I ntn. J. Math. h. Sol. Vol. 6 No. 3 (Iq83) 409-418 409 DOT PRODUCT REARRANGEMENTS PAUL ERDOS and GARY WEISS Mathematics Institute, Budapest; University of Cincinnati (Received September 18, 1980) ABSTRACT. Let a (a), x n (x) n denote nonnegative sequences; x the rearranged sequence determined by the permutation Za x n n examine and S(a,x) denotes {a n, a" (x (n) denotes x denotes the dot product is a permuation of the positive integers}. We x S(a,x) as a subset of the nonnegative real line in certain special circum- stances. The main result is that if if and only if an+l/an KEY WORDS AND PHRASES. a n +, then S(a,x) x [a" x, ] for every , n 0 is uniformly bounded. Dot product, series rearrangements, conditional convergence. 1982 MATHEMATICS SUBJECT CLASSIFICATION CODE. 40A05. An elementary classical result of Riemann on infinite series states that a condi- tionally convergent series that is not absolutely convergent can be rearranged to sum any extended real number. A slightly similar group of questions arose in connec- tion with certain formulas in operator theory [i, p. 181]. Namely, if we le’t a M (x) denote any two non-negative sequences and n x denote the sequence (x (a), n (n) where n is any permutation of the positive integers, then what can be said about the Set of non-negative real numbers ntegers}. S(a,x) {a" x n is a permutation of the positive More specifically, which subsets of the non-negative real line can be realized as the form S(a,x) Various facts about for some such S (a, x) are obvious a and x? 410 P. ERDOS and G. WEISS (i) S(a,x) (2) If a 0 The values may be obtained. and are strictly positive sequences or are at most finitely zero, x and S(a,x) then (3) [0,]. (0,]. [0,] Not all subsets of a cardinality argument. If of the class of sequences S(a,x) (4) If either [0,] (5) An example: is x and (0,2,0,2 a if [0,], and c then the but the cardinality c c- c S(a,x) is finitely non-zero then x 2 is This follows by and thus the cardinality of the c is less than or equal to or a a set. denotes the cardinality of c cardinality of the class of subsets of subsets S(a,x) are realizable as an (3 x -n ), is countable. is precisely S(a,x) then the Cantor set except for those non-negative real numbers whose ternary expansion consists of a tail of O’s 2’s or a tail of (i.e., a subset of the rational numbers. For this It seems too ambitious to consider the general question at this time. reason we shall restrict our attention to the cases when sequence and If x and S(a,x) then x n + x or 0 the problem is trivial and 0, t x ]. n [0, M {0}. S(a,x) {} S(a,x) the problem is trivial and 0, n is a non-increasing sequence, x a is a non-decreasing a If a a If 0 I is bounded by n In any case, hereafter we shall assume a n " M, and unless otherwise specified. 0, The Lemma that follows is a well-known fact, but we give a proof for complete- ness and because the proof contains some of the ideas used in the main result. LEMMA. and if x n a 0, n If + a n and . Define 0 n . + S(a,x) [a. x,]. then for all n or if a n % and In addition, a > 0 n It suffices to show that for every permutation integers, we have every x n a. x e S(a,x), for some n e S(a,x). then PROOF. x and % a. x _< anXz(n) or, equivalently, The rest of the lemma is clear. nl in terms of n as follows. Set a- x n of the positive --< az(n)Xn for and DOT PRODUCT REARRANGEMENTS 1 zl(n) n= 1 (I) -i(i) n (n) It is straightforward to verify that otherwise z is also a permutation of the positive I integers (one-to-one and onto) which fixes To see this, note that X1 X -i > 0. z(1) > 1 411 and -i n We assert that 1 x < a a x 1 (i) Hence i. a a (i) 0 I and Therefore (i) [(a(n) al(n))Xn (a (a n(1) (i) -a I (i) )Xl+ (a ( -i (i)) a I) (xi x -i a H (n I -i )x (I)) z -i (I) (i) 0 Proceeding inductively, we obtain a sequence of permutations for which x < a a k Hence, for every x we obtain k n=l a Zk (n) x < a z n k, k x for which The main question of this paper is: S(a,x) with a,x The main result of this paper gives a partial answer. a acterize which x n +, n + have the property that it is quite rare. be [a- x,] a n 2 2 S(a,x) [a- x, ] for every S(a,x) [a" x, ] can never be for every x such that x n 0. That was first decided by an example due to Robert Young. n and n Namely, we can char- The first result in this direction was that if [a- x, ] S(a,x) then x x n 2 _2n+l Both results are unpublished. a n n or that for every S(a,x) may not Namely, let The succeeding results and techniques are due to the work of the authors in collaboration with Hugh Mont gomery. 0 x and n 0. On first sight, it might appear that n, a x,]? [a such that ,k k n n n=l is that fix 1,2 k k-i k Letting n 412 P. ERDOS and G. WEISS (The Main Theorem) THEOREM i. . a n Let (a) a n > 0 for every n and Consider the following conditions: is bounded. (i) a (2) For the non-negative sequence n+i/an and (x (a) a n [ (b) of x k as (x), there exist subsequences n x k and x and + n 0 (x) n x that tends to that for every positive integer and a ,x n n where for all 0 x= (x) hat tends to n 0 then (2) implies (3). n, To prove that (i) implies that (2) holds for every strictly positive PROOF. sequence k [a- x,] + a n n respectively such that Then (i) implies (2)for every strictly positive sequence Also if (a . x nk mk S(a,x) (3) x and a 0 m k a k n_K a where n (k+l) for which Kn < a n+ i/an for all M We assert n there exist arbitrarily large positive integers k, -i suppose 0, < Mk ankX -i If this assertion were true, then clearly we could choose two strictly increasing subsequences of positive integers (n k) k) and (m such that 0 ankxmk For each fixed positive integer x m (a (k+l)) e -i n for which n,m large M(a k) n x -i (k+l) -i < a x n m < Mk -i if and only if All we need show is that there exist arbitrarily ]. [(a (k+l)) e m k, to prove the assertion. k as -i M(a k) n n -i ]. x m x m [(a (k+l)) U [(a n>N M(a k)-l] n x m -i n 0 n M(a k) -i for every n (k+l))-i M(a n k)-l]. as U [(a For each -i (0,e) N, < n (k+l))-i m > N, Xm let M(a k) n n m there exists 0 e e for some n _> (k+l))-I 0, since such that -1].). denote the least positlve integer which exists since an as n N, Indeed, the proof below can be However, this is not the case. m [(a This would imply that cannot contain any interval of the form n>N k)__ M(an+l (Note: m In other words, for every n>N used to show that for every (0 e) n,m > N. for which N Suppose to the contrary that there exists a positive integer and hence n such that M(an+ik) -i 0 DOT PRODUCT REARRANGEMENTS as . n For Also, since a n +i M(a we have x n +1 k) < M(a m-- -i < (a -i M k) m M(a (a < m k a x nk n n k) k implies m sufficiently large, m n -i] k) m Therefore (k+l)) m, for This implies that m m, for infinitely many or equivalently, which contradicts our assumption m, Hence (2) is proved. n. whenever m k a- x can assume that an x k} > m n . n a + n x n n > m n 2 k m k k 2 ankXn (n determine subsequences of k), (m for which n To see this, let k. {k 1 ankXmk + nan x i k k) [ k <_ mk} x a keZ 2 which satisfy conditions addition satisfy n k a > m k x Then by n n nk and b for all Z denote the set 1 Then m k Therefore n i k e Z (mk), and n a n 1 in the 2 nd n k m k Let m k Z k) (x and of a and x m k k condition of the theorem, and in (a n k. n k > m k for all are strictly increasing (a property of subsequences). k 1 (in increasing order) Next we assert that without loss of generality we may assume To see this, note that we have x k which for simplicity we again call (n This gives us subsequences 1 a a- x respectively, by taking only those entries k e Z a we next assert that without loss of generality we [ keZ Z (a denote the set Z2 x ke Z suppose (2) holds for 0, n and hence (3) holds. for every k and let {} , .a k x and To see this suppose S(a,x) [kank mk k,j. M(a m-- Hence, for infinitely many ]. (k+l))-i x we have for all m k) x 0 as ’and (x such a x m n m k k k k We first assert that without loss of generality we may Assuming that and n +1 m n > N m-- for infinitely many for all the lemma we have that k [(a n , k) m < M and n M(a m M(a n so that there exist subsequences assume that {k x we have M(k+l)/k x, k (k+l)) n k To prove (2)/ (3) and Therefore m and (k+l)) n =. m m as 0 and nm+ i m an+l/an_ that [(a k) n m -I m, m an m+l/an n x x m m, x m infinitely many M(a k)_ n +i m and hence and for these M(a sufficiently large, we have m 413 and that Therefore if n # k < n n k for all m0 k > and < m k> m. for 414 P. ERDOS and G. WEISS some k, j, then and k most once among the m0 ’s. n M k (nk2’mk2) mk2+l exists, such that (n S mk3) k 3 that mk 3+i e S n such that 1 k n mkl+l where 2 j. kl ,m I,S 2 S 1 for some k <_ kl+l e S or 1 S <_ k I or 2. and in each set, no + a $2 n S and so x k m k appears as an 2 <_ k < k < either > m k Z a S1 a n a <-- ankXmk a" a m k 0 n as a x a m k Put (nk2+l, mk2+l) Continuing in this way, if k3+l. n S or 1 S Othez-ise is finite. 2 (n when both mj k n . ankX x or m k k a $2 n k k) (nj,mj) (nk,mk) m a m k n 1 Choosing x m k k x k S n M k for all m0 + m k or 1 (i.e., with the desired properties, (nk,mk) a" x (x m k n m . x < k We shall now show that for every tive integers such that and k (a x k m k n + a n x k k x a m k a a m n k k (a 0 n n x k x k n n n and k,j n (x a nk mk k, a (x m 0 x m n k k x m k k k [ k an x k n n k k we have k a k ). x mk we h ave m k x m k there exists a subsequence (a k [ k an amkX 0, e [ k e n Since and so x x m n k k k Furthermore since . x k(an k amk) (xmk xnk). 0 < x and k k k m m k k [k(ank + I. is the k.). for every k kl+l Put k2+l. a) and b) in Theorem 1 and also satisfying Now consider the series 0 where I Therefore m0 accordingly we produce the sequence 2 k k k satisfying n n k < k can occur at k is a disjoint partition of the set of all SI,S 2 Then clearly 1 n is the least positive integer, if it k2+l For either case, no are infinite. S k no such least positive integer exists, then either both That is, is the least positive integer, if it exists, such k3+l for some nk (n for some k i (nl,mI) Put least posltive integer such that (nkl+l’ mkl+l) for all m.l of posi- (k) n This follows from the nk following more general fact. Suppose [d(k) that (d(k)) is a non-negative sequence for which We assert that very every e [d(kn ). d(k) 0 as e > 0, there exists a subsequence and k (k) such n The proof of this fact proceeds along the same lines as the proof of Riemann’stheorem on rearrangments of conditionally convergent series. Fix DOT PRODUCT REARRANGEMENTS e > 0 and choose n >_ N I I d(k) < e so that I k=N <_ d(k) < e 1 Choose N choose n n 2 > n I k=N d(k). 1 so that I I [ d(k). n 2 [ Hence y d(k) < e k=N 1 > N k > N p > n d(k) < e d(k) d(k))/2 . 2 d(k) as 0 for every N d(k) k/ and N k n (n) P 2 d(k) k=N d(k). e 2 Proceeding inductively of positive integers for which q d(k))/2 k=N q=l and then n k=N2 and P 2 such that 2 [ < k=Nl (N) 0 < d(k) < (e p-i I Hence n2+l I p-i p-- n 1 [ d(k) < (e in this way, we obtain sequences n and so that k=Nl This can be done since n 1 to be the largest integer greater than 2 [ such that 1 >_ N l, k 1 k=N n k=N N nl+l nl for every n is the greatest integer greater than 415 p-I for every and every p q and p n p n p-i I q=l d(k) < e k=N P n +i q p < d(k) k=N d(k) k=N q P This implies that p n d(k) < d(n q=l k=N n p-i q [ 0 < e P + i) I q=l (e q e/2 p-I q k=N 0 d(k))/2P-i q as . p n Therefore q e d(k). q=l k=N Hence, if we choose (k) to be the strictly increasing n q sequence of positive integers U {k N p=l < k < n p-- k, we have p k where e d(k n Applying this result to the sequence negative, tends to 0, exist subsequences of ()’K for which e and sums to k) Yk (ank (n and (mk), Let and e so th at e S(a,x). r- a. x Suppose (x x ), since it is nona m m n k k k k > 0, there we obtain that for every (a n which we shall again denote by (n k) and amk) (xmk xnk) Now recall that we wish to show that a- x is taken to range over the set a. x S(a,x) r <. [a- x, ]. We already know It suffices to show and choose subsequences which we again call (n k) r e S(a,x). and k) (m 416 P. ERDOS and G. WEISS k(an k amk) (xmk xnk) e We now choose and n ni k) z(m n for which for each k n I n k, and let for every nk,mk for all n . I anX(n) n I a x + n n Z [ a x + n n a-x+ r e S(a,x) and so THEOREM 2. Z (a (a) a If a (a [ (a n n x + a m /a x k n + a )(x m Conversely, if /a n+ i’ a n S(a,x) [a I S(a,x). letting Indeed, n. x a a k anXn(n) (x m k x m k n )) k k a . n x an+l/a n- Then S(a,x) 0 n x 0, n a for every (x) n x is [a + n (x) n a and [a x n 0, 1 ]. x, for which ]. x, x then S(a,x) That is, x. > a z _i is not bounded. an+i/an >_ k x Define [a" x, x (k) Z + if > a >_ 4 k h (k) k 3- k ak+l/ak and n (a n z h (n) h(n) Let >_ 4 3n) -I n % 0, we -i +iXk > 1 3 n)-I (k) _< Then 3 -n h(k) denote the h (n) Clearly x n In fact, we claim that ]. anXn -7an(ah(n) be any permuation of I n Also by Theorem i, since x, S(a,x) we claim that z n for which n for which k a non-decreasing function of n for some m k n must remain bounded. least positive integer (x), x m k and 0 1 (x) x Suppose to the contrary that x or k Q.E.D. condition (3) of the theorem is satisfied by a n + (a x m m k k k k k satisfies condition (2) of the theorem. claim that n is bounded, then by Theorem i, if n+l’ n (i.e., those n x m n k k k k a n where n k is well-defined since which proves (3). Let m r g bounded if and only if, for every PROOF. {n k) Then a x + n n nZ n n(n Let as follows. The permutation denote the set n + fix all other integers k). Z Let i,j. (n) Hence Z the requisite permutation on for every m.3 k n 1/2 for some ah (k) +i (ah (k) 3k) -i is 0. Letting a" x 1 but i. Furthermore, k, then DOT PRODUCT REARRANGEMENTS -i On the other hand, if (k) h(k) lanXz(n) lan_ 1 (n) xn [anXn(n) In any case, i, for every < a ).3-n x 1/2 < 1 S(a,x). 1 hence then k, h (n) n 417 Q.E.D. In the proof of Theorem i, each time we constructed a permutation NOTE. [anX(n) solve the equation that we constructed was the product of disjoint 2-cycles. is, each such That it sufficed to use only disjoint 2-cycles. r, to w This seems odd and leads us to ask if there are any circumstances in which the use of In other words, is it always true that infinite-cycles or n-cycles yields more. S(a,x) [anX(n) is the same as disjoint Z z is a permutation of + which is a product of 2-cycles} ? The following question seems likely to have an affirmative answer. would give a characterization for those sequences and 0 x n which satisfy QUESTION I. If S(a,x) x and an ’ where / > 0, al However, it remains unsolved [a- x, are as above, does (3) x and a a If so, this = (2) in Theorem i? Finally, we wish to point out that Theorems i and 2 imply analogous theorems in which a and x Indeed, the proofs of the following two corollaries switch roles. follow naturally along the same lines as those of Theorems 1 and 2. Let COROLLARY 3. (x) x where n x n for all > 0 n, and x n 0 as . n Consider the following conditions. (i) Xn/Xn+ 1 (2) For the non-negative sequence and a) b) is bounded below. (x a n of x k m k .aKnk a 0 (a), n there exist subsequences (a k , k and x m k Then (i) implies that (2) holds for every strictly positive sequence that tends to n x, respectively, such that and as a a (a) n 418 P. ERDOS and G. ISS Let COROLLARY 4. x (x) n be a non-negative sequence. bounded below if and only if, for every S(a,x) [a x, QUESTION 2. S(a,x)? a (an) for which Then an + x /x n and is n+l a I > 0, ]. Is there anything to be said about the qualitative nature of Is it always a Borel set, measurable, F G ? REFE REN CE i. WEISS, GARY "Commutators and operator ideals", dissertation, University of Michigan, 19 75.