2 Continued fractions We are now ready to begin our study of continued fractions. 2.1 A n-overview of the theory In the years BC—before calculators was 22/7 and in the years AD —after decimals~ x became 3.141 59265... Clearly, x is reasonably well approxi- ‘mated by the vulgar fraction 22/7, and some of us know that 355/113 does a yet better jab since it yields as many as seven correct decimal digits. ‘The ‘why this is so” of the matter is as follows. It happens that en and, in particular, that nyressh hile sssitsess Obviously, the notation takes too much space. We also note that truncations of the neverending fraction (2.1) seem to provide very good rational approxi- ‘mations. (We have coined the term neverending fraction as a synonym for an Infinite continued fraction, which is what equations like (2.1) represent, We will of course need to consider convergence issues.) 2B24 Continued fractions In example (2.1) it is only the partial quotients 3,7, 15,... that matter, so we may conveniently write (2.1) as m= [3:7,15,1,292,1,..-]. In general, given an irrational number a, define its sequence 09,01, of complete quotients by setting a = a and aq.1 = f(a — ay). Here, the sequence dp, a1, d2,... of partial quotients of a is given by dy = an, where LJ is the greatest integer function, introduced in Section 1.1, The truncations [ao a:,-.+de) are plainly rational numbers p,/dy- Indeed, the continuants py and qy are defined by the matrix identities fay Wfar 1) (dn 1\_ (Pr Pro Tol of (Too) ( a ‘This olows realy by indston on nan the dsiniton 1 ° Viana ln Taking dterinat inthe max conespondene above immediatly plies that the convergent pn/qy satisfy [ao as, ---+ de Br Bet CUP a2) de Geet dontde and so Pas ae showing thet the convergent do converge ta limit, namely Sep"! a=ayt . (2.3) ae ‘ and also that ify Pee tot o 1/(24). Exmecise 2.2 (See [51], [22, Section 1.1] and [105)) Verify that f way 2 rt x8 ~ 0825 4 816.2) 355 LSS eB and use the integrals to give an alternative proof of the estimates (2.4). Note that the integrands are strictly positive on the open interval and hence the inte- gral is strictly positive. I appears to be largely coincidental that this works for the two early conver- gents of x. That is, we know of no methodical way to undertake such a process of representing continued fraction errors by integrals of positive rational func: 2.2 Finite continued fractions and the Euclidean algorithm A finite continued fraction is the expression Lag; 4.035.» 2s) where dy € Zand aj,...,dy € N are the partial quotients of the continued fraction Asa result of the finite number of arithmetic operations involved in its def- inition, a finite continued fraction is clearly a rational number. This number is called the value of the continued fraction, Exarcist 2.3 Find the value of the finite continued fraction [1;2,3,4, 5]. Theorem 2.4 Any rational number admits a representation in the form of continued fraction (2.5) with last term ay > 1 ifm > 1 We shall discuss the uniqueness of such representations later.26 Continued fractions Proof Let a/bbe a rational number with a € Z, b ¢ N, and (a, b) = 1. Setting 1 = @and rq = b, write the Euclidean algorithm (successive application of division with remainder) as ras rom tn, mW=na tr, n= Ma +r, 26 Toad = Facet + Fay Toc = Falls where we have Oey 1 then a is not divisible by b, since (a,b) = 1. Therefore, r; > 0 and the sequence (2.6) consists of more than one iteration. It follows from the first equality that a < 0 implies ay < 0, while if a > 0 ‘and the fraction a/b is regular (that is, a < B) then ay = 0. As for the other equalities in (2.6), by the inequalities (2.7) we deduce that ay,...,d, € N and 4a, > 1 (the latter follows from the equality r)-1 = rpdy and the inequality Fa < Tat): Dividing both sides of each equality in (2.6) by ro, 71,---+Fasts ta respectively, we get 1 malta Starting from the last equality and substituting each equality into the previous ‘one, we finally arrive at the desired expansion (2.5) o We now see that the terms dp, aj,..., ay are ealled partial quotients because the successive divisions in the Euclidean algorithm involve these numbers as partial quotients of the corresponding quantities Exuxerst 2.5 (The Euclid game [39]) The game Euclid is played with a pair of nonnegative integers. Two players move alternately, each subiracting a positive2.3 Algebraic theory of continued fractions 27 Integer multiple of one integer from the other integer without making the result negative. The player who reduces one of the integers to zero wins Ina subtle variation, the game stops when the two numbers are equal. ‘When does a winning strategy exist for the first player? 2.3 Algebraic theory of continued fractions Let us now take a slightly different approach. Viewing dy.d;,...,dp,..+ as independent variables, define consecutively pairs of polynomials in these vari- ables Pa = Pro: Giy. By). Gu = Gul Oi.) by setting py = ao, qo = I and Py = Prot sta, 54) + Gyoa Qt a, Gn = Pras d2, +e) 2.8) We now introduce a convenient notation for the quotient of the polynomials: P= fagsas,... au] ra Then (ait = 06+ a9 Developing this equality further we obtain Fag; a... 54) = ag + = ay + 2.10) This is the expression we have agreed to cal a finite continued fraction, Example 2.6 Using the above definition we find that py=aoa)+1 and ge =a28 Continued fractions Resank 2.7 In fact, it is often helpful to define the polynomials p, and qy in accordance with (2.8) but starting from the initial data p-; = 1 and q-1 = for n= =I (when the set of variables of the corresponding polynomials is empty) ‘Then py = ap and go = 1 by (2.8), and this agrees with the initial data above, ‘Summarising the construction above, we associate the sequence of contin- uued fractions (2.10) with a set of independent variables ay.ay,...,dy,... IE ‘we replace the variables ap,q),...,4y,-.- With (not necessarily integer) real numbers such that the values of gq do not vanish then we get a sequence of numbers written in the form of continued fractions, ‘A simpler (and, in a certain sense, more natural) way of writing the paired ‘equations (2.8) is the matrix form ()-( Ge) en where the prime means that the set of variables is shifted by one: », (03425...) aNd gg = Aye) 3A2. My) erating the matrix relation (2.11) we obtain re) ao 1)fav 1) (anes 1) (ay ()-FF lt dF dl) on fay 1\fay 1) (deer 1)fay 1)/1 “Et lt dei Aft all) em Combining the representations (212) form and (213) for n~ L, we ave at [Lemma 28 (Key lemma) For each n > 1, we have the relations Pe Pes) _foo 1)fas 1) fama (an 1) _pyfay 2 (i moet ali ) (i ali dt H) ew ‘The matrix form (2.14) is a very useful basis for proving practically e thing about the formal continued fraction (2.9). Lemma 2.9 For each n > 1, we have the relations Pe = Anant Pras n= Annet + ed 15) Proof This follows immediately from (ee) erg) 82.3 Algebraic theory of continued fractions 29 Exercise 2,10 (Repeated Is) Find the value of the finite continued fraction ("= Osh) Tim oa (C1) At present we are interested in the ease when ay, aa... are positive real numbers; their positivity guarantees that g, > 0 for n > 0. Then the fraction a/da isa certain real number, and the definition 2.10) of a continued fraction together with Lemma 2.9 imply the following statements and compute the limit [(1)* Lemma 2.11 Leta, i tients (the ‘tails’) «n, be positive reals, Define the complete quo- r= [anders ‘ Then fags. sda] = Lays, ---see ra] = PME Pie $A = [aoe anor = DE "la; 1 _yfay 1\ yfay 1 fit tite AG Namely, we have ry = pj, where CG dle deb la then (r a fan NY far 1) (pi) _ (per pe-a\(P% al Kt oo) C1 olla Goer ana) Nat = (Pork me iP, + e204) hence fags ay... yey) = Be = PECNPL * Poca a Metal 0 Gm Gad eit +a Lemma 242 Let agi... and Bosbiy sb be postive real numbers satisfying a; > | and b; > | fori = 1 n. Assume that the numbers a;,b; are integers for =0,...,n—1. Then the equality [ag 2.5 g] = [oi BI. Pa implies that a; = b; for all i= 0,1,....30 Continued fractions Proof For n= lana... .q] = a + an] ‘we have ry > 1, and the same reasoning implies that s; = [by;b2,..., Bq] 2 1 By the hypothesis, weboned a6 fry = I then the left-hand side of (216) is an integer. Hence 1/5; is an integer as well, implying s; = I and by = ay. Otherwise, if 7; > 1 then the left- hand side of (2.16) is notin Z and we deduce that 51 > 1 and, consequently, day = Bo, Since both ay and by are the largest integers which do not exceed {a;a1,.--y0q) = [Bosbi,---,Ppl In all cases we get ap = by hence m1 = Induction on n completes the proof of the statement. a Lemma 2.12 actually shows the uniqueness of the continued fraction repre- sentation (2.10) (with dg > 1 if > 1) of a rational number; the existence was proved in Theorem 2.4, Exurcist 2.13 Deduce the uniqueness of the representation from Theorem 2.4, 2.4 Relations for continuants of a continued fraction In the statements in this section we consider ay, a, as indetermi- nates, which may even lie in function fields. ‘Theorem 2.14 Forn > 0, the relation nPa-t ~ Paens = (-1" holds: Proof This is the determinant evaluation of (2.14) after noting that Corollary 2.18 For n> 1, the following relation holds: Prot Pr _ 1" Get de anid Corollary 216 If ap, a;,43,... are positive integers then py and qq are co- prime and 0 < qi I, the relation Pret ~ Pores = (1)! ay holds. Proof To prove this relation, we can obtain from (2.13) the matrix identity [Pn Pra) _ (a0 ie 1) far A)[ernge 1 am wal \1 okt of V1 ofl am and then compute the determinants of both sides. Alternatively, multiply the first expression in Lemma 2.9 by q,-2 and the second by pyu2 and subtract the first equality from the second, In accordance with Theorem 2.14 we get GpPr-a ~ Peltn-d = Qn Gre Prod ~ Pr-idn-2) = (=D, finishing the proof. 0 Corollary 219 Form = 2we have Prod Py _ (0 m2 Gn Gre Corollary 2.20 If ay.4r,... are positive (not necessarily integer) numbers then the sequence Py/qn restricted 10 even n is strictly increasing, while re- stricted to odd n itis strictly decreasing, Corollary 2.21 Consider a = [ag;.a),...,dq.2] as a rational flnetion of the variables ao, dy,....dy43- Then gut — py = Dan Pe geal * In Proof Replace n in Theorem 2.18 with n + 2 and divide both sides of the formula by qy.2-Note that pys2/qn42 = @ by definition and gy. = dn424ne1*9n by Lemma 2.9. This implies the desired statement. 032 Continued fractions Exercise 2.22 Prove the following identities for n > 1 (@) alah ©) Pa/ Prot leiden sale Solution Transpose the matrix identity (2.14). o Exexcise 2.23 (H.J.S. Smith; see [46]) Forcontinuants py = ppl; ai,..-, dn) and gy = del 1... x) Show the following determinant expressions: Cn) o 0 -la 1 o 0 Oo -1 a 0 0 n= det o 0 1 0 0 ay, a) o 0 -la 1 o 0 oO -1 a 0 0 qn = det 0 0 2.5 Continued fraction of a real number ‘We have already seen, in Section 2.2, thatthe (finite) continued fraction of @ rational number a can be found using the Euclidean algorithm (Theorem 2.4) take a = la and, if@ is not an integer, then write it in the form a = ay+1/a, Where ay > 1 is again a rational number. Inductively, we choose a = Lata) and a = ay + Letns: With ane: > 1 if ay is not an integer. The procedure terminates at some step (that is, eventually we get an integer am = a > 1), so that er = [ap;a;,... dp]. If we discard the condition dy > 1 for m > 1 then the number a can be also represented as a = [ap;),...,aq— 1, 1]. This fact is sometimes useful for manipulating the parity of a particular length of a finite ‘continued fraction and for other similar reasons, ‘The recursive algorithm above extends to the case of an irrational number @ ‘with no trouble; however, at each step we obtain irrational a, > 1, so that the continued fraction cannot be finite, We will use the notation = fap, 4dgy 12.5 Continued fraction of a real number 33 for this infinite ease, From Corollary 2.16 we obtain a sequence of coprime integers py and ny defined as the numerator and denominator of each finite continued fraction ‘ag; a),..-, dy]; im addition, 0 < gy < gz <0 < qq <>. In this case, the equality Pn 4% becomes a relation between real numbers rather than between functions involv ing indeterminates ap, a\,.... Moreover pe/qa is an irreducible fraction, which is called the nth (principal) convergent of a. The number ais called the nth Partial quotient of the number a. In this notation, the statements of Corol- Tries 2.17 and 2.21 (after replacement of aqy2 With @y,2) take the following form. lay, = Lemma 2.24 For n > 0, we have the equalities pes Gne2dnet + Gn We stress that 1 1 nv = Past = ‘Theorem 2.25 (Monotonicity and estimation of convergents) For even 1, the nth convergents of « form a strictly increasing sequence converging to c; for add n, the rth convergent of form asritty decreasing sequence converging 10.0, Furthermore, 1 1 Bedi Glade Goad) Gdn Syst Proof The first part of the claim follows from Corollaties 2.20 and 2.15; Corollary 2.15 also justifies the estimate from above for la — pa/gal. To show the validity of the estimate from below we use Corollary 2.19: Ps]. [Pea _ Pal _ tno aa dn dnv2— Gel nna — nl Qne2Gnvt + Gud On dividing the numerator and denominator ofthe last faction by dasa 2 1a the theorem follows. a Since gyi > gy» We conclude that the principal convergents py/qy satisfy the inequality Exancise 2.26 (Dirichlet’s theorem) Using continued fractions, give a third proof of Dirichlet's theorem (Theorem 1.36) for an ierational real number34 Continued fractions Hint For a given integer n > 0, take the index j such that qy << qyy1 and show that the convergent p/q = pj/qj satisties 0; this is (a) Using (2.3) show that the partial quotients in the continued fraction ex- pansion of are all squares. (b) Prove that hhas a continued fraction [a9;a5,...,d,,..-] where for k > 2 each ay/2 is a square, (©) Finally, show that 2C = C" + 1 ‘The number C = 0.643.410546 28... is sometimes called Cahen’s con- stant; see [56] and [58, Section 6.7}2.6 A taste of Diophantine approximation 35, 2.6 A taste of Diophantine approximation Lemma 2.30 Let a = [ao,4),d3....] be a continued fraction and Paid n> 0, the sequence of its convergents. Then the inequalities 1 Pal 1 —_. nei, 2.17) Gat DR | Gel” Goaith en yy Sy S Felt) qn A 2.18) hold, where F, denotes the Fibonacei sequence (defined in Section 1.3), Proof By Lemma 2.24 we have the equality 1 Gnd Ge | Ps % Where on = [dqs1+dni2.---] is the complete quotient of the continued frac- tion. Therefore, the inequalities ayy) < opy1 < dyer +1 and 0 < duet < dy imply the estimates (2.17) To prove (2.18), we proceed by induction. The inequalities (2.18) are clearly true for m= 1 and n = 2, while for n & 2 the equality dest = david + det plies dy.idu S dost S Gnst (Ga + daet)- The latter inequalities together with the inductive hypothesis and the recurrence relation Fay = Fy + Frat for the Fibonacei numbers result in (2.18) a Exurcist 2.31 (Repeated Is; see Exercise 2.10) (2) Show that {1} = (1 + V5)/2. Thus the golden mean (or ratio), ¢, yields the most basic periodic continued fraction, (b) Evaluate (k}°] for2 < k = 10 Lemma 2.32 Let py-1/dy-i and ppd, be two successive convergents of an irrational number a = (ao;4y,43....]. Then at least one of these fractions satisfies the inequality Using the fact that a lies between pyai/de-1 and py/de (Theorem 2.25) we oblain36 Continued fractions This contradicts the inequality xy < (2 + y*)/2 for x > y > 0, applied with X= Ligue and y = Un o ‘Our next statement shows that, in a certain sense, the converse of Lemma 2.32 holds as well Lemma 2.33 (Legendre’s theorem) Let p and q be coprime integers, q > 0, and let 4| 1 al” 2g Then piq isa convergent of Proof Write the continued fraction expansion of the rational number p/ lq = (aniai,.-..4gh Let praildens and lq = Py/dy be the last two con- vvergents of this expansion, where we assume that both «and pyui/daci aC simultaneously greater or smaller than the number p/g (if this does not hap- pen then we replace the continued fraction with p/q = aoa, ....dy~ 1.1). Consider the number ei oe) ca int “Pm ew in Pe for which we have lox a, @e= pala” FR=p implying det) duct 5 yt 2 [p+ Set] Set 2 det 220 tas [pe Sa Sea oan ‘Comparing the latter inequality with (2.19) we deduce that roa = gn] > ipa — eal hence Get However, the numbers a and p,-/qy-1 are both either greater or smaller than ig, that is, o lies between Py-1/qr-t and p/q = Pr/qy. But then 6 > 0 in accordance with (2.19), and so 8 > 1 by (2.20).2.7 Equivalent numbers 37 (Sole in other words, p/q = Pp (qn is indeed a convergent of a, 0 Leo sos dyedipstss T= [aos ayy on 2.7 Equivalent numbers la 6) Ye dl with integer entries a, , c,d and having determinant 1 (hat is, ad — be = 1 or -1) is a multiplicative group with identity (neutral) element 10 e-( 4) Indeed the produto any two such matrices and ihe inverse of sucha marx again has neger ens and determinant equal to +1. This group is known the special linear group (over the rng Z) and is denoted by SL_(2); in what folows we reserve the neationT fo this group For an ational number o, the action ofan element € is defined by the rile The set of matrices tant 234 Show ata sel dnd mney. at Ee = ean y(éa) = (yd)a for all y, 6 €T. We sy hao oa ane and fa ue if = 6e Exec 235 Show har hielo ind an que Example 2.36 For an irrational number a we have the representation Preity + Pent Gna dea in accordance with Lemma 2.11. Define the (n— 1)h continued transformation ofthe number @ by the equality yy alms Pea) f(a meal i A38 Continued fractions note that y,-1 €T' by Theorem 2,14, Then a = yy-1dq and hence a is equiva- lent to a for any n > 1. In other words, all complete quotients a, = 1,2,..., are equivalent to each other. ‘The following theorem characterises the situation considered in Example 2.36. ‘Theorem 237. Let a,f €R\Qand prom raf et Assume that f > 1 and ¢ > d > 0. Then b/d and aje are two consecutive convergents of , Say, Py-2!dyen Gnd Py {qyaii furthermore, B= dy Proof Note that a and c are relatively prime since ad —be = +1, Write a/c as the finite continued fraction Pre Sm fags, sagen] = Get > e ai where @ = pyot and c = gps. Inreasing by 1, if required, the length of the continued fraction for a/c (namely, replacing a, with a,-y ~ 1,1), we obtain the equality Pridn-2 = Go-iPr-a 22) where e= ad ~ be. Then ad ~ be = py-id ~ qn-ib = € (2.23) ‘and, comparing equations (2.22) and (2.23), we deduce that Proild = Goa) = rib ~ Pra) 2.2%) Since p)-1 and q,-1 are coprime, we infer from (2.24) that qy-1 divides d—q,-23 Dut gna Squat and 0 < d < ¢ = dys that is, ld ~ qual < gyoi, and hence d= qp-2 = 0. Then (2.24) implies that b ~ p,.2 = 0. Therefore WB +b Pai + Bd Gn iB + Ge By the hypothesis > 1, so the resulting expression is the continued fraction representing the number @ and we have 8 = a, In other words b/d and a/c are Laosa1,-.-+da0. Bi consecutive convergents of a. a ‘Theorem 2.38 (Serret) Two numbers a, € R\ Qare equivalent if and only if there exist integers n,m > L such that aq = Bm. Equivalently, o and fare equivalent if and only if their continued fractions are Fasas,a2,..-1 and B= [bo:b1. bx...2.7 Equivalent numbers 39 dand ay = bys for some I< Zand all n > N. Proof First assume that for some n,m 2 1 we have ay = fp that is, = Lana... dgeas ahs — B= [oibI, Pret Bl and that a = By. Since @ is equivalent to a, and f is equivalent to fm (cf Example 2,36), we conclude that and f are equivalent Conversely, suppose that o and f are equivalent, that is, = ye, ad—be = +1 Changing, if necessary, the signs of all entries of y to their opposites, we may assume that ca-+d > 0. Let yp. be the (n~ I)th continued transformation of @; thus a = ¥_.1@. Then f = y7_.12,, and srs = (et 2 css) *) (Pat dderi ePear + dna) lel a Wehave crm tdnas =n 225) + dines = dua{eP=2 +d) =a Pana + da = (eB +d) = Take n large enough that both py2/do-2 and Pr-i/det a close to a. Then > 0," > O and, in addition, a, > 1. Finally, Corellary 2.20 allows us to manipulate the parity of nin such a way that 2 2 oP Gnd “ee arm . then from (2.25) we have c! > das daa < diet (by Corollary 2.16). Thus ll the conditions of Theorem 2.37 ae fulfilled, and we conclude that a = for some m. This completes our proof of the theorem, a Exmeetse 2.39 Let a = {a} 41, 43,43, ..] be an irrational real number. Show the following: (a) IO. 1 then a = [agjay,...] and Lar = [0; 4.41, (©) W=1/2.< ar < Othen = [15 1,02, 4, al and fa = [fas +2); ay me Note: this collapses to [=(a2 + 2):a4 + Tyas. Jifay = 1) (@) Wet < a < -1/2 then a = [-Iyar.a2,as,...}, where ay > 2, and Mer =[-2; 1a, — 2,4, ote: this collapses to [-2;a3 + 1,as,...Jitay = 2.)40 Continued fractions (©) Ie < =I then a = [apsay.ay,...J, where ay < -2 and Lfa = [=I 1, (a9 +2), 1,4) = 1a, (Note: this collapses as follows: to [=1;2, a ~ 1, @2,...] if ay = ~2 and ay > 250 [-15 1, (ap + 2),a2 + 1,a3,...]ifay = and ay < ~3; and to [-Iyaz +2.43,..Lif'ay = ~2 and ay = 1.) 2.8 Continued fr: ction of a quadratic irrational Let d be a positive integer. It can be seen that the set (x+yVd : x.y € Q} forms. 2 field, ln what follows, we assume that this field does not coincide with Q, in ‘other words, that d is not a perfect square. Morcover, without loss of generality wwe may assume thatthe number dis square-ree (that is, isnot divisible by a square > 1) Note that 1 and Vai are linearly independent over Q (otherwise Vd would bbe rational). This implies that each element of the field possesses a unique representation in the form x + y¥d with x.y © Q. Let this field be denoted by QV) and define the conjugate of a number a = x +yVd to be @ = x— yVid Exercise 2.40 Verify that TiA=T4+F and =a. ‘Now define the trace and the norm of a number « € Qa) by Tea) =a+F=2reQ, — Norm(a)= P-dteQ ‘Then a and its conjugate @ are the roots of the quadratic polynomial (x-a\(-2) Ix(@) x + Norm(a) ‘with rational coefficients; this characterises a as a quadratic irrational. Thus, a defining equation for the quadratic irrational « can be written in the form 2 ~ Daa + (x2 ~dy) = 0; taking x? — dy? = c/a and -2x = b/a, where a, b,c are coprime integers and ‘a> 0, we can represent the quadratic equation as, aa? +ba+e=0 ‘with coprime integers a, b,c, a > 0. Such a, b,c are determined by a uniquely. Finally, define the discriminant of a quadratic irrational a by the formula D(a) = 6 — 4ac = Aaya. Since a is a rea irrational number, we have D(a) > 0.2.8 Continued fraction of a quadratic irrational 4 ‘We shall call @ a reduced quadratic irrational if @ > I and -1 < @ <0 (equivalently, -1/3 > 1), Exancise 241 Ifa is a reduced quadratic irrational, show that ~1/@ is re duced as wel Theorem 242. Fora given positive integer D, there exist at most finitely many reduced elements of the field Q(Wa) whose discriminant is equal to D. Proof Leta bea reduced number having discriminant D(a) = D. Then bP ang 1 NP, aH where €= Lor “I Ife = —1 en we obtain o < 0, which is impossible Therefore €= 1 an, in accordance witha > Oand (220). b+ VD <2a<-b+ VD. (227) This means that b < 0; furthermore, the second inequality in (2.26) implies =b < YD. From these bounds on [| we conclude that there ar finitely many possibilities for the quantity b to satisfy the inequalities (2.26) In turn, the inequality (2.27) retains only finitely many possibilities for the quantity a > 0 as well. Finally, the quantity c € Z (if i exists) is subject to the relation 6? — ac = D, and hence it is determined uniquely by the three quantities D, a and b a Lemma 243. Ifa has discriminant D > 0 and is equivalent to a then has the same discriminant D. Proof Fora = x+yVd write Ap+B , AP BE= EB+F Then is equivalent to the quadratic equation (AB + BY + BAB + BEB + F) + {EB + FY = (aA? + DAE + cE?)6" + (2aAB + bAF + bBE + 2cEF YB + (aB? + DBE + F*) =0a Continued fractions whose discriminant is equal to (QaAB + bAF + BBE + 2cE I ~ 4(aA? + DAE + cE?\(aB? + bBF + cF*) ac = D(a), and whose coefficients are coprime. (If there is a common multiple of the co- cfficients then the inverse transformation Fa-B leads to the original quadratic equation for a, with coefficients a, b,c having the same common multiple.) o ‘Theorem 2.44 Let « be a real quadratic irrational number. Then (i) the number ay, m & 1, in the continued fraction has the same discriminant as a (ii) if a is a reduced number then a, is reduced for any n > | as well; and (ii) if eis not necessarily reduced then ay is reduced for all n sufciently large. Proof Chaim (i) follows from Lemma 2.43. Moreover, the defining procedure ofthe continued fraction for o implies ay > I forall > L (i) If ais reduced then a = a + 1/8 for an integer a > 1 and areal > 1: thisimplies -1/B = a—@ > 1, since a2 1 and # < 0, Therefore isa reduced number as wel ii) By Lemma 2.11, Genie * ed hence oy = 2.28) Gai Pr Therefore yw feted __ Yr T= Palen GesF= Pre eet B= Pra e (2.29) Eventually the fractions py-2/de-2 and p-1/qy-1 become close to, s0 that2.9 The Euler-Lagrange theorem B both the numerator and denominator of the last fraction are close to @ ~ @; in particular, they are of the same sign. Consequently, 2, < 0, Furthermore, 4 = Prailent Prot! Gn- oY Gree 2F — Proi nat) ale where we have used Corollary 2.15. Continuing (2.29) we find that 1 cor tte Do (tite Sapte) The expression 1 Tae Pala tends to 0 as n+ 00, and hence its absolute value is less than 1 for all n suliciently large. This implies %, + 1 > Oand demonstrates claim (i). AAs a somewhat tangential application, we may iterate (2.28) to derive Theorem 2.45 (Distance formula) Form > O we have cy aoa 230) Prot = Gi with our previous conventions that p: = 1, 4) = 0 and the empty product is interpreted as | Ieturms out that one may usefully think of [log [yes ~4o-il a8 measuring a weighted distance thatthe continued fraction hes trversed in moving from a 0, 2.9 The Euler-Lagrange theorem Let be areal irrational number. We say that its continued fraction fags, a2,...] is periodic if there exists an integer k such that dqy4 = dy for all n sufficiently large and purely periodic if ayy = dy for all n > 0; we call & the primitive period itis the smallest positive integer with the above property. "The following standard notation is used for periodic continued fractions:“4 Continued fractions where the vinculum (overbar) denotes the periodic repetition of the core- sponding part. A continued fraction is purely periodic iff it can be wittn in the form [ia] Lemma 246 Leta be a reduced quadratic irrational anda an integer Write a= a+ 1/f. Then B is reduced iffa 1 and -1/F = a~7 > a = [a] = 1, hence pis reduced Conversely, if < a then 6 < 0, and if a+ 1 < arthen < 1; thus 6 cannot be reduced if a # [al]. o Ruwark 2.47 We point out that the relation between a and f in Lemma 2.46 determines one of these numbers in terms of the other. Indeed, -Wp=at which implies that a = [-1/B]. Moreover, ® (and hence a itself) is uniquely determined by or, hence, by f. ‘We now come toa central result characterising quadratic rationals in terms ofthe periodicity of their continued fractions. Recall, in contrast, that the even tual periodicity ofits base-b expansion characterises the rationality ofthe num ber. This points tothe power of continued fraction representations over b-ary ‘Theorem 2.48 (Euler-Lagrange theorem) Let a be a real irrational number The continued fraction for a is periodic iff « is a quadratic irrational. In the latter case, a is reduced iff ts continued fraction is purely periodic. Proof First assume that a is a quadratic irrational. By Theorem 2.4 the cor- responding tails ay ate reduced for all n > no, while Theorem 2.42, together with Lemma 2.43, implies the finiteness of the reduced numbers that are equiv alent to a. Therefore, for some m > ng and k > 1 we have ay = anya. This immediately implics the periodicity of the continued fraction. Furthermore, assume that a itself is reduced; by part (ii) of Theorem 2.44 all the ary are re- duced as well. As we already know, ay = aqy4 for some n and k > 1. From Lemma 2.46 and Remark 2.47 we conclude that a1 is uniquely determined by ap and hence that ay-1 = ay.4-1. Applying this descent » times, we finally arrive at «= ag = ay; in other words, the continued fraction is purely periodic. Conversely, if continued fraction is purely periodic then it may be written oF (Tapa) = [aos ay. a,c2.10 Examples of non-periodie continued fractions 45 The relation a = ya implies that a is a root of a quadratic equation with integer coefficients, while by claim (iii) of Theorem 2.44 the number yfer = @ is reduced. In the case of a periodic continued fraction, we write 01 [a5 44.0 Tea Ara] = [0 A, erst] where the purely periodic continued fraction a, = {Grrr 7 Vis, by the above argument, a (reduced) quadratic irrational. Since a and a.) are equiva Jent, the number a is a quadratic irrational as wel. a Exercise 2.49 Show that, if a is reduced and a = [Zgidy—ai] then “WG = aa Exancise 2.50 ({125, Satz 3.9, p. 79]; see also Section 4.3 below) Let be a real number. Show that 8 isthe square root of a rational number > 1 iff there exist an integer by > O and a finite (possibly empty) palindromic lis of positive integers by,....by such that f= [by; By... 03, 269}, (More about palindromes in continued fractions will be revealed in Section 2.14.) tI Sketch of solution An equivalent way of saying that a list by,b2,....e is palindromic is that the matrix la b)_{br ‘\( 1) fmt e dtr oli of (1 of is symmetric (that is, b = ¢). Writing B= [Doi Dinss bi Bol = [oi is. u,B + Dol =e L = bp 4 Barbed SO" ba BiB bo)?” alB+ Dy) +b we obtain a quadratic equation for, OB? + = 0B ~ bolaby 4b +c whose linear term vanishes iff b = ¢ a 2.10 Examples of non-periodic continued fractions As we now know, continued fractions of quadratic irrationalities follow a sim- ple pattern; they are periodic. Are there other examples of continued fractions that satisfy a clear law and correspond to ‘meaningful’ real numbers? This is in general an unresolved problem, but we will uncover some partial solutions; ‘our immediate goal is to construct the continued fraction fore.46 Continued fractions Consider the funetion 1s, ot er xa” feels Gea rn= 1 al where we assume ¢ to be an arbitrary real number different from 0, ~1, (Go ensure thatthe series does not terminate). It is not hard to see that Fle.2)= fle + Lad + TAT Het 2.ay; this can be writen as fe+l» f(c+1,x) flex c+ 1+ —fle+ 2.» Flom FHMe+2.9 Tera er flerio Now, this looks like a continued fraction expansion, but the term 1 in the denominator isnot in the ‘ight’ place. We set x = z* and transform the above relation further fete) 1 eI We+2.2) et fer) This identity now has the right form to allow us to apply the general algebraic theory of continued fractions from Seation 2.3 By induction we get 2 flc+ 12) ofa 231) where ze flesn+2, Tus, we do indeed gota continued fraction if we can specialise c andzin such ‘a way that g.2 2 [and all partial quotients (c+ )/z are positive integers for cach n > 0. These conditions are met fore = 1/2 and : = 1/(2y), with y 2 Yan arbitrary integer. The resulting continued fraction LQ) = iy. Dred QFL23.09 is known as the Lambert continued fraction (which Was actually known (0 Euler)2.10 Examples of non-periodie continued fractions 47 ‘The quantity on the left-hand side of (2.31) admits a more familiar form. Indeed, for real w we have oe sinh w since the coefficient of (¥*)* in both power series is equal to a Tel @4n-1x4nl 3X5 Qt DX2x4 On) 1 Ga A similar argument leads to coshw = © & 1 ws 2 -4(37) Finally, letting w = 1 /y, we arrive at the following conclusion. Theorem 2.51 (Euler-Lambert) For every integer y = 1, we have tans (4) SF = 1089.39.95 Is y) Tele in particular for y = 2, 2 10,2,6,10,14,.041 (2.32) = 10:2,6,10,14, ‘The continued fraction fore itself is obtained in the next section and is de- rived from (2.32), Note that the numbers (e—1)/(e+1) and e are not equivalent, since the determinant of the matrix (1 5!) is 2 ‘The recursive method used in the proof of Theorem 2.51 can be applied for a more general clas of functions—the so-called hypergeometric functions However, just as in Theorem 2.51, one obtains continued fractions only for special cases of the parameters, We retum to ths inthe final chapter, where we study the Gauss continued fraction. Exanctse 2.52 (164); see also the chapter notes below) Find an expression for the valve ofthe continued fraction 10; y.2y,3y, 4y,..-1 eN in terms of Bessel fictions48 Continued fractions 2.11 The continued fraction for ¢ Following Buler we now prove the following statement. ‘Theorem 2.53 Ifthe continued fraction for e is given by = [aos ays a2, 26) = [1,211.4 161,611.80.) then ag = 2, daya2 = ing = Land dayat = 2m for m > I. Proof Let ry/%, denote the nth convergent of the number T 0:2, 6,10, 14... J,and hence By Theorem 2.51, we have o~ = [2;6,10,14,...] by Exercise 2.39. In addition, eo ott (2.33) a Let € = (2) 1.2.1, 1.4.1, 1.6.1... } and let py/dy denote the nth convergent of € Let us show that, for > 0, the following relation holds: Pani =TatSn — Ganet = Ta ~ Se (2.34) ‘These relations are easily checked form = 0, 1. I'm > 2 then r= (ANH Dry breed and — Sy = AMA D5pa4 Seeds in accordance with Lemma 2.9. Now we multiply the recurrence relations for py (as well as for q,) as follows, Pit = Pitt Pros by 1. Pana = Pana + Pine bY = Pat = 2nprg-a + Pixs by 2. Pan = Pane + Pa by 1, Pant = Pan Pet bY ‘and sum them up. This gives us Part = A+ Dpann + Paws — and — Gays = (4+ 2gana + dans ‘The equalities (2.34) are then deduced by induction on n. From (2.34) we con- clude that Prost _ Tat Sn _ Tale #1 Gut Tan FolSn= 1 (2.35)2.12 The order of approximation of e by rationals 49 Letting m — 09 in (2.35) we finally obtain atl oa 236) and comparing (2.33) and (2.36) produces the required equality =e. a As a corollary of Theorem 2.53 we deduce that ¢ is neither rational nor a quadratic irrational. Exaneise 2.54 ((125, Bd. 1, Section 34]) Use the argument above to derive the Perron continued fraction, eM" = (Lyn 11 13m 11S 111 In, for integers n > 1, Deduce that e is not an nth root. Exanciss 2.55 ({143, p. 132) Let a = L/y for a positive integer n > 1 Show that Geae 7 [53n.5,7n,9, Un 13,..] Conclude that e4¥ is irrational for all imtegers n > 1 Exercise 2.57 ([88]) Show that for integers n > O we have eNO) = En, 12m +6, 5m + 2, 1, 1, Tn + 3, 36n + 18, In +5, 1, 1, 13 + 6,60n +30, 19n+8,1,1,...] = [(1, 3Qn + 1) +n, 62k-+ Qn +1), Bk +2Qn4 1) +0, V 2.12 The order of approximation of ¢ by rationals The aim ofthis section i to prove the following more refined result of C. Davis 54; the use of the continued fraction for e from Theorem 2.53 is a erucial ingredient ofthe proof ‘Theorem 2.58 (Davis' theorem) For any € > 0, there exist inginitely many rationals p/q that satisfy the inequality 2) (1, <\eslosa bE <(5+ hone30 Continued fractions conversely, the inequality loglogg @ logg ‘has only finitely many solutions. ‘Theorem 2.58 means that the function gig) = Seed is the best approximation order of e by rationals (recall the definition from Section 1.4). Lemma 2.59 For the continued fraction (asa, 225 1,2,1,1,4,1,1,6,1,1,8,.-] {for the number e, we have the asymptotic estimate nlogn 3 Proof Set k = |(n= 1)/3]. We have og(ayaa ay) = +O(n) ano. logtayay---ay) = Yogt2h = klog? + 9° tog ‘To estimate the later sum we use the inequalities, [asses fetes [gia (= 1...85 log xdx, then klogk —k foeste< S081 < f esdr= es Doge + D =k 0. (On noting that k= n/3 + O(1) we deduce the required asymptotics, a Proof of Theorem 2.58 Let us start with the second part of the theorem. If rational number pq is not a convergent of e then, by Lemma 2.33, we have2.12 The order of approximation of e by rationals st and, in particular, the inequality (2.38) al «(4 gues al \2 Tlogg does not have solutions in such rationals. In the case p/ n= 3m-—3or n= 3m—1 we have dys = Is hence olde with either kk by Lemma 2.30, which contradicts (2.38) for q > 2. Finally, if p/q = Pa/dn with n = 3m—2 then ays = diy-1 = 2m, implying that 1 —1 2.39) GDF ee Using the estimates in (2.18), log F, = O(n) and Lemma 2,59, we deduce that log qin-2 = 108 de = log(ay »--ay) + Om) nlogn + O(n) = mlogm + O(n). Hence os a Tog log dina and substitution of this asymptotic estimate into (2.39) gives 2. eles -|> >> (1) feof ]> gapELA + op This demonstrates the second part ofthe theorem |As for the first part, note that choosing p/q = pin-2/qin-2 in Lemma 2.30 (ie estimate in (2.17) from above) and using Lemma 2.9 gives us loglogg 2g?ogq fee (1+ 011) ql This completes the proof of the theorem. a Exuncist 2.60 (Estimation of e'!*) Using Perron’s continued fraction for e" from Exercise 2.54, extend Theorem 2.58 to rational approximations to with ye,2 Continued fractions 2.13 Bounded partial quotients Let er be areal irational number with continued fraction [aya 42, .--IeIP K(@) = sp a is finite, we say that « has bounded partial quotients (or is of finite type (94] or is badly approximable |149)). Otherwise, if K(a) = e0 then we say that a has unbounded partial quotients. We define 8 to be the set of all real numbers with ‘bounded partial quotients. As we will see in Remark 3.2 itis an uncountable set of Lebesgue measure zero, Not many ‘naturally occurring’ irrational numbers are known to be in 8 ‘other than the quadratic irrationals. We will see some additional examples in Chapter 6. The status of the algebraic numbers of degree > 2 is currently un- resolved, It is conjectured that all these numbers have unbounded partial quo- tients Inthis section we prove some properties of the numbers with bounded partial quotients. In particular, we show that 8 is closed under the linear fractional transformations, that is, the maps a (aa + b)/(ca + d) for integers a,b. ¢,d with ad ~ be #0. Recall that if a = [a9;41,2,...] then ay denotes the nth complete quotient [ani usa} We detine the norm lla = min(a — [ar fa o) = min({a}, 1 - (a), that is, the distance from «to the nearest integer. Lemma 2.61 Suppose that qligal| > 1/r for some r Vand all q > 1. Then Ka) 1. Since mr from Theorem 2.25, we see that dnlPq ~ dual < U/aye1. Now, clearly lal = dn In = qual so that 1 5S aullqctl $ dele — duel < hence dys O and the estimate K(a) < r follows, 2 Lemma 2.62 Suppose that K(w) 1. Then for all q > 1 we have qigall > 1/(¢ +2.2.13 Bounded partial quotients: 3 Proof We will proceed by contradiction, Suppose there exists a q > 1 such that glgar| < 1/(¢ + 2). Let p be an integer such that l[gal| = [ga ~ pl Then algar~ pl < 1/(r +2), implying that bl q from which it follows by Lemma 2.33 that p/qis a convergent to a, say p/q = elo Thus, p = apy and q = agg for Some integer a > 1 Now, from Lemma 2.24 we have 1 |ana— Pal = —————+ Bonide de hence 1 > alaeil = alae ~ pl = aga | Fay > Hleell = dae pl = agulagger— apy Sadao -p)> ——! 1 © deldne’~ Pol = nt FGa-1/Ga (Ges + +1 ‘Thus ay.; > r and so K(q) > r. a We say aris of ype < rif qlgal| > L/r for all integers q > 1 Lemma 2.63 Leta and b be integers with a > 1, [bl > 1. Ifa is of ype 1. Now there exists an integer p such that béel-beo—a Thus, a 1 pena Owe have K($.0) < abl K(@) +2)4 Continued fractions Lemma 2.65 Leta and b be integers with al > 1, b > 1. Ifa is of type 1 ‘There exists an integer p such that g(r + a/b) = g(a + a/b) ~ pl thus, a 1 alle ~ i) 7 PS eM and so, multiplying by b*, we get gblgba + ga ~ pb| < 1/M. Hence we do indeed have qbjlgbal) < 1/M. 0 Carley 2.66. For iepers ab hb > Owe hve Ko 8) 1 the following hold: 1. x; has a simple continued fraction ofthe form (15), where by isa finite Tis of Is and 2s that starts with aI and ends with 2; 2. the continued fraction for x; isa prefi ofthat for tints 3. Pisa rational number 4. 32, isan excellent approximation to y2~so good, infact, thatthe con- tinued fractions of y2,, and y? coincide at least as far as the penultimate partial quotient in the expansion for y? if dhe expansions first differ at ppattial quotient a; then either the (j + 1th or (j + 2)th partial quotient of Yh is very large Our construction uses strings (or lists) of partial quotients. If b is a string then by “6 we mean the reverse (or transpose) of bs for example, (1,2, G,2,1). If = "b then we say that b isa palindrome (see Exercise 2.50). We define [bo be the number of terms in the string b. By (6}* we mean the string in which b is repeated & times (cf. Exercise 2.10), so that [(6}!| = Albl; by {b)" we denote the string 6, 6,6, ... which is infinite if [| > 0. Reprising the definitions in Section 2.9, if an infinite string e can be written inthe form (b}* for some finite nonempty string & then we say that ¢ is purely periodic. The period of a purely periodic string ¢ is the length of the shortest string b such that ¢ = (6) Here are more details. Set by = (1,2). Now, for i = 1,2,3,..., define the following: mi = Wi. x: = (15 Bil, ys = (15(0)"] = (156;,b:,Bi,..-] and ry =56 Continued fractions yf = el fs where ged, f) = 1. Choose an integer n; > 1 such that if gy = (1; (64) then 1? > 4(100' +2)/2. Finally, let busy = a! 2), (2%, 2) ‘We then have the following theorem. ‘Theorem 2.69 (Bounded quotients with unbounded squares) Let x). be de: fined as in the preceding construction, and define xelimx; and r= lim 7 Then the partial quotients of x are all 1s and 2s, the partial quotients of r are unbounded and r= x. To prove the correctness of the construction of the theorem, we need some Jemmas, Let r = [do:ai,..-,4e] # 0 be the continued fraction of a rational number, with ay 2 2 if k= 1. Then define length(r) = k-+ 1 Lemma 2.70 Suppose that r isa rational number and that [ay ay,.... a] is its continued fraction, where the last partial quotient ax > 2 (if k # 0), Let algx be the kth convergent to r. Let tbe an integer > 2, and suppose that sis «areal number, with O s and k is odd, we have s = (by; by by, where bj = a; for0 SiS kand bys: ®t; Gi) fr < sand k is odd, or ifr > s and k is even, we have s = [bps by, ba,..- where bj = aj for 0 < is k= 1, by = ay — 1, buss = Land bea 2h Furthermore, i sis rational then length(r) < lengths). Proof We will teat only the case r < s: the case r > sis similar and is left to the reader, i) Assume k is even. Then we have that the first & +2 terms of the continued. fraction expansion for s are dp,ay,...,a4,b iff Laos... 1S 8 $ laosay,...,a%,0) Hence the first & +2 terms of the continued fraction for s are of the form ap, dy, de,b, with b > fi fay; a,,---.4] $8 $ [aosay,-.-,au,t,2.14 In the footsteps of Maillet 7 it and iff a alae a where we have used Theorem 2.14, Hence, if 1 one it follows thatthe continued fraction for s begins with the partial quotients 4,2, With Bt Gi) Now assume that k is odd, Then the fist & +3 terms ofthe continued fraction expansion for sare ap,a},...,a4-1,04 1,1, biff O 6 iff fap; a3,--sAgetsde 11) $$$ [api i,..-sgeas de 1 1th it po Phe ge C+D Pu gS Dae ae and iff oes ere CDP pes 1 Damar e+ Dae 4)" where again we have used Theorem 2.14. Hence, if 1 O a Lemma 2.71 Leta and be two positive real numbers such that the first k+ 1 terms of their continued fractions coincide and are equal t0 a. 4,,....d,. Let Pal qu = [ays ... 04) Then fa ~ fl < Via Proof Assume that k is odd. (The case where k is even is similar.) Then La3a1,-.-544, 1] $ af $ Lao, ay,..., a4)58 Continued fractions Hence we have Pat Piet pe PARP caps Ph Gin % Ie follows that Pi Pie leas Oe Where we have used Theorem 2.14, 0 ‘We now prove the correctness of the construction Proof of Theorem 2.69 First, we claim the following are true for all> 1 () by is a string of Is and 2s which ends in a 2; i) bisa palindrome if the last symbol is removed; (ii) 6, is a prefix of Bier; Gv) me < misss () Le any <2 (vi) ris a rational number. Parts ()-(¥) follow easily from the definition of ys for pat (vi), combine (i) with Exercise 2.50. Next, we argue that all they, are distinet, that i, if # j then y, # yy. For this, it sufices to show that (B,)" # (b;)”. We do this by showing that the period p of (5) satisies m1 < pi < my for i > 2. The upper bound is clear, since m, = [i For the lower bound, observe that by construction starts with a 1 but contains a block of my consecutive 2s. It follows that pis atleast my +1 Hence py < pz < py < +++, and so the y; are all distinc. Itis clear that slim, and y= limy, exist. By construction, the continued fractions for y; and x41 agree on partial ‘quotients a through day, and so by Lemma 2.71 we have 1 mails < 1007) Ps It follows that x = y, and so if r = lim;,.. 7; then r = 2°. Final it remains to see whether r has unbounded paril quotients. We ap- Ply Lemma 270 with r~ rand s~ ry. Since they ar all distinct, we know that r# sand, by result (v) above, risa rational number. By construction, 1 1 BS Mose2.14 In the footsteps of Maillet 39 using result (iv) above, we get Wr sl= i rinll = Lo? —Yhal 1 = 04 ObIH, TOF wil <4lyi yall By Lemma 2.70, it now follows that the continued fractions for rand riy1 coincide up to the penultimate partial quotient of r;, and the continued frac- tion for ris1 contains partial quotient b of size > 100/ + 1. Furthermore, Jength(yi.2) > length(ri.1) > length(r), so this new large partial quotient 5 introduced in the continued fraction for 7) ean be reduced by at most I in the continued fraction for ri42 (and this reduction can occur only if b is the last partial quotient in the continued fraction r,.). The partial quotient 6 cannot be reduced further in r:.s,ri44,--- It follows that the continued fraction for ri.2 contains a partial quotient > 100 and that this partial quotient also appears in r, Hence r has unbounded partial quotients. a Rustarx 2.72 It seems worthwhile pointing out which parts of the above con: struction are crucial and which parts are somewhat arbitrary. The definition of by is more or less arbitrary as long as it begins with 1, ends with 2 and is 1 palindrome when the last term is deleted. Similarly, the choice of the term 100/ in the denominator is more or less arbitrary; any function tending to in- finity sufficiently quickly could be substituted. Example 2.73 Let us look at an example of the construction. We have by = 1.2 xy = [pba] = [151,2] = 5/3: ye Usb, bi. ba, n=3=3/1= 13h; m=3; by = (by, bby, 2,2, "1, "By,"Bs,2) = (1,2, 1,2, 1,2,2,2,2, 1,2, 1,2, 1,2) xy = 52472/30 297; y2 = YOOSSTIIIB: ry = 6653/2218 = [25 1, 2217]; 550 Continued fractions by = (Ba, ba, (2)! "Ba, "ba, 2) 1x3 = 400 183 051 571 243 862.647 721/2310 631 559 280.479 906 375 689; ys = YT0STL919 435 937 731 341 509]23 494 170 646 161 12017202 rs = [25 1.2216, 1, 365316604 851, 1, 1,2,3,5,4,1,3,3,2, 1, 136, 1,2, 1,2I; ms = 76; ny = 2. Continuing in this fashion, we construct a number x = lim. given by x= (05 1,2,1,2,1,2,2,2,2, 1,2, 1,2,1,2,1,2,1,2, 1,2,2,2,21,2,1,2,1,2,2,2,2,2,2,2,2,2,.0 and for which ¥ = [2;1,2216, 1, 365316604851, 1, 1,2,3,5.4,1,3,3,2, 1,136, 12111 where N = 110 153056 861 410 353 112 260 022 760 667 309 348 898 319 235 606 578. 1. By choosing a more rapidly growing function one can easily ensure that the resulting value of r is transcendental (not a root of any algebraic equation with rational coeflicients). 2, By choosing (o introduce either a block of 2s or 3s instead of a block of 2s in the middle at each step, the construction can be easily modified to give uncountably many numbers x with the desired properties. 3. By replacing y? with y! in the construction given at the start of this sec- tion, and suitably modifying the inequality test, we can generate exam- ples of numbers x with bounded partial quotients for which 2 has un- bounded partial quotients for any even positive integer d. In fact, we can generate numbers x with bounded partial quotients for which all nonzero even powers of x have unbounded partial quotients, To do this, we ‘modify the construction as follows. Set by = (1,2). Now, for i = 1,2,3,..., define the following: m) = [bl 2 = (sbi) and yy = [15 (i)"] = [15 bi, bi, bi,-..]- For j = 1,2, i define Fi) = Guy fiy = 97. Choose an integer n, > I such that if gf = [1548] then Ae > max j2™ (100! + 2)f3, Finally, let Bios = (Ubi), (2), (8 +2).Notes 6 ‘Theorem 2.74 (Unbounded even powers) If; is defined as given in the pre- ceding construction, and x = lim.yx %, then x has partial quotients consisting of Is and 2s, but every nonzero even power of x has unbounded partial quo- tients Proof The proof is largely the same as the proof of Theorem 2.69 and is left to the reader. 0 Even more possible. By altering the construction of the previous theo- rem to utilise a list of all rational functions with integer coefficients, not just powers, we can prove the existence of a real number x with bounded partial quotients such that every rational function of x? has unbounded partial quo- tients. Notes ‘The 2 x 2 matrix approach to continued fractions was originally developed by Hurwitz before 1916 (matrices themselves were then quite new); see [76]. It was later rediscovered by Frame [63] and Kolden [84], independently. ‘Smith's representation of the continuants given in Exercise 2.23 allowed bim to produce an elegant and constructive proof of Fermat's ‘two squares’ theorem: every prime p congruent to 1 (mod 4) can be expressed as a sum of ‘vo squares, We refer the reader to [46] for the details of this derivation and some historical remarks. ‘The correspondence between continued fractions and matrices discussed in Section 2.7 in effect identifies all nonsingular 2 x 2 matrices _(a 2 le (not necessarily from = SLo(Z)) with linear fractional transformations (2.21), because ky for an arbitrary scalar k # 0 acts on a real number a independently ok ‘Then any sequence an bg ™ (: ) of nonsingular 2x 2 matics sch that ayy and by/dy havea common init {as n-» 20 yields a continued faction expansion. For example if Cs )-TCa TS ont)2 Continued fractions then dydy ~ bycy = (=1)"*124%D shows that the z-formal power series a/¢y and by /d, coincide inthe limit. Here ag(z) = da(z) and By(z) = ¢q(~2) and we need to confirm only that both a,(2) and b,(2) times e-*?nl/(n-+ 1)! converge to 1 as n> c0; so here the common limit is e eet aye ee ey wwe obtain the main result of Section 2.11 12 U612114,11,6,118,...1= Usb 201, Another recent self-contained proof by Cohn of the continued fraction for ¢ is to be found in [47] A lovely classical result (125), re-observed by Schroeppel in 1972, is that for real a and 6 # 0 any arithmetic simple continued fraction is expressible as a ratio of modified Bessel functions of the second kind: Tays(21B) ar bat2hat3h.d= po In particular, (2) 0.2.3...1= 7G In our derivation of Davis’ Theorem 2.58 we followed Tasoev [159], who proves a variant of the result for more general pseudo- periodic continued frac tions, In a related way, [28, Theorem 11.1] yields integers _ Qn-W! yy Gna bt me am yy DGS Bie (2.40) such that, Pal _ loglog dn — Pal = OBB ny 5 o(1 dn 2g 10g an XI). ‘The majority of other mathematical constants, including x, have no known simple patterns in their continued fraction expansions. However, suitable gen- cralizations of continued fractions, such as the irregular continued fractions we consider in Chapter 9, do sometimes exhibit simple patterns; one example of such a continued fraction is given in (9.16) For more results regarding bounded partial quotients, one may consult the survey [151]. Exercise 2.39 was based on [83, Exercise 4.5.3.10]. A different proof of Theorem 2.68 is given in [92]. Some results in Section 2.14 can be found in [37]Notes 8 ‘A weaker form of Lemma 2.62, with a worse constant, was given by Hardy and Wright [72, Theorem 188], while Lemmas 2.61 and 2.62 above were es- sentially proved by W. Schmidt [149, pp. 22-23]. One can also deduce these results by filling in the details in [10, p. 47] ‘Theorems similar to Lemmas 2,63 and 2.65 were given by Cusick and Men- ds France [49]. Instead of considering sup, algal, they studied Jim sup gla, which is somewhat more natural; se also Perron [124]. Chowla [45] proved in 1931 that K(aa/b) < 2ab(K(a) + 1)°, a bound much weaker, however, than our Corollary 2.64. ‘The Maillet continued fraction construction continues to be of interest; see for example [1] Finally the real mapping G: x +> 1/x (mod 1) for x # 0 with GO) often called the Gauss map. It has fascinating dynamics, some of which may be studied in [48].’ Figure 2.1 shows the map drawn on the torus and hints at subtlety. OF course, drawn in the real plane, it has singularities at 1/n for n= 1,2, Figure 2.1 The Gauss map drawn on the toms, 1 See aio nts: / wr. coom.afa.c8/organica/papers/corleas/