Ferkans — Interactive Telecom Tutor

Four Views of the Same Channel

The discrete multipath channel $h(\tau, t) = \sum_i h_i(t)\,\delta(\tau - \tau_i(t))$ is already a complete description. But depending on what we want to do with it — equalize, detect, estimate, sense — we want to look at it from different angles. Bello (1963) organized these angles into four equivalent functions, any two of which are related by a one-dimensional Fourier transform. The point is that the same channel looks simple in exactly one of these four views, and that view is the delay-Doppler spreading function $h(\tau, \nu)$ — the subject of the next section. But to see why that is, we need all four pictures first.

Definition:
Delay-Time Function (Time-Variant Impulse Response)

The delay-time function $h(\tau, t)$ is the time-varying impulse response: the output at time $t$ of the channel in response to an impulse applied at time $t - \tau$ . For a signal $x(t)$ transmitted, the received signal is $y(t) \;=\; \int_{-\infty}^{\infty} h(\tau, t)\,x(t - \tau)\,d\tau \;+\; w(t),$ where $\mathbf{w}(t)$ is AWGN. The function $h(\tau, t)$ captures how the channel's impulse response drifts with $t$ .

When the channel is time-invariant ( $h$ does not depend on $t$ ), this reduces to the familiar LTI convolution from Telecom Ch. 4.

Definition:
Frequency-Time Function (Time-Varying Transfer Function)

Fourier-transforming $h(\tau, t)$ with respect to the delay variable yields the frequency-time function: $H(f, t) \;=\; \int_{-\infty}^{\infty} h(\tau, t)\,e^{-j 2\pi f \tau}\,d\tau.$ This is the "snapshot transfer function" at time $t$ : if we froze the channel at time $t$ , $H(f, t)$ would be its frequency response. In OFDM, this is the coefficient that multiplies each subcarrier during one symbol.

Definition:
Delay-Doppler Spreading Function

Fourier-transforming $h(\tau, t)$ with respect to the time variable yields the delay-Doppler spreading function: $h(\tau, \nu) \;=\; \int_{-\infty}^{\infty} h(\tau, t)\,e^{-j 2\pi \nu t}\,dt.$ For the discrete multipath model, $h(\tau, \nu) = \sum_{i=1}^P h_i\,\delta(\tau - \tau_i)\,\delta(\nu - \nu_i)$ — a sum of $P$ Dirac impulses in the 2D delay-Doppler plane. This is the object around which the entire book is organized.

,

Definition:
Output Doppler Spread Function

The fourth Bello function is obtained by Fourier-transforming $H(f, t)$ with respect to time: $H(f, \nu) \;=\; \int_{-\infty}^{\infty} H(f, t)\,e^{-j 2\pi \nu t}\,dt.$ This represents how each frequency component of the transmitted signal is spread in Doppler. It is related to $h(\tau, \nu)$ by a Fourier transform in $\tau \leftrightarrow f$ .

Theorem: Bello's Four-Function Diagram

The four system functions are connected by one-dimensional Fourier transforms: $\begin{array}{ccc} h(\tau, t) & \xrightarrow{\mathcal{F}_\tau} & H(f, t) \\ \big\downarrow \mathcal{F}_t & & \big\downarrow \mathcal{F}_t \\ h(\tau, \nu) & \xrightarrow{\mathcal{F}_\tau} & H(f, \nu) \end{array}$ where $\mathcal{F}_\tau$ is Fourier transform in delay (with conjugate variable $f$ ), and $\mathcal{F}_t$ is Fourier transform in time (with conjugate variable $\nu$ ). All four functions are equivalent descriptions of the same channel.

Think of a 2D function indexed by $(\text{delay-or-frequency}, \text{time-or-Doppler})$ . You can choose either variable in each slot, and moving across each axis swaps the two choices via a 1D Fourier transform.

Proof

Setup

Start from the definition $h(\tau, t)$ . Apply $\mathcal{F}_\tau$ (Fourier transform in delay): this is the definition of $H(f, t)$ . Apply $\mathcal{F}_t$ to $h(\tau, t)$ : this is the definition of $h(\tau, \nu)$ .

Commutativity

The two Fourier transforms act on independent variables, hence they commute: $\mathcal{F}_t \mathcal{F}_\tau h(\tau, t) = \mathcal{F}_\tau \mathcal{F}_t h(\tau, t)$ . Both expressions equal $H(f, \nu)$ .

Inverses close the square

The inverse Fourier transforms $\mathcal{F}_\tau^{-1}, \mathcal{F}_t^{-1}$ provide the return arrows. Iterating around the square returns the original function, confirming self-consistency. $\blacksquare$

,

Bello's Four-Function Relationship Diagram — The four system functions are arranged at the corners of a square, with Fourier transforms in delay ( $\tau \leftrightarrow f$ ) along horizontal edges and Fourier transforms in time ( $t \leftrightarrow \nu$ ) along vertical edges. The delay-Doppler function $h(\tau, \nu)$ is diagonally opposite the physical time-varying impulse response $h(\tau, t)$ — it is the view obtained by Fourier-transforming in time but not in delay.

Example: The Four Functions of a Single Path

Consider a single path with complex gain $h_1$ , delay $\tau_1$ , and Doppler $\nu_1$ . Write down all four Bello functions.

Solution

Time-variant impulse response

By definition, the path produces a delayed, Doppler-rotating echo: $h(\tau, t) \;=\; h_1\,e^{j2\pi \nu_1 t}\,\delta(\tau - \tau_1).$ Note the time-dependence is entirely in the phase $e^{j2\pi \nu_1 t}$ .

Frequency-time function

Fourier-transform in $\tau$ : $H(f, t) \;=\; h_1\,e^{j2\pi \nu_1 t}\,e^{-j2\pi f \tau_1}.$ At every frequency $f$ the transfer function is a complex exponential in time — this is the channel's Doppler modulation, visible in the TF picture.

Spreading function

Fourier-transform in $t$ : $h(\tau, \nu) \;=\; h_1\,\delta(\tau - \tau_1)\,\delta(\nu - \nu_1).$ A single Dirac impulse in the 2D plane. This is by far the simplest of the four representations — and it will be our workhorse.

Output Doppler spread function

Either route — $\mathcal{F}_\tau$ of $h(\tau, \nu)$ or $\mathcal{F}_t$ of $H(f, t)$ — gives $H(f, \nu) \;=\; h_1\,e^{-j2\pi f \tau_1}\,\delta(\nu - \nu_1).$ For each Doppler, the transfer function is a pure exponential in $f$ .

When to Use Which Bello Function

Function	Natural for	Pitfall
$h(\tau, t)$	Physical modeling, ray tracing, system simulation	Time-varying — not diagonalized by any fixed basis
$H(f, t)$	OFDM / TF-domain analysis, subcarrier equalization	Loses the sparsity of the underlying channel
$h(\tau, \nu)$	OTFS modulation, DD-domain detection, sensing	Requires finite observation window; continuous-time idealization
$H(f, \nu)$	Doppler radar analysis, ambiguity function	Rarely used directly in communications

Historical Note: Philip Bello's 1963 Paper

1963

P. A. Bello, then at Adcom, Inc., published "Characterization of Randomly Time-Variant Linear Channels" in the June 1963 issue of the IEEE Transactions on Communications Systems — an astonishingly durable piece of work. Bello was motivated by the emerging need to model HF ionospheric and tropospheric scatter links, which were critical for long-range military communication before satellites took over. He did not foresee OFDM, let alone OTFS. What he did do was realize that the time-varying channel was best described by a family of functions related by Fourier duality, and that which member of the family one preferred depended on the physical question at hand.

The delay-Doppler spreading function $h(\tau, \nu)$ was, for Bello, simply one corner of the diagram — useful for studying scatter-produced multipath in troposcatter links. Sixty years later, it would become the signal space of a candidate 6G waveform.

Why $h(\tau, \nu)$ Is Time-Invariant

Notice a subtle but essential point: the spreading function $h(\tau, \nu)$ has no time argument. Yet the physical channel is manifestly time-varying — reflectors move, paths come and go. How is this consistent?

The answer is that $h(\tau, \nu)$ is the time-integrated signature of the channel over the observation window. Within that window, the channel's temporal variation is captured entirely by the Doppler axis $\nu$ . We trade a one-argument time-varying function $h(\tau, t)$ for a two-argument static function $h(\tau, \nu)$ . This is the same trick by which the Fourier transform converts a time-varying signal into a static frequency spectrum — except now we do it for the channel itself, and the "spectrum" is 2D.

Chapter 4 will show that input-output in the DD domain takes the form $y(\tau, \nu) = h(\tau, \nu) \star\star x(\tau, \nu)$ — a 2D convolution with a static kernel. This is precisely the formal sense in which OTFS turns a time-varying channel into a time-invariant one.

Common Mistake: Delay vs Time Ordering

Mistake:

Some references write $h(t, \tau)$ — time-argument first — so that $y(t) = \int h(t, \tau)\,x(t - \tau)\,d\tau$ . This leads to sign confusion in Fourier transforms when converting between Bello functions.

Correction:

Throughout this book, and in the OTFS literature (Hadani 2017, Raviteja 2018), the convention is delay first: $h(\tau, t)$ . Under this convention, the Fourier transform pair is $h(\tau, t) \leftrightarrow h(\tau, \nu)$ via $\nu = \mathcal{F}_t$ , and not $h(t, \tau) \leftrightarrow h(\nu, \tau)$ (which would be a different function). When cross-referencing papers, always check the ordering convention before applying formulas.

Bello's Four System Functions