Computers & Geosciences 28 (2002) 929937

Classical tidal harmonic analysis including error estimates in

MATLAB using T TIDE$
Rich Pawlowicza,*, Bob Beardsleyb, Steve Lentzb
a

Department of Earth and Ocean Sciences, University of British Columbia, 6270 University Blvd., Vancouver, BC, Canada V6T 1Z4
b
Woods Hole Oceanographic Institution, Woods Hole, MA 02543 USA
Received 24 September 2001; received in revised form 13 February 2002; accepted 28 February 2002

Abstract
A standard part of any oceanic pressure gauge or current meter analysis is the separation of tidal from non-tidal
components of the signal. The tidal signal can either be discarded, or its characteristics described in some fashion useful
for further analysis. Although tidal signals can be removed by standard high or bandpass ltering techniques, their
relatively deterministic character and large amplitude make special techniques more effective. In classical harmonic
analysis, the tidal signal is modelled as the sum of a nite set of sinusoids at specic frequencies related to astronomical
parameters. A set of programs has been written in MATLAB to (a) perform classical harmonic analysis for periods of
about 1 year or shorter, (b) account for (some) unresolved constituents using nodal corrections, and (c) compute
condence intervals for the analyzed components. r 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Tides; Condence interval; Oceanography

1. Introduction
As the earth rotates on its axis, spatially varying
gravitational forces from the moon and the sun act on
the ocean, generating a forced elevation and current
response primarily (but not solely) at diurnal and semidiurnal frequencies. Body forces act directly on deep
oceanic waters. Tidal effects in coastal regions are not
directly forced by these astronomical forces. Instead
they arise as a side-effect of deep oceanic variability,
propagating through shallower coastal waters as a wave
or a combination of waves. In a typical oceanic time
series, tidal variability is often the largest signal. Power
spectra for such time series are often characterized by
a broad hump with a low-frequency maximum and a
decline at higher frequencies. Superimposed are a
$

Code available from server at http://www.iamg.org/CGEditor/index.htm or http://www.ocgy.ubc.ca/~rich.

*Corresponding author. Tel.: +1-604-822-1356; fax: +1604-822-6091.
E-mail address: [email protected] (R. Pawlowicz).

number of sharp tidal peaks near diurnal and semidiurnal frequencies, and sometimes a broader peak
associated with Coriolis or inertial effects. Dynamical
analysis requires the separation of the tidal signal from
sub or super-tidal variations, or in some cases separation
of tidal effects from inertial effects at a nearby
frequency. The tidal information is either discarded or
kept for further analysis.
Standard high/low/bandpass ltering techniques (e.g.,
Jackson 1986) can be used but in general these are
inefcient because fairly narrow lters with a great deal
of rejection are needed. Also, although these are useful
in analyzing non-tidal variability, they provide no
compression of the tidal information. Specialized
techniques have been devised to take advantage of the
deterministic nature of tidal processes. In classical
harmonic analysis, the tidal forcing is modelled as a set
of spectral lines, i.e., the sum of a nite set of sinusoids
at specic frequencies. These frequencies are specied by
various combinations of sums and differences of integer
multiples of 6 fundamental frequencies arising from
planetary motions (Godin, 1972). These fundamental

0098-3004/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 9 8 - 3 0 0 4 ( 0 2 ) 0 0 0 1 3 - 4

930

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

parameters represent effects of rotation of the earth

(lunar day of 24:8 h), the orbit of the moon around the
earth (lunar month of 27 days) and the earth around the
sun (tropical year), and periodicities in the location of
lunar perigee (8.85 years), lunar orbital tilt (18.6 years),
and the location of perihelion (E21 000 years). The set
of 6 signed integers required to describe a particular
frequency are called the Doodson numbers. Many of the
more important frequencies have names such as M2 ,
K1 , etc. From astronomical considerations alone, an
equilibrium response can be predicted; this is the
phase and amplitude that would be observed if
the response of the earth was fast enough that the
surface deformation was effectively in equilibrium with
the forcing at all times. The real ocean is denitely not in
equilibrium with the tidal forcing. However, as tidal
amplitudes are small compared with the total ocean
depth, the dynamics are very nearly linear, implying that
the forced response contains only those frequencies
present in the forcing. A least-squares t can be used to
determine the relative phase and amplitude of each
frequency in the response. This phase/amplitude data
thus provides a compression of the data in the complete
time series, which can then be compared with similar
data at other locations to understand the characteristics
of tidal dynamics, or can be used to synthesize time
series of tidal effects at other times for predictive
purposes.
There are several drawbacks to classical harmonic
analysis. The rst is that, ignoring the modulation of
perihelion which is effectively constant over historical
time, an E18:6 year time series is required to resolve all
of the listed frequencies (that is, the number of
wavelengths of each constituent in the record is at least
1 different from all other constituents). In practice,
record lengths are often 1 year or shorter. In order to
handle this issue an assumption is made that the phase/
amplitudes of response sinusoids with similar frequencies (i.e., those whose rst three Doodson numbers are
identical) are in the same proportion as those of the
equilibrium response under the reasonable premise that
the ocean response should be similar at similar
frequencies. In such a cluster large equilibrium peaks
are surrounded by small subsidiary peaks in frequencyspace which provide nodal modulations (or more
correctly, satellite modulations) to the main peak.
The appearance of the total signal will be a sinusoid
whose phase and amplitude varies slowly with time.
These changes are slow enough to be considered
effectively constant for record lengths of up to 1 year.
At much shorter record lengths another problem arises.
The frequency resolution further degrades until even
dissimilar constituents are unresolvable. The best solution is to apply inference. This technique for nding the
absolute phase/amplitude requires that the relative
differences in phase/amplitude between the two unre-

solved constituents is known from other nearby data. If

this is not the case, it is thought best to either discard the
smaller constituents and t only to the largest in a given
frequency interval, or to use the equilibrium response to
establish the desired differences.
Another drawback of classical analysis is that it
provides no easy way to determine whether the resulting
phase/amplitude of a given sinusoid is meaningful in a
deterministic way (i.e., it is truly a tidal line), or whether
it results from tting to a component of the non-tidal
broad-spectrum variability. In general a t is likely to
include elements of both and some kind of condence
interval for the deterministic part is useful. To address
this issue, the response method was invented (Munk
and Cartwright, 1966). Although this provides better
results than classical harmonic analysis, it has not found
widespread use.
Further problems with classical harmonic analysis
arise in coastal regions where the tidal response is in the
form of a wave propagating onshore. In large estuaries,
the seasonal change in salinity and ow may change the
dynamic response but as these changes can vary from
year to year the tidal process is not really stationary.
Instead spectral peaks are broadened so that they are no
longer pure lines, but, depending on the situation, such
variations may be treated as lines in the analysis. Within
smaller estuaries, tidal height variations may be signicant compared to water column depth and a variety
of non-linear effects can occur. For example, ood
periods shorten and intensify and ebbs lengthen. As long
as these effects are reasonably deterministic they may be
handled by adding extra shallow water constituents
which occur at sum/difference frequencies of the major
constituents. More problematic in these regions are the
effects of internal variability. Tidal interactions with
varying topography can produce large internal waves
and bores whose characteristics are highly sensitive to
ambient stratication. In such cases the assumption of
line frequencies becomes questionable and other
techniques such as wavelet analysis have been suggested
(Jay and Flinchem, 1999). More comprehensive descriptions of analysis techniques, their use, and their
limitations is given in, e.g., Foreman et al. (1995) and
Godin (1991).
Here, we describe T TIDE, a package of routines that
can be used to perform classical harmonic analysis with
nodal corrections, inference, and a variety of userspecied options. Predictions can also be made using the
analyzed constituents. There are several novel features.
First, although the harmonic analysis algorithm with
nodal corrections, etc., itself is not original (other than
conversion to complex algebra), it is implemented in
MATLAB, an analysis package widely used by oceanographers. This allows for easy use within the framework
of a complete analysis involving plotting of raw data,
scatter plots, and so forth. Second, the code is written

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

directly in matrix terms and thus relatively easy to

understand and modify if required. Finally, in order to
differentiate between true deterministic (line) frequencies
and broad-spectrum variability, condence intervals for
the estimated tidal parameters are computed using one
of several user-selectable algorithms. The package is
made up of a number of les each of which contain one
or more functions. User-callable functions generally
have a t prex to prevent namespace collisions.
The paper is composed of four parts. In the rst the
form of the equilibrium potential is described. In the
second part the mathematical basis of the technique for
making phase/amplitude estimates is described. In the
third part the generation of condence intervals is
outlined. Finally an illustrative example is discussed.

2. Tidal potential
The effect of gravitational force vectors from the sun
and moon, F; can be written as the gradient of a scalar
potential V ; F rV : The magnitude of this potential
at the earths surface at any time obviously is dependent
on the relative positions of the earth, moon, and sun. In
Doodsons development (Doodson, 1954, described in
Godin, 1972) the potential is written as a function of
lunar time t (dened to begin at lunar midnight) and
other astronomical variables (which are also functions
of time):
s is the mean longitude of moon, h the mean longitude
of sun, p the longitude of perigee, N 0 the negative of
the longitude of the ascending node, and p0 is the
longitude of perihelion, where all terms are in units of
cycles. These variables can be evaluated for a given
Julian date using the function t astron which implements
formulas in Seidelmann (1992). Their effects are
combined with the Doodson number set for a particular
constituent fi0 ; j 0 ; k0 ; l 0 ; m0 ; n0 g into the astronomical
argument Va i0 t j 0 s k0 h l 0 p m0 N 0 n0 p0 : Sets
with a common i0 are called a species (thus the slow,
diurnal, and semidiurnal species for i0 0; 1; and 2;
respectively), and sets with common i0 j 0 k0 are called a
subgroup. The constituent frequency s is dened as s
2p dVa =dt: The tidal potential is then written in the form
"
3
X
X
Gi0 y
A0i0 j 0 k0 l 0 m0 n0 cos2pVa
V
i0 0

Gi00 y

j 0 k0 l 0 m0 n0

B0i0 j 0 k0 l 0 m0 n0

#
sin2pVa :

j 0 k0 l 0 m0 n0

For a given Doodson number set either A0 or B0 is nonzero, but not both. These constants are tabulated and
stored in data structures that can be loaded using
t getconsts. The geodetic functions Gi0 and Gi00 vary with
species i0 and latitude y; and also depend on such

931

constants as the radius of the earth and the masses and

separations of the earth, moon, and sun. The equilibrium amplitude for a constituent is dened as
GA0 =g or G0 B0 =g; where g is the gravitational acceleration, and can be generated for a particular latitude using
t equilib.

3. Phase/amplitude estimates
The algorithm used here for making phase and
amplitude estimates is based on algorithms and FORTRAN code described by Godin (1972), Foreman
(1977), and Foreman (1978). However, unlike those
authors we use complex algebra directly rather than deal
with sine and cosine ts separately. This has the
advantage of unifying the treatment for scalar (e.g.,
pressure) and vector (e.g., horizontal currents) time
series which are represented as complex numbers u iv:
Note that the complex form for currents is based on a
physical model of a rotating current vector, and is only
valid for linear or nearly linear tidal waves. In some
cases it may be better to treat, e.g., along and acrosschannel currents as two separate scalar time series.
Consider a time series of observations yt; t
t1 ; t2 ; y; tM arranged in a vector, where the observation
times are regularly spaced at an interval Dt (default 1 h)
and M is an odd number (an endpoint is discarded if
required). The time axis is dened such that the origin
(or central time) is at tM1=2 : Some missing observations can be handled by using a missing data marker
in the input vector (by MATLAB convention this is
NaN, the IEEE arithmetic representation for Not-aNumber). This regular interval restriction does not arise
from the least-squares t itself but rather from the
automated constituent-selection algorithm and is also a
requirement when spectra are estimated in one of
the condence interval algorithms. This time series
may be composed of either real or complex numbers.
The time series is passed to the analysis program t tide
along with a variety of (mostly optional) parameters.
The tidal response is modelled as
X
xt b0 b1 t
ak eisk t ak eisk t ;
2
k1;y;N

where N constituents (each with unique Doodson

number sets) are used. Each constituent has a frequency
sk which is known from the development of the
potential, and a complex amplitude ak which is not
known, although if yt is a real time series ak and ak
are complex conjugates. A possible offset and (optional)
linear drift is handled by the rst two terms. The
traditional approach uses real sinusoids:
X
Ak cossk t Bk sinsk t
3
xt b0 b1 t
k1;y;N

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

932

and is related to Eq. (2) by Ak ak ak and Bk

iak  ak : The real representation is more convenient
for the linear error analysis described later.
Constituents can be chosen from a list of 45
astronomical and 101 shallow-water constituents. Data
structures containing names and other information
about these constituents are loaded using t getconsts.
There are several alternatives for selecting constituents.
For general use, an automated selection algorithm
(following Foreman, 1977) is in place, which works as
follows. A basis of all astronomical and 24 of the most
important shallow-water constituents are gathered
together. All constituents are listed in order of predened importance based on equilibrium amplitudes.
Less important constituents whose frequencies are less
than a Rayleigh resolution limit aNDt1 (with default
a 1) apart from more important constituents in
frequency are discarded. Additional shallow-water constituents can be specied if required. If the relative
phase/amplitude of two constituents that are otherwise
unresolvable is known from other sources, then an
inference procedure can be carried out. Alternatively,
constituent lists can be explicitly specied.
The least-squares t are the coefcients minimizing
X
E
jxtm  ytm j2 jjTa  yjj2 ;
4
m

where y yt1 ; yt2 ; y; ytM 0 ; a b0 ; b1 ; a1 ; a1 ; a2 ;

a2 ; y; aN 0 ; and T is an M 2N 2 matrix of linear
and sinusoidal basis functions evaluated at observation
times. The solution is found using the Matlab W
matrix division operator.
Once the t has been performed, various corrections
are applied. These are generated in t vuf. First, the phase
of the constituent response is usually reported as
Greenwich phase gk ; that is, phase referenced to the
phase of the equilibrium response at 01 longitude (the
Greenwich meridian). This can be interpreted as
reporting the phase of the response at the time when
the equilibrium forcing is at its largest positive value at
01 longitude. It is simplest to nd the tted phase at the
central time of the record t 0; the equilibrium phase
vk is then just Va for the given constituent computed at
the Julian date corresponding to this central time, with
possible adjustments of 14; 12; or 34 cycle depending on
whether A or B is non-zero, and their signs.
Second, if a latitude is specied, then nodal or satellite
corrections are computed as follows. Consider a main
peak of index k with satellites with indices kj: The effect
of the different satellites will be to slowly modulate the
phase/amplitude of the main peak over various periods,
usually more than 8 years. Our tted response a# k over
some period can then be written as a modication of the
true response of the main constituent ak ; in which the
amplitude is changed by a factor fk and the phase by an
angle uk due to the presence of the satellites. fk and uk

are called the nodal correction amplitude and phase,

respectively. That is,
X
akj eiskj t :
5
a# k eisk t fk ak eisk tiuk ak eisk t
j

Cancelling common terms, we have

X akj
X akj
fk eiuk 1
eiskj sk t E1
:
a
ak
k
j
j

The nal approximation will hold as long as skj  sk t

remains small (i.e., Ndt58 years). In general the true
phases and amplitudes of the satellites are not known.
However, since their frequencies are very similar to that
of the main peak it is standard to assume the ratio of
true amplitudes is the same as the ratio of amplitudes in
the equilibrium response, and the difference in true
phases will be equal to the difference in equilibrium
phases. The nodal corrections are thus computed from
the equilibrium response Eq. (1). A latitudinal dependence arises from the geodetic functions. G10 is zero at
the equator and a crude limiting is used to prevent some
diurnal corrections from getting overly large. The
validity of using the latitude-dependent equilibrium
response to predict an aspect of the dynamic behavior
in one part of an ocean basin in such a simple way is not
clear. If the record length is longer than 1 year the
comparison of successive 1 year analyses with and
without nodal correction can be used to test the validity
of this process. Note that if the time series to be analyzed
is longer than 18.6 years in length then the true
satellite amplitude/phase terms can be estimated directly
(Foreman and Neufeld, 1991) but this is not currently
possible in T TIDE.
The products of the analysis above are a pair of
complex values fak ; ak g; possibly corrected for nodal
modulation, for each constituent k: These are converted
into standard parameters:
Lk jak j jak j;

lk jak j  jak j;

yk

angak angak
mod 180;
2

gk vk  angak yk :

9
10

For horizontal currents, these parameters describe the

features of an ellipse traced out by the tip of the velocity
vector: the length of the semi-major and semi-minor axis
of the ellipse (Lk ; lk ; respectively), the inclination of the
northern semi-major axis counter-clockwise from due
east yk ; and the Greenwich phase gk : If lk > 0 o0 then
the ellipse is traced in a counter-clockwise (clockwise)
direction. For scalar time series, the parameter Lk is the
amplitude, and lk ; yk  0 (the ellipse degenerates to a
line along the positive axis). Note that the restriction of
the denition of inclination to the northern axis (via the

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

modulo operator) means that analyses of constituents

whose ellipses are aligned in an east/west direction may
have inclinations that uctuate between close to 01 and
nearly 1801 due to noise. These apparently large jumps
are an artifact of the restriction but do not represent
similarly large changes in the physical behavior.
Once ellipse parameters are found, these can be used
for further analysis. They can also be used to generate
predictions at other times using t predic. Nodal corrections in t predic are computed at the time series midpoint so that it is an exact inverse of t tide.

4. Condence intervals
One drawback of classical harmonic analysis is that
the degree to which a given constituent represents true
tidal energy as opposed to the energy of a broad-band
non-tidal process is not determined. This is useful
information for two reasons: rst, it allows one to make
better estimates of the tidal behavior, and second, it can
allow one to quantitatively compare different analyses.
There are two steps to producing condence intervals.
First, we must form an estimate of the characteristics of
non-tidal or residual noise affecting the ak (or Ak ; Bk ).
Second, we must convert these estimates into condence
intervals for the standard parameters through a nonlinear mapping. We discuss the situation of real time
series rst.
4.1. Residual noise (real)
After the harmonic analysis for an N-point real time
series yt is performed, we examine the structure of
the residual series. In the simplest situation, the residuals
are statistically Gaussian and uncorrelated in time. If
this is the case then the total residual power PT s2x
P=Dt; where P is the two-sided spectral density. The
amplitude of the t to sine and cosine terms (A and B;
respectively) will be contaminated by errors arising from
unresolved noise components within a frequency interval of Df NDt1 around the line. Thus s2A s2B
PDf s2x =N: It is unlikely that a geophysical series will
be spectrally at, and a more sophisticated approach
used in t tide is to nd a local value of P suitable for
constituents in that neighborhood by making a spectral
estimate from the residual time series (i.e., after the
removal of all tted constituents) and averaging
the power over frequency bins in a window around the
frequency of any constituent, neglecting bins in which
tted constituents reside. Here we chose a sequence of
windows of width 0.4 cpd centered on 1; 2; 3; y cpd
(actually on multiples of the M2 frequency, see the code
for details). The value of P appropriate to, say, semidiurnal constituents would be estimated from the second
of these bins.

933

4.2. Conversion to standard parameters (real)

A conversion from errors in the cos/sine amplitudes to
errors in standard parameters (amplitude and phase) can
be done using a linearized analysis. Consider a
constituent k: Let x F Ak ; Bk be a non-linear function
of these parameters, either the amplitude or the Greenwich phase. Then if fAk ; Bk g are independent random
variables, we can nd a linearized estimate of the
standard error of x in terms of the standard errors of the
sinusoid amplitudes:




@F 2 2
@F 2 2
s2x
sA
sB ;
11
@Ak
@Bk
where the partial derivatives can be derived exactly (but
tediously) from Eqs. (7)(10).
Alternatively the non-linear mapping can be handled
directly using a parametric bootstrap (Efron and
Tibshirani, 1993). In this situation the residual variance
estimates are used by the code to simulate a number of
realizations or replications of the analysis by taking the
estimates of the sinusoid amplitudes and adding
Gaussian noise with the appropriate variance to them.
All of these realizations are then converted non-linearly
to standard parameters using Eqs. (11)(10) and an
estimate of the standard error computed from this
replicate data set directly.
Once a standard error is determined, 95% condence
intervals can be estimated using standard techniques.
Alternatively, a signal-to-noise power ratio (SNR) can
be computed based on the square of the ratio of
amplitude to amplitude error. Simulations performed in
t synth (and described in the text le t errors) in which
the variability of analyses carried out on a xed data set
with different noise realization are compared with
estimated condence intervals show that the linear
procedure appears to be adequate for real time series
(e.g. tidal height), as long as the SNR > 10; and is
probably not bad for SNR as low as 2 or 3. The nonlinear procedure gives similar results to the linearized
procedure at high SNR, and is more accurate at low
SNR.
4.3. Residual noise (complex)
A complex residual time series u iv can be modelled
as bivariate white noise and variances s2u ; s2v and
covariance suv computed. If we assume further that the
noise in both components is not correlated suv E0; then
a coloured bivariate noise model can be used and
variances assigned to real and imaginary parts of
constituent amplitude separately on the basis of local
spectral densities as described above. If it is suspected
that suv a0 then it is recommended that the time series
be rotated into a coordinate system in which this is true

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

934
6

Tuktoyuktuk Elevations 1975

Elevation (m)

5
4
3
2
1
0
190

200

210

220
Yearday

230

240

250

210

220
Yearday

230

240

250

(A)
6

Tidal and residual series

Elevation (m)

4
2
0
2
190

(B)

200

10
Amplitude (m)

Analysis with 95% significance level

1

10

10

10

Greenwich Phase (deg)

(C)
360

0.4

Phase for significant constituents

270
180
90
0
0

(D)

0.2
frequency (cph)

0.2
frequency (cph)

0.4

Fig. 1. Tuktoyuktuk analysis example. (A) Raw time series. (B) Upper curve is residual time series after removal of tidal signal. Lower
curve is synthesized tidal series using signicant constituents. (C) Amplitude of all analyzed components with 95% signicance level.
Note frequency dependence. Signicant constituents are marked with solid circle. (D) Phase of signicant constituents with 95%
condence interval.

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

Table 1
Analysis results for Tuktoyuktuk data set
File name: PAPEROUT.txt
Date: 17-Aug-2001
Nobs 1584; ngood 1510; record length (days)66.00
Start time: 06-Jul-1975 01:00:00
Rayleigh criterion 1:0
Greenwich phase computed with nodal corrections applied to amplitude and phase relative to center time
x0 1:98; x trend 0
Varx 0:82196 varxp 0:21224 varxres 0:60972
Percent var predicted 25:8%
Tidal amplitude and phase with 95% CI estimates
Tide
Freq
Amp
Amp err
Pha
Pha err
Snr
MM
0.00151
0.2121
0.503
263.34
161.41
0.18
MSF
0.00282
0.1561
0.526
133.80
188.82
0.088
ALP1
0.03440
0.0152
0.044
334.95
150.82
0.12
2Q1
0.03571
0.0246
0.044
82.69
106.21
0.31
Q1
0.03722
0.0158
0.045
65.74
160.30
0.12
* O1
0.03873
0.0764
0.055
74.23
43.35
1.9
NO1
0.04027
0.0290
0.035
238.14
74.68
0.69
* P1
0.04155
0.0465
0.045
71.88
70.96
1.1
* K1
0.04178
0.1405
0.059
64.81
23.49
5.7
J1
0.04329
0.0253
0.050
7.32
129.76
0.25
OO1
0.04483
0.0531
0.059
235.75
72.96
0.81
UPSI
0.04634
0.0298
0.055
91.73
137.06
0.29
EPS2
0.07618
0.0211
0.030
184.59
104.65
0.51
* MU2
0.07769
0.0419
0.034
83.23
48.82
1.5
* N2
0.07900
0.0838
0.035
44.52
25.54
5.9
* M2
0.08051
0.4904
0.035
77.70
4.51
1.9e02
L2
0.08202
0.0213
0.037
35.22
113.22
0.33
* S2
0.08333
0.2197
0.038
126.72
9.14
34
* K2
0.08356
0.0598
0.043
149.12
46.60
2
ETA2
0.08507
0.0071
0.033
246.05
207.25
0.048
* MO3
0.11924
0.0148
0.014
234.97
67.38
1.1
M3
0.12077
0.0123
0.014
261.57
62.11
0.81
MK3
0.12229
0.0049
0.012
331.60
144.92
0.18
SK3
0.12511
0.0023
0.010
237.69
219.86
0.054
MN4
0.15951
0.0092
0.011
256.47
69.76
0.68
* M4
0.16102
0.0126
0.011
291.78
65.09
1.4
SN4
0.16233
0.0083
0.011
270.85
91.22
0.54
MS4
0.16384
0.0010
0.008
339.35
248.82
0.015
S4
0.16667
0.0047
0.010
299.56
142.32
0.23
2MK5
0.20280
0.0013
0.005
310.10
181.03
0.067
2SK5
0.20845
0.0045
0.006
104.00
99.71
0.64
2MN6
0.24002
0.0035
0.007
271.24
133.30
0.22
M6
0.24153
0.0017
0.006
158.88
197.43
0.093
2MS6
0.24436
0.0056
0.008
306.10
90.03
0.54
2SM6
0.24718
0.0023
0.007
298.92
175.13
0.11
* 3MK7
0.28331
0.0086
0.006
212.25
44.21
2.1
MS
0.32205
0.0030
0.004
42.43
75.29
0.55
M10
0.40256
0.0009
0.003
198.23
209.99
0.089

935

936

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

(e.g., into along/across channel axes or into the principal

axes).
4.4. Conversion to standard parameters (complex)
The linearized analysis now involves functions of four
variables, since both Ak Ar iAi and Bk Br iBi
have real and imaginary parts, and analytic expressions
for partial derivatives


 2
 2
@F 2 2
@F
@F
sAr
s2Ai
s2Br
s2x
@Ar
@Ai
@Br
 2
@F

s2Bi
12
@Bi
become large. Some analytical simplication is possible
by assuming that all four variables are independent.
The bootstrap approach can also be applied to the
complex coefcients ak : One minor complication that
arises is that unless the noise is circular s2u s2v ; suv
0; the errors in ak and ak are correlated with each
other. The bootstrap process requires the generation of
correlated noise replicates.

5. Example
The analysis of an example data set provided in
Foreman (1977) is shown in Fig. 1 and Table 1. This
example is included in datale t example.mat. The
example data set consists of 66 days of hourly elevations
with a 3 day gap. A tidal variation is visible superimposed on subtidal variability. The time series can be
loaded and analyzed using the demonstration script
t demo. Code within this script illustrates how the
programs are called. In this example, the automated
constituent selection algorithm is used and it selects 35
constituents. In addition, one shallow water constituent
M10 is manually added and two other constituents
analyzed via inference. The P1 constituent is inferred
from K1 ; and K2 is inferred from S2 : Nodal corrections
are performed. A linear trend is not included in the
analysis. The coloured bootstrap analysis is used to
determine signicance and condence intervals. Table 1
gives the output of the program. In the rst column the
name of the constituent is given. Signicant constituents
(those with SNR in the last column > 1) are marked with
a *. The SNR is the squared ratio of amplitude (third
column) to the error in amplitude (fourth column). The
error factors (and hence SNR) will change slightly in
repeated analyses due to the stochastic nature of the
bootstrap procedure but amplitudes and phases themselves will be invariant. Frequencies (rst column) are
listed in cph and Greenwich phase/phase error (fth and
sixth columns) in degrees. Eleven constituents were
judged to be signicant (only 6 would be signicant at

an SNR cutoff of 2). Fig. 1B shows the residual series

and an elevation series synthesized from the signicant
constituents (note that it spans the 3-day gap). Results
of the analysis are shown in a spectrum in Fig. 1C. Most
of the signicant constituents are in the diurnal and
semidiurnal bands (E0:04; and E0:08 cph; respectively)
although several higher-frequency constituents also
appear to be marginally signicant. In spite of the large
amount of energy in the fortnightly band E0:002 cph
the tted constituents there are apparently not signicant. The analyzed phases of signicant constituents are
shown in Fig. 1D. The signicant constituents generally
have reasonably small phase errors.

6. Summary
Separation of tidal and non-tidal energy is an
important task in any analysis of oceanic time series.
Here, we discuss the theoretical foundation and implementation details of a MATLAB package for
classical harmonic analysis. The package can also
compute condence intervals for the tidal parameters
using one of three different sets of assumptions about
the structure of residual noise. An example is provided
to show typical results. The code is available at http://
www.ocgy.ubc.ca/~rich, or the IAMG Server.

Acknowledgements
Elements of this MATLAB toolbox for tidal harmonic analysis were developed over the last few years
as part of the authors research into tidal phenomena.
We want to rst acknowledge Mike Foremans major
contribution in providing the FORTRAN codes
and documentation to the community in 197778 that
served as our starting point. Rich Signell rst derived
the formulae for the linearized error analysis used here,
and Julio Candela and Jim Irish also provided useful
input. Support for this effort was provided by Natural
Sciences and Engineering Research Council of Canada
under grant OGPO194270 (RP) and the National
Science Foundation and Ofce of Naval Research (SL
and RB).

References
Doodson, A.T., 1954. Appendix to circular-letter 4-H. The
harmonic development of the tide-generating potential.
International Hydrographic Review 31, 3761.
Efron, B., Tibshirani, R.J., 1993. An Introduction to the
Bootstrap, Vol. 57. Monographs on Statistics and Applied
Probability. Chapman & Hall, New York, 436pp.

R. Pawlowicz et al. / Computers & Geosciences 28 (2002) 929937

Foreman, M.G.G., 1977. Manual for tidal heights analysis and
prediction. Pacic Marine Science Report 77-10, Institute of
Ocean Sciences, Patricia Bay, Sidney, BC, 97pp.
Foreman, M.G.G., 1978. Manual for tidal currents analysis and
prediction. Pacic Marine Science Report 78-6, Institute of
Ocean Sciences, Patricia Bay, Sidney, BC, 57pp.
Foreman, M.G.G., Crawford, W.R., Marsden, R.F., 1995. Detiding: theory and practise. In: Lynch, D.R., Davies, A.M.
(Eds.), Quantitative Skill Assessment for Coastal Ocean
Models. Vol. 47. Coastal and Estuarine Studies. American
Geophysical Union, pp. 203239.
Foreman, M.G.G., Neufeld, E.T., 1991. Harmonic tidal
analyses of long time series. International Hydrographic
Review 68 (1), 85108.
Godin, G., 1972. The Analysis of Tides. University of Toronto
Press, Toronto, 264pp.

937

Godin, G., 1991. The analysis of tides and currents. In: Parker,
B.B (Ed.), Tidal Hydrodynamics. Wiley, New York,
pp. 675709.
Jackson, L.E., 1986. Digital Filters and Signal Processing.
Kluwer Academic Publishers, New York, 259pp.
Jay, D.A., Flinchem, E.P., 1999. A comparison of methods for
analysis of tidal records containing multi-scale non-tidal
background energy. Continental Shelf Research 19,
16951732.
Munk, W.H., Cartwright, D.E., 1966. Tidal spectroscopy and
predication. Philosophical Transactions of the Royal
Society of London, Series A 259, 533581.
Seidelmann, P.K. (Ed.), 1992. Explanatory Supplement to
the Astronomical Almanac. United States Naval Observatory, Mill Valley, Calif., University Science Books.
752pp.