Entrega IV Premio “Jóvenes Investigadores Joan Prats”
A.05 Gestión del talento en la administración pública
We now wish to introduce long-range dependence into the source model and derive explicit estimates of boundsq xandQ xof Proposition 5.4.1 for such a case. As shown below, a natural way to do this is to assume a heavy tail for the distribution of either the silence period, the activity period, or both. In the rest of this section, we ®rst specify the LRD behavior of the rate process associated with heavy-tailed distributions for on=off durations, then provide the estimates for bounds q xand
Q x in such a case. We conclude this section by providing a weak convergence theorem for the limiting input process fOtg obtained in Section 5.3, which, once rescaled in space and time, converges toward the so-called fractional Brownian motion.
5.5.1 LRD Properties
Recall [6] that a stationary processfltg is said to exhibit short- (resp. long-) range dependence if its correlation functionrde®ned in Eq. (5.6) is such that the integral
1
0 r tdtis convergent (resp. divergent). If the distributions of on and off periods
A and B are Coxian, that is, their Laplace transforms a* and b* are rational functions, then standard results for Laplace transform inversion applied to expres- sion (5.7) entail that the correlation functiont7 !r ttends exponentially fast to 0 ast" 1, thus corresponding to short-term dependence for processfltg. Now, the situation appears totally different in the case when the distribution of either off durationAor on durationBis heavy tailed. A distribution functionFis here de®ned to be heavy tailed if
1 F t h ttr 5:28
fort>0 and 1<r<2, where functionhhas a ®nite limith at in®nity. A typical example is the Pareto distribution, where
F t 1 tyy
r
for some y>0. As a consequence of de®nition (5.28), the Laplace transform
f*:p>07 !01e pt dF tof a heavy-tailed distributionFwith ®nite meanmand powerr2 1;2veri®es
f* p 1 mphGr 21 rpro pr 5:29 for small positivep, whereGis Euler's function (see Section 5.7.3). For 1<r<2, such an expression for the Laplace transform near 0 in fact characterizes the heavy- tailed behavior of the corresponding distribution F (in the case when r2, Eq. (5.29) is identical to a classical Taylor expansion, corresponding to a regular distribution having a ®nite variance). We can then show the following.
Proposition 5.5.1. Assume that the Laplace transforms a*and b*of the off and on periods have expansions
a* p 1 p
aarpro pr;
respectively, withpositive coef®cients ar and bs and where powers r;s21;2are suchthat rs64. The covariance function of ratefltg then veri®es
r t r0
tq 1 5:31
for large t, withqmin r;sand
r0 bn 3a r G 2 r resp:r0 a 1 n3bs G 2 s ! if r<s(resp. s<r)and r0bn3ara 1 n3bs G 2 r if rs.
In other terms, assuming such heavy-tailed distributions for A and B with
r;s2 1;2 entails long-range dependence for process fltg since 01r tdt
diverges for 0<q 11. The proof of Proposition 5.5.1 can be performed simply by inserting expansions (5.30) for a* and b* into expression (5.7) and then using Tauberian Theorem 5.7.1 recalled in Section 5.7.3 for deriving the asymptotics ofr tfor larget. Concerning input processfotgwithot0tludu, it is easy to express its varianceD2 tde®ned in Eq. (5.3) as
D2 t 2 t
0 t ur udu 5:32 in terms of the covariance functionrof processfltg. As a consequence, it readily follows that if
1
0 r udu<1;
then D2 t O t as t" 1. On the other hand, if behavior (5.31) holds with 1<q<2, we have instead
D2 t lt3 q 5:33 ast" 1, where
l 2r0
3 q 2 q 5:34
r0 being expressed as in Proposition 5.5.1.
5.5.2 Impact on Queue Content Distribution
To make boundsq x and Q xof Proposition 5.4.1 explicit in the case of heavy- tailed distributions for on and off periods, we ®rst need some preliminary estimation of intermediate quantitiesm tandD tde®ned in expressions (5.21).
Lemma 5.5.2. Assume that expansions (5.30) hold for the Laplace transforms a*
and b*. Then quantitym tde®ned in Eq. (5.21) veri®es
m t 3 2 q
l
s2t2 q 5:35
for large t, withqmin r;s and constants s2 and l de®ned in Eqs. (5.2) and
(5.34), respectively.
The proof is deferred to Section 5.7.4. Note that the average differencem t of input due to different initial conditionsl00 orl01 grows with increasing t rather than tending to a constant, an eloquent testimony to long-range dependence. It further follows from estimate (5.33) and the latter result that the variance D2 t
introduced in Eq. (5.21) satis®es
D2 t D2 t m2 ts2D2 t 5:36 for largetsince 2 2 q<3 q, whenq>1. These observations lead to the next proposition, the main result in this chapter.
Proposition 5.5.3. Assume that the distributions of either on or off periods have heavy tails with respective power r;s;2 1;2and that expansions (5.30) hold. Then limiting distributionx7 !h xof Proposition 5.4.1 has the logarithmic estimate
lim x"1 1 x2 1 Hlogh x k2 2 5:37 where H 3 q=2; k 1 HH 1 H1 H gH l p and withlspeci®ed in Eq. (5.34)
Proof. First, consider lower bound (5.22). As expansions (5.30) hold, estimate (5.33) applies to thatD2 t lt2H for large t and 2H3 q. Use the variable change txu and write D mx pl uxH1d ux, where d t !0 as
t" 1. Lower bound (5.22) then reads
q x sup u>0 F x1H l p r u 1d ux ! ; 5:38
where
r u 1uHgu:
It is easily veri®ed that function u7 !r u has a unique minimum at point
u*H=g 1 H with r u* gH=HH 1 H1 H. Supremum (5.38) entails, in particular, thatq x q ~ xwith
~ q x F x1H l p r u* 1d u*x ! :
Using the asymptotics F z exp z2=2=zp2p for large z, we may write that logF z z2=2o z2. Since d u*x !0 as x" 1, lower bound q ~ x thus veri®es
logq ~ x 12 1x1d uH*xr u*
l p !2 o 1x1d uH*x !2 12 x1 Hr u* l p 2 o x1 H2: The latter consequently implies that
lim inf
x"1 1
x2 1 Hlogq x lim infx"1 1
x2 1 Hlogq ~ x
k22 5:39
withkr u*=pl:
Second, consider upper bound (5.23). Performing again the variable change
txuin the integral gives
Q x x1 H 2pl p 1 0 exp x2 1 H 2l Jx u ! Lx udu; 5:40 where Jx u r u 2 1d ux2; Lx u R ux;x uH1d ux;
and with functionsu7 !r uand u7 !d ux de®ned as in Eq. (5.38). To evaluate integral (5.40), we apply the Laplace method [12] for the estimation of integrals with
exponential integrand. For ®xedx, the continuous functionu7 !Jx utends to1
asu#0 andu" 1(in fact, it isO u 2Hfor small positiveuandO u2 1 Hfor largeu). It therefore has a minimum at some pointux*. AsJx u !r2 uasx" 1 for boundedu, it is clear thatux*!u* as x" 1, where u* is the minimum of
u7 !r2 u. The contribution of the exponential factor to the value of the integral is thus preponderant in the neighborhood of that minimum ux*. Concerning the nonexponential factor Lx u in Eq. (5.40), we note from Lemma 5.5.2 with
H 3 q=2 that
x 1gum uxs2 D ux2 !g
as x" 1 and u!u*. As D ux D ux for large x and ®xed u in view of estimate (5.36), we successively deduce from the above and expressions (5.21) that
S ux;x !0 and therefore R ux;x !s=p2p as x" 1 and u!u*. These limiting results entail thatLx u !LwithLs= u*H
2p
p
. FactorLx utending to
a nonzero limit, integral (5.40) is therefore, up to powers ofx, of order exp x2 1 H
2l Jx ux
!
: We therefore deduce from the above that
lim x"1 1 x2 1 HlogQ x xlim"1 Jx ux* 2l ; 5:41
but the latter simply equalsr u*2=2lk2=2 by continuity. The conjunction of (5.39) and (5.41) together with the boundsq x h x Q xprovide the desired
logarithmic estimate (5.37). j
The latter result consequently entails that the distribution of the content of a queue at heavy load, when fed by the superposition of an in®nite number of on=off sources with LRD characteristics, has a Weibullian tail at in®nity, namely,
h x exp k22x2 1 Hox2 1 H
5:42
for large x. As shown in the proof of Proposition 5.5.3, only the leading term
D2 t lt2H is used to derive the logarithmic estimate for h x. Re®ned asymptotics for bounds q x and Q x, and thus estimates for the ox2 1 H term in Eq. (5.42), can further be derived from theasymptotic expansionof the variance
D2 tfor larget; that is,
D2 t lt2Hl0t2H0
o t2H0
as t" 1, with H0<H. Using generalized Tauberian theorems [7, p. 142] and identify (5.32), such an expansion can be derived from that of the Laplace transform
p7 !2r* p=p2 of t7 !D2 t near p0 and the use of expansions (5.30) in formula (5.7) forr* p. Second-order powerH0proves, however, dif®cult to specify for arbitrary powersr and s in expansions (5.30). By way of illustration, we just mention without proof that ifs<minr; r1=2, then a second-order expansion (5.43) forD2 tcan be written with
2H3 s;l 3 2s a 21 snG3b 2s s; 2H04 2s;l0 a2 1 n3b2s
2 sG 4 2s:
Using Eq. (5.43) in formulas (5.22) and (5.23) then enables one to derive complete asymptotics forq xandQ xin the form
q x ek 0 kp2px1 Hexp k2 2 x2 1 H ; Q x sek 0 gp2p H 1 H r exp k22x2 1 H ; 5:44 respectively, where k0 l0 2H2l2g2:
We illustrate these approximations in the following numerical examples. Consider
N100 on=off sources, assuming that 100 sources is suf®cient for a useful application of our asymptotic results. We take the mean burst volume as unity
1=b1. The mean activity probabilityna= abis set to 0.1 a1
9and the multiplexer load Nn=C 1g=npN 1 to 0.9, implying C11:11 g1
9. Figure 5.1 shows a log plot of the complementary distribution function (interpreted as the over¯ow probability against buffer size) measured in units of mean burst volume, when sources have activity periods distributed as Pareto random variables. We also consider two values for H H0:8 and H 0:85). Curve A (B, resp.) corresponds to the limiting upper bound Q (lower bound q, resp.) given by Eq. (5.44). We observe that the buffer size required for a loss probability of 10 9 is around 3104 times the mean burst size forH 0:8 and more than 106times the mean burst size forH0:85. These values have to be compared to the case where both silent and activity periods are exponentially distributed. In the latter case, a buffer size of around 50 times the mean burst size guarantees a loss probability less than 10 9.
5.5.3 Convergence to the FBM
The limiting Gaussian processfOtgde®ned in Proposition (5.3.6) of Section 5.3 has been derived as the integral of limiting rate processfYtg, that is,
Ot
t
0Yu du; t0:
Now, introduce the family of Gaussian centered processesfZt ygde®ned by
Zt y 1
yHOyt; t0; 5:45 for all parameter values y>0 and with ®xed constant H. Process fZt yg is thus deduced fromfOtg by the time and space scaling change t;x 7 ! yt;xyH. Each processfYt Ngde®ned by Eq. (5.15) being the superposition of i.i.d. rate processes distributed as fltg, its covariance function is independent of N and is therefore identical, along with that of its limiting processfYtg, to the correlation functionrof the single rate processfltg. Similarly, in view of de®nition (5.18), the variance ofOt equals that of anyO tNand is identical toD2 t Var otde®ned in Eq. (5.3). The second-order properties of processesfZ tyg can then be readily derived as follows. Formula (5.32) ®rst entails that the variance ofZt y can be written as
Var Zt y 2y2 1 H
t
0dy
y
0ry y xdx: 5:46 Fig. 5.1 Over¯ow probability as a function of the buffer size with Pareto activity distribu- tion.
On the other hand, it is known that for any centered process fOtg with stationary increments, the covariance E OuOt can be expressed in terms of the variance functiont7 !D t2Var O
tonly as
E OuOt 1
2D t2D u2 D t u2 5:47 withut. The covariance structure and, therefore, the distribution of each Gaussian processfZt ygare thus entirely de®ned through formulas (5.46) and (5.47).
From expression (5.46), in particular, we deduce that the convergence of the family of processesfZt ygasy" 1toward some limiting process is governed by the behavior of the covariance functionrin the neighborhood of in®nity. In view of factory2 1 H in Eq. (5.46), it is then natural to expect that the variance Var Z y
t
and, therefore, the sequence of processesfZt yghave a nontrivial limit asy" 1in the case wherer tbehaves as a power of tfor larget. Such a situation has been exempli®ed in Proposition 5.5.1 of this section, where it is shown that the covariance functionrof rate process fltgis preciselyO t1 qfor larget and someq2 1;2, provided the distribution of either on or off periods is heavy tailed. A weak convergence result for the family of processesfZt yg,y>0, is therefore plausible in this framework.
Before stating the formal result, it is useful to recall here that [14, p. 318] there exists a unique centered Gaussian processfZtghaving
continuous paths withZ00,
stationary increments,
and with covariance function de®ned by
E ZuZt 1
2jtj2H juj2H jt uj2H 5:48 for allu;t2Rand for some parameterH 2 1
2;1.
Note that, inviewof the stationarity of increments offZtgand of general property (5.47), this de®nition is equivalent to stating thatE Z2
t jtj2Hfor allt2R. This processfZtg is known as the fractional Brownian motion (FBM) with Hurst parameterH.
Proposition 5.5.4. Assume that the covariance function t7 !r tof rate process fltgveri®es asymptotics (5.31) for large t and some power q2 1;2. The processes fZt yg de®ned in Eq. (5.45) with
H3 q
2
then converge weakly asy" 1toward processpl fZtg, wherefZtgis an FBM
withHurst parameter H andlis de®ned in Eq. (5.34).
Proof. To ensure the convergence of the sequence of centered Gaussian processesfZt yg, one could, for instance, rely on Theorem 5.3.2. Here we content
ourselves with the veri®cation of the ®rst half of this theorem's hypotheses, that is, convergence of the ®nite-dimensional distributions. The latter holds if the covariance functions converge toward the covariance function of some identi®ed centered Gaussian process. In the present case, using estimate (5.31) in (5.46), we readily obtain Var Zt y 2y2 1 H t 0dy y 0 r0 y y xq 1dx; 5:49
where the conditionq2 1;2ensures that the integral is ®nite. As
t 0dy y 0 dx y xq 1 t 0 y2 q 2 qdy t3 q 3 q 2 q;
the right-hand side of Eq. (5.49), withH 3 q=2, is consequently asymptotic to yq 1y2rq 01 3 q t3 2q qlt2H
asy" 1. Using general property (5.47) for processes with stationary increments, we then deduce that the covariance function offZt ygconverges to
1
2ljtj2Hljsj2H ljt sj2H
as y" 1. Up to multiplicative constant l, the latter is equal to the covariance function of the FBM introduced in Eq. (5.48). We therefore conclude that the sequence of Gaussian centered processesfZt ygconverges weakly toward the latter
Gaussian process. j
5.6 CONCLUSION
As noted by Boxma and Dumas [4] in the case of a ®nite number N of on=off sources, no matter how heavy the tail of the off duration distribution, the distribution ofV0 N is light tailed (i.e., has ®nite moments) whenever the on duration distribution is light tailed. On the other hand, for an on duration distribution with Pareto tail, the distribution ofV0 N also has a Pareto tail whatever the off duration distribution. In contrast, the Weibullian distribution derived in this chapter forV0 N asN" 1in