J Westerweel
Figure 1. The displacement of the tracer particles is an
approximation of the fluid velocity (after Adrian 1995).
average velocity along the trajectory over a time 1t. This
is illustrated in figure 1.
Thus, D cannot lead to an exact representation of u,
but approximates it within a finite error ε:
kD − u · 1tk <ε. (2)
The associated error is often negligible, provided that the
spatial and temporal scales of the flow are large with respect
to the spatial resolution and the exposure time delay, and
the dynamics of the tracer particles. A further analysis of
these aspects is given by Adrian (1995).
The flow information is only obtained from the
locations at which the tracer particles are present.
Since these are distributed randomly over the flow, the
displacement of individual tracer particles constitutes a
random sampling of the displacement field, and different
realizations yield different estimates of D. Obviously,
these differences can be neglected as long as the
reconstructed displacement field satisfies equation (2). This
implies that the displacement field should be sampled
at a density that matches the smallest length scale of
the spatial variations in D. Since D can be regarded
as a low-pass filtered representation of u, with a cut-
off filter length that is equal to kDk, the displacement
field should be sampled with an average distance that is
smaller than the particle displacement. This implies that
a measurement in which the average distance between
distinct particle images is larger than the displacement
(as is the case in conventional particle tracking; see
figure 2(a)) cannot resolve the full displacement field.
However, when the seeding concentration is high (so that
the mean spacing between tracer particles is smaller than
the displacement) it is not possible to identify matching
particle pairs unambiguously; see figure 2(b). It is therefore
(a) individual tracer (b) tracer pattern
Figure 2. (a) At low seeding density individual tracers yield the fluid motion; (b) at high seeding density the tracers constitute
a pattern that is advected by the flow.
more convenient to describe the tracer particles in terms of
a pattern.
2.2. The tracer pattern
The tracer particles constitute a random pattern that is ‘tied’
to the fluid and the fluid motion is visible through changes
of the tracer pattern. The tracer pattern in X at time t is
defined as:
G(X,t) =
N
X
i=1
δ[X −X
i
(t)] (3)
where N is the total number of particles in the flow, δ(X) is
the Dirac δ-function and X
i
(t) the position vector of the
particle with index i at time t. Integration of G(X ,t) over
a volume yields the number of particles in that volume.
The tracer pattern at time t
0
can be viewed as a spatial
signal G
0
(X) = G(X,t
0
) at the input of a ‘black-box’
system (representing the flow) that acts on the input signal,
and returns a new signal G
00
(X) = G(X ,t
00
) at the output;
see figure 3. For ideal tracer particles the addition of a new
particle does not affect the action of the system on the other
tracer particles, i.e. the system is linear. Consequently, the
output signal can be written as a convolution of the input
signal with the impulse response H of the system:
G
00
(X) =
Z
H(X,X
0
)G
0
(X
0
) dX
0
. (4)
The impulse response is a shift of the input by the local
displacement D in equation (1):
H(X
0
,X
00
) = δ[X
00
− X
0
− D]. (5)
The shift formally depends on X, but under equation (2) it
can be assumed that D is locally uniform, so that H can be
regarded as shift invariant, i.e. H(X
0
,X
00
) = H(X
00
−X
0
).
According to linear system theory, the impulse response
of a black-box system can be obtained from the cross-
covariance R
G
0
G
00
of a random input signal with the
corresponding output signal:
R
G
0
G
00
(s) = H ∗ R
G
0
(s) (6)
(Priestley 1992), where ∗ denotes a convolution integral,
and R
G
0
is the auto-covariance of the input signal. For the
special case where the input signal is a homogeneous white
process (i.e. R
G
0
(s) ∝ δ(s)), the cross-correlation directly
yields the impulse response.
1380