Title: Total Least Squares and ErrorsinVariables Modeling : Problem formulation, Algorithms, and Applicatio
1Total Least Squares and Errors-in-Variables
Modeling Problem formulation, Algorithms, and
Applications
- By Sabine Van Huffel
- Katholieke Universiteit Leuven
- Dept Electrical Engineering, Division
ESAT/SCD(SISTA) - Leuven, Belgium
2Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
3Introduction
Linear parameter estimation problem
?1?1 ? 2?2 ? p?p ?
?j parameters, ?j,? variables
Take n measurements xij, yj , j1,,n(ngtp)
solve for ? ?1,, ?pT
Xnxp ? ? ynx1
Classical approach
Linear Least Squares
(X exact)
Unrealistic if errors occur in both X and ye.g.
in automatic control, signal processing, system
theory, statistics, biology, chemistry,
economics, physics, ... Requires a more general
approach Total Least Squares (TLS)
4The geometry of least squares inverse least
squares total least squares
5Introduction (contd.)
Consider the simplest linear model between 2
variables ? and ? with underlying
relationship
The real measurements x and y may have errors ?
and ?, i.e., x ? ? and y ? ?
(errors-in-variables model)
A wide variety of situations exist, depending on
the error structures in x and y. We will examine
a number of these.
6Introduction (contd.)
Least Squares vs.
Total Least Squares
Orthogonal Projection Sensitive to Scale
Change Invariant to Exchange of Axes ML solution
whensx sy const.
Vertical Projection Invariant to Scale
Change Sensitive to Exchange of Axes ML solution
whensx 0, sy const.
7Problem Formulation
8Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
9Univariate EIV regression a statistical approach
- Model formulation fit n measurements (xi,yi),
i1,,n to a straight line - Standard regression xi error-free, yi with error
- no error
- Estimate ?0, ?1 using Least Squares (LS)
- EIV regression xi, yi with error
10Univariate EIV regression a statistical approach
(contd.)
- Three EIV models functional
fixed values - structural
random E( )?, var( )?2 - ultrastructural
random E( )?i, var( )?2 - Rewrite EIV regression as
- EIV regression ? standard regression xi random
and cov(xi,?i)-?1??2?0 - Standard regression estimator not consistent
- EIV model not identifiable unless side conditions
are imposed - ratio of error variances ???2 /??2 known
- ??2 known
- ??2 known
- both ??2 and ??2 known
- Assumption 1 most popular (Adcock, 1877)
Orthogonal Regression (OR) if ?1
11Univariate EIV regression a statistical approach
(contd.)
- EIV model parameter estimation
- Orthogonal regression (Adcock, 1877) maximizing
likelihood function (?1) reduces to minimizing
the sum of squared orthogonal distances from data
points to regression line - Weighted LS (Lindley, 1947) or generalized LS
(Sprent, 1966) weight residuals with reciprocal
of their variance and minimize sum -
- only holds for no-equation-error model
- Modified LS (Cheng, 1999) ?most general approach
minimize unbiased and consistent estimator of the
appropriate unknown error variance, e.g. estimate
??2 (? known)
12Univariate EIV regression a statistical approach
(contd.)
- Minimizing with respect to ?0, ?1 yields
13Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
14TLS and EIV regression a computational approach
- Statistical approach estimate EIV parameters ?0,
?1 by optimizing statistical properties (remove
bias, max. likelihood, consistency) - Applied mathematics approach (TLS) estimate ?0,
?1 by minimally correcting measured data xi,yi
such that the corrected data -
- exactly satisfy the imposed model, i.e.
- TLS for univariate EIV model OR
- For more than one explanatory variable (X
matrix) no analytical solution, use SVD or EVD
approach
15Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
16Historical remarks TLS (re) discovered many times
- Statistics, regression
- errors-in-variables regression (Gleser, 1981)
- measurement error models (Fuller, 1987)
- orthogonal (distance) regression (Pearson ,1902
Adcock,1877) York 1966) - generalized (weighted, modified) least squares
(Madansky, 1959 Lindley 1947 Sprent 1966, Cheng
van Ness 1999 ) - Numerical analysis Total Least Squares (Golub
1973 Golub Van Loan, 1980) - System Identification
- global linear least squares (Staar, 1981)
- eigenvector method (Levin, 1964)
- Koopmans-Levin method (Fernando Nicholson,
1985) - Compensated least squares (Stoica Söderström,
1982) - Experimental modal analysis Hv technique
(Leuridan et al., 1986) - Signal Processing minimum norm method (Kumaresan
Tufts, 1983) - Chemometrics Maximum Likelihood PCA (Wentzell et
al., 1997)
17Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
18The SVD revisited
- Given nxp complex matrix X
- with
- r rank number of independent columns/rows of
X - U1 basis of column space of X, U2?U1
- V1 basis of row space of X, V2?V1
19The SVD revisited (contd.) SVD example
forward map
Easy to compute Matlab, calculator
inverse map
20The SVD revisited (contd.) SVD properties
21The SVD revisited (contd.) Oriented energy
Consider a matrix A representing a set of p
vector samples from an n-dimensional signal space
Axes of symmetry singular vectors Length axes
squared singular values
22Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
23Recall Problem Formulation
24Basic TLS algorithm and computational issues
(contd.)
25Basic TLS algorithm and computational issues
(contd.)Extensions TLS algorithm
1. not unique
minimum norm solution 2.
non-generic case additional
constraint TLS solution
26Basic TLS algorithm and computational issues
(contd.) Extensions TLS algorithm
3. multivariate case
p d
software available (SLICOT, NETLIB)or
equivalently,
27Basic TLS algorithm and computational issues
(contd.)More efficient TLS algorithms
TLS solution entirely determined by - lowest
right singular vector of Xy (d1)- more
generally basis of lowest right singular
subspace of XY corresponding to
smallest singular values. ? compute ONLY those
vectors instead of whole SVD a. directly
compute SVD of XY partially ? partial TLS
(halves CPU time), rank-revealing URV/ULV
(Van Huffel, Stewart, Zha, Hansen,
Fierro,...) b. iteratively recommended for
slowly varying TLS problems. ? inverse iteration,
Rayleigh quotient iteration, conjugate gradient,
(inverse) Chebyshev iteration, (implicitly
restarted) Lanczos methods, neural computing, ...
(Van Huffel, Fierro, Hansen, Barlow,
Cirrincione, Björck, Kamm, Nagy,...)
28Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
29 TLS Properties
- Compare TLS directly with LS (solutions,
residuals, approximation efforts, subspaces)
- LS
- TLS
- TLS minimizes a sum of weighted squared residuals
30TLS Properties (contd.) Algebraic properties
and Sensitivity
- Differences increase when ?p (Xy)/ ?p (X)
grows - X? ? y less compatible
- vector y growing
- X (nearly) rank-deficient
- y oriented along the lowest singular vectors
of X - Accuracy TLS maximal when y parallel with
lowest singular vector up of X
31TLS Properties (contd.) Sensitivity
32TLS Properties (contd.) Consistency TLS
solution
Errors-in-variables model ? ? ? X ? ? y ?
? error matrix ?? Theorem (Gleser, 1981
Fuller, 1987) if ?i?i i.i.d. with
zero mean and E(?i?i ?i?i T) ?I,
(i.e. all errors equally sized
and uncorrelated) and
exists, positive definite then TLS
strongly consistent (LS inconsistent)
33TLS Properties (contd.) Consistency TLS
solution
34TLS Properties (contd.) extend consistency
TLS
- ? ? ?
- X ? ?
- y ? ?
- Assume errors ?i ?i are i.i.d. with zero
mean but correlated - Error covariance matrix E(?i ?i ?i ?i
T)?CCTC known, positive definite - Consistent solution ?
- 1. transform data XyXyC-1
- 2. compute TLS solution
3. transform back - implicit transformations via Generalized SVD /
QSVD of (Xy,C) - Solve GENERALIZED TLS problem ? strongly
consistent
(Van Huffel and Vandewalle, 1989 Gallo, 1981
Gleser, 1981 Fuller, 1987)
35Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- Case Study TLS in near-infrared spectroscopy
- Case Study TLS in magnetic resonance
spectroscopy - TLS extensions
- Structured TLS
- STLS applications
- Conclusions
36 Total Least Squares Applications
- CASE STUDIES IN SIGNAL PROCESSING SYSTEM
IDENTIFICATION - Subset selection in EIV models
- Blackbox modeling in signal processing and system
identification - Linear prediction problem
- Impulse response estimation by discrete
deconvolution - Parameter estimation in transfer function models
- TLS based estimation of frequency response
functions in Modal Analysis - STLS based linear prediction modeling in Speech
Compression - CASE STUDIES IN BIOMEDICAL SIGNAL PROCESSING
- Renography
- Magnetic Resonance Spectroscopy
- Near-Infrared Spectroscopy
- Polysomnography
- Fetal Electrocardiography
- OTHER TLS APPLICATIONS
37Some Biomedical Examples
- HOW? preprocessing --gt modeling --gt parameter
estimation --gt validation --gt GUI - TOPICS in collaboration with UZ Leuven
- Renogram deconvolution structured TLS,
displacement rank - Fetal ECG extraction generalized SVD,
Independent Component Analysis - MR Spectroscopic quantitation TLS, FIR filter,
nonlinear LS fitting - Near-Infrared Spectroscopy weighted TLS, PCA,
CCA, constrained LS - Polysomnography multichannel SVD based rank
reduction TLS LP modeling - Preoperative cancer diagnosis neural Bayesian
networks (MLP) - Osseoperception, brain tumor recognition, .
38NEAR-INFRARED SPECTROSCOPY quantitation of
neonatal oxygenation
- In collaboration with the Department of
Paediatrics and Neonatal Medicine, Univ.
Hospitals Leuven, Belgium - Co-workers G. Morren, G. Naulaers, P. Casaer,
H. Devlieger
39Near Infrared Spectroscopy (NIRS)
- Non-invasive measurement of cerebral oxygenation
of infants to study - the effects of apneas, drugs, temperature
changes, sucking, - the correlation with other physiological
parameters (heart rate, respiration, EMG )
40Beer-Lambert Law Model
d
41NIRS algorithm
42Algorithm refinements
43Total least squares
44 MAGNETIC RESONANCE SPECTROSCOPY in-vivo
quantitation of metabolite concentrations and
images
- In collaboration with the biomedical NMR unit,
Department of radiology, Univ. Hospitals Leuven,
Belgium and the Spin Imaging Group, Fac. Of Appl.
Physics, Delft Univ. of Technology, The
Netherlands - Co-workers L. Vanhamme, R. in t Zandt, P. Van
Hecke, Y. Wang, D. van Ormondt, N. Mastronardi,
T. Sundin, H. Chen, D. Graveron, I. Dologlou,
E. Heyvaert, C. Decanniere
45Nuclear magnetic resonance
static magnetic field
46Different applications of NMR
47MR Spectroscopic Imaging
Survey Image and 1H Spectroscopic Images of a
Patient Suffering from a Glioblastoma
48Model function
Theoretical Model
49The real world ...
- accuracy
- automation
- efficiency
50Proton MRS
51Preprocessing Data Quantification
HSVD/HTLS
subspace- based approach
? minimal interaction expertise, limited
possibilities, suboptimal
52Results using Simulated MRS data
53Results using Simulated MRS data
54Results using Simulated MRS data
55Results using in-vivo MRS data
56Preprocessing
HTLS filtering
HLR replace SVD by low rank revealing ULV ? 2
to 100 times faster
57Extension to time series
58Time Series Quantitation
extension of HTLS
HTLSstack
- Select for every signal the corresponding
frequencies dampings
- Determine for every signal the corresponding
amplitudes phases
59HTLSstack versus HTLS
60Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
61TLS extensions
- Mixed LS-TLS and extensions submatrix of XY
error-free (Demmel, 87 88) - Generalized TLS XX1X2, X1 error-free,
errors on X2 correlated
(Gallo 81, Fuller 87,
Golub, Hoffman, Stewart 87, -
Van
Huffel Vandewalle, 91) - Weighted (scaled) TLS DX Y T, D and T
diagonal, unequal error variances in X and Y
(Golub, Van Loan 81, Rao 97, Paige and
Strakos, 01-02) - Restricted TLS error matrix of the form EDEC
(includes equality constraints, LS, TLS, mixed
LS-TLS, ) (Van Huffel, Zha 91) - Total (Lp) approximation uses other norms
(Watson, Späth, Osborne, 82) - Nonlinear measurement error model (Caroll et al.
95) XBG?Y bilinear TLS
approach min
inconsistent (Fuller 87) adjusted LS
estimator
consistentcorrection for small
samples (A. Kukush, I. Markovsky, Van Huffel,
01) Other nonlinear models semi-linear,
quadratic, polynomial (Caroll et al, S. Zwanzig,
A. Kukush, I. Markovsky, Amari 2002)
62TLS extensions (contd.)
- Elementwise-Weighted TLS (differently sized
errors) (M.L.
Rastello, A. Premoli, Markovsky, Kukush, Van
Huffel 2002) - Rowwise-Weighted TLS (rowwise different error
covariance Matrix)
(Markovsky, Kukush, Van Huffel 2004) - Bounded Uncertainties (El Ghaoui 1997, Sayed,
Chandrasekaran, Golub 1997) - Structured TLS (e.g. Toeplitz/Hankel,
displacement rank, regularisation)
- (Rosen, Park, Glick, 96, Lemmerling, De
Moor, N. Mastronardi, Van Huffel, M. Schuermans) - Regularized TLS (truncated TLS, quadratic
eigenvalue problems,)
(Guo, Renaut 2002, P.C. Hansen, D. OLeary, G.
Golub, R. Fierro 1997, D. Sima et al, 2004) - Latency error (equation error) (De Moor,
Lemmerling, A. Yeredor 2002) - Cox proportional Hazards model with EIV (H.
Kuchenhoff 2002) - TLS for large scale problems
(using a preconditioned conjugate gradient
method proposed by A. Björck ,1997) - TLS for large scale Toeplitz systems of equations
(estimate no longer ML)
(J. Kamm and J. Nagy, 1998, applied Newton
iterations combined with a bisection scheme and
circulant factorization preconditioners) - (S. Van Huffel and P. Lemmerling, eds, TLS and
EIV modeling, Kluwer 2002)
63Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
64Structured TLS
Park, Rosen, Glick 94, Lemmerling 99,
Mastronardi 01)
(Abatzoglou,Mendel 87, De Moor 92,
Structured TLS
- Why structured TLS ?
- is structured and noise on different entries of S
is i.i.d. Gaussian white noise - Example Toeplitz matrix
- Computation constrained nonlinear optimization
(Newton) exploit matrix structure ? displacement
rank - Note STLS solution consistent for affine
structures (Kukush 02)
65Structured TLS Is STLS a simple extension
of TLS ?
- fTLS(z) is the objective function we have to
minimize for solving the TLS problem - fSTLS(z) is the objective function we have to
minimize for solving the structured TLS problem
66Structured TLS (contd.) for structured X, y
unstructured
Assume that q lt np different elements of X are
subject to error, e.g. X Toeplitz q lt np-1, X
sparse qltltnp represents the
corrections applied to these elements Vector ?
and correction matrix E are equivalent ry-(XE)?
gives rr(?, ?) STLS problem
Dqxq diagonal matrix of positive
weights Equivalent to TLS when qnp and Lp2 In
order to solve STLS, an nxq matrix X is needed
such that E ? B? B has the following
characteristics - Elements of B are the ? is
with suitable repetition- Number of non-zeros in
B equals number of non-zeros in E- B and E have
similar structure
67Structured TLS (contd.) Construction of E
and B
If ?k is (i,j)th element of E then ? j is (i,k)th
element of B Example
68Structured TLS (contd.) linearize
r(?,?)
- Let ??, ?E and ?? represent small changes
- Use (?E) ? B?? and neglect 2nd order terms in
??, ??. - Linearization gives
- r(???, ???) y-(XE?E)(???)
- ?r(?, ?)-B??-(XE)??
- At each iteration, solve the linearized
minimization - with ,
rank (M)pq if (XE) is of full rank
69Structured TLS (contd.) for structured X
unstructured y
- Input X, D,y, structure on X, tolerance ?
- Output correction matrix E, solution ?
- 1. Set E0, ?0, compute ? from B from ?, set
ry-X ? - 2. repeat
- (a)
- (b) set (c) construct E from ? and B
from ? compute ry-(XE) ? until - p2?2(a)LS problem
70Structured TLS (contd.) exploit structure
of
- assume Lp2 kernel step of basic algorithm LS
problem, X Toeplitz, DI ? exploit low
displacement rank of involved matrices
sparsity of generators ? O(npp2) flops - comparison in efficiency, using simulation
example- alg1 see above (exploiting
displacement structure sparsity)- alg2
basic algorithm without exploitation of structure
(O(np)3)
71Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Case Study STLS in renography
- Other TLS applications
- Conclusions
72Case Study STLS in RENOGRAPHY renogram
deconvolution in kidneys
- In collaboration with the division of nuclear
medicine, Univ. Hospital Leuven, Belgium - co-workers P. Lemmerling, N. Mastronardi, J.
Baetens
73Measurement setup
74Measurement setup
75Overview of the renal scintigraphy
76Used renal regions of interest
Heart
Right kidney
Left kidney
Background region
77TAC (Time-Activity) curves
Kidneys y(t)
Heart or renal artery u(t) y(t)u(t)h(t) where
y(t)renal TAC (OUT) u(t)heart TAC
(IN) h(t)impulse response (unknown) convolution
operator
78Impulse response estimation by discrete
deconvolution
- assume system - linear - time invariant -
causal - zero initial state - finite state
dimension
d(t)
1
t
impulse response
impulse
u(t)
79Convolution illustrated
80Example Impulse response estimation
by discrete deconvolution
Measure u(t) and y(t), find h(t)discrete
deconvolution
0
Ymx1
H
Umxn
u(t) and y(t) noisy ? TLS recommended exploit
matrix structure of U ? STLS (max. likelihood)
81Simulation setup
82Comparison in accuracy
MA versus TLS (MAMatrix Algorithm Solves a
square system YUH Via Gaussian Elimination
with Partial pivoting)
- TLS more accurate than MA, even if curves are
smoothed- accuracy of MA depends heavily on the
number of smoothings- TLS needs no smoothing-
overdetermination not possible with MA- MA fails
to solve rank - deficient problems TLS more
reliable
83Comparison in accuracy
Average relative error of MA and TLS in function
of ?v for 4 different degrees of smoothing
84Renogram deconvolution via STLS
- relations between in- and output at time t
y(t)u(t)h(0)u(t-1)h(1)...u(1)h(t-1)u(0)h(t
)
model selection criterion statistically optimal
relation between in- and outputs
structure to preserve
85STLS solution H can be computed via STLS
algorithm provided Lp2, nM, pN, qM, DI
weighted matrix, ? set to 10-6, XEU?U,
yrY?Y, ?H
0
0
86TLS versus STLS in renal deconvolution
TLS more reliable robust than currently used
algorithm (Gaussian elim. with partial pivoting)
Add regularisation as noise st. dev. increases
(Mastronardi et al, 2004)
87Other STLS Applications
- Medical diagnosis (renography, brain tumor
recognition) - System identification (EIV Kalman filtering,
Hidden Markov) - Modal analysis
- Signal Processing (audio, NMR, speech)
- Astronomy
- Information Retrieval
- Image (Reconstruction, Deblurring, )
- Computer Vision
- Metrology
- Multivariate calibration
- (S. Van Huffel and P. Lemmerling,
eds, TLS and EIV modeling, Kluwer 2002)
88Overview
- Introduction and Problem Formulation
- Univariate EIV regression a statistical approach
- TLS and EIV regression a computational approach
- Historical remarks
- The SVD revisited
- Basic TLS algorithm and computational issues
- TLS properties
- TLS applications
- TLS extensions
- Structured TLS
- STLS applications
- Conclusions
89Conclusions Collaboration must continue ...
between STATISTICS, COMPUTATIONAL
MATHEMATICS and ENGINEERING