Mathematics 2010, Daublebsky in 1895 found that there are precisely 228 types of collections of 12 lines and 12 points, each incident with three of the others. In fact, as found by Gropp in 1990, the correct number is 229.

The Gropp reference is H Gropp, On the existence and nonexistence of configurations $n_k$, J Combin Inform System Sci 15 (1990) 34-48.

Grunbaum has some other examples in this paper, which I may write up for this question.

The Busemann-Petty problem (posed in 1956) has an interesting history. It asks the following question: if $K$ and $L$ are two origin-symmetric convex bodies in $mathbb{R}^n$ such that the volume of each central hyperplane section of $K$ is less than the volume of the corresponding section of $L$: $$operatorname{Vol}_{n-1}(Kcap xi^perp)le operatorname{Vol}_{n-1}(Lcap xi^perp)qquadtext{for all } xiin S^{n-1},$$ does it follow that the volume of $K$ is less than the volume of $L$: $operatorname{Vol}_n(K)le operatorname{Vol}_n(L)?$

Many mathematician's gut reaction to the question is that the answer must be yes and Minkowski's uniqueness theorem provides some mathematical justification for such a belief---Minkwoski's uniqueness theorem implies that an origin-symmetric star body in $mathbb{R}^n$ is completely determined by the volumes of its central hyperplane sections, so these volumes of central hyperplane sections do contain a vast amount of information about the bodies. It was widely believed that the answer to the Busemann-Problem must be true, even though it was still a largely unopened conjecture.

Nevertheless, in 1975 everyone was caught off-guard when Larman and Rogers produced a counter-example showing that the assertion is false in $n ge 12$ dimensions. Their counter-example was quite complicated, but in 1986, Keith Ball proved that the maximum hyperplane section of the unit cube is $sqrt{2}$ regardless of the dimension, and a consequence of this is that the centered unit cube and a centered ball of suitable radius provide a counter-example when $n ge 10$. Some time later Giannopoulos and Bourgain (independently) gave counter-examples for $nge 7$, and then Papadimitrakis and Gardner (independently) gave counter-examples for $n=5,6$.

By 1992 only the three and four dimensional cases of the Busemann-Petty problem remained unsolved, since the problem is trivially true in two dimensions and by that point counter-examples had been found for all $nge 5$. Around this time theory had been developed connecting the problem with the notion of an "intersection body". Lutwak proved that if the body with smaller sections is an intersection body then the conclusion of the Busemann-Petty problem follows. Later work by Grinberg, Rivin, Gardner, and Zhang strengthened the connection and established that the Busemann-Petty problem has an affirmative answer in $mathbb{R}^n$ iff every origin-symmetric convex body in $mathbb{R}^n$ is an intersection body. But the question of whether a body is an intersection body is closely related to the positivity of the inverse spherical Radon transform. In 1994, Richard Gardner used geometric methods to invert the spherical Radon transform in three dimensions in such a way to prove that the problem has an affirmative answer in three dimensions (which was surprising since all of the results up to that point had been negative). Then in 1994, Gaoyong Zhang published a paper (in the Annals of Mathematics) which claimed to prove that the unit cube in $mathbb{R}^4$ is not an intersection body and as a consequence that the problem has a negative answer in $n=4$.

For three years everyone believed the problem had been solved, but in 1997 Alexander Koldobsky (who was working on completely different problems) provided a new Fourier analytic approach to convex bodies and in particular established a very convenient Fourier analytic characterization of intersection bodies. Using his new characterization he showed that the unit cube in $mathbb{R}^4$ is an intersection body, contradicting Zhang's earlier claim. It turned out that Zhang's paper was incorrect and this re-opened the Busemann-Petty problem again.

After learning that Koldobsky's results contradicted his claims, Zhang quickly proved that in fact every origin-symmetric convex body in $mathbb{R}^4$ is an intersection body and hence that the Busemann-Petty problem has an affirmative answer in $mathbb{R}^4$---the opposite of what he had previously claimed. This later paper was also published in the Annals, and so Zhang may be perhaps the only person to have published in such a prestigious journal both that $P$ and that $neg P$!

Here is a list of counterexamples to once accepted theorems on Clifford algebras.

Edit: The original link is broken, I now replaced it by a pointer to the wayback machine. Alternatively, here are two of Lounesto's articles:

P. Lounesto: Counterexamples in Clifford algebras with CLICAL, pp. 3-30 in R. Ablamowicz et al. (eds.): Clifford Algebras with Numeric and Symbolic Computations. Birkh"auser, Boston, 1996.

P. Lounesto: Counterexamples in Clifford algebras. Advances in Applied Clifford Algebras 6 (1996), 69-104.

In a 1966 paper (Rational surfaces over perfect fields, Publ. Math. IHES), Manin gave examples of cubic surfaces with Brauer group of order 2. In 1996, Urabe proved a conjecture of Tate on The bilinear form of the Brauer group of a surface (this is the title of his Invent. Math. 1996 paper) after noticing that Manin's examples, that were in contradiction with Tate's conjecture, were false (this he noted in Calculation of Manin's invariant for Del Pezzo surfaces, Math. of Computation 1996). Read a bit more on this story in Liu, Lorenzini, Raynaud, On the Brauer group of a surface, Invent. Math. 2005.

In 1803, Gian Francesco Malfatti proposed a solution to the problem of how to cut out three circular columns of marble of maximal area from a triangular piece of stone. Malfatti's solution was three circles that are tangent to each other and to the sides of the triangle (known as Malfatti circles). His solution was believed to be correct until 1930, when it was shown that Malfatti circles are not always the best solution. Then, in 1967, Goldberg showed that Malfatti circles are never the optimal solution. Finally, in 1992, Zalgaller and Los' gave a complete solution to the problem.

http://en.wikipedia.org/wiki/Malfatti_circles

In 1993, Pat Gilmer asserted as Theorem 1 of Classical knot and link concordance, that certain Casson-Gordon invariants vanish for all slice knots, which would be true if the kernel of the inclusion $H_1(M_K;mathbb{Z}[t^{pm1}])rightarrow H_1(N_D;mathbb{Z}[t^{pm1}])$ were a metabolizer for the Blanchfield pairing. There, $M_K$ is the $3$--manifold obtained from zero-surgery on a knot K and $N_D$ is the complement of a slice disc in $D^4$.
The statement was believed, and many papers based statements on this theorem, which was taken for granted. It looks plausible, and the similar-looking statements of Levine or of Cochran-Orr-Teichner are certainly true. But it was shown a decade later in Stefan Friedl's 2004 thesis, Sections 8.3 and 8.4, that Gilmer's proof assumes that tensoring with $mathbb{Q}/mathbb{Z}$ is exact, which is false. Stefan is forced to do something unnatural and ugly to get his results, and to show that for each choice of Seifert surface, the Casson-Gordon invariants in question vanish for all but a finite number of primes (Theorem 8.6).
I believe that Gilmer's theorem is still open, which is very irritating for people studying knot concordance; because surely it MUST be true, and it is quite fundamental.

The (in)famous Jacobian Conjecture was considered a theorem since a 1939 publication by Keller (who claimed to prove it). Then Shafarevich found a new proof and published it in some conference proceedings paper (in early 1950-ies). This conjecture states that any polynomial map from C^2 to C^2 is invertible if its Jacobian is nowhere zero. In 1960-ies, Vitushkin found a counterexample to all the proofs known to date, by constructing a complex analytic map, not invertible and with nowhere vanishing Jacobian. It is still a main source of embarrassment for arxiv.org contributors, who publish about 3-5 false proofs yearly. Here is a funny refutation for one of the proofs: http://arxiv.org/abs/math/0604049

"The problem of Jacobian Conjecture is very hard. Perhaps it will take human being another 100 years to solve it. Your attempt is noble, Maybe the Gods of Olympus will smile on you one day. Do not be too disappointed. B. Sagre has the honor of publishing three wrong proofs and C. Chevalley mistakes a wrong proof for a correct one in the 1950's in his Math Review comments, and I.R. Shafarevich uses Jacobian Conjecture (to him it is a theorem) as a fact..."

This question reminded me of the following article of A. Neeman with an appendix by P. Deligne:

Neeman, Amnon, A counterexample to a 1961 “theorem” in homological algebra, Invent. Math. 148, No. 2, 397-420 (2002). ZBL1025.18007.

Verma proved that the multiplicities of all simple modules in a verma module are 1 or 0. When BGG tried to repeat his proof for some other case they found an error. This led to the study of multiplicities in category O etc.

William Shanks's calculation of pi to 707 digits in 1873 seems to have been accepted for 72 years before Ferguson discovered an error in the 528th place.

R. B. Kershner's paper "On Paving the Plane," Amer. Math. Monthly 75 (1968), 839–844, announced the classification of all convex pentagons that tile the plane. Kershner said that "The proof...is extremely laborious and will be given elsewhere." As far as I know the proof was never published, but the claim was apparently accepted at least until 1975 when Martin Gardner wrote about the subject. Then, as explained in detail by Doris Schattschneider ("In Praise of Amateurs," in The Mathematical Gardner, ed. David A. Klarner, Wadsworth International, 1981, pages 140–166), Richard James III and Marjorie Rice found examples that had been missed by Kershner.

In the 1960s, John Horton Conway verified the Nineteenth Century efforts of Tait and Little to tabulate all the knots through alternating 11 crossings (1). He found several omissions and one duplication, but somewhat famously failed to discover another one. This error would propagate when Dale Rolfsen added a knot table (as Appendix C) in his influential book Knots and Links in 1976, based on Conway's work. This addition happened in spite that Kenneth Perko had noticed, in 1974, the other pair of entries in classical knot tables that actually represent the same knot (2). In Rolfsen's knot table, this supposed pair of distinct knots is labeled $10_{161}$ and $10_{162}$. Now this pair is called the Perko pair, for obvious reasons :)

(1) An enumeration of knots and links, and some of their algebraic properties, Computational Problems in Abstract Algebra (Proc. Conf., Oxford, 1967), Pergamon, pp. 329–358

(2) On the classification of knots. Proc. Amer. Math. Soc. 45 (1974), 262--266

(The information of this post is quoted from the following Wikipedia articles:

• http://en.wikipedia.org/wiki/Knot_theory#Tabulating_knots
• http://en.wikipedia.org/wiki/Perko_pair

)

EDIT: The episode I had in mind turns out to be the work of Robert Coleman repairing a gap in a paper by Manin about the Mordell conjecture over function fields. See comments by KConrad below, giving specific references. Note that this is not about a false result, it is about an accepted proof with a gap that was found 20 to 25 years later and repaired.

Original:Requesting assistance with a memory.

This being community wiki, I will give my vague memory. I think someone who was actually there could tell a good story. I have been searching with combinations of words on google with no success.

Anyhoo, when I was in graduate school at Berkeley in the 1980's, a professor, whom I think was likely Robert Coleman, told us a story about a celebrated result on "function fields" or the function field version of something... The accepted proof was by someone really big, on google I kept running across the name Manin but I am not at all sure about the name. Prof. Coleman decided to present the proof to a class/seminar. Partway through it became clear that the accepted proof just did not work. I have a sense that the class and professor were able to clean up the proof but I have no idea what publication may have come of this. There is also the chance that the seminar did not occur at Berkeley, rather at an earlier job of the professor concerned. Sigh.

So, there are a few ways this story could be filled in. Many MO people are students or postdocs at Berkeley, somebody could walk down the hall and ask Prof. Coleman if that was really him, and if so what actually happened, or ask Ken Ribet, etc.. Again, someone on MO with encyclopedic knowledge of every possible use of the phrase "function field" might be able to say. Or someone very old, yea, verily stricken in years, like unto me.

Finally, note that the title and text of the OP's question disagree a little, and people have posted both "results" that remained false and correct results with incorrect proofs. Also, my memory is really quite good, but I heard this story once and did my dissertation on differential geometry and minimal submanifolds.

(I don't have enough rep to comment on KConrad's answer, hence this additional answer.)

On the matter of Cauchy's "mistaken" proof that a convergent infinite series of continuous functions is continuous, Detlev Laugwitz argues in his paper "Definite values of infinite sums: Aspects of the foundations of infinitesimal analysis around 1820" (in particular pages 211-212) that Cauchy was well aware of the issue that $displaystylesum_{k=1}^{infty} frac{sin(kx)}k$ is not continuous at $x=0$, and that it's not a counterexample to his theorem.

Basically, Laugwitz argues that the mistake is not in Cauchy's proof, but in its interpretation by others; in particular, a direct translation of Cauchy's notions of infinitesimal quantities and convergence into epsilons and deltas fails to capture the intended meaning. The point is that Cauchy understood the series to converge for infinitesimal $x$ as well, which is tantamount to requiring uniform convergence in the modern sense. His line of reasoning can be made rigorous by using non-standard analysis.

Edit: To elaborate, here a faithful reproduction of Cauchy's theorem and Cauchy's (1853) discussion of this trigonometric series.

Theorem: Let $S_m(x)$ be the partial sums of a series on the interval $a leq x leq b$. If

• $S_m(x)$ is continuous for all finite $m$
• and $S_m(xi)$ converges to $S(xi)$ for all numbers $xi$ in the interval (including non-standard numbers!)

then the sum $S(x)$ is also continuous. (Continuity in the sense of Cauchy, which is defined with infinitesimals and also very sensitive to $x$ being non-standard or not, but that's not relevant here.) $square$

Now, consider the series $sum frac{sin(kx)}k$. It's not a counterexample to this theorem because it does not converge for infinitesimal $x$. Namely, let $n=mu$ infinitely large and $x = omega := frac1mu$ infinitesimally small. Then, the residual sum is

$$S(omega) - S_{mu-1}(omega) = sum_{k=mu}^{infty} frac{sin(komega)}k = sum_{k=mu}^{infty} frac{sin(komega)}{komega}omega approx int_{omegamu}^{infty} frac{sin t}{t} dt = int_1^{infty} frac{sin t}{t} dt$$

Clearly, the integral is finite and not negligible; hence, the series does not converge for $x=omegaapprox 0$.

One part of Hilbert's 16th problem is to determine whether a polynomial vector field in $mathbb R^2$, $$V(x,y) = (P(x,y),Q(x,y)),$$ has at most a finite number of limit cycles.

In 1923, Dulac published a paper supposedly proving this.

Around 1980–81, Ecalle and Ilyashenko independently recognized that the proof had serious gaps.

In 1991–92, Ilyashenko and Ecalle independently published (quite different) proofs that a polynomial vector field in the plane does indeed have at most a finite number of limit cycles.

See Ilyashenko's paper, "A centennial history of Hilbert's 16th problem".

(Many related questions remain unsolved, such as finding sharp or even good upper bounds for the maximum number of limit cycles in terms of the degrees of the polynomials $P$ and $Q$.)

The Euler Characteristic V-E+F has an interesting history. It was initially stated that, for all polyhedra,

$$V(ertices)-E(dges)+F(aces)=2$$

and its proof was widely accepted, until people found counter-examples.

Imre Lakatos' book Proofs and Refutations has an imagined dialogue between teacher and student giving arguments and counter-examples leading to the correct formulation, which, he explains in his footnotes, traces the actual historical development of the statement and proof of the theorem.

Mathematicians used to hold plenty of false, but intuitively reasonable, ideas in analysis that were backed up with proofs of one kind or another (understood in the context of those times). Coming to terms with the counterexamples led to important new ideas in analysis.

1. A convergent infinite series of continuous functions is continuous. Cauchy gave a proof of this (1821). See Theorem 1 in Cours D'Analyse Chap. VI Section 1. Five years later Abel pointed out that certain Fourier series are counterexamples. A consequence is that the concept of uniform convergence was isolated and, going back to Cauchy's proof, it was seen that he had really proved a uniformly convergent series of continuous functions is continuous. For a nice discussion of this as an educational tool, see "Cauchy's Famous Wrong Proof" by V. Fred Rickey. [Edit: This may not be historically fair to Cauchy. See Graviton's answer for another assessment of Cauchy's work, which operated with continuity using infinitesimals in such a way that Abel's counterexample was not a counterexample to Cauchy's theorem.]

2. Lagrange, in the late 18th century, believed any function could be expanded into a power series except at some isolated points and wrote an entire book on analysis based on this assumption. (This was a time when there wasn't a modern definition of function; it was just a "formula".) His goal was to develop analysis without using infinitesmals or limits. This approach to analysis was influential for quite a few years. See Section 4.7 of Jahnke's "A History of Analysis". Work in the 19th century, e.g., Dirichlet's better definition of function, blew the whole work of Lagrange apart, although in a reverse historical sense Lagrange was saved since the title of his book is "Theory of Analytic Functions..."

3. Any continuous function (on a real interval, with real values) is differentiable except at some isolated points. Ampere gave a proof (1806) and the claim was repeated in lots of 19th century calculus books. See pp. 43--44, esp. footnote 11 on page 44, of Hawkins's book "Lebesgue's theory of integration: its origins and development". Here is a Google Books link. In 1872 Weierstrass killed the whole idea with his continuous nowhere differentiable function, which was one of the first fractal curves in mathematics. For a survey of different constructions of such functions, see "Continuous Nowhere Differentiable Functions" by Johan Thim.

4. A solution to an elliptic PDE with a given boundary condition could be solved by minimizing an associated "energy" functional which is always nonnegative. It could be shown that if the associated functional achieved a minimum at some function, then that function was a solution to a certain PDE, and the minimizer was believed to exist for the false reason that any set of nonnegative numbers has an infimum. Dirichlet gave an electrostatic argument to justify this method, and Riemann accepted it and made significant use of it in his development of complex analysis (e.g., proof of Riemann mapping theorem). Weierstrass presented a counterexample to the Dirichlet principle in 1870: a certain energy functional could have infimum 0 with there being no function in the function space under study at which the functional is 0. This led to decades of uncertainty about whether results in complex analysis or PDEs obtained from Dirichlet's principle were valid. In 1900 Hilbert finally justified Dirichlet's principle as a valid method in the calculus of variations, and the wider classes of function spaces in which Dirichlet's principle would be valid eventually led to Sobolev spaces. A book on this whole story is A. F. Monna, "Dirichlet's principle: A mathematical comedy of errors and its influence on the development of analysis" (1975), which is not reviewed on MathSciNet.

Hilbert's 21st problem, on the existence of linear DEs with prescribed monodromy group, was for a long time thought to have been solved by Plemelj in 1908. In fact, Plemelj died in 1967 still believing he had solved the problem.

However, in 1989, Bolibruch discovered a counterexample. Details are in the book The Riemann-Hilbert Problem by Anosov and Bolibruch (Vieweg-Teubner 1994), and a nice popular recounting of the story is in Ben Yandell's The Honors Class (A K Peters 2002).

Euclid's proofs were accepted for two thousand years. Only in the late 19th century was it noticed by Hilbert and others that Euclid was making a lot of implicit assumptions and that if you don't make those assumptions the results are false. The text by Prenowitz and Jordan is a good source for details.

In 1882 Kronecker proved that every algebraic subset in $mathbb P^n$ can be cut out by $n+1$ polynomial equations.

In 1891 Vahlen asserted that the result was best possible by exhibiting a curve in $mathbb P^3$ which he claimed was not the zero locus of 3 equations. It is only 50 years later, in 1941, that Perron gave 3 equations defining Vahlen's curve, thus refuting Vahlen's claim which had been accepted for half a century.

Finally, in 1973 Eisenbud and Evans proved that $n$ equations always suffice to describe (set-theoretically) any algebraic subset of $mathbb P^n$

Kempe's "proof" of the four-color theorem springs to mind. Wikipedia says that Kempe published it in 1879 and it was proven to be incorrect by Heawood in 1890. As I recall, the flaw in the original argument was as follows: Kempe defined a structure on a planar graph called a Kempe chain, and argued that certain of these chains could not intersect. There was a subtle flaw in this argument (which I can't seem to find a decent explanation of) and it failed for certain large graphs - the chains can in fact intersect. Heawood provided a 25-node example of intersecting chains; the smallest counterexamples are the Fritsch and Soifer graphs on 9 nodes.

Edit: I didn't address the renown of Kempe's proof. Wikipedia says that it was "widely acclaimed" (interesting coincidence of wording) while Thomas 1998 provides an excellent history but says little on this matter. I don't know if this could be truly considered "widely acclaimed" based on an uncited Wikipedia entry.