NOTE: This transcription was contributed by Martin P.M. van der Burgt. Although its markup is incomplete, we believe it serves a useful purpose by virtue of its searchability and its accessibility to text-reading software. It will be replaced by a fully marked-up version when time permits. —HR


On one-sided smoothing of event sequences.

We consider a finite number of events Ei (i from 0 through N) occurring at moments ti (i > j ⇒ ti > tj). The problem I found myself faced with was how to define a continuous function f(t) that could be interpreted as “the event frequency at moment t.” (I encountered this problem when trying to discover a strategy for deciding how much primary store should be allocated to independent programs in a multiprogrammed environment when for each program a “target page fault frequency” had been decided.)

A possible definition of the event frequency is in terms of the Dirac function:

N
f(t) = Σ δ (t - ti)(1)
i=0
which has the obvious advantage that it satisfies
+∞
∫ f’(t).dt = N + 1(2)
-∞
but the equally obvious disadvantage of severe discontinuities, which makes the comparison of f0(t) and f1(t) — two different frequencies corresponding to different values of the parameter I could control — rather difficult. Hence our desire to smooth the data. The smoothing, however, had to be one-sided in the sense that the smoothed value f’(t) may at any moment t only depend on the set of values ti with ti ≤ t. In order to make a sensible choice among the wealth of possibilities for f(t) we should impose upon f(t) as many “sensible” constraints as we can think of. To start with, property (2) seems a good choice. The next one is that we should realize that frequencies —as usually understood— have a dimension “time -1”, i.e. if we consider at moments t’i = ati and define the corresponding frequency f’(t) for that second event sequence, then
f’(at) = (1/a) f(t)(3)
should hold. Poperty (3) is not in conflict with requirement (2), for
+∞+∞+∞
∫ f’(t’).dt’ =∫ f’(at).d(at) =∫ a.f’(at).dt =
-∞-∞-∞
 +∞ 
 ∫ f(t).dt = N + 1 
 -∞ 
Reasonable as requirement (3) may seem it is (for finite sequences) too much to hope for: if t0 = 0 the relation (3) will not hold for t0 ≤ t ≤ t1, for how are we going to guess “a”? Therefore we must loosen it and can only require (3) to hold asymptotically for t > ti with sufficiently large i.

Relation (2) however, still stands rather firmly. Because in (2) we have only used

+∞
∫ δ(t) dt = 1
-∞
We could consider functions wi(t) such that

w(t) = 0 for t < 0(4)
and ∫ w(t).dt = 1(5)
0
 N
and replace (1) by f(t) =Σ wi (t - ti)
 i=0
(6)
where (4) does credit to the one-sidedness of the required smoothing , (5) guarantees (2) and (6) makes the whole definition invariant for shifting of the origin. The question is whether we can choose the wi in such a way — satisfying (4) and (5) — as to approximate (3) at least asymptotically.

A reasonable first guess — corresponding to fading out of the past in an exponential way — for wi(t) is with ki > 0:

wi(t) = 0 for t < 0
      = ki.e-ki.t for t ≥ 0(7)
where we can try to adapt ki — on account of past history, if any — towards a better approximation of (3). I called (7) a first guess, because we also wanted a continuous function f(t) and with (6) and (7), f(t) is discontinuous for any t = ti. A continuous f would follow from the proper linear compositum — proper in view of (4), (5) and wi(t) = 0 —
wi(t) = 0 for t < 0
      = (e-ki.t - e-hi.t).(hi.ki)/(hi - ki) for t ≥ 0(8)
and now we are free in the choice of both hi and ki, provided that they are both positive and different from each other. I am severely tempted to restrict myself to
hi = c * ki with c ≠ 1 and c > 0(9)
the reason being that both hi and ki are of dimension “time-1” and that I cannot forget my target (3).

*              *
*

In order to come to grips with the constant c I studied (omitting subscripts “i”) according to (8) & (9)

      w(t) = k . c/(c-1) . (e-kt - e-ckt)
and tried to determine c so as to minimize the maximum value of
                        c/(c-1) . (e-kt - e-ckt)
(I wanted my ripples as smooth as possible.) As this led to the inacceptable value c = 1, I took the limit for c → 1 and found — and this is my final suggestion —
w(t) = 0 for t < 0
      = ki2t.e-kit for t ≥ 0(10)
 
Also (10) has the property that∫ wi(t).dt = 1
 0

*              *
*

The next thing to decide is, how to choose the value ki. A constant value for ki seems in view of (9) out of the question, because ki has the dimension of a frequency. Apart from initial difficulties it seems reasonable to choose ki proportional to f(ti) i.e.

ki = d * f(ti)(11)
To get some feeling for a reasonable value for d, we consider f(t) for t≥0 as results from an infinite sequence of events that have occurred at t=0, -1, -2, -3, .... etc; for all passed events we may assume the same value ki = k and we find for f(t), according to (6) and (10)
 
f(t) = k.Σ k.(t+i).e-k.(t+i)
 i=0
which, indeed satisfies f(0) = f(1) as was to be expected.

Note. I summed the sequence and found

    k2.e-k.t                 e-k
f(t) =(————) . (t + ————)
    1 - e-k                  1 - e-k
(End of note)

The functions (10) have all one maximum, viz. at t = ki-1. We would like to attribute the maximum of f(t) for 0 ≤ t ≤ 1 to the term with i = 0 (i.e. the last event). For i > 0 all the terms should be decreasing functions of t for t > 0. We can expect the smoothest maximum of f(t) if it is in the neighbourhood of t = ½ and this suggests k in the neighbourhood of 2 and

d = 2(12)
For, because in the average f(t) = 1, we know that f(0) = f(1) < 1, therefore with d = 2 we shall find k < 2 and the term with i = 0 will reach its maximum at t0 > ½. Because the contribution of the remaining terms is a decreasing function of t, this will be compensated and f(t) will find its maximum at tf < t0.

Note. Using I sliderule I have investigated the consequences of the choice d = 2. Salvo errore et omissione I found for k=1.616, f(0) = f(1) = 0.808 as minimum value for f(t) and f(0.41) = 1.105 as maximum value. Anyone that would like to rely on these figures should reestablish them in a more reliable manner, I only mention these results because they represent the outcome of an hour’s work and give some impression of how smooth f(t) becomes in the case of equally spaced events. If it were of importance, I would try slightly smaller values of d as well. (End of note)

19th February 1975prof.dr.Edsger W. Dijkstra
Plataanstraat 5Burroughs Research Fellow
NUENEN - 4565
The Netherlands

Transcribed by Martin P.M. van der Burgt
Last revision 2014-12-08 .