Statistical Independence and Normal Numbers: An Aftermath to
Transcription
Statistical Independence and Normal Numbers: An Aftermath to
Statistical Independence and Normal Numbers: An Aftermath to Mark Kac's Carus Monograph Author(s): Gerald S. Goodman Source: The American Mathematical Monthly, Vol. 106, No. 2 (Feb., 1999), pp. 112-126 Published by: Mathematical Association of America Stable URL: http://www.jstor.org/stable/2589048 . Accessed: 12/08/2013 16:04 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp . JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected]. . Mathematical Association of America is collaborating with JSTOR to digitize, preserve and extend access to The American Mathematical Monthly. http://www.jstor.org This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions StatisticalIndependenceand Normal Numbers:An Aftermath to Mark Kac's Carus Monograph Gerald S. Goodman 1. INTRODUCTION. In 1958, Mark Kac deliveredthe prestigiousPhilips Lectures at HaverfordCollege, whichformedthe basis forhis Carus Monographon StatisticalIndependence [2]. As a student,I had the privilegeof attendingthose lectures.By foragingin theiraftermath, meaningliterally,"that whichgrowsafter the harvest,"I have been led to the followingideas. They may be regarded as comprisinga long footnote,or possiblya missingchapter,of that Monograph, composed in memoryof its authorby one of his erstwhile"guinea pigs." In [2], Kac presentsan arithmeticalmodel for coin-tossing,due to Borel [1], whichculminatesin an analyticalproofof the StrongLaw of Large Numbersfor Bernoullitrials.The idea is to identify the digitsoccurringin the binaryexpansion of a point w E Q = [0, 1), Ea 2' ' (1.1) j=l withthe outcomesof a faircoin-tossing experiment.The j-th digit,aj(w), equals 1 if tails occurson the j-th toss,and equals 0 if the outcomeis heads. Then Sn( w) = a,( w) + a2( w) + -- +an( w) (1.2) is the total numberof tails in the firstn tosses,and Borel's StrongLaw of Large Numbersassertsthat,when fQis endowedwithLebesgue measure, 1 Sn(w&) - -> as n -> oo. n 2 a.e., (1.3) Kac's proofof Borel's theoremmakes use of the Rademacherfunctions.These EE N and w E fQ,as are defined, foreach j rj( w) = 1- 2aj( w), (1.4) and theyrepresentthe net gain of a gamblerwho bets one florinon the outcome heads at the j-th toss.The orthonormality of the Rademacherfunctionson the unit intervalfQ,and, indeed, of theirfiniteproducts,is a consequence of the factthat the binarydigitsare statistically that Kac independent.It is this orthonormality exploitsto give a simpleproofof the Law of Large Numbers. Kac goes on to interpretthisresultin termsof the existenceof simplynormal numbers.For such a number,the relativefrequencyof the l's, among the firstn digitsin the binaryexpansion of its fractionalpart, tends to 1/2 as n -> oo.It followsfromBorel's theoremthat almostall real numbershave thisproperty.Kac pointsout thata similarresultholds forany integralbase b > 1, sketchesa proof, and concludes,withBorel,thatalmostall numbersare simplynormalto everysuch base. 112 STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions [February Now, it is known[8] thatanynumberthatis simplynormalto all bases thatare positivepowers of an integralbase b has the followingproperty:Everystringof finitely manyb-adicdigitsoccursin the base b expansionof its fractionalpartwith asymptoticrelativefrequencyl/bm, where m is the lengthof the string.Partial overlaps of the stringwith itselfare counted as distinctoccurrences.Numbers exhibiting thispropertyare termednormalto base b, and the factthatalmostevery real numberhas thispropertyis knownas TheNormalNumberTheoremforb-adic digits. Since Kac's use of orthonormality leads to such a simple proof of the case m = 1, b = 2, it is naturalto ask whetherhis method can be adapted to give a directproofof the NormalNumberTheorem,at least forstringsof binarydigits.It turnsout thatthiscan be done by replacingRademacherfunctionsby theirfinite products,which themselvesform an orthonormalsystemknown as the Walsh fuinctions. Kac presentsthemin an exercise,withoutany mentionof theirpossible connectionwithnormalnumbers[2, p. 11]. That discoverywas made by Mendes-France [6], who showed how base 2 normality can be characterizedin termsof the Walsh functions.He thenused this characterizationto prove the Normal Number Theorem for binarydigits. His approach, which, as he showed, generalizes to arbitraryintegralbases b > 1, makes use of Haar measure,group characters,a generalizedWeyl Criterionfor asymptotically equidistributed sequences,and Birkhoff's ErgodicTheorem,applied to the dyadicmap. We shall show, instead,how the orthonormality of the Walsh functionsleads directlyto a simple proof,a la Kac, of the Normal NumberTheorem forbinary digits.The only additional idea required is the well-knownobservation,due to Wall [10] and used by Mendes-France,that a numberis normalto base 2 if and only if the iterates,under the dyadic map, of its fractionalpart are uniformly distributedin the unit interval.Using this idea, our proofproceeds in the same spiritas one suggestedby Kac-again, in the formof an exercise-for proving-the classical theoremof Weyl [11] on the equidistributionof the fractionalparts of multiplesof irrationalnumbers[2, p. 41]. We can extendthisapproach to integralbases b > 1 by defining, withMendesFrance [6], the b-adicRademacherfunctionsas ) forall w E Q, I R b E r1( )=exp2 of the point w, wherethe bj are the b-adic coefficients v0 bj() bJ1 and definingthe b-adic Walsh functionsas theirpower products.When b = 2, of rj agreeswiththe bE {0, 1} and Euler's formulaensuresthatthe new definition old one. These b-adic Rademacher functionshave mean zero, and all that is needed to establishtheirorthogonality (over the complexfield) and furthermultiplicativity propertiesis the formula k Hrexp(itLjbj( __1 ())dw = k fH j=10 exp(i,tjbj( w)) dw, values of k EE N. But forsuitablevalues of the real parameters,Ujand arbitrary the validityof such a formulais guaranteedby the statisticalindependenceof the 1999] STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions 113 as Kac well knew [3], and its proofis virtuallythe same as one foundin the openingpages of [2]. In Section 8, we givethe details,and show how the reasoningused to establish the base 2 normalityof almost everyreal numbergoes throughin the case of generalintegralbases b. Once done, it remainsonlyto collectthe exceptionalsets in orderto arriveat a new proofof The NormalNumberTheorem,in itsfullforce: almosteveryreal numberis normalto everybase. We then go on to examine more carefullythe connectionbetween the multiplicativityof the b-adic Rademacher functions,understoodas the vanishingof theirmixedmoments,and the statisticalindependenceof the b-adic coefficients. Using a device of Renyi's[9] we are able to draw a remarkableconclusion-the propertycan be two notions are entirelyequivalent! Since the multiplicativity establishedeasilyby elementaryanalysis,thisyieldsa new proofof the independence of the b-adic coefficients. The same ideas can be applied to b-adic Walsh functions.While theythemindependent,we findthatthe ones whose indicesforma selves are not statistically geometricalprogressionmade up of fixed multiplesof powers of b are. The statisticalindependence of the b-adic Rademacher functions,and thus of the followsas a special case. b-adic coefficients, bj's, 2. THE RADEMACHER FUNCTIONS. The functionsrj are defined by (1.4). Since the binarycoefficientsai satisfy ~~~1 1 jaj(w) do) = -for j EJ, frj( ) dw = 0 forj E N. it followsthat The statisticalindependenceof the a1 impliesthatthe rj are also independent, Formula and, since theyeach have mean zero, theysatisfythe Multiplicativity f1rjl( (l) &w)rj2( ... rik( @) dw = 0, k EEN, (2.1) wheneverthe subscriptsare distinct.Since rf2= 1 foreveryj, it followsfrom(2.1) systemon fQ,thatis, thatthe Rademacherfunctionsforman orthonormal f1rj( w)rk( w) dw =8jk forj, k e . (2.2) 3. KAC'S PROOF [2], [3]. In viewof (1.4), the limitingrelation(1.3) is equivalent to the assertionthat Rn(o) -Oa.e., as n oo, (3.1) where Rn( w) = rl( w) + r2( w) + ?+rn(w) (3.2) is the difference betweenthe numberof heads and the numberof tails in the first n tosses. 114 STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions [February A directcomputationusing(2.1) and (2.2) showsthat = n + 3n(n - 1). f[Rn()]d,=n+2!2!d(2) Consequently, fl[ Rn(W) ] d E < (3.3) 00 and it followsfromBeppo Levi's Theorem that 4-'[Rn~w) n=l ]4d w < oo a.e. This impliesthat [R~(w) ->0 a.e., as n ]4 -> oo, whichis clearlyequivalentto (3.1). 4. A VARIANT OF KAC'S PROOF. The precedingproofuses (2.1) up to fourfold of products,whichis evidentlya strongerpropertythan the mere orthonormality functions expressed by (2.2). However, a proofof (3.1) thatuses the Rademacher of the rj (along with the uniformboundedness of their only the orthonormality absolute values) can be based upon an argumentemployedby H. Weyl [11] in a similarcontext. (2.2) impliesthat,foreach n E , The orthonormality 1 [Rn( W)]2 dw = n. It followsthat fl[Rn2(W) ]2 E du < oo, O n=1 and, therefore,reasoningas before, 2 n > 0 a.e., asn-> (4.1) c. Now, to each value of n, not a perfectsquare, thereis a unique positiveinteger mn such that m2 < n < (ma + 1)2. Clearly, m ' and the factthat Ir, = 1 forall j, oo as n -> oo. In view of (3.2) + rM2 + Irnl< IRm21+ 2mn. +11+ IRnI < IRm21 Dividingby m2 and using(4.1) with n2 replaced by m2 yields IRn( mn w,)I 2 -Oa.e.,asn ->oo. Since m2 < n, (3.1) follows. 5. SHIFT DYNAMICS. Define the binaryshift(or dyadicmap) T on fQ by the formula[2, p. 19 and p. 93] (5.1) T( w) = 2 w (mod 1). 1999] STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions 115 Then, when a is expressedin termsof its binaryexpansion(1.1), T(w) takes the form T(wa) = Z 2' L _a _ j=l1j=l Makingthe conventionthatbinaryrationalshave finiteexpansions,the uniqueness of the binarycoefficients gives aj1+( w) = aj o T( w) forj E N. Iterationyields aj+k(w) = aj o Tk( W) forjE N, k E No, (5.2) and, in particular, al+k(wt) =a, o Tk() fork E N0. (5.3) Now, a, is the indicatorfunctionof the interval[2, 1). In view of (5.3) and the definition(1.2) of Sn, +a, o Tn-1 S,j( ) = a1( ) + a1 o T( )) + Thus, Sn( W) countsthe numberof timesthatthe orbitof w under T, thatis, ,T(Z).. Tk(@).. is foundin [2,1) duringthe firstn steps,startingfromstep 0. The StrongLaw of Large Numberscan thereforebe interpretedas sayingthat therelativetimethatthe orbitof w occupies[1/2,1) tendsto 1/2 as n -> oo,for almost all startingpoints EE fQ,in accordancewith[2, p. 93]. of the Law of Large Numbersleads to a formulaThis dynamicalinterpretation tion of the NormalNumberTheorem in an analogous way. Let al,, 2 *.... a* (5.4) n denote a stringof binarydigitsof lengthm. Set m aj I= E j=l ~2-m* and let II [= ) *(5.5) 2m n S2m Then I, is the l-th binaryintervalof orderm, and it consistsof those pointsin fQ whose binaryexpansionstartsout with (5.4). With this enumeration,the lexicographicorderof the strings(5.4) is expressedby the naturalorderof the intervals (5.5), in termsof increasingvalues of 1. If we now denote by XI, the indicatorfunctionof the intervalII, 1 - n-i E XI!o T(k) nk= is the average numberof timesthe string(5.4) occurs among the firstn + m - 1 of w,whereoverlappingscountas multipleoccurrences,and an binarycoefficients occurrenceis marked at the momentwhen the stringstartsto appear. Any real numberwhose fractionalpart is a will thenbe normalif,foreach m E , in-1 -E k=O 1 XI,Tk(w)'> 212 , as n ,oo This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions (5.6) where XII is the indicatorfunctionof a genericbinaryintervalI, of order m, and be a conse1/2m is its length.The Normal NumberTheorem will, accordingly, quence of the assertionthat(5.6) holds a.e. on fQforeach such binaryintervalI,. 6. THE WALSH FUNCTIONS. Kac introducesthe Walsh functions,as follows. EE R, let For each j 2j = 2i1 + 212 (6.1) .+2is + < j] are in R, be the unique expansionof the integer2j in where il < j2 <... base 2. Then, foreach j EE N, definethe Walshfunctionsby means of the formula &j(w) = rjl( w)rj2( w) ... forw E rj(w) fQ, (6.2) and set w0- 1. The Walsh functionsthus constitutean enumerationof all the finiteproductsof Rademacherfunctions. The Walsh functions,like the Rademacher functions,have a probabilistic in termsof coin tossing.For each j EE R, wj representsthe net gain interpretation of the gamblerwho bets one florinon the outcome thatthe total numberof tails , ]s is even. occurringon the tosses 11, , Because of (2.1), the Walsh functionsare orthogonalon fQ,while the identity lw21 -1 implies that they are orthonormal.For each mEE N the 2' functions W0,W1,... I,W21 are constanton binaryintervalsof order m. The range of any such functionthus correspondsto a vector whose l-th componentis the value of the functionsmakes takenby the functionon the intervalI,. The orthogonality these vectorsorthogonaland thereforelinearlyindependent.It followsthat the functionsthemselvesare linearly independent. Consequently,any real-valued functionf thatis constanton binaryintervalsof orderm can be writtenas a linear combinationof the first2' Walsh functions: 2n f(f) - 1 j=0 ((t) , @ Ajwj E Q (6.3) of f: coefficients wherethe weightsAi are the Fourier-Walsh Aj = |f0)f( ) ij( d@, j = O,..., 2m - 1. In particular,the functionXI, can be writtenin such a form.In thiscase, Ao = 1 1 XI,((w)wO(w)dw = 2m (6.4) while Aj = J1lx11(w)w1(w)dw = 1 ?-m ]= 1, ,2' - 1, where the value of the + signis such as to make sgn[Ajw] = + 1 on I,. 7. PROOF OF THE NORMAL NUMBER THEOREM FOR BINARY DIGITS. In order to prove that (5.6) holds a.e. for any binaryintervalof the form(5.5) with m ERN arbitrary, observethat(1.4) and (5.2) implythat rj+k( w) = rj o Tk( w) 1999] forallw E Q, j E , k E No, STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions (7.1) 117 so that,in view of (6.1) and (6.2), W2kj(w)= wj]oT ( w) forall EE f, j, kE No. (7.2) It is easy to verifythat,for each j EE R, the subset Wj, w2j, ... , W2kj. . . . formsa multiplicative,orthonormalsystem,so the reasoning of the previous sections, applied to the sums wj(W) + w2j(W) + .+W2`1-i(w), insteadof to the R", shows that,foreach fixedj EE F, 1 n-i n- E wjoTk(W) -> O a.e. as n> k=O , while,trivially, 1 n-i >1asn EwooTk(w) n k=O - >oo. Consequently,using(6.3) withf replaced by XI,, 1 n-1 - E nk=O 1 n-1 x11oTk() =- n k=O 2in= j=0 2m-1 A1woP() j=O k l n-1 Ai _ , wjo Tk( n oJ k= ) AOa.e. as n >oo. Since, by (6.4), AO= 1/2m,(5.6) holds a.e., and the theoremis proved. 8. GENERALIZATION TO b-ADIC DIGITS. Suppose that b > 1 is any integral base. Let the b-adic expansionof a genericpoint w E fl be g i b1(wo) (8.1) 1 with the conventionthat b-adic rationals have finiteexpansions.Following [6], definethe b-adicRademacherfunctions r (w) = exp forall w E Q, j ( EE N. (8.2) Thus, the rj assumevalues in the cyclotomicgroupof b-throotsof unity,and each root is takenon a subsetof fl havingmeasure 1/b. It followsat once that,foreach j EE N and any integerd, dw=O, frjd(w) unless d (8.3) 0 mod b. It is also evidentthat,when d is in the rangefrom0 to b - 1, rd( () = rb-d( w)) forall E fQ,j E , so the complexconjugatesof powersof the rj are expressibleas positivepowersof (8.3) holds also when rj is replaced by its complexconjugate,as rj. Accordingly, could have been seen directly. As statedin the Introduction, we shall make use of the formula 1j=i | 1 k Hlexp(itub;( cw))dw = ~~~~~~k H flexp(i,ubjw(w)) dw ]=i 0 This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions (8.4) k EE N. The left-hand forcertainvaluesof the real parameters ,Ujand arbitrary function(Fourier characteristic integralis one formof the multidimensional considered as randomvariableson fQ ofthefirstk b-adiccoefficients, transform) thatassurestheequalityof the independence [2, p. 42],and it is theirstatistical proofcan be based uponKac's demonstration of An elementary twoexpressions. [2,p. 7]. an analogousformula Indeed,letP denoteLebesguemeasureon f. Then f H exp(i,ujbj(c)) k k E = dcv = 1 **....bk(c H| exp(i,8j,3)P[{b1(c) 1,XPk j=1 ) 1=0} overtheset0,1,... , b - 1. Usingthestatistical wheretheP3irangeindependently becomes ofthebj's,thelastexpression independence k E i. . k Hl exp(i,uj,813)HP{bJ( c) .k J= j=1 k = 83j}= I Hexp(i,uj1P)P{bJ(c) Pk =1 1 . = 18j, which,in turn,reducesto k n b-1 cP b38} ) Hl E exp(i Pj8)P{bj( j=1 = p1j=o Hrj= f1exp(i,jbj (co)) dcv, ] =10 as required. between0 and b - 1, Settingyj = 2irdj/bin (8.4),wherethe dj are integers Formula equation(8.2),theMultiplicativity yields,inviewof(8.3) and thedefining functions, fortheb-adicRademacher f'r'd(c drkk() d)r'2(cv) dc = Oforall k E , (8.5) hasthevalue1. Recallingour unlessall ofthed'svanish,inwhichcase theintegral to holdwhenany we see that(8.5) continues comments aboutcomplexconjugates, Whenb = 2, are replacedbytheirconjugates. in theintegrand numberoffactors chosenand the in thepresentform, whenk is suitably (2.1) can also be expressed all are setequal to 1,whiletheotherexponents oftheselectedfactors exponents vanish. takeanyj E N and makethepartition To generalizetheWalshfunctions, bj = dj1bil + d12bi2+ *.. (8.6) +dj1bis, in therange1,. . ., b - 1. Thenset wo 1 and define wherethe d's are integers theb-adicWalshfunctions wj(wo)) = rdi,u( )rf'i2() r' rdis(w), cv E Ql, j E N. (8.7) [6,p. 44]. Thisdefinition agreeswiththatofMendes-France thattheb-adicWalsh By(8.2) and(8.7), 1wj12= 1 forall ] E N0.It thenfollows functions theorthonormality relations satisfy f1Wj(cv)cvk(c)dcv= 8jk forj,k E N ? (8.8) for,by (8.5) and (8.7), thevalue of the integralcan be 1 onlyif wj and Wk are and theircorresponding productsof the same b-adic Rademacherfunctions, powersagree. 1999] STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions 119 Instead of the dyadicmap (5.1), considerthe b-adic map T(w) = bw (mod 1) and note that the uniqueness of the expansion (8.1) implies that bj+k(&w) bj? Tk(w) forj EE R, k EE No.Thus, in view of (8.2), we also have rj o Tk( = rj+k( (a) (a) forall wEE , k E No, f, j E the same as (7.1). Consequently,by (8.7), whichis formally wjoTk( &J) = rdl+Joj) r J2 + ... r &) (8.9)fl (fik ( &) I and so, in view of (8.7) and (8.6), Wbkj( w) = wjo Tk( w) forall wEE f, j, k E No, (8.10) whichgeneralizes(7.2). Let m be any naturalnumber,and take any string ( 8.11) 1311 1321 -* *,* O m of m b-adic digits.Let I denote the b-adic intervalof order m made up of those points in fQwhose b-adic expansionstartsout with(8.11). Clearly,I has length 1/bm. In view of (8.8), the firstbm b-adic Walsh functions,wO,w1,... , Wbm-l, are linearlyindependentover the complex numbers,and their span consists of all complex-valuedfunctionsthat are constanton b-adic intervalsof order m. In particular,XI can be writtenas bi-1 X,(a) = E j=o AjWj(w), (E fQ, (8.12) wherethe weightsAi are givenby the formulas XI( )wj( (D)d Q)= bm, j = 1, . .. I bm_1 Ai Here, -qjis a b-throot of unityforeach j, and i*1 J X.( w)wo( w) di Ao= ~~~~~1 = bm* (8.13) It follows,again from(8.8), that,foreach j (E R, the functions (8.14) W; , Wb ,I... I,Wbkj , . . . themselvesforman orthonormalsystemon fQ.Accordingly, in view of (8.10), we can establishthat,foreach fixedj EE I, 1 - n-1 n k=O wj o Tk( w) a.e. as n ->0 (8.15) -0, by appealing to the reasoning of Section 4, provided we replace the second momentsof the Rn(W) thereby the second momentsof the absolutevaluesof the sums Wj(W) + Wb](W) + forj = 1, . . . , bf+1 foreach k E N, and n EE N. Wbkk (k( ) 120 STATISTICAL = ... +Wbf-l(W) A more involvedcalculation,usingthe fact that, +k((J) INDEPENDENCE j2 .+k()) r k ) AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions [February whichfollowsfrom(8.9) and (8.10), and using,mutatismutandis,the Multiplicativand that ityFormula (8.5), showsthatthe functions(8.14) are multiplicative f1W( Wbl-lj(w)4 dw i) + Wbj( W) + < n + 3n(n - 1), so Kac's originalargumentof Section 3 can stillbe used. Once again, 1 n-i ) EwOoTk( - k=0 >oo. lasn Consequently,because of (8.12) and (8.15), 1 n-i EXioTk(w)X n k=0 AOa.e.,asn - >oo, just as in Section7. But, in view of (8.13), Ao= 1/bm.This provesthatthe relative frequencywith which the string(8.11) occurs in the b-adic expansion of the fractionalpart of almost everyreal numbertends to l/bmas n -> oo,and that is enoughto establishthe NormalNumberTheorem. 9. INDEPENDENCE AND MULTIPLICATIVITY IN BASE 2. Returningnow to base 2, observethat 1 + yN(9.1) 81/71 2 whenevery, q = ?1, as noted by Renyi [9, p. 130]. Consequently,forany j 1 + y1r1 2 =Arj= E (9.2) yj}, where rj is the j-th Rademacherfunctionand yj = + 1. Since m 17x{1=y,j} = J=1 X{rl=y ,. .r, =,=yI1} it followsthatforany m E NJ Xrl=y r (=iy , ) = (9.3) Renyiuses thisformulain the followingway.He expandsthe productin (9.3) to get jj~1Yi ri) /rl~~~~~ ( 2 ) = i(+ 2( r+ Y j 94 + -+ jH yjrj . (9 .4) yj^ylrjr, Setting (9.5) yj 1=1-2a, for =1, ...,m, whereaj = O orl, oftheindicator function thengives,inviewof(1.2) and(9.3),an expansion XI,ofa 2mWalshfunctions, as (5.5) oforderm in termsofthefirst genericbinaryinterval anyrecourseto orthonormality. in our(6.3) et seq., butwithout of offiniteproducts theintroduction Renyi'sapproachcan be usedto motivate to deal with of the Walsh functions, Rademacherfunctions, and, therefore, in itis also possibleto employtheseformulas However, strings. involving problems 19991 STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions 121 anotherway to establishan unexpectedconnectionbetweenthe multiplicativity of the Rademacherfunctionsand theirstatisticalindependence. Recall that,in Section2, we used thestatisticalindependenceoftheRademacher functions, togetherwiththe vanishingof theirmeans,to justifythe Multiplicativity Formula (2.1). However,as noted by Kaczmarz and Steinhaus,this formulacan also be established directlyby calculus, without appealing to the notion of statisticalindependence[4, p. 125f]. To do so, note firstthat(7.1) impliesthat rj = r1oTT-1 for]j E N, whereT is the binaryshiftgivenby(5.1). For fixedj, let 1 varyin the rangefrom0 to 2j-' - 1, and denote by I, the l-thdyadicintervalof orderj - 1. Since Tj-1 is linear on each such I, and maps it onto fl, while the derivativeof T;1 on I, is 2i-1, the change-of-variable formulagives r o Ti"- ( fr1(2w)dow I ) dw = r,(w) dw = 0 (9.6) foreach I,. Summingon 1 thenyields f'rj() dw = 0 for]jE N, (9.7) as assertedat the startof Section 2. To arriveat the Multiplicativity Formula(2.1), we can now employan argument used by Kac in a different connection[2, p. 27]. Order the subscriptsso that Ik is the largest,and thenwrite &J)rj2( &J) ... f1r1(w)r1(w) |Orjl( rik( d ) d -1 2ik-1 = E | JSJ(t)ri2((0) ... r1()1w 2j1E ())d, rik( 1w dw, where the I, are the dyadic intervalsof order Ik - 1. Since the first k - 1 functionsrjl,r1_, ... , r are constanton each such interval,while,by (9.6) with j = Ik,the integralof rik over each one is zero, the integralsin the sum at rightall Formula follows. vanish,and the Multiplicativity With this resultin hand, we can establishthe statisticalindependenceof the Rademacherfunctionsbyproceedingin the followingway.We integrate(9.2) over fl and use (9.7) to findthatforeveryj E NJ 1 Pt r = zyI (9.8) -2whenevery,= +1. Integrating(9.3) over fl gives P{rt = 'y,,**, rm= I) = ffl (+2 ) Now integrating Formula along with (9.4) and makinguse of the Multiplicativity (9.8) gives m 1 = = HP{r P{r= (9.9) yi) rm X yj}. ]1 2 Since m E NJis arbitrary, it followsthat multiplicativity of theRademacherfunctions impliestheirstatistical independence. 122 STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions [February The precedingreasoningis not limitedto the case of Rademacherfunctions.As a furtherapplication,recall fromSection 7 thatthe Walsh functions ... * I** W2k ... Wj , W2j (9.10) multiplicative. When j = 1, these funcare, for fixedj E%N and varyingk E %N, tions reduce to the Rademacher functions,as is evident from(6.1) and (6.2). Notingthattheyhave mean zero, we can repeat our argumentand conclude that independentforeveryfixedj E N. the functions(9.10) are, in fact,statistically (9.5) in (9.9), and do the same in (9.8), then we make the substitution If, finally, (9.9) becomes m 1 = n P{aj = aj}. P{a1 = a,,..., am= am) = Thus, statisticalindependenceof the binarydigitsis, itself,a directconsequenceof the two notionsare of the Rademacherfunctions,and, therefore, multiplicativity equivalent. entirely 10. EXTENSION TO GENERAL BASES b. All the reasoningof the preceding section can be carried over to integralbases b > 1, once we find a suitable generalizationof Renyi'sidentity(9.1). To this end, let -yand 7qbe b-throots of unityand considerthe expression b-i E yb-dd = d=O d b-1 E _1 d=O ') Since yb = _qb = 1, the quotient-q/y is also a b-throot of unity,and, therefore, the right-hand sum vanishesunless -y= 7q.Consequently, b-i 1+ E eb-d d d=1 b [q and we have foundthe identitywe need. Proceedingas before,foreach j E NJwe get b-i 1+ E yb-drd d=1 b =Xrj (10.1) Yj}X and, since,by (8.3), d1... I b - 1, da = Oforj E- N and d frd()) (10.1) gives integrating JX{rj= } j} = Pr= (10.2) = This time, b-i m X{r1=yj.. S=S rm=yn,} [ _ 1 r+=dj} Xm m 1 + E j-d d B This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions (10.3) Expandingthe second product,we get an expressionof the form 1 of the rj), bm(1 + sumsofweightedpower-products (10.4) wherethe weightassignedto each rj is )jThus, integrating (10.3) over fl and Formula (8.5) for the b-adic Rademacher functions, using the Multiplicativity yields,in view of (10.2), m 1 HPr = y1} (10.5) P{r1 = yi,...,rm = Ymi) Accordingly,multiplicativity of theb-adicRademacherfunctionsimpliestheirstatistical independence, just as in the case b = 2. In Section 8, we introducedsubsetsof b-adic Walsh functionsof the form (10.6) Wj, Wbi, ..., Wbkj, . . ., for fixedvalues of j E NJand varyingk E-N, and pointed out that theywere multiplicative, in the sense that Wbjlj( (I) ... Wbjkj( WbJ2j( ) (I) dw = 0 wheneverthe exponentsare distinct.We can now show that theyare statistically independentforeach j E N byextendingthisformulain the followingway.Let b' be the smallestpositiveintegersuch that WJb 1. Then, as a consequence of (8.5) in the extendedsense and the choice of b', the functions(10.6) are multiplicative that I d()d .(). w* bk(( w) dw = 0 foreveryk E N, where the d's are restrictedto the values 0, 1, . . ., b' - 1 and do not all vanish. It is thus possible to apply the foregoingreasoningto them,by replacingb by b' in the formulasabove. This yieldsthe statisticalindependenceof the functions(10.6), and it generalizes (10.5), which,in view of (8.6) and (8.7), correspondsto the case j = 1. Returningnow to (10.3) and setting = exp( 2y3') (10.7) for ] 1 . . ., m, with fj3Ee {0,1, ...,b - 1}, we get, in view of (8.2) and the expansion(10.4), a formulaforthe indicatorfunctions = P3,n} X{b 1= ,l 1 . . ., b,m of the rj. The of b-adic intervalsof order m, in termsof weightedpower-products resultingformulais, clearly,comparableto our (8.5), et seq., and it can be used to motivatethe introduction of power-products of the b-adic Rademacherfunctions, and, thus,of the b-adic Walsh functions,in orderto deal withproblemsinvolving b-adic strings. It can also be employedto establishan equivalence betweenmultiplicativity of the b-adic Rademacher functionsand the statisticalindependenceof the b-adic digits,generalizingthe case b = 2 treated in the previous section. Indeed, it is 124 STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions [February enoughto make the substitution (10.7) in (10.5) to get P{b= b. = m 1 bt= = FI1P(bj =[3.j 1m =1 Consequently,multiplicativity of the b-adic Rademacherfunctionsimpliesstatistical independence of theb-adicdigits. in extendingto a generalbase b the argumentused Since thereis no difficulty in the previoussection to establishthe Multiplicativity Formula for Rademacher functions,withouthavingrecourseto theirstatisticalindependence,we conclude that statisticalindependenceof the b-adic digitsis equivalentto the multiplicativity property (8.5) possessedbytheb-adicRademacherfunctions. One wayto look at thisresultis as follows.Recall that,in Section 8, we derived the Multiplicativity Formula (8.5) fromthe formula w)) dw = fl exp(i1ttjbj( H f1exp(ipjbj ( w)) dw, (8.4) by setting,tj = 2iTdj/b,where the dj are integersin the range from0 to b - 1 and j varies from1 to k. As shown there,(8.4) is a directconsequence of the statisticalindependenceof the functionsbj. However, it is known that the validityof (8.4) for all real values of the real parameters ,aj impliesstatisticalindependenceof the otherwisequite arbitrary functionsbj, and Kac himselfwas among the firstto prove it, by use of Fourier analysis[3]. In effect,what we have foundis a refinementof Kac's sufficiency result,in a more algebraicsetting,wherethe additivegroupof the reals has been replaced by the group of integersmod b. For, in restricting the bj's to take integralvalues in the range from0 to b - 1, instead of arbitraryreal values, we have established that the validityof (8.4) for the narrowrange of values of ,Ujindicatedabove is alreadyenoughto ensure statisticalindependenceof the bj's. The argumentwe have used cannot be regarded as new, even tho-ughits applicationto Rademacherand Walsh functionsis. It is a reductionto the case of the cyclotomicgroupof a methodsuggestedbyRenyi[9, p. 171] forprovinga more general resultabout the statisticalindependenceof familiesof randomvariables, attributedto Kantorovic. Renyi would, no doubt, have appreciated that the method he proposed could be applied, trivially,to the classical Rademacher functionsin orderto establishtheirstatisticalindependence,but he seems to have been unaware, as was Kac, of Mendes-France's extensionof the Rademacher functionsto the b-adic case and, thus,missedthe opportunity to applyhis method there. The reader interestedin learning more about normal numbers can consult Niven's Carus Monograph [7] and the treatiseby Kuipers and Niederreiter[5], whichcontainsan extensivebibliography. in partbytheConsiglioNazionaledelle Ricerche Thisworkwas supported ACKNOWLEDGMENT. Research. andtheItalianMinistry oftheUniversities andScientific andTechnological REFERENCES e leursapplications Rend.Circ.Mat. 1. E. Borel,Surles probabilites denombrables arithmetiques, Palermo27 (1909)247-271. 2. M. Kac, StatisticalIndependencein Probability, Analysisand NumberTheory,Carus Math. Mono- graphs, no. 12,Wiley,NewYork1959. 1999] STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions 125 3. M. Kac,Surles fonctions independentes, I, StudiaMath.6 (1936)46-58. 4. S. Kaczmarz, and H. Steinhaus, Theorie derOrthogonalreihen, Monografie Matematyczne, vol. 6, WarsawandLwow,1935. 5. L. Kuipers,andH. Niederreiter, Uniform Distribution ofSequences, PureandAppliedMath.,no. 29,Wiley,NewYork,1974. 6. M. Mendes-France, Nombres normaux. Applications auxfonctions pseudo-aleatoires, J. d'Anal. 7. 8. 9. 10. 11. Math. 20 (1967) 1-56. I. Niven,Irrational Numbers, CarusMath.Monographs, no. 11,Wiley,NewYork,1956. S. S. Pillai,On normalnumbers, Proc.IndianAcad. Sci.,Sect.A 12 (1940)179-184. A. Renyi,Foundations ofProbability, Holden-Day, San Francisco, 1970. D. D. Wall,OnNormalNumbers, Ph.D. Thesis,Univ.ofCalifornia, 1949. Berkeley, H. Weyl,Uberdie Gleichverteilung vonZahlenmod.Eins,Math.Ann.77 (1916)313-352. GERALDS. GOODMANtookdegreesin mathematics at Haverford, BrynMawr,andStanford, where he was a pupilof Loewner.His doctoralthesiswas devotedto Loewner'stheoryof transformation semigroups, appliedto conformal A chanceencounter mapping. withRenyiled himto use thoseideas to obtainnewresults Markovchains.Lateron,he appliedprobabilistic concerning toolsto Loewnerian semigroups, suchas totally matrices. he haspublished control positive Although papersinrealanalysis, hisfinest theory, andfractal imagegeneration, someconsider mathematical achievement wasfinding a permanent postin a Renaissance city. Statistics Department, University of Florence,VialeMorgagni,59, 50134 Florence,ITALY [email protected] .. . it did not take the Miloslavskyslong to find a pretextfor was foolishenoughto Matveev.Thateducatedgentleman arresting be foundwitha book of algebrain his baggage,whichwas, takento be a formofblackmagic. naturally, EvenNikita,whenhe heardofthearrestofhismentor, couldonly "He was askingfortrouble.Whatdid shakehishead and remark: he wantwithsuchstuff anyway?" Edward Rutherfurd, Ruisska,TheNovelofRussia,CrownPublishers,Inc., New York,1991,pp. 347-348. ContributedbyGeorge Dorner,WilliamRaineyHarperCollege,Palatine,IL 126 STATISTICAL INDEPENDENCE AND NORMAL NUMBERS This content downloaded from 143.107.128.41 on Mon, 12 Aug 2013 16:04:15 PM All use subject to JSTOR Terms and Conditions [February