The harmonic analysis of tidal model time series

M. G. G. Foreman and R. F. Henry

Institute of Ocean Sciences, P.O. Box 6000, Sidney, B.C., V8L 4B2, Canada

The harmonic analysis of tides is reviewed and guidelines for the analysis of tidal model time series
are presented. The analysis technique introduced by Godin and circulated by Foreman is
summarized. The related issues of constituent selection, time series length, confidence regions, and
inference are discussed. It is shown that since numerical model time series have a much lower
non-tidal content than observational time series, constituents can be separated with shorter record
lengths.

INTRODUCTION interactions of astronomical tides can produce secondary

tidal components. These components are often referred to
Over the centuries many theories have been proposed to
as compound tides or overtides. However since the
explain and predict oceanic tides. The major break-
nonlinear interactions often arise in regions with shallow
through in our present understanding of tides came about
depths, the term shallow water tides is also commonly
with the development of Newton's 1 theory of gravitation.
used. Details on the generation of these additional
He not only showed that tides originate from the
components can be found in Godin s (pp. 154-164).
gravitational forces of the sun and moon acting on a
All tidal frequencies are linear combinations, termed
rotating earth, but he also expressed the variations of
harmonics, of the rates of change of r, mean lunar time,
these forces by means of the equilibrium tide which would
and the following five astronomical variables that
result under certain idealized conditions. Laplace 2'3 made
uniquely specify the position of the sun and moon:
further major contributions. He derived the mathematical
expression for the astronomical forces that cause tides, the - s, the mean longitude of the moon
tidal potential. He also formulated the equations of h, the mean longitude of the sun
motion for tides on a rotating earth and solved them for - p, the mean longitude of the lunar perigee
the special case of an ocean completely covering the earth. - n', the negative of the longitude of the moon's
The harmonic method for tidal analysis and prediction ascending node
was independently and almost simultaneously developed - p', the mean longitude of the solar perigee.
by Thomson (later known as Lord Kelvin) and Ferrel.
The approximate periods for these six variables are 24.84
(See Doodson and Warburg 4 for further details and a
hours, 27 days, 1 year, 8.8 years, 18.6 years, and 21 000
more complete historical review.) This method was
years respectively. For each constituent, the integer
expanded by Darwin s and further refinements were made
coefficients of these 6 harmonics are called the Doodson
by Harris 6 and Doodson 7. The harmonic analysis
numbers. For example, the Doodson numbers o f M 2 and
technique introduced by Godin s and presented here is
K 1, the largest semi-diurnal and diurnal constituents
based on Doodson's work.
(approximate periods of 12 and 24 hours respectively), are
Doodson's 7 tidal potential expression reveals that the
(2, 0, 0, 0, 0, 0) and (1, 1, 0, 0, 0, 0) respectively.
astronomical forcing for tides can be written as a linear
In essence, a harmonic analysis requires calculating the
combination of sinusoidal terms, each having a distinct
amplitudes and phases for a finite number of sinusoidal
amplitude, phase, and temporal frequency. The oceanic
functions with known frequencies. The precise manner in
response to this forcing may be expressed in the same
which this is done will now be outlined.
manner, where each sinusoid is referred to as a tidal
constituent. Although the constituent frequencies are the
same as those in the tidal potential, the amplitudes and GODIN'S APPROACH
phases of the constituents can vary greatly due to the
irregular coastal boundaries and bathymetry of the As there are hundreds of tidal frequencies, many so close
earth's oceans. For example, the natural frequency of the that several years are required to separate neighbours by
Gulf of Maine and Bay of Fundy region is very close to the one cycle, and others which always have very small
frequency of constituent M 2. This causes resonance and amplitudes, it is neither practical nor mathematically
very large tides. On the other hand, in the southern North feasible to include all constituents in every analysis.
Sea an interference of M z signals causes an amphidrome Godin 8 resolved this dilemma by defining constituent
where the constituent amplitude is zero. clusters. Specifically he lumped together constituents with
There are two types of ocean tides. Components of the the same first three Doodson numbers and assumed that
tide that can be traced to the gravitational forcing of the for the first stage of the analysis, each cluster can be
sun and moon are termed astronomical. Nonlinear replaced by a single sinusoid having the same frequency as
the major constituent (in terms of tidal potential
Paper accepted March 1989. Discussion closes February 1990. amplitude) in the cluster. Consequently, constituents in

"~) 1989 Computational Mechanics Publications Adv. Water Resources, 1989, Volume 12, September 109
The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry
the same cluster have frequencies that differ by linear time series. This point will be discussed in more detail
multiples of the rates of change of p, n', and p'. The major later.
constituent gives its name to the cluster while the lesser In addition to a criterion that restricts constituents
constituents are referred to as satellites. For example, according to the length of record, a complementary
Godin's M 2 cluster contains the major constituent M2, criterion is also required for deciding the order of
with Doodson numbers (2, 0, 0, 0, 0, 0); and seven inclusion. For example, if the Rayleigh criterion has
satellites with Doodson numbers (2, 0, 0, - 1 , 0, 0), determined that the constituents associated with
(2, 0, 0, 0 , - 2 , 0), (2, 0, 0, 0, - 1 , 0), (2, 0, 0, 1, 0, 0), frequencies al and a2 cannot both be included in the
(2,0,0, 1, 1,0), (2,0,0, 2,0, 0),and (2,0,0,2, 1,0). Given a analysis, a second criterion is then needed to decide which
time series, Godin's harmonic analysis first calculates the one should be included. Godin makes this second decision
amplitude and phase for the sinusoid representing each on the basis of tidal potential amplitude. That is, the
cluster. Then it accounts for the presence of the satellites constituent with the larger expected amplitude is chosen
and calculates the amplitude and phase of each major first. The decision trees used by Foreman 9 for constituent
constituent. inclusion in the diurnal and semi-diurnal species are
With short records, it is not wise from a mathematical shown in Tables 1 and 2. The tidal potential amplitudes in
perspective to include all possible constituent clusters in a these tables do not include a latitudinal dependence factor
harmonic analysis. Although linear algebra only requires which is constant throughout each species. At latitude ~b,
that the number of unknowns (two for each constituent the factor is sin 2q~ for diurnal constituents and cos 2 q~for
cluster, and amplitude and phase) be no larger than the semi-diurnal constituents. See Doodson ¢ for further
number of equations (one for each observation), short details. The values in the record length columns are
records and close frequencies result in an ill-conditioned calculated from equation (1) with R = l.
matrix equation. Although such equations can be solved,
the confidence intervals for the solutions are large. In
particular, a small change in the observations can THE LEAST SQUARES ANALYSIS
produce a large change in the solutions.
In order to narrow the confidence intervals, a criterion Assume that a selection procedure has chosen M
is needed to restrict the inclusion of constituent clusters in constituents for inclusion in the analysis. The next stage is
the analysis. Godin uses the Rayleigh separation equation to calculate the amplitude and phase for the sinusoids
representing each constituent cluster. This is done by
la2-al[r> R (1) solving a system of linear equations. For a one
where (71 and a2 are major constituent frequencies dimensional time series (such as would arise from surface
(cycles/unit time), T is the record length, and R is the elevations), the analysis objective is to solve the system of
Rayleigh constant. This criterion includes both constitu- equations
ents in the analysis if the number of cycles separating al M

and a2 over the period T is greater than R. For yi=Ao + ~ Ajcos(ajti-(I)j) (2)
observational time series, R is commonly set to 1. j=1
However it can be smaller if the non-tidal content of the for the unknowns Aj and (:I)j,j = 1, M. Aj, aj, (I)j are the
signal is small. Such is generally the case with tidal model amplitude, frequency and phase of constituent j; and Yi,

Table 1. Order of Diurnal Constituent Selection in the Foreman 9 harmonic analysis programs
', Link Royleigh Comparison P a i r s
( ) Tidal Potential Amplitudes for Main Constituents

Length of Frequency Differences (cycles/hr)XlO 3 Between Neighbouring Constituents

Record (hr) oJ P-- co 03
cq _ 03 o o o to ~ h-
Required
0 0 0 0 ~ 04 CO o,I 0 ',~ ~" cO 0 0 C~
for r,~ N re) (M ~ N o N N ~b -- o o~ ~ ~
Constituent -' 6 - 6 -: - o 6 o 6 o 0 o 0
Inclusion TAUI~BET/NOI~CHII~PII '~ PI U S i ~ K' ~ P S I I P H I '
(53011)
24
(376941
01-=
_._.._---/
J
328
//
651 / /.'~
/
\ /
(1624)
ooi
t ,,
(955) (7217) / (2964 ~(2964)/ / {3,r)
662

764
ALP~
(278)
"\
2QI- OI
\ \
TAU I BETI
/
NOI'
/\
\ P'/
',jI *

//
sd,
UPSli

4383 {493) (278) (17543 (755)i /

CHII / ITHEI
4942 (1152) (13711 1567: /

;1[
SI PSII r 5671
8767
I P[JI
(~o28) (416) (422}
I L

110 Adv. Water Resources, 1989, Volume 12, September

 The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

Table 2. Order of Semi-Diurnal Constituent Selection in the Foreman9 harmonic analysis programs
Link Rayleigh Comparison Pairs
( ) Tidal Potential Amplitudes for Main Constituents

Length of Frequency Differences (cycles/hr)× iO 3 Between Neighbouring Constituents

Record (hr) P.- o~ ~- o~ ~- o ¢o = oo ¢~ ~ ~ _
~ ~ r~ t~ ,~. ~ O 0 0
Required
r~
~ o~ N o~ oJ F- Qo ,~ ,¢ ~ -tD
ao g ~
~
0
'*
0
,¢
0
'~
~
co
for o o o o o o =
o : - - N 0~ - - - e~

Constituent
Inclusion 002 IEPS2 2N2 M R2 K2 MSN?-- I

(
13

355

662

764

4 38 3 ') / / S2 / / SN~

4942 (259) (2sol) (3s02) ,

HI H2 T2 R2
8767 (s=3) {277) (2476) (355)

GAM'~
I I :3 2 6 (273}

i = 1, N, is the observation at time t~. Each equation can be sampling, the M 2 harmonic whose frequency is closest but
made linear in the new unknowns C~ and St by rewriting still less than the Nyquist threshold, is m]2.
Ajcos(cTjti-(Yl)/)=Cjcos(ajti)+ Sjsin(~yjtl) (3)
where STATISTICS O F T H E LEAST SQUARES
A~ = (C 2 + 82) 1/2 and (I)j = arctan(SflC). SOLUTION
The statistical properties of the least squares solution can
As the number of equations, N, is greater than the number
of unknowns, 2 M + l , the system of equations is be found in any analysis of variance or regression model
text (e.g., Draper and Smith11). They are summarized as
overdetermined and all the equations cannot be solved
follows.
exactly. The solution must therefore be calculated by
The system of equations given by (2) and (3) can be
minimizing the equation residuals in some sense. The least
re-expressed in matrix form as
squares approach, wherein the sum of the squares of the
residuals is minimized, has been adopted by Godin. This y = Ax + e (4)
approach produces the so-called normal equations which
where y, x, e are vectors of the observations, the C i and Sj
are efficiently solved with the Cholesky algorithm (see
Ortega1°). coefficients, and the residuals. The normal equations that
result from minimizin9 e'e are
It should be mentioned that a harmonic analysis which
uses the least squares approach permits the time series to AtAx = Aty (5)
be irregularly sampled, or to be regularly sampled and
and their solution is
contain gaps. Such would not be the case if a Fourier
transform were used to calculate the amplitudes and x s=(AtA) tA'y. (6)
phases. However, if the time series is regularly sampled
and has no gaps (as would be the case with tidal model The total sum of squares is y'y and the sum of squares
time series), the terms in the normal equations matrix can due to regression is x'sAty. Their difference is the residual
be calculated efficiently and accurately and the matrix can error sum of squares, and this difference divided by the
be partitioned. This partitioning reduces the solution time degrees of freedom is the residual mean square error, a2mse.
by a factor of four. See Foreman 9 for further details. With an analysis of N observations that includes M
The recommended sampling interval for the time series constituents and the constant term Zo, a,,~e2 is then
is one hour. In addition to convenience, such an interval 2 yty _ xtsAty
- -

permits at least two observations per period for the high a,,~e - N - 2M - 1 (7)
frequency shallow water tides such as M4, M6, and M s. It
also avoids aliasin9 wherein constituents that are sampled If it is assumed that e is distributed normally with 0
fewer than twice per cycle become irretrievably confused mean and 0-2I variance (where I is the unit diagonal
with lower frequency constituents. With one hour matrix), then the variance ofxs is (AtA) - ~a z. Since ~rz~eis

Adv. Water Resources, 1989, Volume 12, September 111

The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

major and minor axes are approximately equal to the Cj

I and S~ confidence intervals. Figure 1 illustrates these
I
I points.
I
I
E I
I CORRECTIONS TO THE LEAST SQUARES
0 SOLUTION
0 In order to calculate the amplitude and phase of the major
o constituent from the representative sinusoid for each
r- . . . . . . . . . . . . . . . . . . . . . .

constituent cluster, at least two corrections to the least

squares solutions are required.
The first is usually termed a nodal correction and it
accounts for the fact that the amplitudes and phases of the
reresentative sinusoid that arise from the least squares
solution actually represent the cumulative effect of all
Cosine Coefficient constituents in the cluster. Since these constituents differ
Fig. 1. An example of confidence intervals for Cj and Sj slightly in frequency, the amplitude and phase of the
and the associated confidence ellipse for the complex replacement sinusoid vary slowly in time and thus do not
amplitude Aj e- ~, provide a basis for prediction at a subsequent time. In
order to avoid this difficulty, the time-invariant amplitude
and phase of the major constituent in each cluster are
calculated from those of the replacement sinusoid.
an unbiased estimator of a 2, an estimate of the standard
So to summarize, the outcome of the least squares
deviation of xi, the ith element of xs is then
analysis is an amplitude and phase pair for each
ai = (/~(A'A)- l ~ a 2 m s e ) l / 2 (8) constituent cluster. The nodal correction calculation
transforms this pair into an amplitude and phase for the
where/~ is the vector with 1 in the ith position and O's
major constituent. When a prediction is required, the
elsewhere. Such estimates are easily calculated (c.g., in the
reverse process occurs. That is, the major constituent
program associated with Foreman 9) as (AtA) - 1/~ is the
amplitudes and phases are first transformed into cluster
solution of the modified normal equations A~Ax=/~.
values, and these in turn are used to produce the time
Confidence intervals can be computed from these
series.
estimates. For example, the 1 - ~ confidence limits for xi,
Algebraically, the nodal correction process re-expresses
the ith element in x, are
the least squares result for each representative sinusoid
x i ++t(v, 1 - ½oOai (9) A cos(at-@) as
where al is given by equation (8) and t(v, 1 - : ) is the f(t)a cos(at + u(t)- 0). (12)
Student's t distribution coefficient at the 1-½~ percentile
for v degrees of freedom. f(t) and u(t) are the amplitude and phase correction
When Cj and S i are combined as in equation (3), it is factors that account for the presence of satellites in the
generally not correct to simply take the cross product of cluster; and a, 0 are the amplitude and phase of the major
the one dimensional confidence intervals to form a constituent in the cluster. The nodal modulation terms
confidence rectangle for the complex amplitude A~ e -i*j. f(t) and u(t) vary slowly with time and, for records up to
Such a combination does not take into account either the one year in length, very little error is introduced by
covariance of C i and Sj or the covariance of these two assuming them to be constant and equal to their value at
parameters with their counterparts arising from the other the midpoint of the record.
constituents that were included in the least squares With appropriate assumptions one can solve for f and
analysis. The correct joint 1 - ~ confidence region for all u. Godin assumes that the amplitude ratios and phase
parameters in the least squares analysis is given by differences between the satellites and the major
constituent are the same as predicted by the tidal
(x-xs)~AtA(x-xs)<-..pa2seF(p, v, 1 - ~ ) (10)
potential, and that the satellite constituents do not
where F ( p , v , l - s ) is the upper c¢ point of the F interfere with the major constituent. This latter
distribution with p parameters and v degrees of freedom. assumption loses validity with analyses over periods
This inequality describes a p-dimensional ellipse whose longer than one year. Further details of these assumptions
shape will depend on the relative magnitudes of the entries and the calculation of f and u can be found on pages
in the covariance matrix (AtA) -1. With a harmonic 167-171 of Godin s.
analysis of regularly sampled data and a time origin at the It should be mentioned that with Godin's harmonic
center of the analysis period, matrix A~A can be analysis, the term nodal correction is a misnomer. The
partitioned so that the Cj and S~ parameters can be solved phrase and the symbols f and u were first used before the
separately. Thus advent of modern computers to designate corrections for
cov(C~, S j) = 0. (11 ) the moon's nodal progression that were not incorporated
into the astronomical argument calculations for the main
Furthermore, if the analysis period is sufficiently long to constituents. However, now the term satellite modulation
resolve constituents j and i, then both cov(Cj, C~) and is more appropriate because the correction is not only for
coy(Sj, Si) will be small and the (Cj, S j) confidence region the effects of the moon's ascending node, but also for the
will be essentially unaffected by Ci and S~. This means that effects of the solar and lunar perigee.
the joint (Ci, Sj) confidence region will be an ellipse whose The second correction to the least squares solution is

112 Adv. Water Resources, 1989, Volume 12, September

 The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

referred to as the astronomical argument correction. It With good inference parameters, that is accurate
simply re-expresses phase lags with respect to an absolute amplitude ratios and phase differences for P~/K 1 and
time and space origin. Instead of regarding each tidal Kz/S z, both these problems would be greatly reduced.
constituent as the result of a particular component in the The question of when constituents should be included
tidal potential, an artificial causal agent can be attributed directly in the least squares analysis and when they should
to each constituent in the form of a fictitious star which be inferred is not easily answered. The presumption that a
travels around the equator with an angular speed equal to record of length T is required to distinguish constituents
that of its corresponding constituent. Making use of this with a frequency separation of T - ~ (i.e., equation (1) with
conceptual aid and setting t o to be a reference time and er R = 1) is incomplete. In fact it conflicts with the algebraic
to be the frequency, the astronomical argument of a given viewpoint whereby, for any n independent observations,
tidal constituent, one can obtain n equations and solve for n unknowns
V(t)=•(t-to)+ V(to) (13) (e.g., n/2 amplitudes and n/2 phases) regardless of the
frequency separation.
can then be viewed as the angular position (longitude) of The missing consideration in both of these viewpoints is
its fictitious star. For historical reasons, all such that the observations contain other non-tidal contribu-
arguments or longitudes are expressed relative to the tions, henceforth referred to as noise, in addition to the
Greenwich meridian and can consequently be expressed discrete tidal signals. With actual observations, this noise
as functions of time only. The phase lag for a particular may be comprised of meteorological effects, non-tidal
constituent can then be viewed as the time between the waves, and random errors in recording the observations.
overhead passage of the fictitious star and the time of With tidal model time series, the noise may be comprised
maximum tide. So the contribution of each major of transient solutions and rounding errors. Taking these
constituent can be expressed as effects into account, Munk and Hasselmann 13 showed
a cos(a(t - to) + V(to)-9) (14) that meaningful statements can be made about a~ and a2
provided
where 9 is a phase lag that is measured with respect to an
absolute time and space origin. standard deviation of noise
(16)
It should be pointed out that many types of phase lag ''laZ-allT>standard deviation of signal
can be used in harmonic analyses depending on whether
or not one wishes to take into account the longitude and (Notice that this equation gives a guideline to the choice
time zone of the observation site. Schureman 12 of R in equation (1).) Munk and Hasselmann refer to any
lpp. 75 78) discusses several common phase lags. The resolution improvement over and above the frequency
analysis programs distributed by Foreman 9 calculate G, separation T - 1 as super-resolution. They also show that
the Greenwich epoch, if the time series is in Greenwich for two neighbouring spectral lines at frequencies fo and
lime and g, the modified epoch, if the time series is in local f0 + N', the variance in the estimate of either amplitude is
lime. They do not employ any longitude correction.
Consequently, the analysis results from neighbouring 3S ( f o)/(zr2(Af)2 T 3) (17)
sites that were recorded in the same time zone can be where S(f) is the underlying power spectrum of the noise.
compared directly. And the conversion of a phase lag Munk and Bullard 14 estimate S ( f ) = l cm2/cycle/day
from one time zone to another is straightforward. For based on measurements of the noise spectrum well to one
example, if g is calculated from observations that were side or the other of the tidal line clusters. However when
recorded in PST, then cusps in the noise spectrum are taken into account,
G=g+8a (15) estimates of S(f) increases dramatically to as much as
300cm2/cycle/day (Munk and Hasselmann 13) within a
for a constituent with frequency a measured in band of _+0.01 cycles/day around the M 2 frequency.
degrees~hour. With S(f)= 300 cmZ/cycle/day and frequencies separ-
ated by A f = 1/180 cycles/day (the approximate separ-
ation between K 1 and P1, and S 2 and K 2 ) , the amplitude
INFERENCE variances for each constituent arising from analyses of 30,
60, 90, and 180 days, are 109, 13.7, 4.05, and 0.51 cm 2
A final correction that is sometimes made to the least respectively. However with tidal model observations,
squares solution is referred to as inference. It involves the both S(f) and these variance estimates will be much
extraction of important constituents, excluded from the smaller.
analysis because the record was too short, but deduced Uncertainties in the analysis results can also be viewed
afterwards from known relationships with constituents from the perspective of matrix theory (see Ortega ~° for
that were included. If it is done properly, inference serves example). Assume B x = b and B x ' = b ' are the normal
the dual purpose of further reducing the residual error in equations (see equation (5)) arising from a least squares fit
the least squares solution, and eliminating a periodic for the amplitudes and phases of several tidal
behaviour in the estimated amplitudes and phases of the constituents. In particular, assume the right-hand sides b
reference constituents. For example, a 182-day time series and b' are respectively calculated from observations
is required to separate P~ from K 1 (the third largest and without and with background noise. That is, b assumes a
largest diurnal constituents), and K z from S z (the fourth pure tidal time series whereas b' assumes tide and noise. If
and second largest semi-diurnal constituents) by one for some norm II'l], K ( B ) = [IBII lIB-11] is the condition
cycle. Without inference, a series of monthly analyses number of matrix B, then
would show a periodic behaviour in the amplitudes and
phases o f K 1 and S z and have substantial contributions to [Ix--x'll ~<K(B)Ilb-h'[I (18)
the residual errors due to the P1 and K 2 constituents. rlxll Ilbll

Adv. Water Resources, 1989, Volume 12, September 113

 The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry
Table 3. Harmonic analysis results with and without inference. Amplitudes are in centimeters and phases are in degrees. Analysis type A denotes the direct
inclusion of all constituents; type B denotes the direct inclusion of only Zo, K1, M2,' type C denotes the direct inclusion Zo, K1, M2, and the inference of 0 l
and S2. Analyses I ~ 4 were with the time series comprised of only tidal signals whereas analyses 5 --* 8 were with the time series comprised of tides and
random noise

test hours type statistic Zo O1 K1 $2

nO. amp amp ph amp ph amp ph amp ph
1 72O A 25.000 3.700 297.5 5.250 179.1 9.065 220.3 4.225 69.9
illean 25.000 3.700 297.5 5.250 179.1 9.066 220.3 4.225 69.9
2 6O A [ max 25.000 3.702 297.6 5.251 179.1 9.069 220.3 4.229 69.9
I min 25.000 3.698 297.5 5.247 179.1 9.062 220.3 4.222 69.8
[
I s.d. .000 .002 .03 .001 .02 .002 .02 .002 .03
mean 25.044 5.901 176.8 9.494 220.5 I
3 60 B max 25.118 8.515 223.6 12.845 244.0 F
rain 24.990 2.308 136.6 5.404 194.2
s.d. .055 2.285 31.3 2.873 19.5
mean 25.044 3.754 298.3 5.327 179.9 9.062 220.2 4.224 69.7
4 60 C max 25.118 4.197 303.9 5.955 185.4 9.247 220.8 4.310 70.4
min 24•990 3.594 296.3 5.099 177.9 8.901 219.5 4.151 69.1
!
! s.d. •055 .169 2.13 •240 2,13 .125 .42 .058 .42
5 720 A 25.074 3.832 295.7 5.138 179•0 9.078 220.3 4.280 69.6
mean 25.070 3.530 297.3 5.005 174.9 9.188 220.0 4.487 69.3
6 60 A l max 25.370 5.143 310.1 6.121 188.3 9.917 227.0 5.388 82.7
[ rain 24.665 2.291 272.2 4;1:: 161.2 8.148 214.2 3.157 58.3
i s.d. •244 .840 12.3 • 9.9 .506 4.5 .615 8.]
p F

7
t
] 60 g
r mean
max
25.118
25.429
I! 85.856
.722
176.5
226.8
9.483
13.000
220.5
245.1
i min 24.743 ~2.131 i 132.2 5.158 193.5
s.d. .252 i 2.398 ! 34.2 2.974 19.5
mean 25.118 3.681 298.0 i5.223t179.6 r
9.014 220.2 4.201 69.8
8 60 C max 25.429 4.052 310.3 i 5 . 7 5 0 i 191.9 9.323 221.3 4.345 70.9
lllii1 24.743 3.350 292.4 j 4.753 , 174.0 8.561 218.3 3.990 67.9
o96J lO
I

, i
s.d. .252 .244 4.8 i .346 4.8 .206 1.0

The effect of seeking amplitudes and phases corre- constituents Z o, 01, K 1, M2, and $2, and the other with
sponding to frequencies al and a2 that are relatively close, these same constituents plus some random background
i.e. [al-a2[ < T-1, is to make the appropriate rows in B noise. Time periods of 328 hours and 355 hours are
more linearly dependent and increase K(B). So in the required to respectively separate O 1 from K 1 and S 2 from
presence of substantial background noise, one can expect M 2 by one cycle. (See Tables 1 and 2.) The tidal
significant differences between the calculated set of amplitudes (cm) and phases (degrees) were chosen to be
parameters x' and their true values x. If accurate inference (25.000,0.0), (3.70,297.5), (5.25, 179.1), (9.065,220.3),
parameters are available, inference in such a case should and (4.225, 69.9), for constituents Z o, O1, K1, M2, and S 2
yield better results, because solving for the parameters of respectively; while the noise was uniformly distributed in
only one frequency, either al or a2, would remove the the range [ - 1 . 0 , 1.0]. Nodal corrections were not
rows that are almost linearly dependent and reduce K(B). included in any of the time series syntheses or analyses.
On the other hand, if the noise level were very small the Three sets of twelve consecutive 60-hour harmonic
effect of a large K(B) would be counteracted and a analyses were executed: the first searching directly for all
reasonably accurate set of parameters x' could be constituents; the second searching for only Z o, K l, and
expected without inference. This should be the case with M2; and the third extending the second by inferring O 1
time series from numerical tidal models that have been and S 2 from K 1 and m 2 respectively. (Inference
run sufficiently long that the transient solutions have parameters were calculated from the true amplitudes and
essentially died away. (Such models are usually said to phases.) In order to compare performances, means,
have reached a steady state.) maximum and minimum values, and the standard
Table 3 gives the results of tests designed to deviations were calculated for each amplitude and phase
demonstrate these points. Two 30-day records of hourly from the twelve analysis results in each series.
tidal heights were simulated, one using only the The standard deviation ratio of noise to signal (without

114 Adv. Water Resources, 1989, Volume 12, September

 The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

Zo) in the second time series is approximately 0.07. So NORTH

with [0"2 --0"1[ ~ 1/355 hours in equation (16), meaningful
results should be possible with analyses of only
T = 60 hours.
Results from the analyses of the tidal record with no
background noise (Tests 1 to 4) demonstrate a clear
advantage to including all constituents directly in the least
squares fit. The small non-zero standard deviations in
Test 2 are wholly attributable to the fact that the data
were rounded to four digits, thereby making Hb-b'll
slightly larger than zero. The standard deviations for the
runs with inference (Test 4) are not zero because of
simplifying assumptions in the inference calculations. (See
Foreman 9 for further details.) The semi-diurnal standard
deviations are generally smaller than the diurnal values
because they have more cycles sampled over the analysis
period.
However, when random noise is included in the time
series (Tests 5 to 8), the standard deviations for the
inference runs (Test 8) are consistently less than those
obtained by the direct inclusion of all constituents in the
least squares fit (Test 6). This is a consequence of reducing
K(B) (using the l~ norm) from 120.1 for Test 6 to 2.884 for
Test 7.
The largest standard deviations in Tests 3 and 7 should
also be noted. The presence of O1 and $2 in the time series
but not in the harmonic analysis, has caused significant
fluctuations in the amplitudes and phases of K 1 and M 2.
However, as demonstrated by the standard deviations in
Tests 4 and 8, these fluctuations can be significantly
reduced through inference.
The standard deviations in test 6 can be compared to Fig. 2. Current ellipse notation
those estimated by equation (17). With the uniformly
distributed noise applied to hourly values, S ( f ) = current contribution for each constituent can also be
1/72cm2/cycles/day and the standard deviation asso- represented as a vector whose tip traces out an ellipse over
ciated with the amplitude estimates are approximately the period of the constituent, it is customary to re-express
0.22 cm for O 1 and K 1, and 0.24 cm for M 2 and S 2. The the amplitudes and phases for the two components in
larger values in Table 3 are probably due to interference terms of four ellipse parameters. As illustrated in Fig. 2,
between the two constituent pairs. these are: the amplitudes of the major and minor
With reference to equation 10, it is interesting to note semi-axes (Ma, Mi); the angle of inclination of the
the covariance estimates arising with the different tests. As northern major semi-axis, 0; and the phase lag of the
all the tests involved a uniformly sampled time series maximum current behind the maximum tidal potential of
whose origin t o could be defined (for the purpose of the the individual constituent, 9. Since an ellipse is invariant
least squares analysis) at the center of the analysis period, through rotations of 180 °, there is an ambiguity of 180° in
cov(Cj, S i ) = 0 for all j, i. For Tests 1 and 5, B was the ellipse orientation. Godin avoids this ambiguity by
diagonally dominant with the diagonal elements at least a requiring 0 to be in the range 0 to 180 °. This means that as
factor of ten larger than the off-diagonal elements. In 0 passes through the value 180 °, both 0 and 9 will be
particular, cov(Cj, Ci)/var(Cj)~0.1 and cov(Sj, Si)/ altered by 180 ° .
var(Si)~0.1 within species and these ratios were much One can also consider the current contribution for each
smaller between species. However for Tests 2 and 6, the constituent as the sum of two rotating vectors, one
covariances and variances were approximately equal in counterclockwise with amplitude and phase a + and 9+,
magnitude within species and slightly smaller between and the other clockwise with amplitude and phase a - and
species. So the effect of trying to separate constituent O 1 g - . For example, Crawford and Thomson a5 use g - and
from K 1 and S 2 from M 2 with too short a record is to the ratio a+/a - to determine the shelf wave contribution
increase the covariance estimates and joint confidence of the diurnal constituents off Vancouver Island. See
regions of the C and S parameters. Foreman 16 for further details on these various current
representations.

TIDAL CURRENTS C O M M E N T S O N T I D A L M O D E L ANALYSES

If the time series is two dimensional, such as occurs with The preceding general discussion is applicable to the
horizontal tidal currents, an extension of the one harmonic analysis of time series arising either from
dimensional harmonic analysis procedure can be carried observations or numerical models. The following specific
out. Specifically each component (usually in the comments are intended for analyses of tidal model
north/south and east/west directions) is analysed records. They apply directly to analyses with Foreman's
separately as a one dimensional time series. Since the tidal 9.16 set of computer programs. In all cases, it is assumed

Adv. Water Resources, 1989, Volume 12, September 115

The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

that the tidal model has either been forced directly with required to separate P1 from K 1 and g 2 from S 2 by a full
astronomical body forcing or indirectly by specifying cycle, these constituents can be adequately separated with
tides, either surface elevations or currents, on some of the much shorter time series. However it is important to note
model boundaries. It is also assumed that the tidal model that the non-tidal content in the time series is partially
has been run sufficiently long for it to have reached an determined by the constituents chosen for the analysis,
approximately steady state. This is probably best and these may only be a subset of the constituents in the
determined by producing several time series plots at model. For example, if the advective terms are included in
representative points in the model domain. a model simulation but the compound tides generated by
The first comment is that it is important to maintain these terms are not included in the analysis, the
consistency between model forcing and the harmonic compound tide energy is effectively noise.
analysis. This not only means that the same constituent These points are now illustrated with two runs using
frequency values (i.e., to the same precision) should be Flather's18 Vancouver Island model. The advective terms
used in both the model and the analysis, but also that the and the forcing constituents Q1, O1, P~, K1, N2, m 2 , $2,
nature of the forcing should be consistent. Specifically, if and K 2 are included in the first run, but only the forcing
single major constituents (rather than clusters) are used in constituents are included in the analysis. Figure 3 shows
the forcing, then nodal corrections should not be included the P~ amplitude contours based on a 30-day time series
in the analysis. On the other hand, if the model results are analysis. (The model was run for six days prior to this
to be compared with observations or to predict tides, the analysis period in order to permit the transient solutions
satellite effects should be included, though their to decay.) The wiggliness of the contours indicates an
contribution is minor for many constituents. This can be unsatisfactory variability in the analysis results. As
done in either of two ways. The first way is to force the predicted by equation (18), a combination of too short a
model with the representative sinusoids for the record and too much noise has perturbed the results from
constituent clusters, rather than just major constituents. their correct values. This wiggliness decreases slightly
The second way is to use the major constituents for the when thirteen shallow water constituents are also
forcing and analysis and adjust the harmonic analysis included in the analysis, but the results are still
results so that conventional nodal modulation correc- unsatisfactory. However, as seen in Fig. 3, with a 60-day
tions are applied to account for the presence of the other harmonic analysis the contours become much smoother.
constituents in the cluster. When the model is rerun without the advective terms, the
A second important point is that the constituents to be time series are less noisy and P~ and K~ (and K : and Sz)
included in the harmonic analysis should be chosen by the are adequately separated with a 30-day analysis.
user and not by the analysis program. Specifically, the Nevertheless, from a statistical viewpoint a longer record
constituents in the analysis should be those included in will reduce the confidence intervals around the solutions.
the forcing and those generated in the model interior by So, computing resources permitting, a longer simulation
nonlinear interactions. Constituents chosen by the is always preferable,
Rayleigh selection criterion in the analysis program may In summary, ifa model can be run sufficiently long that
not necessarily include all of the constituents in the model all transients have decayed below machine precision at
time series. For example, the simulation period may not
be sufficiently long that the selection criterion includes P~.
Consequently, some of the energy from the constituents Table 4. Shallowwater constituent and astronomical constituent twins
that are not included in the analysis will spill over and
contaminate the results of the constituents that have been Astronomical Constituent [ Shallow Water Constituenl~
1

included. Name Frequency (cycles/hr) NaYglc Composition |

As discussed in Lynch and Werner 17, it is also Q1 .03721850 Nh-1 N~ - K1 |
preferable to include in the boundary forcing the O1 .03873065 MK~ M2 t(1
significant shallow water constituents generated by the r~ .03895881 M P1 M2 P1
model. If this is not done, the model assumes zero NO~ .04026859 NOz iV,_ Oi
boundary forcing for these constituents thereby making PI .04155259 SK1 $2- K1
the interior amplitudes and phases incorrect. K1 .04178075 M0I M2 - 01
It should also be mentioned that some shallow water J1 .04329290 MQI M2 O~
constituents have the same (or almost the same) frequency SO .04460268 & 01
as those of astronomical origin (see Table 4). Foreman's OQ .07597495 OQ2
harmonic analysis program generally looks for the e: .07617732 MNS2 M2 + N2 $2
astronomical constituent, although it could be easily 2N .07748710 02 01 + 0i
altered to look for the shallow water twin. If the tidal #2 .07768947 2MS~ M, + M~ - S~
model has not been forced with the astronomical N2 .07899925 KQ. k-~ + Ol
constituent but there is energy at its frequency, obviously 72 .08030903 O~ 01+P1
the signal must then be due to shallow water activity. As M~ .08051140 K02 K~ + 01
the nodal corrections for the astronomical and shallow L: .08202355 2MA_~ M2 ~ 2142- N~
water constituents differ slightly, it is important to $2 .08333333 KP2 K~ =P1
correctly identify the energy source. K2 .08356149 E2 K1 + K1
Time series arising from tidal models obviously have a q2 .08507364 K J: K1 + ./1
low non-tidal content. Consequently, as discussed in the MOa .11924206 MOa M2 + 01
preceding Inference section, a record long enough to M3 .12076710 NKa N2 + K1
separate all frequencies by one cycle may not be required.
For example, although approximately 182 days are * M1 is a satellite of NO1.

116 Adv. Water Resources, 1989, Volume 12, September

 The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

15.0

~ b,x':

15.0 %
14.5

PACIFIC
OCEAN

Fig. 3. P1 amplitude (cm) contours based on a 30-day analysis of Flather's TM model with advection

%
PACIFIC \ (
OCEAN ~.

Fig. 4. P1 amplitude (cm) contours based on a 60-day analysis of Flather's 18 model with advection

Adv. Water Resources, 1989, Volume 12, September 117

The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

128 123
50 I 50

Vancouver

PACIFIC
0
v

I.U OCEAN
J

..-I

47 I I I 47
(a) 128 LONGITUDE (ow) 123

O,)

Fig. 5. (a) Freeland's 21 CODE-II current meter sites. Barotropic K 1 (b), and M 2 (c) current ellipses at the CODE-II sites.
The largest semi-major axis is 15.3 cm/s for M 2. Lines within each ellipse denote g. Arrows denote the direction of rotation
around the ellipse. Observed - - , model - - -

118 Adv. Water Resources, 1989, Volume 12, September

 The harmonic analysis of tidal model time series: M. G. G. Foreman and R. F. Henry

it t)

(c)

the beginning of the analysis period, and the analysis the model interior but not included in the forcing can
period itself is long enough to resolve constituents such as disrupt a comparison. A safer technique is to compare
Px and K 1, then including all constituents in the analysis amplitudes and phases for individual constituents.
is the best strategy. However, it may either be difficult to Special mention is also warranted for comparing model
determine the time required for this decay, or it may not and observational currents. Obviously if the model is two
be economically feasible to run a model this long. So the dimensional (i.e. barotropic), the current observations
noise may not be as small as one would hope, and should also be averaged in some sense. This requires
inference may be required to resolve closely separated simultaneous observations at several depths. If baroclinic
constituents. The numerical tests whose results are shown processes are suspected in the observations, some attempt
in Table 3 suggest that neighbouring constituents can be should be made to remove them before comparing the
adequately separated with R ~>1/6 in equation (1) when observations with barotropic model results. For example,
the noise level is less than 0.5 × 10 - 4 for each observation. using baroclinic theory such as presented by Phillips 19, it
This translates to a 30-day period for separating P1 and is sometimes possible to fit the observations with the
K 1. The Flather ~8 model tests with no advection confirm barotropic and one or more baroclinic mode shape
this value. functions.
If computer storage is at a premium during a model As with one dimensional tidal signals, it is generally
run, the information to be retained for a subsequent preferable with tidal currents to compare model results
harmonic analysis can be reduced significantly if at each and observations on a constituent by constituent basis.
model grid point, the right hand sides for the normal Figures 5b and 5c show K 1 and M z current ellipse
equations, rather than hourly time series, are stored. For comparisons for an array of six current meter sites in the
example, if a model with 500 grid points is run for 60 days, CODE-II experiment off the west coast of Vancouver
storage of all hourly time series requires 720 000 words per Island. (Figure 5a shows the locations of these sites.) The
variable. If on the other hand, the mean and 15 vertically-averaged model currents were obtained with
constituents are to be included in the harmonic analysis, the Walters 2° finite element method and the barotropic
storage of the right hand sides requires only 15 500 words. currents were calculated by forming weighted averages of
(Each right hand side entry is the summation of a series of observations (courtesy of Freeland 2~) at a minimum of
products of a sinusoid with a grid point value. See two depths. The lines within each ellipse are the phase lag
Foreman 9 for further details.) However a major g, measured in the same manner as shown in Fig. 2, and
disadvantage with storing right hand sides is that the the arrows denote the direction of rotation around the
constituents to be included in the analysis must be chosen ellipse. The K~ model ellipses are too thin because the
prior to running the model. In other words, the harmonic barotropic model cannot accurately reproduce the
analysis cannot be repeated with additional constituents clockwise rotary component of the current that is largely
unless the model run is also repeated. So when calculating due to a baroclinic shelf wave. Although the model ellipses
right hand sides during a model execution, one must for M 2 are more accurate, the discrepancies are still not
anticipate all the constituents that could be present in the acceptable. As the moorings are near the edge of a canyon,
model time series. it is likely that the model grid requires a refinement in
In comparing model time series with observations there order to more accurately represent the topography and
are several points of caution. Obviously it is important to the correct tidal behaviour. (See Foreman z2 for further
ensure that the two time series contain the same details.)
constituents. As demonstrated by Lynch and Werner 17, When comparing model results and observations on a
the presence of shallow water constituents generated in constituent by constituent basis, one may either compare

Adv. Water Resources, 1989, Volume 12, September 119

The harmonic cnalysis o f tidal model time series: M . G. G. Foreman and R. F. Henry

Table 5. Combinedstandard deviations (cm) for the harmonic analysis ACKNOWLEDGEMENTS

results with and without inference. Types A, B, C and test numbers 2 --* 4
and 6 --*8 are as in Table 3 We thank Gabriel G o d i n for helpful discussions; Roger
Flather and Roy Walters for the use of their model codes;
teat houra type Zo 01 K1 M2 $2 Patricia K i m b e r for drafting the figures; and the reviewers
2 60 A .000 .002 .002 .003 .003 for helpful comments.
3 60 B .05 3.65 4.18
4 60 C .05 .22 .31 .14 .07
6 60 A .24 1.14 1.02 .89 .88 / REFERENCES
7 60 B .25 3.82 4.23 1 Newton, I. Philosophia Naturalis Principia Mathematica, 1687
8 I 60 C .25 .40 .56 .26 .12 2 Laplace,P. S. Recherches sur plusieurs points du syst6me du
monde, M~moires de rAcadkmie royale des Sciences, 1775, 88,
75-182
3 Laplace, P. S. Recherches sur plusieurs points du syst6me du
amplitudes with amplitudes and phases with phases (as is monde, Mdmoires de rAcadbmie royale des Sciences, 1776, 89,
illustrated in Table 3), or form a complex-valued 177-267
amplitude and c o m p a r e both simultaneously. Each 4 Doodson, A. T. and Warburg, H. D. Admiralty Manual of Tides,
His Majesty's Stationery Office, London, 1941
a p p r o a c h has its advantages. Whereas the first a p p r o a c h 5 Darwin, G. H. Reports of a Committee,for the Harmonic Analysis
permits separate measures of the magnitude and timing of Tides, British Association for the Advancement of Science,
accuracy, the second a p p r o a c h condenses the error to one 1883 86
value, the distance between two complex numbers. W h e n 6 Harris, R. A. Manual of Tides, Appendices to Reports of the U.S.
Coast and Geodetic Survey, 1897-1907
average amplitudes and phases are c o m p u t e d from a 7 Doodson, A. T. The Harmonic Development of the Tide-
series of analyses (as in Table 3), it is safer to do the Generating Potential, Proc. Roy. Soc. Series A, 1921, 100,
calculation in complex form. This avoids any potential 306 323. Re-issued in the International Hydrographic Review,
difficulties due to the phase discontinuity at 360 °. Table 5 May 1954
shows the complex amplitude standard deviations 8 Godin, G. The Analyis of Tides, University of Toronto Press,
1972
associated with the h a r m o n i c analysis results shown in 9 Foreman, M. G. G. Manual Jot Tidal Heiyhts Analysis and
Table 3. Prediction, Pacific Marine Science Report 77-10, Institute of
As confidence intervals or ellipses can be c o m p u t e d Ocean Sciences, Patricia Bay, Sidney, B.C., 1977, Unpublished
during h a r m o n i c analyses (e.g., Fig. 1), it is logical not manuscript
10 Ortega, J. M. Numerical Analysis A Seeond Course, Academic
only to calculate the difference between model results and Press, New York, 1972, 193 pp.
observations, hut also to determine whether or not each 11 Draper, N. R. and Smith, H. Applied Re,qression Analysis, John
model result lies within a particular confidence region Wiley & Sons, New York, 1966
a r o u n d the observation. As discussed previously, due to a 12 Schureman, P. Manual of Harmonic Analysis and Prediction of
Tides, U.S. Department of Commerce Special Publication
c o m b i n a t i o n of high b a c k g r o u n d noise and a short
No. 98, Washington, 1958
observation period, these confidence regions m a y be 13 Munk, W. H. and Hasselmann, K. Super-resolution of Tides,
large. So even t h o u g h a model prediction m a y not seem to Studies on Oceanography (Hikada volume), Tokyo, 1964, 339-344
be close to an observation, it m a y in fact be well within an 14 Munk, W. H. and Bullard, E. C. Patching the long-wave
acceptable confidence region. spectrum across the tides, Journal qf Geophysical Research, 1963,
68(12), 3627-3634
Finally, it should also be mentioned that when using 15 Crawford, W. R. and Thomson, R. E. Diurnal Period Shelf
F o r e m a n ' s p r o g r a m s for analysing model time series, the Waves along Vancouver Island: A Comparison of Observations
input routines should be changed to acept more than 3 or with Theoretical Models, Journal of Physical Oceanoyraphy,
4 digits for each observation. R o u n d i n g to this n u m b e r of 1984, 14(10), 1629 1646
digits creates a slight non-tidal noise and m a y thus affect 16 Foreman, M. G. G. Manual jor Tidal Currents Analysis and
Prediction, Pacific Marine Science Report 78-6, Institute of
the length of record required to separate close Ocean Sciences, Patricia Bay, Sidney, B.C., 1978, Unpublished
constituents. R o u n d i n g will also affect the accuracy of the manuscript
analysis result. F o r example, an analysis of the time series 17 Lynch, D. R. and Werner, F. E. Long term simulation and
at forced b o u n d a r y points will not return exactly the same harmonic analysis of North Sea/English Channel tides. In:
Developments in Water Sciences, 35. Computational Methods in
amplitudes and phases as were used in the forcing. This is Water Resources. Volume I: Modeling Surface and Sub-Surface
illustrated by the Test 2 results in Table 3. Flows. Proceedings of the Vllth International Conference, MIT,
Boston, USA, June 1988. M. A. Celia et al., editors,
Computational Mechanics Publications, Elsevier, 1988, 257 266
SUMMARY 18 Flather, R. A. A Numerical Model Investigation of Tides and
Diurnal-Period Continental Shelf Waves along Vancouver
This paper has briefly reviewed the d e v e l o p m e n t of the Island, Journal of Physical Oceanography, 1988, 18(1), 115-139
h a r m o n i c analysis of tides and described the particular 19 Phillips, O. M. The Dynamics of the Upper Ocean, 2nd ed.,
h a r m o n i c analysis technique introduced by G o d i n 8 and Cambridge University Press, London, 1977
circulated by F o r e m a n 9. The related issues of constituent 20 Waiters,R. A. A Model for Tides and Currents in the English
Channel and Southern North Sea, Advances in Water Resources,
selection, time series length, confidence regions, and 1987, 10, 138-148
inference were discussed. It was also shown that since 21 Freeland, H. J. Derived Lagrangian Statistics on the Vancouver
numerical model time series have a m u c h lower non-tidal Island Continental Shelfand Implications for Salmon Migration,
content than observational time series, constituents can Atmosphere-Ocean, 1988, 26(2), 267 281
22 Foreman, M. G. G. Two Finite Element Tidal Models for the
be separated with shorter records. Finally several Southwest Coast of Vancouver Island, submitted to Proceedings
guidelines for the analysis of tidal model time series were of the International Conference on Tidal Hydrodynamics,
discussed. Gaithersburg, Maryland, Nov. 14~18, 1988

120 Adv. Water Resources, 1989, Volume 12, September