American Mathematical Monthly - 1972-6

THE AMERICAN MATHEMATICAL MONTHLY (FounDED IN 1894 By BeNsAMIN F. FINKEL) THE OFFICIAL JOURNAL OF THE MATHEMATICAL ASSOCIATION OF AMERICA VOLUME 79 NUMBER 6 CONTENTS What is a Reciprocity Law? oe . . . BLE, WyMAN 571 A Map of Sources, Sinks, and Saddles. aon D. M. JORDAN AND H.L. Porteous 587 Reconstructing the History and Geography of an Evolutionary Tree. ee ee ee ee . Davip SANKOFF 596 Lipschitzian Points . . . . . LE. M. ‘Berstey, A. P. Morse, AND D.C. PrFAFF 603 Professor Leo Moser — Reflections of a Visit. . . . . . .W.E. MIENTKA 609 MATHEMATICAL NOTES The Logarithmic Mean. . . 2. BG. Carison 615 On the Convergence of the L? Norm to the L* Norm. » RA HANDELSMAN | AND L 8. Lew 618 Extension of Mappings in "Finite Abelian Groups... . .K.D, WALLACE 622 A Proof of Gandhi’s Formula for the nth Prime . . CHARLES VANDEN EYNDEN 625 RESEARCH PROBLEMS The Hadamard Maximum Determinant Problem JoEL BRENNER AND LARRY CUMMINGS 626 The Union of Arithmetic Progressions with Differences not Less than k . R. B. CRITTENDEN AND C. L. VANDEN EYNDEN 630. ‘CLASSROOM NOTES Regularity as a Relaxation of Paracompactness . . «+. JAMES CHew 630 A Simple Example on Some Properties of Normal Random Variables Be ee ee ee . . JAVAD BrHBOODIAN 632 An Alternative to the Integral Test for Infinite Series +» + + + «GJ. PorTER 634 (Continued on inside cover) JUNE-JULY 1972MATHEMATICAL EDUCATION Mathematics Courses in 1984 . . . ae J.B. ROsseR 635 ee ners A erect) cee ee : . GARRETT BIRKHOFF 648 Undergraduate Mathematics Training in n 1984 — Some Predictions . arene rere rer nT TEE Ee TT Murray GeRsTENHABER 658 ELEMENTARY PROBLEMS AND SOLUTIONS... . - 7 ee ee es 662 ADVANCED PROBLEMS AND SOLUTIONS 2 2... 2) ee ee es 667 REVIEWS 2 2 ee News AND Norices. . Co es 696 MATHEMATICAL ASSOCIATION OF AMERICA | 1. ss ss ls ss O97 Mathematical Sciences Employment Register. . . 2... 2. 2 2. 697 Calendars of Future Meetings » 2 2. 2 2. eee ee. O88 NOTICE TO AUTHORS Specialized research is usually unsuitable; see Statement of Policy (vol. 76, p.2). Manuscript preparation: Please use the Manual for Monthly Authors (vol. 78, p. 1) and follow the format in current issues of the MONTHLY. Manuscripts should be typewritten, triple-spaced with wide margins; submit two copies and keep one for protection against loss. Backlog: Main Articles 11 months, Math. Notes 10 months, Research Problems 6 months, Classroom Notes 7 months, Math. Education 7 months. EDITORIAL CORRESPONDENCE AND MAIN ARTICLES: to Hartey FLANpers, American Mathe- matical Monthly, Tel Aviv University, Ramat Aviv, Israel (see Notice, vol. 77, 1970, p. 555); NOTES, ete. : to the corresponding Associate Editor; ADVERTISING CORRESPONDENCE: to RAouL HAILPERN, Mathematical Association of America, SUNY at Buffalo, Buffalo, N. Y. 14214; CHANGE OF ADDRESS and SUBSCRIPTIONS: to A. B. Wittcox, Mathematical Association of America, 1225 Connecticut Ave., N.W., Washington, D. C. 20036. HaRrLey FLANDERS, Editor ASSOCIATE EDITORS JOSHUA BARLAZ J. G. HARVEY SEYMOUR SCHUSTER E.R. BERLEKAMP ERIC S$, LANGFORD J, A. SEEBACH, Jr. JANE W. DI PAOLA P.D. LAX E, P. STARKE ROBERT GILMER ARTHUR MATTUCK. LYNN A. STEEN RICHARD GUY M. W. POWNALL JAMES WENDEL RAOUL HAILPERN GIAN-CARLO ROTA Annual dues for members of the Association (including a subscription to the American Mathematical Monthly) are $12.50. For nonmembers the subscription price is $18.00. PUBLISHED BY THE ASSOCIATION at Washington, D. C., and Menasha, Wisconsin, during the months of January, February, March, April, May, June-July, August-September, October, November, December, Second-class postage paid at Washington, D. C., and additional mailing offices. Copyright © The Mathematical Association of America (Incorporated), 1972 PRINTED IN THE UNITED STATES OF AMERICAWHAT IS A RECIPROCITY LAW? B. F. WYMAN, Stanford University 1. Introduction. The Law of Quadratic Reciprocity has fascinated mathematicians for over 300 years, and its generalizations and analogues occupy a central place in number theory today. Fermat’s glimmerings (1640) and Gauss’s proof (1796) have been distilled to an amazing abstract edifice called class field theory. As a graduate student I learned the great cohomological machine and studied ‘Artin’s Reciprocity Law, one form of which gives an isomorphism between two cohomology groups. A little later 1 read Shimura’s paper [19], called “A non- solvable reciprocity law,’ and couldn’t understand the title at all. Where were the cohomology groups? Why was Shimura’s theorem a reciprocity law? It was an embarrassing, but healthy ignorance, because it made me go back and figure out the number theory that lay behind all those cohomology groups. Such a reassessment is especially important nowadays, because it seems more and more certain that the next generalization of the Law of Quadratic Reciprocity will require new techniques, and nobody is quite sure which techniques will work. In this paper I should like to discuss reciprocity laws from a rather general but very concrete point of view. Suppose /(X) is a monic irreducible polynomial with integral coefficients, and suppose p is a prime number, Reducing the coefficients of {(X) modulo p gives a polynomial f,(X) with coefficients in the field F, of p elements. The polynomial f,(X) may factor (even though the original /(X) was irreducible). If f,(X) factors over F, into a product of distinct linear factors, we say that f(X) splits completely modulo p, and we define Spl(f}) to be the set of all primes such that f(X) splits completely modulo p. The general reciprocity problem we shall be considering is: Given f(X) as above, describe the factorization of f,(X) as a function of the prime p. Sometimes we ask for less: give a rule to determine which primes belong to Spl(f). This vague question is hard to make precise until it is answered. What is a ‘“‘rule’*? What is an acceptable method for describing the factorization of f,(X)? Anyway, a satisfactory answer to this unsatisfactory question will be called a reciprocity law. Quadratic polynomials are easiest to handle, and Section 2 shows how the usual Law of Quadratic Reciprocity gives a reciprocity law. (If it did not, our language would be all wrong.) Section 3 treats cyclotomic polynomials, and Sections 4 and 5 take up general results. It turns out that the reciprocity problem has been solved satisfactorily for polynomials which have an abelian Galois group, but that very little is known about polynomials whose Galois group is not abelian. Bostwick Wyman received his Ph.D. at Berkeley in 1966 under G, Hochschild and A. Ogg. He was an Instructor at Princeton for two years and has been an Assistant Professor at Stanford since then. He spent a year’s leave at the University of Oslo, His main research interest is algebraic number theory. Editor. sm572 BF. WYMAN [June-July For an arbitrary polynomial /(X) and a specific prime p, it only takes a finite number of steps to decide whether p is in Spl(f). Sections 6 and 7 give a description of an efficient algorithm for doing this calculation and report on results obtained for a family of quintic polynomials. These results probably do not constitute a reciprocity law, and the last section tries to answer the main question, “What is a reciprocity law??? Prerequisites. Section 2 assumes only knowledge of the Law of Quadratic Reci- procity. The later sections assume somewhat more: acquaintance with cyclotomic polynomials, Galois groups, and the division algorithm in polynomial rings. Parts of Sections 4 and 5 assume the rudiments of algebraic number theory, but they can be skipped. Notation. We use Z, Q, and C for the integers, rational numbers, and complex numbers, respectively. If q is a prime or prime power, then F, is the field with elements, If R is a ring, then R[X] is the ring of polynomials with coefficients in R; mostly we deal with Z[X] and F,[X]. 2. Quadratic Polynomials. Suppose that /(X)is an irreducible quadratic polynomial with integral coefficients. If p is a prime number, let f,(X) be the corresponding polynomial in F,[X] obtained by reducing the coefficients of f(X) modulo p. The reduced polynomial f,(X) can factor in one of three ways: (0) f,(X) = I(X)?, where I(X) is linear. () F(X) = 1X) + LQ), where 1,(X) and 1,(X) are two distinct linear polynomials. In this case we say that /(X) splits modulo p. (2) f,(X) is irreducible in F,[X]. In this paper we shall stick to polynomials of the form X? — q, where q is prime. If f(x) =X? —q, then Case (0) occurs modulo p when p=q, and also when p=2. (The prime 2 behaves strangely for quadratic polynomials.) To distinguish Cases (1) and (2) we need to know whether q is a quadratic residue modulo p. If q is a quadratic residue, and q = a? (mod p), we get X* — q =(X +a) (X —a) (mod p). This puts us in Case (1) if p #2. If q is not a quadratic residue, we are in Case (2). Using the Legendre symbol, and ignoring the prime 2 and the exceptional Case (0) (a widespread practice!), we summarize (1) X? —q splits modulo p if (q/p) = + 1. (2) X? —q is irreducible modulo p if (q/p) = — 1. Remember that we are trying to describe the set Spl(X? — q) of primes p such that X? — q splits modulo p, and now we know that p is in Spl(X? — q) if and only F(qip)= +1. The reader should still be skeptical, because this translation of the problem does not do much for us. The symbol (q/p) is not easy to evaluate, and besides, if we change p we have to start all over again. Since there are infinitely many primes p,1972) WHAT IS A RECIPROCITY LAW? 573 this naive approach requires an infinite amount of work to describe Spl(X? — q). Can we find a better description? Since q is fixed and p varies, things would be better if we could use the symbol (p/q) instead of (q/p). For fixed q, the value of (p/q) depends only on the residue class of p modulo q. There are only q residue classes, and therefore only q symbols to evaluate, This suggests looking for a relationship between (p /q) and (q/p) in hopes of using (p/q) to describe Spl(X? — q). Now you can guess where we are; we have sneaked up behind the Law of Quadratic Reciprocity. Legendre’s statement goes like this [10, p. 455 ff.]: Tutorem 2-1 (Law of Quadratic Reciprocity): Let p be an odd prime. Then 1. (A/D) — 1)", where P=4(p—1). 2. (2/p) =(— DR, where R =(p? — 1)/8. 3. If q is another odd prime, then (q/p)=(—1)"%(p/q), where P =4(p —1) and Q=4(q¢-1). Gauss gave the first proof of this theorem [6, Article 131 ff.], and a modern proof can be found in almost any number theory text, for example, Niven and Zuckerman [17, p. 74]. This venerable law is really exactly what we need to compute Spl(X? — q). We start with a less fancy but quite useful form of the theorem. ‘THEOREM 2-2. Let p and q be distinct odd primes. 1. If q = 1 (mod 4), then (q/p) = (P/4). 2. If q=3 (mod 4), then (q/p) (pees to’® . If q =3 (mod 4), then (q/p) = — (P/q) if p =3 (mod 4). The derivation of Theorem 2-2 from Theorem 2-1 is an easy exercise. Now we are ready to give a prescription for computing (q/p) for fixed q and variable p: First, compute (b/q) for all integers b such that 1 S$ b Sq — 1. Second, given p, find the b such that 1 $b Sq—1 and b =p (mod q). We have therefore (b/q) = (p/q). Third, use the tables in Theorem 2 to convert knowledge of (p/q) into knowledge of (q/p). Example 1. q=17. The squares modulo 17 are 1, 2, 4, 8, 9, 13, 15, and 16, so that we have (b/17) = + 1 for b equal one of these, and (b/17) = — 1 for b = 3, 5,6, 7, 10, 11, 12, or 14, That is (second step), (p/17) = +1 if and only if p= 1, 2,4, 8, 9, 13, 15, of 16 (mod 17). Finally, (third step), 17 = 1 (mod 4) so that (17/p) = (p/17). If we return to the language of polynomials splitting modulo a prime, we can say that pe Spl(X? — 17) if and only if p= 1,2,4,8,9,13, 15, or 16 (mod 17). That is, the set Spl(X? — 17) can be defined by “congruence conditions modulo 17.”574 B.F. WYMAN [June-July Example 2. q = 11. By finding the quadratic residues modulo 11, we conclude that (p/11) = +1 if and only if p = 1,3,4,5, or 9 (mod11). In this case 11 = (mod 4) so (11 /p) = + (p/11) with a sign that depends on the residue of p modulo 4. For example, 23 = 1 (mod 11), and 23 = 3 (mod 4), so that (11/23) = — (23/11) = — (1/11) = — 1. On the other hand, 89 = 1 (mod 11) but 89 = 1 (mod 4) and (11/89) = +(89/11) = +(I/I1) = +1. Using the Chinese Remainder Theorem, we sce that the value of (11/p) depends on the residue class of p modulo 44, and after some calculation we get: pe Spl(X? — 11) if and only if p=1,5,7,9, 19,25, 35, 37,39, or 43 (mod 44). In this case the set Spl(X? — 11) can be described by congruence conditions modulo 44, The results of the last two examples are actually quite general. THEOREM 2-3. Suppose that q is an odd prime. Then the set Spl(X? — q) can be defined by congruence conditions modulo q if q=1 (mod 4) and modulo 4q if q=3 (mod 4). Furthermore, Spl(X? — 2) can be described by congruence conditions modulo 8. In this theorem the phrase “congruence conditions” is interpreted as in the examples. The first part follows from Theorem 2-2, and the second part from Theorem 2-1, part 2. Details are left as an exercise for the reader. Théorem 2-3 shows that the Law of Quadratic Reciprocity gives a “reciprocity law” in the sense of Section 1. That is, it yields a nice description of sets Spl(/) for quadratic polynomials. In the next section we shall try to find such a reciprocity law for certain special polynomials (the cyclotomic ones) of higher degree. 3. Cyclotomic polynomials. Suppose ¢ is a primitive nth root of unity; for instance, { = e**""is one choice. Then the minimal polynomial of { over Q is written ®,(X) and is called the n-th cyclotomic polynomial. One knows that ®,(X) has coefficients in Z and has degree $(n), where ¢ is the Euler phi-function. It can be computed conveniently from the formula xX" -1 = |] o(x), ain where the product runs are all divisors of n, including 1 and n itself. For example, ©,(X) = X — 1, and if p is a prime, then X°— 1 =(X — 1) - @,(X) and OX) = XP + XP pee EK HL Proofs of these facts and more information about ®,(X) can be found in Lang [14, p. 206], van der Waerden [20, Sec. 53] and in many other algebra textbooks.1972] WHAT IS A RECIPROCITY LAW? 575 The goal of this section is a “reciprocity law”” for these cyclotomic polynomials. We want a description of the set Spl(®,(X)), and, just as in the quadratic case, the description will be given in terms of congruence conditions with respect to a modulus which depends on the polynomial. The theorem follows. TueoREM (Cyclotomic Reciprocity Law). The cyclotomic polynomial ®,(X) factors into distinct linear factors modulo p if and only if p= 1 (mod n). First we give a lemma about finite fields, and then use the lemma to prove the theorem. To avoid excessive notation we also use the symbol ®,(X) to denote the cyclotomic polynomial with coefficients reduced modulo a prime p. Lemma. Suppose p is a prime number, and a is an element of F, with a" If a! #1 for all proper divisors d of n, then X —a divides ©,(X) in F,[X]. Proof. The relation X"—1=[],,®(X) holds in F,, so that a"—1=0 =[Tun@,(a). Since F, is a field, it follows that ©,,(a) = 0 for some divisor m of n, and that a”— 1 = | Jij,a) = 0. This gives a” = 1 which can only happen if m =n. Therefore, ®,(a) = 0, and X — a divides ®,(X). Proof of theorem. Recall that the multiplicative group F* of non-zero elements of F, is cyclic of order p — 1. Therefore, F¥ has a cyclic subgroup of order n if and only if n divides p — 1, Such a subgroup has #(n) generators, so that F¥ contains $(n) distinct primitive nth roots of 1 (these generators!) if and only if it contains one, and this happens exactly when p = 1 (mod n). Now assume p = | (mod n), so that F, contains $(n) distinct primitive roots of 1. These must be roots of ©,(X), by the lemma, so that ©,(X) splits into a product of distinct linear factors, Conversely, assume that ®,(X) splits into linear factors modulo p. If these factors are distinct, then p cannot divide n (exercise: start from Lang [14, p. 206]), and it follows easily that X" — 1 also has distinct roots modulo p. Let a be a root of ®,(X) in F,, so that a" = 1. If d is the smallest divisor of n such that a‘ = 1, then ®(a) = 0 by the lemma. If d # n, the basic relationship X" — 1 = [],j,,(X) shows that a is at least a double root of X" — 1, a contradiction. Therefore a generates a cyclic subgroup of order n in F* , and p|n—1. This completes the proof of the “cyclotomic reciprocity law.” 4. Abelian polynomials. In the first two sections we saw that if (X) is a quadratic or cyclotomic polynomial, then the set Spl(/) can be described by congruences with respect to a certain modulus. This gives a rather precise solution to the vague “reciprocity problem.” Unfortunately, such a nice description of Spl(f) is not always possible. We can, however, describe exactly the set of polynomials for which congruence conditions give the answer we need. First we must recall some Galois theory. Associated to each polynomial of degree516 BF, WYMAN [June-July nis the root field Ky = Q(a,,+-,%,), Where a1,--+,a, are the complex roots of f(X). (We avoid the more common term, “splitting field,”” because of possible confusion with polynomials “splitting modulo p.””) The field K, is a finite Galois extension of Q, uniquely determined by f(x). The Galois group of K,/Q is often called the Galois group of f(X), and f(X) is called an abelian polynomial if its Galois group is abelian. The next theorem shows the importance this notion has for the reciprocity problem. ABELIAN POLYNOMIAL THEOREM. The set Spl({) can be described by congruences with respect to a modulus depending only on f(X) if and only if f(X) is an abelian polynomial. Why should Galois groups have anything to do with polynomials splitting modulo primes? What are “congruence conditions” exactly? Enough machinery is developed in the rest of this section to establish the importance of the Galois groups, and to give a precise form of the theorem. A complete proof is far beyond the scope of this paper. In fact, the proof of the theorem involves almost all of “class field theory over the rationals.’ Perhaps the best avenue for an ambitious reader is to work through a basic text in algebraic number theory, and then go on to the cohomological treatment in Cassels and Frohlich [3], or the analytic approaches of Lang [15], Weil [21], or Goldstein [7]. At this point we must escalate the prerequisites: the reader should be familiar with integral dependence, Dedekind domains, and the factorization of prime ideals in Galois extensions, or else be willing to suspend his disbelief. It is safe to skip this discussion and go on to Section 5. Let K be an algebraic extension of Q. The elements of K whose (monic) minimal polynomial has coefficients in Z make up the ring of algebraic integers in K, written Ox, The ring Ox is a Dedekind domain if K/Q is finite. If pis a prime in Z, the ideal pO, factors uniquely into a product of prime ideals: PO = Py P,. If P is one of the factors of p, the residue class ring Ox /SP is a finite field extension of Z|[pZ. This residue class field extension is cyclic, with Galois group generated by the Frobenius map @: $(a) = a? for alla in Ox/P. Except for a finite number of exceptions (called ramified primes) the 8, appearing in pO, are all distinct. If K /Q is Galois with group G, and p is not ramified, then for each §R; there is a unique o G such that o reduces to the Frobenius map modulo §;. This automorphism is called the Artin symbol corresponding to SB. We denote it by q, so that the defining formula is 9() = a%(mod P) for all ae Ox.1972] WHAT IS A RECIPROCITY LAW? 577 These Artin symbols cy are not good enough for our purposes. We need to define an Artin symbol ¢, corresponding to a prime number p ‘“‘downstairs.” This is not possible in general, because different choices of the ideal $8 may give different Gg in G. How are these various og related? If B and Q are two factors of pOx, then there is an automorphism t in G such that 1($) = Q. It turns out that og = toyz". All the og corresponding to a single p are conjugate, and we call this conjugacy class the Artin symbol corresponding to p. In the good case that G is abelian, we can identify a conjugacy class with its unique member, so that the Artin symbol for p is an element o, in G. EXERCISE, If you are familiar with number theory in quadratic fields, try to work out the Artin symbols for them. Start with the field Q(/q) where q is an odd prime, and identify the Galois group with {+ 1}. Check that after this identification, the Artin symbol o, is exactly the Legendre symbol (q /p). (Were you wondering why o, is called a “‘symbol’*?) What about more complicated quadratic fields? Finally, try to compute the Artin symbols a, for the cyclotomic field Q(é,,). (Goldstein [7, p. 96 ff.] is one of many possible references.) From here on, K/Q is an abelian extension with group G. We denote by Q* the multiplicative group of non-zero rational numbers, and we think of Q* as the (multiplicative) free abelian group generated by the primes. For a fixed field K, let TI G, called the Artin map. Can we find the kernel and image of this homomorphism? The image is easy to describe: the Artin map is surjective. We shall get some idea of the proof in the next section. What about the kernel? ‘The result here is more complicated and requires some more terminology. If a is an integer, the ray group I’, is defined as follows: a rational number r # 0 is in I’, if r can be written as c/d with c and d prime to a and c=d (mod a). Then the kernel of the Artin map for K/Q contains the ray group, for some a = p'+- pf, where p,,-+,P, are the ramified primes in K, and e; = 1. The two italicized statements above make up the Artin Reciprocity Law. Emil Artin conjectured it in 1923 [1, p. 98], and proved it in 1927 [1, p. 131]. (Artin worked over arbitrary number fields, not just over Q.) The theorem is central in all modern treatments of class field theory. It is proved in all the books recommended above, and in many others as well. We state it again for reference. ArRTIN Reciprocity Law: Let K/Q be a finite abelian extension with Galois group G, and let T be the subgroup of O* generated by the primes unramified in K. Then the Artin symbol gives a surjective group homomorphism o: T + G whose kernel contains the ray group T,, where a is an appropriate product of the ramified primes.578 B.F. WYMAN (June-July The Artin Reciprocity Law is a precise form of half the Abelian Polynomial Theorem: if f(X) is an abelian polynomial, then Spl(f) can be described by congruence conditions. To see why, we start with a crucial lemma. Lemma. Suppose f(X) is an abelian polynomial with root field K, Galois group G, and Artin map o:-+G. Then except perhaps for a finite number of exceptional primes, f(X) splits modulo p if and only if o, is trivial. Proof. We can only give an outline here. If p is unramified and pO, =B1-- By then the Chinese Remainder Theorem gives x/POx = @i=1Ox/ Bi. On the other hand, except for a finite number of p, Ox/POx =ELX FX), where f,(X) is the reduction of f(X) modulo p. (This is a hard exercise; the exceptions all divide the discriminant of f(X).) Therefore, except for finitely many p, FLX] (4X) © Oi=1%K/ Be When is the Artin symbol ¢, trivial? By definition o, induces the Frobenius map, xx’, on each direct summand. The Frobenius map is trivial on Ox/B only if Ox |B = Fp 80 that F,[X]/(/,(X)) = Ft when o, is trivial, and this is only possible when /,(X) factors into linear factors, All the steps are reversible, so the converse holds too. This: lemma, combined with the Artin Reciprocity Law, guarantees that the set Spl(f) contains all primes p such that p=1 (mod a), with at most finitely many exceptions. (Check this!) We need to change Spl(/) slightly at this point. Add to Spl(f) any primes p (mod a) not already there, and throw away any divisors of a. Call the resulting set this is the set we can describe by explicit congruence conditions. Let Q*(a) be the multiplicative subgroup generated by all primes p which do not divide a. (A fraction b/c in lowest terms is in Q*(a) if both b and c are prime to a.) Let S’ be the subgroup of Q* generated by S. The set S has been chosen so that T, ¢S’ €Q", and the importance of these inclusions comes out in the next lemma. Lemma. Q*(a)/T, = (Z/aZ)*, where (Z /aZ)* is the group of invertible elements in ZlaZ. Proof. Define 0:Q*(a)/T,-»(Z/aZ)* by 0(b/c) = be-! (mod a). Check as an exercise that @ is a surjective homomorphism with kernel exactly I’. This lemma supplies us with congruence conditions. Starting with Spl(f), pass to S, and consider the set 0(S’) of residue classes modulo a. A given prime p will lie in S if and only if its residue class modulo a lies in 6(S’). Since S and Spl(f) differ in at most a finite number of primes, we shall be content with this result.1972] WHAT IS A RECIPROCITY LAW? 379 Next we attack the other half of the Abelian Polynomial Theorem: If Spl(f) can be defined by congruences, then f(X) must be an abelian polynomial. We shall need a hard theorem which says (roughly) that the root field K, of f(X) is uniquely determined by the set Spl(/). We introduce some notation: If $ and T are two sets of primes, then SC*T means that except for at most a finite number of exceptions every member of S is a member of T. The precise statement is then: INCLUSION THEOREM. Suppose f(X) and g(X) are polynomials with root fields K, and K,, respectively. Then K, & K, if and only if Spl(g) S*Spl(/). Note the reversal! The similarity to Galois theory can be made very precise for abelian polynomials and is an important part of class field theory. The theorem itself holds for arbitrary f(X) and g(X). It is not hard to prove that K,& K, implies Spl(g) *Spl(/). The converse requires analytic techniques, and is a corollary of the Tchebotarev Density Theorem discussed in the next section. See Cassels and Fréhlich [3, Exercise 6.1, p. 362] or Goldstein [7, Theorem 9-1-13, p. 164] for a proof. ‘Assume now that Spl(f) can be defined by congruences modulo an integer a. ‘Actually we assume more: namely, that Spl(f) contains the ray group I’,. (Exercise: What’s the difference between these assumptions?) According to Section 3, I, is Spl(@,(X)), and the root field of ,(X) is the cyclotomic field Q(C,), which is abelian over Q. Since I’, © Spl(f), the Inclusion Theorem gives K, < Q(£,), so that K, must also be abelian over Q. One corollary of this discussion deserves special mention. KRONECKER’S THEOREM. Every abelian extension of Q is contained in a cyclotomic extension. Proof. Exercise: Combine the Artin Reciprocity Law with the argument above. (There is an elementary proof in Gaal [5, p. 242].) 5. General polynomials. The Tchebotarev Density Theorem. If /(X)is an irreducible polynomial in ZX] which is not abelian, then very little can be said about the set Spl(/). The best general result is a statement about the relative “‘size’” of Spl(/). First we describe a numerical measure of sets of primes called the density. Let II be the set of all prime numbers, and let TS II be any subset. For any real x 2 1, let card {peT|p z” on A, which we also denote by ¢, is a ring isomorphism useful in studying the structure of A. For example, if A = F, ® -- © F, (n summands), then @ =identity. More generally, the smallest d such that $= identity on A (the order of ) equals the least common multiple of the d;. Since x generates A as aring, the order is the smallest d such that (x) =x, so it is easy to compute. We shall see in the next section that the order can give a lot of information in special cases. In general, however, we need a refinement. Suppose y denotes the isomorphism in the Chinese Remainder Theorem: VAZk, O- Ok, Then it is easy to see that y(ker(@ -— D) =F, ®-- @F,, r summands, where I: A A is the identity map, and ker(¢ — 1) is the kernel of the linear trans- formation ( — 1): A> A. Similarly, ker? -D)=h ®- OL, where I, = Fy: if Fy2 Sk, and 1 = F, =k; otherwise. Therefore, ker(#? — 1) has F,-dimension equal to 2r—(the number of summands with d,=1). DEFINITION. For each integer i, let v;= nullity (¢! — 1) =dim(ker($' — D), where ‘dim’? denotes vector space dimension over the prime field F,. For each integer j, let 4; =the number of factors in the decomposition (*) which have dimension exactly equal to j. In this notation v, = r, the total number of factors, and v, = 2r — ,. The reader should verify that V3 = My + 2uy + 30 My — Ha) = 3r—2y, — Ha.1972] WHAT IS A RECIPROCITY LAW? 583 Generally, it is not hard to see that # Ye = kr — (k= Wy —(k — 2a — 2 — en This relationship is very important. Knowing the y; is the same as knowing d,,dy,-+,d, 80 they give the factorization of f,(X). On the other hand, we shall see below that the y, are relatively easy to compute. The reader should use equation (#) to verify the following inversion formula: (##) My = 2¥5 = Vena — Verte We summarize these facts in the theorem. THEOREM. Suppose, given A=F,[X]/(/,(X))=k,®- @k,, and let d; = [k::F,]. Let ¢ be the Frobenius automorphism of A, and let v; = nullity ($' — 1). Then r=v,, and there are exactly pj = 2v; — v;~-4 — ¥j41 summands with d; j=i,--,d. Here d is the smallest integer such that $* =I. This theorem forms the basis of an efficient algorithm, First compute the matrix [#] with respect to the basis {1,x,--,x""!} of A. (Berlekamp calls this the Q- matrix.) Then compute successively ; = nullity ({@]'— J). Finally, compute the 1, from the theorem, If, =, then p belongs to Spl(f), and in more complicated situations the ; give information about the Artin symbol belonging to p. Of course, we must examine this proposed algorithm, How hard is it? How long does it take? Can it produce significant results and lead to a better theoretical understanding of the problem? First of all I have to admit that it is completely unreasonable to do the algorithm by hand, I worked on f(X) = X5 — X —1 with p= 11 for an hour and could not make it come out. It is much easier to factor by trial and error when p is small, but large primes are impossible. Fortunately it is not too difficult to write a FORTRAN program which will do calculations in the ring A. Since A is an n-dimensional vector space over F, with a nice basis {1,x,x?,---,x""1}, its elements can be represented as a 1 x n FORTRAN array. The program written for the next section uses FORTRAN’s integer arithmetic and works modulo a variable prime p. The algorithm is very efficient in that the number of operations required to factor f(X) modulo p is proportional to log p. In fact, the only part of the algorithm that depends essentially on p is computing x? in the ring A. Abstractly speaking, how many steps does it take to compute x”? Certainly less than 2 - log, p, since x” can be computed by successively squaring together some multiplications by x. (Are you skeptical? If p= 23 = 10111 (binary), the steps are x, x?, x4, x5, x!®, x!4, x?2, x29, which requires 7 <2 log,(23) steps.) The fascinating subject of number theory algorithms and the time needed to do them is discussed in Lehmer [16]. Knuth [13, p. 388 ff.] goes into more detail and discusses algorithms very similar to this one,584 B.F. WYMAN (June-July 7, Numerical results. With the help of R. W. Latzer I have written a FORTRAN program to carry out the algorithm for the polynomials X* — X — a, where a isan integer. This is the “‘Bring-Jerrard Quintic”” which has the non-solvable Galois group (5) for general a, and in particular for a =1, and a =2. The program factored X$ — X — 1 for all p up to 23,099 in about two minutes, at which time the program overflowed the FORTRAN integer capacity. (I have learned that Professor J. D. Brillhart, using other methods, has factored many members of a more general family of quintics up to p = 1000.) If f(X) is any irreducible quintic polynomial, then f,(X) can factor in one of eight ways: Type 0: P| D(A) Type 1: Five linear factors 1/120 Type 23: (Quadratic) (Quadratic) (Linear) 15/120 Type 24: (Quadratic) (Three Linear) 10/120 Type 3: (Cubic) (Linear) (Linear) 20/120 Type 4: (Quartic) (Linear) 30/120 Type 5: (Quintic) 24/120 Type 6: (Quadratic) (Cubic) 20/120 Thé factors are irreducible and distinct, when displayed. The type is the order d of the Frobenius map when p does not divide D(f) except that Type 23 means (order = 2, nullity v, =3) and Type 24 means (order = 2, nullity v; = 4). Thus, no nullities have to be computed, except when the order = 2. The fractions give the density of primes of each type, according to the Tchebotarev Density Theorem. Finally, we give some examples of actual numerical results. 1 f(X) =X -X-1. (a) D(f) = 19-151, so 19 and 151 are bad, (b) The primes of Type 1 (those in Spl(f)) which are less than 23099 are 1973, 3769, 5101, 7727, 8161, 9631, 11093, 14629, 16903, 17737, 17921, 18097, 19477, 20759, 21727, and 22717. There are 16 primes in this list, giving a ratio of 16 /2350 0068, as compared with a density of 1/120 ~ .00833. (c) The primes less than 500 are classified as follows: Type 0. 19, 151. Type 1. None. Type 23. 67,71, 239, 251, 313, 421, 433, and 491. Type 24. 163, 193, 227, 307, 467, 487, and 499. Type 3. 11,41, 43, 47, 53, 107, 113, 179, 181, 191, 229, 281, 293, 311, 317, 347, 349, 373, 409, 457, and 463.1972] WHAT IS A RECIPROCITY LAW? 585 Type 4. 23,29, 31, 61,97, 101, 127, 131, 157, 173, 223, 241, 263, 269, 331, 359, 389, 439, 443, and 479. Type 5. 3, 5, 11, 13, 79, 89, 109, 137, 139, 211, 257, 337, 379, 397, 431, 449, and 461. Type 6. 2,7, 37, 59, 73, 83, 103, 149, 167, 197, 199, 233, 271, 277, 283, 353, 367, 383, 401, and 419. 2. f(X)=X5-X-2. (a) D(f) = 2* + 3109, so 2 and 3109 are bad. (b) The primes of type 1 less than 23099 are 229, 271, 1637, 2647, 2857, 3673, 6323, 7103, 8123, 8999, 11161, 12197, 14341, 14503, 14929, 17183, 18679, 19457 and 20563. There are 19 primes in this list, giving 19/2350 ~ .00809. 3. It is also possible to fix p and let the coefficient a in XS — X — a vary modulo p. So far I have done this for all p up to 239. For example, if p = 31, we get: Type 0. a= 11, 20 (mod 31). Type 1. None. Type 23. a =2, 3, 28, 29 (mod 31). Type 24. 0, 15, 16 (mod 31). Type 3. a=7,24 (mod 31). Type 4. a=1,5,8,9, 14, 17, 22, 23, 26, 30 (mod 31). Type 5. 6, 10, 12, 13, 18, 19, 21, 25 (mod 31). Type 6. 4, 27 (mod 31). 8. What is a reciprocity law? A general reciprocity law should provide a deserip- tion of the set Spl(/) associated with a polynomial f(X). The algorithm discussed in this paper is such a description, but few number theorists would consider it a reciprocity law. More is wanted, but the exact requirements are still vague and undefined. ‘A good general reciprocity law should specialize to the Artin Reciprocity Law in the case of abelian polynomials. A very good reciprocity law should include a one-to- one correspondence between certain sets of prime numbers and field extensions, giving more substance to the Inclusion Theorem in Section 5. Such a correspondence should generalize the known abelian theorems of class field theory. Y. Ihara [11] is beginning to make some progress toward this goal in the function field case. Even if a good correspondence cannot be set up, any reciprocity law must be set in a general framework, and should unify various kinds of number theoretic phenom- ena, The examples in Shimura [19] are related to the theory of elliptic curves, but they are very special, and it is not clear how to use them as a foundation for a general reciprocity law. (The specialist should look at Ihara’s discussion of this question.) I would like to mention briefly another direction of research which may lead to reciprocity laws. The Artin Reciprocity Law can be interpreted as a theorem about certain classes of analytic functions: see Artin’s original paper [1, p. 97] or the section ‘Abelian L-functions are Hecke L-functions’’ in Goldstein [7, p. 182]. There seem to be important non-abelian analogues to this viewpoint which involve group586 B. F. WYMAN representations and automorphic forms, and the interested reader should look at the introduction to Jacquet-Langlands’ book [12] or Shalika’s paper [18]. Finally, I have to confess that I still do not know what a reciprocity law is, or what one should be. The reciprocity problem, like so many other number theory problems, can be stated in a fairly simple and concrete way. However, the simply stated problems are often the hardest, and a complete solution seems to be far out of reach. In fact, we probably will not know what we are looking for until we have found it. This research was supported in part by a grant from the National Science Foundation under grant GP 29696. References 1. E, Artin, Collected Papers, Addison-Wesley, Reading, Mass., 1965. 2. E. R, Berlekamp, Algebraic Coding Theory, McGraw-Hill, New York, 1968, 3. J. W. S. Cassels and A. Frohlich, Algebraic Number Theory, Thompson, Washington, 1967. 4. H. Davenport, Multiplicative Number Theory, Markham, Chicago, 1971. 5. L. Gaal, Classical Galois Theory with Examples, Markham, Chicago, 1971. 6. C. F. Gauss, Disquisitiones Arithmeticae, transl. A. A. Clarke, Yale, New Haven, 1966. 7. L. Goldstein, Analytic Number Theory, Prentice-Hall, Englewood Cliffs, N. J., 1971. 8. G. Hardy and E. Wright, An Introduction to the Theory of Numbers, 4th ed, Oxford, 1960. 9, H. Hasse, Bericht iiber neuere Untersuchungen und Problemen aus der Theorie der algebrai- schen Zahlkérper, Teil I, Ia, Il, 2nd edit., Physica-Verlag, Wiirzburg, 1965. 10. A. M. Legendre, Recherches d’Analyse Indéterminée, Hist. Acad., Paris, 1785. 11. Y. Ihara, Non-abelian class fields over function fields in special cases, to appear in Proc. of the Intern. Congress of Math., Nice, 1970. 12. H. Jacquet and R. P. Langlands, Automorphic Forms on GI), Springer-Verlag Lecture Notes in Mathematics, No. 114, Berlin, 1970, 13. D. Knuth, The Art of Computer Programming, Volume 2: Seminumerical Algorithms, Addison-Wesley, Reading, Mass., 1969. 14, S. Lang, Algebra, Addison-Wesley, Reading, Mass., 1965. 15. ——, Algebraic Number Theory, Addison-Wesley, Reading, Mass., 1971. 16. D. H. Lehmer, Computer Technology Applied to the Theory of Numbers, Studies in Num- ber Theory, MAA, Prentice-Hall, Englewood Cliffs, N.J., 1969. 17. I, Niven and H.S. Zuckerman, An Introduction to the Theory of Numbers, 2nd ed., Wiley, New York, 1966. 18. J. Shalika, Some Conjectures in Class Field Theory, in AMS, Proc. of Symposia in Pure Math., Volume XX: Stony Brook Number Theory Institute, Providence, 1971. 19. G. Shimura, A non-solvable reciprocity law, J. Reine Angew. Math., 221 (1966) 209-220. 20. B, van der Waerden, Modern Algebra, Vol. I, Revised English Edition, Ungar, New York, 1953. 21. A. Weil, Basic Number Theory, Springer, New York, 1967.A MAP OF SOURCES, SINKS AND SADDLES D. M. JORDAN anp H. L. PORTEOUS, University of Hull The system of linear differential equations &, = ax, + bx2, y= cx, + dx, yields a plane flow which has been traditionally classified into one of a number of types including source, sink, saddle, spiral, centre, and node. We define a geomet- rical equivalence relation which gives this classification, the equivalence classes being called directed-orbit-types. Also we give a map of this classification using a topology which makes the set of these flows homeomorphic to R*. 1. Classifications. The system %, = ax, + bx,, y= cx, + dx, where a, b,c, R, gives rise to the plane flow @: R? x R- R? which sends (x,1) to the point g(t) where g is that solution of the system for which g(0) = x. We associate with this flow the matrix (2 3) and w denote the set of such flows by ®. For pictures of these flows see, for example, [1]. From now on, all flows considered are taken to be in ®. An orbit of a flow ¢ is a set {4(x,1):teR} for some x. By a ($,y)-mapping, where ¢ and y are flows, we mean a mapping of the plane which sends each orbit of onto an orbit of . Flows ¢ and are said to be orbit-equivalent if there is a (,)- homeomorphism of the plane onto itself. We give definitions of six types of flow which together make up the whole of ®, and we subsequently show that the six types are the equivalence classes under orbit- equivalence. A non-singular matrix gives a centre if it has purely imaginary eigen- values, a saddle if the determinant is negative, and a topological source otherwise. The flow given by a matrix of rank 1 has a line of equilibrium points, the other orbits forming a set of parallel lines; if these lines are parallel to the line of equilibrium points we call the flow a shear, otherwise a line. It is easy to show that, in the latter case only, the matrix has a non-zero eigenvalue. The zero flow, given by the zero matrix, is the only member of the sixth type. Consequently, in terms of the coefficients of the matrix, the type of a flow is determined by the sign, positive, negative or zero, of ad—be, and whether or not a +d, b —¢ are zero. Flows in different types cannot be orbit-equivalent because they provide different sets of answers to the following three questions: Are there infinitely many point or- D.M. Jordan studied category theory under A. Frdhlick at King's College, London, and he currently is a Lecturer at Hull. H.L. Porteous recently finished his Ph. D. studies at the Univ. of Warwick under D. B. A. Ep- stein and M. Shub. He was a Fullbright Scholar at Berkeley, and has held positions at Hull and Liverpool Universities. Editor. 587588 D.M. JORDAN AND H. L. PORTEOUS (June-July bits? Is there an orbit which is homeomorphic to R? Is there an orbit which is homeomorphic to R and is a closed subset of the plane? It is more convenient to postpone the proof that flows of the same type are orbit-equivalent. Orbit-equivalence gives a rather coarse classification, and a topological source is further classified by the number of its straight orbits: a spiral has no straight orbits, a node has four, a degenerate node has two and a star has infinitely many. Because of the correspondence between the straight orbits of a flow and the eigenvectors of the associated matrix, we can determine to which of these new types a topological source belongs if we know the sign of (a + d)*— 4(ad — bc) and whether or not b — ¢ is zero. We search for a relation whose equivalence classes are the nine types. A fairly subtle test for such a relation is that it distinguishes between stars and spirals and yet identifies all spirals. We show below that stars are distinguished from spirals by the relation which holds between flows $ and y when there is a (,1/)-ditfeomor- phism. Nevertheless this relation fails: to see this we introduce the useful relation of orbit-similarity. Flows $ and y are orbit-similar if there is a ($,y)-homeomorphism which is linear. Accordingly, two matrices give orbit-similar flows if and only if there is a non-zero multiple of one which is similar to the other. Now suppose that 4, ¢ © and that fis a (g,Y)-diffeomorphism, Then it turns out that ¢ andy are orbit-similar, the differential of f at the origin being a linear (¢,¥)-homeomorphism: we have failed to find a very simple proof of this plausible assertion but give a proof in section 4. In consequence, although stars are distinguished from spirals, a classification based on diffeomorphisms has infinitely many classes of nodes, saddles and spirals. We call a mapping dependable if it preserves linear dependence; in the case of a plane homeomorphism this means that lines through the origin are sent into lines through the origin. For any flow ¢ in® and any t in R, the mapping x (x, 1), being a linear homeomorphism, is dependable. We define flows $,¥ to be of the same orbit-typeif there is a dependable (¢, y)- homeomorphism. Thus flows in ® of the same orbit-type are orbit-equivalent and have the same number of straight orbits. Because linear homeomorphisms are dependable, orbit-similar flows are of the same orbit-type. Inspection of the similarity classes of the 2 x 2 real matrices shows that any two centers are orbit-similar, as are any two shears, lines, degenerate nodes or stars. Consequently we establish that there are exactly nine orbit-types with the following proofs that the nodes, spirals, and saddles, each form a single orbit-type. Again, inspection of similarity classes shows that each node is orbit-similar to the flow given by a diagonal matrix which satisfies d — a = 1 and a > 0. Any two such flows ¢, have restrictions s,s which are homeomorphisms from S' x R onto R?\{0}, where S# is the unit circle. Because a >0 for both ¢ and y, mapping the origin to itself extends ss ‘to a homeomorphism of the plane. Lines through the origin are fixed lines of this homeomorphism since both ¢ and w have the property1972] A MAP OF SOURCES, SINKS, AND SADDLES 589 that, in polar coordinates, = 4sin20. By the construction of ¢5,i/s, we see that sds! maps orbits of $ onto orbits of y. The same argument applies to the spirals because each spiral is orbit-similar to a flow for which 6 = 1 and a > 0. Further inspection shows that any two saddles are orbit-similar to flows ¢ and y (3) 6 0) respectively, where 2,1 > 0. Put given by Wy = {(x1%2) 2x1 BX > 0}, Wa = (482): m2 2 > 0}, denote the closures of W, and W, by C, and C3, and let L be the line common to W;, and W. The flows ¢ and have restrictions ¢,,Y,, which are homeomorphisms from L x {t: 120} onto W;. Since for each t in R the mappings x (x, )and xo (x,t) are dependable homeomorphisms, it follows that y),7 lis dependable. Al- s0%, = x, for both $andy ; so zhi! keeps vertical lines fixed. Consequently the identity on the x,-axis extends 7,67 'to a dependable homeomorphism of C,. On C, we can use a similar construction since the orbits of and y are respectively those of -1h 0 =1n 0 Cry CD ou O =a These homeomorphisms on C, and C;, both being the identity on L, produce a dependable homeomorphism of the first quadrant, which maps an orbit of @ meeting L at x to the orbit of y through x. The required plane homeomorphism is now obtained by symmetry. With our results on orbit-types we complete the proof that there are just six equivalence classes under orbit-equivalence by showing that any two topological sources are orbit-equivalent. Matrix inspection shows that two such flows are orbit-similar to flows ¢ and ip, whose orbits, apart from the equilibrium orbit, go outwards from the origin and meet S! just once. As before, ¢ and y have restrictions @s and Ws such that mapping the origin to itself extends ss ‘to a (¢,/)-homeomorphism. We now distinguish between inward and outward flows. Each orbit of a flow ¢ is made into a directed set by the relation, which we call the direction, in which (x, 1) = x whenever t 2 0. The direction is therefore an order on all the orbits, apart from the exceptional, periodic orbits. Flows ¢ and y are of the same directed-orbit- type if there is a dependable ($,y))-homeomorphism which is direction preserving. Thus a topological source may be a stable or unstable node, degenerate node or spiral, or a source or a sink, and a line may be stable or unstable. Now if flows ¢ and y arise from matrices kA and PAP™*, then590 D.M, JORDAN AND H.L. PORTEOUS [June-July Pd(x, 1) = W(Px, kt), showing that $ and y are of the same directed-orbit-type if k > 0. With our previous discussion this shows that the shears, saddles and centres form single directed-orbit- types, making just fourteen types in all. Because any eigenvalue of the matrix of a topological source has the same sign as a + d, it follows that a + d <0 fora sink or stable flow and a + d > 0 for a source or unstable flow. In the maps of © described below, six of the directed-orbit-types split further, each into a clockwise and an anti-clockwise component. It is straightforward to show that the resulting twenty types are obtained by defining @ and to be of the same type if there is a direction preserving dependable (¢,1/)-homeomorphism which preserves the sign of 6(x) for all non-zero x. Because 6(x) has the same sign as ex? +(d — a)x,x, — bx} when x # 0, it follows that b — ¢ > 0 for a clockwise flow and b — c¢ <0 for an anticlockwise flow. An alternative aproach for both the preceding concepts is to use orientation, first for the orbits, then for the plane; the same classification results. We feel, however, that a precise formulation of this would be unnecessarily elaborate here. 2. Topologies. Up to now we have regarded @ just as a set; by giving it additional structure, we can describe in more detail how it splits into directed-orbit-types. In fact we put a topology on it. There are several reasonable ways of doing this, but it turns out, because of the special properties of the flows in®, that the resulting topologies are all the same; we devote the rest of this section to showing this. We call (a,b, c,d) the coefficient vector of the flow arising from the system Xp =axy +x, fy =x, +dxp. We give® the coefficient topology by requiring that the correspondence between a flow and its coefficient vector is a homeomorphism. The distance d(¢, ) between flows ¢ and y is defined to be the distance between their coefficient vectors. Next, define a subset N of to be a neighbourhood of a flow ¢ if for some ¢, R, T, where ¢ > 0, N contains every flow y such that |¥@,0 - $6,0| <2, whenever |x| < R and | ¢| < T. This yields the compact open topology on ® regarded as a set of functions from R? x R to R?. Now the mapping (¢,(x, 1))=— (x, f)) from ® x (R?xR) to R? is continuous in the coefficient topology: this is a particular case of a general result frequently referred to as continuity of the solution with initial conditions and parameters. We deduce that the coefficient topology is larger than the compact open topology either from [2] or by using the uniform continuity of the mapping(¢,(x,)—> (x, ) on {Wi dy, 4) S 1} x (0): SR, |t|ST} for any RT.1972] ‘A MAP OF SOURCES, SINKS, AND SADDLES 591 We now show that the coefficient topology is smaller than the compact open topology. For any flow @ write (x,t) for the derivative at tof the mapping t~— $(x,1). Thus, if has matrix A, then $(x,1) = AG(x,1) for all x and t. Suppose now that and yp are flows which satisfy d(w, $) 2 2n > 0, let A and B be the matrices of $ and y, and let E equal B— A, It follows that ||E || >. Let x be a point such that |x| =1 and |Ex| =| E||, and take a neighbourhood N(x,26) of x where 26 < § and 26 || < 41. Accordingly, if we N(x,23), then |Ew — Ex| <4] E] and |Aw — Ax| < $n. Hence, if w,z € N(x, 28), then |(Bw — Az) — Ex| $ |Aw — Az| + |Ew — Ex| 0 and exp (Al|T) — 1 <6 ensures that $(x, 1) € N(x,5) whenever |t| $ T. From the previous inequality we therefore deduce that either [We 7) — 66, T)| > ST EI, or WO.9 — 6,0] > 6 for some ¢ satisfying |¢|<. Thus any neighbourhood in the coefficient topology is a neighbourhood in the compact open topology, and the two topologies are the same. From [2] again it now follows that another way of characterizing the coefficient topology on is as the smallest topology which makes the mapping (#,(x,1))—> #(~, ) continuous. Alternatively, let a subset N of @ be a neighbourhood of ¢ if for some @,R, T, where e > 0, N contains every flow y such that ¥O,0 — 6,0] + WO, - 60,0] <2, whenever |x| < R and |t|<7. It is clear that every neighbourhood in the coefficient topology is a neighbourhood in this new topology. Since, in the coefficient topology, the mapping (¢,(x, 0) ~> 4(x, t) is continuous, it is uniformly continuous on (W:d.9) $1} x (0: |x| SR, |t| ST} for any R, T; consequently the coefficient topology is larger than, and so the same as, the new topology.592 D.M. JORDAN AND H. L. PORTEOUS (June-July It is not satisfactory to give ® the smallest topology making ¢~— $(x, 1) continuous for each (x,1). This, the pointwise topology on ®, fails to make the mapping ($.(x, 1)» (x, f) continuous: to see this we show that the pointwise topology on ® is strictly smaller than the coefficient topology. Let D be the set of pairs (X, e) where Xis a finite subset of R? x R and ¢ > 0. Dis directed by saying that (X,2) <(Y,5) if and only if X ¢ Y and ¢ 2 6. It is easy to show that for each member (X,e) of D there is a centre whose distance from the zero flow is at least 1 such that |(x,1) — x| < for all (x,1) in X. It follows that this net converges to the zero flow in the pointwise topology but not in the coefficient topology . 3. Maps. Flows have the same directed orbits if their coefficient vectors are on the same open ray from the origin in R*. Since each such ray meets S°, the unit sphere in R*, just once, we see how ® splits into directed-orbit-types if we describe the splitting of the set Y of flows whose coefficient vectors are in S*. We give first a two- dimensional map of ¥ and later a more informative three-dimensional map. UNSTABLE SOURCE, DEGENERATE NODES 3SIMAI0ID STABLE. Fig. 11972] A MAP OF SOURCES, SINKS, AND SADDLES 593 Let v(#) be the coefficient vector of a flow @, let k be the orthogonal transforma- tion (a,b, c,d)» (b — c,a-+d,b + c,a—d)//2 and let p be the orthogonal projection (u,, uz, t3,t4)~— (wy, Uz). Define the continuous mapping « from ¥ to the closed unit plane disc D? by taking a(4) to be p(k(u(4))). We have shown that, if @ has coefficient vector (a, b, c,d), then the directed-orbit-type of ¢ is determined by the signs of ad — be, (a + d)? — 4(ad — be), b—c anda +d. Thus, if ¢ and k(u($)) = (u,, uz, Us, U4), the directed-orbit-type of ¢ is determined by the signs of 2u?+2u2—1, 2u?-+u3—1, uw, and u . Figure 1 shows the directed-orbit-type of the flows mapped by « to each point of D?. The following properties of « can be readily checked. Flows ¢ and w in are mapped to the same point of D? if and only if some rotation sends ¢ into y: that is, there is a plane rotation r about the origin such that (r(x), t) = r(@(x, #)) for all x and t. A flow ¢ is mapped to the boundary of D? if and only if ¢ is isotropic: that is, every rotation about the origin sends ¢ into itself. Because all stars are isotropic, this two-dimensional map has the misleading feature that shears and stars are both represented by two points although their dimensions in ¥ are different. The homeomorphism £ from ¥ to R? U {co} is constructed by taking A(¢) to be the stereographic projection of k(v(@)) from (0,0,0,1) to the hyperplane u, = 0, Figure 2 shows the directed-orbit-type of the flow mapped to each point of R°. The following properties of B can be readily checked. The matrix of ¢ has zero determinant if and only if B(¢) is on the torus obtained by rotating the vertical circle with centre(,/2,0,0)and radius 1 about theu3-axis. This torus together with its inside is the image of the whole of ¥ except the saddles. As before, isotropic flows are mapped to S!, the unit circle in the u,,u plane. Let # be a non-isotropic flow. Then the set of flows obtained by rotating ¢ is mapped to a circle which is linked with S* and lies in a plane through the u3-axis; the u3-axis together with co is to be regarded as such a circle. The surface 2u? +u2 —1=0, which corresponds to matrices with a single eigenvalue, is projected stereographically into a surface which meets the plane u, = 0 in two circles of radius ./2 with centres (1,0,0) and ( — 1,0,0). Because of the close connection between the type of a flow and standard proper- of the associated matrix, it is possible to regard both of the above maps as pro- ing information about the linear plane endomorphisms in their own right. ti vi 4, The failure of diffeomorphisms. We now give the promised proof that if f is a ($,W)-diffeomorphism, where $,y/€®, then the differential of f at the origin is a ($,)-mapping. In fact suppose that and y arise from matrices A and B and let f be a (¢,)-homeomorphism which is differentiable at the origin with a non-singular differential there. We will prove that this differential is a (4, /)-mapping. Itis straightforward to deduce this result from the special case, proved below, in which this differential is the identity mapping.594 D. M. JORDAN AND H. L, PORTEOUS (June-July SADDLES LINES SADDLES We must prove that ¢ and have the same orbits, Since the inverse of fis a(W, $)- homeomorphism which has the identity differential at the origin, it is sufficient to prove that each non-equilibrium orbit of ¢ is contained in an orbit of p. Further, it is enough to prove that, if Ax #0, then (x,t) is on the orbit of y through x for all sufficiently small 1. Suppose that Ax #0. Because the inverse of f has the identity differential at the origin and maps equilibrium points of y to equilibrium points of ¢, we deduce that Bx x 0. Hence there is a strictly positive number 6 such that, if we N(x,26), then |Bw — Bx| <4] Bx|. Because the mapping t»—> $(~x, t) is continuous, there is a number T such that T > 0 and, if |t|ST, then ¢(x,1)€ N(x, 6). Take anyewhich satisfies 0 max { — 0X, #), — 0(¥, W)}. Then X and W have a more recent common ancestor population Z than do X and Y, and Y and W have a more recent common ancestor Z than do X and Y. But by Definitions 1 and 2, the ancestors of W form a nested sequence of subsets of S. Therefore one of Z“ or Z®) must be a common ancestor to both X and Y, contrary to our supposition. Hence the ultrametric inequality holds. Proposition 4. Let S represent a finite set of populations existing at time zero. An ultrametric d on S determines a unique evolutionary tree where, if X, Y eS, then — d(X, Y) is the date of the most recent population ancestral to both X and Y. Proof. There are a finite number of different values of d, say 0 M, is derived from .#, by the subdivision of the face corresponding to X into the faces corresponding to X), ..., X™,1972] RECONSTRUCTING AN EVOLUTIONARY TREE 601 This is the simplest way of constructing planar graphs by extension and hence is the simplest model of the evolution of territorial configurations of related populations. How do populations interact? Instead of just replacing types at sites in T’ with completely new types, we now allow, in addition, the adoption of types from neighboring populations. Two neighboring populations are, of course, populations whose corresponding faces share an edge. If X and ¥ are neighbors, we write XeNyo YeNy. We can construct models where the total replacement rate is constant but the proposition of adoptions depends on the number of neighbors, other models where new replacements occur at a constant rate but the adoption rate depends on the number of neighbors, or models where the adoption rate is constant. Mathematically speaking, these all lead to the same type of problem, and so we study just the last one. We shall describe the genetic divergence process between two successive splits. In this interval , and Y, are fixed, so we can suppress the time subscripts on populations without risking ambiguity. For each population X, we assume a probability rate r for new replacements as before, and probability rate a/k(X) for adoptions from each of its k(X) neighbors. Suppose X ¢ Ny. Then ds(X,, ¥,)/dt, the rate of change in the similarity between the two simultaneously evolving populations X and Y, has several components. There is the change due to new replacements, into X and into Y; the change due to adoptions from X into Yand vice-versa; and finally the change due to adoptions into X and Y from their other neighbors. For the first component, the same arguments which justify (2) and (3) in the case of a single evolutionary line, also imply that the change rate due to new replacements into the two populations is — 2rs(X,, ¥,). (Were all other components zero, this could also be derived directly from Proposition 2, where t =u.) For the next component, the total adoption rate between X and Y is a(1/k(X) + 1/k(¥)) but a proportion s(X,, ¥,) of types adopted are already identical in the two populations so that the change rate due to this process will be (1 —s(X,, ¥,)) a(1/k(X) + 1/k( ¥)). In addition we must take into account adoptions from the remaining k(X) — 1 neighbors of X and the remaining k(Y) — 1 neighbors of Y, Adoptions from neighbors of X change the similarity at a rate ©) Wal X [A -s(X, Y)s(¥.2,) — s(X YA - (XZ), ZeNx-¥ following the same line of reasoning, and adoptions from neighbors of Y have an analogous effect, but with Y and X interchanged in (6). Collecting terms, we find that @ BEe WD. — x, Ys(%, YO +0, V,602 DAVID SANKOFF (June-July where 1 1 a 0 = alice; ecm] + RGTTT 2 Fy ‘ Y RAVAT ne tye 2H 8) BCX, Y)=2r+0(X,¥) + oe ZC -XoZo) + Zz (1-s(¥,Z)). 4 KY) —1 zeny-x For two populations X and Y which are not neighbors, coefficients «(X, Y) and A(X, ¥) are as in (8) but without the term a(1 /k(X) + 1/k(¥)). Proposition 5. Let {P,}°7 be an evolutionary tree with associated geography {.M,}°r. If genetic divergence proceeds according to (7), then Po, 8(Xo, Yo) for all Xo, Yo@Po, and Mo uniquely determine the tree and its geography. Proof. The graph Mo summarizes all neighboring relations between populations in Po. These relationships are fixed as far back as P, remains unchanged. Therefore we can write down equation (7) explicitly, with initial conditions s(X, Yo). The system of first-order equations so obtained satisfies conditions for a unique solution and can be solved by successive approximation. We write the solution as s’. Suppose the most recent population split was at time v, when populations Wand Z were formed from population {W,Z}. Immediately after v, and any time s(W,,Z,) is close to 1, AWE) <9, dt as can be seen in (6) or (8). Thus, s( W,, Z,) < 1 on (v,0]. But, by condition (1) at o, lim s(W,, Z,) = 1. tle Then time v, and the populations Wand Z can be found as v= max {t| 3X, YePo, s'(X,, ¥,) = 1}, and s’ = son(v,0]. The graph /, is then constructed by deleting the edge between the faces corresponding to W and Z, Note that by continuity lim s(X,, W,) = lim s(X,,Z,) for all XeS, de the since s measures proportions of types shared by two populations,1972] LIPSCHITZIAN. POINTS 603 We now have 9,, s(X,, ¥,) for all X, Ye, and M,. We can then set up a.new system of equations (7) with initial conditions s(X,, Y,) and solve as before. The new solution s” will be valid as far back as the second most recent split, and so on. The generalization to n-way splits, and the case where more than one population splits at the same instant, are obvious. We continue the solution procedure until we have deleted the last non-exterior line of the graph, which gives us —T and M_. This construction, for which each step is uniquely determined, proves the proposition. This last proposition means that a biologist, equipped with similarity data as well as a knowledge of the geographical configuration of a number of currently existing related populations, can reconstruct the entire evolutionary tree of the populations, as well as the geographical configuration at all times in [— T,0]. 4, Discussion. There are a number of practical problems associated with the theory of both Section II and Section III. One is that I" is too small to ignore statistical fluctuation. Another is that r and a are not universal constants but may change somewhat from site to site in and from population to population. The hypotheses about non-recurrence of innovation are not always justified. When these factors are taken into account, the reconstruction methods we have described must be bolstered by search algorithms and statistical estimation. Some useful references are: Dayhoff [2], Sokal and Sneath [3], Lerman [4], and Jardine and Sibson [5]. References 1, William Feller, An Introduction to Probability Theory and Its Applications, 2nd ed, Wiley, New York, 1957. 2. M.O. Dayhoff, Computer analysis of protein evolution, Scientific American, 221 (1969) 86-95, 3. Robert R. Sokal and Peter H. Sneath, Principles of Numerical Taxonomy, Freeman, San Francisco, 1963, (2nd ed. forthcoming), 4, I. C. Lerman, Les Bases de la Classification Automatique, Gauthier-Villars, Paris, 1970. 5. N, Jardine and R. Sibson, Mathematical Taxonomy, Wiley, New York, 1971. LIPSCHITZIAN POINTS E. M. BEESLEY, University of Nevada, Reno, A. P. MORSE, University of California, Berkeley, and D. C. PFAFF, University of Nevada, Reno 1, Introduction. Notations are explained in the next two paragraphs and the first paragraph of Section 2, Throughout we understand: that R is the set of real finite numbers; that Rp is the set of real finite positive numbers; that @ is the set of nonnegative integers; with fractions in mind, that F is the set of rational numbers; that J is the open604 E, M, BEESLEY, A. P, MORSE, AND D.C. PFAFF (June-July unit interval; and, for re F, that den r is the smallest positive integer q such that, for some integer p, r = p/q. We shall assume that W and w are such functions on J that, for each xeJ, W(x) = {reF:0 Sr Sx} and w(x) = YZ (denr)-3. remix) On page 408 of [2] we find the following theorem: Fort’s THeoreM. If f is on R to R with dense discontinuities, then the points at which f is differentiable form a set of the first category. Boas, [1, pp. 126-7] has reaffirmed this theorem and Heuer [3] has modified Fort’s proof to obtain the result which follows. Hever’s THeoreM. If f is on R to R with dense discontinuities and if0 5, tox Q) A is of Lebesgue measure zero; @) if xeB, then xeA and Tim 99 — 90) = tox = Proof. Letting M= Zs@ ant and observing, for x J, that ao =e Mend 2 FIO yy <0, remy) denr ~ ger q we notice that g is a bounded increasing pure jump function on J to Rp. Since such a function must have zero derivative almost everywhere in J [1, p. 130], we see that (2) is a consequence of (1). We see that (1) and (3) are consequences of the statement below. SrateeNt. If xeJ, 2€Rp, and there are infinitely many r¢F for which f(denr) denr a|x then x is irrational and g(t) — 96 t—x lim tox. Proof. From F we can and do select such a univalent sequence y that, for new, (den y, deny, = 1 (4) 0<2-+|y,—x] < We infer that lim deny, () lim f(den y, lim y, = x608 E.M. BEESLEY, A. P. MORSE, AND D.C. PFAFF We let z be such a sequence that, for new, z= 2+y,—x. Hence for new, 2 — X= 2°(n— Xp Yn = (Z + X)/2 Ip is strictly between z, and x, Accordingly lim, z, = x. Also, for sufficiently large new, z,¢J and _ f(denr) f(denr) | (2) - 9)| = | Re dear aden a Consequently jim a) — 9x) = 42. tox FOX If x were rational then the multiplication of the first two inequalities in (4) by the positive factor den y, *denx 2 and the use of (5) would lead us by well-known reasoning to the absurdity that 150. Application 3.2. Herein we suppose f(x) = x-? whenever x Rp. Accordingly g = wand the Liouville number ¥%_9(1/2)" belongs to A. The invalidity of 1.1 follows from 3.1(1). Moreover, if x is any Liouville number in J, then x€B and Application 3.3. Turning to Khinchin’s Theorem, we let K = {xeR: there are infinitely many r¢F for which |x —r| y. The left-hand inequality was stated by Carlson [3, Eq. (3.1)], who obtained (2) by specializing some rather general integral inequalities to the case of the repre- sentation 1 1 du wy Ly) if ux+(—wy" In the present note we first prove and sharpen (2) by an elementary method which treats x and y symmetrically. THEOREM 1. If the positive numbers x and y are unequal, then : Es @ out < yy EYE < 165,99 < (24) ap Proof. If t > 0 the inequality of the arithmetic and geometric means implies that Prieens LY >P+ixty)txy >? + 2uxy)' + xy. 615616 B,C, CARLSON (June-July Thus cf dt fr af? at o (22) o GFHE+Y) i} (t+ yxy)?” 2 Evaluating the middle integral by the method of partial fractions, we find cL tim flogtt +) — loge +8 < x+y Y Row cee 0 xy" which implies (2). We now sharpen (2) by replacing x by /x and y by //y: AYx- VV). VE +vy logx — logy 2° @yyF < Multiplication by (\/x + /y)/2 proves the two inner inequalities in (4). The two outer ones follow from the inequality of the arithmetic and geometric means. The process by which (2) was sharpened can be repeated to obtain (8). Instead of taking this route we prove a more general inequality first. For any real t 4 0 and any positive x and y, we define x+y x-y x-y _ xy Ay) =t > = ye Gx, y) = t(xy}? x#y, ©) G(x,x) = A((x,x) = x. If we further define G(x, y) = Ao(x,y) = L(x,y), it is easy to verify that G, and A, are continuous in t. They are also positive and even in t. THEOREM 2. If x and y are positive and t is real, then © Gx, y) < Lx, 9) < A(x y), x —y) #0. The first and third members are respectively decreasing and increasing functions of |t|, and the sharpness of the inequalities is measured by M APY) — GY) = 4° — yy. Proof. In (2) replace x by x' and y by y* and multiply by the positive quantity 1(x — y)(x' — y') to get (6). By straightforward calculation, dG, _ 4) dA, Gi ea Gt - 4), ETAT from which it follows by (6) that A, increases with | t | while G, decreases. Incidentally, a second differentiation shows that A, is convex and 1/G, is log convex in t. Coroutary 1. If x and y are positive and unequal and n is a nonnegative1972] MATHEMATICAL NOTES 617 integer, then ®) ca" TY aul) <9) < 0465.99 FT ant), where Onl) = ~ The products are taken to be unity if n = 0. The first and third members of (8) are respectively increasing and decreasing functions of n, and the difference of their squares is 2-**~?(x— y)*. ~" in Theorem 2 and note that Proof. Choose t = soys yy TT ey). ‘The inequalities (8) reduce to (2) if m = 0 and to the inner inequalities of (4) if n = 1. As n> 00 we obtain the following infinite product. Coro.iary 2. If x and y are positive numbers, then 2) Ley) = TT amGey)- ‘An: equality or inequality for L(x,y) of course implies a corresponding result for logx obtained by putting y = 1. For example, (6) gives (10) 140, x>1, with reversed inequalities if 0 0, 7 es A logx =(x—1) [| —~—. (uy ex @- DT Finally we give an algorithm for computing L(x, y) or logx by recurrence relations. As t+ 0 we find by developing (5) in powers of ¢ that A, and G, differ from L by terms of order 7”, but (12) Lx, y) = 4{A(x, y) + 2G (x, y)} {1 + 6%, y)}, where 5, is of order *. Since (13) Ayn = HA + Gs Gyr = (AG)? extraction of one square root cuts t in half and ultimately reduces the fractional618 R.A, HANDELSMAN AND J. 8. LEW (June-July error 6, by a factor of 16. For small t it is difficult to calculate A, and G, directly from (5) owing to cancellation in x‘ — y', but use of (13) avoids this problem. We define a, = A, and g, = G,, where t = 2'~", and proceed as follows. AvcoritHM. If x and y are positive numbers, let = = dl (14) a, = x+y), a1 = oy, nv1 = HAn +n)» Inv = nsiGn)*s = 1,2,3,-°. Then the common limit of a, and g, as n -+ co is the logarithmic mean L = L(x, y) defined by (1). Moreover, (15) L = 3a, + 2g,)(1 + ¢,) where tg yf “ a) 8 The recurrence relations (14) are those of Borchardt’s algorithm [4]. We omit the proof of the error bounds (16) by expansion in power series, because a method of further speeding the convergence will be discussed elsewhere [5]. An algorithm with slower convergence is given in [4, Eq. (2.4)]. Note added in proof. Corollary 1 provides a solution of the second of two problems proposed by D.S. Mitrinovié, Problem 5626, this MonTHLY, 75 (1968) 911-912. See also [2, pp. 383-384]. This work was performed in the Ames Laboratory of the U. S, Atomic Energy Commission, References 1, B. Ostle and H. L. Terwilliger, A comparison of two means, Proc, Montana Acad. Sci., 17 (1957) 69-70. 2. D.S. Mitrinovié, Analytic Inequalities, Springer-Verlag, Berlin, 1970, p. 273. 3. B. C. Carlson, Some inequalities for hypergeometric functions, Proc. Amer. Math, Soc., 17 (1966) 32-39. 4, ——,, Algorithms involving arithmetic and geometric means, this MONTHLY, 78 (1971) 496-505. 5. ——,, An algorithm for computing logarithms and arctangents, Math. Comp., April 1972, ON THE CONVERGENCE OF THE L? NORM TO THE L® NORM R. A. HANDELSMAN, University of Illinois and J. S. Lew, IBM Thomas J. Watson Research Center If f(x) is a real- or complex-valued Lebesgue-measurable function on the finite or infinite interval [a,b], then its L? norm is defined by ish = (fecal) ”1972] MATHEMATICAL NOTES 619 for 1 S$ p<, and its L® norm is defined by | flo =ess. sup. {[f(%)|: a Sx Sb} with respect to Lebesgue measure. We admit co as a possible value for these norms, so that || f||, is well-defined for 1 < p $ co, and we recall the following well-known result of integration theory. Turorem 1. If f(x) is Lebesgue-measurable on [a,b] and ||f\|, is finite for some q < 00, then @ lim, ol Sl = [le Proof: [4, p. 39]. In this note we shall amplify the result just stated by using a basic tool of asymptotic analysis, a standard Abelian theorem for Laplace transforms which is usually called Watson’s Lemma. To do this we must impose stronger assumptions on F(x) (which are justified in many practical situations), but in return we can exhibit not merely an alternative proof of (1) but additionally an estimate for the resulting order of convergence. Moreover, we obtain as a by-product the Stirling approximation for p! We recall first the statement of this lemma. Watson’s Lemma. Let g(t) be a locally integrable function on [0,00) which is O[exp(rt)] as t + 00 for some finite constant r, and which has a finite asymptotic expansion Q g(t) = z apt * + 0(t"*) ast+0+ with 0 < Re(co) < +» a that f(x) is continuous and positive on [a,r] while f’(x) is defined, continuous, and negative on (a,r); and that (5) f(a) > M = ess, sup {f(x): r Sx 0 that © S(@) — f(x) = ka — a)" + of(x — @)""] as xa + and that (6) may be differentiated to yield @ FQ) = — ke Me — ay tM + of(x — ay "J asxsat. Then as p-> 0, ®) |f |» =S@[1 — ep-* log p + p~*logC + o(p~)], where C is given by 9) C= MOL @)/k]*. Proof. To estimate || f||, we write , : (ln? = 10) + 120) = f'soxrax + [sere and consider I,(p) as poo. If we introduce a new variable ¢ through f(x) = f(a) exp(— 1) and a new function g(t) through w= (i on [0,log(f(a) /f0)], (0 elsewhere on [0, c0), then g(t) is well-defined and locally integrable on [0, oo), so that I,(p) takes the form 1) =S0y [° exp — ph ata Moreover, by computation with (6) and (7) we find Gt) = [f(a [kf + oft") ast 0+ so that, by Watson’s Lemma and definition (9), we obtain (10) 1,(p) =f(@)[Cp~* + o(P-)]_ as p>. However, in [r,b] clearly f(x)/M S 1 by assumption (5), so that for all p> q ’ > (11) 1,(p) = M? f L/@) /M]’dx so that I,(p) and I,(p) coincide to lowest order. If we substitute (14) and (16) into (13), and take the p’-th root of (13) as before, then finally we obtain622 K.D. WALLACE (June-July |f|»=fOL —4p~" log p + 4 p~"log| af(e)/f"(o)| + o(-)] as p> co. In this not unrepresentative case, the convergence to |||. is obviously slow, so that |/f||.. is often a poor estimate for | f|,. As a further specialization let f(x) = xexp(— x) on [0, + 00), so that f(x) has a unique maximum at x = 1 and a convergent expansion SQ) =e" 4G 1)? +4 PF + Then (|f|,)” can be calculated exactly in terms of the gamma function, and approximately by means of our results; indeed by (13)-(16) T+ p)=(|J||pPp'*? ~ 2Cp~4f(1)?p +” = (Axp)tpre?. This is the Stirling approximation for p!, which may, of course, be obtained directly through Watson’s lemma. The initial work leading to this paper took place while both authors were staff members of the Division of Applied Mathematics at Brown University, at which time this work was supported by the Advanced Research Projects Agency, Department of Defense, under ARPA Contract SD-86 with Brown University. References 1. E. T, Copson, Asymptotic Expansions, Cambridge Univ, Press, Cambridge, 1965. 2. A. Erdelyi, Asymptotic Expansions, Dover, New York, 1956, 3. R. A. Handelsman and J. S. Lew, Asymptotic expansion of a class of integral transforms via Mellin transforms, Arch, Rational Mech. Anal, 35 (1969) 382-396, 4, L.H. Loomis, An Introduction to Abstract Harmonic Analysis, Van Nostrand, Princeton, J, 1951. EXTENSION OF MAPPINGS IN FINITE ABELIAN GROUPS Kyte D, Watace, Western Kentucky University The purpose of this note is to provide a simple example to resolve a question that may arise in an introductory course in abstract algebra, and further to provide a means for injecting more elementary structure theory in such a course. Considera- tion shall thus be restricted to finite abelian groups (written additively). If A is a subgroup of the abelian group G and ¢ is an automorphism of G, then @ induces isomorphisms which yield A © (A) and G/A © G/(A). In fact, the diagram A>>G>> GIA | } } | | $A) >> G >> GIG(A) is commutative, where the vertical maps are induced by ¢ and the horizontal maps are canonical. The problem to be considered deals with the converse of this result,1972] MATHEMATICAL NOTES 623 Question. Suppose A and B are subgroups of the finite abelian group G such that A = B and G/A = G/B, Does there exist an automorphism « of G such that (A) = B? We note that it suffices to consider only finite abelian p-groups for p a prime. Before giving an example to illustrate that such an automorphism need not exist, we shall require the concept of the height of an clement in a p-group and a few elementary properties. Let G be an abelian p-group. We define inductively: p°G=G, pG={px|xeG}, + p""'G = p(p"G). For xeG, the height of x in G, hg(x), is the non-negative integer n if xe p"G but x¢ p"*'G (that is, if x is divisible by p" but not by p"*?). We say that x has infinite height in G if x € p"G for each positive integer n. Note that for a finite p-group G, 0 is the only element of infinite height. We write hg(0) = 00 and co > n for each positive integer n. If x lies in a subgroup A of G, we may define two heights for x; namely h,(x) and hg(x) the height of x in A and G, respectively. The following properties are easily established. PL: (a) If he(x) # hg(y) then ho(x + y) = min {hg(x), he(y)}- (b) If ho(x) = hey) then hig(x + y) = ho(x), and may be strictly larger. P2: IfG is a direct sum of subgroups, then height is computed componentwise. (That is, if G = @rer Ais 8 = Lier a then ho(g) = min{h,(a)}). P3: (a) If fis a homomorphism of G into G then he(x) S hg( f(x) for xe G. (b) An automorphism preserves height. Example. Let G = be a direct sum of cyclic groups with {a> = Z,, ® B is an exact sequence then it is split exact. References 1. L. Fuchs, A. Kertesz, and T. Szele, Abelian groups in which every serving subgroup is a direct summand, Publications Math. Debrecen, 3 (1953) 95-105, 2. P. Griffith, Infinite Abelian Group Theory, The University of Chicago Press, Chicago, 1970. 3. P. Hill, On the classification of abelian groups, (Preprint). 4. I. Kaplansky, Infinite Abelian Groups, University of Michigan Press, Ann Arbor, 1954,1972] MATHEMATICAL NOTES 625 A PROOF OF GANDHI’S FORMULA FOR THE nth PRIME CHARLES VANDEN EyNpen,, Illinois State University Let Q denote the product of the primes less than the odd prime p, and let 1: be the Mébius function. The inequalities 1 o) 1<2 (545 <2, eeorat! which were announced by J.M. Gandhi at theAugust 1966, International Mathematics Conference in Moscow, allow p to be calculated from the primes preceding it, since for any real number « the inequalities 1 < 2'« <2 hold for at most one integer k. Gandhi's proof involved generating functions; in this note I present a more elementary argument. Denote the summation by ¢. Then 2e-1 (Q22-1)¢ = Lud) = Lp(a +24 +274 + 420-4), ala ~ alo Since for 0 x€0,> S(x,V)cU for some Ue®, If St(x,W) co X—B, then Vy ¢X—B or b’eX—B, a contradiction. If Si(x,¥) ¢ U, for y#a, then V, < U, for some y x a. This contradicts the choice of U, since we chose U, such that a ¢ U,. Hence P, and P, are disjoint open sets containing a and B respectively. Reference 1, J. L. Kelley, General Topology, Van Nostrand, Princeton, N, J., 1961. A SIMPLE EXAMPLE ON SOME PROPERTIES OF NORMAL RANDOM VARIABLES Javad BewBoop1aN, University of North Carolina, Chapel Hill, and Pahlavi University, Shiraz, Iran In recent issues of this MONTHLY, several authors presented different examples on non-normality of a linear combination of normally distributed random variables [1,2,3]. The purpose of this note is to give a simple example of a non-normal multivariate density for demonstrating some of the properties of normal random variables. Consider 2 S15 X29 +05 %q) = 2 Prfi(X1>¥ 25°" Xn)s where 01. Thus the given series converges if and only if ¢ > 1. Example 2: Xf. 1/n(logn)*. The condensed series is m2 2 1 5 1 * log)" (jlog2)*— log2)" zy, —— 2(log(2))* which converges if and only if « > 1 by Example 1. Example 3. © 2-21/nlogn(log(logn))*. The condensed series is 2 1 1 zt ———_—___ 2!log(2/) log(log 2/))* J(log 2) (log(j log 2))* 1 1 ~ jog2 ~ j(logj+log 2)" which converges if and only if «>1 by comparison with Example 2. MATHEMATICAL EDUCATION Eprrep py J.G, HARVEY AND M. W. POWNALL. Material for this Department should be sent to either of the editors: J. G. Harvey, Department of Mathematics, University of Wisconsin, Madison, W153706; M. W. Pownall, Department of Mathematics, Colgate University, Hamilton, NY 13346, MATHEMATICS COURSES IN 1984 J. Barkxey Rosser, University of Wisconsin To answer the question, “What Undergraduate Courses Will be Taught in 19842” one proceeds through four assessments: (a) what will be taught if past trends simply continue? (b) what ought to be taught? (c) what steps can be taken to convert some of (a) to (b)? (d) how many of the steps proposed under (c) will be undertaken, and what success will they have?662 ELEMENTARY PROBLEMS AND SOLUTIONS. (June-July 2. If for “omitting proofs” we read “omitting demonstrations” then this method of teaching is common in all the sciences, Most students in physics and chemistry, for example, seldom perform the basic experiments which support modern theory. The analogy to mathematics is not precise, since to the practising mathematician the method of proof is frequently as valuable as the final result. But the non-mathematical scientist generally must know only when to apply certain mathematical techniques. Here an intuitive understanding of the rationale without precise demonstrations will go a long way towards satisfying his needs. 3. There are several good examples of such “boot strap” operations of which the University of Pennsylvania is one. Starting with one experimental computer-assisted section serving approximately 10% of the freshmen, it has become possible to offer integrated caleulus-with-computer courses to all who desire it, and these are in the majority. The main difficulty is not in professors learning to program — enough will volunteer — but in getting enough teaching assistants to correct the students’ mistakes, 4. Some view the computer as the greatest single emerging threat to civil liberties. There is little possibility of turning computers off. Intelligent regulation of their use will require a body of informed izens who understand what they can do. Some instruction in the use of computers is probably therefore in order even for pre-law students. For a well-documented account of present dangers see the work’of Professor Arthur Miller of the University of Michigan Law School, “The Assault on Privacy — Computers, Data Banks, and Dossiers”, University of Michigan Press, Ann Arbor, 1971. PROBLEMS AND SOLUTIONS, Eprrep By Emory P, STARKE Associate Eprrors: JosuuA BARLAZ, ERIC S. LANGFORD. COLLABORATING Eprrors: LEONARD Caruitz, GULBANK D. CHAKERIAN, HASKELL COHEN, S. AsHBy Fore, Israet N, HERSTEIN, Murray S. KLAMKIN, DaNteL J. KLermman, Rocer C. LYNDON, MARVIN Marcus, CHRISTOPH ‘NEUGEBAUER, ALBERT WILANSKY, and UNIVERSITY oF MAINE PROBLEMS GrouP: GrorcE S. ConNINGHAM, CLayTon W. Dopce, Howarp W. Eves, WILLIAM R. GriGER, GARY Hac- GARD, Prtuir M. Locke, Joun C. Marnuuser, Curtis $. Morse, Epwarp S. NorTHAM, and Witutam L, Soute, Jr. All problems (both elementary and advanced) proposed for inclusion in this Department should be sent to E. P. Starke, 1000 Kensington Ave., Plainfield, NJ 07060. Proposers of problems are urged to enclose any solutions or information that will assist the editors. Ordinarily, problems in well-known textbooks and results in generally accessible sources are not appropriate for this Department. No solutions (except those accompanying proposals) should be sent to Professor Starke, ELEMENTARY PROBLEMS ‘Solutions of Elementary Problems should be sent to Problems Group, Mathematics Department, University of Maine, Orono, ME 04473. To facilitate their consideration, solutions of Elementary Problems in this issue should be typed (with double spacing) and should be mailed before September 30, 1972. Contributors (in the United States) who desire acknowledgment of receipt of their solutions are asked to enclose self-addressed stamped postcards. An asterisk (*) means neither the proposer nor the editors supplied a solution. E 2361. Proposed by Richard Johnsonbaugh, Morehouse College Prove that the following series converge conditionally:1972] ELEMENTARY PROBLEMS AND SOLUTIONS 663 L(y" 1and © (-1)"[e- (1 + 1/n)"). nat nat E 2362.* Proposed by C. H. Kimberling, University of Evansville Suppose that in some probability space, E,, Ez,--- are events with common probability p. Let m 22 be a fixed integer. Prove or disprove that p™ S sup {P(E, VE, 0 OE, )} where the supremum is taken over all m-tuples (i,,i2-yi,) of distinct natural numbers. E 2363. Proposed by Hiiseyin Demir, Middle East Technical University, Ankara, Turkey Characterize pairs of spherical triangles ABC and A’BC’ for which A ,Ci=c,A=a', B=b',C=c. E 2364. Proposed by G. J. Michaelides, University of South Florida. Suppose that r is a positive integer and that (i,,i2,°--si,) is a partition of r into nonnegative integers. Show that if p is a prime factor of n which is relatively prime to r, then the number of (distinct) permutations of (i,,i2,---,4,) is divisible by p. E 2365. Proposed (part I) by Erwin Just and Kenneth Fogarty, Bronx Com- munity College, and (part II) by J. B. Wilker, University of Toronto 1. Let S be a finite set of points in the plane in which no three points are collinear and not all points are concyclic, Define a common point of S to be a point which lies on some circle which passes through precisely two other points of S. Must each point of S be a common point? Il. Let S be a set of four or more points lying on a sphere but not on a circle, Prove that each point of S is on some circle containing precisely two of the other points of S. E 2366.* Proposed by B. P. Gill, Demarest, N.J. Let V be the set of vertices of a regular 2n-gon and let A* and B* be convex n-gons whose vertices are subsets A and B of V. If the set of lengths of all chords with both ends in A (with each chord length being counted according to its multi- plicity) is identical to the like set for B, then is B* necessarily congruent to either A* or (V\A)*?664 ELEMENTARY PROBLEMS AND SOLUTIONS (June-July SOLUTIONS OF ELEMENTARY PROBLEMS Wilsonian Products in a Group E 2303 [1971, 674]. Proposed by Charles Lindner, Auburn University Let G be a finite group of odd order. Then the set of products of all elements of G, taken in any order, is in the commutator subgroup. [As a corollary we have the well-known result that G abelian implies L,eg x = identity. Query: Does the set of all such products exhaust the commutator group?—Ed.] I. Solution by G. A. Heuer, Concordia College. Let G = {g1,82,--,8,}, let N be the commutator subgroup of G, and suppose that the order of N is m. Then (g1g2-- 8) = (g:N)(g2N)--(g,N) = X™, where X is the product of all elements of G/N; since G/N is an abelian group of odd order, X = N, the identity of G/N. Thus g,g2°"g,€N. IL. Comment by Solomon Golomb, University of Southern California, While the problem posed is indeed elementary, the “‘Editor’s Query” is anything but elementary. I have been interested in this question (as applied to all finite groups) since it first occurred to me in 1951; finally around 1967, having despaired of solving it myself, { submitted it as a Research Problem to the Bulletin of the AMS, where it languished for several years, finally appearing there in vol. 76, no. 5, September 1970, as problem 8 on pp. 973-974, entitled ‘““Wilsonian products in groups.” I have received no solutions to date. Also solved (first part only) by Ram Avtar (India), Anders Bager (Denmark), Michael Barr (Denmark), S. Baskaran (India), California Polytechnic Solutions Group, Fred Clare, John Coolidge, Harold Donnelly, S. F. Ebey, Daniel Farkas, Bruce Ferrero, Zbigniew Fiedorowicz, 8. W. Golomb, M. G. Greening (Australia), J. W. Grossman, Elgin Johnston, Geoffrey Kandall, David Kelly, ‘Yuriko Kojima, Harry Lass, C. B. A. Peck, Ernest Propes, Simeon Reich (Israel), Azriel Rosenfeld, Daniel Shapiro, Stephen Spindler, Glenn Stevens, John Stout, D. P. Sumner, E.T. Wong, and the Proposer, Editor's Comment. It is widely known that if G is abelian, then the product of the elements of G is the identity, except in the case that there is a unique element x € G of order two; in this case the product is x. Applying this to the multiplicative group of numbers mod p, where p is a prime, we have that (p — 1)! = —1 (mod p), which is Wilson's Theorem. This is the origin of the term “Wil- sonian product”, 8. Baskaran and C. B. A. Peck refer to A. R. Rhemtulla, On a problem of L. Fuchs, Studia Scientarum Mathematicarum Hungarica, 4 (1969), 195-200. If G has order n and if S is the set of all products of n distinct elements (i. e., the set of all Wilsonian products), then it is not hard to show that S is contained in a single coset of the commutator subgroup. Rhemtulla shows that Sis equal to that coset in the case that G is solvable. In this paper, mention is made of work by J. Dénes which shows that the same result holds when every member of the commutator subgroup is actually a commutator and when in addition, G has at most 4 elements of order two.1972] ELEMENTARY PROBLEMS AND SOLUTIONS 665 Cardinality of Intersecting Cosets E2304 [1971, 674]. Proposed by J. C. Owings, Jr., University of Maryland Let G be a finite group, H a subgroup of G. Show that, given any left coset L of H, there exists an integer k such that, for any right coset R of H, LA R is either empty or has cardinality k. Solution by the Bennett College Team. We prove the more general assertion: Let G be a group (finite or not) and let H be a subgroup of G. If L is any left coset of H, then there exists a cardinal number k such that if R is any right coset of H, then LOR is either empty or has cardinality k. To prove this, suppose that L = aH. The right cosets that meet L are precisely those of the form Hah where h € H; one of them is Ha. The proposition will follow if we can show for each heH there exists a bijection from aH Ha to aH M Hah. Given such an h, let F(x) = xh forall x €aH 0 Ha. Clearly Fis an injection from aH Ha to aH Hah. We now show that F is a surjection: If yeaH q Hah, then y = ah, = hzah for some h,, h,¢H. Let x = h,a; then xe¢Ha and x = haa = ahyh-*eaH. But F(x) = xh = h,ah = y, and the assertion is proved. Also solved by twenty-four other readers and the proposer. Several other solvers also dispensed with the finiteness condition. Stern Rediscovered E2305 [1971, 674]. Proposed by M. D. Hendy, University of New England, Australia In the system of reduced residues modulo p, where p is a prime, for each e| p—1, there are (e) elements of order e. Prove that the sum of these $(e) elements = p(e) (mod p), where ¢(e) is the Euler totient function and j(e) is the Mébius function. I. Solution by Leonard Carlitz, Duke University. If e divides p—1, let S(e) denote the sum of the $(¢) numbers belonging to the exponent ¢ (mod p) and let T(e) denote the sum of the e numbers satisfying x*= 1 (all congruences being mod p). Obviously T(e) = © q.S(d) so that (as in the Mdbius inversion formula), SQ) = DaeT@ule/d). But T(e) is the sum of all solutions of P(x) = x*—1 = 0 in the integers mod p; this is the negative of the coefficient of x‘! in P(x) so that T(e) = 0, if e>1, and T(1) = 1, Substitution shows that S(¢) = p(e). If we let S,(¢), denote the sum of the kth powers of the $(¢) numbers belonging to the exponent e(mod p) and define 7,(e) analogously, then we can show by a similar argument that S,(¢) = Ldy(e/d), the sum being taken over all divisors d of (¢,k). (Cf. E. Landau, Vorlesungen iiber Zahlentheorie I, Satz 220, p. 188.) Il. Solution by Solomon Golomb, University of Southern California. Let ©,(x) denote the cyclotomic polynomial of order n. (That is, ®,(x) =MI(x — 0,666 ELEMENTARY PROBLEMS AND SOLUTIONS [June-July where { runs through the primitive nth roots of unity.) It is known that ®,(x) has integer coefficients and if s = (n), then ®,(x) = x* — y(n)x*"' + +++; that is, the degree of ,(x) is @(n), and w(n) = ZL, where { runs through the primitive nth roots of unity. (This is the “‘algebraist’s definition” of the Mébius function.) For any prime p, the natural modulo p mapping from Z[x] to Z,[x] is a ring homomorphism, and ©,(x) with coefficients reduced mod p is still the polynomial for the primitive mth roots of unity over Z,, provided (n, p) = 1. (We remark that it is possible to have coefficients other than 0, 1, or —1 in the cyclotomic polynomial, despite occasional printed statements to the contrary.) In particular, if n divides p—1, then surely (n, p) = 1; moreover, the (n) primitive nth roots of unity are elements of Z, itself, and their sum must equal the negative of the second coefficient of ©,(x) (mod p), which is p(n). Also solved by twenty-seven other readers and the proposer. Editor's comment. According to L. E. Dickson, History of the Theory of Numbers, Vol. I, Chelsea, New York, 1952 (p. 184), this result was known to M. A. Stern (Jour. fiir Math. 6 (1830), p. 258). Ney Borba comments that the case e = p — 1 was proved by Gauss (Disg. Arith. I, Article 81). A Singular Problem E2306 [1971, 674]. Proposed by Anon, Erehwon-upon-Wabash Let A be an n x n matrix, wa 1 x n row vector, » an n x 1 column vector, and A —Av a [a uA el (a) Prove that 0 is a characteristic root of B. (b) Suppose detA = 0. Show that ¢? divides the characteristic polynomial det(#I — B) of B. () Discuss the converse of (b). Solution by W. G. Leavitt, University of Nebraska. To show (a), simply note that (u|1)B = 0. To show (b), suppose that det A = 0 so that xA = 0 for some row vector x # 0. Then (x|0)B = 0; but obviously (x|0) and (u| 1) are inde- pendent, so that 1? divides the characteristic polynomial P(!) = det(tl — B) of B. For (c), define P as follows: TO eels Then B has the same characteristic polynomial as PBP-'; it is easily seen that (“" +o) —Av | 0 ; te PBP' = 01972] ADVANCED PROBLEMS AND SOLUTIONS 667 so that P(t) = det(t! — B) = tdet(#! — A(I + vu). Now # divides P(®) if and only if t divides det(t! — A(I + vu), which means that cither A or (I + vu) is singular. Since det(I + ou) = 1+ up, it follows that 1 divides P(#) if and only if det A = 0 or uw =-1. Also solved by thirty-three other readers and the proposer. ADVANCED PROBLEMS All solutions of Advanced Problems should be sent to J. Barlaz, Rutgers — The State University, New Brunswick, N.J. 08903. Solutions of Advanced Problems in this issue should be typed (with double spacing) on separate, signed sheets and should be mailed before September 30, 1972. Con- tributors (in the United States) who desire acknowledgment of receipt of their solutions are asked to enclose self-addressed, stamped postcards. An asterisk (*) means neither the proposer nor the editors supplied a solution. Problem 5837 [1972, 94] is withdrawn. Inexplicably Problem 5797 reappears with a new number. 5860*. Proposed by L.-S. Hahn, University of New Mexico Let f(z) be a measurable function in the plane, assumed integrable on all circles with radius 1. Suppose f(z) has the property that its value at an arbitrary point in the plane is the average of its values on the circle of radius 1 centered at that point, viz. Lope, fO=% f f(z +e) dO for allzeC. Is the function f(z) necessarily continuous? 5861". Proposed by Michael Slater, University of Bristol, England Let F be an ordered field. (a) If pe F[x];a,beF, a 0 for aS x 0, such that ( X ¥) (R\E) > Ofor every rectangle R of positive (u x ¥)-measure. (2) Schwartz and Oxtoby refer us for a solution to the note of Darst and Goffman, A Borel set which contains no rectangles, this MONTHLY, 77 (1970) 728. @) In addition, Oxtoby notes that the solution may also be found in a paper by himself and P. Erdés, Trans. A. M. S., 79 (1955) 91-102, Theorem 1, In this paper it is also proved that subsets of the square which are equivalent modulo nullsets to some countable union of measurable rectangles constitute only a set of first category in the space of measurable subsets of the square. ‘Skew-Symmetric Second Order Directional Differential 5791[1971, 410]. Proposed by M. Z. Nashed, Georgia Institute of Technology For f: R? + R, xo€R°, and nonzero hy, h, € R°, the first and second directional derivatives are defined by flo; hy) = tim 1 {VC + thy) — 0}, 0670 ADVANCED PROBLEMS AND SOLUTIONS (June-July and 89 (%05 ya) = lim 4 (8f C0 + thas hy) ~ 8fC%05 hd} whenever these limits exist. Construct a function f: R? +R for which 52f(xo; ht,,h2) is a skew-symmetric nonzero bilinear form at some x9 €R? (i.e., 5°f(x03 hy, hia) = — 5%f(%o3 hia, hy) for all hy,hy €R°, and 62f(x9; hy, hz) is linear in hy and h, separately), or show that such a function does not exist. Solution by R. L. Van de Wetering, San Diego State College. Let f: R° + R be given by yx —2 Sf, waaxy Te +e Serie ta? or £(0,0,z) =f(0, y,0) = f(x,0, 0) = 0. Now m 12492) ~$0,y,2) x =F(0,y,2) = —y¥ —z. x0 Similarly we get f,(x,0,z) =x —z and f(x, y,0) =x + y. From this it follows that Sx(0,0,0) = —1, fyx(0,0,0) = 1, fr2(0,0,0) = —1, fex(0,0,0)=1, f2,(0,0,0) = 1, Jfy:(0,0,0) = — 1. We also have, after calculating f,(x, y,z), that F(%,0, 0) — F.(0,0, 0) Sxs(0,0,0) = lim x0 Similarly, f,y(0,0,0) = f,,(0,0,0) = 0, Finally —62f(0; h,k) = hgky + hgky — hyky + hykz — bik —htgky which is a skew-symmetric bilinear form given by the matrix 0 1 1 -1 0 1 -1 -1 0 ‘Fundamental Group of Non-orientable Manifolds 5792 [1971, 411]. Proposed by W. S. Massey, Yale University It is well known that given any finitely presented group G and any integer n > 4, there exists a compact, orientable n-manifold M" such that its fundamental group, ™(M"), is isomorphic to G, Is an analogous theorem true for non-orientable1972] ADVANCED PROBLEMS AND SOLUTIONS 671 manifolds? An obvious necessary condition is that G have a subgroup of index 2, since the set of orientation preserving path classes in a non-orientable manifold constitutes a subgroup of its fundamental group which is of index 2, Solution by the proposer. The following theorem is the desired analogue: TuroreM. Let G be a finitely presented group, H a subgroup of G of index 2, and n an integer = 4, Then there exists a compact non-orientable n-manifold M and an isomorphism @ of 2,(M) onto G such that @ maps the subgroup of orientation preserving paths onto H. Sketch of Proof: The proof is somewhat similar to the proof of the analogous result for orientable manifolds (cf. Algebraic Topology: An Introduction, by W. S. Massey, pp. 143-144). Let Z, denote a cyclic group of order 2 and w,: G+ Z, the unique homomorphism that has kernel 1; we shall arrange the construction so that w, will be the first Stiefel-Whitney class of M. Let g1,--,g, be generators for G and r1,-+,?y felations for G; each r, is an element of the free group F generated by 81)"""s8q Let h: FG denote the natural homomorphism; the kernel of F is the normal subgroup generated by the relations r;. Corresponding to each generator g,, we shall choose a compact n-manifold M, as follows: if w,h(g;)=0, then M,=S* x S"-4, while if w,h(g) #0, M,=an n- dimensional Klein bottle (j.e., a non-orientable (n — 1)-sphere bundle over S*), In cither case, the fundamental group of M, is infinite cyclic. Let M’ denote the con- nected sum of the M,; then 7,(M’) =F, and the first Stiefel-Whitney class of M’ “‘realizes”” the homomorphism w,h: F + Zz. Corresponding to each relation r,¢F* choose a smooth imbedding f,: S'+ M’ which represents the corresponding element of 7(M’). Since h(r;) = 0, it follows that w,h(r,)=0; hence the closed path f, is orientation preserving and the normal bundle of the imbedding f, is trivial. Thus we can do surgery on each of the imbeddings f, (see e.g., C. T. C. Wall, Surgery on Compact Manifolds, New York, 1970) and obtain a manifold M with the desired fundamental group; the details are similar to the proof of the analogous theorem in the orientable case. We note a corollary of this result. Given any finitely presented group H and integer n > 4, there exists a compact, orientable n-manifold whose fundamental group is isomorphic to H, and which admits a fixed point free orientation reversing smooth involution.

American Mathematical Monthly - 1972-6

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

American Mathematical Monthly - 1972-6

Enviado por

Direitos autorais:

Formatos disponíveis

Você também pode gostar