Foundations of Audio Signal Processing 2022‑02‑14

Contents

Complex Numbers 2

Signal Spaces 2
Lebesgue spaces for DT signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Lebesgue spaces for CT signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Lebesgue spaces for periodic CT signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

Fourier Transform 4
A Fundamental Isomorphism of Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 5
Fourier Transform for Non‑periodic CT signals . . . . . . . . . . . . . . . . . . . . . . . . . 5
Properties of the Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Fourier Transform of DT Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Discrete Fourier Transform (DFT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

Analog to Digital Conversion 7

Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

Systems and Filters 9

Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
𝑧‑Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

Filter Design 11

Windowed Fourier Transform 12

Discrete WFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2D Signal Processing 13
2D Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

Signal Processing for Communications 14

Jan‑Niklas 1
Foundations of Audio Signal Processing 2022‑02‑14

Complex Numbers

⎛𝑎 −𝑏⎞
• As 𝑎 + 𝑖𝑏 or (𝑎, 𝑏) or ⎜
⎜ ⎟
⎟ with 𝑎, 𝑏 ∈ ℝ
⎝ 𝑏 𝑎 ⎠
• In polar coordinates: (|𝑧| , arg(𝑧)) or 𝑟𝑒𝑖𝜑

• Eulers formula: 𝑒𝑖𝑧 = cos(𝑧) + 𝑖 sin(𝑧)

• 𝑧 is an 𝑛th root of unity if 𝑧𝑛 = 1. It is primitive if it is not an 𝑚th root of unity for any smaller 𝑚.

– The 𝑛th roots of unity are exp( 2𝑘𝜋𝑖

𝑛 ) for 𝑘 = 0, 1, … , 𝑛 − 1.

– If 𝑛 is prime, all 𝑛th roots of unity (except for 1) are primitive.

Signal Spaces

• A signal is a function 𝑓 ∶ 𝑋 → ℝ or 𝑓 ∶ 𝑋 → ℂ, where

– Continuous‑time signal: 𝑋 is an interval in ℝ𝑛

– Discrete‑time signal: 𝑋 is an interval in ℤ𝑛

𝑠𝑖𝑛(𝜋𝑡)
• Sinc function: sinc(𝑡) ≔ 𝜋𝑡 and sinc(0) = 1

– Continuous

– Roots exactly at the non‑zero integers

⎧
{1, |𝑡| ≤ 𝑤2
• Box function: 𝑏𝑤 (𝑡) ≔
⎨
{
⎩0, otherwise
• Frequency signal of frequency 𝜔 ∈ [0, 1): 𝑘 ↦ 𝑒2𝜋𝑖𝜔𝑘

• Sampling: CT‑signal 𝑓 is transformed to DT‑signal 𝑥 by 𝑇 ‑sampling, 𝑥(𝑛) ≔ 𝑓(𝑇 ⋅ 𝑛)

⎧
{1, 𝑛=0
• Unit impulse 𝛿 ∶ ℤ → {0, 1} , 𝑛 ↦ ⎨
{
⎩0, 𝑛≠0

Jan‑Niklas 2
Foundations of Audio Signal Processing 2022‑02‑14

Figure 1: Graph of the sinc function

Hilbert Space: Complete vector space with scalar product

Hilbert Basis: ON‑system (𝑒𝑖 ) of a Hilbert space 𝑉 is called a Hilbert basis, if the following equivalent
conditions are true:

• Completeness: If a vector is orthogonal to all members of the base, it is zero

2 2
• Parseval equality: ∀𝑥 ∈ 𝑉 ∶ ‖𝑥‖ = ∑𝑖 |⟨𝑥, 𝑒𝑖 ⟩|

• Fourier series: Every 𝑥 ∈ 𝑉 has a representation 𝑥 = ∑𝑖 ⟨𝑥, 𝑒𝑖 ⟩ 𝑒𝑖 , where the number of non‑
zero terms is countable

Lebesgue spaces for DT signals

𝑝
ℓ𝑝 (ℤ) = { 𝑥 ∶ ℤ → ℂ ∣ ∑ |𝑥𝑛 | < ∞ }
𝑛

Jan‑Niklas 3
Foundations of Audio Signal Processing 2022‑02‑14

And ℓ∞ (ℤ) is the space of all bounded signals in ℤ. These have the norms
1
𝑝
𝑝
‖𝑥‖𝑝 ≔ (∑ |𝑥𝑛 | )
𝑛

‖𝑥‖∞ ≔ sup |𝑥𝑛 | .

𝑛

ℓ1 (ℤ) ⊂ ℓ2 (ℤ) ⊂ ⋯ ⊂ ℓ∞ (ℤ)

‖𝑥‖1 ≥ ‖𝑥‖2 ≥ ⋯ ≥ ‖𝑥‖∞

ℓ2 (ℤ) is the only Hilbert space among these, with scalar product ⟨𝑥, 𝑦⟩ = ∑𝑛 𝑥(𝑛)𝑦(𝑛).

Lebesgue spaces for CT signals

𝑝
𝐿𝑝 (ℝ) = { 𝑓 ∶ ℝ → ℂ measurable ∣ ∫ |𝑓(𝑡)| d𝑡 < ∞ }
ℝ
∞
And 𝐿 (ℝ) is the space of all measurable signals that are essentially bounded in ℝ. These have the
norms 1
𝑝
𝑝
‖𝑓‖𝑝 ≔ (∫ |𝑓(𝑡)| d𝑡)
ℝ

‖𝑓‖∞ ≔ ess sup |𝑓(𝑡)| .

𝑡∈ℝ

𝐿2 (ℝ) is the only Hilbert space among these, with scalar product ⟨𝑓, 𝑔⟩ = ∫ℝ 𝑓(𝑡)𝑔(𝑡) d𝑡.

Lebesgue spaces for periodic CT signals

1
𝐿2 ([0, 1]) is a Hilbert space with scalar product ⟨𝑓, 𝑔⟩ = ∫0 𝑓(𝑡)𝑔(𝑡) d𝑡

Fourier Transform

Hilbert space 𝐿2 ([0, 1]) has (among others) these Hilbert bases:
√ √
• { 1, 𝐴𝑘 , 𝐵𝑘 | 𝑘 ∈ ℕ } with 𝐴𝑘 (𝑡) = 2 cos(2𝜋𝑘𝑡) and 𝐵𝑘 (𝑡) = 2 sin(2𝜋𝑘𝑡)

• { e𝑘 | 𝑘 ∈ ℤ } with e𝑘 (𝑡) = 𝑒2𝜋𝑖𝑘𝑡

Jan‑Niklas 4
Foundations of Audio Signal Processing 2022‑02‑14

A Fundamental Isomorphism of Hilbert Spaces

The function 𝑓 ↦ 𝑓 ̂ ≔ (⟨𝑓, e𝑘 ⟩)𝑘∈ℤ , which assigns every signal the sequence of its Fourier coefficients,
is an isomorphism of the Hilbert spaces 𝐿2 ([0, 1]) and ℓ2 (ℤ), that also keeps the scalar product invari‑
ant.

Fourier Transform for Non‑periodic CT signals

For 𝑓 ∈ 𝐿1 (ℝ) ∩ 𝐿2 (ℝ) and 𝜔 ∈ ℝ we define

̂
𝑓(𝜔) ≔ ∫ 𝑓(𝑡)𝑒−2𝜋𝑖𝜔𝑡 d𝑡 .
ℝ

This definition can be continued to all of 𝐿2 (ℝ) to define a unitary transform 𝑓 ↦ 𝑓 ̂ on 𝐿2 (ℝ). Its
inverse is
̌ ≔ ∫ 𝑓(𝜔)𝑒2𝜋𝑖𝜔𝑡 d𝜔 .
𝑓(𝑡)
ℝ

Might be helpful for calculations:

̂
𝑓(𝜔) = ∫ 𝑓(𝑡) ⋅ (cos(2𝜋𝜔𝑡) − 𝑖 sin(2𝜋𝜔𝑡)) d𝑡
ℝ

Properties of the Fourier Transform

• 𝑡0 ‑translation: 𝑓𝑡0 (𝑡) ≔ 𝑓(𝑡 − 𝑡0 ). Then 𝑓̂

𝑡0 (𝜔) = 𝑒
−2𝜋𝑖𝜔𝑡0 ̂
𝑓(𝜔)

• 𝜔0 ‑modulation: 𝑓 𝜔0 (𝑡) ≔ 𝑒−2𝜋𝑖𝜔0 𝑡 𝑓(𝑡). Then 𝑓̂ ̂ +𝜔 )

𝜔0 (𝜔) = 𝑓(𝜔
0

• 𝑓̂′ (𝜔) = 2𝜋𝑖𝜔𝑓(𝜔)

̂

• 𝑔(𝑡) ≔ 𝑡𝑓(𝑡). Then 𝑓 ′̂ (𝜔) = −2𝜋𝑖𝑔(𝜔)

̂
̂
• 𝑓( ⋅ ̂
𝑠 )(𝜔) = |𝑠| ⋅ 𝑓(𝜔𝑠)
̌ ̂
• 𝑓 ̂= 𝑓 = 𝑓 ̌

Fourier Transform of DT Signals

For 𝑥 ∈ ℓ2 (ℤ):
𝑥(𝜔)
̂ ≔ ∑ 𝑥(𝑘)𝑒−2𝜋𝑖𝜔𝑘 ∈ 𝐿2 ([0, 1])
𝑘

Properties:

Jan‑Niklas 5
Foundations of Audio Signal Processing 2022‑02‑14

̂
• 𝑥(𝜔) = 𝑥(−𝜔)
̂

• 𝑦(𝑛) = 𝑥(−𝑛) ⟹ 𝑦(𝜔)

̂ ̂
= 𝑓(−𝜔) for all 𝑛 ∈ ℤ

Table 1: Formulas for Fourier transforms. All of these are unitary isomorphisms.

Name Space (from) Space(to) Formula

Fourier Transform 𝐿2 (ℝ) 𝐿2 (ℝ) ̂

𝑓(𝜔) ≔ ∫ℝ 𝑓(𝑡)𝑒−2𝜋𝑖𝜔𝑡 d𝑡
Inverse Fourier Transform 𝐿2 (ℝ) 𝐿2 (ℝ) ̌ ≔ ∫ 𝑓(𝜔)𝑒2𝜋𝑖𝜔𝑡 d𝜔
𝑓(𝑡) ℝ

Fourier Transform for DT Signals ℓ2 (ℤ) 𝐿2 ([0, 1]) 𝑥(𝜔)

̂ ≔ ∑𝑘 𝑥(𝑘)𝑒−2𝜋𝑖𝜔𝑘
Fourier Series Analysis 𝐿2 ([0, 1]) ℓ2 (ℤ) ̂ ≔ ∫1 𝑓(𝑡)𝑒−2𝜋𝑖𝑘𝑡 d𝑡
𝑓(𝑘) 0

Table 2: Fourier Transform Properties.

Property Signal Fourier Transform

Linearity 𝛼𝑥(𝑡) + 𝛽𝑦(𝑡) 𝛼𝑋(𝜔) + 𝛽𝑌 (𝜔)

Duality 𝑋(𝑡) 𝑥(−𝜔)
Conjugacy 𝑥(𝑡) 𝑋(−𝜔)
1 𝜔
Time‑scaling 𝑥(𝑎𝑡) |𝑎| 𝑋( 𝑎 )

Time‑shift 𝑥(𝑡 − 𝑡0 ) 𝑒−2𝜋𝑖𝜔𝑡0 𝑋(𝜔)

Modulation 𝑒2𝜋𝑖𝜔0 𝑡 𝑥(𝑡) 𝑋(𝜔 − 𝜔0 )
Convolution 𝑥(𝑡) ∗ 𝑦(𝑡) 𝑋(𝜔)𝑌 (𝜔)
Multiplication 𝑥(𝑡)𝑦(𝑡) 𝑋(𝜔) ∗ 𝑌 (𝜔)
Differentiation 𝑥′ (𝑡) 2𝜋𝑖𝜔𝑋(𝜔)
𝑡 𝑋(𝜔)
Integration ∫−∞ 𝑥(𝜏 ) d𝜏 2𝜋𝑖𝜔 + 12 𝑋(0)𝛿(𝜔)

∞ ∞
Parseval’s Theorem: ∫−∞ 𝑥(𝑡)𝑦(𝑡) d𝑡 = ∫−∞ 𝑋(𝜔)𝑌 (𝜔) d𝜔

Jan‑Niklas 6
Foundations of Audio Signal Processing 2022‑02‑14

1
Table 3: Fourier Transform Pairs, with 𝜃(𝑡) = 2 sgn(𝑡) + 12 .

Time Domain Frequency Domain

𝑏1 (𝑡) sinc(𝜔)
sinc(𝑡) 𝑏1 (𝜔)
2 2
𝑒−𝜋𝑡 𝑒−𝜋𝜔
𝛿(𝑡) 1
1 𝛿(𝜔)
𝛿(𝑡 − 𝑡0 ) 𝑒−2𝜋𝑖𝜔𝑡0
𝑒2𝜋𝑖𝜔0 𝑡 𝛿(𝜔 − 𝜔0 )
1
sgn(𝑡) 𝜋𝑖𝜔
1
𝜋𝑡 −𝑖 sgn(𝜔)
1
𝜃(𝑡) 2𝜋𝑖𝜔 + 12 𝛿(𝜔)
1 𝑖
2 𝛿(𝑡) + 2𝜋𝑡 𝜃(𝜔)

• 𝑓 real: Re(𝑓)̂ even and Im(𝑓)̂ odd

• 𝑓 purely imaginary: Re(𝑓)̂ odd and Im(𝑓)̂ even
• 𝑓 even: 𝑓 ̂ real
• 𝑓 odd: 𝑓 ̂ purely imaginary

Discrete Fourier Transform (DFT)

Fourier coefficients of a finite DT signal can be computed via matrix‑vector multiplication with the DFT
matrix.

Analog to Digital Conversion

• Analog to digital conversion (ADC): Sampling ((ℝ → ℝ) → (ℤ → ℝ)) and Quantization ((ℤ →
ℝ) → (ℤ → ℤ))

Jan‑Niklas 7
Foundations of Audio Signal Processing 2022‑02‑14

Sampling

𝑇 ‑sampling: CT‑signal 𝑓 ∶ ℝ → ℂ transformed into DT‑signal 𝑥 ∶ ℤ → ℂ with

𝑥(𝑛) ≔ 𝑓(𝑇 ⋅ 𝑛) .

1
𝑇 is the sampling rate.
Synthesis function: Approximate signal 𝑓 given the sampled signal 𝑥.
Bandlimited function: Signal which only contains frequencies up to a certain threshold.
Shannon Sampling Theorem: Bandlimited function can be perfectly reconstructed from a suitable
1
set of samples. Let 𝑓 ∈ 𝐿2 (ℝ) be Ω‑bandlimited and 𝑥 the 𝑇 ‑sampled version of 𝑓 with 𝑇 = 2Ω . Then
𝑓 can be reconstructed from 𝑥:
∞ ∞
𝑡 − 𝑛𝑇 𝑛
𝑓(𝑡) = ∑ 𝑥(𝑛) sinc ( ) = ∑ 𝑓 ( ) sinc(2Ω𝑡 − 𝑛)
𝑛=−∞
𝑇 𝑛=−∞
2Ω

Sampling rate of 2Ω Hz is sufficient for a perfect reconstruction, called Nyquist‑Rate. Ω itself is called
Nyquist frequency.
𝑥 is 𝑇 ‑sampled 𝑓, then:
1 𝜔+𝑘
𝑥(𝜔)
̂ = ∑ 𝑓 ̂( )
𝑇 𝑘∈ℤ 𝑇

Aliasing: In an undersampled DT‑signal 𝑥, high frequency components of 𝑓 outside the interval

1
[−Ω, Ω] (with 𝑇 = 2Ω ) cannot be distinguished from certain low frequency components of 𝑓 withing
the interval, i.e. high‑frequency components act as lower frequencies. High frequencies above |Ω|
are “folded” into the interval [−Ω, Ω]. This effect is called aliasing.

Quantization

Partition ℝ into contiguous intervals, represent each interval by a codeword. Coder maps real value
to codeword, decoder maps back to real value.
Uniform Scalar Quantizer: Uniform width 𝑄 of quantizer levels, except the first and last level which
extend to infinity.
Non‑uniform quantization: Coder compresses the amplitude range before performing a uniform
quantization, decoder expands amplitude range after reversing the quantization. 𝜇‑Law compressor
function for 𝜇 > 0:
ln (1 + 𝑥𝜇|𝑥| )
𝑐𝜇 (𝑥) ≔ 𝑥max max
⋅ sign(𝑥)
ln(1 + 𝜇)

Jan‑Niklas 8
Foundations of Audio Signal Processing 2022‑02‑14

Systems and Filters

A system or operator is a mapping that transforms signals. In our case only DT‑signals, i.e. the signals
spaces are ℓ𝑝 (ℤ). A system is linear, if the input and output spaces are linear and the mapping is
ℂ‑linear.

Common operators:

• Time‑shift 𝑇 𝑘 , 𝑘 ∈ ℤ is stable LTI, defined by

𝑇 𝑘 [𝑥](𝑛) ≔ 𝑥(𝑛 − 𝑘) .

• Downsampling ↓ 𝑀 , 𝑀 ∈ ℕ is linear stable non‑time‑invariant, defined by

↓ 𝑀 [𝑥](𝑛) ≔ 𝑥(𝑀 ⋅ 𝑛) .

• Upsampling ↑ 𝑀 , 𝑀 ∈ ℕ is linear stable non‑time‑invariant, defined by

⎧
{𝑥 ( 𝑀 ) , 𝑀 |𝑛
↑ 𝑀 [𝑥](𝑛) ≔ ⎨ 𝑛 .
{
⎩0, else

• Amplitude modulation 𝑚𝑐 with respect to a carrier signal 𝑐 ∈ ℓ∞ (ℤ) is linear non‑continuous

non‑time‑invariant, defined by

𝑚𝑐 [𝑥](𝑛) ≔ 𝑐(𝑛) ⋅ 𝑥(𝑛) .

– Special case frequency‑shift 𝐸𝜔 , 𝜔 ∈ (0, 1), defined by

𝐸𝜔 [𝑥](𝑛) ≔ 𝑒−2𝜋𝑖𝜔𝑛 𝑥(𝑛) .

• Rounding operators, non‑linear but time‑invariant:

– Ceil‑operator: ⌈𝑥⌉ (𝑛) ≔ ⌈𝑥(𝑛)⌉

– Floor‑operator: ⌊𝑥⌋ (𝑛) ≔ ⌊𝑥(𝑛)⌋

• Cut‑off operator Cut𝜆 , 𝜆 > 0 is non‑linear but time‑invariant, defined by

⎧
{𝑥(𝑛), |𝑥(𝑛)| ≤ 𝜆
Cut𝜆 [𝑥](𝑛) ≔ ⎨ .
{
⎩sgn(𝑥(𝑛)) ⋅ 𝜆, else

Jan‑Niklas 9
Foundations of Audio Signal Processing 2022‑02‑14

A linear system 𝑇 is continuous if it transforms a convergent sequence of input signals into a con‑
vergent sequence of output signals. A linear system 𝑇 is time invariant if it commutes with every
time‑shift. Linear, time invariant systems are called LTI‑systems.

Convolution

For two signals 𝑥, 𝑦 ∶ ℤ → ℂ, the convolution of 𝑥 and 𝑦 is defined by

(𝑥 ∗ 𝑦)(𝑛) ≔ ∑ 𝑥(𝑘)𝑦(𝑛 − 𝑘) .
𝑘∈ℤ

𝐶𝑦 [𝑥] ≔ 𝑥 ∗ 𝑦 defines the convolution operator, which is linear and continuous.

Properties of the convolution:

• Commutativity, Associativity, Distributivity (with +)

̂
• Convolution theorem: 𝑥 ∗ 𝑦 = 𝑥̂ ⋅ 𝑦 ̂

• 𝑇 𝑘 [𝑥] = 𝑥𝑘 = 𝑥 ∗ 𝛿 𝑘 , i.e. 𝑇 𝑘 = 𝐶𝛿𝑘

A continuous LTI‑system 𝑇 coincides with the convolution operator of its impulse response ℎ ≔ 𝑇 [𝛿],
i.e. 𝑇 [𝑥] = 𝐶ℎ [𝑥] = ℎ ∗ 𝑥. ℎ(𝑛) is called the 𝑛‑th filter coefficient of 𝑇 .

A linear system 𝑇 ∶ ℓ𝑝 (ℤ) → ℓ𝑝 (ℤ) is called stable if it is continuous and 𝑇 [𝛿 𝑘 ] ∈ ℓ1 (ℤ) for all 𝑘 ∈ ℤ.
Special case for a system on ℓ∞ (ℤ): BIBO‑stable (bounded input, bounded output). 𝑇 is stable and
time‑invariant iff there exists an ℎ ∈ ℓ1 (ℤ) such that 𝑇 = 𝐶ℎ .

A continuous LTI‑system is called FIR‑system if only finitely many filter coefficients are non‑zero. Oth‑
erwise it is called IIR‑system. The system is called causal if ∀𝑛 < 0 ∶ ℎ(𝑛) = 0. The length ℓ(ℎ) of a
FIR‑filter ℎ is the number of elements from its first to its last non‑zero element (inclusive). The order
of a causal FIR‑filter ℎ is ℓ(ℎ) − 1.

Fourier transform of impulse response ℎ is called frequency response ℎ̂ or 𝐻. For each frequency
𝜔, the frequency sequence 𝑓𝜔 is an eigenvector of the convolution operator 𝐶ℎ for the eigenvalue
𝐻(𝜔):
𝐶ℎ [𝑓𝜔 ] = 𝐻(𝜔)𝑓𝜔 , where 𝑓𝜔 ≔ (𝑒2𝜋𝑖𝜔𝑘 )𝑘∈ℤ

𝜔 ↦ |𝐻(𝜔)| is called amplitude response, 𝜔 ↦ arg(𝐻(𝜔)) = Φℎ (𝜔) is called phase response.

Properties:

• For real‑valued ℎ: 𝐻(𝜔) = 𝐻(−𝜔), and thus |𝐻(𝜔)| = |𝐻(−𝜔)| and Φℎ (−𝜔) = −Φℎ (𝜔)

Jan‑Niklas 10
Foundations of Audio Signal Processing 2022‑02‑14

• Frequency response is 1‑periodic. In case of real‑valued ℎ, all information on 𝐻 is contained in

its restriction to the interval [0, 12 ], with low and high frequencies corresponding to 𝜔 near 0 and
1
2 respectively.

𝑧‑Transform

Value of discrete signals interpreted as coefficients of a polynomial (for finite signals) or formal power
series (for infinite signals) in the variable 𝑧. The 𝑧‑transform formalizes this idea: let 𝑥 ∶ ℤ → ℂ. The
𝑧‑transform 𝑍[𝑥] ≕ 𝑋 of 𝑥 is the function 𝑋 ∶ 𝐷𝑥 → ℂ defined by

𝑋(𝑧) = ∑ 𝑥(𝑛) ⋅ 𝑧 −𝑛 ,
𝑛∈ℤ

where 𝐷𝑥 denotes the region of convergence. Properties of the 𝑧‑transform:

• Linearity, on the common region of convergence

• Delay‑property: 𝑍[𝑥𝑘 ](𝑧) = 𝑧 −𝑘 ⋅ 𝑍[𝑥](𝑧)

• Multiplication by exponential: 𝑍[𝑛 ↦ 𝑎𝑛 𝑥(𝑛)](𝑧) = 𝑍[𝑥]( 𝑎𝑧 ) for 0 ≠ 𝑎 ∈ ℂ, converges on

|𝑎| ⋅ 𝐷𝑥

• Conjugation‑property: 𝑍[𝑥](𝑧) = 𝑍[𝑥](𝑧), converges on 𝐷𝑥

• Time‑reversal: 𝑍[𝑛 ↦ 𝑥(−𝑛)](𝑧) = 𝑍[𝑥](𝑧 −1 ), converges on 𝐷𝑥−1

• Convolution theorem: 𝑍[𝑥 ∗ 𝑦](𝑧) = 𝑍[𝑥](𝑧) ⋅ 𝑍[𝑦](𝑧), converges on a superset of 𝐷𝑥 ∩ 𝐷𝑦

The 𝑧‑transform of the impulse response of a continuous LTI‑system is called the transfer‑function
of the system.

Filter Design

Coefficients of an ideal low‑pass filter with cut‑off frequency 𝜔0 ∈ (0, 12 ]: 𝑘 ↦ 2𝜔0 sinc(2𝜔0 𝑘)

Amplitude response of real low‑pass filter: pass band oscillates around 1, transition band goes straight
from 1 to 0, stop band oscillates around 0. Pass band is at 1 ± 𝛿1 , stop band at 0 + 𝛿2 . Pass band to
transition band at frequency 𝜔𝑝 , transition band to stop band at frequency 𝜔𝑠 .

In general, the phase response of a filter induces a time‑shift in the filtered signal depending on the
frequency. A filter has linear phase when Φℎ (𝜔) = 𝑐𝜔 mod 1 for some 𝑐 ∈ ℝ. In this case, the time‑
delay is independent of the frequency. This time‑delay is formalized using the group delay defined

Jan‑Niklas 11
Foundations of Audio Signal Processing 2022‑02‑14

by
dΦℎ
𝜏ℎ (𝜔) ≔ − (𝜔) .
d𝜔
Filters with linear phase have a constant group delay.

Even filters have a purely real frequency response, odd filters a purely imaginary. The frequency re‑
sponse of a real‑valued filter has even real part and odd imaginary part.

A causal FIR‑filter ℎ that satisfies the requirement of symmetry ℎ(𝑘) = ℎ(2𝑁 − 𝑘) or the require‑
ment of anti‑symmetry ℎ(𝑘) = −ℎ(2𝑁 − 𝑘) has linear phase.
1
Averaging filter: ℎ𝑁 (𝑛) = 𝑁 if 0 ≤ 𝑛 < 𝑁 and ℎ𝑁 (𝑛) = 0 otherwise

Gibbs‑Phenomenon: The amplitude response of an approximate causal linear‑phase FIR low‑pass

filter oscillates at the point of the cut‑off frequency. For increasing filter length, the ripples move
closer to the discontinuity, but their magnitude does not tend to zero.

Windowed Fourier Transform

The information encoded in the Fourier coefficients contains no obvious temporal information about
the signal and is thus kind of the averaged frequency content. The windowed Fourier transform is
used to also obtain temporal information.

A window function is any 𝑔 ∈ 𝐿2 (ℝ) with ‖𝑔‖2 > 0. For a window function 𝑔 we define (musical)
notes of frequency 𝜔 and time position 𝑡 as

𝑔𝜔,𝑡 (𝑢) ≔ 𝑔(𝑢 − 𝑡) ⋅ 𝑒2𝜋𝑖𝜔𝑢 , 𝑢 ∈ ℝ.

For a window function 𝑔 and an 𝑓 ∈ 𝐿2 (ℝ) the function 𝑓 ∶̃ ℝ2 → ℂ defined by

̃ 𝑡) ≔ ⟨𝑓, 𝑔 ⟩ = ∫ 𝑓(𝑢)𝑔(𝑢 − 𝑡)𝑒−2𝜋𝑖𝜔𝑢 d𝑢

𝑓(𝜔, 𝜔,𝑡
ℝ

is called Windowed Fourier Transform of 𝑓 with regards to 𝑔.

For a window function 𝑔 with ‖𝑔‖ = 1 we define:

• The center of 𝑔:
2
𝑡0 (𝑔) ≔ ∫ 𝑡 |𝑔(𝑡)| d𝑡
ℝ

• The width of 𝑔:
2
𝑇 (𝑔) ≔ √∫(𝑡 − 𝑡0 )2 |𝑔(𝑡)| d𝑡
ℝ

Jan‑Niklas 12
Foundations of Audio Signal Processing 2022‑02‑14

• The center of 𝑔:̂

2
𝜔0 (𝑔) ≔ ∫ 𝜔 |𝑔(𝜔)|
̂ d𝜔
ℝ

• The width of 𝑔:̂

2
Ω(𝑔) ≔ √∫(𝜔 − 𝜔0 )2 |𝑔(𝜔)|
̂ d𝜔
ℝ

1
The product of the widths of 𝑔 and 𝑔 ̂ is at least 4𝜋 .

Reconstruction: For 𝑔 window function and 𝑓 ∈ 𝐿2 (ℝ):

1 ̃ 𝑡)𝑔 (𝑢) d𝜔 d𝑡
𝑓(𝑢) = 2
∫ ∫ 𝑓(𝜔, 𝜔,𝑡
‖𝑔‖ ℝ ℝ

Discrete WFT
𝑔(𝑢 − 𝑛𝜏 ) ̃
𝑓(𝑢) = 𝜏 𝜈 ∑ ∑ 𝑓(𝑚𝜈, 𝑛𝜏 )𝑒2𝜋𝑖𝑚𝜈𝑢
𝑛∈ℤ 𝑚∈ℤ
𝐻 𝜏 (𝑢)

For:

• Window function 𝑔 with compact support in [𝑎, 𝑏]

1 1
• Frequency step size 𝜈 ≔ 𝑏−𝑎 = 𝑇

• Sufficiently small temporal step size 𝜏 > 0.

The sample density

1 𝑏−𝑎
𝜌(𝜈, 𝜏 ) ≔ =
𝜈⋅𝜏 𝜏
is a measure of the redundancy of the representation of 𝑓 in the discrete WFT reconstruction.

2D Signal Processing

CT signals 𝑓 ∶ ℝ2 → ℂ, DT signals 𝑥 ∶ ℤ2 → ℂ.

• Unit impulse: 𝛿(𝑛1 , 𝑛2 ) = 1 if 𝑛1 = 𝑛2 = 0, else 0

• Unit step function: 𝑢(𝑛1 , 𝑛2 ) = 1 if 𝑛1 ≥ 0, 𝑛2 ≥ 0, else 0
𝑊1 𝑊2
• Box function: 𝑏𝑊1 ,𝑊2 (𝑡1 , 𝑡2 ) = 1 if |𝑡1 | ≤ 2 , |𝑡2 | ≤ 2 , else 0

A 2D signal is called separable if it can be expressed as a product of two one‑dimensional signals:

𝑥(𝑛1 , 𝑛2 ) = 𝑥1 (𝑛1 ) ⋅ 𝑥2 (𝑛2 )

Jan‑Niklas 13
Foundations of Audio Signal Processing 2022‑02‑14

Lebesgue spaces are completely analogously to 1D signals:

𝑝
ℓ𝑝 (ℤ2 ) = { 𝑥 ∶ ℤ2 → ℂ ∣ ∑ ∑ |𝑥(𝑛1 , 𝑛2 )| < ∞ }
𝑛1 𝑛2

𝑝
𝐿𝑝 (ℝ2 ) = { 𝑓 ∶ ℝ2 → ℂ measurable ∣ ∫ ∫ |𝑓(𝑡1 , 𝑡2 )| d𝑡1 d𝑡2 < ∞ }
ℝ ℝ

Linearity, continuity and time‑invariance of systems also apply in the 2D case, with the 2D shift oper‑
ator 𝜏𝑘1 ,𝑘2 [𝑥](𝑛1 , 𝑛2 ) ≔ 𝑥(𝑛1 − 𝑘1 , 𝑛2 − 𝑘2 ).

Convolution of two signals 𝑥, 𝑦 ∶ ℤ2 → ℂ:

(𝑥 ∗ 𝑦)(𝑛1 , 𝑛2 ) ≔ ∑ ∑ 𝑥(𝑘1 , 𝑘2 )𝑦(𝑛1 − 𝑘1 , 𝑛2 − 𝑘2 )

𝑘1 ∈ℤ 𝑘2 ∈ℤ

2D Fourier Transform

2D frequency functions: (𝑡1 , 𝑡2 ) ⟼ 𝑒2𝜋𝑖(𝜔1 𝑡1 +𝜔2 𝑡2 )

Fourier transform is the same, just with double sum/integral:

̂ , 𝜔 ) ≔ ∫ ∫ 𝑓(𝑡 , 𝑡 )𝑒−2𝜋𝑖(𝜔1 𝑡1 +𝜔2 𝑡2 ) d𝑡 d𝑡 ,

𝑓(𝜔 for 𝐿2 (ℝ2 ) → 𝐿2 (ℝ2 )
1 2 1 2 1 2
ℝ ℝ

̂ 1 , 𝜔2 ) ≔ ∑ ∑ 𝑥(𝑘1 , 𝑘2 )𝑒−2𝜋𝑖(𝜔1 𝑘1 +𝜔2 𝑘2 ) ,

𝑥(𝜔 for ℓ2 (ℤ2 ) → 𝐿2 ([0, 1]2 )
𝑘1 ∈ℤ 𝑘2 ∈ℤ

2D Fourier transform can be written as a cascade of two 1D Fourier transforms.

𝑧‑transform can also simply be generalized to 2D.

When sampling 2D signals we measure the sampling rate in each dimension separately.

2D sampling theorem: Let 𝑓 ∈ 𝐿2 (ℝ2 ) be an (Ω1 , Ω2 )‑bandlimited function and 𝑥 the (𝑇1 , 𝑇2 )‑
sampled version of 𝑓 with 𝑇1 = 2Ω1 , 𝑇2 = 2Ω1 . Then 𝑓 can be reconstructed from 𝑥 using
1 2

𝑡1 − 𝑛1 𝑇1 𝑡 − 𝑛 2 𝑇2
𝑓(𝑡1 , 𝑡2 ) = ∑ ∑ 𝑥(𝑛1 , 𝑛2 ) sinc ( ) sinc ( 2 ).
𝑛1 ∈ℤ 𝑛2 ∈ℤ
𝑇1 𝑇2

Signal Processing for Communications

• Modulator ‑> Noisy channel ‑> Demodulator

Jan‑Niklas 14
Foundations of Audio Signal Processing 2022‑02‑14

• Lowpass or baseband signal: spectrum located around zero frequency, bandwidth 𝑊 .

• Modulated to bandpass signal by translating to higher frequencies, bandwidth 𝑊 and central

frequency 𝜔0 , to match spectral characteristics of the channel.
1
• Hilbert Transform: H[𝑥](𝑡) = 𝑥(𝑡)
̃ ≔ 𝜋𝑡 ∗ 𝑥(𝑡)

– Continuous, bounded linear operator on 𝐿𝑝 (ℝ)

̂̃
– 𝑥(𝜔) = −𝑖 sgn(𝜔)𝑥(𝜔)
̂

– H2 𝑥 = −𝑥

– H−1 = −H

– ⟨H𝑥, 𝑦⟩ = − ⟨𝑥, H𝑦⟩ and ⟨H𝑥, H𝑦⟩ = ⟨𝑥, 𝑦⟩ and ⟨H𝑥, 𝑥⟩ = 0

• Analytic Signal corresponding to 𝑥: 𝑥+ (𝑡) = 12 𝑥(𝑡) + 12 𝑥(𝑡)

̃

Jan‑Niklas 15