betaFeedback

WorksMathematics

The Laplace transform of the time of ruin in the Cramér–Lundberg model with Erlang E2(μ)E_2(μ) claims: spectrum, exact moments and matrix reduction

Abstract

We study the Laplace transform of the time of ruin in the classical Cramér–Lundberg model with claims having an Erlang distribution of shape parameter 2. We derive an integro-differential equation with two boundary conditions. For every positive discount rate, we show that its characteristic polynomial has one negative root and two distinct positive roots. Boundedness excludes the growing exponential term and yields a representation of the transform as a sum of two decaying exponentials. We construct a finite-dimensional matrix reduction of the convolution chain. For a model parameter set, we obtain exact expressions for the ruin probability, the mean time of ruin, and higher moments, including the third moment at zero initial surplus. We prove analyticity in the discount rate and justify the resulting moment series. We consider the analytic continuation of the representation to negative discount rates separately. For Erlang claims with an arbitrary positive integer shape parameter, we establish a closed linear system and the asymptotics of the root continuing the zero root at zero discount.

manuscript No. (will be inserted by the editor)

Mathematics Subject Classification (2020) 91G05 · 60G51 · 45J05 · 34A30 · 60G40

Introduction

Let TT be the first time the surplus of an insurance company enters the negative half-line. The object of study is the Laplace transform of the time of ruin

φδ(u)=Eu[e−δT; T<∞],δ≥0,T=inf⁡{t>0:R(t)<0}.(1)\varphi_\delta(u)=\mathbb{E}_u[e^{-\delta T};\,T<\infty],\qquad\delta\ge0,\qquad T=\inf\{t>0:R(t)<0\}. \tag*{(1)}

At δ=0\delta=0 this transform coincides with the ruin probability Ψ0(u)=Pu(T<∞)\Psi_0(u)=\mathbb{P}_u(T<\infty), and its derivatives with respect to the discount rate are related to the moments of the time of ruin. Such transforms belong to the Gerber–Shiu framework of the time value of ruin [5, 6]. Sundt and Teugels [12] studied the model in which the surplus earns interest at a constant force of interest.

For claims with a rational Laplace transform (in particular, for Erlang-distributed claims), the representation of φδ\varphi_\delta in terms of the roots of the Lundberg equation was obtained by Gerber and Shiu [5]. The defective renewal equations arising in such problems were studied in [8], and recurrence relations for the moments of the time of ruin were obtained in [9]. Here the same case is treated by elementary means. The additions are exact rational expressions for the mean and the higher moments in the model case, analyticity in the discount rate with an explicit radius of the moment series, and a finite-dimensional matrix reduction of the convolution chain with an explicit spectrum.

For Erlang-distributed claims, the integro-differential equation (IDE) admits a reduction to an equation with constant coefficients. The reduction of an IDE to an ordinary differential equation (ODE) is also used in risk models with investment [1, 2, 3, 7]; unlike the case considered here, the equations arising there have variable coefficients. Promyslov [10, 11] also studies the construction and the asymptotics of the solutions of such IDEs. We focus on the case of claims with the Erlang distribution E2(μ)E_2(\mu) (shape 2, rate μ\mu), in which an equation of the third order arises. For it one must write out the characteristic polynomial and then establish which of its roots are admitted by the boundary conditions and by the boundedness of the solution.

We prove that for every δ>0\delta>0 the characteristic polynomial has one negative and two distinct positive roots, and as δ→0\delta\to0 the negative root continues the zero root of the limiting problem. On this basis we obtain the representation of the transform as a sum of two exponentials. For the model parameter set we find the exact ruin probability and the moments of the time of ruin. We additionally construct the matrix reduction of the convolution chain and the analytic expansion in δ\delta, and for claims with the distribution Ek(μ)E_k(\mu) we obtain a closed linear system.

§ 2 derives the equation and the boundary conditions, § 3 studies the roots, and § 4 provides the exact solution and the matrix reduction. § 5 treats the moments, § 6 treats the numerical illustrations, and § 7 and § 8 treat the general case of Ek(μ)E_k(\mu) and the properties of the transform in the parameter δ\delta.

The equation for the Laplace transform and the boundary relations

The surplus process is R(t)=u+ct−S(t)R(t)=u+ct-S(t), where S(t)S(t) is a compound Poisson process of intensity λ\lambda, the jump sizes XX have the density and the survival function

f(x)=μ2xe−μx,F‾(u)=e−μu(1+μu),mX,1=2μ,mX,2=6μ2,(2)f(x)=\mu^{2}x e^{-\mu x},\qquad\overline{F}(u)=e^{-\mu u}(1+\mu u), \qquad m_{X,1}=\frac{2}{\mu},\qquad m_{X,2}=\frac{6}{\mu^{2}}, \tag*{(2)}

and the net profit condition ρ=λmX,1/c<1\rho=\lambda m_{X,1}/c<1 holds. We extend φδ\varphi_{\delta} to u<0u<0 by setting it equal to 1, since ruin has already occurred there.

Proposition 1 (the equation and the boundary relations). For δ≥0\delta\geq0 the function φδ\varphi_{\delta} belongs to C2([0,∞))∩C3((0,∞))C^{2}([0,\infty))\cap C^{3}((0,\infty)), tends to zero as u→∞u\to\infty, and satisfies

cφδ′(u)−(λ+δ)φδ(u)+λ∫0uφδ(u−x)f(x) dx+λF‾(u)=0,u>0,(3)c\varphi_{\delta}'(u)-(\lambda+\delta)\varphi_{\delta}(u)+\lambda\int_{0}^{u}\varphi_{\delta}(u-x)f(x)\,\mathrm{d}x+\lambda\overline{F}(u)=0,\qquad u>0, \tag*{(3)}

whence at u=0u=0

cφδ′(0)−(λ+δ)φδ(0)+λ=0,cφδ′′(0)−(λ+δ)φδ′(0)=0.(4)c\varphi_{\delta}'(0)-(\lambda+\delta)\varphi_{\delta}(0)+\lambda=0,\qquad c\varphi_{\delta}''(0)-(\lambda+\delta)\varphi_{\delta}'(0)=0. \tag*{(4)}

Proof. Conditioning on the first claim time τ1\tau_{1}, which is exponentially distributed with parameter λ\lambda, gives for a bounded measurable φδ\varphi_{\delta} the integral equation

φδ(u)=∫0∞λe−(λ+δ)t[∫0u+ctφδ(u+ct−x)f(x) dx+F‾(u+ct)]dt=λce(λ+δ)u/c∫u∞e−(λ+δ)s/c[(φδ∗f)(s)+F‾(s)]ds,\begin{aligned} \varphi_{\delta}(u)=\int_{0}^{\infty}\lambda e^{-(\lambda+\delta)t}\left[\int_{0}^{u+ct}\varphi_{\delta}(u+ct-x)f(x)\,\mathrm{d}x+\overline{F}(u+ct)\right]\mathrm{d}t \\ &=\frac{\lambda}{c}e^{(\lambda+\delta)u/c}\int_{u}^{\infty}e^{-(\lambda+\delta)s/c}\left[(\varphi_{\delta}*f)(s)+\overline{F}(s)\right]\mathrm{d}s, \end{aligned}

where u≥0u\geq0, and the contribution of the region x>u+ctx>u+ct equals F‾(u+ct)\overline{F}(u+ct) thanks to the extension by 1 for u<0u<0. The right-hand side is continuous in uu on [0,∞)[0,\infty), hence so is φδ\varphi_{\delta}. The convolution (φδ∗f)(u)=∫0uφδ(v)f(u−v) dv(\varphi_{\delta}*f)(u)=\int_{0}^{u}\varphi_{\delta}(v)f(u-v)\,\mathrm{d}v with f(0)=0f(0)=0 belongs to C1([0,∞))C^{1}([0,\infty)) and its derivative is ∫0uφδ(v)f′(u−v) dv\int_{0}^{u}\varphi_{\delta}(v)f'(u-v)\,\mathrm{d}v; therefore the right-hand side is continuously differentiable, and differentiating in uu we obtain (3) on [0,∞)[0,\infty) together with the inclusion φδ∈C1([0,∞))\varphi_{\delta}\in C^{1}([0,\infty)). Differentiating (3) repeatedly, we obtain in turn φδ∈C2([0,∞))\varphi_{\delta}\in C^{2}([0,\infty)) and φδ∈C3((0,∞))\varphi_{\delta}\in C^{3}((0,\infty)), the derivatives at zero being understood as one-sided. The first relation in (4) is the identity (3) at u=0u=0. The derivatives (φδ∗f)′(u)(\varphi_{\delta}*f)'(u) and F‾′(u)\overline{F}'(u) vanish at u=0u=0 since f(0)=0f(0)=0, whence the second relation. Finally, convergence to zero as u→∞u\to\infty follows from ρ<1\rho<1: almost surely R(t)/t→c−λmX,1>0R(t)/t\to c-\lambda m_{X,1}>0, and therefore Pu(T<∞)→0\mathbb{P}_{u}(T<\infty)\to0. □\square

Remark 1. The second relation in (4) uses the condition k≥2k\geq2 in an essential way. For exponential claims (k=1k=1) it takes the form cφδ′′(0)−(λ+δ)φδ′(0)=λμ[1−φδ(0)]c\varphi_{\delta}''(0)-(\lambda+\delta)\varphi_{\delta}'(0)=\lambda\mu[1-\varphi_{\delta}(0)], because (φ∗f)′(0)=φ(0)f(0)=μφ(0)(\varphi*f)'(0)=\varphi(0)f(0)=\mu\varphi(0) and F‾′(0)=−μ\overline{F}'(0)=-\mu, and the linear system for the constants of integration changes. This case is checked in Section 6.

The characteristic equation and the signs of its roots

Lemma 1 (raising the order). The operator (D+μ)2(D+\mu)^2, D=d/duD=\mathrm{d}/\mathrm{d}u, turns (3) into the linear ordinary differential equation

cφ′′′+(2cμ−(λ+δ))φ′′+(cμ2−2μ(λ+δ))φ′−δμ2φ=0,(5)c\varphi''' + (2c\mu-(\lambda+\delta))\varphi'' + (c\mu^2-2\mu(\lambda+\delta))\varphi' - \delta\mu^2\varphi= 0, \tag*{(5)}

because (D+μ)2Fˉ=0(D+\mu)^2\bar{F}=0 and (D+μ)2(φ∗f)=μ2φ(D+\mu)^2(\varphi*f)=\mu^2\varphi. Conversely, if φ∈C3\varphi\in C^3 satisfies (5), then the residual G:=cφ′−(λ+δ)φ+λ(φ∗f)+λFˉG:=c\varphi'-(\lambda+\delta)\varphi+\lambda(\varphi*f)+\lambda\bar{F} lies in ker⁡(D+μ)2=span⁡{e−μu,ue−μu}\ker(D+\mu)^2=\operatorname{span}\{e^{-\mu u},ue^{-\mu u}\}, and the values G(0)G(0) and G′(0)G'(0) coincide with the left-hand sides of (4), since (φ∗f)(0)=(φ∗f)′(0)=0(\varphi*f)(0)=(\varphi*f)'(0)=0, Fˉ(0)=1\bar{F}(0)=1 and Fˉ′(0)=0\bar{F}'(0)=0. Therefore a solution of (5) satisfies (3) if and only if both relations (4) hold.

Proof. From (2): (D+μ)Fˉ=μe−μu(D+\mu)\bar{F}=\mu e^{-\mu u} and (D+μ)2Fˉ=0(D+\mu)^2\bar{F}=0. For g(u)=∫0uφ(u−x)xe−μx dxg(u)=\int_0^u\varphi(u-x)xe^{-\mu x}\,\mathrm{d}x direct differentiation gives (D+μ)g=∫0uφ(u−x)e−μx dx(D+\mu)g=\int_0^u\varphi(u-x)e^{-\mu x}\,\mathrm{d}x, (D+μ)2g=φ(D+\mu)^2g=\varphi, and the factor λμ2\lambda\mu^2 accounts for f=μ2xe−μxf=\mu^2xe^{-\mu x}. The kernel of (D+μ)2(D+\mu)^2 is two-dimensional, as stated. □\square

Substituting φ=e−xu\varphi=e^{-xu} into (5) gives the characteristic polynomial

qδ(x)=cx3+(λ+δ−2cμ)x2+(cμ2−2μ(λ+δ))x+δμ2.(6)q_\delta(x)=cx^3+(\lambda+\delta-2c\mu)x^2+(c\mu^2-2\mu(\lambda+\delta))x+\delta\mu^2. \tag*{(6)}

Theorem 1 (the roots of qδq_\delta). Let cμ>2λc\mu>2\lambda. (i) At δ=0\delta=0 the polynomial q0q_0 has the root x=0x=0 and two distinct positive roots x1,2x_{1,2}. The term corresponding to the zero root is excluded by the condition φ0(∞)=0\varphi_0(\infty)=0; therefore φ0=A1e−x1u+A2e−x2u\varphi_0=A_1e^{-x_1u}+A_2e^{-x_2u}. (ii) For each δ>0\delta>0 the polynomial qδq_\delta has exactly one negative root x3(δ)x_3(\delta) and two distinct positive roots x1(δ),x2(δ)x_1(\delta),x_2(\delta), and

x3(δ)=−δμcμ−2λ+O(δ2),δ→0.(7)x_3(\delta)=-\frac{\delta\mu}{c\mu-2\lambda}+O(\delta^2),\qquad\delta\to0. \tag*{(7)}

The corresponding exponential term e−x3ue^{-x_3u} grows, therefore its coefficient vanishes by the boundedness condition, and

φδ(u)=A1(δ)e−x1(δ)u+A2(δ)e−x2(δ)u.(8)\varphi_\delta(u)=A_1^{(\delta)}e^{-x_1(\delta)u}+A_2^{(\delta)}e^{-x_2(\delta)u}. \tag*{(8)}

Proof. (i) We have q0(0)=0q_0(0)=0, q0′(0)=μ(cμ−2λ)>0q_0'(0)=\mu(c\mu-2\lambda)>0, q0(μ)=−λμ2<0q_0(\mu)=-\lambda\mu^2<0 and q0(x)→+∞q_0(x)\to+\infty as x→+∞x\to+\infty. Hence one nonzero root lies in (0,μ)(0,\mu) and the other in (μ,∞)(\mu,\infty). Together with x=0x=0 these are the three roots of the cubic polynomial.

(ii) To prove the uniqueness of the negative root, we put y=−x≥0y=-x\ge0 and consider

H(y):=qδ(−y)=(λ+δ−cy)(μ+y)2−λμ2.H(y):=q_\delta(-y)=(\lambda+\delta-cy)(\mu+y)^2-\lambda\mu^2.

Then H(0)=δμ2>0H(0)=\delta\mu^2>0 and H(y)→−∞H(y)\to-\infty as y→+∞y\to+\infty. Differentiating gives

H′(y)=(μ+y)[2(λ+δ)−cμ−3cy];H'(y)=(\mu+y)[2(\lambda+\delta)-c\mu-3cy];

the second factor decreases linearly and has at most one zero, while μ+y>0\mu+y>0. Hence H′H' changes sign at most once, and after that the sign remains negative. ends_mid=1 After its maximum HH decreases strictly, so the equation H(y)=0H(y)=0 has exactly one positive root y3y_3, and x3=−y3x_3=-y_3.

Next, qδ(0)=δμ2>0q_\delta(0)=\delta\mu^2>0, qδ(μ)=−λμ2<0q_\delta(\mu)=-\lambda\mu^2<0 and qδ(x)→+∞q_\delta(x)\to+\infty as x→+∞x\to+\infty. Therefore there exist roots x1∈(0,μ)x_1\in(0,\mu) and x2∈(μ,∞)x_2\in(\mu,\infty). Together with x3x_3 these are the three roots, and all of them are simple. The growing exponential term e−x3ue^{-x_3u} is excluded by the boundedness of φδ\varphi_\delta, and the two remaining ones give (8).

Formula (7) follows from the implicit function theorem:

x3′(0)=−∂qδ/∂δ∂q0/∂x∣(x,δ)=(0,0).x_3'(0)=-\left.\frac{\partial q_\delta/\partial\delta}{\partial q_0/\partial x}\right|_{(x,\delta)=(0,0)}.

where ∂qδ/∂δ=(μ−x)2\partial q_\delta/\partial\delta=(\mu-x)^2 and ∂q0/∂x=μ(cμ−2λ)\partial q_0/\partial x=\mu(c\mu-2\lambda) at x=0x=0. We note that the root count in (ii) does not use the condition cμ>2λc\mu>2\lambda; the latter is needed only for (i) and (7). □\square

Corollary 1. The two-exponential representation of the solution (8) holds for all δ≥0\delta\geq0, and the exponents are the two positive roots of (6). The third exponential term grows for δ>0\delta>0 and does not enter the bounded solution. For example, for λ=μ=3\lambda=\mu=3, c=6c=6 and δ=0.05\delta=0.05 the roots are −0.012461-0.012461, 1.5049601.504960, 3.9991683.999168 to the digits shown.

Remark 2. The growing exponential term does not obstruct the computation of φδ(0)\varphi_\delta(0) via the truncated boundary value problem with the condition φδ(U)=0\varphi_\delta(U)=0 (the truncation error has order e−(x1+∣x3∣)Ue^{-(x_1+|x_3|)U}), but it makes the direct computation of the profile of the solution unstable for large uu. Therefore the closed form (8) with the two relations (4) is preferable for computations.

The exact solution of the model case

From now on λ=3\lambda=3, μ=3\mu=3, k=2k=2, c=6c=6. Then mX,1=2/3m_{X,1}=2/3, ρ=1/3\rho=1/3, cμ=18>6=2λc\mu=18>6=2\lambda. The polynomial q0(x)=6x3−33x2+36x=3x(2x−3)(x−4)q_0(x)=6x^3-33x^2+36x=3x(2x-3)(x-4) has the roots 00, 3/23/2, 44.

Theorem 2 For this parameter set

Ψ0(u)=φ0(u)=25e−3u/2−115e−4u,u≥0,(9)\Psi_0(u)=\varphi_0(u)=\frac{2}{5}e^{-3u/2}-\frac{1}{15}e^{-4u},\qquad u\geq0, \tag*{(9)}

and Ψ0(0)=1/3=ρ=λmX,1/c\Psi_0(0)=1/3=\rho=\lambda m_{X,1}/c follows from the boundary conditions and is not imposed from outside.

Proof. The constants A1=2/5A_1=2/5, A2=−1/15A_2=-1/15 solve the linear system (4) at x1,2=3/2,4x_{1,2}=3/2,4. The first relation gives −6(A1x1+A2x2)−3(A1+A2)+3=0-6(A_1x_1+A_2x_2)-3(A_1+A_2)+3=0, that is, 12A1+27A2=312A_1+27A_2=3. The second gives 6(A1x12+A2x22)+3(A1x1+A2x2)=06(A_1x_1^2+A_2x_2^2)+3(A_1x_1+A_2x_2)=0, that is, 18A1+108A2=018A_1+108A_2=0; hence A1=−6A2A_1=-6A_2, and then A2=−1/15A_2=-1/15, A1=2/5A_1 = 2/5. To check the original equation (3) we use the exact convolution of an exponential term:

∫0ue−x(u−y)μ2ye−μy dy=μ2(μ−x)2e−xu−μ2e−μu(uμ−x+1(μ−x)2).(10)\int_0^u e^{-x(u-y)}\mu^2 y e^{-\mu y}\,dy = \frac{\mu^2}{(\mu-x)^2}e^{-xu}-\mu^2e^{-\mu u}\left(\frac{u}{\mu-x}+\frac{1}{(\mu-x)^2}\right). \tag*{(10)}

The coefficients of e−xjue^{-x_j u} vanish after the substitution because of q0(xj)=0q_0(x_j)=0. The remaining terms have the form λe−μu(αu+β)\lambda e^{-\mu u}(\alpha u+\beta), where

α=μ(1−∑j=12Ajμμ−xj),β=1−∑j=12Ajμ2(μ−xj)2.\alpha=\mu\left(1-\sum_{j=1}^{2}A_j\frac{\mu}{\mu-x_j}\right),\qquad \beta=1-\sum_{j=1}^{2}A_j\frac{\mu^2}{(\mu-x_j)^2}.

For the constants found above, ∑jAjμ/(μ−xj)=2/5⋅2+(−1/15)(−3)=1\sum_j A_j\mu/(\mu-x_j)=2/5\cdot2+(-1/15)(-3)=1 and ∑jAjμ2/(μ−xj)2=2/5⋅4+(−1/15)⋅9=1\sum_j A_j\mu^2/(\mu-x_j)^2=2/5\cdot4+(-1/15)\cdot9=1. Hence the residual of the original IDE vanishes identically, including all tail terms of the truncated convolution. Finally, Ψ0(0)=A1+A2=1/3\Psi_0(0)=A_1+A_2=1/3. ⊓⊔

We now turn to an independent matrix reduction of the convolution chain.

Lemma 2 (the convolution chain). For a locally integrable φ\varphi put

B1=∫0uφ(u−x)e−μx dx,B2=∫0uφ(u−x)xe−μx dx,B_1=\int_0^u\varphi(u-x)e^{-\mu x}\,dx,\qquad B_2=\int_0^u\varphi(u-x)xe^{-\mu x}\,dx,
E=e−μu,U1=ue−μu.E=e^{-\mu u},\qquad U_1=ue^{-\mu u}.

Then B1′=φ−μB1B_1'=\varphi-\mu B_1, B2′=B1−μB2B_2'=B_1-\mu B_2, E′=−μEE'=-\mu E, U1′=E−μU1U_1'=E-\mu U_1, (φ∗f)=μ2B2(\varphi*f)=\mu^2B_2, Fˉ=E+μU1\bar F=E+\mu U_1. At u=0u=0 the quantities B1,B2,U1B_1,B_2,U_1 vanish and E(0)=1E(0)=1.

Proof. Differentiation with respect to the upper limit gives B1′=φ−μB1B_1'=\varphi-\mu B_1. For B2B_2 it is more convenient to change the variable v=u−xv=u-x: B2=∫0uφ(v)(u−v)e−μ(u−v) dvB_2=\int_0^u\varphi(v)(u-v)e^{-\mu(u-v)}\,dv, whence

B2′=∫0uφ(v)[e−μ(u−v)−μ(u−v)e−μ(u−v)] dv=B1−μB2.B_2'=\int_0^u\varphi(v)\left[e^{-\mu(u-v)}-\mu(u-v)e^{-\mu(u-v)}\right]\,dv=B_1-\mu B_2.

(the boundary term φ(u)⋅0\varphi(u)\cdot0 is zero). The relations for EE and U1U_1 and the representation Fˉ=E+μU1\bar F=E+\mu U_1 are immediate, and (φ∗f)=μ2B2(\varphi*f)=\mu^2B_2 follows from f=μ2xe−μxf=\mu^2xe^{-\mu x}. ⊓⊔

Put z=(φ,B1,B2,E,U1)T\boldsymbol{z}=(\varphi,B_1,B_2,E,U_1)^{\mathsf T}.

Proposition 2 (the closed linear system). The vector z\boldsymbol{z} satisfies the system z′=Mz\boldsymbol{z}'=M\boldsymbol{z} with constant coefficients

M=(λ+δc0−λμ2c−λc−λμc1−μ00001−μ00000−μ00001−μ).(11)M= \begin{pmatrix} \frac{\lambda+\delta}{c} & 0 & -\frac{\lambda\mu^2}{c} & -\frac{\lambda}{c} & -\frac{\lambda\mu}{c}\\ 1 & -\mu& 0 & 0 & 0\\ 0 & 1 & -\mu& 0 & 0\\ 0 & 0 & 0 & -\mu& 0\\ 0 & 0 & 0 & 1 & -\mu \end{pmatrix}. \tag*{(11)}

and z(u)=eMuz(0)z(u)=e^{Mu}z(0), where z(0)=(φδ(0),0,0,1,0)⊤z(0)=(\varphi_\delta(0),0,0,1,0)^\top. The value φδ(0)\varphi_\delta(0) is determined by the boundedness of the solution and by the condition φδ(u)→0\varphi_\delta(u)\to0 as u→∞u\to\infty. For δ>0\delta>0 the latter condition follows already from boundedness, because the third exponential term grows. In both cases one uses the condition ℓ⊤z(0)=0\ell^\top z(0)=0 for a left eigenvector ℓ\ell of the matrix MM corresponding to the eigenvalue −x3-x_3. In a numerical computation on [0,U][0,U] it is replaced by the truncated condition φ(U)=0\varphi(U)=0, with an error of order e−(x1+∣x3∣)Ue^{-(x_1+|x_3|)U}.

Proof. Substituting (φ∗f)=μ2B2(\varphi*f)=\mu^2B_2 and Fˉ=E+μU1\bar{F}=E+\mu U_1 into (3) by Lemma 2 gives the first row of (11); the remaining rows are definitions. The matrix is constant, hence z(u)=eMuz(0)z(u)=e^{Mu}z(0), and the truncated boundary conditions are linear in z(0)z(0); hence the system has size 5. The condition ℓ⊤z(0)=0\ell^\top z(0)=0 is equivalent to the vanishing of the coefficient of the exponential term e−x3ue^{-x_3u}. The eigenvalues of MM are −μ-\mu (of multiplicity two) and −xj-x_j, where xjx_j are the roots of (6). For δ>0\delta>0 one of them is positive, which produces the instability discussed after Corollary 1. □\square

The mean time of ruin and resonance

Denote

L~[y]:=cy′−λy+λ(y∗f),m(u):=Eu[T;T<∞].\widetilde{L}[y]:=cy'-\lambda y+\lambda(y*f),\qquad m(u):=\mathbb{E}_u[T;T<\infty].

If the mean is finite, differentiating (3) with respect to δ\delta at δ=0\delta=0 gives

L~[m]=−Ψ0,m(∞)=0,(12)\widetilde{L}[m]=-\Psi_0,\qquad m(\infty)=0, \tag*{(12)}

and from the relations (4) it follows that

−cm′(0)+λm(0)=Ψ0(0),−cm′′(0)+λm′(0)=Ψ0′(0).(13)-cm'(0)+\lambda m(0)=\Psi_0(0),\qquad-cm''(0)+\lambda m'(0)=\Psi'_0(0). \tag*{(13)}

Theorem 3 (the identity for the mean time of ruin). Let ρ<1\rho<1 and let mm be integrable on (0,∞)(0,\infty). Then

cE0[T;T<∞]=∫0∞Ψ0(u) du,c\mathbb{E}_0[T;T<\infty]=\int_0^\infty\Psi_0(u)\,\mathrm{d}u,
E0[T∣T<∞]=1λmX,1∫0∞Ψ0(u) du.\mathbb{E}_0[T\mid T<\infty]=\frac{1}{\lambda m_{X,1}}\int_0^\infty\Psi_0(u)\,\mathrm{d}u.

Proof. Integrating (3) at δ=0\delta=0 (the function Ψ0\Psi_0 is integrable by Theorem 1(i)) and using Fubini’s theorem together with ∫0∞Fˉ(u) du=mX,1\int_0^\infty\bar{F}(u)\,\mathrm{d}u=m_{X,1} gives −cΨ0(0)+λmX,1=0-c\Psi_0(0)+\lambda m_{X,1}=0, that is, Ψ0(0)=ρ\Psi_0(0)=\rho. Next we integrate (12). We have c∫0∞m′(u) du=−cm(0)c\int_0^\infty m'(u)\,\mathrm{d}u=-cm(0) and

∫0∞(m∗f)(u) du=∫0∞f(y)[∫y∞m(u−y) du]dy=∫0∞m(v) dv.\int_0^\infty(m*f)(u)\,\mathrm{d}u=\int_0^\infty f(y)\left[\int_y^\infty m(u-y)\,\mathrm{d}u\right]\mathrm{d}y=\int_0^\infty m(v)\,\mathrm{d}v.

Therefore the terms with λ\lambda cancel and we are left with −cm(0)=−∫0∞Ψ0(u) du-cm(0)=-\int_0^\infty\Psi_0(u)\,\mathrm{d}u. The conditional mean is obtained by dividing by Ψ0(0)=ρ=λmX,1/c\Psi_0(0)=\rho=\lambda m_{X,1}/c. In the model example the integrability of mm follows from the exact formula of the next theorem. □\square

Theorem 4 (the resonant structure of mm in the model case) For λ=3\lambda=3, μ=3\mu=3, c=6c=6

m(u)=Eu[T;T<∞]=360u+3279000e−3u/2+10u+489000e−4u,(14)m(u)=\mathbb{E}_u[T;T<\infty]=\frac{360u+327}{9000}e^{-3u/2}+\frac{10u+48}{9000}e^{-4u}, \tag*{(14)}

and m(0)=1/24m(0)=1/24 and E0[T∣T<∞]=1/8\mathbb{E}_0[T\mid T<\infty]=1/8.

Proof The finiteness of the mean follows from the analyticity established in Proposition 3 by the argument given in the proof of Theorem 5; therefore (12) applies. We apply to it the operator (D+μ)2(D+\mu)^2. Unlike the substitution of a single exponential directly into L~\widetilde{L}, this step accounts for the truncated convolution completely. We obtain

cm′′′+(2cμ−λ)m′′+(cμ2−2λμ)m′=−∑j=12Aj(μ−xj)2e−xju=−910e−3u/2+115e−4u.\begin{aligned} cm''' +(2c\mu-\lambda)m'' +(c\mu^2-2\lambda\mu)m'=-\sum_{j=1}^{2} A_j(\mu-x_j)^2e^{-x_ju} \\ &=-\frac{9}{10}e^{-3u/2}+\frac{1}{15}e^{-4u}. \end{aligned}

Both exponents on the right are positive roots of q0q_0; therefore the particular solution carries a factor uu. The condition m(∞)=0m(\infty)=0 excludes the constant term, and the bounded solution has the form

m(u)=(C1u+b1)e−3u/2+(C2u+b2)e−4u.m(u)=(C_1u+b_1)e^{-3u/2}+(C_2u+b_2)e^{-4u}.

The substitution into the resulting ODE gives C1=1/25C_1=1/25, C2=1/900C_2=1/900, after which the boundary conditions (13) give b1=109/3000b_1=109/3000, b2=2/375b_2=2/375. This yields (14). Substituting back into the original equation (12) with the exact convolution (10) and its xx-derivative (for the terms ue−xjuue^{-x_ju}) shows that both the coefficients of e−xjue^{-x_ju} and the terms of the form e−μu(αu+β)e^{-\mu u}(\alpha u+\beta) vanish. Hence no extraneous solutions were introduced by raising the order. Finally, m(0)=327/9000+48/9000=1/24m(0)=327/9000+48/9000=1/24 and

∫0∞Ψ0(u) du=25⋅23−115⋅14=14.\int_0^\infty\Psi_0(u)\,du=\frac{2}{5}\cdot\frac{2}{3}-\frac{1}{15}\cdot\frac{1}{4}=\frac{1}{4}.

so that cm(0)=1/4cm(0)=1/4 agrees with Theorem 3. Division by Ψ0(0)=1/3\Psi_0(0)=1/3 gives the conditional mean 1/81/8. ⊓⊔

Theorem 5 (the hierarchy of moments) For claims with the Erlang distribution of shape parameter 22, all the moments

mn(u)=Eu[Tn;T<∞]m_n(u)=\mathbb{E}_u[T^n;T<\infty]

are finite. Put m0=Ψ0m_0=\Psi_0. Then for n≥1n\ge1 we have

L~[mn]=−nmn−1,cmn(0)=n∫0∞mn−1(u) du.\widetilde{L}[m_n]=-nm_{n-1},\qquad cm_n(0)=n\int_0^\infty m_{n-1}(u)\,du.

Proof. The two-exponential representation (8) extends analytically in δ\delta to a neighborhood of zero; this is justified in Proposition 3, whose proof does not rely on the results of this section. The finiteness of the derivatives of the continuation at zero proves in turn the finiteness of all moments: if, the previous moments being already finite, the next moment were infinite, the corresponding difference quotient as δ↓0\delta\downarrow0 would also diverge by monotone convergence, which contradicts the differentiability of the continuation. Therefore differentiation in δ\delta under the expectation sign is allowed.

Differentiating the representation (8) also shows that mnm_n is a finite sum of terms uje−xluu^j e^{-x_l u} with j≤nj \le n and positive xlx_l. Hence mnm_n is integrable and mn(∞)=0m_n(\infty)=0. Further,

∂δn[(λ+δ)φδ]=(λ+δ)∂δnφδ+n∂δn−1φδ\partial_\delta^n\left[(\lambda+\delta)\varphi_\delta\right]=(\lambda+\delta)\partial_\delta^n\varphi_\delta+n\partial_\delta^{n-1}\varphi_\delta

while the density ff and the tail F‾\overline{F} do not depend on δ\delta. Since ∂δnφδ∣δ=0=(−1)nmn\left.\partial_\delta^n\varphi_\delta\right|_{\delta=0}=(-1)^n m_n, differentiating (3) gives the first statement. The integration over u∈(0,∞)u\in(0,\infty) with Fubini’s theorem, as in the proof of Theorem 3, gives the second. ◻

Corollary 2 (the second moment and the variance in the model case) From (14) we have ∫0∞m1(u) du=25/576\int_0^\infty m_1(u)\,\mathrm{d}u=25/576; hence

E0[T2;T<∞]=2625576=251728,E0[T2∣T<∞]=25576,\mathbb{E}_0[T^2; T<\infty]=\frac{2}{6}\frac{25}{576}=\frac{25}{1728},\qquad\mathbb{E}_0[T^2\mid T<\infty]=\frac{25}{576},
Var⁡(T∣T<∞)=25576−(18)2=136,Var⁡=16.\operatorname{Var}(T\mid T<\infty)=\frac{25}{576}-\left(\frac{1}{8}\right)^2=\frac{1}{36},\qquad\sqrt{\operatorname{Var}}=\frac{1}{6}.

The subsequent moments are computed in the same way: first one solves the equation L[mn]=−nmn−1\mathcal{L}[m_n]=-n m_{n-1} (a resonant ODE with terms uje−xluu^j e^{-x_l u}), and then integrates mnm_n.

Remark 3. The natural guess m(u)=(1/c)∫u∞Ψ0(v) dvm(u)=(1/c)\int_u^\infty\Psi_0(v)\,\mathrm{d}v fails for u>0u>0. Denoting the right-hand side by m~\widetilde{m}, from L~[m~]=−Ψ0+λ[(m~∗f)−m~]\widetilde{\mathcal{L}}[\widetilde{m}]=-\Psi_0+\lambda[(\widetilde{m}*f)-\widetilde{m}] we see that the identity would require the invariance of m~\widetilde{m} with respect to the convolution with ff. Table 2 shows that the relative discrepancy grows with uu. The integral identity of Theorem 3 is not a pointwise identity in uu.

uum(u)m(u) by (14)(1/c)∫u∞Ψ0(1/c)\int_u^\infty\Psi_0ratio
0.00.041666670.041666671.0000
1.00.017150300.009866020.5753
2.00.005794430.002211830.3817
5.00.000130710.000024580.1881
10.00.000000130.000000010.1019

Table 2. The mean time of ruin by the exact formula (14) and the naive continuation (1/c)∫u∞Ψ0(1/c)\int_u^\infty\Psi_0.

Numerical illustrations

The values below are computed from the characteristic equation and the two-exponential representation (8). They illustrate the proved formulas rather than replace their derivation.

Table 1 agrees with Theorem 1: the root x3x_3 is negative for δ>0\delta>0, and its linear estimate is most accurate for small δ\delta. The matrix boundary value problem from Proposition 2 reproduces the values φδ(0)\varphi_\delta(0) within the printed digits. In a direct computation of the profile through eMue^{M u}, however, the growing component may accumulate an error near the right edge of the truncated domain; therefore the accuracy of the value at the initial point should not be carried over to the whole profile without a separate estimate.

δ\deltax3(δ)x_{3}(\delta)−cμcμ−2λ-\frac{c\mu}{c\mu-2\lambda}x1(δ)x_{1}(\delta)x2(δ)x_{2}(\delta)A1(δ)A_{1}^{(\delta)}φδ(0)\varphi_{\delta}(0)
0.000.0000000.0000001.5000004.0000000.400000000.3333333333
0.05−0.012461−0.0125001.5049603.9991680.398198790.3312678779
0.10−0.024847−0.0250001.5098443.9983370.396427790.3292373739
0.20−0.049403−0.0500001.5193883.9966820.392972430.3252766458
0.50−0.121499−0.1250001.5464063.9917590.383235970.3141241706
1.00−0.237275−0.2500001.5869093.9837000.368768930.2975812980
2.00−0.456733−0.5000001.6552983.9681010.344709700.2701787511
5.00−1.060387−1.2500001.8019743.9250790.294826700.2141231862

Table 1. The roots of the characteristic equation (6) and the constants (8) for λ=3\lambda= 3, μ=3\mu= 3, c=6c = 6. The third column is the linear estimate (7).

For the test set λ=μ=3\lambda= \mu= 3, c=4c = 4 we have ρ=1/2\rho= 1/2, the positive roots 1.078835 and 4.171165, A1=0.55316953A_1 = 0.55316953, A2=−0.05316953A_2 = -0.05316953. Hence φ0(0)=0.5000000000=ρ\varphi_0(0) = 0.5000000000 = \rho, ∫0∞Ψ0(u) du=0.5000000000\int_0^\infty\Psi_0(u)\,\mathrm{d}u = 0.5000000000, E0[T;T<∞]=0.12500000\mathbb{E}_0[T; T < \infty] = 0.12500000 and E0[T∣T<∞]=0.25000000\mathbb{E}_0[T \mid T < \infty] = 0.25000000.

For exponential claims (k=1k = 1, μ=λ=1\mu= \lambda= 1, c=2c = 2) the characteristic equation (cx+λ)(μ−x)=λμ(cx + \lambda)(\mu- x) = \lambda\mu has the roots 0 and 1/21/2, and the solution Ψ0(u)=12e−u/2\Psi_0(u) = \frac{1}{2}e^{-u/2} makes the residual of the equation vanish identically. Here E0[T;T<∞]=1/2\mathbb{E}_0[T; T < \infty] = 1/2 and the conditional mean equals 1. The second boundary relation from Remark 1 becomes the equality 1/2=1/21/2 = 1/2, whereas the form (4) for k≥2k \ge2 would give here the wrong equality 1/2=01/2 = 0.

Finally, the slope of φδ(0)\varphi_\delta(0) at zero equals −E0[T;T<∞]=−1/24-\mathbb{E}_0[T; T < \infty] = -1/24 in the main model case. This gives an independent consistency check of Table 1 against Theorem 4 without finite-difference estimates of the derivative.

The case of claims with the Ek(μ)E_k(\mu) distribution

The constructions of Proposition 2 and § 2 extend to the claims fk(x)=μkxk−1e−μx(k−1)!f_k(x) = \frac{\mu^k x^{k-1}e^{-\mu x}}{(k - 1)!}. Put qδ,k(x)=(cx+λ+δ)(μ−x)k−λμkq_{\delta,k}(x) = (cx + \lambda+ \delta)(\mu- x)^k - \lambda\mu^k, a polynomial of degree $k + 1.ats

Theorem 6 (the finite-dimensional system and the root shift for general kk). For claims with the Ek(μ)E_k(\mu) distribution, the system

cφδ′=(λ+δ)φδ−λμkBk(k−1)!−λ∑j=0k−1μjUjj!,B1′=φδ−μB1,Bj′=(j−1)Bj−1−μBj(j=2,…,k),U0′=−μU0,Uj′=jUj−1−μUj(j=1,…,k−1),(15)\begin{aligned} c\varphi_\delta'&=(\lambda+\delta)\varphi_\delta-\lambda\frac{\mu^k B_k}{(k-1)!}-\lambda\sum_{j=0}^{k-1}\frac{\mu^jU_j}{j!},\\ B_1'&=\varphi_\delta-\mu B_1,\qquad B_j'=(j-1)B_{j-1}-\mu B_j\quad(j=2,\ldots,k),\\ U_0'&=-\mu U_0,\qquad U_j'=jU_{j-1}-\mu U_j\quad(j=1,\ldots,k-1), \tag*{(15)} \end{aligned}

with the initial conditions Bj(0)=0B_j(0)=0, U0(0)=1U_0(0)=1, Uj(0)=0U_j(0)=0 (j≥1j\ge1), is closed and has dimension 2k+12k+1; one has (φδ∗fk)=μkBk/(k−1)!(\varphi_\delta*f_k)=\mu^kB_k/(k-1)! and F‾k=∑j<kμjUj/j!\overline{F}_k=\sum_{j<k}\mu^jU_j/j!. The characteristic polynomial of its matrix AkA_k is

det⁡(ζI−Ak)=−(ζ+μ)kcqδ,k(−ζ),(16)\det(\zeta I-A_k)=-\frac{(\zeta+\mu)^k}{c}q_{\delta,k}(-\zeta), \tag*{(16)}

that is, the spectrum consists of −μ-\mu of multiplicity kk and the numbers −x-x, where xx are the roots of qδ,kq_{\delta,k}. At δ=0\delta=0 the number x=0x=0 is a root of q0,kq_{0,k}, and for δ>0\delta>0 this root moves into the negative half-line,

x−(δ)=−δc(1−ρ)+O(δ2),ρ=λkμc<1.(17)x_-(\delta)=-\frac{\delta}{c(1-\rho)}+O(\delta^2),\qquad\rho=\frac{\lambda k}{\mu c}<1. \tag*{(17)}

For a fixed value of ρ\rho the formula is universal, but through ρ=λk/(μc)\rho=\lambda k/(\mu c) the shift depends on kk.

Proof. For Bj(u)=∫0uφδ(u−y)yj−1e−μy dyB_j(u)=\int_0^u\varphi_\delta(u-y)y^{j-1}e^{-\mu y}\,\mathrm{d}y differentiation with respect to the upper limit gives B1′=φδ−μB1B_1'=\varphi_\delta-\mu B_1 and Bj′=(j−1)Bj−1−μBjB_j'=(j-1)B_{j-1}-\mu B_j for j≥2j\ge2, and Uj(u)=uje−μuU_j(u)=u^j e^{-\mu u} gives the second recurrence. The identities (φδ∗fk)=μkBk/(k−1)!(\varphi_\delta*f_k)=\mu^kB_k/(k-1)! and F‾k=∑j<kμjUj/j!\overline{F}_k=\sum_{j<k}\mu^jU_j/j! follow from the definitions of fkf_k and F‾k(u)=e−μu∑j<k(μu)j/j!\overline{F}_k(u)=e^{-\mu u}\sum_{j<k}(\mu u)^j/j!. The matrix AkA_k of the system has a block structure: the BB- and UU-blocks are lower bidiagonal matrices with diagonal −μ-\mu, the first row couples them with φδ\varphi_\delta, and the expansion of the determinant along the first row gives (16). For (17) we use q0,k(0)=0q_{0,k}(0)=0, ∂xq0,k(0)=μk−1(cμ−λk)=cμk(1−ρ)\partial_xq_{0,k}(0)=\mu^{k-1}(c\mu-\lambda k)=c\mu^k(1-\rho) and ∂δq0,k(0)=μk\partial_\delta q_{0,k}(0)=\mu^k. The implicit function theorem gives x−′(0)=−1/[c(1−ρ)]x_-'(0)=-1/[c(1-\rho)], as claimed. □

Table 3 checks (15): for λ=μ=3\lambda=\mu=3, c=6c=6 the matrix method reproduces the classical value φ0,k(0)=ρ=λk/(μc)\varphi_{0,k}(0)=\rho=\lambda k/(\mu c) up to a term of order e−x1Ue^{-x_1U}, where x1x_1 is the smallest positive root of q0,kq_{0,k}, and confirms the estimate (17).

The deviation 4.4⋅10−64.4\cdot10^{-6} at k=5k=5 is explained by the truncation of the computation domain: the smallest positive root of q0,5q_{0,5} is 0.17360.1736, hence the condition φ0,k(U)=0\varphi_{0,k}(U)=0 is imposed with an error of order e−0.1736⋅60≈3⋅10−5e^{-0.1736\cdot60}\approx3\cdot10^{-5}. At k=5k=5 the right-hand side of (17) differs from the exact root by 20%20\%. The reason is not a slow convergence: at δ∗=−(cμ/k)[k+ρ−(k+1)ρ1/(k+1)]≈−0.0465\delta_* =-(c\mu/k)[k+\rho-(k+1)\rho^{1/(k+1)}]\approx-0.0465 the root x−(δ)x_-(\delta) coalesces with the smallest positive root of qδ,5q_{\delta,5}; therefore δ=0.05\delta=0.05 lies outside the disk of convergence of the series in δ\delta, and (17) has here only an asymptotic meaning (at k=2k = 2 the same formula gives δ∗=−γ1\delta_* = -\gamma_1 of Proposition 4).

kkφ0,k(0)\varphi_{0,k}(0)ρ\rhodiscrepancyx−(0.05)x_-(0.05)by (17)
10.1666666666670.1666666667<10−15< 10^{-15}−0.009993−0.010000
20.3333333333330.3333333333<10−15< 10^{-15}−0.012461−0.012500
30.5000000000000.5000000000<10−15< 10^{-15}−0.016487−0.016667
40.6666666666650.66666666671.2⋅10−121.2 \cdot10^{-12}−0.024051−0.025000
50.8333289764200.83333333334.4⋅10−64.4 \cdot10^{-6}−0.041614−0.050000

Table 3. The values φ0,k(0)\varphi_{0,k}(0) (the matrix method, U=60U = 60) and the root shift at δ=0.05\delta= 0.05 for λ=μ=3\lambda= \mu= 3, c=6c = 6.

Remark 4. For k≤2k \le2 all the roots of qδ,kq_{\delta,k} are real, and the transform φδ\varphi_\delta is a combination of exponentials without oscillations. Already at k=3k = 3 and k=4k = 4, in the example considered, at δ=0\delta= 0 the complex conjugate pairs 3.82976±1.18122i3.82976 \pm1.18122 i and 3.19681±1.76237i3.19681 \pm1.76237 i appear, respectively. Therefore the result of Theorem 1 on the two positive roots applies specifically to k=2k = 2. For k≥3k \ge3 Rouché’s theorem, applied to the equation λ+δ−cs=λμk/(μ+s)k\lambda+ \delta- cs = \lambda\mu^k/(\mu+ s)^k (s=−xs = -x) in the half-plane Re⁡s≥0\operatorname{Re} s \ge0, gives for δ>0\delta> 0 exactly one root with Re⁡x<0\operatorname{Re} x < 0 and exactly kk roots with Re⁡x>0\operatorname{Re} x > 0, among which complex conjugate pairs appear (damped oscillations). The matrix reduction (15) remains valid.

Properties of the Laplace transform and the expansion in powers of $\delta`

The properties of φδ(0)\varphi_\delta(0) as a function of δ\delta follow from the definition (1) itself and do not depend on the method of computation. Recall that a function gg is called completely monotone on (0,∞)(0,\infty) if it is infinitely differentiable and (−1)ng(n)(δ)≥0(-1)^n g^{(n)}(\delta) \ge0 for all n=0,1,2,…n = 0,1,2,\ldots and all δ>0\delta> 0.

Lemma 3 (completely monotone functions) The function δ↦φδ(0)\delta\mapsto\varphi_\delta(0) is infinitely differentiable and completely monotone on (0,∞)(0,\infty):

(−1)ndndδnφδ(0)=E0[Tne−δT; T<∞]≥0,n=0,1,2,….(-1)^n \frac{\mathrm{d}^n}{\mathrm{d}\delta^n}\varphi_\delta(0) = \mathbb{E}_0\left[T^n e^{-\delta T};\,T < \infty\right] \ge0,\qquad n=0,1,2,\ldots.

In particular, φδ(0)\varphi_\delta(0) decreases, is convex, and lies below its value at δ=0\delta= 0.

Proof By (1) we have φδ(0)=∫[0,∞)e−δtν(dt)\varphi_\delta(0) = \int_{[0,\infty)} e^{-\delta t}\nu(\mathrm{d}t), where ν\nu is the positive measure of the distribution of the time of ruin on the set {T<∞}\{T < \infty\}. That the Laplace transform of a positive measure is a completely monotone function is checked by differentiating under the integral sign (the converse is Bernstein’s theorem, [4 Ch. XIII, § 4]). For a fixed δ>0\delta> 0 pick 0<δ0<δ0 < \delta_0 < \delta. Then tne−δt≤Cn,δ,δ0e−δ0tt^n e^{-\delta t} \le C_{n,\delta,\delta_0}e^{-\delta_0t}, and the integral ∫e−δ0tν(dt)=φδ0(0)≤1\int e^{-\delta_0t}\nu(\mathrm{d}t) = \varphi_{\delta_0}(0) \le1 is finite. This justifies the differentiation under the integral sign. □

Corollary 3 In the model case λ=μ=3\lambda=\mu=3, c=6c=6, for all δ≥0\delta\geq0 the following estimates hold:

0≤φδ(0)≤φ0(0)=13,φδ(0)≥φ0(0)+δ∂δφδ(0)∣δ=0+=13−δ24.0\leq\varphi_\delta(0)\leq\varphi_0(0)=\frac{1}{3},\qquad\varphi_\delta(0)\geq\varphi_0(0)+\left.\delta\partial_\delta\varphi_\delta(0)\right|_{\delta=0+}=\frac{1}{3}-\frac{\delta}{24}.

The second estimate follows from convexity; for δ>8\delta>8 it is trivial.

We turn to the analyticity at zero. Put dj(δ)=cxj(δ)+λ+δd_j(\delta)=cx_j(\delta)+\lambda+\delta. The substitution of the two-exponential representation (8) into the boundary system (4) gives

A1(δ)=λx2(x2−x1)d1,A2(δ)=−λx1(x2−x1)d2.(18)A_1^{(\delta)}=\frac{\lambda x_2}{(x_2-x_1)d_1},\qquad A_2^{(\delta)}=-\frac{\lambda x_1}{(x_2-x_1)d_2}. \tag*{(18)}

and hence

φδ(0)=λ(c(x1+x2)+λ+δ)d1d2.(19)\varphi_\delta(0)=\frac{\lambda(c(x_1+x_2)+\lambda+\delta)}{d_1d_2}. \tag*{(19)}

Proposition 3 (analyticity of the two-exponential representation) The branches x1(δ),x2(δ)x_1(\delta),x_2(\delta), the coefficients A1,2(δ)A_{1,2}^{(\delta)} and the function φδ(u)\varphi_\delta(u), for each fixed uu, are analytic in δ\delta in a neighborhood of zero.

Proof The positive roots x1(0)x_1(0) and x2(0)x_2(0) are simple and distinct by Theorem 1; therefore the implicit function theorem in the complex domain gives analytic branches x1(δ)x_1(\delta) and x2(δ)x_2(\delta), which remain distinct in a small neighborhood of zero. Moreover, dj≠0d_j\ne0 on the corresponding root: otherwise, from qδ(xj)=(cxj+λ+δ)(μ−xj)2−λμ2q_\delta(x_j)=(cx_j+\lambda+\delta)(\mu-x_j)^2-\lambda\mu^2 it would follow that qδ(xj)=−λμ2≠0q_\delta(x_j)=-\lambda\mu^2\ne0. Therefore the denominators in (18) and (19) do not vanish, which proves the statement. □\square

For negative δ\delta inside the domain of analyticity, the formulas (8)–(19) define the continuation of the branches selected at δ≥0\delta\geq0. For δ<0\delta<0 boundedness no longer excludes the third exponential term; therefore the identification of this continuation with the expectation (1) requires a separate justification; at u=0u=0 it is given after Proposition 4 (at u>0u>0 the argument is the same).

Theorem 5 and Proposition 3 give the moment series

φδ(0)=∑n≥0(−δ)nn!E0[Tn;T<∞],\varphi_\delta(0)=\sum_{n\geq0}\frac{(-\delta)^n}{n!}\mathbb{E}_0[T^n;T<\infty],

and in particular, in the model case,

φδ(0)=13−δ24+25δ23456−δ36E0[T3;T<∞]+O(δ4).(20)\varphi_\delta(0)=\frac{1}{3}-\frac{\delta}{24}+\frac{25\delta^2}{3456}-\frac{\delta^3}{6}\mathbb{E}_0[T^3;T<\infty]+O(\delta^4). \tag*{(20)}

Proposition 4 In the model case the singularity of the constructed analytic continuation of the function δ↦φδ(0)\delta\mapsto\varphi_\delta(0) closest to zero lies at δ=−γ1\delta=-\gamma_1, where

γ1=c(μ−ρ3)+λ−λμ2ρ32,\gamma_1=c(\mu-\rho_3)+\lambda-\frac{\lambda\mu^2}{\rho_3^2},
ρ3=(2λμ2c)1/3,x∗=μ−ρ3.\rho_3=\left(\frac{2\lambda\mu^2}{c}\right)^{1/3},\qquad x_*=\mu-\rho_3.

Numerically γ1=21−993=2.2792455925…\gamma_1=21-9\sqrt[3]{9}=2.2792455925\ldots and x∗=0.9199161769…x_*=0.9199161769\ldots.

Proof. A singularity of the formulas (18)–(19) can arise only from a coalescence of branches, since djd_j does not vanish at the root. Hence the equations qδ(x)=∂xqδ(x)=0q_\delta(x)=\partial_xq_\delta(x)=0 must hold simultaneously. From qδ(x)=(cx+λ+δ)(μ−x)2−λμ2q_\delta(x)=(cx+\lambda+\delta)(\mu-x)^2-\lambda\mu^2 and ∂xqδ=c(μ−x)2−2(cx+λ+δ)(μ−x)\partial_xq_\delta=c(\mu-x)^2-2(cx+\lambda+\delta)(\mu-x), at y=μ−x≠0y=\mu-x\ne0 we obtain cy3=2λμ2cy^3=2\lambda\mu^2 and δ=λμ2/y2−c(μ−y)−λ\delta=\lambda\mu^2/y^2-c(\mu-y)-\lambda. The real solution y=ρ3y=\rho_3 gives x∗=μ−ρ3x_*=\mu-\rho_3 and δ=−γ1\delta=-\gamma_1. The case y=0y=0 is impossible, since qδ(μ)=−λμ2≠0q_\delta(\mu)=-\lambda\mu^2\ne0. The two remaining complex conjugate solutions are obtained by replacing ρ3\rho_3 with the other roots of the cubic equation; the modulus of the corresponding δ\delta equals 34.41805468…>γ134.41805468\ldots>\gamma_1. As we move from zero to the left, at the real singularity the branch x1x_1 coalesces with the branch continuing the zero root, while the branch x2x_2 remains simple.

The point found is a genuine singularity, not a removable one. There qxx=−3cρ3≠0q_{xx}=-3c\rho_3\ne0 and ∂δq=(μ−x∗)2≠0\partial_\delta q=(\mu-x_*)^2\ne0, hence

x1(δ)=x∗+a(δ+γ1)1/2+O(δ+γ1),a≠0.x_1(\delta)=x_*+a(\delta+\gamma_1)^{1/2}+O(\delta+\gamma_1),\qquad a\ne0.

If FF is the right-hand side of (19), then direct differentiation gives at the critical point ∂F/∂x1=−λc2x2/(d12d22)≠0\partial F/\partial x_1=-\lambda c^2x_2/(d_1^2d_2^2)\ne0. Hence the branch point of square-root type survives. □\square

For illustration, the values of the analytic continuation itself are φ−0.5(0)=0.356219\varphi_{-0.5}(0)=0.356219, φ−1(0)=0.384591\varphi_{-1}(0)=0.384591, φ−1.5(0)=0.422364\varphi_{-1.5}(0)=0.422364, φ−2(0)=0.482496\varphi_{-2}(0)=0.482496 and φ−2.25(0)=0.551167\varphi_{-2.25}(0)=0.551167 to the digits shown. For −γ1<δ≤0-\gamma_1<\delta\le0 this continuation coincides with E0[e∣δ∣T;T<∞]\mathbb{E}_0[e^{|\delta|T};T<\infty]: by Theorem 5 the coefficients of the moment series equal E0[Tn;T<∞]/n!≥0\mathbb{E}_0[T^n;T<\infty]/n!\ge0, the series converges in the disk ∣δ∣<γ1|\delta|<\gamma_1, and it remains to apply Tonelli’s theorem. For ∣δ∣>γ1|\delta|>\gamma_1 the corresponding exponential moment is infinite, so that γ1\gamma_1 is the rate of exponential decay of the distribution of TT on {T<∞}\{T<\infty\}.

Proposition 5. In the model case, exactly,

E0[T3;T<∞]=13913824=0.01005497685185…,\mathbb{E}_0[T^3;T<\infty]=\frac{139}{13824}=0.01005497685185\ldots,
E0[T3∣T<∞]=1394608.\mathbb{E}_0[T^3\mid T<\infty]=\frac{139}{4608}.

Proof. We expand both positive roots x1(δ),x2(δ)x_1(\delta),x_2(\delta) from (6) to the third order in δ\delta and substitute the two decaying exponential terms from (8) into the boundary system (4). Solving for φδ(0)=A1(δ)+A2(δ)\varphi_\delta(0)=A_1^{(\delta)}+A_2^{(\delta)} gives

φδ(0)=13−δ24+25δ23456−139δ382944+O(δ4).\varphi_\delta(0)=\frac{1}{3}-\frac{\delta}{24}+\frac{25\delta^2}{3456}-\frac{139\delta^3}{82944}+O(\delta^4).

The comparison with (20) shows that E0[T3;T<∞]=139/13824\mathbb{E}_0[T^3;T<\infty]=139/13824. The conditional moment is obtained by dividing by P0(T<∞)=1/3\mathbb{P}_0(T<\infty)=1/3. The numerical values of Table 4 agree with this exact result. □\square

δ\deltaQ(δ)Q(\delta)
0.0050.0100411
0.010.0100272
0.020.0099995
0.030.0099721
0.050.0099177

Table 4. Extraction of the third moment: the quantities Q(δ)=6(13−δ24+25δ23456−φδ(0))/δ3Q(\delta)=6\left(\frac{1}{3}-\frac{\delta}{24}+\frac{25\delta^2}{3456}-\varphi_\delta(0)\right)/\delta^3 approach E0[T3;T<∞]\mathbb{E}_0[T^3;T<\infty] linearly in δ\delta.

Remark 5. The approximation by the terms up to the second order in (20) has a cubic remainder. Indeed, the difference between φδ(0)\varphi_\delta(0) and 1/3−δ/24+25δ2/34561/3-\delta/24+25\delta^2/3456 at δ=0.01,0.05,0.1,0.2\delta=0.01,0.05,0.1,0.2 equals respectively −1.67⋅10−9-1.67\cdot10^{-9}, −2.07⋅10−7-2.07\cdot10^{-7}, −1.63⋅10−6-1.63\cdot10^{-6}, −1.27⋅10−5-1.27\cdot10^{-5} to the digits shown. The division of these values by −δ3-\delta^3 gives quantities close to E0[T3;T<∞]/6=1.68⋅10−3\mathbb{E}_0[T^3;T<\infty]/6=1.68\cdot10^{-3}.

Conclusion and open questions

For claims with the Erlang distribution of shape parameter 2, we obtain a closed description of the Laplace transform of the time of ruin: the two admissible exponential terms are determined by the characteristic equation and the boundary conditions. This makes it possible to compute both the ruin probability and the moments of the time of ruin, keeping the full truncated convolution in all the checks.

For the general Ek(μ)E_k(\mu) distribution a finite-dimensional system is constructed; the structure of the roots, however, is more complicated: in addition to real exponential terms, complex conjugate pairs are possible, and the explicit formulas for the coefficients of the kk decaying terms require a separate analysis. Extensions to models with proportional investment and two-sided jumps, where the variable coefficients or additional unbounded terms require a different boundary value setting, are separate directions of further work.

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.

References

  1. [1]Belkina, T.A., Konyukhova, N.B., Kurochkin, S.V.: Singular boundary value problem for the integrodifferential equation in an insurance model with stochastic premiums: analysis and numerical solution. Comput. Math. Math. Phys. 52, 1384–1416 (2012)
  2. [2]Belkina, T.A., Konyukhova, N.B., Kurochkin, S.V.: Singular initial-value and boundary-value problems for integrodifferential equations in dynamical insurance models with investments. J. Math. Sci. 218, 369–394 (2016). Transl. from Sovrem. Mat. Fundam. Napravl. 53, 5–29 (2014)DOI
  3. [3]Belkina, T.A., Konyukhova, N.B., Kurochkin, S.V.: Dynamical insurance models with investment: constrained singular problems for integrodifferential equations. Comput. Math. Math. Phys. 56, 43–92 (2016)DOI
  4. [4]Feller, W.: An Introduction to Probability Theory and Its Applications. Vol. II, 2nd ed. Wiley, New York (1971)
  5. [5]Gerber, H.U., Shiu, E.S.W.: On the time value of ruin. N. Am. Actuar. J. 2, 48–72 (1998)DOI
  6. [6]Gerber, H.U., Shiu, E.S.W.: The time value of ruin in a Sparre Andersen model. N. Am. Actuar. J. 9, 49–69 (2005)DOI
  7. [7]Kabanov, Yu., Pukhlakov, N.: Ruin probabilities with investments: smoothness, integro-differential and ordinary differential equations, asymptotic behavior. J. Appl. Probab. 59, 556–570 (2022)DOI
  8. [8]Lin, X.S., Willmot, G.E.: Analysis of a defective renewal equation arising in ruin theory. Insur. Math. Econ. 25, 63–84 (1999)DOI
  9. [9]Lin, X.S., Willmot, G.E.: The moments of the time of ruin, the surplus before ruin, and the deficit at ruin. Insur. Math. Econ. 27, 19–44 (2000)DOI
  10. [10]Promyslov, P.: Existence of a classical solution to the integro-differential equation arising in the Cramér–Lundberg non-life insurance model with proportional investment. arXiv:2604.05143 (2026)arxiv.org/abs/2604.05143
  11. [11]Promyslov, P.: On the integro-differential equation arising in the ruin problem for annuity payment models. arXiv:2601.01447 (2026)arxiv.org/abs/2601.01447
  12. [12]Sundt, B., Teugels, J.L.: Ruin estimates under interest force. Insur. Math. Econ. 16, 7–22 (1995)DOI

Paper details

vlnc:W_qyav748c
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents