LIBRARY OF THE UNIVERSITY OF ILLINOIS AT URBANA-CHAMPAIGN 510c84 I£6r no. 111-130 cop .3 The person charging this material is re- sponsible for its return to the library from which it was withdrawn on or before the Latest Date stamped below. Theft, mutilation, and underlining of books are reasons for disciplinary action and may result in dismissal from the University. To renew call Telephone Center, 333-8400 UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN BUILDING USE ONLY SEP21IS80 SEP 2 1930 L161— -O-1096 Digitized by the Internet Archive in 2013 http://archive.org/details/remarksonroundof116desc 5/0.84 no. U6 cop. 3 REMARKS ON THE ROUND- OFF ERRORS IN ITERATIVE PROCESSES FOR FIXED- POINT COMPUTERS * by J„ Descloux UNIVERSITY OF ILLINOIS DIGITAL COMPUTER LABORATORY URBANA, ILLINOIS Report No. Il6 May 10, 1962 * This work was supported in part "by the National Science Foundation under Grant Gl6^89o re -j >— 3 REMARKS ON THE ROUND-OFF ERRORS IN ITERATIVE PROCESSES FOR FIXED-POINT COMPUTERS Abstract This report develops some ideas suggested by A. H. Taub on the round- off errors in iterative processes. It will be shown that in certain cases the convergence of the process can be improved by using special method for rounding Part I is concerned with simple first order processes. In part II, the 2 Aitken's 5 process will be considered. PART I. ERRORS IN FIRST ORDER ITERATIVE PROCESSES Introduction Let G , ... G be m real functions of the real variables x , ... x . For any set of m numbers p , . . . p , we shall use the vectorial notations: =? ( (1) (m)> P = (P V--P i-4 |p| = y(p ') + ... (p ) We consider the iterative process *n+l = G (*r? > n = °> 1} '" (l) and suppose there exists a vector r and a number b (0 < b < l) such that |G(x) - r| — > Theorem 1 . For any V , the sequence V given by (k) is bounded and all its -> points of accumulation V satisfy the inequality .— > ->i i—» l a 1 -> 1 (l) ( m )\ V - r < -t— r 1 a = (a' '. ... a' ') Theorem 2 . The process (h) is the best possible in the following sense: for given a and b, there exist m functions H (x) , . . .H (x ) for which it is impossible to find an algorithm using only H, a, b , providing closer points of accumulation to r than the algorithm (k) — > — >\ , -» — » Proof : Let G(x) = bx + a, H(x ) = bx G'(x) = bx* - a -» H(x) is an approximation for both G(x) and G'(x) with limits r = — T and - -, ■ a r = l^b -2- If any sequence W has a point of accumulation W such that |W - ?| < '* ' l-b ' then by the triangular inequality, |W - r'l > M 1-b and the process (k-) presides in this case a better information. Round-off errors For the computer, the process (l) can be written in the form .(i) = rn(i), (i) y n+1 = [^-(y n ) + |-] R ; (5) (i) • y is an integer, n [ ] is called a rounding procedure. [x]„ is any integer function of x satis- K K fying the inequality: |[x] R - x| < 1. We consider two particular types of rounding procedures: 1) normal rounding: [x] = [x + 0.5] ; 2) anomalous rounding: W/r for l x l < 1> ItxJ/J > l x l for |x| > 1, |[x] J < |x| -> -> Theorem 3 . Let G and £ satisfy the equations (2) and (3) . If £1 = CG (l) ( yn ) ♦ !*]„, 1-1,2...., (6) then for any y , there exists N such that K - ri vn 1 ^ bb + 271^7 for n > N ; furthermore, for given a and b, there exist a function G and the errors | for which the bound is attained. Now, we restrict ourselves to the particular case m = 1, i.e., the process (l) becomes scalar. Equations (l), (2), (3), and (5) can be written as: x _,, = G(x ) (7) n+1 n |G(x) - r| < b |x-r| (8) ^n + l - [G ^n> + ^n ] R (9) It I N. J n+1 ' 1-b Let us compare the theorem h with the theorem 3 for m = 1. In both cases, the bounds of errors have a common part which can be recognized from theorems 1 and 2 as provided by the truncation errors . The part due to the round-off errors is independent of b for the anomalous rounding; in particular, if a = 0, the error is less than 1 and if the limit r is an integer, it is reached after a finite number of steps. When the convergence is slow, i.e., b ~ 1, the errors can be very large for the normal rounding, even if a = 0; however, if b < 0.5, the normal rounding provides slightly better results than the anomalous rounding. Remark. The condition (2) insures a first-order convergence for the process (l) If we assume higher convergence, i.e., if |G (x) - r I < b |x - r | P , p > 1, we get results which are quite similar, but generaly not simple to formulate. Rather roughly, the theorem k becomes: if y is computed by (ll), then ly - rl < B + 1, where B is due to the truncation error. -h- Proofs — > — > Lemma . Let V and V satisfy the equation (k-) under the assumption (2) and (3) i->i _» i-*i a) if |V - r | < J3- , then |v. - ?| < M 1 ' o ' — 1-b ' '1 ' — 1-b M o * ' ' 1-b ' "*"*" ' "1 * ' ' ' "o b) if |V - r I > -\^£ , then Iv, - r I < |v - r* 1 Proof. Since V, = G (V ) + | : 1 00 .— » l v i — > V ■ — »l - r | < |a | 1-b — » V - r | < 1— >i |a | ■"*-"* ->| IT* I „ , I'lT ->l |-»l < |G(V ) - r I + |T I < b |V - r| + |a| (12) a) |V - r| < ■*—. rj- ; we have by (12) b , I a ^1^ + 1 | = tb ' ^- e ' d - .-» -* \a\ b) V - r > ■!—£ ', ve have by (12): ' 1 o ' 1-b i - * — >i i - * — >i / \ i - * — »i 1— >i \~~ * — >i 1— >i 1 — >i 1 — >i V n - r < V - r - (l-b) V -r + a < V - r - a + a = V -r ; '1 ' — ' o ' /| o l ' ' ' o ' ' ' ' ' ' o ' q.e .d. I-* ">l la I Proof of the Theorem 1. 1st Case: there is N such that V, T - r < N ' - 1-b ' by lemma a the same inequality holds for all n > N and the theorem is proved . i->l 2nd Case: for all n= 0, 1, 2. ...: |v - r*| > 4ht ', ' ' ' ' n ' 1-b 1 — >i by lemma b ; the positive sequence |V - r| is monotone decreasing and converges therefore to a limit I. 1— *i Q j V Suppose that i = -?—£■ + d where d > 0; since b < 1, there exists y such |y- - r| < JS +§ ; by 12: n ->i .. d i-»i , i-»i a Vn - r < — — a + d + a = -^—r}- + c = i, which is a contradiction. • 1 ' 1-b ' ' ' ' 1-b Proof of the Theorem 3 - Since |[x] - x| C 0.5, we can write the equation (6) in the form -5- y (i) = G (i) (?) M W i = 1, 2, ... m where It/ 1 ) I < a' 1 ' + 0.5 , 1 n ' and therefore U | < lal +0.5 vm . — > — > i— >i i— >i r Replacing £ by tj and |a| by |a| +0.5 vm , we can apply the theorem 1: for any £> there exists N such that i-> ->i . 11*1 + 0-5 vm _ . .__ y - r < J — L r rr- for n > N; ,J n ' 1 - b but since the y s are integers, there exists a particular e for which the preceding inequality implies i-> -* |i*| + .5 Vm ^ T .-. y - r < J — L ^ — — ; for n > N, as desired. 1 n ' — 1 - b We have still to show an example valid for every a and b where the bound of error is attained. Let G (i) (3 =bx (l) - a (i) - 0.5 —> — > — ¥ and suppose that for the particular vector y = we have £ = a . rm, ~* n * \^*~~ " ~*\ |a*l + v m • 0.5 „ . _ Then y = and y - r = J — ■— - — — ; — for n > 0. J n |J n ' 1 - b - Proof of the Theorem h . We use the two simple properties of the anomalous rounding procedures: 1) x - 1 < [x] < x + 1 2) if p < x < q and q - p > 1, then p then ' y l " r ' < l y o " r ' -6- Statement I: by lemma a: r - 1-b - J o < y„ + G(y J +| - y < r + o o o — 1-b ' by property 1: r - a —- - 1 < y + [G(y ) + | - y ]. < r + -7— + 1, 1-b J o w o' b o J o A 1-b i.e. jy. - r| < r + — T" + 1 > q.e.d, a Statement II : We suppose r + -r-~T < Y < r + — jT + 1 ("the proof is analogous when r - -r— — - 1 < y < r - — — ) ; by lemma b: 1-b J o 1-b p = r - -7— r- - 1 < y + G(y )+| -y r, q - p > 1 and we apply the property 2: r - t^t- - 1 < y + [G(y +| -y] A ly-L r < r + — - + 1. q.e.d. ' 1-b Statement III : We suppose y > r + — — - (the proof is analogous when y o $ r " l^b ) ; by lemma b: p=2r-y 1: 2r - y < y + [G(y ) + I - y ]. < y J o J o yj o / b o J o A J o 1 .e |y 1 - r| < |y Q - r q.e .d, -7- PART II. ROUND-OFF ERRORS IN THE AITKEN'S 6 2 PROCESS Let G(x) a real continuous function of the real variable x such that (1) the sequence x defined "by x = G(x ) n+1 n' converges to the limit x = r, By the Aitkens 5 process, we define another sequence: 3n+l v 3n' V = 0( W V V V -V 3n 3 n +2 3 n +I 3n+2 " V x +V -2V . 3n 3n+2 3n+l (2) Let us suppose we want to realize the process 2 on a fixed-point computer with the following conditions: a) We use only one "word" for repre- senting the V.'s; we may consider the content of the word as an integer ; b) we may use higher precision for computing G(V.). We cannot expect to compute G(V.) without error; furthermore, if usin^ higher precision, the result must be rounded to an integer . Definition o A rounding procedure denoted by [x] is any function of the real variable x satisfying the inequality [x] R x < 1. We shall use the following particular rounding procedures: 1) [x] : rounding away from zero ; it is defined by the inequality \[xf\ > |x| 2) [x] : rounding toward zero ; it is defined by the inequality |[xf I < 1*1 Example . Let G(x) = 7/8 x and V = 8; by (2), we have v i -7 v 2 = 6,125 V 5 = °- -8- If we ■want to represent the V. 's only "by integers and if we use the normal rounding procedure, we shall find: v l = 7 \ = 6 s = 00 It will he shown that this situation can he improved "by using the following integer process: W, , = W_ + [G(W_ ) + |_ - W_ f 3n+i 3 n 3 n 3 n 3 n VL _ = w + [G(W_ -) + |_ , - W_ ] 3n+2 3n 3 n +l 3^+1 3n W_. _ = W_ + 3n+3 3n (w_ - w_ _) 3n 3n+l 2W^ . - W, - W_ OJ 3n+l 3n 3n+2- ) (3) I, and ^ . are the errors of computation of G(W_, ) and G(W_, .,): 3n 3n+l 3 n 3 n +l since the numerator and the denominator are integers, it is possible with the help of the remainder to compute W without any error; if the numerator and the denominator are simultaneously equal to zero, then W = W = W and we set VL , = W_ . 3n+3 3 n Theorem 1 , We suppose there exists the numbers < h < 1, < c < 1, 5 > 0, l > 1, such that : 1) |x n - r| N. 1 3 n 1-b Theorem 2 . We make the assumptions : l) The convergence of the process (l) is alternating, i.e. for r-£ 0, if X - r < 0, if X = r; 2) 'The errors £ and £_ ' Jn 3n+l in (3) satisfy the inequality I . < a < — , where a is a fixed number. J ■ - - 3 Then , for any W belonging to the interval [r - I + — , r + I - —} , there exists a finite number N such that |W, - rl < 1 + a for n > N, 1 3 n Remark . The assumption (l) of the theorem 2 is sufficient for providing the convergence of the V 's satisfying the equations (2) for any r-£ < o 2 1 The integers p, q_, s, t are defined "by q = p + 1; s = q + 1; t = s + 1; q < r < s. Lemas The following lemmas, except lemma 1 and 2, are valid only under the assumptions of the theorem 1. Lemma 1 . The relations 3 are invariant for the transformation W. = -V. , g'(x) = -G(x) , i ± = -i ± , i.e., if W', Wl, W' are computed from W' = -W by replacing G by G' and £. by £! in 3, then W^ = -W , W' = -W , W' = -W . Proof. W* = W + [G'(W') - W + £'1* — 1 o L x o' o o 1 = W + [-(G(W ) - W + I )]* o L v v o' o o /J = -W - [G(W ) - W + if -- o L x o' o o J W • the proofs, based on the properties [-x] = -[x] and [-x] = -[x] and the same for W' and W' . x o X 2 " X l Lemma 2. Let x , x n , x^ and real numbers and x^ = ■ — r— — — — o' 1' 2 3 x + x - 2x ax o > ; ^ > 0; 2 1 bx (x 1 - x Q )(x 1 - x 2 ) Sx 2 - 1 (x + x_ o 2 2x 1 )' b) If x > x. , there is the following scheme of variations for x 7 , o 1 ° 3 as function of x 2' ■11- UNIVERSIT ILLINOIS LI8RARY f 2x x - Xq f x f t * x T + °° o ' - / x x ;|(x o + Xl ) ; c) If x > x > x , then x > x > x ; if x Q = x 2 if x ± = x 2 then x = - (x Q + x ± ) ; then x^ = x_ . 3 d) Considering x as a function of x . x , x , one has x + xx - 2x x ~dx dx 2 x " x 3 ^ x o ~ ^ x 3 ~ Lemma 3 . If W Q > r + 1, then W < W Q . Proof. By assumption 1; u n - r < b(W - r) 1 — v o u < W - (I - b)(W - r) < W - (l - b) l—o o ' — o u n + £ n < W - (1 - b) + a < W - (1 - b)(l 1 1 — o — o 1 (1-^)(1-C) N < „ W l = W o + t u l + h - W ol < W o ^ e ' d ' Lemma k . Let V > r + 1, V } V , V satisfy the relations (2) and let V.,, V rt , V^ be such that 123 V < V < V + d (1 + 8) V 2 > V 2 - d (1 + 8) V, = V V - V, 02 1 5 V + V - 2V n o21 then V^ > 2r - V 3 o ■12- Proof . By assumption 2: V, > 2r - V . By assumption (l) : V < V . Let x = V + a x 2 = v 2 - a f s V o X 2 - x l y v o+ x 2 - 2Xl We have: x,(0) = V^ > 2r - V 3 3 o x 3 ( j ) = - - Since x^(oj) is continuous in the interval V o + V 2 " 2Y l\ 0, — J , there exists V + V -2V < B < — i for which x,(b) = 2r - V . 3 3 o It is easy to check that V n + S < V . 1 o By lemma 2a , for every V < V < B and V > V - B , one has V^ > x,(B) = 2r - V . 3 3 o For proving the lemma k, we have to show that d(l+8) 0; (l-c«) .\2 if we replace c' "by -c and (V -r) by 1, we get the inequality: P> {-(2+V) + J ? + 2b' 2 ^^ 3C + fib'c | ; We set A = (2+b : ) 2 and B = J ' so that > Vp - *Ja . It is easy to show that l6 > B > A; by the mean value theorem: / B - nTa > a. (b-a) > (1 -^V c) (1 r^ )2( ^ c) aTi B 4(l+c) Ml+c) = d(l+6) , i.e d(l+6) < p as desired. Lemma 5 • If W > u.. > w n > r , there exists z such that - oil G(z) = W n and W > z > W n . 1 o 1 Proof. G(x) is a continuous function with G(W ) = u n and G(r) = r: o 1 fo r u, > W., > r, there exists W > z > r for which G(z) = W n . By 11 o 1 assumption 1, z > W. 1 Lemma 6 . For any z satisfying the conditions W >r + l>z>r and G(z) = W n , o — 1 we have W p > W . Proof. Let u zu 2 - WJ 3 z + u 2 - 2W (5) By assumption 2: u_, > 2r - z > r - 1: "by lemma 2d, if we replace in (5) u by r - 1 < u and z "by r letting W unchanged , we have to replace u by y with u > y + 1 (r + 1) y - W 1 r + 1 + y - 2W n r - 1, -14- W 1 - 7 = 1 - (W 1 - r) 2 -l <2 > consequently: u > J > W 1 " 2 ' W = W + [u_ - W + | n Y > W + [W_ 2 o'2 o "1 — o 1 w„ - if , by assumption 1: W < Z < W ; therefore W n - W < and [W n - W 1 o 1 o w 2 > W i » q.e.d. I ] ' - W l " W o ' Lemma 7, IfW >r+l+ o - — , then W„ < W : if the sign = ' holds, 1 - b ' 2 — o ' W - W n > 2, O -L — Proof . By assumption 1: |u x - r| < b(W Q - r), |w_ - r| < b(W - r) + a + 1, 1 1 o ' i i 2 u_ - r < b (W - r) + ab + b, '2 o u_ + i_ < r + "b (W - r) + ab + b + a, 2 1 o u 2 + | 1 - W q < (b 2 - i)(W q - r) + ab + b + a < (b 2 - l)(l + 3-^) 2 + ab+b + a^b - 1 + b < 1, consequently: [u + | - W J < 0, W 2 = W o + [u 2 + l x - W o T < 0, q.e.d. For proving the second part of lemma 7j> suppose that W - W , but W - W < 2, By lemma 3> "the only possibility is W = W - 1 for which W - r > - — — r- . By assumption 1: |u - r| < b(W - r), u 2 < r + b(W x - r), u 2 + i x - W 1 < (b - l)(W L - r) + a < (b - l) Y^Tb + a = °> U 2 + h < W l W = [u + I ] < W < W , which is a contradiction. 2 2 IK — 1 o -15- Lemma 8. If W > r + 1 + and W, < 2r - W , then o - 1 - b 1 — o' W_ > W_ > 2r - -W - 1 2 1 o Proof. By assumption 1 u > r - b(W - r), W = [u + | ] > r - b(W -r)-a-l = 2r-W +(l- b)(W - r) W ' > 2r - W + (l - b)(l + 1 o - a - 1; a \ a - 1 = 2r - W - b > 2r - W - 1. o o 1 - V For the second part of the lemma, we use again the assumption 1: u 2 > r - b(r - W 1 ), u 2 > W 1 + (1 - b)(r - W x ) > (1 - b)(l + 3-f^) y U 2 + ^L - U 2 a > W + 1 - b > W , u + e -W>W - W 2 ? 1 o 1 o' by lemma 3 • W - W < 0; [u 2 + | x -W o ]"> Wl -W o , W 2 = W q + [u 2 + l x - W q ]* > W x , q.e.d. Lemma 9' If W > r > W n , then W > W n , o — — 1' 2-1 Proof, By assumption 1, u > W ; u 2 + \ > \ a > W^ 1 k ' W^ = W + [n + | n - W ] > W + [W T 2 o 2 1 o — o 1 \'\f = \, q.e.d, Scheme of the proof of Theorem 1 The theorem results from the two statements : l) If |W - r| > 1 + n a . , then |w^ - r| < |w - r| for r o — 1 - b ' '3 o + i - 1. I + 1 < W < r — o — 2) If [W - r| < 1 + , then |w - r| < 1 + — -16- We can restrict ourselves to the case W > r. Indeed, suppose that the state- ments i) and n) are right for W > r and consider a particular value W < r. We set W ' = -W , G'(x) = - G(-x), |.' = - |. and compute W ' W ' , W ' by the equations 3 with the ' values. Since |.', G'(x), -r satisfy the same hypothesis as I., G(x), r, ve have: I) If |W ' + r| > 1 + o |W_ - r| < |W - r 1 3 o — i then W'+r r + 1 + o — I — - , then |W 3 r I < I W - rl o Proof We distinguish three cases : 1) W > W 1 — o 2) W > W, > W„ o 1 2 3) W > W n and W„ > W n o 1 2—1 l) By lemma 3> this case never occurs. 2a) W > W > W and W n > u n = G(W ) . o 1 2 1-1 o o 2 V x 2 V + V - 2V n o 2 1 We define V = W , V = u n = G(W ), V = G(u n ). V n o o 1 1 o 2 13 Since W_, = W + [u n + | - W and W n < W : 1 o 1 " o o 1 o W < u + a < u + d , By assumption 3: |u -V | = |g(W ) - G(u )| < o |w - u | < 5 d; Since W = W + [u_ + £.. - W and W^ < W we have: 2 o 2 1 o 2 o W 2 > u 2 " a > u 2 " d > V 2 " d (! + 5 ) • We set V. = W. and apply the lemma h, the V.'s keeping their signification W n > 2r - W . 3 o By lemma 2b : u < W ; W_ = [u l < W o < W so that 3 3 R _ 2 o |W 3 - r| < |W q - r|, q.e.d. -17- 2b ) W > W., > W_ and W n < u = G(W ) o 1 2 1 1 o By lemma 9 : W > r, By lemma 5, there exists Z such that W > Z > W_ and W n = G(Z). o 1 1 By lemma 6 : Z > r + 1 „ Since W^=W + [u„ + i_ - W < W , we have W^ > u. - a > u_ - d. 2 o 2 1 o o' 2-2 -2 We set Z = V , W = V , u = 1 , W = V , W = V and apply the lemma K ZW - W 2 V = * 77 ^77" > 2r - Z > 2r - W ; 3 Z + W^ - 2W n o 2 1 Since W > Z, by lemma 2a: W W - W 2 W = — - — — > V > 2r - W . * W + VL - 2W, ^ ° o 2 1 By lemma 2b : u < W ; W_ = [u_]_, < W < W so that 3 3 R ~ 2 o W^ - r < W - r , q „e„d. '3 o 3) By lemma 7> W _ < W , we distinguish: 3a) W > W > W, > 2r - W : o 2-1 o ' 3b) W q = W 2 ; 3c) W > W^ and W n < 2r - W . o 2 1 — o 3a) By lemma 2c: W > u > W ; since W , W are integers: W. > W_ > W_, i.e., |w_ - r| < |w - rl q.e.d. 2 — 3 — 1 3 o 3b) By lemma 2c: u = |(W + W ); by lemma 7 ': u. < ^(2W - 2) = W - 1; 3 — 2 o o W_ = [u_]_ < W - 1 < W . 3 3 R - o o By lemma 8: W > 2r - W - 1; u 3 > r - § ; W = W + [u_ - W ] >W + [r - =■ - W ] >r-^>2r-W; 3 o 3 o o 2 o — 2 o' therefore: W > VL > 2r - W , i.e., |w„ - rl < |w - rl q.e.d. o 3 o 3 o -18- 3c) By lemma 8: W > W_ > W_ > 2r - W - 1; -> ' J o 2 1 o by lemma 2c: W > W > u_ > W. > 2r - W - 1: J o231 o since W„ = W + [u_ - W ] : 3o 3o W > W > 2r - W , i.e., W_ o' 3 2 - 3 Statement II r < W - r o q.e.d. If r < W < 1 + - — o 1 -^ , then |W 3 - r| < 1 + ^—-^ Proof. We distinguish two cases and describe for each of them all the possibilities without any computation: 1) W Q - s; 2 ) W = t and s - r < a . o l) W = s o W 2 = s W 3 la) W = t /W_ = q : W = s w 2 = p 3 V =s 2a) W = s W = s (u does not depend on the value of W ), 3 3 w 2 = t 3a) W = q: / W = s W 2 = q W 3 =s W 3 = s W 3 =q ka) W = p w 2 = t W 2 = s W ? = q W 3 = s W 3 =q W„ = q 2) W = t and s - r < a 2a) W w 2 = t W^ = s W„= t W„ - s 2b) W = q W« = t VL = s W, W. -19- w„ = t 2c) W n = p h. = w o = q w. W 3 = s w 3 = , Proof of the Theorem 2 We shall use the same notations as in the proof of the theorem 1. Lemmas The following lemmas are valid only with the assumptions of the theorem 2 . Lemma 10 . If W > t, then W < s; the sign "=" holds only if s - r < — . Proof . By assumption 1: u < r; u n +! < r + - < t; 1 o - 3 W . = W + [tl. + I - W 3 < t; 1 o loo W can equal s only if r +^> s, i.e., s - r < — , q.e.d, Lemma 11. If W > r + 1, then V/ > W n . o — 2—1 Proof . a) Suppose W < q; then: u 2 > v > u 2 + h > P ' W 2 = W q + [u 2 + l x - W q ] > P > q > W v q.e.d. b) Suppose W > s; since W > t, by lemma 10, W = s and s by assumption 1: u > r -~^> q; 1 r< 3 ; W 2 = W q + [u 2 + ^ - W q ] > q i.e. W > s = W q.e.d. -20- Lemma 12 , Proof, If W > r, then W < W + 1, o — ' 2-o By assumption 1: |u - r| < (W - r); |W 1 - r| - |[u x + | q ] r - r| < \u ± - r| + k/ 3 < (W Q - r) + h/3; |u " r| < |W 1 - r| < (W q - r) + k/3', U 2 < W o + ^ /3; u 2 + fi x < w o + 5/3; V7 = W + [u_ + I. -W f < W + [5/3] = W + 1, q.e, o 2 1 o — o ' o Lemma 13- If w = w +1 and W < W - 3, then W n < W . 2 o l-o 3 o Proof. If we keep W constant, u is an increasing function of W by lemma 2a: since W_ = W + [u_ - W , W_ will have the same 3o3o3 property and it suffices to prove the lemma for the case W = w - 3» one finds: 1 o J ' u = W - 9/7 < W - 1 3 o o W Q = [uJ D < W - 1 < W , q.e.d. 3 3 R - o o' Lemma ik . If W„ = W > r + 1 - a, then W n < r and W_ < W . 2 o - 1 3o Proof . Suppose W > r; by assumption 1: u < r, u_ + (L-l, i.e.", W - W, > 2: o 1 o 1 - by lemma 2c: u = J(W + W n ) = W + ^(W n - W ) < W - 1: 3 2 o 1 o 21 — w 3 = [u 3 ] R < W q - 1 < W q , q.e.d. Lemma 15. If W. > W > r + 1 + a, then W n < r - 1 and W n < W , 2 o — 1 3o -21- Proof, Suppose W > r - 1; by assumption 1: u < r + 1, u^ + L 2. i.e., W - ¥ > 3; o 1 ' ' o 1 — ~" by lemma 12 : ¥^ = W + 1: J 2 o ' by lemma 13: ¥ < ¥ , q.e.d. Lemma l6. If W > r + k/3, then ¥_ > 2r - W o - ' -" 3 c Proof, a) Suppose ¥ > r; by lemma 11 : ¥ > ¥ > r; by lemma 2b : u > ¥ , ¥_ = [u_]_, > W n > r > 2r - ¥ , q.e.d. 3 3R-1- o' b) Suppose ¥ < r; we have the inequalities: |u x - r| < (¥ q - r), u > 2r - ¥ , 1 o n + | > 2r - ¥ - \ , 1 o o 3 W l = [u i + l o\ > 2r ' W o " ^ 3 - \ ' u 2 > r > u 2 + I-l > r - i ¥_ = ¥ + [u + L - W f > W + [r - -| - W ] > r - ^ = ¥ . 2o 2 1 o — o 3° — 32 By lemma 11, ¥ > ¥ ; by lemma 2a, for fixed ¥ , u_ decreases when ¥_ and ¥ decrease; consequently: u, ¥ W„ o 2 ¥ n > ¥ + ¥ o 2 2W, ¥ VL - ¥ n 2 o 2 1 ¥ + ¥ - 2¥ n o 2 1 = B -22- k The fact that B > 2r - W for W > r + - results from the o o - 3 three statements : 1) B is a continuous function of W for r < W < °° : o — o ' -W 2 ) for W -> oo W and therefore B > 2r - W when W o 3 3 o c is large enough. 1 + ^33 ^ , k — r t ~ o r + 7 < r + — , is the only value in [r, °°] for which B = 2r - W . o h We have therefore established that u_ > 2r - W for w > r + — . Now 3 o o - 3 W„ = W + [u_ - W Y > 2r - W , q.e.d. 3 o 3 o o' Lemma 17. If W > r + 1 + a, then W„ > 2r - W . o — 2 o r k Proof . By lemma 16, we have only to establish the lemma for W < r + — , i.e., 1 W = t, s - r < -. o 3 a) Suppose W > r; by the same argument used in lemma l6, we con- clude that W n > 2r - W . 3 o b) Suppose W < r; o u>2r-t>t--, u 1 +l 1 >t-|=p, W [U 1 + Ur ^ P ■ u 2 > r > u 2 -i - r " 3 q 3 ; W 2 = W o + [U 2 + 6 1 ~ W o ] > q + \ ' i • e • , W 2 - S; by lemma 11 : W > W ; by lemma 2a : 2 2 W W - W n ts - p o 2 1 ~ 1 U 3 " W + W - 2W n - t + s - 2p q + 5 ; o 2 1 ^ therefore: W_>s>r>2r-W, q.e.d. 3 - - o' -23- Scheme of the proof of the theorem 2 The theorem results from the two statements : 1) If |w - r| > 1 + a, then |w_ - rl < |w - rl for r -,#+-< W < r + I - -. 1 o - 3 o 3-0- 3 2) If|w - r| < 1 + a, then |w - r| < 1 + a. Using the same argument as in the proof of the theorem 1, we can restrict ourselves to the case W > r. o — Statement I: If W > r + a, then W_ - r < W - r, i.e.. o - 3 o 1 ) W < W 3 o 2) W., > 2r - W . 3 o 1) We prove that W < W ; we distinguish three cases: a) W < W ; by lemma 11, W > W : by lemma 2c, u < W ; since W_ = [u_] D , W_ < W < W . 3 3 *< 3 — ^ o b) W^ = W ; by lemma 1^, W„ < W . 2 o 3 o c) W > W ; by lemma 15, W < W . 2) By lemma 17, W > 2r - W . 2 o Statement II. If r < W < r + 1 + a, then Uj - r < 1 + a. - o 3 We distinguish two cases; for each of them, we describe all the possibilities without any computations: 1) W = s o 2 ) W = t and s - r < a . o l) W = s o lb) W = s : W = s (u does not depend on W ) -2k- W„ = s W^ = s lc) W = q W = q : W = q Id) W = p W, w ^ = q vi = w„ = 2 ) W = t and s - r < a 2a) W = s W^ = t W„ = s W = t 3 W = s 2b) W W„ = t W^ = s W 3 =s W = s 'W,= t + 1 . : W,= s 2c) W n = p J W^ = t W, = s W. ■25-