Mathematical Assessment Synthetic Hydrology: Vol. $, No. 4 Water Resources Research Fourth Quarter 1967
Mathematical Assessment Synthetic Hydrology: Vol. $, No. 4 Water Resources Research Fourth Quarter 1967
Mathematical Assessment Synthetic Hydrology: Vol. $, No. 4 Water Resources Research Fourth Quarter 1967
Mathematical
Assessment Synthetic
Hydrology
N. C. MATALAS
U.S. GeologicalSurvey
Washington,D.C.
tO generatesynthetic events that resemblethe ess [Papoulis, 1965] that may be used to gen-
historic events in terms of these estimates
erate synthetic events that will resemble the
in the followingmanner. The most recent his- historicevents in terms of , , x, and (1).
toric event is representedby xi, and a value If the lag-one Markov processis taken to
of ei+ is randomly selectedfrom a population representa strictly stationary process,then the
that has zero mean and unit variance. With probability distributions of x and x+ must be
these two values, equation i yields a value for considered.With a strictly stationary process,
x+, the first synthetic event. This synthetic the distribution of xi is identical to x+ for
event now assumesthe role of x and with the all valuesof k, where]kl -- 0, 1, ... [Papoul;s,
random selection of a new value of e+, equa- 1965]. If x is assumed to follow a skewed dis-
tion i yields a new value for x+, the second tribution, then the assumption of strict sta-
synthetic event. This procedure is repeated N fionarity canbe usedto generatesyntheticevents
timesto obtaina syntheticsequenceof N events. that will resemble the historic events in terms
As N tends to infinity, the estimatesof x, , of , , , and x(1). Whether a particular
and px(1) obtained from the synthetic events skeweddistributionhas any physicalrelevance
will approach, , and p(1). is immaterial as long as interest is confined to
In generatingthe synthetic events, the only , , , and (1). In line with this interest,the
requirementimposedon the values of e+ was reasonablenessof stationarity of order greater
that the values be randomly selected from a than 3 need not be questioned.The assumption
populationthat has zero mean and unit variance of strict stationarity is imposedmerelyto provide
for all valuesof i. This requirementlimits equa- a means of generating synthetic events, in a
tion i to representingwhat is called a weakly rather straightforward manner, that will resem-
stationaryprocess[Wold,1954].For sucha proc- ble the historic events in terms of x, , ,
ess, and r are independentof the absolute and (1). Becausethe gamma and log-normal
valuesof i, and the lag-k serialcorrelationcoeffi- distributionsare widely usedto approximatethe
cient,px(k), ]k] - 0, 1, 2, .-. , that pertainsto distributions of hydrologic events, these two
the events x and xi+ dependsonly upon the distributions will be considered to illustrate the
absolutetime difference[i -- (i q- k)] for all use of strict stationarity in the generationof
values of i and k. synthetic events.
If resemblancebetween the synthetic events Let Yi denote a set of random variates, j =
and the historic events is to be extended to the 1, .-. , m, where each variate is generatedby a
skewhessof x, equation i must be modified. lag-oneMarkov process
This modification can take one of several forms,
amongwhich are the following. Y,+,,i = P,(1)yl ,i 'q'-[1 - p,'(1)"/= (4)
To considerskewness,Thomas and Fier{ng
[1963] replacedthe random componente+, by wherey is normally distributedwith zero mean,
+,, whichis definedas unit variance, and has lag-one serial correlation
coefficientequal to pv(1) for all values of j.
e+,.i is normally distributed with zero mean
and unit variance and is independentof y.
wherethe skewness of $, denotedby q,e,is related for all values of j. In terms of first and second
to the estimate of the skewhessof x, denoted order moments of Yi, equation 4 representsa
by -, by weakly stationary process;however, with the
assumptionthat Yi is normallydistributed,equa-
[1 -- a(1)]
(3) tion 4 representsa strictly stationary process,
'Y= [1 -- ,8'(1)]
a/'% sincefor a normal processthe processis com-
If r/+ is assumedto be normally distributed pletely defined by the first and secondorder
with zero mean and unit variance, then + is moments[Papcults, 1965].
approximatelydistributedas gamma with zero For each value of j, equation4 may be used
mean, unit variance, and skewhess3'. With to generatea sequence of y's, wherethe sequences
+ as the random component, the lag-one are mutually independent,so that the random
Markov processis a third-orderstationaryproc- variate z, definedas
SyntheticHydrology 939
where the x's have mean , sandard deviation p(1) = [exp[ap] - 1]/[exp [a2] - 1]
, coecient of skewess = (8/m)TM,and (12)
lag-one serial correlation coecient p%(1)
The procedurefor generatingsyntheticevents
p(1) : p(1). If , d, and (1) denote he that will resemble the historic events in terms
parametersfor a historic sequence,then the
of x, dx,, and (1) is as follows.The valuesof
sthetic events generatedby equation 6 will
resemble the historic events in ter of these
, , , and x(1) are set equal to the right-
hand sidesof equations7, 8, 9, and 12, where-
parameters.For the events generatedby equa-
upon the solutionsof these equationsgive the
tion 6, the coecient of skewness ' is a function
valuesof a, g, a, and p(1). With the valuesfor
of m, and becausem must be an integer,
/, ay, and py(1), equation 10 may be used to
cannot be set equal, in general, to the historic
generate a sequenceof y's. To thej antilog of
coecientof skewhess . If 2 , then eachvalue of y, the value of a is added to obtain
the procedureoutlined above cannot be used
the synthetic sequenceof x's that will resemble
to generatesynCeric eventsthat are distributed
the historic sequencesin terms of , , ,
as gamma.
and (1). If the valuesof a, g, a, and p(1) had
If the historic events are assumed to follow
been obtained from the logarithms of the his-
a 3-parameterlog-normaldistribution,then syn- toric events rather than in the manner out-
thetic events may be generated that are so
lined above, then equation 10 would lead to a
distributed and that resemble the historic
syntheticsequencethat would not resemblethe
eventsin ter of , d, , and (1).
Assume a to be the lower bound of a random
historicsequencein termsof , dx,, and (1).
The importanceof synthetic events following
variate x, where (x- a) is log-normally dis-
a particular probability distributionis hard to
tributed, so that y: log (x -- a), where loga- evaluate. In studies that deal with truncated
rithm is to the base e, is normally distributed.
events, as would be the case when low flow
For the randomvariate x, the mean, variance
augmentationis being considered,the probabil-
, and the coecient of skewness are related
ity distribution of the synthetic events might
to the lower bound a and to the mean and
be quite important. In such studies,interest is
variancea for the random variate y by focused on the distribution of the durations and
volume deficiencies that are associated with flows
= a exp[1/2a ] (7)
2
that are lessthan certainspecifiedvalues.These
a exp[2(a )] -- exp[a 2y] flow characteristics
for the syntheticeventsmay
(s) not resemble those for the historic events if
only the parameters, d, x, and (1) are con-
exp[3a] -- 3 explay] 2 sideredin the generationof the syntheticevents.
= {exp
[a
]-- 1}/ (9)
A strictly mathematicalapproachto determining
[Aitchisonand Brown, 1957]. Assumey instead whether the probability distribution of the his-
of x to be generatedby lag-oneMarkov process toric events must be consideredin the genera-
tion of the syntheticeventsdoesnot appear to
- = - be feasible. Perhaps the best approach is an
operational one' in a particular situation use
+ [1 - (10)
syntheticsequences that do and do not consider
where + is normally distributed with zero the probability distribution of the historic
90 . C. MATALAS
whose elements are the eigenvaluesof M0 -- whoseeventsoccurat the time points (i q- 1).
MM M. Thus, if px)(q)(1)is replacedby x)(q)(1)in the
Insofar as synthetic hydrologyis concerned, matrix M, then the multivariate synthetic se-
the elementsof B carry no physicalsignificance. quencesgeneratedby equation13 will resemble
Let B denote the matrix obtainedby the prin- the multivariate historic sequences in terms of
cipal componentsolution of equation 17 and x(p),,(), )(1), and ,)()(0), but not in terms
let 0 denotean (m X m) matrix suchthat 00 - of p)(q)(1). If )(q)(1) is of no interest,p(p)(q)(1)
I, where I is an (m X m) identity matrix. A may be taken to be x)(q)(1),definedby equa-
matrix B*, definedas tion 23, thus avoidingthe actual computationof
)(o(1) from the historicevents.
B* -- BO (19) To account for the coefficients of skewness
where ) the assumptionis made that x) followsa
3-parameterlog-normaldistributionwith lower
B'B* ' = BB' ' (20) bound a(p), so that y) = log (x(p) -- at)) is
may be usedin placeof B in equation13. There normallydistributed,wherelogarithmsare taken
existsmore than one matrix 0 suchthat 00 = I, to the base e. The mean, standard deviation,
coefficientof skewhess,and the lag-one serial
and thereforemanyB* matricesmay be defined,
correlation coefficientfor x) are denoted by
any one of which may be usedin place of B in
equation 13. ), ), %,
a ) and x)(1), respectively,and
The matfleesA and B are functionsof only two for y), the mean, standarddeviation, and lag-
one serial correlation coefficientare denoted by
matflees, Mo and M, that involve only second
order moments. The first order moments tz,), a,(p),andp,(p)(1),respectively.The relations
havebeenaccounted
for in the definitions
of betweenthe parametersof x(p)and y(p)for each
matflees x+ and x. The matrix M involves value of p are given by equations7, 8, and 9.
the lag-one serial correlation coefficientsand The lag-zerocross-correlation betweenx(p) and
the lag-one cross-correlation coefficients.If the x(q>,>(,(0), is related to the lag-zero cross-
lag-onecross-correlations are of no interest,then correlationbetweeny) and y(O,or p?)(q)(0), by
the task of computingthe elementsof M may
be facilitated in the followingway.
The matrix A may be taken to be a diagonal exp[a(P)a()p(p)()(0)]-- 1
matrix whosediagonalelementsare the lag-one {exr [a,(>
] -- 1}/{ exr [,'q'] -- 1}/
serial correlation coefficients'the (p, p)th ele-
ment of A is p ) ()1. With A so defined, the
pth and qth elementsof x+ are The y(p),p = 1, ... , m, are assumedto be
representedby the multivariate weakly sta-
() () () ' () (21) tionary process
Y+ = Y + s' + (25)
- + (22)
which is analogousto the processrepresentedby
whereb., and b., are the (p, s)th and (q, s)th equation13, wherethe elementsof A and B
may be detersned in the manner indicated
elements of B, respectively.If both sides of
above for the mirices A and B of equation 13.
equation21 are multipliedby x(q),then the The solutionfor B' vol yes lhe (m X m) mtrk
expectationsof the various productslead to
M/, whoseelementsare composedof the lag-one
- ()( a ()( ()
(1) serial correlationcoefficientsof y(p),p = 1, .-- ,
m, and the lag-onecross-correlation coefficients
where )(q)(1) denotes the lag-one cross-cor- of y(p)and y(O,p, q = 1, ... , m. If the lag-one
relation for the events generatedby equations cross-correlationsof x() and x(q)are of no interest,
21 and 22. Equation23 showsat for lag-one then the lag-one cross-correlations of y>) and
Markov processes, e lag-onecross-correlationy{q),p)<q)(1), may be definedas
the product of the lag-zerocross-correlation
and the lag-oneserialcorrelationof the variate
942 . c. MATALAS