See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227304422

LAMBDA: FAQs

Article  in  GPS Solutions · November 2002

DOI: 10.1007/s10291-002-0004-8

CITATIONS READS

21 290

2 authors, including:

C.C.J.M. Tiberius
Delft University of Technology
112 PUBLICATIONS   2,394 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Taking the Fast Lane View project

SuperGPS View project

All content following this page was uploaded by C.C.J.M. Tiberius on 11 November 2015.

The user has requested enhancement of the downloaded file.

 Review article

LAMBDA: FAQs
Peter Joosten Æ Christian Tiberius

the special integer nature of the ambiguities. The result so

Abstract Since its introduction in 1993, the obtained is often referred to as the ‘float solution’. The
LAMBDA method has found widespread use across parameters are usually estimated using a standard least-
the world. The method has been employed in many squares (LSQ) algorithm, or in case of moving receivers, a
geodetic and navigation applications, with lots of Kalman-filter.
satisfied users. Independent tests show that it is Up to then the fact that the ambiguities are of integer
considered the best method for integer carrier phase nature is not yet exploited. Two extra steps are necessary
ambiguity resolution available. But every now and to incorporate this information. In the second step, the
then we still notice some misunderstandings con- float solution of the ambiguities is used to estimate the
cerning the principles and potential of the method. integer ambiguity values. A numerical example is given in
In this contribution we will briefly summarize the Fig. 2. It is this step for which the LAMBDA method has
principles underlying the LAMBDA method, go into been developed.
some of the frequently asked questions on the Finally, in the third step, the computed integer ambiguities
LAMBDA method and try to clarify some of the are used to improve the first-step solution for the
existing misunderstandings. remaining parameters, like baseline coordinates and/or
atmospheric delays. These parameters are recomputed,
again in a least-squares sense, but this time with the am-
biguities constrained to the integer values as obtained
from the second step. This final result is referred to as the
Background ‘fixed’ solution and it generally inherits a much higher
precision than the previously obtained ‘float solution’.
Global navigation satellite system (GNSS) ambiguity res-
olution is the process of effectively accounting for the
integer nature of the unknown cycle ambiguities of dou- Question 1. How do I compute baseline coordinates
ble-difference (DD) carrier phase data. The (sole) purpose using the LAMBDA-method?
of ambiguity resolution is to use the integer ambiguity The LAMBDA method on its own is not meant to compute
constraints as a means of improving significantly on the baseline coordinates. Instead, it is a single, important step
precision of the remaining model parameters, such as in the process of determining very precise coordinates
baseline coordinates and/or atmospheric delays. and/or other parameters of interest from raw GNSS
For the purpose of ambiguity resolution, GNSS data pro- observations. As stated above, a three-step procedure has
cessing is usually carried out in three steps, as shown in to be applied. First you have to compute the float solution.
Fig. 1. In the first step no distinction is made between Any suitable method can be used here, for example an
ambiguities and other parameters. The parameter ordinary least-squares, or in recursion a Kalman filter
estimation problem is solved without taking into account approach. Next the LAMBDA method can be used to
determine the integer values for the ambiguities. The
LAMBDA method takes the estimated float solution â,
Received: 15 February 2002 / Accepted: 7 March 2002 together with its variance–covariance matrix Qâ as input,
Published online: 11 October 2002 and delivers estimated integer ambiguities a
. Finally, again
ª Springer-Verlag 2002 using your own favourite algorithm, the eventual fixed
solution can be computed, using the values output by the
LAMBDA method.
P. Joosten (&) Æ C. Tiberius The LAMBDA algorithm is an autonomous module for the
Mathematical Geodesy and Positioning,
middle arrow in Fig. 1. As such it can easily be embedded
Delft University of Technology,
Thijsseweg 11, 2629 JA Delft, The Netherlands in any of the existing GPS baseline/network software
E-mail: [email protected] packages. In fact, it has been implemented in a number of
Tel.: +31-15-2782713 commercially available software packages, for instance
Fax: +31-15-2783711 SKI-Pro by Leica Geosystems.

DOI 10.1007/s10291-002-0004-8 GPS Solutions (2002) 6:109–114 109

 Review article

Fig. 1 least-squares estimator, has distinct and well-defined

Procedure to solve a problem in which some of the parameters are optimality properties.
known to be integer, like double-difference combinations of
ambiguities in the case of GNSS carrier phase observations. y
represents the vector of carrier phase observations and optionally
pseudo-range code observations, Qy represents the variance–covari- Question 2. Is the LAMBDA method the best
ance matrix of these observations, a represents the vector of unknown method around for ambiguity resolution?
integer carrier phase ambiguities, b represents the vector of unknown The LAMBDA method is a strict implementation of the
baseline coordinates, atmospheric delays and possibly other param- integer least-squares principle. How you rank the LAMB-
eters, the matrices A and B relate the observations y to the unknowns
a and b, and are referred to as ‘design matrices’. â and b^ are the float DA method therefore depends on your appreciation of the
estimates for a and b, Qâ and Qb^ are the corresponding variance– optimality property of integer least-squares. This principle
covariance matrices. Qa^b^ represents the covariances between â and b^. qualifies as ‘best’ as it offers the highest probability of
a
represents the integer estimate of the carrier phase ambiguities. coming up with the correct integer values for the ambi-
Finally, b
represents the fixed solution, with Qb
its variance–
covariance matrix
guities. ‘Best’ is measured by a probability, and is thereby
a statistical property.
In order to understand this, one has to be familiar with the
Estimation principle fact that ambiguity estimators are stochastic quantities,
simply because they are determined from noisy data. Only
In data processing one should distinguish between the in the hypothetical case of perfect observations without
estimation principle and the implementation in an algo- any noise or errors would the float solution always yield
rithm. The estimation principle is the theoretical mandate the correct integer ambiguity values. In reality, however,
that prescribes how estimation values for unknown this is not the case. Any uncertainty in the observations
parameters shall be obtained from observed data values. will propagate and manifest itself as uncertainty in the
Over more than 200 years the least-squares principle has integer ambiguity values.
been in use. It tells that the (weighted) sum of squared Figure 4 shows a repeated single-frequency experiment
residuals is to be minimized. It is a very versatile criterion based on a geometry-free GNSS model. The figure illus-
to obtain a solution that presents a best fit to observed trates empirically how uncertainty in the data (left)
data. A typical example is fitting a regression line through propagates into the ambiguity float estimate (middle) and
a ‘cloud’ of points in a two-dimensional graph (see Fig. 3); finally into the integer ambiguity estimate (right). The
the two real-valued parameters to be determined are the correct integer for the ambiguity is value 4 in this case,
offset and slope of the line. but, as one can see from the graph at right, also other
The outcome, the least-squares estimate, generally consists integer values are frequently obtained. A more extensive
of real-valued numbers for all of the desired parameters. discussion of this can be found in Joosten and Tiberius
Recently, in Teunissen (1993), least-squares theory has (2000).
been extended to deal also with integer parameters. The This concept is formalized in a probabilistic measure, re-
resulting integer estimator, analogously to the ‘ordinary’ ferred to as the ambiguity success-rate. The success-rate is
a number between 0 and 1, or 0 and 100%, and it expresses
the chance, or probability, that the whole set of integer
Fig. 2 ambiguities is correctly estimated. This ambiguity success-
Numerical example of the integer least-squares estimator; notice that
the result (

a) is not equal to the float solution (â) rounded to the
rate depends on three contributing factors: the observation
nearest integers. The transformation matrix ZT is discussed in the equations (functional model), the precision of the ob-
implementation section of this paper servables (the stochastic model) and the chosen principle

110 GPS Solutions (2002) 6:109–114

 Review article

• Amount of data (observations/measurements): not en-

ough data have been collected to estimate the unknown
parameters with sufficient accuracy. More data would
yield more precise (float) estimates. Going back to
Fig. 4, the average of all 1,800 pseudo-range residuals in
Fig. 4 has a much better precision than each of the

Fig. 3
One physical parameter (length) depends on another (temperature)
and the relationship is known or assumed to be linear. By means of
least-squares a regression line is fitted through a set of observed
values

of integer estimation. Changes in any one of these will

affect the success-rate. The first two contributing factors
reflect the strength of the data model and they are given
once the measurement setup is known. As to the principle
of integer estimation, one has a variety of options avail-
able, but integer least-squares maximizes this success-rate,
and can thus be qualified as ‘best’ (see Teunissen 1999a,
1999b).
Question 3. I thought the LAMBDA method is the
best method available, but still I get wrong results,
how is this possible, and what can I do about it?
The LAMBDA method provides the highest probability of
estimating the integer ambiguities correctly, and as such is
indeed the best possible method available. But ‘best’ does
still not imply perfect, i.e. a 100% success-rate. As long as
our data are subject to noise, the success-rate will never
achieve the full 100%. It is still possible to get wrong re-
sults. Some possible causes for this are:

• Quality of data: the less precise (the more noisy) the

measurements are, the less precise the resulting (float)
estimates will be.
c
Fig. 4
Using single-frequency pseudo-range and carrier phase data, the
phase ambiguity of the geometry-free GPS model is estimated in 1,800
single epoch experiments at a 1-s interval. The histogram at left shows
the residuals of the (double-difference) pseudo-range measurements.
Repeating the experiment yields slightly different outcomes each time;
the noise is at the decimetre level. The histogram in the middle
concerns the float ambiguity. It is primarily the noise in the pseudo-
range which is reflected in the noise of the float ambiguity, and as the
L1 wavelength is about 2 dm, the corresponding uncertainty in the
float ambiguity is at the cycle level. In both the graph at the left and in
the middle, the formal Gaussian probability distribution is also
shown. Finally, the integer ambiguity was computed for each
experiment (just by nearest integer rounding, in this simple case with
just one ambiguity), and yields the ‘histogram’ at right. In this case
the integer ambiguity is estimated correctly (value 4) in only 43% of
the experiments

GPS Solutions (2002) 6:109–114 111

 Review article

individual observations. This average is pretty close to Implementation

0.00 m, and when using this average to estimate the
ambiguity as integer, the success-rate would be much An estimation principle, as for instance least-squares,
closer to 100%. provides a mathematical formulation that tells how esti-
• Strength of model: data are fused into parameter esti- mation values are to be obtained from observed data. This
mates according to a certain (functional) relationship formula can commonly be evaluated in different ways. The
that is assumed to exist between observations and un- formula ‘14+3*5’ is equivalent to ‘14+15’, first evaluating
known parameters. The weaker the relationship repre- the multiplication, but also to ‘14+5+5+5’, evaluating the
sented by the model, the less precise the resulting (float) addition three times.
estimates will be. Solving a linear system of equations N x=r for unknown
vector x, with square and full rank matrix N and given
The success-rate depends on the strength of the (mathe- vector r, can be done by computing the matrix inverse.
matical) model. A simple bridge across a river constructed Multiplication of the right hand side r yields the solution
from just a few trees is not to be taken by a 40-tonne truck. x=N–1r. A much more elegant (and efficient) algorithm is
Don’t ask for the unachievable. Don’t expect a 99.99% provided by Gaussian elimination. The system N x=r is
success-rate at a permanently moving rover with five sat- handled by solving, in two steps, an equivalent triangular
ellites over a 100 km baseline with a cheap, single- system by forward and back substitution, with triangular
frequency engine. When the rover receiver is in permanent matrices L and U, as N=L U. Similarly to Gaussian elimi-
motion, new position coordinate unknowns have to be nation to solve a general linear system of equations for
introduced every epoch and this weakens the mathemati- unknown real-valued parameters, the LAMBDA method is
cal model. On long distances (differential) atmospheric an efficient implementation to solve integer least-squares
delays start to play a role, and they are to be accounted for. problems.
As an example, Fig. 5 shows the impact of using dual- Unfortunately, the integer least-squares principle does not
frequency data instead of single-frequency data on ambi- present us with a formula for the integer estimate as an
guity resolution. For visualization purposes the fail-rate is explicit function of the observed pseudo-range and carrier
shown instead of the success-rate. The fail-rate represents phase measurements. The (integer) least-squares norm,
the probability of estimating the wrong integer values for sometimes referred to as the cost-function, that needs to
the carrier phase ambiguities, and thus equals 1 (one) be minimized has an implicit formulation and reads
minus the success-rate. The example is based on the GPS
satellite configuration as of 1 January 2002, as seen in minða^
aÞT Q
1
^ ða
a ^
aÞ
a
Delft. The figure clearly shows the advantage of using dual-
frequency data. In case of dual-frequency data the fail-rate with â the float estimate and Qâ the corresponding vari-
very rarely reaches a level of 0.01 (or 1%), whereas in the ance–covariance matrix. The minimization is solved by a
single-frequency case the fail-rate is above 0.1 (or 10%) for search over grid points a, each representing an ambiguity
most of the day. This clearly shows that instantaneous vector with all integer values, in an ellipsoidal region in the
ambiguity resolution even over short baselines with a ambiguity parameter space (Fig. 6). The vector found is
single-frequency receiver is unlikely to be successful. the integer least-squares estimate a
.
In practice, the estimated (float) ambiguities are highly
correlated and the ellipsoidal region stretches over a
considerable range of cycles. A search for integer vectors
Fig. 5 inside this region can be terribly inefficient. To improve
Ambiguity fail-rates in case of single- (left) and dual- (right)
frequency data, geometry-based, single epoch, short baseline over a
the computational efficiency of the search, the float am-
full 24 h period. The ‘fail-rate’ is the probability of wrong integer biguities are transformed, whereby the elongated ellipsoid
ambiguity estimation, and thus equals 1 minus the success-rate turns into a sphere-like search space. The search is

112 GPS Solutions (2002) 6:109–114

 Review article

Fig. 7
Example of a decorrelating transformation matrix ZT that brings
original ambiguities a into decorrelated ambiguities z, according to
z=ZTa. This example is based on a single epoch of dual-frequency
GPS data to seven satellites on a short baseline. Consequently there
are six L1 double-difference ambiguities, followed by six L2 double-
difference ambiguities in the same order. Zero entries in the matrix ZT
have been left blank
Fig. 6
Two-dimensional example of ellipsoidal search region in ambiguity
parameter space. The boundary of the ellipse represents locations a last-but-one row, one finds a ‘1’ for the L1 ambiguity and a
that all have equal norm ‘–1’ for the corresponding L2 ambiguity, or vice versa, like
the traditional wide-lane combination. But here the total
transformed ambiguity is eventually a combination of five
executed and the eventual result is transformed back. The wide-lane combinations across satellites. Other, similar,
LAMBDA method includes both steps, the transformation combinations occur in the transformation matrix ZT. A
and the actual search (see e.g. Joosten and Tiberius 2000). linear combination can be formed across frequencies and
It should be noted that the transformation is not required across satellites, thus exploiting the actual satellite–
by the (theoretical) estimation principle; it is only to receiver geometry in order to achieve a better decorrela-
achieve a considerable gain in speed in the computation tion than could be achieved by traditional wide-lane
process. Counting the exact amount of money in a big bag combinations.
of coins is more efficient once you first sort the different In short: wide-laning is just one special (and usually
pieces, instead of taking them piece by piece and adding suboptimal) case of the LAMBDA Z-transformation of
their value to the overall running sum. The eventual out- ambiguities.
comes of the transformation, search and back-transfor-
mation together, and of just the search alone, are identical, Question 5. I heard/believe LAMBDA is not suitable
but for the latter approach a ‘little’ more patience has to be for [my application], is this true?
exercised. No, this is not true. The LAMBDA method can intrin-
sically handle any integer estimation problem, as long as
Question 4. Lots of people are always talking you can provide a float solution for your problem. In
about ‘wide-laning’, how is this different other words, as long as you can manage to solve the first
from the LAMBDA method? step of the procedure outlined in Fig. 1, the LAMBDA
Both the LAMBDA method and ‘wide-laning’ techniques method is suitable for solving the second step of this
aim at creating linear combinations of ambiguities that procedure, independent of your application. This im-
have better precision and are less correlated than the plies, for example, that the LAMBDA method will also be
original ambiguities. The way this goal is achieved, suitable for future triple-frequency systems like mod-
however, is different. Where wide-laning makes use of ernized GPS and the envisioned European Galileo. In
certain predefined linear combinations, the LAMBDA fact, the use of the LAMBDA method is not even re-
method creates linear combinations based on the mea- stricted to satellite navigation, as it is currently being
surement precision and the structure of the (mathemat- used to solve the problem of phase-unwrapping as
ical) model employed, for instance the receiver–satellite encountered in the field of INSAR [Interferometric
geometry. Synthetic Aperture Radar, see Hanssen et al. (2001)].
As an example, Fig. 7 shows the decorrelating transfor- LAMBDA is suitable for any problem in which all or part
mation matrix ZT for a case with a single epoch of dual- of the unknown parameters in the model are of an in-
frequency GPS data to seven satellites. Consequently, there teger nature. Maybe there will exist applications even in
are 12 ambiguities to be estimated. For example, in the fields like chemistry or biology.

GPS Solutions (2002) 6:109–114 113

 Review article

Question 6. With the future introduction of least-squares and the LAMBDA method (Teunissen 1993)
modernized triple-frequency GPS and/or Galileo, dates from the early 1990s. An extensive explanation of the
do we need to invent a new method of carrier algorithm can be found in De Jonge and Tiberius (1996).
ambiguity resolution? Joosten and Tiberius (2000) is not directly related to the
No, on the contrary. Compared with present dual-fre- LAMBDA method, but gives an explanation of the success-
quency GPS, ambiguity success-rates will go up, when rate and its importance. Also, most of the textbooks
more signals and more satellites become available (see e.g. dealing with GPS positioning either mention or explain the
Eissfeller et al. 2001). Just provide a float solution, integral LAMBDA method (e.g. Teunissen and Kleusberg 1998;
of both the GPS and Galileo ambiguities, and feed it to Strang and Borre 1997; Hofmann-Wellenhof 1997; Misra
LAMBDA to produce the optimal integer estimate for and Enge 2001). On the website of the department of
them. LAMBDA is ready for it, and will still be the best Mathematical Geodesy and Positioning of the Delft Uni-
method around for ambiguity resolution when perfor- versity of Technology, an extensive list of literature is
mance of the method is measured in terms of success- available, and most papers are also available for download.
rates. The website can be found at the following URL: http://
www.geo.tudelft.nl/mgp. Finally, you can contact the
authors/maintainers of the LAMBDA method via email at
[email protected], but please make sure your questions
LAMBDA method: information are specific rather than general, and indicate in the
‘subject’ of your email that your question concerns the
and feedback LAMBDA method.
Question 7. I have an idea to improve the LAMBDA
method, are you interested?
Of course we would be interested, as there is always a
possibility for improvements. However, before you claim
References
to have improved the LAMBDA method, or to have found De Jonge P, Tiberius CCJM (1996) The LAMBDA method for
a better way of solving the integer ambiguity estimation integer ambiguity estimation: implementation aspects. Techni-
problem, make sure you actually did. By improving the cal report LGR Series, No 12, Delft Geodetic Computing Centre,
mathematical model, by, for example, finding a great way Delft University of Technology, The Netherlands
of dealing with multipath or atmospheric disturbances, Eissfeller B, Tiberius C, Pany T, Biberger R, Schüler T, Heinrichs
you will find that the performance of your integer ambi- G (2001) Real-time kinematic in the light of GPS modernization
guity resolution module will increase. This might suggest and Galileo. Proceedings of ION-GPS-2001, Salt Lake City, Utah,
pp 560–682
you have improved the LAMBDA method, or even found a Hanssen RF, Teunissen PJG, Joosten P (2001) Phase ambiguity
better way of solving the integer ambiguity estimation resolution for stacked radar interferometric data. Proceedings
problem. In such case, you would not have improved the KIS2001, International Symposium on Kinematic Systems in
integer resolution module itself, but you would have Geodesy, Geomatics and Navigation, 5–8 June, Banff, Canada,
served it with higher quality input. Although this certainly pp 317–320
would be very valuable, it would not be an improvement of Hofmann-Wellenhof B, Lichtenegger H, Collins J (1997) Global
the LAMBDA method. Positioning System: theory and practice, 4th edn. Springer,
Berlin Heidelberg New York
Joosten P, Tiberius CCJM (2000) Fixing the ambiguities, are you
Question 8. I have received the LAMBDA software, sure they’re right? GPS World 11(5):46–51
but it seems to be incomplete, what is wrong? Misra P, Enge P (2001) Global Positioning System – signals,
If you have received the software directly from the measurements and performance. Ganga-Jamuna Press, Massa-
Mathematical Geodesy and Positioning Group of the Delft chusetts
Strang G, Borre K (1997) Linear algebra, geodesy, and GPS.
University of Technology, it will be complete. It consists Wellesley-Cambridge Press, Wellesley, Massachusetts
of several routines, which will perform an integer least- Teunissen PJG (1993) Least-squares estimation of the integer GPS
squares estimation using a float solution, together with ambiguities. IAG General Meeting, Invited Lecture, Section IV:
its quality description in the form of a variance–covari- Theory and methodology, Beijing, China
ance matrix as input. Computation of this float solution Teunissen PJG (1999a) A theorem on maximising the probability
is not part of the LAMBDA method, as explained with of correct integer estimation. Artif Satellites 34(2):3–9
Fig. 1. Teunissen PJG (1999b) An optimality property of the integer
least-squares estimator. J Geodesy 73(11):587–593
Teunissen PJG, Kleusberg A (eds) (1998) GPS for geodesy, 2nd
Question 9. I want to know more, where do I get edn. Springer, Berlin Heidelberg New York
more information?
First, there is an extensive list of papers available dealing
with the LAMBDA method. The original paper on integer

114 GPS Solutions (2002) 6:109–114

View publication stats