mathematics of computation volume. number october x. pages 6-63 A Counterexample to a Conjecture of Mahler on Best P-Adic Diophantine Approximation Constants By Alice A. Deanin Abstract. In 0, Mahler proposed a conjecture regarding the value of best P-adic Diophantine approximation constants. In this paper, a computational technique which tests the conjecture for any particular P is described. A computer search verified the conjecture for all P jg 0, except 83. The case P = 83 is discussed. A counterexample is given.. Introduction to Mahler's Algorithm and Mahler's Conjecture. Mahler [] presented an algorithm which yields sequences of approximations to P-adic integers by rational numbers which are best with respect to a real reduced positive-definite binary quadratic form of determinant -. Let X be a fixed complex number in F, the fundamental domain of the modular group, i.e., the set of complex numbers in the upper half-plane which satisfy Set -{- < Rez < {- and \z\ > or -\ < Rez < 0 and \z\ =. ^ty)^{x-xy){x-\y)~^\k-kï\\ Let f be a P-adic integer. For every n > 0, let An be the unique (rational) integer which satisfies 0^An<P" and Ç = An(modPn). The algorithm defines three sequences. The first of these, z(f ), is a sequence of complex numbers. Let zn e z(l) be defined as the unique complex number in F which is equivalent by an element of the modular group to (An + X)/P". That is, for some integer matrix n r r' of determinant, (An + X)/P" = (r z + r'n)/(qnzn + q' ). The second sequence T(Ç) is a sequence of X integer matrices where, for each n > 0, Tn e T(Ç) is defined by T = P" and the action of Tn as a linear fractional transforma- Then Tn has determinant tion is ti. Pn Pn Pn = r"p" ~ 0 An, In q'nï pn = r;p"-q' A. T.z. n n = X for all n > 0. Received October, 8. 80 Mathematics Subject Classification. Primary 0F. 6 8 American Mathematical Society 00-8/8 $.00 + $. per page
6 ALICE A. DEANIN It is clear that \qj + pn\p < P~". Mahler showed in [] that if \q + p\p*í P~", (P< Q) > 0> then $(/?, q) ^ $(/?, q ). Thus P(f) determines approximations to f which are best with respect to <&( X, Y). This suggests the following Definition. The best P-adic Diophantine approximation constant is the real number cp, the supremum over all c > 0 such that \ql+p\p^p-n, 0< *(/>,?) <i/>" has solutions for infinitely many «, for all X g F and all P-adic integers f. Since T~XX = zn, it is easily seen that yn = Im zn = P"'/<b(pn, qn). Thus z G F implies that yn > i/3/, and it is concluded that cp > v/3~/ for all P. Mahler showed that cp = fh / if and only if P = (mod 6). Using a result of Davenport [], Mahler proved that lim^^c,, = i/3 /. de Weger [6] generalized this last result to a larger class of norms. For a P-adic integer f, set (f) = limj,,. (This value is dependent on X.) It is clear that () cp= min r(f). In order to study the relationship between the successive elements of z(f ), a third sequence is introduced. For every ol,,e (f ) is defined by = T~}XT. Each is a matrix of determinant P with integer entries satisfying z = zn_x. Let M(P) denote the set of X integer matrices of determinant P which satisfy p n F = 0, where acts as a transformation. Any matrix which can occur in the sequence (f ) for some P-adic integer f and some X g F is necessarily in M(P). Mahler showed that the set M(P) is finite for any P and that any g M(P) can occur as some g (f ) for some P-adic integer f (although possibly only for special choices of X). Now let m(p) denote the subset of M(P) of integer matrices of determinant P which have a fixed point in F and trace = 0 (mod P). The elements of m(p) are precisely those which can appear repeated successively in a sequence (f ) for some f (see []). For g m(p), let fa denote its fixed point in F. Let Y(P)= min Im/n. aem(p) Mahler conjectured that cp = Y(P). Moreover, he showed that for every e > 0 there is a P-adic integer f such that Y(Ç) < Y(P) + e. So with () the conjecture can be stated as Mahler's Conjecture. For any X g F and any P-adic integer f, Y(t ) > Y(P). Mahler proved that Y(P) = \/3~/ if and only if P = (mod6), thus for these values the conjecture is trivially verified. Mahler also verified the conjecture for P =,3 and. These verifications checked that, for each g M(P), if both z g F and z G F then max{im z, Im z} > (P). (For some matrices in M() and M(3) this was not the case, but such matrices could not occur more than once in succession in a sequence (f ) for any f.) This suffices to verify the conjecture, for it
COUNTEREXAMPLE TO MAHLER'S CONJECTURE 63 implies that for every -adic integer and every index n, max{yn, y +x] > (). (Similarly, for any -adic (or 3-adic) integer and every index n, max{yn, yn+x, yn+) > () (or (3)).) For this paper, the analysis is carried further by adapting the method for computer search. The conjecture was thus checked for all P < 0 and, in fact, verified for all such P except P = 83. Additional investigation showed that Mahler's conjecture is, in fact, false for P = 83. A periodic (f ) sequence can be constructed as a counterexample. Additional information about periodic sequences can be found in [, [3, [6]. This paper is organized as follows. In Section, the matrices in the finite set M(P) are described by an explicit set of inequalities which their entries must satisfy. In Section 3, a method is described that determines a value () < max{im z, Imz} when z, z g F, for each g M(P). The values of Y(P) are tabulated in an appendix. If for every g M(P), () > Y(P), then the conjecture is verified. There are primes for which this verification fails; the first such is P = 83. In Section, this failure is discussed. In Section, the counterexample for P = 83 is given.. The Set M(P). In this section, inequalities are given which the entries of a matrix in M(P) must satisfy. Since these inequalities are intended to assist in studying Mahler's conjecture, the inequalities are specifically for P = (mod 6). The derivation of these inequalities is described and a sample derivation is given. For full details, see []. Throughout the paper, let = ' - + v/3 i Because and - represent the same transformation, the inequalities will assume that > 0. It is clear that if g M(P) then -P"\ an inverse of as a transformation, g M(P) as well. Definition. The isometric circle, I, of a transformation where i= 0 is the circle centered at -'/ of radius {P /. The isometric circle, P, of the inverse transformation is the circle centered at a/ of radius P /. (A complete discussion of isometric circles can be found in Chapter of [].) The transformation acts on the complex plane by inversion through the circle / and then through the line x (a ')/. The circle / is mapped to the circle P. The region inside (outside) the circle / is mapped to the region outside (inside) the circle /'. In order to guarantee the condition p n F + 0, it is necessary that at least one of the isometric circles / or /' intersect P. Thus p or + p is inside one of the circles, which happens when () + \'\+' ^ P or a + \a\ + < P. This can be used to show that if # 0, then (, ') =.
6 ALICE A. DEANIN In [], Mahler divided the set M(P) into three subsets, MX(P),M(P) and M3(P) according to whether p n F is a region of nonempty interior, a curve, or a single point. The elements of the sets M(P) and M3(P) can be listed explicitly. The justification for these listings is detailed in []. A transformation g M(P) maps a portion of the boundary of F to a portion of the boundary of P. This happens for precisely the matrices - -(P + l) (-P) and its inverse as a transformation (l-p) ( + P) (P-l) -(P + l) and its inverse as a transformation -(P+l) (-P) A transformation g M3(P) when the corner point p of F is mapped to a boundary point of F by or its inverse. This happens precisely for the matrices -(P + l) 0 and its inverse as a transformation 0 -P i (l + p) The case for g MX(P) with = 0 is also straightforward. those of the form These matrices are all a P and its inverse as a transformation for (-P + l)/ ^a' ^(P - l)/. Let F denote the interior of P. g Mx(P) if and only if SiFnF# 0. This condition is easier to work with because of the symmetry of P. Clearly, if satisfies this condition, then so does -P", the inverse transformation. Furthermore, when > 0, the isometric circles of ' = [^ "a J and its inverse are just reflections across the imaginary axis of those of " and. Thus 'P n F + 0 whenever p n F # 0. Consequently, when /} > 0, of the four matrices a a ' -' ' a' a ~' it is only necessary to consider one. The inequalities for the entries of with > 0 are determined assuming that a > 0 and '\ > a. Assuming that > 0 and \'\ > a ^ 0, () reduces to MX(P) - a + a < P.
COUNTEREXAMPLE TO MAHLER-S CONJECTURE 6 This yields an upper bound for in terms of P, as well as upper and lower bounds for a in terms of and P. However, to determine ranges for ', it is effective to break the problem into four cases: (i) a + a + < P and a/ > i (ii) a + a + < P and a/ < \. (iii) a + a + > P and a/ > f (iv) a + a + > P and a/ < \. In each case, bounds for a, and ' can be determined by examination of the image of F under ". The inequalities in case (i) are derived below as an example; for the other cases, see []. The inequalities for all four cases are summarized in Figure. Case (i). The conditions defining case (i) imply - JP - 3> < a < I ^ This interval is nonempty when < {P / y/l. If ' > a, XP n F =h 0 if and only if at least one of _(p), -(i + p) has real part > - I- (Because a + a + < P, their imaginary parts are large enough.) Thus, -{-< max{re^(p),_(i + p)} -maxa"^. P- =l+a^- -\ a + a + ' /3? _ a + : which becomes ( a + P a- P \ ' < max +?- --, + -, a + a + ' a-a + j' The former term is larger when a <. If ' < -a, then _P lies entirely to the right of x = a/ > \ and cannot intersect P. The entries a, and ' of a case (i) matrix must satisfy a < /?' < o</3<-^;p~, V /3 < -, VP - 3/3' ^ a + P a + a/3 + /3 P, a- P a - a + Figure gives a flow diagram for deciding when a matrix with > 0 and /3' > a > 0 has P d F = 0. Find the range on the left line where falls. For the a scale assigned to that range, find the range in which a falls. The case inequalities named for that range give allowable choices for '. The case inequalities are listed at the right.
66 ALICE A. DEANIN case i *L g, q+p _P rt az+o.+ + r-r + R! a-g a'-a+f case i.m -g, VP^3^ V? /g VP-3g, ^+Vp^ ^+a^ F or a+a/?+/? P a-a/3+/ *VT VP-3ff ^ ff! VP-36 a</3'< f- + _J* s)* i'-* ' J cp-a+* if c>/3 «^/3'< f-+/3 /-a 3]* _ _ il a az-~z *-a< V3 /?! VP-3ff P op-a-rf Figure 3. Testing Mahler's Conjecture. In this section, techniques used to determine the value () for each in M(P), P = (mod6), are discussed. The computations require an involved case analysis. Sample computations are given. Definition. For any matrix G M(P) () = inf{max{imz,imz}}, where the inf is taken over all z for which z g F and z g F. The values of () for in M(P) or M3(P) are easy to determine. These values are quite large compared to Y(P) and do not figure in the conjecture testing. The values of () for upper-triangular are also large; such matrices can thus be eliminated from serious conjecture testing work. The complicated testing arises in computing () for = I,\ g MX(P) which have > 0. The value of () will be the same for the four matrices a a ' -a a so only those for which > 0 and \'\ > a > 0 are studied. -' ' a' a -' a' -a
COUNTEREXAMPLE TO MAHLER'S CONJECTURE 6 For g MX(P), () = inf{max{im z, Im z}}, where the inf is taken over all z for which z and z g F, the closure of P. The boundary points of P frequently yield the value for (). The matrices are divided into cases depending on the position of the isometric circles with respect to the boundary of P. (i) a + a + < P and a < ' < a +. (ii) a + a + > P and a < ' < (a + )/ + (P - )/a. (iii) all other matrices with a < '. (iv) a + a + < P and -a > ' > a -. (v) all other matrices with -a ^ '. For each, a lower bound for () can be established. For most matrices, the value established equals (), but calculating this exact value is impractical in some "extreme cases", e.g., when ' is large. Two techniques are described. For cases (i), (ii) and (iv), both the isometric circles / and /' of and " intersect the boundary of P. If a point z is on / (resp., /'), then Im z = Im z (resp., Im z = Im_z). If z lies inside / (resp., I'), then Im z < Imz (resp., Im z < Im _z). Thus a good candidate for () is the imaginary part of the lowest z on an isometric circle for which both z and z or z and ~'z are in P. Denote this value It is still possible that there is a z g P with Im z < Yx and z (resp., _z) g F with Im z (resp., Im _z) <,. The corner points p and + p and low boundary points of F are likely candidates for such a z. The behavior of these points must be compared with that of the isometric circle point which produced. The best comparison value is denoted Y. Sometimes the value easily identified for is not the best possible, but (0) > min{yx,y). In cases (iii) and (v), the points on the isometric circles in F do not map (under or ") to points in F, or such points are difficult to determine. A valuey(q) may be computed instead. Let G be the set of points of the lower-boundary arcs of _P which are contained in P. Set ^() = max/ inf {Imz}, inf (Imz}}. ' zec zeg ' Since inf.eg{ Imz} < () and inlec{imz} < (), y(q)< (). The details of case (i) are presented as an example, with figures at the end; the details of the other cases are included in []. Notation. The intersection point of two curves Cx and C will be denoted Cx d C. Case (i). Let zq = F d(x=-\), >_ild(x=\) if + ' + '<P, Z \ld(\z\=l) if + ' + '>P. These are the possible isometric circle points described earlier. Set Yx = max(im z0, Im z'0}. Yx must be compared to the values of Im _z, where z, l~z g p are inside /'. For each of six subcases, a value is defined which is < the least such Im _z.
68 ALICE A. DEANIN The condition implies that Re _(p) > - { n a + P '<a + <^ + -r- -- a + a + (a)_(p)g F: set = Im_(p). Then () = min{,}. (b)re_(p) > \: let z, = Q~\x = - \) d (x = \). Set = max{im zx, Imzj}. Then () = min{, }. In (c)-(f), _(p) lies in the small region where - \ < Rez < \, Im z > t/3/ and z <. In each case, < () < Yx. (c) Im(_(x = -j)d(x = -\))< \ 3 /: let Zj = _(x = - ) n ( z = ) and z = -( z = l)n( z = l). In (d)-(f), Imt-H* = - è) H (x = - I)) < /3/. (d)re"(l + p)> i: let Zj = B_( * = ) n ( z = ) and z = Q_( z = ) n (* = *) Case (i) Case (ia) Case (ib) w^ Case (ic) T Ä & Case (id) Case (ie) Figure Case (if)
COUNTEREXAMPLE TO MAHLER'S CONJECTURE 6 (e) Re"'(l + p) < -\: let zx -x( z = ) n (\z\ = ) and z = tt~l(\z\ = ) d(x= - ). (f) Re "'(l + p) g P: let Zj = ü~\\z\ = ) n = ) and z = _(l + p). In each of the above four cases, set = max{min{im zx, Im z}, min{imz,imz}}.. The Results of Computer Testing. In the computational testing of Mahler's conjecture, a program was written in PASCAL, and the UNIVAC 80 at the University of Maryland was persuaded to run it. The program tested each matrix whose entries satisfy the inequalities in Figure. For each such matrix, a value *() min{ j, } y(&) in cases (i), (ii), and (iv) in cases (iii) and (v) was established. In every case, *() < (). Let Y*(P) = minh *(). The program was run for all P = (mod 6) up to P = 0. It was found that Y*(P) = Y(P) for all the values of P tested except P = 83, so the conjecture is true for these values. The failure of the test for P = 83 does not show that the conjecture is false. The program tested the much stronger statement, that () > *() ^ Y(P) for all G M(P), and this fails for P = 83 in an obvious way. Consider the isometric circles / and /' for = [ J whose fixed point has imaginary part (83). Let z0 = I' n (x = - j); then z0, _z0 g F and Im z0 = Im _z0 = v^83~/8 < v/6 /8 = (83). Such a point exists because the isometric circles intersect the boundary of F below F(83); this happens because the fixed point of does not lie on the unit circle. The same problem arises for other P, whenever a fixed point / of an elliptic matrix has Im/= Y(P) and / ^. Since / = -a'/, it is easy to identify such P by examination of the table in the appendix. For P < 000, they are 83, 6,,, 3, 6,, 8, 6, 3, 3,, 8,, and.. A Counterexample. It is not possible to disprove the conjecture using only the fact that () < (83) for the matrix [" " ]. However, the computer testing found four other matrices satisfying the inequalities in Figure for which () < (83); they are This yields 0 matrices in Mx(%3) having () < (83). An examination of products of pairs of these matrices found -8 8-0 an elliptic matrix whose fixed point - + \/3003 //8 has imaginary part less than (83), and - - 3 3 8-3 -3. -3 - an elliptic matrix whose fixed point (- + v/3003/)/6 has imaginary part less than (83). For a specific X G F, a periodic sequence of the above matrix factors can be shown to be the (f ) sequence of a P-adic integer f with (f) < (83), disproving the conjecture.
630 ALICE A. DEANIN In [], Mahler proved that a sequence { }^ = is the sequence (f ) for some f if and only if, for a given X, and setting T =, n n Pn P'n Tn-lX<=Fand(q,q' )=l. For the counterexample, let and let = X = - + /3003, 8 for n odd, for n even. Then, by definition, T = 8-8 -0-8 8-0 (n-l)/,,/ -3 for n odd, for n even. Note that T~lX = - + i/3003, g F 6 - + y/3003 ; 8 F for n odd, for n even. Furthermore, by a congruence relation in Section of [], for even n, T = T k 8-8 -0J k-l ( - 0) 8-8 -0 (mod 83), hence q and q'n are not both divisible by 83 and must be relatively prime. Also, for odd n, T = T- k + \ 8-8 -0 = ( - 0)' 8-8 -0 (mod 83) so that again, (q, q' ) =. This verifies that the sequence { } is a legitimate (f ) sequence; it is in fact the sequence for f = -X = ( - v3003 /)/8 with the indicated X. This is a counterexample to the conjecture because (f) = lim Im T~lX = max I v/3003 v/3ööt Ï v/ 8 6 (83). It is likely that such constructions exist for other P for which Y(P) occurs as the imaginary part of a fixed point of a matrix in m(p) where the fixed point is not on the unit circle.
COUNTEREXAMPLE TO MAHLER'S CONJECTURE 63 The nature of the construction of the counterexample suggests that the value of the best P-adic Diophantine approximation constant may be determined by the imaginary part of a fixed point in P on the unit circle of an elliptic matrix of determinant Pg. This possibility is being studied. Acknowledgments. The author would like to thank W. Adams for suggesting the problem and for his help and encouragement before, during and after the preparation of the dissertation in which much of this work orginally appeared. The author is also indebted to D. Shanks for his probing comments on the validity of Mahler's conjecture. Appendix. The table below gives Y(P) for all P = (mod 6) less than 000, along with a matrix in m(p) whose fixed point has imaginary part Y(P). P Y(P) 3 3 83 8 0 0 3 3 3 6 _ 6 Û. 3 y/ 0 y/ 0 y/ïïï y/ï8 fñ 0 8 y/ 8 fjâ_ irj_ y/6 i/ô3 /ÖT /03 /Ï33 3 Ï33 3 /Ï3 matrix - 3-3 - - - 8 - - - -X -8 3-0 3 -'. -X - -0-0 - 0 - - - 0-3' -3-3 P 3 33 3 63 6 X 3 3 3 3 33 3 3X3 Y(P) y/0j y/ 8 y/0t y/ï33 3 y/ït y/ï33 3 y/3t 3 y/ y/l6 y/3t 3 y/6 y/80 y/3 30 y/03" y/ 38 y/ïî 38 y/l03 y/8 8 y/3 0 3-3 3-0 - -3 - IX -0 0 - - - IX - 0-3 - - -3-6 -3 6 - -8-6 - -8 - X - - - - - X -'. -0 /' 38 0 3 3 6 6 03 0 63 6 8 3 Y(P) y/ï33 3 y/80 y/6 6~ y/6 6 y/3 3 y/80 y/6 6 y/3 3 y/ y/ y/03 y/6 3 y/0 y/l33 3 y/3t 30 y/60 y/8 y/ y/ 0 3-3 - 3-3 - 3 - - -0 3-3 - 0-3 0 - - -] J 3 - J - -6 6 3-6 -6-0 -3 - -8 3 - X - 3 J
63 ALICE A. DEANIN P 6 6 6 63 6 6 6X3 0 3 Y(P) 6 /6Ö~ /63 8 /33~ 8 /8 8 /3Ï 0 /3 30. - - 8-3 6 - -3-0 0-30 - 8 - - -6-3 - 6 - - -0-3 P 6 3 X0 X X 83 8 X63 XXI Y(P) 36 3 Im /6 3 /8Ô /Ï33 6 /36 3 / 3 AsT -0 3-30 - 3 6-0 3-6 3-3 3 8-3 0-3 -3 3-3 -6-3 3-3 3-3 -3-3 XX 3 83 Y(P) y/03 y/6 3 y/6 y/ 6 y/866 6 /83T 0 3-3.30-3 3-33 3 3-8 33-3 - 3-3 3-30 -33-3 - -3 3 6-3 3-33 -3 Department of Mathematical Sciences Villanova University Villanova, Pennsylvania 08. H. Davenport, "Note on linear fractional substitutions with large determinant," Ann. of Math., v., 0, pp. -6.. A. A. Deanin, Mahler's P-Adic Continued Fraction Algorithm, Ph.D. Dissertation, University of Maryland, 83. 3. A. A. Deanin, "Periodicity of P-adic continued fraction expansions." (Preprint).. L. R. Ford, Automorphic Functions, nd ed., Chelsea, New York,.. K. Mahler, "On a geometrical representation of/>-adic numbers," Ann. of Math., v.. 0, pp. 8-6. 6. B. M. M. de Weoer, Approximation Lattices ofp- A die Numbers, Report no., Math. Inst. R. U. Leiden, 8.