Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

[J. Res. Natl. Inst. Stand. Technol. 103, 633 (1998)]

Least-Squares Fitting Algorithms of the

NIST Algorithm Testing System

Volume 103
Craig M. Shakarji
National Institute of Standards and
Technology,
Gaithersburg, MD 20899-0001

1.

Number 6

NovemberDecember 1998

This report describes algorithms for fitting

certain curves and surfaces to points in
three dimensions. All fits are based on orthogonal distance regression. The algorithms were developed as reference software for the National Institute of
Standards and Technologys Algorithm Testing System, which has been used for 5
years by NIST and by members of the
American Society of Mechanical Engineers B89.4.10 standards committee. The
Algorithm Testing System itself is described only briefly; the main part of this
paper covers the general linear algebra,

numerical analysis, and optimization methods it employs. Most of the fitting routines rely on the Levenberg-Marquardt optimization routine.
Key words: coordinate measuring machine; curve fitting; least-squares fitting;
Levenberg-Marquardt; orthogonal distance regression; surface fitting.
Accepted: August 7, 1998
Available online: http://www.nist.gov/jres

Introduction

Mathematical software, particularly curve and surface fitting routines, is a critical component of coordinate metrology systems. An important but difficult task
is to assess the performance of such routines [1,2]. The
National Institute of Standards and Technology has developed a software package, the NIST Algorithm Testing System (ATS), that can assist in assessing the performance of geometry-fitting routines [3]. This system has
been aiding in the development of a U. S. standard
(American Society of Mechanical Engineers (ASME)
B89.4.10) for software performance evaluation of coordinate measuring systems [4]. The ATS is also the basis
of NISTs Algorithm Testing and Evaluation Program
for Coordinate Measuring Systems (ATEP-CMS) [5,6].
The ATS incorporates three core modules: a data
generator [7] for defining and generating test data sets,
a collection of reference algorithms which provides a

performance baseline for fitting algorithms, and a comparator for analyzing the results of fitting algorithms
versus the reference algorithms. This paper concentrates
on the development of highly accurate reference algorithms.
The paper is organized as follows. We first introduce
notation and certain key functions and derivatives that
will be used throughout the paper. Section 2 describes
fitting algorithms for linear geometriesplanes and
lines. Lagrange multipliers [8] are used in solving the
constrained minimization problems, which are developed into standard eigenvector problems. Section 3 deals
with nonlinear geometry fitting. We use an unconstrained optimization method that requires derivatives of
appropriate distance functions. These required functions
and derivatives are provided for the reader. Appendix A
gives an outline of the unconstrained optimization

633

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

algorithm (Levenberg-Marquardt) that is modified to
allow for normalization of fitting parameters within the
routine. Appendix B gives, for all the geometries, the
appropriate derivatives needed to create a valuable
check for a local minimum.
1.1

2.1

Defining parameters:
x a point on the plane.
a the direction cosines of the normal to the plane.
Distance equation:
di = d (xi ) = d (xi , x , a ) = a ? (xi 2 x )

Notation and Preliminary Remarks

Assume we are fitting a set of data points, {xi },

i = 1, 2,..., N , that have been translated so that their centroid is the origin. Usually scalar quantities are represented in plain type and matrix or vector quantities with
boldface type. Other notation:
x = (x , y , z )

A point in 3-dimensional space.

|?|

The Euclidean (L2) norm. E. g., |x | = x 2 + y 2 + z 2.

xi = (xi , yi , zi)

The i th data point.

x = (x , y , z)

Objective function:
J (x , a ) = [a ? (xi 2 x )]2
Description:
The centroid of the data must lie on the least-squares
plane. This can be seen because =J = 0 at the least
squares solution, yielding
a ? (xi 2 x ) = 0. Multia
b
plying by 1/N gives
(xi 2 x ) +
(yi 2 y )
N
N
c
(zi 2 z ) = 0. Distributing the summation gives
+
N
a (x 2 x ) + b (y 2 y ) + c (z 2 z ) = 0, which is to say
d (x , x , A ) = 0, i.e., x lies on the least-squares plane.
Since by assumption the data points have been translated to the origin, and since the centroid of the data
must be a point on the least squares plane, we can set
x = 0.
The direction of the fitted plane, a , can be found by
solving the constrained minimization problem, namely,
minimizing J subject to the constraint that | a | = 1. Define a function, G = | a |2 2 1, so that the problem is to
minimize J subject to the constraint that G = 0. The
method of Lagrange multipliers [8] tells us that the
minimum occurs at a point where =J = l =G , for some
real number l . (Here, a , b , and c are treated as independent variables, since the constraint is accounted for in
G . Therefore, = = (/a , /b , /c ).) But =G = 2a ,
and =J = 2(M TM )a , yielding the eigen-problem,
(M TM )a = la , referred to as the normal equations .
This 3 3 3 eigenvector problem can be easily solved
using well-established routines (e.g., Jacobi iterations
[9]). However, we note that the eigenvectors of M TM
are also the singular vectors (from the singular value
decomposition) of M [9]. This allows us to gain numerical stability by applying the singular value decomposition (SVD) to M without ever computing M TM , which
is the method implemented in the ATS.
Finally, we must determine how to select the correct
eigenvector (i.e., singular vector) of the three produced
by the SVD. The normal equations can be written as
follows:

SO O O D

A = (A , B , C )

Direction numbers that specify an orientation, A 0.

a = (a , b , c )

Direction cosines that specify an orientation. Note:

| a | = 1. An orientations direction numbers can be
converted into direction cosines by: a = A /| A |.

The objective function. J is the sum of the squares of

the distances from the data points to the geometry.
J = di2 .

The N 3 3
x 1 y1
x2 y2
...
...
xN yN

3
=

matrix containing the data points:

z1
z2
...
zN

The gradient of a scalar function.

h h h
E.g., =h (x , y , z ) = (hx , hy , hz ) =
,
,
.
x y z

For each geometry, we show the defining parameters,

the equation for the orthogonal distance from a single
data point to the geometry, the objective function, and
a brief description of the steps in the calculation.

2.

O
O

1
xi , yi , zi .
N
(Note: These and all other sums in this paper are
taken from i = 1, 2,..., N .)
The centroid of the data,

Plane Fitting

Linear Geometries

O
O
O

Linear geometries (lines and planes) are solved using

Lagrange multipliers on a constrained minimization
problem. Both cases reduce to a rather simple eigenproblem.

634

xi (a ? xi ) = la
yi (a ? xi ) = lb
zi (a ? xi ) = lc

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

Multiplying these three equations by a , b , and c ,
respectively, then summing the equations gives
(a ? xi )2 = l | a |2 = l .
But the sum on the left is just the objective function,
i.e., the sum of the squares of the residuals. Thus the
value of l is the objective function, hence the correct
eigenvector for the least-squares solution corresponds to
the smallest eigenvalue. When using the SVD, we
choose the singular vector corresponding to the smallest
singular value, since under these conditions the singular
values are the square roots of the eigenvalues [10].

and g respectively, giving necessary derivatives, which

are used throughout the rest of this paper. We compute
the nonlinear fits using unconstrained minimization algorithms, so we define the line and plane distance functions in terms of direction numbers rather than direction
cosines.
Let g (xi , x , A ) denote the distance from the point, xi ,
to the plane defined by the point, x , and the normal
direction, a = A /| A |. The value of g is given by:
gi = g (xi , x , A ) = a ? (xi 2 x ) = a (xi 2 x ) + b (yi 2 y ) +
c (zi 2 z ).
Let f (xi , x , A ) denote the distance from the point, xi ,
to the line defined by the point, x , and the direction,
a = A /| A |. The value of f is given by: fi = f (xi , x , A ) =
| a 3 (xi 2 x )|. That is,

2.2

Line Fitting

Defining parameters:
x a point on the line.
a the direction cosines of the line.

u = c (yi 2 y ) 2 b (zi 2 z )
fi = u 2 + v 2 + w 2, where v = a (zi 2 z ) 2 c (xi 2 x )
w = b (xi 2 x ) 2 a (yi 2 y )

Distance equation:
d (xi ) = d (xi , x , a ) = | a 3 (xi 2 x )| (which also equals
| xi 2 x |2 2 [a ? (xi 2 x )]2) where 3 denotes the
vector cross product.

This expression for f is used because of its numerical

stability. One should note that f could also be expressed
(for derivative calculations) as fi = | xi 2 x |2 2 gi2 .
Note: A , B , and C are independent variables, whereas
a , b , and c are not, because the constraint
a 2 + b 2 + c 2 = 1 causes a , b , and c to depend on each
other. When dealing with the nonlinear geometries we
treat f and g as functions dependent on A , as opposed to
a , in order to use unconstrained minimization algorithms. Treating f and g as functions dependent on a
would force us to restrict ourselves to using constrained
minimization solvers. In the linear cases, we did solve
constrained minimization problems. So when we differentiate with respect to A , for example, we treat a , b , and
c all as functions of A , B , and C (e.g., a = A /
A 2 + B 2 + C 2). This yields the following array of
derivatives:

Objective function:
J (x , a ) = | a 3 (xi 2 x )|2
Description:
Since the data has been translated to the origin, and
since the centroid of the data must be a point on the least
squares plane (similar to the case above with the least
squares plane), we set x = 0.
The direction a can be found by following the same
strategy as in the case of a plane. For line fitting, the
normal equations are M TMa = la just as in the case of
plane fitting. Once again, the correct eigenvector must
be chosen to minimize the sum-of-squares of the residuals. As shown with planes, we obtain l = (a ? x )2, so
J = 2 l + | xi |2, meaning that J is minimized when l
is maximized. Thus the correct eigenvector choice is the
one corresponding to the largest eigenvalue. As in the
case of plane fitting, numerical stability is gained by
finding the eigenvectors of M through the SVD, rather
than by solving the normal equations. Since the singular
values are the square roots of the eigenvalues [10], we
choose the eigenvector corresponding to the largest singular value.

3.

Nonlinear Geometries

3.1

Utility Functions f and g

34 3

a a a
A B C
12a 2 2ab
2ac
b b b
1
2
=b =
=
2ab
12b 2bc
A B C
|A|
=c
c c c
2ac
2bc
12c 2
A B C

=a

These algorithms normalize A at every step, so for

simplicity of expressing derivatives, assume | A | = 1. A
remains unconstrained; we just assume it happens to
have unit magnitude.

The line and plane distance functions arise quite often

in this paper, thus we define them here, calling them f
635

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

The derivatives for fi and gi are then

In the ATS, nonlinear geometries are fit in an unconstrained manner using the Levenberg-Marquardt algorithm. The algorithm requires an initial guess as well
as the first derivatives of the distance function. In practice it converges quickly and accurately even with a wide
range of initial guesses. Details of this algorithm are
given in appendix A. Additionally, the code allows us to
normalize the fitting parameters after every iteration of
the algorithm. For each geometry we list the distance
and objective functions, the appropriate derivatives, and
the parameter normalization we use.

gi
=2a
x
gi
=2b
y
gi
=2c
z
gi
= (xi 2 x ) 2 agi
A
gi
= (yi 2 y ) 2 bgi
B
gi
= (zi 2 z ) 2 cgi
C
fi
= [agi 2 (xi 2 x )]/fi
x
fi
= [bgi 2 (yi 2 y )]/fi
y
fi
= [cgi 2 (zi 2 z )]/fi
z
fi
= gi [agi 2 (xi 2 x )]/fi
A
fi
= gi [bgi 2 (yi 2 y )]/fi
B
fi
= gi [cgi 2 (zi 2 z )]/fi
C

3.3

Defining parameters:
x the center of the sphere.
r the radius of the sphere.
Distance equation:
d (xi ) = | xi 2 x | 2 r

Objective function:
J (x , r ) = (| xi 2 x | 2 r )2
Normalization:
(None)

The above derivatives of fi are undefined when fi = 0

(i.e., when xi is on the line.) In this rarely needed case
the gradient is given by:

S1 2 a , 1 2 b , 1 2 c ,
2

Derivatives:
di
= 2 (xi 2 x )/| xi 2 x |
x
di
= 2 (yi 2 y )/| xi 2 x |
y
di
= 2 (zi 2 z )/| xi 2 x |
z
di
=21
r

g 1 2 a 2, g 1 2 b 2, g 1 2 c 2 .
For cylinders, cones, and tori, the line associated with
f is the geometrys axis. For cones the plane associated
with g is the plane through the point, x , perpendicular to
the axis. For tori, the plane associated with g is the plane
perpendicular to the axis that divides the torus in half.
3.2

Sphere Fitting

3.4

Two-Dimensional Circle Fitting

This case is simply the sphere fit (above) restricted to

two-dimensions:
Defining parameters:
x the x -coordinate of the center of the circle.
y the y -coordinate of the center of the circle.
r the radius of the circle.

Choice of Algorithm

Good optimization algorithms can readily be found to

minimize the objective function, J . Usually such an
algorithm will require an initial guess, along with partial derivatives, either of J itself or of the distance function, di . Both sets of derivatives are given in this paper
(those for J in the appendix). These should enable a
reader to implement a least-squares algorithm even if the
optimization algorithm used differs from the authors
choice, which follows.

Distance equation:
d (xi , yi ) = (xi 2 x )2 + (yi 2 y )2 2 r
Objective function:
J (x , y , r ) =

636

OS(x 2 x ) + (y 2 y ) 2 rD
i

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

Normalization:
(None)

2. Rotate the data such that the least-squares plane is

the x -y plane.
3. Project the rotated data points onto the x -y plane.
4. Compute the 2D circle fit in the x -y plane.
5. Rotate back to the original orientation.
6. Perform a full 3D minimization search over all the
parameters.

Derivatives:
di
= 2 (xi 2 x )/(di + r )
x
di
= 2 (yi 2 y )/(di + r )
y
di
=21
r
3.5

Some coordinate measuring system software packages stop at step (5) and report the orientation, center,
and radius as the least-squares circle in 3D. This approach is valid when the projection onto the plane is
done simply to compensate for measurement errors on
points which would otherwise be coplanar. But this
method does not in general produce the circle yielding
the least sum-of-squares possible (even though it is usually a good approximation.) In order to achieve the true
3D least-squares fit, the circle computed at step (5) is
used as an initial guess in the Levenberg-Marquardt
algorithm, which optimizes over all the parameters
simultaneously [step (6)].
Step (2) is carried out using the appropriate rotation
matrix to rotate the direction, a , to the z -direction,
namely,

Three-Dimensional Circle Fitting

Defining parameters:
x the center of the circle.
A the direction numbers of the normal to circles
plane.
r the radius of the circle.
Distance equation:
d (xi ) = gi2 + (fi 2 r )2
Objective function:
J (x , A , r ) =

OSg + (f 2 r ) D
1
i

12

Normalization:
A A /| A | (Here and elsewhere, denotes assignment of value. In this case, the
value of A is replaced by the value
A /| A |.)
Derivatives:
di
= [gi (gi )x + fi (fi )x ]/di
x
di
= [gi (gi )y + fi (fi )y ]/di
y
di
= [gi (gi )z + fi (fi )z ]/di
z
di
= [gi (gi )A + fi (fi )A ]/di
A
di
= [gi (gi )B + fi (fi )B ]/di
B
di
= [gi (gi )C + fi (fi )C ]/di
C
di
= 2 (fi 2 r )/di
r

a2
1+c

2 ab
1+c

2 ab
1+c
12

b2
1+c

2a

2b

If c < 0, a is replaced with 2 a , thus rotating the

direction to the minus z -direction (which is adequate for
our purposes.) Step (5) is carried out using the appropriate rotation matrix to rotate the z -direction to the direction, a . Namely,

12

3.6

Description:
We use a multi-step process to accomplish 3D circle
fitting:
1. Compute the least-squares plane of the data.

a2
1+c

2 ab
1+c
2a

2 ab
1+c
12

b2
1+c

2b

b
c

Cylinder Fitting

Defining parameters:
x a point on the cylinder axis.
A the direction numbers of the cylinder axis.
r the radius of the cylinder.

637

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

Distance equation:
d (xi ) = fi 2 r

Derivatives:
di
= (fi )x cosc + (gi )x sinc
x
di
= (fi )y cosc + (gi )y sinc
y
di
= (fi )z cosc + (gi )z sinc
z
di
= (fi )A cosc + (gi )A sinc
A
di
= (fi )x cosc + (gi )x sinc
B
di
= (fi )x cosc + (gi )x sinc
C
di
=21
s
di
= 2 fi sinc + gi cosc
c

Objective function:
J (x , A , r ) = (fi 2 r )2
Normalization:
A A /| A |
x (point on axis closest to origin)
Derivatives:
di
= (fi )x
x
di
= (fi )y
y
di
= (fi )z
z
di
= (fi )A
A
di
= (fi )B
B
di
= (fi )C
C
di
=21
r
3.7

3.8

Torus Fitting

Defining parameters:
x the torus center.
A the direction numbers of the torus axis.
r the major radius.
R the minor radius.
Distance equation:
d (xi ) = gi2 + (fi 2 r )2 2 R

Cone Fitting

Defining parameters:
x a point on the cone axis (not the apex).
A the direction numbers of the cone axis (pointing
toward the apex).
s the orthogonal distance from the point, x , to the
cone.
c the cones apex semi-angle.

Objective function:
J (x , A , r , R ) =

OFg + (f 2 r ) 2 RG
2
i

Normalization:
A A /| A |

Distance equation:
d (xi ) = fi cosc + gi sinc 2 s

Derivatives:
di
= [gi (gi )x + (fi 2 r )(fi )x ]/(di + R )
x
di
= [gi (gi )y + (fi 2 r )(fi )y ]/(di + R )
y
di
= [gi (gi )z + (fi 2 r )(fi )z ]/(di + R )
z
di
= [gi (gi )A + (fi 2 r )(fi )A ]/(di + R )
A
di
= [gi (gi )B + (fi 2 r )(fi )B ]/(di + R )
B
di
= [gi (gi )C + (fi 2 r )(fi )C ]/(di + R )
C
di
= 2 (fi 2 r )(di + R )
r
di
=21
R

Objective function:
J (x , A , s , c ) = (fi cosc + gi sinc 2 s )2
Normalization:
A A /| A |
x (point on axis closest to origin)
c c (mod 2p)
if c > p then [c c (mod p); A 2 A ]
p
if c > then c p 2 c
2
if s < 0 then [s 2 s ; A 2 A ]

638

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

4.

(x , y , z , A , B , C , r ) and di ( p ) = fi 2 r . One derivation

of the algorithm [12] begins by approximating J as a
linear function of p , J ( p ):

Discussion

The algorithms have been implemented in the ATS

and have been used for 5 years by NIST and by members
of the ASME B89.4.10 Working Group. In general they
have performed extremely well. They have successfully
solved a number of difficult fitting problems that could
not be solved by many commercial software packages
used on Coordinate Measuring Systems (CMSs). (Some
of the most difficult problems are cylinders or cones
sampled over a small patch.) The ATS algorithms have
an extremely broad range of convergence. Failure to
converge has only been observed for pathological fitting
problems (e.g., fitting a circle to collinear points). Special checks can detect most of these situations.
The ATS algorithms are generally robust. For most
fits, a good starting guess is not required to reach the
global minimum. This is due, in part, to the careful
choice of fitting parameters, the use of certain constraints, and, for cylinders and cones, the technique of
restarting a search after an initial solution is found.

J ( p ) J ( p ) =

O(d ( p ) + =d ( p ) ? p )
i

where =di ( p0) is the gradient of di ( p ) evaluated at an

initial guess, p0. This approximation is valid within a
certain trust region radius. The derivation then considers
how to minimize J ( p ). A search direction is calculated
based on J ( p ), and a search is made in that direction
within the limits of the trust region radius for a point,
pnew, such that J ( pnew) < J ( p0). When pnew is found, it
becomes the new p0 for another iteration of the above
process.
The basis for the algorithm is the result that the solution, p *, to each iteration can be expressed as
p * = p (l ) = 2 (F0TF0 + lD TD )21 F0Td ( p0) where F0 is
a matrix having =di ( p0) as its i th row, D is an appropriate weighting matrix, d ( p0) is the vector of residuals,
di ( p0), and l $ 0 is a variable called the LevenbergMarquardt parameter. This parameter can be considered
the Lagrange multiplier for the constraint that each
search be limited to the trust region radius.
The Levenberg-Marquardt algorithm is presented in
the figure below. The matrix F0TF0 + lD TD is named H
and the vector F0Td ( p0) is called v . The algorithm includes a modification suggested by Nash [11], which is
to use a weighting matrix defined so that D TD is the
identity matrix plus the diagonal of F0TF0. Nashs use of

5. Appendix A. The Levenberg-Marquardt

Algorithm
The Levenberg-Marquardt algorithm [11] finds the
vector, p , that minimizes an objective function, J ( p ) =
S di2 ( p ), where di ( p ) is the distance from the point, xi ,
to the geometry defined by the vector of parameters, p .
In the case of a cylinder, for example, p = (x , A , r ) =

Fig. 1. Levenberg-Marquardt algorithm.

639

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

O
O
O
O

J
= 2 fi (fi )z
z
J
= 2 fi (fi )A
A
J
= 2 fi (fi )B
B
J
= 2 fi (fi )C
C

the identity in the definition of D forces H to be positive

definite. (Note that D is never calculated.) Since H is
also symmetric, the system Hx = 2 v can be reliably
solved using the Cholesky decomposition [9]. We chose
10 and 0.04 as factors with which to increment and
decrement l respectively. Finally, note that we normalize the parameter vector, p , at every iteration. However,
normalization of p never changes the value of the objective function, J ( p ). This routine is the basis of the ATS
implementation of the Levenberg-Marquardt algorithm.

6.

Appendix B.
Correctness

Spheres

Od (x 2 x )/| x 2 x |
Od (y 2 y )/| x 2 x |
Od (z 2 z )/| x 2 x |
Od

J
=22
x
J
=22
y
J
=22
z
J
=22
r

Gradient Test for

One helpful check is to compute the gradient of the

objective function at the computed minimum. The correct least-squares solution yields =J = 0. Also, these
derivatives are important because several minimization
algorithms require that these derivatives be included.
The definitions of f and g are kept the same as they were
in Sec. 3. For every nonlinear geometry, assume the
same definition for the objective function, J , as specified in Sec. 3. For the linear geometries, the objective
functions are here defined in terms of unconstrained
parameters.

2D Circles

Od (x 2 x )/(d + r )
Od (y 2 y )/(d + r )
Od

J
=22
x
J
=22
y
J
=22
r

Planes
J (x , A ) =

Og

3D Circles

2
i

O
O
O
O
O
O
O

J
= 2 [gi (gi )x + (fi 2 r )(fi )x ]
x
J
= 2 [gi (gi )y + (fi 2 r )(fi )y ]
y
J
= 2 [gi (gi )z + (fi 2 r )(fi )z ]
z
J
= 2 [gi (gi )A + (fi 2 r )(fi )A ]
A
J
= 2 [gi (gi )B + (fi 2 r )(fi )B ]
B
J
= 2 [gi (gi )C + (fi 2 r )(fi )C ]
C
J
= 2 2 (fi 2 r )
r

O
O
O
O
O
O

J
= 2 gi (gi )x
x
J
= 2 gi (gi )y
y
J
= 2 gi (gi )z
z
J
= 2 gi (gi )A
A
J
= 2 gi (gi )B
B
J
= 2 gi (gi )C
C

Lines
J (x , A ) =
J
=2
x
J
=2
y

Of

Cylinders

O(f 2 r )(f )
O(f 2 r )(f )
O(f 2 r )(f )

2
i

J
=2
x
J
=2
y
J
=2
z

Of (f )
Of (f )
i

i x

i y

640

i x

i y

i z

Volume 103, Number 6, NovemberDecember 1998

Journal of Research of the National Institute of Standards and Technology

O
O
O
O

J
= 2 (fi 2 r )(fi )A
A
J
= 2 (fi 2 r )(fi )B
B
J
= 2 (fi 2 r )(fi )C
C
J
= 2 2 (fi 2 r )
r

7.

References

[1] T. H. Hopp, Computational MetrologyVol. 6, No. 4, Manufacturing Review, American Society of Mechanical Engineers,
New York (1993) pp. 295304.
[2] R. Walker, CMM Form Tolerance Algorithm Testing, GIDEP
Alert X1-A-88-01, Government-Industry Data Exchange Program, Department of Defense, Washington, D.C. (1988).
[3] D. A. Rosenfeld, Users Guide for the Algorithm Testing System Version 2.0, NISTIR 5674, National Institute of Standards
and Technology, Gaithersburg, MD (1995).
[4] ASME, Methods for Performance Evaluation of Coordinate
Measuring System Software, B89.4.10-199x, DRAFT, American Society of Mechanical Engineers, New York (1998).
[5] C. Diaz and T. H. Hopp, NIST Special Test Service: The Algorithm Testing and Evaluation Program for Coordinate Measuring Systems, NISTSP 250-41, National Institute of Standards
and Technology, Gaithersburg, MD (to be published).
[6] C. Diaz and T. H. Hopp, Testing Coordinate Measuring Systems
Software, Proceedings of ASQC Measurement Quality Conference, October 2627, National Institutes of Standards and Technology, Gaithersburg, MD (1993).
[7] M. E. A. Algeo and T. H. Hopp, Form Error Models of the NIST
Algorithm Testing System, NISTIR 4740, National Institute of
Standards and Technology, Gaithersburg, MD, January 1992.
[8] T. M. Apostol, Calculus, Volume II, Second Edition, John Wiley & Sons, Inc., New York (1969) pp. 314318.
[9] G. H. Golub and C. F. van Loan, Matrix Computations, The
Johns Hopkins University Press, Baltimore, MD (1983).
[10] G. W. Stewart, Matrix Computations, Academic Press, Inc.,
Orlando, FL (1973) p. 320.
[11] J. C. Nash, Compact Numerical Methods for Computers: Linear
Algebra and Function Minimisation, Adam Hilger, Ltd., Bristol,
England (1979).
[12] J. J. More, The Levenberg-Marquardt Algorithm: Implementation and Theory, in Numerical Analysis: Proc. Biennial Conf. on
Numerical Analysis, Dold and Eckmann, eds., Dundee, Scotland
(1977).

Cones

O
O
O
O
O
O
O
O

J
= 2 di [(fi )x cosc + (gi )x sinc ]
x
J
= 2 di [(fi )y cosc + (gi )y sinc ]
y
J
= 2 di [(fi )z cosc + (gi )z sinc ]
z
J
= 2 di [(fi )A cosc + (gi )A sinc ]
A
J
= 2 di [(fi )B cosc + (gi )B sinc ]
B
J
= 2 di [(fi )C cosc + (gi )C sinc ]
C
J
= 2 di [2 fi sinc + gi cosc ]
c
J
= 2 2 di
s

Tori

O
O
O
O
O
O
O
O

J
di
=2
[g (g ) + (fi 2 r )(fi )x ]
x
di + R i i x
di
J
=2
[g (g ) + (fi 2 r )(fi )y ]
y
di + R i i y
J
di
=2
[g (g ) + (fi 2 r )(fi )z ]
z
di + R i i z
di
J
=2
[g (g ) + (fi 2 r )(fi )A ]
A
di + R i i A
J
di
=2
[g (g ) + (fi 2 r )(fi )B ]
B
di + R i i B
J
di
=2
[g (g ) + (fi 2 r )(fi )C ]
C
di + R i i C
J
di
=2
[g (g ) 2 (fi 2 r )]
di + R i i r
r
J
= 2 2 di
R

About the author: Craig Shakarji is a mathematician

in the Manufacturing Systems Integration Division of
the NIST Manufacturing Engineering Laboratory. The
National Institute of Standards and Technology is an
agency of the Technology Administration, U.S. Department of Commerce.

641