Frechet vs. Caratheodory Emesto Acosta G.; Cesar Delgado G. The American Mathematical Monthly, Vol. 101, No. 4. (Apr. 1994), pp. 332-338. Stable URL http: flinks.jstor-org/sici%sici=0002-9890% 28 199404% 29 101 %3A4%3C332%3 AFVC%3E2.0.CO%3B2-5 The American Mathematical Monthly is currently published by Mathematical Association of America, Your use of the ISTOR archive indicates your acceptance of JSTOR’s Terms and Conditions of Use, available at bhupulwww.jstororg/about/terms.hunl. JSTOR’s Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of « journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial us. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http: www jstor.org/journals/maa. html, Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission, JSTOR is an independent not-for-profit organization dedicated to ereating and preserving a digital archive of scholarly journals. For more information regarding JSTOR, please contact support @jstor.org, bupsvwwjstororg/ Tue Oct 24 17:49:06 2006Fréchet vs. Carathéodory Ernesto Acosta G. and Cesar Delgado G. INTRODUCTION. When we read Kuhn’s paper, The Derivative a la Carathéodory [Kuh], we were very impressed how Carathéodory’s formulation of derivative simplifies the proofs of the basic differentiability theorems for real-val- ued functions of one variable, in particular the proof of the Chain Rule. We saw that the strength of Carathéodory's formulation relies on the concept of continuity and that the proofs strongly use the properties of continuous functions. Another advantage about this formulation is that it does not require the difference quotient, which is the key to generalize it to functions of several variables, Let us recall the usual definition of differentiability. Let f be a real valued function defined on Rand a be a real number. We say that f is differentiable at a ifthe limit im LILO rs exists. We can say this in another way. Let # be defined by fx) =f) 0 a Then, f is differentiable at a if and only if # has a removable discontinuity at a. This is « motivation to state the following characterization of differentiabilty Fis differentiable at a if there exists a function & which is continuous at a and such that f(x) ~ f(a) = b(x)(x ~ 2) The latter is Carathéodory’s characterization of differentiability [Car]. We will extend it to functions f: R" -» R and explore some of the advantages it has over Fréchet’s characterization [Fre]. The reader could see without difficulty that all we do in §2, §3, and §4 can be redone replacing R" and R" by two arbitrary Hilbert spaces, and that Carathéodory’s definition makes perfect sense in general linear topological spaces [Aco, Del]. qa 2. THE TWO FORMULATIONS. First we extend definition (1) to functions J: R" + R™ and second we show it is equivalent to the corresponding one in the Fréchet sense. If we read (1) thinking of f as a function from R* to R™ we have to think of x—a and f(x) ~ f(a) as points (or vectors) in R" and R™ respectively. The question is: what is (x)? When we multiply r-a eR" by A(x) we get fl) — fla) = B™. Thus we can think of @(x) as am X n matrix which gives us ‘2 good interpretation of (1). So, we consider ¢ as a function on R* taking values in the space Myc, of real matrices. 332 FRECHET Vs. CARATHEODORY [Aprilthe space Myycq Of real matrices. Now we can give the following definition: Let f : RX +R" and a © R". We say that f is differentiable at a if there exists a function 6: R" > Myxq Which is continuous at @ and satis- fies @ F(x) ~ f(a) = 6(x)(a — a) We call such a function ¢ a slope function for f at a. With respect to the second, let us recall Fréchet’s definition [Spi, page 14}: fis differentiable at a if there exists a linear transformation A: R" > R™ such that @) tim Wf) ~f@) ~~ ah jig a x = all If f satisfies (2) we will say that f is Carathéodory differentiable at @ and if satisfies (3) that f is Fréchet differentiable at a. However we have the following theorem which states the equivalency between (2) and (3). ‘Theorem 1. Every Fréchet differentiable function is Carathéodory differentiable and Proof: We can see without difficulty that (2) implies (3). In fact, if we assume the existence of , we have x) = f(a) — d(a)(x —a)| (x) — o(a))(x-a@ ies) fed dois 9) | _ Ma OU OM iets) -4cant ‘Due to the continuity of # at a we get (3). Let us show now that (3) implies (2). Assume that A exists and define $' by 1 6(2) = ipa) MO A=) 8a) +A, xe A, rea ‘We can see immediately from the definition of # that (xXx — a) = f(x) — fla). ‘We have to prove the continuity of at a. But fC) ~ fa) ~ a(x - a) 14x) - 6¢a)I and since f satisfies (3) we get (2). 3. A UNICITY THEOREM, We know that if A exists in (3), it is unique [Spi, Theorem 2-1], A is called the derivative of f at a and is denoted by Df(a). Unfortunately this uniqueness is not true for @ in (2), as we can see in the following example. Let f: 8? + R-be defined by f(x, y) =.xy, and pick a point presents tensor product, Le, if u € RM and vw € RY then uw © YE Myyy is defined by (wou (o-w)u Where 0 = w isthe inner product of v and w. 1994] FRECHET VS. CARATHEODORY 333,(a,b) © R?. Then F(x, 9) ~ fa, b) = (b,x)(x — ayy ~ b) =(,a)(4-a,y~ 6). ‘Therefore (x,y) = (b, x) and W(x, y) = (y, a) are two different slope functions for f at (a, b). However observe that (a,b) = (a,b) and we can ask if this is true in general. The answer is yes! ‘Theorem 2. If and qs are two slope functions for f at a then $a) = Wa). Proof: Assume that and y are two slope functions for f at a and let a(x) = C0) = YG). Then a x)(x a) and so In(a)(x~ a) | =[(n(a) ~ (2) ~ a) R" are differentiable at a € B" then af + Bg, a, B eR, is also differentiable at a and D(af + Bg)(a) = aDf(a) + BDg(a) Proof: Let f, g: R" + RM be differentiable at a € ®* and a, B € R. Then (af + Bg)(x) — (af + Bg)(a) = a f(x) — f(a)) + B(a(x) - a(a)) = (ad + BY)(x)(x - a), where and y are slope functions for f and g respectively. Since and are continuous at a, ab + BY is also continuous at a. Therefore af + Bg is differen- tiable at a and D(af + Ba)(a) = ad(a) + BY(a) = aDf(a) + BDg(a) ‘More impressive is the proof of the chain rule, Theorem 4. If fis differentiable at a and g is differentiable at f(a) then g © f is differentiable at a and D(a f)(a) = Del f(a) f(a). 334 FRECHET vs. CARATHEODORY [AprilProof: Let f: R" +R" be differentiable at aR" and let g: RR" be differentiable at f(a) € R". Then 8(f(2)) ~ 8(F(a)) = WF))(F(2) ~ F(a) W(f(2))O(x)(x ~ a), where @ is a slope function for f at a and y is a slope function for g at fla). Since and f are continuous? at a and ¢ is continuous at f(a) we have that & © f is differentiable at a. Even more D(g° f(a) = (f(a) 4(4) = Da( f(a) Df(a), To finish this paragraph let us prove the critical point theorem. ‘Theorem S. Let f be a real-valued function defined on an open set UCB. If fis differentiable at xq € U and f (x9) is a extreme value then Df(xo) = 0. Proof: Suppose that 9 € U is such that f(x) < f(x) for all x € U and that f is differentiable at xq. Then 0 = f(x) — f(x) = o(2)(2 — x0) 4) for all x € U, where @ isa slope function for f at xp, Let h be a fixed point in R", and ¢ small enough so that x9 + th € U for all ¢ © (—e,e). By (4) we have that (xo + thyh > 0 for all ¢ © (—e, e). Hence (xy + m)h>0, 16> 0; and (xo + th) <0, £<0, Since @ is continuous at xp we get rh = 0. But h was chosen arbitrarily, then 4p) = 0, Le, Df) = 0. Observe that formally there isa little difference between the proofs of theorems 3, 4 and 5 and those in [Kuh]. Compare them also with the proofs of the same in {[Spil. To finish this paragraph we invite the reader to compute with this new formulation the derivative of a constant function, the derivative of a linear function and to prove the Leibniz rule. 5, THE INVERSE FUNCTION THEOREM. We do not pretend our proof of the inverse function theorem to be easy, but we believe ours is more direct than most of the proofs of the same we have seen (see for example [Spi], [Apo]), and that the reader can follow it without getting lost in too many details. We start by extending the concept of differentiability at a point to differentiabil- in an open set, Let U be an open set in R" and f: U > R™. We say that f is differentiable in U if for each y € U there is a slope function , for f at y. In such a case we write (x, y) = 6,(x). Observe that (x, x) = Dflx). In the following we denote Df(x) by Ay. "Continuity of f follows immediatly from (2) 1994) FRECHET VS. CARATHEODORY 335Let us define now continuous differentiability: J is continuously differentiable in U if there exists a continuous function : U X U-* Myc, such that (5) f(2) = f(9) = ¥0, ya -y)- Observe that our definition is not the same as the conventional one: J is continuously differentiable in U if f is differentiable in U and Df is continuous in U. © One sees clearly that (5) implies (6). Let us show that (6) implies (5). Suppose that f satisfies (6) and define Yi UX U > Myson bY 1 Sony) = Hcp) £0) G9) © =} Hae HY des x It is clear that W(x, yXx ~ y) = f(x) ~ f(y) and that Plyxv-s and pla are continuous, where A is the diagonal of U x U. What itis left to prove is that q is continuous on A. If a € U, 1G) = f0) = .0= y)Ih w(x, 9) = W(a,a)l) < 211A, Agll + or 2A, — Aull + Hg ~ Aall where £ is in the segment [x, y]. The last inequality is obtained by the mean value theorem. We get then, from the continuity of A, the continuity of y. Therefore f satisfies (5). With definition (5) we can now prove the inverse function theorem. Theorem 6. Let f: U + R" be defined in an open set UC R*. If fis continuously differentiable in U and det ya, a) 0 for some point a © U, then there exist neighborhoods'V of a and W of f(a) such that f-!: W -» V exists, is continuously differentiable and (DFW F(a) = [f(a]. Proof: We give the proof in three steps : (A) f is injective in some neighborhood of a, (B) f is onto some neighborhood of f(a), and (C) f- is continuously ferentiable in some neighborhood of f(a). Let w be as in (5) and assume that det Y(a, a) #0, a € U. A) By continuity of y there is a neighborhood V of a such that det YC, y) # 0, W(x, ») is bounded and [yx, y)I-! is bounded for all x,y © V. By (5) we conclude immediately that f is injective in V. BY’ Let B,, be an open ball centered at @ and contained in V. By continuity of J, {(6B,)* is compact. By A, fla) € f(OB,). Let d be the distance from f(a) to This ep I taken from [Spi “aby, denotes the border of 8, 336 PRECHET vs. cARATHEODORY [AprilJ(OB,) and let By be the open ball centered at f(a) and radius 4/2. Let us fix y © B, and define the function g on B,° by a(x) = ly - SOP. ‘Then g reaches its minimum at some point x» € B, since it is continuous and B, is compact. By the definition of By, x» is not in aB,. Now, g is differentiable at x (8) ~ 8(%0) = lly ~ F(2)IP ~ lly ~ F(a) IP = y= F(a) ~F(%0))Wex0)(4 = 0) and (2) = 2y ~ f(x) ~ flxo)\Cx, x9) is continuous at x9. By Theorem 5 0 = B(x) = 2(y ~ f(x) )W( 40> ¥0) and therefore y = f(x9) (see A). Thus f maps B, onto B. ©) Restrict f to W= f-(B,) 0 B,. By A and B, f: W > By is bijective and then we can define f-!: B, > W. By (5), for z,w € By we have FA) F408) = [WF FO)" =»). 7) Now, since [W(f-(z), f-'w))I-! is bounded (see A), f-! is continuous and therefore continuously differentiable in B, by (7). Also by (7) (DF YF(a)) = [DF@)1™ We invite the reader to compare Spivak’s proof of this theorem with our proof, One can see that the proof of step (B) is the same in both, but the proofs of parts (A) and (©), especiatly part (C), clearly show the advantages of Carathéodory's formulation. CONCLUSIONS. From a pedagogical point of view Carathéodory’s characteriza- tion is interesting. In R, it enhances the geometrical sense of the derivative as the continuous approximation to the tangent line by means of secant lines to the graph of a function, showing that continuity is an essential element for differentiability. This formulation allows us to prove with economy and elegance, the elementary facts of differentiability, not only of real functions of real variable, but also of functions from &" to R™ where the arguments are the same ones used for functions in one variable, ACKNOWLEDGMENTS. We want to thank Carlos Rodriguez and Jlrgen Tischer for their interest fand helpful suggestion. We also thank Lais Recalde who pointed out the reference [Apo] where ‘Apostol uses CarathGodory’s formulation for rea functions of one real variable. REFERENCES (Aco,Del] Acosta, C. Delgado, La Derivada de Carathédory en Espacios Vectoraes Topoligeas, Preprint. ‘Tom M. Apostol, Andlits Matemético, Bitoril Reverté, Barcelona, (1986). Constantin Carathéodory, Theory of Functions of a Complex Variable, Vol. I, Chelsea, New ‘York, (1958). "B, denotes the closure of By, 1994) FRECHET vs. CARATHEODORY 337Ure] M, Fréchet, La notion de diffremele dans Vanalyse générale, C. R. Acad. Sei. (Pais, 180 (4925), 806-909, [Kuh] Stephen Kuhn, The Derivative ala Carathéodory, American Mathematical Monthy, Vol. 98, No.1, January, (199). Departamento de Mateméticas Universidad del Valle Apértado Aéreo 2188 Santiago de Cali COLOMBIA 255433 — 1 is Prime David Slowinski and Paul Gage have verfied yet another Mersenne prime. Using a Cray C916 at the Cray Research computing facility, they verfied that 2° — 1 js prime. The run time on 16 processors was about 30 minutes. To verify that a random integer of this size is prime would be well beyond the range of present computers (as well as those for years to come). Mersenne primes, however, are relatively easy to test using the Lucas-Lehemer test. To test N = 2? — 1, start with 4 and repeatedly square and subtract 2. Then N is prime precisely when, after p — 2 repetitions, the result is 0 modulo N. ‘Sounds easy. The problem is squaring (and reducing) numbers with thousands of digits. Performing multiprecision multiplication quickly requires some sophistication, and Slowinski and Gage use an algo- rithm of Schonhage and Strassen, implemented using the Fast Fourier ‘Transform. Why do they continue the search? Running this particular set of programs seems to be a good test of reliability for both hardware and software on new machines. In addition, of course, there is a fascination with finding the largest known prime—at least, largest for the moment. Slowinski and Gage comment that they have not checked all ‘exponents below 858, 433, There may be other Mersenne primes to be iscovered. But chances are that people who have the time and inclination (and a Cray) will forge on to higher values. After all, there’s nothing remarkable about finding the second largest known prime number. 338 PRECHET VS. CARATHEODORY [April