Sei sulla pagina 1di 56

A C T U A R I A L RESEARCH C L E A R I N G HOUS E

1 9 9 7 VOL . 1
On the Ti me Val ue of Rui n
Hans U. Ger ber
Ecol e des hautes 6tudes commer ci al es
Universit6 de Lausanne
CH- 1015 Lausanne, Swi t zer l and
Phone: 41 21 692 3371
Fax: 41 21 692 3305
E-mai l : hgerber@hec. uni l . ch
El i as S. W. Shiu
Depart ment of St at i st i cs and Act uari al Sci ence
The Uni versi t y of I owa
I owa Ci t y, I owa 52242, U. S. A.
Phone: 319 335 2580
Fax: 319 335 3017
E-mai l : eshi u @ st at . ui owa. edu
Abstract
Thi s paper studies the j oi nt di st ri but i on of the t i me of ruin, the surpl us i mmedi at el y bef or e
ruin, and the defi ci t at ruin. The cl assi cal model is general i zed by di scount i ng with
respect to the t i me of ruin. We show how to cal cul at e an expect ed di scount ed penal t y,
whi ch is due at ruin, and may depend on the defi ci t at ruin and the surpl us i mmedi at el y
before ruin. The expect ed di scount ed penalty, consi dered as a function of the initial
surplus, sat i sfi es a certain renewal equation, whi ch has a pr obabi l i st i c i nt erpret at i on.
Expl i ci t answers are obt ai ned for zero initial surplus, very large initial surplus, and
arbi t rary initial surplus if the cl ai m amount di st ri but i on is exponent i al or a mi xt ure of
exponent i al s. We general i ze D. C. M. Di cks on' s formula, whi ch expr esses the j oi nt
di st ri but i on of the surplus i mmedi at el y pri or to and at ruin in t erms of the pr obabi l i t y of
ul t i mat e ruin. Expl i ci t resul t s are obt ai ned when di vi dends are pai d out to the
st ockhol der s accordi ng to a const ant barri er strategy.
145
1. Int roduct i on
CoUective risk theory' has started in 1903 with the doctoral thesis of Fi l i p Lundber g,
a Swedi sh act uary, and it has been devel oped t hroughout this cent ury. It is now an area
rich in useful ideas and sophi st i cat ed t echni ques. Many of its tools can be appl i ed to sol ve
pr obl ems in ot her areas. A recent exampl e is the met hod of Esscher tran~fforms, whi ch was
used by Ger ber and Shi u [23] to price f i nanci al deri vat i ves.
Two par t i cul ar quest i ons of rui n theory are (a) the severi~' qfruin, and (b) the
time of ruin, bot h of whi ch have been treated separat el y in the literature. In this paper
cert ai n answer s to bot h quest i ons are gi ven at the same time. We st udy the j oi nt
di st r i but i on of the defi ci t and the t i me of rui n. Fr om a mat hemat i cal poi nt of vi ew a
cruci al role is pl ayed by the surpl us i mmedi at el y before rui n occurs. We i ncorporat e the
t i me of r ui n in the model by di scount i ng. We show how to cal cul at e an expected
discounted penalty, whi ch is due at rui n and may depend on the defi ci t at rui n and the
sur pl us i mmedi at el y pri or to ruin. The expect ed di scount ed penal t y, consi der ed as a
f unct i on of the i ni t i al surpl us, satisfies a cert ai n renewal equat i on. The renewal equat i on
and its convol ut i on series sol ut i on have nat ural probabi l i st i c i nt erpret at i ons. For the
f or mer one consi der s the first t i me when the surpl us falls bel ow the i ni t i al level, and
di st i ngui shes accor di ng to whet her or not rui n takes pl ace at that time. For the latter the
di st i nct i on is made accor di ng to whi ch of the record lows of the surpl us process causes
rui n. Expl i ci t answer s are obt ai ned for zero i ni t i al surpl us, very large i ni t i al surpl us, and
arbi t rary i ni t i al surpl us if the cl ai m amount di st r i but i on is exponent i al or a mi xt ure of
exponent i al s. Addi t i onal i nsi ght is obt ai ned from a pai r of exponent i al mart i ngal es. As
an appl i cat i on we gener al i ze D. C. M. Di cks on' s [7] formul a, whi ch expresses the j oi nt
di st r i but i on of the surpl us i mmedi at el y pri or to and at rui n in t erms of the pr obabi l i t y of
ul t i mat e rui n. Si mi l ar l y, expl i ci t resul t s are obt ai ned when di vi dends are pai d out to the
st ockhol der s accor di ng to a const ant barri er strategy.
146
The paper generalizes and adds to a better understanding of classical ruin theory,
which can be retrieved by setting the interest rate equal to zero. For example, in the
classical model, the adjustment coefficient is the solution of an implicit equation, which has
0 as another solution. If the interest rate is positive, the situation is suddenly symmetric:
the corresponding equation, called Lundberg'sfimdamental equation, has a positive
solution and a negative solution. Both solutions are important and are used to construct
exponential martingales. We also present some results concerning the finite-time ruin
function and its Laplace transforms.
This paper was motivated by the problem of pricing American options. The
classical model uses the geometric Brownian motion to model the stock price process.
Such a process has continuous sample paths, which facilitate the analysis of an American
option: the option is exercised as soon as the stock price arrives on the optimal exercise
boundary, and the price of the option is the expected discounted payoff. On the other hand,
we would like to price an American option in a perhaps more realistic model where the
stock price may have jumps. The resulting mathematical problem is more intricate,
because now, at the time of the exercise, the stock price is not on but beyond the optimal
exercise boundary. If the logarithm of the stock price is modeled by a shifted compound
Poisson process, this leads to the type of problems that are discussed in this paper.
Evidently "penalty at rain" has to be replaced by "payoff at exercise." Thus the paper lays
the mathematical bases for a financial application. Since it is substantial and of an
independent interest, we decided to present the application to the pricing of American
options in a subsequent paper.
2. Whe n and How Does Rui n Occur
We follow the notation in Chapter 12 of Actuarial Mathematics [4]. Thus u _> 0 is
the i nsurer' s initial surplus. The premiums are received continuously at a constant rate c
per unit time. The aggregate claims constitute a compound Poisson process, {S(t)}, given
147
by t he Poi s s on par amet er ~. and i ndi vi dual cl ai m a mount di s t r i but i on f unct i on P(x) wi t h
P(O) = O. That is,
N(t)
s( t ) = Y _ , x j , ( 2. 1)
j = l
wher e {N(t)} is a Poi sson pr ocess wi t h mean per uni t t i me X and {Xj} are i nde pe nde nt
r a ndom var i abl es wi t h c o mmo n di s t r i but i on P(x). Then
U(t ) -- u + ct - S(t) (2. 2)
is t he s ur pl us at t i me t, t _> O. For si mpl i ci t y we as s ume that P(x) is di f f er ent i abl e, wi t h
P' (x) = p(x)
bei ng t he i ndi vi dual cl ai m a mount pr obabi l i t y densi t y f unct i on.
Let T denot e t he time of ruin,
T = i nf {t [ U( t ) < 0} (2. 3)
( T = o~ i f r ui n does not occur). We cons i der the pr obabi l i t y of ul t i mat e r ui n as a f unct i on
of t he i ni t i al sur pl us U(0) = u _> 0,
~g(u) = Pr [ T < oo I U( 0) = u]. (2. 4)
Let p! denot e t he me a n of the i ndi vi dual cl ai m a mount di st r i but i on,
Pl = IO x p(x) dx = E(Xj ).
We as s ume
c > ~-Pt (2. 5)
to ens ur e t hat {U(t)} has a posi t i ve drift; hence
l i m U(t ) = oo (2. 6)
t - - ) ~
wi t h cer t ai nt y, and
~ff(u) < 1. (2. 7)
We al so cons i der t he r a ndom var i abl es U( T- ) , the sur pl us i mmedi at el y bef or e r ui n,
and U( T) , t he s ur pl us at rui n. See Fi gur e 1. For gi ven U(0) = u _> 0, let f(x, y, t I u) denot e
t he j oi nt pr obabi l i t y densi t y f unct i on of U( T- ) , IU(T)I and T. The n
148
J oloJ of( x , y , t l u) dx dy dt = P r I T < o~1 u( 0 ) = ul = v ( u> .
Becaus e of (2. 7), f(x, y, t I u) is a defective pr obabi l i t y densi t y f unct i on.
f ( x , y , t l u) = 0
for x > u +c t .
(2. 8)
We not e t hat
~ , ' ( 0
: / :T
, . . . . . ~ (1(z' -)
Fi gur e 1. The Sur pl us I mmedi at el y bef or e and at Rui n
It is easi er to anal yze the f ol l owi ng f unct i on, the st udy of whi ch is a cent ral t heme
in this paper. For 5 _> O, def i ne
f( x , Y l u) = l o e ~ t f ( x , y , t I u) dt. ( 2. 9 )
Her e ~ can be i nt er pr et ed as a force of i nt erest , or, in t he cont ext of Lapl ace t r ansf or ms, as
a d u mmy vari abl e. Not e that the s ymbol f(x, y l u) does not exhi bi t the dependence on 5.
If 8 = 0, (2.9) is t he def ect i ve j oi nt pr obabi l i t y densi t y f unct i on of U( T- ) and IU(T)I, gi ven
U(0) = u. Al so, if t3 > 0, t hen
e ~Sw = e --'ST I(T < o~),
where I denot es t he i ndi cat or f unct i on, i.e., I(A) = I i f A is true and I(A) = 0 if A is false.
149
Let w(x, y) be a nonnegative f unct i on of x > 0 and y > 0. We consi der , for tl _> 0,
~(u) = E[ w( U( T- ) , IU(T)I) e -ST I(T < oo)[ U(0) = u] (2. 10)
= G T j 2 w ( x , t t u ) d t d x d ,
= J , Tl w ( x , y ) f ( x , Y l u ) d x d y . , 2 . , , )
Wi t h 8 = 0 and w(x, y) = w( - y) , 0( u) is denot ed as ~( u ; w) in the Pr oof of The or e m 12.4
of Actuarial Mathematics [4]. I f we i nt erpret ~ as a force of i nt er est and w as s ome ki nd
of penal t y when r ui n occurs, t hen O(u) is the expect at i on of the di s count ed penal t y. We
s houl d cl ar i f y that, whi l e it can be very hel pful to cons i der 8 as a force of i nt er est in t hi s
paper, we are deal i ng wi t h the cl assi cal model in whi ch the sur pl us does not ear n any
i nt erest .
Our i mmedi at e goal . i s to der i ve a f unct i onal equat i on for O(u) by appl yi ng t he l aw
of i t erat ed expect at i ons to t he r i ght - hand side of (2.10). For h > O, cons i der the t i me
i nt er val (0, h), and condi t i on on the t i me t and the amount x of the first cl ai m in t hi s t i me
i nt er val . Not e that the pr obabi l i t y that t here is no cl ai m up to t i me h is e -xh, t he
pr obabi l i t y that the first cl ai m occur s bet ween t i me t and t i me t + dt is e-Xt)~dt, and
x >u + c t
me a ns t hat r ui n has occur r ed wi t h tile first cl ai m. Hence
th .tu+ct _ x) p( x) dx] e- ( 6 + ~.)t ~vdt
( u) = e -(5 +x) hc ~( u+c h) + ]o[]o ( u + c t
+ l,",llu+c w(u + ct, x - u - c t ) p ( x ) d x ] e (5+ ~)' )~dt. (2. 12)
Di f f er ent i at i ng (2. 12) wi t h respect to h and set t i ng h = 0, we obt ai n
= - ( 8 + k ) * ( u) + c ' ( u) + ~ o O( u - x ) p ( x )
0 dx
+ Kl u w( u, x - u ) p ( x ) d x
. u
= - ( 5 +X) O( u ) + c0' ( u) + ~0 ~ ( u - x ) p ( x ) d x +
)~o)(u), (2. 13)
wher e
j
~
~( u) = w(u, x - u) p(x) dx.
U
Wi t h X denot i ng the i ndi vi dual cl ai m a mount r andom var i abl e,
(2. 14)
150
C0(u) = E[ w( u, X - u) I(X > u)].
Al so, by a change of variable,
o o
Co(u) = IO w(u, y) p(u + y) dy.
For furt her anal ysi s, we use the t echni que of integrating factors.
(~p(U) = e-P u O(u),
(2. 15)
(2. 16)
Let
(2. 17)
wher e P is a nonnegat i ve number to be speci fi ed later. Mul t i pl yi ng (2. 13) wi t h e-P u,
appl yi ng t he pr oduct rule for di fferent i at i on, and rearrangi ng yi el ds
U
cO~'p(u) = (8 + ~ - cp)Op(U) - kloOp(U - x)e-pxp(x)dx - )~e-puol(u).
Def i ne
(2. 18)
7(~) = ~i + ~. - c~; (2. 19)
hence t he coef f i ci ent of ~p(u) in (2. 18) is [~(p). In this paper we let f denot e the Lapl ace
t r ansf or m of a funct i on f,
f(~) = I oe- ~ x fi x) dx. (2. 20)
The Lapl ace t r ansf or m of p, ~(~), is defi ned for all nonnegat i ve number s ~,, and is a
decr easi ng convex funct i on because
= -J oe-~X xp( x) dx < 0
and
~"(~) = ] oe- ~ x x 2 p(x) dx > 0.
Consi der the equat i on
Si nce
~(~) = Xt3(~). (2. 21)
0(0) = 8 + ; ~ > _ X= X0(0)
and by (2. 5)
~ ' ( 0 ) = - c < - ~ , P l = ~ 0 ' ( 0 ) ,
equat i on (2. 21) has a uni que nonnegat i ve root, say ~1. It is obvi ous f r om Fi gure 2 ( whi ch
cor r esponds t o Fi gure I 1.7.1 of Panj er and Wi l l mot [27]) that ~1 is an i ncreasi ng f unct i on
151
of 5, wi t h ~1 = 0 when 8 = 0. Fur t her mor e, if the i ndi vi dual cl ai m amount dens i t y
f unct i on p is s uf f i ci ent l y r egul ar , equat i on (2.21 ) has one mor e root , say ~2, whi ch is
negat i ve. Thi s negat i ve root , whi ch wi l l be denot ed as - R, pl ays an i mpor t ant r ol e l at er.
As we shal l see in Sect i on 8, bot h r oot s are r el at ed to the const r uct i on of exponent i al
mar t i ngal es . When ~5 = 0, (2.21 ) is equi val ent to (12. 3. 1) in Actuarial Mathematics [4] and
R is t he adjustment coefficient. Equat i on (2. 21) is equi val ent to Be e kma n [3, p. 41, t op
equat i on] , Panj er and Wi l l mot [27, (11. 7. 8)], and Seal [34, (4. 24)]. Lundber g [26, p. 144]
poi nt s out t hat the equat i on is "f undament al to the whol e of col l ect i ve ri sk t heor y, " and
Seal [34, p. 111] cal l s it "Lu n d b e r g ' s (1928) ' f undament aL' equat i on. " (It is i ncor r ect for
Seal [34, p. 112] t o asser t t hat the s econd r oot is al so posi t i ve. )
L
Fi gur e 2. The Two Root s of Lundbe r g' s Fundament al Equat i on
The t r i ck f or s ol vi ng 12.18) is to choose
p = Y-,l,
so t hat ( 2. 18) be c ome s
(2. 22)
152
U
c(~'p(u) = X# ( p) %( u) - ~'fo*p(U - x)e-pXp(x)dx - ~.e-PU00(u)
U
= ) ~[ 0( p) %( u) - t0 * P( x) e- p( u- x) p( u- x) dx - e-pUoXu)]. (2. 23)
For z > 0, we i nt egrat e (2. 23) f r om u = 0 to u = z. Af t er a di vi si on by )~, the resul t i ng
equat i on is
~ [ % ( z ) - %(0)1
^ Z Z II Z
= p( p) / 0%( u) du - 1 0 [ I 0 dOp(x)e-p(u-x)p(u - x) dx] du - J 0e- pu~( u) du
Z Z Z Z
: P( P) f 0%( u) du - f0[Ixe-p(u-x)p(u - x) du] , p( x) dx - f0e-puc (u)du
Z
= 10%(x)[lz _xe-PYp(y)dyldx-J 0e-pUo3(u)du. (2. 24)
For z ---) co, the first t erms on bot h sides of (2. 24) vani sh, whi ch shows that
%( 0) = ~ I0 e-put-(u)du = ~ ~( P) ' (2. 25)
Fi nal l y, subst i t ut i ng (2. 25) in (2. 24) and si mpl i f yi ng yi el ds
L z ~ lz e - pu~( u) du}' z > O. (2. 26)
, p( Z) = E-{lo*p(X)[lz_xe-PYP(y)dy]dx + o~
For t wo i nt egrabl e fun~:tions fl and f2 defi ned on [0, oo), the convolution of fl and
f2 is the f unct i on
(fl *f2)(x) = / of l ( y) f2(x - y) d y , x > o. (2. 27)
Not e t hat
Wi t h t he defi ni t i ons
and
f l * f 2 = f2*fl .
gO(x) = ~I ? e- PYp( y) dy, x _> 0, (2. 28)
hp( x) = a ~ l?e--puo3(u)du, x > O, (2. 29)
equat i on (2. 26) can be wri t t en mor e conci sel y as
~)p : *p*gp + hp. (2. 30)
In t he literature of integral equat i ons, (2. 30) is cl assi fi ed as a Volterra equation of the
second kind. The f unct i on gp is a nonnegat i ve funct i on on [0, o~) and hence may be
153
i nt er pr et ed as a (not neces s ar i l y pr oper ) pr obabi l i t y dens i t y f unct i on; in pr oba bi l i t y t heor y,
(2. 30) is known as a renewal equation f or the f unct i on 09.
Unl i ke 0, t he f unct i on 0p does not have a pr obabi l i s t i c i nt er pr et at i on. Hence it is
pr ef er abl e to wor k wi t h the f unct i on 0. It f ol l ows f r om (2. 27) t hat for each const ant k
ekx( f l * f 2) = (ekxfl )*(ekxf2). (2. 3 I )
Thi s e na bl e s us to conver t (2. 30) i nt o a r enewal equat i on for 0 usi ng (2. 17), whi ch is
~( x) = ePXOp(x), x _> O.
Wi t h t he def i ni t i on
and
we have
g( x) = epX gO(x)
= 12e P( Y- X) p( y) dy (2. 32)
= ~12e~:~Zp(x + z ) dz, x_>O, ( 2. 33,
h( x )
= eP x hp(x)
= ~ f xe- p( u- x) co( u) du ( 2. 34)
X I o e pz ~o(x + z) dz, x > O, ( 2. 35)
0 = 0* g + h. (2. 36)
The s ol ut i on of (2. 36) can be expr es s ed as an i nf i ni t e ser i es of f unct i ons, s ome t i me s cal l ed
a Neumann series,
0 = h + g* h + g* g* h + g* g* g* h + g* g* g* g* h + . . . . (2. 37)
Thi s ge ne r a l i z e s what is cal l ed Beekman' s convolution series in the act uar i al l i t er at ur e.
One ma y obt ai n (2. 37) f r om (2. 36) by t he met hod of s ucces s i ve subst i t ut i on.
Remark s (i) Two usef ul expr es s i ons f or h are
h( u) : ~ l uJ O e- p( x- u) w( x, y ) p ( x + y ) d y dx.
and
( 2. 38)
154
h( u) = ~J oJ oe- PZ w(u + z, y ) p ( u + z + y ) d y dz. ( 2. 39)
(i i ) Wi t h 5 = 0 and hence p = 0, it is wel l known [4, Theor em 12.4] t hat t he di f f er ent i al
g( y) dy = ~[ 1 - P( y) ] dy, ( 2. 40)
can be i nt er pr et ed as t he pr obabi l i t y t hat t he sur pl us wi l l ever fal l bel ow its i ni t i al l evel u,
and wi l l be bet ween u - y and u - y - dy when it happens f or t he fi rst t i me. Fur t her mor e,
wi t h 5 -- 0 and w --- 1, we have
h(x) = l ; g ( y ) d y ,
whi ch is t he pr oba bi l i t y t hat the sur pl us wi l l e ve r fall bel ow its i ni t i al l evel u, and wi l l be
be l ow u - x when it happens for t he fi rst t i me. The r enewal equat i on (2. 36) gener al i zes
Exer ci s e 12.11 of Actuarial Mathematics [4]; see al so (9. 44) bel ow. In Sect i on 5 we shal l
see how t he f unct i ons g and h can be i nt er pr et ed for 3 > O, and hence pr obabi l i s t i c
e xpl a na t i ons of (2. 36) and (2. 37) can be gi ven.
(i i i ) It f ol l ows f r om t he condi t i onal pr obabi l i t y f or mul a,
Pr ( A n B) = Pr ( A) Pr ( B] A),
t hat t he j oi nt pr obabi l i t y dens i t y f unct i on of U( T- ) , IU(T)I and T at the poi nt (x, y, t) is t he
j oi nt pr oba bi l i t y dens i t y f unct i on of U( T- ) and T at t he poi nt (x, t) mul t i pl i ed by t he
condi t i onal pr oba bi l i t y densi t y f unct i on of IU(T)I at y, gi ven t hat U( T- ) = x and T = t. The
l at t er does not depend on t and is
p(x + y) p(x + y)
- 1 - P ( x ) ' Y > 0
f ~ p(x + y ) d y
Hence
f ( x , y , t [ u ) = [ f : f ( x , z , t l u ) dzl P( X + y )
- P ( x ) " ( 2 . 4 1 )
Wi t h t he def i ni t i on
f ( xl u)
: f : f ( x , y l u ) dy
= JoJge -at f(x, y , t [ u) dt dy, ( 2. 42)
155
mul t i pl yi ng (2. 41) with e -St and then i nt egrat i ng with respect to t yi el ds
f(x, y[ u) = f(x[ u) p(___x
+
Y)
P(x) " (2. 43)
Wi t h 6 = 0, (2. 43) was first poi nt ed out by Duf r esne and Ger ber [I 3, (3)]; anot her pr oof
can be f ound in Di ckson and Egf di o dos Rei s [9]. Al so, it fol l ows f r om (2. 10), (2. 41),
(2. 42) and (2. 16) that
~(u) = J ot oJ oW(X, y ) e ~St f(x, y , t I n ) d t dx dy
~, oo ~, . I p ( x + y )
= 1010 I0 w(x, y)e~St [ lof(X, z , t l u ) a z ] ~ dt dx dy
. p( x + y)
= J o/oW(X, y ) f ( x I u ) q ~ f f ~ dx dy
r ~ . . f(x l u)
= ]0 c ~ ' x ) T ~ - ~ dx. (2. 44)
(iv) It f ol l ows f r om an i nt egrat i on by parts that
- ~/oe-~X[I - P(x)] dx, (2. 45)
~(~)
1
wi t h whi ch we can rewri t e Lundbe r g' s f undament al equat i on (2. 21) as
6 = c~ - )~[1 - ~(~)] (2. 46)
~4oe-~X[1 - P(x)] dx}. (2. 47) t i c
Hence
- - c - fot' - P ( x ) l d x
= c - )~Pl, (2. 48)
whi ch is the drift of {U(t)}. Wi t h i5 = 0, the negat i ve root ~ = ~2 of (2. 47) or (2. 21) is
det er mi ned by the equat i on
loe-~X[1 - P(x)] dx = ~ ; (2. 49)
this condi t i on to i dent i cal to the one in Exer ci se 12.7 of Actuarial Mathematics [4].
(v) Equat i on (2. 36) may be sol ved by the met hod of Lapl ace t r ansf or ms (Spi egel [37]).
Taki ng Lapl ace t ransforms, we have
o r
$ ( ~ ) = $ ( ~ ) ~ ( ~ ) + fi( ~ ) , (2. 50)
156
~ ( 0 - f i ( O ( 2 . 5 l )
1 - ~ ( 0
Hence ~ can obt ai ned by inverting or identifying the right-hand side of (2.51 ). In t erms of
compl ex integration,
1_2._ i b+i~ /a(~) e~ u d~, (2.52)
~(u) = 2/ti Jb-io~ 1 --- ~-(~)
where i = dZ7-, and the path of integration is parallel to the i magi nary axis in the compl ex
plane, with the real number b being chosen so that all the singularities of the integrand lie
to the left of the line of integration (Spiegel [37, p. 201]). Furt hermore, if we expand the
right-hand side of (2.51) as a geomet ri c series, we obtain
oo
~(O = Z ~(O ~(O, ( 2.53)
I 1 =
which is the Lapl ace t ransform of (2.37).
(vi) Fr om (2.32) and by changi ng of the order of integration we see that
= l o e - X t l x e - . , y - , p ( y )
dy] dx
- c( o . fole'P -Oy - le- P Y p ( y ) dy
--- ~ [ ~ ( ~ ) - ~(p)]. (2.54)
Because p satisfies Lundber g' s fundamental equation (2.21), it follows that
~ . ~ ( O + c o - 8 - Z,
f ~( O = c ( p - ~,) ' ( 2 . 5 5 )
o r
~ . [ 1 - ~ ( O] + ~ 5 - c ~
1 - ~(~) = c ( p - ~) (2.56)
We note that (2.21) is the condition that the numerat or on the right-hand side of (2.56)
vanishes. Hence the negative root F~2 of (2.21) is det ermi ned by the condition that
f~(~,) = 1. ( 2 . 5 7 )
(vii) Wri t i ng (2.54) as
157
and di f f er ent i at i ng wi t h r espect to ~, we obt ai n
g' (~) = p - - ~ [ g ( ~ ) + ~0' ( ~) ] . (2. 58)
Si nce t he negat i ve root ~,2 sat i sf i es (2. 57), a par t i cul ar case of (2. 58) is
g' (~2) = p _~ 2 [ I + ~0' ( ~2) ] . (2. 59)
(vi i i ) Fr om (2. 34) and by changi ng t he or der of i nt egr at i on, we get
l~(~) = ~ 10 e -~ X [ t ? e p(u-x, c o( u) du] dx
= c ( p ~ ) I O[ O( e 4 U- e - p u ) w( u, y) p(u + y ) d u dy. (2. 60)
(ix) Cons i der t he speci al cas e wi t h t he penal t y f unct i on w( x, y) --- I. Then (2. 60) b e c o me s
X I~l%e_~U
fa(~) = ~ 1 0 J 0 t - e - PU) p ( u + y ) d u d y
- c ( ? _~ l o ( e - ~ U - e-PU) [ 1 - P(u)] du
1
du + ~ - c }, (2. 61)
- c( p ~ - J o
be c a us e p sat i sf i es (2. 47). Appl yi ng (2. 45) yi el ds
- + - c } 262)
Hence, wi t h w( x, y) - 1,
f i ( ~ . )
, ~ ( ~ . ) -
l - ~ . ( ~ )
= kp[ l - 0(~)1 + ~ ( 8 - cp) (2. 63)
~p{a.ll-0(~)] + 6- c~}
by (2. 62) and (2. 56). In der i vi ng (2. 63) it is a s s ume d t hat & and hence 13, are posi t i ve.
The cas e wher e 8 = p = 0, and hence (~ = ~, is best t r eat ed as a l i mi t i ng case: Fr om (2. 63)
and ( 2. 48) we obt ai n
Z[I - Pl ~ - t3(~)] (2. 64)
~ ( ~ ) = ~ { ~ [ I - 0 ( ~ ) 1 - c ~ } '
whi ch can be r econci l ed wi t h (12. 6. 9) in Actuarial Mathematics [4] by the f or mul a
]c~ e ~ u v ' ( , ) du = - V~ 0 ) + ~, ~ , ( ~ ) . (2. 65)
158
3. F i ni t e - Ti me Rui n P r obabi l i t y
Adopting the notation in Actuarial Mathematics [4, ( 12.1.4)], we let
~(u, t) = Pr[T _< t ] U(0) = u] (3.1)
be the probability of ruin by time t, t _> 0. Then
~(u, t) = I 0[ f 0] 0 f(x, y, s] u) dxdy] ds , (3.2)
or
Hence
~ ( u , t) = Io]of(X, y, t] u)dxdy.
(3.3)
E[e~ST
I ( T < o o 3 [ U ( 0 ) = u ] = JO ~ t v ( u ' t ) e-& dt
= 8f0 ~(u, t) dt (3.4)
e ~t
by an integration by parts. Formula (3.4) shows how the Laplace transform of the
defective distribution of the time of ruin can be expressed in terms of the single Lapl ace
transform of the finite-time ruin function. Because the left-hand side of (3.4) is ,(u) with
w(x, y) --- I, the double Laplace transform of the finite-time ruin function is
,( ~, 55 = I o f o e~U - ~' V ~u , t5 dt du
= -~Ioe-~ u (u)
du
= Lp[ 1- O( ~) ] + ~ ( ~ - c p ) ~ >0 , 5 >0 , (3.5t
by (2.63).
Historically, some actuarial researchers have preferred to study the survival
( 3 . 6 )
function
cy(u,t) = 1 - ~/(u,t).
Its double Laplace transform is
~(~, 85 : 1 2 j 2 e + - ~ t ~ u , t) dt du
1
= ~ - q , ( ~ , 8 5
159
p - ~
~ p{ ) . l l -~ ( ~ ) ] + 8 - c ~ , }
1 1
~ - ~
= ~[1 - ~(~)] + 8 - c~, ' ~ > 0, 8 > 0, ( 3. 7)
whi ch is t he s ame as Seal [34, (4. 26)] and Panj er and Wi t l mo t [27, The or e m 11.7.4].
Thr ough an i nt egr at i on by par t s, (3. 7) is equi val ent to Be e kma n [3, p. 40, Cor ol l a r y 1].
Not e t hat t he de nomi na t or on t he r i ght - hand si de of (3. 7) is t he di f f er ence of t he t wo si des
in Lu n d b e r g ' s f undament al equat i on (2. 21).
A par t i cul ar l y el egant r esul t is t he Lapl ace t r ans f or m of the sur vi val f unct i on wi t h
zer o i ni t i al sur pl us, cy(0, t), whi ch can be r eadi l y obt ai ned wi t h t he I ni t i al Val ue The or e m
of La pl a c e Tr a ns f or ms ( Spi egel [37, p. 5, The or e m 1-16]). The t heor em st at es t hat , f or a
s uf f i ci ent l y r egul ar f unct i on f,
ulim0 f(u) = ~ l i r n ~ f(~). (3. 8)
Ap p l y i n g it t o (3. 7) yi el ds
f 0 ~ e -St ~( 0, t) dt
= ~ ~, ~-( ~, a)
1
= ~-p, 8 >0 . (3. 9)
Fo r mu l a (3. 9), whi ch is the s ame as Seal [34, (4. 25)] and Panj er and Wi l l mo t [27,
Th e o r e m 11. 7. 2], may al so be der i ved usi ng (4. 9) bel ow. An expl i ci t f or mul a f or ~( 0, t) is
gi ven by (8. 23).
Mos t of the r esul t s in t hi s s ect i on can be f ound in Ar f we ds on [1]. Howe ve r , Seal
[34, p. 111] poi nt s out t hat t hey are "i mpl i c i t in Se ge r da hl ' s ( 1939) wor k on t he mome nt s
of t he t i me t o r ui n. " To c onc l ude t hi s sect i on, l et us suggest an anal ogy wi t h Li f e
Cont i nge nc i e s whi ch may be usef ul to s ome r eader s. I nt er pr et 8 as a f or ce of i nt erest .
I dent i f y T, t he t i me of rui n, wi t h T( x) , t he r emai ni ng t i me till deat h f or a per son now aged
x. Then ~( u, t) and o( u, t) c or r e s pond to t he pr obabi l i t es tqx and tPx, r es pect i vel y. Wi t h
w -= 1, 0( u) c or r e s ponds to A x, whi l e JO e~St ~( u, t )dt c or r e s ponds t o a x.
160
4. Zer o initial surplus
In this section we study functions such as f(x I 0) [defined by (2.42)] and f(x, y] 0)
[defined by (2.9)]. With initial surplus U(0) = u = 0, some very explicit results can be
obtained. Since 0 satisfies the renewal equation (2.36), it follows that
0(0) = h(O). (4.1)
Appl yi ng (2.11) and (2.38) to (4.1) yields
l ofoW(X, y) f(x, yl 0) d x d y = ~J o I o e - p Xw( x , y ) p ( x + y ) d x d y . (4.2)
Because the function w is arbitrary, it follows that
f(x, Y] 0) = ~ e-P x p(x + y), x > 0, y > 0. (4.3)
This formul a plays a central role; an alternative proof and additional insight will be given
in Section 8. Some immediate consequences can be obtained by integrating over x, y, and
both:
~ r
y I o ) d x = p ( x + y ) d x
= g ( y ) , ( 4 . 4 )
as defined by (2.33);
f(x I 0) = Iof(X, Y l 0) dr
= ~e-PX JO p(x + y) dy
= ~e- px[ l - P(x)]; (4.5)
E[e -ST I(T < oo) 1U( 0) = 0] = Io/of(X, y 10) dy dx
/ oe- X
= [1 - P(x)] dx. (4.6)
As a check, note that (4.3) and (4.5) satisfy (2.43) with u = 0.
With 8 = 0, and hence p = 0, (4.3) reduces to a result of Dufresne and Gerber [13,
(10)]. Here
f(x, Y l 0) = f(y, x I 0). (4.7)
Di ckson [7] has pointed out that this symmet ry can be explained in terms of "duality. "
Figures 3.a and 3.b illustrate the duality. Further discussion can be found in Dickson and
16 1
Egfdi o dos Rei s [9], and in Sect i on 9 bel ow. For ~5 > 0, f or mul a (4.7) does not hol d any
l onger.
For 8 = O, (4.6) reduces to the f amous f or mul a
= P ( x ) l d x
kP ~ (4.8)
c ,
whi ch, of course, can also be der i ved from (2.25). For ~5 > O, we can use (4.6) and the fact
that p is a sol ut i on of (2. 47) to see that
E[e~ST I U(O) = O] = E[e ~ST I(T < co) [ U(O) = O]
= 1 6
cp" (4.9)
Thi s resul t is equi val ent to (3.9). For mul a (4.8) can be obt ai ned as a l i mi t i ng case of (4. 9)
because of (2.48).
Ex a mp l e Let us l ook at the case of an exponent i al i ndi vi dual cl ai m amount di st r i but i on,
p(x) = 13e-[ 3x, x > 0, (4. 10)
wi t h [3 > 0 and c > ~,Pl = ~. The numbe r !3 is ~j , the nonnegat i ve sol ut i on of (2. 21), whi ch
M
is
o r
Hence
8 + k - c % - 13 + %'
c, ~ 2 + ( c l 3 - 6 - Z.) , 5, . - 136 = o.
P = % 1
X + 8 - c [ 3 + ~ / ( c l 3 - 8 - K ) 2 + 4c [ 3 8
2c
( N o t e t h at , i f 8 = O, t h e n 13 = ~ 1 = 0 . ) T h e n
fix, y l 0) = ~ - e-(P + 13)x- [~y : ~ e-(P + 13)x p(y);
~.13 ~
g( y) = ~ e-13y = c ( - ~ - - ~ p ( y ) ,
f ( xl O) = ~ e-(O + ~)x:
( 4. 11)
( 4. 12 )
(4. 13)
16 2
E[e 6T] ( T<oo) I U( 0 ) = 0 ] = t o l o f ( x , y [ 0 ) d y d x
k
- c( ~ + p ) ( 4 . 14 )
= 2~ (4.15)
C~ + ~ "t- ~, + 4 ( C ~ -- ~ -- ~) 2 + 4c[38 "
An al t ernat i ve to (4.14) and (4.15) is formul a (4.9), which is si mpl e and general at the
same t i me. In Section 7 we shall show that
E[e ~ST I(T < ~o) I U(0) = u] = E[e -fiT I(T < oo) I U(0) = 0] e{2U (4.16)
where ~2 is the negat i ve root of (4.11); see (8.33) and (8.38). Hence it fol l ows from (3.4),
(4.16) and (4.14) that
l oe_& ~/Ku, t) dt = k
eg2U. (4.17)
On the ot her hand, using (4.9) instead of (4.14) yi el ds
I0e- t , ( u, t) dt _- _ + e 2U ( 4 , 8 )
Fi nal l y, we note that (2.43) can be si mpl i f i ed to
f ( x , y [ u ) = f(x[ u) p(y), u >_O, x >O, y >O. (4.19)
5 . P os i t i v e I ni t i al S ur p l us
Resul t s concerni ng "rui n" for zero initial surplus can be t ransl at ed into results that
are rel at ed to the event that the surplus fal l s bel ow the initial level in the more general
situation where the initial surplus is posi t i ve. We can use (4.3) and (4.4) to deri ve the
renewal equat i on (2.36) by probabi l i st i c reasoning. We condi t i on on the first t i me when
the surplus falls bel ow the initial level. For given initial surplus U(O) = u > O, the
pr obabi l i t y that this event occurs bet ween t i me t and t i me t + dt, with
u + x < U( t - ) < u + x + d x
and
is
u - y - dy < U( t ) _< u - y ,
f(x, y, t ] O) dx dy dt. (5.1)
163
Fur t her mor e, the occur r ence
y >u
me a ns t hat r ui n al so t akes pl ace wi t h this cl ai m. Thus
~( u) = I oI oI oe- St ~ ( u - y) f ( x, y, t[ 0 ) d t dx dy
+ LJ oJ oe~St w(x + u , y - u ) f ( x , y, t[ 0 ) d t dx dy
U
= 1010 ~ ( u - y) f ( x, Yl 0 ) d x d r + t ul oW(X + u , y - u) f ( x, Yl 0 ) d x d e. (5. 2)
Appl yi ng (4. 4) and (4.3) to the r i ght - hand side of (5. 2) yi el ds
~(u) = 0( u - y) g( y) dy + w(x + u, y - u)e-pXp(x + y) dxdy
Fl, TW( X
= ( ~* g) ( u) + 0 + u, s)e-pXp(x + u + s)dxds
= ( ~* g) ( u) + h(u)
by (2. 39). Thi s is the pr obabi l i st i c pr oof of (2.36).
Let 8 be cons i der ed as a force of interest. Then O(u) is t he expect at i on of the
di s count ed penal t y. Fr om t he cal cul at i ons above we see that h(u) is t he expect at i on of the
di s count ed penal t y if r ui n occur s at the first t i me when the sur pl us fal l s bel ow t he i ni t i al
l evel u. Becaus e
g(y) = J oJ oe~St f(x, y, t l 0 ) dx dt, (5.3)
the di f f er ent i al g( y) dy is the di scount ed pr obabi l i t y that the sur pl us wi l l ever fall bel ow its
i ni t i al l evel u and wi l l be bet ween u - y and u - y - dy when it happens for the first t i me.
The t wo t er ms on the r i ght - hand side of the r enewal equat i on (2. 36) cor r es pond to whet her
or not r ui n occur s at the first t i me when t he sur pl us fal l s bel ow the i ni t i al l evel u.
For mul a (2. 37), the r epr esent at i on of ~ as a series of f unct i ons, has a nat ur al
pr obabi l i st i c i nt er pr et at i on as fol l ows. We obser ve that r ui n occur s at a t i me when t he
sur pl us pr ocess { U(t ) } at t ai ns a record low (cf. Sect i on 12.6 of Actuarial Mat hemat i cs
[4]). For j = 1, 2, 3 . . . . . let zj denot e the t i me of the j - t h record low of the sur pl us process.
The n
I(T < co) = ~ I(T = xj), (5. 4)
j = =1
164
and hence
( u) = E[e -ST w( U( T- ) , IU(T)I) I(T < oo)[ U( 0) = u]
= E[ e -ST w( U( T- ) , IU(T)I) I(T = xj) [ U( 0) = u], u > 0. (5. 5)
j = = l
It r e ma i ns t o s how by mat hemat i cal i nduct i on t hat t he j - t h t er m in (5. 5) i s i dent i cal t o t he
j - t h t er m in (2. 37). Fr om t he cal cul at i ons in (5. 2) and (5. 3), we see t hat
h( u) = f uI O/O e-6t w(x + u , y - u ) f i x, y, t[ 0 ) d t dx
dy
= E[ e ~' T w( U( T- ) , IU(T)I) I ( T = Xl)[ U( 0) = u], u > 0. (5. 6)
Thus ( 2. 37) and (5. 5) have t he s ame f i r st t erm. Let j be an i nt eger wi t h j > 2. As s u me t hat
it has been pr oved t hat (2. 37) and (5. 5) have t he same (j - l ) - t h t er m,
E[ e -ST w( U( T- ) , IU(T)I) I(T = "~j-l) [ U( 0) = u] = ( g . g * . . . * g* g *h)(u). (5. 7)
j - 2
We are t o s how t hat (2. 37) and (5. 5) have t he s ame j - t h t erm. Be c a us e zj - "c I has t he
s a me di s t r i but i on as Xj -l , we have
E[ e ~5T w( U( T- ) , IU(T)I) I(T = zj ) [ U( 0) = u]
= E[e-'SxJ w( U( T- ) , IU(T)I) I(T = xj )[ U( 0) = u]
= E[e-~Xt +z, - x, ) w( U( T- ) , IU(T)I) I ( T = xj )[ U( 0) = u]
= [ g* ( g* g* . . . * g* g * h)](u)
j - 2
= ( g . g * . . . * g* g *h)(u), (5. 8)
j - I
whi ch is t he j - t h t er m in (2. 37). Hence (2. 37) has a pr obabi l i s t i c i nt er pr et at i on.
I f we c ons i de r f(x [ u) and f(x, y[ u) as f unct i ons of u, t hey sat i sf y r enewal
equat i ons s i mi l ar t o (2. 36). By di s t i ngui s hi ng whet her or not rui n occur s at t he f i r st t i me
when t he sur pl us f al l s bel ow the i ni t i al val ue u, we see t hat
uf
f ( x , y [ u ) = I0 ( x , y [ u - z ) g ( z ) d z + f ( x - u , y + u ] 0 ) , 0 _< u < x , (5. 9)
and
U
f(x, y I u) = 10f(x, y[ u - z) g( z) dz,
O< x < u . ( 5. 10)
165
By (4. 3),
f(x - u, y + u I O) = ~ e-P(x-u) p( x + y). (5.1 I)
Hence, f or u _> 0, x > O, y > 0,
U
f(x, y l u ) = /0f ( x, y l u - z ) g ( z ) d z + ~e- p( x- u) p( x + y ) I ( x >u ) . (5. 12)
Fr om t hi s and (2. 43) it f ol l ows t hat , for u _> O, x > 0,
f ( x l u ) - - / , Li f ( x l u - z ) g ( z ) d z + ~ e - P ( x - u ) [ l - P ( x t ] I I x >u ) . (5. 13)
We obs er ve t hat , as a f unct i on of x, f(x ] u) has a di s cont i nui t y of amount
~ [ l - P ( u) ] ( 5 . 14 )
at x = u. Re ma r ka bl y, it does not depend on 8.
Equat i ons (5. 12) and (5. 13) can be vi ewed as speci al cases of (2. 36). Let xo be a
pos i t i ve number . Cons i de r w( x, y) as t he "ge ne r a l i z e d" dens i t y f unct i on wi t h mas s 1 f or
x = xo and 0 f or x : xo ( and i ndependent of y). Then
, ( u ) = f(x0[ u), ( 5. 15)
and, by ( 2. 38) ,
h(u) = "~c 1~o1,~o l ( x > u) e- P( x- u) w( x, y ) p ( x + y ) d y
dx
= ~ e-P , : X , , u) [ 1 - P ( x 0 ) ] I( x 0 > u)
= f ( xoI 0) epUI(x0 > u). (5. 16)
He nc e ( 5. 13) is a s peci al case of (2. 36).
Fr o m (5. 13) it f ol l ows t hat
f ( x l u ) -- ht u) + ( g* h) ( u) + ( g* g* h) ( u) + ( g* g* g* h) ( u) + . . . , (5. 17)
wi t h
h( u ) = f ( x ] 0 ) e pu I ( x > u ) . ( 5. 18)
N o t e t hat t hi s se r i e s has a pr o b ab i l i s t i c i nt e r pr e t at i o n s i mi l ar t o ( 2. 3 7) : T h e j -t h t e r m
r epr es ent s t he cont r i but i on to f(x I u) of the event t hat ~j = T. Wi t h t he def i ni t i on
rl (u ) = eP u I(x > u), ( 5. 19)
( 5. 18) b e c o me s
h(u) = f(x t 0) rl (u), ( 5. 20)
1 6 6
and we can r ewr i t e ( 5. 17) as
f ( x l u ) = f ( x[ 0) [ r l ( u ) + ( g* r l ) ( u) + (g*g*~l )(u) + ( g* g* g* r l ) ( u) + . . . ] . (5. 21)
6. Key Renewal Theorem
Let f ( x) and z( x) be t wo nonnegat i ve f unct i ons on [0, oo). Cons i de r t he i nt egr al
equat i on
Z( x) = ( f * Z) ( x) + z( x) , x _> 0, (6. 1)
whi ch is a r enewal equat i on for Z( x) . The s o- cal l ed key renewal theorem, or i gi nal l y
f or mul a t e d by Wa l t e r L. Smi t h, gi ves the as ympt ot i c be ha vi or of t he sol ut i on of a r enewal
equat i on. It st at es t hat , i f f is a pr oper pr obabi l i t y dens i t y f unct i on, i. e. ,
I of ( x) dx = 1, (6. 2)
and t he f unct i on z is suf f i ci ent l y r egul ar , t hen t he sol ut i on of t he r enewal equat i on (6. I )
sat i sf i es
f o z( y) dy
xli~rn Z( x) - oo (6. 3)
0 y f l y) dy
In ear l i er days t here was s ome conf us i on about t hi s r esul t becaus e it was gi ven under a
var i et y of hypot hes es . Fi nal l y, Fel l er [17, p. 362] cl ar i f i ed t he si t uat i on wi t h what he
cal l ed t he direct Riemann integrability condi t i on. The condi t i on r equi r es t hat t he f unct i on
z be Ri e ma n n i nt egr abl e and not os ci l l at e t oo much in a ne i ghbor hood of i nf i ni t y. Thi s
not i on can be f ound in Nor ber t Wi e n e r ' s wor k on Tauber i an t heor ems. In our
appl i cat i ons , t hi s condi t i on is al ways sat i sf i ed. A r ecent book wi t h a l ong chapt er on
r enewal t heor y is Res ni ck [321. Sever al act uar i al books ([51, [19], [34]) cont ai n
di s c us s i ons on a ppl yi ng r enewal t heor y to r i sk t heor y.
If t he f unct i on f is not a pr oper pr obabi l i t y dens i t y f unct i on, i.e., i f
f ( 0) = J of ( x) dx , I, ( 6. 4)
t hen we t ry t o f i nd a real numbe r R such t hat
16 7
~ ( - R ) = J oeRXf(x)dx = 1. (6. 5)
f ~
Ap p l y i n g ( 2. 31) wi t h k = R to (6. 1) yi el ds a pr oper r enewal equat i on f or the f unct i on
eRxZ( x) ,
eRxZ = (eRxf)*(eRxZ) + eRxz. (6. 6)
The key r enewal t heor em is appl i cabl e to (6. 6), yi el di ng
f o eRy z( y) dy
x l i ~ eRxZ(x) = o=
f O y f(y) d y
e R Y
2( - R)
- - f ' ( - R) ' ( 6 . 7 )
The numbe r R s at i s f yi ng (6. 5) is uni que, becaus e
d--~f(-~) = /oe~Xx f ( x ) d x > 0.
I f 7(0) < 1 ( f a def ect i ve dens i t y) , t hen R > 0; i f f ( 0) > 1 ( f an e xc e s s i ve dens i t y) , t hen
R < 0. Let f l ( x) and r e( x) be t wo f unct i ons; we wr i t e
f l ( x ) ~ f 2 ( x ) f o r x ~ oo ( 6 . 8 )
i f
l i m f l ( x) 1.
x -~ ~ f2(x) -
Then (6. 7) can be r es t at ed as
Z( x ) ~ - -
2( - R) e_Rx f or x - ~ . (6. 9)
- f ' ( - R)
R e ma r k s (i) The r eader who is f ami l i ar wi t h c ompl e x anal ys i s mi ght r ecogni ze t hat t he
r i ght - hand si de of ( 6. 9) can be i nt er pr et ed as a residue. As poi nt ed out in Re ma r k (v) of
Sect i on 2, a r enewal equat i on can s ol ved by t he met hod of La pl a c e t r ans f or ms . Fr om (6. 1)
we have
and he nc e
7,, = ? Z + 2, ( 6. 10)
2
= 1 - V (6. 11)
1 6 8
Consequent l y,
(b+ioo ~(~) e~ x d E, (6.12)
Z(x) = 12/tidb_i~ 1 - ~ )
The residue of the integrand at the si mpl e pole ~ = - R is:
I~rnR (~ - - R) z(~)_ e~ x 2(-R) e_Rx ' (6.13)
1 - f ( ) = - f ' ( - R)
(ii) With f bei ng a proper density function, formul a (6.3) can be obt ai ned by appl yi ng the
Final Value Theorem of Lapl ace Transforms (Spiegel [37, p. 6, Theor em 1-17]), The
t heorem states that, for a sufficiently regular function Z,
x l i ~ Z(x) = li m ~ Z(). (6.14)
Here
. 2 ( )
~(o)
-?' (o)'
because f(0) = I.
7 . L a r g e I n i t i a l S u r p l u s
We now appl y the key renewal t heorem to find an asympt ot i c expressi on for 0
satisfying (2.36),
= *g + h,
where g and h are defined by (2.33) and (2.34), respectively. Since p _> O, we have
~(0) _< ~-p~ < 1,
whi ch means that (2.36) is a defect i ve renewal equation. Thus we seek R > 0 such that
= ~( - R) = l0 g(x) dx, (7.1)
1
eRx
which is equation (2.57). Hence
R = -~2,
where ~2 is the negative root of Lundber g' s fundamental equation (2.21). Note that both p
(or El) and R (or 1~21) are increasing functions of 15 and do not depend on the penal t y
16 9
function w. When confusion may arise, we wri t e p(~5) for 9 and R(6) for R.
that p( 0) = 0, and Ri 0) is the adj ust ment coeffi ci ent in cl assi cal risk theory.
It f ol l ows from the key renewal t heorem that
la(-R) e_Ru
do(u) ~ - ~ ' ( - R)
By (2.60),
~ ( - R) -
By (2.59),
Hence
We obser ve
for u --~ o ~ . (7.2)
c ( R~ - ~ [ o f ( 7 ( e Ru - e- 9 u) w(u, y) p( u + y ) d y du. (7.3)
l [ I + ~ I3' ( - R) ] . ( 7 . 4 )
g ' ( - R ) - 9 + R
X f / f 0 ~ w(x, y) ( e Rx- e ~x) p(x + y) dx dy
d0(u ) ~ e -Ru for u --~ oo. (7.5)
-XO' (-R) - c
Now, consi der the special case where w(x, y) -- 1 and S = 0. Then O = W, and the
renewal equat i on (2.36) is
= ~* g + h,
wi t h
( 7 . 6 )
and
g(x) = ~[1 - P(x)] (7.7)
Equat i on (7.6) is the same as Exerci se 12.11 in Actuarial Mathematics [4]. Because R is
the sol ut i on of (2.49), we have
/,T/o(e R x - I ) p ( x + y ) d x d y = ] o( e R x - 1 ) [ l - P ( x ) ] d x
- - t dx
= ~ - Pl. (7.9)
170
Hence (7.5) si mpl i fi es as
c - LPl e_Ru for u - - ) ~ . (7. 10)
~( u) - ~ ' ( - R ) - c
Seal [34, p. 131] poi nt s out that the asympt ot i c f or mul a (7. 10) was first publ i shed by the
Swedi sh act uary F. Lundber g in 1926; the special case of (7. 10) for const ant { Xj } was
deri ved by A. K. Er l ang in 1909 in the cont ext of t el ephone calls.
We can also obt ai n an asympt ot i c f or mul a for e 4t ~( u, t) dt, 8 > 0. By (3.4),
I
f o r e St ~( u, t) dt - (7. 11)
O(u)
8 '
wher e w(x, y) - I. It fol l ows f r om (2. 62) and (2. 46) that
Hence
,
ta(-R) - c ( R + l 9 )
_ ( 5 t l , 1 ~
(7. 12)
5 1 1 e_Ru for u ---) (7. 13)
, ( u ) - - ~. 0' ( - R) - c(R- + O)
[ whi ch can al so be deri ved by appl yi ng (2. 47) to (7.5)]. Subst i t ut i ng (7. 13) in (7. I 1)
yi el ds
f o ~ 1 1 1 e_Ru foru---~ (7. 14)
e -& ~( u, t) dt - X~' ( - R) - c ( R + -0-) o~.
Lundbe r g' s asympt ot i c f or mul a (7. 10) is, of course, a special case of (7. 13):
v ( u) - - , ( u)
l t5 ~ e_R(8)u,
- - ~l~rn0 - ) ~' ( - R( 8 ) ) - c [ R ~ + 9 - ~ ] for u ---) oo, (7. 15)
whi ch, by (2. 48), is (7.10). A mor e interesting way to retrieve (7. 10) is by appl yi ng the
Final Val ue Theor em (6. 14):
~( u) = lira ~( u , t )
t ----> ~
= ~ o 8J o e - a v ( u ' t)
dt,
whi ch, vi a (7. 14), is (7. 15) again.
Si nce w(x, y) can be an arbi t rary funct i on, compar i ng (2.11 ) with (7.5) yi el ds the
asympt ot i c f or mul a
171
f(x, y [ u) ~ ~'(eRX - ewx) p(x + y) e_Ru for u --9 ~, (7. 16)
-~. O' (-R) - c
whi ch gener al i zes Duf r es ne and Ger ber [13, (24)]. Because (5. 13) is a r enewal equat i on
for f(x, y [ u) (as a f unct i on of u), it can al so be used to der i ve (7. 16) here
h(u) = f(x, y ] 0) eP u I(x > u), (7. 17)
and
h( - R) = fi x, y I 0) / O eRu e9 u I(x > u) du
e(R + p)x _ 1
= f(x, y l 0 ) ~ - 7 ~ -
~(e Rx - e-PX)p(x + y)
c(R + 9)
(7. 18)
Ex a mp l e For the exponent i al i ndi vi dual cl ai m exampl e in Sect i on 3, we have
~. + (5 - cl3 - V/(cl3 - 8 - ~)2 + 4c13 8
- R = ~2 = 2 c
(7. 19)
and the adj us t ment coef f i ci ent is
c [ 3 -
R(O) - c
(7. 20)
Fr om (4. 13)
x13
fg(F.,) = c(13 + p)(13 + ~) ' ( 7. 21)
and hence
- ~ ' ( - R) = c([3 + 9) ( [ 3- R) 2 (7. 22)
Now, let us cons i der the par t i cul ar case where w(x, y) = w(y), a f unct i on not dependi ng on
x. The n
re(x) = loW(Y) p(x + y) dy
= [3e-13Xlow(Y ) e-!3y dy
= [],e-13x ~v(13), (7. 23)
and
172
h ( x ) = ~ ] o e ~ Zm( u + z ) d z
=
_ )~, 13 ,v(13) e -13x
c(13 + p )
( 7. 24)
He n c e
f i ( ~ ) =
c([3 + o) ( [ 3 + ~ )
( 7. 25)
It f o l l o ws f r o m ( 7. 25) and ( 7. 22) t hat
f i ( - R)
-~-W~-R) - ' ~(13) (13 - R) .
Thus , wi t h w( x, y) = w( y) , a nd P( x) = I - e-!3x,
* ( u) ~ ,'(13) (13 - R) e - Ru f or u --4 ,,o.
( 7. 26)
( 7. 27)
In t he ne xt s e c t i on, we s hal l s ee t hat ( 7. 27) is in f act an e q u a l i t y va l i d f or all u > 0.
Fu r t h e r mo r e , ( 7. 16) i s
f(x, Y l u)
~,(e Rx - e ~ x ) 13e-~( x + y)
2.13(~ - R) - 2 - c
e - Ru f or u --~ oo.
( 7. 28)
Be c a u s e p and - R ar e t he r oot s o f ( 4. 11) , we h a v e
c(13 + 0)(13 - R) = c ( - D 2 + (c13 - 8 - ~. )(-13) - 138
= ; q 3 . ( 7. 29)
It f o l l o ws f r o m ( 7. 29) and s o me a l g e b r a t hat ( 7. 28) can be r e wr i t t e n as
f ( x, y[ u) ~ ;~13(13 - R) [eRx _ e - px] e_13(x+Y)e_R u f or u --4 oo.
c ( R + p)
( 7. 30)
Ap p l y i n g ( 4. 19) t o ( 7. 30) yi e l ds
)~(13 - R) [eRx _ e - px] e_!3xe_R u
f ( x l u ) ~ c ( R + p)
f or u- - + o, , . ( 7. 31)
It t ur ns out t hat ( 7. 30) a nd ( 7. 31) ar e e x a c t f or 0 < x < u; s e e ( 9. 40) b e l o w. IIII
To c o n c l u d e t hi s s e c t i on we l ook at t he La p l a c e t r a n s f o r m o f T, g i v e n t hat r ui n
oc c ur s :
173
E[e 4 T I(T < ~) I U(0) = u]
E[e ~ST ] T < oo, U( 0) = u] =
E[ I ( T < ~o) I U(0) = u]
~(u)
~ ( u ) '
f i ( - r )
wher e w =- 1. Cons i der ~ as a f unct i on o f & and write it as C(6).
(7.2) that, for u - + oo
~(u) ~ C( 6) e- R(;6)u
~( u) C( 0) e- R~0~u
If 8 > 0, t hen R(8) > R(0), and hence
whi ch means that
Thus, for each t > 0,
C(6)
- C ( 0 ) e [ R( 8 ) - R( 0 ) l u .
ul i ~n e [R(~5)-R(O)lu = O,
u l ~ E[e-"ST I T < oo U(0) = u] = 0.
(7. 32)
It fol l ows f r om
(7. 33)
(7. 34)
u l ~ Pr [ T < t I T < oo, U(0) = u] = 0, (7. 35)
whi ch i mpl i es that, for a large i ni t i al surpl us u and gi ven that r ui n occurs, it occurs late.
The resul t (7. 35) is compat i bl e with the obser vat i on that the condi t i onal expect at i on
E[T I T < ~, U(0) = u]
is essent i al l y a l i near f unct i on in u in some cases; see (8. 43) bel ow, Ger ber [19, p. 138,
Exampl e 3. 2], and Seal [34, p. 114].
8 . M a r t i n g a l e s
Fur t her i nsi ght can be pr ovi ded to the reader who has some f ami l i ar i t y wi t h
martingales. Let ~ be a number . Because {U(t)}t_> 0 is a st ochast i c process wi t h st at i onar y
and i ndependent i ncr ement s, a process of the form
{ e - & + ~U(t)}t_>0 ( 8. 1)
is a mar t i ngal e if and onl y if, for each t > 0, its expect at i on at t i me t is equal to its i ni t i al
val ue, i.e., if and onl y if
E[e~St + ~,u(t) I U(0) = u] = e 4 0 + ~u
= e~ u. (8.2)
174
Si nce
E[ e ~ t + ~,u(t) ] U( 0) = ul = exp( - ~t + ~u + ~ct + ~,t[~(~) - 1]),
the mar t i ngal e condi t i on is that
0 = - ~ + c~ + Z, [ ~( ~, ) - l l ,
whi ch is (2. 21), Lundbe r g' s f undament al equat i on. Thus, for (8.1) to be a mar t i ngal e, t he
coef f i ci ent of U(t ) in (8.1) is ei t her ~1 = 9 -> 0 or ~2 = - R < 0.
Wi t h such a ~, (8.2) hol ds for each fi xed t, t _> 0. However , i f we repl ace t by a
st oppi ng t i me whi ch is a r andom vari abl e, t hen t here is no guar ant ee that (8.2) wi l l hol d.
For t unat el y, it hol ds in t wo i mpor t ant cases, as we shall see in this and the next paragraph.
I f the st oppi ng t i me is T, the t i me of rui n, the optional sampling theorem is appl i cabl e to
the mar t i ngal e wi t h ~ = - R. For 0 _< t < T,
& + RU(t ) > 0,
and hence
0 < e ~ t - RU( t ) < 1.
Wi t h {e -St -RU(t ); t < T} bei ng bounded, the opt i onal sampl i ng t heor em is appl i cabl e and
we obt ai n
E [ e - f T - RU(T) [ U(0) = u] = e -Ru . (8. 3)
Fur t her mor e, it f ol l ows f r om (2.6) that, even if ~ = 0,
E[ e~ST- RU(T) I(T = oo) 1 U(0) = u] = 0.
Cons equent l y, we can rewri t e (8.3) as
e-Ru = E [ e ~ T - RU(T) I(T < ~o) I U(0) = u], ~5 > 0, u > 0. (8. 4)
The above is a pr oof by mar t i ngal e t heory of a gener al i zat i on of Theor em 12.1 in
Actuarial Mathematics [4].
We now show that the quant i t y e-O(x-u), whi ch appears t hr oughout t hi s paper
( usual l y wi t h u = 0), has a pr obabi l i st i c i nt erpret at i on. For x > U(0) = u, let
Tx = rai n {t l U( t ) = x} (8. 5)
175
be t he first t i me when the sur pl us reaches the level x. We can use equal i t y to def i ne the
st oppi ng t i me Tx because t he process {U(t)} is ski p-free ( j ump- f r ee) upwar d. Then, for
0 - < t < T x ,
e-qSt + pU(t) < epX. (8.6)
Hence we can appl y the opt i onal s ampl i ng t heor em to the mar t i ngal e {e -'St + 9u(t)} to
obt ai n
e-,50 + 9u = E[e~Tx + pufro I U(0) = u]
= E[ e~T~ ] U(0) = u] eP x,
or
e - 0 ( x - u) = E[e-6Tx [ U(0) = u]. (8.7)
Wi t h ,5 i nt er pr et ed as a force of interest, the quant i t y e-P( x - u) is the expect ed di s count ed
val ue of a pa yme nt of l due at the t i me when U(t ) = x for the first t i me. We not e that
(8. 7) r emai ns val i d even if u is negat i ve. The r equi r ed condi t i on is x > u; t he condi t i on
u >_ 0 is not needed anywher e in the deri vat i on. For mul a (8.7) was pr obabl y first gi ven by
Kendal l [25, (14)], al t hough he di d not provi de a compl et e proof. It can al so be f ound in
Cox and Mi l l er [6, p. 245, (184)], Ger ber [21, (1 I)], Pr abhu ([30], [31, p. 79, Theor em
50) ; p. 105, #41), and Tak~cs [38, p. 88, Theor em 8].
For mul a (8.7) can be used to gi ve an al t er nat i ve pr oof of the i mpor t ant f or mul a
(4.3). For x > u = U(0), let gl (X, t t u), t > 0, denot e the pr obabi l i t y densi t y f unct i on of the
r andom var i abl e Tx. Hence (8.7) is
f
0 e ~ t ~l(X, t] u) dt = e- p( x- u) . (8. 8)
The di f f er ent i al ~l (X, t l u)dt is t he pr obabi l i t y that the sur pl us process upcr osses l evel x
bet ween t and t +dt and that t hen this happens for the first time. For U(0) = u _> 0, x > 0,
let r~2(x, t I u), t > 0, be the f unct i on def i ned by the condi t i on that rc2(x, t] u)dt is the
pr obabi l i t y that rui n does not occur by t i me t and that t here is an upcr ossi ng of the sur pl us
process at l evel x bet ween t and t+dt. It can be pr oved by dual i t y, a not i on to be
di scussed in the next sect i on, that
176
~ l ( x , t [ 0 ) = r t 2( x, t [ 0) , x >0 , t >0 . (8. 9)
Now, f(x, y, t I u) dt dxdy can be i nt er pr et ed as t he pr obabi l i t y of the event t hat "r ui n" does
not t ake pl ace by t i me t, t hat t he sur pl us pr oces s upcr os s es t hr ough l evel x bet ween t i me t
and t i me t +dt , but does not at t ai n l evel x+dx, i. e. , t hat t her e is a cl ai m wi t hi n @ t i me
uni t s af t er Tx, and t hat t he si ze of t hi s c l a i m is bet ween x+y and x+y+dy. Thus
f(x, y, t l u) dt dx dy = [rt2(x, t I u) dt ] [% _d~] [p(x + y) dy] , ( 8. 10)
f r om whi ch it f ol l ows t hat
f(x, y, t I u) = ~ p( x + y) 7t2(x, t] u). (8. I 1)
Thi s f or mul a is par t i cul ar l y usef ul i f u = 0: It f ol l ows f r om (8. 9) t hat
f(x, y, t I 0) = ~ p( x + y) 7tl(x, t l 0). ( 8. 12)
I f we mul t i pl y (8. 12) by e ~St, i nt egr at e f r om t = 0 to t = oo, and appl y (8. 8) wi t h u = 0, we
obt ai n (4. 3),
f(x, y[ 0) = ~ p( x + y) e- P x,
once agai n.
R e ma r k s (i) For x > u = U( 0) > 0, t he f unct i ons tel(X, t] u) and 7t2(x, t[ 0) can be
e xpr e s s e d in t er ms of 7t3(x, t I u), t he passage time density of t he sur pl us pr oces s at t he
l evel x. The di f f er ent i al rt3(x, t ] u) dt is t he pr obabi l i t y t hat t he sur pl us pr oces s upcr os s es
l evel x bet ween t and t +dt . Thi s is t he s ame as t he pr obabi l i t y t hat t he sur pl us at t i me t is
bet ween x--dx and x wi t h dx = c dt. Hence, we have
7t3(x, t I u) = c fs(t)(u + ct - x), (8. 13)
wher e
f s( t ) ( s) = e - k t ~ ' ( ) ~ t) n
n=/'U0 n! P*n(s) ( 8. 14)
is t he pr oba bi l i t y dens i t y f unct i on of S(t ), t he aggr egat e c l a i ms up t o t i me t. The
f ol l owi ng ver s i on of t he ballot theorem,
r c | ( x , t [ 0 ) = xt ~3( x, t l 0 ) , x >0 , t >0 , ( 8. 15)
177
was f i r st gi ven by Kendal l [25, (17)]; see al so Cox and Mi l l er [6, p. 251, # 12] , Di nges
[12], Kei l s on [241, Pr abhu [31, p. 81, The or e m 61, Seal [35, p. 471 and Tak~ics [38, p. 87,
Th e o r e m 6]. For x > u and t > 0, because
rtl(x, t l u) = ~ l ( x - u , t [ 0 ) ( 8 .16)
and
we have
n 3 ( x , t ] u ) = r t 3 ( x - u , t l 0 ) , (8. 17)
x - u t t u ) (8.18)
rtl( , tl u) = ~T---re3(x,
Ge r be r [20, Th e o r e m 2] has gi ven a pr oof of
r t l ( x, t I 0) = ~tt rt3(x, t] 0) = rt2(x, t I 0) ( 8. 19)
by mar t i ngal es . The s econd equal i t y of (8. 19) is equi val ent to equat i on (2. 1) on page 112
of Ge r b e r [ 19].
(i i ) The di f f er ent i al 7t2(x, t t u)dt can be i nt er pr et ed as the pr obabi l i t y t hat rui n does not
occur by t i me t and t hat t he sur pl us at t i me t is bet ween x - d x and x, wher e dx = c dt. By
(3. 6),
c ( u , t) = Pr [ T > t l U(0) = u]
= f o~z ( X, tl u) - ~. ( 8. 20)
Hence, it f ol l ows f r om t he s econd equal i t y of (8. 19) and f r om (8. 13) t hat
f [ x 3( x , tl
~ ( 0 , t) = ~
= - ~ x fs(t)(ct - x) dx, (8.21 )
whi ch is a r esul t fi rst gi ven by Pr abhu [30, (4. 6)]; see al so Seal [34, (4. 8)]. Let
Fs( t ) ( s) = e - kt ~" (~t)n P*n(s) (8. 22)
. ~ 0 n!
be t he pr oba bi l i t y di s t r i but i on f unct i on of S(t ). Then i nt egr at i ng the r i ght - hand si de of
( 8. 21) by par t s and not i ng t hat Fs(t )(s) = 0 f or s < 0 yi el ds
178
~3(O,t) = 0 - 0 + ~ F s a ) ( c t - x ) d
1 [-ct
= ~ 3 o F s ( o ( c t - x) dx
I [-ct
= c-t J 0 Fs( ( s) ds, ( 8. 23)
whi ch is the same as Ge r be r [19, p. 113, ( 2A) ] , Panj er and Wi l l mot [27, The or e m 11.7.3],
and Seal [35, p. 48, (6)].
(i i i ) By (8. 20), (8. 9), and (8. 8), t he Lapl ace t r ansf or m of the sur vi val f unct i on cy(0, t) is
e 'St 0( 0, t ) d t = f ~ e ~t [ ~ 0 a:2(x, t} O) ~8 . ] d t
dO
lfo~fo
= 6- e-St ~;l(X, t ] 0) dt dx
= ~- e-P x dx
= ~ , ( 8. 24)
whi ch is t he s ame as (3. 9).
(i v) For u > 0, t > O, cons i der t he pr obabi l i t y
Pr [ U( t ) >_ 01 U( 0) = u] = Fs(tX u + ct).
By c ondi t i oni ng on whet her or not rui n occur s bef or e t i me t and di s t i ngui s hi ng a c c or di ng
to t he t i me "r when t he sur pl us pr oces s upcr os s es t he l evel 0 f or t he l ast t i me, we have t he
f ol l owi ng equat i on f or t he sur vi val f unct i on o( u, t),
Fs( t ) ( u + ct) = cJ(u, t) + n3(0, z l u) o( 0, t - x) d'~, ( 8. 25)
whi ch, in t he cont ext of Ri sk Theor y, was fi rst gi ven by Pr abhu [30, (3. 3)]. Wi t h (8. 13),
f or mul a (8. 25) is the same as Ger ber [19, p. 114, (2. 13)], Panj er and Wi l l mo t [27,
Th e o r e m 11. 7. 5], Seal [34, (4. 16)] and Seal [35, p. 44, (1)]. As t --~ oo, (8. 25) b e c o me s
1 = ~( u, oo) + rt3(O, "t I u) 0( 0, oo) dz,
or
179
w( u ) = [ 1 - y( O)1
J
x, ( 0 , I I u ) dT
( 8. 26 )
0
c-hp, =
=c
s
%( O, -c I u ) d%
u 2 0 . ( 8. 27)
0
One may v i e w f o r mu l a ( 8. 27) as ano t he r v e r si o n o f t he bal l o t t he o r e m. Se ah [ 3 3 , p. 426 1
has po i nt e d o u t t hat ( 8. 27) i s no t pr act i cal f o r co mpu t i ng.
( v ) I t f o l l o ws f r o m ( 8. 26 ) t hat , f o r x < u = U ( O),
x3 ( x, t 1 u ) dz = n3 ( 0 , z] u -x)dr
w( u - xl
I I ---
1 - w( O)
( 8. 28)
Fo r t he case x > u = U ( O), be cau se ( U ( t )] has a po si t i v e dr i f t . t he su r pl u s wi l l r e ach
l e v e l x wi t h pr o babi l i t y I . He nce , f o r x > u ,
x3 ( x, z( u ) dz = 1 +
x3 ( x, z I x) dT
( 8. 29)
= 1 + wx-x)
1 - w( O)
=&i g
( 8. 3 0 )
by ( 8. 28). Fu r t he r mo r e , ( 8. 29) can be ge ne r al i ze d as
n3 ( x, ~ 1 u ) e -6 r dr = e -P( x-u )[ 1 +
s
o m x3 ( x. 5 1 x) e &r dr ]
=
e -P( X -U )[ l + ~ u ~ ~ ~ ( O, . ~ I O)e dr dr ] , x>u . &>O. ( 8. 3 1)
Fu r t he r r e su l t s ar e gi v e n i n Re mar k ( v i ) o f Se ct i o n 9.
Exampl e Agai n, co nsi de r t he case whe r e t he i ndi v i du al cl ai m amo u nt r ando m v ar i abl e
has an e xpo ne nt i al di st r i bu t i o n, p( x) = pe -o x. T he n R i s gi v e n by ( 7. 19). Appl yi ng ( 4. 19)
t o ( 8. 4) yi e l ds
e -Ru =
[ I , e RWy)dy] We
AT I ( T < -) I U ( 0 ) = u ]
= 0 ( -R) E[ e AT I ( T < -) 1 U ( 0 ) = u ]
He nce
= ---!-E[ e dT I ( T < -) I U ( 0 ) = u ] .
P-R
( 8. 3 2)
E[ e qb( T c -) ( U ( 0 ) = u ] = ye -Ru . ( 8. 3 3 )
180
For mul a (8. 33) shoul d be compar ed wi t h the first l i ne of (12. 3. 8) in Actuarial
Mathematics [4] whi ch is onl y for (5 = 0. To r econci l e (8. 33) for u = 0 wi t h (4.9) we need
to show that, for (3 > 0,
R 6
]3 - cp" (8. 34)
Equat i on (8. 34) hol ds because the pr oduct of the t wo root s of t he quadrat i c equat i on (4. 11)
-138
is ~ - - . As a furt her check, we want to see that (8. 33) wi t h u = 0 is consi st ent wi t h (4. 14);
here we need the i dent i t y
[ 5 - R ~,
13 - c ( ~- TN' (8.35)
whi ch is true because of (7.29). It f ol l ows f r om (4. 13) and (8. 35) that
g(y) = ( [ 3- R)e-13y - p(y)
~ ( - R ) (8. 36)
In t he par t i cul ar case where w(x, y) = w(y), a f unct i on not dependi ng on x, we can
appl y (4. 19) and ( 8 3 3 ) to obt ai n an expl i ci t expr essi on for 0(u):
0(u) = E[e -ST w(IU(T)I) I ( T < ,,~)1 U( 0) = u]
o o
= [f0 w( y) p( y) dy] E[ e - ~ I(T < ~) 1U( 0) = ul
= [IoW(y)e-13ydy]([3- R)e-Ru
= q(13)(13 - R)e -Ru. (8. 37)
Thi s s hows that t he asympt ot i c f or mul a (7. 27) is act ual l y an exact f or mul a, and
~(u) = ~(0)e -Ru. (8. 38)
Fur t her mor e, for (5 > 0,
E[e ~ST w(IU(T)I) I ( T < oo) ] U(0) = u]
E[e~STw(IU(T)I) ] T < o% U(0) = u] = E[ I ( T < oo) I U(0) = u]
(u)
~( u)
_ ~,(13)[13 - R(8)] e_[R(8 )_ R(0)lu. (8. 39)
[3[[3 - R(0)I
Becaus e of (8. 35) [or (7.29)], we can rewri t e (8. 39) in t erms of 9((5) and p(0). Not i ng that
i)(0) = 0, we have
E[e-6Tw(IU(T)I) ] T < ~,, U( 0) = u] - ~, (13) exp(_L[ 1 1]u). (8. 40)
13 + p( ( 5) c I + p( ( 5)I 13
181
Di f f er ent i at i ng ( 8. 39) wi t h r espect to 8 and t hen set t i ng 8 = 0 yi el ds
E[ T w(IU(T)I)I T < ~, U( 0) = u] -
~ ( 13) R' ( O) + ~( fl) R'( O)
]3I 6 - R( O) ] [3
whi ch is a l i near f unct i on in u. By (7. 20)
- R( o ) :
Fr o m (7. l 9)
- - u , (8. 41)
He nc e
X
R' ( 0) - (8. 42)
c(c13 - X)
E[ T w(IU(T)I) I T < oo, U(O) = ul - ff(13)--R'(O)[3 '-~[c + u]
~ ( 13) z . c u] .
- +
(8. 43)
Her e,
9 . G e n e r a l i z a t i o n o f Di c k s o n ' s F o r mu l a
In [7] Di cks on gi ves t he f ol l owi ng r esul t f or t he case 8 = 0:
1 - ~ ( u )
/ f i x I o) i = V~05'
f ( xl
u)
\
tC(u - x) - tg(u)
- f ( x I o) ~ - ~ - o ~ '
x >u ~ 0 , (9. 1)
0 < x S u . (9. 2)
f ( xl o) = ~X[l - P( x) ] ; (9. 3)
see (4. 5). The pur pos e of t hi s sect i on is to gener al i ze (9. 1) and (9. 2) t o the cas e wher e
8 >0 .
A fi rst ques t i on is how t o ext end the def i ni t i on of ~( u) for 8 > 0. It t urns out t hat
t he a ppr opr i a t e def i ni t i on is
~( u) = E[e -"ST + pU(T) I(T < co) [ U( 0) = u], u > 0. (9. 4)
Thus ~( u) = ~(u) wi t h w( x, y) = e-PY; see (2. 10). ] Compar e (9. 4) wi t h (8. 4). ] Then
Di c k s o n ' s f or mul a can be gener al i zed as
182
wi t h
eP u - ~( u)
; f(x { 0) x > u >_ 0, (9. 5)
f ( x) u) = ( I - V(0) '
\
eOXgt(u - x) - gt(u) 0 < x < u, (9. 6)
f ( x 10) 1 - ~t(0) ' -
f ( x l 0 ) = ~e- pX[ l - P(x)l (9. 7)
accor di ng t o (4.5). Hence, as a f unct i on of x, f(x I u) has a di scont i nui t y of amount
f(u I 0) eP u = ~ [ I - P(u)] (9. 8)
at x = u. Thi s is t he same resul t as (5. 14). [ Compar e f or mul as (9. 5) and (9. 6) wi t h (5. 21). ]
To pr ove (9.5) and (9.6) we need some mor e concept s. We begi n by ext endi ng
t he def i ni t i on of t he st oppi ng t i me Tx as gi ven by (8.5). For a real number x, we now let
Tx denot e t he t i me of the first upcr ossi ng of t he surpl us t hr ough t he l evel x; we set
Tx = oo i f t he surpl us never upcr osses t hr ough t he l evel x. For x > U( 0) , t hi s is t he s a me
as (8. 5). For x < U( 0) , t he surpl us has to dr op bel ow the l evel x bef or e it can e ve r upcr os s
t hr ough x. We cal l the st oppi ng t i me T0 t he time ofrecoveo,; it is the first t i me when t he
sur pl us r eaches zer o af t er ruin. It f ol l ows f r om (8. 7) that, for a < b,
E[e-&Tb-T~) [ Ta < Tb] = e - P( b - a). (9. 9)
He n c e
EleCT,, I ( T < ,,o) I U( 0) = u] = E[ e ~ST e ~ [ 0 - U(T)] I ( T < oo) [ U( 0) = u]
= ~ ( u ) . ( 9 . 1 0 )
Thi s f or mul a s hows t hat ~( u ) can be i nt er pr et ed as t he expect ed pr esent val ue of a
pa yme nt of 1 that is made at t he t i me of r ecover y, i f ruin t akes pl ace.
F o r a _< u < b ,
Pr[Ta < ' ~ ] U( 0) = u] < 1 (9. 11)
and
Pr[Tb < ,,o I U( 0) = u] = 1 (9. 12)
becaus e t he sur pl us pr ocess {U(t)} has a posi t i ve drift. We def i ne t he s t oppi ng t i me
Ta,b = mi n(Ta, Tb), (9. 13)
18 3
and c ons i de r t he f unct i ons
A( a, b[ u) = E[e-~Y,.b I(U(Ta, b) = a) I U(0) = u]
= E[ e -;6T. I(Ta < Tb)[ U( 0) = u],
and
(9. 14)
B(a, b[ u) = E[e4T,.b I(U(Ta, b) = b)[ U( 0) = u]
= E[ e ~'Tb I(Ta > Tb) l U( 0) = u]. (9. 15)
Wi t h 8 i nt er pr et ed as a f or ce of i nt erest , A(a, b [ u) is t he expect ed pr esent val ue of a
pa yme nt of 1 whi ch is made when t he surpl us upcr osses t he l evel a f or t he first t i me,
pr ovi de d that t he sur pl us has not r eached the l evel b in t he meant i me. Si mi l ar l y,
B(a, b [ u) is t he expect ed pr esent val ue of a payment of 1 whi ch is made when t he sur pl us
r eaches t he l evel b f or t he first t i me, pr ovi ded that t he surpl us has not dr opped bel ow the
l evel a in t he meant i me. Not e that, f or each const ant k,
A( a, b[ u) = A( a+k, b+k[ u+k) (9. 16)
and
B(a, b [ u) = B(a+k, b+k [ u+k). (9. 17)
It f ol l ows f r om (9. 10) that, for u _> a,
A(a, oo[ u) = lira A( a , b l u)
b --q. ~
= l i m A( O, b - a [ u - a )
b ---) ~
= E[e ~ v . I(To < oo)] U(O) = u - a]
= ~g(u - a). (9. 18)
Si mi l ar l y, it f ol l ows f r om (9. 12) and (8.7) that, for u < b,
B(-oo, b [ u ) = l i m B( a , b l u)
a - - q, ~
= e~( b- u) . (9. 19)
Not e that, wi t h ~i = 0 and 0 <_ u < b, A(0, b ] u) is t he pr obabi l i t y of ruin f r om an initial
sur pl us u in t he pr es ence of an absor bi ng upper bar r i er at b.
For a' < a < u < b < b' , by cons i der i ng whet her T a < T b or T a > T b, we obt ai n t he
i dent i t i es
184
A( a, b' l u) = A( a, b l u) + B ( a, b l u) A( a, b' l b) ( 9 . 20 )
and
B(a' , b l u ) = A(a, b[ u)B(a' , b I a) + B(a, b[ u). (9. 21)
Wi t h a = 0, b = x, b' = o~ and because of (9. 18), (9. 20) be c ome s
~( u) = A(0, x] u) + B(0, x I u) ~( x) . (9. 22)
Wi t h a' = ~,~, a = 0, b = x, b' = oo and because of (9.19), (9. 21) be c ome s
e-P(x-u) = A(0, x [ u)e-P x + B(0, x I u). (9. 23)
For 0 < u < x, f or mul as (9. 22) and (9. 23) are t wo l i near equat i ons for A( 0, x ] u) and
B(0, x I u); t hei r sol ut i on is
A(0, x[ u) =
and
ePX~(u) - epU~(x)
epX _ ~( x)
(9. 24)
B ( 0 , x l u ) - e0 U _ W( u)
epX _ W(x)" (9. 25)
Wi t h 8 = 0, Seger dahl [361 denot es A( 0, x ] u) and B(0, x] u) as ~(u, x) and X(U, x),
r es pect i vel y. For mul as (9. 24) and (9. 25) ext end Di c ks on' s [7] f or mul as (1. 3) and (1. 4) to
t he gener al case of fi > 0.
To pr ove (9.5) let 0 _< u < x. I f ruin shoul d occur wi t h U( 0) = 0 such t hat t he sur pl us
i mme di a t e l y bef or e ruin is x, t hen t he surpl us must attain the l evel u pr i or to ruin. He n c e
f(x[ 0) = B(0, u I 0) f(x] u), (9. 26)
o r
1
f(x[ u) = f(xl 0) B(0, u I 0)
mo p' ; - ~(u)
= f ( x l - " l - ~( 0) ' (9. 27)
whi ch is (9. 5).
For mul a (9. 6) is mor e i nt ri cat e because t he condi t i on U( 0) = u _> x = U( T- ) > 0
means t hat t he surpl us is to dr op bel ow t he l evel x s ome t i me bef or e ruin occur s. Its p r o o f
is based on t he not i on of duality, whi ch as poi nt ed out by Fel l er [17, p. 395] enabl es one
"t o prove in an el ement ar y way t heor ems that woul d ot herwi se requi re deep anal yt i c
met hods . " We shal l der i ve the i dent i t y
B(0, u[ 0) f ( x] u ) f ~ p(__x P(x) + Y)e-Or+ d y = g ( x ) A( - u , 01 - x) e-flu, (9. 28)
val i d for 0 < x < u. Sol vi ng for f(x I u) and us i ng (4.4) and (9. 16) we get
f(x] u) = 811 - P( x) l e- P u
A(0,
u
l u x)
B(0, u 10) (9. 29)
Appl yi ng (9. 24) and (9. 25) to (9. 29) yi el ds
f(x[ u) = ~[ 1 - P(x)l
~t/(u
x) e
9X~(u)
1 -- ~( 0) (9. 30)
whi ch is i ndeed (9.6).
To prove the i dent i t y (9. 28), we mul t i pl y it by dx; the expr essi on on the l ef t - hand
side can be i nt erpret ed as
E[e -;6T,, I(Tu < T < oo x < U( T- ) < x+dx) I U(0) = 0], (9.31 )
whi l e t he expr essi on on the r i ght - hand side is
E[e-ST. I ( - x - d x _< U( T) < - x , rain U(t ) < u) I U(0) = 0]. (9. 32)
T<t<T~
The equal i t y of (9. 31) and (9. 32) can be expl ai ned by dual i t y. A dual process {U*(t)} of
t he pr ocess {U(t)} with U(0) = 0 is def i ned as fol l ows: I f T = ,:,o, we set U*(t ) = U(t), and i f
T < oo, we set
~
- U( T o-- t) for 0 <_ t <_ T o
U*(t ) = (9. 33)
U(t ) for t > T 0
See Fi gur es 3. a and 3.b. In ot her words, suppose that {U(t)} has n j umps before t i me TO,
and that t he j u mp of size Xi occur s at t i me ti, ti < To, i = 1 . . . . . n; t hen {U*(t)} has the
s ame n j u mp s bef or e t i me TO, except that the j u mp of size Xi occur s at t i me TO - ti, i = 1,
. . . . n. Thi s is a meas ur e- pr es er vi ng cor r espondence, and hence the process {U*(t)}
f ol l ows the same pr obabi l i t y law as the process {U(t)}. That is, i f a cert ai n event in t erms
of {U(t)} is t r ansl at ed as an event that is f or mul at ed in t erms of {U*(t )}, the pr obabi l i t i es,
or, as in the case of (9. 31) and (9. 32), the cont i ngent expect at i ons, are i dent i cal .
18 6
l U(t)
u !_ .......... ~ /
-31 ..........
Figure 3.a. A sample path of the process {U(t)} which contributes to (9.31)
0
- - X
- I A
i i i i i i i i . . . . . . .
U*(t)
Incidentally, this duality also explains the symmetric formula (4.7), which is for the case
8 = 0 .
Figure 3.b. The dual sample path which contributes to (9.32)
187
Us i ng (2. 43) we obt ai n f r om (9. 5) and (9. 6) t he f or mul a
f ( x, y I O) epu - ~( u)
f(x, Y l u) -- ( / I - v ( o ) '
\
f(x, y I 0) ePX~(u - x) - ~( u)
I - v ( o )
wi t h
f(x, Y l O) = ~e - pxp( x + y)
a c c or di ng t o (4. 3).
x > u >_ O, ( 9. 34)
O< x < u , (9. 35)
(9. 36)
E x a mp l e One c ons e que nc e of (9. 34) and (9. 35) is t hat t her e is an expl i ci t f or mul a f or
f ( x, y[ u) whe ne ve r t her e is an expl i ci t expr es s i on f or t he f unct i on ~t(u). Thi s is t he case
f or an exponent i al c l a i m amount di st r i but i on,
p( x) = 13e-13x,
Her e we have, f or u > 0,
~(u)
x>_O.
= E[e~ST + pU(T) I ( T < ~ ) [ U( 0) = u]
[~ - R e_Ru
= ~ + p
(9. 37)
a c c or di ng to (8. 37) [ wi t h w( y) = e-PY]. Then
Ilt(u - x) = eRx~( u) . (9. 38)
Hence, by (9. 5) and (9. 6) we obt ai n
2, e-fp + 13)x [([3 + p)e 0u _ ([3 - R) e-RU], x > u > 0, (9. 39)
I c( R + p)
f ( x I u) = M[3 - R) e-(P + fi)x [e(R + p)x _ 1] e -Ru, 0 < x -< u. (9. 40)
c( R + p)
To de t e r mi ne f(x, Y l u) we a ppl y (4. 19). We ma y use t he f or mul a
o f ( xl u) dx = E[ e ~ST I(T < oo) I U( 0) = u]
as a c he c k f or t he val i di t y of (9. 39) and (9. 40). Af t er s ome cal cul at i on t he l ef t - hand si de
s i mpl i f i e s as
)" e -Ru, (9.41 )
c(--ff7-
1 8 8
whi l e t he r i ght - hand si de is
- ~ - - ~ e -Ru ( 9. 42)
by (8. 33). Thes e t wo t er ms are t he s ame becaus e of (8. 35). It is amus i ng t o not e t hat t he
i nt egr al of e xpr e s s i on (9. 40) f r om x = 0 t o x = ,,o is al so (9. 41). IIII
Re ma r k s (i) Wi t h w( x, y) = e-PY,
h( u ) = / u e -P( z-u ) g( z) dz ( 9. 43 )
by (2. 34), (2. 16) and (4. 4). It f ol l ows f r om (2. 36) [wi t h ~( u) = gt(u)] t hat
~l/(u) = ( gt * g) ( u) + h( u)
U
I0 gt(u - z) g( z) dz + /ue- P( z- u) g(z) dz, ( 9. 44)
whi ch gener al i zes Exer ci s e 12.11 of Actuarial Mathematics [4]. Wi t h u = 0, ( 9. 44)
b e c o me s
i,
~( 0) = ]0 e - PZg ( z ) d z = ~( p) ; ( 9. 45)
r ecal l t hat ~( - R) = 1.
(i i ) As an al t er nat i ve pr oof , we woul d l i ke t o s how t hat (9. 5) and (9. 6) sat i sf y t he r enewal
equat i on (5. 13), or equi val ent l y, wi t h t he def i ni t i on
eP u, x >u _>0 ,
q0(u) = ( 9. 46)
\ ePX~l/(u - x), 0 < x _< u,
t hat
q~(u) - ~( u) = [ ( {p- Ig)* g](u) + [t - q/(0)] eP u I(x > u) ( 9. 47)
hol ds. A di r ect ver i f i cat i on of (9. 47) s eems di f f i cul t . However , we can conf i r m its
val i di t y by means of Lapl ace t r ansf or ms. Taki ng Lapl ace t r ans f or ms of (9. 44) yi el ds
~ ( ~ ) = ~ ( ~ ), ~ ( ~ ) + ' ~ ( ~ ), ( 9. 48)
wher e
~(~)
= , - r , ~ j oe+L j e_p(z_0)g(z) d z ] d u
~ ( ~ ) - f ~ ( p )
( 9. 49)
189
by c ha ngi ng the or der of i nt egr at i on. Hence
~ ( ~ ) - f i ( ~ )
i - ~ ( ~ )
~(~) - ~( p)
[ l -~(~)](p-~)
( 9. 50)
F r o m( 9 . 4 6 )
e(p -,~)x _ l
~)(~) - P _ ~ + e ( P - ~ ) x ~ ( ~ ) . ( 9. 51)
Thus
e(O-~)x _ I I - ~( p)
,(~) - ~ ( ~ ) -
p - ~ l - ~(~)
e (p-~)X - I 1 - ~( 0)
p - ~ l - ~ ( ~ )
( 9. 52)
by (9. 45). We now see t hat (9. 47) hol ds.
(i i i ) As we poi nt e d out ear l i er , it f ol l ows f r om our gener al i zat i on of Di c k s o n ' s f or mul a t hat
t her e is an expl i ci t f or mul a f or f(x I u) [and hence f i x, Y l u)] whe ne ve r t here is an expl i ci t
e xpr e s s i on f or t he f unct i on ~( u) . If u~(~) is a r at i onal f unct i on, t hen, by l ocat i ng its pol es
( s i ngul ar i t i es ) , we can det er mi ne ~( u) . It f ol l ows f r om (9. 50) t hat ~( ~) is a r at i onal
f unct i on i f and onl y i f ~(~) is a r at i onal f unct i on; by (2. 54) ~(~) is a r at i onal f unct i on i f and
onl y i f 0(~) is a r at i onal f unct i on. It al so f ol l ows f r om (9. 50) t hat t he s i ngul ar i t i es of ~ ( ~ )
ar e e xa c t l y t he r oot s of t he equat i on
~ ( ~ ) - - 1. (9. 53)
We s houl d cl ar i f y t hat her e t he f unct i ons ~( ~) , ~(~) and 0(~) are def i ned on the whol e
c o mp l e x pl ane by anal yt i c cont i nuat i on. Cons i de r the e xa mpl e wher e p( x) = [3e-13x;
al t hough t he i nt egr al
/oe-gX p( x) dx
is not de f i ne d f or c ompl e x number s ~ wi t h Re(G ) _< --[3, the r at i onal f unct i on ~1
~ i s .
Cons e que nt l y, whi l e (2. 57) has at mos t one sol ut i on, (9. 53) can have mul t i pl e sol ut i ons.
Now, l et - r l , - r 2 . . . . . - r m be the di st i nct r oot s of (9. 53) and n i, n2 . . . . . nm be t hei r
190
mul t i pl i ci t i es . Then it f ol l ows f r om Heaviside's expansion formula ( Spi egel [37, p. 73] )
t hat
dnk- I
1 l_~im_rk ~ [ (~ + rk) nk i~/(~) e~U], ( 9. 54) ~( u)
k =l ( n k - 1 ) ! ~,
wher e qt(~,) is gi ven by (9. 50). In t he speci al cas e wher e all pol es of ~( ~) are s i mpl e, i. e. ,
n I = n 2 . . . . . n m = 1, t hen (9. 54) s i mpl i f i es as
= ~-I"-~1 f i ( - r k ) e _r k u
= - ~ )
k~a'l~ g( - r k) -- ~ ( p ) e - r k u . ( 9 . 5 5 )
=
By ( 9. 53) and (9. 45),
Si mi l a r t o (2. 59), we have
f~(-rk) - 2( P) = I - ~I/(0).
' l
- ~' ( - r k ) ( p + rk) = - ~ P ' ( - r k ) - 1.
He nc e ( 9. 55) s i mpl i f i es as
e --l"kH
i l l
v ( u ) = [ l - v ( O ) l k ~ ' = - ~ - ' ( - r k ) ~ P - l '
P u t t i n g u = 0 i n ( 9 . 5 6 ) a n d r e a r r a n g i n g , w e o b t a i n
I
1 - W( 0 ) =
1
+
k= l _~ o , ( _r k ) _ 1
whi ch can be s ubs t i t ut ed in (9. 56) yi el di ng
c ~ e-rk u
qt(u) = x ~ C k= 1 -)LP' (-rk) - c
I + k -z '- -~, -'( ---P -rk)
= |
c
Cons i de r t he cas e wher e p( x) is a mi xt ur e of exponent i al di st r i but i ons,
wher e
p( x) = --~1Aj ~j e-I~jx, x _> 0,
j =
o < ~ l < ~ : < ... < ~ n ,
( 9. 56)
( 9. 57)
( 9. 58)
( 9. 59)
191
and
Then
__IAJ
= 1.
j=
~ ( ~ ) = _ _ ~ ] ~ , (9.60)
J
and Lundber g' s fundamental equation (2.21) becomes
a j [ 3j
8 + ) v - c ~ = Kj__~ 1 ~j - - - f ~. (9.61)
The nonnegat i ve solution of (9.61) is p and the negative solutions are the poles of ~t(~).
We now impose the condition that Aj > 0, j = 1, 2 . . . . . n. Then (9.61) has n distinct
negative roots {-rk} with
O< r l = R < ~ l < r 2 < ~ 2 < ... < r n < ~ n .
(Inequalities (12.6.15) of Actuarial Mathematics [4] are for the case 8 = 0; see also Figure
12.7 of Actuarial Mathematics [4].) It follows from
--~1 Aj lSj
P'(~) = - i = ( 13j + ~ ) 2
(9.62)
and (9.58) that, given the roots {-rk}, we have an explicit formula for ~(u) [and hence
explicit formulas for both f(x I u) and f(x, y I u)]. On the other hand, by (2.54) and (9.60),
~ ( ~ ) = c o -~ _ ~ [ P( ~ ) - PI P)]
- j= l (~j + ~)([~j + P) ' (9.63)
from which we obtain
- - _ Aj
p - ~ j= (~j + ~)([~j + p/2
and
(9.64)
--~1 Aj ~ j
C ' ~ j= ( J3 j + ~ )2( ~ j + p)
(9.65)
It follows from (9.55), (9.64) and (9.65) that we have the following alternative formula for
u/(u),
192
Aj6 j
- - - rk) ( j +
p) 2
= Aj ~j
J ~ ( 13j r0 2( 13j + p )
Not e t hat in t he speci al case n = 1 we obt ai n agai n (9. 37).
( i v) Subs t i t ut i ng t he as ympt ot i c expr es s i on of ~( u) ,
tl/(u ) '~' Ce-Ru
m ( 9. 35) yi el ds
f(x, Y l u)
Be c a us e
C
e - - I ' k u .
C [ePXe_R(u_x ) _ e_RU]
f(x, y[ 0) 1 - ~?(0)
c [ l ~C~( 0) ] (eRx - e-pX) p( x + y) e -Ru
( 9. 66)
f or u --~ oo, ( 9. 67)
f or u ---) oo. (9. 68)
f~ ( - R) - f~ ( p )
- f f ( - R ) ( p + R )
1 - ~ ( 0 )
- [ ~' (- R )(p + R)
1 - ~ ( 0 )
- , ( 9. 69)
- ~ ' ( - R) - 1
t he a s ympt ot i c f or mul as (9. 68) and (7. 16) are t he same.
(v) The e xpr e s s i on
E[ v T a t e =, ~ I ( T < ~ ) l U(O) = u] ( 9. 70)
is t he e xpe c t e d pr esent val ue of a cont i nuous annui t y at a r at e of 1 per uni t t i me bet ween
t he t i me of rui n and t he t i me of r ecover y f or a gi ven i ni t i al sur pl us u. Be c a us e
( 9. 70) is
v T a,r-ffT-,ll = ~ ( e ~ST - e-81~,,), ( 9. 71)
[J0r~ f ( x I u) dx - ~( u) ] . ( 9. 72)
We not e t hat
~i m E[ v T a ~ l i T < oo) 1 U( 0) = u] = El ( T0 - T) l i T < oo) I U( 0) = u]. ( 9. 73)
Al t e r na t i ve l y, (9. 70) is
193
E[ vT 1 - eP U(T)
be c a us e of (9. 9). Thi s yi el ds
E[ ( T0 - T) I ( T < ,,o) I U( 0) = u]
I ( T < oo) I U( 0) = u] (9. 74)
= ~i m ~E[ I U( T) I I ( T < oo) I U(O) = u]
- c - l k p / E[ I U( T) I I ( T < ~ ) l U( 0) = u] (9. 75)
by (2. 48). For mul a (9. 75) is i nt ui t i vel y cl ear becaus e c - Kpl is t he dri ft of {U(t)}. For
r el at ed r esul t s see Egf di o dos Rei s [16].
(%) Recal l t he f unct i on 7t3(x, t I u), t he pas s age t i me densi t y of the sur pl us pr oces s at t he
l evel x, an expl i ci t f or mul a f or whi ch is gi ven by (8. 13) and (8. 14). Si mi l ar l y to (8. 31),
we have, f or x <_ u and 5 >- 0,
f o ~ 3 ( x , ~ l u ) e~SZd~ = / l t ( x - u ) [ l + f o r t 3 ( x , ' t ] x ) e - & t d ~ ]
- u ) [ I + f ~ n 3 ( O, ' c I O) e- ~Xdx] . ~( x
d O
(9. 76)
Put t i ng x = u = 0 in (9. 76) and sol vi ng f or t he i nt egral yi el ds
_ ~ ( o )
f ) ~3(0, x I 0) e -5"t d' r 1 - ~ ( 0 ) '
(9. 77)
Appl yi ng (9. 77) t o (8. 31) and (9. 76), we obt ai n
e- p( x- u)
I I - ~ ( 0 ) ' x > u
f ~" x3( x, x I u) e -&t d"c =
~( u - x)
x < u
1 - v ( O) '
(9. 78)
The r i ght - hand si de of (9. 78) can be wr i t t en as a pai r of i nfi ni t e seri es usi ng t he ge ome t r i c
ser i es f or mul a
1 _ n__~ 0 ilt(0)n;
l - ~ ( 0 )
t he j - t h t er m of ei t her ser i es r epr esent s t he cont r i but i on of the j t h upcr os s i ng at t he l evel x
t o t he i nt egral .
194
10. Optimal Div idends
We now consi der a pr obl em that is due to Br uno de Fi net t i , has been t reat ed by
Kar l Bor ch and ot hers, and can be f ound in t he t ext books of Bi i hl mann [5, Sect i on 6.4]
and Ger ber [ 19, Sect i on 10.1 ]. Here the sur pl us model is modi f i ed in that di vi dends are
pai d to t he sharehol ders of the i nsur ance company. We assume that the di vi dends are pai d
accor di ng to a barrier strategy cor r espondi ng to a barri er at the level b. Thus whenever
t he sur pl us is on t he barri er b, di vi dends are pai d cont i nuous l y, at a rate of c so t hat t he
sur pl us st ays on the barri er, unt i l the next cl ai m occurs and the sur pl us falls bel ow b. If
the sur pl us is bel ow b, no di vi dends are bei ng paid. Evi dent l y, rui n wi l l occur wi t h
cer t ai nt y in this model . For 0 _< u _< b, let V(u, b) denot e the expect ed pr esent val ue of t he
di vi dend payment s unt i l rui n.
Si nce no di vi dends are paid, unl ess the sur pl us reaches the l evel b before r ui n
occurs, we have, for 0 <_ u _< b,
V(u, b) = B(0, bl u) V( b, b), ( t 0. 1)
or, by (9. 25)
V(u, b) - epu - u/(u) V(b, b). (10. 2)
e ob ~( b)
To det er mi ne V(b, b), we need a boundar y condi t i on at u = b. To obt ai n it, we compar e
t wo si t uat i ons, one wi t h i ni t i al sur pl us b, and the ot her wi t h i ni t i al sur pl us u = b - e, E > 0.
If E is suf f i ci ent l y smal l , we can be "al most sure" that in the second si t uat i on the sur pl us
wi l l reach the barri er before a cl ai m occurs; hence the di vi dends pai d in the second case
wi l l be "es s ent i al l y" the di vi dends pai d in the first case reduced by ~:, f r om whi ch it
f ol l ows t hat
V ( b - e, b) - - V(b, b) - ~. ( 10. 3)
In the l i mi t , t hi s means that
--' o3_~_V [ = 1 . ( l 0. 4)
OU I
u = b
Di f f er ent i at i ng (10. 2) wi t h respect to u and t hen set t i ng u = b yi el ds
195
He nc e
1 - pepb - gt' (b) V( b, b).
eP b tlt(b )
eP u - ~( u) 0 < u < b . ( 10. 5)
V(u, b) - pePb _ ~ ' ( b ) '
Thi s f or mul a s houl d be c ompa r e d wi t h ( 1. t 3) in Chapt er 10 of Ge r be r [ 19]. In Sect i on
10.1 of [19], t he f unct i on B( 0, b[ u) is denot ed as W( u, b).
Let b be t he opt i mal bar r i er , i.e., t he val ue of b t hat ma xi mi z e s t he e xpe c t e d
pr es ent val ue of t he di vi dends . In vi ew of (10. 5), b is t he val ue t hat mi ni mi z e s t he
de nomi na t or , i. e. , b sat i sf i es
pZePb - t g"(b) = 0. ( 10. 6)
An e qui va l e nt condi t i on is t hat
02V
~ U2 u = b = 0; ( 10. 7)
t hi s f ol l ows f r om t he e xpl i c i t f or m of (10. 5).
Ex ample In t he cas e of an exponent i al c l a i m amount di st r i but i on, t here is an expl i ci t
e xpr e s s i on f or ~( u) . Subs t i t ut i ng (9. 37) i nt o (10. 5) yi el ds
V ( u , b ) = ( [ 3+p) e P u - ( [ 3 - R) e -Ru (10. 8)
p([3 + p) e pb + R([3 - R)e -Rb"
The opt i ma l val ue b is obt a i ne d f r om t he condi t i on t hat
p2([3 + p)eP b - R2(~ - R)e -Rb = 0. ( 10, 9)
Thus
1 , R2( [ 3- R)
- p + R , n p - ~ + ~ ( 10 . 10 )
is t he opt i ma l bar r i er . Illl
196
11. Concl udi ng Re ma r ks
This paper studies the joint distribution of the time of ruin, the surplus
immediately before ruin, and the deficit at ruin. Motivated by the problem of pricing
American options on stocks with jumps, we incorporate the time of ruin in the classical
model by discounting. New results are derived, many of which have a probabilistic
interpretation, and additional insight is gained for existing results in the classical model.
Our next goal is to treat the option pricing problem mentioned in the Introduction.
The results presented can be generalized in various directions. For example,
several formulas can be extended to the case where the compound Poisson process is
replaced by a more general process with positive, independent and stationary increments,
such as the gamma process or the inverse Gaussian process. In the references there are
some recently published related papers.
Acknowledgment
Elias Shiu gratefully acknowledges the support from the Principal Financial
Group Foundation.
References
[ 1 ] Arfwedson, G. "Research in Collective Risk Theory. Part I," Skandi navi sk
Akt uari et i dskri f t 37 ( 1954): 191-223.
[2] Arfwedson, G. "Research in Collective Risk Theory. Part II, " Skandi navi sk
Akt uari et i dskri f t 38 ( 1955): 53-100.
[3] Beekman, J.A. Two Stochastic Processes. Stockholm: AImqvist & Wiksell, 1974.
[4] Bowers, N.L., Jr., Gerber, H.U., Hickman, J.C., Jones, D.A., and Nesbitt, C.J.
Act uari al Mathematics. Itasca, Ill.: Society of Actuaries, 1986.
[5] Btihlmann, H. Mat hemat i cal Met hods in Ri sk Theory. New York: Springer-Verlag,
1970.
[6] Cox, D.R., and Miller, H.D. The Theory of Stochastic Processes. London: Metbuen,
1965.
197
[7] Dickson, D.C.M. "On the Distribution of Surplus prior to Ruin," &surance."
Mathematics and Economics I 1 (1992): 191-207.
[8] Dickson, D.C.M. "On the Distribution of the Claim Causing Ruin," hlsurance:
Mathematics and Econ omics 12 ( [ 993): 143-154.
[9] Dickson, D.C.M., and Egldio dos Reis, A.D. "Ruin Problems and Dual Events,"
Insurance: Mathematics and Economics 14 ( 1994): 51-60.
[ 10] Dickson, D.C.M., Egidio dos Reis, A.D., and Waters, HR. "Some Stable Algorithms
in Ruin Theory and Their Applications," ASTINBul l et i n 25 (1995): 153-175.
[I 1] Dickson, D.C.M., and Waters, H.R. "The Probability and Severity of Ruin in Finite
and Infinite Time, " ASTIN Bulletin 22 ( 1992): 177-190.
[ ] 2] Dinges, H., "Eine kombinatorische lJberlegung und ihre mal3lheoretiscbe
Erweiterung, " Zeitschrift fiir Wahrscheinlichkeitstheorie und verwandte Gebiete 1
(1963): 278-287.
[ 13] Dufresne, F., and Gerber, H.U. "The Surpluses Immediately before and at Ruin, and
the Amount of the Claim Causing Ruin," Insurance: Mathematics and Economics 7
(1988): 193-199.
[14] Dufresne, F., and Gerber, H.U. "The Probability of Ruin for the Inverse Gaussian and
Related Processes, " Insurance: Mathematics and Economics 12 (1993): 9-22.
[15] Dufresne, F., Gerber H.U., and Shiu, E.S.W. "Risk Theory with the Gamma Process,"
ASTI NBul l et i n 21 (1991): [77-192.
[ 16] Egfdio dos Reis, A.D. "How Long is the Surplus Below Zero'?" Insurance:
Mathematics and Economics 12 (1993): 23-38.
[ 17] Feller, W. An Introduction to Probability Theor3 ' and Its Applications, Vol. 2, 2nd ed.
New York: Wiley, 197 I.
[18] Frey, A., and Schmidt, V. "Taylor-Series Expansion for Multivariate Characteristics
of Classical Risk Processes," Insurance: Mathematics attd Economics 18 ( 1996): I-12.
[ 19] Gerber, H.U. An Introduction to Mathematical Risk Theoo,. S.S. Huebner Foundation
Monograph Series No. 8. Homewood, Ill.: Irwin, 1979.
[20] Gerber, H.U. "Mathematical Fun with Ruin Theory, " Insurance: Mathematics and
Economics 7 (1988): 15-23.
[21] Gerber, H.U. "When Does the Surplus Reach a Given Target'?" Insurance:
Mathematics and Economics 9 (1990): 115-119.
[22] Gerber, H.U., Goovaerts, M., and Kaas, R. "On the Probability and Severity of Ruin,"
ASTI NBul l et i n 17 (1987): 151-163.
[23] Gerber H.U., and Shiu, E.S.W. "Option Pricing by Esscher Transforms, "
Transactions, Socie~., of Actuaries 46 ( 1994): 99-140; Discussion 141 - 191.
198
[24] Keilson, J. "The First Passage Time Density for Homogeneous Skip-Free Walks on
the Continuum," Annals of Mathematical Statistics 34 (1963): 1003-1011.
[25] Kendall, D.G. "Some Problems in the Theory of Dams," Journal of the Royal
Statistical Society Series B 19 (1957): 207-212.
[26] Lundberg, F. "Some Supplementary Researches on the Collective Risk Theory,"
Skandinavisk Aktuarietidskrift 15 ( 1932): 137-158.
[27] Panjer, H.H., and Willmot, G.E. Insurance Risk Models. Schaumburg, Ill.: Society of
Actuaries, 1992.
[28] Picard, P. "On Some Measures of the Severity of Ruin in the Classical Poisson
Model," h~surance : Mathematics and Economics 14 (1994): 107-115.
[29] Picard, P., and Lef6vre, C. "On the First Crossing of the Surplus Process with a Given
Upper Barrier," Insurance: Mathematics and Economics 14 (1994): 163-179.
[30] Prabhu, N.U. "On the Ruin Problem of Collective Risk Theory," Annals of
Mathematical Statistics 32 ( 1961 ): 757-764.
[31 ] Prabhu, N.U. Stochastic Storage Processes: Queues, Insurance Risk, and Dams. New
York: Springer-Verlag, 1980.
[32] Resnick, S.1. Adventures in Stochastic Processes. Boston: Birkh~user, 1992.
[33] Seah, E.S. "Computing the Probability of Eventual Ruin," Transactions, Society of
Actuaries, 42 (1990), 421-431 ; Discussion: 433-446.
[34] Seal, H.L. Stochastic Theo~ of a Risk Business. New York: Wiley, 1969.
[35] Seal, H.L. Survival Probabilities: The Goal of Risk Theory,. New York: Wiley, 1978.
[36] Segerdahl, C.-O. "On Some Distributions in Time Connected with the Collective
Theory of Risk," Skandinavisk Aktuarietidskrift 53 ( 1970): 167-192.
[37] Spiegel, M.R. Schaum' s Outline of Theor) , and Problems of Laplace Transforms.
New York: Schaum, 1965.
[38] Tak~cs, L. Combinatorial Methods in the Theory. of Stochastic Processes. New York:
Wiley, 1967. Reprinted by Krieger, Huntington, N.Y., 1977.
199

Potrebbero piacerti anche