MMARKOV SPECTRA FOR MODULAR BILLIARDS
NICKOLAS ANDERSEN AND WILLIAM DUKE
In memory of Harvey Cohn (1923–2014)
Abstract.
We introduce some analogues of the Markov spectrum defined in terms ofmodular billiards and consider the problem of characterizing that part of the spectrumbelow the lowest limit point. Introduction
The abstract triangle group usually denoted by ∆(2 , , ∞ ) is generated by A, B, C subjectto the relations A = B = C = ( AB ) = ( AC ) = 1. The extended modular groupΓ = PGL(2 , Z ) gives a faithful representation of this triangle group when we make theidentifications:(1.1) A = ± (cid:18) (cid:19) , B = ± (cid:18) − (cid:19) , C = ± (cid:18) − (cid:19) . The usual modular group PSL(2 , Z ) is the subgroup of index 2 consisting of all matrices inΓ with determinant one.Let H be the upper half-plane with its hyperbolic metric given by ds = | dz | y . It is wellknown that M = ± ( a bc d ) ∈ Γ with det M = 1 acts as an orientation preserving isometry of H through(1.2) z (cid:55)→ M ( z ) = az + bcz + d , while when det M = − z (cid:55)→ M ( z ) as an orientation reversing isometry.The generators A, B, C give reflections across the unit circle, the y -axis, and the line x = ,respectively; Γ acts as a reflection group. A convenient fundamental domain for Γ is thesolid hyperbolic triangle T = { z ∈ H ; 0 ≤ Re z ≤ , | z | ≥ } , whose sides are fixed by the generating reflections and which is the shaded region depictedin Figure 1.Let S be an oriented geodesic in H . Thus S is given either by a directed vertical half-lineor a directed semi-circle that is orthogonal to R and is uniquely determined by ordering itsendpoints, say α, β , which are distinct elements of R ∪ {∞} . More generally, for z , z ∈ H ∪ R ∪ {∞} let (cid:104) z , z (cid:105) denote the geodesic segment connecting z to z . Hence we maywrite S = (cid:104) α, β (cid:105) .The set of all geodesics splits into orbits Γ S under the action of Γ, where S is a fixedgeodesic. Let B denote the set of distinct directed geodesic segments in T of an orbit Γ S . Date : March 15, 2018.Supported by NSF grant DMS 1701638. a r X i v : . [ m a t h . N T ] M a r NICKOLAS ANDERSEN AND WILLIAM DUKE
Figure 1.
The modular triangle T and a modular billiardWe will refer to B as the trajectory of a modular billiard, but usually call it simply a modularbilliard . We will say that B is induced by S for any S in the orbit. Note that B can bethought of as the path of a point acting like a billiard ball bouncing off the sides of T , withwell-defined bounces from the corners of T , which are at z = i and z = ρ = + √− . Suppose that B is induced by (cid:104) α, β (cid:105) . Define its reversal B ∗ to be the billiard induced by (cid:104) β, α (cid:105) . We say the billiard B is non-orientable if B = B ∗ , orientable otherwise. If B containsa vertical segment we say it is improper , otherwise proper . If the total hyperbolic length ofthe segments in B is finite, we call the billiard periodic . Clearly a periodic billiard is proper.The billiard illustrated in Figure 1 is non-orientable and periodic.The simplest modular billiard, which we will denote C , is that induced by the imaginaryaxis (cid:104) , ∞(cid:105) . It covers the segment connecting i to infinity. The billiard induced by (cid:104) , ∞(cid:105) ,denoted C , covers the rest of the boundary of T . Both C and C are improper and non-orientable.In a prescient article of 1924, Artin [2] observed that properties of continued fractionsimply that a generic modular billiard is dense in T . On the other hand, the behavior of Figure 2.
Increasing segments of a billiard exhibiting generic behaviora non-generic billiard is subtle and can be quite interesting arithmetically. For instance, amodular billiard B has a maximal height, possibly infinite, defined to be the supremum ofimaginary parts of points on B . Let λ ∞ ( B ) be twice this maximal height. Consider the set M ∞ = { λ ∞ ( B ); B is a modular billiard } . This is the
Markov spectrum , which is usually defined (equivalently) in terms of the minimaof indefinite binary quadratic forms. The Markov numbers are those positive integers p for ARKOV SPECTRA FOR MODULAR BILLIARDS 3 which there are q, r ∈ Z + such that p + q + r = 3 pqr. These may be ordered into an infinite increasing sequence whose n th term is denoted by p n : { , , , , , , . . . , p n , . . . } . The following result is a consequence of the fundamental work of A. A. Markov [19]:
Theorem 1.
For any fixed κ < there are only finitely many modular billiards B with λ ∞ ( B ) < κ. The points in M ∞ less than are given by the sequence (cid:26) √ , √ , √ , √ , √ , . . . , √ p n − p n , . . . . (cid:27) , which is monotone increasing to the limit ∈ M ∞ . Figure 3.
Billiards associated to the points √ , √ , √ , √ ∈ M ∞ It is also known that each of the points < M ∞ is actually attained by a non-orientableperiodic billiard (see Theorem 75 of [11]) and it was conjectured by Frobenius [15], but isstill open, that the multiplicity of each of these points is one, meaning that the associatedbilliard is unique. The part of the Markov spectrum that is > M ∞ and that M ∞ is closed, butthere are few completely definitive results known. Building on pioneering work of Hall [16],Freiman [13] obtained one such result. He showed that [ µ, ∞ ) ⊂ M ∞ , where µ = √ = 4 . . . . , with no smaller value of µ being possible.We remark that the seminal work of Harvey Cohn, beginning with his 1955 paper [7],revealed among other things a completely unexpected relation between the Markov spectrumand the length spectrum of simple closed geodesics on the modular torus. His work has hada lasting impact on the study of simple closed geodesics on Riemann surfaces. It has alsoled to a better understanding of the Markov spectrum itself. See [8] for a summary of someof his contributions.The value λ ∞ ( B ) − may be thought of as a measure of how close the billiard B gets tothe corner of T at the cusp i ∞ . It is natural to ask how close a modular billiard must getto each of the other corners i and ρ of T . By the distance of a billiard from a point z ∈ T ,denoted by δ z ( B ), we mean the infimum of the hyperbolic distance between points on thebilliard and z . Let(1.3) λ z ( B ) = (sinh δ z ( B )) − . NICKOLAS ANDERSEN AND WILLIAM DUKE
A natural analogue of the Markov spectrum is(1.4) M z = { λ z ( B ); B is a modular billiard } for a fixed z ∈ T . In this paper we will give results about M ρ and M i that correspond to Markov’s for M ∞ .The result for z = ρ is quite easy to prove. Theorem 2.
The smallest value in M ρ is √ , which is attained by C . The value √ is alimit point of M ρ . The result for z = i is deeper and most of this paper is devoted to its proof. Theorem 3.
The three smallest values in M i are { √ , √ , (3 + √ } = { . ..., . ..., . ... } . These three values are attained, respectively, by unique billiards C , C and C , each properand non-orientable. Here C and C are periodic billiards, while C is not periodic. The value (3 + √ is a limit point of M i . Figure 4.
Billiards C , C and C Explicitly, C is induced by the geodesic (cid:104) (1 − √ , (1 + √ (cid:105) , C is induced by thegeodesic (cid:104) (2 −√ , (2+ √ (cid:105) and C is induced by the geodesic (cid:104) (3 −√ , (5+ √ (cid:105) .In both cases the rest of the spectrum invites investigation. It is also of interest to considerthe Markov spectrum M z for other points in T , in particular CM points. In addition todistances from a fixed point, there are other geometric quantities associated to non-genericmodular billiards whose sets of values define Markov-type spectra. The purpose of thispaper is to initiate a study of these generalizations by concentrating on the simplest andmost natural examples and giving the analogues of Markov’s results for them.In the next section we give a geometric interpretation of Theorems 1–3 in terms of thepacking of discs in tessellations formed by geodesic segments and prove the first statementof Theorem 2. In § § § λ i ( B ) in terms of them. Then we give in § ARKOV SPECTRA FOR MODULAR BILLIARDS 5 between proper modular billiards and doubly-infinite sequences of positive integers thatconnects billiards to simple continued fractions. This connection is exploited in § § § Acknowledgement:
The second author thanks Alex Kontorovich for some enlightening discussions on thetopics of this paper. Packing discs in hyperbolic tessellations
Elementary geometric considerations provide some useful insight into Theorems 1–3 andserve to establish “trivial” bounds for λ z ( B ). The problem of finding points of M z is equiva-lent to the problem of fitting geodesics in H between discs of varying radii around the imagesunder Γ of z .Consider the case of the original Markov spectrum M ∞ . A Ford circle is the horocyclearound the reduced rational number p/q with radius q . The set of all Ford circles forma packing of the tessellation Γ (cid:104) i, ρ (cid:105) . See the left hand side of Figure 5. It is obvious thatevery geodesic S must intersect infinitely many Ford circles. This gives that λ ∞ ( B ) ≥ , or M ∞ ⊂ [2 , ∞ ) . Ford [12] proved that if we reduce the radii of the Ford circles to any r ≥ r = √ q it still forces intersection but if r < r there are geodesics that intersect no circle. See theright hand side of Figure 5. Figure 5.
Ford circles
Proof of first statement of Theorem 2.
The first statement of Theorem 2 may be proventhis way. To show that λ ρ ( B ) ≥ √ log 3 around the points Γ ρ form tangent sequences that approach a dense subset of R . Itis straightforward to show that they are tangent to the Farey triangulation Γ (cid:104) , ∞(cid:105) . Anygeodesic must intersect these circles if their radius is made any larger since its endpoints willbe separated by a sequence of circles. See the left hand side of Figure 6.As we will prove below and is illustrated in the right hand side of Figure 6, if the radiiare reduced by any positive amount there are infinitely many inequivalent geodesics thatintersect no circle.Turning to Theorem 3, we can pack Γ (cid:104) i, ∞(cid:105) by geodesic circles centered at the points Γ i ofradius log( √ ) = 0 . . . . . This implies that λ i ( B ) ≥
2, as is illustrated in the left handside of Figure 7. This is weaker than the consequence of Theorem 3 that λ i ( B ) ≥ . . . . . as the right hand side of Figure 7 illustrates.This point of view sheds light on why the Markov-type result for the distance problem iseasier for z = ρ. The corresponding tessellation in this case comprises complete geodesics,while in the other two cases only geodesic segments.
NICKOLAS ANDERSEN AND WILLIAM DUKE
Figure 6.
Illustrating Theorem 2 by disks around images of ρ Figure 7.
Disks around images of i For example, we easily get the first statement of the following result using the tessellationΓ (cid:104) , ∞(cid:105) = Γ (cid:104) i, ρ (cid:105) ∪ Γ (cid:104) ρ, ∞(cid:105) , whose associated billiard is C . Figure 8 illustrates the packingof this tessellation by disks around images of √− log 22 . Theorem 4.
The smallest value in M i is √ , which is attained by C . The value √ is alimit point of M i . Figure 8.
Disks around images of √− Binary quadratic forms and billiards
To go beyond this basic geometric method we need a usable formula for the distancebetween a billiard and a point. Binary quadratic forms provide the key. In this section weestablish their relation to modular billiards of the various kinds.For a, b, c ∈ R with d = disc( Q ) = b − ac (cid:54) = 0 let Q ( x, y ) = ax + bxy + cy , ARKOV SPECTRA FOR MODULAR BILLIARDS 7 which is a non-singular real binary quadratic form. Sometimes we will write Q = ( a, b, c ) . Now M = ± (cid:0) a (cid:48) b (cid:48) c (cid:48) d (cid:48) (cid:1) ∈ Γ = PGL(2 , Z ) acts on Q by(3.1) ( Q | M )( x, y ) def = (det M ) Q ( a (cid:48) x + b (cid:48) y, c (cid:48) x + d (cid:48) y ) . Clearly Q | ( M M ) = ( Q | M ) | M for M , M ∈ Γ. We say that two such forms Q and Q are equivalent if there is an M ∈ Γ such that( Q | M )( x, y ) = Q ( x, y ) . If M ∈ PSL(2 , Z ) then we say that Q and Q are properly equivalent. The class of formsthat are equivalent to Q , but not necessarily properly equivalent to Q , will be denoted by[ Q ]. The discriminant disc( Q ) is an invariant of [ Q ].If Q = (0 , b, c ) with b > α Q = − cb and β Q = ∞ , while if b < β Q = − cb and α Q = ∞ . Otherwise the roots of Q ( z,
1) = 0 are given by(3.2) α Q = − b + √ d a and β Q = − b − √ d a . In all cases α Q will be called the first root and β Q the second root of Q . One checks that d , α Q and β Q uniquely determine Q. Furthermore, using the generators
A, B, C from (1.1), itfollows that for each j = 1 , M = ± (cid:0) a (cid:48) b (cid:48) c (cid:48) d (cid:48) (cid:1) ∈ Γ α Q | M = M − ( α Q ) and β Q | M = M − ( β Q ) , with M ( z ) given in the definition around (1.2) extended to all of C . It is important that theaction of Γ on quadratic forms defined in (3.1) preserves the first and second roots.The proofs of the following two results are straightforward. Proposition 1. (i) The map Q (cid:55)→ α Q determines a bijection between classes [ Q ] with a fixednegative discriminant that are represented by positive definite Q and points of T .(ii) The map Q (cid:55)→ (cid:104) α Q , β Q (cid:105) determines a bijection [ Q ] ↔ B between the classes [ Q ] with afixed positive discriminant and the set of all modular billiards, where the associated billiard B is induced by (cid:104) α Q , β Q (cid:105) . Say Q represents zero if Q ( x, y ) = 0 for some x, y ∈ Z not both zero, that Q is reciprocal if Q is equivalent to − Q and that Q = ( a, b, c ) is primitive integral if a, b, c ∈ Z withgcd( a, b, c ) = 1 . We have the following characterization of improper, non-orientable andperiodic billiards in terms of quadratic forms.
Proposition 2.
Under the bijection [ Q ] ↔ B of (ii) of Proposition 1, for any Q ∈ [ Q ] (a) Q represents zero if and only if B is improper,(b) Q is reciprocal if and only if B is non-orientable,(c) For some non-zero real κ the form κQ does not represent zero and is primitive integralif and only if B is periodic. The study of periodic billiards is the same as that of (primitive) integral binary quadraticforms. It follows from Proposition 2 (c) that each periodic modular billiard B may beassigned a unique positive integer given by d = disc( B ) def = κ disc( Q )for any Q ∈ [ Q ]. There are only finitely many periodic billiards with a given discriminant d and each has the same length. NICKOLAS ANDERSEN AND WILLIAM DUKE
A formula for the length of B is determined by finding that solution ( t, u ) with t, u ∈ Z + of t − du = ± ε = ( t + u √ d ) is minimal. Then the length is 2 log ε. If t − du = − Q is improperly equivalent to itself when Q | M = − Q where M ∈ Γ with det M = − x, y ) (cid:55)→ ( x, y ) M t preserves Q . It follows from Theorem 90 of [10] that this holds if and only ifsome ( a, b, c ) ∈ [ Q ] has a | b . These forms are called ambiguous and one may say that thebilliard B associated to [ Q ] is ambiguous. By Proposition 2 (b) we have that an ambiguous B is non-orientable. If t − du = − Q is properly equivalent to − Q . Markov’sbilliards are examples. If d is fundamental, the number of non-orientable billiards is 2 ω ( d ) − ,where ω ( d ) is the number of distinct prime factors dividing d .4. Quaternary quadratic forms
A consequence of Proposition 2, one that is crucial for us, is a formula forsinh δ z ( B ) = ( λ z ( B )) − from (1.3). This involves finding the minimum of a certain indefinite quaternary quadraticform. For a fixed z ∈ T let Q (cid:48) ( x, y ) = a (cid:48) x + b (cid:48) xy + c (cid:48) y with a (cid:48) > d (cid:48) = b (cid:48) − a (cid:48) c (cid:48) < z . Also, let Q ( x, y ) = ax + bxy + cy with d = b − ac > B . Proposition 3.
Notation as above, sinh δ z ( B ) = ( d | d (cid:48) | ) − inf [ Q ] | c (cid:48) a + 2 a (cid:48) c − b (cid:48) b | . Proof.
A standard exercise in hyperbolic geometry shows that the hyperbolic distance δ ( z, S )from z to the geodesic S = (cid:104) α Q , β Q (cid:105) satisfies(4.1) sinh δ ( z, S ) = ( d | d (cid:48) | ) − | c (cid:48) a + 2 a (cid:48) c − b (cid:48) b | . See e.g. [3, p. 162]. The result follows. (cid:3)
We are now able to justify the second statements of Theorems 2 and 4. For fixed (cid:96) ∈ Z + let A (cid:96) be the billiard associated to the quadratic form Q ( x, y ) = x − (cid:96)xy − y . The case when (cid:96) = 5 is illustrated in Figure 1, which is typical in that of those geodesicsin the orbit intersecting T , the one corresponding to Q is the one that gives the closestapproach to ρ. By Proposition 3 we havesinh δ ρ ( A (cid:96) ) = (cid:96) √ (cid:96) + 12 . Therefore λ ρ ( A (cid:96) ) = √ (cid:96) +12 (cid:96) , which decreases to the limit √ (cid:96) → ∞ . This completes theproof of Theorem 2. The second statement in Theorem 4 follows in like manner.Note that we may rewrite (4.1) using (1.3) as(4.2) λ z ( B ) − = ( d | d (cid:48) | ) − inf x x − x x = ± | Q (cid:48)(cid:48) ( x , x , x , x ) | , ARKOV SPECTRA FOR MODULAR BILLIARDS 9 where for a fixed choice of Q ( x, y ) = ax + bxy + cy representing B we have Q (cid:48)(cid:48) ( x , x , x , x ) =2 a (cid:48) cx + 2 a (cid:48) ax + 2 c (cid:48) cx + 2 c (cid:48) ax + 2 b (cid:48) ax x + 2 b (cid:48) cx x − c (cid:48) bx x − a (cid:48) bx x − b (cid:48) bx x − b (cid:48) bx x . Here Q (cid:48)(cid:48) is an indefinite quaternary quadratic form of signature (2 , Q (cid:48)(cid:48) ( x , x , x , x ) subject to(4.3) x x − x x = ± x , x , x , x ) (cid:54) = (0 , , , . The formula for λ ∞ ( B ) from the Markov spectrum corresponding to (4.2) is simply(4.4) λ ∞ ( B ) − = d − inf ( x ,x ) (cid:54) =(0 , | Q ( x , x ) | . In this sense the problem of finding λ i ( B ) is more difficult than that of finding λ ∞ ( B ). Thestudy of the minima of certain indefinite quaternary forms subject to (4.3) goes back at leastto a 1913 paper of Schur [20], which was an inspiration for this paper and deserves to bebetter known.At this point we may obtain a good lower bound for λ i ( B ) when B is improper. It iseasy to check that an improper billiard is determined by some (cid:104) α, ∞(cid:105) , where 0 ≤ α ≤ , orequivalently by the form Q = y ( x − αy ) for this α. Proposition 4.
For B an improper billiard we have that λ i ( B ) ≥ and this is attained by the billiard determined by (cid:104) , ∞(cid:105) .Proof. Let Q ( x, y ) = y ( x − αy ). Then λ i ( B ) − ≤ min( α, − α ), which is found by applyingProposition 3 in the form (4.2) to Q (cid:48)(cid:48) ( x , x , x , x ) = − α ( x + x ) − x x − x x and taking ( x , x , x , x ) = (1 , , , , (1 , − , , α = 1 / (cid:3) Proper billiards and reduced forms
We say that a form Q = ( a, b, c ) with discriminant d > reduced if − < β Q < α Q > , where α Q and β Q were defined in (3.2). A classical argument given in the proof of Theorem76 in [10] may be adapted to prove that for any proper billiard B the corresponding class[ Q ] (as in Proposition 2) contains such reduced forms.Given a form Q = ( a, b, c ), let(5.1) Q ∗ ( x, y ) = − Q ( − y, x ) . If Q is reduced then Q ∗ is a reduced form that is properly equivalent to − Q. Note that the geodesics associated to the reduced forms do not necessarily account for allof the geodesic segments comprising a modular billiard. This fact is illustrated in Figure 1,where the single geodesic associated to a reduced form is shown in black. In the proof ofMarkov’s Theorem 1 the maximal height of a billiard B will be approached by the heights of geodesics associated to reduced forms. Thus by (4.4) it follows that we have the simpleformula(5.2) λ ∞ ( B ) − = d − inf Q ∈ [ Q ] reduced | a | . To obtain an analogous formula for λ i ( B ) − we must consider some transforms of reducedforms. This motivates the following definition. For any for Q = ( a, b, c ) with discriminant d > ν ( Q ) = d − min (cid:0) | a + c | , | a + b + c | , | c + b + a | (cid:1) . Proposition 5.
For a proper B we have λ i ( B ) − = inf Q ∈ [ Q ] reduced ν ( Q ) , where the class [ Q ] corresponds to B .Proof. To prove this we will show that for any geodesic in H that intersects T , we can find aform Q = ( a, b, c ) such that either Q or − Q is reduced and that the geodesic correspondingto Q ( x, y ), Q ( x + y, y ), or − Q ( x, x − y ) is as close or closer to z = i . The result then followsby Proposition 3 applied with Q (cid:48) = (1 , ,
1) since the terms in (5.3) correspond exactly tothese three cases.Note that we may restrict our attention to geodesics S that either (i) cross both verticalsides of the boundary of T or (ii) cross the right vertical side and the circular arc of thisboundary. This is because the reflection across the y -axis of a geodesic that crosses the lefthand vertical side and the circular arc will cross both vertical sides and will also have thesame distance from z = i .In case (i) we may assume that the apex of the geodesic S lies on or to the right of the y -axis; if not, the reflection across the y -axis of S will have that property and be the samedistance from i . Let (cid:101) α, (cid:101) β be the roots of a form (cid:101) Q associated to S . Then we have (cid:101) α > , (cid:101) β < , (cid:101) α − (cid:101) β ≥ , and (cid:101) α + (cid:101) β ≥ . Thus there is a unique integer n ≥ Q ( x, y ) = (cid:101) Q ( x − ny, y ), obtained byshifting S to the right n units, is reduced. If n = 0 or 1 then we are done because either (cid:101) Q is reduced already or (cid:101) Q ( x, y ) = Q ( x + y, y ). Suppose that n ≥
2. Then (cid:101) β < − (cid:101) α > α, β of Q satisfy − < β < α > . If (cid:104) α, β (cid:105) intersects the y -axis above i , then a simple geometric argument shows that (cid:104) α, β (cid:105) is closer than S to the point z = i . If not, then the geodesic (cid:104) α − , β − (cid:105) associated to Q ( x + y, y ) crosses the y -axis above i (since ( α − β − ≤ − (cid:104) α − , β − (cid:105) or (cid:104) α, β (cid:105) is closer than S to the point z = i .In case (ii) we may assume that either Q = (cid:101) Q is reduced or that0 < (cid:101) β < and (cid:101) α > , in which case Q ( x, y ) = − (cid:101) Q ( x, x − y ) is reduced. Since this is equivalent to the identity (cid:101) Q ( x, y ) = − Q ( x, x − y ) , we are done. (cid:3) ARKOV SPECTRA FOR MODULAR BILLIARDS 11 Billiards and Sequences
Simple continued fractions are crucial in Markov’s proof of Theorem 1 and in our proof ofTheorem 3. We denote one by[ k , k , k , k , . . . ] = k + 1 k + 1 k + 1 k + . . ., whose finite version ending in k n is written [ k , k , . . . , k n ] . Here we will present the beginningsof Markov’s method in a form that we adapt in the next section to prove Theorem 3. Themethod relates chains of reduced quadratic forms to doubly-infinite sequences of positiveintegers. We give a somewhat novel treatment of this correspondence based on equivalencerather than proper equivalence.For a reduced Q define the doubly infinite sequence K Q by expanding(6.1) α Q = [ k , k , k , . . . ] and − β Q = [0 , k , k − , k − , . . . ]into simple continued fractions and setting K Q = ( . . . , k − , k , k , k , . . . ) . We shall refer to k as the first entry of K .Say two doubly infinite sequences of positive integers K = ( k n ) and L = ( (cid:96) n ) are equivalent if there is a j ∈ Z such that k n = (cid:96) n + j for all n ∈ Z and properly equivalent if there is a j ∈ Z such that k n = (cid:96) n + j for all n ∈ Z . If K = ( k n ) define the reversal of K to be K ∗ = ( k − n ). Proposition 6.
The map Q (cid:55)→ K Q determines a bijection between classes of forms with afixed positive discriminant that do not represent zero and equivalence classes of sequences ofpositive integers. It also determines a bijection between proper equivalence classes of formswith a fixed positive discriminant that do not represent zero and proper equivalence classesof sequences of positive integers. Furthermore K ∗ Q = K Q ∗ . Proof.
For each n ∈ Z let(6.2) r n = [ k n , k n +1 , . . . ] and s n = [0 , k n − , k n − , . . . ] . Thus r = α Q and s = − β Q and also(6.3) r n − = 1 r n + k n − and s n − = 1 s n − k n − . Define a n , b n > n ∈ Z by(6.4) a n √ d = 1 r n − + s n − and b n √ d = r n − s n r n + s n . It can be seen that using (6.3) that d = disc( Q ) = b n + 4 a n a n +1 for all n ∈ Z . Let Q n = (cid:0) a n +1 , − b n , − a n (cid:1) . A calculation shows that α Q n = r n and − β Q n = s n . It follows that each Q n is reduced and equivalent to Q . We claim that every such form occursas a Q n . Further, each Q n +1 is properly equivalent to Q and every reduced form that isproperly equivalent to Q is one of the Q n +1 . Again, these statements follow from variationson the arguments given in Chaper VII of [10]. That the claimed bijections are well-definedand injective follows. Clearly every sequence K arises from some reduced form so the mapsare also surjective.Turning to the last statement, recall that Q ∗ was defined in (5.1) and observe that β Q ∗ = − α Q and α Q ∗ = − β Q . Let r ∗ n and s ∗ n correspond to Q ∗ as in (6.2). By (6.1) we have r ∗ = − β Q = [ k , k − , k − , k − , . . . ] and s ∗ = α Q = [0 , k , k , k . . . ] , giving the result.This completes the proof of Proposition 6. (cid:3) Say that a sequence K = ( k n ) is periodic if there is an N ∈ Z + so that k n + N = k n for all n ∈ Z and palindromic if K ∗ is equivalent to K . Combining Propositions 2 and 6 we derivethe following correspondence. Theorem 5.
There is a bijection between proper modular billiards and equivalence classes ofdoubly infinite sequences of positive integers. Under this correspondence a billiard is periodicprecisely when the sequence is periodic and non-orientable precisely when the correspondingsequence is palindromic.
From the first formula of (6.4) and (5.2) we have(6.5) λ ∞ ( B ) = sup n ∈ Z ( r n + s n ) . This is the starting point of the proof of Markov’s Theorem 1. The main difficulty is inunderstanding which K cannot have any small values of r n + s n . The first observation isthat if any k m > λ ∞ ( B ) >
3. The complete result requires an ingeniousanalysis of continued fractions all of whose partial quotients are either 1 or 2. A treatment ofMarkov’s method and a proof of Theorem 1 based on it can be found in Dickson’s book [11].Other useful references are [1], [4], [5] and [6].7.
Sequences and the spectrum
To return to the proof of Theorem 3, recall from (1.4) that M i is defined in terms of λ i ( B ),which was given in (1.3). We now find a formula for λ i ( B ) that is analogous to (6.5) when B is a proper modular billiard. For our problem we are led to estimate certain quantitiesinvolving pairs of successive values of r n and s n from (6.2), rather than simply r n + s n . Let K be a doubly-infinite sequence of positive integers. The quantities we need are thefollowing: µ (cid:48) n ( K ) = (cid:12)(cid:12)(cid:12)(cid:12) r n + s n − r n − + s n − (cid:12)(cid:12)(cid:12)(cid:12) (7.1) µ (cid:48)(cid:48) n ( K ) = (cid:12)(cid:12)(cid:12)(cid:12) − r n + s n r n + s n − r n − + s n − (cid:12)(cid:12)(cid:12)(cid:12) (7.2) µ (cid:48)(cid:48)(cid:48) n ( K ) = (cid:12)(cid:12)(cid:12)(cid:12) r n − + s n − + r n − s n − r n + s n (cid:12)(cid:12)(cid:12)(cid:12) . (7.3) ARKOV SPECTRA FOR MODULAR BILLIARDS 13
Also set(7.4) µ (cid:48) ( K ) = inf n ∈ Z µ (cid:48) n ( K ) , µ (cid:48)(cid:48) ( K ) = inf n ∈ Z µ (cid:48)(cid:48) n ( K ) , µ (cid:48)(cid:48)(cid:48) ( K ) = inf n ∈ Z µ (cid:48)(cid:48)(cid:48) n ( K )and(7.5) µ ( K ) = min (cid:0) µ (cid:48) ( K ) , µ (cid:48)(cid:48) ( K ) , µ (cid:48)(cid:48)(cid:48) ( K ) , µ (cid:48) ( K ∗ ) , µ (cid:48)(cid:48) ( K ∗ ) , µ (cid:48)(cid:48)(cid:48) ( K ∗ ) (cid:1) . Proposition 7.
Let K correspond to a proper B . Then we have λ i ( B ) − = µ ( K ) . Proof.
By Proposition 6 and (6.4) µ (cid:48) n ( K ) = d − | a n +1 − a n | , µ (cid:48)(cid:48) n ( K ) = d − | a n +1 − b n − a n | , µ (cid:48)(cid:48)(cid:48) n ( K ) = d − | a n + b n − a n +1 | . The result follows from Proposition 5 since every reduced form of the class [ Q ] is foundamong the Q n . (cid:3) In the following proposition we show that { √ , √ , (3 + √ } ∈ M i , and that each value is attained. For j = 1 , , K = (1 , , K = (1 , , , , K = (3 , , , , . Here an overlined subsequence adjacent to a parenthesis indicates that one must concatenatethe subsequence infinitely many times in the direction of the parenthesis.
Proposition 8.
Let C j be the modular billiard associated to K j for j = 1 , , . Then (7.7) λ i ( C ) = √ , λ i ( C ) = √ , λ i ( C ) = (3 + √ . Each value λ i ( C j ) is attained. All three billiards are proper and non-orientable; C , C areperiodic, while C is not periodic.Proof. That (7.7) holds and that each λ i ( C j ) is attained is a straightforward application ofProposition 7. Clearly K , K are periodic and palindromic and K is not periodic but ispalindromic. Thus the final statement follows from Theorem 5. (cid:3) Exceptional sequences
In this and the next section we complete the proof of Theorem 3. By Proposition 4 wemay assume that the billiard B is proper. The main result of this section is the followingproposition, which (together with Proposition 7) shows that the only proper billiards thatstay farther away from z = i than C are C and C . The method is completely elementaryand amounts to finding inequalities determined by continued fractions. Proposition 9.
Unless K is equivalent to K j for j = 1 , , we have that µ ( K ) < (cid:16) √ − (cid:17) = 0 . . . . . We will say that any K with µ ( K ) ≥ . . . . is exceptional. The proof of Proposition 9consists of a series of results that successively eliminate configurations of subsequences in a K that force it to not be exceptional. Since µ ( K ) = µ ( K ∗ ), it is clearly permissible to onlyprove it for either K or K ∗ . Hence we will often only provide estimates for one of them andmight not mention when reversals must also be considered in order to cover all cases. Proposition 10.
An exceptional K must have the form K = ( . . . , , m , , m , , . . . ) where m j ≥ . If any m j = 2 then K is equivalent to K = (1 , , , . Proposition 10 will be proven in the four lemmas that follow.
Lemma 1.
Given any K , if K contains any subsequence of the form (8.1) (1 , , (2 , , (2 , , (2 , , (2 , , (2 , , ( m, m (cid:48) ) for ≤ m ≤ m (cid:48) then µ (cid:48) ( K ) ≤ and therefore K is not exceptional.Proof. If K contains the subsequence ( m, m (cid:48) ) for any m, m (cid:48) ∈ Z + there is an n ∈ Z so that r n − = [ m, m (cid:48) , . . . ] , r n = [ m (cid:48) , . . . ] , s n − = [0 , . . . ] , s n = [0 , m, . . . ] . Thus [ m, m (cid:48) , < r n − < [ m, m (cid:48) ] , < s n − < , [ m (cid:48) ] < r n < [ m (cid:48) , , [0 , m, < s n < [0 , m ]and so it follows that mmm (cid:48) + m + 1 < r n + s n < m + 1 mm (cid:48) + m (cid:48) + 1and m (cid:48) mm (cid:48) + m (cid:48) + 1 < r n − + s n − < m (cid:48) + 1 mm (cid:48) + m + 1 . The result now follows from (7.1) and (7.4) since for the pairs in (8.1) it can be easily verifiedthat (cid:12)(cid:12)(cid:12)(cid:12) m − m (cid:48) − mm (cid:48) + m + 1 (cid:12)(cid:12)(cid:12)(cid:12) , (cid:12)(cid:12)(cid:12)(cid:12) m (cid:48) − m − mm (cid:48) + m (cid:48) + 1 (cid:12)(cid:12)(cid:12)(cid:12) ≤ . (cid:3) A useful consequence of Lemma 1 is that (1 ,
1) cannot occur as a subsequence in anexceptional K . We will in several places use this fact without further mention. Lemma 2. If K contains the sequence ( (cid:96) (cid:48) , , (cid:96) ) where (cid:96) (cid:54) = 1 or (cid:96) (cid:48) (cid:54) = 1 then K is notexceptional.Proof. By Lemma 1 we may assume that either (i) (cid:96), (cid:96) (cid:48) ≥ (cid:96) (cid:48) = 1 and (cid:96) ≥ . Incase (i) there is an n ∈ Z so that r n − = [ (cid:96) (cid:48) , , (cid:96), . . . ] , r n = [2 , (cid:96), . . . ] , s n − = [0 , . . . ] , s n = [0 , (cid:96) (cid:48) , . . . ] , hence r n − > [7 , ,
7] = , < s n − < , < r n < [2 ,
7] = , < s n < [0 ,
7] = . It follows that 0 < r n − + s n − < and − < − r n + s n r n + s n < so that by (7.2) we have that µ (cid:48)(cid:48) n ( K ) < = 0 . . . . . In case (ii) either (a) K contains ( m (cid:48) , , m, , , (cid:96) ) where m, m (cid:48) ≥ K contains( m, , , , (cid:96) ) where m ≥ n ∈ Z so that r n − = [ m, , , (cid:96), . . . ] , r n = [1 , , (cid:96), . . . ] , s n − = [0 , , m (cid:48) , . . . ] , s n = [0 , m, , m (cid:48) , . . . ] . ARKOV SPECTRA FOR MODULAR BILLIARDS 15
Thus m + < r n − < m + , < s n − < , < r n < , m +1 < s n < m + . Hence ( m + 3)(3 m + 2)3( m + 1)(3 m + 4) < − r n + s n r n + s n < ( m + 1)(24 m + 61)(3 m + 2)(22 m + 37)and 1 m + < r n − + s n − < m + . It now follows easily that µ (cid:48)(cid:48) n ( K ) < = 0 . . . . . The same kind of computation shows that in case (b) we have < r n − < , < s n − < , < r n < , < s n < and so µ (cid:48)(cid:48) n ( K ) < = 0 . . . . . By (7.4) and (7.5) the result follows. (cid:3)
Lemmas 1 and 2 prove the first statement of Proposition 10.
Lemma 3.
Suppose that K contains the subsequence (1 , , , m ) . If m (cid:54) = 6 then K is notexceptional.Proof. Suppose that 1 < m < . By Lemmas 1 and 2 we may assume that for some n ∈ Z r n − = [2 , , m, , m (cid:48) , . . . ] , s n − = [0 , , m (cid:48)(cid:48) , . . . ] , r n = [1 , m, , m (cid:48) , . . . ] and s n = [0 , , , m (cid:48)(cid:48) , . . . ] , where m (cid:48) , m (cid:48)(cid:48) ≥
2. It follows that[2 , , m, , < r n − < [2 , , m,
1] and < s n − < , while [1 , m, < r n < [1 , m, ,
2] and [0 , , < s n < [0 , , , . Hence m + 24 m + 7 < r n − + s n − < m + 1533 m + 46 and 8(3 m + 2)33 m + 46 < r n + s n < m + 1)4 m + 7 . A calculation now shows that for m ≤ (cid:12)(cid:12)(cid:12)(cid:12) r n + s n − r n − + s n − (cid:12)(cid:12)(cid:12)(cid:12) < m + 14 m + 7 <
923 = 0 . . . . and the statement of the Lemma in this case follows by (7.1), (7.4) and (7.5).Now assume that m ≥ . For this we will apply (7.3) to the reversed sequence ( m, , , r n − = [ m, , , , m (cid:48) , , . . . ] , s n − = [0 , , m (cid:48)(cid:48) , . . . ] , r n = [1 , , , m (cid:48) , . . . ] , s n = [0 , m, , m (cid:48)(cid:48) , . . . ]where m (cid:48) ≥ m (cid:48)(cid:48) ≥
2. Then a calculation using r n − > [ m, , , , , [1 , , < r n < [1 , , , , s n − > [0 , , , [0 , m, < s n < [0 , m, , < r n − + s n − < m + and 3 m − m + 6 < r n − s n − r n + s n < m − m + 7) . Thus we have (cid:12)(cid:12)(cid:12)(cid:12) r n − + s n − + r n − s n − r n + s n (cid:12)(cid:12)(cid:12)(cid:12) < m + 2951 m + 4407)17(4 m + 7)(69 m + 97) ≤ . . . . for m ≥ . Suppose now that K contains (1 , , , K contains(1 , , , , , ,
1) or (1 , , , , , , K contains (1 , , , , , ,
1) we have for some n that[1 , , , , < r n − < [1 , , , , , , [0 , , < s n − < [0 , , , , , , , < r n < [2 , , , , [0 , , , , < s n < [0 , , , . This gives 0 . · · · = < r n − + s n − < = 0 . . . . and0 . · · · = < r n + s n < = 0 . . . . and hence µ (cid:48) n ( K ) < = 0 . . . . . Similarly, if K contains (1 , , , , , ,
1) we have for some n the same inequalities for s n − and s n while[1 , , , , < r n − < [1 , , , , ,
2] and [2 , , , , < r n < [2 , , , . This gives 0 . · · · = < r n − + s n − < = 0 . . . . and0 . · · · = < r n + s n < = 0 . . . . . This gives µ (cid:48) n ( K ) < = 0 . . . . . Hence by (7.4) and (7.5) we are done. (cid:3)
If we now assume that K contains (1 , , ,
6) then in fact we may assume that K contains(1 , , , , , m,
1) where m ≥ . Lemma 4.
Suppose that K contains (1 , , , , , m, where m > . Then K is not excep-tional.Proof. We may assume that for some n [6 , , m, , < r n − , [0 , , < s n − and [1 , m, < r n < [1 , m, , , [0 , , , , < s n < [0 , , , . Thus we have that 0 < r n − + s n − < m + 5)69 m + 106 ARKOV SPECTRA FOR MODULAR BILLIARDS 17 and − m + 45 m + 34( m + 1)(69 m + 106) < r n − s n − r n + s n < − ( m + 1)(12 m − m + 2)(31 m + 58) . Hence (cid:12)(cid:12)(cid:12)(cid:12) r n − + s n − + r n − s n − r n + s n (cid:12)(cid:12)(cid:12)(cid:12) < m + 9975 m + 20671 m + 11218(3 m + 2)(31 m + 58)(69 m + 106) ≤ . . . . . for m ≥ . (cid:3) This gives the second statement and thus completes the proof of Proposition 10.Suppose now that K = ( . . . , , m , , m , , . . . ) is exceptional with K (cid:54) = K and K (cid:54) = K .By Proposition 10 and (7.6) we have that m j ≥ j with at least one m j >
3. Nowfor K given in (7.6) we have that µ (cid:48)(cid:48)(cid:48) n ( K ) = (cid:12)(cid:12)(cid:12)(cid:12) r n − + s n − + r n − s n − r n + s n (cid:12)(cid:12)(cid:12)(cid:12) = 14 (cid:16) √ − (cid:17) , when r n = [1 , , , , , . . . ] and s n = [0 , , , , , , . . . ] . Since µ ( K ) ≤ µ (cid:48)(cid:48)(cid:48) n ( K ) by (7.4) and (7.5), Proposition 9 is a consequence of the followinglemma, which implies that µ (cid:48)(cid:48)(cid:48) n ( K ) < µ (cid:48)(cid:48)(cid:48) n ( K )unless K is equivalent to K . Lemma 5.
Suppose that r n = [1 , m , , m , , . . . ] , s n = [0 , m , , m − , , . . . ] and r (cid:48) n = [1 , m (cid:48) , , m (cid:48) , , . . . ] , s (cid:48) n = [0 , m (cid:48) , , m (cid:48)− , , . . . ] with m (cid:48) j ≥ m j ≥ for all j ∈ Z . Then (cid:12)(cid:12)(cid:12)(cid:12) r (cid:48) n − + s (cid:48) n − + r (cid:48) n − s (cid:48) n − r (cid:48) n + s (cid:48) n (cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:12)(cid:12)(cid:12)(cid:12) r n − + s n − + r n − s n − r n + s n (cid:12)(cid:12)(cid:12)(cid:12) , with equality if and only if m (cid:48) j = m j for all j ∈ Z . Proof.
Under our assumptions we have r n − ≤ r (cid:48) n − and r n ≥ r (cid:48) n . Now s n − = s n − m and r n − = r n + m and similarly for r (cid:48) , s (cid:48) . Thus2 r n − + s n − + r n − s n − r n + s n = 2 r n s n + r n − s n − r n + s n . The result now follows since the function F ( x, y ) = 2 xy + x − y − x + y satisfies 0 < F ( x , y ) ≤ F ( x , y ) whenever 1 < x ≤ x < < y ≤ y <
1, withequality if and only if x = x and y = y . To see this, use that the gradient of F ( x, y ) isgiven by ∇ F ( x, y ) = (cid:16) y ( y +1)+1( x + y ) , x − x +1( x + y ) (cid:17) . (cid:3) Approximating C We conclude by justifying the final statement of Theorem 3. For (cid:96) ≥
1, let K (cid:96) = (3 , , , , , , . . . , , , , , , k = k (cid:96) = 4 , where the number of 1’s between the 4’s is given by (cid:96) . Propositions 7 and 9 show that thebilliard B (cid:96) associated to K (cid:96) satisfies λ i ( B (cid:96) ) > (3 + √ Proposition 11.
Let B (cid:96) be the billiard associated to the class of K (cid:96) by Theorem 5. Then lim (cid:96) →∞ λ i ( B (cid:96) ) = (3 + √ . Proof.
For a fixed (cid:96) , define r n = r n ( (cid:96) ) and s n = s n ( (cid:96) ) as in (6.2) for the sequence K (cid:96) . Then r = [1 , , , . . . , , , , ,
3] and s = [0 , , ,
3] = (5 − √ , where the bar indicates a repeated sequence, and the number of 1’s before the 4 in r is givenby (cid:96) . One can easily compute an explicit formula for r using the recurrence relation(9.1) r ( (cid:96) + 1) = 1 + 13 + r ( (cid:96) ) , r (1) = (7 − √ . Let ε = (5 + √
21) denote the fundamental unit in Q ( √ µ (cid:48)(cid:48)(cid:48) ( K (cid:96) ) = ( − √ ε (cid:96) − ε − (cid:96) (11 + √ ε (cid:96) − ε − (cid:96) . We claim that µ ( K (cid:96) ) = µ (cid:48)(cid:48)(cid:48) ( K (cid:96) ); the result then follows easily since λ i ( B (cid:96) ) = µ ( K (cid:96) ) − .It is straightforward to check that µ ( K (cid:96) ) = µ (cid:48)(cid:48)(cid:48) ( K (cid:96) ) for small (cid:96) , so we assume (cid:96) ≥
3. Toprove that µ ( K (cid:96) ) = µ (cid:48)(cid:48)(cid:48) ( K (cid:96) ), we first show that µ (cid:48) n ( K (cid:96) ) , µ (cid:48)(cid:48) n ( K (cid:96) ) > − √ = . . . . for all n and that µ (cid:48)(cid:48)(cid:48) n ( K (cid:96) ) > − √ if n (cid:54) = ± , (cid:96) ±
1. If n is odd, then √ = [1 , ≤ r n ≤ [1 ,
3] = √ , − √ = [0 , , ≤ s n ≤ [0 , ,
1] = − √ . If n is even and k n = 4, then √ = [4 , , , ≤ r n ≤ [4 ,
1] = 2 + 2 √ , − √ = [0 , , ≤ s n ≤ [0 , , , ,
4] = −√ , while if n is even and k n (cid:54) = 4, then √ = [3 , ≤ r n ≤ [3 , , ,
1] = 1 + 2 √ , − √ = [0 , , ≤ s n ≤ [0 , ,
4] = − √ , It follows from (7.1), (7.2), and (6.3) that in each case, µ (cid:48) n ( K (cid:96) ) , µ (cid:48)(cid:48) n ( K (cid:96) ) > . µ (cid:48)(cid:48)(cid:48) n ( K (cid:96) ) > .
399 if n is even. If n is odd and k n +1 , k n − (cid:54) = 4 then r n ≥ [1 , , ,
4] = √ , s n ≥ [0 , , , ,
1] = − √ , from which it follows that µ (cid:48)(cid:48)(cid:48) n ( K (cid:96) ) > . ARKOV SPECTRA FOR MODULAR BILLIARDS 19
It remains to show that µ − ( K (cid:96) ) > µ ( K (cid:96) ) since, by symmetry, we have µ (cid:96) +1 ( K (cid:96) ) >µ (cid:96) − ( K (cid:96) ) = µ ( K (cid:96) ). By (6.3) we have µ (cid:48)(cid:48)(cid:48)− ( K (cid:96) ) = µ (cid:48)(cid:48)(cid:48) ( K (cid:96) ) + 6(4 r s + s − r ) r + s . Thus µ − ( K (cid:96) ) > µ ( K (cid:96) ) if and only if r < s − s = (3 + √ (cid:3) References [1] Aigner, Martin Markov’s theorem and 100 years of the uniqueness conjecture. A mathematical journeyfrom irrational numbers to perfect matchings. Springer, Cham, 2013. x+257 pp.[2] E. Artin, Ein mechanisches System mit quasiergodischen Bahnen, Hamb. Math. Abh. 3 (1924), 170–177[3] A.F. Beardon, The geometry of discrete groups. Graduate Texts in Mathematics, 91. Springer-Verlag,New York, 1983. xii+337 pp.[4] J. Berstel; A. Lauve; C. Reutenauer; F. Saliola, Combinatorics on words. Christoffel words and repetitionsin words. CRM Monograph Series, 27. AMS, Providence, RI, 2009. xii+147 pp.[5] Bombieri, Enrico Continued fractions and the Markoff tree. Expo. Math. 25 (2007), no. 3, 187–213.[6] Cassels, J. W. S. An introduction to Diophantine approximation. Cambridge Tracts in Mathematics andMathematical Physics, No. 45. Cambridge University Press, New York, 1957. x+166 pp.[7] Cohn, Harvey, Approach to Markoff’s minimal forms through modular functions. Ann. of Math. (2) 61,(1955). 1–12.[8] Cohn, Harvey, Markoff geodesics in matrix theory. Number theory with an emphasis on the Markoffspectrum (Provo, UT, 1991), 69–82, Lecture Notes in Pure and Appl. Math., 147, Dekker, New York,1993.[9] Cusick, Thomas W.; Flahive, Mary E. The Markoff and Lagrange spectra. Mathematical Surveys andMonographs, 30. American Mathematical Society, Providence, RI, 1989. x+97 pp.[10] L.E. Dickson, Modern Elementary Theory of Numbers. University of Chicago Press, Chicago, 1939.vii+309 pp.[11] L.E. Dickson, Studies in the Theory of Numbers, Chicago University Press, 1930.[12] L.R. Ford, A geometrical proof of a theorem of Hurwitz, Proc. Edinburgh Math. Soc. 35 (1917), 59–65.[13] G.A. Freiman, Diofantovy priblizheniya i geometriya chisel (zadacha Markova). (Russian) [Diophantineapproximations and the geometry of numbers (Markov’s problem)] Kalinin. Gosudarstv. Univ., Kalinin,1975. 144 pp.[14] G. Frobenius, ¨Uber die Reduktion der indefiniten bin¨aren quadratischen Formen. (1913)[15] G. Frobenius, ¨Uber die Markoffschen Zahlen, Preuss. Akad. Wiss. Sitzungberichte (1913) 458–487. (Alsoin: G. Frobenius, Gesammelte Abhandlungen, Bd. 3, Springer, Berlin, Heidelberg, New York, 1968, pp.598–627).[16] M. Hall, Jr., On the sum and product of continued fractions. Ann. of Math. (2) 48, (1947). 966–993.[17] M. Hall, Jr., The Markoff spectrum. Acta Arith. 18 (1971), 387–399.[18] A.V. Malyshev, Markov and Lagrange spectra [Survey of the literature], Zap. Nauch. Sem. Lenin. Otd.Math. Inst. V.A. Steklova AN SSSR 67 (1977) 5–38. (English translation in J. Soviet Math. 16 (1981)767–788).[19] A. Markoff, Sur les formes quadratiques binaires ind´efinies, Math. Ann. 15 (1879) 381–409, 17 (1880)379–399.[20] I. Schur, Zur Theorie der indefiniten bin¨aren quadratischen Formen, Sitzungberichte der PreussischenAkademie Wiss. 1913, 212–231.
UCLA Mathematics Department, Box 951555, Los Angeles, CA 90095-1555
E-mail address : [email protected] UCLA Mathematics Department, Box 951555, Los Angeles, CA 90095-1555
E-mail address ::