2020 MSC: 05A15; 05A20; 15B48; 44A60.
Introduction
Total positivity of matrices has been deeply studied in various branches of mathematics. For example, Brenti [4 ] developed total positivity to study unimodality and log-concavity of sequences. Lusztig extended total positivity to reductive groups [24 ] . Fomin and Zelevinsky [11 ] studied totally nonnegative parts of double Bruhat cells and their parametrizations. Rietsch related totally positive Toeplitz matrices to the quantum cohomology of partial flag varieties [34 ] . Kodama and Williams [19 ] connected the totally nonnegative Grassmannian with regular soliton solutions of the Kadomtsev–Petviashvili equation. Belton, Guillot, Khare and Putinar [2 ] obtained a classification for entrywise polynomial preservers of positive semidefiniteness in fixed dimension. In their work on moment-sequence transforms [3 ] , they classified entrywise maps preserving moment sequences and totally nonnegative Hankel matrices, and also treated multivariable transforms. We also refer the reader to monographs [18 , 32 ] for more details.
Recent years, there many important progress on coefficientwise Hankel-total positivity. For example, in [31 ] , Pétréolle, Sokal and Zhu developed the theory for coefficientwise Hankel-total positivity based on lattice paths and branched continued fractions. See Zhu [48 , 49 ] for methods from production matrices and Riordan arrays in combinatorics. There are still many difficult open problems on coefficientwise total positivity. We also refer the reader to [7 , 31 , 38 , 30 , 31 , 44 , 45 , 46 , 48 , 49 ] for more details. In this present paper, we will develop the probabilistic method to prove coefficientwise Hankel-total positivity. We address the functional-digraph problem within this development by characterizing the exponential cumulant families for which a universal coefficientwise lifting property holds.
Functional digraphs and Sokal’s conjecture
A functional digraph on [ n ] = { 1 , … , n } [n] = \{1, \ldots, n\} [ n ] = { 1 , … , n } is the directed graph of a map from [ n ] [n] [ n ] to itself: every vertex has out-degree one. Each weakly connected component consists of a directed cycle with rooted trees attached to its cyclic vertices. Let ψ n , k \psi_{n,k} ψ n , k be the number of such digraphs with k k k components, and put
ψ n ( y ) = ∑ k = 0 n ψ n , k y k , ψ 0 ( y ) = 1. (1) \psi_n(y) = \sum_{k=0}^{n} \psi_{n,k} y^k,\qquad\psi_0(y) = 1. \tag*{(1)} ψ n ( y ) = k = 0 ∑ n ψ n , k y k , ψ 0 ( y ) = 1. (1)
We call ψ n ( y ) \psi_n(y) ψ n ( y ) the functional digraph component polynomial . For rooted labeled trees on n n n vertices, we write T ( t ) T(t) T ( t ) for the corresponding exponential generating function. It is known that
T ( t ) = ∑ n ≥ 1 n n − 1 t n n ! , T ( t ) = t e T ( t ) . (2) T(t) = \sum_{n\geq1} n^{n-1}\frac{t^n}{n!},\qquad T(t) = t e^{T(t)}. \tag*{(2)} T ( t ) = n ≥ 1 ∑ n n − 1 n ! t n , T ( t ) = t e T ( t ) . (2)
By the exponential formula [41 Corollary 5.1.6] , we obtain
∑ n ≥ 0 ψ n ( y ) t n n ! = ( 1 − T ( t ) ) − y . (3) \sum_{n\geq0}\psi_n(y)\frac{t^n}{n!}=(1-T(t))^{-y}. \tag*{(3)} n ≥ 0 ∑ ψ n ( y ) n ! t n = ( 1 − T ( t ) ) − y . (3)
Sokal studied these functional digraph component polynomials and proposed the next conjecture [39 Conjecture 6.9(b)] .
Conjecture 1.1 ([39 ] ). The functional digraph component polynomial sequence ( ψ n ( y ) ) n ≥ 0 (\psi_n(y))_{n\geq0} ( ψ n ( y ) ) n ≥ 0 is coefficientwise Hankel-totally positive in y y y .
Chen and Sokal [5 ] studied related row-generating polynomials associated with the entries ( n k ) n n − k \binom{n}{k}n^{n-k} ( k n ) n n − k , using a model of partial functional digraphs. To describe the progress on this conjecture, we assign an additional weight q q q to each cyclic vertex and denote the resulting polynomials by ψ n ( q , y ) \psi_n(q,y) ψ n ( q , y ) . Their generating functions are
∑ n ≥ 0 ψ n ( q , y ) t n n ! = ( 1 − q T ( t ) ) − y , (4) \sum_{n\geq0}\psi_n(q,y)\frac{t^n}{n!}=(1-qT(t))^{-y}, \tag*{(4)} n ≥ 0 ∑ ψ n ( q , y ) n ! t n = ( 1 − q T ( t ) ) − y , (4)
C q ( t ) : = − log ( 1 − q T ( t ) ) = ∑ n ≥ 1 c n ( q ) t n n ! . (5) C_q(t):=-\log(1-qT(t))=\sum_{n\geq1}c_n(q)\frac{t^n}{n!}. \tag*{(5)} C q ( t ) := − log ( 1 − q T ( t )) = n ≥ 1 ∑ c n ( q ) n ! t n . (5)
By (22) , we extract the coefficient of y y y in ψ n ( q , y ) \psi_n(q,y) ψ n ( q , y ) to select precisely the digraphs with one weakly connected component. Hence c n ( q ) = [ y ] ψ n ( q , y ) c_n(q)=[y]\psi_n(q,y) c n ( q ) = [ y ] ψ n ( q , y ) . At q = 1 q=1 q = 1 , we recover ψ n ( 1 , y ) = ψ n ( y ) \psi_n(1,y)=\psi_n(y) ψ n ( 1 , y ) = ψ n ( y ) . Conjecture 1.1 was also stated as Conjecture 6.16(ii) in [48 ] and the following result was proved in [48 Proposition 6.18] .
Theorem 1.2 ([48 ] ). For fixed q , y ≥ 0 q,y\geq0 q , y ≥ 0 , the sequence ( ψ n ( q , y ) ) n ≥ 0 (\psi_n(q,y))_{n\geq0} ( ψ n ( q , y ) ) n ≥ 0 is a numerical Stieltjes moment sequence. In particular, for each fixed q ≥ 0 q\geq0 q ≥ 0 , ( c n + 1 ( q ) ) n ≥ 0 (c_{n+1}(q))_{n\geq0} ( c n + 1 ( q ) ) n ≥ 0 is a Stieltjes moment sequence.
The result concerns fixed nonnegative parameter values. Coefficientwise Hankel-total positivity is stronger than Stieltjes moment property. In Conjecture 1.1 , we require every coefficient of each Hankel minor to be nonnegative. We extend Theorem 1.2 to the following result for coefficientwise Hankel-total positivity, which in particular implies Conjecture 1.1 .
Theorem 1.3 (Functional-digraph Hankel-total positivity). For every fixed q ≥ 0 q\geq0 q ≥ 0 , the sequence ( ψ n ( q , y ) ) n ≥ 0 (\psi_n(q,y))_{n\geq0} ( ψ n ( q , y ) ) n ≥ 0 is coefficientwise Hankel-totally positive in y y y . In particular, ( ψ n ( y ) ) n ≥ 0 (\psi_n(y))_{n\geq0} ( ψ n ( y ) ) n ≥ 0 is coefficientwise Hankel-totally positive in y y y .
We will extend Theorem 1.3 by assigning independent variables to different component colors.
Definition 1.4. Fix m ≥ 1 m \ge1 m ≥ 1 and q = ( q 1 , … , q m ) ∈ [ 0 , ∞ ) m \mathbf{q}=(q_1,\ldots,q_m)\in[0,\infty)^m q = ( q 1 , … , q m ) ∈ [ 0 , ∞ ) m . Assign one color ℓ ∈ { 1 , … , m } \ell\in\{1,\ldots,m\} ℓ ∈ { 1 , … , m } to each component of a functional digraph. Give that component weight y ℓ y_\ell y ℓ and each of its cyclic vertices weight q ℓ q_\ell q ℓ . The weight of a colored digraph is the product of its component and vertex weights. Let Ψ n ( q ; y ) \Psi_n(\mathbf{q};\mathbf{y}) Ψ n ( q ; y ) be the sum of these weights over all digraphs on [ n ] [n] [ n ] and all component colorings, where y = ( y 1 , … , y m ) \mathbf{y}=(y_1,\ldots,y_m) y = ( y 1 , … , y m ) . We call Ψ n ( q ; y ) \Psi_n(\mathbf{q};\mathbf{y}) Ψ n ( q ; y ) the colored functional digraph polynomial .
By the exponential formula, we obtain the generating function
∑ n ≥ 0 Ψ n ( q ; y ) t n n ! = ∏ ℓ = 1 m ( 1 − q ℓ T ( t ) ) − y ℓ = exp ( ∑ ℓ = 1 m y ℓ C q ℓ ( t ) ) . (6) \sum_{n\ge0}\Psi_n(\mathbf{q};\mathbf{y})\frac{t^n}{n!}=\prod_{\ell=1}^{m}(1-q_\ell T(t))^{-y_\ell}=\exp\left(\sum_{\ell=1}^{m}y_\ell C_{q_\ell}(t)\right). \tag*{(6)} n ≥ 0 ∑ Ψ n ( q ; y ) n ! t n = ℓ = 1 ∏ m ( 1 − q ℓ T ( t ) ) − y ℓ = exp ( ℓ = 1 ∑ m y ℓ C q ℓ ( t ) ) . (6)
Taking m = 1 m=1 m = 1 , q 1 = q q_1=q q 1 = q , and y 1 = y y_1=y y 1 = y , we obtain Ψ n ( ( q ) ; ( y ) ) = ψ n ( q , y ) \Psi_n((q);(y))=\psi_n(q,y) Ψ n (( q ) ; ( y )) = ψ n ( q , y ) . We extend Theorem 1.3 to the following joint positivity for the colored model.
Theorem 1.5 (Colored functional-digraph Hankel-total positivity). For every fixed q ∈ [ 0 , ∞ ) m \mathbf{q}\in[0,\infty)^m q ∈ [ 0 , ∞ ) m , ( Ψ n ( q ; y ) ) n ≥ 0 (\Psi_n(\mathbf{q};\mathbf{y}))_{n\ge0} ( Ψ n ( q ; y ) ) n ≥ 0 is coefficientwise Hankel-totally positive jointly in y 1 , … , y m y_1,\ldots,y_m y 1 , … , y m .
Theorems 1.3 and 1.5 follow from the Stieltjes lifting criterion below; their proof is given in § 6 .
A universal lifting criterion for coefficientwise Hankel-total positivity
We will use Stieltjes moment property to present a new universal lifting criterion for coefficientwise Hankel-total positivity. To place the lifting criterion in context, we first recall the role of classical cumulants in probability and statistics. For a real random variable X X X with finite absolute moments of every order, we define the moment and cumulant formal power series by
A X ( t ) = ∑ n ≥ 0 E [ X n ] t n n ! , K X ( t ) = log A X ( t ) = ∑ n ≥ 1 κ n ( X ) t n n ! . (7) A_X(t)=\sum_{n\ge0}\mathbb{E}[X^n]\frac{t^n}{n!},\qquad K_X(t)=\log A_X(t)=\sum_{n\ge1}\kappa_n(X)\frac{t^n}{n!}. \tag*{(7)} A X ( t ) = n ≥ 0 ∑ E [ X n ] n ! t n , K X ( t ) = log A X ( t ) = n ≥ 1 ∑ κ n ( X ) n ! t n . (7)
We use the normalization A X ( 0 ) = 1 A_X(0)=1 A X ( 0 ) = 1 , and require no convergence of either formal power series. In particular, κ 1 ( X ) = E [ X ] \kappa_1(X)=\mathbb{E}[X] κ 1 ( X ) = E [ X ] and κ 2 ( X ) = E [ X 2 ] − ( E [ X ] ) 2 \kappa_2(X)=\mathbb{E}[X^2]-(\mathbb{E}[X])^2 κ 2 ( X ) = E [ X 2 ] − ( E [ X ] ) 2 ; thus the first two cumulants are the mean and variance. For independent random variables X , Y X,Y X , Y with all absolute moments finite, we have K X + Y ( t ) = K X ( t ) + K Y ( t ) K_{X+Y}(t)=K_X(t)+K_Y(t) K X + Y ( t ) = K X ( t ) + K Y ( t ) , as a formal identity, and hence κ n ( X + Y ) = κ n ( X ) + κ n ( Y ) \kappa_n(X+Y)=\kappa_n(X)+\kappa_n(Y) κ n ( X + Y ) = κ n ( X ) + κ n ( Y ) . For the classical theory and its formulation through partition lattices, see [40 ] .
In enumerative combinatorics, we use the same moment–cumulant relation to pass between arbitrary labeled structures and their connected components. By summing products of cumulants over set partitions, we recover the moments and give the independently colored version in (29) . By the exponential formula [41 Corollary 5.1.6] , we may therefore regard the connected-component coefficients c n ( q ) c_n(q) c n ( q ) in (5) as cumulants of the functional-digraph family. In this way, we connect Sokal’s conjecture with the positivity properties of moment–cumulant transformations.
In the classical moment problem, for a real sequence, it is a Stieltjes moment sequence , that is,
a n = ∫ 0 ∞ x n d μ ( x ) ( n ≥ 0 ) (8) a_n = \int_0^\infty x^n\,d\mu(x) \quad(n \ge0) \tag*{(8)} a n = ∫ 0 ∞ x n d μ ( x ) ( n ≥ 0 ) (8)
for a nonnegative measure with all moments finite, where the zero measure is allowed. For Stieltjes moment sequences, one requires support in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) ; for Hamburger moments, one allows support on R \mathbb{R} R . One also studies these questions through the linear functional L ( x n ) = a n \mathcal{L}(x^n)=a_n L ( x n ) = a n : for a Stieltjes representation, one requires L ( p ( x ) 2 ) ≥ 0 \mathcal{L}(p(x)^2)\ge0 L ( p ( x ) 2 ) ≥ 0 and L ( x p ( x ) 2 ) ≥ 0 \mathcal{L}(xp(x)^2)\ge0 L ( x p ( x ) 2 ) ≥ 0 for every real polynomial p ( x ) p(x) p ( x ) . We recall the equivalent matrix criteria in Lemma 2.1; see Schmüdgen [36 Chapter 3] .
In probability, one encounters the shifted cumulant condition through compound-Poisson distributions. Let d , λ ≥ 0 d,\lambda\ge0 d , λ ≥ 0 , let N y N_y N y have Poisson distribution with mean λ y \lambda y λ y , and let J 1 , J 2 , … J_1,J_2,\ldots J 1 , J 2 , … be independent copies of a nonnegative random variable J J J , independent of N y N_y N y , with all moments finite. For y ≥ 0 y\ge0 y ≥ 0 , set X y = d y + ∑ h = 1 N y J h X_y=dy+\sum_{h=1}^{N_y}J_h X y = d y + ∑ h = 1 N y J h , where an empty sum equals zero. By the standard compound-Poisson cumulant formula [22 Chapter 15] , we have
κ n ( X y ) = y ( d 1 { n = 1 } + λ E [ J n ] ) ( n ≥ 1 ) . (9) \kappa_n(X_y)=y\left(d\,1_{\{n=1\}}+\lambda\mathbb{E}[J^n]\right) \quad(n\ge1). \tag*{(9)} κ n ( X y ) = y ( d 1 { n = 1 } + λ E [ J n ] ) ( n ≥ 1 ) . (9)
Here 1 { n = 1 } 1_{\{n=1\}} 1 { n = 1 } equals one for n = 1 n=1 n = 1 and zero otherwise. Writing δ 0 \delta_0 δ 0 for unit mass at zero and P J P_J P J for the distribution of J J J , we obtain κ n + 1 ( X 1 ) = ∫ 0 ∞ x n ν ( d x ) \kappa_{n+1}(X_1)=\int_0^\infty x^n\,\nu(dx) κ n + 1 ( X 1 ) = ∫ 0 ∞ x n ν ( d x ) with the finite measure ν ( d x ) = d δ 0 ( d x ) + λ x P J ( d x ) \nu(dx)=d\,\delta_0(dx)+\lambda xP_J(dx) ν ( d x ) = d δ 0 ( d x ) + λ x P J ( d x ) . Thus we can interpret the once-shifted cumulants as Stieltjes moments, including the deterministic drift as an atom at zero.
More generally, probabilists construct nondecreasing Lévy processes, called subordinators , from Bernstein functions. For a Bernstein function ϕ ( s ) \phi(s) ϕ ( s ) with ϕ ( 0 ) = 0 \phi(0)=0 ϕ ( 0 ) = 0 , one obtains a subordinator ( S y ) y ≥ 0 (S_y)_{y\ge0} ( S y ) y ≥ 0 with
E [ e − s S y ] = e − y ϕ ( s ) ( s , y ≥ 0 ) . (10) \mathbb{E}[e^{-sS_y}]=e^{-y\phi(s)} \quad(s,y\ge0). \tag*{(10)} E [ e − s S y ] = e − y ϕ ( s ) ( s , y ≥ 0 ) . (10)
Here one calls ϕ ( s ) \phi(s) ϕ ( s ) a Bernstein function when it is nonnegative, smooth for s > 0 s>0 s > 0 , and satisfies ( − 1 ) r − 1 ϕ ( r ) ( s ) ≥ 0 (-1)^{r-1}\phi^{(r)}(s)\ge0 ( − 1 ) r − 1 ϕ ( r ) ( s ) ≥ 0 for r ≥ 1 r\ge1 r ≥ 1 . For the Lévy representation and the convolution-semigroup correspondence, see Schilling, Song and Vondraček [35 Theorems 3.2 and 5.2] . We recall the corresponding cumulant representation, under the additional requirement of finite moments of every order, in Lemma 2.7. Through this representation, we also explain the shift by one in the lifting criterion.
For fixed nonnegative indeterminates, one can use independent sums to combine such moment families. In this paper, we retain the intensities as independent indeterminates and prove nonnegativity of every coefficient in every Hankel minor. We also allow an arbitrary initial Stieltjes moment sequence, without normalization of its zeroth moment. In Theorem 1.7 , we characterize this universal coefficientwise preservation by the shifted cumulant condition. We work with formal power series throughout; in particular, we do not require an exponential moment or uniqueness of a representing measure.
Let K ( t ) = ∑ n ≥ 1 κ n t n / n ! K(t)=\sum_{n\geq1}\kappa_n t^n/n! K ( t ) = ∑ n ≥ 1 κ n t n / n ! be any real formal power series, and write e y K ( t ) = ∑ n ≥ 0 B n ( y ) t n / n ! e^{yK(t)}=\sum_{n\geq0}B_n(y)t^n/n! e y K ( t ) = ∑ n ≥ 0 B n ( y ) t n / n ! . To formulate the universal lifting property, we state the following equivalence.
Theorem 1.6 (One-parameter lifting criterion). The following statements are equivalent:
(i) The shifted cumulant sequence ( κ n + 1 ) n ≥ 0 (\kappa_{n+1})_{n\geq0} ( κ n + 1 ) n ≥ 0 is a Stieltjes moment sequence.
(ii) The polynomial sequence ( B n ( y ) ) n ≥ 0 (B_n(y))_{n\geq0} ( B n ( y ) ) n ≥ 0 is coefficientwise Hankel-totally positive in y y y .
(iii) For every real Stieltjes moment sequence a = ( a n ) n ≥ 0 \mathbf{a}=(a_n)_{n\geq0} a = ( a n ) n ≥ 0 , the sequence
F n ( y ) = ∑ k = 0 n ( n k ) a n − k B k ( y ) ( n ≥ 0 ) (11) F_n(y)=\sum_{k=0}^{n}\binom{n}{k}a_{n-k}B_k(y)\qquad(n\geq0) \tag*{(11)} F n ( y ) = k = 0 ∑ n ( k n ) a n − k B k ( y ) ( n ≥ 0 ) (11) is coefficientwise Hankel-totally positive in y y y .
The last assertion is the universal lifting property : arbitrary initial Stieltjes moments may be convolved with the component family. Closure of real Stieltjes moment sequences under binomial convolution is classical (see [33 ] , [42 Corollary 4.3] for instance). The lifting assertion extends it to coefficientwise Hankel-total positivity. We obtain this theorem from the multivariate result below with m = 1 m=1 m = 1 .
We now retain independent intensities for different formal power series. Fix m ≥ 1 m\geq1 m ≥ 1 . For ℓ = 1 , … , m \ell=1,\ldots,m ℓ = 1 , … , m , let κ 1 ( ℓ ) , κ 2 ( ℓ ) , … \kappa^{(\ell)}_1,\kappa^{(\ell)}_2,\ldots κ 1 ( ℓ ) , κ 2 ( ℓ ) , … be real numbers, and define
∑ n ≥ 0 M n ( y ) t n n ! = exp ( ∑ ℓ = 1 m y ℓ K ℓ ( t ) ) , K ℓ ( t ) = ∑ n ≥ 1 κ n ( ℓ ) t n n ! . (12) \sum_{n\geq0}M_n(\mathbf{y})\frac{t^n}{n!}=\exp\left(\sum_{\ell=1}^{m}y_\ell K_\ell(t)\right),\qquad K_\ell(t)=\sum_{n\geq1}\kappa_n^{(\ell)}\frac{t^n}{n!}. \tag*{(12)} n ≥ 0 ∑ M n ( y ) n ! t n = exp ( ℓ = 1 ∑ m y ℓ K ℓ ( t ) ) , K ℓ ( t ) = n ≥ 1 ∑ κ n ( ℓ ) n ! t n . (12)
For a real sequence a = ( a n ) n ≥ 0 \mathbf{a}=(a_n)_{n\geq0} a = ( a n ) n ≥ 0 , define
F n ( y ) : = ( T y a ) n = ∑ k = 0 n ( n k ) a n − k M k ( y ) . (13) F_n(\mathbf{y}):=(T_{\mathbf{y}}\mathbf{a})_n=\sum_{k=0}^{n}\binom{n}{k}a_{n-k}M_k(\mathbf{y}). \tag*{(13)} F n ( y ) := ( T y a ) n = k = 0 ∑ n ( k n ) a n − k M k ( y ) . (13)
If A ( t ) = ∑ n ≥ 0 a n t n / n ! A(t)=\sum_{n\geq0}a_n t^n/n! A ( t ) = ∑ n ≥ 0 a n t n / n ! , then (13) is equivalent to the formal identity
∑ n ≥ 0 F n ( y ) t n n ! = A ( t ) exp ( ∑ ℓ = 1 m y ℓ K ℓ ( t ) ) . (14) \sum_{n\geq0}F_n(\mathbf{y})\frac{t^n}{n!}=A(t)\exp\left(\sum_{\ell=1}^{m}y_\ell K_\ell(t)\right). \tag*{(14)} n ≥ 0 ∑ F n ( y ) n ! t n = A ( t ) exp ( ℓ = 1 ∑ m y ℓ K ℓ ( t ) ) . (14)
In particular, F 0 ( y ) = a 0 F_0(\mathbf{y})=a_0 F 0 ( y ) = a 0 .
We characterize the admissible cumulants as follows, which for m = 1 m=1 m = 1 implies Theorem 1.6 .
Theorem 1.7 (Universal multivariate Stieltjes–cumulant equivalence). For the polynomial sequence ( M n ( y ) ) n ≥ 0 (M_n(\mathbf{y}))_{n\geq0} ( M n ( y ) ) n ≥ 0 defined by (1.9) , the following conditions are equivalent:
(i) For each ℓ = 1 , … , m \ell=1,\ldots,m ℓ = 1 , … , m , the shifted sequence ( κ n + 1 ( ℓ ) ) n ≥ 0 (\kappa_{n+1}^{(\ell)})_{n\geq0} ( κ n + 1 ( ℓ ) ) n ≥ 0 is a Stieltjes moment sequence.
(ii) ( M n ( y ) ) n ≥ 0 (M_n(\mathbf{y}))_{n\geq0} ( M n ( y ) ) n ≥ 0 is coefficientwise Hankel-totally positive jointly in y 1 , … , y m y_1,\ldots,y_m y 1 , … , y m .
(iii) For every fixed y ∈ [ 0 , ∞ ) m \mathbf{y}\in[0,\infty)^m y ∈ [ 0 , ∞ ) m , ( M n ( y ) ) n ≥ 0 (M_n(\mathbf{y}))_{n\geq0} ( M n ( y ) ) n ≥ 0 is a Stieltjes moment sequence.
(iv) For every real Stieltjes moment sequence a \mathbf{a} a , ( F n ( y ) ) n ≥ 0 (F_n(\mathbf{y}))_{n\geq0} ( F n ( y ) ) n ≥ 0 in (13) is coefficientwise Hankel-totally positive jointly in y 1 , … , y m y_1,\ldots,y_m y 1 , … , y m .
(v) For every real Stieltjes moment sequence a \mathbf{a} a and every fixed y ∈ [ 0 , ∞ ) m \mathbf{y}\in[0,\infty)^m y ∈ [ 0 , ∞ ) m , ( F n ( y ) ) n ≥ 0 (F_n(\mathbf{y}))_{n\geq0} ( F n ( y ) ) n ≥ 0 is a Stieltjes moment sequence.*
Under any of these conditions, we have κ n ( ℓ ) ≥ 0 \kappa_n^{(\ell)}\geq0 κ n ( ℓ ) ≥ 0 for all n , ℓ n,\ell n , ℓ , so we need no separate sign assumption on the cumulants. In condition (iv), we allow every initial real Stieltjes moment sequence and assert coefficientwise positivity in y \mathbf{y} y .
We will prove Theorem 1.7 in § 3 . For sufficiency of Theorem 1.7 , the proof combines the Poisson square-integral identity of Last and Penrose [21 ] , Schur alternants, and a partitioned Mecke expansion. The resulting kernels are independent of the indeterminates. Necessity follows by extracting the lowest-degree terms of shifted Hankel minors.
A Hamburger extension on the real line
A real sequence ( a n ) n ≥ 0 (a_n)_{n\geq0} ( a n ) n ≥ 0 is a Hamburger moment sequence if
a n = ∫ − ∞ ∞ x n μ ( d x ) a_n=\int_{-\infty}^{\infty}x^n\,\mu(\mathrm{d}x) a n = ∫ − ∞ ∞ x n μ ( d x )
for a nonnegative measure μ \mu μ on R \mathbb{R} R with all absolute moments finite. The zero measure is allowed. The Hamburger moment theorem says that this is equivalent to positive semidefiniteness of every finite Hankel matrix [ a i + j ] i , j = 0 r [a_{i+j}]_{i,j=0}^{r} [ a i + j ] i , j = 0 r ; see [36 Chapter 3] .
For arbitrary real support, the universal coefficientwise criterion, stated in the following as in Theorem 1.7 , involves twice-shifted cumulants and leading principal determinants. In addition, Remark 3 explains why its coefficientwise assertion does not extend to arbitrary principal minors.
Theorem 1.8 (Universal Hamburger lifting). Let M n M_n M n and F n F_n F n be defined by (12) and (13) with real cumulants and an arbitrary initial real sequence a = ( a n ) n ≥ 0 \mathbf{a}=(a_n)_{n\geq0} a = ( a n ) n ≥ 0 , and put
H r a ( y ) = det [ F i + j ( y ) ] i , j = 0 r − 1 . (15) H_r^{\mathbf{a}}(\mathbf{y})=\det\left[F_{i+j}(\mathbf{y})\right]_{i,j=0}^{r-1}. \tag*{(15)} H r a ( y ) = det [ F i + j ( y ) ] i , j = 0 r − 1 . (15)
The following conditions are equivalent:
(i) For every ℓ \ell ℓ , the twice-shifted sequence ( κ n + 2 ( ℓ ) ) n ≥ 0 (\kappa_{n+2}^{(\ell)})_{n\ge0} ( κ n + 2 ( ℓ ) ) n ≥ 0 is a Hamburger moment sequence.
(ii) For every initial Hamburger moment sequence and r ≥ 1 r \ge1 r ≥ 1 , H r a ( y ) ∈ R ≥ 0 [ y ] H_r^{a}(y) \in\mathbb{R}_{\ge0}[y] H r a ( y ) ∈ R ≥ 0 [ y ] .
(iii) For every initial Hamburger moment sequence and every y ∈ [ 0 , ∞ ) m \mathbf{y} \in[0,\infty)^m y ∈ [ 0 , ∞ ) m , ( F n ( y ) ) n ≥ 0 (F_n(\mathbf{y}))_{n\ge0} ( F n ( y ) ) n ≥ 0 is a Hamburger moment sequence.
(iv) For every y ∈ [ 0 , ∞ ) m \mathbf{y} \in[0,\infty)^m y ∈ [ 0 , ∞ ) m , ( M n ( y ) ) n ≥ 0 (M_n(\mathbf{y}))_{n\ge0} ( M n ( y ) ) n ≥ 0 is a Hamburger moment sequence.
Under these conditions, fix an initial Hamburger moment sequence and let D = ( r 2 ) D = \binom{r}{2} D = ( 2 r ) and v ( y ) = ∑ ℓ y ℓ κ 2 ( ℓ ) v(\mathbf{y}) = \sum_{\ell} y_{\ell}\kappa_{2}^{(\ell)} v ( y ) = ∑ ℓ y ℓ κ 2 ( ℓ ) . Then
deg y H r a ≤ D , [ H r a ] D = a 0 r ( ∏ j = 0 r − 1 j ! ) v ( y ) D , (16) \deg_{\mathbf{y}} H_r^a \le D,\qquad[H_r^a]_D = a_0^r\left(\prod_{j=0}^{r-1}j!\right)v(\mathbf{y})^D, \tag*{(16)} deg y H r a ≤ D , [ H r a ] D = a 0 r ( j = 0 ∏ r − 1 j ! ) v ( y ) D , (16)
where [ P ] D [P]_D [ P ] D denotes the homogeneous part of total degree D D D . The determinant is independent of all first cumulants κ 1 ( ℓ ) \kappa_1^{(\ell)} κ 1 ( ℓ ) .
Using the common moment and Poisson tools from § 3 , we will prove Theorem 1.8 in § 4 . To clarify the scope of Theorem 1.8 , we record four remarks on drift invariance, Gaussian components, principal minors, and composition.
Remark 1 (Invariance under common polynomial drifts). The fact that these determinants do not depend on the first cumulants has an algebraic explanation. If any polynomial c ( y ) ∈ R [ y ] c(\mathbf{y}) \in\mathbb{R}[\mathbf{y}] c ( y ) ∈ R [ y ] is used to replace F n F_n F n by its common binomial translate
F ~ n ( y ) = ∑ k = 0 n ( n k ) c ( y ) n − k F k ( y ) , (17) \widetilde{F}_n(\mathbf{y})=\sum_{k=0}^{n}\binom{n}{k}c(\mathbf{y})^{n-k}F_k(\mathbf{y}), \tag*{(17)} F n ( y ) = k = 0 ∑ n ( k n ) c ( y ) n − k F k ( y ) , (17)
then H ~ r a : = det [ F ~ i + j ] i , j = 0 r − 1 = H r a \widetilde{H}_r^a := \det[\widetilde{F}_{i+j}]_{i,j=0}^{r-1} = H_r^a H r a := det [ F i + j ] i , j = 0 r − 1 = H r a . This is the parameter form of Layman’s binomial invariance [23 Theorem 1] . Indeed, writing H r = [ F i + j ] i , j = 0 r − 1 H_r=[F_{i+j}]_{i,j=0}^{r-1} H r = [ F i + j ] i , j = 0 r − 1 , one has H ~ r = P r H r P r T \widetilde{H}_r=P_rH_rP_r^{\mathsf{T}} H r = P r H r P r T , where P r P_r P r is the r × r r\times r r × r lower triangular matrix with ( P r ) i , j = ( i j ) c ( y ) i − j (P_r)_{i,j}=\binom{i}{j}c(\mathbf{y})^{i-j} ( P r ) i , j = ( j i ) c ( y ) i − j for 0 ≤ j ≤ i < r 0\le j\le i<r 0 ≤ j ≤ i < r and zero entries for j > i j>i j > i . Its diagonal entries are one. Hence the determinant assertion also survives arbitrary common polynomial drifts, including drifts with negative coefficients.
Remark 2 (Gaussian cumulants and moment determinants). For the pure exponential family with
K ℓ ( t ) = b ℓ t + σ ℓ 2 t 2 2 , b ℓ ∈ R , σ ℓ 2 ≥ 0 , (18) K_{\ell}(t)=b_{\ell}t+\frac{\sigma_{\ell}^{2}t^{2}}{2},\qquad b_{\ell}\in\mathbb{R},\quad\sigma_{\ell}^{2}\ge0, \tag*{(18)} K ℓ ( t ) = b ℓ t + 2 σ ℓ 2 t 2 , b ℓ ∈ R , σ ℓ 2 ≥ 0 , (18)
the representing measure in (i) is σ ℓ 2 δ 0 \sigma_\ell^2\delta_0 σ ℓ 2 δ 0 . Take the initial measure δ 0 \delta_0 δ 0 . By Gaussian scaling and (16) , we obtain the exact formula
det [ M i + j ( y ) ] i , j = 0 r − 1 = ( ∏ j = 0 r − 1 j ! ) ( ∑ ℓ σ ℓ 2 y ℓ ) ( r 2 ) . (19) \det\left[M_{i+j}(y)\right]_{i,j=0}^{r-1}=\left(\prod_{j=0}^{r-1}j!\right)\left(\sum_{\ell}\sigma_\ell^2 y_\ell\right)^{\binom{r}{2}}. \tag*{(19)} det [ M i + j ( y ) ] i , j = 0 r − 1 = ( j = 0 ∏ r − 1 j ! ) ( ℓ ∑ σ ℓ 2 y ℓ ) ( 2 r ) . (19)
This is also the classical Gaussian moment determinant, obtained from the monic Hermite squared norms; see [29 Table 18.3.1 and Section 18.2 (ix)] . This family lies beyond the Stieltjes theorem: when the mean is zero and the variance is positive, the once-shifted cumulant sequence starts with 0 , σ 2 0,\sigma^2 0 , σ 2 , which cannot be Stieltjes.
Notation and concepts
We use the following notation for coefficient extraction and positivity. For a formal power series f ( x ) = ∑ j ≥ 0 f j x j f(x)=\sum_{j\ge0}f_jx^j f ( x ) = ∑ j ≥ 0 f j x j , we write
[ x k ] f ( x ) = f k ( k ≥ 0 ) . (22) [x^k]f(x)=f_k\qquad(k\ge0). \tag*{(22)} [ x k ] f ( x ) = f k ( k ≥ 0 ) . (22)
Thus, for an exponential generating function F ( t ) = ∑ n ≥ 0 a n t n / n ! F(t)=\sum_{n\ge0}a_nt^n/n! F ( t ) = ∑ n ≥ 0 a n t n / n ! , we have a n = n ! [ t n ] F ( t ) a_n=n![t^n]F(t) a n = n ! [ t n ] F ( t ) . For several variables, we use [ z α ] [z^\alpha] [ z α ] to extract the coefficient of z 1 α 1 ⋯ z d α d z_1^{\alpha_1}\cdots z_d^{\alpha_d} z 1 α 1 ⋯ z d α d . We write R , Q , Z \mathbb{R},\mathbb{Q},\mathbb{Z} R , Q , Z for the real numbers, rational numbers and integers, respectively, and put R ≥ 0 = [ 0 , ∞ ) \mathbb{R}_{\ge0}=[0,\infty) R ≥ 0 = [ 0 , ∞ ) and N 0 = N 0 = { 0 , 1 , 2 , … } \mathbb{N}_0=\mathbb{N}_0=\{0,1,2,\ldots\} N 0 = N 0 = { 0 , 1 , 2 , … } . For a multi-index α = ( α 1 , … , α d ) ∈ N 0 d \boldsymbol{\alpha}=(\alpha_1,\ldots,\alpha_d)\in\mathbb{N}_0^d α = ( α 1 , … , α d ) ∈ N 0 d , we use
∣ α ∣ = ∑ j = 1 d α j , α ! = ∏ j = 1 d α j ! , z α = ∏ j = 1 d z j α j . (23) |\boldsymbol{\alpha}|=\sum_{j=1}^{d}\alpha_j,\qquad\boldsymbol{\alpha}!=\prod_{j=1}^{d}\alpha_j!,\qquad z^{\boldsymbol{\alpha}}=\prod_{j=1}^{d}z_j^{\alpha_j}. \tag*{(23)} ∣ α ∣ = j = 1 ∑ d α j , α ! = j = 1 ∏ d α j ! , z α = j = 1 ∏ d z j α j . (23)
We write R [ z ] \mathbb{R}[z] R [ z ] for the polynomial ring and R [ [ z ] ] \mathbb{R}[[z]] R [[ z ]] for the ring of formal power series, and use
R ≥ 0 [ z ] = { ∑ α c α z α : c α ≥ 0 , only finitely many c α ≠ 0 } . (24) \mathbb{R}_{\ge0}[z]=\left\{\sum_{\boldsymbol{\alpha}}c_{\boldsymbol{\alpha}}z^{\boldsymbol{\alpha}}:c_{\boldsymbol{\alpha}}\ge0,\ \text{only finitely many }c_{\boldsymbol{\alpha}}\ne0\right\}. \tag*{(24)} R ≥ 0 [ z ] = { α ∑ c α z α : c α ≥ 0 , only finitely many c α = 0 } . (24)
For polynomials or formal power series P , Q P,Q P , Q in the same indeterminates, we write P ⪰ c o e f f Q P\succeq_{\mathrm{coeff}}Q P ⪰ coeff Q when every coefficient of P − Q P-Q P − Q is nonnegative. In particular, P ⪰ c o e f f 0 P\succeq_{\mathrm{coeff}}0 P ⪰ coeff 0 for a polynomial means membership in (24) . When we assert coefficientwise positivity, we require this condition for every minor under consideration.
Throughout, a matrix is totally positive (TP) if its every minor is nonnegative and coefficientwise totally positive if its every minor is a nonnegative coefficient polynomial. For a sequence of polynomials a n ( z ) a_n(z) a n ( z ) , it is called coefficientwise Hankel-totally positive if the Hankel matrix [ a i + j ( z ) ] i , j ≥ 0 [a_{i+j}(z)]_{i,j\ge0} [ a i + j ( z ) ] i , j ≥ 0 is coefficientwise totally positive. In other words, for every r ≥ 1 r\ge1 r ≥ 1 and strictly increasing lists I = ( i 1 < ⋯ < i r ) I=(i_1<\cdots<i_r) I = ( i 1 < ⋯ < i r ) , J = ( j 1 < ⋯ < j r ) J=(j_1<\cdots<j_r) J = ( j 1 < ⋯ < j r ) ,
det [ a i a + j b ( z ) ] a , b = 1 r ∈ R ≥ 0 [ z ] . (25) \det\left[a_{i_a+j_b}(z)\right]_{a,b=1}^{r}\in\mathbb{R}_{\ge0}[z]. \tag*{(25)} det [ a i a + j b ( z ) ] a , b = 1 r ∈ R ≥ 0 [ z ] . (25)
For a real sequence, Hankel total positivity is equivalent to being a Stieltjes moment sequence. Coefficientwise positivity implies positivity at every nonnegative parameter value z z z . For a general polynomial, the converse fails: ( u − 1 ) 2 ≥ 0 (u-1)^2\ge0 ( u − 1 ) 2 ≥ 0 for every u ≥ 0 u\ge0 u ≥ 0 , although its coefficient of u u u is negative. The distinction is therefore essential when we pass from numerical moment sequences to combinatorial polynomials.
We write S n S_n S n for the permutations of [ 1 , n ] [1,n] [ 1 , n ] , with sgn ( σ ) ∈ { − 1 , 1 } \operatorname{sgn}(\sigma)\in\{-1,1\} sgn ( σ ) ∈ { − 1 , 1 } their signs. For a set A A A , we use 1 A \mathbf{1}_A 1 A for its indicator, and δ x \delta_x δ x for the unit point mass at x x x . In a finite-dimensional coordinate space, we write 1 = ( 1 , … , 1 ) \mathbf{1}=(1,\ldots,1) 1 = ( 1 , … , 1 ) and use e j \mathbf{e}_j e j for the vector whose j j j -th entry is one and whose other entries are zero.
For a matrix M M M and increasing index lists I = ( i 1 , … , i r ) I=(i_1,\ldots,i_r) I = ( i 1 , … , i r ) , J = ( j 1 , … , j r ) J=(j_1,\ldots,j_r) J = ( j 1 , … , j r ) , let
M [ I , J ] : = [ M i a , j b ] a , b = 1 r . (26) M[I,J]:=[M_{i_a,j_b}]_{a,b=1}^{r}. \tag*{(26)} M [ I , J ] := [ M i a , j b ] a , b = 1 r . (26)
We use M T M^{T} M T for the transpose, I I I for an identity matrix of the required size, and E i , j E_{i,j} E i , j for the matrix with its only nonzero entry, equal to one, in position ( i , j ) (i,j) ( i , j ) . We use diag \operatorname{diag} diag for a diagonal or block-diagonal matrix and adj ( M ) \operatorname{adj}(M) adj ( M ) for the classical adjugate, the transpose of the cofactor matrix. Unless a finite range is specified, we index infinite matrices by N 0 \mathbb{N}_0 N 0 . Empty products and determinants of size zero have value one.
For a ∈ R a \in\mathbb{R} a ∈ R and n ∈ N 0 n \in\mathbb{N}_0 n ∈ N 0 , our factorial conventions are
( a ) n : = ∏ h = 0 n − 1 ( a + h ) , a n : = ∏ h = 0 n − 1 ( a − h ) , ( a n ) : = a n n ! . (27) (a)_n := \prod_{h=0}^{n-1}(a+h), \qquad a^n := \prod_{h=0}^{n-1}(a-h), \qquad\binom{a}{n} := \frac{a^n}{n!}. \tag*{(27)} ( a ) n := h = 0 ∏ n − 1 ( a + h ) , a n := h = 0 ∏ n − 1 ( a − h ) , ( n a ) := n ! a n . (27)
For a positive measure μ \mu μ , we use interval limits for one-dimensional integrals. Our endpoint convention is
∫ a b f ( x ) μ ( d x ) : = ∫ [ a , b ] f ( x ) μ ( d x ) , ∫ a + b f ( x ) μ ( d x ) : = ∫ ( a , b ] f ( x ) μ ( d x ) . (28) \int_a^b f(x)\,\mu(\mathrm{d}x) := \int_{[a,b]} f(x)\,\mu(\mathrm{d}x), \qquad\int_{a+}^b f(x)\,\mu(\mathrm{d}x) := \int_{(a,b]} f(x)\,\mu(\mathrm{d}x). \tag*{(28)} ∫ a b f ( x ) μ ( d x ) := ∫ [ a , b ] f ( x ) μ ( d x ) , ∫ a + b f ( x ) μ ( d x ) := ∫ ( a , b ] f ( x ) μ ( d x ) . (28)
At an infinite endpoint we integrate over the corresponding half-line. Thus ∫ 0 ∞ f ( x ) μ ( d x ) \int_0^\infty f(x)\,\mu(\mathrm{d}x) ∫ 0 ∞ f ( x ) μ ( d x ) includes a possible atom at zero. We extend measures originally defined on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) or R ∖ { 0 } \mathbb{R} \setminus\{0\} R ∖ { 0 } by assigning zero mass to the omitted points. We retain a set below the integral sign for a multidimensional domain or a general measurable space.
Organization of the paper
In § 3 , we collect the common moment and Poisson tools and prove the Stieltjes lifting criterion in Theorem 1.7 . In § 4 , we prove the Hamburger lifting criterion by using the real-line Vandermonde expansion. In § 5 , we establish binomial-convolution, composition, and substitution results, discuss logarithms of ordinary Stieltjes series, and collect moment inputs for the applications.
In § 6 , we apply the lifting criteria to functional digraphs and their extensions. In Section 6, we derive general results for implicit equations and autonomous differential equations from Pólya-frequency series. In § 8 , we use known moment representations for tree and forest sequences to verify the lifting hypotheses and obtain families that are coefficientwise Hankel-totally positive in the lifting variables. We treat the Ward sequence and several related tree families in this way.
In § 9 , we obtain further general families from polynomial moment inputs, including extensions of the Bessel, Lah, and Laguerre polynomials and a fixed-shape result for the Schläfli-Gessel-Seo polynomials. We also consider set partitions, lognormal moments, Poisson differences, and matrix consequences. For each application, we distinguish the fixed parameters used to verify the moment condition from the variables in which we prove coefficientwise positivity. We not only give new proofs of coefficientwise Hankel-total positivity of many combinatorial polynomials in a unified approach, but also prove coefficientwise Hankel-total positivity for the component polynomials of functional digraphs confirming a conjecture of Sokal and establish fixed-shape Schläfli-Gessel-Seo polynomials positivity supporting another coefficientwise Hankel-total positivity conjecture of Sokal. In § 10 , we discuss the remaining questions about positivity jointly in the structural parameters.
Proof of the Stieltjes lifting theorem
Proof of the Stieltjes lifting theorem
In this section, we will prove Theorem 1.7 . We first collect the moment criteria, alternant identities, and Poisson formulas used. For the Stieltjes sufficiency of Theorem 1.7 , we construct square-integral expansions with variable-independent kernels and integrate them against an arbitrary initial moment measure. The lowest-degree terms of shifted Hankel minors then give necessity.
By the classical exponential formula of Stanley [41 Corollary 5.1.6] , we may express each moment polynomial as a finite sum over set partitions. Let P 0 = { ∅ } \mathcal{P}_0=\{\varnothing\} P 0 = { ∅ } and P n \mathcal{P}_n P n denote the set of partitions of [ n ] [n] [ n ] . With block weight ∑ ℓ y ℓ κ ∣ B ∣ ( ℓ ) \sum_{\ell}y_\ell\kappa^{(\ell)}_{|B|} ∑ ℓ y ℓ κ ∣ B ∣ ( ℓ ) , we obtain
M n ( y ) = ∑ π ∈ P n ∏ B ∈ π ( ∑ ℓ = 1 m y ℓ κ ∣ B ∣ ( ℓ ) ) . (29) M_n(\mathbf{y})=\sum_{\pi\in\mathcal{P}_n}\prod_{B\in\pi}\left(\sum_{\ell=1}^{m}y_\ell\kappa^{(\ell)}_{|B|}\right). \tag*{(29)} M n ( y ) = π ∈ P n ∑ B ∈ π ∏ ( ℓ = 1 ∑ m y ℓ κ ∣ B ∣ ( ℓ ) ) . (29)
Thus M 0 = 1 M_0=1 M 0 = 1 and deg y M n ≤ n \deg_{\mathbf{y}}M_n\le n deg y M n ≤ n .
Common tools
We begin with the classical Hamburger and Stieltjes moment criteria (see Schmüdgen [36 Chapter 3] , Pinkus [32 ] and Karlin [18 ] for instance).
Lemma 2.1 (Classical moment criteria). Let ( a n ) n ≥ 0 (a_n)_{n\geq0} ( a n ) n ≥ 0 be a real sequence, and, for h ∈ { 0 , 1 } h\in\{0,1\} h ∈ { 0 , 1 } , define
H d ( h ) ( a ) : = [ a i + j + h ] i , j = 0 d ( d ≥ 0 ) . (30) H_d^{(h)}(\mathbf{a}):=[a_{i+j+h}]_{i,j=0}^{d}\qquad(d\geq0). \tag*{(30)} H d ( h ) ( a ) := [ a i + j + h ] i , j = 0 d ( d ≥ 0 ) . (30) (i) The sequence is a Hamburger moment sequence if and only if H d ( 0 ) ( a ) H_d^{(0)}(\mathbf{a}) H d ( 0 ) ( a ) is positive semidefinite for every d ≥ 0 d\geq0 d ≥ 0 .
(ii) The sequence is a Stieltjes moment sequence if and only if both H d ( 0 ) ( a ) H_d^{(0)}(\mathbf{a}) H d ( 0 ) ( a ) and H d ( 1 ) ( a ) H_d^{(1)}(\mathbf{a}) H d ( 1 ) ( a ) are positive semidefinite for every d ≥ 0 d\geq0 d ≥ 0 . Equivalently, its infinite Hankel matrix [ a i + j ] i , j ≥ 0 [a_{i+j}]_{i,j\geq0} [ a i + j ] i , j ≥ 0 is totally positive.
(iii) If a representing Stieltjes measure has infinitely many positive support points, then every Hankel minor with strictly increasing row and column index lists is strictly positive.
To convert a determinant of integrals into an integral of two determinants, we use Andréief’s integration identity [1 ] . We integrate over the full Cartesian product, so we include the normalizing factor 1 / r ! 1/r! 1/ r ! .
Lemma 2.2 (Andréief’s integration identity). Let μ \mu μ be a positive measure on a measurable space X X X , and let f 1 ( t ) , … , f r ( t ) f_1(t),\ldots,f_r(t) f 1 ( t ) , … , f r ( t ) and g 1 ( t ) , … , g r ( t ) g_1(t),\ldots,g_r(t) g 1 ( t ) , … , g r ( t ) be measurable functions on X X X such that ∫ X ∣ f i ( t ) g j ( t ) ∣ d μ ( t ) < ∞ \int_X |f_i(t)g_j(t)|\,\mathrm{d}\mu(t)<\infty ∫ X ∣ f i ( t ) g j ( t ) ∣ d μ ( t ) < ∞ for all i , j i,j i , j . Then we have
det [ ∫ X f i ( t ) g j ( t ) d μ ( t ) ] i , j = 1 r = 1 r ! ∫ X r det [ f i ( t j ) ] i , j = 1 r det [ g i ( t j ) ] i , j = 1 r ∏ j = 1 r d μ ( t j ) . (31) \det\left[\int_X f_i(t)g_j(t)\,\mathrm{d}\mu(t)\right]_{i,j=1}^{r}
=\frac{1}{r!}\int_{X^r}\det[f_i(t_j)]_{i,j=1}^{r}\det[g_i(t_j)]_{i,j=1}^{r}\prod_{j=1}^{r}\mathrm{d}\mu(t_j). \tag*{(31)} det [ ∫ X f i ( t ) g j ( t ) d μ ( t ) ] i , j = 1 r = r ! 1 ∫ X r det [ f i ( t j ) ] i , j = 1 r det [ g i ( t j ) ] i , j = 1 r j = 1 ∏ r d μ ( t j ) . (31)
We next recall the notation for Schur polynomials. For a partition λ = ( λ 1 ≥ ⋯ ≥ λ r ≥ 0 ) \lambda=(\lambda_1\ge\cdots\ge\lambda_r\ge0) λ = ( λ 1 ≥ ⋯ ≥ λ r ≥ 0 ) , we draw λ i \lambda_i λ i boxes in row i i i . Let SSYT r ( λ ) \operatorname{SSYT}_r(\lambda) SSYT r ( λ ) be the set of fillings of these boxes by 1 , … , r 1,\ldots,r 1 , … , r that are weakly increasing along rows and strictly increasing down columns. Let m j ( P ) m_j(P) m j ( P ) be the number of entries equal to j j j in a filling P P P , and we define
s λ ( t 1 , … , t r ) = ∑ P ∈ SSYT r ( λ ) ∏ j = 1 r t j m j ( P ) . (32) s_\lambda(t_1,\ldots,t_r)=\sum_{P\in\operatorname{SSYT}_r(\lambda)}\prod_{j=1}^{r}t_j^{m_j(P)}. \tag*{(32)} s λ ( t 1 , … , t r ) = P ∈ SSYT r ( λ ) ∑ j = 1 ∏ r t j m j ( P ) . (32)
For the empty partition, we set s ∅ = 1 s_{\varnothing}=1 s ∅ = 1 . From (32) , we have nonnegative integer coefficients. To factor the generalized Vandermonde determinants, we use the alternant formula in Macdonald [25 ] , Chapter I, Sections 3 and 5, or Stanley [41 ] , Sections 7.10 and 7.15.
Lemma 2.3 (Schur alternants and coefficient positivity). For integers 0 ≤ n 1 < ⋯ < n r 0\le n_1<\cdots<n_r 0 ≤ n 1 < ⋯ < n r , set λ i = n r + 1 − i − ( r − i ) \lambda_i=n_{r+1-i}-(r-i) λ i = n r + 1 − i − ( r − i ) with 1 ≤ i ≤ r 1\le i\le r 1 ≤ i ≤ r . Then we have
det [ t j n i ] i , j = 1 r = Δ ( t ) s λ ( t ) , Δ ( t ) = ∏ 1 ≤ i < j ≤ r ( t j − t i ) . (33) \det[t_j^{n_i}]_{i,j=1}^{r}=\Delta(t)s_\lambda(t),\qquad\Delta(t)=\prod_{1\le i<j\le r}(t_j-t_i). \tag*{(33)} det [ t j n i ] i , j = 1 r = Δ ( t ) s λ ( t ) , Δ ( t ) = 1 ≤ i < j ≤ r ∏ ( t j − t i ) . (33) In particular, the determinant in (33) is positive when 0 < t 1 < ⋯ < t r 0<t_1<\cdots<t_r 0 < t 1 < ⋯ < t r .
For the subsequent matrix factorizations, we record three standard closure rules. We cite the finite matrix results in the corresponding parts.
Lemma 2.4 (Classical matrix closure rules). The following hold.
(i) (Finite products.) The product of coefficientwise totally positive matrices is coefficientwise totally positive whenever the Cauchy–Binet sums for its minors are finite [32 Proposition 1.4] .
(ii) (Bidiagonal matrices.) An upper or lower bidiagonal matrix with coefficientwise nonnegative entries is coefficientwise totally positive [31 Lemma 9.1] .
Stieltjes representations and compound-Poisson moments
We will use the Poisson exponential formula [22 Theorem 3.9] and the Fock isometry [21 Theorem 1.1] . We allow atomic intensity measures. For the exponential formula, we use bounded integrands. For the Fock isometry, we use polynomial functions of finite Poisson counts with bounded jumps; consequently, all required polynomial moments and second moments are finite.
Lemma 2.5 (Poisson exponential formula [22 ] and Fock isometry [21 ] ). Let η \eta η be a Poisson random measure on a measurable space E E E , with finite intensity measure Λ \Lambda Λ .
(i) For a bounded measurable function f : E → R f:E\to\mathbb{R} f : E → R ,
E exp ( ∫ E f ( z ) η ( d z ) ) = exp ( ∫ E ( e f ( z ) − 1 ) Λ ( d z ) ) . (34) \mathbb{E}\exp\left(\int_E f(z)\,\eta(\mathrm{d}z)\right)=\exp\left(\int_E\left(e^{f(z)}-1\right)\Lambda(\mathrm{d}z)\right). \tag*{(34)} E exp ( ∫ E f ( z ) η ( d z ) ) = exp ( ∫ E ( e f ( z ) − 1 ) Λ ( d z ) ) . (34) (ii) For a square-integrable real function F F F of the point measure η \eta η , define D z F ( η ) = F ( η + δ z ) − F ( η ) D_zF(\eta)=F(\eta+\delta_z)-F(\eta) D z F ( η ) = F ( η + δ z ) − F ( η ) . Then we have
E [ ( F ( η ) ) 2 ] = ∑ k ≥ 0 1 k ! ∫ E k ( E [ D z 1 ⋯ D z k F ( η ) ] ) 2 ∏ h = 1 k Λ ( d z h ) . (35) \mathbb{E}\left[(F(\eta))^2\right]=\sum_{k\geq0}\frac{1}{k!}\int_{E^k}\left(\mathbb{E}[D_{z_1}\cdots D_{z_k}F(\eta)]\right)^2\prod_{h=1}^k\Lambda(\mathrm{d}z_h). \tag*{(35)} E [ ( F ( η ) ) 2 ] = k ≥ 0 ∑ k ! 1 ∫ E k ( E [ D z 1 ⋯ D z k F ( η )] ) 2 h = 1 ∏ k Λ ( d z h ) . (35) For k = 0 k=0 k = 0 , the summand is ( E F ) 2 (\mathbb{E}F)^2 ( E F ) 2 .
To handle products of Poisson integrals, we use the multivariate Mecke equation [22 Theorem 4.4] . We distinguish point occurrences from their locations: two distinct occurrences may be situated at the same location. By distinguishing occurrences from locations, we can work with atomic intensity measures throughout.
Lemma 2.6 (Factorial Mecke identity [22 ] ). For a Poisson random measure η \eta η on a measurable space E E E with intensity measure Λ \Lambda Λ , let η ( k ) \eta^{(k)} η ( k ) count ordered k k k -tuples of distinct point occurrences. For every nonnegative measurable function H ( η , z ) H(\eta,z) H ( η , z ) , we have
E ∫ E k H ( η , z ) d η ( k ) ( z ) = ∫ E k E H ( η + ∑ i = 1 k δ z i , z ) ∏ h = 1 k Λ ( d z h ) . (36) \mathbb{E}\int_{E^k}H(\eta,z)\,\mathrm{d}\eta^{(k)}(z)=\int_{E^k}\mathbb{E}H\left(\eta+\sum_{i=1}^k\delta_{z_i},z\right)\prod_{h=1}^k\Lambda(\mathrm{d}z_h). \tag*{(36)} E ∫ E k H ( η , z ) d η ( k ) ( z ) = ∫ E k E H ( η + i = 1 ∑ k δ z i , z ) h = 1 ∏ k Λ ( d z h ) . (36)
Before applying the Poisson exponential formula (34) , the Fock isometry (35) , and the factorial Mecke identity (36) to the cumulant families, we first separate the mass at zero in each representing measure. For x ≥ 0 x\geq0 x ≥ 0 , we write 1 ∧ x : = min { 1 , x } 1\wedge x:=\min\{1,x\} 1 ∧ x := min { 1 , x } . We obtain the following moment form of the classical subordinator representation; for the corresponding Laplace exponents, see [35 Theorems 3.2 and 5.2] . We include the elementary measure conversion because we shall use the individual formal moments.
Lemma 2.7 (Lévy cumulants in moment form [35 ] ). Assume that the shifted sequence ( κ n + 1 ( ℓ ) ) n ≥ 0 (\kappa_{n+1}^{(\ell)})_{n\geq0} ( κ n + 1 ( ℓ ) ) n ≥ 0 is a Stieltjes moment sequence for each ℓ = 1 , … , m \ell=1,\ldots,m ℓ = 1 , … , m . Then there exist b ℓ ≥ 0 b_\ell\geq0 b ℓ ≥ 0 and nonnegative measures ρ ℓ \rho_\ell ρ ℓ on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) such that, for every ℓ \ell ℓ ,
κ n ( ℓ ) = b ℓ 1 { n = 1 } + ∫ 0 ∞ x n ρ ℓ ( d x ) ( n ≥ 1 ) . (37) \kappa_n^{(\ell)}=b_\ell\mathbf{1}_{\{n=1\}}+\int_0^\infty x^n\,\rho_\ell(\mathrm{d}x)\qquad(n\geq1). \tag*{(37)} κ n ( ℓ ) = b ℓ 1 { n = 1 } + ∫ 0 ∞ x n ρ ℓ ( d x ) ( n ≥ 1 ) . (37) All moments in (37) are finite, and
∫ 0 ∞ ( 1 ∧ x ) ρ ℓ ( d x ) < ∞ . (38) \int_0^\infty(1\wedge x)\,\rho_\ell(\mathrm{d}x)<\infty. \tag*{(38)} ∫ 0 ∞ ( 1 ∧ x ) ρ ℓ ( d x ) < ∞. (38)
Proof. For each ℓ \ell ℓ , choose a nonnegative representing measure ν ℓ \nu_\ell ν ℓ on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that
κ n + 1 ( ℓ ) = ∫ 0 ∞ x n ν ℓ ( d x ) ( n ≥ 0 ) . (39) \kappa_{n+1}^{(\ell)}=\int_0^\infty x^n\,\nu_\ell(\mathrm{d}x)\qquad(n\geq0). \tag*{(39)} κ n + 1 ( ℓ ) = ∫ 0 ∞ x n ν ℓ ( d x ) ( n ≥ 0 ) . (39)
Taking n = 0 n=0 n = 0 in (39) , we obtain ν ℓ ( [ 0 , ∞ ) ) = κ 1 ( ℓ ) < ∞ \nu_\ell([0,\infty))=\kappa_1^{(\ell)}<\infty ν ℓ ([ 0 , ∞ )) = κ 1 ( ℓ ) < ∞ . Define
b ℓ = ν ℓ ( { 0 } ) , ρ ℓ ( d x ) = x − 1 ν ℓ ( d x ) ( x > 0 ) . (40) b_\ell=\nu_\ell(\{0\}),\qquad\rho_\ell(\mathrm{d}x)=x^{-1}\nu_\ell(\mathrm{d}x)\quad(x>0). \tag*{(40)} b ℓ = ν ℓ ({ 0 }) , ρ ℓ ( d x ) = x − 1 ν ℓ ( d x ) ( x > 0 ) . (40)
For n = 1 n=1 n = 1 , we recover (37) by adding the mass at zero to ∫ 0 ∞ x ρ ℓ = ν ℓ ( ( 0 , ∞ ) ) \int_0^\infty x\,\rho_\ell=\nu_\ell((0,\infty)) ∫ 0 ∞ x ρ ℓ = ν ℓ (( 0 , ∞ )) . On the other hand, for n ≥ 2 n\geq2 n ≥ 2 , the atom at zero contributes nothing to the ( n − 1 ) (n-1) ( n − 1 ) -st moment of ν ℓ \nu_\ell ν ℓ . Hence ∫ 0 ∞ x n d ρ ℓ = ∫ 0 ∞ x n − 1 d ν ℓ = κ n ( ℓ ) \int_0^\infty x^n\,\mathrm{d}\rho_\ell=\int_0^\infty x^{n-1}\,\mathrm{d}\nu_\ell=\kappa_n^{(\ell)} ∫ 0 ∞ x n d ρ ℓ = ∫ 0 ∞ x n − 1 d ν ℓ = κ n ( ℓ ) . Finally, by (40) , we have
∫ 0 1 x ρ ℓ ( d x ) = ν ℓ ( ( 0 , 1 ] ) < ∞ , ∫ 1 + ∞ ρ ℓ ( d x ) = ∫ 1 + ∞ x − 1 ν ℓ ( d x ) ≤ ν ℓ ( ( 1 , ∞ ) ) < ∞ . (41) \int_0^1 x\,\rho_\ell(\mathrm{d}x)=\nu_\ell((0,1])<\infty,\qquad\int_{1+}^\infty\rho_\ell(\mathrm{d}x)=\int_{1+}^\infty x^{-1}\nu_\ell(\mathrm{d}x)\leq\nu_\ell((1,\infty))<\infty. \tag*{(41)} ∫ 0 1 x ρ ℓ ( d x ) = ν ℓ (( 0 , 1 ]) < ∞ , ∫ 1 + ∞ ρ ℓ ( d x ) = ∫ 1 + ∞ x − 1 ν ℓ ( d x ) ≤ ν ℓ (( 1 , ∞ )) < ∞. (41)
Adding such two finite integrals in (41) , we obtain (38) .
Until the truncation step, we assume that each ρ ℓ \rho_\ell ρ ℓ is finite and compactly supported in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . For the probabilistic calculation we fix y ∈ ( 0 , ∞ ) m \boldsymbol{y}\in(0,\infty)^m y ∈ ( 0 , ∞ ) m and define
ρ y = ∑ ℓ = 1 m y ℓ ρ ℓ , b ( y ) = ∑ ℓ = 1 m b ℓ y ℓ . (42) \rho_{\boldsymbol{y}}=\sum_{\ell=1}^m y_\ell\rho_\ell,\qquad b(\boldsymbol{y})=\sum_{\ell=1}^m b_\ell y_\ell. \tag*{(42)} ρ y = ℓ = 1 ∑ m y ℓ ρ ℓ , b ( y ) = ℓ = 1 ∑ m b ℓ y ℓ . (42)
We write N y N_{\boldsymbol{y}} N y for a Poisson random measure of ρ y \rho_{\boldsymbol{y}} ρ y , and set
J y = ∫ 0 ∞ x N y ( d x ) , X y = b ( y ) + J y . (43) J_{\boldsymbol{y}}=\int_0^\infty x\,N_{\boldsymbol{y}}(\mathrm{d}x),\qquad X_{\boldsymbol{y}}=b(\boldsymbol{y})+J_{\boldsymbol{y}}. \tag*{(43)} J y = ∫ 0 ∞ x N y ( d x ) , X y = b ( y ) + J y . (43)
Lemma 2.8 (Compound-Poisson moment representation). Let b ℓ ≥ 0 b_\ell\geq0 b ℓ ≥ 0 and ρ ℓ \rho_\ell ρ ℓ in Lemma 2.7 , and suppose that each ρ ℓ \rho_\ell ρ ℓ is finite and compactly supported in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . For fixed y ∈ ( 0 , ∞ ) m \boldsymbol{y}\in(0,\infty)^m y ∈ ( 0 , ∞ ) m , let X y X_{\boldsymbol{y}} X y be defined by (42) –(43) . Then X y ≥ 0 X_{\boldsymbol{y}}\geq0 X y ≥ 0 , its moment generating function is finite for every real t t t , and M n ( y ) = E X y n M_n(\boldsymbol{y})=\mathbb{E}X_{\boldsymbol{y}}^n M n ( y ) = E X y n for every n ≥ 0 n\geq0 n ≥ 0 . In particular, ( M n ( y ) ) n ≥ 0 (M_n(\boldsymbol{y}))_{n\geq0} ( M n ( y ) ) n ≥ 0 is a Stieltjes moment sequence.
Proof. For clarity, we will derive the moment generating function explicitly. Put λ y = ρ y ( ( 0 , ∞ ) ) \lambda_y=\rho_y((0,\infty)) λ y = ρ y (( 0 , ∞ )) . If λ y > 0 \lambda_y>0 λ y > 0 , we may write J y = ∑ j = 1 P Y j J_y=\sum_{j=1}^{P}Y_j J y = ∑ j = 1 P Y j , where P P P has the Poisson distribution of mean λ y \lambda_y λ y , and the Y j Y_j Y j are independent with common law ρ y / λ y \rho_y/\lambda_y ρ y / λ y , independently of P P P . Conditioning on P P P , and then summing its probability mass function, we obtain
E e t J y = ∑ k ≥ 0 e − λ y λ y k k ! ( 1 λ y ∫ 0 ∞ e t x ρ y ( d x ) ) k = exp ( ∫ 0 ∞ ( e t x − 1 ) ρ y ( d x ) ) . (44) \mathbb{E}e^{tJ_y}=\sum_{k\geq0}e^{-\lambda_y}\frac{\lambda_y^k}{k!}\left(\frac{1}{\lambda_y}\int_0^\infty e^{tx}\rho_y(\mathrm{d}x)\right)^k=\exp\left(\int_0^\infty(e^{tx}-1)\rho_y(\mathrm{d}x)\right). \tag*{(44)} E e t J y = k ≥ 0 ∑ e − λ y k ! λ y k ( λ y 1 ∫ 0 ∞ e t x ρ y ( d x ) ) k = exp ( ∫ 0 ∞ ( e t x − 1 ) ρ y ( d x ) ) . (44)
If λ y = 0 \lambda_y=0 λ y = 0 , we take J y = 0 J_y=0 J y = 0 , and both sides of (44) are one. For bounded jump sizes, we may expand e t x − 1 e^{tx}-1 e t x − 1 under the finite integral. By (37) and (42) , we have
b ( y ) t + ∫ 0 ∞ ( e t x − 1 ) ρ y ( d x ) = ∑ ℓ = 1 m y ℓ ( b ℓ t + ∑ n ≥ 1 t n n ! ∫ 0 ∞ x n ρ ℓ ( d x ) ) = ∑ ℓ = 1 m y ℓ K ℓ ( t ) . (45) b(y)t+\int_0^\infty(e^{tx}-1)\rho_y(\mathrm{d}x)=\sum_{\ell=1}^{m}y_\ell\left(b_\ell t+\sum_{n\geq1}\frac{t^n}{n!}\int_0^\infty x^n\rho_\ell(\mathrm{d}x)\right)=\sum_{\ell=1}^{m}y_\ell K_\ell(t). \tag*{(45)} b ( y ) t + ∫ 0 ∞ ( e t x − 1 ) ρ y ( d x ) = ℓ = 1 ∑ m y ℓ ( b ℓ t + n ≥ 1 ∑ n ! t n ∫ 0 ∞ x n ρ ℓ ( d x ) ) = ℓ = 1 ∑ m y ℓ K ℓ ( t ) . (45)
Multiplying (44) by e b ( y ) t e^{b(y)t} e b ( y ) t , and using (45) , we obtain
E e t X y = E e b ( y ) t + t J y = exp ( b ( y ) t + ∫ 0 ∞ ( e t x − 1 ) ρ y ( d x ) ) = exp ( ∑ ℓ = 1 m y ℓ K ℓ ( t ) ) . (46) \mathbb{E}e^{tX_y}=\mathbb{E}e^{b(y)t+tJ_y}=\exp\left(b(y)t+\int_0^\infty(e^{tx}-1)\rho_y(\mathrm{d}x)\right)=\exp\left(\sum_{\ell=1}^{m}y_\ell K_\ell(t)\right). \tag*{(46)} E e t X y = E e b ( y ) t + t J y = exp ( b ( y ) t + ∫ 0 ∞ ( e t x − 1 ) ρ y ( d x ) ) = exp ( ℓ = 1 ∑ m y ℓ K ℓ ( t ) ) . (46)
We may differentiate at zero because the moment generating function is finite for every real t t t . Comparing (46) with (12) , we conclude that
M n ( y ) = E X y n ( n ≥ 0 ) . (47) M_n(y)=\mathbb{E}X_y^n \qquad(n\geq0). \tag*{(47)} M n ( y ) = E X y n ( n ≥ 0 ) . (47)
We will use only finite polynomial identities in y \boldsymbol{y} y . Once we have proved such an identity on y > 0 \boldsymbol{y}>0 y > 0 , we may therefore extend it to all boundary specializations as an identity in the polynomial ring R [ y ] \mathbb{R}[\boldsymbol{y}] R [ y ] .
A multivariate shifted Vandermonde square
For x = ( x 1 , … , x r ) \boldsymbol{x}=(x_1,\ldots,x_r) x = ( x 1 , … , x r ) , we continue to use Δ ( x ) = ∏ i < j ( x j − x i ) \Delta(\boldsymbol{x})=\prod_{i<j}(x_j-x_i) Δ ( x ) = ∏ i < j ( x j − x i ) , as in (33) . Take independent copies J y ( 1 ) , … , J y ( r ) J_y^{(1)},\ldots,J_y^{(r)} J y ( 1 ) , … , J y ( r ) of J y J_y J y , and write J y = ( J y ( 1 ) , … , J y ( r ) ) \boldsymbol{J}_y=(J_y^{(1)},\ldots,J_y^{(r)}) J y = ( J y ( 1 ) , … , J y ( r ) ) . We shall obtain nonnegative coefficients by expressing the second moment of the Vandermonde as a finite sum of integrals of squares.
Lemma 2.9 (Multivariate Poisson–Vandermonde positivity). For every fixed s ∈ R r \boldsymbol{s}\in\mathbb{R}^r s ∈ R r , define
V s ( y ) : = E [ Δ ( J y + s ) 2 ] . (48) V_{\boldsymbol{s}}(\boldsymbol{y}):=\mathbb{E}\left[\Delta(\boldsymbol{J}_y+\boldsymbol{s})^2\right]. \tag*{(48)} V s ( y ) := E [ Δ ( J y + s ) 2 ] . (48) Then we have V s ( y ) ∈ R ≥ 0 [ y ] V_{\boldsymbol{s}}(\boldsymbol{y})\in\mathbb{R}_{\geq0}[\boldsymbol{y}] V s ( y ) ∈ R ≥ 0 [ y ] and deg y V s ≤ D : = ( r 2 ) \deg_{\boldsymbol{y}}V_{\boldsymbol{s}}\leq D:=\binom{r}{2} deg y V s ≤ D := ( 2 r ) . Every coefficient is a polynomial in s \boldsymbol{s} s of degree at most 2 D 2D 2 D . The same assertions hold for finite compactly supported nonnegative jump measures on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } .
Proof. To remove the indeterminates from the kernels in the Fock expansion, we first compute the mean of the Vandermonde. For this purpose, we establish the following mean identity.
Claim 1. For every s ∈ R r s \in\mathbb{R}^{r} s ∈ R r , we have
E [ Δ ( J y + s ) ] = Δ ( s ) . (49) \mathbb{E}[\Delta(J_{\mathbf{y}}+s)] = \Delta(s). \tag*{(49)} E [ Δ ( J y + s )] = Δ ( s ) . (49)
Proof. Choose B < ∞ B < \infty B < ∞ that bounds the absolute jump sizes, and let P j P_j P j be the Poisson count in coordinate j j j . Then ∣ J y ( j ) ∣ ≤ B P j \lvert J_{\mathbf{y}}^{(j)}\rvert\le BP_j ∣ J y ( j ) ∣ ≤ B P j . For k ≥ 0 k \ge0 k ≥ 0 , define
p k ( u ) : = E ( J y + u ) k = ∑ h = 0 k ( k h ) E J y h u k − h . (50) p_k(u) := \mathbb{E}(J_{\mathbf{y}}+u)^k = \sum_{h=0}^{k} \binom{k}{h} \mathbb{E}J_{\mathbf{y}}^h u^{k-h}. \tag*{(50)} p k ( u ) := E ( J y + u ) k = h = 0 ∑ k ( h k ) E J y h u k − h . (50)
Since a Poisson count has finite moments of every order, we may integrate every polynomial above. In particular, p k ( u ) p_k(u) p k ( u ) is monic of degree k k k , because E J y 0 = 1 \mathbb{E}J_{\mathbf{y}}^0=1 E J y 0 = 1 .
Expanding the determinant and then using independence in each product, we obtain
E Δ ( J y + s ) = ∑ σ ∈ S r sgn ( σ ) ∏ i = 1 r E ( s σ ( i ) + J y ( σ ( i ) ) ) i − 1 = det [ p i − 1 ( s j ) ] i , j = 1 r . (51) \mathbb{E}\,\Delta(J_{\mathbf{y}}+s) = \sum_{\sigma\in S_r} \operatorname{sgn}(\sigma)\prod_{i=1}^{r}\mathbb{E}(s_{\sigma(i)}+J_{\mathbf{y}}^{(\sigma(i))})^{i-1} = \det[p_{i-1}(s_j)]_{i,j=1}^{r}. \tag*{(51)} E Δ ( J y + s ) = σ ∈ S r ∑ sgn ( σ ) i = 1 ∏ r E ( s σ ( i ) + J y ( σ ( i )) ) i − 1 = det [ p i − 1 ( s j ) ] i , j = 1 r . (51)
Here we use independence because the indices σ ( i ) \sigma(i) σ ( i ) are distinct.
Write A y ( s ) = [ p i − 1 ( s j ) ] i , j = 1 r A_{\mathbf{y}}(s)=[p_{i-1}(s_j)]_{i,j=1}^{r} A y ( s ) = [ p i − 1 ( s j ) ] i , j = 1 r and V ( s ) = [ s j i − 1 ] i , j = 1 r V(s)=[s_j^{i-1}]_{i,j=1}^{r} V ( s ) = [ s j i − 1 ] i , j = 1 r . By (50) , we have
A y ( s ) = L y V ( s ) , ( L y ) i , h = { ( i − 1 h − 1 ) E J y i − h , 1 ≤ h ≤ i ≤ r , 0 , 1 ≤ i < h ≤ r . (52) A_{\mathbf{y}}(s)=L_{\mathbf{y}}V(s), \qquad(L_{\mathbf{y}})_{i,h} =
\begin{cases}
\binom{i-1}{h-1}\mathbb{E}J_{\mathbf{y}}^{i-h}, & 1\le h\le i\le r,\\
0, & 1\le i<h\le r.
\end{cases} \tag*{(52)} A y ( s ) = L y V ( s ) , ( L y ) i , h = { ( h − 1 i − 1 ) E J y i − h , 0 , 1 ≤ h ≤ i ≤ r , 1 ≤ i < h ≤ r . (52)
Since L y L_{\mathbf{y}} L y is lower triangular with diagonal entries equal to one, we obtain
det A y ( s ) = det L y det V ( s ) = Δ ( s ) . (53) \det A_{\mathbf{y}}(s)=\det L_{\mathbf{y}}\det V(s)=\Delta(s). \tag*{(53)} det A y ( s ) = det L y det V ( s ) = Δ ( s ) . (53)
Combining (51) and (53) yields (49) .
Using the mean identity, we now express the second moment as an integral of squares with kernels independent of y \mathbf{y} y . We use this expansion to establish coefficientwise nonnegativity and the degree bound in the following claim.
Claim 2. For every s ∈ R r s \in\mathbb{R}^{r} s ∈ R r and α ∈ N 0 m \alpha\in\mathbb{N}_{0}^{m} α ∈ N 0 m , we have
[ y α ] V s ( y ) ≥ 0 ( ∣ α ∣ ≤ D ) , [ y α ] V s ( y ) = 0 ( ∣ α ∣ > D ) . (54) [\mathbf{y}^{\alpha}]V_s(\mathbf{y})\ge0 \quad(|\alpha|\le D), \qquad[\mathbf{y}^{\alpha}]V_s(\mathbf{y})=0 \quad(|\alpha|>D). \tag*{(54)} [ y α ] V s ( y ) ≥ 0 ( ∣ α ∣ ≤ D ) , [ y α ] V s ( y ) = 0 ( ∣ α ∣ > D ) . (54)
Proof. Step 1: One point measure for all coordinates. Take E = { 1 , … , r } × ( 0 , ∞ ) E=\{1,\ldots,r\}\times(0,\infty) E = { 1 , … , r } × ( 0 , ∞ ) . For Borel A ⊆ ( 0 , ∞ ) A\subseteq(0,\infty) A ⊆ ( 0 , ∞ ) , define
Λ ℓ ( { j } × A ) = ρ ℓ ( A ) , Λ y = ∑ ℓ = 1 m y ℓ Λ ℓ . (55) \Lambda_\ell(\{j\}\times A)=\rho_\ell(A),\qquad\Lambda_y=\sum_{\ell=1}^{m}y_\ell\Lambda_\ell. \tag*{(55)} Λ ℓ ({ j } × A ) = ρ ℓ ( A ) , Λ y = ℓ = 1 ∑ m y ℓ Λ ℓ . (55)
Let η \eta η be a Poisson random measure with intensity Λ y \Lambda_y Λ y . The restrictions of η \eta η to the coordinate copies are independent; hence we may realize J y ( j ) J_y^{(j)} J y ( j ) as the jump sum on copy j j j . For a function f : R r → R f:\mathbb{R}^r\to\mathbb{R} f : R r → R and z = ( j , x ) ∈ E z=(j,x)\in E z = ( j , x ) ∈ E , put
D ( j , x ) f ( s ) = f ( s + x e j ) − f ( s ) . (56) \mathcal{D}_{(j,x)}f(s)=f(s+xe_j)-f(s). \tag*{(56)} D ( j , x ) f ( s ) = f ( s + x e j ) − f ( s ) . (56)
Step 2: Expected difference kernels. Set F ( η ) = Δ ( J y + s ) F(\eta)=\Delta(J_y+s) F ( η ) = Δ ( J y + s ) . On adding an occurrence ( j , x ) (j,x) ( j , x ) , we increase only the j j j -th jump sum by x x x . Therefore, iterating (56) and then taking expectations, we obtain
E [ D z 1 ⋯ D z k F ( η ) ] = D z 1 ⋯ D z k E [ Δ ( J y + s ) ] = D z 1 ⋯ D z k Δ ( s ) . (57) \mathbb{E}[D_{z_1}\cdots D_{z_k}F(\eta)]=\mathcal{D}_{z_1}\cdots\mathcal{D}_{z_k}\mathbb{E}[\Delta(J_y+s)]=\mathcal{D}_{z_1}\cdots\mathcal{D}_{z_k}\Delta(s). \tag*{(57)} E [ D z 1 ⋯ D z k F ( η )] = D z 1 ⋯ D z k E [ Δ ( J y + s )] = D z 1 ⋯ D z k Δ ( s ) . (57)
For the first equality, we interchange expectation with a finite signed sum of integrable polynomials. For the second, we use Claim 1 at every translated shift. We have thus obtained kernels that do not depend on y \boldsymbol{y} y .
Step 3: A finite expansion into integrals of squares. For fixed jump vectors, a nonzero translation difference reduces the degree in s s s by at least one. Since Δ \Delta Δ has degree D D D , the last expression in (57) vanishes for k > D k>D k > D . Write P j P_j P j for the Poisson count on copy j j j , and choose B B B that bounds the absolute jump sizes. Then, for a constant C s , B C_{s,B} C s , B ,
∣ F ( η ) ∣ 2 ≤ C s , B ( 1 + ∑ j = 1 r P j ) 2 D , E ∣ F ( η ) ∣ 2 < ∞ . (58) \lvert F(\eta)\rvert^2\le C_{s,B}\left(1+\sum_{j=1}^{r}P_j\right)^{2D},\qquad\mathbb{E}\lvert F(\eta)\rvert^2<\infty. \tag*{(58)} ∣ F ( η ) ∣ 2 ≤ C s , B ( 1 + j = 1 ∑ r P j ) 2 D , E ∣ F ( η ) ∣ 2 < ∞. (58)
Applying the Fock isometry (35) to F ( η ) = Δ ( J y + s ) F(\eta)=\Delta(J_y+s) F ( η ) = Δ ( J y + s ) , and then using (57) , we have
V s ( y ) = ∑ k = 0 D 1 k ! ∫ E k ( D z 1 ⋯ D z k Δ ( s ) ) 2 ∏ h = 1 k Λ y ( d z h ) . (59) V_s(y)=\sum_{k=0}^{D}\frac{1}{k!}\int_{E^k}\left(\mathcal{D}_{z_1}\cdots\mathcal{D}_{z_k}\Delta(s)\right)^2\prod_{h=1}^{k}\Lambda_y(\mathrm{d}z_h). \tag*{(59)} V s ( y ) = k = 0 ∑ D k ! 1 ∫ E k ( D z 1 ⋯ D z k Δ ( s ) ) 2 h = 1 ∏ k Λ y ( d z h ) . (59)
For k = 0 k=0 k = 0 , we interpret the integral as Δ ( s ) 2 \Delta(s)^2 Δ ( s ) 2 .
Step 4: Coefficient extraction and polynomial dependence on the shifts. Let ∣ α ∣ = k ≤ D \lvert\alpha\rvert=k\le D ∣ α ∣ = k ≤ D and α ! = ∏ ℓ = 1 m α ℓ ! \alpha!=\prod_{\ell=1}^{m}\alpha_\ell! α ! = ∏ ℓ = 1 m α ℓ ! . Choose an ordered color list ( c 1 , … , c k ) (c_1,\ldots,c_k) ( c 1 , … , c k ) containing α ℓ \alpha_\ell α ℓ copies of ℓ \ell ℓ , with all copies of color 1 first, then color 2, and so on. We integrate z h z_h z h against Λ c h \Lambda_{c_h} Λ c h . Applying the expansion (55) to (59) , we have
[ y α ] V s ( y ) = 1 α ! ∫ E k ( D z 1 ⋯ D z k Δ ( s ) ) 2 ∏ h = 1 k Λ c h ( d z h ) ≥ 0. (60) [y^\alpha]V_s(y)=\frac{1}{\alpha!}\int_{E^k}\left(\mathcal{D}_{z_1}\cdots\mathcal{D}_{z_k}\Delta(s)\right)^2\prod_{h=1}^{k}\Lambda_{c_h}(\mathrm{d}z_h)\ge0. \tag*{(60)} [ y α ] V s ( y ) = α ! 1 ∫ E k ( D z 1 ⋯ D z k Δ ( s ) ) 2 h = 1 ∏ k Λ c h ( d z h ) ≥ 0. (60)
Indeed, the difference operators commute, so the kernel is symmetric in z 1 , … , z k z_1,\ldots,z_k z 1 , … , z k . The k ! / α ! k!/\alpha! k ! / α ! assignments of the colors give equal integrals; after multiplication by 1 / k ! 1/k! 1/ k ! , we obtain 1 / α ! 1/\alpha! 1/ α ! . No term with ∣ α ∣ > D |\alpha|>D ∣ α ∣ > D occurs in (59) . Finally, the squared kernel has degree at most 2 ( D − k ) 2(D-k) 2 ( D − k ) in s s s . Since the measures are finite with bounded jump sizes, we may integrate its finitely many polynomial coefficients. Thus every coefficient of V s V_s V s is a polynomial in s s s of degree at most 2 D 2D 2 D .
Combining such two claims, we prove the lemma for nonnegative jump locations. We used no sign restriction on the shifts or on the jump locations in the determinant and square-integral calculations. Applying the same calculations to E = { 1 , … , r } × ( R ∖ { 0 } ) E=\{1,\ldots,r\}\times(\mathbb{R}\setminus\{0\}) E = { 1 , … , r } × ( R ∖ { 0 }) , we have the real-line extension as well.
A partitioned form of the Mecke formula
For the next step, we need a power of a Poisson integral multiplied by a function of the same process. We record the following standard consequence of the multivariate Mecke equation [22 Theorem 4.4] , recalled in Lemma 2.6 . We obtain it by grouping the factors in the power according to set partitions. We include the complete reduction to specify the treatment of repeated locations and the absence of any additional factorial.
Lemma 2.10 (Partitioned consequence of the Mecke formula). Let N N N be a Poisson random measure on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) with finite, compactly supported measure θ \theta θ , and put J = ∫ 0 ∞ x N ( d x ) J=\int_0^\infty x\,N(\mathrm{d}x) J = ∫ 0 ∞ x N ( d x ) . For p ∈ N 0 p\in\mathbb{N}_0 p ∈ N 0 and a nonnegative measurable function G G G on the space of finite point measures on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , we have
E [ J p G ( N ) ] = ∑ π ∈ P p ∫ ( 0 , ∞ ) ∣ π ∣ ( ∏ B ∈ π x B ∣ B ∣ ) E G ( N + ∑ B ∈ π δ x B ) ∏ B ∈ π θ ( d x B ) . (61) \mathbb{E}\left[J^pG(N)\right]=\sum_{\pi\in\mathcal{P}_p}\int_{(0,\infty)^{|\pi|}}\left(\prod_{B\in\pi}x_B^{|B|}\right)\mathbb{E}G\left(N+\sum_{B\in\pi}\delta_{x_B}\right)\prod_{B\in\pi}\theta(\mathrm{d}x_B). \tag*{(61)} E [ J p G ( N ) ] = π ∈ P p ∑ ∫ ( 0 , ∞ ) ∣ π ∣ ( B ∈ π ∏ x B ∣ B ∣ ) E G ( N + B ∈ π ∑ δ x B ) B ∈ π ∏ θ ( d x B ) . (61) For p = 0 p=0 p = 0 , we use the single empty-partition term. The identity holds in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] .
Proof. Step 1: Grouping the factors by distinct occurrences. We first prove a deterministic identity for a finite point measure. Write η = ∑ j = 1 M δ x j \eta=\sum_{j=1}^{M}\delta_{x_j} η = ∑ j = 1 M δ x j , where the indices j j j distinguish occurrences even when some locations x j x_j x j coincide. For p ≥ 1 p\geq1 p ≥ 1 , we expand the power and obtain
( ∫ 0 ∞ x η ( d x ) ) p = ∑ a 1 , … , a p = 1 M x a 1 ⋯ x a p . (62) \left(\int_0^\infty x\,\eta(\mathrm{d}x)\right)^p=\sum_{a_1,\ldots,a_p=1}^{M}x_{a_1}\cdots x_{a_p}. \tag*{(62)} ( ∫ 0 ∞ x η ( d x ) ) p = a 1 , … , a p = 1 ∑ M x a 1 ⋯ x a p . (62)
For a tuple ( a 1 , … , a p ) (a_1,\ldots,a_p) ( a 1 , … , a p ) , place two positions u , v ∈ [ p ] u,v\in[p] u , v ∈ [ p ] in the same block exactly when a u = a v a_u=a_v a u = a v . In this way we associate one partition π ∈ P p \pi\in\mathcal{P}_p π ∈ P p with the tuple. Order the blocks by their least elements, say B 1 , … , B k B_1,\ldots,B_k B 1 , … , B k . For a fixed partition, we must choose k k k distinct occurrence indices j 1 , … , j k j_1,\ldots,j_k j 1 , … , j k , one for each ordered block. The product in (62) is then ∏ h = 1 k x j h ∣ B h ∣ \prod_{h=1}^{k} x_{j_h}^{|B_h|} ∏ h = 1 k x j h ∣ B h ∣ . Conversely, from the partition and the ordered distinct indices, we recover exactly one original tuple. Therefore, using the factorial measure of distinct occurrences, we deduce
( ∫ 0 ∞ x η ( d x ) ) p = ∑ π ∈ P p ∫ ( 0 , ∞ ) ∣ π ∣ ( ∏ B ∈ π x B ∣ B ∣ ) η ( ∣ π ∣ ) ( d x π ) . (63) \left(\int_{0}^{\infty} x\,\eta(\mathrm{d}x)\right)^{p}
=
\sum_{\pi\in\mathcal{P}_{p}}
\int_{(0,\infty)^{|\pi|}}
\left(\prod_{B\in\pi}x_{B}^{|B|}\right)\eta^{(|\pi|)}(\mathrm{d}x_{\pi}). \tag*{(63)} ( ∫ 0 ∞ x η ( d x ) ) p = π ∈ P p ∑ ∫ ( 0 , ∞ ) ∣ π ∣ ( B ∈ π ∏ x B ∣ B ∣ ) η ( ∣ π ∣ ) ( d x π ) . (63)
Step 2: The factorial Mecke identity for one partition. Because θ \theta θ is finite, the Poisson random measure N N N is finite almost surely. Applying (63) to N N N , and then multiplying by G ( N ) G(N) G ( N ) , we may take expectations term by term. For a partition π = { B 1 , … , B k } \pi=\{B_1,\ldots,B_k\} π = { B 1 , … , B k } , with the blocks in their canonical order, write x = ( x 1 , … , x k ) x=(x_1,\ldots,x_k) x = ( x 1 , … , x k ) and set
w π ( x ) = ∏ h = 1 k x h ∣ B h ∣ , g k ( x ) = E G ( N + ∑ h = 1 k δ x h ) . (64) w_{\pi}(x)=\prod_{h=1}^{k}x_h^{|B_h|},
\qquad
g_k(x)=\mathbb{E}G\left(N+\sum_{h=1}^{k}\delta_{x_h}\right). \tag*{(64)} w π ( x ) = h = 1 ∏ k x h ∣ B h ∣ , g k ( x ) = E G ( N + h = 1 ∑ k δ x h ) . (64)
Thus w π ( x ) w_{\pi}(x) w π ( x ) is the product of the block weights, and g k ( x ) g_k(x) g k ( x ) is the expected value of G G G after inserting one occurrence at each of x 1 , … , x k x_1,\ldots,x_k x 1 , … , x k . Applying the factorial Mecke identity (36) to H π ( η , x ) = G ( η ) w π ( x ) H_{\pi}(\eta,x)=G(\eta)w_{\pi}(x) H π ( η , x ) = G ( η ) w π ( x ) , we obtain
E ∫ ( 0 , ∞ ) k G ( N ) w π ( x ) N ( k ) ( d x ) = ∫ ( 0 , ∞ ) k w π ( x ) g k ( x ) ∏ h = 1 k θ ( d x h ) . (65) \mathbb{E}\int_{(0,\infty)^k}G(N)w_{\pi}(x)N^{(k)}(\mathrm{d}x)
=
\int_{(0,\infty)^k}w_{\pi}(x)g_k(x)\prod_{h=1}^{k}\theta(\mathrm{d}x_h). \tag*{(65)} E ∫ ( 0 , ∞ ) k G ( N ) w π ( x ) N ( k ) ( d x ) = ∫ ( 0 , ∞ ) k w π ( x ) g k ( x ) h = 1 ∏ k θ ( d x h ) . (65)
Step 3: Summation and the empty partition. Summing (65) over all partitions, we recover (61) . We require no factor k ! k! k ! : we fixed one block order, and we already count all ordered distinct occurrence choices through the factorial measure. Moreover, since the integrands are nonnegative, we may interchange expectation and integration by Tonelli’s theorem without assuming that E G ( N ) < ∞ \mathbb{E}G(N)<\infty E G ( N ) < ∞ . For p = 0 p=0 p = 0 , both sides are E G ( N ) \mathbb{E}G(N) E G ( N ) under the empty-product convention.
For example, when p = 2 p=2 p = 2 , we have exactly two partitions. Therefore, using g 1 ( x ) = E G ( N + δ x ) g_1(x)=\mathbb{E}G(N+\delta_x) g 1 ( x ) = E G ( N + δ x ) and g 2 ( x , y ) = E G ( N + δ x + δ y ) g_2(x,y)=\mathbb{E}G(N+\delta_x+\delta_y) g 2 ( x , y ) = E G ( N + δ x + δ y ) , we obtain
E [ J 2 G ( N ) ] = ∫ 0 ∞ x 2 g 1 ( x ) θ ( d x ) + ∫ ( 0 , ∞ ) 2 x y g 2 ( x , y ) θ ( d x ) θ ( d y ) . (66) \mathbb{E}\left[J^2G(N)\right]
=
\int_{0}^{\infty}x^2g_1(x)\,\theta(\mathrm{d}x)
+
\int_{(0,\infty)^2}xy\,g_2(x,y)\,\theta(\mathrm{d}x)\theta(\mathrm{d}y). \tag*{(66)} E [ J 2 G ( N ) ] = ∫ 0 ∞ x 2 g 1 ( x ) θ ( d x ) + ∫ ( 0 , ∞ ) 2 x y g 2 ( x , y ) θ ( d x ) θ ( d y ) . (66)
In the second integral we also include x = y x=y x = y when the intensity has atoms; in that case we add two occurrences at that location. For θ = λ δ c \theta=\lambda\delta_c θ = λ δ c and G ( N ) = 1 G(N)=1 G ( N ) = 1 , we obtain E J 2 = c 2 ( λ + λ 2 ) \mathbb{E}J^2=c^2(\lambda+\lambda^2) E J 2 = c 2 ( λ + λ 2 ) , as required by the Poisson second moment. This completes the proof.
Universal sufficiency of Theorem 1.7
We now prove preservation for an arbitrary initial Stieltjes moment sequence. We first reduce a Hankel minor to weighted Poisson moment; we then prove coefficientwise nonnegativity and justify the passage from compactly supported jumps to general jumps. Proof of Theorem 1.7 : (i)⇒ \Rightarrow ⇒ (iv). Assume (i), and fix an initial Stieltjes moment sequence a = ( a n ) n ≥ 0 a=(a_n)_{n\geq0} a = ( a n ) n ≥ 0 . Choose a finite nonnegative measure μ \mu μ on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that
a n = ∫ 0 ∞ z n μ ( d z ) < ∞ ( n ≥ 0 ) . (67) a_n=\int_{0}^{\infty}z^n\,\mu(\mathrm{d}z)<\infty\qquad(n\geq0). \tag*{(67)} a n = ∫ 0 ∞ z n μ ( d z ) < ∞ ( n ≥ 0 ) . (67)
If a 0 = 0 a_0=0 a 0 = 0 , then μ \mu μ has zero total mass, so every a n a_n a n and every F n ( y ) F_n(\boldsymbol{y}) F n ( y ) is zero. The conclusion is immediate in that case. We may therefore assume a 0 > 0 a_0>0 a 0 > 0 , without normalizing its value. This preserves the full range of initial moment sequences in the statement.
Applying Lemma 2.7 to each shifted cumulant sequence, we have the drift and jump representation (37) . We first assume that every ρ ℓ \rho_\ell ρ ℓ is finite and compactly supported in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) .
Step 1: An integral representation of a fixed minor. Fix strictly increasing lists I = ( i 1 , … , i r ) I=(i_1,\ldots,i_r) I = ( i 1 , … , i r ) and J = ( j 1 , … , j r ) J=(j_1,\ldots,j_r) J = ( j 1 , … , j r ) of nonnegative integers, and let
D I , J μ ( y ) : = det [ F i a + j b ( y ) ] a , b = 1 r . (68) D^\mu_{I,J}(\boldsymbol{y}):=\det[F_{i_a+j_b}(\boldsymbol{y})]_{a,b=1}^r. \tag*{(68)} D I , J μ ( y ) := det [ F i a + j b ( y ) ] a , b = 1 r . (68)
For numerical y ∈ ( 0 , ∞ ) m \boldsymbol{y}\in(0,\infty)^m y ∈ ( 0 , ∞ ) m , we use X y X_{\boldsymbol{y}} X y and J y J_{\boldsymbol{y}} J y from (43) . Applying the binomial theorem to ( z + X y ) n (z+X_{\boldsymbol{y}})^n ( z + X y ) n , and using (47) and (13) , we have
∫ 0 ∞ E ( z + X y ) n μ ( d z ) = ∑ k = 0 n ( n k ) ( ∫ 0 ∞ z n − k μ ( d z ) ) E X y k = ∑ k = 0 n ( n k ) a n − k M k ( y ) = F n ( y ) . (69) \begin{aligned}
\int_{0}^{\infty}\mathbb{E}(z+X_{\boldsymbol{y}})^n\,\mu(\mathrm{d}z)
&=\sum_{k=0}^{n}\binom{n}{k}\left(\int_{0}^{\infty}z^{n-k}\,\mu(\mathrm{d}z)\right)\mathbb{E}X_{\boldsymbol{y}}^k\\
&=\sum_{k=0}^{n}\binom{n}{k}a_{n-k}M_k(\boldsymbol{y})=F_n(\boldsymbol{y}). \tag*{(69)}
\end{aligned} ∫ 0 ∞ E ( z + X y ) n μ ( d z ) = k = 0 ∑ n ( k n ) ( ∫ 0 ∞ z n − k μ ( d z ) ) E X y k = k = 0 ∑ n ( k n ) a n − k M k ( y ) = F n ( y ) . (69)
Every term is finite; we require no convergence assumption on the formal power series A ( t ) A(t) A ( t ) .
Take independent copies J y ( 1 ) , … , J y ( r ) J_{\boldsymbol{y}}^{(1)},\ldots,J_{\boldsymbol{y}}^{(r)} J y ( 1 ) , … , J y ( r ) , and write
W y ( z ) = J y + z + b ( y ) 1. (70) W_{\boldsymbol{y}}(z)=J_{\boldsymbol{y}}+z+b(\boldsymbol{y})\mathbf{1}. \tag*{(70)} W y ( z ) = J y + z + b ( y ) 1 . (70)
The measure whose moments occur in (69) is the convolution of μ \mu μ with the law of X y X_{\boldsymbol{y}} X y and it has all moments finite. Applying Andréief’s identity (31) to that measure and the monomials with exponents I , J I,J I , J , and then using the Schur alternant formula (33) , we obtain
D I , J μ ( y ) = 1 r ! ∫ [ 0 , ∞ ) r E [ Δ ( J y + z ) 2 s λ ( W y ( z ) ) s η ( W y ( z ) ) ] ∏ j = 1 r μ ( d z j ) , (71) D^\mu_{I,J}(\boldsymbol{y})=\frac{1}{r!}\int_{[0,\infty)^r}\mathbb{E}\left[\Delta(J_{\boldsymbol{y}}+z)^2s_\lambda(W_{\boldsymbol{y}}(z))s_\eta(W_{\boldsymbol{y}}(z))\right]\prod_{j=1}^{r}\mu(\mathrm{d}z_j), \tag*{(71)} D I , J μ ( y ) = r ! 1 ∫ [ 0 , ∞ ) r E [ Δ ( J y + z ) 2 s λ ( W y ( z )) s η ( W y ( z )) ] j = 1 ∏ r μ ( d z j ) , (71)
where λ a = i r + 1 − a − ( r − a ) \lambda_a=i_{r+1-a}-(r-a) λ a = i r + 1 − a − ( r − a ) and η a = j r + 1 − a − ( r − a ) \eta_a=j_{r+1-a}-(r-a) η a = j r + 1 − a − ( r − a ) . To remove the drift from the Vandermonde, we used the identity
Δ ( x + c 1 ) = ∏ a < b ( ( x b + c ) − ( x a + c ) ) = Δ ( x ) . (72) \Delta(\boldsymbol{x}+c\mathbf{1})=\prod_{a<b}\bigl((x_b+c)-(x_a+c)\bigr)=\Delta(\boldsymbol{x}). \tag*{(72)} Δ ( x + c 1 ) = a < b ∏ ( ( x b + c ) − ( x a + c ) ) = Δ ( x ) . (72)
Step 2: Schur and multinomial expansions. Using the tableau definition (32) , we write
s λ ( w ) s η ( w ) = ∑ α ∈ N 0 r c α w α , c α ≥ 0 , (73) s_{\lambda}(\mathbf{w})s_{\eta}(\mathbf{w})=\sum_{\boldsymbol{\alpha}\in\mathbb{N}_0^r}c_{\boldsymbol{\alpha}}\mathbf{w}^{\boldsymbol{\alpha}},\qquad c_{\boldsymbol{\alpha}}\ge0, \tag*{(73)} s λ ( w ) s η ( w ) = α ∈ N 0 r ∑ c α w α , c α ≥ 0 , (73)
with finitely many nonzero coefficients. For multi-indices in N 0 r \mathbb{N}_0^r N 0 r , put α ! = ∏ j α j ! \boldsymbol{\alpha}!=\prod_j\alpha_j! α ! = ∏ j α j ! . Applying the multinomial formula to each factor, we have
∏ j = 1 r ( J y ( j ) + z j + b ( y ) ) α j = ∑ β + γ + δ = α α ! β ! γ ! δ ! ( ∏ j = 1 r ( J y ( j ) ) β j ) z γ b ( y ) ∣ δ ∣ . (74) \prod_{j=1}^r\left(J_y^{(j)}+z_j+b(\mathbf{y})\right)^{\alpha_j}
=\sum_{\boldsymbol{\beta}+\boldsymbol{\gamma}+\boldsymbol{\delta}=\boldsymbol{\alpha}}
\frac{\boldsymbol{\alpha}!}{\boldsymbol{\beta}!\boldsymbol{\gamma}!\boldsymbol{\delta}!}
\left(\prod_{j=1}^r\left(J_y^{(j)}\right)^{\beta_j}\right)\mathbf{z}^{\boldsymbol{\gamma}}b(\mathbf{y})^{|\boldsymbol{\delta}|}. \tag*{(74)} j = 1 ∏ r ( J y ( j ) + z j + b ( y ) ) α j = β + γ + δ = α ∑ β ! γ ! δ ! α ! ( j = 1 ∏ r ( J y ( j ) ) β j ) z γ b ( y ) ∣ δ ∣ . (74)
Substituting these expansions into (71) , and taking the deterministic factors outside the expectation, we obtain
D I , J μ ( y ) = 1 r ! ∑ α c α ∑ β + γ + δ = α α ! β ! γ ! δ ! b ( y ) ∣ δ ∣ × ∫ [ 0 , ∞ ) r z γ E [ Δ ( J y + z ) 2 ∏ j = 1 r ( J y ( j ) ) β j ] ∏ j = 1 r μ ( d z j ) . (75) \begin{aligned}
D_{I,J}^{\mu}(\mathbf{y})
&=\frac{1}{r!}\sum_{\boldsymbol{\alpha}}c_{\boldsymbol{\alpha}}
\sum_{\boldsymbol{\beta}+\boldsymbol{\gamma}+\boldsymbol{\delta}=\boldsymbol{\alpha}}
\frac{\boldsymbol{\alpha}!}{\boldsymbol{\beta}!\boldsymbol{\gamma}!\boldsymbol{\delta}!}
b(\mathbf{y})^{|\boldsymbol{\delta}|}\\
&\quad\times\int_{[0,\infty)^r}\mathbf{z}^{\boldsymbol{\gamma}}\mathbb{E}\left[\Delta(\mathbf{J}_{\mathbf{y}}+\mathbf{z})^2\prod_{j=1}^r\left(J_y^{(j)}\right)^{\beta_j}\right]\prod_{j=1}^r\mu(dz_j). \tag*{(75)}
\end{aligned} D I , J μ ( y ) = r ! 1 α ∑ c α β + γ + δ = α ∑ β ! γ ! δ ! α ! b ( y ) ∣ δ ∣ × ∫ [ 0 , ∞ ) r z γ E [ Δ ( J y + z ) 2 j = 1 ∏ r ( J y ( j ) ) β j ] j = 1 ∏ r μ ( d z j ) . (75)
Thus we need to treat finitely many expressions of the form
b ( y ) h ∫ [ 0 , ∞ ) r z γ E [ Δ ( J y + z ) 2 ∏ j = 1 r ( J y ( j ) ) β j ] ∏ j = 1 r μ ( d z j ) , h ∈ N 0 . (76) b(\mathbf{y})^h\int_{[0,\infty)^r}\mathbf{z}^{\boldsymbol{\gamma}}\mathbb{E}\left[\Delta(\mathbf{J}_{\mathbf{y}}+\mathbf{z})^2\prod_{j=1}^r\left(J_y^{(j)}\right)^{\beta_j}\right]\prod_{j=1}^r\mu(dz_j),\qquad h\in\mathbb{N}_0. \tag*{(76)} b ( y ) h ∫ [ 0 , ∞ ) r z γ E [ Δ ( J y + z ) 2 j = 1 ∏ r ( J y ( j ) ) β j ] j = 1 ∏ r μ ( d z j ) , h ∈ N 0 . (76)
To separate the positivity argument from the integrability argument, we first keep z \mathbf{z} z fixed and prove coefficientwise nonnegativity. We then justify integration against μ \mu μ in each initial coordinate.
Claim 3. For z ∈ [ 0 , ∞ ) r \mathbf{z}\in[0,\infty)^r z ∈ [ 0 , ∞ ) r and β ∈ N 0 r \boldsymbol{\beta}\in\mathbb{N}_0^r β ∈ N 0 r , we have
P β , z ( y ) : = E [ Δ ( J y + z ) 2 ∏ j = 1 r ( J y ( j ) ) β j ] ∈ R ≥ 0 [ y ] . (77) P_{\boldsymbol{\beta},\mathbf{z}}(\mathbf{y}):=\mathbb{E}\left[\Delta(\mathbf{J}_{\mathbf{y}}+\mathbf{z})^2\prod_{j=1}^r\left(J_y^{(j)}\right)^{\beta_j}\right]\in\mathbb{R}_{\ge0}[\mathbf{y}]. \tag*{(77)} P β , z ( y ) := E [ Δ ( J y + z ) 2 j = 1 ∏ r ( J y ( j ) ) β j ] ∈ R ≥ 0 [ y ] . (77)
Proof. Step 1: Coordinate processes and tagged partition blocks. Fix z \mathbf{z} z and β \boldsymbol{\beta} β , and first take numerical y ∈ ( 0 , ∞ ) m \mathbf{y}\in(0,\infty)^m y ∈ ( 0 , ∞ ) m . Note
J y ( j ) = ∫ 0 ∞ x N y ( j ) ( d x ) , ρ y = ∑ ℓ = 1 m y ℓ ρ ℓ , (78) J_y^{(j)}=\int_0^\infty x\,N_y^{(j)}(dx),\qquad\rho_y=\sum_{\ell=1}^m y_\ell\rho_\ell, \tag*{(78)} J y ( j ) = ∫ 0 ∞ x N y ( j ) ( d x ) , ρ y = ℓ = 1 ∑ m y ℓ ρ ℓ , (78)
using independent Poisson random measures N y ( j ) N_y^{(j)} N y ( j ) of common intensity ρ y \rho_y ρ y . Let π = ( π 1 , … , π r ) \boldsymbol{\pi}=(\pi_1,\ldots,\pi_r) π = ( π 1 , … , π r ) , with π j ∈ P β j \pi_j\in\mathcal{P}_{\beta_j} π j ∈ P β j . To distinguish blocks in different coordinates, use the tagged set
B π : = { ( j , B ) : 1 ≤ j ≤ r , B ∈ π j } , k π : = ∑ j = 1 r ∣ π j ∣ . (79) B_{\boldsymbol{\pi}}:=\{(j,B):1\le j\le r,\ B\in\pi_j\},\qquad k_{\boldsymbol{\pi}}:=\sum_{j=1}^r|\pi_j|. \tag*{(79)} B π := {( j , B ) : 1 ≤ j ≤ r , B ∈ π j } , k π := j = 1 ∑ r ∣ π j ∣. (79)
For one integration variable x j , B x_{j,B} x j , B per tagged block, put
w π ( x ) : = ∏ ( j , B ) ∈ B π x j , B ∣ B ∣ , s π , j ( x ) : = ∑ B ∈ π j x j , B , s π ( x ) : = ( s π , 1 ( x ) , … , s π , r ( x ) ) . (80) w_{\pi}(\boldsymbol{x}) := \prod_{(j,B)\in\mathcal{B}_{\pi}} x_{j,B}^{|B|},
\qquad
s_{\pi,j}(\boldsymbol{x}) := \sum_{B\in\pi_j} x_{j,B},
\qquad
s_{\pi}(\boldsymbol{x}) := \left(s_{\pi,1}(\boldsymbol{x}),\ldots,s_{\pi,r}(\boldsymbol{x})\right). \tag*{(80)} w π ( x ) := ( j , B ) ∈ B π ∏ x j , B ∣ B ∣ , s π , j ( x ) := B ∈ π j ∑ x j , B , s π ( x ) := ( s π , 1 ( x ) , … , s π , r ( x ) ) . (80)
On inserting the occurrences indexed by π j \pi_j π j , we increase the j j j -th jump sum by s π , j ( x ) s_{\pi,j}(\boldsymbol{x}) s π , j ( x ) .
Step 2: Successive uses of the partitioned Mecke identity. For fixed realizations of all other coordinates, the Vandermonde square is a nonnegative measurable function of N y ( j ) N_y^{(j)} N y ( j ) . Applying Lemma 2.10 to that function and ( J y ( j ) ) β j (J_y^{(j)})_{\beta_j} ( J y ( j ) ) β j , successively for j = 1 , … , r j=1,\ldots,r j = 1 , … , r , we have
P β , z ( y ) = ∑ π ∫ ( 0 , ∞ ) κ π w π ( x ) E [ Δ ( J y + z + s π ( x ) ) 2 ] ∏ ( j , B ) ∈ B π ρ y ( d x j , B ) = ∑ π ∫ ( 0 , ∞ ) κ π w π ( x ) V z + s π ( x ) ( y ) ∏ ( j , B ) ∈ B π ρ y ( d x j , B ) . (81) \begin{aligned}
P_{\beta,\boldsymbol{z}}(\boldsymbol{y})
&= \sum_{\pi}\int_{(0,\infty)^{\kappa_{\pi}}} w_{\pi}(\boldsymbol{x})\mathbb{E}\left[\Delta(J_y+\boldsymbol{z}+s_{\pi}(\boldsymbol{x}))^2\right]\prod_{(j,B)\in\mathcal{B}_{\pi}}\rho_y(\mathrm{d}x_{j,B}) \\
&= \sum_{\pi}\int_{(0,\infty)^{\kappa_{\pi}}} w_{\pi}(\boldsymbol{x})V_{\boldsymbol{z}+s_{\pi}(\boldsymbol{x})}(\boldsymbol{y})\prod_{(j,B)\in\mathcal{B}_{\pi}}\rho_y(\mathrm{d}x_{j,B}). \tag*{(81)}
\end{aligned} P β , z ( y ) = π ∑ ∫ ( 0 , ∞ ) κ π w π ( x ) E [ Δ ( J y + z + s π ( x ) ) 2 ] ( j , B ) ∈ B π ∏ ρ y ( d x j , B ) = π ∑ ∫ ( 0 , ∞ ) κ π w π ( x ) V z + s π ( x ) ( y ) ( j , B ) ∈ B π ∏ ρ y ( d x j , B ) . (81)
We may interchange the successive integrations by Tonelli’s theorem, because all the integrands in this step are nonnegative.
Step 3: An explicit nonnegative coefficient formula. For a color assignment c : B π → [ m ] c:\mathcal{B}_{\pi}\to[m] c : B π → [ m ] , define
n ℓ ( c ) : = ∣ { ( j , B ) ∈ B π : c ( j , B ) = ℓ } ∣ , n ( c ) : = ( n 1 ( c ) , … , n m ( c ) ) . (82) n_{\ell}(c):=\left|\{(j,B)\in\mathcal{B}_{\pi}:c(j,B)=\ell\}\right|,
\qquad
n(c):=\left(n_1(c),\ldots,n_m(c)\right). \tag*{(82)} n ℓ ( c ) := ∣ {( j , B ) ∈ B π : c ( j , B ) = ℓ } ∣ , n ( c ) := ( n 1 ( c ) , … , n m ( c ) ) . (82)
Expanding the product of measure functions, we obtain
∏ ( j , B ) ∈ B π ρ y ( d x j , B ) = ∑ c : B π → [ m ] y n ( c ) ∏ ( j , B ) ∈ B π ρ c ( j , B ) ( d x j , B ) . (83) \prod_{(j,B)\in\mathcal{B}_{\pi}}\rho_y(\mathrm{d}x_{j,B})
=
\sum_{c:\mathcal{B}_{\pi}\to[m]}y^{n(c)}
\prod_{(j,B)\in\mathcal{B}_{\pi}}\rho_{c(j,B)}(\mathrm{d}x_{j,B}). \tag*{(83)} ( j , B ) ∈ B π ∏ ρ y ( d x j , B ) = c : B π → [ m ] ∑ y n ( c ) ( j , B ) ∈ B π ∏ ρ c ( j , B ) ( d x j , B ) . (83)
By Claim 2 and its proof, we may write
V u ( y ) = ∑ ∣ ξ ∣ ≤ D v ξ ( u ) y ξ , v ξ ( u ) ≥ 0 ( u ∈ [ 0 , ∞ ) r ) . (84) V_{\boldsymbol{u}}(\boldsymbol{y})=\sum_{|\xi|\le D}v_{\xi}(\boldsymbol{u})\boldsymbol{y}^{\xi},
\qquad
v_{\xi}(\boldsymbol{u})\ge0
\quad(\boldsymbol{u}\in[0,\infty)^r). \tag*{(84)} V u ( y ) = ∣ ξ ∣ ≤ D ∑ v ξ ( u ) y ξ , v ξ ( u ) ≥ 0 ( u ∈ [ 0 , ∞ ) r ) . (84)
Each v ξ ( u ) v_{\xi}(\boldsymbol{u}) v ξ ( u ) is a polynomial in u \boldsymbol{u} u of degree at most 2 D 2D 2 D . For fixed z \boldsymbol{z} z , the inserted jumps range over a compact set, so these polynomials and the weights w π w_{\pi} w π are bounded on the integration domains. Since every ρ ℓ \rho_{\ell} ρ ℓ is finite, all coefficient integrals are finite. We may therefore extract coefficients in the finite polynomial expansions. Combining (81) –(84) , we obtain, for ν ∈ N 0 m \nu\in\mathbb{N}_0^m ν ∈ N 0 m ,
[ y ν ] P β , z ( y ) = ∑ π ∑ c : B π → [ m ] n ( c ) ≤ ν ∫ ( 0 , ∞ ) κ π w π ( x ) v ν − n ( c ) ( z + s π ( x ) ) ∏ ( j , B ) ∈ B π ρ c ( j , B ) ( d x j , B ) ≥ 0. (85) [\boldsymbol{y}^{\nu}]P_{\beta,\boldsymbol{z}}(\boldsymbol{y})
=
\sum_{\pi}\sum_{\substack{c:\mathcal{B}_{\pi}\to[m]\\ n(c)\le\nu}}
\int_{(0,\infty)^{\kappa_{\pi}}}
w_{\pi}(\boldsymbol{x})v_{\nu-n(c)}\left(\boldsymbol{z}+s_{\pi}(\boldsymbol{x})\right)
\prod_{(j,B)\in\mathcal{B}_{\pi}}\rho_{c(j,B)}(\mathrm{d}x_{j,B})
\ge0. \tag*{(85)} [ y ν ] P β , z ( y ) = π ∑ c : B π → [ m ] n ( c ) ≤ ν ∑ ∫ ( 0 , ∞ ) κ π w π ( x ) v ν − n ( c ) ( z + s π ( x ) ) ( j , B ) ∈ B π ∏ ρ c ( j , B ) ( d x j , B ) ≥ 0. (85)
Here we compare multi-indices componentwise and set v ξ = 0 v_{\xi}=0 v ξ = 0 for ∣ ξ ∣ > D |\xi|>D ∣ ξ ∣ > D . Every integrand and every measure in the last expression is nonnegative, which proves the coefficientwise assertion. There are only finitely many partitions and colors, and the degree is at most D + ∣ β ∣ D+|\beta| D + ∣ β ∣ . Thus the expansions are polynomial identities; their validity on ( 0 , ∞ ) m (0,\infty)^{m} ( 0 , ∞ ) m also gives validity on the boundary. When k π = 0 k_{\pi}=0 k π = 0 , we use the empty-product convention, with n ( c ) = 0 n(c)=0 n ( c ) = 0 and s π = 0 s_{\pi}=0 s π = 0 .
It remains to justify integration against the initial measure and to remove the restriction on the jump measures. For general measures ρ ℓ \rho_{\ell} ρ ℓ in (37) , define, for integers L ≥ 1 L\ge1 L ≥ 1 ,
ρ ℓ , L ( d x ) : = 1 [ 1 / L , L ] ( x ) ρ ℓ ( d x ) , κ n , L ( ℓ ) : = b ℓ 1 { n = 1 } + ∫ 1 / L L x n ρ ℓ ( d x ) ( n ≥ 1 ) . (86) \rho_{\ell,L}(\mathrm{d}x):=\mathbf{1}_{[1/L,L]}(x)\rho_{\ell}(\mathrm{d}x),\qquad\kappa_{n,L}^{(\ell)}:=b_{\ell}\mathbf{1}_{\{n=1\}}+\int_{1/L}^{L}x^{n}\rho_{\ell}(\mathrm{d}x)\qquad(n\ge1). \tag*{(86)} ρ ℓ , L ( d x ) := 1 [ 1/ L , L ] ( x ) ρ ℓ ( d x ) , κ n , L ( ℓ ) := b ℓ 1 { n = 1 } + ∫ 1/ L L x n ρ ℓ ( d x ) ( n ≥ 1 ) . (86)
Let M n , L M_{n,L} M n , L be given by (29) with κ n ( ℓ ) \kappa_{n}^{(\ell)} κ n ( ℓ ) replaced by κ n , L ( ℓ ) \kappa_{n,L}^{(\ell)} κ n , L ( ℓ ) , and put
F n , L ( y ) : = ∑ k = 0 n ( n k ) a n − k M k , L ( y ) , D I , J , L μ ( y ) : = det [ F i a + j b , L ( y ) ] a , b = 1 r . (87) F_{n,L}(\boldsymbol{y}):=\sum_{k=0}^{n}\binom{n}{k}a_{n-k}M_{k,L}(\boldsymbol{y}),\qquad D_{I,J,L}^{\mu}(\boldsymbol{y}):=\det\left[F_{i_a+j_b,L}(\boldsymbol{y})\right]_{a,b=1}^{r}. \tag*{(87)} F n , L ( y ) := k = 0 ∑ n ( k n ) a n − k M k , L ( y ) , D I , J , L μ ( y ) := det [ F i a + j b , L ( y ) ] a , b = 1 r . (87)
We now justify the coefficientwise limit as the jump truncation is removed. In the following claim, we also verify finiteness and nonnegativity after integration against the initial measure.
Claim 4. For every α ∈ N 0 m \boldsymbol{\alpha}\in\mathbb{N}_{0}^{m} α ∈ N 0 m , we have
0 ≤ [ y α ] D I , J μ ( y ) = lim L → ∞ [ y α ] D I , J , L μ ( y ) < ∞ . (88) 0\le[\boldsymbol{y}^{\boldsymbol{\alpha}}]D_{I,J}^{\mu}(\boldsymbol{y})=\lim_{L\to\infty}[\boldsymbol{y}^{\boldsymbol{\alpha}}]D_{I,J,L}^{\mu}(\boldsymbol{y})<\infty. \tag*{(88)} 0 ≤ [ y α ] D I , J μ ( y ) = L → ∞ lim [ y α ] D I , J , L μ ( y ) < ∞. (88)
Proof. Step 1: Polynomial bounds and integration against μ \mu μ . First assume that the jump measures are finite and compactly supported. For β ∈ N 0 r \boldsymbol{\beta}\in\mathbb{N}_{0}^{r} β ∈ N 0 r and ν ∈ N 0 m \boldsymbol{\nu}\in\mathbb{N}_{0}^{m} ν ∈ N 0 m , write
p β , ν ( z ) : = [ y ν ] P β , z ( y ) ≥ 0 ( z ∈ [ 0 , ∞ ) r ) . (89) p_{\boldsymbol{\beta},\boldsymbol{\nu}}(\boldsymbol{z}):=[\boldsymbol{y}^{\boldsymbol{\nu}}]P_{\boldsymbol{\beta},\boldsymbol{z}}(\boldsymbol{y})\ge0\qquad(\boldsymbol{z}\in[0,\infty)^{r}). \tag*{(89)} p β , ν ( z ) := [ y ν ] P β , z ( y ) ≥ 0 ( z ∈ [ 0 , ∞ ) r ) . (89)
In (85) , we integrate translates of the polynomials v ξ v_{\xi} v ξ , whose degrees are at most 2 D 2D 2 D , over bounded jump domains with finite measures. Consequently, p β , ν ( z ) p_{\boldsymbol{\beta},\boldsymbol{\nu}}(\boldsymbol{z}) p β , ν ( z ) is a polynomial in z \boldsymbol{z} z of degree at most 2 D 2D 2 D . For a constant C C C depending on the fixed indices and jump measures, we therefore have
0 ≤ z γ p β , ν ( z ) ≤ C ( 1 + z 1 + ⋯ + z r ) 2 D + ∣ γ ∣ ( z ∈ [ 0 , ∞ ) r ) . (90) 0\le z^{\boldsymbol{\gamma}}p_{\boldsymbol{\beta},\boldsymbol{\nu}}(\boldsymbol{z})\le C(1+z_{1}+\cdots+z_{r})^{2D+|\boldsymbol{\gamma}|}\qquad(\boldsymbol{z}\in[0,\infty)^{r}). \tag*{(90)} 0 ≤ z γ p β , ν ( z ) ≤ C ( 1 + z 1 + ⋯ + z r ) 2 D + ∣ γ ∣ ( z ∈ [ 0 , ∞ ) r ) . (90)
To verify integrability of this bound, let q ∈ N 0 q\in\mathbb{N}_{0} q ∈ N 0 . Applying the multinomial formula to ( 1 + z 1 + ⋯ + z r ) q (1+z_{1}+\cdots+z_{r})^{q} ( 1 + z 1 + ⋯ + z r ) q , we have
∫ [ 0 , ∞ ) r ( 1 + z 1 + ⋯ + z r ) q ∏ j = 1 r μ ( d z j ) = ∑ δ 0 + ⋯ + δ r = q q ! δ 0 ! ⋯ δ r ! ∏ j = 1 r ( ∫ 0 ∞ z j δ j μ ( d z j ) ) = ∑ δ 0 + ⋯ + δ r = q q ! δ 0 ! ⋯ δ r ! ∏ j = 1 r a δ j < ∞ . (91) \begin{aligned}
\int_{[0,\infty)^{r}}(1+z_{1}+\cdots+z_{r})^{q}\prod_{j=1}^{r}\mu(\mathrm{d}z_{j})
=\sum_{\delta_{0}+\cdots+\delta_{r}=q}\frac{q!}{\delta_{0}!\cdots\delta_{r}!}\prod_{j=1}^{r}\left(\int_{0}^{\infty}z_{j}^{\delta_{j}}\mu(\mathrm{d}z_{j})\right) \\
&=\sum_{\delta_{0}+\cdots+\delta_{r}=q}\frac{q!}{\delta_{0}!\cdots\delta_{r}!}\prod_{j=1}^{r}a_{\delta_{j}}<\infty. \tag*{(91)}
\end{aligned} ∫ [ 0 , ∞ ) r ( 1 + z 1 + ⋯ + z r ) q j = 1 ∏ r μ ( d z j ) = δ 0 + ⋯ + δ r = q ∑ δ 0 ! ⋯ δ r ! q ! j = 1 ∏ r ( ∫ 0 ∞ z j δ j μ ( d z j ) ) = δ 0 + ⋯ + δ r = q ∑ δ 0 ! ⋯ δ r ! q ! j = 1 ∏ r a δ j < ∞. (91)
Using q = 2 D + ∣ γ ∣ q=2D+|\gamma| q = 2 D + ∣ γ ∣ , we obtain
0 ≤ ∫ [ 0 , ∞ ) r z γ p β , ν ( z ) ∏ j = 1 r μ ( d z j ) < ∞ . (92) 0\leq\int_{[0,\infty)^r}z^\gamma p_{\beta,\nu}(z)\prod_{j=1}^{r}\mu(\mathrm{d}z_j)<\infty. \tag*{(92)} 0 ≤ ∫ [ 0 , ∞ ) r z γ p β , ν ( z ) j = 1 ∏ r μ ( d z j ) < ∞. (92)
Thus we may integrate the finite polynomial expansion coefficient by coefficient. In particular, no compact-support assumption on μ \mu μ is needed.
Step 2: Nonnegative coefficients of the minor in the compact case. Recall b ( y ) = ∑ ℓ = 1 m b ℓ y ℓ b(\mathbf{y})=\sum_{\ell=1}^{m}b_\ell y_\ell b ( y ) = ∑ ℓ = 1 m b ℓ y ℓ , with b ℓ ≥ 0 b_\ell\geq0 b ℓ ≥ 0 . For h ∈ N 0 h\in\mathbb{N}_0 h ∈ N 0 , applying the multinomial formula to b ( y ) h b(\mathbf{y})^h b ( y ) h , we have
b ( y ) h = ∑ ∣ ω ∣ = h h ! ω ! ( ∏ ℓ = 1 m b ℓ ω ℓ ) y ω . (93) b(\mathbf{y})^h=\sum_{|\boldsymbol{\omega}|=h}\frac{h!}{\boldsymbol{\omega}!}\left(\prod_{\ell=1}^{m}b_\ell^{\omega_\ell}\right)\mathbf{y}^{\boldsymbol{\omega}}. \tag*{(93)} b ( y ) h = ∣ ω ∣ = h ∑ ω ! h ! ( ℓ = 1 ∏ m b ℓ ω ℓ ) y ω . (93)
Here and below we interpret a zeroth power as one, including 0 0 0^0 0 0 in these polynomial coefficient formulas. Combining (75) , (92) , and (93) , we obtain the explicit formula
[ y α ] D I , J μ ( y ) = 1 r ! ∑ τ c τ ∑ β + γ + δ = τ τ ! β ! γ ! δ ! × ∑ ∣ ω ∣ = ∣ δ ∣ ω ≤ α ∣ δ ∣ ! ω ! ( ∏ ℓ = 1 m b ℓ ω ℓ ) [\mathbf{y}^{\alpha}]D^\mu_{I,J}(\mathbf{y})=\frac{1}{r!}\sum_{\tau}c_\tau\sum_{\beta+\gamma+\delta=\tau}\frac{\tau!}{\beta!\gamma!\delta!}\times\sum_{\substack{|\omega|=\lvert\delta\rvert\\ \omega\leq\alpha}}\frac{|\delta|!}{\omega!}\left(\prod_{\ell=1}^{m}b_\ell^{\omega_\ell}\right) [ y α ] D I , J μ ( y ) = r ! 1 τ ∑ c τ β + γ + δ = τ ∑ β ! γ ! δ ! τ ! × ∣ ω ∣ = ∣ δ ∣ ω ≤ α ∑ ω ! ∣ δ ∣ ! ( ℓ = 1 ∏ m b ℓ ω ℓ )
× ∫ [ 0 , ∞ ) r z γ p β , α − ω ( z ) ∏ j = 1 r μ ( d z j ) ≥ 0. (94) {}\times\int_{[0,\infty)^r}z^\gamma p_{\beta,\alpha-\omega}(z)\prod_{j=1}^{r}\mu(\mathrm{d}z_j)\geq0. \tag*{(94)} × ∫ [ 0 , ∞ ) r z γ p β , α − ω ( z ) j = 1 ∏ r μ ( d z j ) ≥ 0. (94)
We used τ , β , γ , δ ∈ N 0 r \tau,\beta,\gamma,\delta\in\mathbb{N}_0^r τ , β , γ , δ ∈ N 0 r and α , ω ∈ N 0 m \alpha,\omega\in\mathbb{N}_0^m α , ω ∈ N 0 m to distinguish the two kinds of multi-indices. Every sum is finite and every factor is nonnegative. We have therefore proved the desired assertion for finite compactly supported jump measures.
Step 3: Finite compact truncations of general jump measures. Return to the general representation (37) . For ρ ℓ , L \rho_{\ell,L} ρ ℓ , L in (86) , we have
ρ ℓ , L ( ( 0 , ∞ ) ) = ρ ℓ ( [ 1 / L , L ] ) ≤ L ∫ 0 1 x ρ ℓ ( d x ) + ρ ℓ ( ( 1 , ∞ ) ) < ∞ . (95) \rho_{\ell,L}((0,\infty))=\rho_\ell([1/L,L])\leq L\int_0^1x\,\rho_\ell(\mathrm{d}x)+\rho_\ell((1,\infty))<\infty. \tag*{(95)} ρ ℓ , L (( 0 , ∞ )) = ρ ℓ ([ 1/ L , L ]) ≤ L ∫ 0 1 x ρ ℓ ( d x ) + ρ ℓ (( 1 , ∞ )) < ∞. (95)
We used 1 ≤ L x 1\leq Lx 1 ≤ Lx on [ 1 / L , 1 ] [1/L,1] [ 1/ L , 1 ] and (41) . The support lies in [ 1 / L , L ] [1/L,L] [ 1/ L , L ] , so the truncated jump measures satisfy the assumptions of Steps 1 and 2. More explicitly, the nonnegative measure
ν ℓ , L ( d x ) : = b ℓ δ 0 ( d x ) + x ρ ℓ , L ( d x ) (96) \nu_{\ell,L}(\mathrm{d}x):=b_\ell\delta_0(\mathrm{d}x)+x\rho_{\ell,L}(\mathrm{d}x) \tag*{(96)} ν ℓ , L ( d x ) := b ℓ δ 0 ( d x ) + x ρ ℓ , L ( d x ) (96)
has finite mass and compact support, and
∫ 0 ∞ x n ν ℓ , L ( d x ) = b ℓ 1 { n = 0 } + ∫ 1 / L L x n + 1 ρ ℓ ( d x ) = κ n + 1 , L ( ℓ ) ( n ≥ 0 ) . (97) \int_0^\infty x^n\nu_{\ell,L}(\mathrm{d}x)=b_\ell\mathbf{1}_{\{n=0\}}+\int_{1/L}^{L}x^{n+1}\rho_\ell(\mathrm{d}x)=\kappa^{(\ell)}_{n+1,L}\qquad(n\geq0). \tag*{(97)} ∫ 0 ∞ x n ν ℓ , L ( d x ) = b ℓ 1 { n = 0 } + ∫ 1/ L L x n + 1 ρ ℓ ( d x ) = κ n + 1 , L ( ℓ ) ( n ≥ 0 ) . (97)
Applying the compact-case conclusion (94) to these truncated result, we have
[ y α ] D I , J , L μ ( y ) ≥ 0 ( L ≥ 1 ) . (98) [\mathbf{y}^{\alpha}]D^\mu_{I,J,L}(\mathbf{y})\geq0\qquad(L\geq1). \tag*{(98)} [ y α ] D I , J , L μ ( y ) ≥ 0 ( L ≥ 1 ) . (98)
Step 4: Coefficientwise convergence of moments and minors. Since [ 1 / L , L ] [1/L,L] [ 1/ L , L ] increases to ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , applying the monotone convergence theorem to the nonnegative functions x n 1 [ 1 / L , L ] ( x ) x^n\mathbf{1}_{[1/L,L]}(x) x n 1 [ 1/ L , L ] ( x ) , we have
κ n , L ( ℓ ) ⟶ b ℓ 1 { n = 1 } + ∫ 0 ∞ x n ρ ℓ ( d x ) = κ n ( ℓ ) < ∞ . (99) \kappa_{n,L}^{(\ell)} \longrightarrow b_\ell\mathbf{1}_{\{n=1\}}+\int_0^\infty x^n\,\rho_\ell(\mathrm{d}x)=\kappa_n^{(\ell)}<\infty. \tag*{(99)} κ n , L ( ℓ ) ⟶ b ℓ 1 { n = 1 } + ∫ 0 ∞ x n ρ ℓ ( d x ) = κ n ( ℓ ) < ∞. (99)
For each fixed n n n , the set-partition formula (29) expresses every coefficient of M n , L M_{n,L} M n , L as a finite polynomial in the cumulants with indices at most n n n . Using that formula and (87) , we obtain
[ y α ] M n , L ( y ) ⟶ [ y α ] M n ( y ) , [ y α ] F n , L ( y ) ⟶ [ y α ] F n ( y ) . (100) [y^\alpha]M_{n,L}(y)\longrightarrow[y^\alpha]M_n(y),\qquad[y^\alpha]F_{n,L}(y)\longrightarrow[y^\alpha]F_n(y). \tag*{(100)} [ y α ] M n , L ( y ) ⟶ [ y α ] M n ( y ) , [ y α ] F n , L ( y ) ⟶ [ y α ] F n ( y ) . (100)
To make the last passage to a determinant explicit, expand by permutations and then extract the coefficient of y α y^\alpha y α :
[ y α ] D I , J , L μ = ∑ σ ∈ S r sgn ( σ ) ∑ α 1 + ⋯ + α r = α ∏ a = 1 r [ y α a ] F i a + j σ ( a ) , L ⟶ ∑ σ ∈ S r sgn ( σ ) ∑ α 1 + ⋯ + α r = α ∏ a = 1 r [ y α a ] F i a + j σ ( a ) = [ y α ] D I , J μ . (101) \begin{aligned}
[y^\alpha]D^\mu_{I,J,L}
&=\sum_{\sigma\in S_r}\operatorname{sgn}(\sigma)
\sum_{\alpha_1+\cdots+\alpha_r=\alpha}
\prod_{a=1}^{r}[y^{\alpha_a}]F_{i_a+j_{\sigma(a)},L}\\
&\longrightarrow\sum_{\sigma\in S_r}\operatorname{sgn}(\sigma)
\sum_{\alpha_1+\cdots+\alpha_r=\alpha}
\prod_{a=1}^{r}[y^{\alpha_a}]F_{i_a+j_{\sigma(a)}}\\
&=[y^\alpha]D^\mu_{I,J}. \tag*{(101)}
\end{aligned} [ y α ] D I , J , L μ = σ ∈ S r ∑ sgn ( σ ) α 1 + ⋯ + α r = α ∑ a = 1 ∏ r [ y α a ] F i a + j σ ( a ) , L ⟶ σ ∈ S r ∑ sgn ( σ ) α 1 + ⋯ + α r = α ∑ a = 1 ∏ r [ y α a ] F i a + j σ ( a ) = [ y α ] D I , J μ . (101)
All sums in this calculation are finite. Moreover, deg F n , L ≤ n \deg F_{n,L}\le n deg F n , L ≤ n and deg D I , J , L μ ≤ ∑ a i a + ∑ a j a \deg D^\mu_{I,J,L}\le\sum_a i_a+\sum_a j_a deg D I , J , L μ ≤ ∑ a i a + ∑ a j a , independently of L L L . Thus we obtain a finite limiting coefficient, and from (98) we retain its nonnegativity.
For the fixed index lists I , J I,J I , J , put N = i r + j r N=i_r+j_r N = i r + j r . Only the initial moments a 0 , … , a N a_0,\ldots,a_N a 0 , … , a N and the cumulants κ n ( ℓ ) \kappa_n^{(\ell)} κ n ( ℓ ) with 1 ≤ n ≤ N 1\le n\le N 1 ≤ n ≤ N occur in this calculation. Each coefficient of the minor is a polynomial in these finitely many real numbers. Thus we use monotone convergence for the nonnegative cumulant integrals, followed by continuity of finite polynomial expressions for the minor coefficients. We require no monotonicity of determinant coefficients, no interchange of an infinite formal power series with an integral, and no convergence of a moment generating function. This proves (88) .
Since I , J I,J I , J , the minor order r r r , and the initial Stieltjes moment sequence are arbitrary, Claim 4 yields (iv).
Necessity and completion of the equivalence of Theorem 1.7
We next recover the shifted cumulant minors from the lowest nonzero terms of the moment minors. This will prove necessity using only numerical Stieltjes moment sequences in (iii).
Completion of the proof of Theorem 1.7 . We have proved (i)⇒ \Rightarrow ⇒ (iv). For the initial measure δ 0 \delta_0 δ 0 , the moments are a 0 = 1 a_0=1 a 0 = 1 and a n = 0 a_n=0 a n = 0 for n ≥ 1 n\geq1 n ≥ 1 , so (13) gives F n = M n F_n=M_n F n = M n . Hence (iv)⇒ \Rightarrow ⇒ (ii) and (v)⇒ \Rightarrow ⇒ (iii). Under (iv) or (ii), evaluating every nonnegative polynomial minor at y ∈ [ 0 , ∞ ) m \boldsymbol{y}\in[0,\infty)^m y ∈ [ 0 , ∞ ) m gives a numerical nonnegative minor. Applying Lemma 2.1(ii) to the resulting numerical Hankel matrices, we have (iv)⇒ \Rightarrow ⇒ (v) and (ii)⇒ \Rightarrow ⇒ (iii). It remains to prove (iii)⇒ \Rightarrow ⇒ (i).
Step 1: Restriction to one color and the one-block term. Assume that (iii) holds, fix ℓ ∈ { 1 , … , m } \ell\in\{1,\ldots,m\} ℓ ∈ { 1 , … , m } , and let e ℓ \boldsymbol{e}_\ell e ℓ be the ℓ \ell ℓ -th coordinate vector. Define
M n [ ℓ ] ( u ) : = M n ( u e ℓ ) , ∑ n ≥ 0 M n [ ℓ ] ( u ) t n n ! = e u κ ℓ ( t ) . (102) M_n^{[\ell]}(u):=M_n(u\boldsymbol{e}_\ell),\qquad\sum_{n\geq0}M_n^{[\ell]}(u)\frac{t^n}{n!}=e^{u\kappa_\ell(t)}. \tag*{(102)} M n [ ℓ ] ( u ) := M n ( u e ℓ ) , n ≥ 0 ∑ M n [ ℓ ] ( u ) n ! t n = e u κ ℓ ( t ) . (102)
Substituting y = u e ℓ \boldsymbol{y}=u\boldsymbol{e}_\ell y = u e ℓ into the finite partition formula (29) , we obtain, for n ≥ 1 n\geq1 n ≥ 1 ,
M n [ ℓ ] ( u ) = ∑ π ∈ P n u ∣ π ∣ ∏ B ∈ π κ ∣ B ∣ ( ℓ ) = u κ n ( ℓ ) + ∑ π ∈ P n ∣ π ∣ ≥ 2 u ∣ π ∣ ∏ B ∈ π κ ∣ B ∣ ( ℓ ) . (103) \begin{aligned}
M_n^{[\ell]}(u)=\sum_{\pi\in\mathcal{P}_n}u^{|\pi|}\prod_{B\in\pi}\kappa_{|B|}^{(\ell)} \\
&= u\kappa_n^{(\ell)}+\sum_{\substack{\pi\in\mathcal{P}_n\\|\pi|\geq2}}u^{|\pi|}\prod_{B\in\pi}\kappa_{|B|}^{(\ell)}. \tag*{(103)}
\end{aligned} M n [ ℓ ] ( u ) = π ∈ P n ∑ u ∣ π ∣ B ∈ π ∏ κ ∣ B ∣ ( ℓ ) = u κ n ( ℓ ) + π ∈ P n ∣ π ∣ ≥ 2 ∑ u ∣ π ∣ B ∈ π ∏ κ ∣ B ∣ ( ℓ ) . (103)
There is exactly one partition with one block, namely { [ n ] } \{[n]\} {[ n ]} . Every other partition contributes a power of u u u at least two. Thus we may write
M n [ ℓ ] ( u ) = u κ n ( ℓ ) + u 2 R n , ℓ ( u ) , R n , ℓ ( u ) : = ∑ π ∈ P n ∣ π ∣ ≥ 2 u ∣ π ∣ − 2 ∏ B ∈ π κ ∣ B ∣ ( ℓ ) ∈ R [ u ] . (104) M_n^{[\ell]}(u)=u\kappa_n^{(\ell)}+u^2R_{n,\ell}(u),\qquad R_{n,\ell}(u):=\sum_{\substack{\pi\in\mathcal{P}_n\\|\pi|\geq2}}u^{|\pi|-2}\prod_{B\in\pi}\kappa_{|B|}^{(\ell)}\in\mathbb{R}[u]. \tag*{(104)} M n [ ℓ ] ( u ) = u κ n ( ℓ ) + u 2 R n , ℓ ( u ) , R n , ℓ ( u ) := π ∈ P n ∣ π ∣ ≥ 2 ∑ u ∣ π ∣ − 2 B ∈ π ∏ κ ∣ B ∣ ( ℓ ) ∈ R [ u ] . (104)
For n = 1 n=1 n = 1 , this is an empty sum and R 1 , ℓ = 0 R_{1,\ell}=0 R 1 , ℓ = 0 . We assume no sign condition on the coefficients of R n , ℓ R_{n,\ell} R n , ℓ .
Step 2: The lowest-degree term of a shifted minor. Take arbitrary strictly increasing nonnegative integer lists I = ( i 1 , … , i r ) I=(i_1,\ldots,i_r) I = ( i 1 , … , i r ) and J = ( j 1 , … , j r ) J=(j_1,\ldots,j_r) J = ( j 1 , … , j r ) , and set
Q I , J [ ℓ ] ( u ) : = det [ M i a + j b + 1 [ ℓ ] ( u ) ] a , b = 1 r . (105) Q_{I,J}^{[\ell]}(u):=\det\left[M_{i_a+j_b+1}^{[\ell]}(u)\right]_{a,b=1}^r. \tag*{(105)} Q I , J [ ℓ ] ( u ) := det [ M i a + j b + 1 [ ℓ ] ( u ) ] a , b = 1 r . (105)
This is a minor of the ordinary moment Hankel matrix with row list I I I and column list ( j 1 + 1 , … , j r + 1 ) (j_1+1,\ldots,j_r+1) ( j 1 + 1 , … , j r + 1 ) . Under (iii), ( M n [ ℓ ] ( u ) ) n ≥ 0 (M_n^{[\ell]}(u))_{n\geq0} ( M n [ ℓ ] ( u ) ) n ≥ 0 is a Stieltjes moment sequence for every u > 0 u>0 u > 0 . Applying Lemma 2.1(ii) to this sequence, we have
Q I , J [ ℓ ] ( u ) ≥ 0 ( u > 0 ) . (106) Q_{I,J}^{[\ell]}(u)\geq0\qquad(u>0). \tag*{(106)} Q I , J [ ℓ ] ( u ) ≥ 0 ( u > 0 ) . (106)
Every index i a + j b + 1 i_a+j_b+1 i a + j b + 1 is positive, so we may use (104) for every entry. Factoring u u u from each row, and then expanding the determinant by permutations, we obtain
u − r Q I , J [ ℓ ] ( u ) = det [ κ i a + j b + 1 ( ℓ ) + u R i a + j b + 1 , ℓ ( u ) ] a , b = 1 r = ∑ σ ∈ S r sgn ( σ ) ∏ a = 1 r ( κ i a + j σ ( a ) + 1 ( ℓ ) + u R i a + j σ ( a ) + 1 , ℓ ( u ) ) . (107) \begin{aligned}
u^{-r}Q_{I,J}^{[\ell]}(u)=\det\left[\kappa_{i_a+j_b+1}^{(\ell)}+uR_{i_a+j_b+1,\ell}(u)\right]_{a,b=1}^r \\
&=\sum_{\sigma\in S_r}\operatorname{sgn}(\sigma)\prod_{a=1}^r\left(\kappa_{i_a+j_{\sigma(a)}+1}^{(\ell)}+uR_{i_a+j_{\sigma(a)}+1,\ell}(u)\right). \tag*{(107)}
\end{aligned} u − r Q I , J [ ℓ ] ( u ) = det [ κ i a + j b + 1 ( ℓ ) + u R i a + j b + 1 , ℓ ( u ) ] a , b = 1 r = σ ∈ S r ∑ sgn ( σ ) a = 1 ∏ r ( κ i a + j σ ( a ) + 1 ( ℓ ) + u R i a + j σ ( a ) + 1 , ℓ ( u ) ) . (107)
In each product, choosing only cumulant factors gives the constant term. Every other choice contributes at least one factor of u u u . In consequence, for a polynomial S I , J , ℓ ( u ) S_{I,J,\ell}(u) S I , J , ℓ ( u ) , we have
Q I , J [ ℓ ] ( u ) = u r det [ κ i a + j b + 1 ( ℓ ) ] a , b = 1 r + u r + 1 S I , J , ℓ ( u ) . (108) Q^{[\ell]}_{I,J}(u)=u^r\det\left[\kappa^{(\ell)}_{i_a+j_b+1}\right]_{a,b=1}^{r}+u^{r+1}S_{I,J,\ell}(u). \tag*{(108)} Q I , J [ ℓ ] ( u ) = u r det [ κ i a + j b + 1 ( ℓ ) ] a , b = 1 r + u r + 1 S I , J , ℓ ( u ) . (108)
Dividing by the positive number u r u^r u r , and then taking u → 0 + u\to0^+ u → 0 + in (107) , we obtain
det [ κ i a + j b + 1 ( ℓ ) ] a , b = 1 r = lim u → 0 + u − r Q I , J [ ℓ ] ( u ) ≥ 0. (109) \det\left[\kappa^{(\ell)}_{i_a+j_b+1}\right]_{a,b=1}^{r}=\lim_{u\to0^+}u^{-r}Q^{[\ell]}_{I,J}(u)\ge0. \tag*{(109)} det [ κ i a + j b + 1 ( ℓ ) ] a , b = 1 r = u → 0 + lim u − r Q I , J [ ℓ ] ( u ) ≥ 0. (109)
We use nonnegativity of the numerical values for u > 0 u>0 u > 0 , together with the fact that all terms of degree below r r r vanish; we do not assume coefficientwise nonnegativity in (iii).
Step 3: The Stieltjes criterion for the shifted cumulants. Put c n ( ℓ ) : = κ n + 1 ( ℓ ) c_n^{(\ell)}:=\kappa_{n+1}^{(\ell)} c n ( ℓ ) := κ n + 1 ( ℓ ) for n ≥ 0 n\ge0 n ≥ 0 . For every choice of I , J I,J I , J , we have
det [ c i a + j b ( ℓ ) ] a , b = 1 r = det [ κ i a + j b + 1 ( ℓ ) ] a , b = 1 r ≥ 0. (110) \det\left[c_{i_a+j_b}^{(\ell)}\right]_{a,b=1}^{r}=\det\left[\kappa_{i_a+j_b+1}^{(\ell)}\right]_{a,b=1}^{r}\ge0. \tag*{(110)} det [ c i a + j b ( ℓ ) ] a , b = 1 r = det [ κ i a + j b + 1 ( ℓ ) ] a , b = 1 r ≥ 0. (110)
Thus every minor of the infinite Hankel matrix [ c i + j ( ℓ ) ] i , j ≥ 0 \left[c_{i+j}^{(\ell)}\right]_{i,j\ge0} [ c i + j ( ℓ ) ] i , j ≥ 0 is nonnegative. Applying the total-positivity characterization in Lemma 2.1 (ii) to ( c n ( ℓ ) ) n ≥ 0 \left(c_n^{(\ell)}\right)_{n\ge0} ( c n ( ℓ ) ) n ≥ 0 , we have a nonnegative measure ν ℓ \nu_\ell ν ℓ on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that
κ n + 1 ( ℓ ) = ∫ 0 ∞ x n ν ℓ ( d x ) < ∞ ( n ≥ 0 ) . (111) \kappa_{n+1}^{(\ell)}=\int_0^\infty x^n\nu_\ell(\mathrm{d}x)<\infty\qquad(n\ge0). \tag*{(111)} κ n + 1 ( ℓ ) = ∫ 0 ∞ x n ν ℓ ( d x ) < ∞ ( n ≥ 0 ) . (111)
Since ℓ \ell ℓ was arbitrary, we obtain (i). Combining the implications in Step 1 with (iii)⇒ \Rightarrow ⇒ (i), we prove the equivalence of all five conditions. We used only finite polynomial identities and limits of finite determinants. This completes the proof.
Proof of the Hamburger lifting theorem
In this section, we will prove Theorem 1.8 stated in § 1.3 . The proof uses the Hamburger moment criterion in Lemma 2.1 and the Poisson Vandermonde expansion of Lemma 2.9 .
Proof of Theorem 1.8. We first derive the necessary moment condition from (iv). For sufficiency, we treat compactly supported jumps, compute the highest-degree component, and then approximate arbitrary twice-shifted moment measures. Finally, we use Gaussian initial measures to deduce (iv) from (ii), thereby completing the equivalence.
To recover the twice-shifted cumulants, we test the moment function on squares of polynomials divisible by the coordinate variable and let a single intensity tend to zero. In the following claim, we make this limiting argument explicit and recover the required nonnegative quadratic forms.
Claim 5. Under condition (iv), for each ℓ ∈ { 1 , … , m } \ell\in\{1,\ldots,m\} ℓ ∈ { 1 , … , m } , we have
∑ i , j = 0 d c i c j κ i + j + 2 ( ℓ ) ≥ 0 ( d ≥ 0 , ( c 0 , … , c d ) ∈ R d + 1 ) . (112) \sum_{i,j=0}^{d} c_i c_j \kappa_{i+j+2}^{(\ell)} \ge0 \qquad(d \ge0,\ (c_0,\ldots,c_d) \in\mathbb{R}^{d+1}). \tag*{(112)} i , j = 0 ∑ d c i c j κ i + j + 2 ( ℓ ) ≥ 0 ( d ≥ 0 , ( c 0 , … , c d ) ∈ R d + 1 ) . (112) In consequence, ( κ n + 2 ( ℓ ) ) n ≥ 0 (\kappa_{n+2}^{(\ell)})_{n\ge0} ( κ n + 2 ( ℓ ) ) n ≥ 0 is a Hamburger moment sequence.
Proof. For the implication (iv)⇒ \Rightarrow ⇒ (i), fix ℓ \ell ℓ and any polynomial p ( x ) = ∑ i = 0 d c i x i p(x)=\sum_{i=0}^{d}c_i x^i p ( x ) = ∑ i = 0 d c i x i . For every u > 0 u>0 u > 0 , applying the Hamburger moment property to the square of the product of the coordinate variable and this polynomial, we obtain
1 u ∑ i , j = 0 d c i c j M i + j + 2 ( u e ℓ ) ≥ 0 , (113) \frac{1}{u}\sum_{i,j=0}^{d} c_i c_j M_{i+j+2}(u e_\ell) \ge0, \tag*{(113)} u 1 i , j = 0 ∑ d c i c j M i + j + 2 ( u e ℓ ) ≥ 0 , (113)
since the numerator is the integral of x 2 p ( x ) 2 x^2p(x)^2 x 2 p ( x ) 2 against a nonnegative measure. For every n ≥ 1 n\ge1 n ≥ 1 , M n ( u e ℓ ) = u κ n ( ℓ ) + O ( u 2 ) M_n(u e_\ell)=u\kappa_n^{(\ell)}+O(u^2) M n ( u e ℓ ) = u κ n ( ℓ ) + O ( u 2 ) . Only finitely many moments occur in the quadratic form. Thus, letting u ↓ 0 u\downarrow0 u ↓ 0 and passing to the limit term by term, we obtain
∑ i , j = 0 d c i c j κ i + j + 2 ( ℓ ) ≥ 0. (114) \sum_{i,j=0}^{d} c_i c_j \kappa_{i+j+2}^{(\ell)} \ge0. \tag*{(114)} i , j = 0 ∑ d c i c j κ i + j + 2 ( ℓ ) ≥ 0. (114)
Applying the Hamburger moment theorem to these nonnegative quadratic forms, we obtain the representing measure required in (i). Since the quadratic forms involve only cumulants of order at least two, we impose no condition on the first cumulant.
We now turn to sufficiency. We begin with finite compactly supported jump measures, for which we can construct the output measure directly and then obtain coefficientwise positivity from an expansion in squares. We carry out this construction of a representing measure and the subsequent square-integral expansion in the following claim.
Claim 6. For finite compactly supported jump measures on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } ,
H r a ( y ) ∈ R ≥ 0 [ y ] , deg y H r a ≤ D , D = ( r 2 ) . (115) H_r^a(\boldsymbol{y})\in\mathbb{R}_{\ge0}[\boldsymbol{y}],\qquad\deg_{\boldsymbol{y}} H_r^a\le D,\qquad D=\binom{r}{2}. \tag*{(115)} H r a ( y ) ∈ R ≥ 0 [ y ] , deg y H r a ≤ D , D = ( 2 r ) . (115) Moreover, for each y ∈ [ 0 , ∞ ) m \boldsymbol{y}\in[0,\infty)^m y ∈ [ 0 , ∞ ) m , there is a nonnegative measure μ y \mu_{\boldsymbol{y}} μ y on R \mathbb{R} R such that
F n ( y ) = ∫ − ∞ ∞ x n μ y ( d x ) ( n ≥ 0 ) . (116) F_n(\boldsymbol{y})=\int_{-\infty}^{\infty}x^n\,\mu_{\boldsymbol{y}}(dx)\qquad(n\ge0). \tag*{(116)} F n ( y ) = ∫ − ∞ ∞ x n μ y ( d x ) ( n ≥ 0 ) . (116)
Proof. To prove (i)⇒ \Rightarrow ⇒ (ii) and (iii) in this case, take finite compactly supported jump measures ρ 1 , … , ρ m \rho_1,\ldots,\rho_m ρ 1 , … , ρ m on R ∖ { 0 } \mathbb{R}\setminus\{0\} R ∖ { 0 } . Let b 1 , … , b m b_1,\ldots,b_m b 1 , … , b m be arbitrary real drifts, and set
K ℓ ( t ) = b ℓ t + ∫ − ∞ ∞ ( e t x − 1 ) ρ ℓ ( d x ) . (117) K_\ell(t)=b_\ell t+\int_{-\infty}^{\infty}(e^{tx}-1)\rho_\ell(\mathrm{d}x). \tag*{(117)} K ℓ ( t ) = b ℓ t + ∫ − ∞ ∞ ( e t x − 1 ) ρ ℓ ( d x ) . (117)
For a numerical y ≥ 0 \mathbf{y}\ge0 y ≥ 0 , let J y J_{\mathbf{y}} J y be the compound-Poisson sum with intensity ρ y = ∑ ℓ y ℓ ρ ℓ \rho_{\mathbf{y}}=\sum_\ell y_\ell\rho_\ell ρ y = ∑ ℓ y ℓ ρ ℓ , and put b ( y ) = ∑ ℓ y ℓ b ℓ b(\mathbf{y})=\sum_\ell y_\ell b_\ell b ( y ) = ∑ ℓ y ℓ b ℓ . If μ \mu μ represents the initial Hamburger sequence, then
F n ( y ) = ∫ − ∞ ∞ E ( z + J y + b ( y ) ) n μ ( d z ) . (118) F_n(\mathbf{y})=\int_{-\infty}^{\infty}\mathbb{E}(z+J_{\mathbf{y}}+b(\mathbf{y}))^n\,\mu(\mathrm{d}z). \tag*{(118)} F n ( y ) = ∫ − ∞ ∞ E ( z + J y + b ( y ) ) n μ ( d z ) . (118)
By integrating the law of the translated compound-Poisson variable against the initial measure, we obtain a finite nonnegative measure on R \mathbb{R} R with these moments. This proves (iii) for compactly supported jumps.
To prove coefficientwise positivity, we next express each leading determinant as an integral of a squared Vandermonde. Let J y J_{\mathbf{y}} J y consist of r r r independent copies of J y J_{\mathbf{y}} J y . By Andréief’s identity (31) and the common-translation invariance of the Vandermonde, we obtain
H r a ( y ) = 1 r ! ∫ R r E [ ( Δ ( s + J y ) ) 2 ] ∏ j = 1 r μ ( d s j ) . (119) H_r^a(\mathbf{y})=\frac{1}{r!}\int_{\mathbb{R}^r}\mathbb{E}\left[(\Delta(\mathbf{s}+J_{\mathbf{y}}))^2\right]\prod_{j=1}^{r}\mu(\mathrm{d}s_j). \tag*{(119)} H r a ( y ) = r ! 1 ∫ R r E [ ( Δ ( s + J y ) ) 2 ] j = 1 ∏ r μ ( d s j ) . (119)
We have thus removed the common drift before expanding in the intensities. Applying the real-line version of Lemma 2.9 to Δ ( s + J y ) \Delta(\mathbf{s}+J_{\mathbf{y}}) Δ ( s + J y ) , with E = { 1 , … , r } × ( R ∖ { 0 } ) E=\{1,\ldots,r\}\times(\mathbb{R}\setminus\{0\}) E = { 1 , … , r } × ( R ∖ { 0 }) and Λ ℓ ( { b } × d x ) = ρ ℓ ( d x ) \Lambda_\ell(\{b\}\times\mathrm{d}x)=\rho_\ell(\mathrm{d}x) Λ ℓ ({ b } × d x ) = ρ ℓ ( d x ) , and then integrating each initial coordinate against μ \mu μ , we obtain
H r a ( y ) = 1 r ! ∑ k = 0 D 1 k ! ∫ R r ∫ E k ( D z 1 ⋯ D z k Δ ( s ) ) 2 ∏ h = 1 k Λ y ( d z h ) ∏ j = 1 r μ ( d s j ) . (120) H_r^a(\mathbf{y})=\frac{1}{r!}\sum_{k=0}^{D}\frac{1}{k!}\int_{\mathbb{R}^r}\int_{E^k}\left(\mathcal{D}_{z_1}\cdots\mathcal{D}_{z_k}\Delta(\mathbf{s})\right)^2\prod_{h=1}^{k}\Lambda_{\mathbf{y}}(\mathrm{d}z_h)\prod_{j=1}^{r}\mu(\mathrm{d}s_j). \tag*{(120)} H r a ( y ) = r ! 1 k = 0 ∑ D k ! 1 ∫ R r ∫ E k ( D z 1 ⋯ D z k Δ ( s ) ) 2 h = 1 ∏ k Λ y ( d z h ) j = 1 ∏ r μ ( d s j ) . (120)
Expanding the product of measure functions, we express each coefficient as an integral of a square against nonnegative measures. The difference kernels are independent of the indeterminates and polynomial in the initial positions and bounded jump sizes. By the absolute moment hypothesis on μ \mu μ , every coefficient integral is finite. We therefore obtain (ii), with degree at most D D D , as required.
Having established the degree bound in the compact-jump case, we now compute the highest-degree component. At this degree, the differences exhaust the degree of the Vandermonde, so the result depends only on the second jump moments. We now evaluate this component by counting the nonzero derivatives of the Vandermonde, as described in the following claim.
Claim 7. For finite compactly supported jump measures, we have
[ H r a ] D ( y ) = a 0 r ( ∏ j = 0 r − 1 j ! ) ( ∑ ℓ = 1 m y ℓ κ 2 ( ℓ ) ) D . (121) [H_r^a]_D(\mathbf{y})=a_0^r\left(\prod_{j=0}^{r-1}j!\right)\left(\sum_{\ell=1}^{m}y_\ell\kappa_2^{(\ell)}\right)^D. \tag*{(121)} [ H r a ] D ( y ) = a 0 r ( j = 0 ∏ r − 1 j ! ) ( ℓ = 1 ∑ m y ℓ κ 2 ( ℓ ) ) D . (121)
Proof. In the square-integral expansion, only the term k = D k=D k = D contributes to the degree-D D D part. Write ∂ j = ∂ / ∂ s j \partial_j=\partial/\partial s_j ∂ j = ∂ / ∂ s j . Since Δ ( s ) \Delta(s) Δ ( s ) is homogeneous of degree D D D , each successive difference lowers the degree by at least one. Consequently, only the linear term in each jump can contribute after all differences have been taken, and we obtain
D ( b 1 , x 1 ) ⋯ D ( b D , x D ) Δ = ( ∏ h = 1 D x h ) ∂ b 1 ⋯ ∂ b D Δ . (122) \mathcal{D}_{(b_1,x_1)}\cdots\mathcal{D}_{(b_D,x_D)}\Delta=\left(\prod_{h=1}^{D}x_h\right)\partial_{b_1}\cdots\partial_{b_D}\Delta. \tag*{(122)} D ( b 1 , x 1 ) ⋯ D ( b D , x D ) Δ = ( h = 1 ∏ D x h ) ∂ b 1 ⋯ ∂ b D Δ. (122)
On the right side, we obtain a constant derivative. To sum its square over the coordinate choices, note that in the monomial expansion of Δ \Delta Δ , the exponent vectors are precisely the r ! r! r ! permutations of ( 0 , 1 , … , r − 1 ) (0,1,\ldots,r-1) ( 0 , 1 , … , r − 1 ) . A nonzero derivative occurs only when the coordinate multiplicities agree with one of these vectors. For each such vector, the absolute value of the derivative is ∏ j = 0 r − 1 j ! \prod_{j=0}^{r-1}j! ∏ j = 0 r − 1 j ! , and the number of ordered derivative lists producing it is D ! / ∏ j = 0 r − 1 j ! D!/\prod_{j=0}^{r-1}j! D ! / ∏ j = 0 r − 1 j ! . In consequence,
∑ b 1 , … , b D = 1 r ( ∂ b 1 ⋯ ∂ b D Δ ) 2 = r ! D ! ∏ j = 0 r − 1 j ! . (123) \sum_{b_1,\ldots,b_D=1}^{r}\left(\partial_{b_1}\cdots\partial_{b_D}\Delta\right)^2=r!D!\prod_{j=0}^{r-1}j!. \tag*{(123)} b 1 , … , b D = 1 ∑ r ( ∂ b 1 ⋯ ∂ b D Δ ) 2 = r ! D ! j = 0 ∏ r − 1 j ! . (123)
For each jump-square factor, integration gives ∫ − ∞ ∞ x 2 ρ y ( d x ) = v ( y ) \int_{-\infty}^{\infty}x^2\rho_y(\mathrm{d}x)=v(y) ∫ − ∞ ∞ x 2 ρ y ( d x ) = v ( y ) . Since the remaining kernel is constant in the initial positions, integration against the product initial measure gives a 0 r a_0^r a 0 r . Substituting these evaluations into (120) , we obtain (16) in the compact-jump case. □
It remains to remove the compact-jump assumption and to include a possible mass at zero in the twice-shifted moment measure. We approximate that mass by symmetric small jumps and adjust the drift to preserve the first cumulant. To justify this approximation, we verify convergence of the cumulants and of every fixed determinant coefficient in the following claim.
Claim 8. Under condition (i), there are drift-plus-compact-jump exponents K ℓ , N ( t ) K_{\ell,N}(t) K ℓ , N ( t ) such that
K ℓ , N ( n ) ( 0 ) ⟶ κ n ( ℓ ) ( n ≥ 1 , 1 ≤ ℓ ≤ m ) , K_{\ell,N}^{(n)}(0)\longrightarrow\kappa_n^{(\ell)}\qquad(n\ge1,\ 1\le\ell\le m), K ℓ , N ( n ) ( 0 ) ⟶ κ n ( ℓ ) ( n ≥ 1 , 1 ≤ ℓ ≤ m ) , [ y α ] H r a , N ( y ) ⟶ [ y α ] H r a ( y ) ( r ≥ 1 , α ∈ N 0 m ) . (124) [y^\alpha]H_r^{a,N}(y)\longrightarrow[y^\alpha]H_r^a(y)\qquad(r\ge1,\ \alpha\in\mathbb{N}_0^m). \tag*{(124)} [ y α ] H r a , N ( y ) ⟶ [ y α ] H r a ( y ) ( r ≥ 1 , α ∈ N 0 m ) . (124) Here N → ∞ N\to\infty N → ∞ , and H r a , N ( y ) H_r^{a,N}(y) H r a , N ( y ) denotes the determinant formed from K 1 , N ( t ) , … , K m , N ( t ) K_{1,N}(t),\ldots,K_{m,N}(t) K 1 , N ( t ) , … , K m , N ( t ) . In particular, we retain (115) and (121) in the limit.
Proof. To treat arbitrary data in (i), we choose nonnegative measures ν ℓ \nu_\ell ν ℓ on R \mathbb{R} R , with all absolute moments finite, such that
κ n + 2 ( ℓ ) = ∫ − ∞ ∞ x n ν ℓ ( d x ) ( n ≥ 0 ) , g ℓ = ν ℓ ( { 0 } ) . (125) \kappa_{n+2}^{(\ell)}=\int_{-\infty}^{\infty}x^n\,\nu_\ell(\mathrm{d}x)\qquad(n\ge0),\qquad g_\ell=\nu_\ell(\{0\}). \tag*{(125)} κ n + 2 ( ℓ ) = ∫ − ∞ ∞ x n ν ℓ ( d x ) ( n ≥ 0 ) , g ℓ = ν ℓ ({ 0 }) . (125)
To obtain finite jump measures, we truncate away from zero and infinity before dividing by the square of the jump size. We represent the atom at zero by two symmetric jumps. For integers N ≥ 1 N \ge1 N ≥ 1 , define
ρ ℓ , N ( d x ) = 1 { 1 / N ≤ ∣ x ∣ ≤ N } x − 2 ν ℓ ( d x ) + g ℓ N 2 2 ( δ 1 / N + δ − 1 / N ) ( d x ) , b ℓ , N = κ 1 ( ℓ ) − ∫ − ∞ ∞ x ρ ℓ , N ( d x ) . (126) \rho_{\ell,N}(\mathrm{d}x)=\mathbf{1}_{\{1/N\leq\lvert x\rvert\leq N\}}x^{-2}\nu_{\ell}(\mathrm{d}x)+\frac{g_{\ell}N^2}{2}(\delta_{1/N}+\delta_{-1/N})(\mathrm{d}x),
\qquad
b_{\ell,N}=\kappa_{1}^{(\ell)}-\int_{-\infty}^{\infty}x\,\rho_{\ell,N}(\mathrm{d}x). \tag*{(126)} ρ ℓ , N ( d x ) = 1 { 1/ N ≤ ∣ x ∣ ≤ N } x − 2 ν ℓ ( d x ) + 2 g ℓ N 2 ( δ 1/ N + δ − 1/ N ) ( d x ) , b ℓ , N = κ 1 ( ℓ ) − ∫ − ∞ ∞ x ρ ℓ , N ( d x ) . (126)
The jump measure in (126) is finite: its total mass is at most N 2 ν ℓ ( R ) + g ℓ N 2 N^2\nu_{\ell}(\mathbb{R})+g_{\ell}N^2 N 2 ν ℓ ( R ) + g ℓ N 2 , since x − 2 ≤ N 2 x^{-2}\leq N^2 x − 2 ≤ N 2 on the truncated domain. It also has bounded support. Hence its first absolute moment is finite and b ℓ , N b_{\ell,N} b ℓ , N is a well-defined real number.
Let K ℓ , N K_{\ell,N} K ℓ , N be the associated drift-plus-jump exponent. Its first cumulant is exactly κ 1 ( ℓ ) \kappa_{1}^{(\ell)} κ 1 ( ℓ ) , while for n ≥ 2 n\geq2 n ≥ 2 its n n n -th cumulant is
∫ − N − 1 / N x n − 2 ν ℓ ( d x ) + ∫ 1 / N N x n − 2 ν ℓ ( d x ) + g ℓ 2 N 2 − n ( 1 + ( − 1 ) n ) . (127) \int_{-N}^{-1/N}x^{n-2}\nu_{\ell}(\mathrm{d}x)+\int_{1/N}^{N}x^{n-2}\nu_{\ell}(\mathrm{d}x)+\frac{g_{\ell}}{2}N^{2-n}\left(1+(-1)^n\right). \tag*{(127)} ∫ − N − 1/ N x n − 2 ν ℓ ( d x ) + ∫ 1/ N N x n − 2 ν ℓ ( d x ) + 2 g ℓ N 2 − n ( 1 + ( − 1 ) n ) . (127)
For the second cumulant, the final term reproduces the atom at zero; for every higher cumulant, it tends to zero. The two truncated integrals converge by finiteness of the corresponding absolute moments. We therefore obtain convergence to κ n ( ℓ ) \kappa_{n}^{(\ell)} κ n ( ℓ ) , including the possible Gaussian contribution.
Since each coefficient of a moment polynomial depends on only finitely many cumulants, we obtain coefficientwise convergence of every fixed moment polynomial F n , N ( y ) F_{n,N}(y) F n , N ( y ) to F n ( y ) F_n(y) F n ( y ) . We then pass to the limit in each determinant, which is a finite polynomial expression in these moments, with uniformly bounded degree. Thus we retain coefficientwise positivity, the degree bound, and the formula for the degree-D D D component.
For each numerical y ≥ 0 y\geq0 y ≥ 0 , we also have positive semidefinite finite Hankel matrices for ( F n , N ( y ) ) (F_{n,N}(y)) ( F n , N ( y )) . Taking entrywise limits, we retain positive semidefiniteness. Applying the Hamburger moment theorem to the limiting sequence, we obtain (iii) without an analytic assumption. Finally, since every approximating determinant is independent of the chosen first cumulants, we retain common-drift invariance in the limit. This proves all the asserted consequences of (i). □
We can now complete the equivalence. Choosing the initial measure δ 0 \delta_0 δ 0 in condition (iii), we obtain (iv). It remains to deduce (iv) directly from (ii). Assume (ii) and fix y ≥ 0 y\geq0 y ≥ 0 . To use Sylvester’s criterion, we first make all leading determinants strictly positive by choosing a Gaussian initial measure. For ε > 0 \varepsilon>0 ε > 0 , take a centered Gaussian probability measure of variance ε \varepsilon ε , and denote its moment sequence by a ε = ( a n ε ) n ≥ 0 \mathbf{a}^{\varepsilon}=(a_n^{\varepsilon})_{n\geq0} a ε = ( a n ε ) n ≥ 0 . For this initial measure, every leading Hankel determinant is strictly positive, since the integral of the square of a nonzero polynomial is positive. By (ii), the output determinants have nonnegative coefficients and hence are bounded below by their constant terms. Thus, for every r r r , we obtain
H r a ε ( y ) ≥ H r a ε ( 0 ) > 0. (128) H_r^{a\varepsilon}(\boldsymbol{y}) \ge H_r^{a\varepsilon}(\boldsymbol{0}) > 0. \tag*{(128)} H r a ε ( y ) ≥ H r a ε ( 0 ) > 0. (128)
Fix d ≥ 0 d \ge0 d ≥ 0 and form the symmetric matrix H d ε ( y ) = [ F i + j a ε ( y ) ] i , j = 0 d H_d^{\varepsilon}(\boldsymbol{y}) = [F_{i+j}^{a\varepsilon}(\boldsymbol{y})]_{i,j=0}^{d} H d ε ( y ) = [ F i + j a ε ( y ) ] i , j = 0 d . For each 1 ≤ r ≤ d + 1 1 \le r \le d+1 1 ≤ r ≤ d + 1 , its leading principal minor is strictly positive by (128) . Applying Sylvester’s criterion [15 Section 7.2] to H d ε ( y ) H_d^{\varepsilon}(\boldsymbol{y}) H d ε ( y ) , we obtain positive definiteness.
As ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 , the Gaussian initial moments converge to those of δ 0 \delta_0 δ 0 . Using the finite convolution formula (13) , we therefore obtain F n a ε ( y ) ⟶ M n ( y ) F_n^{a\varepsilon}(\boldsymbol{y}) \longrightarrow M_n(\boldsymbol{y}) F n a ε ( y ) ⟶ M n ( y ) for every fixed n n n . In particular, for every real vector c = ( c 0 , … , c d ) \boldsymbol{c}=(c_0,\ldots,c_d) c = ( c 0 , … , c d ) , we have ∑ i , j = 0 d c i c j M i + j ( y ) = lim ε ↓ 0 c T H d ε ( y ) c ≥ 0 \sum_{i,j=0}^{d}c_i c_jM_{i+j}(\boldsymbol{y})=\lim_{\varepsilon\downarrow0}\boldsymbol{c}^{\mathsf T}H_d^{\varepsilon}(\boldsymbol{y})\boldsymbol{c}\ge0 ∑ i , j = 0 d c i c j M i + j ( y ) = lim ε ↓ 0 c T H d ε ( y ) c ≥ 0 . Since d d d and c \boldsymbol{c} c are arbitrary, applying the Hamburger moment criterion in Lemma 2.1(i) to ( M n ( y ) ) n ≥ 0 (M_n(\boldsymbol{y}))_{n\ge0} ( M n ( y ) ) n ≥ 0 , we obtain (iv). This completes the proof.
Composition, convolution, and substitution
We now study the operators associated with Theorem 1.7. We first prove a binomial-convolution identity for two independently lifted families and recall a classical logarithm criterion. We then establish composition and substitution closure for the admissible exponents.
Define the class of admissible formal exponents by
C = { K ( t ) = ∑ n ≥ 1 κ n t n n ! ∈ t R [ [ t ] ] : ( κ n + 1 ) n ≥ 0 is a Stieltjes moment sequence } . (129) \mathcal{C}=\left\{K(t)=\sum_{n\ge1}\kappa_n\frac{t^n}{n!}\in t\mathbb{R}[[t]]:(\kappa_{n+1})_{n\ge0}\text{ is a Stieltjes moment sequence}\right\}. \tag*{(129)} C = { K ( t ) = n ≥ 1 ∑ κ n n ! t n ∈ t R [[ t ]] : ( κ n + 1 ) n ≥ 0 is a Stieltjes moment sequence } . (129)
Recall the operator T y T_{\boldsymbol{y}} T y defined in (13) . The finite sum in (13) also defines these operators on polynomial-valued sequences. They satisfy the formal identities
T y T u = T y + u , T 0 = Id . (130) T_{\boldsymbol{y}}T_{\boldsymbol{u}}=T_{\boldsymbol{y}+\boldsymbol{u}},\qquad T_{\boldsymbol{0}}=\operatorname{Id}. \tag*{(130)} T y T u = T y + u , T 0 = Id . (130)
Indeed, their exponential generating functions are multiplied by exp ( ∑ ℓ y ℓ K ℓ ) \exp(\sum_{\ell}y_{\ell}K_{\ell}) exp ( ∑ ℓ y ℓ K ℓ ) and exp ( ∑ ℓ u ℓ K ℓ ) \exp(\sum_{\ell}u_{\ell}K_{\ell}) exp ( ∑ ℓ u ℓ K ℓ ) , respectively. All coefficient sums are finite. By Theorem 1.7, after substituting y ℓ ↦ y ℓ + u ℓ y_{\ell}\mapsto y_{\ell}+u_{\ell} y ℓ ↦ y ℓ + u ℓ , we obtain joint coefficientwise positivity for the output in both parameter vectors. More generally, substituting arbitrary polynomials y ℓ = p ℓ ( w ) ∈ R ≥ 0 [ w ] y_{\ell}=p_{\ell}(\boldsymbol{w})\in\mathbb{R}_{\ge0}[\boldsymbol{w}] y ℓ = p ℓ ( w ) ∈ R ≥ 0 [ w ] preserves the coefficientwise conclusion.
Binomial convolution of lifted families
Wang and Zhu [42 Corollary 4.3] recalled the classical fact that binomial convolution preserves real Stieltjes moment sequences. By joint lifting for the exponential families specified, we will extend it to the coefficientwise Hankel-total positivity in the following Proposition 4.1 . In fact, at fixed nonnegative indeterminates, we can apply it to F ( y ) F(y) F ( y ) and G ( z ) G(z) G ( z ) to obtain a numerical Stieltjes sequence.
Proposition 4.1 (Binomial convolution of two lifted families). Let A , B A, B A , B be initial Stieltjes moment series, and let K 1 , … , K m , L 1 , … , L p ∈ C K_1,\ldots,K_m,L_1,\ldots,L_p \in\mathbb{C} K 1 , … , K m , L 1 , … , L p ∈ C . Define F n ( y ) F_n(y) F n ( y ) and G n ( z ) G_n(z) G n ( z ) by
∑ n ≥ 0 F n ( y ) t n n ! = A ( t ) e ∑ i y i K i ( t ) , ∑ n ≥ 0 G n ( z ) t n n ! = B ( t ) e ∑ j z j L j ( t ) . (131) \sum_{n\geq0} F_n(\mathbf{y}) \frac{t^n}{n!}=A(t)e^{\sum_i y_iK_i(t)}, \qquad\sum_{n\geq0} G_n(\mathbf{z}) \frac{t^n}{n!}=B(t)e^{\sum_j z_jL_j(t)}. \tag*{(131)} n ≥ 0 ∑ F n ( y ) n ! t n = A ( t ) e ∑ i y i K i ( t ) , n ≥ 0 ∑ G n ( z ) n ! t n = B ( t ) e ∑ j z j L j ( t ) . (131) Then the binomial convolution
H n ( y , z ) = ∑ k = 0 n ( n k ) F k ( y ) G n − k ( z ) ( n ≥ 0 ) (132) H_n(\mathbf{y},\mathbf{z})=\sum_{k=0}^{n}\binom{n}{k}F_k(\mathbf{y})G_{n-k}(\mathbf{z}) \qquad(n\geq0) \tag*{(132)} H n ( y , z ) = k = 0 ∑ n ( k n ) F k ( y ) G n − k ( z ) ( n ≥ 0 ) (132) is jointly coefficientwise Hankel-totally positive in y , z \mathbf{y},\mathbf{z} y , z . Write a ⋆ b \mathbf{a}\star\mathbf{b} a ⋆ b for binomial convolution and T K \mathcal{T}^{K} T K for the lifting operator with exponent list K K K . We have
( T y K a ) ⋆ ( T z L b ) = T ( y , z ) ( K , L ) ( a ⋆ b ) . (133) \left(\mathcal{T}^{K}_{\mathbf{y}}\mathbf{a}\right)\star\left(\mathcal{T}^{L}_{\mathbf{z}}\mathbf{b}\right)=\mathcal{T}^{(K,L)}_{(\mathbf{y},\mathbf{z})}(\mathbf{a}\star\mathbf{b}). \tag*{(133)} ( T y K a ) ⋆ ( T z L b ) = T ( y , z ) ( K , L ) ( a ⋆ b ) . (133) For the same exponent list, the right side is T y + z K ( a ⋆ b ) \mathcal{T}^{K}_{\mathbf{y}+\mathbf{z}}(\mathbf{a}\star\mathbf{b}) T y + z K ( a ⋆ b ) . The coefficientwise conclusion is preserved by identifying variables.
Proof. Let μ , ν \mu,\nu μ , ν represent the initial sequences. Integrating ( x + s ) n (x+s)^n ( x + s ) n against μ ( d x ) ν ( d s ) \mu(\mathrm{d}x)\nu(\mathrm{d}s) μ ( d x ) ν ( d s ) , we obtain the moments of their binomial convolution. Thus A ( t ) B ( t ) A(t)B(t) A ( t ) B ( t ) is an initial Stieltjes moment series, including arbitrary total masses and the zero-measure cases. Multiplying the two generating functions, we obtain A ( t ) B ( t ) exp ( ∑ i y i K i + ∑ j z j L j ) A(t)B(t)\exp(\sum_i y_iK_i+\sum_j z_jL_j) A ( t ) B ( t ) exp ( ∑ i y i K i + ∑ j z j L j ) . Applying Theorem 1.7 to this initial series and the concatenated exponent list, we derive the joint coefficientwise assertion. Extracting n ! [ t n ] n![t^n] n ! [ t n ] , we obtain (133) . For a common list, we combine its intensities by addition. Finally, substituting the same indeterminate for selected intensity variables, we retain nonnegative coefficients.
Logarithms of ordinary Stieltjes series
We use the following classical consequence of the infinite divisibility of mixtures of exponential distributions [14 ] . If X X X has Stieltjes moments b n b_n b n , and E E E is an independent unit exponential variable, then E ( X E ) n = n ! b n \mathbb{E}(XE)^n=n!b_n E ( X E ) n = n ! b n . Applying Goldie’s theorem to X E XE X E and the Lévy representation to its nonnegative infinitely divisible law, we obtain the shifted cumulant moment property below. We include an analytic proof using [35 Theorems 3.2, 7.3 and Corollary 3.8] , so that we also cover ordinary formal series with zero convergence radius.
Nonnegative mixtures and composition
Proposition 4.2 (A logarithm criterion). Let ( b n ) n ≥ 0 (b_n)_{n\geq0} ( b n ) n ≥ 0 be a real Stieltjes moment sequence with b 0 = 1 b_0=1 b 0 = 1 . Define the ordinary formal power series B ( t ) B(t) B ( t ) and the exponential coefficients κ n \kappa_n κ n of its logarithm by
B ( t ) = ∑ n ≥ 0 b n t n , log B ( t ) = ∑ n ≥ 1 κ n t n n ! . (134) B(t)=\sum_{n\geq0}b_nt^n,\qquad\log B(t)=\sum_{n\geq1}\kappa_n\frac{t^n}{n!}. \tag*{(134)} B ( t ) = n ≥ 0 ∑ b n t n , log B ( t ) = n ≥ 1 ∑ κ n n ! t n . (134) Then ( κ n + 1 ) n ≥ 0 (\kappa_{n+1})_{n\geq0} ( κ n + 1 ) n ≥ 0 is a Stieltjes moment sequence; equivalently, log B ∈ C \log B\in\mathcal{C} log B ∈ C .
Proof. Choose a probability measure μ \mu μ on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with moments b n b_n b n , and, for s ≥ 0 s\geq0 s ≥ 0 , let
L ( s ) = ∫ 0 ∞ μ ( d x ) 1 + s x , h ( s ) = 1 L ( s ) − 1 , ϕ ( s ) = − log L ( s ) . (135) L(s)=\int_0^\infty\frac{\mu(\mathrm{d}x)}{1+sx},\qquad h(s)=\frac{1}{L(s)}-1,\qquad\phi(s)=-\log L(s). \tag*{(135)} L ( s ) = ∫ 0 ∞ 1 + s x μ ( d x ) , h ( s ) = L ( s ) 1 − 1 , ϕ ( s ) = − log L ( s ) . (135)
We have 0 < L ( s ) ≤ 1 0<L(s)\leq1 0 < L ( s ) ≤ 1 and L ( 0 ) = 1 L(0)=1 L ( 0 ) = 1 . For x > 0 x>0 x > 0 , we can write ( 1 + s x ) − 1 = x − 1 ( s + x − 1 ) − 1 (1+sx)^{-1}=x^{-1}(s+x^{-1})^{-1} ( 1 + s x ) − 1 = x − 1 ( s + x − 1 ) − 1 and a mass of μ \mu μ at zero contributes a nonnegative constant. Thus L ( s ) L(s) L ( s ) is a Stieltjes function in the classical terminology of [35 Chapter 2] . By the reciprocal characterization in [35 Theorem 7.3] , its reciprocal is a Bernstein function. Since L ( s ) − 1 ≥ 1 L(s)^{-1}\geq1 L ( s ) − 1 ≥ 1 , we may subtract its value at zero and conclude that h ( s ) h(s) h ( s ) is also a Bernstein function. We also know that log ( 1 + s ) \log(1+s) log ( 1 + s ) is a Bernstein function. Applying the composition rule in [35 Corollary 3.8] to h ( s ) h(s) h ( s ) and log ( 1 + s ) \log(1+s) log ( 1 + s ) , we obtain that ϕ ( s ) = log ( 1 + h ( s ) ) \phi(s)=\log(1+h(s)) ϕ ( s ) = log ( 1 + h ( s )) is a Bernstein function with ϕ ( 0 ) = 0 \phi(0)=0 ϕ ( 0 ) = 0 .
To identify all derivatives at zero, we differentiate the first integral in (135) . For every n ≥ 0 n\geq0 n ≥ 0 , using the integrable bound n ! x n n!x^n n ! x n and applying dominated convergence to the differentiated integrals, we have
L ( n ) ( s ) = ( − 1 ) n n ! ∫ 0 ∞ x n ( 1 + s x ) n + 1 μ ( d x ) , L ( n ) ( 0 + ) = ( − 1 ) n n ! b n . (136) L^{(n)}(s)=(-1)^n n!\int_0^\infty\frac{x^n}{(1+sx)^{n+1}}\mu(\mathrm{d}x),\qquad L^{(n)}(0+)=(-1)^n n!b_n. \tag*{(136)} L ( n ) ( s ) = ( − 1 ) n n ! ∫ 0 ∞ ( 1 + s x ) n + 1 x n μ ( d x ) , L ( n ) ( 0 + ) = ( − 1 ) n n ! b n . (136)
In consequence, by the finite coefficient formulas for a logarithm, ϕ ( n + 1 ) ( 0 + ) = ( − 1 ) n κ n + 1 \phi^{(n+1)}(0+)=(-1)^n\kappa_{n+1} ϕ ( n + 1 ) ( 0 + ) = ( − 1 ) n κ n + 1 . At each order, we use only finitely many derivatives in this comparison, so no convergence assumption on B ( t ) B(t) B ( t ) is needed.
By the classical Lévy representation [35 Theorem 3.2] , there are a drift d ≥ 0 d\geq0 d ≥ 0 and a positive measure ρ \rho ρ on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) such that ϕ ( s ) = d s + ∫ 0 ∞ ( 1 − e − s x ) ρ ( d x ) \phi(s)=ds+\int_0^\infty(1-e^{-sx})\rho(\mathrm{d}x) ϕ ( s ) = d s + ∫ 0 ∞ ( 1 − e − s x ) ρ ( d x ) , with ∫ 0 ∞ min { 1 , x } ρ ( d x ) < ∞ \int_0^\infty\min\{1,x\}\rho(\mathrm{d}x)<\infty ∫ 0 ∞ min { 1 , x } ρ ( d x ) < ∞ . For s > 0 s>0 s > 0 , we may differentiate under this integral. Then, by monotone convergence as s ↓ 0 s\downarrow0 s ↓ 0 , we obtain
κ n + 1 = ( − 1 ) n ϕ ( n + 1 ) ( 0 + ) = d 1 { n = 0 } + ∫ 0 ∞ x n + 1 ρ ( d x ) ( n ≥ 0 ) . (137) \kappa_{n+1}=(-1)^n\phi^{(n+1)}(0+)=d\mathbf{1}_{\{n=0\}}+\int_0^\infty x^{n+1}\rho(\mathrm{d}x)\qquad(n\geq0). \tag*{(137)} κ n + 1 = ( − 1 ) n ϕ ( n + 1 ) ( 0 + ) = d 1 { n = 0 } + ∫ 0 ∞ x n + 1 ρ ( d x ) ( n ≥ 0 ) . (137)
By (136) , all derivatives at zero are finite, and hence so are the moment integrals above. We may therefore use the finite positive measure ν ( d x ) = d δ 0 ( d x ) + x ρ ( d x ) \nu(\mathrm{d}x)=d\delta_0(\mathrm{d}x)+x\rho(\mathrm{d}x) ν ( d x ) = d δ 0 ( d x ) + x ρ ( d x ) to represent ( κ n + 1 ) n ≥ 0 (\kappa_{n+1})_{n\geq0} ( κ n + 1 ) n ≥ 0 . This completes the proof. ∎
Nonnegative mixtures and composition
We begin with closure under nonnegative mixtures. The proof uses quadratic forms and requires no measurable choice of representing measures.
Proposition 4.3 (Nonnegative mixtures). Let f n ( s ) f_n(s) f n ( s ) be measurable functions on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , and suppose that ( f n ( s ) ) n ≥ 0 (f_n(s))_{n\geq0} ( f n ( s ) ) n ≥ 0 is a Stieltjes moment sequence for every s s s . Let σ \sigma σ be a nonnegative measure on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that ∫ 0 ∞ f n ( s ) σ ( d s ) < ∞ \int_0^\infty f_n(s)\,\sigma(\mathrm{d}s)<\infty ∫ 0 ∞ f n ( s ) σ ( d s ) < ∞ for every n ≥ 0 n\geq0 n ≥ 0 . Then
g n = ∫ 0 ∞ f n ( s ) σ ( d s ) (138) g_n=\int_0^\infty f_n(s)\,\sigma(\mathrm{d}s) \tag*{(138)} g n = ∫ 0 ∞ f n ( s ) σ ( d s ) (138) is a Stieltjes moment sequence.
Proof. For h = 0 , 1 h=0,1 h = 0 , 1 and any finite real vector ( c 0 , … , c d ) (c_0,\ldots,c_d) ( c 0 , … , c d ) , we have
∑ i , j = 0 d c i c j g i + j + h = ∫ 0 ∞ ∑ i , j = 0 d c i c j f i + j + h ( s ) d σ ( s ) ≥ 0. (139) \sum_{i,j=0}^{d}c_i c_j g_{i+j+h}=\int_0^\infty\sum_{i,j=0}^{d}c_i c_j f_{i+j+h}(s)\,\mathrm{d}\sigma(s)\geq0. \tag*{(139)} i , j = 0 ∑ d c i c j g i + j + h = ∫ 0 ∞ i , j = 0 ∑ d c i c j f i + j + h ( s ) d σ ( s ) ≥ 0. (139)
The integrand is nonnegative by the Stieltjes property at each s s s . Since every term is integrable by hypothesis, we may integrate the finite sums term by term. Applying Lemma 2.1 to the resulting quadratic forms, we obtain the Stieltjes property.
For K ( t ) , L ( t ) ∈ t R [ [ t ] ] K(t),L(t)\in t\mathbb{R}[[t]] K ( t ) , L ( t ) ∈ t R [[ t ]] , with K ( t ) = ∑ k ≥ 1 κ k t k / k ! K(t)=\sum_{k\geq1}\kappa_k t^k/k! K ( t ) = ∑ k ≥ 1 κ k t k / k ! , we use ∘ \circ ∘ for composition by formal substitution:
( K ∘ L ) ( t ) : = K ( L ( t ) ) = ∑ k ≥ 1 κ k k ! L ( t ) k . (140) (K\circ L)(t):=K(L(t))=\sum_{k\geq1}\frac{\kappa_k}{k!}L(t)^k. \tag*{(140)} ( K ∘ L ) ( t ) := K ( L ( t )) = k ≥ 1 ∑ k ! κ k L ( t ) k . (140)
Because L ( 0 ) = 0 L(0)=0 L ( 0 ) = 0 , only finitely many powers contribute to each coefficient of t t t ; thus we have a well-defined formal power series. Using the mixture property together with this definition, we establish closure under composition.
Theorem 4.4 (Composition closure). If K , L ∈ C K,L\in\mathcal{C} K , L ∈ C , then K ∘ L ∈ C K\circ L\in\mathcal{C} K ∘ L ∈ C .
Proof. Fix a real number u ≥ 0 u\geq0 u ≥ 0 , and write
e u K ( w ) = ∑ k ≥ 0 B k ( u ) w k k ! . (141) e^{uK(w)}=\sum_{k\geq0}B_k(u)\frac{w^k}{k!}. \tag*{(141)} e u K ( w ) = k ≥ 0 ∑ B k ( u ) k ! w k . (141)
By Theorem 1.7 , there is a nonnegative measure μ u \mu_u μ u on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , with all moments finite, such that B k ( u ) = ∫ 0 ∞ s k μ u ( d s ) B_k(u)=\int_0^\infty s^k\,\mu_u(\mathrm{d}s) B k ( u ) = ∫ 0 ∞ s k μ u ( d s ) . Also define the moment polynomials associated with L L L by
e s L ( t ) = ∑ n ≥ 0 P n ( s ) t n n ! . (142) e^{sL(t)}=\sum_{n\geq0}P_n(s)\frac{t^n}{n!}. \tag*{(142)} e s L ( t ) = n ≥ 0 ∑ P n ( s ) n ! t n . (142)
For every s ≥ 0 s \ge0 s ≥ 0 , ( P n ( s ) ) n ≥ 0 (P_n(s))_{n \ge0} ( P n ( s ) ) n ≥ 0 is a Stieltjes moment sequence, again by Theorem 1.7 . Moreover, P n ( s ) P_n(s) P n ( s ) is a polynomial of degree at most n n n . Since L ( 0 ) = 0 L(0)=0 L ( 0 ) = 0 , only powers L ( t ) k L(t)^k L ( t ) k with k ≤ n k \le n k ≤ n contribute to a fixed coefficient, giving
n ! [ t n ] e u K ( L ( t ) ) = ∫ 0 ∞ P n ( s ) μ u ( d s ) . (143) n![t^n]e^{uK(L(t))}=\int_0^\infty P_n(s)\,\mu_u(ds). \tag*{(143)} n ! [ t n ] e u K ( L ( t )) = ∫ 0 ∞ P n ( s ) μ u ( d s ) . (143)
Since P n ( s ) P_n(s) P n ( s ) is a polynomial and μ u \mu_u μ u has all moments finite, the integral is finite. By Proposition 4.3 , we conclude that the sequence on the left-hand side is a Stieltjes moment sequence. Since this holds for every u ≥ 0 u \ge0 u ≥ 0 , applying the implication (iii)⇒ \Rightarrow ⇒ (i) of Theorem 1.7 to the formal exponent K ( L ( t ) ) K(L(t)) K ( L ( t )) , we obtain K ∘ L ∈ C K\circ L\in\mathcal{C} K ∘ L ∈ C . We use only finite coefficient calculations and polynomial integrals. Hence we can use this argument even when the formal power series have zero radius of convergence.
Universal substitution–convolution preservers
The next theorem characterizes formal substitutions of the initial series that preserve universal coefficientwise Hankel-total positivity.
Theorem 4.5 (Substitution and convolution). Fix L ∈ t R [ [ t ] ] L\in t\mathbb{R}[[t]] L ∈ t R [[ t ]] and K 1 , … , K m ∈ t R [ [ t ] ] K_1,\ldots,K_m\in t\mathbb{R}[[t]] K 1 , … , K m ∈ t R [[ t ]] . For an initial real sequence a = ( a n ) n ≥ 0 \mathbf{a}=(a_n)_{n\ge0} a = ( a n ) n ≥ 0 with formal exponential generating function A ( t ) = ∑ n ≥ 0 a n t n / n ! A(t)=\sum_{n\ge0}a_nt^n/n! A ( t ) = ∑ n ≥ 0 a n t n / n ! , define Q n ( y ) Q_n(\mathbf{y}) Q n ( y ) by
∑ n ≥ 0 Q n ( y ) t n n ! = A ( L ( t ) ) exp ( ∑ ℓ = 1 m y ℓ K ℓ ( t ) ) . (144) \sum_{n\ge0}Q_n(\mathbf{y})\frac{t^n}{n!}=A(L(t))\exp\left(\sum_{\ell=1}^{m}y_\ell K_\ell(t)\right). \tag*{(144)} n ≥ 0 ∑ Q n ( y ) n ! t n = A ( L ( t )) exp ( ℓ = 1 ∑ m y ℓ K ℓ ( t ) ) . (144) The following conditions are equivalent:
L , K 1 , … , K m ∈ C L,K_1,\ldots,K_m\in\mathcal{C} L , K 1 , … , K m ∈ C .
For every initial real Stieltjes moment sequence, ( Q n ( y ) ) n ≥ 0 (Q_n(\mathbf{y}))_{n\ge0} ( Q n ( y ) ) n ≥ 0 is jointly coefficientwise Hankel-totally positive in y \mathbf{y} y .
For every initial real Stieltjes moment sequence and every y ∈ [ 0 , ∞ ) m \mathbf{y}\in[0,\infty)^m y ∈ [ 0 , ∞ ) m , ( Q n ( y ) ) n ≥ 0 (Q_n(\mathbf{y}))_{n\ge0} ( Q n ( y ) ) n ≥ 0 is a Stieltjes moment sequence.
Proof. Suppose (i) holds, and let μ \mu μ represent the initial sequence. Define P n ( s ) P_n(s) P n ( s ) by e s L ( t ) = ∑ n ≥ 0 P n ( s ) t n / n ! e^{sL(t)}=\sum_{n\ge0}P_n(s)t^n/n! e s L ( t ) = ∑ n ≥ 0 P n ( s ) t n / n ! . For each s ≥ 0 s\ge0 s ≥ 0 , we know from Theorem 1.7 that ( P n ( s ) ) n ≥ 0 (P_n(s))_{n\ge0} ( P n ( s ) ) n ≥ 0 is a Stieltjes moment sequence. Since P n ( s ) P_n(s) P n ( s ) is a polynomial of degree at most n n n , we obtain by formal substitution the finite coefficient identity
b n : = n ! [ t n ] A ( L ( t ) ) = ∫ 0 ∞ P n ( s ) μ ( d s ) . (145) b_n:=n![t^n]A(L(t))=\int_0^\infty P_n(s)\,\mu(ds). \tag*{(145)} b n := n ! [ t n ] A ( L ( t )) = ∫ 0 ∞ P n ( s ) μ ( d s ) . (145)
The integral is finite. The numerical preservation step is the row-moment criterion of Wang and Zhu [42 Theorem 4.2] , or equivalently Proposition 4.3 . Thus ( b n ) (b_n) ( b n ) is a Stieltjes moment sequence. Applying Theorem 1.7 to the initial Stieltjes sequence ( b n ) n ≥ 0 (b_n)_{n\geq0} ( b n ) n ≥ 0 and the exponents K 1 ( t ) , … , K m ( t ) K_1(t),\ldots,K_m(t) K 1 ( t ) , … , K m ( t ) , we obtain (ii). By nonnegative specialization, we have (ii)⇒ \Rightarrow ⇒ (iii).
Conversely, assume that (iii) holds. Taking the initial measure δ 0 \delta_0 δ 0 , we obtain the exponential family of Theorem 1.7 , so K 1 , … , K m ∈ C K_1,\ldots,K_m\in\mathcal{C} K 1 , … , K m ∈ C . Set y = 0 \mathbf{y}=\mathbf{0} y = 0 , and use the initial measure δ s \delta_s δ s for every real s ≥ 0 s\geq0 s ≥ 0 . Its initial series is A ( t ) = e s t A(t)=e^{st} A ( t ) = e s t , so the coefficients of e s L ( t ) e^{sL(t)} e s L ( t ) form a Stieltjes moment sequence for every s ≥ 0 s\geq0 s ≥ 0 . By the converse direction of Theorem 1.7 , we now obtain L ∈ C L\in\mathcal{C} L ∈ C . We complete the proof.
For formal power series B ( t ) , L ( t ) B(t),L(t) B ( t ) , L ( t ) with L ( 0 ) = 0 L(0)=0 L ( 0 ) = 0 , write S B , L A = e B ( t ) A ( L ( t ) ) S_{B,L}A=e^{B(t)}A(L(t)) S B , L A = e B ( t ) A ( L ( t )) . These operators satisfy
S B , L S H , R = S B + H ∘ L , R ∘ L . (146) S_{B,L}S_{H,R}=S_{B+H\circ L,R\circ L}. \tag*{(146)} S B , L S H , R = S B + H ∘ L , R ∘ L . (146)
The class C \mathcal{C} C is closed under nonnegative linear combinations, by adding representing measures, and under composition by Theorem 4.4 . Thus the admissible operators are closed under finite composition. With the substitution series fixed, independent convolution intensities may be retained in each factor. We obtain joint coefficientwise positivity by Theorem 4.5 . Only finite coefficient identities are involved.
Moment inputs for component exponents
The next construction turns any Stieltjes moment sequence into admissible component exponents. Its specializations organize the tree and forest applications below.
Theorem 4.6 (Moment inputs and universal component families). Let b = ( b n ) n ≥ 0 \boldsymbol{b}=(b_n)_{n\geq0} b = ( b n ) n ≥ 0 be a real Stieltjes moment sequence, represented by a nonnegative measure ν \nu ν on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with all moments finite. Define the formal series
I b ( t ) = ∑ n ≥ 0 b n t n + 1 ( n + 1 ) ! , I_b(t)=\sum_{n\geq0}b_n\frac{t^{n+1}}{(n+1)!}, I b ( t ) = n ≥ 0 ∑ b n ( n + 1 )! t n + 1 , J b ( t ) = ∑ n ≥ 1 b n t n n ! , J_b(t)=\sum_{n\geq1}b_n\frac{t^n}{n!}, J b ( t ) = n ≥ 1 ∑ b n n ! t n , O b ( t ) = t ∑ n ≥ 0 b n t n . (147) O_b(t)=t\sum_{n\geq0}b_nt^n. \tag*{(147)} O b ( t ) = t n ≥ 0 ∑ b n t n . (147) Then we have:
(i) All three formal series in (147) belong to C \mathcal{C} C ;
(ii) For Stieltjes sequences b ( ℓ ) b^{(\ell)} b ( ℓ ) , choose K ℓ ∈ { I b ( ℓ ) , J b ( ℓ ) , O b ( ℓ ) } K_\ell\in\{I_{b^{(\ell)}}, J_{b^{(\ell)}}, O_{b^{(\ell)}}\} K ℓ ∈ { I b ( ℓ ) , J b ( ℓ ) , O b ( ℓ ) } . For fixed q ℓ ≥ 0 q_\ell\ge0 q ℓ ≥ 0 , S ∈ C S \in\mathcal{C} S ∈ C , and any initial Stieltjes moment series A ( t ) A(t) A ( t ) , the exponential coefficients of
A ( S ( t ) ) exp ( ∑ ℓ = 1 m z ℓ K ℓ ( t ) ) ∏ ℓ = 1 m ( 1 − q ℓ K ℓ ( t ) ) − y ℓ (148) A(S(t)) \exp\left(\sum_{\ell=1}^{m} z_\ell K_\ell(t)\right) \prod_{\ell=1}^{m}(1-q_\ell K_\ell(t))^{-y_\ell} \tag*{(148)} A ( S ( t )) exp ( ℓ = 1 ∑ m z ℓ K ℓ ( t ) ) ℓ = 1 ∏ m ( 1 − q ℓ K ℓ ( t ) ) − y ℓ (148)
are coefficientwise Hankel-totally positive jointly in the indeterminates y , z \mathbf{y},\mathbf{z} y , z .
Proof. The shifted moments of I b I_b I b and J b J_b J b are represented by ν \nu ν and x ν ( d x ) x\nu(dx) xν ( d x ) , respectively. Both measures have all moments finite. For O b O_b O b , applying Euler’s integral [27 Eq. 5.2.1] to the gamma factor, we obtain
( n + 1 ) ! b n = ∫ 0 ∞ ∫ 0 ∞ ( g x ) n g e − g d g ν ( d x ) . (149) (n+1)!b_n=\int_{0}^{\infty}\int_{0}^{\infty}(gx)^n g e^{-g}\,dg\,\nu(dx). \tag*{(149)} ( n + 1 )! b n = ∫ 0 ∞ ∫ 0 ∞ ( g x ) n g e − g d g ν ( d x ) . (149)
The image of ν ( d x ) g e − g d g \nu(dx)g e^{-g}dg ν ( d x ) g e − g d g under ( x , g ) ↦ x g (x,g)\mapsto xg ( x , g ) ↦ xg therefore represents the third shifted sequence. This proves all three admissibility assertions.
By Proposition 4.2 and Theorem 4.4, respectively, we obtain − log ( 1 − q ℓ t ) ∈ C -\log(1-q_\ell t)\in\mathcal{C} − log ( 1 − q ℓ t ) ∈ C and − log ( 1 − q ℓ K ℓ ) ∈ C -\log(1-q_\ell K_\ell)\in\mathcal{C} − log ( 1 − q ℓ K ℓ ) ∈ C . Thus, applying Theorem 4.5 to these exponents, K ℓ K_\ell K ℓ , and substitution S S S , we obtain positivity for (148) . □
Functional digraphs and colored components
We apply Theorems 1.7 and 1.8 to functional digraphs, iterated tree series, and Gaussian perturbations. In Sections 5–8, an initial Stieltjes moment series means a formal series A ( t ) = ∑ n ≥ 0 a n t n / n ! A(t)=\sum_{n\ge0}a_n t^n/n! A ( t ) = ∑ n ≥ 0 a n t n / n ! whose real coefficient sequence is Stieltjes; its total mass a 0 a_0 a 0 need not be one. Hamburger initial sequences are specified explicitly. Unless stated otherwise, structural parameters are fixed, and coefficientwise positivity refers to the indicated independent indeterminates.
Connected-component moments and colored functional digraphs
To verify the moment hypothesis for the connected coefficients c n ( q ) c_n(q) c n ( q ) in (1.5), we use the second assertion of Theorem 1.2. Writing the rooted-forest numbers as
f n , k = ( n − 1 k − 1 ) n n − k ( 1 ≤ k ≤ n ) , (150) f_{n,k}=\binom{n-1}{k-1}n^{n-k}\qquad(1\le k\le n), \tag*{(150)} f n , k = ( k − 1 n − 1 ) n n − k ( 1 ≤ k ≤ n ) , (150)
we obtain the coefficient identity
c n ( q ) = ∑ k = 1 n ( k − 1 ) ! f n , k q k ; (151) c_n(q)=\sum_{k=1}^{n}(k-1)!f_{n,k}q^k; \tag*{(151)} c n ( q ) = k = 1 ∑ n ( k − 1 )! f n , k q k ; (151)
as in the expansion used in [48 Proposition 6.18(iii)] . By the component interpretation, we obtain (1.6) with exponent ∑ ℓ y ℓ C q ℓ ( t ) \sum_\ell y_\ell C_{q_\ell}(t) ∑ ℓ y ℓ C q ℓ ( t ) . Comparing this identity with the colored partition formula (2.1), we identify the cumulants as κ n ( ℓ ) = c n ( q ℓ ) \kappa_n^{(\ell)}=c_n(q_\ell) κ n ( ℓ ) = c n ( q ℓ ) .
Proof of Theorem 1.3. Applying Theorem 1.2 at each fixed q ℓ ≥ 0 q_\ell\ge0 q ℓ ≥ 0 , we obtain the Stieltjes sequences ( c n + 1 ( q ℓ ) ) n ≥ 0 (c_{n+1}(q_\ell))_{n\ge0} ( c n + 1 ( q ℓ ) ) n ≥ 0 . We identify these as the once-shifted cumulants in (1.6). Applying Theorem 1.7, (i)⇒ \Rightarrow ⇒ (ii), to these inputs, we prove Theorem 1.5.
We may introduce a common component parameter by substituting y 1 = ⋯ = y m = y y_1=\cdots=y_m=y y 1 = ⋯ = y m = y , since this substitution preserves coefficientwise nonnegativity. Setting q 1 = ⋯ = q m = q q_1=\cdots=q_m=q q 1 = ⋯ = q m = q , we also obtain the identity
Ψ n ( ( q , … , q ) ; y ) = ψ n ( q , y 1 + ⋯ + y m ) . (152) \Psi_n((q,\ldots,q);y)=\psi_n(q,y_1+\cdots+y_m). \tag*{(152)} Ψ n (( q , … , q ) ; y ) = ψ n ( q , y 1 + ⋯ + y m ) . (152)
To illustrate the multivariate conclusion, put
U = ∑ ℓ = 1 m q ℓ y ℓ , V = ∑ ℓ = 1 m ( 2 q ℓ + q ℓ 2 ) y ℓ . (153) U=\sum_{\ell=1}^{m}q_\ell y_\ell,\qquad V=\sum_{\ell=1}^{m}(2q_\ell+q_\ell^2)y_\ell. \tag*{(153)} U = ℓ = 1 ∑ m q ℓ y ℓ , V = ℓ = 1 ∑ m ( 2 q ℓ + q ℓ 2 ) y ℓ . (153)
Then Ψ 0 = 1 \Psi_0=1 Ψ 0 = 1 , Ψ 1 = U \Psi_1=U Ψ 1 = U , Ψ 2 = U 2 + V \Psi_2=U^2+V Ψ 2 = U 2 + V , and
det ( Ψ 0 Ψ 1 Ψ 1 Ψ 2 ) = V ∈ R ≥ 0 [ y ] . (154) \det\begin{pmatrix}\Psi_0&\Psi_1\\\Psi_1&\Psi_2\end{pmatrix}=V\in\mathbb{R}_{\ge0}[\mathbf{y}]. \tag*{(154)} det ( Ψ 0 Ψ 1 Ψ 1 Ψ 2 ) = V ∈ R ≥ 0 [ y ] . (154)
For m = 1 m=1 m = 1 , q 1 = 1 q_1=1 q 1 = 1 , we recover the original polynomials
ψ 1 ( y ) = y , ψ 2 ( y ) = y 2 + 3 y , ψ 3 ( y ) = y 3 + 9 y 2 + 17 y , (155) \psi_1(y)=y,\qquad\psi_2(y)=y^2+3y,\qquad\psi_3(y)=y^3+9y^2+17y, \tag*{(155)} ψ 1 ( y ) = y , ψ 2 ( y ) = y 2 + 3 y , ψ 3 ( y ) = y 3 + 9 y 2 + 17 y , (155)
and
ψ 4 ( y ) = y 4 + 18 y 3 + 95 y 2 + 142 y , det ( ψ 1 ( y ) ψ 3 ( y ) ψ 2 ( y ) ψ 4 ( y ) ) = y 2 ( 6 y 2 + 51 y + 91 ) . (156) \psi_4(y)=y^4+18y^3+95y^2+142y,\qquad\det\begin{pmatrix}\psi_1(y)&\psi_3(y)\\\psi_2(y)&\psi_4(y)\end{pmatrix}=y^2(6y^2+51y+91). \tag*{(156)} ψ 4 ( y ) = y 4 + 18 y 3 + 95 y 2 + 142 y , det ( ψ 1 ( y ) ψ 2 ( y ) ψ 3 ( y ) ψ 4 ( y ) ) = y 2 ( 6 y 2 + 51 y + 91 ) . (156)
General connected series and iterated tree functions
The following logarithmic composition extends the functional-digraph construction to arbitrary exponents in C \mathbb{C} C in (4.1).
Corollary 5.1 (Logarithmic and iterated-tree families). Fix L 1 , … , L m ∈ C L_1,\ldots,L_m\in\mathbb{C} L 1 , … , L m ∈ C and q 1 , … , q m ≥ 0 q_1,\ldots,q_m\ge0 q 1 , … , q m ≥ 0 . Let A ( t ) A(t) A ( t ) be any initial Stieltjes moment series. Define F n ( y ) F_n(\mathbf{y}) F n ( y ) by
∑ n ≥ 0 F n ( y ) t n n ! = A ( t ) ∏ ℓ = 1 m ( 1 − q ℓ L ℓ ( t ) ) − y ℓ . (157) \sum_{n\ge0}F_n(\mathbf{y})\frac{t^n}{n!}=A(t)\prod_{\ell=1}^{m}(1-q_\ell L_\ell(t))^{-y_\ell}. \tag*{(157)} n ≥ 0 ∑ F n ( y ) n ! t n = A ( t ) ℓ = 1 ∏ m ( 1 − q ℓ L ℓ ( t ) ) − y ℓ . (157) Then ( F n ( y ) ) n ≥ 0 (F_n(\mathbf{y}))_{n\ge0} ( F n ( y ) ) n ≥ 0 is coefficientwise Hankel-totally positive jointly in y 1 , … , y m y_1,\ldots,y_m y 1 , … , y m .
Proof. Put G q ( w ) = − log ( 1 − q w ) G_q(w)=-\log(1-qw) G q ( w ) = − log ( 1 − q w ) . For q > 0 q>0 q > 0 , applying Euler’s gamma integral [27 Eq. 5.2.1] to the shifted coefficients, we obtain
G q ( n + 1 ) ( 0 ) = n ! q n + 1 = ∫ 0 ∞ x n e − x / q d x ( n ≥ 0 ) . (158) G_q^{(n+1)}(0)=n!q^{n+1}=\int_0^\infty x^n e^{-x/q}\,dx\qquad(n\ge0). \tag*{(158)} G q ( n + 1 ) ( 0 ) = n ! q n + 1 = ∫ 0 ∞ x n e − x / q d x ( n ≥ 0 ) . (158)
Thus G q ∈ C G_q \in C G q ∈ C ; for q = 0 q=0 q = 0 , its shifted sequence is represented by the zero measure. Applying Theorem 4.4 to G q ℓ G_{q_\ell} G q ℓ and L ℓ L_\ell L ℓ , we obtain G q ℓ ∘ L ℓ ∈ C G_{q_\ell} \circ L_\ell\in C G q ℓ ∘ L ℓ ∈ C . Applying Theorem 1.7 to these exponents and A A A , we obtain the claimed positivity for (157) .
To include iterated Cayley tree functions in this construction, we first verify their admissibility.
Proposition 5.2. The tree function T T T belongs to C C C . Consequently, T ∘ d ∈ C T^{\circ d} \in C T ∘ d ∈ C for every integer d ≥ 1 d \ge1 d ≥ 1 .
Proof. For every n ≥ 1 n \ge1 n ≥ 1 , expanding (151) at zero, we obtain
lim q ↓ 0 c n ( q ) q = f n , 1 = n n − 1 . (159) \lim_{q \downarrow0} \frac{c_n(q)}{q} = f_{n,1} = n^{n-1}. \tag*{(159)} q ↓ 0 lim q c n ( q ) = f n , 1 = n n − 1 . (159)
For each q > 0 q > 0 q > 0 , we know from Theorem 1.2 that ( c n + 1 ( q ) / q ) n ≥ 0 (c_{n+1}(q)/q)_{n\geq0} ( c n + 1 ( q ) / q ) n ≥ 0 is a Stieltjes moment sequence. Passing to the limit in each finite Hankel minor, we obtain a nonnegative minor for ( ( n + 1 ) n ) n ≥ 0 ((n+1)^n)_{n\geq0} (( n + 1 ) n ) n ≥ 0 . Hence this sequence is a Stieltjes moment sequence. Since T ( n + 1 ) ( 0 ) = ( n + 1 ) n T^{(n+1)}(0)=(n+1)^n T ( n + 1 ) ( 0 ) = ( n + 1 ) n , we obtain T ∈ C T \in C T ∈ C . Applying Theorem 4.4 successively to these exponents, we obtain the assertion for all iterates. For the Stieltjes moment property of ( ( n + 1 ) n ) ((n+1)^n) (( n + 1 ) n ) , see also [39 , 48 ] .
Taking L ℓ = T ∘ d ℓ L_\ell=T^{\circ d_\ell} L ℓ = T ∘ d ℓ for d ℓ ≥ 1 d_\ell\geq1 d ℓ ≥ 1 , in Corollary 5.1, we obtain the iterated-tree families. Taking A = 1 A=1 A = 1 and d ℓ = 1 d_\ell=1 d ℓ = 1 , we recover Theorem 1.5 . As before, we may identify variables by substitution.
For example, take the initial measure 1 2 δ 0 + 1 2 δ 1 \frac{1}{2}\delta_0+\frac{1}{2}\delta_1 2 1 δ 0 + 2 1 δ 1 , with formal generating function
A ( t ) = 1 + e t 2 , (160) A(t)=\frac{1+e^t}{2}, \tag*{(160)} A ( t ) = 2 1 + e t , (160)
and, applying Corollary 5.1 to the two tree exponents, we obtain coefficientwise Hankel total positivity for
∑ n ≥ 0 F n ( y , z ) t n n ! = 1 + e t 2 ( 1 − T ( T ( t ) ) ) − y ( 1 − 2 T ( t ) ) − z . (161) \sum_{n\geq0} F_n(y,z)\frac{t^n}{n!}
=\frac{1+e^t}{2}(1-T(T(t)))^{-y}(1-2T(t))^{-z}. \tag*{(161)} n ≥ 0 ∑ F n ( y , z ) n ! t n = 2 1 + e t ( 1 − T ( T ( t )) ) − y ( 1 − 2 T ( t ) ) − z . (161)
For this family, the first moments and the first nontrivial leading Hankel minor are
F 0 = 1 , F 1 = 1 2 + y + 2 z , F_0=1,\qquad F_1=\frac{1}{2}+y+2z, F 0 = 1 , F 1 = 2 1 + y + 2 z ,
F 2 = ( y + 2 z ) 2 + 6 y + 10 z + 1 2 , F 0 F 2 − F 1 2 = 1 4 + 5 y + 8 z . (162) F_2=(y+2z)^2+6y+10z+\frac{1}{2},\qquad F_0F_2-F_1^2=\frac{1}{4}+5y+8z. \tag*{(162)} F 2 = ( y + 2 z ) 2 + 6 y + 10 z + 2 1 , F 0 F 2 − F 1 2 = 4 1 + 5 y + 8 z . (162)
The initial moment series need not be the exponential of an admissible exponent. For H ( t ) = log ( ( 1 + e t ) / 2 ) H(t)=\log((1+e^t)/2) H ( t ) = log (( 1 + e t ) /2 ) , we have
H ′ ( 0 ) = 1 2 , H ′ ′ ( 0 ) = 1 4 , H ′ ′ ′ ( 0 ) = 0 , det ( H ′ ( 0 ) H ′ ′ ( 0 ) H ′ ′ ( 0 ) H ′ ′ ′ ( 0 ) ) = − 1 16 . (163) H'(0)=\frac{1}{2},\qquad H''(0)=\frac{1}{4},\qquad H'''(0)=0,\qquad
\det\begin{pmatrix} H'(0) & H''(0) \\ H''(0) & H'''(0) \end{pmatrix}=-\frac{1}{16}. \tag*{(163)} H ′ ( 0 ) = 2 1 , H ′′ ( 0 ) = 4 1 , H ′′′ ( 0 ) = 0 , det ( H ′ ( 0 ) H ′′ ( 0 ) H ′′ ( 0 ) H ′′′ ( 0 ) ) = − 16 1 . (163)
Thus H ∉ C H\notin C H ∈ / C , although its exponential is an allowed initial moment series in Theorem 1.7 .
Gaussian perturbations and centered component families`
Gaussian perturbations and centered component families
We next use the real-line theorem to allow an arbitrary initial Hamburger measure, real centering drifts, and an independent Gaussian component in the functional-digraph construction.
Corollary 5.3 (Centered and Gaussian-perturbed tree families). Fix integers d ℓ ≥ 1 d_\ell\ge1 d ℓ ≥ 1 , real numbers q ℓ ≥ 0 q_\ell\ge0 q ℓ ≥ 0 , β ℓ ∈ R \beta_\ell\in\mathbb{R} β ℓ ∈ R , and σ 2 ≥ 0 \sigma^2 \ge0 σ 2 ≥ 0 . Let A ( t ) A(t) A ( t ) be the formal exponential generating function of any initial Hamburger moment sequence ( a n ) (a_n) ( a n ) . Define G n ( y , z ) G_n(\mathbf{y},z) G n ( y , z ) by
∑ n ≥ 0 G n ( y , z ) t n n ! = A ( t ) exp ( σ 2 z t 2 2 − t ∑ ℓ = 1 m β ℓ y ℓ ) ∏ ℓ = 1 m ( 1 − q ℓ T ∘ d ℓ ( t ) ) − y ℓ . (164) \sum_{n\ge0}G_n(\mathbf{y},z)\frac{t^n}{n!}=A(t)\exp\left(\frac{\sigma^2zt^2}{2}-t\sum_{\ell=1}^{m}\beta_\ell y_\ell\right)\prod_{\ell=1}^{m}\left(1-q_\ell T^{\circ d_\ell}(t)\right)^{-y_\ell}. \tag*{(164)} n ≥ 0 ∑ G n ( y , z ) n ! t n = A ( t ) exp ( 2 σ 2 z t 2 − t ℓ = 1 ∑ m β ℓ y ℓ ) ℓ = 1 ∏ m ( 1 − q ℓ T ∘ d ℓ ( t ) ) − y ℓ . (164) For r ≥ 1 r \ge1 r ≥ 1 , we have
H r ( y , z ) : = det [ G i + j ( y , z ) ] i , j = 0 r − 1 ∈ R ≥ 0 [ y , z ] . (165) H_r(\mathbf{y},z):=\det\left[G_{i+j}(\mathbf{y},z)\right]_{i,j=0}^{r-1}\in\mathbb{R}_{\ge0}[\mathbf{y},z]. \tag*{(165)} H r ( y , z ) := det [ G i + j ( y , z ) ] i , j = 0 r − 1 ∈ R ≥ 0 [ y , z ] . (165) With D = ( r 2 ) D=\binom{r}{2} D = ( 2 r ) , its total degree is at most D D D , and
[ H r ] D = a 0 r ( ∏ j = 0 r − 1 j ! ) ( σ 2 z + ∑ ℓ = 1 m ( 2 d ℓ q ℓ + q ℓ 2 ) y ℓ ) D . (166) [H_r]_D=a_0^r\left(\prod_{j=0}^{r-1}j!\right)\left(\sigma^2z+\sum_{\ell=1}^{m}(2d_\ell q_\ell+q_\ell^2)y_\ell\right)^D. \tag*{(166)} [ H r ] D = a 0 r ( j = 0 ∏ r − 1 j ! ) ( σ 2 z + ℓ = 1 ∑ m ( 2 d ℓ q ℓ + q ℓ 2 ) y ℓ ) D . (166) The determinant is independent of β 1 , … , β m \beta_1,\ldots,\beta_m β 1 , … , β m .
Proof. Put C ℓ = − log ( 1 − q ℓ T ∘ d ℓ ) C_\ell=-\log(1-q_\ell T^{\circ d_\ell}) C ℓ = − log ( 1 − q ℓ T ∘ d ℓ ) . Applying Corollary 5.1 to the iterates from Proposition 5.2, we obtain C ℓ ∈ C C_\ell\in\mathcal{C} C ℓ ∈ C . If ν ℓ \nu_\ell ν ℓ represents its once-shifted cumulants, then x ν ℓ ( d x ) x\nu_\ell(dx) x ν ℓ ( d x ) represents its twice-shifted cumulants. Subtracting β ℓ t \beta_\ell t β ℓ t leaves these unchanged. The Gaussian exponent has twice-shifted measure σ 2 δ 0 \sigma^2\delta_0 σ 2 δ 0 . Applying Theorem 1.8 to this exponent list and A A A , we obtain (5.16), the degree bound, and drift invariance. Using the chain rule, we obtain ( T ∘ d ) ′ ( 0 ) = 1 (T^{\circ d})'(0)=1 ( T ∘ d ) ′ ( 0 ) = 1 , ( T ∘ d ) ′ ′ ( 0 ) = 2 d (T^{\circ d})''(0)=2d ( T ∘ d ) ′′ ( 0 ) = 2 d , hence C ℓ ′ ′ ( 0 ) = 2 d ℓ q ℓ + q ℓ 2 C_\ell''(0)=2d_\ell q_\ell+q_\ell^2 C ℓ ′′ ( 0 ) = 2 d ℓ q ℓ + q ℓ 2 . Substituting these derivatives in (1.13), we obtain (5.17).
Choosing β ℓ = q ℓ \beta_\ell=q_\ell β ℓ = q ℓ centers each added component family. So the first moment of the output is then G 1 = a 1 G_1=a_1 G 1 = a 1 . For y ≥ 0 \mathbf{y}\ge0 y ≥ 0 and z ≥ 0 z\ge0 z ≥ 0 , we also obtain a Hamburger moment sequence. In the coefficientwise assertion, we consider leading principal determinants, as in the real-line theorem.
Implicit and autonomous equations We will first characterize the implicit equations whose solutions belong to C \mathcal{C} C . The criteria [49 Theorem 4.2] and [48 Theorem 6.6] then supply moment inputs for universal families with independent colors and powers of prefactors.
Implicit tree functions and an exact lifting criterion
We now replace the Cayley tree equation (2) by a general implicit equation. Let Φ ( u ) ∈ R [ [ u ] ] \Phi(u) \in\mathbb{R}[[u]] Φ ( u ) ∈ R [[ u ]] with Φ ( 0 ) > 0 \Phi(0)>0 Φ ( 0 ) > 0 , and define the unique formal power series U ( t ) ∈ t R [ [ t ] ] U(t) \in t\mathbb{R}[[t]] U ( t ) ∈ t R [[ t ]] by
U ( t ) = t Φ ( U ( t ) ) . (167) U(t)=t\Phi(U(t)). \tag*{(167)} U ( t ) = t Φ ( U ( t )) . (167)
When Φ ( u ) = ∑ k ≥ 0 w k u k / k ! \Phi(u)=\sum_{k\geq0}w_k u^k/k! Φ ( u ) = ∑ k ≥ 0 w k u k / k ! with w k ≥ 0 w_k\geq0 w k ≥ 0 , we interpret U U U as the exponential generating function of labeled rooted trees with weight w k w_k w k at a vertex with k k k children. To compute the coefficients, we recall the classical Lagrange inversion formula; see Stanley [41 Section 5.4] .
Lemma 6.1 (Lagrange inversion). Under (6.1), for n ≥ 1 n\geq1 n ≥ 1 and H ( u ) ∈ R [ [ u ] ] H(u)\in\mathbb{R}[[u]] H ( u ) ∈ R [[ u ]] , we have
[ t n ] H ( U ( t ) ) = 1 n [ u n − 1 ] H ′ ( u ) Φ ( u ) n . (168) [t^n]H(U(t))=\frac{1}{n}[u^{n-1}]H'(u)\Phi(u)^n. \tag*{(168)} [ t n ] H ( U ( t )) = n 1 [ u n − 1 ] H ′ ( u ) Φ ( u ) n . (168) In particular, if U ( t ) = ∑ n ≥ 1 τ n t n / n ! U(t)=\sum_{n\geq1}\tau_n t^n/n! U ( t ) = ∑ n ≥ 1 τ n t n / n ! , then
τ n + 1 = n ! [ u n ] Φ ( u ) n + 1 ( n ≥ 0 ) . (169) \tau_{n+1}=n![u^n]\Phi(u)^{n+1}\qquad(n\geq0). \tag*{(169)} τ n + 1 = n ! [ u n ] Φ ( u ) n + 1 ( n ≥ 0 ) . (169)
Applying Theorem 1.7 to the shifted coefficients (6.3), we obtain the following exact criterion for logarithmic components of an implicit tree series.
Theorem 6.2 (Implicit-tree lifting criterion). For U U U in (6.1), the following conditions are equivalent:
(i) The sequence ( n ! [ u n ] Φ ( u ) n + 1 ) n ≥ 0 \left(n![u^n]\Phi(u)^{n+1}\right)_{n\geq0} ( n ! [ u n ] Φ ( u ) n + 1 ) n ≥ 0 is a Stieltjes moment sequence.
(ii) For every fixed q > 0 q>0 q > 0 , the sequence ( Ψ n U ( q , y ) ) n (\Psi_n^U(q,y))_n ( Ψ n U ( q , y ) ) n defined by
∑ n ≥ 0 Ψ n U ( q , y ) t n n ! = ( 1 − q U ( t ) ) − y (170) \sum_{n\geq0}\Psi_n^U(q,y)\frac{t^n}{n!}=(1-qU(t))^{-y} \tag*{(170)} n ≥ 0 ∑ Ψ n U ( q , y ) n ! t n = ( 1 − q U ( t ) ) − y (170) is coefficientwise Hankel-totally positive in y y y .
(iii) For every m ≥ 1 m\geq1 m ≥ 1 , fixed q 1 , … , q m ≥ 0 q_1,\ldots,q_m\geq0 q 1 , … , q m ≥ 0 , and real initial Stieltjes moment series A A A , the sequence defined by
∑ n ≥ 0 F n ( y ) t n n ! = A ( t ) ∏ ℓ = 1 m ( 1 − q ℓ U ( t ) ) − y ℓ (171) \sum_{n\geq0}F_n(y)\frac{t^n}{n!}=A(t)\prod_{\ell=1}^{m}(1-q_\ell U(t))^{-y\ell} \tag*{(171)} n ≥ 0 ∑ F n ( y ) n ! t n = A ( t ) ℓ = 1 ∏ m ( 1 − q ℓ U ( t ) ) − y ℓ (171) is coefficientwise Hankel-totally positive in y y y .
Proof. By (6.3), condition (i) is equivalent to U ∈ C U\in\mathcal{C} U ∈ C . Applying Corollary 5.1 to U U U , we obtain (i)⇒ \Rightarrow ⇒ (iii). Taking A = 1 A=1 A = 1 and m = 1 m=1 m = 1 , we obtain (iii)⇒ \Rightarrow ⇒ (ii). Assume (i) holds and let C q U = − log ( 1 − q U ) C_q^U=-\log(1-qU) C q U = − log ( 1 − q U ) . Applying Theorem 1.6 to C q U C_q^U C q U and scaling the representing measure by 1 / q 1/q 1/ q , we obtain the Stieltjes sequence ( ( C q U ) ( n + 1 ) ( 0 ) / q ) n ≥ 0 ((C_q^U)^{(n+1)}(0)/q)_{n\geq0} (( C q U ) ( n + 1 ) ( 0 ) / q ) n ≥ 0 for every q > 0 q>0 q > 0 . Expanding the formal logarithm, we have
C q U ( t ) q = U ( t ) + ∑ k ≥ 2 q k − 1 k U ( t ) k , lim q → 0 ( C q U ) ( n + 1 ) ( 0 ) q = τ n + 1 . (172) \frac{C_q^U(t)}{q}=U(t)+\sum_{k\geq2}\frac{q^{k-1}}{k}U(t)^k,\qquad\lim_{q\to0}\frac{(C_q^U)^{(n+1)}(0)}{q}=\tau_{n+1}. \tag*{(172)} q C q U ( t ) = U ( t ) + k ≥ 2 ∑ k q k − 1 U ( t ) k , q → 0 lim q ( C q U ) ( n + 1 ) ( 0 ) = τ n + 1 . (172)
Each finite Hankel minor is a polynomial in finitely many entries, so its limit is nonnegative. Applying Lemma 2.1 to the limiting sequence, we obtain (i). □
To show why the moment condition is needed, we give an example with nonnegative tree weights for which the positivity assertion fails.
Implicit equations with Pólya frequency series
A sufficient moment condition from implicit-equation criterion was showed in [49 ] . For the related branched-continued-fraction approach, see Pétréolle, Sokal and Zhu [31, Theorems 9.8 and 9.9]. We distinguish ordinary coefficients in this input from the exponential coefficients required by lifting.
For a real formal power series h ( u ) = ∑ n ≥ 0 h n u n h(u)=\sum_{n\geq0}h_nu^n h ( u ) = ∑ n ≥ 0 h n u n , put h n = 0 h_n=0 h n = 0 for n < 0 n<0 n < 0 . We call h h h a Pólya frequency series , or a PF series, if the Toeplitz matrix [ h i − j ] i , j ≥ 0 [h_{i-j}]_{i,j\geq0} [ h i − j ] i , j ≥ 0 is totally positive. In particular, h n ≥ 0 h_n\geq0 h n ≥ 0 . By the Cauchy–Binet formula, we preserve this property when multiplying two such series, since the corresponding lower triangular Toeplitz matrices multiply by ordinary convolution.
For the ordinary-coefficient input, we use the real, infinite-order case of the criterion [49 pp. 19–20, Theorems 4.2 and 4.2′, Eq. (4.12)] . We record its specialization and coefficient convention explicitly as follows:
Lemma 6.3 (Implicit-equation criterion). Let φ ( u ) , ψ ( u ) ∈ R [ [ u ] ] \varphi(u),\psi(u) \in\mathbb{R}[[u]] φ ( u ) , ψ ( u ) ∈ R [[ u ]] be PF series with φ ( 0 ) ψ ( 0 ) > 0 \varphi(0)\psi(0)>0 φ ( 0 ) ψ ( 0 ) > 0 , and let U ( t ) = t φ ( U ( t ) ) ψ ( U ( t ) ) U(t)=t\varphi(U(t))\psi(U(t)) U ( t ) = tφ ( U ( t )) ψ ( U ( t )) . Define
R n ( q ) : = [ t n ] ψ ( U ( t ) ) 1 − q U ( t ) ( n ≥ 0 ) . (175) R_n(q):=[t^n]\frac{\psi(U(t))}{1-qU(t)} \qquad(n\geq0). \tag*{(175)} R n ( q ) := [ t n ] 1 − q U ( t ) ψ ( U ( t )) ( n ≥ 0 ) . (175) Then ( R n ( q ) ) n ≥ 0 (R_n(q))_{n\geq0} ( R n ( q ) ) n ≥ 0 is coefficientwise Hankel-totally positive in q q q . In particular, for n ≥ 0 n\geq0 n ≥ 0 , [ t n ] ψ ( U ( t ) ) [t^n]\psi(U(t)) [ t n ] ψ ( U ( t )) form a real Stieltjes moment sequence ( b n ) n ≥ 0 (b_n)_{n\geq0} ( b n ) n ≥ 0 .
Applying the lifting criterion to this moment input, we obtain the following universal implicit-equation family.
Theorem 6.4 (Universal lifting for implicit Pólya frequency equations). For 1 ≤ ℓ ≤ m 1\leq\ell\leq m 1 ≤ ℓ ≤ m , fix real PF series φ ℓ ( u ) , ψ ℓ ( u ) \varphi_\ell(u),\psi_\ell(u) φ ℓ ( u ) , ψ ℓ ( u ) with φ ℓ ( 0 ) ψ ℓ ( 0 ) > 0 \varphi_\ell(0)\psi_\ell(0)>0 φ ℓ ( 0 ) ψ ℓ ( 0 ) > 0 , and define
U ℓ ( t ) = t φ ℓ ( U ℓ ( t ) ) ψ ℓ ( U ℓ ( t ) ) . (176) U_\ell(t)=t\varphi_\ell(U_\ell(t))\psi_\ell(U_\ell(t)). \tag*{(176)} U ℓ ( t ) = t φ ℓ ( U ℓ ( t )) ψ ℓ ( U ℓ ( t )) . (176) Then the following assertions hold.
(i) Each U ℓ ∈ C U_\ell\in\mathcal{C} U ℓ ∈ C ; equivalently, with Φ ℓ = φ ℓ ψ ℓ \Phi_\ell=\varphi_\ell\psi_\ell Φ ℓ = φ ℓ ψ ℓ , the sequence
( n ! [ u n ] Φ ℓ ( u ) n + 1 ) n ≥ 0 = ( U ℓ ( n + 1 ) ( 0 ) ) n ≥ 0 (177) \left(n![u^n]\Phi_\ell(u)^{n+1}\right)_{n\geq0}=\left(U_\ell^{(n+1)}(0)\right)_{n\geq0} \tag*{(177)} ( n ! [ u n ] Φ ℓ ( u ) n + 1 ) n ≥ 0 = ( U ℓ ( n + 1 ) ( 0 ) ) n ≥ 0 (177) is a Stieltjes moment sequence.
(ii) For any initial Stieltjes moment series A ( t ) A(t) A ( t ) and fixed q 1 , … , q m ≥ 0 q_1,\ldots,q_m\geq0 q 1 , … , q m ≥ 0 , define
∑ n ≥ 0 F n ( x , y ) t n n ! = A ( t ) ∏ ℓ = 1 m { ψ ℓ ( U ℓ ( t ) ) e x ℓ U ℓ ( t ) ( 1 − q ℓ U ℓ ( t ) ) − y ℓ } . (178) \sum_{n\geq0}F_n(x,y)\frac{t^n}{n!}
=A(t)\prod_{\ell=1}^{m}\left\{\psi_\ell(U_\ell(t))e^{x_\ell U_\ell(t)}(1-q_\ell U_\ell(t))^{-y_\ell}\right\}. \tag*{(178)} n ≥ 0 ∑ F n ( x , y ) n ! t n = A ( t ) ℓ = 1 ∏ m { ψ ℓ ( U ℓ ( t )) e x ℓ U ℓ ( t ) ( 1 − q ℓ U ℓ ( t ) ) − y ℓ } . (178) Then ( F n ) (F_n) ( F n ) is jointly coefficientwise Hankel-totally positive in x , y x,y x , y . We may omit any of the prefactors ψ ℓ ( U ℓ ( t ) ) \psi_\ell(U_\ell(t)) ψ ℓ ( U ℓ ( t )) and retain the same conclusion.
Proof. Step 1: Ordinary shifted moments. Suppress the color index and put Φ = φ ψ \Phi=\varphi\psi Φ = φ ψ . Applying the Toeplitz product rule to φ \varphi φ and ψ \psi ψ , we obtain Φ ∈ P F \Phi\in\mathrm{PF} Φ ∈ PF . Applying Lemma 6.3 to the factorization Φ = Φ ⋅ 1 \Phi=\Phi\cdot1 Φ = Φ ⋅ 1 , we obtain the resolvent polynomials
Q n ( q ) : = [ t n ] ( 1 − q U ( t ) ) − 1 , Q n ( q ) = q [ t n ] U ( t ) + O ( q 2 ) ( n ≥ 1 ) . (179) Q_n(q):=[t^n](1-qU(t))^{-1},\qquad Q_n(q)=q[t^n]U(t)+O(q^2)\qquad(n\geq1). \tag*{(179)} Q n ( q ) := [ t n ] ( 1 − q U ( t ) ) − 1 , Q n ( q ) = q [ t n ] U ( t ) + O ( q 2 ) ( n ≥ 1 ) . (179)
Here O ( q 2 ) O(q^2) O ( q 2 ) denotes divisibility by q 2 q^2 q 2 . For increasing nonnegative index lists I , J I,J I , J of length r r r , each entry of the shifted minor has zero constant term. Hence
[ q r ] det [ Q i a + j b + 1 ( q ) ] a , b = 1 r = det [ [ t i a + j b + 1 ] U ( t ) ] a , b = 1 r ≥ 0. (180) [q^r]\det\left[Q_{i_a+j_b+1}(q)\right]_{a,b=1}^{r}
=\det\left[[t^{i_a+j_b+1}]U(t)\right]_{a,b=1}^{r}\geq0. \tag*{(180)} [ q r ] det [ Q i a + j b + 1 ( q ) ] a , b = 1 r = det [ [ t i a + j b + 1 ] U ( t ) ] a , b = 1 r ≥ 0. (180)
Thus d n = [ t n + 1 ] U ( t ) d_n=[t^{n+1}]U(t) d n = [ t n + 1 ] U ( t ) has nonnegative Hankel minors and, by Lemma 2.1 , a Stieltjes measure ν \nu ν .
Step 2: Exponential coefficients and the prefactor. Applying Euler’s integral to the product measure, we obtain
U ( n + 1 ) ( 0 ) = ( n + 1 ) ! d n = ∫ 0 ∞ ∫ 0 ∞ ( s x ) n s e − s d s ν ( d x ) . (181) U^{(n+1)}(0)=(n+1)!d_n=\int_0^\infty\int_0^\infty(sx)^nse^{-s}\,\mathrm{d}s\,\nu(\mathrm{d}x). \tag*{(181)} U ( n + 1 ) ( 0 ) = ( n + 1 )! d n = ∫ 0 ∞ ∫ 0 ∞ ( s x ) n s e − s d s ν ( d x ) . (181)
Taking the image of ν ( d x ) s e − s d s \nu(\mathrm{d}x)se^{-s}\,\mathrm{d}s ν ( d x ) s e − s d s under ( x , s ) ↦ s x (x,s)\mapsto sx ( x , s ) ↦ s x , we represent the once-shifted cumulants and obtain U ∈ C U\in\mathcal{C} U ∈ C . Applying Lagrange inversion to U = t Φ ( U ) U=t\Phi(U) U = t Φ ( U ) , we obtain (177) .
Applying Lemma 6.3 to the original factorization at q = 0 q=0 q = 0 yields a measure σ \sigma σ representing b n = [ t n ] ψ ( U ( t ) ) b_n=[t^n]\psi(U(t)) b n = [ t n ] ψ ( U ( t )) . We then have
n ! b n = ∫ 0 ∞ ∫ 0 ∞ ( s x ) n e − s d s σ ( d x ) ( n ≥ 0 ) . (182) n!b_n=\int_0^\infty\int_0^\infty(sx)^n e^{-s}\,\mathrm{d}s\,\sigma(\mathrm{d}x)\qquad(n\geq0). \tag*{(182)} n ! b n = ∫ 0 ∞ ∫ 0 ∞ ( s x ) n e − s d s σ ( d x ) ( n ≥ 0 ) . (182)
These are the exponential coefficients of the prefactor. The distinct gamma weights in (181) and (182) account for the factors ( n + 1 ) ! (n+1)! ( n + 1 )! and n ! n! n ! , respectively.
Step 3: Initial moments and lifting. For moment series represented by μ 1 , μ 2 \mu_1,\mu_2 μ 1 , μ 2 , binomial convolution is the image measure under addition:
∑ k = 0 n ( n k ) ( ∫ 0 ∞ x k μ 1 ( d x ) ) ( ∫ 0 ∞ z n − k μ 2 ( d z ) ) = ∫ 0 ∞ ∫ 0 ∞ ( x + z ) n μ 1 ( d x ) μ 2 ( d z ) . (183) \sum_{k=0}^{n}\binom{n}{k}\left(\int_0^\infty x^k\,\mu_1(\mathrm{d}x)\right)\left(\int_0^\infty z^{n-k}\,\mu_2(\mathrm{d}z)\right)=\int_0^\infty\int_0^\infty(x+z)^n\,\mu_1(\mathrm{d}x)\mu_2(\mathrm{d}z). \tag*{(183)} k = 0 ∑ n ( k n ) ( ∫ 0 ∞ x k μ 1 ( d x ) ) ( ∫ 0 ∞ z n − k μ 2 ( d z ) ) = ∫ 0 ∞ ∫ 0 ∞ ( x + z ) n μ 1 ( d x ) μ 2 ( d z ) . (183)
In consequence, B ( t ) = A ( t ) ∏ ℓ ψ ℓ ( U ℓ ( t ) ) B(t)=A(t)\prod_{\ell}\psi_\ell(U_\ell(t)) B ( t ) = A ( t ) ∏ ℓ ψ ℓ ( U ℓ ( t )) is an initial Stieltjes moment series. By Step 2 and Corollary 5.1, both U ℓ U_\ell U ℓ and − log ( 1 − q ℓ U ℓ ) -\log(1-q_\ell U_\ell) − log ( 1 − q ℓ U ℓ ) belong to C \mathcal{C} C . Applying Theorem 1.7 to these 2 m 2m 2 m exponents and B B B , we obtain (ii). Omitting prefactors simply removes the corresponding factors from B B B .
For polynomial equation coefficients, we may also prescribe a nonzero initial value. For the corresponding ordinary Riordan-array results, see [49 Theorem 6.1] . We now give the universal cumulant-lifting counterpart.
Corollary 6.5 (A nonzero polynomial initial value). Fix κ ≥ 0 \kappa\geq0 κ ≥ 0 and ρ ( u ) = ∏ h = 1 d ( α h + β h u ) \rho(u)=\prod_{h=1}^{d}(\alpha_h+\beta_hu) ρ ( u ) = ∏ h = 1 d ( α h + β h u ) , where α h , β h ≥ 0 \alpha_h,\beta_h\geq0 α h , β h ≥ 0 , and ρ ( κ ) > 0 \rho(\kappa)>0 ρ ( κ ) > 0 . Define the unique formal power series
f ( t ) = κ + t ρ ( f ( t ) ) , U ( t ) = f ( t ) − κ . (184) f(t)=\kappa+t\rho(f(t)),\qquad U(t)=f(t)-\kappa. \tag*{(184)} f ( t ) = κ + tρ ( f ( t )) , U ( t ) = f ( t ) − κ . (184) Then U ∈ C U\in\mathcal{C} U ∈ C . In addition, for every real initial Stieltjes moment series A ( t ) A(t) A ( t ) and fixed q 1 , … , q m ≥ 0 q_1,\ldots,q_m\geq0 q 1 , … , q m ≥ 0 , the exponential coefficients of
A ( t ) exp ( ∑ ℓ = 1 m x ℓ U ( t ) ) ∏ ℓ = 1 m ( 1 − q ℓ U ( t ) ) − y ℓ (185) A(t)\exp\left(\sum_{\ell=1}^{m}x_\ell U(t)\right)\prod_{\ell=1}^{m}(1-q_\ell U(t))^{-y_\ell} \tag*{(185)} A ( t ) exp ( ℓ = 1 ∑ m x ℓ U ( t ) ) ℓ = 1 ∏ m ( 1 − q ℓ U ( t ) ) − y ℓ (185) are coefficientwise Hankel-totally positive jointly in x , y x,y x , y .
Proof. Subtracting κ \kappa κ in (184) , we obtain
U ( t ) = t ρ ~ ( U ( t ) ) , ρ ~ ( u ) = ∏ h = 1 d ( α h + κ β h + β h u ) . (186) U(t)=t\widetilde{\rho}(U(t)),\qquad\widetilde{\rho}(u)=\prod_{h=1}^{d}(\alpha_h+\kappa\beta_h+\beta_hu). \tag*{(186)} U ( t ) = t ρ ( U ( t )) , ρ ( u ) = h = 1 ∏ d ( α h + κ β h + β h u ) . (186)
The Toeplitz matrix of each nonnegative linear factor is bidiagonal and totally positive, by Lemma 2.4 . By the Toeplitz product rule, ρ ~ \widetilde{\rho} ρ is PF, and ρ ~ ( 0 ) = ρ ( κ ) > 0 \widetilde{\rho}(0)=\rho(\kappa)>0 ρ ( 0 ) = ρ ( κ ) > 0 . Applying Theorem 6.4 to the factorization ρ ~ = ρ ~ ⋅ 1 \widetilde{\rho}=\widetilde{\rho}\cdot1 ρ = ρ ⋅ 1 , we obtain the result.
In this corollary, we substitute into a polynomial. For a general formal power series ρ ( u ) \rho(u) ρ ( u ) , substitution at a nonzero κ \kappa κ requires a separate definition or suitable convergence hypotheses; see [49 Remark 1.2] .
Autonomous differential equations with Pólya frequency series
We next obtain an analogous universal result for autonomous differential equations. As the known input, we use [48 Theorem 6.6 (iv) and (iii)] with real coefficients. If V ‾ \overline{V} V denotes the formal compositional inverse, then 1 / V ‾ ′ ( u ) = φ ( u ) ψ ( u ) 1/\overline{V}'(u)=\varphi(u)\psi(u) 1/ V ′ ( u ) = φ ( u ) ψ ( u ) is equivalent to V ′ ( t ) = φ ( V ( t ) ) ψ ( V ( t ) ) V'(t)=\varphi(V(t))\psi(V(t)) V ′ ( t ) = φ ( V ( t )) ψ ( V ( t )) . We therefore restate the criterion in autonomous-equation form and refer to the cited work for its proof.
Lemma 6.6 (Autonomous-equation criterion). Let φ ( u ) , ψ ( u ) ∈ R [ [ u ] ] \varphi(u),\psi(u)\in\mathbb{R[[u]]} φ ( u ) , ψ ( u ) ∈ R [[ u ]] be PF series with φ ( 0 ) ψ ( 0 ) > 0 \varphi(0)\psi(0)>0 φ ( 0 ) ψ ( 0 ) > 0 . Let V ( 0 ) = 0 V(0)=0 V ( 0 ) = 0 and
V ′ ( t ) = φ ( V ( t ) ) ψ ( V ( t ) ) . (187) V'(t)=\varphi(V(t))\psi(V(t)). \tag*{(187)} V ′ ( t ) = φ ( V ( t )) ψ ( V ( t )) . (187) Then ( V ( n + 1 ) ( 0 ) ) n ≥ 0 (V^{(n+1)}(0))_{n\geq0} ( V ( n + 1 ) ( 0 ) ) n ≥ 0 is a Stieltjes moment sequence. Moreover, the exponential coefficients of φ ( V ( t ) ) e λ V ( t ) \varphi(V(t))e^{\lambda V(t)} φ ( V ( t )) e λV ( t ) form a coefficientwise Hankel-totally positive sequence in λ \lambda λ . In particular, the exponential coefficients of φ ( V ( t ) ) \varphi(V(t)) φ ( V ( t )) form a real Stieltjes moment sequence.
Applying the autonomous criterion to the equations below, we obtain the parallel universal family.
Theorem 6.7 (Universal lifting for autonomous Pólya frequency equations). For 1 ≤ ℓ ≤ m 1\leq\ell\leq m 1 ≤ ℓ ≤ m , fix real PF series φ ℓ ( u ) , ψ ℓ ( u ) \varphi_\ell(u),\psi_\ell(u) φ ℓ ( u ) , ψ ℓ ( u ) with φ ℓ ( 0 ) ψ ℓ ( 0 ) > 0 \varphi_\ell(0)\psi_\ell(0)>0 φ ℓ ( 0 ) ψ ℓ ( 0 ) > 0 , and define the unique formal power series V ℓ ( t ) V_\ell(t) V ℓ ( t ) by
V ℓ ′ ( t ) = φ ℓ ( V ℓ ( t ) ) ψ ℓ ( V ℓ ( t ) ) , V ℓ ( 0 ) = 0. (188) V_\ell'(t)=\varphi_\ell(V_\ell(t))\psi_\ell(V_\ell(t)),\qquad V_\ell(0)=0. \tag*{(188)} V ℓ ′ ( t ) = φ ℓ ( V ℓ ( t )) ψ ℓ ( V ℓ ( t )) , V ℓ ( 0 ) = 0. (188) Then the following assertions hold:
V ℓ ∈ C V_\ell\in\mathcal{C} V ℓ ∈ C for every ℓ \ell ℓ .
(ii) For any initial Stieltjes moment series A ( t ) A(t) A ( t ) and fixed q 1 , … , q m ≥ 0 q_1,\ldots,q_m \ge0 q 1 , … , q m ≥ 0 , define
∑ n ≥ 0 F n ( x , y ) t n n ! = A ( t ) ∏ ℓ = 1 m { φ ℓ ( V ℓ ( t ) ) e x ℓ V ℓ ( t ) ( 1 − q ℓ V ℓ ( t ) ) − y ℓ } . (189) \sum_{n\ge0} F_n(\mathbf{x},\mathbf{y})\frac{t^n}{n!}=A(t)\prod_{\ell=1}^{m}\left\{\varphi_\ell(V_\ell(t))e^{x_\ell V_\ell(t)}(1-q_\ell V_\ell(t))^{-y_\ell}\right\}. \tag*{(189)} n ≥ 0 ∑ F n ( x , y ) n ! t n = A ( t ) ℓ = 1 ∏ m { φ ℓ ( V ℓ ( t )) e x ℓ V ℓ ( t ) ( 1 − q ℓ V ℓ ( t ) ) − y ℓ } . (189)
Then ( F n ) (F_n) ( F n ) is coefficientwise Hankel-totally positive jointly in x , y \mathbf{x},\mathbf{y} x , y . We may omit any of the prefactors φ ℓ ( V ℓ ( t ) ) \varphi_\ell(V_\ell(t)) φ ℓ ( V ℓ ( t )) and retain the same conclusion.
Proof. Applying Lemma 6.6 to each equation, we obtain V ℓ ∈ C V_\ell\in\mathcal{C} V ℓ ∈ C and a Stieltjes moment series φ ℓ ( V ℓ ( t ) ) \varphi_\ell(V_\ell(t)) φ ℓ ( V ℓ ( t )) . By (183) , B ( t ) = A ( t ) ∏ ℓ φ ℓ ( V ℓ ( t ) ) B(t)=A(t)\prod_\ell\varphi_\ell(V_\ell(t)) B ( t ) = A ( t ) ∏ ℓ φ ℓ ( V ℓ ( t )) is also a Stieltjes moment series. Applying Corollary 5.1 to V ℓ V_\ell V ℓ , we obtain − log ( 1 − q ℓ V ℓ ) ∈ C -\log(1-q_\ell V_\ell)\in\mathcal{C} − log ( 1 − q ℓ V ℓ ) ∈ C . Hence, by Theorem 1.7 , we obtain the asserted joint positivity. The same argument allows any prefactors to be omitted.
Independent powers of Riordan prefactors
We now retain independent powers of admissible prefactors. Applying the classical logarithm criterion in Proposition 4.2 to the ordinary moment inputs below, we obtain the exponents needed for the implicit case.
We apply Lemmas 6.3 and 6.6 to positive linear factors in either equation.
Theorem 6.8 (Independent powers and universal lifting). Fix integers p , r ≥ 0 p,r\ge0 p , r ≥ 0 , c 1 , … , c p ≥ 0 c_1,\ldots,c_p\ge0 c 1 , … , c p ≥ 0 and d 1 , … , d r ≥ 0 d_1,\ldots,d_r\ge0 d 1 , … , d r ≥ 0 . Let Γ ( u ) \Gamma(u) Γ ( u ) be a real PF series with Γ ( 0 ) > 0 \Gamma(0)>0 Γ ( 0 ) > 0 , and put
Φ ( u ) = Γ ( u ) ∏ i = 1 p ( 1 + c i u ) . (190) \Phi(u)=\Gamma(u)\prod_{i=1}^{p}(1+c_i u). \tag*{(190)} Φ ( u ) = Γ ( u ) i = 1 ∏ p ( 1 + c i u ) . (190) Let U ( t ) U(t) U ( t ) be the unique formal power series with U ( 0 ) = 0 U(0)=0 U ( 0 ) = 0 satisfying either U ( t ) = t Φ ( U ( t ) ) U(t)=t\Phi(U(t)) U ( t ) = t Φ ( U ( t )) or U ′ ( t ) = Φ ( U ( t ) ) U'(t)=\Phi(U(t)) U ′ ( t ) = Φ ( U ( t )) . Then the following hold.
(i) For 1 ≤ i ≤ p 1\le i\le p 1 ≤ i ≤ p and 1 ≤ j ≤ r 1\le j\le r 1 ≤ j ≤ r , we have
{ U ( t ) , log ( 1 + c i U ( t ) ) , − log ( 1 − d j U ( t ) ) } ⊆ C . (191) \{U(t),\log(1+c_iU(t)),-\log(1-d_jU(t))\}\subseteq\mathcal{C}. \tag*{(191)} { U ( t ) , log ( 1 + c i U ( t )) , − log ( 1 − d j U ( t ))} ⊆ C . (191) (ii) For any initial Stieltjes moment series A ( t ) A(t) A ( t ) and independent intensities y , z = ( z 1 , … , z p ) y,\mathbf{z}=(z_1,\ldots,z_p) y , z = ( z 1 , … , z p ) , w = ( w 1 , … , w r ) \mathbf{w}=(w_1,\ldots,w_r) w = ( w 1 , … , w r ) , define
∑ n ≥ 0 F n ( y , z , w ) t n n ! = A ( t ) e y U ( t ) ∏ i = 1 p ( 1 + c i U ( t ) ) z i ∏ j = 1 r ( 1 − d j U ( t ) ) w j . (192) \sum_{n\ge0}F_n(y,\mathbf{z},\mathbf{w})\frac{t^n}{n!}
=A(t)e^{yU(t)}
\frac{\prod_{i=1}^{p}(1+c_iU(t))^{z_i}}
{\prod_{j=1}^{r}(1-d_jU(t))^{w_j}}. \tag*{(192)} n ≥ 0 ∑ F n ( y , z , w ) n ! t n = A ( t ) e y U ( t ) ∏ j = 1 r ( 1 − d j U ( t ) ) w j ∏ i = 1 p ( 1 + c i U ( t ) ) z i . (192) Then ( F n ) (F_n) ( F n ) is coefficientwise Hankel-totally positive jointly in y , z , w y,\mathbf{z},\mathbf{w} y , z , w . The conclusion also holds for finite products from different equations with independent indeterminates for each equation.
Proof. If c i = 0 c_i=0 c i = 0 , the logarithm is zero. For c i > 0 c_i>0 c i > 0 , write
Ψ i ( u ) = Γ ( u ) ∏ h ≠ i ( 1 + c h u ) , Φ ( u ) = ( 1 + c i u ) Ψ i ( u ) . (193) \Psi_i(u)=\Gamma(u)\prod_{h\ne i}(1+c_hu),\qquad\Phi(u)=(1+c_iu)\Psi_i(u). \tag*{(193)} Ψ i ( u ) = Γ ( u ) h = i ∏ ( 1 + c h u ) , Φ ( u ) = ( 1 + c i u ) Ψ i ( u ) . (193)
Both factors are PF. For the implicit equation, applying Lemma 6.3 to φ = Ψ i \varphi=\Psi_i φ = Ψ i , ψ ( u ) = 1 + c i u \psi(u)=1+c_iu ψ ( u ) = 1 + c i u , with resolvent parameter zero, we obtain the Stieltjes ordinary coefficients of 1 + c i U 1+c_iU 1 + c i U . We have positive product Γ ( 0 ) \Gamma(0) Γ ( 0 ) of the source constant terms and constant term one for 1 + c i U 1+c_iU 1 + c i U . Applying Proposition 4.2 to this ordinary moment series, we obtain log ( 1 + c i U ) ∈ C \log(1+c_iU)\in\mathcal{C} log ( 1 + c i U ) ∈ C .
For the autonomous equation,
d d t log ( 1 + c i U ( t ) ) = c i Φ ( U ( t ) ) 1 + c i U ( t ) = c i Ψ i ( U ( t ) ) . (194) \frac{\mathrm{d}}{\mathrm{d}t}\log(1+c_iU(t))=\frac{c_i\Phi(U(t))}{1+c_iU(t)}=c_i\Psi_i(U(t)). \tag*{(194)} d t d log ( 1 + c i U ( t )) = 1 + c i U ( t ) c i Φ ( U ( t )) = c i Ψ i ( U ( t )) . (194)
Applying Lemma 6.6 to the prefactor Ψ i \Psi_i Ψ i gives the Stieltjes exponential coefficients of Ψ i ( U ) \Psi_i(U) Ψ i ( U ) . Multiplying the representing measure by c i c_i c i , we derive the required shifted moments.
Applying Theorem 6.4 or 6.7 to the given equation, we obtain U ∈ C U\in\mathcal{C} U ∈ C . From Corollary 5.1, we have − log ( 1 − d j U ) ∈ C -\log(1-d_jU)\in\mathcal{C} − log ( 1 − d j U ) ∈ C . Rewrite the product in (6.26) as
A ( t ) exp ( y U + ∑ i z i log ( 1 + c i U ) − ∑ j w j log ( 1 − d j U ) ) . A(t)\exp\left(yU+\sum_i z_i\log(1+c_iU)-\sum_j w_j\log(1-d_jU)\right). A ( t ) exp ( y U + i ∑ z i log ( 1 + c i U ) − j ∑ w j log ( 1 − d j U ) ) .
Hence, by Theorem 1.7, we obtain the conclusion including finite collections of equations with independent indeterminates.
For the underlying examples, coefficientwise Hankel total positivity is already established: for multivariate Fuss–Narayana polynomials, see Pétréolle, Sokal and Zhu [31 Theorem 10.1] ; for the Riordan versions, see Zhu [49 Theorem 5.1] ; for higher Eulerian and network families, see [48 Proposition 6.22 and Theorem 6.25] . We now apply Theorem 6.8 to record the corresponding consequences with independent powers and arbitrary initial moments.
Corollary 6.9 (Fuss–Narayana–Riordan and higher Eulerian families). Let A ( t ) A(t) A ( t ) be any initial Stieltjes moment series. The exponential coefficient sequences below are jointly coefficientwise Hankel-totally positive in the indicated intensities.
(i) For fixed m ≥ 1 m\ge1 m ≥ 1 , c 0 , … , c m c_0,\ldots,c_m c 0 , … , c m , q ≥ 0 q\ge0 q ≥ 0 , let U ( t ) = t ∏ i = 0 m ( 1 + c i U ( t ) ) U(t)=t\prod_{i=0}^{m}(1+c_iU(t)) U ( t ) = t ∏ i = 0 m ( 1 + c i U ( t )) . Then we may take
A ( t ) e y U ( t ) ( 1 − q U ( t ) ) − w ∏ i = 0 m ( 1 + c i U ( t ) ) z i , indeterminates y , w , z 0 , … , z m . (195) A(t)e^{yU(t)}(1-qU(t))^{-w}\prod_{i=0}^{m}(1+c_iU(t))^{z_i},\qquad\text{indeterminates }y,w,z_0,\ldots,z_m. \tag*{(195)} A ( t ) e y U ( t ) ( 1 − q U ( t ) ) − w i = 0 ∏ m ( 1 + c i U ( t ) ) z i , indeterminates y , w , z 0 , … , z m . (195) (ii) For fixed integer r ≥ 1 r\ge1 r ≥ 1 and x , q ≥ 0 x,q\ge0 x , q ≥ 0 , let V ′ ( t ) = ( 1 + V ( t ) ) ( 1 + x V ( t ) ) r V'(t)=(1+V(t))(1+xV(t))^r V ′ ( t ) = ( 1 + V ( t )) ( 1 + x V ( t ) ) r and V ( 0 ) = 0 V(0)=0 V ( 0 ) = 0 . Then we may take
A ( t ) e y V ( t ) ( 1 + V ( t ) ) u ( 1 + x V ( t ) ) v ( 1 − q V ( t ) ) − w , indeterminates y , u , v , w . (196) A(t)e^{yV(t)}(1+V(t))^u(1+xV(t))^v(1-qV(t))^{-w},\qquad\text{indeterminates }y,u,v,w. \tag*{(196)} A ( t ) e y V ( t ) ( 1 + V ( t ) ) u ( 1 + x V ( t ) ) v ( 1 − q V ( t ) ) − w , indeterminates y , u , v , w . (196)
(iii) For fixed a , b ≥ 0 a,b \ge0 a , b ≥ 0 , let S ′ ( t ) = ( 1 + a S ( t ) ) / ( 1 − b S ( t ) ) S'(t)=(1+aS(t))/(1-bS(t)) S ′ ( t ) = ( 1 + a S ( t )) / ( 1 − b S ( t )) and S ( 0 ) = 0 S(0)=0 S ( 0 ) = 0 . Then we may take
A ( t ) e y S ( t ) ( 1 + a S ( t ) ) u ( 1 − b S ( t ) ) − v , indeterminates y , u , v . (197) A(t)e^{yS(t)}(1+aS(t))^u(1-bS(t))^{-v},\qquad\text{indeterminates }y,u,v. \tag*{(197)} A ( t ) e y S ( t ) ( 1 + a S ( t ) ) u ( 1 − b S ( t ) ) − v , indeterminates y , u , v . (197)
Proof. For (i), applying Theorem 6.8 to Γ = 1 \Gamma=1 Γ = 1 and the m + 1 m+1 m + 1 factors 1 + c i u 1+c_i u 1 + c i u , we obtain the asserted family. In [49 ] , the ordinary Riordan family uses the same equation with prefactors ∏ i = 0 s − 1 ( 1 + c i U ) \prod_{i=0}^{s-1}(1+c_iU) ∏ i = 0 s − 1 ( 1 + c i U ) , where s s s is an integer, 0 ≤ s ≤ m + 1 0\le s\le m+1 0 ≤ s ≤ m + 1 , and the empty product is one [49 pp. 23–24, Eqs. (5.1)–(5.2) and Theorem 5.1] . Its ordinary row generating series is this prefactor divided by 1 − q U 1-qU 1 − q U . The independent powers and arbitrary initial series A A A in our conclusion follow from the admissible logarithms in Theorem 6.8 .
For (ii), use the factors 1 + u 1+u 1 + u and r r r copies of 1 + x u 1+xu 1 + xu , then identify the intensities of the repeated factors. The source equation and prefactors, with powers i ∈ { 0 , … , r } i\in\{0,\ldots,r\} i ∈ { 0 , … , r } , j ∈ { 0 , 1 } j\in\{0,1\} j ∈ { 0 , 1 } , occur in [48 Section 6.4, Eqs. (6.145)–(6.147) and Proposition 6.22] .
For (iii), take Γ ( u ) = ( 1 − b u ) − 1 \Gamma(u)=(1-bu)^{-1} Γ ( u ) = ( 1 − b u ) − 1 , which is PF, and the factor 1 + a u 1+au 1 + a u . Applying Theorem 6.8 to these factors, we obtain the result. The source prefactor powers are i , j ∈ { 0 , 1 } i,j\in\{0,1\} i , j ∈ { 0 , 1 } [48 Eq. (6.161) and Theorem 6.25] . The choices a = b = 1 a=b=1 a = b = 1 and a = b = 2 a=b=2 a = b = 2 give the series-parallel network and nondegenerate fanout-free function series, respectively [48 Section 6.6] .
Here c i , x , a , b , q c_i,x,a,b,q c i , x , a , b , q remain fixed. The extension concerns independent powers and arbitrary initial moments; it makes no joint coefficientwise assertion in these structural parameters.
Tree moment inputs and universal component families
Fuss–Raney tree functions and colored components
For integer p ≥ 1 p\ge1 p ≥ 1 , we count rooted trees with p p p ordered child slots at each vertex, allowing slots to be empty. For a distinguished root, we may instead prescribe r r r slots. Giving each edge weight b b b , we obtain the series t B p ( b t ) r tB_p(bt)^r t B p ( b t ) r . We use the known Raney moments to extend this model to real p ≥ 1 p\ge1 p ≥ 1 , 0 ≤ r ≤ p 0\le r\le p 0 ≤ r ≤ p , and then apply lifting to obtain a universal colored component family. For the related multivariate Fuss-Narayana positivity, see [31 Theorem 10.1] . Define the unique formal series B p ( z ) ∈ 1 + z R [ [ z ] ] B_p(z)\in1+z\mathbb{R}[[z]] B p ( z ) ∈ 1 + z R [[ z ]] by
B p ( z ) = 1 + z B p ( z ) p . (198) B_p(z)=1+zB_p(z)^p. \tag*{(198)} B p ( z ) = 1 + z B p ( z ) p . (198)
We use the generalized binomial expansion for real powers. Define R p , r ( 0 ) = 1 R_{p,r}(0)=1 R p , r ( 0 ) = 1 and
R p , r ( n ) = r n ! ∏ j = 1 n − 1 ( p n + r − j ) (199) R_{p,r}(n)=\frac{r}{n!}\prod_{j=1}^{n-1}(pn+r-j) \tag*{(199)} R p , r ( n ) = n ! r j = 1 ∏ n − 1 ( p n + r − j ) (199)
for n ≥ 1 n \ge1 n ≥ 1 . The following representation is due to Młotkowski [26 Eqs. (3.2)–(3.3) and Theorem 5.1] `.
Lemma 7.1 (Classical Raney moment measures). For p ≥ 1 p \ge1 p ≥ 1 and 0 ≤ r ≤ p 0 \le r \le p 0 ≤ r ≤ p , there is a compactly supported probability measure μ p , r \mu_{p,r} μ p , r on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that
∫ 0 ∞ x n d μ p , r ( x ) = R p , r ( n ) , B p ( z ) r = ∑ n ≥ 0 R p , r ( n ) z n . (200) \int_{0}^{\infty} x^n \,\mathrm{d}\mu_{p,r}(x) = R_{p,r}(n), \qquad B_p(z)^r = \sum_{n \ge0} R_{p,r}(n)z^n. \tag*{(200)} ∫ 0 ∞ x n d μ p , r ( x ) = R p , r ( n ) , B p ( z ) r = n ≥ 0 ∑ R p , r ( n ) z n . (200) For r = 0 r=0 r = 0 , we take μ p , 0 = δ 0 \mu_{p,0}=\delta_0 μ p , 0 = δ 0 .
To obtain the shifted exponential coefficients required for lifting from these ordinary moment coefficients, we introduce an independent gamma factor of shape two. For α > 0 \alpha>0 α > 0 , we write
Γ ( α ) = ∫ 0 ∞ x α − 1 e − x d x . \Gamma(\alpha)=\int_{0}^{\infty}x^{\alpha-1}e^{-x}\,\mathrm{d}x. Γ ( α ) = ∫ 0 ∞ x α − 1 e − x d x .
For α , c > 0 \alpha,c>0 α , c > 0 , the notation Gamma( α , c ) (\alpha,c) ( α , c ) means the probability distribution on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) with density
x α − 1 e − x / c Γ ( α ) c α ( x > 0 ) . (201) \frac{x^{\alpha-1}e^{-x/c}}{\Gamma(\alpha)c^\alpha}\qquad(x>0). \tag*{(201)} Γ ( α ) c α x α − 1 e − x / c ( x > 0 ) . (201)
Thus the second parameter is the scale, and unit scale means c = 1 c=1 c = 1 .
Fix p ≥ 1 p\ge1 p ≥ 1 , 0 ≤ r ≤ p 0\le r\le p 0 ≤ r ≤ p , and b ≥ 0 b\ge0 b ≥ 0 . Applying Theorem 4.6 to the moments b n = b n R p , r ( n ) b_n=b^nR_{p,r}(n) b n = b n R p , r ( n ) of a scaled Raney variable, we obtain
U p , r , b ( t ) : = t B p ( b t ) r ∈ C . (202) U_{p,r,b}(t):=tB_p(bt)^r\in\mathcal{C}. \tag*{(202)} U p , r , b ( t ) := t B p ( b t ) r ∈ C . (202)
More precisely, if X X X has law μ p , r \mu_{p,r} μ p , r and G G G is an independent gamma variable of shape two and unit scale, then
U p , r , b ( n + 1 ) ( 0 ) = ( n + 1 ) ! b n R p , r ( n ) = E ( b G X ) n ( n ≥ 0 ) . (203) U_{p,r,b}^{(n+1)}(0)=(n+1)!b^nR_{p,r}(n)=\mathbb{E}(bGX)^n \qquad(n\ge0). \tag*{(203)} U p , r , b ( n + 1 ) ( 0 ) = ( n + 1 )! b n R p , r ( n ) = E ( b GX ) n ( n ≥ 0 ) . (203)
Corollary 7.2 (Fuss–Raney tree lifting). For fixed p ℓ ≥ 1 p_\ell\ge1 p ℓ ≥ 1 , 0 ≤ r ℓ ≤ p ℓ 0\le r_\ell\le p_\ell 0 ≤ r ℓ ≤ p ℓ , b ℓ , q ℓ ≥ 0 b_\ell,q_\ell\ge0 b ℓ , q ℓ ≥ 0 , and any initial Stieltjes moment series A A A , define
∑ n ≥ 0 F n ( y ) t n n ! = A ( t ) ∏ ℓ = 1 m ( 1 − q ℓ U p ℓ , r ℓ , b ℓ ( t ) ) − y ℓ . (204) \sum_{n\ge0}F_n(\mathbf{y})\frac{t^n}{n!}
=A(t)\prod_{\ell=1}^{m}\left(1-q_\ell U_{p_\ell,r_\ell,b_\ell}(t)\right)^{-y_\ell}. \tag*{(204)} n ≥ 0 ∑ F n ( y ) n ! t n = A ( t ) ℓ = 1 ∏ m ( 1 − q ℓ U p ℓ , r ℓ , b ℓ ( t ) ) − y ℓ . (204) Then ( F n ) (F_n) ( F n ) is jointly coefficientwise Hankel-totally positive in y \mathbf{y} y .
Proof. We have U p , r , b = t ∑ n ≥ 0 b n t n = O b U_{p,r,b}=t\sum_{n\ge0}b_nt^n=O_b U p , r , b = t ∑ n ≥ 0 b n t n = O b for the moments verified above. Applying Theorem 4.6 to these inputs, with zero exponential indeterminates, we obtain positivity for (7.7). This includes r = 0 r=0 r = 0 or b = 0 b=0 b = 0 , when U p , r , b = t U_{p,r,b}=t U p , r , b = t .
For integer p ≥ 1 p \ge1 p ≥ 1 and r ∈ { 0 , … , p } r \in\{0,\ldots,p\} r ∈ { 0 , … , p } , we interpret B p ( b t ) r B_p(bt)^r B p ( b t ) r as a rooted tree with r r r distinguished child slots at its root and p p p slots at every other vertex; we give each edge weight b b b . After labeling the vertices, we regard this series as an exponential generating function. Using the labeled cycle construction [41 Section 5.1] , we obtain (204) for directed cycles decorated by these rooted trees, with weight q ℓ q_\ell q ℓ at each cyclic vertex and y ℓ y_\ell y ℓ at each component of color ℓ \ell ℓ .
For nonintegral parameters, the construction uses the formal series and moment representation (203) .
To compute the connected coefficients explicitly, we expand the logarithm. Applying (168) to B p ( z ) − 1 = z ( 1 + B p ( z ) − 1 ) p B_p(z)-1=z(1+B_p(z)-1)^p B p ( z ) − 1 = z ( 1 + B p ( z ) − 1 ) p , we obtain, for every s ≥ 0 s \ge0 s ≥ 0 and j ≥ 1 j \ge1 j ≥ 1 ,
[ z j ] B p ( z ) s = s j [ u j − 1 ] ( 1 + u ) p j + s − 1 = R p , s ( j ) . (205) [z^j]B_p(z)^s=\frac{s}{j}[u^{j-1}](1+u)^{pj+s-1}=R_{p,s}(j). \tag*{(205)} [ z j ] B p ( z ) s = j s [ u j − 1 ] ( 1 + u ) p j + s − 1 = R p , s ( j ) . (205)
For j = 0 j=0 j = 0 , both coefficients are one. Using (205) with s = r k s=rk s = r k , we obtain
n ! [ t n ] [ − log ( 1 − q U p , r , b ( t ) ) ] = n ! ∑ k = 1 n q k b n − k k R p , r k ( n − k ) ( n ≥ 1 ) . (206) n![t^n]\left[-\log(1-qU_{p,r,b}(t))\right]=n!\sum_{k=1}^{n}\frac{q^k b^{n-k}}{k}R_{p,rk}(n-k)\qquad(n\ge1). \tag*{(206)} n ! [ t n ] [ − log ( 1 − q U p , r , b ( t )) ] = n ! k = 1 ∑ n k q k b n − k R p , r k ( n − k ) ( n ≥ 1 ) . (206)
In (206) , we evaluate (199) for the coefficient parameter r k rk r k ; we use it only in the coefficient formula, without a moment interpretation for the auxiliary parameter.
We recover two familiar implicit tree equations by selecting subfamilies of (202) .
Corollary 7.3 (Two implicit tree families). For b ≥ 0 b \ge0 b ≥ 0 , each of the following unique formal solutions belongs to C C C :
U ( t ) = t ( 1 + b U ( t ) ) p ( p ≥ 1 ) , (207) U(t)=t(1+bU(t))^p\qquad(p\ge1), \tag*{(207)} U ( t ) = t ( 1 + b U ( t ) ) p ( p ≥ 1 ) , (207) V ( t ) = t ( 1 − b V ( t ) ) − a ( a > 0 ) . (208) V(t)=t(1-bV(t))^{-a}\qquad(a>0). \tag*{(208)} V ( t ) = t ( 1 − bV ( t ) ) − a ( a > 0 ) . (208) Their shifted coefficients are, respectively,
U ( n + 1 ) ( 0 ) = n ! ( p ( n + 1 ) n ) b n , (209) U^{(n+1)}(0)=n!\binom{p(n+1)}{n}b^n, \tag*{(209)} U ( n + 1 ) ( 0 ) = n ! ( n p ( n + 1 ) ) b n , (209) V ( n + 1 ) ( 0 ) = b n ( a ( n + 1 ) n ) ( n ≥ 0 ) . (210) V^{(n+1)}(0)=b^n\binom{a(n+1)}{n}\qquad(n\ge0). \tag*{(210)} V ( n + 1 ) ( 0 ) = b n ( n a ( n + 1 ) ) ( n ≥ 0 ) . (210)
Proof. For (207) , we take U = t B p ( b t ) p U=tB_p(bt)^p U = t B p ( b t ) p . From (198) , we have B p ( b t ) = 1 + b U B_p(bt)=1+bU B p ( b t ) = 1 + b U , proving the equation. For (208) , take V = t B a + 1 ( b t ) a V=tB_{a+1}(bt)^a V = t B a + 1 ( b t ) a . Dividing B a + 1 ( b t ) = 1 + b t B a + 1 ( b t ) a + 1 B_{a+1}(bt)=1+btB_{a+1}(bt)^{a+1} B a + 1 ( b t ) = 1 + b t B a + 1 ( b t ) a + 1 by B a + 1 ( b t ) B_{a+1}(bt) B a + 1 ( b t ) , we obtain 1 − b V = B a + 1 ( b t ) − 1 1-bV=B_{a+1}(bt)^{-1} 1 − bV = B a + 1 ( b t ) − 1 , and hence the stated equation for V V V . By Corollary 7.2 , both series are admissible. Using (169) , we obtain (209) and (210) . For a = 1 a=1 a = 1 , we obtain the usual plane-tree equation V = t / ( 1 − b V ) V=t/(1-bV) V = t / ( 1 − bV ) .
Finally, we recover the original Cayley tree function as a limit of finite-slot tree functions. Put U p = U p , p , 1 / p U_p=U_{p,p,1/p} U p = U p , p , 1/ p . By (207) and (209) , we have
U p = t ( 1 + U p / p ) p , U p ( n ) ( 0 ) = ∏ j = 0 n − 2 ( n − j / p ) ⟶ n n − 1 ( n ≥ 1 ) . (211) U_p=t(1+U_p/p)^p,\qquad U_p^{(n)}(0)=\prod_{j=0}^{n-2}(n-j/p)\longrightarrow n^{n-1}\quad(n\geq1). \tag*{(211)} U p = t ( 1 + U p / p ) p , U p ( n ) ( 0 ) = j = 0 ∏ n − 2 ( n − j / p ) ⟶ n n − 1 ( n ≥ 1 ) . (211)
We interpret an empty product as one. We therefore obtain coefficientwise convergence U p → T U_p\to T U p → T , where T = t e T T=te^T T = t e T . We may also take nonnegative finite sums and finite compositions of the series in (202) , by adding their representing measures and applying Theorem 4.4 . Using these series in Corollary 5.1, or as substitution series in Theorem 4.5 , we obtain further colored tree families with arbitrary initial Stieltjes moments.
Integrated tree moment kernels
We count sets of rooted-tree components and connected functional-digraph components on a labeled vertex set, marking the two component types by fixed weights a , b ≥ 0 a,b\geq0 a , b ≥ 0 . The exponential coefficients of e a T + b L e^{aT+b\mathcal{L}} e a T + b L are the resulting weighted counting sequence. Applying Theorem 1.7 to T , L T,\mathcal{L} T , L , we obtain its numerical Stieltjes property. We then integrate this moment series and obtain the general component family below. We first record the scale closure used in that construction.
Lemma 7.4 (Positive combinations and argument scales). If K 1 , … , K s ∈ C K_1,\ldots,K_s\in\mathcal{C} K 1 , … , K s ∈ C and λ j , c j ≥ 0 \lambda_j,c_j\geq0 λ j , c j ≥ 0 are fixed real numbers, then
∑ j = 1 s λ j K j ( c j t ) ∈ C . (212) \sum_{j=1}^{s}\lambda_jK_j(c_jt)\in\mathcal{C}. \tag*{(212)} j = 1 ∑ s λ j K j ( c j t ) ∈ C . (212) Proof. Let ν j \nu_j ν j represent the once-shifted coefficients of K j K_j K j . Differentiating the scaled series, we obtain
d n + 1 d t n + 1 ∑ j = 1 s λ j K j ( c j t ) ∣ t = 0 = ∑ j = 1 s λ j c j ∫ 0 ∞ ( c j x ) n ν j ( d x ) . (213) \left.\frac{\mathrm{d}^{n+1}}{\mathrm{d}t^{n+1}}\sum_{j=1}^{s}\lambda_jK_j(c_jt)\right|_{t=0}
=\sum_{j=1}^{s}\lambda_jc_j\int_{0}^{\infty}(c_jx)^n\,\nu_j(\mathrm{d}x). \tag*{(213)} d t n + 1 d n + 1 j = 1 ∑ s λ j K j ( c j t ) t = 0 = j = 1 ∑ s λ j c j ∫ 0 ∞ ( c j x ) n ν j ( d x ) . (213) For each j j j , take the image of λ j c j ν j \lambda_jc_j\nu_j λ j c j ν j under x ↦ c j x x\mapsto c_jx x ↦ c j x , and add these positive measures. We thereby obtain the moments in (213) . When c j = 0 c_j=0 c j = 0 , we take the zero measure, consistently with K j ( 0 ) = 0 K_j(0)=0 K j ( 0 ) = 0 . This proves (212) .
Define
L ( t ) : = − log ( 1 − T ( t ) ) . (214) \mathcal{L}(t):=-\log(1-T(t)). \tag*{(214)} L ( t ) := − log ( 1 − T ( t )) . (214)
By Proposition 5.2 and Corollary 5.1, we have T , L ∈ C T,\mathcal{L}\in\mathcal{C} T , L ∈ C . We use these two exponents to construct the following family of integrated kernels.
Theorem 7.5 (Integrated tree kernels and component lifting). For fixed a , b ≥ 0 a,b \ge0 a , b ≥ 0 , define the formal primitive
K a , b ( t ) : = ∫ 0 t e a T ( s ) ( 1 − T ( s ) ) b d s = ∫ 0 T ( t ) e ( a − 1 ) u ( 1 − u ) 1 − b d u . (215) \mathcal{K}_{a,b}(t) := \int_0^t \frac{e^{aT(s)}}{(1-T(s))^b}\,\mathrm{d}s = \int_0^{T(t)} e^{(a-1)u}(1-u)^{1-b}\,\mathrm{d}u. \tag*{(215)} K a , b ( t ) := ∫ 0 t ( 1 − T ( s ) ) b e a T ( s ) d s = ∫ 0 T ( t ) e ( a − 1 ) u ( 1 − u ) 1 − b d u . (215) Then K a , b ∈ C \mathcal{K}_{a,b} \in\mathcal{C} K a , b ∈ C . More generally, fix a ℓ , b ℓ , c ℓ ≥ 0 a_\ell, b_\ell, c_\ell\ge0 a ℓ , b ℓ , c ℓ ≥ 0 for 1 ≤ ℓ ≤ m 1 \le\ell\le m 1 ≤ ℓ ≤ m , and let A ( t ) A(t) A ( t ) be any initial Stieltjes moment series. Define
∑ n ≥ 0 F n ( y ) t n n ! = A ( t ) exp ( ∑ ℓ = 1 m y ℓ K a ℓ , b ℓ ( c ℓ t ) ) . (216) \sum_{n\ge0} F_n(\mathbf{y})\frac{t^n}{n!} = A(t)\exp\left(\sum_{\ell=1}^{m} y_\ell\mathcal{K}_{a_\ell,b_\ell}(c_\ell t)\right). \tag*{(216)} n ≥ 0 ∑ F n ( y ) n ! t n = A ( t ) exp ( ℓ = 1 ∑ m y ℓ K a ℓ , b ℓ ( c ℓ t ) ) . (216) Then ( F n ) (F_n) ( F n ) is jointly coefficientwise Hankel-totally positive in y \mathbf{y} y .
Proof. Define m n ( a , b ) m_n(a,b) m n ( a , b ) by
K a , b ′ ( t ) = e a T ( t ) + b L ( t ) = ∑ n ≥ 0 m n ( a , b ) t n n ! . (217) \mathcal{K}'_{a,b}(t) = e^{aT(t)+b\mathcal{L}(t)} = \sum_{n\ge0}m_n(a,b)\frac{t^n}{n!}. \tag*{(217)} K a , b ′ ( t ) = e a T ( t ) + b L ( t ) = n ≥ 0 ∑ m n ( a , b ) n ! t n . (217)
Since T , L ∈ C T,\mathcal{L} \in\mathcal{C} T , L ∈ C , applying Theorem 1.7 (iii) to these exponents at fixed a , b ≥ 0 a,b \ge0 a , b ≥ 0 , we obtain the Stieltjes sequence ( m n ( a , b ) ) (m_n(a,b)) ( m n ( a , b )) . Applying Theorem 4.6 to this sequence, we obtain K a , b = I m ∈ C \mathcal{K}_{a,b}=I_m \in\mathcal{C} K a , b = I m ∈ C .
Making the formal change of variables s = u e − u s=ue^{-u} s = u e − u , with d s = ( 1 − u ) e − u d u \mathrm{d}s=(1-u)e^{-u}\,\mathrm{d}u d s = ( 1 − u ) e − u d u , we obtain the second expression in (215) ; both primitives have zero constant term. Applying Lemma 7.4 to the argument scales and Theorem 1.7 to the resulting exponents, we obtain the scaled component family.
Several familiar connected series occur in this family:
K 1 , 0 = T − 1 2 T 2 , K 1 , 1 = T , K 1 , 2 = L , K 1 , 3 = T 1 − T . (218) \mathcal{K}_{1,0}=T-\frac{1}{2}T^2,\qquad\mathcal{K}_{1,1}=T,\qquad\mathcal{K}_{1,2}=\mathcal{L},\qquad\mathcal{K}_{1,3}=\frac{T}{1-T}. \tag*{(218)} K 1 , 0 = T − 2 1 T 2 , K 1 , 1 = T , K 1 , 2 = L , K 1 , 3 = 1 − T T . (218)
These are the series for unrooted labeled trees, rooted trees, connected functional digraphs, and nonempty functional digraphs, respectively; the last equals ( 1 − T ) − 1 − 1 (1-T)^{-1}-1 ( 1 − T ) − 1 − 1 .
Exponential tree composites and unrooted forests
For rooted forests, Sokal introduced a weight 1 + w 1+w 1 + w for each root descent, namely an edge from a root to a child with a smaller label [39 Section 6.7] . For that model, we express the connected series as an exponential composite of T T T . Ding and Zeng obtained the same coefficient identity in their study of Ramanujan’s series [8 Eq. (1.3)] . For unrooted trees, we use the classical relation recorded by Josuat-Vergès [17 Section 2] . We collect these known identities in the following lemma.
Lemma 7.6 (Classical rooted and unrooted tree identities). Define B w ( t ) = ( e w T ( t ) − 1 ) / w B_w(t)=(e^{wT(t)}-1)/w B w ( t ) = ( e w T ( t ) − 1 ) / w for w ≠ 0 w\ne0 w = 0 , and B 0 ( t ) = T ( t ) B_0(t)=T(t) B 0 ( t ) = T ( t ) . Then
B w ( t ) = ∑ n ≥ 1 ( n + w ) n n − 1 t n n ! , B w ′ ( t ) = e ( w + 1 ) T ( t ) 1 − T ( t ) . (219) B_w(t)=\sum_{n\ge1}(n+w)n^{n-1}\frac{t^n}{n!},\qquad B'_w(t)=\frac{e^{(w+1)T(t)}}{1-T(t)}. \tag*{(219)} B w ( t ) = n ≥ 1 ∑ ( n + w ) n n − 1 n ! t n , B w ′ ( t ) = 1 − T ( t ) e ( w + 1 ) T ( t ) . (219) Also, the unrooted-tree series satisfies
U ( t ) : = T ( t ) − 1 2 T ( t ) 2 = ∑ n ≥ 1 n n − 2 t n n ! , U ′ ( t ) = e T ( t ) . (220) U(t):=T(t)-\frac{1}{2}T(t)^2=\sum_{n\ge1}n^{n-2}\frac{t^n}{n!},\qquad U'(t)=e^{T(t)}. \tag*{(220)} U ( t ) := T ( t ) − 2 1 T ( t ) 2 = n ≥ 1 ∑ n n − 2 n ! t n , U ′ ( t ) = e T ( t ) . (220) The coefficient identities use 0 0 = 1 0^0=1 0 0 = 1 when needed.
We now extend these two forest models by allowing arbitrary initial moments and independent component colors.
Corollary 7.7 (Universal forest component families). For every fixed w ≥ − 1 w\ge-1 w ≥ − 1 , we have B w ∈ C B_w\in\mathcal{C} B w ∈ C , and U ∈ C U\in\mathcal{C} U ∈ C . Consequently, for fixed w ℓ ≥ − 1 w_\ell\ge-1 w ℓ ≥ − 1 , c ℓ , d ℓ ≥ 0 c_\ell,d_\ell\ge0 c ℓ , d ℓ ≥ 0 , and any real initial Stieltjes moment series A ( t ) A(t) A ( t ) , the exponential coefficients of
A ( t ) exp ( ∑ ℓ = 1 m { y ℓ B w ℓ ( c ℓ t ) + z ℓ U ( d ℓ t ) } ) (221) A(t)\exp\left(\sum_{\ell=1}^{m}\{y_\ell B_{w_\ell}(c_\ell t)+z_\ell U(d_\ell t)\}\right) \tag*{(221)} A ( t ) exp ( ℓ = 1 ∑ m { y ℓ B w ℓ ( c ℓ t ) + z ℓ U ( d ℓ t )} ) (221) are coefficientwise Hankel-totally positive jointly in y , z \mathbf{y},\mathbf{z} y , z . In particular, for fixed w ≥ − 1 w\ge-1 w ≥ − 1 , the polynomials defined by
e x B w ( t ) = ∑ n ≥ 0 F n ♯ ( x , w ) t n n ! (222) e^{xB_w(t)}=\sum_{n\ge0}F_n^\sharp(x,w)\frac{t^n}{n!} \tag*{(222)} e x B w ( t ) = n ≥ 0 ∑ F n ♯ ( x , w ) n ! t n (222) are coefficientwise Hankel-totally positive in x x x .
Proof. Using the derivative identities in Lemma 7.6 and the zero constant terms, we obtain
B w = K w + 1 , 1 ( w ≥ − 1 ) , U = K 1 , 0 . (223) B_w=\mathcal{K}_{w+1,1}\quad(w\ge-1),\qquad U=\mathcal{K}_{1,0}. \tag*{(223)} B w = K w + 1 , 1 ( w ≥ − 1 ) , U = K 1 , 0 . (223)
This also holds at w = 0 w=0 w = 0 . By Theorem 7.5 , we derive admissibility. Applying Lemma 7.4 to the scales and Theorem 1.7 to the resulting exponents, we obtain (7.24). Taking A = 1 A=1 A = 1 , one color, and unit scale, we have (7.25).
Sokal’s joint assertions in ( x , w ) (x,w) ( x , w ) and in ( x , v ) (x,v) ( x , v ) , with w = v − 1 w=v-1 w = v − 1 , remain open [39 Conjectures 6.11(b) and 6.12(b)] . The corollary fixes w w w and does not imply either assertion.
For an explicit check, put v = w + 1 v=w+1 v = w + 1 . From B w ( n + 1 ) ( 0 ) = ( n + w + 1 ) n B_w^{(n+1)}(0)=(n+w+1)^n B w ( n + 1 ) ( 0 ) = ( n + w + 1 ) n , we obtain
det [ F i + j ♯ ( x , v − 1 ) ] i , j = 0 2 = 2 x 3 ( v + 1 ) 3 + x 2 ( 2 v 3 + 12 v 2 + 22 v + 11 ) . (224) \det\left[F_{i+j}^{\sharp}(x,v-1)\right]^2_{i,j=0}=2x^3(v+1)^3+x^2(2v^3+12v^2+22v+11). \tag*{(224)} det [ F i + j ♯ ( x , v − 1 ) ] i , j = 0 2 = 2 x 3 ( v + 1 ) 3 + x 2 ( 2 v 3 + 12 v 2 + 22 v + 11 ) . (224)
For unrooted forests, using the usual labeled set construction, we obtain
e y U ( t ) = ∑ n ≥ 0 F n ( y ) t n n ! . (225) e^{yU(t)}=\sum_{n\geq0}\mathcal{F}_n(y)\frac{t^n}{n!}. \tag*{(225)} e y U ( t ) = n ≥ 0 ∑ F n ( y ) n ! t n . (225)
Here we mark the number of tree components in F n ( y ) \mathcal{F}_n(y) F n ( y ) by the exponent of the component variable. For example,
F 1 = y , F 2 = y 2 + y , F 3 = y 3 + 3 y 2 + 3 y , F 4 = y 4 + 6 y 3 + 15 y 2 + 16 y . (226) \mathcal{F}_1=y,\qquad\mathcal{F}_2=y^2+y,\qquad\mathcal{F}_3=y^3+3y^2+3y,\qquad\mathcal{F}_4=y^4+6y^3+15y^2+16y. \tag*{(226)} F 1 = y , F 2 = y 2 + y , F 3 = y 3 + 3 y 2 + 3 y , F 4 = y 4 + 6 y 3 + 15 y 2 + 16 y . (226)
Applying Corollary 7.7 to this family, we obtain coefficientwise Hankel-total positivity.
Shifted tree functions and Ramanujan–Shor and Greg trees
Sokal [39, Theorem 1.3] proved coefficientwise Hankel total positivity for the underlying rooted-forest polynomials. We begin with their counted objects and equation, verify the shifted moment condition from a tilted tree measure, and then extend the fixed-weight model to arbitrary initial moments and independent component intensities.
Ramanujan–Shor polynomials refine rooted labeled trees by proper and improper edges. For an edge directed from a parent to a child, we call the edge improper if the parent label exceeds the smallest label in the child’s subtree, and proper otherwise. Zeng developed this refinement in connection with Ramanujan’s series [43 ] . From Sokal’s generating function with two edge weights [39 Proposition 3.1(d) and Corollary 3.2] , we obtain the differential equation below. To connect this equation with our moment criterion, we first study translations of the tree function.
For this purpose, we use the principal real continuation of T T T to ( − ∞ , 1 / e ) (-\infty,1/e) ( − ∞ , 1/ e ) : it is the inverse of u ↦ u e − u u\mapsto ue^{-u} u ↦ u e − u on u < 1 u<1 u < 1 . In the next lemma, we evaluate this analytic continuation at nonzero arguments.
Lemma 7.8 (Tilted tree moment measures). Let ν \nu ν be a probability measure representing ( ( n + 1 ) n ) n ≥ 0 ((n+1)^n)_{n\geq0} (( n + 1 ) n ) n ≥ 0 . For every fixed real c < 1 / e c<1/e c < 1/ e , the Taylor series
T ~ c ( s ) : = T ( c + s ) − T ( c ) (227) \widetilde{T}_c(s):=T(c+s)-T(c) \tag*{(227)} T c ( s ) := T ( c + s ) − T ( c ) (227) belongs to C \mathcal{C} C , with shifted moment measure e c x ν ( d x ) e^{cx}\nu(\mathrm{d}x) e c x ν ( d x ) .
Proof. To expand about a nonzero center, we must justify differentiating the moment representation and show that the resulting integrals are finite. We prove these assertions in the following claim.
Claim 9. For c < 1 / e c<1/e c < 1/ e and n ≥ 0 n\geq0 n ≥ 0 , we have
T ( n + 1 ) ( c ) = ∫ 0 ∞ x n e c x ν ( d x ) < ∞ . (228) T^{(n+1)}(c)=\int_{0}^{\infty}x^n e^{cx}\,\nu(\mathrm{d}x)<\infty. \tag*{(228)} T ( n + 1 ) ( c ) = ∫ 0 ∞ x n e c x ν ( d x ) < ∞. (228)
Proof. For 0 ≤ c < 1 / e 0 \le c < 1/e 0 ≤ c < 1/ e , applying Tonelli’s theorem to the nonnegative exponential expansion, we obtain
∫ 0 ∞ e c x ν ( d x ) = ∑ k ≥ 0 c k k ! ( k + 1 ) k = T ′ ( c ) < ∞ . (229) \int_0^\infty e^{cx}\,\nu(\mathrm{d}x)=\sum_{k\geq0}\frac{c^k}{k!}(k+1)^k=T'(c)<\infty. \tag*{(229)} ∫ 0 ∞ e c x ν ( d x ) = k ≥ 0 ∑ k ! c k ( k + 1 ) k = T ′ ( c ) < ∞. (229)
From the defining coefficients, we know that the series on the right has radius 1 / e 1/e 1/ e and is the derivative of T T T . For any fixed c < 1 / e c<1/e c < 1/ e , choose r ∈ ( max { c , 0 } , 1 / e ) r\in(\max\{c,0\},1/e) r ∈ ( max { c , 0 } , 1/ e ) . For every n ≥ 0 n\geq0 n ≥ 0 , there is a finite constant C n , c , r C_{n,c,r} C n , c , r such that
x n e c x ≤ C n , c , r e r x ( x ≥ 0 ) . (230) x^n e^{cx}\leq C_{n,c,r}e^{rx}\qquad(x\geq0). \tag*{(230)} x n e c x ≤ C n , c , r e r x ( x ≥ 0 ) . (230)
By (229) , we have an integrable bound. Using this estimate on a neighborhood of c c c , we may differentiate under the integral any finite number of times. More explicitly, for ∣ h ∣ < r − c |h|<r-c ∣ h ∣ < r − c , the exponential expansion around c c c is dominated in absolute value by e r x e^{rx} e r x . We therefore obtain a real analytic function throughout ( − ∞ , 1 / e ) (-\infty,1/e) ( − ∞ , 1/ e ) .
On the other hand, d d u ( u e − u ) = ( 1 − u ) e − u > 0 \frac{\mathrm{d}}{\mathrm{d}u}(ue^{-u})=(1-u)e^{-u}>0 d u d ( u e − u ) = ( 1 − u ) e − u > 0 for u < 1 u<1 u < 1 . We therefore obtain a real analytic inverse T T T , with real analytic derivative, on the interval of centers under consideration. By the identity theorem for real analytic functions, we extend (229) to every c < 1 / e c<1/e c < 1/ e . Differentiating this identity under the integral, justified by (230) , we obtain (228) . □
Since T ^ c ( n + 1 ) ( 0 ) = T ( n + 1 ) ( c ) \widehat{T}_c^{(n+1)}(0)=T^{(n+1)}(c) T c ( n + 1 ) ( 0 ) = T ( n + 1 ) ( c ) , we obtain a finite positive measure representing all the shifted coefficients by (228) . Hence T ^ c ∈ C \widehat{T}_c\in\mathcal{C} T c ∈ C , as asserted. □
We now state the known rooted-tree equation and its centered tree-function form. In [39 Eq. (3.18)] , the two parameters are denoted by y , z y,z y , z . We use η , ζ \eta,\zeta η , ζ to distinguish them from the component intensities.
Lemma 7.9 (Sokal’s rooted-tree equation). For η , ζ ≥ 0 \eta,\zeta\geq0 η , ζ ≥ 0 , the rooted-tree series R η , ζ ( t ) R_{\eta,\zeta}(t) R η , ζ ( t ) satisfies
R η , ζ ′ ( t ) = e ζ R η , ζ ( t ) 1 − η R η , ζ ( t ) , R η , ζ ( 0 ) = 0. (231) R'_{\eta,\zeta}(t)=\frac{e^{\zeta R_{\eta,\zeta}(t)}}{1-\eta R_{\eta,\zeta}(t)},\qquad R_{\eta,\zeta}(0)=0. \tag*{(231)} R η , ζ ′ ( t ) = 1 − η R η , ζ ( t ) e ζ R η , ζ ( t ) , R η , ζ ( 0 ) = 0. (231) For η , ζ > 0 \eta,\zeta>0 η , ζ > 0 , let
a = 1 − ζ / η < 1 , c = a e − a < 1 / e , b = ( ζ 2 / η ) e − a > 0. (232) a=1-\zeta/\eta<1,\qquad c=ae^{-a}<1/e,\qquad b=(\zeta^2/\eta)e^{-a}>0. \tag*{(232)} a = 1 − ζ / η < 1 , c = a e − a < 1/ e , b = ( ζ 2 / η ) e − a > 0. (232) Then
R η , ζ ( t ) = T ( c + b t ) − a ζ = 1 ζ T ^ c ( b t ) . (233) R_{\eta,\zeta}(t)=\frac{T(c+bt)-a}{\zeta}=\frac{1}{\zeta}\widehat{T}_c(bt). \tag*{(233)} R η , ζ ( t ) = ζ T ( c + b t ) − a = ζ 1 T c ( b t ) . (233)
We now use these admissible series in the lifting theorem with arbitrary initial moments and independent component colors.
Corollary 7.10 (Rooted-tree component lifting). For every fixed η , ζ ≥ 0 \eta,\zeta\ge0 η , ζ ≥ 0 , we have R η , ζ ∈ C R_{\eta,\zeta} \in\mathcal{C} R η , ζ ∈ C . For independently chosen fixed η ℓ , ζ ℓ , c ℓ ≥ 0 \eta_\ell,\zeta_\ell,c_\ell\ge0 η ℓ , ζ ℓ , c ℓ ≥ 0 and every real initial Stieltjes moment series A ( t ) A(t) A ( t ) , the exponential coefficients of
A ( t ) exp ( ∑ ℓ = 1 m y ℓ R η ℓ , ζ ℓ ( c ℓ t ) ) (234) A(t)\exp\left(\sum_{\ell=1}^{m}y_\ell R_{\eta_\ell,\zeta_\ell}(c_\ell t)\right) \tag*{(234)} A ( t ) exp ( ℓ = 1 ∑ m y ℓ R η ℓ , ζ ℓ ( c ℓ t ) ) (234) are coefficientwise Hankel-totally positive jointly in y \mathbf{y} y .
Proof. For η , ζ > 0 \eta,\zeta>0 η , ζ > 0 , combine (233) with Lemmas 7.8 and 7.4. On the boundary, integration gives
R 0 , ζ ( t ) = − log ( 1 − ζ t ) ζ ( ζ > 0 ) , R η , 0 ( t ) = 1 − 1 − 2 η t η ( η > 0 ) . (235) R_{0,\zeta}(t)=-\frac{\log(1-\zeta t)}{\zeta}\quad(\zeta>0),\qquad R_{\eta,0}(t)=\frac{1-\sqrt{1-2\eta t}}{\eta}\quad(\eta>0). \tag*{(235)} R 0 , ζ ( t ) = − ζ log ( 1 − ζ t ) ( ζ > 0 ) , R η , 0 ( t ) = η 1 − 1 − 2 η t ( η > 0 ) . (235)
The shifted coefficients are n ! ζ n n!\zeta^n n ! ζ n and ( 1 / 2 ) n ( 2 η ) n (1/2)_n(2\eta)^n ( 1/2 ) n ( 2 η ) n , the moments of Gamma( 1 , ζ ) (1,\zeta) ( 1 , ζ ) and Gamma( 1 / 2 , 2 η ) (1/2,2\eta) ( 1/2 , 2 η ) , respectively. At η = ζ = 0 \eta=\zeta=0 η = ζ = 0 , R 0 , 0 = t R_{0,0}=t R 0 , 0 = t has shifted measure δ 0 \delta_0 δ 0 . Hence every stated exponent is admissible. Applying scale closure to these exponents and Theorem 1.7 to the initial series and scaled exponents, we obtain the conclusion.
In a rooted Greg tree, we allow additional unlabeled branching vertices alongside the vertices labeled 1 , … , n 1,\ldots,n 1 , … , n . Every unlabeled vertex has degree at least three, except that an unlabeled root may have degree two. We regard the trees as non-plane and identify them by isomorphisms preserving the root and all labels. Josuat-Vergès [17, Definitions 3.1–3.2 and Theorem 4.2 shows that their generating polynomial, with weight q q q for each unlabeled vertex, is G n ( q ) = R 1 + q , 1 ( n ) ( 0 ) \mathfrak{G}_n(q)=R^{(n)}_{1+q,1}(0) G n ( q ) = R 1 + q , 1 ( n ) ( 0 ) . For the first few polynomials, we have
G 1 = 1 , G 2 = q + 2 , G 3 = 3 q 2 + 10 q + 9 , G 4 = 15 q 3 + 70 q 2 + 113 q + 64. (236) \mathfrak{G}_1=1,\qquad\mathfrak{G}_2=q+2,\qquad\mathfrak{G}_3=3q^2+10q+9,\qquad\mathfrak{G}_4=15q^3+70q^2+113q+64. \tag*{(236)} G 1 = 1 , G 2 = q + 2 , G 3 = 3 q 2 + 10 q + 9 , G 4 = 15 q 3 + 70 q 2 + 113 q + 64. (236)
Thus, from (234) , we also obtain Greg component families for independently fixed q ℓ ≥ 0 q_\ell\ge0 q ℓ ≥ 0 . We retain the component intensities as indeterminates and use the established scalar edge-weight positivity as an input.
Ward moment inputs and phylogenetic component families
We count non-plane rooted trees with n + 1 n+1 n + 1 labeled leaves and unlabeled internal vertices, each having at least two children. Giving each internal vertex weight u u u , we obtain the Ward polynomial W n ( u ) W_n(u) W n ( u ) . Elvey Price and Sokal [9 Section 1, item 4, pp. 11–12, and Appendix A] related these trees to augmented perfect matchings and proved coefficientwise Hankel total positivity in u u u ; see also [48 Section 6.5] . We recall the generating identity from [9 Eqs. (1.17)–(1.19)] .
Lemma 7.11 (Ward counting series and its moment input). We have
H u ( t ) : = ∑ n ≥ 0 W n ( u ) t n + 1 ( n + 1 ) ! = t + u ( e H u ( t ) − 1 − H u ( t ) ) . (237) H_u(t) := \sum_{n\ge0} W_n(u)\frac{t^{n+1}}{(n+1)!}=t+u\left(e^{H_u(t)}-1-H_u(t)\right). \tag*{(237)} H u ( t ) := n ≥ 0 ∑ W n ( u ) ( n + 1 )! t n + 1 = t + u ( e H u ( t ) − 1 − H u ( t ) ) . (237) Thus H u ( n + 1 ) ( 0 ) = W n ( u ) H_u^{(n+1)}(0)=W_n(u) H u ( n + 1 ) ( 0 ) = W n ( u ) . For each fixed u ≥ 0 u\ge0 u ≥ 0 , ( W n ( u ) ) n ≥ 0 (W_n(u))_{n\ge0} ( W n ( u ) ) n ≥ 0 is a Stieltjes moment sequence.
Applying the real moment criterion to the known Ward sequence at a fixed u u u , we choose a measure ν u \nu_u ν u with moments W n ( u ) W_n(u) W n ( u ) . We therefore have the shifted moment representation
H u ( n + 1 ) ( 0 ) = ∫ 0 ∞ x n ν u ( d x ) . H_u^{(n+1)}(0)=\int_0^\infty x^n\nu_u(dx). H u ( n + 1 ) ( 0 ) = ∫ 0 ∞ x n ν u ( d x ) . Applying the primitive moment construction to this input, we obtain H u ∈ C H_u\in\mathcal{C} H u ∈ C . We now retain independent component intensities, logarithmic powers, an admissible substitution, and arbitrary initial moments in one general result.
Theorem 7.12 (Universal lifting of Ward moment inputs). Fix u ℓ , c ℓ , θ ℓ ≥ 0 u_\ell,c_\ell,\theta_\ell\ge0 u ℓ , c ℓ , θ ℓ ≥ 0 , put H ℓ ( t ) = H u ℓ ( c ℓ t ) H_\ell(t)=H_{u_\ell}(c_\ell t) H ℓ ( t ) = H u ℓ ( c ℓ t ) , and choose S ∈ C S\in\mathcal{C} S ∈ C . For any initial Stieltjes moment series A A A , the exponential coefficients of
A ( S ( t ) ) exp ( ∑ ℓ = 1 m y ℓ H ℓ ( t ) ) ∏ ℓ = 1 m ( 1 − θ ℓ H ℓ ( t ) ) − z ℓ (238) A(S(t))\exp\left(\sum_{\ell=1}^{m}y_\ell H_\ell(t)\right)\prod_{\ell=1}^{m}(1-\theta_\ell H_\ell(t))^{-z_\ell} \tag*{(238)} A ( S ( t )) exp ( ℓ = 1 ∑ m y ℓ H ℓ ( t ) ) ℓ = 1 ∏ m ( 1 − θ ℓ H ℓ ( t ) ) − z ℓ (238) are jointly coefficientwise Hankel-totally positive in y , z \mathbf{y},\mathbf{z} y , z .
Proof. By Lemma 7.4, we derive the shifted Stieltjes inputs of H ℓ H_\ell H ℓ . Applying Theorem 4.6 to these inputs, with substitution S S S and logarithmic parameters θ ℓ \theta_\ell θ ℓ , we obtain the assertion.
Taking S = t S=t S = t and z = 0 \mathbf{z}=0 z = 0 , we have the exponential component family
A ( t ) exp ( ∑ ℓ = 1 m y ℓ H u ℓ ( c ℓ t ) ) (239) A(t)\exp\left(\sum_{\ell=1}^{m}y_\ell H_{u_\ell}(c_\ell t)\right) \tag*{(239)} A ( t ) exp ( ℓ = 1 ∑ m y ℓ H u ℓ ( c ℓ t ) ) (239)
At A = 1 A=1 A = 1 , we count forests of independently weighted phylogenetic components; we mark their colors by y \mathbf{y} y . In the one-color case at unit scale, we recover W n ( u ) = [ y ] F n + 1 ( y ; u ) W_n(u)=[y]F_{n+1}(y;u) W n ( u ) = [ y ] F n + 1 ( y ; u ) . Our new coefficientwise conclusion concerns the component intensities; we keep the internal-vertex weights fixed. We therefore distinguish this universal extension from the cited positivity in the structural variable u u u .
We also identify the component series explicitly. For fixed u ≥ 0 u\ge0 u ≥ 0 , put
q = u 1 + u , d = 1 1 + u , c = q e − q < 1 / e . (240) q=\frac{u}{1+u},\qquad d=\frac{1}{1+u},\qquad c=qe^{-q}<1/e. \tag*{(240)} q = 1 + u u , d = 1 + u 1 , c = q e − q < 1/ e . (240)
We claim that
H u ( t ) = d t − q + T ( c e d t ) = d t + T ^ c ( c ( e d t − 1 ) ) . (241) H_u(t)=dt-q+T(ce^{dt})=dt+\widehat{T}_c\left(c(e^{dt}-1)\right). \tag*{(241)} H u ( t ) = d t − q + T ( c e d t ) = d t + T c ( c ( e d t − 1 ) ) . (241)
To verify the formula, set Z ( t ) = T ( c e d t ) Z(t)=T(ce^{dt}) Z ( t ) = T ( c e d t ) and J ( t ) = d t − q + Z ( t ) J(t)=dt-q+Z(t) J ( t ) = d t − q + Z ( t ) , using the Taylor expansion at c c c . Since T ( c ) = q T(c)=q T ( c ) = q , we have J ( 0 ) = 0 J(0)=0 J ( 0 ) = 0 . By T = t e T T=te^T T = t e T , we have
Z ( t ) = c e d t + Z ( t ) = q e J ( t ) , ( 1 + u ) J ( t ) = t − u + u e J ( t ) . (242) Z(t)=ce^{dt+Z(t)}=qe^{J(t)},\qquad(1+u)J(t)=t-u+ue^{J(t)}. \tag*{(242)} Z ( t ) = c e d t + Z ( t ) = q e J ( t ) , ( 1 + u ) J ( t ) = t − u + u e J ( t ) . (242)
Rearranging the second identity, we deduce the defining equation for H u H_u H u . Comparing coefficients successively, we obtain J = H u J=H_u J = H u , because e J − 1 − J e^J-1-J e J − 1 − J starts in degree two. At u = 0 u=0 u = 0 , we obtain H 0 = t H_0=t H 0 = t , with shifted measure δ 0 \delta_0 δ 0 .
For example, at unit internal-vertex weight we have
H 1 ( t ) = t − 1 2 + T ( 1 2 e ( t − 1 ) / 2 ) = t + t 2 2 ! + 4 t 3 3 ! + 26 t 4 4 ! + ⋯ . (243) H_1(t)=\frac{t-1}{2}+T\left(\frac{1}{2}e^{(t-1)/2}\right)=t+\frac{t^2}{2!}+4\frac{t^3}{3!}+26\frac{t^4}{4!}+\cdots. \tag*{(243)} H 1 ( t ) = 2 t − 1 + T ( 2 1 e ( t − 1 ) /2 ) = t + 2 ! t 2 + 4 3 ! t 3 + 26 4 ! t 4 + ⋯ . (243)
Combining the unrooted-tree and Ward inputs, we obtain
e y U ( t ) + z H 1 ( t ) = ∑ n ≥ 0 M n ( y , z ) t n n ! . (244) e^{yU(t)+zH_1(t)}=\sum_{n\geq0}M_n(y,z)\frac{t^n}{n!}. \tag*{(244)} e y U ( t ) + z H 1 ( t ) = n ≥ 0 ∑ M n ( y , z ) n ! t n . (244)
with joint coefficientwise Hankel total positivity in y , z y,z y , z . The first leading determinants are
det [ M i + j ] i , j = 0 1 = y + z , det [ M i + j ] i , j = 0 2 = 2 ( y + z ) 3 + 7 y 2 + 18 y z + 10 z 2 . (245) \det[M_{i+j}]_{i,j=0}^{1}=y+z,\qquad\det[M_{i+j}]_{i,j=0}^{2}=2(y+z)^3+7y^2+18yz+10z^2. \tag*{(245)} det [ M i + j ] i , j = 0 1 = y + z , det [ M i + j ] i , j = 0 2 = 2 ( y + z ) 3 + 7 y 2 + 18 y z + 10 z 2 . (245)
We may include arbitrary initial moments in this mixed family by applying Theorem 1.7 to U U U and H 1 H_1 H 1 .
Gamma damping of tree coefficients and an excluded composite
The family T p ( t ) = ∑ n ≥ 1 n n − p t n / n ! T_p(t)=\sum_{n\geq1}n^{n-p}t^n/n! T p ( t ) = ∑ n ≥ 1 n n − p t n / n ! includes rooted and unrooted tree series at p = 1 , 2 p=1,2 p = 1 , 2 [17 Section 2] . Using a gamma damping factor, we verify the shifted moment condition for every real p ≥ 1 p\geq1 p ≥ 1 and obtain a general component family by lifting.
Proposition 7.13 (Gamma-damped tree exponents). For every real p ≥ 1 p\geq1 p ≥ 1 , the formal power series
T p ( t ) : = ∑ n ≥ 1 n n − p t n n ! (246) T_p(t):=\sum_{n\geq1}n^{n-p}\frac{t^n}{n!} \tag*{(246)} T p ( t ) := n ≥ 1 ∑ n n − p n ! t n (246) belongs to C C C . For fixed p ℓ ≥ 1 p_\ell\geq1 p ℓ ≥ 1 and c ℓ ≥ 0 c_\ell\geq0 c ℓ ≥ 0 , the exponential coefficients of
A ( t ) exp ( ∑ ℓ = 1 m y ℓ T p ℓ ( c ℓ t ) ) (247) A(t)\exp\left(\sum_{\ell=1}^{m}y_\ell T_{p_\ell}(c_\ell t)\right) \tag*{(247)} A ( t ) exp ( ℓ = 1 ∑ m y ℓ T p ℓ ( c ℓ t ) ) (247) are coefficientwise Hankel-totally positive jointly in y y y for every real initial Stieltjes moment series A ( t ) A(t) A ( t ) .
Proof. Let X X X have moments ( n + 1 ) n (n+1)^n ( n + 1 ) n . For p > 1 p>1 p > 1 , take an independent G ∼ Gamma ( p − 1 , 1 ) G \sim\operatorname{Gamma}(p-1,1) G ∼ Gamma ( p − 1 , 1 ) and put Q = e − G Q=e^{-G} Q = e − G . Applying Euler’s integral to the gamma density, we obtain
E Q n = 1 Γ ( p − 1 ) ∫ 0 ∞ s p − 2 e − ( n + 1 ) s d s = ( n + 1 ) − ( p − 1 ) . (248) \mathbb{E}Q^n=\frac{1}{\Gamma(p-1)}\int_0^\infty s^{p-2}e^{-(n+1)s}\,\mathrm{d}s=(n+1)^{-(p-1)}. \tag*{(248)} E Q n = Γ ( p − 1 ) 1 ∫ 0 ∞ s p − 2 e − ( n + 1 ) s d s = ( n + 1 ) − ( p − 1 ) . (248)
Using independence, we then have
E ( X Q ) n = ( n + 1 ) n ( n + 1 ) − ( p − 1 ) = ( n + 1 ) n + 1 − p = T p ( n + 1 ) ( 0 ) . (249) \mathbb{E}(XQ)^n=(n+1)^n(n+1)^{-(p-1)}=(n+1)^{n+1-p}=T_p^{(n+1)}(0). \tag*{(249)} E ( X Q ) n = ( n + 1 ) n ( n + 1 ) − ( p − 1 ) = ( n + 1 ) n + 1 − p = T p ( n + 1 ) ( 0 ) . (249)
All moments are finite since 0 < Q ≤ 1 0<Q\le1 0 < Q ≤ 1 . For p = 1 p=1 p = 1 , take Q = 1 Q=1 Q = 1 . Thus T p ∈ C T_p\in\mathcal{C} T p ∈ C . Applying scale closure and then Theorem 1.7 to these exponents yields the component assertion. □ \square □
Nonnegative coefficients alone do not imply admissibility. For K ( t ) = T ( t ) 2 K(t)=T(t)^2 K ( t ) = T ( t ) 2 , the first two cumulants are κ 1 = 0 \kappa_1=0 κ 1 = 0 and κ 2 = 2 \kappa_2=2 κ 2 = 2 . A positive measure with zeroth moment zero is the zero measure, so its first moment cannot equal two. Hence T 2 ∉ C T^2\notin\mathcal{C} T 2 ∈ / C . Equivalently, defining P n ( y ) P_n(y) P n ( y ) by
e y T ( t ) 2 = ∑ n ≥ 0 P n ( y ) t n n ! , (250) e^{yT(t)^2}=\sum_{n\ge0}P_n(y)\frac{t^n}{n!}, \tag*{(250)} e y T ( t ) 2 = n ≥ 0 ∑ P n ( y ) n ! t n , (250)
we have
P 1 ( y ) P 3 ( y ) − P 2 ( y ) 2 = − 4 y 2 . (251) P_1(y)P_3(y)-P_2(y)^2=-4y^2. \tag*{(251)} P 1 ( y ) P 3 ( y ) − P 2 ( y ) 2 = − 4 y 2 . (251)
The shifted moment condition is therefore essential.
Polynomial moment inputs and further lifted families
We will obtain general coefficientwise Hankel-positive families from the shifted moment conditions of concrete counting sequences. For the Schläfli-Gessel-Seo polynomials polynomials, we first prove that condition and obtain fixed-shape positivity. For Bessel, Lah, and Laguerre polynomials, we recover known Hankel positivity by a different proof and allow arbitrary initial moments, independent colors, and independently chosen fixed shape parameters. We specify these extensions below and attribute the earlier polynomial results at each specialization. We also obtain lognormal jump families from their classical moments; we use formal series even when exponential moments are infinite.
Schläfli-Gessel-Seo polynomials
For n ≥ 1 n\ge1 n ≥ 1 , let
P 0 ( x ; a , b ) = 1 , P n ( x ; a , b ) = x ∏ h = 1 n − 1 ( x + h a + ( n − h ) b ) . (252) P_0(x;a,b)=1,\qquad P_n(x;a,b)=x\prod_{h=1}^{n-1}\bigl(x+ha+(n-h)b\bigr). \tag*{(252)} P 0 ( x ; a , b ) = 1 , P n ( x ; a , b ) = x h = 1 ∏ n − 1 ( x + ha + ( n − h ) b ) . (252)
We first recall the combinatorial interpretation [39 Section 6.1, Eqs. (6.10)–(6.11)] . For a rooted labeled forest F F F , call a vertex proper if every other vertex in its descendant subtree has a larger label. Let k ( F ) k(F) k ( F ) and p ( F ) p(F) p ( F ) be its numbers of components and proper vertices. We obtain P n ( x ; a , b ) P_n(x;a,b) P n ( x ; a , b ) by giving F F F weight x k ( F ) a n − p ( F ) b p ( F ) − k ( F ) x^{k(F)}a^{n-p(F)}b^{p(F)-k(F)} x k ( F ) a n − p ( F ) b p ( F ) − k ( F ) . Sokal conjectured joint coefficientwise Hankel total positivity in x , a , b x,a,b x , a , b . We verify the shifted moment condition below for every fixed a , b ≥ 0 a,b \ge0 a , b ≥ 0 , and then obtain a new universal coefficientwise result in the component intensities.
Writing P n , k ( a , b ) = [ x k ] P n ( x ; a , b ) P_{n,k}(a,b)=[x^k]P_n(x;a,b) P n , k ( a , b ) = [ x k ] P n ( x ; a , b ) , we recall the Hankel assertions of the original conjecture before giving the conclusion obtained from real cumulant lifting.
Conjecture 8.1 (Sokal, Conjecture 6.1(b),(c) in [39 ] ). The sequence ( P n ( x ; a , b ) ) n ≥ 0 (P_n(x;a,b))_{n\ge0} ( P n ( x ; a , b ) ) n ≥ 0 is coefficientwise Hankel-totally positive jointly in x , a , b x,a,b x , a , b . Moreover, the sequence ( P n + 1 , 1 ( a , b ) ) n ≥ 0 (P_{n+1,1}(a,b))_{n\ge0} ( P n + 1 , 1 ( a , b ) ) n ≥ 0 is coefficientwise Hankel-totally positive jointly in a , b a,b a , b .
For fixed a , b ≥ 0 a,b \ge0 a , b ≥ 0 , we verify the shifted moment condition and apply Theorem 1.7. The source identities below follow from [39 Section 6.1, Eqs. (6.1)–(6.4)] ; extracting the coefficient of x x x , we obtain the final identity.
Lemma 8.2 (Sokal’s generating-function identities). Let G a , b ( t ) ∈ t R [ [ t ] ] G_{a,b}(t)\in t\mathbb{R[[t]]} G a , b ( t ) ∈ t R [[ t ]] be the compositional inverse of
g a , b ( u ) = { e − b u − e − a u a − b , a ≠ b , u e − a u , a = b . (253) g_{a,b}(u)=
\begin{cases}
\frac{e^{-bu}-e^{-au}}{a-b}, & a\ne b,\\
ue^{-au}, & a=b.
\end{cases} \tag*{(253)} g a , b ( u ) = { a − b e − b u − e − a u , u e − a u , a = b , a = b . (253) Then g a , b ( 0 ) = 0 g_{a,b}(0)=0 g a , b ( 0 ) = 0 , g a , b ′ ( 0 ) = 1 g'_{a,b}(0)=1 g a , b ′ ( 0 ) = 1 , and
∑ n ≥ 0 P n ( x ; a , b ) t n n ! = e x G a , b ( t ) . (254) \sum_{n\ge0}P_n(x;a,b)\frac{t^n}{n!}=e^{xG_{a,b}(t)}. \tag*{(254)} n ≥ 0 ∑ P n ( x ; a , b ) n ! t n = e x G a , b ( t ) . (254) Consequently, if c n ( a , b ) : = P n + 1 , 1 ( a , b ) c_n(a,b):=P_{n+1,1}(a,b) c n ( a , b ) := P n + 1 , 1 ( a , b ) , then
c n ( a , b ) = G a , b ( n + 1 ) ( 0 ) = ∏ h = 1 n ( h a + ( n + 1 − h ) b ) ( n ≥ 0 ) . (255) c_n(a,b)=G_{a,b}^{(n+1)}(0)=\prod_{h=1}^{n}\bigl(ha+(n+1-h)b\bigr)\qquad(n\ge0). \tag*{(255)} c n ( a , b ) = G a , b ( n + 1 ) ( 0 ) = h = 1 ∏ n ( ha + ( n + 1 − h ) b ) ( n ≥ 0 ) . (255)
To construct a positive moment measure, we use the classical Bernstein–Widder theorem [35 Theorem 1.4] . We call a function h : ( 0 , ∞ ) → R h:(0,\infty)\to\mathbb{R} h : ( 0 , ∞ ) → R completely monotone if it is infinitely differentiable and ( − 1 ) n h ( n ) ( s ) ≥ 0 (-1)^n h^{(n)}(s)\ge0 ( − 1 ) n h ( n ) ( s ) ≥ 0 for every n ≥ 0 n\ge0 n ≥ 0 and s > 0 s>0 s > 0 .
Lemma 8.3 (Bernstein–Widder representation). A function h h h is completely monotone if and only if h ( s ) = ∫ 0 ∞ e − s z ν ( d z ) h(s)=\int_0^\infty e^{-sz}\,\nu(dz) h ( s ) = ∫ 0 ∞ e − sz ν ( d z ) for a positive measure ν \nu ν on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , with finite integral for every s > 0 s>0 s > 0 . If h ( 0 + ) < ∞ h(0+)<\infty h ( 0 + ) < ∞ , then ν ( [ 0 , ∞ ) ) = h ( 0 + ) \nu([0,\infty))=h(0+) ν ([ 0 , ∞ )) = h ( 0 + ) .
Using this representation and an explicit calculation of derivative signs, we construct the shifted moment measure required for lifting.
Lemma 8.4 (Shifted SGS moments at fixed shape parameters). For fixed a , b ≥ 0 a,b\ge0 a , b ≥ 0 , there is a probability measure ν a , b \nu_{a,b} ν a , b on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) such that
c n ( a , b ) = ∫ 0 ∞ z n ν a , b ( d z ) ( n ≥ 0 ) . (256) c_n(a,b)=\int_0^\infty z^n\,\nu_{a,b}(\mathrm{d}z) \qquad(n\ge0). \tag*{(256)} c n ( a , b ) = ∫ 0 ∞ z n ν a , b ( d z ) ( n ≥ 0 ) . (256) In particular, G a , b ∈ C G_{a,b}\in C G a , b ∈ C .
Proof. Step 1: A global inverse. By symmetry, assume a ≥ b ≥ 0 a\ge b\ge0 a ≥ b ≥ 0 . For u ≥ 0 u\ge0 u ≥ 0 , put
η ( u ) : = − g a , b ( − u ) = { e a u − e b u a − b , a > b , u e a u , a = b . (257) \eta(u):=-g_{a,b}(-u)=
\begin{cases}
\dfrac{e^{au}-e^{bu}}{a-b}, & a>b,\\
ue^{au}, & a=b.
\end{cases} \tag*{(257)} η ( u ) := − g a , b ( − u ) = ⎩ ⎨ ⎧ a − b e a u − e b u , u e a u , a > b , a = b . (257)
Since η ( 0 ) = 0 \eta(0)=0 η ( 0 ) = 0 , η ′ ( u ) > 0 \eta'(u)>0 η ′ ( u ) > 0 , and η ( u ) → ∞ \eta(u)\to\infty η ( u ) → ∞ , it has a smooth inverse ξ : [ 0 , ∞ ) → [ 0 , ∞ ) \xi:[0,\infty)\to[0,\infty) ξ : [ 0 , ∞ ) → [ 0 , ∞ ) . The inverse is analytic near zero because η ′ ( 0 ) = 1 \eta'(0)=1 η ′ ( 0 ) = 1 , and
η ( ξ ( s ) ) = s , ξ ( 0 ) = 0 , G a , b ( − s ) = − ξ ( s ) . (258) \eta(\xi(s))=s,\qquad\xi(0)=0,\qquad G_{a,b}(-s)=-\xi(s). \tag*{(258)} η ( ξ ( s )) = s , ξ ( 0 ) = 0 , G a , b ( − s ) = − ξ ( s ) . (258)
The last identity is used only near zero.
Step 2: Complete monotonicity of the reciprocal derivative. Put R = 1 / η ′ R=1/\eta' R = 1/ η ′ . If a > b a>b a > b , let d = a − b > 0 d=a-b>0 d = a − b > 0 . Then
R ( u ) = d a e a u − b e b u = d a e − a u 1 − ( b / a ) e − d u = d a ∑ k ≥ 0 ( b / a ) k e − ( a + k d ) u . (259) R(u)=\frac{d}{ae^{au}-be^{bu}}=\frac{d}{a}\frac{e^{-au}}{1-(b/a)e^{-du}}=\frac{d}{a}\sum_{k\ge0}(b/a)^k e^{-(a+kd)u}. \tag*{(259)} R ( u ) = a e a u − b e b u d = a d 1 − ( b / a ) e − d u e − a u = a d k ≥ 0 ∑ ( b / a ) k e − ( a + k d ) u . (259)
For b = 0 b=0 b = 0 , only the k = 0 k=0 k = 0 term remains. For b > 0 b>0 b > 0 , every differentiated series converges uniformly on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , since ∑ k ≥ 0 ( b / a ) k ( a + k d ) n < ∞ \sum_{k\ge0}(b/a)^k(a+kd)^n<\infty ∑ k ≥ 0 ( b / a ) k ( a + k d ) n < ∞ . Differentiating term by term, we obtain ( − 1 ) n R ( n ) ≥ 0 (-1)^nR^{(n)}\ge0 ( − 1 ) n R ( n ) ≥ 0 . If a = b = c a=b=c a = b = c ,
R ( u ) = e − c u 1 + c u . (260) R(u)=\frac{e^{-cu}}{1+cu}. \tag*{(260)} R ( u ) = 1 + c u e − c u . (260)
Both factors are completely monotone, including at c = 0 c=0 c = 0 , so their product is also completely monotone by Leibniz’s rule.
Step 3: Complete monotonicity of the inverse derivative. Differentiating the inverse relation, we obtain ξ ′ = R ∘ ξ \xi'=R\circ\xi ξ ′ = R ∘ ξ . Define
H 0 ( u ) = R ( u ) , H n + 1 ( u ) = − R ( u ) H n ′ ( u ) . (261) H_0(u)=R(u),\qquad H_{n+1}(u)=-R(u)H_n'(u). \tag*{(261)} H 0 ( u ) = R ( u ) , H n + 1 ( u ) = − R ( u ) H n ′ ( u ) . (261)
Complete monotonicity is preserved by h ↦ − h ′ h\mapsto-h' h ↦ − h ′ and by products, since
( − 1 ) n ( h 1 h 2 ) ( n ) ( u ) = ∑ k = 0 n ( n k ) ( − 1 ) k h 1 ( k ) ( u ) ( − 1 ) n − k h 2 ( n − k ) ( u ) ≥ 0. (262) (-1)^n(h_1h_2)^{(n)}(u)=\sum_{k=0}^{n}\binom{n}{k}(-1)^kh_1^{(k)}(u)(-1)^{n-k}h_2^{(n-k)}(u)\ge0. \tag*{(262)} ( − 1 ) n ( h 1 h 2 ) ( n ) ( u ) = k = 0 ∑ n ( k n ) ( − 1 ) k h 1 ( k ) ( u ) ( − 1 ) n − k h 2 ( n − k ) ( u ) ≥ 0. (262)
Inductively, every H n H_n H n is completely monotone and nonnegative. Using the chain rule, we obtain
( − 1 ) n d n d s n ξ ′ ( s ) = H n ( ξ ( s ) ) ≥ 0 ( n ≥ 0 , s > 0 ) . (263) (-1)^n \frac{\mathrm{d}^n}{\mathrm{d}s^n}\xi'(s)=H_n(\xi(s))\ge0 \qquad(n\ge0,\,s>0). \tag*{(263)} ( − 1 ) n d s n d n ξ ′ ( s ) = H n ( ξ ( s )) ≥ 0 ( n ≥ 0 , s > 0 ) . (263)
Differentiating and changing the sign, we obtain − H n ′ ( ξ ( s ) ) R ( ξ ( s ) ) = H n + 1 ( ξ ( s ) ) -H_n'(\xi(s))R(\xi(s))=H_{n+1}(\xi(s)) − H n ′ ( ξ ( s )) R ( ξ ( s )) = H n + 1 ( ξ ( s )) . Thus ξ ′ \xi' ξ ′ is completely monotone.
Step 4: Shifted moment representation. Since ξ ′ ( 0 ) = 1 \xi'(0)=1 ξ ′ ( 0 ) = 1 , applying Lemma 8.3 to ξ ′ \xi' ξ ′ , we obtain a probability measure ν a , b \nu_{a,b} ν a , b with
ξ ′ ( s ) = ∫ 0 ∞ e − s z ν a , b ( d z ) ( s > 0 ) . (264) \xi'(s)=\int_0^\infty e^{-sz}\,\nu_{a,b}(\mathrm{d}z) \qquad(s>0). \tag*{(264)} ξ ′ ( s ) = ∫ 0 ∞ e − sz ν a , b ( d z ) ( s > 0 ) . (264)
For fixed s > 0 s>0 s > 0 , boundedness of z n e − s z / 2 z^n e^{-sz/2} z n e − sz /2 and finiteness of the measure justify differentiation:
( − 1 ) n ( ξ ′ ) ( n ) ( s ) = ∫ 0 ∞ z n e − s z ν a , b ( d z ) . (265) (-1)^n(\xi')^{(n)}(s)=\int_0^\infty z^n e^{-sz}\,\nu_{a,b}(\mathrm{d}z). \tag*{(265)} ( − 1 ) n ( ξ ′ ) ( n ) ( s ) = ∫ 0 ∞ z n e − sz ν a , b ( d z ) . (265)
Near zero, ξ ′ ( s ) = G a , b ′ ( − s ) \xi'(s)=G'_{a,b}(-s) ξ ′ ( s ) = G a , b ′ ( − s ) . As s ↓ 0 s\downarrow0 s ↓ 0 , the left side tends to G a , b ( n + 1 ) ( 0 ) = c n ( a , b ) G_{a,b}^{(n+1)}(0)=c_n(a,b) G a , b ( n + 1 ) ( 0 ) = c n ( a , b ) . Applying monotone convergence to the integrals on the right, we obtain (8.5), including finiteness of every moment.
Having verified the moment condition, we apply the lifting theorem to obtain the following general result. We allow different fixed shape parameters in each color and an arbitrary initial moment series; the original one-color family is a specialization.
Theorem 8.5 (Universal lifting for fixed SGS shape parameters). Fix a ℓ , b ℓ ≥ 0 a_\ell,b_\ell\ge0 a ℓ , b ℓ ≥ 0 , and let A ( t ) A(t) A ( t ) be any initial Stieltjes moment series. Define
∑ n ≥ 0 F n ( y ) t n n ! = A ( t ) exp ( ∑ ℓ = 1 m y ℓ G a ℓ , b ℓ ( t ) ) . (266) \sum_{n\ge0}F_n(\boldsymbol{y})\frac{t^n}{n!}
=A(t)\exp\left(\sum_{\ell=1}^{m}y_\ell G_{a_\ell,b_\ell}(t)\right). \tag*{(266)} n ≥ 0 ∑ F n ( y ) n ! t n = A ( t ) exp ( ℓ = 1 ∑ m y ℓ G a ℓ , b ℓ ( t ) ) . (266) Then ( F n ) (F_n) ( F n ) is coefficientwise Hankel-totally positive in y \boldsymbol{y} y . In particular, for fixed a , b ≥ 0 a,b\ge0 a , b ≥ 0 , ( P n ( x ; a , b ) ) (P_n(x;a,b)) ( P n ( x ; a , b )) is coefficientwise Hankel-totally positive in x x x , and is Stieltjes at every numerical x ≥ 0 x\ge0 x ≥ 0 .
Proof. Applying Lemma 8.4 to each fixed pair ( a ℓ , b ℓ ) (a_\ell,b_\ell) ( a ℓ , b ℓ ) , we obtain admissible exponents G a ℓ , b ℓ G_{a_\ell,b_\ell} G a ℓ , b ℓ . Applying Theorem 1.7 to these exponents and A A A , we obtain the joint coefficientwise assertion. Taking A = 1 A=1 A = 1 , one color, and y 1 = x y_1=x y 1 = x , we obtain the assertion for P n P_n P n from (8.3); taking nonnegative numerical x x x , we obtain its Stieltjes property.
We have obtained a fixed-shape Hankel consequence of Conjecture 8.1. The joint assertion in all three variables x , a , b x,a,b x , a , b remains open; our new generality concerns independent component intensities and initial moments.
Bessel and Lah polynomials from gamma moment inputs
We begin with two counting models. We count forests of non-plane full binary trees with leaves labeled by [ n ] [n] [ n ] , giving each tree component weight x x x . A component with j j j leaves has ( 2 j − 3 ) ! ! (2j-3)!! ( 2 j − 3 )!! realizations, with the one-leaf tree included. Using the labeled set construction, we obtain the reverse Bessel polynomials B n ( x ) B_n(x) B n ( x ) . We also count partitions of [ n ] [n] [ n ] into linearly ordered blocks, giving each block weight x x x ; we obtain the Lah polynomials L n ( x ) L_n(x) L n ( x ) . Set B 0 = L 0 = 1 B_0=L_0=1 B 0 = L 0 = 1 .
∑ n ≥ 0 B n ( x ) t n n ! = exp ( x ( 1 − 1 − 2 t ) ) , B n ( x ) = ∑ k = 1 n ( 2 n − k − 1 ) ! x k 2 n − k ( n − k ) ! ( k − 1 ) ! . (267) \sum_{n\geq0} B_n(x)\frac{t^n}{n!}=\exp\left(x\left(1-\sqrt{1-2t}\right)\right),\qquad B_n(x)=\sum_{k=1}^{n}\frac{(2n-k-1)!x^k}{2^{n-k}(n-k)!(k-1)!}. \tag*{(267)} n ≥ 0 ∑ B n ( x ) n ! t n = exp ( x ( 1 − 1 − 2 t ) ) , B n ( x ) = k = 1 ∑ n 2 n − k ( n − k )! ( k − 1 )! ( 2 n − k − 1 )! x k . (267)
∑ n ≥ 0 L n ( x ) t n n ! = exp ( x t 1 − t ) , L n ( x ) = ∑ k = 1 n n ! k ! ( n − 1 k − 1 ) x k . (268) \sum_{n\geq0} L_n(x)\frac{t^n}{n!}=\exp\left(\frac{xt}{1-t}\right),\qquad L_n(x)=\sum_{k=1}^{n}\frac{n!}{k!}\binom{n-1}{k-1}x^k. \tag*{(268)} n ≥ 0 ∑ L n ( x ) n ! t n = exp ( 1 − t x t ) , L n ( x ) = k = 1 ∑ n k ! n ! ( k − 1 n − 1 ) x k . (268)
The finite formulas hold for n ≥ 1 n\geq1 n ≥ 1 . In [47 Propositions 4.4 and 4.7] , we proved coefficientwise Hankel total positivity of these polynomial sequences. We will obtain both conclusions below from gamma moment inputs and a universal lifting result, with a different proof.
For the binary-tree exponent 1 − 1 − 2 t 1-\sqrt{1-2t} 1 − 1 − 2 t , the shifted cumulants are ( 2 n − 1 ) ! ! = 2 n ( 1 / 2 ) n (2n-1)!!=2^n(1/2)_n ( 2 n − 1 )!! = 2 n ( 1/2 ) n . For the ordered-block exponent t / ( 1 − t ) t/(1-t) t / ( 1 − t ) , they are ( n + 1 ) ! = ( 2 ) n (n+1)!=(2)_n ( n + 1 )! = ( 2 ) n . We represent these sequences by the gamma laws Gamma ( 1 / 2 , 2 ) \operatorname{Gamma}(1/2,2) Gamma ( 1/2 , 2 ) and Gamma ( 2 , 1 ) \operatorname{Gamma}(2,1) Gamma ( 2 , 1 ) , respectively. We now use the whole gamma moment family to obtain a common extension of the two models.
For p ∈ R p\in\mathbb{R} p ∈ R and c ≥ 0 c\geq0 c ≥ 0 , define
K p , c ( t ) = ∑ n ≥ 1 ( p ) n − 1 c n − 1 t n n ! , K p , c ′ ( t ) = ( 1 − c t ) − p . (269) K_{p,c}(t)=\sum_{n\geq1}(p)_{n-1}c^{n-1}\frac{t^n}{n!},\qquad K'_{p,c}(t)=(1-ct)^{-p}. \tag*{(269)} K p , c ( t ) = n ≥ 1 ∑ ( p ) n − 1 c n − 1 n ! t n , K p , c ′ ( t ) = ( 1 − c t ) − p . (269)
Here K p , 0 = K 0 , c = t K_{p,0}=K_{0,c}=t K p , 0 = K 0 , c = t . For c > 0 c>0 c > 0 , we have
K p , c ( t ) = { 1 − ( 1 − c t ) 1 − p c ( 1 − p ) , p ≠ 1 , − log ( 1 − c t ) c , p = 1. (270) K_{p,c}(t)=
\begin{cases}
\dfrac{1-(1-ct)^{1-p}}{c(1-p)}, & p\ne1,\\[6pt]
-\dfrac{\log(1-ct)}{c}, & p=1.
\end{cases} \tag*{(270)} K p , c ( t ) = ⎩ ⎨ ⎧ c ( 1 − p ) 1 − ( 1 − c t ) 1 − p , − c log ( 1 − c t ) , p = 1 , p = 1. (270)
For p , c > 0 p,c>0 p , c > 0 , take the gamma measure
ν p , c ( d x ) = x p − 1 e − x / c Γ ( p ) c p d x ( x > 0 ) (271) \nu_{p,c}(\mathrm{d}x)=\frac{x^{p-1}e^{-x/c}}{\Gamma(p)c^p}\,\mathrm{d}x\qquad(x>0) \tag*{(271)} ν p , c ( d x ) = Γ ( p ) c p x p − 1 e − x / c d x ( x > 0 ) (271)
Applying Euler’s integral and the gamma-ratio formula [27 Eqs. 5.2.1 and 5.2.5] to this measure, we obtain ∫ 0 ∞ x n ν p , c ( d x ) = c n ( p ) n \int_0^\infty x^n\nu_{p,c}(\mathrm{d}x)=c^n(p)_n ∫ 0 ∞ x n ν p , c ( d x ) = c n ( p ) n . Thus we have verified the once-shifted moment condition for K p , c K_{p,c} K p , c before applying the lifting theorem.
Corollary 8.6 (Universal gamma component families). Fix p ℓ , c ℓ ≥ 0 p_\ell,c_\ell\geq0 p ℓ , c ℓ ≥ 0 . For any initial Stieltjes moment series A A A , define
∑ n ≥ 0 P n ( y ) t n n ! = A ( t ) exp ( ∑ ℓ = 1 m y ℓ K p ℓ , c ℓ ( t ) ) (272) \sum_{n\geq0}P_n(y)\frac{t^n}{n!}=A(t)\exp\left(\sum_{\ell=1}^{m}y_\ell K_{p_\ell,c_\ell}(t)\right) \tag*{(272)} n ≥ 0 ∑ P n ( y ) n ! t n = A ( t ) exp ( ℓ = 1 ∑ m y ℓ K p ℓ , c ℓ ( t ) ) (272)
Then ( P n ) (P_n) ( P n ) is coefficientwise Hankel-totally positive in y \mathbf{y} y . For c > 0 c>0 c > 0 , the condition p ≥ 0 p\geq0 p ≥ 0 is necessary for K p , c ∈ C K_{p,c}\in\mathcal{C} K p , c ∈ C .
Proof. Using the gamma measures above, we have K p , c = I b K_{p,c}=I_b K p , c = I b , with b n = c n ( p ) n b_n=c^n(p)_n b n = c n ( p ) n . At p = 0 p=0 p = 0 or c = 0 c=0 c = 0 , use the shifted measure δ 0 \delta_0 δ 0 . Applying Theorem 1.7 to these exponents and A ( t ) A(t) A ( t ) , we obtain the joint coefficientwise assertion. For p < 0 < c p<0<c p < 0 < c , we have second cumulant p c < 0 pc<0 p c < 0 , so we cannot have K p , c ∈ C K_{p,c}\in\mathcal{C} K p , c ∈ C .
We allow independent shapes and scales, arbitrary initial moments, and independent component intensities in this result. Taking ( p , c ) = ( 1 / 2 , 2 ) (p,c)=(1/2,2) ( p , c ) = ( 1/2 , 2 ) or ( 2 , 1 ) (2,1) ( 2 , 1 ) , we recover the two classical polynomial families. We may also retain a common scale as an indeterminate.
Corollary 8.7. (A common scale as an indeterminate). Fix p 1 , … , p m ≥ 0 p_1,\ldots,p_m\geq0 p 1 , … , p m ≥ 0 , and let A A A be any initial Stieltjes moment series. The exponential coefficients of
A ( c t ) exp ( ∑ ℓ = 1 m y ℓ K p ℓ , c ( t ) ) (273) A(ct)\exp\left(\sum_{\ell=1}^{m}y_{\ell}K_{p_{\ell},c}(t)\right) \tag*{(273)} A ( c t ) exp ( ℓ = 1 ∑ m y ℓ K p ℓ , c ( t ) ) (273) are jointly coefficientwise Hankel-totally positive in c , y c,\mathbf{y} c , y .
Proof. Let R n ( y ) R_n(\mathbf{y}) R n ( y ) be the coefficients at unit scale. For c > 0 c>0 c > 0 , using K p , c ( t ) = c − 1 K p , 1 ( c t ) K_{p,c}(t)=c^{-1}K_{p,1}(ct) K p , c ( t ) = c − 1 K p , 1 ( c t ) , we obtain coefficients c n R n ( y / c ) c^nR_n(\mathbf{y}/c) c n R n ( y / c ) . For increasing index lists I , J I,J I , J , let S = ∑ a i a + ∑ b j b S=\sum_a i_a+\sum_b j_b S = ∑ a i a + ∑ b j b . Factoring the row and column scales, we obtain
c S det [ R i a + j b ( y / c ) ] a , b = 1 r . (274) c^S\det\left[R_{i_a+j_b}(\mathbf{y}/c)\right]_{a,b=1}^{r}. \tag*{(274)} c S det [ R i a + j b ( y / c ) ] a , b = 1 r . (274) Applying Corollary 8.6 to the unit-scale family, we obtain a determinant with nonnegative coefficients. Since its degree is at most S S S , we multiply each monomial y α \mathbf{y}^{\alpha} y α by a nonnegative power c S − ∣ α ∣ c^{S-|\alpha|} c S − ∣ α ∣ . Thus we obtain a polynomial identity valid also at c = 0 c=0 c = 0 , with the initial factor A ( c t ) A(ct) A ( c t ) specified in the statement.
Corollary 8.8. (Recovery of the Bessel and Lah results). The sequences ( B n ( x ) ) (B_n(x)) ( B n ( x )) and ( L n ( x ) ) (L_n(x)) ( L n ( x )) defined in (8.16)–(8.17) are coefficientwise Hankel-totally positive in x x x .
Proof. Applying Corollary 8.6 to one-component families with A ( t ) = 1 A(t)=1 A ( t ) = 1 and indeterminate x x x , we obtain the two assertions from K 1 / 2 , 2 = 1 − 1 − 2 t K_{1/2,2}=1-\sqrt{1-2t} K 1/2 , 2 = 1 − 1 − 2 t and K 2 , 1 = t / ( 1 − t ) K_{2,1}=t/(1-t) K 2 , 1 = t / ( 1 − t ) , respectively.
For numerical x > 0 x>0 x > 0 , we may also represent B n ( x ) B_n(x) B n ( x ) by an inverse Gaussian law with mean x x x and shape x 2 x^2 x 2 [13 ] . We represent L n ( x ) L_n(x) L n ( x ) by a compound-Poisson sum with rate x x x and unit exponential jumps. More generally, for p > 1 , c > 0 p>1,c>0 p > 1 , c > 0 , we write
K p , c ( t ) = ( 1 − c t ) − ( p − 1 ) − 1 c ( p − 1 ) . (275) K_{p,c}(t)=\frac{(1-ct)^{-(p-1)}-1}{c(p-1)}. \tag*{(275)} K p , c ( t ) = c ( p − 1 ) ( 1 − c t ) − ( p − 1 ) − 1 . (275)
We represent e y K p , c e^{yK_{p,c}} e y K p , c by gamma jumps of shape p − 1 p-1 p − 1 and scale c c c , with Poisson rate y / [ c ( p − 1 ) ] y/[c(p-1)] y / [ c ( p − 1 )] . At p = 1 p=1 p = 1 , we use a gamma law of shape y / c y/c y / c , including its zero-shape degenerate case.
For generalized Lah polynomials, their Hankel total-positivity result was proved in [47 Proposition 4.7] and [48 Section 6.8, Eq. (6.174) and Proposition 6.30] . For fixed a , b , c > 0 a,b,c>0 a , b , c > 0 and independent d , q , λ d,q,\lambda d , q , λ , put
f ( t ) = c b ( ( 1 − a b t ) − 1 / a − 1 ) , A ( t ) ( 1 − a b t ) − d exp ( ( q + λ ) f ( t ) ) . (276) f(t)=\frac{c}{b}\left((1-abt)^{-1/a}-1\right),\qquad A(t)(1-abt)^{-d}\exp((q+\lambda)f(t)). \tag*{(276)} f ( t ) = b c ( ( 1 − ab t ) − 1/ a − 1 ) , A ( t ) ( 1 − ab t ) − d exp (( q + λ ) f ( t )) . (276)
We represent the shifted cumulants of f f f , namely c ( 1 + 1 / a ) n ( a b ) n c(1+1/a)_n(ab)^n c ( 1 + 1/ a ) n ( ab ) n , by c c c times the gamma law of shape 1 + 1 / a 1+1/a 1 + 1/ a and scale a b ab ab . Applying Corollary 8.6 to − log ( 1 − a b t ) -\log(1-abt) − log ( 1 − ab t ) , we also obtain its admissibility. Applying Theorem 1.7 to these two exponents, and replacing the intensity of f f f by q + λ q+\lambda q + λ , we obtain joint coefficientwise Hankel total positivity in d , q , λ d,q,\lambda d , q , λ . We allow continuous a , b , c > 0 a,b,c>0 a , b , c > 0 and arbitrary initial moments; we do not require the restrictions a , a d ∈ N 0 a,ad\in\mathbb{N}_0 a , a d ∈ N 0 and a d ≤ a + 1 ad\leq a+1 a d ≤ a + 1 used in the cited production-matrix result.
Univariate Laguerre polynomials and universal component families
We count assemblies of permutation cycles and linearly ordered blocks on [ n ] [n] [ n ] , with weights β \beta β per cycle and x x x per ordered block. Using the classical exponential formula, we obtain unsigned Laguerre polynomials with the normalization
Λ n ( x ; β ) : = n ! L n ( β − 1 ) ( − x ) . (277) \Lambda_n(x;\beta):=n!L_n^{(\beta-1)}(-x). \tag*{(277)} Λ n ( x ; β ) := n ! L n ( β − 1 ) ( − x ) . (277)
For β > 0 \beta>0 β > 0 , L n ( β − 1 ) ( z ) L_n^{(\beta-1)}(z) L n ( β − 1 ) ( z ) is orthogonal on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with weight z β − 1 e − z z^{\beta-1}e^{-z} z β − 1 e − z . In [47 Proposition 4.14] and [48 Proposition 6.10] , scalar coefficientwise Hankel total positivity was proved the joint parameter case using [48 Theorem 5.2] ; see also the rook-polynomial interpretation in the latter work. We retain the Laguerre parameter as a component indeterminate and obtain the general independently colored extension below. The normalization, finite formula and generating function are classical [28 Table 18.3.1 and Eqs. 18.5.12, 18.12.13] . At β = 0 \beta=0 β = 0 , we interpret the finite expression by polynomial continuation. We first record these coefficient identities.
Lemma 8.9 (Classical unsigned Laguerre formulas). For n ≥ 0 n\geq0 n ≥ 0 , we have
Λ n ( x ; β ) = ∑ k = 0 n ( n k ) ( β + k ) n − k x k , ∑ n ≥ 0 Λ n ( x ; β ) t n n ! = ( 1 − t ) − β exp ( x t 1 − t ) . (278) \Lambda_n(x;\beta)=\sum_{k=0}^{n}\binom{n}{k}(\beta+k)_{n-k}x^k,\qquad\sum_{n\geq0}\Lambda_n(x;\beta)\frac{t^n}{n!}=(1-t)^{-\beta}\exp\left(\frac{xt}{1-t}\right). \tag*{(278)} Λ n ( x ; β ) = k = 0 ∑ n ( k n ) ( β + k ) n − k x k , n ≥ 0 ∑ Λ n ( x ; β ) n ! t n = ( 1 − t ) − β exp ( 1 − t x t ) . (278)
For every fixed β ≥ 0 \beta\geq0 β ≥ 0 , the sequence ( Λ n ( x ; β ) ) n ≥ 0 (\Lambda_n(x;\beta))_{n\geq0} ( Λ n ( x ; β ) ) n ≥ 0 is coefficientwise Hankel-totally positive in x x x , as in [48 Proposition 6.10(ii)] .
We first verify the two shifted gamma moment conditions. Applying lifting to these inputs, we obtain arbitrary initial moments, independent scales, and separate intensities for cycles and ordered blocks. Taking one color, unit scale, and A = 1 A=1 A = 1 , we recover the known Laguerre coefficientwise positivity by this different proof.
For fixed c ≥ 0 c \ge0 c ≥ 0 , put
U c ( t ) : = − log ( 1 − c t ) , V c ( t ) : = t 1 − c t . (279) U_c(t):=-\log(1-ct),\qquad V_c(t):=\frac{t}{1-ct}. \tag*{(279)} U c ( t ) := − log ( 1 − c t ) , V c ( t ) := 1 − c t t . (279)
Their shifted coefficients are
U c ( n + 1 ) ( 0 ) = n ! c n + 1 , V c ( n + 1 ) ( 0 ) = ( n + 1 ) ! c n ( n ≥ 0 ) . (280) U_c^{(n+1)}(0)=n!c^{n+1},\qquad V_c^{(n+1)}(0)=(n+1)!c^n\qquad(n\ge0). \tag*{(280)} U c ( n + 1 ) ( 0 ) = n ! c n + 1 , V c ( n + 1 ) ( 0 ) = ( n + 1 )! c n ( n ≥ 0 ) . (280)
For c > 0 c>0 c > 0 , applying Euler’s integral to these shifted coefficients, we obtain
n ! c n + 1 = ∫ 0 ∞ s n e − s / c d s , ( n + 1 ) ! c n = ∫ 0 ∞ s n e − s / c c 2 d s . (281) n!c^{n+1}=\int_0^\infty s^n e^{-s/c}\,\mathrm{d}s,\qquad(n+1)!c^n=\int_0^\infty\frac{s^n e^{-s/c}}{c^2}\,\mathrm{d}s. \tag*{(281)} n ! c n + 1 = ∫ 0 ∞ s n e − s / c d s , ( n + 1 )! c n = ∫ 0 ∞ c 2 s n e − s / c d s . (281)
At c = 0 c=0 c = 0 , use the zero measure for U 0 = 0 U_0=0 U 0 = 0 and δ 0 \delta_0 δ 0 for V 0 = t V_0=t V 0 = t . Thus U c , V c ∈ C U_c,V_c\in\mathcal{C} U c , V c ∈ C .
Theorem 8.10 (Universal Laguerre component lifting). Fix c 1 , … , c m ≥ 0 c_1,\ldots,c_m\ge0 c 1 , … , c m ≥ 0 , and let A ( t ) A(t) A ( t ) be any initial Stieltjes moment series. Introduce independent indeterminates β = ( β 1 , … , β m ) \boldsymbol{\beta}=(\beta_1,\ldots,\beta_m) β = ( β 1 , … , β m ) and x = ( x 1 , … , x m ) \boldsymbol{x}=(x_1,\ldots,x_m) x = ( x 1 , … , x m ) , and define
∑ n ≥ 0 F n ( β , x ) t n n ! = A ( t ) ∏ ℓ = 1 m ( 1 − c ℓ t ) − β ℓ exp ( ∑ ℓ = 1 m x ℓ t 1 − c ℓ t ) . (282) \sum_{n\ge0} F_n(\boldsymbol{\beta},\boldsymbol{x})\frac{t^n}{n!}
=A(t)\prod_{\ell=1}^{m}(1-c_\ell t)^{-\beta_\ell}\exp\left(\sum_{\ell=1}^{m}\frac{x_\ell t}{1-c_\ell t}\right). \tag*{(282)} n ≥ 0 ∑ F n ( β , x ) n ! t n = A ( t ) ℓ = 1 ∏ m ( 1 − c ℓ t ) − β ℓ exp ( ℓ = 1 ∑ m 1 − c ℓ t x ℓ t ) . (282) Then ( F n ) (F_n) ( F n ) is coefficientwise Hankel-totally positive jointly in β , x \boldsymbol{\beta},\boldsymbol{x} β , x . In particular, for m = 1 m=1 m = 1 and c 1 = 1 c_1=1 c 1 = 1 , we obtain
F n ( β , x ) = ∑ k = 0 n ( n k ) a n − k Λ k ( x ; β ) . (283) F_n(\beta,x)=\sum_{k=0}^{n}\binom{n}{k}a_{n-k}\Lambda_k(x;\beta). \tag*{(283)} F n ( β , x ) = k = 0 ∑ n ( k n ) a n − k Λ k ( x ; β ) . (283) which has the same joint positivity in β , x \beta,x β , x .
Proof. Applying Theorem 1.7 to U c ℓ , V c ℓ U_{c_\ell},V_{c_\ell} U c ℓ , V c ℓ , with indeterminates β ℓ , x ℓ \beta_\ell,x_\ell β ℓ , x ℓ and initial series A A A , we obtain joint coefficientwise positivity for (282) . At m = 1 m=1 m = 1 , c 1 = 1 c_1=1 c 1 = 1 , multiplying (278) by A ( t ) A(t) A ( t ) and extracting n ! [ t n ] n![t^n] n ! [ t n ] , we obtain (283) .
For numerical x , β ≥ 0 x,\beta\ge0 x , β ≥ 0 , we represent the pure family at unit scale by the sum of an independent gamma variable of shape β \beta β and a compound-Poisson variable with rate x x x and exponential jumps of mean one. At β = 0 \beta=0 β = 0 , the gamma summand is zero. We thus identify the logarithmic exponent with the gamma component and the rational exponent with the Lah component already considered in Subsection 8.2. The first three polynomials are
Λ 0 = 1 , Λ 1 = x + β , Λ 2 = x 2 + 2 ( β + 1 ) x + β ( β + 1 ) . (284) \Lambda_0=1,\qquad\Lambda_1=x+\beta,\qquad\Lambda_2=x^2+2(\beta+1)x+\beta(\beta+1). \tag*{(284)} Λ 0 = 1 , Λ 1 = x + β , Λ 2 = x 2 + 2 ( β + 1 ) x + β ( β + 1 ) . (284)
For example, we obtain the explicit positive minor
Λ 0 Λ 2 − Λ 1 2 = β + 2 x . (285) \Lambda_0\Lambda_2-\Lambda_1^2=\beta+2x. \tag*{(285)} Λ 0 Λ 2 − Λ 1 2 = β + 2 x . (285)
For the next leading minor, we obtain
det [ Λ i + j ( x ; β ) ] i , j = 0 2 = 2 β 2 ( β + 1 ) + 12 β ( β + 1 ) x + 12 ( 2 β + 1 ) x 2 + 16 x 3 . (286) \det\left[\Lambda_{i+j}(x;\beta)\right]_{i,j=0}^{2}=2\beta^2(\beta+1)+12\beta(\beta+1)x+12(2\beta+1)x^2+16x^3. \tag*{(286)} det [ Λ i + j ( x ; β ) ] i , j = 0 2 = 2 β 2 ( β + 1 ) + 12 β ( β + 1 ) x + 12 ( 2 β + 1 ) x 2 + 16 x 3 . (286)
We may also combine the Laguerre components with the tree models above. In particular, for any real initial Stieltjes moment series A ( t ) A(t) A ( t ) , the exponential coefficients of
A ( t ) ( 1 − t ) − β exp ( x t 1 − t + y U ( t ) + z H 1 ( t ) ) (287) A(t)(1-t)^{-\beta}\exp\left(\frac{xt}{1-t}+yU(t)+zH_1(t)\right) \tag*{(287)} A ( t ) ( 1 − t ) − β exp ( 1 − t x t + y U ( t ) + z H 1 ( t ) ) (287)
are coefficientwise Hankel-totally positive jointly in β , x , y , z \beta,x,y,z β , x , y , z . Applying Theorem 1.7 to U 1 , V 1 , U , H 1 U_1,V_1,U,H_1 U 1 , V 1 , U , H 1 , for which we have already verified the moment hypotheses, we obtain the mixed example.
Set partitions and classical moment transforms
We count partitions whose blocks carry successive layers of nonempty sets, with singleton layers allowed. At depth one we obtain the ordinary Bell polynomials, whose coefficientwise Hankel positivity was already studied in [42 Section 3] . Starting from the moment input 1 , 1 , … 1,1,\ldots 1 , 1 , … , represented by δ 1 \delta_1 δ 1 , we obtain admissible iterated exponents by composition. Applying lifting, we extend the one-depth model to independent depths and colors with arbitrary initial moments.
Define
E 0 ( t ) = t , E d + 1 ( t ) = exp ( E d ( t ) ) − 1 ( d ≥ 0 ) . (288) E_0(t)=t,\qquad E_{d+1}(t)=\exp(E_d(t))-1\quad(d\ge0). \tag*{(288)} E 0 ( t ) = t , E d + 1 ( t ) = exp ( E d ( t )) − 1 ( d ≥ 0 ) . (288)
Iterating the construction of nonempty sets, we obtain the following families.
Corollary 8.11 (Hierarchical Bell families). Fix positive integers d 1 , … , d m d_1,\ldots,d_m d 1 , … , d m . For any initial Stieltjes moment series A ( t ) A(t) A ( t ) , the exponential coefficients of
A ( t ) exp ( ∑ ℓ = 1 m y ℓ E d ℓ ( t ) ) (289) A(t)\exp\left(\sum_{\ell=1}^{m}y_\ell E_{d_\ell}(t)\right) \tag*{(289)} A ( t ) exp ( ℓ = 1 ∑ m y ℓ E d ℓ ( t ) ) (289) are coefficientwise Hankel-totally positive in y \mathbf{y} y .
Proof. We represent the shifted cumulants of t t t and e t − 1 e^t-1 e t − 1 by δ 0 \delta_0 δ 0 and δ 1 \delta_1 δ 1 , respectively. Applying Theorem 4.4 successively to E d + 1 = ( e t − 1 ) ∘ E d E_{d+1}=(e^t-1)\circ E_d E d + 1 = ( e t − 1 ) ∘ E d , we obtain E d ∈ C E_d\in\mathcal{C} E d ∈ C . Applying Theorem 1.7 to the selected depths and A A A , we obtain the conclusion.
For A = 1 A=1 A = 1 , we use the exponential formula to interpret a color-ℓ \ell ℓ component as d ℓ d_\ell d ℓ nested layers of nonempty sets, allowing singleton layers. At depth 1, we obtain ordinary set partitions and Bell polynomials, while depth 2 groups nonempty blocks into nonempty collections. For this standard iteration of the nonempty-set construction, see [6 ] for the exponential formula. Differentiating (8.37) at zero, we obtain
E d ′ ( 0 ) = 1 , E d ′ ′ ( 0 ) = d , E d ′ ′ ′ ( 0 ) = d ( 3 d − 1 ) 2 . (290) E_d'(0)=1,\qquad E_d''(0)=d,\qquad E_d'''(0)=\frac{d(3d-1)}{2}. \tag*{(290)} E d ′ ( 0 ) = 1 , E d ′′ ( 0 ) = d , E d ′′′ ( 0 ) = 2 d ( 3 d − 1 ) . (290)
At the step from d d d to d + 1 d+1 d + 1 , the second and third derivatives increase by 1 and 3 d + 1 3d+1 3 d + 1 , respectively.
Writing Y = ∑ ℓ y ℓ Y=\sum_\ell y_\ell Y = ∑ ℓ y ℓ and V = ∑ ℓ d ℓ y ℓ V=\sum_\ell d_\ell y_\ell V = ∑ ℓ d ℓ y ℓ , the first pure-family moments are
M 1 = Y , M 2 = Y 2 + V , M 3 = Y 3 + 3 Y V + ∑ ℓ d ℓ ( 3 d ℓ − 1 ) 2 y ℓ . (291) M_1=Y,\qquad M_2=Y^2+V,\qquad M_3=Y^3+3YV+\sum_\ell\frac{d_\ell(3d_\ell-1)}{2}y_\ell. \tag*{(291)} M 1 = Y , M 2 = Y 2 + V , M 3 = Y 3 + 3 Y V + ℓ ∑ 2 d ℓ ( 3 d ℓ − 1 ) y ℓ . (291)
In particular, M 0 M 2 − M 1 2 = V M_0M_2-M_1^2=V M 0 M 2 − M 1 2 = V . For an arbitrary initial Stieltjes moment sequence, the component of degree ( r 2 ) \binom{r}{2} ( 2 r ) in its leading determinant of order r r r is
a 0 r ( ∏ j = 0 r − 1 j ! ) ( ∑ ℓ d ℓ y ℓ ) ( r 2 ) , (292) a_0^r\left(\prod_{j=0}^{r-1}j!\right)\left(\sum_\ell d_\ell y_\ell\right)^{\binom{r}{2}}, \tag*{(292)} a 0 r ( j = 0 ∏ r − 1 j ! ) ( ℓ ∑ d ℓ y ℓ ) ( 2 r ) , (292)
by Theorem 1.8 .
For a formal power series L L L with L ( 0 ) = 0 L(0)=0 L ( 0 ) = 0 , put
R n , k ( L ) = n ! k ! [ t n ] L ( t ) k . (293) R_{n,k}(L)=\frac{n!}{k!}[t^n]L(t)^k. \tag*{(293)} R n , k ( L ) = k ! n ! [ t n ] L ( t ) k . (293)
Formal substitution transforms an initial sequence into b n = ∑ k = 0 n R n , k ( L ) a k b_n=\sum_{k=0}^n R_{n,k}(L)a_k b n = ∑ k = 0 n R n , k ( L ) a k . For the criterion based on row moment sequences, see Wang and Zhu [42 Theorem 4.2] . By their Corollary 4.4(i), we obtain the second-kind Stirling case. The first-kind case is the gamma moment family ( s ) n (s)_n ( s ) n at a nonnegative argument. For the Bessel and Lah cases, we may also use [47 Propositions 4.4(v) and 4.7(vi)] . The coefficient arrays are the classical ones in those sources.
Lemma 8.12 (Classical Stieltjes-preserving transforms). Each transform in the following table sends every Stieltjes moment sequence to a Stieltjes moment sequence:
Table 1.
In all four cases, R 0 , 0 = 1 R_{0,0}=1 R 0 , 0 = 1 and R n , 0 = 0 R_{n,0}=0 R n , 0 = 0 for n > 0 n>0 n > 0 .
For each L ( t ) L(t) L ( t ) in four examples above, it is easy to obtain the corresponding shifted cumulant sequence ( κ n + 1 ) n ≥ 0 (\kappa_{n+1})_{n\geq0} ( κ n + 1 ) n ≥ 0 is a Stieltjes moment sequence. After any finite composition of these known transforms, we may convolve with the family defined by
exp ( ∑ ℓ = 1 m y ℓ K ℓ ( t ) ) , K ℓ ∈ C . (294) \exp\left(\sum_{\ell=1}^{m} y_{\ell}K_{\ell}(t)\right), \qquad K_{\ell}\in\mathcal{C}. \tag*{(294)} exp ( ℓ = 1 ∑ m y ℓ K ℓ ( t ) ) , K ℓ ∈ C . (294)
By Theorem 4.5 , we obtain Hankel minors with nonnegative coefficients jointly in the indeterminates. We thus combine the four known numerical preservation results with universal cumulant lifting.
For example, for any initial Stieltjes moment series A A A and fixed q ≥ 0 q\geq0 q ≥ 0 , the coefficients of
A ( e t − 1 ) ( 1 − q T ( T ( t ) ) ) − y exp ( x ( 1 − 1 − 2 t ) + z t 1 − t + w ( e t − 1 − 1 ) ) (295) A(e^{t}-1)(1-qT(T(t)))^{-y}\exp\left(x(1-\sqrt{1-2t})+\frac{zt}{1-t}+w(e^{t-1}-1)\right) \tag*{(295)} A ( e t − 1 ) ( 1 − q T ( T ( t )) ) − y exp ( x ( 1 − 1 − 2 t ) + 1 − t z t + w ( e t − 1 − 1 ) ) (295)
are coefficientwise Hankel-totally positive jointly in x , z , w , y x,z,w,y x , z , w , y .
Here the Stirling substitution and the Bessel, Lah, depth-2 Bell, and iterated-tree exponents all belong to C \mathcal{C} C . Applying Theorem 4.5 to these exponents and the Stirling substitution, we obtain the asserted joint positivity.
Compound-Poisson moment families and lognormal jumps
Let ρ \rho ρ be any probability measure on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) with all moments finite, and write b n = ∫ x n ρ ( d x ) b_n=\int x^n\rho(\mathrm{d}x) b n = ∫ x n ρ ( d x ) . The jump exponent J b ( t ) = ∑ n ≥ 1 b n t n / n ! J_b(t)=\sum_{n\geq1}b_n t^n/n! J b ( t ) = ∑ n ≥ 1 b n t n / n ! in Theorem 4.6 has once-shifted moments represented by x ρ ( d x ) x\rho(\mathrm{d}x) x ρ ( d x ) . Hence arbitrary independently chosen jump laws give universal component families with arbitrary initial Stieltjes moments. At numerical intensities, these are the classical compound-Poisson moment families [22 Chapter 3] .
For any jump moment sequence b n b_n b n , we can also count set partitions with block weight b ∣ B ∣ b_{\lvert B\rvert} b ∣ B ∣ and a component weight for each block. We specialize these weighted counts to the known lognormal moments. Verifying the shifted moment condition, we obtain coefficientwise Hankel positivity in the Poisson-rate variables, including divergent moment series.
Proposition 8.13 (Lognormal jumps). Fix μ ℓ ∈ R \mu_{\ell}\in\mathbb{R} μ ℓ ∈ R and σ ℓ 2 ≥ 0 \sigma_{\ell}^{2}\geq0 σ ℓ 2 ≥ 0 , and put
K ℓ ( t ) = ∑ n ≥ 1 exp ( n μ ℓ + n 2 σ ℓ 2 2 ) t n n ! . (296) K_{\ell}(t)=\sum_{n\geq1}\exp\left(n\mu_{\ell}+\frac{n^{2}\sigma_{\ell}^{2}}{2}\right)\frac{t^{n}}{n!}. \tag*{(296)} K ℓ ( t ) = n ≥ 1 ∑ exp ( n μ ℓ + 2 n 2 σ ℓ 2 ) n ! t n . (296) For any initial Stieltjes moment series A ( t ) A(t) A ( t ) , the series A ( t ) exp ( ∑ ℓ y ℓ K ℓ ( t ) ) A(t)\exp\left(\sum_{\ell}y_{\ell}K_{\ell}(t)\right) A ( t ) exp ( ∑ ℓ y ℓ K ℓ ( t ) ) has an exponential coefficient sequence that is jointly coefficientwise Hankel-totally positive in y y y . Whenever σ ℓ 2 > 0 \sigma_{\ell}^{2}>0 σ ℓ 2 > 0 , the series K ℓ K_{\ell} K ℓ has radius of convergence zero.
Proof. Put σ ℓ = σ ℓ 2 \sigma_{\ell}=\sqrt{\sigma_{\ell}^{2}} σ ℓ = σ ℓ 2 and let X ℓ = exp ( μ ℓ + σ ℓ Z ) X_{\ell}=\exp(\mu_{\ell}+\sigma_{\ell}Z) X ℓ = exp ( μ ℓ + σ ℓ Z ) , where Z Z Z is standard normal. Integrating against its density, we obtain
E X ℓ n = exp ( n μ ℓ + n 2 σ ℓ 2 2 ) . (297) \mathbb{E}X_{\ell}^{n}=\exp\left(n\mu_{\ell}+\frac{n^{2}\sigma_{\ell}^{2}}{2}\right). \tag*{(297)} E X ℓ n = exp ( n μ ℓ + 2 n 2 σ ℓ 2 ) . (297)
Thus b n ( ℓ ) = E X ℓ n b_{n}^{(\ell)}=\mathbb{E}X_{\ell}^{n} b n ( ℓ ) = E X ℓ n is Stieltjes and K ℓ = J b ( ℓ ) K_{\ell}=J_{b^{(\ell)}} K ℓ = J b ( ℓ ) . Applying Theorem 4.6 to these jump moment inputs, with S = t S=t S = t and zero logarithmic intensities, we obtain the positivity assertion. For σ ℓ 2 > 0 \sigma_{\ell}^{2}>0 σ ℓ 2 > 0 , applying Stirling’s formula to the exponential coefficient, we have
( exp ( n μ ℓ + n 2 σ ℓ 2 / 2 ) n ! ) 1 / n ∼ e n exp ( μ ℓ + n σ ℓ 2 / 2 ) ⟶ ∞ , (298) \left(\frac{\exp(n\mu_{\ell}+n^{2}\sigma_{\ell}^{2}/2)}{n!}\right)^{1/n}\sim\frac{e}{n}\exp(\mu_{\ell}+n\sigma_{\ell}^{2}/2)\longrightarrow\infty, \tag*{(298)} ( n ! exp ( n μ ℓ + n 2 σ ℓ 2 /2 ) ) 1/ n ∼ n e exp ( μ ℓ + n σ ℓ 2 /2 ) ⟶ ∞ , (298)
so the radius of convergence is zero. □
For numerical y ≥ 0 y\geq0 y ≥ 0 , the pure-family coefficients are the moments of
S y = ∑ ℓ = 1 m ∑ j = 1 N ℓ X ℓ , j , N ℓ ∼ Poisson ( y ℓ ) , (299) S_y=\sum_{\ell=1}^{m}\sum_{j=1}^{N_{\ell}}X_{\ell,j},\qquad N_{\ell}\sim\operatorname{Poisson}(y_{\ell}), \tag*{(299)} S y = ℓ = 1 ∑ m j = 1 ∑ N ℓ X ℓ , j , N ℓ ∼ Poisson ( y ℓ ) , (299)
where, for each ℓ \ell ℓ , the variables X ℓ , j X_{\ell,j} X ℓ , j are independent copies of the lognormal variable X ℓ X_{\ell} X ℓ constructed above. We take all counts and jumps to be independent. Conditioning on the counts, or using the set-partition formula, we verify finiteness of every moment and recover the same formal coefficients. Applying Theorem 1.7 to the lognormal jump exponents, we obtain nonnegativity of all mixed coefficients in the Poisson rates in every Hankel minor. This remains valid when a lognormal jump of positive variance has positive rate, although the moment generating function is then infinite at every positive argument. Indeed, the event of exactly one such jump and no other jumps has positive probability, and E e t X ℓ = ∞ \mathbb{E}e^{tX_{\ell}}=\infty E e t X ℓ = ∞ for t > 0 t>0 t > 0 : in the normal integral, we compare the term t e μ ℓ + σ ℓ z te^{\mu_{\ell}+\sigma_{\ell}z} t e μ ℓ + σ ℓ z with the quadratic normal exponent and see that the former eventually dominates. We therefore need the formulation in terms of formal power series and finite moments for this application.
Poisson differences and Gaussian perturbations
The difference of independent Poisson counts has the classical Skellam distribution [12 ] . Using its twice-shifted moment measures, we obtain a general Hamburger lifting result with arbitrary initial moments and an independent Gaussian component.
Corollary 8.14 (Poisson differences with an arbitrary initial measure). Let A ( t ) = ∑ n ≥ 0 a n t n / n ! A(t)=\sum_{n\geq0}a_nt^n/n! A ( t ) = ∑ n ≥ 0 a n t n / n ! be any initial Hamburger moment series, and fix σ 2 ≥ 0 \sigma^{2}\geq0 σ 2 ≥ 0 . Define
∑ n ≥ 0 S n ( u , v , w ) t n n ! = A ( t ) exp ( u ( e t − 1 ) + v ( e − t − 1 ) + σ 2 w t 2 2 ) . (300) \sum_{n\geq0}S_n(u,v,w)\frac{t^n}{n!}=A(t)\exp\left(u(e^t-1)+v(e^{-t}-1)+\frac{\sigma^{2}wt^{2}}{2}\right). \tag*{(300)} n ≥ 0 ∑ S n ( u , v , w ) n ! t n = A ( t ) exp ( u ( e t − 1 ) + v ( e − t − 1 ) + 2 σ 2 w t 2 ) . (300)
For r ≥ 1 r \ge1 r ≥ 1 , the leading determinant det [ S i + j ( u , v , w ) ] i , j = 0 r − 1 \det[S_{i+j}(u,v,w)]_{i,j=0}^{r-1} det [ S i + j ( u , v , w ) ] i , j = 0 r − 1 belongs to R ≥ 0 [ u , v , w ] \mathbb{R}_{\ge0}[u,v,w] R ≥ 0 [ u , v , w ] , has degree at most D = ( r 2 ) D=\binom{r}{2} D = ( 2 r ) , and has homogeneous part of degree D D D equal to
a 0 r ( ∏ j = 0 r − 1 j ! ) ( u + v + σ 2 w ) D . (301) a_0^r\left(\prod_{j=0}^{r-1}j!\right)(u+v+\sigma^2w)^D. \tag*{(301)} a 0 r ( j = 0 ∏ r − 1 j ! ) ( u + v + σ 2 w ) D . (301)
Any additional common polynomial drift leaves these determinants unchanged.
Proof. The twice-shifted cumulants of e t − 1 e^t-1 e t − 1 , e − t − 1 e^{-t}-1 e − t − 1 , and σ 2 t 2 / 2 \sigma^2t^2/2 σ 2 t 2 /2 are represented by δ 1 \delta_1 δ 1 , δ − 1 \delta_{-1} δ − 1 , and σ 2 δ 0 \sigma^2\delta_0 σ 2 δ 0 , respectively. Applying Theorem 1.8 to these exponents, we obtain coefficientwise positivity, the degree bound, and (8.51), using second cumulants 1 , 1 , σ 2 1,1,\sigma^2 1 , 1 , σ 2 . Applying Remark 1 to the output family, we obtain drift invariance.
For A = 1 A=1 A = 1 and numerical nonnegative parameters, we obtain the moment sequence of N u − N v + σ w Z N_u-N_v+\sigma\sqrt{w}Z N u − N v + σ w Z , where σ = σ 2 \sigma=\sqrt{\sigma^2} σ = σ 2 , N u ∼ Poisson ( u ) N_u\sim\operatorname{Poisson}(u) N u ∼ Poisson ( u ) , N v ∼ Poisson ( v ) N_v\sim\operatorname{Poisson}(v) N v ∼ Poisson ( v ) , and Z Z Z is standard normal. We take these three variables to be independent. To see the positivity explicitly, put V = u + v + σ 2 w V=u+v+\sigma^2w V = u + v + σ 2 w . The first two nontrivial leading determinants are
det [ S i + j ] i , j = 0 1 = V , (302) \det[S_{i+j}]_{i,j=0}^{1}=V, \tag*{(302)} det [ S i + j ] i , j = 0 1 = V , (302)
det [ S i + j ] i , j = 0 2 = 2 V 3 + 4 u v + σ 2 w ( u + v ) . (303) \det[S_{i+j}]_{i,j=0}^{2}=2V^3+4uv+\sigma^2w(u+v). \tag*{(303)} det [ S i + j ] i , j = 0 2 = 2 V 3 + 4 uv + σ 2 w ( u + v ) . (303)
Writing κ j \kappa_j κ j for the j j j -th cumulant of the pure family in (8.50), these determinants are κ 2 \kappa_2 κ 2 and 2 κ 2 3 + κ 2 κ 4 − κ 3 2 2\kappa_2^3+\kappa_2\kappa_4-\kappa_3^2 2 κ 2 3 + κ 2 κ 4 − κ 3 2 . Here κ 2 = V \kappa_2=V κ 2 = V , κ 3 = u − v \kappa_3=u-v κ 3 = u − v , and κ 4 = u + v \kappa_4=u+v κ 4 = u + v . After cancelling the signed odd-cumulant contribution, we obtain exactly (8.53).
The coefficientwise assertion concerns leading determinants, as in Theorem 1.8 .
Further consequences for lifted Hankel matrices
We conclude the applications with three standard consequences of the coefficientwise Hankel total positivity proved above. For the first, we use the classical Sylvester identity [20 Theorem 1.1] . For the second, we use the standard bidiagonal and Pascal factorizations; see [32 ] and Layman’s binomial-invariance formula [23 Theorem 1] . For the third, we use a bidiagonal factorization of the moment matrices appearing in classical Stieltjes–Padé approximation [37 ] . We include the deductions to explain how we use these classical tools for the output sequences of the lifting theorem.
Corollary 8.15 (Standard consequences for the lifted sequences). Let ( F n ( y ) ) n ≥ 0 (F_n(y))_{n\ge0} ( F n ( y ) ) n ≥ 0 be an output of Theorem 1.7 under its equivalent conditions. Then the following assertions hold.
(i) The centered kernel
K i , j ( y ) : = F 0 ( y ) F i + j ( y ) − F i ( y ) F j ( y ) ( i , j ≥ 1 ) (304) K_{i,j}(y):=F_0(y)F_{i+j}(y)-F_i(y)F_j(y)\qquad(i,j\ge1) \tag*{(304)} K i , j ( y ) := F 0 ( y ) F i + j ( y ) − F i ( y ) F j ( y ) ( i , j ≥ 1 ) (304)
is coefficientwise totally positive in y \mathbf{y} y . For increasing positive index lists I I I , J J J of length r r r , we have
det K [ I , J ] = F 0 r − 1 det [ F u + v ] u ∈ { 0 } ∪ I , v ∈ { 0 } ∪ J . (305) \det K[I,J] = F_0^{r-1}\det[F_{u+v}]_{u\in\{0\}\cup I,\,v\in\{0\}\cup J}. \tag*{(305)} det K [ I , J ] = F 0 r − 1 det [ F u + v ] u ∈ { 0 } ∪ I , v ∈ { 0 } ∪ J . (305)
(ii) Fix h ≥ 0 h\geq0 h ≥ 0 and introduce independent indeterminates c 1 , … , c h , u , v c_1,\ldots,c_h,u,v c 1 , … , c h , u , v . Define
P ( z ) : = ∏ ℓ = 1 h ( z + c ℓ ) = ∑ ℓ = 0 h p ℓ z ℓ , F ~ n : = ∑ ℓ = 0 h p ℓ F n + ℓ , (306) P(z):=\prod_{\ell=1}^{h}(z+c_\ell)=\sum_{\ell=0}^{h}p_\ell z^\ell,\qquad\widetilde{F}_n:=\sum_{\ell=0}^{h}p_\ell F_{n+\ell}, \tag*{(306)} P ( z ) := ℓ = 1 ∏ h ( z + c ℓ ) = ℓ = 0 ∑ h p ℓ z ℓ , F n := ℓ = 0 ∑ h p ℓ F n + ℓ , (306)
and
G n : = ∑ k = 0 n ( n k ) u n − k v k F ~ k . (307) G_n:=\sum_{k=0}^{n}\binom{n}{k}u^{n-k}v^k\widetilde{F}_k. \tag*{(307)} G n := k = 0 ∑ n ( k n ) u n − k v k F k . (307)
Both ( F ~ n ) (\widetilde{F}_n) ( F n ) and ( G n ) (G_n) ( G n ) are coefficientwise Hankel-totally positive jointly in the original and new parameters. Moreover,
det [ G i + j ] i , j = 0 r − 1 = v r ( r − 1 ) det [ F ~ i + j ] i , j = 0 r − 1 . (308) \det[G_{i+j}]_{i,j=0}^{r-1}=v^{r(r-1)}\det[\widetilde{F}_{i+j}]_{i,j=0}^{r-1}. \tag*{(308)} det [ G i + j ] i , j = 0 r − 1 = v r ( r − 1 ) det [ F i + j ] i , j = 0 r − 1 . (308)
(iii) For an independent indeterminate s s s , h ∈ { 0 , 1 } h\in\{0,1\} h ∈ { 0 , 1 } , and r ≥ 1 r\geq1 r ≥ 1 , the polynomial
det [ F i + j + h ( y ) + s F i + j + h + 1 ( y ) ] i , j = 0 r − 1 (309) \det[F_{i+j+h}(y)+sF_{i+j+h+1}(y)]_{i,j=0}^{r-1} \tag*{(309)} det [ F i + j + h ( y ) + s F i + j + h + 1 ( y ) ] i , j = 0 r − 1 (309)
has nonnegative coefficients jointly in s , y s,y s , y .
Proof. For (i), applying Sylvester’s identity to the Hankel matrix H = [ F i + j ] i , j ≥ 0 H=[F_{i+j}]_{i,j\geq0} H = [ F i + j ] i , j ≥ 0 , with the fixed row and column index both equal to zero, we obtain (8.55). Since this is a polynomial identity, we may also use it when F 0 = 0 F_0=0 F 0 = 0 . By Theorem 1.7, every factor on the right has nonnegative coefficients, and we obtain the required positivity.
For (ii), define lower bidiagonal matrices T ( c ) i , j = c 1 { i = j } + 1 { i = j + 1 } T(c)_{i,j}=c\mathbf{1}_{\{i=j\}}+\mathbf{1}_{\{i=j+1\}} T ( c ) i , j = c 1 { i = j } + 1 { i = j + 1 } . Writing H ( F ) = [ F i + j ] i , j ≥ 0 H(F)=[F_{i+j}]_{i,j\geq0} H ( F ) = [ F i + j ] i , j ≥ 0 , we have
H ( F ~ ) = H ( F ) ∏ ℓ = 1 h T ( c ℓ ) . (310) H(\widetilde{F})=H(F)\prod_{\ell=1}^{h}T(c_\ell). \tag*{(310)} H ( F ) = H ( F ) ℓ = 1 ∏ h T ( c ℓ ) . (310)
For every weight factor, we have coefficientwise total positivity. For the affine change, put B n , k = ( n k ) u n − k v k B_{n,k}=\binom{n}{k}u^{n-k}v^k B n , k = ( k n ) u n − k v k when k ≤ n k\leq n k ≤ n , and B n , k = 0 B_{n,k}=0 B n , k = 0 otherwise. By the binomial identity, we obtain
H ( G ) = B H ( F ~ ) B T . (311) H(G)=BH(\widetilde{F})B^{\mathsf{T}}. \tag*{(311)} H ( G ) = B H ( F ) B T . (311)
For positive numerical u , v u,v u , v , every minor of B B B is a minor of the Pascal matrix multiplied by the row and column factors u n u^n u n and ( v / u ) k (v/u)^k ( v / u ) k . For a nonzero minor, we can match row and column indices within the lower triangular support. We therefore obtain a factor u ∑ n − ∑ k u^{\sum n-\sum k} u ∑ n − ∑ k with nonnegative exponent. Thus the scaling gives a polynomial identity with nonnegative coefficients, valid also at u = 0 u=0 u = 0 or v = 0 v=0 v = 0 , and we conclude that B B B is coefficientwise totally positive. Applying the finite Cauchy–Binet formula to (310) and (311) , we obtain the asserted Hankel total positivity. For the leading r × r r \times r r × r block, we use the triangular matrix B B B , with diagonal 1 , v , … , v r − 1 1,v,\ldots,v^{r-1} 1 , v , … , v r − 1 . Taking determinants, we obtain (308) .
For (iii), let H ( h ) = [ F i + j + h ] i , j ≥ 0 H^{(h)}=[F_{i+j+h}]_{i,j\geq0} H ( h ) = [ F i + j + h ] i , j ≥ 0 , and let L L L be the down-shift matrix, L i , j = 1 { i = j + 1 } L_{i,j}=\mathbf{1}\{i=j+1\} L i , j = 1 { i = j + 1 } . The required lower bidiagonal matrix is B ( s ) : = I + s L B(s):=I+sL B ( s ) := I + s L , with diagonal entries one and subdiagonal entries s s s . Indeed,
[ H ( h ) B ( s ) ] i , j = F i + j + h + s F i + j + h + 1 . (312) [H^{(h)}B(s)]_{i,j}=F_{i+j+h}+sF_{i+j+h+1}. \tag*{(312)} [ H ( h ) B ( s ) ] i , j = F i + j + h + s F i + j + h + 1 . (312)
Thus (309) is the determinant of the leading r × r r \times r r × r block of H ( h ) ( I + s L ) H^{(h)}(I+sL) H ( h ) ( I + s L ) . Every minor of H ( h ) H^{(h)} H ( h ) is a minor of H ( F ) H(F) H ( F ) , so it has nonnegative coefficients by Theorem 1.7 .
The matrix I + s L I+sL I + s L is coefficientwise totally positive. For the first r r r columns, its nonzero entries are confined to rows 0 , … , r 0,\ldots,r 0 , … , r . Hence the Cauchy–Binet expansion for the leading r × r r \times r r × r block is a finite sum over r r r -element subsets of { 0 , … , r } \{0,\ldots,r\} { 0 , … , r } , and every term has nonnegative coefficients. This proves (iii), including the specialization s = 0 s=0 s = 0 .
When F 0 = 1 F_0=1 F 0 = 1 and the parameters have numerical nonnegative values, we may interpret (304) as a covariance kernel of powers of a nonnegative random variable. In (309) , we obtain coefficientwise positivity for the moment-matrix denominators used in classical Stieltjes–Padé bounds whenever those approximants are defined. To establish convergence of these bounds, one needs additional moment-determinacy hypotheses. Here we assert the displayed coefficientwise positivity.
Some final remarks
For colored functional digraphs, we keep the cycle weights q 1 , … , q m q_1,\ldots,q_m q 1 , … , q m fixed and nonnegative. In Theorem 1.5 , we obtain joint coefficientwise positivity in the component intensities y 1 , … , y m y_1,\ldots,y_m y 1 , … , y m . It remains to establish joint coefficientwise positivity in both q \mathbf{q} q and y \mathbf{y} y .
We likewise keep the real equation coefficients fixed in Theorems 6.4 and 6.7 . We retain all independent component indeterminates in their conclusions. To retain the equation parameters as well, we would need to extend the lifting argument beyond the real moment-sequence hypotheses used here.
For the Schläfli–Gessel–Seo family, we prove the shifted moment condition for every fixed a , b ≥ 0 a,b\geq0 a , b ≥ 0 in Lemma 8.4 . Applying lifting to those inputs, we obtain coefficientwise Hankel total positivity in x x x , arbitrary initial moments, and independent colors in Theorem 8.5 . We leave the joint x , a , b x,a,b x , a , b assertion of Conjecture 8.1 unresolved. Nonnegative minors for every numerical specialization of a , b a,b a , b do not imply nonnegative coefficients in those parameters.
Throughout the lifting arguments, we use formal power series and finite moments. In the lognormal example, we obtain positivity even though the moment generating series has zero radius of convergence. To deduce uniqueness of a representing measure, one would need additional hypotheses beyond the positivity established here.
For the bivariate forest model in Corollary 7.7 , we keep w ≥ − 1 w \ge-1 w ≥ − 1 fixed. The corresponding joint assertions in ( x , w ) (x,w) ( x , w ) and in the shifted variables ( x , w + 1 ) (x,w+1) ( x , w + 1 ) , discussed by Sokal in [39 Conjectures 6.11(b) and 6.12(b)] , are not proved by fixing that parameter. We use the same convention for structural parameters in the further tree component families. In the Laguerre application, we can express the shape parameter as the intensity of a logarithmic exponent and hence retain it as an indeterminate in Theorem 8.10 .