Introduction
Let X X X denote multiplication by the variable x x x and D = d / d x D=d/dx D = d / d x . The classical identity ( X D ) n = ∑ k = 1 n S ( n , k ) X k D k (XD)^n=\sum_{k=1}^{n}S(n,k)X^kD^k ( X D ) n = ∑ k = 1 n S ( n , k ) X k D k , with the Stirling numbers S ( n , k ) S(n,k) S ( n , k ) of the second kind, is the basic example of a normal-ordering formula; see Katriel [11 ] and the monograph [16 ] . Mansour, Schork, and Shattuck [17 ] introduced generalized Stirling numbers S s ; h ( n , k ) S_{s;h}(n,k) S s ; h ( n , k ) through the expansion
( v u ) n = ∑ k = 1 n S s ; h ( n , k ) v s ( n − k ) + k u k , u v − v u = h v s . (1) (vu)^n=\sum_{k=1}^{n}S_{s;h}(n,k)v^{s(n-k)+k}u^k,\qquad uv-vu=hv^s. \tag*{(1)} ( v u ) n = k = 1 ∑ n S s ; h ( n , k ) v s ( n − k ) + k u k , uv − v u = h v s . (1)
They did this first for s ∈ N 0 s\in\mathbb{N}_0 s ∈ N 0 and then, by a recurrence, for every real s s s . Realizations by differential operators with real exponents were studied in [19 Section 4] and by Maier [15 ] . For the Volterra operator ( V f ) ( x ) = ∫ 0 x f ( t ) d t (Vf)(x)=\int_0^x f(t)\,dt ( V f ) ( x ) = ∫ 0 x f ( t ) d t , Benmoussa [3 Theorem 1] expanded ( X V ) n (XV)^n ( X V ) n in the operators X k V 2 n − k X^kV^{2n-k} X k V 2 n − k ; the coefficients are those of the Bessel polynomials. Normal orderings of words in x x x and integrals of integer order were studied in [2 ] .
This paper is about Riemann–Liouville operators of any real order. The Riemann–Liouville integrals satisfy
X I α − I α X = α I α + 1 ( α > 0 ) . (2) XI^\alpha-I^\alpha X=\alpha I^{\alpha+1}\qquad(\alpha>0). \tag*{(2)} X I α − I α X = α I α + 1 ( α > 0 ) . (2)
Since I α + 1 = I α ( 1 + 1 / α ) I^{\alpha+1}=I^\alpha(1+1/\alpha) I α + 1 = I α ( 1 + 1/ α ) , this is the relation in (1.1) with u = X u=X u = X , v = I α v=I^\alpha v = I α , h = α h=\alpha h = α and s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α . If 1 / α 1/\alpha 1/ α is not an integer, then s s s is not an integer, and the single relation u v − v u = h v s uv-vu=hv^s uv − v u = h v s does not in general determine the normal ordering. In this case one needs the commutators of u u u with all powers v a v^a v a , a ∈ Z + Z s a\in\mathbb{Z}+\mathbb{Z}s a ∈ Z + Z s (Proposition 2.6 ). For Riemann–Liouville operators these powers are given by the semigroup { I r } \{I^r\} { I r } , v a = I α a v^a=I^{\alpha a} v a = I α a , and fractional derivatives are the powers of negative order. Our results are of two kinds.
Fractional operators.
(i) For every s ≠ 1 s\ne1 s = 1 , the assignment v a ↦ I a / ( s − 1 ) v^a\mapsto I^{a/(s-1)} v a ↦ I a / ( s − 1 ) , u ↦ h ( s − 1 ) X u\mapsto h(s-1)X u ↦ h ( s − 1 ) X is a faithful representation of the algebra A s , h A_{s,h} A s , h in which the powers v a v^a v a are defined. It acts on smooth functions that are flat at the origin, and integrals of negative order are Riemann–Liouville derivatives (Theorem 4.1 ). Under the Laplace transform it becomes the well-known differential realization (Remark 4.7).
(ii) For every α > 0 \alpha>0 α > 0 and n ≥ 1 n\ge1 n ≥ 1 , on locally integrable functions,
( X I α ) n = ∑ k = 1 n S 1 + 1 / α ; − α ( n , k ) X k I ( α + 1 ) n − k , (3) (XI^\alpha)^n=\sum_{k=1}^{n}S_{1+1/\alpha;-\alpha}(n,k)X^kI^{(\alpha+1)n-k}, \tag*{(3)} ( X I α ) n = k = 1 ∑ n S 1 + 1/ α ; − α ( n , k ) X k I ( α + 1 ) n − k , (3)
and ( I α X ) n (I^\alpha X)^n ( I α X ) n has the expansion in the opposite order with coefficients S 1 + 1 / α ; α ( n , k ) S_{1+1/\alpha;\alpha}(n,k) S 1 + 1/ α ; α ( n , k ) . For every β > 0 \beta>0 β > 0 , the corresponding expansions of ( X D β ) n (XD^\beta)^n ( X D β ) n and ( D β X ) n (D^\beta X)^n ( D β X ) n , with coefficients S 1 − 1 / β ; ± β ( n , k ) S_{1-1/\beta;\pm\beta}(n,k) S 1 − 1/ β ; ± β ( n , k ) , hold on the range of I n β I^{n\beta} I n β (Theorem 4.4 ). For derivatives, a condition at the origin is needed (Remark 4.5). The coefficients are explicit polynomials in α \alpha α (Proposition 5.1). Products I γ X I δ I^\gamma XI^\delta I γ X I δ lead to the full three-parameter family of Hsu and Shiue [8 ] (Proposition 5.3 ). The cases α = 1 \alpha=1 α = 1 , β = 1 \beta=1 β = 1 and β = 1 2 \beta=\frac12 β = 2 1 give Benmoussa’s expansion, Scherk’s identity and the Bessel numbers of the second kind (Corollary 5.2 ).
A probabilistic formula.
(iii) For α > 0 \alpha>0 α > 0 , the kernel of ( X I α ) n (XI^\alpha)^n ( X I α ) n is x n ( x − t ) n α − 1 P n , α ( 1 − t / x ) / Γ ( n α ) x^n(x-t)^{n\alpha-1}P_{n,\alpha}(1-t/x)/\Gamma(n\alpha) x n ( x − t ) n α − 1 P n , α ( 1 − t / x ) /Γ ( n α ) with P n , α ( y ) = E ∏ j = 1 n − 1 ( 1 − y B j ) P_{n,\alpha}(y)=\mathbb{E}\prod_{j=1}^{n-1}(1-yB_j) P n , α ( y ) = E ∏ j = 1 n − 1 ( 1 − y B j ) . Here the B j B_j B j are the partial sums of a symmetric Dirichlet vector with parameter α \alpha α (Theorem 6.1 ). As a consequence, for every real s > 1 s>1 s > 1 ,
S s ; 1 ( n , n − m ) = E e m ( B 1 , … , B n − 1 ) ∏ i = 0 m − 1 ( n + i ( s − 1 ) ) , (4) S_{s;1}(n,n-m)=\mathbb{E}e_m(B_1,\ldots,B_{n-1})\prod_{i=0}^{m-1}(n+i(s-1)), \tag*{(4)} S s ; 1 ( n , n − m ) = E e m ( B 1 , … , B n − 1 ) i = 0 ∏ m − 1 ( n + i ( s − 1 )) , (4)
where e m e_m e m is the elementary symmetric polynomial and the Dirichlet parameter is 1 / ( s − 1 ) 1/(s-1) 1/ ( s − 1 ) (Corollary 6.2 ). The reversal symmetry of the Dirichlet distribution corresponds to switching between the two normal orders.
(iv) By de Finetti’s theorem, for s > 1 s>1 s > 1 (fractional integrals), (1.4) is a statement about a Pólya urn with n n n colours, in which each drawn colour becomes more likely to be drawn again. The urn model also works for s = 1 s=1 s = 1 , where the draws are independent, and for s = 1 − 1 / K s=1-1/K s = 1 − 1/ K with K ∈ N K\in\mathbb{N} K ∈ N (derivatives of order K K K ), where the balls are drawn without replacement from an urn with K K K balls of each colour (Proposition 6.3 ). The urn probabilities are the terms of a product formula for S s ; 1 ( n , n − m ) S_{s;1}(n,n-m) S s ; 1 ( n , n − m ) . This formula follows from the recurrence and holds for every s s s . At s = 1 s=1 s = 1 , s = 0 s=0 s = 0 and s = 2 s=2 s = 2 it gives the classical formulas for the Stirling numbers of the first and second kinds and for the Bessel numbers (Corollary 6.4). For s ≥ 1 s\ge1 s ≥ 1 the same numbers also describe the number of blocks of the Ewens–Pitman partition, which is a different urn model (Remark 6.5).
(v) The same method gives the kernels of all words X λ n I α n ⋯ X λ 1 I α 1 X^{\lambda_n}I^{\alpha_n}\cdots X^{\lambda_1}I^{\alpha_1} X λ n I α n ⋯ X λ 1 I α 1 with α i > 0 \alpha_i>0 α i > 0 and λ i ∈ N 0 \lambda_i\in\mathbb{N}_0 λ i ∈ N 0 (Theorem 6.7 ). They are expectations over Dirichlet distributions that are not symmetric, the coefficients of the normal ordering are explicit with alternating signs, and the urn model extends to these words. For ( X λ I α ) n (X^\lambda I^\alpha)^n ( X λ I α ) n this extends a recurrence of Benmoussa [2 ] from integer to real orders (Corollary 6.8).
Butzer, Kilbas, and Trujillo [5 ] use fractional calculus to extend the Stirling numbers of the second kind to non-integer values of the first index; this is a different kind of extension. Up to a power of x x x , ( X I α ) n (XI^\alpha)^n ( X I α ) n is a multiple Erdélyi–Kober operator in the sense of Kiryakova [14 ] . Such operators have Meijer G G G -function kernels in general, but by (3) the kernel is elementary here. Under the Laplace transform, the identities of (ii) correspond to identities for differential operators that are contained in [15 ] (Remark 4.7). We do not know of earlier normal-ordering formulas for Riemann–Liouville operators of non-integer order, or of a description of the kernels of ( X I α ) n (XI^\alpha)^n ( X I α ) n and of other words in X X X and fractional integrals by Dirichlet distributions or by Pólya urns.
Throughout, algebras are unital associative C \mathbb{C} C -algebras, homomorphisms preserve the identity, h ∈ C ∖ { 0 } h\in\mathbb{C}\setminus\{0\} h ∈ C ∖ { 0 } , and [ a , b ] = a b − b a [a,b]=ab-ba [ a , b ] = ab − ba . We write N 0 = { 0 , 1 , 2 , … } \mathbb{N}_0=\{0,1,2,\ldots\} N 0 = { 0 , 1 , 2 , … } , N = N 0 ∖ { 0 } \mathbb{N}=\mathbb{N}_0\setminus\{0\} N = N 0 ∖ { 0 } , and ( a ) m = a ( a + 1 ) ⋯ ( a + m − 1 ) (a)_m=a(a+1)\cdots(a+m-1) ( a ) m = a ( a + 1 ) ⋯ ( a + m − 1 ) . All sums are finite, and empty products equal 1.
Generalized Stirling numbers and compatible powers
For s ∈ R s\in\mathbb{R} s ∈ R , define S s ; h ( 0 , 0 ) = 1 S_{s;h}(0,0)=1 S s ; h ( 0 , 0 ) = 1 , S s ; h ( n , 0 ) = 0 S_{s;h}(n,0)=0 S s ; h ( n , 0 ) = 0 for n ≥ 1 n\geq1 n ≥ 1 , S s ; h ( n , k ) = 0 S_{s;h}(n,k)=0 S s ; h ( n , k ) = 0 for k < 0 k<0 k < 0 or k > n k>n k > n , and
S s ; h ( n + 1 , k ) = S s ; h ( n , k − 1 ) + h ( k + s ( n − k ) ) S s ; h ( n , k ) ( n ≥ 0 , k ≥ 1 ) . (5) S_{s;h}(n+1,k)=S_{s;h}(n,k-1)+h\left(k+s(n-k)\right)S_{s;h}(n,k)\qquad(n\geq0,\ k\geq1). \tag*{(5)} S s ; h ( n + 1 , k ) = S s ; h ( n , k − 1 ) + h ( k + s ( n − k ) ) S s ; h ( n , k ) ( n ≥ 0 , k ≥ 1 ) . (5)
as in [17 Definition 3.1] . Induction gives
S s ; h ( n , k ) = h n − k S s ; 1 ( n , k ) , S s ; − h ( n , k ) = ( − 1 ) n − k S s ; h ( n , k ) . (6) S_{s;h}(n,k)=h^{n-k}S_{s;1}(n,k),\qquad S_{s;-h}(n,k)=(-1)^{n-k}S_{s;h}(n,k). \tag*{(6)} S s ; h ( n , k ) = h n − k S s ; 1 ( n , k ) , S s ; − h ( n , k ) = ( − 1 ) n − k S s ; h ( n , k ) . (6)
For example, S s ; h ( n , n ) = 1 S_{s;h}(n,n)=1 S s ; h ( n , n ) = 1 , S s ; h ( 2 , 1 ) = h S_{s;h}(2,1)=h S s ; h ( 2 , 1 ) = h , S s ; h ( 3 , 2 ) = 3 h S_{s;h}(3,2)=3h S s ; h ( 3 , 2 ) = 3 h , and S s ; h ( 3 , 1 ) = h 2 ( s + 1 ) S_{s;h}(3,1)=h^2(s+1) S s ; h ( 3 , 1 ) = h 2 ( s + 1 ) . The numbers S s ; h ( n , k ) S_{s;h}(n,k) S s ; h ( n , k ) are connection coefficients, as the next lemma shows.
Lemma 2.1. For s ∈ R s\in\mathbb{R} s ∈ R and n ∈ N 0 n\in\mathbb{N}_0 n ∈ N 0 , as polynomials in z z z ,
∏ i = 0 n − 1 ( z + h s i ) = ∑ k = 0 n S s ; h ( n , k ) ∏ j = 0 k − 1 ( z − h ( 1 − s ) j ) . (7) \prod_{i=0}^{n-1}(z+hsi)=\sum_{k=0}^{n}S_{s;h}(n,k)\prod_{j=0}^{k-1}(z-h(1-s)j). \tag*{(7)} i = 0 ∏ n − 1 ( z + h s i ) = k = 0 ∑ n S s ; h ( n , k ) j = 0 ∏ k − 1 ( z − h ( 1 − s ) j ) . (7)
Proof. The case n = 0 n=0 n = 0 is clear. Multiply (7) by z + h s n = ( z − h ( 1 − s ) k ) + h ( k + s ( n − k ) ) z+hsn=\left(z-h(1-s)k\right)+h\left(k+s(n-k)\right) z + h s n = ( z − h ( 1 − s ) k ) + h ( k + s ( n − k ) ) and compare with (5) . □
Hsu and Shiue [8 ] defined numbers S ( n , k ; a , b , r ) S(n,k;a,b,r) S ( n , k ; a , b , r ) by
( z ∣ a ) n = ∑ k = 0 n S ( n , k ; a , b , r ) ( z − r ∣ b ) k , ( z ∣ a ) n = ∏ j = 0 n − 1 ( z − j a ) . (8) (z\mid a)_n=\sum_{k=0}^{n}S(n,k;a,b,r)(z-r\mid b)_k,\qquad(z\mid a)_n=\prod_{j=0}^{n-1}(z-ja). \tag*{(8)} ( z ∣ a ) n = k = 0 ∑ n S ( n , k ; a , b , r ) ( z − r ∣ b ) k , ( z ∣ a ) n = j = 0 ∏ n − 1 ( z − j a ) . (8)
Thus (7) says that S s ; h ( n , k ) = S ( n , k ; − h s , h ( 1 − s ) , 0 ) S_{s;h}(n,k)=S(n,k;-hs,h(1-s),0) S s ; h ( n , k ) = S ( n , k ; − h s , h ( 1 − s ) , 0 ) ; this was observed in [19 Theorem 2] .
The algebra A(s,h)
Let Γ s = Z + Z s ⊂ R \Gamma_s=\mathbb{Z}+\mathbb{Z}s\subset\mathbb{R} Γ s = Z + Z s ⊂ R and R s = C [ Γ s ] R_s=\mathbb{C}[\Gamma_s] R s = C [ Γ s ] , where we view Γ s \Gamma_s Γ s as an additive subgroup of R \mathbb{R} R . Thus R s R_s R s has basis { v a : a ∈ Γ s } \{v^a:a\in\Gamma_s\} { v a : a ∈ Γ s } and
v a v b = v a + b , v 0 = 1. (9) v^av^b=v^{a+b},\qquad v^0=1. \tag*{(9)} v a v b = v a + b , v 0 = 1. (9)
If s = p / q s=p/q s = p / q in lowest terms, with q > 0 q>0 q > 0 , then Γ s = q − 1 Z \Gamma_s=q^{-1}\mathbb{Z} Γ s = q − 1 Z and R s = C [ t , t − 1 ] R_s=\mathbb{C}[t,t^{-1}] R s = C [ t , t − 1 ] with t = v 1 / q t=v^{1/q} t = v 1/ q . If s s s is irrational, then R s = C [ x ± 1 , y ± 1 ] R_s=\mathbb{C}[x^{\pm1},y^{\pm1}] R s = C [ x ± 1 , y ± 1 ] with x = v x=v x = v and y = v s y=v^s y = v s . The linear map δ s , h ( v a ) = h a v a + s − 1 \delta_{s,h}(v^a)=ha\,v^{a+s-1} δ s , h ( v a ) = ha v a + s − 1 is a derivation of R s R_s R s , and we consider the differential Ore extension
A s , h = R s [ u ; δ s , h ] , u r = r u + δ s , h ( r ) ( r ∈ R s ) ; (10) A_{s,h}=R_s[u;\delta_{s,h}],\qquad ur=ru+\delta_{s,h}(r)\quad(r\in R_s); \tag*{(10)} A s , h = R s [ u ; δ s , h ] , u r = r u + δ s , h ( r ) ( r ∈ R s ) ; (10)
see [22 , 7 ] . Equivalently, in addition to (2.5), its generators satisfy
[ u , v a ] = h a v a + s − 1 ( a ∈ Γ s ) . (11) [u,v^a]=ha\,v^{a+s-1}\qquad(a\in\Gamma_s). \tag*{(11)} [ u , v a ] = ha v a + s − 1 ( a ∈ Γ s ) . (11)
By the universal property of Ore extensions, any elements of another algebra that satisfy (2.5) and (2.7) determine a unique homomorphism from A s , h A_{s,h} A s , h .
Lemma 2.2. Both { v a u m : a ∈ Γ s , m ∈ N 0 } \{v^a u^m:a\in\Gamma_s,\ m\in\mathbb{N}_0\} { v a u m : a ∈ Γ s , m ∈ N 0 } and { u m v a : a ∈ Γ s , m ∈ N 0 } \{u^m v^a:a\in\Gamma_s,\ m\in\mathbb{N}_0\} { u m v a : a ∈ Γ s , m ∈ N 0 } are C \mathbb{C} C -bases of A s , h A_{s,h} A s , h .
Proof. The first statement is part of the construction of the Ore extension. With δ = δ s , h \delta=\delta_{s,h} δ = δ s , h , induction gives u m r = ∑ j = 0 m ( m j ) δ j ( r ) u m − j u^m r=\sum_{j=0}^m\binom{m}{j}\delta^j(r)u^{m-j} u m r = ∑ j = 0 m ( j m ) δ j ( r ) u m − j and r u m = ∑ j = 0 m ( − 1 ) j ( m j ) u m − j δ j ( r ) ru^m=\sum_{j=0}^m(-1)^j\binom{m}{j}u^{m-j}\delta^j(r) r u m = ∑ j = 0 m ( − 1 ) j ( j m ) u m − j δ j ( r ) for r ∈ R s r\in R_s r ∈ R s . The second identity shows that the elements u m v a u^m v^a u m v a span the algebra. If ∑ m = 0 M u m r m = 0 \sum_{m=0}^M u^m r_m=0 ∑ m = 0 M u m r m = 0 with r M ≠ 0 r_M\ne0 r M = 0 , the first identity shows that the coefficient of u M u^M u M in the first normal form is r M r_M r M , a contradiction. □
Theorem 2.3. Let s ∈ R s\in\mathbb{R} s ∈ R and n ≥ 1 n\ge1 n ≥ 1 . In A s , h A_{s,h} A s , h ,
( v u ) n = ∑ k = 1 n S s , h ( n , k ) v s ( n − k ) + k u k , (12) (vu)^n=\sum_{k=1}^n S_{s,h}(n,k)v^{s(n-k)+k}u^k, \tag*{(12)} ( v u ) n = k = 1 ∑ n S s , h ( n , k ) v s ( n − k ) + k u k , (12) ( u v ) n = ∑ k = 1 n S s , − h ( n , k ) u k v s ( n − k ) + k = ∑ k = 1 n ( − 1 ) n − k S s , h ( n , k ) u k v s ( n − k ) + k . (13) (uv)^n=\sum_{k=1}^n S_{s,-h}(n,k)u^k v^{s(n-k)+k}
=\sum_{k=1}^n(-1)^{n-k}S_{s,h}(n,k)u^k v^{s(n-k)+k}. \tag*{(13)} ( uv ) n = k = 1 ∑ n S s , − h ( n , k ) u k v s ( n − k ) + k = k = 1 ∑ n ( − 1 ) n − k S s , h ( n , k ) u k v s ( n − k ) + k . (13)
Proof. Set e n , k = s ( n − k ) + k e_{n,k}=s(n-k)+k e n , k = s ( n − k ) + k . By (2.7), v u v e n , k u k = v e n , k + 1 u k + 1 + h e n , k v e n , k + s u k vu\,v^{e_{n,k}}u^k=v^{e_{n,k}+1}u^{k+1}+he_{n,k}v^{e_{n,k}+s}u^k v u v e n , k u k = v e n , k + 1 u k + 1 + h e n , k v e n , k + s u k . Since e n , k + 1 = e n + 1 , k + 1 e_{n,k}+1=e_{n+1,k+1} e n , k + 1 = e n + 1 , k + 1 and e n , k + s = e n + 1 , k e_{n,k}+s=e_{n+1,k} e n , k + s = e n + 1 , k , the coefficients satisfy (2.1), and (2.8) follows by induction. In the opposite algebra, with product a ∘ b = b a a\circ b=ba a ∘ b = ba , the powers satisfy (2.5) and u ∘ v a − v a ∘ u = − h a v a + s − 1 u\circ v^a-v^a\circ u=-ha\,v^{a+s-1} u ∘ v a − v a ∘ u = − ha v a + s − 1 . So the identity map on the generators gives a homomorphism A s , − h → A s , h o p A_{s,-h}\to A_{s,h}^{\mathrm{op}} A s , − h → A s , h op , and this homomorphism is bijective by Lemma 2.2. Applying (2.8) in A s , − h A_{s,-h} A s , − h gives (2.9), and (2.2) gives its second form. □
Remark 2.4. For the differential operators u = x s d / d x u=x^s\,d/dx u = x s d / d x and v = x v=x v = x with complex s s s , identities equivalent to (2.8) and (2.9) are special cases of the expansions of Maier [15 Theorem 6.4] . For positive integer s s s , the expansion (2.9) was obtained by Burde [4 Proposition 5.3] and in Benmoussa’s preprint [1 Theorem 4.1] . The simplest case, ( u v ) 2 = u ( u v − h v s ) v = u 2 v 2 − h u v s + 1 (uv)^2=u(uv-hv^s)v=u^2v^2-huv^{s+1} ( uv ) 2 = u ( uv − h v s ) v = u 2 v 2 − h u v s + 1 , shows that the coefficients in [3 Conjecture 2] need − h -h − h in place of h h h .
When are compatible powers needed?
For integer s s s the relation [ u , v ] = h v s [u,v]=hv^s [ u , v ] = h v s is enough.
Corollary 2.5. Let u , v u,v u , v belong to an algebra and satisfy [ u , v ] = h v s [u,v]=hv^s [ u , v ] = h v s . If s ∈ N 0 s\in\mathbb{N}_0 s ∈ N 0 , the identities (2.8) and (2.9) hold without any invertibility assumption. They also hold for negative integer s s s if v v v is invertible.
Proof. For a ∈ N a\in\mathbb{N} a ∈ N , the Leibniz rule gives [ u , v a ] = ∑ j = 0 a − 1 v j [ u , v ] v a − 1 − j = h a v a + s − 1 [u,v^a]=\sum_{j=0}^{a-1}v^j[u,v]v^{a-1-j}=ha\,v^{a+s-1} [ u , v a ] = ∑ j = 0 a − 1 v j [ u , v ] v a − 1 − j = ha v a + s − 1 , and, if v v v is invertible, [ u , v − 1 ] = − v − 1 [ u , v ] v − 1 = − h v s − 2 [u,v^{-1}]=-v^{-1}[u,v]v^{-1}=-hv^{s-2} [ u , v − 1 ] = − v − 1 [ u , v ] v − 1 = − h v s − 2 . The proof of Theorem 2.3 uses only these commutators, because all exponents e n , k e_{n,k} e n , k are integers, and they are positive if s ∈ N 0 s\in\mathbb{N}_0 s ∈ N 0 . □
For non-integer s s s this is no longer true.
Proposition 2.6. Let an algebra B B B contain R s R_s R s as a unital subalgebra, and let u ∈ B u \in B u ∈ B satisfy [ u , v ] = h v s [u,v] = hv^s [ u , v ] = h v s .
If s = p / q s = p/q s = p / q in lowest terms, with q > 0 q > 0 q > 0 , then (2.7) holds for every a ∈ Γ s a \in\Gamma_s a ∈ Γ s if and only if [ u , v 1 / q ] [u,v^{1/q}] [ u , v 1/ q ] commutes with v 1 / q v^{1/q} v 1/ q . For s = 1 / 2 s = 1/2 s = 1/2 , for example, this condition can fail, and then (2.9) can fail too.
If s s s is irrational, then (2.7) holds for every a ∈ Γ s a \in\Gamma_s a ∈ Γ s if and only if
[ u , v s ] = h s v 2 s − 1 . (14) [u,v^s] = hs\,v^{2s-1}. \tag*{(14)} [ u , v s ] = h s v 2 s − 1 . (14) The assumption that all [ u , v a ] [u,v^a] [ u , v a ] , a ∈ Γ s a \in\Gamma_s a ∈ Γ s , commute with R s R_s R s is not enough, and in that case (2.9) can fail.
Proof. (1) Put t = v 1 / q t = v^{1/q} t = v 1/ q and c = [ u , t ] c = [u,t] c = [ u , t ] . If c c c commutes with t t t , then h t p = [ u , t q ] = q t q − 1 c ht^p = [u,t^q] = qt^{q-1}c h t p = [ u , t q ] = q t q − 1 c , so c = h q t p − q + 1 c = \frac{h}{q}t^{p-q+1} c = q h t p − q + 1 , and the Leibniz rule gives [ u , t j ] = ( h j / q ) t j + p − q [u,t^j] = (hj/q)t^{j+p-q} [ u , t j ] = ( hj / q ) t j + p − q for all j ∈ Z j \in\mathbb{Z} j ∈ Z ; this is (2.7) for a = j / q a = j/q a = j / q . Conversely, (2.7) with a = 1 / q a = 1/q a = 1/ q gives this value of c c c . For the example, let s = 1 / 2 s = 1/2 s = 1/2 , so that R s = C [ t , t − 1 ] R_s = \mathbb{C}[t,t^{-1}] R s = C [ t , t − 1 ] , and embed R s R_s R s in End C ( R s ) \operatorname{End}_{\mathbb{C}}(R_s) End C ( R s ) by multiplication operators. Let ( Q f ) ( t ) = f ( − t ) (Qf)(t) = f(-t) ( Q f ) ( t ) = f ( − t ) , so that Q t = − t Q Qt = -tQ Qt = − tQ , and put u = h 2 d d t + κ Q ∘ 1 2 t u = \frac{h}{2}\frac{d}{dt} + \kappa Q \circ\frac{1}{2t} u = 2 h d t d + κ Q ∘ 2 t 1 with κ ≠ 0 \kappa\ne0 κ = 0 . Then c = [ u , t ] = h 2 + κ Q c = [u,t] = \frac{h}{2} + \kappa Q c = [ u , t ] = 2 h + κ Q does not commute with t t t , while [ u , v ] = c t + t c = h t = h v s [u,v] = ct + tc = ht = hv^s [ u , v ] = c t + t c = h t = h v s . For n = 3 n = 3 n = 3 , the left side of (2.9) minus the right side maps 1 1 1 to 1 2 h κ ( 4 h − κ ) t 3 \frac{1}{2}h\kappa(4h-\kappa)t^3 2 1 hκ ( 4 h − κ ) t 3 and t t t to − 1 2 h κ ( 5 h + κ ) t 4 -\frac{1}{2}h\kappa(5h+\kappa)t^4 − 2 1 hκ ( 5 h + κ ) t 4 , and these are not both zero. (2) Since R s R_s R s is generated by v ± 1 v^{\pm1} v ± 1 and v ± s v^{\pm s} v ± s , the two derivations [ u , ⋅ ] [u,\cdot] [ u , ⋅ ] and δ s , h \delta_{s,h} δ s , h agree on R s R_s R s once they agree on v v v and v s v^s v s . For the last statement, embed R s = C [ x ± 1 , y ± 1 ] R_s = \mathbb{C}[x^{\pm1},y^{\pm1}] R s = C [ x ± 1 , y ± 1 ] in End C ( R s ) \operatorname{End}_{\mathbb{C}}(R_s) End C ( R s ) by multiplication operators and take u = h y ∂ / ∂ x u = hy \partial/\partial x u = h y ∂ / ∂ x . Then [ u , v m + n s ] = h m v m + n s + s − 1 [u,v^{m+ns}] = hm\,v^{m+ns+s-1} [ u , v m + n s ] = hm v m + n s + s − 1 for m , n ∈ Z m,n \in\mathbb{Z} m , n ∈ Z : these commutators lie in R s R_s R s , [ u , v ] = h v s [u,v] = hv^s [ u , v ] = h v s , but (2.7) fails for n ≠ 0 n \ne0 n = 0 . For n = 3 n = 3 n = 3 , the left side of (2.9) maps 1 1 1 to h 3 y 3 h^3y^3 h 3 y 3 , while the right side, u 3 v 3 − 3 h u 2 v s + 2 + h 2 ( s + 1 ) u v 2 s + 1 u^3v^3 - 3hu^2v^{s+2} + h^2(s+1)uv^{2s+1} u 3 v 3 − 3 h u 2 v s + 2 + h 2 ( s + 1 ) u v 2 s + 1 , maps 1 1 1 to 6 h 3 y 3 − 6 h 3 y 3 + ( s + 1 ) h 3 y 3 = ( s + 1 ) h 3 y 3 6h^3y^3 - 6h^3y^3 + (s+1)h^3y^3 = (s+1)h^3y^3 6 h 3 y 3 − 6 h 3 y 3 + ( s + 1 ) h 3 y 3 = ( s + 1 ) h 3 y 3 . □ \square □
In the next sections, fractional calculus gives the compatible powers.
Riemann–Liouville operators
Let L l o c 1 L^1_{\mathrm{loc}} L loc 1 be the space of measurable functions f : ( 0 , ∞ ) → C f : (0,\infty) \to\mathbb{C} f : ( 0 , ∞ ) → C that are integrable on ( 0 , T ) (0,T) ( 0 , T ) for every T > 0 T > 0 T > 0 ; we identify functions that are equal almost everywhere. Let X X X denote multiplication by x x x . For r > 0 r > 0 r > 0 , the Riemann–Liouville integral of order r r r is
( I r f ) ( x ) = 1 Γ ( r ) ∫ 0 x ( x − t ) r − 1 f ( t ) d t , (15) (I^r f)(x) = \frac{1}{\Gamma(r)}\int_0^x (x-t)^{r-1}f(t)\,\mathrm{d}t, \tag*{(15)} ( I r f ) ( x ) = Γ ( r ) 1 ∫ 0 x ( x − t ) r − 1 f ( t ) d t , (15)
and I 0 = Id I^0 = \operatorname{Id} I 0 = Id ; see [13 , 24 ] . Both X X X and I r I^r I r map L l o c 1 L^1_{\mathrm{loc}} L loc 1 into itself. For β > 0 \beta> 0 β > 0 and m = ⌈ β ⌉ m = \lceil\beta\rceil m = ⌈ β ⌉ , the Riemann–Liouville derivative is D β f = ( d / d x ) m I m − β f D^\beta f = (\mathrm{d}/\mathrm{d}x)^m I^{m-\beta}f D β f = ( d / d x ) m I m − β f , defined when I m − β f I^{m-\beta}f I m − β f has a locally absolutely continuous derivative of order m − 1 m-1 m − 1 on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Explicitly, the domain Dom ( D β ) \operatorname{Dom}(D^\beta) Dom ( D β ) consists of all f ∈ L l o c 1 f \in L^1_{\mathrm{loc}} f ∈ L loc 1 such that
I m − β f = p + I m ψ (16) I^{m-\beta}f = p + I^m\psi \tag*{(16)} I m − β f = p + I m ψ (16)
for a polynomial p p p of degree less than m m m and some ψ ∈ L l o c 1 \psi\in L^1_{\mathrm{loc}} ψ ∈ L loc 1 , and then D β f = ψ D^\beta f = \psi D β f = ψ . For β ∈ N \beta\in\mathbb{N} β ∈ N this is the ordinary derivative. We also write D − r = I r D^{-r} = I^r D − r = I r for r ≥ 0 r \ge0 r ≥ 0 .
Lemma 3.1. Let r , q ≥ 0 r,q \ge0 r , q ≥ 0 and γ ≥ β ≥ 0 \gamma\ge\beta\ge0 γ ≥ β ≥ 0 . On L l o c 1 L^1_{\mathrm{loc}} L loc 1 ,
(a) I r I q = I r + q I^r I^q = I^{r+q} I r I q = I r + q , and I r I^r I r is injective;
(b) X I r = I r X + r I r + 1 XI^r = I^rX + rI^{r+1} X I r = I r X + r I r + 1 ;
(c) I γ φ ∈ Dom ( D β ) I^\gamma\varphi\in\operatorname{Dom}(D^\beta) I γ φ ∈ Dom ( D β ) and D β I γ φ = I γ − β φ D^\beta I^\gamma\varphi= I^{\gamma-\beta}\varphi D β I γ φ = I γ − β φ for every φ ∈ L l o c 1 \varphi\in L^1_{\mathrm{loc}} φ ∈ L loc 1 .
Proof. (a) The semigroup law follows from Fubini’s theorem and the beta integral [13 Section 2.1] . If I r φ = 0 I^r\varphi= 0 I r φ = 0 , choose an integer m ≥ r m \ge r m ≥ r ; then I m φ = I m − r I r φ = 0 I^m\varphi= I^{m-r}I^r\varphi= 0 I m φ = I m − r I r φ = 0 , and differentiating m m m times gives φ = 0 \varphi=0 φ = 0 . (b) For r > 0 r>0 r > 0 , subtracting the integrals gives
( X I r f − I r X f ) ( x ) = 1 Γ ( r ) ∫ 0 x ( x − t ) r f ( t ) d t = r ( I r + 1 f ) ( x ) , (XI^r f-I^rXf)(x)=\frac{1}{\Gamma(r)}\int_0^x(x-t)^r f(t)\,\mathrm{d}t=r(I^{r+1}f)(x), ( X I r f − I r X f ) ( x ) = Γ ( r ) 1 ∫ 0 x ( x − t ) r f ( t ) d t = r ( I r + 1 f ) ( x ) ,
and for r = 0 r=0 r = 0 both sides are zero. (c) With m = ⌈ β ⌉ m=\lceil\beta\rceil m = ⌈ β ⌉ , part (a) gives I m − β I γ φ = I m ψ I^{m-\beta}I^\gamma\varphi=I^m\psi I m − β I γ φ = I m ψ , where ψ = I γ − β φ ∈ L l o c 1 \psi=I^{\gamma-\beta}\varphi\in L^1_{\mathrm{loc}} ψ = I γ − β φ ∈ L loc 1 . This is (3.2) with p = 0 p=0 p = 0 . □
On the space
E = { f ∈ C ∞ ( [ 0 , ∞ ) ; C ) : f ( j ) ( 0 ) = 0 for all j ∈ N 0 } \mathcal{E}=\{f\in C^\infty([0,\infty);\mathbb{C}):f^{(j)}(0)=0\text{ for all }j\in\mathbb{N}_0\} E = { f ∈ C ∞ ([ 0 , ∞ ) ; C ) : f ( j ) ( 0 ) = 0 for all j ∈ N 0 }
of smooth functions that are flat at the origin, operators of all real orders can be composed, and their orders add. This space contains C c ∞ ( ( 0 , ∞ ) ) C^\infty_c((0,\infty)) C c ∞ (( 0 , ∞ )) and is contained in L l o c 1 L^1_{\mathrm{loc}} L loc 1 . Let D = d / d x D=\mathrm{d}/\mathrm{d}x D = d / d x .
Lemma 3.2. The operators X X X , D D D , and I r I^r I r (r > 0 r>0 r > 0 ) map E \mathcal{E} E into itself, and D I r = I r D DI^r=I^rD D I r = I r D on E \mathcal{E} E . For r ≤ 0 r\leq0 r ≤ 0 and f ∈ E f\in\mathcal{E} f ∈ E define I r f = D m I r + m f I^r f=D^mI^{r+m}f I r f = D m I r + m f , where m ∈ N 0 m\in\mathbb{N}_0 m ∈ N 0 and m ≥ − r m\geq-r m ≥ − r . This does not depend on m m m , and on E \mathcal{E} E
I 0 = Id , I r I q = I r + q , X I r − I r X = r I r + 1 ( r , q ∈ R ) . (17) I^0=\operatorname{Id},\qquad I^rI^q=I^{r+q},\qquad XI^r-I^rX=rI^{r+1}\qquad(r,q\in\mathbb{R}). \tag*{(17)} I 0 = Id , I r I q = I r + q , X I r − I r X = r I r + 1 ( r , q ∈ R ) . (17) In particular, every I r I^r I r is a bijection of E \mathcal{E} E , and I − β I^{-\beta} I − β is the Riemann–Liouville derivative D β D^\beta D β on E \mathcal{E} E for β > 0 \beta>0 β > 0 .
Proof. Let r > 0 r>0 r > 0 and f ∈ E f\in\mathcal{E} f ∈ E . For continuous g g g , the function I r g ( x ) = x r Γ ( r ) − 1 ∫ 0 1 ( 1 − τ ) r − 1 g ( x τ ) d τ I^rg(x)=x^r\Gamma(r)^{-1}\int_0^1(1-\tau)^{r-1}g(x\tau)\,\mathrm{d}\tau I r g ( x ) = x r Γ ( r ) − 1 ∫ 0 1 ( 1 − τ ) r − 1 g ( xτ ) d τ is continuous on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) and vanishes at 0 0 0 . Since f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , we have f = I 1 f ′ f=I^1f' f = I 1 f ′ , hence I r f = I 1 ( I r f ′ ) I^rf=I^1(I^{r}f') I r f = I 1 ( I r f ′ ) by Lemma 3.1(a). Thus I r f ∈ C 1 ( [ 0 , ∞ ) ) I^rf\in C^1([0,\infty)) I r f ∈ C 1 ([ 0 , ∞ )) and D I r f = I r D f DI^rf=I^rDf D I r f = I r D f . As f ′ ∈ E f'\in\mathcal{E} f ′ ∈ E , induction gives I r f ∈ C ∞ ( [ 0 , ∞ ) ) I^rf\in C^\infty([0,\infty)) I r f ∈ C ∞ ([ 0 , ∞ )) and D j I r f = I r D j f D^jI^rf=I^rD^jf D j I r f = I r D j f for all j j j , and these functions vanish at 0 0 0 . Hence I r E ⊆ E I^r\mathcal{E}\subseteq\mathcal{E} I r E ⊆ E and D I r = I r D DI^r=I^rD D I r = I r D , and X X X and D D D clearly map this space into itself.
Since D I 1 = Id DI^1=\operatorname{Id} D I 1 = Id on E \mathcal{E} E , we have D m + 1 I r + m + 1 = D m I r + m D^{m+1}I^{r+m+1}=D^mI^{r+m} D m + 1 I r + m + 1 = D m I r + m whenever r + m ≥ 0 r+m\geq0 r + m ≥ 0 . Hence I r I^r I r does not depend on m m m , and I r = D m I r + m I^r=D^mI^{r+m} I r = D m I r + m for all r ∈ R r\in\mathbb{R} r ∈ R and all integers m ≥ max { 0 , − r } m\geq\max\{0,-r\} m ≥ max { 0 , − r } . As D D D commutes with the integrals of positive order,
I r I q = D m I r + m D m ′ I q + m ′ = D m + m ′ I r + q + m + m ′ = I r + q I^rI^q=D^mI^{r+m}D^{m'}I^{q+m'}=D^{m+m'}I^{r+q+m+m'}=I^{r+q} I r I q = D m I r + m D m ′ I q + m ′ = D m + m ′ I r + q + m + m ′ = I r + q
for large m , m ′ m,m' m , m ′ , so I r I^r I r is bijective with inverse I − r I^{-r} I − r . For the commutator, let m ≥ max { 1 , − r } m\geq\max\{1,-r\} m ≥ max { 1 , − r } . By Lemma 3.1(b) and [ X , D m ] = − m D m − 1 [X,D^m]=-mD^{m-1} [ X , D m ] = − m D m − 1 ,
[ X , D m I r + m ] = − m D m − 1 I r + m + ( r + m ) D m I r + m + 1 = − m I r + 1 + ( r + m ) I r + 1 = r I r + 1 . [X,D^mI^{r+m}]=-mD^{m-1}I^{r+m}+(r+m)D^mI^{r+m+1}=-mI^{r+1}+(r+m)I^{r+1}=rI^{r+1}. [ X , D m I r + m ] = − m D m − 1 I r + m + ( r + m ) D m I r + m + 1 = − m I r + 1 + ( r + m ) I r + 1 = r I r + 1 .
Finally, let β > 0 \beta>0 β > 0 , m = ⌈ β ⌉ m=\lceil\beta\rceil m = ⌈ β ⌉ , and f ∈ E f\in\mathcal{E} f ∈ E . Then g = I m − β f ∈ E g=I^{m-\beta}f\in\mathcal{E} g = I m − β f ∈ E satisfies g = I m D m g g=I^mD^mg g = I m D m g , which is (3.2) with p = 0 p=0 p = 0 ; hence D β f = D m I m − β f = I − β f D^\beta f=D^mI^{m-\beta}f=I^{-\beta}f D β f = D m I m − β f = I − β f . □
The fractional realization and normal ordering
Theorem 4.1. Let s ∈ R ∖ { 1 } s\in\mathbb{R}\setminus\{1\} s ∈ R ∖ { 1 } , h ∈ C ∖ { 0 } h\in\mathbb{C}\setminus\{0\} h ∈ C ∖ { 0 } , and α = 1 / ( s − 1 ) \alpha=1/(s-1) α = 1/ ( s − 1 ) . There is a unique homomorphism ρ : A s , h → End C ( E ) \rho:A_{s,h}\to\operatorname{End}_{\mathbb{C}}(\mathcal{E}) ρ : A s , h → End C ( E ) such that
ρ ( v a ) = I α a ( a ∈ Γ s ) , ρ ( u ) = h α X , \rho(v^a)=I^{\alpha a}\quad(a\in\Gamma_s),\qquad\rho(u)=\frac{h}{\alpha}X, ρ ( v a ) = I α a ( a ∈ Γ s ) , ρ ( u ) = α h X , and ρ \rho ρ is injective. Here ρ ( v ) = I α \rho(v)=I^\alpha ρ ( v ) = I α is a fractional integral if s > 1 s>1 s > 1 and the fractional derivative D − α D^{-\alpha} D − α if s < 1 s<1 s < 1 .
Proof. By Lemma 3.2, the operators I α a I^{\alpha a} I α a satisfy (2.5), and
[ h α X , I α a ] = h a I α a + 1 = h a I α ( a + s − 1 ) , \left[\frac{h}{\alpha}X,I^{\alpha a}\right]=haI^{\alpha a+1}=haI^{\alpha(a+s-1)}, [ α h X , I α a ] = ha I α a + 1 = ha I α ( a + s − 1 ) , because α ( s − 1 ) = 1 \alpha(s-1)=1 α ( s − 1 ) = 1 . This is (2.7), so ρ \rho ρ exists and is unique by the universal property of A s , h A_{s,h} A s , h .
Let P = ∑ ( a , k ) ∈ F c a , k v α a u k P=\sum_{(a,k)\in F}c_{a,k}v^{\alpha a}u^k P = ∑ ( a , k ) ∈ F c a , k v α a u k , with F ⊆ Γ s × N 0 F\subseteq\Gamma_s\times\mathbb{N}_0 F ⊆ Γ s × N 0 finite, and suppose that ρ ( P ) = 0 \rho(P)=0 ρ ( P ) = 0 . With c a , k ′ = c a , k ( h / α ) k c'_{a,k}=c_{a,k}(h/\alpha)^k c a , k ′ = c a , k ( h / α ) k this means that ∑ F c a , k ′ I α a X k = 0 \sum_F c'_{a,k}I^{\alpha a}X^k=0 ∑ F c a , k ′ I α a X k = 0 on E \mathcal E E . For λ > 0 \lambda>0 λ > 0 the dilation ( S λ f ) ( x ) = f ( λ x ) (S_\lambda f)(x)=f(\lambda x) ( S λ f ) ( x ) = f ( λ x ) is a bijection of E \mathcal E E , and
X S λ = λ − 1 S λ X , I r S λ = λ − r S λ I r ( r ∈ R ) ; XS_\lambda=\lambda^{-1}S_\lambda X,\qquad I^rS_\lambda=\lambda^{-r}S_\lambda I^r\quad(r\in\mathbb{R}); X S λ = λ − 1 S λ X , I r S λ = λ − r S λ I r ( r ∈ R ) ;
the second relation follows from (3.1) for r > 0 r>0 r > 0 and then, by taking inverses, for r < 0 r<0 r < 0 . Hence, for every f ∈ E f\in\mathcal E f ∈ E ,
0 = ρ ( P ) S λ f = S λ ∑ ( a , k ) ∈ F λ − α a − k c a , k ′ I α a X k f ( λ > 0 ) . 0=\rho(P)S_\lambda f=S_\lambda\sum_{(a,k)\in F}\lambda^{-\alpha a-k}c'_{a,k}I^{\alpha a}X^k f\qquad(\lambda>0). 0 = ρ ( P ) S λ f = S λ ( a , k ) ∈ F ∑ λ − α a − k c a , k ′ I α a X k f ( λ > 0 ) .
Since S λ S_\lambda S λ is injective and the functions λ ↦ λ − d \lambda\mapsto\lambda^{-d} λ ↦ λ − d with distinct real d d d are linearly independent, we obtain, for every d ∈ R d\in\mathbb{R} d ∈ R ,
∑ k b k I d − k X k = 0 on E , where b k = c ( d − k ) / α , k ′ \sum_k b_k I^{d-k}X^k=0\quad\text{on }\mathcal E,\qquad\text{where }b_k=c'_{(d-k)/\alpha,k} k ∑ b k I d − k X k = 0 on E , where b k = c ( d − k ) / α , k ′
and b k = 0 b_k=0 b k = 0 if ( ( d − k ) / α , k ) ∉ F ((d-k)/\alpha,k)\notin F (( d − k ) / α , k ) ∈ / F . Applying I − d I^{-d} I − d gives ∑ k b k D k X k = 0 \sum_k b_kD^kX^k=0 ∑ k b k D k X k = 0 on E \mathcal E E . Suppose that some b k b_k b k is nonzero, and let K K K be the largest such k k k . Since
D k X k = ∑ j = 0 K ( k j ) k ! j ! X j D j , D^kX^k=\sum_{j=0}^K\binom{k}{j}\frac{k!}{j!}X^jD^j, D k X k = j = 0 ∑ K ( j k ) j ! k ! X j D j ,
the operator ∑ k b k D k X k \sum_k b_kD^kX^k ∑ k b k D k X k has the form ∑ j = 0 K a j X j D j \sum_{j=0}^K a_jX^jD^j ∑ j = 0 K a j X j D j with a K = b K ≠ 0 a_K=b_K\ne0 a K = b K = 0 . Applied to a function f ∈ C c ∞ ( ( 0 , ∞ ) ) f\in C^\infty_c((0,\infty)) f ∈ C c ∞ (( 0 , ∞ )) with f ( x ) = ( x − 1 ) K f(x)=(x-1)^K f ( x ) = ( x − 1 ) K near x = 1 x=1 x = 1 , it gives a function with the value a K K ! ≠ 0 a_KK!\ne0 a K K ! = 0 at x = 1 x=1 x = 1 , a contradiction. Hence all b k b_k b k are zero, for every d d d . Since α ≠ 0 \alpha\ne0 α = 0 , each pair ( a , k ) ∈ F (a,k)\in F ( a , k ) ∈ F occurs for d = α a + k d=\alpha a+k d = α a + k , and therefore P = 0 P=0 P = 0 . □
Thus the relations of A s , h A_{s,h} A s , h are exactly the relations between X X X and a group of fractional integrals and derivatives. For s = 1 s=1 s = 1 this construction does not apply, since [ X , I r ] [X,I^r] [ X , I r ] has order r + 1 r+1 r + 1 .
Corollary 4.2. Let α ∈ R ∖ { 0 } \alpha\in\mathbb{R}\setminus\{0\} α ∈ R ∖ { 0 } , s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α , and n ≥ 1 n\ge1 n ≥ 1 . On E \mathcal E E ,
( I α X ) n = ∑ k = 1 n S s ; α ( n , k ) I ( α + 1 ) n − k X k , (18) (I^\alpha X)^n=\sum_{k=1}^n S_{s;\alpha}(n,k)I^{(\alpha+1)n-k}X^k, \tag*{(18)} ( I α X ) n = k = 1 ∑ n S s ; α ( n , k ) I ( α + 1 ) n − k X k , (18) ( X I α ) n = ∑ k = 1 n S s ; − α ( n , k ) X k I ( α + 1 ) n − k . (19) (XI^\alpha)^n=\sum_{k=1}^n S_{s;-\alpha}(n,k)X^kI^{(\alpha+1)n-k}. \tag*{(19)} ( X I α ) n = k = 1 ∑ n S s ; − α ( n , k ) X k I ( α + 1 ) n − k . (19) The operators X k I r X^kI^r X k I r ( k ∈ N 0 , r ∈ Z + Z α ) (k\in\mathbb{N}_0,\ r\in\mathbb{Z}+\mathbb{Z}\alpha) ( k ∈ N 0 , r ∈ Z + Z α ) are linearly independent on E \mathcal E E , and so are the operators I r X k I^rX^k I r X k . In particular, the coefficients in (4.1) and (4.2) are unique.
Proof. Take h = α h=\alpha h = α in Theorem 4.1. Then ρ ( u ) = X \rho(u)=X ρ ( u ) = X , ρ ( v ) = I α \rho(v)=I^\alpha ρ ( v ) = I α , and ρ ( v s ( n − k ) + k ) = I ( α + 1 ) n − k \rho(v^{s(n-k)+k})=I^{(\alpha+1)n-k} ρ ( v s ( n − k ) + k ) = I ( α + 1 ) n − k , because α s = α + 1 \alpha s=\alpha+1 α s = α + 1 . Applying ρ \rho ρ to (2.8) and (2.9) gives (4.1) and (4.2). The independence follows from Lemma 2.2 and the injectivity of ρ \rho ρ . □
To pass from E \mathcal E E to integrable functions, we use only operators of nonnegative order. Fix α ∈ R ∖ { 0 } \alpha\in\mathbb{R}\setminus\{0\} α ∈ R ∖ { 0 } , let s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α and G = Z + Z α = α Γ s G=\mathbb{Z}+\mathbb{Z}\alpha=\alpha\Gamma_s G = Z + Z α = α Γ s , and write J r = v r / α J_r=v^{r/\alpha} J r = v r / α for r ∈ G r\in G r ∈ G in the algebra A = A s , α A=A_{s,\alpha} A = A s , α . Then (2.5) and (2.7) become
J 0 = 1 , J r J q = J r + q , [ u , J r ] = r J r + 1 ( r , q ∈ G ) , (20) J_0=1,\qquad J_rJ_q=J_{r+q},\qquad[u,J_r]=rJ_{r+1}\quad(r,q\in G), \tag*{(20)} J 0 = 1 , J r J q = J r + q , [ u , J r ] = r J r + 1 ( r , q ∈ G ) , (20)
and v = J α v=J_\alpha v = J α , v s ( n − k ) + k = J ( α + 1 ) n − k v^{s(n-k)+k}=J_{(\alpha+1)n-k} v s ( n − k ) + k = J ( α + 1 ) n − k . For c ≥ 0 c\ge0 c ≥ 0 let A ≥ c A_{\ge c} A ≥ c be the span of { J r u m : r ∈ G , r ≥ c , m ∈ N 0 } \{J_ru^m:r\in G,\ r\ge c,\ m\in\mathbb{N}_0\} { J r u m : r ∈ G , r ≥ c , m ∈ N 0 } .
Lemma 4.3. (a) A ≥ 0 A_{\ge0} A ≥ 0 is a subalgebra of A A A . Moreover, u A ≥ c ⊆ A ≥ c uA_{\ge c}\subseteq A_{\ge c} u A ≥ c ⊆ A ≥ c and J r A ≥ c ⊆ A ≥ c + r J_rA_{\ge c}\subseteq A_{\ge c+r} J r A ≥ c ⊆ A ≥ c + r for c ≥ 0 c\ge0 c ≥ 0 and r ∈ G r\in G r ∈ G with c + r ≥ 0 c+r\ge0 c + r ≥ 0 .
(b) There is a unique homomorphism σ : A ≥ 0 → End C ( L l o c 1 ) \sigma:A_{\ge0}\to\operatorname{End}_{\mathbb{C}}(L^1_{\mathrm{loc}}) σ : A ≥ 0 → End C ( L loc 1 ) such that σ ( J r ) = I r \sigma(J_r)=I^r σ ( J r ) = I r for r ∈ G ∩ [ 0 , ∞ ) r\in G\cap[0,\infty) r ∈ G ∩ [ 0 , ∞ ) and σ ( u ) = X \sigma(u)=X σ ( u ) = X .
(c) Let γ ∈ G \gamma\in G γ ∈ G , γ > 0 \gamma>0 γ > 0 , w ∈ A ≥ γ w\in A_{\ge\gamma} w ∈ A ≥ γ , and φ ∈ L l o c 1 \varphi\in L^1_{\mathrm{loc}} φ ∈ L loc 1 . Then σ ( w ) φ ∈ Dom ( D γ ) \sigma(w)\varphi\in\operatorname{Dom}(D^\gamma) σ ( w ) φ ∈ Dom ( D γ ) and D γ σ ( w ) φ = σ ( J − γ w ) φ D^\gamma\sigma(w)\varphi=\sigma(J_{-\gamma}w)\varphi D γ σ ( w ) φ = σ ( J − γ w ) φ .
Proof. (a) By (4.3), u J r u m = J r u m + 1 + r J r + 1 u m uJ_ru^m=J_ru^{m+1}+rJ_{r+1}u^m u J r u m = J r u m + 1 + r J r + 1 u m and J r J q u m = J r + q u m J_rJ_qu^m=J_{r+q}u^m J r J q u m = J r + q u m . This gives the two inclusions, and they imply that A ≥ 0 A_{\geq0} A ≥ 0 is closed under multiplication. (b) The span R + R_+ R + of { J r : r ∈ G , r ≥ 0 } \{J_r:r\in G,\ r\geq0\} { J r : r ∈ G , r ≥ 0 } is a subalgebra of R s R_s R s with δ s , α ( R + ) ⊆ R + \delta_{s,\alpha}(R_+)\subseteq R_+ δ s , α ( R + ) ⊆ R + , and A ≥ 0 = ⨁ m R + u m A_{\geq0}=\bigoplus_m R_+u^m A ≥ 0 = ⨁ m R + u m is the Ore extension of R + R_+ R + by the restricted derivation. By Lemma 3.1(a), J r ↦ I r J_r\mapsto I^r J r ↦ I r defines a homomorphism R + → End C ( L l o c 1 ) R_+\to\operatorname{End}_{\mathbb{C}}(L^1_{\mathrm{loc}}) R + → End C ( L loc 1 ) , and by Lemma 3.1(b), [ X , I r ] = r I r + 1 [X,I^r]=rI^{r+1} [ X , I r ] = r I r + 1 is the image of δ s , α ( J r ) = r J r + 1 \delta_{s,\alpha}(J_r)=rJ_{r+1} δ s , α ( J r ) = r J r + 1 . The universal property of Ore extensions [7 Chapter 2] gives σ \sigma σ . (c) By linearity we may assume that w = J r u m w=J_ru^m w = J r u m with r ≥ γ r\geq\gamma r ≥ γ . Then σ ( w ) φ = I r ( X m φ ) \sigma(w)\varphi=I^r(X^m\varphi) σ ( w ) φ = I r ( X m φ ) , and Lemma 3.1(c) gives D γ I r ( X m φ ) = I r − γ X m φ = σ ( J − γ w ) φ D^\gamma I^r(X^m\varphi)=I^{r-\gamma}X^m\varphi=\sigma(J_{-\gamma}w)\varphi D γ I r ( X m φ ) = I r − γ X m φ = σ ( J − γ w ) φ .
Theorem 4.4. Let n ≥ 1 n\geq1 n ≥ 1 .
(a) Let α > 0 \alpha>0 α > 0 and s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α . Then (4.1) and (4.2) hold on L l o c 1 L^1_{\mathrm{loc}} L loc 1 .
(b) Let β > 0 \beta>0 β > 0 and f = I n β φ f=I^{n\beta}\varphi f = I n β φ with φ ∈ L l o c 1 \varphi\in L^1_{\mathrm{loc}} φ ∈ L loc 1 . Then every Riemann–Liouville derivative below is applied to a function in its domain. Moreover, D ( β − 1 ) n + k f = I n − k φ D^{(\beta-1)n+k}f=I^{n-k}\varphi D ( β − 1 ) n + k f = I n − k φ for 1 ≤ k ≤ n 1\leq k\leq n 1 ≤ k ≤ n , and
( D β X ) n f = ∑ k = 1 n S 1 − 1 / β ; − β ( n , k ) D ( β − 1 ) n + k X k f , (21) (D^\beta X)^n f=\sum_{k=1}^{n}S_{1-1/\beta;-\beta}(n,k)D^{(\beta-1)n+k}X^k f, \tag*{(21)} ( D β X ) n f = k = 1 ∑ n S 1 − 1/ β ; − β ( n , k ) D ( β − 1 ) n + k X k f , (21) ( X D β ) n f = ∑ k = 1 n S 1 − 1 / β ; β ( n , k ) X k D ( β − 1 ) n + k f . (22) (XD^\beta)^n f=\sum_{k=1}^{n}S_{1-1/\beta;\beta}(n,k)X^kD^{(\beta-1)n+k}f. \tag*{(22)} ( X D β ) n f = k = 1 ∑ n S 1 − 1/ β ; β ( n , k ) X k D ( β − 1 ) n + k f . (22)
Proof. (a) Here u , v = J α u,v=J_\alpha u , v = J α , and all terms of (2.8) and (2.9) lie in A ≥ 0 A_{\geq0} A ≥ 0 , because ( α + 1 ) n − k ≥ 0 (\alpha+1)n-k\geq0 ( α + 1 ) n − k ≥ 0 . Applying σ \sigma σ to Theorem 2.3 for A s , α A_{s,\alpha} A s , α gives the result.
(b) We use Lemma 4.3 with α = − β \alpha=-\beta α = − β . Then s = 1 − 1 / β s=1-1/\beta s = 1 − 1/ β , h = − β h=-\beta h = − β , v = J − β v=J_{-\beta} v = J − β , and f = σ ( J n β ) φ f=\sigma(J_{n\beta})\varphi f = σ ( J n β ) φ . We claim that, for 0 ≤ j ≤ n 0\leq j\leq n 0 ≤ j ≤ n ,
( D β X ) j f = σ ( w j ) φ , w j = ( v u ) j J n β ∈ A ≥ ( n − j ) β . (D^\beta X)^j f=\sigma(w_j)\varphi,\qquad w_j=(vu)^jJ_{n\beta}\in A_{\geq(n-j)\beta}. ( D β X ) j f = σ ( w j ) φ , w j = ( v u ) j J n β ∈ A ≥ ( n − j ) β .
This is clear for j = 0 j=0 j = 0 . If it holds for some j < n j<n j < n , then u w j ∈ A ≥ ( n − j ) β uw_j\in A_{\geq(n-j)\beta} u w j ∈ A ≥ ( n − j ) β and X σ ( w j ) φ = σ ( u w j ) φ X\sigma(w_j)\varphi=\sigma(uw_j)\varphi X σ ( w j ) φ = σ ( u w j ) φ . Since ( n − j ) β ≥ β (n-j)\beta\geq\beta ( n − j ) β ≥ β , Lemma 4.3(c) shows that this function lies in Dom ( D β ) \operatorname{Dom}(D^\beta) Dom ( D β ) and that D β σ ( u w j ) φ = σ ( J − β u w j ) φ = σ ( w j + 1 ) φ D^\beta\sigma(uw_j)\varphi=\sigma(J_{-\beta}uw_j)\varphi=\sigma(w_{j+1})\varphi D β σ ( u w j ) φ = σ ( J − β u w j ) φ = σ ( w j + 1 ) φ , and w j + 1 ∈ A ≥ ( n − j − 1 ) β w_{j+1}\in A_{\geq(n-j-1)\beta} w j + 1 ∈ A ≥ ( n − j − 1 ) β by Lemma 4.3(a). In the same way, ( X D β ) j f = σ ( ( u v ) j J n β ) φ (XD^\beta)^jf=\sigma((uv)^jJ_{n\beta})\varphi ( X D β ) j f = σ (( uv ) j J n β ) φ .
Put γ k = ( β − 1 ) n + k \gamma_k=(\beta-1)n+k γ k = ( β − 1 ) n + k , so that v s ( n − k ) + k = J − γ k v^{s(n-k)+k}=J_{-\gamma_k} v s ( n − k ) + k = J − γ k and n β − γ k = n − k n\beta-\gamma_k=n-k n β − γ k = n − k . Theorem 2.3 for A s , − β A_{s,-\beta} A s , − β gives
( v u ) n J n β = ∑ k = 1 n S s ; − β ( n , k ) J − γ k u k J n β , ( u v ) n J n β = ∑ k = 1 n S s ; β ( n , k ) u k J n − k . (vu)^nJ_{n\beta}=\sum_{k=1}^{n}S_{s;-\beta}(n,k)J_{-\gamma_k}u^kJ_{n\beta},\qquad(uv)^nJ_{n\beta}=\sum_{k=1}^{n}S_{s;\beta}(n,k)u^kJ_{n-k}. ( v u ) n J n β = k = 1 ∑ n S s ; − β ( n , k ) J − γ k u k J n β , ( uv ) n J n β = k = 1 ∑ n S s ; β ( n , k ) u k J n − k .
Here u k J n β ∈ A ≥ n β u^kJ_{n\beta}\in A_{\geq n\beta} u k J n β ∈ A ≥ n β and γ k ≤ n β \gamma_k\leq n\beta γ k ≤ n β . Hence σ ( J − γ k u k J n β ) φ = D γ k X k f \sigma(J_{-\gamma_k}u^kJ_{n\beta})\varphi=D^{\gamma_k}X^kf σ ( J − γ k u k J n β ) φ = D γ k X k f and σ ( J n − k ) φ = D γ k f \sigma(J_{n-k})\varphi=D^{\gamma_k}f σ ( J n − k ) φ = D γ k f , by Lemma 4.3(c) if γ k > 0 \gamma_k>0 γ k > 0 and by the convention D γ k = I − γ k D^{\gamma_k}=I^{-\gamma_k} D γ k = I − γ k if γ k ≤ 0 \gamma_k\leq0 γ k ≤ 0 . Applying σ \sigma σ to the two identities gives (4.4) and (4.5).
The same statements hold on a finite interval ( 0 , T ) (0,T) ( 0 , T ) . The first cases are
( X I α ) 2 = X 2 I 2 α − α X I 2 α + 1 , ( X I α ) 3 = X 3 I 3 α − 3 α X 2 I 3 α + 1 + α ( 2 α + 1 ) X I 3 α + 2 , (XI^\alpha)^2=X^2I^{2\alpha}-\alpha XI^{2\alpha+1},\qquad(XI^\alpha)^3=X^3I^{3\alpha}-3\alpha X^2I^{3\alpha+1}+\alpha(2\alpha+1)XI^{3\alpha+2}, ( X I α ) 2 = X 2 I 2 α − α X I 2 α + 1 , ( X I α ) 3 = X 3 I 3 α − 3 α X 2 I 3 α + 1 + α ( 2 α + 1 ) X I 3 α + 2 ,
( X D β ) 2 = X 2 D 2 β + β X D 2 β − 1 , ( X D β ) 3 = X 3 D 3 β + 3 β X 2 D 3 β − 1 + β ( 2 β − 1 ) X D 3 β − 2 . (XD^\beta)^2=X^2D^{2\beta}+\beta XD^{2\beta-1},\qquad(XD^\beta)^3=X^3D^{3\beta}+3\beta X^2D^{3\beta-1}+\beta(2\beta-1)XD^{3\beta-2}. ( X D β ) 2 = X 2 D 2 β + β X D 2 β − 1 , ( X D β ) 3 = X 3 D 3 β + 3 β X 2 D 3 β − 1 + β ( 2 β − 1 ) X D 3 β − 2 .
Remark 4.5. For β ∈ N \beta\in\mathbb{N} β ∈ N both sides of (4.4) and (4.5) are differential operators with polynomial coefficients; they agree on C ∞ ( ( 0 , ∞ ) ) ⊆ E C^\infty((0,\infty))\subseteq\mathcal{E} C ∞ (( 0 , ∞ )) ⊆ E by Corollary 4.2, hence on C n β ( ( 0 , ∞ ) ) C^{n\beta}((0,\infty)) C n β (( 0 , ∞ )) . For non-integer β \beta β , a condition at the origin is needed. Let 1 < β < 2 1<\beta<2 1 < β < 2 and f ( x ) = x β − 2 f(x)=x^{\beta-2} f ( x ) = x β − 2 . Then I 2 − β f I^{2-\beta}f I 2 − β f is constant, so D β f = 0 D^\beta f=0 D β f = 0 and ( X D β ) n f = 0 (XD^\beta)^nf=0 ( X D β ) n f = 0 for all n ≥ 1 n\geq1 n ≥ 1 . However, f ∉ Dom ( D 2 β ) f\notin\operatorname{Dom}(D^{2\beta}) f ∈ / Dom ( D 2 β ) , because I m − 2 β f I^{m-2\beta}f I m − 2 β f is a nonzero multiple of x m − 2 β x^{m-2\beta} x m − 2 β for m = ⌈ 2 β ⌉ m=\lceil2\beta\rceil m = ⌈ 2 β ⌉ , and this function is not of the form (3.2). So the right-hand side of (4.5) is not defined for n = 2 n=2 n = 2 in the sense of § 3 . If its derivatives are computed pointwise on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , that is, by D γ x μ = Γ ( μ + 1 ) x μ − γ Γ ( μ + 1 − γ ) D^\gamma x^\mu= \frac{\Gamma(\mu+1)x^{\mu-\gamma}}{\Gamma(\mu+1-\gamma)} D γ x μ = Γ ( μ + 1 − γ ) Γ ( μ + 1 ) x μ − γ for μ > − 1 \mu>-1 μ > − 1 , it equals
x 2 D 2 β f + β x D 2 β − 1 f = Γ ( β − 1 ) ( 1 Γ ( − β − 1 ) + β Γ ( − β ) ) x − β = − Γ ( β − 1 ) Γ ( − β ) x − β ≠ 0. x^2D^{2\beta}f+\beta xD^{2\beta-1}f=\Gamma(\beta-1)\left(\frac{1}{\Gamma(-\beta-1)}+\frac{\beta}{\Gamma(-\beta)}\right)x^{-\beta}=-\frac{\Gamma(\beta-1)}{\Gamma(-\beta)}x^{-\beta}\ne0. x 2 D 2 β f + β x D 2 β − 1 f = Γ ( β − 1 ) ( Γ ( − β − 1 ) 1 + Γ ( − β ) β ) x − β = − Γ ( − β ) Γ ( β − 1 ) x − β = 0.
Thus D β D β f ≠ D 2 β f D^\beta D^\beta f\ne D^{2\beta}f D β D β f = D 2 β f , even pointwise. On L l o c 1 L^1_{\mathrm{loc}} L loc 1 the Riemann–Liouville derivatives do not form a semigroup, and the operators D a β a D_a^{\beta a} D a β a cannot be used as the powers v a v^a v a of v = D β v=D^\beta v = D β in (2.5). On the other hand, the range of I n β I^{n\beta} I n β is a sufficient domain but not the largest one. For example, for n = 2 n=2 n = 2 and f ( x ) = x 2 β − 1 f(x)=x^{2\beta-1} f ( x ) = x 2 β − 1 , which is not in the range of I 2 β I^{2\beta} I 2 β , both sides of (4.5) are defined and equal β Γ ( 2 β ) x \beta\Gamma(2\beta)x β Γ ( 2 β ) x .
Remark 4.6. If α = 1 / m \alpha=1/m α = 1/ m with m ∈ N m\in\mathbb{N} m ∈ N , then s = m + 1 s=m+1 s = m + 1 and J = I 1 / m J=I^{1/m} J = I 1/ m satisfies [ X , J ] = 1 m J m + 1 [X,J]=\frac{1}{m}J^{m+1} [ X , J ] = m 1 J m + 1 , so that (4.1) and (4.2) follow from this single relation by Corollary 2.5 . For all other α > 0 \alpha>0 α > 0 , s s s is not an integer, and the proof uses the whole semigroup; compare Proposition 2.6 .
Remark 4.7. The realization of Theorem 4.1 is related by the Laplace transform L f ( p ) = ∫ 0 ∞ e − p t f ( t ) d t \mathcal{L}f(p)=\int_0^\infty e^{-pt}f(t)\,dt L f ( p ) = ∫ 0 ∞ e − pt f ( t ) d t to the differential realization v a ↦ y a v^a\mapsto y^a v a ↦ y a , u ↦ h y s d / d y u\mapsto hy^s\,d/dy u ↦ h y s d / d y of A s , h A_{s,h} A s , h on C ∞ ( ( 0 , ∞ ) ) C^\infty((0,\infty)) C ∞ (( 0 , ∞ )) [17 Eq. (6)] , [19 Section 4] . Let G \mathcal{G} G be the space of all f ∈ E f\in\mathcal{E} f ∈ E whose derivatives have at most polynomial growth; X X X and all I r I^r I r map it into itself. For f ∈ G f\in\mathcal{G} f ∈ G and p > 0 p>0 p > 0 , the convolution theorem and integration by parts give
L ( I r f ) ( p ) = p − r L f ( p ) , L ( X f ) ( p ) = − d d p L f ( p ) . \mathcal{L}(I^r f)(p)=p^{-r}\mathcal{L}f(p),\qquad\mathcal{L}(Xf)(p)=-\frac{d}{dp}\mathcal{L}f(p). L ( I r f ) ( p ) = p − r L f ( p ) , L ( X f ) ( p ) = − d p d L f ( p ) .
and with y = p − α y=p^{-\alpha} y = p − α we have p − α a = y a p^{-\alpha a}=y^a p − α a = y a and − d / d p = α y s d / d y -d/dp=\alpha y^s\,d/dy − d / d p = α y s d / d y . Hence, on G \mathcal{G} G , the Laplace transform carries (4.1) and (4.2) to (2.8) and (2.9) in the differential realization with h = α h=\alpha h = α . These are special cases of the expansions of Maier [15 Theorem 6.4] . This correspondence holds only on G \mathcal{G} G . Theorem 4.4 deals with larger function spaces, and for derivatives this difference matters (Remark 4.5). The probabilistic formulas of § 6 are not obtained in this way.
Coefficients, mixed products and special cases
Proposition 5.1. Let α ∈ R ∖ { 0 } \alpha\in\mathbb{R}\setminus\{0\} α ∈ R ∖ { 0 } , s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α , and n ≥ 1 n\ge1 n ≥ 1 . Then
∏ i = 0 n − 1 ( z + ( α + 1 ) i ) = ∑ k = 1 n S s ; α ( n , k ) ( z ) k , (23) \prod_{i=0}^{n-1}\left(z+(\alpha+1)i\right)=\sum_{k=1}^{n}S_{s;\alpha}(n,k)(z)_k, \tag*{(23)} i = 0 ∏ n − 1 ( z + ( α + 1 ) i ) = k = 1 ∑ n S s ; α ( n , k ) ( z ) k , (23)
S s ; α ( n , k ) = 1 k ! ∑ j = 0 k ( − 1 ) j ( k j ) ∏ i = 0 n − 1 ( ( α + 1 ) i − j ) , (24) S_{s;\alpha}(n,k)=\frac{1}{k!}\sum_{j=0}^{k}(-1)^j\binom{k}{j}\prod_{i=0}^{n-1}\left((\alpha+1)i-j\right), \tag*{(24)} S s ; α ( n , k ) = k ! 1 j = 0 ∑ k ( − 1 ) j ( j k ) i = 0 ∏ n − 1 ( ( α + 1 ) i − j ) , (24)
and S s ; α ( n , k ) S_{s;\alpha}(n,k) S s ; α ( n , k ) is a polynomial in α \alpha α of degree n − k n-k n − k with nonnegative integer coefficients. Moreover, for α > 0 \alpha>0 α > 0 and μ > − 1 \mu>-1 μ > − 1 ,
( I α X ) n x μ = Γ ( μ + 1 ) Γ ( μ + 1 + ( α + 1 ) n ) ∏ i = 0 n − 1 ( μ + 1 + ( α + 1 ) i ) x μ + ( α + 1 ) n , (25) (I^\alpha X)^n x^\mu=\frac{\Gamma(\mu+1)}{\Gamma(\mu+1+(\alpha+1)n)}\prod_{i=0}^{n-1}\left(\mu+1+(\alpha+1)i\right)x^{\mu+(\alpha+1)n}, \tag*{(25)} ( I α X ) n x μ = Γ ( μ + 1 + ( α + 1 ) n ) Γ ( μ + 1 ) i = 0 ∏ n − 1 ( μ + 1 + ( α + 1 ) i ) x μ + ( α + 1 ) n , (25)
and (5.1) with z = μ + 1 z=\mu+1 z = μ + 1 is the identity (4.1) applied to x μ x^\mu x μ .
Proof. Identity (5.1) is Lemma 2.1 with h = α h=\alpha h = α , since h s = α + 1 hs=\alpha+1 h s = α + 1 and h ( 1 − s ) = − 1 h(1-s)=-1 h ( 1 − s ) = − 1 . Replacing z z z by − z -z − z turns its right-hand side into ∑ k ( − 1 ) k S s ; α ( n , k ) z ( z − 1 ) ⋯ ( z − k + 1 ) \sum_k(-1)^kS_{s;\alpha}(n,k)z(z-1)\cdots(z-k+1) ∑ k ( − 1 ) k S s ; α ( n , k ) z ( z − 1 ) ⋯ ( z − k + 1 ) , and Newton’s forward-difference formula gives (5.2). For h = α h=\alpha h = α the recurrence (2.1) has the multiplier α ( k + s ( n − k ) ) = n α + ( n − k ) \alpha(k+s(n-k))=n\alpha+(n-k) α ( k + s ( n − k )) = n α + ( n − k ) , and this gives the statement about polynomials. Finally, I r x ν = Γ ( ν + 1 ) x ν + r / Γ ( ν + 1 + r ) I^r x^\nu=\Gamma(\nu+1)x^{\nu+r}/\Gamma(\nu+1+r) I r x ν = Γ ( ν + 1 ) x ν + r /Γ ( ν + 1 + r ) for ν > − 1 \nu>-1 ν > − 1 and r ≥ 0 r\geq0 r ≥ 0 , so that
I α X x ν = ( ν + 1 ) Γ ( ν + 1 ) Γ ( ν + α + 2 ) x ν + α + 1 . I^\alpha Xx^\nu=(\nu+1)\frac{\Gamma(\nu+1)}{\Gamma(\nu+\alpha+2)}x^{\nu+\alpha+1}. I α X x ν = ( ν + 1 ) Γ ( ν + α + 2 ) Γ ( ν + 1 ) x ν + α + 1 .
We iterate this with ν = μ + ( α + 1 ) i \nu=\mu+(\alpha+1)i ν = μ + ( α + 1 ) i , i = 0 , … , n − 1 i=0,\ldots,n-1 i = 0 , … , n − 1 ; most gamma factors cancel, and we obtain (25) . The right-hand side of (18) maps x μ x^\mu x μ to ∑ k S s ; α ( n , k ) Γ ( μ + k + 1 ) x μ + ( α + 1 ) n / Γ ( μ + 1 + ( α + 1 ) n ) \sum_k S_{s;\alpha}(n,k)\Gamma(\mu+k+1)x^{\mu+(\alpha+1)n}/\Gamma(\mu+1+(\alpha+1)n) ∑ k S s ; α ( n , k ) Γ ( μ + k + 1 ) x μ + ( α + 1 ) n /Γ ( μ + 1 + ( α + 1 ) n ) , and Γ ( μ + k + 1 ) = ( μ + 1 ) k Γ ( μ + 1 ) \Gamma(\mu+k+1)=(\mu+1)_k\Gamma(\mu+1) Γ ( μ + k + 1 ) = ( μ + 1 ) k Γ ( μ + 1 ) .
By (6) , the coefficients of the derivative expansions (21) and (22) are obtained from (24) with α = − β \alpha=-\beta α = − β : S 1 − 1 / β ; β ( n , k ) = ( − 1 ) n − k S 1 − 1 / β ; − β ( n , k ) S_{1-1/\beta;\beta}(n,k)=(-1)^{n-k}S_{1-1/\beta;-\beta}(n,k) S 1 − 1/ β ; β ( n , k ) = ( − 1 ) n − k S 1 − 1/ β ; − β ( n , k ) .
Corollary 5.2. Let n ≥ 1 n\geq1 n ≥ 1 .
(a) ( α = 1. ) (\alpha=1.) ( α = 1. ) Let V = I 1 V=I^1 V = I 1 and c ( m , j ) = ( m + j ) ! 2 j j ! ( m − j ) ! c(m,j)=\frac{(m+j)!}{2^j j!(m-j)!} c ( m , j ) = 2 j j ! ( m − j )! ( m + j )! . On L l o c 1 L^1_{\mathrm{loc}} L loc 1 ,
( X V ) n = ∑ j = 0 n − 1 ( − 1 ) j c ( n − 1 , j ) X n − j V n + j , ( V X ) n = ∑ j = 0 n − 1 c ( n − 1 , j ) V n + j X n − j . (XV)^n=\sum_{j=0}^{n-1}(-1)^j c(n-1,j)X^{n-j}V^{n+j},\qquad(VX)^n=\sum_{j=0}^{n-1}c(n-1,j)V^{n+j}X^{n-j}. ( X V ) n = j = 0 ∑ n − 1 ( − 1 ) j c ( n − 1 , j ) X n − j V n + j , ( V X ) n = j = 0 ∑ n − 1 c ( n − 1 , j ) V n + j X n − j . (b) ( β = 1. ) (\beta=1.) ( β = 1. ) On C n ( ( 0 , ∞ ) ) C^n((0,\infty)) C n (( 0 , ∞ )) ,
( X D ) n = ∑ k = 1 n S ( n , k ) X k D k , ( D X ) n = ∑ k = 1 n ( − 1 ) n − k S ( n , k ) D k X k . (XD)^n=\sum_{k=1}^{n}S(n,k)X^kD^k,\qquad(DX)^n=\sum_{k=1}^{n}(-1)^{n-k}S(n,k)D^kX^k. ( X D ) n = k = 1 ∑ n S ( n , k ) X k D k , ( D X ) n = k = 1 ∑ n ( − 1 ) n − k S ( n , k ) D k X k . (c) ( β = 1 2 . ) (\beta=\frac{1}{2}.) ( β = 2 1 . ) For f f f in the range of I n / 2 I^{n/2} I n /2 on L l o c 1 L^1_{\mathrm{loc}} L loc 1 ,
( X D 1 / 2 ) n f = ∑ n / 2 ≤ k ≤ n 2 k − n B ( n , k ) X k D k − n / 2 f , B ( n , k ) = n ! 2 n − k ( 2 k − n ) ! ( n − k ) ! . (XD^{1/2})^n f=\sum_{n/2\leq k\leq n}2^{k-n}B(n,k)X^kD^{k-n/2}f,\qquad B(n,k)=\frac{n!}{2^{n-k}(2k-n)!(n-k)!}. ( X D 1/2 ) n f = n /2 ≤ k ≤ n ∑ 2 k − n B ( n , k ) X k D k − n /2 f , B ( n , k ) = 2 n − k ( 2 k − n )! ( n − k )! n ! . Proof. (a) Here s = 2 s=2 s = 2 , and S 2 ; 1 ( n , n − j ) = c ( n − 1 , j ) S_{2;1}(n,n-j)=c(n-1,j) S 2 ; 1 ( n , n − j ) = c ( n − 1 , j ) ; by (5) this is equivalent to c ( n , j ) = c ( n − 1 , j ) + ( n + j − 1 ) c ( n − 1 , j − 1 ) c(n,j)=c(n-1,j)+(n+j-1)c(n-1,j-1) c ( n , j ) = c ( n − 1 , j ) + ( n + j − 1 ) c ( n − 1 , j − 1 ) , which can be checked directly; see also [18 , 19 ] . Now apply Theorem 4.4(a) and (6) . (b) Here s = 0 s=0 s = 0 , and (5) with h = 1 h=1 h = 1 is the recurrence of the Stirling numbers of the second kind; see Remark 4.5 for the domain. (c) Here s = − 1 s=-1 s = − 1 and S − 1 ; 1 / 2 ( n , k ) = 2 k − n S − 1 ; 1 ( n , k ) S_{-1;1/2}(n,k)=2^{k-n}S_{-1;1}(n,k) S − 1 ; 1/2 ( n , k ) = 2 k − n S − 1 ; 1 ( n , k ) . The numbers S − 1 ; 1 ( n , k ) S_{-1;1}(n,k) S − 1 ; 1 ( n , k ) satisfy S ( n + 1 , k ) = S ( n , k − 1 ) + ( 2 k − n ) S ( n , k ) S(n+1,k)=S(n,k-1)+(2k-n)S(n,k) S ( n + 1 , k ) = S ( n , k − 1 ) + ( 2 k − n ) S ( n , k ) , which is the recurrence of the number B ( n , k ) B(n,k) B ( n , k ) of partitions of an n n n -set into k k k blocks of size at most two: the new element is either a block by itself or it joins one of the 2 k − n 2k-n 2 k − n singletons. Now apply (22) , noting that ( β − 1 ) n + k = k − n / 2 (\beta-1)n+k=k-n/2 ( β − 1 ) n + k = k − n /2 .
The first formula in (a) is Benmoussa’s expansion [3 Theorem 1] ; the numbers c ( m , j ) c(m,j) c ( m , j ) are the coefficients of the Bessel polynomials [21 A001498] . Part (b) is Scherk’s classical identity and its reversed form [16 ] , and the B ( n , k ) B(n,k) B ( n , k ) in (c) are the Bessel numbers of the second kind [19 ] . The numbers in (c) also appear for a differential operator. Since [ D , x a / 2 ] = a 2 x a / 2 − 1 [D,x^{a/2}]=\frac{a}{2}x^{a/2-1} [ D , x a /2 ] = 2 a x a /2 − 1 for a ∈ Z a\in\mathbb{Z} a ∈ Z , the assignment v a ↦ x a / 2 v^a\mapsto x^{a/2} v a ↦ x a /2 , u ↦ D u\mapsto D u ↦ D realizes A − 1 , 1 / 2 A_{-1,1/2} A − 1 , 1/2 on C ∞ ( ( 0 , ∞ ) ) C^\infty((0,\infty)) C ∞ (( 0 , ∞ )) , and (12) gives ( x 1 / 2 D ) n = ∑ n / 2 ≤ k ≤ n 2 k − n B ( n , k ) x k − n / 2 D k (x^{1/2}D)^n=\sum_{n/2\leq k\leq n}2^{k-n}B(n,k)x^{k-n/2}D^k ( x 1/2 D ) n = ∑ n /2 ≤ k ≤ n 2 k − n B ( n , k ) x k − n /2 D k . After multiplication by 2 ⌊ n / 2 ⌋ 2^{\lfloor n/2\rfloor} 2 ⌊ n /2 ⌋ , these coefficients are the Stirling numbers of order 1 / 2 1/2 1/2 introduced by Katugampola [12 ] , in the normalization of Schork [25 ] . Schork identified them with a scaled subfamily of the numbers (8) .
Products with a fractional operator on each side of X X X involve the full family (8) .
Proposition 5.3. Let γ , δ ∈ R \gamma,\delta\in\mathbb{R} γ , δ ∈ R , α = γ + δ \alpha=\gamma+\delta α = γ + δ , and n ∈ N 0 n\in\mathbb{N}_0 n ∈ N 0 . On E \mathcal{E} E ,
( I γ X I δ ) n = ∑ k = 0 n S ( n , k ; α + 1 , 1 , − γ ) X k I ( α + 1 ) n − k = ∑ k = 0 n ( − 1 ) n − k S ( n , k ; α + 1 , 1 , − δ ) I ( α + 1 ) n − k X k . (26) (I^\gamma X I^\delta)^n=\sum_{k=0}^{n}S(n,k;\alpha+1,1,-\gamma)X^kI^{(\alpha+1)n-k}
=
\sum_{k=0}^{n}(-1)^{n-k}S(n,k;\alpha+1,1,-\delta)I^{(\alpha+1)n-k}X^k. \tag*{(26)} ( I γ X I δ ) n = k = 0 ∑ n S ( n , k ; α + 1 , 1 , − γ ) X k I ( α + 1 ) n − k = k = 0 ∑ n ( − 1 ) n − k S ( n , k ; α + 1 , 1 , − δ ) I ( α + 1 ) n − k X k . (26) If γ , δ ≥ 0 \gamma,\delta\geq0 γ , δ ≥ 0 , these identities hold on L l o c 1 L^1_{\mathrm{loc}} L loc 1 .
Proof. By (17) , I γ X I δ = X I α − γ I α + 1 I^\gamma X I^\delta=XI^\alpha-\gamma I^{\alpha+1} I γ X I δ = X I α − γ I α + 1 and, with N = ( α + 1 ) n N=(\alpha+1)n N = ( α + 1 ) n ,
X k I N − k ( X I α − γ I α + 1 ) = X k + 1 I N + α − k + ( k − N − γ ) X k I N + α + 1 − k . X^kI^{N-k}(XI^\alpha-\gamma I^{\alpha+1})=X^{k+1}I^{N+\alpha-k}+(k-N-\gamma)X^kI^{N+\alpha+1-k}. X k I N − k ( X I α − γ I α + 1 ) = X k + 1 I N + α − k + ( k − N − γ ) X k I N + α + 1 − k .
Hence the coefficients C ( n , k ) C(n,k) C ( n , k ) of the first expansion satisfy C ( 0 , 0 ) = 1 C(0,0)=1 C ( 0 , 0 ) = 1 and C ( n + 1 , k ) = C ( n , k − 1 ) + ( k − ( α + 1 ) n − γ ) C ( n , k ) C(n+1,k)=C(n,k-1)+(k-(\alpha+1)n-\gamma)C(n,k) C ( n + 1 , k ) = C ( n , k − 1 ) + ( k − ( α + 1 ) n − γ ) C ( n , k ) . This is the recurrence S ( n + 1 , k ) = S ( n , k − 1 ) + ( k b − n a + r ) S ( n , k ) S(n+1,k)=S(n,k-1)+(kb-na+r)S(n,k) S ( n + 1 , k ) = S ( n , k − 1 ) + ( k b − na + r ) S ( n , k ) of the numbers S ( n , k ) = S ( n , k ; a , b , r ) S(n,k)=S(n,k;a,b,r) S ( n , k ) = S ( n , k ; a , b , r ) , which follows from (8) after multiplication by z − n a = ( z − r − k b ) + ( k b − n a + r ) z-na=(z-r-kb)+(kb-na+r) z − na = ( z − r − k b ) + ( k b − na + r ) . The second expansion is obtained in the same way from I γ X I δ = I α X + δ I α + 1 I^\gamma X I^\delta=I^\alpha X+\delta I^{\alpha+1} I γ X I δ = I α X + δ I α + 1 and multiplication on the left. If γ , δ ≥ 0 \gamma,\delta\geq0 γ , δ ≥ 0 , only operators of nonnegative order occur, and Lemma 3.1 applies. □
For γ = 0 \gamma=0 γ = 0 or δ = 0 \delta=0 δ = 0 , (5.4) contains (19) and (18) and their expansions in the opposite order. For γ = − β < 0 < δ \gamma=-\beta<0<\delta γ = − β < 0 < δ , it expands ( D β X I δ ) n (D^\beta X I^\delta)^n ( D β X I δ ) n . For δ = − γ = β \delta=-\gamma=\beta δ = − γ = β , it is the Leibniz rule for D β X n I β = ( I − β X I β ) n D^\beta X^nI^\beta=(I^{-\beta}XI^\beta)^n D β X n I β = ( I − β X I β ) n , since S ( n , k ; 1 , 1 , β ) = ( n k ) β ( β − 1 ) ⋯ ( β − n + k + 1 ) S(n,k;1,1,\beta)=\binom{n}{k}\beta(\beta-1)\cdots(\beta-n+k+1) S ( n , k ; 1 , 1 , β ) = ( k n ) β ( β − 1 ) ⋯ ( β − n + k + 1 ) .
A probabilistic formula
The expansion (4.2) has alternating signs, although ( X I α ) n (XI^\alpha)^n ( X I α ) n maps nonnegative functions to nonnegative functions. Its kernel explains this and leads to a probabilistic description of the coefficients. In this section, α > 0 \alpha>0 α > 0 , s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α , and n ≥ 1 n\geq1 n ≥ 1 . Let ( D 1 , … , D n ) (D_1,\ldots,D_n) ( D 1 , … , D n ) have the symmetric Dirichlet distribution with parameter α \alpha α [20 ] , that is, the density Γ ( n α ) Γ ( α ) − n ∏ i = 1 n d i α − 1 \Gamma(n\alpha)\Gamma(\alpha)^{-n}\prod_{i=1}^{n}d_i^{\alpha-1} Γ ( n α ) Γ ( α ) − n ∏ i = 1 n d i α − 1 on the simplex { d i > 0 , d 1 + ⋯ + d n = 1 } \{d_i>0,\ d_1+\cdots+d_n=1\} { d i > 0 , d 1 + ⋯ + d n = 1 } with respect to d d 1 ⋯ d d n − 1 dd_1\cdots dd_{n-1} d d 1 ⋯ d d n − 1 . Put B j = D 1 + ⋯ + D j B_j=D_1+\cdots+D_j B j = D 1 + ⋯ + D j for 1 ≤ j ≤ n − 1 1\leq j\leq n-1 1 ≤ j ≤ n − 1 , let E \mathbb{E} E denote expectation, and let e m e_m e m be the m m m th elementary symmetric polynomial.
Theorem 6.1. For f ∈ L l o c 1 f\in L^1_{\mathrm{loc}} f ∈ L loc 1 and almost every x > 0 x>0 x > 0 ,
( ( X I α ) n f ) ( x ) = ∫ 0 x K n , α ( x , t ) f ( t ) d t , K n , α ( x , t ) = x n ( x − t ) n α − 1 Γ ( n α ) P n , α ( 1 − t / x ) , (27) ((XI^\alpha)^nf)(x)=\int_0^x K_{n,\alpha}(x,t)f(t)\,dt,\qquad K_{n,\alpha}(x,t)=\frac{x^n(x-t)^{n\alpha-1}}{\Gamma(n\alpha)}P_{n,\alpha}(1-t/x), \tag*{(27)} (( X I α ) n f ) ( x ) = ∫ 0 x K n , α ( x , t ) f ( t ) d t , K n , α ( x , t ) = Γ ( n α ) x n ( x − t ) n α − 1 P n , α ( 1 − t / x ) , (27) where
P n , α ( y ) = ∑ m = 0 n − 1 S s ; − α ( n , n − m ) ( n α ) m y m = E ∏ j = 1 n − 1 ( 1 − y B j ) . (28) P_{n,\alpha}(y)=\sum_{m=0}^{n-1}\frac{S_{s;-\alpha}(n,n-m)}{(n\alpha)_m}y^m=\mathbb{E}\prod_{j=1}^{n-1}(1-yB_j). \tag*{(28)} P n , α ( y ) = m = 0 ∑ n − 1 ( n α ) m S s ; − α ( n , n − m ) y m = E j = 1 ∏ n − 1 ( 1 − y B j ) . (28)
Proof. By (15) , the operator X n − m I n α + m X^{n-m}I^{n\alpha+m} X n − m I n α + m has the kernel x n − m ( x − t ) n α + m − 1 / Γ ( n α + m ) x^{n-m}(x-t)^{n\alpha+m-1}/\Gamma(n\alpha+m) x n − m ( x − t ) n α + m − 1 /Γ ( n α + m ) , and Γ ( n α + m ) = ( n α ) m Γ ( n α ) \Gamma(n\alpha+m)=(n\alpha)_m\Gamma(n\alpha) Γ ( n α + m ) = ( n α ) m Γ ( n α ) . Hence (4.2), which holds on L l o c 1 L^1_{\mathrm{loc}} L loc 1 by Theorem 4.4(a), gives (6.1) with the first expression for P n , α P_{n,\alpha} P n , α in (6.2). On the other hand, iterating (15) and using Tonelli’s theorem, we obtain ( ( X I α ) n f ) ( x ) = ∫ 0 x K ~ ( x , t ) f ( t ) d t ((XI^\alpha)^nf)(x)=\int_0^x\widetilde K(x,t)f(t)\,dt (( X I α ) n f ) ( x ) = ∫ 0 x K ( x , t ) f ( t ) d t with
K ~ ( x , t ) = x Γ ( α ) n ∫ t < z n − 1 < ⋯ < z 1 < x ∏ j = 1 n − 1 z j ∏ i = 1 n ( z i − 1 − z i ) α − 1 d z 1 ⋯ d z n − 1 , \widetilde K(x,t)=\frac{x}{\Gamma(\alpha)^n}\int_{t<z_{n-1}<\cdots<z_1<x}\prod_{j=1}^{n-1}z_j\prod_{i=1}^{n}(z_{i-1}-z_i)^{\alpha-1}\,dz_1\cdots dz_{n-1}, K ( x , t ) = Γ ( α ) n x ∫ t < z n − 1 < ⋯ < z 1 < x j = 1 ∏ n − 1 z j i = 1 ∏ n ( z i − 1 − z i ) α − 1 d z 1 ⋯ d z n − 1 ,
where z 0 = x z_0=x z 0 = x and z n = t z_n=t z n = t . Put y = ( x − t ) / x y=(x-t)/x y = ( x − t ) / x and d i = ( z i − 1 − z i ) / ( x − t ) d_i=(z_{i-1}-z_i)/(x-t) d i = ( z i − 1 − z i ) / ( x − t ) . Then d i > 0 d_i>0 d i > 0 , d 1 + ⋯ + d n = 1 d_1+\cdots+d_n=1 d 1 + ⋯ + d n = 1 , z j = x ( 1 − y ( d 1 + ⋯ + d j ) ) z_j=x(1-y(d_1+\cdots+d_j)) z j = x ( 1 − y ( d 1 + ⋯ + d j )) , and d z 1 ⋯ d z n − 1 = ( x − t ) n − 1 d d 1 ⋯ d d n − 1 dz_1\cdots dz_{n-1}=(x-t)^{n-1}dd_1\cdots dd_{n-1} d z 1 ⋯ d z n − 1 = ( x − t ) n − 1 d d 1 ⋯ d d n − 1 , so that
K ~ ( x , t ) = x n ( x − t ) n α − 1 Γ ( n α ) E ∏ j = 1 n − 1 ( 1 − y B j ) . (29) \widetilde K(x,t)=\frac{x^n(x-t)^{n\alpha-1}}{\Gamma(n\alpha)}\mathbb{E}\prod_{j=1}^{n-1}(1-yB_j). \tag*{(29)} K ( x , t ) = Γ ( n α ) x n ( x − t ) n α − 1 E j = 1 ∏ n − 1 ( 1 − y B j ) . (29)
Expanding the product, we see that ( X I α ) n (XI^{\alpha})^{n} ( X I α ) n is a linear combination of the operators X n − m I n α + m X^{n-m}I^{n\alpha+m} X n − m I n α + m , 0 ≤ m ≤ n − 1 0 \le m \le n-1 0 ≤ m ≤ n − 1 . The coefficient of each of these operators is ( n α ) m (n\alpha)_{m} ( n α ) m times the corresponding coefficient of the polynomial E ∏ j ( 1 − y B j ) \mathbb{E}\prod_{j}(1-yB_{j}) E ∏ j ( 1 − y B j ) . By the uniqueness of the coefficients in (4.2) (Corollary 4.2 ), this polynomial is P n , α P_{n,\alpha} P n , α .
Corollary 6.2. (a) For 0 ≤ m ≤ n − 1 0 \le m \le n-1 0 ≤ m ≤ n − 1 ,
S s ; α ( n , n − m ) = ( n α ) m E e m ( B 1 , … , B n − 1 ) . (30) S_{s;\alpha}(n,n-m)=(n\alpha)_{m}\mathbb{E}\,e_{m}(B_{1},\ldots,B_{n-1}). \tag*{(30)} S s ; α ( n , n − m ) = ( n α ) m E e m ( B 1 , … , B n − 1 ) . (30) Equivalently, for every real s > 1 s>1 s > 1 , with Dirichlet parameter 1 / ( s − 1 ) 1/(s-1) 1/ ( s − 1 ) ,
S s ; 1 ( n , n − m ) = E e m ( B 1 , … , B n − 1 ) ∏ i = 0 m − 1 ( n + i ( s − 1 ) ) . (31) S_{s;1}(n,n-m)=\mathbb{E}\,e_{m}(B_{1},\ldots,B_{n-1})\prod_{i=0}^{m-1}(n+i(s-1)). \tag*{(31)} S s ; 1 ( n , n − m ) = E e m ( B 1 , … , B n − 1 ) i = 0 ∏ m − 1 ( n + i ( s − 1 )) . (31) (b) As polynomials, P n , α ( y ) = ( 1 − y ) n − 1 P n , α ( − y / ( 1 − y ) ) P_{n,\alpha}(y)=(1-y)^{n-1}P_{n,\alpha}\left(-y/(1-y)\right) P n , α ( y ) = ( 1 − y ) n − 1 P n , α ( − y / ( 1 − y ) ) . Equivalently, on L l o c 1 L^{1}_{\mathrm{loc}} L loc 1 ,
( X I α ) n = ∑ k = 1 n S s ; α ( n , k ) X I ( α + 1 ) n − k X k − 1 , (32) (XI^{\alpha})^{n}=\sum_{k=1}^{n}S_{s;\alpha}(n,k)XI^{(\alpha+1)n-k}X^{k-1}, \tag*{(32)} ( X I α ) n = k = 1 ∑ n S s ; α ( n , k ) X I ( α + 1 ) n − k X k − 1 , (32) an expansion with positive coefficients.
Proof. (a) Expanding the product in (6.2) gives P n , α ( y ) = ∑ m ( − y ) m E e m ( B 1 , … , B n − 1 ) P_{n,\alpha}(y)=\sum_{m}(-y)^{m}\mathbb{E}\,e_{m}(B_{1},\ldots,B_{n-1}) P n , α ( y ) = ∑ m ( − y ) m E e m ( B 1 , … , B n − 1 ) . Comparing coefficients and using (2.2), we obtain (6.3). Then (6.4) follows from S s ; α ( n , n − m ) = α m S s ; 1 ( n , n − m ) S_{s;\alpha}(n,n-m)=\alpha^{m}S_{s;1}(n,n-m) S s ; α ( n , n − m ) = α m S s ; 1 ( n , n − m ) and α − m ( n α ) m = ∏ i < m ( n + i / α ) \alpha^{-m}(n\alpha)_{m}=\prod_{i<m}(n+i/\alpha) α − m ( n α ) m = ∏ i < m ( n + i / α ) , where 1 / α = s − 1 1/\alpha=s-1 1/ α = s − 1 . (b) The vector ( D n , … , D 1 ) (D_{n},\ldots,D_{1}) ( D n , … , D 1 ) has the same distribution as ( D 1 , … , D n ) (D_{1},\ldots,D_{n}) ( D 1 , … , D n ) , and this reversal replaces ( B 1 , … , B n − 1 ) (B_{1},\ldots,B_{n-1}) ( B 1 , … , B n − 1 ) by ( 1 − B n − 1 , … , 1 − B 1 ) (1-B_{n-1},\ldots,1-B_{1}) ( 1 − B n − 1 , … , 1 − B 1 ) . Hence
E ∏ j = 1 n − 1 ( 1 − y B j ) = E ∏ j = 1 n − 1 ( 1 − y + y B j ) = ( 1 − y ) n − 1 E ∏ j = 1 n − 1 ( 1 + y 1 − y B j ) , \mathbb{E}\prod_{j=1}^{n-1}(1-yB_{j})=\mathbb{E}\prod_{j=1}^{n-1}(1-y+yB_{j})=(1-y)^{n-1}\mathbb{E}\prod_{j=1}^{n-1}\left(1+\frac{y}{1-y}B_{j}\right), E j = 1 ∏ n − 1 ( 1 − y B j ) = E j = 1 ∏ n − 1 ( 1 − y + y B j ) = ( 1 − y ) n − 1 E j = 1 ∏ n − 1 ( 1 + 1 − y y B j ) ,
which is the first statement. With y = ( x − t ) / x y=(x-t)/x y = ( x − t ) / x it reads x n P n , α ( y ) = x t n − 1 P n , α ( − ( x − t ) / t ) x^{n}P_{n,\alpha}(y)=xt^{n-1}P_{n,\alpha}(-(x-t)/t) x n P n , α ( y ) = x t n − 1 P n , α ( − ( x − t ) / t ) , so that, by (6.2) and (2.2),
K n , α ( x , t ) = ∑ m = 0 n − 1 S s ; α ( n , n − m ) x t n − 1 − m ( x − t ) n α + m − 1 Γ ( n α + m ) . K_{n,\alpha}(x,t)=\sum_{m=0}^{n-1}S_{s;\alpha}(n,n-m)\frac{xt^{n-1-m}(x-t)^{n\alpha+m-1}}{\Gamma(n\alpha+m)}. K n , α ( x , t ) = m = 0 ∑ n − 1 S s ; α ( n , n − m ) Γ ( n α + m ) x t n − 1 − m ( x − t ) n α + m − 1 .
The m m m th term is the kernel of X I n α + m X n − 1 − m XI^{n\alpha+m}X^{n-1-m} X I n α + m X n − 1 − m , which gives (6.5) with k = n − m k=n-m k = n − m ; the computation can be reversed.
Identity (6.5), composed with X X X on the right, is (4.1) multiplied by X X X on the left, because X ( I α X ) n = ( X I α ) n X X(I^{\alpha}X)^{n}=(XI^{\alpha})^{n}X X ( I α X ) n = ( X I α ) n X . Thus the reversal symmetry of the Dirichlet distribution corresponds to switching between the two orders (4.1) and (4.2). The kernel also gives simple bounds. Since 0 < B j < 1 0<B_{j}<1 0 < B j < 1 almost surely, we have ( 1 − y ) n − 1 ≤ P n , α ( y ) ≤ 1 (1-y)^{n-1}\le P_{n,\alpha}(y)\le1 ( 1 − y ) n − 1 ≤ P n , α ( y ) ≤ 1 for 0 ≤ y ≤ 1 0\le y\le1 0 ≤ y ≤ 1 . With y = ( x − t ) / x y=(x-t)/x y = ( x − t ) / x , the functions x t n − 1 xt^{n-1} x t n − 1 , x n P n , α ( y ) x^{n}P_{n,\alpha}(y) x n P n , α ( y ) and x n x^{n} x n , multiplied by ( x − t ) n α − 1 / Γ ( n α ) (x-t)^{n\alpha-1}/\Gamma(n\alpha) ( x − t ) n α − 1 /Γ ( n α ) , are the kernels of X I n α X n − 1 XI^{n\alpha}X^{n-1} X I n α X n − 1 , ( X I α ) n (XI^{\alpha})^{n} ( X I α ) n and X n I n α X^{n}I^{n\alpha} X n I n α . Hence, for every nonnegative f ∈ L l o c 1 f\in L^{1}_{\mathrm{loc}} f ∈ L loc 1 , almost everywhere,
X I n α X n − 1 f ≤ ( X I α ) n f ≤ X n I n α f . (33) XI^{n\alpha}X^{n-1}f\le(XI^{\alpha})^{n}f\le X^{n}I^{n\alpha}f. \tag*{(33)} X I n α X n − 1 f ≤ ( X I α ) n f ≤ X n I n α f . (33)
An urn model
In the proof of Theorem 6.1, the Dirichlet density comes from the factors ( z i − 1 − z i ) α − 1 (z_{i-1}-z_{i})^{\alpha-1} ( z i − 1 − z i ) α − 1 of the iterated integral. By de Finetti’s theorem, the Dirichlet distribution also appears in Pólya urns [10 ] , and this leads to a description of the coefficients by urns. For a sequence i = ( i 1 , … , i m ) i=(i_{1},\ldots,i_{m}) i = ( i 1 , … , i m ) of colours i t ∈ { 1 , … , n } i_{t}\in\{1,\ldots,n\} i t ∈ { 1 , … , n } , let c r ( i ) c_{r}(i) c r ( i ) be the number of t t t with i t = r i_{t}=r i t = r , and define
w s ( i ) = ∏ r = 1 n ∏ q = 0 c r ( i ) − 1 ( 1 + q ( s − 1 ) ) . (34) w_{s}(i)=\prod_{r=1}^{n}\prod_{q=0}^{c_{r}(i)-1}(1+q(s-1)). \tag*{(34)} w s ( i ) = r = 1 ∏ n q = 0 ∏ c r ( i ) − 1 ( 1 + q ( s − 1 )) . (34)
Proposition 6.3. Let n ≥ 1 n \ge1 n ≥ 1 and 0 ≤ m ≤ n − 1 0 \le m \le n - 1 0 ≤ m ≤ n − 1 . For every s ∈ R s \in\mathbb{R} s ∈ R ,
S s ; 1 ( n , n − m ) = ∑ 1 ≤ j 1 < ⋯ < j m ≤ n − 1 ∏ t = 1 m ( j t + ( t − 1 ) ( s − 1 ) ) (35) S_{s;1}(n,n-m)=\sum_{1\le j_1<\cdots<j_m\le n-1}\prod_{t=1}^{m}\left(j_t+(t-1)(s-1)\right) \tag*{(35)} S s ; 1 ( n , n − m ) = 1 ≤ j 1 < ⋯ < j m ≤ n − 1 ∑ t = 1 ∏ m ( j t + ( t − 1 ) ( s − 1 ) ) (35) = ∑ 1 ≤ j 1 < ⋯ < j m ≤ n − 1 1 ≤ i 1 ≤ j 1 , … , 1 ≤ i m ≤ j m w s ( i ) . =\sum_{\substack{1\le j_1<\cdots<j_m\le n-1\\1\le i_1\le j_1,\ldots,1\le i_m\le j_m}}w_s(i). = 1 ≤ j 1 < ⋯ < j m ≤ n − 1 1 ≤ i 1 ≤ j 1 , … , 1 ≤ i m ≤ j m ∑ w s ( i ) . Now let s ≥ 1 s \ge1 s ≥ 1 , or s = 1 − 1 / K s=1-1/K s = 1 − 1/ K with K ∈ N K\in\mathbb{N} K ∈ N , and let I 1 , I 2 , … I_1,I_2,\ldots I 1 , I 2 , … be random colours in { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } such that
P ( I t + 1 = r ∣ I 1 , … , I t ) = 1 + ( s − 1 ) c r n + ( s − 1 ) t , (36) \mathbb{P}(I_{t+1}=r\mid I_1,\ldots,I_t)=\frac{1+(s-1)c_r}{n+(s-1)t}, \tag*{(36)} P ( I t + 1 = r ∣ I 1 , … , I t ) = n + ( s − 1 ) t 1 + ( s − 1 ) c r , (36) where c r c_r c r is the number of t ′ ≤ t t' \le t t ′ ≤ t with I t ′ = r I_{t'}=r I t ′ = r . Then
S s ; 1 ( n , n − m ) = ∏ t = 0 m − 1 ( n + t ( s − 1 ) ) ∑ 1 ≤ j 1 < ⋯ < j m ≤ n − 1 P ( I 1 ≤ j 1 , … , I m ≤ j m ) . (37) S_{s;1}(n,n-m)=\prod_{t=0}^{m-1}\left(n+t(s-1)\right)\sum_{1\le j_1<\cdots<j_m\le n-1}\mathbb{P}(I_1\le j_1,\ldots,I_m\le j_m). \tag*{(37)} S s ; 1 ( n , n − m ) = t = 0 ∏ m − 1 ( n + t ( s − 1 ) ) 1 ≤ j 1 < ⋯ < j m ≤ n − 1 ∑ P ( I 1 ≤ j 1 , … , I m ≤ j m ) . (37) For s ≥ 1 s \ge1 s ≥ 1 , (36) describes a Pólya urn. At the start the urn contains one ball of each of the n n n colours. After each draw, the ball is put back together with s − 1 s-1 s − 1 more balls of the same colour (if s − 1 s-1 s − 1 is not an integer, we use weights instead of balls). For s = 1 − 1 / K s=1-1/K s = 1 − 1/ K , (36) reads P ( I t + 1 = r ∣ … ) = ( K − c r ) / ( n K − t ) \mathbb{P}(I_{t+1}=r\mid\ldots)=(K-c_r)/(nK-t) P ( I t + 1 = r ∣ … ) = ( K − c r ) / ( n K − t ) ; that is, the balls are drawn without replacement from an urn with K K K balls of each colour. By Theorem 4.4 , the values s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α correspond to the fractional integrals I α I^\alpha I α , and the values s = 1 − 1 / K s=1-1/K s = 1 − 1/ K correspond to the derivatives D K D^K D K of integer order.
Proof. Put T ( n , m ) = S s ; 1 ( n , n − m ) T(n,m)=S_{s;1}(n,n-m) T ( n , m ) = S s ; 1 ( n , n − m ) . By (5) with h = 1 h=1 h = 1 and k = n + 1 − m k=n+1-m k = n + 1 − m ,
T ( n + 1 , m ) = T ( n , m ) + ( n + ( m − 1 ) ( s − 1 ) ) T ( n , m − 1 ) ( 1 ≤ m ≤ n ) , T(n+1,m)=T(n,m)+\left(n+(m-1)(s-1)\right)T(n,m-1)\qquad(1\le m\le n), T ( n + 1 , m ) = T ( n , m ) + ( n + ( m − 1 ) ( s − 1 ) ) T ( n , m − 1 ) ( 1 ≤ m ≤ n ) , where T ( n , 0 ) = 1 T(n,0)=1 T ( n , 0 ) = 1 , and T ( n , m ) = 0 T(n,m)=0 T ( n , m ) = 0 for m ≥ n ≥ 1 m\ge n\ge1 m ≥ n ≥ 1 . Applying this recurrence repeatedly in n n n gives the first expression in (35) . For the second, fix j j j and sum over i m i_m i m first. Since i 1 , … , i m − 1 ≤ j m − 1 < j m i_1,\ldots,i_{m-1}\le j_{m-1}<j_m i 1 , … , i m − 1 ≤ j m − 1 < j m , the colours r ≤ j m r\le j_m r ≤ j m have the weights 1 + c r ( s − 1 ) 1+c_r(s-1) 1 + c r ( s − 1 ) , where c r c_r c r is the number of t < m t<m t < m with i t = r i_t=r i t = r and ∑ r c r = m − 1 \sum_r c_r=m-1 ∑ r c r = m − 1 . Hence the sum over i m i_m i m multiplies w s ( i 1 , … , i m − 1 ) w_s(i_1,\ldots,i_{m-1}) w s ( i 1 , … , i m − 1 ) by j m + ( m − 1 ) ( s − 1 ) j_m+(m-1)(s-1) j m + ( m − 1 ) ( s − 1 ) , and induction on m m m gives the product.
For s ≥ 1 s\ge1 s ≥ 1 the right side of (36) is nonnegative, and its sum over r r r is 1 because ∑ r c r = t \sum_r c_r=t ∑ r c r = t . For s = 1 − 1 / K s=1-1/K s = 1 − 1/ K it equals ( K − c r ) / ( n K − t ) (K-c_r)/(nK-t) ( K − c r ) / ( n K − t ) , with c r ≤ K c_r\le K c r ≤ K and n K − t > 0 nK-t>0 n K − t > 0 for t < m ≤ n − 1 t<m\le n-1 t < m ≤ n − 1 . In both cases the denominators n + t ( s − 1 ) n+t(s-1) n + t ( s − 1 ) , t < m t<m t < m , are positive. When we multiply the conditional probabilities, the draws of colour r r r give the factors 1 , 1 + ( s − 1 ) , … , 1 + ( c r ( i ) − 1 ) ( s − 1 ) 1,1+(s-1),\ldots,1+(c_r(i)-1)(s-1) 1 , 1 + ( s − 1 ) , … , 1 + ( c r ( i ) − 1 ) ( s − 1 ) , so P ( I 1 = i 1 , … , I m = i m ) = w s ( i ) / ∏ t < m ( n + t ( s − 1 ) ) \mathbb{P}(I_1=i_1,\ldots,I_m=i_m)=w_s(i)/\prod_{t<m}(n+t(s-1)) P ( I 1 = i 1 , … , I m = i m ) = w s ( i ) / ∏ t < m ( n + t ( s − 1 )) . Summing over the i i i with i t ≤ j t i_t\le j_t i t ≤ j t and over j j j , and using (35) , we obtain (37) . □
For s > 1 s>1 s > 1 , (37) and (31) are two forms of the same fact. A Pólya urn is a mixture of independent draws whose colour probabilities form a Dirichlet vector D D D with parameter 1 / ( s − 1 ) 1/(s-1) 1/ ( s − 1 ) [10 ] , and, given D D D , the event { I t ≤ j t for all t } \{I_t\le j_t\text{ for all }t\} { I t ≤ j t for all t } has probability B j 1 ⋯ B j m B_{j_1}\cdots B_{j_m} B j 1 ⋯ B j m . Hence, with α = 1 / ( s − 1 ) \alpha=1/(s-1) α = 1/ ( s − 1 ) ,
P n , α ( y ) = ∑ m = 0 n − 1 ( − y ) m ∑ 1 ≤ j 1 < ⋯ < j m ≤ n − 1 P ( I 1 ≤ j 1 , … , I m ≤ j m ) . P_{n,\alpha}(y)=\sum_{m=0}^{n-1}(-y)^m\sum_{1\le j_1<\cdots<j_m\le n-1}\mathbb{P}(I_1\le j_1,\ldots,I_m\le j_m). P n , α ( y ) = m = 0 ∑ n − 1 ( − y ) m 1 ≤ j 1 < ⋯ < j m ≤ n − 1 ∑ P ( I 1 ≤ j 1 , … , I m ≤ j m ) . so the coefficients of the kernel in Theorem 6.1 are urn probabilities. Also, (31) follows from Proposition 6.3 , independently of Theorem 6.1 . As s ↓ 1 s\downarrow1 s ↓ 1 , the urn adds fewer and fewer balls, so the draws become independent and uniform, and B j B_j B j concentrates at E B j = j / n \mathbb{E}B_j=j/n E B j = j / n . So (31) is a random version of the classical formula in Corollary 6.4(a) below, with the points j / n j/n j / n replaced by the random points B j B_j B j .
Corollary 6.4. Let n ≥ 2 n\ge2 n ≥ 2 and $1\le m\le n-1.
(a) (s = 1 s = 1 s = 1 , independent uniform draws.) S 1 ; 1 ( n , n − m ) = e m ( 1 , 2 , … , n − 1 ) S_{1;1}(n,n-m)=e_m(1,2,\ldots,n-1) S 1 ; 1 ( n , n − m ) = e m ( 1 , 2 , … , n − 1 ) , the unsigned Stirling numbers of the first kind .
S 0 ; 1 ( n , n − m ) = ∑ 1 ≤ j 1 < ⋯ < j m ≤ n − 1 ∏ t = 1 m ( j t − t + 1 ) = h m ( 1 , 2 , … , n − m ) , S_{0;1}(n,n-m)=\sum_{1\leq j_1<\cdots<j_m\leq n-1}\prod_{t=1}^{m}(j_t-t+1)=h_m(1,2,\ldots,n-m), S 0 ; 1 ( n , n − m ) = 1 ≤ j 1 < ⋯ < j m ≤ n − 1 ∑ t = 1 ∏ m ( j t − t + 1 ) = h m ( 1 , 2 , … , n − m ) ,
the Stirling numbers S ( n , n − m ) S(n,n-m) S ( n , n − m ) of the second kind ; here h m h_m h m is the complete homogeneous symmetric polynomial.
(c) (s = 2 s = 2 s = 2 , a Pólya urn adding one ball.) S 2 ; 1 ( n , n − m ) = ∑ j ∑ i ∏ r c r ( i ) ! = c ( n − 1 , m ) S_{2;1}(n,n-m)=\sum_j\sum_i\prod_r c_r(i)!=c(n-1,m) S 2 ; 1 ( n , n − m ) = ∑ j ∑ i ∏ r c r ( i )! = c ( n − 1 , m ) , the Bessel numbers of Corollary 5.2(a) .
(d) As a polynomial in s s s , S s ; 1 ( n , n − m ) S_{s;1}(n,n-m) S s ; 1 ( n , n − m ) has degree m − 1 m-1 m − 1 and leading coefficient ( m − 1 ) ! ( n m + 1 ) (m-1)!\binom{n}{m+1} ( m − 1 )! ( m + 1 n ) ; for example, S s ; 1 ( 5 , 1 ) = ( s + 1 ) ( 2 s + 1 ) ( 3 s + 1 ) S_{s;1}(5,1)=(s+1)(2s+1)(3s+1) S s ; 1 ( 5 , 1 ) = ( s + 1 ) ( 2 s + 1 ) ( 3 s + 1 ) .
Proof. (a) For s = 1 s=1 s = 1 the product in (35) is j 1 ⋯ j m j_1\cdots j_m j 1 ⋯ j m , and (5) with h = 1 h=1 h = 1 is the recurrence S ( n + 1 , k ) = S ( n , k − 1 ) + n S ( n , k ) S(n+1,k)=S(n,k-1)+nS(n,k) S ( n + 1 , k ) = S ( n , k − 1 ) + n S ( n , k ) of the unsigned Stirling numbers of the first kind [17 Section 2] . (b) For s = 0 s=0 s = 0 the product is ∏ t ( j t − t + 1 ) \prod_t(j_t-t+1) ∏ t ( j t − t + 1 ) . In the urn, j t − t + 1 j_t-t+1 j t − t + 1 is the number of colours r ≤ j t r\leq j_t r ≤ j t that have not been drawn before. With l t = j t − t + 1 l_t=j_t-t+1 l t = j t − t + 1 the sum runs over 1 ≤ l 1 ≤ ⋯ ≤ l m ≤ n − m 1\leq l_1\leq\cdots\leq l_m\leq n-m 1 ≤ l 1 ≤ ⋯ ≤ l m ≤ n − m , which is h m ( 1 , … , n − m ) h_m(1,\ldots,n-m) h m ( 1 , … , n − m ) . For s = 0 s=0 s = 0 the numbers are the Stirling numbers of the second kind, as in the proof of Corollary 5.2(b). (c) Here w 2 ( i ) = ∏ r ∏ q < c r ( i ) ( 1 + q ) = ∏ r c r ( i ) ! w_2(i)=\prod_r\prod_{q<c_r(i)}(1+q)=\prod_r c_r(i)! w 2 ( i ) = ∏ r ∏ q < c r ( i ) ( 1 + q ) = ∏ r c r ( i )! , and S 2 ; 1 ( n , n − m ) = c ( n − 1 , m ) S_{2;1}(n,n-m)=c(n-1,m) S 2 ; 1 ( n , n − m ) = c ( n − 1 , m ) by the proof of Corollary 5.2(a). (d) In the first sum of (35) , the term of j j j is j 1 ∏ t = 2 m ( j t + ( t − 1 ) ( s − 1 ) ) j_1\prod_{t=2}^{m}(j_t+(t-1)(s-1)) j 1 ∏ t = 2 m ( j t + ( t − 1 ) ( s − 1 )) , a polynomial in s s s of degree m − 1 m-1 m − 1 with leading coefficient ( m − 1 ) ! j 1 (m-1)!j_1 ( m − 1 )! j 1 . Moreover ∑ j j 1 = ( n m + 1 ) \sum_j j_1=\binom{n}{m+1} ∑ j j 1 = ( m + 1 n ) , because the sets { r − 1 , j 1 , … , j m } \{r-1,j_1,\ldots,j_m\} { r − 1 , j 1 , … , j m } with 1 ≤ r ≤ j 1 1\leq r\leq j_1 1 ≤ r ≤ j 1 are exactly the ( m + 1 ) (m+1) ( m + 1 ) -element subsets of { 0 , 1 , … , n − 1 } \{0,1,\ldots,n-1\} { 0 , 1 , … , n − 1 } . The example follows from (5) , which gives S s ; 1 ( n + 1 , 1 ) = ( 1 + ( n − 1 ) s ) S s ; 1 ( n , 1 ) S_{s;1}(n+1,1)=(1+(n-1)s)S_{s;1}(n,1) S s ; 1 ( n + 1 , 1 ) = ( 1 + ( n − 1 ) s ) S s ; 1 ( n , 1 ) .
Remark 6.5. The same numbers occur in a different urn model. By the identification after (8) and the scaling S ( n , k ; c a , c b , c r ) = c n − k S ( n , k ; a , b , r ) S(n,k;ca,cb,cr)=c^{n-k}S(n,k;a,b,r) S ( n , k ; c a , c b , cr ) = c n − k S ( n , k ; a , b , r ) , which follows from (8) ,
S s ; 1 ( n , k ) = s n − k S ( n , k ; − 1 , − τ , 0 ) , τ = 1 − 1 / s ( s ≠ 0 ) . S_{s;1}(n,k)=s^{n-k}S(n,k;-1,-\tau,0),\qquad\tau=1-1/s\quad(s\ne0). S s ; 1 ( n , k ) = s n − k S ( n , k ; − 1 , − τ , 0 ) , τ = 1 − 1/ s ( s = 0 ) .
Let s ≥ 1 s\geq1 s ≥ 1 , that is, 0 ≤ τ < 1 0\leq\tau<1 0 ≤ τ < 1 , and let θ > − τ \theta>-\tau θ > − τ . Then the numbers on the right give the distribution of the number K n K_n K n of blocks of the Ewens–Pitman partition of { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } with parameters ( τ , θ ) (\tau,\theta) ( τ , θ ) . This random partition is generated by the Chinese restaurant process [23 Sections 3.1–3.2] , and
P ( K n = k ) = ∏ i = 1 k − 1 ( θ + i τ ) ( θ + 1 ) n − 1 S ( n , k ; − 1 , − τ , 0 ) . \mathbb{P}(K_n=k)=\frac{\prod_{i=1}^{k-1}(\theta+i\tau)}{(\theta+1)_{n-1}}S(n,k;-1,-\tau,0). P ( K n = k ) = ( θ + 1 ) n − 1 ∏ i = 1 k − 1 ( θ + i τ ) S ( n , k ; − 1 , − τ , 0 ) .
This model counts the blocks, while (37) is about threshold events for a fixed number n n n of colours. Urn models for other families of generalized Stirling numbers appear in [9 ] .
Remark 6.6. For s = 2 s=2 s = 2 , that is, α = 1 \alpha=1 α = 1 , the Dirichlet distribution is uniform on the simplex, and ( B 1 , … , B n − 1 ) (B_1,\ldots,B_{n-1}) ( B 1 , … , B n − 1 ) has the law of the order statistics of n − 1 n-1 n − 1 independent uniform variables on ( 0 , 1 ) (0,1) ( 0 , 1 ) [6 ] . Since the product in (28) is symmetric in the B j B_j B j , we get P n , 1 ( y ) = ( 1 − y / 2 ) n − 1 P_{n,1}(y)=(1-y/2)^{n-1} P n , 1 ( y ) = ( 1 − y /2 ) n − 1 and K n , 1 ( x , t ) = x ( x 2 − t 2 ) n − 1 / ( 2 n − 1 ( n − 1 ) ! ) K_{n,1}(x,t)=x(x^2-t^2)^{n-1}/(2^{n-1}(n-1)!) K n , 1 ( x , t ) = x ( x 2 − t 2 ) n − 1 / ( 2 n − 1 ( n − 1 )!) . In general P n , α P_{n,\alpha} P n , α is not a power of a linear polynomial; for example, P 3 , α ( y ) = 1 − y + 2 α + 1 3 ( 3 α + 1 ) y 2 P_{3,\alpha}(y)=1-y+\frac{2\alpha+1}{3(3\alpha+1)}y^2 P 3 , α ( y ) = 1 − y + 3 ( 3 α + 1 ) 2 α + 1 y 2 .
Words in X and fractional integrals
The method of Theorem 6.1 works for arbitrary words. Let n ≥ 1 n\geq1 n ≥ 1 , α 1 , … , α n > 0 \alpha_1,\ldots,\alpha_n>0 α 1 , … , α n > 0 , λ 1 , … , λ n ∈ N 0 \lambda_1,\ldots,\lambda_n\in\mathbb{N}_0 λ 1 , … , λ n ∈ N 0 , and
w = X λ n I α n X λ n − 1 I α n − 1 ⋯ X λ 1 I α 1 , A = ∑ i = 1 n α i , Λ = ∑ i = 1 n λ i . w=X^{\lambda_n}I^{\alpha_n}X^{\lambda_{n-1}}I^{\alpha_{n-1}}\cdots X^{\lambda_1}I^{\alpha_1},\qquad A=\sum_{i=1}^{n}\alpha_i,\qquad\Lambda=\sum_{i=1}^{n}\lambda_i. w = X λ n I α n X λ n − 1 I α n − 1 ⋯ X λ 1 I α 1 , A = i = 1 ∑ n α i , Λ = i = 1 ∑ n λ i .
In this subsection, ( D 1 , … , D n ) (D_1,\ldots,D_n) ( D 1 , … , D n ) is a Dirichlet vector with parameters ( α n , … , α 1 ) (\alpha_n,\ldots,\alpha_1) ( α n , … , α 1 ) , so that
E ∏ r = 1 n D r c r = ∏ r = 1 n ( α n + 1 − r ) c r ( A ) c 1 + ⋯ + c n ( c r ∈ N 0 ) \mathbb{E}\prod_{r=1}^{n}D_r^{c_r}=\frac{\prod_{r=1}^{n}(\alpha_{n+1-r})_{c_r}}{(\mathsf{A})_{c_1+\cdots+c_n}}\qquad(c_r\in\mathbb{N}_0) E r = 1 ∏ n D r c r = ( A ) c 1 + ⋯ + c n ∏ r = 1 n ( α n + 1 − r ) c r ( c r ∈ N 0 )
[20 ] , and again B j = D 1 + ⋯ + D j B_j=D_1+\cdots+D_j B j = D 1 + ⋯ + D j . Put
Q w ( y ) = E ∏ j = 1 n − 1 ( 1 − y B j ) λ n − j = ∑ m = 0 Λ − λ n ( − 1 ) m q m y m . (38) Q_w(y)=\mathbb{E}\prod_{j=1}^{n-1}(1-yB_j)^{\lambda_{n-j}}=\sum_{m=0}^{\Lambda-\lambda_n}(-1)^m q_m y^m. \tag*{(38)} Q w ( y ) = E j = 1 ∏ n − 1 ( 1 − y B j ) λ n − j = m = 0 ∑ Λ − λ n ( − 1 ) m q m y m . (38)
Theorem 6.7. In this notation, q m > 0 q_m>0 q m > 0 for 0 ≤ m ≤ Λ − λ n 0\le m\le\Lambda-\lambda_n 0 ≤ m ≤ Λ − λ n , and on L l o c 1 L^1_{\mathrm{loc}} L loc 1
w = ∑ m = 0 Λ − λ n ( − 1 ) m ( A ) m q m X Λ − m I A + m . (39) w=\sum_{m=0}^{\Lambda-\lambda_n}(-1)^m(\mathsf{A})_m q_m X^{\Lambda-m}I^{\mathsf{A}+m}. \tag*{(39)} w = m = 0 ∑ Λ − λ n ( − 1 ) m ( A ) m q m X Λ − m I A + m . (39) Equivalently, for f ∈ L l o c 1 f\in L^1_{\mathrm{loc}} f ∈ L loc 1 and almost every x > 0 x>0 x > 0 ,
( w f ) ( x ) = ∫ 0 x x Λ ( x − t ) A − 1 Γ ( A ) Q w ( x − t x ) f ( t ) d t . (wf)(x)=\int_0^x\frac{x^\Lambda(x-t)^{\mathsf{A}-1}}{\Gamma(\mathsf{A})}Q_w\left(\frac{x-t}{x}\right)f(t)\,\mathrm{d}t. ( w f ) ( x ) = ∫ 0 x Γ ( A ) x Λ ( x − t ) A − 1 Q w ( x x − t ) f ( t ) d t . Explicitly,
q m = ∑ m 1 + ⋯ + m n − 1 = m ∏ j = 1 n − 1 ( λ n − j m j ) E ∏ j = 1 n − 1 B j m j , q_m=\sum_{m_1+\cdots+m_{n-1}=m}\prod_{j=1}^{n-1}\binom{\lambda_{n-j}}{m_j}\mathbb{E}\prod_{j=1}^{n-1}B_j^{m_j}, q m = m 1 + ⋯ + m n − 1 = m ∑ j = 1 ∏ n − 1 ( m j λ n − j ) E j = 1 ∏ n − 1 B j m j , and these mixed moments are finite sums of the moments of the D r D_r D r .
Proof. Iterating (3.1) and using Tonelli’s theorem, we obtain ( w f ) ( x ) = ∫ 0 x K ~ ( x , t ) f ( t ) d t (wf)(x)=\int_0^x\widetilde{K}(x,t)f(t)\,\mathrm{d}t ( w f ) ( x ) = ∫ 0 x K ( x , t ) f ( t ) d t with
K ~ ( x , t ) = x Λ ∫ t < z n − 1 < ⋯ < z 1 < x ∏ j = 1 n − 1 λ j λ n − j ∏ i = 1 n ( z i − 1 − z i ) α n + 1 − i − 1 Γ ( α n + 1 − i ) d z 1 ⋯ d z n − 1 , \widetilde{K}(x,t)=x^\Lambda\int_{t<z_{n-1}<\cdots<z_1<x}\prod_{j=1}^{n-1}\lambda_j^{\lambda_{n-j}}\prod_{i=1}^{n}\frac{(z_{i-1}-z_i)^{\alpha_{n+1-i}-1}}{\Gamma(\alpha_{n+1-i})}\,\mathrm{d}z_1\cdots\mathrm{d}z_{n-1}, K ( x , t ) = x Λ ∫ t < z n − 1 < ⋯ < z 1 < x j = 1 ∏ n − 1 λ j λ n − j i = 1 ∏ n Γ ( α n + 1 − i ) ( z i − 1 − z i ) α n + 1 − i − 1 d z 1 ⋯ d z n − 1 ,
where z 0 = x z_0=x z 0 = x and z n = t z_n=t z n = t . The substitution in the proof of Theorem 6.1 turns this into x Λ ( x − t ) A − 1 Γ ( A ) − 1 Q w ( ( x − t ) / x ) x^\Lambda(x-t)^{\mathsf{A}-1}\Gamma(\mathsf{A})^{-1}Q_w((x-t)/x) x Λ ( x − t ) A − 1 Γ ( A ) − 1 Q w (( x − t ) / x ) . The function x Λ − m ( x − t ) A + m − 1 / Γ ( A ) x^{\Lambda-m}(x-t)^{\mathsf{A}+m-1}/\Gamma(\mathsf{A}) x Λ − m ( x − t ) A + m − 1 /Γ ( A ) is ( A ) m (\mathsf{A})_m ( A ) m times the kernel of X Λ − m I A + m X^{\Lambda-m}I^{\mathsf{A}+m} X Λ − m I A + m , so expanding Q w Q_w Q w gives (6.12). The formula for q m q_m q m follows from the binomial theorem. For m ≤ Λ − λ n = ∑ j = 1 n − 1 λ n − j m\le\Lambda-\lambda_n=\sum_{j=1}^{n-1}\lambda_{n-j} m ≤ Λ − λ n = ∑ j = 1 n − 1 λ n − j the sum that defines q m q_m q m is not empty, and each term is positive because B j > 0 B_j>0 B j > 0 almost surely. □
The urn model also extends to words. Consider a Pólya urn in which colour r r r starts with weight α n + 1 − r \alpha_{n+1-r} α n + 1 − r and gains weight 1 each time it is drawn. By de Finetti’s theorem, given ( D 1 , … , D n ) (D_1,\ldots,D_n) ( D 1 , … , D n ) the draws I 1 , I 2 , … I_1,I_2,\ldots I 1 , I 2 , … are independent with P ( I t = r ) = D r \mathbb{P}(I_t=r)=D_r P ( I t = r ) = D r [10 ] . Hence E B k 1 ⋯ B k M = P ( I 1 ≤ k 1 , … , I M ≤ k M ) \mathbb{E}B_{k_1}\cdots B_{k_M}=\mathbb{P}(I_1\le k_1,\ldots,I_M\le k_M) E B k 1 ⋯ B k M = P ( I 1 ≤ k 1 , … , I M ≤ k M ) for all k 1 , … , k M k_1,\ldots,k_M k 1 , … , k M , and each q m q_m q m is a sum of such probabilities. For α r = α \alpha_r=\alpha α r = α this is the urn (6.9) with s = 1 + 1 / α s=1+1/\alpha s = 1 + 1/ α .
Corollary 6.8. Let λ ∈ N \lambda\in\mathbb{N} λ ∈ N , α > 0 \alpha>0 α > 0 and n ≥ 1 n\ge1 n ≥ 1 , and let the B j B_j B j come from the symmetric Dirichlet distribution with parameter α \alpha α . Define a n , m λ , α a_{n,m}^{\lambda,\alpha} a n , m λ , α as ( n α ) m (n\alpha)_m ( n α ) m times the coefficient of y m y^m y m in E ∏ j = 1 n − 1 ( 1 + y B j ) λ \mathbb{E}\prod_{j=1}^{n-1}(1+yB_j)^\lambda E ∏ j = 1 n − 1 ( 1 + y B j ) λ . Then a n , m λ , α > 0 a_{n,m}^{\lambda,\alpha}>0 a n , m λ , α > 0 for 0 ≤ m ≤ λ ( n − 1 ) 0\le m\le\lambda(n-1) 0 ≤ m ≤ λ ( n − 1 ) , and on L l o c 1 L^1_{\mathrm{loc}} L loc 1
( X λ I α ) n = ∑ m = 0 λ ( n − 1 ) ( − 1 ) m a n , m λ , α X λ n − m I α n + m , (40) (X^\lambda I^\alpha)^n=\sum_{m=0}^{\lambda(n-1)}(-1)^m a_{n,m}^{\lambda,\alpha}X^{\lambda n-m}I^{\alpha n+m}, \tag*{(40)} ( X λ I α ) n = m = 0 ∑ λ ( n − 1 ) ( − 1 ) m a n , m λ , α X λn − m I α n + m , (40) ( I α X λ ) n = ∑ m = 0 λ ( n − 1 ) a n , m λ , α I α n + m X λ n − m . (41) (I^\alpha X^\lambda)^n=\sum_{m=0}^{\lambda(n-1)}a_{n,m}^{\lambda,\alpha}I^{\alpha n+m}X^{\lambda n-m}. \tag*{(41)} ( I α X λ ) n = m = 0 ∑ λ ( n − 1 ) a n , m λ , α I α n + m X λn − m . (41) These numbers satisfy a 1 , 0 λ , α = 1 a_{1,0}^{\lambda,\alpha}=1 a 1 , 0 λ , α = 1 and
a n + 1 , m λ , α = ∑ j ( λ n − j m − j ) ( α ) m − j a n , j λ , α , (42) a_{n+1,m}^{\lambda,\alpha}=\sum_j\binom{\lambda n-j}{m-j}(\alpha)_{m-j}a_{n,j}^{\lambda,\alpha}, \tag*{(42)} a n + 1 , m λ , α = j ∑ ( m − j λn − j ) ( α ) m − j a n , j λ , α , (42)
and a n , m 1 , α = α m S 1 + 1 / α ; 1 ( n , n − m ) a_{n,m}^{1,\alpha}=\alpha^m S_{1+1/\alpha;1}(n,n-m) a n , m 1 , α = α m S 1 + 1/ α ; 1 ( n , n − m ) .
Proof. The expansion (6.13) is Theorem 6.7 with λ i = λ \lambda_i=\lambda λ i = λ and α i = α \alpha_i=\alpha α i = α . By induction on p p p , Lemma 3.1(b) gives
I α X p = ∑ i = 0 p ( − 1 ) i ( p i ) ( α ) i X p − i I α + i . I^\alpha X^p=\sum_{i=0}^{p}(-1)^i\binom{p}{i}(\alpha)_iX^{p-i}I^{\alpha+i}. I α X p = i = 0 ∑ p ( − 1 ) i ( i p ) ( α ) i X p − i I α + i .
Write ( X λ I α ) n + 1 = X λ I α ( X λ I α ) n (X^\lambda I^\alpha)^{n+1}=X^\lambda I^\alpha(X^\lambda I^\alpha)^n ( X λ I α ) n + 1 = X λ I α ( X λ I α ) n , expand each I α X λ n − j I^\alpha X^{\lambda n-j} I α X λn − j in this way, and collect the terms with X λ ( n + 1 ) − m I α ( n + 1 ) + m X^{\lambda(n+1)-m}I^{\alpha(n+1)+m} X λ ( n + 1 ) − m I α ( n + 1 ) + m . This gives (6.15), because these operators are linearly independent (Corollary 4.2). For λ = 1 \lambda=1 λ = 1 , compare (6.13) with (4.2) and use (2.2). Finally, X λ ( I α X λ ) n = ( X λ I α ) n X λ X^\lambda(I^\alpha X^\lambda)^n=(X^\lambda I^\alpha)^nX^\lambda X λ ( I α X λ ) n = ( X λ I α ) n X λ , so the kernel of ( I α X λ ) n (I^\alpha X^\lambda)^n ( I α X λ ) n is ( t / x ) λ (t/x)^\lambda ( t / x ) λ times that of ( X λ I α ) n (X^\lambda I^\alpha)^n ( X λ I α ) n . As in the proof of Corollary 6.2(b), E ∏ j ( 1 − y B j ) λ = ( 1 − y ) λ ( n − 1 ) E ∏ j ( 1 + y ′ B j ) λ \mathbb{E}\prod_j(1-yB_j)^\lambda=(1-y)^{\lambda(n-1)}\mathbb{E}\prod_j(1+y'B_j)^\lambda E ∏ j ( 1 − y B j ) λ = ( 1 − y ) λ ( n − 1 ) E ∏ j ( 1 + y ′ B j ) λ with y ′ = y / ( 1 − y ) = ( x − t ) / t y'=y/(1-y)=(x-t)/t y ′ = y / ( 1 − y ) = ( x − t ) / t . Hence this kernel is t λ n ( x − t ) n α − 1 Γ ( n α ) − 1 E ∏ j ( 1 + y ′ B j ) λ t^{\lambda n}(x-t)^{n\alpha-1}\Gamma(n\alpha)^{-1}\mathbb{E}\prod_j(1+y'B_j)^\lambda t λn ( x − t ) n α − 1 Γ ( n α ) − 1 E ∏ j ( 1 + y ′ B j ) λ , and expanding in powers of y ′ y' y ′ gives (6.14). □
For an integer order α = δ \alpha=\delta α = δ we have ( λ n − j m − j ) ( δ ) m − j = ( λ n − j ) ! ( λ n − m ) ! ( m − j + δ − 1 δ − 1 ) \binom{\lambda n-j}{m-j}(\delta)_{m-j}=\frac{(\lambda n-j)!}{(\lambda n-m)!}\binom{m-j+\delta-1}{\delta-1} ( m − j λn − j ) ( δ ) m − j = ( λn − m )! ( λn − j )! ( δ − 1 m − j + δ − 1 ) , so (6.15) is the recurrence found by Benmoussa for ( x λ I δ ) n (x^\lambda I^\delta)^n ( x λ I δ ) n [2 Section 3] . Theorem 6.7 gives these numbers in closed form, shows that they are positive, and extends them to real orders. For the Volterra operator the closed form is especially simple.
Corollary 6.9. Let λ ∈ N \lambda\in\mathbb{N} λ ∈ N and V = I 1 V=I^1 V = I 1 . On L l o c 1 L^1_{\mathrm{loc}} L loc 1 ,
( X λ V ) n = ∑ m = 0 λ ( n − 1 ) ( − 1 ) m ( n ) m g n , m ( λ ) X λ n − m V n + m , ∑ m g n , m ( λ ) y m = ( ( 1 + y ) λ + 1 − 1 ( λ + 1 ) y ) n − 1 . (X^\lambda V)^n=\sum_{m=0}^{\lambda(n-1)}(-1)^m(n)_m g_{n,m}^{(\lambda)}X^{\lambda n-m}V^{n+m},\qquad\sum_m g_{n,m}^{(\lambda)}y^m=\left(\frac{(1+y)^{\lambda+1}-1}{(\lambda+1)y}\right)^{n-1}. ( X λ V ) n = m = 0 ∑ λ ( n − 1 ) ( − 1 ) m ( n ) m g n , m ( λ ) X λn − m V n + m , m ∑ g n , m ( λ ) y m = ( ( λ + 1 ) y ( 1 + y ) λ + 1 − 1 ) n − 1 .
Proof. For α = 1 \alpha=1 α = 1 , ( B 1 , … , B n − 1 ) (B_1,\ldots,B_{n-1}) ( B 1 , … , B n − 1 ) has the law of the order statistics of n − 1 n-1 n − 1 independent uniform variables U 1 , … , U n − 1 U_1,\ldots,U_{n-1} U 1 , … , U n − 1 on ( 0 , 1 ) (0,1) ( 0 , 1 ) [6 ] . Since ∏ j ( 1 + y B j ) λ \prod_j(1+yB_j)^\lambda ∏ j ( 1 + y B j ) λ is symmetric in the B j B_j B j , its expectation is ( E ( 1 + y U 1 ) λ ) n − 1 (\mathbb{E}(1+yU_1)^\lambda)^{n-1} ( E ( 1 + y U 1 ) λ ) n − 1 , and E ( 1 + y U 1 ) λ = ( ( 1 + y ) λ + 1 − 1 ) / ( ( λ + 1 ) y ) \mathbb{E}(1+yU_1)^\lambda=((1+y)^{\lambda+1}-1)/((\lambda+1)y) E ( 1 + y U 1 ) λ = (( 1 + y ) λ + 1 − 1 ) / (( λ + 1 ) y ) . Now apply Corollary 6.8 with α = 1 \alpha=1 α = 1 . □
For λ = 1 \lambda=1 λ = 1 we get g n , m ( 1 ) = ( n − 1 m ) 2 − m g_{n,m}^{(1)}=\binom{n-1}{m}2^{-m} g n , m ( 1 ) = ( m n − 1 ) 2 − m and ( n ) m g n , m ( 1 ) = c ( n − 1 , m ) (n)_mg_{n,m}^{(1)}=c(n-1,m) ( n ) m g n , m ( 1 ) = c ( n − 1 , m ) , which is Corollary 5.2(a) again.
Concluding remarks
The range of I n β I^{n\beta} I n β in Theorem 4.4(b) is a sufficient domain. It would be interesting to describe all functions for which (4.4) and (4.5) hold, and the correction terms that appear in general. Remark 4.5 suggests that these terms depend on the behaviour of f f f at the origin. The urn interpretation of Proposition 6.3 covers s ≥ 1 s\ge1 s ≥ 1 and s = 1 − 1 / K s=1-1/K s = 1 − 1/ K , that is, fractional integrals of every order and derivatives of integer order. For derivatives of non-integer order, that is, for s < 1 s<1 s < 1 with 1 / ( 1 − s ) ∉ N 1/(1-s)\notin\mathbb{N} 1/ ( 1 − s ) ∈ / N , some weights w s ( i ) w_s(i) w s ( i ) are negative (for s = − 1 s=-1 s = − 1 , a colour drawn twice gives 1 − 2 = − 1 1-2=-1 1 − 2 = − 1 ). A probabilistic interpretation of (6.8) in this range would be welcome. Words that contain fractional derivatives need conditions at the origin, as in Theorem 4.4(b), and the kernel method of Theorem 6.7 does not apply to them directly. Finally, similar questions arise for the Caputo derivative, whose commutator with X X X contains an initial-value term, and for the Hadamard and Erdélyi–Kober operators.
Funding. This work was supported by the National University of Mongolia.