Unified Tidal Analysis and Prediction

Using the UTide Matlab Functions

Daniel L. Codiga

Graduate School of Oceanography

University of Rhode Island

September 2011

GSO Technical Report 2011-01

Funded by National Science Foundation, Physical Oceanography Program,

Award 0826243, “Investigating Tidal Influences on Subtidal Estuary-Coast
Exchange Using Observations and Numerical Simulations”

Full citation:
Codiga, D.L., 2011. Unified Tidal Analysis and Prediction Using the UTide Matlab
Functions. Technical Report 2011-01. Graduate School of Oceanography,
University of Rhode Island, Narragansett, RI. 59pp. ftp://www.po.gso.uri.edu/
pub/downloads/codiga/pubs/2011Codiga-UTide-Report.pdf
Abstract
A unified tidal analysis and prediction framework is developed. A self-consistent
and complete set of equations is presented that incorporates several recent advances, with
emphasis on facilitating applicability to the case of irregularly distributed times, and
includes as special cases nearly all prior methods. The two-dimensional case treated is
suitable for ocean currents, yields current ellipse parameters, and naturally reduces to the
one-dimensional case suitable for sea level. The complex number formulation is used for
matrix solution but relationships to the real formulation, needed for confidence interval
estimation with irregular times, are included. The two-dimensional generalization of
Foreman et al. (2009) leads to expressions (including in-matrix treatment instead of post-
fit corrections) incorporating exact times in nodal/satellite corrections and in calculation
of Greenwich phase from the astronomical argument, as well as exact constituent
inference. Some of the resulting capabilities include accurate nodal/satellite corrections
for records longer than 1-2 years, and inference of multiple constituents from a single
reference. A comprehensive set of constituent selection diagnostics is summarized.
Diagnostics to assess constituent independence are the conventional Rayleigh criterion
and its noise-modified variant, the basis matrix condition number relative to the all-
constituent signal-to-noise ratio (SNR), and a newly defined maximum correlation
between model parameters; diagnostics to assess constituent significance are the SNR
and percent energy. A confidence interval estimation method for current ellipse
parameters, based on complex bivariate normal statistics, is presented that generalizes the
colored Monte Carlo method of Pawlowicz et al (2002): the model parameter covariance
matrix is not constrained to a presumed form and is scaled using both auto- and cross-
spectra of the residual, as computed by fast Fourier transform or Lomb-Scargle
periodogram in the case of regularly or irregularly distributed times respectively.
Descriptions are provided for the functionality and syntax of a pair of Matlab
functions denoted “UTide”—ut_solv() and ut_reconstr()—that implement the unified
analysis and prediction framework. Output of ut_solv() includes a table of all diagnostics,
organized to make constituent selection efficient. The robust iteratively-reweighted least
squares (IRLS) L1/L2 solution method, explored by Leffler and Jay (2009) for the one-
dimensional case with uniformly distributed times, is used because it limits sensitivity to
outliers and can substantially reduce confidence intervals. Prior methods (for example,
capabilities of the t_tide Matlab package of Pawlowicz et al. (2002), including the
automated decision tree of Foreman (1977) for constituent selection) are available using
option flags: ordinary least squares can be used (instead of IRLS); nodal corrections
and/or Greenwich phase lag calculations can be omitted, or carried out using linearized
(instead of exact) times; inference can use the traditional approximate method (instead of
the exact formulation); and confidence intervals can be estimated using the linearized
method (instead of Monte Carlo simulation), and/or using the white noise floor
assumption (instead of scaled by colored residual spectra). Reconstructed superposed
harmonic fits (hind-casts or forecasts/predictions) can be generated by ut_reconstr() at
arbitrarily chosen sets of times, using subsets of constituents (for example, based on
meeting a SNR threshold, or as specified by the user). Finally, the same treatment can be
applied to each record in a group of records—such as observations from multiple buoy
sites and/or multiple depths, or numerical simulation output from multiple model grid
nodes—by a single execution of ut_solv() and ut_reconstr().

1
Table of contents
Abstract ........................................................................................................... 1
I. Introduction ................................................................................................. 4
II. Unified equation development ................................................................... 8
II.A. Model equations ..................................................................................................... 8
II.A.1. General case: complex, two-dimensional........................................................ 8
II.A.1.a. Pre-filtering and nodal/satellite corrections; Greenwich phase lags ........ 9
II.A.1.b. Current ellipse parameters...................................................................... 10
II.A.1.c. Constituent inference.............................................................................. 11
II.A.1.d. Summary ................................................................................................ 13
II.A.2. One-dimensional case, complex.................................................................... 13
II.A.3. Relations to real formulation......................................................................... 14
II.A.4. Prior methods as special cases....................................................................... 16
II.A.4.a. Nodal/satellite corrections using linearized times .................................. 16
II.A.4.b. Greenwich phase lags: linearized times in astronomical argument ....... 17
II.A.4.c. Approximate inferences.......................................................................... 17
II.B. Matrix formulation and solution method.............................................................. 18
II.B.1. Iteratively reweighted least squares robust solution...................................... 19
II.C. Confidence intervals ............................................................................................. 20
II.C.1. Complex bi-variate normal statistics ............................................................. 22
II.C.2. White noise floor case with non-zero cross-correlations............................... 23
II.C.3. Colored case using spectra of residuals......................................................... 26
II.C.4. Implementation: Non-reference and reference consituents ........................... 28
II.C.5. Implementation: Inferred constituents........................................................... 29
II.C.6. One-dimensional case.................................................................................... 30
II.D. Constituent selection diagnostics ......................................................................... 30
II.D.1. Diagnostics related to constituent independence........................................... 31
II.D.1.a. Conventional Rayleigh criterion (RR)..................................................... 31
II.D.1.b. Noise-modified Rayleigh criterion (RNM) .............................................. 32
II.D.1.c. Condition number (K) relative to SNR of entire model (SNRallc).......... 32
II.D.1.d. Maximum correlation (Corrmax) between model parameters.................. 33
II.D.2. Diagnostics related to constituent significance ............................................. 35
II.D.2.a. Signal to noise ratio (SNR)..................................................................... 35
II.D.2.b. Percent energy (PE)................................................................................ 36
II.D.3. Diagnostics characterizing reconstructed fits (PTVall, PTVsnrc).................... 36
II.D.4. Considerations for irregularly distributed times............................................ 37
III. The UTide Matlab functions................................................................... 39
III.A. Obtaining and using UTide ................................................................................. 39
III.B. Quick start suggestions........................................................................................ 39
III.C. Functionality and syntax and for a single record ................................................ 40
III.C.1. Solving for coefficients with ut_solv()......................................................... 40
III.C.1.a. Input parameter descriptions ................................................................. 40
III.C.1.b. Output structure coef ............................................................................. 41
III.C.1.c. Defaults and options .............................................................................. 44
III.C.1.d. Summary diagnostics table and diagnostic plots................................... 48
III.C.2 Reconstructing fits with ut_reconstr() .......................................................... 50

2
 III.C.2.a. Input and output parameter descriptions ............................................... 50
III.C.2.b. Defaults and options.............................................................................. 50
III.D. Functionality and syntax for groups of records .................................................. 51
III.E. Relationships to existing software....................................................................... 54
III.F. Computational demands ...................................................................................... 55
IV. Acknowledgements ................................................................................ 57
V. References................................................................................................ 58

3
I. Introduction
Development of this unified tidal analysis was motivated mostly by the need to
carry out tidal analysis on a multi-year sequence of current observations collected at
irregularly spaced times (Codiga 2007). Observational datasets with these characteristics,
less common not long ago, are increasingly available; observing system developments
mean longer field campaigns are being sustained, and it is also typical for them to have
substantial gaps and/or irregularly distributed temporal sampling. However, it is widely
recognized that many commonly available standard software packages for tidal analysis,
while highly sophisticated and mature in many ways, require special treatment for such a
dataset.

Throughout the following, reference is made to the topics of nodal/satellite

corrections, computation of Greenwich phase using the astronomical argument, and
constituent inference. These issues are explained comprehensively in numerous
publications, including the textbook of Godin (1972) (G72), so they will not be reviewed
here except superficially. Readers unfamiliar with them are referred to Foreman and
Henry (1989) (FH89) and Parker (2007) as examples of accessible entry points to the
literature.

One reason there are limitations to the applicability of traditional tidal methods to
multi-year records is that results of the standard (linearized times) method for the
nodal/satellite corrections becomes inaccurate for records longer than a year or two (e.g.
FH89). This necessitates breaking the record into subsets with durations of about a year,
subjecting each to separate analyses, and then combining the results in a final step, for
which there seems not to be a standard practice. Another limitation of nearly all standard
methods is the requirement of uniformly distributed temporal sampling. For irregular
temporal sampling, while an effective approach has recently been developed for the one-
dimensional case (e.g. sea level) (Foreman et al. 2009) (FCB09), constituent selection
methods remain less well-defined, suggesting the need for new diagnostics. In addition,
to the extent that the solution method or confidence interval calculation relies on auto-
and/or cross-spectral quantities, in the case of irregular times the fast Fourier transform
(FFT) relied on by some methods cannot be applied. A Lomb-Scargle least squares
spectral estimation approach (e.g., Press et al. 1992) is suitable for this but has not been
implemented. These issues are all addressed here.

The primary goal is to develop a tidal analysis approach and accompanying

software tool (“UTide”) that (a) integrates several existing tidal analysis methods with
each other (Table 1), and (b) includes enhancements specifically designed to enable
treatment of multi-year records with irregular temporal sampling. The main foundation is
the theory for harmonic analysis laid out by G72, then extended for practical applications
by (Foreman 1977; 1978) (F77, F78) with accompanying Fortran programs, and further
discussed by FH89. This foundation was coded in to Matlab as the “t_tide” package by
Pawlowicz et al. (2002) (PBL02), which has become a widely accepted standard utility in
the physical oceanographic community. PBL02 added a confidence interval estimation
method that could use the spectral characteristics of the residual instead of presuming a

4
Table 1. Comparison of features of UTide and prior software products.
F77, F78 PBL02 LJ09 FCB09 UTide
Nodal/satellite Post-fit, (same) (same) In-matrix, In-matrix;
corrections linearized exact times Default exact
times times; Optional
linearized times
Astronomical Post-fit, (same) (same) In-matrix, In-matrix;
argument for linearized exact times Default exact
Greenwich times times; Optional
phase linearized times
Consituent Post-fit, (same) (same) In-matrix, In-matrix;
inference approx. exact Default exact;
Optional approx.
Missing-data Linearly (same) (same) In-matrix Default in-
gaps (in interp’d (missing matrix; For
regular time points regular temporal
grid) omitted) sampling,
optional linearly
interpolated
Irregular times No (same) (same) In-matrix In-matrix
Confidence Cosine/ Current (same) Cosine/ Current ellipse
interval sine ellipse sine coeff’s; param’s; lin’zd
method coeff’s; param’s; presumed or Monte-Carlo;
presumed lin’zd or white white or resid.
white Monte- residual spec. (FFT
residual Carlo; spectra regular times,
spectra white or Lomb-Scargle
resid. spec. irregular times);
(FFT); general
simplified covar’nce matrix
covar’nce
Matrix
Solution OLS (same) IRLS OLS Default IRLS;
method Optional OLS
Complex two- Yes (same) (same) No Yes
dimensional
case
Matlab No Yes (same) No Yes
Enhanced No (same) (same) (same) Yes
diagnostics for
constituent
selection
Analyze No (same) (same) (same) Yes
multiple
records with
one execution

5
white noise floor. They also added the capability to generate confidence intervals for the
four current ellipse parameters, from the uncertainties in the cosine and sine model
coefficients, using either a linearized form of the nonlinear underlying relations or Monte
Carlo uncertainty propagation. Leffler and Jay (2009) (LJ09) investigated robust solution
methods and demonstrated that an iteratively-reweighted least squares (IRLS) approach
can minimize the influence of outliers, thereby reducing confidence intervals relative to
the standard ordinary least squares (OLS) method, leading to important consequences for
the constituent selection process. They investigated the one-dimensional case with
uniformly distributed times using a modified version of the PBL02 package that
implements the IRLS approach with the robustfit() function provided in the Matlab
Statistics Toolbox. Finally, Foreman et al (2009; FCB09) presented a method and
accompanying Fortran code to handle irregularly distributed temporal sampling. In
addition, their method includes exact “in-matrix” formulations for nodal/satellite
corrections, the astronomical argument for Greenwich phase calculation, and inferences,
such that the corrections are accurate over multi-year time periods. Their approach also
permits inferring multiple constituents from a single reference constituent. The upper
limit of record length for accurate nodal/satellite corrections by the FCB09 method is
18.6 years, beyond which methods without nodal/satellite corrections are applicable
(Foreman and Neufeld 1991), although FCB09 demonstrated the utility of their
formulation for longer records when nodal/satellite corrections are omitted.

UTide consists of a pair of Matlab functions designed to be easy to understand

and implement: ut_solv() for analysis, and ut_reconstr() to use the analysis results for
reconstruction of a time sequence for a hind-cast or forecast/prediction if needed. They
are intended to be helpful in streamlining the various stages of most typical tidal
analyses, including constituent selection and confidence interval estimation, for both two-
dimensional (e.g., tidal currents) and one-dimensional (e.g., sea level) cases. The analysis
function accepts records with times that are uniformly or irregularly distributed, and can
provide accurate nodal correction results for records with durations of up to 18.6 years.
The reconstruction function accepts arbitrary times and permits generation of
reconstructed fits using a subset of constituents, for example based on a signal-to-noise
(SNR) criterion or as specified by the user. While the functions incorporate a set of
optimal default choices that should help make analysis straightforward for users that are
less familiar with the details of tidal methods, they also accept options that enable
convenient experimentation with different method choices.

In addition to combining most features of the prior approaches described above

together in a single package, the following new contributions are made:
• development of single set of equations, for which (a) each of the prior methods
can be obtained as a special case, and (b) nodal/satellite corrections, Greenwich
phase computation by astronomical argument, and inferences are all included in-
matrix instead of being carried out as post-fit corrections;
• backwards compatibility with each of the prior methods, including “mix and
match” choices for individual methods of corrections and confidence intervals,
which is a useful capability to ground-truth new results against results from prior
analyses, as well as to investigate the sensitivity of results to method choices;

6
• enhanced diagnostics to aid constituent selection, particularly when the input
times are irregularly distributed, by presenting signal to noise ratios with
constituents ordered by their energy level, and by inclusion of indicators for
straightforward application of the conventional Rayleigh criterion, as well as the
noise-modified version proposed by Munk and Hasselmann (1964), and a new
diagnostic based on the correlations among model parameters;
• an improved confidence interval calculation building on that of PBL02 to use both
the auto- and cross-spectral character of the residual, estimate them by Lomb-
Scargle periodograms in the case of irregularly distributed times, and apply Monte
Carlo uncertainty propagation with fully general model parameters covariance
matrix;
• extension of the FCB09 methodology to (a) solution for the two-dimensional case
(e.g., tidal currents) in the complex formulation, (b) computation of reconstructed
tidal series (superposed harmonics, or the “fit”) at a sequence of arbitrary times
other than the input times, and (c) implementation in Matlab; and
• analysis of a group of records by a single execution of the UTide functions, which
proves valuable when the tidal analysis is to be applied to multiple records (e.g.,
an array of fixed observation sites, or output from a simulation at multiple grid
points).

7
II. Unified equation development
II.A. Model equations
This section considers the model equations for a single time sequence, referred to
as the “raw input”, which can be observations, numerical simulation output, synthesized
records, etc. An example for two dimensional raw input is a record of east and north
velocity components from a single depth, or their depth-averages; an example for one
dimensional raw input is a sea level record.

The raw inputs can have uniform or non-uniform temporal sampling, and the
development is intended to incorporate nodal/satellite corrections for records of duration
up to 18.6 years, but could be useful for longer records (as noted above) through
omission of the nodal/satellite corrections.

The set of constituents chosen for inclusion in the model, along with the set of
constituents (if any) chosen to be inferred and their associated reference constituents, are
presumed known in this section, based on a previously completed constituent selection
process. Section II.D below addresses diagnostics useful for constituent selection.

The matrix formulation and solution method is taken up in Section II.B and
Section II.C covers the confidence interval calculations.

II.A.1. General case: complex, two-dimensional

The model equation in the most general case (two-dimensional, complex notation,
including inferences) is presented here first. Section II.A.2 explains how the one-
dimensional case follows as a subset of these general equations, and Section II.A.3
explains the relationships between the complex formulation and the corresponding real
formulation, which are useful for the confidence interval calculation.

The most general case, described in this subsection, is the basis for the UTide
Matlab code, since all prior methods (see Table 1) and all subsets (e.g., one-dimensional
case, real notation case) can be obtained as special cases.

Symbols retain their meaning throughout the document. To the extent possible,
they have been selected for consistency with the prior developments cited in Table 1.

The raw input consists of real-valued u raw (ti ) and real-valued v raw (ti ) , where u
and v are perpendicular Cartesian components of the velocity and the arbitrarily
distributed times are ti , where i = 1...nt . By convention the two components are directed
eastward and northward respectively, but more generally they can be the pair of
components along the first and second axes in any right-handed coordinate system. The
complex form of the raw input
x raw (ti ) = u raw (ti ) + iv raw (ti ) (1)

8
is the quantity for which the model, x mod (ti ) = u mod (ti ) + iv mod (ti ) , is constructed.

The model equation in its simplest form is

x mod (ti ) = ∑ (Eiq aq+ + Eiq* aq− ) + x + x& ⋅ (ti − t ref ) .
nallc
(2)
q =1

The summation is over all q = 1...nallc constituents (non-reference, reference, and

inferred, as explained below) in the model. Each constituent has constant complex
amplitudes aq+ , aq− for components that rotate counter-clockwise and clockwise in time,
which multiply the complex exponential functions Eiq , Eiq* explained in detail in the next
paragraph below. The mean x = u + iv combines the respective means of the two real
components. The trend, if included in the model, has coefficient x& = u& + iv& that similarly
combines those of the two real components, and is computed relative to the reference
time tref ; by convention tref is a time central among the raw input times, and here it is
defined as the average of the first and last raw input times,
tref = (t1 + tnt ) / 2 . (3)

II.A.1.a. Pre-filtering and nodal/satellite corrections; Greenwich phase lags

In the complex plane an individual harmonic constituent of radian frequency ωq
consists of a superposed pair of components counter-rotating in time, with complex
coefficients denoted by + and - superscripts for counterclockwise- and clockwise-
rotation respectively. The counterclockwise- and clockwise-rotating elements that the
complex coefficients multiply take exponential forms Eiq and Eiq* , respectively, where
Eiq = E (ti , ωq ) = P(ωq ) ⋅ F (ti , ωq ) ⋅ exp i[U (ti , ωq ) + V (ti , ωq )]
(4)
= Pq Fiq exp i (U iq + Viq )
and the shorthand expressions
Pq = P(ωq ) , Fiq = F (ti , ωq ) , U iq = U (ti , ωq ) , and Viq = V (ti , ωq ) (5)
represent the following real-valued functions:
• the correction factor for pre-filtering, P (ωq ) , a dimensionless transfer function of
the filter that was applied to the raw input xqraw (ti ) prior to the analysis
o P (ωq ) is set to unity in the case of no pre-filtering
o P (ωq ) can be complex, in which case Re( Pq ) = Im(Pq ) ;
• the nodal/satellite correction amplitude factor F (ti , ωq ) (unitless) and phase offset
U (ti , ωq ) (radians), evaluated at time ti for constituent q
o F (ti , ωq ) and U (ti , ωq ) are set to unity and zero respectively, for the case
of no nodal/satellite corrections
o in addition, as explained more fully in Section II.A.4.a below, in the
traditional linearized times development F (tref , ωq ) and U (tref , ωq )
appear in place of F (ti , ωq ) and U (ti , ωq ) here; and

9
 • the astronomical argument V (ti , ωq ) (radians), which ensures resulting phase lags
( g q+ , g q− , g q , g qu , g qv , introduced below) are relative to the equilibrium tide at
Greenwich
o V (ti , ωq ) is replaced by ωq ⋅ (ti − tref ) in order for reported phase lags to
instead be uncorrected “raw” phase lags relative to reference time tref
o in addition, as explained more fully in Section II.A.4.b below, in the
traditional linearized times development, V (t* , ωq ) + ωq ⋅ (ti − t ref ) appears
in place of V (ti , ωq ) here.
Regarding notation, the traditional symbols for nodal/satellite corrections f and u are
capitalized here in order to reduce ambiguity with their commonplace use in
oceanographic literature as symbols for the Coriolis parameter and eastward velocity
component respectively. In addition, for convenience the Greek symbol ν commonly
used for the astronomical argument is replaced by V —also capitalized, to help reduce
ambiguity with its commonplace use for the northward velocity component.

II.A.1.b. Current ellipse parameters

In the two-dimensional case the amplitude and phase information for each
constituent is conventionally reported as four current ellipse parameters. The complex
coefficients have associated positive, real magnitudes Aq+ , Aq− and associated phases
ε q+ , ε q− ,
aq+ = Aq+ exp iε q+
(6)
aq− = Aq− exp iε q− ,
where
Aq+ =| aq+ | ε q+ = arctan[Im(aq+ ), Re(aq+ )]
(7)
ε q− = arctan[Im(aq− ), Re(aq− )] ,
Aq− =| aq− |
and the Greenwich phase lags for the rotating components (see, e.g., G72) are
g q+ = −ε q+
(8)
g q− = ε q− .

For an individual constituent the tip of the velocity vector in the complex plane
traces out an ellipse during each full period. Current ellipse parameters are expressed in
terms of the magnitudes and phases of the complex amplitudes as
Lsmaj
q = ( Aq+ + Aq− )
η
Lsmi
q = ( Aq+ − Aq− )
⎡ ε q+ + ε q− ⎤ (9)
θ q = mod ⎢ ,π ⎥
⎢⎣ 2 ⎥⎦
g q = −ε q + θ q .
+

and are defined as

10
 • the semi-major axis length Lsmaj
q (positive; same units as u, v );
η
• the semi-minor axis length Lsmi
q (positive for counterclockwise rotation in time,
negative for clockwise rotation in time; same units as u, v );
• the orientation angle θ q (positive counterclockwise from the positive u axis,
which in the conventional case is eastward; radians, range 0 and π ) of the semi-
major axis that is (following PBL02) directed toward the positive v axis, which in
the conventional case is northward; and
• the Greenwich phase lag g q (radians, range 0 to 2π ) of the vector velocity
relative to the time of its alignment with the semi-major axis that has a component
directed toward positive v , which in the conventional case is northward.
Note that g q with no superscript denotes the Greenwich phase lag of the vector velocity,
and care should be taken to avoid confusion between it and the Greenwich phase lags
g q+ , g q− (8) of the counterclockwise- and clockwise-rotating components, as well as the
Greenwich phase lags g qu , g qv of the u and v components, defined in (24) below. The
inverse relations for the complex amplitudes,
η
aq+ = [( Lsmaj
q + Lsmi
q ) / 2] exp i (θ q − g q )
η (10)
aq− = [( Lsmaj
q − Lsmi
q ) / 2] exp i (θ q + g q ) ,
prove useful in the confidence intervals development below. By convention, the four
current ellipse parameters (9) are reported, but the equivalent information could be
reported as complex magnitudes Aq+ , Aq− and Greenwich phase lags g q+ , g q− from (7), or as
the real amplitudes Aqu , Aqv and Greenwich phase lags g qu , g qv of the u and v components,
defined in (24) below.

II.A.1.c. Constituent inference

To make constituent inference clear, model equation (2) is rewritten in the form

( )
nR ⎡ ⎤
( )
nNR nIk
x (ti ) = ∑ Eij a j + Eij a j + ∑ ⎢ Eik aˆ k + Eik aˆ k + ∑ Eilk aˆˆl+k + Eil*k aˆˆl−k ⎥
mod + * − + * −

j =1 ⎢
k =1 ⎣ lk =1 ⎦⎥ (11)
+ x + x& ⋅ (ti − t ref ) .
Here, the first summation represents contributions of a sequence of harmonic
constituents, with complex coefficients a +j , a −j for j = 1...nNR , that are denoted “non-
reference” (subscript NR), because they are not used as reference constituents to infer
other constituents. The second summation is non-zero only if inferred constituents are
included in the model; it represents the combined contributions from both (a) the
sequence of k = 1...nR reference constituents (subscript R), in its first two terms, and (b) a
nR
total of nI = ∑ nIk inferred constituents (subscript I), including a sequence of lk = 1...nIk
k =1
inferred constituents for the k th reference constituent, in the interior summation term.
The hat notation for the complex coefficients âk+ , âk− indicates reference constituents, and

11
the double-hat notation for the complex coefficients aˆˆl+k , aˆˆl−k indicates inferred
constituents. Thus all constituents in the model collectively number nallc = nNR + nR + nI .

It can thus be seen that, in the expressions prior to (11) above, (a) q = 1...nallc
denotes any of the indices j , k , or lk and (b) for a, A, ε , g variables, and current ellipse
parameters, the hat and double-hat notation is implied; for example, aq+ represents a +j ,
âk+ , or âl+k .

By definition, an inferred constituent is characterized by an amplitude ratio and

phase offset that are known, relative to its designated reference constituent, from
auxiliary information prior to the tidal analysis at hand. The real-valued inference
amplitude ratios rlk+ , rlk− and phase offsets ς l+k , ς l−k ,
ˆ ˆ
r l+k = Aˆ l+k / Aˆ k+ r l−k = Aˆ l−k / Aˆ k−
(12)
ς + = gˆ + − gˆˆ +
lk k lk ς − = gˆ − − gˆˆ −
lk k lk

are specified for the lk th inferred constituent relative to the k th reference constituent.
The complex coefficients of the inferred constituents are
aˆˆl+k = Rl+k aˆ k+
(13)
aˆˆl−k = Rl−k aˆ k− .
where the complex inference constants are defined
Rl+k = rlk+ exp iς l+k
(14)
Rl−k = rlk− exp(−iς l−k ) .

The model equation that is recast in matrix form below, and solved in practice,
results on substitution of (13) into (11) and is
( )
x mod (ti ) = ∑ (Eij a +j + Eij* a −j ) + ∑ Eik+ aˆ k+ + Eik− aˆ k− + x + x& ⋅ (t − t ref ) ,
nNR nR
~ ~ *
(15)
j =1 k =1

in which the modified exponential functions are, following FCB09 but here in the
complex formulation,
~+ ⎛ nIk ⎞
Eik = Eik ⎜1 + ∑ Qilk Rl+k ⎟⎟
⎜
⎝ lk =1 ⎠
(16)
⎛ ⎞
k
nI
~−
Eik = Eik ⎜⎜1 + ∑ Qilk Rl−k ⎟⎟ .
*

⎝ lk =1 ⎠
The latter expressions include summations over the lk = 1...nIk constituents to be inferred
from the k th reference constituent, and the unitless complex weighting parameter for
inferences

12
 E (ti , ωlk ) Eilk Plk Filk
Qilk = Q(ti , ωlk ) = = =
exp i (U ilk − U ik + Vilk − Vik ) . (17)
E (ti , ωk ) Eik Pk Fik
Note that (15) is cast in terms of (a) the complex coefficients for the non-reference and
reference constituents, and (b) the complex inference constants (14), which appear in the
modified exponential functions of (16). Because the complex coefficients of the inferred
constituents do not appear in (15), the inferred constituents are solved for indirectly.

It should be emphasized that this formulation for inferences is exact. In contrast to

the traditional approximate method (details provided in Section II.A.4.c below), in which
the amplitude and phase of the reference frequency are corrected post-solution using an
approximation, here inferences are accomplished inherently as part of the model equation
(FCB09) and can affect all the other constituents (not just their reference constituents).
Furthermore, the present formulation permits inference of multiple constituents from a
single reference frequency, which is not the case in the standard traditional method.
Finally, unlike the traditional method, this treatment of inferences does not break down
where the amplitude of the reference constituent tends toward zero, such as near
amphidromic points, as Godin (1972) pointed out is the case for the traditional method.

II.A.1.d. Summary
The model characteristics can be summarized in terms of the input and output
information. The input information consists of
• complex-valued raw input x raw (ti ) , formed as the combination of the real-valued
raw inputs u raw (ti ) and v raw (ti ) , as in (1);
• the names and frequencies of nNR non-reference constituents to be included;
• if inferred constituents will be included,
o the names and frequencies of nR reference constituents to be included
o the names and frequencies of a total of nI inferred constituents ( nIk
inferred constituents for the k th reference constituent), together with real-
valued inference amplitude ratios rlk+ , rlk− and phase offsets ς l+k , ς l−k for each
inferred constituent relative to its reference constituent; and
• whether or not a trend is to be included in the model.
The output information conventionally consists of
• four current ellipse parameters ( Lsmaj , Lsmiη ,θ , g ) for each of the
nallc = nNR + nR + nI (non-reference, reference, and inferred) constituents;
• mean values u and v ; and
• trend coefficients u& and v& , if the trend was included.

II.A.2. One-dimensional case, complex

Exposition of how the above general equations for two-dimensional raw input are
simplified in the special case of one-dimensional raw input is useful. In this case the raw
input is the real-valued η raw (ti ) , representing any one-dimensional quantity, for example
sea level. In the above equations, substitute η everywhere for u , and take all v

13
parameters to be identically zero. The result is that for any constituent (subscripts and/or
hats dropped),
*
a+ = a− (18)
or equivalently
A+ = A−
(19)
ε + = −ε − ,
and therefore the current ellipse is degenerate ( Lsmiη = 0 , θ = 0 ) and lies along the real
axis with real amplitude and Greenwich phase lag
Aη = Lsmaj = A+ = A−
(20)
g η = −ε + = ε − .

With respect to inference, for known real-valued inference amplitude ratios and
phase offsets
ˆ
rlηk = Aˆlηk / Aˆ kη
(21)
ς lηk = gˆ ηk − gˆˆ ηlk
~ ~ *
it follows that the modified exponential functions are Eik , Eik , where
⎛ ⎞
k
nI
~
Eik = Eik ⎜⎜1 + ∑ Qilk Rlηk ⎟⎟
⎝ lk ⎠ (22)
~+ ~−
( = Eik = Eik ),
and
Rlηk = rlηk exp iς lηk
* (23)
( = Rl+k = Rl−k ).

II.A.3. Relations to real formulation

It is valuable to understand the relation of the above complex formulation to its
real counterparts—both to gain a more intuitive understanding of the one-dimensional
analysis, and also for later use in the derivation of confidence intervals (Section II.C).
The two-dimensional case is treated first, followed by the one-dimensional case.

Underlying the complex form of the model equation above are expressions for the
real-valued components of the q th constituent (non-reference, reference, or inferred),
uqmod (ti ) = Aqu Pq Fiq cos(U iq + Viq − g qu ) + u + u& ⋅ (ti − tref )
(24)
vqmod (ti ) = Aqv Pq Fiq cos(U iq + Viq − g qv ) + v + v& ⋅ (ti − t ref ) ,
where q represents any of the j , k , or lk indices, and throughout this subsection hat and
double-hat variables are not explicitly shown but all relations apply to them. Here,
Aqu , g qu , Aqv , g qv are the real-valued amplitudes and Greenwich phase lags of the respective
Cartesian velocity components. Real-valued cosine and sine coefficients are defined as
X qu = Aqu cos g qu X qv = Aqv cos g qv (25)

14
 Yqu = Aqu sin g qu Yqv = Aqv sin g qv ,
such that corresponding relations to the amplitudes and Greenwich phase lags are
2 2 2 2
Aqu = X qu + Yqu Aqv = X qv + Yqv
(26)
(
g qu = arctan Yqu , X qu ) (
g qv = arctan Yqv , X qv . )
By use of shorthand definitions
Ciq = Pq Fiq cos(U iq + Viq )
(27)
Siq = Pq Fiq sin(U iq + Viq )
and sine and cosine addition formulae, (24) is written
uqmod (t i ) = Ciq X qu + SiqYqu
(28)
vqmod (t i ) = Ciq X qv + SiqYqv ,
leading to the relation between the complex and real formulations,
Ciq X qu + SiqYqv + i (Ciq X qv + SiqYqv )
(29)
= Eiq aq+ + Eiq* aq− .

The complex coefficients can be expressed in terms of complex combinations of

the cosine and sine coefficients,
X q = X qu + iX qv and Yq = Yqu + iYqv , (30)
as
aq+ = ( X q−iYq ) / 2 and aq− = ( X q+iYq ) / 2 , (31)
and therefore take the form
aq+ = [( X qu + Yqv ) + i ( X qv − Yqu )] / 2
(32)
aq− = [( X qu − Yqv ) + i ( X qv + Yqu )] / 2 .
such that
1
Aq+ = ( X qu + Yqv ) 2 + ( X qv − Yqu ) 2
2
ε q = arctan[( X qv − Yqu ), ( X qu + Yqv )]
+

(33)
1
Aq− = ( X qu − Yqv ) 2 + ( X qv + Yqu ) 2
2
ε q = arctan[( X qv + Yqu ), ( X qu − Yqv )] .
−

The inverse relations of (31) are, using (30),

X q = X qu + iX qv = aq+ + aq−
(34)
− iYq = Yqv − iYqu = aq+ − aq− ,
or equivalently
X qu = Re(aq+ + aq− ) Yqu = − Im(aq+ − aq− )
(35)
X qv = Im(aq+ + aq− ) Yqv = Re(aq+ − aq− ) ,
which are useful, for example, to compute the real amplitudes and Greenwich phase lags
of the u and v components via (26), when the complex coefficients are known.

15
 *
In the one-dimensional case, for all v variables identically zero and a + = a −
(18), replacing u by η , it follows that
X q = X qη = aq+ + aq− = aq+ + aq+ = 2 Re(aq+ ) = 2 Re(aq− )
*

(36)
Yq = Yqη = i (aq+ − aq− ) = i (aq+ − aq+ ) = 2i Im(aq+ ) = −2i Im(aq− ) ,
*

which results in some simplification of the above relations. With respect to inference,
~ ~ *
the complex form Eiq+ aq+ + Eiq− aq− is recast as
~ ~
Ciqη X qη + SiqηYqη (37)
where (as shown by FCB09)

( )
nIk
~η
Cik = Cik + ∑ Cilk Rlck − Silk Rlsk
lk =1
(38)
( )
k
nI
~
Sikη = Sik + ∑ Cilk Rlsk + Silk Rlck ,
lk =1

using inference constants Rlck , Rlsk defined as

Rlck = rlηk cos ς lηk
(39)
Rlsk = rlηk sin ς lηk .

II.A.4. Prior methods as special cases

The equations in the preceding sections generalize those of FCB09 and differ
from various earlier methods, including PBL02, in three main ways. The first difference
is that nodal/satellite corrections use exact times, instead of estimated linearized times.
The second difference is that the astronomical argument uses exact times. The third
difference is that inferences are handled in an exact way instead of using an
approximation. This subsection explains the modifications to the above development that
are required in order to recover results from the earlier methods.

The decision to use an earlier method in the UTide functions can be made
independently for some or all (in ‘mix and match’ fashion) of these three differences.
This enables complete sensitivity analyses to be carried out when investigating the
relative importance of the differences.

II.A.4.a. Nodal/satellite corrections using linearized times

In order to implement the nodal/satellite corrections using linearized times instead
of exact times, throughout the above development replace F (ti , ωq ) and U (ti , ωq ) with
F (t ref , ωq ) and U (tref , ωq ) , respectively. As above, q represents any of the j , k , or lk
indices. The fixed reference time tref is arbitrary but usually taken to be a time that is
central among the raw input times, in order to increase the accuracy of the corrections;
here tref is as defined in (3).

16
 As noted above, in order to omit nodal/satellite corrections entirely, F (ti , ωq ) and
U (ti , ωq ) are replaced everywhere by one and zero, respectively.

II.A.4.b. Greenwich phase lags: linearized times in astronomical argument

In order for linearized times to be used instead of exact times, in the conversion to
Greenwich phase lag by correction with the astronomical argument, replace V (ti , ωq )
everywhere by V (t ref , ωq ) + ωq ⋅ (ti − t ref ) .

As noted above, in order to entirely omit the conversion to Greenwich phase lags,
V (ti , ωq ) should be replaced everywhere by ωq ⋅ (ti − t ref ) , with the result that g values
are raw phase lags relative to the reference time.

II.A.4.c. Approximate inferences

In cases where the nodal/satellite corrections are either omitted or carried out
using linearized times, and the Greenwich phase calculations are also either omitted or
carried out using linearized times—that is, in cases for which neither of these two
corrections are exact—inferences by the approximate method (see G72, F77, and F78 for
the derivation and explanation) can be recovered as follows. Note that the customary
application of the approximate method precludes use of a single reference constituent for
multiple inference constituents. As a result the index lk never differs from 1, or
equivalently nIk = 1 for any reference constituent k .
• First, solve the model equation using Rl+k = R l−k = 0 for all k . This has the effect
that the non-reference and reference constituents are treated identically in the
model, such that the first and second summations in (15) differ only in the ranges
of their indices.
• Then, use the results aˆ k+ , aˆ k− from that solution, for the complex coefficients of the
reference constituents, to compute corrected complex coefficients of the reference
constituents
1
aˆ k+,corr = aˆ k+
1 + β lk Rlk Q(tref , ωlk )
+

(40)
− − 1
aˆ k ,corr = aˆ k ,
1 + β lk Rl−k Q* (tref , ωlk )
where Q is as defined in (17) and

β lk =
(
sin (tnt − t1 )(ωlk − ωk )
.
)
(41)
(tnt − t1 )(ωlk − ωk )
• Finally, use aˆ k+,corr , aˆ k−,corr instead of aˆ k+ , aˆ k− (i) in (7) and (9) to compute the current
ellipse parameters for the reference constituents, and (ii) in (13) to calculate the
complex coefficients of the inferred constituents, from which the current ellipse
parameters for the inferred constituents are computed.

17
II.B. Matrix formulation and solution method
The model equation (15) as developed above is readily cast in matrix form
x mod = Bm , (42)
where symbols without subscripts are understood to be matrices, column vectors, or row
vectors throughout this section. The modeled values x mod are a nt × 1 column vector,
complex in the two-dimensional case and real in the one-dimensional case. The basis
functions comprise the complex-valued nt × nm matrix B , where nm = 2(n NR + nR ) + 2 is
the number of model parameters directly solved for when the trend is included in the
model, that has form
~ ~ *
B = [ E E * E + E − I (nt ,1) t ] (43)
with sub-matrices
• E , a complex-valued nt × nNR matrix with elements Eij defined as in (4);
~ ~ ~ ~
• E + , E − , each a complex-valued nt × nR matrix with elements Eik+ , Eik− defined as
in (16);
• I (nt ,1) , an nt × 1 column vector of unit values; and
• t , an nt × 1 column vector with real elements (ti − tref ) /(tnt − t1 ) , normalized such
that they are order one and unitless as are the other elements of B , in order to
keep it well-conditioned.
The model parameters vector m is a nm ×1 complex-valued column vector of form, when
the trend is included in the model,
m = [a1+ L an+NR a1− L an−NR aˆ1+ L aˆ n+R aˆ1− L aˆ n−R x x&′]T , (44)
where x&′ = (t nt − t1 ) ⋅ x& in order to accommodate the normalization of t . When the trend is
not included in the model, the final column of (43) and the final element of (44) are
omitted, and the number of model parameters is nm = 2(nNR + nR ) + 1 .

The matrix formulation is cast in terms of complex-valued matrices. For most

cases an equivalent formulation using real-valued matrices exists; an example, though its
equations were not explicitly presented by PBL02, is the t_tide Matlab code. The
complex formulation is used here because it facilitates solution of the case of exact
inferences with two-dimensional raw input, for which there is no equivalent in real-
valued matrices, and because the confidence interval development is in terms of complex
bi-variate normal statistics of the coefficients. The complex formulation is also the most
general, and proves to be convenient, although it should be noted that in some cases it
may not be the most efficient computationally.

The problem reduces to determining the set of model parameters that minimizes a
suitable measure of the residual, or misfit between the raw input and the model,
e = x raw − x mod = x raw − Bm . (45)
The residual is an nt × 1 complex-valued column vector with real-valued corresponding
Cartesian components
eu = u raw − u mod = Re( x raw − Bm) (46)

18
 e v = v raw − v mod = Im( x raw − Bm)
that are of use in the confidence interval calculation below. As long as nt exceeds nm ,
the system is over-determined with nt − nm degrees of freedom and the standard solution
method (e.g., F77, F78, PBL02) is ordinary least squares (OLS). The OLS solution
minimizes the L2 norm of e and takes the form
m = ( B H B) −1 B H x raw , (47)
where superscript H indicates the transpose-conjugate or Hermitian adjoint. To
determine the OLS solution, the UTide code uses the built-in Matlab ‘backslash’
operator.

Once the solution is complete, the resulting model parameters vector (44)
comprises a set of complex coefficients for the non-reference and reference constituents.
The complex coefficients of the inferred constituents are then determined from those of
the reference constituents using (13) with the known complex inference constants (14).
Finally, for all constituents, the complex magnitudes and phases are then determined
using (7), from which the current ellipse parameters are computed from (9).

From the solution a hind-cast or a forecast/prediction time sequence can be

reconstructed at any arbitrary set of times. The simplest way to carry this out is using (2).
The arbitrarily distributed times ti* at which the reconstructed values are to be computed
are used, together with the pre-filtering coefficient, the nodal/satellite amplitude/phase
corrections and the astronomical argument, to compute Eij , Eik , Eilk (4). The
reconstruction uses Eij , Eik , Eilk in (2) with the known model outputs, namely (i) the
complex amplitudes a +j , a −j , aˆk+ , âk− of non-reference and reference constituents that were
solved for directly, (ii) the complex amplitudes of inferred constituents aˆˆ + , aˆˆ − , computed
lk lk

using (13) with the known amplitude ratios and phase offsets, (iii) the mean x , and (iv) if
included, the trend x& and its reference time tref .

II.B.1. Iteratively reweighted least squares robust solution

Robust methods using L1/L2 hybrid norms offer a number of advantages, as
explored by LJ09. Confidence intervals can be reduced substantially, relative to those for
the OLS method. In practice, the reduction in confidence intervals relative to the OLS
solution method is commonly larger than the differences among confidence interval
results when calculated by the white or colored methods, and/or the linearized or Monte
Carlo methods (described in detail in section II.C below). In addition, because reduced
confidence intervals increase SNR, this can mean that a substantially larger number of
constituents will be considered significant (for example, when using a fixed minimum
SNR threshold, Section II.D), and therefore selected for inclusion in the model, as
compared to the OLS case.

In the iteratively-reweighted least squares (IRLS) approach, a weighting of the

observations is determined as part of the solution, such that the influence of outliers is

19
minimized. In this case, the minimized quantity is a measure of the weighted residual, or
weighted misfit between the raw input and the model,
ew = We = W ( x raw − Bm), (48)
and similarly
ewu = Weu ewv = Wev , (49)
where W is an nt × nt diagonal matrix with the scalar weight values on the diagonal. In
this case the general solution is
m = ( B H WB) −1 B H Wx raw . (50)
In the case of equally-weighted observations, W is the identity matrix and (50) reduces
to the OLS solution as expected.

For IRLS cases, following LJ09, the UTide code uses the Matlab Statistics
Toolbox function robustfit(), an iterative solver which implements a user-specified
weight function selected from among a range of common shapes, with a corresponding
constant value of the tuning parameter (see Matlab documentation for robustfit()). The
output, for given choices of weight function and tuning parameter, includes the set of
optimal model parameters with the corresponding weight matrix W for the raw inputs.

There are no established guidelines regarding the choice of weight function and
tuning parameter that is appropriate for any given analysis, and it should be expected that
the optimal choices will vary depending on the nature of the raw input (one-dimensional
or two-dimensional, noise conditions), as well as the nature of the sampling duration and
resolution (e.g., uniformly or irregularly distributed times), in addition to the model
configuration (number of constituents, etc). A lower tuning parameter generally causes a
greater penalty against outliers and requires a higher number of iterations. Based on
analysis of a sea level record from a single location in a tidal estuary, with uniformly
distributed times of hourly resolution and duration 6 months, LJ09 concluded that the
best weight function was Cauchy and the best tuning parameter was 0.795 (the Matlab
default value 2.385, reduced by a factor of 3).

In UTide the default weight function is Cauchy and alternative weight functions
can be specifed by an optional input (e.g., ‘Huber’,…). The default tuning parameter used
by UTide is that provided by Matlab for the given weight function (see Matlab
documentation for robustfit()), and an optional UTide input (‘TunRdn’) can reduce the
tuning parameter by a specified factor (e.g. 3, as found optimal by LJ09 for their example
record). Empirical experimentation will be needed for analysis of any given record, to
determine the appropriate weighting function and tuning parameter.

II.C. Confidence intervals

A method to compute uncertainties of the cosine/sine model parameters using the
basis function matrix was explained by G72, F77, and FH89. In contrast to that method,
which did not distinguish the spectral nature of the residual and can thus be referred to as
a presumed “white noise floor” approach, the method put forth by PBL02 and denoted
“colored” was based on using spectral properties computed from the actual residual. In
addition, PBL02 presented two methods to determine confidence intervals for the current

20
ellipse parameters, which are not model parameters themselves but are nonlinear
functions of the model parameters: a linearization approach, and Monte Carlo uncertainty
propagation (which they denoted the “bootstrap” method).

This subsection presents a new method, the default case in UTide, that builds on
earlier approaches and (a) generates confidence intervals of the current ellipse
parameters, using Monte Carlo uncertainty propagation, but generalizes the PBL02
method by relaxing the simplifying assumptions it made about the form of the covariance
matrix of functions of the model parameters, (b) is colored and generalized PBL02 to
incorporate both the auto-spectra and cross-spectra of the u and v component residuals,
using their weighted forms (48) in the case of the IRLS solution method, and (c) is
amenable to use of spectral estimates for the latter that have been computed either by
FFT for evenly distributed times, or by Lomb-Scargle periodogram for irregularly
distributed times.

The default “colored, Monte Carlo” UTide method is explained in the remainder
of this section. However, UTide allows the alternative to specify (a) use of the white
noise floor assumption instead of the colored residual spectra, and/or (b) use of the
linearized approach instead of Monte Carlo. The development below makes clear how the
white case can be implemented. The linearized method used is that of PBL02 (and
therefore not explicitly presented here), in which it is presumed that there are no non-zero
correlations among the coefficients.

For any of these methods, confidence interval estimates are based on the
assumption that all energy in the residual is noise. In most typical analyses, the record
contains a sub-tidal (i.e. non-tidal, low frequency, and/or weather-band) signal at
frequencies lower than tidal, in addition to the random noise and the tidal component. It
should therefore be recognized that the sub-tidal signal, if not somehow removed prior to
the analysis, contributes to the confidence intervals and will in general make them
artificially higher. If not removed, the sub-tidal signal will also tend to make the statistics
of the residual deviate from the normal statistics presumed by the confidence interval
calculations. For these reasons, to obtain the most accurate confidence intervals, a simple
and effective strategy is first to compute the low-pass of the raw inputs and subtract it
from them, then to perform the tidal analysis on the resulting difference. This approach,
in effect, accomplishes high-pass filtering. It is simpler than formal application of a high-
pass or band-pass pre-filter, and should in general not require a pre-filtering correction
( Pq in (5) remains unity).

As described above, relevant context is provided by the fact that using the IRLS
solution method instead of OLS can have a substantial impact on confidence intervals. In
many cases they can be reduced to a greater degree than the differences between when
they are computed using the linearized or Monte Carlo method, and/or the white or
colored method.

Finally, it is important to note that certain assumptions underlying the

development in this section are strictly valid, as discussed in FH89, only when the raw

21
input times are uniformly spaced. Therefore the results in the case of irregular times
should be considered potentially reasonable and approximate first estimates, but should
be compared against the results for uniform times whenever possible, and used with a
measure of caution.

The subsections that follow develop expressions for a key covariance matrix
needed by the Monte Carlo uncertainty propagation method. Implementation of the
confidence interval calculation, using that result to generate random realizations for
Monte Carlo simulation, is then explained.

II.C.1. Complex bi-variate normal statistics

The model parameters vector m (44), which includes the complex coefficients for
the non-reference and reference constituents, is assumed to be a complex normal vector;
its individual elements are assumed to be complex random variables related to each other
by complex bivariate normal statistics. Because the complex coefficients of inferred
constituents do not appear in the model parameters vector, due to the fact that inferred
constituents are included in the model indirectly (see above), computation of confidence
intervals for inferred constituents follows a different approach based on the statistics of
the corresponding reference constituent, as explained in Section II.C.5.

A complex normal vector m of size nm × 1 is characterized by its complex-valued

(i) mean μ = E[m] , an nm × 1 vector,
(ii) covariance ΓC = E[(m − μ )(m − μ ) H ] , a Hermitian nm × nm matrix, and
(iii) pseudo-covariance ΓP = E[(m − μ )(m − μ )T ] , a symmetric nm × nm matrix,
where E[] is the expectation operator and superscripts H and T indicate the complex
conjugate transpose or Hermitian adjoint, and the ordinary transpose, respectively. Full
generality is retained, in the sense that no presumption is made that the complex variables
are “circular” or “proper”, thus allowing for non-zero pseudocovariance ΓP .

A property of the complex bivariate normal statistics (e.g., Goodman 1963) is that
the variance-covariance matrices between the complex random vector formed from the
real part of the model parameters vector and the complex random vector formed from its
imaginary part are
cov[Re(m), Re(m)] = Re(ΓC + ΓP ) / 2
cov[Re(m), Im(m)] = − Im(ΓC − ΓP ) / 2
(51)
cov[Im(m), Re(m)] = Im(ΓC + ΓP ) / 2
cov[Im(m), Im(m)] = Re(ΓC − ΓP ) / 2 .
As will be seen below, these expressions allow the remainder of the development to
proceed in terms of real-valued quantities. This is needed to facilitate the use of the
Lomb-Scargle periodogram for spectral estimates when the distribution of times is
irregular, because they are computed using the real-valued u raw , v raw .

22
II.C.2. White noise floor case with non-zero cross-correlations
The following generalizes the development of G72, F77, and FH89 to include the
two dimensional case, non-zero cross-correlation terms, and the use of complex model
parameters instead of real cosine/sine coefficients. The weighted complex error ew (48)
is assumed be a zero-mean complex normal variable. The development is carried out in
terms of the IRLS case, including the weight matrix, throughout this section; the OLS
case is recovered through use of the identity matrix for the weight matrix. The total error
variance σ C2 is a real scalar representing the variance of the complex error, equivalent to
the mean square residual or mean square misfit (MSM) between the raw input and the
model output,
H
x raw Wx raw − m H B HWx raw
σ =σ
2 2
= . (52)
nt − nm
C MSM

By direct analogy with the expression for the total error variance σ C2 , the total error
pseudo-variance σ P2 is the complex constant
T
x raw Wx raw − mT BTWx raw
σ =
2
. (53)
nt − nm
P

It follows that estimates of the nm × nm covariance matrix and pseudo-covariance matrix

for the model parameters are
ΓC = ( B H WB) −1σ C2
(54)
ΓP = ( BTWB) −1σ P2 .
Implementation of formulas based on (51) is simplified below by defining the sum and
difference of the covariance and pseudo-covariance matrices, which here incorporate all
model parameters, as
G all = ΓC + ΓP
(55)
H all = ΓC − ΓP .

Each individual constituent, indexed by c = 1...nc where nc = nNR + nR , has two

complex model parameters
ac+ = mc and ac− = mc+nc . (56)
Throughout Sections II.C.2-4, the symbols a +c , ac− are used to denote the complex
coefficients of either a non-reference constituent or a reference constituent; that is, for
reference constituents, the hats on these variables in the previous sections are dropped,
such that ac+ represents either a +j or âk+ , for example. The 2 x 2 sub-matrices of G all and
H all pertinent to constituent c are
⎡ Gcall,c Gcall,c+nc ⎤
G cc = ⎢ all ⎥
⎢⎣Gc+nc ,c Gcall+nc ,c+nc ⎥⎦
(57)
⎡ H call,c H call,c+nc ⎤
H cc = ⎢ all ⎥.
⎢⎣ H c+nc ,c H call+nc ,c+nc ⎥⎦

23
It is presumed that the constituent selection process (Section II.D) has been completed
such that the covariances and pseudo-covariances between model parameters of different
constituents can be neglected.

Because the statistics of the real and imaginary parts of the model parameters (51)
are known, it proves convenient to compute the current ellipse parameters for a given
constituent from the vector of 4 real-valued (superscript R) parameters
mcR = [Re(ac+ ) Im(ac+ ) Re(ac− ) Im(ac− )] (58)
rather than from the complex coefficients ac+ , ac− themselves (in m ). By definition, the
variance-covariance matrix of mcR is
var cov(mcR ) = cov(mcR , mcR ) =
⎡σ Re(
2
a+ )
cov[Re(a + ), Im(a + )] cov[Re(a + ), Re(a − )] cov[Re(a + ), Im(a − )]⎤
⎢ ⎥
⎢ σ Im(
2
a+ )
cov[Im(a + ), Re(a − )] cov[Im(a + ), Im(a − )]⎥ (59)
⎢ σ Re( a − )
2
cov[Re(a ), Im(a )]⎥
− −
⎢ ⎥
⎢⎣ σ Im(
2
−
a )
⎥⎦
where the lower left triangle is left blank because the matrix is real-valued and symmetric
with 10 unique values. In (59), as in similar expressions in the remainder of this
subsection, for clarity the c subscripts are dropped from elements within the matrix.
Whereas for simplicity the method of PBL02 (see t_tide code) presumed specific
relationships among the elements of var cov(mcR ) , for example that σ Re( 2
a+ )
= σ Im(
2
a+ )
and
that cov[Re(a + ), Im(a + )] = 0 , these assumptions are relaxed in the present development.

It follows from the statistics of real and imaginary parts of the model parameters
(51) that under the white noise floor assumption the unique terms in var cov(mcR ) are, in
the upper left quadrant,
σ Re(
2
a+ )
= cov[Re(ac+ ), Re(ac+ )] = Re(G11cc ) / 2
c

cov[Re(ac+ ), Im(ac+ )] = − Im(H11cc ) / 2 (60)

σ Im(
2
a )
= cov[Im(ac+ ), Im(ac+ )] = Re( H11cc ) / 2 ;
+
c

in the lower right quadrant,

σ Re(
2
a− )
= cov[Re(ac− ), Re(ac− )] = Re(G22
cc
)/2
c

cov[Re(ac− ), Im(ac− )] = − Im(H 22

cc
)/2 (61)
σ 2
Im( ac− )
−
c
−
= cov[Im(a ), Im(a )] = Re( H ) / 2 ;
c
cc
22

and in the upper right quadrant,

cov[Re(ac+ ), Re(ac− )] = Re(G12cc ) / 2
cov[Re(ac+ ), Im(ac− )] = − Im(H12cc ) / 2
(62)
cov[Im(ac+ ), Re(ac− )] = Im(G12cc ) / 2
cov[Im(ac+ ), Im(ac− )] = Re( H12cc ) / 2 .

24
The variance-covariance matrix of mcR can thus be expressed in terms of G cc and H cc as
⎡Re(G11cc ) − Im(H11cc ) Re(G12cc ) − Im(H12cc )⎤
⎢ ⎥
⎢ Re( H11cc ) Im(G12cc ) Re( H12cc ) ⎥
var cov (mc ) =
white R
cc ⎥
/2. (63)
⎢ cc
Re(G22 ) − Im(H 22 )
⎢ cc ⎥
⎢⎣ Re( H 22 ) ⎥⎦

In order to facilitate more convenient computation of spectral estimates in the

next subsection using the Lomb-Scargle periodogram for the case of irregular times, the
spectra of Cartesian ( u and v ) components of the residual are computed, as opposed to,
for example, rotary spectra. However, the Cartesian spectral quantities are not directly
suitable for scaling var cov white (mcR ) (63), which is cast in terms of the complex
coefficients; the Cartesian spectral quantities are instead appropriate to scale the
variance-covariance matrix of the vector of the four Cartesian (superscript C) cosine/sine
coefficients of the constituent (35),
mcC = [ X cu Ycu X cv Ycv ] . (64)
The corresponding symmetric 4x4 variance-covariance matrix needed is, by definition,
⎡D Duv ⎤
var cov(mcC ) = cov(mcC , mcC ) = ⎢ uuT ⎥, (65)
⎣ Duv Dvv ⎦
where
⎡σ 2 u cov( X u , Y u )⎤ ⎡σ X2 v cov( X v , Y v )⎤
Duu = ⎢ X ⎥ , D = ⎢ ⎥ , and
⎢⎣ σ Y2u ⎥⎦
vv
⎢⎣ σ Y2v ⎥⎦
(66)
⎡cov( X , X ) cov( X , Y )⎤
u v u v
Duv = ⎢ u v u v ⎥
.
⎣ cov(Y , X ) cov(Y , Y ) ⎦
By the defining relations (34) and (35), the 10 unique elements of var cov(mcC ) consist of
those in Duu ,
σ X2 = σ Re(
u
2
+
a )
+ σ Re(
2
−
a )
+ 2 cov[Re(a + ), Re(a − )]
cov( X u , Y u ) = − cov[Re(a + ), Im(a + )] + cov[Re(a + ), Im(a − )]
(67)
− cov[Re(a − ), Im(a + )] + cov[Re(a − ), Im(a − )]
σ Y2u = σ Im(
2
a+ )
+ σ Im(
2
a− )
− 2 cov[Im(a + ), Im(a − )] ;
those in Dvv
σ X2 = σ Im(
v
2
+
a )
+ σ Im(
2
−
a )
+ 2 cov[Im(a + ), Im(a − )]
cov( X v , Y v ) = cov[Im(a + ), Re(a + )] − cov[Im(a + ), Re(a − )]
(68)
+ cov[Im(a − ), Re(a + )] − cov[Im(a − ), Re(a − )]
σ Y2v = σ Re(
2
a+ )
+ σ Re(
2
a− )
− 2 cov[Re(a + ), Re(a − )] ;
and those in Duv ,
cov( X u , X v ) = cov[Re(a + ), Im(a + )] + cov[Re(a + ), Im(a − )] (69)

25
 + cov[Re( a − ), Im(a + )] + cov[Re( a − ), Im(a − )]
cov( X u , Y v ) = cov[Re(a + ), Re(a + )] − cov[Re(a − ), Re(a − )]
cov(Y u , X v ) = − cov[Im(a + ), Im(a + )] + cov[Im(a − ), Im(a − )]
cov(Y u , Y v ) = − cov[Im(a + ), Re(a + )] + cov[Im(a + ), Re(a − )]
+ cov[Im(a − ), Re( a + )] − cov[Im(a − ), Re( a − )] .

Finally, based on combining (60)-(62) and (67)-(69), and dropping the cc

superscripts on G and H for clarity, the unique elements of var cov white (mcC ) in Duu are
σ X2 = [Re(G11 ) + Re(G22 ) + 2 Re(G12 )] / 2
u

cov( X , Y ) = [Im(H11 ) − Im(H12 ) + Im(H 21 ) − Im(H 22 )] / 2

u u
(70)
σ Y2u = [Re( H11 ) + Re( H 22 ) − 2 Re( H12 )] / 2 ;
in Dvv are
σ X2 = [Re( H11 ) + Re( H 22 ) + 2 Re( H12 )] / 2
v

cov( X v , Y v ) = [Im(G11 ) − Im(G12 ) + Im(G21 ) − Im(G22 )] / 2 (71)

σ Y2v = [Re(G11 ) + Re(G22 ) − 2 Re(G12 )] / 2 ;
and in Duv are
cov( X u , X v ) = [− Im(H11 ) − Im(H12 ) − Im(H 21 ) − Im(H 22 )] / 2
cov( X u , Y v ) = [Re(G11 ) − Re(G22 )] / 2
(72)
cov(Y u , X v ) = [− Re( H11 ) + Re( H 22 )] / 2
cov(Y u , Y v ) = [− Im(G11 ) + Im(G12 ) + Im(G21 ) − Im(G22 )] / 2 .

II.C.3. Colored case using spectra of residuals

The residual in general has a non-white, or colored, spectral nature typified by
redness. A means by which to incorporate this (PBL02) in the confidence intervals is to
scale elements of the covariance matrix derived above for the white noise approach using
estimates of the actual residual spectrum at a frequency appropriate to the given
constituent. In the IRLS solution case the appropriate residual is the weighted residual
(48), as pointed out by LJ09.

Following the approach of PBL02, the spectra are considered “locally white”, and
averaged over a fixed set of frequency bands encompassing the main groups of
constituents, after lines of constituents included in the model that fall within that band are
omitted. The nine averaging bands, in cycles per day (cpd), are M0 ± 0.1 cpd, M1 ± 0.2
cpd, M2 ± 0.2 cpd; M3 ± 0.2 cpd, M4 ± 0.2 cpd, M5 ± 0.2 cpd, M6 ± 0.21 cpd; 0.26-0.29 cpd
(includes M7), and 0.30-0.50 cpd (includes M8). For each averaging band, this
computation yields three real values, each a line-decimated, band-averaged, one-sided (2
times the two-sided density, except at the zero and Nyquist frequencies, for which the
one-sided and two-sided values are the same) spectral density: (i) Pe1u esu , the auto-spectral
w w

u 1s
density of e , the weighted u component of the residual; (ii) P
w ewv ewv
, the auto-spectral

26
density of ewv , the weighted v component of the residual; and (iii) Pe1u esv , the rel part of
w w

u v
the cross-spectral density between e and e or cospectrum. The overbars denote the
w w
result of the line-decimation and frequency-band averaging.

The auto- and cross-spectral power terms (not densities) that contribute to the
collective uncertainties in the model parameters of a constituent, index c = 1...nc , with
frequency ωc that lies within the averaged band, are
Pcuu = Pe1u seu Δω
w w

P = Pe1v es v Δω
c
vv
(73)
w w

Pcuv = Pe1u es v Δω ,
w w

where Δω is the frequency resolution of the spectral calculation. For raw input with
length of record (LOR) t nt −t 1 , the frequency resolution is Δω = 1 / LOR e where the
effective LOR is
LORe = (nt /(nt − 1)) LOR , (74)
so formulated because LORe = nt ⋅ Δt in the case of evenly spaced times with time
separation Δt .

In the case of uniformly distributed times, the spectral quantities are computed by
the FFT method after application of a record-length Hanning weighting, using the Matlab
Signal Processing Toolbox function pwelch(). In the case of irregular times, the spectral
estimates are made using the un-normalized, mean-removed, Lomb-Scargle periodogram
(Lomb 1976; Scargle 1982; Press et al. 1992). Calculation of the Lomb-Scargle
periodogram requires specification of a frequency oversampling factor, the amount by
which the grid of frequencies at which the periodogram estimates are computed is more
dense than an equivalent FFT. In common applications of the Lomb-Scargle periodogram
such as astronomy, the goal is to resolve a spectral peak with a certain degree of
confidence, and high oversampling factors (for example 4 or more) are used in order to
increase confidence in peak detection. In the present application peak detection is not the
goal and an oversampling factor of one is used. This approach is taken for practical
reasons as well: for long records the computational burden of the Lomb-Scargle
periodogram increases dramatically, meaning that higher oversampling factors will
require significantly more computing resources, particularly in terms of memory but also
with respect to processing time. Prior to the Lomb-Scargle periodogram calculation, in a
manner similar to the approach of Schulz and Stattegger (1997) for irregularly distributed
times, a record-length Hanning weighting is applied. As a result, in the case of equally
spaced times the UTide Lomb-Scargle periodogram function with oversampling factor of
one returns exactly the same result as does pwelch().

The colored variance-covariance matrix is computed by scaling var cov white (mcC )
(defined in (70)-(72)) using the residual spectral power. The spectral power Pcuu , for the
weighted residual ewu of the u component at the frequency of constituent c , contributes

27
to the variances of X cu and Ycu . Because the elements of Duu consist of these two
variances on the diagonal, and the associated covariance off the diagonal, all elements of
Duu are normalized by its trace and then scaled by Pcuu . Similarly, Dvv is scaled by Pcvv
after normalization by its trace. The cross-spectral power Pcuv contributes to covariances
between one u coefficient ( X cu or Ycu ) and one v coefficient ( X cv or Ycv ). Because all
elements of Duv are covariances (including the elements on the diagonal, unlike for Duu
and Dvv ), the elements of Duv are normalized by the sum of the absolute values of all
elements (instead of its trace), and then scaled by | Pcuv | . These relations are expressed
⎡ Pcuu Duu / tr ( Duu ) | Pcuv | Duv / ∑4 Duv ⎤
var cov colored (mcC ) = ⎢ uv ⎥, (75)
⎢⎣| Pc | Duv / ∑4 Duv
T
Pcvv Dvv / tr ( Dvv ) ⎥⎦
where tr () is the trace operator and ∑4 indicates the element-wise summation. In the
special case that all elements of Duv are zero, in order to avoid division by zero the upper
right four elements and the lower left four elements (all the Duv terms) in (75) are set to
zero.

In certain cases, depending on the model configuration and the spectral

characteristics of the residual, the estimated var cov(mcC ) is not positive semi-definite.
Although generally the difference from positive semi-definiteness is minor, when this
occurs it is not a valid variance-covariance matrix. In this situation, the nearest positive
semi-definite covariance matrix is used instead, as determined using the method of
Higham (2002) with identity weight matrix. The method is iterative and if it does not
converge, a warning message is provided and the off-diagonal elements are simply set to
zero, as an ad-hoc solution.

II.C.4. Implementation: Non-reference and reference consituents

Once the model solution is determined, the optimal complex-valued model
parameter vector m (44) is known. For each non-reference or reference constituent of
index c (where c = 1...nc ), two elements of m are the complex coefficients mc = ac+ and
mc+nc = ac− . These are used, by (35), to compute the four Cartesian cosine/sine
coefficients mcC (64) for the constituent. Next, the associated variance-covariance matrix
var cov colored (mcC ) is computed from (75); this makes use of the spectral quantities (73)
from the weighted residuals to scale the results of the white case (70)-(72), which is in
turn based on expressions for the complex bivariate normal statistics using the model
equation basis function matrix and associated weight matrix (57). Random realizations,
denoted { mcC }, are then generated from the known mcC and var cov colored (mcC ) . In the
UTide code, the random realizations of mcC are generated using the Matlab Statistics
Toolbox function mvnrnd(). By Monte Carlo uncertainty propagation through (9) and
(33), random realizations of the current ellipse parameters, {Lsmaj },{Lsmiη },{θ },{g} , are

28
generated. Standard errors of current ellipse parameters are computed using the median-
average-deviation formulation,
σ Lsmaj = Median[ {Lsmaj } − Median[{Lsmaj }] ] / 0.6745
σL smiη = Median[ {Lsmiη } − Median[{Lsmiη }] ] / 0.6745
(76)
σ θ = Median[ {θ } − Median[{θ }] ] / 0.6745
σ g = Median[ {g} − Median[{g}] ] / 0.6745 .
Finally, the 95% confidence intervals CI are 1.96 times these standard errors, such that,
for example, it is 95% probable that Lsmaj lies between Lsmaj − CI Lsmaj and Lsmaj + CI Lsmaj .
For the white noise case the only change is to use var cov white (mcC ) (70)-(72) instead of
var cov colored (mcC ) (75).

II.C.5. Implementation: Inferred constituents

For each inferred constituent, in the Monte Carlo method, the following approach
is used. First, the realizations {Lˆsmaj },{Lˆsmiη },{θˆ},{gˆ } of the current ellipse parameters for
its reference constituent, computed as just described, are used to compute realizations
{aˆ + },{aˆ − } of the complex coefficients for the reference constituent, by (10). Next,
realizations {aˆˆ + },{aˆˆ − } of the complex coefficients for the inferred constituent are
computed from those of the reference constituent using (13), then converted to
ˆ ˆ ˆ ˆ
realizations { Xˆ u },{Yˆ u },{ Xˆ v }, {Yˆ v } of the cosine/sine coefficients of the inferred
ˆ ˆ ˆ
constituent by (35), and finally to realizations {Lˆsmaj },{Lˆsmiη },{θˆ},{gˆˆ } of the current
ellipse parameters of the inferred constituent, by (9) and (33). From the realizations of the
current ellipse parameters of the inferred constituent, the confidence intervals are
computed as in (76).

Confidence intervals for inference constituents can also be computed when using
the linearized method instead of Monte Carlo. For any constituent (inferred or not), the
linearized method computes the variances of the current ellipse parameters,
σ L2smaj , σ L2smiη , σ θ2 , σ g2 , under the presumption of zero cross-correlations and by error
propagation rules, from the variances of the cosine/since coefficients, σ X2 u , σ Y2u , σ X2 v , σ Y2v
(PBL02). The latter are known for the reference constituents, as described above for the
white or colored case, but need to be computed for the inference constituents. From (35)
and (13), in the case of the lk th constituent inferred from the k th reference constituent,
ˆ
Xˆ luk = Re(aˆˆl+k + aˆˆl−k ) = Re( Rl+k aˆ k+ ) + Re( Rl−k aˆ k− )
ˆ
Yˆlku = − Im(aˆˆl+k − aˆˆl−k ) = − Im(Rl+k aˆ k+ ) + Im(Rl−k aˆ k− )
(77)
ˆ
Xˆ lvk = Im(aˆˆl+k + aˆˆl−k ) = Im(Rl+k aˆ k+ ) + Im(Rl−k aˆ k− )
ˆ
Yˆlkv = Re(aˆˆl+k − aˆˆl−k ) = Re( Rl+k aˆ k+ ) − Re( Rl−k aˆ k− )

29
where the double-hat variables represent inferred constituents and the hat variables
represent reference constituents, as in Section II.A. The properties of complex numbers
z1 and z2 , that Re( z1 z 2 ) = Re( z1 ) Re( z2 ) − Im( z1 ) Im( z2 ) and
Im( z1 z2 ) = Im( z1 ) Re( z2 ) + Re( z1 ) Im( z2 ) , mean these expressions can be written
ˆ
Xˆ luk = Re( Rl+k ) Re( aˆ k+ ) − Im(Rl+k ) Im(aˆk+ ) + Re( Rl−k ) Re( aˆ k− ) − Im( Rl−k ) Im(aˆ k− )
ˆ
Yˆlku = − Im(Rl+k ) Re(aˆ k+ ) − Re( Rl+k ) Im(aˆ k+ ) + Im(Rl−k ) Re(aˆ k− ) + Re( Rl−k ) Im(aˆ k− )
(78)
ˆ
Xˆ lvk = Im( Rl+k ) Re( aˆ k+ ) + Re( Rl+k ) Im(aˆ k+ ) + Im( Rl−k ) Re( aˆ k− ) + Re( Rl−k ) Im(aˆ k− )
ˆ
Yˆlkv = Re( Rl+k ) Re(aˆ k+ ) − Im(Rl+k ) Im(aˆ k+ ) − Re( Rl−k ) Re( aˆ k− ) + Im( Rl−k ) Im(aˆk− ) .
By uncertainty propagation rules, and ignoring cross-correlations just as in the linearized
case for non-inferred constituents, it follows that
σ 2ˆˆ u = σ 2ˆˆv = [Re( Rl+k )]2 σ Re(
2
aˆ + )
+ [Im(Rl+k )]2 σ Im(
2
aˆ + )
X lk Ylk k k

+ [Re( R )] σ −
lk
2 2
Re( aˆ k− )
+ [Im(R )]2 σ Im(
2 −
lkaˆ − ) k
(79)
and σ 2
ˆ =σ 2
ˆ = [Im(R )] σ
+
lk
2 2
Re( aˆk+ )
+ [Re( R )] σ +
lk
2 2
Im( aˆk+ )
Yˆ u
lk Xˆ v
lk

+ [Im(Rl−k )]2 σ Re(

2
aˆ − )
+ [Re( Rl−k )]2 σ Im(
2
aˆ − )
.
k k

By similar application of uncertainty propagation to the real and imaginary parts of (32),
σ Re(
2
aˆk+ )
= σ Re(
2
aˆk− )
= (σ X2ˆ u + σ Y2ˆv ) / 4
k k
(80)
σ Im(aˆ + ) = σ Im(aˆ − ) = (σ Xˆ v + σ Yˆ u ) / 4 .
2 2 2 2
k k k k

It is clear then that the known values of σ , σ , σ , σ Y2ˆv , together with the known
2
Xˆ ku
2
Yˆku
2
Xˆ kv k

+ −
values of R , R , are sufficient with (79) and (80) to compute the needed variances of the
lk lk

inferred cosine/sine coefficients, and hence by the linearization formulae the variances
and confidence intervals of the current ellipse parameters of the inferred constituents.

II.C.6. One-dimensional case

In the case of one-dimensional raw input (real x raw ; a + = a −* ) the above
development applies unmodified except that only the first two elements of the Cartesian
cosine/sine coefficients (64) are non-zero, only Duu is non-zero in (65) and (66), only
(67) and (70) need be considered, and the scaling in (75) is based solely on Pcuu from
(73). With the additional assumption that cov( X u , Y u ) = 0 , the white noise floor result in
the one-dimensional case corresponds to that of F77 and FH89.

II.D. Constituent selection diagnostics

This section reviews various diagnostics that are useful to determine which
constituents are to be included in a model solution. All diagnostics are defined and
explained first for the case of uniformly distributed times, then comments regarding the
case of irregular times are made. An explanation of the structure of the diagnostic table
included by default in the UTide output, which is designed to present the diagnostics

30
collectively in a concise fashion for convenient inspection and usage, is given in Section
III.C.1.d.

Introductory guidance on the constituent selection process can be found in a

number of references, including FH89 and Section 5.5 of the textbook by Emery and
Thomson (1998). In general constituent selection is an iterative process, in which initial
iterations of the analysis include more constituents than are expected to capture
significant energy and/or to be resolved from the other included constituents. This initial
set of constituents tends to be created using auxiliary information such as results from
other, prior analyses based on similar raw input from the same region. Then by inspecting
diagnostics of the output, the decision to keep certain constituents in the model or remove
them from it, or to infer them in a subsequent solution, can be made and the next iteration
carried out.

The decision to keep, omit, or infer (presuming sufficient needed auxiliary

information is available) a given constituent rests on two main, related criteria: (1) the
extent to which it is independent from the other included constituents, and (2) the extent
to which it is significant, relative to noise and to the other included constituents. By
iterative solutions that include different combinations of constituents, one can determine
the extent to which omitting a constituent affects the diagnostics for the remaining
included constituents. In general, the goal is to remove constituents that are not
sufficiently independent and significant, which can be confirmed by verifying that
omitting them causes sufficiently small changes to diagnostics of the remaining
constituents.

On this basis, diagnostics quantities are put in to three groups here: one group
related to confirming independence of constituents, one group related to ascertaining
significance of constituents, and one group to characterize reconstructed harmonic fits
that superpose some or all of the constituents.

II.D.1. Diagnostics related to constituent independence

II.D.1.a. Conventional Rayleigh criterion (RR)
In the case of raw input with uniformly distributed times the widely accepted
approach to constituent selection is an automated decision tree method, based on the
equilibrium tide and the conventional Rayleigh criterion, as developed by F77 and
summarized in FH89. The conventional Rayleigh criterion (superscript R) states that two
constituents, of frequencies ωq1 and ωq2 , are resolvable by a record with uniformly
distributed times if
⎛ LORe ⎞
R R (q1 , q2 ) = ⎜ ⎟ / R ≥ 1, (81)
⎜ 1 / | ωq − ωq | ⎟ min
⎝ 2 1 ⎠
where Rmin is a minimum threshold taken to be 1 in most cases. For Rmin =1, the criterion
is equivalent to requiring that the record length (numerator) is sufficiently long that the
two frequencies are resolved from each other with respect to traditional spectral
estimation. The F77 decision tree essentially compares all constituent pairs and omits any

31
constituent that does not meet the conventional Rayleigh criterion, relative to a
constituent that has larger equilibrium tide amplitude, for a user-specified Rmin value.
The decision tree can be applied for any choice of Rmin , for example a value greater than
one in order to conservatively omit a larger number of constituents (as one example, this
could be appropriate in the case that the condition number of the basis function matrix is
high, as explained below). The decision tree is a standard option in UTide, as it is for
t_tide (PBL02).

II.D.1.b. Noise-modified Rayleigh criterion (RNM)

It is recognized that the conventional Rayleigh criterion (81) is incomplete,
because it does not take in to account the fact that noise in a record affects its ability to
resolve constituents from each other. As a consequence, in the case of a strongly tidal
record, the conventional Rayleigh criterion is overly conservative, rejecting constituents
that may be well resolved from each other. For these reasons, Munk and Hasselmann
(1964) suggested a modified Rayleigh criterion such that the record duration required to
resolve a pair of frequencies was scaled by the square root of the SNR. Munk and
Hasselmann did not provide a particular expression for SNR, nor has one been commonly
adopted in the literature, so it is suggested here to use the constituent-specific noise-
modified (superscript NM) Rayleigh criterion for constituent q1 relative to constituent
q2 , defined as
( )
R NM (q1 , q2 ) = R R (q1 , q2 ) ( SNRq1 + SNRq2 ) / 2 ≥ 1 .
(82)
The denominator is the average of the SNRs of the two constituents, where SNR for an
individual constituent is
Lsmaj + Lsmiη
2 2

SNRq = 2 , (83)
σ Lsmaj + σ L2smiη
in which, for convenience, the q subscripts are dropped on the right hand side. (In the
one-dimensional case this expression for SNR is the same as used by PBL02, but in the
two-dimensional case it is slightly more general due to inclusion of the terms related to
the minor axis.)

In effect, criterion (82) states that the constituents are resolved even for record
lengths that are not long enough for R R to be greater than 1, as long as the SNR factor in
(82) is greater than 1. Conversely, if the SNR factor is less than 1, the record length must
be longer than for the conventional Rayleigh criterion, since R R is then required to be
greater than 1. It should be borne in mind that even the noise-modified criterion is an
incomplete metric, since in the limit of no noise it incorrectly suggests all constituents
will be resolvable. Nonetheless it can provide useful information, in combination with the
conventional Rayleigh criterion, particularly for analysis of records in which the non-tidal
signal is comparable to or larger than the tidal signal.

II.D.1.c. Condition number (K) relative to SNR of entire model (SNRallc)

FH89 pointed out that by matrix theory there is an upper bound on the fractional
error of the norm of the model parameters,

32
 ′
|| m′ − m || || x raw − x raw ||
≤K , (84)
|| m || || x raw ||
where || ⋅ || is a suitable norm, the primed variables indicate a case including random
noise, the unprimed variables indicate a case with no random noise, and K is the
condition number (ratio of largest and smallest singular values) of the matrix B . A
reasonable interpretation of the fraction on the RHS of (84), for L2 norm, is the inverse
of the SNR for the entire model including contributions from all directly modeled
constituents,
m H B HWBM
SNRallc = . (85)
σ MSM 2

It follows that
SNRallc / K > 1 (86)
is a criterion that, when met, means the upper bound of the fractional uncertainty in the
norm of the model parameters is of order one or less. Comparison of the condition
number and SNRallc can therefore in some cases be a useful diagnostic of whether or not
the constituents are, collectively, well resolved by the model; in a model that includes
constituents with small differences in frequency, the condition number will be higher, so
in order to resolve the constituents well by this criterion, SNRallc must be high enough to
exceed the condition number.

By this line of reasoning an additional criterion for whether the constituents are
independent from each other, collectively, could be the requirement that SNRallc is at
least as high as K . However, in practice there is a limit to the usefulness of this approach
because (a) (84) gives an upper bound only and hence the model parameter uncertainties
may be sufficiently small even when SNRallc is less than K; and, of somewhat less
importance, (b) (86) is based on the L2 norm but the hybrid L1 / L2 norm applies in the
case of the IRLS solution. Furthermore, it is clear that because both SNRallc and K are
defined using the entire model, as opposed to a specific constituent pair, comparing them
can provide guidance about the model as a whole (including its mean and, if included,
trend), but does not provide information relevant to any specific constituent or pair of
constituents.

II.D.1.d. Maximum correlation (Corrmax) between model parameters

A related means by which to gauge the extent to which a pair of constituents is
independent is to use the cross-covariances among their model parameters. High cross-
covariance between model parameters of a pair of constituents indicates they that are less
independent, so the less energetic of the pair should be considered for removal from the
model, or for inclusion by inference if possible. Here, a new diagnostic for cross-
covariances is presented and expressed using elements of the confidence interval
development of Section II.C.

33
 The “maximum correlation” diagnostic for constituents with indices c1 and c2 is
the maximum magnitude of the 16 correlation coefficients between the elements of their
Cartesian model parameter vectors mcC1 = [ X cu1 Ycu1 X cv1 Yc1v ] and
mcC2 = [ X cu2 Ycu2 X cv2 Ycv2 ] (64). It can be written
corrmax (c1 , c2 ) = max[ corr ( X cu1 , X cu2 ) , corr ( X cu1 , Ycu1 ) , L corr (Yc1v , Yc1v ) ] , (87)
where the correlations are defined in the standard way, for example,
cov( X cu1 , X cu2 )
corr ( X c1 , X c2 ) =
u u
. (88)
σX σX
u u
c1 c2

The standard deviations in the denominator of (88) are calculated using the
diagonal elements of var cov white (mcC ) , expressions for which are given in (70)-(72). The
covariances in the numerator of (88) are the elements of (65) when generalized to a
constituent pair,
⎡ D c1c2 Duvc1c2 ⎤
cov white (mcC1 , mcC2 ) = ⎢ uuc c c1c2 ⎥
, (89)
⎣ vu
D D vv ⎦
1 2

where the 2 x 2 D submatrices are

⎡cov( X cu1 , X cu2 ) cov( X cu1 , Ycu2 )⎤ ⎡cov( X cu1 , X cv2 ) cov( X cu1 , Ycv2 )⎤
Duu = ⎢
c1c2
u u u u ⎥
Duv = ⎢
c1c2
u v u v ⎥
⎣⎢ cov(Yc1 , X c2 ) cov(Yc1 , Yc2 ) ⎦⎥ ⎣⎢ cov(Yc1 , X c2 ) cov(Yc1 , Yc2 ) ⎦⎥
(90)
⎡cov( X cv1 , X cu2 ) cov( X cv1 , Ycu2 )⎤ ⎡cov( X cv1 , X cv2 ) cov( X cv1 , Ycv2 )⎤
Dvu = ⎢
c1c2
v u v u ⎥
Dvv = ⎢
c1c2
v v v v ⎥
.
⎣⎢ cov(Yc1 , X c2 ) cov(Yc1 , Yc2 ) ⎦⎥ ⎣⎢ cov(Yc1 , X c2 ) cov(Yc1 , Yc2 ) ⎦⎥
The expressions in (57) are generalized to their two-constituent forms,
⎡ Gcall Gcall
1 ,c2 + nc
⎤
G = ⎢ all
c1c2 1 ,c2
all ⎥
⎢⎣Gc1+nc ,c2 Gc1+nc ,c2 +nc ⎥⎦
(91)
⎡ H all
H all
c1 ,c2 + nc
⎤
H c1c2 = ⎢ all1 2
c ,c
all ⎥.
⎢⎣ c1+nc ,c2
H H c1 + nc ,c2 + nc ⎥
⎦
It follows from the definitions (34) and (35) with relations of the same form as (60)-(62),
and omitting the c1c2 superscripts from G and H for clarity, that the terms of Duuc1c2 are
cov( X cu1 , X cu2 ) = [Re(G11 ) + Re(G12 ) + Re(G21 ) + Re(G22 )] / 2
cov( X cu1 , Ycu2 ) = [Im(H11 ) − Im(H12 ) + Im(H 21 ) − Im(H 22 )] / 2
(92)
cov(Ycu1 , X cu2 ) = [− Im(G11 ) − Im(G12 ) + Im(G21 ) + Im(G22 )] / 2
cov(Ycu1 , Ycu2 ) = [Re( H11 ) − Re( H12 ) − Re( H 21 ) + Re( H 22 )] / 2 ;
the terms of Duvc1c2 are
cov( X cu1 , X cv2 ) = [− Im(H11 ) − Im(H12 ) − Im(H 21 ) − Im(H 22 )] / 2
cov( X cu1 , Ycv2 ) = [Re(G11 ) − Re(G12 ) + Re(G21 ) − Re(G22 )] / 2 (93)
cov(Y , X ) = [− Re( H11 ) − Re( H12 ) + Re( H 21 ) + Re( H 22 )] / 2
u
c1
v
c2

34
 cov(Ycu1 , Ycv2 ) = [− Im(G11 ) + Im(G12 ) + Im(G21 ) − Im(G22 )] / 2 ;
the terms of Dvuc1c2 are
cov( X cv1 , X cu2 ) = [Im(G11 ) + Im(G12 ) + Im(G21 ) + Im(G22 )] / 2
cov( X cv1 , Ycu2 ) = [− Re( H11 ) + Re( H12 ) − Re( H 21 ) + Re( H 22 )] / 2
(94)
cov(Ycu1 , X cv2 ) = [Re(G11 ) + Re(G12 ) − Re(G21 ) − Re(G22 )] / 2
cov(Ycu1 , Ycv2 ) = [Im(H11 ) − Im(H12 ) − Im(H 21 ) + Im(H 22 )] / 2 ;
and the terms of Dvvc1c2 are
cov( X cv1 , X cv2 ) = [Re( H11 ) + Re( H12 ) + Re( H 21 ) + Re( H 22 )] / 2
cov( X cv1 , Ycv2 ) = [Im(G11 ) − Im(G12 ) + Im(G21 ) − Im(G22 )] / 2
(95)
cov(Yc1v , X cv2 ) = [− Im(H11 ) − Im(H12 ) + Im(H 21 ) + Im(H 22 )] / 2
cov(Yc1v , Ycv2 ) = [Re(G11 ) − Re(G12 ) − Re(G21 ) + Re(G22 )] / 2 .
(The case c1 = c2 , not considered in this section, would yield the three groups of
expressions (70)-(72).)

In the case of one-dimensional raw input, the Cartesian model parameter vectors
C C
m , m have only two elements each, so corrmax is the maximum among 4 (not 16)
c1 c2

correlations. The above expressions are unchanged except for the fact that only the Duu
portions (upper left quadrants) of the matrices are nonzero and need be considered.

The corrmax diagnostic is analogous to the correlation diagnostic for one-

dimensional raw input based on the singular value decomposition, developed by
Cherniawsky et al (2001) and used in FCB09. In the analysis of FCB09, correlations up
to about 0.2 were considered acceptable. However, experimentation is required in order
to determine acceptable levels of corrmax for each individual analysis.

In the UTide constituent selection diagnostic table (described in Section III.C.1.d

below), corrmax values are computed only for adjacent pairs of constituents in the model,
that is, those pairs that have frequencies nearer to each other than to any other
constituents (excluding inferred constituents). While this is not as complete as computing
and examining corrmax values for every pair of constituents, it is a practical approach
based on the expectation that constituent pairs with adjacent frequencies are typically
more likely to lack independence.

II.D.2. Diagnostics related to constituent significance

II.D.2.a. Signal to noise ratio (SNR)
Constituents q = 1...nallc for a given solution are considered to be significant with
respect to the noise in the raw input if their SNRq (83) obeys

35
 SNRq ≥ SNRmin , (96)
where SNRmin is a minimum threshold value. Common practice takes SNRmin to be 1 or
2, but in certain situations other values may be appropriate. For example, a higher value
might be used in order to conservatively neglect marginally significant constituents if the
estimates of the standard deviations of the model parameters, on which the SNR values
are founded (83), are thought to be biased low.

II.D.2.b. Percent energy (PE)

The model solution x mod (42) is a reconstructed harmonic fit that superposes all
the constituents. Independently of their significance with respect to the SNR threshold,
the relative importance of a constituent can be gauged by the percent energy (Codiga and
Rear 2004) it contributes to the model solution. For constituent q , the percent energy is
E
PEq = 100 nallc q ,
∑ Eq
(97)
q =1

where
Eq = ( L2smaj + L2smiη ) , (98)
the q subscripts have been dropped on the right hand side of (98), and the summed Eq
values equal 100. In the two-dimensional case of horizontal velocity components, Eq is
proportional to the kinetic energy; in the one-dimensional case of sea level it is a gauge of
potential energy. It is useful to rank the constituents by their percent energy so that the
importance of the constituents in an amplitude-weighted sense is clear. This ranking
usually parallels the SNR ranking but it can be less sensitive to the confidence interval
calculation method, and in certain cases provides important complementary information
to SNR.

II.D.3. Diagnostics characterizing reconstructed fits (PTVall, PTVsnrc)

A diagnostic of the model solution is its percent tidal variance,
TVallc | x mod − x ⋅ I (nt ,1) − x& ⋅ t |2
PTVallc = 100 = 100 raw , (99)
TVraw | x − x ⋅ I (nt ,1) − x& ⋅ t |2
where TVallc and TVraw (each in units of squared raw input units) are the tidal variance,
after removal of the mean and trend, of the (all-constituent) model solution and the raw
input, respectively.

Reconstructed fits other than the model solution can be calculated based on
inclusion of a subset of the constituents. Denote by qincl the subset of nincl ≤ nallc
constituents (among non-reference, reference, and inferred constituents) that are chosen,
based on some criteria, to be included in a reconstructed fit. Following model equation
(2), the reconstructed fit computed using that subset of constituents is
nincl

∑ (E
qincl =1
iqincl aq+incl + Eiq* incl aq−incl ) + x + x& ⋅ (ti − t ref ) . (100)

36
Substituting this in (99) for x mod yields the percent tidal variance of the reconstructed fit,
nincl

TVincl
|
q =1
∑ (E iqincl aq+incl + Eiq* incl aq−incl ) |2
(101)
PTVincl = 100 = 100 incl ,
TVraw TVraw
based on the ratio of the tidal variance of the reconstructed fit to that of the raw input. A
straightforward example is the reconstructed fit using only the constituents that meet the
SNR criterion (“snrc”) for significance (96), denoted by indices qsnrc . The percent tidal
variance of the corresponding reconstructed fit is
nsnrc

TVsnrc
| ∑ (E
qsnrc =1
iqsnrc aq+snrc + Eiq* snrc aq−snrc ) |2
(102)
PTVsnrc = 100 = 100 .
TVraw TVraw

II.D.4. Considerations for irregularly distributed times

When the times of the raw input are distributed irregularly, some of the
underlying assumptions behind the above diagnostics are violated, making proper
constituent selection a major challenge. Irregular times can be viewed loosely as if certain
parts of the record have more highly concentrated temporal sampling and could resolve a
higher number of constituents, whereas the opposite is true for other parts of the record. It
follows that to be conservative one should select constituents based on the limitations of
the portions of the record where temporal sampling is most sparse. However, there are no
guidelines or accepted practices for how to carry out this goal on the basis of knowledge
about the distribution of irregular times.

Even though strictly speaking the underlying assumptions are violated, for the
case of irregularly distributed times it is nonetheless straightforward to follow the above
approach, without modification, for both the implementation of the automated decision
tree and the computation of all of the above diagnostics. As a result, although it is
certainly not rigorously justified by the underlying statistics, it is straightforward to
calculate and inspect all the same diagnostics in the case of irregularly distributed times
as are used for regularly distributed times. This approach is at least a starting point, in the
absence of suitable diagnostics that are well-defined in terms of the characteristics of the
distribution of the irregular times.

It will of course be most justified to make use of diagnostics so calculated in cases

when the irregularity in the distribution of the times is modest. An example of modest
irregularity is a distribution of times that deviates from an equispaced time series only by
small random deviations, as opposed to including numerous gaps that are long, in the
sense that their duration spans at least several samples in a comparable regular time
series.

When the times are irregular a crude but practical approach to being conservative,
in the sense of omitting constituents that might not be resolved from each other, is to use
the same diagnostics but judge them in relation to different threshold values. For
example, if the Rayleigh criteria ((81),(82)) are used with Rmin =1 for a uniformly

37
sampled record, then for an irregularly sampled record an Rmin value higher than 1 can be
used, in order to be more stringent in omitting constituents.

There is no accepted practice for determining the multiple by which to increase

Rmin based on the arbitrary, but known, irregular distribution of the times. As a starting
point, consider the comparison between a record with uniform sampling of time
difference Δt between samples, and a modestly irregular record with time differences
between samples that are variable but Gaussian with mean Δt and standard deviation σ Δt
less than the mean. A reasonable choice for the appropriate ratio by which to increase
Rmin is
irregular
Rmin / Rmin = Δt /(Δt − ασ Δt ) (103)
where α is a constant, nominally 1, that can be increased (as long as the denominator
remains positive) in order to implement more conservative constituent rejection. This
approach is equivalent to decreasing the numerator in the Rayleigh criterion (81), from
the length of record or (nt − 1)Δt in the uniformly distributed times case, to
(nt − 1)(Δt − ασ Δt ) . Most real-world raw inputs with irregular temporal sampling are
likely to have non-Gaussian distributions of the time differences, in which case some
improved robustness should follow from using the median (instead of the mean) for Δt
and the median-absolute-deviation (instead of the standard deviation) for σ Δt .

However, the deviation of the irregular sampling distribution from Gaussian is

commonly very severe, with numerous long gaps, such that in general (103) may not be
applicable and it is expected that some empirical experimentation will be necessary to
irregular
arrive at an acceptable value of Rmin .

38
III. The UTide Matlab functions
UTide consists of a pair of Matlab functions designed to be easy to understand
and implement: ut_solv() to carry out the analysis, the results of which are passed to
ut_reconstr() for reconstructing a hind-cast or forecast/prediction, or “fit”, as needed.

III.A. Obtaining and using UTide

There are three UTide files: ut_solv.m, ut_reconstr.m, and ut_constants.mat. The
current version can be downloaded in a compressed bundle, together with this report, at
ftp://www.po.gso.uri.edu/pub/downloads/codiga/utide/UTideCurrentVersion.zip; the
version history, and the bundle file for older versions, will be available in that same
folder. Uncompress the zipfile contents to a single folder/directory and make sure it is on
the Matlab path. No other formal installation is needed.

UTide makes use of functions from both the Signal Processing Toolbox
(pwelch(), cpsd(), hanning()) and the Statistics Toolbox (robustfit(), mvnrnd()). Thus if
either of these toolboxes are not available, executing UTide in its default configuration
will result in errors. If the Signal Processing Toolbox is not available, the colored method
for confidence intervals will not be possible, so to avoid such errors UTide must be run
using the ‘White’ option flag (explained below). Similarly, if the Statistics Toolbox is not
available, the IRLS solution method and the Monte Carlo confidence interval approach
will not be possible, so to avoid such errors UTide must be run using both the ‘OLS’ and
‘LinCI’ option flags (explained below).

As noted by LJ09, calls to robustfit() to carry out the IRLS solution commonly
result in a warning or error related to reaching the interation limit. This can be remedied
by editing the line iterlim = 50; in the file statrobustfit.m, found in the
MATLABROOT\toolbox\stats\private folder, to replace 50 by a sufficiently larger
maximum number of iterations, for example 500. Other than permitting a higher number
of iterations, this has no effect on the functionality of robustfit() or any other aspects of
Matlab. In general, increasing the tuning parameter can be a remedy, in cases for which
the iteration limit is reached. When the iteration limit is reached, the results for current
ellipse parameters (amplitude/phase parameters in the 1D case), means, slopes (if trend
included), and confidence intervals are set to NaN.

III.B. Quick start suggestions

The following steps outline the most efficient way to get started quickly with an
initial computation (for analysis of a single record) using the default settings of UTide.
First, read the opening portion of Section III.C, which briefly summarizes the syntax.
Next, read section III.C.1.a, where the formats of the input parameters to ut_solv() are
explained, and manipulate the raw values you wish to analyze in order to create t_raw,
u_raw, v_raw, lat, and cnstit in the needed formats (for the latter, a typical initial choice
is ‘auto’, which invokes the F77 decision tree to carry out constituent selection). Then,
pass them in to ut_solv() to generate the structure coef. If the computational burden is too
much for the available resources, try again but also pass in the ‘OLS’, ‘white’, and
‘LinCI’ flags, after the cnstit input, as explained in Section III.F. By default the

39
diagnostics table coef.diagn.table (explained in Section III.C.1.d) will be displayed at
runtime, and can be inspected in order to iterate towards a more refined constituent
selection for a subsequent call to ut_solv(). Read Section III.C.1.b to understand the
contents and formats of the field in coef, which include all analysis results and are
available for manipulation in further custom analysis or plotting (of the constituent
statistics, coefficients and confidence intervals, current ellipses, etc). Finally, if
reconstructing a hind-cast or forecast/prediction fit (superposed harmonics) using the
resulting coefficients is desired, construct a vector t_fit (in the same format as t_raw) of
times and pass them in to ut_reconstr(), with specification of the subset of constituents to
include (default includes constituents with SNR≥2), as explained in Section III.C.2. To
treat a group of records, see Section III.D.

III.C. Functionality and syntax and for a single record

Syntax for two-dimensional raw input, such as velocities, is

coef = ut_solv ( t_raw, u_raw, v_raw, lat, cnstit , {options} );

[ u_fit, v_fit ] = ut_reconstr ( t_fit, coef , {options} );

and syntax for one-dimensional raw input, such as sea level, is

coef = ut_solv ( t_raw, sl_raw, [], lat, cnstit , {options} );

[ sl_fit, ~ ] = ut_reconstr ( t_fit, coef , {options} );

for which a brief overview of the various variables/parameters is as follows:

• coef is the output structure generated by ut_solv() and accepted by ut_reconstr();
• the raw input records have times t_raw, current components u_raw/v_raw or sea
level values sl_raw, and latitude lat;
• cnstit specifies the constituents to be included in the model;
• t_fit contains the arbitrary times at which the reconstructed output (u_fit/v_fit or
sl_fit) is computed; and
• {options} represents optional inputs, as explained in more detail below.
More detailed explanations of variables that must be passed in to ut_solv() are given in
Section III.C.1.a; a detailed explanation of the contents of the output from ut_solv(), the
structure coef, is given in Section III.C.1.b; explanations of the default configuration and
option flags for ut_solv() are given in Section III.C.1.d. Information regarding the inputs
and outputs of ut_reconstr(), and its default configuration and option flags, is given in
Section III.C.2.

III.C.1. Solving for coefficients with ut_solv()

III.C.1.a. Input parameter descriptions
The times are in a real-valued column vector t_raw that
• contains Matlab “datenum” values for the sampled times in coordinated universal
time UTC (Greenwich mean time, GMT), with units of days. For the unfamiliar,
the definition and characteristics of “datenum” values are explained in the Matlab
documentation; see the help descriptions for functions datenum(), datevec(), and

40
 datestr(). For example, to generate an N-day time vector with hourly resolution,
1:(1/24):N will not suffice; instead use, e.g., for a start time of 8:15AM on Nov.
1, 2001, datenum(2001,11,1,8,15,0)+(1:(1/24):N);
• can have NaN values, but they will be removed during analysis, along with the
corresponding (NaN or non-NaN) associated u_raw/v_raw or sl_raw values;
• contains values that can be either regularly/uniformly distributed (“equispaced”)
or irregularly distributed;
• is considered equispaced if (after NaNs are removed) the Matlab expression
var(unique(diff(t_raw)))<eps is true [where var(), unique(), diff() , and eps() are
built-in functions], in which case FFT methods ( pwelch(), cpsd() ) are used for
the periodogram of the residual for colored confidence intervals; and
• is considered irregularly distributed if (after NaNs are removed)
var(unique(diff(t_raw)))≥eps, in which case, the Lomb-Scargle periodogram of
the residual is used for colored confidence intervals.
Raw input vectors u_raw and v_raw, or sl_raw
• are real-valued column vectors of the same size as t_raw; and
• are permitted to include NaNs, in which case if the times are equispaced (see
above) the NaNs will be filled by linear interpolation prior to FFT spectral
analysis.
The scalar input lat is
• the latitude (in decimal degrees, positive North and negative South);
• a required input, because nodal/satellite corrections are implemented by default;
and
• not used if nodal/satellite corrections are omitted (‘NodsatNone’ option).
The input cnstit determines the constituents included in the model and is one of the
following:
• the string ‘auto’ (not case-sensitive)
o the F77 automated decision tree (with default value Rmin (81) of 1, unless
a different value is specified with the ‘Rmin’ option) is implemented;
o if inference/reference constituents are specified together with this option,
the inference/reference constituents are included in the model whether or
not they are selected by the decision tree; a constituent selected by the
decision tree will be removed from the non-reference group of constituents
if it is designated as a reference constituent or an inferred constituent.
• a cell array of 4-character strings (not case-sensitive)
o each string contains the name of a non-reference constituent to be
included, including trailing blanks if needed to fill out 4 characters;
o the constituents available are those (including shallow-water constituents)
in the const.name variable in the “ut_constants.mat” file, for example ‘M2
’, ‘MSF ’, etc.

III.C.1.b. Output structure coef

The main output of the call to ut_solv() is a single structure, coef. It consists of
various scalar, vector, and string array fields, which are described here.

41
 The sizes of the cell array and numeric fields and subfields of coef are either 1x1
or nallc x1. For fields with size nallc x1, there is one value for each of the constituents, and
they are ordered such that the results for a given constituent are in the same element of
each such field. By default the elements are ordered by decreasing percent energy
PEq (97). Using the ‘OrderCnstit’ option flag they can instead be ordered by decreasing
SNR, increasing frequency, or in a user-specified sequence (the last option is possible as
long as the automatic decision tree is not used for constituent selection, that is, cnstit is
not ‘auto’).

There are three main groups of results in coef: primary results, auxiliary results,
and diagnostic results. The primary results are the nallc x1 cell array
• coef.name, an array of 4-character constituent names,
the following real-valued nallc x1 vectors,
• in the two-dimensional case ((9),(76)),
o coef.Lsmaj, the current ellipse major axis length (units of u_raw/v_raw)
o coef.Lsmaj_ci, the 95% confidence interval for coef.Lsmaj,
o coef.Lsmin, the current ellipse minor axis length,
o coef.Lsmin_ci, the 95% confidence interval for coef.Lsmin,
o coef.theta, the current ellipse orientation angle (degrees)
o coef.theta_ci, the 95% confidence interval for coef.theta,
o coef.g, the Greenwich phase lag (degrees) of the vector velocity,
o coef.g_ci, the 95% confidence interval for coef.g,
• in the one-dimensional case ((20),(76))
o coef.A, the amplitude (units of sl_raw)
o coef.A_ci, the 95% confidence interval for coef.A
o coef.g, the Greenwich phase lag (degrees),
o coef.g_ci, the 95% confidence interval for coef.g,
and the following real-valued scalars,
• in the two-dimensional case,
o coef.umean and coef.vmean, the mean values u , v (2) (u_raw/v_raw
units) for the u/v components,
o coef.uslope and coef.vslope, the trend slope u&, v& (2) (u_raw/v_raw units
per day) for the u/v components (omitted if the trend is not included in the
model),
• in the one-dimensional case,
o coef.mean, the mean value (sl_raw units),
o coef.slope, the trend slope (sl_raw units per day), omitted if the trend is
not included in the model.
The field coef.results is a character array that presents all the above fields in an easy to
read format and is displayed during runtime by default. If the IRLS solution method is
used and it does not converge, a warning is given in coef.results, and the values of the
above fields are set to NaN.

The auxiliary results are included as fields of coef.aux and consist of

42
 • coef.aux.rundescr, a cell string array with a descriptive explanation of the run,
• coef.aux.opt, a series of fields containing the option settings and input
parameters,
the nallc x1 vectors
• coef.aux.frq, the frequencies of the constituents (cycles per hour),
• coef.aux.lind, the list indices (used by ut_reconstr()) of the constituents as
referenced in the file ut_constants.mat,
and the scalars
• coef.aux.lat, the latitude,
• coef.aux.reftime, the reference time (datenum UTC/GMT as for t_raw).

The diagnostic results are included in coef.diagn, which is created unless the
‘NoDiagn’ option is selected, and consists of the following nallc x1 fields, each with the
same (default decreasing PEq (97); otherwise can be specified, using the ‘OrderCnstit’
option, to be decreasing SNR, increasing frequency, or user-specified) element order:
• coef.diagn.name, the four-character constituent names, in order of decreasing
PEq (identical to coef.name; except when the ‘OrderCnstit’ option is used, in
which case it is the same list of constituents but ordered differently),
• coef.diagn.PE, the percent energy (97),
• coef.diagn.SNR, the signal to noise ratio (83),
• coef.diagn.lo.name, the name of the constituent with the nearest lower frequency
(NaN if no other constituent has a lower frequency),
• coef.diagn.lo.RR, the conventional Rayleigh criterion ratio R R (81) relative to the
constituent with the nearest lower frequency (NaN if no other constituent has a
lower frequency),
• coef.diagn.lo.RNM, the noise-modified Rayleigh criterion ratio R NM (82) relative
to the constituent with the nearest lower frequency (NaN if no other constituent
has a lower frequency),
• coef.diagn.lo.CorMx, the maximum model parameter correlation corrmax (87)
relative to the constituent with the nearest lower frequency (NaN if no other
constituent has a lower frequency),
• coef.diagn.hi.name, coef.diagn.hi.RR, coef.diagn.hi.RNM, and
coef.diagn.hi.CorMx, which are the same as the corresponding above four fields,
but for the nearest higher frequency,
and the following scalars:
• coef.diagn.K, the condition number of the basis function matrix (84),
• coef.diagn.SNRallc, the all-constituent signal to noise ratio (85),
• coef.diagn.TVraw, the tidal variance (99) of the raw inputs, with units
u_raw/v_raw units squared (two-dimensional case) or sl_raw units squared (one-
dimensional case),
• coef.diagn.TVallc, the tidal variance of the model solution (all constituents
superposed) (99), with units u_raw/v_raw units squared (two-dimensional case)
or sl_raw units squared (one-dimensional case),

43
 • coef.diagn.TVsnrc, the tidal variance of the reconstructed fit using only
constituents that meet the SNR criterion (102),
• coef.diagn.PTVallc, the percent tidal variance captured by the (all-constituent)
model solution (99), and
• coef.diagn.PTVsnrc, the percent tidal variance of the reconstructed fit using only
constituents that meet the SNR criterion (102).
The main diagnostic results, in addition to appearing in the above fields, are summarized
in the constituent selection diagnostics table coef.diagn.table. This is a table (described in
detail in Section III.1.C.d) that is formatted for easy viewing within Matlab to aid in the
constituent selection process. Unlike the above nallc x1 fields of coef.diagn, for which the
ordering can be changed using the ‘OrderCnstit’ option flag, the rows in coef.diagn.table
are always ordered by decreasing PEq .

Runtime display. All results of a call to ut_solv() are stored in coef. For convenience, by
default when its execution is complete ut_solv() outputs a three-part runtime display (a
key subset of the contents of coef). The runtime display consists of (a) the coefficients
and confidence intervals, coef.results; (b) the run description meta-information
coef.aux.rundescr; and (c) the constituent selection diagnostics table coef.diagn.table, if
it has been generated (unless the ‘NoDiagn’ option has been selected). The runtime
display can be omitted entirely, or can include subsets of the three components, as
controlled by the opt.RunTimeDisp option to ut_solv() as described below.

III.C.1.c. Defaults and options

The default configuration for a call to the function ut_solv(), which is
implemented when no {options} parameters are passed in, is as follows.
• The linear (secular, non-tidal) trend (2) is included in the model .
• No pre-filtering correction is made ( Pq = 1 , in (5)).
• Nodal/satellite corrections with exact times (4) are implemented.
• Greenwich phase lags are computed by use of the astronomical argument with
exact times (4).
• The model includes no inferred constituents.
• If cnstit is ‘auto’ then the automated decision tree constituent selection method is
applied with Rmin =1 (81).
• The solution method is robust IRLS with the Cauchy weight function and tuning
parameter 2.385, which is the Matlab default value (TunRdn = 1).
• The Monte Carlo uncertainty propagation method (Section II.C) is used, with 200
realizations (Nrlzn=200), to determine confidence intervals of the current ellipse
parameters from those of the model parameters.
• Confidence intervals are computed based on the (colored) spectra computed from
the actual residuals (Section II.C.3). If the input times are uniformly distributed,
the spectra are computed using FFT methods, otherwise they are computed using
the Lomb-Scargle periodogram with frequency oversampling factor 1
(LSFrqOSmp=1).

44
 • Computation of the constituent selection diagnostics table (Section III.C.1.d) is
carried out, using SNRmin = 2 .
• The order of constituents in the output variables and diagnostics table is based on
decreasing percent energy PEq (97).
• All three components of the runtime display are presented.

To change these defaults, the following option flags can be passed in. Option
flags must be passed in after the cnstit argument. The option flags are not case-sensitive,
but they cannot be abbreviated. For some flags, accompanying variables are passed in, as
noted in boldface italics. For those options appearing in a list headed by “One of the
following”, an error will result if more than one on the list is specified.
• ‘NoTrend’
o This will omit the linear/secular trend term from the model.
• ‘PreFilt’, PreFilt
o This will implement the correction to account for pre-filtering that was
applied to the raw inputs before the analysis; PreFilt is a structure that
specifies the pre-filter transfer function (not the inverse transfer function,
as is input to t_tide, PBL02) as
ƒ PreFilt.P, an n frq x1 vector with real-valued P (4) for the one
dimensional case, and complex P + iP for the two-dimensional
case, where n frq is an arbitrary number of frequencies by which
the filter shape is to be specified
ƒ PreFilt.frq, an n frq x1 vector of the frequencies (in cycles per hour)
of the PreFilt.P values, and
ƒ PreFilt.rng, a two-value vector with the range (minimum and
maximum) of acceptable P magnitudes, for example [0.01 100];
values outside this range will be set to 1.
• One of the following:
o ‘NodsatLinT’
ƒ This will cause nodal/satellite corrections to use linearized times,
instead of the default exact formulation.
o ‘NodsatNone’
ƒ This will cause nodal/satellite corrections to be omitted, instead of
the default exact formulation.
• One of the following:
o ‘GwchLinT’
ƒ This will cause the astronomical argument in the Greenwich phase
lag calculation to use linearized times, instead of the default exact
formulation.
o ‘GwchNone’
ƒ This will omit the astronomical argument, such that the reported
phase lags are “raw” (not Greenwich-referenced) relative to the
reference time tref .
• ‘Infer’, Infer

45
 o This causes a total of nI inference constituents, and nR reference
constituents ( 1 ≤ nR ≤ nI ), to be included in the model (11). If any of the
specified reference or inference constituents are in the group of non-
reference constituents, as determined by the automatic decision tree or as
specified manually by the cnstit input, they are removed from the group of
non-reference constituents. Infer is a structure with elements
ƒ Infer.infnam, a cell-array of nI 4-character names of the
constituents to be inferred,
ƒ Infer.refnam, a cell-array of nI 4-character names of the
corresponding reference constituents, not all of which need be
unique from each other (unless ‘InferAprx’ is chosen, see below)
because multiple constituents can be inferred from a single
reference constituent,
ƒ Infer.amprat, which is
• for the two-dimensional case, a 2nI x1 vector of real-valued
unitless amplitude ratios, r + , r − (12) , with r + values in the
first nI elements and r − values in the second nI elements
• for the one-dimensional case, an nI x1 vector of real-valued
unitless amplitude ratios rη (21),
ƒ Infer.phsoff, which is
• for the two-dimensional case, a 2nI x1 array of real-valued
phase offsets (in degrees), ς + , ς − (12), with ς + values in
the first nI elements and ς − values in the second nI
elements
• for the one-dimensional case, an nI x1 vector of real-valued
phase offsets ς η (21).
• ‘InferAprx’
o This causes the inference calculation to follow the approximate method
(Section II.A.4.c). Ignored unless ‘Infer’ also selected. With this option
selected, an error will result if the constituents in infer.refnam are not
unique, i.e. when nR < nI (as when inferring multiple constituents from a
single reference constituent, not possible for the approximate method).
• ‘Rmin’, Rmin
o This will specify the Rmin (81) value (positive) to be used in automated
constituent selection. The default is Rmin =1. Ignored if cnstit is not ‘auto’.
• One of the following:
o ‘OLS’
ƒ This will change the solution method to the Matlab “backslash”
operator, to implement ordinary least squares analysis instead of
the default IRLS.
o ‘Andrews’, ‘Bisquare’, ‘Fair’, ‘Huber’, ‘Logistic’, ‘Talwar’, OR ‘Welsch’

46
 ƒ This will cause the robust IRLS method to be implemented by the
robustfit() function using the named weight function (instead of
the default Cauchy weight function), and using a tuning parameter
that is the Matlab default value for that weight function divided by
the tuning factor reduction parameter TunRdn.
• ‘TunRdn’, TunRdn
o To reduce the IRLS tuning parameter relative to the Matlab default value,
for the given weight function (default Cauchy; otherwise specified by an
option input), by the tuning parameter reduction factor TunRdn (the
tuning parameter used is the default tuning parameter divided by
TunRdn.) The default is TunRdn = 1. Ignored if using ‘OLS’.
• ‘LinCI’
o This causes the confidence intervals on the current ellipse parameters to be
computed from the uncertainties in the model parameters by the linearized
method, instead of the Monte Carlo method.
• ‘White’
o This causes the white noise floor assumption to be implemented in the
confidence interval calculation such that spectra of the residual are
presumed white instead of calculated from the actual colored residual.
• ‘Nrlzn’, Nrlzn
o This will cause the Monte Carlo calculations to use Nrlzn realizations
instead of the default, which is 200. Ignored if ‘LinCI’ flag is passed in.
• ‘LSFrqOSmp’, LSFrqOSmp
o This will cause the Lomb-Scargle periodogram calculation to use
frequency oversampling factor of LSFrqOSmp instead of the default,
which is 1. If LSFrqOSmp is not an integer it is rounded. Ignored if raw
input times are uniformly distributed or if the ‘White’ flag is passed in.
• ‘DiagnMinSNR’, MinSNR
o This will specify the SNRmin value used in (i) calculation of TVsnrc and
PTVsnrc , and (ii) in the constituents included in the reconstructed fits of
the diagnostic figures. Default value is 2.
• One of the following:
o ‘NoDiagn’
ƒ Skip both the summary diagnostics table and the diagnostic
figures.
o ‘DiagnPlots’
ƒ Generate the diagnostic figures (described in Section III.C.1.d) in
addition to the diagnostics table.
• ‘OrderCnstit’, CnstitSeq
o This will override the default PEq -ranked ordering by which the
constituent-based parameters in the coef output structure (explained
below, e.g., coef.name, coef.g, etc; not the diagnostics in coef.diagn) are
listed. The parameter CnstitSeq is one of:
ƒ ‘snr’, to order by decreasing SNRq , OR
ƒ ‘frq’, to order by increasing frequency, OR

47
 ƒ(allowed only in the case that cnstit is not ‘auto’) a cell array of 4-
character strings that, if it differs from cnstit, only differs in the
order of its rows.
o The row order of the fields in coef.diagn, and the rows in the summary
diagnostics table, are always by decreasing PEq and are not affected by
the ‘OrderCnstit’ option.
• ‘RunTimeDisp’, RunTimeDisp
o To suppress all three components of the default runtime display
information (described at the end of section III.C.1.b.) use RunTimeDisp
= ‘nnn’ (not case sensitive). The default is ‘yyy’. To suppress one or two
of the components, replace y by n in the three-letter string; for example to
show only the coefficients and confidence intervals (first component) use
‘ynn’, to show only the constituent selection diagnostic table (third
component) use ‘nny’, etc. If the ‘NoDiagn’ option is selected then no
constituent selection diagnostic table will be computed nor shown at
runtime, regardless of the third character in RunTimeDisp.

III.C.1.d. Summary diagnostics table and diagnostic plots

Based on explanation of diagnostics in Section II.D above, by default UTide
generates a summary table of diagnostics (coef.diagn.table) that provides information
useful in the constituent selection process and is a character array that can be viewed
easily within Matlab after execution of ut_solv(). The table is computed for uniformly or
irregularly distributed times, as explained above. Its computation can be skipped using
the ‘NoDiagn’ option.

The heading lines of the table show quantities not specific to individual
constituents or constituent pairs. The first heading line shows the user-specified Rmin (81)
and SNRmin (96) values. The second heading line shows the basis matrix condition
number K (84) and the all-constituent SNRallc (85). The third heading line shows the tidal
variances (a) TVallc (99) of the model solution, (b) TVsnrc (102) of the reconstructed fit
using constituents meeting the SNR criterion (96), and (c) TVraw (99) of the raw input.
The fourth and final heading line shows the percent tidal variances (a) PTVallc (99) of the
all-constituent model solution, and (b) PTVsnrc (102) of the reconstructed fit using
constituents meeting the SNR criterion.

Within the table there is one row for each constituent, of frequency ωq , and the
rows are ordered by decreasing percent energy PEq (97) values. The first three columns
of the table are the name of the constituent, the PEq value, and the constituent-specific
SNRq (83); at the far left an asterisk appears adjacent to the constituent name if it meets
the SNR criterion (96). Next there is a group of columns presenting diagnostics related to
the constituent with the next-lower frequency compared to ωq ; finally, there is a group of
columns with the same diagnostics related to the constituent with the next-higher

48
frequency compared to ωq . In these latter two groups, the columns include the name of
the neighboring constituent followed by R R (81), R NM (82), and corrmax (87), each
computed for the respective constituent pair.

All types of constituents—non-reference, reference, and inferred—are listed

together in the table. Diagnostics based on constituent pairs ( R R , R NM , and corrmax ) are
computed only between non-reference and reference constituents, with pairs chosen
based on frequencies that are nearest to each other, regardless of whether an inferred
constituent has frequency between them. If the solution included inference of one or more
constituents, a list of them is shown at the bottom of the table with their respective
reference constituents.

The layout of the table columns is such that the importance of the diagnostics
within the columns is generally highest in the columns toward the left. That is, the PEq
and SNRq diagnostics are likely to be of the most use, with the R R , R NM , and
corrmax values also providing relevant information, but each of increasingly lower
importance for most typical situations.

The PEq rank-order of the constituents in the table makes it visually apparent
which among them have captured the most energy. The SNR values of the higher-ranked
constituents typically decrease in a similar manner to PEq , but for the lower- PEq
constituents it is useful to inspect the SNR values carefully, and if they are too low,
consider omitting or inferring the associated constituents.

The table is designed so that it is also easy to scan for and identify any R R and
NM
R values that are lower than 1. Such values indicate violations of the Rayleigh criteria
and suggest that consideration should be given to removing these constituents from the
model or inferring them. When the decision tree method of F77 has been used, the R R
values will all be greater than 1, but the R NM values will provide useful additional
information. Similarly, the corrmax values can be scanned easily for the relatively higher
values, which will help identify potential pairs of constituents that may not be sufficiently
independent from each other to both be included in the model unless one is inferred.

Through use of the ‘DiagnPlots’ option, two diagnostic figures (in addition to the
diagnostics table) can be generated by ut_solv(). Each figure is generated with a call to
the built-in function figure() without application of any rescaling, so the user will need to
manually maximize them (e.g. by mouse) to make the plots more legible on the screen.
The two figure windows will typically overlie each other when initially created.

The first figure has four frames in the two-dimensional case and three frames in
the one-dimensional case. The top frame shows the text field coef.diagn.rundescr, to
provide a descriptive summary of the run characteristics. The second and third frames, in
the two-dimensional case, show time series for the u and v components respectively: the

49
raw input, the reconstructed fit using constituents that meet the SNR criterion, and the
residual; in the one-dimensional case there is one such frame. The bottom frame is a
semi-logarithmic plot that shows, in the two-dimensional case, a vertical bar extending to
Lsmaj + Lsmiη for each constituent in order of increasing frequency, colored red if the SNR
2 2

criterion is met and blue if not, along with a green dotted line showing
(σ L2smaj + σ L2smiη ) ⋅ SNRmin to indicate the height required for each bar to meet the SNR
criterion. In the one-dimensional case the vertical bar heights are A2 and the green dotted
line is σ A2 ⋅ SNRmin .

The second figure has four frames in the two-dimensional case and two frames in
the one-dimensional case. It shows information about only the constituents that meet the
SNR criterion, and they are ordered by decreasing PEq from left to right in each of the
frames. In the two-dimensional case the four frames show the current ellipse parameter
values, together with their 95% confidence intervals. In the one-dimensional case, the
two frames show the amplitude and phase, together with their 95% confidence intervals.

III.C.2 Reconstructing fits with ut_reconstr()

The ut_reconstr() function has two main purposes. The first purpose is to enable
calculation of reconstructed fits at an arbitrary set of times. The second purpose is to
enable calculation of reconstructed fits that include a user-specified subset of constituents
(see (100)), for example as identified based on other criteria in addition to, or in place of,
the SNR threshold.

III.C.2.a. Input and output parameter descriptions

The ut_reconstr() input t_fit is a column vector of arbitrary times that
• contains Matlab datenum values, with units of days (as for t_raw, see above
description);
• can be either equispaced or irregular; and
• can include NaNs and if so the outputs (u_fit/v_fit or sl_fit) will have
corresponding NaNs.
The reconstructed fit (u_fit/v_fit or sl_fit) are column vectors that have
• the same size as t_fit; and
• the same units as u_raw/v_raw or sl_raw.

III.C.2.b. Defaults and options

In a call to the function ut_reconstr() (see opening portion of Section III.C for
syntax) the default implementation, which is executed when no {options} parameters are
passed in, includes in the reconstruction of the fit only the constituents for which SNR is
greater than 2. This default behavior can be changed by use of the following option flag
choices (which, as for the options to ut_solv(), are case-insensitive but cannot be
abbreviated or truncated):
• ‘MinSNR’, MinSNR
o This causes only those constituents with SNRq ≥ MinSNR (a real scalar)
to be used in the reconstruction. The default value is 2.

50
 • ‘MinPE’, MinPE
o This causes only those constituents with percent energy PEq ≥ MinPE (a
real scalar) to be used in the reconstruction. The default value is zero.
If both of ‘MinSNR’ and ‘MinPE’ are selected then no constituent with either
SNR or PE values lower than their respective specified thresholds will be
included in the reconstruction. Constituents will be included in the reconstruction,
or removed from it, by the MinSNR and/or MinPE criteria regardless of whether
they were non-reference, reference, or inferred constituents in the solution.
• ‘Cnstit’, Cnstit
o This causes only those constituents named in Cnstit, which must be
selected from those which were included in the model during the ut_solv
calculation that generated coef, to be used in the reconstruction. Cnstit is a
cell array of 4-character strings, of the same format as the cnstit input to
ut_solv() described above. Constituents that are listed in, or omitted from,
Cnstit are included in the reconstruction, or not, regardless of whether
they were non-reference, reference, or inferred constituents in the solution.
They are also included regardless of their SNR and PE values; if ‘Cnstit’
is used then MinSNR and MinPE are ignored.

All other attributes of the reconstructed fit computed by ut_reconstr() are

determined based on their configuration during the call to ut_solv() that created the coef
input to ut_reconstr(). This includes, for example, whether the trend is included, whether
nodal/satellite corrections use exact or linearized times, etc; this information is stored in
coef.aux.opt and summarized in coef.aux.rundescr. To compute a reconstruction with
any of these attributes changed, an additional run of ut_solv() must be made, and the
resulting coef passed to ut_reconstr().

III.D. Functionality and syntax for groups of records

A group of multiple time sequences can be analyzed with a single execution of
ut_solv() and, if needed, a corresponding group of hindcast/forecasts can be calculated
with a single execution of ut_reconstr(). Each record in the group can have a different
number and distribution of time values, a different latitude, and different inference
constants. The associated modifications to the functionality and syntax of the inputs and
outputs are described in this section.

In an analysis of a group of records, the following should be borne in mind.

o The automated constituent selection option is not available, so the group of
constituents to be included must be manually specified. This should be
straightforward to overcome, by determining a suitable set of constituents based on
some preliminary runs with a few representative members of the group.
o Each record in the group must have the same number of times. This can be facilitated,
if necessary, by padding shorter records with NaNs. At the start and end of the
shorter records the padded values of t_raw must be NaNs (rather than non-NaN
time values); this is to ensure that the corresponding u_raw/v_raw or sl_raw values
are not filled, in the equispaced times case.

51
o Diagnostics can be computed for each member in the group but the diagnostic figures
cannot be generated; the ‘DiagnPlots’ option is not allowed.
o No runtime display is generated; ‘RunTimeDisp’ other than ‘nnn’ is not allowed;
however, all potential runtime display information (coef.results, coef.aux.rundescr,
and coef.diagn.table) for each record in the group is included in the output.
o The order of the elements in the output fields is by increasing frequency and the
‘OrderCnstit’ option is not allowed. This due to the fact that frequencies are the
only ordering (unlike ordering by SNR or PE value) that is certain to be uniform
across all records in the group, which is required by the form of the output fields as
explained below.
o Any other options, if passed in, are applied to all members of the group: ‘NoTrend’,
‘PreFilt’, ‘NodSatLint’, ‘NodSatNone’, ‘GwchLint’, ‘GwchNone’, ‘Method’,
‘TunRdn’, ‘LinCI’, ‘White’, ‘Nrlzn’, ‘LSFrqOSmp’, ‘NoDiagn’, and
‘DiagnMinSNR’.

The group of ns time sequences to be analyzed is indexed as an nd -dimensional

array of size n1 × n2 × n3 K× nnd , where each n value gives the size of that dimension of
the array, and ns = n1n2 n3 L nnd . For example, if the group consists of numerical
simulation time series of sea level from a 20x10 array of lat-lon gridpoints, there are
ns = 200 time sequences, and a valid choice would be n1 = 20 and n2 = 10 , for nnd =2.
Alternatively, they could be treated using n1 = 200 and nnd =1. As another example, if
the group consists of a current observations from bottom-mounted acoustic Doppler
current profilers (ADCPs) deployed along two across-shelf lines, each line having 5
ADCPs and each ADCP collecting current measurements from 50 depth bins, one
configuration to treat the ns = 500 time sequences would be n1 = 2 , n2 = 5 and n3 = 50 ,
for nnd =3.

The inputs to ut_solv() are just as in the single-record case described in the
previous section, except for the following changes.
• t_raw can be either
o a single nt × 1 vector of times that applies to all time sequences in the
group, in which is case it is specified exactly as in the single-record case
described above, or
o an nt × n1 × n2 × n3 K × nnd array of times, as necessary when more than one
record in the group has a different set of times; in this case the number of
times must be the same ( nt ) for each record, necessitating that if there are
records with fewer times their time vectors must be padded with NaNs and
their corresponding u_raw/v_raw or sl_raw must be padded with NaNs.
Note also that none of n1 , n2 ,..., nn d can be 1.
• u_raw and v_raw, or sl_raw, are each nt × n1 × n2 × n3 K × nnd arrays (again, none
of n1 , n2 ,..., nn d can be 1).

52
 • lat is either a scalar, in which case the analysis of all the records will use the same
value, or a n1 × n2 × n3 K × nnd array.
• cnstit cannot be ‘auto’, but rather must be a specific list of constituents to include;
this is required in order that the same group of constituents is included for each
individual analysis, which enables convenient grouping of the results fields in
coef (explained below).
• If constituents are to be inferred, then the same inference and reference
constituents will be used (same Infer.infnam and Infer.refnam) in each
individual analysis, and either
o Infer.amprat and Infer.phsoff are the same size as each other and are
2nI x1 (two-dimensional case) or nI x1 (one-dimensional case) as
explained above for treatment of a single record, in which case the same
inference constants will be applied to every record, OR
o Infer.amprat and Infer.phsoff are the same size as each other and are
2nI × n1 × n2 × n3 K × nnd (two-dimensional case) or nI × n1 × n2 × n3 K× nnd
(one-dimensional case), such that different inference constants can be
applied to each record.
• Option flags (other than ‘infer’ as just noted) will be applied identically to each
individual time sequence analysis, with the exception of ‘OrderCnstit’ and
‘DiagnPlots’, which will be ignored; ordering of constituent-indexed outputs is
always by increasing frequency.

The output coef from ut_solv() is as in the single-record case described above except
for the following changes.
• Fields of size nallc x1 in the single-record case have size nallc × n1 × n2 × n3 K × nnd
in the case of a group analysis.
• Fields of size 1x1 in the single-record case have size n1 × n2 × n3 K × nnd in the case
of a group analysis.
• The ordering of the elements in each of these fields is by increasing constituent
frequency, as explained above.
The exceptions are that, to avoid redundancy in fields of the output coef,
• coef.name, coef.aux.frq, and coef.aux.lind are each nallc x1 (as determined by the
fixed set of constituents in cnstit, which are included identically in the analysis of
each record in the group),
• the only fields of coef.aux.opt that potentially differ from the single-record case
are equi, infer.amprat and infer.phsoff, and the latter two are the same size as
their corresponding inputs,
• if a single lat value was passed in then coef.aux.lat is a scalar.

The inputs to ut_reconstr() in the group case are as for the single-record case
described above, with the following exception.
• t_fit can be either

53
 o a single nt × 1 vector of times to be used for all time sequences in the
group, in which is case it is specified exactly as in the single-record case
described above, or
o an nt × n1 × n2 × n3 K × nnd array of times, in which case each record is
computed at its own set of times, though the number of times must be the
same ( nt ) for each record.
The output from ut_reconstr() in the group case is the same as for the single-record case
except that for u_fit and v_fit, or sl_fit, each is an nt × n1 × n2 × n3 K × nnd array instead of
an nt x1 column vector.

III.E. Relationships to existing software

The UTide code incorporates features of (a) the t_tide Matlab functions (PBL02),
including use of certain important t_tide components unmodified (the database of
harmonic constants, the constituent selection decision tree code, the linearized confidence
interval calculations, the band-averaging of residual spectra); (b) the r_t_tide Matlab
functions (LJ09); and (c) the “versatile tidal analysis” Fortran program (FCB09).

UTide includes all functionality of t_tide except for its XTide capabilities.
However, exact agreement with t_tide in the case of a record with an even number of
points cannot be achieved by UTide because the way t_tide drops the last point is not
compatible in general with the capability of UTide to accept irregular times. In addition,
the confidence interval calculations of UTide cannot exactly recover those of t_tide. In
t_tide the scaling of spectral quantities for the colored case included an extra factor of
two, and simplifying assumptions were made (as explained in detail above) about the
covariance matrix (59), that affect both the white and colored case, but are not made in
UTide. An additional, though minor, contribution to the differences in Monte Carlo cases
is due to the stochastic nature of the calculation, which causes each run (of either UTide
or t_tide) to yield slightly different results. As a result of these relationships, in order to
achieve the closest agreement of UTide results with those of t_tide, for testing purposes,
there are a number of requirements. First, UTide must be executed with an odd number of
points. Second, if t_tide is called with both the start_time and lat inputs, the equivalent
call to UTide requires the ‘NodsatLinT’ and ‘GwchLinT’ options; if t_tide is called with
only the start_time input, the equivalent call to UTide requires the ‘NodsatNone’ and
‘GwchLinT’ options; if t_tide is called without the start_time input input, the equivalent
call to UTide requires the ‘NodsatNone’ and ‘GwchNone’ options; and if inference
calculations are done by t_tide the equivalent call to UTide requires the ‘InferAprx’
option. Third, the ‘OLS’ option to UTide must be used because t_tide does not implement
the IRLS method that is the default for UTide. Finally, it should be noted that calling
UTide with the ‘OrderCnstit’,‘frq’ option will make comparisons to t_tide output more
convenient.

The IRLS features of r_t_tide are included in UTide, including their application to
two-dimensional raw input and irregularly distributed times. Because r_t_tide is a
modification of t_tide to include the IRLS solution method, the above noted relationships

54
between UTide and t_tide generally apply when comparisons between UTide and r_t_tide
outputs are made.

All features of the FCB09 Fortran code for the one-dimensional case are available
in UTide (in addition, UTide includes their generalization to the two-dimensional case),
with the exceptions that (a) UTide does not include the same covariance-based
constituent selection diagnostics developed from the FCB09 singular value
decomposition, and (b) UTide confidence intervals are based on the new formulation and
reported for the current ellipse parameters, while those of the FCB09 code are for the
cosine/sine coefficients and based on the white noise presumption. The only change to
the default configuration of UTide, in order to create results with the closest agreement to
those of the “versatile tidal analysis” program, is to use the option ‘OLS’.

III.F. Computational demands

The computational demands of UTide are significantly higher than some previous
tidal analysis software, for the following reasons: (i) the IRLS solution method is used,
which involves multiple iterations of solutions each with demand similar to an OLS
solution; (ii) treating nodal/satellite corrections and Greenwich phase calculations using
the exact times substantially increases the memory requirements and the number of
computations, (iii) the complex-valued formulation of the matrix system is solved, for
reasons explained above, which can be less efficient than solving comparable real-valued
formulations; (iv) in the case of irregular times, the Lomb-Scargle periodogram
calculations are slower than their FFT counterparts for uniformly distributed times; (v)
the generality of the new confidence interval calculation is slightly more costly than
earlier versions; and (vi) the constituent selection diagnostics require additional
computation. The relative importance of each of these factors in contributing to the
increased computation demand will of course differ depending on the particular analysis
at hand (number of raw input times, whether they are irregular, number of constituents,
one-dimensional or two-dimensional raw input, whether Monte Carlo is used for
confidence intervals, etc). However, the above list of reasons for increased burden is
roughly in order of decreasing importance, very generally.

Crude guidelines for the computational burden result from summarizing the
results of numerous analyses (each an execution of ut_solv() then ut_reconstr() in
sequence; the large majority of the time is spent on the former) of various test datasets of
hourly sea level (Newport, RI) and currents (Martha’s Vineyard Coastal Observatory).
The datasets were sampled uniformly or irregularly for durations between a month and 5
years. Using a modest-capability 2007-era laptop PC with Matlab 2010a, when
configured to mimic t_tide computations (i.e. ‘OLS’, ‘NodsatLinT’, ‘GwchLinT’) UTide
used a comparable amount of memory, and was approximately twice as slow, compared
to t_tide. This meant run times from between about 1-2 seconds and 20-30 seconds, for
the one month and 5 year records respectively. When the exact formulations for
nodal/satellite and Greenwich phase lag calculations were implemented, the run times
increased by about 2-5 times, but remained comparable in speed or faster than the Fortran
code of FCB09, except for records of a year or more. Such records become significantly
slower and, notably when the Lomb-Scargle periodogram is calculated (colored case with

55
irregular times), very memory-intensive. When the IRLS solution method was also
implemented, with the Cauchy weight function and default tuning parameter, the run
times increased by an additional factor of 2 for the shorter records and by a higher
amount, more than an order of magnitude in some cases, for the longer records.

These results just described are of course only applicable for the particular test
records and computational system used; in general, results will vary depending on signal
to noise characteristics of the raw inputs, the nature of the sampling and duration of the
records, and the configuration of the runtime options, as well as the computing resources.
Nonetheless, it is clear that relative to other tidal analysis software, the additional burden
of UTide is at most a few orders of magnitude higher, for records up to 5 years long, and
in many cases a smaller increase. Considering that availability of computing power for a
typical researcher goes well beyond a typical 2007 laptop, in most applications these
costs seem modest enough not to be a major constraint.

In this context, UTide has been developed based on the view that the additional
features it incorporates are sufficiently valuable that they offset the undesirable increase
in computational demands. However, in order to lessen the computational burden, to the
extent it is possible, certain aspects can be omitted from a calculation. For example, the
‘OLS’ option will forego the cost of the IRLS computations, the ‘White’ option will
eliminate the spectral calculations (notably the Lomb-Scargle algorithm for irregular
times), the ‘LinCI’ option obviates the Monte Carlo random realizations, the ‘NoDiagn’
option omits computation of constituent selection diagnostics, and a smaller ‘Nrlzn’ value
will require less cycles when using Monte Carlo. If computational demands are a
constraint, then a good strategy will be to do initial calculations using these options (e.g.,
‘OLS’, ‘White’, ‘LinCI’, possibly ‘NoDiagn’) and then later carry out a select few runs
that involve the slower features as needed.

There are aspects of UTide for which future modifications could potentially
increase its computational efficiency substantially. Examples would be to (a) create a
precompiled executable in a different language, (b) implement more efficient code,
including the matrix solution formulations, and/or (c) use a ‘fast’ Lomb-Scargle
algorithm, such as that of Press and Rybicki (1989), which will reduce processing time
although, as explained above, the major demand of Lomb-Scargle calculations is on
memory. There are also potential modifications that would lessen its speed in return for
reduced memory demands, which is useful when memory limitations are more of a
constraint than processing speed; for example, an alternative loop arrangement for the
Lomb-Scargle code could reduce the memory demand with the trade-off of slower
runtimes. Pursuit of such improvements can follow on an as-needed basis.

56
IV. Acknowledgements
The effort would not have been possible without code made available by Mike
Foreman and colleagues, Rich Pawlowicz and colleagues, Keith Leffler, and David Jay,
all of whom were generous with their advice. Comments from Wendy Callendar on an
early version resulted in improvements and motivated the group analysis capability. Even
Haug (Hydrographic Service, Norwegian Mapping Authority) made several helpful
suggestions, including an improved handling of cross-spectral quantities in the colored
scaling calculation. This material is based upon work supported by the National Science
Foundation, Physical Oceanography, under Grant No. 0826243, “Investigating Tidal
Influences on Subtidal Estuary-Coast Exchange Using Observations and Numerical
Simulations”. Any opinions, findings, and conclusions or recommendations expressed in
this material are those of the author(s) and do not necessarily reflect the views of NSF.

57
V. References
Cherniawsky, J.Y., M.G.G. Foreman, W.R. Crawford, R.F. Henry. 2001. Ocean Tides
from TOPEX/Poseidon Sea Level Data. J. Phys. Ocean. 18, 649-664.
Codiga, D.L., L.V. Rear. 2004. Observed tidal currents outside Block Island Sound:
Offshore decay and effects of estuarine outflow. J. Geophys. Res. 109,
doi:10.1029/2003JC001804.
Codiga, D.L. 2007. FOSTER-LIS Gridded Data Products: Observed Current Profiles and
Near-Surface Water Properties from Ferry-based Oceanographic Sampling in
Eastern Long Island Sound. In, Graduate School of Oceanography, University of
Rhode Island, Narragansett, RI., pp. 14.
Emery, W.J., R.E. Thomson. 1998. Data Analysis Methods in Physical Oceanography.
Pergamon, New York, 634pp.
Foreman, M.G.G. 1977. Manual for tidal heights analysis and prediction. . Pacific
Marine Science Rep. 77-10, Institute of Ocean Sciences, Patricia Bay, 101 pp.
[Revised 2004; Available online at http://www.pac.dfo-
mpo.gc.ca/SCI/osap/publ/online/heights.pdf].
Foreman, M.G.G. 1978. Manual for tidal currents analysis and prediction. . Pacific
Marine Science Rep. 78-6, Institute of Ocean Sciences, Patricia Bay, 70 pp.
[Revised 2004; Available online at http://www.pac.dfo-
mpo.gc.ca/SCI/osap/publ/online/heights.pdf].
Foreman, M.G.G., R.F. Henry. 1989. The harmonic analysis of tidal model time series.
Adv. Wat. Res. 12, 109-120.
Foreman, M.G.G., E.M. Neufeld. 1991. Harmonic tidal analyses of long time series. Int.
Hydrogr. Rev. LXVIII, 85–108.
Foreman, M.G.G., J.Y. Cherniawsky, V.A. Ballantyne. 2009. Versatile Harmonic Tidal
Analysis: Improvements and Applications. J. Atmos. Oceanic Tech. 26, 806-817.
DOI: 810.1175/2008JTECHO1615.1171.
Godin, G. 1972. The analysis of tides. University of Toronto Press, Toronto.
Goodman, N.R. 1963. Statistical analysis based on a certain multivariate complex
Gaussian distribution (an introduction). Ann. Math. Statist. 34, 152–177.
Higham, N.J. 2002. Computing the nearest correlation matrix—A problem from finance.
IMA J. Numer. Anal. 22, 329–343.
Leffler, K.E., D.A. Jay. 2009. Enhancing tidal harmonic analysis: Robust (hybrid L-1/L-
2) solutions Cont. Shelf Res. 29, 78-88. DOI: 10.1016/j.csr.2008.1004.1011
Lomb, N.R. 1976. Least-squares frequency analysis of unequally spaced data. Astrophys.
Space Sci. 39, 447-462.
Munk, W., K. Hasselmann. 1964. Super-resolution of tides. Studies on Oceanography - A
Collection of Papers Dedicated to Koji Hidaka K. Yoshida, Ed., University of
Tokyo, 339-344.
Parker, B.B. 2007. Tidal analysis and prediction. NOAA Special Publication NOS CO-
OPS 3, U.S. Department of Commerce, 378 pp.
Pawlowicz, R., B. Beardsley, S. Lentz. 2002. Classical tidal harmonic analysis including
error estimates in MATLAB using T-TIDE. Computers & Geosciences 28, 929-
937.
Press, W.H., G.B. Rybicki. 1989. Fast algorithm for spectral analysis of unevenly
sampled data. Astrophysical Journal 338, 277-280.

58
Press, W.H., S.A. Teukolsky, W.T. Vetterling, B.P. Flannery. 1992. Numerical Recipes
in FORTRAN: The Art of Scientific Computing. Cambridge University Press,
Cambridge, U. K.
Scargle, J.D. 1982. Studies in astronomical time series analysis. II - Statistical aspects of
spectral analysis of unevenly spaced data. Astrophys. J. 263, 835-853.
Schulz, M., K. Stattegger. 1997. SPECTRUM: Spectral analysis of unevenly spaced
paleoclimatic time series. Comp. Geosci. 23(9), 929-945.

59