manuscript No.
(will be inserted by the editor)
Mathematics Subject Classification (2020)91G05 · 60G51 · 45J05 · 34E05 · 34M35
Introduction
The analytical study of the ruin probability Ψ ( u ) \Psi(u) Ψ ( u ) of an insurance company that invests its surplus in a risky asset relies on the integro-differential equation (IDE) for the survival function Φ = 1 − Ψ \Phi= 1 - \Psi Φ = 1 − Ψ . The power of this approach depends on the possibility of eliminating the nonlocal term and obtaining an ordinary differential equation (ODE). For exponentially distributed jumps such a reduction is known: in one-sided models it yields a third-order equation for Φ \Phi Φ (of second order for Φ ′ \Phi' Φ ′ ) [7 , 13 , 30 ] , and in the two-sided model, one of fourth order [17 ] . In this paper we carry this reduction through for an arbitrary Erlang distribution Γ ( k , θ ) \Gamma(k,\theta) Γ ( k , θ ) and three types of risk process. For an arbitrary k k k the IDE implies the ODE. The criterion in terms of the residual of the IDE at zero covers the converse inclusion (Proposition 1). For bounded solutions of the ODE it reduces to at most k k k conditions in the non-life insurance model, holds automatically in the annuity model, and requires at most k 2 k_2 k 2 conditions in the mixed model (Remark 3).
Interest in models with investments is due to the fact that they make the tail of the ruin probability heavier: in contrast to the light tail of the classical Cramér–Lundberg model, here Ψ \Psi Ψ acquires a power tail Ψ ( u ) ∼ C u − β \Psi(u) \sim Cu^{-\beta} Ψ ( u ) ∼ C u − β with exponent β = 2 a / σ 2 − 1 \beta= 2a/\sigma^2 - 1 β = 2 a / σ 2 − 1 , which is determined solely by the drift and the volatility of the asset and does not depend on the claim distribution, provided that it has a finite moment of order β \beta β (for heavier claim tails the asymptotics are determined by the claims themselves, see [29 Theorem 1] ). Discovered by Paulsen [24 ] and studied in detail in [7 , 18 ] , this effect is investigated within two complementary approaches. The probabilistic approach uses the Kesten–Goldie implicit renewal theory for the tails of solutions of affine stochastic recursions [4 , 10 , 19 ] and its applications to insurance with investments [6 , 14 , 15 , 16 , 25 , 26 ] : it computes the exponent β \beta β and, under the Cramér-type condition on the logarithm of the price of the risky asset and moment conditions on P P P (both are satisfied for geometric Brownian motion and Erlang-distributed jumps), the exact asymptotics Ψ ( u ) ∼ C u − β \Psi(u) \sim Cu^{-\beta} Ψ ( u ) ∼ C u − β , C > 0 C > 0 C > 0 , for jumps of both signs [3 , 22 , 25 ] , while for a general Lévy return process only two-sided estimates of order u − β u^{-\beta} u − β are known [14 ] . The analytical approach, to which this paper belongs, treats Φ \Phi Φ as a solution of the IDE [7 , 11 , 17 , 30 ] and aims at exact and explicit results. For jump densities of both signs with a rational Laplace transform (an Erlang distribution is among them), Antipov [1 ] eliminated the integral term by a differential operator in the variable u u u and obtained an ODE of order 2 n + 2 2n + 2 2 n + 2 with quadratic coefficients in u u u . Then, applying the Laplace transform and a Tauberian theorem, he derived the asymptotics Ψ ( u ) ∼ C u − β \Psi(u) \sim Cu^{-\beta} Ψ ( u ) ∼ C u − β , C > 0 C > 0 C > 0 , for β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) .
The analytical approach requires smoothness of the survival function: without it, writing the integro-differential equation in the classical sense is unjustified. Existence of a classical C 2 C^2 C 2 -smooth solution under proportional investment was established by Promislow for the annuity model (life insurance) [27 ] under E [ ξ ] < ∞ \mathbb{E}[\xi] < \infty E [ ξ ] < ∞ , and for non-life insurance [29 ] under the minimal moment condition E [ ξ ε ] < ∞ \mathbb{E}[\xi^\varepsilon] < \infty E [ ξ ε ] < ∞ and continuity of the distribution function. This strengthens the result of Grandits [11 ] on the Sobolev regularity W l o c 2 , 1 W_{\mathrm{loc}}^{2,1} W loc 2 , 1 of the solution and the result [2 ] on C 2 C^2 C 2 -smoothness under E [ ξ β ] < ∞ \mathbb{E}[\xi^{\beta}]<\infty E [ ξ β ] < ∞ . Both papers [27 , 29 ] also obtain the asymptotics Ψ ( u ) ∼ C u − β \Psi(u)\sim Cu^{-\beta} Ψ ( u ) ∼ C u − β , C > 0 C>0 C > 0 (for non-life insurance, under E [ ξ β ] < ∞ \mathbb{E}[\xi^{\beta}]<\infty E [ ξ β ] < ∞ ). An alternative approach through viscosity solutions is developed in [28 ] . For exponentially distributed jumps the problem is carried through to an exact solution in terms of doubly confluent Heun functions in [30 ] .
Kabanov and Pukhlyakov [17 ] proved C 2 C^2 C 2 -smoothness and carried out such a reduction for a model with investments and two-sided exponentially distributed jumps combining life and non-life insurance, arriving at a linear fourth-order equation. The first reduction of the IDE to an ODE for Erlang-distributed jumps appeared in the diploma thesis of I. V. Krechin, for annuity payments (see [20 ] ). In this paper, for the sake of completeness, we provide a revised version of Krechin’s reduction for this case. For one-sided models with exponentially distributed jumps, an analogous reduction is contained in [7 ] (non-life insurance) and [13 ] (annuity payments), and its detailed exposition with an exact solution through the doubly confluent Heun equation is given in [30 ] . The method rests on a differential operator with constant coefficients that annihilates the integral term: for exponentially distributed one-sided jumps it is of first order (of second order in the two-sided case), which leads to a third-order (respectively fourth-order) equation whose asymptotics recovers the exponent β = 2 a / σ 2 − 1 \beta=2a/\sigma^2-1 β = 2 a / σ 2 − 1 . The possibility of such annihilation for all distributions with a rational Laplace transform (Erlang, mixtures of exponentials, phase-type distributions) dates back to the work of Gerber and Shiu [8 , 9 ] and was noted in [30 ] . In full generality it was implemented by Antipov [1 ] : for k 1 = k 2 k_1=k_2 k 1 = k 2 his equation (6) coincides with the operator form (25) of model (C) (for arbitrary θ 1 , θ 2 \theta_1,\theta_2 θ 1 , θ 2 ), and the reductions of one-sided models and the case k 1 ≠ k 2 k_1\ne k_2 k 1 = k 2 follow from the same construction. The contribution of this paper with respect to [1 ] is as follows: closed binomial formulas for the coefficients for one-sided models with an arbitrary k k k (for the mixed model with k 1 = k 2 k_1=k_2 k 1 = k 2 they are an explicit specialization of equation (6) of [1 ] to the Erlang family), the operator form for k 1 ≠ k 2 k_1\ne k_2 k 1 = k 2 , an asymptotic analysis by Thomé series for all β > 0 \beta>0 β > 0 (without the restriction β < 1 \beta<1 β < 1 ) with a complete expansion for non-life insurance, a converse criterion, and an analysis of the singular point u = 0 u=0 u = 0 .
The passage to the Erlang distribution F ( k , θ ) F(k,\theta) F ( k , θ ) with an arbitrary k ≥ 1 k\ge1 k ≥ 1 does not reduce to replacing one density with another and creates two new difficulties. First, the annihilating operator is of order k k k , so the order of the resulting equation grows together with k k k . Second, because of the investments the coefficients of the equation depend on the capital and contain the terms u 2 Ψ ′ ′ u^2\Psi'' u 2 Ψ ′′ and u Ψ ′ u\Psi' u Ψ ′ , so that an iterated application of an operator of order k k k via the Leibniz rule produces a large number of cross terms, so a closed form of the coefficients requires exact recurrence identities for the derivatives of the integral operator.
The main results of the paper are as follows.
We construct annihilating differential operators N \mathcal{N} N and N ~ \widetilde{\mathcal{N}} N of order k k k with constant coefficients that reduce the integro-differential equation to a linear
ordinary differential equation of order k + 2 k+2 k + 2 for one-sided models and k 1 + k 2 + 2 k_1+k_2+2 k 1 + k 2 + 2 for the mixed one.
(ii) We write all coefficients of the reduced equations explicitly in closed form in terms of binomial coefficients and the parameters of the model (for the mixed model, in expanded form for k 1 = k 2 k_1=k_2 k 1 = k 2 and through the composition of operators in the general case), which makes the equations suitable for numerical and qualitative analysis for an arbitrary k k k .
(iii) By the method of formal expansions near the irregular singular point u = ∞ u=\infty u = ∞ (Thomé series) we prove that the exponent of the power branch does not depend on the shape parameter k k k and coincides with the exponential case, β = 2 a / σ 2 − 1 \beta=2a/\sigma^2-1 β = 2 a / σ 2 − 1 . The nonvanishing of the connection coefficient of this branch follows from the power-law lower bound for the ruin probability [17 Proposition 4.1] (in the annuity model, from the ODE and the condition Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 ), so that Ψ ( u ) ∼ C k u − β \Psi(u)\sim C_k u^{-\beta} Ψ ( u ) ∼ C k u − β with C k > 0 C_k>0 C k > 0 for all k k k and all β > 0 \beta>0 β > 0 (Proposition 4). For one-sided models this gives a different proof of the asymptotic part of the known results [14 , 27 , 29 ] . For the mixed model the asymptotics Ψ ( u ) ∼ C u − β \Psi(u)\sim C u^{-\beta} Ψ ( u ) ∼ C u − β , C > 0 C>0 C > 0 , are known from the probabilistic approach [3 , 22 , 25 ] , while analytically it was obtained for k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 [17 ] and for β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) , c ≠ 0 c\ne0 c = 0 [1 ] . The new elements are an analytical proof for all k 1 , k 2 , β > 0 k_1,k_2,\beta>0 k 1 , k 2 , β > 0 , c ∈ R c\in\mathbb{R} c ∈ R (apart from smoothness and the condition Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 , only the above lower bound is taken from probabilistic results) and a description of the solution space of the ODE. In all three models we obtain a complete asymptotic expansion up to a multiplicative constant C k C_k C k (for non-life insurance it is known for k = 1 k=1 k = 1 from [30 ] , and for k = 2 k=2 k = 2 we give its coefficients by an explicit recurrence). For k = 1 k=1 k = 1 this agrees with [7 , 30 ] (in the annuity model, with [13 ] ).
(iv) We obtain a converse criterion: a bounded solution of the reduced ODE satisfies the original IDE if and only if the residual of the IDE vanishes as u → 0 + u\to0+ u → 0 + together with derivatives up to order K − 1 K-1 K − 1 , where K K K is the order of the annihilator (Proposition 1). The number of effective conditions is at most k k k in model (A), zero in model (B), and at most k 2 k_2 k 2 in model (C) (Remark 3). For k = 1 k=1 k = 1 in model (A) this is the known relation c Φ ′ ( 0 + ) = α Φ ( 0 + ) c\Phi'(0+)=\alpha\Phi(0+) c Φ ′ ( 0 + ) = α Φ ( 0 + ) [30 ] .
We conclude that the tail exponent is stable under deformation of the jump-shape within the Erlang family: it is fixed by the parameters of the risky asset ( a , σ ) (a,\sigma) ( a , σ ) .
The paper is organized as follows. In Sect. 2 we describe the model and write out integro-differential equations for the survival function. In Sect. 3 we introduce annihilating operators and establish the combinatorial identities underlying them. In Sect. 4 we prove reduction theorems for all three models. In Sect. 5 we analyze formal solutions at the irregular singular point u = ∞ u=\infty u = ∞ and prove the realization of the power tail. § 6 contains the test cases k = 1 k=1 k = 1 and k = 2 k=2 k = 2 , and Sect. 7 concludes with a discussion and open questions.
The model and assumptions
All random objects are defined on a stochastic basis ( Ω , F , F , P ) (\Omega,\mathcal{F},\mathbb{F},\mathbb{P}) ( Ω , F , F , P ) with a filtration F = ( F t ) t ≥ 0 \mathbb{F}=(\mathcal{F}_t)_{t\geq0} F = ( F t ) t ≥ 0 satisfying the usual conditions. It carries the following mutually independent objects: a standard F \mathbb{F} F -adapted Wiener process W = ( W t ) W=(W_t) W = ( W t ) , a Poisson process N = ( N t ) N=(N_t) N = ( N t ) with intensity α > 0 \alpha>0 α > 0 and jump times 0 < T 1 < T 2 < ⋯ 0<T_1<T_2<\cdots 0 < T 1 < T 2 < ⋯ , and an i.i.d. sequence ( ξ i ) i ≥ 1 (\xi_i)_{i\geq1} ( ξ i ) i ≥ 1 of positive random variables. In the mixed model we also have two independent Poisson processes N ( 1 ) , N ( 2 ) N^{(1)},N^{(2)} N ( 1 ) , N ( 2 ) with intensities α 1 , α 2 > 0 \alpha_1,\alpha_2>0 α 1 , α 2 > 0 (if α 1 = 0 \alpha_1=0 α 1 = 0 or α 2 = 0 \alpha_2=0 α 2 = 0 with the corresponding sign of c c c , the model degenerates into (A) or (B)) and two sequences of jumps ξ ( 1 ) , ξ ( 2 ) \xi^{(1)},\xi^{(2)} ξ ( 1 ) , ξ ( 2 ) .
The price of the risky asset S S S is a geometric Brownian motion
d S t = S t ( a d t + σ d W t ) , S 0 > 0 , \mathrm{d}S_t=S_t(a\,\mathrm{d}t+\sigma\,\mathrm{d}W_t),\qquad S_0>0, d S t = S t ( a d t + σ d W t ) , S 0 > 0 ,
with parameters a ∈ R a\in\mathbb{R} a ∈ R , σ > 0 \sigma>0 σ > 0 . The return of the asset is described by the process d R t = d S t / S t = a d t + σ d W t \mathrm{d}R_t=\mathrm{d}S_t/S_t=a\,\mathrm{d}t+\sigma\,\mathrm{d}W_t d R t = d S t / S t = a d t + σ d W t . The surplus of a company investing the whole reserve in S S S satisfies the stochastic equation
X t = u + ∫ 0 t X s d R s + P t , X 0 = u ≥ 0 , X_t=u+\int_0^t X_s\,\mathrm{d}R_s+P_t,\qquad X_0=u\geq0, X t = u + ∫ 0 t X s d R s + P t , X 0 = u ≥ 0 ,
where P = ( P t ) P=(P_t) P = ( P t ) is a risk process independent of R R R . The time of ruin is τ u : = inf { t ≥ 0 : X t u ≤ 0 } \tau^u:=\inf\{t\geq0:X_t^u\leq0\} τ u := inf { t ≥ 0 : X t u ≤ 0 } , the survival function is Φ ( u ) : = P ( τ u = ∞ ) \Phi(u):=\mathbb{P}(\tau^u=\infty) Φ ( u ) := P ( τ u = ∞ ) , and the ruin probability is Ψ ( u ) : = 1 − Φ ( u ) \Psi(u):=1-\Phi(u) Ψ ( u ) := 1 − Φ ( u ) .
The risk process is one of the following three kinds.
(A) Non-life insurance:
P t = c t − ∑ i = 1 N t ξ i , c > 0. (1) P_t=ct-\sum_{i=1}^{N_t}\xi_i,\qquad c>0. \tag*{(1)} P t = c t − i = 1 ∑ N t ξ i , c > 0. (1)
(B) Annuity payments:
P t = c t + ∑ i = 1 N t ξ i , c < 0. (2) P_t=ct+\sum_{i=1}^{N_t}\xi_i,\qquad c<0. \tag*{(2)} P t = c t + i = 1 ∑ N t ξ i , c < 0. (2)
(C) The mixed model:
P t = c t + ∑ i = 1 N t ( 1 ) ξ i ( 1 ) − ∑ i = 1 N t ( 2 ) ξ i ( 2 ) , c ∈ R . (3) P_t=ct+\sum_{i=1}^{N_t^{(1)}}\xi_i^{(1)}-\sum_{i=1}^{N_t^{(2)}}\xi_i^{(2)},\qquad c\in\mathbb{R}. \tag*{(3)} P t = c t + i = 1 ∑ N t ( 1 ) ξ i ( 1 ) − i = 1 ∑ N t ( 2 ) ξ i ( 2 ) , c ∈ R . (3)
where ξ i ( j ) ∼ Γ ( k j , θ j ) \xi_i^{(j)}\sim\Gamma(k_j,\theta_j) ξ i ( j ) ∼ Γ ( k j , θ j ) , j = 1 , 2 j=1,2 j = 1 , 2 , and the two sequences are mutually independent.
We impose the following assumptions.
(A1) (non-degeneracy ) a − 1 2 σ 2 > 0 a-\frac{1}{2}\sigma^2>0 a − 2 1 σ 2 > 0 , equivalently β : = 2 a / σ 2 − 1 > 0 \beta:=2a/\sigma^2-1>0 β := 2 a / σ 2 − 1 > 0 .
(A2) (smoothness ) Φ ∈ C 2 ( ( 0 , ∞ ) ) \Phi\in C^2((0,\infty)) Φ ∈ C 2 (( 0 , ∞ )) .
Existence of a classical solution for models (A) and (B) follows from independent results on non-life and annuity models: [29 Theorem 1] (under E [ ξ ε ] < ∞ \mathbb{E}[\xi^\varepsilon] < \infty E [ ξ ε ] < ∞ and continuity of the distribution function) and [27 Theorem 1] (under E [ ξ ] < ∞ \mathbb{E}[\xi] < \infty E [ ξ ] < ∞ ). For the Erlang distribution both conditions hold.
For the mixed model (C), the smoothness of the survival function follows from [17 Theorem 5.1] , and the validity of the IDE in the classical sense from [17 Proposition 6.1] . The conditions on the jump densities and their derivatives are satisfied for the Erlang distribution. Indeed, the density
f k ( x ) = x k − 1 e − x / θ θ k ( k − 1 ) ! , x ≥ 0 , (4) f_k(x) = \frac{x^{k-1}e^{-x/\theta}}{\theta^k(k-1)!}, \qquad x \ge0, \tag*{(4)} f k ( x ) = θ k ( k − 1 )! x k − 1 e − x / θ , x ≥ 0 , (4)
and its derivatives f k ′ f'_k f k ′ , f k ′ ′ f''_k f k ′′ are of the form “a polynomial in x x x multiplied by e − x / θ e^{-x/\theta} e − x / θ ”, and therefore belong to L 1 ( R + ) L^1(\mathbb{R}_+) L 1 ( R + ) for any k ∈ N k \in\mathbb{N} k ∈ N . The regularity is then bootstrapped within the paper itself, without circularity. By Lemma 2, the highest derivative J k ( k ) J_k^{(k)} J k ( k ) is expressed through the convolutions themselves J 0 = Φ , … , J k J_0 = \Phi,\ldots,J_k J 0 = Φ , … , J k rather than through the derivatives of Φ \Phi Φ . Hence J k Φ , I k Φ ∈ C k + r J_k\Phi,I_k\Phi\in C^{k+r} J k Φ , I k Φ ∈ C k + r whenever Φ ∈ C r \Phi\in C^r Φ ∈ C r , and in particular J k Φ , I k Φ ∈ C k + 2 J_k\Phi,I_k\Phi\in C^{k+2} J k Φ , I k Φ ∈ C k + 2 . On any compact subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , where 1 2 σ 2 u 2 \frac{1}{2}\sigma^2u^2 2 1 σ 2 u 2 is bounded away from zero, the IDE of model (A) gives Φ ′ ′ = ( α Φ − α J k Φ − ( a u + c ) Φ ) / ( 1 2 σ 2 u 2 ) ∈ C 1 \Phi'' = (\alpha\Phi-\alpha J_k\Phi-(au+c)\Phi)/(\frac{1}{2}\sigma^2u^2) \in C^1 Φ ′′ = ( α Φ − α J k Φ − ( a u + c ) Φ ) / ( 2 1 σ 2 u 2 ) ∈ C 1 , whence Φ ∈ C 3 \Phi\in C^3 Φ ∈ C 3 . Repeating, we obtain Φ ∈ C ∞ ( ( 0 , ∞ ) ) \Phi\in C^\infty((0,\infty)) Φ ∈ C ∞ (( 0 , ∞ )) (in model (B) similarly with I k Φ I_k\Phi I k Φ and in model (C) with I k 1 Φ I_{k_1}\Phi I k 1 Φ and J k 2 Φ J_{k_2}\Phi J k 2 Φ , and the boundedness of Φ \Phi Φ guarantees the convergence of the integrals). The argument is local (smoothness is a local property), and therefore does not depend on the growth of the coefficients at infinity. It justifies the k k k -fold differentiation of the equations in Sect. 4. Furthermore, the coefficients of the reduced ODEs of Sect. 4 are analytic and the leading coefficient does not vanish on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and therefore Φ \Phi Φ is real-analytic on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . Smoothness under more general conditions weakening f ′ , f ′ ′ ∈ L 1 f',f'' \in L^1 f ′ , f ′′ ∈ L 1 (for distributions with a less regular density, for example) is the subject of a separate study and is not used in this paper.
(A3) (non-degeneracy of the risk process ) the process P P P is not an increasing process. In model (A) this holds automatically (α > 0 \alpha> 0 α > 0 , downward jumps), in model (B) because c < 0 c < 0 c < 0 , and in model (C) because α 2 > 0 \alpha_2 > 0 α 2 > 0 (downward jumps). With unbounded (Erlang-distributed) downward jumps we have Ψ ( u ) > 0 \Psi(u) > 0 Ψ ( u ) > 0 for all u u u . The case Ψ ≡ 0 \Psi\equiv0 Ψ ≡ 0 is possible only for an increasing P P P , cf. [17 ] .
In what follows we use the following properties of the physical survival function.
(i) Φ \Phi Φ is nondecreasing: under a common noise X t u ′ − X t u = ( u ′ − u ) E t > 0 X_t^{u'}-X_t^u=(u'-u)\mathcal{E}_t>0 X t u ′ − X t u = ( u ′ − u ) E t > 0 for u ′ > u u'>u u ′ > u , where E t = exp ( ( a − 1 2 σ 2 ) t + σ W t ) \mathcal{E}_t=\exp((a-\frac{1}{2}\sigma^2)t+\sigma W_t) E t = exp (( a − 2 1 σ 2 ) t + σ W t ) . Hence τ u ′ ≥ τ u \tau^{u'} \ge\tau^u τ u ′ ≥ τ u , and in particular G : = Φ ′ ≥ 0 G:=\Phi' \ge0 G := Φ ′ ≥ 0 . (ii) Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 under (A1): in models (A), (B) this is part of Theorem 1 in [29 ] and [27 ] , and in model (C) of Lemma 2.1 in [17 ] (in the printed statement of that lemma the condition on the moment of order β ′ ∈ ( 0 , β ∧ 1 ) \beta' \in(0,\beta\wedge1) β ′ ∈ ( 0 , β ∧ 1 ) of the downward claim size reads “< 1 <1 < 1 ”, while only its finiteness is used in the proof). If the monetary unit is rescaled, u , c , ξ ↦ λ u , λ c , λ ξ u,c,\xi\mapsto\lambda u,\lambda c,\lambda\xi u , c , ξ ↦ λ u , λ c , λ ξ , the condition holds for a small λ \lambda λ , and the property Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 is invariant: Φ λ ( λ u ) = Φ ( u ) \Phi_\lambda(\lambda u)=\Phi(u) Φ λ ( λ u ) = Φ ( u ) .
(iii) Φ ( 0 + ) ∈ ( 0 , 1 ) \Phi(0+)\in(0,1) Φ ( 0 + ) ∈ ( 0 , 1 ) in model (A) [29 Theorem 1] , and Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 in model (B) [27 Theorem 1] . In model (C), Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 for c < 0 c<0 c < 0 : before the first jump, X t = E t ( u + c ∫ 0 t E s − 1 d s ) X_t=\mathcal{E}_t\left(u+c\int_0^t\mathcal{E}_s^{-1}\,ds\right) X t = E t ( u + c ∫ 0 t E s − 1 d s ) reaches zero in time of order u / ∣ c ∣ u/|c| u /∣ c ∣ , while the probability of an upward jump before that moment tends to zero as u → 0 u\to0 u → 0 (the argument of [27 Lemma 3] carries over verbatim). For c ≥ 0 c\ge0 c ≥ 0 we have Φ ( 0 + ) > 0 \Phi(0+)>0 Φ ( 0 + ) > 0 (Proposition 5, Remark 7).
(iv) In model (C) with c ≠ 0 c\ne0 c = 0 the derivatives Φ ′ \Phi' Φ ′ , Φ ′ ′ \Phi'' Φ ′′ are bounded on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) [17 Theorem 5.1] (the statement of that theorem contains no proviso about c c c , but the boundedness in the proof is obtained precisely for c ≠ 0 c\ne0 c = 0 , and for c = 0 c=0 c = 0 it is in general false, see Remark 7). For c ≠ 0 c\ne0 c = 0 it also follows from the local structure of solutions: Proposition 5 with the asymptotic existence theorem at zero, and Lemma 6 at infinity. For c = 0 c=0 c = 0 they are only locally bounded on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and as u → 0 + u\to0+ u → 0 + , Φ ′ = O ( 1 + u ν + − 1 ∣ ln u ∣ ) \Phi'=O(1+u^{\nu_+-1}|\ln u|) Φ ′ = O ( 1 + u ν + − 1 ∣ ln u ∣ ) and Φ ′ ′ = O ( 1 + u ν + − 2 ∣ ln u ∣ ) \Phi''=O(1+u^{\nu_+-2}|\ln u|) Φ ′′ = O ( 1 + u ν + − 2 ∣ ln u ∣ ) with the exponent ν + > 0 \nu_+>0 ν + > 0 from Remark 7, which may be smaller than 1 (the logarithmic factor is needed only for an integer ν + \nu_+ ν + ).
Under assumptions (A1)–(A3) the function Φ \Phi Φ satisfies the following IDEs (the derivation is standard, based on the generator of the process, see [17 , 29 ] ).
Model (A): the boundary conditions Φ ( z ) = 0 \Phi(z)=0 Φ ( z ) = 0 for z < 0 z<0 z < 0 and Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 , and
1 2 σ 2 u 2 Φ ′ ′ ( u ) + ( a u + c ) Φ ′ ( u ) + α ∫ 0 u Φ ( u − y ) f k ( y ) d y − α Φ ( u ) = 0 , u > 0. (5) \frac{1}{2}\sigma^2u^2\Phi''(u)+(au+c)\Phi'(u)+\alpha\int_0^u\Phi(u-y)f_k(y)\,dy-\alpha\Phi(u)=0,\qquad u>0. \tag*{(5)} 2 1 σ 2 u 2 Φ ′′ ( u ) + ( a u + c ) Φ ′ ( u ) + α ∫ 0 u Φ ( u − y ) f k ( y ) d y − α Φ ( u ) = 0 , u > 0. (5)
Model (B): the boundary conditions Φ ( 0 ) = 0 \Phi(0)=0 Φ ( 0 ) = 0 and Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 , and
1 2 σ 2 u 2 Φ ′ ′ ( u ) + ( a u + c ) Φ ′ ( u ) + α ∫ 0 ∞ ( Φ ( u + y ) − Φ ( u ) ) f k ( y ) d y = 0. (6) \frac{1}{2}\sigma^2u^2\Phi''(u)+(au+c)\Phi'(u)+\alpha\int_0^\infty\left(\Phi(u+y)-\Phi(u)\right)f_k(y)\,dy=0. \tag*{(6)} 2 1 σ 2 u 2 Φ ′′ ( u ) + ( a u + c ) Φ ′ ( u ) + α ∫ 0 ∞ ( Φ ( u + y ) − Φ ( u ) ) f k ( y ) d y = 0. (6)
Model (C): the boundary conditions Φ ( z ) = 0 \Phi(z)=0 Φ ( z ) = 0 for z < 0 z<0 z < 0 and Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 , the behavior at zero being determined by the sign of the drift (for c > 0 c>0 c > 0 survival from u = 0 u=0 u = 0 is possible, as in model (A), while for c < 0 c<0 c < 0 the drift takes the capital from zero into the negative region and Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 , as in (B)), and
1 2 σ 2 u 2 Φ ′ ′ ( u ) + ( a u + c ) Φ ′ ( u ) + α 1 ∫ 0 ∞ ( Φ ( u + y ) − Φ ( u ) ) f k 1 ( 1 ) ( y ) d y + α 2 ∫ 0 u Φ ( u − y ) f k 2 ( 2 ) ( y ) d y − α 2 Φ ( u ) = 0. (7) \frac{1}{2}\sigma^2u^2\Phi''(u)+(au+c)\Phi'(u)+\alpha_1\int_0^\infty\left(\Phi(u+y)-\Phi(u)\right)f_{k_1}^{(1)}(y)\,dy
+\alpha_2\int_0^u\Phi(u-y)f_{k_2}^{(2)}(y)\,dy-\alpha_2\Phi(u)=0. \tag*{(7)} 2 1 σ 2 u 2 Φ ′′ ( u ) + ( a u + c ) Φ ′ ( u ) + α 1 ∫ 0 ∞ ( Φ ( u + y ) − Φ ( u ) ) f k 1 ( 1 ) ( y ) d y + α 2 ∫ 0 u Φ ( u − y ) f k 2 ( 2 ) ( y ) d y − α 2 Φ ( u ) = 0. (7)
Annihilating operators and basic identities
We introduce the second-order differential operator
L Φ ( u ) : = 1 2 σ 2 u 2 Φ ′ ′ ( u ) + ( a u + c ) Φ ′ ( u ) . (8) \mathcal{L}\Phi(u) := \frac{1}{2}\sigma^2u^2\Phi''(u) + (au+c)\Phi'(u). \tag*{(8)} L Φ ( u ) := 2 1 σ 2 u 2 Φ ′′ ( u ) + ( a u + c ) Φ ′ ( u ) . (8)
For the Erlang density (4) we define two integral operators:
J k Φ ( u ) : = 1 θ k ( k − 1 ) ! ∫ 0 u Φ ( u − y ) y k − 1 e − y / θ d y = e − u / θ θ k ( k − 1 ) ! ∫ 0 u Φ ( z ) ( u − z ) k − 1 e z / θ d z , (9) \begin{aligned}
\mathcal{J}_k\Phi(u) &:= \frac{1}{\theta^k(k-1)!}\int_0^u \Phi(u-y)y^{k-1}e^{-y/\theta}\,\mathrm{d}y \\
&= \frac{e^{-u/\theta}}{\theta^k(k-1)!}\int_0^u \Phi(z)(u-z)^{k-1}e^{z/\theta}\,\mathrm{d}z, \tag*{(9)}
\end{aligned} J k Φ ( u ) := θ k ( k − 1 )! 1 ∫ 0 u Φ ( u − y ) y k − 1 e − y / θ d y = θ k ( k − 1 )! e − u / θ ∫ 0 u Φ ( z ) ( u − z ) k − 1 e z / θ d z , (9)
I k Φ ( u ) : = 1 θ k ( k − 1 ) ! ∫ 0 ∞ Φ ( u + y ) y k − 1 e − y / θ d y = e u / θ θ k ( k − 1 ) ! ∫ u ∞ Φ ( z ) ( z − u ) k − 1 e − z / θ d z . (10) \begin{aligned}
\mathcal{I}_k\Phi(u) &:= \frac{1}{\theta^k(k-1)!}\int_0^\infty\Phi(u+y)y^{k-1}e^{-y/\theta}\,\mathrm{d}y \\
&= \frac{e^{u/\theta}}{\theta^k(k-1)!}\int_u^\infty\Phi(z)(z-u)^{k-1}e^{-z/\theta}\,\mathrm{d}z. \tag*{(10)}
\end{aligned} I k Φ ( u ) := θ k ( k − 1 )! 1 ∫ 0 ∞ Φ ( u + y ) y k − 1 e − y / θ d y = θ k ( k − 1 )! e u / θ ∫ u ∞ Φ ( z ) ( z − u ) k − 1 e − z / θ d z . (10)
Notation: J k ( u ) : = J k Φ ( u ) J_k(u):=\mathcal{J}_k\Phi(u) J k ( u ) := J k Φ ( u ) and I k ( u ) : = I k Φ ( u ) I_k(u):=\mathcal{I}_k\Phi(u) I k ( u ) := I k Φ ( u ) , with the convention J 0 : = Φ J_0:=\Phi J 0 := Φ , I 0 : = Φ I_0:=\Phi I 0 := Φ .
The annihilating differential operators of order k k k with constant coefficients are
N ( f ) : = ∑ n = 0 k ( k n ) θ n f ( n ) , N ~ ( f ) : = ∑ n = 0 k ( − 1 ) n ( k n ) θ n f ( n ) . (11) \mathcal{N}(f):=\sum_{n=0}^{k}\binom{k}{n}\theta^n f^{(n)},\qquad\widetilde{\mathcal{N}}(f):=\sum_{n=0}^{k}(-1)^n\binom{k}{n}\theta^n f^{(n)}. \tag*{(11)} N ( f ) := n = 0 ∑ k ( n k ) θ n f ( n ) , N ( f ) := n = 0 ∑ k ( − 1 ) n ( n k ) θ n f ( n ) . (11)
In symbolic form, N = ( I + θ D ) k \mathcal{N}=(I+\theta D)^k N = ( I + θ D ) k and N ~ = ( I − θ D ) k \widetilde{\mathcal{N}}=(I-\theta D)^k N = ( I − θ D ) k , where D = d / d u D=\mathrm{d}/\mathrm{d}u D = d / d u . In this notation the IDEs (5)–(7) take the compact form
(A): L Φ + α J k − α Φ = 0 , (12) \text{(A):}\quad\mathcal{L}\Phi+\alpha J_k-\alpha\Phi=0, \tag*{(12)} (A): L Φ + α J k − α Φ = 0 , (12)
(B): L Φ + α I k − α Φ = 0 , (13) \text{(B):}\quad\mathcal{L}\Phi+\alpha I_k-\alpha\Phi=0, \tag*{(13)} (B): L Φ + α I k − α Φ = 0 , (13)
(C): L Φ + α 1 I k 1 + α 2 J k 2 − ( α 1 + α 2 ) Φ = 0. (14) \text{(C):}\quad\mathcal{L}\Phi+\alpha_1 I_{k_1}+\alpha_2 J_{k_2}-(\alpha_1+\alpha_2)\Phi=0. \tag*{(14)} (C): L Φ + α 1 I k 1 + α 2 J k 2 − ( α 1 + α 2 ) Φ = 0. (14)
Lemma 1 For u > 0 u>0 u > 0 and k ≥ 1 k\geq1 k ≥ 1 :
J k ′ ( u ) = J k − 1 ( u ) − J k ( u ) θ , I k ′ ( u ) = I k ( u ) − I k − 1 ( u ) θ , (15) J_k'(u)=\frac{J_{k-1}(u)-J_k(u)}{\theta},\qquad I_k'(u)=\frac{I_k(u)-I_{k-1}(u)}{\theta}, \tag*{(15)} J k ′ ( u ) = θ J k − 1 ( u ) − J k ( u ) , I k ′ ( u ) = θ I k ( u ) − I k − 1 ( u ) , (15) with the convention J 0 = I 0 = Φ J_0=I_0=\Phi J 0 = I 0 = Φ .
Proof. We write J k J_k J k in the equivalent representation
J k ( u ) = e − u / θ θ k ( k − 1 ) ! Λ k ( u ) , where Λ k ( u ) = ∫ 0 u Φ ( z ) ( u − z ) k − 1 e z / θ d z . J_k(u)=\frac{e^{-u/\theta}}{\theta^k(k-1)!}\Lambda_k(u),\qquad\text{where}\qquad\Lambda_k(u)=\int_0^u\Phi(z)(u-z)^{k-1}e^{z/\theta}\,\mathrm{d}z. J k ( u ) = θ k ( k − 1 )! e − u / θ Λ k ( u ) , where Λ k ( u ) = ∫ 0 u Φ ( z ) ( u − z ) k − 1 e z / θ d z .
Differentiating, we get
Λ k ′ ( u ) = Φ ( u ) ( u − u ) k − 1 e u / θ + ( k − 1 ) ∫ 0 u Φ ( z ) ( u − z ) k − 2 e z / θ d z . \Lambda_k'(u)=\Phi(u)(u-u)^{k-1}e^{u/\theta}+(k-1)\int_0^u\Phi(z)(u-z)^{k-2}e^{z/\theta}\,\mathrm{d}z. Λ k ′ ( u ) = Φ ( u ) ( u − u ) k − 1 e u / θ + ( k − 1 ) ∫ 0 u Φ ( z ) ( u − z ) k − 2 e z / θ d z .
For k ≥ 2 k \ge2 k ≥ 2 the boundary term Φ ( u ) ( u − u ) k − 1 e u / θ \Phi(u)(u-u)^{k-1}e^{u/\theta} Φ ( u ) ( u − u ) k − 1 e u / θ vanishes, and the remaining integral equals ( k − 1 ) θ k − 1 ( k − 2 ) ! e u / θ J k − 1 ( u ) (k-1)\theta^{k-1}(k-2)!e^{u/\theta}J_{k-1}(u) ( k − 1 ) θ k − 1 ( k − 2 )! e u / θ J k − 1 ( u ) . The case k = 1 k=1 k = 1 is treated separately (bypassing the ambiguity 0 0 0^0 0 0 ): then Λ 1 ( u ) = ∫ 0 u Φ ( z ) e z / θ d z \Lambda_1(u)=\int_0^u\Phi(z)e^{z/\theta}\,dz Λ 1 ( u ) = ∫ 0 u Φ ( z ) e z / θ d z and Λ 1 ′ ( u ) = Φ ( u ) e u / θ \Lambda_1'(u)=\Phi(u)e^{u/\theta} Λ 1 ′ ( u ) = Φ ( u ) e u / θ directly. In both cases
J k ′ ( u ) = − 1 θ J k ( u ) + e − u / θ θ k ( k − 1 ) ! Λ k ′ ( u ) = 1 θ ( J k − 1 − J k ) . J_k'(u)=-\frac{1}{\theta}J_k(u)+\frac{e^{-u/\theta}}{\theta^k(k-1)!}\Lambda_k'(u)=\frac{1}{\theta}(J_{k-1}-J_k). J k ′ ( u ) = − θ 1 J k ( u ) + θ k ( k − 1 )! e − u / θ Λ k ′ ( u ) = θ 1 ( J k − 1 − J k ) .
with the convention J 0 = Φ J_0=\Phi J 0 = Φ .
For I k I_k I k , by (10) , we write
I k ( u ) = e u / θ θ k ( k − 1 ) ! M k ( u ) . I_k(u)=\frac{e^{u/\theta}}{\theta^k(k-1)!}\mathcal{M}_k(u). I k ( u ) = θ k ( k − 1 )! e u / θ M k ( u ) .
where
M k ( u ) = ∫ u ∞ Φ ( z ) ( z − u ) k − 1 e − z / θ d z . \mathcal{M}_k(u)=\int_u^\infty\Phi(z)(z-u)^{k-1}e^{-z/\theta}\,dz. M k ( u ) = ∫ u ∞ Φ ( z ) ( z − u ) k − 1 e − z / θ d z .
The integral converges because Φ \Phi Φ is bounded, and the differentiation under the integral sign is legitimate. For k ≥ 2 k\ge2 k ≥ 2 the boundary term vanishes and
M k ′ ( u ) = − ( k − 1 ) ∫ u ∞ Φ ( z ) ( z − u ) k − 2 e − z / θ d z = − ( k − 1 ) θ k − 1 ( k − 2 ) ! e − u / θ I k − 1 ( u ) , \begin{aligned}
\mathcal{M}_k'(u)=-(k-1)\int_u^\infty\Phi(z)(z-u)^{k-2}e^{-z/\theta}\,dz \\
&=-(k-1)\theta^{k-1}(k-2)!e^{-u/\theta}I_{k-1}(u),
\end{aligned} M k ′ ( u ) = − ( k − 1 ) ∫ u ∞ Φ ( z ) ( z − u ) k − 2 e − z / θ d z = − ( k − 1 ) θ k − 1 ( k − 2 )! e − u / θ I k − 1 ( u ) ,
while for k = 1 k=1 k = 1 , M 1 ′ ( u ) = − Φ ( u ) e − u / θ \mathcal{M}_1'(u)=-\Phi(u)e^{-u/\theta} M 1 ′ ( u ) = − Φ ( u ) e − u / θ . In both cases
I k ′ = 1 θ I k + e u / θ θ k ( k − 1 ) ! M k ′ = 1 θ ( I k − I k − 1 ) . I_k'=\frac{1}{\theta}I_k+\frac{e^{u/\theta}}{\theta^k(k-1)!}\mathcal{M}_k'=\frac{1}{\theta}(I_k-I_{k-1}). I k ′ = θ 1 I k + θ k ( k − 1 )! e u / θ M k ′ = θ 1 ( I k − I k − 1 ) .
Lemma 2 For 0 ≤ n ≤ k 0\le n\le k 0 ≤ n ≤ k :
J k ( n ) = θ − n ∑ m = 0 n ( − 1 ) n + m ( n m ) J k − m , (16) J_k^{(n)}=\theta^{-n}\sum_{m=0}^{n}(-1)^{n+m}\binom{n}{m}J_{k-m}, \tag*{(16)} J k ( n ) = θ − n m = 0 ∑ n ( − 1 ) n + m ( m n ) J k − m , (16) I k ( n ) = θ − n ∑ m = 0 n ( − 1 ) m ( n m ) I k − m . (17) I_k^{(n)}=\theta^{-n}\sum_{m=0}^{n}(-1)^m\binom{n}{m}I_{k-m}. \tag*{(17)} I k ( n ) = θ − n m = 0 ∑ n ( − 1 ) m ( m n ) I k − m . (17)
Proof. Induction on n n n . The base n = 1 n=1 n = 1 is the content of Lemma 1, rewritten as J k ′ = θ − 1 ∑ m = 0 1 ( − 1 ) 1 + m ( 1 m ) J k − m J_k'=\theta^{-1}\sum_{m=0}^{1}(-1)^{1+m}\binom{1}{m}J_{k-m} J k ′ = θ − 1 ∑ m = 0 1 ( − 1 ) 1 + m ( m 1 ) J k − m and similarly for I k I_k I k .
Induction step for J k J_k J k . Differentiating the assumed formula for n − 1 n-1 n − 1 and applying Lemma 1 to each J k − m J_{k-m} J k − m , we obtain
J k ( n ) = θ − n ∑ m = 0 n − 1 ( − 1 ) n − 1 + m ( n − 1 m ) ( J k − m − 1 − J k − m ) . (18) J_k^{(n)}=\theta^{-n}\sum_{m=0}^{n-1}(-1)^{n-1+m}\binom{n-1}{m}(J_{k-m-1}-J_{k-m}). \tag*{(18)} J k ( n ) = θ − n m = 0 ∑ n − 1 ( − 1 ) n − 1 + m ( m n − 1 ) ( J k − m − 1 − J k − m ) . (18)
ends_mid=1 We split the sum into two:
J k ( n ) = θ − n [ ∑ m = 0 n − 1 ( − 1 ) n − 1 + m ( n − 1 m ) J k − m − 1 − ∑ m = 0 n − 1 ( − 1 ) n − 1 + m ( n − 1 m ) J k − m ] J_k^{(n)}=\theta^{-n}\left[\sum_{m=0}^{n-1}(-1)^{n-1+m}\binom{n-1}{m}J_{k-m-1}-\sum_{m=0}^{n-1}(-1)^{n-1+m}\binom{n-1}{m}J_{k-m}\right] J k ( n ) = θ − n [ m = 0 ∑ n − 1 ( − 1 ) n − 1 + m ( m n − 1 ) J k − m − 1 − m = 0 ∑ n − 1 ( − 1 ) n − 1 + m ( m n − 1 ) J k − m ]
In the first sum we change the index m + 1 ↦ m m+1\mapsto m m + 1 ↦ m . Both sums then contain J k − m J_{k-m} J k − m with the factors ( − 1 ) n + m ( n − 1 m − 1 ) (-1)^{n+m}\binom{n-1}{m-1} ( − 1 ) n + m ( m − 1 n − 1 ) and ( − 1 ) n + m ( n − 1 m ) (-1)^{n+m}\binom{n-1}{m} ( − 1 ) n + m ( m n − 1 ) , and the Pascal identity ( n − 1 m − 1 ) + ( n − 1 m ) = ( n m ) \binom{n-1}{m-1}+\binom{n-1}{m}=\binom{n}{m} ( m − 1 n − 1 ) + ( m n − 1 ) = ( m n ) for 1 ≤ m ≤ n − 1 1\leq m\leq n-1 1 ≤ m ≤ n − 1 (the extreme terms m = 0 m=0 m = 0 and m = n m=n m = n enter once each, with ( n 0 ) = ( n n ) = 1 \binom{n}{0}=\binom{n}{n}=1 ( 0 n ) = ( n n ) = 1 ) yields (16). For I k I_k I k the proof is analogous with the replacement of signs: ( − 1 ) n + m ↦ ( − 1 ) m (-1)^{n+m}\mapsto(-1)^m ( − 1 ) n + m ↦ ( − 1 ) m .
Remark 1 Below, we apply the operator N = ( I + θ D ) k \mathcal{N}=(I+\theta D)^k N = ( I + θ D ) k (and its mirror ( I − θ D ) k (I-\theta D)^k ( I − θ D ) k ) to functions of class C k C^k C k on the half-line. To formal series at the irregular singular point u = ∞ u=\infty u = ∞ (Thomé series, Sect. 5) we apply it termwise, and nowhere below do we use the convergence of these series as an established fact.
Lemma 3 For all k ≥ 1 k\geq1 k ≥ 1 :
N J k = Φ , N ~ I k = Φ . (19) \mathcal{N}J_k=\Phi,\qquad\widetilde{\mathcal{N}}I_k=\Phi. \tag*{(19)} N J k = Φ , N I k = Φ. (19) Proof By Lemma 1 we have θ J k ′ + J k = J k − 1 \theta J'_k+J_k=J_{k-1} θ J k ′ + J k = J k − 1 , that is, ( I + θ D ) J k = J k − 1 (I+\theta D)J_k=J_{k-1} ( I + θ D ) J k = J k − 1 for all k ≥ 1 k\geq1 k ≥ 1 (with the convention J 0 = Φ J_0=\Phi J 0 = Φ ). Since N = ( I + θ D ) k \mathcal{N}=(I+\theta D)^k N = ( I + θ D ) k , telescoping gives
N J k = ( I + θ D ) k J k = ( I + θ D ) k − 1 J k − 1 = ⋯ = ( I + θ D ) 0 J 0 = Φ . \mathcal{N}J_k=(I+\theta D)^kJ_k=(I+\theta D)^{k-1}J_{k-1}=\cdots=(I+\theta D)^0J_0=\Phi. N J k = ( I + θ D ) k J k = ( I + θ D ) k − 1 J k − 1 = ⋯ = ( I + θ D ) 0 J 0 = Φ. Similarly, from I k ′ = ( I k − I k − 1 ) / θ I'_k=(I_k-I_{k-1})/\theta I k ′ = ( I k − I k − 1 ) / θ we obtain ( I − θ D ) I k = I k − 1 (I-\theta D)I_k=I_{k-1} ( I − θ D ) I k = I k − 1 , whence
N ~ I k = ( I − θ D ) k I k = I 0 = Φ . \widetilde{\mathcal{N}}I_k=(I-\theta D)^kI_k=I_0=\Phi. N I k = ( I − θ D ) k I k = I 0 = Φ. The same follows by substituting (16) directly into N J k = ∑ n = 0 k ( k n ) θ n J k ( n ) \mathcal{N}J_k=\sum_{n=0}^k\binom{k}{n}\theta^nJ_k^{(n)} N J k = ∑ n = 0 k ( n k ) θ n J k ( n ) (Lemma 2 applies up to n = k n=k n = k , and the term m = k m=k m = k in (16) at n = k n=k n = k is exactly J 0 = Φ J_0=\Phi J 0 = Φ , so that no J − 1 J_{-1} J − 1 is required):
N J k = ∑ m = 0 k J k − m ∑ n = m k ( − 1 ) n + m ( k n ) ( n m ) \mathcal{N}J_k=\sum_{m=0}^kJ_{k-m}\sum_{n=m}^k(-1)^{n+m}\binom{k}{n}\binom{n}{m} N J k = m = 0 ∑ k J k − m n = m ∑ k ( − 1 ) n + m ( n k ) ( m n ) = ∑ m = 0 k J k − m ( k m ) ∑ j = 0 k − m ( − 1 ) j ( k − m j ) = ∑ m = 0 k J k − m ( k m ) δ m , k = Φ , =\sum_{m=0}^kJ_{k-m}\binom{k}{m}\sum_{j=0}^{k-m}(-1)^j\binom{k-m}{j}
=\sum_{m=0}^kJ_{k-m}\binom{k}{m}\delta_{m,k}=\Phi, = m = 0 ∑ k J k − m ( m k ) j = 0 ∑ k − m ( − 1 ) j ( j k − m ) = m = 0 ∑ k J k − m ( m k ) δ m , k = Φ , where we used the identity ( k n ) ( n m ) = ( k m ) ( k − m n − m ) \binom{k}{n}\binom{n}{m}=\binom{k}{m}\binom{k-m}{n-m} ( n k ) ( m n ) = ( m k ) ( n − m k − m ) and ( 1 − 1 ) k − m = δ m , k (1-1)^{k-m}=\delta_{m,k} ( 1 − 1 ) k − m = δ m , k . For N ~ I k \widetilde{\mathcal{N}}I_k N I k the factor ( − 1 ) n (-1)^n ( − 1 ) n of N ~ \widetilde{\mathcal{N}} N together with the factor ( − 1 ) m (-1)^m ( − 1 ) m of (17) gives the same sign ( − 1 ) n + m (-1)^{n+m} ( − 1 ) n + m , and the computation repeats verbatim.
Lemma 4 The operators N k , θ \mathcal{N}_{k,\theta} N k , θ and N ~ k ′ , θ ′ \widetilde{\mathcal{N}}_{k',\theta'} N k ′ , θ ′ commute for arbitrary parameters, being polynomials in the operator D = d / d u D=d/du D = d / d u with constant coefficients.
Proof. We write N k , θ = ∑ n = 0 k a n D n \mathcal{N}_{k,\theta}=\sum_{n=0}^{k}a_nD^n N k , θ = ∑ n = 0 k a n D n and N ~ k ′ , θ ′ = ∑ m = 0 k ′ b m D m \widetilde{\mathcal{N}}_{k',\theta'}=\sum_{m=0}^{k'}b_mD^m N k ′ , θ ′ = ∑ m = 0 k ′ b m D m , where a n = ( k n ) θ n a_n=\binom{k}{n}\theta^n a n = ( n k ) θ n and b m = ( − 1 ) m ( k ′ m ) ( θ ′ ) m b_m=(-1)^m\binom{k'}{m}(\theta')^m b m = ( − 1 ) m ( m k ′ ) ( θ ′ ) m . For any f ∈ C k + k ′ f\in C^{k+k'} f ∈ C k + k ′ ,
N k , θ N ~ k ′ , θ ′ f = ∑ n = 0 k ∑ m = 0 k ′ a n b m D n + m f = N ~ k ′ , θ ′ N k , θ f . \mathcal{N}_{k,\theta}\widetilde{\mathcal{N}}_{k',\theta'}f=\sum_{n=0}^{k}\sum_{m=0}^{k'}a_nb_mD^{n+m}f=\widetilde{\mathcal{N}}_{k',\theta'}\mathcal{N}_{k,\theta}f. N k , θ N k ′ , θ ′ f = n = 0 ∑ k m = 0 ∑ k ′ a n b m D n + m f = N k ′ , θ ′ N k , θ f .
The required smoothness is provided by assumption (A2) and the subsequent bootstrap of regularity, so both orders of applying the operators are legitimate and coincide. □ \square □
Reduction to an ODE for all three models
In this section we establish the main reduction of the IDE to an ODE. Throughout we use the Leibniz rule
( u 2 Φ ′ ′ ) ( n ) = u 2 Φ ( n + 2 ) + 2 n u Φ ( n + 1 ) + n ( n − 1 ) Φ ( n ) , ( u Φ ′ ) ( n ) = u Φ ( n + 1 ) + n Φ ( n ) . (20) (u^2\Phi'')^{(n)}=u^2\Phi^{(n+2)}+2nu\Phi^{(n+1)}+n(n-1)\Phi^{(n)},\qquad(u\Phi')^{(n)}=u\Phi^{(n+1)}+n\Phi^{(n)}. \tag*{(20)} ( u 2 Φ ′′ ) ( n ) = u 2 Φ ( n + 2 ) + 2 n u Φ ( n + 1 ) + n ( n − 1 ) Φ ( n ) , ( u Φ ′ ) ( n ) = u Φ ( n + 1 ) + n Φ ( n ) . (20)
We begin with the non-life insurance model (A).
Theorem 1. Under assumption (A2) the survival function Φ \Phi Φ satisfies on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) a linear homogeneous ODE of order k + 2 k+2 k + 2 :
g k + 2 ( u ) Φ ( k + 2 ) + g k + 1 ( u ) Φ ( k + 1 ) + ∑ n = 2 k g n ( u ) Φ ( n ) + g 1 ( u ) Φ ′ = 0 , (21) g_{k+2}(u)\Phi^{(k+2)}+g_{k+1}(u)\Phi^{(k+1)}+\sum_{n=2}^{k}g_n(u)\Phi^{(n)}+g_1(u)\Phi'=0, \tag*{(21)} g k + 2 ( u ) Φ ( k + 2 ) + g k + 1 ( u ) Φ ( k + 1 ) + n = 2 ∑ k g n ( u ) Φ ( n ) + g 1 ( u ) Φ ′ = 0 , (21) with the coefficients
g k + 2 ( u ) = 1 2 σ 2 θ k u 2 , g_{k+2}(u)=\frac{1}{2}\sigma^2\theta^k u^2, g k + 2 ( u ) = 2 1 σ 2 θ k u 2 , g k + 1 ( u ) = 1 2 σ 2 k θ k − 1 u 2 + θ k ( σ 2 k u + a u + c ) , g_{k+1}(u)=\frac{1}{2}\sigma^2 k\theta^{k-1}u^2+\theta^k(\sigma^2ku+au+c), g k + 1 ( u ) = 2 1 σ 2 k θ k − 1 u 2 + θ k ( σ 2 k u + a u + c ) , g n ( u ) = 1 2 σ 2 ( k n − 2 ) θ n − 2 u 2 + ( k n − 1 ) θ n − 1 ( σ 2 ( n − 1 ) u + a u + c ) + ( k n ) θ n ( 1 2 σ 2 n ( n − 1 ) + a n − α ) , 2 ≤ n ≤ k , g_n(u)=\frac{1}{2}\sigma^2\binom{k}{n-2}\theta^{n-2}u^2+\binom{k}{n-1}\theta^{n-1}\bigl(\sigma^2(n-1)u+au+c\bigr)+\binom{k}{n}\theta^n\left(\frac{1}{2}\sigma^2n(n-1)+an-\alpha\right),\qquad2\leq n\leq k, g n ( u ) = 2 1 σ 2 ( n − 2 k ) θ n − 2 u 2 + ( n − 1 k ) θ n − 1 ( σ 2 ( n − 1 ) u + a u + c ) + ( n k ) θ n ( 2 1 σ 2 n ( n − 1 ) + an − α ) , 2 ≤ n ≤ k , g 1 ( u ) = a u + c + k θ ( a − α ) . g_1(u)=au+c+k\theta(a-\alpha). g 1 ( u ) = a u + c + k θ ( a − α ) . The term in Φ ( 0 ) \Phi^{(0)} Φ ( 0 ) is absent, which ensures the reduction of the order to k + 1 k+1 k + 1 for G : = Φ ′ G:=\Phi' G := Φ ′ .
Proof. Applying N \mathcal{N} N to (12) and using N J k = Φ \mathcal{N}J_k=\Phi N J k = Φ (Lemma 3), we obtain
N L Φ − α N Φ + α Φ = 0. (22) \mathcal{N}\mathcal{L}\Phi-\alpha\mathcal{N}\Phi+\alpha\Phi=0. \tag*{(22)} N L Φ − α N Φ + α Φ = 0. (22)
We expand N L Φ \mathcal{N}\mathcal{L}\Phi N L Φ via (20) :
N L Φ = ∑ n = 0 k ( k n ) θ n [ 1 2 σ 2 u 2 Φ ( n + 2 ) + ( σ 2 n u + a u + c ) Φ ( n + 1 ) + ( 1 2 σ 2 n ( n − 1 ) + a n ) Φ ( n ) ] . (23) \mathcal{N}\mathcal{L}\Phi= \sum_{n=0}^{k} \binom{k}{n}\theta^{n}\left[\frac{1}{2}\sigma^{2}u^{2}\Phi^{(n+2)}+\left(\sigma^{2}nu+au+c\right)\Phi^{(n+1)}+\left(\frac{1}{2}\sigma^{2}n(n-1)+an\right)\Phi^{(n)}\right]. \tag*{(23)} N L Φ = n = 0 ∑ k ( n k ) θ n [ 2 1 σ 2 u 2 Φ ( n + 2 ) + ( σ 2 n u + a u + c ) Φ ( n + 1 ) + ( 2 1 σ 2 n ( n − 1 ) + an ) Φ ( n ) ] . (23)
The term − α N Φ + α Φ -\alpha\mathcal{N}\Phi+\alpha\Phi − α N Φ + α Φ is written as − α ∑ n = 1 k ( k n ) θ n Φ ( n ) -\alpha\sum_{n=1}^{k}\binom{k}{n}\theta^{n}\Phi^{(n)} − α ∑ n = 1 k ( n k ) θ n Φ ( n ) (the constant terms cancel). We collect the coefficients of Φ ( j ) \Phi^{(j)} Φ ( j ) :
Leading derivatives. j = k + 2 j=k+2 j = k + 2 : the contribution comes only from Φ ( n + 2 ) \Phi^{(n+2)} Φ ( n + 2 ) at n = k n=k n = k , which gives 1 2 σ 2 θ k u 2 = g k + 2 \frac{1}{2}\sigma^{2}\theta^{k}u^{2}=g_{k+2} 2 1 σ 2 θ k u 2 = g k + 2 .
j = k + 1 j=k+1 j = k + 1 : the contributions from Φ ( n + 2 ) \Phi^{(n+2)} Φ ( n + 2 ) at n = k − 1 n=k-1 n = k − 1 ( 1 2 σ 2 k θ k − 1 u 2 ) \left(\frac{1}{2}\sigma^{2}k\theta^{k-1}u^{2}\right) ( 2 1 σ 2 k θ k − 1 u 2 ) and from Φ ( n + 1 ) \Phi^{(n+1)} Φ ( n + 1 ) at n = k n=k n = k ( θ k ( σ 2 k u + a u + c ) ) \left(\theta^{k}(\sigma^{2}ku+au+c)\right) ( θ k ( σ 2 k u + a u + c ) ) . The sum is g k + 1 g_{k+1} g k + 1 .
Intermediate derivatives. For 2 ≤ j ≤ k 2\leq j\leq k 2 ≤ j ≤ k we collect the contributions from four sources:
– from Φ ( j ) \Phi^{(j)} Φ ( j ) at n = j n=j n = j : ( k j ) θ j ( 1 2 σ 2 j ( j − 1 ) + a j ) \binom{k}{j}\theta^{j}\left(\frac{1}{2}\sigma^{2}j(j-1)+aj\right) ( j k ) θ j ( 2 1 σ 2 j ( j − 1 ) + aj ) ,
– from Φ ( j ) \Phi^{(j)} Φ ( j ) at n = j − 1 n=j-1 n = j − 1 (through Φ ( n + 1 ) \Phi^{(n+1)} Φ ( n + 1 ) ): ( k j − 1 ) θ j − 1 ( σ 2 ( j − 1 ) u + a u + c ) \binom{k}{j-1}\theta^{j-1}\left(\sigma^{2}(j-1)u+au+c\right) ( j − 1 k ) θ j − 1 ( σ 2 ( j − 1 ) u + a u + c ) ,
– from Φ ( j ) \Phi^{(j)} Φ ( j ) at n = j − 2 n=j-2 n = j − 2 (through Φ ( n + 2 ) \Phi^{(n+2)} Φ ( n + 2 ) ): ( k j − 2 ) θ j − 2 1 2 σ 2 u 2 \binom{k}{j-2}\theta^{j-2}\frac{1}{2}\sigma^{2}u^{2} ( j − 2 k ) θ j − 2 2 1 σ 2 u 2 ,
– from − α N Φ -\alpha\mathcal{N}\Phi − α N Φ : − α ( k j ) θ j -\alpha\binom{k}{j}\theta^{j} − α ( j k ) θ j .
We put these together and obtain g j g_j g j from the statement of the theorem.
The first derivative. j = 1 j=1 j = 1 : Φ ( 1 ) \Phi^{(1)} Φ ( 1 ) arises from Φ ( n + 1 ) \Phi^{(n+1)} Φ ( n + 1 ) at n = 0 n=0 n = 0 (the contribution ( k 0 ) θ 0 ( a u + c ) = a u + c \binom{k}{0}\theta^{0}(au+c)=au+c ( 0 k ) θ 0 ( a u + c ) = a u + c ), from Φ ( n ) \Phi^{(n)} Φ ( n ) at n = 1 n=1 n = 1 (the contribution ( k 1 ) θ 1 ⋅ a = k θ a \binom{k}{1}\theta^{1}\cdot a=k\theta a ( 1 k ) θ 1 ⋅ a = k θ a ), and from − α N -\alpha\mathcal{N} − α N at n = 1 n=1 n = 1 (the contribution − α k θ -\alpha k\theta − α k θ ). In total g 1 = a u + c + k θ ( a − α ) g_1=au+c+k\theta(a-\alpha) g 1 = a u + c + k θ ( a − α ) .
The term in Φ ( 0 ) \Phi^{(0)} Φ ( 0 ) . j = 0 j=0 j = 0 : the coefficient of Φ \Phi Φ is − α ( k 0 ) θ 0 + α = 0 -\alpha\binom{k}{0}\theta^{0}+\alpha=0 − α ( 0 k ) θ 0 + α = 0 , so the term in Φ \Phi Φ is absent, and the ODE reduces: G = Φ ′ G=\Phi' G = Φ ′ satisfies an equation of order k + 1 k+1 k + 1 obtained by the substitution Φ ( j ) = G ( j − 1 ) \Phi^{(j)}=G^{(j-1)} Φ ( j ) = G ( j − 1 ) .
We now turn to the annuity model (B): the premium is negative there, c < 0 c<0 c < 0 , and the jumps are directed upward.
Theorem 2 Under assumption (A2), for c < 0 c<0 c < 0 the function Φ \Phi Φ satisfies on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) a linear homogeneous ODE of order k + 2 k+2 k + 2 :
g ~ k + 2 ( u ) Φ ( k + 2 ) + g ~ k + 1 ( u ) Φ ( k + 1 ) + ∑ m = 2 k g ~ m ( u ) Φ ( m ) + g ~ 1 ( u ) Φ ′ = 0 , (24) \widetilde{g}_{k+2}(u)\Phi^{(k+2)}+\widetilde{g}_{k+1}(u)\Phi^{(k+1)}+\sum_{m=2}^{k}\widetilde{g}_{m}(u)\Phi^{(m)}+\widetilde{g}_{1}(u)\Phi'=0, \tag*{(24)} g k + 2 ( u ) Φ ( k + 2 ) + g k + 1 ( u ) Φ ( k + 1 ) + m = 2 ∑ k g m ( u ) Φ ( m ) + g 1 ( u ) Φ ′ = 0 , (24)
with the coefficients (after an overall multiplication by ( − 1 ) k (-1)^k ( − 1 ) k to normalize the sign of the leading coefficient):
g ~ k + 2 ( u ) = 1 2 σ 2 θ k u 2 , \widetilde{g}_{k+2}(u)=\frac{1}{2}\sigma^2\theta^k u^2, g k + 2 ( u ) = 2 1 σ 2 θ k u 2 ,
g ~ k + 1 ( u ) = θ k ( a u + c ) + 1 2 k θ k − 1 σ 2 u ( 2 θ − u ) , \widetilde{g}_{k+1}(u)=\theta^k(au+c)+\frac{1}{2}k\theta^{k-1}\sigma^2u(2\theta-u), g k + 1 ( u ) = θ k ( a u + c ) + 2 1 k θ k − 1 σ 2 u ( 2 θ − u ) ,
g ~ m ( u ) = 1 2 σ 2 θ m − 2 ( − 1 ) k − m ( ( k m − 2 ) u 2 − 2 ( k m − 1 ) θ ( m − 1 ) u + ( k m ) θ 2 m ( m − 1 ) ) + ( − 1 ) k − m + 1 θ m − 1 a ( ( k m − 1 ) u − ( k m ) θ m ) + ( − 1 ) k − m + 1 θ m − 1 ( ( k m − 1 ) c + ( k m ) θ α ) , 2 ≤ m ≤ k , \begin{aligned}
\widetilde{g}_m(u)={}&\frac{1}{2}\sigma^2\theta^{m-2}(-1)^{k-m}\left(\binom{k}{m-2}u^2-2\binom{k}{m-1}\theta(m-1)u+\binom{k}{m}\theta^2m(m-1)\right)\\
&+(-1)^{k-m+1}\theta^{m-1}a\left(\binom{k}{m-1}u-\binom{k}{m}\theta m\right)\\
&+(-1)^{k-m+1}\theta^{m-1}\left(\binom{k}{m-1}c+\binom{k}{m}\theta\alpha\right),\qquad2\le m\le k,
\end{aligned} g m ( u ) = 2 1 σ 2 θ m − 2 ( − 1 ) k − m ( ( m − 2 k ) u 2 − 2 ( m − 1 k ) θ ( m − 1 ) u + ( m k ) θ 2 m ( m − 1 ) ) + ( − 1 ) k − m + 1 θ m − 1 a ( ( m − 1 k ) u − ( m k ) θ m ) + ( − 1 ) k − m + 1 θ m − 1 ( ( m − 1 k ) c + ( m k ) θ α ) , 2 ≤ m ≤ k ,
g ~ 1 ( u ) = ( − 1 ) k ( a u + c + k θ ( α − a ) ) . \widetilde{g}_1(u)=(-1)^k(au+c+k\theta(\alpha-a)). g 1 ( u ) = ( − 1 ) k ( a u + c + k θ ( α − a )) .
Proof. Applying N ~ \widetilde{\mathcal{N}} N to (13) and using N ~ I k = Φ \widetilde{\mathcal{N}}I_k=\Phi N I k = Φ , we obtain N ~ L Φ − α N ~ Φ + α Φ = 0 \widetilde{\mathcal{N}}L\Phi-\alpha\widetilde{\mathcal{N}}\Phi+\alpha\Phi=0 N L Φ − α N Φ + α Φ = 0 . The expansion differs from the proof of Theorem 1 only by the replacement ( k n ) ↦ ( − 1 ) n ( k n ) \binom{k}{n}\mapsto(-1)^n\binom{k}{n} ( n k ) ↦ ( − 1 ) n ( n k ) in each term. The coefficient of Φ ( j ) \Phi^{(j)} Φ ( j ) for 2 ≤ j ≤ k 2\le j\le k 2 ≤ j ≤ k takes the form
T j = ( − 1 ) j − 2 ( k j − 2 ) θ j − 2 1 2 σ 2 u 2 + ( − 1 ) j − 1 ( k j − 1 ) θ j − 1 ( σ 2 ( j − 1 ) u + a u + c ) + ( − 1 ) j ( k j ) θ j ( 1 2 σ 2 j ( j − 1 ) + a j − α ) . (25) \begin{aligned}
T_j={}&(-1)^{j-2}\binom{k}{j-2}\theta^{j-2}\frac{1}{2}\sigma^2u^2+(-1)^{j-1}\binom{k}{j-1}\theta^{j-1}\left(\sigma^2(j-1)u+au+c\right)\\
&+(-1)^j\binom{k}{j}\theta^j\left(\frac{1}{2}\sigma^2j(j-1)+aj-\alpha\right). \tag*{(25)}
\end{aligned} T j = ( − 1 ) j − 2 ( j − 2 k ) θ j − 2 2 1 σ 2 u 2 + ( − 1 ) j − 1 ( j − 1 k ) θ j − 1 ( σ 2 ( j − 1 ) u + a u + c ) + ( − 1 ) j ( j k ) θ j ( 2 1 σ 2 j ( j − 1 ) + aj − α ) . (25)
Factoring out ( − 1 ) j (-1)^j ( − 1 ) j and multiplying the whole equation by ( − 1 ) k (-1)^k ( − 1 ) k (to normalize the sign of the leading coefficient), we arrive at the formula for g ~ j \widetilde{g}_j g j equivalent to the one above. For j = k + 1 j=k+1 j = k + 1 we similarly obtain
g ~ k + 1 ( u ) = θ k ( a u + c ) + 1 2 k θ k − 1 σ 2 u ( 2 θ − u ) , \widetilde{g}_{k+1}(u)=\theta^k(au+c)+\frac{1}{2}k\theta^{k-1}\sigma^2u(2\theta-u), g k + 1 ( u ) = θ k ( a u + c ) + 2 1 k θ k − 1 σ 2 u ( 2 θ − u ) ,
where the expression k θ k − 1 σ 2 u ( 2 θ − u ) / 2 k\theta^{k-1}\sigma^2u(2\theta-u)/2 k θ k − 1 σ 2 u ( 2 θ − u ) /2 arises by combining two terms with opposite signs: − k θ k − 1 σ 2 u 2 / 2 + θ k σ 2 k u = k θ k − 1 σ 2 u ( 2 θ − u ) / 2 -k\theta^{k-1}\sigma^2u^2/2+\theta^k\sigma^2ku=k\theta^{k-1}\sigma^2u(2\theta-u)/2 − k θ k − 1 σ 2 u 2 /2 + θ k σ 2 k u = k θ k − 1 σ 2 u ( 2 θ − u ) /2 . □ \square □
Finally, we consider the mixed model (C).
Theorem 3 Under assumption (A2) the function Φ \Phi Φ satisfies on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) a linear homogeneous ODE of order k 1 + k 2 + 2 k_1+k_2+2 k 1 + k 2 + 2 , obtained by applying the composition of the operators N ~ k 1 , θ 1 N k 2 , θ 2 \widetilde{\mathcal{N}}_{k_1,\theta_1}\mathcal{N}_{k_2,\theta_2} N k 1 , θ 1 N k 2 , θ 2 to the IDE (14) . In operator form:
N ~ k 1 , θ 1 N k 2 , θ 2 L Φ + α 1 N k 2 , θ 2 Φ + α 2 N ~ k 1 , θ 1 Φ − ( α 1 + α 2 ) N ~ k 1 , θ 1 N k 2 , θ 2 Φ = 0 (26) \widetilde{\mathcal{N}}_{k_1,\theta_1}\mathcal{N}_{k_2,\theta_2}L\Phi+\alpha_1\mathcal{N}_{k_2,\theta_2}\Phi+\alpha_2\widetilde{\mathcal{N}}_{k_1,\theta_1}\Phi-(\alpha_1+\alpha_2)\widetilde{\mathcal{N}}_{k_1,\theta_1}\mathcal{N}_{k_2,\theta_2}\Phi=0 \tag*{(26)} N k 1 , θ 1 N k 2 , θ 2 L Φ + α 1 N k 2 , θ 2 Φ + α 2 N k 1 , θ 1 Φ − ( α 1 + α 2 ) N k 1 , θ 1 N k 2 , θ 2 Φ = 0 (26)
(below we omit the indices of the operators). In the special case k 1 = k 2 = k k_1=k_2=k k 1 = k 2 = k , θ 1 = θ 2 = θ \theta_1=\theta_2=\theta θ 1 = θ 2 = θ the expanded form reads
∑ n , m = 0 k ( − 1 ) n ( k n ) ( k m ) θ n + m [ 1 2 σ 2 u 2 Φ ( ℓ + 2 ) + ( σ 2 ℓ u + a u + c ) Φ ( ℓ + 1 ) + ( 1 2 σ 2 ℓ ( ℓ − 1 ) + a ℓ − ( α 1 + α 2 ) ) Φ ( ℓ ) ] + ∑ n = 0 k ( k n ) θ n ( α 1 + ( − 1 ) n α 2 ) Φ ( n ) = 0 , (27) \sum_{n,m=0}^{k}(-1)^n\binom{k}{n}\binom{k}{m}\theta^{n+m}\left[\frac{1}{2}\sigma^2u^2\Phi^{(\ell+2)}+(\sigma^2\ell u+au+c)\Phi^{(\ell+1)}+\left(\frac{1}{2}\sigma^2\ell(\ell-1)+a\ell-(\alpha_1+\alpha_2)\right)\Phi^{(\ell)}\right]+\sum_{n=0}^{k}\binom{k}{n}\theta^n\left(\alpha_1+(-1)^n\alpha_2\right)\Phi^{(n)}=0, \tag*{(27)} n , m = 0 ∑ k ( − 1 ) n ( n k ) ( m k ) θ n + m [ 2 1 σ 2 u 2 Φ ( ℓ + 2 ) + ( σ 2 ℓ u + a u + c ) Φ ( ℓ + 1 ) + ( 2 1 σ 2 ℓ ( ℓ − 1 ) + a ℓ − ( α 1 + α 2 ) ) Φ ( ℓ ) ] + n = 0 ∑ k ( n k ) θ n ( α 1 + ( − 1 ) n α 2 ) Φ ( n ) = 0 , (27)
where ℓ : = n + m \ell:=n+m ℓ := n + m . For k 1 = k 2 = k k_1=k_2=k k 1 = k 2 = k , θ 1 ≠ θ 2 \theta_1\ne\theta_2 θ 1 = θ 2 it suffices to replace θ n + m \theta^{n+m} θ n + m by θ 1 n θ 2 m \theta_1^n\theta_2^m θ 1 n θ 2 m , and θ n ( α 1 + ( − 1 ) n α 2 ) \theta^n(\alpha_1+(-1)^n\alpha_2) θ n ( α 1 + ( − 1 ) n α 2 ) by α 1 θ 2 n + ( − 1 ) n α 2 θ 1 n \alpha_1\theta_2^n+(-1)^n\alpha_2\theta_1^n α 1 θ 2 n + ( − 1 ) n α 2 θ 1 n (precisely this form is taken by the coefficients of equation (6) in [1 ] ).
Proof. Applying N ~ N \widetilde{\mathcal N}\mathcal N N N to (14) :
N ~ N L Φ + α 1 N ~ N I k 1 + α 2 N ~ N J k 2 − ( α 1 + α 2 ) N ~ N Φ = 0. \widetilde{\mathcal N}\mathcal N\mathcal L\Phi+\alpha_1\widetilde{\mathcal N}\mathcal N I_{k_1}+\alpha_2\widetilde{\mathcal N}\mathcal N J_{k_2}-(\alpha_1+\alpha_2)\widetilde{\mathcal N}\mathcal N\Phi=0. N N L Φ + α 1 N N I k 1 + α 2 N N J k 2 − ( α 1 + α 2 ) N N Φ = 0.
By Lemmas 3 and 4,
N ~ N I k 1 = N N ~ I k 1 = N Φ , N ~ N J k 2 = N ~ Φ , \widetilde{\mathcal N}\mathcal N I_{k_1}=\mathcal N\widetilde{\mathcal N}I_{k_1}=\mathcal N\Phi,\qquad\widetilde{\mathcal N}\mathcal N J_{k_2}=\widetilde{\mathcal N}\Phi, N N I k 1 = N N I k 1 = N Φ , N N J k 2 = N Φ ,
whence (26) follows. In the special case k 1 = k 2 = k k_1=k_2=k k 1 = k 2 = k , θ 1 = θ 2 = θ \theta_1=\theta_2=\theta θ 1 = θ 2 = θ the composition expands explicitly: N ~ N f = ∑ n , m = 0 k ( − 1 ) n ( k n ) ( k m ) θ n + m f ( n + m ) \widetilde{\mathcal N}\mathcal Nf=\sum_{n,m=0}^{k}(-1)^n\binom{k}{n}\binom{k}{m}\theta^{n+m}f^{(n+m)} N N f = ∑ n , m = 0 k ( − 1 ) n ( n k ) ( m k ) θ n + m f ( n + m ) . Applying it to L Φ \mathcal L\Phi L Φ with the use of (20) (with ℓ = n + m \ell=n+m ℓ = n + m in place of n n n ) gives the first double sum in (27) , and the replacement a ℓ ↦ a ℓ − ( α 1 + α 2 ) a\ell\mapsto a\ell-(\alpha_1+\alpha_2) a ℓ ↦ a ℓ − ( α 1 + α 2 ) inside the bracket accounts for the term − ( α 1 + α 2 ) N ~ N Φ -(\alpha_1+\alpha_2)\widetilde{\mathcal N}\mathcal N\Phi − ( α 1 + α 2 ) N N Φ . Finally, α 1 N Φ + α 2 N ~ Φ \alpha_1\mathcal N\Phi+\alpha_2\widetilde{\mathcal N}\Phi α 1 N Φ + α 2 N Φ collapses into the single sum ∑ n ( k n ) θ n ( α 1 + ( − 1 ) n α 2 ) Φ ( n ) \sum_n\binom{k}{n}\theta^n(\alpha_1+(-1)^n\alpha_2)\Phi^{(n)} ∑ n ( n k ) θ n ( α 1 + ( − 1 ) n α 2 ) Φ ( n ) .
In the general case k 1 ≠ k 2 k_1\ne k_2 k 1 = k 2 or θ 1 ≠ θ 2 \theta_1\ne\theta_2 θ 1 = θ 2 the operators N k 2 , θ 2 \mathcal N_{k_2,\theta_2} N k 2 , θ 2 and N ~ k 1 , θ 1 \widetilde{\mathcal N}_{k_1,\theta_1} N k 1 , θ 1 still commute as polynomials in d / d u d/du d / d u , and the order of the resulting ODE is k 1 + k 2 + 2 k_1+k_2+2 k 1 + k 2 + 2 . Note the crossed arrangement of the intensities in (26) : the intensity α 1 \alpha_1 α 1 (upward jumps, the annihilator N ~ k 1 , θ 1 \widetilde{\mathcal N}_{k_1,\theta_1} N k 1 , θ 1 ) multiplies the operator N k 2 , θ 2 \mathcal N_{k_2,\theta_2} N k 2 , θ 2 , and α 2 \alpha_2 α 2 multiplies the operator N ~ k 1 , θ 1 \widetilde{\mathcal N}_{k_1,\theta_1} N k 1 , θ 1 , since N ~ N I k 1 = N ( N ~ I k 1 ) = N Φ \widetilde{\mathcal N}\mathcal N I_{k_1}=\mathcal N(\widetilde{\mathcal N}I_{k_1})=\mathcal N\Phi N N I k 1 = N ( N I k 1 ) = N Φ . The part of the equation without L \mathcal L L , i.e., − ( α 1 + α 2 ) N ~ N Φ + α 1 N Φ + α 2 N ~ Φ -(\alpha_1+\alpha_2)\widetilde{\mathcal N}\mathcal N\Phi+\alpha_1\mathcal N\Phi+\alpha_2\widetilde{\mathcal N}\Phi − ( α 1 + α 2 ) N N Φ + α 1 N Φ + α 2 N Φ , does not contain Φ ( 0 ) \Phi^{(0)} Φ ( 0 ) (the constant terms of all three operators are equal to 1). Hence, as in the one-sided models, equation (26) is an equation of order k 1 + k 2 + 1 k_1+k_2+1 k 1 + k 2 + 1 for G = Φ ′ G=\Phi' G = Φ ′ .
Remark 2. Theorems 1–3 establish that the survival function Φ \Phi Φ satisfies the corresponding ODE. The converse inclusion does not follow automatically: the solution space of a linear ODE of order k + 2 k+2 k + 2 is wider than the original probabilistic problem. The physical solution additionally satisfies conditions consistent with the nature of Φ \Phi Φ :
(i) Φ ( z ) ≡ 0 \Phi(z)\equiv0 Φ ( z ) ≡ 0 for z < 0 z<0 z < 0 ,
(ii) lim u → ∞ Φ ( u ) = 1 \lim_{u\to\infty}\Phi(u)=1 lim u → ∞ Φ ( u ) = 1 (from the non-degeneracy condition (A1)), (iii) Φ ( 0 + ) ∈ ( 0 , 1 ) \Phi(0+) \in(0,1) Φ ( 0 + ) ∈ ( 0 , 1 ) in model (A) and in model (C) for c ≥ 0 c \ge0 c ≥ 0 (the positivity follows from Proposition 5 and Remark 7, and Φ ( 0 + ) < 1 \Phi(0+) < 1 Φ ( 0 + ) < 1 because Ψ ( 0 + ) > 0 \Psi(0+) > 0 Ψ ( 0 + ) > 0 ), and Φ ( 0 + ) = 0 \Phi(0+) = 0 Φ ( 0 + ) = 0 in model (B) and in model (C) for c < 0 c < 0 c < 0 (property (iii) of Sect. 2),
(iv) G = Φ ′ ≥ 0 G = \Phi' \ge0 G = Φ ′ ≥ 0 (the monotonicity of Φ \Phi Φ ),
(v) G ∈ L 1 ( ( 0 , ∞ ) ) G \in L^1((0,\infty)) G ∈ L 1 (( 0 , ∞ )) (the integrability of the density, equivalent to 1 − Φ ( 0 + ) = ∫ 0 ∞ G d u < ∞ 1-\Phi(0+) = \int_0^\infty G\,du < \infty 1 − Φ ( 0 + ) = ∫ 0 ∞ G d u < ∞ ).
Conditions (iv)–(v) narrow the class of admissible solutions but do not by themselves prove uniqueness: in model (A) they do not separate the power branch from the exponentially small modes (all of them are integrable). The following proposition gives an exact criterion for an ODE solution to satisfy the original IDE. For k = 1 k=1 k = 1 it reduces to the condition c Φ ′ ( 0 + ) = α Φ ( 0 + ) c\Phi'(0+) = \alpha\Phi(0+) c Φ ′ ( 0 + ) = α Φ ( 0 + ) which closes the verification theorem in [30 ] .
Proposition 1 (converse criterion) Let Φ ∈ C K + 2 ( ( 0 , ∞ ) ) \Phi\in C^{K+2}((0,\infty)) Φ ∈ C K + 2 (( 0 , ∞ )) be bounded, where K K K is the order of the annihilator (K = k K=k K = k in models (A), (B), and K = k 1 + k 2 K=k_1+k_2 K = k 1 + k 2 in model (C)), and let R \mathcal{R} R be the residual of the IDE:
R : = L Φ + α J k − α Φ in model (A) , R : = L Φ + α I k − α Φ in model (B) , R : = L Φ + α 1 I k 1 + α 2 J k 2 − ( α 1 + α 2 ) Φ in model (C) . \begin{aligned}
\mathcal{R} &:= \mathcal{L}\Phi+\alpha J_k-\alpha\Phi&& \text{in model (A)},\\
\mathcal{R} &:= \mathcal{L}\Phi+\alpha I_k-\alpha\Phi&& \text{in model (B)},\\
\mathcal{R} &:= \mathcal{L}\Phi+\alpha_1 I_{k_1}+\alpha_2 J_{k_2}-(\alpha_1+\alpha_2)\Phi&& \text{in model (C)}.
\end{aligned} R R R := L Φ + α J k − α Φ := L Φ + α I k − α Φ := L Φ + α 1 I k 1 + α 2 J k 2 − ( α 1 + α 2 ) Φ in model (A) , in model (B) , in model (C) . Then Φ \Phi Φ satisfies the reduced ODE (Theorems 1–3) if and only if R \mathcal{R} R lies in the kernel of the annihilator A \mathcal{A} A (A = N \mathcal{A}=\mathcal{N} A = N , N ~ \widetilde{\mathcal{N}} N , N ~ N \widetilde{\mathcal{N}}\mathcal{N} N N respectively), i.e., is a quasipolynomial: in (A) R = e − u / θ Q ( u ) \mathcal{R}=e^{-u/\theta}Q(u) R = e − u / θ Q ( u ) with deg Q ≤ k − 1 \deg Q\le k-1 deg Q ≤ k − 1 , in (B) R = e u / θ Q ( u ) \mathcal{R}=e^{u/\theta}Q(u) R = e u / θ Q ( u ) with deg Q ≤ k − 1 \deg Q\le k-1 deg Q ≤ k − 1 , and in (C) R = e u / θ 1 Q 1 ( u ) + e − u / θ 2 Q 2 ( u ) \mathcal{R}=e^{u/\theta_1}Q_1(u)+e^{-u/\theta_2}Q_2(u) R = e u / θ 1 Q 1 ( u ) + e − u / θ 2 Q 2 ( u ) with deg Q i ≤ k i − 1 \deg Q_i\le k_i-1 deg Q i ≤ k i − 1 . In particular, a solution of the reduced ODE satisfies the IDE if and only if
lim u → 0 + R ( j ) ( u ) = 0 , j = 0 , 1 , … , K − 1 , (28) \lim_{u\to0+}\mathcal{R}^{(j)}(u)=0,\qquad j=0,1,\ldots,K-1, \tag*{(28)} u → 0 + lim R ( j ) ( u ) = 0 , j = 0 , 1 , … , K − 1 , (28) and these limits exist for every bounded solution of class C K + 2 ( ( 0 , ∞ ) ) C^{K+2}((0,\infty)) C K + 2 (( 0 , ∞ )) of the ODE.
Proof. Lemmas 3 and 4 are identities valid for any bounded continuous Φ \Phi Φ , so A R \mathcal{A}\mathcal{R} A R coincides with the left-hand side of the reduced ODE: for example, in model (A) N R = N L Φ + α Φ − α N Φ \mathcal{N}\mathcal{R}=\mathcal{N}\mathcal{L}\Phi+\alpha\Phi-\alpha\mathcal{N}\Phi NR = N L Φ + α Φ − α N Φ , and in model (C) N ~ N R = N ~ N L Φ + α 1 N Φ + α 2 N ~ Φ − ( α 1 + α 2 ) N ~ N Φ \widetilde{\mathcal{N}}\mathcal{N}\mathcal{R}=\widetilde{\mathcal{N}}\mathcal{N}\mathcal{L}\Phi+\alpha_1\mathcal{N}\Phi+\alpha_2\widetilde{\mathcal{N}}\Phi-(\alpha_1+\alpha_2)\widetilde{\mathcal{N}}\mathcal{N}\Phi N NR = N N L Φ + α 1 N Φ + α 2 N Φ − ( α 1 + α 2 ) N N Φ . The ODE is therefore equivalent to A R = 0 \mathcal{A}\mathcal{R}=0 A R = 0 , i.e., R ∈ ker A \mathcal{R}\in\ker\mathcal{A} R ∈ ker A , and the kernel of an operator of order K K K with constant coefficients consists precisely of the indicated quasi-polynomials. Every element of the kernel is an entire function uniquely determined by the values R ( j ) ( 0 ) \mathcal{R}^{(j)}(0) R ( j ) ( 0 ) , j < K j<K j < K (the uniqueness of the solution of the Cauchy problem for A R = 0 \mathcal{A}\mathcal{R}=0 A R = 0 ). Hence R ≡ 0 \mathcal{R}\equiv0 R ≡ 0 if and only if (27) holds. The existence of the limits follows from the fact that on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) the function R \mathcal{R} R coincides with an entire function. □ \square □
Remark 3 Conditions (27) are expressed through Φ \Phi Φ and the convolutions. In model (A) we have J k ( u ) = O ( u k ) J_k(u)=O(u^k) J k ( u ) = O ( u k ) and, by Lemma 2, J k ( j ) ( 0 + ) = 0 J_k^{(j)}(0+)=0 J k ( j ) ( 0 + ) = 0 for j < k j<k j < k , so that (28) takes the form lim u → 0 + [ ( L − α ) Φ ] ( j ) ( u ) = 0 \lim_{u\to0+}[(\mathcal{L}-\alpha)\Phi]^{(j)}(u)=0 lim u → 0 + [( L − α ) Φ ] ( j ) ( u ) = 0 , j < k j<k j < k . For k = 1 k=1 k = 1 and u 2 Φ ′ ′ → 0 u^2\Phi^{\prime\prime}\to0 u 2 Φ ′′ → 0 this is the relation c Φ ′ ( 0 + ) = α Φ ( 0 + ) c\Phi^\prime(0+)=\alpha\Phi(0+) c Φ ′ ( 0 + ) = α Φ ( 0 + ) , which coincides with the condition of equivalence of the ODE and the IDE in [30 ] . In models (B), (C) the conditions involve the values I k ( j ) ( 0 + ) I_k^{(j)}(0+) I k ( j ) ( 0 + ) , i.e., functionals of Φ \Phi Φ on the whole half-line. The number of effective conditions, however, does not exceed K K K and depends on the model. For a bounded solution of the ODE the residual R \mathcal{R} R is bounded as u → ∞ u\to\infty u → ∞ : by Lemma 6 and the proof of Proposition 4 in Sect. 5 (the exclusion of the growing modes), such a solution is a combination of a constant, the power branch, and exponentially decaying modes with termwise differentiable asymptotics, so that L Φ → 0 \mathcal{L}\Phi\to0 L Φ → 0 and the convolutions are bounded. The growing quasi-polynomials e u / θ Q ( u ) e^{u/\theta}Q(u) e u / θ Q ( u ) in (B) and e u / θ 1 Q 1 ( u ) e^{u/\theta_1}Q_1(u) e u / θ 1 Q 1 ( u ) in (C) cannot be bounded. In model (B), then, every bounded solution of the ODE satisfies the IDE (all k k k conditions hold automatically). In model (C), conditions (28) reduce to at most k 2 k_2 k 2 conditions Q 2 ≡ 0 Q_2\equiv0 Q 2 ≡ 0 (the coefficients of the polynomial Q 2 Q_2 Q 2 ), and in model (A) to at most k k k conditions Q ≡ 0 Q\equiv0 Q ≡ 0 . The independence of these conditions is not used below. In model (A) all k k k conditions are independent: solutions bounded at zero have asymptotic series ∑ n ≥ 0 b n u n \sum_{n\geq0}b_nu^n ∑ n ≥ 0 b n u n there with free b 0 , … , b k b_0,\ldots,b_k b 0 , … , b k , because the equation of Theorem 1 turns out to have at zero the same structure as equation (26) in Proposition 5 (the leading coefficient 1 2 σ 2 θ k u 2 \frac{1}{2}\sigma^2\theta^ku^2 2 1 σ 2 θ k u 2 , the coefficient of Φ ( k + 1 ) \Phi^{(k+1)} Φ ( k + 1 ) equal to c θ k ≠ 0 c\theta^k\ne0 c θ k = 0 at zero, the indicial equation ν ( ν − 1 ) ⋯ ( ν − k ) = 0 \nu(\nu-1)\cdots(\nu-k)=0 ν ( ν − 1 ) ⋯ ( ν − k ) = 0 , and the mode of the form (47) excluded for c > 0 c>0 c > 0 by boundedness), so that the proof together with the asymptotic existence theorem at zero carries over verbatim. Here [ ( L − α ) Φ ] ( j ) ( 0 + ) = j ! [ c ( j + 1 ) b j + 1 + ( 1 2 σ 2 j ( j − 1 ) + a j − α ) b j ] [(\mathcal{L}-\alpha)\Phi]^{(j)}(0+)=j![c(j+1)b_{j+1}+(\frac{1}{2}\sigma^2j(j-1)+aj-\alpha)b_j] [( L − α ) Φ ] ( j ) ( 0 + ) = j ! [ c ( j + 1 ) b j + 1 + ( 2 1 σ 2 j ( j − 1 ) + aj − α ) b j ] , so that the conditions form a triangular system in b 1 , … , b k b_1,\ldots,b_k b 1 , … , b k with a nonzero diagonal c ( j + 1 ) ! c(j+1)! c ( j + 1 )! . In model (C) the conditions are independent in the numerically tested examples. The dimension K K K of the kernel of the annihilator agrees with the structure of the kernel in [17 Sect. 7] : for k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 it is shown there that the kernel of the ODE operator is four-dimensional, with a basis consisting of the constant, the function Φ \Phi Φ , and two solutions that do not satisfy the IDE, while the kernel of the annihilator (the operator T T T in the notation of [17 ] ) is two-dimensional, in agreement with K = 2 K=2 K = 2 .
Asymptotic analysis
The case k = 2 k=2 k = 2 : formal power branch
For k = 2 k=2 k = 2 the ODE for G : = Φ ′ G:=\Phi^\prime G := Φ ′ is of the third order:
1 2 σ 2 θ 2 u 2 G ′ ′ ′ + g 3 ( u ) G ′ ′ + g 2 ( u ) G ′ + g 1 ( u ) G = 0 , (29) \frac{1}{2}\sigma^2\theta^2u^2G^{\prime\prime\prime}+g_3(u)G^{\prime\prime}+g_2(u)G^\prime+g_1(u)G=0, \tag*{(29)} 2 1 σ 2 θ 2 u 2 G ′′′ + g 3 ( u ) G ′′ + g 2 ( u ) G ′ + g 1 ( u ) G = 0 , (29)
with
g 3 ( u ) = σ 2 θ u 2 + θ 2 ( 2 σ 2 u + a u + c ) , g_3(u)=\sigma^2\theta u^2+\theta^2(2\sigma^2u+au+c), g 3 ( u ) = σ 2 θ u 2 + θ 2 ( 2 σ 2 u + a u + c ) ,
g 2 ( u ) = 1 2 σ 2 u 2 + 2 θ ( σ 2 u + a u + c ) + θ 2 ( σ 2 + 2 a − α ) , g_2(u)=\frac{1}{2}\sigma^2u^2+2\theta(\sigma^2u+au+c)+\theta^2(\sigma^2+2a-\alpha), g 2 ( u ) = 2 1 σ 2 u 2 + 2 θ ( σ 2 u + a u + c ) + θ 2 ( σ 2 + 2 a − α ) ,
g 1 ( u ) = 2 θ a + a u + c − 2 θ α . g_1(u)=2\theta a+au+c-2\theta\alpha. g 1 ( u ) = 2 θ a + a u + c − 2 θ α .
In normalized form (G ′ ′ ′ + q 3 G ′ ′ + q 2 G ′ + q 1 G = 0 G^{\prime\prime\prime}+q_3G^{\prime\prime}+q_2G^\prime+q_1G=0 G ′′′ + q 3 G ′′ + q 2 G ′ + q 1 G = 0 , where g 4 : = g k + 2 ∣ k = 2 = 1 2 σ 2 θ 2 u 2 g_4:=g_{k+2}|_{k=2}=\frac{1}{2}\sigma^2\theta^2u^2 g 4 := g k + 2 ∣ k = 2 = 2 1 σ 2 θ 2 u 2 ), the expansions of the coefficients in powers of 1 / u 1/u 1/ u have the explicit form:
q 3 ( u ) = g 3 ( u ) g 4 ( u ) = 2 θ + 2 ( a + 2 σ 2 ) σ 2 1 u + 2 c σ 2 1 u 2 , (30) q_3(u)=\frac{g_3(u)}{g_4(u)}=\frac{2}{\theta}+\frac{2(a+2\sigma^2)}{\sigma^2}\frac{1}{u}+\frac{2c}{\sigma^2}\frac{1}{u^2}, \tag*{(30)} q 3 ( u ) = g 4 ( u ) g 3 ( u ) = θ 2 + σ 2 2 ( a + 2 σ 2 ) u 1 + σ 2 2 c u 2 1 , (30)
q 2 ( u ) = g 2 ( u ) g 4 ( u ) = 1 θ 2 + 4 ( σ 2 + a ) σ 2 θ 1 u + 2 ( 2 c + θ ( σ 2 + 2 a − α ) ) θ σ 2 1 u 2 , (31) q_2(u)=\frac{g_2(u)}{g_4(u)}=\frac{1}{\theta^2}+\frac{4(\sigma^2+a)}{\sigma^2\theta}\frac{1}{u}+\frac{2(2c+\theta(\sigma^2+2a-\alpha))}{\theta\sigma^2}\frac{1}{u^2}, \tag*{(31)} q 2 ( u ) = g 4 ( u ) g 2 ( u ) = θ 2 1 + σ 2 θ 4 ( σ 2 + a ) u 1 + θ σ 2 2 ( 2 c + θ ( σ 2 + 2 a − α )) u 2 1 , (31)
q 1 ( u ) = g 1 ( u ) g 4 ( u ) = 2 a σ 2 θ 2 1 u + 2 ( 2 θ a + c − 2 θ α ) σ 2 θ 2 1 u 2 . (32) q_1(u)=\frac{g_1(u)}{g_4(u)}=\frac{2a}{\sigma^2\theta^2}\frac{1}{u}+\frac{2(2\theta a+c-2\theta\alpha)}{\sigma^2\theta^2}\frac{1}{u^2}. \tag*{(32)} q 1 ( u ) = g 4 ( u ) g 1 ( u ) = σ 2 θ 2 2 a u 1 + σ 2 θ 2 2 ( 2 θ a + c − 2 θ α ) u 2 1 . (32)
For brevity, we write
q 3 ( u ) = 2 θ + c 31 1 u + c 32 1 u 2 , q 2 ( u ) = 1 θ 2 + c 21 1 u + c 22 1 u 2 , q 1 ( u ) = c 11 1 u + c 12 1 u 2 . q_3(u)=\frac{2}{\theta}+c_{31}\frac{1}{u}+c_{32}\frac{1}{u^2},\qquad q_2(u)=\frac{1}{\theta^2}+c_{21}\frac{1}{u}+c_{22}\frac{1}{u^2},\qquad q_1(u)=c_{11}\frac{1}{u}+c_{12}\frac{1}{u^2}. q 3 ( u ) = θ 2 + c 31 u 1 + c 32 u 2 1 , q 2 ( u ) = θ 2 1 + c 21 u 1 + c 22 u 2 1 , q 1 ( u ) = c 11 u 1 + c 12 u 2 1 .
Theorem 4 (formal power branch, k = 2 k=2 k = 2 ). Let k = 2 k=2 k = 2 (model (A)). (29) admits a unique formal solution of power type (a Thomé series)
G f ( u ) = u − p ∑ j ≥ 0 a j u − j , a 0 = 1 , p = β + 1 = 2 a σ 2 > 1 , (33) G_f(u)=u^{-p}\sum_{j\geq0}a_ju^{-j},\qquad a_0=1,\qquad p=\beta+1=\frac{2a}{\sigma^2}>1, \tag*{(33)} G f ( u ) = u − p j ≥ 0 ∑ a j u − j , a 0 = 1 , p = β + 1 = σ 2 2 a > 1 , (33) where the exponent p p p is determined by the dominant balance, and the coefficients ( a j ) j ≥ 1 (a_j)_{j\geq1} ( a j ) j ≥ 1 are uniquely determined by the recurrence
0 = a j ( c 11 − p + j θ 2 ) + a j − 1 ( c 12 − c 21 ( p + j − 1 ) + 2 θ ( p + j − 1 ) ( p + j − 2 ) + 4 θ ( p + j − 1 ) ) + a j − 2 ( − c 22 ( p + j − 2 ) + 2 c 31 ( p + j − 2 ) − 6 ( p + j − 2 ) − 6 ( p + j − 2 ) ( p + j − 3 ) + c 31 ( p + j − 2 ) ( p + j − 3 ) − ( p + j − 2 ) ( p + j − 3 ) ( p + j − 4 ) ) + a j − 3 ( 2 c 32 ( p + j − 3 ) + c 32 ( p + j − 3 ) ( p + j − 4 ) ) , (34) \begin{aligned}
0={}&a_j\left(c_{11}-\frac{p+j}{\theta^2}\right)\\
&+a_{j-1}\left(c_{12}-c_{21}(p+j-1)+\frac{2}{\theta}(p+j-1)(p+j-2)+\frac{4}{\theta}(p+j-1)\right)\\
&+a_{j-2}\left(-c_{22}(p+j-2)+2c_{31}(p+j-2)-6(p+j-2)-6(p+j-2)(p+j-3)\right.\\
&\qquad\left.+c_{31}(p+j-2)(p+j-3)-(p+j-2)(p+j-3)(p+j-4)\right)\\
&+a_{j-3}\left(2c_{32}(p+j-3)+c_{32}(p+j-3)(p+j-4)\right), \tag*{(34)}
\end{aligned} 0 = a j ( c 11 − θ 2 p + j ) + a j − 1 ( c 12 − c 21 ( p + j − 1 ) + θ 2 ( p + j − 1 ) ( p + j − 2 ) + θ 4 ( p + j − 1 ) ) + a j − 2 ( − c 22 ( p + j − 2 ) + 2 c 31 ( p + j − 2 ) − 6 ( p + j − 2 ) − 6 ( p + j − 2 ) ( p + j − 3 ) + c 31 ( p + j − 2 ) ( p + j − 3 ) − ( p + j − 2 ) ( p + j − 3 ) ( p + j − 4 ) ) + a j − 3 ( 2 c 32 ( p + j − 3 ) + c 32 ( p + j − 3 ) ( p + j − 4 ) ) , (34) with the notation c i j c_{ij} c ij from the expansions of q i ( u ) q_i(u) q i ( u ) . The convention a n = 0 a_n=0 a n = 0 for n < 0 n<0 n < 0 means that for j = 1 , 2 j=1,2 j = 1 , 2 some terms of (33) are absent. For clarity we
display the first two steps separately:
0 = a 1 ( c 11 − p + 1 θ 2 ) + ( c 12 − c 21 p + 2 θ p ( p − 1 ) + 4 θ p ) (35) 0=a_1\left(c_{11}-\frac{p+1}{\theta^2}\right)+\left(c_{12}-c_{21}p+\frac{2}{\theta}p(p-1)+\frac{4}{\theta}p\right) \tag*{(35)} 0 = a 1 ( c 11 − θ 2 p + 1 ) + ( c 12 − c 21 p + θ 2 p ( p − 1 ) + θ 4 p ) (35)
0 = a 2 ( c 11 − p + 2 θ 2 ) + a 1 ( c 12 − c 21 ( p + 1 ) + 2 θ ( p + 1 ) p + 4 θ ( p + 1 ) ) + ( − c 22 p + 2 c 31 p − 6 p − 6 p ( p − 1 ) + c 31 p ( p − 1 ) − p ( p − 1 ) ( p − 2 ) ) (36) 0=a_2\left(c_{11}-\frac{p+2}{\theta^2}\right)+a_1\left(c_{12}-c_{21}(p+1)+\frac{2}{\theta}(p+1)p+\frac{4}{\theta}(p+1)\right)
+\left(-c_{22}p+2c_{31}p-6p-6p(p-1)+c_{31}p(p-1)-p(p-1)(p-2)\right) \tag*{(36)} 0 = a 2 ( c 11 − θ 2 p + 2 ) + a 1 ( c 12 − c 21 ( p + 1 ) + θ 2 ( p + 1 ) p + θ 4 ( p + 1 ) ) + ( − c 22 p + 2 c 31 p − 6 p − 6 p ( p − 1 ) + c 31 p ( p − 1 ) − p ( p − 1 ) ( p − 2 ) ) (36)
while the general formula (33) applies for j ≥ 3 j \ge3 j ≥ 3 . The step j = 0 j=0 j = 0 of (33) is the indicial equality c 11 = p / θ 2 c_{11}=p/\theta^2 c 11 = p / θ 2 . Hence the factor multiplying a j a_j a j equals c 11 − ( p + j ) / θ 2 = − j / θ 2 ≠ 0 c_{11}-(p+j)/\theta^2=-j/\theta^2\ne0 c 11 − ( p + j ) / θ 2 = − j / θ 2 = 0 for all j ≥ 1 j\ge1 j ≥ 1 , and the recursion is solvable uniquely. The series (32) is, in general, divergent (a Gevrey class of order 1), but it serves as the asymptotic expansion of the actual solution G p o w G_{\mathrm{pow}} G pow of equation (28) (Lemma 6 ), and the physical density G = Φ ′ G=\Phi' G = Φ ′ contains this branch with a coefficient ϖ > 0 \varpi>0 ϖ > 0 (Proposition 4), so that
Ψ ( u ) ∼ C u − β , u → ∞ , C = ϖ β > 0. (37) \Psi(u)\sim Cu^{-\beta},\qquad u\to\infty,\qquad C=\frac{\varpi}{\beta}>0. \tag*{(37)} Ψ ( u ) ∼ C u − β , u → ∞ , C = β ϖ > 0. (37)
Proof Step 1. The nature of the singular point. We substitute z = 1 / u z=1/u z = 1/ u , w ( z ) : = G ( 1 / z ) w(z):=G(1/z) w ( z ) := G ( 1/ z ) , which converts (28) to the equation
− z 6 w ′ ′ ′ + ( − 6 z 5 + q ~ 3 z 4 ) w ′ ′ + ( − 6 z 4 + 2 q ~ 3 z 3 − q ~ 2 z 2 ) w ′ + q ~ 1 w = 0 , (38) -z^6w'''+(-6z^5+\widetilde{q}_3z^4)w''+(-6z^4+2\widetilde{q}_3z^3-\widetilde{q}_2z^2)w'+\widetilde{q}_1w=0, \tag*{(38)} − z 6 w ′′′ + ( − 6 z 5 + q 3 z 4 ) w ′′ + ( − 6 z 4 + 2 q 3 z 3 − q 2 z 2 ) w ′ + q 1 w = 0 , (38)
where q ~ i ( z ) : = q i ( 1 / z ) \widetilde{q}_i(z):=q_i(1/z) q i ( z ) := q i ( 1/ z ) . Since q ~ 3 ( z ) = 2 / θ + O ( z ) \widetilde{q}_3(z)=2/\theta+O(z) q 3 ( z ) = 2/ θ + O ( z ) , q ~ 2 ( z ) = 1 / θ 2 + O ( z ) \widetilde{q}_2(z)=1/\theta^2+O(z) q 2 ( z ) = 1/ θ 2 + O ( z ) , and q ~ 1 ( z ) = O ( z ) \widetilde{q}_1(z)=O(z) q 1 ( z ) = O ( z ) as z → 0 z\to0 z → 0 , the point z = 0 z=0 z = 0 (corresponding to u = ∞ u=\infty u = ∞ ) is an irregular singular point of rank 1: the characteristic equation for the exponential solutions e μ u e^{\mu u} e μu of (28) has the form μ ( μ + 1 / θ ) 2 = 0 \mu(\mu+1/\theta)^2=0 μ ( μ + 1/ θ ) 2 = 0 , and the root μ = 0 \mu=0 μ = 0 , corresponding to the power branch, is simple (see [32 Sects. 12–15] ).
Step 2. The indicial equation. We substitute the power expansion
G ( u ) = u − p ∑ j ≥ 0 a j u − j , a 0 = 1 , G(u)=u^{-p}\sum_{j\ge0}a_ju^{-j},\qquad a_0=1, G ( u ) = u − p j ≥ 0 ∑ a j u − j , a 0 = 1 ,
into the normalized form of (28). As u → ∞ u\to\infty u → ∞ we obtain the following orders of the contributions:
Table 1.
The dominant balance (the highest power of u u u ) occurs at u − p − 1 u^{-p-1} u − p − 1 : comparing u − p − 1 u^{-p-1} u − p − 1 with u − p − 2 u^{-p-2} u − p − 2 and u − p − 3 u^{-p-3} u − p − 3 as u → ∞ u\to\infty u → ∞ , we find that the leading terms are q 2 G ′ q_2G' q 2 G ′ (the constant part 1 / θ 2 1/\theta^2 1/ θ 2 ) and q 1 G q_1G q 1 G (the term c 11 / u c_{11}/u c 11 / u ). All other contributions (G ′ ′ ′ G''' G ′′′ , the subdominant parts of q 3 G ′ ′ q_3G'' q 3 G ′′ and q 2 G ′ q_2G' q 2 G ′ ) belong to higher powers of 1 / u 1/u 1/ u and play no role in determining p p p . We equate to zero the coefficient C p + 1 C_{p+1} C p + 1 at u − p − 1 u^{-p-1} u − p − 1 (with a 0 = 1 a_0=1 a 0 = 1 ) and obtain
C p + 1 = c 11 − p θ 2 = 0 ⟹ p = c 11 θ 2 = 2 a σ 2 = β + 1. C_{p+1}=c_{11}-\frac{p}{\theta^2}=0 \quad\Longrightarrow\quad p=c_{11}\theta^2=\frac{2a}{\sigma^2}=\beta+1. C p + 1 = c 11 − θ 2 p = 0 ⟹ p = c 11 θ 2 = σ 2 2 a = β + 1.
Step 3. The recurrence. We systematically count all contributions to the coefficient of u − p − j − 1 u^{-p-j-1} u − p − j − 1 from each of the four terms of (28) in normalized form, which gives (33). With P : = p + j P:=p+j P := p + j the contributions of the four terms are
G ′ ′ ′ : − a j − 2 ( P − 2 ) ( P − 1 ) P , G''':\quad-a_{j-2}(P-2)(P-1)P, G ′′′ : − a j − 2 ( P − 2 ) ( P − 1 ) P ,
q 3 G ′ ′ : 2 θ a j − 1 ( P − 1 ) P + c 31 a j − 2 ( P − 2 ) ( P − 1 ) + c 32 a j − 3 ( P − 3 ) ( P − 2 ) , q_3G'':\quad\frac{2}{\theta}a_{j-1}(P-1)P+c_{31}a_{j-2}(P-2)(P-1)+c_{32}a_{j-3}(P-3)(P-2), q 3 G ′′ : θ 2 a j − 1 ( P − 1 ) P + c 31 a j − 2 ( P − 2 ) ( P − 1 ) + c 32 a j − 3 ( P − 3 ) ( P − 2 ) ,
q 2 G ′ : − θ − 2 a j P − c 21 a j − 1 ( P − 1 ) − c 22 a j − 2 ( P − 2 ) , q_2G':\quad-\theta^{-2}a_jP-c_{21}a_{j-1}(P-1)-c_{22}a_{j-2}(P-2), q 2 G ′ : − θ − 2 a j P − c 21 a j − 1 ( P − 1 ) − c 22 a j − 2 ( P − 2 ) ,
q 1 G : c 11 a j + c 12 a j − 1 . q_1G:\quad c_{11}a_j+c_{12}a_{j-1}. q 1 G : c 11 a j + c 12 a j − 1 .
Expanding the products, we arrive at the form (33).
Step 4. The asymptotics of Ψ \Psi Ψ . By Lemma 6, to the formal series (32) there corresponds an actual solution G p o w G_{\mathrm{pow}} G pow , and the physical density has the form G = ϖ G p o w + O ( e − δ u ) G=\varpi G_{\mathrm{pow}}+O(e^{-\delta u}) G = ϖ G pow + O ( e − δ u ) for some δ > 0 \delta>0 δ > 0 (the remaining solutions are exponentially small). By Proposition 4, ϖ ≠ 0 \varpi\ne0 ϖ = 0 , and from G ≥ 0 G\ge0 G ≥ 0 (the monotonicity of Φ \Phi Φ ) we get ϖ > 0 \varpi>0 ϖ > 0 . From Ψ ( ∞ ) = 0 \Psi(\infty)=0 Ψ ( ∞ ) = 0 , by integration,
Ψ ( u ) = ∫ u ∞ G ( v ) d v ∼ ϖ p − 1 u 1 − p = ϖ β u − β . \Psi(u)=\int_u^\infty G(v)\,\mathrm{d}v\sim\frac{\varpi}{p-1}u^{1-p}=\frac{\varpi}{\beta}u^{-\beta}. Ψ ( u ) = ∫ u ∞ G ( v ) d v ∼ p − 1 ϖ u 1 − p = β ϖ u − β .
We fix the normalization a 0 = 1 a_0=1 a 0 = 1 only for the formal series: the amplitude ϖ \varpi ϖ (the connection coefficient) is determined by the original problem, not by the reduced equation. □ \square □
Remark 4 The exponent p p p should not be extracted from the coefficient of the lowest power u − p − 3 u^{-p-3} u − p − 3 , which is occupied by the term G ′ ′ ′ G''' G ′′′ : at this power the coefficients a 1 , a 2 a_1,a_2 a 1 , a 2 of the series appear along with a 0 a_0 a 0 , so the corresponding relation enters the recurrence and does not constrain p p p . For an irregular singular point, the equation for the exponent of the power branch is given by the dominant balance: since − p − 1 > − p − 2 > − p − 3 -p-1>-p-2>-p-3 − p − 1 > − p − 2 > − p − 3 , it is formed by the terms q 2 G ′ q_2G' q 2 G ′ and q 1 G q_1G q 1 G , while the highest derivative G ′ ′ ′ G''' G ′′′ is subdominant. The form (33) is equivalent to the compact one: with P : = p + j P:=p+j P := p + j the factors multiplying a j − 1 a_{j-1} a j − 1 , a j − 2 a_{j-2} a j − 2 , a j − 3 a_{j-3} a j − 3 equal c 12 − c 21 ( P − 1 ) + 2 θ ( P − 1 ) P c_{12}-c_{21}(P-1)+\frac{2}{\theta}(P-1)P c 12 − c 21 ( P − 1 ) + θ 2 ( P − 1 ) P , − c 22 ( P − 2 ) + c 31 ( P − 2 ) ( P − 1 ) − ( P − 2 ) ( P − 1 ) P -c_{22}(P-2)+c_{31}(P-2)(P-1)-(P-2)(P-1)P − c 22 ( P − 2 ) + c 31 ( P − 2 ) ( P − 1 ) − ( P − 2 ) ( P − 1 ) P and c 32 ( P − 3 ) ( P − 2 ) c_{32}(P-3)(P-2) c 32 ( P − 3 ) ( P − 2 ) respectively. For example, with a 0 = 1 a_0=1 a 0 = 1 we get a 1 = 2 ( c − 2 α θ ) / σ 2 a_1=2(c-2\alpha\theta)/\sigma^2 a 1 = 2 ( c − 2 α θ ) / σ 2 and a 2 = − 2 ( 3 a α θ 2 − 4 α 2 + 4 α c θ − c 2 ) / σ 4 a_2=-2(3a\alpha\theta^2-4\alpha^2+4\alpha c\theta-c^2)/\sigma^4 a 2 = − 2 ( 3 a α θ 2 − 4 α 2 + 4 α c θ − c 2 ) / σ 4 .
The exponent of the power branch for a general k k k
In normalized form the ODE for G = Φ ′ G=\Phi' G = Φ ′ for general k k k reads
G ( k + 1 ) + q k + 1 ( u ) G ( k ) + ⋯ + q 2 ( u ) G ′ + q 1 ( u ) G = 0. G^{(k+1)}+q_{k+1}(u)G^{(k)}+\cdots+q_2(u)G'+q_1(u)G=0. G ( k + 1 ) + q k + 1 ( u ) G ( k ) + ⋯ + q 2 ( u ) G ′ + q 1 ( u ) G = 0.
The coefficients q 1 q_1 q 1 , q 2 q_2 q 2 have the expansions
q 1 ( u ) = 2 a σ 2 θ k ⋅ 1 u + 2 ( θ k a + c − θ k α ) σ 2 θ k ⋅ 1 u 2 , (39) q_1(u)=\frac{2a}{\sigma^2\theta^k}\cdot\frac{1}{u}+\frac{2(\theta ka+c-\theta k\alpha)}{\sigma^2\theta^k}\cdot\frac{1}{u^2}, \tag*{(39)} q 1 ( u ) = σ 2 θ k 2 a ⋅ u 1 + σ 2 θ k 2 ( θ k a + c − θ k α ) ⋅ u 2 1 , (39)
q 2 ( u ) = 1 θ k + 2 k ( σ 2 + a ) σ 2 θ k − 1 ⋅ 1 u + k ( k − 1 ) θ ( σ 2 + 2 a − α ) + 2 k c σ 2 θ k − 1 ⋅ 1 u 2 , (40) q_2(u)=\frac{1}{\theta^k}+\frac{2k(\sigma^2+a)}{\sigma^2\theta^{k-1}}\cdot\frac{1}{u}+\frac{k(k-1)\theta(\sigma^2+2a-\alpha)+2kc}{\sigma^2\theta^{k-1}}\cdot\frac{1}{u^2}, \tag*{(40)} q 2 ( u ) = θ k 1 + σ 2 θ k − 1 2 k ( σ 2 + a ) ⋅ u 1 + σ 2 θ k − 1 k ( k − 1 ) θ ( σ 2 + 2 a − α ) + 2 k c ⋅ u 2 1 , (40)
obtained by the direct division g n / g k + 2 g_n/g_{k+2} g n / g k + 2 (taking into account g k + 2 = 1 2 σ 2 θ k u 2 g_{k+2}=\frac{1}{2}\sigma^2\theta^k u^2 g k + 2 = 2 1 σ 2 θ k u 2 ). Here and below c 11 : = 2 a / ( σ 2 θ k ) c_{11}:=2a/(\sigma^2\theta^k) c 11 := 2 a / ( σ 2 θ k ) is the coefficient of 1 / u 1/u 1/ u in q 1 q_1 q 1 (for k = 2 k=2 k = 2 this is the previous c 11 c_{11} c 11 ). The formal power expansion (a Thomé series) G ( u ) = u − p ∑ j ≥ 0 a j u − j G(u)=u^{-p}\sum_{j\geq0}a_j u^{-j} G ( u ) = u − p ∑ j ≥ 0 a j u − j with a 0 = 1 a_0=1 a 0 = 1 gives G ( i ) ( u ) ∼ ( − 1 ) i p ( p + 1 ) ⋯ ( p + i − 1 ) u − p − i G^{(i)}(u)\sim(-1)^i p(p+1)\cdots(p+i-1)u^{-p-i} G ( i ) ( u ) ∼ ( − 1 ) i p ( p + 1 ) ⋯ ( p + i − 1 ) u − p − i as u → ∞ u\to\infty u → ∞ . The dominant balance now reads as follows:
Here q m 0 q_{m0} q m 0 are the leading constant values of q m ( u ) = g m / g k + 2 q_m(u)=g_m/g_{k+2} q m ( u ) = g m / g k + 2 as u → ∞ u\to\infty u → ∞ . For 2 ≤ m ≤ k + 1 2\leq m\leq k+1 2 ≤ m ≤ k + 1 they are finite, because the g m g_m g m are polynomials of degree at most 2 2 2 , while g k + 2 ∝ u 2 g_{k+2}\propto u^2 g k + 2 ∝ u 2 .
The highest power of u u u among all terms is u − p − 1 u^{-p-1} u − p − 1 (since − p − 1 > − p − m + 1 -p-1>-p-m+1 − p − 1 > − p − m + 1 for m ≥ 3 m\geq3 m ≥ 3 ). The dominant balance is therefore determined exclusively by q 2 G ′ q_2G' q 2 G ′ and q 1 G q_1G q 1 G :
1 θ k ⋅ ( − p ) + c 11 = 0 ⟹ p = c 11 θ k = 2 a σ 2 = β + 1. \frac{1}{\theta^k}\cdot(-p)+c_{11}=0\quad\Longrightarrow\quad p=c_{11}\theta^k=\frac{2a}{\sigma^2}=\beta+1. θ k 1 ⋅ ( − p ) + c 11 = 0 ⟹ p = c 11 θ k = σ 2 2 a = β + 1.
All the remaining terms (G ( k + 1 ) G^{(k+1)} G ( k + 1 ) and q m G ( m − 1 ) q_mG^{(m-1)} q m G ( m − 1 ) for m ≥ 3 m\geq3 m ≥ 3 ) contribute O ( u − p − 2 ) O(u^{-p-2}) O ( u − p − 2 ) and do not enter the determining equation. The exponent p p p thus does not depend on k k k , c c c , α \alpha α , θ \theta θ : it is determined solely by the parameters of the risky asset ( a , σ ) (a,\sigma) ( a , σ ) . For the same reason the term a j a_j a j enters the coefficient of u − p − j − 1 u^{-p-j-1} u − p − j − 1 only through q 2 G ′ q_2G' q 2 G ′ and q 1 G q_1G q 1 G , so the factor multiplying a j a_j a j in the recurrence equals c 11 − ( p + j ) / θ k = − j / θ k ≠ 0 c_{11}-(p+j)/\theta^k=-j/\theta^k\ne0 c 11 − ( p + j ) / θ k = − j / θ k = 0 , and the formal series is determined uniquely for any k k k . The step j = 1 j=1 j = 1 of the same recurrence (the coefficient of u − p − 2 u^{-p-2} u − p − 2 is assembled from q 3 G ′ ′ q_3G'' q 3 G ′′ , q 2 G ′ q_2G' q 2 G ′ and q 1 G q_1G q 1 G ) gives a 1 = θ k [ q 30 p ( p + 1 ) − q 21 p + c 12 ] a_1=\theta^k[q_{30}p(p+1)-q_{21}p+c_{12}] a 1 = θ k [ q 30 p ( p + 1 ) − q 21 p + c 12 ] , where q 30 = k / θ k − 1 q_{30}=k/\theta^{k-1} q 30 = k / θ k − 1 is the limit of q 3 = g 3 / g k + 2 q_3=g_3/g_{k+2} q 3 = g 3 / g k + 2 as u → ∞ u\to\infty u → ∞ (for k = 1 k=1 k = 1 the role of q 3 q_3 q 3 is played by the unit coefficient of G ′ ′ G'' G ′′ ), while q 21 = 2 k ( σ 2 + a ) / ( σ 2 θ k − 1 ) q_{21}=2k(\sigma^2+a)/(\sigma^2\theta^{k-1}) q 21 = 2 k ( σ 2 + a ) / ( σ 2 θ k − 1 ) and c 12 = 2 ( θ k a + c − θ k α ) / ( σ 2 θ k ) c_{12}=2(\theta ka+c-\theta k\alpha)/(\sigma^2\theta^k) c 12 = 2 ( θ k a + c − θ k α ) / ( σ 2 θ k ) are the coefficients of 1 / u 1/u 1/ u in q 2 q_2 q 2 and of 1 / u 2 1/u^2 1/ u 2 in q 1 q_1 q 1 . Substituting p = 2 a / σ 2 p=2a/\sigma^2 p = 2 a / σ 2 , we get a 1 = 2 ( c − k α θ ) / σ 2 a_1=2(c-k\alpha\theta)/\sigma^2 a 1 = 2 ( c − k α θ ) / σ 2 . We have thus proved the following proposition.
Proposition 2 (formal power branch for a general k k k ) For model (A) with an arbitrary k ≥ 1 k \ge1 k ≥ 1 the reduced ODE for G = Φ ′ G = \Phi' G = Φ ′ admits a unique formal solution of the form u − p ∑ j ≥ 0 a j u − j u^{-p}\sum_{j\geq0}a_j u^{-j} u − p ∑ j ≥ 0 a j u − j with a 0 = 1 a_0=1 a 0 = 1 and the exponent p = β + 1 = 2 α / σ 2 p=\beta+1=2\alpha/\sigma^2 p = β + 1 = 2 α / σ 2 , which does not depend on k k k , c c c , α \alpha α or θ \theta θ . The factor multiplying a j a_j a j in the recurrence is − j / θ k ≠ 0 -j/\theta^k\ne0 − j / θ k = 0 , and the first coefficient is a 1 = 2 ( c − k α θ ) / σ 2 a_1=2(c-k\alpha\theta)/\sigma^2 a 1 = 2 ( c − k α θ ) / σ 2 .
Proposition 4 establishes the realization of this branch in the physical solution and the asymptotics Ψ ( u ) ∼ C k u − β \Psi(u)\sim C_k u^{-\beta} Ψ ( u ) ∼ C k u − β , C k > 0 C_k>0 C k > 0 .
The irregular singular point u = ∞ u=\infty u = ∞ of rank 1 generates a fundamental set of solutions consisting of the power solution and exponentially small solutions (formal Thomé series, see [32 ] , and for second-order equations [23 Chapter 7] , [31 Sect. 3.1] ): the characteristic polynomial of the equation for G G G is μ ( 1 + θ μ ) k \mu(1+\theta\mu)^k μ ( 1 + θ μ ) k , the root μ = 0 \mu=0 μ = 0 is simple for all k k k , and the root μ = − 1 / θ \mu=-1/\theta μ = − 1/ θ has multiplicity k k k . For example, for k = 1 k=1 k = 1 (model (A)) the ODE (48) has two fundamental solutions near u = ∞ u=\infty u = ∞ : the power one ∼ u − p \sim u^{-p} ∼ u − p and the exponentially small one ∼ e − u / θ u − 2 \sim e^{-u/\theta}u^{-2} ∼ e − u / θ u − 2 . Substituting G = e μ u u ρ G=e^{\mu u}u^\rho G = e μu u ρ directly into (48), we confirm this: the order u 2 u^2 u 2 gives 1 2 σ 2 μ ( θ μ + 1 ) = 0 \frac{1}{2}\sigma^2\mu(\theta\mu+1)=0 2 1 σ 2 μ ( θ μ + 1 ) = 0 , i.e., μ = − 1 / θ \mu=-1/\theta μ = − 1/ θ , and the order u u u gives − 1 2 σ 2 ( ρ + 2 ) = 0 -\frac{1}{2}\sigma^2(\rho+2)=0 − 2 1 σ 2 ( ρ + 2 ) = 0 , i.e., ρ = − 2 \rho=-2 ρ = − 2 . For k = 1 k=1 k = 1 , the connection coefficient of the power branch in [30 ] is identified with the connection coefficient of the doubly confluent Heun equation. Its positivity is proved there from a first integral, and for an integer 2 α / σ 2 2\alpha/\sigma^2 2 α / σ 2 it is written explicitly in closed form.
Remark 5 The statement Ψ ( u ) ∼ C u − β \Psi(u)\sim Cu^{-\beta} Ψ ( u ) ∼ C u − β with C > 0 C>0 C > 0 is equivalent to a nonzero coefficient of the power branch in the expansion of the physical density G = Φ ′ G=\Phi' G = Φ ′ in the actual fundamental set (Lemma 6 ). For k = 1 k=1 k = 1 this is established in [7 ] and, from a first integral of the ODE, in [30 ] . For general claim distributions with E [ ξ β ] < ∞ \mathbb{E}[\xi^\beta]<\infty E [ ξ β ] < ∞ the asymptotics with C > 0 C>0 C > 0 were obtained by the integral-equation method in [29 Theorem 1] . Here, for all k k k , the nonvanishing follows from the power-law lower bound for Ψ \Psi Ψ (Proposition 4). Conditions (iv)–(v) of Remark 2 (the monotonicity and integrability of G G G ) narrow the solution space but do not separate the power branch u − p u^{-p} u − p from the exponentially small modes (both are integrable). This is exactly why a separation requires an external argument (the lower bound). The criterion of Proposition 1 selects solutions of the IDE among solutions of the ODE, but by itself it does not imply the nonvanishing of the coefficient c β c_\beta c β at the power branch (Proposition 4). Borel summability of the series (32) is not needed for the asymptotics. For k = 2 k=2 k = 2 , in addition, the coefficients for generic parameters (with nonzero Stokes multipliers relating G p o w G_{\mathrm{pow}} G pow to the exponentially small modes) grow factorially, lim j → ∞ ( ∣ a j ∣ / j ! ) 1 / j = θ \lim_{j\to\infty}(|a_j|/j!)^{1/j}=\theta lim j → ∞ ( ∣ a j ∣/ j ! ) 1/ j = θ (for 8 α ≤ σ 2 8\alpha\leq\sigma^2 8 α ≤ σ 2 also a j ( j a j − 1 ) → θ a_j(j a_{j-1})\to\theta a j ( j a j − 1 ) → θ , whereas for 8 α > σ 2 8\alpha>\sigma^2 8 α > σ 2 the exponents s s s of the exponentially small modes e − u / θ u s e^{-u/\theta}u^s e − u / θ u s , σ 2 ( s + 1 ) ( s + 2 ) + 2 α = 0 \sigma^2(s+1)(s+2)+2\alpha=0 σ 2 ( s + 1 ) ( s + 2 ) + 2 α = 0 , are complex, and a j a_j a j changes sign infinitely often — log-periodically, i.e., more and more rarely — and all this is confirmed numerically), so that the Borel transform of the series has a singularity at the point 1 / θ 1/\theta 1/ θ of the positive half-axis, the direction u → + ∞ u\to+\infty u → + ∞ is a Stokes line, and the formal series determines the actual solution only up to exponentially small terms e − u / θ + o ( u ) e^{-u/\theta+o(u)} e − u / θ + o ( u ) .
All three models: the exponent, the fundamental set, the tail realization
The detailed asymptotic analysis above was carried out for model (A), for which the most complete basis for comparison is available [7 , 17 , 30 ] . The exponent of the power asymptotics, however, is the same for all three models.
Proposition 3. In each of the three models the equation for the exponent s s s of the algebraic (power) branch Φ ∼ u s \Phi\sim u^{s} Φ ∼ u s of the reduced ODE at the irregular singular point u = ∞ u = \infty u = ∞ is one and the same:
s ( 1 2 σ 2 ( s − 1 ) + a ) = 0 , (41) s\left(\frac{1}{2}\sigma^{2}(s-1)+a\right)=0, \tag*{(41)} s ( 2 1 σ 2 ( s − 1 ) + a ) = 0 , (41) with roots s = 0 s=0 s = 0 and s = 1 − 2 a / σ 2 = − β s=1-2a/\sigma^{2}=-\beta s = 1 − 2 a / σ 2 = − β . The exponent β = 2 a / σ 2 − 1 \beta=2a/\sigma^{2}-1 β = 2 a / σ 2 − 1 depends neither on the shape parameter k k k nor on the parameters of the risk process c c c , α \alpha α , θ \theta θ .
Proof. Substitute Φ ∼ u s \Phi\sim u^{s} Φ ∼ u s into the reduced equation. The annihilating operators take u s u^{s} u s again into u s u^{s} u s with leading coefficient 1, so the coefficient of the highest power u s u^{s} u s is provided by the operator L \mathcal{L} L alone and equals 1 2 σ 2 s ( s − 1 ) + a s \frac{1}{2}\sigma^{2}s(s-1)+as 2 1 σ 2 s ( s − 1 ) + a s . The integral terms contribute to no such power: the differences − α N Φ + α Φ -\alpha\mathcal{N}\Phi+\alpha\Phi − α N Φ + α Φ in model (A), − α N ~ Φ + α Φ -\alpha\widetilde{\mathcal{N}}\Phi+\alpha\Phi − α N Φ + α Φ in model (B), and − ( α 1 + α 2 ) N N ~ Φ + α 1 N Φ + α 2 N ~ Φ -(\alpha_{1}+\alpha_{2})\mathcal{N}\widetilde{\mathcal{N}}\Phi+\alpha_{1}\mathcal{N}\Phi+\alpha_{2}\widetilde{\mathcal{N}}\Phi − ( α 1 + α 2 ) N N Φ + α 1 N Φ + α 2 N Φ in model (C) start at order u s − 1 u^{s-1} u s − 1 . We equate to zero the coefficient of u s u^{s} u s and obtain (40). The argument does not depend on k k k : all three operators N \mathcal{N} N , N ~ \widetilde{\mathcal{N}} N , N N ~ \mathcal{N}\widetilde{\mathcal{N}} N N are polynomials in D = d / d u D=d/du D = d / d u with constant term 1, so for Φ ∼ u s \Phi\sim u^{s} Φ ∼ u s we have N Φ = Φ + O ( u s − 1 ) \mathcal{N}\Phi=\Phi+O(u^{s-1}) N Φ = Φ + O ( u s − 1 ) and so on, and the integral part at order u s u^{s} u s cancels exactly (− α N Φ + α Φ = O ( u s − 1 ) -\alpha\mathcal{N}\Phi+\alpha\Phi=O(u^{s-1}) − α N Φ + α Φ = O ( u s − 1 ) in (A), (B), and − ( α 1 + α 2 ) N N ~ Φ + α 1 N Φ + α 2 N ~ Φ = ( − ( α 1 + α 2 ) + α 1 + α 2 ) Φ + O ( u s − 1 ) = O ( u s − 1 ) -(\alpha_{1}+\alpha_{2})\mathcal{N}\widetilde{\mathcal{N}}\Phi+\alpha_{1}\mathcal{N}\Phi+\alpha_{2}\widetilde{\mathcal{N}}\Phi= (-(\alpha_{1}+\alpha_{2})+\alpha_{1}+\alpha_{2})\Phi+O(u^{s-1})=O(u^{s-1}) − ( α 1 + α 2 ) N N Φ + α 1 N Φ + α 2 N Φ = ( − ( α 1 + α 2 ) + α 1 + α 2 ) Φ + O ( u s − 1 ) = O ( u s − 1 ) in (C)). □ \square □
In the annuity model (B) all k k k exponential modes grow (the symbol of the annihilator ( 1 − θ μ ) k (1-\theta\mu)^{k} ( 1 − θ μ ) k has the unique root μ = 1 / θ > 0 \mu=1/\theta>0 μ = 1/ θ > 0 ). Hence the bounded solutions of the reduced ODE form a two-dimensional space spanned by 1 and the power branch, and we obtain the asymptotics Ψ ( u ) ∼ C u − β \Psi(u)\sim Cu^{-\beta} Ψ ( u ) ∼ C u − β from the ODE alone (Proposition 4), the coefficients of the power series being given by a recurrence analogous to (33). In the mixed model (C) the solution space of the reduced ODE (25) of order k 1 + k 2 + 2 k_{1}+k_{2}+2 k 1 + k 2 + 2 at the irregular point u = ∞ u=\infty u = ∞ is wider than in the one-sided models. We introduce the symbol of the annihilator
A ( μ ) : = ( 1 − θ 1 μ ) k 1 ( 1 + θ 2 μ ) k 2 , so that N ~ k 1 N k 2 e μ u = A ( μ ) e μ u . (42) A(\mu):=(1-\theta_{1}\mu)^{k_{1}}(1+\theta_{2}\mu)^{k_{2}},\qquad\text{so that }\widetilde{\mathcal{N}}_{k_{1}}\mathcal{N}_{k_{2}}e^{\mu u}=A(\mu)e^{\mu u}. \tag*{(42)} A ( μ ) := ( 1 − θ 1 μ ) k 1 ( 1 + θ 2 μ ) k 2 , so that N k 1 N k 2 e μu = A ( μ ) e μu . (42)
Lemma 5 (fundamental set of model (C)). The ( k 1 + k 2 + 2 ) (k_{1}+k_{2}+2) ( k 1 + k 2 + 2 ) -dimensional space of formal solutions (Thomé series, generalized at repeated roots by fractional powers of u u u in the exponent and logarithmic terms) of equation (25) at the irregular singular point u = ∞ u=\infty u = ∞ splits into three groups:
two algebraic branches u s u^{s} u s , s ∈ { 0 , − β } s\in\{0,-\beta\} s ∈ { 0 , − β } , generated by the diffusion part L \mathcal{L} L (Proposition 3),
(b) k 1 k_1 k 1 solutions with leading exponential factor e u / θ 1 e^{u/\theta_1} e u / θ 1 (exponentially growing), corresponding to the root μ = 1 / θ 1 \mu= 1/\theta_1 μ = 1/ θ 1 of the symbol (41) of multiplicity k 1 k_1 k 1 (upward jumps),
(c) k 2 k_2 k 2 solutions with leading exponential factor e − u / θ 2 e^{-u/\theta_2} e − u / θ 2 (exponentially decaying), corresponding to the root μ = − 1 / θ 2 \mu= -1/\theta_2 μ = − 1/ θ 2 of multiplicity k 2 k_2 k 2 (downward jumps).
Proof. Here k 1 , k 2 ∈ N k_1,k_2 \in\mathbb{N} k 1 , k 2 ∈ N , so that N ~ k 1 N k 2 \widetilde{\mathcal{N}}_{k_1}\mathcal{N}_{k_2} N k 1 N k 2 is a differential operator of order k 1 + k 2 k_1+k_2 k 1 + k 2 with constant coefficients. We substitute Φ = e μ u v \Phi= e^{\mu u}v Φ = e μu v with v ∼ u ρ v \sim u^\rho v ∼ u ρ . By the Leibniz formula,
N ~ k 1 N k 2 [ e μ u v ] = e μ u A ( μ + D ) v = e μ u ( A ( μ ) v + A ′ ( μ ) v ′ + ⋯ ) . \widetilde{\mathcal{N}}_{k_1}\mathcal{N}_{k_2}[e^{\mu u}v] = e^{\mu u}A(\mu+D)v = e^{\mu u}\left(A(\mu)v + A'(\mu)v' + \cdots\right). N k 1 N k 2 [ e μu v ] = e μu A ( μ + D ) v = e μu ( A ( μ ) v + A ′ ( μ ) v ′ + ⋯ ) .
while
L Φ = e μ u [ 1 2 σ 2 μ 2 u ρ + 2 + μ ( σ 2 ρ + a ) u ρ + 1 + O ( u ρ ) ] . \mathcal{L}\Phi= e^{\mu u}\left[\frac{1}{2}\sigma^2\mu^2u^{\rho+2}+\mu(\sigma^2\rho+a)u^{\rho+1}+O(u^\rho)\right]. L Φ = e μu [ 2 1 σ 2 μ 2 u ρ + 2 + μ ( σ 2 ρ + a ) u ρ + 1 + O ( u ρ ) ] .
The leading term is
N ~ k 1 N k 2 L Φ = 1 2 σ 2 μ 2 A ( μ ) u ρ + 2 e μ u ( 1 + O ( u − 1 ) ) . \widetilde{\mathcal{N}}_{k_1}\mathcal{N}_{k_2}\mathcal{L}\Phi= \frac{1}{2}\sigma^2\mu^2A(\mu)u^{\rho+2}e^{\mu u}\left(1+O(u^{-1})\right). N k 1 N k 2 L Φ = 2 1 σ 2 μ 2 A ( μ ) u ρ + 2 e μu ( 1 + O ( u − 1 ) ) .
whereas the remaining terms of (26) are of order u ρ e μ u u^\rho e^{\mu u} u ρ e μu . For μ ≠ 0 \mu\ne0 μ = 0 the condition that the leading term vanish is equivalent to A ( μ ) = 0 A(\mu) = 0 A ( μ ) = 0 , i.e., μ = 1 / θ 1 \mu= 1/\theta_1 μ = 1/ θ 1 or μ = − 1 / θ 2 \mu= -1/\theta_2 μ = − 1/ θ 2 with multiplicities k 1 k_1 k 1 and k 2 k_2 k 2 . The first family of modes grows, and the second one decays. To the sector μ = 0 \mu= 0 μ = 0 there correspond the two algebraic branches with the equation s ( 1 2 σ 2 ( s − 1 ) + a ) = 0 s(\frac{1}{2}\sigma^2(s-1)+a) = 0 s ( 2 1 σ 2 ( s − 1 ) + a ) = 0 of Proposition 3.
The algebraic prefactor only refines the subdominant structure. For a simple root (k i = 1 k_i = 1 k i = 1 : then A ( μ ) = 0 A(\mu) = 0 A ( μ ) = 0 but A ′ ( μ ) ≠ 0 A'(\mu) \ne0 A ′ ( μ ) = 0 — for example, A ′ ( 1 / θ 1 ) = − θ 1 ( 1 + θ 2 / θ 1 ) k 2 ≠ 0 A'(1/\theta_1) = -\theta_1(1+\theta_2/\theta_1)^{k_2} \ne0 A ′ ( 1/ θ 1 ) = − θ 1 ( 1 + θ 2 / θ 1 ) k 2 = 0 ), the next order u ρ + 1 u^{\rho+1} u ρ + 1 gives 1 2 σ 2 μ 2 ( ρ + 2 ) A ′ ( μ ) = 0 \frac{1}{2}\sigma^2\mu^2(\rho+2)A'(\mu) = 0 2 1 σ 2 μ 2 ( ρ + 2 ) A ′ ( μ ) = 0 , whence ρ = − 2 \rho= -2 ρ = − 2 . This agrees with the branch e − u / θ 2 u − 2 e^{-u/\theta_2}u^{-2} e − u / θ 2 u − 2 of (49) (the case k = 1 k = 1 k = 1 , cf. [30 ] ), and for k 1 = k 2 = 1 k_1 = k_2 = 1 k 1 = k 2 = 1 we obtain the whole fundamental set { u 0 , u − β , e u / θ 1 u − 2 , e − u / θ 2 u − 2 } \{u^0, u^{-\beta}, e^{u/\theta_1}u^{-2}, e^{-u/\theta_2}u^{-2}\} { u 0 , u − β , e u / θ 1 u − 2 , e − u / θ 2 u − 2 } by a direct substitution. For a multiplicity k i ≥ 2 k_i \ge2 k i ≥ 2 the root of the symbol becomes multiple, and the exact subexponential structure of the modes (the exponent ρ \rho ρ , possible logarithmic factors and fractional-exponential rank-lowering corrections) depends on the parameters. It is given by the standard theory of normal solutions (Thomé series) at an irregular singular point [32 ] (for second-order equations, see [23 Chapter 7] , [31 Sect. 3.1] ) and is inessential for our purposes: any such correction is o ( u ) o(u) o ( u ) in the exponent, so the leading exponential factor e μ u e^{\mu u} e μu is preserved, and the sign of Re μ \operatorname{Re}\mu Re μ — growth at μ = 1 / θ 1 \mu= 1/\theta_1 μ = 1/ θ 1 and decay at μ = − 1 / θ 2 \mu= -1/\theta_2 μ = − 1/ θ 2 , the only fact used in the selection — remains valid under any multiplicity. □ \square □
Remark 6. The exponent β = 2 a / σ 2 − 1 \beta= 2a/\sigma^2 - 1 β = 2 a / σ 2 − 1 here and the exponent γ = 2 a κ / σ 2 \gamma= 2a\kappa/\sigma^2 γ = 2 aκ / σ 2 in [30 ] (with κ ∈ ( 0 , 1 ] \kappa\in(0,1] κ ∈ ( 0 , 1 ] the fraction of the reserve invested in the risky asset, so that γ = 2 a / σ 2 \gamma= 2a/\sigma^2 γ = 2 a / σ 2 for κ = 1 \kappa= 1 κ = 1 ) are one and the same object in different parametrizations: β = γ − 1 \beta= \gamma- 1 β = γ − 1 , and the power of u u u in the tail is the same, u − β = u 1 − γ u^{-\beta} = u^{1-\gamma} u − β = u 1 − γ , and Φ ′ ( u ) ∼ D 0 u − γ \Phi'(u) \sim D_0u^{-\gamma} Φ ′ ( u ) ∼ D 0 u − γ .
Lemma 6 (actual fundamental set). Let u = ∞ u=\infty u = ∞ be an irregular singular point of rank 1 of a linear ODE of order n n n with polynomial coefficients (such are all reduced equations of Sects. 4–5), and let μ 1 , … , μ r \mu_1,\ldots,\mu_r μ 1 , … , μ r be the roots of its characteristic polynomial with multiplicities n 1 , … , n r n_1,\ldots,n_r n 1 , … , n r . Then, in any closed sector with vertex u = ∞ u=\infty u = ∞ of sufficiently small opening containing the ray [ u 0 , ∞ ) [u_0,\infty) [ u 0 , ∞ ) , there exists a fundamental set of solutions that splits into groups of n i n_i n i solutions of the form e μ i u + o ( u ) e^{\mu_i u+o(u)} e μ i u + o ( u ) . To a simple root μ i \mu_i μ i there corresponds a solution y i = e μ i u u ρ i ( 1 + o ( 1 ) ) y_i=e^{\mu_i u}u^{\rho_i}(1+o(1)) y i = e μ i u u ρ i ( 1 + o ( 1 )) whose formal Thomé series is its Poincaré asymptotic expansion, which can be differentiated termwise. A solution carrying a given asymptotic expansion is fixed up to terms subdominant on the ray (on the positive half-axis, exponentially small). In particular:
(a) To the root μ = 0 \mu=0 μ = 0 of the equation for G = Φ ′ G=\Phi' G = Φ ′ of models ( A ) (A) ( A ) and ( B ) (B) ( B ) (the characteristic polynomials are μ ( 1 + θ μ ) k \mu(1+\theta\mu)^k μ ( 1 + θ μ ) k and μ ( 1 − θ μ ) k \mu(1-\theta\mu)^k μ ( 1 − θ μ ) k ), simple for all k k k , there corresponds the solution G p o w ( u ) = ∑ j < L a j u − j + O ( u − L ) G_{\mathrm{pow}}(u)=\sum_{j<L}a_j u^{-j}+O(u^{-L}) G pow ( u ) = ∑ j < L a j u − j + O ( u − L ) for each L ≥ 1 L\geq1 L ≥ 1 , where the a j a_j a j are the coefficients of the formal series of the corresponding model.
(b) In model ( C ) (C) ( C ) , equation (26) , which contains no term with Φ ( 0 ) \Phi^{(0)} Φ ( 0 ) , is an equation of order k 1 + k 2 + 1 k_1+k_2+1 k 1 + k 2 + 1 for G = Φ ′ G=\Phi' G = Φ ′ with the characteristic polynomial μ A ( μ ) \mu A(\mu) μ A ( μ ) and the simple root μ = 0 \mu=0 μ = 0 . To it there corresponds the solution G p o w ( u ) = u − β − 1 ( 1 + o ( 1 ) ) G_{\mathrm{pow}}(u)=u^{-\beta-1}(1+o(1)) G pow ( u ) = u − β − 1 ( 1 + o ( 1 )) , and the algebraic branches of Lemma 5 correspond to the exact solution Φ ≡ 1 \Phi\equiv1 Φ ≡ 1 and
Φ β ( u ) : = β ∫ u ∞ G p o w ( v ) d v = u − β ( 1 + o ( 1 ) ) , Φ β ′ = − β G p o w . (43) \Phi_\beta(u):=\beta\int_u^\infty G_{\mathrm{pow}}(v)\,\mathrm{d}v=u^{-\beta}(1+o(1)),\qquad\Phi'_\beta=-\beta G_{\mathrm{pow}}. \tag*{(43)} Φ β ( u ) := β ∫ u ∞ G pow ( v ) d v = u − β ( 1 + o ( 1 )) , Φ β ′ = − β G pow . (43)
Since on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) the reduced equations have no singular points, we may extend the fundamental set to the whole half-line.
Proof. These are the classical asymptotic existence theorems for solutions at an irregular singular point. For pairwise distinct roots (k = 1 k=1 k = 1 or k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 ), see [5 Chapter V, Sect. 4, Theorem 4.1] , [12 Theorem VII-5-3] , [32 Sects. 12–15] . Theorem VII-5-3 of [12 ] (a Levinson-type theorem for systems with asymptotically constant coefficients and distinct eigenvalues) is used for the same purpose in [17 Sect. 8] and in [30 ] .
We now verify the conditions of these theorems for the reduced equations. Their coefficients are polynomials in u u u of degree at most 2 2 2 , the leading coefficient is proportional to u 2 u^2 u 2 and does not vanish on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . After division by the leading coefficient, we write an equation of order n n n as the system Y ′ = A ( u ) Y Y'=\mathcal{A}(u)Y Y ′ = A ( u ) Y for Y = ( y , y ′ , … , y ( n − 1 ) ) Y=(y,y',\ldots,y^{(n-1)}) Y = ( y , y ′ , … , y ( n − 1 ) ) with the matrix A ( u ) = A 0 + A 1 u − 1 + A 2 u − 2 \mathcal{A}(u)=\mathcal{A}_0+\mathcal{A}_1u^{-1}+\mathcal{A}_2u^{-2} A ( u ) = A 0 + A 1 u − 1 + A 2 u − 2 . Thus the point u = ∞ u=\infty u = ∞ has Poincaré rank 1 1 1 , and A 0 \mathcal{A}_0 A 0 is the companion matrix of the normalized characteristic polynomial (for the equation for Φ \Phi Φ it is p 2 ( μ ) p_2(\mu) p 2 ( μ ) , see below), its eigenvalues being the roots μ i \mu_i μ i with multiplicities n i n_i n i . All roots are real: 0 0 0 and − 1 / θ -1/\theta − 1/ θ in ( A ) (A) ( A ) , 0 0 0 , 1 / θ 1 1/\theta_1 1/ θ 1 , − 1 / θ 2 -1/\theta_2 − 1/ θ 2 in ( B ) (B) ( B ) , 0 0 0 , 1 / θ 1 1/\theta_1 1/ θ 1 , − 1 / θ 2 -1/\theta_2 − 1/ θ 2 in ( C ) (C) ( C ) . The directions in which two exponentials e μ i u e^{\mu_i u} e μ i u , e μ j u e^{\mu_j u} e μ j u with μ i ≠ μ j \mu_i\ne\mu_j μ i = μ j are comparable in modulus (Re ( ( μ i − μ j ) u ) = 0 \operatorname{Re}((\mu_i-\mu_j)u)=0 Re (( μ i − μ j ) u ) = 0 ) are the rays arg u = ± π / 2 \arg u=\pm\pi/2 arg u = ± π /2 , and the positive half-axis lies strictly inside the sector ∣ arg u ∣ < π / 2 |\arg u|<\pi/2 ∣ arg u ∣ < π /2 , which contains no such rays. In the case of multiple roots, the formal system is first split into blocks corresponding to the distinct roots (the Hukuhara–Turrittin theorem, [32 Sect. 19] ), and the splitting is realized analytically in a closed sector containing the ray by a transformation Y = Π ( u ) Z Y=\Pi(u)Z Y = Π ( u ) Z with Π ( u ) ∼ Π 0 + Π 1 u − 1 + … \Pi(u)\sim\Pi_0+\Pi_1u^{-1}+\ldots Π ( u ) ∼ Π 0 + Π 1 u − 1 + … , det Π 0 ≠ 0 \det\Pi_0\ne0 det Π 0 = 0 [32 Sects. 12–15, 18–19] . The block of the simple root μ = 0 \mu=0 μ = 0 is one-dimensional: z ′ = b ( u ) z z'=b(u)z z ′ = b ( u ) z with b ( u ) ∼ ∑ j ≥ 1 b j u − j b(u)\sim\sum_{j\ge1}b_ju^{-j} b ( u ) ∼ ∑ j ≥ 1 b j u − j , and its solution z = u b 1 exp ( ∫ ∞ u ( b ( v ) − b 1 v − 1 ) d v ) z=u^{b_1}\exp\left(\int_{\infty}^{u}(b(v)-b_1v^{-1})\,dv\right) z = u b 1 exp ( ∫ ∞ u ( b ( v ) − b 1 v − 1 ) d v ) has a complete asymptotic expansion, termwise differentiable by virtue of the equation itself. The block of a multiple root μ i \mu_i μ i has the form z ′ = ( μ i I + B ( u ) ) z z'=(\mu_i I+B(u))z z ′ = ( μ i I + B ( u )) z , where after a rescaling of the Jordan blocks ∥ B ( u ) ∥ ≤ ε \lVert B(u)\rVert\le\varepsilon ∥ B ( u )∥ ≤ ε for u ≥ u 0 ( ε ) u\ge u_0(\varepsilon) u ≥ u 0 ( ε ) , so that all its solutions satisfy ∣ z ( u ) ∣ = e μ i u + o ( u ) |z(u)|=e^{\mu_i u+o(u)} ∣ z ( u ) ∣ = e μ i u + o ( u ) (cf. the proof of Proposition 4). This suffices for all assertions of the lemma and their applications below: the fine structure of the modes of multiple roots (ramified modes with fractional powers of u u u in the exponent, whose exponents do not exceed 1 − 1 / n i 1-1/n_i 1 − 1/ n i , and logarithmic factors) is not used below, and the fact that inside the block of a multiple root the positive half-axis may be a ray of equal moduli for a pair of conjugate ramified modes does not affect the estimate e μ i u + o ( u ) e^{\mu_i u+o(u)} e μ i u + o ( u ) . For the equations of this paper the branching exponents are 1 − 2 / n i 1-2/n_i 1 − 2/ n i for n i ≥ 3 n_i\ge3 n i ≥ 3 , while for n i = 2 n_i=2 n i = 2 there is no branching (modes e μ i u u s e^{\mu_i u}u^s e μ i u u s with exponents s s s from a quadratic equation, and with a logarithmic factor if it has a multiple root). Indeed, the symbol of the equation for Φ \Phi Φ has the form u 2 p 2 ( μ ) + u p 1 ( μ ) + p 0 ( μ ) u^2p_2(\mu)+up_1(\mu)+p_0(\mu) u 2 p 2 ( μ ) + u p 1 ( μ ) + p 0 ( μ ) , where A ( μ ) A(\mu) A ( μ ) is the symbol of the annihilator (( ( 1 + θ μ ) k ((1+\theta\mu)^k (( 1 + θ μ ) k in (A), ( 1 − θ μ ) k (1-\theta\mu)^k ( 1 − θ μ ) k in (B), (42) in (C)). By the Leibniz rule p 2 = 1 2 σ 2 μ 2 A ( μ ) p_2=\frac{1}{2}\sigma^2\mu^2A(\mu) p 2 = 2 1 σ 2 μ 2 A ( μ ) and p 1 = ( μ a A ( μ ) + σ 2 μ A ′ ( μ ) ) p_1=(\mu aA(\mu)+\sigma^2\mu A'(\mu)) p 1 = ( μ a A ( μ ) + σ 2 μ A ′ ( μ )) , so that at a root μ i \mu_i μ i of the symbol A A A of multiplicity n i n_i n i the polynomial p 1 p_1 p 1 has a zero of order n i − 1 n_i-1 n i − 1 , whereas p 0 ( μ i ) ≠ 0 p_0(\mu_i)\ne0 p 0 ( μ i ) = 0 (for n i ≥ 3 n_i\ge3 n i ≥ 3 this is α \alpha α in the one-sided models and α 1 ( 1 + θ 2 μ i ) k 2 \alpha_1(1+\theta_2\mu_i)^{k_2} α 1 ( 1 + θ 2 μ i ) k 2 or α 2 ( 1 − θ 1 μ i ) k 1 \alpha_2(1-\theta_1\mu_i)^{k_1} α 2 ( 1 − θ 1 μ i ) k 1 in model (C), where α 1 , α 2 > 0 \alpha_1,\alpha_2>0 α 1 , α 2 > 0 is essential). The balance u 2 ( μ − μ i ) n i ∼ p 0 ( μ i ) u^2(\mu-\mu_i)^{n_i}\sim p_0(\mu_i) u 2 ( μ − μ i ) n i ∼ p 0 ( μ i ) gives μ − μ i ∼ u − 2 / n i \mu-\mu_i\sim u^{-2/n_i} μ − μ i ∼ u − 2/ n i , and the term u p 1 up_1 u p 1 is subdominant for n i ≥ 3 n_i\ge3 n i ≥ 3 . On the positive half-axis, the term with a smaller μ j \mu_j μ j is maximally subdominant relative to the one with a larger μ i \mu_i μ i , that is, the half-axis is a Stokes line in the sense of Borel summation (Remark 5). On a ray, a solution with a given asymptotic expansion is determined only up to terms subdominant on it: the difference of two solutions with the same expansion has zero power expansion and is therefore subdominant. The continuation to ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) is possible since the leading coefficient does not vanish there.
For the equations of models (A) and (B) the characteristic polynomials are μ ( 1 ± θ μ ) k \mu(1\pm\theta\mu)^k μ ( 1 ± θ μ ) k , the root μ = 0 \mu=0 μ = 0 is simple, and the next order of the dominant balance gives ρ = − p \rho=-p ρ = − p (§ 5.2 , and in model (B) the same computation goes through with the replacement θ ↦ − θ \theta\mapsto-\theta θ ↦ − θ in the coefficients of the annihilator). In model (C), the absence of the term with Φ ( 0 ) \Phi^{(0)} Φ ( 0 ) was noted after Theorem 3. The characteristic polynomial μ 2 A ( μ ) \mu^2A(\mu) μ 2 A ( μ ) of the equation for Φ \Phi Φ (Lemma 5), after passing to G = Φ ′ G=\Phi' G = Φ ′ , becomes μ A ( μ ) \mu A(\mu) μ A ( μ ) with the simple root μ = 0 \mu=0 μ = 0 , and the exponent − β − 1 -\beta-1 − β − 1 of the branch G p o w G_{\mathrm{pow}} G pow corresponds to the branch s = − β s=-\beta s = − β for Φ \Phi Φ (Proposition 3). We integrate the asymptotic expansion of G p o w G_{\mathrm{pow}} G pow over ( u , ∞ ) (u,\infty) ( u , ∞ ) (legitimate for β > 0 \beta>0 β > 0 ) and obtain (42). The function Φ β \Phi_\beta Φ β satisfies the equation for Φ \Phi Φ , because its derivative satisfies the equation for G G G . ▫
Proposition 4 (realization of the power branch). Let (A1)–(A2) hold, and, in model (C), also (A3). Then in each of the three models the coefficient c β c_{\beta} c β of the power branch Φ β \Phi_{\beta} Φ β in the expansion of the physical solution in the actual fundamental set of Lemma 6 is nonzero, and
Ψ ( u ) ∼ C k u − β , u → ∞ , C k = − c β > 0 , β = 2 a σ 2 − 1. (44) \Psi(u) \sim C_k u^{-\beta}, \qquad u \to\infty, \qquad C_k=-c_{\beta}>0, \qquad\beta=\frac{2a}{\sigma^2}-1. \tag*{(44)} Ψ ( u ) ∼ C k u − β , u → ∞ , C k = − c β > 0 , β = σ 2 2 a − 1. (44) In model (A) we also have, for each L L L , Ψ ( u ) = C k u − β ( ∑ j < L β β + j a j u − j + O ( u − L ) ) \Psi(u)=C_k u^{-\beta}\left(\sum_{j<L}\frac{\beta}{\beta+j}a_j u^{-j}+O(u^{-L})\right) Ψ ( u ) = C k u − β ( ∑ j < L β + j β a j u − j + O ( u − L ) ) , where the a j a_j a j are the coefficients of the formal series (for k = 2 k=2 k = 2 , from (34) ).
Proof. By Theorems 1–3 the function G = Φ ′ G=\Phi' G = Φ ′ satisfies the reduced equation for G G G on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and is therefore a linear combination of the actual fundamental set of Lemma 6 . Integrating, we find that Φ \Phi Φ is a linear combination of a constant, the branch Φ β \Phi_{\beta} Φ β (in models (A), (B) we define it by the same formula (43) from the corresponding G p o w G_{\mathrm{pow}} G pow ), and the exponential modes. The growing modes (e u / θ + o ( u ) e^{u/\theta+o(u)} e u / θ + o ( u ) in (B), e u / θ 1 + o ( u ) e^{u/\theta_1+o(u)} e u / θ 1 + o ( u ) in (C)) are excluded by the boundedness of Φ \Phi Φ .
To see this, we write the equation for Φ \Phi Φ (of order K + 2 K+2 K + 2 , with the same growing group as that of the equation for G G G ) as the system Y ′ = A ( u ) Y Y'=\mathcal{A}(u)Y Y ′ = A ( u ) Y for Y = ( Φ , Φ ′ , … , Φ ( K + 1 ) ) Y=(\Phi,\Phi',\ldots,\Phi^{(K+1)}) Y = ( Φ , Φ ′ , … , Φ ( K + 1 ) ) . The analytic block diagonalization in a sector containing the ray [32 Sects. 12–15, 18–19] , Y = Π ( u ) Z Y=\Pi(u)Z Y = Π ( u ) Z with Π ( u ) = Π 0 + O ( u − 1 ) \Pi(u)=\Pi_0+O(u^{-1}) Π ( u ) = Π 0 + O ( u − 1 ) , det Π 0 ≠ 0 \det\Pi_0\ne0 det Π 0 = 0 , separates the growing block z ′ = ( μ I + B ( u ) ) z z'=(\mu I+B(u))z z ′ = ( μ I + B ( u )) z with μ = 1 / θ \mu=1/\theta μ = 1/ θ (respectively 1 / θ 1 1/\theta_1 1/ θ 1 ), where after a suitable rescaling of the Jordan blocks ∥ B ( u ) ∥ ≤ ε \lVert B(u)\rVert\le\varepsilon ∥ B ( u )∥ ≤ ε for u ≥ u 0 ( ε ) u\ge u_0(\varepsilon) u ≥ u 0 ( ε ) . Then d d u ∣ z ∣ 2 = 2 ⟨ z , ( μ I + B ( u ) ) z ⟩ ≥ 2 ( μ − ε ) ∣ z ∣ 2 \frac{d}{du}|z|^2=2\langle z,(\mu I+B(u))z\rangle\ge2(\mu-\varepsilon)|z|^2 d u d ∣ z ∣ 2 = 2 ⟨ z , ( μ I + B ( u )) z ⟩ ≥ 2 ( μ − ε ) ∣ z ∣ 2 whence ∣ z ( u ) ∣ ≥ e ( μ − ε ) ( u − u 0 ) ∣ z ( u 0 ) ∣ |z(u)|\ge e^{(\mu-\varepsilon)(u-u_0)}|z(u_0)| ∣ z ( u ) ∣ ≥ e ( μ − ε ) ( u − u 0 ) ∣ z ( u 0 ) ∣ for every nonzero solution of the block, and hence ∣ Y ( u ) ∣ ≥ c e ( μ − ε ) u |Y(u)|\ge ce^{(\mu-\varepsilon)u} ∣ Y ( u ) ∣ ≥ c e ( μ − ε ) u for every solution Φ g \Phi_g Φ g with a nonzero growing component. On the other hand, if such Φ g \Phi_g Φ g were bounded on [ u 0 , ∞ ) [u_0,\infty) [ u 0 , ∞ ) , then so would its derivatives: the normalized coefficients of the equation are bounded for u ≥ u 0 u\ge u_0 u ≥ u 0 , and by the interpolation inequality on the intervals [ x , x + 1 ] [x,x+1] [ x , x + 1 ] (here ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ is the uniform norm on such an interval, and the constants do not depend on x x x ), ∥ Φ g ( j ) ∥ ≤ η ∥ Φ g ( K + 2 ) ∥ + C η ∥ Φ g ∥ \lVert\Phi_g^{(j)}\rVert\le\eta\lVert\Phi_g^{(K+2)}\rVert+C_\eta\lVert\Phi_g\rVert ∥ Φ g ( j ) ∥ ≤ η ∥ Φ g ( K + 2 ) ∥ + C η ∥ Φ g ∥ , 0 < j < K + 2 0<j<K+2 0 < j < K + 2 , so that, together with the equation (all the interval norms are finite), ∥ Φ g ( K + 2 ) ∥ ≤ C ∥ Φ g ∥ \lVert\Phi_g^{(K+2)}\rVert\le C\lVert\Phi_g\rVert ∥ Φ g ( K + 2 ) ∥ ≤ C ∥ Φ g ∥ with a constant independent of x x x — in contradiction with the growth of ∣ Y ∣ |Y| ∣ Y ∣ . Thus the growing component of a bounded solution vanishes, whatever the logarithmic factors and the complex exponents inside the block may be, and the bounded solutions are exactly the linear combinations of a constant, the branch Φ β \Phi_\beta Φ β , and the decaying modes. The decaying modes have the form e − u / θ + o ( u ) e^{-u/\theta+o(u)} e − u / θ + o ( u ) (in (C): e − u / θ 2 + o ( u ) e^{-u/\theta_2+o(u)} e − u / θ 2 + o ( u ) ) and are therefore O ( e − δ u ) O(e^{-\delta u}) O ( e − δ u ) for some δ > 0 \delta>0 δ > 0 . In all three models this yields
Φ = c 0 + c β Φ β + O ( e − δ u ) , \Phi=c_0+c_\beta\Phi_\beta+O(e^{-\delta u}), Φ = c 0 + c β Φ β + O ( e − δ u ) ,
and in model (B) there are no decaying modes and the remainder is absent. From Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 it follows that c 0 = 1 c_0=1 c 0 = 1 , i.e., Ψ = − c β Φ β + O ( e − δ u ) \Psi=-c_\beta\Phi_\beta+O(e^{-\delta u}) Ψ = − c β Φ β + O ( e − δ u ) . We still need to prove that c β ≠ 0 c_\beta\ne0 c β = 0 .
Model (B). The solutions bounded as u → ∞ u\to\infty u → ∞ form span { 1 , Φ β } \operatorname{span}\{1,\Phi_\beta\} span { 1 , Φ β } , and Φ = 1 + c β Φ β \Phi=1+c_\beta\Phi_\beta Φ = 1 + c β Φ β . If c β = 0 c_\beta=0 c β = 0 , then Φ ≡ 1 \Phi\equiv1 Φ ≡ 1 , which contradicts Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 (and Ψ > 0 \Psi>0 Ψ > 0 ). Thus c β ≠ 0 c_\beta\ne0 c β = 0 .
::: Models (A), (C). If c β = 0 c_{\beta}=0 c β = 0 , then Ψ = O ( e − δ u ) \Psi=O(e^{-\delta u}) Ψ = O ( e − δ u ) . This contradicts the power-law lower bound [17 Proposition 4.1] : if c < 0 c<0 c < 0 or the downward jumps are unbounded, then there exists a β ∗ > 0 \beta_{*}>0 β ∗ > 0 such that lim inf u → ∞ u β ∗ Ψ ( u ) > 0 \liminf_{u\to\infty}u^{\beta_{*}}\Psi(u)>0 lim inf u → ∞ u β ∗ Ψ ( u ) > 0 . In model (C) the conditions of that proposition are satisfied for any c c c , including c = 0 c=0 c = 0 : the risk process is a compound Poisson process with drift and jumps of both signs, as in [17 ] , and the downward jumps ξ ( 2 ) ∼ Γ ( k 2 , θ 2 ) \xi^{(2)}\sim\Gamma(k_{2},\theta_{2}) ξ ( 2 ) ∼ Γ ( k 2 , θ 2 ) are unbounded.
Model (A) does not satisfy the standard assumption of [17 ] about jumps of both signs, but the proof of Proposition 4.1 of [17 ] does not use upward jumps: it relies only on the fact that a jump together with the drift contribution before it is not bounded from below (in model (A) this is provided by the unbounded claims ξ ∼ Γ ( k , θ ) \xi\sim\Gamma(k,\theta) ξ ∼ Γ ( k , θ ) ), and on the Brownian factor between the jumps. Therefore it carries over verbatim to model (A). For jumps with a finite moment of order β \beta β we also obtain, from [14 Theorem 1.1] , 0 < lim inf u → ∞ u β Ψ ( u ) ≤ lim sup u → ∞ u β Ψ ( u ) < ∞ 0<\liminf_{u\to\infty}u^{\beta}\Psi(u)\leq\limsup_{u\to\infty}u^{\beta}\Psi(u)<\infty 0 < lim inf u → ∞ u β Ψ ( u ) ≤ lim sup u → ∞ u β Ψ ( u ) < ∞ . Power-law lower bounds were also obtained in [18 ] (general models with investments) and, for model (A), in [7 ] (exponential jumps) and [26 ] .
The sign and the expansion. From Ψ ≥ 0 \Psi\geq0 Ψ ≥ 0 and Φ β > 0 \Phi_{\beta}>0 Φ β > 0 for large u u u it follows that − c β ≥ 0 -c_{\beta}\geq0 − c β ≥ 0 , hence C k : = − c β > 0 C_{k}:=-c_{\beta}>0 C k := − c β > 0 in all three models, which gives (44) . In model (A), the termwise integration of the asymptotic expansion of G p o w G_{\mathrm{pow}} G pow over ( u , ∞ ) (u,\infty) ( u , ∞ ) , legitimate for power-law asymptotics with exponents p + j > 1 p+j>1 p + j > 1 , gives Φ β = β ∑ j < L a j β + j u − β − j + O ( u − β − L ) \Phi_{\beta}=\beta\sum_{j<L}\frac{a_{j}}{\beta+j}u^{-\beta-j}+O(u^{-\beta-L}) Φ β = β ∑ j < L β + j a j u − β − j + O ( u − β − L ) , whence the complete expansion follows, and the amplitude ϖ \varpi ϖ of Theorem 4 equals β C k \beta C_{k} β C k . □ \square □
For one-sided models the asymptotics Ψ ( u ) ∼ C u − β \Psi(u)\sim Cu^{-\beta} Ψ ( u ) ∼ C u − β , C > 0 C>0 C > 0 , is known for general claim distributions: in model (A) under E [ ξ β ] < ∞ \mathbb{E}[\xi^{\beta}]<\infty E [ ξ β ] < ∞ [29 Theorem 1] , in model (B) under E [ ξ ] < ∞ \mathbb{E}[\xi]<\infty E [ ξ ] < ∞ [27 Theorem 1] , and, under E [ ξ β ] < ∞ \mathbb{E}[\xi^{\beta}]<\infty E [ ξ β ] < ∞ , [14 Theorem 1.1] . Proposition 4 gives for them (for Erlang-distributed claims) a different proof of the asymptotic part. The smoothness, the value of Φ ( 0 + ) \Phi(0+) Φ ( 0 + ) , and the lower bound are taken from [17 , 27 , 29 ] . For the mixed model the asymptotics with C > 0 C>0 C > 0 are also known: they follow from [3 Theorem 2.1] (the Cramér-type condition on the logarithm of the price of the risky asset with root β \beta β , the moment conditions, and the absolute continuity of its law are satisfied, see also [22 ] and, for independent Lévy processes, [25 Theorem 3.2] ), whereas for a general Lévy return process with jumps of both signs, only two-sided estimates of order u − β u^{-\beta} u − β were established in [14 Theorem 1.1] . Analytically it was obtained for k 1 = k 2 = 1 k_{1}=k_{2}=1 k 1 = k 2 = 1 [17 ] and, for rational Laplace transforms, for β ∈ ( 0 , 1 ) \beta\in(0,1) β ∈ ( 0 , 1 ) , c ≠ 0 c\ne0 c = 0 [1 ] . Proposition 4 gives an analytical proof for all k 1 , k 2 k_{1},k_{2} k 1 , k 2 , β > 0 \beta>0 β > 0 , c ∈ R c\in\mathbb{R} c ∈ R , in which, apart from smoothness and the boundary condition Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 (Sect. 2), only the power-law lower bound [17 Proposition 4.1] is taken from probabilistic results (the nonvanishing of c β c_{\beta} c β ). The new element is the complete expansion up to a constant factor C C C in all three models (in model (A) with k = 1 k=1 k = 1 it was obtained in [30 ] ).
Theorem 5 (the physical solution and its tail, model (C)) Under assumptions (A1)–(A3) the physical survival function Ψ \Psi Ψ of model (C) is singled out in terms of the actual fundamental set (Lemmas 5, 6) as follows: the coefficients of all k 1 k_{1} k 1 growing branches e u / θ 1 + o ( u ) e^{u/\theta_{1}+o(u)} e u / θ 1 + o ( u ) vanish (the boundedness of $\Phi),
the coefficient of the constant branch is 1 (the condition Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 ), and the coefficient c β c_\beta c β of the power branch is nonzero (Proposition 4), so that
Ψ ( u ) ∼ C u − β , u → ∞ , C = − c β > 0 , β = 2 a σ 2 − 1 , (45) \Psi(u) \sim C u^{-\beta}, \qquad u \to\infty, \qquad C=-c_\beta>0, \qquad\beta=\frac{2a}{\sigma^2}-1, \tag*{(45)} Ψ ( u ) ∼ C u − β , u → ∞ , C = − c β > 0 , β = σ 2 2 a − 1 , (45)
with the same exponent as in the one-sided models, while the contribution of the k 2 k_2 k 2 decaying branches is O ( e − ( 1 / θ 2 − ε ) u ) O(e^{-(1/\theta_2-\varepsilon)u}) O ( e − ( 1/ θ 2 − ε ) u ) for any ε > 0 \varepsilon>0 ε > 0 .
Proof. The growing branches e u / θ 1 + o ( u ) e^{u/\theta_1+o(u)} e u / θ 1 + o ( u ) are excluded by the boundedness of Φ ∈ [ 0 , 1 ] \Phi\in[0,1] Φ ∈ [ 0 , 1 ] — this is the feature distinguishing models (B) and (C) from model (A): the upward-jump operator I k 1 \mathcal{I}_{k_1} I k 1 generates k 1 k_1 k 1 growing modes, and all of them are cut off. After their exclusion, Φ = c 0 + c β Φ β + O ( e − ( 1 / θ 2 − ε ) u ) \Phi=c_0+c_\beta\Phi_\beta+O(e^{-(1/\theta_2-\varepsilon)u}) Φ = c 0 + c β Φ β + O ( e − ( 1/ θ 2 − ε ) u ) (for a multiplicity k 2 ≥ 2 k_2\ge2 k 2 ≥ 2 the decaying modes contain ramified corrections e o ( u ) e^{o(u)} e o ( u ) , hence the ε \varepsilon ε in the remainder). The condition Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 gives c 0 = 1 c_0=1 c 0 = 1 , and then Ψ = 1 − Φ = − c β u − β ( 1 + o ( 1 ) ) + O ( e − ( 1 / θ 2 − ε ) u ) \Psi=1-\Phi=-c_\beta u^{-\beta}(1+o(1))+O(e^{-(1/\theta_2-\varepsilon)u}) Ψ = 1 − Φ = − c β u − β ( 1 + o ( 1 )) + O ( e − ( 1/ θ 2 − ε ) u ) . From Ψ ≥ 0 \Psi\ge0 Ψ ≥ 0 and Φ β > 0 \Phi_\beta>0 Φ β > 0 for large u u u it follows that c β ≤ 0 c_\beta\le0 c β ≤ 0 , while c β ≠ 0 c_\beta\ne0 c β = 0 is given by Proposition 4, which yields (44). The reduced ODE expresses only a necessary condition (Remark 2): its ( k 1 + k 2 + 2 ) (k_1+k_2+2) ( k 1 + k 2 + 2 ) -dimensional solution space is wider than that of the original IDE, and the data at infinity (the boundedness, Φ ( ∞ ) = 1 \Phi(\infty)=1 Φ ( ∞ ) = 1 ) fix only k 1 + 1 k_1+1 k 1 + 1 degrees of freedom. The IDE itself determines which of the k 2 k_2 k 2 subdominant modes enter the physical Φ \Phi Φ and what the amplitude c β c_\beta c β is: this is fixed by at most k 2 k_2 k 2 effective conditions (27) of Proposition 1 (Remark 3) and by the behavior at u = 0 u=0 u = 0 : for c < 0 c<0 c < 0 by the condition Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 , for c > 0 c>0 c > 0 by boundedness, which excludes the mode (46) (Proposition 5), and for c = 0 c=0 c = 0 by boundedness at zero (Remark 7). □
As in model (A), we obtain a complete asymptotic expansion up to a constant factor C C C . The root μ = 0 \mu=0 μ = 0 is simple, so the formal Thomé series G p o w = u − β − 1 ∑ j ≥ 0 a j u − j G_{\mathrm{pow}}=u^{-\beta-1}\sum_{j\ge0}a_j u^{-j} G pow = u − β − 1 ∑ j ≥ 0 a j u − j with a 0 = 1 a_0=1 a 0 = 1 and coefficients uniquely computed by the recurrence of the equation for G = Φ ′ G=\Phi' G = Φ ′ (in normalized form the factor multiplying a j a_j a j in it equals − j / ( ( − θ 1 ) k 1 θ 2 k 2 ) ≠ 0 -j/((- \theta_1)^{k_1}\theta_2^{k_2})\ne0 − j / (( − θ 1 ) k 1 θ 2 k 2 ) = 0 for j ≥ 1 j\ge1 j ≥ 1 , as in Proposition 2) is its asymptotic expansion (Lemma 6 ), and the termwise integration, as in the proof of Proposition 4, gives Ψ ( u ) = C u − β ( 1 + ∑ 1 ≤ j < L β a j β + j u − j ) + O ( u − β − L ) \Psi(u)=Cu^{-\beta}(1+\sum_{1\le j<L}\frac{\beta a_j}{\beta+j}u^{-j})+O(u^{-\beta-L}) Ψ ( u ) = C u − β ( 1 + ∑ 1 ≤ j < L β + j β a j u − j ) + O ( u − β − L ) for any L ≥ 1 L\ge1 L ≥ 1 . The same is true in model (B), where Ψ = − c β Φ β \Psi=-c_\beta\Phi_\beta Ψ = − c β Φ β exactly.
The singular point u = 0 u=0 u = 0 in the mixed model
Proposition 5 (behavior at the singular point u = 0 u=0 u = 0 , model (C)). Let c ≠ 0 c\ne0 c = 0 . The point u = 0 u=0 u = 0 is an irregular singular point of rank 1 of the reduced ODE (25): the leading coefficient is proportional to u 2 u^2 u 2 , while the coefficient of Φ ( M − 1 ) \Phi^{(M-1)} Φ ( M − 1 ) , M : = k 1 + k 2 + 2 M:=k_1+k_2+2 M := k 1 + k 2 + 2 , is nonzero at zero. Formal power solutions u ν ∑ n ≥ 0 b n u n u^\nu\sum_{n\ge0}b_nu^n u ν ∑ n ≥ 0 b n u n exist for ν = 0 , 1 , … , k 1 + k 2 \nu=0,1,\ldots,k_1+k_2 ν = 0 , 1 , … , k 1 + k 2 — the lowest-order condition is the indicial equation
ν ( ν − 1 ) ⋯ ( ν − ( k 1 + k 2 ) ) = 0 , (46) \nu(\nu-1)\cdots\bigl(\nu-(k_1+k_2)\bigr)=0, \tag*{(46)} ν ( ν − 1 ) ⋯ ( ν − ( k 1 + k 2 ) ) = 0 , (46)
and they form a ( k 1 + k 2 + 1 ) (k_1+k_2+1) ( k 1 + k 2 + 1 ) -parameter family. The convergent series form a subspace of it whose dimension is at least K = k 1 + k 2 K=k_1+k_2 K = k 1 + k 2 and which contains the constant. The dimension equals K K K exactly when at least one formal series diverges. In that case (the factorially growing solution of the recurrence gives ∣ b n + 1 / b n ∣ ∼ n σ 2 / ( 2 ∣ c ∣ ) |b_{n+1}/b_n|\sim n^{\sigma2}/(2|c|) ∣ b n + 1 / b n ∣ ∼ n σ 2 / ( 2∣ c ∣ ) , a Gevrey class of order 1) the divergent tails of all formal solutions are proportional. In either case, since K = k 1 + k 2 ≥ 2 K=k_1+k_2\ge2 K = k 1 + k 2 ≥ 2 , there exist nonconstant solutions analytic at zero (and entire, because the equation has no other finite singular points). The solution corresponding to the formal solution of exponential type has an essential singularity at zero,
Φ i r r ( u ) = u 1 − β exp ( 2 c σ 2 u ) ( 1 + o ( 1 ) ) , u → 0 + . (47) \Phi_{\mathrm{irr}}(u)=u^{1-\beta}\exp\left(\frac{2c}{\sigma^2u}\right)(1+o(1)),\qquad u\to0+. \tag*{(47)} Φ irr ( u ) = u 1 − β exp ( σ 2 u 2 c ) ( 1 + o ( 1 )) , u → 0 + . (47)
For c > 0 c>0 c > 0 it grows without bound and is excluded by the boundedness of Φ \Phi Φ , while for c < 0 c<0 c < 0 it is exponentially small together with all its derivatives and is not ruled out by any regularity condition. The limit u → 0 + u\to0+ u → 0 + in the IDE (7) gives the boundary relation
c Φ ′ ( 0 + ) = ( α 1 + α 2 ) Φ ( 0 + ) − α 1 I k 1 Φ ( 0 ) . (48) c\Phi'(0+)=(\alpha_1+\alpha_2)\Phi(0+)-\alpha_1 I_{k_1}\Phi(0). \tag*{(48)} c Φ ′ ( 0 + ) = ( α 1 + α 2 ) Φ ( 0 + ) − α 1 I k 1 Φ ( 0 ) . (48)
For c < 0 c<0 c < 0 we have Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 (property (iii) of Sect. 2), and then (48) determines Φ ′ ( 0 + ) = α 1 I k 1 Φ ( 0 ) / ∣ c ∣ > 0 \Phi'(0+)=\alpha_1I_{k_1}\Phi(0)/|c|>0 Φ ′ ( 0 + ) = α 1 I k 1 Φ ( 0 ) /∣ c ∣ > 0 . For c > 0 c>0 c > 0 , from (48) , Φ ′ ≥ 0 \Phi'\ge0 Φ ′ ≥ 0 and α 1 > 0 \alpha_1>0 α 1 > 0 it follows that Φ ( 0 + ) ≥ α 1 I k 1 Φ ( 0 ) / ( α 1 + α 2 ) > 0 \Phi(0+)\ge\alpha_1I_{k_1}\Phi(0)/(\alpha_1+\alpha_2)>0 Φ ( 0 + ) ≥ α 1 I k 1 Φ ( 0 ) / ( α 1 + α 2 ) > 0 .
Proof. The leading coefficient of the ODE (26) is ( − 1 ) k 1 1 2 σ 2 θ 1 k 1 θ 2 k 2 u 2 (-1)^{k_1}\frac{1}{2}\sigma^2\theta_1^{k_1}\theta_2^{k_2}u^2 ( − 1 ) k 1 2 1 σ 2 θ 1 k 1 θ 2 k 2 u 2 , and the coefficient of Φ ( M − 1 ) \Phi^{(M-1)} Φ ( M − 1 ) at zero is ( − 1 ) k 1 c θ 1 k 1 θ 2 k 2 (-1)^{k_1}c\theta_1^{k_1}\theta_2^{k_2} ( − 1 ) k 1 c θ 1 k 1 θ 2 k 2 (the contribution of the drift ( a u + c ) Φ ′ (au+c)\Phi' ( a u + c ) Φ ′ through the leading term of the annihilator). They share the sign factor. The normalized coefficient of Φ ( M − 1 ) \Phi^{(M-1)} Φ ( M − 1 ) has a second-order pole at zero. Hence, by the Fuchs criterion, the point is irregular, and after the substitution t = 1 / u t=1/u t = 1/ u the exponential factor is linear in t t t , i.e., the rank is 1. Substituting Φ ∼ u ν \Phi\sim u^\nu Φ ∼ u ν , we find the lowest order u ν − ( M − 1 ) u^{\nu-(M-1)} u ν − ( M − 1 ) given by a single term c Φ ( M − 1 ) c\Phi^{(M-1)} c Φ ( M − 1 ) , which yields (46) . For ν ∈ { 0 , … , k 1 + k 2 } \nu\in\{0,\ldots,k_1+k_2\} ν ∈ { 0 , … , k 1 + k 2 } the coefficients b 0 , … , b k 1 + k 2 b_0,\ldots,b_{k_1+k_2} b 0 , … , b k 1 + k 2 are free, while for n > k 1 + k 2 n>k_1+k_2 n > k 1 + k 2 the principal part of the recurrence has the form c n M − 1 b n = 1 2 σ 2 ( n − 1 ) M b n − 1 + ⋯ c\,n^{M-1}b_n=\frac{1}{2}\sigma^2(n-1)^M b_{n-1}+\cdots c n M − 1 b n = 2 1 σ 2 ( n − 1 ) M b n − 1 + ⋯ (x m x^m x m is the falling factorial), whence b n / b n − 1 ∼ − n σ 2 / ( 2 c ) b_n/b_{n-1}\sim-n\sigma^2/(2c) b n / b n − 1 ∼ − n σ 2 / ( 2 c ) for solutions of the recurrence that contain the factorially growing component. The constant Φ ≡ 1 \Phi\equiv1 Φ ≡ 1 is an exact solution, because the term in Φ ( 0 ) \Phi^{(0)} Φ ( 0 ) is absent. The dimension of the subspace of convergent series is given by the Malgrange index theorem [21 ] : if T = ∑ i = 0 M t i ( u ) D i \mathcal{T}=\sum_{i=0}^{M}t_i(u)D^i T = ∑ i = 0 M t i ( u ) D i is an operator with coefficients holomorphic at zero, t M ≢ 0 t_M\not\equiv0 t M ≡ 0 , and ord t i \operatorname{ord}t_i ord t i is the order of the zero of t i t_i t i at u = 0 u=0 u = 0 (ord 0 = + ∞ \operatorname{ord}0=+\infty ord 0 = + ∞ ), then T \mathcal{T} T is Fredholm both on C [ [ u ] ] \mathbb{C}[[u]] C [[ u ]] and on the convergent power series C { u } \mathbb{C}\{u\} C { u } , with the indices
χ ( T , C [ [ u ] ] ) = max 0 ≤ i ≤ M ( i − ord t i ) , χ ( T , C { u } ) = M − ord t M , \chi(\mathcal{T},\mathbb{C}[[u]])=\max_{0\le i\le M}(i-\operatorname{ord}t_i),\qquad
\chi(\mathcal{T},\mathbb{C}\{u\})=M-\operatorname{ord}t_M, χ ( T , C [[ u ]]) = 0 ≤ i ≤ M max ( i − ord t i ) , χ ( T , C { u }) = M − ord t M ,
and their difference equals the irregularity of the point u = 0 u=0 u = 0 . For the operator of equation (26) we have ord t M = 2 \operatorname{ord}t_M=2 ord t M = 2 , ord t M − 1 = 0 \operatorname{ord}t_{M-1}=0 ord t M − 1 = 0 (the coefficient ( − 1 ) k 1 c θ 1 k 1 θ 2 k 2 ≠ 0 (-1)^{k_1}c\theta_1^{k_1}\theta_2^{k_2}\ne0 ( − 1 ) k 1 c θ 1 k 1 θ 2 k 2 = 0 for c ≠ 0 c\ne0 c = 0 ) and ord t i ≥ 0 \operatorname{ord}t_i\ge0 ord t i ≥ 0 for i < M − 1 i<M-1 i < M − 1 , so that χ ( T , C [ [ u ] ] ) = M − 1 = k 1 + k 2 + 1 \chi(\mathcal{T},\mathbb{C}[[u]])=M-1=k_1+k_2+1 χ ( T , C [[ u ]]) = M − 1 = k 1 + k 2 + 1 , χ ( T , C { u } ) = M − 2 = k 1 + k 2 \chi(\mathcal{T},\mathbb{C}\{u\})=M-2=k_1+k_2 χ ( T , C { u }) = M − 2 = k 1 + k 2 , and the irregularity equals 1 (the unique exponential mode (46) of rank 1, and the Newton polygon has one sloping edge of length 1 and slope 1). This agrees with a direct count: on C [ [ u ] ] \mathbb{C}[[u]] C [[ u ]] we find the operator surjective (the coefficient c n M − 1 c n^{M-1} c n M − 1 of the leading unknown b n b_n b n is nonzero for n ≥ M − 1 n \ge M-1 n ≥ M − 1 ), and its kernel has dimension k 1 + k 2 + 1 k_1+k_2+1 k 1 + k 2 + 1 . The dimension of the kernel on C { u } \mathbb{C}\{u\} C { u } is therefore at least the index k 1 + k 2 k_1+k_2 k 1 + k 2 . It equals k 1 + k 2 k_1+k_2 k 1 + k 2 exactly when the kernel on C { u } \mathbb{C}\{u\} C { u } is strictly smaller than the kernel on C [ [ u ] ] \mathbb{C}[[u]] C [[ u ]] , that is, when at least one formal series diverges. In this case the divergent tails of all formal solutions are proportional to a single factorially growing tail, and the divergence is governed by one linear functional of the free parameters b 0 , … , b k 1 + k 2 b_0,\ldots,b_{k_1+k_2} b 0 , … , b k 1 + k 2 (the Stokes multiplier). An index count alone cannot rule out the vanishing of the Stokes multiplier: when the irregularity equals 1 all formal series may converge (as for the operator ( u 2 D + 1 ) D (u^2D+1)D ( u 2 D + 1 ) D ). For (25) this would mean the convergence of all k 1 + k 2 + 1 k_1+k_2+1 k 1 + k 2 + 1 formal series. At none of the tested parameter values is this observed (for k = 1 k=1 k = 1 in model (A) the divergence for generic parameters was noted in [30 ] ). However, we do not have a proof for all parameter values (for fixed k 1 k_1 k 1 , k 2 k_2 k 2 and the sign of c c c , the set of parameters at which all the series converge is the zero set of the Stokes multiplier, apparently analytic in the parameters, and the fact that it is a proper subset of this component would follow from a single rigorously established example of divergence in it).
For the solution with an essential singularity we substitute Φ = e S \Phi=e^S Φ = e S with S = λ / u + m ln u S=\lambda/u+m\ln u S = λ / u + m ln u . The two leading terms 1 2 u 2 Φ ( M ) + c Φ ( M − 1 ) \frac{1}{2}u^2\Phi^{(M)}+c\Phi^{(M-1)} 2 1 u 2 Φ ( M ) + c Φ ( M − 1 ) give, in the most singular order u 2 − 2 M u^{2-2M} u 2 − 2 M , a factor proportional to λ M − 1 ( 2 c − λ σ 2 ) \lambda^{M-1}(2c-\lambda\sigma^2) λ M − 1 ( 2 c − λ σ 2 ) , whence λ = 2 c / σ 2 \lambda=2c/\sigma^2 λ = 2 c / σ 2 . Equivalently, for y = Φ ( M − 1 ) y=\Phi^{(M-1)} y = Φ ( M − 1 ) the balance 1 2 σ 2 u 2 y ′ + c y = 0 \frac{1}{2}\sigma^2u^2y'+cy=0 2 1 σ 2 u 2 y ′ + cy = 0 integrates exactly: y = exp ( 2 c / ( σ 2 u ) ) y=\exp(2c/(\sigma^2u)) y = exp ( 2 c / ( σ 2 u )) . The next order u 3 − 2 M u^{3-2M} u 3 − 2 M receives contributions from the term m u − 1 m u^{-1} m u − 1 of S ′ S' S ′ , the terms ( M 2 ) S ′ M ′ ′ \binom{M}{2}S'M^{\prime\prime} ( 2 M ) S ′ M ′′ and ( M − 1 2 ) S ′ M ′ ′ \binom{M-1}{2}S'M^{\prime\prime} ( 2 M − 1 ) S ′ M ′′ of the expansions of Φ ( M ) / Φ \Phi^{(M)}/\Phi Φ ( M ) /Φ and Φ ( M − 1 ) / Φ \Phi^{(M-1)}/\Phi Φ ( M − 1 ) /Φ , the term ( M − 2 ) σ 2 u Φ ( M − 1 ) (M-2)\sigma^2u\Phi^{(M-1)} ( M − 2 ) σ 2 u Φ ( M − 1 ) from the Leibniz rule, and the drift a u Φ ( M − 1 ) au\Phi^{(M-1)} a u Φ ( M − 1 ) . Their sum vanishes at m = 2 − 2 a / σ 2 = 1 − β m=2-2a/\sigma^2=1-\beta m = 2 − 2 a / σ 2 = 1 − β , independently of k 1 k_1 k 1 , k 2 k_2 k 2 , which gives (46). The relation (47) is the limit of (7) as u → 0 + u\to0+ u → 0 + : by property (iv) of Sect. 2, the derivatives Φ ′ \Phi' Φ ′ , Φ ′ ′ \Phi'' Φ ′′ are bounded, so that 1 2 σ 2 u 2 Φ ′ ′ → 0 \frac{1}{2}\sigma^2u^2\Phi''\to0 2 1 σ 2 u 2 Φ ′′ → 0 and a u Φ ′ → 0 au\Phi'\to0 a u Φ ′ → 0 , while J k 2 Φ ( u ) → 0 J_{k_2}\Phi(u)\to0 J k 2 Φ ( u ) → 0 . The limit Φ ′ ( 0 + ) \Phi'(0+) Φ ′ ( 0 + ) exists because Φ ′ ′ \Phi'' Φ ′′ is bounded (property (iv), c ≠ 0 c\ne0 c = 0 ), so Φ ′ \Phi' Φ ′ is Lipschitz on ( 0 , 1 ) (0,1) ( 0 , 1 ) . For c > 0 c>0 c > 0 the growing mode (46) is excluded by boundedness. The assertions about Φ ( 0 + ) \Phi(0+) Φ ( 0 + ) follow from (47) immediately: for c > 0 c>0 c > 0 and Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 we would have c Φ ′ ′ ( 0 + ) = − α 1 J k 1 Φ ( 0 ) < 0 c\Phi''(0+)=-\alpha_1J_{k_1}\Phi(0)<0 c Φ ′′ ( 0 + ) = − α 1 J k 1 Φ ( 0 ) < 0 contrary to Φ ′ ′ ≥ 0 \Phi''\ge0 Φ ′′ ≥ 0 .
Remark 7 For c = 0 c=0 c = 0 the coefficient of Φ ( M − 1 ) \Phi^{(M-1)} Φ ( M − 1 ) vanishes at u = 0 u=0 u = 0 , the Fuchs condition holds, and the point u = 0 u=0 u = 0 becomes a regular singular one: all solutions are convergent Frobenius series (possibly with logarithms), and the mode (46) is absent. The indicial equation has the roots 0 , 1 , … , k 1 + k 2 − 1 0,1,\ldots,k_1+k_2-1 0 , 1 , … , k 1 + k 2 − 1 and two roots ν − < 0 < ν + \nu_-<0<\nu_+ ν − < 0 < ν + of the quadratic trinomial σ 2 ν 2 + ( 2 a − σ 2 ) ν − 2 ( α 1 + α 2 ) \sigma^2\nu^2+(2a-\sigma^2)\nu-2(\alpha_1+\alpha_2) σ 2 ν 2 + ( 2 a − σ 2 ) ν − 2 ( α 1 + α 2 ) . The root ν − \nu_- ν − corresponds to the branch u ν − u^{\nu_-} u ν − , unbounded at the origin and therefore excluded. The physical solution is bounded and is therefore a combination of the Frobenius branches with exponents 0 , 1 , … , k 1 + k 2 − 1 0,1,\ldots,k_1+k_2-1 0 , 1 , … , k 1 + k 2 − 1 and ν + \nu_+ ν + : Φ ( u ) = Φ ( 0 + ) + O ( u min ( 1 , ν + ) ∣ ln u ∣ ) \Phi(u)=\Phi(0+)+O(u^{\min(1,\nu_+)}|\ln u|) Φ ( u ) = Φ ( 0 + ) + O ( u m i n ( 1 , ν + ) ∣ ln u ∣ ) . After termwise differentiation of the Frobenius expansions, u Φ ′ ( u ) → 0 u\Phi'(u)\to0 u Φ ′ ( u ) → 0 and u 2 Φ ′ ′ ( u ) → 0 u^2\Phi''(u)\to0 u 2 Φ ′′ ( u ) → 0 as u → 0 + u\to0+ u → 0 + . The limit in (7) gives ( α 1 + α 2 ) Φ ( 0 + ) = α 1 I k 1 Φ ( 0 ) (\alpha_1+\alpha_2)\Phi(0+) = \alpha_1 I_{k_1}\Phi(0) ( α 1 + α 2 ) Φ ( 0 + ) = α 1 I k 1 Φ ( 0 ) , i.e., Φ ( 0 + ) = α 1 I k 1 Φ ( 0 ) / ( α 1 + α 2 ) > 0 \Phi(0+) = \alpha_1 I_{k_1}\Phi(0)/(\alpha_1+\alpha_2) > 0 Φ ( 0 + ) = α 1 I k 1 Φ ( 0 ) / ( α 1 + α 2 ) > 0 : survival from zero initial capital is possible only if the first jump is directed upward. The first part of property (iv), the boundedness of the derivatives, does not apply here: for ν + < 1 \nu_+ < 1 ν + < 1 , i.e., for a > α 1 + α 2 a > \alpha_1+\alpha_2 a > α 1 + α 2 , the derivatives Φ ′ \Phi' Φ ′ , Φ ′ ′ \Phi'' Φ ′′ are unbounded near zero. For c = 0 c = 0 c = 0 the IDE (7) is then an Euler-type equation 1 2 σ 2 u 2 Φ ′ ′ + a u Φ ′ − α ˉ Φ = − h \frac{1}{2}\sigma^2u^2\Phi''+au\Phi'-\bar{\alpha}\Phi=-h 2 1 σ 2 u 2 Φ ′′ + a u Φ ′ − α ˉ Φ = − h with α ˉ : = α 1 + α 2 \bar{\alpha} := \alpha_1+\alpha_2 α ˉ := α 1 + α 2 and h : = α 1 I k 1 Φ + α 2 J k 2 Φ h := \alpha_1 I_{k_1}\Phi+\alpha_2 J_{k_2}\Phi h := α 1 I k 1 Φ + α 2 J k 2 Φ . Here h h h is bounded and nondecreasing, as a sum of convolutions of a nondecreasing function Φ \Phi Φ . For u > 0 u > 0 u > 0 the inequality h ( u ) − h ( 0 ) ≥ α 2 J k 2 Φ ( u ) > 0 h(u)-h(0) \ge\alpha_2J_{k_2}\Phi(u) > 0 h ( u ) − h ( 0 ) ≥ α 2 J k 2 Φ ( u ) > 0 holds and, by Lemma 1, h ′ = α 1 ( I k 1 − I k 1 − 1 ) / θ 1 + α 2 ( J k 2 − 1 − J k 2 ) / θ 2 h'=\alpha_1(I_{k_1}-I_{k_1-1})/\theta_1+\alpha_2(J_{k_2-1}-J_{k_2})/\theta_2 h ′ = α 1 ( I k 1 − I k 1 − 1 ) / θ 1 + α 2 ( J k 2 − 1 − J k 2 ) / θ 2 , so that ∣ h ′ ∣ ≤ α 1 / θ 1 + α 2 / θ 2 |h'| \le\alpha_1/\theta_1+\alpha_2/\theta_2 ∣ h ′ ∣ ≤ α 1 / θ 1 + α 2 / θ 2 since 0 ≤ Φ ≤ 1 0 \le\Phi\le1 0 ≤ Φ ≤ 1 , and h h h is Lipschitz (the boundedness of Φ ′ \Phi' Φ ′ is not needed for this). The unique bounded solution of this Euler-type equation is given by
Φ ( u ) = 2 σ 2 ( ν + − ν − ) [ u ν − ∫ 0 u x − ν − − 1 h ( x ) d x + u ν + ∫ u ∞ x − ν + − 1 h ( x ) d x ] . \Phi(u)=\frac{2}{\sigma^2(\nu_+-\nu_-)}\left[u^{\nu_-}\int_0^u x^{-\nu_--1}h(x)\,dx+u^{\nu_+}\int_u^\infty x^{-\nu_+-1}h(x)\,dx\right]. Φ ( u ) = σ 2 ( ν + − ν − ) 2 [ u ν − ∫ 0 u x − ν − − 1 h ( x ) d x + u ν + ∫ u ∞ x − ν + − 1 h ( x ) d x ] .
whence, taking into account − ν − ν + = 2 α ˉ / σ 2 -\nu_-\nu_+=2\bar{\alpha}/\sigma^2 − ν − ν + = 2 α ˉ / σ 2 , Φ ( u ) = h ( 0 ) / α ˉ + 2 H σ 2 ( ν + − ν − ) u ν + + O ( u ) \Phi(u)=h(0)/\bar{\alpha}+\frac{2H}{\sigma^2(\nu_+-\nu_-)}u^{\nu_+}+O(u) Φ ( u ) = h ( 0 ) / α ˉ + σ 2 ( ν + − ν − ) 2 H u ν + + O ( u ) , where H = ∫ 0 ∞ x − ν + − 1 ( h ( x ) − h ( 0 ) ) d x > 0 H=\int_0^\infty x^{-\nu_+-1}(h(x)-h(0))\,dx>0 H = ∫ 0 ∞ x − ν + − 1 ( h ( x ) − h ( 0 )) d x > 0 for ν + < 1 \nu_+<1 ν + < 1 . This expansion also gives the value Φ ( 0 + ) = h ( 0 ) / α ˉ \Phi(0+)=h(0)/\bar{\alpha} Φ ( 0 + ) = h ( 0 ) / α ˉ . For ν + = 1 \nu_+=1 ν + = 1 the growth of Φ ′ \Phi' Φ ′ is logarithmic, and for 1 < ν + < 2 1<\nu_+<2 1 < ν + < 2 , Φ ′ \Phi' Φ ′ stays bounded, while Φ ′ ′ \Phi'' Φ ′′ , in general, does not.
Remark 8 In Theorem 5, after excluding the growing modes and normalizing c 0 = 1 c_0=1 c 0 = 1 , there remain k 2 + 1 k_2+1 k 2 + 1 unknowns: the amplitude c β c_\beta c β and the coefficients of the k 2 k_2 k 2 decaying modes, and they are not determined by the reduced ODE. They are fixed by at most k 2 k_2 k 2 effective conditions (28) (Proposition 1, Remark 3), i.e., by a return to the IDE, and by one condition at zero: Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 for c < 0 c<0 c < 0 , boundedness for c > 0 c>0 c > 0 (the exclusion of the mode (47) ) and for c = 0 c=0 c = 0 (the exclusion of the branch u ν − u^{\nu_-} u ν − , Remark 7). For k 1 = k 2 = 1 k_1=k_2=1 k 1 = k 2 = 1 (exponential jumps of both signs) a complete description of the solution space and the asymptotics with C > 0 C>0 C > 0 are given in [17 Sects. 7–8] . The exact solution through Heun functions [30 ] refers to the one-sided model (A) with exponentially distributed claims. For general k 1 , k 2 k_1,k_2 k 1 , k 2 , an explicit solution of the connection problem between u = 0 u=0 u = 0 and u = ∞ u=\infty u = ∞ — the computation of the coefficients c β c_\beta c β and C k C_k C k themselves — remains open. This affects neither the exponent β \beta β nor the positivity of C k C_k C k (Proposition 4).
Test cases
For k = 1 k=1 k = 1 : N ( f ) = f + θ f ′ \mathcal{N}(f)=f+\theta f' N ( f ) = f + θ f ′ and J 1 ( u ) = θ − 1 ∫ 0 u Φ ( u − y ) e − y / θ d y J_1(u)=\theta^{-1}\int_0^u\Phi(u-y)e^{-y/\theta}\,dy J 1 ( u ) = θ − 1 ∫ 0 u Φ ( u − y ) e − y / θ d y . The formula of Theorem 1 for k = 1 k=1 k = 1 gives an ODE of order 3 for Φ \Phi Φ or, equivalently, of order 2 for G = Φ ′ G=\Phi' G = Φ ′ :
1 2 σ 2 θ u 2 G ′ ′ + ( 1 2 σ 2 u 2 + θ ( σ 2 u + a u + c ) ) G ′ + ( θ a + a u + c − θ α ) G = 0 (49) \frac{1}{2}\sigma^2\theta u^2G''+\left(\frac{1}{2}\sigma^2u^2+\theta(\sigma^2u+au+c)\right)G'+(\theta a+au+c-\theta\alpha)G=0 \tag*{(49)} 2 1 σ 2 θ u 2 G ′′ + ( 2 1 σ 2 u 2 + θ ( σ 2 u + a u + c ) ) G ′ + ( θ a + a u + c − θ α ) G = 0 (49)
which coincides with equation (13) of [7 ] and admits an exact solution through Heun functions, as shown in [30 ] . We use this as a consistency check of the formulas of Theorem 1 in the simplest case.
For k = 2 k=2 k = 2 the operator is N ( f ) = f + 2 θ f ′ + θ 2 f ′ ′ \mathcal{N}(f)=f+2\theta f^{\prime}+\theta^{2}f^{\prime\prime} N ( f ) = f + 2 θ f ′ + θ 2 f ′′ . By Theorem 1:
g 4 ( u ) = 1 2 σ 2 θ 2 u 2 , g_{4}(u)=\frac{1}{2}\sigma^{2}\theta^{2}u^{2}, g 4 ( u ) = 2 1 σ 2 θ 2 u 2 ,
g 3 ( u ) = σ 2 θ u 2 + θ 2 ( 2 σ 2 u + a u + c ) , g_{3}(u)=\sigma^{2}\theta u^{2}+\theta^{2}(2\sigma^{2}u+au+c), g 3 ( u ) = σ 2 θ u 2 + θ 2 ( 2 σ 2 u + a u + c ) ,
g 2 ( u ) = 1 2 σ 2 u 2 + 2 θ ( σ 2 u + a u + c ) + θ 2 ( σ 2 + 2 a − α ) , g_{2}(u)=\frac{1}{2}\sigma^{2}u^{2}+2\theta(\sigma^{2}u+au+c)+\theta^{2}(\sigma^{2}+2a-\alpha), g 2 ( u ) = 2 1 σ 2 u 2 + 2 θ ( σ 2 u + a u + c ) + θ 2 ( σ 2 + 2 a − α ) ,
g 1 ( u ) = 2 θ a + a u + c − 2 θ α , g_{1}(u)=2\theta a+au+c-2\theta\alpha, g 1 ( u ) = 2 θ a + a u + c − 2 θ α ,
which agrees with the coefficients of (29) . By Theorem 4 and Proposition 4, Ψ ( u ) ∼ C 2 u − β \Psi(u)\sim C_{2}u^{-\beta} Ψ ( u ) ∼ C 2 u − β with C 2 > 0 C_{2}>0 C 2 > 0 , and we compute the coefficients of the complete expansion from the recurrence (34) : a 1 = 2 ( c − 2 α θ ) / σ 2 a_{1}=2(c-2\alpha\theta)/\sigma^{2} a 1 = 2 ( c − 2 α θ ) / σ 2 , a 2 = − 2 ( 3 a α θ 2 − 4 α θ 2 + 4 α c θ − c 2 ) / σ 4 a_{2}=-2(3a\alpha\theta^{2}-4\alpha\theta^{2}+4\alpha c\theta-c^{2})/\sigma^{4} a 2 = − 2 ( 3 a α θ 2 − 4 α θ 2 + 4 α c θ − c 2 ) / σ 4 (Remark 4).
For k 1 = k 2 = 1 k_{1}=k_{2}=1 k 1 = k 2 = 1 the operator form (26) gives an ODE of order 4 4 4 for Φ \Phi Φ (of order 3 3 3 for G G G ), which for θ 1 = θ 2 \theta_{1}=\theta_{2} θ 1 = θ 2 coincides with the expanded form (27) . This case agrees with equation (7.2) for G = Φ ′ G=\Phi^{\prime} G = Φ ′ of the mixed model in [17 ] for exponential jumps with arbitrary θ 1 , θ 2 \theta_{1},\theta_{2} θ 1 , θ 2 (up to an overall sign and a renumbering of the indices: there, index 1 corresponds to downward jumps), which provides a consistency check for Theorem 3.
Discussion
In this paper we carried out the reduction of the IDE to a linear ODE as a necessary consequence for all three variants of the risk model with investments and Erlang-distributed jumps. The main results:
(i) We constructed annihilating operators N \mathcal{N} N and N ~ \widetilde{\mathcal{N}} N of order k k k that explicitly eliminate the nonlocal integral in the IDE.
(ii) The coefficients of the resulting ODEs are given in closed form in terms of binomial coefficients for the one-sided models and for the mixed model with k 1 = k 2 k_{1}=k_{2} k 1 = k 2 (in the general case, through the composition of operators). For k 1 = k 2 k_{1}=k_{2} k 1 = k 2 the equation of the mixed model coincides with equation (6) of Antipov [1 ] for densities with rational Laplace transform. The remaining equations follow from the same construction.
(iii) We established the exponent of the power branch β = 2 a / σ 2 − 1 \beta=2a/\sigma^{2}-1 β = 2 a / σ 2 − 1 , independent of k k k , and proved the realization of the tail Ψ ( u ) ∼ C k u − β \Psi(u)\sim C_{k}u^{-\beta} Ψ ( u ) ∼ C k u − β , C k > 0 C_{k}>0 C k > 0 , for all k k k and all β > 0 \beta>0 β > 0 (Proposition 4). For the one-sided models this gives a different proof of the asymptotic part of the results [14 , 27 , 29 ] (in the annuity model, without any external tail bound, from the ODE and the condition Φ ( 0 + ) = 0 \Phi(0+)=0 Φ ( 0 + ) = 0 , and in non-life insurance with the power-law lower bound [17 Proposition 4.1] ). For the mixed model it gives an analytical proof, with the same lower bound, of a result known from the probabilistic approach [3 , 22 , 25 ] , extending [17 ] to Erlang-distributed jumps and [1 ] to
β ≥ 1 \beta\ge1 β ≥ 1 and c = 0 c = 0 c = 0 . In all three models we obtained a complete asymptotic expansion up to a multiplicative constant C k C_k C k (for non-life insurance it was known for k = 1 k = 1 k = 1 from [30 ] , and for k = 2 k = 2 k = 2 we give its coefficients by an explicit recurrence).
(iv) For the mixed model (C) we classified the fundamental set at u = ∞ u = \infty u = ∞ . The boundedness of the physical solution excludes the k 1 k_1 k 1 growing modes, and the remaining connection coefficients are fixed by at most k 2 k_2 k 2 effective conditions of Proposition 1 (Remark 3) and one condition at zero: Φ ( 0 + ) = 0 \Phi(0+) = 0 Φ ( 0 + ) = 0 for c < 0 c < 0 c < 0 , boundedness for c > 0 c > 0 c > 0 (the exclusion of the mode (47) ) and for c = 0 c = 0 c = 0 (the exclusion of the unbounded Frobenius branch).
(v) We obtained a converse criterion for bounded solutions: IDE ⇔ \Leftrightarrow ⇔ ODE together with the vanishing of the residual of the IDE and its derivatives up to order K − 1 K - 1 K − 1 at zero (here K K K is the order of the annihilator, and the number of effective conditions is at most k k k , zero, and at most k 2 k_2 k 2 in models (A), (B), (C) respectively, Remark 3), generalizing the relation c Φ ′ ( 0 + ) = α Φ ( 0 + ) c\Phi'(0+) = \alpha\Phi(0+) c Φ ′ ( 0 + ) = α Φ ( 0 + ) of [30 ] .
Natural directions for further work are: (a) explicit solutions (analogues of the Heun functions) and the computation of the constants C k C_k C k for k ≥ 2 k \ge2 k ≥ 2 , (b) the asymptotics as u → 0 + u \to0+ u → 0 + and the summation of divergent series at zero (a Gevrey class of order 1, cf. [30 ] ), (c) weakening the smoothness conditions (A2) in the spirit of [29 ] for distributions with a less regular density, and (d) for the mixed model (C), an explicit solution of the connection problem between u = 0 u = 0 u = 0 and u = ∞ u = \infty u = ∞ (Remark 8) and the structure of ramified modes at multiple roots of the symbol.
Declarations
Conflict of interest. The author declares no conflict of interest.
Availability of data and materials. No datasets were generated or analyzed during the current study.