# Misteaks of mathematicians

Misteaks.

(As remembered by Prof Walter Rudin in his autobiography, “The Way I Remember it”.)

Mathematicians are human. Humans make mistakes. Therefore…

This is no cause for alarm. I have no figures to back this up, but compared to the flood of published papers, the number of serious errors in the literature must be tiny. Most are probably caught by referees. And, if there is a serious error in a paper that is important enough to be studied by a significant number of interested mathematicians, that error will be discovered. Even better, the one who made it won’t be able to argue his way out of it.

There is an amusing article by Geoffrey K. Pullum in Natural Language and Linguistic Theory 5, 1987, 303-309, which compares this social aspect of mathematics with what happens in linguistics. It describes the story of Rourke’s claim to have proved the Poincare conjecture. (The article was sent to me by Catherine, my linguist daughter.)

My first encounter with this sort of thing started with the letter on the following page (typed a bit later after two paragraphs):

Needless to say, I was totally amazed. Here was Dieudonne, a world class mathematician and one of the founders of Bourbaki, not telling me, a young upstart, “you are wrong, because here is what I proved a few years ago” but asking me, instead to tell him what he had done wrong! Actually, it took me a while to find the error, and if I had not proved earlier (in J. Math. Mech. 7, 1958, 103-116) that convolution-factorization is always possible in $L^{1}$ I would have accepted his conclusion with no hesitation, not because he was famous, but because his argument was simple, and, perfectly correct, as far as it went.

He proved, correctly, that every convolution of non-negative functions coincides almost everywhere with one that is lower semicontinuous. But (and this is what he ignored) that function may have $+\infty$ among its values.

NORTHWESTERN UNIVERSITY

Evanston, Illinois.

58, Rue de Verneuil, Paris 7e (France).

Paris, December 17

Dear Professor Rudin:

In the last issue of the Bulletin AMS, I see that you announce in abstract 7311, p.382, that in the algebra $L^{1}(R^{n})$, any element is the convolution of two elements of that algebra. I am rather amazed at the statement, for a few years ago I had made a simple remark which seemed to me to disprove your theorem (Compositio Mathematica, 12 (1954) p.17, footnote 3). I reproduce the proof for your convenience.

Suppose f, g are in $L^{1}$ and greater than or equal to zero, and for each n consider the usual “truncated” functions $f_{n}=\inf(f,n)$ and $g_{n}=\inf(g,n)$; f(resp. g) is the limit of the increasing sequence $(f_{n})$ (resp. g_{n}), hence, by the usual Lebesgue convergence theorem, $h=f*g$ is a.e. the limit of

$h_{n}=f_{n}*g_{n}$, which is obviously an increasing sequence. Moreover, $f_{n}$ and $g_{n}$ are both in $L^{2}$, hence it is well known that $h_{n}$ can be taken continuous and bounded. It follows that h is a.e. equal to a Baire function of the first class. However, it is well-known that there are integrable functions which do not have that property, and therefore they cannot be convolutions.

I am unable to find any flaw in that argument, and if you can do so, I would very much appreciate if you can tell me where I am wrong.

Sincerely yours

J. Dieudonne

His argument proves that there are non-negative functions h in

$L^{1}$ that are not representable as $h=f*g$ with $f \geq 0$ and $g \geq 0$. (I had also observed this). In the general real-valued case, if $h=f*g$, each of f and g is a difference of two non negative functions, so that $f*g$ breaks into four convolutions of the type considered by Dieudonne. Of these, two are greater than or equal to zero, two are less than or equal to zero, and one may therefore run into the problem of subtracting $\infty$ from $\infty$ (which is at least as much as of a no-no as is dividing 0 by 0). Hence, one can no longer conclude that h coincides almost everywhere with a function of Baire class one, that is, with a real-valued function which is everywhere the pointwise limit of a sequence of continuous ones.

Dieudonne had fallen into the “without loss of generality” trap by restricting himself to $f \geq 0$ and $g \geq 0$, and tacitly assuming that the general case would follow.

Here is his  reply to my explanation.

Paris, January 12, 1952.

Dear Professor Rudin:

Thank you for pointing out my error; as, it is of a very common type, I suppose I should have been able to detect it myself, but you know how hard it is to see one’s own mistakes, when you have once become convinced that some result must be true!!

Your proof is very ingenious; I hope you will be able to generalize that result to arbitrary locally compact abelian groups, but I suppose this would require a somewhat different type of proof.

With my congratulations for your nice result and my best thanks, I am

Sincerely yours

J. Dieudonne

The factorization theorem was indeed extended, even further than he had hoped. When Paul Cohen saw my rather complicated proof about $L^{1}(\Re^{n})$ (the case n=1 was much easier) he said: “Aha, approximate identities” and quickly produced a very general factorization theorem in Banach algebras with approximate left identities (Duke Math. J. 26, 1959, 199-205). Ed Hewitt (Math. Scand. 15, 1964, 147-155) extended Cohen’s proof so as to include convolution operators on $L^{p}$ ($1 \leq p < \infty$). Every h in $L^{p}$ is $f*g$ with f in $L^{1}$, g in $L^{p}$.

In my next story, I was the one who goofed, but it ended well. This concerned the open unit ball B in the n-dimensional complex space $C^{n}$. A one-to-one holomorphic map from B onto B will be called an automorphism of B. (When $n=1$, B is the unit disc in C, and its automorphisms are the familiar Moebius transformations that send z to $e^{i\theta}\frac{(z-\alpha)}{(1-\overline{\alpha} z)}$ ). The automorphisms of B are also explicitly known for all n.

Can a space X of complex valued functions on B(or on the boundary of B)  Moebius-invariant (or $\mathcal{M}$ invariant) spaces of certain types (Duke Math. J. 43, 10976, 841-861) but the following was not answered:

Which closed subalgebras of C(B) are $\mathcal{M}- invariant$?

Here, C(B) is the algebra of all complex valued continuous (possibly unbounded) functions on B, with the topology of uniform convergence on compact sets, and with pointwise addition and multiplication.

The five obvious possibilities are: $\{ 0 \}$, the constants, the holomorphic functions on B, those whose complex conjugates are holomorphic and C(B) itself. The answer (Ann. Inst. Fourier 23, 1983, 19-41) is:

Theorem. There are no others.

I believe that this is the most difficult theorem that I ever proved. It was new even for $n=1$. I started with a proof of the one-dimensional case, and then used that to derive the same conclusion in n dimensions. Fairly soon after I submitted the paper, Malgrange, who was an editor of Annales Fourier, wrote that the referee did not understand how I passed from 1 to n. When I looked at it, I couldn’t understand it either! What I had written simply made no sense, and there seemed to be no way to repair it. I had to go back and instead of first dealing with the one-variable case I had to do the whole thing in n variables from the start. Fortunately, it worked. But it took a whole summer.

When I sent the corrected (much longer) version to Malgrange, I wrote that I should like to  thank the referee in the corrected paper, but only if I could mention his or her name. I saw no virtue in anonymous thanks. The referee agreed to this; it turned out to be my friend Jean-Pierre Rosay.

A few years later, we (i.e., the Madison Mathematics Department) wanted to invite Rosay for a whole year’s visit. I had a so-called Vilas Professorship which provided research funds for worthy projects. So I sent a request to the appropriate committee; to strengthen the case, I mentioned that not only I but several of my colleagues (Ahern, Forelli, Nagel, Wainger among them) would find him very stimulating. My letter was returned by no other than Irv Shain, the Chancellor, saying that Vilas Professorships were only for the benefit of those who  had one and for no one else’s, and that I should write a different letter, explaining how Rosay’s presence would benefit me. I did that, and as a clincher enclosed a reprint of the paper in which I thanked him. That did it.

Soon after he arrived, he and I started  to work on several questions about holomorphic maps. This resulted in a long paper. (Trans. AMS 310, 1988, 47-86), the first of several that we wrote together. Our collaboration went so well that I suggested that we ought to try to  keep him. We succeeded in this, and it could well be that my role in getting him to stay here was one of thei best things I ever did for the Department.

I know of only one totally absurd paper that was published in a respectable journal, namely the one by Nikola Pandeski in Math., Annalen 287, 1990, 185-192. In one variable, the “corona theorem” asserts that the open unit disc U is dense in the maximal ideal space of the Banach algebra of all bounded holomorphic functions in U. Its original proof, by Lennart Carleson (Annals of Math. 76, 1963, 547-559) involved a great deal of difficult “hard” analysis. A much simpler one was later found by Tom Wolff. But the n-variable analogue, which Pandeski claimed to have proved, is still wide open.

This paper appeared during a several complex variables conference in Oberwolfach. I heard that it caused great hilarity because nothing in it makes sense. For example, the proof starts by covering a sphere with a finite disjoint collection of small balls! I have heard several explanations about how this absurdity got into print, none of them convincing. I was quite annoyed about the whole affair because Pandeski attributed several absolutely false assertions to me, and because Granert, the editor who  had (mis)handled this paper refused to publish my protest.

********************************************************************************************

Never mind. To err is human …

Nalin Pithwa