Sampling Signals
Overview:
¯ We use the Fourier transform to understand the discrete sampling and re-sampling of signals.
¯ One key question is when does sampling or re-sampling provide an adequate representation of the original signal?
Terminology:
¯ sampling – creating a discrete signal from a continuous process.
¯ downsampling (decimation) – subsampling a discrete signal
¯ upsampling – introducing zeros between samples to create a longer signal
¯ aliasing – when sampling or downsampling, two signals have same sampled representation but differ between sample locations.
Matlab Tutorials: samplingTutorial.m, upSample.m
320: Sampling Signals c A.D. Jepson and D.J. Fleet, 2005 Page: 1 Sampling Signals
I(x) continuous signal
x
I[n] ● ● ● ● ● sampled ● ● ● ● ● ● ● ● ● signal ● ● ● ● n
g[m] downsampling ● ● ● ● ● ● ● I[n] 2 g[m] ● ● m
f[n] ● ● upsampling ● ● ● ● ● ● ● g[m] 2 f[n] ● ● ● ● ● ● ● ● ● n
320: Sampling Signals Page: 2
Up-Sampling
[Ò] Æ
Consider up-sampling a signal Á of length :
¯ Æ Ò
Increase number of samples by a factor of × .
¯ Ò ½ Á [Ò] f [Ò] ¼
Step 1. Place × zeros after every sample of , to form
Ò Æ
of length × , namely
´
Á [Ò=Ò ] Ò ÑÓ d Ò = ¼; ×
× for
f [Ò] =
¼ (1)
¼ otherwise.
Signal Raw Upsampled Signal 1.5 2
1.5 1
1 0.5 0.5 0 0 f 0 I[n] −0.5 −0.5 −1
−1.5 −1
−2 −1.5 0 5 10 15 0 5 10 15 20 25 30 n n
320: Sampling Signals Page: 3
Up-Sampling (Cont.)
¯ f
Step 2. Interpolate signal ¼ :
f = Ë £ f ; Ë [Ò]:
¼ for some smoothing kernel (2)
f [Ò] f [Ò] Ë [¼] = ½
Here, in order for to interpolate ¼ , we require that
Ë [jÒ ] = ¼ j
and × for nonzero integers .
Sinc Filtered Upsampled Signal (b) 1.5
1
0.5
0 f[n]
−0.5
−1
−1.5 0 5 10 15 20 25 30 n
¯ Many smoothing kernels can be used (see upSample.m).
320: Sampling Signals Page: 4
Frequency Analysis of Up-Sampling
f [Ò]
Step 1. Fourier transform of the raw up-sampled signal ¼ :
ÆÒ ½
Æ ½
×
X X
× ×
Ò i! jÒ i!
×
k k
F ´f [Ò]µ[k ] = f [Ò]e f [jÒ ]e
¼ ¼ ¼ ×
Ò=¼
j =¼
Æ ½
X
×
i! jÒ
×
k
= Á [j ]e
j =¼
¾
×
k =
Here ! , so the last line above becomes
k
ÆÒ
×
Æ ½
X
¾
ÆÒ ÆÒ
× ×
kj i
Æ
F ´Á µ[k ]; F ´f µ[k ] = Á [j ]e k < :
¼ for (3)
¾ ¾
j =¼
F ´Á µ Æ F ´f µ
Since is -periodic, equation (3) implies that ¼ consists of
Ò F ´Á µ
× copies of concatenated together.
Amplitude Spectrum of I[n] Amplitude Spectrum of Raw Upsampled Signal 6 6
5 5
4 4
3 3 |FT| |FT|
2 2
1 1
0 0 −5 0 5 −15 −10 −5 0 5 10 15 k k
320: Sampling Signals Page: 5
Fourier Analysis of Up-Sampling Step 2
f [Ò] = Ë £ f Ë
Recall Step 2 is to form ¼ , for some interpolation filter .
However, notice from the inverse Fourier transform that (for Æ even)
Æ=¾ ½
X
¾
½
kj i
^
Æ
Á [j ] = Á [k ]e
Æ
k = Æ=¾
Æ=¾ ½
X
¾
Ò
×
i kj Ò
×
^
ÆÒ
×
f [k ]e = f [jÒ ]: = ¼
× (4)
ÆÒ
×
k = Æ=¾
f ´jÒ µ = Á ´j µ
Here we used × in the last line. Notice the left term in the
^
Ò B [k ]f [k ] ¼
last line above is × times the inverse Fourier transform of
B [k ] = ½ Æ=¾ k < Æ=¾
where B is the box function, for and
[k ] = ¼
B otherwise. We can therefore evaluate this inverse Fourier
Ò jÒ
transform at every pixel , and not just at the interpolation values × ,
[Ò]
to construct a possible interpolating function f
½
^
f [Ò] = Ò F ´B ´k µf [k ]µ: ¼ × (5)
Sinc Filtered Upsampled Signal (b) Amp. Spec. (Upsampled Sig.(r), Sinc Filter(g), Filtered Sig.(b) 1.5 12
1 10
0.5 8
0 6 f[n] |FT|
4 −0.5
2 −1
0 −1.5 −15 −10 −5 0 5 10 15 0 5 10 15 20 25 30 k n 320: Sampling Signals Page: 6
Down-Sampling
[Ò] Æ
Consider down-sampling a signal Á of length :
¯ Æ Ò Ò × Reduce number of samples by a factor of × , where is a divi-
sor of Æ .
¯ Define the comb function:
´Æ=Ò µ ½
×
X
Æ C ´Ò; Ò µ =
Ò;ÑÒ ×
×
Ñ=¼
I [n] ns ns = 4 ●●●●●●●●●●●●●●●●●
n
Á [Ò]
¯ Step 1. Introduce zeros in at unwanted samples.
g [Ò] = C [Ò; Ò ]Á [Ò]: ×
¼ (6)
¯ g
Step 2. Downsample signal ¼ :
g [Ñ] = g [ÑÒ ]; ¼ Ñ < Æ=Ò :
× × ¼ for (7)
320: Sampling Signals Page: 7 Frequency Domain Analysis of Down-Sampling
Proposition 1. The Fourier transform of the comb function is another
comb function:
Æ
F ´C [Ò; Ò ]µ = C [k ; Æ=Ò ]: ×
× (8)
Ò ×
∧ ω C( ) 2π n = 4 ns s
−π 0 π ω
¾
k ! = k
Recall the frequency ! and wave number are related by .So
Æ
¾ Æ ¾
! = =
the spacing in the plot above is × .
Æ Ò Ò
× ×
Proposition 2. Pointwise product and convolution of Fourier trans-
[Ò] g [Ò] Æ forms. Suppose f and are two signals of length (extended to
be Æ -periodic). Then
½
F ´f µ £F´g µ F ´f [Ò]g [Ò]µ =
Æ
Æ=¾ ½
X
½
^
f [j ] g^[k j ]:
(9)
Æ
j = Æ=¾
^
g^ f g where f and denote the Fourier transforms of and , respectively.
The proofs of these two propositions are straight forward applications of the definition of the Fourier transform given in the preceeding notes, and are left as exercises.
320: Sampling Signals Page: 8
Fourier Analysis of Down-Sampling Step 1
g [Ò] = C [Ò; Ò ]Á [Ò] × Recall Step 1 is to form ¼ . By Prop. 1 and 2 above,
we have
F ´g µ = F ´C [Ò; Ò ]Á [Ò]µ
¼ ×
½ Æ
^
C [¡; Æ=Ò ] £ Á [¡] =
×
Ò Æ
×
Æ=¾
X
½
^
C [j ; Æ=Ò ]Á [k j ]
×
Ò
×
j = Æ=¾·½
Ò Ö
×
¼
X
½ Æ
^
= Á [k Ö ;
] (10)
Ò Ò
× ×
Ö = Ö ·½
¼
^
Á [k ] Ö
where, due to the periodicity of , we can use any integer ¼ (eg.
Ö = Ò =¾ Ò
× × ¼ for even ).
∧ ∧ I(ω) g (ω) 2π 0 n s ns = 4
−π 0 π ω −π 0 π ω
^
g^ [k ] Á [k ÖÆ=Ò ] ¼
Therefore consists of the sum of replicas × of the
Á Æ=Ò
Fourier transform of the original signal , spaced by wavenumber ×
¾
! =
or, equivalently, by frequency × .
Ò
×
g^ ¾
Note the Fourier transform ¼ has period , so the contribution sketched
¾ above for ! can be shifted to the left.
320: Sampling Signals Page: 9
Fourier Analysis of Down-Sampling Step 2
g [Ò]
In Step 2 we simply drop the samples from ¼ which were set to zero
C [Ò; Ò ]
by the comb function × . That is
g [Ñ] = g [ÑÒ ]; ¼ Ñ < Æ=Ò :
× × ¼ for
In terms of the Fourier transform, it is easy to show
F ´g µ[k ] = g^[k ] = g^ [k ]; Æ=´¾Ò µ k < Æ=´¾Ò µ:
× ×
¼ for (11)
¾
! = k g [Ñ]
Rewriting this in terms of the frequency ×;k (note is
´Æ=Ò µ
×
¾
Æ=Ò ! = k k
a signal of length × ), and the corresponding frequency
Æ
g [Ò] = ! Ò
¼ ×
×;k of the longer signal ,wehave
g^[! ] = g^ [! ] = g^ [! =Ò ]; ! < :
¼ k ¼ ×;k × ×;k ×;k for (12)
∧ ω f( ) ns = 4 original signal −π ω π ω 0 s
∧ g(ω) downsampled signal
−π 0 π 4π ω
∧ ω f( ) upsampled and filtered signal −π 0 π ω
320: Sampling Signals Page: 10
Nyquist Sampling Theorem
[Ò]
Sampling Theorem: Let f be a band-limited signal such that
^
f [! ] = ¼ j! j > !
for all ¼
! f [Ò] g [Ñ] =
for some ¼ . Then is uniquely determined by its samples
f [ÑÒ ]
× when
¼
> ! ! =¾ = Ò <
¼ ×
× or equivalently
Ò ¾
×
= ¾=! ¼ where ¼ . In words, the distance between samples must be
smaller than half a wavelength of the highest frequency in the signal. !
In terms of the previous figure, note that the maximum frequency ¼ ! must be smaller than one-half of the spacing, × , between the replicas
introduced by the sampling. This ensures the replicas do not overlap.
[Ñ]
When the replicas do not overlap, we can up-sample the signal g
[Ò]
and interpolate it to recover the signal f , as discussed above.
g [k ]
Otherwise, when the replicas overlap, the Fourier transform ^ con-
^
[k ]
tains contributions from more than one replica of f . Due to these
[Ò] aliased contributions, we cannot then recover the original signal f .
320: Sampling Signals Page: 11
Sampling Continuous Signals
´Üµ Ü ¾ [¼;ĵ A similar theorem holds for sampling signals f for .We
can represent f as the Fourier series
½
X
i! Ü
^
k
f [k ]e ; f ´Üµ =
a:e:
k = ½
¾
! = k = a:e:
where k and denotes equals almost everywhere. Suppose
Ä
f ´Üµ ! > ¼
is band-limited so that for some ¼
^
j! j > ! : f [k ] = ¼ ¼
for all k
´Üµ Á [Ñ] = f ´Ñµ
Then f is uniquely determined by its samples when
¼ <
(13)
¾
= ¾=! ¼ where ¼ . In words, the distance between samples must be
smaller than half the wavelength of the highest frequency in the signal.
´Üµ The link with the preceeding analysis is that sampling f withasam-
ple spacing of causes replicas in the Fourier transform to appear with
! = ¾=
spacing × . As before, the condition that these replicas do not
! < ! =¾ × overlap is ¼ , which is equivalent to condition (13).
320: Sampling Signals Page: 12 Aliasing
Aliasing occurs when replicas overlap:
Consider a perspective image of an infinite checkerboard. The signal is dominated by high frequencies in the image near the horizon. Properly designed cameras blur the signal before sampling, using
¯ the point spread function due to diffraction,
¯ imperfect focus,
¯ averaging the signal over each CCD element. These operations attenuate high frequency components in the signal. Without this (physical) preprocessing, the sampled image can be severely aliased (corrupted):
320: Sampling Signals Page: 13 Dimensionality A guiding principal throughout signal transforms, sampling, and alias- ing is the underlying dimension of the signal, that is, the number of linearly independent degress of freedom (dof). This helps clarify many
issues that might otherwise appear mysterious.
Æ Æ ¯ Real-valued signals with samples have dof. We need a basis
of dimension Æ to represent them uniquely.
Æ Æ ¯ Why did the DFT of a signal of length use sinusoids? Be-
cause Æ sinusoids are linearly independent, providing a minimal Æ spanning set for signals of length Æ . We need no more than .
¯ But wait: Fourier coefficients are complex-valued, and therefore Æ have ¾ dofs. This matches the dof needed for complex signals of
length Æ but not real-valued signals. For real signals the Fourier spectra are symmetric, so we keep half of the coefficients.
¯ When we down-sample a signal by a factor of two we are moving ¾
to a basis with Æ= dimensions. The Nyquist theorem says that ¾ the original signal should lie in an Æ= dimensional space before you down-sample. Otherwise information is corrupted (i.e. signal
structure in multiple dimensions of the original Æ -D space appear ¾ thesameintheÆ= -D space).
¯ The Nyquist theorem is not primarily about highest frequencies and bandwidth. The issue is really one of having a model for the signal; that is, how many non-zero frequency components are in the signal (i.e., the dofs), and which frequencies are they.
320: Sampling Signals Page: 14