Next2

Principal Curves Revisited
Robert Tibshirani
Department of Preventive Medicine and Biostatistics
and
Department of Statistics
University of Toronto

Abstract
A principal curve (Hastie and Stuetzle, 1989) is a smooth curve passing
through the middle" of a distribution or data cloud, and is a generaliza-
tion of linear principal components. We give an alternative de nition of
a principal curve, based on a mixture model. Estimation is carried out
through an EM algorithm. Some comparisons are made to the Hastie-
Stuetzle de nition.

1 Introduction
Suppose we have a random vector Y = (Y1 ; Y2 ; : : : Yp ) with density gY (y). How
can we draw a smooth curve f (s) through the middle" of the distribution of
Y? Hastie (1984) and Hastie and Stuetzle (1989) (hereafter HS) proposed a
generalization of linear principal components known as principal curves. Let
f (s) = (f1 (s); : : : fp(s)) be a curve in Rp parametrized by a real argument s and
de ne the projection index sf (y) to be the value of s corresponding to the point
on f (s) that is closest to y. Then HS de ne a principal curve to be a curve
satisfying the self-consistency property
f (s) = E(Yjsf (y) = s) (1)
If we think of projecting each point Y to the curve f (s), this says that f (s)
is the average of all points that project to it. In a sense, f (s) passes through
the middle" of the distribution of Y. This is illustrated in Figure 1.
P
HS showed that a principal curve is a critical point of the squared distance
function E p (Yj ? fj (s)) , and in this sense, it generalizes the minimum dis-
2
1
tance property of linear principal components.
HS proposed the following alternating algorithm for determining f and s:

1

y

T

f(s)

Figure 1: Principal curve schematic. The grey area is the distribution of Y and
the solid curve is a principal curve of this density. Given any point on the curve
(e.g. T"), the principal curve is such that the average of all points that project
to T (solid line) is T itself.

HS principal curve algorithm
a) Start with f (s) = E(Y)+ ds where d is the rst eigenvector of the covariance
matrix of Y and s = sf (y) for each y.
b) Fix s and minimize E kY ? f (s)k by setting fj (s) = E(Yj jsf (y) = s) for
2

each j .
c) Fix f and set s = sf (y) for each y.
d) Iterate steps b and c until the change in E kY ? f (s)k2 is less than some
threshold.
In the data case, the conditional expectations in step (b) are replaced by a
smoother or nonparametric regression estimate. HS use locally weighted running
lines or cubic smoothing splines. In step (c) sf (y) is found by projecting y
(numerically) to the curve f .
While this de nition seems appealing, HS note the following (somewhat
unsettling) property. Suppose Y satis es
Yj = fj (S ) + j ; j = 1; 2; : : : p

where S and j ; j = 1; 2 : : : p are independent with E( j ) = 0 for all j . Then
f = (f1; f2; : : : fp) is not in general a principal curve of the distribution of Y.
2

f(s)

Figure 2: Bias in principal curve: the solid curve is the generating function, and
the grey circle indicates a bivariate normal distribution centred at the black dot.
The two straight lines indicate the part of the distribution that projects to the
target point (black dot). Since there is more mass (between the lines) outside
of the arc than inside, the principal curve (dashed curve) will fall outside of the
generating curve. The resulting principal curve (broken curve) falls outside of
the generating curve.

They give as an illustration the situation in Figure 2. S is uniform on the arc of
a circle (solid curve), and the errors are circular normal. The two straight lines
indicate the part of the distribution that projects to the target point (black dot).
Since there is more mass (between the lines) outside of the arc than inside, the
principal curve (dashed curve) will fall outside of the generating curve. This
continues to hold in the limit, as the distance between the two straight lines
goes to zero. As HS note, however, the estimation bias in the data case tends
to cancel out this model bias, so that it is not clear whether this bias is a real
problem in practice.
In this paper we view the principal curve problem in terms of a mixture
model. This leads to a di erent de nition of principal curves and a new al-
gorithm for their estimation. The new de nition does not share the di culty
mentioned above.

3

2 Principal curves for distributions
Let Y = (Y1 ; Y2 ; : : : Yp ) be a random vector with density gY (y). In order to
de ne a principal curve, we imagine that each Y value was generated in two
stages: 1) a latent variable S was generated according to some distribution
gS (s), and 2) Y = (Y1 ; : : : ; Yp ) was generated from a conditional distribution
gYjs having mean f (S ), a point on a curve in Rp , with Y1 ; : : : Yp conditionally
independent given s. Hence we de ne a principal curve of gY to be a triplet
fgS ; gYjS ; f g satisfying the following conditions:
R
I. gS (s) and gYjS (yjs) are consistent with gY (y), that is, gY (y) = y gYjs (yjs)gS (s)ds
II. Y1 ; : : : Yp are conditionally independent given s.
III. f (s) is a curve in Rp parametrized over s 2 ?, a closed interval in R1 ,
satisfying f (s) = E(YjS = s)
Notice that this de nition involves not only a curve f but a decomposition
of gY into gS and gYjS . From a conceptual standpoint, assumption (II) is
not really necessary; however it is an important simplifying assumption for the
estimation procedures described later in the paper.
One obvious advantage of this new de nition is that it does not su er from
R y gYjs(yjs)gS (s)ds where a latent variable S has density gs, and Y gYjs
the problem of Figure 2. Suppose we de ne a distribution gY by gY (y) =
with mean f (s). Then by de nition the generating triplet fgS ; gYjS ; f g satis es
properties I, II, and III and hence is a principal curve. Thus for example the
solid curve in Figure 2 is a principal curve according to I, II, and III.
When do the HS de nition (1) and the new de nition agree? In general they
are not the same except in special cases. Suppose S gS , f (s) is linear, and
the support of the distribution gYjs is only on the projection line orthogonal to
f (s) at s: (?)
Then the Y values generated from the point S = s are exactly the Y values on
the projection line orthogonal to f (s) at s, and therefore
E(YjS = s) = E(Yjsf (y) = s)
The assumption (?) is somewhat unnatural, however, and violates our earlier
assumption (II) .
There are other special cases for which a curve f (s) is a principal curve under
both de nitions. One can check that for a multivariate normal distribution,
the principal components are principal curves under the new de nition. HS
note that for a multivariate normal distribution, the principal components are
principal curves under their de nition as well.

4

3 Principal curves for datasets
Suppose we have observations yi = (yi1 ; : : : yip ), i = 1; 2; : : : n and unobserved
latent data s1 ; : : : sn . We assume the model of the previous section
si gs (S ); yi gYjs ; f (s) = E(Yjs)
i
(2)
with (yi1 ; : : : yip ) conditionally independent given si .
First notice that if we consider the unobserved values s = (s1 ; : : : sn ) as
xed parameters rather than random variables, and assume that the conditional
distributions gY jS are normal with equal known variance, then is easily seen
that the HS principal curve algorithm for datasets can be derived as a penalized
maximum likelihood estimator. More speci cally, maximization of the penalized
least squares criterion
X X(
n n
yij ? fj (si )) +
2
X Z
p 1
fj s( ) ]2 ds
00
(3)
j
i=1 j =1 j =1 0

leads to the HS algorithm with cubic smoothing splines for estimating each fj .
The parameter > 0 governs the tradeo between t and smoothness of the
coordinate functions. Details are given in HS section 5.5. We note that in
problems where the number of nuisance" parameters goes to in nity with the
sample size, maximization of the likelihood over all of the parameters can lead
to inconsistent or ine cient estimates.
Our approach, on the other hand, is to work with the likelihood implied
by model (2). The form of gS (s) is left completely unspeci ed but gYjS is
assumed to be some parametric family. We allow additional parameters (s)
in the speci cation of gYjS and let the complete set of unknowns be = (s) =
(f (s); (s)).
Consider maximum likelihood estimation f (s) and (s), and nonparametric
maximum likelihood estimation of gS . The log-likelihood has the form of a
mixture:
`( )=
X log Z
n
gY jS (yi j )gS (s)ds;
1
(4)
A general theorem on mixtures given by Lindsay (1983) implies that for xed
f (s) and (s), the nonparametric maximum likelihood estimate of the mixing
density gS is discrete with at most n support points. Denote these support
points by a1 ; a2 ; : : : an .
Our approach to maximization of ` is via the EM algorithm (Dempster et
al. 1977, sect. 4.3). EM uses the complete data log-likelihood
X log
n X log
n
`0 ( )= s(yi j (si )) +
gY j ( )
g S si
1 1

5

The E step" starts with a value f 0 and computes the function
Q ( j 0 ) = E f`0( )jy; 0 g
where y denotes the observations (yi ); i = 1; : : : n. Q is considered a function
of with 0 xed. The M step maximizes Q( j 0 ) over to give 1 and the
process is iterated until convergence.
Let wik = P r(si = ak jy; 0 ), vk = gS (ak ) = P r(s = ak ).
We may write Q as

Q (j 0
)=
XX
n n
wik log gY jS (yi j (ak )) +
XX
n n
wik log vk
i=1 k=1 i=1 k=1
Now by Bayes theorem
wik gY jS (yi j (ak ))
Y gY (yh )vk (5)
h6=i

The quantity gY (yh ) is computed from gY (yh ) = n=1 gYjS (yh jak ; 0 )vk
k
P
A Gaussian form for gYjS is most convenient and gives

(yi j (s)) =
Y
p
gY jS fj (s); s (yij )
j( )
where
j =1
(y) = 1 p exp ?(y ? f (a ))2 =2 2(a )] (6)
fj (ak ); a j k j k
j (ak ) 2
j ( k)

Maximizing Q( j 0 ) gives
Pn
^
fj (ak ) = Pn
i =1 wik yij

Pn i =1 wik
? f^j (ak )) ; j = 1 : : : p
^j (ak ) =
2 i =1
P(n
wik yij 2

1X
i =1 w ik
n
^
vk = n wik (7)
i=1
^
The rst expression in (7) says that fj (ak ) is a weighted average of fyij ; i =
1; 2; : : : ng. The weights are the relative probability under the current model
that s = ak gave rise to yij . If j s are equal, the weight is a function of the
Euclidean distance from y to f (ak ).
These equations are certainly not new: they are EM steps for tting a multi-
variate normal mixture model. See for example Titterington et al. (1985, page
86-89). Notice that the log-likelihood is not a function of the support values
a1 ; a2 ; : : : an but only the means fj (ak ) and variances j (ak ) at each of these
2

points.

6

As an example, we generated 100 observations from the circle model
Y1 = 3 sin( ) + :5 1
Y2 = 3 cos( ) + :5 2
where U =4; 5 =4] and N (0; 1). The results of the rst 18 iterations
of the EM procedure (showing every second iteration) are in Figure 3.
The procedure does a fair job of approximating the structure of the data.
Although the approximating points seem to follow a fairly smooth path, there
is no guarantee that this will happen in general. We show how to rectify this in
the next section.

4 Multiple minima and regularization
Consider the log-likelihood (4) in the Gaussian case (6). As mentioned in sec-
tion 3, a theorem of Lindsay (1983) tells us that for xed f (s) and (s), the
nonparametric maximum likelihood estimate of the mixing distribution gS is
discrete with at most n support points. We note the following:
a) The log-likelihood has a global maximum of +1 when ^(s) is any curve that
f
interpolates the data, f ^(ak ) = yk ; k = 1; 2; : : : n, P r(S = ak ) > 0; k =
1; 2; : : : n and ^j (ak ) = 0; k = 1; 2; : : : n.
2

b) For any candidate solution ~, the log-likelihood is only a function of the
f
values of ~ at the support points with positive probability. Hence any
f
curve that agrees with ~ at these points gives the same value of the log-
f
likelihood.
In view of (a), the global maximum is uninteresting and we might hope that
the EM algorithm converges instead to an interesting" local maximum. This
happened in Example 1 above, but may or may not occur in general.
A remedy for these problems is add a regularization component to the log-
likelihood. We seek to maximize the penalized log-likelihood

j ( ) = `( ) ? (c2 ? c1 )
X Zc
p 2
( ) ]2 ds
fj s
00
(8)
j
j =1 c1

For su ciently large j , the smoothness penalty forces each fj to be smooth
so that the global maximum does not occur for j (ak ) ! 0. We assume that f
2

is a parametrized curve over a nite interval, and the values c1 and c2 are the
endpoints of the smallest interval containing the support of gS (s). (See Remark
A below for more details). The corresponding Q function for the application of
the EM algorithm is

7

. . .
4

4

4
. . .. . . .. . . ..
.. . . . .. . . . .. . . .
... .
. . . . .... .
. . . . .... .
. . . .
3

3

3
. .
. . . . .
. . . . .
. . .
. . .. . . . . .. . . . . .. . .
. . . . . .
. . .. . . .. . . ..
2

2

2
. . . . . . .
. .. . . . .... . .. .
. . . .... . .. .
. . . ....
oo oo o
oo o oo oo o
o oo o o
.
• • • • • •••••• .
. o o o oo oo
oo
oo . .
. ooo
oo
oo o
o . .
1

1

1
•.••• ••••••• ••••• • • ••• •••• • • ••••••••.• •• • • •
•. •••• • • • • .. oo .. o o
oo
y2

y2

y2
.. . . .. . . .. . .
. .
. . . .
. . . .
. .
. . . . . . . . .
0

0

0
. .. . . .
. . . .. . . .
. . . . .. . .
. .
. . . . . . . . .
. .. . .. . ..
-1

-1

-1
. . .
.. . . .
.. . . .
.. . .
. .. . .. . .. .
. . .
... . . . ... . . . .. .
. . .
-2

-2

-2
. . .

-2 0 2 4 -2 0 2 4 -2 0 2 4

y1 y1 y1

. . .
4

4

4
. . .. . . .. . . ..
.. . . . .. . . . .. . . .
... .
. . . . .... .
. . . . .... .
. . . .
3

3

3
. . . . . . . . . o .
.oo . .
. . .. . . .
. . . .. . . .
. . . .. o ooo . .
.
. . o . o
. . .. . . .. .oo o. . .
2

2

oooo .
o oo 2 o
. o . oo . .
. . . . ....
. .o o oo
o . .... o . . ....
. ... . o oo oo oo
o oo o . ... . ooo oo o . . .. . o o o
ooo
oo
oo
ooooo . . o oo
oo . .
. o o
o . .
1

1

1

.. o
o
o
o o
oo
o
oo .. o o
o .. o o
o o
y2

y2

y2

o o . . o o . .
.. .. . .. o . . .o
o o .
o
o
. . . . o . .
. . . . . . . . .
0

0

0

. . .. . .
. . . . .. . .
. . . . .. . .
. .
. . . . . . . . .
. .. . .. . ..
-1

-1

-1

. . .
.. . . .. . . .. . .
. .. .
. . . .. .
. . . .. .
. .
... . . . ... . . . ... . . .
-2

-2

-2

. . .

-2 0 2 4 -2 0 2 4 -2 0 2 4

y1 y1 y1

. . .
4

4

4

. . .. . . .. . . ..
. .. . . . .. . . . .. . . .
... . .
. . . .... .
. . . . .... .
. . . .
3

3

3

oo
oooo .
o oo
oo
.
. .. o . o o
. .
. .
. o oo o .
. . . oo . o o . . . o .
. .
. ..o
o oooo oo .. .
o oo .
.
o
. o
o . . . o o . .
. o . o . . .
o
. . . . . . . . ..
2

2

2

o . o .
. o . . o . o . . o . o .
. .. . o . ... . .. . o . ... . .. . o o. . .
..
. o
o o . .
o . o o . .
o . o o
o
o. . o. . o. .
1

1

1

.. oo .. o o . .. o
y2

y2

y2

.o o . . . o. . o. .
o
.o .o .o
oo .o .
. o.
o o .o .
. o.
o . o .
. o.
o
. o.o o. . o .
o oo . . o .
o o
o .
0

0

0

. o. . o
o. .
. . .. . . . . .. . . . . .. . . .
. . . . . . . . .
. .. . .. . ..
-1

-1

-1

. . .
.. . . .. . . .. . .
. .. . . .. . . .. .
. . .
... . . . ... . . . ... . . .
-2

-2

-2

. . .

-2 0 2 4 -2 80 2 4 -2 0 2 4

y1 y1 y1

Figure 3: Circle approximations: dots are the data points, circles are the suc-
cessive approximations using the EM procedure (7).

Q( j 0) =
XX
n n
wik log gY jS (yi j (ak )) +
XX
n n
wik log vk ? (c2 ? c1 )
X Zc
p 2
( ) ]2 ds
fj s
00
j
i=1 k=1 i=1 k=1 j =1 c1

Pi
The solutions are easily found in closed form. The last two equations in (7) give
2
^ Pi
^j and vk ; let bk = n=1 wik , D be a diagonal matrix with entries b1 ; b2 ; : : : bn
and yj be an n-vector with kth component n=1 wik yij . Then

^ = (D + (c2 ? c1 )
fj j Kj )
?1 DfD?1 yj g (9)
The matrix Kj is the usual quadratic penalty matrix associated with a cubic
smoothing spline (see e.g. Hastie and Tibshirani, 1990 sec. 2.10). Equation (9)
says that f^j is obtained by applying a weighted cubic smoothing to the quantity
Pn=1 wik yij = Pn=1 wik which is exactly
D?1 yj ; this latter quantity has elements i i
the estimate that appears in the (unregularized) version (7).
The only remaining obstacle is how to nd the maximizing values a1 ; : : : an ,
^ ^
the location of the support points. We use a Newton-Raphson procedure for
this. Note however that full maximization of Q( j 0 ) with respect to f , j , 2

vk and ak would involve iteration between (9), the last two equations in (7)
and the Newton-Raphson procedure for the ak 's, and this is computationally
unattractive. We therefore seek only to increase Q( j 0 ) at each iteration: this
is called a generalized EM algorithm by Dempster el al. (1977) and the theory
of that paper and Wu (1983) guarantees its convergence to a xed point of the
gradient equations. To ensure an increase in Q( j 0 ), we apply (7) together
with one Newton-Raphson step for the ak 's, using step size halving if necessary.
We summarize the algorithm below.

A new principal curve algorithm
a) Start with ^(s) = E(Y)+ ds where d is the rst eigenvector of the covariance
f
matrix of Y, vk = 1=n. and S = a ; a ; : : : an the projected values onto
^ ^ ^ ^ 1 2
this line.
b) Compute
^
wik gY jS (yi jak ; )
Y gY (yh )vk
h6=i
under the Gaussian assumption (6) and normalize so that n=1 wik = 1.k ^
P
c) Fix ^; ^j ; and vk and apply a Newton-Raphson step to obtain a new set of
f 2 ^
support points a1 ; a2 ; : : : an . Reparametrize to arc-length parametrization.
^ ^ ^

9

o o
0 1 2 3 4

0 1 2 3 4
o o o ooo o o o o o ooo o o
o o o oo o o o
oo o o o o oo o o o
oo o
o oo oo o o
o o ooo o o oo oo o o
o o ooo o
o o o oo o o o oo
oo
oo o o ooo
o oo
oo o o ooo
o
o o ooo o oo ooo
y2

y2
o o o
o o o oo o o o oo
o oo o
oo o o o o o oo o
oo o o o o
o o o
o o o o
o
ooo
o o oo
oo ooo
o o oo
oo
oo o
o o o oo o
o o o
-2

-2
o o

-2 0 2 4 -2 0 2 4

y1 y1

Figure 4: On the left is the result of the new principal curve procedure, applied
to the circle example. On the right is the result of the HS principal curve
procedure

d) For k = 1; 2; : : : n compute
^ ?1 f ?1 y g
Pn
fj = (D + j Kj ) D D
^?
j

^j (ak ) =
2 i=1 ^ P
wik (yij
n
^
fj (ak ))2
; j = 1:::p
1X^
i=1 wik
n
^
vk = n wik
i=1

e) Iterate steps (b), (c), and (d) until the change in the log-likelihood is less
than some threshold.
Step (b) is the E step of the EM algorithm, while (c) and (d) are the M step.
Figure 4 shows, for the earlier circle example, the result of this procedure
(left panel) and the HS principal curve procedure (right panel). Notice that
the HS estimate extends further in the bottom arms of the data cloud. This is
to be expected: under model (2) specifying bivariate normal errors around the
generating curve, there is insu cient data to determine the bottom arms of the
generating curve.

10

REMARK A. The distribution gS (s) is only meaningful relative to some
xed parametrization of f (s). As HS (Sec. 5.5) point out, it is possible to nd a
R f (s)]2 is arbitrarily small, and such a curve can visit every data point. To
curve de ned over an arbitrarily large interval such that the curvature measure
00

j
avoid this anomaly, HS restrict the curve to be de ned over the unit interval,
as in (3). However they do not say which parametrization they use over the
unit interval. We include the additional multiplier c2 ? c1 to account for the
length of f (s); the e ect of this is to make the penalty term invariant to scale
changes s ! cs. After each iteration, the s1 ; s2 ; : : : sn values are rescaled so that
they correspond to arc-length along the current curve. As pointed out by Trevor
Hastie (personal communication) a unit-speed parametrized curve (kf 0 (s)k = 1),
cannot be the minimizer of Q( j 0 ) since a curve with cubic spline coordinate
functions does not have unit speed. Thus we are not able to make explicit the
parametrization that is being used here.

REMARK B. We choose the parameters j via a xed degrees of freedom"
strategy (see Hastie and Tibshirani, 1990 chap. 3). An automatic choice via
cross-validation or generalized cross-validation might be useful in practice and is
worthy of investigation. Most smoothers allow an automatic choice of smooth-
ing parameter; one simple approach then would be to allow each smoother
to choose its smoothing parameter in each coordinate function smoothing op-
eration. However HS demonstrate that this results in a nearly interpolating
curve after a number of iterations; they speculate that this may be due to au-
tocorrelation in the errors. Altman (1990) studies this problem in the usual
nonparametric regression situation.

REMARK C. The weights (5) and estimates (7) are dependent on the as-
sumed Gaussian form of the distribution of Y given s. To make the estimates
more robust to outliers one could use a resistant form for the weighted mean
and variance in (7).

REMARK D. If the variables Y1 ; Y2 ; : : : Yp are measured in di erent units
it would be appropriate, as in principal components analysis, to standardize
each variable rst. If they are measured in the same units, no standardization
is needed. In that instance, the procedure presented here allows for di er-
ent variances in the di erent dimensions. The HS procedure seems to assume
equal variances: however di erent variances could be facilitated by using a non-
orthogonal projection in step (b) of their algorithm (see section 1)

11

. .
.. . ..
5

5
.......... . ........... .
. . . . .. ..... .
... . . . . . .. ..... .
... .
outside assay

outside assay
4

4
. . . ... .. .
... ....... .... . . . . ... .. .
... ........ .... .
. . . . . .. ... .... .. . . . . . . .. ... .... .. .
. . .. ... .................. . . . . . . . . .. ... .......... ...... . . . . . .
3

3
. . .. .. . . .. ..
.. .
. . . ............ . . . . ............ ..
.
.. ..... .... .
2

2
. . .. ........... .. ..
.. . . .. .. .... . .
... . ... .
. . . .... ..... ... .. . . . . . . . ......... ... .. . . . .
. . . . ... . .
. .. . . . . . ... . .
. .. .
1

1
.. . . .
. . ... . . .. . . .
. . . .. . .
0

0
0 1 2 3 4 0 1 2 3 4

inhouse assay inhouse assay

Figure 5: Left panel shows the gold pairs data and principal curve for the HS
procedure. Right panel shows the gold pairs data and principal curve from the
new principal curve procedure.

5 An Example
HS analyzed some data on the gold content of computer chip waste produced by
two laboratories. Two hundred and forty nine samples were each split in two,
and assays were carried out by an inhouse lab and an outside lab to determine
the gold content. The objective was to investigate which laboratory produces
lower or higher gold content measurements. A full description in given in HS
(section 7.2). (HS report 250 pairs, but we received only 249 from Trevor
Hastie). The assumed model has the form Yj = fj (s) + j , j = 1; 2, with
f2 (s) constrained to be the identity function. The only change needed in the
estimation procedure is that the smooth of Y2 versus s is not needed, and
replaced by the identity function.
A scatterplot of the (log) measurements is shown in Figure (5) along with
the estimated principal curve from the HS algorithm.
As HS report, the outside assay seems to be producing systematically higher
gold levels than the inside assay, in the interval 1.4 to 4. However the curve es-
timated from the procedure of this paper (right panel) suggests that the outside
lab gives higher gold content below 1. In this example, then, it seems that the
choice of underlying model has an e ect on the substantive conclusions.

12

6 Relationship between the HS and new princi-
pal curve algorithms
In this section we explore in detail the relationship between the HS and new
algorithm for principal curves, in the data case.
The HS principal curve procedure, when applied to a dataset, uses a non-
parametric regression estimate in place of the conditional expectation E(Yj js).
One nonparametric regression estimate used by HS is a cubic smoothing spline,
resulting from the penalized least squares criterion (3).
The penalized form of the mixture likelihood, on the other hand, leads to
the estimate given by (9). This di ers from a cubic smoothing spline applied to
Pi
the observations yij in that a) the weights wik are rst applied to the yij , and
b) a weighted cubic smoothing spline is used, with weights n=1 wik . Hence
the relationship between the two approaches hinges on the form of the weights.
Consider then estimation of f (s) at a point s = s0 . We assume that s
measures arc-length along the curve f (s). The weight in the new principal
curve procedure is proportional to the Euclidean distance between y and f (s0 )
(at least when the j s are equal). Now suppose Y projects to the curve f at
S = s. The Euclidean distance between y and f (s0 ) is not in general the same
as (s ? s0 )2 . However, when f (s) is a straight line, a simple application of the
Pythagorean theorem shows that shows that the two distances are proportional
to one another.
The left panel of Figure 6 illustrates this. It shows the contours of constant
weight for estimation of f (s0 ) (indicated by T") and these are also the lines
of constant s value. Hence the weighting operation averages the yij 's by their
s value, and can be expressed as a kernel smoothing of the data. The cubic
smoothing spline is then applied to the averaged values yj , and since a cu-
bic smoothing spline is approximately a (variable) kernel smoother (Silverman,
1984), the result is the approximately the convolution of two kernel smoothers
applied to the data, which is just another kernel smoother. Hence when f (s) is
a straight line, the HS and new procedures will give similar results.
The story is di erent however when f (s) is curved, as in the right panel of
Figure 6.
The contours are curved in the opposite direction to the curvature of f (s).
In both examples in Figure 6, we have assumed 1 (ak ) = 2 (ak )= constant.
If 1 6= 2 , the contours are tilted.

7 Discussion
We have presented an alternative de nition of principal curves based on a mix-
ture model. Although there are theoretical reasons for preferring this de nition
to that given by HS, we have no evidence that the resulting procedure estimation
procedure works any better in practice.

13

0 1 2 3 4

0.010.0150.02 0.004
0.006
0.005
1 2 3 4 5 6 0.016
0.014
T T 0.012
0.01

0.020.0150.01 0.005 0.008
-2

0.008 0.006 0.004

-2 -1 0 1 2 3 4 1 2 3 4 5 6

Figure 6: Contours of constant weight for computing the estimate at the target
point T". On the left the underlying function (broken line) is straight, while
on the right it is curved.

14

We have focussed on one-dimensional curves, for which s is real-valued. Ex-
tensions to globally parametrized surfaces (vector-valued s) seem straightfor-
ward in principle but we have not studied the details. A Bayesian analysis of
the model might also prove to be interesting. The Gibbs sampler (Geman and
Geman 1984, Tanner and Wong 1987, Gelfond and Smith 1990) could be used
to simulate from the posterior distribution. Verdinelli and Wasserman (1990)
carry this out for a simpler normal mixture problem and show the structure of
the mixture model lends itself naturally to the Gibbs sampler.

ACKNOWLEDGMENTS
This work was stimulated by questions raised by Geo Hinton and Radford
Neal, and I thank them for their questions and some fruitful discussions. I would
also like to thank Mike Leblanc, Mary L'Esperance, Trevor Hastie and two
referees for helpful comments, and Trevor for a copy of his program for tting
principal curves. Support by the Natural Sciences and Engineering Research
Council of Canada is gratefully acknowledged.

References
Altman, N.S. (1990). Kernel smoothing of data with autocorrelated errors. J.
Amer. Statist. Assoc., 85:749{759.
Dempster, A.P., Laird, N.M., and Rubin, D.B (1977). Maximum likelihood from
incomplete data via the EM algorithm (with discussion). J. R. Statist. Soc.
B, 39:1{38.
Gelfand, A. E. and Smith. A. F. M. (1990). Sampling based approaches to
calculating marginal densities. J. Amer. Statist. Assoc, 85:398{409.
Geman, S. and Geman, D. (1984). Stochastic relaxation, gibbs distributions and
the bayesian restoration of images. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 6:721{741.
Hastie, T. (1984). Principal curves and surfaces. Technical report, Stanford
University.
Hastie, T. and Stuetzle, W. (1989). Principal curves. Journal of the American
Statistical Association, 84(406):502{516.
Hastie, T. and Tibshirani, R. (1990). Generalized Additive Models. Chapman
and Hall.

15

Lindsay, B. G. (1983). The geometry of mixture likelihoods: a general theory.
Annal of Statistics, 11:86{94.
Silverman, B.W. (1984). Spline smoothing: the equivalent kernel method. Annal
of Statistics, 12:898{9164.
Tanner, M. and Wong, W. (1987). The calculation of posterior distributions by
data augmentation (with discussion). J. Amer. Statist. Assoc., 82:528{550.
Titterington, D.M., Smith, A.F.M., and Makov, U.E. (1985). Statistical analysis
of nite mixture distributions. Wiley, New York.
Verdinelli I, and Wasserman L. (1990). Bayesian analysis of outlier problems,
using the gibbs sampler. Technical report, Carnegie-Mellon niversity.
Wu, C.F.J. (1983). On the convergence properties of the em algorithm. Annal
of Statistics, 11:95{103.

16

Next2

More Related Content

What's hot

Viewers also liked

Similar to Next2

Next2