The operation of downsampling by factor
M∈ℕM
describes the process of keeping every
M
t
h
M
t
h
sample and discarding the rest. This is denoted by "
↓M
↓
M
" in block diagrams, as in Figure 1.
Formally, downsampling can be written as
yn=xnM
y
n
x
n
M
In the zz domain,
Yz=∑nynz-n=∑nxnMz-n=∑mxm1M∑p=0M-1ⅇⅈ2πMpmz-mM
Y
z
n
n
y
n
z
n
n
n
x
n
M
z
n
m
m
x
m
1
M
p
0
M
1
2
M
p
m
z
m
M
(1)
where
∑p=0M-1ⅇⅈ2πMpm=1ifm is a multiple of M0otherwise
p
0
M
1
2
M
p
m
1
m is a multiple of M
0
Yz=1M∑p=0M-1∑mxmⅇ-ⅈ2πMpz1M-m=1M∑p=0M-1Xⅇ-ⅈ2πMpz1M
Y
z
1
M
p
0
M
1
m
m
x
m
2
M
p
z
1
M
m
1
M
p
0
M
1
X
2
M
p
z
1
M
(2)
Translating to the frequency domain,
Yⅇⅈω=1M∑p=0M-1Xⅇⅈω-2πpM
Y
ω
1
M
p
0
M
1
X
ω
2
p
M
(3)
As shown in Figure 2, downsampling expands each
2π2
-periodic repetition of
Xⅇⅈω
X
ω
by a factor of MM along the
ωω axis, and reduces the gain
by a factor of MM. If
xm
xm is not bandlimited to
πMM,
aliasing may result from spectral overlap.
When performing a frequency-domain analysis
of systems with up/downsamplers, it is strongly recommended to
carry out the analysis in the z
z-domain until the last step, as done above. Working
directly in the
ⅇⅈω
ω
-domain can easily lead to errors.