# Time-varying “impulse response”

According to many references [1,2], the time-varying "impulse response" can compute wireless channel output $y(t)$ at time $t$ using the following expression: $$y(t) = \int h(\tau, t) x(t - \tau) d\tau$$

In both references, they state that this represents the response of the channel at time $t$ to an impulse applied at time $t-\tau$.

It seems reasonable to assume that there is some version of x(t) that involves a delta function that we can apply as an input that returns $h(\tau,t)$ as the output.

Trying: $$x(t) = \delta(t)$$ $$\implies y(t) = \int h(\tau, t) \delta(t - \tau) d\tau = h(t,t) \qquad \text{nope}$$

Trying: $$x(t) = \delta(t-\tau')$$ $$\implies y(t) = \int h(\tau, t) \delta(t - \tau - \tau') d\tau = h(t-\tau',t) \qquad \text{nope, but closer}$$

Is there a way to generate something resembling $h(\tau,t)$ as the output?



References:

 Proakis, Digital Communications, 5th ed, p.832

 Goldsmith, Wireless Communications, 1st ed, p.67

• I think the input should be $x(t)=\delta(t-\tau)$, that is, a delta at time time $\tau$. – MBaz Jun 20 '17 at 18:20
• @MBaz-- how is this different than the second integral? – Robert L. Jun 20 '17 at 18:23
• i wouldn't quote "impulse response" in the context of time-varying. $$h(\tau,t)$$ is the linear system response (at time $t$) of an impulse applied at time $\tau$. it is an impulse response. – robert bristow-johnson Jun 20 '17 at 19:11
• @robertbristow-johnson did you just answer by editing the question? – AlexTP Jun 20 '17 at 19:29
• @robertbristow-johnson if you want to make a comment, just comment instead of changing my post to change the meaning completely. It is extremely rude to put words in my mouth. – Robert L. Jun 20 '17 at 20:01

changing $\tau$ to $\tau'$ for ease of notation, we have $$y(t) = \int h(\tau', t) x(t - \tau') d\tau' \tag{1}$$

Now to get the output $h(\tau, t)$ it suffices to select

$$\color{red}{x(\tau')=\delta(\tau'-t+\tau)}$$

which is an impulse at $\tau'=t-\tau$.

To substitute it in the convolution equation: \begin{align} x(t-\tau')&= x(-\tau'+t)\\ &=\delta(-\tau'-t+\tau+t)\\ &=\delta(\tau-\tau') \end{align}

Hence the result of convolution in $(1)$ is $$y= \int h(\tau', t) \delta(\tau-\tau') d\tau'=h(\tau,t)$$.

which is resulted from the sifting property of Dirac delta.

In writing $h(\tau,t)$ with $t$ is time and $\tau$ is delay, we are in the model that $\tau$ varies "differently" from $t$. Or in other words, they are different notions in spite of the fact that they are both time unit.

Similarly, in this equation $$y(t) = \int h(\tau, t) x(t - \tau) d\tau$$

The time index of $y(t)$ varies "differently" from the time index of $h(\tau,t)$.

It is easier to write $h(\tau, t')$ and fix the measuring moment $t'=t_0$ for $h(\tau,t_0)$ and applying an input $\delta(t)$

$$y(t) = \int h(\tau, t_0) \delta(t - \tau) d\tau = h(t,t_0)$$

The index $t$ of $h(t,t_0)$ plays the role of delay $\tau$ that is exactly the same expression $h(\tau,t_0)$.

Redo the experiment at $t'=t_1, t_2, ...,$ we have the two dimensional time varying impulse reponse $h(\tau,t)$.

• "we are in the model that $\tau$ varies much faster than $t$"-- reference? – Robert L. Jun 20 '17 at 20:06
• @CarlosDanger for an academic vigor, may be "varies differently" sounds better lol. To be serious, I dont find any paradox in your both approaches, just dont be lost in unuseful notations. – AlexTP Jun 20 '17 at 22:04
• @CarlosDanger And if you need some reference for the concept, I suggest "Introduction to spread-spectrum antimultipath techniques and their application to urban digital radio" of Turin that shows how we pratically measure this time varying IR. – AlexTP Jun 20 '17 at 22:14

With $\tau$ being the integration variable in the convolution integral, your output (i.e., the result of the integration) will never be $h(\tau,t)$. Your second approach leads to the correct interpretation: since $h(t-\tau',t)$ is the response to an impulse at time $\tau'$, the function $h(\tau,t)$ is the response at time $t$ to an impulse at time $\tau'=t-\tau$, as already stated in your question.

Note that the function $h(\tau,t)$ as (implicitly) defined by the input-output relation in your question is often referred to as input delay-spread function. More than you probably ever want to know about the characterization of time-varying channels can be found in Fundamentals of Time-Varying Communication Channels by G.Matz and F.Hlawatsch.

The notation here is somewhat sloppy. The whole concept of a time variant impulse response only works if the time scale on which time variation happens is significantly slower than the length of the impulse response. That allows treating the impulse response as piece wise "time invariant" at least during the interval of the convolution.

The time invariant impulse responses has two arguments: one is really fast and one is really slow. Think about cell phone communication in a car: the first argument is nano seconds or thereabouts (GHz) bandwidth), the second is more in milliseconds or seconds (channel varies by the car driving).

Both have units of time, but the second should not be denoted by t, but rather some other symbol since it elapses at a much much smaller rate than the t in x(t) and y(t), which are again in the GHz range

• "The notation here is somewhat sloppy"-- sources or equations please. This was taken straight out of standard textbooks in the field, so if you are going to suggest an improvement, you need to have an authoritative reason for doing so. – Robert L. Jun 20 '17 at 20:04
• Carlos, you don't make us wanna help you. my Linear Systems book with time-variant has been in a box for 3 decades. i'll try to answer just from my memory (or i'll cheat and use Google/Wikipedia, which is something you can do, too). – robert bristow-johnson Jun 20 '17 at 23:03
• i canceled the downvote, Hilmar. the reason i modified his post was because the axiomatic equations weren't right. – robert bristow-johnson Jun 20 '17 at 23:04
• "The whole concept of a time variant impulse response only works if the time scale on which time variation happens is significantly slower than the length of the impulse response." ----- actually, Hilmar, i think that is incorrect. i believe the mathematical theory works even if the time variation of the system is any speed. it need not happen on a time scale significantly slower than the length of the impulse response. – robert bristow-johnson Jun 21 '17 at 4:27

so, for reference (and to remind myself), the LTI convolution is:

\begin{align} y(t) &= h(t) \circledast x(t) \\ &= \int_{-\infty}^{\infty} h(t-u) \, x(u) \, du \\ &= \int_{-\infty}^{\infty} x(t-u) \, h(u) \, du \\ \end{align}

if $x(t) = \delta(t)$ then the output is $y(t)=h(t)$.

if $x(t) = \delta(t-\tau)$ then the output is $y(t)=h(t-\tau)$.

i'm gonna generalize the first integral to linear time-varying, extending $h(t)$ to two arguments. so

\begin{align} h(t) &\leftarrow h(0,t) \\ h(t-\tau) &\leftarrow h(0,t-\tau) = h(\tau,t) \qquad \text{(LTI)} \\ \end{align}

$h(\tau,t)$ is the system response to an impulse applied at time $\tau$. so, the more general convolution integral is

\begin{align} y(t) &= \int_{-\infty}^{\infty} h(u,t) \, x(u) \, du \\ \end{align}

so, i think your first equation is wrong.

• Any explanation why the downvotes? – jojek Jul 18 '18 at 8:38
• at first i wondered. but that was a year ago. there's a difference in definition, but i think what i wrote is more like what was in my Linear System Theory textbook. – robert bristow-johnson Jul 18 '18 at 8:43