betaFeedback

WorksMathematics

An Exact Poisson Mixture Representation of Siler Remaining Life Expectancy

Abstract

The Siler model describes mortality from birth to old age as the sum of a declining immaturity hazard, a constant background hazard, and an increasing senescence hazard. Its remaining life expectancy has no known closed form, and numerical evaluation alone does not reveal how the immaturity hazard enters it. We show that the remaining-lifetime distribution at age xx is Gompertz–Makeham with a Poisson-distributed Makeham term. Successive Makeham terms differ by the rate at which immaturity mortality declines, and the Poisson mean is the cumulative immaturity hazard remaining after age xx. In a latent-clock construction, individuals carry a Poisson number of constant-rate clocks, and selection removes those carrying more clocks first. Integrating the mixture yields an exact series for remaining life expectancy with a closed-form truncation bound. The life expectancy lost to immaturity mortality is close to the probability of carrying a clock times life expectancy with that hazard removed, with a closed-form error bound. In annual fits to Swedish female mortality, this loss at birth fell from 4.3 years in 1900 to less than 0.01 years in 2020, a decline equal to 15% of the gain in life expectancy at birth, and the approximation was always within 0.06 years.

Introduction

Siler [18] wrote mortality as the sum of three competing risks: an immaturity hazard that falls exponentially from birth, a constant background hazard, and a senescence hazard that rises exponentially with age. With five parameters, the model describes the whole life course and has been fitted to human and animal populations alike [19, 10]. Remaining life expectancy summarizes such a fit. Under the competing-risk interpretation, it also determines the probability of eventually dying from the background risk, which equals the background hazard times remaining life expectancy [15]. The Siler survival function is explicit, but its integral has no known closed form.

Without the immaturity hazard, the Siler model reduces to the Gompertz–Makeham model [12], whose remaining life expectancy is an incomplete gamma function of the senescence hazard at the starting age. The immaturity hazard prevents this reduction. It declines with age while the senescence hazard grows, so the two exponential terms do not combine into a single incomplete gamma function. Numerical quadrature evaluates remaining life expectancy readily, but it does not show how the declining early-life hazard shapes the remaining-lifetime distribution, why that hazard declines, or how much life expectancy it costs.

We address these questions with a latent-clock representation of the immaturity hazard. It treats each individual as if they carried, from birth, a random number of hidden vulnerabilities, or latent clocks. Each clock can cause death at the same constant rate. An individual’s immaturity hazard then does not decline with age: it equals the number of clocks carried times their common rate. Conditional on that number, the remaining lifetime is Gompertz–Makeham, with the combined clock rate added to the background hazard. The number of clocks is Poisson distributed, so the Siler remaining-lifetime distribution is a Poisson mixture of Gompertz–Makeham distributions. Its mixing mean is the cumulative immaturity hazard that remains at the starting age, which falls with age because individuals carrying more clocks die first.

This identity reduces calculations under the Siler model to Gompertz–Makeham ones. We integrate it to obtain an exact series for remaining life expectancy and bound its truncation error in years. We then use the same bound to relate the life expectancy lost to immaturity mortality to the probability of carrying at least one clock. We illustrate the results with annual Siler fits to Swedish female mortality from 1900 to 2020.

The Siler model

Let TT denote age at death. The Siler model specifies the hazard

μ(t)=μ1(t)+a2+μ3(t),μ1(t)=a1e−b1t,μ3(t)=a3eb3t,\mu(t) = \mu_1(t) + a_2 + \mu_3(t), \qquad\mu_1(t) = a_1 e^{-b_1 t}, \qquad\mu_3(t) = a_3 e^{b_3 t},

with a1,a2≥0a_1, a_2 \ge0 and a3,b1,b3>0a_3, b_1, b_3 > 0. The immaturity hazard μ1\mu_1 declines at rate b1b_1, a2a_2 is the background hazard, and the senescence hazard μ3\mu_3 increases at rate b3b_3. For a life aged xx, the remaining lifetime Tx=T−x∣T>xT_x = T - x \mid T > x has survival function

tpx=exp⁡{−∫xx+tμ(u) du}=exp⁡{−Ax(1−e−b1t)−a2t−Gx(eb3t−1)}.(1){}_t p_x = \exp\left\{-\int_x^{x+t} \mu(u)\,\mathrm{d}u\right\} = \exp\left\{-A_x\left(1-e^{-b_1t}\right)-a_2t-G_x\left(e^{b_3t}-1\right)\right\}. \tag*{(1)}

where

Ax=∫x∞μ1(u) du=a1e−b1xb1,Gx=μ3(x)b3=a3eb3xb3.A_x = \int_x^\infty\mu_1(u)\,\mathrm{d}u = \frac{a_1e^{-b_1x}}{b_1}, \qquad G_x = \frac{\mu_3(x)}{b_3} = \frac{a_3e^{b_3x}}{b_3}.

Here AxA_x is the remaining cumulative immaturity hazard after age xx, and e−Axe^{-A_x} is the probability of surviving it. The remaining life expectancy at age xx is

ex=E(Tx)=∫0∞tpx dt.(2)e_x = \mathbb{E}(T_x) = \int_0^\infty{}_t p_x\,\mathrm{d}t. \tag*{(2)}

Conditioning on survival to age xx replaces a1a_1 and a3a_3 by μ1(x)\mu_1(x) and μ3(x)\mu_3(x) and leaves b1b_1, a2a_2, and b3b_3 unchanged. The remaining lifetime at every age therefore follows a Siler law, and the results below hold for every xx.

The Siler model as a Poisson mixture

A hazard that declines in the population can arise from individual hazards that do not decline, provided individuals differ in their level of risk. We use this idea to write the Siler immaturity hazard as heterogeneity in the number of latent clocks that individuals carry.

Let N0∼Poisson⁡(A0)N_0 \sim\operatorname{Poisson}(A_0) be the number of latent clocks carried at birth. Conditional on N0N_0, the clocks ring, that is, cause death, at independent exponential times C1,…,CN0C_1,\ldots,C_{N_0} with rate b1b_1, and they are independent of the background and senescence failure times. The immaturity failure time is the first ring, C=min⁡jCjC=\min_j C_j, with C=∞C=\infty when N0=0N_0=0. Its survival function is

Pr⁡(C>t)=E(e−b1N0t)=exp⁡{−A0(1−e−b1t)}.\Pr(C>t)=\mathbb{E}(e^{-b_1N_0t})=\exp\{-A_0(1-e^{-b_1t})\}.

which is the survival function of the Siler immaturity hazard. Survival to age xx carries information about the clock count. By Bayes’ rule, with Ax=A0e−b1xA_x=A_0e^{-b_1x}, it exponentially tilts the Poisson probabilities:

Pr⁡(N0=n∣C>x)∝Pr⁡(N0=n)Pr⁡(C>x∣N0=n)=e−A0(A0e−b1x)nn!∝(Ax)nn!.\Pr(N_0=n\mid C>x)\propto\Pr(N_0=n)\Pr(C>x\mid N_0=n) =e^{-A_0}\frac{(A_0e^{-b_1x})^n}{n!}\propto\frac{(A_x)^n}{n!}.

Survival from the background and senescence risks carries no information about N0N_0, so the same law holds among those who survive all three risks to age xx. Let KxK_x denote the number of clocks carried by a survivor at age xx; then Kx∼Poisson⁡(Ax)K_x\sim\operatorname{Poisson}(A_x). A survivor carrying KxK_x clocks has immaturity hazard b1Kxb_1K_x, and the population immaturity hazard is its average among survivors:

μ1(x)=b1E(Kx)=b1Ax=a1e−b1x.\mu_1(x)=b_1\mathbb{E}(K_x)=b_1A_x=a_1e^{-b_1x}.

This identity gives the declining hazard a selection interpretation. Each clock has constant rate b1b_1, but individuals carrying more clocks die sooner, so the mean clock count among survivors, and with it the population hazard, falls exponentially at rate b1b_1. Vaupel and Yashin [21] show that a mixture of constant individual hazards produces a declining population hazard in this way. This heterogeneity differs from the gamma frailty of Vaupel et al. [20], which multiplies the individual hazard and slows the rise of senescent mortality. Here the clocks add to the immaturity hazard alone, and their number follows a Poisson rather than a gamma distribution. Not every declining hazard admits a clock representation of this kind. Appendix § 10.§ 2 shows that an immaturity hazard can be generated by clocks with constant, possibly different, rates if and only if it is completely monotone, that is, a mixture of exponentially declining hazards. The Siler hazard is the case in which all clocks share one rate.

In evolutionary demography, the decline of mortality during development is called ontogenescence, and Levitis [11] lists heterogeneous frailty among its explanations. Under this reading, a clock is a persistent early-life vulnerability that raises the individual hazard by b1b_1 until death. The construction is an exact representation of the population survival function rather than an identified biological mechanism, because aggregate survival cannot distinguish it from mechanisms in which individual risk itself declines with development.

Proposition 1 (Distributional Poisson mixture). For every x,t≥0x,t\ge0,

tpx=E[exp⁡{−(a2+b1Kx)t−Gx(eb3t−1)}],Kx∼Poisson⁡(Ax).(3){}_{t}p_x=\mathbb{E}\left[\exp\left\{-(a_2+b_1K_x)t-G_x\left(e^{b_3t}-1\right)\right\}\right],\qquad K_x\sim\operatorname{Poisson}(A_x). \tag*{(3)}

Thus the Siler remaining lifetime at age xx is a Poisson mixture of Gompertz–Makeham remaining lifetimes with Makeham term a2+b1Kxa_2+b_1K_x.

Proof. By memorylessness, conditional on Kx=nK_x=n, the residual clock times are independent exponentials with rate b1b_1, so their joint survival over the next tt years is e−nb1te^{-nb_1t}. By the assumed independence of the competing risks, multiplying this factor by the background and senescence survival functions gives the integrand in Equation (3); averaging over KxK_x gives tpx{}_{t}p_x.

Proposition 1 applies immediately to functionals that are linear in the remaining-lifetime distribution, which include remaining life expectancy, moments, the continuous life annuity, and the Laplace transform. Quantiles and other nonlinear functionals are determined by the mixture distribution but are not, in general, averages of their component values.

This representation complements the competing-risk mixture in Patricio and Missov [15]. Their mixture conditions on the cause that produces death, whereas Proposition 1 mixes over the number of clocks carried at age xx. The constant-shift argument does not depend on the Gompertz form of the senescence hazard, and Appendix § 10.§ 3 gives the corresponding result for a general adult hazard and its gamma–Gompertz specialization.

To obtain the remaining life expectancy explicitly, we write the relevant integrals in terms of the generalized exponential, or integro-exponential, function

Eν(z)=∫1∞u−νe−zu du=zν−1Γ(1−ν,z),z>0, ν∈R,E_\nu(z)=\int_1^\infty u^{-\nu}e^{-zu}\,\mathrm{d}u=z^{\nu-1}\Gamma(1-\nu,z),\qquad z>0,\ \nu\in\mathbb{R},

where

Γ(s,z)=∫z∞vs−1e−v dv\Gamma(s,z)=\int_z^\infty v^{s-1}e^{-v}\,\mathrm{d}v

is the upper incomplete gamma function (Milgram [13]; Olver [14]). Writing exGM(c)e_x^{\mathrm{GM}}(c) for remaining life expectancy under the Gompertz–Makeham hazard c+a3eb3tc+a_3e^{b_3t}, the substitution u=eb3tu=e^{b_3t} gives

exGM(c)=eGxb3E1+c/b3(Gx)=eGxb3(Gx)c/b3Γ(−cb3,Gx).(4)e_x^{\mathrm{GM}}(c)=\frac{e^{G_x}}{b_3}E_{1+c/b_3}(G_x)=\frac{e^{G_x}}{b_3}(G_x)^{c/b_3}\Gamma\left(-\frac{c}{b_3},G_x\right). \tag*{(4)}

the known expression for Gompertz–Makeham remaining life expectancy (Castellares et al. [5]).

Corollary 1 (Siler remaining life expectancy). Let wn(x)=Pr⁡(Kx=n)=e−Ax(Ax)n/n!w_n(x)=\Pr(K_x=n)=e^{-A_x}(A_x)^n/n!, and Qn(x)=exGM(a2+nb1)Q_n(x)=e_x^{\mathrm{GM}}(a_2+nb_1). Then

ex=E[exGM(a2+b1Kx)]=∑n=0∞wn(x)Qn(x),(5)e_x=\mathbb{E}\left[e_x^{\mathrm{GM}}(a_2+b_1K_x)\right]=\sum_{n=0}^{\infty}w_n(x)Q_n(x), \tag*{(5)}

and

ex=eGx−Axb3∑n=0∞(Ax)nn!E1+(a2+nb1)/b3(Gx).(6)e_x=\frac{e^{G_x-A_x}}{b_3}\sum_{n=0}^{\infty}\frac{(A_x)^n}{n!}E_{1+(a_2+nb_1)/b_3}(G_x). \tag*{(6)}

Proof. Integrating Equation (3) and applying Tonelli’s theorem gives the expectation in Equation (5); expanding that expectation gives the sum. Substituting Equation (4) into the sum gives Equation (6).

When a1=0a_1 = 0, Ax=0A_x = 0 and only the component n=0n = 0 remains. In general, the leading term

ex(0)=e−AxexGM(a2)e_x^{(0)} = e^{-A_x}e_x^{\mathrm{GM}}(a_2)

is the Gompertz–Makeham life expectancy with Makeham term a2a_2, weighted by the probability of carrying no clock. As xx increases, AxA_x decays at rate b1b_1, so the mixing law concentrates on Kx=0K_x = 0 and exe_x approaches the Gompertz–Makeham value.

The clock representation also separates carrying an immaturity risk from dying from it. Let π1(x)\pi_1(x) denote the probability that a life aged xx eventually dies from the immaturity risk. Conditional on Kx=nK_x = n, the immaturity hazard is the constant nb1nb_1, so this probability is nb1Qn(x)nb_1Q_n(x). Averaging over KxK_x gives

π1(x)=b1E[KxQKx(x)]=b1AxE[QKx+1(x)].\pi_1(x) = b_1\mathbb{E}[K_xQ_{K_x}(x)] = b_1A_x\mathbb{E}[Q_{K_x+1}(x)].

The second equality uses nwn(x)=Axwn−1(x)nw_n(x) = A_xw_{n-1}(x).

By contrast, Pr⁡(Kx≥1)=1−e−Ax\Pr(K_x \ge1) = 1 - e^{-A_x} is the probability of carrying at least one clock, not the probability that a clock rings before the competing risks. The background probability is a2exa_2e_x, as in [15], and the senescence probability is the remainder.

Remark 1 (Laplace transform). Discounting at force of interest δ≥0\delta\ge0 multiplies tpx{}_tp_x by e−δte^{-\delta t}, which by Equation (1) gives the survival function of a Siler model with background hazard a2+δa_2 + \delta and the same AxA_x and GxG_x. Corollary 1 therefore gives the continuous life annuity

∫0∞e−δttpx dt=E[exGM(a2+δ+b1Kx)].\int_0^\infty e^{-\delta t}{}_tp_x\,\mathrm{d}t = \mathbb{E}\left[e_x^{\mathrm{GM}}(a_2 + \delta+ b_1K_x)\right].

as a Poisson mixture of the Gompertz–Makeham annuities of Castellares et al. [4]. Because E(e−δTx)=1−δ∫0∞e−δttpx dt\mathbb{E}(e^{-\delta T_x}) = 1 - \delta\int_0^\infty e^{-\delta t}{}_tp_x\,\mathrm{d}t, the Laplace transform of the remaining lifetime is the same mixture of Gompertz–Makeham Laplace transforms.

Truncation error

Each mixture component is a life expectancy under a total hazard of at least a2+nb1a_2 + nb_1, so it has a simple bound.

Lemma 1. The sequence Qn(x)Q_n(x) is nonincreasing in nn, and

Qn(x)≤1a2+nb1,Q_n(x) \le\frac{1}{a_2 + nb_1},

whenever a2+nb1>0a_2 + nb_1 > 0.

Proof. By the definition of Qn(x)Q_n(x),

Qn(x)=∫0∞exp⁡{−(a2+nb1)t−Gx(eb3t−1)} dt.Q_n(x) = \int_0^\infty\exp\left\{-(a_2 + nb_1)t - G_x\left(e^{b_3t} - 1\right)\right\}\,\mathrm{d}t.

The integrand decreases in nn and is bounded by e−(a2+nb1)te^{-(a_2+nb_1)t}.

Proposition 2. For m≥0m \ge0, let ex(m)=∑n=0mwn(x)Qn(x)e_x^{(m)} = \sum_{n=0}^{m} w_n(x)Q_n(x). Then ex(m)↑exe_x^{(m)} \uparrow e_x as m→∞m \to\infty, and

0≤ex−ex(m)≤Pr⁡(Kx>m)a2+(m+1)b1.(7)0 \le e_x - e_x^{(m)} \le\frac{\Pr(K_x > m)}{a_2 + (m+1)b_1}. \tag*{(7)}

Proof. By Corollary 1, ex−ex(m)=∑n>mwn(x)Qn(x)e_x-e_x^{(m)}=\sum_{n>m}w_n(x)Q_n(x). All terms are nonnegative, so ex(m)e_x^{(m)} is nondecreasing in mm and converges to exe_x. By Lemma 1, Qn(x)≤1/[a2+(m+1)b1]Q_n(x)\le1/[a_2+(m+1)b_1] for every n>mn>m. Hence

0≤ex−ex(m)≤1a2+(m+1)b1∑n>mwn(x)=Pr⁡(Kx>m)a2+(m+1)b1.0 \le e_x - e_x^{(m)} \le\frac{1}{a_2+(m+1)b_1}\sum_{n>m}w_n(x)=\frac{\Pr(K_x>m)}{a_2+(m+1)b_1}.

For the leading term, Equation (7) reads

0≤ex−ex−AxexGM(a2)≤1−e−Axa2+b1≤Axa2+b1.(8)0 \le e_x-e_x^{-\mathit{A}_x}e_x^{\mathrm{GM}}(a_2)\le\frac{1-e^{-\mathit{A}_x}}{a_2+b_1}\le\frac{\mathit{A}_x}{a_2+b_1}. \tag*{(8)}

The leading term is therefore accurate when little cumulative immaturity hazard remains or when immaturity mortality declines rapidly. The same bounds hold for the integral in Remark 1, with a2a_2 replaced by a2+δa_2+\delta.

Removing the immaturity hazard turns the Siler model into the Gompertz–Makeham model. Thus exGM(a2)e_x^{\mathrm{GM}}(a_2) is the cause-deleted remaining life expectancy, and Δx=exGM(a2)−ex\Delta_x=e_x^{\mathrm{GM}}(a_2)-e_x is the gain in remaining life expectancy from deleting immaturity mortality [9, 1]. Rearranging Equation (8) gives

(1−e−Ax)[exGM(a2)−1a2+b1]≤Δx≤(1−e−Ax)exGM(a2).(9)(1-e^{-\mathit{A}_x})\left[e_x^{\mathrm{GM}}(a_2)-\frac{1}{a_2+b_1}\right]\le\Delta_x\le(1-e^{-\mathit{A}_x})e_x^{\mathrm{GM}}(a_2). \tag*{(9)}

Here 1−e−Ax=Pr⁡(Kx≥1)1-e^{-\mathit{A}_x}=\Pr(K_x\ge1) is the probability of carrying at least one clock. Dividing the remainder in Equation (8) by this probability gives E[QKx(x)∣Kx≥1]\mathbb{E}[Q_{K_x}(x)\mid K_x\ge1], the mean remaining lifetime under all three risks among clock carriers. Every carrier has background-plus-immaturity hazard at least a2+b1a_2+b_1, so this conditional mean is at most 1/(a2+b1)1/(a_2+b_1). Hence the mean gain from deleting immaturity mortality among clock carriers lies between exGM(a2)−1/(a2+b1)e_x^{\mathrm{GM}}(a_2)-1/(a_2+b_1) and exGM(a2)e_x^{\mathrm{GM}}(a_2).

Application to Swedish female mortality

We use female death counts DxD_x and exposures ExE_x for Sweden, 1900–2020, at single ages 0–109, from the Human Mortality Database [8]. For each year, we assume Dx∼Poisson⁡(Exμ(x))D_x\sim\operatorname{Poisson}(E_x\mu(x)) independently across ages [2] and estimate θ=(a1,b1,a2,a3,b3)\boldsymbol{\theta}=(a_1,b_1,a_2,a_3,b_3) by maximizing the log-likelihood

ℓ(θ)=∑x{Dxlog⁡μ(x)−Exμ(x)}.\ell(\boldsymbol{\theta})=\sum_x\{D_x\log\mu(x)-E_x\mu(x)\}.

Figure 1 compares the fitted Siler remaining life expectancy with HMD values at ages 0, 30, and 60. The fitted model overstates e0e_0 by up to 2.6 years before 1940 and understates e60e_{60} by 1.7–3.1% throughout. Part of the gap is definitional, since HMD life tables smooth death rates at old ages and use specific assumptions for the average time lived in the first year of life by infants who die. We use the fits to evaluate the series and the loss bounds rather than to assess the Siler model as a description of Swedish mortality.

Figure 1. Remaining life expectancy for Swedish females, 1900–2020, at ages 0, 30, and 60. Top row: HMD life-table values and fitted Siler values. Middle row: fitted minus HMD, in years. Bottom row: the same difference as a percentage of the HMD value.

The loss to immaturity mortality nearly vanished

The life expectancy lost to immaturity mortality fell from 4.28 years in 1900 to 0.49 years in 1950 and 0.004 years in 2020 (Figure 2). Over the same period, e0e_0 rose by 27.9 years and e0GM(a2)e_0^{\mathrm{GM}}(a_2) by 23.6 years, so under this cause-deleted decomposition the decline of immaturity mortality accounts for 15% of the gain. The leading-term approximation is close in every year: Δ0\Delta_0 lies between 98.6% and 99.99% of its upper bound (1−e−A0)e0GM(a2)(1-e^{-A_0})e_0^{\mathrm{GM}}(a_2) in Equation (9) and differs from that bound by at most 0.060 years. The probability of carrying at least one clock, multiplied by life expectancy with the immaturity hazard removed, therefore approximates the loss at birth closely.

Figure 2. Swedish females, 1900–2020. Left: fitted Siler life expectancy at birth e0e_0 and the Gompertz–Makeham life expectancy with the immaturity hazard removed, e0GM(a2)e_0^{\mathrm{GM}}(a_2). Right: the loss Δ0=e0GM(a2)−e0\Delta_0=e_0^{\mathrm{GM}}(a_2)-e_0 and its leading-term value (1−e−A0)e0GM(a2)(1-e^{-A_0})e_0^{\mathrm{GM}}(a_2).

A few terms of the series suffice

We evaluate Eν(z)E_\nu(z) through the incomplete gamma function in the R package expint [7], falling back to numerical integration of

ezEν(z)=∫0∞exp⁡{−z(es−1)−(ν−1)s} dse^z E_\nu(z)=\int_0^\infty\exp\{-z(e^s-1)-(\nu-1)s\}\,\mathrm{d}s

when the routine returns a nonfinite value. We stop the series at the smallest mm for which the bound in Equation (7) is below 10−1210^{-12} years. For the years 1920, 1940, …, 2020 and ages 0, 30, and 60, the series and adaptive quadrature of Equation (2) differ by at most 1.08×10−121.08\times10^{-12} years (Table S1). These differences come from the numerical routines, not from truncation: at ages 30 and 60, the truncation bound is below 10−1610^{-16} years.

YearAgeValueAbsolute differenceAnalytic bound
1920061.93012.70×10−132.70 \times10^{-13}2.64×10−132.64 \times10^{-13}
19203040.81162.13×10−142.13 \times10^{-14}<10−16< 10^{-16}
19206017.08741.42×10−141.42 \times10^{-14}<10−16< 10^{-16}
1940069.0828<10−16< 10^{-16}1.59×10−131.59 \times10^{-13}
19403042.49864.97×10−144.97 \times10^{-14}<10−16< 10^{-16}
19406016.75879.95×10−149.95 \times10^{-14}<10−16< 10^{-16}
1960075.11189.95×10−149.95 \times10^{-14}3.09×10−153.09 \times10^{-15}
19603046.29702.42×10−132.42 \times10^{-13}<10−16< 10^{-16}
19606018.82745.79×10−135.79 \times10^{-13}<10−16< 10^{-16}
1980078.71067.11×10−147.11 \times10^{-14}4.57×10−144.57 \times10^{-14}
19803049.48738.74×10−138.74 \times10^{-13}<10−16< 10^{-16}
19806021.50618.88×10−148.88 \times10^{-14}<10−16< 10^{-16}
2000081.71008.38×10−138.38 \times10^{-13}1.22×10−141.22 \times10^{-14}
20003052.13818.88×10−138.88 \times10^{-13}<10−16< 10^{-16}
20006023.59451.08×10−121.08 \times10^{-12}<10−16< 10^{-16}
2020083.93847.11×10−137.11 \times10^{-13}1.92×10−161.92 \times10^{-16}
20203054.32524.55×10−134.55 \times10^{-13}<10−16< 10^{-16}
20206025.32992.91×10−132.91 \times10^{-13}<10−16< 10^{-16}

Table S1. Numerical verification of the exact Siler series. The value column reports the series to four decimal places; adaptive quadrature gives the same displayed values. The absolute difference compares the series, stopped at an absolute truncation tolerance of 10−1210^{-12} years, with adaptive quadrature. The final column is the analytic bound on the omitted series tail. Values below 10−1610^{-16} are printed as <10−16< 10^{-16}. Differences can exceed the bound because they also include the numerical error of the special-function routine and of the quadrature, which the bound does not cover.

Figure 3 shows the error of the partial sums ex(m)e_x^{(m)}, m=0,…,4m=0,\ldots,4, at ages 0–50 in 1950, and Figure S1 in Appendix § 11 shows its maximum over all years at each age from 0 to 100. The natural logarithm of the error declines almost linearly with age, with slope close to −(m+1)b1-(m+1)b_1, because the omitted tail is dominated by a term proportional to (Ax)m+1(A_x)^{m+1}. On the base-10 scale of the figures, the slope is −(m+1)b1/log⁡10-(m+1)b_1/\log10. Over all years and ages, the leading term alone is within 0.0601 years (0.107%) of exe_x. Each further term lowers the maximum error by a factor of 50 to 85, to 1.12×10−31.12 \times10^{-3}, 1.85×10−51.85 \times10^{-5}, 2.60×10−72.60 \times10^{-7}, and 3.12×10−93.12 \times10^{-9} years for m=1,…,4m=1,\ldots,4.

Figure 3. Truncation error of the Siler series for Swedish females in 1950: ∣ex(m)−ex∣\lvert e_x^{(m)}-e_x\rvert in years (left) and as a percentage of exe_x (right), m=0,…,4m=0,\ldots,4. The errors are evaluated by summing the nonnegative mixture terms beyond mm. Logarithmic scales; values below 10−5010^{-50} are omitted.

Figure S1. Largest truncation error of the Siler series across all fitted years, Swedish females, 1900–2020: ∣ex(m)−ex∣\lvert e_x^{(m)} - e_x \rvert in years (left) and as a percentage of exe_x (right), m=0,…,4m = 0, \ldots, 4. The errors are evaluated by summing the nonnegative mixture terms beyond mm. Logarithmic scales; values below 10−5010^{-50} are omitted.

Discussion and conclusion

The Siler remaining lifetime is Gompertz–Makeham with a Poisson-distributed Makeham term. This identity separates the two roles of the immaturity hazard. Its remaining cumulative hazard AxA_x sets the expected clock count among survivors, while b1b_1 is both the rate of each clock and the increment in the Makeham term between successive components. Adaptive quadrature suffices for numerical evaluation, but the series adds a guaranteed error in years and isolates the immaturity contribution.

The same representation yields a simple demographic approximation. The life expectancy lost to immaturity mortality is close to the probability of carrying at least one clock multiplied by life expectancy with the immaturity hazard removed. Equation (9) bounds the approximation for every parameter vector, because carriers survive on average for at most 1/(a2+b1)1/(a_2+b_1) years under all three risks. In the Swedish fits, the approximation is within 0.06 years at every age and in every year. Under the cause-deleted decomposition, the decline in immaturity mortality contributed about 4.3 of the 27.9 years gained in life expectancy at birth between 1900 and 2020.

The construction also gives the declining limb of the Siler curve an exact frailty representation, which corresponds to one of the explanations for ontogenescence reviewed by Levitis [11]. The characterization in Appendix § 10.§ 2 gives a check for any fitted immaturity hazard: if the hazard is not completely monotone, its decline cannot arise from constant-rate clocks. For a stretched-exponential immaturity hazard with 0<p<10 < p < 1, clock rates are random, so the persistent vulnerabilities may differ in lethality, and the proportional rate of decline of the immaturity hazard decreases with age.

Three caveats bound these results. First, the biological interpretation is not identifiable from aggregate mortality. The same population curve may arise when individual risks decline through maturation, acquired robustness, or developmental timing, and our analysis cannot distinguish these mechanisms from selection. Second, the construction retains the Siler model’s assumption that its three risks act independently, and the cause-deleted decomposition inherits this assumption. Third, the Swedish estimates are conditional on the fitted Siler model, which overstates e0e_0 by up to 2.6 years before 1940.

The immaturity component of the Siler model is a Poisson-distributed addition to the Makeham term. Results known for the Gompertz–Makeham model therefore extend to the whole life course, and the life expectancy lost to immaturity mortality becomes a simple quantity with an explicit bound.

Funding

This work was funded by the European Union (ERC, Born Once – Die Once, Grant agreement ID 101043983). Views and opinions expressed are those of the author only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.

Data and code availability

Female death counts, exposures, and life-table values were obtained from the Human Mortality Database. The underlying data are available from https://www.mortality.org/ under its access conditions. The materials required to reproduce the analyses, tables, and figures are openly available at https://github.com/scpatricio/siler-poisson-mixture/.

Author contributions

Silvio C. Patricio: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – original draft, Writing – review and editing, and Visualization.

Extensions beyond the Siler specification

The clock argument of § 3 uses neither the exponential form of the immaturity hazard nor the Gompertz form of the senescence hazard. We therefore state it for a general immaturity hazard μ1\mu_1 and a general adult hazard hh in place of μ3\mu_3, so that the total hazard is μ(t)=μ1(t)+a2+h(t)\mu(t)=\mu_1(t)+a_2+h(t). These results generalize the representation and are not proposals for new mortality models. Both μ1\mu_1 and hh are nonnegative and locally integrable. The immaturity hazard has finite and positive cumulative hazard A0=∫0∞μ1(u) duA_0=\int_0^\infty\mu_1(u)\,\mathrm{d}u, so Ax=∫x∞μ1(u) duA_x=\int_x^\infty\mu_1(u)\,\mathrm{d}u is finite for every x≥0x\ge0. If Ax=0A_x=0, as happens at ages beyond which μ1\mu_1 vanishes, survivors at age xx carry no clocks, Kx=0K_x=0 almost surely, and the clock terms below equal one. For a constant c≥0c\ge0, we write

ex,h(c)=∫0∞exp⁡{−ct−∫xx+th(u) du} dt(10)e_{x,h}(c)=\int_0^\infty\exp\left\{-ct-\int_x^{x+t}h(u)\,\mathrm{d}u\right\}\,\mathrm{d}t \tag*{(10)}

for remaining life expectancy under the adult hazard hh and the additive constant hazard cc. We assume ex,h(a2)<∞e_{x,h}(a_2)<\infty, which holds, for instance, when a2>0a_2>0 or when hh is bounded below by a positive constant. Because μ1≥0\mu_1\ge0, the remaining life expectancy exe_x is then finite as well.

Clocks with age-dependent ring times

Any immaturity hazard with finite cumulative hazard admits a clock representation if the clocks are allowed to ring at age-dependent rates. Let N0∼Poisson⁡(A0)N_0\sim\operatorname{Poisson}(A_0) and, conditional on N0N_0, let the clocks ring at independent times C1,…,CN0C_1,\ldots,C_{N_0} with common density μ1(t)/A0\mu_1(t)/A_0, independently of the background and adult failure times. The first ring C=min⁡jCjC=\min_j C_j, with C=∞C=\infty when N0=0N_0=0, has survival function

Pr⁡(C>t)=exp⁡{−A0(1−AtA0)}=exp⁡{−∫0tμ1(u) du}.\Pr(C>t)=\exp\left\{-A_0\left(1-\frac{A_t}{A_0}\right)\right\}=\exp\left\{-\int_0^t\mu_1(u)\,\mathrm{d}u\right\}.

Each clock has not rung by age xx with probability Ax/A0A_x/A_0, so the Bayes argument of § 3 shows that a survivor at age xx carries Kx∼Poisson⁡(Ax)K_x\sim\operatorname{Poisson}(A_x) clocks. When Ax>0A_x>0, their residual ring times are independent, with survival function Ax+t/AxA_{x+t}/A_x and, wherever Ax+t>0A_{x+t}>0, hazard

ρx(t)=μ1(x+t)Ax+t.\rho_x(t)=\frac{\mu_1(x+t)}{A_{x+t}}.

Proposition S1 (Clock representation). For every x,t≥0x,t\ge0,

tpx=E[(Ax+tAx)Kxexp⁡{−a2t−∫xx+th(u) du}],Kx∼Poisson⁡(Ax),(11){}_t p_x=\mathbb{E}\left[\left(\frac{A_{x+t}}{A_x}\right)^{K_x}\exp\left\{-a_2t-\int_x^{x+t}h(u)\,\mathrm{d}u\right\}\right],\qquad K_x\sim\operatorname{Poisson}(A_x), \tag*{(11)}

and

ex=e−Ax∑n=0∞(Ax)nn!∫0∞(Ax+tAx)nexp⁡{−a2t−∫xx+th(u) du} dt.(12)e_x=e^{-A_x}\sum_{n=0}^{\infty}\frac{(A_x)^n}{n!}\int_0^\infty\left(\frac{A_{x+t}}{A_x}\right)^n\exp\left\{-a_2t-\int_x^{x+t}h(u)\,\mathrm{d}u\right\}\,\mathrm{d}t. \tag*{(12)}

Proof. Conditional on Kx=nK_x=n, the residual ring times are independent with survival function Ax+t/AxA_{x+t}/A_x, so none of them rings in the next tt years with probability (Ax+t/Ax)n(A_{x+t}/A_x)^n. Multiplying this factor by the background and adult survival functions and averaging over KxK_x gives Equation (S2). Integrating over tt and applying Tonelli’s theorem gives Equation (S3).

Component nn in Equation (S3) is a remaining life expectancy under the total hazard a2+h(x+t)+nρx(t)a_2+h(x+t)+n\rho_x(t). The clock hazard ρx\rho_x generally varies with time since age xx, so the components are not obtained by adding a constant to the background hazard.

Constant-rate clocks

The clocks of Proposition S1 may have hazards that change with age. The selection reading of § 3 needs more: each clock rings at a constant rate, although rates may differ between clocks. A clock whose rate is a positive random variable RR with distribution function FF rings at a time with density E(Re−tR)=∫0∞ze−tz dF(z)\mathbb{E}(R e^{-tR})=\int_0^\infty z e^{-tz}\,dF(z), a mixture of exponential densities. We call μ1\mu_1 completely monotone if its derivatives alternate in sign, that is, (−1)kμ1(k)(t)≥0(-1)^k\mu_1^{(k)}(t)\ge0 for every k≥0k\ge0 and t>0t>0.

Proposition S2 (Constant-rate clocks). The immaturity hazard is generated by Poisson clocks whose constant rates are independent draws from a common distribution if and only if μ1\mu_1 is completely monotone. In that case there is a positive random rate RR, unique in distribution, such that

μ1(t)=A0E(Re−tR),Ax=A0E(e−xR).\mu_1(t)=A_0\mathbb{E}(R e^{-tR}),\qquad A_x=A_0\mathbb{E}(e^{-xR}).

A survivor at age xx carries Kx∼Poisson⁡(Ax)K_x\sim\operatorname{Poisson}(A_x) clocks whose rates Λx,1,Λx,2,…\Lambda_{x,1},\Lambda_{x,2},\ldots are independent of KxK_x and independent and identically distributed as a random rate Λx\Lambda_x whose distribution satisfies

E{g(Λx)}=E{g(R)e−xR}E(e−xR)\mathbb{E}\{g(\Lambda_x)\}=\frac{\mathbb{E}\{g(R)e^{-xR}\}}{\mathbb{E}(e^{-xR})}

for every bounded measurable function gg. With Yx=∑j=1KxΛx,jY_x=\sum_{j=1}^{K_x}\Lambda_{x,j}, and the empty sum equal to zero,

tpx=E[exp⁡{−(a2+Yx)t−∫xx+th(u) du}],ex=E[ex,h(a2+Yx)].(13){}^tp_x=\mathbb{E}\left[\exp\left\{-(a_2+Y_x)t-\int_x^{x+t}h(u)\,du\right\}\right],\qquad e_x=\mathbb{E}[e_{x,h}(a_2+Y_x)]. \tag*{(13)}

Proof. For the necessity, differentiating E(Re−tR)\mathbb{E}(R e^{-tR}) kk times gives (−1)kE(Rk+1e−tR)(-1)^k\mathbb{E}(R^{k+1}e^{-tR}), so every mixture of exponential densities is completely monotone. For the sufficiency, Bernstein’s theorem ([6]) writes a completely monotone μ1\mu_1 as μ1(t)=∫0∞e−tz dΨ(z)\mu_1(t)=\int_0^\infty e^{-tz}\,d\Psi(z), with Ψ\Psi nondecreasing and determined by μ1\mu_1. Tonelli’s theorem gives A0=∫0∞z−1 dΨ(z)A_0=\int_0^\infty z^{-1}\,d\Psi(z), so a finite A0A_0 excludes a jump of Ψ\Psi at z=0z=0, and F(z)=A0−1∫0zu−1 dΨ(u)F(z)=A_0^{-1}\int_0^z u^{-1}\,d\Psi(u) is a distribution function on (0,∞)(0,\infty). A rate RR with distribution function FF satisfies μ1(t)=A0E(Re−tR)\mu_1(t)=A_0\mathbb{E}(R e^{-tR}), and integrating from xx to ∞\infty gives Ax=A0E(e−xR)A_x=A_0\mathbb{E}(e^{-xR}). A clock with rate zz has not rung by age xx with probability e−xze^{-xz}, so Bayes’ rule reweights the rates of the clocks carried at age xx by e−xRe^{-xR}, which gives the law of Λx\Lambda_x. Proposition S1 gives Kx∼Poisson⁡(Ax)K_x\sim\operatorname{Poisson}(A_x). By memorylessness, conditional on KxK_x and the rates, none of the clocks rings in the next tt years with probability e−tYxe^{-tY_x}. Multiplying by the background and adult survival functions, averaging, and integrating with Tonelli’s theorem gives Equation (S4).

Proposition S2 restates a classical result in mortality terms. The cumulative immaturity hazard A0{1−E(e−tR)}A_0\{1-\mathbb{E}(e^{-tR})\} is a bounded Bernstein function, the Laplace exponent of a compound Poisson process whose jumps have distribution function FF. A Bernstein function is a nonnegative function with a completely monotone derivative, and here that derivative is μ1\mu_1 [17]. The reweighting by e−xRe^{-xR} is selection among clocks: fast clocks ring early, so the clocks still carried at older ages are slower on average. Proposition S2 therefore sets the limit of the selection reading. A completely monotone immaturity hazard can always be read as selection among individuals carrying constant-rate clocks, whereas any other immaturity hazard requires clocks whose hazard changes with age.

In the Siler model, R=b1R=b_1 with probability one, so every clock has rate b1b_1 and Yx=b1KxY_x=b_1K_x. Proposition S2 then reduces to Proposition 1, with the senescence hazard μ3\mu_3 replaced by a general adult hazard hh.

The stretched-exponential immaturity hazard μ1(t)=a1e−b1tp\mu_1(t)=a_1e^{-b_1t^p}, with a1,b1,p>0a_1,b_1,p>0, shows where the condition binds. Its proportional rate of decline, −dlog⁡μ1(t)/dt=b1ptp−1-\mathrm{d}\log\mu_1(t)/\mathrm{d}t=b_1pt^{p-1}, decreases with age when 0<p<10<p<1, is constant when p=1p=1, and increases with age when p>1p>1. The quantity (b1)−1/p(b_1)^{-1/p} is the age at which the hazard has fallen from a1a_1 to a1/ea_1/e, and the remaining cumulative hazard is

Ax=a1p(b1)1/pΓ(1p,b1xp).A_x=\frac{a_1}{p(b_1)^{1/p}}\Gamma\left(\frac{1}{p},b_1x^p\right).

For 0<p<10<p<1, e−b1tp=E(e−tZp)e^{-b_1t^p}=\mathbb{E}(e^{-tZ_p}) for a positive stable random variable ZpZ_p [16], so μ1\mu_1 is completely monotone, and RR has the distribution of ZpZ_p reweighted by 1/Zp1/Z_p. The clock rates then have a continuous distribution, and the mixing distribution of YxY_x retains an atom at zero. For p>1p>1, the second derivative of e−b1tpe^{-b_1t^p} is negative for all sufficiently small positive tt, so μ1\mu_1 is not completely monotone and constant-rate clocks are unavailable. Proposition S1 still applies, with clocks whose hazard ρx\rho_x varies with age.

Exponential immaturity and general adult hazards

With exponential immaturity mortality, every clock has the same constant rate, and the truncation bound of § 4 carries over to any adult hazard.

Corollary S1 (General adult hazard). If μ1(t)=a1e−b1t\mu_1(t)=a_1e^{-b_1t}, then

ex=E[ex,h(a2+b1Kx)],Kx∼Poisson⁡(Ax).(14)e_x=\mathbb{E}[e_{x,h}(a_2+b_1K_x)],\qquad K_x\sim\operatorname{Poisson}(A_x). \tag*{(14)}

The components Qn,h(x)=ex,h(a2+nb1)Q_{n,h}(x)=e_{x,h}(a_2+nb_1) are nonincreasing in nn and satisfy Qn,h(x)≤1/(a2+nb1)Q_{n,h}(x)\le1/(a_2+nb_1) whenever the denominator is positive, so the truncation bound in Equation (7) holds with Qn(x)Q_n(x) replaced by Qn,h(x)Q_{n,h}(x).

Proof. Equation (S5) is Proposition S2 with R=b1R=b_1 with probability one. The integrand in Equation (10) decreases in cc and is bounded by e−cte^{-ct}, which gives the monotonicity and the bound. The truncation argument is then that of Proposition 2.

For the gamma–Gompertz case, set

h(t)=αeβt1+σ2(α/β)(eβt−1),α,β,σ2>0.h(t)=\frac{\alpha e^{\beta t}}{1+\sigma^2(\alpha/\beta)(e^{\beta t}-1)},\qquad\alpha,\beta,\sigma^2>0.

and write exGGM(c)=ex,h(c)e_x^{\mathrm{GGM}}(c)=e_{x,h}(c) for gamma–Gompertz–Makeham remaining life expectancy with Makeham term cc. Define

q=σ2αβ,qx=qeβx1+q(eβx−1),sc=cβ+1σ2.q=\frac{\sigma^2\alpha}{\beta},\qquad q_x=\frac{q e^{\beta x}}{1+q(e^{\beta x}-1)},\qquad s_c=\frac{c}{\beta}+\frac{1}{\sigma^2}.

The conditional gamma–Gompertz survival function is

[1+qx(eβt−1)]−1/σ2.\left[1+q_x(e^{\beta t}-1)\right]^{-1/\sigma^2}.

Substituting y=e−βty=e^{-\beta t} into Equation (10) gives

exGGM(c)=1β∫01ysc−1{qx+(1−qx)y}−1/σ2 dy=(qx)−1/σ2βsc 2F1(1σ2,sc;sc+1;1−1qx).\begin{aligned} e_x^{\mathrm{GGM}}(c)&=\frac{1}{\beta}\int_0^1 y^{s_c-1}\{q_x+(1-q_x)y\}^{-1/\sigma^2}\,dy\\ &=\frac{(q_x)^{-1/\sigma^2}}{\beta s_c}\,{}_2F_1\left(\frac{1}{\sigma^2},s_c;s_c+1;1-\frac{1}{q_x}\right). \end{aligned}

This expression agrees with the closed form derived by Castellares et al. [3].

Corollary S2 (Gamma–Gompertz adult mortality). For the gamma–Gompertz adult hazard above,

ex=E[exGGM(a2+b1Kx)],Kx∼Poisson⁡(Ax).e_x=\mathbb{E}\left[e_x^{\mathrm{GGM}}(a_2+b_1K_x)\right],\qquad K_x\sim\operatorname{Poisson}(A_x).

The truncation bound in Equation (7) also holds with Qn(x)Q_n(x) replaced by exGGM(a2+nb1)e_x^{\mathrm{GGM}}(a_2+nb_1).

Proof. Substitute ex,h(c)=exGGM(c)e_{x,h}(c)=e_x^{\mathrm{GGM}}(c) into Corollary S1.

Numerical verification

Table S1 compares the series with adaptive quadrature in six benchmark years at ages 0, 30, and 60. The bound in the last column covers truncation only, not the error of the numerical routines. Figure S1 gives, at each age, the largest truncation error over all fitted years from 1900 to 2020.

References

  1. [1]Beltrán-Sánchez, H., Preston, S. H., and Canudas-Romo, V. (2008). An integrated approach to cause-of-death analysis: cause-deleted life tables and decompositions of life expectancy. Demographic research, 19:1323.
  2. [2]Brillinger, D. R. (1986). The natural variability of vital rates and associated statistics. Biometrics, 42 4:693–734.
  3. [3]Castellares, F., Patricio, S., and Lemonte, A. (2024). On gamma-gompertz-makeham assurances and life annuities. Austrian Journal of Statistics, 53(4):1–24.DOI
  4. [4]Castellares, F., Patrício, S., and Lemonte, A. J. (2022). On the gompertz–makeham law. Brazilian Journal of Probability and Statistics, 36(3):613–639.DOI
  5. [5]Castellares, F., Patrício, S. C., Lemonte, A. J., and Queiroz, B. L. (2020). On closed-form expressions to gompertz–makeham life expectancy. Theoretical Population Biology, 134:53–60.
  6. [6]Feller, W. (1971). An introduction to probability theory and its applications, volume II. Wiley, New York, 2 edition.
  7. [7]Goulet, V. (2022). expint: Exponential integral and incomplete gamma function.
  8. [8]Human Mortality Database (2026). Human Mortality Database. Max Planck Institute for Demographic Research (Germany), University of California, Berkeley (USA), and French Institute for Demographic Studies (France). Available at https://www.mortality.org/. Data downloaded on 3 November 2026.
  9. [9]Keyfitz, N. (1977). What difference would it make if cancer were eradicated? an examination of the taueber paradox. Demography, 14(4):411–418.pubmed.ncbi.nlm.nih.gov/913728
  10. [10]Lemaître, J.-F., Ronget, V., Tidière, M., Allainé, D., Berger, V., Cohas, A., Colchero, F., Conde, D. A., Garratt, M., Liker, A., et al. (2020). Sex differences in adult lifespan and aging rates of mortality across wild mammals. Proceedings of the National Academy of Sciences, 117(15):8546–8553.DOI
  11. [11]Levitis, D. A. (2011). Before senescence: the evolutionary demography of ontogenesis. Proceedings of the Royal Society B: Biological Sciences, 278(1707):801–809.DOI
  12. [12]Makeham, W. M. (1860). On the law of mortality and the construction of annuity tables. The Assurance Magazine, and Journal of the Institute of Actuaries, 8(6):301–310.DOI
  13. [13]Milgram, M. (1985). The generalized integro-exponential function. Mathematics of computation, 44(170):443–458.DOI
  14. [14]Olver, F. W. (2010). NIST handbook of mathematical functions hardback and CD-ROM. Cambridge university press.
  15. [15]Patricio, S. C. and Missov, T. I. (2024). Makeham mortality models as mixtures. Demographic Research, 51:595–624.
  16. [16]Pollard, H. (1946). The representation of exp(−xᵃ) as a laplace integral. Bulletin of the American Mathematical Society, 52(10):908–910.
  17. [17]Schilling, R. L., Song, R., and Vondraček, Z. (2012). Bernstein Functions: Theory and Applications, volume 37 of De Gruyter Studies in Mathematics. De Gruyter, Berlin, second edition.
  18. [18]Siler, W. (1979). A competing-risk model for animal mortality. Ecology, 60(4):750–757.DOI
  19. [19]Siler, W. (1983). Parameters of mortality in human populations with widely varying life spans. Statistics in medicine, 2(3):373–380.pubmed.ncbi.nlm.nih.gov/6648150
  20. [20]Vaupel, J. W., Manton, K. G., and Stallard, E. (1979). The impact of heterogeneity in individual frailty on the dynamics of mortality. Demography, 16(3):439–454.pubmed.ncbi.nlm.nih.gov/510638
  21. [21]Vaupel, J. W. and Yashin, A. I. (1985). Heterogeneity’s ruses: some surprising effects of selection on population dynamics. The American Statistician, 39(3):176–185.pubmed.ncbi.nlm.nih.gov/12267300

Paper details

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

views · 0

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