Why an Exponential Process Bounds Ruin
In the compound Poisson surplus model, ruin occurs when accumulated claims overwhelm initial capital and premium income. The adjustment coefficient measures how rapidly an exponential function of surplus can penalize downward movements. When that coefficient exists, an exponential martingale converts the event of ruin into a probability bound.
The result is an upper bound, not generally an exact ruin probability. Its proof depends on the model assumptions and on stopping the exponential process at ruin. In particular, the bound requires a positive adjustment coefficient; the net profit condition alone does not guarantee that one exists.
The strict inequality in this definition means that surplus equal to zero is not yet ruin. A claim that takes surplus below zero causes ruin. The same bound also holds if ruin is instead defined by \(U(t)\le 0\), with the corresponding threshold inequality at ruin.
The Adjustment Coefficient and Exponential Process
Let \(M_X(r)=E[e^{rX}]\) be the moment-generating function of claim severity, for values of \(r\) where it is finite. The adjustment coefficient \(R\) is a positive solution of $$ \lambda\bigl(M_X(R)-1\bigr)=cR. $$ This tutorial assumes such a positive root exists and \(M_X(R)<\infty\). The net profit condition \(c>\lambda E[X]\) describes positive expected surplus drift, but does not by itself ensure a positive root: the claim moment-generating function may fail to be finite at positive arguments or may not cross the line required by the equation.
For any \(r\ge 0\) with \(M_X(r)<\infty\), define $$ Z_r(t)=\exp\bigl(-r(U(t)-u)\bigr)=\exp\bigl(-rct+rS(t)\bigr). $$ The process starts at \(Z_r(0)=1\). It discounts the exponential impact of accumulated claims by the premium income earned over time.
Why the Exponential Process Is a Martingale
The claim total over a future time interval is independent of the claims and surplus observed so far. For an interval of length \(h\), the compound Poisson moment-generating function is $$ E[e^{r(S(t+h)-S(t))}]=\exp\left(h\lambda(M_X(r)-1)\right). $$ This follows by conditioning on the number of claims: given \(N(t+h)-N(t)=n\), the moment-generating function of the sum is \(M_X(r)^n\), and averaging over a Poisson count gives the displayed expression.
Using this independent-increment result to calculate the conditional expectation of \(Z_r(t+h)\) gives $$ E[Z_r(t+h)\mid \mathcal{F}_t] =Z_r(t)\exp\left(h\left[\lambda(M_X(r)-1)-cr\right]\right), $$ where \(\mathcal{F}_t\) represents the model information observed through time \(t\). At \(r=R\), the bracket is zero by the adjustment-coefficient equation. Therefore \(E[Z_R(t+h)\mid\mathcal{F}_t]=Z_R(t)\), so \(Z_R(t)\) is a martingale.
More generally, if \(r>0\), \(M_X(r)<\infty\), and \(\lambda(M_X(r)-1)-cr\le 0\), then the same conditional-expectation calculation shows that \(Z_r(t)\) is a nonnegative supermartingale. The root gives the martingale used for the stated bound; a smaller \(r\) with a negative bracket gives a supermartingale and can also produce a bound.
Worked Example: Identifying the Adjustment-Root Martingale
A hypothetical insurer has Poisson claim frequency \(\lambda=2\) claims per year. Claim amounts are exponential with mean 1 thousand dollars, and the premium rate is \(c=3\) thousand dollars per year. Verify the adjustment coefficient and identify the exponential process.
Method. For an exponential claim amount with mean 1 thousand dollars, using claim amounts measured in thousands, \(M_X(r)=1/(1-r)\) for \(r<1\). The adjustment equation is $$ 2\left(\frac{1}{1-R}-1\right)=3R. $$ Since \(\frac{1}{1-R}-1=\frac{R}{1-R}\), the equation becomes \(2R/(1-R)=3R\). A positive root satisfies \(2/(1-R)=3\), so \(1-R=2/3\) and \(R=1/3\) per thousand dollars.
The net profit condition holds: \(\lambda E[X]=2(1)=2\) thousand dollars per year, and \(c=3>2\). The process $$ Z_R(t)=\exp\left(-\frac{1}{3}\left(U(t)-u\right)\right) $$ is a martingale. The units are consistent: \(R\) is per thousand dollars, while \(U(t)-u\) is in thousands of dollars.
Check. At \(R=1/3\), \(M_X(R)=1/(2/3)=3/2\), so \(\lambda(M_X(R)-1)=2(1/2)=1\) per year and \(cR=3(1/3)=1\) per year. The two sides agree.
Stopping at Ruin: The Proof of the Bound
The essential point is that the martingale becomes large if ruin occurs. At the ruin time, $$ Z_R(\tau_u)=\exp\bigl(-R(U(\tau_u)-u)\bigr)>e^{Ru}, $$ because \(U(\tau_u)<0\). Thus, ruin requires the martingale to reach a level greater than \(e^{Ru}\).
For a finite time \(T\), stop the martingale at \(\tau_u\wedge T\), the earlier of ruin and time \(T\). The optional stopping theorem for a martingale stopped at a bounded time gives \(E[Z_R(\tau_u\wedge T)]=Z_R(0)=1\). On the event \(\{\tau_u\le T\}\), the stopped value is greater than \(e^{Ru}\), and it is nonnegative on every other event. Consequently, $$ 1=E[Z_R(\tau_u\wedge T)] \ge E\left[Z_R(\tau_u\wedge T)\mathbf{1}_{\{\tau_u\le T\}}\right] \ge e^{Ru}P(\tau_u\le T). $$ Therefore \(P(\tau_u\le T)\le e^{-Ru}\). As \(T\) increases, the events \(\{\tau_u\le T\}\) increase to \(\{\tau_u<\infty\}\). Taking their probability limit proves \(\psi(u)\le e^{-Ru}\).
Use \(Z_R(t)=e^{-R(U(t)-u)}\), where \(R\) solves the adjustment equation.
Independent compound Poisson increments give a conditional expectation multiplier of \(\exp(t[\lambda(M_X(R)-1)-cR])=1\).
At ruin, \(Z_R(\tau_u)>e^{Ru}\); the stopped martingale still has expectation 1.
Bound ruin by each finite time \(T\), then let \(T\) increase to infinity to obtain \(\psi(u)\le e^{-Ru}\).
Worked Example: A Discrete Severity With a Known Root
A hypothetical portfolio has Poisson claim frequency \(\lambda=1\) claim per year, and every claim is exactly 1 thousand dollars. The premium rate is \(c=1/\ln 2\) thousand dollars per year. Verify that the adjustment coefficient is \(R=\ln 2\) per thousand dollars, and obtain the bound when initial capital is 2 thousand dollars.
Method. For a deterministic claim of 1 thousand dollars, \(M_X(r)=e^r\) when claim amounts are measured in thousands. At \(R=\ln 2\), \(M_X(R)=2\), so $$ \lambda(M_X(R)-1)=1(2-1)=1, \qquad cR=\frac{1}{\ln 2}\ln 2=1. $$ Thus the adjustment equation holds. Also, \(E[X]=1\) thousand dollars, and \(c=1/\ln 2\approx1.4427>1=\lambda E[X]\), so the net profit condition holds.
For \(u=2\) thousand dollars, Lundberg’s inequality gives $$ \psi(2)\le e^{-2\ln 2}=\frac{1}{4}=0.2500. $$ The bound says ultimate ruin probability is at most 0.2500; it does not assert that the probability equals 0.2500.
Check. The exponential process starts at 1, and its conditional expectation multiplier is \(\exp(h[1(2-1)-(1/\ln2)(\ln2)])=\exp(0)=1\), as required for a martingale.
Using a Supermartingale and Interpreting the Result
The proof also explains a useful extension. Suppose \(r>0\) is such that \(M_X(r)<\infty\) and \(\lambda(M_X(r)-1)-cr\le0\). Then \(Z_r(t)\) is a nonnegative supermartingale. Stopping at \(\tau_u\wedge T\) gives \(E[Z_r(\tau_u\wedge T)]\le1\). The same ruin-event comparison yields \(P(\tau_u\le T)\le e^{-ru}\), and hence \(\psi(u)\le e^{-ru}\). If the adjustment root is approached by such values \(r\), taking the limit recovers the bound with \(R\).
Worked Example: A Smaller Exponent Gives a Weaker Bound
A hypothetical insurer has \(\lambda=1\) claim per year, exponential claim amounts with mean 1 thousand dollars, and \(c=2\) thousand dollars per year. Compare the adjustment-root bound and a supermartingale bound using \(r=0.25\) per thousand dollars when \(u=2\) thousand dollars.
Find the adjustment coefficient. Here \(M_X(r)=1/(1-r)\). The equation \(1/(1-R)-1=2R\) reduces, for a positive root, to \(1/(1-R)=2\), giving \(R=0.5\) per thousand dollars. The net profit condition holds because \(c=2>1=\lambda E[X]\).
Check the smaller value. At \(r=0.25\), \(M_X(r)=1/0.75=4/3\). The exponent coefficient is $$ \lambda(M_X(r)-1)-cr =1\left(\frac{4}{3}-1\right)-2(0.25) =\frac{1}{3}-\frac{1}{2} =-\frac{1}{6}\le0. $$ Thus \(Z_r\) is a supermartingale, and it gives \(\psi(2)\le e^{-0.25(2)}=e^{-0.5}\approx0.6065\). Using the adjustment root gives the stronger bound \(\psi(2)\le e^{-0.5(2)}=e^{-1}\approx0.3679\).
Check. Both values are upper bounds, and the smaller one is tighter because \(R=0.5\) is larger than \(r=0.25\), so \(e^{-Ru}<e^{-ru}\) for \(u>0\).
Common Exam Trap
A frequent error is to treat the adjustment equation as if it were \(\lambda(M_X(R)-1)=c\), omitting the factor \(R\). That equation is incorrect: the premium contribution to the exponential process is \(e^{-Rct}\), so its exponent contributes \(-cR\), not \(-c\). The correct balance is \(\lambda(M_X(R)-1)-cR=0\).
A related error is to use \(\exp(-Ru)\) with \(u\) stated in dollars when \(R\) was calculated per thousand dollars. Convert the capital to the same units as the claim amounts before multiplying. The bound depends on the dimensionless product \(Ru\), and inconsistent units change the answer.
Finally, a positive drift does not automatically justify Lundberg’s inequality. Check that the model has a positive adjustment coefficient and that the claim moment-generating function is finite at the value used. The result is an upper bound derived from the model, not an exact ruin probability. This proof is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
Check Your Understanding
For each problem, identify the relevant adjustment coefficient or admissible exponent before applying the bound.
- A hypothetical insurer has \(\lambda=0.5\) claims per year, exponential claim amounts with mean 2 thousand dollars, and \(c=1.5\) thousand dollars per year. Find \(R\) and the Lundberg bound at \(u=6\) thousand dollars. (Answer: \(R=1/6\) per thousand dollars; \(\psi(6)\le e^{-1}\approx0.3679\).)
- Claims are deterministically 1 thousand dollars, \(\lambda=2\) per year, and \(c=2/\ln 2\) thousand dollars per year. Verify a candidate root \(R=\ln 2\) and find the bound at \(u=1.5\) thousand dollars. (Answer: the adjustment equation holds; \(\psi(1.5)\le2^{-1.5}\approx0.3536\).)
- Claim amounts are exponential with mean 1 thousand dollars, \(\lambda=3\) per year, and \(c=4.5\) thousand dollars per year. Find the adjustment coefficient and the bound at \(u=3\) thousand dollars. (Answer: \(R=1/3\) per thousand dollars; \(\psi(3)\le e^{-1}\approx0.3679\).)
- For a model with adjustment coefficient \(R=0.2\) per thousand dollars, a candidate uses \(u=5{,}000\) dollars. What exponent must be used after converting units? (Answer: \(Ru=0.2(5)=1\), so the bound is \(e^{-1}\approx0.3679\).)
- At \(r=0.3\) per thousand dollars, suppose \(M_X(r)<\infty\) and \(\lambda(M_X(r)-1)-cr=-0.04\) per year. What process property follows, and what bound results for \(u=4\) thousand dollars? (Answer: \(Z_r\) is a nonnegative supermartingale; \(\psi(4)\le e^{-1.2}\approx0.3010\).)