Scribd Logo
Upload

Welcome to Scribd!

  • Waring's Problem. W. J. Ellison

    Uploaded by

    Pablo Menéndez Menéndez
    43 views28 pages
    d

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document
    d

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    43 views28 pages

    Waring's Problem. W. J. Ellison

    Uploaded by

    Pablo Menéndez Menéndez
    d

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Flag for inappropriate content
    of 28
    Waring’s Problem W. J. Ellison The American Mathematical Monthly, Vol. 78, No. 1 (Jan., 1971), 10-36. Stable URL: bttp//links jstor.org/sic¥?sici=0002-9890% 28197 101%2978%3A1%3C 10%3AWP%3E2.0,CO%3B2-Q The American Mathematical Monuhly is currently published by Mathematical Association of America, ‘Your use of the ISTOR archive indicates your acceptance of JSTOR’s Terms and Conditions of Use, available at hhup:/www.jstororg/about/terms.hml. JSTOR’s Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at hhup:/www jstor.org/journals/maa. hon, Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the sereen or printed page of such transmission, STOR is an independent not-for-profit organization dedicated to creating and preserving a digital archive of scholarly journals, For more information regarding JSTOR, please contact support @jstor.org. bupswww jstor.org/ Mon Feb 21 16:12:08 2005 10 W. J. ELLISON Wanuary Choose a title for your paper that is short (not an abstract of the paper) but informative. Remember that each page of your article has a running head, which, hopefully, is most of your title. pap: Remark on a Theorem of Hilbert BAD: A generalization of Hilbert's Norm Theorem to Abelian Extensions of Skew Fields of Characteristic Zero Goon: Hilbert’s Norm Theorem for Skew Fields. Avoid mathematical symbols in your title. 8. Referenct 4, A manual for authors of mathematical papers, revised edition, Amer. Math, Soc. 1966, 2. Abbreviations of the names of scientific periodicals, Math. Reviews (reprints of the 1965 edition are sold by the Amer. Math. Soe). ‘3. William Strunk, Jr. and E. B. White, The Elements of Style, Macmillan Paperbacks 107, 1962, $.95. ‘4. A Manual of Style, 12th edition, Univ. of Chicago Press, 198. 5. P.G. Persia, Writer's Guide and Index to English, 4th edition, Scott Foresman, 1965, 6. H, W. Fowler, Modern English Usage, 2nd edition corrected, Oxford Univ. Press, 1968. 17. The preparation and typing of mathematical manuscripts, Bell Telephone Labs, 1963, (Available from: Mathematical Manuscripts, Bell Telephone Labs, Inc. 463 West Street, New York, N. ¥. 10014.) ‘8 Mathematies in Type, William Byrd Press, 1954. 9. Arthur Phillips, Setting Mathematics: A’ Guide to Printers Interested in the Art, John ‘Wright and Sons, 1956 10. T. W. Chaundy, P. R. Barrett, and C. Batey, The Print Pres, 1954, 11, Rita de Clereg Zubli and Cynthia B. Wong, Guide to Technical Typing, D. H. Marks Publishing Company, Braintree, Mass, 1969, 7.95, 1g of Mathematics, Oxford Univ. WARING’S PROBLEM. W, J. ELLISON, University of Michigan is book Meditationes Algebraicae (1770 edition, pages 203-204) makes the following statement: “Omnis integer numerus vel est cubus; vel e duobus, tridus, 4, 5, 6, 7, 8, vel novem cubus compositus: est etiam quadratoquadratus; vel e duobus,tribus &c. usque ad novemdecim compositus &sic deinceps.” In the 1782 edition, page 349, he adds guardedly *. similia etiam affirmari possunt (exceptis excipiendis) de codem numero quantita tum earundem dimensionum.” It has become traditional to interpret these assertions as: “Can every posi- 1, Introduction. Edward Waring in W, J. Ellison took hie B.A. at Cambridge University and is continuing work there for @ Ph.D. under Prof. J. W. S. Cassels. He spent 1969-70 at the University of Michigan with D. J. Lewis, [Evidently he is a promising young number theorist. Editor. 1971] WARING’S PROBLEM rt tive integer be expressed as a sum of at most g(8) kth powers of positive inte- gers, where g(Z) depends only on &, not on the number being represented?” and to call the resulting problem “Waring’s problem.” There seems little doubt that Waring had only limited numerical evidence in favor of his assertions and no shadow of a proof. ‘The case £=2 had been stated by Fermat in 1640 and was attacked unsuc- cessfully by Euler for a very long time. It was finally proved by Lagrange in 1770, who showed that each positive integer could be expressed as a sum of at most four squares of positive integers. During the next 139 years, special cases of the problem were solved for k=3, 4, 5, 6, 7, 8, 10. It was in 1909 that Hilbert solved the problem in the affirmative for all &. His proof was extremely compli- cated in its detailed arguments. The key result was a proof of the following lemma. Lema 1.1. For each pair of positive integers k and n there are: an integer M=(2k+1) - ++ 2k-+n—1)/(n—1)4, positive rationals de, +++, day and inte- ers any © * > Qyar, ayy © * + nar SUCH that et)” Gitta = Dade + The reader may wish to test his ingenuity by proving the lemma in the following special cases: n= 2, k=2; and n=3, £=3. In section 5 of this survey we shall give a complete, short, elementary version of Hilbert's solution of Waring’s problem. ‘Once one knows that the answer to Waring’s problem is “yes!” it is natural to ask “How big is g(k)?” We can easily see that g(k) & [(3/2)*]+2*—2 because the integer m = 2*[(3/2)*]~1 is less than 3*, and its minimal expression as a sum of Ath powers isa sum of ([(3/2)*]—1) kth powers of Zand (2*—1) kth powers of 1. We shall see later that [(3/2)*]+2'—2 is probably the correct value of e(b) for all &. Obtaining an upper bound looks (and is) very much harder. Hilbert’s proof, as it stood in 1909, was not very amenable to giving an ex- plicit upper bound for g(&). Stridsberg [13] however, gave an explicit proof of Lemma 1.1 and the way was open for obtaining an explicit upper bound for (2). Strangely enough nobody took it until 1953 when Rieger [11] worked out the details and showed that (8) S (2k + 1)oorn In the intervening period Hardy and Littlewood published a series of papers during the 1920's in which they used a powerful new analytic technique to re- solve Waring’s problem and to show that g(b) =O(é2'+1). This upper bound, though large compared to the trivial lower bound for g(#), heralded the begin ning of an era in the theory of numbers. From the work of Hardy and Littlewood it became apparent that a more fundamental number than g(#) was G(E), which is defined to be the least posi- co) 12 WW. J. ELLISON January tive integer such that all sufficiently large integers can be represented as a sum of at most G(k) kth powers of positive integers. That is, there are infinitely ‘many integers which actually need G(8) lth powers. For example, each integer of the form 8n-++7 really does need 4 squares in its representation as a sum of squares. For the squares modulo 8 are congruent to 0, 1 or 4, s0 if n=7 (mod 8), then cannot be written as a sum of three squares; hence G(2)24. As G(2) S4(2)=4 we have 4=4(2)=G(4). As a further illustration, it was proved by Dickson [38] that ¢(3)=9, but 23 and 239 are the only integers which actually need 9 cubes; each integer greater than 239 can be written as.a sum of at most 8 cubes. But only a finite number of integers really need 8 cubes, from some point ‘on 7 cubes suflice. The tables seem to indicate that the point is 8042. The pre- cise value of G(3) is not known; the best result that I know is 43 G(3) $7. Hardy and Littlewood did much more than obtain an upper bound for G(&); they obtained an asymptotic formula for the number of integral solutions of the equation x}+ ++ + +x}=W,™120, - + + , #420. In section 3 I shall show how the following theorem is proved. ‘Tuwonem 1.2. If s%2+4, then the number ru.4(N) of solutions in integers of Ht + tae N m1 BO ++ +, ¥,B0, satisfies ra +(/a) TG/é) where ©(N) &c>0 isa certain arithmetical function of N. tual) NGW-G(W) + (WOM) og No, Vinogradov made great technical improvements to the Hardy-Littlewood method, and he was able to show that the conclusion of Theorem 1.2 holds if s>altlog k, where cis a positive real number. For large values of i this is a ‘much weaker condition than s22-+1. Vinogradov also obtained upper bounds for G(&). In [62] he proved G(2) 56% log k-+# log 216. No substantial improve- ments on this estimate have been made, though in recent years the numerical factors have been reduced slightly. Before going on to describe the Hardy-Littlewood-Vinogradov method in detail, I shall discuss Linnik’s solution [17] of Waring’s problem. Though his proof is strictly arithmetical, it was clearly very much influenced by the analyt- ic method. Linnik’s proof uses methods from the general theory of sequences. For a beautiful introduction to this fascinating branch of number theory one cannot do better than to read Halberstam and Roth's book Sequences. 2. Linnik's method. If & is a sequence of positive integers the Schnirelmann density o(Q) is defined by'o(Q) = greatest lower bound of A (z)/x for #21, where A(@) is the counting function of $f, that is A(z) is the number of elements of & less than or equal to x. Clearly 0So(Q) $1. If o(&)>0 then 16%. If of) =1 then {= 2, the set of positive integers. Let &{ and & be two sequences of positive integers, then %@ is to consist of all positive integers which are either a or 6 or a+b, where a€% and JEB, 1974] WARING’S PROBLEM 13 ‘each counted only once. For example, if {= {1, 2, 5,8} and B= [3, 7, 10, 11}, then 1OB = {1, 2, 3,4, 5, 7, 8,9, 10, 11, 12, 13, 14, 15, 16, 18, 19}. ‘Taronem 2.1. For any sequences & ond B, o(L © B) & o(M) + o(B) — o(Mo(B). Proof. Without any loss of generality, we can assume o(Mf)>0; thus 128%. Let 1,3, «+ + ax be the part of & up tox. Whenever ar-+1 1, then L@B consists of all the positive integers. Proof. Suppose n£%+, so in particular n€U. Consider the integers n—a, where a,E% and 1Sa;Sn—1. The number of such integers is A(n)Zan>n —fn2n—B(n). The integers n—a, lie between 1 and n—: ‘lusive. The num- ber of elements of 8 in this range is B(n—1). So the total number of integers of ‘the form —a; and the number of members of B in the range 1 to (n—1) is A(n) + B(n — 1) > mn — Bin) + Bn — 1) Bn 1. ‘Thus atleast one member ofthe act of integers (nae) must belong to 8, 80 we have 2=a;+bE% @Q, a contradiction. A sequence 2 is called a basis of order h if U@ - - - 2, taken i times, consists of all the positive integers. That is, each positive integer can be ex- pressed in the form a,+ +++ -Hay with £Sh, and the a's all belonging to 9. “4 W. J. ELLISON Uanuary ‘THwoREM 2.3. If o(8) >0, then U is a basis. Proof. Define = OY and %-41=Y-OM for 7&2. By Theorem 2.1, we have o() Z2a—a*=1—(1-a)%; by ‘induction we easily prove that o(@%) 21 =(1=ayr. We choose r so that (1—a)’ <4, hence o(@,)>4. Then by Theorem 2.2 we see Yar= ¥, ©4, consists of all the positive integers. We shall take as our sequence {= {nt:n=1, 2,3, ++ - }. Now o(& butif we ean show that for some integers the sequence 1" =%@ «= - OX, taken s times, has positive density, then by Theorem 2.3 we can find an integer r such that 2, is the sequence of positive integers. This will imply that g(&) S2rs. Let us now see how Linnik proved o(2{") >0 for some s. Denote by r(WV) ‘the number of integral solutions of the equation Bites bat aN withasz0,---, 220, and by R,(N) the number of integral solutions of the inequalities OSnt stats W withy2o,---,m20. Thus, Ri(V)=r(0)+ + + + -+rs(N) and Linnik’s fundamental resul lowing lemma. Lenata 2.4, There exists s=5(b) such that r4(n) ScN for 0SnSN, where €>0 depends only on b. Unfortunately, I do not know a proof of this lemma that is not long and com- plicated. However, one can prove quite easily that on average the value of r4(n) is O(n). The hard partis to prove that if is sufficiently large, then r,(n) does not differ from the average value. As an exercise show that }.¥_, r(n)=R,(N) i is the volume of the # dimensional solid defined by related to the number of integral solutions of the inequalities and to the sum XL rn?) It is now straightforward to show that ¢(") >0, where sis as in Lemma 2.4 By Theorem 2.3 this is sufficient to solve Waring's problem. We shall suppose o(%°) =0 and deduce a contradiction. This assumption implies that for any €>0 there are an infinity of V such that A%(N)0 and this solves Waring’s problem. Linnik did not give an explicit estimate for (8). This was done later by Rieger [20], who proved that g(8) 22" a+0! 1971] WARING'S PROBLEM 15 Schnirelmann took the sequence $= {P, primes}, which also has o(8) and proved the analogue of Lemma 2.4. From this he was able to deduce, as above, that each integer can be expressed as a sum of a bounded number of primes. (An estimate for the number of primes needed is $210") This was the first step towards a proof of Goldbach's conjecture. The last step has not been taken yet! Asa generalization of Waring's problem, Rieger [21] proved that if & is a monotone sequence of positive integers with o(%) >0, then the sequence $4 = {a', cE%}, EZ, forms a basis for Z+. The classical Waring problem is the special case = 2+, 3. The Hardy-Littlewood-Vinogradov method. It is not possible in a short space to give a detailed account, with full proofs, of the method. I propose there- fore to give the skeleton structure, indicating how all the bits fit together and refer the reader to Davenport [1] for the proofs. Many of the proofs are straight- forward manipulations and applications of standard analytic and number- theoretic arguments. Enthusiastic readers are urged to attempt proofs of all the lemmas for themselves, before looking at Davenport to see how it should be done For convenience we introduce the notation e(#) =e**. Consider the func- tion f(a) = D2, e(xa), where @ is any real number. Then. [s(a)}" = 2 Reeve), where Re(N) is the number of integral solutions of the equation xf+ +++ +2 =N, with 0Sx,=P and1sigs. Te P(N], then Re()~R(W), the number of integral solutions of the equation af-f «+ + tatmi¥ with 20,» +, 20. By elementary calculus we have: (@.) Re(N) = f {fle)}"e(—Na)de. The idea of the method is to show that if s, depending only on &, is chosen sufficiently large, then the integral is positive for all Nz No(B). This means that the equation x!-+ - + - +<4=W has at least one solution for all Vz No. Since all integers less than No can be written as a sum of at most No 1's we shall have solved Waring’s problem. To show that Rp(W) is positive if sis large enough, we shall obtain an asymptotic formula for Rp(N) of the shape Ta + c/a) TG/#) Thus, if we take P=[1V¥*], then Ro(W) =R(N); and if N is sufficiently large, then the first term is always larger than the error term, so R() >0. Re(W) PrHHS(N)+0o(Pr4) provided +B 2+ 1. 16 W. J. ELLISON Vanuary Unfortunately the integral (3.1) is extremely complicated, and we can do nothing with it as it stands. What we are going to do is to divide the range of integration into two disjoint sets IM and m, called the major and minor sets re- spectively. In the major set we shall make a series of approximations to f(a). It will turn out that the approximating functions are comparatively easy to handle, and we can explicitly evaluate the integrals which arise. On the minor set we shall show that the integral f Lol" -Nadda is comparatively small, and it can be absorbed into the error term of the asymp- totic formula. ‘The details are as follows. Consider the rational numbers a/g with (a, @)=1, 1Sa0, where ¢ depends only on &. This will ensure that the expression we have actually written down as a main term is the leading term in an asymptotic expansion for RV). ‘The estimation of the integral on the minor set is the hardest part of the whole argument. The key results are due to Weyl and Hua. They are our next two lemmas. Leama 3.4 (Weyl). For a&m, | f(a)| =O(P-arh+s), Lema 3.5 (Hua). For v2, f5|f(@)| "dee=O(P*—+ ‘These hold for any €>0, the implied ‘O" constants depend only on ¢, &, ». ‘The proofs are by induction on & and y respectively and they are by no means easy. Readers should try and prove the cases when k= 1 and »=1 respectively. We can now easily deduce our next result, Leama 3.6. If s22*+1, then fu {f(a)}*e(—Na)da=O(P-*) where 6>0. Proof. Jf tier} Neraa 5 f LHe as mas sto) fs) "ta = Ofpe-ontinpty = OPH, where 3=(s/24) ~e(s—2-+1) > 0, if eis sufficiently small, Now it only remains to show that S(WV) 2 c>0 for all N, We recall that 18 'W. J. ELLISON Uanuary 6m) =D @'Ss, ‘—Na/a), where S.,q= Di-1 e(ax*/q). We simpli defining y this rather complicated expression by A@ = and x0) =14 540), where p is a prime number. It is now easy to check that SW) = Tx), the product being over all the primes. One now proves the following lemma. Lewnta 3.7. (i) G(N) és an absolutely convergent series. (i) For any positive there isa prime number pa, depending only on, such that taa< [[x@) 0 for each prime number , then we can quickly show that (N)Zc> 0 forall W, since SW) = TI x) IIx) = —2) Tee) =e>0. we To show that x(¢)>c(P) for all prime numbers p, we define f(g) to be the num- ber of solutions of the congruence aj-+ - - » +x (mod g) and then note the following lemma. Lewma 3.8. 1971] WARING'S PROBLEM 19 and one then verifies that AP = EA x0) ws. po LEMMA 3.9. If for any prime p the congruence st-+ + + + -Fat =N(p") és soluble with not all the ‘0 (mod p), then x(p) 2 c(p)>0. Here y is defined as follows: if Pilb, then yar f p>2, y=742 if p= 2, From one solution of the congruence xj-+ » «+ -+x7=21V (mod 9”) itis possible to derive many solutions of the congruence xj+ + + - +x}=N(mod p’) for y>y. In particular one can show A(p”) 2 6°, so by Lemma 3.8 we have e(P) Sp >0, ‘The final link in the chain is the next lemma which tells us when the con- gruence af+ +++ -+af=N (mod #*) does have solutions with not all the x '=0 (mod 9). Luana 3.10. The congruence A+ +++ -tat=N (mod p') és soluble with xu +* + X4not all congruent to 0 (mod p) provided sz 2k if bis odd or s 4k if k is even, The proof of this lemma is just an exercise in the theory of congruences. ‘Thus, the final link in the chain of arguments leading to a proof of Theorem 1.2 is completed. The reader is once again urged to read Professor Davenport's book so that all the details of the argument can be filled in. ‘The result which we now go on to discuss is Vinogradov's estimate for G(R). The theorem we shall prove is the following. ‘TueoREM 3.11. G() SR(6 log k+4+log 216). In order to get this estimate Vinogradov solves the following, apparently harder, problem. Let r(N) denote the number of solutions of the equation Abe ta teban toy =, where 1SaSP, 0Su, mu S4P!, 0SySP¥%, 0S0S}PH, and wm, 0 can all be expressed’ as the sum of at most J th powers of positive integers. We shall choose / as a function of & later. If we can show that r(NV)>0 for all NV >No(e), then we shall have shown that G(#) $48+3I. Just as before we have an expression for r(N) as an integral: na = f'eo}*(ve}{ri@}«—wedde, where He) = Zea), Veo) = Lalas), and Ve) = Tee. 20 W. J, ELLISON Uanuary We divide the interval [0, 1] into the major and minor sets Mf and m and so obtain: 10) = ff (se0}*(V@PYCe(—wadda+ f {sod} {VCP VO)e—Nedda, = Tut Im Saye The basic idea is to show that if / is chosen appropriately, then r(N)>0 for all sufficiently large NW. We do this by showing that limy+» | In| /Zu=0, for this means that from a certain point on Iur> | In| and sor(N)=Iut+In> 0. Of course one cannot evaluate the integrals Zy and I~ directly. What we are going to do is obtain a lower bound By for Iv and an upper bound bn for | Im|, and then prove that ba/Bu—0 as N-« if J is chosen appropriately. This is sufficient because [nl . bn oslel sin. Siw © Bw Luna 3.12. Lx >aP*U(0)*V(0), where cis a positive real number depending only on b. Proof. By direct substitution we see that tum Ef ieia—weddo, Where Nv=—u——oy! and APES MS PS From Lemma 3.2 we know that f [1@) }#e(—Wia)da = (VS) + O(P*). Now yy a PUG) yan. FO) = ee, where ¢=«,(k)>0, and by Lemmas 3.7-3.10 €(Ni)2c>0, with @ depending only on k. Hence IZ eoP™ 3) = oP*U(0)V(0). We now investigate the integral over the minor set: [nl =| f irol*oe@rind—vewel < f [tol vol V0) ae Now |f(@)| =| X21 e(ax*)| SDF |elext)| =P, so we can give an upper bound for the last expressio 1971] WARING'S PROBLEM a J lael#l veo P| reo | des Pemex |v] f | 000 hee S P# max | V(a)| U(0)*. To investigate maxace | V(a)| we need the following two lemmas: Lenata 3.13. Let Dy(X) denote the number of distinct numbers up to X that can bbe represented as sums of 1 positive kth powers, Then Di(X)>aX"™, where 5 = (1=(1/2))!, «> 0 depends only on i. Proof. We use induction on 1, For 1=1, Di(x) = [x*!*]>ce"', Consider the numbers 2-8, where Ga) << Gx) and O< 5 < gets, and s is a sum of J~1 positive integral kth powers. The integers 2'++s are sums of J positive Ath powers, and they are all distinct. Hence Dit) = 2D abe) B eye, by the induction hypothesis on Dis (Jx!-), Lewnta 3.14 Suppose x runs through Xe distinct integer values in an interoat of length X, and Y runs through Yo distinct integer values in a set of length Y. Sup- ose further that = (¢/q)+O(1/4"), where (a, q) =1. Then |EXeomn[ s exw, Se (@+X)@+ ¥). ‘This lemma was proved by Vinogradov and applied by him in several im- portant investigations in number theory, such as his proof that each sufficiently large odd number can be expressed as a sum of at most three primes. However, the lemma does occur in the book Inequalities by Hardy, Littlewood and Polya, page 205. They seem to have ‘We now return to our investigation of maxace | V(a)|. Since «Gm we have Ja-(a/)| $1/(2kgP*) with P!*0, since log @0. Therefore 2 W. J. ELLISON Uanuary mmax | V@)| < ecxi*2™, | tal seat”. veo 00% Thus [iol ¢ eatvmetrnn Tae Pov Now V(0)2aXeP!* and U(0) =Dy(t- P+), so we have 63) Lal Penns Ta XP DGPS) Now Xo=DigP*¥) 2a(§P*)4 and D,QP*)2a(¢P*)'. On substituting these lower bounds into (3.3) we obtain Lal es pe Tar Sep, where A=k/2—2k—4-+e+h(1—(1/8))'—4—P(1— (1 (1/B)9. We now choose / so that A<0; this will ensure that | J_| /Zv—0. Simplify the expression for A; it becomes 1 = (1/8))((34/2) = 2) = Bebe Thus A will be negative if (1—(1/k))'3k/2<4k, ice., if (1 = (1/8)! < exp(—6 log 4"). Now (1—(1/))'=exp{! log(1—(1/k))} —log 62%, i.e., 1<2k log k+k log 6. In this case r(N)>0 for all sufficiently large values of N, and the number of kth powers required is 4k-+3/<4k+6k log k+k log 216. Therefore G(R) 5 6b log & + H(A + log 216). Since Vinogradov’s proof appeared, a great deal of work has been done in trying to improve the estimates for G(E). This work is exceedingly complicated involving intricate analysis and very delicate estimations of trigonometric sums. The best known results are the following: GQ)=4 6B)ET; GU) =16; GO) SB; GO) 5 36; GM) 52; G®) S73; G0) 599; GIAO) 5 122. Chen [52] showed that G(&) (3 log £+5.2) and Vinogradov [63] showed that G(R) 0, M>0 and positive rationals Duy ++ Man depending only om k, such that each integer N2A can be writien in the form N= Daron, where m2" for sis M. We now show that Theorems 5.1 and 5.2 are equivalent. Obviously if ‘Theorem 5.1 is true, it implies that Theorem 5.2 is true with A=1, M=e(k), and \= - + =Ay=1. Now suppose Theorem 5.2 is true. Let ¢ be the least common multiple of the denominators of the A. Then os=0,EZ". If X is a positive integer and XZeA, then X=No-+0, where 0S6, ctxry + scars such that Git al = Dean t+ ase”. Proof. To prove this lemma we use some elementary properties of convex bodies in R¥, The properties which we need are very intuitive, so we put their proofs (Propositions 5.4 and 5.5) into an appendix, so as not to interrupt the main argument. It is immediate that the set of homogeneous forms of degree 2k in 5 vari ables with real coefficients forms a vector space V of dimension N = (2k-+1) (2k+4)/24 over R. Here N is just the number of coefficients in the general form of degree 2k in x, +++, 1971] WARING'S PROBLEM 25 Consider now the set of vectors S of V given by the forms L=L(a) = (aim + + + +ee%)* for all sets of a, ++ +, eyEQ, the rational numbers. Let ‘4(S) be the convex hull of S (i., h(S) is the smallest convex set in V that con- tains S). We now need our first geometric result. Proposition 5.4, If S is any subset of a real vector space V, then any vector aSh(S) can be written in the form a= Pilg isi, where N=dim V, ES, MER, D=1, and M20 for i= 0,» ++, N. Moreover if a is a rational vector, i.e., all its coordinates are rational, and all the vectors in S are rational, then all the numbers do, - + - , dw can be chosen to be rationals ‘Thus, to prove the lemma it will suffice to show that a rational multiple of the vector (x?+ - + - +28)* is in h(S). For if it is, then Proposition 5.4 tells us that we can write (+--+ +2)' as Dio M(aart «++ Haws)" where Yar 2 AEG" and any += ans © - + ese. To show that («i++ - - - -+x3)*Gh(S) we need our second geometric result. Proposition 5.5. The centre of gravity of a continuous mass distribution in a bounded subset S of a real vector space V always lies in the interior of the convex: hull of the set. Let Tbe the set of vectors in V given by the forms L= (axi-+ + + + -bavs)* with ER for 1SiS5 and oft +++ -+a3S1. Then TCH(T)CH(S). So the centre of gravity of the mass distribution of unit density throughout 7 and zero elsewhere certainly lies in h(S). The centre of gravity of such a mass dis- tribution is given by 6 filo t + baer Mdoy +++ dan / fdas + das where R is the region of R® defined by al+ + + +aiS1. To evaluate the inte- gral we change the variables as follows. Put fy = Buoy + +++ Biss + Bases, where Bu=aiai +++ +98) The remaining entries in the (Bc) which defines the transformation are chosen in such a way that the ‘matrix (Bu) is orthogonal. The expression for g now becomes: mate ball fl Bate aly dt,> 0. Thus we have, aaeeite ta, 26 W. J. ELLISON anuary where cmofe dts + dt4>0. Since OEh(S) and h(S) is convex, therefore MfEH(S), where ) is any num- ber in the range [0, 1]. In particular we take N=r/c, where r is a positive ra- tional in the range 0 - , ay" Bi, + + +, By and positive rationals do, - - - , kw, with N and the ds depending only on, such that (x3-+y)*= Doh hile +B)". Proof. This follows immediately from the lemma and Lagrange's four squares theorem. Conor any 2. If Theorem 5.2 is true for k=m, then itis true for k= 2m. Proof. This followsimmediately from Corollary 1 on putting x1 =0. Fo integer P= A can be written as P= Da ut, where the wEQ* for 0515 depend only on & and the ys€Z* for 051M depend on P, then P can be written as P= Lm Dun = Dome where the 1 depend only on B. Before going on to Corollary 3 we shall introduce a convenient shorthand notation. If we have a sum of the form n= SoM, Nn, where A and the WEQt depend only on k not on n, then we denote it by n= 5 (k). With this notation ifa~ D (@ andb= ¥ (&) then a-+6= D (8) and ifa= Do (2k) thena= D> (2). If the reader prefers he can always write out the expressions in full. (Warning: Have a large sheet of paper ready.) Conottany 3. If r, m, x, TEZ* and rNy(b), then each integer in the range AT*SnA(T+1), for we then have the result by writing min the scale of T. The inequality is AP +E -T> ATH), and this is certainly satisfied if T is large enough, say T=. We now go on to show that if T is large enough, then each integer of the form AT*+hT™1+ «+ +d,47, where the by are integers in the range 0Sb, te ST~1, can be wr the form 3° (t). The proof is by induction on &. For =1 the result is trivial and for #=? the result is contained in Lagrange's four squares theorem. Lagrange’s theorem asserts that each positive integer can be expressed as a sum of at most four squares of integers. A proof can be found in almost any ‘elementary book on number theory. There is a very neat geometrical proof JW. S. Cassels's book Geomeiry of Numbers, page 99. ‘Suppose that Theorem 5.2 is true for all integers #:Sm—1. Then by Corollary 2 of Lemma 5.5, Theorem 5.2 is true for all even integers less than 2m. As ‘Theorem 5.2 is equivalent to Theorem 5.1, our hypothesis implies that Waring's theorem holds for all even integers less than 2m. Let T, Nar (P=1, 2, «+ +, m—1) be positive integers with mar ST, which will be chosen explicitly later. By our induction hypothesis there is an integer r, depending only on m, and integers 4420 such that Oy Dela Mey forj=t,2---5m We can take r=max(g(2k), 1SkSm—1). Substitute these values of ij for # in the identity of Corollary 3. Add the resulting equalities. We obtain S por ad @ Lait Lai + Vest = Tm), 28 W. J. ELLISON (January for j=1,-++, m—1, Let 6,5=B,; Dini Hy, and sum (2) over all j, to obtain ® = {Zero t weet = Lim). Write (3) as a polynomial in 7, say @ ag +++ tat = Dm), Net Dict mamas for #=2, +++, m—1 and aq=r 1)=4,—1 say. Note that A: depends only on m, since r depends only on m: As Lint af? ST we have xyis T!0™-® for all {and j. Hence we certainly have (in Dv ser fori SvSm—2 and 1Sj Brlm —1)T* OID, ie, T> [Br(m — 1}, This means that T>a,—N,20 for 1Si [Brm— 1}, The proof of Theorem 5.2 now follows easily. We replace T by (T'-+1) and have the particular result: © A(T +1)" +en(T + 1) =m), where ¢m is any integer in the range 0S¢qNe=max{ {Br(m—1)}", Mi(24i+1)}, we have shown that each integer n in the range AT*Sn A.NG lies in such an interval, by an appropri- ate choice of T we have proved Theorem 5.2 and hence Theorem 5.1. Appendix on Convexity: The best general reference is H. G. Eggleston, Con- sexity, Cambridge Tract number 47, We refer to this book for the proofs of the following plausible assertions. ‘When thinking of a convex body C in Rr it is usually convenient to consider it as a subset of the “smallest” linear variety containing it; for example we always think of a circle as lying in a plane rather than as a subset of R® or R®, etc, It is then a theorem that a convex body has an interior with respect to this space (see H. G. E. page 13). A support hyperplane to a convex body Cis a hyperplane which intersects the closure of C but does not intersect the interior of C. It is an easy lemma that through each point of the frontier of C there passes at least one support hyperplane (see H. G. E. page 20). If Sis any point set in RO, the convex hull, h(S), of Sis the intersection of all convex sets 5 whi contain S. Carathéodory’s theorem asserts that if yEM(S), then y can be ex- pressed in the form y= of. Miss, where 8,5 and 4ER* for OSiSn. In addi- tion Df-o = 1. (See H. G. E. page 35.) Proposition 5.4 is a trivial consequence of Carathéodory’s theorem, for if y and all the elements of S are rational vectors in R* then we have a set of linear ‘equations for Ao, + + -, Xe with rational coefficients, hence Noy - «be are posi- tive rationals, Our second geometrical result, Proposition §.5, is also easy to prove. For if the centre of gravity g of the mass distribution lies in the exterior ‘of A(S) or on the frontier of h(S), then in the frst case we can pass a hyperplane between g and i(S) and in the second case we can pass a support hyperplane through g. Taking moments about the hyperplane leads to a contradiction in either case, 6. Generalizations of Waring’s problem. The problems which one could say are generalizations of the original Waring problem are legion. There just is not enough space here to describe them. Consequently, I shall confine myself to just a few chosen more or less at random. ‘The first may be called the prime Waring problem; itis just like the classical problem only we restrict the integers i; in -¥., x4 to be prime numbers. The analytic method can be used to solve this problem and a good account of its felon, together with it many ramicaton ito be fund inthe book by lua ani]. A fairly natural generalization is to ask: “If f(x) is an integral valued poly- nomial which takes the value 1, then can each positive integer be expressed as a sum of a bounded number of values of f(s)?” The classical Waring’s problem corresponds to the case f(x) =x%, In a sense this problem goes back to Fermat, 30 W. J. ELLISON (January. who in 1640 characteristically asserted: “A positive integer is triangular or the sum of 2 or 3 triangular numbers; square or the sum of 2, 3, or 4squares; pentag- onal or the sum of 2, 3, 4, 5 pentagonal numbers; etc.” ‘The expression for the mth r-gonal number is Pls) = Hr — (nt — an) bm fore > 2. ‘This assertion also occurs in Waring's book, just preceding ‘Waring’s problem’! Fermat's problem was solved by Cauchy, the proof is quite elementary. Kamke [104] gave a solution of the more general problem, his argument being based on Hilbert's method. Later, the analytic machinery was brought to bear on the problem and a number analogous to G(k) appeared. Hua [97] gives upper bounds for this number, similar in character to Vinogradov's estimates for G(2). Perhaps the most natural generalization of Waring’s problem is to ask the question about algebraic number fields or even about arbitrary fields. Siegel [77] and [78] tackled the problem for number fields. He showed that if As is the set of algebraic integers of the number field K which can be written as a sum of Ath powers of algebraic integers, then there is a bound g(t, K) depending on and K such that each integer @EAs can be written in the form 0= DN. al, where NSe(b, K) and ay, ++, a4 are algebraic integers. In general Ax does not consist of all the algebraic integers of the field K. For example take K=@(12) and look at 4s. The integers of K are all of the form e+5,/2, a, EZ; the square of an integer is of the form a!-+25"+4-2ab-/2 so if @=u-+0-V2 is an integer and 9740 (mod 2) then a cannot be written as a sum of squares of integers in K. Siegel also obtained an asymptotic formula of the number of solutions of the equation @= 0%, af analogous to Theorem 1.2. He also conjectured that the bound g(, K) was independent of the number field K, a result subsequently proved by Birch, [69] and [81]. ‘Waring’s problem for general fields did not receive much treatment in the literature until recently when Ellison [86] showed by very elementary methods the following two theorems: ‘TueoreM 6.1. Let b be a fixed positive integer. Suppose that K is a real field with the following properties: (1) Eack totally positive element in K can be expressed as a sum of s squares in K, where s depends only on K. (2) For each totally positive element a in K, there exists a 8, depending on a such that sai (s-+ 28) ~<8'~

    You might also like