The Development of Continuity
Dr. Paul Ehrlich
February, 2002

Abstract -- Why do we have the precise, but hard to learn at first glance, delta-epsilon definition of continuity and differentiability in our calculus texts ? During much of the 1800's, the most popular texts were free of this bothersome unpleasantness. The answer is partly that work by Joseph Fourier in his 1807 manuscript on the heat equation (using what are now called Fourier series to solve that equation) contradicted widely accepted (but erroneous) results for infinite series of functions which had been obtained without the full precision of the delta-epsilon approach and related concepts such as uniform convergence.


1. Fourier against the establishment

Here is how the concept of continuity was defined in one of the standard calculus texts of the 1800's, S. F. Lacroix's Traite elementaire de calcul differentiel et de calcul integral -- as translated in [B, p. 91] --

By the law of continuity is meant that which is observed in the description of lines by motion, and according to which the consecutive points of the same line succeed each other without any interval.
This definition was first given in the first edition of this text, published in 1802, and remained unchanged even up through the sixth edition, which was published in 1858.

In this same general time frame, the eminent French mathematician Augustin-Louis Cauchy was active as the Professor of Mathematics at the eminent French engineering academy in Paris, the Ecole Royale Polytechnique. As such, Cauchy had the duty of giving a lecture course in calculus to the student body. As is done even today in France, it was customary for lecture notes of such courses to be produced (nowdays by graduate students) and so it is possible to see what Cauchy himself taught by looking in his Oeuvres completes contained in any University library. With definitions such as the above imprecise notion to work with, it is entirely forgivable if errors were made enroute to the development of the rigorous calculus.

Here is one of the results Cauchy gave in his lecture course, hence it must have been widely believed. Let

f1(x), f2(x), ..., fn(x), ...

be a countable sequence of continuous functions. Then the sum

S(x) = f1(x) + f2(x) + ... + fn(x) + ...


is continuous.

We teach nowdays in calculus that the sum of two (and hence any finite number) of continuous functions is always continuous, but here Cauchy was asserting that this finite result could be extended to any COUNTABLY infinite number of functions !

The notational machinery with which Cauchy formulated his proof would be familiar to any reader of a current calculus text. Let Sn(x) denote the nth partial sum

Sn(x) = f1(x) + f2(x) + ... + fn(x)


and let Rn(x) denote the remainder

Rn(x) = S(x) - Sn(x) = fn + 1(x) + fn + 2(x) + ....



What is 100 % correct is that Sn(x) is a continuous function for any n as a finite sum of continuous functions.

As transcribed by Bressoud [B, p. 187], Cauchy gave the following proof of the result extended to infinite sums of continuous functions :

Let us consider the changes in these three functions when we increase x by an infinitely small value alpha. For all possible values of n, the change in Sn(x) will be infinitely small; the change in Rn(x) will be as insignificant (or insensible) as the size of Rn(x) when n is made very large. It follows that the change in the function S(x) can only be an infinitely small quantity. From this remark we immediately deduce the following proposition:

THEREOM I -- When the terms of a series are functions of a single variable x and are continuous with respect to this variable in the neighborhood of a particular value where the series converges, the sum S(x) of the series is also, in the neighborhood of this particular value, a continuous function of x.
According to Bressoud, Cauchy's claimed proof of this result can be understood in the following way. Since S(y) = Sn(y) + Rn(y) for any y in |R, we may derive the following inequality:

|S(x) - S(a)| = |(Sn(x) + Rn(x)) - (Sn(a) + Rn(a))|

= |(Sn(x) - Sn(a)) + Rn(x) - Rn(a)|


which is by the triangle inequality, less than or equal to

(*) |Sn(x) - Sn(a)| + |Rn(x)| + |Rn(a)|.


Now make the first term abitrarily small by the continuity of Sn (taking x close to a) and make the second two terms arbitrarily small by taking n as large as needed.

In December, 1807, Joseph Fourier submitted a manuscript for publication to the Institut de France in Paris, which in translation, had the title Theory of Propagation of Heat in Solid Bodies. Applying the separation of variables method to the heat equation (e.g., assuming that there is a solution of the form z = f(x) g(t) ), Fourier obtained solutions to the heat equation with initial condition 1 for -1 < x < 1 which he wrote in terms of what we now call Fourier series . A particular example is given by (in the choice in the above notation) of

fn(x) = (-1)n + 1(4/pi) cos((2n - 1)pi x / 2) /(2n - 1)


and S(x) the associated infinite sum. Clearly if any functions fn (x) are continuous for all real numbers, those formed out of cosines certainly are. Thus, according to the result of Cauchy's calculus lectures, the associated function S(x) should be continuous for all real numbers. However, Fourier's analysis of this function (cf. [Bressoud, p. 7] for a graph) revealed that it was a function of period 2 with jump discontinuities at the odd integers. Thus, for example, S(x) = 1 for -1 < x < 1, but then S(x) = -1 for 1 < x < 3, so S jumps from the value +1 to the value -1 at x = 1. Such behavior, of course, precludes continuity. On these (and other) grounds, scientists were reluctant to accept Fourier series for several decades.

In Bressoud, [B, pp. 188 - 190], a nice telescoping series is used to illustrate what goes wrong with Cauchy's analysis of (*) above. We choose

fk(x) = x2/[(1 + kx2) (1 + (k - 1)x2)]


which by partial fraction techniques may also be written as

fk(x) = [1/(1 + (k -1)x2] - [ 1/(1 + kx2)].


Accordingly, the partial sums Sn(x) display the telescoping behavior beloved of calculus textbook writers, and we have

(**) Sn(x) = 1 - [1/(1 + nx2)


hence, also,

(***) Sn(x) = [nx2]/[1 + nx2].


From (**), we have for any nonzero x that Sn(x) approaches 1 as n tends to + infinity, thus S(x) = 1 for any nonzero x. But from (***), we have Sn(0) = 0 for any n, so that S(0) = 0. Hence the power series S(x) with kth term defined as above converges for all x, but is discontinuous at x = 0, since S(x) = 1 for x nonzero, but S(0) = 0.

We are now in a position to investigate the terms of decomposition (*) in this particular example, with a = 0. Recall that Sn (0) = S(0) = 0, so that Rn(0) = 0 also. We obtain the expression

|Sn(x) - Sn(0)| + |Rn(x)| + |Rn(0)| = |Sn(x) - 0| + |Rn(x)| + |0|
= |Sn(x)| + |Rn(x)|
(****) = [nx2/(1 + nx2)] +[1/(1 +nx2)]


Intuitively, Cauchy's idea in this example is to make the first term in (****) arbitrarily small by taking x close to zero. But then notice that the second term in the expression (****) becomes arbitrarily close to 1, unless n can then be taken larger and larger. But doing that undoes the work accomplished in making the first term small by making x small, given n. From a slightly different viewpoint, we have bounded |S(x) - S(0)| by expression (****) in carrying out the Cauchy argument, but in this particular case, one can even observe by doing the algebra that (****) = 1, hence no juggling of x and n whatsoever will make (****) arbitrarily small and the proof fails.

What we see emerging here is the need for the concept of uniform convergence (which was not widely appreciated until the 1860's), and indeed, nowdays we teach in Advanced Calculus the result that if the infinite series S(x) converges uniformly in an interval (a,b) and each of the summands fn(x) is continous at every point of the interval (a,b), then the series S is itself continuous at every point of (a,b).

But what became of the Fourier manuscript, submitted for publication in December, 1807 ? Bressoud in his book reveals what happened at the beginning of Chapter 6 ([B, pp. 219]):

"In the spring of 1808, Simeon Dennis Poisson wrote up the committee's report on Fourier's Theory of Propagation of Heat in Solid Bodies. The conclusion was that it contained nothing that was new or interesting. Behind this opinion lay Lagrange's opposition to the admission of Fourier's trigonometric series and his conviction that they must not converge. In the following years, Fourier attempted to meet Lagrange's objections and to convince him that his series did in fact converge. In the meantime, he conducted experiments, comparing the predictions of his mathematical models with observed phenomena.

The problem of modeling the flow of heat was of concern to many scientists of the time. In 1811, the Institut de France announced a competition for the best explanation of heat diffusion. Fourier reworked his earlier manuscript and submitted it. Despite continuing objections from Lagrange, he was awarded the prize. Lagrange could not deny him the award, but he could postpone publication. Even after Lagrange died in 1813, Fourier's manuscript continued to languish at the Institut. Fourier began to prepare a book to disseminate his ideas.

After the second fall of Napoleon, Fourier came back to Paris as the Director of the Bureau of Statistics for the department of the Seine. He was back at the center of intellectual life. His book, Theorie analytique de la chaleur (Analytic theory of heat), appeared in 1822. That same year he was elected perpetual secretary of the Academie des Sciences, the highest of scientific honors. He used that position in succeeding years to encourage and promote the careers of emerging mathematicians. Gustav Dirichlet, Sophie Germain, Joseph Liouville, Claude Navier, and Charles Sturm were among those who received his assistance and would remember him fondly."

2. The Special Case of Power Series

In the case of a power series, in the notation of the previous section, we have

fn(x) = cn (x - a) n.


Also, in our calculus texts, we encounter the following result on differentiation and integration of power series, term by term (cf. [S, pp. 746 - 747]) --

THEOREM 2 -- If the power series Summation { cn (x - a) n} has radius of convergence R > 0, then the function

f(x) = co + c1(x - a) + c2 (x - a) 2+ ... + cn (x - a)n + ...

is differentiable (and therefore continuous) on the interval (a - R,a + R) and

(i) f '(x) = c1 + 2c2(x - a) + 3c3 (x - a)2 + ...


(ii) Integral{ f(x) dx } = C + co(x - a) + c1 (x - a)2/2 + ....


The radii of convergence of the power series in Equations (i) and (ii) are both R.


Unfortunately, for series of more general functions fn(x) than cn (x - a)n, the question of when term-by-term differentiation is permissable is much more complicated, and indeed, such questions were at the heart of some of the questioning of the legitimacy of Fourier's methods when his paper was being reviewed in 1807 - 1808 before this theory had been well understood or even developed.

Here is a typical result from the general theory (cf. [B, pp. 199]):


Theorem -- Let F(x) = f 1(x) + f 2(x) + ... + fn(x) + ... be an infinite series convergent at x = a and such that each fk(x) is differentiable at every point in an open interval I containing x = a. If

Summation { fk '(x) }


converges uniformly over the interval I, then F(x) is differentiable at x = a and

F '(a) = f1'(a) + f2'(a) + ... + fn'(a) + ...


For power series, with f k(x) = ck (x - a) k, and the wonderful elementary limit

lim (k) 1/k = 1 as k tends to + infinity


one has

lim sup (|ck|1/k) = lim sup ((k |c k|)1/k).


Hence, the radius of convergence of any power series is the same as the radius of convergence of the series of derivatives. At each point inside the interval of absolute convergence, we are inside an interval at which the series and series of derivatives both converge uniformly. Hence, the conditions required for "good" behavior of a series of functions, as given in the above sample Theorem, are always automatically satisfied in the case of power series, and the more complicated behavior of Fourier series, which was so troubling to the 19th century mathematicians, was not hinted at from their earlier understanding of power series.

References

1). Bressoud, David, A Radical Approach to Real Analysis, Mathematical Association of America, Classroom Resource Materials Series, Vol. 2, 1994.

2). Stewart, James, Calculus, Early Transcendentals , Fourth Edition, Brooks/Cole Publishing Company, New York, 1999.