5: Gibbs

⊲ Phenomenon
Discontinuities
Discontinuous
Waveform
Gibbs Phenomenon
Integration
Rate at which
coefficients decrease
with m
Differentiation
Periodic Extension
t2 Periodic
Extension: Method
5: Gibbs Phenomenon
(a)
t2 Periodic
Extension: Method
(b)
Summary

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 1 / 11

 Discontinuities

5: Gibbs Phenomenon
⊲ Discontinuities
A function, v(t), has a discontinuity of amplitude b at t = a if
Discontinuous
Waveform lime→0 (v(a + e) − v(a − e)) = b 6= 0
Gibbs Phenomenon
Integration
Rate at which Conversely, v(t), is continuous at t = a if the limit, b, equals zero.
coefficients decrease
with m
Differentiation
Periodic Extension
t2 Periodic
Examples:
Extension: Method b b
(a)
u(t)

v(t)
t2 Periodic
0 0
Extension: Method
a–e a a+e a–e a a+e
(b) Time (t) Time (t)
Summary
Continuous Discontinuous

We will see that if a periodic function, v(t), is discontinuous, then its

Fourier series behaves in a strange way.

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 2 / 11

 Discontinuous Waveform

1
5: Gibbs Phenomenon
Discontinuities
Pulse: T = = 20, width= 12 T , height A = 1
F
1
0.5
Discontinuous 0
⊲ Waveform 1 0.5T
Ae−i2πmF t dt
R 0 5 10 15 20
Gibbs Phenomenon Um = T 0 1
max(u0)=0.500
Integration  −i2πmF t 0.5T 0.5
i
Rate at which
coefficients decrease
= 2πmF T e 0
0
0
N=0

5 10 15 20
with m
i −iπm

((−1)m −1)i
Differentiation = 2πm e −1 = 2πm
1
0.5
max(u1)=1.137

Periodic Extension  0
N=1

t2 Periodic
0 m 6= 0, even 0 5 10 15 20
Extension: Method

1
(a) max(u3)=1.100

t2 Periodic
= 0.5 m = 0 0.5
N=3
0
Extension: Method  −i

(b)
mπ m odd 0 5 10 15 20

1
Summary max(u5)=1.094
0.5
1
So, u(t) = 2 + π2 sin 2πF t + 31 sin
6πF t 0
0
N=5

5 10 15 20

+ 15 sin 10πF t + . . . 1
max(u41)=1.089
0.5
N=41
0
PN i2πmF t 0 5 10 15 20
Define: uN (t) = m=−N Um e 1
max(u41)=1.089
0.5
uN (0) = 0.5 ∀N 0
-1 -0.5 0 0.5 1
1 1 π sin t
R
maxt uN (t) −→ 2 + π 0 t dt ≈ 1.0895 [Enlarged View: u41 (t)]
N →∞

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 3 / 11

[Powers of −1 and i]

Expressions involving (−1)m or, less commonly, im arise quite frequently and it is worth becoming
familiar with them. They can arise in several guises:

m
e−iπm = eiπm = eiπ = cos (πm) = (−1)m
1 πm
 1 m
i2
e = ei 2 π = im
1 πm
m
−i 1

e−i 2
= e 2 π
= (−i)m

As m increases these expressions repeat with periods of 2 or 4. Simple expressions involving these
quantities make useful sequences.

m −4 −3 −2 −1 0 1 2 3 4
(−1)m = cos πm = eiπm 1 −1 1 −1 1 −1 1 −1 1
im = ei0.5πm 1 i −1 −i 1 i −1 −i 1
(−i)m = e−i0.5πm 1 −i −1 i 1 −i −1 i 1
1
2
(1 + (−1)m ) 1 0 1 0 1 0 1 0 1
1
2
(1 − (−1)m ) 0 1 0 1 0 1 0 1 0
1 m
2
(i + (−i)m ) = cos 0.5πm 1 0 −1 0 1 0 −1 0 1
1
4
(1 + (−1)m + im + (−i)m ) 1 0 0 0 1 0 0 0 1

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – note 1 of slide 3
 Gibbs Phenomenon

5: Gibbs Phenomenon PN
Discontinuities Truncated Fourier Series: uN (t) = m=−N Um ei2πmF t
Discontinuous
Waveform If u(t) has a discontinuity of height b at t = a then:
⊲ Gibbs Phenomenon
u(a−e)+u(a+e)
Integration
Rate at which
(1) uN (a) −→ lime→0 2
N →∞
coefficients decrease
with m
Differentiation
(2) uN (t) has an overshoot of about 9% of b at the discontinuity. For
Periodic Extension large N the overshoot moves closer to the discontinuity but does
t2 Periodic
Extension: Method not get smaller (Gibbs phenomenon). In the limit the overshoot
π
(a)
equals − 21 + π1 0 sint t dt b ≈ 0.0895b.
R

t2 Periodic
Extension: Method −1
(b) (3) For large m, the coefficients, Um decrease no faster than |m| .
Summary

Example:
1
uN (0) −→ 0.5 0.5
N →∞ 0
0 5 10 15 20

maxt uN (t) −→ 1.0895 . . . 1

N →∞ 0.5
max(u41)=1.089

 N=41
0

0
 m 6= 0, even 0 5 10 15 20

1
Um = 0.5 m = 0 0.5
max(u41)=1.089

0
 −i

mπ m odd -1 -0.5 0 0.5 1

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 4 / 11

[Origin of Gibbs Phenomenon]

This topic is included for interest but is not examinable.

Our first goal is to express the partial Fourier series, uN (t), in terms of the original signal, u(t). We
begin by substituting the integral expression for Un in the partial Fourier series
 R 
P+N i2πnF t
P+N 1 T τ
uN (t) = n=−N Un e = n=−N T 0 u(τ )e −i2πnF dτ ei2πnF t
Now we swap the order of the integration and the finite summation (OK if the integral converges ∀n)
P 
1
RT +N i2πnF (t−τ )
uN (t) = T 0 u(τ ) n=−N e dτ

Now apply the formula for the sum of a geometric progression with z = ei2πF (t−τ ) :
P+N n = z
−N
−z N +1 z −(N +0.5) −z N +0.5
n=−N z 1−z
= z −0.5 −z 0.5
i2π(N +0.5)F (τ −t)
−e−i2π(N +0.5)F (τ −t)
uN (t) = T1 0T u(τ ) e
R
dτ
ei2π0.5F (τ −t) −e−i2π0.5F (τ −t)
sin π(2N +1)F (τ −t)
= T1 0T u(τ )
R
sin πF (τ −t)
dτ
sin((N +0.5)x)
So if we define the Dirichlet Kernel to be DN (x) = sin 0.5x
, and set x = 2πF (τ − t), we obtain
uN (t) = T1 0T u(τ )DN (2πF (τ − t)) dτ
R

So what we have shown is that uN (t) can be obtained by multiplying u(τ ) by a time-shifted Dirichlet
Kernel and then integrating over one period. Next we will look at the properties of the Dirichlet Kernel.

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – note 1 of slide 4
[Dirichlet Kernel]

This topic is included for interest but is not examinable.

sin((N +0.5)x)
Dirichlet Kernel definition: DN (x) = +N
P inx = 1 + 2
PN
n=−N e n=1 cos nx = sin 0.5x
DN (x) is plotted below for N = {2, 5, 10, 21}. The vertical red lines at ±π mark one period.
D2(x) 10 D5(x) 20 D10(x) 40 D21(x)
4

D10(x)

D21(x)
D2(x)

D5(x)
2 5 10 20

0 0 0 0
-6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6 -6 -4 -2 0 2 4 6
x x x x

• Periodic: with period 2π

1
R +π
• Average value: hDN (x)i = 2π −π DN (x)dx = 1
π π π
• First Zeros: DN (x) = 0 at x = ± N +0.5 define the main lobe as − N +0.5 <x< N +0.5
• Peak value: 2N + 1 at x = 0. The main lobe gets narrower but higher as N increases.
• Main Lobe semi-integral:
π π
R N +0.5 R N +0.5 sin((N +0.5)x) 1
Rπ sin y
x=0 DN (x)dx = x=0 sin 0.5x
dx = N +0.5 y=0 sin y dy[y = (N + 0.5)x]
2N +1
where we substituted y = (N + 0.5)x. Now, for large N , we can approximate sin 2Ny+1 ≈ 2Ny+1 :
π
R N +0.5 1
Rπ sin y R π sin y
x=0 D N (x)dx ≈ N +0.5 y=0 y dy = 2 y=0 y dy ≈ 3.7038741 ≈ 2π × 0.58949
2N +1
We see that, for large enough N , the main lobe semi-integral is independent of N .
[In MATLAB DN (x) = (2N + 1) × diric(x, 2N + 1)]

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – note 2 of slide 4
[Gibbs Phenomenon Overshoot]

This topic is included for interest but is not examinable. 1

u41(t)
0.5 T=20

The partial Fourier Series, uN (t), can be obtained by multiplying u(t) 0

-1 -0.5 0 0.5 1 1.5 2
by a shifted Dirichlet Kernel and integrating over one period:
1
RT
uN (t) = T 0 u(τ )DN (2πF (τ − t)) dτ 80 t=0
60
40
For the special case when u(t) is a pulse of height 1 and width 0.5T : 20
0
-20

uN (t) = T1 00.5T DN (2πF (τ − t)) dτ

R -1 -0.5 0 0.5 1 1.5 2

80 t=0.24
60
40
Substitute x = 2πF (τ − t) 20
0
R πF T −2πF t R π−2πF t -20
1 1
uN (t) = 2πF T −2πF t
DN (x) dx = 2π −2πF t DN (x) dx -1 -0.5 0 0.5 1 1.5 2

80 t=0.96
60
• For t = 0 (theRblue integral and the blue circle on the upper graph): 40
1 π 20
uN (0) = 2π 0 DN (x) dx = 0.5
0
-20
-1 -0.5 0 0.5 1 1.5 2
T
• For t = (the red integral and the red circle on the upper graph):
2N +1
  π
R π− N +0.5 R0 π
R π− N +0.5
T 1 1 1
uN 2N +1 = 2π − π DN (x) dx= 2π − π DN (x) dx + 2π 0 DN (x) dx
N +0.5 N +0.5
For large N , we replace the first term by a constant (since it is the semi-integral of the main lobe)
and replace the upper limit of theR second term by π:
1 π
≈ 0.58949 + 2π 0 DN (x) dx = 1.08949
• For 0 ≪ t ≪ 0.5T , uN (t) ≈ 1 (the green integral and the green circle on the upper graph).

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – note 3 of slide 4
 Integration

P∞
5: Gibbs Phenomenon
Discontinuities
Suppose u(t) = m=−∞ Um ei2πmF t
Discontinuous
Waveform Define v(t) to be the integral of u(t) [boundedness requires U0 = 0]
Gibbs Phenomenon Rt R t P∞
⊲ Integration v(t) = u(τ )dτ = m=−∞ Um e
i2πmF τ
dτ
Rate at which
coefficients decrease P∞ R t i2πmF τ R P
with m = m=−∞ Um e dτ [assume OK to swap and ]
Differentiation P∞ 1
Periodic Extension = c + m=−∞ Um i2πmF ei2πmF t
t2 Periodic P∞
Extension: Method
(a) = c + m=−∞ Vm ei2πmF t where c is an integration constant
t2 Periodic −i
Extension: Method Hence Vm = 2πmF Um except for V0 = c (arbitrary constant)
(b)
Summary
Example:
Square wave: Um = −2i
mπ for odd m (0 for even m)
−i
Triangle wave: Vm = 2πmF × −2i
mπ = −1
π m2 F for odd m (0 for even m)
2

1 5
u7(t) v7(t)
0 0

-1 -5
0 5 10 15 20 0 5 10 15 20
1
Convergence: v(t) always converges if u(t) does since Vm ∝ m Um
vN (t) is a good approximation even for small N

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 5 / 11

 Rate at which coefficients decrease with m

5: Gibbs Phenomenon
Discontinuities
Square wave: Um = −2i
π m −1
for odd m (0 for even m)
Discontinuous
Waveform
Triangle wave: Vm = π−1
2F m
−2
for odd m (0 for even m)
Gibbs Phenomenon 1 5
u7(t) v7(t)
Integration
0 0
Rate at which
coefficients -1 -5
⊲ decrease with m 0 5 10 15 20 0 5 10 15 20 Integrating
Differentiation −i
Periodic Extension u(t) multiplies the Um by 2πF × m−1 ⇒ they decrease faster.
t2 Periodic
Extension: Method
(a)
The rate at which the coefficients, Um , decrease with m depends on the
t2 Periodic
Extension: Method lowest derivative that has a discontinuity:
(b)
Summary • Discontinuity in u(t) itself (e.g. square wave)
−1
For large |m|, Um decreases as |m|
• Discontinuity in u′ (t) (e.g. triangle wave)
For large |m|, Um decreases as |m|−2
• Discontinuity in u(n) (t)
−(n+1)
For large |m|, Um decreases as |m|
• No discontinuous derivatives
For large |m|, Um decreases faster than any power (e.g. e−|m| )

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 6 / 11

 Differentiation

5: Gibbs Phenomenon −i
Discontinuities
Integration multiplies Um by 2πmF .
Discontinuous
Waveform 2πmF
Gibbs Phenomenon
Hence differentiation multiplies Um by −i = i2πmF
Integration
Rate at which
coefficients decrease
If u(t) is a continuous differentiable function and w(t) = du(t)
dt then,
with m
⊲ Differentiation
provided that w(t) satisfies the Dirichlet conditions, its Fourier coefficients
Periodic Extension are: (
t2 Periodic
Extension: Method 0 m=0
(a) Wm = .
t2 Periodic i2πmF Um m 6= 0
Extension: Method
(b)
Summary Since we are multiplying Um by m the coefficients Wm decrease more
slowly with m and so the Fourier series for w(t) may not converge (i.e.
w(t) may not satisfy the Dirichlet conditions).

d d
dt dt
−→ −→
−2 −1 −0
Um ∝ |m| Um ∝ |m| Um ∝ |m|
Differentiation makes waveforms spikier and less smooth.

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 7 / 11

 Periodic Extension

5: Gibbs Phenomenon
Discontinuities
Suppose y(t) is only defined over a finite interval (a, b).
Discontinuous
Waveform You have two reasonable choices to make a periodic version:
Gibbs Phenomenon
Integration
Rate at which (a) T = b − a, u(t) = y(t) for a ≤ t < b
coefficients decrease (
with m
Differentiation y(t) a≤t≤b
⊲ Periodic Extension (b) T = 2(b − a), u(t) =
2
t Periodic y(2b − t) b ≤ t ≤ 2b − a
Extension: Method
(a)
t2 Periodic
Example:
Extension: Method
(b) y(t) = t2 for 0 ≤ t < 2
Summary
4 4 4

2 2 2

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

y(t) (a) T = 2 (b) T = 4

Option (b) has twice the period, no discontinuities, no Gibbs phenomenon
overshoots and if y(t) is continuous the coefficients decrease at least as fast
−2
as |m| .

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 8 / 11

 t2 Periodic Extension: Method (a)

5: Gibbs Phenomenon
Discontinuities
y(t) = t2 for 0 ≤ t < 2
Discontinuous
Waveform
Gibbs Phenomenon
Method (a): T = F1 = 2
Integration
1 T 2 −i2πmF t 1
RT 4
t2 dt =
R
Rate at which
coefficients decrease
Um = T 0 t e dt U0 = T 0 3
with m h iT
Differentiation 1 t2 e−i2πmF t 2te−i2πmF t 2e−i2πmF t
Periodic Extension = T −i2πmF − (−i2πmF )2
+ (−i2πmF )3 0
t2 Periodic
Extension: Method
⊲ (a) Substitute e−i2πmF 0 = e−i2πmF T = 1 [for integer m]
t2 Periodic
Extension: Method
h 2
i
(b) = T1 −i2πmF
T 2T
− (−i2πmF )2
Summary

2i 2
= πm + π 2 m2

4 4 4
K=1 K=3 K=6

2 2 2

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

U0:3 = [1.333, 0.203 + 0.637i, 0.051 + 0.318i, 0.023 + 0.212i]

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 9 / 11

 t2 Periodic Extension: Method (b)

5: Gibbs Phenomenon
Discontinuities
y(t) = t2 for 0 ≤ t < 2
Discontinuous
Waveform
Gibbs Phenomenon
Method (b): T = F1 = 4
Integration
1 0.5T 1
R 0.5T 4
Um = T −0.5T t2 e−i2πmF t dt t2 dt =
R
Rate at which
coefficients decrease
U0 = T −0.5T 3
with m
Differentiation
h 2 −i2πmF t i0.5T
1 t e 2te−i2πmF t −i2πmF t
2e
Periodic Extension = T −i2πmF − (−i2πmF )2 + (−i2πmF )3 −0.5T
t2 Periodic
Extension: Method
m
(a)
Substitute e±iπmF T = e±iπm = (−1) [for integer m]
t2 Periodic
Extension: Method h i
⊲ (b) (−1)m −2T
Summary
= T (−i2πmF )2
[all even powers of t cancel out]
(−1)m T 2 (−1)m 8
= 2π 2 m2 = π 2 m2

4 4 4
K=1 K=3 K=6

2 2 2

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

U0:3 = [1.333, −0.811, 0.203, −0.090] [u(t) real+even ⇒ Um real]

Convergence is noticeably faster than for method (a)
E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 10 / 11
 Summary

5: Gibbs Phenomenon
Discontinuities
• Discontinuity at t = a
Discontinuous
Waveform
◦ Gibbs phenomenon: uN (t) overshoots by 9% of iump
Gibbs Phenomenon ◦ uN (a) → mid point of iump
Integration Rt −i
Rate at which
coefficients decrease
• Integration: If v(t) = u(τ )dτ , then Vm = 2πmF Um
with m and V0 = c, an arbitrary constant. U0 must be zero.
Differentiation
Periodic Extension
t2 Periodic
• Differentiation: If w(t) = du(t)
dt , then Wm = i2πmF Um provided
Extension: Method
(a)
w(t) satisfies Dirichlet conditions (it might not)
t2 Periodic
Extension: Method
• Rate of decay:
(b) −(k+1) dk u(t)
⊲ Summary
◦ For large n, Un decreases at a rate |n| where dtk
is
the first discontinuous
D derivative
E P
2
◦ Error power: (u(t) − uN (t)) = |n|>N |Un |2
• Periodic Extension of finite domain signal of length L
◦ (a) Repeat indefinitely with period T = L
◦ (b) Reflect alternate repetitions for period T = 2L
no discontinuities or Gibbs phenomenon
For further details see RHB Chapter 12.4, 12.5, 12.6

E1.10 Fourier Series and Transforms (2014-5559) Gibbs Phenomenon: 5 – 11 / 11