A question worth asking
What do the derivatives of the tangent function, a family of orthogonal polynomials, and a continued fraction have to do with one another? A three-term recurrence gives a surprisingly economical answer. Fix η > 0 \eta> 0 η > 0 and δ ∈ R \delta\in\mathbb{R} δ ∈ R , and define monic polynomials by
M n + 1 ( x ; δ , η ) = [ x − ( 2 n + η ) δ ] M n ( x ; δ , η ) − n ( n + η − 1 ) ( 1 + δ 2 ) M n − 1 ( x ; δ , η ) , (1) M_{n+1}(x;\delta,\eta) = [x - (2n+\eta)\delta]M_n(x;\delta,\eta) - n(n+\eta-1)(1+\delta^2)M_{n-1}(x;\delta,\eta), \tag*{(1)} M n + 1 ( x ; δ , η ) = [ x − ( 2 n + η ) δ ] M n ( x ; δ , η ) − n ( n + η − 1 ) ( 1 + δ 2 ) M n − 1 ( x ; δ , η ) , (1)
with M − 1 = 0 M_{-1}=0 M − 1 = 0 and M 0 = 1 M_0=1 M 0 = 1 . The first few are
M 0 = 1 , M 1 = x − η δ , M 2 = x 2 − 2 ( η + 1 ) x δ + η ( η + 1 ) δ 2 − η . \begin{aligned}
M_0 &= 1,\qquad M_1 = x-\eta\delta,\\
M_2 &= x^2-2(\eta+1)x\delta+\eta(\eta+1)\delta^2-\eta.
\end{aligned} M 0 M 2 = 1 , M 1 = x − η δ , = x 2 − 2 ( η + 1 ) x δ + η ( η + 1 ) δ 2 − η .
These are the Meixner polynomials of the second kind, more commonly encountered today in the Meixner–Pollaczek normalization. Already their name conceals a small history. Their recurrence conceals rather more.
We shall find that their normalized moments have the exponential generating function
H δ , η ( u ) = ( cos u − δ sin u ) − η . (2) H_{\delta,\eta}(u)=(\cos u-\delta\sin u)^{-\eta}. \tag*{(2)} H δ , η ( u ) = ( cos u − δ sin u ) − η . (2)
At δ = 0 \delta=0 δ = 0 , η = 1 \eta=1 η = 1 , the moments are the secant numbers. At δ = 0 \delta=0 δ = 0 , η = 2 \eta=2 η = 2 , they are the derivatives of tan u \tan u tan u , shifted by one place. Thus classical counting sequences arise from a perfectly ordinary positive measure. We will see the measure, prove its orthogonality, and understand why its continued fraction has such simple coefficients.
Then we will turn the same recurrence sideways. The Meixner-II are polynomials both in δ \delta δ and in x x x . What becomes of orthogonality when the parameter becomes the variable? The coefficient 1 + δ 2 1+\delta^2 1 + δ 2 hints that the answer should involve rational functions. It does: the second family is the complementary Romanovski–Routh family, and its continued fraction belongs to Wall’s theory. Better still, both measures come from one explicit positive kernel. With the correct normalization, they are the two conditional laws of a joint probability measure with Cauchy marginals.
The purpose of this paper is to make these connections understandable without requiring the reader to learn several theories before seeing what they explain. The order is deliberate. We begin with the classification problem that led Meixner to the polynomials. Generating functions then supply the polynomials and their moments together. Fourier inversion produces the measure, and a product of two generating functions proves orthogonality. Carlitz’s work provides a second point of entry: start with familiar numbers, treat them as a moment functional, and recover the polynomials. Paths make the coefficient arrays visible; finite Jacobi matrices make continued-fraction convergence precise. Only then do we need the second spectral variable.
The second measure will, in turn, raise two questions of its own. Does the joint law already exist before the continued fractions are allowed to go to infinity? And what makes two Cauchy variables with fixed marginals depend on one another? Finite matrices answer the first question exactly; a gamma law at the edge answers the second. The same Laguerre polynomials describe both the directions of the zero curves and the limiting fluctuations. Two appendices pursue the corresponding operator and zero geometry in greater detail.
Most of the individual ingredients are classical. We distinguish their sources from the direct derivations that assemble them here; in particular, we make no novelty claim for the CRR identification or the general Wall correspondence. The point is to see how much of the picture follows from a few calculations, and where a change of viewpoint reveals something hidden.
We use ( a ) n = a ( a + 1 ) ⋯ ( a + n − 1 ) (a)_n = a(a+1)\cdots(a+n-1) ( a ) n = a ( a + 1 ) ⋯ ( a + n − 1 ) , with ( a ) 0 = 1 (a)_0 = 1 ( a ) 0 = 1 , and write d x dx d x for integration with respect to Lebesgue measure. All formal power-series identities are explicitly distinguished from analytic identities. A polynomial sequence is called an orthogonal polynomial sequence when its members have successive degrees, are pairwise orthogonal, and have nonzero squared norms for the indicated moment functional; a positive orthogonality measure is stated separately. These distinctions will matter in the discussion of Carlitz.
Meixner and the lives of his polynomials
An eponym can make a mathematical object seem to have had a single birth. The Meixner polynomials have a more interesting history. They arise from classification, from limits, from moment problems, and from the search for useful formulas. These approaches explain different features of the same objects. Keeping them in view will help us understand why changing the variable in a recurrence can reveal something that its original name hides.
A life in mathematical physics
Josef Meixner (1908–1994) was trained in Munich, where he studied mathematics, physics, and physical chemistry and completed his doctorate under Arnold Sommerfeld in 1931. After assistantships in Munich and Giessen, he habilitated in 1936/37. His appointment to Aachen in 1942 was interrupted by military service; he began work there after the war, became a regular chairholder in 1951, and retired in 1974. Butzer and Koornwinder give a documented account of his career [4 §2] .
That account also records Meixner’s membership in the SA from 1934 and the NSDAP from 1937. His postwar employment required a denazification certificate, and Sommerfeld supplied a supporting letter in December 1945 [4 §2] .
The mathematical range of his work is best seen through the problems he chose. His dissertation concerned the Green function of the wave-mechanical Kepler problem [26 ] . A Green function packages solutions of an equation into a kernel: the choice of solutions, their normalization, and their behavior at the endpoints all matter at once. This is a natural meeting place for mathematical physics and the theory of special functions.
Meixner’s later publications continued that conversation. His book with Friedrich Wilhelm Schäfke, published in 1954, develops Mathieu and spheroidal functions and their physical and technical applications [28 ] . The organization is revealing: foundations, the two families of functions, then applications. Explicit formulas and structural understanding are meant to support each other.
Another substantial part of his work concerned irreversible processes. The 1959 Handbuch der Physik chapter with H. G. Reik develops a macroscopic thermodynamics extending the equilibrium theory to evolving continuous media [27 ] . Its scope includes fluid motion, elasticity, and media in electromagnetic fields. Here the unifying question is how different physical processes fit into a common mathematical description. With Fritz Bopp, Meixner also edited Sommerfeld’s Thermodynamik und Statistik , published in 1952 [34 ] . His scientific life thus extends well beyond the eight-page paper that attached his name to two polynomial families.
The question behind the classification
That paper, received in May 1933 and published in 1934, begins with a question one can ask before knowing any special-function terminology. Suppose an exponential generating function has the form A ( t ) e x U ( t ) A(t)e^{xU(t)} A ( t ) e xU ( t ) , with A ( 0 ) = 1 A(0)=1 A ( 0 ) = 1 , U ( 0 ) = 0 U(0)=0 U ( 0 ) = 0 , and U ′ ( 0 ) = 1 U'(0)=1 U ′ ( 0 ) = 1 . When can its polynomials also be orthogonal for a positive measure on the real line? Hermite and Laguerre polynomials provide affirmative examples. Bernoulli and Euler polynomials have equally attractive generating functions, but fail the orthogonality requirement. The resemblance of their formulas is therefore a reason to ask the question, not an answer to it [25 §1] .
The starting point was the simpler case U ( t ) = t U(t)=t U ( t ) = t , now called the Appell case. Differentiation then gives p n ′ = n p n − 1 p_n'=np_{n-1} p n ′ = n p n − 1 : every polynomial lowers to the previous one. Meixner explicitly credits Salomon Bochner with participation in the result for this special case, where only the Hermite family survives. His acknowledgment should not be enlarged into a joint authorship of the full classification [25 p. 7, footnote] .
Allowing U U U to vary changes the lowering operator, while orthogonality still demands a three-term recurrence. Meixner’s decisive step is to make these two requirements speak to each other. The resulting restrictions reduce the problem to a quadratic polynomial and its root configurations. The answer consists, up to the usual changes of scale and origin, of Hermite, Laguerre, Charlier, discrete Meixner, and the continuous family now called Meixner–Pollaczek [25 §§4–6] .
This is a classification of infinite positive orthogonal systems : Meixner explicitly assumes a measure with infinitely many points of increase. Finite Krawtchouk systems require a change of the question. They are not a counterexample to the theorem as he stated it. Nor are the two families bearing his name both discrete: their measures live on a lattice and on the real line, respectively.
Meixner also recovered the moment transform by integrating the generating function and using orthogonality to the constant polynomial. Inverting the inner series led to Fourier or Laplace inversion and the measure itself [25 §6] . We carry out this calculation in § 4 –§ 5 . The classification tells us when the encoding in the generating function is compatible with positivity.
Discovery, rediscovery, and a neighboring family
The discrete family had appeared in Ladislav Truksa’s 1931 work under the name generalized Kummer polynomials [37 Part III, §6, pp. 198–199] . Truksa approached them through limiting transitions in hypergeometric orthogonal systems. He gave their weight, hypergeometric and Rodrigues formulas, recurrence, and difference equation. His viewpoint makes the connection with probability particularly visible: the weight is a generalized Poisson law, the negative binomial distribution in present terminology. Meixner’s later classification supplies a different explanation for their existence. One construction locates them among limiting families; the other shows that they are forced by the coexistence of a generating function and orthogonality.
The continuous family acquired another life in Félix Pollaczek’s 1950 paper, whose title advertises Hermite and Laguerre polynomials as limiting cases [30 ] . Pollaczek obtained it by a limiting passage from his broader family of orthogonal polynomials. The name Meixner–Pollaczek preserves both routes. For our purposes the distinction between the continuous and discrete measures matters more than the shared part of their names: analytic continuation of formulas does not automatically continue positivity of a measure.
The generating functions themselves also received a new vocabulary. Sheffer’s 1939 paper develops polynomial sets of type zero [32 ] ; today we speak of Sheffer sequences. The generating-function and lowering-operator descriptions needed for Meixner’s classification were already present in the 1934 paper. The useful modern distinction is between the broad algebraic class of Sheffer sequences and its much smaller positive orthogonal subclass. It is the intersection that makes Meixner’s theorem so restrictive.
Finally, Bateman’s 1933 polynomials belong to a neighboring story [1 ] . Their route through Legendre polynomials and Fourier transforms leads to the continuous Hahn family. Koelink traces this development through Pasternack, Touchard, and Carlitz [21 §1] . This distinction will be essential when we read Carlitz: the Bernoulli and Euler sides of his paper need not land in the same family just because their moment formulas are similar. History here provides a guide for the mathematics: compare the recurrence and its measure before identifying the polynomials by name.
Why only five families?
Meixner’s question can be expressed in one line. Which orthogonal polynomial sequences have an exponential generating function
G ( x ; t ) = f ( t ) e x g ( t ) = ∑ n ≥ 0 p n ( x ) t n n ! , f ( 0 ) = 1 , g ( 0 ) = 0 , g ′ ( 0 ) = 1 ? (3) G(x;t) = f(t)e^{xg(t)} = \sum_{n\ge0} p_n(x)\frac{t^n}{n!}, \qquad f(0) = 1,\quad g(0) = 0,\quad g'(0) = 1? \tag*{(3)} G ( x ; t ) = f ( t ) e xg ( t ) = n ≥ 0 ∑ p n ( x ) n ! t n , f ( 0 ) = 1 , g ( 0 ) = 0 , g ′ ( 0 ) = 1 ? (3)
The normalization makes the polynomials monic. Such sequences are now called Sheffer sequences . The exponential in (3) allows differentiation in x x x to be replaced by multiplication by g ( t ) g(t) g ( t ) . Orthogonality, on the other hand, asks multiplication by x x x to have only three diagonals in the polynomial basis. The classification comes from making these two requirements agree.
Indeed, differentiation with respect to t t t gives
x G = A ( t ) G t + B ( t ) G , A ( t ) = 1 g ′ ( t ) , B ( t ) = − f ′ ( t ) f ( t ) g ′ ( t ) . (4) xG = A(t)G_t + B(t)G,\qquad A(t) = \frac{1}{g'(t)},\qquad B(t) = -\frac{f'(t)}{f(t)g'(t)}. \tag*{(4)} x G = A ( t ) G t + B ( t ) G , A ( t ) = g ′ ( t ) 1 , B ( t ) = − f ( t ) g ′ ( t ) f ′ ( t ) . (4)
Write A ( t ) = ∑ A j t j A(t) = \sum A_jt^j A ( t ) = ∑ A j t j , B ( t ) = ∑ B j t j B(t) = \sum B_jt^j B ( t ) = ∑ B j t j , with A 0 = 1 A_0 = 1 A 0 = 1 . For j ≥ 2 j \ge2 j ≥ 2 , the coefficient of p n − j p_{n-j} p n − j in x p n xp_n x p n is
n ! ( n − j ) ! [ ( n − j ) A j + 1 + B j ] . \frac{n!}{(n-j)!}\left[(n-j)A_{j+1}+B_j\right]. ( n − j )! n ! [ ( n − j ) A j + 1 + B j ] .
A three-term recurrence requires this to vanish for every n ≥ j n \ge j n ≥ j . Taking n = j n = j n = j and then n = j + 1 n = j+1 n = j + 1 gives B j = A j + 1 = 0 B_j = A_{j+1} = 0 B j = A j + 1 = 0 . Thus necessarily
A ( t ) = 1 + a t + b t 2 , B ( t ) = α + β t . (5) A(t) = 1 + at + bt^2,\qquad B(t) = \alpha+ \beta t. \tag*{(5)} A ( t ) = 1 + a t + b t 2 , B ( t ) = α + β t . (5)
Conversely, these forms give exactly
x p n = p n + 1 + ( a n + α ) p n + n [ b ( n − 1 ) + β ] p n − 1 . (6) xp_n = p_{n+1} + (an+\alpha)p_n + n\left[b(n-1)+\beta\right]p_{n-1}. \tag*{(6)} x p n = p n + 1 + ( an + α ) p n + n [ b ( n − 1 ) + β ] p n − 1 . (6)
For a positive measure on the real line the coefficients a a a , b b b , α \alpha α , β \beta β are real. For an infinite positive orthogonality measure, the last coefficient must be positive for every n ≥ 1 n \ge1 n ≥ 1 . This is equivalent to β > 0 \beta> 0 β > 0 and b ≥ 0 b \ge0 b ≥ 0 . Favard’s theorem gives the converse [6 ] .
We have reduced the problem to a quadratic polynomial. Its possibilities, up to affine changes of the variable and harmless normalizations, are
Table 1.
This is the substance of Meixner’s five-family classification [25 ] . It is a classification of infinite positive systems; finite systems require separate endpoint conditions.
The case of interest to us has
a = 2 δ , b = 1 + δ 2 , α = η δ , β = η ( 1 + δ 2 ) . a = 2\delta,\qquad b = 1 + \delta^2,\qquad\alpha= \eta\delta,\qquad\beta= \eta(1 + \delta^2). a = 2 δ , b = 1 + δ 2 , α = η δ , β = η ( 1 + δ 2 ) .
Its discriminant is always − 4 -4 − 4 . More generally, if a 2 < 4 b a^2 < 4b a 2 < 4 b , put
ω = b − a 2 / 4 , δ = a 2 ω , η = β b . \omega= \sqrt{b-a^2/4},\qquad\delta= \frac{a}{2\omega},\qquad\eta= \frac{\beta}{b}. ω = b − a 2 /4 , δ = 2 ω a , η = b β .
The monic rescaling
q n ( X ) = ω − n p n ( ω X + α − η a / 2 ) q_n(X) = \omega^{-n}p_n(\omega X + \alpha- \eta a/2) q n ( X ) = ω − n p n ( ω X + α − η a /2 )
reduces (6) to (1) in X X X . The quadratic factor 1 + δ 2 1 + \delta^2 1 + δ 2 is therefore a consequence of a particularly convenient normalization of the negative-discriminant case. It is also the factor that will later turn the parameter recurrence into an R I I R_{\mathrm{II}} R II recurrence.
The calculation above is worth remembering. A rather grand classification theorem begins with the observation that a linear function of n n n cannot vanish at every integer unless both of its coefficients vanish. The special functions enter only after this restriction has done its work.
One recurrence, two generating functions
The polynomials
Let
G ( x ; t ) = ∑ n ≥ 0 M n ( x ; δ , η ) t n n ! , D ( t ) = 1 + 2 δ t + ( 1 + δ 2 ) t 2 . G(x;t) = \sum_{n\ge0} M_n(x;\delta,\eta)\frac{t^n}{n!},\qquad D(t) = 1 + 2\delta t + (1+\delta^2)t^2. G ( x ; t ) = n ≥ 0 ∑ M n ( x ; δ , η ) n ! t n , D ( t ) = 1 + 2 δ t + ( 1 + δ 2 ) t 2 .
Multiply (1) by t n / n ! t^n/n! t n / n ! and sum. The only term whose index deserves a second look is
∑ n ≥ 1 n ( n + η − 1 ) M n − 1 t n n ! = t ∑ m ≥ 0 ( m + η ) M m t m m ! = t 2 G t + η t G . \sum_{n\ge1} n(n+\eta-1)M_{n-1}\frac{t^n}{n!}
= t\sum_{m\ge0}(m+\eta)M_m\frac{t^m}{m!}
= t^2G_t+\eta tG. n ≥ 1 ∑ n ( n + η − 1 ) M n − 1 n ! t n = t m ≥ 0 ∑ ( m + η ) M m m ! t m = t 2 G t + η tG .
Consequently,
D ( t ) G t = [ x − η δ − η ( 1 + δ 2 ) t ] G , G ( x ; 0 ) = 1. (7) D(t)G_t = [x-\eta\delta-\eta(1+\delta^2)t]G,\qquad G(x;0)=1. \tag*{(7)} D ( t ) G t = [ x − η δ − η ( 1 + δ 2 ) t ] G , G ( x ; 0 ) = 1. (7)
This is an ordinary first-order differential equation in t t t ; x x x is a parameter. Its logarithmic derivative separates immediately:
G t G = x D ( t ) − η D ′ ( t ) 2 D ( t ) . \frac{G_t}{G}=\frac{x}{D(t)}-\frac{\eta D'(t)}{2D(t)}. G G t = D ( t ) x − 2 D ( t ) η D ′ ( t ) .
We obtain
G ( x ; t ) = D ( t ) − η / 2 e x g ( t ) , g ( t ) = ∫ 0 t d v D ( v ) = arctan t 1 + δ t . (8) G(x;t)=D(t)^{-\eta/2}e^{xg(t)},\qquad g(t)=\int_0^t\frac{dv}{D(v)}=\arctan\frac{t}{1+\delta t}. \tag*{(8)} G ( x ; t ) = D ( t ) − η /2 e xg ( t ) , g ( t ) = ∫ 0 t D ( v ) d v = arctan 1 + δ t t . (8)
The last expression means the branch analytic at 0, with value 0 there. Equivalently, the two scalar equations in the Sheffer form G = f e x g G=f e^{xg} G = f e xg are
g ′ = D − 1 , f ′ f = − η D ′ ( t ) 2 D ( t ) . g'=D^{-1},\qquad\frac{f'}{f}=-\frac{\eta D'(t)}{2D(t)}. g ′ = D − 1 , f f ′ = − 2 D ( t ) η D ′ ( t ) .
Thus the generating function is not a guess. The recurrence has already specified its two factors.
Inversion, and what the word means
Because the M n M_n M n are monic, there are unique numbers q n , k q_{n,k} q n , k such that
x n = ∑ k = 0 n q n , k M k ( x ; δ , η ) . (9) x^n=\sum_{k=0}^{n}q_{n,k}M_k(x;\delta,\eta). \tag*{(9)} x n = k = 0 ∑ n q n , k M k ( x ; δ , η ) . (9)
The inverse polynomials are Q n ( y ) = ∑ k = 0 n q n , k y k Q_n(y)=\sum_{k=0}^{n}q_{n,k}y^k Q n ( y ) = ∑ k = 0 n q n , k y k . This is inversion of a triangular coefficient matrix, not inversion of a polynomial as a function.
The compositional inverse of g g g in (8) is
τ ( u ) = g − 1 ( u ) = tan u 1 − δ tan u = sin u cos u − δ sin u . (10) \tau(u)=g^{-1}(u)=\frac{\tan u}{1-\delta\tan u}=\frac{\sin u}{\cos u-\delta\sin u}. \tag*{(10)} τ ( u ) = g − 1 ( u ) = 1 − δ tan u tan u = cos u − δ sin u sin u . (10)
Substituting t = τ ( u ) t=\tau(u) t = τ ( u ) in G = f e x g G=f e^{xg} G = f e xg gives
e x u = 1 f ( τ ( u ) ) ∑ k ≥ 0 M k ( x ) τ ( u ) k k ! . e^{xu}=\frac{1}{f(\tau(u))}\sum_{k\geq0}M_k(x)\frac{\tau(u)^k}{k!}. e xu = f ( τ ( u )) 1 k ≥ 0 ∑ M k ( x ) k ! τ ( u ) k .
Comparison with (9) proves the Sheffer inversion formula in this case:
∑ n ≥ 0 Q n ( y ) u n n ! = H ( u ) e y τ ( u ) , H ( u ) = 1 f ( τ ( u ) ) = ( cos u − δ sin u ) − η . (11) \sum_{n\geq0}Q_n(y)\frac{u^n}{n!}=H(u)e^{y\tau(u)},\qquad H(u)=\frac{1}{f(\tau(u))}=(\cos u-\delta\sin u)^{-\eta}. \tag*{(11)} n ≥ 0 ∑ Q n ( y ) n ! u n = H ( u ) e y τ ( u ) , H ( u ) = f ( τ ( u )) 1 = ( cos u − δ sin u ) − η . (11)
We used the simple identity D ( τ ( u ) ) = ( cos u − δ sin u ) − 2 D(\tau(u))=(\cos u-\delta\sin u)^{-2} D ( τ ( u )) = ( cos u − δ sin u ) − 2 . In particular, the exponent in the prefactor of the inverse generating function is η \eta η , not η / 2 \eta/2 η /2 .
This is one useful meaning of umbral calculus : a linear map sends the monomial basis to a polynomial basis, and substitution and composition organize the map and its inverse. There is also a lowering operator. If D x = d / d x D_x=d/dx D x = d / d x , then, as a formal differential operator on polynomials,
τ ( D x ) M n = n M n − 1 . (12) \tau(D_x)M_n=nM_{n-1}. \tag*{(12)} τ ( D x ) M n = n M n − 1 . (12)
Indeed, τ ( D x ) G = τ ( g ( t ) ) G = t G \tau(D_x)G=\tau(g(t))G=tG τ ( D x ) G = τ ( g ( t )) G = tG . No convergence question is hidden here: on a polynomial only finitely many derivatives survive. This operator language belongs to the same tradition as Sheffer’s classification by lowering operators [32 ] .
The inverse sequence is again Sheffer, but need not be an ordinary orthogonal sequence. For example, when δ = 0 \delta=0 δ = 0 , η = 1 \eta=1 η = 1 ,
Q 0 = 1 , Q 1 = y , Q 2 = y 2 + 1 , Q 3 = y 3 + 5 y , Q 4 = y 4 + 14 y 2 + 5. Q_0=1,\qquad Q_1=y,\qquad Q_2=y^2+1,\qquad Q_3=y^3+5y,\qquad Q_4=y^4+14y^2+5. Q 0 = 1 , Q 1 = y , Q 2 = y 2 + 1 , Q 3 = y 3 + 5 y , Q 4 = y 4 + 14 y 2 + 5.
Here y Q 3 = Q 4 − 9 Q 2 + 4 Q 0 yQ_3=Q_4-9Q_2+4Q_0 y Q 3 = Q 4 − 9 Q 2 + 4 Q 0 , which rules out a three-term recurrence. Triangular inversion preserves the Sheffer structure, not automatically orthogonality.
The moments are already in the inverse matrix
Define a linear functional L \mathcal{L} L on polynomials by L ( M 0 ) = 1 \mathcal{L}(M_0)=1 L ( M 0 ) = 1 and L ( M n ) = 0 \mathcal{L}(M_n)=0 L ( M n ) = 0 for $n\geq 1. At this point the definition is algebraic; the measure will come next. Applying L \mathcal{L} L to (9) gives
μ n : = L ( x n ) = q n , 0 = Q n ( 0 ) . \mu_n:=\mathcal{L}(x^n)=q_{n,0}=Q_n(0). μ n := L ( x n ) = q n , 0 = Q n ( 0 ) .
Putting y = 0 y=0 y = 0 in (11) , we get all moments at once:
∑ n ≥ 0 μ n ( δ , η ) u n n ! = H δ , η ( u ) = ( cos u − δ sin u ) − η . (13) \sum_{n\ge0}\mu_n(\delta,\eta)\frac{u^n}{n!}=H_{\delta,\eta}(u)=(\cos u-\delta\sin u)^{-\eta}. \tag*{(13)} n ≥ 0 ∑ μ n ( δ , η ) n ! u n = H δ , η ( u ) = ( cos u − δ sin u ) − η . (13)
Equivalently, L ( G ( x ; t ) ) = 1 \mathcal{L}(G(x;t))=1 L ( G ( x ; t )) = 1 says f ( t ) H ( g ( t ) ) = 1 f(t)H(g(t))=1 f ( t ) H ( g ( t )) = 1 . The inversion formula and the moment calculation are two readings of the same identity.
For later checks, the first moments are
μ 0 = 1 , μ 1 = η δ , \mu_0=1,\qquad\mu_1=\eta\delta, μ 0 = 1 , μ 1 = η δ ,
μ 2 = η + η ( η + 1 ) δ 2 , \mu_2=\eta+\eta(\eta+1)\delta^2, μ 2 = η + η ( η + 1 ) δ 2 ,
μ 3 = η ( 3 η + 2 ) δ + η ( η + 1 ) ( η + 2 ) δ 3 . \mu_3=\eta(3\eta+2)\delta+\eta(\eta+1)(\eta+2)\delta^3. μ 3 = η ( 3 η + 2 ) δ + η ( η + 1 ) ( η + 2 ) δ 3 .
A convenient recurrence follows by differentiating (13) :
μ n + 1 = ( 1 + δ 2 ) ∂ δ μ n + η δ μ n , μ 0 = 1. (14) \mu_{n+1}=(1+\delta^2)\partial_\delta\mu_n+\eta\delta\mu_n,\qquad\mu_0=1. \tag*{(14)} μ n + 1 = ( 1 + δ 2 ) ∂ δ μ n + η δ μ n , μ 0 = 1. (14)
For δ = 0 \delta=0 δ = 0 , the generating function is sec η u \sec^\eta u sec η u . It is important not to confuse this with G ( 0 ; t ) = ( 1 + t 2 ) − η / 2 G(0;t)=(1+t^2)^{-\eta/2} G ( 0 ; t ) = ( 1 + t 2 ) − η /2 . The former generates moments; the latter generates the values of the orthogonal polynomials at the origin. They answer different questions.
A beta integral finds the measure
The moments have appeared algebraically. We now construct their positive measure and prove orthogonality directly. The main calculation is a Fourier transform that becomes a beta integral after one substitution.
The symmetric density and its exponential tilt
For η > 0 \eta>0 η > 0 , set
C η ( x ) = 2 η − 2 π Γ ( η ) ∣ Γ ( η + i x 2 ) ∣ 2 . (15) C_\eta(x)=\frac{2^{\eta-2}}{\pi\Gamma(\eta)}\left\lvert\Gamma\left(\frac{\eta+ix}{2}\right)\right\rvert^2. \tag*{(15)} C η ( x ) = π Γ ( η ) 2 η − 2 Γ ( 2 η + i x ) 2 . (15)
To discover this expression, put v = e 2 u v=e^{2u} v = e 2 u in the Fourier integral:
∫ R e − i x u sech η u d u = 2 η − 1 ∫ 0 ∞ v ( η − i x ) / 2 − 1 ( 1 + v ) η d v = 2 η − 1 Γ ( η ) Γ ( η − i x 2 ) Γ ( η + i x 2 ) . (16) \begin{aligned}
\int_{\mathbb{R}}e^{-ixu}\operatorname{sech}^{\eta}u\,du
=2^{\eta-1}\int_0^\infty\frac{v^{(\eta-ix)/2-1}}{(1+v)^\eta}\,dv \\
&=\frac{2^{\eta-1}}{\Gamma(\eta)}
\Gamma\left(\frac{\eta-ix}{2}\right)
\Gamma\left(\frac{\eta+ix}{2}\right). \tag*{(16)}
\end{aligned} ∫ R e − i xu sech η u d u = 2 η − 1 ∫ 0 ∞ ( 1 + v ) η v ( η − i x ) /2 − 1 d v = Γ ( η ) 2 η − 1 Γ ( 2 η − i x ) Γ ( 2 η + i x ) . (16)
Fourier inversion therefore gives
∫ R e i u x C η ( x ) d x = sech η u . (17) \int_{\mathbb{R}}e^{iux}C_\eta(x)\,dx=\operatorname{sech}^{\eta}u. \tag*{(17)} ∫ R e i ux C η ( x ) d x = sech η u . (17)
All integrals are ordinary absolutely convergent integrals. For example, Stirling’s formula gives
C η ( x ) ∼ ∣ x ∣ η − 1 Γ ( η ) e − π ∣ x ∣ / 2 ( ∣ x ∣ → ∞ ) . (18) C_\eta(x)\sim\frac{|x|^{\eta-1}}{\Gamma(\eta)}e^{-\pi|x|/2}\qquad(|x|\to\infty). \tag*{(18)} C η ( x ) ∼ Γ ( η ) ∣ x ∣ η − 1 e − π ∣ x ∣/2 ( ∣ x ∣ → ∞ ) . (18)
The density is positive, and its integral is one by (17) at u = 0 u=0 u = 0 .
Write θ = arctan δ \theta=\arctan\delta θ = arctan δ , so that − π / 2 < θ < π / 2 -\pi/2<\theta<\pi/2 − π /2 < θ < π /2 . Exponentially tilting this symmetric probability gives our central kernel:
K η ( x , δ ) = C η ( x ) ( 1 + δ 2 ) − η / 2 e x arctan δ . (19) K_\eta(x,\delta)=C_\eta(x)(1+\delta^2)^{-\eta/2}e^{x\arctan\delta}. \tag*{(19)} K η ( x , δ ) = C η ( x ) ( 1 + δ 2 ) − η /2 e x a r c t a n δ . (19)
Indeed, the exponential tails allow analytic continuation of (17) to ∣ ℑ u ∣ < π / 2 |\Im u|<\pi/2 ∣ℑ u ∣ < π /2 , and hence
∫ R e θ x C η ( x ) d x = cos − η θ . \int_{\mathbb{R}}e^{\theta x}C_\eta(x)\,dx=\cos^{-\eta}\theta. ∫ R e θ x C η ( x ) d x = cos − η θ .
Thus
d μ δ ( x ) = K η ( x , δ ) d x (20) d\mu_\delta(x)=K_\eta(x,\delta)\,dx \tag*{(20)} d μ δ ( x ) = K η ( x , δ ) d x (20)
is a probability measure. Its moment generating function is
∫ R e u x d μ δ ( x ) = cos η θ cos η ( θ + u ) = ( cos u − δ sin u ) − η , − π 2 < θ + u < π 2 . (21) \begin{aligned}
\int_{\mathbb{R}}e^{ux}\,d\mu_\delta(x)=\frac{\cos^\eta\theta}{\cos^\eta(\theta+u)} \\
&=\left(\cos u-\delta\sin u\right)^{-\eta},\qquad-\frac{\pi}{2}<\theta+u<\frac{\pi}{2}. \tag*{(21)}
\end{aligned} ∫ R e ux d μ δ ( x ) = cos η ( θ + u ) cos η θ = ( cos u − δ sin u ) − η , − 2 π < θ + u < 2 π . (21)
This agrees with the formal series (13) . Our algebraic moment functional has found its positive realization.
The measure is also unique among positive measures with these moments. Here is a short reason. For sufficiently small r > 0 r>0 r > 0 , ∑ n ≥ 0 μ 2 n r 2 n / ( 2 n ) ! < ∞ \sum_{n\geq0}\mu_{2n}r^{2n}/(2n)!<\infty ∑ n ≥ 0 μ 2 n r 2 n / ( 2 n )! < ∞ . If another positive measure has the same moments, monotone convergence applied to the series of cosh ( r x ) \cosh(rx) cosh ( r x ) gives the same finite integral. Its moment generating function is consequently analytic near zero, with the same Taylor coefficients as (21) . Equality extends through their common vertical strip, in particular along the imaginary axis. Uniqueness of the Fourier transform then gives the same measure. This observation will also settle the convergence of our Jacobi fraction.
Orthogonality in one line, once the line is understood
Let G ( x ; t ) = D ( t ) − η / 2 e x g ( t ) G(x;t)=D(t)^{-\eta/2}e^{xg(t)} G ( x ; t ) = D ( t ) − η /2 e xg ( t ) be the polynomial generating function from (8) , with D ( t ) = 1 + 2 δ t + ( 1 + δ 2 ) t 2 D(t)=1+2\delta t+(1+\delta^2)t^2 D ( t ) = 1 + 2 δ t + ( 1 + δ 2 ) t 2 . For small s , t s,t s , t , equation (21) gives
∫ G ( x ; s ) G ( x ; t ) d μ δ ( x ) = D ( s ) − η / 2 D ( t ) − η / 2 [ cos ( g ( s ) + g ( t ) ) − δ sin ( g ( s ) + g ( t ) ) ] − η . \int G(x;s)G(x;t)\,d\mu_\delta(x)=D(s)^{-\eta/2}D(t)^{-\eta/2}\left[\cos(g(s)+g(t))-\delta\sin(g(s)+g(t))\right]^{-\eta}. ∫ G ( x ; s ) G ( x ; t ) d μ δ ( x ) = D ( s ) − η /2 D ( t ) − η /2 [ cos ( g ( s ) + g ( t )) − δ sin ( g ( s ) + g ( t )) ] − η .
The trigonometry has been arranged for us by the recurrence:
cos g ( t ) = 1 + δ t D ( t ) , sin g ( t ) = t D ( t ) . \cos g(t)=\frac{1+\delta t}{\sqrt{D(t)}},\qquad\sin g(t)=\frac{t}{\sqrt{D(t)}}. cos g ( t ) = D ( t ) 1 + δ t , sin g ( t ) = D ( t ) t .
The addition formulas therefore reduce the bracket to
1 − ( 1 + δ 2 ) s t D ( s ) D ( t ) . \frac{1-(1+\delta^2)st}{\sqrt{D(s)D(t)}}. D ( s ) D ( t ) 1 − ( 1 + δ 2 ) s t .
Everything cancels except
∫ R G ( x ; s ) G ( x ; t ) d μ δ ( x ) = ( 1 − ( 1 + δ 2 ) s t ) − η . (22) \int_{\mathbb{R}}G(x;s)G(x;t)\,d\mu_\delta(x)=\left(1-(1+\delta^2)st\right)^{-\eta}. \tag*{(22)} ∫ R G ( x ; s ) G ( x ; t ) d μ δ ( x ) = ( 1 − ( 1 + δ 2 ) s t ) − η . (22)
Comparing coefficients proves both orthogonality and the norms:
∫ R M m ( x ; δ , η ) M n ( x ; δ , η ) K η ( x , δ ) d x = n ! ( η ) n ( 1 + δ 2 ) n δ m n . (23) \int_{\mathbb{R}}M_m(x;\delta,\eta)M_n(x;\delta,\eta)K_\eta(x,\delta)\,dx=n!(\eta)_n(1+\delta^2)^n\delta_{mn}. \tag*{(23)} ∫ R M m ( x ; δ , η ) M n ( x ; δ , η ) K η ( x , δ ) d x = n ! ( η ) n ( 1 + δ 2 ) n δ mn . (23)
The coefficient comparison is justified analytically in a neighborhood of ( s , t ) = ( 0 , 0 ) (s,t)=(0,0) ( s , t ) = ( 0 , 0 ) by the exponential tails; it may equally be read coefficient by coefficient after the moments have been identified. No measure has been guessed from the name of the polynomials.
For comparison with the conventional notation, put
λ = η 2 , ϕ = π 2 + arctan δ , δ = − cot ϕ . (24) \lambda=\frac{\eta}{2},\qquad\phi=\frac{\pi}{2}+\arctan\delta,\qquad\delta=-\cot\phi. \tag*{(24)} λ = 2 η , ϕ = 2 π + arctan δ , δ = − cot ϕ . (24)
Then the standard Meixner–Pollaczek polynomials satisfy
M n ( x ; δ , η ) = n ! sin n ϕ P n ( λ ) ( x / 2 ; ϕ ) . (25) M_n(x;\delta,\eta)=\frac{n!}{\sin^n\phi}P_n^{(\lambda)}(x/2;\phi). \tag*{(25)} M n ( x ; δ , η ) = sin n ϕ n ! P n ( λ ) ( x /2 ; ϕ ) . (25)
The identity follows either from their generating function or by matching the recurrence and the initial values [19 §9.7] . The angle convention fixes the sign of the exponential tilt in (19) .
The two parameters have simple jobs
At fixed δ \delta δ , multiplying the moment generating functions gives
μ δ , η 1 ∗ μ δ , η 2 = μ δ , η 1 + η 2 , η 1 , η 2 > 0. (26) \mu_{\delta,\eta_1} * \mu_{\delta,\eta_2}=\mu_{\delta,\eta_1+\eta_2},\qquad\eta_1,\eta_2>0. \tag*{(26)} μ δ , η 1 ∗ μ δ , η 2 = μ δ , η 1 + η 2 , η 1 , η 2 > 0. (26)
Thus η \eta η adds independent summands. The parameter δ \delta δ , or equivalently θ \theta θ , tilts the symmetric law toward one end of the line. In particular,
E [ X ] = η δ , Var ( X ) = η ( 1 + δ 2 ) . \mathbb{E}[X]=\eta\delta,\qquad\operatorname{Var}(X)=\eta(1+\delta^2). E [ X ] = η δ , Var ( X ) = η ( 1 + δ 2 ) .
There is a parallel polynomial addition formula. Multiplying G η 1 ( x ; t ) G_{\eta_1}(x;t) G η 1 ( x ; t ) and G η 2 ( y ; t ) G_{\eta_2}(y;t) G η 2 ( y ; t ) and comparing coefficients gives
M n ( x + y ; δ , η 1 + η 2 ) = ∑ k = 0 n ( n k ) M k ( x ; δ , η 1 ) M n − k ( y ; δ , η 2 ) . (27) M_n(x+y;\delta,\eta_1+\eta_2)=\sum_{k=0}^{n}\binom{n}{k}M_k(x;\delta,\eta_1)M_{n-k}(y;\delta,\eta_2). \tag*{(27)} M n ( x + y ; δ , η 1 + η 2 ) = k = 0 ∑ n ( k n ) M k ( x ; δ , η 1 ) M n − k ( y ; δ , η 2 ) . (27)
The probability law and the polynomial sequence express the same additive structure. Carlitz’s higher-order Euler numbers will make this correspondence especially concrete.
The shape parameter becomes time
There is another way to read the addition law: let the shape parameter measure elapsed time. Fix δ \delta δ , and write
μ t ( δ ) ( d x ) = K t ( x , δ ) d x , t > 0. \mu_t^{(\delta)}(dx)=K_t(x,\delta)\,dx,\qquad t>0. μ t ( δ ) ( d x ) = K t ( x , δ ) d x , t > 0.
These probabilities form a convolution semigroup by (26) . Their characteristic functions tend to one as t ↓ 0 t\downarrow0 t ↓ 0 , so the limiting law is a unit mass at zero. They therefore describe a process ( X t ) t ≥ 0 (X_t)_{t\geq0} ( X t ) t ≥ 0 starting at zero, with stationary independent increments and X t − X s ∼ μ t − s ( δ ) X_t-X_s\sim\mu_{t-s}^{(\delta)} X t − X s ∼ μ t − s ( δ ) for s < t s<t s < t . This is the Meixner process . For our calculations, only its joint laws at finitely many times are needed: prescribe independent increments with these densities; the convolution identity makes the prescriptions consistent. The connection with orthogonal polynomial martingales is classical [31 ] , but its proof here is already contained in the addition formula.
Abbreviate M n ( t ) ( x ) = M n ( x ; δ , t ) M_n^{(t)}(x)=M_n(x;\delta,t) M n ( t ) ( x ) = M n ( x ; δ , t ) and set
h n ( t ) = n ! ( t ) n ( 1 + δ 2 ) n . (28) h_n(t)=n!(t)_n(1+\delta^2)^n. \tag*{(28)} h n ( t ) = n ! ( t ) n ( 1 + δ 2 ) n . (28)
Proposition 5.1. For 0 < s < t 0 < s < t 0 < s < t and n ≥ 0 n \ge0 n ≥ 0 ,
E [ M n ( t ) ( X t ) ∣ F s ] = M n ( s ) ( X s ) , (29) \mathbb{E}[M_n^{(t)}(X_t) \mid\mathcal{F}_s] = M_n^{(s)}(X_s), \tag*{(29)} E [ M n ( t ) ( X t ) ∣ F s ] = M n ( s ) ( X s ) , (29) E [ M n ( s ) ( X s ) ∣ X t ] = ( s ) n ( t ) n M n ( t ) ( X t ) , (30) \mathbb{E}[M_n^{(s)}(X_s) \mid X_t] = \frac{(s)_n}{(t)_n}M_n^{(t)}(X_t), \tag*{(30)} E [ M n ( s ) ( X s ) ∣ X t ] = ( t ) n ( s ) n M n ( t ) ( X t ) , (30) where F s \mathcal{F}_s F s records the process up to time s s s .
Proof. Write X t = X s + Y X_t = X_s + Y X t = X s + Y , with Y Y Y independent of F s \mathcal{F}_s F s and having law μ t − s ( δ ) \mu_{t-s}^{(\delta)} μ t − s ( δ ) . Expand the polynomial by (27) . Orthogonality to the constant says E [ M j ( Y ; δ , t − s ) ] = 0 \mathbb{E}[M_j(Y;\delta,t-s)] = 0 E [ M j ( Y ; δ , t − s )] = 0 for j > 0 j > 0 j > 0 , so conditional expectation leaves exactly the term M n ( s ) ( X s ) M_n^{(s)}(X_s) M n ( s ) ( X s ) .
For the other direction, the identity just proved and orthogonality give
E [ M n ( s ) ( X s ) M m ( t ) ( X t ) ] = E [ M n ( s ) ( X s ) M m ( s ) ( X s ) ] = δ n m h n ( s ) . \mathbb{E}[M_n^{(s)}(X_s)M_m^{(t)}(X_t)] = \mathbb{E}[M_n^{(s)}(X_s)M_m^{(s)}(X_s)] = \delta_{nm}h_n(s). E [ M n ( s ) ( X s ) M m ( t ) ( X t )] = E [ M n ( s ) ( X s ) M m ( s ) ( X s )] = δ nm h n ( s ) .
Thus the orthogonal expansion of E [ M n ( s ) ( X s ) ∣ X t ] \mathbb{E}[M_n^{(s)}(X_s) \mid X_t] E [ M n ( s ) ( X s ) ∣ X t ] has only its n n n th term, with coefficient h n ( s ) / h n ( t ) = ( s ) n / ( t ) n h_n(s)/h_n(t) = (s)_n/(t)_n h n ( s ) / h n ( t ) = ( s ) n / ( t ) n .
To justify that expansion, the polynomials are complete in L 2 ( μ t ( δ ) ) L^2(\mu_t^{(\delta)}) L 2 ( μ t ( δ ) ) . Indeed, if f f f is orthogonal to every polynomial, Cauchy–Schwarz and the exponential tails make ∫ f ( x ) e w x μ t ( δ ) ( d x ) \int f(x)e^{wx}\,\mu_t^{(\delta)}(dx) ∫ f ( x ) e w x μ t ( δ ) ( d x ) analytic in a strip about the imaginary axis. All its derivatives at zero vanish. The identity theorem, followed by uniqueness of the Fourier transform, gives f = 0 f = 0 f = 0 almost everywhere. Conditional expectation belongs to this L 2 L^2 L 2 space by Jensen’s inequality, so its expansion identifies it uniquely. □
The first identity says that the polynomial’s present value is its conditional future average: it is a martingale. The second gives the reverse statement after division by ( t ) n (t)_n ( t ) n . More precisely, if F t = σ ( X r : r ≥ t ) \mathcal{F}^t = \sigma(X_r : r \ge t) F t = σ ( X r : r ≥ t ) records the future, independence of the increments after t t t shows that M n ( t ) ( X t ) / ( t ) n M_n^{(t)}(X_t)/(t)_n M n ( t ) ( X t ) / ( t ) n is a reverse martingale with respect to these decreasing sigma-fields.
The first two degrees already give useful conditional moments. Since
M 1 ( t ) ( x ) = x − t δ , M 2 ( t ) ( x ) = x 2 − 2 ( t + 1 ) δ x + t ( t + 1 ) δ 2 − t , M_1^{(t)}(x) = x - t\delta,\qquad M_2^{(t)}(x) = x^2 - 2(t+1)\delta x + t(t+1)\delta^2 - t, M 1 ( t ) ( x ) = x − t δ , M 2 ( t ) ( x ) = x 2 − 2 ( t + 1 ) δ x + t ( t + 1 ) δ 2 − t ,
substitution in (30) yields
Table 5.17.
The tilt has disappeared. In the next section we will see why the entire conditional density shares this property, and why Carlitz’s Bernoulli construction appears in it. These conditional moments are also recorded in [8 ] . Appendix A uses the same two polynomial identities to determine exactly how much information survives conditioning.
Carlitz's route: numbers first, polynomials later
Carlitz enters this story from the opposite direction. He starts with familiar numbers—Bernoulli numbers, Euler numbers, and their relatives—and looks for polynomials that make those numbers into moments. His 1959 paper [5 ] moves repeatedly between a symbolic calculation, a three-term recurrence, and an integral or sum. Reading it through Meixner’s classification reveals how much of the same landscape these different starting points uncover.
This is more than a change of notation. A moment sequence can look rather unpromising when inspected term by term, yet become transparent in a basis adapted to it. Carlitz repeatedly finds that basis by replacing powers with binomial coefficients. The resulting cancellation is elementary; its consequences include orthogonality, explicit norms, and eventually a measure. We follow that thread through the parts of his paper that meet the Meixner family, and then briefly examine the Bernoulli branch that goes beyond it.
The symbolic language deserves a brief explanation. If numbers a n a_n a n are given, the notation a n = a n a^n=a_n a n = a n means that the linear functional A [ u n ] = a n \mathcal{A}[u^n]=a_n A [ u n ] = a n is to be applied after expanding a polynomial. It does not assert that the a n a_n a n are powers of one number. Likewise, e a t = ∑ a n t n / n ! e^{at}=\sum a_n t^n/n! e a t = ∑ a n t n / n ! is shorthand for A [ e u t ] \mathcal{A}[e^{ut}] A [ e u t ] . Such a functional need not be positive. That distinction will explain both the imaginary changes of variable below and the signs in Carlitz’s formulas.
Eulerian moments: one calculation, two kinds of Meixner polynomials
Carlitz’s “Eulerian” numbers are defined by
R λ [ e z t ] = ∑ j ≥ 0 R j ( λ ) t j j ! = 1 − λ e t − λ , λ ≠ 1. (31) \mathcal{R}_{\lambda}[e^{zt}]=\sum_{j\geq0}R_j(\lambda)\frac{t^j}{j!}=\frac{1-\lambda}{e^t-\lambda},\qquad\lambda\ne1. \tag*{(31)} R λ [ e z t ] = j ≥ 0 ∑ R j ( λ ) j ! t j = e t − λ 1 − λ , λ = 1. (31)
The associated polynomials have generating function
∑ n ≥ 0 A n ( λ ) ( z ) t n = ( 1 − λ t ) z ( 1 − t ) z + 1 . (32) \sum_{n\geq0}A_n^{(\lambda)}(z)t^n=\frac{(1-\lambda t)^z}{(1-t)^{z+1}}. \tag*{(32)} n ≥ 0 ∑ A n ( λ ) ( z ) t n = ( 1 − t ) z + 1 ( 1 − λ t ) z . (32)
Here the series is ordinary in t t t ; the leading coefficient of A n ( λ ) A_n^{(\lambda)} A n ( λ ) is ( 1 − λ ) n / n ! (1-\lambda)^n/n! ( 1 − λ ) n / n ! . These are [5 (1.17), (1.19)] ; the parameter λ \lambda λ here is Carlitz’s parameter, not our shape parameter η \eta η .
Carlitz’s cancellation is particularly pleasant. Substituting t = log ( 1 + v ) t=\log(1+v) t = log ( 1 + v ) in (31) gives
R λ [ ( z r ) ] = ( λ − 1 ) − r . \mathcal{R}_{\lambda}\left[\binom{z}{r}\right]=(\lambda-1)^{-r}. R λ [ ( r z ) ] = ( λ − 1 ) − r .
On the other hand, expanding (32) gives
A n ( λ ) ( z ) = ∑ r = 0 n ( 1 − λ ) r ( n r ) ( z r ) = 2 F 1 ( − n , − z ; 1 ; 1 − λ ) . (33) A_n^{(\lambda)}(z)=\sum_{r=0}^{n}(1-\lambda)^r\binom{n}{r}\binom{z}{r}={}_2F_1(-n,-z;1;1-\lambda). \tag*{(33)} A n ( λ ) ( z ) = r = 0 ∑ n ( 1 − λ ) r ( r n ) ( r z ) = 2 F 1 ( − n , − z ; 1 ; 1 − λ ) . (33)
Applying R λ \mathcal{R}_{\lambda} R λ leaves ∑ r ( − 1 ) r ( n r ) = 0 \sum_r(-1)^r\binom{n}{r}=0 ∑ r ( − 1 ) r ( r n ) = 0 for n > 0 n>0 n > 0 . This is the essential step in [5 (5.1)–(5.4)] .
Differentiating the generating function supplies the recurrence
( n + 1 ) A n + 1 ( λ ) ( z ) = [ ( 1 − λ ) z + 1 + n ( 1 + λ ) ] A n ( λ ) ( z ) − n λ A n − 1 ( λ ) ( z ) . (34) (n+1)A_{n+1}^{(\lambda)}(z)=\left[(1-\lambda)z+1+n(1+\lambda)\right]A_n^{(\lambda)}(z)-n\lambda A_{n-1}^{(\lambda)}(z). \tag*{(34)} ( n + 1 ) A n + 1 ( λ ) ( z ) = [ ( 1 − λ ) z + 1 + n ( 1 + λ ) ] A n ( λ ) ( z ) − nλ A n − 1 ( λ ) ( z ) . (34)
Multiplication by z z z can lower the polynomial index by at most one. Therefore R λ [ z r A n ( λ ) ] = 0 \mathcal{R}_{\lambda}[z^rA_n^{(\lambda)}]=0 R λ [ z r A n ( λ ) ] = 0 for r < n r<n r < n : fewer than n n n multiplications cannot reach the constant polynomial. At the first possible step there is just one contribution: every multiplication must lower the index. Solving (34) for z A n ( λ ) zA_n^{(\lambda)} z A n ( λ ) shows that the lowering coefficient is n λ / ( 1 − λ ) n\lambda/(1-\lambda) nλ / ( 1 − λ ) , so
R λ [ z n A n ( λ ) ] = ∏ j = 1 n j λ 1 − λ = n ! λ n ( 1 − λ ) n . \mathcal{R}_{\lambda}[z^nA_n^{(\lambda)}]=\prod_{j=1}^{n}\frac{j\lambda}{1-\lambda}=\frac{n!\lambda^n}{(1-\lambda)^n}. R λ [ z n A n ( λ ) ] = j = 1 ∏ n 1 − λ j λ = ( 1 − λ ) n n ! λ n .
Multiplying by the leading coefficient of A n ( λ ) A_n^{(\lambda)} A n ( λ ) gives
R λ [ A m ( λ ) A n ( λ ) ] = λ n δ m n . (35) \mathcal{R}_{\lambda}\left[A_m^{(\lambda)}A_n^{(\lambda)}\right]=\lambda^n\delta_{mn}. \tag*{(35)} R λ [ A m ( λ ) A n ( λ ) ] = λ n δ mn . (35)
This illustrates a useful feature of Carlitz’s method. Once the average of each nonconstant polynomial is zero, the recurrence does the remaining work.
For λ > 1 \lambda>1 λ > 1 , the functional is already a positive probability law:
R λ [ p ] = ( λ − 1 ) ∑ j = 0 ∞ p ( j ) λ − j − 1 . (36) \mathcal{R}_{\lambda}[p]=(\lambda-1)\sum_{j=0}^{\infty}p(j)\lambda^{-j-1}. \tag*{(36)} R λ [ p ] = ( λ − 1 ) j = 0 ∑ ∞ p ( j ) λ − j − 1 . (36)
Indeed, summing the geometric series for p ( z ) = e z t p(z)=e^{zt} p ( z ) = e z t gives (31) . Equation (33) now identifies A n ( λ ) A_n^{(\lambda)} A n ( λ ) with the Meixner polynomials of the first kind having shape parameter 1 and c = 1 / λ c=1/\lambda c = 1/ λ . Carlitz gives this discrete orthogonality in [5 §8] .
The same functional also has the contour representation
R λ [ p ] = 1 − λ 2 i λ ∫ − 1 / 2 − i ∞ − 1 / 2 + i ∞ p ( z ) ( − λ ) − z sin π z d z , (37) \mathcal{R}_{\lambda}[p]=\frac{1-\lambda}{2i\lambda}\int_{-1/2-i\infty}^{-1/2+i\infty}\frac{p(z)(-\lambda)^{-z}}{\sin\pi z}\,dz, \tag*{(37)} R λ [ p ] = 2 iλ 1 − λ ∫ − 1/2 − i ∞ − 1/2 + i ∞ sin π z p ( z ) ( − λ ) − z d z , (37)
for λ ∈ C ∖ [ 0 , ∞ ) \lambda\in\mathbb{C}\setminus[0,\infty) λ ∈ C ∖ [ 0 , ∞ ) , with the principal logarithm of − λ -\lambda − λ . This is an absolutely convergent integral for every polynomial: on a vertical line the reciprocal sine decays like e − π ∣ Im z ∣ e^{-\pi|\operatorname{Im}z|} e − π ∣ Im z ∣ , whereas the exponential factor contributes e arg ( − λ ) Im z e^{\arg(-\lambda)\operatorname{Im}z} e a r g ( − λ ) Im z , with ∣ arg ( − λ ) ∣ < π |\arg(-\lambda)|<\pi ∣ arg ( − λ ) ∣ < π . Carlitz proves the identity by shifting the contour by one unit. The residue fixes the mass, and the shift gives
R λ [ p ( z + 1 ) − λ p ( z ) ] = ( 1 − λ ) p ( 0 ) , \mathcal{R}_{\lambda}[p(z+1)-\lambda p(z)]=(1-\lambda)p(0), R λ [ p ( z + 1 ) − λ p ( z )] = ( 1 − λ ) p ( 0 ) ,
which also follows directly from (31) . For p ( z ) = z n p(z)=z^n p ( z ) = z n , the coefficient of the unknown n n n th moment is 1 − λ 1-\lambda 1 − λ , and all remaining moments have smaller index. The identity therefore determines the functional inductively, including its mass when n = 0 n=0 n = 0 . This is the moment argument behind [5 (7.1)–(7.5)] .
Now put
λ = δ − i δ + i , z = − 1 + i x 2 , δ ∈ R . (38) \lambda=\frac{\delta-i}{\delta+i},\qquad z=\frac{-1+ix}{2},\qquad\delta\in\mathbb{R}. \tag*{(38)} λ = δ + i δ − i , z = 2 − 1 + i x , δ ∈ R . (38)
Since 1 − λ = 2 i / ( δ + i ) 1-\lambda=2i/(\delta+i) 1 − λ = 2 i / ( δ + i ) and 1 + λ = 2 δ / ( δ + i ) 1+\lambda=2\delta/(\delta+i) 1 + λ = 2 δ / ( δ + i ) , the recurrence (34) changes into our monic Meixner recurrence with η = 1 \eta=1 η = 1 . The initial polynomial becomes M 1 = x − δ M_1=x-\delta M 1 = x − δ , fixing the normalization, and we obtain
M n ( x ; δ , 1 ) = n ! ( − δ − i ) n A n ( ( δ − i ) / ( δ + i ) ) ( − 1 + i x 2 ) . (39) M_n(x;\delta,1)=n!(-\delta-i)^nA_n^{((\delta-i)/(\delta+i))}\left(\frac{-1+ix}{2}\right). \tag*{(39)} M n ( x ; δ , 1 ) = n ! ( − δ − i ) n A n (( δ − i ) / ( δ + i )) ( 2 − 1 + i x ) . (39)
The contour becomes the real x x x -axis. If θ = arctan δ \theta=\arctan\delta θ = arctan δ , then − λ = e 2 i θ -\lambda=e^{2i\theta} − λ = e 2 i θ , and its weight becomes
cos θ e θ x 2 cosh ( π x / 2 ) d x = K 1 ( x , δ ) d x . \frac{\cos\theta\,e^{\theta x}}{2\cosh(\pi x/2)}\,dx=K_1(x,\delta)\,dx. 2 cosh ( π x /2 ) cos θ e θ x d x = K 1 ( x , δ ) d x .
Thus Carlitz’s contour construction is exactly the positive Meixner-II measure at η = 1 \eta=1 η = 1 . Even the apparently complex norm in (35) becomes our positive norm, because ( − δ − i ) 2 λ = 1 + δ 2 (-\delta-i)^2\lambda=1+\delta^2 ( − δ − i ) 2 λ = 1 + δ 2 . The discrete and continuous Meixner families have appeared from the same symbolic calculation, with different realizations of its functional. In particular, Carlitz’s connection with the continuous family is not confined to the symmetric case δ = 0 \delta=0 δ = 0 . His Eulerian parameter already contains every real tilt δ \delta δ at shape 1. The complex quantities in (39) are coordinates for real polynomials and a positive measure; the recurrence and the transformed weight make that reality visible.
Higher Euler numbers and the meaning of the imaginary rotation
Carlitz next considers the signed Euler numbers of order η \eta η , which we write as the moments of E η \mathcal{E}_{\eta} E η :
E η [ e u t ] = ∑ j ≥ 0 E j ( η ) t j j ! = sech η t . (40) \mathcal{E}_{\eta}[e^{ut}] = \sum_{j\ge0} \mathcal{E}^{(\eta)}_j \frac{t^j}{j!} = \operatorname{sech}^{\eta} t. \tag*{(40)} E η [ e u t ] = j ≥ 0 ∑ E j ( η ) j ! t j = sech η t . (40)
In particular, E 2 ( η ) = − η \mathcal{E}^{(\eta)}_2=-\eta E 2 ( η ) = − η . For η > 0 \eta>0 η > 0 this cannot be the second moment of a positive measure on the real line. The umbral identity is nevertheless perfectly meaningful, and positivity will emerge after rotation.
His monic polynomials f n ( η ) f_n^{(\eta)} f n ( η ) satisfy
f n + 1 ( η ) ( u ) = u f n ( η ) ( u ) + n ( n + η − 1 ) f n − 1 ( η ) ( u ) , f 0 ( η ) = 1 , f 1 ( η ) = u . (41) f_{n+1}^{(\eta)}(u) = u f_n^{(\eta)}(u) + n(n+\eta-1)f_{n-1}^{(\eta)}(u), \qquad f_0^{(\eta)}=1,\quad f_1^{(\eta)}=u. \tag*{(41)} f n + 1 ( η ) ( u ) = u f n ( η ) ( u ) + n ( n + η − 1 ) f n − 1 ( η ) ( u ) , f 0 ( η ) = 1 , f 1 ( η ) = u . (41)
The plus sign is intentional. If Φ η ( u ; t ) = ∑ f n ( η ) ( u ) t n / n ! \Phi_{\eta}(u;t)=\sum f_n^{(\eta)}(u)t^n/n! Φ η ( u ; t ) = ∑ f n ( η ) ( u ) t n / n ! , the recurrence gives the first-order equation
( 1 − t 2 ) ∂ t Φ η = ( u + η t ) Φ η . (1-t^2)\partial_t\Phi_{\eta}=(u+\eta t)\Phi_{\eta}. ( 1 − t 2 ) ∂ t Φ η = ( u + η t ) Φ η .
Integration with Φ η ( u ; 0 ) = 1 \Phi_{\eta}(u;0)=1 Φ η ( u ; 0 ) = 1 yields
Φ η ( u ; t ) = ( 1 − t 2 ) − η / 2 e u arctanh t = ( 1 + t ) ( u − η ) / 2 ( 1 − t ) ( u + η ) / 2 . (42) \Phi_{\eta}(u;t)=(1-t^2)^{-\eta/2}e^{u\operatorname{arctanh}t}
=\frac{(1+t)^{(u-\eta)/2}}{(1-t)^{(u+\eta)/2}}. \tag*{(42)} Φ η ( u ; t ) = ( 1 − t 2 ) − η /2 e u arctanh t = ( 1 − t ) ( u + η ) /2 ( 1 + t ) ( u − η ) /2 . (42)
This is Carlitz’s derivation in [5 §9, (9.3)–(9.7)] .
In his proof of orthogonality, another binomial identity first makes the nonconstant averages vanish, and the recurrence again propagates the result. The mechanism can be compressed into one generating-function calculation:
E η [ Φ η ( u ; s ) Φ η ( u ; t ) ] = ( 1 − s 2 ) − η / 2 ( 1 − t 2 ) − η / 2 sech η ( arctanh s + arctanh t ) = ( 1 + s t ) − η . (43) \mathcal{E}_{\eta}[\Phi_{\eta}(u;s)\Phi_{\eta}(u;t)]
=(1-s^2)^{-\eta/2}(1-t^2)^{-\eta/2}
\operatorname{sech}^{\eta}(\operatorname{arctanh}s+\operatorname{arctanh}t)
=(1+st)^{-\eta}. \tag*{(43)} E η [ Φ η ( u ; s ) Φ η ( u ; t )] = ( 1 − s 2 ) − η /2 ( 1 − t 2 ) − η /2 sech η ( arctanh s + arctanh t ) = ( 1 + s t ) − η . (43)
Coefficient comparison gives the exact signed norm
E η [ f m ( η ) f n ( η ) ] = ( − 1 ) n n ! ( η ) n δ m n . (44) \mathcal{E}_{\eta}[f_m^{(\eta)}f_n^{(\eta)}]=(-1)^n n!(\eta)_n\delta_{mn}. \tag*{(44)} E η [ f m ( η ) f n ( η ) ] = ( − 1 ) n n ! ( η ) n δ mn . (44)
This argument takes place in formal power series and uses no positivity. There is no convergence question hidden in the coefficient comparison: each coefficient uses only finitely many moments. The hyperbolic addition formula is what removes all mixed terms except those with equal polynomial indices. It is the same mechanism as our earlier bilinear calculation for M n M_n M n , now expressed in the signed Euler convention.
The conversion to our polynomials is now immediate:
M n ( x ; 0 , η ) = i − n f n ( η ) ( i x ) , μ n ( 0 , η ) = i − n E n ( η ) . (45) M_n(x;0,\eta)=i^{-n}f_n^{(\eta)}(ix),\qquad\mu_n(0,\eta)=i^{-n}\mathcal{E}^{(\eta)}_n. \tag*{(45)} M n ( x ; 0 , η ) = i − n f n ( η ) ( i x ) , μ n ( 0 , η ) = i − n E n ( η ) . (45)
The imaginary rotation changes the sign of the recurrence coefficient and turns the signed norms into n ! ( η ) n n!(\eta)_n n ! ( η ) n . In probabilistic terms, E η [ p ( u ) ] = ∫ p ( i x ) K η ( x , 0 ) d x \mathcal{E}_{\eta}[p(u)]=\int p(ix)K_{\eta}(x,0)\,dx E η [ p ( u )] = ∫ p ( i x ) K η ( x , 0 ) d x . This explains the symbolic substitution without pretending that the signed Euler numbers themselves are moments of a positive real measure.1
Equivalently, the ordinary moment generating function of the positive law is sec η t \sec^\eta t sec η t near t = 0 t=0 t = 0 , while its characteristic function is sech η t \operatorname{sech}^\eta t sech η t . This distinction matters: Carlitz’s generating function becomes a characteristic function on the real line, not the generating function of its ordinary real moments. The factors i i i in (45) say precisely this.
For integer η = k ≥ 1 \eta=k\ge1 η = k ≥ 1 , Carlitz realizes the measure as a convolution of k k k hyperbolic-secant laws. In our scale,
p 1 ( x ) = 1 2 cosh ( π x / 2 ) , p k = p 1 ∗ k , ∫ R e i t x p k ( x ) d x = sech k t . (46) p_1(x)=\frac{1}{2\cosh(\pi x/2)},\qquad p_k=p_1^{*k},\qquad\int_{\mathbb{R}}e^{itx}p_k(x)\,\mathrm{d}x=\operatorname{sech}^k t. \tag*{(46)} p 1 ( x ) = 2 cosh ( π x /2 ) 1 , p k = p 1 ∗ k , ∫ R e i t x p k ( x ) d x = sech k t . (46)
Our gamma density identifies p k ( x ) = K k ( x , 0 ) p_k(x)=K_k(x,0) p k ( x ) = K k ( x , 0 ) and extends this positive realization to every real η > 0 \eta>0 η > 0 , as in (26) . Carlitz’s constructions illuminate both parameters: the order of the Euler numbers controls convolution, while the continuous Eulerian parameter realizes the exponential tilt in (19) .
The Bernoulli branch and its intersection with Meixner
The first part of Carlitz’s paper extends Touchard’s work on Bernoulli moments. For a new parameter a a a , define
β j ( a ) = B j + 1 ( a ) − B j + 1 ( j + 1 ) a , ∑ j ≥ 0 β j ( a ) t j j ! = e a t − 1 a ( e t − 1 ) . (47) \beta_j(a)=\frac{B_{j+1}(a)-B_{j+1}}{(j+1)a},\qquad\sum_{j\geq0}\beta_j(a)\frac{t^j}{j!}=\frac{e^{at}-1}{a(e^t-1)}. \tag*{(47)} β j ( a ) = ( j + 1 ) a B j + 1 ( a ) − B j + 1 , j ≥ 0 ∑ β j ( a ) j ! t j = a ( e t − 1 ) e a t − 1 . (47)
The limit a → 0 a\to0 a → 0 gives the Bernoulli numbers. At a positive integer a = N a=N a = N , Faulhaber’s formula says that β j ( N ) \beta_j(N) β j ( N ) is the average of 0 j , 1 j , … , ( N − 1 ) j 0^j,1^j,\ldots,(N-1)^j 0 j , 1 j , … , ( N − 1 ) j . Thus the same expression interpolates a familiar finite probability law and the Bernoulli moment functional.
Carlitz identifies the corresponding polynomials with the Bateman–Pasternack family. In his normalization their recurrence is
Ω n + 1 ( a ) ( z ) = ( 2 z + 1 − a ) Ω n ( a ) ( z ) + n 2 ( n 2 − a 2 ) 4 n 2 − 1 Ω n − 1 ( a ) ( z ) . (48) \Omega_{n+1}^{(a)}(z)=(2z+1-a)\Omega_n^{(a)}(z)+\frac{n^2(n^2-a^2)}{4n^2-1}\Omega_{n-1}^{(a)}(z). \tag*{(48)} Ω n + 1 ( a ) ( z ) = ( 2 z + 1 − a ) Ω n ( a ) ( z ) + 4 n 2 − 1 n 2 ( n 2 − a 2 ) Ω n − 1 ( a ) ( z ) . (48)
The finite support leaves a visible trace: at a = N a=N a = N the recurrence coefficient vanishes when n = N n=N n = N . The nondegenerate orthogonal family then has exactly N N N members, corresponding to the N N N support points of the uniform law. This is already a different phenomenon from the infinite positive Meixner families considered above. After the change z = ( a − 1 + i x ) / 2 z=(a-1+ix)/2 z = ( a − 1 + i x ) /2 , the monic polynomials Q n ( a ) ( x ) = i − n Ω n ( a ) ( ( a − 1 + i x ) / 2 ) Q_n^{(a)}(x)=i^{-n}\Omega_n^{(a)}((a-1+ix)/2) Q n ( a ) ( x ) = i − n Ω n ( a ) (( a − 1 + i x ) /2 ) satisfy
Q n + 1 ( a ) ( x ) = x Q n ( a ) ( x ) − n 2 ( n 2 − a 2 ) 4 n 2 − 1 Q n − 1 ( a ) ( x ) . Q_{n+1}^{(a)}(x)=xQ_n^{(a)}(x)-\frac{n^2(n^2-a^2)}{4n^2-1}Q_{n-1}^{(a)}(x). Q n + 1 ( a ) ( x ) = x Q n ( a ) ( x ) − 4 n 2 − 1 n 2 ( n 2 − a 2 ) Q n − 1 ( a ) ( x ) .
For 0 < a < 1 0<a<1 0 < a < 1 , Carlitz’s contour integral [5 (1.21), (6.1)] becomes the positive density
sin π a 2 a d x cos π a + cosh π x . (49) \frac{\sin\pi a}{2a}\frac{\mathrm{d}x}{\cos\pi a+\cosh\pi x}. \tag*{(49)} 2 a sin π a cos π a + cosh π x d x . (49)
This is a continuous Hahn family, neighboring Meixner in the larger hypergeometric picture [21 ] . Its recurrence coefficient is generally different from the quadratic Meixner coefficient. At a = 1 / 2 a=1/2 a = 1/2 , however, it reduces to n 2 / 4 n^2/4 n 2 /4 , and
Q n ( 1 / 2 ) ( x ) = 2 − n M n ( 2 x ; 0 , 1 ) . Q_n^{(1/2)}(x)=2^{-n}M_n(2x;0,1). Q n ( 1/2 ) ( x ) = 2 − n M n ( 2 x ; 0 , 1 ) .
The density then becomes 1 / cosh π x 1/\cosh\pi x 1/ cosh π x . This is the intersection between the Bernoulli and Euler constructions that Carlitz develops in [5 §§3–4] . It explains why both sorts of numbers occur in his paper without requiring that all of its polynomial families be Meixner families.
The Bernoulli density is a Meixner connection
The neighboring continuous Hahn family has an even closer connection to Meixner: it appears when we condition two independent Meixner summands on their sum. Let ( X t ) (X_t) ( X t ) be the process of § 5.4 , with fixed tilt δ \delta δ . For 0 < s < t 0 < s < t 0 < s < t , independence of X s X_s X s and X t − X s X_t-X_s X t − X s gives the conditional density
b s , t ; x ( y ) = K s ( y , δ ) K t − s ( x − y , δ ) K t ( x , δ ) = C s ( y ) C t − s ( x − y ) C t ( x ) . (50) b_{s,t;x}(y)=\frac{K_s(y,\delta)K_{t-s}(x-y,\delta)}{K_t(x,\delta)}=\frac{C_s(y)C_{t-s}(x-y)}{C_t(x)}. \tag*{(50)} b s , t ; x ( y ) = K t ( x , δ ) K s ( y , δ ) K t − s ( x − y , δ ) = C t ( x ) C s ( y ) C t − s ( x − y ) . (50)
The exponential tilts cancel, as do their normalizing factors. Thus the whole conditional law of X s X_s X s given X t = x X_t=x X t = x is independent of δ \delta δ , explaining the cancellation in (5.17). The convolution identity gives its normalization. Since the denominator is positive, this formula supplies a continuous version of the conditional density for every real endpoint x x x , including the endpoint zero.
With r = t − s r=t-s r = t − s and z = y / 2 z=y/2 z = y /2 , the density in z z z is proportional to
∣ Γ ( s 2 + i z ) Γ ( r − i x 2 + i z ) ∣ 2 . \left|\Gamma\left(\frac{s}{2}+iz\right)\Gamma\left(\frac{r-ix}{2}+iz\right)\right|^2. Γ ( 2 s + i z ) Γ ( 2 r − i x + i z ) 2 .
This is the continuous Hahn weight with parameters ( s / 2 , ( r − i x ) / 2 , s / 2 , ( r + i x ) / 2 ) (s/2,(r-ix)/2,s/2,(r+ix)/2) ( s /2 , ( r − i x ) /2 , s /2 , ( r + i x ) /2 ) . The same disintegration occurs in Koelink and Van der Jeugt’s positive discrete tensor-product construction [20 Proposition 3.3 and its proof] : continuous Hahn polynomials resolve the two summands at fixed sum. In a statistical setting, Devroye and Hill call the conditional law the betaized Meixner–Morris distribution [8 §11] . Formula (50) makes the common calculation visible.
Proposition 6.1 (Carlitz’s weight as a connector). Let q a q_a q a denote the density in (49) , extended to − 1 < a < 1 -1<a<1 − 1 < a < 1 and interpreted continuously at a = 0 a=0 a = 0 . Then
P ( X 1 + a ∈ d y ∣ X 2 = 0 ) = q a ( y ) d y = C 1 + a ( y ) C 1 − a ( y ) C 2 ( 0 ) d y . (51) \mathbb{P}(X_{1+a}\in dy\mid X_2=0)=q_a(y)\,dy=\frac{C_{1+a}(y)C_{1-a}(y)}{C_2(0)}\,dy. \tag*{(51)} P ( X 1 + a ∈ d y ∣ X 2 = 0 ) = q a ( y ) d y = C 2 ( 0 ) C 1 + a ( y ) C 1 − a ( y ) d y . (51)
Proof. Only the identification of the density remains. Since C 2 ( 0 ) = 1 / π C_2(0)=1/\pi C 2 ( 0 ) = 1/ π , Euler’s reflection formula gives
∣ Γ ( 1 + a + i y 2 ) Γ ( 1 − a + i y 2 ) ∣ 2 = π 2 ∣ cos ( π ( a + i y ) / 2 ) ∣ 2 = 2 π 2 cos π a + cosh π y . \left|\Gamma\left(\frac{1+a+iy}{2}\right)\Gamma\left(\frac{1-a+iy}{2}\right)\right|^2
=\frac{\pi^2}{\left|\cos\left(\pi(a+iy)/2\right)\right|^2}
=\frac{2\pi^2}{\cos\pi a+\cosh\pi y}. Γ ( 2 1 + a + i y ) Γ ( 2 1 − a + i y ) 2 = ∣ cos ( π ( a + i y ) /2 ) ∣ 2 π 2 = cos π a + cosh π y 2 π 2 .
Insert the constants in (15) and use Γ ( 1 + a ) Γ ( 1 − a ) = π a / sin π a \Gamma(1+a)\Gamma(1-a)=\pi a/\sin\pi a Γ ( 1 + a ) Γ ( 1 − a ) = π a / sin π a . The result is (49) for a ≠ 0 a\ne0 a = 0 , and continuity supplies a = 0 a=0 a = 0 .
Carlitz’s Euler branch therefore describes an unconditioned Meixner law, while his Bernoulli/Pasternack branch is obtained by conditioning the same process on its endpoint. At the midpoint the connector has the logistic density
q 0 ( y ) = π 4 cosh 2 ( π y / 2 ) . q_0(y)=\frac{\pi}{4\cosh^2(\pi y/2)}. q 0 ( y ) = 4 cosh 2 ( π y /2 ) π .
at time 3 / 2 3/2 3/2 it has the hyperbolic-secant density q 1 / 2 ( y ) = 1 / cosh π y q_{1/2}(y)=1/\cosh\pi y q 1/2 ( y ) = 1/ cosh π y . The symmetry q a = q − a q_a=q_{-a} q a = q − a expresses time reversal of the zero-endpoint connector, together with its spatial symmetry. Finally, (5.17) gives
Var q a ( Y ) = 1 − a 2 3 , \operatorname{Var}_{q_a}(Y)=\frac{1-a^2}{3}, Var q a ( Y ) = 3 1 − a 2 ,
exactly the first coefficient of the monic Bernoulli/Pasternack recurrence. A conditional probability has explained both the family and its simplest recurrence coefficient.
Carlitz explicitly says that the higher-order Euler construction was suggested by Stieltjes’ continued fraction for the Laplace transform of sech η t \operatorname{sech}\eta t sech η t , citing Wall’s book [5 §9, (9.2)] . We will derive that fraction in § 8.3 , rather than treat it as a recognition trick. The general relation between recurrences, moment fractions, and Hankel determinants is established in § 7 . Carlitz’s contribution is especially illuminating here: familiar numbers lead to a functional, the functional finds its polynomials, and a suitable realization of the functional reveals the measure.
Paths, permutations, and the formal fraction
The recurrence contains a small combinatorial machine. Write
b n = ( 2 n + η ) δ , λ n = n ( n + η − 1 ) ( 1 + δ 2 ) . (52) b_n=(2n+\eta)\delta,\qquad\lambda_n=n(n+\eta-1)(1+\delta^2). \tag*{(52)} b n = ( 2 n + η ) δ , λ n = n ( n + η − 1 ) ( 1 + δ 2 ) . (52)
Multiplication by x x x takes M k M_k M k to M k + 1 + b k M k + λ k M k − 1 M_{k+1}+b_kM_k+\lambda_kM_{k-1} M k + 1 + b k M k + λ k M k − 1 . Think of these three terms as an up step, a level step, and a down step, starting at height k k k . Their weights are respectively 1 1 1 , b k b_k b k , and λ k \lambda_k λ k . A down step from height zero is forbidden.
After n n n multiplications, the coefficient q n , k q_{n,k} q n , k in (9) is the sum of the weights of all paths of length n n n from height zero to height k k k . The moment μ n = q n , 0 \mu_n=q_{n,0} μ n = q n , 0 counts the returning paths, usually called Motzkin paths. This is the mechanism of Viennot’s interpretation of orthogonal polynomials [12 , 38 ] . It follows directly from matrix multiplication, so remains an identity for weights of both signs. Here all weights are nonnegative when δ ≥ 0 \delta\ge0 δ ≥ 0 ; for δ < 0 \delta<0 δ < 0 the horizontal weights are signed.
First return gives the continued fraction
Let S j ( z ) S_j(z) S j ( z ) be the formal generating function of excursions based at height j j j that never go below j j j . Such an excursion is a sequence of two possible pieces: a level step of weight b j z b_jz b j z , or an arch consisting of an up step, an excursion at height j + 1 j+1 j + 1 , and a down step. The arch has weight λ j + 1 z 2 S j + 1 ( z ) \lambda_{j+1}z^2S_{j+1}(z) λ j + 1 z 2 S j + 1 ( z ) . Consequently
S j ( z ) = 1 1 − b j z − λ j + 1 z 2 S j + 1 ( z ) . S_j(z)=\frac{1}{1-b_jz-\lambda_{j+1}z^2S_{j+1}(z)}. S j ( z ) = 1 − b j z − λ j + 1 z 2 S j + 1 ( z ) 1 .
Starting at height zero gives
∑ n ≥ 0 μ n ( δ , η ) z n = 1 1 − η δ z − η ( 1 + δ 2 ) z 2 1 − ( η + 2 ) δ z − 2 ( η + 1 ) ( 1 + δ 2 ) z 2 1 − ( η + 4 ) δ z − ⋯ . (53) \sum_{n\ge0}\mu_n(\delta,\eta)z^n
=
\cfrac{1}{1-\eta\delta z-\cfrac{\eta(1+\delta^2)z^2}{1-(\eta+2)\delta z-\cfrac{2(\eta+1)(1+\delta^2)z^2}{1-(\eta+4)\delta z-\cdots}}}. \tag*{(53)} n ≥ 0 ∑ μ n ( δ , η ) z n = 1 − η δ z − 1 − ( η + 2 ) δ z − 1 − ( η + 4 ) δ z − ⋯ 2 ( η + 1 ) ( 1 + δ 2 ) z 2 η ( 1 + δ 2 ) z 2 1 . (53)
The formal meaning is precise: at any specified degree, only finitely many heights can be reached, so every sufficiently deep truncation has the same coefficient. This is the path form of the classical Jacobi fraction correspondence [9 , 41 ] .
It is important to keep this ordinary generating function separate from the exponential generating function (13) . In fact, the ordinary series in (53) has radius zero. For any R > 0 R>0 R > 0 , positivity of the density on the whole line gives
μ 2 n ≥ R 2 n μ δ { ∣ x ∣ > R } . \mu_{2n}\ge R^{2n}\mu_\delta\{|x|>R\}. μ 2 n ≥ R 2 n μ δ { ∣ x ∣ > R } .
Taking 2 n 2n 2 n th roots and then allowing R R R to increase shows μ 2 n 1 / ( 2 n ) → ∞ \mu_{2n}^{1/(2n)} \to\infty μ 2 n 1/ ( 2 n ) → ∞ . The continued fraction nevertheless has a natural analytic meaning after the spectral variable is placed in its denominators. We will prove that meaning in the next section.
The same recurrence also evaluates the Hankel determinants of our moment sequence. Let
Δ N = det [ μ j + k ] j , k = 0 N . \Delta_N=\det[\mu_{j+k}]_{j,k=0}^{N}. Δ N = det [ μ j + k ] j , k = 0 N .
Changing from the monomials to monic orthogonal polynomials is a triangular change of basis with determinant one. It diagonalizes this Gram matrix, so (23) gives
Δ N = ∏ j = 0 N j ! ( η ) j ( 1 + δ 2 ) j = ( 1 + δ 2 ) N ( N + 1 ) / 2 ∏ j = 0 N j ! ( η ) j . (54) \Delta_N=\prod_{j=0}^{N}j!(\eta)_j(1+\delta^2)^j=(1+\delta^2)^{N(N+1)/2}\prod_{j=0}^{N}j!(\eta)_j. \tag*{(54)} Δ N = j = 0 ∏ N j ! ( η ) j ( 1 + δ 2 ) j = ( 1 + δ 2 ) N ( N + 1 ) /2 j = 0 ∏ N j ! ( η ) j . (54)
The product evaluations are another form of the norm calculation, rather than a separate miracle of determinants.
Why powers of the secant count alternating permutations
At δ = 0 \delta=0 δ = 0 , horizontal steps disappear. The odd moments vanish, and the even moments have generating function sec η t \sec^\eta t sec η t . There is a particularly attractive interpretation of the parameter η \eta η . Let Z 2 n \mathcal{Z}_{2n} Z 2 n be the permutations satisfying
σ 1 < σ 2 > σ 3 < ⋯ < σ 2 n , \sigma_1<\sigma_2>\sigma_3<\cdots<\sigma_{2n}, σ 1 < σ 2 > σ 3 < ⋯ < σ 2 n ,
and let lrmin ( σ ) \operatorname{lrmin}(\sigma) lrmin ( σ ) count their left-to-right minima, including the first entry. Then
μ 2 n ( 0 , η ) = ∑ σ ∈ Z 2 n η lrmin ( σ ) . (55) \mu_{2n}(0,\eta)=\sum_{\sigma\in\mathcal{Z}_{2n}}\eta^{\operatorname{lrmin}(\sigma)}. \tag*{(55)} μ 2 n ( 0 , η ) = σ ∈ Z 2 n ∑ η lrmin ( σ ) . (55)
This refinement is due to Viennot; Hetyei uses it to connect the Meixner moments with representation theory [12 §2] .
Here is a generating-function proof that also explains the statistic. Let E 2 n , k E_{2n,k} E 2 n , k count the permutations with k k k such minima, and let T 2 j + 1 T_{2j+1} T 2 j + 1 be the number of alternating permutations of odd length 2 j + 1 2j+1 2 j + 1 . Their exponential generating function is tan t \tan t tan t . For completeness, the usual decomposition of an odd alternating permutation at its largest entry gives T ′ ( t ) = 1 + T ( t ) 2 T'(t)=1+T(t)^2 T ′ ( t ) = 1 + T ( t ) 2 , T ( 0 ) = 0 T(0)=0 T ( 0 ) = 0 , and hence T ( t ) = tan t T(t)=\tan t T ( t ) = tan t .
In a permutation from Z 2 n \mathcal{Z}_{2n} Z 2 n , the entry 1 is in an odd position, say 2 m + 1 2m+1 2 m + 1 . The left block has even length 2 m 2m 2 m and carries k − 1 k-1 k − 1 left-to-right minima. The entry 1 is the last new minimum; the right block has odd length and contributes no further minima. Its alternating pattern is reversed, but complementation of its labels gives the same tangent-number count. Choosing the labels in the left block gives
E 2 n , k = ∑ m = 0 n − 1 ( 2 n − 1 2 m ) E 2 m , k − 1 T 2 n − 2 m − 1 . E_{2n,k}=\sum_{m=0}^{n-1}\binom{2n-1}{2m}E_{2m,k-1}T_{2n-2m-1}. E 2 n , k = m = 0 ∑ n − 1 ( 2 m 2 n − 1 ) E 2 m , k − 1 T 2 n − 2 m − 1 .
With Φ k ( t ) = ∑ n ≥ 0 E 2 n , k t 2 n / ( 2 n ) ! \Phi_k(t)=\sum_{n\geq0}E_{2n,k}t^{2n}/(2n)! Φ k ( t ) = ∑ n ≥ 0 E 2 n , k t 2 n / ( 2 n )! , this becomes
Φ 0 ( t ) = 1 , Φ k ′ ( t ) = Φ k − 1 ( t ) tan t , Φ k ( 0 ) = 0 ( k > 0 ) . \Phi_0(t)=1,\qquad\Phi_k'(t)=\Phi_{k-1}(t)\tan t,\qquad\Phi_k(0)=0\quad(k>0). Φ 0 ( t ) = 1 , Φ k ′ ( t ) = Φ k − 1 ( t ) tan t , Φ k ( 0 ) = 0 ( k > 0 ) .
Integration yields
Φ k ( t ) = ( − log cos t ) k k ! , ∑ k ≥ 0 η k Φ k ( t ) = sec η t . \Phi_k(t)=\frac{(-\log\cos t)^k}{k!},\qquad\sum_{k\geq0}\eta^k\Phi_k(t)=\sec^\eta t. Φ k ( t ) = k ! ( − log cos t ) k , k ≥ 0 ∑ η k Φ k ( t ) = sec η t .
proving (55) . For example,
μ 2 = η , μ 4 = 3 η 2 + 2 η , μ 6 = 15 η 3 + 30 η 2 + 16 η . \mu_2=\eta,\qquad\mu_4=3\eta^2+2\eta,\qquad\mu_6=15\eta^3+30\eta^2+16\eta. μ 2 = η , μ 4 = 3 η 2 + 2 η , μ 6 = 15 η 3 + 30 η 2 + 16 η .
At η = 1 \eta=1 η = 1 these are the secant numbers. At η = 2 \eta=2 η = 2 they are the odd tangent numbers, since ( tan t ) ′ = sec 2 t (\tan t)'=\sec^2t ( tan t ) ′ = sec 2 t . The latter specialization will bring the even values of the zeta function into the picture.
What the continued fraction converges to
The analytic companion of the formal moment series is its Cauchy transform
C δ ( ζ ) = ∫ R K η ( x , δ ) ζ − x d x , ζ ∈ C ∖ R . (56) C_\delta(\zeta)=\int_{\mathbb R}\frac{K_\eta(x,\delta)}{\zeta-x}\,dx,\qquad\zeta\in\mathbb C\setminus\mathbb R. \tag*{(56)} C δ ( ζ ) = ∫ R ζ − x K η ( x , δ ) d x , ζ ∈ C ∖ R . (56)
With this sign convention, C δ C_\delta C δ has negative imaginary part in the upper half-plane. The Nevanlinna convention, which we will use for the parameter fraction, has the opposite sign.
Finite matrices give an infinite limit
Theorem 8.1. For η > 0 \eta>0 η > 0 and real δ \delta δ , the finite convergents of
C δ ( ζ ) = 1 ζ − η δ − η ( 1 + δ 2 ) ζ − ( η + 2 ) δ − 2 ( η + 1 ) ( 1 + δ 2 ) ζ − ( η + 4 ) δ − ⋯ (57) C_\delta(\zeta)=\cfrac{1}{\zeta-\eta\delta-\cfrac{\eta(1+\delta^2)}{\zeta-(\eta+2)\delta-\cfrac{2(\eta+1)(1+\delta^2)}{\zeta-(\eta+4)\delta-\cdots}}} \tag*{(57)} C δ ( ζ ) = ζ − η δ − ζ − ( η + 2 ) δ − ζ − ( η + 4 ) δ − ⋯ 2 ( η + 1 ) ( 1 + δ 2 ) η ( 1 + δ 2 ) 1 (57) converge locally uniformly on C ∖ R \mathbb C\setminus\mathbb R C ∖ R to the integral (56) .
Proof. Let J N J_N J N be the real symmetric N × N N\times N N × N matrix with diagonal b 0 , … , b N − 1 b_0,\ldots,b_{N-1} b 0 , … , b N − 1 and off-diagonal entries λ 1 , … , λ N − 1 \sqrt{\lambda_1},\ldots,\sqrt{\lambda_{N-1}} λ 1 , … , λ N − 1 . Successive Schur complements give
C δ , N ( ζ ) = ⟨ e 0 , ( ζ I − J N ) − 1 e 0 ⟩ = 1 ζ − b 0 − λ 1 ζ − b 1 − λ 2 ⋱ − λ N − 1 ζ − b N − 1 . (58) C_{\delta,N}(\zeta)=\left\langle e_0,(\zeta I-J_N)^{-1}e_0\right\rangle=\cfrac{1}{\zeta-b_0-\cfrac{\lambda_1}{\zeta-b_1-\cfrac{\lambda_2}{\ddots-\cfrac{\lambda_{N-1}}{\zeta-b_{N-1}}}}}. \tag*{(58)} C δ , N ( ζ ) = ⟨ e 0 , ( ζ I − J N ) − 1 e 0 ⟩ = ζ − b 0 − ζ − b 1 − ⋱ − ζ − b N − 1 λ N − 1 λ 2 λ 1 1 . (58)
The finite-dimensional spectral theorem makes this the Cauchy transform of a probability measure μ δ , N \mu_{\delta,N} μ δ , N supported on the eigenvalues of J N J_N J N .
Its first 2 N 2N 2 N moments agree with those of μ δ \mu_\delta μ δ :
∫ x j d μ δ , N ( x ) = μ j , 0 ≤ j ≤ 2 N − 1. (59) \int x^j\,d\mu_{\delta,N}(x)=\mu_j,\qquad0\leq j\leq2N-1. \tag*{(59)} ∫ x j d μ δ , N ( x ) = μ j , 0 ≤ j ≤ 2 N − 1. (59)
Indeed, the matrix entry ⟨ e 0 , J N j e 0 ⟩ \left\langle e_0,J_N^j e_0\right\rangle ⟨ e 0 , J N j e 0 ⟩ is a sum over walks. A walk that reaches the missing height N N N and returns to zero has length at least 2 N 2N 2 N . Before that length, truncation changes nothing. This is Gaussian quadrature seen through the paths.
The common second moment makes the measures tight for N ≥ 2 N \ge2 N ≥ 2 . Every weakly convergent subsequence has a limit with all the moments μ j \mu_j μ j : for each fixed j j j , use a common higher even moment in (59) to control the tails uniformly. The determinacy proved in § 5 forces this limit to be μ δ \mu_\delta μ δ . Thus the whole sequence converges weakly to μ δ \mu_\delta μ δ .
The function x ↦ ( ζ − x ) − 1 x \mapsto(\zeta-x)^{-1} x ↦ ( ζ − x ) − 1 is bounded and continuous off the real axis, so the Cauchy transforms converge pointwise. On a compact set at distance ε > 0 \varepsilon> 0 ε > 0 from R \mathbb{R} R , they and their derivatives are bounded by ε − 1 \varepsilon^{-1} ε − 1 and ε − 2 \varepsilon^{-2} ε − 2 . A finite-net argument gives local uniform convergence.
The denominator of (58) is det ( ζ I − J N ) = M N ( ζ ; δ , η ) \det(\zeta I-J_N)=M_N(\zeta;\delta,\eta) det ( ζ I − J N ) = M N ( ζ ; δ , η ) . For later comparison, define the numerator polynomials by
N 0 = 0 , N 1 = 1 , N n + 1 = ( ζ − b n ) N n − λ n N n − 1 ( n ≥ 1 ) . (60) N_0=0,\qquad N_1=1,\qquad N_{n+1}=(\zeta-b_n)N_n-\lambda_nN_{n-1}\quad(n\ge1). \tag*{(60)} N 0 = 0 , N 1 = 1 , N n + 1 = ( ζ − b n ) N n − λ n N n − 1 ( n ≥ 1 ) . (60)
Then C δ , N = N N / M N C_{\delta,N}=N_N/M_N C δ , N = N N / M N . These associated polynomials are different from the inverse Sheffer polynomials of § 4 .
Although the ordinary moment series diverges, its coefficients still describe the transform at infinity:
C δ ( ζ ) ∼ ∑ n ≥ 0 μ n ζ n + 1 (61) C_\delta(\zeta)\sim\sum_{n\ge0}\frac{\mu_n}{\zeta^{n+1}} \tag*{(61)} C δ ( ζ ) ∼ n ≥ 0 ∑ ζ n + 1 μ n (61)
in every closed angular sector separated from the real axis. To prove any finite order of this expansion, use the finite geometric identity for ( ζ − x ) − 1 (\zeta-x)^{-1} ( ζ − x ) − 1 and bound the remainder by an absolute moment. Thus the formal and analytic interpretations fit together without treating a divergent series as a convergent one.
A hypergeometric evaluation, with the half-plane specified
The fraction can also be summed explicitly. We use 2 F 1 ( a , b ; c ; q ) = ∑ j ≥ 0 ( a ) j ( b ) j q j / ( ( c ) j j ! ) {}_2F_1(a,b;c;q)=\sum_{j\ge0}(a)_j(b)_j q^j/((c)_j j!) 2 F 1 ( a , b ; c ; q ) = ∑ j ≥ 0 ( a ) j ( b ) j q j / (( c ) j j !) near q = 0 q=0 q = 0 , with its usual continuation cut along [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) .
Proposition 8.2. For ℑ ζ < 0 \Im\zeta<0 ℑ ζ < 0 ,
C δ ( ζ ) = 1 ζ − i η 2 F 1 ( 1 , η ; 1 + η + i ζ 2 ; 1 + i δ 2 ) . (62) C_\delta(\zeta)=\frac{1}{\zeta-i\eta}{}_2F_1\left(1,\eta;1+\frac{\eta+i\zeta}{2};\frac{1+i\delta}{2}\right). \tag*{(62)} C δ ( ζ ) = ζ − i η 1 2 F 1 ( 1 , η ; 1 + 2 η + i ζ ; 2 1 + i δ ) . (62) For ℑ ζ > 0 \Im\zeta>0 ℑ ζ > 0 , its conjugate form is
C δ ( ζ ) = 1 ζ + i η 2 F 1 ( 1 , η ; 1 + η − i ζ 2 ; 1 − i δ 2 ) . (63) C_\delta(\zeta)=\frac{1}{\zeta+i\eta}{}_2F_1\left(1,\eta;1+\frac{\eta-i\zeta}{2};\frac{1-i\delta}{2}\right). \tag*{(63)} C δ ( ζ ) = ζ + i η 1 2 F 1 ( 1 , η ; 1 + 2 η − i ζ ; 2 1 − i δ ) . (63)
Proof. For ℑ ζ < 0 \Im\zeta<0 ℑ ζ < 0 ,
1 ζ − x = i ∫ 0 ∞ e − i ( ζ − x ) u d u . \frac{1}{\zeta-x}=i\int_0^\infty e^{-i(\zeta-x)u}\,du. ζ − x 1 = i ∫ 0 ∞ e − i ( ζ − x ) u d u .
Absolute Fubini is justified by the integrable majorant e ( ℑ ζ ) u e^{(\Im\zeta)u} e ( ℑ ζ ) u . The characteristic function obtained from (21) therefore gives
C δ ( ζ ) = i ∫ 0 ∞ e − i ζ u ( cosh u − i δ sinh u ) − η d u . (64) C_\delta(\zeta)=i\int_0^\infty e^{-i\zeta u}(\cosh u-i\delta\sinh u)^{-\eta}\,du. \tag*{(64)} C δ ( ζ ) = i ∫ 0 ∞ e − i ζ u ( cosh u − i δ sinh u ) − η d u . (64)
Put r = e − 2 u r=e^{-2u} r = e − 2 u , a = ( η + i ζ ) / 2 a=(\eta+i\zeta)/2 a = ( η + i ζ ) /2 , and q = ( 1 + i δ ) / 2 q=(1+i\delta)/2 q = ( 1 + i δ ) /2 . Then ℜ a > 0 \Re a>0 ℜ a > 0 and
C δ ( ζ ) = i 2 ∫ 0 1 r a − 1 [ ( 1 − q ) + q r ] − η d r = i 2 a ( 1 − q ) − η 2 F 1 ( η , a ; a + 1 ; − q 1 − q ) = i 2 a 2 F 1 ( η , 1 ; a + 1 ; q ) . \begin{aligned}
C_{\delta}(\zeta)=\frac{i}{2}\int_{0}^{1}r^{a-1}\left[(1-q)+qr\right]^{-\eta}\,dr \\
&=\frac{i}{2a}(1-q)^{-\eta}{}_2F_1\left(\eta,a;a+1;-\frac{q}{1-q}\right) \\
&=\frac{i}{2a}{}_2F_1(\eta,1;a+1;q).
\end{aligned} C δ ( ζ ) = 2 i ∫ 0 1 r a − 1 [ ( 1 − q ) + q r ] − η d r = 2 a i ( 1 − q ) − η 2 F 1 ( η , a ; a + 1 ; − 1 − q q ) = 2 a i 2 F 1 ( η , 1 ; a + 1 ; q ) .
The second line is Euler’s beta-integral representation; the third is Pfaff’s transformation. Both follow with the branches continued from small q q q to ℜ q = 1 / 2 \Re q=1/2 ℜ q = 1/2 , where no cut is crossed [29 §§15.6, 15.8] . Since i / ( 2 a ) = 1 / ( ζ − i η ) i/(2a)=1/(\zeta-i\eta) i / ( 2 a ) = 1/ ( ζ − i η ) , this proves (62) . Conjugating the real density proves (63) . □ \square □
The proof gives a summation of the Jacobi fraction by identifying its analytic function first. It also prepares a useful surprise: the same hypergeometric expression will appear when δ \delta δ , rather than x x x , becomes the spectral variable.
Stieltjes’ fraction and Carlitz’s starting point
At δ = 0 \delta=0 δ = 0 , the fraction takes a familiar form. For ℜ s > 0 \Re s>0 ℜ s > 0 , let
L η ( s ) = ∫ 0 ∞ e − s u sech η u d u . L_{\eta}(s)=\int_{0}^{\infty}e^{-su}\operatorname{sech}^{\eta}u\,du. L η ( s ) = ∫ 0 ∞ e − s u sech η u d u .
Since C 0 ( − i s ) = i L η ( s ) C_0(-is)=iL_{\eta}(s) C 0 ( − i s ) = i L η ( s ) , Theorem 8.1 gives the convergent continued fraction
L η ( s ) = 1 s + η s + 2 ( η + 1 ) s + 3 ( η + 2 ) s + ⋯ , ℜ s > 0. (65) L_{\eta}(s)=\cfrac{1}{s+\cfrac{\eta}{s+\cfrac{2(\eta+1)}{s+\cfrac{3(\eta+2)}{s+\cdots}}}},\qquad\Re s>0. \tag*{(65)} L η ( s ) = s + s + s + s + ⋯ 3 ( η + 2 ) 2 ( η + 1 ) η 1 , ℜ s > 0. (65)
The convergence is locally uniform in this half-plane. This is the Stieltjes fraction that Carlitz cites as the impetus for his higher-order Euler construction [5 §9] .
Its coefficients can be derived directly by integration by parts. For real s > 0 s>0 s > 0 , define
A n ( s ) = ∫ 0 ∞ e − s u tanh n u sech η u d u . A_n(s)=\int_{0}^{\infty}e^{-su}\tanh^n u\,\operatorname{sech}^{\eta}u\,du. A n ( s ) = ∫ 0 ∞ e − s u tanh n u sech η u d u .
Differentiating tanh n u sech η u \tanh^n u\,\operatorname{sech}^{\eta}u tanh n u sech η u gives
d d u ( tanh n u sech η u ) = n tanh n − 1 u sech η u − ( n + η ) tanh n + 1 u sech η u . \frac{d}{du}\left(\tanh^n u\,\operatorname{sech}^{\eta}u\right)
=n\tanh^{n-1}u\,\operatorname{sech}^{\eta}u-(n+\eta)\tanh^{n+1}u\,\operatorname{sech}^{\eta}u. d u d ( tanh n u sech η u ) = n tanh n − 1 u sech η u − ( n + η ) tanh n + 1 u sech η u .
The endpoint terms yield
1 = s A 0 + η A 1 , n A n − 1 = s A n + ( n + η ) A n + 1 ( n ≥ 1 ) . (66) 1=sA_0+\eta A_1,\qquad nA_{n-1}=sA_n+(n+\eta)A_{n+1}\quad(n\ge1). \tag*{(66)} 1 = s A 0 + η A 1 , n A n − 1 = s A n + ( n + η ) A n + 1 ( n ≥ 1 ) . (66)
Since A n > 0 A_n>0 A n > 0 , their ratios may be substituted successively:
A n A n − 1 = n s + ( n + η ) A n + 1 / A n . \frac{A_n}{A_{n-1}}=\frac{n}{s+(n+\eta)A_{n+1}/A_n}. A n − 1 A n = s + ( n + η ) A n + 1 / A n n .
Together with the first identity, this produces exactly (65) . The preceding Jacobi argument supplies the convergence that a succession of formal substitutions alone would not establish. The two derivations have different jobs: the integrals explain the coefficients, and the finite matrices prove the infinite limit.
A tangent-number encore: the even zeta values
The specialization η = 2 \eta=2 η = 2 , δ = 0 \delta=0 δ = 0 has a particularly simple density. The reflection formula for the gamma function, or the convolution in Carlitz’s construction, gives
K 2 ( x , 0 ) = x 2 sinh ( π x / 2 ) . (67) K_2(x,0)=\frac{x}{2\sinh(\pi x/2)}. \tag*{(67)} K 2 ( x , 0 ) = 2 sinh ( π x /2 ) x . (67)
with its continuous value 1 / π 1/\pi 1/ π at x = 0 x=0 x = 0 . The convolution is worth seeing once. If p 1 ( x ) = 1 / ( 2 cosh ( π x / 2 ) ) p_1(x)=1/(2\cosh(\pi x/2)) p 1 ( x ) = 1/ ( 2 cosh ( π x /2 )) and v = π x / 2 v=\pi x/2 v = π x /2 , then
( p 1 ∗ p 1 ) ( x ) = 1 2 π ∫ R d u cosh u cosh ( v − u ) . (p_1*p_1)(x)=\frac{1}{2\pi}\int_{\mathbb{R}}\frac{du}{\cosh u\cosh(v-u)}. ( p 1 ∗ p 1 ) ( x ) = 2 π 1 ∫ R cosh u cosh ( v − u ) d u .
The identity
1 cosh u cosh ( v − u ) = tanh u + tanh ( v − u ) sinh v \frac{1}{\cosh u\cosh(v-u)}=\frac{\tanh u+\tanh(v-u)}{\sinh v} cosh u cosh ( v − u ) 1 = sinh v tanh u + tanh ( v − u )
reduces the integral to 2 v / sinh v 2v/\sinh v 2 v / sinh v by evaluating log cosh u − log cosh ( v − u ) \log\cosh u-\log\cosh(v-u) log cosh u − log cosh ( v − u ) at the endpoints. This proves (67) directly from Carlitz’s two summands.
Write tan t = ∑ m ≥ 0 T 2 m + 1 t 2 m + 1 / ( 2 m + 1 ) ! \tan t=\sum_{m\geq0}T_{2m+1}t^{2m+1}/(2m+1)! tan t = ∑ m ≥ 0 T 2 m + 1 t 2 m + 1 / ( 2 m + 1 )! . Since sec 2 t = ( tan t ) ′ \sec^2t=(\tan t)' sec 2 t = ( tan t ) ′ , our moment calculation says
T 2 m + 1 = ∫ R x 2 m K 2 ( x , 0 ) d x . T_{2m+1}=\int_{\mathbb{R}}x^{2m}K_2(x,0)\,dx. T 2 m + 1 = ∫ R x 2 m K 2 ( x , 0 ) d x .
For x > 0 x>0 x > 0 , expand
1 2 sinh ( π x / 2 ) = ∑ j ≥ 0 e − ( j + 1 / 2 ) π x . \frac{1}{2\sinh(\pi x/2)}=\sum_{j\geq0}e^{-(j+1/2)\pi x}. 2 sinh ( π x /2 ) 1 = j ≥ 0 ∑ e − ( j + 1/2 ) π x .
Every term is nonnegative, so termwise integration follows from monotone convergence. The elementary gamma integral then gives
T 2 m + 1 = 2 ∑ j ≥ 0 ∫ 0 ∞ x 2 m + 1 e − ( j + 1 / 2 ) π x d x = 2 ( 2 m + 1 ) ! π 2 m + 2 ( 2 2 m + 2 − 1 ) ζ ( 2 m + 2 ) . (68) \begin{aligned}
T_{2m+1}
&=2\sum_{j\geq0}\int_0^\infty x^{2m+1}e^{-(j+1/2)\pi x}\,dx\\
&=\frac{2(2m+1)!}{\pi^{2m+2}}(2^{2m+2}-1)\zeta(2m+2). \tag*{(68)}
\end{aligned} T 2 m + 1 = 2 j ≥ 0 ∑ ∫ 0 ∞ x 2 m + 1 e − ( j + 1/2 ) π x d x = π 2 m + 2 2 ( 2 m + 1 )! ( 2 2 m + 2 − 1 ) ζ ( 2 m + 2 ) . (68)
The odd integers in the geometric expansion explain the factor 2 2 m + 2 − 1 2^{2m+2}-1 2 2 m + 2 − 1 .
To recover Euler’s usual form, define the Bernoulli numbers by
z e z − 1 = ∑ n ≥ 0 B n z n n ! . \frac{z}{e^z-1}=\sum_{n\geq0}B_n\frac{z^n}{n!}. e z − 1 z = n ≥ 0 ∑ B n n ! z n .
Substituting z = 2 i t z=2it z = 2 i t and using tan t = cot t − 2 cot ( 2 t ) \tan t=\cot t-2\cot(2t) tan t = cot t − 2 cot ( 2 t ) gives, by coefficient comparison,
T 2 r − 1 = ( − 1 ) r − 1 2 2 r ( 2 2 r − 1 ) 2 r B 2 r , r ≥ 1. (69) T_{2r-1}=(-1)^{r-1}\frac{2^{2r}(2^{2r}-1)}{2r}B_{2r},\qquad r\geq1. \tag*{(69)} T 2 r − 1 = ( − 1 ) r − 1 2 r 2 2 r ( 2 2 r − 1 ) B 2 r , r ≥ 1. (69)
Combining (68) and (69) yields
ζ ( 2 r ) = ( − 1 ) r + 1 ( 2 π ) 2 r B 2 r 2 ( 2 r ) ! . (70) \zeta(2r)=(-1)^{r+1}\frac{(2\pi)^{2r}B_{2r}}{2(2r)!}. \tag*{(70)} ζ ( 2 r ) = ( − 1 ) r + 1 2 ( 2 r )! ( 2 π ) 2 r B 2 r . (70)
The measure was obtained from a beta integral, so this argument does not use the zeta evaluation it is proving. It places a classical formula at the meeting point of a positive measure, alternating permutations, and Carlitz’s Euler umbra. For other elementary routes to the same values, see [13 ] ; for the arithmetic of tangent and secant numbers, see [3 ] .
The same density measures the jumps
The tangent density also has a role in the Meixner process of § 5.4 . Put θ = arctan δ \theta= \arctan\delta θ = arctan δ and write its cumulant exponent per unit time as
Ψ δ ( u ) = − log ( cos u − δ sin u ) , log E e u X t = t Ψ δ ( u ) . \Psi_\delta(u)=-\log(\cos u-\delta\sin u),\qquad\log\mathbb{E}e^{uX_t}=t\Psi_\delta(u). Ψ δ ( u ) = − log ( cos u − δ sin u ) , log E e u X t = t Ψ δ ( u ) .
The positive measure
ν δ ( d y ) = e θ y 2 y sinh ( π y / 2 ) d y = ( 1 + δ 2 ) K 2 ( y , δ ) y 2 d y (71) \nu_\delta(dy)=\frac{e^{\theta y}}{2y\sinh(\pi y/2)}\,dy=(1+\delta^2)\frac{K_2(y,\delta)}{y^2}\,dy \tag*{(71)} ν δ ( d y ) = 2 y sinh ( π y /2 ) e θ y d y = ( 1 + δ 2 ) y 2 K 2 ( y , δ ) d y (71)
is its Lévy measure: it records the intensity of the jumps. This classical formula [24 §3.1] follows directly from the density already in hand. Indeed,
y 2 ν δ ( d y ) = e θ y C 2 ( y ) d y , ∫ R y 2 e u y ν δ ( d y ) = sec 2 ( θ + u ) = Ψ δ ′ ′ ( u ) . y^2\nu_\delta(dy)=e^{\theta y}C_2(y)\,dy,\qquad\int_{\mathbb{R}} y^2e^{uy}\nu_\delta(dy)=\sec^2(\theta+u)=\Psi_\delta^{\prime\prime}(u). y 2 ν δ ( d y ) = e θ y C 2 ( y ) d y , ∫ R y 2 e u y ν δ ( d y ) = sec 2 ( θ + u ) = Ψ δ ′′ ( u ) .
The measure satisfies ∫ min ( 1 , y 2 ) ν δ ( d y ) < ∞ \int\min(1,y^2)\nu_\delta(dy)<\infty ∫ min ( 1 , y 2 ) ν δ ( d y ) < ∞ : its density is asymptotic to 1 / ( π y 2 ) 1/(\pi y^2) 1/ ( π y 2 ) at zero and has exponentially decreasing tails. Integrating twice and fixing the value and derivative at zero therefore gives
Ψ δ ( u ) = δ u + ∫ R ( e u y − 1 − u y ) ν δ ( d y ) . (72) \Psi_\delta(u)=\delta u+\int_{\mathbb{R}}(e^{uy}-1-uy)\nu_\delta(dy). \tag*{(72)} Ψ δ ( u ) = δ u + ∫ R ( e u y − 1 − u y ) ν δ ( d y ) . (72)
The compensation cancels the singularity at zero, and the integral converges throughout ∣ θ + u ∣ < π / 2 |\theta+u|<\pi/2 ∣ θ + u ∣ < π /2 . This is the Lévy–Khintchine representation; no Gaussian term is present.
If κ n ( δ , η ) \kappa_n(\delta,\eta) κ n ( δ , η ) denotes the n n n th cumulant of our law, then log H ( u ) = η Ψ δ ( u ) \log H(u)=\eta\Psi_\delta(u) log H ( u ) = η Ψ δ ( u ) , and differentiation gives
κ n + 2 ( δ , η ) = η ( 1 + δ 2 ) μ n ( δ , 2 ) , n ≥ 0. (73) \kappa_{n+2}(\delta,\eta)=\eta(1+\delta^2)\mu_n(\delta,2),\qquad n\ge0. \tag*{(73)} κ n + 2 ( δ , η ) = η ( 1 + δ 2 ) μ n ( δ , 2 ) , n ≥ 0. (73)
Thus the shape-two moments are the higher cumulants of the entire Meixner convolution family. At zero tilt they are the tangent numbers just used to evaluate the even zeta values. The same small density connects moments, cumulants, and jumps.
Turning the recurrence sideways
So far x x x has been the variable and δ \delta δ a parameter. Nothing in the recurrence prevents us from reversing those roles. The question is what kind of orthogonality survives the reversal. The factor 1 + δ 2 1+\delta^2 1 + δ 2 tells us in advance that the answer will involve rational functions.
The parameter recurrence
Fix x ∈ R x\in\mathbb{R} x ∈ R and define
P n ( z ) = ( − 1 ) n M n ( x ; z , η ) 2 n ( η / 2 ) n . (74) P_n(z)=\frac{(-1)^nM_n(x;z,\eta)}{2^n(\eta/2)_n}. \tag*{(74)} P n ( z ) = 2 n ( η /2 ) n ( − 1 ) n M n ( x ; z , η ) . (74)
The recurrence becomes
P n + 1 ( z ) = ( z − r n ) P n ( z ) − c n ( 1 + z 2 ) P n − 1 ( z ) , (75) P_{n+1}(z)=(z-r_n)P_n(z)-c_n(1+z^2)P_{n-1}(z), \tag*{(75)} P n + 1 ( z ) = ( z − r n ) P n ( z ) − c n ( 1 + z 2 ) P n − 1 ( z ) , (75)
where
r n = x 2 n + η , c n = n ( n + η − 1 ) ( 2 n + η − 2 ) ( 2 n + η ) ( n ≥ 1 ) . (76) r_n=\frac{x}{2n+\eta},\qquad c_n=\frac{n(n+\eta-1)}{(2n+\eta-2)(2n+\eta)}\qquad(n\ge1). \tag*{(76)} r n = 2 n + η x , c n = ( 2 n + η − 2 ) ( 2 n + η ) n ( n + η − 1 ) ( n ≥ 1 ) . (76)
Here P 0 = 1 P_0=1 P 0 = 1 and P 1 = z − r 0 P_1=z-r_0 P 1 = z − r 0 . These polynomials are not generally monic in z z z : the quadratic factor also contributes to the leading coefficient. A recurrence with a quadratic factor multiplying the preceding polynomial is of type R I I R_{II} R I I ; a linear factor would give type R I R_I R I [14 ] .
Why Wall’s fraction is the right question
The Jacobi fraction in § 8 encoded moments through an expansion at infinity. The factor 1 + z 2 1+z^2 1 + z 2 suggests a different expansion point: it vanishes at i i i and − i -i − i . In the rational orthogonality these same points will be the fixed poles of the test functions. They are natural when the real line is viewed as a circle. The Cayley map
ζ = i − z i + z , z = i 1 − ζ 1 + ζ (77) \zeta=\frac{i-z}{i+z},\qquad z=i\frac{1-\zeta}{1+\zeta} \tag*{(77)} ζ = i + z i − z , z = i 1 + ζ 1 − ζ (77)
sends the upper half-plane to the unit disc, i i i to its center, and the real line to its boundary, with infinity represented by − 1 -1 − 1 . Taylor interpolation at i i i is therefore the half-plane counterpart of Taylor interpolation at the center of the disc. The conjugate point − i -i − i enters when the approximants have real coefficients.
Here is the measure-theoretic part of the picture. A Nevanlinna function , also called a Herglotz function, is an analytic function h h h on C + \mathbb{C}_+ C + with positive imaginary part. Such a function has the representation
h ( z ) = a z + b + ∫ R 1 + t z t − z d σ ( t ) , a ≥ 0 , b ∈ R , σ ≥ 0 , σ ( R ) < ∞ . (78) h(z)=az+b+\int_{\mathbb{R}}\frac{1+tz}{t-z}\,d\sigma(t),\qquad a\geq0,\quad b\in\mathbb{R},\quad\sigma\geq0,\quad\sigma(\mathbb{R})<\infty. \tag*{(78)} h ( z ) = a z + b + ∫ R t − z 1 + t z d σ ( t ) , a ≥ 0 , b ∈ R , σ ≥ 0 , σ ( R ) < ∞. (78)
The measure here is finite. In the usual subtracted Cauchy representation the measure is d ρ ( t ) = ( 1 + t 2 ) d σ ( t ) d\rho(t)=(1+t^2)\,d\sigma(t) d ρ ( t ) = ( 1 + t 2 ) d σ ( t ) , which need not have finite mass, and the integral becomes
∫ R ( 1 t − z − t 1 + t 2 ) d ρ ( t ) . \int_{\mathbb{R}}\left(\frac{1}{t-z}-\frac{t}{1+t^2}\right)d\rho(t). ∫ R ( t − z 1 − 1 + t 2 t ) d ρ ( t ) .
The subtraction ensures convergence when the two terms cannot be integrated separately. Positivity is visible in the original kernel:
ℜ 1 + t z t − z = ( 1 + t 2 ) ℜ z ∣ t − z ∣ 2 . (79) \Re\frac{1+tz}{t-z}=\frac{(1+t^2)\Re z}{\lvert t-z\rvert^2}. \tag*{(79)} ℜ t − z 1 + t z = ∣ t − z ∣ 2 ( 1 + t 2 ) ℜ z . (79)
For completeness, the appearance of a measure can be seen directly on the disc. Put H ( ζ ) = − i h ( z ) H(\zeta)=-ih(z) H ( ζ ) = − ih ( z ) with z z z as in (77) . Then u = ℜ H u=\Re H u = ℜ H is positive and harmonic. The measures u ( r e i θ ) d θ / ( 2 π ) u(re^{i\theta})\,d\theta/(2\pi) u ( r e i θ ) d θ / ( 2 π ) have the common mass u ( 0 ) u(0) u ( 0 ) . A weakly convergent subsequence as r ↑ 1 r\uparrow1 r ↑ 1 gives a positive boundary measure μ \mu μ . The Poisson formula on each smaller disc, followed by this limit, represents u u u as its Poisson integral. Adding the harmonic conjugate gives
H ( ζ ) = i ℑ H ( 0 ) + ∫ ∣ ξ ∣ = 1 ξ + ζ ξ − ζ d μ ( ξ ) . H(\zeta)=i\Im H(0)+\int_{\lvert\xi\rvert=1}\frac{\xi+\zeta}{\xi-\zeta}\,d\mu(\xi). H ( ζ ) = i ℑ H ( 0 ) + ∫ ∣ ξ ∣ = 1 ξ − ζ ξ + ζ d μ ( ξ ) .
For ξ t = ( i − t ) / ( i + t ) \xi_t=(i-t)/(i+t) ξ t = ( i − t ) / ( i + t ) , a calculation gives
i ξ t + ζ ξ t − ζ = 1 + t z t − z . i\frac{\xi_t+\zeta}{\xi_t-\zeta}=\frac{1+tz}{t-z}. i ξ t − ζ ξ t + ζ = t − z 1 + t z .
The atom at ξ = − 1 \xi=-1 ξ = − 1 contributes a z az a z , and the rest of μ \mu μ pushes forward to σ \sigma σ on R \mathbb{R} R . This proves (10.5); the real constant is b = ℜ h ( i ) b=\Re h(i) b = ℜ h ( i ) . Uniqueness follows from the Fourier coefficients of the boundary measure, or from the Poisson formula itself.
At z = i z=i z = i , the integral kernel is exactly i i i . Thus
h ( i ) = i ⟺ b = 0 , a + σ ( R ) = 1. (80) h(i)=i \Longleftrightarrow b=0,\qquad a+\sigma(\mathbb{R})=1. \tag*{(80)} h ( i ) = i ⟺ b = 0 , a + σ ( R ) = 1. (80)
A normalized Nevanlinna function therefore represents a probability on the real line together with a possible atom at infinity. It represents a probability on R \mathbb{R} R itself when a = 0 a=0 a = 0 . The coefficient of the atom is
a = lim y → ∞ ℑ h ( i y ) y , a=\lim_{y\to\infty}\frac{\Im h(iy)}{y}, a = y → ∞ lim y ℑ h ( i y ) ,
as follows from (78) by dominated convergence. This small qualification will matter when we compare normalizations.
Wall found a continued-fraction description of this correspondence [39 , 40 , 41 ] . In the disc, a function H H H with positive real part and H ( 0 ) = 1 H(0)=1 H ( 0 ) = 1 can be written
H ( ζ ) = 1 + ζ f ( ζ ) 1 − ζ f ( ζ ) , ∣ f ( ζ ) ∣ ≤ 1. H(\zeta)=\frac{1+\zeta f(\zeta)}{1-\zeta f(\zeta)},\qquad|f(\zeta)|\le1. H ( ζ ) = 1 − ζ f ( ζ ) 1 + ζ f ( ζ ) , ∣ f ( ζ ) ∣ ≤ 1.
Indeed, ( H − 1 ) / ( H + 1 ) (H-1)/(H+1) ( H − 1 ) / ( H + 1 ) maps the disc into itself and vanishes at zero, so Schwarz’s lemma supplies the factor ζ \zeta ζ . The same argument can be repeated: if γ = f ( 0 ) \gamma=f(0) γ = f ( 0 ) and ∣ γ ∣ < 1 |\gamma|<1 ∣ γ ∣ < 1 , then
f 1 ( ζ ) = f ( ζ ) − γ ζ [ 1 − γ ‾ f ( ζ ) ] f_1(\zeta)=\frac{f(\zeta)-\gamma}{\zeta[1-\overline{\gamma}f(\zeta)]} f 1 ( ζ ) = ζ [ 1 − γ f ( ζ )] f ( ζ ) − γ
again maps the disc into its closure. Iterating this step is the Schur algorithm; if ∣ γ ∣ = 1 |\gamma|=1 ∣ γ ∣ = 1 , the maximum principle makes f f f constant and the procedure terminates. Wall reorganized the resulting fractions. After the Cayley change of variable, his canonical half-plane form is
Φ ( z ) = − 1 z − g 1 ( 1 + z 2 ) z − r 1 − ( 1 − g 1 ) g 2 ( 1 + z 2 ) z − r 2 − ( 1 − g 2 ) g 3 ( 1 + z 2 ) z − r 3 − ⋯ . (81) \Phi(z)=-\frac{1}{z-\dfrac{g_1(1+z^2)}{z-r_1-\dfrac{(1-g_1)g_2(1+z^2)}{z-r_2-\dfrac{(1-g_2)g_3(1+z^2)}{z-r_3-\cdots}}}}. \tag*{(81)} Φ ( z ) = − z − z − r 1 − z − r 2 − z − r 3 − ⋯ ( 1 − g 2 ) g 3 ( 1 + z 2 ) ( 1 − g 1 ) g 2 ( 1 + z 2 ) g 1 ( 1 + z 2 ) 1 . (81)
For real r n r_n r n and 0 < g n < 1 0<g_n<1 0 < g n < 1 , this fraction converges locally uniformly on C + \mathbb{C}_+ C + to a Nevanlinna function with Φ ( i ) = i \Phi(i)=i Φ ( i ) = i . This is the form of Wall’s theorem presented by Derevyagin [7 Theorem 3.1, formula (3.12), and Corollary 3.2] , where the associated integral representation is also displayed in Theorem 5.3. The first denominator z z z records the normalization at i i i ; the factors 1 + z 2 1+z^2 1 + z 2 record the interpolation points.
Our coefficients have exactly Wall’s chain form:
c n = ( 1 − g n − 1 ) g n , g n = n 2 n + η , g 0 = 0. (82) c_n=(1-g_{n-1})g_n,\qquad g_n=\frac{n}{2n+\eta},\qquad g_0=0. \tag*{(82)} c n = ( 1 − g n − 1 ) g n , g n = 2 n + η n , g 0 = 0. (82)
In particular c 1 = g 1 c_1=g_1 c 1 = g 1 , and 0 < g n < 1 0<g_n<1 0 < g n < 1 for every n ≥ 1 n\ge1 n ≥ 1 and every η > 0 \eta>0 η > 0 . The only initial difference is the shift r 0 r_0 r 0 in the first denominator. We shall first identify the transform with that natural shift, deriving its measure and convergence directly. Afterwards a single fractional linear transformation will put it into (81) . The general theory has now told us what to look for; the following calculations prove the needed instance.
The positive pencil
The two spectral directions can already be seen in finite matrices. Index rows by k = 0 , … , N − 1 k=0,\ldots,N-1 k = 0 , … , N − 1 , and put a k = ( k + 1 ) ( k + η ) a_k=\sqrt{(k+1)(k+\eta)} a k = ( k + 1 ) ( k + η ) . Let A N A_N A N have diagonal 2 k + η 2k+\eta 2 k + η and upper and lower entries a k a_k a k . Let B N B_N B N have zero diagonal, upper entries − i a k -ia_k − i a k , and lower entries i a k ia_k i a k . Expansion of the tridiagonal determinant gives
M N ( x ; z , η ) = det ( x I N − B N − z A N ) . (83) M_N(x;z,\eta)=\det(xI_N-B_N-zA_N). \tag*{(83)} M N ( x ; z , η ) = det ( x I N − B N − z A N ) . (83)
Moreover A N = L N L N ∗ A_N=L_NL_N^* A N = L N L N ∗ , where L N L_N L N is lower bidiagonal with diagonal k + η \sqrt{k+\eta} k + η and subdiagonal k \sqrt{k} k . Thus A N A_N A N is positive definite, while B N B_N B N is Hermitian.
For real z = δ z=\delta z = δ , the matrix B N + δ A N B_N+\delta A_N B N + δ A N is Hermitian. A diagonal unitary change of basis makes its off-diagonal entries a k 1 + δ 2 a_k\sqrt{1+\delta^2} a k 1 + δ 2 , recovering the Jacobi matrix of § 8 . For fixed real x x x , the same determinant describes the generalized eigenvalue problem
( x I N − B N ) v = z A N v . (84) (xI_N-B_N)v=zA_Nv. \tag*{(84)} ( x I N − B N ) v = z A N v . (84)
Its eigenvalues are real because A N > 0 A_N>0 A N > 0 : conjugating by A N − 1 / 2 A_N^{-1/2} A N − 1/2 gives an ordinary Hermitian problem. In particular, the zeros of M N M_N M N in the parameter are real, just as its zeros in the argument are real. Related unit-circle and pencil constructions are developed in [7 , 16 ] .
A Rodrigues formula in the parameter
Keep x x x fixed and write
s ( t ) = 1 + t 2 , w ( t ) = s ( t ) − η / 2 e x arctan t , K η ( x , t ) = C η ( x ) w ( t ) . (85) s(t)=1+t^2,\qquad w(t)=s(t)^{-\eta/2}e^{x\arctan t},\qquad K_\eta(x,t)=C_\eta(x)w(t). \tag*{(85)} s ( t ) = 1 + t 2 , w ( t ) = s ( t ) − η /2 e x a r c t a n t , K η ( x , t ) = C η ( x ) w ( t ) . (85)
The logarithmic derivative is remarkably simple:
s ( t ) w ′ ( t ) = ( x − η t ) w ( t ) . (86) s(t)w'(t)=(x-\eta t)w(t). \tag*{(86)} s ( t ) w ′ ( t ) = ( x − η t ) w ( t ) . (86)
The generating function gives a Rodrigues formula without any guesswork. Taylor-expand w ( t + u s ( t ) ) / w ( t ) w(t+us(t))/w(t) w ( t + u s ( t )) / w ( t ) . The elementary identities
1 + ( t + u s ( t ) ) 2 1 + t 2 = 1 + 2 t u + ( 1 + t 2 ) u 2 , arctan ( t + u s ( t ) ) − arctan t = arctan u 1 + t u \frac{1+(t+us(t))^2}{1+t^2}=1+2tu+(1+t^2)u^2,\qquad\arctan(t+us(t))-\arctan t=\arctan\frac{u}{1+tu} 1 + t 2 1 + ( t + u s ( t ) ) 2 = 1 + 2 t u + ( 1 + t 2 ) u 2 , arctan ( t + u s ( t )) − arctan t = arctan 1 + t u u
hold near u = 0 u=0 u = 0 . Consequently this Taylor series is exactly G ( x ; u ) G(x;u) G ( x ; u ) with δ = t \delta=t δ = t , and
M n ( x ; t , η ) = s ( t ) n w ( t ) d n w ( t ) d t n . (87) M_n(x;t,\eta)=\frac{s(t)^n}{w(t)}\frac{d^n w(t)}{dt^n}. \tag*{(87)} M n ( x ; t , η ) = w ( t ) s ( t ) n d t n d n w ( t ) . (87)
These are the complementary Romanovski–Routh polynomials in the normalization of our recurrence [23 ] . The identification explains the family; the calculation proves the formula in the conventions used here.
The Rodrigues formula also gives a lowering operation in the parameter. Differentiate it, use (86) , and recognize the next Rodrigues expression:
∂ t M n = M n + 1 − [ x − ( η + 2 n ) t ] M n 1 + t 2 . \partial_tM_n=\frac{M_{n+1}-[x-(\eta+2n)t]M_n}{1+t^2}. ∂ t M n = 1 + t 2 M n + 1 − [ x − ( η + 2 n ) t ] M n .
The recurrence therefore reduces the derivative to
∂ δ M n = − n ( n + η − 1 ) M n − 1 . (88) \partial_\delta M_n=-n(n+\eta-1)M_{n-1}. \tag*{(88)} ∂ δ M n = − n ( n + η − 1 ) M n − 1 . (88)
Thus ( − 1 ) n M n / ( η ) n (-1)^n M_n/(\eta)_n ( − 1 ) n M n / ( η ) n is an Appell sequence in the parameter. Differentiate once more and use the recurrence at index n − 1 n-1 n − 1 to obtain
( 1 + δ 2 ) ∂ δ 2 M n + [ x − ( η + 2 n − 2 ) δ ] ∂ δ M n + n ( n + η − 1 ) M n = 0. (89) (1+\delta^2)\partial_\delta^2 M_n+[x-(\eta+2n-2)\delta]\partial_\delta M_n+n(n+\eta-1)M_n=0. \tag*{(89)} ( 1 + δ 2 ) ∂ δ 2 M n + [ x − ( η + 2 n − 2 ) δ ] ∂ δ M n + n ( n + η − 1 ) M n = 0. (89)
For n ≥ 2 n\geq2 n ≥ 2 , if c = n ( n + η − 1 ) c=n(n+\eta-1) c = n ( n + η − 1 ) and e = ( n − 1 ) ( n + η − 2 ) e=(n-1)(n+\eta-2) e = ( n − 1 ) ( n + η − 2 ) , the left side is
c { ( 1 + δ 2 ) e M n − 2 − [ x − ( η + 2 n − 2 ) δ ] M n − 1 + M n } = 0. c\{(1+\delta^2)eM_{n-2}-[x-(\eta+2n-2)\delta]M_{n-1}+M_n\}=0. c {( 1 + δ 2 ) e M n − 2 − [ x − ( η + 2 n − 2 ) δ ] M n − 1 + M n } = 0.
The cases n = 0 , 1 n=0,1 n = 0 , 1 follow directly. This also derives the complementary Romanovski–Routh differential equation in our normalization.
Integration by parts now gives rational orthogonality:
∫ R t k M n ( x ; t , η ) ( 1 + t 2 ) n K η ( x , t ) d t = 0 , 0 ≤ k < n , η > 0. (90) \int_{\mathbb{R}}\frac{t^kM_n(x;t,\eta)}{(1+t^2)^n}K_\eta(x,t)\,\mathrm{d}t=0,\qquad0\leq k<n,\quad\eta>0. \tag*{(90)} ∫ R ( 1 + t 2 ) n t k M n ( x ; t , η ) K η ( x , t ) d t = 0 , 0 ≤ k < n , η > 0. (90)
Indeed, the integral is C η ( x ) ∫ t k w ( n ) ( t ) d t C_\eta(x)\int t^k w^{(n)}(t)\,\mathrm{d}t C η ( x ) ∫ t k w ( n ) ( t ) d t . At either endpoint w ( j ) ( t ) = O ( ∣ t ∣ − η − j ) w^{(j)}(t)=O(|t|^{-\eta-j}) w ( j ) ( t ) = O ( ∣ t ∣ − η − j ) . Every boundary term after integration by parts is therefore O ( ∣ t ∣ k − n + 1 − η ) O(|t|^{k-n+1-\eta}) O ( ∣ t ∣ k − n + 1 − η ) , which tends to zero for k < n k<n k < n . After n n n integrations the derivative of t k t^k t k vanishes. The integrand itself is absolutely integrable, since its growth is O ( ∣ t ∣ k − n − η ) O(|t|^{k-n-\eta}) O ( ∣ t ∣ k − n − η ) .
This is the promised second orthogonality. It is not ordinary polynomial orthogonality for a fixed measure with all moments: the denominator grows with the degree. That denominator is exactly what makes the heavy-tailed parameter measure useful at every degree.
Two normalization integrals
We will need the mass of the parameter weight and its mass after division by 1 + t 2 1+t^2 1 + t 2 . Both follow from one Fourier calculation. For ℜ a > 1 / 2 \Re a>1/2 ℜ a > 1/2 , the gamma integral gives the Fourier transform
∫ 0 ∞ y a − 1 e − y Γ ( a ) e − i t y d y = ( 1 + i t ) − a . \int_0^\infty\frac{y^{a-1}e^{-y}}{\Gamma(a)}e^{-ity}\,\mathrm{d}y=(1+it)^{-a}. ∫ 0 ∞ Γ ( a ) y a − 1 e − y e − i t y d y = ( 1 + i t ) − a .
Plancherel’s identity and another gamma integral therefore give
∫ R ∣ ( 1 + i t ) − a ∣ 2 d t = 2 π ∣ Γ ( a ) ∣ 2 ∫ 0 ∞ y 2 ℜ a − 2 e − 2 y d y = π Γ ( 2 ℜ a − 1 ) 2 2 ℜ a − 2 ∣ Γ ( a ) ∣ 2 . (91) \int_{\mathbb{R}}|(1+it)^{-a}|^2\,\mathrm{d}t
=\frac{2\pi}{|\Gamma(a)|^2}\int_0^\infty y^{2\Re a-2}e^{-2y}\,\mathrm{d}y
=\frac{\pi\Gamma(2\Re a-1)}{2^{2\Re a-2}|\Gamma(a)|^2}. \tag*{(91)} ∫ R ∣ ( 1 + i t ) − a ∣ 2 d t = ∣Γ ( a ) ∣ 2 2 π ∫ 0 ∞ y 2ℜ a − 2 e − 2 y d y = 2 2ℜ a − 2 ∣Γ ( a ) ∣ 2 π Γ ( 2ℜ a − 1 ) . (91)
This is a beta integral in Fourier form [29 §5.12] . Choose a = ( η + i x ) / 2 a=(\eta+ix)/2 a = ( η + i x ) /2 for the first identity below, and a = 1 + ( η + i x ) / 2 a=1+(\eta+ix)/2 a = 1 + ( η + i x ) /2 for the second. Multiplying by C η ( x ) C_\eta(x) C η ( x ) and cancelling gamma factors yields
∫ R K η ( x , t ) d t = 1 η − 1 , η > 1. (92) \int_{\mathbb{R}}K_\eta(x,t)\,\mathrm{d}t=\frac{1}{\eta-1},\qquad\eta>1. \tag*{(92)} ∫ R K η ( x , t ) d t = η − 1 1 , η > 1. (92)
∫ R K η ( x , t ) 1 + t 2 d t = η x 2 + η 2 , η > 0. (93) \int_{\mathbb{R}}\frac{K_\eta(x,t)}{1+t^2}\,\mathrm{d}t=\frac{\eta}{x^2+\eta^2},\qquad\eta>0. \tag*{(93)} ∫ R 1 + t 2 K η ( x , t ) d t = x 2 + η 2 η , η > 0. (93)
The first integral is infinite for 0 < η ≤ 1 0<\eta\leq1 0 < η ≤ 1 , since w ( t ) w(t) w ( t ) has tails proportional to ∣ t ∣ − η |t|^{-\eta} ∣ t ∣ − η . When η > 1 \eta>1 η > 1 , we may introduce the probability measure
d ν x ( t ) = ( η − 1 ) K η ( x , t ) d t . (94) \mathrm{d}\nu_x(t)=(\eta-1)K_\eta(x,t)\,\mathrm{d}t. \tag*{(94)} d ν x ( t ) = ( η − 1 ) K η ( x , t ) d t . (94)
The preceding integration by parts then also includes the endpoint k = n k=n k = n :
∫ R t k M n ( x ; t , η ) ( 1 + t 2 ) n d ν x ( t ) = ( − 1 ) n n ! δ k n , 0 ≤ k ≤ n . (95) \int_{\mathbb{R}}\frac{t^kM_n(x;t,\eta)}{(1+t^2)^n}\,\mathrm{d}\nu_x(t)=(-1)^n n!\,\delta_{kn},\qquad0\leq k\leq n. \tag*{(95)} ∫ R ( 1 + t 2 ) n t k M n ( x ; t , η ) d ν x ( t ) = ( − 1 ) n n ! δ k n , 0 ≤ k ≤ n . (95)
The parameter fraction and its exact measure
For z z z in the upper half-plane C + \mathbb{C}_{+} C + , define
F x ( z ) = η ∫ R K η ( x , t ) t − z d t . (96) F_x(z)=\eta\int_{\mathbb{R}}\frac{K_\eta(x,t)}{t-z}\,dt. \tag*{(96)} F x ( z ) = η ∫ R t − z K η ( x , t ) d t . (96)
This converges for every η > 0 \eta>0 η > 0 , even when the measure in its numerator has infinite total mass: the integrand is O ( ∣ t ∣ − η − 1 ) O(|t|^{-\eta-1}) O ( ∣ t ∣ − η − 1 ) . Moreover
ℑ F x ( z ) = η ℑ z ∫ R K η ( x , t ) ∣ t − z ∣ 2 d t > 0. (97) \Im F_x(z)=\eta\Im z\int_{\mathbb{R}}\frac{K_\eta(x,t)}{|t-z|^2}\,dt>0. \tag*{(97)} ℑ F x ( z ) = η ℑ z ∫ R ∣ t − z ∣ 2 K η ( x , t ) d t > 0. (97)
Thus F x F_x F x is a Nevanlinna (or Herglotz) function.
Theorem 10.1. For every real x x x and η > 0 \eta>0 η > 0 ,
F x ( z ) = − 1 z − r 0 − c 1 ( 1 + z 2 ) z − r 1 − c 2 ( 1 + z 2 ) z − r 2 − ⋯ , z ∈ C + . (98) F_x(z)=-\frac{1}{z-r_0-\cfrac{c_1(1+z^2)}{z-r_1-\cfrac{c_2(1+z^2)}{z-r_2-\cdots}}},\qquad z\in\mathbb{C}_{+}. \tag*{(98)} F x ( z ) = − z − r 0 − z − r 1 − z − r 2 − ⋯ c 2 ( 1 + z 2 ) c 1 ( 1 + z 2 ) 1 , z ∈ C + . (98) The finite convergents converge locally uniformly on C + \mathbb{C}_{+} C + to the integral (96) .
Proof. We first find an equation that characterizes the integral. Put I j = ∫ t j w ( t ) / ( 1 + t 2 ) d t I_j=\int t^j w(t)/(1+t^2)\,dt I j = ∫ t j w ( t ) / ( 1 + t 2 ) d t for j = 0 , 1 j=0,1 j = 0 , 1 and κ = η C η ( x ) \kappa=\eta C_\eta(x) κ = η C η ( x ) . Integrating (86) after division by s ( t ) s(t) s ( t ) gives x I 0 = η I 1 xI_0=\eta I_1 x I 0 = η I 1 , since w w w vanishes at both endpoints. Formula (93) gives κ I 0 = η 2 / ( x 2 + η 2 ) \kappa I_0=\eta^2/(x^2+\eta^2) κ I 0 = η 2 / ( x 2 + η 2 ) . In particular,
F x ( i ) = κ ( I 1 + i I 0 ) = η x − i η . (99) F_x(i)=\kappa(I_1+iI_0)=\frac{\eta}{x-i\eta}. \tag*{(99)} F x ( i ) = κ ( I 1 + i I 0 ) = x − i η η . (99)
Integration by parts in w ( t ) / ( t − z ) w(t)/(t-z) w ( t ) / ( t − z ) gives F x ′ ( z ) = κ ∫ w ′ ( t ) / ( t − z ) d t F_x'(z)=\kappa\int w'(t)/(t-z)\,dt F x ′ ( z ) = κ ∫ w ′ ( t ) / ( t − z ) d t . Substituting (86) and cancelling t − z t-z t − z gives
( 1 + z 2 ) F x ′ ( z ) − ( x − η z ) F x ( z ) = − κ [ η I 0 + x I 1 + z ( x I 0 − η I 1 ) ] = − η . (1+z^2)F_x'(z)-(x-\eta z)F_x(z)=-\kappa[\eta I_0+xI_1+z(xI_0-\eta I_1)]
=-\eta. ( 1 + z 2 ) F x ′ ( z ) − ( x − η z ) F x ( z ) = − κ [ η I 0 + x I 1 + z ( x I 0 − η I 1 )] = − η .
All boundary terms vanish for η > 0 \eta>0 η > 0 . Thus
( 1 + z 2 ) F x ′ ( z ) = ( x − η z ) F x ( z ) − η . (100) (1+z^2)F_x'(z)=(x-\eta z)F_x(z)-\eta. \tag*{(100)} ( 1 + z 2 ) F x ′ ( z ) = ( x − η z ) F x ( z ) − η . (100)
Differentiate n n n times, for n ≥ 1 n\geq1 n ≥ 1 :
( 1 + z 2 ) F x ( n + 1 ) = [ x − ( η + 2 n ) z ] F x ( n ) − n ( n + η − 1 ) F x ( n − 1 ) . (1+z^2)F_x^{(n+1)}=[x-(\eta+2n)z]F_x^{(n)}-n(n+\eta-1)F_x^{(n-1)}. ( 1 + z 2 ) F x ( n + 1 ) = [ x − ( η + 2 n ) z ] F x ( n ) − n ( n + η − 1 ) F x ( n − 1 ) .
It follows that U n = ( 1 + z 2 ) n F x ( n ) U_n=(1+z^2)^nF_x^{(n)} U n = ( 1 + z 2 ) n F x ( n ) satisfies the original Meixner recurrence in n n n . If N 0 = 0 N_0=0 N 0 = 0 , N 1 = 1 N_1=1 N 1 = 1 , and N n N_n N n satisfies that recurrence with δ = z \delta=z δ = z for n ≥ 1 n\geq1 n ≥ 1 , the initial values U 0 = F x U_0=F_x U 0 = F x , U 1 = M 1 F x − η U_1=M_1F_x-\eta U 1 = M 1 F x − η give the exact identity
M n ( x ; z , η ) F x ( z ) − η N n ( z ) = ( 1 + z 2 ) n F x ( n ) ( z ) . (101) M_n(x;z,\eta)F_x(z)-\eta N_n(z)=(1+z^2)^nF_x^{(n)}(z). \tag*{(101)} M n ( x ; z , η ) F x ( z ) − η N n ( z ) = ( 1 + z 2 ) n F x ( n ) ( z ) . (101)
The quotient F x , n = η N n / M n F_{x,n}=\eta N_n/M_n F x , n = η N n / M n is the n n n -level fraction in (98) ; this follows by the ordinary continuant recurrence, or by taking the first Schur complement in (83) .
It remains to prove convergence. The same Schur complement gives
F x , N ( z ) = η ⟨ e 0 , ( x I N − B N − z A N ) − 1 e 0 ⟩ . (102) F_{x,N}(z)=\eta\langle e_0,(xI_N-B_N-zA_N)^{-1}e_0\rangle. \tag*{(102)} F x , N ( z ) = η ⟨ e 0 , ( x I N − B N − z A N ) − 1 e 0 ⟩ . (102)
If H = x I N − B N − z A N H=xI_N-B_N-zA_N H = x I N − B N − z A N , then ℑ H = − ( ℑ z ) A N < 0 \Im H=-(\Im z)A_N<0 ℑ H = − ( ℑ z ) A N < 0 , so H H H is invertible and ℑ H − 1 = ( ℑ z ) H − ∗ A N H − 1 > 0 \Im H^{-1}=(\Im z)H^{-*}A_NH^{-1}>0 ℑ H − 1 = ( ℑ z ) H −∗ A N H − 1 > 0 . Thus each of these functions has positive imaginary part in C + \mathbb{C}_+ C + . At z = i z=i z = i the matrix is lower triangular, and
F x , N ( i ) = η x − i η = F x ( i ) . F_{x,N}(i)=\frac{\eta}{x-i\eta}=F_x(i). F x , N ( i ) = x − i η η = F x ( i ) .
Herglotz functions with this fixed value form a normal family: map their common value to zero by a half-plane-to-disc map and apply Schwarz’s lemma, followed by Montel’s theorem. The fixed interior value also prevents a subsequence from tending to infinity.
Finally,
M n ( x ; i , η ) = ∏ j = 0 n − 1 [ x − i ( η + 2 j ) ] ≠ 0. M_n(x;i,\eta)=\prod_{j=0}^{n-1}[x-i(\eta+2j)]\ne0. M n ( x ; i , η ) = j = 0 ∏ n − 1 [ x − i ( η + 2 j )] = 0.
Equation (101) shows that F x , n F_{x,n} F x , n and F x F_x F x have the same first n n n Taylor coefficients at i i i . Every locally uniform subsequential limit therefore agrees with F x F_x F x to every order at i i i . The identity theorem identifies that limit with F x F_x F x throughout C + \mathbb{C}_+ C + . Since every subsequential limit is the same, the whole sequence converges.
This proof makes the role of positivity quite concrete. The recurrence alone produces rational approximants; the positive pencil makes them a controlled analytic family; the differential identity selects their unique limit.
Expansion of (100) at z = i z=i z = i also gives
F x ( z ) = η x − i η 2 F 1 ( 1 , η ; 1 + η + i x 2 ; 1 + i z 2 ) . (103) F_x(z)=\frac{\eta}{x-i\eta}\,{}_2F_1\left(1,\eta;1+\frac{\eta+ix}{2};\frac{1+iz}{2}\right). \tag*{(103)} F x ( z ) = x − i η η 2 F 1 ( 1 , η ; 1 + 2 η + i x ; 2 1 + i z ) . (103)
For example, with q = ( 1 + i z ) / 2 q=(1+iz)/2 q = ( 1 + i z ) /2 and F x = ∑ j ≥ 0 a j q j F_x=\sum_{j\geq0}a_jq^j F x = ∑ j ≥ 0 a j q j , the equation gives a 0 = η / ( x − i η ) a_0=\eta/(x-i\eta) a 0 = η / ( x − i η ) and a j + 1 / a j = ( η + j ) / ( 1 + ( η + i x ) / 2 + j ) a_{j+1}/a_j=(\eta+j)/(1+(\eta+ix)/2+j) a j + 1 / a j = ( η + j ) / ( 1 + ( η + i x ) /2 + j ) . This is precisely the displayed hypergeometric series. Analytic continuation covers C + \mathbb{C}_+ C + because ℜ q < 1 / 2 \Re q<1/2 ℜ q < 1/2 there.
The match with (62) is now exact: the same finite continued fraction, viewed in x x x , is an ordinary Jacobi fraction; viewed in δ \delta δ , it is 1 / η 1/\eta 1/ η times this Wall fraction. The lower-half-plane branch in the first variable leads to the upper-half-plane Nevanlinna function in the second. The distinction between the two branches is essential when identifying their boundary measures.
One can make this analytic connection without merely comparing formulas. The function
C ( ζ , z ) = i ∫ 0 ∞ e − i ζ u ( cosh u − i z sinh u ) − η d u (104) C(\zeta,z)=i\int_0^\infty e^{-i\zeta u}(\cosh u-iz\sinh u)^{-\eta}\,du \tag*{(104)} C ( ζ , z ) = i ∫ 0 ∞ e − i ζ u ( cosh u − i z sinh u ) − η d u (104)
is jointly holomorphic for ℑ ζ < η \Im\zeta<\eta ℑ ζ < η , ℑ z > − 1 \Im z>-1 ℑ z > − 1 . Indeed, the expression in parentheses has positive real part and, locally uniformly in this domain, its modulus is bounded below by a positive constant times cosh u \cosh u cosh u . We use its principal power. Equation (64) identifies C ( ζ , δ ) = C δ ( ζ ) C(\zeta,\delta)=C_\delta(\zeta) C ( ζ , δ ) = C δ ( ζ ) for real δ \delta δ and ℑ ζ < 0 \Im\zeta<0 ℑ ζ < 0 ; the hypergeometric formulas identify η C ( x , z ) = F x ( z ) \eta C(x,z)=F_x(z) η C ( x , z ) = F x ( z ) for real x x x and z ∈ C + z\in\mathbb{C}_+ z ∈ C + . In particular, the compatible boundary values satisfy
η C δ ( x − i 0 ) = F x ( δ + i 0 ) , ℑ C δ ( x − i 0 ) = π K η ( x , δ ) . \eta C_\delta(x-i0)=F_x(\delta+i0),\qquad\Im C_\delta(x-i0)=\pi K_\eta(x,\delta). η C δ ( x − i 0 ) = F x ( δ + i 0 ) , ℑ C δ ( x − i 0 ) = π K η ( x , δ ) .
The upper x x x boundary is the complex conjugate. Thus the same analytic continuation carries the two spectral readings, with their boundary branches explicitly specified.
Two measures, or one measure in two variables?
Both descriptions are useful. One positive kernel gives the two spectral realizations, and a Cauchy weighting turns them into the conditional laws of one joint probability. The factor 1 + t 2 1+t^2 1 + t 2 determines which parameter measure appears: a Cauchy integral and a normalized Nevanlinna integral use closely related, but different, measures.
The Cauchy measure and the finite Nevanlinna measure
The measure in (96) is
d ρ x ( t ) = η K η ( x , t ) d t . (105) d\rho_x(t)=\eta K_\eta(x,t)\,dt. \tag*{(105)} d ρ x ( t ) = η K η ( x , t ) d t . (105)
Its mass is η / ( η − 1 ) \eta/(\eta-1) η / ( η − 1 ) when η > 1 \eta>1 η > 1 and infinite when 0 < η ≤ 1 0<\eta\le1 0 < η ≤ 1 . Nevertheless the Cauchy integral is absolutely convergent for every η > 0 \eta>0 η > 0 .
For the Nevanlinna kernel, divide this measure by 1 + t 2 1+t^2 1 + t 2 :
d σ x ( t ) = d ρ x ( t ) 1 + t 2 = η K η ( x , t ) 1 + t 2 d t . (106) d\sigma_x(t)=\frac{d\rho_x(t)}{1+t^2}=\eta\frac{K_\eta(x,t)}{1+t^2}\,dt. \tag*{(106)} d σ x ( t ) = 1 + t 2 d ρ x ( t ) = η 1 + t 2 K η ( x , t ) d t . (106)
The identity
1 t − z = 1 1 + t 2 ( 1 + t z t − z + t ) \frac{1}{t-z}=\frac{1}{1+t^2}\left(\frac{1+tz}{t-z}+t\right) t − z 1 = 1 + t 2 1 ( t − z 1 + t z + t )
and the value (99) give
F x ( z ) = η x x 2 + η 2 + ∫ R 1 + t z t − z d σ x ( t ) , σ x ( R ) = η 2 x 2 + η 2 . (107) F_x(z)=\frac{\eta x}{x^2+\eta^2}+\int_{\mathbb{R}}\frac{1+tz}{t-z}\,d\sigma_x(t),\qquad\sigma_x(\mathbb{R})=\frac{\eta^2}{x^2+\eta^2}. \tag*{(107)} F x ( z ) = x 2 + η 2 η x + ∫ R t − z 1 + t z d σ x ( t ) , σ x ( R ) = x 2 + η 2 η 2 . (107)
There is no linear term. More directly than an appeal to the general representation theorem, dominated convergence gives F x ( i y ) → 0 F_x(iy)\to0 F x ( i y ) → 0 as y → ∞ y\to\infty y → ∞ : for y ≥ 1 y\ge1 y ≥ 1 , ∣ t − i y ∣ − 1 ≤ ( 1 + t 2 ) − 1 / 2 \lvert t-iy\rvert^{-1}\le(1+t^2)^{-1/2} ∣ t − i y ∣ − 1 ≤ ( 1 + t 2 ) − 1/2 , and K η ( x , t ) ( 1 + t 2 ) − 1 / 2 K_\eta(x,t)(1+t^2)^{-1/2} K η ( x , t ) ( 1 + t 2 ) − 1/2 is integrable for every η > 0 \eta>0 η > 0 .
To normalize at i i i , set
φ x ( z ) = x 2 + η 2 η 2 F x ( z ) − x η . (108) \varphi_x(z)=\frac{x^2+\eta^2}{\eta^2}F_x(z)-\frac{x}{\eta}. \tag*{(108)} φ x ( z ) = η 2 x 2 + η 2 F x ( z ) − η x . (108)
Then φ x ( i ) = i \varphi_x(i)=i φ x ( i ) = i and
φ x ( z ) = ∫ R 1 + t z t − z d σ ^ x ( t ) , d σ ^ x ( t ) = x 2 + η 2 η K η ( x , t ) 1 + t 2 d t . (109) \varphi_x(z)=\int_{\mathbb{R}}\frac{1+tz}{t-z}\,d\widehat{\sigma}_x(t),\qquad d\widehat{\sigma}_x(t)=\frac{x^2+\eta^2}{\eta}\frac{K_\eta(x,t)}{1+t^2}\,dt. \tag*{(109)} φ x ( z ) = ∫ R t − z 1 + t z d σ x ( t ) , d σ x ( t ) = η x 2 + η 2 1 + t 2 K η ( x , t ) d t . (109)
The measure σ ^ x \widehat{\sigma}_x σ x is a probability for all η > 0 \eta>0 η > 0 . Gamma cancellation in (15) also gives
d σ ^ x ( t ) = ( η + 1 ) K η + 2 ( x , t ) d t . (110) d\widehat{\sigma}_x(t)=(\eta+1)K_{\eta+2}(x,t)\,dt. \tag*{(110)} d σ x ( t ) = ( η + 1 ) K η + 2 ( x , t ) d t . (110)
Thus this normalized probability is the unweighted parameter probability with the shape parameter shifted by two. For the original polynomials, its rational orthogonality is equivalently
∫ R t k M n ( x ; t , η ) ( 1 + t 2 ) n − 1 d σ ^ x ( t ) = 0 , n ≥ 1 , 0 ≤ k < n . (111) \int_{\mathbb{R}}\frac{t^kM_n(x;t,\eta)}{(1+t^2)^{n-1}}\,d\widehat{\sigma}_x(t)=0,\qquad n\ge1,\quad0\le k<n. \tag*{(111)} ∫ R ( 1 + t 2 ) n − 1 t k M n ( x ; t , η ) d σ x ( t ) = 0 , n ≥ 1 , 0 ≤ k < n . (111)
The canonical Wall normalization
We can now return to Wall’s canonical fraction (81) . Its first denominator is z z z ; the natural Meixner fraction (98) starts with z − r 0 z-r_0 z − r 0 , where r 0 = x / η r_0=x/\eta r 0 = x / η . Removing just that shift means adding r 0 r_0 r 0 to the reciprocal denominator:
− 1 W x ( z ) = − 1 F x ( z ) + r 0 . -\frac{1}{W_x(z)}=-\frac{1}{F_x(z)}+r_0. − W x ( z ) 1 = − F x ( z ) 1 + r 0 .
Solving this equation derives the required transformation:
W x ( z ) = F x ( z ) 1 − ( x / η ) F x ( z ) = − 1 z − c 1 ( 1 + z 2 ) z − r 1 − c 2 ( 1 + z 2 ) z − r 2 − ⋯ . (112) W_x(z)=\frac{F_x(z)}{1-(x/\eta)F_x(z)}=-\cfrac{1}{z-\cfrac{c_1(1+z^2)}{z-r_1-\cfrac{c_2(1+z^2)}{z-r_2-\cdots}}}. \tag*{(112)} W x ( z ) = 1 − ( x / η ) F x ( z ) F x ( z ) = − z − z − r 1 − z − r 2 − ⋯ c 2 ( 1 + z 2 ) c 1 ( 1 + z 2 ) 1 . (112)
The same identity holds for every finite convergent, so the local uniform convergence proved in Theorem 10.1 also justifies this infinite fraction. It retains the entire continued-fraction tail. The denominator in the fractional linear transformation cannot vanish on C + \mathbb{C}_+ C + , since F x F_x F x has strictly positive imaginary part. Moreover
ℑ W x ( z ) = ℑ F x ( z ) ∣ 1 − ( x / η ) F x ( z ) ∣ 2 > 0 , W x ( i ) = i . \Im W_x(z)=\frac{\Im F_x(z)}{\left|1-(x/\eta)F_x(z)\right|^2}>0,\qquad W_x(i)=i. ℑ W x ( z ) = ∣ 1 − ( x / η ) F x ( z ) ∣ 2 ℑ F x ( z ) > 0 , W x ( i ) = i .
The latter identity follows at once from F x ( i ) = η / ( x − i η ) F_x(i)=\eta/(x-i\eta) F x ( i ) = η / ( x − i η ) . Thus this is precisely the canonical normalization described by Wall and Derevyagin, for every η > 0 \eta>0 η > 0 .
There are now two natural ways to normalize at i i i . The affine transformation (108) merely multiplies the representing measure by a constant, producing the explicit probability σ ~ x \widetilde{\sigma}_x σ x . The fractional linear transformation (112) instead preserves the continued-fraction tail. These operations generally give different functions and different measures. For the latter, the representing probability has density
d σ W x d t ( t ) = η K η ( x , t ) ( 1 + t 2 ) ∣ 1 − ( x / η ) F x ( t + i 0 ) ∣ 2 . (113) \frac{d\sigma_{W_x}}{dt}(t)=\frac{\eta K_\eta(x,t)}{(1+t^2)\left|1-(x/\eta)F_x(t+i0)\right|^2}. \tag*{(113)} d t d σ W x ( t ) = ( 1 + t 2 ) ∣ 1 − ( x / η ) F x ( t + i 0 ) ∣ 2 η K η ( x , t ) . (113)
To verify it, combine the formula for ℑ W x \Im W_x ℑ W x with the Cauchy boundary formula ℑ F x ( t + i 0 ) = π η K η ( x , t ) \Im F_x(t+i0)=\pi\eta K_\eta(x,t) ℑ F x ( t + i 0 ) = π η K η ( x , t ) and divide the Cauchy density by 1 + t 2 1+t^2 1 + t 2 . The hypergeometric expression, or the differential equation away from ± i \pm i ± i , continues the upper branch of F x F_x F x analytically through each real point. Its strictly positive boundary imaginary part prevents the denominator in (113) from vanishing there. Stieltjes inversion on each compact interval therefore leaves no singular component on the real line. Since F x ( i y ) → 0 F_x(iy)\to0 F x ( i y ) → 0 , also W x ( i y ) → 0 W_x(iy)\to0 W x ( i y ) → 0 , so the coefficient a a a in (78) is zero. The normalization at i i i then proves that (113) has mass one.
This identifies the measure of the canonical fraction as well as that of the original Meixner fraction. It also explains why the restriction η > 1 \eta>1 η > 1 belongs only to the unweighted probability ν x \nu_x ν x : both normalized Nevanlinna probabilities exist for every η > 0 \eta>0 η > 0 .
A joint law with Cauchy marginals
There is a particularly simple way to put both spectral directions in one probability space.
Theorem 11.1. For every η > 0 \eta>0 η > 0 , the density
d Π η ( x , δ ) = K η ( x , δ ) π ( 1 + δ 2 ) d x d δ (114) d\Pi_\eta(x,\delta)=\frac{K_\eta(x,\delta)}{\pi(1+\delta^2)}\,dx\,d\delta \tag*{(114)} d Π η ( x , δ ) = π ( 1 + δ 2 ) K η ( x , δ ) d x d δ (114)
defines a probability measure. Its marginals are
d Π η δ ( δ ) = d δ π ( 1 + δ 2 ) , d Π η x ( x ) = η d x π ( x 2 + η 2 ) . (115) d\Pi_{\eta}^{\delta}(\delta)=\frac{d\delta}{\pi(1+\delta^2)},\qquad d\Pi_{\eta}^{x}(x)=\frac{\eta\,dx}{\pi(x^2+\eta^2)}. \tag*{(115)} d Π η δ ( δ ) = π ( 1 + δ 2 ) d δ , d Π η x ( x ) = π ( x 2 + η 2 ) η d x . (115)
Its conditional laws are
Π η ( d x ∣ δ ) = d μ δ ( x ) , Π η ( d δ ∣ x ) = d σ ^ x ( δ ) . (116) \Pi_{\eta}(dx\mid\delta)=d\mu_{\delta}(x),\qquad\Pi_{\eta}(d\delta\mid x)=d\widehat{\sigma}_{x}(\delta). \tag*{(116)} Π η ( d x ∣ δ ) = d μ δ ( x ) , Π η ( d δ ∣ x ) = d σ x ( δ ) . (116)
Proof. Integrate first in x x x . The kernel has mass one, so the first marginal and the total mass follow from Tonelli’s theorem. Integrate first in δ \delta δ and use (93) to get the second marginal. Dividing the joint density by either marginal gives the two conditional densities.
This gives a literal answer to the question posed at the start. The ordinary Meixner measure and the affine-normalized Nevanlinna measure are the two conditional laws of one joint distribution. The first marginal is standard Cauchy, and the second is Cauchy with scale η \eta η . In the angle δ = tan θ \delta=\tan\theta δ = tan θ , the first marginal is simply uniform on ( − π / 2 , π / 2 ) (-\pi/2,\pi/2) ( − π /2 , π /2 ) .
The unweighted probability ν x \nu_x ν x , when η > 1 \eta>1 η > 1 , has a slightly different role. It and μ δ \mu_\delta μ δ are normalized slices of the same kernel:
d ν x d δ = ( η − 1 ) K η ( x , δ ) , d μ δ d x = K η ( x , δ ) . \frac{d\nu_x}{d\delta}=(\eta-1)K_\eta(x,\delta),\qquad\frac{d\mu_\delta}{dx}=K_\eta(x,\delta). d δ d ν x = ( η − 1 ) K η ( x , δ ) , d x d μ δ = K η ( x , δ ) .
The unweighted two-variable measure K η ( x , δ ) d x d δ K_\eta(x,\delta)\,dx\,d\delta K η ( x , δ ) d x d δ has infinite total mass. The Cauchy reference factor in (114) turns it into a probability and converts the second slice into σ ^ x \widehat{\sigma}_x σ x . Orthogonality in each variable does not by itself supply a joint probability law; the Cauchy reference factor supplies the required compatibility.
As a useful check, at η = 2 \eta=2 η = 2 and x = 0 x=0 x = 0 ,
K 2 ( 0 , t ) = 1 π ( 1 + t 2 ) , F 0 ( z ) = − 2 z + i , d σ ^ 0 ( t ) = 2 d t π ( 1 + t 2 ) 2 . K_2(0,t)=\frac{1}{\pi(1+t^2)},\qquad F_0(z)=-\frac{2}{z+i},\qquad d\widehat{\sigma}_0(t)=\frac{2\,dt}{\pi(1+t^2)^2}. K 2 ( 0 , t ) = π ( 1 + t 2 ) 1 , F 0 ( z ) = − z + i 2 , d σ 0 ( t ) = π ( 1 + t 2 ) 2 2 d t .
The Cauchy measure 2 K 2 ( 0 , t ) d t 2K_2(0,t)\,dt 2 K 2 ( 0 , t ) d t has mass two, while the finite Nevanlinna measure has mass one. This small example makes an omitted factor 1 + t 2 1+t^2 1 + t 2 impossible to overlook. In the other direction, the same kernel gives the tangent density (67) .
The parameter law as a conjugate posterior
The angle also gives the joint law a familiar statistical meaning. For a fixed parameter θ \theta θ , the density of the observation x x x is
K η ( x , tan θ ) = C η ( x ) e x θ cos η θ . K_\eta(x,\tan\theta)=C_\eta(x)e^{x\theta}\cos^\eta\theta. K η ( x , tan θ ) = C η ( x ) e x θ cos η θ .
The dependence on θ \theta θ is entirely through an exponential and a power of the cosine. It is therefore natural to give θ \theta θ a prior density of the same form,
p r , s ( θ ) = e s θ cos r θ Z ( r , s ) , − π 2 < θ < π 2 , r > − 1 , s ∈ R . (117) p_{r,s}(\theta)=\frac{e^{s\theta}\cos^r\theta}{Z(r,s)},\qquad-\frac{\pi}{2}<\theta<\frac{\pi}{2},\qquad r>-1,\quad s\in\mathbb{R}. \tag*{(117)} p r , s ( θ ) = Z ( r , s ) e s θ cos r θ , − 2 π < θ < 2 π , r > − 1 , s ∈ R . (117)
The normalizing constant is
Z ( r , s ) = ∫ − π / 2 π / 2 e s θ cos r θ d θ = π Γ ( r + 1 ) 2 r Γ ( 1 + ( r + i s ) / 2 ) Γ ( 1 + ( r − i s ) / 2 ) . (118) Z(r,s)=\int_{-\pi/2}^{\pi/2}e^{s\theta}\cos^r\theta\,d\theta=\frac{\pi\Gamma(r+1)}{2^r\Gamma(1+(r+is)/2)\Gamma(1+(r-is)/2)}. \tag*{(118)} Z ( r , s ) = ∫ − π /2 π /2 e s θ cos r θ d θ = 2 r Γ ( 1 + ( r + i s ) /2 ) Γ ( 1 + ( r − i s ) /2 ) π Γ ( r + 1 ) . (118)
No new integral is needed: substitute t = tan θ t=\tan\theta t = tan θ into (91) and take a = 1 + ( r + i s ) / 2 a=1+(r+is)/2 a = 1 + ( r + i s ) /2 . Its condition ℜ a > 1 / 2 \Re a>1/2 ℜ a > 1/2 becomes exactly r > − 1 r>-1 r > − 1 .
Multiplication by the likelihood merely adds exponents. Bayes’ formula gives the posterior update and the predictive density
( r , s ) ⟼ ( r + η , s + x ) , p η ; r , s ( x ) = C η ( x ) Z ( r + η , s + x ) Z ( r , s ) . (119) (r,s)\longmapsto(r+\eta,s+x),\qquad p_{\eta;r,s}(x)=C_\eta(x)\frac{Z(r+\eta,s+x)}{Z(r,s)}. \tag*{(119)} ( r , s ) ⟼ ( r + η , s + x ) , p η ; r , s ( x ) = C η ( x ) Z ( r , s ) Z ( r + η , s + x ) . (119)
Thus the family of priors is conjugate : observing x x x keeps the posterior within the same two-parameter family. For r = s = 0 r=s=0 r = s = 0 the prior is uniform in angle and hence standard Cauchy in δ \delta δ . Formula (93) , or gamma cancellation in (118) , gives the scale-η \eta η Cauchy predictive density. The posterior, expressed in δ \delta δ , is exactly σ ^ x \widehat{\sigma}_x σ x in (109) . This conjugate family and its predictive distributions are discussed in [8 , 36 ] . Here they explain why the two spectral measures fit so simply into one joint probability.
Bounded rational moments
Under σ ^ x \widehat{\sigma}_x σ x , the angle has density e x θ cos η θ / Z ( η , x ) e^{x\theta}\cos^\eta\theta/Z(\eta,x) e x θ cos η θ / Z ( η , x ) . Its Fourier coefficients have a particularly simple product form. For every integer n ≥ 0 n\geq0 n ≥ 0 ,
∫ R ( 1 + i t 1 − i t ) n d σ ^ x ( t ) = ∏ j = 0 n − 1 η − 2 j + i x η + 2 j + 2 − i x . (120) \int_{\mathbb{R}}\left(\frac{1+it}{1-it}\right)^n\,d\widehat{\sigma}_x(t)=\prod_{j=0}^{n-1}\frac{\eta-2j+ix}{\eta+2j+2-ix}. \tag*{(120)} ∫ R ( 1 − i t 1 + i t ) n d σ x ( t ) = j = 0 ∏ n − 1 η + 2 j + 2 − i x η − 2 j + i x . (120)
The empty product is one, and negative powers give the complex conjugates. To prove the formula, observe that
e 2 i θ = 1 + i t 1 − i t , t = tan θ . e^{2i\theta}=\frac{1+it}{1-it},\qquad t=\tan\theta. e 2 i θ = 1 − i t 1 + i t , t = tan θ .
The required integral is therefore Z ( η , x + 2 i n ) / Z ( η , x ) Z(\eta,x+2in)/Z(\eta,x) Z ( η , x + 2 in ) / Z ( η , x ) . For a fixed r > − 1 r>-1 r > − 1 , both sides of (118) are entire in s s s : on the integral side the interval is bounded, and on the gamma side one uses the entire reciprocal gamma function. The identity for real s s s therefore extends to complex s s s . Gamma recurrence now gives the displayed product; if a reciprocal gamma factor vanishes, the result follows by continuity.
The integrand in (120) has modulus one on the real line. These rational moments exist at every degree even though sufficiently high polynomial moments do not. The circle coordinate is precisely the Cayley coordinate of (77) : what looked like a difficulty at infinity on the line becomes ordinary Fourier analysis on the circle. This is the natural setting of the complementary Romanovski–Routh and unit-circle connections [16 , 23 ] .
A joint Cauchy law at every finite degree
The joint law of the preceding section is already present in every finite continued fraction. At finite degree it lives on the zero curves of a polynomial, yet its two marginals are exactly the same Cauchy laws as in the limit. To see why, we need only follow a zero as the parameter moves.
The zeros form increasing spectral curves
Return to the positive pencil of (83) , and write
H N ( δ ) = B N + δ A N , x 1 ( δ ) < ⋯ < x N ( δ ) H_N(\delta)=B_N+\delta A_N,\qquad x_1(\delta)<\cdots<x_N(\delta) H N ( δ ) = B N + δ A N , x 1 ( δ ) < ⋯ < x N ( δ )
for its eigenvalues. They are simple because a diagonal unitary conjugation makes H N ( δ ) H_N(\delta) H N ( δ ) an irreducible real Jacobi matrix. If v j ( δ ) v_j(\delta) v j ( δ ) is a unit eigenvector, differentiating its eigenvalue equation and taking the inner product with v j v_j v j gives the Hellmann–Feynman formula
x j ′ ( δ ) = v j ( δ ) ∗ A N v j ( δ ) > 0. (121) x_j'(\delta)=v_j(\delta)^*A_Nv_j(\delta)>0. \tag*{(121)} x j ′ ( δ ) = v j ( δ ) ∗ A N v j ( δ ) > 0. (121)
Positivity of the pencil has become motion in a definite direction.
There are exact, familiar bounds for this motion. The characteristic polynomials of the leading blocks of A N A_N A N satisfy
L n + 1 ( u ) = ( u − 2 n − η ) L n ( u ) − n ( n + η − 1 ) L n − 1 ( u ) , L 0 = 1 , L 1 = u − η . \mathcal{L}_{n+1}(u)=(u-2n-\eta)\mathcal{L}_n(u)-n(n+\eta-1)\mathcal{L}_{n-1}(u),\qquad\mathcal{L}_0=1,\qquad\mathcal{L}_1=u-\eta. L n + 1 ( u ) = ( u − 2 n − η ) L n ( u ) − n ( n + η − 1 ) L n − 1 ( u ) , L 0 = 1 , L 1 = u − η .
These are the monic Laguerre polynomials L n ( u ) = ( − 1 ) n n ! L n ( η − 1 ) ( u ) \mathcal{L}_n(u)=(-1)^n n!L_n^{(\eta-1)}(u) L n ( u ) = ( − 1 ) n n ! L n ( η − 1 ) ( u ) [19 §9.12] . Thus, if 0 < ℓ 1 < ⋯ < ℓ N 0<\ell_1<\cdots<\ell_N 0 < ℓ 1 < ⋯ < ℓ N are the zeros of L N ( η − 1 ) L_N^{(\eta-1)} L N ( η − 1 ) , the Rayleigh quotient in (121) gives
ℓ 1 ≤ x j ′ ( δ ) ≤ ℓ N . (122) \ell_1\le x_j'(\delta)\le\ell_N. \tag*{(122)} ℓ 1 ≤ x j ′ ( δ ) ≤ ℓ N . (122)
Each zero curve is consequently a real analytic increasing bijection of R \mathbb{R} R onto R \mathbb{R} R . The two spectral problems are the vertical and horizontal readings of the same N N N disjoint curves.
The pencil gives a two-variable statement as well. A real polynomial is called real stable if it is nonzero whenever all its variables have positive imaginary parts. Here p N ( x , y ) = M N ( x ; − y , η ) p_N(x,y)=M_N(x;-y,\eta) p N ( x , y ) = M N ( x ; − y , η ) is real stable: the imaginary part of its matrix x I N + y A N − B N xI_N+yA_N-B_N x I N + y A N − B N is ( ℑ x ) I N + ( ℑ y ) A N > 0 (\Im x)I_N+(\Im y)A_N>0 ( ℑ x ) I N + ( ℑ y ) A N > 0 , so that matrix cannot have a null vector. More generally, every oblique slice
t ⟼ M N ( a + u t ; b − v t , η ) , a , b ∈ R , u , v ≥ 0 , u + v > 0 , (123) t\longmapsto M_N(a+ut;b-vt,\eta),\qquad a,b\in\mathbb{R},\quad u,v\ge0,\quad u+v>0, \tag*{(123)} t ⟼ M N ( a + u t ; b − v t , η ) , a , b ∈ R , u , v ≥ 0 , u + v > 0 , (123)
has only real zeros. Its matrix is a Hermitian pencil with positive coefficient u I N + v A N uI_N+vA_N u I N + v A N ; conjugating by the inverse square root of that coefficient turns its zeros into ordinary Hermitian eigenvalues. The reality of the zeros in either variable is part of this more coherent geometric fact.
A Jacobian relates the two quadrature weights
The spectral weights in the x x x direction are
w j ( δ ) = ∣ ⟨ e 0 , v j ( δ ) ⟩ ∣ 2 , μ δ , N = ∑ j = 1 N w j ( δ ) δ x j ( δ ) . (124) w_j(\delta)=|\langle e_0,v_j(\delta)\rangle|^2,\qquad\mu_{\delta,N}=\sum_{j=1}^{N}w_j(\delta)\delta_{x_j(\delta)}. \tag*{(124)} w j ( δ ) = ∣ ⟨ e 0 , v j ( δ )⟩ ∣ 2 , μ δ , N = j = 1 ∑ N w j ( δ ) δ x j ( δ ) . (124)
Here δ a \delta_a δ a denotes a unit point mass at a a a . This is the Gaussian quadrature measure used in § 8 : the diagonal unitary change of basis preserves the first-coordinate weights. Each weight is positive, and their sum is one.
For the parameter direction, factor the matrix in (102) through A N 1 / 2 A_N^{1/2} A N 1/2 . If
T x , N = A N − 1 / 2 ( x I N − B N ) A N − 1 / 2 , T_{x,N}=A_N^{-1/2}(xI_N-B_N)A_N^{-1/2}, T x , N = A N − 1/2 ( x I N − B N ) A N − 1/2 ,
the spectral measure of T x , N T_{x,N} T x , N at the vector η A N − 1 / 2 e 0 \sqrt{\eta}A_N^{-1/2}e_0 η A N − 1/2 e 0 is a positive discrete measure ρ x , N \rho_{x,N} ρ x , N , and
F x , N ( z ) = ∫ R ρ x , N ( d t ) t − z . (125) F_{x,N}(z)=\int_{\mathbb{R}}\frac{\rho_{x,N}(\mathrm{d}t)}{t-z}. \tag*{(125)} F x , N ( z ) = ∫ R t − z ρ x , N ( d t ) . (125)
Its support consists of the parameter zeros of M N ( x ; t , η ) M_N(x;t,\eta) M N ( x ; t , η ) .
Proposition 12.1. At the crossing x = x j ( t ) x=x_j(t) x = x j ( t ) , the two weights satisfy
ρ x , N ( { t } ) = η w j ( t ) x j ′ ( t ) . (126) \rho_{x,N}(\{t\})=\frac{\eta w_j(t)}{x_j'(t)}. \tag*{(126)} ρ x , N ({ t }) = x j ′ ( t ) η w j ( t ) . (126) Consequently, as locally finite measures on R 2 \mathbb{R}^2 R 2 ,
μ δ , N ( d x ) d δ = 1 η ρ x , N ( d δ ) d x . (127) \mu_{\delta,N}(\mathrm{d}x)\,\mathrm{d}\delta=\frac{1}{\eta}\rho_{x,N}(\mathrm{d}\delta)\,\mathrm{d}x. \tag*{(127)} μ δ , N ( d x ) d δ = η 1 ρ x , N ( d δ ) d x . (127)
Proof. At a crossing, the corresponding unit eigenvector of T x , N T_{x,N} T x , N is
q = A N 1 / 2 v j ( t ) v j ( t ) ∗ A N v j ( t ) . q=\frac{A_N^{1/2}v_j(t)}{\sqrt{v_j(t)^*A_Nv_j(t)}}. q = v j ( t ) ∗ A N v j ( t ) A N 1/2 v j ( t ) .
Its spectral weight at η A N − 1 / 2 e 0 \sqrt{\eta}\,A_N^{-1/2}e_0 η A N − 1/2 e 0 is η w j ( t ) / ( v j ( t ) ∗ A N v j ( t ) ) \eta w_j(t)/(v_j(t)^*A_Nv_j(t)) η w j ( t ) / ( v j ( t ) ∗ A N v j ( t )) . Formula (121) proves (126) .
Now substitute x = x j ( δ ) x=x_j(\delta) x = x j ( δ ) on each increasing curve. For a compactly supported continuous function f f f ,
∑ j ∫ R f ( x j ( δ ) , δ ) w j ( δ ) d δ = ∫ R ∑ j = 1 N f ( x , t j ( x ) ) w j ( t j ( x ) ) x j ′ ( t j ( x ) ) d x , \sum_{j}\int_{\mathbb{R}}f(x_j(\delta),\delta)w_j(\delta)\,\mathrm{d}\delta
=
\int_{\mathbb{R}}\sum_{j=1}^{N}f(x,t_j(x))\frac{w_j(t_j(x))}{x_j'(t_j(x))}\,\mathrm{d}x, j ∑ ∫ R f ( x j ( δ ) , δ ) w j ( δ ) d δ = ∫ R j = 1 ∑ N f ( x , t j ( x )) x j ′ ( t j ( x )) w j ( t j ( x )) d x ,
where t j = x j − 1 t_j=x_j^{-1} t j = x j − 1 . This is (127) .
The Jacobian is the essential point. The two finite measures assign mass along the same curve, using different coordinates on that curve.
The finite coupling and its limit
Theorem 12.2. For every N ≥ 1 N\geq1 N ≥ 1 and η > 0 \eta>0 η > 0 , the probability measure
Π η , N ( d x , d δ ) = μ δ , N ( d x ) d δ π ( 1 + δ 2 ) (128) \Pi_{\eta,N}(\mathrm{d}x,\mathrm{d}\delta)=\mu_{\delta,N}(\mathrm{d}x)\frac{\mathrm{d}\delta}{\pi(1+\delta^2)} \tag*{(128)} Π η , N ( d x , d δ ) = μ δ , N ( d x ) π ( 1 + δ 2 ) d δ (128) is supported on M N ( x ; δ , η ) = 0 M_N(x;\delta,\eta)=0 M N ( x ; δ , η ) = 0 and has the exact marginals
Π η , N δ ( d δ ) = d δ π ( 1 + δ 2 ) , Π η , N x ( d x ) = η d x π ( x 2 + η 2 ) . (129) \Pi_{\eta,N}^{\delta}(\mathrm{d}\delta)=\frac{\mathrm{d}\delta}{\pi(1+\delta^2)},\qquad
\Pi_{\eta,N}^{x}(\mathrm{d}x)=\frac{\eta\,\mathrm{d}x}{\pi(x^2+\eta^2)}. \tag*{(129)} Π η , N δ ( d δ ) = π ( 1 + δ 2 ) d δ , Π η , N x ( d x ) = π ( x 2 + η 2 ) η d x . (129) Its reverse conditional law is
σ ^ x , N ( d t ) = x 2 + η 2 η 2 ( 1 + t 2 ) ρ x , N ( d t ) . (130) \widehat{\sigma}_{x,N}(\mathrm{d}t)=\frac{x^2+\eta^2}{\eta^2(1+t^2)}\rho_{x,N}(\mathrm{d}t). \tag*{(130)} σ x , N ( d t ) = η 2 ( 1 + t 2 ) x 2 + η 2 ρ x , N ( d t ) . (130) As N → ∞ N\to\infty N → ∞ , Π η , N \Pi_{\eta,N} Π η , N converges weakly to the joint law Π η \Pi_\eta Π η of (114) .
Proof. The first marginal follows from μ δ , N ( R ) = 1 \mu_{\delta,N}(\mathbb{R})=1 μ δ , N ( R ) = 1 . The common value F x , N ( i ) = F x ( i ) = η / ( x − i η ) F_{x,N}(i)=F_x(i)=\eta/(x-i\eta) F x , N ( i ) = F x ( i ) = η / ( x − i η ) , obtained in the proof of Theorem 10.1, gives
∫ R ρ x , N ( d t ) 1 + t 2 = ℑ F x , N ( i ) = η 2 x 2 + η 2 . \int_{\mathbb{R}}\frac{\rho_{x,N}(\mathrm{d}t)}{1+t^2}=\Im F_{x,N}(i)=\frac{\eta^2}{x^2+\eta^2}. ∫ R 1 + t 2 ρ x , N ( d t ) = ℑ F x , N ( i ) = x 2 + η 2 η 2 . Using (127) , the x x x -marginal density of (128) is therefore
1 π η ∫ R ρ x , N ( d t ) 1 + t 2 = η π ( x 2 + η 2 ) . \frac{1}{\pi\eta}\int_{\mathbb{R}}\frac{\rho_{x,N}(\mathrm{d}t)}{1+t^2}=\frac{\eta}{\pi(x^2+\eta^2)}. π η 1 ∫ R 1 + t 2 ρ x , N ( d t ) = π ( x 2 + η 2 ) η .
Dividing the same measure identity by this density proves (130) .
For fixed δ \delta δ , the convergence μ δ , N ⇒ μ δ \mu_{\delta,N} \Rightarrow\mu_\delta μ δ , N ⇒ μ δ was proved in Theorem 8.1 . Apply it to f ( ⋅ , δ ) f(\cdot,\delta) f ( ⋅ , δ ) , for a bounded continuous function f f f on R 2 \mathbb{R}^2 R 2 , and then use dominated convergence against the fixed Cauchy law in δ \delta δ . This proves the joint weak convergence.
For N = 1 N=1 N = 1 , the coupling is simply x = η δ x=\eta\delta x = η δ . Higher degrees change the dependence while preserving both marginals. The smooth density of (114) is approached by singular measures on algebraic curves, as Figure 1 illustrates.
Figure 1. Finite spectral curves and the limiting density, with η = 2 \eta=2 η = 2 , in the angles θ = arctan δ \theta=\arctan\delta θ = arctan δ and φ = arctan ( x / η ) \varphi=\arctan(x/\eta) φ = arctan ( x / η ) . Both angular marginals are uniform on ( − π / 2 , π / 2 ) (-\pi/2,\pi/2) ( − π /2 , π /2 ) at every stage. The first three panels show support only: the conditional weights are those in (124) , and are not encoded by line thickness. The last panel shows the joint density in the two angles.
This is spectral averaging in a particularly explicit form; general matrix-valued spectral averaging is developed in [22 ] . Here the positive pencil makes the mechanism visible: a zero moves with velocity v ∗ A N v v^*A_Nv v ∗ A N v , and its spectral mass changes by the reciprocal velocity when the two coordinates are flipped. The finite fractions therefore retain both measures and the compatibility between them.
What the finite fractions tell us quantitatively
The finite measures also give practical formulas and convergence bounds. The general relationship between RII recurrences and rational quadrature is established [2 ] ; here all the constants are explicit.
Positive rational Gaussian weights
The preceding section identified the finite measures geometrically. We can now write each of their masses in closed form. Recall that F x , N = η N N / M N F_{x,N}=\eta N_N/M_N F x , N = η N N / M N , with numerator polynomials N 0 = 0 N_0=0 N 0 = 0 , N 1 = 1 N_1=1 N 1 = 1 from the continuant recurrence in § 10 . The exact identity (101) gives
F x ( z ) − F x , N ( z ) = O ( ( z − i ) N ) ( z → i ) . F_x(z)-F_{x,N}(z)=O((z-i)^N)\qquad(z\to i). F x ( z ) − F x , N ( z ) = O (( z − i ) N ) ( z → i ) .
This interpolation property will provide both quadrature exactness and a convergence bound.
Proposition 13.1. Let t 1 < ⋯ < t N t_1<\cdots<t_N t 1 < ⋯ < t N be the parameter zeros of M N ( x ; t , η ) M_N(x;t,\eta) M N ( x ; t , η ) , and let a prime denote differentiation in t t t . Then
ρ x , N ( { t j } ) = η N ! ( η ) N ( 1 + t j 2 ) N − 1 [ M N ′ ( x ; t j , η ) ] 2 , (131) \rho_{x,N}(\{t_j\})=\frac{\eta N!(\eta)_N(1+t_j^2)^{N-1}}{[M_N'(x;t_j,\eta)]^2}, \tag*{(131)} ρ x , N ({ t j }) = [ M N ′ ( x ; t j , η ) ] 2 η N ! ( η ) N ( 1 + t j 2 ) N − 1 , (131) ω j : = σ ^ x , N ( { t j } ) = x 2 + η 2 η N ! ( η ) N ( 1 + t j 2 ) N − 2 [ M N ′ ( x ; t j , η ) ] 2 . (132) \omega_j:=\widehat{\sigma}_{x,N}(\{t_j\})=\frac{x^2+\eta^2}{\eta}\frac{N!(\eta)_N(1+t_j^2)^{N-2}}{[M_N'(x;t_j,\eta)]^2}. \tag*{(132)} ω j := σ x , N ({ t j }) = η x 2 + η 2 [ M N ′ ( x ; t j , η ) ] 2 N ! ( η ) N ( 1 + t j 2 ) N − 2 . (132) All weights are positive and ∑ j ω j = 1 \sum_j\omega_j=1 ∑ j ω j = 1 . For every polynomial p p p of degree at most 2 N − 1 2N-1 2 N − 1 ,
∫ R p ( t ) ( 1 + t 2 ) N − 1 σ ^ x ( d t ) = ∑ j = 1 N ω j p ( t j ) ( 1 + t j 2 ) N − 1 . (133) \int_{\mathbb{R}}\frac{p(t)}{(1+t^2)^{N-1}}\widehat{\sigma}_x(\mathrm{d}t)=\sum_{j=1}^{N}\omega_j\frac{p(t_j)}{(1+t_j^2)^{N-1}}. \tag*{(133)} ∫ R ( 1 + t 2 ) N − 1 p ( t ) σ x ( d t ) = j = 1 ∑ N ω j ( 1 + t j 2 ) N − 1 p ( t j ) . (133) These statements hold for every η > 0 \eta>0 η > 0 .
Proof. The recurrence gives the continuant identity
M N − 1 N N − M N N N − 1 = ( N − 1 ) ! ( η ) N − 1 ( 1 + t 2 ) N − 1 . M_{N-1}N_N-M_NN_{N-1}=(N-1)!(\eta)_{N-1}(1+t^2)^{N-1}. M N − 1 N N − M N N N − 1 = ( N − 1 )! ( η ) N − 1 ( 1 + t 2 ) N − 1 .
The parameter lowering identity (88) then gives
N N ( x ; t j , η ) = − N ! ( η ) N ( 1 + t j 2 ) N − 1 M N ′ ( x ; t j , η ) . N_N(x;t_j,\eta)=-\frac{N!(\eta)_N(1+t_j^2)^{N-1}}{M_N'(x;t_j,\eta)}. N N ( x ; t j , η ) = − M N ′ ( x ; t j , η ) N ! ( η ) N ( 1 + t j 2 ) N − 1 .
Taking the negative residue of F x , N = η N N / M N F_{x,N}=\eta N_N/M_N F x , N = η N N / M N proves (131) ; (130) proves (132) .
Taylor interpolation at i i i gives equality of the integrals of ( t − i ) − k (t-i)^{-k} ( t − i ) − k against ρ x \rho_x ρ x and ρ x , N \rho_{x,N} ρ x , N for 1 ≤ k ≤ N 1\le k\le N 1 ≤ k ≤ N . Reflection gives the same assertion at − i -i − i . Every p ( t ) / ( 1 + t 2 ) N p(t)/(1+t^2)^N p ( t ) / ( 1 + t 2 ) N with deg p ≤ 2 N − 1 \deg p\le2N-1 deg p ≤ 2 N − 1 is a linear combination of those 2 N 2N 2 N partial fractions. Divide the measures by 1 + t 2 1+t^2 1 + t 2 and use their common normalization to obtain (133) . All integrals converge: at the largest allowed degree the integrand against ρ x \rho_x ρ x is O ( ∣ t ∣ − η − 1 ) O(|t|^{-\eta-1}) O ( ∣ t ∣ − η − 1 ) .
This is genuine Gaussian exactness in a rational space. The denominator provides the moments which the heavy-tailed measure does not possess as ordinary polynomial moments.
The mass transition at η = 1 \eta= 1 η = 1
The total mass of a finite Wall measure is independent of x x x :
ρ x , N ( R ) = η ( A N − 1 ) 00 = η ∑ k = 0 N − 1 k ! ( η ) k + 1 = { η η − 1 ( 1 − N ! ( η ) N ) , η ≠ 1 , H N , η = 1. (134) \rho_{x,N}(\mathbb{R}) = \eta(A_N^{-1})_{00} = \eta\sum_{k=0}^{N-1} \frac{k!}{(\eta)_{k+1}} =
\begin{cases}
\frac{\eta}{\eta-1}\left(1-\frac{N!}{(\eta)_N}\right), & \eta\ne1,\\
H_N, & \eta= 1.
\end{cases} \tag*{(134)} ρ x , N ( R ) = η ( A N − 1 ) 00 = η k = 0 ∑ N − 1 ( η ) k + 1 k ! = { η − 1 η ( 1 − ( η ) N N ! ) , H N , η = 1 , η = 1. (134)
Here H N = ∑ j = 1 N 1 / j H_N = \sum_{j=1}^{N} 1/j H N = ∑ j = 1 N 1/ j is the N N N th harmonic number. To prove the formula, solving L N u = e 0 L_Nu=e_0 L N u = e 0 gives ∣ u k ∣ 2 = k ! / ( η ) k + 1 |u_k|^2 = k!/(\eta)_{k+1} ∣ u k ∣ 2 = k ! / ( η ) k + 1 , and the sum telescopes because
k ! ( η ) k − ( k + 1 ) ! ( η ) k + 1 = ( η − 1 ) k ! ( η ) k + 1 . \frac{k!}{(\eta)_k} - \frac{(k+1)!}{(\eta)_{k+1}} = (\eta-1)\frac{k!}{(\eta)_{k+1}}. ( η ) k k ! − ( η ) k + 1 ( k + 1 )! = ( η − 1 ) ( η ) k + 1 k ! .
Thus the finite measures approach total mass η / ( η − 1 ) \eta/(\eta-1) η / ( η − 1 ) when η > 1 \eta> 1 η > 1 , grow as log N \log N log N when η = 1 \eta= 1 η = 1 , and grow as
η Γ ( η ) 1 − η N 1 − η ( 0 < η < 1 ) . \frac{\eta\Gamma(\eta)}{1-\eta}N^{1-\eta}\qquad(0<\eta<1). 1 − η η Γ ( η ) N 1 − η ( 0 < η < 1 ) .
The threshold for the infinite measure is already encoded in the first column of a finite triangular inverse. The normalized probabilities (130) have mass one throughout.
A geometric error bound for the Wall fraction
Proposition 13.2. For z ∈ C + z \in\mathbb{C}_+ z ∈ C + , set
q = ∣ z − i z + i ∣ < 1 , α = η 2 x 2 + η 2 . q=\left|\frac{z-i}{z+i}\right|<1,\qquad\alpha=\frac{\eta^2}{x^2+\eta^2}. q = z + i z − i < 1 , α = x 2 + η 2 η 2 . Then
∣ F x , N ( z ) − F x ( z ) ∣ ≤ 4 α q N ( 1 − q ) 2 . (135) |F_{x,N}(z)-F_x(z)|\le\frac{4\alpha q^N}{(1-q)^2}. \tag*{(135)} ∣ F x , N ( z ) − F x ( z ) ∣ ≤ ( 1 − q ) 2 4 α q N . (135) Proof. Put c = F x ( i ) = F x , N ( i ) c=F_x(i)=F_{x,N}(i) c = F x ( i ) = F x , N ( i ) , so ℑ c = α \Im c=\alpha ℑ c = α , and map the upper half-plane to the disk by w = ( z − i ) / ( z + i ) w=(z-i)/(z+i) w = ( z − i ) / ( z + i ) . The functions
S ( w ) = F x ( z ) − c F x ( z ) − c ‾ , S N ( w ) = F x , N ( z ) − c F x , N ( z ) − c ‾ S(w)=\frac{F_x(z)-c}{F_x(z)-\overline{c}},\qquad S_N(w)=\frac{F_{x,N}(z)-c}{F_{x,N}(z)-\overline{c}} S ( w ) = F x ( z ) − c F x ( z ) − c , S N ( w ) = F x , N ( z ) − c F x , N ( z ) − c map the disk into itself and vanish at zero. Schwarz’s lemma gives ∣ S ∣ , ∣ S N ∣ ≤ ∣ w ∣ |S|,|S_N|\le|w| ∣ S ∣ , ∣ S N ∣ ≤ ∣ w ∣ . Their difference vanishes to order N N N by (101) ; the same lemma applied to ( S N − S ) / 2 (S_N-S)/2 ( S N − S ) /2 gives ∣ S N − S ∣ ≤ 2 ∣ w ∣ N |S_N-S|\le2|w|^N ∣ S N − S ∣ ≤ 2∣ w ∣ N . Inverting the fractional linear transformation yields
∣ F x , N − F x ∣ = 2 α ∣ S N − S ∣ ∣ 1 − S N ∣ ∣ 1 − S ∣ , |F_{x,N}-F_x|=\frac{2\alpha|S_N-S|}{|1-S_N||1-S|}, ∣ F x , N − F x ∣ = ∣1 − S N ∣∣1 − S ∣ 2 α ∣ S N − S ∣ , which proves the bound.
The interpolation point i i i is the natural center of this estimate: its Cayley coordinate q q q measures how far we have moved toward the boundary. On every compact subset of the upper half-plane the continued fraction therefore converges geometrically. No asymptotic analysis of the hypergeometric expression is required.
What the two Cauchy variables do at infinity
The two marginals remain Cauchy while the shape parameter η \eta η changes. What changes is the dependence between the coordinates. Ordinary correlation cannot measure it, since the Cauchy marginals have no second moments. We can ask a more direct question: if one coordinate is large and positive, how likely is the other to be comparably large? The answer leads back to the gamma measure and the Laguerre polynomials that bounded the velocities of our zeros.
Rescale the first coordinate by putting Y = X / η Y=X/\eta Y = X / η and write D D D for the parameter coordinate. Both D D D and Y Y Y now have the standard Cauchy law
m ( d d ) = d d π ( 1 + d 2 ) . m(\mathrm{d}d)=\frac{\mathrm{d}d}{\pi(1+d^2)}. m ( d d ) = π ( 1 + d 2 ) d d .
Their conditional kernel and joint density are
Q η ( d , d y ) = η K η ( η y , d ) d y , p η ( d , y ) = η K η ( η y , d ) π ( 1 + d 2 ) . (136) Q_\eta(d,\mathrm{d}y)=\eta K_\eta(\eta y,d)\,\mathrm{d}y,\qquad p_\eta(d,y)=\frac{\eta K_\eta(\eta y,d)}{\pi(1+d^2)}. \tag*{(136)} Q η ( d , d y ) = η K η ( η y , d ) d y , p η ( d , y ) = π ( 1 + d 2 ) η K η ( η y , d ) . (136)
The integral operator Q η f ( d ) = ∫ f ( y ) Q η ( d , d y ) Q_\eta f(d)=\int f(y)Q_\eta(d,\mathrm{d}y) Q η f ( d ) = ∫ f ( y ) Q η ( d , d y ) is the 1 + d 2 1+d^2 1 + d 2 member of the Ismail–May family of approximation operators [15 ] ; its kernel appears explicitly in [11 ] , Eq. (1). Our concern here is its behavior under its invariant Cauchy measure.
The Meixner transform gives
E [ e i t Y ∣ D = d ] = [ cosh ( t / η ) − i d sinh ( t / η ) ] − η , E [ Y ∣ D = d ] = d , Var ( Y ∣ D = d ) = 1 + d 2 η . (137) \mathbb{E}[e^{itY}\mid D=d]=\left[\cosh(t/\eta)-id\sinh(t/\eta)\right]^{-\eta},\qquad\mathbb{E}[Y\mid D=d]=d,\qquad\operatorname{Var}(Y\mid D=d)=\frac{1+d^2}{\eta}. \tag*{(137)} E [ e i t Y ∣ D = d ] = [ cosh ( t / η ) − i d sinh ( t / η ) ] − η , E [ Y ∣ D = d ] = d , Var ( Y ∣ D = d ) = η 1 + d 2 . (137)
The complex power is the continuous branch equal to one at t = 0 t=0 t = 0 . The conditional moments here are moments of the probability Q η ( d , d y ) Q_\eta(d,\mathrm{d}y) Q η ( d , d y ) at a fixed d d d . They are finite even though the stationary pair ( D , Y ) (D,Y) ( D , Y ) has no integrable coordinates.
A gamma law at the edge
Proposition 14.1. Fix η > 0 \eta>0 η > 0 . As d → + ∞ d\to+\infty d → + ∞ ,
L ( Y / d ∣ D = d ) ⟹ G η , P ( G η ∈ d u ) = η η Γ ( η ) u η − 1 e − η u d u , u > 0. (138) \mathcal{L}(Y/d\mid D=d)\Longrightarrow G_\eta,\qquad\mathbb{P}(G_\eta\in\mathrm{d}u)=\frac{\eta^\eta}{\Gamma(\eta)}u^{\eta-1}e^{-\eta u}\,\mathrm{d}u,\qquad u>0. \tag*{(138)} L ( Y / d ∣ D = d ) ⟹ G η , P ( G η ∈ d u ) = Γ ( η ) η η u η − 1 e − η u d u , u > 0. (138) The same ratio limit holds as d → − ∞ d\to-\infty d → − ∞ .
Proof. Replace t t t by t / d t/d t / d in (137) . Then
E [ e i t Y / d ∣ D = d ] = [ cosh ( t η d ) − i d sinh ( t η d ) ] − η ⟶ ( 1 − i t / η ) − η . \mathbb{E}[e^{itY/d}\mid D=d]=\left[\cosh\left(\frac{t}{\eta d}\right)-id\sinh\left(\frac{t}{\eta d}\right)\right]^{-\eta}\longrightarrow(1-it/\eta)^{-\eta}. E [ e i t Y / d ∣ D = d ] = [ cosh ( η d t ) − i d sinh ( η d t ) ] − η ⟶ ( 1 − i t / η ) − η . The limit is the characteristic function of the displayed gamma law. Lévy’s continuity theorem proves the assertion. Reflection, K η ( − x , − d ) = K η ( x , d ) K_\eta(-x,-d)=K_\eta(x,d) K η ( − x , − d ) = K η ( x , d ) , gives the other endpoint.
The limiting relative fluctuation has mean one and variance 1 / η 1/\eta 1/ η . Its gamma law is the Laguerre orthogonality measure, with the scale changed to give mean one.
Reversing the conditioning gives another classical distribution. It is useful to state the reverse limit separately, since it explains why the two directions have different normalizations.
Proposition 14.2. As y → + ∞ y \to+\infty y → + ∞ , the conditional law of D / y D/y D / y given Y = y Y=y Y = y converges in total variation to the inverse-gamma law with density
η η + 1 Γ ( η + 1 ) v − η − 2 e − η / v , v > 0. (139) \frac{\eta^{\eta+1}}{\Gamma(\eta+1)}v^{-\eta-2}e^{-\eta/v}, \qquad v>0. \tag*{(139)} Γ ( η + 1 ) η η + 1 v − η − 2 e − η / v , v > 0. (139) The same assertion holds as y → − ∞ y \to-\infty y → − ∞ .
Proof. The reverse conditional law (110) is
P ( D ∈ d d ∣ Y = y ) = ( η + 1 ) K η + 2 ( η y , d ) d d . \mathbb{P}(D \in\mathrm{d}d \mid Y=y)=(\eta+1)K_{\eta+2}(\eta y,d)\,\mathrm{d}d. P ( D ∈ d d ∣ Y = y ) = ( η + 1 ) K η + 2 ( η y , d ) d d . For positive y y y , the density of D / y D/y D / y is therefore y ( η + 1 ) K η + 2 ( η y , y v ) y(\eta+1)K_{\eta+2}(\eta y,yv) y ( η + 1 ) K η + 2 ( η y , y v ) . Use the gamma asymptotic (18) and, for v > 0 v>0 v > 0 ,
arctan ( y v ) = π 2 − 1 y v + O ( y − 3 ) . \arctan(yv)=\frac{\pi}{2}-\frac{1}{yv}+O(y^{-3}). arctan ( y v ) = 2 π − y v 1 + O ( y − 3 ) . Substitution gives (139) pointwise. For v < 0 v<0 v < 0 the two exponential factors instead leave a factor e − π η y + O ( 1 ) e^{-\pi\eta y+O(1)} e − π η y + O ( 1 ) , so the limit is zero. The limiting density has mass one, as the substitution u = 1 / v u=1/v u = 1/ v shows. Scheffé’s lemma proves total-variation convergence. Central reflection gives the other endpoint. □
This inverse-gamma variable is the reciprocal of a size-biased G η G_\eta G η : since E G η = 1 \mathbb{E}G_\eta=1 E G η = 1 , size bias means multiplying its density by u u u , which changes its shape from η \eta η to η + 1 \eta+1 η + 1 . Thus changing the direction of conditioning at the boundary has a precise and elementary cost.
An exact answer for simultaneous extremes
Theorem 14.3. For every η > 0 \eta>0 η > 0 and a , b > 0 a,b>0 a , b > 0 ,
lim R → ∞ π R P ( D > a R , Y > b R ) = E min ( 1 a , G η b ) . (140) \lim_{R\to\infty}\pi R\mathbb{P}(D>aR,Y>bR)=\mathbb{E}\min\left(\frac{1}{a},\frac{G_\eta}{b}\right). \tag*{(140)} R → ∞ lim π R P ( D > a R , Y > b R ) = E min ( a 1 , b G η ) . (140) The same formula holds for the two lower tails. Opposite-sign tails satisfy
lim R → ∞ π R P ( D > a R , Y < − b R ) = 0. (141) \lim_{R\to\infty}\pi R\mathbb{P}(D>aR,Y<-bR)=0. \tag*{(141)} R → ∞ lim π R P ( D > a R , Y < − b R ) = 0. (141) Proof. Disintegrate over D D D and put d = R r d=Rr d = R r :
π R P ( D > a R , Y > b R ) = ∫ a ∞ R 2 1 + R 2 r 2 P ( Y R r > b r | D = R r ) d r . \pi R\mathbb{P}(D>aR,Y>bR)=\int_a^\infty\frac{R^2}{1+R^2r^2}\mathbb{P}\left(\frac{Y}{Rr}>\frac{b}{r}\,\middle|\,D=Rr\right)\,\mathrm{d}r. π R P ( D > a R , Y > b R ) = ∫ a ∞ 1 + R 2 r 2 R 2 P ( R r Y > r b D = R r ) d r . Proposition 14.1 identifies the pointwise limit. The integrand is bounded by r − 2 r^{-2} r − 2 , which is integrable on [ a , ∞ ) [a,\infty) [ a , ∞ ) . Dominated convergence and Tonelli’s theorem give
lim R → ∞ π R P ( D > a R , Y > b R ) = ∫ a ∞ r − 2 P ( G η > b / r ) d r = E [ ∫ a ∞ r − 2 1 { r > b / G η } d r ] = E min ( 1 / a , G η / b ) . \begin{aligned}
\lim_{R\to\infty}\pi R\mathbb{P}(D>aR,Y>bR)
&=\int_a^\infty r^{-2}\mathbb{P}(G_\eta>b/r)\,\mathrm{d}r\\
&=\mathbb{E}\left[\int_a^\infty r^{-2}\mathbf{1}_{\{r>b/G_\eta\}}\,\mathrm{d}r\right]
=\mathbb{E}\min(1/a,G_\eta/b).
\end{aligned} R → ∞ lim π R P ( D > a R , Y > b R ) = ∫ a ∞ r − 2 P ( G η > b / r ) d r = E [ ∫ a ∞ r − 2 1 { r > b / G η } d r ] = E min ( 1/ a , G η / b ) . For opposite signs the limiting probability is zero, since G η > 0 G_\eta>0 G η > 0 . Reflection proves the lower-tail assertion. □
For an explicit unequal-threshold formula, let P ( s , z ) = γ ( s , z ) / Γ ( s ) P(s,z)=\gamma(s,z)/\Gamma(s) P ( s , z ) = γ ( s , z ) /Γ ( s ) and Q ( s , z ) = 1 − P ( s , z ) Q(s,z)=1-P(s,z) Q ( s , z ) = 1 − P ( s , z ) denote the regularized incomplete gamma functions. Splitting the expectation at G η = b / a G_\eta=b/a G η = b / a gives
E min ( 1 / a , G η / b ) = 1 a Q ( η , η b a ) + 1 b P ( η + 1 , η b a ) . (142) \mathbb{E}\min(1/a,G_\eta/b)=\frac{1}{a}Q\left(\eta,\frac{\eta b}{a}\right)+\frac{1}{b}P\left(\eta+1,\frac{\eta b}{a}\right). \tag*{(142)} E min ( 1/ a , G η / b ) = a 1 Q ( η , a η b ) + b 1 P ( η + 1 , a η b ) . (142)
Corollary 14.4. The probability that Y Y Y exceeds the same high threshold as D D D has the limit
λ η : = lim R → ∞ P ( Y > R ∣ D > R ) = 1 − η η e − η Γ ( η + 1 ) . (143) \lambda_\eta:=\lim_{R\to\infty}\mathbb{P}(Y>R\mid D>R)=1-\frac{\eta^\eta e^{-\eta}}{\Gamma(\eta+1)}. \tag*{(143)} λ η := R → ∞ lim P ( Y > R ∣ D > R ) = 1 − Γ ( η + 1 ) η η e − η . (143)
It is strictly increasing in η \eta η , from zero to one.
Proof. Use π R P ( D > R ) → 1 \pi R\mathbb{P}(D>R)\to1 π R P ( D > R ) → 1 and set a = b = 1 a=b=1 a = b = 1 in (142) . The incomplete-gamma recurrence reduces the answer to (143) .
For monotonicity, take 0 < α < β 0<\alpha<\beta 0 < α < β and independent gamma variables A ∼ Γ ( α , 1 ) A\sim\Gamma(\alpha,1) A ∼ Γ ( α , 1 ) , B ∼ Γ ( β − α , 1 ) B\sim\Gamma(\beta-\alpha,1) B ∼ Γ ( β − α , 1 ) , where the second parameter is the rate. Changing variables from ( A , B ) (A,B) ( A , B ) to ( A + B , A / ( A + B ) ) (A+B,A/(A+B)) ( A + B , A / ( A + B )) in their product density shows that the ratio is independent of the sum and has beta mean α / β \alpha/\beta α / β . Consequently,
E [ A / α ∣ A + B ] = ( A + B ) / β . \mathbb{E}[A/\alpha\mid A+B]=(A+B)/\beta. E [ A / α ∣ A + B ] = ( A + B ) / β .
Apply conditional Jensen to the concave function u ↦ min ( 1 , u ) u\mapsto\min(1,u) u ↦ min ( 1 , u ) and take expectations: E min ( 1 , A / α ) ≤ E min ( 1 , ( A + B ) / β ) \mathbb{E}\min(1,A/\alpha)\leq\mathbb{E}\min(1,(A+B)/\beta) E min ( 1 , A / α ) ≤ E min ( 1 , ( A + B ) / β ) . The inequality is strict: when A + B > α A+B>\alpha A + B > α , the conditional law of A / α A/\alpha A / α places positive mass on both sides of its corner at one. The endpoint limits follow from the displayed formula.
For example, λ 1 = 1 − e − 1 \lambda_1=1-e^{-1} λ 1 = 1 − e − 1 . More precisely, the gamma expansion at zero and Stirling’s formula give
λ η = η log ( 1 / η ) + ( 1 − γ ) η + O ( η 2 log 2 η ) \lambda_\eta=\eta\log(1/\eta)+(1-\gamma)\eta+O\left(\eta^2\log^2\eta\right) λ η = η log ( 1/ η ) + ( 1 − γ ) η + O ( η 2 log 2 η )
( η ↓ 0 ) (\eta\downarrow0) ( η ↓ 0 ) . {#eq:14.9}
1 − λ η = 1 2 π η ( 1 − 1 12 η + O ( η − 2 ) ) 1-\lambda_\eta=\frac{1}{\sqrt{2\pi\eta}}\left(1-\frac{1}{12\eta}+O(\eta^{-2})\right) 1 − λ η = 2 π η 1 ( 1 − 12 η 1 + O ( η − 2 ) )
( η → ∞ ) (\eta\to\infty) ( η → ∞ ) . {#eq:14.10}
Here γ \gamma γ is Euler’s constant. These finite quantities describe dependence even though a covariance cannot do so.
From independence to coincidence
Proposition 14.5. As η ↓ 0 \eta\downarrow0 η ↓ 0 , the law of ( D , Y ) (D,Y) ( D , Y ) tends in total variation to the product of two standard Cauchy laws. As η → ∞ \eta\to\infty η → ∞ , it tends weakly to the diagonal law ( D , D ) (D,D) ( D , D ) , with D ∼ m D\sim m D ∼ m .
Proof. At the first endpoint, Γ ( z ) ∼ 1 / z \Gamma(z)\sim1/z Γ ( z ) ∼ 1/ z gives, for fixed y y y ,
η C η ( η y ) ⟶ 1 π ( 1 + y 2 ) . \eta C_\eta(\eta y)\longrightarrow\frac{1}{\pi(1+y^2)}. η C η ( η y ) ⟶ π ( 1 + y 2 ) 1 . The remaining factor ( 1 + d 2 ) − η / 2 e η y arctan d (1+d^2)^{-\eta/2}e^{\eta y\arctan d} ( 1 + d 2 ) − η /2 e η y a r c t a n d tends to one. Hence p η ( d , y ) p_\eta(d,y) p η ( d , y ) converges pointwise to the product Cauchy density; Scheffé’s lemma gives total-variation convergence.
At the other endpoint, (137) and Chebyshev’s inequality give, for each ε > 0 \varepsilon>0 ε > 0 ,
P ( ∣ Y − d ∣ > ε ∣ D = d ) ≤ min ( 1 , 1 + d 2 η ε 2 ) . \mathbb{P}(|Y-d|>\varepsilon\mid D=d)\leq\min\left(1,\frac{1+d^2}{\eta\varepsilon^2}\right). P ( ∣ Y − d ∣ > ε ∣ D = d ) ≤ min ( 1 , η ε 2 1 + d 2 ) . Integrating against m m m and applying dominated convergence shows that Y − D → 0 Y-D\to0 Y − D → 0 in probability. This proves the weak limit using only the bounded conditional probability estimate.
Laguerre quadrature at the boundary
The finite couplings on spectral curves have the same boundary mechanism. Let ℓ j , N \ell_{j,N} ℓ j , N and α j , N \alpha_{j,N} α j , N be the nodes and normalized weights of N N N -point Gaussian quadrature for u η − 1 e − u d u / Γ ( η ) u^{\eta-1}e^{-u}\,du/\Gamma(\eta) u η − 1 e − u d u /Γ ( η ) . Equivalently, they are the eigenvalues and first-coordinate spectral weights of A N A_N A N .
Corollary 14.6. Under Π η , N \Pi_{\eta,N} Π η , N put Y N = X / η Y_N=X/\eta Y N = X / η , and let
G η , N ∼ ∑ j = 1 N α j , N δ ℓ j , N / η . G_{\eta,N} \sim\sum_{j=1}^{N}\alpha_{j,N}\delta_{\ell_{j,N}/\eta}. G η , N ∼ j = 1 ∑ N α j , N δ ℓ j , N / η . For a , b > 0 a,b>0 a , b > 0 ,
lim R → ∞ π R P ( D > a R , Y N > b R ) = E min ( 1 / a , G η , N / b ) . (144) \lim_{R\to\infty}\pi R\mathbb{P}(D>aR,Y_N>bR)=\mathbb{E}\min(1/a,G_{\eta,N}/b). \tag*{(144)} R → ∞ lim π R P ( D > a R , Y N > b R ) = E min ( 1/ a , G η , N / b ) . (144) In particular,
λ η , N : = lim R → ∞ P ( Y N > R ∣ D > R ) \lambda_{\eta,N}:=\lim_{R\to\infty}\mathbb{P}(Y_N>R\mid D>R) λ η , N := R → ∞ lim P ( Y N > R ∣ D > R ) = ∑ j = 1 N α j , N min ( 1 , ℓ j , N / η ) ⟶ λ η ( N → ∞ ) . (145) =\sum_{j=1}^{N}\alpha_{j,N}\min(1,\ell_{j,N}/\eta)\longrightarrow\lambda_\eta\qquad(N\to\infty). \tag*{(145)} = j = 1 ∑ N α j , N min ( 1 , ℓ j , N / η ) ⟶ λ η ( N → ∞ ) . (145)
Proof. As d → + ∞ d\to+\infty d → + ∞ , we have d − 1 H N ( d ) = A N + d − 1 B N → A N d^{-1}H_N(d)=A_N+d^{-1}B_N\to A_N d − 1 H N ( d ) = A N + d − 1 B N → A N in matrix norm. Continuity of simple eigenvalues and their spectral projections gives
x j ( d ) / d → ℓ j , N , w j ( d ) → α j , N . x_j(d)/d\to\ell_{j,N},\qquad w_j(d)\to\alpha_{j,N}. x j ( d ) / d → ℓ j , N , w j ( d ) → α j , N .
Thus Y N / d Y_N/d Y N / d , conditional on D = d D=d D = d , converges to G η , N G_{\eta,N} G η , N . Repeat the proof of Theorem 14.3 . The limiting conditional probability can jump only when b / r b/r b / r equals one of its finitely many atoms; these exceptional r r r ’s do not affect dominated convergence.
The laws of G η , N G_{\eta,N} G η , N are the Gaussian quadrature measures for G η G_\eta G η and converge weakly to it by the same moment argument used in Theorem 8.1 : their common mean gives tightness on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , higher matched moments give uniform integrability for each fixed power, and the gamma law is determined by its moment generating function near zero. Finally, u ↦ min ( 1 , u ) u\mapsto\min(1,u) u ↦ min ( 1 , u ) is bounded and continuous. □ \square □
The first quadrature has G η , 1 = 1 G_{\eta,1}=1 G η , 1 = 1 , hence λ η , 1 = 1 \lambda_{\eta,1}=1 λ η , 1 = 1 . The convergence in (145) follows from the boundary quadrature calculation; weak convergence of the joint measures alone would not imply convergence of their tail coefficients.
For a small explicit check, the two Laguerre nodes are η + 1 ∓ η + 1 \eta+1\mp\sqrt{\eta+1} η + 1 ∓ η + 1 , with weights ( 1 ± 1 / η + 1 ) / 2 (1\pm1/\sqrt{\eta+1})/2 ( 1 ± 1/ η + 1 ) /2 . They give
λ η , 2 = 1 − 1 2 η + 1 . \lambda_{\eta,2}=1-\frac{1}{2\sqrt{\eta+1}}. λ η , 2 = 1 − 2 η + 1 1 .
There is a second useful connection with the finite fraction:
E 1 G η , N = η ∑ j α j , N ℓ j , N = η ( A N − 1 ) 00 = ρ x , N ( R ) . (146) \mathbb{E}\frac{1}{G_{\eta,N}}=\eta\sum_j\frac{\alpha_{j,N}}{\ell_{j,N}}=\eta(A_N^{-1})_{00}=\rho_{x,N}(\mathbb{R}). \tag*{(146)} E G η , N 1 = η j ∑ ℓ j , N α j , N = η ( A N − 1 ) 00 = ρ x , N ( R ) . (146)
The finite Wall mass is exactly the reciprocal moment of its gamma quadrature. In the limit, the gamma reciprocal moment equals η / ( η − 1 ) \eta/(\eta-1) η / ( η − 1 ) for η > 1 \eta>1 η > 1 and is infinite for η ≤ 1 \eta\leq1 η ≤ 1 . This gives the threshold in (134) a probabilistic meaning. The explicit sum there, rather than weak convergence alone, justifies the convergence or divergence of these unbounded moments.
Why the conditional transform is not compact
The boundary gamma law also answers an operator-theoretic question. Invariance of m m m and conditional Jensen imply
∥ Q η f ∥ L 2 ( m ) 2 ≤ ∫ ∣ Q η ∣ f ∣ ∣ 2 d m = ∥ f ∥ L 2 ( m ) 2 . \lVert Q_\eta f\rVert_{L^2(m)}^2 \le\int|Q_\eta|f||^2\,\mathrm{d}m = \lVert f\rVert_{L^2(m)}^2. ∥ Q η f ∥ L 2 ( m ) 2 ≤ ∫ ∣ Q η ∣ f ∣ ∣ 2 d m = ∥ f ∥ L 2 ( m ) 2 .
Thus Q η Q_\eta Q η preserves nonnegative functions and is a contraction. Its smooth kernel might suggest compactness, but its behavior far out in the tails prevents it.
Proposition 14.7. For every η > 0 \eta> 0 η > 0 , Q η Q_\eta Q η is not compact on L 2 ( m ) L^2(m) L 2 ( m ) .
Proof. Set
f R = 1 [ R , 2 R ] m ( [ R , 2 R ] ) . f_R = \frac{\mathbf{1}_{[R,2R]}}{\sqrt{m([R,2R])}}. f R = m ([ R , 2 R ]) 1 [ R , 2 R ] . These functions have norm one and converge weakly to zero: for h ∈ L 2 ( m ) h \in L^2(m) h ∈ L 2 ( m ) ,
∣ ⟨ f R , h ⟩ ∣ ≤ ( ∫ R ∞ ∣ h ∣ 2 d m ) 1 / 2 ⟶ 0. |\langle f_R,h\rangle| \le\left(\int_R^\infty|h|^2\,\mathrm{d}m\right)^{1/2} \longrightarrow0. ∣ ⟨ f R , h ⟩ ∣ ≤ ( ∫ R ∞ ∣ h ∣ 2 d m ) 1/2 ⟶ 0. Put c η = 1 2 P ( 1 < G η < 4 / 3 ) > 0 c_\eta= \frac{1}{2}\mathbb{P}(1 < G_\eta< 4/3) > 0 c η = 2 1 P ( 1 < G η < 4/3 ) > 0 . Proposition 14.1 shows that, for all sufficiently large d d d ,
P ( 1 < Y / d < 4 / 3 ∣ D = d ) ≥ c η . \mathbb{P}(1 < Y/d < 4/3 \mid D=d) \ge c_\eta. P ( 1 < Y / d < 4/3 ∣ D = d ) ≥ c η . If R ≤ d ≤ 3 R / 2 R \le d \le3R/2 R ≤ d ≤ 3 R /2 , this event implies R < Y < 2 R R < Y < 2R R < Y < 2 R . Therefore
∥ Q η f R ∥ 2 2 ≥ c η 2 m ( [ R , 3 R / 2 ] ) m ( [ R , 2 R ] ) ⟶ 2 3 c η 2 > 0. \lVert Q_\eta f_R\rVert_2^2 \ge c_\eta^2 \frac{m([R,3R/2])}{m([R,2R])} \longrightarrow\frac{2}{3}c_\eta^2 > 0. ∥ Q η f R ∥ 2 2 ≥ c η 2 m ([ R , 2 R ]) m ([ R , 3 R /2 ]) ⟶ 3 2 c η 2 > 0. A compact operator would send this bounded weakly null sequence to a norm-null sequence, a contradiction.
In particular, the transform is not Hilbert–Schmidt. The two conditional laws therefore provide a useful positive transform and its adjoint without automatically supplying a compact spectral decomposition.
The two measures, their finite quadratures, and their behavior at infinity have now emerged from the same positive kernel. We can return to the word bispectral with a clearer picture of what each spectral direction means.
What is spectral, and what is bispectral?
The two-variable kernel is concrete enough that we can distinguish several structures often gathered under the word “bispectrality.” There is the classical relation between degree and argument, the parameter duality of the positive pencil, and a representation-theoretic explanation for the recurrence itself. Each answers a different question.
The classical difference operator
In the argument x x x , define the fixed difference operator
D δ , η f ( x ) = ( 1 + i δ ) ( η − i x ) 4 [ f ( x + 2 i ) − f ( x ) ] + ( 1 − i δ ) ( η + i x ) 4 [ f ( x − 2 i ) − f ( x ) ] . (147) \mathcal{D}_{\delta,\eta}f(x)=\frac{(1+i\delta)(\eta-ix)}{4}[f(x+2i)-f(x)]+\frac{(1-i\delta)(\eta+ix)}{4}[f(x-2i)-f(x)]. \tag*{(147)} D δ , η f ( x ) = 4 ( 1 + i δ ) ( η − i x ) [ f ( x + 2 i ) − f ( x )] + 4 ( 1 − i δ ) ( η + i x ) [ f ( x − 2 i ) − f ( x )] . (147)
Then
D δ , η M n = n M n . (148) \mathcal{D}_{\delta,\eta}M_n=nM_n. \tag*{(148)} D δ , η M n = n M n . (148)
The imaginary shifts are harmless on polynomials. To prove the identity, factor the generating function as
G ( x ; t ) = A − a B − b , A = 1 + ( δ + i ) t , B = 1 + ( δ − i ) t , a = η + i x 2 , b = η − i x 2 . (149) G(x;t)=A^{-a}B^{-b},\qquad A=1+(\delta+i)t,\qquad B=1+(\delta-i)t,\qquad a=\frac{\eta+ix}{2},\qquad b=\frac{\eta-ix}{2}. \tag*{(149)} G ( x ; t ) = A − a B − b , A = 1 + ( δ + i ) t , B = 1 + ( δ − i ) t , a = 2 η + i x , b = 2 η − i x . (149)
Then G ( x + 2 i ; t ) / G ( x ; t ) = A / B G(x+2i;t)/G(x;t)=A/B G ( x + 2 i ; t ) / G ( x ; t ) = A / B and G ( x − 2 i ; t ) / G ( x ; t ) = B / A G(x-2i;t)/G(x;t)=B/A G ( x − 2 i ; t ) / G ( x ; t ) = B / A . Substitution in (147) gives
D δ , η G = − t ( ( δ + i ) a A + ( δ − i ) b B ) G = t ∂ t G . \mathcal{D}_{\delta,\eta}G=-t\left(\frac{(\delta+i)a}{A}+\frac{(\delta-i)b}{B}\right)G=t\partial_tG. D δ , η G = − t ( A ( δ + i ) a + B ( δ − i ) b ) G = t ∂ t G .
Coefficient comparison proves (148) . Together with the three-term recurrence in n n n , this is the usual bispectral structure of Meixner–Pollaczek polynomials [19 §9.7] .
In the parameter direction, the lowering identity and differential equation proved in Section 10 tell a different story. The first-derivative coefficient in (89) depends on n n n . This equation therefore does not exhibit one fixed second-order operator in δ \delta δ with all M n M_n M n as eigenfunctions. What we have established in the pair ( x , δ ) (x,\delta) ( x , δ ) is a precise parameter duality: an ordinary Jacobi spectral problem in one direction and a positive RII pencil in the other. Calling this bispectrality is useful provided that this distinction remains explicit.
Why the Lie algebra has these coefficients
There is a concise representation-theoretic reason for the factors n ( n + η − 1 ) n(n+\eta-1) n ( n + η − 1 ) . On the basis e 0 , e 1 , … e_0,e_1,\ldots e 0 , e 1 , … of ℓ 2 ( Z ≥ 0 ) \ell^2(\mathbb{Z}_{\ge0}) ℓ 2 ( Z ≥ 0 ) , define
K 0 e n = ( n + η / 2 ) e n , K_0e_n=(n+\eta/2)e_n, K 0 e n = ( n + η /2 ) e n ,
K + e n = ( n + 1 ) ( n + η ) e n + 1 , K_+e_n=\sqrt{(n+1)(n+\eta)}\,e_{n+1}, K + e n = ( n + 1 ) ( n + η ) e n + 1 ,
K − e n = n ( n + η − 1 ) e n − 1 , K − e 0 = 0. K_-e_n=\sqrt{n(n+\eta-1)}\,e_{n-1},\qquad K_-e_0=0. K − e n = n ( n + η − 1 ) e n − 1 , K − e 0 = 0.
Direct calculation gives
[ K 0 , K + ] = K + , [ K 0 , K − ] = − K − , [ K + , K − ] = − 2 K 0 , [K_0,K_+]=K_+,\qquad[K_0,K_-]=-K_-,\qquad[K_+,K_-]=-2K_0, [ K 0 , K + ] = K + , [ K 0 , K − ] = − K − , [ K + , K − ] = − 2 K 0 ,
and K + ∗ = K − K_+^*=K_- K + ∗ = K − on finite sequences. These are the positive discrete-series relations for s u ( 1 , 1 ) \mathfrak{su}(1,1) su ( 1 , 1 ) with lowest weight η / 2 \eta/2 η /2 . For arbitrary η > 0 \eta>0 η > 0 , the group statement is naturally made for the universal covering group; no integrality condition is needed for the Lie algebra construction [12 , 20 ] .
The Jacobi operator is simply
J δ = 2 δ K 0 + 1 + δ 2 ( K + + K − ) . (150) J_\delta=2\delta K_0+\sqrt{1+\delta^2}(K_++K_-). \tag*{(150)} J δ = 2 δ K 0 + 1 + δ 2 ( K + + K − ) . (150)
Normalize the polynomials by their norms:
p n ( x ) = M n ( x ; δ , η ) ( 1 + δ 2 ) n / 2 n ! ( η ) n . p_n(x)=\frac{M_n(x;\delta,\eta)}{(1+\delta^2)^{n/2}\sqrt{n!(\eta)_n}}. p n ( x ) = ( 1 + δ 2 ) n /2 n ! ( η ) n M n ( x ; δ , η ) .
The recurrence says exactly that the map e n ↦ p n e_n\mapsto p_n e n ↦ p n takes J δ J_\delta J δ to multiplication by x x x in L 2 ( μ δ ) L^2(\mu_\delta) L 2 ( μ δ ) . This map is onto by the polynomial completeness proved in § 5.4 . The Jacobi Carleman condition
∑ n ≥ 1 λ n − 1 / 2 = ∞ \sum_{n\geq1}\lambda_n^{-1/2}=\infty n ≥ 1 ∑ λ n − 1/2 = ∞
also gives essential self-adjointness on finite sequences [6 ] . The density K η ( x , δ ) K_\eta(x,\delta) K η ( x , δ ) is the spectral measure of the vector e 0 e_0 e 0 . The numbers p n ( x ) p_n(x) p n ( x ) are components of generalized eigenvectors; for fixed real x x x they are not being asserted to form an ℓ 2 \ell^2 ℓ 2 eigenvector.
The parameter δ \delta δ changes the operator inside the same representation. Put K 1 = ( K + + K − ) / 2 K_1=(K_++K_-)/2 K 1 = ( K + + K − ) /2 , K 2 = ( K + − K − ) / ( 2 i ) K_2=(K_+-K_-)/(2i) K 2 = ( K + − K − ) / ( 2 i ) , and δ = sinh r \delta=\sinh r δ = sinh r . The commutators give
e − i r K 2 K 1 e i r K 2 = cosh r K 1 + sinh r K 0 . e^{-irK_2}K_1e^{irK_2}=\cosh r K_1+\sinh r K_0. e − i r K 2 K 1 e i r K 2 = cosh r K 1 + sinh r K 0 .
One obtains this by differentiating the two conjugates of K 1 K_1 K 1 and K 0 K_0 K 0 ; they satisfy the two equations X ′ = Y X'=Y X ′ = Y , Y ′ = X Y'=X Y ′ = X . The identity holds first on the common analytic core and then for the self-adjoint closures. Hence
J δ = e − i r K 2 J 0 e i r K 2 . (151) J_\delta=e^{-irK_2}J_0e^{irK_2}. \tag*{(151)} J δ = e − i r K 2 J 0 e i r K 2 . (151)
The operators are unitarily conjugate, while their spectral measures at the fixed vector e 0 e_0 e 0 vary. This makes the exponential tilt compatible with the unchanged continuous spectrum.
The wider element 2 c K 0 + K + + K − 2cK_0+K_++K_- 2 c K 0 + K + + K − leads, after elementary affine changes, to Meixner-II, Laguerre, and Meixner-I according as ∣ c ∣ < 1 |c|<1 ∣ c ∣ < 1 , ∣ c ∣ = 1 |c|=1 ∣ c ∣ = 1 , and ∣ c ∣ > 1 |c|>1 ∣ c ∣ > 1 [12 ] . The same discriminant that appeared in the Sheffer classification has reappeared in the Lie algebra. Principal-series representations enter a different, larger story: mixed tensor products of positive and negative discrete series, continuous dual Hahn coefficients, and nonpolynomial Meixner functions [10 ] . They are not needed to explain the positive measure studied here.
What carries over from moment duality?
Ismail and Stanton’s Moment duality for orthogonal polynomials develops parameter duality of type R I R_{\mathrm{I}} R I [17 ] . Their Meixner example varies the shape parameter of the first-kind family. The coefficient multiplying the preceding polynomial is linear in that parameter. Here the conventional second-kind parametrization gives 1 + δ 2 1+\delta^2 1 + δ 2 , so the corresponding problem is of type R I I R_{\mathrm{II}} R II . The analogy is real, but the analytic consequences are different.
In the x x x direction, the density has exponential tails, and every ordinary moment exists. In the parameter direction,
K η ( x , t ) ∼ C η ( x ) e ± π x / 2 ∣ t ∣ − η ( t → ± ∞ ) . (152) K_\eta(x,t)\sim C_\eta(x)e^{\pm\pi x/2}|t|^{-\eta}\qquad(t\to\pm\infty). \tag*{(152)} K η ( x , t ) ∼ C η ( x ) e ± π x /2 ∣ t ∣ − η ( t → ± ∞ ) . (152)
For η > 1 \eta>1 η > 1 , the ordinary absolute moments of ν x \nu_x ν x exist only in orders k < η − 1 k<\eta-1 k < η − 1 ; those of σ x \sigma_x σ x exist only in orders k < η + 1 k<\eta+1 k < η + 1 . At fixed finite η \eta η , neither probability provides all ordinary polynomial moments in the parameter. What survives at every degree is the rational orthogonality (90) or (111) .
This observation also guides the combinatorics. The Motzkin model in § 7 naturally describes an ordinary Jacobi recurrence. Kim and Song’s theory of generalized R I I R_{\mathrm{II}} R II orthogonality provides the relevant broader path model [18 ] : vertical-down and backward-down steps join the usual up, level, and down steps, recording the additional terms of the quadratic coefficient. Their formal framework does not require a heavy-tailed positive measure to have moments of all orders. Applying their model to the present coefficients is a concrete combinatorial specialization, distinct from the positive-measure statement proved here.
The new probabilistic readings sharpen this distinction. The finite pencil already couples two exact Cauchy laws, before any limit is taken. Its zero curves approach Laguerre rays, and their quadrature weights approach the gamma law governing simultaneous extremes. At the other end of the paper, independent increments explain why Carlitz’s continuous Hahn measure appears when a Meixner process is conditioned on its endpoint. These are consequences of the same kernel, reached by asking different questions of it.
There are still concrete questions to ask. What is the essential spectrum of the Cauchy conditional operator Q η Q_\eta Q η ? How quickly does its Laguerre quadrature approximate the tail coefficient, especially when the corner of u ↦ min ( 1 , u ) u \mapsto\min(1,u) u ↦ min ( 1 , u ) lies close to a node? Which other positive R I I R_{II} R I I pencils admit finite couplings with two explicit marginals? The formulas above give starting points for these questions.
The kernel gives a useful test for any further analogy. In one direction it has exponential tails and ordinary orthogonality; in the other it has algebraic tails and rational orthogonality. The joint Cauchy law makes these two descriptions compatible without erasing their difference. What looked like a mystery about the number of measures is a question about how one kernel is read.
How much information survives conditioning?
The two polynomial identities in Proposition 5.1 determine an entire operator. Fix δ \delta δ and 0 < s < t 0 < s < t 0 < s < t , retaining the notation μ t ( δ ) ( d x ) = K t ( x , δ ) d x \mu_t^{(\delta)}(dx) = K_t(x,\delta)\,dx μ t ( δ ) ( d x ) = K t ( x , δ ) d x from Section 5.4. Define
T s , t : L 2 ( μ t ( δ ) ) ⟶ L 2 ( μ s ( δ ) ) , ( T s , t f ) ( y ) = ∫ R f ( y + z ) K t − s ( z , δ ) d z . T_{s,t}:L^2(\mu_t^{(\delta)}) \longrightarrow L^2(\mu_s^{(\delta)}), \qquad(T_{s,t}f)(y)=\int_{\mathbb{R}} f(y+z)K_{t-s}(z,\delta)\,dz. T s , t : L 2 ( μ t ( δ ) ) ⟶ L 2 ( μ s ( δ ) ) , ( T s , t f ) ( y ) = ∫ R f ( y + z ) K t − s ( z , δ ) d z .
This is conditional expectation from time t t t to time s s s ; its adjoint is conditional expectation in the reverse direction. Both are contractions by Jensen’s inequality. The complete polynomial bases make their action particularly simple.
Proposition A.1. Put p n ( t ) = M n ( t ) / h n ( t ) p_n^{(t)}=M_n^{(t)}/\sqrt{h_n(t)} p n ( t ) = M n ( t ) / h n ( t ) , with h n ( t ) h_n(t) h n ( t ) as in (28) . A complete singular system is
T s , t p n ( t ) = σ n p n ( s ) , T s , t ∗ p n ( s ) = σ n p n ( t ) , σ n = ( s ) n ( t ) n . (153) T_{s,t}p_n^{(t)}=\sigma_n p_n^{(s)}, \qquad T_{s,t}^{*}p_n^{(s)}=\sigma_n p_n^{(t)}, \qquad\sigma_n=\sqrt{\frac{(s)_n}{(t)_n}}. \tag*{(153)} T s , t p n ( t ) = σ n p n ( s ) , T s , t ∗ p n ( s ) = σ n p n ( t ) , σ n = ( t ) n ( s ) n . (153) The operator is compact for every t > s t>s t > s , and, for 0 < p < ∞ 0<p<\infty 0 < p < ∞ ,
T s , t ∈ S p ⟺ p ( t − s ) > 2. (154) T_{s,t}\in S_p \Longleftrightarrow p(t-s)>2. \tag*{(154)} T s , t ∈ S p ⟺ p ( t − s ) > 2. (154) Here S p S_p S p means that the p p p th powers of the singular values are summable.
Proof. Normalize the two identities of Proposition 5.1. The completeness proved there gives the entire singular system. Moreover,
σ n + 1 2 σ n 2 = n + s n + t < 1 , σ n ∼ Γ ( t ) Γ ( s ) n − ( t − s ) / 2 . \frac{\sigma_{n+1}^{2}}{\sigma_n^{2}}=\frac{n+s}{n+t}<1, \qquad\sigma_n\sim\sqrt{\frac{\Gamma(t)}{\Gamma(s)}}\,n^{-(t-s)/2}. σ n 2 σ n + 1 2 = n + t n + s < 1 , σ n ∼ Γ ( s ) Γ ( t ) n − ( t − s ) /2 .
The second assertion follows from the gamma ratio asymptotic. Thus the singular values tend to zero, proving compactness; summing their p p p th powers proves (A.2). □
In particular, the Hilbert–Schmidt threshold is t − s > 1 t-s>1 t − s > 1 , and the trace class threshold is t − s > 2 t-s>2 t − s > 2 . On mean-zero functions the exact norm is σ 1 = s / t \sigma_1=\sqrt{s/t} σ 1 = s / t . Equivalently, for square-integrable f , g f,g f , g ,
∣ Cov ( f ( X s ) , g ( X t ) ) ∣ ≤ s t Var ( f ( X s ) ) Var ( g ( X t ) ) . (155) \left|\operatorname{Cov}(f(X_s),g(X_t))\right|\le\sqrt{\frac{s}{t}}\sqrt{\operatorname{Var}(f(X_s))\operatorname{Var}(g(X_t))}. \tag*{(155)} ∣ Cov ( f ( X s ) , g ( X t )) ∣ ≤ t s Var ( f ( X s )) Var ( g ( X t )) . (155)
Equality is attained by the centered linear functions. The tilt changes the laws and their orthogonal bases, but leaves every singular value unchanged.
There is also an exact measure of dependence. For probability measures P ≪ Q P\ll Q P ≪ Q , define
χ 2 ( P ∥ Q ) = ∫ ( d P d Q − 1 ) 2 d Q , \chi^2(P\mathbin\Vert Q)=\int\left(\frac{dP}{dQ}-1\right)^2dQ, χ 2 ( P ∥ Q ) = ∫ ( d Q d P − 1 ) 2 d Q ,
allowing the value + ∞ +\infty + ∞ .
Corollary A.2. If P s , t P_{s,t} P s , t is the joint law of ( X s , X t ) (X_s,X_t) ( X s , X t ) , then
χ 2 ( P s , t ∥ μ s ( δ ) ⊗ μ t ( δ ) ) = { s / ( t − s − 1 ) , t − s > 1 , + ∞ , 0 < t − s ≤ 1. (156) \chi^2\left(P_{s,t}\mathbin\Vert\mu_s^{(\delta)}\otimes\mu_t^{(\delta)}\right)=
\begin{cases}
s/(t-s-1), & t-s>1,\\
+\infty, & 0<t-s\le1.
\end{cases} \tag*{(156)} χ 2 ( P s , t ∥ μ s ( δ ) ⊗ μ t ( δ ) ) = { s / ( t − s − 1 ) , + ∞ , t − s > 1 , 0 < t − s ≤ 1. (156) When finite, ∥ T s , t ∥ H S 2 = ( t − 1 ) / ( t − s − 1 ) \lVert T_{s,t}\rVert_{\mathrm{HS}}^2=(t-1)/(t-s-1) ∥ T s , t ∥ HS 2 = ( t − 1 ) / ( t − s − 1 ) .
Proof. Relative to the product of the marginal laws, the joint density is
R ( y , x ) = K t − s ( x − y , δ ) K t ( x , δ ) . R(y,x)=\frac{K_{t-s}(x-y,\delta)}{K_t(x,\delta)}. R ( y , x ) = K t ( x , δ ) K t − s ( x − y , δ ) .
For fixed y y y , the tail estimate (18) gives R ( y , x ) = O ( ∣ x ∣ − s ) R(y,x)=O(|x|^{-s}) R ( y , x ) = O ( ∣ x ∣ − s ) at infinity. Thus R ( y , ⋅ ) R(y,\cdot) R ( y , ⋅ ) is bounded and belongs to L 2 ( μ t ( δ ) ) L^2(\mu_t^{(\delta)}) L 2 ( μ t ( δ ) ) . Its n n n th orthogonal coefficient is T s , t p n ( t ) ( y ) = σ n p n ( s ) ( y ) T_{s,t}p_n^{(t)}(y)=\sigma_np_n^{(s)}(y) T s , t p n ( t ) ( y ) = σ n p n ( s ) ( y ) . Parseval followed by Tonelli gives, with infinite values allowed,
∫ R 2 d ( μ s ( δ ) ⊗ μ t ( δ ) ) = ∑ n ≥ 0 σ n 2 . \int R^2\,d\left(\mu_s^{(\delta)}\otimes\mu_t^{(\delta)}\right)=\sum_{n\ge0}\sigma_n^2. ∫ R 2 d ( μ s ( δ ) ⊗ μ t ( δ ) ) = n ≥ 0 ∑ σ n 2 .
The constant term is one, leaving χ 2 = ∑ n ≥ 1 σ n 2 \chi^2=\sum_{n\ge1}\sigma_n^2 χ 2 = ∑ n ≥ 1 σ n 2 .
To sum the series, put r = t − s r=t-s r = t − s and let B B B have beta density u s − 1 ( 1 − u ) r − 1 / B ( s , r ) u^{s-1}(1-u)^{r-1}/B(s,r) u s − 1 ( 1 − u ) r − 1 / B ( s , r ) on ( 0 , 1 ) (0,1) ( 0 , 1 ) . Then
∑ n ≥ 0 ( s ) n ( t ) n = E ∑ n ≥ 0 B n = E 1 1 − B = B ( s , r − 1 ) B ( s , r ) = t − 1 r − 1 \sum_{n\ge0}\frac{(s)_n}{(t)_n}=\mathbb{E}\sum_{n\ge0}B^n=\mathbb{E}\frac{1}{1-B}=\frac{B(s,r-1)}{B(s,r)}=\frac{t-1}{r-1} n ≥ 0 ∑ ( t ) n ( s ) n = E n ≥ 0 ∑ B n = E 1 − B 1 = B ( s , r ) B ( s , r − 1 ) = r − 1 t − 1
if r > 1 r>1 r > 1 . If r ≤ 1 r\le1 r ≤ 1 , the same integral diverges at 1 1 1 . This proves both assertions. □
The associated density expansion is an L 2 L^2 L 2 expansion precisely when t − s > 1 t-s>1 t − s > 1 ; the operator identities (A.1) hold for every positive time gap. Thus the amount of smoothing is measured exactly by elapsed time, degree by degree.
What the zero curves approach
The positive pencil does more than keep the zeros real: it identifies their directions at infinity. The eigenvalues of A N A_N A N are the zeros of L N ( n − 1 ) L_N^{(n-1)} L N ( n − 1 ) . Thus the Laguerre family appears as the asymptotic geometry of the same curves that support the finite coupling in Theorem 12.2.
A first correction to the Laguerre limit
Fix N ≥ 1 N \ge1 N ≥ 1 and η > 0 \eta> 0 η > 0 , and write
L N ( y ) = ( − 1 ) N N ! L N ( η − 1 ) ( y ) , 0 < ℓ 1 < ⋯ < ℓ N \mathcal{L}_N(y)=(-1)^N N!L_N^{(\eta-1)}(y), \qquad0<\ell_1<\cdots<\ell_N L N ( y ) = ( − 1 ) N N ! L N ( η − 1 ) ( y ) , 0 < ℓ 1 < ⋯ < ℓ N
for the monic Laguerre polynomial and its zeros. The leading Meixner–Pollaczek-to-Laguerre limit is classical; the corresponding reciprocal-Laguerre approximation for the parameter zeros appears in [33 Theorem 4.2] . Full polynomial expansions are available in [35 §4] . In the present normalization, the first correction has a particularly short derivation and an explicit expression involving only the limiting zero.
Proposition B.1. Let x 1 ( δ ) < ⋯ < x N ( δ ) x_1(\delta)<\cdots<x_N(\delta) x 1 ( δ ) < ⋯ < x N ( δ ) be the zeros of M N ( x ; δ , η ) M_N(x;\delta,\eta) M N ( x ; δ , η ) in x x x . As δ → + ∞ \delta\to+\infty δ → + ∞ ,
x j ( δ ) = δ ℓ j + β j δ + O ( δ − 3 ) , β j = 2 ℓ j − 2 N − η + η ( 2 − η ) / ℓ j 6 . (157) x_j(\delta)=\delta\ell_j+\frac{\beta_j}{\delta}+O(\delta^{-3}),\qquad\beta_j=\frac{2\ell_j-2N-\eta+\eta(2-\eta)/\ell_j}{6}. \tag*{(157)} x j ( δ ) = δ ℓ j + δ β j + O ( δ − 3 ) , β j = 6 2 ℓ j − 2 N − η + η ( 2 − η ) / ℓ j . (157) If d 1 ( x ) < ⋯ < d N ( x ) d_1(x)<\cdots<d_N(x) d 1 ( x ) < ⋯ < d N ( x ) are the zeros in the parameter, then
d k ( x ) = x ℓ N + 1 − k − β N + 1 − k x + O ( x − 3 ) , x → + ∞ . (158) d_k(x)=\frac{x}{\ell_{N+1-k}}-\frac{\beta_{N+1-k}}{x}+O(x^{-3}),\qquad x\to+\infty. \tag*{(158)} d k ( x ) = ℓ N + 1 − k x − x β N + 1 − k + O ( x − 3 ) , x → + ∞. (158) All error estimates here are for fixed N N N and η \eta η .
Proof. Put h = δ − 2 h=\delta^{-2} h = δ − 2 . Rescale the generating function by replacing x x x with δ y \delta y δ y and its series variable with u / δ u/\delta u / δ . This gives
∑ n ≥ 0 δ − n M n ( δ y ; δ , η ) u n n ! = [ ( 1 + u ) 2 + h u 2 ] − η / 2 exp { y h arctan h u 1 + u } = F 0 ( y , u ) [ 1 − h ( η r 2 2 + y r 3 3 ) + O ( h 2 ) ] , \begin{aligned}
\sum_{n\ge0}\delta^{-n}M_n(\delta y;\delta,\eta)\frac{u^n}{n!}
=\left[(1+u)^2+hu^2\right]^{-\eta/2}\exp\left\{\frac{y}{\sqrt{h}}\arctan\frac{\sqrt{h}u}{1+u}\right\} \\
&=F_0(y,u)\left[1-h\left(\frac{\eta r^2}{2}+\frac{yr^3}{3}\right)+O(h^2)\right],
\end{aligned} n ≥ 0 ∑ δ − n M n ( δ y ; δ , η ) n ! u n = [ ( 1 + u ) 2 + h u 2 ] − η /2 exp { h y arctan 1 + u h u } = F 0 ( y , u ) [ 1 − h ( 2 η r 2 + 3 y r 3 ) + O ( h 2 ) ] ,
where r = u / ( 1 + u ) r=u/(1+u) r = u / ( 1 + u ) and
F 0 ( y , u ) = ( 1 + u ) − y e y r = ∑ n ≥ 0 L n ( y ) u n n ! . F_0(y,u)=(1+u)^{-y}e^{yr}=\sum_{n\ge0}\mathcal{L}_n(y)\frac{u^n}{n!}. F 0 ( y , u ) = ( 1 + u ) − y e y r = n ≥ 0 ∑ L n ( y ) n ! u n .
The apparent square root causes no difficulty: the displayed expression is analytic in h h h near zero. Since ∂ y m F 0 = ( u / ( 1 + u ) ) m F 0 \partial_y^mF_0=(u/(1+u))^mF_0 ∂ y m F 0 = ( u / ( 1 + u ) ) m F 0 for every nonnegative integer m m m , coefficient comparison yields
δ − N M N ( δ y ; δ , η ) = L N ( y ) − δ − 2 ( η 2 L N ′ ′ ( y ) + y 3 L N ′ ′ ′ ( y ) ) + O ( δ − 4 ) . (159) \delta^{-N}M_N(\delta y;\delta,\eta)=\mathcal{L}_N(y)-\delta^{-2}\left(\frac{\eta}{2}\mathcal{L}_N''(y)+\frac{y}{3}\mathcal{L}_N'''(y)\right)+O(\delta^{-4}). \tag*{(159)} δ − N M N ( δ y ; δ , η ) = L N ( y ) − δ − 2 ( 2 η L N ′′ ( y ) + 3 y L N ′′′ ( y ) ) + O ( δ − 4 ) . (159)
This is a coefficientwise polynomial expansion. Each simple Laguerre zero therefore moves analytically with h h h , and its first displacement is
β j = η 2 L N ′ ′ ( ℓ j ) L N ′ ( ℓ j ) + ℓ j 3 L N ′ ′ ′ ( ℓ j ) L N ′ ( ℓ j ) . \beta_j=\frac{\eta}{2}\frac{\mathcal{L}_N''(\ell_j)}{\mathcal{L}_N'(\ell_j)}+\frac{\ell_j}{3}\frac{\mathcal{L}_N'''(\ell_j)}{\mathcal{L}_N'(\ell_j)}. β j = 2 η L N ′ ( ℓ j ) L N ′′ ( ℓ j ) + 3 ℓ j L N ′ ( ℓ j ) L N ′′′ ( ℓ j ) .
The Laguerre equation y L N ′ ′ + ( η − y ) L N ′ + N L N = 0 y\mathcal{L}_N''+(\eta-y)\mathcal{L}_N'+N\mathcal{L}_N=0 y L N ′′ + ( η − y ) L N ′ + N L N = 0 and its derivative give, at y = ℓ j y=\ell_j y = ℓ j ,
L N ′ ′ L N ′ = ℓ j − η ℓ j , L N ′ ′ ′ L N ′ = ( ℓ j − η − 1 ) ( ℓ j − η ) − ℓ j ( N − 1 ) ℓ j 2 . \frac{\mathcal{L}_N''}{\mathcal{L}_N'}=\frac{\ell_j-\eta}{\ell_j},\qquad
\frac{\mathcal{L}_N'''}{\mathcal{L}_N'}=\frac{(\ell_j-\eta-1)(\ell_j-\eta)-\ell_j(N-1)}{\ell_j^2}. L N ′ L N ′′ = ℓ j ℓ j − η , L N ′ L N ′′′ = ℓ j 2 ( ℓ j − η − 1 ) ( ℓ j − η ) − ℓ j ( N − 1 ) .
Substitution simplifies to (B.1). Finally, inverting x = ℓ j δ + β j / δ + O ( δ − 3 ) x=\ell_j\delta+\beta_j/\delta+O(\delta^{-3}) x = ℓ j δ + β j / δ + O ( δ − 3 ) gives δ = x / ℓ j − β j / x + O ( x − 3 ) \delta=x/\ell_j-\beta_j/x+O(x^{-3}) δ = x / ℓ j − β j / x + O ( x − 3 ) . The order reverses because the positive numbers 1 / ℓ j 1/\ell_j 1/ ℓ j decrease. □
Remark B.2. Reflection, M N ( − x ; − δ , η ) = ( − 1 ) N M N ( x ; δ , η ) M_N(-x;-\delta,\eta)=(-1)^NM_N(x;\delta,\eta) M N ( − x ; − δ , η ) = ( − 1 ) N M N ( x ; δ , η ) , gives the negative-infinity formulas with the indices reversed.
The angle reveals a convex equilibrium
The parameter zeros also admit an electrostatic interpretation, already established for complementary Romanovski–Routh polynomials in [23 §3] . The substitution d j = tan θ j d_j=\tan\theta_j d j = tan θ j makes both uniqueness and motion of the equilibrium particularly clear.
Proposition B.3. The angles θ j ( x ) = arctan d j ( x ) \theta_j(x)=\arctan d_j(x) θ j ( x ) = arctan d j ( x ) uniquely minimize
E x ( θ ) = − 2 ∑ j < k log sin ( θ k − θ j ) − η ∑ j log cos θ j − x ∑ j θ j (160) E_x(\theta)=-2\sum_{j<k}\log\sin(\theta_k-\theta_j)-\eta\sum_j\log\cos\theta_j-x\sum_j\theta_j \tag*{(160)} E x ( θ ) = − 2 j < k ∑ log sin ( θ k − θ j ) − η j ∑ log cos θ j − x j ∑ θ j (160) on − π / 2 < θ 1 < ⋯ < θ N < π / 2 -\pi/2<\theta_1<\cdots<\theta_N<\pi/2 − π /2 < θ 1 < ⋯ < θ N < π /2 . Moreover,
0 < θ j ′ ( x ) = d j ′ ( x ) 1 + d j ( x ) 2 ≤ 1 η . (161) 0<\theta'_j(x)=\frac{d'_j(x)}{1+d_j(x)^2}\le\frac{1}{\eta}. \tag*{(161)} 0 < θ j ′ ( x ) = 1 + d j ( x ) 2 d j ′ ( x ) ≤ η 1 . (161) The upper bound is strict when N > 1 N>1 N > 1 .
Proof. At a parameter zero, the differential equation (89) gives
2 ∑ k ≠ j 1 d j − d k = ( η + 2 N − 2 ) d j − x 1 + d j 2 . 2\sum_{k\ne j}\frac{1}{d_j-d_k}=\frac{(\eta+2N-2)d_j-x}{1+d_j^2}. 2 k = j ∑ d j − d k 1 = 1 + d j 2 ( η + 2 N − 2 ) d j − x .
Using cot ( θ j − θ k ) = ( 1 + d j 2 ) / ( d j − d k ) − d j \cot(\theta_j-\theta_k)=(1+d_j^2)/(d_j-d_k)-d_j cot ( θ j − θ k ) = ( 1 + d j 2 ) / ( d j − d k ) − d j , this becomes
2 ∑ k ≠ j cot ( θ j − θ k ) = η tan θ j − x , 2\sum_{k\ne j}\cot(\theta_j-\theta_k)=\eta\tan\theta_j-x, 2 k = j ∑ cot ( θ j − θ k ) = η tan θ j − x ,
which is exactly ∇ E x = 0 \nabla E_x=0 ∇ E x = 0 . If H H H is the Hessian, then
v T H v = η ∑ j v j 2 sec 2 θ j + 2 ∑ j < k ( v j − v k ) 2 csc 2 ( θ j − θ k ) > 0 ( v ≠ 0 ) . v^{\mathsf{T}}Hv=\eta\sum_jv_j^2\sec^2\theta_j+2\sum_{j<k}(v_j-v_k)^2\csc^2(\theta_j-\theta_k)>0\qquad(v\ne0). v T H v = η j ∑ v j 2 sec 2 θ j + 2 j < k ∑ ( v j − v k ) 2 csc 2 ( θ j − θ k ) > 0 ( v = 0 ) .
Thus E x E_x E x is strictly convex. It tends to infinity at every boundary of the ordered chamber, proving existence and uniqueness.
Differentiating the equilibrium equations gives H θ ′ = 1 H\theta'=1 H θ ′ = 1 . Write u = θ ′ u=\theta' u = θ ′ . At an index where u j u_j u j is smallest,
1 = ( H u ) j = η sec 2 θ j u j + 2 ∑ k ≠ j csc 2 ( θ j − θ k ) ( u j − u k ) ≤ η sec 2 θ j u j , 1=(Hu)_j=\eta\sec^2\theta_j\,u_j+2\sum_{k\ne j}\csc^2(\theta_j-\theta_k)(u_j-u_k)\le\eta\sec^2\theta_j\,u_j, 1 = ( H u ) j = η sec 2 θ j u j + 2 k = j ∑ csc 2 ( θ j − θ k ) ( u j − u k ) ≤ η sec 2 θ j u j ,
so every component is positive. At a largest component the inequality reverses, giving u j ≤ cos 2 θ j / η ≤ 1 / η u_j\le\cos^2\theta_j/\eta\le1/\eta u j ≤ cos 2 θ j / η ≤ 1/ η . Equality in the last bound would force all components of u u u to agree and all angles to vanish, which is impossible for N > 1 N>1 N > 1 . □
The energy describes repelling points confined between two endpoints. Increasing x x x applies a uniform force to the right; strict convexity ensures a unique response, and (B.5) bounds the speed of every point.
Acknowledgement
The author is grateful to Maxim Derevyagin for asking about bispectrality of Meixner-II polynomials and for pointing out his relevant work.
AI Declaration
Artificial intelligence (AI) tools were used to generate the graphics for this paper. An LLM was also used to audit its earlier draft and assist in the rewriting. The author assumes full responsibility for all results presented here.